@tryhamster/gerbil 1.0.1 → 1.0.2

package/dist/{moonshine-stt-4ojLtMq7.mjs.map → moonshine-stt-BzQRl-BO.mjs.map}

@@ -1 +1 @@
-{"version":3,"file":"moonshine-stt-4ojLtMq7.mjs","names":["MAP_MODE_READ","requiredFeatures: GPUFeatureName[]","requestLimits: Record<string, number>","device: GPUDevice","layout: GPUPipelineLayout | \"auto\"","details: string[]","mismatches: string[]","addSpec: KernelSpec","mulSpec: KernelSpec","sliceLastRowSpec: KernelSpec","meanPoolSpec: KernelSpec","scaleSpec: KernelSpec","softcapSpec: KernelSpec","l2NormSpec: KernelSpec","swigluSpec: KernelSpec","residualRmsnormSpec: KernelSpec","siluSpec: KernelSpec","geluSpec: KernelSpec","geluErfSpec: KernelSpec","addBiasSpec: KernelSpec","sliceColsSpec: KernelSpec","mulColsSpec: KernelSpec","applyRotarySpec: KernelSpec","matmulSpec: KernelSpec","M: number","matmulBiasSpec: KernelSpec","poolMatMulSpec: KernelSpec","clippedMatMulSpec: KernelSpec","MATMUL_BIAS_F16C_SPEC: KernelSpec","matmulInt4Spec: KernelSpec","MATVEC_SPEC: KernelSpec","MATVEC_INT4_SPEC: KernelSpec","GATED_MATVEC_INT4_SPEC: KernelSpec","SWIGLU_GATED_MATVEC_INT4_SPEC: KernelSpec","MATVEC_SUBGROUPS_SPEC: KernelSpec","MATVEC_INT4_SUBGROUPS_SPEC: KernelSpec","SWIGLU_MATVEC_SPEC: KernelSpec","SWIGLU_MATVEC_INT4_SPEC: KernelSpec","DUAL_MATVEC_INT4_SPEC: KernelSpec","ARGMAX_SPEC: KernelSpec","rmsnormSpec: KernelSpec","seq_len: number","DUAL_RMSNORM_SPEC: KernelSpec","layernormSpec: KernelSpec","embeddingSpec: KernelSpec","embeddingInt4Spec: KernelSpec","ropeSpec: KernelSpec","ROPE_INTERLEAVED_SPEC: KernelSpec","mropeSpec: KernelSpec","embedSpliceSpec: KernelSpec","attentionSpec: KernelSpec","crossAttentionSpec: KernelSpec","KV_CACHE_APPEND_PACKED_F16_SPEC: KernelSpec","T: number","ATTENTION_PACKED_F16_SPEC: KernelSpec","softmaxSpec: KernelSpec","causalConv1dSpec: KernelSpec","causalConv1dSiluSpec: KernelSpec","causalConv1dGatedSpec: KernelSpec","sigmoidGateSpec: KernelSpec","mambaSSMSpec: KernelSpec","MAMBA_SSM_F16_SPEC: KernelSpec","kvCacheAppendSpec: KernelSpec","DUAL_KV_CACHE_APPEND_SPEC: KernelSpec","DUAL_KV_CACHE_APPEND_F16_SPEC: KernelSpec","DUAL_KV_CACHE_APPEND_PACKED_F16_SPEC: KernelSpec","KV_CACHE_APPEND_F16_SPEC: KernelSpec","ATTENTION_F16_SPEC: KernelSpec","convStateUpdateSpec: KernelSpec","LAYERNORM_NO_BIAS_SPEC: KernelSpec","tanhSpec: KernelSpec","transposeSpec: KernelSpec","groupnormSpec: KernelSpec","conv1dFullSpec: KernelSpec","convTranspose1dSpec: KernelSpec","snake1dSpec: KernelSpec","fsqDequantSpec: KernelSpec","halfSnake1dSpec: KernelSpec","convTranspose1dDepthwiseSpec: KernelSpec","KERNEL_REGISTRY: Partial<Record<OpType, KernelSpec>>","MAP_MODE_READ","prefillEntry: DispatchEntry","runtimeContext: RuntimeContext","dispatchSizes: Array<[number, number, number]>","vocabSize","dispatchSizes: number[][]","argmaxDispatchSizes: Array<[number, number, number]>","results: Array<{\n      idx: number;\n      nodeId: string;\n      opType: string;\n      tensor: string;\n      sum: number;\n      first4: number[];\n      uniformParams?: number[];\n    }>","uniformParams: number[] | undefined","resolved: Record<string, number[]>","bufferEntries: Array<{ buffer: GPUBuffer }>","fusedEntry: DispatchEntry","fusedSpec: KernelSpec","fusedNode: OpNode","entries: Array<{ buffer: GPUBuffer }>","buffer: GPUBuffer | undefined","_dbPromise: Promise<IDBDatabase | null> | null","_client: OpfsClient | null","stale: string[]","header: Record<string, unknown>","entries: SafetensorEntry[]","metadata: Record<string, string> | null","bs: number[]","ids: number[]","pieces: string[]","byteRun: number[]","chars: number[]","parts: string[]","parts: Array<{ text: string; special: boolean }>","match: RegExpExecArray | null","chunks: string[]","buf: ArrayBuffer | null","_fs: typeof import(\"node:fs\") | null","_path: typeof import(\"node:path\") | null","buf: Buffer","headers: Record<string, string>","chunks: Uint8Array[]","resp: Response","ranges: ByteRange[]","current: ByteRange","uncached: SafetensorEntry[]","resolvedDtype: GraphDType | undefined","store: MutableWeightStore","cachedBuffer: ArrayBuffer | null","data: ArrayBufferView","isFullAttn: boolean","normKeys: string[]","gemmaNormKeys: string[]","N: number","K: number","pleSource: PleSource | undefined","missingTensors: string[]","sizes: Array<[number, number, number]>","enc","p","encK: Float32Array[]","encV: Float32Array[]","resolved: Record<string, number[]>","entries: Array<{ buffer: GPUBuffer }>","encK: Float32Array[]","encV: Float32Array[]","frames: number","tokens: number[]"],"sources":["../src/gpu/device.ts","../src/gpu/kernels/registry.ts","../src/gpu/executor.ts","../src/gpu/gptq-adapter.ts","../src/gpu/idb-cache.ts","../src/gpu/mlx-adapter.ts","../src/gpu/opfs-cache.ts","../src/gpu/safetensors.ts","../src/gpu/tokenizer.ts","../src/gpu/weight-source.ts","../src/gpu/model-loader.ts","../src/gpu/moonshine-executor.ts","../src/gpu/moonshine-stt.ts"],"sourcesContent":["/**\n * WebGPU device abstraction layer.\n *\n * Wraps GPUDevice with helpers for buffer allocation, pipeline compilation,\n * compute dispatch, and readback. All GPU interaction flows through here.\n */\n\n// WebGPU buffer usage flags (numeric for Node.js Dawn polyfill compatibility)\nconst STORAGE = 0x0080;\nconst COPY_SRC = 0x0004;\nconst COPY_DST = 0x0008;\nconst UNIFORM = 0x0040;\nconst MAP_READ = 0x0001;\nconst MAP_MODE_READ = 0x0001;\n\nexport interface GPUContext {\n  /** The underlying WebGPU device. */\n  device: GPUDevice;\n  /** Device limits (max buffer size, workgroup size, etc.). */\n  limits: GPUSupportedLimits;\n  /** Whether f16 is supported as a shader type. */\n  hasF16: boolean;\n  /**\n   * Whether the WebGPU `subgroups` feature is available (Chrome 134+, Safari 26+).\n   * When true, kernels may use `subgroupAdd`/`subgroupBroadcast` etc. (requires\n   * `enable subgroups;` in the shader). Absence falls back to the portable\n   * shared-memory reductions — never assume this is present.\n   */\n  hasSubgroups: boolean;\n  /** True when the WebGPU implementation is WebKit's (Safari, all iOS/iPadOS browsers). */\n  isWebKitWebGPU: boolean;\n  /** Whether the `timestamp-query` feature is available (per-pass GPU timing). Used\n   * only by the env-gated decode profiler; never on the normal inference path. */\n  hasTimestamp: boolean;\n  /** Raw adapter info string for diagnostics. */\n  adapterDescription: string;\n}\n\nexport interface InitGPUOptions {\n  /** Called when the GPU device is lost (e.g. tab backgrounded on iOS). */\n  onDeviceLost?: (reason: string, message: string) => void;\n}\n\n/**\n * Initialize WebGPU and request a device with the features we need.\n *\n * In Node.js, initializes Dawn's WebGPU polyfill if navigator.gpu is absent.\n * In the browser, uses the native WebGPU API directly.\n *\n * Throws a clear error if WebGPU is unavailable.\n */\nexport async function initGPU(options?: InitGPUOptions): Promise<GPUContext> {\n  // Node.js: initialize Dawn polyfill if needed\n  if ((typeof navigator === \"undefined\" || !navigator.gpu) && typeof process !== \"undefined\") {\n    try {\n      const dynamicImport = new Function(\"specifier\", \"return import(specifier)\");\n      const webgpuModule = await dynamicImport(\"webgpu\");\n      const { create } = webgpuModule;\n      if (!(globalThis as any).navigator) {\n        (globalThis as any).navigator = {} as Navigator;\n      }\n      (globalThis as any).navigator.gpu = create([]);\n    } catch {\n      // Dawn not available\n    }\n  }\n\n  if (typeof navigator === \"undefined\" || !navigator.gpu) {\n    throw new Error(\n      \"WebGPU is not available in this environment. \" +\n        \"Requires Chrome 113+, Safari 26+ (iOS 26+), or Firefox 141+.\",\n    );\n  }\n\n  // Detect mobile for power preference\n  const isMobile =\n    typeof navigator !== \"undefined\" && /iPhone|iPad|iPod|Android/i.test(navigator.userAgent);\n\n  const adapter = await navigator.gpu.requestAdapter({\n    powerPreference: isMobile ? \"low-power\" : \"high-performance\",\n  });\n\n  if (!adapter) {\n    throw new Error(\n      \"No WebGPU adapter found. Your GPU may not support WebGPU, \" +\n        \"or another tab may be using it. Try closing other tabs.\",\n    );\n  }\n\n  // Check for f16 support\n  const hasF16 = adapter.features.has(\"shader-f16\");\n\n  // Check for subgroups support (Chrome 134+, Safari 26+). Optional — kernels\n  // fall back to portable shared-memory reductions when absent.\n  const hasSubgroups = adapter.features.has(\"subgroups\" as GPUFeatureName);\n\n  // timestamp-query: per-pass GPU timing for the env-gated decode profiler. Only\n  // requested when GERBIL_PROFILE is set so the normal path is untouched.\n  const wantTimestamp = typeof process !== \"undefined\" && process.env?.GERBIL_PROFILE != null;\n  const hasTimestamp = wantTimestamp && adapter.features.has(\"timestamp-query\");\n\n  // Request device with f16 if available\n  const requiredFeatures: GPUFeatureName[] = [];\n  if (hasF16) {\n    requiredFeatures.push(\"shader-f16\");\n  }\n  if (hasSubgroups) {\n    requiredFeatures.push(\"subgroups\" as GPUFeatureName);\n  }\n  if (hasTimestamp) {\n    requiredFeatures.push(\"timestamp-query\");\n  }\n\n  // Request limits conservatively — don't exceed adapter limits\n  const requestLimits: Record<string, number> = {\n    maxBufferSize: adapter.limits.maxBufferSize,\n    maxStorageBufferBindingSize: adapter.limits.maxStorageBufferBindingSize,\n    maxComputeWorkgroupSizeX: Math.min(256, adapter.limits.maxComputeWorkgroupSizeX),\n    maxComputeWorkgroupSizeY: Math.min(256, adapter.limits.maxComputeWorkgroupSizeY),\n    maxComputeWorkgroupStorageSize: adapter.limits.maxComputeWorkgroupStorageSize,\n    maxStorageBuffersPerShaderStage: adapter.limits.maxStorageBuffersPerShaderStage,\n  };\n\n  let device: GPUDevice;\n  try {\n    device = await adapter.requestDevice({\n      requiredFeatures,\n      requiredLimits: requestLimits,\n    });\n  } catch (e) {\n    // Retry with minimal limits if the first attempt fails (e.g. on low-end iPad)\n    try {\n      device = await adapter.requestDevice({ requiredFeatures });\n    } catch {\n      throw new Error(\n        `Failed to create WebGPU device: ${e instanceof Error ? e.message : String(e)}. ` +\n          \"Your GPU may not meet the minimum requirements.\",\n      );\n    }\n  }\n\n  // Handle device loss\n  device.lost.then((info) => {\n    console.error(`[Gerbil] WebGPU device lost: ${info.reason}`, info.message);\n    options?.onDeviceLost?.(info.reason, info.message);\n  });\n\n  // Surface errors that no error scope caught — without this, failed buffer\n  // allocations and invalid bind groups on Safari are completely silent and\n  // the affected dispatches just no-op (reads as zero logits).\n  try {\n    device.onuncapturederror = (ev: GPUUncapturedErrorEvent) => {\n      console.error(`[Gerbil] Uncaptured GPU error: ${ev.error?.message ?? ev.error}`);\n    };\n  } catch {\n    // not supported by this implementation (e.g. older Dawn polyfill)\n  }\n\n  // Detect WebKit's WebGPU implementation, NOT Apple GPU hardware: Dawn on Apple\n  // Silicon (node-dawn, Chrome on macOS) also reports vendor \"apple\"/arch \"common-*\",\n  // so adapter info cannot distinguish implementations. Only WebKit's implementation\n  // needs the Safari coherence workarounds; hardware-based detection routed desktop\n  // Dawn through them (161 -> 19.7 tok/s regression).\n  let adapterDescription = \"\";\n  try {\n    const info = (adapter as any).info ?? (adapter as any).getInfo?.();\n    if (info) {\n      adapterDescription = `vendor=${info.vendor ?? \"\"} arch=${info.architecture ?? \"\"} device=${info.device ?? \"\"} desc=${info.description ?? \"\"}`;\n    }\n  } catch {\n    // adapter.info not available\n  }\n\n  // All iOS/iPadOS browsers are WebKit; on macOS only Safari is (Chrome/Chromium UAs\n  // contain \"Chrome/\"). Node (Dawn) has no AppleWebKit UA and resolves to false.\n  const ua = typeof navigator !== \"undefined\" && navigator.userAgent ? navigator.userAgent : \"\";\n  const isAppleWebKit = ua.includes(\"AppleWebKit\");\n  const isIOSOrIPadOS =\n    /iPhone|iPad|iPod/.test(ua) ||\n    (isAppleWebKit && typeof navigator !== \"undefined\" && (navigator.maxTouchPoints ?? 0) > 1);\n  const isSafariMac = isAppleWebKit && !ua.includes(\"Chrome/\") && !isIOSOrIPadOS;\n  const isWebKitWebGPU = isIOSOrIPadOS || isSafariMac;\n  if (!adapterDescription) {\n    adapterDescription = `UA-fallback: ${ua.slice(0, 120)}`;\n  }\n\n  return {\n    device,\n    limits: device.limits,\n    hasF16,\n    hasSubgroups,\n    isWebKitWebGPU,\n    hasTimestamp,\n    adapterDescription,\n  };\n}\n\n// ── Buffer helpers ────────────────────────────────────────────────────\n\n/**\n * Create a GPU storage buffer and optionally upload data to it.\n */\nexport function createStorageBuffer(\n  ctx: GPUContext,\n  label: string,\n  sizeBytes: number,\n  data?: ArrayBuffer | ArrayBufferView,\n): GPUBuffer {\n  const usage = STORAGE | COPY_SRC | COPY_DST;\n\n  // Align size to 4 bytes (WebGPU requirement)\n  const alignedSize = Math.ceil(sizeBytes / 4) * 4;\n\n  const buffer = ctx.device.createBuffer({\n    label,\n    size: alignedSize,\n    usage,\n  });\n\n  // Use writeBuffer instead of mappedAtCreation — Safari/WebKit's mappedAtCreation\n  // may not properly persist data after unmap(), causing all-zero weight buffers.\n  if (data) {\n    if (data instanceof ArrayBuffer) {\n      ctx.device.queue.writeBuffer(buffer, 0, data);\n    } else {\n      ctx.device.queue.writeBuffer(buffer, 0, data.buffer, data.byteOffset, data.byteLength);\n    }\n  }\n\n  return buffer;\n}\n\n/**\n * Create a uniform buffer for passing parameters to shaders.\n */\nexport function createUniformBuffer(\n  ctx: GPUContext,\n  label: string,\n  data: ArrayBuffer | ArrayBufferView,\n): GPUBuffer {\n  const byteLength = data instanceof ArrayBuffer ? data.byteLength : data.byteLength;\n\n  // Use STORAGE instead of UNIFORM to bypass Safari/Metal bugs with uniform buffer\n  // visibility across many compute passes. WGSL kernels use var<storage, read> for params.\n  // Keep 16-byte alignment for compatibility.\n  const alignedSize = Math.ceil(byteLength / 16) * 16;\n\n  const buffer = ctx.device.createBuffer({\n    label,\n    size: alignedSize,\n    usage: STORAGE | COPY_SRC | COPY_DST,\n  });\n\n  if (data instanceof ArrayBuffer) {\n    ctx.device.queue.writeBuffer(buffer, 0, data);\n  } else {\n    ctx.device.queue.writeBuffer(buffer, 0, data.buffer, data.byteOffset, data.byteLength);\n  }\n\n  return buffer;\n}\n\n/**\n * Create a staging buffer for reading GPU results back to CPU.\n */\nexport function createReadbackBuffer(ctx: GPUContext, label: string, sizeBytes: number): GPUBuffer {\n  const alignedSize = Math.ceil(sizeBytes / 4) * 4;\n  return ctx.device.createBuffer({\n    label,\n    size: alignedSize,\n    usage: MAP_READ | COPY_DST,\n  });\n}\n\n/**\n * Update a uniform buffer's contents.\n */\nexport function writeUniformBuffer(\n  ctx: GPUContext,\n  buffer: GPUBuffer,\n  data: ArrayBuffer | ArrayBufferView,\n): void {\n  if (data instanceof ArrayBuffer) {\n    ctx.device.queue.writeBuffer(buffer, 0, data);\n  } else {\n    ctx.device.queue.writeBuffer(buffer, 0, data.buffer, data.byteOffset, data.byteLength);\n  }\n}\n\n// ── Pipeline helpers ──────────────────────────────────────────────────\n\n/**\n * Per-device pipeline cache. Pipelines are bound to the GPUDevice that created\n * them — using a pipeline from device1 with device2 causes WebGPU validation\n * errors. WeakMap ensures the cache is GC'd when the device is destroyed.\n */\nconst perDeviceCache = new WeakMap<GPUDevice, Map<string, GPUComputePipeline>>();\n\n/** Get or initialize the pipeline cache for a specific device. */\nfunction getPipelineCache(device: GPUDevice): Map<string, GPUComputePipeline> {\n  let cache = perDeviceCache.get(device);\n  if (!cache) {\n    cache = new Map();\n    perDeviceCache.set(device, cache);\n  }\n  return cache;\n}\n\n/**\n * Get or create a compute pipeline from WGSL source code.\n */\nexport function getOrCreatePipeline(\n  ctx: GPUContext,\n  label: string,\n  wgslCode: string,\n  entryPoint: string = \"main\",\n  bindGroupLayoutEntries?: GPUBindGroupLayoutEntry[],\n): GPUComputePipeline {\n  const pipelineCache = getPipelineCache(ctx.device);\n  const cacheKey = `${wgslCode}::${entryPoint}`;\n  const cached = pipelineCache.get(cacheKey);\n  if (cached) return cached;\n\n  let code = wgslCode;\n\n  // Only prepend f16 enable if the shader actually uses f16 types.\n  // Enabling it unconditionally can trigger buggy codepaths in Safari/WebKit's\n  // WGSL→Metal compiler even when no f16 types are used.\n  const usesF16 = code.includes(\": f16\") || code.includes(\"<f16>\");\n  if (ctx.hasF16 && usesF16 && !code.includes(\"enable f16\")) {\n    code = `enable f16;\\n\\n${code}`;\n  }\n\n  const shaderModule = ctx.device.createShaderModule({\n    label: `${label}_shader`,\n    code,\n  });\n\n  let layout: GPUPipelineLayout | \"auto\" = \"auto\";\n  if (bindGroupLayoutEntries) {\n    const bindGroupLayout = ctx.device.createBindGroupLayout({\n      label: `${label}_layout`,\n      entries: bindGroupLayoutEntries,\n    });\n    layout = ctx.device.createPipelineLayout({\n      label: `${label}_pipeline_layout`,\n      bindGroupLayouts: [bindGroupLayout],\n    });\n  }\n\n  const pipeline = ctx.device.createComputePipeline({\n    label,\n    layout,\n    compute: {\n      module: shaderModule,\n      entryPoint,\n    },\n  });\n\n  pipelineCache.set(cacheKey, pipeline);\n  return pipeline;\n}\n\n/**\n * Clear the pipeline cache for a specific device, or no-op if called without\n * a device (WeakMap handles cleanup automatically when the device is GC'd).\n */\nexport function clearPipelineCache(device?: GPUDevice): void {\n  if (device) {\n    perDeviceCache.delete(device);\n  }\n}\n\n// ── Bind group helpers ────────────────────────────────────────────────\n\nexport interface BindGroupEntry {\n  buffer: GPUBuffer;\n  offset?: number;\n  size?: number;\n}\n\n/**\n * Create a bind group from a list of buffers.\n */\nexport function createBindGroup(\n  ctx: GPUContext,\n  pipeline: GPUComputePipeline,\n  entries: BindGroupEntry[],\n  label?: string,\n): GPUBindGroup {\n  return ctx.device.createBindGroup({\n    label,\n    layout: pipeline.getBindGroupLayout(0),\n    entries: entries.map((entry, i) => ({\n      binding: i,\n      resource: {\n        buffer: entry.buffer,\n        offset: entry.offset ?? 0,\n        size: entry.size ?? entry.buffer.size,\n      },\n    })),\n  });\n}\n\n// ── Dispatch helpers ──────────────────────────────────────────────────\n\n/**\n * Encode and submit a single compute dispatch.\n * Useful for testing; the executor batches dispatches in practice.\n */\nexport function dispatchOnce(\n  ctx: GPUContext,\n  pipeline: GPUComputePipeline,\n  bindGroup: GPUBindGroup,\n  workgroupCountX: number,\n  workgroupCountY: number = 1,\n  workgroupCountZ: number = 1,\n): void {\n  const encoder = ctx.device.createCommandEncoder();\n  const pass = encoder.beginComputePass();\n  pass.setPipeline(pipeline);\n  pass.setBindGroup(0, bindGroup);\n  pass.dispatchWorkgroups(workgroupCountX, workgroupCountY, workgroupCountZ);\n  pass.end();\n  ctx.device.queue.submit([encoder.finish()]);\n}\n\n// ── Readback helpers ──────────────────────────────────────────────────\n\n/**\n * Copy data from a storage buffer to a readback buffer, then map and return\n * the contents as a Float32Array.\n *\n * This is the primary way to get results (logits) back to CPU for sampling.\n */\nexport async function readbackFloats(\n  ctx: GPUContext,\n  source: GPUBuffer,\n  readback: GPUBuffer,\n  byteOffset: number = 0,\n  byteLength?: number,\n): Promise<Float32Array> {\n  const length = byteLength ?? source.size - byteOffset;\n\n  const encoder = ctx.device.createCommandEncoder();\n  encoder.copyBufferToBuffer(source, byteOffset, readback, 0, length);\n  ctx.device.queue.submit([encoder.finish()]);\n\n  await readback.mapAsync(MAP_MODE_READ, 0, length);\n  const mapped = readback.getMappedRange(0, length);\n  const result = new Float32Array(mapped.slice(0));\n  readback.unmap();\n\n  return result;\n}\n\n/**\n * Read back raw bytes from a GPU buffer.\n */\nexport async function readbackBytes(\n  ctx: GPUContext,\n  source: GPUBuffer,\n  readback: GPUBuffer,\n  byteOffset: number = 0,\n  byteLength?: number,\n): Promise<ArrayBuffer> {\n  const length = byteLength ?? source.size - byteOffset;\n\n  const encoder = ctx.device.createCommandEncoder();\n  encoder.copyBufferToBuffer(source, byteOffset, readback, 0, length);\n  ctx.device.queue.submit([encoder.finish()]);\n\n  await readback.mapAsync(MAP_MODE_READ, 0, length);\n  const mapped = readback.getMappedRange(0, length);\n  const result = mapped.slice(0);\n  readback.unmap();\n\n  return result;\n}\n\n// ── Memory helpers ────────────────────────────────────────────────────\n\n/**\n * Destroy a list of GPU buffers and release their memory.\n */\nexport function destroyBuffers(buffers: GPUBuffer[]): void {\n  for (const buf of buffers) {\n    buf.destroy();\n  }\n}\n\n// ── GPU diagnostics ──────────────────────────────────────────────────\n\nexport interface GPUDiagnosticResult {\n  /** Whether the basic buffer upload → readback round-trip works. */\n  bufferIntegrity: boolean;\n  /** Whether a trivial compute shader executes correctly. */\n  computeWorks: boolean;\n  /** Whether shared memory + workgroupBarrier() works. */\n  sharedMemoryWorks: boolean;\n  /** Detailed messages for each test. */\n  details: string[];\n}\n\n/**\n * Run GPU diagnostics to verify basic WebGPU operations work correctly.\n * Useful for isolating Safari/WebKit-specific issues.\n */\nexport async function verifyGPU(ctx: GPUContext): Promise<GPUDiagnosticResult> {\n  const details: string[] = [];\n  let bufferIntegrity = false;\n  let computeWorks = false;\n  let sharedMemoryWorks = false;\n\n  // Report device limits\n  try {\n    details.push(\n      `Limits: wgSize=${ctx.limits.maxComputeWorkgroupSizeX}, ` +\n        `wgStorage=${ctx.limits.maxComputeWorkgroupStorageSize}, ` +\n        `maxBuf=${(ctx.limits.maxBufferSize / 1024 / 1024).toFixed(0)}MB, ` +\n        `f16=${ctx.hasF16}, subgroups=${ctx.hasSubgroups}`,\n    );\n  } catch {}\n\n  // Test 1: Buffer upload via writeBuffer → readback\n  try {\n    const testData = new Float32Array([1.0, 2.0, 3.0, 4.0, -1.5, 0.0, 100.0, 0.001]);\n    const srcBuf = createStorageBuffer(ctx, \"diag_src\", testData.byteLength, testData);\n    const readBuf = createReadbackBuffer(ctx, \"diag_read\", testData.byteLength);\n    const result = await readbackFloats(ctx, srcBuf, readBuf, 0, testData.byteLength);\n\n    let match = true;\n    for (let i = 0; i < testData.length; i++) {\n      if (Math.abs(result[i] - testData[i]) > 1e-6) {\n        details.push(`Buffer mismatch at [${i}]: expected ${testData[i]}, got ${result[i]}`);\n        match = false;\n      }\n    }\n    bufferIntegrity = match;\n    details.push(\n      bufferIntegrity ? \"✓ Buffer upload/readback OK\" : \"✗ Buffer upload/readback FAILED\",\n    );\n    srcBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ Buffer test error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // ── Isolation tests ──────────────────────────────────────────────\n  // Test 2a PASSED (single encoder, wg=1, no input, onSubmittedWorkDone).\n  // Test 2b FAILED (separate encoders, wg=4, has input, no wait).\n  // Need to isolate: is it (A) separate encoders, (B) input binding, or (C) workgroup size?\n  //\n  // Matrix of tests — each changes ONE variable from the passing baseline:\n  //   Baseline: single encoder, wg=1, no input, with onSubmittedWorkDone → PASSES\n  //   Test A: SEPARATE encoders, wg=1, no input, no wait → tests queue ordering\n  //   Test B: single encoder, wg=1, HAS input, with wait → tests input binding\n  //   Test C: single encoder, wg=4, no input, with wait → tests workgroup size\n  //   Test D: single encoder, wg=256 + shared mem, no input, with wait → tests shared memory\n\n  /** Helper: run a compute shader and read back result using a SINGLE command encoder + onSubmittedWorkDone. */\n  const runSingleEncoder = async (\n    label: string,\n    code: string,\n    bindings: GPUBuffer[],\n    readBuf: GPUBuffer,\n    outBuf: GPUBuffer,\n    readBytes: number,\n    workgroups: number = 1,\n  ): Promise<Float32Array> => {\n    const pipeline = getOrCreatePipeline(ctx, label, code, \"main\");\n    const bg = createBindGroup(\n      ctx,\n      pipeline,\n      bindings.map((b) => ({ buffer: b })),\n      `${label}_bg`,\n    );\n    const encoder = ctx.device.createCommandEncoder();\n    const pass = encoder.beginComputePass();\n    pass.setPipeline(pipeline);\n    pass.setBindGroup(0, bg);\n    pass.dispatchWorkgroups(workgroups);\n    pass.end();\n    encoder.copyBufferToBuffer(outBuf, 0, readBuf, 0, readBytes);\n    ctx.device.queue.submit([encoder.finish()]);\n    if (ctx.device.queue.onSubmittedWorkDone) {\n      await ctx.device.queue.onSubmittedWorkDone();\n    }\n    await readBuf.mapAsync(MAP_MODE_READ, 0, readBytes);\n    const mapped = readBuf.getMappedRange(0, readBytes);\n    const result = new Float32Array(mapped.slice(0));\n    readBuf.unmap();\n    return result;\n  };\n\n  /** Helper: run compute with SEPARATE encoders (like dispatchOnce + readbackFloats). */\n  const runSeparateEncoders = async (\n    label: string,\n    code: string,\n    bindings: GPUBuffer[],\n    readBuf: GPUBuffer,\n    outBuf: GPUBuffer,\n    readBytes: number,\n    workgroups: number = 1,\n  ): Promise<Float32Array> => {\n    const pipeline = getOrCreatePipeline(ctx, label, code, \"main\");\n    const bg = createBindGroup(\n      ctx,\n      pipeline,\n      bindings.map((b) => ({ buffer: b })),\n      `${label}_bg`,\n    );\n    // Encoder 1: compute dispatch\n    dispatchOnce(ctx, pipeline, bg, workgroups);\n    // Encoder 2: copy + readback\n    return readbackFloats(ctx, outBuf, readBuf, 0, readBytes);\n  };\n\n  // Test A: Separate encoders, wg=1, no input (isolates queue ordering)\n  try {\n    const shader = `\n      @group(0) @binding(0) var<storage, read_write> out: array<f32>;\n      @compute @workgroup_size(1)\n      fn main() { out[0] = 42.0; }\n    `;\n    const outBuf = createStorageBuffer(ctx, \"dA_out\", 4);\n    const readBuf = createReadbackBuffer(ctx, \"dA_read\", 4);\n    const result = await runSeparateEncoders(\"dA\", shader, [outBuf], readBuf, outBuf, 4);\n    const ok = Math.abs(result[0] - 42.0) < 1e-3;\n    details.push(\n      ok\n        ? `✓ A: separate encoders OK (${result[0]})`\n        : `✗ A: separate encoders FAILED (got ${result[0]})`,\n    );\n    outBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ A error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test B: Single encoder, wg=1, HAS input (isolates input binding)\n  try {\n    const shader = `\n      @group(0) @binding(0) var<storage, read> src: array<f32>;\n      @group(0) @binding(1) var<storage, read_write> dst: array<f32>;\n      @compute @workgroup_size(1)\n      fn main() { dst[0] = src[0] * 2.0 + 1.0; }\n    `;\n    const input = new Float32Array([3.0]);\n    const inBuf = createStorageBuffer(ctx, \"dB_in\", 4, input);\n    const outBuf = createStorageBuffer(ctx, \"dB_out\", 4);\n    const readBuf = createReadbackBuffer(ctx, \"dB_read\", 4);\n    const result = await runSingleEncoder(\"dB\", shader, [inBuf, outBuf], readBuf, outBuf, 4);\n    const ok = Math.abs(result[0] - 7.0) < 1e-3;\n    details.push(\n      ok ? `✓ B: input binding OK (${result[0]})` : `✗ B: input binding FAILED (got ${result[0]})`,\n    );\n    computeWorks = ok;\n    inBuf.destroy();\n    outBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ B error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test C: Single encoder, wg=4, no input (isolates workgroup size)\n  try {\n    const shader = `\n      @group(0) @binding(0) var<storage, read_write> out: array<f32>;\n      @compute @workgroup_size(4)\n      fn main(@builtin(local_invocation_id) lid: vec3u) {\n        out[lid.x] = f32(lid.x) + 1.0;\n      }\n    `;\n    const outBuf = createStorageBuffer(ctx, \"dC_out\", 16);\n    const readBuf = createReadbackBuffer(ctx, \"dC_read\", 16);\n    const result = await runSingleEncoder(\"dC\", shader, [outBuf], readBuf, outBuf, 16);\n    const ok = Math.abs(result[0] - 1.0) < 1e-3 && Math.abs(result[3] - 4.0) < 1e-3;\n    details.push(\n      ok\n        ? `✓ C: wg=4 OK (${Array.from(result).join(\",\")})`\n        : `✗ C: wg=4 FAILED (${Array.from(result).join(\",\")})`,\n    );\n    outBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ C error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test D: Single encoder, wg=4 + shared memory (isolates shared mem)\n  // Uses a small workgroup (4 threads) to avoid 256-thread driver issues in diagnostics.\n  // Each thread writes to shared memory, barrier, thread 0 sums and writes to output.\n  try {\n    const shader = `\n      var<workgroup> smem: array<f32, 4>;\n      @group(0) @binding(0) var<storage, read_write> dst: array<f32>;\n      @compute @workgroup_size(4)\n      fn main(@builtin(local_invocation_id) lid: vec3u) {\n        smem[lid.x] = f32(lid.x) + 1.0;\n        workgroupBarrier();\n        if (lid.x == 0u) {\n          dst[0] = smem[0] + smem[1] + smem[2] + smem[3];\n        }\n      }\n    `;\n    const outBuf = createStorageBuffer(ctx, \"dD_out\", 4);\n    const readBuf = createReadbackBuffer(ctx, \"dD_read\", 4);\n    const result = await runSingleEncoder(\"dD\", shader, [outBuf], readBuf, outBuf, 4);\n    sharedMemoryWorks = Math.abs(result[0] - 10.0) < 1e-3;\n    details.push(\n      sharedMemoryWorks\n        ? `✓ D: shared mem OK (${result[0]})`\n        : `✗ D: shared mem FAILED (got ${result[0]})`,\n    );\n    outBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ D error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test E: Separate encoders + input + wg=4 (the full failing combo, to confirm)\n  try {\n    const shader = `\n      @group(0) @binding(0) var<storage, read> src: array<f32>;\n      @group(0) @binding(1) var<storage, read_write> dst: array<f32>;\n      @compute @workgroup_size(4)\n      fn main(@builtin(global_invocation_id) gid: vec3u) {\n        if (gid.x < 4u) { dst[gid.x] = src[gid.x] * 2.0 + 1.0; }\n      }\n    `;\n    const input = new Float32Array([3.0, 7.0, -2.0, 0.5]);\n    const inBuf = createStorageBuffer(ctx, \"dE_in\", 16, input);\n    const outBuf = createStorageBuffer(ctx, \"dE_out\", 16);\n    const readBuf = createReadbackBuffer(ctx, \"dE_read\", 16);\n    const result = await runSeparateEncoders(\"dE\", shader, [inBuf, outBuf], readBuf, outBuf, 16);\n    const ok = Math.abs(result[0] - 7.0) < 1e-3 && Math.abs(result[1] - 15.0) < 1e-3;\n    details.push(\n      ok ? `✓ E: full combo OK` : `✗ E: full combo FAILED (${Array.from(result).join(\",\")})`,\n    );\n    inBuf.destroy();\n    outBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ E error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test F: pack2x16float / unpack2x16float round-trip (tests the packed-f16 KV path)\n  // Packs pairs of f32 → u32 via pack2x16float, then unpacks back via unpack2x16float.\n  // If this fails on Safari/Metal, the packed KV cache will produce gibberish.\n  try {\n    const shader = `\n      @group(0) @binding(0) var<storage, read> src: array<f32>;\n      @group(0) @binding(1) var<storage, read_write> dst: array<f32>;\n      @compute @workgroup_size(1)\n      fn main() {\n        // Pack pairs of f32 into u32, then unpack back\n        let packed0 = pack2x16float(vec2f(src[0], src[1]));\n        let packed1 = pack2x16float(vec2f(src[2], src[3]));\n        let packed2 = pack2x16float(vec2f(src[4], src[5]));\n        let packed3 = pack2x16float(vec2f(src[6], src[7]));\n\n        let u0 = unpack2x16float(packed0);\n        let u1 = unpack2x16float(packed1);\n        let u2 = unpack2x16float(packed2);\n        let u3 = unpack2x16float(packed3);\n\n        dst[0] = u0.x; dst[1] = u0.y;\n        dst[2] = u1.x; dst[3] = u1.y;\n        dst[4] = u2.x; dst[5] = u2.y;\n        dst[6] = u3.x; dst[7] = u3.y;\n      }\n    `;\n    // Values chosen to cover: positive, negative, zero, small, >1\n    const input = new Float32Array([1.0, -0.5, 0.0, 3.14, 0.001, -2.0, 100.0, 0.25]);\n    const inBuf = createStorageBuffer(ctx, \"dF_in\", input.byteLength, input);\n    const outBuf = createStorageBuffer(ctx, \"dF_out\", input.byteLength);\n    const readBuf = createReadbackBuffer(ctx, \"dF_read\", input.byteLength);\n    const result = await runSingleEncoder(\n      \"dF\",\n      shader,\n      [inBuf, outBuf],\n      readBuf,\n      outBuf,\n      input.byteLength,\n    );\n    // f16 has limited precision, so use larger tolerance\n    let ok = true;\n    const mismatches: string[] = [];\n    for (let i = 0; i < input.length; i++) {\n      const tol = Math.max(Math.abs(input[i]) * 0.01, 0.001); // 1% or 0.001\n      if (Math.abs(result[i] - input[i]) > tol) {\n        ok = false;\n        mismatches.push(`[${i}]:${input[i]}→${result[i]}`);\n      }\n    }\n    details.push(\n      ok\n        ? `✓ F: pack/unpack2x16float OK`\n        : `✗ F: pack/unpack2x16float FAILED (${mismatches.join(\", \")})`,\n    );\n    inBuf.destroy();\n    outBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ F error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test G: Packed KV-like pattern — pack into array<u32>, read back with elem indexing\n  // Mimics the actual KV cache append + attention load_k/load_v pattern.\n  try {\n    const shader = `\n      @group(0) @binding(0) var<storage, read> src: array<f32>;\n      @group(0) @binding(1) var<storage, read_write> packed: array<u32>;\n      @group(0) @binding(2) var<storage, read_write> dst: array<f32>;\n      @compute @workgroup_size(1)\n      fn main() {\n        // Step 1: Pack f32 pairs into u32 array (like KVCacheAppend)\n        packed[0] = pack2x16float(vec2f(src[0], src[1]));\n        packed[1] = pack2x16float(vec2f(src[2], src[3]));\n        packed[2] = pack2x16float(vec2f(src[4], src[5]));\n        packed[3] = pack2x16float(vec2f(src[6], src[7]));\n\n        // Step 2: Read back using element-level indexing (like load_k/load_v)\n        for (var i: u32 = 0u; i < 8u; i += 1u) {\n          let pair = unpack2x16float(packed[i >> 1u]);\n          if ((i & 1u) == 0u) {\n            dst[i] = pair.x;\n          } else {\n            dst[i] = pair.y;\n          }\n        }\n      }\n    `;\n    const input = new Float32Array([1.5, -0.75, 2.0, 0.0, -1.0, 0.5, 3.0, -3.0]);\n    const inBuf = createStorageBuffer(ctx, \"dG_in\", input.byteLength, input);\n    const packedBuf = createStorageBuffer(ctx, \"dG_packed\", 16); // 4 u32 = 16 bytes\n    const outBuf = createStorageBuffer(ctx, \"dG_out\", input.byteLength);\n    const readBuf = createReadbackBuffer(ctx, \"dG_read\", input.byteLength);\n    const result = await runSingleEncoder(\n      \"dG\",\n      shader,\n      [inBuf, packedBuf, outBuf],\n      readBuf,\n      outBuf,\n      input.byteLength,\n    );\n    let ok = true;\n    const mismatches: string[] = [];\n    for (let i = 0; i < input.length; i++) {\n      const tol = Math.max(Math.abs(input[i]) * 0.01, 0.001);\n      if (Math.abs(result[i] - input[i]) > tol) {\n        ok = false;\n        mismatches.push(`[${i}]:${input[i]}→${result[i]}`);\n      }\n    }\n    details.push(\n      ok\n        ? `✓ G: packed KV round-trip OK`\n        : `✗ G: packed KV round-trip FAILED (${mismatches.join(\", \")})`,\n    );\n    inBuf.destroy();\n    packedBuf.destroy();\n    outBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ G error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test H: Tree reduction with 256 threads + shared memory (RMSNorm/softmax pattern)\n  // 256 threads each write a value, then tree-reduce to sum. Tests the exact\n  // shared memory reduction pattern used in RMSNorm, softmax, and dot products.\n  try {\n    const shader = `\n      var<workgroup> smem: array<f32, 256>;\n      @group(0) @binding(0) var<storage, read> src: array<f32>;\n      @group(0) @binding(1) var<storage, read_write> dst: array<f32>;\n      @compute @workgroup_size(256)\n      fn main(@builtin(local_invocation_id) lid: vec3u) {\n        let tid = lid.x;\n        smem[tid] = src[tid] * src[tid];\n        workgroupBarrier();\n        var stride: u32 = 128u;\n        while (stride > 0u) {\n          if (tid < stride) {\n            smem[tid] += smem[tid + stride];\n          }\n          workgroupBarrier();\n          stride /= 2u;\n        }\n        // Thread 0 writes sum, all threads write their squared input for verification\n        dst[tid] = src[tid] * src[tid];\n        if (tid == 0u) { dst[256] = smem[0]; }\n      }\n    `;\n    // Input: 1..256, expected sum of squares = n(n+1)(2n+1)/6 = 5559680\n    const input = new Float32Array(256);\n    for (let i = 0; i < 256; i++) input[i] = i + 1;\n    const expectedSum = (256 * 257 * 513) / 6; // 5559680\n    const inBuf = createStorageBuffer(ctx, \"dH_in\", input.byteLength, input);\n    const outBuf = createStorageBuffer(ctx, \"dH_out\", 257 * 4);\n    const readBuf = createReadbackBuffer(ctx, \"dH_read\", 257 * 4);\n    const result = await runSingleEncoder(\"dH\", shader, [inBuf, outBuf], readBuf, outBuf, 257 * 4);\n    const sum = result[256];\n    const ok = Math.abs(sum - expectedSum) / expectedSum < 0.001;\n    details.push(\n      ok\n        ? `✓ H: 256-thread tree reduce OK (${sum})`\n        : `✗ H: 256-thread tree reduce FAILED (got ${sum}, expected ${expectedSum})`,\n    );\n    inBuf.destroy();\n    outBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ H error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test I: INT4 dequantize pattern (bit manipulation + multiply-add)\n  // Tests the exact u32→nibble extraction used in MatMulInt4 and SwiGLU INT4 kernels.\n  try {\n    const shader = `\n      @group(0) @binding(0) var<storage, read> packed: array<u32>;\n      @group(0) @binding(1) var<storage, read> scales: array<f32>;\n      @group(0) @binding(2) var<storage, read> zeros: array<f32>;\n      @group(0) @binding(3) var<storage, read_write> dst: array<f32>;\n      @compute @workgroup_size(1)\n      fn main() {\n        // Unpack 8 nibbles from one u32, dequantize: (nibble - zero) * scale\n        let word = packed[0];\n        let scale = scales[0];\n        let zero = zeros[0];\n        for (var i: u32 = 0u; i < 8u; i += 1u) {\n          let nibble = (word >> (i * 4u)) & 0xFu;\n          dst[i] = (f32(nibble) - zero) * scale;\n        }\n      }\n    `;\n    // Pack nibbles 0,1,2,...,7 into one u32: 0x76543210\n    const packedData = new Uint32Array([0x76543210]);\n    const scaleData = new Float32Array([0.5]);\n    const zeroData = new Float32Array([2.0]);\n    // Expected: (0-2)*0.5=-1, (1-2)*0.5=-0.5, (2-2)*0.5=0, (3-2)*0.5=0.5, ...\n    const expected = [0, 1, 2, 3, 4, 5, 6, 7].map((n) => (n - 2.0) * 0.5);\n\n    const packedBuf = createStorageBuffer(ctx, \"dI_packed\", 4, packedData);\n    const scaleBuf = createStorageBuffer(ctx, \"dI_scale\", 4, scaleData);\n    const zeroBuf = createStorageBuffer(ctx, \"dI_zero\", 4, zeroData);\n    const outBuf = createStorageBuffer(ctx, \"dI_out\", 32);\n    const readBuf = createReadbackBuffer(ctx, \"dI_read\", 32);\n    const result = await runSingleEncoder(\n      \"dI\",\n      shader,\n      [packedBuf, scaleBuf, zeroBuf, outBuf],\n      readBuf,\n      outBuf,\n      32,\n    );\n    let ok = true;\n    const mismatches: string[] = [];\n    for (let i = 0; i < 8; i++) {\n      if (Math.abs(result[i] - expected[i]) > 0.001) {\n        ok = false;\n        mismatches.push(`[${i}]:${expected[i]}→${result[i]}`);\n      }\n    }\n    details.push(\n      ok ? `✓ I: INT4 dequant OK` : `✗ I: INT4 dequant FAILED (${mismatches.join(\", \")})`,\n    );\n    packedBuf.destroy();\n    scaleBuf.destroy();\n    zeroBuf.destroy();\n    outBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ I error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test J: Multi-dispatch chain (tests sequential dispatch ordering)\n  // Runs 3 dispatches in sequence within one command encoder: A→B→C.\n  // If dispatch ordering is broken, intermediate results will be wrong.\n  try {\n    const shaderAdd = `\n      @group(0) @binding(0) var<storage, read> a: array<f32>;\n      @group(0) @binding(1) var<storage, read_write> b: array<f32>;\n      @compute @workgroup_size(4)\n      fn main(@builtin(global_invocation_id) gid: vec3u) {\n        if (gid.x < 4u) { b[gid.x] = a[gid.x] + b[gid.x]; }\n      }\n    `;\n    const shaderMul = `\n      @group(0) @binding(0) var<storage, read_write> x: array<f32>;\n      @compute @workgroup_size(4)\n      fn main(@builtin(global_invocation_id) gid: vec3u) {\n        if (gid.x < 4u) { x[gid.x] = x[gid.x] * 2.0; }\n      }\n    `;\n    // Chain: buf = [1,2,3,4], add [10,20,30,40] → [11,22,33,44], mul ×2 → [22,44,66,88], add again → [32,64,96,128]... actually simpler:\n    // Start: buf = [0,0,0,0]\n    // Dispatch 1: buf += src ([1,2,3,4])\n    // Dispatch 2: buf *= 2\n    // Dispatch 3: buf += src again\n    // Expected: ([1,2,3,4] * 2) + [1,2,3,4] = [3,6,9,12]\n    const srcData = new Float32Array([1, 2, 3, 4]);\n    const srcBuf = createStorageBuffer(ctx, \"dJ_src\", 16, srcData);\n    const bufBuf = createStorageBuffer(ctx, \"dJ_buf\", 16, new Float32Array(4)); // zeros\n    const readBuf = createReadbackBuffer(ctx, \"dJ_read\", 16);\n\n    const addPipeline = getOrCreatePipeline(ctx, \"dJ_add\", shaderAdd, \"main\");\n    const mulPipeline = getOrCreatePipeline(ctx, \"dJ_mul\", shaderMul, \"main\");\n\n    const addBg = createBindGroup(\n      ctx,\n      addPipeline,\n      [{ buffer: srcBuf }, { buffer: bufBuf }],\n      \"dJ_add_bg\",\n    );\n    const mulBg = createBindGroup(ctx, mulPipeline, [{ buffer: bufBuf }], \"dJ_mul_bg\");\n\n    const encoder = ctx.device.createCommandEncoder();\n    const pass = encoder.beginComputePass();\n    pass.setPipeline(addPipeline);\n    pass.setBindGroup(0, addBg);\n    pass.dispatchWorkgroups(1); // buf = [1,2,3,4]\n    pass.setPipeline(mulPipeline);\n    pass.setBindGroup(0, mulBg);\n    pass.dispatchWorkgroups(1); // buf = [2,4,6,8]\n    pass.setPipeline(addPipeline);\n    pass.setBindGroup(0, addBg);\n    pass.dispatchWorkgroups(1); // buf = [3,6,9,12]\n    pass.end();\n    encoder.copyBufferToBuffer(bufBuf, 0, readBuf, 0, 16);\n    ctx.device.queue.submit([encoder.finish()]);\n    if (ctx.device.queue.onSubmittedWorkDone) {\n      await ctx.device.queue.onSubmittedWorkDone();\n    }\n    await readBuf.mapAsync(MAP_MODE_READ, 0, 16);\n    const mapped = readBuf.getMappedRange(0, 16);\n    const result = new Float32Array(mapped.slice(0));\n    readBuf.unmap();\n\n    const expected = [3, 6, 9, 12];\n    let ok = true;\n    for (let i = 0; i < 4; i++) {\n      if (Math.abs(result[i] - expected[i]) > 0.01) ok = false;\n    }\n    details.push(\n      ok\n        ? `✓ J: multi-dispatch chain OK (${Array.from(result).join(\",\")})`\n        : `✗ J: multi-dispatch chain FAILED (got ${Array.from(result).join(\",\")}, expected ${expected.join(\",\")})`,\n    );\n    srcBuf.destroy();\n    bufBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ J error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test K: exp() clamp safety (Metal fast-math NaN check)\n  // Tests that exp(large_negative) returns ~0 and not NaN.\n  try {\n    const shader = `\n      @group(0) @binding(0) var<storage, read_write> dst: array<f32>;\n      @compute @workgroup_size(1)\n      fn main() {\n        dst[0] = exp(max(-1000.0, -80.0));      // should be exp(-80) ≈ 1.8e-35\n        dst[1] = exp(max(-1e30, -80.0));         // should be exp(-80) ≈ 1.8e-35\n        dst[2] = exp(-10.0);                     // should be ~4.54e-5\n        dst[3] = 1.0 / (1.0 + exp(max(-50.0, -80.0)));  // sigmoid(50) ≈ 1.0\n        // Direct exp without clamp — what does Metal actually do?\n        dst[4] = exp(-100.0);                    // should be ~3.7e-44 or 0\n        dst[5] = exp(-1000.0);                   // might be NaN on Metal!\n        dst[6] = exp(-1e10);                     // might be NaN on Metal!\n        // Check for NaN: NaN != NaN\n        let v5 = exp(-1000.0);\n        if (v5 != v5) { dst[7] = -999.0; } else { dst[7] = v5; }\n      }\n    `;\n    const outBuf = createStorageBuffer(ctx, \"dK_out\", 32);\n    const readBuf = createReadbackBuffer(ctx, \"dK_read\", 32);\n    const result = await runSingleEncoder(\"dK\", shader, [outBuf], readBuf, outBuf, 32);\n    const clampedOk = result[0] > 0 && result[0] < 1e-30 && result[1] > 0 && result[1] < 1e-30;\n    const normalOk = Math.abs(result[2] - Math.exp(-10)) < 1e-8;\n    const sigmoidOk = Math.abs(result[3] - 1.0) < 0.01;\n    const nanDetected =\n      result[7] === -999 || Number.isNaN(result[5]) || Number.isNaN(result[6]) || result[5] === 0;\n\n    const detail =\n      `clamped=${result[0].toExponential(2)},${result[1].toExponential(2)} ` +\n      `exp(-10)=${result[2].toExponential(2)} sig=${result[3].toFixed(4)} ` +\n      `raw: exp(-100)=${result[4]} exp(-1000)=${result[5]} exp(-1e10)=${result[6]} ` +\n      `NaN?=${result[7]}`;\n\n    details.push(\n      clampedOk && normalOk\n        ? `✓ K: exp() clamp OK (${detail})`\n        : `✗ K: exp() clamp FAILED (${detail})`,\n    );\n    outBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ K error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test L: Large buffer integrity (1 MB upload + readback)\n  // Weights are uploaded as large storage buffers. If Safari's writeBuffer silently\n  // corrupts data for large buffers or views with non-zero byteOffset, this will catch it.\n  try {\n    const MB = 256 * 1024; // 256K floats = 1 MB\n    const testData = new Float32Array(MB);\n    // Fill with a deterministic non-trivial pattern (not all-zeros or sequential)\n    for (let i = 0; i < MB; i++) {\n      testData[i] = Math.sin(i * 0.001) * (i % 1000) * 0.01;\n    }\n    const srcBuf = createStorageBuffer(ctx, \"dL_src\", testData.byteLength, testData);\n    const readBuf = createReadbackBuffer(ctx, \"dL_read\", testData.byteLength);\n    const result = await readbackFloats(ctx, srcBuf, readBuf, 0, testData.byteLength);\n\n    let mismatches = 0;\n    let firstMismatchIdx = -1;\n    for (let i = 0; i < MB; i++) {\n      if (Math.abs(result[i] - testData[i]) > 1e-6) {\n        if (firstMismatchIdx === -1) firstMismatchIdx = i;\n        mismatches++;\n      }\n    }\n    const ok = mismatches === 0;\n    details.push(\n      ok\n        ? `✓ L: 1MB buffer integrity OK (${MB} floats)`\n        : `✗ L: 1MB buffer FAILED (${mismatches} mismatches, first at [${firstMismatchIdx}]: expected ${testData[firstMismatchIdx]}, got ${result[firstMismatchIdx]})`,\n    );\n    srcBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ L error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test M: Large buffer with byteOffset (simulates weight tensor views)\n  // Weight tensors from safetensors are ArrayBufferViews with non-zero byteOffset\n  // into a larger buffer. Safari might handle the offset incorrectly in writeBuffer.\n  try {\n    // Create a 2 MB backing buffer, write a known pattern at offset 1 MB\n    const backingSize = 512 * 1024; // 512K floats = 2 MB\n    const backing = new Float32Array(backingSize);\n    const sliceStart = 256 * 1024; // Start at 1 MB offset\n    const sliceLen = 128 * 1024; // 128K floats = 512 KB\n    for (let i = 0; i < sliceLen; i++) {\n      backing[sliceStart + i] = Math.cos(i * 0.002) * (i % 500) * 0.005;\n    }\n    // Create a view with non-zero byteOffset (like getTensorData returns)\n    const view = new Float32Array(backing.buffer, sliceStart * 4, sliceLen);\n\n    const srcBuf = createStorageBuffer(ctx, \"dM_src\", view.byteLength, view);\n    const readBuf = createReadbackBuffer(ctx, \"dM_read\", view.byteLength);\n    const result = await readbackFloats(ctx, srcBuf, readBuf, 0, view.byteLength);\n\n    let mismatches = 0;\n    let firstMismatchIdx = -1;\n    for (let i = 0; i < sliceLen; i++) {\n      if (Math.abs(result[i] - view[i]) > 1e-6) {\n        if (firstMismatchIdx === -1) firstMismatchIdx = i;\n        mismatches++;\n      }\n    }\n    const ok = mismatches === 0;\n    details.push(\n      ok\n        ? `✓ M: offset buffer integrity OK (512KB at 1MB offset)`\n        : `✗ M: offset buffer FAILED (${mismatches} mismatches, first at [${firstMismatchIdx}]: expected ${view[firstMismatchIdx]}, got ${result[firstMismatchIdx]})`,\n    );\n    srcBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ M error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test N: Real RMSNorm kernel (actual WGSL from registry, not simplified)\n  // Tests the exact shader code used in inference. If Safari's WGSL→MSL compiler\n  // miscompiles the real kernel, this will catch it.\n  try {\n    const rmsnormShader = `\nstruct Params {\n  seq_len: u32,\n  hidden_size: u32,\n  eps_bits: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read> weight: array<f32>;\n@group(0) @binding(2) var<storage, read_write> output: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\nvar<workgroup> shared_sum: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let row = wid.x;\n  if (row >= params.seq_len) { return; }\n\n  let tid = lid.x;\n  let row_offset = row * params.hidden_size;\n  let eps = bitcast<f32>(params.eps_bits);\n\n  var local_sum: f32 = 0.0;\n  var i = tid;\n  while (i < params.hidden_size) {\n    let val = input[row_offset + i];\n    local_sum += val * val;\n    i += 256u;\n  }\n  shared_sum[tid] = local_sum;\n  workgroupBarrier();\n\n  var stride: u32 = 128u;\n  while (stride > 0u) {\n    if (tid < stride) {\n      shared_sum[tid] += shared_sum[tid + stride];\n    }\n    workgroupBarrier();\n    stride = stride / 2u;\n  }\n\n  let rms = sqrt(shared_sum[0] / f32(params.hidden_size) + eps);\n\n  i = tid;\n  while (i < params.hidden_size) {\n    let val = input[row_offset + i];\n    output[row_offset + i] = (val / rms) * weight[i];\n    i += 256u;\n  }\n}\n    `;\n\n    // Set up test data: hidden_size=1024, seq_len=1, eps=1e-6\n    const hiddenSize = 1024;\n    const inputData = new Float32Array(hiddenSize);\n    const weightData = new Float32Array(hiddenSize);\n    for (let i = 0; i < hiddenSize; i++) {\n      inputData[i] = Math.sin(i * 0.01) * 2.0; // range [-2, 2]\n      weightData[i] = 1.0 + Math.cos(i * 0.005) * 0.1; // range [0.9, 1.1]\n    }\n\n    // Compute expected output in JS\n    let sumSq = 0;\n    for (let i = 0; i < hiddenSize; i++) sumSq += inputData[i] * inputData[i];\n    const rms = Math.sqrt(sumSq / hiddenSize + 1e-6);\n    const expectedOutput = new Float32Array(hiddenSize);\n    for (let i = 0; i < hiddenSize; i++) {\n      expectedOutput[i] = (inputData[i] / rms) * weightData[i];\n    }\n\n    // Encode eps as u32 bits\n    const epsF32 = new Float32Array(1);\n    epsF32[0] = 1e-6;\n    const epsBits = new Uint32Array(epsF32.buffer)[0];\n\n    // Build uniform buffer: {seq_len=1, hidden_size=1024, eps_bits, _pad=0}\n    const paramsBuf = new ArrayBuffer(16);\n    const paramsView = new DataView(paramsBuf);\n    paramsView.setUint32(0, 1, true);\n    paramsView.setUint32(4, hiddenSize, true);\n    paramsView.setUint32(8, epsBits, true);\n    paramsView.setUint32(12, 0, true);\n\n    const inBuf = createStorageBuffer(ctx, \"dN_in\", inputData.byteLength, inputData);\n    const wBuf = createStorageBuffer(ctx, \"dN_w\", weightData.byteLength, weightData);\n    const outBuf = createStorageBuffer(ctx, \"dN_out\", hiddenSize * 4);\n    const uniformBuf = createUniformBuffer(ctx, \"dN_params\", paramsBuf);\n    const readBuf = createReadbackBuffer(ctx, \"dN_read\", hiddenSize * 4);\n\n    const pipeline = getOrCreatePipeline(ctx, \"dN_rmsnorm\", rmsnormShader, \"main\");\n    const bg = createBindGroup(\n      ctx,\n      pipeline,\n      [{ buffer: inBuf }, { buffer: wBuf }, { buffer: outBuf }, { buffer: uniformBuf }],\n      \"dN_bg\",\n    );\n\n    const encoder = ctx.device.createCommandEncoder();\n    const pass = encoder.beginComputePass();\n    pass.setPipeline(pipeline);\n    pass.setBindGroup(0, bg);\n    pass.dispatchWorkgroups(1); // 1 row\n    pass.end();\n    encoder.copyBufferToBuffer(outBuf, 0, readBuf, 0, hiddenSize * 4);\n    ctx.device.queue.submit([encoder.finish()]);\n    if (ctx.device.queue.onSubmittedWorkDone) {\n      await ctx.device.queue.onSubmittedWorkDone();\n    }\n    await readBuf.mapAsync(MAP_MODE_READ, 0, hiddenSize * 4);\n    const mapped = readBuf.getMappedRange(0, hiddenSize * 4);\n    const result = new Float32Array(mapped.slice(0));\n    readBuf.unmap();\n\n    let maxErr = 0;\n    let maxErrIdx = 0;\n    for (let i = 0; i < hiddenSize; i++) {\n      const err = Math.abs(result[i] - expectedOutput[i]);\n      if (err > maxErr) {\n        maxErr = err;\n        maxErrIdx = i;\n      }\n    }\n    // Allow small floating-point tolerance (tree reduce accumulation order differs)\n    const ok = maxErr < 0.01;\n    details.push(\n      ok\n        ? `✓ N: real RMSNorm kernel OK (maxErr=${maxErr.toExponential(2)} at [${maxErrIdx}])`\n        : `✗ N: real RMSNorm kernel FAILED (maxErr=${maxErr.toExponential(2)} at [${maxErrIdx}]: expected ${expectedOutput[maxErrIdx].toFixed(4)}, got ${result[maxErrIdx].toFixed(4)})`,\n    );\n    inBuf.destroy();\n    wBuf.destroy();\n    outBuf.destroy();\n    uniformBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ N error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test O: Real MatVec kernel (actual WGSL from registry)\n  // Tests K-parallel matrix-vector multiply with 256 threads = 8×32 layout.\n  try {\n    const matvecShader = `\nconst N_TILE: u32 = 8u;\nconst K_THREADS: u32 = 32u;\n\nstruct Params {\n  K: u32,\n  N: u32,\n}\n\n@group(0) @binding(0) var<storage, read> A: array<f32>;\n@group(0) @binding(1) var<storage, read> B: array<f32>;\n@group(0) @binding(2) var<storage, read_write> C: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\nvar<workgroup> shared_sums: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let tid = lid.x;\n  let n_idx = tid / K_THREADS;\n  let k_tid = tid % K_THREADS;\n  let col = wid.x * N_TILE + n_idx;\n  let K = params.K;\n\n  var sum: f32 = 0.0;\n  if (col < params.N) {\n    let b_off = col * K;\n    var k = k_tid * 4u;\n    while (k + 3u < K) {\n      let a_v = vec4f(A[k], A[k+1u], A[k+2u], A[k+3u]);\n      let b_v = vec4f(B[b_off+k], B[b_off+k+1u], B[b_off+k+2u], B[b_off+k+3u]);\n      sum += dot(a_v, b_v);\n      k += K_THREADS * 4u;\n    }\n  }\n\n  shared_sums[tid] = sum;\n  workgroupBarrier();\n\n  if (k_tid == 0u && col < params.N) {\n    let base = n_idx * K_THREADS;\n    var total: f32 = shared_sums[base];\n    for (var i: u32 = 1u; i < K_THREADS; i++) {\n      total += shared_sums[base + i];\n    }\n    C[col] = total;\n  }\n}\n    `;\n\n    // Test: A[256] * B[64, 256] → C[64]  (small but realistic dimensions)\n    const K_dim = 256;\n    const N_dim = 64;\n    const aData = new Float32Array(K_dim);\n    const bData = new Float32Array(N_dim * K_dim);\n    for (let i = 0; i < K_dim; i++) aData[i] = Math.sin(i * 0.05) * 0.5;\n    for (let i = 0; i < N_dim * K_dim; i++) bData[i] = Math.cos(i * 0.003) * 0.1;\n\n    // Expected: C[n] = sum_k A[k] * B[n*K + k]\n    const expectedC = new Float32Array(N_dim);\n    for (let n = 0; n < N_dim; n++) {\n      let s = 0;\n      for (let k = 0; k < K_dim; k++) s += aData[k] * bData[n * K_dim + k];\n      expectedC[n] = s;\n    }\n\n    const paramBuf2 = new ArrayBuffer(8);\n    const pv2 = new DataView(paramBuf2);\n    pv2.setUint32(0, K_dim, true);\n    pv2.setUint32(4, N_dim, true);\n\n    const aBuf = createStorageBuffer(ctx, \"dO_a\", aData.byteLength, aData);\n    const bBuf = createStorageBuffer(ctx, \"dO_b\", bData.byteLength, bData);\n    const cBuf = createStorageBuffer(ctx, \"dO_c\", N_dim * 4);\n    const uBuf = createUniformBuffer(ctx, \"dO_params\", paramBuf2);\n    const rBuf = createReadbackBuffer(ctx, \"dO_read\", N_dim * 4);\n\n    const mvPipeline = getOrCreatePipeline(ctx, \"dO_matvec\", matvecShader, \"main\");\n    const mvBg = createBindGroup(\n      ctx,\n      mvPipeline,\n      [{ buffer: aBuf }, { buffer: bBuf }, { buffer: cBuf }, { buffer: uBuf }],\n      \"dO_bg\",\n    );\n\n    // Dispatch: N_dim / 8 = 8 workgroups\n    const encoder2 = ctx.device.createCommandEncoder();\n    const pass2 = encoder2.beginComputePass();\n    pass2.setPipeline(mvPipeline);\n    pass2.setBindGroup(0, mvBg);\n    pass2.dispatchWorkgroups(Math.ceil(N_dim / 8));\n    pass2.end();\n    encoder2.copyBufferToBuffer(cBuf, 0, rBuf, 0, N_dim * 4);\n    ctx.device.queue.submit([encoder2.finish()]);\n    if (ctx.device.queue.onSubmittedWorkDone) {\n      await ctx.device.queue.onSubmittedWorkDone();\n    }\n    await rBuf.mapAsync(MAP_MODE_READ, 0, N_dim * 4);\n    const mapped2 = rBuf.getMappedRange(0, N_dim * 4);\n    const resultC = new Float32Array(mapped2.slice(0));\n    rBuf.unmap();\n\n    let maxErr2 = 0;\n    let maxErrIdx2 = 0;\n    for (let i = 0; i < N_dim; i++) {\n      const err = Math.abs(resultC[i] - expectedC[i]);\n      if (err > maxErr2) {\n        maxErr2 = err;\n        maxErrIdx2 = i;\n      }\n    }\n    const ok2 = maxErr2 < 0.01;\n    details.push(\n      ok2\n        ? `✓ O: real MatVec kernel OK (maxErr=${maxErr2.toExponential(2)} at [${maxErrIdx2}])`\n        : `✗ O: real MatVec kernel FAILED (maxErr=${maxErr2.toExponential(2)} at [${maxErrIdx2}]: expected ${expectedC[maxErrIdx2].toFixed(4)}, got ${resultC[maxErrIdx2].toFixed(4)})`,\n    );\n    aBuf.destroy();\n    bBuf.destroy();\n    cBuf.destroy();\n    uBuf.destroy();\n    rBuf.destroy();\n  } catch (e) {\n    details.push(`✗ O error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test P: Many dispatches in one compute pass (300 sequential dispatches)\n  // The model runs ~300 dispatches per forward pass in a single compute pass.\n  // If Safari silently drops or corrupts dispatches beyond a limit, only this test catches it.\n  // Uses a simple chain: buf[i] += 1.0 for 300 dispatches → should equal 300.0.\n  try {\n    const manyShader = `\n      @group(0) @binding(0) var<storage, read_write> data: array<f32>;\n      @compute @workgroup_size(64)\n      fn main(@builtin(global_invocation_id) gid: vec3u) {\n        if (gid.x < 4u) { data[gid.x] += 1.0; }\n      }\n    `;\n    const DISPATCH_COUNT = 300;\n    const manyBuf = createStorageBuffer(ctx, \"dP_data\", 16, new Float32Array(4));\n    const manyReadBuf = createReadbackBuffer(ctx, \"dP_read\", 16);\n    const manyPipeline = getOrCreatePipeline(ctx, \"dP\", manyShader, \"main\");\n    const manyBg = createBindGroup(ctx, manyPipeline, [{ buffer: manyBuf }], \"dP_bg\");\n\n    const enc = ctx.device.createCommandEncoder();\n    const cpass = enc.beginComputePass();\n    for (let d = 0; d < DISPATCH_COUNT; d++) {\n      cpass.setPipeline(manyPipeline);\n      cpass.setBindGroup(0, manyBg);\n      cpass.dispatchWorkgroups(1);\n    }\n    cpass.end();\n    enc.copyBufferToBuffer(manyBuf, 0, manyReadBuf, 0, 16);\n    ctx.device.queue.submit([enc.finish()]);\n    if (ctx.device.queue.onSubmittedWorkDone) {\n      await ctx.device.queue.onSubmittedWorkDone();\n    }\n    await manyReadBuf.mapAsync(MAP_MODE_READ, 0, 16);\n    const manyMapped = manyReadBuf.getMappedRange(0, 16);\n    const manyResult = new Float32Array(manyMapped.slice(0));\n    manyReadBuf.unmap();\n\n    const allCorrect = manyResult.every((v) => Math.abs(v - DISPATCH_COUNT) < 0.1);\n    details.push(\n      allCorrect\n        ? `✓ P: ${DISPATCH_COUNT} dispatches OK (${manyResult[0]})`\n        : `✗ P: ${DISPATCH_COUNT} dispatches FAILED (got [${Array.from(manyResult).join(\",\")}], expected ${DISPATCH_COUNT})`,\n    );\n    manyBuf.destroy();\n    manyReadBuf.destroy();\n  } catch (e) {\n    details.push(`✗ P error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test Q: Same shader, different bind groups (Metal argument buffer caching bug)\n  // This is the EXACT pattern that triggers the Safari/Metal bug:\n  // Two dispatches use the same shader code but different bind groups pointing\n  // to different output buffers. If Metal caches argument buffers per compiled\n  // function, the second dispatch will write to the wrong buffer.\n  try {\n    const qShader = `\n      @group(0) @binding(0) var<storage, read> src: array<f32>;\n      @group(0) @binding(1) var<storage, read_write> dst: array<f32>;\n      @compute @workgroup_size(4)\n      fn main(@builtin(global_invocation_id) gid: vec3u) {\n        if (gid.x < 4u) { dst[gid.x] = src[gid.x] * 2.0; }\n      }\n    `;\n    const qSrc = createStorageBuffer(ctx, \"dQ_src\", 16, new Float32Array([1, 2, 3, 4]));\n    const qDst1 = createStorageBuffer(ctx, \"dQ_dst1\", 16, new Float32Array(4)); // zeros\n    const qDst2 = createStorageBuffer(ctx, \"dQ_dst2\", 16, new Float32Array(4)); // zeros\n    const qRead1 = createReadbackBuffer(ctx, \"dQ_read1\", 16);\n    const qRead2 = createReadbackBuffer(ctx, \"dQ_read2\", 16);\n\n    // Same shader, same pipeline — different bind groups with different output buffers\n    const qPipeline = getOrCreatePipeline(ctx, \"dQ\", qShader, \"main\");\n    const qBg1 = createBindGroup(ctx, qPipeline, [{ buffer: qSrc }, { buffer: qDst1 }], \"dQ_bg1\");\n    const qBg2 = createBindGroup(ctx, qPipeline, [{ buffer: qSrc }, { buffer: qDst2 }], \"dQ_bg2\");\n\n    const qEnc = ctx.device.createCommandEncoder();\n    const qPass = qEnc.beginComputePass();\n    qPass.setPipeline(qPipeline);\n    qPass.setBindGroup(0, qBg1);\n    qPass.dispatchWorkgroups(1); // dst1 should be [2,4,6,8]\n    qPass.setPipeline(qPipeline); // SAME pipeline\n    qPass.setBindGroup(0, qBg2); // DIFFERENT bind group → different output buffer\n    qPass.dispatchWorkgroups(1); // dst2 should be [2,4,6,8]\n    qPass.end();\n    qEnc.copyBufferToBuffer(qDst1, 0, qRead1, 0, 16);\n    qEnc.copyBufferToBuffer(qDst2, 0, qRead2, 0, 16);\n    ctx.device.queue.submit([qEnc.finish()]);\n    if (ctx.device.queue.onSubmittedWorkDone) {\n      await ctx.device.queue.onSubmittedWorkDone();\n    }\n    await qRead1.mapAsync(MAP_MODE_READ, 0, 16);\n    const qR1 = new Float32Array(qRead1.getMappedRange(0, 16).slice(0));\n    qRead1.unmap();\n    await qRead2.mapAsync(MAP_MODE_READ, 0, 16);\n    const qR2 = new Float32Array(qRead2.getMappedRange(0, 16).slice(0));\n    qRead2.unmap();\n\n    const qExpected = [2, 4, 6, 8];\n    const dst1Ok = qExpected.every((v, i) => Math.abs(qR1[i] - v) < 0.01);\n    const dst2Ok = qExpected.every((v, i) => Math.abs(qR2[i] - v) < 0.01);\n    details.push(\n      dst1Ok && dst2Ok\n        ? `✓ Q: same-shader diff-bindgroup OK (dst1=[${Array.from(qR1)}] dst2=[${Array.from(qR2)}])`\n        : `✗ Q: same-shader diff-bindgroup FAILED (dst1=[${Array.from(qR1)}] dst2=[${Array.from(qR2)}], expected [${qExpected}])`,\n    );\n    qSrc.destroy();\n    qDst1.destroy();\n    qDst2.destroy();\n    qRead1.destroy();\n    qRead2.destroy();\n  } catch (e) {\n    details.push(`✗ Q error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  return { bufferIntegrity, computeWorks, sharedMemoryWorks, details };\n}\n","/**\n * Kernel registry — maps IR op types to WGSL shader code and dispatch logic.\n *\n * Each entry defines:\n *  - The WGSL source (embedded as string constants for browser compatibility)\n *  - The compute entry point name\n *  - How to calculate workgroup dispatch dimensions\n *  - How to build the uniform buffer matching the WGSL Params struct\n *  - The binding layout (storage-read, storage-read-write, uniform) in order\n *\n * Only ops with WGSL implementations are included. Stubbed ops (MoERouter,\n * ExpertMatMul, Conv2d, AvgPool2d, CrossAttention, Gather, Reshape,\n * Transpose, Concat) are omitted.\n */\n\nimport type { OpNode, OpType } from \"../ir.js\";\n\n// ── KernelSpec interface ────────────────────────────────────────────────\n\n/** Runtime context passed from the executor to kernels at dispatch time. */\nexport interface RuntimeContext {\n  /** Current sequence position (for autoregressive decode). */\n  seqPos: number;\n  /**\n   * Query-grid base offset for windowed Attention dispatches. When set, the\n   * Attention kernel treats workgroup row r as query position r + qOffset, so a\n   * caller can process a sub-range of query rows in one dispatch. Only the WebKit\n   * vision encoder sets it (to chunk the O(N²) ViT attention); defaults to 0.\n   */\n  qOffset?: number;\n}\n\nexport interface KernelSpec {\n  /** WGSL shader source code */\n  shaderCode: string;\n  /** Compute entry point name */\n  entryPoint: string;\n  /** Calculate dispatch workgroup counts from op attributes and resolved shapes */\n  getDispatchSize: (\n    op: OpNode,\n    resolvedShapes: Record<string, number[]>,\n    context?: RuntimeContext,\n  ) => [number, number, number];\n  /** Build the uniform buffer data for this op */\n  buildParams: (\n    op: OpNode,\n    resolvedShapes: Record<string, number[]>,\n    context?: RuntimeContext,\n  ) => ArrayBuffer;\n  /** List of binding entries (storage vs uniform) in order */\n  bindings: Array<{\n    type: \"storage-read\" | \"storage-read-write\" | \"uniform\";\n  }>;\n}\n\n// ── Helpers ─────────────────────────────────────────────────────────────\n\n/** Ceiling division: ceil(a / b) */\nfunction cdiv(a: number, b: number): number {\n  return Math.ceil(a / b);\n}\n\n/** Reinterpret an f32 value as its u32 bit pattern. */\nfunction f32BitsToU32(value: number): number {\n  const f = new Float32Array(1);\n  f[0] = value;\n  return new Uint32Array(f.buffer)[0];\n}\n\n/**\n * Create an ArrayBuffer from an array of u32 values.\n * Each value is written at 4-byte intervals using DataView.\n */\nfunction buildUniformBuffer(values: number[]): ArrayBuffer {\n  const buf = new ArrayBuffer(values.length * 4);\n  const view = new DataView(buf);\n  for (let i = 0; i < values.length; i++) {\n    view.setUint32(i * 4, values[i], true /* littleEndian */);\n  }\n  return buf;\n}\n\n/**\n * Resolve the first dimension of a tensor referenced by an attribute.\n * Falls back to the direct attribute value if no tensor reference is set.\n */\nfunction _resolveFirstDim(\n  op: OpNode,\n  attrName: string,\n  resolvedShapes: Record<string, number[]>,\n): number {\n  const tensorRef = op.attributes[`${attrName}_tensor`] as string | undefined;\n  if (tensorRef && resolvedShapes[tensorRef]) {\n    return resolvedShapes[tensorRef][0];\n  }\n  return op.attributes[attrName] as number;\n}\n\n/**\n * Resolve total element count of a tensor referenced by an attribute.\n */\nfunction _resolveTotalElements(\n  _op: OpNode,\n  tensorRef: string,\n  resolvedShapes: Record<string, number[]>,\n): number {\n  const shape = resolvedShapes[tensorRef];\n  if (!shape) return 0;\n  return shape.reduce((a, b) => a * b, 1);\n}\n\n/**\n * Resolve a dynamic seq_len from tensor references, output shape, or static attribute.\n *\n * Tries (in order):\n *  1. Explicit tensor reference via `seq_len_tensor` attribute — computes total / hidden\n *  2. First output tensor's first dimension (for 2D+ outputs)\n *  3. Static `seq_len` attribute\n */\nfunction resolveSeqLen(\n  op: OpNode,\n  resolvedShapes: Record<string, number[]>,\n  hidden: number,\n): number {\n  // Try explicit tensor reference first\n  const ref = op.attributes.seq_len_tensor as string | undefined;\n  if (ref && resolvedShapes[ref]) {\n    const shape = resolvedShapes[ref];\n    const total = shape.reduce((a: number, b: number) => a * b, 1);\n    return total / hidden;\n  }\n  // Fall back to output tensor shape\n  const outShape = resolvedShapes[op.outputs[0]];\n  if (outShape && outShape.length >= 2) {\n    return outShape[0];\n  }\n  return op.attributes.seq_len as number;\n}\n\n// ── Embedded WGSL shader sources ────────────────────────────────────────\n\nconst WGSL_EMBEDDING = `\\\n// Embedding lookup: gather rows from weight matrix by token ID.\n\nstruct Params {\n  seq_len: u32,\n  hidden_size: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input_ids: array<u32>;\n@group(0) @binding(1) var<storage, read> weight: array<f32>;\n@group(0) @binding(2) var<storage, read_write> output: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  let total = params.seq_len * params.hidden_size;\n  if (idx >= total) { return; }\n\n  let token_pos = idx / params.hidden_size;\n  let dim = idx % params.hidden_size;\n\n  let token_id = input_ids[token_pos];\n  let weight_idx = token_id * params.hidden_size + dim;\n\n  output[idx] = weight[weight_idx];\n}\n`;\n\nconst WGSL_EMBEDDING_INT4 = `\\\n// INT4 embedding lookup: dequantize rows from packed weight matrix by token ID.\n// Weight is packed as 8 nibbles per u32, with per-group scales and zeros.\n// Dequant formula: output = (nibble - zero) * scale\n//\n// Bindings:\n//   @group(0) @binding(0) input_ids: array<u32>     — token IDs, length T\n//   @group(0) @binding(1) weight_q: array<u32>      — packed INT4 [vocab_size * hidden_size / 8]\n//   @group(0) @binding(2) scales: array<f32>         — per-group scales\n//   @group(0) @binding(3) zeros: array<f32>          — per-group zeros\n//   @group(0) @binding(4) output: array<f32>         — output [T, hidden_size]\n//   @group(0) @binding(5) params: { seq_len, hidden_size, group_size }\n\nstruct Params {\n  seq_len: u32,\n  hidden_size: u32,\n  group_size: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input_ids: array<u32>;\n@group(0) @binding(1) var<storage, read> weight_q: array<u32>;\n@group(0) @binding(2) var<storage, read> scales: array<f32>;\n@group(0) @binding(3) var<storage, read> zeros: array<f32>;\n@group(0) @binding(4) var<storage, read_write> output: array<f32>;\n@group(0) @binding(5) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  let total = params.seq_len * params.hidden_size;\n  if (idx >= total) { return; }\n\n  let token_pos = idx / params.hidden_size;\n  let dim = idx % params.hidden_size;\n\n  let token_id = input_ids[token_pos];\n\n  // Flat index into the [vocab_size, hidden_size] embedding table\n  let flat_idx = token_id * params.hidden_size + dim;\n  let packed_idx = flat_idx / 8u;\n  let nibble_pos = flat_idx % 8u;\n  let packed = weight_q[packed_idx];\n  let shift = nibble_pos * 4u;\n  let raw_val = f32((packed >> shift) & 0xFu);\n\n  let group_idx = flat_idx / params.group_size;\n  let scale = scales[group_idx];\n  let zero = zeros[group_idx];\n\n  output[idx] = (raw_val - zero) * scale;\n}\n`;\n\nconst WGSL_MATMUL = `\\\n// Tiled matrix multiply with 4x2 register blocking.\n// C[M,N] = A[M,K] * B^T[N,K]; B stored in [N,K] layout (HF [out,in]).\n//\n// A 16x16 workgroup computes a 64-row x 32-col output tile: each thread owns a\n// 4x2 block of outputs (rows {r,r+16,r+32,r+48} x cols {c,c+16}). Each B column\n// loaded into shared memory is reused for 4 output rows, cutting B (weight)\n// global traffic 4x vs the scalar version — the dominant cost of the wide f32\n// ViT matmuls. A/B tile fills use vec4 loads (K is always %4==0).\n\nconst MT: u32 = 64u;     // output rows per workgroup\nconst NT: u32 = 32u;     // output cols per workgroup\nconst KT: u32 = 16u;     // K-dimension tile depth\nconst KT4: u32 = 4u;     // KT / 4 (vec4 columns along K)\n\nstruct Params {\n  M: u32,\n  K: u32,\n  N: u32,\n}\n\n@group(0) @binding(0) var<storage, read> A: array<vec4f>;\n@group(0) @binding(1) var<storage, read> B: array<vec4f>;\n@group(0) @binding(2) var<storage, read_write> C: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\n// tileA: [MT rows x KT] = 1024, tileB: [KT x NT cols] = 512\nvar<workgroup> tileA: array<f32, 1024>;\nvar<workgroup> tileB: array<f32, 512>;\n\n@compute @workgroup_size(16, 16)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let lr = lid.y;\n  let lc = lid.x;\n  let mBase = wid.y * MT;\n  let nBase = wid.x * NT;\n  // This thread's 4x2 output block: rows {lr, lr+16, lr+32, lr+48}, cols {lc, lc+16}.\n  let row0 = mBase + lr;\n  let col0 = nBase + lc;\n\n  var acc: array<f32, 8>; // [4 rows x 2 cols]\n  for (var i = 0u; i < 8u; i = i + 1u) { acc[i] = 0.0; }\n\n  let numTiles = (params.K + KT - 1u) / KT;\n  let tid = lr * 16u + lc; // 0..255 linear thread id\n  let K4 = params.K / 4u;  // K in vec4 units\n\n  for (var t: u32 = 0u; t < numTiles; t = t + 1u) {\n    let kv0 = t * KT4; // first vec4 column of this K-tile\n    // Load A tile [MT x KT] = 64 rows * 4 vec4 = 256 vec4 loads (1 per thread).\n    {\n      let ar = tid / KT4;            // 0..63 row within tile\n      let av4 = tid % KT4;           // 0..3 vec4-col within tile (along K)\n      let gRow = mBase + ar;\n      var v = vec4f(0.0, 0.0, 0.0, 0.0);\n      if (gRow < params.M && kv0 + av4 < K4) {\n        v = A[gRow * K4 + kv0 + av4];\n      }\n      let base = ar * KT + av4 * 4u;\n      tileA[base] = v.x;\n      tileA[base + 1u] = v.y;\n      tileA[base + 2u] = v.z;\n      tileA[base + 3u] = v.w;\n    }\n    // Load B tile [KT x NT cols]: 32 cols * 4 vec4 = 128 vec4 loads (threads 0..127).\n    if (tid < 128u) {\n      let bc = tid / KT4;            // 0..31 col within tile\n      let bv4 = tid % KT4;           // 0..3 vec4-col within tile (along K)\n      let gCol = nBase + bc;\n      var v = vec4f(0.0, 0.0, 0.0, 0.0);\n      if (gCol < params.N && kv0 + bv4 < K4) {\n        v = B[gCol * K4 + kv0 + bv4];\n      }\n      let krow = bv4 * 4u;\n      tileB[(krow + 0u) * NT + bc] = v.x;\n      tileB[(krow + 1u) * NT + bc] = v.y;\n      tileB[(krow + 2u) * NT + bc] = v.z;\n      tileB[(krow + 3u) * NT + bc] = v.w;\n    }\n\n    workgroupBarrier();\n\n    for (var k: u32 = 0u; k < KT; k = k + 1u) {\n      let b0 = tileB[k * NT + lc];\n      let b1 = tileB[k * NT + lc + 16u];\n      let a0 = tileA[lr * KT + k];\n      let a1 = tileA[(lr + 16u) * KT + k];\n      let a2 = tileA[(lr + 32u) * KT + k];\n      let a3 = tileA[(lr + 48u) * KT + k];\n      acc[0] += a0 * b0;\n      acc[1] += a0 * b1;\n      acc[2] += a1 * b0;\n      acc[3] += a1 * b1;\n      acc[4] += a2 * b0;\n      acc[5] += a2 * b1;\n      acc[6] += a3 * b0;\n      acc[7] += a3 * b1;\n    }\n\n    workgroupBarrier();\n  }\n\n  for (var rr = 0u; rr < 4u; rr = rr + 1u) {\n    let gRow = row0 + rr * 16u;\n    if (gRow < params.M) {\n      if (col0 < params.N) {\n        C[gRow * params.N + col0] = acc[rr * 2u];\n      }\n      if (col0 + 16u < params.N) {\n        C[gRow * params.N + col0 + 16u] = acc[rr * 2u + 1u];\n      }\n    }\n  }\n}\n`;\n\n// Fused MatMul + row-broadcast bias: C[r,c] = A[r,:]·B[c,:] + bias[c].\n// Identical tiling to WGSL_MATMUL (4x2 register blocking, vec4 loads) but adds\n// the bias at store time, eliminating a full wide read+write of the matmul\n// output that a separate AddBias dispatch would incur — one fewer round-trip per\n// ViT linear layer (qkv/proj/fc1/fc2/patch_embed/merger fc1/fc2).\nconst WGSL_MATMUL_BIAS = `\\\nconst MT: u32 = 64u;\nconst NT: u32 = 32u;\nconst KT: u32 = 16u;\nconst KT4: u32 = 4u;\n\nstruct Params {\n  M: u32,\n  K: u32,\n  N: u32,\n}\n\n@group(0) @binding(0) var<storage, read> A: array<vec4f>;\n@group(0) @binding(1) var<storage, read> B: array<vec4f>;\n@group(0) @binding(2) var<storage, read> bias: array<f32>;\n@group(0) @binding(3) var<storage, read_write> C: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\nvar<workgroup> tileA: array<f32, 1024>;\nvar<workgroup> tileB: array<f32, 512>;\n\n@compute @workgroup_size(16, 16)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let lr = lid.y;\n  let lc = lid.x;\n  let mBase = wid.y * MT;\n  let nBase = wid.x * NT;\n  let row0 = mBase + lr;\n  let col0 = nBase + lc;\n\n  var acc: array<f32, 8>;\n  for (var i = 0u; i < 8u; i = i + 1u) { acc[i] = 0.0; }\n\n  let numTiles = (params.K + KT - 1u) / KT;\n  let tid = lr * 16u + lc;\n  let K4 = params.K / 4u;\n\n  for (var t: u32 = 0u; t < numTiles; t = t + 1u) {\n    let kv0 = t * KT4;\n    {\n      let ar = tid / KT4;\n      let av4 = tid % KT4;\n      let gRow = mBase + ar;\n      var v = vec4f(0.0, 0.0, 0.0, 0.0);\n      if (gRow < params.M && kv0 + av4 < K4) {\n        v = A[gRow * K4 + kv0 + av4];\n      }\n      let base = ar * KT + av4 * 4u;\n      tileA[base] = v.x;\n      tileA[base + 1u] = v.y;\n      tileA[base + 2u] = v.z;\n      tileA[base + 3u] = v.w;\n    }\n    if (tid < 128u) {\n      let bc = tid / KT4;\n      let bv4 = tid % KT4;\n      let gCol = nBase + bc;\n      var v = vec4f(0.0, 0.0, 0.0, 0.0);\n      if (gCol < params.N && kv0 + bv4 < K4) {\n        v = B[gCol * K4 + kv0 + bv4];\n      }\n      let krow = bv4 * 4u;\n      tileB[(krow + 0u) * NT + bc] = v.x;\n      tileB[(krow + 1u) * NT + bc] = v.y;\n      tileB[(krow + 2u) * NT + bc] = v.z;\n      tileB[(krow + 3u) * NT + bc] = v.w;\n    }\n\n    workgroupBarrier();\n\n    for (var k: u32 = 0u; k < KT; k = k + 1u) {\n      let b0 = tileB[k * NT + lc];\n      let b1 = tileB[k * NT + lc + 16u];\n      let a0 = tileA[lr * KT + k];\n      let a1 = tileA[(lr + 16u) * KT + k];\n      let a2 = tileA[(lr + 32u) * KT + k];\n      let a3 = tileA[(lr + 48u) * KT + k];\n      acc[0] += a0 * b0;\n      acc[1] += a0 * b1;\n      acc[2] += a1 * b0;\n      acc[3] += a1 * b1;\n      acc[4] += a2 * b0;\n      acc[5] += a2 * b1;\n      acc[6] += a3 * b0;\n      acc[7] += a3 * b1;\n    }\n\n    workgroupBarrier();\n  }\n\n  var bias0 = 0.0;\n  if (col0 < params.N) { bias0 = bias[col0]; }\n  var bias1 = 0.0;\n  if (col0 + 16u < params.N) { bias1 = bias[col0 + 16u]; }\n  for (var rr = 0u; rr < 4u; rr = rr + 1u) {\n    let gRow = row0 + rr * 16u;\n    if (gRow < params.M) {\n      if (col0 < params.N) {\n        C[gRow * params.N + col0] = acc[rr * 2u] + bias0;\n      }\n      if (col0 + 16u < params.N) {\n        C[gRow * params.N + col0 + 16u] = acc[rr * 2u + 1u] + bias1;\n      }\n    }\n  }\n}\n`;\n\n// MIXED-precision MatMulBias: f16 weight + f16 shared tiles, but products are\n// summed in an f32 accumulator (acc += f32(a*b), with a*b done in f16). Keeps the\n// fast f16 MULTIPLY (the ~2% win measured in b4-r5) while accumulating in f32 to\n// stay inside the cos gate that full-f16 accumulation (r5) broke.\nconst WGSL_MATMUL_BIAS_F16MIX = `\\\nenable f16;\n\nconst MT: u32 = 64u;\nconst NT: u32 = 32u;\nconst KT: u32 = 16u;\nconst KT4: u32 = 4u;\n\nstruct Params {\n  M: u32,\n  K: u32,\n  N: u32,\n}\n\n@group(0) @binding(0) var<storage, read> A: array<vec4f>;\n@group(0) @binding(1) var<storage, read> B: array<vec4<f16>>;\n@group(0) @binding(2) var<storage, read> bias: array<f32>;\n@group(0) @binding(3) var<storage, read_write> C: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\nvar<workgroup> tileA: array<f16, 1024>;\nvar<workgroup> tileB: array<f16, 512>;\n\n@compute @workgroup_size(16, 16)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let lr = lid.y;\n  let lc = lid.x;\n  let mBase = wid.y * MT;\n  let nBase = wid.x * NT;\n  let row0 = mBase + lr;\n  let col0 = nBase + lc;\n\n  var acc: array<f32, 8>;\n  for (var i = 0u; i < 8u; i = i + 1u) { acc[i] = 0.0; }\n\n  let numTiles = (params.K + KT - 1u) / KT;\n  let tid = lr * 16u + lc;\n  let K4 = params.K / 4u;\n\n  for (var t: u32 = 0u; t < numTiles; t = t + 1u) {\n    let kv0 = t * KT4;\n    {\n      let ar = tid / KT4;\n      let av4 = tid % KT4;\n      let gRow = mBase + ar;\n      var v = vec4f(0.0, 0.0, 0.0, 0.0);\n      if (gRow < params.M && kv0 + av4 < K4) {\n        v = A[gRow * K4 + kv0 + av4];\n      }\n      let hv = vec4<f16>(v);\n      let base = ar * KT + av4 * 4u;\n      tileA[base] = hv.x;\n      tileA[base + 1u] = hv.y;\n      tileA[base + 2u] = hv.z;\n      tileA[base + 3u] = hv.w;\n    }\n    if (tid < 128u) {\n      let bc = tid / KT4;\n      let bv4 = tid % KT4;\n      let gCol = nBase + bc;\n      var v = vec4<f16>(f16(0.0), f16(0.0), f16(0.0), f16(0.0));\n      if (gCol < params.N && kv0 + bv4 < K4) {\n        v = B[gCol * K4 + kv0 + bv4];\n      }\n      let krow = bv4 * 4u;\n      tileB[(krow + 0u) * NT + bc] = v.x;\n      tileB[(krow + 1u) * NT + bc] = v.y;\n      tileB[(krow + 2u) * NT + bc] = v.z;\n      tileB[(krow + 3u) * NT + bc] = v.w;\n    }\n\n    workgroupBarrier();\n\n    for (var k: u32 = 0u; k < KT; k = k + 1u) {\n      let b0 = tileB[k * NT + lc];\n      let b1 = tileB[k * NT + lc + 16u];\n      let a0 = tileA[lr * KT + k];\n      let a1 = tileA[(lr + 16u) * KT + k];\n      let a2 = tileA[(lr + 32u) * KT + k];\n      let a3 = tileA[(lr + 48u) * KT + k];\n      acc[0] += f32(a0 * b0);\n      acc[1] += f32(a0 * b1);\n      acc[2] += f32(a1 * b0);\n      acc[3] += f32(a1 * b1);\n      acc[4] += f32(a2 * b0);\n      acc[5] += f32(a2 * b1);\n      acc[6] += f32(a3 * b0);\n      acc[7] += f32(a3 * b1);\n    }\n\n    workgroupBarrier();\n  }\n\n  var bias0 = 0.0;\n  if (col0 < params.N) { bias0 = bias[col0]; }\n  var bias1 = 0.0;\n  if (col0 + 16u < params.N) { bias1 = bias[col0 + 16u]; }\n  for (var rr = 0u; rr < 4u; rr = rr + 1u) {\n    let gRow = row0 + rr * 16u;\n    if (gRow < params.M) {\n      if (col0 < params.N) {\n        C[gRow * params.N + col0] = acc[rr * 2u] + bias0;\n      }\n      if (col0 + 16u < params.N) {\n        C[gRow * params.N + col0 + 16u] = acc[rr * 2u + 1u] + bias1;\n      }\n    }\n  }\n}\n`;\n\nconst WGSL_MATMUL_INT4 = `\\\n// Tiled INT4 dequantize + matrix multiply.\n// C[M,N] = A[M,K] * dequant(B_q[N,K])\n// B_q is packed nibbles in [N,K] layout (same as F32 MatMul convention).\n// 8 nibbles per u32, little-nibble-first.\n\nconst TILE: u32 = 16u;\n\nstruct Params {\n  M: u32,\n  K: u32,\n  N: u32,\n  group_size: u32,\n}\n\n@group(0) @binding(0) var<storage, read> A: array<f32>;\n@group(0) @binding(1) var<storage, read> B_q: array<u32>;\n@group(0) @binding(2) var<storage, read> scales: array<f32>;\n@group(0) @binding(3) var<storage, read> zeros: array<f32>;\n@group(0) @binding(4) var<storage, read_write> C: array<f32>;\n@group(0) @binding(5) var<storage, read> params: Params;\n\nvar<workgroup> tileA: array<f32, 256>;\nvar<workgroup> tileB: array<f32, 256>;\n\n@compute @workgroup_size(16, 16)\nfn main(\n  @builtin(global_invocation_id) gid: vec3u,\n  @builtin(local_invocation_id) lid: vec3u,\n) {\n  let row = gid.y;\n  let col = gid.x;\n  let lr = lid.y;\n  let lc = lid.x;\n\n  var sum: f32 = 0.0;\n  let numTiles = (params.K + TILE - 1u) / TILE;\n\n  for (var t: u32 = 0u; t < numTiles; t = t + 1u) {\n    // Load tile of A into shared memory\n    let a_col = t * TILE + lc;\n    if (row < params.M && a_col < params.K) {\n      tileA[lr * TILE + lc] = A[row * params.K + a_col];\n    } else {\n      tileA[lr * TILE + lc] = 0.0;\n    }\n\n    // Load tile of B (INT4 dequantized) into shared memory\n    // B is stored as [N,K] packed nibbles. We need B[col, t*TILE+lr].\n    let b_k = t * TILE + lr;\n    if (b_k < params.K && col < params.N) {\n      let flat_idx = col * params.K + b_k;\n      let packed_idx = flat_idx / 8u;\n      let nibble_pos = flat_idx % 8u;\n      let packed = B_q[packed_idx];\n      let shift = nibble_pos * 4u;\n      let raw_val = f32((packed >> shift) & 0xFu);\n\n      let group_idx = flat_idx / params.group_size;\n      let scale = scales[group_idx];\n      let zero = zeros[group_idx];\n      tileB[lr * TILE + lc] = (raw_val - zero) * scale;\n    } else {\n      tileB[lr * TILE + lc] = 0.0;\n    }\n\n    workgroupBarrier();\n\n    for (var k: u32 = 0u; k < TILE; k = k + 1u) {\n      sum += tileA[lr * TILE + k] * tileB[k * TILE + lc];\n    }\n\n    workgroupBarrier();\n  }\n\n  if (row < params.M && col < params.N) {\n    C[row * params.N + col] = sum;\n  }\n}\n`;\n\nconst WGSL_MATVEC = `\\\n// K-parallel matrix-vector multiply: C[N] = A[K] * B^T[N,K]\n// Optimized for M=1 decode on Apple Silicon (SIMD width 32).\n//\n// Design: 256 threads = 8 output columns × 32 K-threads per column.\n// Each group of 32 threads cooperates along K for one output element.\n// A reads go through L1 cache (no shared memory — maximizes occupancy).\n// B reads are coalesced: 32 adjacent threads read 32 adjacent f32 values\n// = 128 bytes = one full cache line per iteration.\n// Single-barrier reduction (thread 0 of each group sums 32 partials).\n\nconst N_TILE: u32 = 8u;\nconst K_THREADS: u32 = 32u;\n\nstruct Params {\n  K: u32,\n  N: u32,\n}\n\n@group(0) @binding(0) var<storage, read> A: array<f32>;\n@group(0) @binding(1) var<storage, read> B: array<f32>;\n@group(0) @binding(2) var<storage, read_write> C: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\nvar<workgroup> shared_sums: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let tid = lid.x;\n  let n_idx = tid / K_THREADS;\n  let k_tid = tid % K_THREADS;\n  let col = wid.x * N_TILE + n_idx;\n  let K = params.K;\n\n  // K-parallel accumulation with vec4 dot products\n  // A reads hit L1 cache (same addresses across all 16 output columns)\n  var sum: f32 = 0.0;\n  if (col < params.N) {\n    let b_off = col * K;\n    var k = k_tid * 4u;\n    while (k + 3u < K) {\n      let a_v = vec4f(A[k], A[k+1u], A[k+2u], A[k+3u]);\n      let b_v = vec4f(B[b_off+k], B[b_off+k+1u], B[b_off+k+2u], B[b_off+k+3u]);\n      sum += dot(a_v, b_v);\n      k += K_THREADS * 4u;\n    }\n  }\n\n  // Store partials and reduce\n  shared_sums[tid] = sum;\n  workgroupBarrier();\n\n  // Thread 0 of each K-group sequentially sums 16 partials\n  if (k_tid == 0u && col < params.N) {\n    let base = n_idx * K_THREADS;\n    var total: f32 = shared_sums[base];\n    for (var i: u32 = 1u; i < K_THREADS; i++) {\n      total += shared_sums[base + i];\n    }\n    C[col] = total;\n  }\n}\n`;\n\n// ── Subgroups variant of WGSL_MATVEC (desktop-only fast path) ──\n//\n// Identical math and memory-access pattern to WGSL_MATVEC, but the per-column\n// partial-sum reduction is done with subgroupAdd() instead of a shared-memory\n// write + barrier + thread-0 sequential sum. Requires the \"subgroups\" device\n// feature (Chrome 134+, Safari 26+) and \"enable subgroups;\".\n//\n// Layout: 256 threads = 8 output columns × 32 K-threads. The kernel is hand-\n// tuned for SIMD width 32 (Apple Silicon / desktop GPUs), where each group of\n// 32 contiguous K-threads forms exactly one subgroup. We do NOT rely on the\n// (unspecified) lane↔invocation mapping for correctness across arbitrary\n// subgroup sizes: subgroupAdd() reduces over whatever subgroup the lane belongs\n// to, then the subgroup leaders deposit their partials into a tiny shared array\n// (one slot per column), and the column leader combines them. When subgroup\n// size == K_THREADS (the common desktop case) there is exactly one subgroup per\n// column and the shared-memory combine collapses to a single store + load.\n//\n// This variant is selected only on non-WebKit devices that expose subgroups\n// (see executor.ts) — the portable WGSL_MATVEC remains the universal fallback.\nconst WGSL_MATVEC_SUBGROUPS = `\\\nenable subgroups;\n\nconst N_TILE: u32 = 8u;\nconst K_THREADS: u32 = 32u;\n\nstruct Params {\n  K: u32,\n  N: u32,\n}\n\n@group(0) @binding(0) var<storage, read> A: array<f32>;\n@group(0) @binding(1) var<storage, read> B: array<f32>;\n@group(0) @binding(2) var<storage, read_write> C: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\n// One partial slot per column per possible subgroup-within-column. K_THREADS=32\n// so at most 32 subgroups per column in the pathological size-1 case; size\n// 8*32 = 256 covers it. In practice (size 32) only [n_idx*32] is used.\nvar<workgroup> col_partials: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let tid = lid.x;\n  let n_idx = tid / K_THREADS;\n  let k_tid = tid % K_THREADS;\n  let col = wid.x * N_TILE + n_idx;\n  let K = params.K;\n\n  var sum: f32 = 0.0;\n  if (col < params.N) {\n    let b_off = col * K;\n    var k = k_tid * 4u;\n    while (k + 3u < K) {\n      let a_v = vec4f(A[k], A[k+1u], A[k+2u], A[k+3u]);\n      let b_v = vec4f(B[b_off+k], B[b_off+k+1u], B[b_off+k+2u], B[b_off+k+3u]);\n      sum += dot(a_v, b_v);\n      k += K_THREADS * 4u;\n    }\n  }\n\n  // Reduce within the subgroup. Each lane gets the sum over its subgroup.\n  let sg_sum = subgroupAdd(sum);\n\n  // Subgroup leaders deposit their partial into a per-column slot. Slot index is\n  // the leader's k_tid (its offset within the column), guaranteeing distinct\n  // slots per column even when a column spans multiple subgroups.\n  col_partials[tid] = 0.0;\n  workgroupBarrier();\n  if (subgroupElect()) {\n    col_partials[n_idx * K_THREADS + k_tid] = sg_sum;\n  }\n  workgroupBarrier();\n\n  // Column leader sums the (few) deposited subgroup partials for its column.\n  if (k_tid == 0u && col < params.N) {\n    let base = n_idx * K_THREADS;\n    var total: f32 = 0.0;\n    for (var i: u32 = 0u; i < K_THREADS; i++) {\n      total += col_partials[base + i];\n    }\n    C[col] = total;\n  }\n}\n`;\n\n// ── Subgroups variant of WGSL_MATVEC_INT4 (desktop-only fast path) ──\n//\n// Same approach as WGSL_MATVEC_SUBGROUPS applied to the INT4 dequant matvec:\n// subgroupAdd() replaces the shared-memory tree reduction. Layout is widened to\n// 256 threads = 8 columns × 32 K-threads (vs the portable kernel's 128/16) so\n// each column maps to exactly one 32-wide subgroup on desktop GPUs. N_TILE stays\n// 8 so getDispatchSize is unchanged. Selected only on non-WebKit + subgroups.\nconst WGSL_MATVEC_INT4_SUBGROUPS = `\\\nenable subgroups;\n\nconst N_TILE: u32 = 8u;\nconst K_THREADS: u32 = 32u;\n\nstruct Params {\n  K: u32,\n  N: u32,\n  group_size: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> A: array<vec4f>;\n@group(0) @binding(1) var<storage, read> B_q: array<vec4u>;\n@group(0) @binding(2) var<storage, read> scales: array<f32>;\n@group(0) @binding(3) var<storage, read> zeros: array<f32>;\n@group(0) @binding(4) var<storage, read_write> C: array<f32>;\n@group(0) @binding(5) var<storage, read> params: Params;\n\nvar<workgroup> col_partials: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let tid = lid.x;\n  let n_idx = tid / K_THREADS;\n  let k_tid = tid % K_THREADS;\n  let col = wid.x * N_TILE + n_idx;\n  let K = params.K;\n  let k_vec = K / 32u;\n  let groups_per_col = K / params.group_size;\n  let packed_per_group = params.group_size / 8u;\n\n  var sum: f32 = 0.0;\n  if (col < params.N) {\n    let b_off_v = col * k_vec;\n    let col_g_base = col * groups_per_col;\n\n    var v = k_tid;\n    while (v < k_vec) {\n      let bq = B_q[b_off_v + v];\n      let g = (v * 4u) / packed_per_group;\n      let scale = scales[col_g_base + g];\n      let zero = zeros[col_g_base + g];\n      let a_base = v * 8u;\n\n      for (var j: u32 = 0u; j < 4u; j++) {\n        let packed = bq[j];\n        let n0 = f32(packed & 0xFu);\n        let n1 = f32((packed >> 4u) & 0xFu);\n        let n2 = f32((packed >> 8u) & 0xFu);\n        let n3 = f32((packed >> 12u) & 0xFu);\n        let n4 = f32((packed >> 16u) & 0xFu);\n        let n5 = f32((packed >> 20u) & 0xFu);\n        let n6 = f32((packed >> 24u) & 0xFu);\n        let n7 = f32((packed >> 28u) & 0xFu);\n\n        let b0 = vec4f(n0 - zero, n1 - zero, n2 - zero, n3 - zero) * scale;\n        let b1 = vec4f(n4 - zero, n5 - zero, n6 - zero, n7 - zero) * scale;\n\n        sum += dot(A[a_base + j * 2u], b0) + dot(A[a_base + j * 2u + 1u], b1);\n      }\n\n      v += K_THREADS;\n    }\n  }\n\n  let sg_sum = subgroupAdd(sum);\n\n  col_partials[tid] = 0.0;\n  workgroupBarrier();\n  if (subgroupElect()) {\n    col_partials[n_idx * K_THREADS + k_tid] = sg_sum;\n  }\n  workgroupBarrier();\n\n  if (k_tid == 0u && col < params.N) {\n    let base = n_idx * K_THREADS;\n    var total: f32 = 0.0;\n    for (var i: u32 = 0u; i < K_THREADS; i++) {\n      total += col_partials[base + i];\n    }\n    C[col] = total;\n  }\n}\n`;\n\nconst WGSL_MATVEC_INT4 = `\\\n// K-parallel INT4 matrix-vector multiply: C[N] = A[K] * dequant(B_q[N,K])\n// Optimized for M=1 decode on Apple Silicon (SIMD width 32).\n//\n// CONSTRAINT: N_TILE must equal workgroup_size / K_THREADS\n// Design: 256 threads = 8 output columns × 32 K-threads per column.\n\nconst N_TILE: u32 = 16u;\nconst K_THREADS: u32 = 16u;\n\nstruct Params {\n  K: u32,\n  N: u32,\n  group_size: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> A: array<vec4f>;\n@group(0) @binding(1) var<storage, read> B_q: array<vec4u>;\n@group(0) @binding(2) var<storage, read> scales: array<f32>;\n@group(0) @binding(3) var<storage, read> zeros: array<f32>;\n@group(0) @binding(4) var<storage, read_write> C: array<f32>;\n@group(0) @binding(5) var<storage, read> params: Params;\n\nvar<workgroup> shared_sums: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let tid = lid.x;\n  let n_idx = tid / K_THREADS;\n  let k_tid = tid % K_THREADS;\n  let col = wid.x * N_TILE + n_idx;\n  let K = params.K;\n  let k_vec = K / 32u;\n  let groups_per_col = K / params.group_size;\n  let packed_per_group = params.group_size / 8u;\n\n  // K-parallel accumulation with vec4 loads: each thread reads one vec4u\n  // (4 packed u32s = 32 INT4 weights) per iteration. group_size (128) spans\n  // 16 packed u32s, so a vec4u never straddles a quantization group — one\n  // scale/zero load per iteration instead of four.\n  // A reads hit L1 cache (same addresses across all 8 output columns in workgroup)\n  var sum: f32 = 0.0;\n  if (col < params.N) {\n    let b_off_v = col * k_vec;\n    let col_g_base = col * groups_per_col;\n\n    var v = k_tid;\n    while (v < k_vec) {\n      let bq = B_q[b_off_v + v];\n      let g = (v * 4u) / packed_per_group;\n      let scale = scales[col_g_base + g];\n      let zero = zeros[col_g_base + g];\n      let a_base = v * 8u;\n\n      for (var j: u32 = 0u; j < 4u; j++) {\n        let packed = bq[j];\n        // Manual nibble extraction (avoids unpack4xU8 — broken on Safari/WebKit)\n        let n0 = f32(packed & 0xFu);\n        let n1 = f32((packed >> 4u) & 0xFu);\n        let n2 = f32((packed >> 8u) & 0xFu);\n        let n3 = f32((packed >> 12u) & 0xFu);\n        let n4 = f32((packed >> 16u) & 0xFu);\n        let n5 = f32((packed >> 20u) & 0xFu);\n        let n6 = f32((packed >> 24u) & 0xFu);\n        let n7 = f32((packed >> 28u) & 0xFu);\n\n        let b0 = vec4f(n0 - zero, n1 - zero, n2 - zero, n3 - zero) * scale;\n        let b1 = vec4f(n4 - zero, n5 - zero, n6 - zero, n7 - zero) * scale;\n\n        sum += dot(A[a_base + j * 2u], b0) + dot(A[a_base + j * 2u + 1u], b1);\n      }\n\n      v += K_THREADS;\n    }\n  }\n\n  // Store partials and reduce\n  shared_sums[tid] = sum;\n  workgroupBarrier();\n\n  if (k_tid == 0u && col < params.N) {\n    let base = n_idx * K_THREADS;\n    var total: f32 = shared_sums[base];\n    for (var i: u32 = 1u; i < K_THREADS; i++) {\n      total += shared_sums[base + i];\n    }\n    C[col] = total;\n  }\n}\n`;\n\n// ── Gated MatVecInt4 (fused attn-gate + INT4 o_proj, M=1 decode) ──\n//\n// out[n] = sum_k (attn[k] * sigmoid(gate[k])) * dequant(W[n,k])\n// Folds the SigmoidGate (attn_out * sigmoid(gate)) into the o_proj input so the\n// gate elementwise + o_proj run in ONE dispatch. Same INT4 dequant and\n// K-parallel reduction as WGSL_MATVEC_INT4 — only the A vector is built from two\n// inputs with an inline sigmoid. Numerically identical to SigmoidGate→MatVecInt4.\nconst WGSL_GATED_MATVEC_INT4 = `\\\nconst N_TILE: u32 = 8u;\nconst K_THREADS: u32 = 16u;\n\nstruct Params {\n  K: u32,\n  N: u32,\n  group_size: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> attn: array<vec4f>;\n@group(0) @binding(1) var<storage, read> gate: array<vec4f>;\n@group(0) @binding(2) var<storage, read> B_q: array<vec4u>;\n@group(0) @binding(3) var<storage, read> scales: array<f32>;\n@group(0) @binding(4) var<storage, read> zeros: array<f32>;\n@group(0) @binding(5) var<storage, read_write> C: array<f32>;\n@group(0) @binding(6) var<storage, read> params: Params;\n\nvar<workgroup> shared_sums: array<f32, 128>;\n\nfn gated(i: u32) -> vec4f {\n  let a = attn[i];\n  let g = gate[i];\n  let s = vec4f(1.0) / (vec4f(1.0) + exp(max(-g, vec4f(-80.0))));\n  return a * s;\n}\n\n@compute @workgroup_size(128)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let tid = lid.x;\n  let n_idx = tid / K_THREADS;\n  let k_tid = tid % K_THREADS;\n  let col = wid.x * N_TILE + n_idx;\n  let K = params.K;\n  let k_vec = K / 32u;\n  let groups_per_col = K / params.group_size;\n  let packed_per_group = params.group_size / 8u;\n\n  var sum: f32 = 0.0;\n  if (col < params.N) {\n    let b_off_v = col * k_vec;\n    let col_g_base = col * groups_per_col;\n\n    var v = k_tid;\n    while (v < k_vec) {\n      let bq = B_q[b_off_v + v];\n      let g = (v * 4u) / packed_per_group;\n      let scale = scales[col_g_base + g];\n      let zero = zeros[col_g_base + g];\n      let a_base = v * 8u;\n\n      for (var j: u32 = 0u; j < 4u; j++) {\n        let packed = bq[j];\n        let n0 = f32(packed & 0xFu);\n        let n1 = f32((packed >> 4u) & 0xFu);\n        let n2 = f32((packed >> 8u) & 0xFu);\n        let n3 = f32((packed >> 12u) & 0xFu);\n        let n4 = f32((packed >> 16u) & 0xFu);\n        let n5 = f32((packed >> 20u) & 0xFu);\n        let n6 = f32((packed >> 24u) & 0xFu);\n        let n7 = f32((packed >> 28u) & 0xFu);\n\n        let b0 = vec4f(n0 - zero, n1 - zero, n2 - zero, n3 - zero) * scale;\n        let b1 = vec4f(n4 - zero, n5 - zero, n6 - zero, n7 - zero) * scale;\n\n        sum += dot(gated(a_base + j * 2u), b0) + dot(gated(a_base + j * 2u + 1u), b1);\n      }\n\n      v += K_THREADS;\n    }\n  }\n\n  shared_sums[tid] = sum;\n  workgroupBarrier();\n\n  if (k_tid == 0u && col < params.N) {\n    let base = n_idx * K_THREADS;\n    var total: f32 = shared_sums[base];\n    for (var i: u32 = 1u; i < K_THREADS; i++) {\n      total += shared_sums[base + i];\n    }\n    C[col] = total;\n  }\n}\n`;\n\n// ── SwiGLU-gated MatVecInt4 (fused SwiGLU + INT4 projection, M=1 decode) ──\n//\n// out[n] = sum_k (silu(gate[k]) * up[k]) * dequant(W[n,k])\n// Folds a SwiGLU (silu(gate) * up) into the INT4 projection that consumes it —\n// used for the Mamba block's mamba_swiglu (silu(z) * norm_out) feeding out_proj.\n// Same INT4 dequant/reduction as WGSL_MATVEC_INT4; only the A vector is built\n// from two inputs with an inline SiLU. Identical to SwiGLU→MatVecInt4.\nconst WGSL_SWIGLU_GATED_MATVEC_INT4 = `\\\nconst N_TILE: u32 = 8u;\nconst K_THREADS: u32 = 16u;\n\nstruct Params {\n  K: u32,\n  N: u32,\n  group_size: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> gate: array<vec4f>;\n@group(0) @binding(1) var<storage, read> up: array<vec4f>;\n@group(0) @binding(2) var<storage, read> B_q: array<vec4u>;\n@group(0) @binding(3) var<storage, read> scales: array<f32>;\n@group(0) @binding(4) var<storage, read> zeros: array<f32>;\n@group(0) @binding(5) var<storage, read_write> C: array<f32>;\n@group(0) @binding(6) var<storage, read> params: Params;\n\nvar<workgroup> shared_sums: array<f32, 128>;\n\nfn swiglu_a(i: u32) -> vec4f {\n  let g = gate[i];\n  let u = up[i];\n  let s = g / (vec4f(1.0) + exp(max(-g, vec4f(-80.0))));\n  return s * u;\n}\n\n@compute @workgroup_size(128)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let tid = lid.x;\n  let n_idx = tid / K_THREADS;\n  let k_tid = tid % K_THREADS;\n  let col = wid.x * N_TILE + n_idx;\n  let K = params.K;\n  let k_vec = K / 32u;\n  let groups_per_col = K / params.group_size;\n  let packed_per_group = params.group_size / 8u;\n\n  var sum: f32 = 0.0;\n  if (col < params.N) {\n    let b_off_v = col * k_vec;\n    let col_g_base = col * groups_per_col;\n\n    var v = k_tid;\n    while (v < k_vec) {\n      let bq = B_q[b_off_v + v];\n      let g = (v * 4u) / packed_per_group;\n      let scale = scales[col_g_base + g];\n      let zero = zeros[col_g_base + g];\n      let a_base = v * 8u;\n\n      for (var j: u32 = 0u; j < 4u; j++) {\n        let packed = bq[j];\n        let n0 = f32(packed & 0xFu);\n        let n1 = f32((packed >> 4u) & 0xFu);\n        let n2 = f32((packed >> 8u) & 0xFu);\n        let n3 = f32((packed >> 12u) & 0xFu);\n        let n4 = f32((packed >> 16u) & 0xFu);\n        let n5 = f32((packed >> 20u) & 0xFu);\n        let n6 = f32((packed >> 24u) & 0xFu);\n        let n7 = f32((packed >> 28u) & 0xFu);\n\n        let b0 = vec4f(n0 - zero, n1 - zero, n2 - zero, n3 - zero) * scale;\n        let b1 = vec4f(n4 - zero, n5 - zero, n6 - zero, n7 - zero) * scale;\n\n        sum += dot(swiglu_a(a_base + j * 2u), b0) + dot(swiglu_a(a_base + j * 2u + 1u), b1);\n      }\n\n      v += K_THREADS;\n    }\n  }\n\n  shared_sums[tid] = sum;\n  workgroupBarrier();\n\n  if (k_tid == 0u && col < params.N) {\n    let base = n_idx * K_THREADS;\n    var total: f32 = shared_sums[base];\n    for (var i: u32 = 1u; i < K_THREADS; i++) {\n      total += shared_sums[base + i];\n    }\n    C[col] = total;\n  }\n}\n`;\n\nconst WGSL_ARGMAX = `\\\n// GPU-side argmax: finds the index of the maximum value.\n// Single workgroup, 256 threads, parallel reduction.\n\nstruct Params {\n  count: u32,\n  offset: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<vec4f>;\n@group(0) @binding(1) var<storage, read_write> result: array<u32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\nvar<workgroup> shared_max: array<f32, 256>;\nvar<workgroup> shared_idx: array<u32, 256>;\n\n@compute @workgroup_size(256)\nfn main(@builtin(local_invocation_id) lid: vec3u) {\n  let tid = lid.x;\n  // base/count are in elements; argmax always runs offset=0 with count a\n  // multiple of 4 (vocab_size), so we scan the logits as vec4f (128-bit loads,\n  // 4 logits per fetch) to better saturate memory bandwidth.\n  let count4 = params.count >> 2u;\n\n  // Each thread scans its chunk of vec4 groups\n  var local_max: f32 = -3.402823e+38;\n  var local_idx: u32 = 0u;\n\n  var i = tid;\n  while (i < count4) {\n    let v = input[i];\n    let g = i << 2u;\n    if (v.x > local_max) { local_max = v.x; local_idx = g; }\n    if (v.y > local_max) { local_max = v.y; local_idx = g + 1u; }\n    if (v.z > local_max) { local_max = v.z; local_idx = g + 2u; }\n    if (v.w > local_max) { local_max = v.w; local_idx = g + 3u; }\n    i += 256u;\n  }\n\n  shared_max[tid] = local_max;\n  shared_idx[tid] = local_idx;\n  workgroupBarrier();\n\n  // Tree reduction\n  if (tid < 128u) { if (shared_max[tid + 128u] > shared_max[tid]) { shared_max[tid] = shared_max[tid + 128u]; shared_idx[tid] = shared_idx[tid + 128u]; } }\n  workgroupBarrier();\n  if (tid < 64u) { if (shared_max[tid + 64u] > shared_max[tid]) { shared_max[tid] = shared_max[tid + 64u]; shared_idx[tid] = shared_idx[tid + 64u]; } }\n  workgroupBarrier();\n  if (tid < 32u) { if (shared_max[tid + 32u] > shared_max[tid]) { shared_max[tid] = shared_max[tid + 32u]; shared_idx[tid] = shared_idx[tid + 32u]; } }\n  workgroupBarrier();\n  if (tid < 16u) { if (shared_max[tid + 16u] > shared_max[tid]) { shared_max[tid] = shared_max[tid + 16u]; shared_idx[tid] = shared_idx[tid + 16u]; } }\n  workgroupBarrier();\n  if (tid < 8u) { if (shared_max[tid + 8u] > shared_max[tid]) { shared_max[tid] = shared_max[tid + 8u]; shared_idx[tid] = shared_idx[tid + 8u]; } }\n  workgroupBarrier();\n  if (tid < 4u) { if (shared_max[tid + 4u] > shared_max[tid]) { shared_max[tid] = shared_max[tid + 4u]; shared_idx[tid] = shared_idx[tid + 4u]; } }\n  workgroupBarrier();\n  if (tid < 2u) { if (shared_max[tid + 2u] > shared_max[tid]) { shared_max[tid] = shared_max[tid + 2u]; shared_idx[tid] = shared_idx[tid + 2u]; } }\n  workgroupBarrier();\n  if (tid == 0u) {\n    if (shared_max[1u] > shared_max[0u]) {\n      result[0] = shared_idx[1u];\n    } else {\n      result[0] = shared_idx[0u];\n    }\n  }\n}\n`;\n\nconst WGSL_RMSNORM = `\\\n// RMS Normalization: output = (x / rms(x)) * weight\n\nstruct Params {\n  seq_len: u32,\n  hidden_size: u32,\n  eps_bits: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read> weight: array<f32>;\n@group(0) @binding(2) var<storage, read_write> output: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\nvar<workgroup> shared_sum: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let row = wid.x;\n  if (row >= params.seq_len) { return; }\n\n  let tid = lid.x;\n  let row_offset = row * params.hidden_size;\n  let eps = bitcast<f32>(params.eps_bits);\n\n  var local_sum: f32 = 0.0;\n  var i = tid;\n  while (i < params.hidden_size) {\n    let val = input[row_offset + i];\n    local_sum += val * val;\n    i += 256u;\n  }\n  shared_sum[tid] = local_sum;\n  workgroupBarrier();\n\n  var stride: u32 = 128u;\n  while (stride > 0u) {\n    if (tid < stride) {\n      shared_sum[tid] += shared_sum[tid + stride];\n    }\n    workgroupBarrier();\n    stride = stride / 2u;\n  }\n\n  let rms = sqrt(shared_sum[0] / f32(params.hidden_size) + eps);\n\n  i = tid;\n  while (i < params.hidden_size) {\n    let val = input[row_offset + i];\n    output[row_offset + i] = (val / rms) * weight[i];\n    i += 256u;\n  }\n}\n`;\n\nconst WGSL_DUAL_RMSNORM = `\\\n// Two independent per-row RMSNorms in one dispatch, sharing hidden_size + eps.\n// Workgroups [0, rows0) normalize input0→output0 with weight0; workgroups\n// [rows0, rows0+rows1) normalize input1→output1 with weight1. Used to fuse the\n// per-head q_norm and k_norm in full-attention decode (same head_dim, same eps;\n// q has num_heads rows, k has num_kv_heads rows). Numerically identical to two\n// separate RMSNorm dispatches.\n\nstruct Params {\n  rows0: u32,        // workgroups handling input0 (e.g. T*num_heads)\n  rows1: u32,        // workgroups handling input1 (e.g. T*num_kv_heads)\n  hidden_size: u32,  // shared per-row width (head_dim)\n  eps_bits: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input0: array<f32>;\n@group(0) @binding(1) var<storage, read> weight0: array<f32>;\n@group(0) @binding(2) var<storage, read> input1: array<f32>;\n@group(0) @binding(3) var<storage, read> weight1: array<f32>;\n@group(0) @binding(4) var<storage, read_write> output0: array<f32>;\n@group(0) @binding(5) var<storage, read_write> output1: array<f32>;\n@group(0) @binding(6) var<storage, read> params: Params;\n\nvar<workgroup> shared_sum: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let wg = wid.x;\n  if (wg >= params.rows0 + params.rows1) { return; }\n\n  let tid = lid.x;\n  let hs = params.hidden_size;\n  let eps = bitcast<f32>(params.eps_bits);\n\n  // Route this workgroup to input0 (first rows0) or input1 (remaining).\n  let is1 = wg >= params.rows0;\n  let row = select(wg, wg - params.rows0, is1);\n  let row_offset = row * hs;\n\n  var local_sum: f32 = 0.0;\n  var i = tid;\n  while (i < hs) {\n    var val: f32;\n    if (is1) { val = input1[row_offset + i]; } else { val = input0[row_offset + i]; }\n    local_sum += val * val;\n    i += 256u;\n  }\n  shared_sum[tid] = local_sum;\n  workgroupBarrier();\n\n  var stride: u32 = 128u;\n  while (stride > 0u) {\n    if (tid < stride) {\n      shared_sum[tid] += shared_sum[tid + stride];\n    }\n    workgroupBarrier();\n    stride = stride / 2u;\n  }\n\n  let rms = sqrt(shared_sum[0] / f32(hs) + eps);\n\n  i = tid;\n  while (i < hs) {\n    if (is1) {\n      output1[row_offset + i] = (input1[row_offset + i] / rms) * weight1[i];\n    } else {\n      output0[row_offset + i] = (input0[row_offset + i] / rms) * weight0[i];\n    }\n    i += 256u;\n  }\n}\n`;\n\nconst WGSL_LAYERNORM = `\\\n// Layer Normalization: output = ((x - mean) / sqrt(var + eps)) * weight + bias\n\nstruct Params {\n  seq_len: u32,\n  hidden_size: u32,\n  eps_bits: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read> weight: array<f32>;\n@group(0) @binding(2) var<storage, read> bias: array<f32>;\n@group(0) @binding(3) var<storage, read_write> output: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\nvar<workgroup> shared_sum: array<f32, 256>;\nvar<workgroup> shared_sq_sum: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let row = wid.x;\n  if (row >= params.seq_len) { return; }\n\n  let tid = lid.x;\n  let row_offset = row * params.hidden_size;\n  let eps = bitcast<f32>(params.eps_bits);\n\n  var local_sum: f32 = 0.0;\n  var i = tid;\n  while (i < params.hidden_size) {\n    local_sum += input[row_offset + i];\n    i += 256u;\n  }\n  shared_sum[tid] = local_sum;\n  workgroupBarrier();\n\n  var stride: u32 = 128u;\n  while (stride > 0u) {\n    if (tid < stride) {\n      shared_sum[tid] += shared_sum[tid + stride];\n    }\n    workgroupBarrier();\n    stride /= 2u;\n  }\n  let mean = shared_sum[0] / f32(params.hidden_size);\n\n  var local_sq: f32 = 0.0;\n  i = tid;\n  while (i < params.hidden_size) {\n    let diff = input[row_offset + i] - mean;\n    local_sq += diff * diff;\n    i += 256u;\n  }\n  shared_sq_sum[tid] = local_sq;\n  workgroupBarrier();\n\n  stride = 128u;\n  while (stride > 0u) {\n    if (tid < stride) {\n      shared_sq_sum[tid] += shared_sq_sum[tid + stride];\n    }\n    workgroupBarrier();\n    stride /= 2u;\n  }\n  let inv_std = 1.0 / sqrt(shared_sq_sum[0] / f32(params.hidden_size) + eps);\n\n  i = tid;\n  while (i < params.hidden_size) {\n    let val = (input[row_offset + i] - mean) * inv_std;\n    output[row_offset + i] = val * weight[i] + bias[i];\n    i += 256u;\n  }\n}\n`;\n\nconst WGSL_ROPE = `\\\n// Rotary Position Embeddings (RoPE) with partial rotation and rotate_half convention.\n// Uses HuggingFace \"rotate_half\" pairing: (x[i], x[i + rope_half]) NOT adjacent (x[2i], x[2i+1]).\n//\n// Three decoupled knobs (so we match both Qwen-style partial RoPE and Gemma-style\n// \"proportional\" RoPE exactly):\n//   - rope_half          : the rotate_half pairing offset AND the number of pairs we\n//                          iterate per head. For HF this is head_dim/2 (Gemma) or\n//                          rope_dim/2 (Qwen partial). Pairs dim i with i+rope_half.\n//   - rope_denom         : denominator in the inv_freq exponent base^(2i/rope_denom).\n//                          Gemma \"proportional\" uses head_dim; Qwen partial uses rope_dim.\n//   - rope_active_pairs  : pairs [0, rope_active_pairs) get a real frequency; pairs\n//                          >= rope_active_pairs have inv_freq=0 (identity, the \"nope\"\n//                          zero-padded tail of HF _compute_proportional_rope_parameters).\n//\n// Defaults (set in buildParams) reproduce the legacy single-rope_dim behavior:\n//   rope_half = rope_dim/2, rope_denom = rope_dim, rope_active_pairs = rope_dim/2.\n\nstruct Params {\n  seq_len: u32,\n  num_q_heads: u32,\n  num_kv_heads: u32,\n  head_dim: u32,\n  rope_base_bits: u32,\n  position_offset: u32,\n  rope_dim: u32,\n  rope_half: u32,\n  rope_denom: u32,\n  rope_active_pairs: u32,\n  _pad0: u32,\n  _pad1: u32,\n}\n\n@group(0) @binding(0) var<storage, read_write> q: array<f32>;\n@group(0) @binding(1) var<storage, read_write> k: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let rope_half = params.rope_half;\n  let total_q_pairs = params.seq_len * params.num_q_heads * rope_half;\n  let total_k_pairs = params.seq_len * params.num_kv_heads * rope_half;\n\n  let idx = gid.x;\n\n  let rope_base = bitcast<f32>(params.rope_base_bits);\n\n  // rotate_half convention:\n  //   For pair_idx i (0..rope_half-1):\n  //     x_lo = x[i],  x_hi = x[i + rope_half]\n  //     x_lo' = x_lo * cos - x_hi * sin\n  //     x_hi' = x_lo * sin + x_hi * cos\n  //   inv_freq[i] = 1 / base^(2i/rope_denom)  for i < rope_active_pairs, else 0.\n\n  if (idx < total_q_pairs) {\n    let pos = idx / (params.num_q_heads * rope_half);\n    let remainder = idx % (params.num_q_heads * rope_half);\n    let head = remainder / rope_half;\n    let pair_idx = remainder % rope_half;\n\n    let position = f32(pos + params.position_offset);\n    var cos_val = 1.0;\n    var sin_val = 0.0;\n    if (pair_idx < params.rope_active_pairs) {\n      let freq = position / pow(rope_base, f32(2u * pair_idx) / f32(params.rope_denom));\n      cos_val = cos(freq);\n      sin_val = sin(freq);\n    }\n\n    let q_stride = params.num_q_heads * params.head_dim;\n    let head_base = pos * q_stride + head * params.head_dim;\n    let lo = head_base + pair_idx;\n    let hi = head_base + pair_idx + rope_half;\n\n    let q_lo = q[lo];\n    let q_hi = q[hi];\n    q[lo] = q_lo * cos_val - q_hi * sin_val;\n    q[hi] = q_lo * sin_val + q_hi * cos_val;\n  }\n\n  if (idx < total_k_pairs) {\n    let pos = idx / (params.num_kv_heads * rope_half);\n    let remainder = idx % (params.num_kv_heads * rope_half);\n    let head = remainder / rope_half;\n    let pair_idx = remainder % rope_half;\n\n    let position = f32(pos + params.position_offset);\n    var cos_val = 1.0;\n    var sin_val = 0.0;\n    if (pair_idx < params.rope_active_pairs) {\n      let freq = position / pow(rope_base, f32(2u * pair_idx) / f32(params.rope_denom));\n      cos_val = cos(freq);\n      sin_val = sin(freq);\n    }\n\n    let k_stride = params.num_kv_heads * params.head_dim;\n    let head_base = pos * k_stride + head * params.head_dim;\n    let lo = head_base + pair_idx;\n    let hi = head_base + pair_idx + rope_half;\n\n    let k_lo = k[lo];\n    let k_hi = k[hi];\n    k[lo] = k_lo * cos_val - k_hi * sin_val;\n    k[hi] = k_lo * sin_val + k_hi * cos_val;\n  }\n}\n`;\n\nconst WGSL_ROPE_INTERLEAVED = `\\\n// Rotary Position Embeddings — INTERLEAVED (adjacent-pair) convention, used by\n// Moonshine. Unlike the HF-Llama \"rotate_half\" split kernel above (which pairs\n// dim i with dim i+rope_half), Moonshine pairs ADJACENT dims (2p, 2p+1) with a\n// single frequency inv_freq[p] = 1 / base^(2p/rope_denom). Verified against the HF\n// MoonshineRotaryEmbedding (cos[..., :half].repeat_interleave(2) + interleaved\n// rotate_half). Only the first 2*rope_half dims of each head are rotated.\n//\n// Shares the Params layout with WGSL_ROPE (same buildParams). rope_half is the\n// number of adjacent pairs; rope_denom/rope_active_pairs default to the legacy\n// values so Moonshine is byte-identical to before.\n\nstruct Params {\n  seq_len: u32,\n  num_q_heads: u32,\n  num_kv_heads: u32,\n  head_dim: u32,\n  rope_base_bits: u32,\n  position_offset: u32,\n  rope_dim: u32,\n  rope_half: u32,\n  rope_denom: u32,\n  rope_active_pairs: u32,\n  _pad0: u32,\n  _pad1: u32,\n}\n\n@group(0) @binding(0) var<storage, read_write> q: array<f32>;\n@group(0) @binding(1) var<storage, read_write> k: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let half_rope = params.rope_half;\n  let total_q_pairs = params.seq_len * params.num_q_heads * half_rope;\n  let total_k_pairs = params.seq_len * params.num_kv_heads * half_rope;\n  let idx = gid.x;\n  let rope_base = bitcast<f32>(params.rope_base_bits);\n\n  if (idx < total_q_pairs) {\n    let pos = idx / (params.num_q_heads * half_rope);\n    let remainder = idx % (params.num_q_heads * half_rope);\n    let head = remainder / half_rope;\n    let pair_idx = remainder % half_rope;\n\n    let position = f32(pos + params.position_offset);\n    var cos_val = 1.0;\n    var sin_val = 0.0;\n    if (pair_idx < params.rope_active_pairs) {\n      let freq = position / pow(rope_base, f32(2u * pair_idx) / f32(params.rope_denom));\n      cos_val = cos(freq);\n      sin_val = sin(freq);\n    }\n\n    let q_stride = params.num_q_heads * params.head_dim;\n    let head_base = pos * q_stride + head * params.head_dim;\n    // Adjacent-pair (interleaved): dims 2*pair_idx and 2*pair_idx+1.\n    let lo = head_base + 2u * pair_idx;\n    let hi = lo + 1u;\n\n    let q_lo = q[lo];\n    let q_hi = q[hi];\n    q[lo] = q_lo * cos_val - q_hi * sin_val;\n    q[hi] = q_hi * cos_val + q_lo * sin_val;\n  }\n\n  if (idx < total_k_pairs) {\n    let pos = idx / (params.num_kv_heads * half_rope);\n    let remainder = idx % (params.num_kv_heads * half_rope);\n    let head = remainder / half_rope;\n    let pair_idx = remainder % half_rope;\n\n    let position = f32(pos + params.position_offset);\n    var cos_val = 1.0;\n    var sin_val = 0.0;\n    if (pair_idx < params.rope_active_pairs) {\n      let freq = position / pow(rope_base, f32(2u * pair_idx) / f32(params.rope_denom));\n      cos_val = cos(freq);\n      sin_val = sin(freq);\n    }\n\n    let k_stride = params.num_kv_heads * params.head_dim;\n    let head_base = pos * k_stride + head * params.head_dim;\n    let lo = head_base + 2u * pair_idx;\n    let hi = lo + 1u;\n\n    let k_lo = k[lo];\n    let k_hi = k[hi];\n    k[lo] = k_lo * cos_val - k_hi * sin_val;\n    k[hi] = k_hi * cos_val + k_lo * sin_val;\n  }\n}\n`;\n\nconst WGSL_ATTENTION = `\\\n// Tiled online-softmax attention (Flash-Attention style).\n//\n// Processes KV cache in tiles of TILE_S=32 positions. Per tile:\n//   1. Cooperative K tile load into shared memory (256 threads)\n//   2. Parallel Q·K dot products (8 threads per position, reduce via smem)\n//   3. Online softmax: update running max, rescale, compute tile weights\n//   4. Cooperative V tile load into shared memory (reuse K buffer)\n//   5. Parallel V accumulation: each of 128 threads handles one output dim\n//\n// Uses online softmax to maintain running (max, sum_exp, output) across tiles.\n// No global scratch buffer needed — everything in shared memory + registers.\n//\n// Safari/Metal compatibility:\n//   - Uses if/else instead of select() (Safari has select() bugs)\n//   - Clamps exp() args to -80 (Metal can return NaN for exp(-1e30))\n//\n// Shared memory: 4096 f32 = 16384 bytes = exactly 16 KB (minimum WebGPU guarantee)\n\nconst TILE_S: u32 = 16u;\nconst WG: u32 = 256u;\n// 16 threads cooperate on each dot product (256 / 16 = 16)\nconst DOT_THREADS: u32 = 16u;\n\nstruct Params {\n  T: u32,\n  S: u32,\n  num_q_heads: u32,\n  num_kv_heads: u32,\n  head_dim: u32,\n  position_offset: u32,\n  is_causal: u32,\n  // Query-grid base offset. Lets a caller process a window of query positions\n  // [q_offset, q_offset + dispatched_workgroups_x) in a single dispatch without a\n  // workgroup-id offset (WebGPU has none). Default 0 → full grid from row 0, byte-\n  // identical to the un-windowed path. Used only by the WebKit vision encoder to\n  // split the O(N²) ViT attention into watchdog-safe chunks.\n  q_offset: u32,\n  // Softmax QK temperature, supplied by buildParams as an f32. Defaults to\n  // 1/sqrt(head_dim) (byte-identical to the legacy hardcoded scale), so every\n  // existing caller is unchanged. Gemma 4 sets it to 1.0 because the per-head\n  // QK-norm already absorbs the 1/sqrt(head_dim) factor (HF scaling = 1.0).\n  attn_scale: f32,\n}\n\n@group(0) @binding(0) var<storage, read> Q: array<f32>;\n@group(0) @binding(1) var<storage, read> K: array<f32>;\n@group(0) @binding(2) var<storage, read> V: array<f32>;\n@group(0) @binding(3) var<storage, read_write> output: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\n// Shared memory layout (all phases share a single 4096-element array):\n//   K phase:    smem[0..TILE_S*hd) = K tile data (16×256 = 4096 f32)\n//   Dot phase:  smem[0..WG) = partial dot products for reduction\n//               smem[0..TILE_S) reused for final scores after reduction\n//   Softmax:    smem[0..TILE_S) = scores → exp weights\n//   V phase:    smem[TILE_S..TILE_S+TILE_S*hd) = V tile data\n//               But TILE_S+TILE_S*hd = 16+4096 > 4096, so we must save scores\n//               to per-thread registers before V load, then use full smem for V.\n// Total: 4096 f32 = 16384 bytes = exactly 16 KB (minimum WebGPU guarantee)\nvar<workgroup> smem: array<f32, 4096>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let q_pos = wid.x + params.q_offset;\n  let q_head = wid.y;\n\n  if (q_pos >= params.T || q_head >= params.num_q_heads) { return; }\n\n  let tid = lid.x;\n  let heads_per_kv = params.num_q_heads / params.num_kv_heads;\n  let kv_head = q_head / heads_per_kv;\n  let scale = params.attn_scale;\n\n  let q_stride = params.num_q_heads * params.head_dim;\n  let kv_stride = params.num_kv_heads * params.head_dim;\n  let q_base = q_pos * q_stride + q_head * params.head_dim;\n  let kv_head_offset = kv_head * params.head_dim;\n  let causal_limit = q_pos + params.position_offset + 1u;\n  let hd = params.head_dim;\n  // Causal (text) attends to keys up to its own position; bidirectional (ViT)\n  // attends to all S keys. Default is_causal=1 keeps text decoding unchanged.\n  var S_eff = params.S;\n  if (params.is_causal != 0u) { S_eff = min(params.S, causal_limit); }\n\n  // Online softmax running state (per-thread for output dims)\n  var running_max: f32 = -1e30;\n  var running_sum: f32 = 0.0;\n  // Per-thread output accumulators. Thread tid handles output dim tid; a second\n  // accumulator covers dim tid+WG so head_dim up to 2*WG (512) is supported with\n  // WG=256 threads. For hd<=256, tid+WG>=hd is never written → byte-identical.\n  var out_acc: f32 = 0.0;\n  var out_acc2: f32 = 0.0;\n  let d2 = tid + WG;\n\n  // Tile size capped so a K/V tile (tile_cap * hd f32) fits the 4096-f32 smem.\n  // hd<=256 → tile_cap=16=TILE_S (unchanged). hd=512 → tile_cap=8.\n  let tile_cap = min(TILE_S, 4096u / hd);\n\n  // Process KV cache in tiles of tile_cap positions\n  let num_tiles = (S_eff + tile_cap - 1u) / tile_cap;\n\n  for (var tile: u32 = 0u; tile < num_tiles; tile += 1u) {\n    let tile_start = tile * tile_cap;\n    let tile_end = min(tile_start + tile_cap, S_eff);\n    let tile_len = tile_end - tile_start;\n\n    // ── Step 1: Cooperative K tile load ──\n    // Load tile_len × hd values from K cache into smem\n    let total_k_elems = tile_len * hd;\n    var load_idx = tid;\n    while (load_idx < total_k_elems) {\n      let s_local = load_idx / hd;\n      let d = load_idx % hd;\n      let s_global = tile_start + s_local;\n      let k_addr = s_global * kv_stride + kv_head_offset + d;\n      smem[s_local * hd + d] = K[k_addr];\n      load_idx += WG;\n    }\n    workgroupBarrier();\n\n    // ── Step 2: Parallel Q·K dot products ──\n    // 256 threads = TILE_S positions × DOT_THREADS per position\n    let pos_in_tile = tid / DOT_THREADS;\n    let dot_tid = tid % DOT_THREADS;\n\n    var my_partial: f32 = 0.0;\n    if (pos_in_tile < tile_len) {\n      let dims_per_thread = hd / DOT_THREADS;\n      let d_start = dot_tid * dims_per_thread;\n      let d_end = d_start + dims_per_thread;\n      let kv_row_base = pos_in_tile * hd;\n\n      var d = d_start;\n      while (d + 3u < d_end) {\n        let q_v = vec4f(Q[q_base + d], Q[q_base + d + 1u],\n                         Q[q_base + d + 2u], Q[q_base + d + 3u]);\n        let k_v = vec4f(smem[kv_row_base + d], smem[kv_row_base + d + 1u],\n                         smem[kv_row_base + d + 2u], smem[kv_row_base + d + 3u]);\n        my_partial += dot(q_v, k_v);\n        d += 4u;\n      }\n      while (d < d_end) {\n        my_partial += Q[q_base + d] * smem[kv_row_base + d];\n        d += 1u;\n      }\n    }\n\n    // K tile no longer needed — reuse smem for dot reduction\n    workgroupBarrier();\n    smem[tid] = my_partial;\n    workgroupBarrier();\n\n    // Thread 0 of each dot group reduces and writes score to smem[0..TILE_S).\n    // Two-phase: leader 0 reads smem[1..15], which are exactly the slots\n    // leaders 1..15 write — reduce into a local first, barrier (in uniform\n    // control flow), then write. Read+write in one block is a data race.\n    var group_total: f32 = 0.0;\n    if (dot_tid == 0u && pos_in_tile < tile_len) {\n      group_total = smem[tid];\n      for (var i: u32 = 1u; i < DOT_THREADS; i += 1u) {\n        group_total += smem[tid + i];\n      }\n    }\n    workgroupBarrier();\n    if (dot_tid == 0u && pos_in_tile < tile_len) {\n      smem[pos_in_tile] = group_total * scale;\n    }\n    workgroupBarrier();\n\n    // ── Step 3: Online softmax ──\n    // Scores are in smem[0..tile_len). Find max via tree reduce in smem[TILE_S..].\n    // NOTE: Uses if/else instead of select() — Safari/Metal has bugs with select().\n    if (tid < TILE_S) {\n      if (tid < tile_len) {\n        smem[TILE_S + tid] = smem[tid];\n      } else {\n        smem[TILE_S + tid] = -1e30;\n      }\n    }\n    workgroupBarrier();\n    if (tid < 8u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 8u]); }\n    workgroupBarrier();\n    if (tid < 4u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 4u]); }\n    workgroupBarrier();\n    if (tid < 2u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 2u]); }\n    workgroupBarrier();\n    if (tid < 1u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 1u]); }\n    workgroupBarrier();\n    let tile_max = smem[TILE_S];\n\n    let new_max = max(running_max, tile_max);\n    // Clamp exp() argument to avoid NaN on Metal (exp(-87) ≈ 0 in f32)\n    let old_correction = exp(max(running_max - new_max, -80.0));\n\n    // Compute exp weights in smem[0..TILE_S)\n    if (tid < tile_len) {\n      smem[tid] = exp(max(smem[tid] - new_max, -80.0));\n    }\n    workgroupBarrier();\n\n    // Sum tile weights via tree reduce in smem[TILE_S..]\n    // NOTE: Uses if/else instead of select() — Safari/Metal has bugs with select().\n    if (tid < TILE_S) {\n      if (tid < tile_len) {\n        smem[TILE_S + tid] = smem[tid];\n      } else {\n        smem[TILE_S + tid] = 0.0;\n      }\n    }\n    workgroupBarrier();\n    if (tid < 8u) { smem[TILE_S + tid] += smem[TILE_S + tid + 8u]; }\n    workgroupBarrier();\n    if (tid < 4u) { smem[TILE_S + tid] += smem[TILE_S + tid + 4u]; }\n    workgroupBarrier();\n    if (tid < 2u) { smem[TILE_S + tid] += smem[TILE_S + tid + 2u]; }\n    workgroupBarrier();\n    if (tid < 1u) { smem[TILE_S + tid] += smem[TILE_S + tid + 1u]; }\n    workgroupBarrier();\n    let tile_sum = smem[TILE_S];\n\n    // ── Save scores to per-thread registers before V overwrites smem ──\n    // Each thread saves the scores it needs (all TILE_S values)\n    var w0: f32 = 0.0; var w1: f32 = 0.0; var w2: f32 = 0.0; var w3: f32 = 0.0;\n    var w4: f32 = 0.0; var w5: f32 = 0.0; var w6: f32 = 0.0; var w7: f32 = 0.0;\n    var w8: f32 = 0.0; var w9: f32 = 0.0; var w10: f32 = 0.0; var w11: f32 = 0.0;\n    var w12: f32 = 0.0; var w13: f32 = 0.0; var w14: f32 = 0.0; var w15: f32 = 0.0;\n    if (tid < hd) {\n      if (0u < tile_len) { w0 = smem[0]; }\n      if (1u < tile_len) { w1 = smem[1]; }\n      if (2u < tile_len) { w2 = smem[2]; }\n      if (3u < tile_len) { w3 = smem[3]; }\n      if (4u < tile_len) { w4 = smem[4]; }\n      if (5u < tile_len) { w5 = smem[5]; }\n      if (6u < tile_len) { w6 = smem[6]; }\n      if (7u < tile_len) { w7 = smem[7]; }\n      if (8u < tile_len) { w8 = smem[8]; }\n      if (9u < tile_len) { w9 = smem[9]; }\n      if (10u < tile_len) { w10 = smem[10]; }\n      if (11u < tile_len) { w11 = smem[11]; }\n      if (12u < tile_len) { w12 = smem[12]; }\n      if (13u < tile_len) { w13 = smem[13]; }\n      if (14u < tile_len) { w14 = smem[14]; }\n      if (15u < tile_len) { w15 = smem[15]; }\n    }\n    workgroupBarrier();\n\n    // ── Step 4: Cooperative V tile load (full smem reuse) ──\n    let total_v_elems = tile_len * hd;\n    load_idx = tid;\n    while (load_idx < total_v_elems) {\n      let s_local = load_idx / hd;\n      let d = load_idx % hd;\n      let s_global = tile_start + s_local;\n      let v_addr = s_global * kv_stride + kv_head_offset + d;\n      smem[s_local * hd + d] = V[v_addr];\n      load_idx += WG;\n    }\n    workgroupBarrier();\n\n    // ── Step 5: Parallel V accumulation ──\n    // Thread tid handles output dim tid and (for hd>WG) dim tid+WG via out_acc2.\n    // Uses saved score registers instead of smem[0..TILE_S).\n    if (tid < hd) {\n      out_acc = out_acc * old_correction;\n      if (0u < tile_len) { out_acc += w0 * smem[0u * hd + tid]; }\n      if (1u < tile_len) { out_acc += w1 * smem[1u * hd + tid]; }\n      if (2u < tile_len) { out_acc += w2 * smem[2u * hd + tid]; }\n      if (3u < tile_len) { out_acc += w3 * smem[3u * hd + tid]; }\n      if (4u < tile_len) { out_acc += w4 * smem[4u * hd + tid]; }\n      if (5u < tile_len) { out_acc += w5 * smem[5u * hd + tid]; }\n      if (6u < tile_len) { out_acc += w6 * smem[6u * hd + tid]; }\n      if (7u < tile_len) { out_acc += w7 * smem[7u * hd + tid]; }\n      if (8u < tile_len) { out_acc += w8 * smem[8u * hd + tid]; }\n      if (9u < tile_len) { out_acc += w9 * smem[9u * hd + tid]; }\n      if (10u < tile_len) { out_acc += w10 * smem[10u * hd + tid]; }\n      if (11u < tile_len) { out_acc += w11 * smem[11u * hd + tid]; }\n      if (12u < tile_len) { out_acc += w12 * smem[12u * hd + tid]; }\n      if (13u < tile_len) { out_acc += w13 * smem[13u * hd + tid]; }\n      if (14u < tile_len) { out_acc += w14 * smem[14u * hd + tid]; }\n      if (15u < tile_len) { out_acc += w15 * smem[15u * hd + tid]; }\n    }\n    if (d2 < hd) {\n      out_acc2 = out_acc2 * old_correction;\n      if (0u < tile_len) { out_acc2 += w0 * smem[0u * hd + d2]; }\n      if (1u < tile_len) { out_acc2 += w1 * smem[1u * hd + d2]; }\n      if (2u < tile_len) { out_acc2 += w2 * smem[2u * hd + d2]; }\n      if (3u < tile_len) { out_acc2 += w3 * smem[3u * hd + d2]; }\n      if (4u < tile_len) { out_acc2 += w4 * smem[4u * hd + d2]; }\n      if (5u < tile_len) { out_acc2 += w5 * smem[5u * hd + d2]; }\n      if (6u < tile_len) { out_acc2 += w6 * smem[6u * hd + d2]; }\n      if (7u < tile_len) { out_acc2 += w7 * smem[7u * hd + d2]; }\n      if (8u < tile_len) { out_acc2 += w8 * smem[8u * hd + d2]; }\n      if (9u < tile_len) { out_acc2 += w9 * smem[9u * hd + d2]; }\n      if (10u < tile_len) { out_acc2 += w10 * smem[10u * hd + d2]; }\n      if (11u < tile_len) { out_acc2 += w11 * smem[11u * hd + d2]; }\n      if (12u < tile_len) { out_acc2 += w12 * smem[12u * hd + d2]; }\n      if (13u < tile_len) { out_acc2 += w13 * smem[13u * hd + d2]; }\n      if (14u < tile_len) { out_acc2 += w14 * smem[14u * hd + d2]; }\n      if (15u < tile_len) { out_acc2 += w15 * smem[15u * hd + d2]; }\n    }\n\n    // Update running state\n    running_sum = running_sum * old_correction + tile_sum;\n    running_max = new_max;\n    workgroupBarrier();\n  }\n\n  // ── Write normalized output ──\n  var inv_sum: f32 = 0.0;\n  if (running_sum > 0.0) { inv_sum = 1.0 / running_sum; }\n  let out_base = q_pos * q_stride + q_head * hd;\n  if (tid < hd) {\n    output[out_base + tid] = out_acc * inv_sum;\n  }\n  if (d2 < hd) {\n    output[out_base + d2] = out_acc2 * inv_sum;\n  }\n}\n`;\n\n// ── Cross-Attention (encoder-decoder) ──────────────────────────────────\n//\n// Decoder queries attend to a FIXED encoder output. This is the distinct\n// coherence regime from the causal self-attention KVCache path:\n//\n//   Q  = decoder hidden state    [T_dec, num_q_heads  * head_dim]\n//   K  = encoder output (K proj) [S_enc, num_kv_heads * head_dim]\n//   V  = encoder output (V proj) [S_enc, num_kv_heads * head_dim]\n//\n// The encoder K/V are computed ONCE and frozen for the whole decode — there\n// is no position_offset and no growing cache. Attention is UNMASKED /\n// bidirectional over the full S_enc encoder sequence (every decoder query\n// position sees every encoder position). One workgroup per (q_pos, q_head).\n//\n// Algorithm is the identical online-softmax tiling to WGSL_ATTENTION (the\n// validated flash-attention kernel) with the causal-limit logic removed, so\n// it inherits the same Safari/Metal-safe patterns:\n//   - if/else instead of select()\n//   - exp() args clamped to -80 (Metal returns NaN for exp(-1e30))\n//   - <= 16 KB workgroup memory (single 4096 f32 smem array)\n//   - no `enable f16`\nconst WGSL_CROSS_ATTENTION = `\\\nconst TILE_S: u32 = 16u;\nconst WG: u32 = 256u;\nconst DOT_THREADS: u32 = 16u;\n\nstruct Params {\n  T: u32,            // decoder query length\n  S: u32,            // encoder (key/value) length\n  num_q_heads: u32,\n  num_kv_heads: u32,\n  head_dim: u32,\n}\n\n@group(0) @binding(0) var<storage, read> Q: array<f32>;\n@group(0) @binding(1) var<storage, read> K: array<f32>;\n@group(0) @binding(2) var<storage, read> V: array<f32>;\n@group(0) @binding(3) var<storage, read_write> output: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\nvar<workgroup> smem: array<f32, 4096>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let q_pos = wid.x;\n  let q_head = wid.y;\n\n  if (q_pos >= params.T || q_head >= params.num_q_heads) { return; }\n\n  let tid = lid.x;\n  let heads_per_kv = params.num_q_heads / params.num_kv_heads;\n  let kv_head = q_head / heads_per_kv;\n  let scale = 1.0 / sqrt(f32(params.head_dim));\n\n  let q_stride = params.num_q_heads * params.head_dim;\n  let kv_stride = params.num_kv_heads * params.head_dim;\n  let q_base = q_pos * q_stride + q_head * params.head_dim;\n  let kv_head_offset = kv_head * params.head_dim;\n  let hd = params.head_dim;\n  // Unmasked: every query attends to all S_enc encoder positions.\n  let S_eff = params.S;\n\n  var running_max: f32 = -1e30;\n  var running_sum: f32 = 0.0;\n  // out_acc2 covers dim tid+WG for head_dim up to 2*WG; unused for hd<=WG.\n  var out_acc: f32 = 0.0;\n  var out_acc2: f32 = 0.0;\n  let d2 = tid + WG;\n\n  // Tile capped so a K/V tile (tile_cap*hd f32) fits the 4096-f32 smem.\n  let tile_cap = min(TILE_S, 4096u / hd);\n  let num_tiles = (S_eff + tile_cap - 1u) / tile_cap;\n\n  for (var tile: u32 = 0u; tile < num_tiles; tile += 1u) {\n    let tile_start = tile * tile_cap;\n    let tile_end = min(tile_start + tile_cap, S_eff);\n    let tile_len = tile_end - tile_start;\n\n    // ── Step 1: Cooperative K tile load ──\n    let total_k_elems = tile_len * hd;\n    var load_idx = tid;\n    while (load_idx < total_k_elems) {\n      let s_local = load_idx / hd;\n      let d = load_idx % hd;\n      let s_global = tile_start + s_local;\n      let k_addr = s_global * kv_stride + kv_head_offset + d;\n      smem[s_local * hd + d] = K[k_addr];\n      load_idx += WG;\n    }\n    workgroupBarrier();\n\n    // ── Step 2: Parallel Q·K dot products ──\n    let pos_in_tile = tid / DOT_THREADS;\n    let dot_tid = tid % DOT_THREADS;\n\n    var my_partial: f32 = 0.0;\n    if (pos_in_tile < tile_len) {\n      let kv_row_base = pos_in_tile * hd;\n      // Strided over head_dim so any head_dim works (incl. ones not divisible\n      // by DOT_THREADS, e.g. Moonshine's 52). Each thread sums its dims.\n      var d = dot_tid;\n      while (d < hd) {\n        my_partial += Q[q_base + d] * smem[kv_row_base + d];\n        d += DOT_THREADS;\n      }\n    }\n\n    workgroupBarrier();\n    smem[tid] = my_partial;\n    workgroupBarrier();\n\n    var group_total: f32 = 0.0;\n    if (dot_tid == 0u && pos_in_tile < tile_len) {\n      group_total = smem[tid];\n      for (var i: u32 = 1u; i < DOT_THREADS; i += 1u) {\n        group_total += smem[tid + i];\n      }\n    }\n    workgroupBarrier();\n    if (dot_tid == 0u && pos_in_tile < tile_len) {\n      smem[pos_in_tile] = group_total * scale;\n    }\n    workgroupBarrier();\n\n    // ── Step 3: Online softmax (max reduce, no select()) ──\n    if (tid < TILE_S) {\n      if (tid < tile_len) {\n        smem[TILE_S + tid] = smem[tid];\n      } else {\n        smem[TILE_S + tid] = -1e30;\n      }\n    }\n    workgroupBarrier();\n    if (tid < 8u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 8u]); }\n    workgroupBarrier();\n    if (tid < 4u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 4u]); }\n    workgroupBarrier();\n    if (tid < 2u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 2u]); }\n    workgroupBarrier();\n    if (tid < 1u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 1u]); }\n    workgroupBarrier();\n    let tile_max = smem[TILE_S];\n\n    let new_max = max(running_max, tile_max);\n    let old_correction = exp(max(running_max - new_max, -80.0));\n\n    if (tid < tile_len) {\n      smem[tid] = exp(max(smem[tid] - new_max, -80.0));\n    }\n    workgroupBarrier();\n\n    if (tid < TILE_S) {\n      if (tid < tile_len) {\n        smem[TILE_S + tid] = smem[tid];\n      } else {\n        smem[TILE_S + tid] = 0.0;\n      }\n    }\n    workgroupBarrier();\n    if (tid < 8u) { smem[TILE_S + tid] += smem[TILE_S + tid + 8u]; }\n    workgroupBarrier();\n    if (tid < 4u) { smem[TILE_S + tid] += smem[TILE_S + tid + 4u]; }\n    workgroupBarrier();\n    if (tid < 2u) { smem[TILE_S + tid] += smem[TILE_S + tid + 2u]; }\n    workgroupBarrier();\n    if (tid < 1u) { smem[TILE_S + tid] += smem[TILE_S + tid + 1u]; }\n    workgroupBarrier();\n    let tile_sum = smem[TILE_S];\n\n    // ── Save scores to per-thread registers before V overwrites smem ──\n    var w0: f32 = 0.0; var w1: f32 = 0.0; var w2: f32 = 0.0; var w3: f32 = 0.0;\n    var w4: f32 = 0.0; var w5: f32 = 0.0; var w6: f32 = 0.0; var w7: f32 = 0.0;\n    var w8: f32 = 0.0; var w9: f32 = 0.0; var w10: f32 = 0.0; var w11: f32 = 0.0;\n    var w12: f32 = 0.0; var w13: f32 = 0.0; var w14: f32 = 0.0; var w15: f32 = 0.0;\n    if (tid < hd) {\n      if (0u < tile_len) { w0 = smem[0]; }\n      if (1u < tile_len) { w1 = smem[1]; }\n      if (2u < tile_len) { w2 = smem[2]; }\n      if (3u < tile_len) { w3 = smem[3]; }\n      if (4u < tile_len) { w4 = smem[4]; }\n      if (5u < tile_len) { w5 = smem[5]; }\n      if (6u < tile_len) { w6 = smem[6]; }\n      if (7u < tile_len) { w7 = smem[7]; }\n      if (8u < tile_len) { w8 = smem[8]; }\n      if (9u < tile_len) { w9 = smem[9]; }\n      if (10u < tile_len) { w10 = smem[10]; }\n      if (11u < tile_len) { w11 = smem[11]; }\n      if (12u < tile_len) { w12 = smem[12]; }\n      if (13u < tile_len) { w13 = smem[13]; }\n      if (14u < tile_len) { w14 = smem[14]; }\n      if (15u < tile_len) { w15 = smem[15]; }\n    }\n    workgroupBarrier();\n\n    // ── Step 4: Cooperative V tile load ──\n    let total_v_elems = tile_len * hd;\n    load_idx = tid;\n    while (load_idx < total_v_elems) {\n      let s_local = load_idx / hd;\n      let d = load_idx % hd;\n      let s_global = tile_start + s_local;\n      let v_addr = s_global * kv_stride + kv_head_offset + d;\n      smem[s_local * hd + d] = V[v_addr];\n      load_idx += WG;\n    }\n    workgroupBarrier();\n\n    // ── Step 5: Parallel V accumulation ──\n    if (tid < hd) {\n      out_acc = out_acc * old_correction;\n      if (0u < tile_len) { out_acc += w0 * smem[0u * hd + tid]; }\n      if (1u < tile_len) { out_acc += w1 * smem[1u * hd + tid]; }\n      if (2u < tile_len) { out_acc += w2 * smem[2u * hd + tid]; }\n      if (3u < tile_len) { out_acc += w3 * smem[3u * hd + tid]; }\n      if (4u < tile_len) { out_acc += w4 * smem[4u * hd + tid]; }\n      if (5u < tile_len) { out_acc += w5 * smem[5u * hd + tid]; }\n      if (6u < tile_len) { out_acc += w6 * smem[6u * hd + tid]; }\n      if (7u < tile_len) { out_acc += w7 * smem[7u * hd + tid]; }\n      if (8u < tile_len) { out_acc += w8 * smem[8u * hd + tid]; }\n      if (9u < tile_len) { out_acc += w9 * smem[9u * hd + tid]; }\n      if (10u < tile_len) { out_acc += w10 * smem[10u * hd + tid]; }\n      if (11u < tile_len) { out_acc += w11 * smem[11u * hd + tid]; }\n      if (12u < tile_len) { out_acc += w12 * smem[12u * hd + tid]; }\n      if (13u < tile_len) { out_acc += w13 * smem[13u * hd + tid]; }\n      if (14u < tile_len) { out_acc += w14 * smem[14u * hd + tid]; }\n      if (15u < tile_len) { out_acc += w15 * smem[15u * hd + tid]; }\n    }\n    if (d2 < hd) {\n      out_acc2 = out_acc2 * old_correction;\n      if (0u < tile_len) { out_acc2 += w0 * smem[0u * hd + d2]; }\n      if (1u < tile_len) { out_acc2 += w1 * smem[1u * hd + d2]; }\n      if (2u < tile_len) { out_acc2 += w2 * smem[2u * hd + d2]; }\n      if (3u < tile_len) { out_acc2 += w3 * smem[3u * hd + d2]; }\n      if (4u < tile_len) { out_acc2 += w4 * smem[4u * hd + d2]; }\n      if (5u < tile_len) { out_acc2 += w5 * smem[5u * hd + d2]; }\n      if (6u < tile_len) { out_acc2 += w6 * smem[6u * hd + d2]; }\n      if (7u < tile_len) { out_acc2 += w7 * smem[7u * hd + d2]; }\n      if (8u < tile_len) { out_acc2 += w8 * smem[8u * hd + d2]; }\n      if (9u < tile_len) { out_acc2 += w9 * smem[9u * hd + d2]; }\n      if (10u < tile_len) { out_acc2 += w10 * smem[10u * hd + d2]; }\n      if (11u < tile_len) { out_acc2 += w11 * smem[11u * hd + d2]; }\n      if (12u < tile_len) { out_acc2 += w12 * smem[12u * hd + d2]; }\n      if (13u < tile_len) { out_acc2 += w13 * smem[13u * hd + d2]; }\n      if (14u < tile_len) { out_acc2 += w14 * smem[14u * hd + d2]; }\n      if (15u < tile_len) { out_acc2 += w15 * smem[15u * hd + d2]; }\n    }\n\n    running_sum = running_sum * old_correction + tile_sum;\n    running_max = new_max;\n    workgroupBarrier();\n  }\n\n  // ── Write normalized output ──\n  var inv_sum: f32 = 0.0;\n  if (running_sum > 0.0) { inv_sum = 1.0 / running_sum; }\n  let out_base = q_pos * q_stride + q_head * hd;\n  if (tid < hd) {\n    output[out_base + tid] = out_acc * inv_sum;\n  }\n  if (d2 < hd) {\n    output[out_base + d2] = out_acc2 * inv_sum;\n  }\n}\n`;\n\nconst WGSL_SOFTMAX = `\\\n// Row-wise softmax: output[i] = exp(input[i] - max) / sum(exp(input - max))\n\nstruct Params {\n  num_rows: u32,\n  row_size: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read_write> output: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\nvar<workgroup> shared_data: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let row = wid.x;\n  if (row >= params.num_rows) { return; }\n\n  let tid = lid.x;\n  let row_offset = row * params.row_size;\n\n  var local_max: f32 = -1e30;\n  var i = tid;\n  while (i < params.row_size) {\n    local_max = max(local_max, input[row_offset + i]);\n    i += 256u;\n  }\n  shared_data[tid] = local_max;\n  workgroupBarrier();\n\n  var stride: u32 = 128u;\n  while (stride > 0u) {\n    if (tid < stride) {\n      shared_data[tid] = max(shared_data[tid], shared_data[tid + stride]);\n    }\n    workgroupBarrier();\n    stride /= 2u;\n  }\n  let row_max = shared_data[0];\n  workgroupBarrier();\n\n  var local_sum: f32 = 0.0;\n  i = tid;\n  while (i < params.row_size) {\n    local_sum += exp(max(input[row_offset + i] - row_max, -80.0));\n    i += 256u;\n  }\n  shared_data[tid] = local_sum;\n  workgroupBarrier();\n\n  stride = 128u;\n  while (stride > 0u) {\n    if (tid < stride) {\n      shared_data[tid] += shared_data[tid + stride];\n    }\n    workgroupBarrier();\n    stride /= 2u;\n  }\n  let inv_sum = 1.0 / shared_data[0];\n  workgroupBarrier();\n\n  i = tid;\n  while (i < params.row_size) {\n    output[row_offset + i] = exp(max(input[row_offset + i] - row_max, -80.0)) * inv_sum;\n    i += 256u;\n  }\n}\n`;\n\nconst WGSL_SILU = `\\\n// SiLU (Sigmoid Linear Unit): output = x * sigmoid(x) = x / (1 + exp(-x))\n\nstruct Params {\n  count: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read_write> output: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n\n  let x = input[idx];\n  output[idx] = x / (1.0 + exp(max(-x, -80.0)));\n}\n`;\n\nconst WGSL_GELU = `\\\n// GELU (Gaussian Error Linear Unit) - approximate version\n// output = 0.5 * x * (1 + tanh(sqrt(2/pi) * (x + 0.044715 * x^3)))\n\nstruct Params {\n  count: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read_write> output: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\nconst SQRT_2_OVER_PI: f32 = 0.7978845608;\nconst GELU_COEFF: f32 = 0.044715;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n\n  let x = input[idx];\n  // Clamp the tanh argument: tanh saturates to ±1 by |arg|≈10, but on Metal/Dawn\n  // tanh() of a large argument returns NaN (the internal exp overflows). The ViT\n  // MLP can produce |x|>30 → x^3 → arg in the thousands, so clamp to a safe ±15.\n  let inner = clamp(SQRT_2_OVER_PI * (x + GELU_COEFF * x * x * x), -15.0, 15.0);\n  output[idx] = 0.5 * x * (1.0 + tanh(inner));\n}\n`;\n\nconst WGSL_ADD = `\\\n// Element-wise addition: output = a + b\n\nstruct Params {\n  count: u32,\n}\n\n@group(0) @binding(0) var<storage, read> a: array<f32>;\n@group(0) @binding(1) var<storage, read> b: array<f32>;\n@group(0) @binding(2) var<storage, read_write> output: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n  output[idx] = a[idx] + b[idx];\n}\n`;\n\nconst WGSL_MUL = `\\\n// Element-wise multiplication: output = a * b\n\nstruct Params {\n  count: u32,\n}\n\n@group(0) @binding(0) var<storage, read> a: array<f32>;\n@group(0) @binding(1) var<storage, read> b: array<f32>;\n@group(0) @binding(2) var<storage, read_write> output: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n  output[idx] = a[idx] * b[idx];\n}\n`;\n\nconst WGSL_SLICE_LAST_ROW = `\\\n// Copy the last row of a [T, width] tensor into a [1, width] tensor.\n// Lets lm_head run with M=1, so the logits buffer is [1, vocab] instead\n// of [T, vocab] (485MB at T=512 for a 248k vocab).\n\nstruct Params {\n  width: u32,\n  last_row_offset: u32,\n}\n\n@group(0) @binding(0) var<storage, read> src: array<f32>;\n@group(0) @binding(1) var<storage, read_write> output: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.width) { return; }\n  output[idx] = src[params.last_row_offset + idx];\n}\n`;\n\nconst WGSL_MEAN_POOL = `\\\n// Mean-pool a [T, width] tensor over the T tokens into a [1, width] tensor:\n//   output[c] = (1/T) * sum_{t=0..T-1} src[t * width + c]\n// Used for bidirectional encoder embeddings (EmbeddingGemma) where the sentence\n// vector is the average of all token hidden states. We process a single unpadded\n// sequence, so every one of the T rows is a valid (non-pad) token — the mean is\n// over the full T with no attention-mask gather needed.\n// One thread per output channel.\n\nstruct Params {\n  seq_len: u32,\n  width: u32,\n}\n\n@group(0) @binding(0) var<storage, read> src: array<f32>;\n@group(0) @binding(1) var<storage, read_write> output: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let c = gid.x;\n  if (c >= params.width) { return; }\n  var acc = 0.0;\n  var t = 0u;\n  loop {\n    if (t >= params.seq_len) { break; }\n    acc += src[t * params.width + c];\n    t += 1u;\n  }\n  let inv_t = 1.0 / max(f32(params.seq_len), 1.0);\n  output[c] = acc * inv_t;\n}\n`;\n\nconst WGSL_SCALE = `\\\n// Multiply every element of a tensor by a scalar constant:\n//   output[i] = input[i] * scale\n// Used for Gemma's embedding normalizer (× sqrt(hidden_size)), which is applied\n// to the residual stream after the token embedding lookup.\n\nstruct Params {\n  count: u32,\n  scale_bits: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read_write> output: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n  output[idx] = input[idx] * bitcast<f32>(params.scale_bits);\n}\n`;\n\n// ── Softcap (Gemma final logit softcapping) ──\n// output[i] = cap * tanh(input[i] / cap)\n// Squashes logits into (-cap, cap). Mobile-safe: tanh is a WGSL builtin, no\n// select(), no exp (tanh is bounded), no f16. The divide-by-cap argument to tanh\n// is finite for real logits, so no clamping is required.\nconst WGSL_SOFTCAP = `\\\nstruct Params {\n  count: u32,\n  cap_bits: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read_write> output: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n  let cap = bitcast<f32>(params.cap_bits);\n  output[idx] = cap * tanh(input[idx] / cap);\n}\n`;\n\n// (SliceCols WGSL + spec already defined above for the ViT qkv split; Gemma 4\n// PLE reuses it for the per-layer-input column slice.)\n\nconst WGSL_L2NORM = `\\\n// L2-normalize each row of a [rows, width] tensor.\n// embedding[r, :] = x[r, :] / max(sqrt(sum(x[r, :]^2)), eps)\n// One workgroup per row; threads cooperatively reduce the sum of squares.\n\nstruct Params {\n  rows: u32,\n  width: u32,\n}\n\n@group(0) @binding(0) var<storage, read> src: array<f32>;\n@group(0) @binding(1) var<storage, read_write> output: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\nvar<workgroup> partial: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(workgroup_id) wid: vec3u,\n  @builtin(local_invocation_id) lid: vec3u,\n) {\n  let row = wid.x;\n  if (row >= params.rows) { return; }\n  let tid = lid.x;\n  let base = row * params.width;\n\n  // Each thread accumulates a strided slice of the row.\n  var acc = 0.0;\n  var i = tid;\n  loop {\n    if (i >= params.width) { break; }\n    let v = src[base + i];\n    acc += v * v;\n    i += 256u;\n  }\n  partial[tid] = acc;\n  workgroupBarrier();\n\n  // Tree reduction over the 256 partials.\n  var stride = 128u;\n  loop {\n    if (stride == 0u) { break; }\n    if (tid < stride) {\n      partial[tid] += partial[tid + stride];\n    }\n    workgroupBarrier();\n    stride = stride >> 1u;\n  }\n\n  // inv_norm = 1 / max(sqrt(sumsq), eps)\n  let inv_norm = 1.0 / max(sqrt(partial[0]), 1e-12);\n  workgroupBarrier();\n\n  var j = tid;\n  loop {\n    if (j >= params.width) { break; }\n    output[base + j] = src[base + j] * inv_norm;\n    j += 256u;\n  }\n}\n`;\n\nconst WGSL_ADD_BIAS = `\\\n// Row-broadcast bias add: output[r * width + c] = input[r * width + c] + bias[c].\n// Used by the ViT linear layers (qkv/proj/fc1/fc2/patch_embed) whose MatMul has\n// no bias term — the bias vector (length = width) is added across all rows.\n\nstruct Params {\n  count: u32,\n  width: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read> bias: array<f32>;\n@group(0) @binding(2) var<storage, read_write> output: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n  let c = idx % params.width;\n  output[idx] = input[idx] + bias[c];\n}\n`;\n\nconst WGSL_GELU_ERF = `\\\n// Exact GELU (erf form): output = 0.5 * x * (1 + erf(x / sqrt(2))).\n// The ViT merger uses nn.GELU() (exact), unlike the block MLP which uses the\n// tanh approximation. WGSL has no erf(), so we use a high-accuracy rational\n// approximation (Abramowitz & Stegun 7.1.26), giving < 1e-6 abs error.\n\nstruct Params {\n  count: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read_write> output: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\nconst INV_SQRT2: f32 = 0.7071067811865476;\n\nfn erf_approx(x: f32) -> f32 {\n  // A&S 7.1.26: erf(x) for x >= 0; odd-extended for x < 0.\n  // if/else instead of select() — Safari/Metal has select() bugs.\n  var sign = 1.0;\n  if (x < 0.0) { sign = -1.0; }\n  let ax = abs(x);\n  let t = 1.0 / (1.0 + 0.3275911 * ax);\n  let y = 1.0 - (((((1.061405429 * t - 1.453152027) * t) + 1.421413741) * t - 0.284496736) * t + 0.254829592) * t * exp(-ax * ax);\n  return sign * y;\n}\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n  let x = input[idx];\n  output[idx] = 0.5 * x * (1.0 + erf_approx(x * INV_SQRT2));\n}\n`;\n\nconst WGSL_SLICE_COLS = `\\\n// Extract a contiguous column range from a [rows, in_width] tensor into a\n// [rows, out_width] tensor:  output[r, c] = input[r, col_offset + c].\n// Used to split the fused ViT qkv projection [N, 3*768] into Q/K/V [N, 768].\n\nstruct Params {\n  rows: u32,\n  in_width: u32,\n  out_width: u32,\n  col_offset: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read_write> output: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  let total = params.rows * params.out_width;\n  if (idx >= total) { return; }\n  let r = idx / params.out_width;\n  let c = idx % params.out_width;\n  output[idx] = input[r * params.in_width + params.col_offset + c];\n}\n`;\n\nconst WGSL_MUL_COLS = `\\\n// Multiply two contiguous column ranges of a single [rows, in_width] tensor:\n//   output[r, c] = src[r, off_a + c] * src[r, off_b + c]\n// LFM2 short-conv pre-gate: Bx = B * x where B and x are column slices of the\n// fused in_proj output — fuses two SliceCols + a Mul into one dispatch.\n\nstruct Params {\n  rows: u32,\n  in_width: u32,\n  out_width: u32,\n  off_a: u32,\n  off_b: u32,\n  _pad0: u32,\n  _pad1: u32,\n  _pad2: u32,\n}\n\n@group(0) @binding(0) var<storage, read> src: array<f32>;\n@group(0) @binding(1) var<storage, read_write> output: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  let total = params.rows * params.out_width;\n  if (idx >= total) { return; }\n  let r = idx / params.out_width;\n  let c = idx % params.out_width;\n  let base = r * params.in_width;\n  output[idx] = src[base + params.off_a + c] * src[base + params.off_b + c];\n}\n`;\n\nconst WGSL_APPLY_ROTARY = `\\\n// Apply precomputed rotary embeddings to a [T, num_heads*head_dim] tensor.\n// rotate_half convention (matches HF apply_rotary_pos_emb_vision):\n//   out = x * cos + rotate_half(x) * sin\n//   rotate_half(x) = concat(-x[half:], x[:half]) over each head's head_dim.\n// cos/sin are [T, head_dim], broadcast across all heads of a row.\n// In-place safe is NOT assumed — writes to a separate output buffer.\n\nstruct Params {\n  seq_len: u32,\n  num_heads: u32,\n  head_dim: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> x: array<f32>;\n@group(0) @binding(1) var<storage, read> cos_t: array<f32>;\n@group(0) @binding(2) var<storage, read> sin_t: array<f32>;\n@group(0) @binding(3) var<storage, read_write> output: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  let hd = params.head_dim;\n  let total = params.seq_len * params.num_heads * hd;\n  if (idx >= total) { return; }\n\n  let d = idx % hd;          // dim within head\n  let row = idx / (params.num_heads * hd); // token position (for cos/sin lookup)\n  let half = hd / 2u;\n\n  let xv = x[idx];\n  // rotate_half: for d < half → pairs with (d+half) as -x2; for d >= half → x1\n  var rot: f32;\n  if (d < half) {\n    rot = -x[idx + half];\n  } else {\n    rot = x[idx - half];\n  }\n\n  let cs = cos_t[row * hd + d];\n  let sn = sin_t[row * hd + d];\n  output[idx] = xv * cs + rot * sn;\n}\n`;\n\nconst WGSL_MROPE = `\\\n// Multimodal RoPE (M-RoPE) with partial rotation, rotate_half convention.\n// Identical math to WGSL_ROPE but the per-token angles come from host-precomputed\n// cos/sin tables [seq, rope_dim] (built from 3D position ids) instead of pos*inv_freq.\n// Only the first rope_dim of each head is rotated; the rest passes through.\n// For text-only this is fed linear-position cos/sin and reduces to standard 1D RoPE.\n//\n// Q: [T, num_q_heads*head_dim]   K: [T, num_kv_heads*head_dim]\n// cos/sin: [seq, rope_dim]   (rope_dim == 2*half_rope; emb=cat(freqs,freqs))\n\nstruct Params {\n  seq_len: u32,\n  num_q_heads: u32,\n  num_kv_heads: u32,\n  head_dim: u32,\n  rope_dim: u32,\n  pos_offset: u32,  // row offset into cos/sin for decode (single-token steps)\n  _pad0: u32,\n  _pad1: u32,\n}\n\n@group(0) @binding(0) var<storage, read_write> q: array<f32>;\n@group(0) @binding(1) var<storage, read_write> k: array<f32>;\n@group(0) @binding(2) var<storage, read> cos_t: array<f32>;\n@group(0) @binding(3) var<storage, read> sin_t: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let half_rope = params.rope_dim / 2u;\n  let total_q_pairs = params.seq_len * params.num_q_heads * half_rope;\n  let total_k_pairs = params.seq_len * params.num_kv_heads * half_rope;\n  let idx = gid.x;\n  let rd = params.rope_dim;\n\n  if (idx < total_q_pairs) {\n    let pos = idx / (params.num_q_heads * half_rope);\n    let remainder = idx % (params.num_q_heads * half_rope);\n    let head = remainder / half_rope;\n    let pair_idx = remainder % half_rope;\n\n    let cos_row = (pos + params.pos_offset) * rd;\n    let cs = cos_t[cos_row + pair_idx];\n    let sn = sin_t[cos_row + pair_idx];\n\n    let q_stride = params.num_q_heads * params.head_dim;\n    let head_base = pos * q_stride + head * params.head_dim;\n    let lo = head_base + pair_idx;\n    let hi = head_base + pair_idx + half_rope;\n    let q_lo = q[lo];\n    let q_hi = q[hi];\n    // rotate_half: out_lo = lo*cos - hi*sin ; out_hi = hi*cos + lo*sin\n    q[lo] = q_lo * cs - q_hi * sn;\n    q[hi] = q_hi * cs + q_lo * sn;\n  }\n\n  if (idx < total_k_pairs) {\n    let pos = idx / (params.num_kv_heads * half_rope);\n    let remainder = idx % (params.num_kv_heads * half_rope);\n    let head = remainder / half_rope;\n    let pair_idx = remainder % half_rope;\n\n    let cos_row = (pos + params.pos_offset) * rd;\n    let cs = cos_t[cos_row + pair_idx];\n    let sn = sin_t[cos_row + pair_idx];\n\n    let k_stride = params.num_kv_heads * params.head_dim;\n    let head_base = pos * k_stride + head * params.head_dim;\n    let lo = head_base + pair_idx;\n    let hi = head_base + pair_idx + half_rope;\n    let k_lo = k[lo];\n    let k_hi = k[hi];\n    k[lo] = k_lo * cs - k_hi * sn;\n    k[hi] = k_hi * cs + k_lo * sn;\n  }\n}\n`;\n\nconst WGSL_EMBED_SPLICE = `\\\n// Overwrite image-token rows of the text embeddings with merged vision tokens.\n// embed_out[row, c] = vision[map[row], c]  when map[row] >= 0, else embed_in[row,c].\n// row_map: i32 per token; -1 for text rows, else the index into the vision buffer.\n// Writes to a separate output buffer (pooled-alias-safe).\n\nstruct Params {\n  seq_len: u32,\n  hidden: u32,\n}\n\n@group(0) @binding(0) var<storage, read> embed_in: array<f32>;\n@group(0) @binding(1) var<storage, read> vision: array<f32>;\n@group(0) @binding(2) var<storage, read> row_map: array<i32>;\n@group(0) @binding(3) var<storage, read_write> embed_out: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  let total = params.seq_len * params.hidden;\n  if (idx >= total) { return; }\n  let row = idx / params.hidden;\n  let col = idx % params.hidden;\n  let vrow = row_map[row];\n  if (vrow >= 0) {\n    embed_out[idx] = vision[u32(vrow) * params.hidden + col];\n  } else {\n    embed_out[idx] = embed_in[idx];\n  }\n}\n`;\n\nconst WGSL_SWIGLU = `\\\n// Fused SwiGLU: output = SiLU(gate) * up = (gate / (1 + exp(-gate))) * up\n// Saves 2 dispatches per MLP block (SiLU + Mul → single kernel).\n\nstruct Params {\n  count: u32,\n}\n\n@group(0) @binding(0) var<storage, read> gate: array<f32>;\n@group(0) @binding(1) var<storage, read> up: array<f32>;\n@group(0) @binding(2) var<storage, read_write> output: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n\n  let g = gate[idx];\n  output[idx] = (g / (1.0 + exp(max(-g, -80.0)))) * up[idx];\n}\n`;\n\nconst WGSL_SWIGLU_MATVEC = `\\\n// Fused gate+up projection + SwiGLU for M=1 decode.\n// output[d] = SiLU(A · B_gate[d,:]) * (A · B_up[d,:])\n// Reads input vector A once from L1 cache, computes both projections,\n// and applies SiLU gating in a single dispatch.\n// Saves 2 dispatches per MLP block vs separate gate_proj + up_proj + SwiGLU.\n\nconst N_TILE: u32 = 8u;\nconst K_THREADS: u32 = 32u;\n\nstruct Params {\n  K: u32,\n  N: u32,\n}\n\n@group(0) @binding(0) var<storage, read> A: array<f32>;\n@group(0) @binding(1) var<storage, read> B_gate: array<f32>;\n@group(0) @binding(2) var<storage, read> B_up: array<f32>;\n@group(0) @binding(3) var<storage, read_write> output: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\nvar<workgroup> shared_gate: array<f32, 256>;\nvar<workgroup> shared_up: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let tid = lid.x;\n  let n_idx = tid / K_THREADS;\n  let k_tid = tid % K_THREADS;\n  let col = wid.x * N_TILE + n_idx;\n  let K = params.K;\n\n  var sum_gate: f32 = 0.0;\n  var sum_up: f32 = 0.0;\n  if (col < params.N) {\n    let b_off = col * K;\n    var k = k_tid * 4u;\n    while (k + 3u < K) {\n      let a_v = vec4f(A[k], A[k+1u], A[k+2u], A[k+3u]);\n      sum_gate += dot(a_v, vec4f(B_gate[b_off+k], B_gate[b_off+k+1u], B_gate[b_off+k+2u], B_gate[b_off+k+3u]));\n      sum_up += dot(a_v, vec4f(B_up[b_off+k], B_up[b_off+k+1u], B_up[b_off+k+2u], B_up[b_off+k+3u]));\n      k += K_THREADS * 4u;\n    }\n  }\n\n  shared_gate[tid] = sum_gate;\n  shared_up[tid] = sum_up;\n  workgroupBarrier();\n\n  if (k_tid == 0u && col < params.N) {\n    let base = n_idx * K_THREADS;\n    var tg: f32 = shared_gate[base];\n    var tu: f32 = shared_up[base];\n    for (var i: u32 = 1u; i < K_THREADS; i++) {\n      tg += shared_gate[base + i];\n      tu += shared_up[base + i];\n    }\n    let g = tg;\n    output[col] = (g / (1.0 + exp(max(-g, -80.0)))) * tu;\n  }\n}\n`;\n\nconst WGSL_SWIGLU_MATVEC_INT4 = `\\\n// CONSTRAINT: N_TILE must equal workgroup_size / K_THREADS\nconst N_TILE: u32 = 8u;\nconst K_THREADS: u32 = 32u;\n\nstruct Params {\n  K: u32,\n  N: u32,\n  group_size: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> A: array<vec4f>;\n@group(0) @binding(1) var<storage, read> B_gate_q: array<vec4u>;\n@group(0) @binding(2) var<storage, read> B_gate_s: array<f32>;\n@group(0) @binding(3) var<storage, read> B_gate_z: array<f32>;\n@group(0) @binding(4) var<storage, read> B_up_q: array<vec4u>;\n@group(0) @binding(5) var<storage, read> B_up_s: array<f32>;\n@group(0) @binding(6) var<storage, read> B_up_z: array<f32>;\n@group(0) @binding(7) var<storage, read_write> output: array<f32>;\n@group(0) @binding(8) var<storage, read> params: Params;\n\nvar<workgroup> shared_gate: array<f32, 256>;\nvar<workgroup> shared_up: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let tid = lid.x;\n  let n_idx = tid / K_THREADS;\n  let k_tid = tid % K_THREADS;\n  let col = wid.x * N_TILE + n_idx;\n  let K = params.K;\n  let k_vec = K / 32u;\n  let groups_per_col = K / params.group_size;\n  let packed_per_group = params.group_size / 8u;\n\n  // vec4 loads: each thread reads one vec4u per matrix per iteration\n  // (4 packed u32s = 32 INT4 weights). group_size (128) spans 16 packed u32s,\n  // so a vec4u never straddles a quantization group.\n  var sum_gate: f32 = 0.0;\n  var sum_up: f32 = 0.0;\n  if (col < params.N) {\n    let b_off_v = col * k_vec;\n    let col_g_base = col * groups_per_col;\n\n    var v = k_tid;\n    while (v < k_vec) {\n      let group_idx = col_g_base + (v * 4u) / packed_per_group;\n      let gq = B_gate_q[b_off_v + v];\n      let g_scale = B_gate_s[group_idx];\n      let g_zero = B_gate_z[group_idx];\n      let uq = B_up_q[b_off_v + v];\n      let u_scale = B_up_s[group_idx];\n      let u_zero = B_up_z[group_idx];\n      let a_base = v * 8u;\n\n      for (var j: u32 = 0u; j < 4u; j++) {\n        // Gate weight dequantization\n        let g_packed = gq[j];\n        let gn0 = f32(g_packed & 0xFu); let gn1 = f32((g_packed >> 4u) & 0xFu);\n        let gn2 = f32((g_packed >> 8u) & 0xFu); let gn3 = f32((g_packed >> 12u) & 0xFu);\n        let gn4 = f32((g_packed >> 16u) & 0xFu); let gn5 = f32((g_packed >> 20u) & 0xFu);\n        let gn6 = f32((g_packed >> 24u) & 0xFu); let gn7 = f32((g_packed >> 28u) & 0xFu);\n        let gb0 = vec4f(gn0 - g_zero, gn1 - g_zero, gn2 - g_zero, gn3 - g_zero) * g_scale;\n        let gb1 = vec4f(gn4 - g_zero, gn5 - g_zero, gn6 - g_zero, gn7 - g_zero) * g_scale;\n\n        // Up weight dequantization\n        let u_packed = uq[j];\n        let un0 = f32(u_packed & 0xFu); let un1 = f32((u_packed >> 4u) & 0xFu);\n        let un2 = f32((u_packed >> 8u) & 0xFu); let un3 = f32((u_packed >> 12u) & 0xFu);\n        let un4 = f32((u_packed >> 16u) & 0xFu); let un5 = f32((u_packed >> 20u) & 0xFu);\n        let un6 = f32((u_packed >> 24u) & 0xFu); let un7 = f32((u_packed >> 28u) & 0xFu);\n        let ub0 = vec4f(un0 - u_zero, un1 - u_zero, un2 - u_zero, un3 - u_zero) * u_scale;\n        let ub1 = vec4f(un4 - u_zero, un5 - u_zero, un6 - u_zero, un7 - u_zero) * u_scale;\n\n        let a0 = A[a_base + j * 2u];\n        let a1 = A[a_base + j * 2u + 1u];\n\n        sum_gate += dot(a0, gb0) + dot(a1, gb1);\n        sum_up += dot(a0, ub0) + dot(a1, ub1);\n      }\n\n      v += K_THREADS;\n    }\n  }\n\n  shared_gate[tid] = sum_gate;\n  shared_up[tid] = sum_up;\n  workgroupBarrier();\n\n  if (k_tid == 0u && col < params.N) {\n    let base = n_idx * K_THREADS;\n    var tg: f32 = shared_gate[base];\n    var tu: f32 = shared_up[base];\n    for (var i: u32 = 1u; i < K_THREADS; i++) {\n      tg += shared_gate[base + i];\n      tu += shared_up[base + i];\n    }\n    let gg = tg;\n    output[col] = (gg / (1.0 + exp(max(-gg, -80.0)))) * tu;\n  }\n}\n`;\n\n// ── Dual MatVecInt4 (two independent INT4 projections sharing input A, M=1) ──\n//\n// out0[d] = A · dequant(B0[d,:]),  out1[d] = A · dequant(B1[d,:])\n// Identical dequant + reduction to WGSL_SWIGLU_MATVEC_INT4, but writes the two\n// projection sums to SEPARATE output buffers instead of SiLU-combining them.\n// Used to fuse q_proj+gate_proj and k_proj+v_proj in full-attention decode:\n// both projections share the same input_layernorm activation and the same K/N,\n// so reading A once and computing both halves removes one full GPU round-trip\n// (one submit+drain on Safari/iOS) per fused pair.\nconst WGSL_DUAL_MATVEC_INT4 = `\\\nconst N_TILE: u32 = 8u;\nconst K_THREADS: u32 = 32u;\n\nstruct Params {\n  K: u32,\n  N: u32,\n  group_size: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> A: array<vec4f>;\n@group(0) @binding(1) var<storage, read> B0_q: array<vec4u>;\n@group(0) @binding(2) var<storage, read> B0_s: array<f32>;\n@group(0) @binding(3) var<storage, read> B0_z: array<f32>;\n@group(0) @binding(4) var<storage, read> B1_q: array<vec4u>;\n@group(0) @binding(5) var<storage, read> B1_s: array<f32>;\n@group(0) @binding(6) var<storage, read> B1_z: array<f32>;\n@group(0) @binding(7) var<storage, read_write> out0: array<f32>;\n@group(0) @binding(8) var<storage, read_write> out1: array<f32>;\n@group(0) @binding(9) var<storage, read> params: Params;\n\nvar<workgroup> shared_0: array<f32, 256>;\nvar<workgroup> shared_1: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let tid = lid.x;\n  let n_idx = tid / K_THREADS;\n  let k_tid = tid % K_THREADS;\n  let col = wid.x * N_TILE + n_idx;\n  let K = params.K;\n  let k_vec = K / 32u;\n  let groups_per_col = K / params.group_size;\n  let packed_per_group = params.group_size / 8u;\n\n  var sum_0: f32 = 0.0;\n  var sum_1: f32 = 0.0;\n  if (col < params.N) {\n    let b_off_v = col * k_vec;\n    let col_g_base = col * groups_per_col;\n\n    var v = k_tid;\n    while (v < k_vec) {\n      let group_idx = col_g_base + (v * 4u) / packed_per_group;\n      let q0 = B0_q[b_off_v + v];\n      let s0 = B0_s[group_idx];\n      let z0 = B0_z[group_idx];\n      let q1 = B1_q[b_off_v + v];\n      let s1 = B1_s[group_idx];\n      let z1 = B1_z[group_idx];\n      let a_base = v * 8u;\n\n      for (var j: u32 = 0u; j < 4u; j++) {\n        let p0 = q0[j];\n        let a0n0 = f32(p0 & 0xFu); let a0n1 = f32((p0 >> 4u) & 0xFu);\n        let a0n2 = f32((p0 >> 8u) & 0xFu); let a0n3 = f32((p0 >> 12u) & 0xFu);\n        let a0n4 = f32((p0 >> 16u) & 0xFu); let a0n5 = f32((p0 >> 20u) & 0xFu);\n        let a0n6 = f32((p0 >> 24u) & 0xFu); let a0n7 = f32((p0 >> 28u) & 0xFu);\n        let b0a = vec4f(a0n0 - z0, a0n1 - z0, a0n2 - z0, a0n3 - z0) * s0;\n        let b0b = vec4f(a0n4 - z0, a0n5 - z0, a0n6 - z0, a0n7 - z0) * s0;\n\n        let p1 = q1[j];\n        let a1n0 = f32(p1 & 0xFu); let a1n1 = f32((p1 >> 4u) & 0xFu);\n        let a1n2 = f32((p1 >> 8u) & 0xFu); let a1n3 = f32((p1 >> 12u) & 0xFu);\n        let a1n4 = f32((p1 >> 16u) & 0xFu); let a1n5 = f32((p1 >> 20u) & 0xFu);\n        let a1n6 = f32((p1 >> 24u) & 0xFu); let a1n7 = f32((p1 >> 28u) & 0xFu);\n        let b1a = vec4f(a1n0 - z1, a1n1 - z1, a1n2 - z1, a1n3 - z1) * s1;\n        let b1b = vec4f(a1n4 - z1, a1n5 - z1, a1n6 - z1, a1n7 - z1) * s1;\n\n        let av0 = A[a_base + j * 2u];\n        let av1 = A[a_base + j * 2u + 1u];\n\n        sum_0 += dot(av0, b0a) + dot(av1, b0b);\n        sum_1 += dot(av0, b1a) + dot(av1, b1b);\n      }\n\n      v += K_THREADS;\n    }\n  }\n\n  shared_0[tid] = sum_0;\n  shared_1[tid] = sum_1;\n  workgroupBarrier();\n\n  if (k_tid == 0u && col < params.N) {\n    let base = n_idx * K_THREADS;\n    var t0: f32 = shared_0[base];\n    var t1: f32 = shared_1[base];\n    for (var i: u32 = 1u; i < K_THREADS; i++) {\n      t0 += shared_0[base + i];\n      t1 += shared_1[base + i];\n    }\n    out0[col] = t0;\n    out1[col] = t1;\n  }\n}\n`;\n\nconst WGSL_RESIDUAL_RMSNORM = `\\\n// Fused residual add + RMS normalization.\n// sum_output = a + b  (residual sum, kept for downstream residual)\n// norm_output = RMSNorm(a + b) * weight\n\nstruct Params {\n  seq_len: u32,\n  hidden_size: u32,\n  eps_bits: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> a: array<f32>;\n@group(0) @binding(1) var<storage, read> b: array<f32>;\n@group(0) @binding(2) var<storage, read> weight: array<f32>;\n@group(0) @binding(3) var<storage, read_write> sum_output: array<f32>;\n@group(0) @binding(4) var<storage, read_write> norm_output: array<f32>;\n@group(0) @binding(5) var<storage, read> params: Params;\n\nvar<workgroup> shared_sum: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let row = wid.x;\n  if (row >= params.seq_len) { return; }\n\n  let tid = lid.x;\n  let row_offset = row * params.hidden_size;\n  let eps = bitcast<f32>(params.eps_bits);\n\n  // Pass 1: add residual, store sum, accumulate sum-of-squares\n  var local_sum: f32 = 0.0;\n  var i = tid;\n  while (i < params.hidden_size) {\n    let val = a[row_offset + i] + b[row_offset + i];\n    sum_output[row_offset + i] = val;\n    local_sum += val * val;\n    i += 256u;\n  }\n  shared_sum[tid] = local_sum;\n  workgroupBarrier();\n\n  // Tree reduction\n  var stride: u32 = 128u;\n  while (stride > 0u) {\n    if (tid < stride) {\n      shared_sum[tid] += shared_sum[tid + stride];\n    }\n    workgroupBarrier();\n    stride = stride / 2u;\n  }\n\n  let rms = sqrt(shared_sum[0] / f32(params.hidden_size) + eps);\n\n  // Pass 2: normalize and scale\n  i = tid;\n  while (i < params.hidden_size) {\n    let val = sum_output[row_offset + i];\n    norm_output[row_offset + i] = (val / rms) * weight[i];\n    i += 256u;\n  }\n}\n`;\n\nconst WGSL_CAUSAL_CONV1D = `\\\n// Causal 1D depthwise convolution with state buffer for autoregressive decode.\n// conv_state stores the last (kernel_size-1) input timesteps from prior forward passes.\n\nstruct Params {\n  seq_len: u32,\n  channels: u32,\n  kernel_size: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read> weight: array<f32>;\n@group(0) @binding(2) var<storage, read> conv_state: array<f32>;\n@group(0) @binding(3) var<storage, read_write> output: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  let total = params.seq_len * params.channels;\n  if (idx >= total) { return; }\n\n  let t = idx / params.channels;\n  let c = idx % params.channels;\n  let state_size = params.kernel_size - 1u;\n\n  var sum: f32 = 0.0;\n  for (var k: u32 = 0u; k < params.kernel_size; k = k + 1u) {\n    let src_t = i32(t) - i32(k);\n    var in_val: f32 = 0.0;\n    if (src_t >= 0) {\n      in_val = input[u32(src_t) * params.channels + c];\n    } else {\n      // Read from state: state[0]=oldest, state[state_size-1]=most recent\n      let state_t = i32(state_size) + src_t;\n      if (state_t >= 0) {\n        in_val = conv_state[u32(state_t) * params.channels + c];\n      }\n    }\n    // PyTorch Conv1d stores weights reversed: w[0]=most distant, w[K-1]=current.\n    // Our loop looks back k steps, so we need w[K-1-k] to match PyTorch.\n    let w_val = weight[c * params.kernel_size + (params.kernel_size - 1u - k)];\n    sum += in_val * w_val;\n  }\n\n  output[t * params.channels + c] = sum;\n}\n`;\n\nconst WGSL_CAUSAL_CONV1D_SILU = `\\\n// Causal 1D depthwise convolution with fused SiLU activation on the output.\n// Identical to WGSL_CAUSAL_CONV1D but applies output = silu(sum), eliminating a\n// separate SiLU dispatch + a full read/write of the conv output (Qwen3.5 linear\n// attention: F.silu(conv1d(x))).\n\nstruct Params {\n  seq_len: u32,\n  channels: u32,\n  kernel_size: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read> weight: array<f32>;\n@group(0) @binding(2) var<storage, read> conv_state: array<f32>;\n@group(0) @binding(3) var<storage, read_write> output: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  let total = params.seq_len * params.channels;\n  if (idx >= total) { return; }\n\n  let t = idx / params.channels;\n  let c = idx % params.channels;\n  let state_size = params.kernel_size - 1u;\n\n  var sum: f32 = 0.0;\n  for (var k: u32 = 0u; k < params.kernel_size; k = k + 1u) {\n    let src_t = i32(t) - i32(k);\n    var in_val: f32 = 0.0;\n    if (src_t >= 0) {\n      in_val = input[u32(src_t) * params.channels + c];\n    } else {\n      let state_t = i32(state_size) + src_t;\n      if (state_t >= 0) {\n        in_val = conv_state[u32(state_t) * params.channels + c];\n      }\n    }\n    let w_val = weight[c * params.kernel_size + (params.kernel_size - 1u - k)];\n    sum += in_val * w_val;\n  }\n\n  // Fused SiLU: silu(x) = x / (1 + exp(-x))\n  output[t * params.channels + c] = sum / (1.0 + exp(max(-sum, -80.0)));\n}\n`;\n\nconst WGSL_CAUSAL_CONV1D_GATED = `\\\n// Causal 1D depthwise convolution with a fused multiplicative output gate:\n//   output = gate * conv(input)\n// LFM2 short-conv: y = C * depthwise_conv(Bx). Fusing the C-gate into the conv\n// eliminates a separate Mul dispatch + a conv_dim-wide read/write per layer.\n\nstruct Params {\n  seq_len: u32,\n  channels: u32,\n  kernel_size: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read> weight: array<f32>;\n@group(0) @binding(2) var<storage, read> conv_state: array<f32>;\n@group(0) @binding(3) var<storage, read> gate: array<f32>;\n@group(0) @binding(4) var<storage, read_write> output: array<f32>;\n@group(0) @binding(5) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  let total = params.seq_len * params.channels;\n  if (idx >= total) { return; }\n\n  let t = idx / params.channels;\n  let c = idx % params.channels;\n  let state_size = params.kernel_size - 1u;\n\n  var sum: f32 = 0.0;\n  for (var k: u32 = 0u; k < params.kernel_size; k = k + 1u) {\n    let src_t = i32(t) - i32(k);\n    var in_val: f32 = 0.0;\n    if (src_t >= 0) {\n      in_val = input[u32(src_t) * params.channels + c];\n    } else {\n      let state_t = i32(state_size) + src_t;\n      if (state_t >= 0) {\n        in_val = conv_state[u32(state_t) * params.channels + c];\n      }\n    }\n    let w_val = weight[c * params.kernel_size + (params.kernel_size - 1u - k)];\n    sum += in_val * w_val;\n  }\n\n  output[t * params.channels + c] = gate[idx] * sum;\n}\n`;\n\nconst WGSL_SIGMOID_GATE = `\\\n// Sigmoid gate: output = x * sigmoid(gate)\n\nstruct Params {\n  count: u32,\n}\n\n@group(0) @binding(0) var<storage, read> x: array<f32>;\n@group(0) @binding(1) var<storage, read> gate: array<f32>;\n@group(0) @binding(2) var<storage, read_write> output: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n\n  let g = gate[idx];\n  let sig = 1.0 / (1.0 + exp(max(-g, -80.0)));\n  output[idx] = x[idx] * sig;\n}\n`;\n\nconst WGSL_MAMBA_SSM = `\\\n// Gated Delta Net recurrence (Qwen3.5 linear attention).\n//\n// For each head, maintains state [key_dim, val_dim] across timesteps.\n// Recurrence per timestep:\n//   g = -exp(A_log) * softplus(a + dt_bias)   (log-space decay)\n//   decay = exp(g)                             (actual decay in (0,1))\n//   state *= decay                             (decay state)\n//   q_norm, k_norm = L2_normalize(q), L2_normalize(k)\n//   kv_mem = state^T @ k_norm                  (retrieve from memory)\n//   delta = (v - kv_mem) * sigmoid(b)          (delta rule)\n//   state += k_norm @ delta^T                  (rank-1 update)\n//   output = state^T @ q_norm / sqrt(key_dim)  (query output)\n\nstruct Params {\n  T: u32,\n  num_heads: u32,\n  key_dim: u32,\n  val_dim: u32,\n  qkv_dim: u32,\n  _pad1: u32,\n  _pad2: u32,\n  _pad3: u32,\n}\n\n@group(0) @binding(0) var<storage, read> qkv_conv: array<f32>;\n@group(0) @binding(1) var<storage, read> a_input: array<f32>;\n@group(0) @binding(2) var<storage, read> b_input: array<f32>;\n@group(0) @binding(3) var<storage, read> A_log: array<f32>;\n@group(0) @binding(4) var<storage, read> dt_bias: array<f32>;\n@group(0) @binding(5) var<storage, read_write> ssm_state: array<f32>;\n@group(0) @binding(6) var<storage, read_write> output: array<f32>;\n@group(0) @binding(7) var<storage, read> params: Params;\n\n// Shared memory for Q and K vectors (L2-normalized) and reduction\nvar<workgroup> shared_q: array<f32, 128>;\nvar<workgroup> shared_k: array<f32, 128>;\nvar<workgroup> shared_norm: array<f32, 128>;\n\n@compute @workgroup_size(128)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let h = wid.x;  // uniform within workgroup\n  if (h >= params.num_heads) { return; }\n\n  let d = lid.x;  // 0..127, all threads participate in barriers\n\n  let A_val = -exp(A_log[h]);\n  let state_base = h * params.key_dim * params.val_dim;\n  let scale = 1.0 / sqrt(f32(params.key_dim));\n\n  let q_off = h * params.key_dim;\n  let k_off = params.num_heads * params.key_dim + h * params.key_dim;\n  let v_off = 2u * params.num_heads * params.key_dim + h * params.val_dim;\n  let out_stride = params.num_heads * params.val_dim;\n\n  for (var t: u32 = 0u; t < params.T; t = t + 1u) {\n    let qkv_base = t * params.qkv_dim;\n\n    // ── Load Q and K into shared memory ──\n    shared_q[d] = qkv_conv[qkv_base + q_off + d];\n    shared_k[d] = qkv_conv[qkv_base + k_off + d];\n    workgroupBarrier();\n\n    // ── L2-normalize Q (parallel reduction) ──\n    shared_norm[d] = shared_q[d] * shared_q[d];\n    workgroupBarrier();\n    var stride: u32 = 64u;\n    while (stride > 0u) {\n      if (d < stride) { shared_norm[d] += shared_norm[d + stride]; }\n      workgroupBarrier();\n      stride /= 2u;\n    }\n    let q_inv_norm = 1.0 / sqrt(shared_norm[0] + 1e-12);\n    shared_q[d] *= q_inv_norm;\n\n    // ── L2-normalize K (parallel reduction) ──\n    shared_norm[d] = shared_k[d] * shared_k[d];\n    workgroupBarrier();\n    stride = 64u;\n    while (stride > 0u) {\n      if (d < stride) { shared_norm[d] += shared_norm[d + stride]; }\n      workgroupBarrier();\n      stride /= 2u;\n    }\n    let k_inv_norm = 1.0 / sqrt(shared_norm[0] + 1e-12);\n    shared_k[d] *= k_inv_norm;\n    workgroupBarrier();\n\n    // ── Compute decay and beta ──\n    let a_val = a_input[t * params.num_heads + h];\n    let b_val = b_input[t * params.num_heads + h];\n    let dt_in = a_val + dt_bias[h];\n    let sp = log(1.0 + exp(max(dt_in, -80.0)));\n    let g = A_val * sp;\n    let decay = exp(max(g, -80.0));\n    let beta = 1.0 / (1.0 + exp(max(-b_val, -80.0)));\n\n    // ── 1+2. Retrieve from memory (decay is a per-step scalar, so it factors\n    // out of the dot product): kv_mem[d] = decay * Σ_k state[k,d] * k_norm[k] ──\n    let v_val = qkv_conv[qkv_base + v_off + d];\n    var kv_dot: f32 = 0.0;\n    for (var k: u32 = 0u; k < params.key_dim; k = k + 1u) {\n      kv_dot += ssm_state[state_base + k * params.val_dim + d] * shared_k[k];\n    }\n    let kv_mem = kv_dot * decay;\n\n    // ── 3. Delta rule: delta[d] = (v[d] - kv_mem[d]) * beta ──\n    let delta_val = (v_val - kv_mem) * beta;\n\n    // ── 4+5. Fused decay + rank-1 update + output in ONE pass over state\n    // (one read + one write instead of three reads + two writes):\n    //   state'[k,d] = state[k,d] * decay + k_norm[k] * delta[d]\n    //   out[d]      = Σ_k state'[k,d] * q_norm[k] * scale ──\n    var out_val: f32 = 0.0;\n    for (var k: u32 = 0u; k < params.key_dim; k = k + 1u) {\n      let idx = state_base + k * params.val_dim + d;\n      let s = ssm_state[idx] * decay + shared_k[k] * delta_val;\n      ssm_state[idx] = s;\n      out_val += s * shared_q[k];\n    }\n    output[t * out_stride + h * params.val_dim + d] = out_val * scale;\n\n    workgroupBarrier();  // sync before next timestep loads shared_q/k\n  }\n}\n`;\n\nconst WGSL_MAMBA_SSM_F16 = `\\\n// Gated Delta Net recurrence with f16 SSM state storage (Dawn only).\n// Compute stays f32; only the persistent state reads/writes are half-precision,\n// halving the dominant state bandwidth. Buffer is allocated as f32-sized and\n// reinterpreted (extra capacity unused).\n//\n// For each head, maintains state [key_dim, val_dim] across timesteps.\n// Recurrence per timestep:\n//   g = -exp(A_log) * softplus(a + dt_bias)   (log-space decay)\n//   decay = exp(g)                             (actual decay in (0,1))\n//   state *= decay                             (decay state)\n//   q_norm, k_norm = L2_normalize(q), L2_normalize(k)\n//   kv_mem = state^T @ k_norm                  (retrieve from memory)\n//   delta = (v - kv_mem) * sigmoid(b)          (delta rule)\n//   state += k_norm @ delta^T                  (rank-1 update)\n//   output = state^T @ q_norm / sqrt(key_dim)  (query output)\n\nstruct Params {\n  T: u32,\n  num_heads: u32,\n  key_dim: u32,\n  val_dim: u32,\n  qkv_dim: u32,\n  _pad1: u32,\n  _pad2: u32,\n  _pad3: u32,\n}\n\n@group(0) @binding(0) var<storage, read> qkv_conv: array<f32>;\n@group(0) @binding(1) var<storage, read> a_input: array<f32>;\n@group(0) @binding(2) var<storage, read> b_input: array<f32>;\n@group(0) @binding(3) var<storage, read> A_log: array<f32>;\n@group(0) @binding(4) var<storage, read> dt_bias: array<f32>;\n@group(0) @binding(5) var<storage, read_write> ssm_state: array<f16>;\n@group(0) @binding(6) var<storage, read_write> output: array<f32>;\n@group(0) @binding(7) var<storage, read> params: Params;\n\n// Shared memory for Q and K vectors (L2-normalized) and reduction\nvar<workgroup> shared_q: array<f32, 128>;\nvar<workgroup> shared_k: array<f32, 128>;\nvar<workgroup> shared_norm: array<f32, 128>;\n\n@compute @workgroup_size(128)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let h = wid.x;  // uniform within workgroup\n  if (h >= params.num_heads) { return; }\n\n  let d = lid.x;  // 0..127, all threads participate in barriers\n\n  let A_val = -exp(A_log[h]);\n  let state_base = h * params.key_dim * params.val_dim;\n  let scale = 1.0 / sqrt(f32(params.key_dim));\n\n  let q_off = h * params.key_dim;\n  let k_off = params.num_heads * params.key_dim + h * params.key_dim;\n  let v_off = 2u * params.num_heads * params.key_dim + h * params.val_dim;\n  let out_stride = params.num_heads * params.val_dim;\n\n  for (var t: u32 = 0u; t < params.T; t = t + 1u) {\n    let qkv_base = t * params.qkv_dim;\n\n    // ── Load Q and K into shared memory ──\n    shared_q[d] = qkv_conv[qkv_base + q_off + d];\n    shared_k[d] = qkv_conv[qkv_base + k_off + d];\n    workgroupBarrier();\n\n    // ── L2-normalize Q (parallel reduction) ──\n    shared_norm[d] = shared_q[d] * shared_q[d];\n    workgroupBarrier();\n    var stride: u32 = 64u;\n    while (stride > 0u) {\n      if (d < stride) { shared_norm[d] += shared_norm[d + stride]; }\n      workgroupBarrier();\n      stride /= 2u;\n    }\n    let q_inv_norm = 1.0 / sqrt(shared_norm[0] + 1e-12);\n    shared_q[d] *= q_inv_norm;\n\n    // ── L2-normalize K (parallel reduction) ──\n    shared_norm[d] = shared_k[d] * shared_k[d];\n    workgroupBarrier();\n    stride = 64u;\n    while (stride > 0u) {\n      if (d < stride) { shared_norm[d] += shared_norm[d + stride]; }\n      workgroupBarrier();\n      stride /= 2u;\n    }\n    let k_inv_norm = 1.0 / sqrt(shared_norm[0] + 1e-12);\n    shared_k[d] *= k_inv_norm;\n    workgroupBarrier();\n\n    // ── Compute decay and beta ──\n    let a_val = a_input[t * params.num_heads + h];\n    let b_val = b_input[t * params.num_heads + h];\n    let dt_in = a_val + dt_bias[h];\n    let sp = log(1.0 + exp(max(dt_in, -80.0)));\n    let g = A_val * sp;\n    let decay = exp(max(g, -80.0));\n    let beta = 1.0 / (1.0 + exp(max(-b_val, -80.0)));\n\n    // ── 1+2. Retrieve from memory (decay is a per-step scalar, so it factors\n    // out of the dot product): kv_mem[d] = decay * Σ_k state[k,d] * k_norm[k] ──\n    let v_val = qkv_conv[qkv_base + v_off + d];\n    var kv_dot: f32 = 0.0;\n    for (var k: u32 = 0u; k < params.key_dim; k = k + 1u) {\n      kv_dot += f32(ssm_state[state_base + k * params.val_dim + d]) * shared_k[k];\n    }\n    let kv_mem = kv_dot * decay;\n\n    // ── 3. Delta rule: delta[d] = (v[d] - kv_mem[d]) * beta ──\n    let delta_val = (v_val - kv_mem) * beta;\n\n    // ── 4+5. Fused decay + rank-1 update + output in ONE pass over state\n    // (one read + one write instead of three reads + two writes):\n    //   state'[k,d] = state[k,d] * decay + k_norm[k] * delta[d]\n    //   out[d]      = Σ_k state'[k,d] * q_norm[k] * scale ──\n    var out_val: f32 = 0.0;\n    for (var k: u32 = 0u; k < params.key_dim; k = k + 1u) {\n      let idx = state_base + k * params.val_dim + d;\n      let s = f32(ssm_state[idx]) * decay + shared_k[k] * delta_val;\n      ssm_state[idx] = f16(s);\n      out_val += s * shared_q[k];\n    }\n    output[t * out_stride + h * params.val_dim + d] = out_val * scale;\n\n    workgroupBarrier();  // sync before next timestep loads shared_q/k\n  }\n}\n`;\n\nconst WGSL_KV_CACHE_APPEND = `\\\n// Copy T new K or V vectors into cache at position offset.\n// src: [T, width], dst: [max_seq_len, width]\n\nstruct Params {\n  count: u32,       // T * width (total elements to copy)\n  dst_offset: u32,  // seqPos * width (element offset into cache)\n}\n\n@group(0) @binding(0) var<storage, read> src: array<f32>;\n@group(0) @binding(1) var<storage, read_write> dst: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n  dst[params.dst_offset + idx] = src[idx];\n}\n`;\n\nconst WGSL_DUAL_KV_CACHE_APPEND = `\\\n// Append both K and V into their f32 caches in one dispatch.\n// Both share the same width, T, and dst_offset (same kv_dim, same seqPos),\n// so a single param set + a single grid over 'count' elements drives both\n// copies. Pure memcpy — numerically identical to two separate appends.\n\nstruct Params {\n  count: u32,       // T * width (per-cache element count)\n  dst_offset: u32,  // seqPos * width (element offset into each cache)\n}\n\n@group(0) @binding(0) var<storage, read> src_k: array<f32>;\n@group(0) @binding(1) var<storage, read> src_v: array<f32>;\n@group(0) @binding(2) var<storage, read_write> dst_k: array<f32>;\n@group(0) @binding(3) var<storage, read_write> dst_v: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n  dst_k[params.dst_offset + idx] = src_k[idx];\n  dst_v[params.dst_offset + idx] = src_v[idx];\n}\n`;\n\nconst WGSL_DUAL_KV_CACHE_APPEND_F16 = `\\\n// Append both K and V into their native-f16 caches in one dispatch.\n// Same per-element conversion as the single f16 append, doubled.\n\nstruct Params {\n  count: u32,\n  dst_offset: u32,\n}\n\n@group(0) @binding(0) var<storage, read> src_k: array<f32>;\n@group(0) @binding(1) var<storage, read> src_v: array<f32>;\n@group(0) @binding(2) var<storage, read_write> dst_k: array<f16>;\n@group(0) @binding(3) var<storage, read_write> dst_v: array<f16>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n  dst_k[params.dst_offset + idx] = f16(src_k[idx]);\n  dst_v[params.dst_offset + idx] = f16(src_v[idx]);\n}\n`;\n\nconst WGSL_DUAL_KV_CACHE_APPEND_PACKED_F16 = `\\\n// Append both K and V into packed-f16 caches in one dispatch (Safari-safe).\n// Each u32 packs 2 f16 via pack2x16float; one thread handles one f16 pair per\n// cache. Same math as the single packed append, doubled — no 'enable f16'.\n\nstruct Params {\n  count: u32,       // T * width (per-cache f32 element count)\n  dst_offset: u32,  // seqPos * width (f32-element offset)\n}\n\n@group(0) @binding(0) var<storage, read> src_k: array<f32>;\n@group(0) @binding(1) var<storage, read> src_v: array<f32>;\n@group(0) @binding(2) var<storage, read_write> dst_k: array<u32>;\n@group(0) @binding(3) var<storage, read_write> dst_v: array<u32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let pair_idx = gid.x;\n  let num_pairs = params.count >> 1u;\n  if (pair_idx >= num_pairs) { return; }\n\n  let src_base = pair_idx * 2u;\n  let dst_idx = (params.dst_offset >> 1u) + pair_idx;\n\n  dst_k[dst_idx] = pack2x16float(vec2f(src_k[src_base], src_k[src_base + 1u]));\n  dst_v[dst_idx] = pack2x16float(vec2f(src_v[src_base], src_v[src_base + 1u]));\n}\n`;\n\nconst WGSL_KV_CACHE_APPEND_F16 = `\\\n// Copy T new K or V vectors into f16 cache at position offset.\n// src: [T, width] (f32), dst: [max_seq_len, width] (f16)\n// Converts f32 activations to f16 on write for reduced memory traffic.\n\nstruct Params {\n  count: u32,       // T * width (total elements to copy)\n  dst_offset: u32,  // seqPos * width (element offset into cache)\n}\n\n@group(0) @binding(0) var<storage, read> src: array<f32>;\n@group(0) @binding(1) var<storage, read_write> dst: array<f16>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n  dst[params.dst_offset + idx] = f16(src[idx]);\n}\n`;\n\nconst WGSL_KV_CACHE_APPEND_PACKED_F16 = `\\\n// Copy T new K or V vectors into packed f16 cache (Safari-safe).\n// src: [T, width] (f32), dst: [max_seq_len, width/2] (u32, packed f16 pairs)\n// Each u32 holds 2 f16 values via pack2x16float. Does NOT require 'enable f16'.\n// Params use f32-element-level addressing; kernel converts to u32 pair indices.\n\nstruct Params {\n  count: u32,       // T * width (total f32 elements)\n  dst_offset: u32,  // seqPos * width (f32-element offset into cache)\n}\n\n@group(0) @binding(0) var<storage, read> src: array<f32>;\n@group(0) @binding(1) var<storage, read_write> dst: array<u32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let pair_idx = gid.x;\n  let num_pairs = params.count >> 1u;\n  if (pair_idx >= num_pairs) { return; }\n\n  let src_base = pair_idx * 2u;\n  let dst_idx = (params.dst_offset >> 1u) + pair_idx;\n\n  dst[dst_idx] = pack2x16float(vec2f(src[src_base], src[src_base + 1u]));\n}\n`;\n\nconst WGSL_ATTENTION_F16 = `\\\n// Tiled online-softmax attention with f16 KV cache (Flash-Attention style).\n//\n// Same algorithm as WGSL_ATTENTION but K/V are stored as f16 in global memory.\n// Shared memory stays f32 for computation accuracy — only global reads change.\n// Halves KV cache memory traffic for ~1.3-1.8x decode speedup.\n\nconst TILE_S: u32 = 16u;\nconst WG: u32 = 256u;\nconst DOT_THREADS: u32 = 16u;\n\nstruct Params {\n  T: u32,\n  S: u32,\n  num_q_heads: u32,\n  num_kv_heads: u32,\n  head_dim: u32,\n  position_offset: u32,\n  // Softmax QK temperature (f32). Default 1/sqrt(head_dim); Gemma 4 sets 1.0.\n  attn_scale: f32,\n}\n\n@group(0) @binding(0) var<storage, read> Q: array<f32>;\n@group(0) @binding(1) var<storage, read> K: array<f16>;\n@group(0) @binding(2) var<storage, read> V: array<f16>;\n@group(0) @binding(3) var<storage, read_write> output: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\nvar<workgroup> smem: array<f32, 4096>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let q_pos = wid.x;\n  let q_head = wid.y;\n\n  if (q_pos >= params.T || q_head >= params.num_q_heads) { return; }\n\n  let tid = lid.x;\n  let heads_per_kv = params.num_q_heads / params.num_kv_heads;\n  let kv_head = q_head / heads_per_kv;\n  let scale = params.attn_scale;\n\n  let q_stride = params.num_q_heads * params.head_dim;\n  let kv_stride = params.num_kv_heads * params.head_dim;\n  let q_base = q_pos * q_stride + q_head * params.head_dim;\n  let kv_head_offset = kv_head * params.head_dim;\n  let causal_limit = q_pos + params.position_offset + 1u;\n  let hd = params.head_dim;\n  let S_eff = min(params.S, causal_limit);\n\n  var running_max: f32 = -1e30;\n  var running_sum: f32 = 0.0;\n  // Per-thread output accumulators; out_acc2 covers dim tid+WG so head_dim up to\n  // 2*WG (512) works with WG=256 threads. hd<=256 → out_acc2 unused (identical).\n  var out_acc: f32 = 0.0;\n  var out_acc2: f32 = 0.0;\n  let d2 = tid + WG;\n\n  // Tile capped so a K/V tile (tile_cap*hd f32) fits the 4096-f32 smem.\n  let tile_cap = min(TILE_S, 4096u / hd);\n  let num_tiles = (S_eff + tile_cap - 1u) / tile_cap;\n\n  for (var tile: u32 = 0u; tile < num_tiles; tile += 1u) {\n    let tile_start = tile * tile_cap;\n    let tile_end = min(tile_start + tile_cap, S_eff);\n    let tile_len = tile_end - tile_start;\n\n    // ── Step 1: Cooperative K tile load (f16 → f32 conversion) ──\n    let total_k_elems = tile_len * hd;\n    var load_idx = tid;\n    while (load_idx < total_k_elems) {\n      let s_local = load_idx / hd;\n      let d = load_idx % hd;\n      let s_global = tile_start + s_local;\n      let k_addr = s_global * kv_stride + kv_head_offset + d;\n      smem[s_local * hd + d] = f32(K[k_addr]);\n      load_idx += WG;\n    }\n    workgroupBarrier();\n\n    // ── Step 2: Parallel Q·K dot products ──\n    let pos_in_tile = tid / DOT_THREADS;\n    let dot_tid = tid % DOT_THREADS;\n\n    var my_partial: f32 = 0.0;\n    if (pos_in_tile < tile_len) {\n      let dims_per_thread = hd / DOT_THREADS;\n      let d_start = dot_tid * dims_per_thread;\n      let d_end = d_start + dims_per_thread;\n      let kv_row_base = pos_in_tile * hd;\n\n      var d = d_start;\n      while (d + 3u < d_end) {\n        let q_v = vec4f(Q[q_base + d], Q[q_base + d + 1u],\n                         Q[q_base + d + 2u], Q[q_base + d + 3u]);\n        let k_v = vec4f(smem[kv_row_base + d], smem[kv_row_base + d + 1u],\n                         smem[kv_row_base + d + 2u], smem[kv_row_base + d + 3u]);\n        my_partial += dot(q_v, k_v);\n        d += 4u;\n      }\n      while (d < d_end) {\n        my_partial += Q[q_base + d] * smem[kv_row_base + d];\n        d += 1u;\n      }\n    }\n\n    workgroupBarrier();\n    smem[tid] = my_partial;\n    workgroupBarrier();\n\n    // Two-phase: leader 0 reads smem[1..15], which are exactly the slots\n    // leaders 1..15 write — reduce into a local first, barrier (in uniform\n    // control flow), then write. Read+write in one block is a data race.\n    var group_total: f32 = 0.0;\n    if (dot_tid == 0u && pos_in_tile < tile_len) {\n      group_total = smem[tid];\n      for (var i: u32 = 1u; i < DOT_THREADS; i += 1u) {\n        group_total += smem[tid + i];\n      }\n    }\n    workgroupBarrier();\n    if (dot_tid == 0u && pos_in_tile < tile_len) {\n      smem[pos_in_tile] = group_total * scale;\n    }\n    workgroupBarrier();\n\n    // ── Step 3: Online softmax ──\n    if (tid < TILE_S) {\n      if (tid < tile_len) {\n        smem[TILE_S + tid] = smem[tid];\n      } else {\n        smem[TILE_S + tid] = -1e30;\n      }\n    }\n    workgroupBarrier();\n    if (tid < 8u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 8u]); }\n    workgroupBarrier();\n    if (tid < 4u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 4u]); }\n    workgroupBarrier();\n    if (tid < 2u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 2u]); }\n    workgroupBarrier();\n    if (tid < 1u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 1u]); }\n    workgroupBarrier();\n    let tile_max = smem[TILE_S];\n\n    let new_max = max(running_max, tile_max);\n    let old_correction = exp(max(running_max - new_max, -80.0));\n\n    if (tid < tile_len) {\n      smem[tid] = exp(max(smem[tid] - new_max, -80.0));\n    }\n    workgroupBarrier();\n\n    if (tid < TILE_S) {\n      if (tid < tile_len) {\n        smem[TILE_S + tid] = smem[tid];\n      } else {\n        smem[TILE_S + tid] = 0.0;\n      }\n    }\n    workgroupBarrier();\n    if (tid < 8u) { smem[TILE_S + tid] += smem[TILE_S + tid + 8u]; }\n    workgroupBarrier();\n    if (tid < 4u) { smem[TILE_S + tid] += smem[TILE_S + tid + 4u]; }\n    workgroupBarrier();\n    if (tid < 2u) { smem[TILE_S + tid] += smem[TILE_S + tid + 2u]; }\n    workgroupBarrier();\n    if (tid < 1u) { smem[TILE_S + tid] += smem[TILE_S + tid + 1u]; }\n    workgroupBarrier();\n    let tile_sum = smem[TILE_S];\n\n    // ── Save scores to registers before V overwrites smem ──\n    var w0: f32 = 0.0; var w1: f32 = 0.0; var w2: f32 = 0.0; var w3: f32 = 0.0;\n    var w4: f32 = 0.0; var w5: f32 = 0.0; var w6: f32 = 0.0; var w7: f32 = 0.0;\n    var w8: f32 = 0.0; var w9: f32 = 0.0; var w10: f32 = 0.0; var w11: f32 = 0.0;\n    var w12: f32 = 0.0; var w13: f32 = 0.0; var w14: f32 = 0.0; var w15: f32 = 0.0;\n    if (tid < hd) {\n      if (0u < tile_len) { w0 = smem[0]; }\n      if (1u < tile_len) { w1 = smem[1]; }\n      if (2u < tile_len) { w2 = smem[2]; }\n      if (3u < tile_len) { w3 = smem[3]; }\n      if (4u < tile_len) { w4 = smem[4]; }\n      if (5u < tile_len) { w5 = smem[5]; }\n      if (6u < tile_len) { w6 = smem[6]; }\n      if (7u < tile_len) { w7 = smem[7]; }\n      if (8u < tile_len) { w8 = smem[8]; }\n      if (9u < tile_len) { w9 = smem[9]; }\n      if (10u < tile_len) { w10 = smem[10]; }\n      if (11u < tile_len) { w11 = smem[11]; }\n      if (12u < tile_len) { w12 = smem[12]; }\n      if (13u < tile_len) { w13 = smem[13]; }\n      if (14u < tile_len) { w14 = smem[14]; }\n      if (15u < tile_len) { w15 = smem[15]; }\n    }\n    workgroupBarrier();\n\n    // ── Step 4: Cooperative V tile load (f16 → f32 conversion, full smem reuse) ──\n    let total_v_elems = tile_len * hd;\n    load_idx = tid;\n    while (load_idx < total_v_elems) {\n      let s_local = load_idx / hd;\n      let d = load_idx % hd;\n      let s_global = tile_start + s_local;\n      let v_addr = s_global * kv_stride + kv_head_offset + d;\n      smem[s_local * hd + d] = f32(V[v_addr]);\n      load_idx += WG;\n    }\n    workgroupBarrier();\n\n    // ── Step 5: Parallel V accumulation ──\n    if (tid < hd) {\n      out_acc = out_acc * old_correction;\n      if (0u < tile_len) { out_acc += w0 * smem[0u * hd + tid]; }\n      if (1u < tile_len) { out_acc += w1 * smem[1u * hd + tid]; }\n      if (2u < tile_len) { out_acc += w2 * smem[2u * hd + tid]; }\n      if (3u < tile_len) { out_acc += w3 * smem[3u * hd + tid]; }\n      if (4u < tile_len) { out_acc += w4 * smem[4u * hd + tid]; }\n      if (5u < tile_len) { out_acc += w5 * smem[5u * hd + tid]; }\n      if (6u < tile_len) { out_acc += w6 * smem[6u * hd + tid]; }\n      if (7u < tile_len) { out_acc += w7 * smem[7u * hd + tid]; }\n      if (8u < tile_len) { out_acc += w8 * smem[8u * hd + tid]; }\n      if (9u < tile_len) { out_acc += w9 * smem[9u * hd + tid]; }\n      if (10u < tile_len) { out_acc += w10 * smem[10u * hd + tid]; }\n      if (11u < tile_len) { out_acc += w11 * smem[11u * hd + tid]; }\n      if (12u < tile_len) { out_acc += w12 * smem[12u * hd + tid]; }\n      if (13u < tile_len) { out_acc += w13 * smem[13u * hd + tid]; }\n      if (14u < tile_len) { out_acc += w14 * smem[14u * hd + tid]; }\n      if (15u < tile_len) { out_acc += w15 * smem[15u * hd + tid]; }\n    }\n    if (d2 < hd) {\n      out_acc2 = out_acc2 * old_correction;\n      if (0u < tile_len) { out_acc2 += w0 * smem[0u * hd + d2]; }\n      if (1u < tile_len) { out_acc2 += w1 * smem[1u * hd + d2]; }\n      if (2u < tile_len) { out_acc2 += w2 * smem[2u * hd + d2]; }\n      if (3u < tile_len) { out_acc2 += w3 * smem[3u * hd + d2]; }\n      if (4u < tile_len) { out_acc2 += w4 * smem[4u * hd + d2]; }\n      if (5u < tile_len) { out_acc2 += w5 * smem[5u * hd + d2]; }\n      if (6u < tile_len) { out_acc2 += w6 * smem[6u * hd + d2]; }\n      if (7u < tile_len) { out_acc2 += w7 * smem[7u * hd + d2]; }\n      if (8u < tile_len) { out_acc2 += w8 * smem[8u * hd + d2]; }\n      if (9u < tile_len) { out_acc2 += w9 * smem[9u * hd + d2]; }\n      if (10u < tile_len) { out_acc2 += w10 * smem[10u * hd + d2]; }\n      if (11u < tile_len) { out_acc2 += w11 * smem[11u * hd + d2]; }\n      if (12u < tile_len) { out_acc2 += w12 * smem[12u * hd + d2]; }\n      if (13u < tile_len) { out_acc2 += w13 * smem[13u * hd + d2]; }\n      if (14u < tile_len) { out_acc2 += w14 * smem[14u * hd + d2]; }\n      if (15u < tile_len) { out_acc2 += w15 * smem[15u * hd + d2]; }\n    }\n\n    running_sum = running_sum * old_correction + tile_sum;\n    running_max = new_max;\n    workgroupBarrier();\n  }\n\n  // ── Write normalized output ──\n  var inv_sum: f32 = 0.0;\n  if (running_sum > 0.0) { inv_sum = 1.0 / running_sum; }\n  let out_base = q_pos * q_stride + q_head * hd;\n  if (tid < hd) {\n    output[out_base + tid] = out_acc * inv_sum;\n  }\n  if (d2 < hd) {\n    output[out_base + d2] = out_acc2 * inv_sum;\n  }\n}\n`;\n\nconst WGSL_ATTENTION_PACKED_F16 = `\\\n// Tiled online-softmax attention with packed f16 KV cache (Safari-safe).\n//\n// K/V stored as array<u32> — each u32 holds 2 f16 values via pack2x16float.\n// Does NOT require 'enable f16' — only uses u32 + f32 + unpack2x16float.\n// Same algorithm as native f16 attention, same memory savings.\n//\n// Safari/Metal compatibility:\n//   - Uses if/else instead of select() (Safari has select() bugs)\n//   - Clamps exp() args to -80 (Metal can return NaN for exp(-1e30))\n//   - No f16 types avoids WebKit WGSL compiler miscompilation\n//\n// Shared memory: 4096 f32 = 16384 bytes = exactly 16 KB (minimum WebGPU guarantee)\n\nconst TILE_S: u32 = 16u;\nconst WG: u32 = 256u;\nconst DOT_THREADS: u32 = 16u;\n\nstruct Params {\n  T: u32,\n  S: u32,\n  num_q_heads: u32,\n  num_kv_heads: u32,\n  head_dim: u32,\n  position_offset: u32,\n  // Softmax QK temperature (f32). Default 1/sqrt(head_dim); Gemma 4 sets 1.0.\n  attn_scale: f32,\n}\n\n@group(0) @binding(0) var<storage, read> Q: array<f32>;\n@group(0) @binding(1) var<storage, read> K: array<u32>;\n@group(0) @binding(2) var<storage, read> V: array<u32>;\n@group(0) @binding(3) var<storage, read_write> output: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\nvar<workgroup> smem: array<f32, 4096>;\n\n// Load a single f32 from packed u32 K cache at element-level index\nfn load_k(elem_idx: u32) -> f32 {\n  let pair = unpack2x16float(K[elem_idx >> 1u]);\n  if ((elem_idx & 1u) == 0u) {\n    return pair.x;\n  } else {\n    return pair.y;\n  }\n}\n\n// Load a single f32 from packed u32 V cache at element-level index\nfn load_v(elem_idx: u32) -> f32 {\n  let pair = unpack2x16float(V[elem_idx >> 1u]);\n  if ((elem_idx & 1u) == 0u) {\n    return pair.x;\n  } else {\n    return pair.y;\n  }\n}\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let q_pos = wid.x;\n  let q_head = wid.y;\n\n  if (q_pos >= params.T || q_head >= params.num_q_heads) { return; }\n\n  let tid = lid.x;\n  let heads_per_kv = params.num_q_heads / params.num_kv_heads;\n  let kv_head = q_head / heads_per_kv;\n  let scale = params.attn_scale;\n\n  let q_stride = params.num_q_heads * params.head_dim;\n  let kv_stride = params.num_kv_heads * params.head_dim;\n  let q_base = q_pos * q_stride + q_head * params.head_dim;\n  let kv_head_offset = kv_head * params.head_dim;\n  let causal_limit = q_pos + params.position_offset + 1u;\n  let hd = params.head_dim;\n  let S_eff = min(params.S, causal_limit);\n\n  var running_max: f32 = -1e30;\n  var running_sum: f32 = 0.0;\n  // Per-thread output accumulators; out_acc2 covers dim tid+WG so head_dim up to\n  // 2*WG (512) works with WG=256 threads. hd<=256 → out_acc2 unused (identical).\n  var out_acc: f32 = 0.0;\n  var out_acc2: f32 = 0.0;\n  let d2 = tid + WG;\n\n  // Tile capped so a K/V tile (tile_cap*hd f32) fits the 4096-f32 smem.\n  let tile_cap = min(TILE_S, 4096u / hd);\n  let num_tiles = (S_eff + tile_cap - 1u) / tile_cap;\n\n  for (var tile: u32 = 0u; tile < num_tiles; tile += 1u) {\n    let tile_start = tile * tile_cap;\n    let tile_end = min(tile_start + tile_cap, S_eff);\n    let tile_len = tile_end - tile_start;\n\n    // ── Step 1: Cooperative K tile load (packed u32 → f32 via unpack2x16float) ──\n    let total_k_elems = tile_len * hd;\n    var load_idx = tid;\n    while (load_idx < total_k_elems) {\n      let s_local = load_idx / hd;\n      let d = load_idx % hd;\n      let s_global = tile_start + s_local;\n      let k_addr = s_global * kv_stride + kv_head_offset + d;\n      smem[s_local * hd + d] = load_k(k_addr);\n      load_idx += WG;\n    }\n    workgroupBarrier();\n\n    // ── Step 2: Parallel Q·K dot products ──\n    let pos_in_tile = tid / DOT_THREADS;\n    let dot_tid = tid % DOT_THREADS;\n\n    var my_partial: f32 = 0.0;\n    if (pos_in_tile < tile_len) {\n      let dims_per_thread = hd / DOT_THREADS;\n      let d_start = dot_tid * dims_per_thread;\n      let d_end = d_start + dims_per_thread;\n      let kv_row_base = pos_in_tile * hd;\n\n      var d = d_start;\n      while (d + 3u < d_end) {\n        let q_v = vec4f(Q[q_base + d], Q[q_base + d + 1u],\n                         Q[q_base + d + 2u], Q[q_base + d + 3u]);\n        let k_v = vec4f(smem[kv_row_base + d], smem[kv_row_base + d + 1u],\n                         smem[kv_row_base + d + 2u], smem[kv_row_base + d + 3u]);\n        my_partial += dot(q_v, k_v);\n        d += 4u;\n      }\n      while (d < d_end) {\n        my_partial += Q[q_base + d] * smem[kv_row_base + d];\n        d += 1u;\n      }\n    }\n\n    workgroupBarrier();\n    smem[tid] = my_partial;\n    workgroupBarrier();\n\n    // Two-phase: leader 0 reads smem[1..15], which are exactly the slots\n    // leaders 1..15 write — reduce into a local first, barrier (in uniform\n    // control flow), then write. Read+write in one block is a data race.\n    var group_total: f32 = 0.0;\n    if (dot_tid == 0u && pos_in_tile < tile_len) {\n      group_total = smem[tid];\n      for (var i: u32 = 1u; i < DOT_THREADS; i += 1u) {\n        group_total += smem[tid + i];\n      }\n    }\n    workgroupBarrier();\n    if (dot_tid == 0u && pos_in_tile < tile_len) {\n      smem[pos_in_tile] = group_total * scale;\n    }\n    workgroupBarrier();\n\n    // ── Step 3: Online softmax ──\n    if (tid < TILE_S) {\n      if (tid < tile_len) {\n        smem[TILE_S + tid] = smem[tid];\n      } else {\n        smem[TILE_S + tid] = -1e30;\n      }\n    }\n    workgroupBarrier();\n    if (tid < 8u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 8u]); }\n    workgroupBarrier();\n    if (tid < 4u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 4u]); }\n    workgroupBarrier();\n    if (tid < 2u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 2u]); }\n    workgroupBarrier();\n    if (tid < 1u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 1u]); }\n    workgroupBarrier();\n    let tile_max = smem[TILE_S];\n\n    let new_max = max(running_max, tile_max);\n    let old_correction = exp(max(running_max - new_max, -80.0));\n\n    if (tid < tile_len) {\n      smem[tid] = exp(max(smem[tid] - new_max, -80.0));\n    }\n    workgroupBarrier();\n\n    if (tid < TILE_S) {\n      if (tid < tile_len) {\n        smem[TILE_S + tid] = smem[tid];\n      } else {\n        smem[TILE_S + tid] = 0.0;\n      }\n    }\n    workgroupBarrier();\n    if (tid < 8u) { smem[TILE_S + tid] += smem[TILE_S + tid + 8u]; }\n    workgroupBarrier();\n    if (tid < 4u) { smem[TILE_S + tid] += smem[TILE_S + tid + 4u]; }\n    workgroupBarrier();\n    if (tid < 2u) { smem[TILE_S + tid] += smem[TILE_S + tid + 2u]; }\n    workgroupBarrier();\n    if (tid < 1u) { smem[TILE_S + tid] += smem[TILE_S + tid + 1u]; }\n    workgroupBarrier();\n    let tile_sum = smem[TILE_S];\n\n    // ── Save scores to registers before V overwrites smem ──\n    var w0: f32 = 0.0; var w1: f32 = 0.0; var w2: f32 = 0.0; var w3: f32 = 0.0;\n    var w4: f32 = 0.0; var w5: f32 = 0.0; var w6: f32 = 0.0; var w7: f32 = 0.0;\n    var w8: f32 = 0.0; var w9: f32 = 0.0; var w10: f32 = 0.0; var w11: f32 = 0.0;\n    var w12: f32 = 0.0; var w13: f32 = 0.0; var w14: f32 = 0.0; var w15: f32 = 0.0;\n    if (tid < hd) {\n      if (0u < tile_len) { w0 = smem[0]; }\n      if (1u < tile_len) { w1 = smem[1]; }\n      if (2u < tile_len) { w2 = smem[2]; }\n      if (3u < tile_len) { w3 = smem[3]; }\n      if (4u < tile_len) { w4 = smem[4]; }\n      if (5u < tile_len) { w5 = smem[5]; }\n      if (6u < tile_len) { w6 = smem[6]; }\n      if (7u < tile_len) { w7 = smem[7]; }\n      if (8u < tile_len) { w8 = smem[8]; }\n      if (9u < tile_len) { w9 = smem[9]; }\n      if (10u < tile_len) { w10 = smem[10]; }\n      if (11u < tile_len) { w11 = smem[11]; }\n      if (12u < tile_len) { w12 = smem[12]; }\n      if (13u < tile_len) { w13 = smem[13]; }\n      if (14u < tile_len) { w14 = smem[14]; }\n      if (15u < tile_len) { w15 = smem[15]; }\n    }\n    workgroupBarrier();\n\n    // ── Step 4: Cooperative V tile load (packed u32 → f32 via unpack2x16float) ──\n    let total_v_elems = tile_len * hd;\n    load_idx = tid;\n    while (load_idx < total_v_elems) {\n      let s_local = load_idx / hd;\n      let d = load_idx % hd;\n      let s_global = tile_start + s_local;\n      let v_addr = s_global * kv_stride + kv_head_offset + d;\n      smem[s_local * hd + d] = load_v(v_addr);\n      load_idx += WG;\n    }\n    workgroupBarrier();\n\n    // ── Step 5: Parallel V accumulation ──\n    if (tid < hd) {\n      out_acc = out_acc * old_correction;\n      if (0u < tile_len) { out_acc += w0 * smem[0u * hd + tid]; }\n      if (1u < tile_len) { out_acc += w1 * smem[1u * hd + tid]; }\n      if (2u < tile_len) { out_acc += w2 * smem[2u * hd + tid]; }\n      if (3u < tile_len) { out_acc += w3 * smem[3u * hd + tid]; }\n      if (4u < tile_len) { out_acc += w4 * smem[4u * hd + tid]; }\n      if (5u < tile_len) { out_acc += w5 * smem[5u * hd + tid]; }\n      if (6u < tile_len) { out_acc += w6 * smem[6u * hd + tid]; }\n      if (7u < tile_len) { out_acc += w7 * smem[7u * hd + tid]; }\n      if (8u < tile_len) { out_acc += w8 * smem[8u * hd + tid]; }\n      if (9u < tile_len) { out_acc += w9 * smem[9u * hd + tid]; }\n      if (10u < tile_len) { out_acc += w10 * smem[10u * hd + tid]; }\n      if (11u < tile_len) { out_acc += w11 * smem[11u * hd + tid]; }\n      if (12u < tile_len) { out_acc += w12 * smem[12u * hd + tid]; }\n      if (13u < tile_len) { out_acc += w13 * smem[13u * hd + tid]; }\n      if (14u < tile_len) { out_acc += w14 * smem[14u * hd + tid]; }\n      if (15u < tile_len) { out_acc += w15 * smem[15u * hd + tid]; }\n    }\n    if (d2 < hd) {\n      out_acc2 = out_acc2 * old_correction;\n      if (0u < tile_len) { out_acc2 += w0 * smem[0u * hd + d2]; }\n      if (1u < tile_len) { out_acc2 += w1 * smem[1u * hd + d2]; }\n      if (2u < tile_len) { out_acc2 += w2 * smem[2u * hd + d2]; }\n      if (3u < tile_len) { out_acc2 += w3 * smem[3u * hd + d2]; }\n      if (4u < tile_len) { out_acc2 += w4 * smem[4u * hd + d2]; }\n      if (5u < tile_len) { out_acc2 += w5 * smem[5u * hd + d2]; }\n      if (6u < tile_len) { out_acc2 += w6 * smem[6u * hd + d2]; }\n      if (7u < tile_len) { out_acc2 += w7 * smem[7u * hd + d2]; }\n      if (8u < tile_len) { out_acc2 += w8 * smem[8u * hd + d2]; }\n      if (9u < tile_len) { out_acc2 += w9 * smem[9u * hd + d2]; }\n      if (10u < tile_len) { out_acc2 += w10 * smem[10u * hd + d2]; }\n      if (11u < tile_len) { out_acc2 += w11 * smem[11u * hd + d2]; }\n      if (12u < tile_len) { out_acc2 += w12 * smem[12u * hd + d2]; }\n      if (13u < tile_len) { out_acc2 += w13 * smem[13u * hd + d2]; }\n      if (14u < tile_len) { out_acc2 += w14 * smem[14u * hd + d2]; }\n      if (15u < tile_len) { out_acc2 += w15 * smem[15u * hd + d2]; }\n    }\n\n    running_sum = running_sum * old_correction + tile_sum;\n    running_max = new_max;\n    workgroupBarrier();\n  }\n\n  // ── Write normalized output ──\n  var inv_sum: f32 = 0.0;\n  if (running_sum > 0.0) { inv_sum = 1.0 / running_sum; }\n  let out_base = q_pos * q_stride + q_head * hd;\n  if (tid < hd) {\n    output[out_base + tid] = out_acc * inv_sum;\n  }\n  if (d2 < hd) {\n    output[out_base + d2] = out_acc2 * inv_sum;\n  }\n}\n`;\n\nconst WGSL_CONV_STATE_UPDATE = `\\\n// Update conv1d rolling state buffer with latest input timesteps.\n// For T >= state_size: copy last state_size timesteps from input.\n// For T < state_size: shift old state left by T, append T new values.\n\nstruct Params {\n  seq_len: u32,\n  channels: u32,\n  state_size: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read_write> conv_state: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let c = gid.x;\n  if (c >= params.channels) { return; }\n\n  let ss = params.state_size;\n\n  if (params.seq_len >= ss) {\n    // Large batch (prefill): copy last state_size timesteps from input\n    for (var s: u32 = 0u; s < ss; s = s + 1u) {\n      let src_t = params.seq_len - ss + s;\n      conv_state[s * params.channels + c] = input[src_t * params.channels + c];\n    }\n  } else {\n    // Small batch (decode): shift old state left, append new values\n    // Read old state for this channel first to avoid read-write conflict\n    var old0: f32 = 0.0;\n    var old1: f32 = 0.0;\n    var old2: f32 = 0.0;\n    if (ss > 0u) { old0 = conv_state[0u * params.channels + c]; }\n    if (ss > 1u) { old1 = conv_state[1u * params.channels + c]; }\n    if (ss > 2u) { old2 = conv_state[2u * params.channels + c]; }\n\n    let shift = ss - params.seq_len;\n    for (var s: u32 = 0u; s < ss; s = s + 1u) {\n      var val: f32 = 0.0;\n      if (s < shift) {\n        // Read from shifted old state\n        let src_s = s + params.seq_len;\n        if (src_s == 0u) { val = old0; }\n        else if (src_s == 1u) { val = old1; }\n        else if (src_s == 2u) { val = old2; }\n      } else {\n        // Read from new input\n        let input_idx = s - shift;\n        val = input[input_idx * params.channels + c];\n      }\n      conv_state[s * params.channels + c] = val;\n    }\n  }\n}\n`;\n\n// PoolMatMul: pooled[Np, W] = poolW[Np, N] @ hidden[N, W].\n// The Gemma 4 ViT spatial average-pool, expressed as a host-built pooling matrix\n// (poolW[r, k] = 1/k² when patch k falls in pooled-cell r, else 0) times the\n// encoder output. One thread per output element accumulates over N. Deliberately\n// simple (no workgroup memory, no select(), no vec4 alignment assumption) — it\n// runs ONCE per image over tiny matrices (Np≤~512, N≤~4096, W=768), so a scalar\n// kernel is plenty fast and maximally mobile-safe.\nconst WGSL_POOL_MATMUL = `\\\nstruct Params {\n  Np: u32,  // pooled rows (output rows)\n  N: u32,   // patch count (contraction dim)\n  W: u32,   // hidden width (output cols)\n}\n\n@group(0) @binding(0) var<storage, read> poolW: array<f32>;   // [Np, N]\n@group(0) @binding(1) var<storage, read> hidden: array<f32>;  // [N, W]\n@group(0) @binding(2) var<storage, read_write> outp: array<f32>; // [Np, W]\n@group(0) @binding(3) var<storage, read> params: Params;\n\n@compute @workgroup_size(64)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  let total = params.Np * params.W;\n  if (idx >= total) { return; }\n  let r = idx / params.W;   // pooled row\n  let c = idx % params.W;   // hidden col\n  var acc: f32 = 0.0;\n  let prow = r * params.N;\n  for (var k: u32 = 0u; k < params.N; k = k + 1u) {\n    let w = poolW[prow + k];\n    if (w != 0.0) {\n      acc = acc + w * hidden[k * params.W + c];\n    }\n  }\n  outp[idx] = acc;\n}\n`;\n\n// ClippedMatMul: out = clamp(clamp(A, imin, imax) @ B^T, omin, omax).\n// Gemma 4 ViT Gemma4ClippableLinear: each projection clamps its input to a\n// calibrated [imin, imax] before the linear and its output to [omin, omax] after.\n// Same tiling as WGSL_MATMUL (4×2 register blocking, vec4 loads), with the input\n// clamp applied at tile load and the output clamp at store. Mobile-safe: clamp =\n// min(max(x, lo), hi), no select(); no `enable f16`. The four clip scalars are\n// passed as f32 bit patterns in the uniform.\nconst WGSL_CLIPPED_MATMUL = `\\\nconst MT: u32 = 64u;\nconst NT: u32 = 32u;\nconst KT: u32 = 16u;\nconst KT4: u32 = 4u;\n\nstruct Params {\n  M: u32,\n  K: u32,\n  N: u32,\n  imin: f32,\n  imax: f32,\n  omin: f32,\n  omax: f32,\n}\n\n@group(0) @binding(0) var<storage, read> A: array<vec4f>;\n@group(0) @binding(1) var<storage, read> B: array<vec4f>;\n@group(0) @binding(2) var<storage, read_write> C: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\nvar<workgroup> tileA: array<f32, 1024>;\nvar<workgroup> tileB: array<f32, 512>;\n\nfn clip4(v: vec4f, lo: f32, hi: f32) -> vec4f {\n  return min(max(v, vec4f(lo)), vec4f(hi));\n}\n\n@compute @workgroup_size(16, 16)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let lr = lid.y;\n  let lc = lid.x;\n  let mBase = wid.y * MT;\n  let nBase = wid.x * NT;\n  let row0 = mBase + lr;\n  let col0 = nBase + lc;\n\n  var acc: array<f32, 8>;\n  for (var i = 0u; i < 8u; i = i + 1u) { acc[i] = 0.0; }\n\n  let numTiles = (params.K + KT - 1u) / KT;\n  let tid = lr * 16u + lc;\n  let K4 = params.K / 4u;\n\n  for (var t: u32 = 0u; t < numTiles; t = t + 1u) {\n    let kv0 = t * KT4;\n    {\n      let ar = tid / KT4;\n      let av4 = tid % KT4;\n      let gRow = mBase + ar;\n      var v = vec4f(0.0, 0.0, 0.0, 0.0);\n      if (gRow < params.M && kv0 + av4 < K4) {\n        v = clip4(A[gRow * K4 + kv0 + av4], params.imin, params.imax);\n      }\n      let base = ar * KT + av4 * 4u;\n      tileA[base] = v.x;\n      tileA[base + 1u] = v.y;\n      tileA[base + 2u] = v.z;\n      tileA[base + 3u] = v.w;\n    }\n    if (tid < 128u) {\n      let bc = tid / KT4;\n      let bv4 = tid % KT4;\n      let gCol = nBase + bc;\n      var v = vec4f(0.0, 0.0, 0.0, 0.0);\n      if (gCol < params.N && kv0 + bv4 < K4) {\n        v = B[gCol * K4 + kv0 + bv4];\n      }\n      let krow = bv4 * 4u;\n      tileB[(krow + 0u) * NT + bc] = v.x;\n      tileB[(krow + 1u) * NT + bc] = v.y;\n      tileB[(krow + 2u) * NT + bc] = v.z;\n      tileB[(krow + 3u) * NT + bc] = v.w;\n    }\n\n    workgroupBarrier();\n\n    for (var k: u32 = 0u; k < KT; k = k + 1u) {\n      let b0 = tileB[k * NT + lc];\n      let b1 = tileB[k * NT + lc + 16u];\n      let a0 = tileA[lr * KT + k];\n      let a1 = tileA[(lr + 16u) * KT + k];\n      let a2 = tileA[(lr + 32u) * KT + k];\n      let a3 = tileA[(lr + 48u) * KT + k];\n      acc[0] += a0 * b0;\n      acc[1] += a0 * b1;\n      acc[2] += a1 * b0;\n      acc[3] += a1 * b1;\n      acc[4] += a2 * b0;\n      acc[5] += a2 * b1;\n      acc[6] += a3 * b0;\n      acc[7] += a3 * b1;\n    }\n\n    workgroupBarrier();\n  }\n\n  for (var rr = 0u; rr < 4u; rr = rr + 1u) {\n    let gRow = row0 + rr * 16u;\n    if (gRow < params.M) {\n      if (col0 < params.N) {\n        C[gRow * params.N + col0] = min(max(acc[rr * 2u], params.omin), params.omax);\n      }\n      if (col0 + 16u < params.N) {\n        C[gRow * params.N + col0 + 16u] = min(max(acc[rr * 2u + 1u], params.omin), params.omax);\n      }\n    }\n  }\n}\n`;\n\n// ── Kernel specs ────────────────────────────────────────────────────────\n\n/**\n * Resolve the total element count of the first output tensor.\n * Used by elementwise ops whose dispatch is based on output size.\n */\nfunction outputElementCount(op: OpNode, resolvedShapes: Record<string, number[]>): number {\n  const outShape = resolvedShapes[op.outputs[0]];\n  let count = 1;\n  for (const d of outShape) count *= d;\n  return count;\n}\n\n// ── Add ──\n\nconst addSpec: KernelSpec = {\n  shaderCode: WGSL_ADD,\n  entryPoint: \"main\",\n\n  // @binding(0) a: storage read\n  // @binding(1) b: storage read\n  // @binding(2) output: storage read_write\n  // @binding(3) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    // Params { count: u32 }\n    return buildUniformBuffer([count]);\n  },\n};\n\n// ── Mul ──\n\nconst mulSpec: KernelSpec = {\n  shaderCode: WGSL_MUL,\n  entryPoint: \"main\",\n\n  // @binding(0) a: storage read\n  // @binding(1) b: storage read\n  // @binding(2) output: storage read_write\n  // @binding(3) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    // Params { count: u32 }\n    return buildUniformBuffer([count]);\n  },\n};\n\n// ── SliceLastRow ──\n\nconst sliceLastRowSpec: KernelSpec = {\n  shaderCode: WGSL_SLICE_LAST_ROW,\n  entryPoint: \"main\",\n\n  // @binding(0) src: storage read\n  // @binding(1) output: storage read_write\n  // @binding(2) params: uniform\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n\n  getDispatchSize(op) {\n    const width = op.attributes.width as number;\n    return [cdiv(width, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const width = op.attributes.width as number;\n    const T = resolvedShapes[op.inputs[0]]?.[0] ?? 1;\n    // Params { width: u32, last_row_offset: u32 }\n    return buildUniformBuffer([width, (T - 1) * width]);\n  },\n};\n\n// ── MeanPool (mean over T tokens → [1, width], bidirectional embeddings) ──\n\nconst meanPoolSpec: KernelSpec = {\n  shaderCode: WGSL_MEAN_POOL,\n  entryPoint: \"main\",\n\n  // @binding(0) src: storage read\n  // @binding(1) output: storage read_write\n  // @binding(2) params: uniform\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n\n  getDispatchSize(op) {\n    const width = op.attributes.width as number;\n    return [cdiv(width, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const width = op.attributes.width as number;\n    // seq_len = rows of the input activation (T).\n    const T = resolvedShapes[op.inputs[0]]?.[0] ?? 1;\n    // Params { seq_len: u32, width: u32 }\n    return buildUniformBuffer([T, width]);\n  },\n};\n\n// ── Scale (multiply tensor by a scalar constant) ──\n\nconst scaleSpec: KernelSpec = {\n  shaderCode: WGSL_SCALE,\n  entryPoint: \"main\",\n\n  // @binding(0) input: storage read\n  // @binding(1) output: storage read_write\n  // @binding(2) params: uniform\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    const scale = op.attributes.scale as number;\n    // Params { count: u32, scale_bits: u32 }\n    return buildUniformBuffer([count, f32BitsToU32(scale)]);\n  },\n};\n\n// ── Softcap (Gemma final logit softcapping) ──\n\nconst softcapSpec: KernelSpec = {\n  shaderCode: WGSL_SOFTCAP,\n  entryPoint: \"main\",\n\n  // @binding(0) input: storage read\n  // @binding(1) output: storage read_write\n  // @binding(2) params: uniform\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    const cap = op.attributes.cap as number;\n    // Params { count: u32, cap_bits: u32 }\n    return buildUniformBuffer([count, f32BitsToU32(cap)]);\n  },\n};\n\n// ── L2Norm (row-wise L2 normalization, used for embeddings) ──\n\nconst l2NormSpec: KernelSpec = {\n  shaderCode: WGSL_L2NORM,\n  entryPoint: \"main\",\n\n  // @binding(0) src: storage read\n  // @binding(1) output: storage read_write\n  // @binding(2) params: uniform\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n\n  getDispatchSize(op, resolvedShapes) {\n    // One workgroup per row.\n    const rows = resolvedShapes[op.outputs[0]]?.[0] ?? 1;\n    return [rows, 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const width = op.attributes.width as number;\n    const rows = resolvedShapes[op.outputs[0]]?.[0] ?? 1;\n    // Params { rows: u32, width: u32 }\n    return buildUniformBuffer([rows, width]);\n  },\n};\n\n// ── SwiGLU (fused SiLU + Mul) ──\n\nconst swigluSpec: KernelSpec = {\n  shaderCode: WGSL_SWIGLU,\n  entryPoint: \"main\",\n\n  // @binding(0) gate: storage read\n  // @binding(1) up: storage read\n  // @binding(2) output: storage read_write\n  // @binding(3) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return buildUniformBuffer([count]);\n  },\n};\n\n// ── ResidualRMSNorm (fused Add + RMSNorm) ──\n\nconst residualRmsnormSpec: KernelSpec = {\n  shaderCode: WGSL_RESIDUAL_RMSNORM,\n  entryPoint: \"main\",\n\n  // @binding(0) a: storage read\n  // @binding(1) b: storage read\n  // @binding(2) weight: storage read\n  // @binding(3) sum_output: storage read_write\n  // @binding(4) norm_output: storage read_write\n  // @binding(5) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const hidden = op.attributes.hidden_size as number;\n    const seqLen = resolveSeqLen(op, resolvedShapes, hidden);\n    return [seqLen, 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const hidden = op.attributes.hidden_size as number;\n    const seqLen = resolveSeqLen(op, resolvedShapes, hidden);\n    const eps = (op.attributes.eps as number) ?? 1e-6;\n    return buildUniformBuffer([seqLen, hidden, f32BitsToU32(eps), 0]);\n  },\n};\n\n// ── SiLU ──\n\nconst siluSpec: KernelSpec = {\n  shaderCode: WGSL_SILU,\n  entryPoint: \"main\",\n\n  // @binding(0) input: storage read\n  // @binding(1) output: storage read_write\n  // @binding(2) params: uniform\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    // Params { count: u32 }\n    return buildUniformBuffer([count]);\n  },\n};\n\n// ── GELU ──\n\nconst geluSpec: KernelSpec = {\n  shaderCode: WGSL_GELU,\n  entryPoint: \"main\",\n\n  // @binding(0) input: storage read\n  // @binding(1) output: storage read_write\n  // @binding(2) params: uniform\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    // Params { count: u32 }\n    return buildUniformBuffer([count]);\n  },\n};\n\n// ── GELU (exact erf form — ViT merger) ──\n\nconst geluErfSpec: KernelSpec = {\n  shaderCode: WGSL_GELU_ERF,\n  entryPoint: \"main\",\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n  buildParams(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return buildUniformBuffer([count]);\n  },\n};\n\n// ── AddBias (row-broadcast bias add) ──\n\nconst addBiasSpec: KernelSpec = {\n  shaderCode: WGSL_ADD_BIAS,\n  entryPoint: \"main\",\n\n  // @binding(0) input, (1) bias, (2) output, (3) params\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n  buildParams(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    const width = op.attributes.width as number;\n    return buildUniformBuffer([count, width]);\n  },\n};\n\n// ── SliceCols (extract column range) ──\n\nconst sliceColsSpec: KernelSpec = {\n  shaderCode: WGSL_SLICE_COLS,\n  entryPoint: \"main\",\n\n  // @binding(0) input, (1) output, (2) params\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n  buildParams(op, resolvedShapes) {\n    const in_width = op.attributes.in_width as number;\n    const out_width = op.attributes.out_width as number;\n    const col_offset = op.attributes.col_offset as number;\n    const outShape = resolvedShapes[op.outputs[0]];\n    const rows = outShape[0];\n    return buildUniformBuffer([rows, in_width, out_width, col_offset]);\n  },\n};\n\n// ── MulCols (multiply two column ranges of one tensor) ──\n\nconst mulColsSpec: KernelSpec = {\n  shaderCode: WGSL_MUL_COLS,\n  entryPoint: \"main\",\n\n  // @binding(0) src, (1) output, (2) params\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n  buildParams(op, resolvedShapes) {\n    const in_width = op.attributes.in_width as number;\n    const out_width = op.attributes.out_width as number;\n    const off_a = op.attributes.off_a as number;\n    const off_b = op.attributes.off_b as number;\n    const outShape = resolvedShapes[op.outputs[0]];\n    const rows = outShape[0];\n    return buildUniformBuffer([rows, in_width, out_width, off_a, off_b, 0, 0, 0]);\n  },\n};\n\n// ── ApplyRotaryEmb (precomputed cos/sin, rotate_half) ──\n\nconst applyRotarySpec: KernelSpec = {\n  shaderCode: WGSL_APPLY_ROTARY,\n  entryPoint: \"main\",\n\n  // @binding(0) x, (1) cos, (2) sin, (3) output, (4) params\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n  buildParams(op, resolvedShapes) {\n    const num_heads = op.attributes.num_heads as number;\n    const head_dim = op.attributes.head_dim as number;\n    const outShape = resolvedShapes[op.outputs[0]];\n    const total = outShape.reduce((a, b) => a * b, 1);\n    const seq_len = total / (num_heads * head_dim);\n    return buildUniformBuffer([seq_len, num_heads, head_dim, 0]);\n  },\n};\n\n// ── MatMul ──\n\nconst matmulSpec: KernelSpec = {\n  shaderCode: WGSL_MATMUL,\n  entryPoint: \"main\",\n\n  // @binding(0) A: storage read\n  // @binding(1) B: storage read\n  // @binding(2) C: storage read_write\n  // @binding(3) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    // C[M,N] = A[M,K] * B[K,N]\n    // 16x16 workgroup, 4x2 register blocking -> each workgroup covers a\n    // 64-row x 32-col output tile. Dispatch X covers N, Y covers M.\n    const N = op.attributes.N as number;\n    const mTensor = op.attributes.M_tensor as string | undefined;\n    let M: number;\n    if (mTensor && resolvedShapes[mTensor]) {\n      M = resolvedShapes[mTensor][0];\n    } else {\n      M = op.attributes.M as number;\n    }\n    return [cdiv(N, 32), cdiv(M, 64), 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    // Params { M: u32, K: u32, N: u32 }\n    const K = op.attributes.K as number;\n    const N = op.attributes.N as number;\n    // M is dynamic — derive from referenced tensor or static attribute\n    const mTensor = op.attributes.M_tensor as string | undefined;\n    let M: number;\n    if (mTensor && resolvedShapes[mTensor]) {\n      M = resolvedShapes[mTensor][0];\n    } else {\n      M = op.attributes.M as number;\n    }\n    return buildUniformBuffer([M, K, N]);\n  },\n};\n\n// ── MatMulBias (fused MatMul + bias) ──\n\nconst matmulBiasSpec: KernelSpec = {\n  shaderCode: WGSL_MATMUL_BIAS,\n  entryPoint: \"main\",\n\n  // @binding(0) A, (1) B, (2) bias, (3) C, (4) params\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const N = op.attributes.N as number;\n    const mTensor = op.attributes.M_tensor as string | undefined;\n    let M: number;\n    if (mTensor && resolvedShapes[mTensor]) {\n      M = resolvedShapes[mTensor][0];\n    } else {\n      M = op.attributes.M as number;\n    }\n    return [cdiv(N, 32), cdiv(M, 64), 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const K = op.attributes.K as number;\n    const N = op.attributes.N as number;\n    const mTensor = op.attributes.M_tensor as string | undefined;\n    let M: number;\n    if (mTensor && resolvedShapes[mTensor]) {\n      M = resolvedShapes[mTensor][0];\n    } else {\n      M = op.attributes.M as number;\n    }\n    return buildUniformBuffer([M, K, N]);\n  },\n};\n\n// ── PoolMatMul (Gemma 4 ViT spatial average-pool as host matrix · encoder out) ──\n\nconst poolMatMulSpec: KernelSpec = {\n  shaderCode: WGSL_POOL_MATMUL,\n  entryPoint: \"main\",\n\n  // @binding(0) poolW [Np,N], (1) hidden [N,W], (2) out [Np,W], (3) params\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const out = resolvedShapes[op.outputs[0]]; // [Np, W]\n    const Np = out[0];\n    const W = out[1];\n    return [cdiv(Np * W, 64), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const out = resolvedShapes[op.outputs[0]]; // [Np, W]\n    const Np = out[0];\n    const W = out[1];\n    // N = contraction dim = rows of the hidden tensor (op.inputs[1]).\n    const N = resolvedShapes[op.inputs[1]]?.[0] ?? (op.attributes.hidden_size as number);\n    // Params { Np: u32, N: u32, W: u32 }\n    return buildUniformBuffer([Np, N, W]);\n  },\n};\n\n// ── ClippedMatMul (Gemma 4 ViT Gemma4ClippableLinear) ──\n\nconst clippedMatMulSpec: KernelSpec = {\n  shaderCode: WGSL_CLIPPED_MATMUL,\n  entryPoint: \"main\",\n\n  // @binding(0) A, (1) B, (2) C, (3) params\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const N = op.attributes.N as number;\n    const mTensor = op.attributes.M_tensor as string | undefined;\n    const M =\n      mTensor && resolvedShapes[mTensor] ? resolvedShapes[mTensor][0] : (op.attributes.M as number);\n    return [cdiv(N, 32), cdiv(M, 64), 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const K = op.attributes.K as number;\n    const N = op.attributes.N as number;\n    const mTensor = op.attributes.M_tensor as string | undefined;\n    const M =\n      mTensor && resolvedShapes[mTensor] ? resolvedShapes[mTensor][0] : (op.attributes.M as number);\n    // Clip scalars default to ±inf (identity) until the loader patches them.\n    const imin = (op.attributes.imin as number) ?? Number.NEGATIVE_INFINITY;\n    const imax = (op.attributes.imax as number) ?? Number.POSITIVE_INFINITY;\n    const omin = (op.attributes.omin as number) ?? Number.NEGATIVE_INFINITY;\n    const omax = (op.attributes.omax as number) ?? Number.POSITIVE_INFINITY;\n    // Params { M, K, N, imin, imax, omin, omax } — first 3 u32, rest f32 bits.\n    return buildUniformBuffer([\n      M,\n      K,\n      N,\n      f32BitsToU32(imin),\n      f32BitsToU32(imax),\n      f32BitsToU32(omin),\n      f32BitsToU32(omax),\n    ]);\n  },\n};\n\n// Mixed-precision (f16 multiply, f32 accumulate) MatMulBias for ViT linear layers.\n// Same bindings/dispatch/params as matmulBiasSpec; binding(1) B is an f16 weight.\n// The VisionExecutor swaps this in when the device has shader-f16.\nexport const MATMUL_BIAS_F16C_SPEC: KernelSpec = {\n  ...matmulBiasSpec,\n  shaderCode: WGSL_MATMUL_BIAS_F16MIX,\n};\n\n// ── MatMulInt4 ──\n\nconst matmulInt4Spec: KernelSpec = {\n  shaderCode: WGSL_MATMUL_INT4,\n  entryPoint: \"main\",\n\n  // @binding(0) A: storage read\n  // @binding(1) B_q: storage read\n  // @binding(2) scales: storage read\n  // @binding(3) zeros: storage read\n  // @binding(4) C: storage read_write\n  // @binding(5) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    // Workgroup size is (16, 16), dispatch X covers N, Y covers M\n    const N = op.attributes.N as number;\n    const mTensor = op.attributes.M_tensor as string | undefined;\n    let M: number;\n    if (mTensor && resolvedShapes[mTensor]) {\n      M = resolvedShapes[mTensor][0];\n    } else {\n      M = op.attributes.M as number;\n    }\n    return [cdiv(N, 16), cdiv(M, 16), 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    // Params { M: u32, K: u32, N: u32, group_size: u32 }\n    const K = op.attributes.K as number;\n    const N = op.attributes.N as number;\n    const group_size = op.attributes.group_size as number;\n    const mTensor = op.attributes.M_tensor as string | undefined;\n    let M: number;\n    if (mTensor && resolvedShapes[mTensor]) {\n      M = resolvedShapes[mTensor][0];\n    } else {\n      M = op.attributes.M as number;\n    }\n    return buildUniformBuffer([M, K, N, group_size]);\n  },\n};\n\n// ── MatVec (F32, M=1 decode) ──\n\nexport const MATVEC_SPEC: KernelSpec = {\n  shaderCode: WGSL_MATVEC,\n  entryPoint: \"main\",\n\n  // Same binding layout as MatMul: A, B, C, params\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op) {\n    const N = op.attributes.N as number;\n    const N_TILE = 8;\n    return [cdiv(N, N_TILE), 1, 1];\n  },\n\n  buildParams(op) {\n    const K = op.attributes.K as number;\n    const N = op.attributes.N as number;\n    return buildUniformBuffer([K, N]);\n  },\n};\n\n// ── MatVecInt4 (K-parallel INT4, M=1 decode) ──\n\nexport const MATVEC_INT4_SPEC: KernelSpec = {\n  shaderCode: WGSL_MATVEC_INT4,\n  entryPoint: \"main\",\n\n  // Same binding layout as MatMulInt4: A, B_q, scales, zeros, C, params\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op) {\n    const N = op.attributes.N as number;\n    // Must match WGSL N_TILE = workgroup_size / K_THREADS = 256 / 16 = 16\n    const N_TILE = 16;\n    return [cdiv(N, N_TILE), 1, 1];\n  },\n\n  buildParams(op) {\n    const K = op.attributes.K as number;\n    const N = op.attributes.N as number;\n    const group_size = op.attributes.group_size as number;\n    return buildUniformBuffer([K, N, group_size, 0]);\n  },\n};\n\n// ── Gated MatVecInt4 (fused attn sigmoid-gate + INT4 o_proj, M=1 decode) ──\n\nexport const GATED_MATVEC_INT4_SPEC: KernelSpec = {\n  shaderCode: WGSL_GATED_MATVEC_INT4,\n  entryPoint: \"main\",\n\n  // @binding(0) attn, (1) gate, (2) B_q, (3) scales, (4) zeros, (5) C, (6) params\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op) {\n    const N = op.attributes.N as number;\n    const N_TILE = 8;\n    return [cdiv(N, N_TILE), 1, 1];\n  },\n\n  buildParams(op) {\n    const K = op.attributes.K as number;\n    const N = op.attributes.N as number;\n    const group_size = op.attributes.group_size as number;\n    return buildUniformBuffer([K, N, group_size, 0]);\n  },\n};\n\n// ── SwiGLU-gated MatVecInt4 (fused SwiGLU + INT4 projection, M=1 decode) ──\n\nexport const SWIGLU_GATED_MATVEC_INT4_SPEC: KernelSpec = {\n  ...GATED_MATVEC_INT4_SPEC,\n  shaderCode: WGSL_SWIGLU_GATED_MATVEC_INT4,\n};\n\n// ── MatVec subgroups variants (desktop-only fast path) ──\n//\n// Same binding layout, dispatch size, and params as the portable MatVec specs —\n// only the shader differs (subgroupAdd reduction). The executor selects these\n// when the device exposes the \"subgroups\" feature and is NOT WebKit. Kept as\n// distinct specs so the portable kernels remain the universal fallback.\n\nexport const MATVEC_SUBGROUPS_SPEC: KernelSpec = {\n  ...MATVEC_SPEC,\n  shaderCode: WGSL_MATVEC_SUBGROUPS,\n};\n\nexport const MATVEC_INT4_SUBGROUPS_SPEC: KernelSpec = {\n  ...MATVEC_INT4_SPEC,\n  shaderCode: WGSL_MATVEC_INT4_SUBGROUPS,\n};\n\n// ── SwiGLU MatVec (fused gate+up+SwiGLU, M=1 decode) ──\n\nexport const SWIGLU_MATVEC_SPEC: KernelSpec = {\n  shaderCode: WGSL_SWIGLU_MATVEC,\n  entryPoint: \"main\",\n\n  // @binding(0) A: storage read\n  // @binding(1) B_gate: storage read\n  // @binding(2) B_up: storage read\n  // @binding(3) output: storage read_write\n  // @binding(4) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op) {\n    const N = op.attributes.N as number;\n    const N_TILE = 8;\n    return [cdiv(N, N_TILE), 1, 1];\n  },\n\n  buildParams(op) {\n    const K = op.attributes.K as number;\n    const N = op.attributes.N as number;\n    return buildUniformBuffer([K, N]);\n  },\n};\n\n// ── SwiGLU MatVecInt4 (fused gate+up+SwiGLU INT4, M=1 decode) ──\n\nexport const SWIGLU_MATVEC_INT4_SPEC: KernelSpec = {\n  shaderCode: WGSL_SWIGLU_MATVEC_INT4,\n  entryPoint: \"main\",\n\n  // @binding(0) A: storage read\n  // @binding(1) B_gate_q: storage read\n  // @binding(2) B_gate_scales: storage read\n  // @binding(3) B_gate_zeros: storage read\n  // @binding(4) B_up_q: storage read\n  // @binding(5) B_up_scales: storage read\n  // @binding(6) B_up_zeros: storage read\n  // @binding(7) output: storage read_write\n  // @binding(8) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op) {\n    const N = op.attributes.N as number;\n    const N_TILE = 8;\n    return [cdiv(N, N_TILE), 1, 1];\n  },\n\n  buildParams(op) {\n    const K = op.attributes.K as number;\n    const N = op.attributes.N as number;\n    const group_size = op.attributes.group_size as number;\n    return buildUniformBuffer([K, N, group_size, 0]);\n  },\n};\n\n// ── Dual MatVecInt4 (two INT4 projections sharing input A, M=1 decode) ──\n\nexport const DUAL_MATVEC_INT4_SPEC: KernelSpec = {\n  shaderCode: WGSL_DUAL_MATVEC_INT4,\n  entryPoint: \"main\",\n\n  // @binding(0) A\n  // @binding(1..3) B0_q, B0_scales, B0_zeros\n  // @binding(4..6) B1_q, B1_scales, B1_zeros\n  // @binding(7) out0, @binding(8) out1, @binding(9) params\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op) {\n    const N = op.attributes.N as number;\n    const N_TILE = 8;\n    return [cdiv(N, N_TILE), 1, 1];\n  },\n\n  buildParams(op) {\n    const K = op.attributes.K as number;\n    const N = op.attributes.N as number;\n    const group_size = op.attributes.group_size as number;\n    return buildUniformBuffer([K, N, group_size, 0]);\n  },\n};\n\n// ── Argmax (GPU-side, single workgroup) ──\n\nexport const ARGMAX_SPEC: KernelSpec = {\n  shaderCode: WGSL_ARGMAX,\n  entryPoint: \"main\",\n\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n\n  getDispatchSize() {\n    return [1, 1, 1]; // Single workgroup\n  },\n\n  buildParams(_op, _resolvedShapes) {\n    // Params built externally by executor (count + offset)\n    return buildUniformBuffer([0, 0]);\n  },\n};\n\n// ── RMSNorm ──\n\nconst rmsnormSpec: KernelSpec = {\n  shaderCode: WGSL_RMSNORM,\n  entryPoint: \"main\",\n\n  // @binding(0) input: storage read\n  // @binding(1) weight: storage read\n  // @binding(2) output: storage read_write\n  // @binding(3) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    // One workgroup per row (token)\n    const hidden_size = op.attributes.hidden_size as number;\n    const ref = op.attributes.seq_len_tensor as string | undefined;\n    let seq_len: number;\n    if (ref && resolvedShapes[ref]) {\n      const shape = resolvedShapes[ref];\n      const total = shape.reduce((a: number, b: number) => a * b, 1);\n      seq_len = total / hidden_size;\n    } else {\n      seq_len = op.attributes.seq_len as number;\n    }\n    return [seq_len, 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    // Params { seq_len: u32, hidden_size: u32, eps_bits: u32, _pad: u32 }\n    const hidden_size = op.attributes.hidden_size as number;\n    const eps = op.attributes.eps as number;\n    const ref = op.attributes.seq_len_tensor as string | undefined;\n    let seq_len: number;\n    if (ref && resolvedShapes[ref]) {\n      const shape = resolvedShapes[ref];\n      const total = shape.reduce((a: number, b: number) => a * b, 1);\n      seq_len = total / hidden_size;\n    } else {\n      seq_len = op.attributes.seq_len as number;\n    }\n    return buildUniformBuffer([\n      seq_len,\n      hidden_size,\n      f32BitsToU32(eps),\n      0, // _pad\n    ]);\n  },\n};\n\n// ── Dual RMSNorm (two per-row RMSNorms, one dispatch) ──\n\nfunction dualRmsnormRows(\n  op: OpNode,\n  resolvedShapes: Record<string, number[]>,\n  refKey: string,\n  fallbackKey: string,\n): number {\n  const hidden_size = op.attributes.hidden_size as number;\n  const ref = op.attributes[refKey] as string | undefined;\n  if (ref && resolvedShapes[ref]) {\n    const total = resolvedShapes[ref].reduce((a: number, b: number) => a * b, 1);\n    return total / hidden_size;\n  }\n  return op.attributes[fallbackKey] as number;\n}\n\nexport const DUAL_RMSNORM_SPEC: KernelSpec = {\n  shaderCode: WGSL_DUAL_RMSNORM,\n  entryPoint: \"main\",\n\n  // input0, weight0, input1, weight1, output0, output1, params\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const rows0 = dualRmsnormRows(op, resolvedShapes, \"seq_len_tensor0\", \"seq_len0\");\n    const rows1 = dualRmsnormRows(op, resolvedShapes, \"seq_len_tensor1\", \"seq_len1\");\n    return [rows0 + rows1, 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const rows0 = dualRmsnormRows(op, resolvedShapes, \"seq_len_tensor0\", \"seq_len0\");\n    const rows1 = dualRmsnormRows(op, resolvedShapes, \"seq_len_tensor1\", \"seq_len1\");\n    const hidden_size = op.attributes.hidden_size as number;\n    const eps = op.attributes.eps as number;\n    return buildUniformBuffer([rows0, rows1, hidden_size, f32BitsToU32(eps)]);\n  },\n};\n\n// ── LayerNorm ──\n\nconst layernormSpec: KernelSpec = {\n  shaderCode: WGSL_LAYERNORM,\n  entryPoint: \"main\",\n\n  // @binding(0) input: storage read\n  // @binding(1) weight: storage read\n  // @binding(2) bias: storage read\n  // @binding(3) output: storage read_write\n  // @binding(4) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    // One workgroup per row (token)\n    const hidden_size = op.attributes.hidden_size as number;\n    const ref = op.attributes.seq_len_tensor as string | undefined;\n    let seq_len: number;\n    if (ref && resolvedShapes[ref]) {\n      const shape = resolvedShapes[ref];\n      const total = shape.reduce((a: number, b: number) => a * b, 1);\n      seq_len = total / hidden_size;\n    } else {\n      seq_len = op.attributes.seq_len as number;\n    }\n    return [seq_len, 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    // Params { seq_len: u32, hidden_size: u32, eps_bits: u32, _pad: u32 }\n    const hidden_size = op.attributes.hidden_size as number;\n    const eps = op.attributes.eps as number;\n    const ref = op.attributes.seq_len_tensor as string | undefined;\n    let seq_len: number;\n    if (ref && resolvedShapes[ref]) {\n      const shape = resolvedShapes[ref];\n      const total = shape.reduce((a: number, b: number) => a * b, 1);\n      seq_len = total / hidden_size;\n    } else {\n      seq_len = op.attributes.seq_len as number;\n    }\n    return buildUniformBuffer([\n      seq_len,\n      hidden_size,\n      f32BitsToU32(eps),\n      0, // _pad\n    ]);\n  },\n};\n\n// ── Embedding ──\n\nconst embeddingSpec: KernelSpec = {\n  shaderCode: WGSL_EMBEDDING,\n  entryPoint: \"main\",\n\n  // @binding(0) input_ids: storage read\n  // @binding(1) weight: storage read\n  // @binding(2) output: storage read_write\n  // @binding(3) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const hidden_size = op.attributes.hidden_size as number;\n    const outShape = resolvedShapes[op.outputs[0]];\n    const seq_len = outShape ? outShape[0] : (op.attributes.seq_len as number);\n    return [cdiv(seq_len * hidden_size, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    // Params { seq_len: u32, hidden_size: u32 }\n    const hidden_size = op.attributes.hidden_size as number;\n    const outShape = resolvedShapes[op.outputs[0]];\n    const seq_len = outShape ? outShape[0] : (op.attributes.seq_len as number);\n    return buildUniformBuffer([seq_len, hidden_size]);\n  },\n};\n\n// ── EmbeddingInt4 ──\n\nconst embeddingInt4Spec: KernelSpec = {\n  shaderCode: WGSL_EMBEDDING_INT4,\n  entryPoint: \"main\",\n\n  // @binding(0) input_ids: storage read\n  // @binding(1) weight_q: storage read\n  // @binding(2) scales: storage read\n  // @binding(3) zeros: storage read\n  // @binding(4) output: storage read_write\n  // @binding(5) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const hidden_size = op.attributes.hidden_size as number;\n    const outShape = resolvedShapes[op.outputs[0]];\n    const seq_len = outShape ? outShape[0] : (op.attributes.seq_len as number);\n    return [cdiv(seq_len * hidden_size, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    // Params { seq_len: u32, hidden_size: u32, group_size: u32, _pad: u32 }\n    const hidden_size = op.attributes.hidden_size as number;\n    const group_size = op.attributes.group_size as number;\n    const outShape = resolvedShapes[op.outputs[0]];\n    const seq_len = outShape ? outShape[0] : (op.attributes.seq_len as number);\n    return buildUniformBuffer([seq_len, hidden_size, group_size, 0]);\n  },\n};\n\n// ── RoPE ──\n\nconst ropeSpec: KernelSpec = {\n  shaderCode: WGSL_ROPE,\n  entryPoint: \"main\",\n\n  // @binding(0) q: storage read_write\n  // @binding(1) k: storage read_write\n  // @binding(2) params: uniform\n  bindings: [{ type: \"storage-read-write\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n\n  getDispatchSize(op, resolvedShapes) {\n    const num_q_heads = op.attributes.num_q_heads as number;\n    const num_kv_heads = op.attributes.num_kv_heads as number;\n    const head_dim = op.attributes.head_dim as number;\n    const rope_dim = (op.attributes.rope_dim as number) ?? head_dim;\n    // rope_half is the rotate_half pairing offset / pair count. Defaults to\n    // rope_dim/2 (legacy), but Gemma \"proportional\" overrides it to head_dim/2.\n    const rope_half = (op.attributes.rope_half as number) ?? rope_dim / 2;\n\n    const seqRef = op.attributes.seq_len_tensor as string | undefined;\n    let seq_len: number;\n    if (seqRef && resolvedShapes[seqRef]) {\n      seq_len = resolvedShapes[seqRef][0];\n    } else {\n      seq_len = op.attributes.seq_len as number;\n    }\n\n    const total_q_pairs = seq_len * num_q_heads * rope_half;\n    const total_k_pairs = seq_len * num_kv_heads * rope_half;\n    return [cdiv(Math.max(total_q_pairs, total_k_pairs), 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes, context) {\n    // Params { seq_len, num_q_heads, num_kv_heads, head_dim, rope_base_bits,\n    //          position_offset, rope_dim, rope_half, rope_denom, rope_active_pairs, _pad0, _pad1 }\n    const num_q_heads = op.attributes.num_q_heads as number;\n    const num_kv_heads = op.attributes.num_kv_heads as number;\n    const head_dim = op.attributes.head_dim as number;\n    const rope_base = op.attributes.rope_base as number;\n    const rope_dim = (op.attributes.rope_dim as number) ?? head_dim;\n    // Decoupled RoPE knobs (default to legacy single-rope_dim behavior):\n    //   rope_half          = rotate_half pairing offset + pair count (legacy rope_dim/2)\n    //   rope_denom         = inv_freq exponent denominator (legacy rope_dim)\n    //   rope_active_pairs  = pairs with a real frequency; rest are identity (legacy rope_dim/2)\n    const rope_half = (op.attributes.rope_half as number) ?? rope_dim / 2;\n    const rope_denom = (op.attributes.rope_denom as number) ?? rope_dim;\n    const rope_active_pairs = (op.attributes.rope_active_pairs as number) ?? rope_dim / 2;\n    // Use runtime seqPos for position offset during autoregressive decode\n    const position_offset = context?.seqPos ?? (op.attributes.position_offset as number) ?? 0;\n\n    const seqRef = op.attributes.seq_len_tensor as string | undefined;\n    let seq_len: number;\n    if (seqRef && resolvedShapes[seqRef]) {\n      seq_len = resolvedShapes[seqRef][0];\n    } else {\n      seq_len = op.attributes.seq_len as number;\n    }\n\n    return buildUniformBuffer([\n      seq_len,\n      num_q_heads,\n      num_kv_heads,\n      head_dim,\n      f32BitsToU32(rope_base),\n      position_offset,\n      rope_dim,\n      rope_half,\n      rope_denom,\n      rope_active_pairs,\n      0, // _pad0\n      0, // _pad1\n    ]);\n  },\n};\n\n/**\n * Interleaved (adjacent-pair) RoPE, used by Moonshine. Identical dispatch/params\n * to ropeSpec — only the rotation pairing differs (2p/2p+1 vs p/p+half). Selected\n * by the executors when a RoPE node carries `interleaved: true`.\n */\nexport const ROPE_INTERLEAVED_SPEC: KernelSpec = {\n  ...ropeSpec,\n  shaderCode: WGSL_ROPE_INTERLEAVED,\n};\n\n// ── M-RoPE (precomputed cos/sin, partial rotate_half) ──\n\nconst mropeSpec: KernelSpec = {\n  shaderCode: WGSL_MROPE,\n  entryPoint: \"main\",\n  // @binding(0) q rw, (1) k rw, (2) cos, (3) sin, (4) params\n  bindings: [\n    { type: \"storage-read-write\" },\n    { type: \"storage-read-write\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"uniform\" },\n  ],\n  getDispatchSize(op, resolvedShapes) {\n    const num_q_heads = op.attributes.num_q_heads as number;\n    const num_kv_heads = op.attributes.num_kv_heads as number;\n    const rope_dim = op.attributes.rope_dim as number;\n    const half = rope_dim / 2;\n    const seqRef = op.attributes.seq_len_tensor as string | undefined;\n    const seq_len =\n      seqRef && resolvedShapes[seqRef]\n        ? resolvedShapes[seqRef][0]\n        : (op.attributes.seq_len as number);\n    const total = seq_len * Math.max(num_q_heads, num_kv_heads) * half;\n    return [cdiv(total, 256), 1, 1];\n  },\n  buildParams(op, resolvedShapes, context) {\n    const num_q_heads = op.attributes.num_q_heads as number;\n    const num_kv_heads = op.attributes.num_kv_heads as number;\n    const head_dim = op.attributes.head_dim as number;\n    const rope_dim = op.attributes.rope_dim as number;\n    const seqRef = op.attributes.seq_len_tensor as string | undefined;\n    const seq_len =\n      seqRef && resolvedShapes[seqRef]\n        ? resolvedShapes[seqRef][0]\n        : (op.attributes.seq_len as number);\n    // Decode (single-token) reads cos/sin row seqPos; prefill writes all rows 0..T.\n    const pos_offset = seq_len === 1 ? (context?.seqPos ?? 0) : 0;\n    return buildUniformBuffer([\n      seq_len,\n      num_q_heads,\n      num_kv_heads,\n      head_dim,\n      rope_dim,\n      pos_offset,\n      0,\n      0,\n    ]);\n  },\n};\n\n// ── EmbedSplice (scatter vision tokens into image-token rows) ──\n\nconst embedSpliceSpec: KernelSpec = {\n  shaderCode: WGSL_EMBED_SPLICE,\n  entryPoint: \"main\",\n  // @binding(0) embed_in, (1) vision, (2) row_map, (3) embed_out, (4) params\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n  getDispatchSize(op, resolvedShapes) {\n    const hidden = op.attributes.hidden as number;\n    const seqRef = op.attributes.seq_len_tensor as string;\n    const seq_len = resolvedShapes[seqRef][0];\n    return [cdiv(seq_len * hidden, 256), 1, 1];\n  },\n  buildParams(op, resolvedShapes) {\n    const hidden = op.attributes.hidden as number;\n    const seqRef = op.attributes.seq_len_tensor as string;\n    const seq_len = resolvedShapes[seqRef][0];\n    return buildUniformBuffer([seq_len, hidden]);\n  },\n};\n\n// ── Attention ──\n\nconst attentionSpec: KernelSpec = {\n  shaderCode: WGSL_ATTENTION,\n  entryPoint: \"main\",\n\n  // @binding(0) Q: storage read\n  // @binding(1) K: storage read\n  // @binding(2) V: storage read\n  // @binding(3) output: storage read_write\n  // @binding(4) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    // One workgroup per (query_position, q_head) pair\n    const num_q_heads = op.attributes.num_q_heads as number;\n    // Resolve T from output shape or static attribute\n    const outShape = resolvedShapes[op.outputs[0]];\n    const T = outShape ? outShape[0] : (op.attributes.T as number);\n    return [T, num_q_heads, 1];\n  },\n\n  buildParams(op, resolvedShapes, context) {\n    // Params { T, S, num_q_heads, num_kv_heads, head_dim, position_offset, is_causal }\n    const num_q_heads = op.attributes.num_q_heads as number;\n    const num_kv_heads = op.attributes.num_kv_heads as number;\n    const head_dim = op.attributes.head_dim as number;\n    // Use runtime seqPos for position offset during autoregressive decode\n    const position_offset = context?.seqPos ?? (op.attributes.position_offset as number) ?? 0;\n\n    // Resolve T from output shape or static attribute\n    const outShape = resolvedShapes[op.outputs[0]];\n    const T = outShape ? outShape[0] : (op.attributes.T as number);\n\n    // S = total KV length (seqPos + T from KV cache, or T for prefill-only)\n    const kvShape = resolvedShapes[op.inputs[1]]; // K cache shape: [L_max, kv_dim]\n    const S = kvShape ? kvShape[0] : T;\n\n    // Default causal (text). ViT attention sets causal:false → bidirectional.\n    const is_causal = (op.attributes.causal as boolean) === false ? 0 : 1;\n\n    // Query-grid base offset (default 0). The WebKit vision encoder sets this\n    // per-chunk to split the O(N²) ViT attention into watchdog-safe windows; all\n    // other callers leave it 0, keeping their dispatch byte-identical.\n    const q_offset = context?.qOffset ?? 0;\n\n    // Softmax QK temperature. Defaults to 1/sqrt(head_dim) — byte-identical to\n    // the legacy kernel-hardcoded scale, so every existing caller is unchanged.\n    // Gemma 4 sets attn_scale:1.0 (its per-head QK-norm absorbs the scale).\n    const attn_scale = (op.attributes.attn_scale as number) ?? 1 / Math.sqrt(head_dim);\n\n    return buildUniformBuffer([\n      T,\n      S,\n      num_q_heads,\n      num_kv_heads,\n      head_dim,\n      position_offset,\n      is_causal,\n      q_offset,\n      f32BitsToU32(attn_scale),\n    ]);\n  },\n};\n\n// ── Cross-Attention (encoder-decoder) ──\n//\n// Inputs: Q (decoder hidden) [T, num_q_heads*head_dim],\n//         K/V (frozen encoder output) [S, num_kv_heads*head_dim].\n// Output: [T, num_q_heads*head_dim]. One workgroup per (q_pos, q_head).\nconst crossAttentionSpec: KernelSpec = {\n  shaderCode: WGSL_CROSS_ATTENTION,\n  entryPoint: \"main\",\n\n  // @binding(0) Q: storage read\n  // @binding(1) K: storage read  (encoder output, frozen)\n  // @binding(2) V: storage read  (encoder output, frozen)\n  // @binding(3) output: storage read_write\n  // @binding(4) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const num_q_heads = op.attributes.num_q_heads as number;\n    // T = decoder query length (rows of the output / Q tensor)\n    const outShape = resolvedShapes[op.outputs[0]];\n    const T = outShape ? outShape[0] : (op.attributes.T as number);\n    return [T, num_q_heads, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    // Params { T, S, num_q_heads, num_kv_heads, head_dim }\n    const num_q_heads = op.attributes.num_q_heads as number;\n    const num_kv_heads = op.attributes.num_kv_heads as number;\n    const head_dim = op.attributes.head_dim as number;\n\n    // T = decoder query length (rows of Q / output)\n    const outShape = resolvedShapes[op.outputs[0]];\n    const T = outShape ? outShape[0] : (op.attributes.T as number);\n\n    // S = encoder length (rows of the frozen encoder K input)\n    const kShape = resolvedShapes[op.inputs[1]];\n    const S = kShape ? kShape[0] : (op.attributes.S as number);\n\n    return buildUniformBuffer([T, S, num_q_heads, num_kv_heads, head_dim]);\n  },\n};\n\n// ── KVCacheAppend Packed F16 (f32 src → packed u32 dst, Safari-safe) ──\n\nexport const KV_CACHE_APPEND_PACKED_F16_SPEC: KernelSpec = {\n  shaderCode: WGSL_KV_CACHE_APPEND_PACKED_F16,\n  entryPoint: \"main\",\n\n  // @binding(0) src: storage read (f32)\n  // @binding(1) dst: storage read_write (packed u32)\n  // @binding(2) params: uniform\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n\n  getDispatchSize(op, resolvedShapes) {\n    const width = op.attributes.width as number;\n    const srcRef = op.attributes.T_tensor as string | undefined;\n    let T: number;\n    if (srcRef && resolvedShapes[srcRef]) {\n      T = resolvedShapes[srcRef][0];\n    } else {\n      T = op.attributes.T as number;\n    }\n    // Each thread processes a pair of f32 values → half the dispatch count\n    return [cdiv((T * width) / 2, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes, context) {\n    const width = op.attributes.width as number;\n    const seqPos = context?.seqPos ?? 0;\n    const srcRef = op.attributes.T_tensor as string | undefined;\n    let T: number;\n    if (srcRef && resolvedShapes[srcRef]) {\n      T = resolvedShapes[srcRef][0];\n    } else {\n      T = op.attributes.T as number;\n    }\n    // Same params as native f16 — kernel internally divides by 2\n    return buildUniformBuffer([T * width, seqPos * width]);\n  },\n};\n\n// ── Attention Packed F16 (packed u32 KV cache, Safari-safe) ──\n\nexport const ATTENTION_PACKED_F16_SPEC: KernelSpec = {\n  shaderCode: WGSL_ATTENTION_PACKED_F16,\n  entryPoint: \"main\",\n\n  // @binding(0) Q: storage read (f32)\n  // @binding(1) K: storage read (packed u32)\n  // @binding(2) V: storage read (packed u32)\n  // @binding(3) output: storage read_write (f32)\n  // @binding(4) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const num_q_heads = op.attributes.num_q_heads as number;\n    const outShape = resolvedShapes[op.outputs[0]];\n    const T = outShape ? outShape[0] : (op.attributes.T as number);\n    return [T, num_q_heads, 1];\n  },\n\n  buildParams(op, resolvedShapes, context) {\n    const num_q_heads = op.attributes.num_q_heads as number;\n    const num_kv_heads = op.attributes.num_kv_heads as number;\n    const head_dim = op.attributes.head_dim as number;\n    const position_offset = context?.seqPos ?? (op.attributes.position_offset as number) ?? 0;\n\n    const outShape = resolvedShapes[op.outputs[0]];\n    const T = outShape ? outShape[0] : (op.attributes.T as number);\n\n    const kvShape = resolvedShapes[op.inputs[1]];\n    const S = kvShape ? kvShape[0] : T;\n\n    // Default 1/sqrt(head_dim) → byte-identical to the legacy scale; Gemma4 = 1.0.\n    const attn_scale = (op.attributes.attn_scale as number) ?? 1 / Math.sqrt(head_dim);\n\n    return buildUniformBuffer([\n      T,\n      S,\n      num_q_heads,\n      num_kv_heads,\n      head_dim,\n      position_offset,\n      f32BitsToU32(attn_scale),\n    ]);\n  },\n};\n\n// ── Softmax ──\n\nconst softmaxSpec: KernelSpec = {\n  shaderCode: WGSL_SOFTMAX,\n  entryPoint: \"main\",\n\n  // @binding(0) input: storage read\n  // @binding(1) output: storage read_write\n  // @binding(2) params: uniform\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n\n  getDispatchSize(op) {\n    // One workgroup per row\n    const num_rows = op.attributes.num_rows as number;\n    return [num_rows, 1, 1];\n  },\n\n  buildParams(op) {\n    // Params { num_rows: u32, row_size: u32 }\n    const num_rows = op.attributes.num_rows as number;\n    const row_size = op.attributes.row_size as number;\n    return buildUniformBuffer([num_rows, row_size]);\n  },\n};\n\n// ── CausalConv1d ──\n\nconst causalConv1dSpec: KernelSpec = {\n  shaderCode: WGSL_CAUSAL_CONV1D,\n  entryPoint: \"main\",\n\n  // @binding(0) input: storage read\n  // @binding(1) weight: storage read\n  // @binding(2) conv_state: storage read\n  // @binding(3) output: storage read_write\n  // @binding(4) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const channels = op.attributes.channels as number;\n    const ref = op.attributes.seq_len_tensor as string | undefined;\n    let seq_len: number;\n    if (ref && resolvedShapes[ref]) {\n      seq_len = resolvedShapes[ref][0];\n    } else {\n      seq_len = op.attributes.seq_len as number;\n    }\n    return [cdiv(seq_len * channels, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    // Params { seq_len: u32, channels: u32, kernel_size: u32, _pad: u32 }\n    const channels = op.attributes.channels as number;\n    const kernel_size = op.attributes.kernel_size as number;\n    const ref = op.attributes.seq_len_tensor as string | undefined;\n    let seq_len: number;\n    if (ref && resolvedShapes[ref]) {\n      seq_len = resolvedShapes[ref][0];\n    } else {\n      seq_len = op.attributes.seq_len as number;\n    }\n    return buildUniformBuffer([seq_len, channels, kernel_size, 0]);\n  },\n};\n\n// ── CausalConv1dSiLU (conv + fused SiLU) ──\n\nconst causalConv1dSiluSpec: KernelSpec = {\n  shaderCode: WGSL_CAUSAL_CONV1D_SILU,\n  entryPoint: \"main\",\n\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const channels = op.attributes.channels as number;\n    const ref = op.attributes.seq_len_tensor as string | undefined;\n    let seq_len: number;\n    if (ref && resolvedShapes[ref]) {\n      seq_len = resolvedShapes[ref][0];\n    } else {\n      seq_len = op.attributes.seq_len as number;\n    }\n    return [cdiv(seq_len * channels, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const channels = op.attributes.channels as number;\n    const kernel_size = op.attributes.kernel_size as number;\n    const ref = op.attributes.seq_len_tensor as string | undefined;\n    let seq_len: number;\n    if (ref && resolvedShapes[ref]) {\n      seq_len = resolvedShapes[ref][0];\n    } else {\n      seq_len = op.attributes.seq_len as number;\n    }\n    return buildUniformBuffer([seq_len, channels, kernel_size, 0]);\n  },\n};\n\n// ── CausalConv1dGated (conv + fused multiplicative output gate) ──\n\nconst causalConv1dGatedSpec: KernelSpec = {\n  shaderCode: WGSL_CAUSAL_CONV1D_GATED,\n  entryPoint: \"main\",\n\n  // @binding(0) input  @binding(1) weight  @binding(2) conv_state\n  // @binding(3) gate   @binding(4) output  @binding(5) params\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const channels = op.attributes.channels as number;\n    const ref = op.attributes.seq_len_tensor as string | undefined;\n    let seq_len: number;\n    if (ref && resolvedShapes[ref]) {\n      seq_len = resolvedShapes[ref][0];\n    } else {\n      seq_len = op.attributes.seq_len as number;\n    }\n    return [cdiv(seq_len * channels, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const channels = op.attributes.channels as number;\n    const kernel_size = op.attributes.kernel_size as number;\n    const ref = op.attributes.seq_len_tensor as string | undefined;\n    let seq_len: number;\n    if (ref && resolvedShapes[ref]) {\n      seq_len = resolvedShapes[ref][0];\n    } else {\n      seq_len = op.attributes.seq_len as number;\n    }\n    return buildUniformBuffer([seq_len, channels, kernel_size, 0]);\n  },\n};\n\n// ── SigmoidGate ──\n\nconst sigmoidGateSpec: KernelSpec = {\n  shaderCode: WGSL_SIGMOID_GATE,\n  entryPoint: \"main\",\n\n  // @binding(0) x: storage read\n  // @binding(1) gate: storage read\n  // @binding(2) output: storage read_write\n  // @binding(3) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    // Params { count: u32 }\n    return buildUniformBuffer([count]);\n  },\n};\n\n// ── MambaSSM ──\n\nconst mambaSSMSpec: KernelSpec = {\n  shaderCode: WGSL_MAMBA_SSM,\n  entryPoint: \"main\",\n\n  // @binding(0) qkv_conv: storage read\n  // @binding(1) a_input: storage read\n  // @binding(2) b_input: storage read\n  // @binding(3) A_log: storage read\n  // @binding(4) dt_bias: storage read\n  // @binding(5) ssm_state: storage read_write\n  // @binding(6) output: storage read_write\n  // @binding(7) params: uniform\n  bindings: [\n    { type: \"storage-read\" }, // qkv_conv\n    { type: \"storage-read\" }, // a_input\n    { type: \"storage-read\" }, // b_input\n    { type: \"storage-read\" }, // A_log\n    { type: \"storage-read\" }, // dt_bias\n    { type: \"storage-read-write\" }, // ssm_state\n    { type: \"storage-read-write\" }, // output\n    { type: \"uniform\" }, // params\n  ],\n\n  getDispatchSize(op) {\n    const num_heads = op.attributes.num_heads as number;\n    return [num_heads, 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const num_heads = op.attributes.num_heads as number;\n    const key_dim = op.attributes.key_dim as number;\n    const val_dim = op.attributes.val_dim as number;\n    const qkv_dim = op.attributes.qkv_dim as number;\n    // Resolve T from qkv_conv tensor reference\n    const qkvRef = op.attributes.T_tensor as string | undefined;\n    let T: number;\n    if (qkvRef && resolvedShapes[qkvRef]) {\n      T = resolvedShapes[qkvRef][0];\n    } else {\n      T = op.attributes.T as number;\n    }\n    return buildUniformBuffer([T, num_heads, key_dim, val_dim, qkv_dim, 0, 0, 0]);\n  },\n};\n\n// ── MambaSSM with f16 state (Dawn only — selected in executor when hasF16) ──\n\nexport const MAMBA_SSM_F16_SPEC: KernelSpec = {\n  ...mambaSSMSpec,\n  shaderCode: WGSL_MAMBA_SSM_F16,\n};\n\n// ── KVCacheAppend ──\n\nconst kvCacheAppendSpec: KernelSpec = {\n  shaderCode: WGSL_KV_CACHE_APPEND,\n  entryPoint: \"main\",\n\n  // @binding(0) src: storage read\n  // @binding(1) dst: storage read_write\n  // @binding(2) params: uniform\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n\n  getDispatchSize(op, resolvedShapes) {\n    const width = op.attributes.width as number;\n    const srcRef = op.attributes.T_tensor as string | undefined;\n    let T: number;\n    if (srcRef && resolvedShapes[srcRef]) {\n      T = resolvedShapes[srcRef][0];\n    } else {\n      T = op.attributes.T as number;\n    }\n    return [cdiv(T * width, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes, context) {\n    const width = op.attributes.width as number;\n    const seqPos = context?.seqPos ?? 0;\n    const srcRef = op.attributes.T_tensor as string | undefined;\n    let T: number;\n    if (srcRef && resolvedShapes[srcRef]) {\n      T = resolvedShapes[srcRef][0];\n    } else {\n      T = op.attributes.T as number;\n    }\n    return buildUniformBuffer([T * width, seqPos * width]);\n  },\n};\n\n// ── Dual KVCacheAppend (f32 K+V into both caches, one dispatch) ──\n\nexport const DUAL_KV_CACHE_APPEND_SPEC: KernelSpec = {\n  shaderCode: WGSL_DUAL_KV_CACHE_APPEND,\n  entryPoint: \"main\",\n\n  // @binding(0) src_k, (1) src_v, (2) dst_k, (3) dst_v, (4) params\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const width = op.attributes.width as number;\n    const srcRef = op.attributes.T_tensor as string | undefined;\n    const T =\n      srcRef && resolvedShapes[srcRef] ? resolvedShapes[srcRef][0] : (op.attributes.T as number);\n    return [cdiv(T * width, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes, context) {\n    const width = op.attributes.width as number;\n    const seqPos = context?.seqPos ?? 0;\n    const srcRef = op.attributes.T_tensor as string | undefined;\n    const T =\n      srcRef && resolvedShapes[srcRef] ? resolvedShapes[srcRef][0] : (op.attributes.T as number);\n    return buildUniformBuffer([T * width, seqPos * width]);\n  },\n};\n\n// ── Dual KVCacheAppend native-f16 (f32 K+V → f16 caches, one dispatch) ──\n\nexport const DUAL_KV_CACHE_APPEND_F16_SPEC: KernelSpec = {\n  ...DUAL_KV_CACHE_APPEND_SPEC,\n  shaderCode: WGSL_DUAL_KV_CACHE_APPEND_F16,\n};\n\n// ── Dual KVCacheAppend packed-f16 (Safari-safe, one dispatch) ──\n\nexport const DUAL_KV_CACHE_APPEND_PACKED_F16_SPEC: KernelSpec = {\n  ...DUAL_KV_CACHE_APPEND_SPEC,\n  shaderCode: WGSL_DUAL_KV_CACHE_APPEND_PACKED_F16,\n  // Each thread handles one f16 PAIR per cache → grid over count/2.\n  getDispatchSize(op, resolvedShapes) {\n    const width = op.attributes.width as number;\n    const srcRef = op.attributes.T_tensor as string | undefined;\n    const T =\n      srcRef && resolvedShapes[srcRef] ? resolvedShapes[srcRef][0] : (op.attributes.T as number);\n    return [cdiv((T * width) >> 1, 256), 1, 1];\n  },\n};\n\n// ── KVCacheAppend F16 (f32 src → f16 dst) ──\n\nexport const KV_CACHE_APPEND_F16_SPEC: KernelSpec = {\n  shaderCode: WGSL_KV_CACHE_APPEND_F16,\n  entryPoint: \"main\",\n\n  // @binding(0) src: storage read (f32)\n  // @binding(1) dst: storage read_write (f16)\n  // @binding(2) params: uniform\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n\n  getDispatchSize(op, resolvedShapes) {\n    const width = op.attributes.width as number;\n    const srcRef = op.attributes.T_tensor as string | undefined;\n    let T: number;\n    if (srcRef && resolvedShapes[srcRef]) {\n      T = resolvedShapes[srcRef][0];\n    } else {\n      T = op.attributes.T as number;\n    }\n    return [cdiv(T * width, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes, context) {\n    const width = op.attributes.width as number;\n    const seqPos = context?.seqPos ?? 0;\n    const srcRef = op.attributes.T_tensor as string | undefined;\n    let T: number;\n    if (srcRef && resolvedShapes[srcRef]) {\n      T = resolvedShapes[srcRef][0];\n    } else {\n      T = op.attributes.T as number;\n    }\n    return buildUniformBuffer([T * width, seqPos * width]);\n  },\n};\n\n// ── Attention F16 (f16 KV cache) ──\n\nexport const ATTENTION_F16_SPEC: KernelSpec = {\n  shaderCode: WGSL_ATTENTION_F16,\n  entryPoint: \"main\",\n\n  // @binding(0) Q: storage read (f32)\n  // @binding(1) K: storage read (f16)\n  // @binding(2) V: storage read (f16)\n  // @binding(3) output: storage read_write (f32)\n  // @binding(4) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const num_q_heads = op.attributes.num_q_heads as number;\n    const outShape = resolvedShapes[op.outputs[0]];\n    const T = outShape ? outShape[0] : (op.attributes.T as number);\n    return [T, num_q_heads, 1];\n  },\n\n  buildParams(op, resolvedShapes, context) {\n    const num_q_heads = op.attributes.num_q_heads as number;\n    const num_kv_heads = op.attributes.num_kv_heads as number;\n    const head_dim = op.attributes.head_dim as number;\n    const position_offset = context?.seqPos ?? (op.attributes.position_offset as number) ?? 0;\n\n    const outShape = resolvedShapes[op.outputs[0]];\n    const T = outShape ? outShape[0] : (op.attributes.T as number);\n\n    const kvShape = resolvedShapes[op.inputs[1]];\n    const S = kvShape ? kvShape[0] : T;\n\n    // Default 1/sqrt(head_dim) → byte-identical to the legacy scale; Gemma4 = 1.0.\n    const attn_scale = (op.attributes.attn_scale as number) ?? 1 / Math.sqrt(head_dim);\n\n    return buildUniformBuffer([\n      T,\n      S,\n      num_q_heads,\n      num_kv_heads,\n      head_dim,\n      position_offset,\n      f32BitsToU32(attn_scale),\n    ]);\n  },\n};\n\n// ── ConvStateUpdate ──\n\nconst convStateUpdateSpec: KernelSpec = {\n  shaderCode: WGSL_CONV_STATE_UPDATE,\n  entryPoint: \"main\",\n\n  // @binding(0) input: storage read\n  // @binding(1) conv_state: storage read_write\n  // @binding(2) params: uniform\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n\n  getDispatchSize(op) {\n    const channels = op.attributes.channels as number;\n    return [cdiv(channels, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const channels = op.attributes.channels as number;\n    const state_size = op.attributes.state_size as number;\n    const srcRef = op.attributes.T_tensor as string | undefined;\n    let seq_len: number;\n    if (srcRef && resolvedShapes[srcRef]) {\n      seq_len = resolvedShapes[srcRef][0];\n    } else {\n      seq_len = op.attributes.T as number;\n    }\n    return buildUniformBuffer([seq_len, channels, state_size, 0]);\n  },\n};\n\n// ── Registry ────────────────────────────────────────────────────────────\n\n/**\n * Maps each implemented OpType to its kernel spec.\n *\n * Only ops with WGSL shader implementations are included.\n * Stubbed ops (MoERouter, ExpertMatMul, Conv2d, AvgPool2d,\n * CrossAttention, Gather, Reshape, Transpose, Concat) are not present.\n */\n// ── Audio codec-decoder kernels (OmniVoice / HiggsAudioV2 DAC vocoder) ──\n//\n// New vocoder primitives for native TTS decode (token grid -> 24kHz PCM). The\n// DAC decoder is: Conv1dFull (in) -> 5 upsample blocks (ConvTranspose1d + Snake1d\n// + dilated residual Conv1dFull) -> Snake1d -> Conv1dFull (out, NO tanh). These\n// three kernels cover every op in that path.\n//\n// Validated bit-exact vs NumPy reference (max|err| ~2e-7, f32 rounding only) in\n// scripts/engine/test-codec-kernels.mjs across: kernel-7 dilation-1/3, pointwise\n// kernel-1, transposed stride-4 (even) and stride-3 (odd, output_padding=1), and\n// per-channel Snake. Mobile-safe: flat workgroup, no select(), no exp(), no\n// shared memory, 1 thread per output element.\n\nconst WGSL_CONV1D_FULL = `\\\n// Full cross-channel 1D convolution: x[Cin,L] * w[Cout,Cin,K] + bias[Cout]\n//   -> y[Cout,Lout],  Lout = (L + 2*padding - dilation*(K-1) - 1)/stride + 1\n// One thread per (oc, ot) output element.\nstruct Params {\n  Cin: u32, Cout: u32, L: u32, Lout: u32,\n  K: u32, stride: u32, padding: u32, dilation: u32,\n}\n@group(0) @binding(0) var<storage, read> x: array<f32>;\n@group(0) @binding(1) var<storage, read> w: array<f32>;\n@group(0) @binding(2) var<storage, read> bias: array<f32>;\n@group(0) @binding(3) var<storage, read_write> y: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\n@compute @workgroup_size(64)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  let total = params.Cout * params.Lout;\n  if (idx >= total) { return; }\n  let oc = idx / params.Lout;\n  let ot = idx % params.Lout;\n  let base = i32(ot * params.stride) - i32(params.padding);\n  var acc: f32 = 0.0;\n  for (var ic: u32 = 0u; ic < params.Cin; ic = ic + 1u) {\n    let x_row = ic * params.L;\n    let w_row = (oc * params.Cin + ic) * params.K;\n    for (var k: u32 = 0u; k < params.K; k = k + 1u) {\n      let it = base + i32(k * params.dilation);\n      if (it >= 0 && it < i32(params.L)) {\n        acc = acc + x[x_row + u32(it)] * w[w_row + k];\n      }\n    }\n  }\n  y[oc * params.Lout + ot] = acc + bias[oc];\n}\n`;\n\nconst WGSL_CONV_TRANSPOSE1D = `\\\n// Transposed (strided) 1D convolution: x[Cin,L] * w[Cin,Cout,K] + bias[Cout]\n//   -> y[Cout,Lout],  Lout = (L-1)*stride - 2*padding + (K-1) + output_padding + 1\n// Gather form (one thread per output): for full-buffer position op = ot+padding,\n// sum over (ic,k) where op-k is a non-negative multiple of stride -> it=(op-k)/stride.\n// HiggsAudioV2 sets output_padding = stride % 2 on every DAC upsample block.\nstruct Params {\n  Cin: u32, Cout: u32, L: u32, Lout: u32,\n  K: u32, stride: u32, padding: u32, out_pad: u32,\n}\n@group(0) @binding(0) var<storage, read> x: array<f32>;\n@group(0) @binding(1) var<storage, read> w: array<f32>;\n@group(0) @binding(2) var<storage, read> bias: array<f32>;\n@group(0) @binding(3) var<storage, read_write> y: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\n@compute @workgroup_size(64)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  let total = params.Cout * params.Lout;\n  if (idx >= total) { return; }\n  let oc = idx / params.Lout;\n  let ot = idx % params.Lout;\n  let op = i32(ot + params.padding);\n  var acc: f32 = 0.0;\n  for (var ic: u32 = 0u; ic < params.Cin; ic = ic + 1u) {\n    let x_row = ic * params.L;\n    let w_row = (ic * params.Cout + oc) * params.K;\n    for (var k: u32 = 0u; k < params.K; k = k + 1u) {\n      let num = op - i32(k);\n      if (num >= 0 && (num % i32(params.stride)) == 0) {\n        let it = num / i32(params.stride);\n        if (it >= 0 && it < i32(params.L)) {\n          acc = acc + x[x_row + u32(it)] * w[w_row + k];\n        }\n      }\n    }\n  }\n  y[oc * params.Lout + ot] = acc + bias[oc];\n}\n`;\n\nconst WGSL_SNAKE1D = `\\\n// Snake activation (learned per-channel periodic): y = x + (1/(alpha+1e-9)) * sin(alpha*x)^2\nstruct Params { C: u32, L: u32 }\n@group(0) @binding(0) var<storage, read> x: array<f32>;\n@group(0) @binding(1) var<storage, read> alpha: array<f32>;\n@group(0) @binding(2) var<storage, read_write> y: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\n@compute @workgroup_size(64)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  let total = params.C * params.L;\n  if (idx >= total) { return; }\n  let c = idx / params.L;\n  let a = alpha[c];\n  let s = sin(a * x[idx]);\n  y[idx] = x[idx] + (1.0 / (a + 1e-9)) * s * s;\n}\n`;\n\n// ── Moonshine STT frontend kernels (Tanh, Transpose, GroupNorm) ──\n\nconst WGSL_TANH = `\\\n// Elementwise tanh (Moonshine conv1 activation). Clamp the argument: on Metal/Dawn\n// tanh() of a large magnitude returns NaN (internal exp overflows). |arg|>=10\n// already saturates to +/-1, so clamping to +/-20 is numerically exact and safe.\nstruct Params { count: u32 }\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read_write> output: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n  output[idx] = tanh(clamp(input[idx], -20.0, 20.0));\n}\n`;\n\nconst WGSL_TRANSPOSE = `\\\n// Transpose a row-major [rows, cols] tensor into [cols, rows].\n// Moonshine conv output is [C, L] (channels-major); the encoder transformer wants\n// [L, C] (token-major). One thread per output element; bounds-checked.\nstruct Params { rows: u32, cols: u32 }\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read_write> output: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  let total = params.rows * params.cols;\n  if (idx >= total) { return; }\n  // output is [cols, rows]: out[c, r] = in[r, c]\n  let c = idx / params.rows;\n  let r = idx % params.rows;\n  output[idx] = input[r * params.cols + c];\n}\n`;\n\nconst WGSL_GROUPNORM = `\\\n// GroupNorm(num_groups=1) over the channel axis of a [C, L] tensor: normalize each\n// channel-position element by the mean/variance computed across ALL C*L elements\n// (a single group), then scale/shift per-channel: y[c,l] = (x[c,l]-mu)/sqrt(var+eps)\n// * weight[c] + bias[c]. Moonshine's encoder frontend applies this after conv1.\n// One workgroup reduces the whole [C, L] tensor (C*L <= a few k for short clips).\nstruct Params { channels: u32, length: u32, eps_bits: u32, _pad: u32 }\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read> weight: array<f32>;\n@group(0) @binding(2) var<storage, read> bias: array<f32>;\n@group(0) @binding(3) var<storage, read_write> output: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\nvar<workgroup> shared_sum: array<f32, 256>;\nvar<workgroup> shared_sq: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(@builtin(local_invocation_id) lid: vec3u) {\n  let tid = lid.x;\n  let total = params.channels * params.length;\n  let eps = bitcast<f32>(params.eps_bits);\n\n  var local_sum: f32 = 0.0;\n  var local_sq: f32 = 0.0;\n  var i = tid;\n  while (i < total) {\n    let v = input[i];\n    local_sum += v;\n    local_sq += v * v;\n    i += 256u;\n  }\n  shared_sum[tid] = local_sum;\n  shared_sq[tid] = local_sq;\n  workgroupBarrier();\n\n  var stride: u32 = 128u;\n  while (stride > 0u) {\n    if (tid < stride) {\n      shared_sum[tid] += shared_sum[tid + stride];\n      shared_sq[tid] += shared_sq[tid + stride];\n    }\n    workgroupBarrier();\n    stride /= 2u;\n  }\n\n  let n = f32(total);\n  let mean = shared_sum[0] / n;\n  let variance = shared_sq[0] / n - mean * mean;\n  let inv_std = 1.0 / sqrt(variance + eps);\n  workgroupBarrier();\n\n  i = tid;\n  while (i < total) {\n    let c = i / params.length; // channel index (row-major [C, L])\n    output[i] = (input[i] - mean) * inv_std * weight[c] + bias[c];\n    i += 256u;\n  }\n}\n`;\n\nconst WGSL_LAYERNORM_NO_BIAS = `\\\n// LayerNorm without a bias term: output = ((x - mean) / sqrt(var + eps)) * weight.\n// Moonshine's encoder/decoder norms are weight-only (no .bias in the checkpoint),\n// so the node binds only [input, weight] — this variant drops the bias binding to\n// keep the executor's positional input->binding mapping correct.\nstruct Params {\n  seq_len: u32,\n  hidden_size: u32,\n  eps_bits: u32,\n  _pad: u32,\n}\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read> weight: array<f32>;\n@group(0) @binding(2) var<storage, read_write> output: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\nvar<workgroup> shared_sum: array<f32, 256>;\nvar<workgroup> shared_sq_sum: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let row = wid.x;\n  if (row >= params.seq_len) { return; }\n\n  let tid = lid.x;\n  let row_offset = row * params.hidden_size;\n  let eps = bitcast<f32>(params.eps_bits);\n\n  var local_sum: f32 = 0.0;\n  var i = tid;\n  while (i < params.hidden_size) {\n    local_sum += input[row_offset + i];\n    i += 256u;\n  }\n  shared_sum[tid] = local_sum;\n  workgroupBarrier();\n\n  var stride: u32 = 128u;\n  while (stride > 0u) {\n    if (tid < stride) { shared_sum[tid] += shared_sum[tid + stride]; }\n    workgroupBarrier();\n    stride /= 2u;\n  }\n  let mean = shared_sum[0] / f32(params.hidden_size);\n\n  var local_sq: f32 = 0.0;\n  i = tid;\n  while (i < params.hidden_size) {\n    let diff = input[row_offset + i] - mean;\n    local_sq += diff * diff;\n    i += 256u;\n  }\n  shared_sq_sum[tid] = local_sq;\n  workgroupBarrier();\n\n  stride = 128u;\n  while (stride > 0u) {\n    if (tid < stride) { shared_sq_sum[tid] += shared_sq_sum[tid + stride]; }\n    workgroupBarrier();\n    stride /= 2u;\n  }\n  let inv_std = 1.0 / sqrt(shared_sq_sum[0] / f32(params.hidden_size) + eps);\n\n  i = tid;\n  while (i < params.hidden_size) {\n    output[row_offset + i] = (input[row_offset + i] - mean) * inv_std * weight[i];\n    i += 256u;\n  }\n}\n`;\n\n/**\n * LayerNorm with no bias term (Moonshine). Binds [input, weight, output, params]\n * — one fewer storage-read than the standard layernormSpec. The executor selects\n * this variant when a LayerNorm node carries only 2 inputs.\n */\nexport const LAYERNORM_NO_BIAS_SPEC: KernelSpec = {\n  shaderCode: WGSL_LAYERNORM_NO_BIAS,\n  entryPoint: \"main\",\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n  getDispatchSize(op, resolvedShapes) {\n    const hidden_size = op.attributes.hidden_size as number;\n    const ref = op.attributes.seq_len_tensor as string | undefined;\n    let seq_len: number;\n    if (ref && resolvedShapes[ref]) {\n      const shape = resolvedShapes[ref];\n      seq_len = shape.reduce((a, b) => a * b, 1) / hidden_size;\n    } else {\n      seq_len = op.attributes.seq_len as number;\n    }\n    return [seq_len, 1, 1];\n  },\n  buildParams(op, resolvedShapes) {\n    const hidden_size = op.attributes.hidden_size as number;\n    const eps = op.attributes.eps as number;\n    const ref = op.attributes.seq_len_tensor as string | undefined;\n    let seq_len: number;\n    if (ref && resolvedShapes[ref]) {\n      const shape = resolvedShapes[ref];\n      seq_len = shape.reduce((a, b) => a * b, 1) / hidden_size;\n    } else {\n      seq_len = op.attributes.seq_len as number;\n    }\n    return buildUniformBuffer([seq_len, hidden_size, f32BitsToU32(eps), 0]);\n  },\n};\n\nconst tanhSpec: KernelSpec = {\n  shaderCode: WGSL_TANH,\n  entryPoint: \"main\",\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n  buildParams(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return buildUniformBuffer([count]);\n  },\n};\n\nconst transposeSpec: KernelSpec = {\n  shaderCode: WGSL_TRANSPOSE,\n  entryPoint: \"main\",\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n  buildParams(op, resolvedShapes) {\n    // input shape [rows, cols]; output is [cols, rows].\n    const inShape = resolvedShapes[op.inputs[0]];\n    const rows = inShape ? inShape[0] : (op.attributes.rows as number);\n    const cols = inShape ? inShape[1] : (op.attributes.cols as number);\n    return buildUniformBuffer([rows, cols]);\n  },\n};\n\nconst groupnormSpec: KernelSpec = {\n  shaderCode: WGSL_GROUPNORM,\n  entryPoint: \"main\",\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n  // One workgroup reduces the full [C, L] tensor.\n  getDispatchSize() {\n    return [1, 1, 1];\n  },\n  buildParams(op, resolvedShapes) {\n    const channels = op.attributes.channels as number;\n    const inShape = resolvedShapes[op.inputs[0]];\n    const length = inShape ? inShape[1] : (op.attributes.length as number);\n    const eps = (op.attributes.eps as number) ?? 1e-5;\n    return buildUniformBuffer([channels, length, f32BitsToU32(eps), 0]);\n  },\n};\n\nconst conv1dFullSpec: KernelSpec = {\n  shaderCode: WGSL_CONV1D_FULL,\n  entryPoint: \"main\",\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n  getDispatchSize(op) {\n    const Cout = op.attributes.Cout as number;\n    const Lout = op.attributes.Lout as number;\n    return [cdiv(Cout * Lout, 64), 1, 1];\n  },\n  buildParams(op) {\n    return buildUniformBuffer([\n      op.attributes.Cin as number,\n      op.attributes.Cout as number,\n      op.attributes.L as number,\n      op.attributes.Lout as number,\n      op.attributes.K as number,\n      op.attributes.stride as number,\n      op.attributes.padding as number,\n      op.attributes.dilation as number,\n    ]);\n  },\n};\n\nconst convTranspose1dSpec: KernelSpec = {\n  shaderCode: WGSL_CONV_TRANSPOSE1D,\n  entryPoint: \"main\",\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n  getDispatchSize(op) {\n    const Cout = op.attributes.Cout as number;\n    const Lout = op.attributes.Lout as number;\n    return [cdiv(Cout * Lout, 64), 1, 1];\n  },\n  buildParams(op) {\n    return buildUniformBuffer([\n      op.attributes.Cin as number,\n      op.attributes.Cout as number,\n      op.attributes.L as number,\n      op.attributes.Lout as number,\n      op.attributes.K as number,\n      op.attributes.stride as number,\n      op.attributes.padding as number,\n      op.attributes.output_padding as number,\n    ]);\n  },\n};\n\nconst snake1dSpec: KernelSpec = {\n  shaderCode: WGSL_SNAKE1D,\n  entryPoint: \"main\",\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n  getDispatchSize(op) {\n    const C = op.attributes.C as number;\n    const L = op.attributes.L as number;\n    return [cdiv(C * L, 64), 1, 1];\n  },\n  buildParams(op) {\n    return buildUniformBuffer([op.attributes.C as number, op.attributes.L as number]);\n  },\n};\n\n// ════════════════════════════════════════════════════════════════════════\n// Kani-TTS-2 NanoCodec (NVIDIA NeMo, 22 kHz 0.6 kbps 12.5 fps) decoder kernels.\n// Validated bit-exact vs the real MLX NanoCodec via scripts/engine/test-nanocodec\n// -decode.mjs (max|err| 4.2e-6, f32 rounding only). Three new ops beyond the DAC\n// cluster above:\n//   - FSQDequant: code idx -> per-dim continuous latent (finite scalar quantization,\n//     4 groups x 4 dims, levels [9,8,8,7], mixed-radix base [1,9,72,576]).\n//   - HalfSnake1d: snake on the first C/2 channels, leaky-relu(0.01) on the rest.\n//   - ConvTranspose1dDepthwise: causal depthwise (groups=Cout) transposed conv —\n//     insert (stride-1) zeros, cross-correlate the FLIPPED kernel, right-trim.\n// The causal full conv reuses Conv1dFull with padding=(K-1)*dilation, Lout=L.\n// Mobile-safe: flat workgroup_size(64), no select(), no shared memory, no exp(),\n// 1 thread per output element, no `enable f16`.\n// ════════════════════════════════════════════════════════════════════════\n\nconst WGSL_FSQ_DEQUANT = `\\\n// Finite-scalar-quantization decode. codes[groups,T] (u32) -> latent[groups*dims,T]\n// (f32). Channel ch = g*dims + d; per-dim level/base passed as small constants.\nstruct Params {\n  groups: u32, dims: u32, T: u32, _pad: u32,\n  l0: u32, l1: u32, l2: u32, l3: u32,\n  b0: u32, b1: u32, b2: u32, b3: u32,\n}\n@group(0) @binding(0) var<storage, read> codes: array<u32>;        // [groups, T]\n@group(0) @binding(1) var<storage, read_write> latent: array<f32>; // [groups*dims, T]\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(64)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let total = params.groups * params.dims * params.T;\n  if (gid.x >= total) { return; }\n  let ch = gid.x / params.T;\n  let t = gid.x % params.T;\n  let d = ch % params.dims;\n  let g = ch / params.dims;\n  var lvl: u32 = params.l0;\n  var base: u32 = params.b0;\n  if (d == 1u) { lvl = params.l1; base = params.b1; }\n  else if (d == 2u) { lvl = params.l2; base = params.b2; }\n  else if (d == 3u) { lvl = params.l3; base = params.b3; }\n  let idx = codes[g * params.T + t];\n  let nonneg = (idx / base) % lvl;\n  let scale = lvl / 2u;\n  latent[ch * params.T + t] = (f32(nonneg) - f32(scale)) / f32(scale);\n}\n`;\n\nconst WGSL_HALF_SNAKE1D = `\\\n// HalfSnake: first half channels -> snake(x; alpha); rest -> leaky_relu(x, 0.01).\nstruct Params { C: u32, L: u32, half: u32, _pad: u32 }\n@group(0) @binding(0) var<storage, read> x: array<f32>;       // [C, L]\n@group(0) @binding(1) var<storage, read> alpha: array<f32>;   // [half]\n@group(0) @binding(2) var<storage, read_write> y: array<f32>; // [C, L]\n@group(0) @binding(3) var<storage, read> params: Params;\n\n@compute @workgroup_size(64)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let total = params.C * params.L;\n  if (gid.x >= total) { return; }\n  let c = gid.x / params.L;\n  let v = x[gid.x];\n  if (c < params.half) {\n    let a = alpha[c];\n    let s = sin(a * v);\n    y[gid.x] = v + (1.0 / (a + 1e-9)) * s * s;\n  } else {\n    if (v >= 0.0) { y[gid.x] = v; } else { y[gid.x] = 0.01 * v; }\n  }\n}\n`;\n\nconst WGSL_CONV_TRANSPOSE1D_DW = `\\\n// Causal depthwise (groups = Cout) transposed conv. Weight in PyTorch\n// ConvTranspose1d layout [Cin, 1, K] (out/groups = 1); each output channel oc is fed\n// by cpg = Cin/Cout consecutive input channels (oc*cpg .. oc*cpg+cpg-1). Matches MLX\n// CausalConvTranspose1d: insert (stride-1) zeros, pad (K-1) both sides, cross-correlate\n// the FLIPPED kernel, then trim (K-stride) from the right. Output length Lout = L*stride.\n// Gather form (one thread per (oc, ot)): contributions where it*stride = ot+k-(K-1),\n// 0<=k<K, with flipped weight w[K-1-k].\nstruct Params { Cin: u32, Cout: u32, L: u32, Lout: u32, K: u32, stride: u32, cpg: u32, _pad: u32 }\n@group(0) @binding(0) var<storage, read> x: array<f32>;       // [Cin, L]\n@group(0) @binding(1) var<storage, read> w: array<f32>;       // [Cin, 1, K]\n@group(0) @binding(2) var<storage, read> bias: array<f32>;    // [Cout]\n@group(0) @binding(3) var<storage, read_write> y: array<f32>; // [Cout, Lout]\n@group(0) @binding(4) var<storage, read> params: Params;\n\n@compute @workgroup_size(64)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let total = params.Cout * params.Lout;\n  if (gid.x >= total) { return; }\n  let oc = gid.x / params.Lout;\n  let ot = gid.x % params.Lout;\n  var acc: f32 = 0.0;\n  for (var j: u32 = 0u; j < params.cpg; j = j + 1u) {\n    let ic = oc * params.cpg + j;\n    let x_row = ic * params.L;\n    let w_row = ic * params.K;\n    for (var k: u32 = 0u; k < params.K; k = k + 1u) {\n      let num = i32(ot) + i32(k) - (i32(params.K) - 1);\n      if (num >= 0 && (num % i32(params.stride)) == 0) {\n        let it = num / i32(params.stride);\n        if (it >= 0 && it < i32(params.L)) {\n          acc = acc + x[x_row + u32(it)] * w[w_row + (params.K - 1u - k)];\n        }\n      }\n    }\n  }\n  y[oc * params.Lout + ot] = acc + bias[oc];\n}\n`;\n\nconst fsqDequantSpec: KernelSpec = {\n  shaderCode: WGSL_FSQ_DEQUANT,\n  entryPoint: \"main\",\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n  getDispatchSize(op) {\n    const groups = op.attributes.groups as number;\n    const dims = op.attributes.dims as number;\n    const T = op.attributes.T as number;\n    return [cdiv(groups * dims * T, 64), 1, 1];\n  },\n  buildParams(op) {\n    const levels = op.attributes.levels as number[];\n    const base = op.attributes.base as number[];\n    return buildUniformBuffer([\n      op.attributes.groups as number,\n      op.attributes.dims as number,\n      op.attributes.T as number,\n      0,\n      levels[0],\n      levels[1],\n      levels[2],\n      levels[3],\n      base[0],\n      base[1],\n      base[2],\n      base[3],\n    ]);\n  },\n};\n\nconst halfSnake1dSpec: KernelSpec = {\n  shaderCode: WGSL_HALF_SNAKE1D,\n  entryPoint: \"main\",\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n  getDispatchSize(op) {\n    const C = op.attributes.C as number;\n    const L = op.attributes.L as number;\n    return [cdiv(C * L, 64), 1, 1];\n  },\n  buildParams(op) {\n    const C = op.attributes.C as number;\n    return buildUniformBuffer([C, op.attributes.L as number, Math.floor(C / 2), 0]);\n  },\n};\n\nconst convTranspose1dDepthwiseSpec: KernelSpec = {\n  shaderCode: WGSL_CONV_TRANSPOSE1D_DW,\n  entryPoint: \"main\",\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n  getDispatchSize(op) {\n    const Cout = op.attributes.Cout as number;\n    const Lout = op.attributes.Lout as number;\n    return [cdiv(Cout * Lout, 64), 1, 1];\n  },\n  buildParams(op) {\n    const Cin = op.attributes.Cin as number;\n    const Cout = op.attributes.Cout as number;\n    return buildUniformBuffer([\n      Cin,\n      Cout,\n      op.attributes.L as number,\n      op.attributes.Lout as number,\n      op.attributes.K as number,\n      op.attributes.stride as number,\n      Math.floor(Cin / Cout),\n      0,\n    ]);\n  },\n};\n\nexport const KERNEL_REGISTRY: Partial<Record<OpType, KernelSpec>> = {\n  Embedding: embeddingSpec,\n  EmbeddingInt4: embeddingInt4Spec,\n  MatMul: matmulSpec,\n  MatMulBias: matmulBiasSpec,\n  MatMulInt4: matmulInt4Spec,\n  Add: addSpec,\n  Mul: mulSpec,\n  SwiGLU: swigluSpec,\n  SiLU: siluSpec,\n  ResidualRMSNorm: residualRmsnormSpec,\n  GELU: geluSpec,\n  GeluErf: geluErfSpec,\n  AddBias: addBiasSpec,\n  ApplyRotaryEmb: applyRotarySpec,\n  MRoPE: mropeSpec,\n  EmbedSplice: embedSpliceSpec,\n  SliceCols: sliceColsSpec,\n  MulCols: mulColsSpec,\n  PoolMatMul: poolMatMulSpec,\n  ClippedMatMul: clippedMatMulSpec,\n  RMSNorm: rmsnormSpec,\n  LayerNorm: layernormSpec,\n  RoPE: ropeSpec,\n  Attention: attentionSpec,\n  CrossAttention: crossAttentionSpec,\n  Softmax: softmaxSpec,\n  CausalConv1d: causalConv1dSpec,\n  CausalConv1dSiLU: causalConv1dSiluSpec,\n  CausalConv1dGated: causalConv1dGatedSpec,\n  SigmoidGate: sigmoidGateSpec,\n  MambaSSM: mambaSSMSpec,\n  KVCacheAppend: kvCacheAppendSpec,\n  ConvStateUpdate: convStateUpdateSpec,\n  SliceLastRow: sliceLastRowSpec,\n  MeanPool: meanPoolSpec,\n  Scale: scaleSpec,\n  Softcap: softcapSpec,\n  L2Norm: l2NormSpec,\n  // Audio codec-decoder (OmniVoice native TTS) — validated bit-exact, see above.\n  Conv1dFull: conv1dFullSpec,\n  ConvTranspose1d: convTranspose1dSpec,\n  Snake1d: snake1dSpec,\n  // Kani-TTS-2 NanoCodec decoder — validated bit-exact vs MLX, see above.\n  FSQDequant: fsqDequantSpec,\n  HalfSnake1d: halfSnake1dSpec,\n  ConvTranspose1dDepthwise: convTranspose1dDepthwiseSpec,\n  // Moonshine STT encoder frontend.\n  Tanh: tanhSpec,\n  Transpose: transposeSpec,\n  GroupNorm: groupnormSpec,\n};\n","/**\n * Graph executor — runs a ModelGraph on WebGPU.\n *\n * Optimized for minimal per-token overhead:\n * - Persistent bind groups (created once after weight upload, reused every forward)\n * - Pre-allocated input buffer (no per-forward buffer creation)\n * - Cached uniform buffers (contents updated via writeBuffer, no recreation)\n * - GPU-side argmax for greedy decode (4 bytes readback vs vocab_size*4)\n * - Dual dispatch paths: tiled matmul for prefill, K-parallel matvec for decode\n */\n\nimport type { KvMode } from \"./architectures/index.js\";\nimport type { GPUContext } from \"./device.js\";\nimport {\n  createBindGroup,\n  createReadbackBuffer,\n  createStorageBuffer,\n  createUniformBuffer,\n  destroyBuffers,\n  getOrCreatePipeline,\n} from \"./device.js\";\nimport type { ModelGraph, OpNode } from \"./ir.js\";\nimport { DTYPE_BYTES } from \"./ir.js\";\nimport {\n  ARGMAX_SPEC,\n  ATTENTION_F16_SPEC,\n  ATTENTION_PACKED_F16_SPEC,\n  DUAL_KV_CACHE_APPEND_F16_SPEC,\n  DUAL_KV_CACHE_APPEND_PACKED_F16_SPEC,\n  DUAL_KV_CACHE_APPEND_SPEC,\n  DUAL_MATVEC_INT4_SPEC,\n  DUAL_RMSNORM_SPEC,\n  GATED_MATVEC_INT4_SPEC,\n  KERNEL_REGISTRY,\n  type KernelSpec,\n  KV_CACHE_APPEND_F16_SPEC,\n  KV_CACHE_APPEND_PACKED_F16_SPEC,\n  LAYERNORM_NO_BIAS_SPEC,\n  MAMBA_SSM_F16_SPEC,\n  MATVEC_INT4_SPEC,\n  MATVEC_INT4_SUBGROUPS_SPEC,\n  MATVEC_SPEC,\n  MATVEC_SUBGROUPS_SPEC,\n  ROPE_INTERLEAVED_SPEC,\n  type RuntimeContext,\n  SWIGLU_GATED_MATVEC_INT4_SPEC,\n  SWIGLU_MATVEC_INT4_SPEC,\n  SWIGLU_MATVEC_SPEC,\n} from \"./kernels/registry.js\";\nimport type { WeightSource } from \"./weight-source.js\";\n\nconst MAP_MODE_READ = 0x0001;\n\n/**\n * Safari/Metal workaround: shader variant alternation.\n * Metal caches argument buffers per compiled function. When consecutive dispatches\n * use the same WGSL code (same Metal function), Metal reuses the previous dispatch's\n * argument buffer, ignoring setBindGroup(). We alternate between variant 0/1 of each\n * shader (prepending `const _MV: u32 = Xu;`) to force different Metal function\n * specializations, preventing argument buffer reuse.\n */\n\nexport interface ExecutorOptions {\n  maxSeqLen: number;\n  /** KV cache kernel strategy. Defaults to \"native-f16\" when not specified. */\n  kvMode?: KvMode;\n  /**\n   * WebKit only: dispatches per command buffer, with at most one command\n   * buffer in flight (awaited). 1 (default) is the proven-correct floor on\n   * iPad; larger values are faster if this WebKit version keeps storage\n   * writes visible across dispatches within one submission. Sweepable via\n   * the ?group=N URL param.\n   */\n  webkitGroupSize?: number;\n}\n\nexport interface ForwardResult {\n  logits: Float32Array;\n}\n\n/**\n * Pre-compiled dispatch entry for a single graph op.\n * Created once during initBindGroups(), reused every forward pass.\n */\ninterface DispatchEntry {\n  nodeId: string;\n  node: OpNode;\n  spec: KernelSpec;\n  pipeline: GPUComputePipeline;\n  bindGroup: GPUBindGroup;\n  uniformBuffer: GPUBuffer;\n  /** Cached last-written params bytes. Skip writeBuffer when unchanged. */\n  lastParamsBytes: Uint8Array | null;\n  /** Cached last dispatch size. */\n  lastDispatchSize: [number, number, number] | null;\n}\n\nexport class Executor {\n  private ctx: GPUContext;\n  private graph: ModelGraph;\n\n  private weightBuffers: Map<string, GPUBuffer> = new Map();\n  private activationBuffers: Map<string, GPUBuffer> = new Map();\n  private ssmStateBuffers: Map<string, GPUBuffer> = new Map();\n  private kvCacheBuffers: Map<string, GPUBuffer> = new Map();\n\n  /** Pre-allocated input_ids buffer (maxSeqLen * 4 bytes). */\n  private inputIdsBuffer: GPUBuffer;\n\n  /**\n   * CPU-resident Per-Layer-Embeddings (PLE) source for Gemma 4. The PLE table\n   * (`embed_tokens_per_layer`, [vocab, L*256]) is ~1.17GB at 4-bit and is kept\n   * OFF the GPU. Each forward step we gather + dequantize only the rows for the\n   * actual input tokens and upload a tiny [T, L*256] f32 buffer. See setPleSource.\n   */\n  private pleSource: {\n    packed: Uint32Array; // flat row-major INT4 nibbles, 8 per u32\n    scales: Float32Array; // per-group scale (Gerbil convention)\n    zeros: Float32Array; // per-group zero\n    width: number; // L*256 (row width)\n    groupSize: number; // dequant group size\n    targetTensor: string; // activation tensor the gathered rows are written to\n    cache?: {\n      cacheName: string;\n      packedKey: string;\n      scalesKey: string;\n      zerosKey: string;\n      packedLen: number;\n    };\n  } | null = null;\n  /** Reusable scratch for the dequantized PLE rows (resized on demand). */\n  private pleScratch: Float32Array | null = null;\n  /**\n   * One-time promise that materializes a cache-backed PLE table into heap. The\n   * table is read from CacheStorage on the FIRST forward — i.e. AFTER the GPU\n   * weight upload has completed, so the ~1.17 GB does not stack on top of the\n   * upload's transient allocations at the load-time memory high-water mark.\n   */\n  private pleReady: Promise<void> | null = null;\n\n  /**\n   * Dummy GPU buffer bound to an otherwise-aliasing storage-read-write slot.\n   * Used by RoPE-Q-only nodes (Gemma 4 KV-shared layers) whose node lists the\n   * same tensor as input and output: the RoPE kernel always declares two\n   * read_write bindings (q, k), but with num_kv_heads=0 the k slot is never\n   * touched. WebGPU still rejects two read_write bindings aliasing one buffer,\n   * so we bind this throwaway buffer to the unused k slot. Allocated lazily.\n   */\n  private bindingScratchBuffer: GPUBuffer | null = null;\n  /** Readback buffer for logits. */\n  private logitsReadback: GPUBuffer;\n  /** GPU buffer for argmax result (1 u32). */\n  private argmaxResultBuffer: GPUBuffer;\n  /** Readback buffer for argmax result (1 u32). */\n  private argmaxReadback: GPUBuffer;\n  /** Readback ring for pipelined greedy decode (created lazily). */\n  private decodeReadbacks: GPUBuffer[] = [];\n\n  /**\n   * Staging buffer for uniform param updates.\n   * Safari/Metal has weaker visibility guarantees for queue.writeBuffer() to\n   * UNIFORM buffers — early writes get dropped when hundreds are queued.\n   * Instead, we pack all params into this STORAGE staging buffer (1 writeBuffer),\n   * then use encoder.copyBufferToBuffer to distribute to each uniform buffer.\n   * Copies are GPU-sequenced and guaranteed to complete before compute passes.\n   */\n  private uniformStagingBuffer: GPUBuffer | null = null;\n  private uniformStagingCapacity: number = 0;\n\n  /** Dispatch entries for prefill (M>1): uses tiled matmul kernels. */\n  private dispatchEntries: DispatchEntry[] = [];\n  /** Dispatch entries for decode (M=1): uses K-parallel matvec kernels. */\n  private decodeEntries: DispatchEntry[] = [];\n  /** Argmax dispatch entry (created in initBindGroups). */\n  private argmaxEntry: {\n    pipeline: GPUComputePipeline;\n    bindGroup: GPUBindGroup;\n    uniformBuffer: GPUBuffer;\n  } | null = null;\n\n  /** True when running on Safari/WebKit (needs multi-encoder submit). */\n  readonly needsMultiEncoder: boolean;\n\n  private maxSeqLen: number;\n  private kvMode: KvMode;\n  private seqPos: number = 0;\n  private webkitGroupSize: number;\n\n  // ── Env-gated decode profiler (GERBIL_PROFILE) ──\n  // Times each dispatch with timestamp-query (one compute pass per dispatch) and\n  // accumulates GPU nanoseconds per opType. OFF on the normal inference path.\n  private profileEnabled = false;\n  private readonly profileData = new Map<string, { ns: number; count: number }>();\n  private querySet: GPUQuerySet | null = null;\n  private queryResolveBuf: GPUBuffer | null = null;\n  private queryReadbackBuf: GPUBuffer | null = null;\n\n  constructor(ctx: GPUContext, graph: ModelGraph, options: ExecutorOptions) {\n    this.ctx = ctx;\n    this.graph = graph;\n    this.maxSeqLen = options.maxSeqLen;\n    this.kvMode = options.kvMode ?? \"native-f16\";\n    this.webkitGroupSize = Math.max(1, options.webkitGroupSize ?? 1);\n    this.profileEnabled =\n      ctx.hasTimestamp && typeof process !== \"undefined\" && process.env?.GERBIL_PROFILE != null;\n\n    // WebKit's WebGPU fails to make storage writes visible across dispatches\n    // within one submission at production scale (see docs/mobile-failure-diagnosis.md);\n    // Dawn (Chrome, node) does not need the multi-encoder workaround.\n    this.needsMultiEncoder = ctx.isWebKitWebGPU;\n\n    // Pre-allocate input_ids buffer at max size\n    this.inputIdsBuffer = createStorageBuffer(ctx, \"input_ids\", options.maxSeqLen * 4);\n\n    this.allocateActivationBuffers();\n    this.allocateSSMStateBuffers();\n    this.allocateKVCacheBuffers();\n\n    // Guard against vocab_size 0 (non-LM graphs, e.g. the NanoCodec decoder run\n    // via runGraphOutput): a zero-size buffer is invalid in WebGPU.\n    this.logitsReadback = createReadbackBuffer(\n      ctx,\n      \"logits_readback\",\n      Math.max(4, graph.config.vocab_size * 4),\n    );\n    this.argmaxResultBuffer = createStorageBuffer(ctx, \"argmax_result\", 4);\n    this.argmaxReadback = createReadbackBuffer(ctx, \"argmax_readback\", 4);\n  }\n\n  /**\n   * Register the CPU-resident Gemma 4 PLE table. The quantized table is kept in\n   * JS memory (NOT uploaded to a GPU buffer), and {@link forward} gathers +\n   * dequantizes only the rows for the current input tokens each step, uploading a\n   * small [T, width] f32 buffer into `targetTensor`. This is what keeps Gemma 4\n   * mobile-viable: resident GPU memory is just the active transformer weights.\n   */\n  setPleSource(src: {\n    packed: Uint32Array;\n    scales: Float32Array;\n    zeros: Float32Array;\n    width: number;\n    groupSize: number;\n    targetTensor: string;\n    cache?: {\n      cacheName: string;\n      packedKey: string;\n      scalesKey: string;\n      zerosKey: string;\n      packedLen: number;\n    };\n  }): void {\n    this.pleSource = src;\n  }\n\n  /**\n   * Materialize a cache-backed PLE table into heap exactly once. Deferred to the\n   * first forward (after GPU upload) so the ~1.17 GB table is not co-resident\n   * with the upload's transient allocations during the load-time memory peak.\n   */\n  private ensurePleLoaded(): Promise<void> {\n    const src = this.pleSource;\n    if (!src || !src.cache) return Promise.resolve();\n    if (this.pleReady) return this.pleReady;\n    const { cacheName, packedKey, scalesKey, zerosKey } = src.cache;\n    this.pleReady = (async () => {\n      const cache = await caches.open(cacheName);\n      const read = async (key: string): Promise<ArrayBuffer> => {\n        const resp = await cache.match(new Request(key));\n        if (!resp) throw new Error(`PLE cache entry missing: ${key}`);\n        return resp.arrayBuffer();\n      };\n      const [pBuf, sBuf, zBuf] = await Promise.all([\n        read(packedKey),\n        read(scalesKey),\n        read(zerosKey),\n      ]);\n      src.packed = new Uint32Array(pBuf);\n      src.scales = new Float32Array(sBuf);\n      src.zeros = new Float32Array(zBuf);\n    })();\n    return this.pleReady;\n  }\n\n  /**\n   * Gather + dequantize the PLE rows for `inputIds` and upload them into the\n   * target activation buffer. Touches only T rows (T*width floats) — the full\n   * [vocab, width] quantized table never goes to the GPU.\n   */\n  private async streamPleRows(inputIds: Uint32Array): Promise<void> {\n    const src = this.pleSource;\n    if (!src) return;\n    if (src.cache) await this.ensurePleLoaded();\n    const T = inputIds.length;\n    const { packed, scales, zeros, width, groupSize } = src;\n    const needed = T * width;\n    if (!this.pleScratch || this.pleScratch.length < needed) {\n      this.pleScratch = new Float32Array(needed);\n    }\n    const out = this.pleScratch;\n    for (let t = 0; t < T; t++) {\n      const tokenId = inputIds[t];\n      const rowBase = tokenId * width; // flat element offset of this row\n      const outBase = t * width;\n      for (let d = 0; d < width; d++) {\n        const flat = rowBase + d;\n        const packedIdx = flat >>> 3; // flat / 8\n        const nibblePos = (flat & 7) * 4; // (flat % 8) * 4\n        const nibble = (packed[packedIdx] >>> nibblePos) & 0xf;\n        const groupIdx = (flat / groupSize) | 0;\n        out[outBase + d] = (nibble - zeros[groupIdx]) * scales[groupIdx];\n      }\n    }\n    const buffer = this.getBuffer(src.targetTensor);\n    if (!buffer) {\n      throw new Error(`PLE target tensor \"${src.targetTensor}\" has no GPU buffer`);\n    }\n    this.ctx.device.queue.writeBuffer(buffer, 0, out.buffer, out.byteOffset, needed * 4);\n  }\n\n  /**\n   * Stream weights to the GPU one tensor at a time, pulling each from the\n   * `WeightSource` only when it is about to be uploaded and dropping the\n   * reference immediately afterward.\n   *\n   * This is the property that bounds peak JS heap: with a cache-backed source\n   * (browser/mobile), only ONE tensor's bytes are materialized in heap at a time\n   * (read from CacheStorage in `source.get()`), uploaded to its GPU buffer, then\n   * released before the next tensor is fetched. The whole model is never co-\n   * resident in heap. With a heap-backed source (Node/desktop), behavior matches\n   * the old Map path (the source deletes each entry as it is consumed).\n   *\n   * Accepts either a `WeightSource` (new, streamed/async) or a plain Map\n   * (back-compat for callers that build a Map directly, e.g. Kani/Moonshine).\n   */\n  async uploadWeights(\n    source: WeightSource | Map<string, { data: ArrayBufferView; shape: number[] }>,\n  ): Promise<void> {\n    if (source instanceof Map) {\n      this.uploadWeightsMap(source);\n      return;\n    }\n    for (const name of source.keys()) {\n      const entry = await source.get(name);\n      if (!entry) continue;\n      const { data } = entry;\n      const buffer = createStorageBuffer(this.ctx, `weight_${name}`, data.byteLength, data);\n      this.weightBuffers.set(name, buffer);\n      // Each `entry` reference goes out of scope here; with a cache-backed source\n      // its bytes can be GC'd before the next tensor is pulled — bounding peak heap.\n    }\n  }\n\n  /**\n   * Synchronous Map upload (Node/desktop and the Kani/Moonshine sub-executors,\n   * which build small heap Maps). Deletes each entry as it is consumed to free\n   * the JS-side bytes once they are GPU-resident. Safe to call from a constructor.\n   */\n  uploadWeightsMap(weights: Map<string, { data: ArrayBufferView; shape: number[] }>): void {\n    const names = [...weights.keys()];\n    for (const name of names) {\n      const entry = weights.get(name);\n      if (!entry) continue;\n      const buffer = createStorageBuffer(\n        this.ctx,\n        `weight_${name}`,\n        entry.data.byteLength,\n        entry.data,\n      );\n      this.weightBuffers.set(name, buffer);\n      weights.delete(name);\n    }\n  }\n\n  /**\n   * Build all pipelines and bind groups. Call ONCE after uploadWeights().\n   *\n   * Creates two dispatch entry arrays:\n   * - dispatchEntries: tiled matmul for prefill (any M)\n   * - decodeEntries: K-parallel matvec for decode (M=1)\n   */\n  initBindGroups(): void {\n    for (const nodeId of this.graph.executionOrder) {\n      const node = this.graph.nodes.find((n) => n.id === nodeId)!;\n      let spec = KERNEL_REGISTRY[node.opType];\n      if (!spec) throw new Error(`No kernel for op type: ${node.opType}`);\n\n      // Use f16 kernel variants when KV cache tensors are f16.\n      // \"packed-f16\" uses pack2x16float (Safari-safe), \"native-f16\" uses enable f16.\n      if (node.opType === \"KVCacheAppend\") {\n        const hasF16KV = node.outputs.some((out) => this.graph.tensors[out]?.dtype === \"f16\");\n        if (hasF16KV) {\n          spec =\n            this.kvMode === \"packed-f16\"\n              ? KV_CACHE_APPEND_PACKED_F16_SPEC\n              : KV_CACHE_APPEND_F16_SPEC;\n        }\n      } else if (node.opType === \"Attention\") {\n        const hasF16KV = node.inputs.some((inp) => this.graph.tensors[inp]?.dtype === \"f16\");\n        if (hasF16KV) {\n          spec = this.kvMode === \"packed-f16\" ? ATTENTION_PACKED_F16_SPEC : ATTENTION_F16_SPEC;\n        }\n      } else if (node.opType === \"MambaSSM\" && this.ctx.hasF16 && !this.ctx.isWebKitWebGPU) {\n        // f16 SSM state storage halves the dominant state bandwidth.\n        // Dawn only — WebKit keeps the f32 kernel (unverified there).\n        spec = MAMBA_SSM_F16_SPEC;\n      } else if (node.opType === \"LayerNorm\" && node.inputs.length === 2) {\n        // Weight-only LayerNorm (Moonshine encoder/decoder norms have no bias):\n        // the node binds [input, weight] only, so use the no-bias variant whose\n        // binding layout matches.\n        spec = LAYERNORM_NO_BIAS_SPEC;\n      } else if (node.opType === \"RoPE\" && node.attributes.interleaved === true) {\n        // Moonshine rotates adjacent dim pairs (2p, 2p+1), not the [p, p+half]\n        // split the default kernel uses.\n        spec = ROPE_INTERLEAVED_SPEC;\n      }\n\n      const pipeline = getOrCreatePipeline(\n        this.ctx,\n        `kernel_${nodeId}`,\n        spec.shaderCode,\n        spec.entryPoint,\n      );\n\n      // Create uniform buffer with dummy params (will be overwritten before first dispatch)\n      const dummyShapes = this.resolveShapes(1);\n      const paramsData = spec.buildParams(node, dummyShapes, { seqPos: 0 });\n      const uniformBuffer = createUniformBuffer(this.ctx, `uniform_${nodeId}`, paramsData);\n\n      // Gather buffers for bind group\n      const bufferEntries = this.gatherBuffers(spec, node, uniformBuffer);\n      const bindGroup = createBindGroup(this.ctx, pipeline, bufferEntries, `bg_${nodeId}`);\n\n      const prefillEntry: DispatchEntry = {\n        nodeId,\n        node,\n        spec,\n        pipeline,\n        bindGroup,\n        uniformBuffer,\n        lastParamsBytes: null,\n        lastDispatchSize: null,\n      };\n      this.dispatchEntries.push(prefillEntry);\n\n      // For matmul ops, create a separate decode entry with the matvec kernel\n      const isMatmul = node.opType === \"MatMul\" || node.opType === \"MatMulInt4\";\n      if (isMatmul) {\n        // Subgroups-accelerated matvec reduction. GATED OFF by default: on Dawn\n        // (M-series desktop) it measured a large DECODE REGRESSION (~210→47 tok/s\n        // Qwen3.5-0.8B, ~593→146 tok/s LFM2.5-350M) — coherent output, but the\n        // extra workgroup barriers + per-column subgroup combine outweigh the\n        // tiny saved 16/32-element sequential reduction these matvecs do. The\n        // variant is kept registered and selectable behind an opt-in flag for\n        // future hardware/driver re-evaluation; default path is the portable\n        // kernel. WebKit is excluded regardless (mobile-safety rules).\n        const subgroupsOptIn =\n          (globalThis as { process?: { env?: Record<string, string | undefined> } }).process?.env\n            ?.GERBIL_MATVEC_SUBGROUPS === \"1\";\n        const useSubgroups = subgroupsOptIn && this.ctx.hasSubgroups && !this.ctx.isWebKitWebGPU;\n        const mvSpec =\n          node.opType === \"MatMulInt4\"\n            ? useSubgroups\n              ? MATVEC_INT4_SUBGROUPS_SPEC\n              : MATVEC_INT4_SPEC\n            : useSubgroups\n              ? MATVEC_SUBGROUPS_SPEC\n              : MATVEC_SPEC;\n\n        const mvPipeline = getOrCreatePipeline(\n          this.ctx,\n          `matvec_${nodeId}`,\n          mvSpec.shaderCode,\n          mvSpec.entryPoint,\n        );\n        const mvUniformData = mvSpec.buildParams(node, dummyShapes, { seqPos: 0 });\n        const mvUniform = createUniformBuffer(this.ctx, `mv_uniform_${nodeId}`, mvUniformData);\n        const mvBuffers = this.gatherBuffers(mvSpec, node, mvUniform);\n        const mvBindGroup = createBindGroup(this.ctx, mvPipeline, mvBuffers, `bg_mv_${nodeId}`);\n\n        this.decodeEntries.push({\n          nodeId,\n          node,\n          spec: mvSpec,\n          pipeline: mvPipeline,\n          bindGroup: mvBindGroup,\n          uniformBuffer: mvUniform,\n          lastParamsBytes: null,\n          lastDispatchSize: null,\n        });\n      } else {\n        // Non-matmul ops share the same entry for decode\n        this.decodeEntries.push(prefillEntry);\n      }\n    }\n\n    // ── Fuse gate_proj + up_proj + SwiGLU → single SwiGLUMatVec dispatch ──\n    // Scans decode entries for the pattern: MatMul(input→gate) + MatMul(input→up) + SwiGLU(gate,up→out)\n    // Replaces 3 dispatches with 1 fused kernel that reads input once.\n    this.fuseSwiGLUDecodeEntries();\n\n    // ── Fuse adjacent INT4 projections sharing the same input (q+gate, k+v) ──\n    // Full-attention decode runs q_proj/gate_proj/k_proj/v_proj all reading the\n    // same input_layernorm output. Merge the q+gate pair and the k+v pair into\n    // one DualMatVecInt4 dispatch each, removing one GPU round-trip per pair.\n    this.fuseDualMatVecDecodeEntries();\n\n    // ── Fuse adjacent K+V cache appends (full-attention layers) ──\n    this.fuseDualKVCacheAppendEntries();\n\n    // ── Fuse attention SigmoidGate into the INT4 o_proj that consumes it ──\n    this.fuseGatedOProjDecodeEntries();\n\n    // ── Fuse Mamba SwiGLU (silu(z)*norm) into the INT4 out_proj ──\n    this.fuseSwiGLUGatedProjDecodeEntries();\n\n    // ── Fuse adjacent per-head RMSNorms (q_norm + k_norm) ──\n    this.fuseDualRMSNormDecodeEntries();\n\n    // Decode dispatch count per token — the mobile-proportional metric. On WebKit\n    // each dispatch is a submit+drain round-trip, so decode tok/s ≈ 1/this. Logged\n    // so every fusion's dispatch reduction is measurable on desktop/node.\n    console.log(`[executor] decode: ${this.decodeEntries.length} dispatches/token`);\n\n    // Create argmax bind group\n    const logitsBuffer = this.getBuffer(\"logits\");\n    if (logitsBuffer) {\n      const argmaxPipeline = getOrCreatePipeline(\n        this.ctx,\n        \"argmax\",\n        ARGMAX_SPEC.shaderCode,\n        ARGMAX_SPEC.entryPoint,\n      );\n      const argmaxUniform = createUniformBuffer(\n        this.ctx,\n        \"uniform_argmax\",\n        new ArrayBuffer(16), // {count, offset} — will be overwritten\n      );\n      const argmaxBindGroup = createBindGroup(\n        this.ctx,\n        argmaxPipeline,\n        [{ buffer: logitsBuffer }, { buffer: this.argmaxResultBuffer }, { buffer: argmaxUniform }],\n        \"bg_argmax\",\n      );\n      this.argmaxEntry = {\n        pipeline: argmaxPipeline,\n        bindGroup: argmaxBindGroup,\n        uniformBuffer: argmaxUniform,\n      };\n    }\n\n    if (this.needsMultiEncoder) {\n      console.log(`[executor] WebKit WebGPU detected — using per-dispatch submit`);\n    }\n\n    // Create staging buffer for uniform param updates.\n    // Size = max(total prefill uniform bytes, total decode uniform bytes + argmax).\n    let prefillTotal = 0;\n    for (const entry of this.dispatchEntries) {\n      prefillTotal += Math.ceil(entry.uniformBuffer.size / 4) * 4;\n    }\n    let decodeTotal = 0;\n    for (const entry of this.decodeEntries) {\n      decodeTotal += Math.ceil(entry.uniformBuffer.size / 4) * 4;\n    }\n    decodeTotal += 16; // argmax params (8 bytes, padded)\n    // Add space for input_ids (routed through staging too, for Safari safety)\n    const inputIdBytes = this.maxSeqLen * 4;\n    this.uniformStagingCapacity = Math.max(prefillTotal, decodeTotal) + inputIdBytes;\n    this.uniformStagingBuffer = this.ctx.device.createBuffer({\n      label: \"uniform_staging\",\n      size: this.uniformStagingCapacity,\n      usage: 0x0004 | 0x0008, // COPY_SRC | COPY_DST\n    });\n  }\n\n  /**\n   * Run a forward pass. Uses matvec kernels for M=1 (decode), tiled for M>1 (prefill).\n   */\n  async forward(inputIds: Uint32Array): Promise<ForwardResult> {\n    const T = inputIds.length;\n\n    const resolvedShapes = this.resolveShapes(T);\n    const runtimeContext: RuntimeContext = { seqPos: this.seqPos };\n\n    // Use matvec entries for single-token decode, tiled for prefill\n    const entries = T === 1 ? this.decodeEntries : this.dispatchEntries;\n\n    // Write input_ids directly — no staging buffer.\n    this.ctx.device.queue.writeBuffer(this.inputIdsBuffer, 0, inputIds as BufferSource);\n\n    // Gemma 4: stream the CPU-resident PLE rows for these tokens (no-op otherwise).\n    await this.streamPleRows(inputIds);\n\n    // Build params and write directly to each uniform buffer.\n    // Also compute dispatch sizes up front.\n    const dispatchSizes: Array<[number, number, number]> = new Array(entries.length);\n    for (let i = 0; i < entries.length; i++) {\n      const entry = entries[i];\n      const paramsData = entry.spec.buildParams(entry.node, resolvedShapes, runtimeContext);\n      const paramsView = new Uint8Array(paramsData);\n\n      let changed = !entry.lastParamsBytes || entry.lastParamsBytes.length !== paramsView.length;\n      if (!changed) {\n        const cached = entry.lastParamsBytes!;\n        for (let j = 0; j < paramsView.length; j++) {\n          if (paramsView[j] !== cached[j]) {\n            changed = true;\n            break;\n          }\n        }\n      }\n\n      if (changed) {\n        this.ctx.device.queue.writeBuffer(entry.uniformBuffer, 0, paramsData);\n        entry.lastParamsBytes = new Uint8Array(paramsView.slice(0));\n      }\n\n      dispatchSizes[i] = entry.spec.getDispatchSize(entry.node, resolvedShapes, runtimeContext);\n    }\n\n    if (this.profileEnabled) {\n      await this.runProfiledDispatches(entries, dispatchSizes);\n    } else if (this.needsMultiEncoder) {\n      // WebKit: group dispatches into command buffers of webkitGroupSize\n      // (one compute pass per dispatch), with at most ONE command buffer in\n      // flight (awaited). Fire-and-forget per-dispatch submits queue ~400\n      // unbounded in-flight Metal command buffers per token — the documented\n      // WebKit resource-exhaustion anti-pattern (WebKit bug 311598). Batching\n      // many dispatches into one submission zeroed activations on iPad\n      // (docs/metal-safari-intel.md); sweep ?group=N to find this WebKit\n      // version's correct/fast granularity (llama.cpp runs ~64/CB on iOS 26.4).\n      const group = this.webkitGroupSize;\n      for (let start = 0; start < entries.length; start += group) {\n        const enc = this.ctx.device.createCommandEncoder();\n        const end = Math.min(start + group, entries.length);\n        for (let i = start; i < end; i++) {\n          const p = enc.beginComputePass();\n          p.setPipeline(entries[i].pipeline);\n          p.setBindGroup(0, entries[i].bindGroup);\n          p.dispatchWorkgroups(...dispatchSizes[i]);\n          p.end();\n        }\n        this.ctx.device.queue.submit([enc.finish()]);\n        await this.ctx.device.queue.onSubmittedWorkDone();\n      }\n      const logitsBuffer = this.getBuffer(\"logits\");\n      if (logitsBuffer) {\n        const copyEnc = this.ctx.device.createCommandEncoder();\n        const vocabSize = this.graph.config.vocab_size;\n        // logits is [1, vocab] — SliceLastRow feeds lm_head only the last position\n        copyEnc.copyBufferToBuffer(logitsBuffer, 0, this.logitsReadback, 0, vocabSize * 4);\n        this.ctx.device.queue.submit([copyEnc.finish()]);\n      }\n    } else {\n      // Desktop/Dawn: single encoder, single compute pass\n      const encoder = this.ctx.device.createCommandEncoder({ label: \"forward\" });\n      const pass = encoder.beginComputePass({ label: \"fwd_pass\" });\n      for (let i = 0; i < entries.length; i++) {\n        pass.setPipeline(entries[i].pipeline);\n        pass.setBindGroup(0, entries[i].bindGroup);\n        pass.dispatchWorkgroups(...dispatchSizes[i]);\n      }\n      pass.end();\n      const logitsBuffer = this.getBuffer(\"logits\");\n      if (logitsBuffer) {\n        const vocabSize = this.graph.config.vocab_size;\n        // logits is [1, vocab] — SliceLastRow feeds lm_head only the last position\n        encoder.copyBufferToBuffer(logitsBuffer, 0, this.logitsReadback, 0, vocabSize * 4);\n      }\n      this.ctx.device.queue.submit([encoder.finish()]);\n    }\n\n    const vocabSize = this.graph.config.vocab_size;\n    await this.logitsReadback.mapAsync(MAP_MODE_READ, 0, vocabSize * 4);\n    const mapped = this.logitsReadback.getMappedRange(0, vocabSize * 4);\n    const logits = new Float32Array(mapped.slice(0));\n    this.logitsReadback.unmap();\n\n    this.seqPos += T;\n    return { logits };\n  }\n\n  /**\n   * Profiling variant of the desktop dispatch path: one compute pass per dispatch,\n   * each bracketed by timestamp queries, so we get per-op GPU time. Accumulates\n   * into profileData by opType. Only runs under GERBIL_PROFILE with timestamp-query\n   * support — slower than the batched path (it measures relative cost, not tok/s).\n   */\n  private async runProfiledDispatches(\n    entries: DispatchEntry[],\n    dispatchSizes: number[][],\n    copyLogits = true,\n  ): Promise<void> {\n    const n = entries.length;\n    const queryCount = n * 2;\n    if (!this.querySet) {\n      this.querySet = this.ctx.device.createQuerySet({ type: \"timestamp\", count: queryCount });\n      // QUERY_RESOLVE (0x0200) | COPY_SRC (0x0004)\n      this.queryResolveBuf = this.ctx.device.createBuffer({\n        size: queryCount * 8,\n        usage: 0x0200 | 0x0004,\n      });\n      // MAP_READ (0x0001) | COPY_DST (0x0008)\n      this.queryReadbackBuf = this.ctx.device.createBuffer({\n        size: queryCount * 8,\n        usage: 0x0001 | 0x0008,\n      });\n    }\n    const querySet = this.querySet;\n    const resolveBuf = this.queryResolveBuf as GPUBuffer;\n    const readbackBuf = this.queryReadbackBuf as GPUBuffer;\n\n    const encoder = this.ctx.device.createCommandEncoder({ label: \"fwd_profiled\" });\n    for (let i = 0; i < n; i++) {\n      const pass = encoder.beginComputePass({\n        timestampWrites: {\n          querySet,\n          beginningOfPassWriteIndex: i * 2,\n          endOfPassWriteIndex: i * 2 + 1,\n        },\n      });\n      pass.setPipeline(entries[i].pipeline);\n      pass.setBindGroup(0, entries[i].bindGroup);\n      pass.dispatchWorkgroups(dispatchSizes[i][0], dispatchSizes[i][1], dispatchSizes[i][2]);\n      pass.end();\n    }\n    if (copyLogits) {\n      const logitsBuffer = this.getBuffer(\"logits\");\n      const vocabSize = this.graph.config.vocab_size;\n      if (logitsBuffer) {\n        encoder.copyBufferToBuffer(logitsBuffer, 0, this.logitsReadback, 0, vocabSize * 4);\n      }\n    }\n    encoder.resolveQuerySet(querySet, 0, queryCount, resolveBuf, 0);\n    encoder.copyBufferToBuffer(resolveBuf, 0, readbackBuf, 0, queryCount * 8);\n    this.ctx.device.queue.submit([encoder.finish()]);\n\n    await readbackBuf.mapAsync(MAP_MODE_READ, 0, queryCount * 8);\n    const ts = new BigUint64Array(readbackBuf.getMappedRange(0, queryCount * 8).slice(0));\n    readbackBuf.unmap();\n    for (let i = 0; i < n; i++) {\n      const ns = Number(ts[i * 2 + 1] - ts[i * 2]);\n      if (ns <= 0) {\n        continue;\n      }\n      const key = entries[i].node.opType;\n      const cur = this.profileData.get(key) ?? { ns: 0, count: 0 };\n      cur.ns += ns;\n      cur.count += 1;\n      this.profileData.set(key, cur);\n    }\n  }\n\n  /** Per-opType GPU time (ns) + dispatch count accumulated by GERBIL_PROFILE, hottest first. */\n  getProfile(): Array<{ opType: string; ns: number; count: number }> {\n    return [...this.profileData.entries()]\n      .map(([opType, v]) => ({ opType, ns: v.ns, count: v.count }))\n      .sort((a, b) => b.ns - a.ns);\n  }\n\n  /** Clear accumulated profiler data (e.g. to drop warm-up tokens). */\n  resetProfile(): void {\n    this.profileData.clear();\n  }\n\n  /** GPU dispatches per decode token (post-fusion). On mobile this drives the\n   * submit-group count = ceil(dispatchCount / webkitGroupSize). */\n  get decodeDispatchCount(): number {\n    return this.decodeEntries.length;\n  }\n\n  /** Device limit that gates the INT4 projection fusions (they need ≥9). If a\n   * device caps at 8 the dual/gated/swiglu-gated INT4 fusions silently fall back,\n   * inflating the decode dispatch count. */\n  get maxStorageBuffers(): number {\n    return this.ctx.limits.maxStorageBuffersPerShaderStage;\n  }\n\n  /**\n   * Profile ONE real decode step: times the actual `decodeEntries` (the kernels\n   * the pipelined greedy benchmark runs) with per-dispatch timestamps. Timing is\n   * token-independent, so pass any valid id; runs un-pipelined with a synchronous\n   * timestamp readback (measurement only, not for production decode). Argmax (one\n   * tiny dispatch) is intentionally excluded — it is not a hotspot target.\n   */\n  async profileDecodeStep(tokenId: number): Promise<void> {\n    const resolvedShapes = this.resolveShapes(1);\n    const runtimeContext: RuntimeContext = { seqPos: this.seqPos };\n    this.ctx.device.queue.writeBuffer(this.inputIdsBuffer, 0, new Uint32Array([tokenId]));\n\n    const dispatchSizes: number[][] = new Array(this.decodeEntries.length);\n    for (let i = 0; i < this.decodeEntries.length; i++) {\n      const entry = this.decodeEntries[i];\n      const paramsData = entry.spec.buildParams(entry.node, resolvedShapes, runtimeContext);\n      this.ctx.device.queue.writeBuffer(entry.uniformBuffer, 0, paramsData);\n      dispatchSizes[i] = entry.spec.getDispatchSize(entry.node, resolvedShapes, runtimeContext);\n    }\n    await this.runProfiledDispatches(this.decodeEntries, dispatchSizes, false);\n    this.seqPos += 1;\n  }\n\n  /**\n   * Run a single forward pass over `inputIds` and read back the L2-normalized\n   * embedding vector. Requires an embedding graph (one whose output tensor is\n   * \"embedding\", produced by the last-token-pool + L2-norm tail).\n   *\n   * Always runs in a fresh-state single pass (caller should reset() first):\n   * embeddings are non-autoregressive, so the whole sequence is one prefill.\n   */\n  async embed(inputIds: Uint32Array): Promise<Float32Array> {\n    const embeddingBuffer = this.getBuffer(\"embedding\");\n    if (!embeddingBuffer) {\n      throw new Error(\n        \"embed() requires an embedding graph (no 'embedding' output tensor found). \" +\n          \"Load the model with { embedding: true }.\",\n      );\n    }\n    const hiddenSize = this.graph.config.hidden_size;\n    const byteLen = hiddenSize * 4;\n\n    const T = inputIds.length;\n    const resolvedShapes = this.resolveShapes(T);\n    const runtimeContext: RuntimeContext = { seqPos: this.seqPos };\n\n    // Embeddings process the full sequence at once → always use prefill entries.\n    const entries = T === 1 ? this.decodeEntries : this.dispatchEntries;\n\n    this.ctx.device.queue.writeBuffer(this.inputIdsBuffer, 0, inputIds as BufferSource);\n\n    const dispatchSizes: Array<[number, number, number]> = new Array(entries.length);\n    for (let i = 0; i < entries.length; i++) {\n      const entry = entries[i];\n      const paramsData = entry.spec.buildParams(entry.node, resolvedShapes, runtimeContext);\n      const paramsView = new Uint8Array(paramsData);\n\n      let changed = !entry.lastParamsBytes || entry.lastParamsBytes.length !== paramsView.length;\n      if (!changed) {\n        const cached = entry.lastParamsBytes!;\n        for (let j = 0; j < paramsView.length; j++) {\n          if (paramsView[j] !== cached[j]) {\n            changed = true;\n            break;\n          }\n        }\n      }\n      if (changed) {\n        this.ctx.device.queue.writeBuffer(entry.uniformBuffer, 0, paramsData);\n        entry.lastParamsBytes = new Uint8Array(paramsView.slice(0));\n      }\n\n      dispatchSizes[i] = entry.spec.getDispatchSize(entry.node, resolvedShapes, runtimeContext);\n    }\n\n    // Dedicated readback buffer (COPY_DST | MAP_READ).\n    const readback = this.ctx.device.createBuffer({ size: byteLen, usage: 0x0001 | 0x0008 });\n\n    if (this.needsMultiEncoder) {\n      const group = this.webkitGroupSize;\n      for (let start = 0; start < entries.length; start += group) {\n        const enc = this.ctx.device.createCommandEncoder();\n        const end = Math.min(start + group, entries.length);\n        for (let i = start; i < end; i++) {\n          const p = enc.beginComputePass();\n          p.setPipeline(entries[i].pipeline);\n          p.setBindGroup(0, entries[i].bindGroup);\n          p.dispatchWorkgroups(...dispatchSizes[i]);\n          p.end();\n        }\n        this.ctx.device.queue.submit([enc.finish()]);\n        await this.ctx.device.queue.onSubmittedWorkDone();\n      }\n      const copyEnc = this.ctx.device.createCommandEncoder();\n      copyEnc.copyBufferToBuffer(embeddingBuffer, 0, readback, 0, byteLen);\n      this.ctx.device.queue.submit([copyEnc.finish()]);\n    } else {\n      const encoder = this.ctx.device.createCommandEncoder({ label: \"embed\" });\n      const pass = encoder.beginComputePass({ label: \"embed_pass\" });\n      for (let i = 0; i < entries.length; i++) {\n        pass.setPipeline(entries[i].pipeline);\n        pass.setBindGroup(0, entries[i].bindGroup);\n        pass.dispatchWorkgroups(...dispatchSizes[i]);\n      }\n      pass.end();\n      encoder.copyBufferToBuffer(embeddingBuffer, 0, readback, 0, byteLen);\n      this.ctx.device.queue.submit([encoder.finish()]);\n    }\n\n    await readback.mapAsync(MAP_MODE_READ, 0, byteLen);\n    const out = new Float32Array(readback.getMappedRange(0, byteLen).slice(0));\n    readback.unmap();\n    readback.destroy();\n\n    this.seqPos += T;\n    return out;\n  }\n\n  /**\n   * Generic one-shot dispatch for a non-LM graph: run every entry once over a\n   * single forward and read back `elemCount` f32 elements of the named output\n   * tensor. Used to execute the NanoCodec decoder graph (codes→PCM) — its ops use\n   * concrete lengths and it has no \"logits\" output, so the normal forward()\n   * logits-readback path does not apply. Caller writes any inputs (e.g. audio_codes)\n   * via writeInput() and reset()s first.\n   */\n  async runGraphOutput(outputName: string, elemCount: number): Promise<Float32Array> {\n    const outBuffer = this.getBuffer(outputName);\n    if (!outBuffer) {\n      throw new Error(`runGraphOutput: no GPU buffer for output tensor \"${outputName}\".`);\n    }\n    const byteLen = elemCount * 4;\n    // Codec graphs carry concrete lengths; dispatchEntries (M>1 path) builds the\n    // full pipeline. Pass T=1 so shape resolution keeps any symbolic dims minimal\n    // (the codec ops ignore input_ids entirely).\n    const resolvedShapes = this.resolveShapes(1);\n    const runtimeContext: RuntimeContext = { seqPos: this.seqPos };\n    const entries = this.dispatchEntries;\n\n    const dispatchSizes: Array<[number, number, number]> = new Array(entries.length);\n    for (let i = 0; i < entries.length; i++) {\n      const entry = entries[i];\n      const paramsData = entry.spec.buildParams(entry.node, resolvedShapes, runtimeContext);\n      this.ctx.device.queue.writeBuffer(entry.uniformBuffer, 0, paramsData);\n      entry.lastParamsBytes = new Uint8Array(new Uint8Array(paramsData).slice(0));\n      dispatchSizes[i] = entry.spec.getDispatchSize(entry.node, resolvedShapes, runtimeContext);\n    }\n\n    const readback = this.ctx.device.createBuffer({ size: byteLen, usage: 0x0001 | 0x0008 });\n\n    if (this.needsMultiEncoder) {\n      const group = this.webkitGroupSize;\n      for (let start = 0; start < entries.length; start += group) {\n        const enc = this.ctx.device.createCommandEncoder();\n        const end = Math.min(start + group, entries.length);\n        for (let i = start; i < end; i++) {\n          const pass = enc.beginComputePass();\n          pass.setPipeline(entries[i].pipeline);\n          pass.setBindGroup(0, entries[i].bindGroup);\n          pass.dispatchWorkgroups(...dispatchSizes[i]);\n          pass.end();\n        }\n        this.ctx.device.queue.submit([enc.finish()]);\n        await this.ctx.device.queue.onSubmittedWorkDone();\n      }\n      const copyEnc = this.ctx.device.createCommandEncoder();\n      copyEnc.copyBufferToBuffer(outBuffer, 0, readback, 0, byteLen);\n      this.ctx.device.queue.submit([copyEnc.finish()]);\n    } else {\n      const encoder = this.ctx.device.createCommandEncoder({ label: \"runGraphOutput\" });\n      const pass = encoder.beginComputePass({ label: \"run_pass\" });\n      for (let i = 0; i < entries.length; i++) {\n        pass.setPipeline(entries[i].pipeline);\n        pass.setBindGroup(0, entries[i].bindGroup);\n        pass.dispatchWorkgroups(...dispatchSizes[i]);\n      }\n      pass.end();\n      encoder.copyBufferToBuffer(outBuffer, 0, readback, 0, byteLen);\n      this.ctx.device.queue.submit([encoder.finish()]);\n    }\n\n    await readback.mapAsync(MAP_MODE_READ, 0, byteLen);\n    const out = new Float32Array(readback.getMappedRange(0, byteLen).slice(0));\n    readback.unmap();\n    readback.destroy();\n    return out;\n  }\n\n  /**\n   * Greedy decode step: forward + GPU argmax. Returns token ID directly.\n   * Always uses matvec kernels (M=1). Reads back 4 bytes instead of vocab_size*4.\n   */\n  async forwardArgmax(inputIds: Uint32Array): Promise<number> {\n    const T = inputIds.length;\n\n    const resolvedShapes = this.resolveShapes(T);\n    const runtimeContext: RuntimeContext = { seqPos: this.seqPos };\n\n    // Write input_ids directly\n    this.ctx.device.queue.writeBuffer(this.inputIdsBuffer, 0, inputIds as BufferSource);\n\n    // Write uniform params directly to each buffer\n    const argmaxDispatchSizes: Array<[number, number, number]> = new Array(\n      this.decodeEntries.length,\n    );\n    for (let i = 0; i < this.decodeEntries.length; i++) {\n      const entry = this.decodeEntries[i];\n      const paramsData = entry.spec.buildParams(entry.node, resolvedShapes, runtimeContext);\n      const paramsView = new Uint8Array(paramsData);\n\n      let changed = !entry.lastParamsBytes || entry.lastParamsBytes.length !== paramsView.length;\n      if (!changed) {\n        const cached = entry.lastParamsBytes!;\n        for (let j = 0; j < paramsView.length; j++) {\n          if (paramsView[j] !== cached[j]) {\n            changed = true;\n            break;\n          }\n        }\n      }\n\n      if (changed) {\n        this.ctx.device.queue.writeBuffer(entry.uniformBuffer, 0, paramsData);\n        entry.lastParamsBytes = new Uint8Array(paramsView.slice(0));\n      }\n\n      argmaxDispatchSizes[i] = entry.spec.getDispatchSize(\n        entry.node,\n        resolvedShapes,\n        runtimeContext,\n      );\n    }\n\n    // Write argmax params directly.\n    // logits is [1, vocab] (SliceLastRow feeds lm_head only the last position),\n    // so the argmax always reads row 0.\n    if (this.argmaxEntry) {\n      const vocabSize = this.graph.config.vocab_size;\n      const argmaxParams = new ArrayBuffer(8);\n      const view = new DataView(argmaxParams);\n      view.setUint32(0, vocabSize, true);\n      view.setUint32(4, 0, true);\n      this.ctx.device.queue.writeBuffer(this.argmaxEntry.uniformBuffer, 0, argmaxParams);\n    }\n\n    if (this.needsMultiEncoder) {\n      // WebKit: grouped submits with one command buffer in flight\n      // (see forward() comment for rationale)\n      const group = this.webkitGroupSize;\n      for (let start = 0; start < this.decodeEntries.length; start += group) {\n        const enc = this.ctx.device.createCommandEncoder();\n        const end = Math.min(start + group, this.decodeEntries.length);\n        for (let i = start; i < end; i++) {\n          const p = enc.beginComputePass();\n          p.setPipeline(this.decodeEntries[i].pipeline);\n          p.setBindGroup(0, this.decodeEntries[i].bindGroup);\n          p.dispatchWorkgroups(...argmaxDispatchSizes[i]);\n          p.end();\n        }\n        this.ctx.device.queue.submit([enc.finish()]);\n        await this.ctx.device.queue.onSubmittedWorkDone();\n      }\n      // Argmax dispatch and readback copy in SEPARATE submits, mirroring\n      // forward(): the copy reads argmaxResultBuffer written by the dispatch,\n      // so it must sit behind a command buffer boundary on WebKit.\n      if (this.argmaxEntry) {\n        const enc = this.ctx.device.createCommandEncoder();\n        const p = enc.beginComputePass();\n        p.setPipeline(this.argmaxEntry.pipeline);\n        p.setBindGroup(0, this.argmaxEntry.bindGroup);\n        p.dispatchWorkgroups(1, 1, 1);\n        p.end();\n        this.ctx.device.queue.submit([enc.finish()]);\n\n        const copyEnc = this.ctx.device.createCommandEncoder();\n        copyEnc.copyBufferToBuffer(this.argmaxResultBuffer, 0, this.argmaxReadback, 0, 4);\n        this.ctx.device.queue.submit([copyEnc.finish()]);\n      }\n    } else {\n      // Desktop/Dawn: single encoder, single compute pass\n      const encoder = this.ctx.device.createCommandEncoder({ label: \"argmax\" });\n      const pass = encoder.beginComputePass({ label: \"am_pass\" });\n      for (let i = 0; i < this.decodeEntries.length; i++) {\n        const entry = this.decodeEntries[i];\n        pass.setPipeline(entry.pipeline);\n        pass.setBindGroup(0, entry.bindGroup);\n        pass.dispatchWorkgroups(...argmaxDispatchSizes[i]);\n      }\n      if (this.argmaxEntry) {\n        pass.setPipeline(this.argmaxEntry.pipeline);\n        pass.setBindGroup(0, this.argmaxEntry.bindGroup);\n        pass.dispatchWorkgroups(1, 1, 1);\n      }\n      pass.end();\n      encoder.copyBufferToBuffer(this.argmaxResultBuffer, 0, this.argmaxReadback, 0, 4);\n      this.ctx.device.queue.submit([encoder.finish()]);\n    }\n\n    await this.argmaxReadback.mapAsync(MAP_MODE_READ, 0, 4);\n    const mapped = this.argmaxReadback.getMappedRange(0, 4);\n    const tokenId = new Uint32Array(mapped.slice(0))[0];\n    this.argmaxReadback.unmap();\n\n    this.seqPos += T;\n    return tokenId;\n  }\n\n  /** KV-cache positions still available for decode steps. */\n  decodeCapacityRemaining(): number {\n    return this.maxSeqLen - this.seqPos;\n  }\n\n  /** Number of decode steps that may be in flight in the pipelined path. */\n  static readonly PIPELINE_DEPTH = 2;\n\n  /**\n   * Pipelined greedy decode step (Dawn only — WebKit uses forwardArgmax).\n   *\n   * Encodes one full decode forward + argmax and submits WITHOUT awaiting\n   * completion. The input token is taken from `tokenId` for the first step\n   * after prefill; for subsequent steps (tokenId === null) the previous step's\n   * argmax result is copied into input_ids ON THE GPU, so the decode loop\n   * never blocks on a readback before submitting the next step. The argmax\n   * result is copied to a per-slot readback buffer, read later (one step\n   * behind) via readDecodeToken(slot).\n   *\n   * queue.writeBuffer is queue-ordered: uniform updates land after the\n   * previously submitted step's command buffer and before this one's, so\n   * shared uniform buffers are safe with multiple steps in flight.\n   */\n  submitGreedyDecodeStep(tokenId: number | null, slot: number): void {\n    const resolvedShapes = this.resolveShapes(1);\n    const runtimeContext: RuntimeContext = { seqPos: this.seqPos };\n\n    if (this.decodeReadbacks.length === 0) {\n      for (let i = 0; i < Executor.PIPELINE_DEPTH; i++) {\n        this.decodeReadbacks.push(createReadbackBuffer(this.ctx, `decode_readback_${i}`, 4));\n      }\n    }\n\n    if (tokenId !== null) {\n      this.ctx.device.queue.writeBuffer(this.inputIdsBuffer, 0, new Uint32Array([tokenId]));\n      // Argmax params are decode-invariant — write once per generation.\n      if (this.argmaxEntry) {\n        const argmaxParams = new ArrayBuffer(8);\n        const view = new DataView(argmaxParams);\n        view.setUint32(0, this.graph.config.vocab_size, true);\n        view.setUint32(4, 0, true);\n        this.ctx.device.queue.writeBuffer(this.argmaxEntry.uniformBuffer, 0, argmaxParams);\n      }\n    }\n\n    // Update uniforms (same skip-unchanged caching as forwardArgmax)\n    const dispatchSizes: Array<[number, number, number]> = new Array(this.decodeEntries.length);\n    for (let i = 0; i < this.decodeEntries.length; i++) {\n      const entry = this.decodeEntries[i];\n      const paramsData = entry.spec.buildParams(entry.node, resolvedShapes, runtimeContext);\n      const paramsView = new Uint8Array(paramsData);\n\n      let changed = !entry.lastParamsBytes || entry.lastParamsBytes.length !== paramsView.length;\n      if (!changed) {\n        const cached = entry.lastParamsBytes!;\n        for (let j = 0; j < paramsView.length; j++) {\n          if (paramsView[j] !== cached[j]) {\n            changed = true;\n            break;\n          }\n        }\n      }\n\n      if (changed) {\n        this.ctx.device.queue.writeBuffer(entry.uniformBuffer, 0, paramsData);\n        entry.lastParamsBytes = new Uint8Array(paramsView.slice(0));\n      }\n\n      dispatchSizes[i] = entry.spec.getDispatchSize(entry.node, resolvedShapes, runtimeContext);\n    }\n\n    const encoder = this.ctx.device.createCommandEncoder({ label: \"decode_step\" });\n    if (tokenId === null) {\n      // Chain: previous step's argmax result becomes this step's input token.\n      encoder.copyBufferToBuffer(this.argmaxResultBuffer, 0, this.inputIdsBuffer, 0, 4);\n    }\n    const pass = encoder.beginComputePass({ label: \"decode_pass\" });\n    for (let i = 0; i < this.decodeEntries.length; i++) {\n      const entry = this.decodeEntries[i];\n      pass.setPipeline(entry.pipeline);\n      pass.setBindGroup(0, entry.bindGroup);\n      pass.dispatchWorkgroups(...dispatchSizes[i]);\n    }\n    if (this.argmaxEntry) {\n      pass.setPipeline(this.argmaxEntry.pipeline);\n      pass.setBindGroup(0, this.argmaxEntry.bindGroup);\n      pass.dispatchWorkgroups(1, 1, 1);\n    }\n    pass.end();\n    encoder.copyBufferToBuffer(this.argmaxResultBuffer, 0, this.decodeReadbacks[slot], 0, 4);\n    this.ctx.device.queue.submit([encoder.finish()]);\n\n    this.seqPos += 1;\n  }\n\n  /** Read back the token produced by the pipelined step that used `slot`. */\n  async readDecodeToken(slot: number): Promise<number> {\n    const buf = this.decodeReadbacks[slot];\n    await buf.mapAsync(MAP_MODE_READ, 0, 4);\n    const mapped = buf.getMappedRange(0, 4);\n    const tokenId = new Uint32Array(mapped.slice(0))[0];\n    buf.unmap();\n    return tokenId;\n  }\n\n  reset(): void {\n    this.seqPos = 0;\n    // Zero SSM state buffers\n    for (const buf of this.ssmStateBuffers.values()) {\n      this.ctx.device.queue.writeBuffer(buf, 0, new Uint8Array(buf.size));\n    }\n    // Clear cached params — shapes change between prefill (T=prompt_len) and decode (T=1)\n    for (const entry of this.dispatchEntries) {\n      entry.lastParamsBytes = null;\n      entry.lastDispatchSize = null;\n    }\n    for (const entry of this.decodeEntries) {\n      entry.lastParamsBytes = null;\n      entry.lastDispatchSize = null;\n    }\n  }\n\n  /**\n   * Diagnostic: dispatch ONLY the first kernel (EmbeddingInt4) using the\n   * production bind group, pipeline, and buffers — but in isolation (1 dispatch,\n   * no staging, fresh encoder). Compares against full forward pass to isolate\n   * whether the issue is the bind group/pipeline or the multi-dispatch context.\n   */\n  async debugFirstDispatch(inputIds: Uint32Array): Promise<{\n    nodeId: string;\n    opType: string;\n    dispatchSize: [number, number, number];\n    output: Float32Array;\n  }> {\n    const T = inputIds.length;\n    const resolvedShapes = this.resolveShapes(T);\n    const runtimeContext: RuntimeContext = { seqPos: this.seqPos };\n    const entries = T === 1 ? this.decodeEntries : this.dispatchEntries;\n    const entry = entries[0];\n\n    // Write input_ids directly (no staging)\n    this.ctx.device.queue.writeBuffer(this.inputIdsBuffer, 0, inputIds as BufferSource);\n\n    // Write params directly to uniform buffer (no staging)\n    const paramsData = entry.spec.buildParams(entry.node, resolvedShapes, runtimeContext);\n    this.ctx.device.queue.writeBuffer(entry.uniformBuffer, 0, paramsData);\n\n    // Single dispatch with fresh encoder\n    const dispatchSize = entry.spec.getDispatchSize(entry.node, resolvedShapes, runtimeContext) as [\n      number,\n      number,\n      number,\n    ];\n    const encoder = this.ctx.device.createCommandEncoder({ label: \"debug_first\" });\n    const pass = encoder.beginComputePass({ label: \"debug_first_pass\" });\n    pass.setPipeline(entry.pipeline);\n    pass.setBindGroup(0, entry.bindGroup);\n    pass.dispatchWorkgroups(...dispatchSize);\n    pass.end();\n    this.ctx.device.queue.submit([encoder.finish()]);\n\n    // Read back the output tensor (first 16 elements)\n    const outputName = entry.node.outputs[0];\n    const output = await this.debugReadBuffer(outputName, 16);\n\n    return {\n      nodeId: entry.nodeId,\n      opType: entry.node.opType,\n      dispatchSize,\n      output,\n    };\n  }\n\n  /**\n   * Diagnostic: run a single dispatch entry by index, in isolation.\n   * Call after debugFirstDispatch() to test whether entry[1] (RMSNorm)\n   * can read embed_out written by entry[0] (EmbeddingInt4).\n   */\n  async debugDispatchEntry(\n    entryIndex: number,\n    T: number,\n  ): Promise<{ nodeId: string; opType: string; output: Float32Array }> {\n    const resolvedShapes = this.resolveShapes(T);\n    const runtimeContext: RuntimeContext = { seqPos: this.seqPos };\n    const entries = T === 1 ? this.decodeEntries : this.dispatchEntries;\n    const entry = entries[entryIndex];\n\n    // Write params directly\n    const paramsData = entry.spec.buildParams(entry.node, resolvedShapes, runtimeContext);\n    this.ctx.device.queue.writeBuffer(entry.uniformBuffer, 0, paramsData);\n\n    // Single dispatch\n    const dispatchSize = entry.spec.getDispatchSize(entry.node, resolvedShapes, runtimeContext) as [\n      number,\n      number,\n      number,\n    ];\n    const encoder = this.ctx.device.createCommandEncoder({ label: `debug_entry_${entryIndex}` });\n    const pass = encoder.beginComputePass({ label: `debug_pass_${entryIndex}` });\n    pass.setPipeline(entry.pipeline);\n    pass.setBindGroup(0, entry.bindGroup);\n    pass.dispatchWorkgroups(...dispatchSize);\n    pass.end();\n    this.ctx.device.queue.submit([encoder.finish()]);\n\n    const outputName = entry.node.outputs[0];\n    const output = await this.debugReadBuffer(outputName, 16);\n    return { nodeId: entry.nodeId, opType: entry.node.opType, output };\n  }\n\n  /**\n   * Diagnostic: compute the JS-side params for the first N decode entries\n   * WITHOUT dispatching. Shows what buildParams produces.\n   */\n  debugComputeParams(\n    T: number,\n    count = 5,\n  ): Array<{\n    idx: number;\n    nodeId: string;\n    opType: string;\n    paramsU32: number[];\n    dispatchSize: [number, number, number];\n  }> {\n    const resolvedShapes = this.resolveShapes(T);\n    const runtimeContext: RuntimeContext = { seqPos: this.seqPos };\n    const entries = T === 1 ? this.decodeEntries : this.dispatchEntries;\n    const results = [];\n    for (let i = 0; i < Math.min(count, entries.length); i++) {\n      const entry = entries[i];\n      const paramsData = entry.spec.buildParams(entry.node, resolvedShapes, runtimeContext);\n      const u32 = new Uint32Array(paramsData);\n      const dispatchSize = entry.spec.getDispatchSize(\n        entry.node,\n        resolvedShapes,\n        runtimeContext,\n      ) as [number, number, number];\n      results.push({\n        idx: i,\n        nodeId: entry.nodeId,\n        opType: entry.node.opType,\n        paramsU32: Array.from(u32),\n        dispatchSize,\n      });\n    }\n    return results;\n  }\n\n  /**\n   * Diagnostic: after a forward pass, read back output tensors at several points\n   * in the pipeline to find where data drops to zero.\n   */\n  async debugPipelineProbe(T: number): Promise<\n    Array<{\n      idx: number;\n      nodeId: string;\n      opType: string;\n      tensor: string;\n      sum: number;\n      first4: number[];\n      uniformParams?: number[];\n    }>\n  > {\n    const entries = T === 1 ? this.decodeEntries : this.dispatchEntries;\n    // Focus on the first few entries to find where zeros start\n    const totalEntries = entries.length;\n    const indices = [\n      0,\n      1,\n      2,\n      3,\n      Math.floor(totalEntries / 4),\n      Math.floor(totalEntries / 2),\n      Math.floor((totalEntries * 3) / 4),\n      totalEntries - 1,\n    ];\n    const unique = [...new Set(indices)].filter((i) => i >= 0 && i < totalEntries);\n\n    const results: Array<{\n      idx: number;\n      nodeId: string;\n      opType: string;\n      tensor: string;\n      sum: number;\n      first4: number[];\n      uniformParams?: number[];\n    }> = [];\n    for (const i of unique) {\n      const entry = entries[i];\n      const outputName = entry.node.outputs[0];\n      if (!outputName) continue;\n      try {\n        const data = await this.debugReadBuffer(outputName, 16);\n        let sum = 0;\n        for (let j = 0; j < data.length; j++) sum += data[j];\n\n        // Also read back the uniform buffer to verify params\n        let uniformParams: number[] | undefined;\n        try {\n          const ubSize = entry.uniformBuffer.size;\n          const readback = this.ctx.device.createBuffer({\n            size: ubSize,\n            usage: 0x0001 | 0x0008, // MAP_READ | COPY_DST\n          });\n          const enc = this.ctx.device.createCommandEncoder();\n          enc.copyBufferToBuffer(entry.uniformBuffer, 0, readback, 0, ubSize);\n          this.ctx.device.queue.submit([enc.finish()]);\n          await readback.mapAsync(MAP_MODE_READ, 0, ubSize);\n          const u32 = new Uint32Array(readback.getMappedRange(0, ubSize).slice(0));\n          uniformParams = Array.from(u32);\n          readback.unmap();\n          readback.destroy();\n        } catch {\n          /* skip if uniform readback fails */\n        }\n\n        results.push({\n          idx: i,\n          nodeId: entry.nodeId,\n          opType: entry.node.opType,\n          tensor: outputName,\n          sum: Math.round(sum * 1e6) / 1e6,\n          first4: [data[0], data[1], data[2], data[3]].map((v) => Math.round(v * 1e6) / 1e6),\n          uniformParams,\n        });\n      } catch {\n        results.push({\n          idx: i,\n          nodeId: entry.nodeId,\n          opType: entry.node.opType,\n          tensor: outputName,\n          sum: NaN,\n          first4: [NaN, NaN, NaN, NaN],\n        });\n      }\n    }\n    return results;\n  }\n\n  debugWriteBuffer(tensorName: string, data: ArrayBufferView): void {\n    const buffer = this.getBuffer(tensorName);\n    if (!buffer) throw new Error(`No buffer for \"${tensorName}\"`);\n    this.ctx.device.queue.writeBuffer(buffer, 0, data as BufferSource);\n  }\n\n  /**\n   * Write a host-supplied activation input buffer (e.g. multimodal M-RoPE\n   * cos/sin, spliced vision embeddings, image row-map). The buffer must be a\n   * persistent activation tensor in the graph. Call before forward().\n   */\n  writeInput(tensorName: string, data: ArrayBufferView): void {\n    const buffer = this.getBuffer(tensorName);\n    if (!buffer) throw new Error(`No input buffer for \"${tensorName}\"`);\n    this.ctx.device.queue.writeBuffer(buffer, 0, data as BufferSource);\n  }\n\n  /** Write a host activation buffer at a byte offset (for per-row decode updates). */\n  writeInputAt(tensorName: string, data: ArrayBufferView, byteOffset: number): void {\n    const buffer = this.getBuffer(tensorName);\n    if (!buffer) throw new Error(`No input buffer for \"${tensorName}\"`);\n    this.ctx.device.queue.writeBuffer(buffer, byteOffset, data as BufferSource);\n  }\n\n  /** True if the graph has a buffer with this name (multimodal-capability probe). */\n  hasBuffer(tensorName: string): boolean {\n    return this.getBuffer(tensorName) !== undefined;\n  }\n\n  /** Current sequence position (number of tokens processed since reset). */\n  get currentSeqPos(): number {\n    return this.seqPos;\n  }\n\n  async debugReadBuffer(\n    tensorName: string,\n    maxElements?: number,\n    byteOffset?: number,\n  ): Promise<Float32Array> {\n    const buffer = this.getBuffer(tensorName);\n    if (!buffer) throw new Error(`No buffer for \"${tensorName}\"`);\n    const srcOffset = byteOffset ?? 0;\n    const remaining = buffer.size - srcOffset;\n    const byteLen = maxElements ? Math.min(maxElements * 4, remaining) : remaining;\n    const readback = this.ctx.device.createBuffer({\n      size: byteLen,\n      usage: 0x0001 | 0x0008,\n    });\n    const encoder = this.ctx.device.createCommandEncoder();\n    encoder.copyBufferToBuffer(buffer, srcOffset, readback, 0, byteLen);\n    this.ctx.device.queue.submit([encoder.finish()]);\n    await readback.mapAsync(MAP_MODE_READ, 0, byteLen);\n    const data = new Float32Array(readback.getMappedRange(0, byteLen).slice(0));\n    readback.unmap();\n    readback.destroy();\n    return data;\n  }\n\n  destroy(): void {\n    destroyBuffers([...this.weightBuffers.values()]);\n    // Pooled activation buffers are shared by multiple tensor names — dedupe.\n    destroyBuffers([...new Set(this.activationBuffers.values())]);\n    destroyBuffers([...this.ssmStateBuffers.values()]);\n    destroyBuffers([...this.kvCacheBuffers.values()]);\n    for (const entry of this.dispatchEntries) {\n      entry.uniformBuffer.destroy();\n    }\n    // Only destroy decode-specific uniform buffers (matmul nodes have their own)\n    for (let i = 0; i < this.decodeEntries.length; i++) {\n      if (this.decodeEntries[i] !== this.dispatchEntries[i]) {\n        this.decodeEntries[i].uniformBuffer.destroy();\n      }\n    }\n    if (this.argmaxEntry) {\n      this.argmaxEntry.uniformBuffer.destroy();\n    }\n    if (this.uniformStagingBuffer) {\n      this.uniformStagingBuffer.destroy();\n      this.uniformStagingBuffer = null;\n    }\n    this.inputIdsBuffer.destroy();\n    this.logitsReadback.destroy();\n    this.argmaxResultBuffer.destroy();\n    this.argmaxReadback.destroy();\n    destroyBuffers(this.decodeReadbacks);\n    this.decodeReadbacks = [];\n    this.weightBuffers.clear();\n    this.activationBuffers.clear();\n    this.ssmStateBuffers.clear();\n    this.kvCacheBuffers.clear();\n    this.dispatchEntries = [];\n    this.decodeEntries = [];\n  }\n\n  // ── Private ──────────────────────────────────────────────────────\n\n  /**\n   * Allocate activation buffers with liveness-based reuse.\n   *\n   * One dedicated buffer per activation tensor at full maxSeqLen is ~2.3GB for\n   * Qwen3.5-0.8B at T=512 — over the iOS jetsam budget on its own. Instead,\n   * a buffer returns to a size-keyed pool once its tensor's last reader has\n   * executed, so concurrently-live tensors share a small working set.\n   *\n   * Graph outputs and tensors read before they are written (cross-forward\n   * state) keep dedicated buffers. Within a forward, dispatches execute in\n   * executionOrder on every path (single-pass Dawn, per-dispatch WebKit), and\n   * WebGPU synchronizes hazards between dispatches, so reuse is safe.\n   *\n   * Caveat: debugReadBuffer() on an intermediate tensor is only meaningful\n   * before a later op reuses its buffer (probes that stop mid-graph are fine).\n   */\n  private allocateActivationBuffers(): void {\n    const isActivation = (name: string) => this.graph.tensors[name]?.storage === \"activation\";\n    // Debug: give every activation its own buffer so post-hoc debugReadBuffer of\n    // intermediate tensors is reliable (pooling otherwise recycles buffers).\n    const noPool = typeof process !== \"undefined\" && process.env?.GERBIL_NO_ACT_POOL === \"1\";\n    if (noPool) {\n      for (const name of Object.keys(this.graph.tensors)) {\n        if (!isActivation(name)) continue;\n        const desc = this.graph.tensors[name];\n        const shape = desc.shape.map((d) =>\n          d === \"T\" || d === \"L_max\" ? this.maxSeqLen : (d as number),\n        );\n        const bytes = shape.reduce((a, b) => a * b, 1) * DTYPE_BYTES[desc.dtype];\n        this.activationBuffers.set(\n          name,\n          createStorageBuffer(this.ctx, `act_nopool_${name}`, bytes),\n        );\n      }\n      console.log(\n        `[executor] activations: NO-POOL debug mode → ${this.activationBuffers.size} dedicated buffers`,\n      );\n      return;\n    }\n    const resolveBytes = (name: string): number => {\n      const desc = this.graph.tensors[name];\n      const shape = desc.shape.map((d) => {\n        if (d === \"T\") return this.maxSeqLen;\n        if (d === \"L_max\") return this.maxSeqLen;\n        return d as number;\n      });\n      return shape.reduce((a, b) => a * b, 1) * DTYPE_BYTES[desc.dtype];\n    };\n\n    const nodeById = new Map(this.graph.nodes.map((n) => [n.id, n]));\n    const order = this.graph.executionOrder;\n    const persistent = new Set<string>(this.graph.outputs);\n    const firstDef = new Map<string, number>();\n    const lastUse = new Map<string, number>();\n    for (let i = 0; i < order.length; i++) {\n      const node = nodeById.get(order[i])!;\n      for (const inp of node.inputs) {\n        if (!isActivation(inp)) continue;\n        // Read before any write → carries state across forwards (or in-place op);\n        // never pool it.\n        if (!firstDef.has(inp)) persistent.add(inp);\n        lastUse.set(inp, i);\n      }\n      for (const out of node.outputs) {\n        if (isActivation(out) && !firstDef.has(out)) firstDef.set(out, i);\n      }\n    }\n\n    const freePool = new Map<number, GPUBuffer[]>();\n    let created = 0;\n    const released = new Set<string>();\n    for (let i = 0; i < order.length; i++) {\n      const node = nodeById.get(order[i])!;\n      // Acquire outputs first so a node never writes the buffer of a tensor\n      // it also reads.\n      for (const out of node.outputs) {\n        if (!isActivation(out) || persistent.has(out) || this.activationBuffers.has(out)) {\n          continue;\n        }\n        const byteSize = resolveBytes(out);\n        const pool = freePool.get(byteSize);\n        let buffer = pool?.pop();\n        if (!buffer) {\n          buffer = createStorageBuffer(this.ctx, `act_pool_${created}_${byteSize}b`, byteSize);\n          created++;\n        }\n        this.activationBuffers.set(out, buffer);\n      }\n      // Release buffers whose tensor will never be read again.\n      for (const name of [...node.inputs, ...node.outputs]) {\n        if (!isActivation(name) || persistent.has(name) || released.has(name)) continue;\n        const last = lastUse.get(name) ?? firstDef.get(name) ?? i;\n        if (last > i) continue;\n        const buffer = this.activationBuffers.get(name);\n        if (!buffer) continue;\n        released.add(name);\n        const byteSize = resolveBytes(name);\n        const pool = freePool.get(byteSize) ?? [];\n        pool.push(buffer);\n        freePool.set(byteSize, pool);\n      }\n    }\n\n    // Dedicated buffers for graph outputs, cross-forward state, and any\n    // activation tensor not touched by the execution order.\n    for (const name of Object.keys(this.graph.tensors)) {\n      if (!isActivation(name) || this.activationBuffers.has(name)) continue;\n      this.activationBuffers.set(\n        name,\n        createStorageBuffer(this.ctx, `act_${name}`, resolveBytes(name)),\n      );\n    }\n\n    const unique = new Set(this.activationBuffers.values());\n    let totalBytes = 0;\n    for (const b of unique) totalBytes += b.size;\n    console.log(\n      `[executor] activations: ${this.activationBuffers.size} tensors → ${unique.size} buffers (${(totalBytes / 1e6).toFixed(0)} MB)`,\n    );\n  }\n\n  private allocateSSMStateBuffers(): void {\n    for (const [name, desc] of Object.entries(this.graph.tensors)) {\n      if (desc.storage !== \"ssm_state\") continue;\n      const shape = desc.shape as number[];\n      const byteSize = shape.reduce((a, b) => a * b, 1) * DTYPE_BYTES[desc.dtype];\n      const buffer = createStorageBuffer(this.ctx, `ssm_${name}`, byteSize);\n      this.ssmStateBuffers.set(name, buffer);\n    }\n  }\n\n  private allocateKVCacheBuffers(): void {\n    for (const [name, desc] of Object.entries(this.graph.tensors)) {\n      if (desc.storage !== \"kv_cache\") continue;\n      const shape = desc.shape.map((d) => {\n        if (d === \"L_max\") return this.maxSeqLen;\n        return d as number;\n      });\n      const byteSize = shape.reduce((a, b) => a * b, 1) * DTYPE_BYTES[desc.dtype];\n      const buffer = createStorageBuffer(this.ctx, `kv_${name}`, byteSize);\n      this.kvCacheBuffers.set(name, buffer);\n    }\n  }\n\n  private resolveShapes(T: number): Record<string, number[]> {\n    const resolved: Record<string, number[]> = {};\n    for (const [name, desc] of Object.entries(this.graph.tensors)) {\n      resolved[name] = desc.shape.map((d) => {\n        if (d === \"T\") return T;\n        if (d === \"L_max\") return this.seqPos + T;\n        return d as number;\n      });\n    }\n    return resolved;\n  }\n\n  private getBuffer(tensorName: string): GPUBuffer | undefined {\n    return (\n      this.weightBuffers.get(tensorName) ??\n      this.activationBuffers.get(tensorName) ??\n      this.ssmStateBuffers.get(tensorName) ??\n      this.kvCacheBuffers.get(tensorName)\n    );\n  }\n\n  /**\n   * Detect gate_proj + up_proj + SwiGLU patterns in decode entries and replace\n   * with a single fused SwiGLUMatVec dispatch. Saves 2 dispatches per MLP block.\n   */\n  private fuseSwiGLUDecodeEntries(): void {\n    let fusionCount = 0;\n    const maxBindings = this.ctx.limits.maxStorageBuffersPerShaderStage;\n\n    for (let i = 0; i < this.decodeEntries.length - 2; i++) {\n      const e1 = this.decodeEntries[i]; // gate_proj MatVec\n      const e2 = this.decodeEntries[i + 1]; // up_proj MatVec\n      const e3 = this.decodeEntries[i + 2]; // SwiGLU\n\n      // Check op types: two MatMul/MatMulInt4 followed by SwiGLU\n      const isE1Mat = e1.node.opType === \"MatMul\" || e1.node.opType === \"MatMulInt4\";\n      const isE2Mat = e2.node.opType === \"MatMul\" || e2.node.opType === \"MatMulInt4\";\n      if (!isE1Mat || !isE2Mat || e3.node.opType !== \"SwiGLU\") continue;\n\n      // Both projections must use the same op type (both F32 or both INT4)\n      if (e1.node.opType !== e2.node.opType) continue;\n\n      // INT4 fused kernel needs 9 storage bindings — skip if device doesn't support it\n      const isInt4Fusion = e1.node.opType === \"MatMulInt4\";\n      if (isInt4Fusion && maxBindings < 9) continue;\n\n      // Both must share the same input activation (first input tensor)\n      if (e1.node.inputs[0] !== e2.node.inputs[0]) continue;\n\n      // SwiGLU must consume the outputs of both projections\n      const gateOut = e1.node.outputs[0];\n      const upOut = e2.node.outputs[0];\n      if (e3.node.inputs[0] !== gateOut || e3.node.inputs[1] !== upOut) continue;\n\n      // Same K and N dimensions\n      if (e1.node.attributes.K !== e2.node.attributes.K) continue;\n      if (e1.node.attributes.N !== e2.node.attributes.N) continue;\n\n      // Pooled activation aliasing: the fused kernel reads the shared input\n      // and writes the SwiGLU output in ONE dispatch — they must not share\n      // a pooled buffer (separate dispatches are hazard-synchronized; a\n      // single dispatch is not).\n      if (this.getBuffer(e1.node.inputs[0]) === this.getBuffer(e3.node.outputs[0])) continue;\n\n      // Pattern matched — create fused dispatch entry\n      const fusedSpec = isInt4Fusion ? SWIGLU_MATVEC_INT4_SPEC : SWIGLU_MATVEC_SPEC;\n      const fusedPipeline = getOrCreatePipeline(\n        this.ctx,\n        `swiglu_mv_${e1.nodeId}`,\n        fusedSpec.shaderCode,\n        fusedSpec.entryPoint,\n      );\n\n      const fusedUniformData = fusedSpec.buildParams(e1.node, {}, { seqPos: 0 });\n      const fusedUniform = createUniformBuffer(\n        this.ctx,\n        `uniform_swiglu_mv_${e1.nodeId}`,\n        fusedUniformData,\n      );\n\n      // Build bind group manually: input, gate weights, up weights, output, uniform\n      const inputBuf = this.getBuffer(e1.node.inputs[0])!;\n      const outputBuf = this.getBuffer(e3.node.outputs[0])!;\n\n      let bufferEntries: Array<{ buffer: GPUBuffer }>;\n      if (isInt4Fusion) {\n        // INT4: A, gate_q, gate_s, gate_z, up_q, up_s, up_z, output, params\n        const gateQ = this.getBuffer(e1.node.inputs[1])!;\n        const gateS = this.getBuffer(e1.node.inputs[2])!;\n        const gateZ = this.getBuffer(e1.node.inputs[3])!;\n        const upQ = this.getBuffer(e2.node.inputs[1])!;\n        const upS = this.getBuffer(e2.node.inputs[2])!;\n        const upZ = this.getBuffer(e2.node.inputs[3])!;\n        bufferEntries = [\n          { buffer: inputBuf },\n          { buffer: gateQ },\n          { buffer: gateS },\n          { buffer: gateZ },\n          { buffer: upQ },\n          { buffer: upS },\n          { buffer: upZ },\n          { buffer: outputBuf },\n          { buffer: fusedUniform },\n        ];\n      } else {\n        // F32: A, B_gate, B_up, output, params\n        const gateBuf = this.getBuffer(e1.node.inputs[1])!;\n        const upBuf = this.getBuffer(e2.node.inputs[1])!;\n        bufferEntries = [\n          { buffer: inputBuf },\n          { buffer: gateBuf },\n          { buffer: upBuf },\n          { buffer: outputBuf },\n          { buffer: fusedUniform },\n        ];\n      }\n\n      const fusedBindGroup = createBindGroup(\n        this.ctx,\n        fusedPipeline,\n        bufferEntries,\n        `bg_swiglu_mv_${e1.nodeId}`,\n      );\n\n      const fusedEntry: DispatchEntry = {\n        nodeId: `fused_swiglu_${e1.nodeId}`,\n        node: e1.node, // Use gate node for K/N attributes\n        spec: fusedSpec,\n        pipeline: fusedPipeline,\n        bindGroup: fusedBindGroup,\n        uniformBuffer: fusedUniform,\n        lastParamsBytes: null,\n        lastDispatchSize: null,\n      };\n\n      // Replace 3 entries with 1 fused entry\n      this.decodeEntries.splice(i, 3, fusedEntry);\n      fusionCount++;\n      // Don't increment i — check if next position is also fusable\n    }\n\n    if (fusionCount > 0) {\n      console.log(\n        `[executor] Fused ${fusionCount} SwiGLU MLP blocks (saved ${fusionCount * 2} dispatches)`,\n      );\n    }\n  }\n\n  /**\n   * Fuse two adjacent INT4 projections that share the same input activation and\n   * the same K/N (e.g. q_proj+gate_proj and k_proj+v_proj in full-attention\n   * decode) into a single DualMatVecInt4 dispatch. Reads the shared input vector\n   * once and writes both projection outputs — removing one GPU round-trip (one\n   * submit+drain on Safari/iOS) per fused pair.\n   *\n   * Numerics are identical to running the two MatVecInt4 kernels separately\n   * (same dequant, same K-parallel reduction), so this is WebKit-safe: it reuses\n   * the proven INT4 matvec math and only merges two writes into one dispatch.\n   *\n   * Must run AFTER fuseSwiGLUDecodeEntries so the MLP gate+up pair (which is\n   * consumed by a SwiGLU node) is already collapsed and won't be matched here.\n   */\n  private fuseDualMatVecDecodeEntries(): void {\n    let fusionCount = 0;\n    // DualMatVecInt4 needs 10 storage/uniform bindings (9 storage + 1 uniform).\n    const maxBindings = this.ctx.limits.maxStorageBuffersPerShaderStage;\n    if (maxBindings < 9) return;\n\n    for (let i = 0; i < this.decodeEntries.length - 1; i++) {\n      const e0 = this.decodeEntries[i];\n      const e1 = this.decodeEntries[i + 1];\n\n      // Both must be INT4 matmuls.\n      if (e0.node.opType !== \"MatMulInt4\" || e1.node.opType !== \"MatMulInt4\") continue;\n      // Same input activation (first input tensor).\n      if (e0.node.inputs[0] !== e1.node.inputs[0]) continue;\n      // Same K, N, group_size.\n      if (e0.node.attributes.K !== e1.node.attributes.K) continue;\n      if (e0.node.attributes.N !== e1.node.attributes.N) continue;\n      if (e0.node.attributes.group_size !== e1.node.attributes.group_size) continue;\n\n      const out0 = e0.node.outputs[0];\n      const out1 = e1.node.outputs[0];\n      const inputBuf = this.getBuffer(e0.node.inputs[0]);\n      const out0Buf = this.getBuffer(out0);\n      const out1Buf = this.getBuffer(out1);\n      if (!inputBuf || !out0Buf || !out1Buf) continue;\n\n      // Pooled-buffer aliasing guard: the single fused dispatch reads the shared\n      // input and writes both outputs. None of these may share a pooled buffer\n      // (separate dispatches are hazard-synchronized; one dispatch is not), and\n      // the two outputs must be distinct buffers.\n      if (inputBuf === out0Buf || inputBuf === out1Buf || out0Buf === out1Buf) continue;\n\n      const fusedSpec = DUAL_MATVEC_INT4_SPEC;\n      const fusedPipeline = getOrCreatePipeline(\n        this.ctx,\n        `dual_mv_${e0.nodeId}`,\n        fusedSpec.shaderCode,\n        fusedSpec.entryPoint,\n      );\n\n      const fusedUniformData = fusedSpec.buildParams(e0.node, {}, { seqPos: 0 });\n      const fusedUniform = createUniformBuffer(\n        this.ctx,\n        `uniform_dual_mv_${e0.nodeId}`,\n        fusedUniformData,\n      );\n\n      // Bind group: A, B0_q, B0_s, B0_z, B1_q, B1_s, B1_z, out0, out1, params\n      const b0q = this.getBuffer(e0.node.inputs[1])!;\n      const b0s = this.getBuffer(e0.node.inputs[2])!;\n      const b0z = this.getBuffer(e0.node.inputs[3])!;\n      const b1q = this.getBuffer(e1.node.inputs[1])!;\n      const b1s = this.getBuffer(e1.node.inputs[2])!;\n      const b1z = this.getBuffer(e1.node.inputs[3])!;\n\n      const bufferEntries = [\n        { buffer: inputBuf },\n        { buffer: b0q },\n        { buffer: b0s },\n        { buffer: b0z },\n        { buffer: b1q },\n        { buffer: b1s },\n        { buffer: b1z },\n        { buffer: out0Buf },\n        { buffer: out1Buf },\n        { buffer: fusedUniform },\n      ];\n\n      const fusedBindGroup = createBindGroup(\n        this.ctx,\n        fusedPipeline,\n        bufferEntries,\n        `bg_dual_mv_${e0.nodeId}`,\n      );\n\n      const fusedEntry: DispatchEntry = {\n        nodeId: `fused_dual_${e0.nodeId}`,\n        node: e0.node, // K/N/group_size attributes shared\n        spec: fusedSpec,\n        pipeline: fusedPipeline,\n        bindGroup: fusedBindGroup,\n        uniformBuffer: fusedUniform,\n        lastParamsBytes: null,\n        lastDispatchSize: null,\n      };\n\n      this.decodeEntries.splice(i, 2, fusedEntry);\n      fusionCount++;\n      // Don't increment i — the next pair may also be fusable.\n    }\n\n    if (fusionCount > 0) {\n      console.log(\n        `[executor] Fused ${fusionCount} dual INT4 projections (saved ${fusionCount} dispatches)`,\n      );\n    }\n  }\n\n  /**\n   * Fuse the adjacent K-cache and V-cache appends in each full-attention layer\n   * into a single DualKVCacheAppend dispatch. Both are pure memcpys into f32\n   * caches sharing the same width and dst_offset, so one dispatch with two\n   * src/dst buffers writes both — removing one GPU round-trip per layer.\n   *\n   * Supports f32, native-f16, and packed-f16 caches (the dual kernel mirrors the\n   * single-append kernel selected for the active kvMode). Numerically identical\n   * to the separate appends — WebKit-safe (the packed-f16 variant is the Safari\n   * path and uses pack2x16float, no `enable f16`).\n   */\n  private fuseDualKVCacheAppendEntries(): void {\n    let fusionCount = 0;\n\n    for (let i = 0; i < this.decodeEntries.length - 1; i++) {\n      const e0 = this.decodeEntries[i]; // K append\n      const e1 = this.decodeEntries[i + 1]; // V append\n\n      if (e0.node.opType !== \"KVCacheAppend\" || e1.node.opType !== \"KVCacheAppend\") continue;\n      // Same per-element width (kv_dim) → one param set drives both.\n      if (e0.node.attributes.width !== e1.node.attributes.width) continue;\n\n      // Select the dual kernel matching the cache dtype/mode. Both destinations\n      // must share the same dtype (they always do: K and V caches are symmetric).\n      const dstK = e0.node.outputs[0];\n      const dstV = e1.node.outputs[0];\n      const dtK = this.graph.tensors[dstK]?.dtype;\n      const dtV = this.graph.tensors[dstV]?.dtype;\n      if (dtK !== dtV) continue;\n      let fusedSpec: KernelSpec;\n      if (dtK === \"f32\") {\n        fusedSpec = DUAL_KV_CACHE_APPEND_SPEC;\n      } else if (dtK === \"f16\") {\n        fusedSpec =\n          this.kvMode === \"packed-f16\"\n            ? DUAL_KV_CACHE_APPEND_PACKED_F16_SPEC\n            : DUAL_KV_CACHE_APPEND_F16_SPEC;\n      } else {\n        continue;\n      }\n\n      const srcK = this.getBuffer(e0.node.inputs[0]);\n      const srcV = this.getBuffer(e1.node.inputs[0]);\n      const dstKBuf = this.getBuffer(dstK);\n      const dstVBuf = this.getBuffer(dstV);\n      if (!srcK || !srcV || !dstKBuf || !dstVBuf) continue;\n      // Distinct caches; sources distinct from caches.\n      if (dstKBuf === dstVBuf) continue;\n\n      const fusedPipeline = getOrCreatePipeline(\n        this.ctx,\n        `dual_kv_${e0.nodeId}`,\n        fusedSpec.shaderCode,\n        fusedSpec.entryPoint,\n      );\n\n      const dummyShapes = this.resolveShapes(1);\n      const fusedUniformData = fusedSpec.buildParams(e0.node, dummyShapes, { seqPos: 0 });\n      const fusedUniform = createUniformBuffer(\n        this.ctx,\n        `uniform_dual_kv_${e0.nodeId}`,\n        fusedUniformData,\n      );\n\n      const bufferEntries = [\n        { buffer: srcK },\n        { buffer: srcV },\n        { buffer: dstKBuf },\n        { buffer: dstVBuf },\n        { buffer: fusedUniform },\n      ];\n\n      const fusedBindGroup = createBindGroup(\n        this.ctx,\n        fusedPipeline,\n        bufferEntries,\n        `bg_dual_kv_${e0.nodeId}`,\n      );\n\n      const fusedEntry: DispatchEntry = {\n        nodeId: `fused_kv_${e0.nodeId}`,\n        node: e0.node, // width + T_tensor for per-token param rebuild\n        spec: fusedSpec,\n        pipeline: fusedPipeline,\n        bindGroup: fusedBindGroup,\n        uniformBuffer: fusedUniform,\n        lastParamsBytes: null,\n        lastDispatchSize: null,\n      };\n\n      this.decodeEntries.splice(i, 2, fusedEntry);\n      fusionCount++;\n    }\n\n    if (fusionCount > 0) {\n      console.log(\n        `[executor] Fused ${fusionCount} dual KV-cache appends (saved ${fusionCount} dispatches)`,\n      );\n    }\n  }\n\n  /**\n   * Fuse the attention SigmoidGate (attn_out * sigmoid(gate)) into the INT4\n   * o_proj that consumes it: a GatedMatVecInt4 reads attn_out and gate directly,\n   * applies the sigmoid gate to its input vector, and runs the projection in ONE\n   * dispatch — removing the standalone SigmoidGate (one round-trip per\n   * full-attention layer).\n   *\n   * Numerically identical to SigmoidGate→MatVecInt4 (same gate formula, same INT4\n   * dequant + reduction). Slight extra ALU: the gated input is recomputed per\n   * output column, but A reads hit L1 and the saved submit+drain dominates on\n   * mobile. WebKit risk: low — reuses the proven INT4 matvec, only the A vector\n   * is built from two reads + a sigmoid (no new reduction/barrier pattern).\n   *\n   * Runs after the dual fusions so it only sees the post-attention SigmoidGate.\n   */\n  private fuseGatedOProjDecodeEntries(): void {\n    let fusionCount = 0;\n\n    for (let i = 0; i < this.decodeEntries.length - 1; i++) {\n      const eg = this.decodeEntries[i]; // SigmoidGate\n      const eo = this.decodeEntries[i + 1]; // o_proj MatMulInt4\n\n      if (eg.node.opType !== \"SigmoidGate\") continue;\n      if (eo.node.opType !== \"MatMulInt4\") continue;\n\n      // o_proj must consume the gate's output as its activation input.\n      const gatedOut = eg.node.outputs[0];\n      if (eo.node.inputs[0] !== gatedOut) continue;\n\n      // SigmoidGate inputs: [x (attn_out), gate]\n      const attnName = eg.node.inputs[0];\n      const gateName = eg.node.inputs[1];\n      const attnBuf = this.getBuffer(attnName);\n      const gateBuf = this.getBuffer(gateName);\n      const wQ = this.getBuffer(eo.node.inputs[1]);\n      const wS = this.getBuffer(eo.node.inputs[2]);\n      const wZ = this.getBuffer(eo.node.inputs[3]);\n      const outBuf = this.getBuffer(eo.node.outputs[0]);\n      if (!attnBuf || !gateBuf || !wQ || !wS || !wZ || !outBuf) continue;\n\n      // Pooled aliasing: the single fused dispatch reads attn + gate and writes\n      // the projection output — none may share a pooled buffer.\n      if (outBuf === attnBuf || outBuf === gateBuf || attnBuf === gateBuf) continue;\n\n      const fusedSpec = GATED_MATVEC_INT4_SPEC;\n      const fusedPipeline = getOrCreatePipeline(\n        this.ctx,\n        `gated_o_${eo.nodeId}`,\n        fusedSpec.shaderCode,\n        fusedSpec.entryPoint,\n      );\n\n      const fusedUniformData = fusedSpec.buildParams(eo.node, {}, { seqPos: 0 });\n      const fusedUniform = createUniformBuffer(\n        this.ctx,\n        `uniform_gated_o_${eo.nodeId}`,\n        fusedUniformData,\n      );\n\n      const bufferEntries = [\n        { buffer: attnBuf },\n        { buffer: gateBuf },\n        { buffer: wQ },\n        { buffer: wS },\n        { buffer: wZ },\n        { buffer: outBuf },\n        { buffer: fusedUniform },\n      ];\n\n      const fusedBindGroup = createBindGroup(\n        this.ctx,\n        fusedPipeline,\n        bufferEntries,\n        `bg_gated_o_${eo.nodeId}`,\n      );\n\n      const fusedEntry: DispatchEntry = {\n        nodeId: `fused_gated_o_${eo.nodeId}`,\n        node: eo.node, // K/N/group_size attributes for per-token param rebuild\n        spec: fusedSpec,\n        pipeline: fusedPipeline,\n        bindGroup: fusedBindGroup,\n        uniformBuffer: fusedUniform,\n        lastParamsBytes: null,\n        lastDispatchSize: null,\n      };\n\n      this.decodeEntries.splice(i, 2, fusedEntry);\n      fusionCount++;\n    }\n\n    if (fusionCount > 0) {\n      console.log(\n        `[executor] Fused ${fusionCount} gated o_proj blocks (saved ${fusionCount} dispatches)`,\n      );\n    }\n  }\n\n  /**\n   * Fuse a standalone SwiGLU (silu(gate) * up) into the INT4 projection that\n   * consumes its output: a SwiGLUGatedMatVecInt4 reads gate and up directly,\n   * builds the gated input vector, and runs the projection in ONE dispatch.\n   * Targets the Mamba block's mamba_swiglu (silu(z) * norm_out) feeding out_proj\n   * — one round-trip saved per linear-attention layer.\n   *\n   * The MLP SwiGLU is already collapsed into a SwiGLUMatVec entry by\n   * fuseSwiGLUDecodeEntries (it has no surviving standalone SwiGLU node), so only\n   * the Mamba SwiGLU matches here. Numerically identical to SwiGLU→MatVecInt4;\n   * WebKit risk low (reuses the proven INT4 matvec, only the A vector changes).\n   */\n  private fuseSwiGLUGatedProjDecodeEntries(): void {\n    let fusionCount = 0;\n\n    for (let i = 0; i < this.decodeEntries.length - 1; i++) {\n      const es = this.decodeEntries[i]; // SwiGLU\n      const ep = this.decodeEntries[i + 1]; // projection MatMulInt4\n\n      if (es.node.opType !== \"SwiGLU\") continue;\n      if (ep.node.opType !== \"MatMulInt4\") continue;\n\n      const swigluOut = es.node.outputs[0];\n      if (ep.node.inputs[0] !== swigluOut) continue;\n\n      // SwiGLU inputs: [gate, up]\n      const gateName = es.node.inputs[0];\n      const upName = es.node.inputs[1];\n      const gateBuf = this.getBuffer(gateName);\n      const upBuf = this.getBuffer(upName);\n      const wQ = this.getBuffer(ep.node.inputs[1]);\n      const wS = this.getBuffer(ep.node.inputs[2]);\n      const wZ = this.getBuffer(ep.node.inputs[3]);\n      const outBuf = this.getBuffer(ep.node.outputs[0]);\n      if (!gateBuf || !upBuf || !wQ || !wS || !wZ || !outBuf) continue;\n\n      // Pooled aliasing: the single fused dispatch reads gate + up and writes the\n      // projection output — none may share a pooled buffer.\n      if (outBuf === gateBuf || outBuf === upBuf || gateBuf === upBuf) continue;\n\n      const fusedSpec = SWIGLU_GATED_MATVEC_INT4_SPEC;\n      const fusedPipeline = getOrCreatePipeline(\n        this.ctx,\n        `swiglu_gated_${ep.nodeId}`,\n        fusedSpec.shaderCode,\n        fusedSpec.entryPoint,\n      );\n\n      const fusedUniformData = fusedSpec.buildParams(ep.node, {}, { seqPos: 0 });\n      const fusedUniform = createUniformBuffer(\n        this.ctx,\n        `uniform_swiglu_gated_${ep.nodeId}`,\n        fusedUniformData,\n      );\n\n      const bufferEntries = [\n        { buffer: gateBuf },\n        { buffer: upBuf },\n        { buffer: wQ },\n        { buffer: wS },\n        { buffer: wZ },\n        { buffer: outBuf },\n        { buffer: fusedUniform },\n      ];\n\n      const fusedBindGroup = createBindGroup(\n        this.ctx,\n        fusedPipeline,\n        bufferEntries,\n        `bg_swiglu_gated_${ep.nodeId}`,\n      );\n\n      const fusedEntry: DispatchEntry = {\n        nodeId: `fused_swiglu_gated_${ep.nodeId}`,\n        node: ep.node,\n        spec: fusedSpec,\n        pipeline: fusedPipeline,\n        bindGroup: fusedBindGroup,\n        uniformBuffer: fusedUniform,\n        lastParamsBytes: null,\n        lastDispatchSize: null,\n      };\n\n      this.decodeEntries.splice(i, 2, fusedEntry);\n      fusionCount++;\n    }\n\n    if (fusionCount > 0) {\n      console.log(\n        `[executor] Fused ${fusionCount} SwiGLU-gated projections (saved ${fusionCount} dispatches)`,\n      );\n    }\n  }\n\n  /**\n   * Fuse two adjacent per-row RMSNorms sharing hidden_size + eps into a single\n   * DualRMSNorm dispatch (e.g. the per-head q_norm and k_norm in full-attention\n   * decode). One workgroup still handles one row; the fused grid just spans both\n   * inputs' rows, so each row's reduction is unchanged — numerically identical to\n   * two separate RMSNorm dispatches. One round-trip saved per fused pair.\n   *\n   * WebKit risk: low — same single-workgroup reduction as the proven RMSNorm\n   * kernel, only the row→input routing is added.\n   */\n  private fuseDualRMSNormDecodeEntries(): void {\n    let fusionCount = 0;\n\n    for (let i = 0; i < this.decodeEntries.length - 1; i++) {\n      const e0 = this.decodeEntries[i];\n      const e1 = this.decodeEntries[i + 1];\n\n      if (e0.node.opType !== \"RMSNorm\" || e1.node.opType !== \"RMSNorm\") continue;\n      // Must share per-row width and eps (the dual kernel uses one of each).\n      if (e0.node.attributes.hidden_size !== e1.node.attributes.hidden_size) continue;\n      if (e0.node.attributes.eps !== e1.node.attributes.eps) continue;\n\n      const in0 = e0.node.inputs[0];\n      const w0 = e0.node.inputs[1];\n      const out0 = e0.node.outputs[0];\n      const in1 = e1.node.inputs[0];\n      const w1 = e1.node.inputs[1];\n      const out1 = e1.node.outputs[0];\n\n      const in0Buf = this.getBuffer(in0);\n      const w0Buf = this.getBuffer(w0);\n      const out0Buf = this.getBuffer(out0);\n      const in1Buf = this.getBuffer(in1);\n      const w1Buf = this.getBuffer(w1);\n      const out1Buf = this.getBuffer(out1);\n      if (!in0Buf || !w0Buf || !out0Buf || !in1Buf || !w1Buf || !out1Buf) continue;\n\n      // The two normalizations must write distinct buffers. (In-place RMSNorm\n      // where input===output is fine: each row reads then writes its own row,\n      // and the fused kernel preserves per-row ordering.) Cross-aliasing — where\n      // one norm's input shares a pooled buffer with the OTHER norm's output —\n      // would create an unsynchronized read/write hazard within the single\n      // dispatch, so reject it.\n      if (out0Buf === out1Buf) continue;\n      if (in0Buf === out1Buf || in1Buf === out0Buf) continue;\n\n      // Build a synthetic node carrying both row sources so the spec can resolve\n      // per-token row counts each decode step.\n      const fusedNode: OpNode = {\n        id: `fused_dualnorm_${e0.nodeId}`,\n        opType: \"RMSNorm\",\n        inputs: [in0, w0, in1, w1],\n        outputs: [out0, out1],\n        attributes: {\n          hidden_size: e0.node.attributes.hidden_size,\n          eps: e0.node.attributes.eps,\n          seq_len_tensor0: e0.node.attributes.seq_len_tensor,\n          seq_len_tensor1: e1.node.attributes.seq_len_tensor,\n          seq_len0: e0.node.attributes.seq_len,\n          seq_len1: e1.node.attributes.seq_len,\n        },\n      };\n\n      const fusedSpec = DUAL_RMSNORM_SPEC;\n      const fusedPipeline = getOrCreatePipeline(\n        this.ctx,\n        `dual_norm_${e0.nodeId}`,\n        fusedSpec.shaderCode,\n        fusedSpec.entryPoint,\n      );\n\n      const dummyShapes = this.resolveShapes(1);\n      const fusedUniformData = fusedSpec.buildParams(fusedNode, dummyShapes, { seqPos: 0 });\n      const fusedUniform = createUniformBuffer(\n        this.ctx,\n        `uniform_dual_norm_${e0.nodeId}`,\n        fusedUniformData,\n      );\n\n      const bufferEntries = [\n        { buffer: in0Buf },\n        { buffer: w0Buf },\n        { buffer: in1Buf },\n        { buffer: w1Buf },\n        { buffer: out0Buf },\n        { buffer: out1Buf },\n        { buffer: fusedUniform },\n      ];\n\n      const fusedBindGroup = createBindGroup(\n        this.ctx,\n        fusedPipeline,\n        bufferEntries,\n        `bg_dual_norm_${e0.nodeId}`,\n      );\n\n      const fusedEntry: DispatchEntry = {\n        nodeId: fusedNode.id,\n        node: fusedNode,\n        spec: fusedSpec,\n        pipeline: fusedPipeline,\n        bindGroup: fusedBindGroup,\n        uniformBuffer: fusedUniform,\n        lastParamsBytes: null,\n        lastDispatchSize: null,\n      };\n\n      this.decodeEntries.splice(i, 2, fusedEntry);\n      fusionCount++;\n    }\n\n    if (fusionCount > 0) {\n      console.log(\n        `[executor] Fused ${fusionCount} dual RMSNorms (saved ${fusionCount} dispatches)`,\n      );\n    }\n  }\n\n  /**\n   * Gather buffer entries for a bind group, matching the kernel spec's binding layout.\n   * Uses the pre-allocated inputIdsBuffer for the \"input_ids\" tensor.\n   */\n  private gatherBuffers(\n    spec: KernelSpec,\n    node: OpNode,\n    uniformBuffer: GPUBuffer,\n  ): Array<{ buffer: GPUBuffer }> {\n    const entries: Array<{ buffer: GPUBuffer }> = [];\n    const allTensorNames = [...node.inputs, ...node.outputs];\n\n    // Track buffers already bound to a writable slot in this group so we can\n    // detect (and break) read_write aliasing, which WebGPU rejects.\n    const writableBound = new Set<GPUBuffer>();\n\n    for (let i = 0; i < spec.bindings.length; i++) {\n      const binding = spec.bindings[i];\n\n      if (binding.type === \"uniform\") {\n        entries.push({ buffer: uniformBuffer });\n      } else {\n        const tensorIdx = i - spec.bindings.filter((b, j) => j < i && b.type === \"uniform\").length;\n        const tensorName = allTensorNames[tensorIdx];\n        if (!tensorName) throw new Error(`Missing tensor for binding ${i} in op ${node.id}`);\n\n        let buffer: GPUBuffer | undefined;\n        if (tensorName === \"input_ids\") {\n          buffer = this.inputIdsBuffer;\n        } else {\n          buffer = this.getBuffer(tensorName);\n        }\n        if (!buffer) throw new Error(`No GPU buffer for tensor \"${tensorName}\" in op ${node.id}`);\n\n        // Break read_write aliasing: if a writable binding would reuse a buffer\n        // already bound writable in this group (e.g. RoPE-Q-only nodes that list\n        // the same tensor as input+output and leave the k slot unused), bind a\n        // throwaway scratch buffer instead. Safe because the aliasing slot is not\n        // actually accessed at dispatch time (num_kv_heads=0 → no k work).\n        if (binding.type === \"storage-read-write\") {\n          if (writableBound.has(buffer)) {\n            buffer = this.getBindingScratchBuffer(buffer.size);\n          } else {\n            writableBound.add(buffer);\n          }\n        }\n\n        entries.push({ buffer });\n      }\n    }\n\n    return entries;\n  }\n\n  /** Lazily allocate a scratch storage buffer at least `minBytes` large. */\n  private getBindingScratchBuffer(minBytes: number): GPUBuffer {\n    if (!this.bindingScratchBuffer || this.bindingScratchBuffer.size < minBytes) {\n      this.bindingScratchBuffer = createStorageBuffer(\n        this.ctx,\n        \"binding_scratch\",\n        Math.max(minBytes, 256),\n      );\n    }\n    return this.bindingScratchBuffer;\n  }\n}\n","/**\n * GPTQ → Gerbil INT4 format adapter.\n *\n * Repacks GPTQ-quantized weights (qweight/qzeros/scales) into the flat\n * row-major packing expected by Gerbil's MatMulInt4 WGSL kernel.\n *\n * GPTQ layout (AutoGPTQ, per linear layer):\n *   qweight: Int32Array, shape [K/8, N] — 8 nibbles per int32 along INPUT dim (K)\n *   scales:  Float32Array, shape [K/group_size, N] — per-group per-output-column\n *   qzeros:  Int32Array, shape [K/group_size, N/8] — packed 4-bit zero points along N\n *\n * Gerbil layout (per linear layer):\n *   packed:  Uint32Array, flat [N*K/8] — 8 nibbles per u32 in row-major (N,K) order\n *   scales:  Float32Array, flat [N * K/group_size] — sequential in (N,K) row-major\n *   zeros:   Float32Array, flat [N * K/group_size]\n *\n * The kernel indexes: flat_idx = col * K + k, group = flat_idx / group_size\n */\n\nexport interface GPTQTensors {\n  qweight: Int32Array; // [K/8, N]\n  scales: Float32Array; // [K/group_size, N]\n  qzeros: Int32Array; // [K/group_size, N/8]\n  K: number;\n  N: number;\n  groupSize: number;\n}\n\nexport interface GerbilINT4Tensors {\n  packed: Uint32Array;\n  scales: Float32Array;\n  zeros: Float32Array;\n}\n\n/**\n * Repack GPTQ tensors into Gerbil's flat INT4 format.\n *\n * This is pure data reshuffling — no quantization math. The 4-bit values,\n * scales, and zero points are preserved exactly.\n */\nexport function repackGPTQ(gptq: GPTQTensors): GerbilINT4Tensors {\n  const { qweight, scales: gptqScales, qzeros, K, N, groupSize } = gptq;\n  const totalElements = N * K;\n  const groupsPerCol = Math.ceil(K / groupSize);\n  const totalGroups = N * groupsPerCol;\n\n  const packed = new Uint32Array(Math.ceil(totalElements / 8));\n  const scales = new Float32Array(totalGroups);\n  const zeros = new Float32Array(totalGroups);\n\n  // Repack nibbles: GPTQ [K/8, N] → Gerbil flat [N*K/8]\n  // GPTQ: qweight[(k >> 3) * N + col], nibble at (k & 7) * 4\n  // Gerbil: packed[(col*K+k)/8], nibble at ((col*K+k) % 8) * 4\n  for (let k = 0; k < K; k++) {\n    const gptqRow = k >>> 3;\n    const gptqNibblePos = k & 7;\n    const gptqRowBase = gptqRow * N;\n    for (let col = 0; col < N; col++) {\n      // Extract nibble from GPTQ\n      const gptqWord = qweight[gptqRowBase + col];\n      const nibble = (gptqWord >>> (gptqNibblePos * 4)) & 0xf;\n\n      // Write nibble into Gerbil flat layout\n      const flatIdx = col * K + k;\n      const packedIdx = flatIdx >>> 3;\n      const nibblePos = flatIdx & 7;\n      packed[packedIdx] |= nibble << (nibblePos * 4);\n    }\n  }\n\n  // Repack scales: GPTQ [groups_per_col, N] → Gerbil flat [N * groups_per_col]\n  // GPTQ: scales[g * N + col]\n  // Gerbil: scales[col * groups_per_col + g]\n  for (let g = 0; g < groupsPerCol; g++) {\n    const srcRow = g * N;\n    for (let col = 0; col < N; col++) {\n      scales[col * groupsPerCol + g] = gptqScales[srcRow + col];\n    }\n  }\n\n  // Repack zeros: GPTQ [groups_per_col, N/8] packed int4 → Gerbil flat float32\n  // GPTQ: qzeros[g * (N/8) + (col >> 3)], nibble at (col & 7) * 4\n  // Gerbil: zeros[col * groups_per_col + g] as float32\n  //\n  // GPTQ convention: stored qzeros = actual_zero_point - 1.\n  // All standard GPTQ kernels (exllama, marlin) add +1 during dequant.\n  // We bake the +1 into the stored value so the kernel formula stays simple:\n  //   dequant = (nibble - zero) * scale\n  const zerosNDiv8 = N >>> 3;\n  for (let g = 0; g < groupsPerCol; g++) {\n    const srcRow = g * zerosNDiv8;\n    for (let col = 0; col < N; col++) {\n      const zWord = qzeros[srcRow + (col >>> 3)];\n      const zNibblePos = col & 7;\n      const zeroVal = (zWord >>> (zNibblePos * 4)) & 0xf;\n      zeros[col * groupsPerCol + g] = zeroVal + 1; // +1: GPTQ packing convention\n    }\n  }\n\n  return { packed, scales, zeros };\n}\n","/**\n * IndexedDB-backed blob cache for model weights.\n *\n * Why IndexedDB and not the Cache API / OPFS we hand-rolled before:\n *   - The Cache API ignores URL fragments in `match()`, which silently collapsed\n *     our per-tensor `${url}#t:start-len` keys onto one entry → wrong bytes → the\n *     `RRRR` garbage and `wt.data` crashes. IDB keys on the exact string.\n *   - OPFS via Worker (`createSyncAccessHandle`) is finicky on iOS Safari (OOM /\n *     transient failures) and added a lot of moving parts.\n * IndexedDB is the boring, proven large-binary store every browser supports. It\n * keys on arbitrary strings (no fragment footgun), stores ArrayBuffers directly,\n * and is durable (subject to the same per-origin quota as Cache/OPFS — install to\n * Home Screen lifts that ceiling; IDB doesn't change it, it just makes caching\n * actually CORRECT).\n *\n * Inert on Node (no indexedDB) — every export resolves to a miss/no-op so the\n * heap load path is unchanged off the browser.\n */\n\nconst DB_NAME = \"gerbil-weights-v1\";\nconst STORE = \"tensors\";\n\nfunction idbAvailable(): boolean {\n  return typeof indexedDB !== \"undefined\";\n}\n\nlet _dbPromise: Promise<IDBDatabase | null> | null = null;\n\nfunction openDB(): Promise<IDBDatabase | null> {\n  if (_dbPromise) return _dbPromise;\n  _dbPromise = new Promise<IDBDatabase | null>((resolve) => {\n    if (!idbAvailable()) {\n      resolve(null);\n      return;\n    }\n    try {\n      const req = indexedDB.open(DB_NAME, 1);\n      req.onupgradeneeded = () => {\n        const db = req.result;\n        if (!db.objectStoreNames.contains(STORE)) db.createObjectStore(STORE);\n      };\n      req.onsuccess = () => resolve(req.result);\n      req.onerror = () => resolve(null);\n      req.onblocked = () => resolve(null);\n    } catch {\n      resolve(null);\n    }\n  });\n  return _dbPromise;\n}\n\n/** A single transaction operation, promisified. Resolves the fallback on any error. */\nfunction tx<T>(\n  mode: IDBTransactionMode,\n  run: (store: IDBObjectStore) => IDBRequest,\n  fallback: T,\n): Promise<T> {\n  return openDB().then(\n    (db) =>\n      new Promise<T>((resolve) => {\n        if (!db) {\n          resolve(fallback);\n          return;\n        }\n        try {\n          const t = db.transaction(STORE, mode);\n          const req = run(t.objectStore(STORE));\n          req.onsuccess = () => resolve((req.result as T) ?? fallback);\n          req.onerror = () => resolve(fallback);\n          t.onabort = () => resolve(fallback);\n        } catch {\n          resolve(fallback);\n        }\n      }),\n  );\n}\n\n/** Read a key's bytes, or null on miss / no IDB / any error. */\nexport async function idbGet(key: string): Promise<ArrayBuffer | null> {\n  const v = await tx<ArrayBuffer | null>(\"readonly\", (s) => s.get(key), null);\n  return v instanceof ArrayBuffer ? v : null;\n}\n\n/** Cheap presence check (reads only the key via getKey — no body materialized). */\nexport async function idbHas(key: string): Promise<boolean> {\n  const k = await tx<IDBValidKey | undefined>(\"readonly\", (s) => s.getKey(key), undefined);\n  return k !== undefined;\n}\n\n/**\n * All cached keys, as a Set, in ONE transaction. Probing presence per-tensor\n * (hundreds of parallel getKey transactions) storms Safari's IndexedDB and hangs\n * the load; getAllKeys() does it in a single transaction so callers can check\n * membership in memory. Empty set on miss / no IDB / error.\n */\nexport async function idbAllKeys(): Promise<Set<string>> {\n  const keys = await tx<IDBValidKey[]>(\"readonly\", (s) => s.getAllKeys(), []);\n  return new Set((keys ?? []).map((k) => String(k)));\n}\n\n/** Write a key's bytes. Returns true on success, false on quota/error (caller\n * falls back to network — never throws). */\nexport async function idbPut(key: string, buf: ArrayBuffer): Promise<boolean> {\n  // put returns the key on success; map to boolean.\n  const r = await tx<IDBValidKey | null>(\"readwrite\", (s) => s.put(buf, key), null);\n  return r !== null;\n}\n\n/** Delete a key (best-effort). */\nexport async function idbDelete(key: string): Promise<void> {\n  await tx<unknown>(\"readwrite\", (s) => s.delete(key), null);\n}\n\n/** True when IndexedDB is usable in this environment (browser, not Node). */\nexport function idbCacheAvailable(): boolean {\n  return idbAvailable();\n}\n","/**\n * MLX 4-bit → Gerbil INT4 format adapter.\n *\n * Converts MLX quantized weights (weight/scales/biases) into the flat\n * row-major packing expected by Gerbil's MatMulInt4 / EmbeddingInt4 kernels.\n *\n * MLX layout (per quantized layer):\n *   weight: Uint32Array, shape [N, K/8] — 8 nibbles per u32, packed along K (last axis)\n *   scales: Float32Array, shape [N, K/group_size] — per-group scale factors\n *   biases: Float32Array, shape [N, K/group_size] — per-group bias values\n *\n * MLX dequant formula (affine mode):\n *   w = scale * nibble + bias\n *\n * Gerbil dequant formula:\n *   w = (nibble - zero) * scale\n *\n * Conversion: scale_gerbil = scale_mlx, zero_gerbil = -bias_mlx / scale_mlx\n *\n * MLX [N, K/8] is already row-major with nibbles packed along K — same layout\n * as Gerbil flat [N*K/8]. The nibble data is directly usable; only scales/zeros\n * need conversion.\n */\n\nimport type { GerbilINT4Tensors } from \"./gptq-adapter.js\";\n\nexport interface MLXTensors {\n  weight: Uint32Array; // [N, K/8] packed 4-bit\n  scales: Float32Array; // [N, K/group_size]\n  biases: Float32Array; // [N, K/group_size]\n  N: number;\n  K: number;\n  groupSize: number;\n}\n\n/**\n * Convert MLX 4-bit tensors into Gerbil's flat INT4 format.\n *\n * The packed nibbles are already in the correct layout — MLX [N, K/8] row-major\n * matches Gerbil flat [N*K/8]. We just convert the affine scale/bias to\n * Gerbil's (nibble - zero) * scale convention.\n */\nexport function repackMLX(mlx: MLXTensors): GerbilINT4Tensors {\n  const { weight, scales: mlxScales, biases: mlxBiases, N, K, groupSize } = mlx;\n  const groupsPerRow = Math.ceil(K / groupSize);\n  const totalGroups = N * groupsPerRow;\n\n  // Nibbles: MLX [N, K/8] is already flat [N*K/8] in row-major order\n  // The packing is identical: 8 nibbles per u32, little-nibble-first\n  const packed = weight;\n\n  // Convert affine scale/bias to Gerbil scale/zero\n  // MLX: w = scale * nibble + bias\n  // Gerbil: w = (nibble - zero) * scale\n  // Therefore: zero = -bias / scale, scale stays the same\n  const scales = new Float32Array(totalGroups);\n  const zeros = new Float32Array(totalGroups);\n\n  // MLX stores [N, groups_per_row] which matches Gerbil flat [N * groups_per_row]\n  // Both are row-major with groups sequential per output row — same layout.\n  for (let i = 0; i < totalGroups; i++) {\n    const s = mlxScales[i];\n    const b = mlxBiases[i];\n    scales[i] = s;\n    // When scale is zero, bias should also be zero (constant group)\n    zeros[i] = s !== 0 ? -b / s : 0;\n  }\n\n  return { packed, scales, zeros };\n}\n","/**\n * OPFS-backed blob cache for durable model-weight persistence.\n *\n * WHY THIS EXISTS\n * ---------------\n * On iOS Safari the CacheStorage API (`caches.open(...)`) used for model weights\n * is evicted under quota pressure in a plain tab — `navigator.storage.persist()`\n * is not granted outside an installed PWA — so the ~404 MB model is re-downloaded\n * on every visit. The Origin Private File System (OPFS) is the durable path, but:\n *\n *  • Main-thread OPFS `createWritable` throws \"out of quota\" mid-write on iOS\n *    Safari and can leave orphaned, unclearable bytes.\n *  • The ONLY iOS-safe OPFS write path is `FileSystemSyncAccessHandle`\n *    (`createSyncAccessHandle`), which is available ONLY inside a Worker.\n *\n * So this module runs all OPFS I/O on a dedicated Worker created from an inlined\n * blob URL (no build wiring, zero extra deps). The main thread talks to it over\n * postMessage; payload ArrayBuffers are TRANSFERRED (zero-copy) so a write does\n * not double peak memory. Writes are chunked inside the worker so the sync handle\n * never has to hold the whole tensor at once.\n *\n * KEY → FILENAME\n * --------------\n * Cache keys are URLs with `/`, `#`, `?` etc. (e.g.\n * `https://huggingface.co/...model.safetensors#t:8388608-2359296`). OPFS file\n * names cannot contain those, so each key is mapped to a stable, collision-\n * resistant filename via a 53-bit FNV-style hash plus the key length. Files live\n * in a versioned OPFS subdirectory (see OPFS_DIR_NAME).\n *\n * DESKTOP / NODE\n * --------------\n * Everything here is gated by `opfsAvailable()`, which is false in Node (no\n * `navigator.storage.getDirectory`, no `Worker`, no `Blob`). Callers must short-\n * circuit on that before touching any export here. The Node weight path never\n * reaches this module.\n */\n\n/** Versioned OPFS directory. Bump when the on-disk layout/semantics change. */\nconst OPFS_DIR_NAME = \"gerbil-models-v4\";\n\n/** Per-message write chunk inside the worker (bounded peak memory). */\nconst WRITE_CHUNK_BYTES = 8 * 1024 * 1024;\n\n/**\n * True when the durable OPFS+Worker path is usable. False in Node (no navigator/\n * Worker/Blob) and in browsers without OPFS or Workers. Cheap and synchronous so\n * the seam functions can bail before any async work.\n */\nexport function opfsAvailable(): boolean {\n  return (\n    typeof navigator !== \"undefined\" &&\n    typeof (navigator as { storage?: { getDirectory?: unknown } }).storage?.getDirectory ===\n      \"function\" &&\n    typeof Worker !== \"undefined\" &&\n    typeof Blob !== \"undefined\" &&\n    typeof URL !== \"undefined\" &&\n    typeof URL.createObjectURL === \"function\"\n  );\n}\n\n/**\n * Map an arbitrary cache key to a safe OPFS filename. FNV-1a over UTF-16 code\n * units folded into a 53-bit safe integer, combined with the key length so two\n * different keys are astronomically unlikely to collide. Deterministic across\n * reloads (this is what makes a tensor written last visit re-resolve this visit).\n */\nfunction keyToFileName(key: string): string {\n  // FNV-1a (32-bit) mixed twice for a wider hash; emit two hex halves + length.\n  let h1 = 0x81_1c_9d_c5;\n  let h2 = 0x01_00_01_93;\n  for (let i = 0; i < key.length; i++) {\n    const c = key.charCodeAt(i);\n    h1 ^= c;\n    h1 = Math.imul(h1, 0x01_00_01_93);\n    h2 ^= c + i;\n    h2 = Math.imul(h2, 0x85_eb_ca_6b);\n  }\n  const a = (h1 >>> 0).toString(16).padStart(8, \"0\");\n  const b = (h2 >>> 0).toString(16).padStart(8, \"0\");\n  return `${a}${b}-${key.length.toString(16)}.bin`;\n}\n\n// ── Inline worker source ─────────────────────────────────────────────────────\n//\n// Self-contained worker that owns one OPFS directory handle and performs\n// has/read/write/delete via createSyncAccessHandle. Kept as a string so the\n// existing browser tsdown entry emits it inline (no separate worker bundle).\n// The worker has no imports — everything it needs is inlined here.\nconst WORKER_SOURCE = `\nconst DIR_NAME = ${JSON.stringify(OPFS_DIR_NAME)};\nlet dirPromise = null;\n\nfunction getDir() {\n  if (!dirPromise) {\n    dirPromise = navigator.storage\n      .getDirectory()\n      .then((root) => root.getDirectoryHandle(DIR_NAME, { create: true }));\n  }\n  return dirPromise;\n}\n\nasync function handleHas(name) {\n  const dir = await getDir();\n  try {\n    const fh = await dir.getFileHandle(name, { create: false });\n    const access = await fh.createSyncAccessHandle();\n    const size = access.getSize();\n    access.close();\n    // A zero-byte file is a half-written/aborted entry — treat as absent.\n    return size > 0;\n  } catch {\n    return false;\n  }\n}\n\nasync function handleRead(name) {\n  const dir = await getDir();\n  const fh = await dir.getFileHandle(name, { create: false });\n  const access = await fh.createSyncAccessHandle();\n  try {\n    const size = access.getSize();\n    if (size === 0) {\n      access.close();\n      return null;\n    }\n    const buf = new ArrayBuffer(size);\n    access.read(new Uint8Array(buf), { at: 0 });\n    access.close();\n    return buf;\n  } catch (e) {\n    try { access.close(); } catch {}\n    throw e;\n  }\n}\n\nasync function handleWrite(name, buf) {\n  const dir = await getDir();\n  const fh = await dir.getFileHandle(name, { create: true });\n  const access = await fh.createSyncAccessHandle();\n  try {\n    access.truncate(0);\n    const total = buf.byteLength;\n    const view = new Uint8Array(buf);\n    const CHUNK = ${WRITE_CHUNK_BYTES};\n    let off = 0;\n    while (off < total) {\n      const end = Math.min(off + CHUNK, total);\n      // Subarray is a zero-copy view; write returns bytes written.\n      access.write(view.subarray(off, end), { at: off });\n      off = end;\n    }\n    access.flush();\n    const finalSize = access.getSize();\n    access.close();\n    if (finalSize !== total) {\n      // Write came up short (quota) — remove the partial so it never reads back\n      // as a valid (wrong-size) entry.\n      try { await dir.removeEntry(name); } catch {}\n      throw new Error(\"OPFS short write: \" + finalSize + \"/\" + total);\n    }\n  } catch (e) {\n    try { access.close(); } catch {}\n    // Best-effort cleanup of any partial file.\n    try { await dir.removeEntry(name); } catch {}\n    throw e;\n  }\n}\n\nasync function handleDelete(name) {\n  const dir = await getDir();\n  try { await dir.removeEntry(name); } catch {}\n}\n\nself.onmessage = async (ev) => {\n  const { id, op, name, buf } = ev.data;\n  try {\n    if (op === \"has\") {\n      const present = await handleHas(name);\n      self.postMessage({ id, ok: true, present });\n    } else if (op === \"read\") {\n      const out = await handleRead(name);\n      if (out) self.postMessage({ id, ok: true, buf: out }, [out]);\n      else self.postMessage({ id, ok: true, buf: null });\n    } else if (op === \"write\") {\n      await handleWrite(name, buf);\n      self.postMessage({ id, ok: true });\n    } else if (op === \"delete\") {\n      await handleDelete(name);\n      self.postMessage({ id, ok: true });\n    } else {\n      self.postMessage({ id, ok: false, error: \"unknown op: \" + op });\n    }\n  } catch (e) {\n    self.postMessage({ id, ok: false, error: String((e && e.message) || e) });\n  }\n};\n`;\n\ninterface PendingResolver {\n  resolve: (value: { present?: boolean; buf?: ArrayBuffer | null }) => void;\n  reject: (err: Error) => void;\n}\n\n/**\n * Lazily-created singleton worker client. Created on first use; if creation or\n * any handshake fails the whole OPFS path is marked unavailable so callers fall\n * back to CacheStorage cleanly.\n */\nclass OpfsClient {\n  private worker: Worker | null = null;\n  private blobUrl: string | null = null;\n  private pending = new Map<number, PendingResolver>();\n  private nextId = 1;\n  private broken = false;\n\n  private ensureWorker(): Worker | null {\n    if (this.broken) return null;\n    if (this.worker) return this.worker;\n    try {\n      const blob = new Blob([WORKER_SOURCE], { type: \"application/javascript\" });\n      this.blobUrl = URL.createObjectURL(blob);\n      const worker = new Worker(this.blobUrl);\n      worker.onmessage = (ev: MessageEvent) => {\n        const { id, ok, present, buf, error } = ev.data as {\n          id: number;\n          ok: boolean;\n          present?: boolean;\n          buf?: ArrayBuffer | null;\n          error?: string;\n        };\n        const p = this.pending.get(id);\n        if (!p) return;\n        this.pending.delete(id);\n        if (ok) p.resolve({ present, buf });\n        else p.reject(new Error(error || \"OPFS worker error\"));\n      };\n      worker.onerror = () => {\n        // Fatal worker error — reject everything in flight and disable OPFS.\n        this.broken = true;\n        for (const p of this.pending.values()) {\n          p.reject(new Error(\"OPFS worker crashed\"));\n        }\n        this.pending.clear();\n      };\n      this.worker = worker;\n      return worker;\n    } catch {\n      this.broken = true;\n      return null;\n    }\n  }\n\n  private call(\n    op: string,\n    name: string,\n    buf?: ArrayBuffer,\n  ): Promise<{ present?: boolean; buf?: ArrayBuffer | null }> {\n    const worker = this.ensureWorker();\n    if (!worker) return Promise.reject(new Error(\"OPFS unavailable\"));\n    const id = this.nextId++;\n    return new Promise((resolve, reject) => {\n      this.pending.set(id, { resolve, reject });\n      try {\n        if (buf) {\n          // Transfer the buffer (zero-copy). The caller must not reuse it after.\n          worker.postMessage({ id, op, name, buf }, [buf]);\n        } else {\n          worker.postMessage({ id, op, name });\n        }\n      } catch (e) {\n        this.pending.delete(id);\n        reject(e instanceof Error ? e : new Error(String(e)));\n      }\n    });\n  }\n\n  async has(key: string): Promise<boolean> {\n    const { present } = await this.call(\"has\", keyToFileName(key));\n    return Boolean(present);\n  }\n\n  async read(key: string): Promise<ArrayBuffer | null> {\n    const { buf } = await this.call(\"read\", keyToFileName(key));\n    return buf ?? null;\n  }\n\n  async write(key: string, buf: ArrayBuffer): Promise<void> {\n    await this.call(\"write\", keyToFileName(key), buf);\n  }\n\n  async delete(key: string): Promise<void> {\n    await this.call(\"delete\", keyToFileName(key));\n  }\n}\n\nlet _client: OpfsClient | null = null;\nfunction client(): OpfsClient {\n  if (!_client) _client = new OpfsClient();\n  return _client;\n}\n\n/**\n * Presence check for a key in the OPFS store. Returns false (never throws) so the\n * caller can fall through to CacheStorage on any error.\n */\nexport async function opfsHas(key: string): Promise<boolean> {\n  if (!opfsAvailable()) return false;\n  try {\n    return await client().has(key);\n  } catch {\n    return false;\n  }\n}\n\n/**\n * Read a key's bytes from OPFS. Returns null on miss or any error (caller falls\n * back to CacheStorage / network).\n */\nexport async function opfsRead(key: string): Promise<ArrayBuffer | null> {\n  if (!opfsAvailable()) return null;\n  try {\n    return await client().read(key);\n  } catch {\n    return null;\n  }\n}\n\n/**\n * Write a key's bytes to OPFS durably. The buffer is TRANSFERRED to the worker\n * (do not reuse `buf` after calling — pass a copy if you still need it). Returns\n * true on success, false on any failure (quota, worker crash, …) so the caller\n * can fall back to CacheStorage without crashing the load.\n */\nexport async function opfsWrite(key: string, buf: ArrayBuffer): Promise<boolean> {\n  if (!opfsAvailable()) return false;\n  try {\n    await client().write(key, buf);\n    return true;\n  } catch {\n    return false;\n  }\n}\n\n/** Delete a key from OPFS (best-effort, never throws). */\nexport async function opfsDelete(key: string): Promise<void> {\n  if (!opfsAvailable()) return;\n  try {\n    await client().delete(key);\n  } catch {\n    /* best-effort */\n  }\n}\n\n/**\n * Delete superseded OPFS model directories (old namespaces, e.g. a previous\n * `gerbil-models-v3`) to reclaim quota. On iOS the per-origin quota is ~1 GB and\n * orphaned namespaces can fill it, blocking the current model from caching at all\n * (every write then fails with quota/OOM → redownload every load). Directory\n * removal works on the MAIN thread — only createSyncAccessHandle needs a Worker —\n * so this is cheap and best-effort. Inert on Node (no navigator.storage).\n */\nexport async function opfsEvictStale(): Promise<void> {\n  try {\n    const root = (await (\n      navigator as { storage?: { getDirectory?: () => Promise<unknown> } }\n    )?.storage?.getDirectory?.()) as\n      | {\n          entries?: () => AsyncIterable<[string, { kind: string }]>;\n          removeEntry: (n: string, o?: { recursive?: boolean }) => Promise<void>;\n        }\n      | undefined;\n    if (!root) return;\n    const stale: string[] = [];\n    for await (const [name, handle] of root.entries?.() ?? []) {\n      if (\n        handle.kind === \"directory\" &&\n        name.startsWith(\"gerbil-models-\") &&\n        name !== OPFS_DIR_NAME\n      ) {\n        stale.push(name);\n      }\n    }\n    for (const name of stale) {\n      await root.removeEntry(name, { recursive: true }).catch(() => {\n        /* best-effort */\n      });\n    }\n  } catch {\n    /* best-effort */\n  }\n}\n\n/**\n * One-time health probe: actually write a few bytes to OPFS, read them back,\n * verify they round-trip, then clean up. Memoized. iOS Safari can report OPFS as\n * \"available\" (getDirectory/Worker/Blob all present) yet have the Worker's\n * createSyncAccessHandle writes silently fail — in which case we must NOT trust\n * OPFS, because the per-tensor write path transfers (detaches) the buffer and a\n * failed OPFS write would then have no CacheStorage fallback (→ nothing cached →\n * redownload every load). Callers gate OPFS use on THIS, not the cheap synchronous\n * opfsAvailable(). Resolves false if OPFS can't actually persist a round-trip.\n */\nlet _healthy: Promise<boolean> | null = null;\nexport function opfsHealthy(): Promise<boolean> {\n  if (_healthy) return _healthy;\n  _healthy = (async () => {\n    if (!opfsAvailable()) return false;\n    try {\n      const key = \"__opfs_health_probe__\";\n      const ok = await opfsWrite(key, new Uint8Array([7, 13, 42, 99]).buffer);\n      if (!ok) return false;\n      const back = await opfsRead(key);\n      await opfsDelete(key);\n      if (!back || back.byteLength !== 4) return false;\n      const b = new Uint8Array(back);\n      return b[0] === 7 && b[1] === 13 && b[2] === 42 && b[3] === 99;\n    } catch {\n      return false;\n    }\n  })();\n  return _healthy;\n}\n","/**\n * Safetensors binary format parser.\n *\n * Parses the HuggingFace safetensors format into typed array views\n * over the raw buffer — zero-copy where possible.\n */\n\nexport interface SafetensorEntry {\n  /** Tensor name as it appears in the file (e.g. \"model.layers.0.self_attn.q_proj.weight\"). */\n  name: string;\n  /** Data type string from the header. */\n  dtype: SafetensorDType;\n  /** Shape dimensions. */\n  shape: number[];\n  /** Byte offset of this tensor's data relative to the start of the data section. */\n  dataOffset: number;\n  /** Byte length of this tensor's data. */\n  dataLength: number;\n}\n\nexport type SafetensorDType =\n  | \"F16\"\n  | \"F32\"\n  | \"BF16\"\n  | \"I32\"\n  | \"I8\"\n  | \"U8\"\n  | \"F64\"\n  | \"BOOL\"\n  | \"I16\"\n  | \"U16\"\n  | \"U32\"\n  | \"I64\"\n  | \"U64\";\n\nexport interface SafetensorsFile {\n  /** Header length in bytes. */\n  headerLength: number;\n  /** Byte offset where tensor data begins (8 + headerLength). */\n  dataStart: number;\n  /** All tensor entries (excludes __metadata__). */\n  entries: SafetensorEntry[];\n  /** Optional metadata from __metadata__ key. */\n  metadata: Record<string, string> | null;\n}\n\n/**\n * Parse safetensors header from an ArrayBuffer.\n *\n * Can be called with just the header bytes (for streaming — parse header first,\n * then fetch tensor data by offset) or with the entire file.\n */\nexport function parseSafetensorsHeader(buffer: ArrayBuffer): SafetensorsFile {\n  if (buffer.byteLength < 8) {\n    throw new Error(\n      `Buffer too small to contain safetensors header length (need 8 bytes, got ${buffer.byteLength}).`,\n    );\n  }\n\n  const view = new DataView(buffer);\n\n  // First 8 bytes: header length as little-endian uint64.\n  // JS doesn't have native uint64, but headers are always < 4 GB.\n  const headerLength = Number(view.getBigUint64(0, true));\n\n  if (headerLength > buffer.byteLength - 8) {\n    throw new Error(\n      `Safetensors header length (${headerLength}) exceeds buffer size (${buffer.byteLength - 8}). ` +\n        `Pass at least the first ${headerLength + 8} bytes.`,\n    );\n  }\n\n  // Decode the JSON header\n  const headerBytes = new Uint8Array(buffer, 8, headerLength);\n  const headerStr = new TextDecoder().decode(headerBytes);\n  const header: Record<string, unknown> = JSON.parse(headerStr);\n\n  const dataStart = 8 + headerLength;\n  const entries: SafetensorEntry[] = [];\n  let metadata: Record<string, string> | null = null;\n\n  for (const [name, info] of Object.entries(header)) {\n    if (name === \"__metadata__\") {\n      metadata = info as Record<string, string>;\n      continue;\n    }\n\n    const { dtype, shape, data_offsets } = info as {\n      dtype: SafetensorDType;\n      shape: number[];\n      data_offsets: [number, number];\n    };\n\n    entries.push({\n      name,\n      dtype,\n      shape,\n      dataOffset: data_offsets[0],\n      dataLength: data_offsets[1] - data_offsets[0],\n    });\n  }\n\n  // Sort by offset for sequential access patterns\n  entries.sort((a, b) => a.dataOffset - b.dataOffset);\n\n  return { headerLength, dataStart, entries, metadata };\n}\n\n/** Byte alignment required for each dtype. */\nfunction dtypeAlignment(dtype: SafetensorDType): number {\n  switch (dtype) {\n    case \"F64\":\n    case \"I64\":\n    case \"U64\":\n      return 8;\n    case \"F32\":\n    case \"I32\":\n    case \"U32\":\n      return 4;\n    case \"F16\":\n    case \"BF16\":\n    case \"I16\":\n    case \"U16\":\n      return 2;\n    case \"I8\":\n    case \"U8\":\n    case \"BOOL\":\n      return 1;\n    default:\n      return 1;\n  }\n}\n\n/**\n * Create a typed array for the given dtype.\n *\n * If `offset` is aligned to the element size, returns a zero-copy view\n * into `buffer`. Otherwise, copies the relevant slice into a new\n * properly-aligned ArrayBuffer.\n */\nfunction makeTypedView(\n  buffer: ArrayBuffer,\n  offset: number,\n  byteLength: number,\n  dtype: SafetensorDType,\n): ArrayBufferView {\n  const align = dtypeAlignment(dtype);\n  const aligned = offset % align === 0;\n\n  // For single-byte types, alignment is always satisfied.\n  const src = aligned ? buffer : buffer.slice(offset, offset + byteLength);\n  const base = aligned ? offset : 0;\n\n  switch (dtype) {\n    case \"F32\":\n      return new Float32Array(src, base, byteLength / 4);\n    case \"F16\":\n      // F16 has no native TypedArray — return Uint16Array (bitwise).\n      return new Uint16Array(src, base, byteLength / 2);\n    case \"BF16\":\n      // BF16 also has no native TypedArray — return Uint16Array (bitwise).\n      return new Uint16Array(src, base, byteLength / 2);\n    case \"I32\":\n      return new Int32Array(src, base, byteLength / 4);\n    case \"U32\":\n      return new Uint32Array(src, base, byteLength / 4);\n    case \"I8\":\n      return new Int8Array(src, base, byteLength);\n    case \"U8\":\n      return new Uint8Array(src, base, byteLength);\n    case \"I16\":\n      return new Int16Array(src, base, byteLength / 2);\n    case \"U16\":\n      return new Uint16Array(src, base, byteLength / 2);\n    case \"F64\":\n      return new Float64Array(src, base, byteLength / 8);\n    case \"I64\":\n      return new BigInt64Array(src, base, byteLength / 8);\n    case \"U64\":\n      return new BigUint64Array(src, base, byteLength / 8);\n    case \"BOOL\":\n      return new Uint8Array(src, base, byteLength);\n    default:\n      throw new Error(`Unsupported safetensors dtype: ${dtype}`);\n  }\n}\n\n/**\n * Get a typed array view for a tensor entry from a complete safetensors buffer.\n * Zero-copy when the byte offset is properly aligned; copies otherwise.\n */\nexport function getTensorData(\n  buffer: ArrayBuffer,\n  file: SafetensorsFile,\n  entry: SafetensorEntry,\n): ArrayBufferView {\n  const offset = file.dataStart + entry.dataOffset;\n  return makeTypedView(buffer, offset, entry.dataLength, entry.dtype);\n}\n\n/**\n * Find a tensor entry by name (exact match).\n */\nexport function findTensor(file: SafetensorsFile, name: string): SafetensorEntry | undefined {\n  return file.entries.find((e) => e.name === name);\n}\n\n/**\n * Parse just the header from a Response (streaming-friendly).\n *\n * Reads the full response into memory and parses the header.\n * Returns the parsed header and the full buffer for tensor access.\n */\nexport async function parseSafetensorsFromResponse(\n  response: Response,\n): Promise<{ file: SafetensorsFile; fullBuffer: ArrayBuffer }> {\n  const fullBuffer = await response.arrayBuffer();\n  const file = parseSafetensorsHeader(fullBuffer);\n  return { file, fullBuffer };\n}\n\n/**\n * Compute the total byte size of all tensor data in a safetensors file.\n */\nexport function totalTensorBytes(file: SafetensorsFile): number {\n  let total = 0;\n  for (const entry of file.entries) {\n    total += entry.dataLength;\n  }\n  return total;\n}\n\n/**\n * Convert BF16 (bfloat16) data to F32.\n * BF16 is the upper 16 bits of an IEEE 754 float32, so conversion is a simple left-shift.\n */\nexport function bf16ToF32(bf16Bytes: Uint8Array): Float32Array {\n  const bf16 = new Uint16Array(bf16Bytes.buffer, bf16Bytes.byteOffset, bf16Bytes.byteLength / 2);\n  const f32 = new Float32Array(bf16.length);\n  const u32 = new Uint32Array(f32.buffer);\n  for (let i = 0; i < bf16.length; i++) {\n    u32[i] = bf16[i] << 16;\n  }\n  return f32;\n}\n\n/**\n * Convert F16 (IEEE 754 half-precision) data to F32.\n */\nexport function f16ToF32(f16View: ArrayBufferView): Float32Array {\n  const u16 =\n    f16View instanceof Uint16Array\n      ? f16View\n      : new Uint16Array(f16View.buffer, f16View.byteOffset, f16View.byteLength / 2);\n  const f32 = new Float32Array(u16.length);\n  for (let i = 0; i < u16.length; i++) {\n    const h = u16[i];\n    const sign = (h >> 15) & 1;\n    const exp = (h >> 10) & 0x1f;\n    const frac = h & 0x3ff;\n    if (exp === 0) {\n      // Subnormal or zero\n      f32[i] = (sign ? -1 : 1) * 2 ** -14 * (frac / 1024);\n    } else if (exp === 31) {\n      // Inf or NaN\n      f32[i] = frac === 0 ? (sign ? -Infinity : Infinity) : NaN;\n    } else {\n      f32[i] = (sign ? -1 : 1) * 2 ** (exp - 15) * (1 + frac / 1024);\n    }\n  }\n  return f32;\n}\n","/**\n * Pure JavaScript BPE tokenizer.\n *\n * Reads HuggingFace tokenizer.json — no WASM, no external dependencies.\n * Supports encoding, decoding, and chat template application.\n */\n\n// Byte-level BPE unicode mapping (same as HuggingFace's bytes_to_unicode).\n// Maps each byte (0-255) to a unique unicode codepoint such that printable\n// ASCII characters map to themselves, and non-printable bytes are shifted\n// into the U+0100+ range.\nconst BYTE_TO_UNICODE = new Map<number, number>();\nconst UNICODE_TO_BYTE = new Map<number, number>();\n\n{\n  // Printable bytes that map to themselves: 33-126 (! through ~), 161-172, 174-255\n  const bs: number[] = [];\n  for (let i = 33; i <= 126; i++) bs.push(i); // ! through ~\n  for (let i = 161; i <= 172; i++) bs.push(i); // ¡ through ¬\n  for (let i = 174; i <= 255; i++) bs.push(i); // ® through ÿ\n\n  const cs = [...bs]; // unicode codepoints (same for these ranges)\n\n  // Non-printable bytes get shifted to U+0100+\n  let n = 0;\n  for (let b = 0; b < 256; b++) {\n    if (!bs.includes(b)) {\n      bs.push(b);\n      cs.push(256 + n);\n      n++;\n    }\n  }\n\n  for (let i = 0; i < bs.length; i++) {\n    BYTE_TO_UNICODE.set(bs[i], cs[i]);\n    UNICODE_TO_BYTE.set(cs[i], bs[i]);\n  }\n}\n\nexport interface TokenizerConfig {\n  bosToken: string | null;\n  eosToken: string | null;\n  bosTokenId: number | null;\n  eosTokenId: number | null;\n  chatTemplate: string | null;\n  addBosToken: boolean;\n  addEosToken: boolean;\n}\n\nexport interface ChatMessage {\n  role: \"system\" | \"user\" | \"assistant\";\n  content: string;\n}\n\nexport class Tokenizer {\n  private vocab: Map<string, number>; // token string → id\n  private vocabReverse: Map<number, string>; // id → token string\n  private merges: Map<string, number>; // \"tokenA tokenB\" → priority (lower = merge first)\n  private specialTokens: Map<string, number>; // special token string → id\n  private addedTokens: Map<string, number>; // ALL added tokens (special + non-special like <think>)\n  private byteFallback: Map<number, string>; // byte value → byte-level token\n  /**\n   * SentencePiece mode (Gemma/Llama-style). When true, vocab uses U+2581 (▁) for\n   * spaces and raw UTF-8 tokens (NOT the GPT-2 byte-to-unicode \"Ġ\" mapping), and\n   * raw bytes fall back to <0xHH> tokens. When false, GPT-2 byte-level BPE.\n   */\n  private spmMode: boolean;\n  readonly config: TokenizerConfig;\n  readonly vocabSize: number;\n\n  private constructor(\n    vocab: Map<string, number>,\n    vocabReverse: Map<number, string>,\n    merges: Map<string, number>,\n    specialTokens: Map<string, number>,\n    addedTokens: Map<string, number>,\n    config: TokenizerConfig,\n    vocabSize: number,\n    spmMode: boolean,\n  ) {\n    this.vocab = vocab;\n    this.vocabReverse = vocabReverse;\n    this.merges = merges;\n    this.specialTokens = specialTokens;\n    this.addedTokens = addedTokens;\n    this.byteFallback = new Map();\n    this.spmMode = spmMode;\n    this.config = config;\n    this.vocabSize = vocabSize;\n\n    // Build byte fallback map (for handling unknown characters)\n    // HF tokenizers use <0xHH> tokens for raw bytes\n    for (let b = 0; b < 256; b++) {\n      const hex = b.toString(16).toUpperCase().padStart(2, \"0\");\n      const key = `<0x${hex}>`;\n      if (vocab.has(key)) {\n        this.byteFallback.set(b, key);\n      }\n    }\n  }\n\n  /**\n   * Create a tokenizer from HuggingFace JSON files.\n   */\n  static fromJSON(tokenizerJSON: any, tokenizerConfigJSON?: any): Tokenizer {\n    const model = tokenizerJSON.model;\n    if (!model || model.type !== \"BPE\") {\n      throw new Error(\n        `Unsupported tokenizer type: ${model?.type ?? \"unknown\"}. Only BPE is supported.`,\n      );\n    }\n\n    // Build vocab\n    const vocab = new Map<string, number>();\n    const vocabReverse = new Map<number, string>();\n    for (const [token, id] of Object.entries(model.vocab as Record<string, number>)) {\n      vocab.set(token, id);\n      vocabReverse.set(id, token);\n    }\n\n    // Detect SentencePiece mode (Gemma/Llama): the normalizer replaces spaces\n    // with U+2581 (▁) and/or the vocab is dominated by ▁-prefixed tokens. In SPM\n    // mode tokens are raw UTF-8 with ▁ for spaces and bytes fall back to <0xHH>,\n    // rather than the GPT-2 byte-to-unicode (Ġ) mapping.\n    const normalizerReplacesSpace = (() => {\n      const norm = tokenizerJSON.normalizer;\n      const checkOne = (n: any) =>\n        n?.type === \"Replace\" && n?.pattern?.String === \" \" && n?.content === \"▁\";\n      if (!norm) return false;\n      if (checkOne(norm)) return true;\n      if (Array.isArray(norm.normalizers)) return norm.normalizers.some(checkOne);\n      return false;\n    })();\n    const byteFallbackFlag = model.byte_fallback === true;\n    const spmMode = normalizerReplacesSpace || (byteFallbackFlag && \"▁the\" in model.vocab);\n\n    // Build merge priority map. Merges are either \"a b\" strings (GPT-2 style) or\n    // [\"a\", \"b\"] arrays (newer/Gemma tokenizer.json). Normalize both to \"a b\".\n    const merges = new Map<string, number>();\n    const mergeList = (model.merges ?? []) as Array<string | [string, string]>;\n    for (let i = 0; i < mergeList.length; i++) {\n      const m = mergeList[i];\n      const key = Array.isArray(m) ? `${m[0]} ${m[1]}` : m;\n      merges.set(key, i);\n    }\n\n    // Register added tokens — ALL of them go into addedTokensMap for splitting,\n    // only special:true ones go into specialTokens for skip-during-decode\n    const specialTokens = new Map<string, number>();\n    const addedTokensMap = new Map<string, number>();\n    const addedTokensList = tokenizerJSON.added_tokens as\n      | Array<{ id: number; content: string; special?: boolean }>\n      | undefined;\n    if (addedTokensList) {\n      for (const t of addedTokensList) {\n        vocab.set(t.content, t.id);\n        vocabReverse.set(t.id, t.content);\n        addedTokensMap.set(t.content, t.id);\n        if (t.special) {\n          specialTokens.set(t.content, t.id);\n        }\n      }\n    }\n\n    // Parse config\n    const bosToken = tokenizerConfigJSON?.bos_token ?? null;\n    const eosToken = tokenizerConfigJSON?.eos_token ?? null;\n    const chatTemplate = tokenizerConfigJSON?.chat_template ?? null;\n    const addBosToken = tokenizerConfigJSON?.add_bos_token ?? false;\n    const addEosToken = tokenizerConfigJSON?.add_eos_token ?? false;\n\n    const config: TokenizerConfig = {\n      bosToken,\n      eosToken,\n      bosTokenId: bosToken ? (vocab.get(bosToken) ?? null) : null,\n      eosTokenId: eosToken ? (vocab.get(eosToken) ?? null) : null,\n      chatTemplate,\n      addBosToken,\n      addEosToken,\n    };\n\n    return new Tokenizer(\n      vocab,\n      vocabReverse,\n      merges,\n      specialTokens,\n      addedTokensMap,\n      config,\n      vocab.size,\n      spmMode,\n    );\n  }\n\n  /**\n   * Resolve a literal token string (e.g. \"<|endoftext|>\") to its vocab id,\n   * or null if it isn't in the vocabulary.\n   */\n  tokenToId(token: string): number | null {\n    return this.vocab.get(token) ?? null;\n  }\n\n  /**\n   * Encode text into token IDs.\n   */\n  encode(text: string): number[] {\n    if (!text) return [];\n\n    const ids: number[] = [];\n\n    // Add BOS token if configured\n    if (this.config.addBosToken && this.config.bosTokenId !== null) {\n      ids.push(this.config.bosTokenId);\n    }\n\n    // Check for special tokens first — split text around them\n    const parts = this.splitOnSpecialTokens(text);\n\n    for (const part of parts) {\n      if (part.special) {\n        const id = this.addedTokens.get(part.text);\n        if (id !== undefined) {\n          ids.push(id);\n        }\n      } else {\n        // Pre-tokenize: split into words/chunks\n        // Use a simple regex-based pre-tokenizer similar to GPT-style\n        const chunks = this.preTokenize(part.text);\n        for (const chunk of chunks) {\n          const chunkIds = this.bpeEncode(chunk);\n          ids.push(...chunkIds);\n        }\n      }\n    }\n\n    // Add EOS token if configured\n    if (this.config.addEosToken && this.config.eosTokenId !== null) {\n      ids.push(this.config.eosTokenId);\n    }\n\n    return ids;\n  }\n\n  /**\n   * Decode token IDs back to text.\n   */\n  decode(ids: number[], skipSpecialTokens: boolean = true): string {\n    const pieces: string[] = [];\n\n    for (const id of ids) {\n      const token = this.vocabReverse.get(id);\n      if (!token) continue;\n\n      // Skip special tokens if requested\n      if (skipSpecialTokens && this.specialTokens.has(token)) {\n        continue;\n      }\n\n      pieces.push(token);\n    }\n\n    if (this.spmMode) {\n      // SentencePiece (Gemma/Llama): tokens are raw UTF-8 with ▁ for spaces and\n      // <0xHH> byte-fallback pieces. Concatenate the <0xHH> bytes into proper\n      // UTF-8 runs (Fuse), decode, and turn ▁ back into spaces.\n      let out = \"\";\n      let byteRun: number[] = [];\n      const flushBytes = () => {\n        if (byteRun.length === 0) return;\n        out += new TextDecoder().decode(new Uint8Array(byteRun));\n        byteRun = [];\n      };\n      for (const piece of pieces) {\n        const m = /^<0x([0-9A-Fa-f]{2})>$/.exec(piece);\n        if (m) {\n          byteRun.push(Number.parseInt(m[1], 16));\n        } else {\n          flushBytes();\n          out += piece;\n        }\n      }\n      flushBytes();\n      return out.replace(/▁/g, \" \");\n    }\n\n    // Join and decode byte-level BPE tokens.\n    // GPT/Qwen tokenizers use a byte-to-unicode mapping where non-printable\n    // bytes are shifted into the U+0100+ range:\n    //   byte 0x00-0x20 → U+0100-0x0120 (e.g. 0x0A=\\n → U+010A=Ċ, 0x20=space → U+0120=Ġ)\n    //   byte 0x7F-0xA0 → U+0121-0x0142\n    // We decode by reversing this mapping back to raw bytes.\n    let text = pieces.join(\"\");\n\n    // Build the reverse byte-level mapping\n    const chars: number[] = [];\n    for (let i = 0; i < text.length; i++) {\n      const cp = text.codePointAt(i)!;\n      const byte = UNICODE_TO_BYTE.get(cp);\n      if (byte !== undefined) {\n        chars.push(byte);\n      } else {\n        // Multi-byte UTF-8: encode the codepoint as UTF-8 bytes\n        const encoded = new TextEncoder().encode(String.fromCodePoint(cp));\n        for (const b of encoded) chars.push(b);\n        if (cp > 0xffff) i++; // skip surrogate pair\n      }\n    }\n    text = new TextDecoder().decode(new Uint8Array(chars));\n\n    // Handle byte-level fallback tokens like <0x0A> → \\n\n    text = text.replace(/<0x([0-9A-Fa-f]{2})>/g, (_, hex) => {\n      return String.fromCharCode(parseInt(hex, 16));\n    });\n\n    return text;\n  }\n\n  /**\n   * Apply chat template to messages.\n   *\n   * For now, implements the common ChatML format used by Qwen models:\n   * <|im_start|>system\\n{content}<|im_end|>\\n\n   * <|im_start|>user\\n{content}<|im_end|>\\n\n   * <|im_start|>assistant\\n\n   *\n   * TODO: Parse Jinja2 templates from tokenizer_config.json for full generality.\n   */\n  /**\n   * Gemma 4 turn format: `<bos><|turn>user\\n{content}<turn|>\\n<|turn>model\\n`.\n   * Gemma has no \"system\" role, so a system message is folded into the next user\n   * turn (matching the reference chat template).\n   */\n  private applyGemmaTurnTemplate(messages: ChatMessage[], addGenPrompt: boolean): string {\n    const parts: string[] = [];\n    if (this.config.bosToken) parts.push(this.config.bosToken);\n    let systemText = \"\";\n    for (const msg of messages) {\n      if (msg.role === \"system\") {\n        systemText += `${msg.content}\\n\\n`;\n        continue;\n      }\n      const role = msg.role === \"assistant\" ? \"model\" : msg.role;\n      const content = role === \"user\" && systemText ? systemText + msg.content : msg.content;\n      if (role === \"user\") systemText = \"\";\n      parts.push(`<|turn>${role}\\n${content}<turn|>\\n`);\n    }\n    if (addGenPrompt) parts.push(\"<|turn>model\\n\");\n    return parts.join(\"\");\n  }\n\n  applyChatTemplate(messages: ChatMessage[], options?: { addGenerationPrompt?: boolean }): string {\n    const addGenPrompt = options?.addGenerationPrompt ?? true;\n    const parts: string[] = [];\n\n    // ── Gemma 4 turn format ──\n    // Detected by the turn-delimiter tokens `<|turn>`/`<turn|>` (unique to the\n    // Gemma 4 vocab, absent from the ChatML/Qwen vocabs handled below).\n    if (this.addedTokens.has(\"<|turn>\") && this.addedTokens.has(\"<turn|>\")) {\n      return this.applyGemmaTurnTemplate(messages, addGenPrompt);\n    }\n\n    // Models whose chat template prepends bos at the start (e.g. LFM2's\n    // `{{- bos_token -}}` ...). The hardcoded ChatML body is otherwise shared.\n    const template = this.config.chatTemplate;\n    if (template?.includes(\"bos_token\") && this.config.bosToken) {\n      parts.push(this.config.bosToken);\n    }\n\n    for (const msg of messages) {\n      parts.push(`<|im_start|>${msg.role}\\n${msg.content}<|im_end|>\\n`);\n    }\n\n    if (addGenPrompt) {\n      parts.push(`<|im_start|>assistant\\n`);\n      // Some reasoning models (e.g. Qwen3/3.5) pre-fill an (empty) think block\n      // after the generation prompt for non-thinking mode. Only inject it when\n      // the model's own chat template actually emits `<think>` at generation\n      // time — otherwise models that merely *have* a <think> token but don't\n      // pre-fill it (e.g. LFM2.5) would be corrupted into degenerate output.\n      const templateWantsThink = template\n        ? template.includes(\"<think>\")\n        : this.addedTokens.has(\"<think>\");\n      if (templateWantsThink) {\n        parts.push(`<think>\\n\\n</think>\\n\\n`);\n      }\n    }\n\n    return parts.join(\"\");\n  }\n\n  /**\n   * Encode a chat conversation into token IDs.\n   */\n  encodeChat(messages: ChatMessage[], options?: { addGenerationPrompt?: boolean }): number[] {\n    const text = this.applyChatTemplate(messages, options);\n    return this.encode(text);\n  }\n\n  // ── Internal BPE implementation ──────────────────────────────────\n\n  private splitOnSpecialTokens(text: string): Array<{ text: string; special: boolean }> {\n    if (this.addedTokens.size === 0) return [{ text, special: false }];\n\n    const parts: Array<{ text: string; special: boolean }> = [];\n    // Sort added tokens by length (longest first) for greedy matching\n    // Uses ALL added tokens (not just special:true) so <think>/<|tool_call|> etc. are split\n    const sortedSpecials = [...this.addedTokens.keys()].sort((a, b) => b.length - a.length);\n\n    // Build regex that matches any special token\n    const escaped = sortedSpecials.map((s) => s.replace(/[.*+?^${}()|[\\]\\\\]/g, \"\\\\$&\"));\n    const regex = new RegExp(`(${escaped.join(\"|\")})`, \"g\");\n\n    let lastIndex = 0;\n    let match: RegExpExecArray | null;\n    while ((match = regex.exec(text)) !== null) {\n      if (match.index > lastIndex) {\n        parts.push({\n          text: text.slice(lastIndex, match.index),\n          special: false,\n        });\n      }\n      parts.push({ text: match[0], special: true });\n      lastIndex = regex.lastIndex;\n    }\n    if (lastIndex < text.length) {\n      parts.push({ text: text.slice(lastIndex), special: false });\n    }\n\n    return parts;\n  }\n\n  private preTokenize(text: string): string[] {\n    if (!text) return [];\n\n    if (this.spmMode) {\n      // SentencePiece (Gemma/Llama): normalize spaces → ▁ (U+2581), then split on\n      // the ▁ boundary keeping it with the FOLLOWING piece (Split \" \"\n      // MergedWithPrevious on the original, equivalent to ▁-prefixed words).\n      // BPE then runs on the raw ▁-encoded text — no byte-to-unicode mapping.\n      const encoded = text.replace(/ /g, \"▁\");\n      const chunks: string[] = [];\n      let i = 0;\n      while (i < encoded.length) {\n        let j = i + 1;\n        while (j < encoded.length && encoded[j] !== \"▁\") j++;\n        chunks.push(encoded.slice(i, j));\n        i = j;\n      }\n      return chunks;\n    }\n\n    // GPT-style pre-tokenization regex\n    // Splits on whitespace boundaries, keeping the space as a prefix (Ġ convention)\n    // Also splits numbers and punctuation\n    const pattern = /'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+/gu;\n    const matches = text.match(pattern);\n    if (!matches) return [];\n\n    // Convert to the byte-level representation HF uses:\n    // space → Ġ, etc. (the GPT2 byte-to-unicode mapping)\n    return matches.map((chunk) => this.textToTokenRepr(chunk));\n  }\n\n  private textToTokenRepr(text: string): string {\n    // Convert text to the representation used in the vocab\n    // In HF BPE tokenizers, spaces are typically represented as Ġ (U+0120)\n    // and other bytes have specific unicode mappings\n    let result = \"\";\n    for (let i = 0; i < text.length; i++) {\n      const c = text.charCodeAt(i);\n      if (c === 32) {\n        result += \"\\u0120\"; // space → Ġ\n      } else if (c < 33 || (c > 126 && c < 161)) {\n        // Control chars and non-printable → byte-level tokens\n        // Use the GPT2 byte encoder mapping\n        result += String.fromCharCode(c + 256);\n      } else {\n        result += text[i];\n      }\n    }\n    return result;\n  }\n\n  private bpeEncode(word: string): number[] {\n    if (word.length === 0) return [];\n\n    // Check if the whole word is in vocab\n    if (this.vocab.has(word)) {\n      return [this.vocab.get(word)!];\n    }\n\n    // Split into individual characters\n    let symbols = [...word];\n\n    // If any symbol isn't in vocab, try byte fallback\n    // Convert to individual characters and check\n    if (symbols.length === 1) {\n      const id = this.vocab.get(symbols[0]);\n      if (id !== undefined) return [id];\n      // Byte fallback\n      return this.encodeByteFallback(word);\n    }\n\n    // Iteratively merge the most frequent pair\n    while (symbols.length > 1) {\n      // Find the pair with the lowest merge rank (highest priority)\n      let _bestPair = \"\";\n      let bestRank = Infinity;\n      let bestIdx = -1;\n\n      for (let i = 0; i < symbols.length - 1; i++) {\n        const pair = `${symbols[i]} ${symbols[i + 1]}`;\n        const rank = this.merges.get(pair);\n        if (rank !== undefined && rank < bestRank) {\n          bestRank = rank;\n          _bestPair = pair;\n          bestIdx = i;\n        }\n      }\n\n      // No more merges possible\n      if (bestIdx === -1) break;\n\n      // Apply the merge\n      const merged = symbols[bestIdx] + symbols[bestIdx + 1];\n      symbols = [...symbols.slice(0, bestIdx), merged, ...symbols.slice(bestIdx + 2)];\n    }\n\n    // Convert symbols to IDs\n    const ids: number[] = [];\n    for (const sym of symbols) {\n      const id = this.vocab.get(sym);\n      if (id !== undefined) {\n        ids.push(id);\n      } else {\n        // Byte fallback for unknown symbols\n        ids.push(...this.encodeByteFallback(sym));\n      }\n    }\n    return ids;\n  }\n\n  private encodeByteFallback(text: string): number[] {\n    const encoder = new TextEncoder();\n    const bytes = encoder.encode(text);\n    const ids: number[] = [];\n    for (const b of bytes) {\n      const tok = this.byteFallback.get(b);\n      if (tok) {\n        const id = this.vocab.get(tok);\n        if (id !== undefined) ids.push(id);\n      }\n    }\n    return ids;\n  }\n}\n","/**\n * WeightSource — a tensor store the loader/executor can pull from one tensor at a\n * time, so the *whole* model never has to sit in the JS heap simultaneously.\n *\n * Two backends share one async-shaped interface:\n *\n *  • HeapWeightStore (Node / desktop) — wraps a plain `Map`. Identical behavior\n *    to the original code path: tensors are held in heap, gets/sets are O(1) heap\n *    ops (the async methods resolve synchronously). Zero behavior change.\n *\n *  • CacheWeightStore (browser / mobile) — backs every tensor's BYTES in the\n *    CacheStorage API (already populated per-tensor during download). The heap\n *    holds only lightweight descriptors ({ shape, dtype, cacheKey, byteLength })\n *    plus a small set of tiny \"override\" tensors. `get()` reads one tensor's\n *    bytes back from cache on demand (applying any BF16/F16→F32 conversion) and\n *    `set()` writes transform outputs straight back to cache without retaining\n *    the bytes. This bounds peak heap to roughly ONE tensor at a time instead of\n *    the entire weight set — which is what lets a 2.5 GB model load on iOS Safari\n *    without the tab being jetsam-killed.\n *\n * The browser path is selected automatically when `typeof caches !== \"undefined\"`.\n */\n\nimport { opfsAvailable, opfsRead } from \"./opfs-cache.js\";\nimport { bf16ToF32, f16ToF32, type SafetensorDType } from \"./safetensors.js\";\n\n/** A single tensor's data + shape (the unit the executor uploads to a GPU buffer). */\nexport interface WeightEntry {\n  data: ArrayBufferView;\n  shape: number[];\n}\n\n/**\n * Read-side view consumed by the executor's streaming `uploadWeights`. Async by\n * design so a cache-backed store can fetch one tensor's bytes at a time.\n */\nexport interface WeightSource {\n  has(name: string): boolean;\n  keys(): string[];\n  readonly size: number;\n  /** Pull a single tensor (bytes materialized + dtype-converted on demand). */\n  get(name: string): Promise<WeightEntry | undefined>;\n  /**\n   * Release any transient backing storage (e.g. the browser transform-staging\n   * cache) once the consumer has finished uploading. No-op for the heap backend.\n   */\n  dispose?(): Promise<void>;\n}\n\n/**\n * Mutable store used while loading. The loader's post-download transforms\n * (RMSNorm +1 bake, fused-Q split, GPTQ/MLX repack, INT4 quantize, PLE/vision\n * fixups) read tensors, produce new ones, and drop sources through this\n * interface. All ops are async so the cache backend can stream bytes; the heap\n * backend resolves them synchronously.\n */\nexport interface MutableWeightStore extends WeightSource {\n  set(name: string, entry: WeightEntry): Promise<void>;\n  delete(name: string): Promise<void>;\n}\n\n// ── Heap backend (Node / desktop) — unchanged behavior ──────────────────────\n\nexport class HeapWeightStore implements MutableWeightStore {\n  private map: Map<string, WeightEntry>;\n\n  constructor(map?: Map<string, WeightEntry>) {\n    this.map = map ?? new Map();\n  }\n\n  has(name: string): boolean {\n    return this.map.has(name);\n  }\n\n  keys(): string[] {\n    return [...this.map.keys()];\n  }\n\n  get size(): number {\n    return this.map.size;\n  }\n\n  get(name: string): Promise<WeightEntry | undefined> {\n    return Promise.resolve(this.map.get(name));\n  }\n\n  set(name: string, entry: WeightEntry): Promise<void> {\n    this.map.set(name, entry);\n    return Promise.resolve();\n  }\n\n  delete(name: string): Promise<void> {\n    this.map.delete(name);\n    return Promise.resolve();\n  }\n\n  /** Escape hatch for the few Node paths that still want the raw Map. */\n  asMap(): Map<string, WeightEntry> {\n    return this.map;\n  }\n}\n\n// ── Cache backend (browser / mobile) — bounded peak heap ─────────────────────\n\n/**\n * A descriptor of a tensor whose bytes live in CacheStorage rather than the heap.\n * `dtype` is the *source* safetensors dtype so `get()` can reproduce the same\n * BF16/F16→F32 conversion the heap path applies inline during download.\n */\ninterface CacheDescriptor {\n  shape: number[];\n  /** Source safetensors dtype, or \"RAW\" for already-converted/derived bytes. */\n  dtype: SafetensorDType | \"RAW\";\n  /** Which CacheStorage bucket holds the bytes (download cache vs staging cache). */\n  cacheName: string;\n  cacheKey: string;\n  /** TypedArray ctor tag for RAW (transform-produced) bytes. */\n  view?: \"f32\" | \"u32\" | \"i32\" | \"u16\" | \"u8\";\n}\n\n/**\n * Build a typed-array view over a freshly-read (offset-0) ArrayBuffer. A cache\n * read always returns a fresh, offset-0 buffer, so it is aligned for every dtype\n * — no copy needed (unlike the merged-range download path which can be unaligned).\n */\nfunction viewForSourceDtype(buf: ArrayBuffer, dtype: SafetensorDType): ArrayBufferView {\n  switch (dtype) {\n    case \"F32\":\n      return new Float32Array(buf, 0, buf.byteLength / 4);\n    case \"I32\":\n      return new Int32Array(buf, 0, buf.byteLength / 4);\n    case \"U32\":\n      return new Uint32Array(buf, 0, buf.byteLength / 4);\n    case \"F16\":\n    case \"BF16\":\n      return new Uint16Array(buf, 0, buf.byteLength / 2);\n    default:\n      return new Uint8Array(buf, 0, buf.byteLength);\n  }\n}\n\nfunction viewForRaw(buf: ArrayBuffer, tag: CacheDescriptor[\"view\"]): ArrayBufferView {\n  switch (tag) {\n    case \"u32\":\n      return new Uint32Array(buf, 0, buf.byteLength / 4);\n    case \"i32\":\n      return new Int32Array(buf, 0, buf.byteLength / 4);\n    case \"u16\":\n      return new Uint16Array(buf, 0, buf.byteLength / 2);\n    case \"u8\":\n      return new Uint8Array(buf, 0, buf.byteLength);\n    default:\n      return new Float32Array(buf, 0, buf.byteLength / 4);\n  }\n}\n\nfunction rawTag(data: ArrayBufferView): CacheDescriptor[\"view\"] {\n  if (data instanceof Float32Array) return \"f32\";\n  if (data instanceof Uint32Array) return \"u32\";\n  if (data instanceof Int32Array) return \"i32\";\n  if (data instanceof Uint16Array) return \"u16\";\n  return \"u8\";\n}\n\nconst CACHE_BACKED_NAME = \"gerbil-weights-staging-v2\";\n\nexport class CacheWeightStore implements MutableWeightStore {\n  private descriptors = new Map<string, CacheDescriptor>();\n  /**\n   * Small heap-resident tensors that are NOT worth a cache round-trip: synthetic\n   * fill constants, sub-MB norm/scalar tensors, and anything a caller explicitly\n   * pins (e.g. the vision pos-embed table that index.ts snapshots before upload).\n   * Keeping these in heap is fine — they are individually tiny.\n   */\n  private overrides = new Map<string, WeightEntry>();\n  private cacheName: string;\n\n  /** Tensors at or below this size stay in heap rather than round-tripping cache. */\n  private static readonly HEAP_THRESHOLD = 512 * 1024; // 512 KB\n\n  constructor(cacheName: string = CACHE_BACKED_NAME) {\n    this.cacheName = cacheName;\n  }\n\n  private static stagingKey(name: string): string {\n    return `gerbil-staging:${name}`;\n  }\n\n  /**\n   * Register a tensor whose bytes already live in CacheStorage under\n   * `downloadCacheName`/`cacheKey` (the per-tensor download cache). No bytes are\n   * retained in heap. The descriptor records which cache bucket to read from so\n   * `get()` opens the right one (download cache vs this store's staging cache).\n   */\n  registerCached(\n    name: string,\n    shape: number[],\n    dtype: SafetensorDType,\n    cacheKey: string,\n    downloadCacheName: string,\n  ): void {\n    this.overrides.delete(name);\n    this.descriptors.set(name, { shape, dtype, cacheName: downloadCacheName, cacheKey });\n  }\n\n  has(name: string): boolean {\n    return this.overrides.has(name) || this.descriptors.has(name);\n  }\n\n  keys(): string[] {\n    const set = new Set<string>(this.descriptors.keys());\n    for (const k of this.overrides.keys()) set.add(k);\n    return [...set];\n  }\n\n  get size(): number {\n    return this.keys().length;\n  }\n\n  async get(name: string): Promise<WeightEntry | undefined> {\n    const pinned = this.overrides.get(name);\n    if (pinned) return pinned;\n    const desc = this.descriptors.get(name);\n    if (!desc) return undefined;\n\n    // Tensors registered from the durable per-tensor download cache may live in\n    // OPFS (the iOS-durable backend) rather than CacheStorage — model-loader's\n    // browserCacheWrite prefers OPFS. Probe OPFS by the same key first, then fall\n    // back to CacheStorage. RAW transform-staging entries (written via cache.put\n    // in set() below) are never in OPFS, so this probe simply misses for them.\n    let buf: ArrayBuffer | null = null;\n    if (desc.dtype !== \"RAW\" && opfsAvailable()) {\n      buf = await opfsRead(desc.cacheKey);\n    }\n    if (!buf) {\n      const cache = await caches.open(desc.cacheName);\n      const resp = await cache.match(new Request(desc.cacheKey));\n      if (!resp) return undefined;\n      buf = await resp.arrayBuffer();\n    }\n\n    if (desc.dtype === \"RAW\") {\n      return { data: viewForRaw(buf, desc.view), shape: desc.shape };\n    }\n    let data = viewForSourceDtype(buf, desc.dtype);\n    // Mirror the heap path's inline conversion (model-loader.ts ~860-933).\n    if (desc.dtype === \"BF16\") {\n      data = bf16ToF32(new Uint8Array(data.buffer, data.byteOffset, data.byteLength));\n    } else if (desc.dtype === \"F16\") {\n      data = f16ToF32(data);\n    }\n    return { data, shape: desc.shape };\n  }\n\n  async set(name: string, entry: WeightEntry): Promise<void> {\n    this.descriptors.delete(name);\n    // Tiny tensors stay in heap; large ones write through to cache so their bytes\n    // do not accumulate in the heap as transforms run layer-by-layer.\n    if (entry.data.byteLength <= CacheWeightStore.HEAP_THRESHOLD) {\n      this.overrides.set(name, entry);\n      return;\n    }\n    this.overrides.delete(name);\n    const cacheKey = CacheWeightStore.stagingKey(name);\n    const bytes = entry.data.buffer.slice(\n      entry.data.byteOffset,\n      entry.data.byteOffset + entry.data.byteLength,\n    ) as ArrayBuffer;\n    const cache = await caches.open(this.cacheName);\n    await cache.put(new Request(cacheKey), new Response(bytes));\n    this.descriptors.set(name, {\n      shape: entry.shape,\n      dtype: \"RAW\",\n      cacheName: this.cacheName,\n      cacheKey,\n      view: rawTag(entry.data),\n    });\n  }\n\n  async delete(name: string): Promise<void> {\n    this.overrides.delete(name);\n    const desc = this.descriptors.get(name);\n    this.descriptors.delete(name);\n    // Only evict staging entries we created; never evict the durable per-tensor\n    // download cache (a future load reuses those by stable byte-range key).\n    if (desc && desc.dtype === \"RAW\" && desc.cacheName === this.cacheName) {\n      try {\n        const cache = await caches.open(this.cacheName);\n        await cache.delete(new Request(desc.cacheKey));\n      } catch {\n        /* best-effort */\n      }\n    }\n  }\n\n  /** Drop all transform-staging cache entries (call after upload completes). */\n  async dispose(): Promise<void> {\n    try {\n      await caches.delete(this.cacheName);\n    } catch {\n      /* best-effort */\n    }\n  }\n}\n\n/** True when the browser CacheStorage API is available (i.e. not Node). */\nexport function cacheStorageAvailable(): boolean {\n  return typeof caches !== \"undefined\";\n}\n","/**\n * Model loader -- fetches model files from HuggingFace Hub.\n *\n * Handles:\n * - Fetching config.json, tokenizer.json, tokenizer_config.json\n * - Downloading safetensors weight files (with progress callbacks)\n * - HF key -> canonical key mapping\n * - Uploading weight data to GPU buffers via the Executor\n */\n\nimport { type GraphDType, generateGraph, type KVDType } from \"./architectures/index.js\";\nimport { foldNanoCodecWeightNorm, nanoCodecWeightMap } from \"./architectures/kani_tts.js\";\nimport { repackGPTQ } from \"./gptq-adapter.js\";\nimport { idbAllKeys, idbCacheAvailable, idbGet, idbHas, idbPut } from \"./idb-cache.js\";\nimport type { ModelGraph } from \"./ir.js\";\nimport { CANONICAL_KEYS, createDefaultHFKeyMapper, type HFKeyMapper } from \"./ir.js\";\nimport { repackMLX } from \"./mlx-adapter.js\";\nimport { opfsEvictStale } from \"./opfs-cache.js\";\nimport { DEFAULT_GROUP_SIZE, quantizeInt4 } from \"./quantize.js\";\nimport {\n  bf16ToF32,\n  f16ToF32,\n  getTensorData,\n  parseSafetensorsHeader,\n  type SafetensorEntry,\n  type SafetensorsFile,\n} from \"./safetensors.js\";\nimport { Tokenizer } from \"./tokenizer.js\";\nimport {\n  CacheWeightStore,\n  cacheStorageAvailable,\n  HeapWeightStore,\n  type MutableWeightStore,\n  type WeightSource,\n} from \"./weight-source.js\";\n\nexport interface LoadModelOptions {\n  /** HF repo ID (e.g. \"Qwen/Qwen3.5-0.8B\") or full URL. */\n  repo: string;\n  /** Progress callback: (loaded, total, message) */\n  onProgress?: (loaded: number, total: number, message: string) => void;\n  /** Custom HF key mapper (defaults to stripping \"model.\" prefix). */\n  keyMapper?: HFKeyMapper;\n  /** HuggingFace API token for gated models. */\n  hfToken?: string;\n  /** Revision/branch (default: \"main\"). */\n  revision?: string;\n  /** Local cache directory for downloaded files (Node.js only). */\n  cacheDir?: string;\n  /**\n   * Weight dtype:\n   *  - \"f32\"  full precision (or the repo's native quantization, e.g. MLX/GPTQ q4)\n   *  - \"q4\"   on-the-fly INT4 quantization (~4× smaller)\n   *  - \"auto\" (recommended) picks q4 on mobile (iOS/Android) to fit in device\n   *           memory and f32/native on desktop. Already-quantized repos\n   *           (MLX/GPTQ 4-bit) stay q4 regardless.\n   */\n  dtype?: GraphDType | \"auto\";\n  /** KV cache dtype: \"f16\" halves memory traffic during attention. Requires GPU f16 support. */\n  kvDtype?: KVDType;\n  /**\n   * Build an embedding graph (last-token pool + L2 norm) instead of an LM head.\n   * Only valid for Qwen2/Qwen3 CausalLM architectures (e.g. Qwen3-Embedding).\n   */\n  embedding?: boolean;\n  /**\n   * Build the multimodal LM graph variant (M-RoPE + image-embedding splice) so\n   * the text model can consume spliced image tokens. Only meaningful for\n   * Qwen3_5ForConditionalGeneration. Reserves `maxVisionTokens` rows for the\n   * vision-embedding buffer. Text-only generation through this graph is\n   * numerically identical to the non-multimodal graph (M-RoPE fed linear\n   * positions == standard 1D RoPE).\n   */\n  multimodal?: { maxVisionTokens: number };\n  /**\n   * Force-download and key-map the vision tower even without the multimodal LM\n   * graph. `enableVision` (via `multimodal`) already implies this; this flag is\n   * for callers that load weights directly (e.g. the vision-encoder validation\n   * scripts) and build the vision graph/executor themselves. When neither this\n   * nor `multimodal` is set, the ~201MB ViT is excluded from the download.\n   */\n  loadVisionTower?: boolean;\n}\n\n/**\n * Gemma 4 Per-Layer-Embeddings (PLE) source, kept CPU-resident.\n *\n * The PLE table (`embed_tokens_per_layer`, [vocab, num_layers*256]) is ~1.17GB\n * at 4-bit. Uploading it to a GPU buffer would make the model non-mobile-viable\n * and would hit the per-binding size cap. Instead the loader hands the quantized\n * table to the executor in JS memory; the executor gathers + dequantizes only the\n * rows for the current input tokens each forward step (a tiny [T, width] upload).\n */\nexport interface PleSource {\n  /**\n   * Flat row-major INT4 nibbles (Gerbil packing, 8 per u32). HEAP-RESIDENT path\n   * (Node/desktop). In the browser this is empty and `cache` is set instead so\n   * the ~1.17 GB table never sits in the JS heap during load.\n   */\n  packed: Uint32Array;\n  /** Per-group scales (Gerbil (nibble - zero) * scale convention). */\n  scales: Float32Array;\n  /** Per-group zero points. */\n  zeros: Float32Array;\n  /** Row width = num_layers * hidden_size_per_layer_input (E2B: 35*256 = 8960). */\n  width: number;\n  /** Dequant group size (MLX: 64). */\n  groupSize: number;\n  /** Activation tensor the per-step gathered rows are written into. */\n  targetTensor: string;\n  /**\n   * Browser only: when set, the quantized PLE table's bytes live in CacheStorage\n   * (not the heap). The executor reads the slice of nibbles/scales/zeros it needs\n   * for the current tokens on demand. Keeps peak load heap bounded.\n   */\n  cache?: {\n    cacheName: string;\n    packedKey: string;\n    scalesKey: string;\n    zerosKey: string;\n    /** packed.length (u32 count) — for bounds/Range math. */\n    packedLen: number;\n  };\n}\n\nexport interface LoadedModel {\n  /** The generated computation graph (IR). */\n  graph: ModelGraph;\n  /** The tokenizer. */\n  tokenizer: Tokenizer;\n  /**\n   * Weight tensors mapped to canonical names. A `WeightSource` so the executor\n   * can pull one tensor at a time (cache-backed in the browser, heap-backed on\n   * Node) instead of requiring the whole model to sit in heap at once. Use\n   * `get(name)` (async) to materialize a tensor's bytes on demand.\n   */\n  weights: WeightSource;\n  /** Raw config.json for reference. */\n  rawConfig: Record<string, unknown>;\n  /**\n   * CPU-resident Gemma 4 PLE table (set only for Gemma 4). Pass to\n   * `executor.setPleSource()` so the big table never becomes GPU-resident.\n   */\n  pleSource?: PleSource;\n}\n\n/**\n * Resolve a repo string to a base URL for HuggingFace Hub API.\n */\nfunction resolveHFBaseURL(repo: string, revision: string): string {\n  // If it's already a full URL, use it\n  if (repo.startsWith(\"http://\") || repo.startsWith(\"https://\")) {\n    return repo;\n  }\n  // Standard HF Hub URL\n  return `https://huggingface.co/${repo}/resolve/${revision}`;\n}\n\n// ── File cache helpers (Node.js only) ──────────────────────────────────\n\nlet _fs: typeof import(\"node:fs\") | null = null;\nlet _path: typeof import(\"node:path\") | null = null;\nlet _fsInitialized = false;\n\nasync function initFs(): Promise<void> {\n  if (_fsInitialized) return;\n  _fsInitialized = true;\n  try {\n    _fs = await import(\"node:fs\");\n    _path = await import(\"node:path\");\n  } catch {\n    // Not in Node.js (browser) — caching disabled\n  }\n}\n\nfunction getCachePath(cacheDir: string, filename: string): string {\n  return _path!.join(cacheDir, filename.replace(/\\//g, \"_\"));\n}\n\nfunction readCachedFile(cacheDir: string, filename: string): Buffer | null {\n  if (!_fs || !_path) return null;\n  const p = getCachePath(cacheDir, filename);\n  try {\n    return _fs.readFileSync(p);\n  } catch {\n    return null;\n  }\n}\n\nfunction writeCacheFile(\n  cacheDir: string,\n  filename: string,\n  data: ArrayBuffer | Buffer | Uint8Array,\n): void {\n  if (!_fs || !_path) return;\n  _fs.mkdirSync(cacheDir, { recursive: true });\n  const p = getCachePath(cacheDir, filename);\n  let buf: Buffer;\n  if (Buffer.isBuffer(data)) {\n    buf = data;\n  } else if (data instanceof Uint8Array) {\n    buf = Buffer.from(data);\n  } else {\n    buf = Buffer.from(data);\n  }\n  _fs.writeFileSync(p, buf);\n}\n\n// ── Browser cache helpers (OPFS-durable, CacheStorage fallback) ─────────\n//\n// Two-tier durable cache behind a single read/write/has seam:\n//\n//  1. OPFS via a dedicated Worker (`createSyncAccessHandle`, chunked writes) —\n//     the ONLY iOS-safe durable path. On iOS Safari the CacheStorage API is\n//     evicted under quota pressure in a plain tab (`navigator.storage.persist()`\n//     is not granted outside an installed PWA), so weights cached there are\n//     re-downloaded every visit. OPFS survives a full tab close + reopen. Main-\n//     thread OPFS `createWritable` throws \"out of quota\" mid-write on iOS, so all\n//     OPFS I/O runs on a blob-URL worker with sync access handles. See\n//     src/gpu/opfs-cache.ts.\n//\n//  2. CacheStorage (`caches.open(...)`) — fallback when OPFS/Worker is\n//     unavailable (older browsers, some embedded webviews). Best-effort durable;\n//     fine on desktop where quota pressure is rare.\n//\n// Reads prefer OPFS, then fall back to CacheStorage. Writes go to OPFS when\n// available and only to CacheStorage otherwise (no double-write — OPFS is the\n// durable winner and double-storing would waste the ~404 MB twice against quota).\n// All of this is inert on Node: `opfsAvailable()` and `typeof caches` are both\n// false there, so the helpers short-circuit and the Node weight path is\n// unaffected.\n\n// v4: storage layout changed (OPFS subdirectory `gerbil-models-v4` + CacheStorage\n// bucket of the same name). A fresh namespace also abandons any entries poisoned\n// during earlier buggy builds (the reload-loop era wrote right-size-wrong-content\n// tensors that size validation can't detect). The per-tensor cache logic itself\n// is verified correct by scripts/engine/test-cache-roundtrip.mjs (load → cache →\n// reload-from-cache → identical coherent output).\nconst BROWSER_CACHE_NAME = \"gerbil-models-v4\";\n\n// Per-tensor weight caching. ON by default — the round-trip test\n// (scripts/engine/test-cache-roundtrip.mjs) proves the cache read path is\n// correct, and the v3 namespace abandons entries poisoned by earlier buggy\n// builds. Can be force-disabled via GERBIL_TENSOR_CACHE=0 / globalThis flag if a\n// corruption issue ever resurfaces.\nconst TENSOR_CACHE_ENABLED =\n  (typeof process === \"undefined\" || process.env?.GERBIL_TENSOR_CACHE !== \"0\") &&\n  (globalThis as { GERBIL_TENSOR_CACHE?: boolean }).GERBIL_TENSOR_CACHE !== false;\n\nlet _persistRequested = false;\nasync function requestPersistence(): Promise<void> {\n  if (_persistRequested) return;\n  _persistRequested = true;\n  try {\n    // Only granted in an installed PWA on iOS; harmless best-effort elsewhere.\n    await (navigator as any)?.storage?.persist?.();\n  } catch {\n    /* best-effort */\n  }\n}\n\n// Model weights are cached in IndexedDB (see idb-cache.ts for why: the Cache API's\n// fragment-stripping collapsed our per-tensor keys → wrong bytes/garbage, and OPFS\n// via Worker was fragile on iOS). IDB keys on the exact string — no footguns. All\n// three helpers are inert on Node (no indexedDB → miss/no-op), so the heap load\n// path is unchanged off the browser.\n\n/** Read a cache key's bytes from IndexedDB, or null on miss. */\nasync function browserCacheRead(cacheKey: string): Promise<ArrayBuffer | null> {\n  return idbGet(cacheKey);\n}\n\n/** Cheap presence check — true without materializing the body. */\nasync function browserCacheHas(cacheKey: string): Promise<boolean> {\n  return idbHas(cacheKey);\n}\n\n/**\n * Write a cache key's bytes to IndexedDB (best-effort; never throws). The\n * `transferable` flag is accepted for call-site compatibility but unused — IDB\n * structured-clones the value, so the caller's buffer is never detached.\n */\nasync function browserCacheWrite(\n  cacheKey: string,\n  data: ArrayBuffer,\n  _transferable = false,\n): Promise<void> {\n  await requestPersistence();\n  await idbPut(cacheKey, data);\n}\n\n/**\n * Reclaim quota from the LEGACY caches, ONCE per session. The model cache is now\n * IndexedDB; the old CacheStorage (`gerbil-models-*`, fragment-keyed → the `RRRR`\n * corruption) and OPFS-worker stores are dead. iOS Safari's per-origin quota is\n * ~1 GB and that orphaned junk could consume nearly all of it, so we purge ALL of\n * it to free room for the IDB cache. Best-effort; inert on Node.\n */\nlet _evictedStale = false;\nasync function evictStaleCaches(): Promise<void> {\n  if (_evictedStale) return;\n  _evictedStale = true;\n  try {\n    if (typeof caches !== \"undefined\") {\n      const keys = await caches.keys();\n      await Promise.all(\n        keys.filter((k) => k.startsWith(\"gerbil-models-\")).map((k) => caches.delete(k)),\n      );\n    }\n  } catch {\n    /* best-effort */\n  }\n  await opfsEvictStale();\n}\n\n// ── Fetch helpers ──────────────────────────────────────────────────────\n\n/**\n * Fetch a JSON file from HF with optional auth and caching.\n */\nasync function fetchJSON(\n  baseURL: string,\n  filename: string,\n  hfToken?: string,\n  cacheDir?: string,\n): Promise<any> {\n  // Check Node.js cache first\n  if (cacheDir) {\n    const cached = readCachedFile(cacheDir, filename);\n    if (cached) return JSON.parse(cached.toString(\"utf-8\"));\n  }\n\n  // Check browser cache\n  const browserKey = `${baseURL}/${filename}`;\n  const browserCached = await browserCacheRead(browserKey);\n  if (browserCached) {\n    return JSON.parse(new TextDecoder().decode(browserCached));\n  }\n\n  const url = `${baseURL}/${filename}`;\n  const headers: Record<string, string> = {};\n  if (hfToken) {\n    headers.Authorization = `Bearer ${hfToken}`;\n  }\n  const response = await fetch(url, { headers });\n  if (!response.ok) {\n    throw new Error(`Failed to fetch ${url}: ${response.status} ${response.statusText}`);\n  }\n  const text = await response.text();\n\n  // Write to Node.js cache\n  if (cacheDir) {\n    writeCacheFile(cacheDir, filename, Buffer.from(text, \"utf-8\"));\n  }\n\n  // Write to browser cache\n  await browserCacheWrite(browserKey, new TextEncoder().encode(text).buffer);\n\n  return JSON.parse(text);\n}\n\n/**\n * Fetch a UTF-8 text file from HF (with caching). Returns null on 404 / error.\n * Used for sidecar files like `chat_template.jinja` that some models ship\n * instead of an inline `chat_template` in tokenizer_config.json.\n */\nasync function fetchText(\n  baseURL: string,\n  filename: string,\n  hfToken?: string,\n  cacheDir?: string,\n): Promise<string | null> {\n  if (cacheDir) {\n    const cached = readCachedFile(cacheDir, filename);\n    if (cached) return cached.toString(\"utf-8\");\n  }\n  const browserKey = `${baseURL}/${filename}`;\n  const browserCached = await browserCacheRead(browserKey);\n  if (browserCached) {\n    return new TextDecoder().decode(browserCached);\n  }\n  const url = `${baseURL}/${filename}`;\n  const headers: Record<string, string> = {};\n  if (hfToken) {\n    headers.Authorization = `Bearer ${hfToken}`;\n  }\n  const response = await fetch(url, { headers });\n  if (!response.ok) return null;\n  const text = await response.text();\n  if (cacheDir) {\n    writeCacheFile(cacheDir, filename, Buffer.from(text, \"utf-8\"));\n  }\n  await browserCacheWrite(browserKey, new TextEncoder().encode(text).buffer);\n  return text;\n}\n\n/**\n * Fetch a binary file from HF with progress tracking and caching.\n */\nasync function fetchBinary(\n  baseURL: string,\n  filename: string,\n  hfToken?: string,\n  onProgress?: (loaded: number, total: number) => void,\n  cacheDir?: string,\n): Promise<ArrayBuffer> {\n  // Check cache first\n  if (cacheDir) {\n    const cached = readCachedFile(cacheDir, filename);\n    if (cached) {\n      onProgress?.(cached.byteLength, cached.byteLength);\n      return cached.buffer.slice(\n        cached.byteOffset,\n        cached.byteOffset + cached.byteLength,\n      ) as ArrayBuffer;\n    }\n  }\n\n  const url = `${baseURL}/${filename}`;\n  const headers: Record<string, string> = {};\n  if (hfToken) {\n    headers.Authorization = `Bearer ${hfToken}`;\n  }\n\n  const response = await fetch(url, { headers });\n  if (!response.ok) {\n    throw new Error(`Failed to fetch ${url}: ${response.status} ${response.statusText}`);\n  }\n\n  const contentLength = Number(response.headers.get(\"content-length\") || 0);\n\n  if (!response.body || !onProgress) {\n    const buf = await response.arrayBuffer();\n    if (cacheDir) writeCacheFile(cacheDir, filename, buf);\n    return buf;\n  }\n\n  // Stream with progress\n  const reader = response.body.getReader();\n  const chunks: Uint8Array[] = [];\n  let loaded = 0;\n\n  while (true) {\n    const { done, value } = await reader.read();\n    if (done) break;\n    chunks.push(value);\n    loaded += value.byteLength;\n    onProgress(loaded, contentLength);\n  }\n\n  // Combine chunks into single ArrayBuffer\n  const result = new Uint8Array(loaded);\n  let offset = 0;\n  for (const chunk of chunks) {\n    result.set(chunk, offset);\n    offset += chunk.byteLength;\n  }\n\n  if (cacheDir) writeCacheFile(cacheDir, filename, result);\n  return result.buffer;\n}\n\n/**\n * Discover safetensors files in the repo.\n *\n * Models can have a single `model.safetensors` or multiple shards:\n * `model-00001-of-00002.safetensors`, `model-00002-of-00002.safetensors`\n *\n * We check for the index file first, then fall back to single file.\n */\nasync function discoverSafetensorsFiles(\n  baseURL: string,\n  hfToken?: string,\n  cacheDir?: string,\n): Promise<string[]> {\n  // Try the index file first (for sharded models)\n  try {\n    const index = await fetchJSON(baseURL, \"model.safetensors.index.json\", hfToken, cacheDir);\n    if (index.weight_map) {\n      // Get unique filenames from the weight map\n      const files = [...new Set(Object.values(index.weight_map))] as string[];\n      return files.sort();\n    }\n  } catch {\n    // No index file -- try single file\n  }\n\n  return [\"model.safetensors\"];\n}\n\n/**\n * Fetch a byte range from a URL over the network (no caching).\n */\nasync function fetchRangeRaw(\n  url: string,\n  start: number,\n  end: number,\n  hfToken?: string,\n): Promise<ArrayBuffer> {\n  const headers: Record<string, string> = {\n    Range: `bytes=${start}-${end - 1}`,\n  };\n  if (hfToken) {\n    headers.Authorization = `Bearer ${hfToken}`;\n  }\n  // Retry transient failures with backoff. HuggingFace rate-limits unauthenticated\n  // downloads (403/429), and CDN/network hiccups happen; a few backed-off retries\n  // turn those into a slightly slower load instead of a hard failure. (Once the\n  // model is cached in IDB these requests don't happen again.)\n  let lastStatus = 0;\n  for (let attempt = 0; attempt < 4; attempt++) {\n    let resp: Response;\n    try {\n      resp = await fetch(url, { headers });\n    } catch {\n      // Network error — back off and retry.\n      if (attempt < 3) {\n        await new Promise((r) => setTimeout(r, 600 * 2 ** attempt));\n        continue;\n      }\n      throw new Error(`Range request failed: network error for ${url}`);\n    }\n    if (resp.ok || resp.status === 206) return resp.arrayBuffer();\n    lastStatus = resp.status;\n    // 403/429 = rate limit (HF), 5xx = transient server error → retry; else fail.\n    const retriable = resp.status === 403 || resp.status === 429 || resp.status >= 500;\n    if (retriable && attempt < 3) {\n      await new Promise((r) => setTimeout(r, 600 * 2 ** attempt));\n      continue;\n    }\n    break;\n  }\n  throw new Error(\n    `Range request failed: ${lastStatus}${lastStatus === 403 || lastStatus === 429 ? \" (HuggingFace rate limit — wait a bit and retry; once cached this won't recur)\" : \"\"}`,\n  );\n}\n\n/**\n * Fetch a byte range from a URL, caching the exact range.\n *\n * Used for small, stable ranges (e.g. safetensors headers) whose boundaries\n * don't depend on which tensors a particular load needs. Tensor *bodies* are\n * NOT fetched through here — they're cached per-tensor by stable byte-range in\n * fetchSelectiveTensors so a vision load can reuse a text load's tensors.\n */\nasync function fetchRange(\n  url: string,\n  start: number,\n  end: number,\n  hfToken?: string,\n): Promise<ArrayBuffer> {\n  // Check browser cache for this range\n  const browserKey = `${url}?_range=${start}-${end}`;\n  const browserCached = await browserCacheRead(browserKey);\n  if (browserCached) return browserCached;\n\n  const buf = await fetchRangeRaw(url, start, end, hfToken);\n\n  // Cache for next time\n  await browserCacheWrite(browserKey, buf);\n\n  return buf;\n}\n\n/**\n * Fetch the safetensors header only (two range requests: 8 bytes + header).\n */\nasync function fetchSafetensorsHeader(url: string, hfToken?: string): Promise<SafetensorsFile> {\n  // Fetch first 8 bytes to get header length\n  const lengthBuf = await fetchRange(url, 0, 8, hfToken);\n  const headerLength = Number(new DataView(lengthBuf).getBigUint64(0, true));\n\n  // Fetch full header (8 + headerLength bytes)\n  const headerBuf = await fetchRange(url, 0, 8 + headerLength, hfToken);\n  return parseSafetensorsHeader(headerBuf);\n}\n\n/** A contiguous byte range to download. */\ninterface ByteRange {\n  start: number;\n  end: number;\n  entries: SafetensorEntry[];\n}\n\n/**\n * Merge close byte ranges to reduce HTTP requests.\n *\n * - `gapThreshold`: merge two ranges if the gap between them is smaller than this\n *   (avoids an extra request for a tiny skipped gap).\n * - `maxRangeBytes`: hard cap on a single merged range. Without this, a\n *   contiguously-packed safetensors merges into ONE range = the whole model, so\n *   the loader buffers the entire file in a single fetch — no streaming progress\n *   (the bar sits at 0/N until done) and a transient allocation big enough to\n *   crash mobile Safari. Capping splits the download into bounded segments that\n *   report progress as they land and keep peak memory low.\n */\nfunction mergeRanges(\n  entries: SafetensorEntry[],\n  dataStart: number,\n  gapThreshold: number = 1024 * 1024,\n  maxRangeBytes: number = Number.POSITIVE_INFINITY,\n): ByteRange[] {\n  if (entries.length === 0) return [];\n\n  // Sort by offset\n  const sorted = [...entries].sort((a, b) => a.dataOffset - b.dataOffset);\n\n  const ranges: ByteRange[] = [];\n  let current: ByteRange = {\n    start: dataStart + sorted[0].dataOffset,\n    end: dataStart + sorted[0].dataOffset + sorted[0].dataLength,\n    entries: [sorted[0]],\n  };\n\n  for (let i = 1; i < sorted.length; i++) {\n    const entryStart = dataStart + sorted[i].dataOffset;\n    const entryEnd = entryStart + sorted[i].dataLength;\n\n    const withinGap = entryStart - current.end <= gapThreshold;\n    const withinCap = entryEnd - current.start <= maxRangeBytes;\n    if (withinGap && withinCap) {\n      // Merge: extend the range (may include some skipped bytes, but avoids extra request)\n      current.end = Math.max(current.end, entryEnd);\n      current.entries.push(sorted[i]);\n    } else {\n      ranges.push(current);\n      current = { start: entryStart, end: entryEnd, entries: [sorted[i]] };\n    }\n  }\n  ranges.push(current);\n  return ranges;\n}\n\n/**\n * Selectively download needed tensors from a safetensors file using Range requests.\n * Returns a map of tensor name → { data, shape }.\n */\n/** Byte alignment required for a safetensors dtype's typed-array view. */\nfunction dtypeAlign(dtype: string): number {\n  if (dtype === \"F64\" || dtype === \"I64\" || dtype === \"U64\") return 8;\n  if (dtype === \"F32\" || dtype === \"I32\" || dtype === \"U32\") return 4;\n  if (dtype === \"F16\" || dtype === \"BF16\" || dtype === \"I16\" || dtype === \"U16\") return 2;\n  return 1;\n}\n\n/** Build the appropriate typed-array view of a tensor at `byteOffset` in `buf`. */\nfunction makeTensorView(\n  buf: ArrayBuffer,\n  byteOffset: number,\n  entry: SafetensorEntry,\n): ArrayBufferView {\n  // Aligned views can alias the buffer directly; unaligned ones must be copied\n  // into a fresh (offset-0) buffer so the typed-array constructor accepts them.\n  const aligned = byteOffset % dtypeAlign(entry.dtype) === 0;\n  const src = aligned ? buf : buf.slice(byteOffset, byteOffset + entry.dataLength);\n  const off = aligned ? byteOffset : 0;\n  switch (entry.dtype) {\n    case \"F32\":\n      return new Float32Array(src, off, entry.dataLength / 4);\n    case \"I32\":\n      return new Int32Array(src, off, entry.dataLength / 4);\n    case \"U32\":\n      return new Uint32Array(src, off, entry.dataLength / 4);\n    case \"F16\":\n    case \"BF16\":\n      return new Uint16Array(src, off, entry.dataLength / 2);\n    default:\n      return new Uint8Array(src, off, entry.dataLength);\n  }\n}\n\n/**\n * Stable per-tensor browser-cache key. Keyed by the tensor's *absolute* byte\n * range in the file (offset + length), which is fixed regardless of which other\n * tensors a given load requests. This is what lets a vision load reuse the\n * tensors a prior text load already downloaded — they hit the same keys —\n * instead of re-downloading because the merged Range boundaries differ.\n */\nfunction tensorCacheKey(url: string, file: SafetensorsFile, entry: SafetensorEntry): string {\n  const absStart = file.dataStart + entry.dataOffset;\n  return `${url}#t:${absStart}-${entry.dataLength}`;\n}\n\n/**\n * Selectively download needed tensors from a safetensors file using Range\n * requests.\n *\n * Two modes, gated by `cacheStore`:\n *\n *  • `cacheStore` supplied (browser): each tensor's bytes are written to the\n *    per-tensor download cache (as before) and the tensor is REGISTERED in the\n *    store as a cache-backed descriptor under its canonical name — its bytes are\n *    NOT retained in heap. The returned Map is empty. This is the change that\n *    bounds peak heap during download: instead of accumulating the whole model\n *    in the returned Map, the heap only ever holds the current range segment\n *    (≤ maxRangeBytes, 24 MB on mobile) which is released each iteration.\n *\n *  • `cacheStore` omitted (used by direct-load helpers): returns a Map of\n *    name → { data, shape } exactly as before.\n */\nasync function fetchSelectiveTensors(\n  url: string,\n  file: SafetensorsFile,\n  neededEntries: SafetensorEntry[],\n  hfToken?: string,\n  onProgress?: (loaded: number, total: number, message?: string) => void,\n  cacheStore?: { store: CacheWeightStore; keyMapper: HFKeyMapper },\n): Promise<Map<string, { data: ArrayBufferView; shape: number[] }>> {\n  const results = new Map<string, { data: ArrayBufferView; shape: number[] }>();\n  const totalBytes = neededEntries.reduce((sum, e) => sum + e.dataLength, 0);\n  let loadedBytes = 0;\n\n  const isMobile =\n    typeof navigator !== \"undefined\" && /iPhone|iPad|iPod|Android/i.test(navigator.userAgent);\n  // Cap merged-range size so the download streams in bounded segments: keeps\n  // peak transient memory low (critical on iOS) and lets the progress bar move.\n  const maxRangeBytes = isMobile ? 24 * 1024 * 1024 : 256 * 1024 * 1024;\n\n  // Register a downloaded tensor: either retain it in the result Map (Node-ish\n  // direct loads) or — in browser cache mode — record only a cache-backed\n  // descriptor under the canonical name so no bytes stay in heap.\n  const record = (entry: SafetensorEntry, key: string): void => {\n    if (cacheStore) {\n      const canonical = cacheStore.keyMapper(entry.name);\n      if (canonical) {\n        // Bytes live in the per-tensor download cache (BROWSER_CACHE_NAME).\n        cacheStore.store.registerCached(\n          canonical,\n          entry.shape,\n          entry.dtype,\n          key,\n          BROWSER_CACHE_NAME,\n        );\n      }\n      return;\n    }\n    // Fallback (no cache backend): caller still wants bytes in the Map.\n    // (This branch is only reachable when CacheStorage is unavailable.)\n  };\n\n  // Reclaim quota from the dead legacy CacheStorage/OPFS caches before we\n  // probe/write IDB — runs on the default (heap) browser path too, since that's\n  // where per-tensor IDB caching now happens.\n  // Fire-and-forget: legacy-cache cleanup must NEVER block (or hang) the load —\n  // OPFS dir iteration can stall on Safari, and it's only reclaiming dead junk.\n  if (TENSOR_CACHE_ENABLED) void evictStaleCaches();\n\n  // Pass 1: serve any tensor already in the per-tensor browser cache. A tensor\n  // cached by an earlier load (e.g. text-only) is reused here even if this load\n  // (e.g. enableVision) needs a different overall set — the key is per-tensor.\n  // Probe in parallel: hundreds of sequential CacheStorage.match() calls are slow\n  // on mobile and would make a fresh load look frozen before any download starts.\n  // In cache mode we only need to know WHICH tensors are present (HEAD-style), not\n  // their bytes — so probe with `caches.match` but discard bodies immediately.\n  const uncached: SafetensorEntry[] = [];\n  // Probe presence CHEAPLY (getKey, no bodies) in parallel. Reading the actual\n  // bytes for hundreds of tensors at once (the old heap-path Promise.all over\n  // browserCacheRead) stormed IndexedDB with ~696 concurrent transactions AND\n  // pulled ~404 MB into heap in one burst → Safari hung on a warm cache (the lean\n  // harness only ever hit a cold cache, so it never showed). Presence is cheap and\n  // parallel-safe; the heap path then reads the HIT bytes one at a time below.\n  // ONE getAllKeys() transaction, then membership in memory — NOT hundreds of\n  // parallel getKey transactions (that storm hangs Safari on a warm cache).\n  const cachedKeys = TENSOR_CACHE_ENABLED ? await idbAllKeys() : new Set<string>();\n  const presence = neededEntries.map((entry) => cachedKeys.has(tensorCacheKey(url, file, entry)));\n  for (let i = 0; i < neededEntries.length; i++) {\n    const entry = neededEntries[i];\n    const key = tensorCacheKey(url, file, entry);\n    if (!presence[i]) {\n      uncached.push(entry);\n      continue;\n    }\n    if (cacheStore) {\n      // Presence-only (the store reads bytes lazily later).\n      record(entry, key);\n      loadedBytes += entry.dataLength;\n      onProgress?.(loadedBytes, totalBytes);\n      continue;\n    }\n    // Heap path: read the cached bytes ONE AT A TIME (never all in parallel — that\n    // storms IDB and OOMs Safari). VALIDATE size; a partial/corrupt entry is\n    // treated as a miss so pass 2 re-downloads and overwrites it (self-healing).\n    const buf = await browserCacheRead(key);\n    if (buf && buf.byteLength >= entry.dataLength) {\n      results.set(entry.name, {\n        data: makeTensorView(buf, 0, entry),\n        shape: entry.shape,\n      });\n      loadedBytes += entry.dataLength;\n      onProgress?.(loadedBytes, totalBytes);\n    } else {\n      uncached.push(entry);\n    }\n  }\n\n  // Cache visibility: report how many tensors were reused vs must be downloaded,\n  // and which backend served them — so a silent \"redownloads everything\" cache\n  // miss is diagnosable on-device (most users have no Safari console). Surfaces\n  // through onProgress, which the iPad runner streams back to the dashboard.\n  if (TENSOR_CACHE_ENABLED) {\n    const hits = neededEntries.length - uncached.length;\n    const backend = idbCacheAvailable() ? \"idb\" : \"none\";\n    const msg = `cache[${backend}]: ${hits}/${neededEntries.length} tensors reused, ${uncached.length} to download`;\n    console.log(`[gerbil] ${msg}`);\n    onProgress?.(loadedBytes, totalBytes, msg);\n  }\n\n  // Pass 2: download only the uncached tensors (merged into bounded contiguous\n  // ranges), then cache each one individually by its stable key.\n  const ranges = mergeRanges(uncached, file.dataStart, 1024 * 1024, maxRangeBytes);\n  for (const range of ranges) {\n    const rangeBuf = await fetchRangeRaw(url, range.start, range.end, hfToken);\n    // A truncated range response (e.g. an aborted/partial download) would yield\n    // short tensor slices — bad views now and a poisoned cache later. Reject it\n    // loudly instead of silently caching garbage.\n    const expectedRangeBytes = range.end - range.start;\n    if (rangeBuf.byteLength < expectedRangeBytes) {\n      throw new Error(\n        `Incomplete weight download: got ${rangeBuf.byteLength} of ${expectedRangeBytes} bytes for ${url}. Retry.`,\n      );\n    }\n\n    for (const entry of range.entries) {\n      const localOffset = file.dataStart + entry.dataOffset - range.start;\n      const key = tensorCacheKey(url, file, entry);\n      // Cache the exact tensor bytes (a standalone copy) for cross-load reuse.\n      const tensorBytes = rangeBuf.slice(localOffset, localOffset + entry.dataLength);\n      // Only cache a full-size slice so we never write a short (poisoning) entry.\n      if (TENSOR_CACHE_ENABLED && tensorBytes.byteLength === entry.dataLength) {\n        // `tensorBytes` is a standalone slice not reused below (the heap path\n        // views `rangeBuf`, a different buffer), so it can be transferred to the\n        // OPFS worker zero-copy — this is the big-memory path, so avoiding a copy\n        // matters most here.\n        await browserCacheWrite(key, tensorBytes, true);\n      }\n      if (cacheStore) {\n        // Bytes are now durably in cache; do NOT retain them in heap.\n        record(entry, key);\n      } else {\n        results.set(entry.name, {\n          data: makeTensorView(rangeBuf, localOffset, entry),\n          shape: entry.shape,\n        });\n      }\n      loadedBytes += entry.dataLength;\n      onProgress?.(loadedBytes, totalBytes);\n    }\n    // `rangeBuf` (and its slices for the non-cache path) go out of scope here;\n    // in cache mode nothing references it, so the segment is freed before the\n    // next range is fetched — peak heap stays ≈ one range segment.\n  }\n\n  return results;\n}\n\n/**\n * Create the appropriate HF key mapper for a given architecture.\n *\n * Qwen3.5 conditional generation models use \"model.language_model.\" prefix for\n * text model weights and include visual encoder + MTP weights that we skip.\n * Standard models just use \"model.\" prefix.\n *\n * `wantVision` controls whether the vision tower is kept. When false (text-only\n * loads) the ViT is dropped from the selective download so the ~201MB tower is\n * never fetched. When true (enableVision) the tower is mapped to the canonical\n * \"visual.*\" keys the vision graph expects, regardless of whether the source\n * checkpoint names it \"model.visual.*\" (HF BF16) or \"vision_tower.*\" (MLX).\n */\nfunction createKeyMapperForArch(architectureName: string, wantVision = false): HFKeyMapper {\n  // KaniTTS2 ships the LFM2 backbone verbatim — the only structural rename it needs\n  // is the same self_attn.out_proj → o_proj as LFM2. Its extra tensors\n  // (learnable_rope_layers.*, speaker_emb_projection) pass through after stripping\n  // the \"model.\" prefix and are picked up by the codec-LM graph when it lands.\n  if (architectureName === \"Lfm2ForCausalLM\" || architectureName === \"KaniTTS2ForCausalLM\") {\n    // LFM2 weights live under \"model.\" like the others, but the attention\n    // output projection is named \"out_proj\" where the canonical IR (shared with\n    // Qwen) expects \"o_proj\". Rename it so the o_proj graph tensors resolve.\n    // All other LFM2 keys (operator_norm, ffn_norm, embedding_norm, conv.*,\n    // feed_forward.w1/w2/w3, q_layernorm/k_layernorm) are referenced verbatim\n    // by the generator via per-tensor safetensorsKey, so they pass through.\n    return (hfKey: string): string | null => {\n      let key = hfKey;\n      if (key.startsWith(\"model.\")) {\n        key = key.slice(6);\n      }\n      key = key.replace(/\\.self_attn\\.out_proj\\./, \".self_attn.o_proj.\");\n      return key;\n    };\n  }\n  if (architectureName === \"Qwen3_5ForConditionalGeneration\") {\n    return (hfKey: string): string | null => {\n      // Vision tower lives under one of two prefixes depending on the converter:\n      //   • \"model.visual.*\"  — HF BF16 checkpoint (Qwen/Qwen3.5-0.8B)\n      //   • \"vision_tower.*\"   — mlx-vlm 4-bit convert (mlx-community/…-4bit).\n      //     In MLX-4bit repos the text tower is quantized (U32 + scales/biases)\n      //     but the ViT is left BF16, so the suffix structure is identical to\n      //     the HF tower — only the prefix differs. Both map to the canonical\n      //     \"visual.*\" keys the vision graph generator expects.\n      // When vision is not requested we drop the tower so the selective\n      // downloader never fetches its ~201MB of weights.\n      if (hfKey.startsWith(\"model.visual.\")) {\n        return wantVision ? hfKey.slice(6) : null; // \"model.visual.*\" → \"visual.*\"\n      }\n      if (hfKey.startsWith(\"vision_tower.\")) {\n        // \"vision_tower.\".length === 13 → replace with \"visual.\"\n        return wantVision ? `visual.${hfKey.slice(13)}` : null;\n      }\n      // Always skip the multi-token-prediction head (an inference-speed\n      // optimization, not a modality).\n      if (hfKey.startsWith(\"mtp.\")) {\n        return null;\n      }\n      let key = hfKey;\n      // HF/GPTQ uses \"model.language_model.\" prefix, MLX uses \"language_model.model.\"\n      if (key.startsWith(\"model.language_model.\")) {\n        key = key.slice(21); // \"model.language_model.\".length\n      } else if (key.startsWith(\"language_model.model.\")) {\n        key = key.slice(21); // \"language_model.model.\".length (also 21 chars)\n      } else if (key.startsWith(\"model.\")) {\n        key = key.slice(6); // \"model.\".length\n      }\n      return key;\n    };\n  }\n  if (architectureName === \"Gemma4ForConditionalGeneration\") {\n    // Gemma 4 nests the text tower under \"language_model.\" and carries vision +\n    // audio towers we drop for text-only loads. Prefix shapes observed:\n    //   HF BF16:  \"model.language_model.*\"  (+ \"model.vision_tower.*\", \"model.audio_tower.*\",\n    //             \"model.embed_vision.*\", \"model.embed_audio.*\")\n    //   MLX-4bit: \"language_model.model.*\"  (+ \"vision_tower.*\", \"audio_tower.*\")\n    // The text tower keys after stripping match the canonical Gemma layout used\n    // by gemma4.ts (embed_tokens, embed_tokens_per_layer, per_layer_*, layers.N.*,\n    // norm). PLE/projection tensors pass through verbatim.\n    return (hfKey: string): string | null => {\n      // Vision tower: keep + canonicalize when wantVision; otherwise drop so the\n      // selective downloader never fetches it. Both prefixes (\"model.vision_tower.*\"\n      // HF BF16, \"vision_tower.*\" MLX) and the multimodal projector\n      // (\"model.embed_vision.*\" / \"embed_vision.*\") map to the canonical\n      // \"vision_tower.*\" / \"embed_vision.*\" keys gemma4_vision.ts expects.\n      if (hfKey.includes(\"vision_tower.\")) {\n        if (!wantVision) return null;\n        const idx = hfKey.indexOf(\"vision_tower.\");\n        return hfKey.slice(idx); // strip any \"model.\" prefix before vision_tower.\n      }\n      if (hfKey.includes(\"embed_vision.\")) {\n        if (!wantVision) return null;\n        const idx = hfKey.indexOf(\"embed_vision.\");\n        return hfKey.slice(idx);\n      }\n      // Always drop the audio modality (not supported in the vision path).\n      if (hfKey.includes(\"audio_tower.\") || hfKey.includes(\"embed_audio.\")) {\n        return null;\n      }\n      let key = hfKey;\n      if (key.startsWith(\"model.language_model.\")) {\n        key = key.slice(21); // \"model.language_model.\".length\n      } else if (key.startsWith(\"language_model.model.\")) {\n        key = key.slice(21); // \"language_model.model.\".length (also 21)\n      } else if (key.startsWith(\"language_model.\")) {\n        key = key.slice(15); // \"language_model.\".length (lm_head etc.)\n      } else if (key.startsWith(\"model.\")) {\n        key = key.slice(6);\n      }\n      return key;\n    };\n  }\n  if (architectureName === \"MoonshineForConditionalGeneration\") {\n    // Moonshine's HF keys already match the canonical Moonshine graph names once\n    // the \"model.\" prefix is stripped, e.g.\n    //   model.encoder.conv1.weight                       → encoder.conv1.weight\n    //   model.encoder.layers.0.self_attn.q_proj.weight   → encoder.layers.0.self_attn.q_proj.weight\n    //   model.decoder.layers.0.encoder_attn.q_proj.weight→ decoder.layers.0.encoder_attn.q_proj.weight\n    //   model.decoder.embed_tokens.weight                → decoder.embed_tokens.weight (tied lm_head)\n    // The default \"strip model.\" mapper is therefore exactly correct. (Moonshine\n    // weight loading runs through the dual-graph STT loader — remaining work —\n    // not the single-graph generate() path, which throws by design.)\n    return createDefaultHFKeyMapper();\n  }\n  return createDefaultHFKeyMapper();\n}\n\nconst PLE_CACHE_NAME = \"gerbil-ple-v2\";\n\n/**\n * Build a `PleSource` from the quantized PLE table.\n *\n * Node/desktop (`cacheStore` null): keep the arrays heap-resident — unchanged\n * behavior, the executor streams rows directly from them.\n *\n * Browser (`cacheStore` set): write the (already-compact INT4) packed/scales/\n * zeros to CacheStorage and return EMPTY heap arrays plus a `cache` descriptor.\n * This keeps the ~1.17 GB table out of the JS heap during the load/upload peak;\n * the executor lazily materializes it from cache after the GPU upload has\n * completed (past the memory high-water mark). See Executor.setPleSource.\n */\nasync function buildPleSource(\n  packed: Uint32Array,\n  scales: Float32Array,\n  zeros: Float32Array,\n  width: number,\n  groupSize: number,\n  cacheStore: CacheWeightStore | null,\n): Promise<PleSource> {\n  if (!cacheStore) {\n    return { packed, scales, zeros, width, groupSize, targetTensor: \"ple_embed_out\" };\n  }\n  const cache = await caches.open(PLE_CACHE_NAME);\n  const packedKey = \"ple:packed\";\n  const scalesKey = \"ple:scales\";\n  const zerosKey = \"ple:zeros\";\n  const copy = (a: ArrayBufferView): ArrayBuffer =>\n    a.buffer.slice(a.byteOffset, a.byteOffset + a.byteLength) as ArrayBuffer;\n  await cache.put(new Request(packedKey), new Response(copy(packed)));\n  await cache.put(new Request(scalesKey), new Response(copy(scales)));\n  await cache.put(new Request(zerosKey), new Response(copy(zeros)));\n  return {\n    // Empty heap arrays — the bytes live in cache now.\n    packed: new Uint32Array(0),\n    scales: new Float32Array(0),\n    zeros: new Float32Array(0),\n    width,\n    groupSize,\n    targetTensor: \"ple_embed_out\",\n    cache: {\n      cacheName: PLE_CACHE_NAME,\n      packedKey,\n      scalesKey,\n      zerosKey,\n      packedLen: packed.length,\n    },\n  };\n}\n\n/**\n * Load a model from HuggingFace Hub.\n *\n * 1. Fetch config.json -> determine architecture -> generate IR graph\n * 2. Fetch tokenizer.json + tokenizer_config.json -> build tokenizer\n * 3. Download safetensors -> parse headers -> extract weight data\n * 4. Map HF tensor keys -> canonical names\n */\nexport async function loadModel(options: LoadModelOptions): Promise<LoadedModel> {\n  const { repo, onProgress, hfToken, revision = \"main\", cacheDir, dtype } = options;\n\n  // Initialize Node.js fs for caching (no-op in browser)\n  await initFs();\n\n  // Resolve cache directory: use provided path or default to ~/.cache/gerbil/<repo>\n  let resolvedCacheDir = cacheDir;\n  if (!resolvedCacheDir && _fs && _path) {\n    const home = process.env.HOME || process.env.USERPROFILE || \"\";\n    if (home) {\n      resolvedCacheDir = _path.join(home, \".cache\", \"gerbil\", repo.replace(/\\//g, \"_\"), revision);\n    }\n  }\n\n  const baseURL = resolveHFBaseURL(repo, revision);\n\n  // -- Step 1: Fetch config.json --\n  onProgress?.(0, 100, \"Fetching model config...\");\n  const rawConfig = await fetchJSON(baseURL, \"config.json\", hfToken, resolvedCacheDir);\n\n  const architectureName = rawConfig.architectures?.[0];\n  if (!architectureName) {\n    throw new Error(\"config.json missing 'architectures' field\");\n  }\n\n  // Detect pre-quantized formats — force q4 graph when loading quantized models.\n  // MLX-lm writes the quant block under EITHER \"quantization_config\" (newer\n  // converts, often with mode:\"affine\") OR \"quantization\" (standard mlx-lm,\n  // typically just {bits, group_size} with no mode field). Read whichever exists.\n  const quantConfig = (rawConfig.quantization_config ?? rawConfig.quantization) as\n    | Record<string, unknown>\n    | undefined;\n  const isGPTQ = quantConfig?.quant_method === \"gptq\";\n\n  // MLX 4-bit detection. Two shapes both indicate a standard affine MLX convert:\n  //   1. {bits:4, mode:\"affine\", ...}                (newer converts)\n  //   2. {bits:4, group_size:N}  (no mode field)     (standard mlx-lm)\n  // CAUTION: DWQ converts are config-INDISTINGUISHABLE from (2) — both write only\n  // {bits:4, group_size}. The repack treats the weights as affine MLX, which\n  // silently produces garbage for DWQ. So a no-mode MLX config is only accepted\n  // for repos we've verified are standard (non-DWQ) affine converts. A verified\n  // load must still be sanity-checked at runtime (cosine reference) — see the\n  // embedding validation scripts.\n  const quantMode = quantConfig?.mode as string | undefined;\n  const hasMlxShape =\n    !isGPTQ &&\n    quantConfig?.bits === 4 &&\n    (quantMode === \"affine\" ||\n      (quantMode === undefined && typeof quantConfig?.group_size === \"number\"));\n\n  // Allowlist of repos verified to be standard (non-DWQ) MLX-4bit converts.\n  // Match is substring-insensitive so revision-pinned / mirrored repo strings\n  // still resolve. DWQ repos (…-4bit-DWQ) must NOT appear here.\n  const VERIFIED_MLX_REPOS = [\"mlx-community/embeddinggemma-300m-4bit\"];\n  const repoLower = repo.toLowerCase();\n  const isDWQ = repoLower.includes(\"dwq\");\n  const isVerifiedMlxRepo =\n    !isDWQ && VERIFIED_MLX_REPOS.some((r) => repoLower.includes(r.toLowerCase()));\n\n  // Accept MLX when the config explicitly says affine, OR when the no-mode shape\n  // is present AND the repo is on the verified allowlist (DWQ-trap guard).\n  const isMLX = hasMlxShape && (quantMode === \"affine\" || isVerifiedMlxRepo);\n  // Resolve \"auto\": on mobile (iOS/Android) prefer on-the-fly q4 so large f16\n  // checkpoints (e.g. LFM2.5-350M ~676 MB f16 → ~190 MB q4) fit in device memory;\n  // on desktop keep the repo's native precision. Already-quantized repos (MLX/\n  // GPTQ) are always q4.\n  const isMobileUA =\n    typeof navigator !== \"undefined\" && /iPhone|iPad|iPod|Android/i.test(navigator.userAgent);\n  const requestedDtype: GraphDType | undefined =\n    dtype === \"auto\" ? (isMobileUA ? \"q4\" : undefined) : dtype;\n  const resolvedDtype: GraphDType | undefined = isGPTQ || isMLX ? \"q4\" : requestedDtype;\n\n  // MLX uses a different group_size (typically 64 vs Gerbil default 128)\n  const quantGroupSize = isMLX ? (quantConfig?.group_size as number) : undefined;\n\n  if (isGPTQ) {\n    onProgress?.(2, 100, \"Detected GPTQ model — will repack to native INT4...\");\n  } else if (isMLX) {\n    onProgress?.(\n      2,\n      100,\n      `Detected MLX 4-bit model (group_size=${quantGroupSize}) — will repack...`,\n    );\n  }\n\n  // Select key mapper: Qwen3.5 conditional generation models have weights\n  // under \"model.language_model.\" prefix instead of the standard \"model.\" prefix.\n  // Vision is wanted when the caller requested the multimodal graph variant\n  // (enableVision) or explicitly asked for the tower (loadVisionTower — used by\n  // the direct-load vision validation scripts). Gate the ViT key-mapping on this\n  // so text-only loads stay lean (the ~201MB tower is excluded from download).\n  const wantVision = Boolean(options.multimodal || options.loadVisionTower);\n  const keyMapper = options.keyMapper ?? createKeyMapperForArch(architectureName, wantVision);\n\n  // Generate the computation graph (with INT4 ops if dtype is \"q4\" or GPTQ/MLX detected)\n  const graph = generateGraph(\n    architectureName,\n    rawConfig,\n    resolvedDtype,\n    quantGroupSize,\n    options.kvDtype,\n    options.embedding,\n    options.multimodal,\n  );\n\n  // -- Step 2: Fetch tokenizer --\n  onProgress?.(5, 100, \"Fetching tokenizer...\");\n  const [tokenizerJSON, tokenizerConfigJSON, chatTemplateJinja] = await Promise.all([\n    fetchJSON(baseURL, \"tokenizer.json\", hfToken, resolvedCacheDir),\n    fetchJSON(baseURL, \"tokenizer_config.json\", hfToken, resolvedCacheDir).catch(() => null),\n    // Some models (e.g. LFM2) ship the chat template as a sidecar `.jinja` file\n    // instead of inline in tokenizer_config.json. Fetch it as a fallback signal.\n    fetchText(baseURL, \"chat_template.jinja\", hfToken, resolvedCacheDir).catch(() => null),\n  ]);\n\n  // Prefer the inline template; otherwise fall back to the sidecar .jinja.\n  const configWithTemplate =\n    tokenizerConfigJSON && !tokenizerConfigJSON.chat_template && chatTemplateJinja\n      ? { ...tokenizerConfigJSON, chat_template: chatTemplateJinja }\n      : (tokenizerConfigJSON ?? (chatTemplateJinja ? { chat_template: chatTemplateJinja } : null));\n\n  const tokenizer = Tokenizer.fromJSON(tokenizerJSON, configWithTemplate);\n\n  // -- Step 3: Download safetensors --\n  onProgress?.(10, 100, \"Discovering weight files...\");\n  const safetensorsFiles = await discoverSafetensorsFiles(baseURL, hfToken, resolvedCacheDir);\n\n  // Cache-backed streaming store keeps the browser heap bounded (tensor bytes\n  // live in CacheStorage; the heap holds only descriptors) so very large models\n  // load on mobile without OOM. It is the DEFAULT on mobile — the heap-backed\n  // path keeps every tensor resident in the JS heap on top of the GPU buffers\n  // (and expands f16/bf16 -> f32 in-heap), doubling peak memory and OOM-crashing\n  // iOS WebKit right after the safetensors header read. Desktop/Node keep the\n  // proven heap path. Opt in elsewhere with `?stream=1` /\n  // `globalThis.GERBIL_STREAM_WEIGHTS = true`; force the heap path anywhere with\n  // `globalThis.GERBIL_STREAM_WEIGHTS = false`.\n  const streamFlag = (globalThis as { GERBIL_STREAM_WEIGHTS?: boolean }).GERBIL_STREAM_WEIGHTS;\n  const streamOptIn =\n    streamFlag === true ||\n    (typeof location !== \"undefined\" &&\n      typeof URLSearchParams !== \"undefined\" &&\n      new URLSearchParams(location.search).has(\"stream\"));\n  const useCacheBacked =\n    streamFlag !== false && (streamOptIn || isMobileUA) && cacheStorageAvailable();\n  const cacheStore = useCacheBacked ? new CacheWeightStore() : null;\n  const store: MutableWeightStore = cacheStore ?? new HeapWeightStore();\n\n  let filesLoaded = 0;\n  const totalFiles = safetensorsFiles.length;\n\n  for (const filename of safetensorsFiles) {\n    const fileUrl = `${baseURL}/${filename}`;\n\n    // Check cache first — if cached, use the old full-buffer path\n    let cachedBuffer: ArrayBuffer | null = null;\n    if (resolvedCacheDir) {\n      const cached = readCachedFile(resolvedCacheDir, filename);\n      if (cached) {\n        cachedBuffer = cached.buffer.slice(\n          cached.byteOffset,\n          cached.byteOffset + cached.byteLength,\n        ) as ArrayBuffer;\n      }\n    }\n\n    if (cachedBuffer) {\n      // Cached (Node disk cache) — extract all needed tensors from full buffer.\n      const file = parseSafetensorsHeader(cachedBuffer);\n      for (const entry of file.entries) {\n        const canonicalName = keyMapper(entry.name);\n        if (!canonicalName) continue;\n        let data: ArrayBufferView = getTensorData(cachedBuffer, file, entry);\n        if (entry.dtype === \"BF16\") {\n          data = bf16ToF32(new Uint8Array(data.buffer, data.byteOffset, data.byteLength));\n        }\n        if (entry.dtype === \"F16\") {\n          data = f16ToF32(data);\n        }\n        await store.set(canonicalName, { data, shape: entry.shape });\n      }\n    } else {\n      // Not cached — use selective Range-request downloading\n      const fileWeight = 85 / totalFiles;\n      const base = 10 + filesLoaded * fileWeight;\n\n      // Step 1: Fetch header only. This + the index probe above can take a few\n      // seconds with no bytes yet — without this emit the bar looks frozen at 10%\n      // (\"stuck at discovering weight files\") until the first chunk lands.\n      onProgress?.(base, 100, `Reading ${filename} header (${filesLoaded + 1}/${totalFiles})...`);\n      const file = await fetchSafetensorsHeader(fileUrl, hfToken);\n\n      // Step 2: Filter to needed entries\n      const neededEntries = file.entries.filter((e) => keyMapper(e.name) !== null);\n      const skippedBytes = file.entries\n        .filter((e) => keyMapper(e.name) === null)\n        .reduce((sum, e) => sum + e.dataLength, 0);\n      const neededBytes = neededEntries.reduce((sum, e) => sum + e.dataLength, 0);\n\n      const totalMB = (neededBytes / 1048576).toFixed(0);\n      if (skippedBytes > 0) {\n        const savedMB = (skippedBytes / 1048576).toFixed(0);\n        onProgress?.(\n          base,\n          100,\n          `Selective download: ${totalMB} MB needed, skipping ${savedMB} MB (vision/MTP)`,\n        );\n      }\n      // Show the size up front so the first (latency-heavy) chunk doesn't read as a freeze.\n      onProgress?.(base, 100, `Downloading ${filename} (0/${totalMB} MB)`);\n\n      // Step 3: Download needed tensors via Range requests\n      let lastReportedPct = -1;\n\n      const rawTensors = await fetchSelectiveTensors(\n        fileUrl,\n        file,\n        neededEntries,\n        hfToken,\n        (loaded, total) => {\n          const filePct = total > 0 ? (loaded / total) * fileWeight : 0;\n          const overallPct = Math.round(base + filePct);\n          if (overallPct <= lastReportedPct) return;\n          lastReportedPct = overallPct;\n          const sizeMB =\n            total > 0\n              ? ` (${(loaded / 1048576).toFixed(0)}/${(total / 1048576).toFixed(0)} MB)`\n              : \"\";\n          onProgress?.(base + filePct, 100, `Downloading ${filename}${sizeMB}`);\n        },\n        // Browser: register tensors as cache-backed descriptors under their\n        // canonical names — no bytes retained in heap. Node: omitted, so the\n        // returned Map carries the bytes (handled below).\n        cacheStore ? { store: cacheStore, keyMapper } : undefined,\n      );\n\n      // Step 4 (Node / non-cache only): map to canonical names with dtype\n      // conversion. In cache-backed mode `rawTensors` is empty — tensors were\n      // already registered into the store with the canonical name + source dtype,\n      // and `store.get()` applies BF16/F16→F32 on demand.\n      for (const [hfName, { data: rawData, shape }] of rawTensors) {\n        const canonicalName = keyMapper(hfName);\n        if (!canonicalName) continue;\n\n        let data: ArrayBufferView = rawData;\n        // Find the entry to check dtype\n        const entry = neededEntries.find((e) => e.name === hfName);\n        if (entry) {\n          if (entry.dtype === \"BF16\") {\n            data = bf16ToF32(new Uint8Array(data.buffer, data.byteOffset, data.byteLength));\n          }\n          if (entry.dtype === \"F16\") {\n            data = f16ToF32(data);\n          }\n        }\n\n        await store.set(canonicalName, { data, shape });\n      }\n    }\n\n    filesLoaded++;\n  }\n\n  onProgress?.(95, 100, \"Weights loaded.\");\n\n  // -- Split fused Q projection for Qwen3.5 full attention layers --\n  // HF stores q_proj.weight as [num_heads * 2 * head_dim, hidden_size].\n  // The layout is per-head interleaved: for each head h, rows\n  //   [h*2*head_dim .. h*2*head_dim + head_dim) are Q weights, and\n  //   [h*2*head_dim + head_dim .. (h+1)*2*head_dim) are gate weights.\n  // We deinterleave into q_proj.weight [num_heads*head_dim, hidden_size]\n  // and attn_gate.weight [num_heads*head_dim, hidden_size].\n  if (architectureName === \"Qwen3_5ForConditionalGeneration\") {\n    const textCfg = (rawConfig.text_config as Record<string, unknown>) ?? rawConfig;\n    const numLayers = textCfg.num_hidden_layers as number;\n    const layerTypesArr = (textCfg.layer_types as string[]) ?? [];\n    const fullAttnInterval = (textCfg.full_attention_interval as number) ?? 4;\n    const numHeads = textCfg.num_attention_heads as number; // 8\n    const headDim =\n      (textCfg.head_dim as number) ?? Math.floor((textCfg.hidden_size as number) / numHeads);\n\n    for (let i = 0; i < numLayers; i++) {\n      // Check if this is a full attention layer\n      let isFullAttn: boolean;\n      if (layerTypesArr.length > i) {\n        isFullAttn = layerTypesArr[i] === \"full_attention\";\n      } else {\n        isFullAttn = i % fullAttnInterval === fullAttnInterval - 1;\n      }\n      if (!isFullAttn) continue;\n\n      const qProjKey = `layers.${i}.self_attn.q_proj.weight`;\n      const attnGateKey = `layers.${i}.self_attn.attn_gate.weight`;\n      const fusedWeight = await store.get(qProjKey);\n\n      if (fusedWeight && !store.has(attnGateKey) && !isMLX) {\n        const fusedData = fusedWeight.data;\n        const [fusedRows, cols] = fusedWeight.shape;\n        const halfRows = fusedRows / 2; // num_heads * head_dim\n\n        const qData = new Float32Array(halfRows * cols);\n        const gateData = new Float32Array(halfRows * cols);\n\n        const src =\n          fusedData instanceof Float32Array\n            ? fusedData\n            : new Float32Array(fusedData.buffer, fusedData.byteOffset, fusedRows * cols);\n\n        // Deinterleave per-head: each head has a block of 2*head_dim rows\n        // First head_dim rows = Q, second head_dim rows = gate\n        const headBlock = 2 * headDim; // rows per head in fused weight\n        for (let h = 0; h < numHeads; h++) {\n          const srcBase = h * headBlock * cols;\n          const dstBase = h * headDim * cols;\n          qData.set(src.subarray(srcBase, srcBase + headDim * cols), dstBase);\n          gateData.set(src.subarray(srcBase + headDim * cols, srcBase + headBlock * cols), dstBase);\n        }\n\n        await store.set(qProjKey, { data: qData, shape: [halfRows, cols] });\n        await store.set(attnGateKey, { data: gateData, shape: [halfRows, cols] });\n      }\n\n      // GPTQ: split fused q_proj GPTQ tensors into q_proj + attn_gate\n      // GPTQ q_proj has N=4096 (fused), graph expects N=2048 each.\n      // The deinterleave is per-head: each head's block of 2*head_dim output columns\n      // has head_dim Q columns followed by head_dim gate columns.\n      const qProjQweight = `layers.${i}.self_attn.q_proj.qweight`;\n      const attnGateQweight = `layers.${i}.self_attn.attn_gate.qweight`;\n      if (store.has(qProjQweight) && !store.has(attnGateQweight)) {\n        const fusedN = numHeads * 2 * headDim; // 4096\n        const halfN = numHeads * headDim; // 2048\n        const K_val = textCfg.hidden_size as number; // 1024\n\n        // Split qweight [K/8, fusedN] → two [K/8, halfN] with per-head deinterleave\n        const qw = (await store.get(qProjQweight))!;\n        const qwSrc =\n          qw.data instanceof Int32Array\n            ? qw.data\n            : new Int32Array(qw.data.buffer, qw.data.byteOffset, qw.data.byteLength / 4);\n        const kDiv8 = K_val >>> 3;\n        const qwQ = new Int32Array(kDiv8 * halfN);\n        const qwG = new Int32Array(kDiv8 * halfN);\n        const headBlock2 = 2 * headDim; // output columns per head in fused tensor\n        for (let row = 0; row < kDiv8; row++) {\n          for (let h = 0; h < numHeads; h++) {\n            const srcOff = row * fusedN + h * headBlock2;\n            const dstOff = row * halfN + h * headDim;\n            for (let d = 0; d < headDim; d++) {\n              qwQ[dstOff + d] = qwSrc[srcOff + d];\n              qwG[dstOff + d] = qwSrc[srcOff + headDim + d];\n            }\n          }\n        }\n        await store.set(qProjQweight, { data: qwQ, shape: [kDiv8, halfN] });\n        await store.set(attnGateQweight, { data: qwG, shape: [kDiv8, halfN] });\n\n        // Split scales [groups, fusedN] → two [groups, halfN]\n        const qProjScales = `layers.${i}.self_attn.q_proj.scales`;\n        const attnGateScales = `layers.${i}.self_attn.attn_gate.scales`;\n        const sc = (await store.get(qProjScales))!;\n        const scSrc =\n          sc.data instanceof Float32Array\n            ? sc.data\n            : new Float32Array(sc.data.buffer, sc.data.byteOffset, sc.data.byteLength / 4);\n        const groups = sc.shape[0];\n        const scQ = new Float32Array(groups * halfN);\n        const scG = new Float32Array(groups * halfN);\n        for (let g = 0; g < groups; g++) {\n          for (let h = 0; h < numHeads; h++) {\n            const srcOff = g * fusedN + h * headBlock2;\n            const dstOff = g * halfN + h * headDim;\n            for (let d = 0; d < headDim; d++) {\n              scQ[dstOff + d] = scSrc[srcOff + d];\n              scG[dstOff + d] = scSrc[srcOff + headDim + d];\n            }\n          }\n        }\n        await store.set(qProjScales, { data: scQ, shape: [groups, halfN] });\n        await store.set(attnGateScales, { data: scG, shape: [groups, halfN] });\n\n        // Split qzeros [groups, fusedN/8] → two [groups, halfN/8]\n        // qzeros packs 8 output columns per int32. Since headDim (256) is divisible by 8,\n        // the deinterleave works on int32 words directly: headDim/8 words per head.\n        const qProjQzeros = `layers.${i}.self_attn.q_proj.qzeros`;\n        const attnGateQzeros = `layers.${i}.self_attn.attn_gate.qzeros`;\n        const qz = (await store.get(qProjQzeros))!;\n        const qzSrc =\n          qz.data instanceof Int32Array\n            ? qz.data\n            : new Int32Array(qz.data.buffer, qz.data.byteOffset, qz.data.byteLength / 4);\n        const fusedNDiv8 = fusedN >>> 3;\n        const halfNDiv8 = halfN >>> 3;\n        const headBlockDiv8 = headBlock2 >>> 3;\n        const headDimDiv8 = headDim >>> 3;\n        const qzQ = new Int32Array(groups * halfNDiv8);\n        const qzG = new Int32Array(groups * halfNDiv8);\n        for (let g = 0; g < groups; g++) {\n          for (let h = 0; h < numHeads; h++) {\n            const srcOff = g * fusedNDiv8 + h * headBlockDiv8;\n            const dstOff = g * halfNDiv8 + h * headDimDiv8;\n            for (let d = 0; d < headDimDiv8; d++) {\n              qzQ[dstOff + d] = qzSrc[srcOff + d];\n              qzG[dstOff + d] = qzSrc[srcOff + headDimDiv8 + d];\n            }\n          }\n        }\n        await store.set(qProjQzeros, { data: qzQ, shape: [groups, halfNDiv8] });\n        await store.set(attnGateQzeros, { data: qzG, shape: [groups, halfNDiv8] });\n\n        // Clean up g_idx tensors if present (we assume sequential grouping)\n        await store.delete(`layers.${i}.self_attn.q_proj.g_idx`);\n      }\n\n      // MLX: split fused q_proj MLX tensors into q_proj + attn_gate\n      // MLX q_proj.weight is packed [fusedN, K/8], scales/biases are [fusedN, K/gs]\n      // The split is along the N (output/row) dimension — same deinterleave pattern.\n      const mlxQProjWeight = `layers.${i}.self_attn.q_proj.weight`;\n      const mlxAttnGateWeight = `layers.${i}.self_attn.attn_gate.weight`;\n      if (isMLX && store.has(mlxQProjWeight) && !store.has(mlxAttnGateWeight)) {\n        const K_val = textCfg.hidden_size as number; // 1024\n        const fusedN = numHeads * 2 * headDim; // 4096\n        const halfN = numHeads * headDim; // 2048\n        const headBlock2 = 2 * headDim;\n        const mlxGS = (quantConfig?.group_size as number) ?? 64;\n\n        // Deinterleave [fusedN, cols] → two [halfN, cols] with per-head split\n        const deinterleaveMLX = <T extends Float32Array | Uint32Array>(\n          src: T,\n          fusedRows: number,\n          cols: number,\n          Ctor: { new (len: number): T },\n        ): [T, T] => {\n          const halfRows = fusedRows / 2;\n          const qArr = new Ctor(halfRows * cols);\n          const gArr = new Ctor(halfRows * cols);\n          for (let h = 0; h < numHeads; h++) {\n            const srcBase = h * headBlock2 * cols;\n            const dstBase = h * headDim * cols;\n            for (let r = 0; r < headDim; r++) {\n              const s0 = srcBase + r * cols;\n              const s1 = srcBase + (headDim + r) * cols;\n              const d0 = dstBase + r * cols;\n              for (let c = 0; c < cols; c++) {\n                qArr[d0 + c] = src[s0 + c];\n                gArr[d0 + c] = src[s1 + c];\n              }\n            }\n          }\n          return [qArr, gArr];\n        };\n\n        // Split packed weight [fusedN, K/8]\n        const wt = (await store.get(mlxQProjWeight))!;\n        const wtSrc =\n          wt.data instanceof Uint32Array\n            ? wt.data\n            : new Uint32Array(wt.data.buffer, wt.data.byteOffset, wt.data.byteLength / 4);\n        const colsW = K_val >>> 3; // K/8\n        const [wtQ, wtG] = deinterleaveMLX(wtSrc, fusedN, colsW, Uint32Array);\n        await store.set(mlxQProjWeight, { data: wtQ, shape: [halfN, colsW] });\n        await store.set(mlxAttnGateWeight, { data: wtG, shape: [halfN, colsW] });\n\n        // Split scales [fusedN, K/gs]\n        const mlxQProjScales = `layers.${i}.self_attn.q_proj.scales`;\n        const mlxAttnGateScales = `layers.${i}.self_attn.attn_gate.scales`;\n        const sc = (await store.get(mlxQProjScales))!;\n        const scSrc =\n          sc.data instanceof Float32Array\n            ? sc.data\n            : new Float32Array(sc.data.buffer, sc.data.byteOffset, sc.data.byteLength / 4);\n        const colsS = Math.ceil(K_val / mlxGS);\n        const [scQ, scG] = deinterleaveMLX(scSrc, fusedN, colsS, Float32Array);\n        await store.set(mlxQProjScales, { data: scQ, shape: [halfN, colsS] });\n        await store.set(mlxAttnGateScales, { data: scG, shape: [halfN, colsS] });\n\n        // Split biases [fusedN, K/gs]\n        const mlxQProjBiases = `layers.${i}.self_attn.q_proj.biases`;\n        const mlxAttnGateBiases = `layers.${i}.self_attn.attn_gate.biases`;\n        const bi = (await store.get(mlxQProjBiases))!;\n        const biSrc =\n          bi.data instanceof Float32Array\n            ? bi.data\n            : new Float32Array(bi.data.buffer, bi.data.byteOffset, bi.data.byteLength / 4);\n        const [biQ, biG] = deinterleaveMLX(biSrc, fusedN, colsS, Float32Array);\n        await store.set(mlxQProjBiases, { data: biQ, shape: [halfN, colsS] });\n        await store.set(mlxAttnGateBiases, { data: biG, shape: [halfN, colsS] });\n      }\n    }\n  }\n\n  // -- Apply (1 + weight) offset for ALL Qwen3.5 RMSNorm weights --\n  // Qwen3_5RMSNorm initializes weight=0 and applies (1 + weight) * normalized.\n  // ALL norms in the model use this class: input_layernorm, post_attention_layernorm,\n  // final_norm, q_norm, k_norm, and linear_attn.norm. We bake +1 into weights so\n  // the standard RMSNorm kernel (which does weight * normalized) produces correct results.\n  //\n  // EXCEPTION: MLX Community models already absorb the +1 during conversion —\n  // their RMSNorm uses `weight * normalized` (no +1), so the stored weights\n  // are already `1 + original_weight`. Applying +1 again would double-offset.\n  if (architectureName === \"Qwen3_5ForConditionalGeneration\" && !isMLX) {\n    const textCfg = (rawConfig.text_config as Record<string, unknown>) ?? rawConfig;\n    const numLayers = textCfg.num_hidden_layers as number;\n    const layerTypesArr = (textCfg.layer_types as string[]) ?? [];\n    const fullAttnInterval = (textCfg.full_attention_interval as number) ?? 4;\n\n    // Collect all norm weight keys\n    const normKeys: string[] = [];\n    for (let i = 0; i < numLayers; i++) {\n      // All layers have input/post-attention layernorms\n      normKeys.push(`layers.${i}.input_layernorm.weight`);\n      normKeys.push(`layers.${i}.post_attention_layernorm.weight`);\n\n      let isFullAttn: boolean;\n      if (layerTypesArr.length > i) {\n        isFullAttn = layerTypesArr[i] === \"full_attention\";\n      } else {\n        isFullAttn = i % fullAttnInterval === fullAttnInterval - 1;\n      }\n\n      if (isFullAttn) {\n        normKeys.push(`layers.${i}.self_attn.q_norm.weight`);\n        normKeys.push(`layers.${i}.self_attn.k_norm.weight`);\n      }\n      // NOTE: linear_attn.norm.weight uses Qwen3_5RMSNormGated which\n      // initializes weight=1 (not 0), so it does NOT need the +1 offset.\n    }\n    // Final norm\n    normKeys.push(\"norm.weight\");\n\n    for (const normKey of normKeys) {\n      const entry = await store.get(normKey);\n      if (entry) {\n        const data =\n          entry.data instanceof Float32Array\n            ? entry.data\n            : new Float32Array(entry.data.buffer, entry.data.byteOffset, entry.data.byteLength / 4);\n        const adjusted = new Float32Array(data.length);\n        for (let j = 0; j < data.length; j++) {\n          adjusted[j] = 1.0 + data[j];\n        }\n        await store.set(normKey, { data: adjusted, shape: entry.shape });\n      }\n    }\n  }\n\n  // -- Apply (1 + weight) offset for ALL Gemma3 RMSNorm weights --\n  // Gemma3RMSNorm computes `(1.0 + weight) * normalized` (mlx-lm:\n  // mx.fast.rms_norm(x, 1.0 + self.weight, eps)). The stored weight is the raw\n  // trained value centered near 0; we bake +1 so the standard RMSNorm kernel\n  // (weight * normalized) is correct. Unlike Qwen3.5, the MLX convert does NOT\n  // absorb the +1 for Gemma, so this runs for MLX too. Every per-layer norm uses\n  // this class: input_layernorm, post_attention_layernorm, pre_feedforward_\n  // layernorm, post_feedforward_layernorm, q_norm, k_norm, plus the final norm.\n  if (architectureName === \"Gemma3TextModel\" || architectureName === \"Gemma3Model\") {\n    const gemmaNumLayers = rawConfig.num_hidden_layers as number;\n    const gemmaNormKeys: string[] = [];\n    for (let i = 0; i < gemmaNumLayers; i++) {\n      gemmaNormKeys.push(`layers.${i}.input_layernorm.weight`);\n      gemmaNormKeys.push(`layers.${i}.post_attention_layernorm.weight`);\n      gemmaNormKeys.push(`layers.${i}.pre_feedforward_layernorm.weight`);\n      gemmaNormKeys.push(`layers.${i}.post_feedforward_layernorm.weight`);\n      gemmaNormKeys.push(`layers.${i}.self_attn.q_norm.weight`);\n      gemmaNormKeys.push(`layers.${i}.self_attn.k_norm.weight`);\n    }\n    gemmaNormKeys.push(\"norm.weight\");\n\n    for (const normKey of gemmaNormKeys) {\n      const entry = await store.get(normKey);\n      if (entry) {\n        const data =\n          entry.data instanceof Float32Array\n            ? entry.data\n            : new Float32Array(entry.data.buffer, entry.data.byteOffset, entry.data.byteLength / 4);\n        const adjusted = new Float32Array(data.length);\n        for (let j = 0; j < data.length; j++) {\n          adjusted[j] = 1.0 + data[j];\n        }\n        await store.set(normKey, { data: adjusted, shape: entry.shape });\n      }\n    }\n  }\n\n  // NOTE on Gemma 4 RMSNorm: unlike Gemma 3 (weights centered near 0, runtime\n  // `(1+weight)`), the released Gemma 4 / Gemma3n checkpoints store the FULL norm\n  // gain directly (raw input_layernorm values are ~7–68, rms ~12). So NO +1 bake\n  // is applied here — the standard `weight * normalized` kernel is already correct,\n  // and the large residual-stream magnitudes are expected for this model.\n\n  // -- Gemma 4 per-layer output scalar (`layer_scalar`) --\n  // Gemma4TextDecoderLayer ends with `hidden_states *= self.layer_scalar` — a single\n  // learned scalar per layer (shape [1], values ~0.018–0.5, NOT ones). The graph\n  // emits a placeholder Scale node (id `layer{i}_layer_scalar`); we read the scalar\n  // value here and patch the node's `scale` attribute, then drop the weight (it is a\n  // baked constant, not a GPU tensor). Without this the residual stream is unscaled\n  // and generation is incoherent.\n  if (architectureName === \"Gemma4ForConditionalGeneration\") {\n    for (const node of graph.nodes) {\n      const key = node.attributes?.layer_scalar_key as string | undefined;\n      if (!key) continue;\n      const entry = await store.get(key);\n      if (!entry) continue;\n      const data =\n        entry.data instanceof Float32Array\n          ? entry.data\n          : new Float32Array(entry.data.buffer, entry.data.byteOffset, entry.data.byteLength / 4);\n      node.attributes.scale = data[0];\n      await store.delete(key);\n    }\n  }\n\n  // -- Handle tied embeddings --\n  // When safetensorsKey differs from the tensor name (e.g. lm_head.weight\n  // pointing to embed_tokens.weight for tied embeddings), copy the data.\n  for (const [name, desc] of Object.entries(graph.tensors)) {\n    if (\n      desc.storage === \"constant\" &&\n      desc.safetensorsKey &&\n      desc.safetensorsKey !== name &&\n      !store.has(name)\n    ) {\n      const sourceData = await store.get(desc.safetensorsKey);\n      if (sourceData) {\n        await store.set(name, sourceData);\n      }\n    }\n  }\n\n  // -- Repack GPTQ weights into Gerbil INT4 format --\n  if (isGPTQ) {\n    onProgress?.(96, 100, \"Repacking GPTQ weights...\");\n    const gptqGroupSize = (quantConfig?.group_size as number) ?? 128;\n\n    for (const node of graph.nodes) {\n      if (node.opType !== \"MatMulInt4\") continue;\n\n      const K = node.attributes.K as number;\n      const N = node.attributes.N as number;\n\n      const qTensorName = node.inputs[1]; // e.g. layers.0.self_attn.q_proj.weight.q\n      const scalesTensorName = node.inputs[2]; // e.g. layers.0.self_attn.q_proj.weight.scales\n      const zerosTensorName = node.inputs[3]; // e.g. layers.0.self_attn.q_proj.weight.zeros\n\n      // Derive GPTQ tensor names from graph names:\n      // Graph: \"layers.0.self_attn.q_proj.weight.q\" → strip \".q\" → strip \".weight\" → base\n      const sourceName = qTensorName.slice(0, -2); // \"layers.0.self_attn.q_proj.weight\"\n      const gptqBase = sourceName.slice(0, -7); // \"layers.0.self_attn.q_proj\"\n\n      const gptqQweight = await store.get(`${gptqBase}.qweight`);\n      const gptqScales = await store.get(`${gptqBase}.scales`);\n      const gptqQzeros = await store.get(`${gptqBase}.qzeros`);\n\n      // Skip layers not quantized in GPTQ — they'll be handled by on-the-fly quantization below\n      if (!gptqQweight || !gptqScales || !gptqQzeros) continue;\n\n      // Ensure correct typed arrays for the adapter\n      const qweightData =\n        gptqQweight.data instanceof Int32Array\n          ? gptqQweight.data\n          : new Int32Array(\n              gptqQweight.data.buffer,\n              gptqQweight.data.byteOffset,\n              gptqQweight.data.byteLength / 4,\n            );\n      const scalesData =\n        gptqScales.data instanceof Float32Array\n          ? gptqScales.data\n          : new Float32Array(\n              gptqScales.data.buffer,\n              gptqScales.data.byteOffset,\n              gptqScales.data.byteLength / 4,\n            );\n      const qzerosData =\n        gptqQzeros.data instanceof Int32Array\n          ? gptqQzeros.data\n          : new Int32Array(\n              gptqQzeros.data.buffer,\n              gptqQzeros.data.byteOffset,\n              gptqQzeros.data.byteLength / 4,\n            );\n\n      const repacked = repackGPTQ({\n        qweight: qweightData,\n        scales: scalesData,\n        qzeros: qzerosData,\n        K,\n        N,\n        groupSize: gptqGroupSize,\n      });\n\n      await store.set(qTensorName, { data: repacked.packed, shape: [repacked.packed.length] });\n      await store.set(scalesTensorName, { data: repacked.scales, shape: [repacked.scales.length] });\n      await store.set(zerosTensorName, { data: repacked.zeros, shape: [repacked.zeros.length] });\n\n      // Clean up GPTQ source tensors to free memory\n      await store.delete(`${gptqBase}.qweight`);\n      await store.delete(`${gptqBase}.scales`);\n      await store.delete(`${gptqBase}.qzeros`);\n    }\n  }\n\n  // -- Repack MLX 4-bit weights into Gerbil INT4 format --\n  if (isMLX) {\n    onProgress?.(96, 100, \"Repacking MLX weights...\");\n    const mlxGS = (quantConfig?.group_size as number) ?? 64;\n\n    for (const node of graph.nodes) {\n      if (node.opType !== \"MatMulInt4\" && node.opType !== \"EmbeddingInt4\") continue;\n\n      // Derive N and K for this layer\n      let N: number, K: number;\n      if (node.opType === \"EmbeddingInt4\") {\n        N = node.attributes.vocab_size as number;\n        K = node.attributes.hidden_size as number;\n      } else {\n        N = node.attributes.N as number;\n        K = node.attributes.K as number;\n      }\n\n      const qTensorName = node.inputs[1]; // e.g. layers.0.self_attn.q_proj.weight.q\n      const scalesTensorName = node.inputs[2]; // e.g. layers.0.self_attn.q_proj.weight.scales\n      const zerosTensorName = node.inputs[3]; // e.g. layers.0.self_attn.q_proj.weight.zeros\n\n      // Derive MLX source tensor names:\n      // Graph: \"base.weight.q\" → strip \".q\" → \"base.weight\" (MLX packed weight)\n      // MLX scales: \"base.scales\", MLX biases: \"base.biases\"\n      const sourceName = qTensorName.slice(0, -2); // \"base.weight\"\n      const baseName = sourceName.slice(0, -7); // \"base\" (strip \".weight\")\n\n      const mlxWeight = await store.get(sourceName); // packed U32\n      const mlxScales = await store.get(`${baseName}.scales`); // F32 (already converted from BF16)\n      const mlxBiases = await store.get(`${baseName}.biases`); // F32 (already converted from BF16)\n\n      if (!mlxWeight || !mlxScales || !mlxBiases) continue;\n\n      const weightData =\n        mlxWeight.data instanceof Uint32Array\n          ? mlxWeight.data\n          : new Uint32Array(\n              mlxWeight.data.buffer,\n              mlxWeight.data.byteOffset,\n              mlxWeight.data.byteLength / 4,\n            );\n      const scalesData =\n        mlxScales.data instanceof Float32Array\n          ? mlxScales.data\n          : new Float32Array(\n              mlxScales.data.buffer,\n              mlxScales.data.byteOffset,\n              mlxScales.data.byteLength / 4,\n            );\n      const biasesData =\n        mlxBiases.data instanceof Float32Array\n          ? mlxBiases.data\n          : new Float32Array(\n              mlxBiases.data.buffer,\n              mlxBiases.data.byteOffset,\n              mlxBiases.data.byteLength / 4,\n            );\n\n      const repacked = repackMLX({\n        weight: weightData,\n        scales: scalesData,\n        biases: biasesData,\n        N,\n        K,\n        groupSize: mlxGS,\n      });\n\n      await store.set(qTensorName, { data: repacked.packed, shape: [repacked.packed.length] });\n      await store.set(scalesTensorName, { data: repacked.scales, shape: [repacked.scales.length] });\n      await store.set(zerosTensorName, { data: repacked.zeros, shape: [repacked.zeros.length] });\n\n      // Clean up MLX source tensors to free memory\n      await store.delete(sourceName);\n      await store.delete(`${baseName}.scales`);\n      await store.delete(`${baseName}.biases`);\n    }\n  }\n\n  // -- Quantize weights for INT4 if requested (or for non-GPTQ/MLX layers) --\n  if (resolvedDtype === \"q4\") {\n    onProgress?.(96, 100, \"Quantizing weights to INT4...\");\n    for (const node of graph.nodes) {\n      if (node.opType !== \"MatMulInt4\" && node.opType !== \"EmbeddingInt4\") continue;\n\n      const groupSz = (node.attributes.group_size as number) ?? DEFAULT_GROUP_SIZE;\n\n      const qTensorName = node.inputs[1]; // weight.q\n      const scalesTensorName = node.inputs[2]; // weight.scales\n      const zerosTensorName = node.inputs[3]; // weight.zeros\n\n      // Skip nodes already handled by GPTQ or MLX repacking\n      if (store.has(qTensorName)) continue;\n\n      // Derive source weight name by stripping \".q\" suffix\n      const sourceName = qTensorName.slice(0, -2);\n      let sourceWeight = await store.get(sourceName);\n      // Tied embeddings fallback: lm_head.weight may not exist in safetensors\n      // when tie_word_embeddings=true — use embed_tokens.weight instead.\n      if (!sourceWeight && sourceName === \"lm_head.weight\") {\n        sourceWeight = await store.get(\"embed_tokens.weight\");\n      }\n\n      if (sourceWeight) {\n        const f32Data =\n          sourceWeight.data instanceof Float32Array\n            ? sourceWeight.data\n            : new Float32Array(\n                sourceWeight.data.buffer,\n                sourceWeight.data.byteOffset,\n                sourceWeight.data.byteLength / 4,\n              );\n\n        const { packed, scales, zeros } = quantizeInt4(f32Data, groupSz);\n\n        await store.set(qTensorName, { data: packed, shape: [packed.length] });\n        await store.set(scalesTensorName, { data: scales, shape: [scales.length] });\n        await store.set(zerosTensorName, { data: zeros, shape: [zeros.length] });\n\n        // Free the original F32 weight to save memory\n        await store.delete(sourceName);\n      }\n    }\n  }\n\n  // -- MLX vision patch_embed axis fixup --\n  // The ViT patch_embed conv weight is a 5D tensor whose flattening must match\n  // the host patch vector's [C, T, ph, pw] order. The HF BF16 checkpoint stores\n  // it channels-first as [out, C, T, ph, pw] → flattens correctly as-is. The\n  // mlx-vlm convert stores it channels-LAST as [out, T, ph, pw, C], so the\n  // flat [out, 1536] inner order is permuted and the first matmul would be fed\n  // mismatched weights (manifests as low cosine, not NaN). Detect the\n  // channels-last 5D shape and permute back to [out, C, T, ph, pw] so the\n  // flattened weight matches the host patch layout on both checkpoints.\n  if (isMLX) {\n    const pe = await store.get(CANONICAL_KEYS.visPatchEmbedWeight);\n    const s = pe?.shape;\n    // channels-last conv weight: [out, T, ph, pw, C] with C === in_channels (3)\n    if (pe && s && s.length === 5 && s[4] === 3) {\n      const [outC, T, ph, pw, C] = s;\n      const src =\n        pe.data instanceof Float32Array\n          ? pe.data\n          : new Float32Array(pe.data.buffer, pe.data.byteOffset, pe.data.byteLength / 4);\n      const dst = new Float32Array(src.length);\n      // src index (channels-last):  ((((o*T + t)*ph + y)*pw + x)*C + c)\n      // dst index (channels-first): ((((o*C + c)*T + t)*ph + y)*pw + x)\n      for (let o = 0; o < outC; o++) {\n        for (let t = 0; t < T; t++) {\n          for (let y = 0; y < ph; y++) {\n            for (let x = 0; x < pw; x++) {\n              for (let c = 0; c < C; c++) {\n                const si = (((o * T + t) * ph + y) * pw + x) * C + c;\n                const di = (((o * C + c) * T + t) * ph + y) * pw + x;\n                dst[di] = src[si];\n              }\n            }\n          }\n        }\n      }\n      // Store flattened [out, C*T*ph*pw] = [out, 1536] in channels-first order.\n      await store.set(CANONICAL_KEYS.visPatchEmbedWeight, {\n        data: dst,\n        shape: [outC, C * T * ph * pw],\n      });\n    }\n  }\n\n  // -- Gemma 4: extract the PLE table to CPU-resident storage (NOT GPU) --\n  // The per-layer embedding table `embed_tokens_per_layer` is ~1.17GB at 4-bit.\n  // We keep it in JS memory and stream per-token rows at inference, so it never\n  // becomes a GPU buffer. Repack MLX 4-bit → Gerbil INT4 here; for an unquantized\n  // checkpoint, encode the f32 rows with a unit scale (zero=0) so the executor's\n  // single dequant path handles both. The PLE EmbeddingInt4 node was intentionally\n  // omitted from the graph (gemma4.ts), so the standard repack loop skips this\n  // table — it is the loader's job to divert it out of `weights`.\n  let pleSource: PleSource | undefined;\n  const isGemma4 = String(architectureName).startsWith(\"Gemma4\");\n  if (isGemma4) {\n    const pleBase = CANONICAL_KEYS.GEMMA4_PLE_EMBED.slice(0, -7); // \"embed_tokens_per_layer\"\n    const textCfg = (rawConfig.text_config as Record<string, unknown>) ?? rawConfig;\n    const pleDim = (textCfg.hidden_size_per_layer_input as number) ?? 256;\n    const numLayers = textCfg.num_hidden_layers as number;\n    const pleWidth = numLayers * pleDim; // L*256\n    const pleVocab =\n      (textCfg.vocab_size_per_layer_input as number) ?? (textCfg.vocab_size as number);\n\n    const packedEntry = await store.get(`${pleBase}.weight`);\n    const scalesEntry = await store.get(`${pleBase}.scales`);\n    const biasesEntry = await store.get(`${pleBase}.biases`);\n\n    if (isMLX && packedEntry && scalesEntry && biasesEntry) {\n      const mlxGS = (quantConfig?.group_size as number) ?? 64;\n      const packedU32 =\n        packedEntry.data instanceof Uint32Array\n          ? packedEntry.data\n          : new Uint32Array(\n              packedEntry.data.buffer,\n              packedEntry.data.byteOffset,\n              packedEntry.data.byteLength / 4,\n            );\n      const scalesF32 =\n        scalesEntry.data instanceof Float32Array\n          ? scalesEntry.data\n          : new Float32Array(\n              scalesEntry.data.buffer,\n              scalesEntry.data.byteOffset,\n              scalesEntry.data.byteLength / 4,\n            );\n      const biasesF32 =\n        biasesEntry.data instanceof Float32Array\n          ? biasesEntry.data\n          : new Float32Array(\n              biasesEntry.data.buffer,\n              biasesEntry.data.byteOffset,\n              biasesEntry.data.byteLength / 4,\n            );\n      const repacked = repackMLX({\n        weight: packedU32,\n        scales: scalesF32,\n        biases: biasesF32,\n        N: pleVocab,\n        K: pleWidth,\n        groupSize: mlxGS,\n      });\n      pleSource = await buildPleSource(\n        repacked.packed,\n        repacked.scales,\n        repacked.zeros,\n        pleWidth,\n        mlxGS,\n        cacheStore,\n      );\n      // Divert out of the store so uploadWeights() never sends it to the GPU.\n      await store.delete(`${pleBase}.weight`);\n      await store.delete(`${pleBase}.scales`);\n      await store.delete(`${pleBase}.biases`);\n      console.log(\n        `[Gerbil] Gemma4 PLE table kept ${cacheStore ? \"cache-resident\" : \"CPU-resident\"}: ` +\n          `[${pleVocab}, ${pleWidth}] int4 ` +\n          `(~${((repacked.packed.byteLength + repacked.scales.byteLength + repacked.zeros.byteLength) / 1e6).toFixed(0)} MB, 0 MB GPU)`,\n      );\n    } else if (packedEntry?.data instanceof Float32Array) {\n      // Unquantized f32 table: quantize to INT4 for compact storage and a\n      // single shared dequant path. (Quantization-sensitive precision concerns\n      // only apply when forced GPU-resident; here we still dequantize per-row.)\n      const { packed, scales, zeros } = quantizeInt4(packedEntry.data, DEFAULT_GROUP_SIZE);\n      pleSource = await buildPleSource(\n        packed,\n        scales,\n        zeros,\n        pleWidth,\n        DEFAULT_GROUP_SIZE,\n        cacheStore,\n      );\n      await store.delete(`${pleBase}.weight`);\n    } else {\n      console.warn(\n        \"[Gerbil] Gemma4: PLE table not found in weights — per-layer embeddings \" +\n          \"will be zero. Check the checkpoint ships embed_tokens_per_layer.\",\n      );\n    }\n  }\n\n  // -- Materialize synthetic fill-value constants (e.g. Gemma 4 v_norm ones) --\n  // Tensors declared with `fillValue` carry no checkpoint weight; the loader fills\n  // them with a constant at the declared shape. Used for parameter-free norms.\n  for (const [name, desc] of Object.entries(graph.tensors)) {\n    if (desc.storage === \"constant\" && desc.fillValue !== undefined && !store.has(name)) {\n      const length = desc.shape.reduce(\n        (acc: number, dim) => acc * (typeof dim === \"number\" ? dim : 1),\n        1,\n      );\n      const data = new Float32Array(length).fill(desc.fillValue);\n      await store.set(name, {\n        data,\n        shape: desc.shape.map((d) => (typeof d === \"number\" ? d : 1)),\n      });\n    }\n  }\n\n  // -- Verify required tensors --\n  const missingTensors: string[] = [];\n  for (const [name, desc] of Object.entries(graph.tensors)) {\n    if (desc.storage === \"constant\" && !store.has(name)) {\n      missingTensors.push(name);\n    }\n  }\n\n  if (missingTensors.length > 0) {\n    console.warn(\n      `[Gerbil] Missing ${missingTensors.length} weight tensors:`,\n      missingTensors.slice(0, 10),\n    );\n  }\n\n  onProgress?.(100, 100, \"Model loaded.\");\n\n  return { graph, tokenizer, weights: store, rawConfig, pleSource };\n}\n\n// ── Moonshine STT loader ───────────────────────────────────────────────────\n//\n// Moonshine is an encoder-decoder ASR model that cannot be expressed as a single\n// causal-LM graph (the standard loadModel path). It needs two graphs (a conv\n// encoder run once, an AR decoder with cross-attention) sharing one set of f32\n// weights. This loader downloads config + tokenizer + the single model.safetensors\n// and returns the canonical-named f32 weights for both graphs. Keys are mapped by\n// stripping the \"model.\" prefix the checkpoint uses (e.g. \"model.encoder.conv1.\n// weight\" → \"encoder.conv1.weight\"), which is exactly the safetensorsKey the\n// Moonshine graph generators reference.\n\nexport interface LoadedMoonshine {\n  /** Canonical-named f32 weights (data + shape), shared by encoder + decoder graphs. */\n  weights: Map<string, { data: ArrayBufferView; shape: number[] }>;\n  /** The (decode-capable) tokenizer. */\n  tokenizer: Tokenizer;\n  /** Raw config.json. */\n  rawConfig: Record<string, unknown>;\n}\n\nexport async function loadMoonshine(options: {\n  repo: string;\n  revision?: string;\n  hfToken?: string;\n  cacheDir?: string;\n  onProgress?: (loaded: number, total: number, message: string) => void;\n}): Promise<LoadedMoonshine> {\n  const revision = options.revision ?? \"main\";\n  const baseURL = resolveHFBaseURL(options.repo, revision);\n  await initFs();\n\n  options.onProgress?.(0, 100, \"Fetching config + tokenizer…\");\n  const [rawConfig, tokenizerJSON] = await Promise.all([\n    fetchJSON(baseURL, \"config.json\", options.hfToken, options.cacheDir),\n    fetchJSON(baseURL, \"tokenizer.json\", options.hfToken, options.cacheDir),\n  ]);\n  const tokenizerConfig = await fetchJSON(\n    baseURL,\n    \"tokenizer_config.json\",\n    options.hfToken,\n    options.cacheDir,\n  ).catch(() => null);\n  const tokenizer = Tokenizer.fromJSON(tokenizerJSON, tokenizerConfig);\n\n  options.onProgress?.(10, 100, \"Downloading weights…\");\n  const buf = await fetchBinary(\n    baseURL,\n    \"model.safetensors\",\n    options.hfToken,\n    (loaded, total) => options.onProgress?.(10 + (loaded / (total || 1)) * 85, 100, \"weights\"),\n    options.cacheDir,\n  );\n\n  const file = parseSafetensorsHeader(buf);\n  const weights = new Map<string, { data: ArrayBufferView; shape: number[] }>();\n  for (const entry of file.entries) {\n    // Strip the \"model.\" prefix the Moonshine checkpoint uses on every tensor.\n    const canonical = entry.name.startsWith(\"model.\") ? entry.name.slice(6) : entry.name;\n    let data: ArrayBufferView = getTensorData(buf, file, entry);\n    if (entry.dtype === \"BF16\") {\n      data = bf16ToF32(new Uint8Array(data.buffer, data.byteOffset, data.byteLength));\n    } else if (entry.dtype === \"F16\") {\n      data = f16ToF32(data);\n    }\n    weights.set(canonical, { data, shape: entry.shape });\n  }\n\n  options.onProgress?.(100, 100, \"Moonshine loaded.\");\n  return { weights, tokenizer, rawConfig };\n}\n\n// ════════════════════════════════════════════════════════════════════════════\n// Kani-TTS-2 dual-checkpoint loader (LFM2-350M codec-LM + NVIDIA NeMo NanoCodec).\n//\n// Kani-TTS-2 is a two-stage TTS model that, like Moonshine, cannot be one graph:\n//   1. the codec-LM backbone (LFM2-350M body) lives in the main repo\n//      (nineninesix/kani-tts-2-en, model.safetensors, \"model.\" prefix), and\n//   2. the NanoCodec decoder weights live in a SEPARATE repo\n//      (nineninesix/nemo-nano-codec-22khz-0.6kbps-12.5fps-MLX, model.safetensors)\n//      stored as weight-norm parametrizations under verbose NeMo key paths.\n//\n// This loader downloads both, folds the codec's weight-norm (W = g·v/‖v‖) and\n// renames its keys to the canonical `nanocodec.*` names the decoder graph expects,\n// and returns one merged f32 weight map (+ tokenizer + config) for the KaniTTS\n// consumer (src/gpu/kani-tts.ts), which builds the backbone + decoder graphs.\n// ════════════════════════════════════════════════════════════════════════════\n\n/** Default NanoCodec checkpoint Kani-TTS-2 uses (NeMo 22 kHz, 0.6 kbps, 12.5 fps). */\nexport const KANI_NANOCODEC_REPO = \"nineninesix/nemo-nano-codec-22khz-0.6kbps-12.5fps-MLX\";\n\nexport interface LoadedKaniTTS {\n  /** Canonical-named f32 weights for the codec-LM backbone (LFM2 keys). */\n  backboneWeights: Map<string, { data: ArrayBufferView; shape: number[] }>;\n  /** Folded NanoCodec decoder weights under canonical `nanocodec.*` names. */\n  codecWeights: Map<string, { data: ArrayBufferView; shape: number[] }>;\n  /** The text tokenizer. */\n  tokenizer: Tokenizer;\n  /** Raw backbone config.json (LFM2 dims + KaniTTS2 fields). */\n  rawConfig: Record<string, unknown>;\n}\n\n/** Matches the LFM2 attention out_proj key (renamed to o_proj for the canonical map). */\nconst KANI_OUT_PROJ_RE = /\\.self_attn\\.out_proj\\./;\n\n/** Coerce a safetensors tensor view (bf16/f16→f32) to a contiguous Float32Array. */\nfunction toF32(entry: SafetensorEntry, buf: ArrayBuffer, file: SafetensorsFile): Float32Array {\n  let data: ArrayBufferView = getTensorData(buf, file, entry);\n  if (entry.dtype === \"BF16\") {\n    data = bf16ToF32(new Uint8Array(data.buffer, data.byteOffset, data.byteLength));\n  } else if (entry.dtype === \"F16\") {\n    data = f16ToF32(data);\n  }\n  return data instanceof Float32Array\n    ? data\n    : new Float32Array(data.buffer, data.byteOffset, data.byteLength / 4);\n}\n\n/** Fold + rename the raw NanoCodec tensors into the canonical `nanocodec.*` map. */\nfunction buildKaniCodecWeights(\n  codecRaw: Map<string, { data: Float32Array; shape: number[] }>,\n): Map<string, { data: ArrayBufferView; shape: number[] }> {\n  const out = new Map<string, { data: ArrayBufferView; shape: number[] }>();\n  const { convs, plains } = nanoCodecWeightMap();\n  for (const c of convs) {\n    const g = codecRaw.get(c.gKey);\n    const v = codecRaw.get(c.vKey);\n    if (!(g && v)) {\n      throw new Error(\n        `NanoCodec weight-norm pair missing for ${c.graphKey}: ${c.gKey} / ${c.vKey}. ` +\n          \"Check the codec checkpoint (expected weight-norm parametrizations).\",\n      );\n    }\n    out.set(c.graphKey, {\n      data: foldNanoCodecWeightNorm(g.data, v.data, v.shape),\n      shape: v.shape,\n    });\n  }\n  for (const pl of plains) {\n    const t = codecRaw.get(pl.mlxKey);\n    if (!t) {\n      throw new Error(`NanoCodec tensor missing: ${pl.mlxKey} (for ${pl.graphKey}).`);\n    }\n    out.set(pl.graphKey, { data: t.data, shape: t.shape });\n  }\n  return out;\n}\n\nexport async function loadKaniTTS(options: {\n  /** Backbone repo (default nineninesix/kani-tts-2-en). */\n  repo?: string;\n  /** NanoCodec repo (default KANI_NANOCODEC_REPO). */\n  codecRepo?: string;\n  revision?: string;\n  hfToken?: string;\n  cacheDir?: string;\n  onProgress?: (loaded: number, total: number, message: string) => void;\n}): Promise<LoadedKaniTTS> {\n  const revision = options.revision ?? \"main\";\n  const repo = options.repo ?? \"nineninesix/kani-tts-2-en\";\n  const codecRepo = options.codecRepo ?? KANI_NANOCODEC_REPO;\n  const baseURL = resolveHFBaseURL(repo, revision);\n  const codecURL = resolveHFBaseURL(codecRepo, \"main\");\n  await initFs();\n\n  options.onProgress?.(0, 100, \"Fetching config + tokenizer…\");\n  const [rawConfig, tokenizerJSON] = await Promise.all([\n    fetchJSON(baseURL, \"config.json\", options.hfToken, options.cacheDir),\n    fetchJSON(baseURL, \"tokenizer.json\", options.hfToken, options.cacheDir),\n  ]);\n  const tokenizerConfig = await fetchJSON(\n    baseURL,\n    \"tokenizer_config.json\",\n    options.hfToken,\n    options.cacheDir,\n  ).catch(() => null);\n  const tokenizer = Tokenizer.fromJSON(tokenizerJSON, tokenizerConfig);\n\n  // ── Backbone weights (LFM2-350M): strip \"model.\" and rename out_proj→o_proj. ──\n  options.onProgress?.(8, 100, \"Downloading codec-LM backbone…\");\n  const backboneBuf = await fetchBinary(\n    baseURL,\n    \"model.safetensors\",\n    options.hfToken,\n    (loaded, total) =>\n      options.onProgress?.(8 + (loaded / (total || 1)) * 45, 100, \"backbone weights\"),\n    options.cacheDir,\n  );\n  const backboneFile = parseSafetensorsHeader(backboneBuf);\n  const backboneWeights = new Map<string, { data: ArrayBufferView; shape: number[] }>();\n  for (const entry of backboneFile.entries) {\n    const key = (entry.name.startsWith(\"model.\") ? entry.name.slice(6) : entry.name).replace(\n      KANI_OUT_PROJ_RE,\n      \".self_attn.o_proj.\",\n    );\n    backboneWeights.set(key, {\n      data: toF32(entry, backboneBuf, backboneFile),\n      shape: entry.shape,\n    });\n  }\n\n  // ── NanoCodec weights: fold weight-norm + rename to canonical nanocodec.* ──\n  options.onProgress?.(55, 100, \"Downloading NanoCodec decoder…\");\n  // The codec's model.safetensors shares its name with the backbone's, and\n  // fetchBinary caches by filename only — give the codec its own cache subdir so\n  // the two downloads do not collide.\n  const codecCacheDir = options.cacheDir ? `${options.cacheDir}/nanocodec` : undefined;\n  const codecBuf = await fetchBinary(\n    codecURL,\n    \"model.safetensors\",\n    options.hfToken,\n    (loaded, total) =>\n      options.onProgress?.(55 + (loaded / (total || 1)) * 40, 100, \"codec weights\"),\n    codecCacheDir,\n  );\n  const codecFile = parseSafetensorsHeader(codecBuf);\n  const codecRaw = new Map<string, { data: Float32Array; shape: number[] }>();\n  for (const entry of codecFile.entries) {\n    codecRaw.set(entry.name, { data: toF32(entry, codecBuf, codecFile), shape: entry.shape });\n  }\n  const codecWeights = buildKaniCodecWeights(codecRaw);\n\n  options.onProgress?.(100, 100, \"Kani-TTS-2 loaded.\");\n  return { backboneWeights, codecWeights, tokenizer, rawConfig };\n}\n","/**\n * MoonshineEncoderExecutor — runs the Moonshine STT encoder graph on WebGPU.\n *\n * Modeled on VisionExecutor: a single non-autoregressive prefill (conv frontend +\n * bidirectional transformer) run ONCE per utterance. The graph is length-static\n * (the Conv1dFull ops carry concrete L/Lout), so it is regenerated per utterance\n * from the input sample count (moonshineEncoderFrames) and a fresh executor is\n * built for that length.\n *\n * Beyond the encoder output, the graph also projects the frozen encoder hidden\n * state through each DECODER layer's encoder_attn.k_proj / v_proj, producing\n * enc_k_layer{i} / enc_v_layer{i}. The host reads these back and binds them frozen\n * into the decoder's CrossAttention across every AR decode step.\n *\n * WebKit discipline mirrors VisionExecutor: per-dispatch submit + drain, and the\n * bidirectional Attention op is windowed by query-row so no single dispatch runs\n * long enough to trip Metal's GPU watchdog. Dawn uses one submission.\n */\n\nimport type { GPUContext } from \"./device.js\";\nimport {\n  createBindGroup,\n  createStorageBuffer,\n  createUniformBuffer,\n  destroyBuffers,\n  getOrCreatePipeline,\n} from \"./device.js\";\nimport { DTYPE_BYTES, type ModelGraph, type OpNode } from \"./ir.js\";\nimport {\n  KERNEL_REGISTRY,\n  type KernelSpec,\n  LAYERNORM_NO_BIAS_SPEC,\n  ROPE_INTERLEAVED_SPEC,\n} from \"./kernels/registry.js\";\n\nconst MAP_MODE_READ = 0x0001;\n\n/**\n * WebKit only: window the bidirectional encoder Attention grid into chunks of this\n * many query rows, each its own submit + drain (same watchdog mitigation the vision\n * tower uses). Encoder frames are far fewer than ViT patches, but a long clip can\n * still produce a few hundred frames, so we keep the guard.\n */\nconst WEBKIT_ATTN_CHUNK_ROWS = 64;\n\ninterface MoonshineDispatch {\n  node: OpNode;\n  spec: KernelSpec;\n  pipeline: GPUComputePipeline;\n  bindGroup: GPUBindGroup;\n  uniformBuffer: GPUBuffer;\n}\n\nexport interface EncoderResult {\n  /** encoder_out [S_enc, hidden] (debug / parity). */\n  encoderOut: Float32Array;\n  /** Per-decoder-layer frozen K, indexed by layer. */\n  encK: Float32Array[];\n  /** Per-decoder-layer frozen V, indexed by layer. */\n  encV: Float32Array[];\n  /** Encoder length (frames). */\n  sEnc: number;\n}\n\nexport class MoonshineEncoderExecutor {\n  private ctx: GPUContext;\n  private graph: ModelGraph;\n  private decLayers: number;\n  private hidden: number;\n\n  private weightBuffers = new Map<string, GPUBuffer>();\n  private activationBuffers = new Map<string, GPUBuffer>();\n  private dispatches: MoonshineDispatch[] = [];\n\n  constructor(ctx: GPUContext, graph: ModelGraph, decLayers: number) {\n    this.ctx = ctx;\n    this.graph = graph;\n    this.decLayers = decLayers;\n    this.hidden = graph.config.hidden_size;\n    this.allocateActivationBuffers();\n  }\n\n  /**\n   * Upload the encoder constants. `weights` holds canonical-named f32 tensors;\n   * only those referenced by the graph are uploaded (the decoder weights are\n   * uploaded into the decoder Executor separately).\n   */\n  uploadWeights(weights: Map<string, { data: ArrayBufferView; shape: number[] }>): void {\n    for (const [name, desc] of Object.entries(this.graph.tensors)) {\n      if (desc.storage !== \"constant\") continue;\n      const w = weights.get(name);\n      if (w) {\n        const buffer = createStorageBuffer(this.ctx, `menc_${name}`, w.data.byteLength, w.data);\n        this.weightBuffers.set(name, buffer);\n      } else if (!desc.safetensorsKey) {\n        // Synthetic constant with no checkpoint source (e.g. conv1's zero bias —\n        // Conv1dFull requires a bias binding but Moonshine's conv1 has none).\n        // Allocate it zero-filled.\n        const count = (desc.shape as number[]).reduce((a, b) => a * (b as number), 1);\n        const zeros = new Float32Array(count);\n        const buffer = createStorageBuffer(this.ctx, `menc_${name}`, zeros.byteLength, zeros);\n        this.weightBuffers.set(name, buffer);\n      } else {\n        throw new Error(`MoonshineEncoderExecutor: missing weight \"${name}\"`);\n      }\n    }\n  }\n\n  initBindGroups(): void {\n    const shapes = this.resolveShapes();\n    for (const nodeId of this.graph.executionOrder) {\n      const node = this.graph.nodes.find((n) => n.id === nodeId)!;\n      let spec = KERNEL_REGISTRY[node.opType];\n      if (!spec) throw new Error(`MoonshineEncoderExecutor: no kernel for op \"${node.opType}\"`);\n      // Moonshine norms are weight-only — pick the no-bias LayerNorm variant.\n      if (node.opType === \"LayerNorm\" && node.inputs.length === 2) {\n        spec = LAYERNORM_NO_BIAS_SPEC;\n      } else if (node.opType === \"RoPE\" && node.attributes.interleaved === true) {\n        spec = ROPE_INTERLEAVED_SPEC;\n      }\n      const pipeline = getOrCreatePipeline(\n        this.ctx,\n        `mkernel_${nodeId}`,\n        spec.shaderCode,\n        spec.entryPoint,\n      );\n      const paramsData = spec.buildParams(node, shapes, { seqPos: 0 });\n      const uniformBuffer = createUniformBuffer(this.ctx, `muniform_${nodeId}`, paramsData);\n      const bufferEntries = this.gatherBuffers(spec, node, uniformBuffer);\n      const bindGroup = createBindGroup(this.ctx, pipeline, bufferEntries, `mbg_${nodeId}`);\n      this.dispatches.push({ node, spec, pipeline, bindGroup, uniformBuffer });\n    }\n  }\n\n  /** Run the conv frontend + encoder + K/V projection. `pcm` is raw 16kHz mono. */\n  async encode(pcm: Float32Array): Promise<EncoderResult> {\n    const q = this.ctx.device.queue;\n    q.writeBuffer(this.activationBuffers.get(\"pcm\")!, 0, pcm as BufferSource);\n\n    const shapes = this.resolveShapes();\n    const ctxRt = { seqPos: 0 };\n    const sizes: Array<[number, number, number]> = [];\n    for (const d of this.dispatches) {\n      const params = d.spec.buildParams(d.node, shapes, ctxRt);\n      q.writeBuffer(d.uniformBuffer, 0, params);\n      sizes.push(d.spec.getDispatchSize(d.node, shapes, ctxRt));\n    }\n\n    if (this.ctx.isWebKitWebGPU) {\n      for (let i = 0; i < this.dispatches.length; i++) {\n        const d = this.dispatches[i];\n        if (d.node.opType === \"Attention\") {\n          // Window the O(S²) bidirectional attention by query-row (watchdog guard).\n          const [gridX, gridY, gridZ] = sizes[i];\n          for (let start = 0; start < gridX; start += WEBKIT_ATTN_CHUNK_ROWS) {\n            const rows = Math.min(WEBKIT_ATTN_CHUNK_ROWS, gridX - start);\n            const params = d.spec.buildParams(d.node, shapes, { seqPos: 0, qOffset: start });\n            q.writeBuffer(d.uniformBuffer, 0, params);\n            const enc = this.ctx.device.createCommandEncoder();\n            const p = enc.beginComputePass();\n            p.setPipeline(d.pipeline);\n            p.setBindGroup(0, d.bindGroup);\n            p.dispatchWorkgroups(rows, gridY, gridZ);\n            p.end();\n            q.submit([enc.finish()]);\n            await q.onSubmittedWorkDone();\n          }\n          continue;\n        }\n        const enc = this.ctx.device.createCommandEncoder();\n        const p = enc.beginComputePass();\n        p.setPipeline(d.pipeline);\n        p.setBindGroup(0, d.bindGroup);\n        p.dispatchWorkgroups(...sizes[i]);\n        p.end();\n        q.submit([enc.finish()]);\n        await q.onSubmittedWorkDone();\n      }\n    } else {\n      const enc = this.ctx.device.createCommandEncoder({ label: \"moonshine_encode\" });\n      const pass = enc.beginComputePass({ label: \"moonshine_enc_pass\" });\n      for (let i = 0; i < this.dispatches.length; i++) {\n        pass.setPipeline(this.dispatches[i].pipeline);\n        pass.setBindGroup(0, this.dispatches[i].bindGroup);\n        pass.dispatchWorkgroups(...sizes[i]);\n      }\n      pass.end();\n      q.submit([enc.finish()]);\n    }\n\n    const sEnc = this.encoderFrames();\n    const encoderOut = await this.readBack(\"encoder_out\", sEnc * this.hidden);\n    const encK: Float32Array[] = [];\n    const encV: Float32Array[] = [];\n    for (let i = 0; i < this.decLayers; i++) {\n      encK.push(await this.readBack(`enc_k_layer${i}`, sEnc * this.hidden));\n      encV.push(await this.readBack(`enc_v_layer${i}`, sEnc * this.hidden));\n    }\n    return { encoderOut, encK, encV, sEnc };\n  }\n\n  /** Read back a named activation buffer after encode() (debug / parity checks). */\n  async readActivation(name: string, maxElements = 4096): Promise<Float32Array> {\n    return this.readBack(name, maxElements);\n  }\n\n  destroy(): void {\n    destroyBuffers([...this.weightBuffers.values()]);\n    destroyBuffers([...this.activationBuffers.values()]);\n    for (const d of this.dispatches) d.uniformBuffer.destroy();\n    this.weightBuffers.clear();\n    this.activationBuffers.clear();\n    this.dispatches = [];\n  }\n\n  // ── Private ──\n\n  /** Encoder frame count = first dim of encoder_out (resolved, numeric). */\n  private encoderFrames(): number {\n    return this.graph.tensors.encoder_out.shape[0] as number;\n  }\n\n  private resolveShapes(): Record<string, number[]> {\n    // The Moonshine encoder graph is fully concrete (no symbolic dims) — every\n    // tensor shape is numeric for the chosen sample count.\n    const resolved: Record<string, number[]> = {};\n    for (const [name, desc] of Object.entries(this.graph.tensors)) {\n      resolved[name] = desc.shape.map((dim) => dim as number);\n    }\n    return resolved;\n  }\n\n  private allocateActivationBuffers(): void {\n    for (const [name, desc] of Object.entries(this.graph.tensors)) {\n      if (desc.storage !== \"activation\") continue;\n      const bytes =\n        (desc.shape as number[]).reduce((a, b) => a * (b as number), 1) * DTYPE_BYTES[desc.dtype];\n      this.activationBuffers.set(name, createStorageBuffer(this.ctx, `mact_${name}`, bytes));\n    }\n  }\n\n  private async readBack(name: string, elements: number): Promise<Float32Array> {\n    const buffer = this.activationBuffers.get(name);\n    if (!buffer) throw new Error(`MoonshineEncoderExecutor: no buffer \"${name}\"`);\n    const byteLen = Math.min(elements * 4, buffer.size);\n    const readback = this.ctx.device.createBuffer({ size: byteLen, usage: 0x0001 | 0x0008 });\n    const enc = this.ctx.device.createCommandEncoder();\n    enc.copyBufferToBuffer(buffer, 0, readback, 0, byteLen);\n    this.ctx.device.queue.submit([enc.finish()]);\n    await readback.mapAsync(MAP_MODE_READ, 0, byteLen);\n    const data = new Float32Array(readback.getMappedRange(0, byteLen).slice(0));\n    readback.unmap();\n    readback.destroy();\n    return data;\n  }\n\n  private gatherBuffers(\n    spec: KernelSpec,\n    node: OpNode,\n    uniformBuffer: GPUBuffer,\n  ): Array<{ buffer: GPUBuffer }> {\n    const entries: Array<{ buffer: GPUBuffer }> = [];\n    const tensorNames = [...node.inputs, ...node.outputs];\n    let tensorIdx = 0;\n    for (const binding of spec.bindings) {\n      if (binding.type === \"uniform\") {\n        entries.push({ buffer: uniformBuffer });\n      } else {\n        const tensorName = tensorNames[tensorIdx++];\n        const buffer = this.weightBuffers.get(tensorName) ?? this.activationBuffers.get(tensorName);\n        if (!buffer) {\n          throw new Error(\n            `MoonshineEncoderExecutor: no buffer for \"${tensorName}\" in op ${node.id}`,\n          );\n        }\n        entries.push({ buffer });\n      }\n    }\n    return entries;\n  }\n}\n","/**\n * MoonshineSTT — native speech-to-text engine for Gerbil's WebGPU backend.\n *\n * Moonshine is an encoder-decoder ASR model with a raw-waveform conv frontend\n * (no FFT / mel spectrogram). Unlike the causal-LM path, it needs two graphs:\n *\n *   1. ENCODER (run once per utterance): conv frontend → bidirectional transformer\n *      → encoder hidden state, which is then projected through every decoder\n *      layer's cross-attention k_proj/v_proj into FROZEN K/V buffers.\n *   2. DECODER (autoregressive): causal self-attention with a growing KV-cache,\n *      plus cross-attention into the frozen encoder K/V at every step.\n *\n * The conv frontend is length-static (Conv1dFull carries concrete L/Lout), so the\n * encoder graph is regenerated per utterance from the input sample count, and the\n * decoder graph is regenerated with the resulting encoder frame count (S_enc).\n * Weights are uploaded once and reused across utterances via per-call executors.\n *\n * Validated on Dawn (desktop) via scripts/engine/test-moonshine-transcribe.mjs.\n * The kernels are mobile-safe (≤16KB workgroup memory, clamped exp/tanh, no\n * select(), no `enable f16`) and the executors use the WebKit submit/drain\n * discipline, so the same path runs on iPad.\n */\n\nimport {\n  generateMoonshineDecoderGraph,\n  generateMoonshineEncoderGraph,\n  moonshineEncoderFrames,\n  parseMoonshineConfig,\n} from \"./architectures/moonshine.js\";\nimport type { GPUContext } from \"./device.js\";\nimport { initGPU } from \"./device.js\";\nimport { Executor } from \"./executor.js\";\nimport { loadMoonshine } from \"./model-loader.js\";\nimport { MoonshineEncoderExecutor } from \"./moonshine-executor.js\";\nimport type { Tokenizer } from \"./tokenizer.js\";\n\nconst MOONSHINE_SAMPLE_RATE = 16_000;\n/** Hard cap so a stuck decode can't loop forever (config max_position_embeddings). */\nconst DEFAULT_MAX_NEW_TOKENS = 194;\n\nexport interface MoonshineSTTOptions {\n  /** HF repo (default UsefulSensors/moonshine-base). */\n  repo?: string;\n  revision?: string;\n  hfToken?: string;\n  cacheDir?: string;\n  onProgress?: (loaded: number, total: number, message: string) => void;\n}\n\nexport interface TranscribeOptions {\n  /** Stop after this many decoded tokens (default 194). */\n  maxNewTokens?: number;\n  /**\n   * RMS energy below which a clip is treated as silence/noise (`noSpeech`).\n   * Default 0.01 (normalized 16 kHz float PCM).\n   */\n  minRms?: number;\n  /**\n   * Clips shorter than this are treated as no-speech — a sub-threshold clip is\n   * almost always silence, and any transcript on it is a hallucination.\n   * Default 0.35 seconds.\n   */\n  minSpeechSeconds?: number;\n}\n\nexport interface TranscribeResult {\n  text: string;\n  /** Decoded token ids (excluding the start token, including the trailing EOS). */\n  tokens: number[];\n  /** Number of encoder frames produced by the conv frontend. */\n  encoderFrames: number;\n  /** Audio duration in seconds (samples / 16000). */\n  audioSeconds: number;\n  /** RMS energy of the input PCM (~0..1). */\n  speechRms: number;\n  /**\n   * True when the clip almost certainly contains no speech — too quiet, too\n   * short, an empty transcript, or a non-speech marker like \"[BLANK_AUDIO]\".\n   * Skip empty/hallucinated transcripts when this is set.\n   */\n  noSpeech: boolean;\n}\n\n/** Non-speech markers Moonshine can emit on silence, e.g. \"[BLANK_AUDIO]\", \"(silence)\". */\nconst NON_SPEECH_RE = /^\\s*[[(][^)\\]]*[)\\]]\\s*$/;\n\nfunction computeRms(pcm: Float32Array): number {\n  if (pcm.length === 0) {\n    return 0;\n  }\n  let sum = 0;\n  for (let i = 0; i < pcm.length; i++) {\n    sum += pcm[i] * pcm[i];\n  }\n  return Math.sqrt(sum / pcm.length);\n}\n\nexport class MoonshineSTT {\n  private ctx: GPUContext;\n  private weights: Map<string, { data: ArrayBufferView; shape: number[] }>;\n  private tokenizer: Tokenizer;\n  private rawConfig: Record<string, unknown>;\n  private bosTokenId: number;\n  private eosTokenId: number;\n  private decoderStartTokenId: number;\n  private _destroyed = false;\n\n  /** HF architecture string, for parity with WebGPUEngine. */\n  readonly architecture = \"MoonshineForConditionalGeneration\";\n\n  private constructor(\n    ctx: GPUContext,\n    weights: Map<string, { data: ArrayBufferView; shape: number[] }>,\n    tokenizer: Tokenizer,\n    rawConfig: Record<string, unknown>,\n  ) {\n    this.ctx = ctx;\n    this.weights = weights;\n    this.tokenizer = tokenizer;\n    this.rawConfig = rawConfig;\n    this.bosTokenId = (rawConfig.bos_token_id as number) ?? 1;\n    this.eosTokenId = (rawConfig.eos_token_id as number) ?? 2;\n    // Moonshine seeds the decoder with decoder_start_token_id (== bos == 1).\n    this.decoderStartTokenId = (rawConfig.decoder_start_token_id as number) ?? this.bosTokenId;\n  }\n\n  /** Download + initialize a Moonshine STT engine. */\n  static async create(options: MoonshineSTTOptions = {}): Promise<MoonshineSTT> {\n    const ctx = await initGPU();\n    const repo = options.repo ?? \"UsefulSensors/moonshine-base\";\n    const { weights, tokenizer, rawConfig } = await loadMoonshine({\n      repo,\n      revision: options.revision,\n      hfToken: options.hfToken,\n      cacheDir: options.cacheDir,\n      onProgress: options.onProgress,\n    });\n    return new MoonshineSTT(ctx, weights, tokenizer, rawConfig);\n  }\n\n  /**\n   * Transcribe raw 16 kHz mono PCM. Runs the conv frontend + encoder once, then\n   * greedily AR-decodes with cross-attention into the frozen encoder K/V, stopping\n   * on EOS. Returns the detokenized transcript.\n   */\n  async transcribe(pcm: Float32Array, opts: TranscribeOptions = {}): Promise<TranscribeResult> {\n    if (this._destroyed) throw new Error(\"MoonshineSTT has been destroyed.\");\n    if (pcm.length < 127) {\n      throw new Error(`PCM too short: ${pcm.length} samples (need >= conv1 kernel size 127).`);\n    }\n    const d = parseMoonshineConfig(this.rawConfig);\n    const sEnc = moonshineEncoderFrames(pcm.length);\n    if (sEnc < 1) {\n      throw new Error(`Audio too short: produces ${sEnc} encoder frames.`);\n    }\n\n    const audioSeconds = pcm.length / MOONSHINE_SAMPLE_RATE;\n    const speechRms = computeRms(pcm);\n    const minRms = opts.minRms ?? 0.01;\n    const minSpeechSeconds = opts.minSpeechSeconds ?? 0.35;\n\n    // Cheap VAD: skip the encode/decode entirely for clips too quiet or too short\n    // to contain speech (a sub-threshold clip is almost always silence; any\n    // transcript on it is a hallucination).\n    if (speechRms < minRms || audioSeconds < minSpeechSeconds) {\n      return {\n        text: \"\",\n        tokens: [],\n        encoderFrames: 0,\n        audioSeconds,\n        speechRms,\n        noSpeech: true,\n      };\n    }\n\n    // ── Encoder (once) ──\n    const encoderGraph = generateMoonshineEncoderGraph(this.rawConfig, pcm.length);\n    const encExec = new MoonshineEncoderExecutor(this.ctx, encoderGraph, d.dec_layers);\n    let encK: Float32Array[];\n    let encV: Float32Array[];\n    let frames: number;\n    try {\n      encExec.uploadWeights(this.weights);\n      encExec.initBindGroups();\n      const enc = await encExec.encode(pcm);\n      encK = enc.encK;\n      encV = enc.encV;\n      frames = enc.sEnc;\n    } finally {\n      encExec.destroy();\n    }\n\n    // ── Decoder (autoregressive, greedy) ──\n    const decoderGraph = generateMoonshineDecoderGraph(this.rawConfig, frames);\n    // maxSeqLen bounds the self-attn KV cache. The decode can never exceed the\n    // model's max_position_embeddings, so cap there.\n    const maxSeqLen = Math.min(opts.maxNewTokens ?? DEFAULT_MAX_NEW_TOKENS, d.context_length);\n    const decExec = new Executor(this.ctx, decoderGraph, { maxSeqLen, kvMode: \"f32\" });\n    const tokens: number[] = [];\n    try {\n      decExec.uploadWeightsMap(selectGraphWeights(decoderGraph, this.weights));\n      decExec.initBindGroups();\n      // Bind the frozen encoder K/V into each decoder layer's cross-attention.\n      for (let i = 0; i < d.dec_layers; i++) {\n        decExec.writeInput(`enc_k_layer${i}`, encK[i]);\n        decExec.writeInput(`enc_v_layer${i}`, encV[i]);\n      }\n      decExec.reset();\n\n      // Greedy decode: seed with the decoder start token, feed argmax back in.\n      let nextToken = this.decoderStartTokenId;\n      const limit = maxSeqLen;\n      for (let step = 0; step < limit; step++) {\n        const tok = await decExec.forwardArgmax(new Uint32Array([nextToken]));\n        tokens.push(tok);\n        if (tok === this.eosTokenId) break;\n        nextToken = tok;\n      }\n    } finally {\n      decExec.destroy();\n    }\n\n    const text = this.tokenizer.decode(tokens, /* skipSpecialTokens */ true).trim();\n    const noSpeech = text.length === 0 || NON_SPEECH_RE.test(text);\n    return {\n      text,\n      tokens,\n      encoderFrames: frames,\n      audioSeconds,\n      speechRms,\n      noSpeech,\n    };\n  }\n\n  destroy(): void {\n    this._destroyed = true;\n    this.weights.clear();\n  }\n}\n\n/**\n * Build a fresh map of just the constant weights a graph references.\n *\n * Executor.uploadWeights() DELETES entries from the map it is given (to free JS\n * memory as each weight reaches the GPU). Moonshine reuses the same weights across\n * utterances and across the encoder+decoder, so we hand each executor its OWN map\n * (deletes only touch the copy; the shared ArrayBufferViews are uploaded, never\n * mutated). Scoping to graph-referenced constants also avoids uploading the\n * encoder's weights into the decoder executor.\n */\nfunction selectGraphWeights(\n  graph: { tensors: Record<string, { storage: string }> },\n  weights: Map<string, { data: ArrayBufferView; shape: number[] }>,\n): Map<string, { data: ArrayBufferView; shape: number[] }> {\n  const out = new Map<string, { data: ArrayBufferView; shape: number[] }>();\n  for (const [name, desc] of Object.entries(graph.tensors)) {\n    if (desc.storage !== \"constant\") continue;\n    const w = weights.get(name);\n    if (w) out.set(name, w);\n  }\n  return out;\n}\n"],"mappings":";;;AASA,MAAM,WAAW;AACjB,MAAM,WAAW;AAEjB,MAAM,WAAW;AACjB,MAAMA,kBAAgB;;;;;;;;;AAsCtB,eAAsB,QAAQ,SAA+C;AAE3E,MAAK,OAAO,cAAc,eAAe,CAAC,UAAU,QAAQ,OAAO,YAAY,YAC7E,KAAI;EAGF,MAAM,EAAE,WADa,MADC,IAAI,SAAS,aAAa,2BAA2B,CAClC,SAAS;AAElD,MAAI,CAAE,WAAmB,UACvB,CAAC,WAAmB,YAAY,EAAE;AAEpC,EAAC,WAAmB,UAAU,MAAM,OAAO,EAAE,CAAC;SACxC;AAKV,KAAI,OAAO,cAAc,eAAe,CAAC,UAAU,IACjD,OAAM,IAAI,MACR,4GAED;CAIH,MAAM,WACJ,OAAO,cAAc,eAAe,4BAA4B,KAAK,UAAU,UAAU;CAE3F,MAAM,UAAU,MAAM,UAAU,IAAI,eAAe,EACjD,iBAAiB,WAAW,cAAc,oBAC3C,CAAC;AAEF,KAAI,CAAC,QACH,OAAM,IAAI,MACR,oHAED;CAIH,MAAM,SAAS,QAAQ,SAAS,IAAI,aAAa;CAIjD,MAAM,eAAe,QAAQ,SAAS,IAAI,YAA8B;CAKxE,MAAM,eADgB,OAAO,YAAY,eAAe,QAAQ,KAAK,kBAAkB,QACjD,QAAQ,SAAS,IAAI,kBAAkB;CAG7E,MAAMC,mBAAqC,EAAE;AAC7C,KAAI,OACF,kBAAiB,KAAK,aAAa;AAErC,KAAI,aACF,kBAAiB,KAAK,YAA8B;AAEtD,KAAI,aACF,kBAAiB,KAAK,kBAAkB;CAI1C,MAAMC,gBAAwC;EAC5C,eAAe,QAAQ,OAAO;EAC9B,6BAA6B,QAAQ,OAAO;EAC5C,0BAA0B,KAAK,IAAI,KAAK,QAAQ,OAAO,yBAAyB;EAChF,0BAA0B,KAAK,IAAI,KAAK,QAAQ,OAAO,yBAAyB;EAChF,gCAAgC,QAAQ,OAAO;EAC/C,iCAAiC,QAAQ,OAAO;EACjD;CAED,IAAIC;AACJ,KAAI;AACF,WAAS,MAAM,QAAQ,cAAc;GACnC;GACA,gBAAgB;GACjB,CAAC;UACK,GAAG;AAEV,MAAI;AACF,YAAS,MAAM,QAAQ,cAAc,EAAE,kBAAkB,CAAC;UACpD;AACN,SAAM,IAAI,MACR,mCAAmC,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,CAAC,mDAE/E;;;AAKL,QAAO,KAAK,MAAM,SAAS;AACzB,UAAQ,MAAM,gCAAgC,KAAK,UAAU,KAAK,QAAQ;AAC1E,WAAS,eAAe,KAAK,QAAQ,KAAK,QAAQ;GAClD;AAKF,KAAI;AACF,SAAO,qBAAqB,OAAgC;AAC1D,WAAQ,MAAM,kCAAkC,GAAG,OAAO,WAAW,GAAG,QAAQ;;SAE5E;CASR,IAAI,qBAAqB;AACzB,KAAI;EACF,MAAM,OAAQ,QAAgB,QAAS,QAAgB,WAAW;AAClE,MAAI,KACF,sBAAqB,UAAU,KAAK,UAAU,GAAG,QAAQ,KAAK,gBAAgB,GAAG,UAAU,KAAK,UAAU,GAAG,QAAQ,KAAK,eAAe;SAErI;CAMR,MAAM,KAAK,OAAO,cAAc,eAAe,UAAU,YAAY,UAAU,YAAY;CAC3F,MAAM,gBAAgB,GAAG,SAAS,cAAc;CAChD,MAAM,gBACJ,mBAAmB,KAAK,GAAG,IAC1B,iBAAiB,OAAO,cAAc,gBAAgB,UAAU,kBAAkB,KAAK;CAC1F,MAAM,cAAc,iBAAiB,CAAC,GAAG,SAAS,UAAU,IAAI,CAAC;CACjE,MAAM,iBAAiB,iBAAiB;AACxC,KAAI,CAAC,mBACH,sBAAqB,gBAAgB,GAAG,MAAM,GAAG,IAAI;AAGvD,QAAO;EACL;EACA,QAAQ,OAAO;EACf;EACA;EACA;EACA;EACA;EACD;;;;;AAQH,SAAgB,oBACd,KACA,OACA,WACA,MACW;CACX,MAAM,QAAkB;CAGxB,MAAM,cAAc,KAAK,KAAK,YAAY,EAAE,GAAG;CAE/C,MAAM,SAAS,IAAI,OAAO,aAAa;EACrC;EACA,MAAM;EACN;EACD,CAAC;AAIF,KAAI,KACF,KAAI,gBAAgB,YAClB,KAAI,OAAO,MAAM,YAAY,QAAQ,GAAG,KAAK;KAE7C,KAAI,OAAO,MAAM,YAAY,QAAQ,GAAG,KAAK,QAAQ,KAAK,YAAY,KAAK,WAAW;AAI1F,QAAO;;;;;AAMT,SAAgB,oBACd,KACA,OACA,MACW;CACX,MAAM,aAAa,gBAAgB,cAAc,KAAK,aAAa,KAAK;CAKxE,MAAM,cAAc,KAAK,KAAK,aAAa,GAAG,GAAG;CAEjD,MAAM,SAAS,IAAI,OAAO,aAAa;EACrC;EACA,MAAM;EACN,OAAiB;EAClB,CAAC;AAEF,KAAI,gBAAgB,YAClB,KAAI,OAAO,MAAM,YAAY,QAAQ,GAAG,KAAK;KAE7C,KAAI,OAAO,MAAM,YAAY,QAAQ,GAAG,KAAK,QAAQ,KAAK,YAAY,KAAK,WAAW;AAGxF,QAAO;;;;;AAMT,SAAgB,qBAAqB,KAAiB,OAAe,WAA8B;CACjG,MAAM,cAAc,KAAK,KAAK,YAAY,EAAE,GAAG;AAC/C,QAAO,IAAI,OAAO,aAAa;EAC7B;EACA,MAAM;EACN,OAAO,WAAW;EACnB,CAAC;;;;;;;AAyBJ,MAAM,iCAAiB,IAAI,SAAqD;;AAGhF,SAAS,iBAAiB,QAAoD;CAC5E,IAAI,QAAQ,eAAe,IAAI,OAAO;AACtC,KAAI,CAAC,OAAO;AACV,0BAAQ,IAAI,KAAK;AACjB,iBAAe,IAAI,QAAQ,MAAM;;AAEnC,QAAO;;;;;AAMT,SAAgB,oBACd,KACA,OACA,UACA,aAAqB,QACrB,wBACoB;CACpB,MAAM,gBAAgB,iBAAiB,IAAI,OAAO;CAClD,MAAM,WAAW,GAAG,SAAS,IAAI;CACjC,MAAM,SAAS,cAAc,IAAI,SAAS;AAC1C,KAAI,OAAQ,QAAO;CAEnB,IAAI,OAAO;CAKX,MAAM,UAAU,KAAK,SAAS,QAAQ,IAAI,KAAK,SAAS,QAAQ;AAChE,KAAI,IAAI,UAAU,WAAW,CAAC,KAAK,SAAS,aAAa,CACvD,QAAO,kBAAkB;CAG3B,MAAM,eAAe,IAAI,OAAO,mBAAmB;EACjD,OAAO,GAAG,MAAM;EAChB;EACD,CAAC;CAEF,IAAIC,SAAqC;AACzC,KAAI,wBAAwB;EAC1B,MAAM,kBAAkB,IAAI,OAAO,sBAAsB;GACvD,OAAO,GAAG,MAAM;GAChB,SAAS;GACV,CAAC;AACF,WAAS,IAAI,OAAO,qBAAqB;GACvC,OAAO,GAAG,MAAM;GAChB,kBAAkB,CAAC,gBAAgB;GACpC,CAAC;;CAGJ,MAAM,WAAW,IAAI,OAAO,sBAAsB;EAChD;EACA;EACA,SAAS;GACP,QAAQ;GACR;GACD;EACF,CAAC;AAEF,eAAc,IAAI,UAAU,SAAS;AACrC,QAAO;;;;;;AAOT,SAAgB,mBAAmB,QAA0B;AAC3D,KAAI,OACF,gBAAe,OAAO,OAAO;;;;;AAejC,SAAgB,gBACd,KACA,UACA,SACA,OACc;AACd,QAAO,IAAI,OAAO,gBAAgB;EAChC;EACA,QAAQ,SAAS,mBAAmB,EAAE;EACtC,SAAS,QAAQ,KAAK,OAAO,OAAO;GAClC,SAAS;GACT,UAAU;IACR,QAAQ,MAAM;IACd,QAAQ,MAAM,UAAU;IACxB,MAAM,MAAM,QAAQ,MAAM,OAAO;IAClC;GACF,EAAE;EACJ,CAAC;;;;;;AASJ,SAAgB,aACd,KACA,UACA,WACA,iBACA,kBAA0B,GAC1B,kBAA0B,GACpB;CACN,MAAM,UAAU,IAAI,OAAO,sBAAsB;CACjD,MAAM,OAAO,QAAQ,kBAAkB;AACvC,MAAK,YAAY,SAAS;AAC1B,MAAK,aAAa,GAAG,UAAU;AAC/B,MAAK,mBAAmB,iBAAiB,iBAAiB,gBAAgB;AAC1E,MAAK,KAAK;AACV,KAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;;;;;;;;AAW7C,eAAsB,eACpB,KACA,QACA,UACA,aAAqB,GACrB,YACuB;CACvB,MAAM,SAAS,cAAc,OAAO,OAAO;CAE3C,MAAM,UAAU,IAAI,OAAO,sBAAsB;AACjD,SAAQ,mBAAmB,QAAQ,YAAY,UAAU,GAAG,OAAO;AACnE,KAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;AAE3C,OAAM,SAAS,SAASJ,iBAAe,GAAG,OAAO;CACjD,MAAM,SAAS,SAAS,eAAe,GAAG,OAAO;CACjD,MAAM,SAAS,IAAI,aAAa,OAAO,MAAM,EAAE,CAAC;AAChD,UAAS,OAAO;AAEhB,QAAO;;;;;AAgCT,SAAgB,eAAe,SAA4B;AACzD,MAAK,MAAM,OAAO,QAChB,KAAI,SAAS;;;;;;AAqBjB,eAAsB,UAAU,KAA+C;CAC7E,MAAMK,UAAoB,EAAE;CAC5B,IAAI,kBAAkB;CACtB,IAAI,eAAe;CACnB,IAAI,oBAAoB;AAGxB,KAAI;AACF,UAAQ,KACN,kBAAkB,IAAI,OAAO,yBAAyB,cACvC,IAAI,OAAO,+BAA+B,YAC5C,IAAI,OAAO,gBAAgB,OAAO,MAAM,QAAQ,EAAE,CAAC,UACvD,IAAI,OAAO,cAAc,IAAI,eACvC;SACK;AAGR,KAAI;EACF,MAAM,WAAW,IAAI,aAAa;GAAC;GAAK;GAAK;GAAK;GAAK;GAAM;GAAK;GAAO;GAAM,CAAC;EAChF,MAAM,SAAS,oBAAoB,KAAK,YAAY,SAAS,YAAY,SAAS;EAClF,MAAM,UAAU,qBAAqB,KAAK,aAAa,SAAS,WAAW;EAC3E,MAAM,SAAS,MAAM,eAAe,KAAK,QAAQ,SAAS,GAAG,SAAS,WAAW;EAEjF,IAAI,QAAQ;AACZ,OAAK,IAAI,IAAI,GAAG,IAAI,SAAS,QAAQ,IACnC,KAAI,KAAK,IAAI,OAAO,KAAK,SAAS,GAAG,GAAG,MAAM;AAC5C,WAAQ,KAAK,uBAAuB,EAAE,cAAc,SAAS,GAAG,QAAQ,OAAO,KAAK;AACpF,WAAQ;;AAGZ,oBAAkB;AAClB,UAAQ,KACN,kBAAkB,gCAAgC,kCACnD;AACD,SAAO,SAAS;AAChB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,wBAAwB,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;;CAgBpF,MAAM,mBAAmB,OACvB,OACA,MACA,UACA,SACA,QACA,WACA,aAAqB,MACK;EAC1B,MAAM,WAAW,oBAAoB,KAAK,OAAO,MAAM,OAAO;EAC9D,MAAM,KAAK,gBACT,KACA,UACA,SAAS,KAAK,OAAO,EAAE,QAAQ,GAAG,EAAE,EACpC,GAAG,MAAM,KACV;EACD,MAAM,UAAU,IAAI,OAAO,sBAAsB;EACjD,MAAM,OAAO,QAAQ,kBAAkB;AACvC,OAAK,YAAY,SAAS;AAC1B,OAAK,aAAa,GAAG,GAAG;AACxB,OAAK,mBAAmB,WAAW;AACnC,OAAK,KAAK;AACV,UAAQ,mBAAmB,QAAQ,GAAG,SAAS,GAAG,UAAU;AAC5D,MAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;AAC3C,MAAI,IAAI,OAAO,MAAM,oBACnB,OAAM,IAAI,OAAO,MAAM,qBAAqB;AAE9C,QAAM,QAAQ,SAASL,iBAAe,GAAG,UAAU;EACnD,MAAM,SAAS,QAAQ,eAAe,GAAG,UAAU;EACnD,MAAM,SAAS,IAAI,aAAa,OAAO,MAAM,EAAE,CAAC;AAChD,UAAQ,OAAO;AACf,SAAO;;;CAIT,MAAM,sBAAsB,OAC1B,OACA,MACA,UACA,SACA,QACA,WACA,aAAqB,MACK;EAC1B,MAAM,WAAW,oBAAoB,KAAK,OAAO,MAAM,OAAO;AAQ9D,eAAa,KAAK,UAPP,gBACT,KACA,UACA,SAAS,KAAK,OAAO,EAAE,QAAQ,GAAG,EAAE,EACpC,GAAG,MAAM,KACV,EAE+B,WAAW;AAE3C,SAAO,eAAe,KAAK,QAAQ,SAAS,GAAG,UAAU;;AAI3D,KAAI;EACF,MAAM,SAAS;;;;;EAKf,MAAM,SAAS,oBAAoB,KAAK,UAAU,EAAE;EACpD,MAAM,UAAU,qBAAqB,KAAK,WAAW,EAAE;EACvD,MAAM,SAAS,MAAM,oBAAoB,MAAM,QAAQ,CAAC,OAAO,EAAE,SAAS,QAAQ,EAAE;EACpF,MAAM,KAAK,KAAK,IAAI,OAAO,KAAK,GAAK,GAAG;AACxC,UAAQ,KACN,KACI,8BAA8B,OAAO,GAAG,KACxC,sCAAsC,OAAO,GAAG,GACrD;AACD,SAAO,SAAS;AAChB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAI1E,KAAI;EACF,MAAM,SAAS;;;;;;EAOf,MAAM,QAAQ,oBAAoB,KAAK,SAAS,GADlC,IAAI,aAAa,CAAC,EAAI,CAAC,CACoB;EACzD,MAAM,SAAS,oBAAoB,KAAK,UAAU,EAAE;EACpD,MAAM,UAAU,qBAAqB,KAAK,WAAW,EAAE;EACvD,MAAM,SAAS,MAAM,iBAAiB,MAAM,QAAQ,CAAC,OAAO,OAAO,EAAE,SAAS,QAAQ,EAAE;EACxF,MAAM,KAAK,KAAK,IAAI,OAAO,KAAK,EAAI,GAAG;AACvC,UAAQ,KACN,KAAK,0BAA0B,OAAO,GAAG,KAAK,kCAAkC,OAAO,GAAG,GAC3F;AACD,iBAAe;AACf,QAAM,SAAS;AACf,SAAO,SAAS;AAChB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAI1E,KAAI;EACF,MAAM,SAAS;;;;;;;EAOf,MAAM,SAAS,oBAAoB,KAAK,UAAU,GAAG;EACrD,MAAM,UAAU,qBAAqB,KAAK,WAAW,GAAG;EACxD,MAAM,SAAS,MAAM,iBAAiB,MAAM,QAAQ,CAAC,OAAO,EAAE,SAAS,QAAQ,GAAG;EAClF,MAAM,KAAK,KAAK,IAAI,OAAO,KAAK,EAAI,GAAG,QAAQ,KAAK,IAAI,OAAO,KAAK,EAAI,GAAG;AAC3E,UAAQ,KACN,KACI,iBAAiB,MAAM,KAAK,OAAO,CAAC,KAAK,IAAI,CAAC,KAC9C,qBAAqB,MAAM,KAAK,OAAO,CAAC,KAAK,IAAI,CAAC,GACvD;AACD,SAAO,SAAS;AAChB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAM1E,KAAI;EACF,MAAM,SAAS;;;;;;;;;;;;EAYf,MAAM,SAAS,oBAAoB,KAAK,UAAU,EAAE;EACpD,MAAM,UAAU,qBAAqB,KAAK,WAAW,EAAE;EACvD,MAAM,SAAS,MAAM,iBAAiB,MAAM,QAAQ,CAAC,OAAO,EAAE,SAAS,QAAQ,EAAE;AACjF,sBAAoB,KAAK,IAAI,OAAO,KAAK,GAAK,GAAG;AACjD,UAAQ,KACN,oBACI,uBAAuB,OAAO,GAAG,KACjC,+BAA+B,OAAO,GAAG,GAC9C;AACD,SAAO,SAAS;AAChB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAI1E,KAAI;EACF,MAAM,SAAS;;;;;;;;EASf,MAAM,QAAQ,oBAAoB,KAAK,SAAS,IADlC,IAAI,aAAa;GAAC;GAAK;GAAK;GAAM;GAAI,CAAC,CACK;EAC1D,MAAM,SAAS,oBAAoB,KAAK,UAAU,GAAG;EACrD,MAAM,UAAU,qBAAqB,KAAK,WAAW,GAAG;EACxD,MAAM,SAAS,MAAM,oBAAoB,MAAM,QAAQ,CAAC,OAAO,OAAO,EAAE,SAAS,QAAQ,GAAG;EAC5F,MAAM,KAAK,KAAK,IAAI,OAAO,KAAK,EAAI,GAAG,QAAQ,KAAK,IAAI,OAAO,KAAK,GAAK,GAAG;AAC5E,UAAQ,KACN,KAAK,uBAAuB,2BAA2B,MAAM,KAAK,OAAO,CAAC,KAAK,IAAI,CAAC,GACrF;AACD,QAAM,SAAS;AACf,SAAO,SAAS;AAChB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAM1E,KAAI;EACF,MAAM,SAAS;;;;;;;;;;;;;;;;;;;;;;EAuBf,MAAM,QAAQ,IAAI,aAAa;GAAC;GAAK;GAAM;GAAK;GAAM;GAAO;GAAM;GAAO;GAAK,CAAC;EAChF,MAAM,QAAQ,oBAAoB,KAAK,SAAS,MAAM,YAAY,MAAM;EACxE,MAAM,SAAS,oBAAoB,KAAK,UAAU,MAAM,WAAW;EACnE,MAAM,UAAU,qBAAqB,KAAK,WAAW,MAAM,WAAW;EACtE,MAAM,SAAS,MAAM,iBACnB,MACA,QACA,CAAC,OAAO,OAAO,EACf,SACA,QACA,MAAM,WACP;EAED,IAAI,KAAK;EACT,MAAMM,aAAuB,EAAE;AAC/B,OAAK,IAAI,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;GACrC,MAAM,MAAM,KAAK,IAAI,KAAK,IAAI,MAAM,GAAG,GAAG,KAAM,KAAM;AACtD,OAAI,KAAK,IAAI,OAAO,KAAK,MAAM,GAAG,GAAG,KAAK;AACxC,SAAK;AACL,eAAW,KAAK,IAAI,EAAE,IAAI,MAAM,GAAG,GAAG,OAAO,KAAK;;;AAGtD,UAAQ,KACN,KACI,iCACA,qCAAqC,WAAW,KAAK,KAAK,CAAC,GAChE;AACD,QAAM,SAAS;AACf,SAAO,SAAS;AAChB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAK1E,KAAI;EACF,MAAM,SAAS;;;;;;;;;;;;;;;;;;;;;;;EAuBf,MAAM,QAAQ,IAAI,aAAa;GAAC;GAAK;GAAO;GAAK;GAAK;GAAM;GAAK;GAAK;GAAK,CAAC;EAC5E,MAAM,QAAQ,oBAAoB,KAAK,SAAS,MAAM,YAAY,MAAM;EACxE,MAAM,YAAY,oBAAoB,KAAK,aAAa,GAAG;EAC3D,MAAM,SAAS,oBAAoB,KAAK,UAAU,MAAM,WAAW;EACnE,MAAM,UAAU,qBAAqB,KAAK,WAAW,MAAM,WAAW;EACtE,MAAM,SAAS,MAAM,iBACnB,MACA,QACA;GAAC;GAAO;GAAW;GAAO,EAC1B,SACA,QACA,MAAM,WACP;EACD,IAAI,KAAK;EACT,MAAMA,aAAuB,EAAE;AAC/B,OAAK,IAAI,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;GACrC,MAAM,MAAM,KAAK,IAAI,KAAK,IAAI,MAAM,GAAG,GAAG,KAAM,KAAM;AACtD,OAAI,KAAK,IAAI,OAAO,KAAK,MAAM,GAAG,GAAG,KAAK;AACxC,SAAK;AACL,eAAW,KAAK,IAAI,EAAE,IAAI,MAAM,GAAG,GAAG,OAAO,KAAK;;;AAGtD,UAAQ,KACN,KACI,iCACA,qCAAqC,WAAW,KAAK,KAAK,CAAC,GAChE;AACD,QAAM,SAAS;AACf,YAAU,SAAS;AACnB,SAAO,SAAS;AAChB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAM1E,KAAI;EACF,MAAM,SAAS;;;;;;;;;;;;;;;;;;;;;;EAuBf,MAAM,QAAQ,IAAI,aAAa,IAAI;AACnC,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,IAAK,OAAM,KAAK,IAAI;EAC7C,MAAM,cAAe,MAAM,MAAM,MAAO;EACxC,MAAM,QAAQ,oBAAoB,KAAK,SAAS,MAAM,YAAY,MAAM;EACxE,MAAM,SAAS,oBAAoB,KAAK,UAAU,MAAM,EAAE;EAC1D,MAAM,UAAU,qBAAqB,KAAK,WAAW,MAAM,EAAE;EAE7D,MAAM,OADS,MAAM,iBAAiB,MAAM,QAAQ,CAAC,OAAO,OAAO,EAAE,SAAS,QAAQ,MAAM,EAAE,EAC3E;EACnB,MAAM,KAAK,KAAK,IAAI,MAAM,YAAY,GAAG,cAAc;AACvD,UAAQ,KACN,KACI,mCAAmC,IAAI,KACvC,2CAA2C,IAAI,aAAa,YAAY,GAC7E;AACD,QAAM,SAAS;AACf,SAAO,SAAS;AAChB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAK1E,KAAI;EACF,MAAM,SAAS;;;;;;;;;;;;;;;;;EAkBf,MAAM,aAAa,IAAI,YAAY,CAAC,WAAW,CAAC;EAChD,MAAM,YAAY,IAAI,aAAa,CAAC,GAAI,CAAC;EACzC,MAAM,WAAW,IAAI,aAAa,CAAC,EAAI,CAAC;EAExC,MAAM,WAAW;GAAC;GAAG;GAAG;GAAG;GAAG;GAAG;GAAG;GAAG;GAAE,CAAC,KAAK,OAAO,IAAI,KAAO,GAAI;EAErE,MAAM,YAAY,oBAAoB,KAAK,aAAa,GAAG,WAAW;EACtE,MAAM,WAAW,oBAAoB,KAAK,YAAY,GAAG,UAAU;EACnE,MAAM,UAAU,oBAAoB,KAAK,WAAW,GAAG,SAAS;EAChE,MAAM,SAAS,oBAAoB,KAAK,UAAU,GAAG;EACrD,MAAM,UAAU,qBAAqB,KAAK,WAAW,GAAG;EACxD,MAAM,SAAS,MAAM,iBACnB,MACA,QACA;GAAC;GAAW;GAAU;GAAS;GAAO,EACtC,SACA,QACA,GACD;EACD,IAAI,KAAK;EACT,MAAMA,aAAuB,EAAE;AAC/B,OAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IACrB,KAAI,KAAK,IAAI,OAAO,KAAK,SAAS,GAAG,GAAG,MAAO;AAC7C,QAAK;AACL,cAAW,KAAK,IAAI,EAAE,IAAI,SAAS,GAAG,GAAG,OAAO,KAAK;;AAGzD,UAAQ,KACN,KAAK,yBAAyB,6BAA6B,WAAW,KAAK,KAAK,CAAC,GAClF;AACD,YAAU,SAAS;AACnB,WAAS,SAAS;AAClB,UAAQ,SAAS;AACjB,SAAO,SAAS;AAChB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAM1E,KAAI;EACF,MAAM,YAAY;;;;;;;;EAQlB,MAAM,YAAY;;;;;;;EAclB,MAAM,SAAS,oBAAoB,KAAK,UAAU,IADlC,IAAI,aAAa;GAAC;GAAG;GAAG;GAAG;GAAE,CAAC,CACgB;EAC9D,MAAM,SAAS,oBAAoB,KAAK,UAAU,IAAI,IAAI,aAAa,EAAE,CAAC;EAC1E,MAAM,UAAU,qBAAqB,KAAK,WAAW,GAAG;EAExD,MAAM,cAAc,oBAAoB,KAAK,UAAU,WAAW,OAAO;EACzE,MAAM,cAAc,oBAAoB,KAAK,UAAU,WAAW,OAAO;EAEzE,MAAM,QAAQ,gBACZ,KACA,aACA,CAAC,EAAE,QAAQ,QAAQ,EAAE,EAAE,QAAQ,QAAQ,CAAC,EACxC,YACD;EACD,MAAM,QAAQ,gBAAgB,KAAK,aAAa,CAAC,EAAE,QAAQ,QAAQ,CAAC,EAAE,YAAY;EAElF,MAAM,UAAU,IAAI,OAAO,sBAAsB;EACjD,MAAM,OAAO,QAAQ,kBAAkB;AACvC,OAAK,YAAY,YAAY;AAC7B,OAAK,aAAa,GAAG,MAAM;AAC3B,OAAK,mBAAmB,EAAE;AAC1B,OAAK,YAAY,YAAY;AAC7B,OAAK,aAAa,GAAG,MAAM;AAC3B,OAAK,mBAAmB,EAAE;AAC1B,OAAK,YAAY,YAAY;AAC7B,OAAK,aAAa,GAAG,MAAM;AAC3B,OAAK,mBAAmB,EAAE;AAC1B,OAAK,KAAK;AACV,UAAQ,mBAAmB,QAAQ,GAAG,SAAS,GAAG,GAAG;AACrD,MAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;AAC3C,MAAI,IAAI,OAAO,MAAM,oBACnB,OAAM,IAAI,OAAO,MAAM,qBAAqB;AAE9C,QAAM,QAAQ,SAASN,iBAAe,GAAG,GAAG;EAC5C,MAAM,SAAS,QAAQ,eAAe,GAAG,GAAG;EAC5C,MAAM,SAAS,IAAI,aAAa,OAAO,MAAM,EAAE,CAAC;AAChD,UAAQ,OAAO;EAEf,MAAM,WAAW;GAAC;GAAG;GAAG;GAAG;GAAG;EAC9B,IAAI,KAAK;AACT,OAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IACrB,KAAI,KAAK,IAAI,OAAO,KAAK,SAAS,GAAG,GAAG,IAAM,MAAK;AAErD,UAAQ,KACN,KACI,iCAAiC,MAAM,KAAK,OAAO,CAAC,KAAK,IAAI,CAAC,KAC9D,yCAAyC,MAAM,KAAK,OAAO,CAAC,KAAK,IAAI,CAAC,aAAa,SAAS,KAAK,IAAI,CAAC,GAC3G;AACD,SAAO,SAAS;AAChB,SAAO,SAAS;AAChB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAK1E,KAAI;EACF,MAAM,SAAS;;;;;;;;;;;;;;;;;EAiBf,MAAM,SAAS,oBAAoB,KAAK,UAAU,GAAG;EACrD,MAAM,UAAU,qBAAqB,KAAK,WAAW,GAAG;EACxD,MAAM,SAAS,MAAM,iBAAiB,MAAM,QAAQ,CAAC,OAAO,EAAE,SAAS,QAAQ,GAAG;EAClF,MAAM,YAAY,OAAO,KAAK,KAAK,OAAO,KAAK,SAAS,OAAO,KAAK,KAAK,OAAO,KAAK;EACrF,MAAM,WAAW,KAAK,IAAI,OAAO,KAAK,KAAK,IAAI,IAAI,CAAC,GAAG;AACrC,OAAK,IAAI,OAAO,KAAK,EAAI;AAEzC,SAAO,OAAO,QAAQ,OAAO,MAAM,OAAO,GAAG,IAAI,OAAO,MAAM,OAAO,GAAG,IAAI,OAAO;EAErF,MAAM,SACJ,WAAW,OAAO,GAAG,cAAc,EAAE,CAAC,GAAG,OAAO,GAAG,cAAc,EAAE,CAAC,YACxD,OAAO,GAAG,cAAc,EAAE,CAAC,OAAO,OAAO,GAAG,QAAQ,EAAE,CAAC,kBACjD,OAAO,GAAG,cAAc,OAAO,GAAG,cAAc,OAAO,GAAG,QACpE,OAAO;AAEjB,UAAQ,KACN,aAAa,WACT,wBAAwB,OAAO,KAC/B,4BAA4B,OAAO,GACxC;AACD,SAAO,SAAS;AAChB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAM1E,KAAI;EACF,MAAM,KAAK,MAAM;EACjB,MAAM,WAAW,IAAI,aAAa,GAAG;AAErC,OAAK,IAAI,IAAI,GAAG,IAAI,IAAI,IACtB,UAAS,KAAK,KAAK,IAAI,IAAI,KAAM,IAAI,IAAI,OAAQ;EAEnD,MAAM,SAAS,oBAAoB,KAAK,UAAU,SAAS,YAAY,SAAS;EAChF,MAAM,UAAU,qBAAqB,KAAK,WAAW,SAAS,WAAW;EACzE,MAAM,SAAS,MAAM,eAAe,KAAK,QAAQ,SAAS,GAAG,SAAS,WAAW;EAEjF,IAAI,aAAa;EACjB,IAAI,mBAAmB;AACvB,OAAK,IAAI,IAAI,GAAG,IAAI,IAAI,IACtB,KAAI,KAAK,IAAI,OAAO,KAAK,SAAS,GAAG,GAAG,MAAM;AAC5C,OAAI,qBAAqB,GAAI,oBAAmB;AAChD;;EAGJ,MAAM,KAAK,eAAe;AAC1B,UAAQ,KACN,KACI,iCAAiC,GAAG,YACpC,2BAA2B,WAAW,yBAAyB,iBAAiB,cAAc,SAAS,kBAAkB,QAAQ,OAAO,kBAAkB,GAC/J;AACD,SAAO,SAAS;AAChB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAM1E,KAAI;EAGF,MAAM,UAAU,IAAI,aADA,MAAM,KACmB;EAC7C,MAAM,aAAa,MAAM;EACzB,MAAM,WAAW,MAAM;AACvB,OAAK,IAAI,IAAI,GAAG,IAAI,UAAU,IAC5B,SAAQ,aAAa,KAAK,KAAK,IAAI,IAAI,KAAM,IAAI,IAAI,OAAO;EAG9D,MAAM,OAAO,IAAI,aAAa,QAAQ,QAAQ,aAAa,GAAG,SAAS;EAEvE,MAAM,SAAS,oBAAoB,KAAK,UAAU,KAAK,YAAY,KAAK;EACxE,MAAM,UAAU,qBAAqB,KAAK,WAAW,KAAK,WAAW;EACrE,MAAM,SAAS,MAAM,eAAe,KAAK,QAAQ,SAAS,GAAG,KAAK,WAAW;EAE7E,IAAI,aAAa;EACjB,IAAI,mBAAmB;AACvB,OAAK,IAAI,IAAI,GAAG,IAAI,UAAU,IAC5B,KAAI,KAAK,IAAI,OAAO,KAAK,KAAK,GAAG,GAAG,MAAM;AACxC,OAAI,qBAAqB,GAAI,oBAAmB;AAChD;;EAGJ,MAAM,KAAK,eAAe;AAC1B,UAAQ,KACN,KACI,0DACA,8BAA8B,WAAW,yBAAyB,iBAAiB,cAAc,KAAK,kBAAkB,QAAQ,OAAO,kBAAkB,GAC9J;AACD,SAAO,SAAS;AAChB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAM1E,KAAI;EACF,MAAM,gBAAgB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;EA0DtB,MAAM,aAAa;EACnB,MAAM,YAAY,IAAI,aAAa,WAAW;EAC9C,MAAM,aAAa,IAAI,aAAa,WAAW;AAC/C,OAAK,IAAI,IAAI,GAAG,IAAI,YAAY,KAAK;AACnC,aAAU,KAAK,KAAK,IAAI,IAAI,IAAK,GAAG;AACpC,cAAW,KAAK,IAAM,KAAK,IAAI,IAAI,KAAM,GAAG;;EAI9C,IAAI,QAAQ;AACZ,OAAK,IAAI,IAAI,GAAG,IAAI,YAAY,IAAK,UAAS,UAAU,KAAK,UAAU;EACvE,MAAM,MAAM,KAAK,KAAK,QAAQ,aAAa,KAAK;EAChD,MAAM,iBAAiB,IAAI,aAAa,WAAW;AACnD,OAAK,IAAI,IAAI,GAAG,IAAI,YAAY,IAC9B,gBAAe,KAAM,UAAU,KAAK,MAAO,WAAW;EAIxD,MAAM,SAAS,IAAI,aAAa,EAAE;AAClC,SAAO,KAAK;EACZ,MAAM,UAAU,IAAI,YAAY,OAAO,OAAO,CAAC;EAG/C,MAAM,4BAAY,IAAI,YAAY,GAAG;EACrC,MAAM,aAAa,IAAI,SAAS,UAAU;AAC1C,aAAW,UAAU,GAAG,GAAG,KAAK;AAChC,aAAW,UAAU,GAAG,YAAY,KAAK;AACzC,aAAW,UAAU,GAAG,SAAS,KAAK;AACtC,aAAW,UAAU,IAAI,GAAG,KAAK;EAEjC,MAAM,QAAQ,oBAAoB,KAAK,SAAS,UAAU,YAAY,UAAU;EAChF,MAAM,OAAO,oBAAoB,KAAK,QAAQ,WAAW,YAAY,WAAW;EAChF,MAAM,SAAS,oBAAoB,KAAK,UAAU,aAAa,EAAE;EACjE,MAAM,aAAa,oBAAoB,KAAK,aAAa,UAAU;EACnE,MAAM,UAAU,qBAAqB,KAAK,WAAW,aAAa,EAAE;EAEpE,MAAM,WAAW,oBAAoB,KAAK,cAAc,eAAe,OAAO;EAC9E,MAAM,KAAK,gBACT,KACA,UACA;GAAC,EAAE,QAAQ,OAAO;GAAE,EAAE,QAAQ,MAAM;GAAE,EAAE,QAAQ,QAAQ;GAAE,EAAE,QAAQ,YAAY;GAAC,EACjF,QACD;EAED,MAAM,UAAU,IAAI,OAAO,sBAAsB;EACjD,MAAM,OAAO,QAAQ,kBAAkB;AACvC,OAAK,YAAY,SAAS;AAC1B,OAAK,aAAa,GAAG,GAAG;AACxB,OAAK,mBAAmB,EAAE;AAC1B,OAAK,KAAK;AACV,UAAQ,mBAAmB,QAAQ,GAAG,SAAS,GAAG,aAAa,EAAE;AACjE,MAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;AAC3C,MAAI,IAAI,OAAO,MAAM,oBACnB,OAAM,IAAI,OAAO,MAAM,qBAAqB;AAE9C,QAAM,QAAQ,SAASA,iBAAe,GAAG,aAAa,EAAE;EACxD,MAAM,SAAS,QAAQ,eAAe,GAAG,aAAa,EAAE;EACxD,MAAM,SAAS,IAAI,aAAa,OAAO,MAAM,EAAE,CAAC;AAChD,UAAQ,OAAO;EAEf,IAAI,SAAS;EACb,IAAI,YAAY;AAChB,OAAK,IAAI,IAAI,GAAG,IAAI,YAAY,KAAK;GACnC,MAAM,MAAM,KAAK,IAAI,OAAO,KAAK,eAAe,GAAG;AACnD,OAAI,MAAM,QAAQ;AAChB,aAAS;AACT,gBAAY;;;EAIhB,MAAM,KAAK,SAAS;AACpB,UAAQ,KACN,KACI,uCAAuC,OAAO,cAAc,EAAE,CAAC,OAAO,UAAU,MAChF,2CAA2C,OAAO,cAAc,EAAE,CAAC,OAAO,UAAU,cAAc,eAAe,WAAW,QAAQ,EAAE,CAAC,QAAQ,OAAO,WAAW,QAAQ,EAAE,CAAC,GACjL;AACD,QAAM,SAAS;AACf,OAAK,SAAS;AACd,SAAO,SAAS;AAChB,aAAW,SAAS;AACpB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAK1E,KAAI;EACF,MAAM,eAAe;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;EAsDrB,MAAM,QAAQ;EACd,MAAM,QAAQ;EACd,MAAM,QAAQ,IAAI,aAAa,MAAM;EACrC,MAAM,QAAQ,IAAI,aAAa,QAAQ,MAAM;AAC7C,OAAK,IAAI,IAAI,GAAG,IAAI,OAAO,IAAK,OAAM,KAAK,KAAK,IAAI,IAAI,IAAK,GAAG;AAChE,OAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,OAAO,IAAK,OAAM,KAAK,KAAK,IAAI,IAAI,KAAM,GAAG;EAGzE,MAAM,YAAY,IAAI,aAAa,MAAM;AACzC,OAAK,IAAI,IAAI,GAAG,IAAI,OAAO,KAAK;GAC9B,IAAI,IAAI;AACR,QAAK,IAAI,IAAI,GAAG,IAAI,OAAO,IAAK,MAAK,MAAM,KAAK,MAAM,IAAI,QAAQ;AAClE,aAAU,KAAK;;EAGjB,MAAM,4BAAY,IAAI,YAAY,EAAE;EACpC,MAAM,MAAM,IAAI,SAAS,UAAU;AACnC,MAAI,UAAU,GAAG,OAAO,KAAK;AAC7B,MAAI,UAAU,GAAG,OAAO,KAAK;EAE7B,MAAM,OAAO,oBAAoB,KAAK,QAAQ,MAAM,YAAY,MAAM;EACtE,MAAM,OAAO,oBAAoB,KAAK,QAAQ,MAAM,YAAY,MAAM;EACtE,MAAM,OAAO,oBAAoB,KAAK,QAAQ,QAAQ,EAAE;EACxD,MAAM,OAAO,oBAAoB,KAAK,aAAa,UAAU;EAC7D,MAAM,OAAO,qBAAqB,KAAK,WAAW,QAAQ,EAAE;EAE5D,MAAM,aAAa,oBAAoB,KAAK,aAAa,cAAc,OAAO;EAC9E,MAAM,OAAO,gBACX,KACA,YACA;GAAC,EAAE,QAAQ,MAAM;GAAE,EAAE,QAAQ,MAAM;GAAE,EAAE,QAAQ,MAAM;GAAE,EAAE,QAAQ,MAAM;GAAC,EACxE,QACD;EAGD,MAAM,WAAW,IAAI,OAAO,sBAAsB;EAClD,MAAM,QAAQ,SAAS,kBAAkB;AACzC,QAAM,YAAY,WAAW;AAC7B,QAAM,aAAa,GAAG,KAAK;AAC3B,QAAM,mBAAmB,KAAK,KAAK,QAAQ,EAAE,CAAC;AAC9C,QAAM,KAAK;AACX,WAAS,mBAAmB,MAAM,GAAG,MAAM,GAAG,QAAQ,EAAE;AACxD,MAAI,OAAO,MAAM,OAAO,CAAC,SAAS,QAAQ,CAAC,CAAC;AAC5C,MAAI,IAAI,OAAO,MAAM,oBACnB,OAAM,IAAI,OAAO,MAAM,qBAAqB;AAE9C,QAAM,KAAK,SAASA,iBAAe,GAAG,QAAQ,EAAE;EAChD,MAAM,UAAU,KAAK,eAAe,GAAG,QAAQ,EAAE;EACjD,MAAM,UAAU,IAAI,aAAa,QAAQ,MAAM,EAAE,CAAC;AAClD,OAAK,OAAO;EAEZ,IAAI,UAAU;EACd,IAAI,aAAa;AACjB,OAAK,IAAI,IAAI,GAAG,IAAI,OAAO,KAAK;GAC9B,MAAM,MAAM,KAAK,IAAI,QAAQ,KAAK,UAAU,GAAG;AAC/C,OAAI,MAAM,SAAS;AACjB,cAAU;AACV,iBAAa;;;EAGjB,MAAM,MAAM,UAAU;AACtB,UAAQ,KACN,MACI,sCAAsC,QAAQ,cAAc,EAAE,CAAC,OAAO,WAAW,MACjF,0CAA0C,QAAQ,cAAc,EAAE,CAAC,OAAO,WAAW,cAAc,UAAU,YAAY,QAAQ,EAAE,CAAC,QAAQ,QAAQ,YAAY,QAAQ,EAAE,CAAC,GAChL;AACD,OAAK,SAAS;AACd,OAAK,SAAS;AACd,OAAK,SAAS;AACd,OAAK,SAAS;AACd,OAAK,SAAS;UACP,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAO1E,KAAI;EACF,MAAM,aAAa;;;;;;;EAOnB,MAAM,iBAAiB;EACvB,MAAM,UAAU,oBAAoB,KAAK,WAAW,IAAI,IAAI,aAAa,EAAE,CAAC;EAC5E,MAAM,cAAc,qBAAqB,KAAK,WAAW,GAAG;EAC5D,MAAM,eAAe,oBAAoB,KAAK,MAAM,YAAY,OAAO;EACvE,MAAM,SAAS,gBAAgB,KAAK,cAAc,CAAC,EAAE,QAAQ,SAAS,CAAC,EAAE,QAAQ;EAEjF,MAAM,MAAM,IAAI,OAAO,sBAAsB;EAC7C,MAAM,QAAQ,IAAI,kBAAkB;AACpC,OAAK,IAAI,IAAI,GAAG,IAAI,gBAAgB,KAAK;AACvC,SAAM,YAAY,aAAa;AAC/B,SAAM,aAAa,GAAG,OAAO;AAC7B,SAAM,mBAAmB,EAAE;;AAE7B,QAAM,KAAK;AACX,MAAI,mBAAmB,SAAS,GAAG,aAAa,GAAG,GAAG;AACtD,MAAI,OAAO,MAAM,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;AACvC,MAAI,IAAI,OAAO,MAAM,oBACnB,OAAM,IAAI,OAAO,MAAM,qBAAqB;AAE9C,QAAM,YAAY,SAASA,iBAAe,GAAG,GAAG;EAChD,MAAM,aAAa,YAAY,eAAe,GAAG,GAAG;EACpD,MAAM,aAAa,IAAI,aAAa,WAAW,MAAM,EAAE,CAAC;AACxD,cAAY,OAAO;EAEnB,MAAM,aAAa,WAAW,OAAO,MAAM,KAAK,IAAI,IAAI,eAAe,GAAG,GAAI;AAC9E,UAAQ,KACN,aACI,QAAQ,eAAe,kBAAkB,WAAW,GAAG,KACvD,QAAQ,eAAe,2BAA2B,MAAM,KAAK,WAAW,CAAC,KAAK,IAAI,CAAC,cAAc,eAAe,GACrH;AACD,UAAQ,SAAS;AACjB,cAAY,SAAS;UACd,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAQ1E,KAAI;EACF,MAAM,UAAU;;;;;;;;EAQhB,MAAM,OAAO,oBAAoB,KAAK,UAAU,IAAI,IAAI,aAAa;GAAC;GAAG;GAAG;GAAG;GAAE,CAAC,CAAC;EACnF,MAAM,QAAQ,oBAAoB,KAAK,WAAW,IAAI,IAAI,aAAa,EAAE,CAAC;EAC1E,MAAM,QAAQ,oBAAoB,KAAK,WAAW,IAAI,IAAI,aAAa,EAAE,CAAC;EAC1E,MAAM,SAAS,qBAAqB,KAAK,YAAY,GAAG;EACxD,MAAM,SAAS,qBAAqB,KAAK,YAAY,GAAG;EAGxD,MAAM,YAAY,oBAAoB,KAAK,MAAM,SAAS,OAAO;EACjE,MAAM,OAAO,gBAAgB,KAAK,WAAW,CAAC,EAAE,QAAQ,MAAM,EAAE,EAAE,QAAQ,OAAO,CAAC,EAAE,SAAS;EAC7F,MAAM,OAAO,gBAAgB,KAAK,WAAW,CAAC,EAAE,QAAQ,MAAM,EAAE,EAAE,QAAQ,OAAO,CAAC,EAAE,SAAS;EAE7F,MAAM,OAAO,IAAI,OAAO,sBAAsB;EAC9C,MAAM,QAAQ,KAAK,kBAAkB;AACrC,QAAM,YAAY,UAAU;AAC5B,QAAM,aAAa,GAAG,KAAK;AAC3B,QAAM,mBAAmB,EAAE;AAC3B,QAAM,YAAY,UAAU;AAC5B,QAAM,aAAa,GAAG,KAAK;AAC3B,QAAM,mBAAmB,EAAE;AAC3B,QAAM,KAAK;AACX,OAAK,mBAAmB,OAAO,GAAG,QAAQ,GAAG,GAAG;AAChD,OAAK,mBAAmB,OAAO,GAAG,QAAQ,GAAG,GAAG;AAChD,MAAI,OAAO,MAAM,OAAO,CAAC,KAAK,QAAQ,CAAC,CAAC;AACxC,MAAI,IAAI,OAAO,MAAM,oBACnB,OAAM,IAAI,OAAO,MAAM,qBAAqB;AAE9C,QAAM,OAAO,SAASA,iBAAe,GAAG,GAAG;EAC3C,MAAM,MAAM,IAAI,aAAa,OAAO,eAAe,GAAG,GAAG,CAAC,MAAM,EAAE,CAAC;AACnE,SAAO,OAAO;AACd,QAAM,OAAO,SAASA,iBAAe,GAAG,GAAG;EAC3C,MAAM,MAAM,IAAI,aAAa,OAAO,eAAe,GAAG,GAAG,CAAC,MAAM,EAAE,CAAC;AACnE,SAAO,OAAO;EAEd,MAAM,YAAY;GAAC;GAAG;GAAG;GAAG;GAAE;EAC9B,MAAM,SAAS,UAAU,OAAO,GAAG,MAAM,KAAK,IAAI,IAAI,KAAK,EAAE,GAAG,IAAK;EACrE,MAAM,SAAS,UAAU,OAAO,GAAG,MAAM,KAAK,IAAI,IAAI,KAAK,EAAE,GAAG,IAAK;AACrE,UAAQ,KACN,UAAU,SACN,6CAA6C,MAAM,KAAK,IAAI,CAAC,UAAU,MAAM,KAAK,IAAI,CAAC,MACvF,iDAAiD,MAAM,KAAK,IAAI,CAAC,UAAU,MAAM,KAAK,IAAI,CAAC,eAAe,UAAU,IACzH;AACD,OAAK,SAAS;AACd,QAAM,SAAS;AACf,QAAM,SAAS;AACf,SAAO,SAAS;AAChB,SAAO,SAAS;UACT,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAG1E,QAAO;EAAE;EAAiB;EAAc;EAAmB;EAAS;;;;;;ACp+CtE,SAAS,KAAK,GAAW,GAAmB;AAC1C,QAAO,KAAK,KAAK,IAAI,EAAE;;;AAIzB,SAAS,aAAa,OAAuB;CAC3C,MAAM,IAAI,IAAI,aAAa,EAAE;AAC7B,GAAE,KAAK;AACP,QAAO,IAAI,YAAY,EAAE,OAAO,CAAC;;;;;;AAOnC,SAAS,mBAAmB,QAA+B;CACzD,MAAM,sBAAM,IAAI,YAAY,OAAO,SAAS,EAAE;CAC9C,MAAM,OAAO,IAAI,SAAS,IAAI;AAC9B,MAAK,IAAI,IAAI,GAAG,IAAI,OAAO,QAAQ,IACjC,MAAK,UAAU,IAAI,GAAG,OAAO,IAAI,KAAwB;AAE3D,QAAO;;;;;;;;;;AAwCT,SAAS,cACP,IACA,gBACA,QACQ;CAER,MAAM,MAAM,GAAG,WAAW;AAC1B,KAAI,OAAO,eAAe,KAGxB,QAFc,eAAe,KACT,QAAQ,GAAW,MAAc,IAAI,GAAG,EAAE,GAC/C;CAGjB,MAAM,WAAW,eAAe,GAAG,QAAQ;AAC3C,KAAI,YAAY,SAAS,UAAU,EACjC,QAAO,SAAS;AAElB,QAAO,GAAG,WAAW;;AAKvB,MAAM,iBAAiB;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA6BvB,MAAM,sBAAsB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAsD5B,MAAM,cAAc;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA4HpB,MAAM,mBAAmB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAmHzB,MAAM,0BAA0B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAkHhC,MAAM,mBAAmB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAiFzB,MAAM,cAAc;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAsFpB,MAAM,wBAAwB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA4E9B,MAAM,6BAA6B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA0FnC,MAAM,mBAAmB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAsGzB,MAAM,yBAAyB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAiG/B,MAAM,gCAAgC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA0FtC,MAAM,cAAc;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAoEpB,MAAM,eAAe;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA2DrB,MAAM,oBAAoB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA4E1B,MAAM,iBAAiB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA+EvB,MAAM,YAAY;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA4GlB,MAAM,wBAAwB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA8F9B,MAAM,iBAAiB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAyVvB,MAAM,uBAAuB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuP7B,MAAM,eAAe;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAyErB,MAAM,YAAY;;;;;;;;;;;;;;;;;;;;AAqBlB,MAAM,YAAY;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA6BlB,MAAM,WAAW;;;;;;;;;;;;;;;;;;;AAoBjB,MAAM,WAAW;;;;;;;;;;;;;;;;;;;AAoBjB,MAAM,sBAAsB;;;;;;;;;;;;;;;;;;;;;AAsB5B,MAAM,iBAAiB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAkCvB,MAAM,aAAa;;;;;;;;;;;;;;;;;;;;;;AA4BnB,MAAM,eAAe;;;;;;;;;;;;;;;;;;AAsBrB,MAAM,cAAc;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA8DpB,MAAM,gBAAgB;;;;;;;;;;;;;;;;;;;;;;;AAwBtB,MAAM,gBAAgB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAoCtB,MAAM,kBAAkB;;;;;;;;;;;;;;;;;;;;;;;;;;AA2BxB,MAAM,gBAAgB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAiCtB,MAAM,oBAAoB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA+C1B,MAAM,aAAa;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA8EnB,MAAM,oBAAoB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAiC1B,MAAM,cAAc;;;;;;;;;;;;;;;;;;;;;;AAuBpB,MAAM,qBAAqB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAkE3B,MAAM,0BAA0B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAoHhC,MAAM,wBAAwB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAsG9B,MAAM,wBAAwB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAmE9B,MAAM,qBAAqB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAkD3B,MAAM,0BAA0B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAkDhC,MAAM,2BAA2B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAkDjC,MAAM,oBAAoB;;;;;;;;;;;;;;;;;;;;;;AAuB1B,MAAM,iBAAiB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAkIvB,MAAM,qBAAqB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqI3B,MAAM,uBAAuB;;;;;;;;;;;;;;;;;;;;AAqB7B,MAAM,4BAA4B;;;;;;;;;;;;;;;;;;;;;;;;;AA0BlC,MAAM,gCAAgC;;;;;;;;;;;;;;;;;;;;;;;AAwBtC,MAAM,uCAAuC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA8B7C,MAAM,2BAA2B;;;;;;;;;;;;;;;;;;;;;AAsBjC,MAAM,kCAAkC;;;;;;;;;;;;;;;;;;;;;;;;;;;AA4BxC,MAAM,qBAAqB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA8Q3B,MAAM,4BAA4B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAySlC,MAAM,yBAAyB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAkE/B,MAAM,mBAAmB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAsCzB,MAAM,sBAAsB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAwH5B,SAAS,mBAAmB,IAAY,gBAAkD;CACxF,MAAM,WAAW,eAAe,GAAG,QAAQ;CAC3C,IAAI,QAAQ;AACZ,MAAK,MAAM,KAAK,SAAU,UAAS;AACnC,QAAO;;AAKT,MAAMO,UAAsB;CAC1B,YAAY;CACZ,YAAY;CAMZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAGjC,YAAY,IAAI,gBAAgB;AAG9B,SAAO,mBAAmB,CAFZ,mBAAmB,IAAI,eAAe,CAEnB,CAAC;;CAErC;AAID,MAAMC,UAAsB;CAC1B,YAAY;CACZ,YAAY;CAMZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAGjC,YAAY,IAAI,gBAAgB;AAG9B,SAAO,mBAAmB,CAFZ,mBAAmB,IAAI,eAAe,CAEnB,CAAC;;CAErC;AAID,MAAMC,mBAA+B;CACnC,YAAY;CACZ,YAAY;CAKZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CAEzF,gBAAgB,IAAI;EAClB,MAAM,QAAQ,GAAG,WAAW;AAC5B,SAAO;GAAC,KAAK,OAAO,IAAI;GAAE;GAAG;GAAE;;CAGjC,YAAY,IAAI,gBAAgB;EAC9B,MAAM,QAAQ,GAAG,WAAW;AAG5B,SAAO,mBAAmB,CAAC,SAFjB,eAAe,GAAG,OAAO,MAAM,MAAM,KAER,KAAK,MAAM,CAAC;;CAEtD;AAID,MAAMC,eAA2B;CAC/B,YAAY;CACZ,YAAY;CAKZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CAEzF,gBAAgB,IAAI;EAClB,MAAM,QAAQ,GAAG,WAAW;AAC5B,SAAO;GAAC,KAAK,OAAO,IAAI;GAAE;GAAG;GAAE;;CAGjC,YAAY,IAAI,gBAAgB;EAC9B,MAAM,QAAQ,GAAG,WAAW;AAI5B,SAAO,mBAAmB,CAFhB,eAAe,GAAG,OAAO,MAAM,MAAM,GAEjB,MAAM,CAAC;;CAExC;AAID,MAAMC,YAAwB;CAC5B,YAAY;CACZ,YAAY;CAKZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CAEzF,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAGjC,YAAY,IAAI,gBAAgB;EAC9B,MAAM,QAAQ,mBAAmB,IAAI,eAAe;EACpD,MAAM,QAAQ,GAAG,WAAW;AAE5B,SAAO,mBAAmB,CAAC,OAAO,aAAa,MAAM,CAAC,CAAC;;CAE1D;AAID,MAAMC,cAA0B;CAC9B,YAAY;CACZ,YAAY;CAKZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CAEzF,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAGjC,YAAY,IAAI,gBAAgB;EAC9B,MAAM,QAAQ,mBAAmB,IAAI,eAAe;EACpD,MAAM,MAAM,GAAG,WAAW;AAE1B,SAAO,mBAAmB,CAAC,OAAO,aAAa,IAAI,CAAC,CAAC;;CAExD;AAID,MAAMC,aAAyB;CAC7B,YAAY;CACZ,YAAY;CAKZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CAEzF,gBAAgB,IAAI,gBAAgB;AAGlC,SAAO;GADM,eAAe,GAAG,QAAQ,MAAM,MAAM;GACrC;GAAG;GAAE;;CAGrB,YAAY,IAAI,gBAAgB;EAC9B,MAAM,QAAQ,GAAG,WAAW;AAG5B,SAAO,mBAAmB,CAFb,eAAe,GAAG,QAAQ,MAAM,MAAM,GAElB,MAAM,CAAC;;CAE3C;AAID,MAAMC,aAAyB;CAC7B,YAAY;CACZ,YAAY;CAMZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAGjC,YAAY,IAAI,gBAAgB;AAE9B,SAAO,mBAAmB,CADZ,mBAAmB,IAAI,eAAe,CACnB,CAAC;;CAErC;AAID,MAAMC,sBAAkC;CACtC,YAAY;CACZ,YAAY;CAQZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,SAAS,GAAG,WAAW;AAE7B,SAAO;GADQ,cAAc,IAAI,gBAAgB,OAAO;GACxC;GAAG;GAAE;;CAGvB,YAAY,IAAI,gBAAgB;EAC9B,MAAM,SAAS,GAAG,WAAW;AAG7B,SAAO,mBAAmB;GAFX,cAAc,IAAI,gBAAgB,OAAO;GAErB;GAAQ,aAD9B,GAAG,WAAW,OAAkB,KACe;GAAE;GAAE,CAAC;;CAEpE;AAID,MAAMC,WAAuB;CAC3B,YAAY;CACZ,YAAY;CAKZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CAEzF,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAGjC,YAAY,IAAI,gBAAgB;AAG9B,SAAO,mBAAmB,CAFZ,mBAAmB,IAAI,eAAe,CAEnB,CAAC;;CAErC;AAID,MAAMC,WAAuB;CAC3B,YAAY;CACZ,YAAY;CAKZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CAEzF,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAGjC,YAAY,IAAI,gBAAgB;AAG9B,SAAO,mBAAmB,CAFZ,mBAAmB,IAAI,eAAe,CAEnB,CAAC;;CAErC;AAID,MAAMC,cAA0B;CAC9B,YAAY;CACZ,YAAY;CACZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CACzF,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAEjC,YAAY,IAAI,gBAAgB;AAE9B,SAAO,mBAAmB,CADZ,mBAAmB,IAAI,eAAe,CACnB,CAAC;;CAErC;AAID,MAAMC,cAA0B;CAC9B,YAAY;CACZ,YAAY;CAGZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CACD,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAEjC,YAAY,IAAI,gBAAgB;EAC9B,MAAM,QAAQ,mBAAmB,IAAI,eAAe;EACpD,MAAM,QAAQ,GAAG,WAAW;AAC5B,SAAO,mBAAmB,CAAC,OAAO,MAAM,CAAC;;CAE5C;AAID,MAAMC,gBAA4B;CAChC,YAAY;CACZ,YAAY;CAGZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CACzF,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAEjC,YAAY,IAAI,gBAAgB;EAC9B,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,YAAY,GAAG,WAAW;EAChC,MAAM,aAAa,GAAG,WAAW;EAEjC,MAAM,OADW,eAAe,GAAG,QAAQ,IACrB;AACtB,SAAO,mBAAmB;GAAC;GAAM;GAAU;GAAW;GAAW,CAAC;;CAErE;AAID,MAAMC,cAA0B;CAC9B,YAAY;CACZ,YAAY;CAGZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CACzF,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAEjC,YAAY,IAAI,gBAAgB;EAC9B,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,YAAY,GAAG,WAAW;EAChC,MAAM,QAAQ,GAAG,WAAW;EAC5B,MAAM,QAAQ,GAAG,WAAW;EAE5B,MAAM,OADW,eAAe,GAAG,QAAQ,IACrB;AACtB,SAAO,mBAAmB;GAAC;GAAM;GAAU;GAAW;GAAO;GAAO;GAAG;GAAG;GAAE,CAAC;;CAEhF;AAID,MAAMC,kBAA8B;CAClC,YAAY;CACZ,YAAY;CAGZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CACD,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAEjC,YAAY,IAAI,gBAAgB;EAC9B,MAAM,YAAY,GAAG,WAAW;EAChC,MAAM,WAAW,GAAG,WAAW;AAI/B,SAAO,mBAAmB;GAHT,eAAe,GAAG,QAAQ,IACpB,QAAQ,GAAG,MAAM,IAAI,GAAG,EAAE,IACxB,YAAY;GACD;GAAW;GAAU;GAAE,CAAC;;CAE/D;AAID,MAAMC,aAAyB;CAC7B,YAAY;CACZ,YAAY;CAMZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAIlC,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,UAAU,GAAG,WAAW;EAC9B,IAAIC;AACJ,MAAI,WAAW,eAAe,SAC5B,KAAI,eAAe,SAAS;MAE5B,KAAI,GAAG,WAAW;AAEpB,SAAO;GAAC,KAAK,GAAG,GAAG;GAAE,KAAK,GAAG,GAAG;GAAE;GAAE;;CAGtC,YAAY,IAAI,gBAAgB;EAE9B,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,IAAI,GAAG,WAAW;EAExB,MAAM,UAAU,GAAG,WAAW;EAC9B,IAAIA;AACJ,MAAI,WAAW,eAAe,SAC5B,KAAI,eAAe,SAAS;MAE5B,KAAI,GAAG,WAAW;AAEpB,SAAO,mBAAmB;GAAC;GAAG;GAAG;GAAE,CAAC;;CAEvC;AAID,MAAMC,iBAA6B;CACjC,YAAY;CACZ,YAAY;CAGZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,UAAU,GAAG,WAAW;EAC9B,IAAID;AACJ,MAAI,WAAW,eAAe,SAC5B,KAAI,eAAe,SAAS;MAE5B,KAAI,GAAG,WAAW;AAEpB,SAAO;GAAC,KAAK,GAAG,GAAG;GAAE,KAAK,GAAG,GAAG;GAAE;GAAE;;CAGtC,YAAY,IAAI,gBAAgB;EAC9B,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,UAAU,GAAG,WAAW;EAC9B,IAAIA;AACJ,MAAI,WAAW,eAAe,SAC5B,KAAI,eAAe,SAAS;MAE5B,KAAI,GAAG,WAAW;AAEpB,SAAO,mBAAmB;GAAC;GAAG;GAAG;GAAE,CAAC;;CAEvC;AAID,MAAME,iBAA6B;CACjC,YAAY;CACZ,YAAY;CAGZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,MAAM,eAAe,GAAG,QAAQ;EACtC,MAAM,KAAK,IAAI;EACf,MAAM,IAAI,IAAI;AACd,SAAO;GAAC,KAAK,KAAK,GAAG,GAAG;GAAE;GAAG;GAAE;;CAGjC,YAAY,IAAI,gBAAgB;EAC9B,MAAM,MAAM,eAAe,GAAG,QAAQ;EACtC,MAAM,KAAK,IAAI;EACf,MAAM,IAAI,IAAI;AAId,SAAO,mBAAmB;GAAC;GAFjB,eAAe,GAAG,OAAO,MAAM,MAAO,GAAG,WAAW;GAE5B;GAAE,CAAC;;CAExC;AAID,MAAMC,oBAAgC;CACpC,YAAY;CACZ,YAAY;CAGZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,UAAU,GAAG,WAAW;EAC9B,MAAM,IACJ,WAAW,eAAe,WAAW,eAAe,SAAS,KAAM,GAAG,WAAW;AACnF,SAAO;GAAC,KAAK,GAAG,GAAG;GAAE,KAAK,GAAG,GAAG;GAAE;GAAE;;CAGtC,YAAY,IAAI,gBAAgB;EAC9B,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,UAAU,GAAG,WAAW;EAC9B,MAAM,IACJ,WAAW,eAAe,WAAW,eAAe,SAAS,KAAM,GAAG,WAAW;EAEnF,MAAM,OAAQ,GAAG,WAAW,QAAmB,OAAO;EACtD,MAAM,OAAQ,GAAG,WAAW,QAAmB,OAAO;EACtD,MAAM,OAAQ,GAAG,WAAW,QAAmB,OAAO;EACtD,MAAM,OAAQ,GAAG,WAAW,QAAmB,OAAO;AAEtD,SAAO,mBAAmB;GACxB;GACA;GACA;GACA,aAAa,KAAK;GAClB,aAAa,KAAK;GAClB,aAAa,KAAK;GAClB,aAAa,KAAK;GACnB,CAAC;;CAEL;AAKD,MAAaC,wBAAoC;CAC/C,GAAG;CACH,YAAY;CACb;AAID,MAAMC,iBAA6B;CACjC,YAAY;CACZ,YAAY;CAQZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAElC,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,UAAU,GAAG,WAAW;EAC9B,IAAIL;AACJ,MAAI,WAAW,eAAe,SAC5B,KAAI,eAAe,SAAS;MAE5B,KAAI,GAAG,WAAW;AAEpB,SAAO;GAAC,KAAK,GAAG,GAAG;GAAE,KAAK,GAAG,GAAG;GAAE;GAAE;;CAGtC,YAAY,IAAI,gBAAgB;EAE9B,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,aAAa,GAAG,WAAW;EACjC,MAAM,UAAU,GAAG,WAAW;EAC9B,IAAIA;AACJ,MAAI,WAAW,eAAe,SAC5B,KAAI,eAAe,SAAS;MAE5B,KAAI,GAAG,WAAW;AAEpB,SAAO,mBAAmB;GAAC;GAAG;GAAG;GAAG;GAAW,CAAC;;CAEnD;AAID,MAAaM,cAA0B;CACrC,YAAY;CACZ,YAAY;CAGZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI;EAClB,MAAM,IAAI,GAAG,WAAW;AAExB,SAAO;GAAC,KAAK,GADE,EACQ;GAAE;GAAG;GAAE;;CAGhC,YAAY,IAAI;EACd,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,IAAI,GAAG,WAAW;AACxB,SAAO,mBAAmB,CAAC,GAAG,EAAE,CAAC;;CAEpC;AAID,MAAaC,mBAA+B;CAC1C,YAAY;CACZ,YAAY;CAGZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI;EAClB,MAAM,IAAI,GAAG,WAAW;AAGxB,SAAO;GAAC,KAAK,GADE,GACQ;GAAE;GAAG;GAAE;;CAGhC,YAAY,IAAI;EACd,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,aAAa,GAAG,WAAW;AACjC,SAAO,mBAAmB;GAAC;GAAG;GAAG;GAAY;GAAE,CAAC;;CAEnD;AAID,MAAaC,yBAAqC;CAChD,YAAY;CACZ,YAAY;CAGZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI;EAClB,MAAM,IAAI,GAAG,WAAW;AAExB,SAAO;GAAC,KAAK,GADE,EACQ;GAAE;GAAG;GAAE;;CAGhC,YAAY,IAAI;EACd,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,aAAa,GAAG,WAAW;AACjC,SAAO,mBAAmB;GAAC;GAAG;GAAG;GAAY;GAAE,CAAC;;CAEnD;AAID,MAAaC,gCAA4C;CACvD,GAAG;CACH,YAAY;CACb;AASD,MAAaC,wBAAoC;CAC/C,GAAG;CACH,YAAY;CACb;AAED,MAAaC,6BAAyC;CACpD,GAAG;CACH,YAAY;CACb;AAID,MAAaC,qBAAiC;CAC5C,YAAY;CACZ,YAAY;CAOZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI;EAClB,MAAM,IAAI,GAAG,WAAW;AAExB,SAAO;GAAC,KAAK,GADE,EACQ;GAAE;GAAG;GAAE;;CAGhC,YAAY,IAAI;EACd,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,IAAI,GAAG,WAAW;AACxB,SAAO,mBAAmB,CAAC,GAAG,EAAE,CAAC;;CAEpC;AAID,MAAaC,0BAAsC;CACjD,YAAY;CACZ,YAAY;CAWZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI;EAClB,MAAM,IAAI,GAAG,WAAW;AAExB,SAAO;GAAC,KAAK,GADE,EACQ;GAAE;GAAG;GAAE;;CAGhC,YAAY,IAAI;EACd,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,aAAa,GAAG,WAAW;AACjC,SAAO,mBAAmB;GAAC;GAAG;GAAG;GAAY;GAAE,CAAC;;CAEnD;AAID,MAAaC,wBAAoC;CAC/C,YAAY;CACZ,YAAY;CAMZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI;EAClB,MAAM,IAAI,GAAG,WAAW;AAExB,SAAO;GAAC,KAAK,GADE,EACQ;GAAE;GAAG;GAAE;;CAGhC,YAAY,IAAI;EACd,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,aAAa,GAAG,WAAW;AACjC,SAAO,mBAAmB;GAAC;GAAG;GAAG;GAAY;GAAE,CAAC;;CAEnD;AAID,MAAaC,cAA0B;CACrC,YAAY;CACZ,YAAY;CAEZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CAEzF,kBAAkB;AAChB,SAAO;GAAC;GAAG;GAAG;GAAE;;CAGlB,YAAY,KAAK,iBAAiB;AAEhC,SAAO,mBAAmB,CAAC,GAAG,EAAE,CAAC;;CAEpC;AAID,MAAMC,cAA0B;CAC9B,YAAY;CACZ,YAAY;CAMZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAElC,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,MAAM,GAAG,WAAW;EAC1B,IAAIC;AACJ,MAAI,OAAO,eAAe,KAGxB,WAFc,eAAe,KACT,QAAQ,GAAW,MAAc,IAAI,GAAG,EAAE,GAC5C;MAElB,WAAU,GAAG,WAAW;AAE1B,SAAO;GAAC;GAAS;GAAG;GAAE;;CAGxB,YAAY,IAAI,gBAAgB;EAE9B,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,MAAM,GAAG,WAAW;EAC1B,MAAM,MAAM,GAAG,WAAW;EAC1B,IAAIA;AACJ,MAAI,OAAO,eAAe,KAGxB,WAFc,eAAe,KACT,QAAQ,GAAW,MAAc,IAAI,GAAG,EAAE,GAC5C;MAElB,WAAU,GAAG,WAAW;AAE1B,SAAO,mBAAmB;GACxB;GACA;GACA,aAAa,IAAI;GACjB;GACD,CAAC;;CAEL;AAID,SAAS,gBACP,IACA,gBACA,QACA,aACQ;CACR,MAAM,cAAc,GAAG,WAAW;CAClC,MAAM,MAAM,GAAG,WAAW;AAC1B,KAAI,OAAO,eAAe,KAExB,QADc,eAAe,KAAK,QAAQ,GAAW,MAAc,IAAI,GAAG,EAAE,GAC7D;AAEjB,QAAO,GAAG,WAAW;;AAGvB,MAAaC,oBAAgC;CAC3C,YAAY;CACZ,YAAY;CAGZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;AAGlC,SAAO;GAFO,gBAAgB,IAAI,gBAAgB,mBAAmB,WAAW,GAClE,gBAAgB,IAAI,gBAAgB,mBAAmB,WAAW;GACzD;GAAG;GAAE;;CAG9B,YAAY,IAAI,gBAAgB;EAC9B,MAAM,QAAQ,gBAAgB,IAAI,gBAAgB,mBAAmB,WAAW;EAChF,MAAM,QAAQ,gBAAgB,IAAI,gBAAgB,mBAAmB,WAAW;EAChF,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,MAAM,GAAG,WAAW;AAC1B,SAAO,mBAAmB;GAAC;GAAO;GAAO;GAAa,aAAa,IAAI;GAAC,CAAC;;CAE5E;AAID,MAAMC,gBAA4B;CAChC,YAAY;CACZ,YAAY;CAOZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAElC,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,MAAM,GAAG,WAAW;EAC1B,IAAIF;AACJ,MAAI,OAAO,eAAe,KAGxB,WAFc,eAAe,KACT,QAAQ,GAAW,MAAc,IAAI,GAAG,EAAE,GAC5C;MAElB,WAAU,GAAG,WAAW;AAE1B,SAAO;GAAC;GAAS;GAAG;GAAE;;CAGxB,YAAY,IAAI,gBAAgB;EAE9B,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,MAAM,GAAG,WAAW;EAC1B,MAAM,MAAM,GAAG,WAAW;EAC1B,IAAIA;AACJ,MAAI,OAAO,eAAe,KAGxB,WAFc,eAAe,KACT,QAAQ,GAAW,MAAc,IAAI,GAAG,EAAE,GAC5C;MAElB,WAAU,GAAG,WAAW;AAE1B,SAAO,mBAAmB;GACxB;GACA;GACA,aAAa,IAAI;GACjB;GACD,CAAC;;CAEL;AAID,MAAMG,gBAA4B;CAChC,YAAY;CACZ,YAAY;CAMZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,WAAW,eAAe,GAAG,QAAQ;AAE3C,SAAO;GAAC,MADQ,WAAW,SAAS,KAAM,GAAG,WAAW,WACjC,aAAa,IAAI;GAAE;GAAG;GAAE;;CAGjD,YAAY,IAAI,gBAAgB;EAE9B,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,WAAW,eAAe,GAAG,QAAQ;AAE3C,SAAO,mBAAmB,CADV,WAAW,SAAS,KAAM,GAAG,WAAW,SACpB,YAAY,CAAC;;CAEpD;AAID,MAAMC,oBAAgC;CACpC,YAAY;CACZ,YAAY;CAQZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,WAAW,eAAe,GAAG,QAAQ;AAE3C,SAAO;GAAC,MADQ,WAAW,SAAS,KAAM,GAAG,WAAW,WACjC,aAAa,IAAI;GAAE;GAAG;GAAE;;CAGjD,YAAY,IAAI,gBAAgB;EAE9B,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,aAAa,GAAG,WAAW;EACjC,MAAM,WAAW,eAAe,GAAG,QAAQ;AAE3C,SAAO,mBAAmB;GADV,WAAW,SAAS,KAAM,GAAG,WAAW;GACpB;GAAa;GAAY;GAAE,CAAC;;CAEnE;AAID,MAAMC,WAAuB;CAC3B,YAAY;CACZ,YAAY;CAKZ,UAAU;EAAC,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CAE/F,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,eAAe,GAAG,WAAW;EACnC,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,WAAY,GAAG,WAAW,YAAuB;EAGvD,MAAM,YAAa,GAAG,WAAW,aAAwB,WAAW;EAEpE,MAAM,SAAS,GAAG,WAAW;EAC7B,IAAIL;AACJ,MAAI,UAAU,eAAe,QAC3B,WAAU,eAAe,QAAQ;MAEjC,WAAU,GAAG,WAAW;EAG1B,MAAM,gBAAgB,UAAU,cAAc;EAC9C,MAAM,gBAAgB,UAAU,eAAe;AAC/C,SAAO;GAAC,KAAK,KAAK,IAAI,eAAe,cAAc,EAAE,IAAI;GAAE;GAAG;GAAE;;CAGlE,YAAY,IAAI,gBAAgB,SAAS;EAGvC,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,eAAe,GAAG,WAAW;EACnC,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,YAAY,GAAG,WAAW;EAChC,MAAM,WAAY,GAAG,WAAW,YAAuB;EAKvD,MAAM,YAAa,GAAG,WAAW,aAAwB,WAAW;EACpE,MAAM,aAAc,GAAG,WAAW,cAAyB;EAC3D,MAAM,oBAAqB,GAAG,WAAW,qBAAgC,WAAW;EAEpF,MAAM,kBAAkB,SAAS,UAAW,GAAG,WAAW,mBAA8B;EAExF,MAAM,SAAS,GAAG,WAAW;EAC7B,IAAIA;AACJ,MAAI,UAAU,eAAe,QAC3B,WAAU,eAAe,QAAQ;MAEjC,WAAU,GAAG,WAAW;AAG1B,SAAO,mBAAmB;GACxB;GACA;GACA;GACA;GACA,aAAa,UAAU;GACvB;GACA;GACA;GACA;GACA;GACA;GACA;GACD,CAAC;;CAEL;;;;;;AAOD,MAAaM,wBAAoC;CAC/C,GAAG;CACH,YAAY;CACb;AAID,MAAMC,YAAwB;CAC5B,YAAY;CACZ,YAAY;CAEZ,UAAU;EACR,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,WAAW;EACpB;CACD,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,eAAe,GAAG,WAAW;EAEnC,MAAM,OADW,GAAG,WAAW,WACP;EACxB,MAAM,SAAS,GAAG,WAAW;AAM7B,SAAO;GAAC,MAJN,UAAU,eAAe,UACrB,eAAe,QAAQ,KACtB,GAAG,WAAW,WACG,KAAK,IAAI,aAAa,aAAa,GAAG,MAC1C,IAAI;GAAE;GAAG;GAAE;;CAEjC,YAAY,IAAI,gBAAgB,SAAS;EACvC,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,eAAe,GAAG,WAAW;EACnC,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,SAAS,GAAG,WAAW;EAC7B,MAAM,UACJ,UAAU,eAAe,UACrB,eAAe,QAAQ,KACtB,GAAG,WAAW;AAGrB,SAAO,mBAAmB;GACxB;GACA;GACA;GACA;GACA;GANiB,YAAY,IAAK,SAAS,UAAU,IAAK;GAQ1D;GACA;GACD,CAAC;;CAEL;AAID,MAAMC,kBAA8B;CAClC,YAAY;CACZ,YAAY;CAEZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CACD,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,SAAS,GAAG,WAAW;EAE7B,MAAM,UAAU,eADD,GAAG,WAAW,gBACU;AACvC,SAAO;GAAC,KAAK,UAAU,QAAQ,IAAI;GAAE;GAAG;GAAE;;CAE5C,YAAY,IAAI,gBAAgB;EAC9B,MAAM,SAAS,GAAG,WAAW;EAE7B,MAAM,UAAU,eADD,GAAG,WAAW,gBACU;AACvC,SAAO,mBAAmB,CAAC,SAAS,OAAO,CAAC;;CAE/C;AAID,MAAMC,gBAA4B;CAChC,YAAY;CACZ,YAAY;CAOZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAElC,MAAM,cAAc,GAAG,WAAW;EAElC,MAAM,WAAW,eAAe,GAAG,QAAQ;AAE3C,SAAO;GADG,WAAW,SAAS,KAAM,GAAG,WAAW;GACvC;GAAa;GAAE;;CAG5B,YAAY,IAAI,gBAAgB,SAAS;EAEvC,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,eAAe,GAAG,WAAW;EACnC,MAAM,WAAW,GAAG,WAAW;EAE/B,MAAM,kBAAkB,SAAS,UAAW,GAAG,WAAW,mBAA8B;EAGxF,MAAM,WAAW,eAAe,GAAG,QAAQ;EAC3C,MAAM,IAAI,WAAW,SAAS,KAAM,GAAG,WAAW;EAGlD,MAAM,UAAU,eAAe,GAAG,OAAO;AAgBzC,SAAO,mBAAmB;GACxB;GAhBQ,UAAU,QAAQ,KAAK;GAkB/B;GACA;GACA;GACA;GAlBiB,GAAG,WAAW,WAAuB,QAAQ,IAAI;GAKnD,SAAS,WAAW;GAgBnC,aAXkB,GAAG,WAAW,cAAyB,IAAI,KAAK,KAAK,SAAS,CAWxD;GACzB,CAAC;;CAEL;AAOD,MAAMC,qBAAiC;CACrC,YAAY;CACZ,YAAY;CAOZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,cAAc,GAAG,WAAW;EAElC,MAAM,WAAW,eAAe,GAAG,QAAQ;AAE3C,SAAO;GADG,WAAW,SAAS,KAAM,GAAG,WAAW;GACvC;GAAa;GAAE;;CAG5B,YAAY,IAAI,gBAAgB;EAE9B,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,eAAe,GAAG,WAAW;EACnC,MAAM,WAAW,GAAG,WAAW;EAG/B,MAAM,WAAW,eAAe,GAAG,QAAQ;EAC3C,MAAM,IAAI,WAAW,SAAS,KAAM,GAAG,WAAW;EAGlD,MAAM,SAAS,eAAe,GAAG,OAAO;AAGxC,SAAO,mBAAmB;GAAC;GAFjB,SAAS,OAAO,KAAM,GAAG,WAAW;GAEb;GAAa;GAAc;GAAS,CAAC;;CAEzE;AAID,MAAaC,kCAA8C;CACzD,YAAY;CACZ,YAAY;CAKZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CAEzF,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,QAAQ,GAAG,WAAW;EAC5B,MAAM,SAAS,GAAG,WAAW;EAC7B,IAAIC;AACJ,MAAI,UAAU,eAAe,QAC3B,KAAI,eAAe,QAAQ;MAE3B,KAAI,GAAG,WAAW;AAGpB,SAAO;GAAC,KAAM,IAAI,QAAS,GAAG,IAAI;GAAE;GAAG;GAAE;;CAG3C,YAAY,IAAI,gBAAgB,SAAS;EACvC,MAAM,QAAQ,GAAG,WAAW;EAC5B,MAAM,SAAS,SAAS,UAAU;EAClC,MAAM,SAAS,GAAG,WAAW;EAC7B,IAAIA;AACJ,MAAI,UAAU,eAAe,QAC3B,KAAI,eAAe,QAAQ;MAE3B,KAAI,GAAG,WAAW;AAGpB,SAAO,mBAAmB,CAAC,IAAI,OAAO,SAAS,MAAM,CAAC;;CAEzD;AAID,MAAaC,4BAAwC;CACnD,YAAY;CACZ,YAAY;CAOZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,WAAW,eAAe,GAAG,QAAQ;AAE3C,SAAO;GADG,WAAW,SAAS,KAAM,GAAG,WAAW;GACvC;GAAa;GAAE;;CAG5B,YAAY,IAAI,gBAAgB,SAAS;EACvC,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,eAAe,GAAG,WAAW;EACnC,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,kBAAkB,SAAS,UAAW,GAAG,WAAW,mBAA8B;EAExF,MAAM,WAAW,eAAe,GAAG,QAAQ;EAC3C,MAAM,IAAI,WAAW,SAAS,KAAM,GAAG,WAAW;EAElD,MAAM,UAAU,eAAe,GAAG,OAAO;AAMzC,SAAO,mBAAmB;GACxB;GANQ,UAAU,QAAQ,KAAK;GAQ/B;GACA;GACA;GACA;GACA,aATkB,GAAG,WAAW,cAAyB,IAAI,KAAK,KAAK,SAAS,CASxD;GACzB,CAAC;;CAEL;AAID,MAAMC,cAA0B;CAC9B,YAAY;CACZ,YAAY;CAKZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CAEzF,gBAAgB,IAAI;AAGlB,SAAO;GADU,GAAG,WAAW;GACb;GAAG;GAAE;;CAGzB,YAAY,IAAI;EAEd,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,WAAW,GAAG,WAAW;AAC/B,SAAO,mBAAmB,CAAC,UAAU,SAAS,CAAC;;CAElD;AAID,MAAMC,mBAA+B;CACnC,YAAY;CACZ,YAAY;CAOZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,MAAM,GAAG,WAAW;EAC1B,IAAIf;AACJ,MAAI,OAAO,eAAe,KACxB,WAAU,eAAe,KAAK;MAE9B,WAAU,GAAG,WAAW;AAE1B,SAAO;GAAC,KAAK,UAAU,UAAU,IAAI;GAAE;GAAG;GAAE;;CAG9C,YAAY,IAAI,gBAAgB;EAE9B,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,MAAM,GAAG,WAAW;EAC1B,IAAIA;AACJ,MAAI,OAAO,eAAe,KACxB,WAAU,eAAe,KAAK;MAE9B,WAAU,GAAG,WAAW;AAE1B,SAAO,mBAAmB;GAAC;GAAS;GAAU;GAAa;GAAE,CAAC;;CAEjE;AAID,MAAMgB,uBAAmC;CACvC,YAAY;CACZ,YAAY;CAEZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,MAAM,GAAG,WAAW;EAC1B,IAAIhB;AACJ,MAAI,OAAO,eAAe,KACxB,WAAU,eAAe,KAAK;MAE9B,WAAU,GAAG,WAAW;AAE1B,SAAO;GAAC,KAAK,UAAU,UAAU,IAAI;GAAE;GAAG;GAAE;;CAG9C,YAAY,IAAI,gBAAgB;EAC9B,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,MAAM,GAAG,WAAW;EAC1B,IAAIA;AACJ,MAAI,OAAO,eAAe,KACxB,WAAU,eAAe,KAAK;MAE9B,WAAU,GAAG,WAAW;AAE1B,SAAO,mBAAmB;GAAC;GAAS;GAAU;GAAa;GAAE,CAAC;;CAEjE;AAID,MAAMiB,wBAAoC;CACxC,YAAY;CACZ,YAAY;CAIZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,MAAM,GAAG,WAAW;EAC1B,IAAIjB;AACJ,MAAI,OAAO,eAAe,KACxB,WAAU,eAAe,KAAK;MAE9B,WAAU,GAAG,WAAW;AAE1B,SAAO;GAAC,KAAK,UAAU,UAAU,IAAI;GAAE;GAAG;GAAE;;CAG9C,YAAY,IAAI,gBAAgB;EAC9B,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,MAAM,GAAG,WAAW;EAC1B,IAAIA;AACJ,MAAI,OAAO,eAAe,KACxB,WAAU,eAAe,KAAK;MAE9B,WAAU,GAAG,WAAW;AAE1B,SAAO,mBAAmB;GAAC;GAAS;GAAU;GAAa;GAAE,CAAC;;CAEjE;AAID,MAAMkB,kBAA8B;CAClC,YAAY;CACZ,YAAY;CAMZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAGjC,YAAY,IAAI,gBAAgB;AAG9B,SAAO,mBAAmB,CAFZ,mBAAmB,IAAI,eAAe,CAEnB,CAAC;;CAErC;AAID,MAAMC,eAA2B;CAC/B,YAAY;CACZ,YAAY;CAUZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI;AAElB,SAAO;GADW,GAAG,WAAW;GACb;GAAG;GAAE;;CAG1B,YAAY,IAAI,gBAAgB;EAC9B,MAAM,YAAY,GAAG,WAAW;EAChC,MAAM,UAAU,GAAG,WAAW;EAC9B,MAAM,UAAU,GAAG,WAAW;EAC9B,MAAM,UAAU,GAAG,WAAW;EAE9B,MAAM,SAAS,GAAG,WAAW;EAC7B,IAAIP;AACJ,MAAI,UAAU,eAAe,QAC3B,KAAI,eAAe,QAAQ;MAE3B,KAAI,GAAG,WAAW;AAEpB,SAAO,mBAAmB;GAAC;GAAG;GAAW;GAAS;GAAS;GAAS;GAAG;GAAG;GAAE,CAAC;;CAEhF;AAID,MAAaQ,qBAAiC;CAC5C,GAAG;CACH,YAAY;CACb;AAID,MAAMC,oBAAgC;CACpC,YAAY;CACZ,YAAY;CAKZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CAEzF,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,QAAQ,GAAG,WAAW;EAC5B,MAAM,SAAS,GAAG,WAAW;EAC7B,IAAIT;AACJ,MAAI,UAAU,eAAe,QAC3B,KAAI,eAAe,QAAQ;MAE3B,KAAI,GAAG,WAAW;AAEpB,SAAO;GAAC,KAAK,IAAI,OAAO,IAAI;GAAE;GAAG;GAAE;;CAGrC,YAAY,IAAI,gBAAgB,SAAS;EACvC,MAAM,QAAQ,GAAG,WAAW;EAC5B,MAAM,SAAS,SAAS,UAAU;EAClC,MAAM,SAAS,GAAG,WAAW;EAC7B,IAAIA;AACJ,MAAI,UAAU,eAAe,QAC3B,KAAI,eAAe,QAAQ;MAE3B,KAAI,GAAG,WAAW;AAEpB,SAAO,mBAAmB,CAAC,IAAI,OAAO,SAAS,MAAM,CAAC;;CAEzD;AAID,MAAaU,4BAAwC;CACnD,YAAY;CACZ,YAAY;CAGZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,QAAQ,GAAG,WAAW;EAC5B,MAAM,SAAS,GAAG,WAAW;AAG7B,SAAO;GAAC,MADN,UAAU,eAAe,UAAU,eAAe,QAAQ,KAAM,GAAG,WAAW,KAC/D,OAAO,IAAI;GAAE;GAAG;GAAE;;CAGrC,YAAY,IAAI,gBAAgB,SAAS;EACvC,MAAM,QAAQ,GAAG,WAAW;EAC5B,MAAM,SAAS,SAAS,UAAU;EAClC,MAAM,SAAS,GAAG,WAAW;AAG7B,SAAO,mBAAmB,EADxB,UAAU,eAAe,UAAU,eAAe,QAAQ,KAAM,GAAG,WAAW,KACjD,OAAO,SAAS,MAAM,CAAC;;CAEzD;AAID,MAAaC,gCAA4C;CACvD,GAAG;CACH,YAAY;CACb;AAID,MAAaC,uCAAmD;CAC9D,GAAG;CACH,YAAY;CAEZ,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,QAAQ,GAAG,WAAW;EAC5B,MAAM,SAAS,GAAG,WAAW;AAG7B,SAAO;GAAC,MADN,UAAU,eAAe,UAAU,eAAe,QAAQ,KAAM,GAAG,WAAW,KAC9D,SAAU,GAAG,IAAI;GAAE;GAAG;GAAE;;CAE7C;AAID,MAAaC,2BAAuC;CAClD,YAAY;CACZ,YAAY;CAKZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CAEzF,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,QAAQ,GAAG,WAAW;EAC5B,MAAM,SAAS,GAAG,WAAW;EAC7B,IAAIb;AACJ,MAAI,UAAU,eAAe,QAC3B,KAAI,eAAe,QAAQ;MAE3B,KAAI,GAAG,WAAW;AAEpB,SAAO;GAAC,KAAK,IAAI,OAAO,IAAI;GAAE;GAAG;GAAE;;CAGrC,YAAY,IAAI,gBAAgB,SAAS;EACvC,MAAM,QAAQ,GAAG,WAAW;EAC5B,MAAM,SAAS,SAAS,UAAU;EAClC,MAAM,SAAS,GAAG,WAAW;EAC7B,IAAIA;AACJ,MAAI,UAAU,eAAe,QAC3B,KAAI,eAAe,QAAQ;MAE3B,KAAI,GAAG,WAAW;AAEpB,SAAO,mBAAmB,CAAC,IAAI,OAAO,SAAS,MAAM,CAAC;;CAEzD;AAID,MAAac,qBAAiC;CAC5C,YAAY;CACZ,YAAY;CAOZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,WAAW,eAAe,GAAG,QAAQ;AAE3C,SAAO;GADG,WAAW,SAAS,KAAM,GAAG,WAAW;GACvC;GAAa;GAAE;;CAG5B,YAAY,IAAI,gBAAgB,SAAS;EACvC,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,eAAe,GAAG,WAAW;EACnC,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,kBAAkB,SAAS,UAAW,GAAG,WAAW,mBAA8B;EAExF,MAAM,WAAW,eAAe,GAAG,QAAQ;EAC3C,MAAM,IAAI,WAAW,SAAS,KAAM,GAAG,WAAW;EAElD,MAAM,UAAU,eAAe,GAAG,OAAO;AAMzC,SAAO,mBAAmB;GACxB;GANQ,UAAU,QAAQ,KAAK;GAQ/B;GACA;GACA;GACA;GACA,aATkB,GAAG,WAAW,cAAyB,IAAI,KAAK,KAAK,SAAS,CASxD;GACzB,CAAC;;CAEL;AAID,MAAMC,sBAAkC;CACtC,YAAY;CACZ,YAAY;CAKZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CAEzF,gBAAgB,IAAI;EAClB,MAAM,WAAW,GAAG,WAAW;AAC/B,SAAO;GAAC,KAAK,UAAU,IAAI;GAAE;GAAG;GAAE;;CAGpC,YAAY,IAAI,gBAAgB;EAC9B,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,aAAa,GAAG,WAAW;EACjC,MAAM,SAAS,GAAG,WAAW;EAC7B,IAAI3B;AACJ,MAAI,UAAU,eAAe,QAC3B,WAAU,eAAe,QAAQ;MAEjC,WAAU,GAAG,WAAW;AAE1B,SAAO,mBAAmB;GAAC;GAAS;GAAU;GAAY;GAAE,CAAC;;CAEhE;;;;;;;;AAwBD,MAAM,mBAAmB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqCzB,MAAM,wBAAwB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA0C9B,MAAM,eAAe;;;;;;;;;;;;;;;;;;;AAsBrB,MAAM,YAAY;;;;;;;;;;;;;;;;AAiBlB,MAAM,iBAAiB;;;;;;;;;;;;;;;;;;;;AAqBvB,MAAM,iBAAiB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA4DvB,MAAM,yBAAyB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA+E/B,MAAa4B,yBAAqC;CAChD,YAAY;CACZ,YAAY;CACZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CACD,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,MAAM,GAAG,WAAW;EAC1B,IAAI5B;AACJ,MAAI,OAAO,eAAe,KAExB,WADc,eAAe,KACb,QAAQ,GAAG,MAAM,IAAI,GAAG,EAAE,GAAG;MAE7C,WAAU,GAAG,WAAW;AAE1B,SAAO;GAAC;GAAS;GAAG;GAAE;;CAExB,YAAY,IAAI,gBAAgB;EAC9B,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,MAAM,GAAG,WAAW;EAC1B,MAAM,MAAM,GAAG,WAAW;EAC1B,IAAIA;AACJ,MAAI,OAAO,eAAe,KAExB,WADc,eAAe,KACb,QAAQ,GAAG,MAAM,IAAI,GAAG,EAAE,GAAG;MAE7C,WAAU,GAAG,WAAW;AAE1B,SAAO,mBAAmB;GAAC;GAAS;GAAa,aAAa,IAAI;GAAE;GAAE,CAAC;;CAE1E;AAED,MAAM6B,WAAuB;CAC3B,YAAY;CACZ,YAAY;CACZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CACzF,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAEjC,YAAY,IAAI,gBAAgB;AAE9B,SAAO,mBAAmB,CADZ,mBAAmB,IAAI,eAAe,CACnB,CAAC;;CAErC;AAED,MAAMC,gBAA4B;CAChC,YAAY;CACZ,YAAY;CACZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CACzF,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAEjC,YAAY,IAAI,gBAAgB;EAE9B,MAAM,UAAU,eAAe,GAAG,OAAO;AAGzC,SAAO,mBAAmB,CAFb,UAAU,QAAQ,KAAM,GAAG,WAAW,MACtC,UAAU,QAAQ,KAAM,GAAG,WAAW,KACb,CAAC;;CAE1C;AAED,MAAMC,gBAA4B;CAChC,YAAY;CACZ,YAAY;CACZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,kBAAkB;AAChB,SAAO;GAAC;GAAG;GAAG;GAAE;;CAElB,YAAY,IAAI,gBAAgB;EAC9B,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,UAAU,eAAe,GAAG,OAAO;AAGzC,SAAO,mBAAmB;GAAC;GAFZ,UAAU,QAAQ,KAAM,GAAG,WAAW;GAER,aADhC,GAAG,WAAW,OAAkB,KACiB;GAAE;GAAE,CAAC;;CAEtE;AAED,MAAMC,iBAA6B;CACjC,YAAY;CACZ,YAAY;CACZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CACD,gBAAgB,IAAI;EAClB,MAAM,OAAO,GAAG,WAAW;EAC3B,MAAM,OAAO,GAAG,WAAW;AAC3B,SAAO;GAAC,KAAK,OAAO,MAAM,GAAG;GAAE;GAAG;GAAE;;CAEtC,YAAY,IAAI;AACd,SAAO,mBAAmB;GACxB,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACf,CAAC;;CAEL;AAED,MAAMC,sBAAkC;CACtC,YAAY;CACZ,YAAY;CACZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CACD,gBAAgB,IAAI;EAClB,MAAM,OAAO,GAAG,WAAW;EAC3B,MAAM,OAAO,GAAG,WAAW;AAC3B,SAAO;GAAC,KAAK,OAAO,MAAM,GAAG;GAAE;GAAG;GAAE;;CAEtC,YAAY,IAAI;AACd,SAAO,mBAAmB;GACxB,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACf,CAAC;;CAEL;AAED,MAAMC,cAA0B;CAC9B,YAAY;CACZ,YAAY;CACZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CACD,gBAAgB,IAAI;EAClB,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,IAAI,GAAG,WAAW;AACxB,SAAO;GAAC,KAAK,IAAI,GAAG,GAAG;GAAE;GAAG;GAAE;;CAEhC,YAAY,IAAI;AACd,SAAO,mBAAmB,CAAC,GAAG,WAAW,GAAa,GAAG,WAAW,EAAY,CAAC;;CAEpF;AAiBD,MAAM,mBAAmB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAgCzB,MAAM,oBAAoB;;;;;;;;;;;;;;;;;;;;;;;AAwB1B,MAAM,2BAA2B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAwCjC,MAAMC,iBAA6B;CACjC,YAAY;CACZ,YAAY;CACZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CACzF,gBAAgB,IAAI;EAClB,MAAM,SAAS,GAAG,WAAW;EAC7B,MAAM,OAAO,GAAG,WAAW;EAC3B,MAAM,IAAI,GAAG,WAAW;AACxB,SAAO;GAAC,KAAK,SAAS,OAAO,GAAG,GAAG;GAAE;GAAG;GAAE;;CAE5C,YAAY,IAAI;EACd,MAAM,SAAS,GAAG,WAAW;EAC7B,MAAM,OAAO,GAAG,WAAW;AAC3B,SAAO,mBAAmB;GACxB,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACd;GACA,OAAO;GACP,OAAO;GACP,OAAO;GACP,OAAO;GACP,KAAK;GACL,KAAK;GACL,KAAK;GACL,KAAK;GACN,CAAC;;CAEL;AAED,MAAMC,kBAA8B;CAClC,YAAY;CACZ,YAAY;CACZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CACD,gBAAgB,IAAI;EAClB,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,IAAI,GAAG,WAAW;AACxB,SAAO;GAAC,KAAK,IAAI,GAAG,GAAG;GAAE;GAAG;GAAE;;CAEhC,YAAY,IAAI;EACd,MAAM,IAAI,GAAG,WAAW;AACxB,SAAO,mBAAmB;GAAC;GAAG,GAAG,WAAW;GAAa,KAAK,MAAM,IAAI,EAAE;GAAE;GAAE,CAAC;;CAElF;AAED,MAAMC,+BAA2C;CAC/C,YAAY;CACZ,YAAY;CACZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CACD,gBAAgB,IAAI;EAClB,MAAM,OAAO,GAAG,WAAW;EAC3B,MAAM,OAAO,GAAG,WAAW;AAC3B,SAAO;GAAC,KAAK,OAAO,MAAM,GAAG;GAAE;GAAG;GAAE;;CAEtC,YAAY,IAAI;EACd,MAAM,MAAM,GAAG,WAAW;EAC1B,MAAM,OAAO,GAAG,WAAW;AAC3B,SAAO,mBAAmB;GACxB;GACA;GACA,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACd,KAAK,MAAM,MAAM,KAAK;GACtB;GACD,CAAC;;CAEL;AAED,MAAaC,kBAAuD;CAClE,WAAW;CACX,eAAe;CACf,QAAQ;CACR,YAAY;CACZ,YAAY;CACZ,KAAK;CACL,KAAK;CACL,QAAQ;CACR,MAAM;CACN,iBAAiB;CACjB,MAAM;CACN,SAAS;CACT,SAAS;CACT,gBAAgB;CAChB,OAAO;CACP,aAAa;CACb,WAAW;CACX,SAAS;CACT,YAAY;CACZ,eAAe;CACf,SAAS;CACT,WAAW;CACX,MAAM;CACN,WAAW;CACX,gBAAgB;CAChB,SAAS;CACT,cAAc;CACd,kBAAkB;CAClB,mBAAmB;CACnB,aAAa;CACb,UAAU;CACV,eAAe;CACf,iBAAiB;CACjB,cAAc;CACd,UAAU;CACV,OAAO;CACP,SAAS;CACT,QAAQ;CAER,YAAY;CACZ,iBAAiB;CACjB,SAAS;CAET,YAAY;CACZ,aAAa;CACb,0BAA0B;CAE1B,MAAM;CACN,WAAW;CACX,WAAW;CACZ;;;;ACtjOD,MAAMC,kBAAgB;AA8CtB,IAAa,WAAb,MAAa,SAAS;CACpB,AAAQ;CACR,AAAQ;CAER,AAAQ,gCAAwC,IAAI,KAAK;CACzD,AAAQ,oCAA4C,IAAI,KAAK;CAC7D,AAAQ,kCAA0C,IAAI,KAAK;CAC3D,AAAQ,iCAAyC,IAAI,KAAK;;CAG1D,AAAQ;;;;;;;CAQR,AAAQ,YAcG;;CAEX,AAAQ,aAAkC;;;;;;;CAO1C,AAAQ,WAAiC;;;;;;;;;CAUzC,AAAQ,uBAAyC;;CAEjD,AAAQ;;CAER,AAAQ;;CAER,AAAQ;;CAER,AAAQ,kBAA+B,EAAE;;;;;;;;;CAUzC,AAAQ,uBAAyC;CACjD,AAAQ,yBAAiC;;CAGzC,AAAQ,kBAAmC,EAAE;;CAE7C,AAAQ,gBAAiC,EAAE;;CAE3C,AAAQ,cAIG;;CAGX,AAAS;CAET,AAAQ;CACR,AAAQ;CACR,AAAQ,SAAiB;CACzB,AAAQ;CAKR,AAAQ,iBAAiB;CACzB,AAAiB,8BAAc,IAAI,KAA4C;CAC/E,AAAQ,WAA+B;CACvC,AAAQ,kBAAoC;CAC5C,AAAQ,mBAAqC;CAE7C,YAAY,KAAiB,OAAmB,SAA0B;AACxE,OAAK,MAAM;AACX,OAAK,QAAQ;AACb,OAAK,YAAY,QAAQ;AACzB,OAAK,SAAS,QAAQ,UAAU;AAChC,OAAK,kBAAkB,KAAK,IAAI,GAAG,QAAQ,mBAAmB,EAAE;AAChE,OAAK,iBACH,IAAI,gBAAgB,OAAO,YAAY,eAAe,QAAQ,KAAK,kBAAkB;AAKvF,OAAK,oBAAoB,IAAI;AAG7B,OAAK,iBAAiB,oBAAoB,KAAK,aAAa,QAAQ,YAAY,EAAE;AAElF,OAAK,2BAA2B;AAChC,OAAK,yBAAyB;AAC9B,OAAK,wBAAwB;AAI7B,OAAK,iBAAiB,qBACpB,KACA,mBACA,KAAK,IAAI,GAAG,MAAM,OAAO,aAAa,EAAE,CACzC;AACD,OAAK,qBAAqB,oBAAoB,KAAK,iBAAiB,EAAE;AACtE,OAAK,iBAAiB,qBAAqB,KAAK,mBAAmB,EAAE;;;;;;;;;CAUvE,aAAa,KAcJ;AACP,OAAK,YAAY;;;;;;;CAQnB,AAAQ,kBAAiC;EACvC,MAAM,MAAM,KAAK;AACjB,MAAI,CAAC,OAAO,CAAC,IAAI,MAAO,QAAO,QAAQ,SAAS;AAChD,MAAI,KAAK,SAAU,QAAO,KAAK;EAC/B,MAAM,EAAE,WAAW,WAAW,WAAW,aAAa,IAAI;AAC1D,OAAK,YAAY,YAAY;GAC3B,MAAM,QAAQ,MAAM,OAAO,KAAK,UAAU;GAC1C,MAAM,OAAO,OAAO,QAAsC;IACxD,MAAM,OAAO,MAAM,MAAM,MAAM,IAAI,QAAQ,IAAI,CAAC;AAChD,QAAI,CAAC,KAAM,OAAM,IAAI,MAAM,4BAA4B,MAAM;AAC7D,WAAO,KAAK,aAAa;;GAE3B,MAAM,CAAC,MAAM,MAAM,QAAQ,MAAM,QAAQ,IAAI;IAC3C,KAAK,UAAU;IACf,KAAK,UAAU;IACf,KAAK,SAAS;IACf,CAAC;AACF,OAAI,SAAS,IAAI,YAAY,KAAK;AAClC,OAAI,SAAS,IAAI,aAAa,KAAK;AACnC,OAAI,QAAQ,IAAI,aAAa,KAAK;MAChC;AACJ,SAAO,KAAK;;;;;;;CAQd,MAAc,cAAc,UAAsC;EAChE,MAAM,MAAM,KAAK;AACjB,MAAI,CAAC,IAAK;AACV,MAAI,IAAI,MAAO,OAAM,KAAK,iBAAiB;EAC3C,MAAM,IAAI,SAAS;EACnB,MAAM,EAAE,QAAQ,QAAQ,OAAO,OAAO,cAAc;EACpD,MAAM,SAAS,IAAI;AACnB,MAAI,CAAC,KAAK,cAAc,KAAK,WAAW,SAAS,OAC/C,MAAK,aAAa,IAAI,aAAa,OAAO;EAE5C,MAAM,MAAM,KAAK;AACjB,OAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;GAE1B,MAAM,UADU,SAAS,KACC;GAC1B,MAAM,UAAU,IAAI;AACpB,QAAK,IAAI,IAAI,GAAG,IAAI,OAAO,KAAK;IAC9B,MAAM,OAAO,UAAU;IACvB,MAAM,YAAY,SAAS;IAC3B,MAAM,aAAa,OAAO,KAAK;IAC/B,MAAM,SAAU,OAAO,eAAe,YAAa;IACnD,MAAM,WAAY,OAAO,YAAa;AACtC,QAAI,UAAU,MAAM,SAAS,MAAM,aAAa,OAAO;;;EAG3D,MAAM,SAAS,KAAK,UAAU,IAAI,aAAa;AAC/C,MAAI,CAAC,OACH,OAAM,IAAI,MAAM,sBAAsB,IAAI,aAAa,qBAAqB;AAE9E,OAAK,IAAI,OAAO,MAAM,YAAY,QAAQ,GAAG,IAAI,QAAQ,IAAI,YAAY,SAAS,EAAE;;;;;;;;;;;;;;;;;CAkBtF,MAAM,cACJ,QACe;AACf,MAAI,kBAAkB,KAAK;AACzB,QAAK,iBAAiB,OAAO;AAC7B;;AAEF,OAAK,MAAM,QAAQ,OAAO,MAAM,EAAE;GAChC,MAAM,QAAQ,MAAM,OAAO,IAAI,KAAK;AACpC,OAAI,CAAC,MAAO;GACZ,MAAM,EAAE,SAAS;GACjB,MAAM,SAAS,oBAAoB,KAAK,KAAK,UAAU,QAAQ,KAAK,YAAY,KAAK;AACrF,QAAK,cAAc,IAAI,MAAM,OAAO;;;;;;;;CAWxC,iBAAiB,SAAwE;EACvF,MAAM,QAAQ,CAAC,GAAG,QAAQ,MAAM,CAAC;AACjC,OAAK,MAAM,QAAQ,OAAO;GACxB,MAAM,QAAQ,QAAQ,IAAI,KAAK;AAC/B,OAAI,CAAC,MAAO;GACZ,MAAM,SAAS,oBACb,KAAK,KACL,UAAU,QACV,MAAM,KAAK,YACX,MAAM,KACP;AACD,QAAK,cAAc,IAAI,MAAM,OAAO;AACpC,WAAQ,OAAO,KAAK;;;;;;;;;;CAWxB,iBAAuB;AACrB,OAAK,MAAM,UAAU,KAAK,MAAM,gBAAgB;GAC9C,MAAM,OAAO,KAAK,MAAM,MAAM,MAAM,MAAM,EAAE,OAAO,OAAO;GAC1D,IAAI,OAAO,gBAAgB,KAAK;AAChC,OAAI,CAAC,KAAM,OAAM,IAAI,MAAM,0BAA0B,KAAK,SAAS;AAInE,OAAI,KAAK,WAAW,iBAElB;QADiB,KAAK,QAAQ,MAAM,QAAQ,KAAK,MAAM,QAAQ,MAAM,UAAU,MAAM,CAEnF,QACE,KAAK,WAAW,eACZ,kCACA;cAEC,KAAK,WAAW,aAEzB;QADiB,KAAK,OAAO,MAAM,QAAQ,KAAK,MAAM,QAAQ,MAAM,UAAU,MAAM,CAElF,QAAO,KAAK,WAAW,eAAe,4BAA4B;cAE3D,KAAK,WAAW,cAAc,KAAK,IAAI,UAAU,CAAC,KAAK,IAAI,eAGpE,QAAO;YACE,KAAK,WAAW,eAAe,KAAK,OAAO,WAAW,EAI/D,QAAO;YACE,KAAK,WAAW,UAAU,KAAK,WAAW,gBAAgB,KAGnE,QAAO;GAGT,MAAM,WAAW,oBACf,KAAK,KACL,UAAU,UACV,KAAK,YACL,KAAK,WACN;GAGD,MAAM,cAAc,KAAK,cAAc,EAAE;GACzC,MAAM,aAAa,KAAK,YAAY,MAAM,aAAa,EAAE,QAAQ,GAAG,CAAC;GACrE,MAAM,gBAAgB,oBAAoB,KAAK,KAAK,WAAW,UAAU,WAAW;GAGpF,MAAM,gBAAgB,KAAK,cAAc,MAAM,MAAM,cAAc;GACnE,MAAM,YAAY,gBAAgB,KAAK,KAAK,UAAU,eAAe,MAAM,SAAS;GAEpF,MAAMC,eAA8B;IAClC;IACA;IACA;IACA;IACA;IACA;IACA,iBAAiB;IACjB,kBAAkB;IACnB;AACD,QAAK,gBAAgB,KAAK,aAAa;AAIvC,OADiB,KAAK,WAAW,YAAY,KAAK,WAAW,cAC/C;IAYZ,MAAM,eAFH,WAA0E,SAAS,KAChF,4BAA4B,OACK,KAAK,IAAI,gBAAgB,CAAC,KAAK,IAAI;IAC1E,MAAM,SACJ,KAAK,WAAW,eACZ,eACE,6BACA,mBACF,eACE,wBACA;IAER,MAAM,aAAa,oBACjB,KAAK,KACL,UAAU,UACV,OAAO,YACP,OAAO,WACR;IACD,MAAM,gBAAgB,OAAO,YAAY,MAAM,aAAa,EAAE,QAAQ,GAAG,CAAC;IAC1E,MAAM,YAAY,oBAAoB,KAAK,KAAK,cAAc,UAAU,cAAc;IACtF,MAAM,YAAY,KAAK,cAAc,QAAQ,MAAM,UAAU;IAC7D,MAAM,cAAc,gBAAgB,KAAK,KAAK,YAAY,WAAW,SAAS,SAAS;AAEvF,SAAK,cAAc,KAAK;KACtB;KACA;KACA,MAAM;KACN,UAAU;KACV,WAAW;KACX,eAAe;KACf,iBAAiB;KACjB,kBAAkB;KACnB,CAAC;SAGF,MAAK,cAAc,KAAK,aAAa;;AAOzC,OAAK,yBAAyB;AAM9B,OAAK,6BAA6B;AAGlC,OAAK,8BAA8B;AAGnC,OAAK,6BAA6B;AAGlC,OAAK,kCAAkC;AAGvC,OAAK,8BAA8B;AAKnC,UAAQ,IAAI,sBAAsB,KAAK,cAAc,OAAO,mBAAmB;EAG/E,MAAM,eAAe,KAAK,UAAU,SAAS;AAC7C,MAAI,cAAc;GAChB,MAAM,iBAAiB,oBACrB,KAAK,KACL,UACA,YAAY,YACZ,YAAY,WACb;GACD,MAAM,gBAAgB,oBACpB,KAAK,KACL,kCACA,IAAI,YAAY,GAAG,CACpB;AAOD,QAAK,cAAc;IACjB,UAAU;IACV,WARsB,gBACtB,KAAK,KACL,gBACA;KAAC,EAAE,QAAQ,cAAc;KAAE,EAAE,QAAQ,KAAK,oBAAoB;KAAE,EAAE,QAAQ,eAAe;KAAC,EAC1F,YACD;IAIC,eAAe;IAChB;;AAGH,MAAI,KAAK,kBACP,SAAQ,IAAI,gEAAgE;EAK9E,IAAI,eAAe;AACnB,OAAK,MAAM,SAAS,KAAK,gBACvB,iBAAgB,KAAK,KAAK,MAAM,cAAc,OAAO,EAAE,GAAG;EAE5D,IAAI,cAAc;AAClB,OAAK,MAAM,SAAS,KAAK,cACvB,gBAAe,KAAK,KAAK,MAAM,cAAc,OAAO,EAAE,GAAG;AAE3D,iBAAe;EAEf,MAAM,eAAe,KAAK,YAAY;AACtC,OAAK,yBAAyB,KAAK,IAAI,cAAc,YAAY,GAAG;AACpE,OAAK,uBAAuB,KAAK,IAAI,OAAO,aAAa;GACvD,OAAO;GACP,MAAM,KAAK;GACX,OAAO;GACR,CAAC;;;;;CAMJ,MAAM,QAAQ,UAA+C;EAC3D,MAAM,IAAI,SAAS;EAEnB,MAAM,iBAAiB,KAAK,cAAc,EAAE;EAC5C,MAAMC,iBAAiC,EAAE,QAAQ,KAAK,QAAQ;EAG9D,MAAM,UAAU,MAAM,IAAI,KAAK,gBAAgB,KAAK;AAGpD,OAAK,IAAI,OAAO,MAAM,YAAY,KAAK,gBAAgB,GAAG,SAAyB;AAGnF,QAAM,KAAK,cAAc,SAAS;EAIlC,MAAMC,gBAAiD,IAAI,MAAM,QAAQ,OAAO;AAChF,OAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,QAAQ,KAAK;GACvC,MAAM,QAAQ,QAAQ;GACtB,MAAM,aAAa,MAAM,KAAK,YAAY,MAAM,MAAM,gBAAgB,eAAe;GACrF,MAAM,aAAa,IAAI,WAAW,WAAW;GAE7C,IAAI,UAAU,CAAC,MAAM,mBAAmB,MAAM,gBAAgB,WAAW,WAAW;AACpF,OAAI,CAAC,SAAS;IACZ,MAAM,SAAS,MAAM;AACrB,SAAK,IAAI,IAAI,GAAG,IAAI,WAAW,QAAQ,IACrC,KAAI,WAAW,OAAO,OAAO,IAAI;AAC/B,eAAU;AACV;;;AAKN,OAAI,SAAS;AACX,SAAK,IAAI,OAAO,MAAM,YAAY,MAAM,eAAe,GAAG,WAAW;AACrE,UAAM,kBAAkB,IAAI,WAAW,WAAW,MAAM,EAAE,CAAC;;AAG7D,iBAAc,KAAK,MAAM,KAAK,gBAAgB,MAAM,MAAM,gBAAgB,eAAe;;AAG3F,MAAI,KAAK,eACP,OAAM,KAAK,sBAAsB,SAAS,cAAc;WAC/C,KAAK,mBAAmB;GASjC,MAAM,QAAQ,KAAK;AACnB,QAAK,IAAI,QAAQ,GAAG,QAAQ,QAAQ,QAAQ,SAAS,OAAO;IAC1D,MAAM,MAAM,KAAK,IAAI,OAAO,sBAAsB;IAClD,MAAM,MAAM,KAAK,IAAI,QAAQ,OAAO,QAAQ,OAAO;AACnD,SAAK,IAAI,IAAI,OAAO,IAAI,KAAK,KAAK;KAChC,MAAM,IAAI,IAAI,kBAAkB;AAChC,OAAE,YAAY,QAAQ,GAAG,SAAS;AAClC,OAAE,aAAa,GAAG,QAAQ,GAAG,UAAU;AACvC,OAAE,mBAAmB,GAAG,cAAc,GAAG;AACzC,OAAE,KAAK;;AAET,SAAK,IAAI,OAAO,MAAM,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;AAC5C,UAAM,KAAK,IAAI,OAAO,MAAM,qBAAqB;;GAEnD,MAAM,eAAe,KAAK,UAAU,SAAS;AAC7C,OAAI,cAAc;IAChB,MAAM,UAAU,KAAK,IAAI,OAAO,sBAAsB;IACtD,MAAMC,cAAY,KAAK,MAAM,OAAO;AAEpC,YAAQ,mBAAmB,cAAc,GAAG,KAAK,gBAAgB,GAAGA,cAAY,EAAE;AAClF,SAAK,IAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;;SAE7C;GAEL,MAAM,UAAU,KAAK,IAAI,OAAO,qBAAqB,EAAE,OAAO,WAAW,CAAC;GAC1E,MAAM,OAAO,QAAQ,iBAAiB,EAAE,OAAO,YAAY,CAAC;AAC5D,QAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,QAAQ,KAAK;AACvC,SAAK,YAAY,QAAQ,GAAG,SAAS;AACrC,SAAK,aAAa,GAAG,QAAQ,GAAG,UAAU;AAC1C,SAAK,mBAAmB,GAAG,cAAc,GAAG;;AAE9C,QAAK,KAAK;GACV,MAAM,eAAe,KAAK,UAAU,SAAS;AAC7C,OAAI,cAAc;IAChB,MAAMA,cAAY,KAAK,MAAM,OAAO;AAEpC,YAAQ,mBAAmB,cAAc,GAAG,KAAK,gBAAgB,GAAGA,cAAY,EAAE;;AAEpF,QAAK,IAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;;EAGlD,MAAM,YAAY,KAAK,MAAM,OAAO;AACpC,QAAM,KAAK,eAAe,SAASJ,iBAAe,GAAG,YAAY,EAAE;EACnE,MAAM,SAAS,KAAK,eAAe,eAAe,GAAG,YAAY,EAAE;EACnE,MAAM,SAAS,IAAI,aAAa,OAAO,MAAM,EAAE,CAAC;AAChD,OAAK,eAAe,OAAO;AAE3B,OAAK,UAAU;AACf,SAAO,EAAE,QAAQ;;;;;;;;CASnB,MAAc,sBACZ,SACA,eACA,aAAa,MACE;EACf,MAAM,IAAI,QAAQ;EAClB,MAAM,aAAa,IAAI;AACvB,MAAI,CAAC,KAAK,UAAU;AAClB,QAAK,WAAW,KAAK,IAAI,OAAO,eAAe;IAAE,MAAM;IAAa,OAAO;IAAY,CAAC;AAExF,QAAK,kBAAkB,KAAK,IAAI,OAAO,aAAa;IAClD,MAAM,aAAa;IACnB,OAAO;IACR,CAAC;AAEF,QAAK,mBAAmB,KAAK,IAAI,OAAO,aAAa;IACnD,MAAM,aAAa;IACnB,OAAO;IACR,CAAC;;EAEJ,MAAM,WAAW,KAAK;EACtB,MAAM,aAAa,KAAK;EACxB,MAAM,cAAc,KAAK;EAEzB,MAAM,UAAU,KAAK,IAAI,OAAO,qBAAqB,EAAE,OAAO,gBAAgB,CAAC;AAC/E,OAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;GAC1B,MAAM,OAAO,QAAQ,iBAAiB,EACpC,iBAAiB;IACf;IACA,2BAA2B,IAAI;IAC/B,qBAAqB,IAAI,IAAI;IAC9B,EACF,CAAC;AACF,QAAK,YAAY,QAAQ,GAAG,SAAS;AACrC,QAAK,aAAa,GAAG,QAAQ,GAAG,UAAU;AAC1C,QAAK,mBAAmB,cAAc,GAAG,IAAI,cAAc,GAAG,IAAI,cAAc,GAAG,GAAG;AACtF,QAAK,KAAK;;AAEZ,MAAI,YAAY;GACd,MAAM,eAAe,KAAK,UAAU,SAAS;GAC7C,MAAM,YAAY,KAAK,MAAM,OAAO;AACpC,OAAI,aACF,SAAQ,mBAAmB,cAAc,GAAG,KAAK,gBAAgB,GAAG,YAAY,EAAE;;AAGtF,UAAQ,gBAAgB,UAAU,GAAG,YAAY,YAAY,EAAE;AAC/D,UAAQ,mBAAmB,YAAY,GAAG,aAAa,GAAG,aAAa,EAAE;AACzE,OAAK,IAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;AAEhD,QAAM,YAAY,SAASA,iBAAe,GAAG,aAAa,EAAE;EAC5D,MAAM,KAAK,IAAI,eAAe,YAAY,eAAe,GAAG,aAAa,EAAE,CAAC,MAAM,EAAE,CAAC;AACrF,cAAY,OAAO;AACnB,OAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;GAC1B,MAAM,KAAK,OAAO,GAAG,IAAI,IAAI,KAAK,GAAG,IAAI,GAAG;AAC5C,OAAI,MAAM,EACR;GAEF,MAAM,MAAM,QAAQ,GAAG,KAAK;GAC5B,MAAM,MAAM,KAAK,YAAY,IAAI,IAAI,IAAI;IAAE,IAAI;IAAG,OAAO;IAAG;AAC5D,OAAI,MAAM;AACV,OAAI,SAAS;AACb,QAAK,YAAY,IAAI,KAAK,IAAI;;;;CAKlC,aAAmE;AACjE,SAAO,CAAC,GAAG,KAAK,YAAY,SAAS,CAAC,CACnC,KAAK,CAAC,QAAQ,QAAQ;GAAE;GAAQ,IAAI,EAAE;GAAI,OAAO,EAAE;GAAO,EAAE,CAC5D,MAAM,GAAG,MAAM,EAAE,KAAK,EAAE,GAAG;;;CAIhC,eAAqB;AACnB,OAAK,YAAY,OAAO;;;;CAK1B,IAAI,sBAA8B;AAChC,SAAO,KAAK,cAAc;;;;;CAM5B,IAAI,oBAA4B;AAC9B,SAAO,KAAK,IAAI,OAAO;;;;;;;;;CAUzB,MAAM,kBAAkB,SAAgC;EACtD,MAAM,iBAAiB,KAAK,cAAc,EAAE;EAC5C,MAAME,iBAAiC,EAAE,QAAQ,KAAK,QAAQ;AAC9D,OAAK,IAAI,OAAO,MAAM,YAAY,KAAK,gBAAgB,GAAG,IAAI,YAAY,CAAC,QAAQ,CAAC,CAAC;EAErF,MAAMG,gBAA4B,IAAI,MAAM,KAAK,cAAc,OAAO;AACtE,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,cAAc,QAAQ,KAAK;GAClD,MAAM,QAAQ,KAAK,cAAc;GACjC,MAAM,aAAa,MAAM,KAAK,YAAY,MAAM,MAAM,gBAAgB,eAAe;AACrF,QAAK,IAAI,OAAO,MAAM,YAAY,MAAM,eAAe,GAAG,WAAW;AACrE,iBAAc,KAAK,MAAM,KAAK,gBAAgB,MAAM,MAAM,gBAAgB,eAAe;;AAE3F,QAAM,KAAK,sBAAsB,KAAK,eAAe,eAAe,MAAM;AAC1E,OAAK,UAAU;;;;;;;;;;CAWjB,MAAM,MAAM,UAA8C;EACxD,MAAM,kBAAkB,KAAK,UAAU,YAAY;AACnD,MAAI,CAAC,gBACH,OAAM,IAAI,MACR,qHAED;EAGH,MAAM,UADa,KAAK,MAAM,OAAO,cACR;EAE7B,MAAM,IAAI,SAAS;EACnB,MAAM,iBAAiB,KAAK,cAAc,EAAE;EAC5C,MAAMH,iBAAiC,EAAE,QAAQ,KAAK,QAAQ;EAG9D,MAAM,UAAU,MAAM,IAAI,KAAK,gBAAgB,KAAK;AAEpD,OAAK,IAAI,OAAO,MAAM,YAAY,KAAK,gBAAgB,GAAG,SAAyB;EAEnF,MAAMC,gBAAiD,IAAI,MAAM,QAAQ,OAAO;AAChF,OAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,QAAQ,KAAK;GACvC,MAAM,QAAQ,QAAQ;GACtB,MAAM,aAAa,MAAM,KAAK,YAAY,MAAM,MAAM,gBAAgB,eAAe;GACrF,MAAM,aAAa,IAAI,WAAW,WAAW;GAE7C,IAAI,UAAU,CAAC,MAAM,mBAAmB,MAAM,gBAAgB,WAAW,WAAW;AACpF,OAAI,CAAC,SAAS;IACZ,MAAM,SAAS,MAAM;AACrB,SAAK,IAAI,IAAI,GAAG,IAAI,WAAW,QAAQ,IACrC,KAAI,WAAW,OAAO,OAAO,IAAI;AAC/B,eAAU;AACV;;;AAIN,OAAI,SAAS;AACX,SAAK,IAAI,OAAO,MAAM,YAAY,MAAM,eAAe,GAAG,WAAW;AACrE,UAAM,kBAAkB,IAAI,WAAW,WAAW,MAAM,EAAE,CAAC;;AAG7D,iBAAc,KAAK,MAAM,KAAK,gBAAgB,MAAM,MAAM,gBAAgB,eAAe;;EAI3F,MAAM,WAAW,KAAK,IAAI,OAAO,aAAa;GAAE,MAAM;GAAS,OAAO;GAAiB,CAAC;AAExF,MAAI,KAAK,mBAAmB;GAC1B,MAAM,QAAQ,KAAK;AACnB,QAAK,IAAI,QAAQ,GAAG,QAAQ,QAAQ,QAAQ,SAAS,OAAO;IAC1D,MAAM,MAAM,KAAK,IAAI,OAAO,sBAAsB;IAClD,MAAM,MAAM,KAAK,IAAI,QAAQ,OAAO,QAAQ,OAAO;AACnD,SAAK,IAAI,IAAI,OAAO,IAAI,KAAK,KAAK;KAChC,MAAM,IAAI,IAAI,kBAAkB;AAChC,OAAE,YAAY,QAAQ,GAAG,SAAS;AAClC,OAAE,aAAa,GAAG,QAAQ,GAAG,UAAU;AACvC,OAAE,mBAAmB,GAAG,cAAc,GAAG;AACzC,OAAE,KAAK;;AAET,SAAK,IAAI,OAAO,MAAM,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;AAC5C,UAAM,KAAK,IAAI,OAAO,MAAM,qBAAqB;;GAEnD,MAAM,UAAU,KAAK,IAAI,OAAO,sBAAsB;AACtD,WAAQ,mBAAmB,iBAAiB,GAAG,UAAU,GAAG,QAAQ;AACpE,QAAK,IAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;SAC3C;GACL,MAAM,UAAU,KAAK,IAAI,OAAO,qBAAqB,EAAE,OAAO,SAAS,CAAC;GACxE,MAAM,OAAO,QAAQ,iBAAiB,EAAE,OAAO,cAAc,CAAC;AAC9D,QAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,QAAQ,KAAK;AACvC,SAAK,YAAY,QAAQ,GAAG,SAAS;AACrC,SAAK,aAAa,GAAG,QAAQ,GAAG,UAAU;AAC1C,SAAK,mBAAmB,GAAG,cAAc,GAAG;;AAE9C,QAAK,KAAK;AACV,WAAQ,mBAAmB,iBAAiB,GAAG,UAAU,GAAG,QAAQ;AACpE,QAAK,IAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;;AAGlD,QAAM,SAAS,SAASH,iBAAe,GAAG,QAAQ;EAClD,MAAM,MAAM,IAAI,aAAa,SAAS,eAAe,GAAG,QAAQ,CAAC,MAAM,EAAE,CAAC;AAC1E,WAAS,OAAO;AAChB,WAAS,SAAS;AAElB,OAAK,UAAU;AACf,SAAO;;;;;;;;;;CAWT,MAAM,eAAe,YAAoB,WAA0C;EACjF,MAAM,YAAY,KAAK,UAAU,WAAW;AAC5C,MAAI,CAAC,UACH,OAAM,IAAI,MAAM,oDAAoD,WAAW,IAAI;EAErF,MAAM,UAAU,YAAY;EAI5B,MAAM,iBAAiB,KAAK,cAAc,EAAE;EAC5C,MAAME,iBAAiC,EAAE,QAAQ,KAAK,QAAQ;EAC9D,MAAM,UAAU,KAAK;EAErB,MAAMC,gBAAiD,IAAI,MAAM,QAAQ,OAAO;AAChF,OAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,QAAQ,KAAK;GACvC,MAAM,QAAQ,QAAQ;GACtB,MAAM,aAAa,MAAM,KAAK,YAAY,MAAM,MAAM,gBAAgB,eAAe;AACrF,QAAK,IAAI,OAAO,MAAM,YAAY,MAAM,eAAe,GAAG,WAAW;AACrE,SAAM,kBAAkB,IAAI,WAAW,IAAI,WAAW,WAAW,CAAC,MAAM,EAAE,CAAC;AAC3E,iBAAc,KAAK,MAAM,KAAK,gBAAgB,MAAM,MAAM,gBAAgB,eAAe;;EAG3F,MAAM,WAAW,KAAK,IAAI,OAAO,aAAa;GAAE,MAAM;GAAS,OAAO;GAAiB,CAAC;AAExF,MAAI,KAAK,mBAAmB;GAC1B,MAAM,QAAQ,KAAK;AACnB,QAAK,IAAI,QAAQ,GAAG,QAAQ,QAAQ,QAAQ,SAAS,OAAO;IAC1D,MAAM,MAAM,KAAK,IAAI,OAAO,sBAAsB;IAClD,MAAM,MAAM,KAAK,IAAI,QAAQ,OAAO,QAAQ,OAAO;AACnD,SAAK,IAAI,IAAI,OAAO,IAAI,KAAK,KAAK;KAChC,MAAM,OAAO,IAAI,kBAAkB;AACnC,UAAK,YAAY,QAAQ,GAAG,SAAS;AACrC,UAAK,aAAa,GAAG,QAAQ,GAAG,UAAU;AAC1C,UAAK,mBAAmB,GAAG,cAAc,GAAG;AAC5C,UAAK,KAAK;;AAEZ,SAAK,IAAI,OAAO,MAAM,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;AAC5C,UAAM,KAAK,IAAI,OAAO,MAAM,qBAAqB;;GAEnD,MAAM,UAAU,KAAK,IAAI,OAAO,sBAAsB;AACtD,WAAQ,mBAAmB,WAAW,GAAG,UAAU,GAAG,QAAQ;AAC9D,QAAK,IAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;SAC3C;GACL,MAAM,UAAU,KAAK,IAAI,OAAO,qBAAqB,EAAE,OAAO,kBAAkB,CAAC;GACjF,MAAM,OAAO,QAAQ,iBAAiB,EAAE,OAAO,YAAY,CAAC;AAC5D,QAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,QAAQ,KAAK;AACvC,SAAK,YAAY,QAAQ,GAAG,SAAS;AACrC,SAAK,aAAa,GAAG,QAAQ,GAAG,UAAU;AAC1C,SAAK,mBAAmB,GAAG,cAAc,GAAG;;AAE9C,QAAK,KAAK;AACV,WAAQ,mBAAmB,WAAW,GAAG,UAAU,GAAG,QAAQ;AAC9D,QAAK,IAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;;AAGlD,QAAM,SAAS,SAASH,iBAAe,GAAG,QAAQ;EAClD,MAAM,MAAM,IAAI,aAAa,SAAS,eAAe,GAAG,QAAQ,CAAC,MAAM,EAAE,CAAC;AAC1E,WAAS,OAAO;AAChB,WAAS,SAAS;AAClB,SAAO;;;;;;CAOT,MAAM,cAAc,UAAwC;EAC1D,MAAM,IAAI,SAAS;EAEnB,MAAM,iBAAiB,KAAK,cAAc,EAAE;EAC5C,MAAME,iBAAiC,EAAE,QAAQ,KAAK,QAAQ;AAG9D,OAAK,IAAI,OAAO,MAAM,YAAY,KAAK,gBAAgB,GAAG,SAAyB;EAGnF,MAAMI,sBAAuD,IAAI,MAC/D,KAAK,cAAc,OACpB;AACD,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,cAAc,QAAQ,KAAK;GAClD,MAAM,QAAQ,KAAK,cAAc;GACjC,MAAM,aAAa,MAAM,KAAK,YAAY,MAAM,MAAM,gBAAgB,eAAe;GACrF,MAAM,aAAa,IAAI,WAAW,WAAW;GAE7C,IAAI,UAAU,CAAC,MAAM,mBAAmB,MAAM,gBAAgB,WAAW,WAAW;AACpF,OAAI,CAAC,SAAS;IACZ,MAAM,SAAS,MAAM;AACrB,SAAK,IAAI,IAAI,GAAG,IAAI,WAAW,QAAQ,IACrC,KAAI,WAAW,OAAO,OAAO,IAAI;AAC/B,eAAU;AACV;;;AAKN,OAAI,SAAS;AACX,SAAK,IAAI,OAAO,MAAM,YAAY,MAAM,eAAe,GAAG,WAAW;AACrE,UAAM,kBAAkB,IAAI,WAAW,WAAW,MAAM,EAAE,CAAC;;AAG7D,uBAAoB,KAAK,MAAM,KAAK,gBAClC,MAAM,MACN,gBACA,eACD;;AAMH,MAAI,KAAK,aAAa;GACpB,MAAM,YAAY,KAAK,MAAM,OAAO;GACpC,MAAM,+BAAe,IAAI,YAAY,EAAE;GACvC,MAAM,OAAO,IAAI,SAAS,aAAa;AACvC,QAAK,UAAU,GAAG,WAAW,KAAK;AAClC,QAAK,UAAU,GAAG,GAAG,KAAK;AAC1B,QAAK,IAAI,OAAO,MAAM,YAAY,KAAK,YAAY,eAAe,GAAG,aAAa;;AAGpF,MAAI,KAAK,mBAAmB;GAG1B,MAAM,QAAQ,KAAK;AACnB,QAAK,IAAI,QAAQ,GAAG,QAAQ,KAAK,cAAc,QAAQ,SAAS,OAAO;IACrE,MAAM,MAAM,KAAK,IAAI,OAAO,sBAAsB;IAClD,MAAM,MAAM,KAAK,IAAI,QAAQ,OAAO,KAAK,cAAc,OAAO;AAC9D,SAAK,IAAI,IAAI,OAAO,IAAI,KAAK,KAAK;KAChC,MAAM,IAAI,IAAI,kBAAkB;AAChC,OAAE,YAAY,KAAK,cAAc,GAAG,SAAS;AAC7C,OAAE,aAAa,GAAG,KAAK,cAAc,GAAG,UAAU;AAClD,OAAE,mBAAmB,GAAG,oBAAoB,GAAG;AAC/C,OAAE,KAAK;;AAET,SAAK,IAAI,OAAO,MAAM,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;AAC5C,UAAM,KAAK,IAAI,OAAO,MAAM,qBAAqB;;AAKnD,OAAI,KAAK,aAAa;IACpB,MAAM,MAAM,KAAK,IAAI,OAAO,sBAAsB;IAClD,MAAM,IAAI,IAAI,kBAAkB;AAChC,MAAE,YAAY,KAAK,YAAY,SAAS;AACxC,MAAE,aAAa,GAAG,KAAK,YAAY,UAAU;AAC7C,MAAE,mBAAmB,GAAG,GAAG,EAAE;AAC7B,MAAE,KAAK;AACP,SAAK,IAAI,OAAO,MAAM,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;IAE5C,MAAM,UAAU,KAAK,IAAI,OAAO,sBAAsB;AACtD,YAAQ,mBAAmB,KAAK,oBAAoB,GAAG,KAAK,gBAAgB,GAAG,EAAE;AACjF,SAAK,IAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;;SAE7C;GAEL,MAAM,UAAU,KAAK,IAAI,OAAO,qBAAqB,EAAE,OAAO,UAAU,CAAC;GACzE,MAAM,OAAO,QAAQ,iBAAiB,EAAE,OAAO,WAAW,CAAC;AAC3D,QAAK,IAAI,IAAI,GAAG,IAAI,KAAK,cAAc,QAAQ,KAAK;IAClD,MAAM,QAAQ,KAAK,cAAc;AACjC,SAAK,YAAY,MAAM,SAAS;AAChC,SAAK,aAAa,GAAG,MAAM,UAAU;AACrC,SAAK,mBAAmB,GAAG,oBAAoB,GAAG;;AAEpD,OAAI,KAAK,aAAa;AACpB,SAAK,YAAY,KAAK,YAAY,SAAS;AAC3C,SAAK,aAAa,GAAG,KAAK,YAAY,UAAU;AAChD,SAAK,mBAAmB,GAAG,GAAG,EAAE;;AAElC,QAAK,KAAK;AACV,WAAQ,mBAAmB,KAAK,oBAAoB,GAAG,KAAK,gBAAgB,GAAG,EAAE;AACjF,QAAK,IAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;;AAGlD,QAAM,KAAK,eAAe,SAASN,iBAAe,GAAG,EAAE;EACvD,MAAM,SAAS,KAAK,eAAe,eAAe,GAAG,EAAE;EACvD,MAAM,UAAU,IAAI,YAAY,OAAO,MAAM,EAAE,CAAC,CAAC;AACjD,OAAK,eAAe,OAAO;AAE3B,OAAK,UAAU;AACf,SAAO;;;CAIT,0BAAkC;AAChC,SAAO,KAAK,YAAY,KAAK;;;CAI/B,OAAgB,iBAAiB;;;;;;;;;;;;;;;;CAiBjC,uBAAuB,SAAwB,MAAoB;EACjE,MAAM,iBAAiB,KAAK,cAAc,EAAE;EAC5C,MAAME,iBAAiC,EAAE,QAAQ,KAAK,QAAQ;AAE9D,MAAI,KAAK,gBAAgB,WAAW,EAClC,MAAK,IAAI,IAAI,GAAG,IAAI,SAAS,gBAAgB,IAC3C,MAAK,gBAAgB,KAAK,qBAAqB,KAAK,KAAK,mBAAmB,KAAK,EAAE,CAAC;AAIxF,MAAI,YAAY,MAAM;AACpB,QAAK,IAAI,OAAO,MAAM,YAAY,KAAK,gBAAgB,GAAG,IAAI,YAAY,CAAC,QAAQ,CAAC,CAAC;AAErF,OAAI,KAAK,aAAa;IACpB,MAAM,+BAAe,IAAI,YAAY,EAAE;IACvC,MAAM,OAAO,IAAI,SAAS,aAAa;AACvC,SAAK,UAAU,GAAG,KAAK,MAAM,OAAO,YAAY,KAAK;AACrD,SAAK,UAAU,GAAG,GAAG,KAAK;AAC1B,SAAK,IAAI,OAAO,MAAM,YAAY,KAAK,YAAY,eAAe,GAAG,aAAa;;;EAKtF,MAAMC,gBAAiD,IAAI,MAAM,KAAK,cAAc,OAAO;AAC3F,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,cAAc,QAAQ,KAAK;GAClD,MAAM,QAAQ,KAAK,cAAc;GACjC,MAAM,aAAa,MAAM,KAAK,YAAY,MAAM,MAAM,gBAAgB,eAAe;GACrF,MAAM,aAAa,IAAI,WAAW,WAAW;GAE7C,IAAI,UAAU,CAAC,MAAM,mBAAmB,MAAM,gBAAgB,WAAW,WAAW;AACpF,OAAI,CAAC,SAAS;IACZ,MAAM,SAAS,MAAM;AACrB,SAAK,IAAI,IAAI,GAAG,IAAI,WAAW,QAAQ,IACrC,KAAI,WAAW,OAAO,OAAO,IAAI;AAC/B,eAAU;AACV;;;AAKN,OAAI,SAAS;AACX,SAAK,IAAI,OAAO,MAAM,YAAY,MAAM,eAAe,GAAG,WAAW;AACrE,UAAM,kBAAkB,IAAI,WAAW,WAAW,MAAM,EAAE,CAAC;;AAG7D,iBAAc,KAAK,MAAM,KAAK,gBAAgB,MAAM,MAAM,gBAAgB,eAAe;;EAG3F,MAAM,UAAU,KAAK,IAAI,OAAO,qBAAqB,EAAE,OAAO,eAAe,CAAC;AAC9E,MAAI,YAAY,KAEd,SAAQ,mBAAmB,KAAK,oBAAoB,GAAG,KAAK,gBAAgB,GAAG,EAAE;EAEnF,MAAM,OAAO,QAAQ,iBAAiB,EAAE,OAAO,eAAe,CAAC;AAC/D,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,cAAc,QAAQ,KAAK;GAClD,MAAM,QAAQ,KAAK,cAAc;AACjC,QAAK,YAAY,MAAM,SAAS;AAChC,QAAK,aAAa,GAAG,MAAM,UAAU;AACrC,QAAK,mBAAmB,GAAG,cAAc,GAAG;;AAE9C,MAAI,KAAK,aAAa;AACpB,QAAK,YAAY,KAAK,YAAY,SAAS;AAC3C,QAAK,aAAa,GAAG,KAAK,YAAY,UAAU;AAChD,QAAK,mBAAmB,GAAG,GAAG,EAAE;;AAElC,OAAK,KAAK;AACV,UAAQ,mBAAmB,KAAK,oBAAoB,GAAG,KAAK,gBAAgB,OAAO,GAAG,EAAE;AACxF,OAAK,IAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;AAEhD,OAAK,UAAU;;;CAIjB,MAAM,gBAAgB,MAA+B;EACnD,MAAM,MAAM,KAAK,gBAAgB;AACjC,QAAM,IAAI,SAASH,iBAAe,GAAG,EAAE;EACvC,MAAM,SAAS,IAAI,eAAe,GAAG,EAAE;EACvC,MAAM,UAAU,IAAI,YAAY,OAAO,MAAM,EAAE,CAAC,CAAC;AACjD,MAAI,OAAO;AACX,SAAO;;CAGT,QAAc;AACZ,OAAK,SAAS;AAEd,OAAK,MAAM,OAAO,KAAK,gBAAgB,QAAQ,CAC7C,MAAK,IAAI,OAAO,MAAM,YAAY,KAAK,GAAG,IAAI,WAAW,IAAI,KAAK,CAAC;AAGrE,OAAK,MAAM,SAAS,KAAK,iBAAiB;AACxC,SAAM,kBAAkB;AACxB,SAAM,mBAAmB;;AAE3B,OAAK,MAAM,SAAS,KAAK,eAAe;AACtC,SAAM,kBAAkB;AACxB,SAAM,mBAAmB;;;;;;;;;CAU7B,MAAM,mBAAmB,UAKtB;EACD,MAAM,IAAI,SAAS;EACnB,MAAM,iBAAiB,KAAK,cAAc,EAAE;EAC5C,MAAME,iBAAiC,EAAE,QAAQ,KAAK,QAAQ;EAE9D,MAAM,SADU,MAAM,IAAI,KAAK,gBAAgB,KAAK,iBAC9B;AAGtB,OAAK,IAAI,OAAO,MAAM,YAAY,KAAK,gBAAgB,GAAG,SAAyB;EAGnF,MAAM,aAAa,MAAM,KAAK,YAAY,MAAM,MAAM,gBAAgB,eAAe;AACrF,OAAK,IAAI,OAAO,MAAM,YAAY,MAAM,eAAe,GAAG,WAAW;EAGrE,MAAM,eAAe,MAAM,KAAK,gBAAgB,MAAM,MAAM,gBAAgB,eAAe;EAK3F,MAAM,UAAU,KAAK,IAAI,OAAO,qBAAqB,EAAE,OAAO,eAAe,CAAC;EAC9E,MAAM,OAAO,QAAQ,iBAAiB,EAAE,OAAO,oBAAoB,CAAC;AACpE,OAAK,YAAY,MAAM,SAAS;AAChC,OAAK,aAAa,GAAG,MAAM,UAAU;AACrC,OAAK,mBAAmB,GAAG,aAAa;AACxC,OAAK,KAAK;AACV,OAAK,IAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;EAGhD,MAAM,aAAa,MAAM,KAAK,QAAQ;EACtC,MAAM,SAAS,MAAM,KAAK,gBAAgB,YAAY,GAAG;AAEzD,SAAO;GACL,QAAQ,MAAM;GACd,QAAQ,MAAM,KAAK;GACnB;GACA;GACD;;;;;;;CAQH,MAAM,mBACJ,YACA,GACmE;EACnE,MAAM,iBAAiB,KAAK,cAAc,EAAE;EAC5C,MAAMA,iBAAiC,EAAE,QAAQ,KAAK,QAAQ;EAE9D,MAAM,SADU,MAAM,IAAI,KAAK,gBAAgB,KAAK,iBAC9B;EAGtB,MAAM,aAAa,MAAM,KAAK,YAAY,MAAM,MAAM,gBAAgB,eAAe;AACrF,OAAK,IAAI,OAAO,MAAM,YAAY,MAAM,eAAe,GAAG,WAAW;EAGrE,MAAM,eAAe,MAAM,KAAK,gBAAgB,MAAM,MAAM,gBAAgB,eAAe;EAK3F,MAAM,UAAU,KAAK,IAAI,OAAO,qBAAqB,EAAE,OAAO,eAAe,cAAc,CAAC;EAC5F,MAAM,OAAO,QAAQ,iBAAiB,EAAE,OAAO,cAAc,cAAc,CAAC;AAC5E,OAAK,YAAY,MAAM,SAAS;AAChC,OAAK,aAAa,GAAG,MAAM,UAAU;AACrC,OAAK,mBAAmB,GAAG,aAAa;AACxC,OAAK,KAAK;AACV,OAAK,IAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;EAEhD,MAAM,aAAa,MAAM,KAAK,QAAQ;EACtC,MAAM,SAAS,MAAM,KAAK,gBAAgB,YAAY,GAAG;AACzD,SAAO;GAAE,QAAQ,MAAM;GAAQ,QAAQ,MAAM,KAAK;GAAQ;GAAQ;;;;;;CAOpE,mBACE,GACA,QAAQ,GAOP;EACD,MAAM,iBAAiB,KAAK,cAAc,EAAE;EAC5C,MAAMA,iBAAiC,EAAE,QAAQ,KAAK,QAAQ;EAC9D,MAAM,UAAU,MAAM,IAAI,KAAK,gBAAgB,KAAK;EACpD,MAAM,UAAU,EAAE;AAClB,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,IAAI,OAAO,QAAQ,OAAO,EAAE,KAAK;GACxD,MAAM,QAAQ,QAAQ;GACtB,MAAM,aAAa,MAAM,KAAK,YAAY,MAAM,MAAM,gBAAgB,eAAe;GACrF,MAAM,MAAM,IAAI,YAAY,WAAW;GACvC,MAAM,eAAe,MAAM,KAAK,gBAC9B,MAAM,MACN,gBACA,eACD;AACD,WAAQ,KAAK;IACX,KAAK;IACL,QAAQ,MAAM;IACd,QAAQ,MAAM,KAAK;IACnB,WAAW,MAAM,KAAK,IAAI;IAC1B;IACD,CAAC;;AAEJ,SAAO;;;;;;CAOT,MAAM,mBAAmB,GAUvB;EACA,MAAM,UAAU,MAAM,IAAI,KAAK,gBAAgB,KAAK;EAEpD,MAAM,eAAe,QAAQ;EAC7B,MAAM,UAAU;GACd;GACA;GACA;GACA;GACA,KAAK,MAAM,eAAe,EAAE;GAC5B,KAAK,MAAM,eAAe,EAAE;GAC5B,KAAK,MAAO,eAAe,IAAK,EAAE;GAClC,eAAe;GAChB;EACD,MAAM,SAAS,CAAC,GAAG,IAAI,IAAI,QAAQ,CAAC,CAAC,QAAQ,MAAM,KAAK,KAAK,IAAI,aAAa;EAE9E,MAAMK,UAQD,EAAE;AACP,OAAK,MAAM,KAAK,QAAQ;GACtB,MAAM,QAAQ,QAAQ;GACtB,MAAM,aAAa,MAAM,KAAK,QAAQ;AACtC,OAAI,CAAC,WAAY;AACjB,OAAI;IACF,MAAM,OAAO,MAAM,KAAK,gBAAgB,YAAY,GAAG;IACvD,IAAI,MAAM;AACV,SAAK,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,IAAK,QAAO,KAAK;IAGlD,IAAIC;AACJ,QAAI;KACF,MAAM,SAAS,MAAM,cAAc;KACnC,MAAM,WAAW,KAAK,IAAI,OAAO,aAAa;MAC5C,MAAM;MACN,OAAO;MACR,CAAC;KACF,MAAM,MAAM,KAAK,IAAI,OAAO,sBAAsB;AAClD,SAAI,mBAAmB,MAAM,eAAe,GAAG,UAAU,GAAG,OAAO;AACnE,UAAK,IAAI,OAAO,MAAM,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;AAC5C,WAAM,SAAS,SAASR,iBAAe,GAAG,OAAO;KACjD,MAAM,MAAM,IAAI,YAAY,SAAS,eAAe,GAAG,OAAO,CAAC,MAAM,EAAE,CAAC;AACxE,qBAAgB,MAAM,KAAK,IAAI;AAC/B,cAAS,OAAO;AAChB,cAAS,SAAS;YACZ;AAIR,YAAQ,KAAK;KACX,KAAK;KACL,QAAQ,MAAM;KACd,QAAQ,MAAM,KAAK;KACnB,QAAQ;KACR,KAAK,KAAK,MAAM,MAAM,IAAI,GAAG;KAC7B,QAAQ;MAAC,KAAK;MAAI,KAAK;MAAI,KAAK;MAAI,KAAK;MAAG,CAAC,KAAK,MAAM,KAAK,MAAM,IAAI,IAAI,GAAG,IAAI;KAClF;KACD,CAAC;WACI;AACN,YAAQ,KAAK;KACX,KAAK;KACL,QAAQ,MAAM;KACd,QAAQ,MAAM,KAAK;KACnB,QAAQ;KACR,KAAK;KACL,QAAQ;MAAC;MAAK;MAAK;MAAK;MAAI;KAC7B,CAAC;;;AAGN,SAAO;;CAGT,iBAAiB,YAAoB,MAA6B;EAChE,MAAM,SAAS,KAAK,UAAU,WAAW;AACzC,MAAI,CAAC,OAAQ,OAAM,IAAI,MAAM,kBAAkB,WAAW,GAAG;AAC7D,OAAK,IAAI,OAAO,MAAM,YAAY,QAAQ,GAAG,KAAqB;;;;;;;CAQpE,WAAW,YAAoB,MAA6B;EAC1D,MAAM,SAAS,KAAK,UAAU,WAAW;AACzC,MAAI,CAAC,OAAQ,OAAM,IAAI,MAAM,wBAAwB,WAAW,GAAG;AACnE,OAAK,IAAI,OAAO,MAAM,YAAY,QAAQ,GAAG,KAAqB;;;CAIpE,aAAa,YAAoB,MAAuB,YAA0B;EAChF,MAAM,SAAS,KAAK,UAAU,WAAW;AACzC,MAAI,CAAC,OAAQ,OAAM,IAAI,MAAM,wBAAwB,WAAW,GAAG;AACnE,OAAK,IAAI,OAAO,MAAM,YAAY,QAAQ,YAAY,KAAqB;;;CAI7E,UAAU,YAA6B;AACrC,SAAO,KAAK,UAAU,WAAW,KAAK;;;CAIxC,IAAI,gBAAwB;AAC1B,SAAO,KAAK;;CAGd,MAAM,gBACJ,YACA,aACA,YACuB;EACvB,MAAM,SAAS,KAAK,UAAU,WAAW;AACzC,MAAI,CAAC,OAAQ,OAAM,IAAI,MAAM,kBAAkB,WAAW,GAAG;EAC7D,MAAM,YAAY,cAAc;EAChC,MAAM,YAAY,OAAO,OAAO;EAChC,MAAM,UAAU,cAAc,KAAK,IAAI,cAAc,GAAG,UAAU,GAAG;EACrE,MAAM,WAAW,KAAK,IAAI,OAAO,aAAa;GAC5C,MAAM;GACN,OAAO;GACR,CAAC;EACF,MAAM,UAAU,KAAK,IAAI,OAAO,sBAAsB;AACtD,UAAQ,mBAAmB,QAAQ,WAAW,UAAU,GAAG,QAAQ;AACnE,OAAK,IAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;AAChD,QAAM,SAAS,SAASA,iBAAe,GAAG,QAAQ;EAClD,MAAM,OAAO,IAAI,aAAa,SAAS,eAAe,GAAG,QAAQ,CAAC,MAAM,EAAE,CAAC;AAC3E,WAAS,OAAO;AAChB,WAAS,SAAS;AAClB,SAAO;;CAGT,UAAgB;AACd,iBAAe,CAAC,GAAG,KAAK,cAAc,QAAQ,CAAC,CAAC;AAEhD,iBAAe,CAAC,GAAG,IAAI,IAAI,KAAK,kBAAkB,QAAQ,CAAC,CAAC,CAAC;AAC7D,iBAAe,CAAC,GAAG,KAAK,gBAAgB,QAAQ,CAAC,CAAC;AAClD,iBAAe,CAAC,GAAG,KAAK,eAAe,QAAQ,CAAC,CAAC;AACjD,OAAK,MAAM,SAAS,KAAK,gBACvB,OAAM,cAAc,SAAS;AAG/B,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,cAAc,QAAQ,IAC7C,KAAI,KAAK,cAAc,OAAO,KAAK,gBAAgB,GACjD,MAAK,cAAc,GAAG,cAAc,SAAS;AAGjD,MAAI,KAAK,YACP,MAAK,YAAY,cAAc,SAAS;AAE1C,MAAI,KAAK,sBAAsB;AAC7B,QAAK,qBAAqB,SAAS;AACnC,QAAK,uBAAuB;;AAE9B,OAAK,eAAe,SAAS;AAC7B,OAAK,eAAe,SAAS;AAC7B,OAAK,mBAAmB,SAAS;AACjC,OAAK,eAAe,SAAS;AAC7B,iBAAe,KAAK,gBAAgB;AACpC,OAAK,kBAAkB,EAAE;AACzB,OAAK,cAAc,OAAO;AAC1B,OAAK,kBAAkB,OAAO;AAC9B,OAAK,gBAAgB,OAAO;AAC5B,OAAK,eAAe,OAAO;AAC3B,OAAK,kBAAkB,EAAE;AACzB,OAAK,gBAAgB,EAAE;;;;;;;;;;;;;;;;;;CAqBzB,AAAQ,4BAAkC;EACxC,MAAM,gBAAgB,SAAiB,KAAK,MAAM,QAAQ,OAAO,YAAY;AAI7E,MADe,OAAO,YAAY,eAAe,QAAQ,KAAK,uBAAuB,KACzE;AACV,QAAK,MAAM,QAAQ,OAAO,KAAK,KAAK,MAAM,QAAQ,EAAE;AAClD,QAAI,CAAC,aAAa,KAAK,CAAE;IACzB,MAAM,OAAO,KAAK,MAAM,QAAQ;IAIhC,MAAM,QAHQ,KAAK,MAAM,KAAK,MAC5B,MAAM,OAAO,MAAM,UAAU,KAAK,YAAa,EAChD,CACmB,QAAQ,GAAG,MAAM,IAAI,GAAG,EAAE,GAAG,YAAY,KAAK;AAClE,SAAK,kBAAkB,IACrB,MACA,oBAAoB,KAAK,KAAK,cAAc,QAAQ,MAAM,CAC3D;;AAEH,WAAQ,IACN,gDAAgD,KAAK,kBAAkB,KAAK,oBAC7E;AACD;;EAEF,MAAM,gBAAgB,SAAyB;GAC7C,MAAM,OAAO,KAAK,MAAM,QAAQ;AAMhC,UALc,KAAK,MAAM,KAAK,MAAM;AAClC,QAAI,MAAM,IAAK,QAAO,KAAK;AAC3B,QAAI,MAAM,QAAS,QAAO,KAAK;AAC/B,WAAO;KACP,CACW,QAAQ,GAAG,MAAM,IAAI,GAAG,EAAE,GAAG,YAAY,KAAK;;EAG7D,MAAM,WAAW,IAAI,IAAI,KAAK,MAAM,MAAM,KAAK,MAAM,CAAC,EAAE,IAAI,EAAE,CAAC,CAAC;EAChE,MAAM,QAAQ,KAAK,MAAM;EACzB,MAAM,aAAa,IAAI,IAAY,KAAK,MAAM,QAAQ;EACtD,MAAM,2BAAW,IAAI,KAAqB;EAC1C,MAAM,0BAAU,IAAI,KAAqB;AACzC,OAAK,IAAI,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;GACrC,MAAM,OAAO,SAAS,IAAI,MAAM,GAAG;AACnC,QAAK,MAAM,OAAO,KAAK,QAAQ;AAC7B,QAAI,CAAC,aAAa,IAAI,CAAE;AAGxB,QAAI,CAAC,SAAS,IAAI,IAAI,CAAE,YAAW,IAAI,IAAI;AAC3C,YAAQ,IAAI,KAAK,EAAE;;AAErB,QAAK,MAAM,OAAO,KAAK,QACrB,KAAI,aAAa,IAAI,IAAI,CAAC,SAAS,IAAI,IAAI,CAAE,UAAS,IAAI,KAAK,EAAE;;EAIrE,MAAM,2BAAW,IAAI,KAA0B;EAC/C,IAAI,UAAU;EACd,MAAM,2BAAW,IAAI,KAAa;AAClC,OAAK,IAAI,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;GACrC,MAAM,OAAO,SAAS,IAAI,MAAM,GAAG;AAGnC,QAAK,MAAM,OAAO,KAAK,SAAS;AAC9B,QAAI,CAAC,aAAa,IAAI,IAAI,WAAW,IAAI,IAAI,IAAI,KAAK,kBAAkB,IAAI,IAAI,CAC9E;IAEF,MAAM,WAAW,aAAa,IAAI;IAElC,IAAI,SADS,SAAS,IAAI,SAAS,EAChB,KAAK;AACxB,QAAI,CAAC,QAAQ;AACX,cAAS,oBAAoB,KAAK,KAAK,YAAY,QAAQ,GAAG,SAAS,IAAI,SAAS;AACpF;;AAEF,SAAK,kBAAkB,IAAI,KAAK,OAAO;;AAGzC,QAAK,MAAM,QAAQ,CAAC,GAAG,KAAK,QAAQ,GAAG,KAAK,QAAQ,EAAE;AACpD,QAAI,CAAC,aAAa,KAAK,IAAI,WAAW,IAAI,KAAK,IAAI,SAAS,IAAI,KAAK,CAAE;AAEvE,SADa,QAAQ,IAAI,KAAK,IAAI,SAAS,IAAI,KAAK,IAAI,KAC7C,EAAG;IACd,MAAM,SAAS,KAAK,kBAAkB,IAAI,KAAK;AAC/C,QAAI,CAAC,OAAQ;AACb,aAAS,IAAI,KAAK;IAClB,MAAM,WAAW,aAAa,KAAK;IACnC,MAAM,OAAO,SAAS,IAAI,SAAS,IAAI,EAAE;AACzC,SAAK,KAAK,OAAO;AACjB,aAAS,IAAI,UAAU,KAAK;;;AAMhC,OAAK,MAAM,QAAQ,OAAO,KAAK,KAAK,MAAM,QAAQ,EAAE;AAClD,OAAI,CAAC,aAAa,KAAK,IAAI,KAAK,kBAAkB,IAAI,KAAK,CAAE;AAC7D,QAAK,kBAAkB,IACrB,MACA,oBAAoB,KAAK,KAAK,OAAO,QAAQ,aAAa,KAAK,CAAC,CACjE;;EAGH,MAAM,SAAS,IAAI,IAAI,KAAK,kBAAkB,QAAQ,CAAC;EACvD,IAAI,aAAa;AACjB,OAAK,MAAM,KAAK,OAAQ,eAAc,EAAE;AACxC,UAAQ,IACN,2BAA2B,KAAK,kBAAkB,KAAK,aAAa,OAAO,KAAK,aAAa,aAAa,KAAK,QAAQ,EAAE,CAAC,MAC3H;;CAGH,AAAQ,0BAAgC;AACtC,OAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,KAAK,MAAM,QAAQ,EAAE;AAC7D,OAAI,KAAK,YAAY,YAAa;GAElC,MAAM,WADQ,KAAK,MACI,QAAQ,GAAG,MAAM,IAAI,GAAG,EAAE,GAAG,YAAY,KAAK;GACrE,MAAM,SAAS,oBAAoB,KAAK,KAAK,OAAO,QAAQ,SAAS;AACrE,QAAK,gBAAgB,IAAI,MAAM,OAAO;;;CAI1C,AAAQ,yBAA+B;AACrC,OAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,KAAK,MAAM,QAAQ,EAAE;AAC7D,OAAI,KAAK,YAAY,WAAY;GAKjC,MAAM,WAJQ,KAAK,MAAM,KAAK,MAAM;AAClC,QAAI,MAAM,QAAS,QAAO,KAAK;AAC/B,WAAO;KACP,CACqB,QAAQ,GAAG,MAAM,IAAI,GAAG,EAAE,GAAG,YAAY,KAAK;GACrE,MAAM,SAAS,oBAAoB,KAAK,KAAK,MAAM,QAAQ,SAAS;AACpE,QAAK,eAAe,IAAI,MAAM,OAAO;;;CAIzC,AAAQ,cAAc,GAAqC;EACzD,MAAMS,WAAqC,EAAE;AAC7C,OAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,KAAK,MAAM,QAAQ,CAC3D,UAAS,QAAQ,KAAK,MAAM,KAAK,MAAM;AACrC,OAAI,MAAM,IAAK,QAAO;AACtB,OAAI,MAAM,QAAS,QAAO,KAAK,SAAS;AACxC,UAAO;IACP;AAEJ,SAAO;;CAGT,AAAQ,UAAU,YAA2C;AAC3D,SACE,KAAK,cAAc,IAAI,WAAW,IAClC,KAAK,kBAAkB,IAAI,WAAW,IACtC,KAAK,gBAAgB,IAAI,WAAW,IACpC,KAAK,eAAe,IAAI,WAAW;;;;;;CAQvC,AAAQ,0BAAgC;EACtC,IAAI,cAAc;EAClB,MAAM,cAAc,KAAK,IAAI,OAAO;AAEpC,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,cAAc,SAAS,GAAG,KAAK;GACtD,MAAM,KAAK,KAAK,cAAc;GAC9B,MAAM,KAAK,KAAK,cAAc,IAAI;GAClC,MAAM,KAAK,KAAK,cAAc,IAAI;GAGlC,MAAM,UAAU,GAAG,KAAK,WAAW,YAAY,GAAG,KAAK,WAAW;GAClE,MAAM,UAAU,GAAG,KAAK,WAAW,YAAY,GAAG,KAAK,WAAW;AAClE,OAAI,CAAC,WAAW,CAAC,WAAW,GAAG,KAAK,WAAW,SAAU;AAGzD,OAAI,GAAG,KAAK,WAAW,GAAG,KAAK,OAAQ;GAGvC,MAAM,eAAe,GAAG,KAAK,WAAW;AACxC,OAAI,gBAAgB,cAAc,EAAG;AAGrC,OAAI,GAAG,KAAK,OAAO,OAAO,GAAG,KAAK,OAAO,GAAI;GAG7C,MAAM,UAAU,GAAG,KAAK,QAAQ;GAChC,MAAM,QAAQ,GAAG,KAAK,QAAQ;AAC9B,OAAI,GAAG,KAAK,OAAO,OAAO,WAAW,GAAG,KAAK,OAAO,OAAO,MAAO;AAGlE,OAAI,GAAG,KAAK,WAAW,MAAM,GAAG,KAAK,WAAW,EAAG;AACnD,OAAI,GAAG,KAAK,WAAW,MAAM,GAAG,KAAK,WAAW,EAAG;AAMnD,OAAI,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG,KAAK,KAAK,UAAU,GAAG,KAAK,QAAQ,GAAG,CAAE;GAG9E,MAAM,YAAY,eAAe,0BAA0B;GAC3D,MAAM,gBAAgB,oBACpB,KAAK,KACL,aAAa,GAAG,UAChB,UAAU,YACV,UAAU,WACX;GAED,MAAM,mBAAmB,UAAU,YAAY,GAAG,MAAM,EAAE,EAAE,EAAE,QAAQ,GAAG,CAAC;GAC1E,MAAM,eAAe,oBACnB,KAAK,KACL,qBAAqB,GAAG,UACxB,iBACD;GAGD,MAAM,WAAW,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAClD,MAAM,YAAY,KAAK,UAAU,GAAG,KAAK,QAAQ,GAAG;GAEpD,IAAIC;AACJ,OAAI,cAAc;IAEhB,MAAM,QAAQ,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;IAC/C,MAAM,QAAQ,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;IAC/C,MAAM,QAAQ,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;IAC/C,MAAM,MAAM,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;IAC7C,MAAM,MAAM,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;IAC7C,MAAM,MAAM,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;AAC7C,oBAAgB;KACd,EAAE,QAAQ,UAAU;KACpB,EAAE,QAAQ,OAAO;KACjB,EAAE,QAAQ,OAAO;KACjB,EAAE,QAAQ,OAAO;KACjB,EAAE,QAAQ,KAAK;KACf,EAAE,QAAQ,KAAK;KACf,EAAE,QAAQ,KAAK;KACf,EAAE,QAAQ,WAAW;KACrB,EAAE,QAAQ,cAAc;KACzB;UACI;IAEL,MAAM,UAAU,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;IACjD,MAAM,QAAQ,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;AAC/C,oBAAgB;KACd,EAAE,QAAQ,UAAU;KACpB,EAAE,QAAQ,SAAS;KACnB,EAAE,QAAQ,OAAO;KACjB,EAAE,QAAQ,WAAW;KACrB,EAAE,QAAQ,cAAc;KACzB;;GAGH,MAAM,iBAAiB,gBACrB,KAAK,KACL,eACA,eACA,gBAAgB,GAAG,SACpB;GAED,MAAMC,aAA4B;IAChC,QAAQ,gBAAgB,GAAG;IAC3B,MAAM,GAAG;IACT,MAAM;IACN,UAAU;IACV,WAAW;IACX,eAAe;IACf,iBAAiB;IACjB,kBAAkB;IACnB;AAGD,QAAK,cAAc,OAAO,GAAG,GAAG,WAAW;AAC3C;;AAIF,MAAI,cAAc,EAChB,SAAQ,IACN,oBAAoB,YAAY,4BAA4B,cAAc,EAAE,cAC7E;;;;;;;;;;;;;;;;CAkBL,AAAQ,8BAAoC;EAC1C,IAAI,cAAc;AAGlB,MADoB,KAAK,IAAI,OAAO,kCAClB,EAAG;AAErB,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,cAAc,SAAS,GAAG,KAAK;GACtD,MAAM,KAAK,KAAK,cAAc;GAC9B,MAAM,KAAK,KAAK,cAAc,IAAI;AAGlC,OAAI,GAAG,KAAK,WAAW,gBAAgB,GAAG,KAAK,WAAW,aAAc;AAExE,OAAI,GAAG,KAAK,OAAO,OAAO,GAAG,KAAK,OAAO,GAAI;AAE7C,OAAI,GAAG,KAAK,WAAW,MAAM,GAAG,KAAK,WAAW,EAAG;AACnD,OAAI,GAAG,KAAK,WAAW,MAAM,GAAG,KAAK,WAAW,EAAG;AACnD,OAAI,GAAG,KAAK,WAAW,eAAe,GAAG,KAAK,WAAW,WAAY;GAErE,MAAM,OAAO,GAAG,KAAK,QAAQ;GAC7B,MAAM,OAAO,GAAG,KAAK,QAAQ;GAC7B,MAAM,WAAW,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAClD,MAAM,UAAU,KAAK,UAAU,KAAK;GACpC,MAAM,UAAU,KAAK,UAAU,KAAK;AACpC,OAAI,CAAC,YAAY,CAAC,WAAW,CAAC,QAAS;AAMvC,OAAI,aAAa,WAAW,aAAa,WAAW,YAAY,QAAS;GAEzE,MAAM,YAAY;GAClB,MAAM,gBAAgB,oBACpB,KAAK,KACL,WAAW,GAAG,UACd,UAAU,YACV,UAAU,WACX;GAED,MAAM,mBAAmB,UAAU,YAAY,GAAG,MAAM,EAAE,EAAE,EAAE,QAAQ,GAAG,CAAC;GAC1E,MAAM,eAAe,oBACnB,KAAK,KACL,mBAAmB,GAAG,UACtB,iBACD;GAGD,MAAM,MAAM,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAC7C,MAAM,MAAM,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAC7C,MAAM,MAAM,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAC7C,MAAM,MAAM,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAC7C,MAAM,MAAM,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAC7C,MAAM,MAAM,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAE7C,MAAM,gBAAgB;IACpB,EAAE,QAAQ,UAAU;IACpB,EAAE,QAAQ,KAAK;IACf,EAAE,QAAQ,KAAK;IACf,EAAE,QAAQ,KAAK;IACf,EAAE,QAAQ,KAAK;IACf,EAAE,QAAQ,KAAK;IACf,EAAE,QAAQ,KAAK;IACf,EAAE,QAAQ,SAAS;IACnB,EAAE,QAAQ,SAAS;IACnB,EAAE,QAAQ,cAAc;IACzB;GAED,MAAM,iBAAiB,gBACrB,KAAK,KACL,eACA,eACA,cAAc,GAAG,SAClB;GAED,MAAMA,aAA4B;IAChC,QAAQ,cAAc,GAAG;IACzB,MAAM,GAAG;IACT,MAAM;IACN,UAAU;IACV,WAAW;IACX,eAAe;IACf,iBAAiB;IACjB,kBAAkB;IACnB;AAED,QAAK,cAAc,OAAO,GAAG,GAAG,WAAW;AAC3C;;AAIF,MAAI,cAAc,EAChB,SAAQ,IACN,oBAAoB,YAAY,gCAAgC,YAAY,cAC7E;;;;;;;;;;;;;CAeL,AAAQ,+BAAqC;EAC3C,IAAI,cAAc;AAElB,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,cAAc,SAAS,GAAG,KAAK;GACtD,MAAM,KAAK,KAAK,cAAc;GAC9B,MAAM,KAAK,KAAK,cAAc,IAAI;AAElC,OAAI,GAAG,KAAK,WAAW,mBAAmB,GAAG,KAAK,WAAW,gBAAiB;AAE9E,OAAI,GAAG,KAAK,WAAW,UAAU,GAAG,KAAK,WAAW,MAAO;GAI3D,MAAM,OAAO,GAAG,KAAK,QAAQ;GAC7B,MAAM,OAAO,GAAG,KAAK,QAAQ;GAC7B,MAAM,MAAM,KAAK,MAAM,QAAQ,OAAO;AAEtC,OAAI,QADQ,KAAK,MAAM,QAAQ,OAAO,MACrB;GACjB,IAAIC;AACJ,OAAI,QAAQ,MACV,aAAY;YACH,QAAQ,MACjB,aACE,KAAK,WAAW,eACZ,uCACA;OAEN;GAGF,MAAM,OAAO,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAC9C,MAAM,OAAO,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAC9C,MAAM,UAAU,KAAK,UAAU,KAAK;GACpC,MAAM,UAAU,KAAK,UAAU,KAAK;AACpC,OAAI,CAAC,QAAQ,CAAC,QAAQ,CAAC,WAAW,CAAC,QAAS;AAE5C,OAAI,YAAY,QAAS;GAEzB,MAAM,gBAAgB,oBACpB,KAAK,KACL,WAAW,GAAG,UACd,UAAU,YACV,UAAU,WACX;GAED,MAAM,cAAc,KAAK,cAAc,EAAE;GACzC,MAAM,mBAAmB,UAAU,YAAY,GAAG,MAAM,aAAa,EAAE,QAAQ,GAAG,CAAC;GACnF,MAAM,eAAe,oBACnB,KAAK,KACL,mBAAmB,GAAG,UACtB,iBACD;GAED,MAAM,gBAAgB;IACpB,EAAE,QAAQ,MAAM;IAChB,EAAE,QAAQ,MAAM;IAChB,EAAE,QAAQ,SAAS;IACnB,EAAE,QAAQ,SAAS;IACnB,EAAE,QAAQ,cAAc;IACzB;GAED,MAAM,iBAAiB,gBACrB,KAAK,KACL,eACA,eACA,cAAc,GAAG,SAClB;GAED,MAAMD,aAA4B;IAChC,QAAQ,YAAY,GAAG;IACvB,MAAM,GAAG;IACT,MAAM;IACN,UAAU;IACV,WAAW;IACX,eAAe;IACf,iBAAiB;IACjB,kBAAkB;IACnB;AAED,QAAK,cAAc,OAAO,GAAG,GAAG,WAAW;AAC3C;;AAGF,MAAI,cAAc,EAChB,SAAQ,IACN,oBAAoB,YAAY,gCAAgC,YAAY,cAC7E;;;;;;;;;;;;;;;;;CAmBL,AAAQ,8BAAoC;EAC1C,IAAI,cAAc;AAElB,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,cAAc,SAAS,GAAG,KAAK;GACtD,MAAM,KAAK,KAAK,cAAc;GAC9B,MAAM,KAAK,KAAK,cAAc,IAAI;AAElC,OAAI,GAAG,KAAK,WAAW,cAAe;AACtC,OAAI,GAAG,KAAK,WAAW,aAAc;GAGrC,MAAM,WAAW,GAAG,KAAK,QAAQ;AACjC,OAAI,GAAG,KAAK,OAAO,OAAO,SAAU;GAGpC,MAAM,WAAW,GAAG,KAAK,OAAO;GAChC,MAAM,WAAW,GAAG,KAAK,OAAO;GAChC,MAAM,UAAU,KAAK,UAAU,SAAS;GACxC,MAAM,UAAU,KAAK,UAAU,SAAS;GACxC,MAAM,KAAK,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAC5C,MAAM,KAAK,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAC5C,MAAM,KAAK,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAC5C,MAAM,SAAS,KAAK,UAAU,GAAG,KAAK,QAAQ,GAAG;AACjD,OAAI,CAAC,WAAW,CAAC,WAAW,CAAC,MAAM,CAAC,MAAM,CAAC,MAAM,CAAC,OAAQ;AAI1D,OAAI,WAAW,WAAW,WAAW,WAAW,YAAY,QAAS;GAErE,MAAM,YAAY;GAClB,MAAM,gBAAgB,oBACpB,KAAK,KACL,WAAW,GAAG,UACd,UAAU,YACV,UAAU,WACX;GAED,MAAM,mBAAmB,UAAU,YAAY,GAAG,MAAM,EAAE,EAAE,EAAE,QAAQ,GAAG,CAAC;GAC1E,MAAM,eAAe,oBACnB,KAAK,KACL,mBAAmB,GAAG,UACtB,iBACD;GAED,MAAM,gBAAgB;IACpB,EAAE,QAAQ,SAAS;IACnB,EAAE,QAAQ,SAAS;IACnB,EAAE,QAAQ,IAAI;IACd,EAAE,QAAQ,IAAI;IACd,EAAE,QAAQ,IAAI;IACd,EAAE,QAAQ,QAAQ;IAClB,EAAE,QAAQ,cAAc;IACzB;GAED,MAAM,iBAAiB,gBACrB,KAAK,KACL,eACA,eACA,cAAc,GAAG,SAClB;GAED,MAAMA,aAA4B;IAChC,QAAQ,iBAAiB,GAAG;IAC5B,MAAM,GAAG;IACT,MAAM;IACN,UAAU;IACV,WAAW;IACX,eAAe;IACf,iBAAiB;IACjB,kBAAkB;IACnB;AAED,QAAK,cAAc,OAAO,GAAG,GAAG,WAAW;AAC3C;;AAGF,MAAI,cAAc,EAChB,SAAQ,IACN,oBAAoB,YAAY,8BAA8B,YAAY,cAC3E;;;;;;;;;;;;;;CAgBL,AAAQ,mCAAyC;EAC/C,IAAI,cAAc;AAElB,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,cAAc,SAAS,GAAG,KAAK;GACtD,MAAM,KAAK,KAAK,cAAc;GAC9B,MAAM,KAAK,KAAK,cAAc,IAAI;AAElC,OAAI,GAAG,KAAK,WAAW,SAAU;AACjC,OAAI,GAAG,KAAK,WAAW,aAAc;GAErC,MAAM,YAAY,GAAG,KAAK,QAAQ;AAClC,OAAI,GAAG,KAAK,OAAO,OAAO,UAAW;GAGrC,MAAM,WAAW,GAAG,KAAK,OAAO;GAChC,MAAM,SAAS,GAAG,KAAK,OAAO;GAC9B,MAAM,UAAU,KAAK,UAAU,SAAS;GACxC,MAAM,QAAQ,KAAK,UAAU,OAAO;GACpC,MAAM,KAAK,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAC5C,MAAM,KAAK,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAC5C,MAAM,KAAK,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAC5C,MAAM,SAAS,KAAK,UAAU,GAAG,KAAK,QAAQ,GAAG;AACjD,OAAI,CAAC,WAAW,CAAC,SAAS,CAAC,MAAM,CAAC,MAAM,CAAC,MAAM,CAAC,OAAQ;AAIxD,OAAI,WAAW,WAAW,WAAW,SAAS,YAAY,MAAO;GAEjE,MAAM,YAAY;GAClB,MAAM,gBAAgB,oBACpB,KAAK,KACL,gBAAgB,GAAG,UACnB,UAAU,YACV,UAAU,WACX;GAED,MAAM,mBAAmB,UAAU,YAAY,GAAG,MAAM,EAAE,EAAE,EAAE,QAAQ,GAAG,CAAC;GAC1E,MAAM,eAAe,oBACnB,KAAK,KACL,wBAAwB,GAAG,UAC3B,iBACD;GAED,MAAM,gBAAgB;IACpB,EAAE,QAAQ,SAAS;IACnB,EAAE,QAAQ,OAAO;IACjB,EAAE,QAAQ,IAAI;IACd,EAAE,QAAQ,IAAI;IACd,EAAE,QAAQ,IAAI;IACd,EAAE,QAAQ,QAAQ;IAClB,EAAE,QAAQ,cAAc;IACzB;GAED,MAAM,iBAAiB,gBACrB,KAAK,KACL,eACA,eACA,mBAAmB,GAAG,SACvB;GAED,MAAMA,aAA4B;IAChC,QAAQ,sBAAsB,GAAG;IACjC,MAAM,GAAG;IACT,MAAM;IACN,UAAU;IACV,WAAW;IACX,eAAe;IACf,iBAAiB;IACjB,kBAAkB;IACnB;AAED,QAAK,cAAc,OAAO,GAAG,GAAG,WAAW;AAC3C;;AAGF,MAAI,cAAc,EAChB,SAAQ,IACN,oBAAoB,YAAY,mCAAmC,YAAY,cAChF;;;;;;;;;;;;CAcL,AAAQ,+BAAqC;EAC3C,IAAI,cAAc;AAElB,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,cAAc,SAAS,GAAG,KAAK;GACtD,MAAM,KAAK,KAAK,cAAc;GAC9B,MAAM,KAAK,KAAK,cAAc,IAAI;AAElC,OAAI,GAAG,KAAK,WAAW,aAAa,GAAG,KAAK,WAAW,UAAW;AAElE,OAAI,GAAG,KAAK,WAAW,gBAAgB,GAAG,KAAK,WAAW,YAAa;AACvE,OAAI,GAAG,KAAK,WAAW,QAAQ,GAAG,KAAK,WAAW,IAAK;GAEvD,MAAM,MAAM,GAAG,KAAK,OAAO;GAC3B,MAAM,KAAK,GAAG,KAAK,OAAO;GAC1B,MAAM,OAAO,GAAG,KAAK,QAAQ;GAC7B,MAAM,MAAM,GAAG,KAAK,OAAO;GAC3B,MAAM,KAAK,GAAG,KAAK,OAAO;GAC1B,MAAM,OAAO,GAAG,KAAK,QAAQ;GAE7B,MAAM,SAAS,KAAK,UAAU,IAAI;GAClC,MAAM,QAAQ,KAAK,UAAU,GAAG;GAChC,MAAM,UAAU,KAAK,UAAU,KAAK;GACpC,MAAM,SAAS,KAAK,UAAU,IAAI;GAClC,MAAM,QAAQ,KAAK,UAAU,GAAG;GAChC,MAAM,UAAU,KAAK,UAAU,KAAK;AACpC,OAAI,CAAC,UAAU,CAAC,SAAS,CAAC,WAAW,CAAC,UAAU,CAAC,SAAS,CAAC,QAAS;AAQpE,OAAI,YAAY,QAAS;AACzB,OAAI,WAAW,WAAW,WAAW,QAAS;GAI9C,MAAME,YAAoB;IACxB,IAAI,kBAAkB,GAAG;IACzB,QAAQ;IACR,QAAQ;KAAC;KAAK;KAAI;KAAK;KAAG;IAC1B,SAAS,CAAC,MAAM,KAAK;IACrB,YAAY;KACV,aAAa,GAAG,KAAK,WAAW;KAChC,KAAK,GAAG,KAAK,WAAW;KACxB,iBAAiB,GAAG,KAAK,WAAW;KACpC,iBAAiB,GAAG,KAAK,WAAW;KACpC,UAAU,GAAG,KAAK,WAAW;KAC7B,UAAU,GAAG,KAAK,WAAW;KAC9B;IACF;GAED,MAAM,YAAY;GAClB,MAAM,gBAAgB,oBACpB,KAAK,KACL,aAAa,GAAG,UAChB,UAAU,YACV,UAAU,WACX;GAED,MAAM,cAAc,KAAK,cAAc,EAAE;GACzC,MAAM,mBAAmB,UAAU,YAAY,WAAW,aAAa,EAAE,QAAQ,GAAG,CAAC;GACrF,MAAM,eAAe,oBACnB,KAAK,KACL,qBAAqB,GAAG,UACxB,iBACD;GAED,MAAM,gBAAgB;IACpB,EAAE,QAAQ,QAAQ;IAClB,EAAE,QAAQ,OAAO;IACjB,EAAE,QAAQ,QAAQ;IAClB,EAAE,QAAQ,OAAO;IACjB,EAAE,QAAQ,SAAS;IACnB,EAAE,QAAQ,SAAS;IACnB,EAAE,QAAQ,cAAc;IACzB;GAED,MAAM,iBAAiB,gBACrB,KAAK,KACL,eACA,eACA,gBAAgB,GAAG,SACpB;GAED,MAAMF,aAA4B;IAChC,QAAQ,UAAU;IAClB,MAAM;IACN,MAAM;IACN,UAAU;IACV,WAAW;IACX,eAAe;IACf,iBAAiB;IACjB,kBAAkB;IACnB;AAED,QAAK,cAAc,OAAO,GAAG,GAAG,WAAW;AAC3C;;AAGF,MAAI,cAAc,EAChB,SAAQ,IACN,oBAAoB,YAAY,wBAAwB,YAAY,cACrE;;;;;;CAQL,AAAQ,cACN,MACA,MACA,eAC8B;EAC9B,MAAMG,UAAwC,EAAE;EAChD,MAAM,iBAAiB,CAAC,GAAG,KAAK,QAAQ,GAAG,KAAK,QAAQ;EAIxD,MAAM,gCAAgB,IAAI,KAAgB;AAE1C,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,SAAS,QAAQ,KAAK;GAC7C,MAAM,UAAU,KAAK,SAAS;AAE9B,OAAI,QAAQ,SAAS,UACnB,SAAQ,KAAK,EAAE,QAAQ,eAAe,CAAC;QAClC;IAEL,MAAM,aAAa,eADD,IAAI,KAAK,SAAS,QAAQ,GAAG,MAAM,IAAI,KAAK,EAAE,SAAS,UAAU,CAAC;AAEpF,QAAI,CAAC,WAAY,OAAM,IAAI,MAAM,8BAA8B,EAAE,SAAS,KAAK,KAAK;IAEpF,IAAIC;AACJ,QAAI,eAAe,YACjB,UAAS,KAAK;QAEd,UAAS,KAAK,UAAU,WAAW;AAErC,QAAI,CAAC,OAAQ,OAAM,IAAI,MAAM,6BAA6B,WAAW,UAAU,KAAK,KAAK;AAOzF,QAAI,QAAQ,SAAS,qBACnB,KAAI,cAAc,IAAI,OAAO,CAC3B,UAAS,KAAK,wBAAwB,OAAO,KAAK;QAElD,eAAc,IAAI,OAAO;AAI7B,YAAQ,KAAK,EAAE,QAAQ,CAAC;;;AAI5B,SAAO;;;CAIT,AAAQ,wBAAwB,UAA6B;AAC3D,MAAI,CAAC,KAAK,wBAAwB,KAAK,qBAAqB,OAAO,SACjE,MAAK,uBAAuB,oBAC1B,KAAK,KACL,mBACA,KAAK,IAAI,UAAU,IAAI,CACxB;AAEH,SAAO,KAAK;;;;;;;;;;;;ACtzEhB,SAAgB,WAAW,MAAsC;CAC/D,MAAM,EAAE,SAAS,QAAQ,YAAY,QAAQ,GAAG,GAAG,cAAc;CACjE,MAAM,gBAAgB,IAAI;CAC1B,MAAM,eAAe,KAAK,KAAK,IAAI,UAAU;CAC7C,MAAM,cAAc,IAAI;CAExB,MAAM,SAAS,IAAI,YAAY,KAAK,KAAK,gBAAgB,EAAE,CAAC;CAC5D,MAAM,SAAS,IAAI,aAAa,YAAY;CAC5C,MAAM,QAAQ,IAAI,aAAa,YAAY;AAK3C,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,MAAM,UAAU,MAAM;EACtB,MAAM,gBAAgB,IAAI;EAC1B,MAAM,cAAc,UAAU;AAC9B,OAAK,IAAI,MAAM,GAAG,MAAM,GAAG,OAAO;GAGhC,MAAM,SADW,QAAQ,cAAc,SACT,gBAAgB,IAAM;GAGpD,MAAM,UAAU,MAAM,IAAI;GAC1B,MAAM,YAAY,YAAY;GAC9B,MAAM,YAAY,UAAU;AAC5B,UAAO,cAAc,UAAW,YAAY;;;AAOhD,MAAK,IAAI,IAAI,GAAG,IAAI,cAAc,KAAK;EACrC,MAAM,SAAS,IAAI;AACnB,OAAK,IAAI,MAAM,GAAG,MAAM,GAAG,MACzB,QAAO,MAAM,eAAe,KAAK,WAAW,SAAS;;CAYzD,MAAM,aAAa,MAAM;AACzB,MAAK,IAAI,IAAI,GAAG,IAAI,cAAc,KAAK;EACrC,MAAM,SAAS,IAAI;AACnB,OAAK,IAAI,MAAM,GAAG,MAAM,GAAG,OAAO;GAGhC,MAAM,UAFQ,OAAO,UAAU,QAAQ,SACpB,MAAM,KACgB,IAAM;AAC/C,SAAM,MAAM,eAAe,KAAK,UAAU;;;AAI9C,QAAO;EAAE;EAAQ;EAAQ;EAAO;;;;;;;;;;;;;;;;;;;;;;;AChFlC,MAAM,UAAU;AAChB,MAAM,QAAQ;AAEd,SAAS,eAAwB;AAC/B,QAAO,OAAO,cAAc;;AAG9B,IAAIC,aAAiD;AAErD,SAAS,SAAsC;AAC7C,KAAI,WAAY,QAAO;AACvB,cAAa,IAAI,SAA6B,YAAY;AACxD,MAAI,CAAC,cAAc,EAAE;AACnB,WAAQ,KAAK;AACb;;AAEF,MAAI;GACF,MAAM,MAAM,UAAU,KAAK,SAAS,EAAE;AACtC,OAAI,wBAAwB;IAC1B,MAAM,KAAK,IAAI;AACf,QAAI,CAAC,GAAG,iBAAiB,SAAS,MAAM,CAAE,IAAG,kBAAkB,MAAM;;AAEvE,OAAI,kBAAkB,QAAQ,IAAI,OAAO;AACzC,OAAI,gBAAgB,QAAQ,KAAK;AACjC,OAAI,kBAAkB,QAAQ,KAAK;UAC7B;AACN,WAAQ,KAAK;;GAEf;AACF,QAAO;;;AAIT,SAAS,GACP,MACA,KACA,UACY;AACZ,QAAO,QAAQ,CAAC,MACb,OACC,IAAI,SAAY,YAAY;AAC1B,MAAI,CAAC,IAAI;AACP,WAAQ,SAAS;AACjB;;AAEF,MAAI;GACF,MAAM,IAAI,GAAG,YAAY,OAAO,KAAK;GACrC,MAAM,MAAM,IAAI,EAAE,YAAY,MAAM,CAAC;AACrC,OAAI,kBAAkB,QAAS,IAAI,UAAgB,SAAS;AAC5D,OAAI,gBAAgB,QAAQ,SAAS;AACrC,KAAE,gBAAgB,QAAQ,SAAS;UAC7B;AACN,WAAQ,SAAS;;GAEnB,CACL;;;AAIH,eAAsB,OAAO,KAA0C;CACrE,MAAM,IAAI,MAAM,GAAuB,aAAa,MAAM,EAAE,IAAI,IAAI,EAAE,KAAK;AAC3E,QAAO,aAAa,cAAc,IAAI;;;;;;;;AAexC,eAAsB,aAAmC;CACvD,MAAM,OAAO,MAAM,GAAkB,aAAa,MAAM,EAAE,YAAY,EAAE,EAAE,CAAC;AAC3E,QAAO,IAAI,KAAK,QAAQ,EAAE,EAAE,KAAK,MAAM,OAAO,EAAE,CAAC,CAAC;;;;AAKpD,eAAsB,OAAO,KAAa,KAAoC;AAG5E,QADU,MAAM,GAAuB,cAAc,MAAM,EAAE,IAAI,KAAK,IAAI,EAAE,KAAK,KACpE;;;AASf,SAAgB,oBAA6B;AAC3C,QAAO,cAAc;;;;;;;;;;;;ACzEvB,SAAgB,UAAU,KAAoC;CAC5D,MAAM,EAAE,QAAQ,QAAQ,WAAW,QAAQ,WAAW,GAAG,GAAG,cAAc;CAE1E,MAAM,cAAc,IADC,KAAK,KAAK,IAAI,UAAU;CAK7C,MAAM,SAAS;CAMf,MAAM,SAAS,IAAI,aAAa,YAAY;CAC5C,MAAM,QAAQ,IAAI,aAAa,YAAY;AAI3C,MAAK,IAAI,IAAI,GAAG,IAAI,aAAa,KAAK;EACpC,MAAM,IAAI,UAAU;EACpB,MAAM,IAAI,UAAU;AACpB,SAAO,KAAK;AAEZ,QAAM,KAAK,MAAM,IAAI,CAAC,IAAI,IAAI;;AAGhC,QAAO;EAAE;EAAQ;EAAQ;EAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AC9BlC,MAAM,gBAAgB;;AAGtB,MAAM,oBAAoB,IAAI,OAAO;;;;;;AAOrC,SAAgB,gBAAyB;AACvC,QACE,OAAO,cAAc,eACrB,OAAQ,UAAuD,SAAS,iBACtE,cACF,OAAO,WAAW,eAClB,OAAO,SAAS,eAChB,OAAO,QAAQ,eACf,OAAO,IAAI,oBAAoB;;;;;;;;AAUnC,SAAS,cAAc,KAAqB;CAE1C,IAAI,KAAK;CACT,IAAI,KAAK;AACT,MAAK,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;EACnC,MAAM,IAAI,IAAI,WAAW,EAAE;AAC3B,QAAM;AACN,OAAK,KAAK,KAAK,IAAI,SAAc;AACjC,QAAM,IAAI;AACV,OAAK,KAAK,KAAK,IAAI,WAAc;;AAInC,QAAO,IAFI,OAAO,GAAG,SAAS,GAAG,CAAC,SAAS,GAAG,IAAI,IACvC,OAAO,GAAG,SAAS,GAAG,CAAC,SAAS,GAAG,IAAI,CAClC,GAAG,IAAI,OAAO,SAAS,GAAG,CAAC;;AAS7C,MAAM,gBAAgB;mBACH,KAAK,UAAU,cAAc,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;oBAsD7B,kBAAkB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAiEtC,IAAM,aAAN,MAAiB;CACf,AAAQ,SAAwB;CAChC,AAAQ,UAAyB;CACjC,AAAQ,0BAAU,IAAI,KAA8B;CACpD,AAAQ,SAAS;CACjB,AAAQ,SAAS;CAEjB,AAAQ,eAA8B;AACpC,MAAI,KAAK,OAAQ,QAAO;AACxB,MAAI,KAAK,OAAQ,QAAO,KAAK;AAC7B,MAAI;GACF,MAAM,OAAO,IAAI,KAAK,CAAC,cAAc,EAAE,EAAE,MAAM,0BAA0B,CAAC;AAC1E,QAAK,UAAU,IAAI,gBAAgB,KAAK;GACxC,MAAM,SAAS,IAAI,OAAO,KAAK,QAAQ;AACvC,UAAO,aAAa,OAAqB;IACvC,MAAM,EAAE,IAAI,IAAI,SAAS,KAAK,UAAU,GAAG;IAO3C,MAAM,IAAI,KAAK,QAAQ,IAAI,GAAG;AAC9B,QAAI,CAAC,EAAG;AACR,SAAK,QAAQ,OAAO,GAAG;AACvB,QAAI,GAAI,GAAE,QAAQ;KAAE;KAAS;KAAK,CAAC;QAC9B,GAAE,OAAO,IAAI,MAAM,SAAS,oBAAoB,CAAC;;AAExD,UAAO,gBAAgB;AAErB,SAAK,SAAS;AACd,SAAK,MAAM,KAAK,KAAK,QAAQ,QAAQ,CACnC,GAAE,uBAAO,IAAI,MAAM,sBAAsB,CAAC;AAE5C,SAAK,QAAQ,OAAO;;AAEtB,QAAK,SAAS;AACd,UAAO;UACD;AACN,QAAK,SAAS;AACd,UAAO;;;CAIX,AAAQ,KACN,IACA,MACA,KAC0D;EAC1D,MAAM,SAAS,KAAK,cAAc;AAClC,MAAI,CAAC,OAAQ,QAAO,QAAQ,uBAAO,IAAI,MAAM,mBAAmB,CAAC;EACjE,MAAM,KAAK,KAAK;AAChB,SAAO,IAAI,SAAS,SAAS,WAAW;AACtC,QAAK,QAAQ,IAAI,IAAI;IAAE;IAAS;IAAQ,CAAC;AACzC,OAAI;AACF,QAAI,IAEF,QAAO,YAAY;KAAE;KAAI;KAAI;KAAM;KAAK,EAAE,CAAC,IAAI,CAAC;QAEhD,QAAO,YAAY;KAAE;KAAI;KAAI;KAAM,CAAC;YAE/B,GAAG;AACV,SAAK,QAAQ,OAAO,GAAG;AACvB,WAAO,aAAa,QAAQ,IAAI,IAAI,MAAM,OAAO,EAAE,CAAC,CAAC;;IAEvD;;CAGJ,MAAM,IAAI,KAA+B;EACvC,MAAM,EAAE,YAAY,MAAM,KAAK,KAAK,OAAO,cAAc,IAAI,CAAC;AAC9D,SAAO,QAAQ,QAAQ;;CAGzB,MAAM,KAAK,KAA0C;EACnD,MAAM,EAAE,QAAQ,MAAM,KAAK,KAAK,QAAQ,cAAc,IAAI,CAAC;AAC3D,SAAO,OAAO;;CAGhB,MAAM,MAAM,KAAa,KAAiC;AACxD,QAAM,KAAK,KAAK,SAAS,cAAc,IAAI,EAAE,IAAI;;CAGnD,MAAM,OAAO,KAA4B;AACvC,QAAM,KAAK,KAAK,UAAU,cAAc,IAAI,CAAC;;;AAIjD,IAAIC,UAA6B;AACjC,SAAS,SAAqB;AAC5B,KAAI,CAAC,QAAS,WAAU,IAAI,YAAY;AACxC,QAAO;;;;;;AAoBT,eAAsB,SAAS,KAA0C;AACvE,KAAI,CAAC,eAAe,CAAE,QAAO;AAC7B,KAAI;AACF,SAAO,MAAM,QAAQ,CAAC,KAAK,IAAI;SACzB;AACN,SAAO;;;;;;;;;;;AAsCX,eAAsB,iBAAgC;AACpD,KAAI;EACF,MAAM,OAAQ,MACZ,WACC,SAAS,gBAAgB;AAM5B,MAAI,CAAC,KAAM;EACX,MAAMC,QAAkB,EAAE;AAC1B,aAAW,MAAM,CAAC,MAAM,WAAW,KAAK,WAAW,IAAI,EAAE,CACvD,KACE,OAAO,SAAS,eAChB,KAAK,WAAW,iBAAiB,IACjC,SAAS,cAET,OAAM,KAAK,KAAK;AAGpB,OAAK,MAAM,QAAQ,MACjB,OAAM,KAAK,YAAY,MAAM,EAAE,WAAW,MAAM,CAAC,CAAC,YAAY,GAE5D;SAEE;;;;;;;;;;;AC/UV,SAAgB,uBAAuB,QAAsC;AAC3E,KAAI,OAAO,aAAa,EACtB,OAAM,IAAI,MACR,4EAA4E,OAAO,WAAW,IAC/F;CAGH,MAAM,OAAO,IAAI,SAAS,OAAO;CAIjC,MAAM,eAAe,OAAO,KAAK,aAAa,GAAG,KAAK,CAAC;AAEvD,KAAI,eAAe,OAAO,aAAa,EACrC,OAAM,IAAI,MACR,8BAA8B,aAAa,yBAAyB,OAAO,aAAa,EAAE,6BAC7D,eAAe,EAAE,SAC/C;CAIH,MAAM,cAAc,IAAI,WAAW,QAAQ,GAAG,aAAa;CAC3D,MAAM,YAAY,IAAI,aAAa,CAAC,OAAO,YAAY;CACvD,MAAMC,SAAkC,KAAK,MAAM,UAAU;CAE7D,MAAM,YAAY,IAAI;CACtB,MAAMC,UAA6B,EAAE;CACrC,IAAIC,WAA0C;AAE9C,MAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,OAAO,EAAE;AACjD,MAAI,SAAS,gBAAgB;AAC3B,cAAW;AACX;;EAGF,MAAM,EAAE,OAAO,OAAO,iBAAiB;AAMvC,UAAQ,KAAK;GACX;GACA;GACA;GACA,YAAY,aAAa;GACzB,YAAY,aAAa,KAAK,aAAa;GAC5C,CAAC;;AAIJ,SAAQ,MAAM,GAAG,MAAM,EAAE,aAAa,EAAE,WAAW;AAEnD,QAAO;EAAE;EAAc;EAAW;EAAS;EAAU;;;AAIvD,SAAS,eAAe,OAAgC;AACtD,SAAQ,OAAR;EACE,KAAK;EACL,KAAK;EACL,KAAK,MACH,QAAO;EACT,KAAK;EACL,KAAK;EACL,KAAK,MACH,QAAO;EACT,KAAK;EACL,KAAK;EACL,KAAK;EACL,KAAK,MACH,QAAO;EACT,KAAK;EACL,KAAK;EACL,KAAK,OACH,QAAO;EACT,QACE,QAAO;;;;;;;;;;AAWb,SAAS,cACP,QACA,QACA,YACA,OACiB;CAEjB,MAAM,UAAU,SADF,eAAe,MAAM,KACA;CAGnC,MAAM,MAAM,UAAU,SAAS,OAAO,MAAM,QAAQ,SAAS,WAAW;CACxE,MAAM,OAAO,UAAU,SAAS;AAEhC,SAAQ,OAAR;EACE,KAAK,MACH,QAAO,IAAI,aAAa,KAAK,MAAM,aAAa,EAAE;EACpD,KAAK,MAEH,QAAO,IAAI,YAAY,KAAK,MAAM,aAAa,EAAE;EACnD,KAAK,OAEH,QAAO,IAAI,YAAY,KAAK,MAAM,aAAa,EAAE;EACnD,KAAK,MACH,QAAO,IAAI,WAAW,KAAK,MAAM,aAAa,EAAE;EAClD,KAAK,MACH,QAAO,IAAI,YAAY,KAAK,MAAM,aAAa,EAAE;EACnD,KAAK,KACH,QAAO,IAAI,UAAU,KAAK,MAAM,WAAW;EAC7C,KAAK,KACH,QAAO,IAAI,WAAW,KAAK,MAAM,WAAW;EAC9C,KAAK,MACH,QAAO,IAAI,WAAW,KAAK,MAAM,aAAa,EAAE;EAClD,KAAK,MACH,QAAO,IAAI,YAAY,KAAK,MAAM,aAAa,EAAE;EACnD,KAAK,MACH,QAAO,IAAI,aAAa,KAAK,MAAM,aAAa,EAAE;EACpD,KAAK,MACH,QAAO,IAAI,cAAc,KAAK,MAAM,aAAa,EAAE;EACrD,KAAK,MACH,QAAO,IAAI,eAAe,KAAK,MAAM,aAAa,EAAE;EACtD,KAAK,OACH,QAAO,IAAI,WAAW,KAAK,MAAM,WAAW;EAC9C,QACE,OAAM,IAAI,MAAM,kCAAkC,QAAQ;;;;;;;AAQhE,SAAgB,cACd,QACA,MACA,OACiB;AAEjB,QAAO,cAAc,QADN,KAAK,YAAY,MAAM,YACD,MAAM,YAAY,MAAM,MAAM;;;;;;AAuCrE,SAAgB,UAAU,WAAqC;CAC7D,MAAM,OAAO,IAAI,YAAY,UAAU,QAAQ,UAAU,YAAY,UAAU,aAAa,EAAE;CAC9F,MAAM,MAAM,IAAI,aAAa,KAAK,OAAO;CACzC,MAAM,MAAM,IAAI,YAAY,IAAI,OAAO;AACvC,MAAK,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,IAC/B,KAAI,KAAK,KAAK,MAAM;AAEtB,QAAO;;;;;AAMT,SAAgB,SAAS,SAAwC;CAC/D,MAAM,MACJ,mBAAmB,cACf,UACA,IAAI,YAAY,QAAQ,QAAQ,QAAQ,YAAY,QAAQ,aAAa,EAAE;CACjF,MAAM,MAAM,IAAI,aAAa,IAAI,OAAO;AACxC,MAAK,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;EACnC,MAAM,IAAI,IAAI;EACd,MAAM,OAAQ,KAAK,KAAM;EACzB,MAAM,MAAO,KAAK,KAAM;EACxB,MAAM,OAAO,IAAI;AACjB,MAAI,QAAQ,EAEV,KAAI,MAAM,OAAO,KAAK,KAAK,KAAK,OAAO,OAAO;WACrC,QAAQ,GAEjB,KAAI,KAAK,SAAS,IAAK,OAAO,YAAY,WAAY;MAEtD,KAAI,MAAM,OAAO,KAAK,KAAK,MAAM,MAAM,OAAO,IAAI,OAAO;;AAG7D,QAAO;;;;;;;;;;;ACnQT,MAAM,kCAAkB,IAAI,KAAqB;AACjD,MAAM,kCAAkB,IAAI,KAAqB;AAEjD;CAEE,MAAMC,KAAe,EAAE;AACvB,MAAK,IAAI,IAAI,IAAI,KAAK,KAAK,IAAK,IAAG,KAAK,EAAE;AAC1C,MAAK,IAAI,IAAI,KAAK,KAAK,KAAK,IAAK,IAAG,KAAK,EAAE;AAC3C,MAAK,IAAI,IAAI,KAAK,KAAK,KAAK,IAAK,IAAG,KAAK,EAAE;CAE3C,MAAM,KAAK,CAAC,GAAG,GAAG;CAGlB,IAAI,IAAI;AACR,MAAK,IAAI,IAAI,GAAG,IAAI,KAAK,IACvB,KAAI,CAAC,GAAG,SAAS,EAAE,EAAE;AACnB,KAAG,KAAK,EAAE;AACV,KAAG,KAAK,MAAM,EAAE;AAChB;;AAIJ,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,QAAQ,KAAK;AAClC,kBAAgB,IAAI,GAAG,IAAI,GAAG,GAAG;AACjC,kBAAgB,IAAI,GAAG,IAAI,GAAG,GAAG;;;AAmBrC,IAAa,YAAb,MAAa,UAAU;CACrB,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;;;;;;CAMR,AAAQ;CACR,AAAS;CACT,AAAS;CAET,AAAQ,YACN,OACA,cACA,QACA,eACA,aACA,QACA,WACA,SACA;AACA,OAAK,QAAQ;AACb,OAAK,eAAe;AACpB,OAAK,SAAS;AACd,OAAK,gBAAgB;AACrB,OAAK,cAAc;AACnB,OAAK,+BAAe,IAAI,KAAK;AAC7B,OAAK,UAAU;AACf,OAAK,SAAS;AACd,OAAK,YAAY;AAIjB,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,KAAK;GAE5B,MAAM,MAAM,MADA,EAAE,SAAS,GAAG,CAAC,aAAa,CAAC,SAAS,GAAG,IAAI,CACnC;AACtB,OAAI,MAAM,IAAI,IAAI,CAChB,MAAK,aAAa,IAAI,GAAG,IAAI;;;;;;CAQnC,OAAO,SAAS,eAAoB,qBAAsC;EACxE,MAAM,QAAQ,cAAc;AAC5B,MAAI,CAAC,SAAS,MAAM,SAAS,MAC3B,OAAM,IAAI,MACR,+BAA+B,OAAO,QAAQ,UAAU,0BACzD;EAIH,MAAM,wBAAQ,IAAI,KAAqB;EACvC,MAAM,+BAAe,IAAI,KAAqB;AAC9C,OAAK,MAAM,CAAC,OAAO,OAAO,OAAO,QAAQ,MAAM,MAAgC,EAAE;AAC/E,SAAM,IAAI,OAAO,GAAG;AACpB,gBAAa,IAAI,IAAI,MAAM;;EAO7B,MAAM,iCAAiC;GACrC,MAAM,OAAO,cAAc;GAC3B,MAAM,YAAY,MAChB,GAAG,SAAS,aAAa,GAAG,SAAS,WAAW,OAAO,GAAG,YAAY;AACxE,OAAI,CAAC,KAAM,QAAO;AAClB,OAAI,SAAS,KAAK,CAAE,QAAO;AAC3B,OAAI,MAAM,QAAQ,KAAK,YAAY,CAAE,QAAO,KAAK,YAAY,KAAK,SAAS;AAC3E,UAAO;MACL;EACJ,MAAM,mBAAmB,MAAM,kBAAkB;EACjD,MAAM,UAAU,2BAA4B,oBAAoB,UAAU,MAAM;EAIhF,MAAM,yBAAS,IAAI,KAAqB;EACxC,MAAM,YAAa,MAAM,UAAU,EAAE;AACrC,OAAK,IAAI,IAAI,GAAG,IAAI,UAAU,QAAQ,KAAK;GACzC,MAAM,IAAI,UAAU;GACpB,MAAM,MAAM,MAAM,QAAQ,EAAE,GAAG,GAAG,EAAE,GAAG,GAAG,EAAE,OAAO;AACnD,UAAO,IAAI,KAAK,EAAE;;EAKpB,MAAM,gCAAgB,IAAI,KAAqB;EAC/C,MAAM,iCAAiB,IAAI,KAAqB;EAChD,MAAM,kBAAkB,cAAc;AAGtC,MAAI,gBACF,MAAK,MAAM,KAAK,iBAAiB;AAC/B,SAAM,IAAI,EAAE,SAAS,EAAE,GAAG;AAC1B,gBAAa,IAAI,EAAE,IAAI,EAAE,QAAQ;AACjC,kBAAe,IAAI,EAAE,SAAS,EAAE,GAAG;AACnC,OAAI,EAAE,QACJ,eAAc,IAAI,EAAE,SAAS,EAAE,GAAG;;EAMxC,MAAM,WAAW,qBAAqB,aAAa;EACnD,MAAM,WAAW,qBAAqB,aAAa;EACnD,MAAM,eAAe,qBAAqB,iBAAiB;EAC3D,MAAM,cAAc,qBAAqB,iBAAiB;EAC1D,MAAM,cAAc,qBAAqB,iBAAiB;AAY1D,SAAO,IAAI,UACT,OACA,cACA,QACA,eACA,gBAf8B;GAC9B;GACA;GACA,YAAY,WAAY,MAAM,IAAI,SAAS,IAAI,OAAQ;GACvD,YAAY,WAAY,MAAM,IAAI,SAAS,IAAI,OAAQ;GACvD;GACA;GACA;GACD,EASC,MAAM,MACN,QACD;;;;;;CAOH,UAAU,OAA8B;AACtC,SAAO,KAAK,MAAM,IAAI,MAAM,IAAI;;;;;CAMlC,OAAO,MAAwB;AAC7B,MAAI,CAAC,KAAM,QAAO,EAAE;EAEpB,MAAMC,MAAgB,EAAE;AAGxB,MAAI,KAAK,OAAO,eAAe,KAAK,OAAO,eAAe,KACxD,KAAI,KAAK,KAAK,OAAO,WAAW;EAIlC,MAAM,QAAQ,KAAK,qBAAqB,KAAK;AAE7C,OAAK,MAAM,QAAQ,MACjB,KAAI,KAAK,SAAS;GAChB,MAAM,KAAK,KAAK,YAAY,IAAI,KAAK,KAAK;AAC1C,OAAI,OAAO,OACT,KAAI,KAAK,GAAG;SAET;GAGL,MAAM,SAAS,KAAK,YAAY,KAAK,KAAK;AAC1C,QAAK,MAAM,SAAS,QAAQ;IAC1B,MAAM,WAAW,KAAK,UAAU,MAAM;AACtC,QAAI,KAAK,GAAG,SAAS;;;AAM3B,MAAI,KAAK,OAAO,eAAe,KAAK,OAAO,eAAe,KACxD,KAAI,KAAK,KAAK,OAAO,WAAW;AAGlC,SAAO;;;;;CAMT,OAAO,KAAe,oBAA6B,MAAc;EAC/D,MAAMC,SAAmB,EAAE;AAE3B,OAAK,MAAM,MAAM,KAAK;GACpB,MAAM,QAAQ,KAAK,aAAa,IAAI,GAAG;AACvC,OAAI,CAAC,MAAO;AAGZ,OAAI,qBAAqB,KAAK,cAAc,IAAI,MAAM,CACpD;AAGF,UAAO,KAAK,MAAM;;AAGpB,MAAI,KAAK,SAAS;GAIhB,IAAI,MAAM;GACV,IAAIC,UAAoB,EAAE;GAC1B,MAAM,mBAAmB;AACvB,QAAI,QAAQ,WAAW,EAAG;AAC1B,WAAO,IAAI,aAAa,CAAC,OAAO,IAAI,WAAW,QAAQ,CAAC;AACxD,cAAU,EAAE;;AAEd,QAAK,MAAM,SAAS,QAAQ;IAC1B,MAAM,IAAI,yBAAyB,KAAK,MAAM;AAC9C,QAAI,EACF,SAAQ,KAAK,OAAO,SAAS,EAAE,IAAI,GAAG,CAAC;SAClC;AACL,iBAAY;AACZ,YAAO;;;AAGX,eAAY;AACZ,UAAO,IAAI,QAAQ,MAAM,IAAI;;EAS/B,IAAI,OAAO,OAAO,KAAK,GAAG;EAG1B,MAAMC,QAAkB,EAAE;AAC1B,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;GACpC,MAAM,KAAK,KAAK,YAAY,EAAE;GAC9B,MAAM,OAAO,gBAAgB,IAAI,GAAG;AACpC,OAAI,SAAS,OACX,OAAM,KAAK,KAAK;QACX;IAEL,MAAM,UAAU,IAAI,aAAa,CAAC,OAAO,OAAO,cAAc,GAAG,CAAC;AAClE,SAAK,MAAM,KAAK,QAAS,OAAM,KAAK,EAAE;AACtC,QAAI,KAAK,MAAQ;;;AAGrB,SAAO,IAAI,aAAa,CAAC,OAAO,IAAI,WAAW,MAAM,CAAC;AAGtD,SAAO,KAAK,QAAQ,0BAA0B,GAAG,QAAQ;AACvD,UAAO,OAAO,aAAa,SAAS,KAAK,GAAG,CAAC;IAC7C;AAEF,SAAO;;;;;;;;;;;;;;;;;CAkBT,AAAQ,uBAAuB,UAAyB,cAA+B;EACrF,MAAMC,QAAkB,EAAE;AAC1B,MAAI,KAAK,OAAO,SAAU,OAAM,KAAK,KAAK,OAAO,SAAS;EAC1D,IAAI,aAAa;AACjB,OAAK,MAAM,OAAO,UAAU;AAC1B,OAAI,IAAI,SAAS,UAAU;AACzB,kBAAc,GAAG,IAAI,QAAQ;AAC7B;;GAEF,MAAM,OAAO,IAAI,SAAS,cAAc,UAAU,IAAI;GACtD,MAAM,UAAU,SAAS,UAAU,aAAa,aAAa,IAAI,UAAU,IAAI;AAC/E,OAAI,SAAS,OAAQ,cAAa;AAClC,SAAM,KAAK,UAAU,KAAK,IAAI,QAAQ,WAAW;;AAEnD,MAAI,aAAc,OAAM,KAAK,iBAAiB;AAC9C,SAAO,MAAM,KAAK,GAAG;;CAGvB,kBAAkB,UAAyB,SAAqD;EAC9F,MAAM,eAAe,SAAS,uBAAuB;EACrD,MAAMA,QAAkB,EAAE;AAK1B,MAAI,KAAK,YAAY,IAAI,UAAU,IAAI,KAAK,YAAY,IAAI,UAAU,CACpE,QAAO,KAAK,uBAAuB,UAAU,aAAa;EAK5D,MAAM,WAAW,KAAK,OAAO;AAC7B,MAAI,UAAU,SAAS,YAAY,IAAI,KAAK,OAAO,SACjD,OAAM,KAAK,KAAK,OAAO,SAAS;AAGlC,OAAK,MAAM,OAAO,SAChB,OAAM,KAAK,eAAe,IAAI,KAAK,IAAI,IAAI,QAAQ,cAAc;AAGnE,MAAI,cAAc;AAChB,SAAM,KAAK,0BAA0B;AASrC,OAH2B,WACvB,SAAS,SAAS,UAAU,GAC5B,KAAK,YAAY,IAAI,UAAU,CAEjC,OAAM,KAAK,0BAA0B;;AAIzC,SAAO,MAAM,KAAK,GAAG;;;;;CAMvB,WAAW,UAAyB,SAAuD;EACzF,MAAM,OAAO,KAAK,kBAAkB,UAAU,QAAQ;AACtD,SAAO,KAAK,OAAO,KAAK;;CAK1B,AAAQ,qBAAqB,MAAyD;AACpF,MAAI,KAAK,YAAY,SAAS,EAAG,QAAO,CAAC;GAAE;GAAM,SAAS;GAAO,CAAC;EAElE,MAAMC,QAAmD,EAAE;EAM3D,MAAM,UAHiB,CAAC,GAAG,KAAK,YAAY,MAAM,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,SAAS,EAAE,OAAO,CAGxD,KAAK,MAAM,EAAE,QAAQ,uBAAuB,OAAO,CAAC;EACnF,MAAM,QAAQ,IAAI,OAAO,IAAI,QAAQ,KAAK,IAAI,CAAC,IAAI,IAAI;EAEvD,IAAI,YAAY;EAChB,IAAIC;AACJ,UAAQ,QAAQ,MAAM,KAAK,KAAK,MAAM,MAAM;AAC1C,OAAI,MAAM,QAAQ,UAChB,OAAM,KAAK;IACT,MAAM,KAAK,MAAM,WAAW,MAAM,MAAM;IACxC,SAAS;IACV,CAAC;AAEJ,SAAM,KAAK;IAAE,MAAM,MAAM;IAAI,SAAS;IAAM,CAAC;AAC7C,eAAY,MAAM;;AAEpB,MAAI,YAAY,KAAK,OACnB,OAAM,KAAK;GAAE,MAAM,KAAK,MAAM,UAAU;GAAE,SAAS;GAAO,CAAC;AAG7D,SAAO;;CAGT,AAAQ,YAAY,MAAwB;AAC1C,MAAI,CAAC,KAAM,QAAO,EAAE;AAEpB,MAAI,KAAK,SAAS;GAKhB,MAAM,UAAU,KAAK,QAAQ,MAAM,IAAI;GACvC,MAAMC,SAAmB,EAAE;GAC3B,IAAI,IAAI;AACR,UAAO,IAAI,QAAQ,QAAQ;IACzB,IAAI,IAAI,IAAI;AACZ,WAAO,IAAI,QAAQ,UAAU,QAAQ,OAAO,IAAK;AACjD,WAAO,KAAK,QAAQ,MAAM,GAAG,EAAE,CAAC;AAChC,QAAI;;AAEN,UAAO;;EAOT,MAAM,UAAU,KAAK,MADL,qEACmB;AACnC,MAAI,CAAC,QAAS,QAAO,EAAE;AAIvB,SAAO,QAAQ,KAAK,UAAU,KAAK,gBAAgB,MAAM,CAAC;;CAG5D,AAAQ,gBAAgB,MAAsB;EAI5C,IAAI,SAAS;AACb,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;GACpC,MAAM,IAAI,KAAK,WAAW,EAAE;AAC5B,OAAI,MAAM,GACR,WAAU;YACD,IAAI,MAAO,IAAI,OAAO,IAAI,IAGnC,WAAU,OAAO,aAAa,IAAI,IAAI;OAEtC,WAAU,KAAK;;AAGnB,SAAO;;CAGT,AAAQ,UAAU,MAAwB;AACxC,MAAI,KAAK,WAAW,EAAG,QAAO,EAAE;AAGhC,MAAI,KAAK,MAAM,IAAI,KAAK,CACtB,QAAO,CAAC,KAAK,MAAM,IAAI,KAAK,CAAE;EAIhC,IAAI,UAAU,CAAC,GAAG,KAAK;AAIvB,MAAI,QAAQ,WAAW,GAAG;GACxB,MAAM,KAAK,KAAK,MAAM,IAAI,QAAQ,GAAG;AACrC,OAAI,OAAO,OAAW,QAAO,CAAC,GAAG;AAEjC,UAAO,KAAK,mBAAmB,KAAK;;AAItC,SAAO,QAAQ,SAAS,GAAG;GAGzB,IAAI,WAAW;GACf,IAAI,UAAU;AAEd,QAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,SAAS,GAAG,KAAK;IAC3C,MAAM,OAAO,GAAG,QAAQ,GAAG,GAAG,QAAQ,IAAI;IAC1C,MAAM,OAAO,KAAK,OAAO,IAAI,KAAK;AAClC,QAAI,SAAS,UAAa,OAAO,UAAU;AACzC,gBAAW;AAEX,eAAU;;;AAKd,OAAI,YAAY,GAAI;GAGpB,MAAM,SAAS,QAAQ,WAAW,QAAQ,UAAU;AACpD,aAAU;IAAC,GAAG,QAAQ,MAAM,GAAG,QAAQ;IAAE;IAAQ,GAAG,QAAQ,MAAM,UAAU,EAAE;IAAC;;EAIjF,MAAMP,MAAgB,EAAE;AACxB,OAAK,MAAM,OAAO,SAAS;GACzB,MAAM,KAAK,KAAK,MAAM,IAAI,IAAI;AAC9B,OAAI,OAAO,OACT,KAAI,KAAK,GAAG;OAGZ,KAAI,KAAK,GAAG,KAAK,mBAAmB,IAAI,CAAC;;AAG7C,SAAO;;CAGT,AAAQ,mBAAmB,MAAwB;EAEjD,MAAM,QADU,IAAI,aAAa,CACX,OAAO,KAAK;EAClC,MAAMA,MAAgB,EAAE;AACxB,OAAK,MAAM,KAAK,OAAO;GACrB,MAAM,MAAM,KAAK,aAAa,IAAI,EAAE;AACpC,OAAI,KAAK;IACP,MAAM,KAAK,KAAK,MAAM,IAAI,IAAI;AAC9B,QAAI,OAAO,OAAW,KAAI,KAAK,GAAG;;;AAGtC,SAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACzeX,IAAa,kBAAb,MAA2D;CACzD,AAAQ;CAER,YAAY,KAAgC;AAC1C,OAAK,MAAM,uBAAO,IAAI,KAAK;;CAG7B,IAAI,MAAuB;AACzB,SAAO,KAAK,IAAI,IAAI,KAAK;;CAG3B,OAAiB;AACf,SAAO,CAAC,GAAG,KAAK,IAAI,MAAM,CAAC;;CAG7B,IAAI,OAAe;AACjB,SAAO,KAAK,IAAI;;CAGlB,IAAI,MAAgD;AAClD,SAAO,QAAQ,QAAQ,KAAK,IAAI,IAAI,KAAK,CAAC;;CAG5C,IAAI,MAAc,OAAmC;AACnD,OAAK,IAAI,IAAI,MAAM,MAAM;AACzB,SAAO,QAAQ,SAAS;;CAG1B,OAAO,MAA6B;AAClC,OAAK,IAAI,OAAO,KAAK;AACrB,SAAO,QAAQ,SAAS;;;CAI1B,QAAkC;AAChC,SAAO,KAAK;;;;;;;;AA2BhB,SAAS,mBAAmB,KAAkB,OAAyC;AACrF,SAAQ,OAAR;EACE,KAAK,MACH,QAAO,IAAI,aAAa,KAAK,GAAG,IAAI,aAAa,EAAE;EACrD,KAAK,MACH,QAAO,IAAI,WAAW,KAAK,GAAG,IAAI,aAAa,EAAE;EACnD,KAAK,MACH,QAAO,IAAI,YAAY,KAAK,GAAG,IAAI,aAAa,EAAE;EACpD,KAAK;EACL,KAAK,OACH,QAAO,IAAI,YAAY,KAAK,GAAG,IAAI,aAAa,EAAE;EACpD,QACE,QAAO,IAAI,WAAW,KAAK,GAAG,IAAI,WAAW;;;AAInD,SAAS,WAAW,KAAkB,KAA+C;AACnF,SAAQ,KAAR;EACE,KAAK,MACH,QAAO,IAAI,YAAY,KAAK,GAAG,IAAI,aAAa,EAAE;EACpD,KAAK,MACH,QAAO,IAAI,WAAW,KAAK,GAAG,IAAI,aAAa,EAAE;EACnD,KAAK,MACH,QAAO,IAAI,YAAY,KAAK,GAAG,IAAI,aAAa,EAAE;EACpD,KAAK,KACH,QAAO,IAAI,WAAW,KAAK,GAAG,IAAI,WAAW;EAC/C,QACE,QAAO,IAAI,aAAa,KAAK,GAAG,IAAI,aAAa,EAAE;;;AAIzD,SAAS,OAAO,MAAgD;AAC9D,KAAI,gBAAgB,aAAc,QAAO;AACzC,KAAI,gBAAgB,YAAa,QAAO;AACxC,KAAI,gBAAgB,WAAY,QAAO;AACvC,KAAI,gBAAgB,YAAa,QAAO;AACxC,QAAO;;AAGT,MAAM,oBAAoB;AAE1B,IAAa,mBAAb,MAAa,iBAA+C;CAC1D,AAAQ,8BAAc,IAAI,KAA8B;;;;;;;CAOxD,AAAQ,4BAAY,IAAI,KAA0B;CAClD,AAAQ;;CAGR,OAAwB,iBAAiB,MAAM;CAE/C,YAAY,YAAoB,mBAAmB;AACjD,OAAK,YAAY;;CAGnB,OAAe,WAAW,MAAsB;AAC9C,SAAO,kBAAkB;;;;;;;;CAS3B,eACE,MACA,OACA,OACA,UACA,mBACM;AACN,OAAK,UAAU,OAAO,KAAK;AAC3B,OAAK,YAAY,IAAI,MAAM;GAAE;GAAO;GAAO,WAAW;GAAmB;GAAU,CAAC;;CAGtF,IAAI,MAAuB;AACzB,SAAO,KAAK,UAAU,IAAI,KAAK,IAAI,KAAK,YAAY,IAAI,KAAK;;CAG/D,OAAiB;EACf,MAAM,MAAM,IAAI,IAAY,KAAK,YAAY,MAAM,CAAC;AACpD,OAAK,MAAM,KAAK,KAAK,UAAU,MAAM,CAAE,KAAI,IAAI,EAAE;AACjD,SAAO,CAAC,GAAG,IAAI;;CAGjB,IAAI,OAAe;AACjB,SAAO,KAAK,MAAM,CAAC;;CAGrB,MAAM,IAAI,MAAgD;EACxD,MAAM,SAAS,KAAK,UAAU,IAAI,KAAK;AACvC,MAAI,OAAQ,QAAO;EACnB,MAAM,OAAO,KAAK,YAAY,IAAI,KAAK;AACvC,MAAI,CAAC,KAAM,QAAO;EAOlB,IAAIQ,MAA0B;AAC9B,MAAI,KAAK,UAAU,SAAS,eAAe,CACzC,OAAM,MAAM,SAAS,KAAK,SAAS;AAErC,MAAI,CAAC,KAAK;GAER,MAAM,OAAO,OADC,MAAM,OAAO,KAAK,KAAK,UAAU,EACtB,MAAM,IAAI,QAAQ,KAAK,SAAS,CAAC;AAC1D,OAAI,CAAC,KAAM,QAAO;AAClB,SAAM,MAAM,KAAK,aAAa;;AAGhC,MAAI,KAAK,UAAU,MACjB,QAAO;GAAE,MAAM,WAAW,KAAK,KAAK,KAAK;GAAE,OAAO,KAAK;GAAO;EAEhE,IAAI,OAAO,mBAAmB,KAAK,KAAK,MAAM;AAE9C,MAAI,KAAK,UAAU,OACjB,QAAO,UAAU,IAAI,WAAW,KAAK,QAAQ,KAAK,YAAY,KAAK,WAAW,CAAC;WACtE,KAAK,UAAU,MACxB,QAAO,SAAS,KAAK;AAEvB,SAAO;GAAE;GAAM,OAAO,KAAK;GAAO;;CAGpC,MAAM,IAAI,MAAc,OAAmC;AACzD,OAAK,YAAY,OAAO,KAAK;AAG7B,MAAI,MAAM,KAAK,cAAc,iBAAiB,gBAAgB;AAC5D,QAAK,UAAU,IAAI,MAAM,MAAM;AAC/B;;AAEF,OAAK,UAAU,OAAO,KAAK;EAC3B,MAAM,WAAW,iBAAiB,WAAW,KAAK;EAClD,MAAM,QAAQ,MAAM,KAAK,OAAO,MAC9B,MAAM,KAAK,YACX,MAAM,KAAK,aAAa,MAAM,KAAK,WACpC;AAED,SADc,MAAM,OAAO,KAAK,KAAK,UAAU,EACnC,IAAI,IAAI,QAAQ,SAAS,EAAE,IAAI,SAAS,MAAM,CAAC;AAC3D,OAAK,YAAY,IAAI,MAAM;GACzB,OAAO,MAAM;GACb,OAAO;GACP,WAAW,KAAK;GAChB;GACA,MAAM,OAAO,MAAM,KAAK;GACzB,CAAC;;CAGJ,MAAM,OAAO,MAA6B;AACxC,OAAK,UAAU,OAAO,KAAK;EAC3B,MAAM,OAAO,KAAK,YAAY,IAAI,KAAK;AACvC,OAAK,YAAY,OAAO,KAAK;AAG7B,MAAI,QAAQ,KAAK,UAAU,SAAS,KAAK,cAAc,KAAK,UAC1D,KAAI;AAEF,UADc,MAAM,OAAO,KAAK,KAAK,UAAU,EACnC,OAAO,IAAI,QAAQ,KAAK,SAAS,CAAC;UACxC;;;CAOZ,MAAM,UAAyB;AAC7B,MAAI;AACF,SAAM,OAAO,OAAO,KAAK,UAAU;UAC7B;;;;AAOZ,SAAgB,wBAAiC;AAC/C,QAAO,OAAO,WAAW;;;;;;;;;;;;;;;;;AC9J3B,SAAS,iBAAiB,MAAc,UAA0B;AAEhE,KAAI,KAAK,WAAW,UAAU,IAAI,KAAK,WAAW,WAAW,CAC3D,QAAO;AAGT,QAAO,0BAA0B,KAAK,WAAW;;AAKnD,IAAIC,MAAuC;AAC3C,IAAIC,QAA2C;AAC/C,IAAI,iBAAiB;AAErB,eAAe,SAAwB;AACrC,KAAI,eAAgB;AACpB,kBAAiB;AACjB,KAAI;AACF,QAAM,MAAM,OAAO;AACnB,UAAQ,MAAM,OAAO;SACf;;AAKV,SAAS,aAAa,UAAkB,UAA0B;AAChE,QAAO,MAAO,KAAK,UAAU,SAAS,QAAQ,OAAO,IAAI,CAAC;;AAG5D,SAAS,eAAe,UAAkB,UAAiC;AACzE,KAAI,CAAC,OAAO,CAAC,MAAO,QAAO;CAC3B,MAAM,IAAI,aAAa,UAAU,SAAS;AAC1C,KAAI;AACF,SAAO,IAAI,aAAa,EAAE;SACpB;AACN,SAAO;;;AAIX,SAAS,eACP,UACA,UACA,MACM;AACN,KAAI,CAAC,OAAO,CAAC,MAAO;AACpB,KAAI,UAAU,UAAU,EAAE,WAAW,MAAM,CAAC;CAC5C,MAAM,IAAI,aAAa,UAAU,SAAS;CAC1C,IAAIC;AACJ,KAAI,OAAO,SAAS,KAAK,CACvB,OAAM;UACG,gBAAgB,WACzB,OAAM,OAAO,KAAK,KAAK;KAEvB,OAAM,OAAO,KAAK,KAAK;AAEzB,KAAI,cAAc,GAAG,IAAI;;AAiC3B,MAAM,qBAAqB;AAO3B,MAAM,wBACH,OAAO,YAAY,eAAe,QAAQ,KAAK,wBAAwB,QACvE,WAAiD,wBAAwB;AAE5E,IAAI,oBAAoB;AACxB,eAAe,qBAAoC;AACjD,KAAI,kBAAmB;AACvB,qBAAoB;AACpB,KAAI;AAEF,QAAO,WAAmB,SAAS,WAAW;SACxC;;;AAYV,eAAe,iBAAiB,UAA+C;AAC7E,QAAO,OAAO,SAAS;;;;;;;AAazB,eAAe,kBACb,UACA,MACA,gBAAgB,OACD;AACf,OAAM,oBAAoB;AAC1B,OAAM,OAAO,UAAU,KAAK;;;;;;;;;AAU9B,IAAI,gBAAgB;AACpB,eAAe,mBAAkC;AAC/C,KAAI,cAAe;AACnB,iBAAgB;AAChB,KAAI;AACF,MAAI,OAAO,WAAW,aAAa;GACjC,MAAM,OAAO,MAAM,OAAO,MAAM;AAChC,SAAM,QAAQ,IACZ,KAAK,QAAQ,MAAM,EAAE,WAAW,iBAAiB,CAAC,CAAC,KAAK,MAAM,OAAO,OAAO,EAAE,CAAC,CAChF;;SAEG;AAGR,OAAM,gBAAgB;;;;;AAQxB,eAAe,UACb,SACA,UACA,SACA,UACc;AAEd,KAAI,UAAU;EACZ,MAAM,SAAS,eAAe,UAAU,SAAS;AACjD,MAAI,OAAQ,QAAO,KAAK,MAAM,OAAO,SAAS,QAAQ,CAAC;;CAIzD,MAAM,aAAa,GAAG,QAAQ,GAAG;CACjC,MAAM,gBAAgB,MAAM,iBAAiB,WAAW;AACxD,KAAI,cACF,QAAO,KAAK,MAAM,IAAI,aAAa,CAAC,OAAO,cAAc,CAAC;CAG5D,MAAM,MAAM,GAAG,QAAQ,GAAG;CAC1B,MAAMC,UAAkC,EAAE;AAC1C,KAAI,QACF,SAAQ,gBAAgB,UAAU;CAEpC,MAAM,WAAW,MAAM,MAAM,KAAK,EAAE,SAAS,CAAC;AAC9C,KAAI,CAAC,SAAS,GACZ,OAAM,IAAI,MAAM,mBAAmB,IAAI,IAAI,SAAS,OAAO,GAAG,SAAS,aAAa;CAEtF,MAAM,OAAO,MAAM,SAAS,MAAM;AAGlC,KAAI,SACF,gBAAe,UAAU,UAAU,OAAO,KAAK,MAAM,QAAQ,CAAC;AAIhE,OAAM,kBAAkB,YAAY,IAAI,aAAa,CAAC,OAAO,KAAK,CAAC,OAAO;AAE1E,QAAO,KAAK,MAAM,KAAK;;;;;;;AAQzB,eAAe,UACb,SACA,UACA,SACA,UACwB;AACxB,KAAI,UAAU;EACZ,MAAM,SAAS,eAAe,UAAU,SAAS;AACjD,MAAI,OAAQ,QAAO,OAAO,SAAS,QAAQ;;CAE7C,MAAM,aAAa,GAAG,QAAQ,GAAG;CACjC,MAAM,gBAAgB,MAAM,iBAAiB,WAAW;AACxD,KAAI,cACF,QAAO,IAAI,aAAa,CAAC,OAAO,cAAc;CAEhD,MAAM,MAAM,GAAG,QAAQ,GAAG;CAC1B,MAAMA,UAAkC,EAAE;AAC1C,KAAI,QACF,SAAQ,gBAAgB,UAAU;CAEpC,MAAM,WAAW,MAAM,MAAM,KAAK,EAAE,SAAS,CAAC;AAC9C,KAAI,CAAC,SAAS,GAAI,QAAO;CACzB,MAAM,OAAO,MAAM,SAAS,MAAM;AAClC,KAAI,SACF,gBAAe,UAAU,UAAU,OAAO,KAAK,MAAM,QAAQ,CAAC;AAEhE,OAAM,kBAAkB,YAAY,IAAI,aAAa,CAAC,OAAO,KAAK,CAAC,OAAO;AAC1E,QAAO;;;;;AAMT,eAAe,YACb,SACA,UACA,SACA,YACA,UACsB;AAEtB,KAAI,UAAU;EACZ,MAAM,SAAS,eAAe,UAAU,SAAS;AACjD,MAAI,QAAQ;AACV,gBAAa,OAAO,YAAY,OAAO,WAAW;AAClD,UAAO,OAAO,OAAO,MACnB,OAAO,YACP,OAAO,aAAa,OAAO,WAC5B;;;CAIL,MAAM,MAAM,GAAG,QAAQ,GAAG;CAC1B,MAAMA,UAAkC,EAAE;AAC1C,KAAI,QACF,SAAQ,gBAAgB,UAAU;CAGpC,MAAM,WAAW,MAAM,MAAM,KAAK,EAAE,SAAS,CAAC;AAC9C,KAAI,CAAC,SAAS,GACZ,OAAM,IAAI,MAAM,mBAAmB,IAAI,IAAI,SAAS,OAAO,GAAG,SAAS,aAAa;CAGtF,MAAM,gBAAgB,OAAO,SAAS,QAAQ,IAAI,iBAAiB,IAAI,EAAE;AAEzE,KAAI,CAAC,SAAS,QAAQ,CAAC,YAAY;EACjC,MAAM,MAAM,MAAM,SAAS,aAAa;AACxC,MAAI,SAAU,gBAAe,UAAU,UAAU,IAAI;AACrD,SAAO;;CAIT,MAAM,SAAS,SAAS,KAAK,WAAW;CACxC,MAAMC,SAAuB,EAAE;CAC/B,IAAI,SAAS;AAEb,QAAO,MAAM;EACX,MAAM,EAAE,MAAM,UAAU,MAAM,OAAO,MAAM;AAC3C,MAAI,KAAM;AACV,SAAO,KAAK,MAAM;AAClB,YAAU,MAAM;AAChB,aAAW,QAAQ,cAAc;;CAInC,MAAM,SAAS,IAAI,WAAW,OAAO;CACrC,IAAI,SAAS;AACb,MAAK,MAAM,SAAS,QAAQ;AAC1B,SAAO,IAAI,OAAO,OAAO;AACzB,YAAU,MAAM;;AAGlB,KAAI,SAAU,gBAAe,UAAU,UAAU,OAAO;AACxD,QAAO,OAAO;;;;;;;;;;AAWhB,eAAe,yBACb,SACA,SACA,UACmB;AAEnB,KAAI;EACF,MAAM,QAAQ,MAAM,UAAU,SAAS,gCAAgC,SAAS,SAAS;AACzF,MAAI,MAAM,WAGR,QADc,CAAC,GAAG,IAAI,IAAI,OAAO,OAAO,MAAM,WAAW,CAAC,CAAC,CAC9C,MAAM;SAEf;AAIR,QAAO,CAAC,oBAAoB;;;;;AAM9B,eAAe,cACb,KACA,OACA,KACA,SACsB;CACtB,MAAMD,UAAkC,EACtC,OAAO,SAAS,MAAM,GAAG,MAAM,KAChC;AACD,KAAI,QACF,SAAQ,gBAAgB,UAAU;CAMpC,IAAI,aAAa;AACjB,MAAK,IAAI,UAAU,GAAG,UAAU,GAAG,WAAW;EAC5C,IAAIE;AACJ,MAAI;AACF,UAAO,MAAM,MAAM,KAAK,EAAE,SAAS,CAAC;UAC9B;AAEN,OAAI,UAAU,GAAG;AACf,UAAM,IAAI,SAAS,MAAM,WAAW,GAAG,MAAM,KAAK,QAAQ,CAAC;AAC3D;;AAEF,SAAM,IAAI,MAAM,2CAA2C,MAAM;;AAEnE,MAAI,KAAK,MAAM,KAAK,WAAW,IAAK,QAAO,KAAK,aAAa;AAC7D,eAAa,KAAK;AAGlB,OADkB,KAAK,WAAW,OAAO,KAAK,WAAW,OAAO,KAAK,UAAU,QAC9D,UAAU,GAAG;AAC5B,SAAM,IAAI,SAAS,MAAM,WAAW,GAAG,MAAM,KAAK,QAAQ,CAAC;AAC3D;;AAEF;;AAEF,OAAM,IAAI,MACR,yBAAyB,aAAa,eAAe,OAAO,eAAe,MAAM,mFAAmF,KACrK;;;;;;;;;;AAWH,eAAe,WACb,KACA,OACA,KACA,SACsB;CAEtB,MAAM,aAAa,GAAG,IAAI,UAAU,MAAM,GAAG;CAC7C,MAAM,gBAAgB,MAAM,iBAAiB,WAAW;AACxD,KAAI,cAAe,QAAO;CAE1B,MAAM,MAAM,MAAM,cAAc,KAAK,OAAO,KAAK,QAAQ;AAGzD,OAAM,kBAAkB,YAAY,IAAI;AAExC,QAAO;;;;;AAMT,eAAe,uBAAuB,KAAa,SAA4C;CAE7F,MAAM,YAAY,MAAM,WAAW,KAAK,GAAG,GAAG,QAAQ;AAKtD,QAAO,uBADW,MAAM,WAAW,KAAK,GAAG,IAHtB,OAAO,IAAI,SAAS,UAAU,CAAC,aAAa,GAAG,KAAK,CAAC,EAGb,QAAQ,CAC7B;;;;;;;;;;;;;;AAsB1C,SAAS,YACP,SACA,WACA,eAAuB,OAAO,MAC9B,gBAAwB,OAAO,mBAClB;AACb,KAAI,QAAQ,WAAW,EAAG,QAAO,EAAE;CAGnC,MAAM,SAAS,CAAC,GAAG,QAAQ,CAAC,MAAM,GAAG,MAAM,EAAE,aAAa,EAAE,WAAW;CAEvE,MAAMC,SAAsB,EAAE;CAC9B,IAAIC,UAAqB;EACvB,OAAO,YAAY,OAAO,GAAG;EAC7B,KAAK,YAAY,OAAO,GAAG,aAAa,OAAO,GAAG;EAClD,SAAS,CAAC,OAAO,GAAG;EACrB;AAED,MAAK,IAAI,IAAI,GAAG,IAAI,OAAO,QAAQ,KAAK;EACtC,MAAM,aAAa,YAAY,OAAO,GAAG;EACzC,MAAM,WAAW,aAAa,OAAO,GAAG;EAExC,MAAM,YAAY,aAAa,QAAQ,OAAO;EAC9C,MAAM,YAAY,WAAW,QAAQ,SAAS;AAC9C,MAAI,aAAa,WAAW;AAE1B,WAAQ,MAAM,KAAK,IAAI,QAAQ,KAAK,SAAS;AAC7C,WAAQ,QAAQ,KAAK,OAAO,GAAG;SAC1B;AACL,UAAO,KAAK,QAAQ;AACpB,aAAU;IAAE,OAAO;IAAY,KAAK;IAAU,SAAS,CAAC,OAAO,GAAG;IAAE;;;AAGxE,QAAO,KAAK,QAAQ;AACpB,QAAO;;;;;;;AAQT,SAAS,WAAW,OAAuB;AACzC,KAAI,UAAU,SAAS,UAAU,SAAS,UAAU,MAAO,QAAO;AAClE,KAAI,UAAU,SAAS,UAAU,SAAS,UAAU,MAAO,QAAO;AAClE,KAAI,UAAU,SAAS,UAAU,UAAU,UAAU,SAAS,UAAU,MAAO,QAAO;AACtF,QAAO;;;AAIT,SAAS,eACP,KACA,YACA,OACiB;CAGjB,MAAM,UAAU,aAAa,WAAW,MAAM,MAAM,KAAK;CACzD,MAAM,MAAM,UAAU,MAAM,IAAI,MAAM,YAAY,aAAa,MAAM,WAAW;CAChF,MAAM,MAAM,UAAU,aAAa;AACnC,SAAQ,MAAM,OAAd;EACE,KAAK,MACH,QAAO,IAAI,aAAa,KAAK,KAAK,MAAM,aAAa,EAAE;EACzD,KAAK,MACH,QAAO,IAAI,WAAW,KAAK,KAAK,MAAM,aAAa,EAAE;EACvD,KAAK,MACH,QAAO,IAAI,YAAY,KAAK,KAAK,MAAM,aAAa,EAAE;EACxD,KAAK;EACL,KAAK,OACH,QAAO,IAAI,YAAY,KAAK,KAAK,MAAM,aAAa,EAAE;EACxD,QACE,QAAO,IAAI,WAAW,KAAK,KAAK,MAAM,WAAW;;;;;;;;;;AAWvD,SAAS,eAAe,KAAa,MAAuB,OAAgC;AAE1F,QAAO,GAAG,IAAI,KADG,KAAK,YAAY,MAAM,WACZ,GAAG,MAAM;;;;;;;;;;;;;;;;;;;AAoBvC,eAAe,sBACb,KACA,MACA,eACA,SACA,YACA,YACkE;CAClE,MAAM,0BAAU,IAAI,KAAyD;CAC7E,MAAM,aAAa,cAAc,QAAQ,KAAK,MAAM,MAAM,EAAE,YAAY,EAAE;CAC1E,IAAI,cAAc;CAMlB,MAAM,gBAHJ,OAAO,cAAc,eAAe,4BAA4B,KAAK,UAAU,UAAU,GAG1D,KAAK,OAAO,OAAO,MAAM,OAAO;CAKjE,MAAM,UAAU,OAAwB,QAAsB;AAC5D,MAAI,YAAY;GACd,MAAM,YAAY,WAAW,UAAU,MAAM,KAAK;AAClD,OAAI,UAEF,YAAW,MAAM,eACf,WACA,MAAM,OACN,MAAM,OACN,KACA,mBACD;AAEH;;;AAWJ,KAAI,qBAAsB,CAAK,kBAAkB;CASjD,MAAMC,WAA8B,EAAE;CAStC,MAAM,aAAa,uBAAuB,MAAM,YAAY,mBAAG,IAAI,KAAa;CAChF,MAAM,WAAW,cAAc,KAAK,UAAU,WAAW,IAAI,eAAe,KAAK,MAAM,MAAM,CAAC,CAAC;AAC/F,MAAK,IAAI,IAAI,GAAG,IAAI,cAAc,QAAQ,KAAK;EAC7C,MAAM,QAAQ,cAAc;EAC5B,MAAM,MAAM,eAAe,KAAK,MAAM,MAAM;AAC5C,MAAI,CAAC,SAAS,IAAI;AAChB,YAAS,KAAK,MAAM;AACpB;;AAEF,MAAI,YAAY;AAEd,UAAO,OAAO,IAAI;AAClB,kBAAe,MAAM;AACrB,gBAAa,aAAa,WAAW;AACrC;;EAKF,MAAM,MAAM,MAAM,iBAAiB,IAAI;AACvC,MAAI,OAAO,IAAI,cAAc,MAAM,YAAY;AAC7C,WAAQ,IAAI,MAAM,MAAM;IACtB,MAAM,eAAe,KAAK,GAAG,MAAM;IACnC,OAAO,MAAM;IACd,CAAC;AACF,kBAAe,MAAM;AACrB,gBAAa,aAAa,WAAW;QAErC,UAAS,KAAK,MAAM;;AAQxB,KAAI,sBAAsB;EACxB,MAAM,OAAO,cAAc,SAAS,SAAS;EAE7C,MAAM,MAAM,SADI,mBAAmB,GAAG,QAAQ,OACjB,KAAK,KAAK,GAAG,cAAc,OAAO,mBAAmB,SAAS,OAAO;AAClG,UAAQ,IAAI,YAAY,MAAM;AAC9B,eAAa,aAAa,YAAY,IAAI;;CAK5C,MAAM,SAAS,YAAY,UAAU,KAAK,WAAW,OAAO,MAAM,cAAc;AAChF,MAAK,MAAM,SAAS,QAAQ;EAC1B,MAAM,WAAW,MAAM,cAAc,KAAK,MAAM,OAAO,MAAM,KAAK,QAAQ;EAI1E,MAAM,qBAAqB,MAAM,MAAM,MAAM;AAC7C,MAAI,SAAS,aAAa,mBACxB,OAAM,IAAI,MACR,mCAAmC,SAAS,WAAW,MAAM,mBAAmB,aAAa,IAAI,UAClG;AAGH,OAAK,MAAM,SAAS,MAAM,SAAS;GACjC,MAAM,cAAc,KAAK,YAAY,MAAM,aAAa,MAAM;GAC9D,MAAM,MAAM,eAAe,KAAK,MAAM,MAAM;GAE5C,MAAM,cAAc,SAAS,MAAM,aAAa,cAAc,MAAM,WAAW;AAE/E,OAAI,wBAAwB,YAAY,eAAe,MAAM,WAK3D,OAAM,kBAAkB,KAAK,aAAa,KAAK;AAEjD,OAAI,WAEF,QAAO,OAAO,IAAI;OAElB,SAAQ,IAAI,MAAM,MAAM;IACtB,MAAM,eAAe,UAAU,aAAa,MAAM;IAClD,OAAO,MAAM;IACd,CAAC;AAEJ,kBAAe,MAAM;AACrB,gBAAa,aAAa,WAAW;;;AAOzC,QAAO;;;;;;;;;;;;;;;AAgBT,SAAS,uBAAuB,kBAA0B,aAAa,OAAoB;AAKzF,KAAI,qBAAqB,qBAAqB,qBAAqB,sBAOjE,SAAQ,UAAiC;EACvC,IAAI,MAAM;AACV,MAAI,IAAI,WAAW,SAAS,CAC1B,OAAM,IAAI,MAAM,EAAE;AAEpB,QAAM,IAAI,QAAQ,2BAA2B,qBAAqB;AAClE,SAAO;;AAGX,KAAI,qBAAqB,kCACvB,SAAQ,UAAiC;AAUvC,MAAI,MAAM,WAAW,gBAAgB,CACnC,QAAO,aAAa,MAAM,MAAM,EAAE,GAAG;AAEvC,MAAI,MAAM,WAAW,gBAAgB,CAEnC,QAAO,aAAa,UAAU,MAAM,MAAM,GAAG,KAAK;AAIpD,MAAI,MAAM,WAAW,OAAO,CAC1B,QAAO;EAET,IAAI,MAAM;AAEV,MAAI,IAAI,WAAW,wBAAwB,CACzC,OAAM,IAAI,MAAM,GAAG;WACV,IAAI,WAAW,wBAAwB,CAChD,OAAM,IAAI,MAAM,GAAG;WACV,IAAI,WAAW,SAAS,CACjC,OAAM,IAAI,MAAM,EAAE;AAEpB,SAAO;;AAGX,KAAI,qBAAqB,iCASvB,SAAQ,UAAiC;AAMvC,MAAI,MAAM,SAAS,gBAAgB,EAAE;AACnC,OAAI,CAAC,WAAY,QAAO;GACxB,MAAM,MAAM,MAAM,QAAQ,gBAAgB;AAC1C,UAAO,MAAM,MAAM,IAAI;;AAEzB,MAAI,MAAM,SAAS,gBAAgB,EAAE;AACnC,OAAI,CAAC,WAAY,QAAO;GACxB,MAAM,MAAM,MAAM,QAAQ,gBAAgB;AAC1C,UAAO,MAAM,MAAM,IAAI;;AAGzB,MAAI,MAAM,SAAS,eAAe,IAAI,MAAM,SAAS,eAAe,CAClE,QAAO;EAET,IAAI,MAAM;AACV,MAAI,IAAI,WAAW,wBAAwB,CACzC,OAAM,IAAI,MAAM,GAAG;WACV,IAAI,WAAW,wBAAwB,CAChD,OAAM,IAAI,MAAM,GAAG;WACV,IAAI,WAAW,kBAAkB,CAC1C,OAAM,IAAI,MAAM,GAAG;WACV,IAAI,WAAW,SAAS,CACjC,OAAM,IAAI,MAAM,EAAE;AAEpB,SAAO;;AAGX,KAAI,qBAAqB,oCAUvB,QAAO,0BAA0B;AAEnC,QAAO,0BAA0B;;AAGnC,MAAM,iBAAiB;;;;;;;;;;;;;AAcvB,eAAe,eACb,QACA,QACA,OACA,OACA,WACA,YACoB;AACpB,KAAI,CAAC,WACH,QAAO;EAAE;EAAQ;EAAQ;EAAO;EAAO;EAAW,cAAc;EAAiB;CAEnF,MAAM,QAAQ,MAAM,OAAO,KAAK,eAAe;CAC/C,MAAM,YAAY;CAClB,MAAM,YAAY;CAClB,MAAM,WAAW;CACjB,MAAM,QAAQ,MACZ,EAAE,OAAO,MAAM,EAAE,YAAY,EAAE,aAAa,EAAE,WAAW;AAC3D,OAAM,MAAM,IAAI,IAAI,QAAQ,UAAU,EAAE,IAAI,SAAS,KAAK,OAAO,CAAC,CAAC;AACnE,OAAM,MAAM,IAAI,IAAI,QAAQ,UAAU,EAAE,IAAI,SAAS,KAAK,OAAO,CAAC,CAAC;AACnE,OAAM,MAAM,IAAI,IAAI,QAAQ,SAAS,EAAE,IAAI,SAAS,KAAK,MAAM,CAAC,CAAC;AACjE,QAAO;EAEL,QAAQ,IAAI,YAAY,EAAE;EAC1B,QAAQ,IAAI,aAAa,EAAE;EAC3B,OAAO,IAAI,aAAa,EAAE;EAC1B;EACA;EACA,cAAc;EACd,OAAO;GACL,WAAW;GACX;GACA;GACA;GACA,WAAW,OAAO;GACnB;EACF;;;;;;;;;;AAWH,eAAsB,UAAU,SAAiD;CAC/E,MAAM,EAAE,MAAM,YAAY,SAAS,WAAW,QAAQ,UAAU,UAAU;AAG1E,OAAM,QAAQ;CAGd,IAAI,mBAAmB;AACvB,KAAI,CAAC,oBAAoB,OAAO,OAAO;EACrC,MAAM,OAAO,QAAQ,IAAI,QAAQ,QAAQ,IAAI,eAAe;AAC5D,MAAI,KACF,oBAAmB,MAAM,KAAK,MAAM,UAAU,UAAU,KAAK,QAAQ,OAAO,IAAI,EAAE,SAAS;;CAI/F,MAAM,UAAU,iBAAiB,MAAM,SAAS;AAGhD,cAAa,GAAG,KAAK,2BAA2B;CAChD,MAAM,YAAY,MAAM,UAAU,SAAS,eAAe,SAAS,iBAAiB;CAEpF,MAAM,mBAAmB,UAAU,gBAAgB;AACnD,KAAI,CAAC,iBACH,OAAM,IAAI,MAAM,4CAA4C;CAO9D,MAAM,cAAe,UAAU,uBAAuB,UAAU;CAGhE,MAAM,SAAS,aAAa,iBAAiB;CAW7C,MAAM,YAAY,aAAa;CAC/B,MAAM,cACJ,CAAC,UACD,aAAa,SAAS,MACrB,cAAc,YACZ,cAAc,UAAa,OAAO,aAAa,eAAe;CAKnE,MAAM,qBAAqB,CAAC,yCAAyC;CACrE,MAAM,YAAY,KAAK,aAAa;CAEpC,MAAM,oBACJ,CAFY,UAAU,SAAS,MAAM,IAE3B,mBAAmB,MAAM,MAAM,UAAU,SAAS,EAAE,aAAa,CAAC,CAAC;CAI/E,MAAM,QAAQ,gBAAgB,cAAc,YAAY;CAKxD,MAAM,aACJ,OAAO,cAAc,eAAe,4BAA4B,KAAK,UAAU,UAAU;CAG3F,MAAMC,gBAAwC,UAAU,QAAQ,OAD9D,UAAU,SAAU,aAAa,OAAO,SAAa;CAIvD,MAAM,iBAAiB,QAAS,aAAa,aAAwB;AAErE,KAAI,OACF,cAAa,GAAG,KAAK,sDAAsD;UAClE,MACT,cACE,GACA,KACA,wCAAwC,eAAe,oBACxD;CASH,MAAM,aAAa,QAAQ,QAAQ,cAAc,QAAQ,gBAAgB;CACzE,MAAM,YAAY,QAAQ,aAAa,uBAAuB,kBAAkB,WAAW;CAG3F,MAAM,QAAQ,cACZ,kBACA,WACA,eACA,gBACA,QAAQ,SACR,QAAQ,WACR,QAAQ,WACT;AAGD,cAAa,GAAG,KAAK,wBAAwB;CAC7C,MAAM,CAAC,eAAe,qBAAqB,qBAAqB,MAAM,QAAQ,IAAI;EAChF,UAAU,SAAS,kBAAkB,SAAS,iBAAiB;EAC/D,UAAU,SAAS,yBAAyB,SAAS,iBAAiB,CAAC,YAAY,KAAK;EAGxF,UAAU,SAAS,uBAAuB,SAAS,iBAAiB,CAAC,YAAY,KAAK;EACvF,CAAC;CAGF,MAAM,qBACJ,uBAAuB,CAAC,oBAAoB,iBAAiB,oBACzD;EAAE,GAAG;EAAqB,eAAe;EAAmB,GAC3D,wBAAwB,oBAAoB,EAAE,eAAe,mBAAmB,GAAG;CAE1F,MAAM,YAAY,UAAU,SAAS,eAAe,mBAAmB;AAGvE,cAAa,IAAI,KAAK,8BAA8B;CACpD,MAAM,mBAAmB,MAAM,yBAAyB,SAAS,SAAS,iBAAiB;CAW3F,MAAM,aAAc,WAAmD;CACvE,MAAM,cACJ,eAAe,QACd,OAAO,aAAa,eACnB,OAAO,oBAAoB,eAC3B,IAAI,gBAAgB,SAAS,OAAO,CAAC,IAAI,SAAS;CAGtD,MAAM,aADJ,eAAe,UAAU,eAAe,eAAe,uBAAuB,GAC5C,IAAI,kBAAkB,GAAG;CAC7D,MAAMC,QAA4B,cAAc,IAAI,iBAAiB;CAErE,IAAI,cAAc;CAClB,MAAM,aAAa,iBAAiB;AAEpC,MAAK,MAAM,YAAY,kBAAkB;EACvC,MAAM,UAAU,GAAG,QAAQ,GAAG;EAG9B,IAAIC,eAAmC;AACvC,MAAI,kBAAkB;GACpB,MAAM,SAAS,eAAe,kBAAkB,SAAS;AACzD,OAAI,OACF,gBAAe,OAAO,OAAO,MAC3B,OAAO,YACP,OAAO,aAAa,OAAO,WAC5B;;AAIL,MAAI,cAAc;GAEhB,MAAM,OAAO,uBAAuB,aAAa;AACjD,QAAK,MAAM,SAAS,KAAK,SAAS;IAChC,MAAM,gBAAgB,UAAU,MAAM,KAAK;AAC3C,QAAI,CAAC,cAAe;IACpB,IAAIC,OAAwB,cAAc,cAAc,MAAM,MAAM;AACpE,QAAI,MAAM,UAAU,OAClB,QAAO,UAAU,IAAI,WAAW,KAAK,QAAQ,KAAK,YAAY,KAAK,WAAW,CAAC;AAEjF,QAAI,MAAM,UAAU,MAClB,QAAO,SAAS,KAAK;AAEvB,UAAM,MAAM,IAAI,eAAe;KAAE;KAAM,OAAO,MAAM;KAAO,CAAC;;SAEzD;GAEL,MAAM,aAAa,KAAK;GACxB,MAAM,OAAO,KAAK,cAAc;AAKhC,gBAAa,MAAM,KAAK,WAAW,SAAS,WAAW,cAAc,EAAE,GAAG,WAAW,MAAM;GAC3F,MAAM,OAAO,MAAM,uBAAuB,SAAS,QAAQ;GAG3D,MAAM,gBAAgB,KAAK,QAAQ,QAAQ,MAAM,UAAU,EAAE,KAAK,KAAK,KAAK;GAC5E,MAAM,eAAe,KAAK,QACvB,QAAQ,MAAM,UAAU,EAAE,KAAK,KAAK,KAAK,CACzC,QAAQ,KAAK,MAAM,MAAM,EAAE,YAAY,EAAE;GAG5C,MAAM,WAFc,cAAc,QAAQ,KAAK,MAAM,MAAM,EAAE,YAAY,EAAE,GAE5C,SAAS,QAAQ,EAAE;AAClD,OAAI,eAAe,GAAG;IACpB,MAAM,WAAW,eAAe,SAAS,QAAQ,EAAE;AACnD,iBACE,MACA,KACA,uBAAuB,QAAQ,uBAAuB,QAAQ,kBAC/D;;AAGH,gBAAa,MAAM,KAAK,eAAe,SAAS,MAAM,QAAQ,MAAM;GAGpE,IAAI,kBAAkB;GAEtB,MAAM,aAAa,MAAM,sBACvB,SACA,MACA,eACA,UACC,QAAQ,UAAU;IACjB,MAAM,UAAU,QAAQ,IAAK,SAAS,QAAS,aAAa;IAC5D,MAAM,aAAa,KAAK,MAAM,OAAO,QAAQ;AAC7C,QAAI,cAAc,gBAAiB;AACnC,sBAAkB;IAClB,MAAM,SACJ,QAAQ,IACJ,MAAM,SAAS,SAAS,QAAQ,EAAE,CAAC,IAAI,QAAQ,SAAS,QAAQ,EAAE,CAAC,QACnE;AACN,iBAAa,OAAO,SAAS,KAAK,eAAe,WAAW,SAAS;MAKvE,aAAa;IAAE,OAAO;IAAY;IAAW,GAAG,OACjD;AAMD,QAAK,MAAM,CAAC,QAAQ,EAAE,MAAM,SAAS,YAAY,YAAY;IAC3D,MAAM,gBAAgB,UAAU,OAAO;AACvC,QAAI,CAAC,cAAe;IAEpB,IAAIA,OAAwB;IAE5B,MAAM,QAAQ,cAAc,MAAM,MAAM,EAAE,SAAS,OAAO;AAC1D,QAAI,OAAO;AACT,SAAI,MAAM,UAAU,OAClB,QAAO,UAAU,IAAI,WAAW,KAAK,QAAQ,KAAK,YAAY,KAAK,WAAW,CAAC;AAEjF,SAAI,MAAM,UAAU,MAClB,QAAO,SAAS,KAAK;;AAIzB,UAAM,MAAM,IAAI,eAAe;KAAE;KAAM;KAAO,CAAC;;;AAInD;;AAGF,cAAa,IAAI,KAAK,kBAAkB;AASxC,KAAI,qBAAqB,mCAAmC;EAC1D,MAAM,UAAW,UAAU,eAA2C;EACtE,MAAM,YAAY,QAAQ;EAC1B,MAAM,gBAAiB,QAAQ,eAA4B,EAAE;EAC7D,MAAM,mBAAoB,QAAQ,2BAAsC;EACxE,MAAM,WAAW,QAAQ;EACzB,MAAM,UACH,QAAQ,YAAuB,KAAK,MAAO,QAAQ,cAAyB,SAAS;AAExF,OAAK,IAAI,IAAI,GAAG,IAAI,WAAW,KAAK;GAElC,IAAIC;AACJ,OAAI,cAAc,SAAS,EACzB,cAAa,cAAc,OAAO;OAElC,cAAa,IAAI,qBAAqB,mBAAmB;AAE3D,OAAI,CAAC,WAAY;GAEjB,MAAM,WAAW,UAAU,EAAE;GAC7B,MAAM,cAAc,UAAU,EAAE;GAChC,MAAM,cAAc,MAAM,MAAM,IAAI,SAAS;AAE7C,OAAI,eAAe,CAAC,MAAM,IAAI,YAAY,IAAI,CAAC,OAAO;IACpD,MAAM,YAAY,YAAY;IAC9B,MAAM,CAAC,WAAW,QAAQ,YAAY;IACtC,MAAM,WAAW,YAAY;IAE7B,MAAM,QAAQ,IAAI,aAAa,WAAW,KAAK;IAC/C,MAAM,WAAW,IAAI,aAAa,WAAW,KAAK;IAElD,MAAM,MACJ,qBAAqB,eACjB,YACA,IAAI,aAAa,UAAU,QAAQ,UAAU,YAAY,YAAY,KAAK;IAIhF,MAAM,YAAY,IAAI;AACtB,SAAK,IAAI,IAAI,GAAG,IAAI,UAAU,KAAK;KACjC,MAAM,UAAU,IAAI,YAAY;KAChC,MAAM,UAAU,IAAI,UAAU;AAC9B,WAAM,IAAI,IAAI,SAAS,SAAS,UAAU,UAAU,KAAK,EAAE,QAAQ;AACnE,cAAS,IAAI,IAAI,SAAS,UAAU,UAAU,MAAM,UAAU,YAAY,KAAK,EAAE,QAAQ;;AAG3F,UAAM,MAAM,IAAI,UAAU;KAAE,MAAM;KAAO,OAAO,CAAC,UAAU,KAAK;KAAE,CAAC;AACnE,UAAM,MAAM,IAAI,aAAa;KAAE,MAAM;KAAU,OAAO,CAAC,UAAU,KAAK;KAAE,CAAC;;GAO3E,MAAM,eAAe,UAAU,EAAE;GACjC,MAAM,kBAAkB,UAAU,EAAE;AACpC,OAAI,MAAM,IAAI,aAAa,IAAI,CAAC,MAAM,IAAI,gBAAgB,EAAE;IAC1D,MAAM,SAAS,WAAW,IAAI;IAC9B,MAAM,QAAQ,WAAW;IACzB,MAAM,QAAQ,QAAQ;IAGtB,MAAM,KAAM,MAAM,MAAM,IAAI,aAAa;IACzC,MAAM,QACJ,GAAG,gBAAgB,aACf,GAAG,OACH,IAAI,WAAW,GAAG,KAAK,QAAQ,GAAG,KAAK,YAAY,GAAG,KAAK,aAAa,EAAE;IAChF,MAAM,QAAQ,UAAU;IACxB,MAAM,MAAM,IAAI,WAAW,QAAQ,MAAM;IACzC,MAAM,MAAM,IAAI,WAAW,QAAQ,MAAM;IACzC,MAAM,aAAa,IAAI;AACvB,SAAK,IAAI,MAAM,GAAG,MAAM,OAAO,MAC7B,MAAK,IAAI,IAAI,GAAG,IAAI,UAAU,KAAK;KACjC,MAAM,SAAS,MAAM,SAAS,IAAI;KAClC,MAAM,SAAS,MAAM,QAAQ,IAAI;AACjC,UAAK,IAAI,IAAI,GAAG,IAAI,SAAS,KAAK;AAChC,UAAI,SAAS,KAAK,MAAM,SAAS;AACjC,UAAI,SAAS,KAAK,MAAM,SAAS,UAAU;;;AAIjD,UAAM,MAAM,IAAI,cAAc;KAAE,MAAM;KAAK,OAAO,CAAC,OAAO,MAAM;KAAE,CAAC;AACnE,UAAM,MAAM,IAAI,iBAAiB;KAAE,MAAM;KAAK,OAAO,CAAC,OAAO,MAAM;KAAE,CAAC;IAGtE,MAAM,cAAc,UAAU,EAAE;IAChC,MAAM,iBAAiB,UAAU,EAAE;IACnC,MAAM,KAAM,MAAM,MAAM,IAAI,YAAY;IACxC,MAAM,QACJ,GAAG,gBAAgB,eACf,GAAG,OACH,IAAI,aAAa,GAAG,KAAK,QAAQ,GAAG,KAAK,YAAY,GAAG,KAAK,aAAa,EAAE;IAClF,MAAM,SAAS,GAAG,MAAM;IACxB,MAAM,MAAM,IAAI,aAAa,SAAS,MAAM;IAC5C,MAAM,MAAM,IAAI,aAAa,SAAS,MAAM;AAC5C,SAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,IAC1B,MAAK,IAAI,IAAI,GAAG,IAAI,UAAU,KAAK;KACjC,MAAM,SAAS,IAAI,SAAS,IAAI;KAChC,MAAM,SAAS,IAAI,QAAQ,IAAI;AAC/B,UAAK,IAAI,IAAI,GAAG,IAAI,SAAS,KAAK;AAChC,UAAI,SAAS,KAAK,MAAM,SAAS;AACjC,UAAI,SAAS,KAAK,MAAM,SAAS,UAAU;;;AAIjD,UAAM,MAAM,IAAI,aAAa;KAAE,MAAM;KAAK,OAAO,CAAC,QAAQ,MAAM;KAAE,CAAC;AACnE,UAAM,MAAM,IAAI,gBAAgB;KAAE,MAAM;KAAK,OAAO,CAAC,QAAQ,MAAM;KAAE,CAAC;IAKtE,MAAM,cAAc,UAAU,EAAE;IAChC,MAAM,iBAAiB,UAAU,EAAE;IACnC,MAAM,KAAM,MAAM,MAAM,IAAI,YAAY;IACxC,MAAM,QACJ,GAAG,gBAAgB,aACf,GAAG,OACH,IAAI,WAAW,GAAG,KAAK,QAAQ,GAAG,KAAK,YAAY,GAAG,KAAK,aAAa,EAAE;IAChF,MAAM,aAAa,WAAW;IAC9B,MAAM,YAAY,UAAU;IAC5B,MAAM,gBAAgB,eAAe;IACrC,MAAM,cAAc,YAAY;IAChC,MAAM,MAAM,IAAI,WAAW,SAAS,UAAU;IAC9C,MAAM,MAAM,IAAI,WAAW,SAAS,UAAU;AAC9C,SAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,IAC1B,MAAK,IAAI,IAAI,GAAG,IAAI,UAAU,KAAK;KACjC,MAAM,SAAS,IAAI,aAAa,IAAI;KACpC,MAAM,SAAS,IAAI,YAAY,IAAI;AACnC,UAAK,IAAI,IAAI,GAAG,IAAI,aAAa,KAAK;AACpC,UAAI,SAAS,KAAK,MAAM,SAAS;AACjC,UAAI,SAAS,KAAK,MAAM,SAAS,cAAc;;;AAIrD,UAAM,MAAM,IAAI,aAAa;KAAE,MAAM;KAAK,OAAO,CAAC,QAAQ,UAAU;KAAE,CAAC;AACvE,UAAM,MAAM,IAAI,gBAAgB;KAAE,MAAM;KAAK,OAAO,CAAC,QAAQ,UAAU;KAAE,CAAC;AAG1E,UAAM,MAAM,OAAO,UAAU,EAAE,yBAAyB;;GAM1D,MAAM,iBAAiB,UAAU,EAAE;GACnC,MAAM,oBAAoB,UAAU,EAAE;AACtC,OAAI,SAAS,MAAM,IAAI,eAAe,IAAI,CAAC,MAAM,IAAI,kBAAkB,EAAE;IACvE,MAAM,QAAQ,QAAQ;IACtB,MAAM,SAAS,WAAW,IAAI;IAC9B,MAAM,QAAQ,WAAW;IACzB,MAAM,aAAa,IAAI;IACvB,MAAM,QAAS,aAAa,cAAyB;IAGrD,MAAM,mBACJ,KACA,WACA,MACA,SACW;KACX,MAAM,WAAW,YAAY;KAC7B,MAAM,OAAO,IAAI,KAAK,WAAW,KAAK;KACtC,MAAM,OAAO,IAAI,KAAK,WAAW,KAAK;AACtC,UAAK,IAAI,IAAI,GAAG,IAAI,UAAU,KAAK;MACjC,MAAM,UAAU,IAAI,aAAa;MACjC,MAAM,UAAU,IAAI,UAAU;AAC9B,WAAK,IAAI,IAAI,GAAG,IAAI,SAAS,KAAK;OAChC,MAAM,KAAK,UAAU,IAAI;OACzB,MAAM,KAAK,WAAW,UAAU,KAAK;OACrC,MAAM,KAAK,UAAU,IAAI;AACzB,YAAK,IAAI,IAAI,GAAG,IAAI,MAAM,KAAK;AAC7B,aAAK,KAAK,KAAK,IAAI,KAAK;AACxB,aAAK,KAAK,KAAK,IAAI,KAAK;;;;AAI9B,YAAO,CAAC,MAAM,KAAK;;IAIrB,MAAM,KAAM,MAAM,MAAM,IAAI,eAAe;IAC3C,MAAM,QACJ,GAAG,gBAAgB,cACf,GAAG,OACH,IAAI,YAAY,GAAG,KAAK,QAAQ,GAAG,KAAK,YAAY,GAAG,KAAK,aAAa,EAAE;IACjF,MAAM,QAAQ,UAAU;IACxB,MAAM,CAAC,KAAK,OAAO,gBAAgB,OAAO,QAAQ,OAAO,YAAY;AACrE,UAAM,MAAM,IAAI,gBAAgB;KAAE,MAAM;KAAK,OAAO,CAAC,OAAO,MAAM;KAAE,CAAC;AACrE,UAAM,MAAM,IAAI,mBAAmB;KAAE,MAAM;KAAK,OAAO,CAAC,OAAO,MAAM;KAAE,CAAC;IAGxE,MAAM,iBAAiB,UAAU,EAAE;IACnC,MAAM,oBAAoB,UAAU,EAAE;IACtC,MAAM,KAAM,MAAM,MAAM,IAAI,eAAe;IAC3C,MAAM,QACJ,GAAG,gBAAgB,eACf,GAAG,OACH,IAAI,aAAa,GAAG,KAAK,QAAQ,GAAG,KAAK,YAAY,GAAG,KAAK,aAAa,EAAE;IAClF,MAAM,QAAQ,KAAK,KAAK,QAAQ,MAAM;IACtC,MAAM,CAAC,KAAK,OAAO,gBAAgB,OAAO,QAAQ,OAAO,aAAa;AACtE,UAAM,MAAM,IAAI,gBAAgB;KAAE,MAAM;KAAK,OAAO,CAAC,OAAO,MAAM;KAAE,CAAC;AACrE,UAAM,MAAM,IAAI,mBAAmB;KAAE,MAAM;KAAK,OAAO,CAAC,OAAO,MAAM;KAAE,CAAC;IAGxE,MAAM,iBAAiB,UAAU,EAAE;IACnC,MAAM,oBAAoB,UAAU,EAAE;IACtC,MAAM,KAAM,MAAM,MAAM,IAAI,eAAe;IAK3C,MAAM,CAAC,KAAK,OAAO,gBAHjB,GAAG,gBAAgB,eACf,GAAG,OACH,IAAI,aAAa,GAAG,KAAK,QAAQ,GAAG,KAAK,YAAY,GAAG,KAAK,aAAa,EAAE,EACxC,QAAQ,OAAO,aAAa;AACtE,UAAM,MAAM,IAAI,gBAAgB;KAAE,MAAM;KAAK,OAAO,CAAC,OAAO,MAAM;KAAE,CAAC;AACrE,UAAM,MAAM,IAAI,mBAAmB;KAAE,MAAM;KAAK,OAAO,CAAC,OAAO,MAAM;KAAE,CAAC;;;;AAc9E,KAAI,qBAAqB,qCAAqC,CAAC,OAAO;EACpE,MAAM,UAAW,UAAU,eAA2C;EACtE,MAAM,YAAY,QAAQ;EAC1B,MAAM,gBAAiB,QAAQ,eAA4B,EAAE;EAC7D,MAAM,mBAAoB,QAAQ,2BAAsC;EAGxE,MAAMC,WAAqB,EAAE;AAC7B,OAAK,IAAI,IAAI,GAAG,IAAI,WAAW,KAAK;AAElC,YAAS,KAAK,UAAU,EAAE,yBAAyB;AACnD,YAAS,KAAK,UAAU,EAAE,kCAAkC;GAE5D,IAAID;AACJ,OAAI,cAAc,SAAS,EACzB,cAAa,cAAc,OAAO;OAElC,cAAa,IAAI,qBAAqB,mBAAmB;AAG3D,OAAI,YAAY;AACd,aAAS,KAAK,UAAU,EAAE,0BAA0B;AACpD,aAAS,KAAK,UAAU,EAAE,0BAA0B;;;AAMxD,WAAS,KAAK,cAAc;AAE5B,OAAK,MAAM,WAAW,UAAU;GAC9B,MAAM,QAAQ,MAAM,MAAM,IAAI,QAAQ;AACtC,OAAI,OAAO;IACT,MAAM,OACJ,MAAM,gBAAgB,eAClB,MAAM,OACN,IAAI,aAAa,MAAM,KAAK,QAAQ,MAAM,KAAK,YAAY,MAAM,KAAK,aAAa,EAAE;IAC3F,MAAM,WAAW,IAAI,aAAa,KAAK,OAAO;AAC9C,SAAK,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,IAC/B,UAAS,KAAK,IAAM,KAAK;AAE3B,UAAM,MAAM,IAAI,SAAS;KAAE,MAAM;KAAU,OAAO,MAAM;KAAO,CAAC;;;;AAatE,KAAI,qBAAqB,qBAAqB,qBAAqB,eAAe;EAChF,MAAM,iBAAiB,UAAU;EACjC,MAAME,gBAA0B,EAAE;AAClC,OAAK,IAAI,IAAI,GAAG,IAAI,gBAAgB,KAAK;AACvC,iBAAc,KAAK,UAAU,EAAE,yBAAyB;AACxD,iBAAc,KAAK,UAAU,EAAE,kCAAkC;AACjE,iBAAc,KAAK,UAAU,EAAE,mCAAmC;AAClE,iBAAc,KAAK,UAAU,EAAE,oCAAoC;AACnE,iBAAc,KAAK,UAAU,EAAE,0BAA0B;AACzD,iBAAc,KAAK,UAAU,EAAE,0BAA0B;;AAE3D,gBAAc,KAAK,cAAc;AAEjC,OAAK,MAAM,WAAW,eAAe;GACnC,MAAM,QAAQ,MAAM,MAAM,IAAI,QAAQ;AACtC,OAAI,OAAO;IACT,MAAM,OACJ,MAAM,gBAAgB,eAClB,MAAM,OACN,IAAI,aAAa,MAAM,KAAK,QAAQ,MAAM,KAAK,YAAY,MAAM,KAAK,aAAa,EAAE;IAC3F,MAAM,WAAW,IAAI,aAAa,KAAK,OAAO;AAC9C,SAAK,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,IAC/B,UAAS,KAAK,IAAM,KAAK;AAE3B,UAAM,MAAM,IAAI,SAAS;KAAE,MAAM;KAAU,OAAO,MAAM;KAAO,CAAC;;;;AAkBtE,KAAI,qBAAqB,iCACvB,MAAK,MAAM,QAAQ,MAAM,OAAO;EAC9B,MAAM,MAAM,KAAK,YAAY;AAC7B,MAAI,CAAC,IAAK;EACV,MAAM,QAAQ,MAAM,MAAM,IAAI,IAAI;AAClC,MAAI,CAAC,MAAO;EACZ,MAAM,OACJ,MAAM,gBAAgB,eAClB,MAAM,OACN,IAAI,aAAa,MAAM,KAAK,QAAQ,MAAM,KAAK,YAAY,MAAM,KAAK,aAAa,EAAE;AAC3F,OAAK,WAAW,QAAQ,KAAK;AAC7B,QAAM,MAAM,OAAO,IAAI;;AAO3B,MAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,MAAM,QAAQ,CACtD,KACE,KAAK,YAAY,cACjB,KAAK,kBACL,KAAK,mBAAmB,QACxB,CAAC,MAAM,IAAI,KAAK,EAChB;EACA,MAAM,aAAa,MAAM,MAAM,IAAI,KAAK,eAAe;AACvD,MAAI,WACF,OAAM,MAAM,IAAI,MAAM,WAAW;;AAMvC,KAAI,QAAQ;AACV,eAAa,IAAI,KAAK,4BAA4B;EAClD,MAAM,gBAAiB,aAAa,cAAyB;AAE7D,OAAK,MAAM,QAAQ,MAAM,OAAO;AAC9B,OAAI,KAAK,WAAW,aAAc;GAElC,MAAM,IAAI,KAAK,WAAW;GAC1B,MAAM,IAAI,KAAK,WAAW;GAE1B,MAAM,cAAc,KAAK,OAAO;GAChC,MAAM,mBAAmB,KAAK,OAAO;GACrC,MAAM,kBAAkB,KAAK,OAAO;GAKpC,MAAM,WADa,YAAY,MAAM,GAAG,GAAG,CACf,MAAM,GAAG,GAAG;GAExC,MAAM,cAAc,MAAM,MAAM,IAAI,GAAG,SAAS,UAAU;GAC1D,MAAM,aAAa,MAAM,MAAM,IAAI,GAAG,SAAS,SAAS;GACxD,MAAM,aAAa,MAAM,MAAM,IAAI,GAAG,SAAS,SAAS;AAGxD,OAAI,CAAC,eAAe,CAAC,cAAc,CAAC,WAAY;GA4BhD,MAAM,WAAW,WAAW;IAC1B,SAzBA,YAAY,gBAAgB,aACxB,YAAY,OACZ,IAAI,WACF,YAAY,KAAK,QACjB,YAAY,KAAK,YACjB,YAAY,KAAK,aAAa,EAC/B;IAoBL,QAlBA,WAAW,gBAAgB,eACvB,WAAW,OACX,IAAI,aACF,WAAW,KAAK,QAChB,WAAW,KAAK,YAChB,WAAW,KAAK,aAAa,EAC9B;IAaL,QAXA,WAAW,gBAAgB,aACvB,WAAW,OACX,IAAI,WACF,WAAW,KAAK,QAChB,WAAW,KAAK,YAChB,WAAW,KAAK,aAAa,EAC9B;IAML;IACA;IACA,WAAW;IACZ,CAAC;AAEF,SAAM,MAAM,IAAI,aAAa;IAAE,MAAM,SAAS;IAAQ,OAAO,CAAC,SAAS,OAAO,OAAO;IAAE,CAAC;AACxF,SAAM,MAAM,IAAI,kBAAkB;IAAE,MAAM,SAAS;IAAQ,OAAO,CAAC,SAAS,OAAO,OAAO;IAAE,CAAC;AAC7F,SAAM,MAAM,IAAI,iBAAiB;IAAE,MAAM,SAAS;IAAO,OAAO,CAAC,SAAS,MAAM,OAAO;IAAE,CAAC;AAG1F,SAAM,MAAM,OAAO,GAAG,SAAS,UAAU;AACzC,SAAM,MAAM,OAAO,GAAG,SAAS,SAAS;AACxC,SAAM,MAAM,OAAO,GAAG,SAAS,SAAS;;;AAK5C,KAAI,OAAO;AACT,eAAa,IAAI,KAAK,2BAA2B;EACjD,MAAM,QAAS,aAAa,cAAyB;AAErD,OAAK,MAAM,QAAQ,MAAM,OAAO;AAC9B,OAAI,KAAK,WAAW,gBAAgB,KAAK,WAAW,gBAAiB;GAGrE,IAAIC,GAAWC;AACf,OAAI,KAAK,WAAW,iBAAiB;AACnC,QAAI,KAAK,WAAW;AACpB,QAAI,KAAK,WAAW;UACf;AACL,QAAI,KAAK,WAAW;AACpB,QAAI,KAAK,WAAW;;GAGtB,MAAM,cAAc,KAAK,OAAO;GAChC,MAAM,mBAAmB,KAAK,OAAO;GACrC,MAAM,kBAAkB,KAAK,OAAO;GAKpC,MAAM,aAAa,YAAY,MAAM,GAAG,GAAG;GAC3C,MAAM,WAAW,WAAW,MAAM,GAAG,GAAG;GAExC,MAAM,YAAY,MAAM,MAAM,IAAI,WAAW;GAC7C,MAAM,YAAY,MAAM,MAAM,IAAI,GAAG,SAAS,SAAS;GACvD,MAAM,YAAY,MAAM,MAAM,IAAI,GAAG,SAAS,SAAS;AAEvD,OAAI,CAAC,aAAa,CAAC,aAAa,CAAC,UAAW;GA2B5C,MAAM,WAAW,UAAU;IACzB,QAzBA,UAAU,gBAAgB,cACtB,UAAU,OACV,IAAI,YACF,UAAU,KAAK,QACf,UAAU,KAAK,YACf,UAAU,KAAK,aAAa,EAC7B;IAoBL,QAlBA,UAAU,gBAAgB,eACtB,UAAU,OACV,IAAI,aACF,UAAU,KAAK,QACf,UAAU,KAAK,YACf,UAAU,KAAK,aAAa,EAC7B;IAaL,QAXA,UAAU,gBAAgB,eACtB,UAAU,OACV,IAAI,aACF,UAAU,KAAK,QACf,UAAU,KAAK,YACf,UAAU,KAAK,aAAa,EAC7B;IAML;IACA;IACA,WAAW;IACZ,CAAC;AAEF,SAAM,MAAM,IAAI,aAAa;IAAE,MAAM,SAAS;IAAQ,OAAO,CAAC,SAAS,OAAO,OAAO;IAAE,CAAC;AACxF,SAAM,MAAM,IAAI,kBAAkB;IAAE,MAAM,SAAS;IAAQ,OAAO,CAAC,SAAS,OAAO,OAAO;IAAE,CAAC;AAC7F,SAAM,MAAM,IAAI,iBAAiB;IAAE,MAAM,SAAS;IAAO,OAAO,CAAC,SAAS,MAAM,OAAO;IAAE,CAAC;AAG1F,SAAM,MAAM,OAAO,WAAW;AAC9B,SAAM,MAAM,OAAO,GAAG,SAAS,SAAS;AACxC,SAAM,MAAM,OAAO,GAAG,SAAS,SAAS;;;AAK5C,KAAI,kBAAkB,MAAM;AAC1B,eAAa,IAAI,KAAK,gCAAgC;AACtD,OAAK,MAAM,QAAQ,MAAM,OAAO;AAC9B,OAAI,KAAK,WAAW,gBAAgB,KAAK,WAAW,gBAAiB;GAErE,MAAM,UAAW,KAAK,WAAW,cAAyB;GAE1D,MAAM,cAAc,KAAK,OAAO;GAChC,MAAM,mBAAmB,KAAK,OAAO;GACrC,MAAM,kBAAkB,KAAK,OAAO;AAGpC,OAAI,MAAM,IAAI,YAAY,CAAE;GAG5B,MAAM,aAAa,YAAY,MAAM,GAAG,GAAG;GAC3C,IAAI,eAAe,MAAM,MAAM,IAAI,WAAW;AAG9C,OAAI,CAAC,gBAAgB,eAAe,iBAClC,gBAAe,MAAM,MAAM,IAAI,sBAAsB;AAGvD,OAAI,cAAc;IAUhB,MAAM,EAAE,QAAQ,QAAQ,UAAU,aARhC,aAAa,gBAAgB,eACzB,aAAa,OACb,IAAI,aACF,aAAa,KAAK,QAClB,aAAa,KAAK,YAClB,aAAa,KAAK,aAAa,EAChC,EAEiD,QAAQ;AAEhE,UAAM,MAAM,IAAI,aAAa;KAAE,MAAM;KAAQ,OAAO,CAAC,OAAO,OAAO;KAAE,CAAC;AACtE,UAAM,MAAM,IAAI,kBAAkB;KAAE,MAAM;KAAQ,OAAO,CAAC,OAAO,OAAO;KAAE,CAAC;AAC3E,UAAM,MAAM,IAAI,iBAAiB;KAAE,MAAM;KAAO,OAAO,CAAC,MAAM,OAAO;KAAE,CAAC;AAGxE,UAAM,MAAM,OAAO,WAAW;;;;AAcpC,KAAI,OAAO;EACT,MAAM,KAAK,MAAM,MAAM,IAAI,eAAe,oBAAoB;EAC9D,MAAM,IAAI,IAAI;AAEd,MAAI,MAAM,KAAK,EAAE,WAAW,KAAK,EAAE,OAAO,GAAG;GAC3C,MAAM,CAAC,MAAM,GAAG,IAAI,IAAI,KAAK;GAC7B,MAAM,MACJ,GAAG,gBAAgB,eACf,GAAG,OACH,IAAI,aAAa,GAAG,KAAK,QAAQ,GAAG,KAAK,YAAY,GAAG,KAAK,aAAa,EAAE;GAClF,MAAM,MAAM,IAAI,aAAa,IAAI,OAAO;AAGxC,QAAK,IAAI,IAAI,GAAG,IAAI,MAAM,IACxB,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IACrB,MAAK,IAAI,IAAI,GAAG,IAAI,IAAI,IACtB,MAAK,IAAI,IAAI,GAAG,IAAI,IAAI,IACtB,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;IAC1B,MAAM,QAAQ,IAAI,IAAI,KAAK,KAAK,KAAK,KAAK,KAAK,IAAI;IACnD,MAAM,QAAQ,IAAI,IAAI,KAAK,IAAI,KAAK,KAAK,KAAK,KAAK;AACnD,QAAI,MAAM,IAAI;;AAOxB,SAAM,MAAM,IAAI,eAAe,qBAAqB;IAClD,MAAM;IACN,OAAO,CAAC,MAAM,IAAI,IAAI,KAAK,GAAG;IAC/B,CAAC;;;CAYN,IAAIC;AAEJ,KADiB,OAAO,iBAAiB,CAAC,WAAW,SAAS,EAChD;EACZ,MAAM,UAAU,eAAe,iBAAiB,MAAM,GAAG,GAAG;EAC5D,MAAM,UAAW,UAAU,eAA2C;EACtE,MAAM,SAAU,QAAQ,+BAA0C;EAElE,MAAM,WADY,QAAQ,oBACG;EAC7B,MAAM,WACH,QAAQ,8BAA0C,QAAQ;EAE7D,MAAM,cAAc,MAAM,MAAM,IAAI,GAAG,QAAQ,SAAS;EACxD,MAAM,cAAc,MAAM,MAAM,IAAI,GAAG,QAAQ,SAAS;EACxD,MAAM,cAAc,MAAM,MAAM,IAAI,GAAG,QAAQ,SAAS;AAExD,MAAI,SAAS,eAAe,eAAe,aAAa;GACtD,MAAM,QAAS,aAAa,cAAyB;GAyBrD,MAAM,WAAW,UAAU;IACzB,QAxBA,YAAY,gBAAgB,cACxB,YAAY,OACZ,IAAI,YACF,YAAY,KAAK,QACjB,YAAY,KAAK,YACjB,YAAY,KAAK,aAAa,EAC/B;IAmBL,QAjBA,YAAY,gBAAgB,eACxB,YAAY,OACZ,IAAI,aACF,YAAY,KAAK,QACjB,YAAY,KAAK,YACjB,YAAY,KAAK,aAAa,EAC/B;IAYL,QAVA,YAAY,gBAAgB,eACxB,YAAY,OACZ,IAAI,aACF,YAAY,KAAK,QACjB,YAAY,KAAK,YACjB,YAAY,KAAK,aAAa,EAC/B;IAKL,GAAG;IACH,GAAG;IACH,WAAW;IACZ,CAAC;AACF,eAAY,MAAM,eAChB,SAAS,QACT,SAAS,QACT,SAAS,OACT,UACA,OACA,WACD;AAED,SAAM,MAAM,OAAO,GAAG,QAAQ,SAAS;AACvC,SAAM,MAAM,OAAO,GAAG,QAAQ,SAAS;AACvC,SAAM,MAAM,OAAO,GAAG,QAAQ,SAAS;AACvC,WAAQ,IACN,kCAAkC,aAAa,mBAAmB,eAAe,KAC3E,SAAS,IAAI,SAAS,aACnB,SAAS,OAAO,aAAa,SAAS,OAAO,aAAa,SAAS,MAAM,cAAc,KAAK,QAAQ,EAAE,CAAC,gBACjH;aACQ,aAAa,gBAAgB,cAAc;GAIpD,MAAM,EAAE,QAAQ,QAAQ,UAAU,aAAa,YAAY,MAAM,mBAAmB;AACpF,eAAY,MAAM,eAChB,QACA,QACA,OACA,UACA,oBACA,WACD;AACD,SAAM,MAAM,OAAO,GAAG,QAAQ,SAAS;QAEvC,SAAQ,KACN,0IAED;;AAOL,MAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,MAAM,QAAQ,CACtD,KAAI,KAAK,YAAY,cAAc,KAAK,cAAc,UAAa,CAAC,MAAM,IAAI,KAAK,EAAE;EACnF,MAAM,SAAS,KAAK,MAAM,QACvB,KAAa,QAAQ,OAAO,OAAO,QAAQ,WAAW,MAAM,IAC7D,EACD;EACD,MAAM,OAAO,IAAI,aAAa,OAAO,CAAC,KAAK,KAAK,UAAU;AAC1D,QAAM,MAAM,IAAI,MAAM;GACpB;GACA,OAAO,KAAK,MAAM,KAAK,MAAO,OAAO,MAAM,WAAW,IAAI,EAAG;GAC9D,CAAC;;CAKN,MAAMC,iBAA2B,EAAE;AACnC,MAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,MAAM,QAAQ,CACtD,KAAI,KAAK,YAAY,cAAc,CAAC,MAAM,IAAI,KAAK,CACjD,gBAAe,KAAK,KAAK;AAI7B,KAAI,eAAe,SAAS,EAC1B,SAAQ,KACN,oBAAoB,eAAe,OAAO,mBAC1C,eAAe,MAAM,GAAG,GAAG,CAC5B;AAGH,cAAa,KAAK,KAAK,gBAAgB;AAEvC,QAAO;EAAE;EAAO;EAAW,SAAS;EAAO;EAAW;EAAW;;AAuBnE,eAAsB,cAAc,SAMP;CAC3B,MAAM,WAAW,QAAQ,YAAY;CACrC,MAAM,UAAU,iBAAiB,QAAQ,MAAM,SAAS;AACxD,OAAM,QAAQ;AAEd,SAAQ,aAAa,GAAG,KAAK,+BAA+B;CAC5D,MAAM,CAAC,WAAW,iBAAiB,MAAM,QAAQ,IAAI,CACnD,UAAU,SAAS,eAAe,QAAQ,SAAS,QAAQ,SAAS,EACpE,UAAU,SAAS,kBAAkB,QAAQ,SAAS,QAAQ,SAAS,CACxE,CAAC;CACF,MAAM,kBAAkB,MAAM,UAC5B,SACA,yBACA,QAAQ,SACR,QAAQ,SACT,CAAC,YAAY,KAAK;CACnB,MAAM,YAAY,UAAU,SAAS,eAAe,gBAAgB;AAEpE,SAAQ,aAAa,IAAI,KAAK,uBAAuB;CACrD,MAAM,MAAM,MAAM,YAChB,SACA,qBACA,QAAQ,UACP,QAAQ,UAAU,QAAQ,aAAa,KAAM,UAAU,SAAS,KAAM,IAAI,KAAK,UAAU,EAC1F,QAAQ,SACT;CAED,MAAM,OAAO,uBAAuB,IAAI;CACxC,MAAM,0BAAU,IAAI,KAAyD;AAC7E,MAAK,MAAM,SAAS,KAAK,SAAS;EAEhC,MAAM,YAAY,MAAM,KAAK,WAAW,SAAS,GAAG,MAAM,KAAK,MAAM,EAAE,GAAG,MAAM;EAChF,IAAIP,OAAwB,cAAc,KAAK,MAAM,MAAM;AAC3D,MAAI,MAAM,UAAU,OAClB,QAAO,UAAU,IAAI,WAAW,KAAK,QAAQ,KAAK,YAAY,KAAK,WAAW,CAAC;WACtE,MAAM,UAAU,MACzB,QAAO,SAAS,KAAK;AAEvB,UAAQ,IAAI,WAAW;GAAE;GAAM,OAAO,MAAM;GAAO,CAAC;;AAGtD,SAAQ,aAAa,KAAK,KAAK,oBAAoB;AACnD,QAAO;EAAE;EAAS;EAAW;EAAW;;;AAoB1C,MAAa,sBAAsB;;AAcnC,MAAM,mBAAmB;;AAGzB,SAAS,MAAM,OAAwB,KAAkB,MAAqC;CAC5F,IAAIA,OAAwB,cAAc,KAAK,MAAM,MAAM;AAC3D,KAAI,MAAM,UAAU,OAClB,QAAO,UAAU,IAAI,WAAW,KAAK,QAAQ,KAAK,YAAY,KAAK,WAAW,CAAC;UACtE,MAAM,UAAU,MACzB,QAAO,SAAS,KAAK;AAEvB,QAAO,gBAAgB,eACnB,OACA,IAAI,aAAa,KAAK,QAAQ,KAAK,YAAY,KAAK,aAAa,EAAE;;;AAIzE,SAAS,sBACP,UACyD;CACzD,MAAM,sBAAM,IAAI,KAAyD;CACzE,MAAM,EAAE,OAAO,WAAW,oBAAoB;AAC9C,MAAK,MAAM,KAAK,OAAO;EACrB,MAAM,IAAI,SAAS,IAAI,EAAE,KAAK;EAC9B,MAAM,IAAI,SAAS,IAAI,EAAE,KAAK;AAC9B,MAAI,EAAE,KAAK,GACT,OAAM,IAAI,MACR,0CAA0C,EAAE,SAAS,IAAI,EAAE,KAAK,KAAK,EAAE,KAAK,uEAE7E;AAEH,MAAI,IAAI,EAAE,UAAU;GAClB,MAAM,wBAAwB,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM;GACtD,OAAO,EAAE;GACV,CAAC;;AAEJ,MAAK,MAAM,MAAM,QAAQ;EACvB,MAAM,IAAI,SAAS,IAAI,GAAG,OAAO;AACjC,MAAI,CAAC,EACH,OAAM,IAAI,MAAM,6BAA6B,GAAG,OAAO,QAAQ,GAAG,SAAS,IAAI;AAEjF,MAAI,IAAI,GAAG,UAAU;GAAE,MAAM,EAAE;GAAM,OAAO,EAAE;GAAO,CAAC;;AAExD,QAAO;;AAGT,eAAsB,YAAY,SASP;CACzB,MAAM,WAAW,QAAQ,YAAY;CACrC,MAAM,OAAO,QAAQ,QAAQ;CAC7B,MAAM,YAAY,QAAQ,aAAa;CACvC,MAAM,UAAU,iBAAiB,MAAM,SAAS;CAChD,MAAM,WAAW,iBAAiB,WAAW,OAAO;AACpD,OAAM,QAAQ;AAEd,SAAQ,aAAa,GAAG,KAAK,+BAA+B;CAC5D,MAAM,CAAC,WAAW,iBAAiB,MAAM,QAAQ,IAAI,CACnD,UAAU,SAAS,eAAe,QAAQ,SAAS,QAAQ,SAAS,EACpE,UAAU,SAAS,kBAAkB,QAAQ,SAAS,QAAQ,SAAS,CACxE,CAAC;CACF,MAAM,kBAAkB,MAAM,UAC5B,SACA,yBACA,QAAQ,SACR,QAAQ,SACT,CAAC,YAAY,KAAK;CACnB,MAAM,YAAY,UAAU,SAAS,eAAe,gBAAgB;AAGpE,SAAQ,aAAa,GAAG,KAAK,iCAAiC;CAC9D,MAAM,cAAc,MAAM,YACxB,SACA,qBACA,QAAQ,UACP,QAAQ,UACP,QAAQ,aAAa,IAAK,UAAU,SAAS,KAAM,IAAI,KAAK,mBAAmB,EACjF,QAAQ,SACT;CACD,MAAM,eAAe,uBAAuB,YAAY;CACxD,MAAM,kCAAkB,IAAI,KAAyD;AACrF,MAAK,MAAM,SAAS,aAAa,SAAS;EACxC,MAAM,OAAO,MAAM,KAAK,WAAW,SAAS,GAAG,MAAM,KAAK,MAAM,EAAE,GAAG,MAAM,MAAM,QAC/E,kBACA,qBACD;AACD,kBAAgB,IAAI,KAAK;GACvB,MAAM,MAAM,OAAO,aAAa,aAAa;GAC7C,OAAO,MAAM;GACd,CAAC;;AAIJ,SAAQ,aAAa,IAAI,KAAK,iCAAiC;CAI/D,MAAM,gBAAgB,QAAQ,WAAW,GAAG,QAAQ,SAAS,cAAc;CAC3E,MAAM,WAAW,MAAM,YACrB,UACA,qBACA,QAAQ,UACP,QAAQ,UACP,QAAQ,aAAa,KAAM,UAAU,SAAS,KAAM,IAAI,KAAK,gBAAgB,EAC/E,cACD;CACD,MAAM,YAAY,uBAAuB,SAAS;CAClD,MAAM,2BAAW,IAAI,KAAsD;AAC3E,MAAK,MAAM,SAAS,UAAU,QAC5B,UAAS,IAAI,MAAM,MAAM;EAAE,MAAM,MAAM,OAAO,UAAU,UAAU;EAAE,OAAO,MAAM;EAAO,CAAC;CAE3F,MAAM,eAAe,sBAAsB,SAAS;AAEpD,SAAQ,aAAa,KAAK,KAAK,qBAAqB;AACpD,QAAO;EAAE;EAAiB;EAAc;EAAW;EAAW;;;;;ACrrEhE,MAAM,gBAAgB;;;;;;;AAQtB,MAAM,yBAAyB;AAqB/B,IAAa,2BAAb,MAAsC;CACpC,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CAER,AAAQ,gCAAgB,IAAI,KAAwB;CACpD,AAAQ,oCAAoB,IAAI,KAAwB;CACxD,AAAQ,aAAkC,EAAE;CAE5C,YAAY,KAAiB,OAAmB,WAAmB;AACjE,OAAK,MAAM;AACX,OAAK,QAAQ;AACb,OAAK,YAAY;AACjB,OAAK,SAAS,MAAM,OAAO;AAC3B,OAAK,2BAA2B;;;;;;;CAQlC,cAAc,SAAwE;AACpF,OAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,KAAK,MAAM,QAAQ,EAAE;AAC7D,OAAI,KAAK,YAAY,WAAY;GACjC,MAAM,IAAI,QAAQ,IAAI,KAAK;AAC3B,OAAI,GAAG;IACL,MAAM,SAAS,oBAAoB,KAAK,KAAK,QAAQ,QAAQ,EAAE,KAAK,YAAY,EAAE,KAAK;AACvF,SAAK,cAAc,IAAI,MAAM,OAAO;cAC3B,CAAC,KAAK,gBAAgB;IAI/B,MAAM,QAAS,KAAK,MAAmB,QAAQ,GAAG,MAAM,IAAK,GAAc,EAAE;IAC7E,MAAM,QAAQ,IAAI,aAAa,MAAM;IACrC,MAAM,SAAS,oBAAoB,KAAK,KAAK,QAAQ,QAAQ,MAAM,YAAY,MAAM;AACrF,SAAK,cAAc,IAAI,MAAM,OAAO;SAEpC,OAAM,IAAI,MAAM,6CAA6C,KAAK,GAAG;;;CAK3E,iBAAuB;EACrB,MAAM,SAAS,KAAK,eAAe;AACnC,OAAK,MAAM,UAAU,KAAK,MAAM,gBAAgB;GAC9C,MAAM,OAAO,KAAK,MAAM,MAAM,MAAM,MAAM,EAAE,OAAO,OAAO;GAC1D,IAAI,OAAO,gBAAgB,KAAK;AAChC,OAAI,CAAC,KAAM,OAAM,IAAI,MAAM,+CAA+C,KAAK,OAAO,GAAG;AAEzF,OAAI,KAAK,WAAW,eAAe,KAAK,OAAO,WAAW,EACxD,QAAO;YACE,KAAK,WAAW,UAAU,KAAK,WAAW,gBAAgB,KACnE,QAAO;GAET,MAAM,WAAW,oBACf,KAAK,KACL,WAAW,UACX,KAAK,YACL,KAAK,WACN;GACD,MAAM,aAAa,KAAK,YAAY,MAAM,QAAQ,EAAE,QAAQ,GAAG,CAAC;GAChE,MAAM,gBAAgB,oBAAoB,KAAK,KAAK,YAAY,UAAU,WAAW;GACrF,MAAM,gBAAgB,KAAK,cAAc,MAAM,MAAM,cAAc;GACnE,MAAM,YAAY,gBAAgB,KAAK,KAAK,UAAU,eAAe,OAAO,SAAS;AACrF,QAAK,WAAW,KAAK;IAAE;IAAM;IAAM;IAAU;IAAW;IAAe,CAAC;;;;CAK5E,MAAM,OAAO,KAA2C;EACtD,MAAM,IAAI,KAAK,IAAI,OAAO;AAC1B,IAAE,YAAY,KAAK,kBAAkB,IAAI,MAAM,EAAG,GAAG,IAAoB;EAEzE,MAAM,SAAS,KAAK,eAAe;EACnC,MAAM,QAAQ,EAAE,QAAQ,GAAG;EAC3B,MAAMQ,QAAyC,EAAE;AACjD,OAAK,MAAM,KAAK,KAAK,YAAY;GAC/B,MAAM,SAAS,EAAE,KAAK,YAAY,EAAE,MAAM,QAAQ,MAAM;AACxD,KAAE,YAAY,EAAE,eAAe,GAAG,OAAO;AACzC,SAAM,KAAK,EAAE,KAAK,gBAAgB,EAAE,MAAM,QAAQ,MAAM,CAAC;;AAG3D,MAAI,KAAK,IAAI,eACX,MAAK,IAAI,IAAI,GAAG,IAAI,KAAK,WAAW,QAAQ,KAAK;GAC/C,MAAM,IAAI,KAAK,WAAW;AAC1B,OAAI,EAAE,KAAK,WAAW,aAAa;IAEjC,MAAM,CAAC,OAAO,OAAO,SAAS,MAAM;AACpC,SAAK,IAAI,QAAQ,GAAG,QAAQ,OAAO,SAAS,wBAAwB;KAClE,MAAM,OAAO,KAAK,IAAI,wBAAwB,QAAQ,MAAM;KAC5D,MAAM,SAAS,EAAE,KAAK,YAAY,EAAE,MAAM,QAAQ;MAAE,QAAQ;MAAG,SAAS;MAAO,CAAC;AAChF,OAAE,YAAY,EAAE,eAAe,GAAG,OAAO;KACzC,MAAMC,QAAM,KAAK,IAAI,OAAO,sBAAsB;KAClD,MAAMC,MAAID,MAAI,kBAAkB;AAChC,SAAE,YAAY,EAAE,SAAS;AACzB,SAAE,aAAa,GAAG,EAAE,UAAU;AAC9B,SAAE,mBAAmB,MAAM,OAAO,MAAM;AACxC,SAAE,KAAK;AACP,OAAE,OAAO,CAACA,MAAI,QAAQ,CAAC,CAAC;AACxB,WAAM,EAAE,qBAAqB;;AAE/B;;GAEF,MAAM,MAAM,KAAK,IAAI,OAAO,sBAAsB;GAClD,MAAM,IAAI,IAAI,kBAAkB;AAChC,KAAE,YAAY,EAAE,SAAS;AACzB,KAAE,aAAa,GAAG,EAAE,UAAU;AAC9B,KAAE,mBAAmB,GAAG,MAAM,GAAG;AACjC,KAAE,KAAK;AACP,KAAE,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;AACxB,SAAM,EAAE,qBAAqB;;OAE1B;GACL,MAAM,MAAM,KAAK,IAAI,OAAO,qBAAqB,EAAE,OAAO,oBAAoB,CAAC;GAC/E,MAAM,OAAO,IAAI,iBAAiB,EAAE,OAAO,sBAAsB,CAAC;AAClE,QAAK,IAAI,IAAI,GAAG,IAAI,KAAK,WAAW,QAAQ,KAAK;AAC/C,SAAK,YAAY,KAAK,WAAW,GAAG,SAAS;AAC7C,SAAK,aAAa,GAAG,KAAK,WAAW,GAAG,UAAU;AAClD,SAAK,mBAAmB,GAAG,MAAM,GAAG;;AAEtC,QAAK,KAAK;AACV,KAAE,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;;EAG1B,MAAM,OAAO,KAAK,eAAe;EACjC,MAAM,aAAa,MAAM,KAAK,SAAS,eAAe,OAAO,KAAK,OAAO;EACzE,MAAME,OAAuB,EAAE;EAC/B,MAAMC,OAAuB,EAAE;AAC/B,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,WAAW,KAAK;AACvC,QAAK,KAAK,MAAM,KAAK,SAAS,cAAc,KAAK,OAAO,KAAK,OAAO,CAAC;AACrE,QAAK,KAAK,MAAM,KAAK,SAAS,cAAc,KAAK,OAAO,KAAK,OAAO,CAAC;;AAEvE,SAAO;GAAE;GAAY;GAAM;GAAM;GAAM;;;CAIzC,MAAM,eAAe,MAAc,cAAc,MAA6B;AAC5E,SAAO,KAAK,SAAS,MAAM,YAAY;;CAGzC,UAAgB;AACd,iBAAe,CAAC,GAAG,KAAK,cAAc,QAAQ,CAAC,CAAC;AAChD,iBAAe,CAAC,GAAG,KAAK,kBAAkB,QAAQ,CAAC,CAAC;AACpD,OAAK,MAAM,KAAK,KAAK,WAAY,GAAE,cAAc,SAAS;AAC1D,OAAK,cAAc,OAAO;AAC1B,OAAK,kBAAkB,OAAO;AAC9B,OAAK,aAAa,EAAE;;;CAMtB,AAAQ,gBAAwB;AAC9B,SAAO,KAAK,MAAM,QAAQ,YAAY,MAAM;;CAG9C,AAAQ,gBAA0C;EAGhD,MAAMC,WAAqC,EAAE;AAC7C,OAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,KAAK,MAAM,QAAQ,CAC3D,UAAS,QAAQ,KAAK,MAAM,KAAK,QAAQ,IAAc;AAEzD,SAAO;;CAGT,AAAQ,4BAAkC;AACxC,OAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,KAAK,MAAM,QAAQ,EAAE;AAC7D,OAAI,KAAK,YAAY,aAAc;GACnC,MAAM,QACH,KAAK,MAAmB,QAAQ,GAAG,MAAM,IAAK,GAAc,EAAE,GAAG,YAAY,KAAK;AACrF,QAAK,kBAAkB,IAAI,MAAM,oBAAoB,KAAK,KAAK,QAAQ,QAAQ,MAAM,CAAC;;;CAI1F,MAAc,SAAS,MAAc,UAAyC;EAC5E,MAAM,SAAS,KAAK,kBAAkB,IAAI,KAAK;AAC/C,MAAI,CAAC,OAAQ,OAAM,IAAI,MAAM,wCAAwC,KAAK,GAAG;EAC7E,MAAM,UAAU,KAAK,IAAI,WAAW,GAAG,OAAO,KAAK;EACnD,MAAM,WAAW,KAAK,IAAI,OAAO,aAAa;GAAE,MAAM;GAAS,OAAO;GAAiB,CAAC;EACxF,MAAM,MAAM,KAAK,IAAI,OAAO,sBAAsB;AAClD,MAAI,mBAAmB,QAAQ,GAAG,UAAU,GAAG,QAAQ;AACvD,OAAK,IAAI,OAAO,MAAM,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;AAC5C,QAAM,SAAS,SAAS,eAAe,GAAG,QAAQ;EAClD,MAAM,OAAO,IAAI,aAAa,SAAS,eAAe,GAAG,QAAQ,CAAC,MAAM,EAAE,CAAC;AAC3E,WAAS,OAAO;AAChB,WAAS,SAAS;AAClB,SAAO;;CAGT,AAAQ,cACN,MACA,MACA,eAC8B;EAC9B,MAAMC,UAAwC,EAAE;EAChD,MAAM,cAAc,CAAC,GAAG,KAAK,QAAQ,GAAG,KAAK,QAAQ;EACrD,IAAI,YAAY;AAChB,OAAK,MAAM,WAAW,KAAK,SACzB,KAAI,QAAQ,SAAS,UACnB,SAAQ,KAAK,EAAE,QAAQ,eAAe,CAAC;OAClC;GACL,MAAM,aAAa,YAAY;GAC/B,MAAM,SAAS,KAAK,cAAc,IAAI,WAAW,IAAI,KAAK,kBAAkB,IAAI,WAAW;AAC3F,OAAI,CAAC,OACH,OAAM,IAAI,MACR,4CAA4C,WAAW,UAAU,KAAK,KACvE;AAEH,WAAQ,KAAK,EAAE,QAAQ,CAAC;;AAG5B,SAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;AClPX,MAAM,wBAAwB;;AAE9B,MAAM,yBAAyB;;AA8C/B,MAAM,gBAAgB;AAEtB,SAAS,WAAW,KAA2B;AAC7C,KAAI,IAAI,WAAW,EACjB,QAAO;CAET,IAAI,MAAM;AACV,MAAK,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,IAC9B,QAAO,IAAI,KAAK,IAAI;AAEtB,QAAO,KAAK,KAAK,MAAM,IAAI,OAAO;;AAGpC,IAAa,eAAb,MAAa,aAAa;CACxB,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ,aAAa;;CAGrB,AAAS,eAAe;CAExB,AAAQ,YACN,KACA,SACA,WACA,WACA;AACA,OAAK,MAAM;AACX,OAAK,UAAU;AACf,OAAK,YAAY;AACjB,OAAK,YAAY;AACjB,OAAK,aAAc,UAAU,gBAA2B;AACxD,OAAK,aAAc,UAAU,gBAA2B;AAExD,OAAK,sBAAuB,UAAU,0BAAqC,KAAK;;;CAIlF,aAAa,OAAO,UAA+B,EAAE,EAAyB;EAC5E,MAAM,MAAM,MAAM,SAAS;EAE3B,MAAM,EAAE,SAAS,WAAW,cAAc,MAAM,cAAc;GAC5D,MAFW,QAAQ,QAAQ;GAG3B,UAAU,QAAQ;GAClB,SAAS,QAAQ;GACjB,UAAU,QAAQ;GAClB,YAAY,QAAQ;GACrB,CAAC;AACF,SAAO,IAAI,aAAa,KAAK,SAAS,WAAW,UAAU;;;;;;;CAQ7D,MAAM,WAAW,KAAmB,OAA0B,EAAE,EAA6B;AAC3F,MAAI,KAAK,WAAY,OAAM,IAAI,MAAM,mCAAmC;AACxE,MAAI,IAAI,SAAS,IACf,OAAM,IAAI,MAAM,kBAAkB,IAAI,OAAO,2CAA2C;EAE1F,MAAM,IAAI,qBAAqB,KAAK,UAAU;EAC9C,MAAM,OAAO,uBAAuB,IAAI,OAAO;AAC/C,MAAI,OAAO,EACT,OAAM,IAAI,MAAM,6BAA6B,KAAK,kBAAkB;EAGtE,MAAM,eAAe,IAAI,SAAS;EAClC,MAAM,YAAY,WAAW,IAAI;EACjC,MAAM,SAAS,KAAK,UAAU;EAC9B,MAAM,mBAAmB,KAAK,oBAAoB;AAKlD,MAAI,YAAY,UAAU,eAAe,iBACvC,QAAO;GACL,MAAM;GACN,QAAQ,EAAE;GACV,eAAe;GACf;GACA;GACA,UAAU;GACX;EAIH,MAAM,eAAe,8BAA8B,KAAK,WAAW,IAAI,OAAO;EAC9E,MAAM,UAAU,IAAI,yBAAyB,KAAK,KAAK,cAAc,EAAE,WAAW;EAClF,IAAIC;EACJ,IAAIC;EACJ,IAAIC;AACJ,MAAI;AACF,WAAQ,cAAc,KAAK,QAAQ;AACnC,WAAQ,gBAAgB;GACxB,MAAM,MAAM,MAAM,QAAQ,OAAO,IAAI;AACrC,UAAO,IAAI;AACX,UAAO,IAAI;AACX,YAAS,IAAI;YACL;AACR,WAAQ,SAAS;;EAInB,MAAM,eAAe,8BAA8B,KAAK,WAAW,OAAO;EAG1E,MAAM,YAAY,KAAK,IAAI,KAAK,gBAAgB,wBAAwB,EAAE,eAAe;EACzF,MAAM,UAAU,IAAI,SAAS,KAAK,KAAK,cAAc;GAAE;GAAW,QAAQ;GAAO,CAAC;EAClF,MAAMC,SAAmB,EAAE;AAC3B,MAAI;AACF,WAAQ,iBAAiB,mBAAmB,cAAc,KAAK,QAAQ,CAAC;AACxE,WAAQ,gBAAgB;AAExB,QAAK,IAAI,IAAI,GAAG,IAAI,EAAE,YAAY,KAAK;AACrC,YAAQ,WAAW,cAAc,KAAK,KAAK,GAAG;AAC9C,YAAQ,WAAW,cAAc,KAAK,KAAK,GAAG;;AAEhD,WAAQ,OAAO;GAGf,IAAI,YAAY,KAAK;GACrB,MAAM,QAAQ;AACd,QAAK,IAAI,OAAO,GAAG,OAAO,OAAO,QAAQ;IACvC,MAAM,MAAM,MAAM,QAAQ,cAAc,IAAI,YAAY,CAAC,UAAU,CAAC,CAAC;AACrE,WAAO,KAAK,IAAI;AAChB,QAAI,QAAQ,KAAK,WAAY;AAC7B,gBAAY;;YAEN;AACR,WAAQ,SAAS;;EAGnB,MAAM,OAAO,KAAK,UAAU,OAAO,QAAgC,KAAK,CAAC,MAAM;EAC/E,MAAM,WAAW,KAAK,WAAW,KAAK,cAAc,KAAK,KAAK;AAC9D,SAAO;GACL;GACA;GACA,eAAe;GACf;GACA;GACA;GACD;;CAGH,UAAgB;AACd,OAAK,aAAa;AAClB,OAAK,QAAQ,OAAO;;;;;;;;;;;;;AAcxB,SAAS,mBACP,OACA,SACyD;CACzD,MAAM,sBAAM,IAAI,KAAyD;AACzE,MAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,MAAM,QAAQ,EAAE;AACxD,MAAI,KAAK,YAAY,WAAY;EACjC,MAAM,IAAI,QAAQ,IAAI,KAAK;AAC3B,MAAI,EAAG,KAAI,IAAI,MAAM,EAAE;;AAEzB,QAAO"}
+{"version":3,"file":"moonshine-stt-BzQRl-BO.mjs","names":["MAP_MODE_READ","requiredFeatures: GPUFeatureName[]","requestLimits: Record<string, number>","device: GPUDevice","layout: GPUPipelineLayout | \"auto\"","details: string[]","mismatches: string[]","addSpec: KernelSpec","mulSpec: KernelSpec","sliceLastRowSpec: KernelSpec","meanPoolSpec: KernelSpec","scaleSpec: KernelSpec","softcapSpec: KernelSpec","l2NormSpec: KernelSpec","swigluSpec: KernelSpec","residualRmsnormSpec: KernelSpec","siluSpec: KernelSpec","geluSpec: KernelSpec","geluErfSpec: KernelSpec","addBiasSpec: KernelSpec","sliceColsSpec: KernelSpec","mulColsSpec: KernelSpec","applyRotarySpec: KernelSpec","matmulSpec: KernelSpec","M: number","matmulBiasSpec: KernelSpec","poolMatMulSpec: KernelSpec","clippedMatMulSpec: KernelSpec","MATMUL_BIAS_F16C_SPEC: KernelSpec","matmulInt4Spec: KernelSpec","MATVEC_SPEC: KernelSpec","MATVEC_INT4_SPEC: KernelSpec","GATED_MATVEC_INT4_SPEC: KernelSpec","SWIGLU_GATED_MATVEC_INT4_SPEC: KernelSpec","MATVEC_SUBGROUPS_SPEC: KernelSpec","MATVEC_INT4_SUBGROUPS_SPEC: KernelSpec","SWIGLU_MATVEC_SPEC: KernelSpec","SWIGLU_MATVEC_INT4_SPEC: KernelSpec","DUAL_MATVEC_INT4_SPEC: KernelSpec","ARGMAX_SPEC: KernelSpec","rmsnormSpec: KernelSpec","seq_len: number","DUAL_RMSNORM_SPEC: KernelSpec","layernormSpec: KernelSpec","embeddingSpec: KernelSpec","embeddingInt4Spec: KernelSpec","ropeSpec: KernelSpec","ROPE_INTERLEAVED_SPEC: KernelSpec","mropeSpec: KernelSpec","embedSpliceSpec: KernelSpec","attentionSpec: KernelSpec","crossAttentionSpec: KernelSpec","KV_CACHE_APPEND_PACKED_F16_SPEC: KernelSpec","T: number","ATTENTION_PACKED_F16_SPEC: KernelSpec","softmaxSpec: KernelSpec","causalConv1dSpec: KernelSpec","causalConv1dSiluSpec: KernelSpec","causalConv1dGatedSpec: KernelSpec","sigmoidGateSpec: KernelSpec","mambaSSMSpec: KernelSpec","MAMBA_SSM_F16_SPEC: KernelSpec","kvCacheAppendSpec: KernelSpec","DUAL_KV_CACHE_APPEND_SPEC: KernelSpec","DUAL_KV_CACHE_APPEND_F16_SPEC: KernelSpec","DUAL_KV_CACHE_APPEND_PACKED_F16_SPEC: KernelSpec","KV_CACHE_APPEND_F16_SPEC: KernelSpec","ATTENTION_F16_SPEC: KernelSpec","convStateUpdateSpec: KernelSpec","LAYERNORM_NO_BIAS_SPEC: KernelSpec","tanhSpec: KernelSpec","transposeSpec: KernelSpec","groupnormSpec: KernelSpec","conv1dFullSpec: KernelSpec","convTranspose1dSpec: KernelSpec","snake1dSpec: KernelSpec","fsqDequantSpec: KernelSpec","halfSnake1dSpec: KernelSpec","convTranspose1dDepthwiseSpec: KernelSpec","KERNEL_REGISTRY: Partial<Record<OpType, KernelSpec>>","MAP_MODE_READ","prefillEntry: DispatchEntry","runtimeContext: RuntimeContext","dispatchSizes: Array<[number, number, number]>","vocabSize","dispatchSizes: number[][]","argmaxDispatchSizes: Array<[number, number, number]>","results: Array<{\n      idx: number;\n      nodeId: string;\n      opType: string;\n      tensor: string;\n      sum: number;\n      first4: number[];\n      uniformParams?: number[];\n    }>","uniformParams: number[] | undefined","resolved: Record<string, number[]>","bufferEntries: Array<{ buffer: GPUBuffer }>","fusedEntry: DispatchEntry","fusedSpec: KernelSpec","fusedNode: OpNode","entries: Array<{ buffer: GPUBuffer }>","buffer: GPUBuffer | undefined","_dbPromise: Promise<IDBDatabase | null> | null","_client: OpfsClient | null","stale: string[]","header: Record<string, unknown>","entries: SafetensorEntry[]","metadata: Record<string, string> | null","bs: number[]","ids: number[]","pieces: string[]","byteRun: number[]","chars: number[]","parts: string[]","parts: Array<{ text: string; special: boolean }>","match: RegExpExecArray | null","chunks: string[]","buf: ArrayBuffer | null","_fs: typeof import(\"node:fs\") | null","_path: typeof import(\"node:path\") | null","buf: Buffer","headers: Record<string, string>","chunks: Uint8Array[]","resp: Response","ranges: ByteRange[]","current: ByteRange","uncached: SafetensorEntry[]","resolvedDtype: GraphDType | undefined","store: MutableWeightStore","cachedBuffer: ArrayBuffer | null","data: ArrayBufferView","isFullAttn: boolean","normKeys: string[]","gemmaNormKeys: string[]","N: number","K: number","pleSource: PleSource | undefined","missingTensors: string[]","sizes: Array<[number, number, number]>","enc","p","encK: Float32Array[]","encV: Float32Array[]","resolved: Record<string, number[]>","entries: Array<{ buffer: GPUBuffer }>","encK: Float32Array[]","encV: Float32Array[]","frames: number","tokens: number[]"],"sources":["../src/gpu/device.ts","../src/gpu/kernels/registry.ts","../src/gpu/executor.ts","../src/gpu/gptq-adapter.ts","../src/gpu/idb-cache.ts","../src/gpu/mlx-adapter.ts","../src/gpu/opfs-cache.ts","../src/gpu/safetensors.ts","../src/gpu/tokenizer.ts","../src/gpu/weight-source.ts","../src/gpu/model-loader.ts","../src/gpu/moonshine-executor.ts","../src/gpu/moonshine-stt.ts"],"sourcesContent":["/**\n * WebGPU device abstraction layer.\n *\n * Wraps GPUDevice with helpers for buffer allocation, pipeline compilation,\n * compute dispatch, and readback. All GPU interaction flows through here.\n */\n\n// WebGPU buffer usage flags (numeric for Node.js Dawn polyfill compatibility)\nconst STORAGE = 0x0080;\nconst COPY_SRC = 0x0004;\nconst COPY_DST = 0x0008;\nconst UNIFORM = 0x0040;\nconst MAP_READ = 0x0001;\nconst MAP_MODE_READ = 0x0001;\n\nexport interface GPUContext {\n  /** The underlying WebGPU device. */\n  device: GPUDevice;\n  /** Device limits (max buffer size, workgroup size, etc.). */\n  limits: GPUSupportedLimits;\n  /** Whether f16 is supported as a shader type. */\n  hasF16: boolean;\n  /**\n   * Whether the WebGPU `subgroups` feature is available (Chrome 134+, Safari 26+).\n   * When true, kernels may use `subgroupAdd`/`subgroupBroadcast` etc. (requires\n   * `enable subgroups;` in the shader). Absence falls back to the portable\n   * shared-memory reductions — never assume this is present.\n   */\n  hasSubgroups: boolean;\n  /** True when the WebGPU implementation is WebKit's (Safari, all iOS/iPadOS browsers). */\n  isWebKitWebGPU: boolean;\n  /** Whether the `timestamp-query` feature is available (per-pass GPU timing). Used\n   * only by the env-gated decode profiler; never on the normal inference path. */\n  hasTimestamp: boolean;\n  /** Raw adapter info string for diagnostics. */\n  adapterDescription: string;\n}\n\nexport interface InitGPUOptions {\n  /** Called when the GPU device is lost (e.g. tab backgrounded on iOS). */\n  onDeviceLost?: (reason: string, message: string) => void;\n}\n\n/**\n * Initialize WebGPU and request a device with the features we need.\n *\n * In Node.js, initializes Dawn's WebGPU polyfill if navigator.gpu is absent.\n * In the browser, uses the native WebGPU API directly.\n *\n * Throws a clear error if WebGPU is unavailable.\n */\nexport async function initGPU(options?: InitGPUOptions): Promise<GPUContext> {\n  // Node.js: initialize Dawn polyfill if needed\n  if ((typeof navigator === \"undefined\" || !navigator.gpu) && typeof process !== \"undefined\") {\n    try {\n      const dynamicImport = new Function(\"specifier\", \"return import(specifier)\");\n      const webgpuModule = await dynamicImport(\"webgpu\");\n      const { create } = webgpuModule;\n      if (!(globalThis as any).navigator) {\n        (globalThis as any).navigator = {} as Navigator;\n      }\n      (globalThis as any).navigator.gpu = create([]);\n    } catch {\n      // Dawn not available\n    }\n  }\n\n  if (typeof navigator === \"undefined\" || !navigator.gpu) {\n    throw new Error(\n      \"WebGPU is not available in this environment. \" +\n        \"Requires Chrome 113+, Safari 26+ (iOS 26+), or Firefox 141+.\",\n    );\n  }\n\n  // Detect mobile for power preference\n  const isMobile =\n    typeof navigator !== \"undefined\" && /iPhone|iPad|iPod|Android/i.test(navigator.userAgent);\n\n  const adapter = await navigator.gpu.requestAdapter({\n    powerPreference: isMobile ? \"low-power\" : \"high-performance\",\n  });\n\n  if (!adapter) {\n    throw new Error(\n      \"No WebGPU adapter found. Your GPU may not support WebGPU, \" +\n        \"or another tab may be using it. Try closing other tabs.\",\n    );\n  }\n\n  // Check for f16 support\n  const hasF16 = adapter.features.has(\"shader-f16\");\n\n  // Check for subgroups support (Chrome 134+, Safari 26+). Optional — kernels\n  // fall back to portable shared-memory reductions when absent.\n  const hasSubgroups = adapter.features.has(\"subgroups\" as GPUFeatureName);\n\n  // timestamp-query: per-pass GPU timing for the env-gated decode profiler. Only\n  // requested when GERBIL_PROFILE is set so the normal path is untouched.\n  const wantTimestamp = typeof process !== \"undefined\" && process.env?.GERBIL_PROFILE != null;\n  const hasTimestamp = wantTimestamp && adapter.features.has(\"timestamp-query\");\n\n  // Request device with f16 if available\n  const requiredFeatures: GPUFeatureName[] = [];\n  if (hasF16) {\n    requiredFeatures.push(\"shader-f16\");\n  }\n  if (hasSubgroups) {\n    requiredFeatures.push(\"subgroups\" as GPUFeatureName);\n  }\n  if (hasTimestamp) {\n    requiredFeatures.push(\"timestamp-query\");\n  }\n\n  // Request limits conservatively — don't exceed adapter limits\n  const requestLimits: Record<string, number> = {\n    maxBufferSize: adapter.limits.maxBufferSize,\n    maxStorageBufferBindingSize: adapter.limits.maxStorageBufferBindingSize,\n    maxComputeWorkgroupSizeX: Math.min(256, adapter.limits.maxComputeWorkgroupSizeX),\n    maxComputeWorkgroupSizeY: Math.min(256, adapter.limits.maxComputeWorkgroupSizeY),\n    maxComputeWorkgroupStorageSize: adapter.limits.maxComputeWorkgroupStorageSize,\n    maxStorageBuffersPerShaderStage: adapter.limits.maxStorageBuffersPerShaderStage,\n  };\n\n  let device: GPUDevice;\n  try {\n    device = await adapter.requestDevice({\n      requiredFeatures,\n      requiredLimits: requestLimits,\n    });\n  } catch (e) {\n    // Retry with minimal limits if the first attempt fails (e.g. on low-end iPad)\n    try {\n      device = await adapter.requestDevice({ requiredFeatures });\n    } catch {\n      throw new Error(\n        `Failed to create WebGPU device: ${e instanceof Error ? e.message : String(e)}. ` +\n          \"Your GPU may not meet the minimum requirements.\",\n      );\n    }\n  }\n\n  // Handle device loss\n  device.lost.then((info) => {\n    console.error(`[Gerbil] WebGPU device lost: ${info.reason}`, info.message);\n    options?.onDeviceLost?.(info.reason, info.message);\n  });\n\n  // Surface errors that no error scope caught — without this, failed buffer\n  // allocations and invalid bind groups on Safari are completely silent and\n  // the affected dispatches just no-op (reads as zero logits).\n  try {\n    device.onuncapturederror = (ev: GPUUncapturedErrorEvent) => {\n      console.error(`[Gerbil] Uncaptured GPU error: ${ev.error?.message ?? ev.error}`);\n    };\n  } catch {\n    // not supported by this implementation (e.g. older Dawn polyfill)\n  }\n\n  // Detect WebKit's WebGPU implementation, NOT Apple GPU hardware: Dawn on Apple\n  // Silicon (node-dawn, Chrome on macOS) also reports vendor \"apple\"/arch \"common-*\",\n  // so adapter info cannot distinguish implementations. Only WebKit's implementation\n  // needs the Safari coherence workarounds; hardware-based detection routed desktop\n  // Dawn through them (161 -> 19.7 tok/s regression).\n  let adapterDescription = \"\";\n  try {\n    const info = (adapter as any).info ?? (adapter as any).getInfo?.();\n    if (info) {\n      adapterDescription = `vendor=${info.vendor ?? \"\"} arch=${info.architecture ?? \"\"} device=${info.device ?? \"\"} desc=${info.description ?? \"\"}`;\n    }\n  } catch {\n    // adapter.info not available\n  }\n\n  // All iOS/iPadOS browsers are WebKit; on macOS only Safari is (Chrome/Chromium UAs\n  // contain \"Chrome/\"). Node (Dawn) has no AppleWebKit UA and resolves to false.\n  const ua = typeof navigator !== \"undefined\" && navigator.userAgent ? navigator.userAgent : \"\";\n  const isAppleWebKit = ua.includes(\"AppleWebKit\");\n  const isIOSOrIPadOS =\n    /iPhone|iPad|iPod/.test(ua) ||\n    (isAppleWebKit && typeof navigator !== \"undefined\" && (navigator.maxTouchPoints ?? 0) > 1);\n  const isSafariMac = isAppleWebKit && !ua.includes(\"Chrome/\") && !isIOSOrIPadOS;\n  const isWebKitWebGPU = isIOSOrIPadOS || isSafariMac;\n  if (!adapterDescription) {\n    adapterDescription = `UA-fallback: ${ua.slice(0, 120)}`;\n  }\n\n  return {\n    device,\n    limits: device.limits,\n    hasF16,\n    hasSubgroups,\n    isWebKitWebGPU,\n    hasTimestamp,\n    adapterDescription,\n  };\n}\n\n// ── Buffer helpers ────────────────────────────────────────────────────\n\n/**\n * Create a GPU storage buffer and optionally upload data to it.\n */\nexport function createStorageBuffer(\n  ctx: GPUContext,\n  label: string,\n  sizeBytes: number,\n  data?: ArrayBuffer | ArrayBufferView,\n): GPUBuffer {\n  const usage = STORAGE | COPY_SRC | COPY_DST;\n\n  // Align size to 4 bytes (WebGPU requirement)\n  const alignedSize = Math.ceil(sizeBytes / 4) * 4;\n\n  const buffer = ctx.device.createBuffer({\n    label,\n    size: alignedSize,\n    usage,\n  });\n\n  // Use writeBuffer instead of mappedAtCreation — Safari/WebKit's mappedAtCreation\n  // may not properly persist data after unmap(), causing all-zero weight buffers.\n  if (data) {\n    if (data instanceof ArrayBuffer) {\n      ctx.device.queue.writeBuffer(buffer, 0, data);\n    } else {\n      ctx.device.queue.writeBuffer(buffer, 0, data.buffer, data.byteOffset, data.byteLength);\n    }\n  }\n\n  return buffer;\n}\n\n/**\n * Create a uniform buffer for passing parameters to shaders.\n */\nexport function createUniformBuffer(\n  ctx: GPUContext,\n  label: string,\n  data: ArrayBuffer | ArrayBufferView,\n): GPUBuffer {\n  const byteLength = data instanceof ArrayBuffer ? data.byteLength : data.byteLength;\n\n  // Use STORAGE instead of UNIFORM to bypass Safari/Metal bugs with uniform buffer\n  // visibility across many compute passes. WGSL kernels use var<storage, read> for params.\n  // Keep 16-byte alignment for compatibility.\n  const alignedSize = Math.ceil(byteLength / 16) * 16;\n\n  const buffer = ctx.device.createBuffer({\n    label,\n    size: alignedSize,\n    usage: STORAGE | COPY_SRC | COPY_DST,\n  });\n\n  if (data instanceof ArrayBuffer) {\n    ctx.device.queue.writeBuffer(buffer, 0, data);\n  } else {\n    ctx.device.queue.writeBuffer(buffer, 0, data.buffer, data.byteOffset, data.byteLength);\n  }\n\n  return buffer;\n}\n\n/**\n * Create a staging buffer for reading GPU results back to CPU.\n */\nexport function createReadbackBuffer(ctx: GPUContext, label: string, sizeBytes: number): GPUBuffer {\n  const alignedSize = Math.ceil(sizeBytes / 4) * 4;\n  return ctx.device.createBuffer({\n    label,\n    size: alignedSize,\n    usage: MAP_READ | COPY_DST,\n  });\n}\n\n/**\n * Update a uniform buffer's contents.\n */\nexport function writeUniformBuffer(\n  ctx: GPUContext,\n  buffer: GPUBuffer,\n  data: ArrayBuffer | ArrayBufferView,\n): void {\n  if (data instanceof ArrayBuffer) {\n    ctx.device.queue.writeBuffer(buffer, 0, data);\n  } else {\n    ctx.device.queue.writeBuffer(buffer, 0, data.buffer, data.byteOffset, data.byteLength);\n  }\n}\n\n// ── Pipeline helpers ──────────────────────────────────────────────────\n\n/**\n * Per-device pipeline cache. Pipelines are bound to the GPUDevice that created\n * them — using a pipeline from device1 with device2 causes WebGPU validation\n * errors. WeakMap ensures the cache is GC'd when the device is destroyed.\n */\nconst perDeviceCache = new WeakMap<GPUDevice, Map<string, GPUComputePipeline>>();\n\n/** Get or initialize the pipeline cache for a specific device. */\nfunction getPipelineCache(device: GPUDevice): Map<string, GPUComputePipeline> {\n  let cache = perDeviceCache.get(device);\n  if (!cache) {\n    cache = new Map();\n    perDeviceCache.set(device, cache);\n  }\n  return cache;\n}\n\n/**\n * Get or create a compute pipeline from WGSL source code.\n */\nexport function getOrCreatePipeline(\n  ctx: GPUContext,\n  label: string,\n  wgslCode: string,\n  entryPoint: string = \"main\",\n  bindGroupLayoutEntries?: GPUBindGroupLayoutEntry[],\n): GPUComputePipeline {\n  const pipelineCache = getPipelineCache(ctx.device);\n  const cacheKey = `${wgslCode}::${entryPoint}`;\n  const cached = pipelineCache.get(cacheKey);\n  if (cached) return cached;\n\n  let code = wgslCode;\n\n  // Only prepend f16 enable if the shader actually uses f16 types.\n  // Enabling it unconditionally can trigger buggy codepaths in Safari/WebKit's\n  // WGSL→Metal compiler even when no f16 types are used.\n  const usesF16 = code.includes(\": f16\") || code.includes(\"<f16>\");\n  if (ctx.hasF16 && usesF16 && !code.includes(\"enable f16\")) {\n    code = `enable f16;\\n\\n${code}`;\n  }\n\n  const shaderModule = ctx.device.createShaderModule({\n    label: `${label}_shader`,\n    code,\n  });\n\n  let layout: GPUPipelineLayout | \"auto\" = \"auto\";\n  if (bindGroupLayoutEntries) {\n    const bindGroupLayout = ctx.device.createBindGroupLayout({\n      label: `${label}_layout`,\n      entries: bindGroupLayoutEntries,\n    });\n    layout = ctx.device.createPipelineLayout({\n      label: `${label}_pipeline_layout`,\n      bindGroupLayouts: [bindGroupLayout],\n    });\n  }\n\n  const pipeline = ctx.device.createComputePipeline({\n    label,\n    layout,\n    compute: {\n      module: shaderModule,\n      entryPoint,\n    },\n  });\n\n  pipelineCache.set(cacheKey, pipeline);\n  return pipeline;\n}\n\n/**\n * Clear the pipeline cache for a specific device, or no-op if called without\n * a device (WeakMap handles cleanup automatically when the device is GC'd).\n */\nexport function clearPipelineCache(device?: GPUDevice): void {\n  if (device) {\n    perDeviceCache.delete(device);\n  }\n}\n\n// ── Bind group helpers ────────────────────────────────────────────────\n\nexport interface BindGroupEntry {\n  buffer: GPUBuffer;\n  offset?: number;\n  size?: number;\n}\n\n/**\n * Create a bind group from a list of buffers.\n */\nexport function createBindGroup(\n  ctx: GPUContext,\n  pipeline: GPUComputePipeline,\n  entries: BindGroupEntry[],\n  label?: string,\n): GPUBindGroup {\n  return ctx.device.createBindGroup({\n    label,\n    layout: pipeline.getBindGroupLayout(0),\n    entries: entries.map((entry, i) => ({\n      binding: i,\n      resource: {\n        buffer: entry.buffer,\n        offset: entry.offset ?? 0,\n        size: entry.size ?? entry.buffer.size,\n      },\n    })),\n  });\n}\n\n// ── Dispatch helpers ──────────────────────────────────────────────────\n\n/**\n * Encode and submit a single compute dispatch.\n * Useful for testing; the executor batches dispatches in practice.\n */\nexport function dispatchOnce(\n  ctx: GPUContext,\n  pipeline: GPUComputePipeline,\n  bindGroup: GPUBindGroup,\n  workgroupCountX: number,\n  workgroupCountY: number = 1,\n  workgroupCountZ: number = 1,\n): void {\n  const encoder = ctx.device.createCommandEncoder();\n  const pass = encoder.beginComputePass();\n  pass.setPipeline(pipeline);\n  pass.setBindGroup(0, bindGroup);\n  pass.dispatchWorkgroups(workgroupCountX, workgroupCountY, workgroupCountZ);\n  pass.end();\n  ctx.device.queue.submit([encoder.finish()]);\n}\n\n// ── Readback helpers ──────────────────────────────────────────────────\n\n/**\n * Copy data from a storage buffer to a readback buffer, then map and return\n * the contents as a Float32Array.\n *\n * This is the primary way to get results (logits) back to CPU for sampling.\n */\nexport async function readbackFloats(\n  ctx: GPUContext,\n  source: GPUBuffer,\n  readback: GPUBuffer,\n  byteOffset: number = 0,\n  byteLength?: number,\n): Promise<Float32Array> {\n  const length = byteLength ?? source.size - byteOffset;\n\n  const encoder = ctx.device.createCommandEncoder();\n  encoder.copyBufferToBuffer(source, byteOffset, readback, 0, length);\n  ctx.device.queue.submit([encoder.finish()]);\n\n  await readback.mapAsync(MAP_MODE_READ, 0, length);\n  const mapped = readback.getMappedRange(0, length);\n  const result = new Float32Array(mapped.slice(0));\n  readback.unmap();\n\n  return result;\n}\n\n/**\n * Read back raw bytes from a GPU buffer.\n */\nexport async function readbackBytes(\n  ctx: GPUContext,\n  source: GPUBuffer,\n  readback: GPUBuffer,\n  byteOffset: number = 0,\n  byteLength?: number,\n): Promise<ArrayBuffer> {\n  const length = byteLength ?? source.size - byteOffset;\n\n  const encoder = ctx.device.createCommandEncoder();\n  encoder.copyBufferToBuffer(source, byteOffset, readback, 0, length);\n  ctx.device.queue.submit([encoder.finish()]);\n\n  await readback.mapAsync(MAP_MODE_READ, 0, length);\n  const mapped = readback.getMappedRange(0, length);\n  const result = mapped.slice(0);\n  readback.unmap();\n\n  return result;\n}\n\n// ── Memory helpers ────────────────────────────────────────────────────\n\n/**\n * Destroy a list of GPU buffers and release their memory.\n */\nexport function destroyBuffers(buffers: GPUBuffer[]): void {\n  for (const buf of buffers) {\n    buf.destroy();\n  }\n}\n\n// ── GPU diagnostics ──────────────────────────────────────────────────\n\nexport interface GPUDiagnosticResult {\n  /** Whether the basic buffer upload → readback round-trip works. */\n  bufferIntegrity: boolean;\n  /** Whether a trivial compute shader executes correctly. */\n  computeWorks: boolean;\n  /** Whether shared memory + workgroupBarrier() works. */\n  sharedMemoryWorks: boolean;\n  /** Detailed messages for each test. */\n  details: string[];\n}\n\n/**\n * Run GPU diagnostics to verify basic WebGPU operations work correctly.\n * Useful for isolating Safari/WebKit-specific issues.\n */\nexport async function verifyGPU(ctx: GPUContext): Promise<GPUDiagnosticResult> {\n  const details: string[] = [];\n  let bufferIntegrity = false;\n  let computeWorks = false;\n  let sharedMemoryWorks = false;\n\n  // Report device limits\n  try {\n    details.push(\n      `Limits: wgSize=${ctx.limits.maxComputeWorkgroupSizeX}, ` +\n        `wgStorage=${ctx.limits.maxComputeWorkgroupStorageSize}, ` +\n        `maxBuf=${(ctx.limits.maxBufferSize / 1024 / 1024).toFixed(0)}MB, ` +\n        `f16=${ctx.hasF16}, subgroups=${ctx.hasSubgroups}`,\n    );\n  } catch {}\n\n  // Test 1: Buffer upload via writeBuffer → readback\n  try {\n    const testData = new Float32Array([1.0, 2.0, 3.0, 4.0, -1.5, 0.0, 100.0, 0.001]);\n    const srcBuf = createStorageBuffer(ctx, \"diag_src\", testData.byteLength, testData);\n    const readBuf = createReadbackBuffer(ctx, \"diag_read\", testData.byteLength);\n    const result = await readbackFloats(ctx, srcBuf, readBuf, 0, testData.byteLength);\n\n    let match = true;\n    for (let i = 0; i < testData.length; i++) {\n      if (Math.abs(result[i] - testData[i]) > 1e-6) {\n        details.push(`Buffer mismatch at [${i}]: expected ${testData[i]}, got ${result[i]}`);\n        match = false;\n      }\n    }\n    bufferIntegrity = match;\n    details.push(\n      bufferIntegrity ? \"✓ Buffer upload/readback OK\" : \"✗ Buffer upload/readback FAILED\",\n    );\n    srcBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ Buffer test error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // ── Isolation tests ──────────────────────────────────────────────\n  // Test 2a PASSED (single encoder, wg=1, no input, onSubmittedWorkDone).\n  // Test 2b FAILED (separate encoders, wg=4, has input, no wait).\n  // Need to isolate: is it (A) separate encoders, (B) input binding, or (C) workgroup size?\n  //\n  // Matrix of tests — each changes ONE variable from the passing baseline:\n  //   Baseline: single encoder, wg=1, no input, with onSubmittedWorkDone → PASSES\n  //   Test A: SEPARATE encoders, wg=1, no input, no wait → tests queue ordering\n  //   Test B: single encoder, wg=1, HAS input, with wait → tests input binding\n  //   Test C: single encoder, wg=4, no input, with wait → tests workgroup size\n  //   Test D: single encoder, wg=256 + shared mem, no input, with wait → tests shared memory\n\n  /** Helper: run a compute shader and read back result using a SINGLE command encoder + onSubmittedWorkDone. */\n  const runSingleEncoder = async (\n    label: string,\n    code: string,\n    bindings: GPUBuffer[],\n    readBuf: GPUBuffer,\n    outBuf: GPUBuffer,\n    readBytes: number,\n    workgroups: number = 1,\n  ): Promise<Float32Array> => {\n    const pipeline = getOrCreatePipeline(ctx, label, code, \"main\");\n    const bg = createBindGroup(\n      ctx,\n      pipeline,\n      bindings.map((b) => ({ buffer: b })),\n      `${label}_bg`,\n    );\n    const encoder = ctx.device.createCommandEncoder();\n    const pass = encoder.beginComputePass();\n    pass.setPipeline(pipeline);\n    pass.setBindGroup(0, bg);\n    pass.dispatchWorkgroups(workgroups);\n    pass.end();\n    encoder.copyBufferToBuffer(outBuf, 0, readBuf, 0, readBytes);\n    ctx.device.queue.submit([encoder.finish()]);\n    if (ctx.device.queue.onSubmittedWorkDone) {\n      await ctx.device.queue.onSubmittedWorkDone();\n    }\n    await readBuf.mapAsync(MAP_MODE_READ, 0, readBytes);\n    const mapped = readBuf.getMappedRange(0, readBytes);\n    const result = new Float32Array(mapped.slice(0));\n    readBuf.unmap();\n    return result;\n  };\n\n  /** Helper: run compute with SEPARATE encoders (like dispatchOnce + readbackFloats). */\n  const runSeparateEncoders = async (\n    label: string,\n    code: string,\n    bindings: GPUBuffer[],\n    readBuf: GPUBuffer,\n    outBuf: GPUBuffer,\n    readBytes: number,\n    workgroups: number = 1,\n  ): Promise<Float32Array> => {\n    const pipeline = getOrCreatePipeline(ctx, label, code, \"main\");\n    const bg = createBindGroup(\n      ctx,\n      pipeline,\n      bindings.map((b) => ({ buffer: b })),\n      `${label}_bg`,\n    );\n    // Encoder 1: compute dispatch\n    dispatchOnce(ctx, pipeline, bg, workgroups);\n    // Encoder 2: copy + readback\n    return readbackFloats(ctx, outBuf, readBuf, 0, readBytes);\n  };\n\n  // Test A: Separate encoders, wg=1, no input (isolates queue ordering)\n  try {\n    const shader = `\n      @group(0) @binding(0) var<storage, read_write> out: array<f32>;\n      @compute @workgroup_size(1)\n      fn main() { out[0] = 42.0; }\n    `;\n    const outBuf = createStorageBuffer(ctx, \"dA_out\", 4);\n    const readBuf = createReadbackBuffer(ctx, \"dA_read\", 4);\n    const result = await runSeparateEncoders(\"dA\", shader, [outBuf], readBuf, outBuf, 4);\n    const ok = Math.abs(result[0] - 42.0) < 1e-3;\n    details.push(\n      ok\n        ? `✓ A: separate encoders OK (${result[0]})`\n        : `✗ A: separate encoders FAILED (got ${result[0]})`,\n    );\n    outBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ A error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test B: Single encoder, wg=1, HAS input (isolates input binding)\n  try {\n    const shader = `\n      @group(0) @binding(0) var<storage, read> src: array<f32>;\n      @group(0) @binding(1) var<storage, read_write> dst: array<f32>;\n      @compute @workgroup_size(1)\n      fn main() { dst[0] = src[0] * 2.0 + 1.0; }\n    `;\n    const input = new Float32Array([3.0]);\n    const inBuf = createStorageBuffer(ctx, \"dB_in\", 4, input);\n    const outBuf = createStorageBuffer(ctx, \"dB_out\", 4);\n    const readBuf = createReadbackBuffer(ctx, \"dB_read\", 4);\n    const result = await runSingleEncoder(\"dB\", shader, [inBuf, outBuf], readBuf, outBuf, 4);\n    const ok = Math.abs(result[0] - 7.0) < 1e-3;\n    details.push(\n      ok ? `✓ B: input binding OK (${result[0]})` : `✗ B: input binding FAILED (got ${result[0]})`,\n    );\n    computeWorks = ok;\n    inBuf.destroy();\n    outBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ B error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test C: Single encoder, wg=4, no input (isolates workgroup size)\n  try {\n    const shader = `\n      @group(0) @binding(0) var<storage, read_write> out: array<f32>;\n      @compute @workgroup_size(4)\n      fn main(@builtin(local_invocation_id) lid: vec3u) {\n        out[lid.x] = f32(lid.x) + 1.0;\n      }\n    `;\n    const outBuf = createStorageBuffer(ctx, \"dC_out\", 16);\n    const readBuf = createReadbackBuffer(ctx, \"dC_read\", 16);\n    const result = await runSingleEncoder(\"dC\", shader, [outBuf], readBuf, outBuf, 16);\n    const ok = Math.abs(result[0] - 1.0) < 1e-3 && Math.abs(result[3] - 4.0) < 1e-3;\n    details.push(\n      ok\n        ? `✓ C: wg=4 OK (${Array.from(result).join(\",\")})`\n        : `✗ C: wg=4 FAILED (${Array.from(result).join(\",\")})`,\n    );\n    outBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ C error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test D: Single encoder, wg=4 + shared memory (isolates shared mem)\n  // Uses a small workgroup (4 threads) to avoid 256-thread driver issues in diagnostics.\n  // Each thread writes to shared memory, barrier, thread 0 sums and writes to output.\n  try {\n    const shader = `\n      var<workgroup> smem: array<f32, 4>;\n      @group(0) @binding(0) var<storage, read_write> dst: array<f32>;\n      @compute @workgroup_size(4)\n      fn main(@builtin(local_invocation_id) lid: vec3u) {\n        smem[lid.x] = f32(lid.x) + 1.0;\n        workgroupBarrier();\n        if (lid.x == 0u) {\n          dst[0] = smem[0] + smem[1] + smem[2] + smem[3];\n        }\n      }\n    `;\n    const outBuf = createStorageBuffer(ctx, \"dD_out\", 4);\n    const readBuf = createReadbackBuffer(ctx, \"dD_read\", 4);\n    const result = await runSingleEncoder(\"dD\", shader, [outBuf], readBuf, outBuf, 4);\n    sharedMemoryWorks = Math.abs(result[0] - 10.0) < 1e-3;\n    details.push(\n      sharedMemoryWorks\n        ? `✓ D: shared mem OK (${result[0]})`\n        : `✗ D: shared mem FAILED (got ${result[0]})`,\n    );\n    outBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ D error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test E: Separate encoders + input + wg=4 (the full failing combo, to confirm)\n  try {\n    const shader = `\n      @group(0) @binding(0) var<storage, read> src: array<f32>;\n      @group(0) @binding(1) var<storage, read_write> dst: array<f32>;\n      @compute @workgroup_size(4)\n      fn main(@builtin(global_invocation_id) gid: vec3u) {\n        if (gid.x < 4u) { dst[gid.x] = src[gid.x] * 2.0 + 1.0; }\n      }\n    `;\n    const input = new Float32Array([3.0, 7.0, -2.0, 0.5]);\n    const inBuf = createStorageBuffer(ctx, \"dE_in\", 16, input);\n    const outBuf = createStorageBuffer(ctx, \"dE_out\", 16);\n    const readBuf = createReadbackBuffer(ctx, \"dE_read\", 16);\n    const result = await runSeparateEncoders(\"dE\", shader, [inBuf, outBuf], readBuf, outBuf, 16);\n    const ok = Math.abs(result[0] - 7.0) < 1e-3 && Math.abs(result[1] - 15.0) < 1e-3;\n    details.push(\n      ok ? `✓ E: full combo OK` : `✗ E: full combo FAILED (${Array.from(result).join(\",\")})`,\n    );\n    inBuf.destroy();\n    outBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ E error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test F: pack2x16float / unpack2x16float round-trip (tests the packed-f16 KV path)\n  // Packs pairs of f32 → u32 via pack2x16float, then unpacks back via unpack2x16float.\n  // If this fails on Safari/Metal, the packed KV cache will produce gibberish.\n  try {\n    const shader = `\n      @group(0) @binding(0) var<storage, read> src: array<f32>;\n      @group(0) @binding(1) var<storage, read_write> dst: array<f32>;\n      @compute @workgroup_size(1)\n      fn main() {\n        // Pack pairs of f32 into u32, then unpack back\n        let packed0 = pack2x16float(vec2f(src[0], src[1]));\n        let packed1 = pack2x16float(vec2f(src[2], src[3]));\n        let packed2 = pack2x16float(vec2f(src[4], src[5]));\n        let packed3 = pack2x16float(vec2f(src[6], src[7]));\n\n        let u0 = unpack2x16float(packed0);\n        let u1 = unpack2x16float(packed1);\n        let u2 = unpack2x16float(packed2);\n        let u3 = unpack2x16float(packed3);\n\n        dst[0] = u0.x; dst[1] = u0.y;\n        dst[2] = u1.x; dst[3] = u1.y;\n        dst[4] = u2.x; dst[5] = u2.y;\n        dst[6] = u3.x; dst[7] = u3.y;\n      }\n    `;\n    // Values chosen to cover: positive, negative, zero, small, >1\n    const input = new Float32Array([1.0, -0.5, 0.0, 3.14, 0.001, -2.0, 100.0, 0.25]);\n    const inBuf = createStorageBuffer(ctx, \"dF_in\", input.byteLength, input);\n    const outBuf = createStorageBuffer(ctx, \"dF_out\", input.byteLength);\n    const readBuf = createReadbackBuffer(ctx, \"dF_read\", input.byteLength);\n    const result = await runSingleEncoder(\n      \"dF\",\n      shader,\n      [inBuf, outBuf],\n      readBuf,\n      outBuf,\n      input.byteLength,\n    );\n    // f16 has limited precision, so use larger tolerance\n    let ok = true;\n    const mismatches: string[] = [];\n    for (let i = 0; i < input.length; i++) {\n      const tol = Math.max(Math.abs(input[i]) * 0.01, 0.001); // 1% or 0.001\n      if (Math.abs(result[i] - input[i]) > tol) {\n        ok = false;\n        mismatches.push(`[${i}]:${input[i]}→${result[i]}`);\n      }\n    }\n    details.push(\n      ok\n        ? `✓ F: pack/unpack2x16float OK`\n        : `✗ F: pack/unpack2x16float FAILED (${mismatches.join(\", \")})`,\n    );\n    inBuf.destroy();\n    outBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ F error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test G: Packed KV-like pattern — pack into array<u32>, read back with elem indexing\n  // Mimics the actual KV cache append + attention load_k/load_v pattern.\n  try {\n    const shader = `\n      @group(0) @binding(0) var<storage, read> src: array<f32>;\n      @group(0) @binding(1) var<storage, read_write> packed: array<u32>;\n      @group(0) @binding(2) var<storage, read_write> dst: array<f32>;\n      @compute @workgroup_size(1)\n      fn main() {\n        // Step 1: Pack f32 pairs into u32 array (like KVCacheAppend)\n        packed[0] = pack2x16float(vec2f(src[0], src[1]));\n        packed[1] = pack2x16float(vec2f(src[2], src[3]));\n        packed[2] = pack2x16float(vec2f(src[4], src[5]));\n        packed[3] = pack2x16float(vec2f(src[6], src[7]));\n\n        // Step 2: Read back using element-level indexing (like load_k/load_v)\n        for (var i: u32 = 0u; i < 8u; i += 1u) {\n          let pair = unpack2x16float(packed[i >> 1u]);\n          if ((i & 1u) == 0u) {\n            dst[i] = pair.x;\n          } else {\n            dst[i] = pair.y;\n          }\n        }\n      }\n    `;\n    const input = new Float32Array([1.5, -0.75, 2.0, 0.0, -1.0, 0.5, 3.0, -3.0]);\n    const inBuf = createStorageBuffer(ctx, \"dG_in\", input.byteLength, input);\n    const packedBuf = createStorageBuffer(ctx, \"dG_packed\", 16); // 4 u32 = 16 bytes\n    const outBuf = createStorageBuffer(ctx, \"dG_out\", input.byteLength);\n    const readBuf = createReadbackBuffer(ctx, \"dG_read\", input.byteLength);\n    const result = await runSingleEncoder(\n      \"dG\",\n      shader,\n      [inBuf, packedBuf, outBuf],\n      readBuf,\n      outBuf,\n      input.byteLength,\n    );\n    let ok = true;\n    const mismatches: string[] = [];\n    for (let i = 0; i < input.length; i++) {\n      const tol = Math.max(Math.abs(input[i]) * 0.01, 0.001);\n      if (Math.abs(result[i] - input[i]) > tol) {\n        ok = false;\n        mismatches.push(`[${i}]:${input[i]}→${result[i]}`);\n      }\n    }\n    details.push(\n      ok\n        ? `✓ G: packed KV round-trip OK`\n        : `✗ G: packed KV round-trip FAILED (${mismatches.join(\", \")})`,\n    );\n    inBuf.destroy();\n    packedBuf.destroy();\n    outBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ G error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test H: Tree reduction with 256 threads + shared memory (RMSNorm/softmax pattern)\n  // 256 threads each write a value, then tree-reduce to sum. Tests the exact\n  // shared memory reduction pattern used in RMSNorm, softmax, and dot products.\n  try {\n    const shader = `\n      var<workgroup> smem: array<f32, 256>;\n      @group(0) @binding(0) var<storage, read> src: array<f32>;\n      @group(0) @binding(1) var<storage, read_write> dst: array<f32>;\n      @compute @workgroup_size(256)\n      fn main(@builtin(local_invocation_id) lid: vec3u) {\n        let tid = lid.x;\n        smem[tid] = src[tid] * src[tid];\n        workgroupBarrier();\n        var stride: u32 = 128u;\n        while (stride > 0u) {\n          if (tid < stride) {\n            smem[tid] += smem[tid + stride];\n          }\n          workgroupBarrier();\n          stride /= 2u;\n        }\n        // Thread 0 writes sum, all threads write their squared input for verification\n        dst[tid] = src[tid] * src[tid];\n        if (tid == 0u) { dst[256] = smem[0]; }\n      }\n    `;\n    // Input: 1..256, expected sum of squares = n(n+1)(2n+1)/6 = 5559680\n    const input = new Float32Array(256);\n    for (let i = 0; i < 256; i++) input[i] = i + 1;\n    const expectedSum = (256 * 257 * 513) / 6; // 5559680\n    const inBuf = createStorageBuffer(ctx, \"dH_in\", input.byteLength, input);\n    const outBuf = createStorageBuffer(ctx, \"dH_out\", 257 * 4);\n    const readBuf = createReadbackBuffer(ctx, \"dH_read\", 257 * 4);\n    const result = await runSingleEncoder(\"dH\", shader, [inBuf, outBuf], readBuf, outBuf, 257 * 4);\n    const sum = result[256];\n    const ok = Math.abs(sum - expectedSum) / expectedSum < 0.001;\n    details.push(\n      ok\n        ? `✓ H: 256-thread tree reduce OK (${sum})`\n        : `✗ H: 256-thread tree reduce FAILED (got ${sum}, expected ${expectedSum})`,\n    );\n    inBuf.destroy();\n    outBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ H error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test I: INT4 dequantize pattern (bit manipulation + multiply-add)\n  // Tests the exact u32→nibble extraction used in MatMulInt4 and SwiGLU INT4 kernels.\n  try {\n    const shader = `\n      @group(0) @binding(0) var<storage, read> packed: array<u32>;\n      @group(0) @binding(1) var<storage, read> scales: array<f32>;\n      @group(0) @binding(2) var<storage, read> zeros: array<f32>;\n      @group(0) @binding(3) var<storage, read_write> dst: array<f32>;\n      @compute @workgroup_size(1)\n      fn main() {\n        // Unpack 8 nibbles from one u32, dequantize: (nibble - zero) * scale\n        let word = packed[0];\n        let scale = scales[0];\n        let zero = zeros[0];\n        for (var i: u32 = 0u; i < 8u; i += 1u) {\n          let nibble = (word >> (i * 4u)) & 0xFu;\n          dst[i] = (f32(nibble) - zero) * scale;\n        }\n      }\n    `;\n    // Pack nibbles 0,1,2,...,7 into one u32: 0x76543210\n    const packedData = new Uint32Array([0x76543210]);\n    const scaleData = new Float32Array([0.5]);\n    const zeroData = new Float32Array([2.0]);\n    // Expected: (0-2)*0.5=-1, (1-2)*0.5=-0.5, (2-2)*0.5=0, (3-2)*0.5=0.5, ...\n    const expected = [0, 1, 2, 3, 4, 5, 6, 7].map((n) => (n - 2.0) * 0.5);\n\n    const packedBuf = createStorageBuffer(ctx, \"dI_packed\", 4, packedData);\n    const scaleBuf = createStorageBuffer(ctx, \"dI_scale\", 4, scaleData);\n    const zeroBuf = createStorageBuffer(ctx, \"dI_zero\", 4, zeroData);\n    const outBuf = createStorageBuffer(ctx, \"dI_out\", 32);\n    const readBuf = createReadbackBuffer(ctx, \"dI_read\", 32);\n    const result = await runSingleEncoder(\n      \"dI\",\n      shader,\n      [packedBuf, scaleBuf, zeroBuf, outBuf],\n      readBuf,\n      outBuf,\n      32,\n    );\n    let ok = true;\n    const mismatches: string[] = [];\n    for (let i = 0; i < 8; i++) {\n      if (Math.abs(result[i] - expected[i]) > 0.001) {\n        ok = false;\n        mismatches.push(`[${i}]:${expected[i]}→${result[i]}`);\n      }\n    }\n    details.push(\n      ok ? `✓ I: INT4 dequant OK` : `✗ I: INT4 dequant FAILED (${mismatches.join(\", \")})`,\n    );\n    packedBuf.destroy();\n    scaleBuf.destroy();\n    zeroBuf.destroy();\n    outBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ I error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test J: Multi-dispatch chain (tests sequential dispatch ordering)\n  // Runs 3 dispatches in sequence within one command encoder: A→B→C.\n  // If dispatch ordering is broken, intermediate results will be wrong.\n  try {\n    const shaderAdd = `\n      @group(0) @binding(0) var<storage, read> a: array<f32>;\n      @group(0) @binding(1) var<storage, read_write> b: array<f32>;\n      @compute @workgroup_size(4)\n      fn main(@builtin(global_invocation_id) gid: vec3u) {\n        if (gid.x < 4u) { b[gid.x] = a[gid.x] + b[gid.x]; }\n      }\n    `;\n    const shaderMul = `\n      @group(0) @binding(0) var<storage, read_write> x: array<f32>;\n      @compute @workgroup_size(4)\n      fn main(@builtin(global_invocation_id) gid: vec3u) {\n        if (gid.x < 4u) { x[gid.x] = x[gid.x] * 2.0; }\n      }\n    `;\n    // Chain: buf = [1,2,3,4], add [10,20,30,40] → [11,22,33,44], mul ×2 → [22,44,66,88], add again → [32,64,96,128]... actually simpler:\n    // Start: buf = [0,0,0,0]\n    // Dispatch 1: buf += src ([1,2,3,4])\n    // Dispatch 2: buf *= 2\n    // Dispatch 3: buf += src again\n    // Expected: ([1,2,3,4] * 2) + [1,2,3,4] = [3,6,9,12]\n    const srcData = new Float32Array([1, 2, 3, 4]);\n    const srcBuf = createStorageBuffer(ctx, \"dJ_src\", 16, srcData);\n    const bufBuf = createStorageBuffer(ctx, \"dJ_buf\", 16, new Float32Array(4)); // zeros\n    const readBuf = createReadbackBuffer(ctx, \"dJ_read\", 16);\n\n    const addPipeline = getOrCreatePipeline(ctx, \"dJ_add\", shaderAdd, \"main\");\n    const mulPipeline = getOrCreatePipeline(ctx, \"dJ_mul\", shaderMul, \"main\");\n\n    const addBg = createBindGroup(\n      ctx,\n      addPipeline,\n      [{ buffer: srcBuf }, { buffer: bufBuf }],\n      \"dJ_add_bg\",\n    );\n    const mulBg = createBindGroup(ctx, mulPipeline, [{ buffer: bufBuf }], \"dJ_mul_bg\");\n\n    const encoder = ctx.device.createCommandEncoder();\n    const pass = encoder.beginComputePass();\n    pass.setPipeline(addPipeline);\n    pass.setBindGroup(0, addBg);\n    pass.dispatchWorkgroups(1); // buf = [1,2,3,4]\n    pass.setPipeline(mulPipeline);\n    pass.setBindGroup(0, mulBg);\n    pass.dispatchWorkgroups(1); // buf = [2,4,6,8]\n    pass.setPipeline(addPipeline);\n    pass.setBindGroup(0, addBg);\n    pass.dispatchWorkgroups(1); // buf = [3,6,9,12]\n    pass.end();\n    encoder.copyBufferToBuffer(bufBuf, 0, readBuf, 0, 16);\n    ctx.device.queue.submit([encoder.finish()]);\n    if (ctx.device.queue.onSubmittedWorkDone) {\n      await ctx.device.queue.onSubmittedWorkDone();\n    }\n    await readBuf.mapAsync(MAP_MODE_READ, 0, 16);\n    const mapped = readBuf.getMappedRange(0, 16);\n    const result = new Float32Array(mapped.slice(0));\n    readBuf.unmap();\n\n    const expected = [3, 6, 9, 12];\n    let ok = true;\n    for (let i = 0; i < 4; i++) {\n      if (Math.abs(result[i] - expected[i]) > 0.01) ok = false;\n    }\n    details.push(\n      ok\n        ? `✓ J: multi-dispatch chain OK (${Array.from(result).join(\",\")})`\n        : `✗ J: multi-dispatch chain FAILED (got ${Array.from(result).join(\",\")}, expected ${expected.join(\",\")})`,\n    );\n    srcBuf.destroy();\n    bufBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ J error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test K: exp() clamp safety (Metal fast-math NaN check)\n  // Tests that exp(large_negative) returns ~0 and not NaN.\n  try {\n    const shader = `\n      @group(0) @binding(0) var<storage, read_write> dst: array<f32>;\n      @compute @workgroup_size(1)\n      fn main() {\n        dst[0] = exp(max(-1000.0, -80.0));      // should be exp(-80) ≈ 1.8e-35\n        dst[1] = exp(max(-1e30, -80.0));         // should be exp(-80) ≈ 1.8e-35\n        dst[2] = exp(-10.0);                     // should be ~4.54e-5\n        dst[3] = 1.0 / (1.0 + exp(max(-50.0, -80.0)));  // sigmoid(50) ≈ 1.0\n        // Direct exp without clamp — what does Metal actually do?\n        dst[4] = exp(-100.0);                    // should be ~3.7e-44 or 0\n        dst[5] = exp(-1000.0);                   // might be NaN on Metal!\n        dst[6] = exp(-1e10);                     // might be NaN on Metal!\n        // Check for NaN: NaN != NaN\n        let v5 = exp(-1000.0);\n        if (v5 != v5) { dst[7] = -999.0; } else { dst[7] = v5; }\n      }\n    `;\n    const outBuf = createStorageBuffer(ctx, \"dK_out\", 32);\n    const readBuf = createReadbackBuffer(ctx, \"dK_read\", 32);\n    const result = await runSingleEncoder(\"dK\", shader, [outBuf], readBuf, outBuf, 32);\n    const clampedOk = result[0] > 0 && result[0] < 1e-30 && result[1] > 0 && result[1] < 1e-30;\n    const normalOk = Math.abs(result[2] - Math.exp(-10)) < 1e-8;\n    const sigmoidOk = Math.abs(result[3] - 1.0) < 0.01;\n    const nanDetected =\n      result[7] === -999 || Number.isNaN(result[5]) || Number.isNaN(result[6]) || result[5] === 0;\n\n    const detail =\n      `clamped=${result[0].toExponential(2)},${result[1].toExponential(2)} ` +\n      `exp(-10)=${result[2].toExponential(2)} sig=${result[3].toFixed(4)} ` +\n      `raw: exp(-100)=${result[4]} exp(-1000)=${result[5]} exp(-1e10)=${result[6]} ` +\n      `NaN?=${result[7]}`;\n\n    details.push(\n      clampedOk && normalOk\n        ? `✓ K: exp() clamp OK (${detail})`\n        : `✗ K: exp() clamp FAILED (${detail})`,\n    );\n    outBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ K error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test L: Large buffer integrity (1 MB upload + readback)\n  // Weights are uploaded as large storage buffers. If Safari's writeBuffer silently\n  // corrupts data for large buffers or views with non-zero byteOffset, this will catch it.\n  try {\n    const MB = 256 * 1024; // 256K floats = 1 MB\n    const testData = new Float32Array(MB);\n    // Fill with a deterministic non-trivial pattern (not all-zeros or sequential)\n    for (let i = 0; i < MB; i++) {\n      testData[i] = Math.sin(i * 0.001) * (i % 1000) * 0.01;\n    }\n    const srcBuf = createStorageBuffer(ctx, \"dL_src\", testData.byteLength, testData);\n    const readBuf = createReadbackBuffer(ctx, \"dL_read\", testData.byteLength);\n    const result = await readbackFloats(ctx, srcBuf, readBuf, 0, testData.byteLength);\n\n    let mismatches = 0;\n    let firstMismatchIdx = -1;\n    for (let i = 0; i < MB; i++) {\n      if (Math.abs(result[i] - testData[i]) > 1e-6) {\n        if (firstMismatchIdx === -1) firstMismatchIdx = i;\n        mismatches++;\n      }\n    }\n    const ok = mismatches === 0;\n    details.push(\n      ok\n        ? `✓ L: 1MB buffer integrity OK (${MB} floats)`\n        : `✗ L: 1MB buffer FAILED (${mismatches} mismatches, first at [${firstMismatchIdx}]: expected ${testData[firstMismatchIdx]}, got ${result[firstMismatchIdx]})`,\n    );\n    srcBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ L error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test M: Large buffer with byteOffset (simulates weight tensor views)\n  // Weight tensors from safetensors are ArrayBufferViews with non-zero byteOffset\n  // into a larger buffer. Safari might handle the offset incorrectly in writeBuffer.\n  try {\n    // Create a 2 MB backing buffer, write a known pattern at offset 1 MB\n    const backingSize = 512 * 1024; // 512K floats = 2 MB\n    const backing = new Float32Array(backingSize);\n    const sliceStart = 256 * 1024; // Start at 1 MB offset\n    const sliceLen = 128 * 1024; // 128K floats = 512 KB\n    for (let i = 0; i < sliceLen; i++) {\n      backing[sliceStart + i] = Math.cos(i * 0.002) * (i % 500) * 0.005;\n    }\n    // Create a view with non-zero byteOffset (like getTensorData returns)\n    const view = new Float32Array(backing.buffer, sliceStart * 4, sliceLen);\n\n    const srcBuf = createStorageBuffer(ctx, \"dM_src\", view.byteLength, view);\n    const readBuf = createReadbackBuffer(ctx, \"dM_read\", view.byteLength);\n    const result = await readbackFloats(ctx, srcBuf, readBuf, 0, view.byteLength);\n\n    let mismatches = 0;\n    let firstMismatchIdx = -1;\n    for (let i = 0; i < sliceLen; i++) {\n      if (Math.abs(result[i] - view[i]) > 1e-6) {\n        if (firstMismatchIdx === -1) firstMismatchIdx = i;\n        mismatches++;\n      }\n    }\n    const ok = mismatches === 0;\n    details.push(\n      ok\n        ? `✓ M: offset buffer integrity OK (512KB at 1MB offset)`\n        : `✗ M: offset buffer FAILED (${mismatches} mismatches, first at [${firstMismatchIdx}]: expected ${view[firstMismatchIdx]}, got ${result[firstMismatchIdx]})`,\n    );\n    srcBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ M error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test N: Real RMSNorm kernel (actual WGSL from registry, not simplified)\n  // Tests the exact shader code used in inference. If Safari's WGSL→MSL compiler\n  // miscompiles the real kernel, this will catch it.\n  try {\n    const rmsnormShader = `\nstruct Params {\n  seq_len: u32,\n  hidden_size: u32,\n  eps_bits: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read> weight: array<f32>;\n@group(0) @binding(2) var<storage, read_write> output: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\nvar<workgroup> shared_sum: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let row = wid.x;\n  if (row >= params.seq_len) { return; }\n\n  let tid = lid.x;\n  let row_offset = row * params.hidden_size;\n  let eps = bitcast<f32>(params.eps_bits);\n\n  var local_sum: f32 = 0.0;\n  var i = tid;\n  while (i < params.hidden_size) {\n    let val = input[row_offset + i];\n    local_sum += val * val;\n    i += 256u;\n  }\n  shared_sum[tid] = local_sum;\n  workgroupBarrier();\n\n  var stride: u32 = 128u;\n  while (stride > 0u) {\n    if (tid < stride) {\n      shared_sum[tid] += shared_sum[tid + stride];\n    }\n    workgroupBarrier();\n    stride = stride / 2u;\n  }\n\n  let rms = sqrt(shared_sum[0] / f32(params.hidden_size) + eps);\n\n  i = tid;\n  while (i < params.hidden_size) {\n    let val = input[row_offset + i];\n    output[row_offset + i] = (val / rms) * weight[i];\n    i += 256u;\n  }\n}\n    `;\n\n    // Set up test data: hidden_size=1024, seq_len=1, eps=1e-6\n    const hiddenSize = 1024;\n    const inputData = new Float32Array(hiddenSize);\n    const weightData = new Float32Array(hiddenSize);\n    for (let i = 0; i < hiddenSize; i++) {\n      inputData[i] = Math.sin(i * 0.01) * 2.0; // range [-2, 2]\n      weightData[i] = 1.0 + Math.cos(i * 0.005) * 0.1; // range [0.9, 1.1]\n    }\n\n    // Compute expected output in JS\n    let sumSq = 0;\n    for (let i = 0; i < hiddenSize; i++) sumSq += inputData[i] * inputData[i];\n    const rms = Math.sqrt(sumSq / hiddenSize + 1e-6);\n    const expectedOutput = new Float32Array(hiddenSize);\n    for (let i = 0; i < hiddenSize; i++) {\n      expectedOutput[i] = (inputData[i] / rms) * weightData[i];\n    }\n\n    // Encode eps as u32 bits\n    const epsF32 = new Float32Array(1);\n    epsF32[0] = 1e-6;\n    const epsBits = new Uint32Array(epsF32.buffer)[0];\n\n    // Build uniform buffer: {seq_len=1, hidden_size=1024, eps_bits, _pad=0}\n    const paramsBuf = new ArrayBuffer(16);\n    const paramsView = new DataView(paramsBuf);\n    paramsView.setUint32(0, 1, true);\n    paramsView.setUint32(4, hiddenSize, true);\n    paramsView.setUint32(8, epsBits, true);\n    paramsView.setUint32(12, 0, true);\n\n    const inBuf = createStorageBuffer(ctx, \"dN_in\", inputData.byteLength, inputData);\n    const wBuf = createStorageBuffer(ctx, \"dN_w\", weightData.byteLength, weightData);\n    const outBuf = createStorageBuffer(ctx, \"dN_out\", hiddenSize * 4);\n    const uniformBuf = createUniformBuffer(ctx, \"dN_params\", paramsBuf);\n    const readBuf = createReadbackBuffer(ctx, \"dN_read\", hiddenSize * 4);\n\n    const pipeline = getOrCreatePipeline(ctx, \"dN_rmsnorm\", rmsnormShader, \"main\");\n    const bg = createBindGroup(\n      ctx,\n      pipeline,\n      [{ buffer: inBuf }, { buffer: wBuf }, { buffer: outBuf }, { buffer: uniformBuf }],\n      \"dN_bg\",\n    );\n\n    const encoder = ctx.device.createCommandEncoder();\n    const pass = encoder.beginComputePass();\n    pass.setPipeline(pipeline);\n    pass.setBindGroup(0, bg);\n    pass.dispatchWorkgroups(1); // 1 row\n    pass.end();\n    encoder.copyBufferToBuffer(outBuf, 0, readBuf, 0, hiddenSize * 4);\n    ctx.device.queue.submit([encoder.finish()]);\n    if (ctx.device.queue.onSubmittedWorkDone) {\n      await ctx.device.queue.onSubmittedWorkDone();\n    }\n    await readBuf.mapAsync(MAP_MODE_READ, 0, hiddenSize * 4);\n    const mapped = readBuf.getMappedRange(0, hiddenSize * 4);\n    const result = new Float32Array(mapped.slice(0));\n    readBuf.unmap();\n\n    let maxErr = 0;\n    let maxErrIdx = 0;\n    for (let i = 0; i < hiddenSize; i++) {\n      const err = Math.abs(result[i] - expectedOutput[i]);\n      if (err > maxErr) {\n        maxErr = err;\n        maxErrIdx = i;\n      }\n    }\n    // Allow small floating-point tolerance (tree reduce accumulation order differs)\n    const ok = maxErr < 0.01;\n    details.push(\n      ok\n        ? `✓ N: real RMSNorm kernel OK (maxErr=${maxErr.toExponential(2)} at [${maxErrIdx}])`\n        : `✗ N: real RMSNorm kernel FAILED (maxErr=${maxErr.toExponential(2)} at [${maxErrIdx}]: expected ${expectedOutput[maxErrIdx].toFixed(4)}, got ${result[maxErrIdx].toFixed(4)})`,\n    );\n    inBuf.destroy();\n    wBuf.destroy();\n    outBuf.destroy();\n    uniformBuf.destroy();\n    readBuf.destroy();\n  } catch (e) {\n    details.push(`✗ N error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test O: Real MatVec kernel (actual WGSL from registry)\n  // Tests K-parallel matrix-vector multiply with 256 threads = 8×32 layout.\n  try {\n    const matvecShader = `\nconst N_TILE: u32 = 8u;\nconst K_THREADS: u32 = 32u;\n\nstruct Params {\n  K: u32,\n  N: u32,\n}\n\n@group(0) @binding(0) var<storage, read> A: array<f32>;\n@group(0) @binding(1) var<storage, read> B: array<f32>;\n@group(0) @binding(2) var<storage, read_write> C: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\nvar<workgroup> shared_sums: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let tid = lid.x;\n  let n_idx = tid / K_THREADS;\n  let k_tid = tid % K_THREADS;\n  let col = wid.x * N_TILE + n_idx;\n  let K = params.K;\n\n  var sum: f32 = 0.0;\n  if (col < params.N) {\n    let b_off = col * K;\n    var k = k_tid * 4u;\n    while (k + 3u < K) {\n      let a_v = vec4f(A[k], A[k+1u], A[k+2u], A[k+3u]);\n      let b_v = vec4f(B[b_off+k], B[b_off+k+1u], B[b_off+k+2u], B[b_off+k+3u]);\n      sum += dot(a_v, b_v);\n      k += K_THREADS * 4u;\n    }\n  }\n\n  shared_sums[tid] = sum;\n  workgroupBarrier();\n\n  if (k_tid == 0u && col < params.N) {\n    let base = n_idx * K_THREADS;\n    var total: f32 = shared_sums[base];\n    for (var i: u32 = 1u; i < K_THREADS; i++) {\n      total += shared_sums[base + i];\n    }\n    C[col] = total;\n  }\n}\n    `;\n\n    // Test: A[256] * B[64, 256] → C[64]  (small but realistic dimensions)\n    const K_dim = 256;\n    const N_dim = 64;\n    const aData = new Float32Array(K_dim);\n    const bData = new Float32Array(N_dim * K_dim);\n    for (let i = 0; i < K_dim; i++) aData[i] = Math.sin(i * 0.05) * 0.5;\n    for (let i = 0; i < N_dim * K_dim; i++) bData[i] = Math.cos(i * 0.003) * 0.1;\n\n    // Expected: C[n] = sum_k A[k] * B[n*K + k]\n    const expectedC = new Float32Array(N_dim);\n    for (let n = 0; n < N_dim; n++) {\n      let s = 0;\n      for (let k = 0; k < K_dim; k++) s += aData[k] * bData[n * K_dim + k];\n      expectedC[n] = s;\n    }\n\n    const paramBuf2 = new ArrayBuffer(8);\n    const pv2 = new DataView(paramBuf2);\n    pv2.setUint32(0, K_dim, true);\n    pv2.setUint32(4, N_dim, true);\n\n    const aBuf = createStorageBuffer(ctx, \"dO_a\", aData.byteLength, aData);\n    const bBuf = createStorageBuffer(ctx, \"dO_b\", bData.byteLength, bData);\n    const cBuf = createStorageBuffer(ctx, \"dO_c\", N_dim * 4);\n    const uBuf = createUniformBuffer(ctx, \"dO_params\", paramBuf2);\n    const rBuf = createReadbackBuffer(ctx, \"dO_read\", N_dim * 4);\n\n    const mvPipeline = getOrCreatePipeline(ctx, \"dO_matvec\", matvecShader, \"main\");\n    const mvBg = createBindGroup(\n      ctx,\n      mvPipeline,\n      [{ buffer: aBuf }, { buffer: bBuf }, { buffer: cBuf }, { buffer: uBuf }],\n      \"dO_bg\",\n    );\n\n    // Dispatch: N_dim / 8 = 8 workgroups\n    const encoder2 = ctx.device.createCommandEncoder();\n    const pass2 = encoder2.beginComputePass();\n    pass2.setPipeline(mvPipeline);\n    pass2.setBindGroup(0, mvBg);\n    pass2.dispatchWorkgroups(Math.ceil(N_dim / 8));\n    pass2.end();\n    encoder2.copyBufferToBuffer(cBuf, 0, rBuf, 0, N_dim * 4);\n    ctx.device.queue.submit([encoder2.finish()]);\n    if (ctx.device.queue.onSubmittedWorkDone) {\n      await ctx.device.queue.onSubmittedWorkDone();\n    }\n    await rBuf.mapAsync(MAP_MODE_READ, 0, N_dim * 4);\n    const mapped2 = rBuf.getMappedRange(0, N_dim * 4);\n    const resultC = new Float32Array(mapped2.slice(0));\n    rBuf.unmap();\n\n    let maxErr2 = 0;\n    let maxErrIdx2 = 0;\n    for (let i = 0; i < N_dim; i++) {\n      const err = Math.abs(resultC[i] - expectedC[i]);\n      if (err > maxErr2) {\n        maxErr2 = err;\n        maxErrIdx2 = i;\n      }\n    }\n    const ok2 = maxErr2 < 0.01;\n    details.push(\n      ok2\n        ? `✓ O: real MatVec kernel OK (maxErr=${maxErr2.toExponential(2)} at [${maxErrIdx2}])`\n        : `✗ O: real MatVec kernel FAILED (maxErr=${maxErr2.toExponential(2)} at [${maxErrIdx2}]: expected ${expectedC[maxErrIdx2].toFixed(4)}, got ${resultC[maxErrIdx2].toFixed(4)})`,\n    );\n    aBuf.destroy();\n    bBuf.destroy();\n    cBuf.destroy();\n    uBuf.destroy();\n    rBuf.destroy();\n  } catch (e) {\n    details.push(`✗ O error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test P: Many dispatches in one compute pass (300 sequential dispatches)\n  // The model runs ~300 dispatches per forward pass in a single compute pass.\n  // If Safari silently drops or corrupts dispatches beyond a limit, only this test catches it.\n  // Uses a simple chain: buf[i] += 1.0 for 300 dispatches → should equal 300.0.\n  try {\n    const manyShader = `\n      @group(0) @binding(0) var<storage, read_write> data: array<f32>;\n      @compute @workgroup_size(64)\n      fn main(@builtin(global_invocation_id) gid: vec3u) {\n        if (gid.x < 4u) { data[gid.x] += 1.0; }\n      }\n    `;\n    const DISPATCH_COUNT = 300;\n    const manyBuf = createStorageBuffer(ctx, \"dP_data\", 16, new Float32Array(4));\n    const manyReadBuf = createReadbackBuffer(ctx, \"dP_read\", 16);\n    const manyPipeline = getOrCreatePipeline(ctx, \"dP\", manyShader, \"main\");\n    const manyBg = createBindGroup(ctx, manyPipeline, [{ buffer: manyBuf }], \"dP_bg\");\n\n    const enc = ctx.device.createCommandEncoder();\n    const cpass = enc.beginComputePass();\n    for (let d = 0; d < DISPATCH_COUNT; d++) {\n      cpass.setPipeline(manyPipeline);\n      cpass.setBindGroup(0, manyBg);\n      cpass.dispatchWorkgroups(1);\n    }\n    cpass.end();\n    enc.copyBufferToBuffer(manyBuf, 0, manyReadBuf, 0, 16);\n    ctx.device.queue.submit([enc.finish()]);\n    if (ctx.device.queue.onSubmittedWorkDone) {\n      await ctx.device.queue.onSubmittedWorkDone();\n    }\n    await manyReadBuf.mapAsync(MAP_MODE_READ, 0, 16);\n    const manyMapped = manyReadBuf.getMappedRange(0, 16);\n    const manyResult = new Float32Array(manyMapped.slice(0));\n    manyReadBuf.unmap();\n\n    const allCorrect = manyResult.every((v) => Math.abs(v - DISPATCH_COUNT) < 0.1);\n    details.push(\n      allCorrect\n        ? `✓ P: ${DISPATCH_COUNT} dispatches OK (${manyResult[0]})`\n        : `✗ P: ${DISPATCH_COUNT} dispatches FAILED (got [${Array.from(manyResult).join(\",\")}], expected ${DISPATCH_COUNT})`,\n    );\n    manyBuf.destroy();\n    manyReadBuf.destroy();\n  } catch (e) {\n    details.push(`✗ P error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  // Test Q: Same shader, different bind groups (Metal argument buffer caching bug)\n  // This is the EXACT pattern that triggers the Safari/Metal bug:\n  // Two dispatches use the same shader code but different bind groups pointing\n  // to different output buffers. If Metal caches argument buffers per compiled\n  // function, the second dispatch will write to the wrong buffer.\n  try {\n    const qShader = `\n      @group(0) @binding(0) var<storage, read> src: array<f32>;\n      @group(0) @binding(1) var<storage, read_write> dst: array<f32>;\n      @compute @workgroup_size(4)\n      fn main(@builtin(global_invocation_id) gid: vec3u) {\n        if (gid.x < 4u) { dst[gid.x] = src[gid.x] * 2.0; }\n      }\n    `;\n    const qSrc = createStorageBuffer(ctx, \"dQ_src\", 16, new Float32Array([1, 2, 3, 4]));\n    const qDst1 = createStorageBuffer(ctx, \"dQ_dst1\", 16, new Float32Array(4)); // zeros\n    const qDst2 = createStorageBuffer(ctx, \"dQ_dst2\", 16, new Float32Array(4)); // zeros\n    const qRead1 = createReadbackBuffer(ctx, \"dQ_read1\", 16);\n    const qRead2 = createReadbackBuffer(ctx, \"dQ_read2\", 16);\n\n    // Same shader, same pipeline — different bind groups with different output buffers\n    const qPipeline = getOrCreatePipeline(ctx, \"dQ\", qShader, \"main\");\n    const qBg1 = createBindGroup(ctx, qPipeline, [{ buffer: qSrc }, { buffer: qDst1 }], \"dQ_bg1\");\n    const qBg2 = createBindGroup(ctx, qPipeline, [{ buffer: qSrc }, { buffer: qDst2 }], \"dQ_bg2\");\n\n    const qEnc = ctx.device.createCommandEncoder();\n    const qPass = qEnc.beginComputePass();\n    qPass.setPipeline(qPipeline);\n    qPass.setBindGroup(0, qBg1);\n    qPass.dispatchWorkgroups(1); // dst1 should be [2,4,6,8]\n    qPass.setPipeline(qPipeline); // SAME pipeline\n    qPass.setBindGroup(0, qBg2); // DIFFERENT bind group → different output buffer\n    qPass.dispatchWorkgroups(1); // dst2 should be [2,4,6,8]\n    qPass.end();\n    qEnc.copyBufferToBuffer(qDst1, 0, qRead1, 0, 16);\n    qEnc.copyBufferToBuffer(qDst2, 0, qRead2, 0, 16);\n    ctx.device.queue.submit([qEnc.finish()]);\n    if (ctx.device.queue.onSubmittedWorkDone) {\n      await ctx.device.queue.onSubmittedWorkDone();\n    }\n    await qRead1.mapAsync(MAP_MODE_READ, 0, 16);\n    const qR1 = new Float32Array(qRead1.getMappedRange(0, 16).slice(0));\n    qRead1.unmap();\n    await qRead2.mapAsync(MAP_MODE_READ, 0, 16);\n    const qR2 = new Float32Array(qRead2.getMappedRange(0, 16).slice(0));\n    qRead2.unmap();\n\n    const qExpected = [2, 4, 6, 8];\n    const dst1Ok = qExpected.every((v, i) => Math.abs(qR1[i] - v) < 0.01);\n    const dst2Ok = qExpected.every((v, i) => Math.abs(qR2[i] - v) < 0.01);\n    details.push(\n      dst1Ok && dst2Ok\n        ? `✓ Q: same-shader diff-bindgroup OK (dst1=[${Array.from(qR1)}] dst2=[${Array.from(qR2)}])`\n        : `✗ Q: same-shader diff-bindgroup FAILED (dst1=[${Array.from(qR1)}] dst2=[${Array.from(qR2)}], expected [${qExpected}])`,\n    );\n    qSrc.destroy();\n    qDst1.destroy();\n    qDst2.destroy();\n    qRead1.destroy();\n    qRead2.destroy();\n  } catch (e) {\n    details.push(`✗ Q error: ${e instanceof Error ? e.message : String(e)}`);\n  }\n\n  return { bufferIntegrity, computeWorks, sharedMemoryWorks, details };\n}\n","/**\n * Kernel registry — maps IR op types to WGSL shader code and dispatch logic.\n *\n * Each entry defines:\n *  - The WGSL source (embedded as string constants for browser compatibility)\n *  - The compute entry point name\n *  - How to calculate workgroup dispatch dimensions\n *  - How to build the uniform buffer matching the WGSL Params struct\n *  - The binding layout (storage-read, storage-read-write, uniform) in order\n *\n * Only ops with WGSL implementations are included. Stubbed ops (MoERouter,\n * ExpertMatMul, Conv2d, AvgPool2d, CrossAttention, Gather, Reshape,\n * Transpose, Concat) are omitted.\n */\n\nimport type { OpNode, OpType } from \"../ir.js\";\n\n// ── KernelSpec interface ────────────────────────────────────────────────\n\n/** Runtime context passed from the executor to kernels at dispatch time. */\nexport interface RuntimeContext {\n  /** Current sequence position (for autoregressive decode). */\n  seqPos: number;\n  /**\n   * Query-grid base offset for windowed Attention dispatches. When set, the\n   * Attention kernel treats workgroup row r as query position r + qOffset, so a\n   * caller can process a sub-range of query rows in one dispatch. Only the WebKit\n   * vision encoder sets it (to chunk the O(N²) ViT attention); defaults to 0.\n   */\n  qOffset?: number;\n}\n\nexport interface KernelSpec {\n  /** WGSL shader source code */\n  shaderCode: string;\n  /** Compute entry point name */\n  entryPoint: string;\n  /** Calculate dispatch workgroup counts from op attributes and resolved shapes */\n  getDispatchSize: (\n    op: OpNode,\n    resolvedShapes: Record<string, number[]>,\n    context?: RuntimeContext,\n  ) => [number, number, number];\n  /** Build the uniform buffer data for this op */\n  buildParams: (\n    op: OpNode,\n    resolvedShapes: Record<string, number[]>,\n    context?: RuntimeContext,\n  ) => ArrayBuffer;\n  /** List of binding entries (storage vs uniform) in order */\n  bindings: Array<{\n    type: \"storage-read\" | \"storage-read-write\" | \"uniform\";\n  }>;\n}\n\n// ── Helpers ─────────────────────────────────────────────────────────────\n\n/** Ceiling division: ceil(a / b) */\nfunction cdiv(a: number, b: number): number {\n  return Math.ceil(a / b);\n}\n\n/** Reinterpret an f32 value as its u32 bit pattern. */\nfunction f32BitsToU32(value: number): number {\n  const f = new Float32Array(1);\n  f[0] = value;\n  return new Uint32Array(f.buffer)[0];\n}\n\n/**\n * Create an ArrayBuffer from an array of u32 values.\n * Each value is written at 4-byte intervals using DataView.\n */\nfunction buildUniformBuffer(values: number[]): ArrayBuffer {\n  const buf = new ArrayBuffer(values.length * 4);\n  const view = new DataView(buf);\n  for (let i = 0; i < values.length; i++) {\n    view.setUint32(i * 4, values[i], true /* littleEndian */);\n  }\n  return buf;\n}\n\n/**\n * Resolve the first dimension of a tensor referenced by an attribute.\n * Falls back to the direct attribute value if no tensor reference is set.\n */\nfunction _resolveFirstDim(\n  op: OpNode,\n  attrName: string,\n  resolvedShapes: Record<string, number[]>,\n): number {\n  const tensorRef = op.attributes[`${attrName}_tensor`] as string | undefined;\n  if (tensorRef && resolvedShapes[tensorRef]) {\n    return resolvedShapes[tensorRef][0];\n  }\n  return op.attributes[attrName] as number;\n}\n\n/**\n * Resolve total element count of a tensor referenced by an attribute.\n */\nfunction _resolveTotalElements(\n  _op: OpNode,\n  tensorRef: string,\n  resolvedShapes: Record<string, number[]>,\n): number {\n  const shape = resolvedShapes[tensorRef];\n  if (!shape) return 0;\n  return shape.reduce((a, b) => a * b, 1);\n}\n\n/**\n * Resolve a dynamic seq_len from tensor references, output shape, or static attribute.\n *\n * Tries (in order):\n *  1. Explicit tensor reference via `seq_len_tensor` attribute — computes total / hidden\n *  2. First output tensor's first dimension (for 2D+ outputs)\n *  3. Static `seq_len` attribute\n */\nfunction resolveSeqLen(\n  op: OpNode,\n  resolvedShapes: Record<string, number[]>,\n  hidden: number,\n): number {\n  // Try explicit tensor reference first\n  const ref = op.attributes.seq_len_tensor as string | undefined;\n  if (ref && resolvedShapes[ref]) {\n    const shape = resolvedShapes[ref];\n    const total = shape.reduce((a: number, b: number) => a * b, 1);\n    return total / hidden;\n  }\n  // Fall back to output tensor shape\n  const outShape = resolvedShapes[op.outputs[0]];\n  if (outShape && outShape.length >= 2) {\n    return outShape[0];\n  }\n  return op.attributes.seq_len as number;\n}\n\n// ── Embedded WGSL shader sources ────────────────────────────────────────\n\nconst WGSL_EMBEDDING = `\\\n// Embedding lookup: gather rows from weight matrix by token ID.\n\nstruct Params {\n  seq_len: u32,\n  hidden_size: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input_ids: array<u32>;\n@group(0) @binding(1) var<storage, read> weight: array<f32>;\n@group(0) @binding(2) var<storage, read_write> output: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  let total = params.seq_len * params.hidden_size;\n  if (idx >= total) { return; }\n\n  let token_pos = idx / params.hidden_size;\n  let dim = idx % params.hidden_size;\n\n  let token_id = input_ids[token_pos];\n  let weight_idx = token_id * params.hidden_size + dim;\n\n  output[idx] = weight[weight_idx];\n}\n`;\n\nconst WGSL_EMBEDDING_INT4 = `\\\n// INT4 embedding lookup: dequantize rows from packed weight matrix by token ID.\n// Weight is packed as 8 nibbles per u32, with per-group scales and zeros.\n// Dequant formula: output = (nibble - zero) * scale\n//\n// Bindings:\n//   @group(0) @binding(0) input_ids: array<u32>     — token IDs, length T\n//   @group(0) @binding(1) weight_q: array<u32>      — packed INT4 [vocab_size * hidden_size / 8]\n//   @group(0) @binding(2) scales: array<f32>         — per-group scales\n//   @group(0) @binding(3) zeros: array<f32>          — per-group zeros\n//   @group(0) @binding(4) output: array<f32>         — output [T, hidden_size]\n//   @group(0) @binding(5) params: { seq_len, hidden_size, group_size }\n\nstruct Params {\n  seq_len: u32,\n  hidden_size: u32,\n  group_size: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input_ids: array<u32>;\n@group(0) @binding(1) var<storage, read> weight_q: array<u32>;\n@group(0) @binding(2) var<storage, read> scales: array<f32>;\n@group(0) @binding(3) var<storage, read> zeros: array<f32>;\n@group(0) @binding(4) var<storage, read_write> output: array<f32>;\n@group(0) @binding(5) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  let total = params.seq_len * params.hidden_size;\n  if (idx >= total) { return; }\n\n  let token_pos = idx / params.hidden_size;\n  let dim = idx % params.hidden_size;\n\n  let token_id = input_ids[token_pos];\n\n  // Flat index into the [vocab_size, hidden_size] embedding table\n  let flat_idx = token_id * params.hidden_size + dim;\n  let packed_idx = flat_idx / 8u;\n  let nibble_pos = flat_idx % 8u;\n  let packed = weight_q[packed_idx];\n  let shift = nibble_pos * 4u;\n  let raw_val = f32((packed >> shift) & 0xFu);\n\n  let group_idx = flat_idx / params.group_size;\n  let scale = scales[group_idx];\n  let zero = zeros[group_idx];\n\n  output[idx] = (raw_val - zero) * scale;\n}\n`;\n\nconst WGSL_MATMUL = `\\\n// Tiled matrix multiply with 4x2 register blocking.\n// C[M,N] = A[M,K] * B^T[N,K]; B stored in [N,K] layout (HF [out,in]).\n//\n// A 16x16 workgroup computes a 64-row x 32-col output tile: each thread owns a\n// 4x2 block of outputs (rows {r,r+16,r+32,r+48} x cols {c,c+16}). Each B column\n// loaded into shared memory is reused for 4 output rows, cutting B (weight)\n// global traffic 4x vs the scalar version — the dominant cost of the wide f32\n// ViT matmuls. A/B tile fills use vec4 loads (K is always %4==0).\n\nconst MT: u32 = 64u;     // output rows per workgroup\nconst NT: u32 = 32u;     // output cols per workgroup\nconst KT: u32 = 16u;     // K-dimension tile depth\nconst KT4: u32 = 4u;     // KT / 4 (vec4 columns along K)\n\nstruct Params {\n  M: u32,\n  K: u32,\n  N: u32,\n}\n\n@group(0) @binding(0) var<storage, read> A: array<vec4f>;\n@group(0) @binding(1) var<storage, read> B: array<vec4f>;\n@group(0) @binding(2) var<storage, read_write> C: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\n// tileA: [MT rows x KT] = 1024, tileB: [KT x NT cols] = 512\nvar<workgroup> tileA: array<f32, 1024>;\nvar<workgroup> tileB: array<f32, 512>;\n\n@compute @workgroup_size(16, 16)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let lr = lid.y;\n  let lc = lid.x;\n  let mBase = wid.y * MT;\n  let nBase = wid.x * NT;\n  // This thread's 4x2 output block: rows {lr, lr+16, lr+32, lr+48}, cols {lc, lc+16}.\n  let row0 = mBase + lr;\n  let col0 = nBase + lc;\n\n  var acc: array<f32, 8>; // [4 rows x 2 cols]\n  for (var i = 0u; i < 8u; i = i + 1u) { acc[i] = 0.0; }\n\n  let numTiles = (params.K + KT - 1u) / KT;\n  let tid = lr * 16u + lc; // 0..255 linear thread id\n  let K4 = params.K / 4u;  // K in vec4 units\n\n  for (var t: u32 = 0u; t < numTiles; t = t + 1u) {\n    let kv0 = t * KT4; // first vec4 column of this K-tile\n    // Load A tile [MT x KT] = 64 rows * 4 vec4 = 256 vec4 loads (1 per thread).\n    {\n      let ar = tid / KT4;            // 0..63 row within tile\n      let av4 = tid % KT4;           // 0..3 vec4-col within tile (along K)\n      let gRow = mBase + ar;\n      var v = vec4f(0.0, 0.0, 0.0, 0.0);\n      if (gRow < params.M && kv0 + av4 < K4) {\n        v = A[gRow * K4 + kv0 + av4];\n      }\n      let base = ar * KT + av4 * 4u;\n      tileA[base] = v.x;\n      tileA[base + 1u] = v.y;\n      tileA[base + 2u] = v.z;\n      tileA[base + 3u] = v.w;\n    }\n    // Load B tile [KT x NT cols]: 32 cols * 4 vec4 = 128 vec4 loads (threads 0..127).\n    if (tid < 128u) {\n      let bc = tid / KT4;            // 0..31 col within tile\n      let bv4 = tid % KT4;           // 0..3 vec4-col within tile (along K)\n      let gCol = nBase + bc;\n      var v = vec4f(0.0, 0.0, 0.0, 0.0);\n      if (gCol < params.N && kv0 + bv4 < K4) {\n        v = B[gCol * K4 + kv0 + bv4];\n      }\n      let krow = bv4 * 4u;\n      tileB[(krow + 0u) * NT + bc] = v.x;\n      tileB[(krow + 1u) * NT + bc] = v.y;\n      tileB[(krow + 2u) * NT + bc] = v.z;\n      tileB[(krow + 3u) * NT + bc] = v.w;\n    }\n\n    workgroupBarrier();\n\n    for (var k: u32 = 0u; k < KT; k = k + 1u) {\n      let b0 = tileB[k * NT + lc];\n      let b1 = tileB[k * NT + lc + 16u];\n      let a0 = tileA[lr * KT + k];\n      let a1 = tileA[(lr + 16u) * KT + k];\n      let a2 = tileA[(lr + 32u) * KT + k];\n      let a3 = tileA[(lr + 48u) * KT + k];\n      acc[0] += a0 * b0;\n      acc[1] += a0 * b1;\n      acc[2] += a1 * b0;\n      acc[3] += a1 * b1;\n      acc[4] += a2 * b0;\n      acc[5] += a2 * b1;\n      acc[6] += a3 * b0;\n      acc[7] += a3 * b1;\n    }\n\n    workgroupBarrier();\n  }\n\n  for (var rr = 0u; rr < 4u; rr = rr + 1u) {\n    let gRow = row0 + rr * 16u;\n    if (gRow < params.M) {\n      if (col0 < params.N) {\n        C[gRow * params.N + col0] = acc[rr * 2u];\n      }\n      if (col0 + 16u < params.N) {\n        C[gRow * params.N + col0 + 16u] = acc[rr * 2u + 1u];\n      }\n    }\n  }\n}\n`;\n\n// Fused MatMul + row-broadcast bias: C[r,c] = A[r,:]·B[c,:] + bias[c].\n// Identical tiling to WGSL_MATMUL (4x2 register blocking, vec4 loads) but adds\n// the bias at store time, eliminating a full wide read+write of the matmul\n// output that a separate AddBias dispatch would incur — one fewer round-trip per\n// ViT linear layer (qkv/proj/fc1/fc2/patch_embed/merger fc1/fc2).\nconst WGSL_MATMUL_BIAS = `\\\nconst MT: u32 = 64u;\nconst NT: u32 = 32u;\nconst KT: u32 = 16u;\nconst KT4: u32 = 4u;\n\nstruct Params {\n  M: u32,\n  K: u32,\n  N: u32,\n}\n\n@group(0) @binding(0) var<storage, read> A: array<vec4f>;\n@group(0) @binding(1) var<storage, read> B: array<vec4f>;\n@group(0) @binding(2) var<storage, read> bias: array<f32>;\n@group(0) @binding(3) var<storage, read_write> C: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\nvar<workgroup> tileA: array<f32, 1024>;\nvar<workgroup> tileB: array<f32, 512>;\n\n@compute @workgroup_size(16, 16)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let lr = lid.y;\n  let lc = lid.x;\n  let mBase = wid.y * MT;\n  let nBase = wid.x * NT;\n  let row0 = mBase + lr;\n  let col0 = nBase + lc;\n\n  var acc: array<f32, 8>;\n  for (var i = 0u; i < 8u; i = i + 1u) { acc[i] = 0.0; }\n\n  let numTiles = (params.K + KT - 1u) / KT;\n  let tid = lr * 16u + lc;\n  let K4 = params.K / 4u;\n\n  for (var t: u32 = 0u; t < numTiles; t = t + 1u) {\n    let kv0 = t * KT4;\n    {\n      let ar = tid / KT4;\n      let av4 = tid % KT4;\n      let gRow = mBase + ar;\n      var v = vec4f(0.0, 0.0, 0.0, 0.0);\n      if (gRow < params.M && kv0 + av4 < K4) {\n        v = A[gRow * K4 + kv0 + av4];\n      }\n      let base = ar * KT + av4 * 4u;\n      tileA[base] = v.x;\n      tileA[base + 1u] = v.y;\n      tileA[base + 2u] = v.z;\n      tileA[base + 3u] = v.w;\n    }\n    if (tid < 128u) {\n      let bc = tid / KT4;\n      let bv4 = tid % KT4;\n      let gCol = nBase + bc;\n      var v = vec4f(0.0, 0.0, 0.0, 0.0);\n      if (gCol < params.N && kv0 + bv4 < K4) {\n        v = B[gCol * K4 + kv0 + bv4];\n      }\n      let krow = bv4 * 4u;\n      tileB[(krow + 0u) * NT + bc] = v.x;\n      tileB[(krow + 1u) * NT + bc] = v.y;\n      tileB[(krow + 2u) * NT + bc] = v.z;\n      tileB[(krow + 3u) * NT + bc] = v.w;\n    }\n\n    workgroupBarrier();\n\n    for (var k: u32 = 0u; k < KT; k = k + 1u) {\n      let b0 = tileB[k * NT + lc];\n      let b1 = tileB[k * NT + lc + 16u];\n      let a0 = tileA[lr * KT + k];\n      let a1 = tileA[(lr + 16u) * KT + k];\n      let a2 = tileA[(lr + 32u) * KT + k];\n      let a3 = tileA[(lr + 48u) * KT + k];\n      acc[0] += a0 * b0;\n      acc[1] += a0 * b1;\n      acc[2] += a1 * b0;\n      acc[3] += a1 * b1;\n      acc[4] += a2 * b0;\n      acc[5] += a2 * b1;\n      acc[6] += a3 * b0;\n      acc[7] += a3 * b1;\n    }\n\n    workgroupBarrier();\n  }\n\n  var bias0 = 0.0;\n  if (col0 < params.N) { bias0 = bias[col0]; }\n  var bias1 = 0.0;\n  if (col0 + 16u < params.N) { bias1 = bias[col0 + 16u]; }\n  for (var rr = 0u; rr < 4u; rr = rr + 1u) {\n    let gRow = row0 + rr * 16u;\n    if (gRow < params.M) {\n      if (col0 < params.N) {\n        C[gRow * params.N + col0] = acc[rr * 2u] + bias0;\n      }\n      if (col0 + 16u < params.N) {\n        C[gRow * params.N + col0 + 16u] = acc[rr * 2u + 1u] + bias1;\n      }\n    }\n  }\n}\n`;\n\n// MIXED-precision MatMulBias: f16 weight + f16 shared tiles, but products are\n// summed in an f32 accumulator (acc += f32(a*b), with a*b done in f16). Keeps the\n// fast f16 MULTIPLY (the ~2% win measured in b4-r5) while accumulating in f32 to\n// stay inside the cos gate that full-f16 accumulation (r5) broke.\nconst WGSL_MATMUL_BIAS_F16MIX = `\\\nenable f16;\n\nconst MT: u32 = 64u;\nconst NT: u32 = 32u;\nconst KT: u32 = 16u;\nconst KT4: u32 = 4u;\n\nstruct Params {\n  M: u32,\n  K: u32,\n  N: u32,\n}\n\n@group(0) @binding(0) var<storage, read> A: array<vec4f>;\n@group(0) @binding(1) var<storage, read> B: array<vec4<f16>>;\n@group(0) @binding(2) var<storage, read> bias: array<f32>;\n@group(0) @binding(3) var<storage, read_write> C: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\nvar<workgroup> tileA: array<f16, 1024>;\nvar<workgroup> tileB: array<f16, 512>;\n\n@compute @workgroup_size(16, 16)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let lr = lid.y;\n  let lc = lid.x;\n  let mBase = wid.y * MT;\n  let nBase = wid.x * NT;\n  let row0 = mBase + lr;\n  let col0 = nBase + lc;\n\n  var acc: array<f32, 8>;\n  for (var i = 0u; i < 8u; i = i + 1u) { acc[i] = 0.0; }\n\n  let numTiles = (params.K + KT - 1u) / KT;\n  let tid = lr * 16u + lc;\n  let K4 = params.K / 4u;\n\n  for (var t: u32 = 0u; t < numTiles; t = t + 1u) {\n    let kv0 = t * KT4;\n    {\n      let ar = tid / KT4;\n      let av4 = tid % KT4;\n      let gRow = mBase + ar;\n      var v = vec4f(0.0, 0.0, 0.0, 0.0);\n      if (gRow < params.M && kv0 + av4 < K4) {\n        v = A[gRow * K4 + kv0 + av4];\n      }\n      let hv = vec4<f16>(v);\n      let base = ar * KT + av4 * 4u;\n      tileA[base] = hv.x;\n      tileA[base + 1u] = hv.y;\n      tileA[base + 2u] = hv.z;\n      tileA[base + 3u] = hv.w;\n    }\n    if (tid < 128u) {\n      let bc = tid / KT4;\n      let bv4 = tid % KT4;\n      let gCol = nBase + bc;\n      var v = vec4<f16>(f16(0.0), f16(0.0), f16(0.0), f16(0.0));\n      if (gCol < params.N && kv0 + bv4 < K4) {\n        v = B[gCol * K4 + kv0 + bv4];\n      }\n      let krow = bv4 * 4u;\n      tileB[(krow + 0u) * NT + bc] = v.x;\n      tileB[(krow + 1u) * NT + bc] = v.y;\n      tileB[(krow + 2u) * NT + bc] = v.z;\n      tileB[(krow + 3u) * NT + bc] = v.w;\n    }\n\n    workgroupBarrier();\n\n    for (var k: u32 = 0u; k < KT; k = k + 1u) {\n      let b0 = tileB[k * NT + lc];\n      let b1 = tileB[k * NT + lc + 16u];\n      let a0 = tileA[lr * KT + k];\n      let a1 = tileA[(lr + 16u) * KT + k];\n      let a2 = tileA[(lr + 32u) * KT + k];\n      let a3 = tileA[(lr + 48u) * KT + k];\n      acc[0] += f32(a0 * b0);\n      acc[1] += f32(a0 * b1);\n      acc[2] += f32(a1 * b0);\n      acc[3] += f32(a1 * b1);\n      acc[4] += f32(a2 * b0);\n      acc[5] += f32(a2 * b1);\n      acc[6] += f32(a3 * b0);\n      acc[7] += f32(a3 * b1);\n    }\n\n    workgroupBarrier();\n  }\n\n  var bias0 = 0.0;\n  if (col0 < params.N) { bias0 = bias[col0]; }\n  var bias1 = 0.0;\n  if (col0 + 16u < params.N) { bias1 = bias[col0 + 16u]; }\n  for (var rr = 0u; rr < 4u; rr = rr + 1u) {\n    let gRow = row0 + rr * 16u;\n    if (gRow < params.M) {\n      if (col0 < params.N) {\n        C[gRow * params.N + col0] = acc[rr * 2u] + bias0;\n      }\n      if (col0 + 16u < params.N) {\n        C[gRow * params.N + col0 + 16u] = acc[rr * 2u + 1u] + bias1;\n      }\n    }\n  }\n}\n`;\n\nconst WGSL_MATMUL_INT4 = `\\\n// Tiled INT4 dequantize + matrix multiply.\n// C[M,N] = A[M,K] * dequant(B_q[N,K])\n// B_q is packed nibbles in [N,K] layout (same as F32 MatMul convention).\n// 8 nibbles per u32, little-nibble-first.\n\nconst TILE: u32 = 16u;\n\nstruct Params {\n  M: u32,\n  K: u32,\n  N: u32,\n  group_size: u32,\n}\n\n@group(0) @binding(0) var<storage, read> A: array<f32>;\n@group(0) @binding(1) var<storage, read> B_q: array<u32>;\n@group(0) @binding(2) var<storage, read> scales: array<f32>;\n@group(0) @binding(3) var<storage, read> zeros: array<f32>;\n@group(0) @binding(4) var<storage, read_write> C: array<f32>;\n@group(0) @binding(5) var<storage, read> params: Params;\n\nvar<workgroup> tileA: array<f32, 256>;\nvar<workgroup> tileB: array<f32, 256>;\n\n@compute @workgroup_size(16, 16)\nfn main(\n  @builtin(global_invocation_id) gid: vec3u,\n  @builtin(local_invocation_id) lid: vec3u,\n) {\n  let row = gid.y;\n  let col = gid.x;\n  let lr = lid.y;\n  let lc = lid.x;\n\n  var sum: f32 = 0.0;\n  let numTiles = (params.K + TILE - 1u) / TILE;\n\n  for (var t: u32 = 0u; t < numTiles; t = t + 1u) {\n    // Load tile of A into shared memory\n    let a_col = t * TILE + lc;\n    if (row < params.M && a_col < params.K) {\n      tileA[lr * TILE + lc] = A[row * params.K + a_col];\n    } else {\n      tileA[lr * TILE + lc] = 0.0;\n    }\n\n    // Load tile of B (INT4 dequantized) into shared memory\n    // B is stored as [N,K] packed nibbles. We need B[col, t*TILE+lr].\n    let b_k = t * TILE + lr;\n    if (b_k < params.K && col < params.N) {\n      let flat_idx = col * params.K + b_k;\n      let packed_idx = flat_idx / 8u;\n      let nibble_pos = flat_idx % 8u;\n      let packed = B_q[packed_idx];\n      let shift = nibble_pos * 4u;\n      let raw_val = f32((packed >> shift) & 0xFu);\n\n      let group_idx = flat_idx / params.group_size;\n      let scale = scales[group_idx];\n      let zero = zeros[group_idx];\n      tileB[lr * TILE + lc] = (raw_val - zero) * scale;\n    } else {\n      tileB[lr * TILE + lc] = 0.0;\n    }\n\n    workgroupBarrier();\n\n    for (var k: u32 = 0u; k < TILE; k = k + 1u) {\n      sum += tileA[lr * TILE + k] * tileB[k * TILE + lc];\n    }\n\n    workgroupBarrier();\n  }\n\n  if (row < params.M && col < params.N) {\n    C[row * params.N + col] = sum;\n  }\n}\n`;\n\nconst WGSL_MATVEC = `\\\n// K-parallel matrix-vector multiply: C[N] = A[K] * B^T[N,K]\n// Optimized for M=1 decode on Apple Silicon (SIMD width 32).\n//\n// Design: 256 threads = 8 output columns × 32 K-threads per column.\n// Each group of 32 threads cooperates along K for one output element.\n// A reads go through L1 cache (no shared memory — maximizes occupancy).\n// B reads are coalesced: 32 adjacent threads read 32 adjacent f32 values\n// = 128 bytes = one full cache line per iteration.\n// Single-barrier reduction (thread 0 of each group sums 32 partials).\n\nconst N_TILE: u32 = 8u;\nconst K_THREADS: u32 = 32u;\n\nstruct Params {\n  K: u32,\n  N: u32,\n}\n\n@group(0) @binding(0) var<storage, read> A: array<f32>;\n@group(0) @binding(1) var<storage, read> B: array<f32>;\n@group(0) @binding(2) var<storage, read_write> C: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\nvar<workgroup> shared_sums: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let tid = lid.x;\n  let n_idx = tid / K_THREADS;\n  let k_tid = tid % K_THREADS;\n  let col = wid.x * N_TILE + n_idx;\n  let K = params.K;\n\n  // K-parallel accumulation with vec4 dot products\n  // A reads hit L1 cache (same addresses across all 16 output columns)\n  var sum: f32 = 0.0;\n  if (col < params.N) {\n    let b_off = col * K;\n    var k = k_tid * 4u;\n    while (k + 3u < K) {\n      let a_v = vec4f(A[k], A[k+1u], A[k+2u], A[k+3u]);\n      let b_v = vec4f(B[b_off+k], B[b_off+k+1u], B[b_off+k+2u], B[b_off+k+3u]);\n      sum += dot(a_v, b_v);\n      k += K_THREADS * 4u;\n    }\n  }\n\n  // Store partials and reduce\n  shared_sums[tid] = sum;\n  workgroupBarrier();\n\n  // Thread 0 of each K-group sequentially sums 16 partials\n  if (k_tid == 0u && col < params.N) {\n    let base = n_idx * K_THREADS;\n    var total: f32 = shared_sums[base];\n    for (var i: u32 = 1u; i < K_THREADS; i++) {\n      total += shared_sums[base + i];\n    }\n    C[col] = total;\n  }\n}\n`;\n\n// ── Subgroups variant of WGSL_MATVEC (desktop-only fast path) ──\n//\n// Identical math and memory-access pattern to WGSL_MATVEC, but the per-column\n// partial-sum reduction is done with subgroupAdd() instead of a shared-memory\n// write + barrier + thread-0 sequential sum. Requires the \"subgroups\" device\n// feature (Chrome 134+, Safari 26+) and \"enable subgroups;\".\n//\n// Layout: 256 threads = 8 output columns × 32 K-threads. The kernel is hand-\n// tuned for SIMD width 32 (Apple Silicon / desktop GPUs), where each group of\n// 32 contiguous K-threads forms exactly one subgroup. We do NOT rely on the\n// (unspecified) lane↔invocation mapping for correctness across arbitrary\n// subgroup sizes: subgroupAdd() reduces over whatever subgroup the lane belongs\n// to, then the subgroup leaders deposit their partials into a tiny shared array\n// (one slot per column), and the column leader combines them. When subgroup\n// size == K_THREADS (the common desktop case) there is exactly one subgroup per\n// column and the shared-memory combine collapses to a single store + load.\n//\n// This variant is selected only on non-WebKit devices that expose subgroups\n// (see executor.ts) — the portable WGSL_MATVEC remains the universal fallback.\nconst WGSL_MATVEC_SUBGROUPS = `\\\nenable subgroups;\n\nconst N_TILE: u32 = 8u;\nconst K_THREADS: u32 = 32u;\n\nstruct Params {\n  K: u32,\n  N: u32,\n}\n\n@group(0) @binding(0) var<storage, read> A: array<f32>;\n@group(0) @binding(1) var<storage, read> B: array<f32>;\n@group(0) @binding(2) var<storage, read_write> C: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\n// One partial slot per column per possible subgroup-within-column. K_THREADS=32\n// so at most 32 subgroups per column in the pathological size-1 case; size\n// 8*32 = 256 covers it. In practice (size 32) only [n_idx*32] is used.\nvar<workgroup> col_partials: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let tid = lid.x;\n  let n_idx = tid / K_THREADS;\n  let k_tid = tid % K_THREADS;\n  let col = wid.x * N_TILE + n_idx;\n  let K = params.K;\n\n  var sum: f32 = 0.0;\n  if (col < params.N) {\n    let b_off = col * K;\n    var k = k_tid * 4u;\n    while (k + 3u < K) {\n      let a_v = vec4f(A[k], A[k+1u], A[k+2u], A[k+3u]);\n      let b_v = vec4f(B[b_off+k], B[b_off+k+1u], B[b_off+k+2u], B[b_off+k+3u]);\n      sum += dot(a_v, b_v);\n      k += K_THREADS * 4u;\n    }\n  }\n\n  // Reduce within the subgroup. Each lane gets the sum over its subgroup.\n  let sg_sum = subgroupAdd(sum);\n\n  // Subgroup leaders deposit their partial into a per-column slot. Slot index is\n  // the leader's k_tid (its offset within the column), guaranteeing distinct\n  // slots per column even when a column spans multiple subgroups.\n  col_partials[tid] = 0.0;\n  workgroupBarrier();\n  if (subgroupElect()) {\n    col_partials[n_idx * K_THREADS + k_tid] = sg_sum;\n  }\n  workgroupBarrier();\n\n  // Column leader sums the (few) deposited subgroup partials for its column.\n  if (k_tid == 0u && col < params.N) {\n    let base = n_idx * K_THREADS;\n    var total: f32 = 0.0;\n    for (var i: u32 = 0u; i < K_THREADS; i++) {\n      total += col_partials[base + i];\n    }\n    C[col] = total;\n  }\n}\n`;\n\n// ── Subgroups variant of WGSL_MATVEC_INT4 (desktop-only fast path) ──\n//\n// Same approach as WGSL_MATVEC_SUBGROUPS applied to the INT4 dequant matvec:\n// subgroupAdd() replaces the shared-memory tree reduction. Layout is widened to\n// 256 threads = 8 columns × 32 K-threads (vs the portable kernel's 128/16) so\n// each column maps to exactly one 32-wide subgroup on desktop GPUs. N_TILE stays\n// 8 so getDispatchSize is unchanged. Selected only on non-WebKit + subgroups.\nconst WGSL_MATVEC_INT4_SUBGROUPS = `\\\nenable subgroups;\n\nconst N_TILE: u32 = 8u;\nconst K_THREADS: u32 = 32u;\n\nstruct Params {\n  K: u32,\n  N: u32,\n  group_size: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> A: array<vec4f>;\n@group(0) @binding(1) var<storage, read> B_q: array<vec4u>;\n@group(0) @binding(2) var<storage, read> scales: array<f32>;\n@group(0) @binding(3) var<storage, read> zeros: array<f32>;\n@group(0) @binding(4) var<storage, read_write> C: array<f32>;\n@group(0) @binding(5) var<storage, read> params: Params;\n\nvar<workgroup> col_partials: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let tid = lid.x;\n  let n_idx = tid / K_THREADS;\n  let k_tid = tid % K_THREADS;\n  let col = wid.x * N_TILE + n_idx;\n  let K = params.K;\n  let k_vec = K / 32u;\n  let groups_per_col = K / params.group_size;\n  let packed_per_group = params.group_size / 8u;\n\n  var sum: f32 = 0.0;\n  if (col < params.N) {\n    let b_off_v = col * k_vec;\n    let col_g_base = col * groups_per_col;\n\n    var v = k_tid;\n    while (v < k_vec) {\n      let bq = B_q[b_off_v + v];\n      let g = (v * 4u) / packed_per_group;\n      let scale = scales[col_g_base + g];\n      let zero = zeros[col_g_base + g];\n      let a_base = v * 8u;\n\n      for (var j: u32 = 0u; j < 4u; j++) {\n        let packed = bq[j];\n        let n0 = f32(packed & 0xFu);\n        let n1 = f32((packed >> 4u) & 0xFu);\n        let n2 = f32((packed >> 8u) & 0xFu);\n        let n3 = f32((packed >> 12u) & 0xFu);\n        let n4 = f32((packed >> 16u) & 0xFu);\n        let n5 = f32((packed >> 20u) & 0xFu);\n        let n6 = f32((packed >> 24u) & 0xFu);\n        let n7 = f32((packed >> 28u) & 0xFu);\n\n        let b0 = vec4f(n0 - zero, n1 - zero, n2 - zero, n3 - zero) * scale;\n        let b1 = vec4f(n4 - zero, n5 - zero, n6 - zero, n7 - zero) * scale;\n\n        sum += dot(A[a_base + j * 2u], b0) + dot(A[a_base + j * 2u + 1u], b1);\n      }\n\n      v += K_THREADS;\n    }\n  }\n\n  let sg_sum = subgroupAdd(sum);\n\n  col_partials[tid] = 0.0;\n  workgroupBarrier();\n  if (subgroupElect()) {\n    col_partials[n_idx * K_THREADS + k_tid] = sg_sum;\n  }\n  workgroupBarrier();\n\n  if (k_tid == 0u && col < params.N) {\n    let base = n_idx * K_THREADS;\n    var total: f32 = 0.0;\n    for (var i: u32 = 0u; i < K_THREADS; i++) {\n      total += col_partials[base + i];\n    }\n    C[col] = total;\n  }\n}\n`;\n\nconst WGSL_MATVEC_INT4 = `\\\n// K-parallel INT4 matrix-vector multiply: C[N] = A[K] * dequant(B_q[N,K])\n// Optimized for M=1 decode on Apple Silicon (SIMD width 32).\n//\n// CONSTRAINT: N_TILE must equal workgroup_size / K_THREADS\n// Design: 256 threads = 8 output columns × 32 K-threads per column.\n\nconst N_TILE: u32 = 16u;\nconst K_THREADS: u32 = 16u;\n\nstruct Params {\n  K: u32,\n  N: u32,\n  group_size: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> A: array<vec4f>;\n@group(0) @binding(1) var<storage, read> B_q: array<vec4u>;\n@group(0) @binding(2) var<storage, read> scales: array<f32>;\n@group(0) @binding(3) var<storage, read> zeros: array<f32>;\n@group(0) @binding(4) var<storage, read_write> C: array<f32>;\n@group(0) @binding(5) var<storage, read> params: Params;\n\nvar<workgroup> shared_sums: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let tid = lid.x;\n  let n_idx = tid / K_THREADS;\n  let k_tid = tid % K_THREADS;\n  let col = wid.x * N_TILE + n_idx;\n  let K = params.K;\n  let k_vec = K / 32u;\n  let groups_per_col = K / params.group_size;\n  let packed_per_group = params.group_size / 8u;\n\n  // K-parallel accumulation with vec4 loads: each thread reads one vec4u\n  // (4 packed u32s = 32 INT4 weights) per iteration. group_size (128) spans\n  // 16 packed u32s, so a vec4u never straddles a quantization group — one\n  // scale/zero load per iteration instead of four.\n  // A reads hit L1 cache (same addresses across all 8 output columns in workgroup)\n  var sum: f32 = 0.0;\n  if (col < params.N) {\n    let b_off_v = col * k_vec;\n    let col_g_base = col * groups_per_col;\n\n    var v = k_tid;\n    while (v < k_vec) {\n      let bq = B_q[b_off_v + v];\n      let g = (v * 4u) / packed_per_group;\n      let scale = scales[col_g_base + g];\n      let zero = zeros[col_g_base + g];\n      let a_base = v * 8u;\n\n      for (var j: u32 = 0u; j < 4u; j++) {\n        let packed = bq[j];\n        // Manual nibble extraction (avoids unpack4xU8 — broken on Safari/WebKit)\n        let n0 = f32(packed & 0xFu);\n        let n1 = f32((packed >> 4u) & 0xFu);\n        let n2 = f32((packed >> 8u) & 0xFu);\n        let n3 = f32((packed >> 12u) & 0xFu);\n        let n4 = f32((packed >> 16u) & 0xFu);\n        let n5 = f32((packed >> 20u) & 0xFu);\n        let n6 = f32((packed >> 24u) & 0xFu);\n        let n7 = f32((packed >> 28u) & 0xFu);\n\n        let b0 = vec4f(n0 - zero, n1 - zero, n2 - zero, n3 - zero) * scale;\n        let b1 = vec4f(n4 - zero, n5 - zero, n6 - zero, n7 - zero) * scale;\n\n        sum += dot(A[a_base + j * 2u], b0) + dot(A[a_base + j * 2u + 1u], b1);\n      }\n\n      v += K_THREADS;\n    }\n  }\n\n  // Store partials and reduce\n  shared_sums[tid] = sum;\n  workgroupBarrier();\n\n  if (k_tid == 0u && col < params.N) {\n    let base = n_idx * K_THREADS;\n    var total: f32 = shared_sums[base];\n    for (var i: u32 = 1u; i < K_THREADS; i++) {\n      total += shared_sums[base + i];\n    }\n    C[col] = total;\n  }\n}\n`;\n\n// ── Gated MatVecInt4 (fused attn-gate + INT4 o_proj, M=1 decode) ──\n//\n// out[n] = sum_k (attn[k] * sigmoid(gate[k])) * dequant(W[n,k])\n// Folds the SigmoidGate (attn_out * sigmoid(gate)) into the o_proj input so the\n// gate elementwise + o_proj run in ONE dispatch. Same INT4 dequant and\n// K-parallel reduction as WGSL_MATVEC_INT4 — only the A vector is built from two\n// inputs with an inline sigmoid. Numerically identical to SigmoidGate→MatVecInt4.\nconst WGSL_GATED_MATVEC_INT4 = `\\\nconst N_TILE: u32 = 8u;\nconst K_THREADS: u32 = 16u;\n\nstruct Params {\n  K: u32,\n  N: u32,\n  group_size: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> attn: array<vec4f>;\n@group(0) @binding(1) var<storage, read> gate: array<vec4f>;\n@group(0) @binding(2) var<storage, read> B_q: array<vec4u>;\n@group(0) @binding(3) var<storage, read> scales: array<f32>;\n@group(0) @binding(4) var<storage, read> zeros: array<f32>;\n@group(0) @binding(5) var<storage, read_write> C: array<f32>;\n@group(0) @binding(6) var<storage, read> params: Params;\n\nvar<workgroup> shared_sums: array<f32, 128>;\n\nfn gated(i: u32) -> vec4f {\n  let a = attn[i];\n  let g = gate[i];\n  let s = vec4f(1.0) / (vec4f(1.0) + exp(max(-g, vec4f(-80.0))));\n  return a * s;\n}\n\n@compute @workgroup_size(128)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let tid = lid.x;\n  let n_idx = tid / K_THREADS;\n  let k_tid = tid % K_THREADS;\n  let col = wid.x * N_TILE + n_idx;\n  let K = params.K;\n  let k_vec = K / 32u;\n  let groups_per_col = K / params.group_size;\n  let packed_per_group = params.group_size / 8u;\n\n  var sum: f32 = 0.0;\n  if (col < params.N) {\n    let b_off_v = col * k_vec;\n    let col_g_base = col * groups_per_col;\n\n    var v = k_tid;\n    while (v < k_vec) {\n      let bq = B_q[b_off_v + v];\n      let g = (v * 4u) / packed_per_group;\n      let scale = scales[col_g_base + g];\n      let zero = zeros[col_g_base + g];\n      let a_base = v * 8u;\n\n      for (var j: u32 = 0u; j < 4u; j++) {\n        let packed = bq[j];\n        let n0 = f32(packed & 0xFu);\n        let n1 = f32((packed >> 4u) & 0xFu);\n        let n2 = f32((packed >> 8u) & 0xFu);\n        let n3 = f32((packed >> 12u) & 0xFu);\n        let n4 = f32((packed >> 16u) & 0xFu);\n        let n5 = f32((packed >> 20u) & 0xFu);\n        let n6 = f32((packed >> 24u) & 0xFu);\n        let n7 = f32((packed >> 28u) & 0xFu);\n\n        let b0 = vec4f(n0 - zero, n1 - zero, n2 - zero, n3 - zero) * scale;\n        let b1 = vec4f(n4 - zero, n5 - zero, n6 - zero, n7 - zero) * scale;\n\n        sum += dot(gated(a_base + j * 2u), b0) + dot(gated(a_base + j * 2u + 1u), b1);\n      }\n\n      v += K_THREADS;\n    }\n  }\n\n  shared_sums[tid] = sum;\n  workgroupBarrier();\n\n  if (k_tid == 0u && col < params.N) {\n    let base = n_idx * K_THREADS;\n    var total: f32 = shared_sums[base];\n    for (var i: u32 = 1u; i < K_THREADS; i++) {\n      total += shared_sums[base + i];\n    }\n    C[col] = total;\n  }\n}\n`;\n\n// ── SwiGLU-gated MatVecInt4 (fused SwiGLU + INT4 projection, M=1 decode) ──\n//\n// out[n] = sum_k (silu(gate[k]) * up[k]) * dequant(W[n,k])\n// Folds a SwiGLU (silu(gate) * up) into the INT4 projection that consumes it —\n// used for the Mamba block's mamba_swiglu (silu(z) * norm_out) feeding out_proj.\n// Same INT4 dequant/reduction as WGSL_MATVEC_INT4; only the A vector is built\n// from two inputs with an inline SiLU. Identical to SwiGLU→MatVecInt4.\nconst WGSL_SWIGLU_GATED_MATVEC_INT4 = `\\\nconst N_TILE: u32 = 8u;\nconst K_THREADS: u32 = 16u;\n\nstruct Params {\n  K: u32,\n  N: u32,\n  group_size: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> gate: array<vec4f>;\n@group(0) @binding(1) var<storage, read> up: array<vec4f>;\n@group(0) @binding(2) var<storage, read> B_q: array<vec4u>;\n@group(0) @binding(3) var<storage, read> scales: array<f32>;\n@group(0) @binding(4) var<storage, read> zeros: array<f32>;\n@group(0) @binding(5) var<storage, read_write> C: array<f32>;\n@group(0) @binding(6) var<storage, read> params: Params;\n\nvar<workgroup> shared_sums: array<f32, 128>;\n\nfn swiglu_a(i: u32) -> vec4f {\n  let g = gate[i];\n  let u = up[i];\n  let s = g / (vec4f(1.0) + exp(max(-g, vec4f(-80.0))));\n  return s * u;\n}\n\n@compute @workgroup_size(128)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let tid = lid.x;\n  let n_idx = tid / K_THREADS;\n  let k_tid = tid % K_THREADS;\n  let col = wid.x * N_TILE + n_idx;\n  let K = params.K;\n  let k_vec = K / 32u;\n  let groups_per_col = K / params.group_size;\n  let packed_per_group = params.group_size / 8u;\n\n  var sum: f32 = 0.0;\n  if (col < params.N) {\n    let b_off_v = col * k_vec;\n    let col_g_base = col * groups_per_col;\n\n    var v = k_tid;\n    while (v < k_vec) {\n      let bq = B_q[b_off_v + v];\n      let g = (v * 4u) / packed_per_group;\n      let scale = scales[col_g_base + g];\n      let zero = zeros[col_g_base + g];\n      let a_base = v * 8u;\n\n      for (var j: u32 = 0u; j < 4u; j++) {\n        let packed = bq[j];\n        let n0 = f32(packed & 0xFu);\n        let n1 = f32((packed >> 4u) & 0xFu);\n        let n2 = f32((packed >> 8u) & 0xFu);\n        let n3 = f32((packed >> 12u) & 0xFu);\n        let n4 = f32((packed >> 16u) & 0xFu);\n        let n5 = f32((packed >> 20u) & 0xFu);\n        let n6 = f32((packed >> 24u) & 0xFu);\n        let n7 = f32((packed >> 28u) & 0xFu);\n\n        let b0 = vec4f(n0 - zero, n1 - zero, n2 - zero, n3 - zero) * scale;\n        let b1 = vec4f(n4 - zero, n5 - zero, n6 - zero, n7 - zero) * scale;\n\n        sum += dot(swiglu_a(a_base + j * 2u), b0) + dot(swiglu_a(a_base + j * 2u + 1u), b1);\n      }\n\n      v += K_THREADS;\n    }\n  }\n\n  shared_sums[tid] = sum;\n  workgroupBarrier();\n\n  if (k_tid == 0u && col < params.N) {\n    let base = n_idx * K_THREADS;\n    var total: f32 = shared_sums[base];\n    for (var i: u32 = 1u; i < K_THREADS; i++) {\n      total += shared_sums[base + i];\n    }\n    C[col] = total;\n  }\n}\n`;\n\nconst WGSL_ARGMAX = `\\\n// GPU-side argmax: finds the index of the maximum value.\n// Single workgroup, 256 threads, parallel reduction.\n\nstruct Params {\n  count: u32,\n  offset: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<vec4f>;\n@group(0) @binding(1) var<storage, read_write> result: array<u32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\nvar<workgroup> shared_max: array<f32, 256>;\nvar<workgroup> shared_idx: array<u32, 256>;\n\n@compute @workgroup_size(256)\nfn main(@builtin(local_invocation_id) lid: vec3u) {\n  let tid = lid.x;\n  // base/count are in elements; argmax always runs offset=0 with count a\n  // multiple of 4 (vocab_size), so we scan the logits as vec4f (128-bit loads,\n  // 4 logits per fetch) to better saturate memory bandwidth.\n  let count4 = params.count >> 2u;\n\n  // Each thread scans its chunk of vec4 groups\n  var local_max: f32 = -3.402823e+38;\n  var local_idx: u32 = 0u;\n\n  var i = tid;\n  while (i < count4) {\n    let v = input[i];\n    let g = i << 2u;\n    if (v.x > local_max) { local_max = v.x; local_idx = g; }\n    if (v.y > local_max) { local_max = v.y; local_idx = g + 1u; }\n    if (v.z > local_max) { local_max = v.z; local_idx = g + 2u; }\n    if (v.w > local_max) { local_max = v.w; local_idx = g + 3u; }\n    i += 256u;\n  }\n\n  shared_max[tid] = local_max;\n  shared_idx[tid] = local_idx;\n  workgroupBarrier();\n\n  // Tree reduction\n  if (tid < 128u) { if (shared_max[tid + 128u] > shared_max[tid]) { shared_max[tid] = shared_max[tid + 128u]; shared_idx[tid] = shared_idx[tid + 128u]; } }\n  workgroupBarrier();\n  if (tid < 64u) { if (shared_max[tid + 64u] > shared_max[tid]) { shared_max[tid] = shared_max[tid + 64u]; shared_idx[tid] = shared_idx[tid + 64u]; } }\n  workgroupBarrier();\n  if (tid < 32u) { if (shared_max[tid + 32u] > shared_max[tid]) { shared_max[tid] = shared_max[tid + 32u]; shared_idx[tid] = shared_idx[tid + 32u]; } }\n  workgroupBarrier();\n  if (tid < 16u) { if (shared_max[tid + 16u] > shared_max[tid]) { shared_max[tid] = shared_max[tid + 16u]; shared_idx[tid] = shared_idx[tid + 16u]; } }\n  workgroupBarrier();\n  if (tid < 8u) { if (shared_max[tid + 8u] > shared_max[tid]) { shared_max[tid] = shared_max[tid + 8u]; shared_idx[tid] = shared_idx[tid + 8u]; } }\n  workgroupBarrier();\n  if (tid < 4u) { if (shared_max[tid + 4u] > shared_max[tid]) { shared_max[tid] = shared_max[tid + 4u]; shared_idx[tid] = shared_idx[tid + 4u]; } }\n  workgroupBarrier();\n  if (tid < 2u) { if (shared_max[tid + 2u] > shared_max[tid]) { shared_max[tid] = shared_max[tid + 2u]; shared_idx[tid] = shared_idx[tid + 2u]; } }\n  workgroupBarrier();\n  if (tid == 0u) {\n    if (shared_max[1u] > shared_max[0u]) {\n      result[0] = shared_idx[1u];\n    } else {\n      result[0] = shared_idx[0u];\n    }\n  }\n}\n`;\n\nconst WGSL_RMSNORM = `\\\n// RMS Normalization: output = (x / rms(x)) * weight\n\nstruct Params {\n  seq_len: u32,\n  hidden_size: u32,\n  eps_bits: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read> weight: array<f32>;\n@group(0) @binding(2) var<storage, read_write> output: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\nvar<workgroup> shared_sum: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let row = wid.x;\n  if (row >= params.seq_len) { return; }\n\n  let tid = lid.x;\n  let row_offset = row * params.hidden_size;\n  let eps = bitcast<f32>(params.eps_bits);\n\n  var local_sum: f32 = 0.0;\n  var i = tid;\n  while (i < params.hidden_size) {\n    let val = input[row_offset + i];\n    local_sum += val * val;\n    i += 256u;\n  }\n  shared_sum[tid] = local_sum;\n  workgroupBarrier();\n\n  var stride: u32 = 128u;\n  while (stride > 0u) {\n    if (tid < stride) {\n      shared_sum[tid] += shared_sum[tid + stride];\n    }\n    workgroupBarrier();\n    stride = stride / 2u;\n  }\n\n  let rms = sqrt(shared_sum[0] / f32(params.hidden_size) + eps);\n\n  i = tid;\n  while (i < params.hidden_size) {\n    let val = input[row_offset + i];\n    output[row_offset + i] = (val / rms) * weight[i];\n    i += 256u;\n  }\n}\n`;\n\nconst WGSL_DUAL_RMSNORM = `\\\n// Two independent per-row RMSNorms in one dispatch, sharing hidden_size + eps.\n// Workgroups [0, rows0) normalize input0→output0 with weight0; workgroups\n// [rows0, rows0+rows1) normalize input1→output1 with weight1. Used to fuse the\n// per-head q_norm and k_norm in full-attention decode (same head_dim, same eps;\n// q has num_heads rows, k has num_kv_heads rows). Numerically identical to two\n// separate RMSNorm dispatches.\n\nstruct Params {\n  rows0: u32,        // workgroups handling input0 (e.g. T*num_heads)\n  rows1: u32,        // workgroups handling input1 (e.g. T*num_kv_heads)\n  hidden_size: u32,  // shared per-row width (head_dim)\n  eps_bits: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input0: array<f32>;\n@group(0) @binding(1) var<storage, read> weight0: array<f32>;\n@group(0) @binding(2) var<storage, read> input1: array<f32>;\n@group(0) @binding(3) var<storage, read> weight1: array<f32>;\n@group(0) @binding(4) var<storage, read_write> output0: array<f32>;\n@group(0) @binding(5) var<storage, read_write> output1: array<f32>;\n@group(0) @binding(6) var<storage, read> params: Params;\n\nvar<workgroup> shared_sum: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let wg = wid.x;\n  if (wg >= params.rows0 + params.rows1) { return; }\n\n  let tid = lid.x;\n  let hs = params.hidden_size;\n  let eps = bitcast<f32>(params.eps_bits);\n\n  // Route this workgroup to input0 (first rows0) or input1 (remaining).\n  let is1 = wg >= params.rows0;\n  let row = select(wg, wg - params.rows0, is1);\n  let row_offset = row * hs;\n\n  var local_sum: f32 = 0.0;\n  var i = tid;\n  while (i < hs) {\n    var val: f32;\n    if (is1) { val = input1[row_offset + i]; } else { val = input0[row_offset + i]; }\n    local_sum += val * val;\n    i += 256u;\n  }\n  shared_sum[tid] = local_sum;\n  workgroupBarrier();\n\n  var stride: u32 = 128u;\n  while (stride > 0u) {\n    if (tid < stride) {\n      shared_sum[tid] += shared_sum[tid + stride];\n    }\n    workgroupBarrier();\n    stride = stride / 2u;\n  }\n\n  let rms = sqrt(shared_sum[0] / f32(hs) + eps);\n\n  i = tid;\n  while (i < hs) {\n    if (is1) {\n      output1[row_offset + i] = (input1[row_offset + i] / rms) * weight1[i];\n    } else {\n      output0[row_offset + i] = (input0[row_offset + i] / rms) * weight0[i];\n    }\n    i += 256u;\n  }\n}\n`;\n\nconst WGSL_LAYERNORM = `\\\n// Layer Normalization: output = ((x - mean) / sqrt(var + eps)) * weight + bias\n\nstruct Params {\n  seq_len: u32,\n  hidden_size: u32,\n  eps_bits: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read> weight: array<f32>;\n@group(0) @binding(2) var<storage, read> bias: array<f32>;\n@group(0) @binding(3) var<storage, read_write> output: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\nvar<workgroup> shared_sum: array<f32, 256>;\nvar<workgroup> shared_sq_sum: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let row = wid.x;\n  if (row >= params.seq_len) { return; }\n\n  let tid = lid.x;\n  let row_offset = row * params.hidden_size;\n  let eps = bitcast<f32>(params.eps_bits);\n\n  var local_sum: f32 = 0.0;\n  var i = tid;\n  while (i < params.hidden_size) {\n    local_sum += input[row_offset + i];\n    i += 256u;\n  }\n  shared_sum[tid] = local_sum;\n  workgroupBarrier();\n\n  var stride: u32 = 128u;\n  while (stride > 0u) {\n    if (tid < stride) {\n      shared_sum[tid] += shared_sum[tid + stride];\n    }\n    workgroupBarrier();\n    stride /= 2u;\n  }\n  let mean = shared_sum[0] / f32(params.hidden_size);\n\n  var local_sq: f32 = 0.0;\n  i = tid;\n  while (i < params.hidden_size) {\n    let diff = input[row_offset + i] - mean;\n    local_sq += diff * diff;\n    i += 256u;\n  }\n  shared_sq_sum[tid] = local_sq;\n  workgroupBarrier();\n\n  stride = 128u;\n  while (stride > 0u) {\n    if (tid < stride) {\n      shared_sq_sum[tid] += shared_sq_sum[tid + stride];\n    }\n    workgroupBarrier();\n    stride /= 2u;\n  }\n  let inv_std = 1.0 / sqrt(shared_sq_sum[0] / f32(params.hidden_size) + eps);\n\n  i = tid;\n  while (i < params.hidden_size) {\n    let val = (input[row_offset + i] - mean) * inv_std;\n    output[row_offset + i] = val * weight[i] + bias[i];\n    i += 256u;\n  }\n}\n`;\n\nconst WGSL_ROPE = `\\\n// Rotary Position Embeddings (RoPE) with partial rotation and rotate_half convention.\n// Uses HuggingFace \"rotate_half\" pairing: (x[i], x[i + rope_half]) NOT adjacent (x[2i], x[2i+1]).\n//\n// Three decoupled knobs (so we match both Qwen-style partial RoPE and Gemma-style\n// \"proportional\" RoPE exactly):\n//   - rope_half          : the rotate_half pairing offset AND the number of pairs we\n//                          iterate per head. For HF this is head_dim/2 (Gemma) or\n//                          rope_dim/2 (Qwen partial). Pairs dim i with i+rope_half.\n//   - rope_denom         : denominator in the inv_freq exponent base^(2i/rope_denom).\n//                          Gemma \"proportional\" uses head_dim; Qwen partial uses rope_dim.\n//   - rope_active_pairs  : pairs [0, rope_active_pairs) get a real frequency; pairs\n//                          >= rope_active_pairs have inv_freq=0 (identity, the \"nope\"\n//                          zero-padded tail of HF _compute_proportional_rope_parameters).\n//\n// Defaults (set in buildParams) reproduce the legacy single-rope_dim behavior:\n//   rope_half = rope_dim/2, rope_denom = rope_dim, rope_active_pairs = rope_dim/2.\n\nstruct Params {\n  seq_len: u32,\n  num_q_heads: u32,\n  num_kv_heads: u32,\n  head_dim: u32,\n  rope_base_bits: u32,\n  position_offset: u32,\n  rope_dim: u32,\n  rope_half: u32,\n  rope_denom: u32,\n  rope_active_pairs: u32,\n  _pad0: u32,\n  _pad1: u32,\n}\n\n@group(0) @binding(0) var<storage, read_write> q: array<f32>;\n@group(0) @binding(1) var<storage, read_write> k: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let rope_half = params.rope_half;\n  let total_q_pairs = params.seq_len * params.num_q_heads * rope_half;\n  let total_k_pairs = params.seq_len * params.num_kv_heads * rope_half;\n\n  let idx = gid.x;\n\n  let rope_base = bitcast<f32>(params.rope_base_bits);\n\n  // rotate_half convention:\n  //   For pair_idx i (0..rope_half-1):\n  //     x_lo = x[i],  x_hi = x[i + rope_half]\n  //     x_lo' = x_lo * cos - x_hi * sin\n  //     x_hi' = x_lo * sin + x_hi * cos\n  //   inv_freq[i] = 1 / base^(2i/rope_denom)  for i < rope_active_pairs, else 0.\n\n  if (idx < total_q_pairs) {\n    let pos = idx / (params.num_q_heads * rope_half);\n    let remainder = idx % (params.num_q_heads * rope_half);\n    let head = remainder / rope_half;\n    let pair_idx = remainder % rope_half;\n\n    let position = f32(pos + params.position_offset);\n    var cos_val = 1.0;\n    var sin_val = 0.0;\n    if (pair_idx < params.rope_active_pairs) {\n      let freq = position / pow(rope_base, f32(2u * pair_idx) / f32(params.rope_denom));\n      cos_val = cos(freq);\n      sin_val = sin(freq);\n    }\n\n    let q_stride = params.num_q_heads * params.head_dim;\n    let head_base = pos * q_stride + head * params.head_dim;\n    let lo = head_base + pair_idx;\n    let hi = head_base + pair_idx + rope_half;\n\n    let q_lo = q[lo];\n    let q_hi = q[hi];\n    q[lo] = q_lo * cos_val - q_hi * sin_val;\n    q[hi] = q_lo * sin_val + q_hi * cos_val;\n  }\n\n  if (idx < total_k_pairs) {\n    let pos = idx / (params.num_kv_heads * rope_half);\n    let remainder = idx % (params.num_kv_heads * rope_half);\n    let head = remainder / rope_half;\n    let pair_idx = remainder % rope_half;\n\n    let position = f32(pos + params.position_offset);\n    var cos_val = 1.0;\n    var sin_val = 0.0;\n    if (pair_idx < params.rope_active_pairs) {\n      let freq = position / pow(rope_base, f32(2u * pair_idx) / f32(params.rope_denom));\n      cos_val = cos(freq);\n      sin_val = sin(freq);\n    }\n\n    let k_stride = params.num_kv_heads * params.head_dim;\n    let head_base = pos * k_stride + head * params.head_dim;\n    let lo = head_base + pair_idx;\n    let hi = head_base + pair_idx + rope_half;\n\n    let k_lo = k[lo];\n    let k_hi = k[hi];\n    k[lo] = k_lo * cos_val - k_hi * sin_val;\n    k[hi] = k_lo * sin_val + k_hi * cos_val;\n  }\n}\n`;\n\nconst WGSL_ROPE_INTERLEAVED = `\\\n// Rotary Position Embeddings — INTERLEAVED (adjacent-pair) convention, used by\n// Moonshine. Unlike the HF-Llama \"rotate_half\" split kernel above (which pairs\n// dim i with dim i+rope_half), Moonshine pairs ADJACENT dims (2p, 2p+1) with a\n// single frequency inv_freq[p] = 1 / base^(2p/rope_denom). Verified against the HF\n// MoonshineRotaryEmbedding (cos[..., :half].repeat_interleave(2) + interleaved\n// rotate_half). Only the first 2*rope_half dims of each head are rotated.\n//\n// Shares the Params layout with WGSL_ROPE (same buildParams). rope_half is the\n// number of adjacent pairs; rope_denom/rope_active_pairs default to the legacy\n// values so Moonshine is byte-identical to before.\n\nstruct Params {\n  seq_len: u32,\n  num_q_heads: u32,\n  num_kv_heads: u32,\n  head_dim: u32,\n  rope_base_bits: u32,\n  position_offset: u32,\n  rope_dim: u32,\n  rope_half: u32,\n  rope_denom: u32,\n  rope_active_pairs: u32,\n  _pad0: u32,\n  _pad1: u32,\n}\n\n@group(0) @binding(0) var<storage, read_write> q: array<f32>;\n@group(0) @binding(1) var<storage, read_write> k: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let half_rope = params.rope_half;\n  let total_q_pairs = params.seq_len * params.num_q_heads * half_rope;\n  let total_k_pairs = params.seq_len * params.num_kv_heads * half_rope;\n  let idx = gid.x;\n  let rope_base = bitcast<f32>(params.rope_base_bits);\n\n  if (idx < total_q_pairs) {\n    let pos = idx / (params.num_q_heads * half_rope);\n    let remainder = idx % (params.num_q_heads * half_rope);\n    let head = remainder / half_rope;\n    let pair_idx = remainder % half_rope;\n\n    let position = f32(pos + params.position_offset);\n    var cos_val = 1.0;\n    var sin_val = 0.0;\n    if (pair_idx < params.rope_active_pairs) {\n      let freq = position / pow(rope_base, f32(2u * pair_idx) / f32(params.rope_denom));\n      cos_val = cos(freq);\n      sin_val = sin(freq);\n    }\n\n    let q_stride = params.num_q_heads * params.head_dim;\n    let head_base = pos * q_stride + head * params.head_dim;\n    // Adjacent-pair (interleaved): dims 2*pair_idx and 2*pair_idx+1.\n    let lo = head_base + 2u * pair_idx;\n    let hi = lo + 1u;\n\n    let q_lo = q[lo];\n    let q_hi = q[hi];\n    q[lo] = q_lo * cos_val - q_hi * sin_val;\n    q[hi] = q_hi * cos_val + q_lo * sin_val;\n  }\n\n  if (idx < total_k_pairs) {\n    let pos = idx / (params.num_kv_heads * half_rope);\n    let remainder = idx % (params.num_kv_heads * half_rope);\n    let head = remainder / half_rope;\n    let pair_idx = remainder % half_rope;\n\n    let position = f32(pos + params.position_offset);\n    var cos_val = 1.0;\n    var sin_val = 0.0;\n    if (pair_idx < params.rope_active_pairs) {\n      let freq = position / pow(rope_base, f32(2u * pair_idx) / f32(params.rope_denom));\n      cos_val = cos(freq);\n      sin_val = sin(freq);\n    }\n\n    let k_stride = params.num_kv_heads * params.head_dim;\n    let head_base = pos * k_stride + head * params.head_dim;\n    let lo = head_base + 2u * pair_idx;\n    let hi = lo + 1u;\n\n    let k_lo = k[lo];\n    let k_hi = k[hi];\n    k[lo] = k_lo * cos_val - k_hi * sin_val;\n    k[hi] = k_hi * cos_val + k_lo * sin_val;\n  }\n}\n`;\n\nconst WGSL_ATTENTION = `\\\n// Tiled online-softmax attention (Flash-Attention style).\n//\n// Processes KV cache in tiles of TILE_S=32 positions. Per tile:\n//   1. Cooperative K tile load into shared memory (256 threads)\n//   2. Parallel Q·K dot products (8 threads per position, reduce via smem)\n//   3. Online softmax: update running max, rescale, compute tile weights\n//   4. Cooperative V tile load into shared memory (reuse K buffer)\n//   5. Parallel V accumulation: each of 128 threads handles one output dim\n//\n// Uses online softmax to maintain running (max, sum_exp, output) across tiles.\n// No global scratch buffer needed — everything in shared memory + registers.\n//\n// Safari/Metal compatibility:\n//   - Uses if/else instead of select() (Safari has select() bugs)\n//   - Clamps exp() args to -80 (Metal can return NaN for exp(-1e30))\n//\n// Shared memory: 4096 f32 = 16384 bytes = exactly 16 KB (minimum WebGPU guarantee)\n\nconst TILE_S: u32 = 16u;\nconst WG: u32 = 256u;\n// 16 threads cooperate on each dot product (256 / 16 = 16)\nconst DOT_THREADS: u32 = 16u;\n\nstruct Params {\n  T: u32,\n  S: u32,\n  num_q_heads: u32,\n  num_kv_heads: u32,\n  head_dim: u32,\n  position_offset: u32,\n  is_causal: u32,\n  // Query-grid base offset. Lets a caller process a window of query positions\n  // [q_offset, q_offset + dispatched_workgroups_x) in a single dispatch without a\n  // workgroup-id offset (WebGPU has none). Default 0 → full grid from row 0, byte-\n  // identical to the un-windowed path. Used only by the WebKit vision encoder to\n  // split the O(N²) ViT attention into watchdog-safe chunks.\n  q_offset: u32,\n  // Softmax QK temperature, supplied by buildParams as an f32. Defaults to\n  // 1/sqrt(head_dim) (byte-identical to the legacy hardcoded scale), so every\n  // existing caller is unchanged. Gemma 4 sets it to 1.0 because the per-head\n  // QK-norm already absorbs the 1/sqrt(head_dim) factor (HF scaling = 1.0).\n  attn_scale: f32,\n}\n\n@group(0) @binding(0) var<storage, read> Q: array<f32>;\n@group(0) @binding(1) var<storage, read> K: array<f32>;\n@group(0) @binding(2) var<storage, read> V: array<f32>;\n@group(0) @binding(3) var<storage, read_write> output: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\n// Shared memory layout (all phases share a single 4096-element array):\n//   K phase:    smem[0..TILE_S*hd) = K tile data (16×256 = 4096 f32)\n//   Dot phase:  smem[0..WG) = partial dot products for reduction\n//               smem[0..TILE_S) reused for final scores after reduction\n//   Softmax:    smem[0..TILE_S) = scores → exp weights\n//   V phase:    smem[TILE_S..TILE_S+TILE_S*hd) = V tile data\n//               But TILE_S+TILE_S*hd = 16+4096 > 4096, so we must save scores\n//               to per-thread registers before V load, then use full smem for V.\n// Total: 4096 f32 = 16384 bytes = exactly 16 KB (minimum WebGPU guarantee)\nvar<workgroup> smem: array<f32, 4096>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let q_pos = wid.x + params.q_offset;\n  let q_head = wid.y;\n\n  if (q_pos >= params.T || q_head >= params.num_q_heads) { return; }\n\n  let tid = lid.x;\n  let heads_per_kv = params.num_q_heads / params.num_kv_heads;\n  let kv_head = q_head / heads_per_kv;\n  let scale = params.attn_scale;\n\n  let q_stride = params.num_q_heads * params.head_dim;\n  let kv_stride = params.num_kv_heads * params.head_dim;\n  let q_base = q_pos * q_stride + q_head * params.head_dim;\n  let kv_head_offset = kv_head * params.head_dim;\n  let causal_limit = q_pos + params.position_offset + 1u;\n  let hd = params.head_dim;\n  // Causal (text) attends to keys up to its own position; bidirectional (ViT)\n  // attends to all S keys. Default is_causal=1 keeps text decoding unchanged.\n  var S_eff = params.S;\n  if (params.is_causal != 0u) { S_eff = min(params.S, causal_limit); }\n\n  // Online softmax running state (per-thread for output dims)\n  var running_max: f32 = -1e30;\n  var running_sum: f32 = 0.0;\n  // Per-thread output accumulators. Thread tid handles output dim tid; a second\n  // accumulator covers dim tid+WG so head_dim up to 2*WG (512) is supported with\n  // WG=256 threads. For hd<=256, tid+WG>=hd is never written → byte-identical.\n  var out_acc: f32 = 0.0;\n  var out_acc2: f32 = 0.0;\n  let d2 = tid + WG;\n\n  // Tile size capped so a K/V tile (tile_cap * hd f32) fits the 4096-f32 smem.\n  // hd<=256 → tile_cap=16=TILE_S (unchanged). hd=512 → tile_cap=8.\n  let tile_cap = min(TILE_S, 4096u / hd);\n\n  // Process KV cache in tiles of tile_cap positions\n  let num_tiles = (S_eff + tile_cap - 1u) / tile_cap;\n\n  for (var tile: u32 = 0u; tile < num_tiles; tile += 1u) {\n    let tile_start = tile * tile_cap;\n    let tile_end = min(tile_start + tile_cap, S_eff);\n    let tile_len = tile_end - tile_start;\n\n    // ── Step 1: Cooperative K tile load ──\n    // Load tile_len × hd values from K cache into smem\n    let total_k_elems = tile_len * hd;\n    var load_idx = tid;\n    while (load_idx < total_k_elems) {\n      let s_local = load_idx / hd;\n      let d = load_idx % hd;\n      let s_global = tile_start + s_local;\n      let k_addr = s_global * kv_stride + kv_head_offset + d;\n      smem[s_local * hd + d] = K[k_addr];\n      load_idx += WG;\n    }\n    workgroupBarrier();\n\n    // ── Step 2: Parallel Q·K dot products ──\n    // 256 threads = TILE_S positions × DOT_THREADS per position\n    let pos_in_tile = tid / DOT_THREADS;\n    let dot_tid = tid % DOT_THREADS;\n\n    var my_partial: f32 = 0.0;\n    if (pos_in_tile < tile_len) {\n      let dims_per_thread = hd / DOT_THREADS;\n      let d_start = dot_tid * dims_per_thread;\n      let d_end = d_start + dims_per_thread;\n      let kv_row_base = pos_in_tile * hd;\n\n      var d = d_start;\n      while (d + 3u < d_end) {\n        let q_v = vec4f(Q[q_base + d], Q[q_base + d + 1u],\n                         Q[q_base + d + 2u], Q[q_base + d + 3u]);\n        let k_v = vec4f(smem[kv_row_base + d], smem[kv_row_base + d + 1u],\n                         smem[kv_row_base + d + 2u], smem[kv_row_base + d + 3u]);\n        my_partial += dot(q_v, k_v);\n        d += 4u;\n      }\n      while (d < d_end) {\n        my_partial += Q[q_base + d] * smem[kv_row_base + d];\n        d += 1u;\n      }\n    }\n\n    // K tile no longer needed — reuse smem for dot reduction\n    workgroupBarrier();\n    smem[tid] = my_partial;\n    workgroupBarrier();\n\n    // Thread 0 of each dot group reduces and writes score to smem[0..TILE_S).\n    // Two-phase: leader 0 reads smem[1..15], which are exactly the slots\n    // leaders 1..15 write — reduce into a local first, barrier (in uniform\n    // control flow), then write. Read+write in one block is a data race.\n    var group_total: f32 = 0.0;\n    if (dot_tid == 0u && pos_in_tile < tile_len) {\n      group_total = smem[tid];\n      for (var i: u32 = 1u; i < DOT_THREADS; i += 1u) {\n        group_total += smem[tid + i];\n      }\n    }\n    workgroupBarrier();\n    if (dot_tid == 0u && pos_in_tile < tile_len) {\n      smem[pos_in_tile] = group_total * scale;\n    }\n    workgroupBarrier();\n\n    // ── Step 3: Online softmax ──\n    // Scores are in smem[0..tile_len). Find max via tree reduce in smem[TILE_S..].\n    // NOTE: Uses if/else instead of select() — Safari/Metal has bugs with select().\n    if (tid < TILE_S) {\n      if (tid < tile_len) {\n        smem[TILE_S + tid] = smem[tid];\n      } else {\n        smem[TILE_S + tid] = -1e30;\n      }\n    }\n    workgroupBarrier();\n    if (tid < 8u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 8u]); }\n    workgroupBarrier();\n    if (tid < 4u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 4u]); }\n    workgroupBarrier();\n    if (tid < 2u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 2u]); }\n    workgroupBarrier();\n    if (tid < 1u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 1u]); }\n    workgroupBarrier();\n    let tile_max = smem[TILE_S];\n\n    let new_max = max(running_max, tile_max);\n    // Clamp exp() argument to avoid NaN on Metal (exp(-87) ≈ 0 in f32)\n    let old_correction = exp(max(running_max - new_max, -80.0));\n\n    // Compute exp weights in smem[0..TILE_S)\n    if (tid < tile_len) {\n      smem[tid] = exp(max(smem[tid] - new_max, -80.0));\n    }\n    workgroupBarrier();\n\n    // Sum tile weights via tree reduce in smem[TILE_S..]\n    // NOTE: Uses if/else instead of select() — Safari/Metal has bugs with select().\n    if (tid < TILE_S) {\n      if (tid < tile_len) {\n        smem[TILE_S + tid] = smem[tid];\n      } else {\n        smem[TILE_S + tid] = 0.0;\n      }\n    }\n    workgroupBarrier();\n    if (tid < 8u) { smem[TILE_S + tid] += smem[TILE_S + tid + 8u]; }\n    workgroupBarrier();\n    if (tid < 4u) { smem[TILE_S + tid] += smem[TILE_S + tid + 4u]; }\n    workgroupBarrier();\n    if (tid < 2u) { smem[TILE_S + tid] += smem[TILE_S + tid + 2u]; }\n    workgroupBarrier();\n    if (tid < 1u) { smem[TILE_S + tid] += smem[TILE_S + tid + 1u]; }\n    workgroupBarrier();\n    let tile_sum = smem[TILE_S];\n\n    // ── Save scores to per-thread registers before V overwrites smem ──\n    // Each thread saves the scores it needs (all TILE_S values)\n    var w0: f32 = 0.0; var w1: f32 = 0.0; var w2: f32 = 0.0; var w3: f32 = 0.0;\n    var w4: f32 = 0.0; var w5: f32 = 0.0; var w6: f32 = 0.0; var w7: f32 = 0.0;\n    var w8: f32 = 0.0; var w9: f32 = 0.0; var w10: f32 = 0.0; var w11: f32 = 0.0;\n    var w12: f32 = 0.0; var w13: f32 = 0.0; var w14: f32 = 0.0; var w15: f32 = 0.0;\n    if (tid < hd) {\n      if (0u < tile_len) { w0 = smem[0]; }\n      if (1u < tile_len) { w1 = smem[1]; }\n      if (2u < tile_len) { w2 = smem[2]; }\n      if (3u < tile_len) { w3 = smem[3]; }\n      if (4u < tile_len) { w4 = smem[4]; }\n      if (5u < tile_len) { w5 = smem[5]; }\n      if (6u < tile_len) { w6 = smem[6]; }\n      if (7u < tile_len) { w7 = smem[7]; }\n      if (8u < tile_len) { w8 = smem[8]; }\n      if (9u < tile_len) { w9 = smem[9]; }\n      if (10u < tile_len) { w10 = smem[10]; }\n      if (11u < tile_len) { w11 = smem[11]; }\n      if (12u < tile_len) { w12 = smem[12]; }\n      if (13u < tile_len) { w13 = smem[13]; }\n      if (14u < tile_len) { w14 = smem[14]; }\n      if (15u < tile_len) { w15 = smem[15]; }\n    }\n    workgroupBarrier();\n\n    // ── Step 4: Cooperative V tile load (full smem reuse) ──\n    let total_v_elems = tile_len * hd;\n    load_idx = tid;\n    while (load_idx < total_v_elems) {\n      let s_local = load_idx / hd;\n      let d = load_idx % hd;\n      let s_global = tile_start + s_local;\n      let v_addr = s_global * kv_stride + kv_head_offset + d;\n      smem[s_local * hd + d] = V[v_addr];\n      load_idx += WG;\n    }\n    workgroupBarrier();\n\n    // ── Step 5: Parallel V accumulation ──\n    // Thread tid handles output dim tid and (for hd>WG) dim tid+WG via out_acc2.\n    // Uses saved score registers instead of smem[0..TILE_S).\n    if (tid < hd) {\n      out_acc = out_acc * old_correction;\n      if (0u < tile_len) { out_acc += w0 * smem[0u * hd + tid]; }\n      if (1u < tile_len) { out_acc += w1 * smem[1u * hd + tid]; }\n      if (2u < tile_len) { out_acc += w2 * smem[2u * hd + tid]; }\n      if (3u < tile_len) { out_acc += w3 * smem[3u * hd + tid]; }\n      if (4u < tile_len) { out_acc += w4 * smem[4u * hd + tid]; }\n      if (5u < tile_len) { out_acc += w5 * smem[5u * hd + tid]; }\n      if (6u < tile_len) { out_acc += w6 * smem[6u * hd + tid]; }\n      if (7u < tile_len) { out_acc += w7 * smem[7u * hd + tid]; }\n      if (8u < tile_len) { out_acc += w8 * smem[8u * hd + tid]; }\n      if (9u < tile_len) { out_acc += w9 * smem[9u * hd + tid]; }\n      if (10u < tile_len) { out_acc += w10 * smem[10u * hd + tid]; }\n      if (11u < tile_len) { out_acc += w11 * smem[11u * hd + tid]; }\n      if (12u < tile_len) { out_acc += w12 * smem[12u * hd + tid]; }\n      if (13u < tile_len) { out_acc += w13 * smem[13u * hd + tid]; }\n      if (14u < tile_len) { out_acc += w14 * smem[14u * hd + tid]; }\n      if (15u < tile_len) { out_acc += w15 * smem[15u * hd + tid]; }\n    }\n    if (d2 < hd) {\n      out_acc2 = out_acc2 * old_correction;\n      if (0u < tile_len) { out_acc2 += w0 * smem[0u * hd + d2]; }\n      if (1u < tile_len) { out_acc2 += w1 * smem[1u * hd + d2]; }\n      if (2u < tile_len) { out_acc2 += w2 * smem[2u * hd + d2]; }\n      if (3u < tile_len) { out_acc2 += w3 * smem[3u * hd + d2]; }\n      if (4u < tile_len) { out_acc2 += w4 * smem[4u * hd + d2]; }\n      if (5u < tile_len) { out_acc2 += w5 * smem[5u * hd + d2]; }\n      if (6u < tile_len) { out_acc2 += w6 * smem[6u * hd + d2]; }\n      if (7u < tile_len) { out_acc2 += w7 * smem[7u * hd + d2]; }\n      if (8u < tile_len) { out_acc2 += w8 * smem[8u * hd + d2]; }\n      if (9u < tile_len) { out_acc2 += w9 * smem[9u * hd + d2]; }\n      if (10u < tile_len) { out_acc2 += w10 * smem[10u * hd + d2]; }\n      if (11u < tile_len) { out_acc2 += w11 * smem[11u * hd + d2]; }\n      if (12u < tile_len) { out_acc2 += w12 * smem[12u * hd + d2]; }\n      if (13u < tile_len) { out_acc2 += w13 * smem[13u * hd + d2]; }\n      if (14u < tile_len) { out_acc2 += w14 * smem[14u * hd + d2]; }\n      if (15u < tile_len) { out_acc2 += w15 * smem[15u * hd + d2]; }\n    }\n\n    // Update running state\n    running_sum = running_sum * old_correction + tile_sum;\n    running_max = new_max;\n    workgroupBarrier();\n  }\n\n  // ── Write normalized output ──\n  var inv_sum: f32 = 0.0;\n  if (running_sum > 0.0) { inv_sum = 1.0 / running_sum; }\n  let out_base = q_pos * q_stride + q_head * hd;\n  if (tid < hd) {\n    output[out_base + tid] = out_acc * inv_sum;\n  }\n  if (d2 < hd) {\n    output[out_base + d2] = out_acc2 * inv_sum;\n  }\n}\n`;\n\n// ── Cross-Attention (encoder-decoder) ──────────────────────────────────\n//\n// Decoder queries attend to a FIXED encoder output. This is the distinct\n// coherence regime from the causal self-attention KVCache path:\n//\n//   Q  = decoder hidden state    [T_dec, num_q_heads  * head_dim]\n//   K  = encoder output (K proj) [S_enc, num_kv_heads * head_dim]\n//   V  = encoder output (V proj) [S_enc, num_kv_heads * head_dim]\n//\n// The encoder K/V are computed ONCE and frozen for the whole decode — there\n// is no position_offset and no growing cache. Attention is UNMASKED /\n// bidirectional over the full S_enc encoder sequence (every decoder query\n// position sees every encoder position). One workgroup per (q_pos, q_head).\n//\n// Algorithm is the identical online-softmax tiling to WGSL_ATTENTION (the\n// validated flash-attention kernel) with the causal-limit logic removed, so\n// it inherits the same Safari/Metal-safe patterns:\n//   - if/else instead of select()\n//   - exp() args clamped to -80 (Metal returns NaN for exp(-1e30))\n//   - <= 16 KB workgroup memory (single 4096 f32 smem array)\n//   - no `enable f16`\nconst WGSL_CROSS_ATTENTION = `\\\nconst TILE_S: u32 = 16u;\nconst WG: u32 = 256u;\nconst DOT_THREADS: u32 = 16u;\n\nstruct Params {\n  T: u32,            // decoder query length\n  S: u32,            // encoder (key/value) length\n  num_q_heads: u32,\n  num_kv_heads: u32,\n  head_dim: u32,\n}\n\n@group(0) @binding(0) var<storage, read> Q: array<f32>;\n@group(0) @binding(1) var<storage, read> K: array<f32>;\n@group(0) @binding(2) var<storage, read> V: array<f32>;\n@group(0) @binding(3) var<storage, read_write> output: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\nvar<workgroup> smem: array<f32, 4096>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let q_pos = wid.x;\n  let q_head = wid.y;\n\n  if (q_pos >= params.T || q_head >= params.num_q_heads) { return; }\n\n  let tid = lid.x;\n  let heads_per_kv = params.num_q_heads / params.num_kv_heads;\n  let kv_head = q_head / heads_per_kv;\n  let scale = 1.0 / sqrt(f32(params.head_dim));\n\n  let q_stride = params.num_q_heads * params.head_dim;\n  let kv_stride = params.num_kv_heads * params.head_dim;\n  let q_base = q_pos * q_stride + q_head * params.head_dim;\n  let kv_head_offset = kv_head * params.head_dim;\n  let hd = params.head_dim;\n  // Unmasked: every query attends to all S_enc encoder positions.\n  let S_eff = params.S;\n\n  var running_max: f32 = -1e30;\n  var running_sum: f32 = 0.0;\n  // out_acc2 covers dim tid+WG for head_dim up to 2*WG; unused for hd<=WG.\n  var out_acc: f32 = 0.0;\n  var out_acc2: f32 = 0.0;\n  let d2 = tid + WG;\n\n  // Tile capped so a K/V tile (tile_cap*hd f32) fits the 4096-f32 smem.\n  let tile_cap = min(TILE_S, 4096u / hd);\n  let num_tiles = (S_eff + tile_cap - 1u) / tile_cap;\n\n  for (var tile: u32 = 0u; tile < num_tiles; tile += 1u) {\n    let tile_start = tile * tile_cap;\n    let tile_end = min(tile_start + tile_cap, S_eff);\n    let tile_len = tile_end - tile_start;\n\n    // ── Step 1: Cooperative K tile load ──\n    let total_k_elems = tile_len * hd;\n    var load_idx = tid;\n    while (load_idx < total_k_elems) {\n      let s_local = load_idx / hd;\n      let d = load_idx % hd;\n      let s_global = tile_start + s_local;\n      let k_addr = s_global * kv_stride + kv_head_offset + d;\n      smem[s_local * hd + d] = K[k_addr];\n      load_idx += WG;\n    }\n    workgroupBarrier();\n\n    // ── Step 2: Parallel Q·K dot products ──\n    let pos_in_tile = tid / DOT_THREADS;\n    let dot_tid = tid % DOT_THREADS;\n\n    var my_partial: f32 = 0.0;\n    if (pos_in_tile < tile_len) {\n      let kv_row_base = pos_in_tile * hd;\n      // Strided over head_dim so any head_dim works (incl. ones not divisible\n      // by DOT_THREADS, e.g. Moonshine's 52). Each thread sums its dims.\n      var d = dot_tid;\n      while (d < hd) {\n        my_partial += Q[q_base + d] * smem[kv_row_base + d];\n        d += DOT_THREADS;\n      }\n    }\n\n    workgroupBarrier();\n    smem[tid] = my_partial;\n    workgroupBarrier();\n\n    var group_total: f32 = 0.0;\n    if (dot_tid == 0u && pos_in_tile < tile_len) {\n      group_total = smem[tid];\n      for (var i: u32 = 1u; i < DOT_THREADS; i += 1u) {\n        group_total += smem[tid + i];\n      }\n    }\n    workgroupBarrier();\n    if (dot_tid == 0u && pos_in_tile < tile_len) {\n      smem[pos_in_tile] = group_total * scale;\n    }\n    workgroupBarrier();\n\n    // ── Step 3: Online softmax (max reduce, no select()) ──\n    if (tid < TILE_S) {\n      if (tid < tile_len) {\n        smem[TILE_S + tid] = smem[tid];\n      } else {\n        smem[TILE_S + tid] = -1e30;\n      }\n    }\n    workgroupBarrier();\n    if (tid < 8u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 8u]); }\n    workgroupBarrier();\n    if (tid < 4u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 4u]); }\n    workgroupBarrier();\n    if (tid < 2u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 2u]); }\n    workgroupBarrier();\n    if (tid < 1u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 1u]); }\n    workgroupBarrier();\n    let tile_max = smem[TILE_S];\n\n    let new_max = max(running_max, tile_max);\n    let old_correction = exp(max(running_max - new_max, -80.0));\n\n    if (tid < tile_len) {\n      smem[tid] = exp(max(smem[tid] - new_max, -80.0));\n    }\n    workgroupBarrier();\n\n    if (tid < TILE_S) {\n      if (tid < tile_len) {\n        smem[TILE_S + tid] = smem[tid];\n      } else {\n        smem[TILE_S + tid] = 0.0;\n      }\n    }\n    workgroupBarrier();\n    if (tid < 8u) { smem[TILE_S + tid] += smem[TILE_S + tid + 8u]; }\n    workgroupBarrier();\n    if (tid < 4u) { smem[TILE_S + tid] += smem[TILE_S + tid + 4u]; }\n    workgroupBarrier();\n    if (tid < 2u) { smem[TILE_S + tid] += smem[TILE_S + tid + 2u]; }\n    workgroupBarrier();\n    if (tid < 1u) { smem[TILE_S + tid] += smem[TILE_S + tid + 1u]; }\n    workgroupBarrier();\n    let tile_sum = smem[TILE_S];\n\n    // ── Save scores to per-thread registers before V overwrites smem ──\n    var w0: f32 = 0.0; var w1: f32 = 0.0; var w2: f32 = 0.0; var w3: f32 = 0.0;\n    var w4: f32 = 0.0; var w5: f32 = 0.0; var w6: f32 = 0.0; var w7: f32 = 0.0;\n    var w8: f32 = 0.0; var w9: f32 = 0.0; var w10: f32 = 0.0; var w11: f32 = 0.0;\n    var w12: f32 = 0.0; var w13: f32 = 0.0; var w14: f32 = 0.0; var w15: f32 = 0.0;\n    if (tid < hd) {\n      if (0u < tile_len) { w0 = smem[0]; }\n      if (1u < tile_len) { w1 = smem[1]; }\n      if (2u < tile_len) { w2 = smem[2]; }\n      if (3u < tile_len) { w3 = smem[3]; }\n      if (4u < tile_len) { w4 = smem[4]; }\n      if (5u < tile_len) { w5 = smem[5]; }\n      if (6u < tile_len) { w6 = smem[6]; }\n      if (7u < tile_len) { w7 = smem[7]; }\n      if (8u < tile_len) { w8 = smem[8]; }\n      if (9u < tile_len) { w9 = smem[9]; }\n      if (10u < tile_len) { w10 = smem[10]; }\n      if (11u < tile_len) { w11 = smem[11]; }\n      if (12u < tile_len) { w12 = smem[12]; }\n      if (13u < tile_len) { w13 = smem[13]; }\n      if (14u < tile_len) { w14 = smem[14]; }\n      if (15u < tile_len) { w15 = smem[15]; }\n    }\n    workgroupBarrier();\n\n    // ── Step 4: Cooperative V tile load ──\n    let total_v_elems = tile_len * hd;\n    load_idx = tid;\n    while (load_idx < total_v_elems) {\n      let s_local = load_idx / hd;\n      let d = load_idx % hd;\n      let s_global = tile_start + s_local;\n      let v_addr = s_global * kv_stride + kv_head_offset + d;\n      smem[s_local * hd + d] = V[v_addr];\n      load_idx += WG;\n    }\n    workgroupBarrier();\n\n    // ── Step 5: Parallel V accumulation ──\n    if (tid < hd) {\n      out_acc = out_acc * old_correction;\n      if (0u < tile_len) { out_acc += w0 * smem[0u * hd + tid]; }\n      if (1u < tile_len) { out_acc += w1 * smem[1u * hd + tid]; }\n      if (2u < tile_len) { out_acc += w2 * smem[2u * hd + tid]; }\n      if (3u < tile_len) { out_acc += w3 * smem[3u * hd + tid]; }\n      if (4u < tile_len) { out_acc += w4 * smem[4u * hd + tid]; }\n      if (5u < tile_len) { out_acc += w5 * smem[5u * hd + tid]; }\n      if (6u < tile_len) { out_acc += w6 * smem[6u * hd + tid]; }\n      if (7u < tile_len) { out_acc += w7 * smem[7u * hd + tid]; }\n      if (8u < tile_len) { out_acc += w8 * smem[8u * hd + tid]; }\n      if (9u < tile_len) { out_acc += w9 * smem[9u * hd + tid]; }\n      if (10u < tile_len) { out_acc += w10 * smem[10u * hd + tid]; }\n      if (11u < tile_len) { out_acc += w11 * smem[11u * hd + tid]; }\n      if (12u < tile_len) { out_acc += w12 * smem[12u * hd + tid]; }\n      if (13u < tile_len) { out_acc += w13 * smem[13u * hd + tid]; }\n      if (14u < tile_len) { out_acc += w14 * smem[14u * hd + tid]; }\n      if (15u < tile_len) { out_acc += w15 * smem[15u * hd + tid]; }\n    }\n    if (d2 < hd) {\n      out_acc2 = out_acc2 * old_correction;\n      if (0u < tile_len) { out_acc2 += w0 * smem[0u * hd + d2]; }\n      if (1u < tile_len) { out_acc2 += w1 * smem[1u * hd + d2]; }\n      if (2u < tile_len) { out_acc2 += w2 * smem[2u * hd + d2]; }\n      if (3u < tile_len) { out_acc2 += w3 * smem[3u * hd + d2]; }\n      if (4u < tile_len) { out_acc2 += w4 * smem[4u * hd + d2]; }\n      if (5u < tile_len) { out_acc2 += w5 * smem[5u * hd + d2]; }\n      if (6u < tile_len) { out_acc2 += w6 * smem[6u * hd + d2]; }\n      if (7u < tile_len) { out_acc2 += w7 * smem[7u * hd + d2]; }\n      if (8u < tile_len) { out_acc2 += w8 * smem[8u * hd + d2]; }\n      if (9u < tile_len) { out_acc2 += w9 * smem[9u * hd + d2]; }\n      if (10u < tile_len) { out_acc2 += w10 * smem[10u * hd + d2]; }\n      if (11u < tile_len) { out_acc2 += w11 * smem[11u * hd + d2]; }\n      if (12u < tile_len) { out_acc2 += w12 * smem[12u * hd + d2]; }\n      if (13u < tile_len) { out_acc2 += w13 * smem[13u * hd + d2]; }\n      if (14u < tile_len) { out_acc2 += w14 * smem[14u * hd + d2]; }\n      if (15u < tile_len) { out_acc2 += w15 * smem[15u * hd + d2]; }\n    }\n\n    running_sum = running_sum * old_correction + tile_sum;\n    running_max = new_max;\n    workgroupBarrier();\n  }\n\n  // ── Write normalized output ──\n  var inv_sum: f32 = 0.0;\n  if (running_sum > 0.0) { inv_sum = 1.0 / running_sum; }\n  let out_base = q_pos * q_stride + q_head * hd;\n  if (tid < hd) {\n    output[out_base + tid] = out_acc * inv_sum;\n  }\n  if (d2 < hd) {\n    output[out_base + d2] = out_acc2 * inv_sum;\n  }\n}\n`;\n\nconst WGSL_SOFTMAX = `\\\n// Row-wise softmax: output[i] = exp(input[i] - max) / sum(exp(input - max))\n\nstruct Params {\n  num_rows: u32,\n  row_size: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read_write> output: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\nvar<workgroup> shared_data: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let row = wid.x;\n  if (row >= params.num_rows) { return; }\n\n  let tid = lid.x;\n  let row_offset = row * params.row_size;\n\n  var local_max: f32 = -1e30;\n  var i = tid;\n  while (i < params.row_size) {\n    local_max = max(local_max, input[row_offset + i]);\n    i += 256u;\n  }\n  shared_data[tid] = local_max;\n  workgroupBarrier();\n\n  var stride: u32 = 128u;\n  while (stride > 0u) {\n    if (tid < stride) {\n      shared_data[tid] = max(shared_data[tid], shared_data[tid + stride]);\n    }\n    workgroupBarrier();\n    stride /= 2u;\n  }\n  let row_max = shared_data[0];\n  workgroupBarrier();\n\n  var local_sum: f32 = 0.0;\n  i = tid;\n  while (i < params.row_size) {\n    local_sum += exp(max(input[row_offset + i] - row_max, -80.0));\n    i += 256u;\n  }\n  shared_data[tid] = local_sum;\n  workgroupBarrier();\n\n  stride = 128u;\n  while (stride > 0u) {\n    if (tid < stride) {\n      shared_data[tid] += shared_data[tid + stride];\n    }\n    workgroupBarrier();\n    stride /= 2u;\n  }\n  let inv_sum = 1.0 / shared_data[0];\n  workgroupBarrier();\n\n  i = tid;\n  while (i < params.row_size) {\n    output[row_offset + i] = exp(max(input[row_offset + i] - row_max, -80.0)) * inv_sum;\n    i += 256u;\n  }\n}\n`;\n\nconst WGSL_SILU = `\\\n// SiLU (Sigmoid Linear Unit): output = x * sigmoid(x) = x / (1 + exp(-x))\n\nstruct Params {\n  count: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read_write> output: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n\n  let x = input[idx];\n  output[idx] = x / (1.0 + exp(max(-x, -80.0)));\n}\n`;\n\nconst WGSL_GELU = `\\\n// GELU (Gaussian Error Linear Unit) - approximate version\n// output = 0.5 * x * (1 + tanh(sqrt(2/pi) * (x + 0.044715 * x^3)))\n\nstruct Params {\n  count: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read_write> output: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\nconst SQRT_2_OVER_PI: f32 = 0.7978845608;\nconst GELU_COEFF: f32 = 0.044715;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n\n  let x = input[idx];\n  // Clamp the tanh argument: tanh saturates to ±1 by |arg|≈10, but on Metal/Dawn\n  // tanh() of a large argument returns NaN (the internal exp overflows). The ViT\n  // MLP can produce |x|>30 → x^3 → arg in the thousands, so clamp to a safe ±15.\n  let inner = clamp(SQRT_2_OVER_PI * (x + GELU_COEFF * x * x * x), -15.0, 15.0);\n  output[idx] = 0.5 * x * (1.0 + tanh(inner));\n}\n`;\n\nconst WGSL_ADD = `\\\n// Element-wise addition: output = a + b\n\nstruct Params {\n  count: u32,\n}\n\n@group(0) @binding(0) var<storage, read> a: array<f32>;\n@group(0) @binding(1) var<storage, read> b: array<f32>;\n@group(0) @binding(2) var<storage, read_write> output: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n  output[idx] = a[idx] + b[idx];\n}\n`;\n\nconst WGSL_MUL = `\\\n// Element-wise multiplication: output = a * b\n\nstruct Params {\n  count: u32,\n}\n\n@group(0) @binding(0) var<storage, read> a: array<f32>;\n@group(0) @binding(1) var<storage, read> b: array<f32>;\n@group(0) @binding(2) var<storage, read_write> output: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n  output[idx] = a[idx] * b[idx];\n}\n`;\n\nconst WGSL_SLICE_LAST_ROW = `\\\n// Copy the last row of a [T, width] tensor into a [1, width] tensor.\n// Lets lm_head run with M=1, so the logits buffer is [1, vocab] instead\n// of [T, vocab] (485MB at T=512 for a 248k vocab).\n\nstruct Params {\n  width: u32,\n  last_row_offset: u32,\n}\n\n@group(0) @binding(0) var<storage, read> src: array<f32>;\n@group(0) @binding(1) var<storage, read_write> output: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.width) { return; }\n  output[idx] = src[params.last_row_offset + idx];\n}\n`;\n\nconst WGSL_MEAN_POOL = `\\\n// Mean-pool a [T, width] tensor over the T tokens into a [1, width] tensor:\n//   output[c] = (1/T) * sum_{t=0..T-1} src[t * width + c]\n// Used for bidirectional encoder embeddings (EmbeddingGemma) where the sentence\n// vector is the average of all token hidden states. We process a single unpadded\n// sequence, so every one of the T rows is a valid (non-pad) token — the mean is\n// over the full T with no attention-mask gather needed.\n// One thread per output channel.\n\nstruct Params {\n  seq_len: u32,\n  width: u32,\n}\n\n@group(0) @binding(0) var<storage, read> src: array<f32>;\n@group(0) @binding(1) var<storage, read_write> output: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let c = gid.x;\n  if (c >= params.width) { return; }\n  var acc = 0.0;\n  var t = 0u;\n  loop {\n    if (t >= params.seq_len) { break; }\n    acc += src[t * params.width + c];\n    t += 1u;\n  }\n  let inv_t = 1.0 / max(f32(params.seq_len), 1.0);\n  output[c] = acc * inv_t;\n}\n`;\n\nconst WGSL_SCALE = `\\\n// Multiply every element of a tensor by a scalar constant:\n//   output[i] = input[i] * scale\n// Used for Gemma's embedding normalizer (× sqrt(hidden_size)), which is applied\n// to the residual stream after the token embedding lookup.\n\nstruct Params {\n  count: u32,\n  scale_bits: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read_write> output: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n  output[idx] = input[idx] * bitcast<f32>(params.scale_bits);\n}\n`;\n\n// ── Softcap (Gemma final logit softcapping) ──\n// output[i] = cap * tanh(input[i] / cap)\n// Squashes logits into (-cap, cap). Mobile-safe: tanh is a WGSL builtin, no\n// select(), no exp (tanh is bounded), no f16. The divide-by-cap argument to tanh\n// is finite for real logits, so no clamping is required.\nconst WGSL_SOFTCAP = `\\\nstruct Params {\n  count: u32,\n  cap_bits: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read_write> output: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n  let cap = bitcast<f32>(params.cap_bits);\n  output[idx] = cap * tanh(input[idx] / cap);\n}\n`;\n\n// (SliceCols WGSL + spec already defined above for the ViT qkv split; Gemma 4\n// PLE reuses it for the per-layer-input column slice.)\n\nconst WGSL_L2NORM = `\\\n// L2-normalize each row of a [rows, width] tensor.\n// embedding[r, :] = x[r, :] / max(sqrt(sum(x[r, :]^2)), eps)\n// One workgroup per row; threads cooperatively reduce the sum of squares.\n\nstruct Params {\n  rows: u32,\n  width: u32,\n}\n\n@group(0) @binding(0) var<storage, read> src: array<f32>;\n@group(0) @binding(1) var<storage, read_write> output: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\nvar<workgroup> partial: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(workgroup_id) wid: vec3u,\n  @builtin(local_invocation_id) lid: vec3u,\n) {\n  let row = wid.x;\n  if (row >= params.rows) { return; }\n  let tid = lid.x;\n  let base = row * params.width;\n\n  // Each thread accumulates a strided slice of the row.\n  var acc = 0.0;\n  var i = tid;\n  loop {\n    if (i >= params.width) { break; }\n    let v = src[base + i];\n    acc += v * v;\n    i += 256u;\n  }\n  partial[tid] = acc;\n  workgroupBarrier();\n\n  // Tree reduction over the 256 partials.\n  var stride = 128u;\n  loop {\n    if (stride == 0u) { break; }\n    if (tid < stride) {\n      partial[tid] += partial[tid + stride];\n    }\n    workgroupBarrier();\n    stride = stride >> 1u;\n  }\n\n  // inv_norm = 1 / max(sqrt(sumsq), eps)\n  let inv_norm = 1.0 / max(sqrt(partial[0]), 1e-12);\n  workgroupBarrier();\n\n  var j = tid;\n  loop {\n    if (j >= params.width) { break; }\n    output[base + j] = src[base + j] * inv_norm;\n    j += 256u;\n  }\n}\n`;\n\nconst WGSL_ADD_BIAS = `\\\n// Row-broadcast bias add: output[r * width + c] = input[r * width + c] + bias[c].\n// Used by the ViT linear layers (qkv/proj/fc1/fc2/patch_embed) whose MatMul has\n// no bias term — the bias vector (length = width) is added across all rows.\n\nstruct Params {\n  count: u32,\n  width: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read> bias: array<f32>;\n@group(0) @binding(2) var<storage, read_write> output: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n  let c = idx % params.width;\n  output[idx] = input[idx] + bias[c];\n}\n`;\n\nconst WGSL_GELU_ERF = `\\\n// Exact GELU (erf form): output = 0.5 * x * (1 + erf(x / sqrt(2))).\n// The ViT merger uses nn.GELU() (exact), unlike the block MLP which uses the\n// tanh approximation. WGSL has no erf(), so we use a high-accuracy rational\n// approximation (Abramowitz & Stegun 7.1.26), giving < 1e-6 abs error.\n\nstruct Params {\n  count: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read_write> output: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\nconst INV_SQRT2: f32 = 0.7071067811865476;\n\nfn erf_approx(x: f32) -> f32 {\n  // A&S 7.1.26: erf(x) for x >= 0; odd-extended for x < 0.\n  // if/else instead of select() — Safari/Metal has select() bugs.\n  var sign = 1.0;\n  if (x < 0.0) { sign = -1.0; }\n  let ax = abs(x);\n  let t = 1.0 / (1.0 + 0.3275911 * ax);\n  let y = 1.0 - (((((1.061405429 * t - 1.453152027) * t) + 1.421413741) * t - 0.284496736) * t + 0.254829592) * t * exp(-ax * ax);\n  return sign * y;\n}\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n  let x = input[idx];\n  output[idx] = 0.5 * x * (1.0 + erf_approx(x * INV_SQRT2));\n}\n`;\n\nconst WGSL_SLICE_COLS = `\\\n// Extract a contiguous column range from a [rows, in_width] tensor into a\n// [rows, out_width] tensor:  output[r, c] = input[r, col_offset + c].\n// Used to split the fused ViT qkv projection [N, 3*768] into Q/K/V [N, 768].\n\nstruct Params {\n  rows: u32,\n  in_width: u32,\n  out_width: u32,\n  col_offset: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read_write> output: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  let total = params.rows * params.out_width;\n  if (idx >= total) { return; }\n  let r = idx / params.out_width;\n  let c = idx % params.out_width;\n  output[idx] = input[r * params.in_width + params.col_offset + c];\n}\n`;\n\nconst WGSL_MUL_COLS = `\\\n// Multiply two contiguous column ranges of a single [rows, in_width] tensor:\n//   output[r, c] = src[r, off_a + c] * src[r, off_b + c]\n// LFM2 short-conv pre-gate: Bx = B * x where B and x are column slices of the\n// fused in_proj output — fuses two SliceCols + a Mul into one dispatch.\n\nstruct Params {\n  rows: u32,\n  in_width: u32,\n  out_width: u32,\n  off_a: u32,\n  off_b: u32,\n  _pad0: u32,\n  _pad1: u32,\n  _pad2: u32,\n}\n\n@group(0) @binding(0) var<storage, read> src: array<f32>;\n@group(0) @binding(1) var<storage, read_write> output: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  let total = params.rows * params.out_width;\n  if (idx >= total) { return; }\n  let r = idx / params.out_width;\n  let c = idx % params.out_width;\n  let base = r * params.in_width;\n  output[idx] = src[base + params.off_a + c] * src[base + params.off_b + c];\n}\n`;\n\nconst WGSL_APPLY_ROTARY = `\\\n// Apply precomputed rotary embeddings to a [T, num_heads*head_dim] tensor.\n// rotate_half convention (matches HF apply_rotary_pos_emb_vision):\n//   out = x * cos + rotate_half(x) * sin\n//   rotate_half(x) = concat(-x[half:], x[:half]) over each head's head_dim.\n// cos/sin are [T, head_dim], broadcast across all heads of a row.\n// In-place safe is NOT assumed — writes to a separate output buffer.\n\nstruct Params {\n  seq_len: u32,\n  num_heads: u32,\n  head_dim: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> x: array<f32>;\n@group(0) @binding(1) var<storage, read> cos_t: array<f32>;\n@group(0) @binding(2) var<storage, read> sin_t: array<f32>;\n@group(0) @binding(3) var<storage, read_write> output: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  let hd = params.head_dim;\n  let total = params.seq_len * params.num_heads * hd;\n  if (idx >= total) { return; }\n\n  let d = idx % hd;          // dim within head\n  let row = idx / (params.num_heads * hd); // token position (for cos/sin lookup)\n  let half = hd / 2u;\n\n  let xv = x[idx];\n  // rotate_half: for d < half → pairs with (d+half) as -x2; for d >= half → x1\n  var rot: f32;\n  if (d < half) {\n    rot = -x[idx + half];\n  } else {\n    rot = x[idx - half];\n  }\n\n  let cs = cos_t[row * hd + d];\n  let sn = sin_t[row * hd + d];\n  output[idx] = xv * cs + rot * sn;\n}\n`;\n\nconst WGSL_MROPE = `\\\n// Multimodal RoPE (M-RoPE) with partial rotation, rotate_half convention.\n// Identical math to WGSL_ROPE but the per-token angles come from host-precomputed\n// cos/sin tables [seq, rope_dim] (built from 3D position ids) instead of pos*inv_freq.\n// Only the first rope_dim of each head is rotated; the rest passes through.\n// For text-only this is fed linear-position cos/sin and reduces to standard 1D RoPE.\n//\n// Q: [T, num_q_heads*head_dim]   K: [T, num_kv_heads*head_dim]\n// cos/sin: [seq, rope_dim]   (rope_dim == 2*half_rope; emb=cat(freqs,freqs))\n\nstruct Params {\n  seq_len: u32,\n  num_q_heads: u32,\n  num_kv_heads: u32,\n  head_dim: u32,\n  rope_dim: u32,\n  pos_offset: u32,  // row offset into cos/sin for decode (single-token steps)\n  _pad0: u32,\n  _pad1: u32,\n}\n\n@group(0) @binding(0) var<storage, read_write> q: array<f32>;\n@group(0) @binding(1) var<storage, read_write> k: array<f32>;\n@group(0) @binding(2) var<storage, read> cos_t: array<f32>;\n@group(0) @binding(3) var<storage, read> sin_t: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let half_rope = params.rope_dim / 2u;\n  let total_q_pairs = params.seq_len * params.num_q_heads * half_rope;\n  let total_k_pairs = params.seq_len * params.num_kv_heads * half_rope;\n  let idx = gid.x;\n  let rd = params.rope_dim;\n\n  if (idx < total_q_pairs) {\n    let pos = idx / (params.num_q_heads * half_rope);\n    let remainder = idx % (params.num_q_heads * half_rope);\n    let head = remainder / half_rope;\n    let pair_idx = remainder % half_rope;\n\n    let cos_row = (pos + params.pos_offset) * rd;\n    let cs = cos_t[cos_row + pair_idx];\n    let sn = sin_t[cos_row + pair_idx];\n\n    let q_stride = params.num_q_heads * params.head_dim;\n    let head_base = pos * q_stride + head * params.head_dim;\n    let lo = head_base + pair_idx;\n    let hi = head_base + pair_idx + half_rope;\n    let q_lo = q[lo];\n    let q_hi = q[hi];\n    // rotate_half: out_lo = lo*cos - hi*sin ; out_hi = hi*cos + lo*sin\n    q[lo] = q_lo * cs - q_hi * sn;\n    q[hi] = q_hi * cs + q_lo * sn;\n  }\n\n  if (idx < total_k_pairs) {\n    let pos = idx / (params.num_kv_heads * half_rope);\n    let remainder = idx % (params.num_kv_heads * half_rope);\n    let head = remainder / half_rope;\n    let pair_idx = remainder % half_rope;\n\n    let cos_row = (pos + params.pos_offset) * rd;\n    let cs = cos_t[cos_row + pair_idx];\n    let sn = sin_t[cos_row + pair_idx];\n\n    let k_stride = params.num_kv_heads * params.head_dim;\n    let head_base = pos * k_stride + head * params.head_dim;\n    let lo = head_base + pair_idx;\n    let hi = head_base + pair_idx + half_rope;\n    let k_lo = k[lo];\n    let k_hi = k[hi];\n    k[lo] = k_lo * cs - k_hi * sn;\n    k[hi] = k_hi * cs + k_lo * sn;\n  }\n}\n`;\n\nconst WGSL_EMBED_SPLICE = `\\\n// Overwrite image-token rows of the text embeddings with merged vision tokens.\n// embed_out[row, c] = vision[map[row], c]  when map[row] >= 0, else embed_in[row,c].\n// row_map: i32 per token; -1 for text rows, else the index into the vision buffer.\n// Writes to a separate output buffer (pooled-alias-safe).\n\nstruct Params {\n  seq_len: u32,\n  hidden: u32,\n}\n\n@group(0) @binding(0) var<storage, read> embed_in: array<f32>;\n@group(0) @binding(1) var<storage, read> vision: array<f32>;\n@group(0) @binding(2) var<storage, read> row_map: array<i32>;\n@group(0) @binding(3) var<storage, read_write> embed_out: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  let total = params.seq_len * params.hidden;\n  if (idx >= total) { return; }\n  let row = idx / params.hidden;\n  let col = idx % params.hidden;\n  let vrow = row_map[row];\n  if (vrow >= 0) {\n    embed_out[idx] = vision[u32(vrow) * params.hidden + col];\n  } else {\n    embed_out[idx] = embed_in[idx];\n  }\n}\n`;\n\nconst WGSL_SWIGLU = `\\\n// Fused SwiGLU: output = SiLU(gate) * up = (gate / (1 + exp(-gate))) * up\n// Saves 2 dispatches per MLP block (SiLU + Mul → single kernel).\n\nstruct Params {\n  count: u32,\n}\n\n@group(0) @binding(0) var<storage, read> gate: array<f32>;\n@group(0) @binding(1) var<storage, read> up: array<f32>;\n@group(0) @binding(2) var<storage, read_write> output: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n\n  let g = gate[idx];\n  output[idx] = (g / (1.0 + exp(max(-g, -80.0)))) * up[idx];\n}\n`;\n\nconst WGSL_SWIGLU_MATVEC = `\\\n// Fused gate+up projection + SwiGLU for M=1 decode.\n// output[d] = SiLU(A · B_gate[d,:]) * (A · B_up[d,:])\n// Reads input vector A once from L1 cache, computes both projections,\n// and applies SiLU gating in a single dispatch.\n// Saves 2 dispatches per MLP block vs separate gate_proj + up_proj + SwiGLU.\n\nconst N_TILE: u32 = 8u;\nconst K_THREADS: u32 = 32u;\n\nstruct Params {\n  K: u32,\n  N: u32,\n}\n\n@group(0) @binding(0) var<storage, read> A: array<f32>;\n@group(0) @binding(1) var<storage, read> B_gate: array<f32>;\n@group(0) @binding(2) var<storage, read> B_up: array<f32>;\n@group(0) @binding(3) var<storage, read_write> output: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\nvar<workgroup> shared_gate: array<f32, 256>;\nvar<workgroup> shared_up: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let tid = lid.x;\n  let n_idx = tid / K_THREADS;\n  let k_tid = tid % K_THREADS;\n  let col = wid.x * N_TILE + n_idx;\n  let K = params.K;\n\n  var sum_gate: f32 = 0.0;\n  var sum_up: f32 = 0.0;\n  if (col < params.N) {\n    let b_off = col * K;\n    var k = k_tid * 4u;\n    while (k + 3u < K) {\n      let a_v = vec4f(A[k], A[k+1u], A[k+2u], A[k+3u]);\n      sum_gate += dot(a_v, vec4f(B_gate[b_off+k], B_gate[b_off+k+1u], B_gate[b_off+k+2u], B_gate[b_off+k+3u]));\n      sum_up += dot(a_v, vec4f(B_up[b_off+k], B_up[b_off+k+1u], B_up[b_off+k+2u], B_up[b_off+k+3u]));\n      k += K_THREADS * 4u;\n    }\n  }\n\n  shared_gate[tid] = sum_gate;\n  shared_up[tid] = sum_up;\n  workgroupBarrier();\n\n  if (k_tid == 0u && col < params.N) {\n    let base = n_idx * K_THREADS;\n    var tg: f32 = shared_gate[base];\n    var tu: f32 = shared_up[base];\n    for (var i: u32 = 1u; i < K_THREADS; i++) {\n      tg += shared_gate[base + i];\n      tu += shared_up[base + i];\n    }\n    let g = tg;\n    output[col] = (g / (1.0 + exp(max(-g, -80.0)))) * tu;\n  }\n}\n`;\n\nconst WGSL_SWIGLU_MATVEC_INT4 = `\\\n// CONSTRAINT: N_TILE must equal workgroup_size / K_THREADS\nconst N_TILE: u32 = 8u;\nconst K_THREADS: u32 = 32u;\n\nstruct Params {\n  K: u32,\n  N: u32,\n  group_size: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> A: array<vec4f>;\n@group(0) @binding(1) var<storage, read> B_gate_q: array<vec4u>;\n@group(0) @binding(2) var<storage, read> B_gate_s: array<f32>;\n@group(0) @binding(3) var<storage, read> B_gate_z: array<f32>;\n@group(0) @binding(4) var<storage, read> B_up_q: array<vec4u>;\n@group(0) @binding(5) var<storage, read> B_up_s: array<f32>;\n@group(0) @binding(6) var<storage, read> B_up_z: array<f32>;\n@group(0) @binding(7) var<storage, read_write> output: array<f32>;\n@group(0) @binding(8) var<storage, read> params: Params;\n\nvar<workgroup> shared_gate: array<f32, 256>;\nvar<workgroup> shared_up: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let tid = lid.x;\n  let n_idx = tid / K_THREADS;\n  let k_tid = tid % K_THREADS;\n  let col = wid.x * N_TILE + n_idx;\n  let K = params.K;\n  let k_vec = K / 32u;\n  let groups_per_col = K / params.group_size;\n  let packed_per_group = params.group_size / 8u;\n\n  // vec4 loads: each thread reads one vec4u per matrix per iteration\n  // (4 packed u32s = 32 INT4 weights). group_size (128) spans 16 packed u32s,\n  // so a vec4u never straddles a quantization group.\n  var sum_gate: f32 = 0.0;\n  var sum_up: f32 = 0.0;\n  if (col < params.N) {\n    let b_off_v = col * k_vec;\n    let col_g_base = col * groups_per_col;\n\n    var v = k_tid;\n    while (v < k_vec) {\n      let group_idx = col_g_base + (v * 4u) / packed_per_group;\n      let gq = B_gate_q[b_off_v + v];\n      let g_scale = B_gate_s[group_idx];\n      let g_zero = B_gate_z[group_idx];\n      let uq = B_up_q[b_off_v + v];\n      let u_scale = B_up_s[group_idx];\n      let u_zero = B_up_z[group_idx];\n      let a_base = v * 8u;\n\n      for (var j: u32 = 0u; j < 4u; j++) {\n        // Gate weight dequantization\n        let g_packed = gq[j];\n        let gn0 = f32(g_packed & 0xFu); let gn1 = f32((g_packed >> 4u) & 0xFu);\n        let gn2 = f32((g_packed >> 8u) & 0xFu); let gn3 = f32((g_packed >> 12u) & 0xFu);\n        let gn4 = f32((g_packed >> 16u) & 0xFu); let gn5 = f32((g_packed >> 20u) & 0xFu);\n        let gn6 = f32((g_packed >> 24u) & 0xFu); let gn7 = f32((g_packed >> 28u) & 0xFu);\n        let gb0 = vec4f(gn0 - g_zero, gn1 - g_zero, gn2 - g_zero, gn3 - g_zero) * g_scale;\n        let gb1 = vec4f(gn4 - g_zero, gn5 - g_zero, gn6 - g_zero, gn7 - g_zero) * g_scale;\n\n        // Up weight dequantization\n        let u_packed = uq[j];\n        let un0 = f32(u_packed & 0xFu); let un1 = f32((u_packed >> 4u) & 0xFu);\n        let un2 = f32((u_packed >> 8u) & 0xFu); let un3 = f32((u_packed >> 12u) & 0xFu);\n        let un4 = f32((u_packed >> 16u) & 0xFu); let un5 = f32((u_packed >> 20u) & 0xFu);\n        let un6 = f32((u_packed >> 24u) & 0xFu); let un7 = f32((u_packed >> 28u) & 0xFu);\n        let ub0 = vec4f(un0 - u_zero, un1 - u_zero, un2 - u_zero, un3 - u_zero) * u_scale;\n        let ub1 = vec4f(un4 - u_zero, un5 - u_zero, un6 - u_zero, un7 - u_zero) * u_scale;\n\n        let a0 = A[a_base + j * 2u];\n        let a1 = A[a_base + j * 2u + 1u];\n\n        sum_gate += dot(a0, gb0) + dot(a1, gb1);\n        sum_up += dot(a0, ub0) + dot(a1, ub1);\n      }\n\n      v += K_THREADS;\n    }\n  }\n\n  shared_gate[tid] = sum_gate;\n  shared_up[tid] = sum_up;\n  workgroupBarrier();\n\n  if (k_tid == 0u && col < params.N) {\n    let base = n_idx * K_THREADS;\n    var tg: f32 = shared_gate[base];\n    var tu: f32 = shared_up[base];\n    for (var i: u32 = 1u; i < K_THREADS; i++) {\n      tg += shared_gate[base + i];\n      tu += shared_up[base + i];\n    }\n    let gg = tg;\n    output[col] = (gg / (1.0 + exp(max(-gg, -80.0)))) * tu;\n  }\n}\n`;\n\n// ── Dual MatVecInt4 (two independent INT4 projections sharing input A, M=1) ──\n//\n// out0[d] = A · dequant(B0[d,:]),  out1[d] = A · dequant(B1[d,:])\n// Identical dequant + reduction to WGSL_SWIGLU_MATVEC_INT4, but writes the two\n// projection sums to SEPARATE output buffers instead of SiLU-combining them.\n// Used to fuse q_proj+gate_proj and k_proj+v_proj in full-attention decode:\n// both projections share the same input_layernorm activation and the same K/N,\n// so reading A once and computing both halves removes one full GPU round-trip\n// (one submit+drain on Safari/iOS) per fused pair.\nconst WGSL_DUAL_MATVEC_INT4 = `\\\nconst N_TILE: u32 = 8u;\nconst K_THREADS: u32 = 32u;\n\nstruct Params {\n  K: u32,\n  N: u32,\n  group_size: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> A: array<vec4f>;\n@group(0) @binding(1) var<storage, read> B0_q: array<vec4u>;\n@group(0) @binding(2) var<storage, read> B0_s: array<f32>;\n@group(0) @binding(3) var<storage, read> B0_z: array<f32>;\n@group(0) @binding(4) var<storage, read> B1_q: array<vec4u>;\n@group(0) @binding(5) var<storage, read> B1_s: array<f32>;\n@group(0) @binding(6) var<storage, read> B1_z: array<f32>;\n@group(0) @binding(7) var<storage, read_write> out0: array<f32>;\n@group(0) @binding(8) var<storage, read_write> out1: array<f32>;\n@group(0) @binding(9) var<storage, read> params: Params;\n\nvar<workgroup> shared_0: array<f32, 256>;\nvar<workgroup> shared_1: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let tid = lid.x;\n  let n_idx = tid / K_THREADS;\n  let k_tid = tid % K_THREADS;\n  let col = wid.x * N_TILE + n_idx;\n  let K = params.K;\n  let k_vec = K / 32u;\n  let groups_per_col = K / params.group_size;\n  let packed_per_group = params.group_size / 8u;\n\n  var sum_0: f32 = 0.0;\n  var sum_1: f32 = 0.0;\n  if (col < params.N) {\n    let b_off_v = col * k_vec;\n    let col_g_base = col * groups_per_col;\n\n    var v = k_tid;\n    while (v < k_vec) {\n      let group_idx = col_g_base + (v * 4u) / packed_per_group;\n      let q0 = B0_q[b_off_v + v];\n      let s0 = B0_s[group_idx];\n      let z0 = B0_z[group_idx];\n      let q1 = B1_q[b_off_v + v];\n      let s1 = B1_s[group_idx];\n      let z1 = B1_z[group_idx];\n      let a_base = v * 8u;\n\n      for (var j: u32 = 0u; j < 4u; j++) {\n        let p0 = q0[j];\n        let a0n0 = f32(p0 & 0xFu); let a0n1 = f32((p0 >> 4u) & 0xFu);\n        let a0n2 = f32((p0 >> 8u) & 0xFu); let a0n3 = f32((p0 >> 12u) & 0xFu);\n        let a0n4 = f32((p0 >> 16u) & 0xFu); let a0n5 = f32((p0 >> 20u) & 0xFu);\n        let a0n6 = f32((p0 >> 24u) & 0xFu); let a0n7 = f32((p0 >> 28u) & 0xFu);\n        let b0a = vec4f(a0n0 - z0, a0n1 - z0, a0n2 - z0, a0n3 - z0) * s0;\n        let b0b = vec4f(a0n4 - z0, a0n5 - z0, a0n6 - z0, a0n7 - z0) * s0;\n\n        let p1 = q1[j];\n        let a1n0 = f32(p1 & 0xFu); let a1n1 = f32((p1 >> 4u) & 0xFu);\n        let a1n2 = f32((p1 >> 8u) & 0xFu); let a1n3 = f32((p1 >> 12u) & 0xFu);\n        let a1n4 = f32((p1 >> 16u) & 0xFu); let a1n5 = f32((p1 >> 20u) & 0xFu);\n        let a1n6 = f32((p1 >> 24u) & 0xFu); let a1n7 = f32((p1 >> 28u) & 0xFu);\n        let b1a = vec4f(a1n0 - z1, a1n1 - z1, a1n2 - z1, a1n3 - z1) * s1;\n        let b1b = vec4f(a1n4 - z1, a1n5 - z1, a1n6 - z1, a1n7 - z1) * s1;\n\n        let av0 = A[a_base + j * 2u];\n        let av1 = A[a_base + j * 2u + 1u];\n\n        sum_0 += dot(av0, b0a) + dot(av1, b0b);\n        sum_1 += dot(av0, b1a) + dot(av1, b1b);\n      }\n\n      v += K_THREADS;\n    }\n  }\n\n  shared_0[tid] = sum_0;\n  shared_1[tid] = sum_1;\n  workgroupBarrier();\n\n  if (k_tid == 0u && col < params.N) {\n    let base = n_idx * K_THREADS;\n    var t0: f32 = shared_0[base];\n    var t1: f32 = shared_1[base];\n    for (var i: u32 = 1u; i < K_THREADS; i++) {\n      t0 += shared_0[base + i];\n      t1 += shared_1[base + i];\n    }\n    out0[col] = t0;\n    out1[col] = t1;\n  }\n}\n`;\n\nconst WGSL_RESIDUAL_RMSNORM = `\\\n// Fused residual add + RMS normalization.\n// sum_output = a + b  (residual sum, kept for downstream residual)\n// norm_output = RMSNorm(a + b) * weight\n\nstruct Params {\n  seq_len: u32,\n  hidden_size: u32,\n  eps_bits: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> a: array<f32>;\n@group(0) @binding(1) var<storage, read> b: array<f32>;\n@group(0) @binding(2) var<storage, read> weight: array<f32>;\n@group(0) @binding(3) var<storage, read_write> sum_output: array<f32>;\n@group(0) @binding(4) var<storage, read_write> norm_output: array<f32>;\n@group(0) @binding(5) var<storage, read> params: Params;\n\nvar<workgroup> shared_sum: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let row = wid.x;\n  if (row >= params.seq_len) { return; }\n\n  let tid = lid.x;\n  let row_offset = row * params.hidden_size;\n  let eps = bitcast<f32>(params.eps_bits);\n\n  // Pass 1: add residual, store sum, accumulate sum-of-squares\n  var local_sum: f32 = 0.0;\n  var i = tid;\n  while (i < params.hidden_size) {\n    let val = a[row_offset + i] + b[row_offset + i];\n    sum_output[row_offset + i] = val;\n    local_sum += val * val;\n    i += 256u;\n  }\n  shared_sum[tid] = local_sum;\n  workgroupBarrier();\n\n  // Tree reduction\n  var stride: u32 = 128u;\n  while (stride > 0u) {\n    if (tid < stride) {\n      shared_sum[tid] += shared_sum[tid + stride];\n    }\n    workgroupBarrier();\n    stride = stride / 2u;\n  }\n\n  let rms = sqrt(shared_sum[0] / f32(params.hidden_size) + eps);\n\n  // Pass 2: normalize and scale\n  i = tid;\n  while (i < params.hidden_size) {\n    let val = sum_output[row_offset + i];\n    norm_output[row_offset + i] = (val / rms) * weight[i];\n    i += 256u;\n  }\n}\n`;\n\nconst WGSL_CAUSAL_CONV1D = `\\\n// Causal 1D depthwise convolution with state buffer for autoregressive decode.\n// conv_state stores the last (kernel_size-1) input timesteps from prior forward passes.\n\nstruct Params {\n  seq_len: u32,\n  channels: u32,\n  kernel_size: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read> weight: array<f32>;\n@group(0) @binding(2) var<storage, read> conv_state: array<f32>;\n@group(0) @binding(3) var<storage, read_write> output: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  let total = params.seq_len * params.channels;\n  if (idx >= total) { return; }\n\n  let t = idx / params.channels;\n  let c = idx % params.channels;\n  let state_size = params.kernel_size - 1u;\n\n  var sum: f32 = 0.0;\n  for (var k: u32 = 0u; k < params.kernel_size; k = k + 1u) {\n    let src_t = i32(t) - i32(k);\n    var in_val: f32 = 0.0;\n    if (src_t >= 0) {\n      in_val = input[u32(src_t) * params.channels + c];\n    } else {\n      // Read from state: state[0]=oldest, state[state_size-1]=most recent\n      let state_t = i32(state_size) + src_t;\n      if (state_t >= 0) {\n        in_val = conv_state[u32(state_t) * params.channels + c];\n      }\n    }\n    // PyTorch Conv1d stores weights reversed: w[0]=most distant, w[K-1]=current.\n    // Our loop looks back k steps, so we need w[K-1-k] to match PyTorch.\n    let w_val = weight[c * params.kernel_size + (params.kernel_size - 1u - k)];\n    sum += in_val * w_val;\n  }\n\n  output[t * params.channels + c] = sum;\n}\n`;\n\nconst WGSL_CAUSAL_CONV1D_SILU = `\\\n// Causal 1D depthwise convolution with fused SiLU activation on the output.\n// Identical to WGSL_CAUSAL_CONV1D but applies output = silu(sum), eliminating a\n// separate SiLU dispatch + a full read/write of the conv output (Qwen3.5 linear\n// attention: F.silu(conv1d(x))).\n\nstruct Params {\n  seq_len: u32,\n  channels: u32,\n  kernel_size: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read> weight: array<f32>;\n@group(0) @binding(2) var<storage, read> conv_state: array<f32>;\n@group(0) @binding(3) var<storage, read_write> output: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  let total = params.seq_len * params.channels;\n  if (idx >= total) { return; }\n\n  let t = idx / params.channels;\n  let c = idx % params.channels;\n  let state_size = params.kernel_size - 1u;\n\n  var sum: f32 = 0.0;\n  for (var k: u32 = 0u; k < params.kernel_size; k = k + 1u) {\n    let src_t = i32(t) - i32(k);\n    var in_val: f32 = 0.0;\n    if (src_t >= 0) {\n      in_val = input[u32(src_t) * params.channels + c];\n    } else {\n      let state_t = i32(state_size) + src_t;\n      if (state_t >= 0) {\n        in_val = conv_state[u32(state_t) * params.channels + c];\n      }\n    }\n    let w_val = weight[c * params.kernel_size + (params.kernel_size - 1u - k)];\n    sum += in_val * w_val;\n  }\n\n  // Fused SiLU: silu(x) = x / (1 + exp(-x))\n  output[t * params.channels + c] = sum / (1.0 + exp(max(-sum, -80.0)));\n}\n`;\n\nconst WGSL_CAUSAL_CONV1D_GATED = `\\\n// Causal 1D depthwise convolution with a fused multiplicative output gate:\n//   output = gate * conv(input)\n// LFM2 short-conv: y = C * depthwise_conv(Bx). Fusing the C-gate into the conv\n// eliminates a separate Mul dispatch + a conv_dim-wide read/write per layer.\n\nstruct Params {\n  seq_len: u32,\n  channels: u32,\n  kernel_size: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read> weight: array<f32>;\n@group(0) @binding(2) var<storage, read> conv_state: array<f32>;\n@group(0) @binding(3) var<storage, read> gate: array<f32>;\n@group(0) @binding(4) var<storage, read_write> output: array<f32>;\n@group(0) @binding(5) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  let total = params.seq_len * params.channels;\n  if (idx >= total) { return; }\n\n  let t = idx / params.channels;\n  let c = idx % params.channels;\n  let state_size = params.kernel_size - 1u;\n\n  var sum: f32 = 0.0;\n  for (var k: u32 = 0u; k < params.kernel_size; k = k + 1u) {\n    let src_t = i32(t) - i32(k);\n    var in_val: f32 = 0.0;\n    if (src_t >= 0) {\n      in_val = input[u32(src_t) * params.channels + c];\n    } else {\n      let state_t = i32(state_size) + src_t;\n      if (state_t >= 0) {\n        in_val = conv_state[u32(state_t) * params.channels + c];\n      }\n    }\n    let w_val = weight[c * params.kernel_size + (params.kernel_size - 1u - k)];\n    sum += in_val * w_val;\n  }\n\n  output[t * params.channels + c] = gate[idx] * sum;\n}\n`;\n\nconst WGSL_SIGMOID_GATE = `\\\n// Sigmoid gate: output = x * sigmoid(gate)\n\nstruct Params {\n  count: u32,\n}\n\n@group(0) @binding(0) var<storage, read> x: array<f32>;\n@group(0) @binding(1) var<storage, read> gate: array<f32>;\n@group(0) @binding(2) var<storage, read_write> output: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n\n  let g = gate[idx];\n  let sig = 1.0 / (1.0 + exp(max(-g, -80.0)));\n  output[idx] = x[idx] * sig;\n}\n`;\n\nconst WGSL_MAMBA_SSM = `\\\n// Gated Delta Net recurrence (Qwen3.5 linear attention).\n//\n// For each head, maintains state [key_dim, val_dim] across timesteps.\n// Recurrence per timestep:\n//   g = -exp(A_log) * softplus(a + dt_bias)   (log-space decay)\n//   decay = exp(g)                             (actual decay in (0,1))\n//   state *= decay                             (decay state)\n//   q_norm, k_norm = L2_normalize(q), L2_normalize(k)\n//   kv_mem = state^T @ k_norm                  (retrieve from memory)\n//   delta = (v - kv_mem) * sigmoid(b)          (delta rule)\n//   state += k_norm @ delta^T                  (rank-1 update)\n//   output = state^T @ q_norm / sqrt(key_dim)  (query output)\n\nstruct Params {\n  T: u32,\n  num_heads: u32,\n  key_dim: u32,\n  val_dim: u32,\n  qkv_dim: u32,\n  _pad1: u32,\n  _pad2: u32,\n  _pad3: u32,\n}\n\n@group(0) @binding(0) var<storage, read> qkv_conv: array<f32>;\n@group(0) @binding(1) var<storage, read> a_input: array<f32>;\n@group(0) @binding(2) var<storage, read> b_input: array<f32>;\n@group(0) @binding(3) var<storage, read> A_log: array<f32>;\n@group(0) @binding(4) var<storage, read> dt_bias: array<f32>;\n@group(0) @binding(5) var<storage, read_write> ssm_state: array<f32>;\n@group(0) @binding(6) var<storage, read_write> output: array<f32>;\n@group(0) @binding(7) var<storage, read> params: Params;\n\n// Shared memory for Q and K vectors (L2-normalized) and reduction\nvar<workgroup> shared_q: array<f32, 128>;\nvar<workgroup> shared_k: array<f32, 128>;\nvar<workgroup> shared_norm: array<f32, 128>;\n\n@compute @workgroup_size(128)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let h = wid.x;  // uniform within workgroup\n  if (h >= params.num_heads) { return; }\n\n  let d = lid.x;  // 0..127, all threads participate in barriers\n\n  let A_val = -exp(A_log[h]);\n  let state_base = h * params.key_dim * params.val_dim;\n  let scale = 1.0 / sqrt(f32(params.key_dim));\n\n  let q_off = h * params.key_dim;\n  let k_off = params.num_heads * params.key_dim + h * params.key_dim;\n  let v_off = 2u * params.num_heads * params.key_dim + h * params.val_dim;\n  let out_stride = params.num_heads * params.val_dim;\n\n  for (var t: u32 = 0u; t < params.T; t = t + 1u) {\n    let qkv_base = t * params.qkv_dim;\n\n    // ── Load Q and K into shared memory ──\n    shared_q[d] = qkv_conv[qkv_base + q_off + d];\n    shared_k[d] = qkv_conv[qkv_base + k_off + d];\n    workgroupBarrier();\n\n    // ── L2-normalize Q (parallel reduction) ──\n    shared_norm[d] = shared_q[d] * shared_q[d];\n    workgroupBarrier();\n    var stride: u32 = 64u;\n    while (stride > 0u) {\n      if (d < stride) { shared_norm[d] += shared_norm[d + stride]; }\n      workgroupBarrier();\n      stride /= 2u;\n    }\n    let q_inv_norm = 1.0 / sqrt(shared_norm[0] + 1e-12);\n    shared_q[d] *= q_inv_norm;\n\n    // ── L2-normalize K (parallel reduction) ──\n    shared_norm[d] = shared_k[d] * shared_k[d];\n    workgroupBarrier();\n    stride = 64u;\n    while (stride > 0u) {\n      if (d < stride) { shared_norm[d] += shared_norm[d + stride]; }\n      workgroupBarrier();\n      stride /= 2u;\n    }\n    let k_inv_norm = 1.0 / sqrt(shared_norm[0] + 1e-12);\n    shared_k[d] *= k_inv_norm;\n    workgroupBarrier();\n\n    // ── Compute decay and beta ──\n    let a_val = a_input[t * params.num_heads + h];\n    let b_val = b_input[t * params.num_heads + h];\n    let dt_in = a_val + dt_bias[h];\n    let sp = log(1.0 + exp(max(dt_in, -80.0)));\n    let g = A_val * sp;\n    let decay = exp(max(g, -80.0));\n    let beta = 1.0 / (1.0 + exp(max(-b_val, -80.0)));\n\n    // ── 1+2. Retrieve from memory (decay is a per-step scalar, so it factors\n    // out of the dot product): kv_mem[d] = decay * Σ_k state[k,d] * k_norm[k] ──\n    let v_val = qkv_conv[qkv_base + v_off + d];\n    var kv_dot: f32 = 0.0;\n    for (var k: u32 = 0u; k < params.key_dim; k = k + 1u) {\n      kv_dot += ssm_state[state_base + k * params.val_dim + d] * shared_k[k];\n    }\n    let kv_mem = kv_dot * decay;\n\n    // ── 3. Delta rule: delta[d] = (v[d] - kv_mem[d]) * beta ──\n    let delta_val = (v_val - kv_mem) * beta;\n\n    // ── 4+5. Fused decay + rank-1 update + output in ONE pass over state\n    // (one read + one write instead of three reads + two writes):\n    //   state'[k,d] = state[k,d] * decay + k_norm[k] * delta[d]\n    //   out[d]      = Σ_k state'[k,d] * q_norm[k] * scale ──\n    var out_val: f32 = 0.0;\n    for (var k: u32 = 0u; k < params.key_dim; k = k + 1u) {\n      let idx = state_base + k * params.val_dim + d;\n      let s = ssm_state[idx] * decay + shared_k[k] * delta_val;\n      ssm_state[idx] = s;\n      out_val += s * shared_q[k];\n    }\n    output[t * out_stride + h * params.val_dim + d] = out_val * scale;\n\n    workgroupBarrier();  // sync before next timestep loads shared_q/k\n  }\n}\n`;\n\nconst WGSL_MAMBA_SSM_F16 = `\\\n// Gated Delta Net recurrence with f16 SSM state storage (Dawn only).\n// Compute stays f32; only the persistent state reads/writes are half-precision,\n// halving the dominant state bandwidth. Buffer is allocated as f32-sized and\n// reinterpreted (extra capacity unused).\n//\n// For each head, maintains state [key_dim, val_dim] across timesteps.\n// Recurrence per timestep:\n//   g = -exp(A_log) * softplus(a + dt_bias)   (log-space decay)\n//   decay = exp(g)                             (actual decay in (0,1))\n//   state *= decay                             (decay state)\n//   q_norm, k_norm = L2_normalize(q), L2_normalize(k)\n//   kv_mem = state^T @ k_norm                  (retrieve from memory)\n//   delta = (v - kv_mem) * sigmoid(b)          (delta rule)\n//   state += k_norm @ delta^T                  (rank-1 update)\n//   output = state^T @ q_norm / sqrt(key_dim)  (query output)\n\nstruct Params {\n  T: u32,\n  num_heads: u32,\n  key_dim: u32,\n  val_dim: u32,\n  qkv_dim: u32,\n  _pad1: u32,\n  _pad2: u32,\n  _pad3: u32,\n}\n\n@group(0) @binding(0) var<storage, read> qkv_conv: array<f32>;\n@group(0) @binding(1) var<storage, read> a_input: array<f32>;\n@group(0) @binding(2) var<storage, read> b_input: array<f32>;\n@group(0) @binding(3) var<storage, read> A_log: array<f32>;\n@group(0) @binding(4) var<storage, read> dt_bias: array<f32>;\n@group(0) @binding(5) var<storage, read_write> ssm_state: array<f16>;\n@group(0) @binding(6) var<storage, read_write> output: array<f32>;\n@group(0) @binding(7) var<storage, read> params: Params;\n\n// Shared memory for Q and K vectors (L2-normalized) and reduction\nvar<workgroup> shared_q: array<f32, 128>;\nvar<workgroup> shared_k: array<f32, 128>;\nvar<workgroup> shared_norm: array<f32, 128>;\n\n@compute @workgroup_size(128)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let h = wid.x;  // uniform within workgroup\n  if (h >= params.num_heads) { return; }\n\n  let d = lid.x;  // 0..127, all threads participate in barriers\n\n  let A_val = -exp(A_log[h]);\n  let state_base = h * params.key_dim * params.val_dim;\n  let scale = 1.0 / sqrt(f32(params.key_dim));\n\n  let q_off = h * params.key_dim;\n  let k_off = params.num_heads * params.key_dim + h * params.key_dim;\n  let v_off = 2u * params.num_heads * params.key_dim + h * params.val_dim;\n  let out_stride = params.num_heads * params.val_dim;\n\n  for (var t: u32 = 0u; t < params.T; t = t + 1u) {\n    let qkv_base = t * params.qkv_dim;\n\n    // ── Load Q and K into shared memory ──\n    shared_q[d] = qkv_conv[qkv_base + q_off + d];\n    shared_k[d] = qkv_conv[qkv_base + k_off + d];\n    workgroupBarrier();\n\n    // ── L2-normalize Q (parallel reduction) ──\n    shared_norm[d] = shared_q[d] * shared_q[d];\n    workgroupBarrier();\n    var stride: u32 = 64u;\n    while (stride > 0u) {\n      if (d < stride) { shared_norm[d] += shared_norm[d + stride]; }\n      workgroupBarrier();\n      stride /= 2u;\n    }\n    let q_inv_norm = 1.0 / sqrt(shared_norm[0] + 1e-12);\n    shared_q[d] *= q_inv_norm;\n\n    // ── L2-normalize K (parallel reduction) ──\n    shared_norm[d] = shared_k[d] * shared_k[d];\n    workgroupBarrier();\n    stride = 64u;\n    while (stride > 0u) {\n      if (d < stride) { shared_norm[d] += shared_norm[d + stride]; }\n      workgroupBarrier();\n      stride /= 2u;\n    }\n    let k_inv_norm = 1.0 / sqrt(shared_norm[0] + 1e-12);\n    shared_k[d] *= k_inv_norm;\n    workgroupBarrier();\n\n    // ── Compute decay and beta ──\n    let a_val = a_input[t * params.num_heads + h];\n    let b_val = b_input[t * params.num_heads + h];\n    let dt_in = a_val + dt_bias[h];\n    let sp = log(1.0 + exp(max(dt_in, -80.0)));\n    let g = A_val * sp;\n    let decay = exp(max(g, -80.0));\n    let beta = 1.0 / (1.0 + exp(max(-b_val, -80.0)));\n\n    // ── 1+2. Retrieve from memory (decay is a per-step scalar, so it factors\n    // out of the dot product): kv_mem[d] = decay * Σ_k state[k,d] * k_norm[k] ──\n    let v_val = qkv_conv[qkv_base + v_off + d];\n    var kv_dot: f32 = 0.0;\n    for (var k: u32 = 0u; k < params.key_dim; k = k + 1u) {\n      kv_dot += f32(ssm_state[state_base + k * params.val_dim + d]) * shared_k[k];\n    }\n    let kv_mem = kv_dot * decay;\n\n    // ── 3. Delta rule: delta[d] = (v[d] - kv_mem[d]) * beta ──\n    let delta_val = (v_val - kv_mem) * beta;\n\n    // ── 4+5. Fused decay + rank-1 update + output in ONE pass over state\n    // (one read + one write instead of three reads + two writes):\n    //   state'[k,d] = state[k,d] * decay + k_norm[k] * delta[d]\n    //   out[d]      = Σ_k state'[k,d] * q_norm[k] * scale ──\n    var out_val: f32 = 0.0;\n    for (var k: u32 = 0u; k < params.key_dim; k = k + 1u) {\n      let idx = state_base + k * params.val_dim + d;\n      let s = f32(ssm_state[idx]) * decay + shared_k[k] * delta_val;\n      ssm_state[idx] = f16(s);\n      out_val += s * shared_q[k];\n    }\n    output[t * out_stride + h * params.val_dim + d] = out_val * scale;\n\n    workgroupBarrier();  // sync before next timestep loads shared_q/k\n  }\n}\n`;\n\nconst WGSL_KV_CACHE_APPEND = `\\\n// Copy T new K or V vectors into cache at position offset.\n// src: [T, width], dst: [max_seq_len, width]\n\nstruct Params {\n  count: u32,       // T * width (total elements to copy)\n  dst_offset: u32,  // seqPos * width (element offset into cache)\n}\n\n@group(0) @binding(0) var<storage, read> src: array<f32>;\n@group(0) @binding(1) var<storage, read_write> dst: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n  dst[params.dst_offset + idx] = src[idx];\n}\n`;\n\nconst WGSL_DUAL_KV_CACHE_APPEND = `\\\n// Append both K and V into their f32 caches in one dispatch.\n// Both share the same width, T, and dst_offset (same kv_dim, same seqPos),\n// so a single param set + a single grid over 'count' elements drives both\n// copies. Pure memcpy — numerically identical to two separate appends.\n\nstruct Params {\n  count: u32,       // T * width (per-cache element count)\n  dst_offset: u32,  // seqPos * width (element offset into each cache)\n}\n\n@group(0) @binding(0) var<storage, read> src_k: array<f32>;\n@group(0) @binding(1) var<storage, read> src_v: array<f32>;\n@group(0) @binding(2) var<storage, read_write> dst_k: array<f32>;\n@group(0) @binding(3) var<storage, read_write> dst_v: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n  dst_k[params.dst_offset + idx] = src_k[idx];\n  dst_v[params.dst_offset + idx] = src_v[idx];\n}\n`;\n\nconst WGSL_DUAL_KV_CACHE_APPEND_F16 = `\\\n// Append both K and V into their native-f16 caches in one dispatch.\n// Same per-element conversion as the single f16 append, doubled.\n\nstruct Params {\n  count: u32,\n  dst_offset: u32,\n}\n\n@group(0) @binding(0) var<storage, read> src_k: array<f32>;\n@group(0) @binding(1) var<storage, read> src_v: array<f32>;\n@group(0) @binding(2) var<storage, read_write> dst_k: array<f16>;\n@group(0) @binding(3) var<storage, read_write> dst_v: array<f16>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n  dst_k[params.dst_offset + idx] = f16(src_k[idx]);\n  dst_v[params.dst_offset + idx] = f16(src_v[idx]);\n}\n`;\n\nconst WGSL_DUAL_KV_CACHE_APPEND_PACKED_F16 = `\\\n// Append both K and V into packed-f16 caches in one dispatch (Safari-safe).\n// Each u32 packs 2 f16 via pack2x16float; one thread handles one f16 pair per\n// cache. Same math as the single packed append, doubled — no 'enable f16'.\n\nstruct Params {\n  count: u32,       // T * width (per-cache f32 element count)\n  dst_offset: u32,  // seqPos * width (f32-element offset)\n}\n\n@group(0) @binding(0) var<storage, read> src_k: array<f32>;\n@group(0) @binding(1) var<storage, read> src_v: array<f32>;\n@group(0) @binding(2) var<storage, read_write> dst_k: array<u32>;\n@group(0) @binding(3) var<storage, read_write> dst_v: array<u32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let pair_idx = gid.x;\n  let num_pairs = params.count >> 1u;\n  if (pair_idx >= num_pairs) { return; }\n\n  let src_base = pair_idx * 2u;\n  let dst_idx = (params.dst_offset >> 1u) + pair_idx;\n\n  dst_k[dst_idx] = pack2x16float(vec2f(src_k[src_base], src_k[src_base + 1u]));\n  dst_v[dst_idx] = pack2x16float(vec2f(src_v[src_base], src_v[src_base + 1u]));\n}\n`;\n\nconst WGSL_KV_CACHE_APPEND_F16 = `\\\n// Copy T new K or V vectors into f16 cache at position offset.\n// src: [T, width] (f32), dst: [max_seq_len, width] (f16)\n// Converts f32 activations to f16 on write for reduced memory traffic.\n\nstruct Params {\n  count: u32,       // T * width (total elements to copy)\n  dst_offset: u32,  // seqPos * width (element offset into cache)\n}\n\n@group(0) @binding(0) var<storage, read> src: array<f32>;\n@group(0) @binding(1) var<storage, read_write> dst: array<f16>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n  dst[params.dst_offset + idx] = f16(src[idx]);\n}\n`;\n\nconst WGSL_KV_CACHE_APPEND_PACKED_F16 = `\\\n// Copy T new K or V vectors into packed f16 cache (Safari-safe).\n// src: [T, width] (f32), dst: [max_seq_len, width/2] (u32, packed f16 pairs)\n// Each u32 holds 2 f16 values via pack2x16float. Does NOT require 'enable f16'.\n// Params use f32-element-level addressing; kernel converts to u32 pair indices.\n\nstruct Params {\n  count: u32,       // T * width (total f32 elements)\n  dst_offset: u32,  // seqPos * width (f32-element offset into cache)\n}\n\n@group(0) @binding(0) var<storage, read> src: array<f32>;\n@group(0) @binding(1) var<storage, read_write> dst: array<u32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let pair_idx = gid.x;\n  let num_pairs = params.count >> 1u;\n  if (pair_idx >= num_pairs) { return; }\n\n  let src_base = pair_idx * 2u;\n  let dst_idx = (params.dst_offset >> 1u) + pair_idx;\n\n  dst[dst_idx] = pack2x16float(vec2f(src[src_base], src[src_base + 1u]));\n}\n`;\n\nconst WGSL_ATTENTION_F16 = `\\\n// Tiled online-softmax attention with f16 KV cache (Flash-Attention style).\n//\n// Same algorithm as WGSL_ATTENTION but K/V are stored as f16 in global memory.\n// Shared memory stays f32 for computation accuracy — only global reads change.\n// Halves KV cache memory traffic for ~1.3-1.8x decode speedup.\n\nconst TILE_S: u32 = 16u;\nconst WG: u32 = 256u;\nconst DOT_THREADS: u32 = 16u;\n\nstruct Params {\n  T: u32,\n  S: u32,\n  num_q_heads: u32,\n  num_kv_heads: u32,\n  head_dim: u32,\n  position_offset: u32,\n  // Softmax QK temperature (f32). Default 1/sqrt(head_dim); Gemma 4 sets 1.0.\n  attn_scale: f32,\n}\n\n@group(0) @binding(0) var<storage, read> Q: array<f32>;\n@group(0) @binding(1) var<storage, read> K: array<f16>;\n@group(0) @binding(2) var<storage, read> V: array<f16>;\n@group(0) @binding(3) var<storage, read_write> output: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\nvar<workgroup> smem: array<f32, 4096>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let q_pos = wid.x;\n  let q_head = wid.y;\n\n  if (q_pos >= params.T || q_head >= params.num_q_heads) { return; }\n\n  let tid = lid.x;\n  let heads_per_kv = params.num_q_heads / params.num_kv_heads;\n  let kv_head = q_head / heads_per_kv;\n  let scale = params.attn_scale;\n\n  let q_stride = params.num_q_heads * params.head_dim;\n  let kv_stride = params.num_kv_heads * params.head_dim;\n  let q_base = q_pos * q_stride + q_head * params.head_dim;\n  let kv_head_offset = kv_head * params.head_dim;\n  let causal_limit = q_pos + params.position_offset + 1u;\n  let hd = params.head_dim;\n  let S_eff = min(params.S, causal_limit);\n\n  var running_max: f32 = -1e30;\n  var running_sum: f32 = 0.0;\n  // Per-thread output accumulators; out_acc2 covers dim tid+WG so head_dim up to\n  // 2*WG (512) works with WG=256 threads. hd<=256 → out_acc2 unused (identical).\n  var out_acc: f32 = 0.0;\n  var out_acc2: f32 = 0.0;\n  let d2 = tid + WG;\n\n  // Tile capped so a K/V tile (tile_cap*hd f32) fits the 4096-f32 smem.\n  let tile_cap = min(TILE_S, 4096u / hd);\n  let num_tiles = (S_eff + tile_cap - 1u) / tile_cap;\n\n  for (var tile: u32 = 0u; tile < num_tiles; tile += 1u) {\n    let tile_start = tile * tile_cap;\n    let tile_end = min(tile_start + tile_cap, S_eff);\n    let tile_len = tile_end - tile_start;\n\n    // ── Step 1: Cooperative K tile load (f16 → f32 conversion) ──\n    let total_k_elems = tile_len * hd;\n    var load_idx = tid;\n    while (load_idx < total_k_elems) {\n      let s_local = load_idx / hd;\n      let d = load_idx % hd;\n      let s_global = tile_start + s_local;\n      let k_addr = s_global * kv_stride + kv_head_offset + d;\n      smem[s_local * hd + d] = f32(K[k_addr]);\n      load_idx += WG;\n    }\n    workgroupBarrier();\n\n    // ── Step 2: Parallel Q·K dot products ──\n    let pos_in_tile = tid / DOT_THREADS;\n    let dot_tid = tid % DOT_THREADS;\n\n    var my_partial: f32 = 0.0;\n    if (pos_in_tile < tile_len) {\n      let dims_per_thread = hd / DOT_THREADS;\n      let d_start = dot_tid * dims_per_thread;\n      let d_end = d_start + dims_per_thread;\n      let kv_row_base = pos_in_tile * hd;\n\n      var d = d_start;\n      while (d + 3u < d_end) {\n        let q_v = vec4f(Q[q_base + d], Q[q_base + d + 1u],\n                         Q[q_base + d + 2u], Q[q_base + d + 3u]);\n        let k_v = vec4f(smem[kv_row_base + d], smem[kv_row_base + d + 1u],\n                         smem[kv_row_base + d + 2u], smem[kv_row_base + d + 3u]);\n        my_partial += dot(q_v, k_v);\n        d += 4u;\n      }\n      while (d < d_end) {\n        my_partial += Q[q_base + d] * smem[kv_row_base + d];\n        d += 1u;\n      }\n    }\n\n    workgroupBarrier();\n    smem[tid] = my_partial;\n    workgroupBarrier();\n\n    // Two-phase: leader 0 reads smem[1..15], which are exactly the slots\n    // leaders 1..15 write — reduce into a local first, barrier (in uniform\n    // control flow), then write. Read+write in one block is a data race.\n    var group_total: f32 = 0.0;\n    if (dot_tid == 0u && pos_in_tile < tile_len) {\n      group_total = smem[tid];\n      for (var i: u32 = 1u; i < DOT_THREADS; i += 1u) {\n        group_total += smem[tid + i];\n      }\n    }\n    workgroupBarrier();\n    if (dot_tid == 0u && pos_in_tile < tile_len) {\n      smem[pos_in_tile] = group_total * scale;\n    }\n    workgroupBarrier();\n\n    // ── Step 3: Online softmax ──\n    if (tid < TILE_S) {\n      if (tid < tile_len) {\n        smem[TILE_S + tid] = smem[tid];\n      } else {\n        smem[TILE_S + tid] = -1e30;\n      }\n    }\n    workgroupBarrier();\n    if (tid < 8u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 8u]); }\n    workgroupBarrier();\n    if (tid < 4u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 4u]); }\n    workgroupBarrier();\n    if (tid < 2u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 2u]); }\n    workgroupBarrier();\n    if (tid < 1u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 1u]); }\n    workgroupBarrier();\n    let tile_max = smem[TILE_S];\n\n    let new_max = max(running_max, tile_max);\n    let old_correction = exp(max(running_max - new_max, -80.0));\n\n    if (tid < tile_len) {\n      smem[tid] = exp(max(smem[tid] - new_max, -80.0));\n    }\n    workgroupBarrier();\n\n    if (tid < TILE_S) {\n      if (tid < tile_len) {\n        smem[TILE_S + tid] = smem[tid];\n      } else {\n        smem[TILE_S + tid] = 0.0;\n      }\n    }\n    workgroupBarrier();\n    if (tid < 8u) { smem[TILE_S + tid] += smem[TILE_S + tid + 8u]; }\n    workgroupBarrier();\n    if (tid < 4u) { smem[TILE_S + tid] += smem[TILE_S + tid + 4u]; }\n    workgroupBarrier();\n    if (tid < 2u) { smem[TILE_S + tid] += smem[TILE_S + tid + 2u]; }\n    workgroupBarrier();\n    if (tid < 1u) { smem[TILE_S + tid] += smem[TILE_S + tid + 1u]; }\n    workgroupBarrier();\n    let tile_sum = smem[TILE_S];\n\n    // ── Save scores to registers before V overwrites smem ──\n    var w0: f32 = 0.0; var w1: f32 = 0.0; var w2: f32 = 0.0; var w3: f32 = 0.0;\n    var w4: f32 = 0.0; var w5: f32 = 0.0; var w6: f32 = 0.0; var w7: f32 = 0.0;\n    var w8: f32 = 0.0; var w9: f32 = 0.0; var w10: f32 = 0.0; var w11: f32 = 0.0;\n    var w12: f32 = 0.0; var w13: f32 = 0.0; var w14: f32 = 0.0; var w15: f32 = 0.0;\n    if (tid < hd) {\n      if (0u < tile_len) { w0 = smem[0]; }\n      if (1u < tile_len) { w1 = smem[1]; }\n      if (2u < tile_len) { w2 = smem[2]; }\n      if (3u < tile_len) { w3 = smem[3]; }\n      if (4u < tile_len) { w4 = smem[4]; }\n      if (5u < tile_len) { w5 = smem[5]; }\n      if (6u < tile_len) { w6 = smem[6]; }\n      if (7u < tile_len) { w7 = smem[7]; }\n      if (8u < tile_len) { w8 = smem[8]; }\n      if (9u < tile_len) { w9 = smem[9]; }\n      if (10u < tile_len) { w10 = smem[10]; }\n      if (11u < tile_len) { w11 = smem[11]; }\n      if (12u < tile_len) { w12 = smem[12]; }\n      if (13u < tile_len) { w13 = smem[13]; }\n      if (14u < tile_len) { w14 = smem[14]; }\n      if (15u < tile_len) { w15 = smem[15]; }\n    }\n    workgroupBarrier();\n\n    // ── Step 4: Cooperative V tile load (f16 → f32 conversion, full smem reuse) ──\n    let total_v_elems = tile_len * hd;\n    load_idx = tid;\n    while (load_idx < total_v_elems) {\n      let s_local = load_idx / hd;\n      let d = load_idx % hd;\n      let s_global = tile_start + s_local;\n      let v_addr = s_global * kv_stride + kv_head_offset + d;\n      smem[s_local * hd + d] = f32(V[v_addr]);\n      load_idx += WG;\n    }\n    workgroupBarrier();\n\n    // ── Step 5: Parallel V accumulation ──\n    if (tid < hd) {\n      out_acc = out_acc * old_correction;\n      if (0u < tile_len) { out_acc += w0 * smem[0u * hd + tid]; }\n      if (1u < tile_len) { out_acc += w1 * smem[1u * hd + tid]; }\n      if (2u < tile_len) { out_acc += w2 * smem[2u * hd + tid]; }\n      if (3u < tile_len) { out_acc += w3 * smem[3u * hd + tid]; }\n      if (4u < tile_len) { out_acc += w4 * smem[4u * hd + tid]; }\n      if (5u < tile_len) { out_acc += w5 * smem[5u * hd + tid]; }\n      if (6u < tile_len) { out_acc += w6 * smem[6u * hd + tid]; }\n      if (7u < tile_len) { out_acc += w7 * smem[7u * hd + tid]; }\n      if (8u < tile_len) { out_acc += w8 * smem[8u * hd + tid]; }\n      if (9u < tile_len) { out_acc += w9 * smem[9u * hd + tid]; }\n      if (10u < tile_len) { out_acc += w10 * smem[10u * hd + tid]; }\n      if (11u < tile_len) { out_acc += w11 * smem[11u * hd + tid]; }\n      if (12u < tile_len) { out_acc += w12 * smem[12u * hd + tid]; }\n      if (13u < tile_len) { out_acc += w13 * smem[13u * hd + tid]; }\n      if (14u < tile_len) { out_acc += w14 * smem[14u * hd + tid]; }\n      if (15u < tile_len) { out_acc += w15 * smem[15u * hd + tid]; }\n    }\n    if (d2 < hd) {\n      out_acc2 = out_acc2 * old_correction;\n      if (0u < tile_len) { out_acc2 += w0 * smem[0u * hd + d2]; }\n      if (1u < tile_len) { out_acc2 += w1 * smem[1u * hd + d2]; }\n      if (2u < tile_len) { out_acc2 += w2 * smem[2u * hd + d2]; }\n      if (3u < tile_len) { out_acc2 += w3 * smem[3u * hd + d2]; }\n      if (4u < tile_len) { out_acc2 += w4 * smem[4u * hd + d2]; }\n      if (5u < tile_len) { out_acc2 += w5 * smem[5u * hd + d2]; }\n      if (6u < tile_len) { out_acc2 += w6 * smem[6u * hd + d2]; }\n      if (7u < tile_len) { out_acc2 += w7 * smem[7u * hd + d2]; }\n      if (8u < tile_len) { out_acc2 += w8 * smem[8u * hd + d2]; }\n      if (9u < tile_len) { out_acc2 += w9 * smem[9u * hd + d2]; }\n      if (10u < tile_len) { out_acc2 += w10 * smem[10u * hd + d2]; }\n      if (11u < tile_len) { out_acc2 += w11 * smem[11u * hd + d2]; }\n      if (12u < tile_len) { out_acc2 += w12 * smem[12u * hd + d2]; }\n      if (13u < tile_len) { out_acc2 += w13 * smem[13u * hd + d2]; }\n      if (14u < tile_len) { out_acc2 += w14 * smem[14u * hd + d2]; }\n      if (15u < tile_len) { out_acc2 += w15 * smem[15u * hd + d2]; }\n    }\n\n    running_sum = running_sum * old_correction + tile_sum;\n    running_max = new_max;\n    workgroupBarrier();\n  }\n\n  // ── Write normalized output ──\n  var inv_sum: f32 = 0.0;\n  if (running_sum > 0.0) { inv_sum = 1.0 / running_sum; }\n  let out_base = q_pos * q_stride + q_head * hd;\n  if (tid < hd) {\n    output[out_base + tid] = out_acc * inv_sum;\n  }\n  if (d2 < hd) {\n    output[out_base + d2] = out_acc2 * inv_sum;\n  }\n}\n`;\n\nconst WGSL_ATTENTION_PACKED_F16 = `\\\n// Tiled online-softmax attention with packed f16 KV cache (Safari-safe).\n//\n// K/V stored as array<u32> — each u32 holds 2 f16 values via pack2x16float.\n// Does NOT require 'enable f16' — only uses u32 + f32 + unpack2x16float.\n// Same algorithm as native f16 attention, same memory savings.\n//\n// Safari/Metal compatibility:\n//   - Uses if/else instead of select() (Safari has select() bugs)\n//   - Clamps exp() args to -80 (Metal can return NaN for exp(-1e30))\n//   - No f16 types avoids WebKit WGSL compiler miscompilation\n//\n// Shared memory: 4096 f32 = 16384 bytes = exactly 16 KB (minimum WebGPU guarantee)\n\nconst TILE_S: u32 = 16u;\nconst WG: u32 = 256u;\nconst DOT_THREADS: u32 = 16u;\n\nstruct Params {\n  T: u32,\n  S: u32,\n  num_q_heads: u32,\n  num_kv_heads: u32,\n  head_dim: u32,\n  position_offset: u32,\n  // Softmax QK temperature (f32). Default 1/sqrt(head_dim); Gemma 4 sets 1.0.\n  attn_scale: f32,\n}\n\n@group(0) @binding(0) var<storage, read> Q: array<f32>;\n@group(0) @binding(1) var<storage, read> K: array<u32>;\n@group(0) @binding(2) var<storage, read> V: array<u32>;\n@group(0) @binding(3) var<storage, read_write> output: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\nvar<workgroup> smem: array<f32, 4096>;\n\n// Load a single f32 from packed u32 K cache at element-level index\nfn load_k(elem_idx: u32) -> f32 {\n  let pair = unpack2x16float(K[elem_idx >> 1u]);\n  if ((elem_idx & 1u) == 0u) {\n    return pair.x;\n  } else {\n    return pair.y;\n  }\n}\n\n// Load a single f32 from packed u32 V cache at element-level index\nfn load_v(elem_idx: u32) -> f32 {\n  let pair = unpack2x16float(V[elem_idx >> 1u]);\n  if ((elem_idx & 1u) == 0u) {\n    return pair.x;\n  } else {\n    return pair.y;\n  }\n}\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let q_pos = wid.x;\n  let q_head = wid.y;\n\n  if (q_pos >= params.T || q_head >= params.num_q_heads) { return; }\n\n  let tid = lid.x;\n  let heads_per_kv = params.num_q_heads / params.num_kv_heads;\n  let kv_head = q_head / heads_per_kv;\n  let scale = params.attn_scale;\n\n  let q_stride = params.num_q_heads * params.head_dim;\n  let kv_stride = params.num_kv_heads * params.head_dim;\n  let q_base = q_pos * q_stride + q_head * params.head_dim;\n  let kv_head_offset = kv_head * params.head_dim;\n  let causal_limit = q_pos + params.position_offset + 1u;\n  let hd = params.head_dim;\n  let S_eff = min(params.S, causal_limit);\n\n  var running_max: f32 = -1e30;\n  var running_sum: f32 = 0.0;\n  // Per-thread output accumulators; out_acc2 covers dim tid+WG so head_dim up to\n  // 2*WG (512) works with WG=256 threads. hd<=256 → out_acc2 unused (identical).\n  var out_acc: f32 = 0.0;\n  var out_acc2: f32 = 0.0;\n  let d2 = tid + WG;\n\n  // Tile capped so a K/V tile (tile_cap*hd f32) fits the 4096-f32 smem.\n  let tile_cap = min(TILE_S, 4096u / hd);\n  let num_tiles = (S_eff + tile_cap - 1u) / tile_cap;\n\n  for (var tile: u32 = 0u; tile < num_tiles; tile += 1u) {\n    let tile_start = tile * tile_cap;\n    let tile_end = min(tile_start + tile_cap, S_eff);\n    let tile_len = tile_end - tile_start;\n\n    // ── Step 1: Cooperative K tile load (packed u32 → f32 via unpack2x16float) ──\n    let total_k_elems = tile_len * hd;\n    var load_idx = tid;\n    while (load_idx < total_k_elems) {\n      let s_local = load_idx / hd;\n      let d = load_idx % hd;\n      let s_global = tile_start + s_local;\n      let k_addr = s_global * kv_stride + kv_head_offset + d;\n      smem[s_local * hd + d] = load_k(k_addr);\n      load_idx += WG;\n    }\n    workgroupBarrier();\n\n    // ── Step 2: Parallel Q·K dot products ──\n    let pos_in_tile = tid / DOT_THREADS;\n    let dot_tid = tid % DOT_THREADS;\n\n    var my_partial: f32 = 0.0;\n    if (pos_in_tile < tile_len) {\n      let dims_per_thread = hd / DOT_THREADS;\n      let d_start = dot_tid * dims_per_thread;\n      let d_end = d_start + dims_per_thread;\n      let kv_row_base = pos_in_tile * hd;\n\n      var d = d_start;\n      while (d + 3u < d_end) {\n        let q_v = vec4f(Q[q_base + d], Q[q_base + d + 1u],\n                         Q[q_base + d + 2u], Q[q_base + d + 3u]);\n        let k_v = vec4f(smem[kv_row_base + d], smem[kv_row_base + d + 1u],\n                         smem[kv_row_base + d + 2u], smem[kv_row_base + d + 3u]);\n        my_partial += dot(q_v, k_v);\n        d += 4u;\n      }\n      while (d < d_end) {\n        my_partial += Q[q_base + d] * smem[kv_row_base + d];\n        d += 1u;\n      }\n    }\n\n    workgroupBarrier();\n    smem[tid] = my_partial;\n    workgroupBarrier();\n\n    // Two-phase: leader 0 reads smem[1..15], which are exactly the slots\n    // leaders 1..15 write — reduce into a local first, barrier (in uniform\n    // control flow), then write. Read+write in one block is a data race.\n    var group_total: f32 = 0.0;\n    if (dot_tid == 0u && pos_in_tile < tile_len) {\n      group_total = smem[tid];\n      for (var i: u32 = 1u; i < DOT_THREADS; i += 1u) {\n        group_total += smem[tid + i];\n      }\n    }\n    workgroupBarrier();\n    if (dot_tid == 0u && pos_in_tile < tile_len) {\n      smem[pos_in_tile] = group_total * scale;\n    }\n    workgroupBarrier();\n\n    // ── Step 3: Online softmax ──\n    if (tid < TILE_S) {\n      if (tid < tile_len) {\n        smem[TILE_S + tid] = smem[tid];\n      } else {\n        smem[TILE_S + tid] = -1e30;\n      }\n    }\n    workgroupBarrier();\n    if (tid < 8u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 8u]); }\n    workgroupBarrier();\n    if (tid < 4u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 4u]); }\n    workgroupBarrier();\n    if (tid < 2u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 2u]); }\n    workgroupBarrier();\n    if (tid < 1u) { smem[TILE_S + tid] = max(smem[TILE_S + tid], smem[TILE_S + tid + 1u]); }\n    workgroupBarrier();\n    let tile_max = smem[TILE_S];\n\n    let new_max = max(running_max, tile_max);\n    let old_correction = exp(max(running_max - new_max, -80.0));\n\n    if (tid < tile_len) {\n      smem[tid] = exp(max(smem[tid] - new_max, -80.0));\n    }\n    workgroupBarrier();\n\n    if (tid < TILE_S) {\n      if (tid < tile_len) {\n        smem[TILE_S + tid] = smem[tid];\n      } else {\n        smem[TILE_S + tid] = 0.0;\n      }\n    }\n    workgroupBarrier();\n    if (tid < 8u) { smem[TILE_S + tid] += smem[TILE_S + tid + 8u]; }\n    workgroupBarrier();\n    if (tid < 4u) { smem[TILE_S + tid] += smem[TILE_S + tid + 4u]; }\n    workgroupBarrier();\n    if (tid < 2u) { smem[TILE_S + tid] += smem[TILE_S + tid + 2u]; }\n    workgroupBarrier();\n    if (tid < 1u) { smem[TILE_S + tid] += smem[TILE_S + tid + 1u]; }\n    workgroupBarrier();\n    let tile_sum = smem[TILE_S];\n\n    // ── Save scores to registers before V overwrites smem ──\n    var w0: f32 = 0.0; var w1: f32 = 0.0; var w2: f32 = 0.0; var w3: f32 = 0.0;\n    var w4: f32 = 0.0; var w5: f32 = 0.0; var w6: f32 = 0.0; var w7: f32 = 0.0;\n    var w8: f32 = 0.0; var w9: f32 = 0.0; var w10: f32 = 0.0; var w11: f32 = 0.0;\n    var w12: f32 = 0.0; var w13: f32 = 0.0; var w14: f32 = 0.0; var w15: f32 = 0.0;\n    if (tid < hd) {\n      if (0u < tile_len) { w0 = smem[0]; }\n      if (1u < tile_len) { w1 = smem[1]; }\n      if (2u < tile_len) { w2 = smem[2]; }\n      if (3u < tile_len) { w3 = smem[3]; }\n      if (4u < tile_len) { w4 = smem[4]; }\n      if (5u < tile_len) { w5 = smem[5]; }\n      if (6u < tile_len) { w6 = smem[6]; }\n      if (7u < tile_len) { w7 = smem[7]; }\n      if (8u < tile_len) { w8 = smem[8]; }\n      if (9u < tile_len) { w9 = smem[9]; }\n      if (10u < tile_len) { w10 = smem[10]; }\n      if (11u < tile_len) { w11 = smem[11]; }\n      if (12u < tile_len) { w12 = smem[12]; }\n      if (13u < tile_len) { w13 = smem[13]; }\n      if (14u < tile_len) { w14 = smem[14]; }\n      if (15u < tile_len) { w15 = smem[15]; }\n    }\n    workgroupBarrier();\n\n    // ── Step 4: Cooperative V tile load (packed u32 → f32 via unpack2x16float) ──\n    let total_v_elems = tile_len * hd;\n    load_idx = tid;\n    while (load_idx < total_v_elems) {\n      let s_local = load_idx / hd;\n      let d = load_idx % hd;\n      let s_global = tile_start + s_local;\n      let v_addr = s_global * kv_stride + kv_head_offset + d;\n      smem[s_local * hd + d] = load_v(v_addr);\n      load_idx += WG;\n    }\n    workgroupBarrier();\n\n    // ── Step 5: Parallel V accumulation ──\n    if (tid < hd) {\n      out_acc = out_acc * old_correction;\n      if (0u < tile_len) { out_acc += w0 * smem[0u * hd + tid]; }\n      if (1u < tile_len) { out_acc += w1 * smem[1u * hd + tid]; }\n      if (2u < tile_len) { out_acc += w2 * smem[2u * hd + tid]; }\n      if (3u < tile_len) { out_acc += w3 * smem[3u * hd + tid]; }\n      if (4u < tile_len) { out_acc += w4 * smem[4u * hd + tid]; }\n      if (5u < tile_len) { out_acc += w5 * smem[5u * hd + tid]; }\n      if (6u < tile_len) { out_acc += w6 * smem[6u * hd + tid]; }\n      if (7u < tile_len) { out_acc += w7 * smem[7u * hd + tid]; }\n      if (8u < tile_len) { out_acc += w8 * smem[8u * hd + tid]; }\n      if (9u < tile_len) { out_acc += w9 * smem[9u * hd + tid]; }\n      if (10u < tile_len) { out_acc += w10 * smem[10u * hd + tid]; }\n      if (11u < tile_len) { out_acc += w11 * smem[11u * hd + tid]; }\n      if (12u < tile_len) { out_acc += w12 * smem[12u * hd + tid]; }\n      if (13u < tile_len) { out_acc += w13 * smem[13u * hd + tid]; }\n      if (14u < tile_len) { out_acc += w14 * smem[14u * hd + tid]; }\n      if (15u < tile_len) { out_acc += w15 * smem[15u * hd + tid]; }\n    }\n    if (d2 < hd) {\n      out_acc2 = out_acc2 * old_correction;\n      if (0u < tile_len) { out_acc2 += w0 * smem[0u * hd + d2]; }\n      if (1u < tile_len) { out_acc2 += w1 * smem[1u * hd + d2]; }\n      if (2u < tile_len) { out_acc2 += w2 * smem[2u * hd + d2]; }\n      if (3u < tile_len) { out_acc2 += w3 * smem[3u * hd + d2]; }\n      if (4u < tile_len) { out_acc2 += w4 * smem[4u * hd + d2]; }\n      if (5u < tile_len) { out_acc2 += w5 * smem[5u * hd + d2]; }\n      if (6u < tile_len) { out_acc2 += w6 * smem[6u * hd + d2]; }\n      if (7u < tile_len) { out_acc2 += w7 * smem[7u * hd + d2]; }\n      if (8u < tile_len) { out_acc2 += w8 * smem[8u * hd + d2]; }\n      if (9u < tile_len) { out_acc2 += w9 * smem[9u * hd + d2]; }\n      if (10u < tile_len) { out_acc2 += w10 * smem[10u * hd + d2]; }\n      if (11u < tile_len) { out_acc2 += w11 * smem[11u * hd + d2]; }\n      if (12u < tile_len) { out_acc2 += w12 * smem[12u * hd + d2]; }\n      if (13u < tile_len) { out_acc2 += w13 * smem[13u * hd + d2]; }\n      if (14u < tile_len) { out_acc2 += w14 * smem[14u * hd + d2]; }\n      if (15u < tile_len) { out_acc2 += w15 * smem[15u * hd + d2]; }\n    }\n\n    running_sum = running_sum * old_correction + tile_sum;\n    running_max = new_max;\n    workgroupBarrier();\n  }\n\n  // ── Write normalized output ──\n  var inv_sum: f32 = 0.0;\n  if (running_sum > 0.0) { inv_sum = 1.0 / running_sum; }\n  let out_base = q_pos * q_stride + q_head * hd;\n  if (tid < hd) {\n    output[out_base + tid] = out_acc * inv_sum;\n  }\n  if (d2 < hd) {\n    output[out_base + d2] = out_acc2 * inv_sum;\n  }\n}\n`;\n\nconst WGSL_CONV_STATE_UPDATE = `\\\n// Update conv1d rolling state buffer with latest input timesteps.\n// For T >= state_size: copy last state_size timesteps from input.\n// For T < state_size: shift old state left by T, append T new values.\n\nstruct Params {\n  seq_len: u32,\n  channels: u32,\n  state_size: u32,\n  _pad: u32,\n}\n\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read_write> conv_state: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let c = gid.x;\n  if (c >= params.channels) { return; }\n\n  let ss = params.state_size;\n\n  if (params.seq_len >= ss) {\n    // Large batch (prefill): copy last state_size timesteps from input\n    for (var s: u32 = 0u; s < ss; s = s + 1u) {\n      let src_t = params.seq_len - ss + s;\n      conv_state[s * params.channels + c] = input[src_t * params.channels + c];\n    }\n  } else {\n    // Small batch (decode): shift old state left, append new values\n    // Read old state for this channel first to avoid read-write conflict\n    var old0: f32 = 0.0;\n    var old1: f32 = 0.0;\n    var old2: f32 = 0.0;\n    if (ss > 0u) { old0 = conv_state[0u * params.channels + c]; }\n    if (ss > 1u) { old1 = conv_state[1u * params.channels + c]; }\n    if (ss > 2u) { old2 = conv_state[2u * params.channels + c]; }\n\n    let shift = ss - params.seq_len;\n    for (var s: u32 = 0u; s < ss; s = s + 1u) {\n      var val: f32 = 0.0;\n      if (s < shift) {\n        // Read from shifted old state\n        let src_s = s + params.seq_len;\n        if (src_s == 0u) { val = old0; }\n        else if (src_s == 1u) { val = old1; }\n        else if (src_s == 2u) { val = old2; }\n      } else {\n        // Read from new input\n        let input_idx = s - shift;\n        val = input[input_idx * params.channels + c];\n      }\n      conv_state[s * params.channels + c] = val;\n    }\n  }\n}\n`;\n\n// PoolMatMul: pooled[Np, W] = poolW[Np, N] @ hidden[N, W].\n// The Gemma 4 ViT spatial average-pool, expressed as a host-built pooling matrix\n// (poolW[r, k] = 1/k² when patch k falls in pooled-cell r, else 0) times the\n// encoder output. One thread per output element accumulates over N. Deliberately\n// simple (no workgroup memory, no select(), no vec4 alignment assumption) — it\n// runs ONCE per image over tiny matrices (Np≤~512, N≤~4096, W=768), so a scalar\n// kernel is plenty fast and maximally mobile-safe.\nconst WGSL_POOL_MATMUL = `\\\nstruct Params {\n  Np: u32,  // pooled rows (output rows)\n  N: u32,   // patch count (contraction dim)\n  W: u32,   // hidden width (output cols)\n}\n\n@group(0) @binding(0) var<storage, read> poolW: array<f32>;   // [Np, N]\n@group(0) @binding(1) var<storage, read> hidden: array<f32>;  // [N, W]\n@group(0) @binding(2) var<storage, read_write> outp: array<f32>; // [Np, W]\n@group(0) @binding(3) var<storage, read> params: Params;\n\n@compute @workgroup_size(64)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  let total = params.Np * params.W;\n  if (idx >= total) { return; }\n  let r = idx / params.W;   // pooled row\n  let c = idx % params.W;   // hidden col\n  var acc: f32 = 0.0;\n  let prow = r * params.N;\n  for (var k: u32 = 0u; k < params.N; k = k + 1u) {\n    let w = poolW[prow + k];\n    if (w != 0.0) {\n      acc = acc + w * hidden[k * params.W + c];\n    }\n  }\n  outp[idx] = acc;\n}\n`;\n\n// ClippedMatMul: out = clamp(clamp(A, imin, imax) @ B^T, omin, omax).\n// Gemma 4 ViT Gemma4ClippableLinear: each projection clamps its input to a\n// calibrated [imin, imax] before the linear and its output to [omin, omax] after.\n// Same tiling as WGSL_MATMUL (4×2 register blocking, vec4 loads), with the input\n// clamp applied at tile load and the output clamp at store. Mobile-safe: clamp =\n// min(max(x, lo), hi), no select(); no `enable f16`. The four clip scalars are\n// passed as f32 bit patterns in the uniform.\nconst WGSL_CLIPPED_MATMUL = `\\\nconst MT: u32 = 64u;\nconst NT: u32 = 32u;\nconst KT: u32 = 16u;\nconst KT4: u32 = 4u;\n\nstruct Params {\n  M: u32,\n  K: u32,\n  N: u32,\n  imin: f32,\n  imax: f32,\n  omin: f32,\n  omax: f32,\n}\n\n@group(0) @binding(0) var<storage, read> A: array<vec4f>;\n@group(0) @binding(1) var<storage, read> B: array<vec4f>;\n@group(0) @binding(2) var<storage, read_write> C: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\nvar<workgroup> tileA: array<f32, 1024>;\nvar<workgroup> tileB: array<f32, 512>;\n\nfn clip4(v: vec4f, lo: f32, hi: f32) -> vec4f {\n  return min(max(v, vec4f(lo)), vec4f(hi));\n}\n\n@compute @workgroup_size(16, 16)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let lr = lid.y;\n  let lc = lid.x;\n  let mBase = wid.y * MT;\n  let nBase = wid.x * NT;\n  let row0 = mBase + lr;\n  let col0 = nBase + lc;\n\n  var acc: array<f32, 8>;\n  for (var i = 0u; i < 8u; i = i + 1u) { acc[i] = 0.0; }\n\n  let numTiles = (params.K + KT - 1u) / KT;\n  let tid = lr * 16u + lc;\n  let K4 = params.K / 4u;\n\n  for (var t: u32 = 0u; t < numTiles; t = t + 1u) {\n    let kv0 = t * KT4;\n    {\n      let ar = tid / KT4;\n      let av4 = tid % KT4;\n      let gRow = mBase + ar;\n      var v = vec4f(0.0, 0.0, 0.0, 0.0);\n      if (gRow < params.M && kv0 + av4 < K4) {\n        v = clip4(A[gRow * K4 + kv0 + av4], params.imin, params.imax);\n      }\n      let base = ar * KT + av4 * 4u;\n      tileA[base] = v.x;\n      tileA[base + 1u] = v.y;\n      tileA[base + 2u] = v.z;\n      tileA[base + 3u] = v.w;\n    }\n    if (tid < 128u) {\n      let bc = tid / KT4;\n      let bv4 = tid % KT4;\n      let gCol = nBase + bc;\n      var v = vec4f(0.0, 0.0, 0.0, 0.0);\n      if (gCol < params.N && kv0 + bv4 < K4) {\n        v = B[gCol * K4 + kv0 + bv4];\n      }\n      let krow = bv4 * 4u;\n      tileB[(krow + 0u) * NT + bc] = v.x;\n      tileB[(krow + 1u) * NT + bc] = v.y;\n      tileB[(krow + 2u) * NT + bc] = v.z;\n      tileB[(krow + 3u) * NT + bc] = v.w;\n    }\n\n    workgroupBarrier();\n\n    for (var k: u32 = 0u; k < KT; k = k + 1u) {\n      let b0 = tileB[k * NT + lc];\n      let b1 = tileB[k * NT + lc + 16u];\n      let a0 = tileA[lr * KT + k];\n      let a1 = tileA[(lr + 16u) * KT + k];\n      let a2 = tileA[(lr + 32u) * KT + k];\n      let a3 = tileA[(lr + 48u) * KT + k];\n      acc[0] += a0 * b0;\n      acc[1] += a0 * b1;\n      acc[2] += a1 * b0;\n      acc[3] += a1 * b1;\n      acc[4] += a2 * b0;\n      acc[5] += a2 * b1;\n      acc[6] += a3 * b0;\n      acc[7] += a3 * b1;\n    }\n\n    workgroupBarrier();\n  }\n\n  for (var rr = 0u; rr < 4u; rr = rr + 1u) {\n    let gRow = row0 + rr * 16u;\n    if (gRow < params.M) {\n      if (col0 < params.N) {\n        C[gRow * params.N + col0] = min(max(acc[rr * 2u], params.omin), params.omax);\n      }\n      if (col0 + 16u < params.N) {\n        C[gRow * params.N + col0 + 16u] = min(max(acc[rr * 2u + 1u], params.omin), params.omax);\n      }\n    }\n  }\n}\n`;\n\n// ── Kernel specs ────────────────────────────────────────────────────────\n\n/**\n * Resolve the total element count of the first output tensor.\n * Used by elementwise ops whose dispatch is based on output size.\n */\nfunction outputElementCount(op: OpNode, resolvedShapes: Record<string, number[]>): number {\n  const outShape = resolvedShapes[op.outputs[0]];\n  let count = 1;\n  for (const d of outShape) count *= d;\n  return count;\n}\n\n// ── Add ──\n\nconst addSpec: KernelSpec = {\n  shaderCode: WGSL_ADD,\n  entryPoint: \"main\",\n\n  // @binding(0) a: storage read\n  // @binding(1) b: storage read\n  // @binding(2) output: storage read_write\n  // @binding(3) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    // Params { count: u32 }\n    return buildUniformBuffer([count]);\n  },\n};\n\n// ── Mul ──\n\nconst mulSpec: KernelSpec = {\n  shaderCode: WGSL_MUL,\n  entryPoint: \"main\",\n\n  // @binding(0) a: storage read\n  // @binding(1) b: storage read\n  // @binding(2) output: storage read_write\n  // @binding(3) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    // Params { count: u32 }\n    return buildUniformBuffer([count]);\n  },\n};\n\n// ── SliceLastRow ──\n\nconst sliceLastRowSpec: KernelSpec = {\n  shaderCode: WGSL_SLICE_LAST_ROW,\n  entryPoint: \"main\",\n\n  // @binding(0) src: storage read\n  // @binding(1) output: storage read_write\n  // @binding(2) params: uniform\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n\n  getDispatchSize(op) {\n    const width = op.attributes.width as number;\n    return [cdiv(width, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const width = op.attributes.width as number;\n    const T = resolvedShapes[op.inputs[0]]?.[0] ?? 1;\n    // Params { width: u32, last_row_offset: u32 }\n    return buildUniformBuffer([width, (T - 1) * width]);\n  },\n};\n\n// ── MeanPool (mean over T tokens → [1, width], bidirectional embeddings) ──\n\nconst meanPoolSpec: KernelSpec = {\n  shaderCode: WGSL_MEAN_POOL,\n  entryPoint: \"main\",\n\n  // @binding(0) src: storage read\n  // @binding(1) output: storage read_write\n  // @binding(2) params: uniform\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n\n  getDispatchSize(op) {\n    const width = op.attributes.width as number;\n    return [cdiv(width, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const width = op.attributes.width as number;\n    // seq_len = rows of the input activation (T).\n    const T = resolvedShapes[op.inputs[0]]?.[0] ?? 1;\n    // Params { seq_len: u32, width: u32 }\n    return buildUniformBuffer([T, width]);\n  },\n};\n\n// ── Scale (multiply tensor by a scalar constant) ──\n\nconst scaleSpec: KernelSpec = {\n  shaderCode: WGSL_SCALE,\n  entryPoint: \"main\",\n\n  // @binding(0) input: storage read\n  // @binding(1) output: storage read_write\n  // @binding(2) params: uniform\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    const scale = op.attributes.scale as number;\n    // Params { count: u32, scale_bits: u32 }\n    return buildUniformBuffer([count, f32BitsToU32(scale)]);\n  },\n};\n\n// ── Softcap (Gemma final logit softcapping) ──\n\nconst softcapSpec: KernelSpec = {\n  shaderCode: WGSL_SOFTCAP,\n  entryPoint: \"main\",\n\n  // @binding(0) input: storage read\n  // @binding(1) output: storage read_write\n  // @binding(2) params: uniform\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    const cap = op.attributes.cap as number;\n    // Params { count: u32, cap_bits: u32 }\n    return buildUniformBuffer([count, f32BitsToU32(cap)]);\n  },\n};\n\n// ── L2Norm (row-wise L2 normalization, used for embeddings) ──\n\nconst l2NormSpec: KernelSpec = {\n  shaderCode: WGSL_L2NORM,\n  entryPoint: \"main\",\n\n  // @binding(0) src: storage read\n  // @binding(1) output: storage read_write\n  // @binding(2) params: uniform\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n\n  getDispatchSize(op, resolvedShapes) {\n    // One workgroup per row.\n    const rows = resolvedShapes[op.outputs[0]]?.[0] ?? 1;\n    return [rows, 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const width = op.attributes.width as number;\n    const rows = resolvedShapes[op.outputs[0]]?.[0] ?? 1;\n    // Params { rows: u32, width: u32 }\n    return buildUniformBuffer([rows, width]);\n  },\n};\n\n// ── SwiGLU (fused SiLU + Mul) ──\n\nconst swigluSpec: KernelSpec = {\n  shaderCode: WGSL_SWIGLU,\n  entryPoint: \"main\",\n\n  // @binding(0) gate: storage read\n  // @binding(1) up: storage read\n  // @binding(2) output: storage read_write\n  // @binding(3) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return buildUniformBuffer([count]);\n  },\n};\n\n// ── ResidualRMSNorm (fused Add + RMSNorm) ──\n\nconst residualRmsnormSpec: KernelSpec = {\n  shaderCode: WGSL_RESIDUAL_RMSNORM,\n  entryPoint: \"main\",\n\n  // @binding(0) a: storage read\n  // @binding(1) b: storage read\n  // @binding(2) weight: storage read\n  // @binding(3) sum_output: storage read_write\n  // @binding(4) norm_output: storage read_write\n  // @binding(5) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const hidden = op.attributes.hidden_size as number;\n    const seqLen = resolveSeqLen(op, resolvedShapes, hidden);\n    return [seqLen, 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const hidden = op.attributes.hidden_size as number;\n    const seqLen = resolveSeqLen(op, resolvedShapes, hidden);\n    const eps = (op.attributes.eps as number) ?? 1e-6;\n    return buildUniformBuffer([seqLen, hidden, f32BitsToU32(eps), 0]);\n  },\n};\n\n// ── SiLU ──\n\nconst siluSpec: KernelSpec = {\n  shaderCode: WGSL_SILU,\n  entryPoint: \"main\",\n\n  // @binding(0) input: storage read\n  // @binding(1) output: storage read_write\n  // @binding(2) params: uniform\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    // Params { count: u32 }\n    return buildUniformBuffer([count]);\n  },\n};\n\n// ── GELU ──\n\nconst geluSpec: KernelSpec = {\n  shaderCode: WGSL_GELU,\n  entryPoint: \"main\",\n\n  // @binding(0) input: storage read\n  // @binding(1) output: storage read_write\n  // @binding(2) params: uniform\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    // Params { count: u32 }\n    return buildUniformBuffer([count]);\n  },\n};\n\n// ── GELU (exact erf form — ViT merger) ──\n\nconst geluErfSpec: KernelSpec = {\n  shaderCode: WGSL_GELU_ERF,\n  entryPoint: \"main\",\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n  buildParams(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return buildUniformBuffer([count]);\n  },\n};\n\n// ── AddBias (row-broadcast bias add) ──\n\nconst addBiasSpec: KernelSpec = {\n  shaderCode: WGSL_ADD_BIAS,\n  entryPoint: \"main\",\n\n  // @binding(0) input, (1) bias, (2) output, (3) params\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n  buildParams(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    const width = op.attributes.width as number;\n    return buildUniformBuffer([count, width]);\n  },\n};\n\n// ── SliceCols (extract column range) ──\n\nconst sliceColsSpec: KernelSpec = {\n  shaderCode: WGSL_SLICE_COLS,\n  entryPoint: \"main\",\n\n  // @binding(0) input, (1) output, (2) params\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n  buildParams(op, resolvedShapes) {\n    const in_width = op.attributes.in_width as number;\n    const out_width = op.attributes.out_width as number;\n    const col_offset = op.attributes.col_offset as number;\n    const outShape = resolvedShapes[op.outputs[0]];\n    const rows = outShape[0];\n    return buildUniformBuffer([rows, in_width, out_width, col_offset]);\n  },\n};\n\n// ── MulCols (multiply two column ranges of one tensor) ──\n\nconst mulColsSpec: KernelSpec = {\n  shaderCode: WGSL_MUL_COLS,\n  entryPoint: \"main\",\n\n  // @binding(0) src, (1) output, (2) params\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n  buildParams(op, resolvedShapes) {\n    const in_width = op.attributes.in_width as number;\n    const out_width = op.attributes.out_width as number;\n    const off_a = op.attributes.off_a as number;\n    const off_b = op.attributes.off_b as number;\n    const outShape = resolvedShapes[op.outputs[0]];\n    const rows = outShape[0];\n    return buildUniformBuffer([rows, in_width, out_width, off_a, off_b, 0, 0, 0]);\n  },\n};\n\n// ── ApplyRotaryEmb (precomputed cos/sin, rotate_half) ──\n\nconst applyRotarySpec: KernelSpec = {\n  shaderCode: WGSL_APPLY_ROTARY,\n  entryPoint: \"main\",\n\n  // @binding(0) x, (1) cos, (2) sin, (3) output, (4) params\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n  buildParams(op, resolvedShapes) {\n    const num_heads = op.attributes.num_heads as number;\n    const head_dim = op.attributes.head_dim as number;\n    const outShape = resolvedShapes[op.outputs[0]];\n    const total = outShape.reduce((a, b) => a * b, 1);\n    const seq_len = total / (num_heads * head_dim);\n    return buildUniformBuffer([seq_len, num_heads, head_dim, 0]);\n  },\n};\n\n// ── MatMul ──\n\nconst matmulSpec: KernelSpec = {\n  shaderCode: WGSL_MATMUL,\n  entryPoint: \"main\",\n\n  // @binding(0) A: storage read\n  // @binding(1) B: storage read\n  // @binding(2) C: storage read_write\n  // @binding(3) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    // C[M,N] = A[M,K] * B[K,N]\n    // 16x16 workgroup, 4x2 register blocking -> each workgroup covers a\n    // 64-row x 32-col output tile. Dispatch X covers N, Y covers M.\n    const N = op.attributes.N as number;\n    const mTensor = op.attributes.M_tensor as string | undefined;\n    let M: number;\n    if (mTensor && resolvedShapes[mTensor]) {\n      M = resolvedShapes[mTensor][0];\n    } else {\n      M = op.attributes.M as number;\n    }\n    return [cdiv(N, 32), cdiv(M, 64), 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    // Params { M: u32, K: u32, N: u32 }\n    const K = op.attributes.K as number;\n    const N = op.attributes.N as number;\n    // M is dynamic — derive from referenced tensor or static attribute\n    const mTensor = op.attributes.M_tensor as string | undefined;\n    let M: number;\n    if (mTensor && resolvedShapes[mTensor]) {\n      M = resolvedShapes[mTensor][0];\n    } else {\n      M = op.attributes.M as number;\n    }\n    return buildUniformBuffer([M, K, N]);\n  },\n};\n\n// ── MatMulBias (fused MatMul + bias) ──\n\nconst matmulBiasSpec: KernelSpec = {\n  shaderCode: WGSL_MATMUL_BIAS,\n  entryPoint: \"main\",\n\n  // @binding(0) A, (1) B, (2) bias, (3) C, (4) params\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const N = op.attributes.N as number;\n    const mTensor = op.attributes.M_tensor as string | undefined;\n    let M: number;\n    if (mTensor && resolvedShapes[mTensor]) {\n      M = resolvedShapes[mTensor][0];\n    } else {\n      M = op.attributes.M as number;\n    }\n    return [cdiv(N, 32), cdiv(M, 64), 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const K = op.attributes.K as number;\n    const N = op.attributes.N as number;\n    const mTensor = op.attributes.M_tensor as string | undefined;\n    let M: number;\n    if (mTensor && resolvedShapes[mTensor]) {\n      M = resolvedShapes[mTensor][0];\n    } else {\n      M = op.attributes.M as number;\n    }\n    return buildUniformBuffer([M, K, N]);\n  },\n};\n\n// ── PoolMatMul (Gemma 4 ViT spatial average-pool as host matrix · encoder out) ──\n\nconst poolMatMulSpec: KernelSpec = {\n  shaderCode: WGSL_POOL_MATMUL,\n  entryPoint: \"main\",\n\n  // @binding(0) poolW [Np,N], (1) hidden [N,W], (2) out [Np,W], (3) params\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const out = resolvedShapes[op.outputs[0]]; // [Np, W]\n    const Np = out[0];\n    const W = out[1];\n    return [cdiv(Np * W, 64), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const out = resolvedShapes[op.outputs[0]]; // [Np, W]\n    const Np = out[0];\n    const W = out[1];\n    // N = contraction dim = rows of the hidden tensor (op.inputs[1]).\n    const N = resolvedShapes[op.inputs[1]]?.[0] ?? (op.attributes.hidden_size as number);\n    // Params { Np: u32, N: u32, W: u32 }\n    return buildUniformBuffer([Np, N, W]);\n  },\n};\n\n// ── ClippedMatMul (Gemma 4 ViT Gemma4ClippableLinear) ──\n\nconst clippedMatMulSpec: KernelSpec = {\n  shaderCode: WGSL_CLIPPED_MATMUL,\n  entryPoint: \"main\",\n\n  // @binding(0) A, (1) B, (2) C, (3) params\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const N = op.attributes.N as number;\n    const mTensor = op.attributes.M_tensor as string | undefined;\n    const M =\n      mTensor && resolvedShapes[mTensor] ? resolvedShapes[mTensor][0] : (op.attributes.M as number);\n    return [cdiv(N, 32), cdiv(M, 64), 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const K = op.attributes.K as number;\n    const N = op.attributes.N as number;\n    const mTensor = op.attributes.M_tensor as string | undefined;\n    const M =\n      mTensor && resolvedShapes[mTensor] ? resolvedShapes[mTensor][0] : (op.attributes.M as number);\n    // Clip scalars default to ±inf (identity) until the loader patches them.\n    const imin = (op.attributes.imin as number) ?? Number.NEGATIVE_INFINITY;\n    const imax = (op.attributes.imax as number) ?? Number.POSITIVE_INFINITY;\n    const omin = (op.attributes.omin as number) ?? Number.NEGATIVE_INFINITY;\n    const omax = (op.attributes.omax as number) ?? Number.POSITIVE_INFINITY;\n    // Params { M, K, N, imin, imax, omin, omax } — first 3 u32, rest f32 bits.\n    return buildUniformBuffer([\n      M,\n      K,\n      N,\n      f32BitsToU32(imin),\n      f32BitsToU32(imax),\n      f32BitsToU32(omin),\n      f32BitsToU32(omax),\n    ]);\n  },\n};\n\n// Mixed-precision (f16 multiply, f32 accumulate) MatMulBias for ViT linear layers.\n// Same bindings/dispatch/params as matmulBiasSpec; binding(1) B is an f16 weight.\n// The VisionExecutor swaps this in when the device has shader-f16.\nexport const MATMUL_BIAS_F16C_SPEC: KernelSpec = {\n  ...matmulBiasSpec,\n  shaderCode: WGSL_MATMUL_BIAS_F16MIX,\n};\n\n// ── MatMulInt4 ──\n\nconst matmulInt4Spec: KernelSpec = {\n  shaderCode: WGSL_MATMUL_INT4,\n  entryPoint: \"main\",\n\n  // @binding(0) A: storage read\n  // @binding(1) B_q: storage read\n  // @binding(2) scales: storage read\n  // @binding(3) zeros: storage read\n  // @binding(4) C: storage read_write\n  // @binding(5) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    // Workgroup size is (16, 16), dispatch X covers N, Y covers M\n    const N = op.attributes.N as number;\n    const mTensor = op.attributes.M_tensor as string | undefined;\n    let M: number;\n    if (mTensor && resolvedShapes[mTensor]) {\n      M = resolvedShapes[mTensor][0];\n    } else {\n      M = op.attributes.M as number;\n    }\n    return [cdiv(N, 16), cdiv(M, 16), 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    // Params { M: u32, K: u32, N: u32, group_size: u32 }\n    const K = op.attributes.K as number;\n    const N = op.attributes.N as number;\n    const group_size = op.attributes.group_size as number;\n    const mTensor = op.attributes.M_tensor as string | undefined;\n    let M: number;\n    if (mTensor && resolvedShapes[mTensor]) {\n      M = resolvedShapes[mTensor][0];\n    } else {\n      M = op.attributes.M as number;\n    }\n    return buildUniformBuffer([M, K, N, group_size]);\n  },\n};\n\n// ── MatVec (F32, M=1 decode) ──\n\nexport const MATVEC_SPEC: KernelSpec = {\n  shaderCode: WGSL_MATVEC,\n  entryPoint: \"main\",\n\n  // Same binding layout as MatMul: A, B, C, params\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op) {\n    const N = op.attributes.N as number;\n    const N_TILE = 8;\n    return [cdiv(N, N_TILE), 1, 1];\n  },\n\n  buildParams(op) {\n    const K = op.attributes.K as number;\n    const N = op.attributes.N as number;\n    return buildUniformBuffer([K, N]);\n  },\n};\n\n// ── MatVecInt4 (K-parallel INT4, M=1 decode) ──\n\nexport const MATVEC_INT4_SPEC: KernelSpec = {\n  shaderCode: WGSL_MATVEC_INT4,\n  entryPoint: \"main\",\n\n  // Same binding layout as MatMulInt4: A, B_q, scales, zeros, C, params\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op) {\n    const N = op.attributes.N as number;\n    // Must match WGSL N_TILE = workgroup_size / K_THREADS = 256 / 16 = 16\n    const N_TILE = 16;\n    return [cdiv(N, N_TILE), 1, 1];\n  },\n\n  buildParams(op) {\n    const K = op.attributes.K as number;\n    const N = op.attributes.N as number;\n    const group_size = op.attributes.group_size as number;\n    return buildUniformBuffer([K, N, group_size, 0]);\n  },\n};\n\n// ── Gated MatVecInt4 (fused attn sigmoid-gate + INT4 o_proj, M=1 decode) ──\n\nexport const GATED_MATVEC_INT4_SPEC: KernelSpec = {\n  shaderCode: WGSL_GATED_MATVEC_INT4,\n  entryPoint: \"main\",\n\n  // @binding(0) attn, (1) gate, (2) B_q, (3) scales, (4) zeros, (5) C, (6) params\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op) {\n    const N = op.attributes.N as number;\n    const N_TILE = 8;\n    return [cdiv(N, N_TILE), 1, 1];\n  },\n\n  buildParams(op) {\n    const K = op.attributes.K as number;\n    const N = op.attributes.N as number;\n    const group_size = op.attributes.group_size as number;\n    return buildUniformBuffer([K, N, group_size, 0]);\n  },\n};\n\n// ── SwiGLU-gated MatVecInt4 (fused SwiGLU + INT4 projection, M=1 decode) ──\n\nexport const SWIGLU_GATED_MATVEC_INT4_SPEC: KernelSpec = {\n  ...GATED_MATVEC_INT4_SPEC,\n  shaderCode: WGSL_SWIGLU_GATED_MATVEC_INT4,\n};\n\n// ── MatVec subgroups variants (desktop-only fast path) ──\n//\n// Same binding layout, dispatch size, and params as the portable MatVec specs —\n// only the shader differs (subgroupAdd reduction). The executor selects these\n// when the device exposes the \"subgroups\" feature and is NOT WebKit. Kept as\n// distinct specs so the portable kernels remain the universal fallback.\n\nexport const MATVEC_SUBGROUPS_SPEC: KernelSpec = {\n  ...MATVEC_SPEC,\n  shaderCode: WGSL_MATVEC_SUBGROUPS,\n};\n\nexport const MATVEC_INT4_SUBGROUPS_SPEC: KernelSpec = {\n  ...MATVEC_INT4_SPEC,\n  shaderCode: WGSL_MATVEC_INT4_SUBGROUPS,\n};\n\n// ── SwiGLU MatVec (fused gate+up+SwiGLU, M=1 decode) ──\n\nexport const SWIGLU_MATVEC_SPEC: KernelSpec = {\n  shaderCode: WGSL_SWIGLU_MATVEC,\n  entryPoint: \"main\",\n\n  // @binding(0) A: storage read\n  // @binding(1) B_gate: storage read\n  // @binding(2) B_up: storage read\n  // @binding(3) output: storage read_write\n  // @binding(4) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op) {\n    const N = op.attributes.N as number;\n    const N_TILE = 8;\n    return [cdiv(N, N_TILE), 1, 1];\n  },\n\n  buildParams(op) {\n    const K = op.attributes.K as number;\n    const N = op.attributes.N as number;\n    return buildUniformBuffer([K, N]);\n  },\n};\n\n// ── SwiGLU MatVecInt4 (fused gate+up+SwiGLU INT4, M=1 decode) ──\n\nexport const SWIGLU_MATVEC_INT4_SPEC: KernelSpec = {\n  shaderCode: WGSL_SWIGLU_MATVEC_INT4,\n  entryPoint: \"main\",\n\n  // @binding(0) A: storage read\n  // @binding(1) B_gate_q: storage read\n  // @binding(2) B_gate_scales: storage read\n  // @binding(3) B_gate_zeros: storage read\n  // @binding(4) B_up_q: storage read\n  // @binding(5) B_up_scales: storage read\n  // @binding(6) B_up_zeros: storage read\n  // @binding(7) output: storage read_write\n  // @binding(8) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op) {\n    const N = op.attributes.N as number;\n    const N_TILE = 8;\n    return [cdiv(N, N_TILE), 1, 1];\n  },\n\n  buildParams(op) {\n    const K = op.attributes.K as number;\n    const N = op.attributes.N as number;\n    const group_size = op.attributes.group_size as number;\n    return buildUniformBuffer([K, N, group_size, 0]);\n  },\n};\n\n// ── Dual MatVecInt4 (two INT4 projections sharing input A, M=1 decode) ──\n\nexport const DUAL_MATVEC_INT4_SPEC: KernelSpec = {\n  shaderCode: WGSL_DUAL_MATVEC_INT4,\n  entryPoint: \"main\",\n\n  // @binding(0) A\n  // @binding(1..3) B0_q, B0_scales, B0_zeros\n  // @binding(4..6) B1_q, B1_scales, B1_zeros\n  // @binding(7) out0, @binding(8) out1, @binding(9) params\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op) {\n    const N = op.attributes.N as number;\n    const N_TILE = 8;\n    return [cdiv(N, N_TILE), 1, 1];\n  },\n\n  buildParams(op) {\n    const K = op.attributes.K as number;\n    const N = op.attributes.N as number;\n    const group_size = op.attributes.group_size as number;\n    return buildUniformBuffer([K, N, group_size, 0]);\n  },\n};\n\n// ── Argmax (GPU-side, single workgroup) ──\n\nexport const ARGMAX_SPEC: KernelSpec = {\n  shaderCode: WGSL_ARGMAX,\n  entryPoint: \"main\",\n\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n\n  getDispatchSize() {\n    return [1, 1, 1]; // Single workgroup\n  },\n\n  buildParams(_op, _resolvedShapes) {\n    // Params built externally by executor (count + offset)\n    return buildUniformBuffer([0, 0]);\n  },\n};\n\n// ── RMSNorm ──\n\nconst rmsnormSpec: KernelSpec = {\n  shaderCode: WGSL_RMSNORM,\n  entryPoint: \"main\",\n\n  // @binding(0) input: storage read\n  // @binding(1) weight: storage read\n  // @binding(2) output: storage read_write\n  // @binding(3) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    // One workgroup per row (token)\n    const hidden_size = op.attributes.hidden_size as number;\n    const ref = op.attributes.seq_len_tensor as string | undefined;\n    let seq_len: number;\n    if (ref && resolvedShapes[ref]) {\n      const shape = resolvedShapes[ref];\n      const total = shape.reduce((a: number, b: number) => a * b, 1);\n      seq_len = total / hidden_size;\n    } else {\n      seq_len = op.attributes.seq_len as number;\n    }\n    return [seq_len, 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    // Params { seq_len: u32, hidden_size: u32, eps_bits: u32, _pad: u32 }\n    const hidden_size = op.attributes.hidden_size as number;\n    const eps = op.attributes.eps as number;\n    const ref = op.attributes.seq_len_tensor as string | undefined;\n    let seq_len: number;\n    if (ref && resolvedShapes[ref]) {\n      const shape = resolvedShapes[ref];\n      const total = shape.reduce((a: number, b: number) => a * b, 1);\n      seq_len = total / hidden_size;\n    } else {\n      seq_len = op.attributes.seq_len as number;\n    }\n    return buildUniformBuffer([\n      seq_len,\n      hidden_size,\n      f32BitsToU32(eps),\n      0, // _pad\n    ]);\n  },\n};\n\n// ── Dual RMSNorm (two per-row RMSNorms, one dispatch) ──\n\nfunction dualRmsnormRows(\n  op: OpNode,\n  resolvedShapes: Record<string, number[]>,\n  refKey: string,\n  fallbackKey: string,\n): number {\n  const hidden_size = op.attributes.hidden_size as number;\n  const ref = op.attributes[refKey] as string | undefined;\n  if (ref && resolvedShapes[ref]) {\n    const total = resolvedShapes[ref].reduce((a: number, b: number) => a * b, 1);\n    return total / hidden_size;\n  }\n  return op.attributes[fallbackKey] as number;\n}\n\nexport const DUAL_RMSNORM_SPEC: KernelSpec = {\n  shaderCode: WGSL_DUAL_RMSNORM,\n  entryPoint: \"main\",\n\n  // input0, weight0, input1, weight1, output0, output1, params\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const rows0 = dualRmsnormRows(op, resolvedShapes, \"seq_len_tensor0\", \"seq_len0\");\n    const rows1 = dualRmsnormRows(op, resolvedShapes, \"seq_len_tensor1\", \"seq_len1\");\n    return [rows0 + rows1, 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const rows0 = dualRmsnormRows(op, resolvedShapes, \"seq_len_tensor0\", \"seq_len0\");\n    const rows1 = dualRmsnormRows(op, resolvedShapes, \"seq_len_tensor1\", \"seq_len1\");\n    const hidden_size = op.attributes.hidden_size as number;\n    const eps = op.attributes.eps as number;\n    return buildUniformBuffer([rows0, rows1, hidden_size, f32BitsToU32(eps)]);\n  },\n};\n\n// ── LayerNorm ──\n\nconst layernormSpec: KernelSpec = {\n  shaderCode: WGSL_LAYERNORM,\n  entryPoint: \"main\",\n\n  // @binding(0) input: storage read\n  // @binding(1) weight: storage read\n  // @binding(2) bias: storage read\n  // @binding(3) output: storage read_write\n  // @binding(4) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    // One workgroup per row (token)\n    const hidden_size = op.attributes.hidden_size as number;\n    const ref = op.attributes.seq_len_tensor as string | undefined;\n    let seq_len: number;\n    if (ref && resolvedShapes[ref]) {\n      const shape = resolvedShapes[ref];\n      const total = shape.reduce((a: number, b: number) => a * b, 1);\n      seq_len = total / hidden_size;\n    } else {\n      seq_len = op.attributes.seq_len as number;\n    }\n    return [seq_len, 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    // Params { seq_len: u32, hidden_size: u32, eps_bits: u32, _pad: u32 }\n    const hidden_size = op.attributes.hidden_size as number;\n    const eps = op.attributes.eps as number;\n    const ref = op.attributes.seq_len_tensor as string | undefined;\n    let seq_len: number;\n    if (ref && resolvedShapes[ref]) {\n      const shape = resolvedShapes[ref];\n      const total = shape.reduce((a: number, b: number) => a * b, 1);\n      seq_len = total / hidden_size;\n    } else {\n      seq_len = op.attributes.seq_len as number;\n    }\n    return buildUniformBuffer([\n      seq_len,\n      hidden_size,\n      f32BitsToU32(eps),\n      0, // _pad\n    ]);\n  },\n};\n\n// ── Embedding ──\n\nconst embeddingSpec: KernelSpec = {\n  shaderCode: WGSL_EMBEDDING,\n  entryPoint: \"main\",\n\n  // @binding(0) input_ids: storage read\n  // @binding(1) weight: storage read\n  // @binding(2) output: storage read_write\n  // @binding(3) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const hidden_size = op.attributes.hidden_size as number;\n    const outShape = resolvedShapes[op.outputs[0]];\n    const seq_len = outShape ? outShape[0] : (op.attributes.seq_len as number);\n    return [cdiv(seq_len * hidden_size, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    // Params { seq_len: u32, hidden_size: u32 }\n    const hidden_size = op.attributes.hidden_size as number;\n    const outShape = resolvedShapes[op.outputs[0]];\n    const seq_len = outShape ? outShape[0] : (op.attributes.seq_len as number);\n    return buildUniformBuffer([seq_len, hidden_size]);\n  },\n};\n\n// ── EmbeddingInt4 ──\n\nconst embeddingInt4Spec: KernelSpec = {\n  shaderCode: WGSL_EMBEDDING_INT4,\n  entryPoint: \"main\",\n\n  // @binding(0) input_ids: storage read\n  // @binding(1) weight_q: storage read\n  // @binding(2) scales: storage read\n  // @binding(3) zeros: storage read\n  // @binding(4) output: storage read_write\n  // @binding(5) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const hidden_size = op.attributes.hidden_size as number;\n    const outShape = resolvedShapes[op.outputs[0]];\n    const seq_len = outShape ? outShape[0] : (op.attributes.seq_len as number);\n    return [cdiv(seq_len * hidden_size, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    // Params { seq_len: u32, hidden_size: u32, group_size: u32, _pad: u32 }\n    const hidden_size = op.attributes.hidden_size as number;\n    const group_size = op.attributes.group_size as number;\n    const outShape = resolvedShapes[op.outputs[0]];\n    const seq_len = outShape ? outShape[0] : (op.attributes.seq_len as number);\n    return buildUniformBuffer([seq_len, hidden_size, group_size, 0]);\n  },\n};\n\n// ── RoPE ──\n\nconst ropeSpec: KernelSpec = {\n  shaderCode: WGSL_ROPE,\n  entryPoint: \"main\",\n\n  // @binding(0) q: storage read_write\n  // @binding(1) k: storage read_write\n  // @binding(2) params: uniform\n  bindings: [{ type: \"storage-read-write\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n\n  getDispatchSize(op, resolvedShapes) {\n    const num_q_heads = op.attributes.num_q_heads as number;\n    const num_kv_heads = op.attributes.num_kv_heads as number;\n    const head_dim = op.attributes.head_dim as number;\n    const rope_dim = (op.attributes.rope_dim as number) ?? head_dim;\n    // rope_half is the rotate_half pairing offset / pair count. Defaults to\n    // rope_dim/2 (legacy), but Gemma \"proportional\" overrides it to head_dim/2.\n    const rope_half = (op.attributes.rope_half as number) ?? rope_dim / 2;\n\n    const seqRef = op.attributes.seq_len_tensor as string | undefined;\n    let seq_len: number;\n    if (seqRef && resolvedShapes[seqRef]) {\n      seq_len = resolvedShapes[seqRef][0];\n    } else {\n      seq_len = op.attributes.seq_len as number;\n    }\n\n    const total_q_pairs = seq_len * num_q_heads * rope_half;\n    const total_k_pairs = seq_len * num_kv_heads * rope_half;\n    return [cdiv(Math.max(total_q_pairs, total_k_pairs), 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes, context) {\n    // Params { seq_len, num_q_heads, num_kv_heads, head_dim, rope_base_bits,\n    //          position_offset, rope_dim, rope_half, rope_denom, rope_active_pairs, _pad0, _pad1 }\n    const num_q_heads = op.attributes.num_q_heads as number;\n    const num_kv_heads = op.attributes.num_kv_heads as number;\n    const head_dim = op.attributes.head_dim as number;\n    const rope_base = op.attributes.rope_base as number;\n    const rope_dim = (op.attributes.rope_dim as number) ?? head_dim;\n    // Decoupled RoPE knobs (default to legacy single-rope_dim behavior):\n    //   rope_half          = rotate_half pairing offset + pair count (legacy rope_dim/2)\n    //   rope_denom         = inv_freq exponent denominator (legacy rope_dim)\n    //   rope_active_pairs  = pairs with a real frequency; rest are identity (legacy rope_dim/2)\n    const rope_half = (op.attributes.rope_half as number) ?? rope_dim / 2;\n    const rope_denom = (op.attributes.rope_denom as number) ?? rope_dim;\n    const rope_active_pairs = (op.attributes.rope_active_pairs as number) ?? rope_dim / 2;\n    // Use runtime seqPos for position offset during autoregressive decode\n    const position_offset = context?.seqPos ?? (op.attributes.position_offset as number) ?? 0;\n\n    const seqRef = op.attributes.seq_len_tensor as string | undefined;\n    let seq_len: number;\n    if (seqRef && resolvedShapes[seqRef]) {\n      seq_len = resolvedShapes[seqRef][0];\n    } else {\n      seq_len = op.attributes.seq_len as number;\n    }\n\n    return buildUniformBuffer([\n      seq_len,\n      num_q_heads,\n      num_kv_heads,\n      head_dim,\n      f32BitsToU32(rope_base),\n      position_offset,\n      rope_dim,\n      rope_half,\n      rope_denom,\n      rope_active_pairs,\n      0, // _pad0\n      0, // _pad1\n    ]);\n  },\n};\n\n/**\n * Interleaved (adjacent-pair) RoPE, used by Moonshine. Identical dispatch/params\n * to ropeSpec — only the rotation pairing differs (2p/2p+1 vs p/p+half). Selected\n * by the executors when a RoPE node carries `interleaved: true`.\n */\nexport const ROPE_INTERLEAVED_SPEC: KernelSpec = {\n  ...ropeSpec,\n  shaderCode: WGSL_ROPE_INTERLEAVED,\n};\n\n// ── M-RoPE (precomputed cos/sin, partial rotate_half) ──\n\nconst mropeSpec: KernelSpec = {\n  shaderCode: WGSL_MROPE,\n  entryPoint: \"main\",\n  // @binding(0) q rw, (1) k rw, (2) cos, (3) sin, (4) params\n  bindings: [\n    { type: \"storage-read-write\" },\n    { type: \"storage-read-write\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"uniform\" },\n  ],\n  getDispatchSize(op, resolvedShapes) {\n    const num_q_heads = op.attributes.num_q_heads as number;\n    const num_kv_heads = op.attributes.num_kv_heads as number;\n    const rope_dim = op.attributes.rope_dim as number;\n    const half = rope_dim / 2;\n    const seqRef = op.attributes.seq_len_tensor as string | undefined;\n    const seq_len =\n      seqRef && resolvedShapes[seqRef]\n        ? resolvedShapes[seqRef][0]\n        : (op.attributes.seq_len as number);\n    const total = seq_len * Math.max(num_q_heads, num_kv_heads) * half;\n    return [cdiv(total, 256), 1, 1];\n  },\n  buildParams(op, resolvedShapes, context) {\n    const num_q_heads = op.attributes.num_q_heads as number;\n    const num_kv_heads = op.attributes.num_kv_heads as number;\n    const head_dim = op.attributes.head_dim as number;\n    const rope_dim = op.attributes.rope_dim as number;\n    const seqRef = op.attributes.seq_len_tensor as string | undefined;\n    const seq_len =\n      seqRef && resolvedShapes[seqRef]\n        ? resolvedShapes[seqRef][0]\n        : (op.attributes.seq_len as number);\n    // Decode (single-token) reads cos/sin row seqPos; prefill writes all rows 0..T.\n    const pos_offset = seq_len === 1 ? (context?.seqPos ?? 0) : 0;\n    return buildUniformBuffer([\n      seq_len,\n      num_q_heads,\n      num_kv_heads,\n      head_dim,\n      rope_dim,\n      pos_offset,\n      0,\n      0,\n    ]);\n  },\n};\n\n// ── EmbedSplice (scatter vision tokens into image-token rows) ──\n\nconst embedSpliceSpec: KernelSpec = {\n  shaderCode: WGSL_EMBED_SPLICE,\n  entryPoint: \"main\",\n  // @binding(0) embed_in, (1) vision, (2) row_map, (3) embed_out, (4) params\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n  getDispatchSize(op, resolvedShapes) {\n    const hidden = op.attributes.hidden as number;\n    const seqRef = op.attributes.seq_len_tensor as string;\n    const seq_len = resolvedShapes[seqRef][0];\n    return [cdiv(seq_len * hidden, 256), 1, 1];\n  },\n  buildParams(op, resolvedShapes) {\n    const hidden = op.attributes.hidden as number;\n    const seqRef = op.attributes.seq_len_tensor as string;\n    const seq_len = resolvedShapes[seqRef][0];\n    return buildUniformBuffer([seq_len, hidden]);\n  },\n};\n\n// ── Attention ──\n\nconst attentionSpec: KernelSpec = {\n  shaderCode: WGSL_ATTENTION,\n  entryPoint: \"main\",\n\n  // @binding(0) Q: storage read\n  // @binding(1) K: storage read\n  // @binding(2) V: storage read\n  // @binding(3) output: storage read_write\n  // @binding(4) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    // One workgroup per (query_position, q_head) pair\n    const num_q_heads = op.attributes.num_q_heads as number;\n    // Resolve T from output shape or static attribute\n    const outShape = resolvedShapes[op.outputs[0]];\n    const T = outShape ? outShape[0] : (op.attributes.T as number);\n    return [T, num_q_heads, 1];\n  },\n\n  buildParams(op, resolvedShapes, context) {\n    // Params { T, S, num_q_heads, num_kv_heads, head_dim, position_offset, is_causal }\n    const num_q_heads = op.attributes.num_q_heads as number;\n    const num_kv_heads = op.attributes.num_kv_heads as number;\n    const head_dim = op.attributes.head_dim as number;\n    // Use runtime seqPos for position offset during autoregressive decode\n    const position_offset = context?.seqPos ?? (op.attributes.position_offset as number) ?? 0;\n\n    // Resolve T from output shape or static attribute\n    const outShape = resolvedShapes[op.outputs[0]];\n    const T = outShape ? outShape[0] : (op.attributes.T as number);\n\n    // S = total KV length (seqPos + T from KV cache, or T for prefill-only)\n    const kvShape = resolvedShapes[op.inputs[1]]; // K cache shape: [L_max, kv_dim]\n    const S = kvShape ? kvShape[0] : T;\n\n    // Default causal (text). ViT attention sets causal:false → bidirectional.\n    const is_causal = (op.attributes.causal as boolean) === false ? 0 : 1;\n\n    // Query-grid base offset (default 0). The WebKit vision encoder sets this\n    // per-chunk to split the O(N²) ViT attention into watchdog-safe windows; all\n    // other callers leave it 0, keeping their dispatch byte-identical.\n    const q_offset = context?.qOffset ?? 0;\n\n    // Softmax QK temperature. Defaults to 1/sqrt(head_dim) — byte-identical to\n    // the legacy kernel-hardcoded scale, so every existing caller is unchanged.\n    // Gemma 4 sets attn_scale:1.0 (its per-head QK-norm absorbs the scale).\n    const attn_scale = (op.attributes.attn_scale as number) ?? 1 / Math.sqrt(head_dim);\n\n    return buildUniformBuffer([\n      T,\n      S,\n      num_q_heads,\n      num_kv_heads,\n      head_dim,\n      position_offset,\n      is_causal,\n      q_offset,\n      f32BitsToU32(attn_scale),\n    ]);\n  },\n};\n\n// ── Cross-Attention (encoder-decoder) ──\n//\n// Inputs: Q (decoder hidden) [T, num_q_heads*head_dim],\n//         K/V (frozen encoder output) [S, num_kv_heads*head_dim].\n// Output: [T, num_q_heads*head_dim]. One workgroup per (q_pos, q_head).\nconst crossAttentionSpec: KernelSpec = {\n  shaderCode: WGSL_CROSS_ATTENTION,\n  entryPoint: \"main\",\n\n  // @binding(0) Q: storage read\n  // @binding(1) K: storage read  (encoder output, frozen)\n  // @binding(2) V: storage read  (encoder output, frozen)\n  // @binding(3) output: storage read_write\n  // @binding(4) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const num_q_heads = op.attributes.num_q_heads as number;\n    // T = decoder query length (rows of the output / Q tensor)\n    const outShape = resolvedShapes[op.outputs[0]];\n    const T = outShape ? outShape[0] : (op.attributes.T as number);\n    return [T, num_q_heads, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    // Params { T, S, num_q_heads, num_kv_heads, head_dim }\n    const num_q_heads = op.attributes.num_q_heads as number;\n    const num_kv_heads = op.attributes.num_kv_heads as number;\n    const head_dim = op.attributes.head_dim as number;\n\n    // T = decoder query length (rows of Q / output)\n    const outShape = resolvedShapes[op.outputs[0]];\n    const T = outShape ? outShape[0] : (op.attributes.T as number);\n\n    // S = encoder length (rows of the frozen encoder K input)\n    const kShape = resolvedShapes[op.inputs[1]];\n    const S = kShape ? kShape[0] : (op.attributes.S as number);\n\n    return buildUniformBuffer([T, S, num_q_heads, num_kv_heads, head_dim]);\n  },\n};\n\n// ── KVCacheAppend Packed F16 (f32 src → packed u32 dst, Safari-safe) ──\n\nexport const KV_CACHE_APPEND_PACKED_F16_SPEC: KernelSpec = {\n  shaderCode: WGSL_KV_CACHE_APPEND_PACKED_F16,\n  entryPoint: \"main\",\n\n  // @binding(0) src: storage read (f32)\n  // @binding(1) dst: storage read_write (packed u32)\n  // @binding(2) params: uniform\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n\n  getDispatchSize(op, resolvedShapes) {\n    const width = op.attributes.width as number;\n    const srcRef = op.attributes.T_tensor as string | undefined;\n    let T: number;\n    if (srcRef && resolvedShapes[srcRef]) {\n      T = resolvedShapes[srcRef][0];\n    } else {\n      T = op.attributes.T as number;\n    }\n    // Each thread processes a pair of f32 values → half the dispatch count\n    return [cdiv((T * width) / 2, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes, context) {\n    const width = op.attributes.width as number;\n    const seqPos = context?.seqPos ?? 0;\n    const srcRef = op.attributes.T_tensor as string | undefined;\n    let T: number;\n    if (srcRef && resolvedShapes[srcRef]) {\n      T = resolvedShapes[srcRef][0];\n    } else {\n      T = op.attributes.T as number;\n    }\n    // Same params as native f16 — kernel internally divides by 2\n    return buildUniformBuffer([T * width, seqPos * width]);\n  },\n};\n\n// ── Attention Packed F16 (packed u32 KV cache, Safari-safe) ──\n\nexport const ATTENTION_PACKED_F16_SPEC: KernelSpec = {\n  shaderCode: WGSL_ATTENTION_PACKED_F16,\n  entryPoint: \"main\",\n\n  // @binding(0) Q: storage read (f32)\n  // @binding(1) K: storage read (packed u32)\n  // @binding(2) V: storage read (packed u32)\n  // @binding(3) output: storage read_write (f32)\n  // @binding(4) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const num_q_heads = op.attributes.num_q_heads as number;\n    const outShape = resolvedShapes[op.outputs[0]];\n    const T = outShape ? outShape[0] : (op.attributes.T as number);\n    return [T, num_q_heads, 1];\n  },\n\n  buildParams(op, resolvedShapes, context) {\n    const num_q_heads = op.attributes.num_q_heads as number;\n    const num_kv_heads = op.attributes.num_kv_heads as number;\n    const head_dim = op.attributes.head_dim as number;\n    const position_offset = context?.seqPos ?? (op.attributes.position_offset as number) ?? 0;\n\n    const outShape = resolvedShapes[op.outputs[0]];\n    const T = outShape ? outShape[0] : (op.attributes.T as number);\n\n    const kvShape = resolvedShapes[op.inputs[1]];\n    const S = kvShape ? kvShape[0] : T;\n\n    // Default 1/sqrt(head_dim) → byte-identical to the legacy scale; Gemma4 = 1.0.\n    const attn_scale = (op.attributes.attn_scale as number) ?? 1 / Math.sqrt(head_dim);\n\n    return buildUniformBuffer([\n      T,\n      S,\n      num_q_heads,\n      num_kv_heads,\n      head_dim,\n      position_offset,\n      f32BitsToU32(attn_scale),\n    ]);\n  },\n};\n\n// ── Softmax ──\n\nconst softmaxSpec: KernelSpec = {\n  shaderCode: WGSL_SOFTMAX,\n  entryPoint: \"main\",\n\n  // @binding(0) input: storage read\n  // @binding(1) output: storage read_write\n  // @binding(2) params: uniform\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n\n  getDispatchSize(op) {\n    // One workgroup per row\n    const num_rows = op.attributes.num_rows as number;\n    return [num_rows, 1, 1];\n  },\n\n  buildParams(op) {\n    // Params { num_rows: u32, row_size: u32 }\n    const num_rows = op.attributes.num_rows as number;\n    const row_size = op.attributes.row_size as number;\n    return buildUniformBuffer([num_rows, row_size]);\n  },\n};\n\n// ── CausalConv1d ──\n\nconst causalConv1dSpec: KernelSpec = {\n  shaderCode: WGSL_CAUSAL_CONV1D,\n  entryPoint: \"main\",\n\n  // @binding(0) input: storage read\n  // @binding(1) weight: storage read\n  // @binding(2) conv_state: storage read\n  // @binding(3) output: storage read_write\n  // @binding(4) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const channels = op.attributes.channels as number;\n    const ref = op.attributes.seq_len_tensor as string | undefined;\n    let seq_len: number;\n    if (ref && resolvedShapes[ref]) {\n      seq_len = resolvedShapes[ref][0];\n    } else {\n      seq_len = op.attributes.seq_len as number;\n    }\n    return [cdiv(seq_len * channels, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    // Params { seq_len: u32, channels: u32, kernel_size: u32, _pad: u32 }\n    const channels = op.attributes.channels as number;\n    const kernel_size = op.attributes.kernel_size as number;\n    const ref = op.attributes.seq_len_tensor as string | undefined;\n    let seq_len: number;\n    if (ref && resolvedShapes[ref]) {\n      seq_len = resolvedShapes[ref][0];\n    } else {\n      seq_len = op.attributes.seq_len as number;\n    }\n    return buildUniformBuffer([seq_len, channels, kernel_size, 0]);\n  },\n};\n\n// ── CausalConv1dSiLU (conv + fused SiLU) ──\n\nconst causalConv1dSiluSpec: KernelSpec = {\n  shaderCode: WGSL_CAUSAL_CONV1D_SILU,\n  entryPoint: \"main\",\n\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const channels = op.attributes.channels as number;\n    const ref = op.attributes.seq_len_tensor as string | undefined;\n    let seq_len: number;\n    if (ref && resolvedShapes[ref]) {\n      seq_len = resolvedShapes[ref][0];\n    } else {\n      seq_len = op.attributes.seq_len as number;\n    }\n    return [cdiv(seq_len * channels, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const channels = op.attributes.channels as number;\n    const kernel_size = op.attributes.kernel_size as number;\n    const ref = op.attributes.seq_len_tensor as string | undefined;\n    let seq_len: number;\n    if (ref && resolvedShapes[ref]) {\n      seq_len = resolvedShapes[ref][0];\n    } else {\n      seq_len = op.attributes.seq_len as number;\n    }\n    return buildUniformBuffer([seq_len, channels, kernel_size, 0]);\n  },\n};\n\n// ── CausalConv1dGated (conv + fused multiplicative output gate) ──\n\nconst causalConv1dGatedSpec: KernelSpec = {\n  shaderCode: WGSL_CAUSAL_CONV1D_GATED,\n  entryPoint: \"main\",\n\n  // @binding(0) input  @binding(1) weight  @binding(2) conv_state\n  // @binding(3) gate   @binding(4) output  @binding(5) params\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const channels = op.attributes.channels as number;\n    const ref = op.attributes.seq_len_tensor as string | undefined;\n    let seq_len: number;\n    if (ref && resolvedShapes[ref]) {\n      seq_len = resolvedShapes[ref][0];\n    } else {\n      seq_len = op.attributes.seq_len as number;\n    }\n    return [cdiv(seq_len * channels, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const channels = op.attributes.channels as number;\n    const kernel_size = op.attributes.kernel_size as number;\n    const ref = op.attributes.seq_len_tensor as string | undefined;\n    let seq_len: number;\n    if (ref && resolvedShapes[ref]) {\n      seq_len = resolvedShapes[ref][0];\n    } else {\n      seq_len = op.attributes.seq_len as number;\n    }\n    return buildUniformBuffer([seq_len, channels, kernel_size, 0]);\n  },\n};\n\n// ── SigmoidGate ──\n\nconst sigmoidGateSpec: KernelSpec = {\n  shaderCode: WGSL_SIGMOID_GATE,\n  entryPoint: \"main\",\n\n  // @binding(0) x: storage read\n  // @binding(1) gate: storage read\n  // @binding(2) output: storage read_write\n  // @binding(3) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    // Params { count: u32 }\n    return buildUniformBuffer([count]);\n  },\n};\n\n// ── MambaSSM ──\n\nconst mambaSSMSpec: KernelSpec = {\n  shaderCode: WGSL_MAMBA_SSM,\n  entryPoint: \"main\",\n\n  // @binding(0) qkv_conv: storage read\n  // @binding(1) a_input: storage read\n  // @binding(2) b_input: storage read\n  // @binding(3) A_log: storage read\n  // @binding(4) dt_bias: storage read\n  // @binding(5) ssm_state: storage read_write\n  // @binding(6) output: storage read_write\n  // @binding(7) params: uniform\n  bindings: [\n    { type: \"storage-read\" }, // qkv_conv\n    { type: \"storage-read\" }, // a_input\n    { type: \"storage-read\" }, // b_input\n    { type: \"storage-read\" }, // A_log\n    { type: \"storage-read\" }, // dt_bias\n    { type: \"storage-read-write\" }, // ssm_state\n    { type: \"storage-read-write\" }, // output\n    { type: \"uniform\" }, // params\n  ],\n\n  getDispatchSize(op) {\n    const num_heads = op.attributes.num_heads as number;\n    return [num_heads, 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const num_heads = op.attributes.num_heads as number;\n    const key_dim = op.attributes.key_dim as number;\n    const val_dim = op.attributes.val_dim as number;\n    const qkv_dim = op.attributes.qkv_dim as number;\n    // Resolve T from qkv_conv tensor reference\n    const qkvRef = op.attributes.T_tensor as string | undefined;\n    let T: number;\n    if (qkvRef && resolvedShapes[qkvRef]) {\n      T = resolvedShapes[qkvRef][0];\n    } else {\n      T = op.attributes.T as number;\n    }\n    return buildUniformBuffer([T, num_heads, key_dim, val_dim, qkv_dim, 0, 0, 0]);\n  },\n};\n\n// ── MambaSSM with f16 state (Dawn only — selected in executor when hasF16) ──\n\nexport const MAMBA_SSM_F16_SPEC: KernelSpec = {\n  ...mambaSSMSpec,\n  shaderCode: WGSL_MAMBA_SSM_F16,\n};\n\n// ── KVCacheAppend ──\n\nconst kvCacheAppendSpec: KernelSpec = {\n  shaderCode: WGSL_KV_CACHE_APPEND,\n  entryPoint: \"main\",\n\n  // @binding(0) src: storage read\n  // @binding(1) dst: storage read_write\n  // @binding(2) params: uniform\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n\n  getDispatchSize(op, resolvedShapes) {\n    const width = op.attributes.width as number;\n    const srcRef = op.attributes.T_tensor as string | undefined;\n    let T: number;\n    if (srcRef && resolvedShapes[srcRef]) {\n      T = resolvedShapes[srcRef][0];\n    } else {\n      T = op.attributes.T as number;\n    }\n    return [cdiv(T * width, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes, context) {\n    const width = op.attributes.width as number;\n    const seqPos = context?.seqPos ?? 0;\n    const srcRef = op.attributes.T_tensor as string | undefined;\n    let T: number;\n    if (srcRef && resolvedShapes[srcRef]) {\n      T = resolvedShapes[srcRef][0];\n    } else {\n      T = op.attributes.T as number;\n    }\n    return buildUniformBuffer([T * width, seqPos * width]);\n  },\n};\n\n// ── Dual KVCacheAppend (f32 K+V into both caches, one dispatch) ──\n\nexport const DUAL_KV_CACHE_APPEND_SPEC: KernelSpec = {\n  shaderCode: WGSL_DUAL_KV_CACHE_APPEND,\n  entryPoint: \"main\",\n\n  // @binding(0) src_k, (1) src_v, (2) dst_k, (3) dst_v, (4) params\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const width = op.attributes.width as number;\n    const srcRef = op.attributes.T_tensor as string | undefined;\n    const T =\n      srcRef && resolvedShapes[srcRef] ? resolvedShapes[srcRef][0] : (op.attributes.T as number);\n    return [cdiv(T * width, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes, context) {\n    const width = op.attributes.width as number;\n    const seqPos = context?.seqPos ?? 0;\n    const srcRef = op.attributes.T_tensor as string | undefined;\n    const T =\n      srcRef && resolvedShapes[srcRef] ? resolvedShapes[srcRef][0] : (op.attributes.T as number);\n    return buildUniformBuffer([T * width, seqPos * width]);\n  },\n};\n\n// ── Dual KVCacheAppend native-f16 (f32 K+V → f16 caches, one dispatch) ──\n\nexport const DUAL_KV_CACHE_APPEND_F16_SPEC: KernelSpec = {\n  ...DUAL_KV_CACHE_APPEND_SPEC,\n  shaderCode: WGSL_DUAL_KV_CACHE_APPEND_F16,\n};\n\n// ── Dual KVCacheAppend packed-f16 (Safari-safe, one dispatch) ──\n\nexport const DUAL_KV_CACHE_APPEND_PACKED_F16_SPEC: KernelSpec = {\n  ...DUAL_KV_CACHE_APPEND_SPEC,\n  shaderCode: WGSL_DUAL_KV_CACHE_APPEND_PACKED_F16,\n  // Each thread handles one f16 PAIR per cache → grid over count/2.\n  getDispatchSize(op, resolvedShapes) {\n    const width = op.attributes.width as number;\n    const srcRef = op.attributes.T_tensor as string | undefined;\n    const T =\n      srcRef && resolvedShapes[srcRef] ? resolvedShapes[srcRef][0] : (op.attributes.T as number);\n    return [cdiv((T * width) >> 1, 256), 1, 1];\n  },\n};\n\n// ── KVCacheAppend F16 (f32 src → f16 dst) ──\n\nexport const KV_CACHE_APPEND_F16_SPEC: KernelSpec = {\n  shaderCode: WGSL_KV_CACHE_APPEND_F16,\n  entryPoint: \"main\",\n\n  // @binding(0) src: storage read (f32)\n  // @binding(1) dst: storage read_write (f16)\n  // @binding(2) params: uniform\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n\n  getDispatchSize(op, resolvedShapes) {\n    const width = op.attributes.width as number;\n    const srcRef = op.attributes.T_tensor as string | undefined;\n    let T: number;\n    if (srcRef && resolvedShapes[srcRef]) {\n      T = resolvedShapes[srcRef][0];\n    } else {\n      T = op.attributes.T as number;\n    }\n    return [cdiv(T * width, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes, context) {\n    const width = op.attributes.width as number;\n    const seqPos = context?.seqPos ?? 0;\n    const srcRef = op.attributes.T_tensor as string | undefined;\n    let T: number;\n    if (srcRef && resolvedShapes[srcRef]) {\n      T = resolvedShapes[srcRef][0];\n    } else {\n      T = op.attributes.T as number;\n    }\n    return buildUniformBuffer([T * width, seqPos * width]);\n  },\n};\n\n// ── Attention F16 (f16 KV cache) ──\n\nexport const ATTENTION_F16_SPEC: KernelSpec = {\n  shaderCode: WGSL_ATTENTION_F16,\n  entryPoint: \"main\",\n\n  // @binding(0) Q: storage read (f32)\n  // @binding(1) K: storage read (f16)\n  // @binding(2) V: storage read (f16)\n  // @binding(3) output: storage read_write (f32)\n  // @binding(4) params: uniform\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n\n  getDispatchSize(op, resolvedShapes) {\n    const num_q_heads = op.attributes.num_q_heads as number;\n    const outShape = resolvedShapes[op.outputs[0]];\n    const T = outShape ? outShape[0] : (op.attributes.T as number);\n    return [T, num_q_heads, 1];\n  },\n\n  buildParams(op, resolvedShapes, context) {\n    const num_q_heads = op.attributes.num_q_heads as number;\n    const num_kv_heads = op.attributes.num_kv_heads as number;\n    const head_dim = op.attributes.head_dim as number;\n    const position_offset = context?.seqPos ?? (op.attributes.position_offset as number) ?? 0;\n\n    const outShape = resolvedShapes[op.outputs[0]];\n    const T = outShape ? outShape[0] : (op.attributes.T as number);\n\n    const kvShape = resolvedShapes[op.inputs[1]];\n    const S = kvShape ? kvShape[0] : T;\n\n    // Default 1/sqrt(head_dim) → byte-identical to the legacy scale; Gemma4 = 1.0.\n    const attn_scale = (op.attributes.attn_scale as number) ?? 1 / Math.sqrt(head_dim);\n\n    return buildUniformBuffer([\n      T,\n      S,\n      num_q_heads,\n      num_kv_heads,\n      head_dim,\n      position_offset,\n      f32BitsToU32(attn_scale),\n    ]);\n  },\n};\n\n// ── ConvStateUpdate ──\n\nconst convStateUpdateSpec: KernelSpec = {\n  shaderCode: WGSL_CONV_STATE_UPDATE,\n  entryPoint: \"main\",\n\n  // @binding(0) input: storage read\n  // @binding(1) conv_state: storage read_write\n  // @binding(2) params: uniform\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n\n  getDispatchSize(op) {\n    const channels = op.attributes.channels as number;\n    return [cdiv(channels, 256), 1, 1];\n  },\n\n  buildParams(op, resolvedShapes) {\n    const channels = op.attributes.channels as number;\n    const state_size = op.attributes.state_size as number;\n    const srcRef = op.attributes.T_tensor as string | undefined;\n    let seq_len: number;\n    if (srcRef && resolvedShapes[srcRef]) {\n      seq_len = resolvedShapes[srcRef][0];\n    } else {\n      seq_len = op.attributes.T as number;\n    }\n    return buildUniformBuffer([seq_len, channels, state_size, 0]);\n  },\n};\n\n// ── Registry ────────────────────────────────────────────────────────────\n\n/**\n * Maps each implemented OpType to its kernel spec.\n *\n * Only ops with WGSL shader implementations are included.\n * Stubbed ops (MoERouter, ExpertMatMul, Conv2d, AvgPool2d,\n * CrossAttention, Gather, Reshape, Transpose, Concat) are not present.\n */\n// ── Audio codec-decoder kernels (OmniVoice / HiggsAudioV2 DAC vocoder) ──\n//\n// New vocoder primitives for native TTS decode (token grid -> 24kHz PCM). The\n// DAC decoder is: Conv1dFull (in) -> 5 upsample blocks (ConvTranspose1d + Snake1d\n// + dilated residual Conv1dFull) -> Snake1d -> Conv1dFull (out, NO tanh). These\n// three kernels cover every op in that path.\n//\n// Validated bit-exact vs NumPy reference (max|err| ~2e-7, f32 rounding only) in\n// scripts/engine/test-codec-kernels.mjs across: kernel-7 dilation-1/3, pointwise\n// kernel-1, transposed stride-4 (even) and stride-3 (odd, output_padding=1), and\n// per-channel Snake. Mobile-safe: flat workgroup, no select(), no exp(), no\n// shared memory, 1 thread per output element.\n\nconst WGSL_CONV1D_FULL = `\\\n// Full cross-channel 1D convolution: x[Cin,L] * w[Cout,Cin,K] + bias[Cout]\n//   -> y[Cout,Lout],  Lout = (L + 2*padding - dilation*(K-1) - 1)/stride + 1\n// One thread per (oc, ot) output element.\nstruct Params {\n  Cin: u32, Cout: u32, L: u32, Lout: u32,\n  K: u32, stride: u32, padding: u32, dilation: u32,\n}\n@group(0) @binding(0) var<storage, read> x: array<f32>;\n@group(0) @binding(1) var<storage, read> w: array<f32>;\n@group(0) @binding(2) var<storage, read> bias: array<f32>;\n@group(0) @binding(3) var<storage, read_write> y: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\n@compute @workgroup_size(64)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  let total = params.Cout * params.Lout;\n  if (idx >= total) { return; }\n  let oc = idx / params.Lout;\n  let ot = idx % params.Lout;\n  let base = i32(ot * params.stride) - i32(params.padding);\n  var acc: f32 = 0.0;\n  for (var ic: u32 = 0u; ic < params.Cin; ic = ic + 1u) {\n    let x_row = ic * params.L;\n    let w_row = (oc * params.Cin + ic) * params.K;\n    for (var k: u32 = 0u; k < params.K; k = k + 1u) {\n      let it = base + i32(k * params.dilation);\n      if (it >= 0 && it < i32(params.L)) {\n        acc = acc + x[x_row + u32(it)] * w[w_row + k];\n      }\n    }\n  }\n  y[oc * params.Lout + ot] = acc + bias[oc];\n}\n`;\n\nconst WGSL_CONV_TRANSPOSE1D = `\\\n// Transposed (strided) 1D convolution: x[Cin,L] * w[Cin,Cout,K] + bias[Cout]\n//   -> y[Cout,Lout],  Lout = (L-1)*stride - 2*padding + (K-1) + output_padding + 1\n// Gather form (one thread per output): for full-buffer position op = ot+padding,\n// sum over (ic,k) where op-k is a non-negative multiple of stride -> it=(op-k)/stride.\n// HiggsAudioV2 sets output_padding = stride % 2 on every DAC upsample block.\nstruct Params {\n  Cin: u32, Cout: u32, L: u32, Lout: u32,\n  K: u32, stride: u32, padding: u32, out_pad: u32,\n}\n@group(0) @binding(0) var<storage, read> x: array<f32>;\n@group(0) @binding(1) var<storage, read> w: array<f32>;\n@group(0) @binding(2) var<storage, read> bias: array<f32>;\n@group(0) @binding(3) var<storage, read_write> y: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\n@compute @workgroup_size(64)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  let total = params.Cout * params.Lout;\n  if (idx >= total) { return; }\n  let oc = idx / params.Lout;\n  let ot = idx % params.Lout;\n  let op = i32(ot + params.padding);\n  var acc: f32 = 0.0;\n  for (var ic: u32 = 0u; ic < params.Cin; ic = ic + 1u) {\n    let x_row = ic * params.L;\n    let w_row = (ic * params.Cout + oc) * params.K;\n    for (var k: u32 = 0u; k < params.K; k = k + 1u) {\n      let num = op - i32(k);\n      if (num >= 0 && (num % i32(params.stride)) == 0) {\n        let it = num / i32(params.stride);\n        if (it >= 0 && it < i32(params.L)) {\n          acc = acc + x[x_row + u32(it)] * w[w_row + k];\n        }\n      }\n    }\n  }\n  y[oc * params.Lout + ot] = acc + bias[oc];\n}\n`;\n\nconst WGSL_SNAKE1D = `\\\n// Snake activation (learned per-channel periodic): y = x + (1/(alpha+1e-9)) * sin(alpha*x)^2\nstruct Params { C: u32, L: u32 }\n@group(0) @binding(0) var<storage, read> x: array<f32>;\n@group(0) @binding(1) var<storage, read> alpha: array<f32>;\n@group(0) @binding(2) var<storage, read_write> y: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\n@compute @workgroup_size(64)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  let total = params.C * params.L;\n  if (idx >= total) { return; }\n  let c = idx / params.L;\n  let a = alpha[c];\n  let s = sin(a * x[idx]);\n  y[idx] = x[idx] + (1.0 / (a + 1e-9)) * s * s;\n}\n`;\n\n// ── Moonshine STT frontend kernels (Tanh, Transpose, GroupNorm) ──\n\nconst WGSL_TANH = `\\\n// Elementwise tanh (Moonshine conv1 activation). Clamp the argument: on Metal/Dawn\n// tanh() of a large magnitude returns NaN (internal exp overflows). |arg|>=10\n// already saturates to +/-1, so clamping to +/-20 is numerically exact and safe.\nstruct Params { count: u32 }\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read_write> output: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  if (idx >= params.count) { return; }\n  output[idx] = tanh(clamp(input[idx], -20.0, 20.0));\n}\n`;\n\nconst WGSL_TRANSPOSE = `\\\n// Transpose a row-major [rows, cols] tensor into [cols, rows].\n// Moonshine conv output is [C, L] (channels-major); the encoder transformer wants\n// [L, C] (token-major). One thread per output element; bounds-checked.\nstruct Params { rows: u32, cols: u32 }\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read_write> output: array<f32>;\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(256)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let idx = gid.x;\n  let total = params.rows * params.cols;\n  if (idx >= total) { return; }\n  // output is [cols, rows]: out[c, r] = in[r, c]\n  let c = idx / params.rows;\n  let r = idx % params.rows;\n  output[idx] = input[r * params.cols + c];\n}\n`;\n\nconst WGSL_GROUPNORM = `\\\n// GroupNorm(num_groups=1) over the channel axis of a [C, L] tensor: normalize each\n// channel-position element by the mean/variance computed across ALL C*L elements\n// (a single group), then scale/shift per-channel: y[c,l] = (x[c,l]-mu)/sqrt(var+eps)\n// * weight[c] + bias[c]. Moonshine's encoder frontend applies this after conv1.\n// One workgroup reduces the whole [C, L] tensor (C*L <= a few k for short clips).\nstruct Params { channels: u32, length: u32, eps_bits: u32, _pad: u32 }\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read> weight: array<f32>;\n@group(0) @binding(2) var<storage, read> bias: array<f32>;\n@group(0) @binding(3) var<storage, read_write> output: array<f32>;\n@group(0) @binding(4) var<storage, read> params: Params;\n\nvar<workgroup> shared_sum: array<f32, 256>;\nvar<workgroup> shared_sq: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(@builtin(local_invocation_id) lid: vec3u) {\n  let tid = lid.x;\n  let total = params.channels * params.length;\n  let eps = bitcast<f32>(params.eps_bits);\n\n  var local_sum: f32 = 0.0;\n  var local_sq: f32 = 0.0;\n  var i = tid;\n  while (i < total) {\n    let v = input[i];\n    local_sum += v;\n    local_sq += v * v;\n    i += 256u;\n  }\n  shared_sum[tid] = local_sum;\n  shared_sq[tid] = local_sq;\n  workgroupBarrier();\n\n  var stride: u32 = 128u;\n  while (stride > 0u) {\n    if (tid < stride) {\n      shared_sum[tid] += shared_sum[tid + stride];\n      shared_sq[tid] += shared_sq[tid + stride];\n    }\n    workgroupBarrier();\n    stride /= 2u;\n  }\n\n  let n = f32(total);\n  let mean = shared_sum[0] / n;\n  let variance = shared_sq[0] / n - mean * mean;\n  let inv_std = 1.0 / sqrt(variance + eps);\n  workgroupBarrier();\n\n  i = tid;\n  while (i < total) {\n    let c = i / params.length; // channel index (row-major [C, L])\n    output[i] = (input[i] - mean) * inv_std * weight[c] + bias[c];\n    i += 256u;\n  }\n}\n`;\n\nconst WGSL_LAYERNORM_NO_BIAS = `\\\n// LayerNorm without a bias term: output = ((x - mean) / sqrt(var + eps)) * weight.\n// Moonshine's encoder/decoder norms are weight-only (no .bias in the checkpoint),\n// so the node binds only [input, weight] — this variant drops the bias binding to\n// keep the executor's positional input->binding mapping correct.\nstruct Params {\n  seq_len: u32,\n  hidden_size: u32,\n  eps_bits: u32,\n  _pad: u32,\n}\n@group(0) @binding(0) var<storage, read> input: array<f32>;\n@group(0) @binding(1) var<storage, read> weight: array<f32>;\n@group(0) @binding(2) var<storage, read_write> output: array<f32>;\n@group(0) @binding(3) var<storage, read> params: Params;\n\nvar<workgroup> shared_sum: array<f32, 256>;\nvar<workgroup> shared_sq_sum: array<f32, 256>;\n\n@compute @workgroup_size(256)\nfn main(\n  @builtin(local_invocation_id) lid: vec3u,\n  @builtin(workgroup_id) wid: vec3u,\n) {\n  let row = wid.x;\n  if (row >= params.seq_len) { return; }\n\n  let tid = lid.x;\n  let row_offset = row * params.hidden_size;\n  let eps = bitcast<f32>(params.eps_bits);\n\n  var local_sum: f32 = 0.0;\n  var i = tid;\n  while (i < params.hidden_size) {\n    local_sum += input[row_offset + i];\n    i += 256u;\n  }\n  shared_sum[tid] = local_sum;\n  workgroupBarrier();\n\n  var stride: u32 = 128u;\n  while (stride > 0u) {\n    if (tid < stride) { shared_sum[tid] += shared_sum[tid + stride]; }\n    workgroupBarrier();\n    stride /= 2u;\n  }\n  let mean = shared_sum[0] / f32(params.hidden_size);\n\n  var local_sq: f32 = 0.0;\n  i = tid;\n  while (i < params.hidden_size) {\n    let diff = input[row_offset + i] - mean;\n    local_sq += diff * diff;\n    i += 256u;\n  }\n  shared_sq_sum[tid] = local_sq;\n  workgroupBarrier();\n\n  stride = 128u;\n  while (stride > 0u) {\n    if (tid < stride) { shared_sq_sum[tid] += shared_sq_sum[tid + stride]; }\n    workgroupBarrier();\n    stride /= 2u;\n  }\n  let inv_std = 1.0 / sqrt(shared_sq_sum[0] / f32(params.hidden_size) + eps);\n\n  i = tid;\n  while (i < params.hidden_size) {\n    output[row_offset + i] = (input[row_offset + i] - mean) * inv_std * weight[i];\n    i += 256u;\n  }\n}\n`;\n\n/**\n * LayerNorm with no bias term (Moonshine). Binds [input, weight, output, params]\n * — one fewer storage-read than the standard layernormSpec. The executor selects\n * this variant when a LayerNorm node carries only 2 inputs.\n */\nexport const LAYERNORM_NO_BIAS_SPEC: KernelSpec = {\n  shaderCode: WGSL_LAYERNORM_NO_BIAS,\n  entryPoint: \"main\",\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n  getDispatchSize(op, resolvedShapes) {\n    const hidden_size = op.attributes.hidden_size as number;\n    const ref = op.attributes.seq_len_tensor as string | undefined;\n    let seq_len: number;\n    if (ref && resolvedShapes[ref]) {\n      const shape = resolvedShapes[ref];\n      seq_len = shape.reduce((a, b) => a * b, 1) / hidden_size;\n    } else {\n      seq_len = op.attributes.seq_len as number;\n    }\n    return [seq_len, 1, 1];\n  },\n  buildParams(op, resolvedShapes) {\n    const hidden_size = op.attributes.hidden_size as number;\n    const eps = op.attributes.eps as number;\n    const ref = op.attributes.seq_len_tensor as string | undefined;\n    let seq_len: number;\n    if (ref && resolvedShapes[ref]) {\n      const shape = resolvedShapes[ref];\n      seq_len = shape.reduce((a, b) => a * b, 1) / hidden_size;\n    } else {\n      seq_len = op.attributes.seq_len as number;\n    }\n    return buildUniformBuffer([seq_len, hidden_size, f32BitsToU32(eps), 0]);\n  },\n};\n\nconst tanhSpec: KernelSpec = {\n  shaderCode: WGSL_TANH,\n  entryPoint: \"main\",\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n  buildParams(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return buildUniformBuffer([count]);\n  },\n};\n\nconst transposeSpec: KernelSpec = {\n  shaderCode: WGSL_TRANSPOSE,\n  entryPoint: \"main\",\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n  getDispatchSize(op, resolvedShapes) {\n    const count = outputElementCount(op, resolvedShapes);\n    return [cdiv(count, 256), 1, 1];\n  },\n  buildParams(op, resolvedShapes) {\n    // input shape [rows, cols]; output is [cols, rows].\n    const inShape = resolvedShapes[op.inputs[0]];\n    const rows = inShape ? inShape[0] : (op.attributes.rows as number);\n    const cols = inShape ? inShape[1] : (op.attributes.cols as number);\n    return buildUniformBuffer([rows, cols]);\n  },\n};\n\nconst groupnormSpec: KernelSpec = {\n  shaderCode: WGSL_GROUPNORM,\n  entryPoint: \"main\",\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n  // One workgroup reduces the full [C, L] tensor.\n  getDispatchSize() {\n    return [1, 1, 1];\n  },\n  buildParams(op, resolvedShapes) {\n    const channels = op.attributes.channels as number;\n    const inShape = resolvedShapes[op.inputs[0]];\n    const length = inShape ? inShape[1] : (op.attributes.length as number);\n    const eps = (op.attributes.eps as number) ?? 1e-5;\n    return buildUniformBuffer([channels, length, f32BitsToU32(eps), 0]);\n  },\n};\n\nconst conv1dFullSpec: KernelSpec = {\n  shaderCode: WGSL_CONV1D_FULL,\n  entryPoint: \"main\",\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n  getDispatchSize(op) {\n    const Cout = op.attributes.Cout as number;\n    const Lout = op.attributes.Lout as number;\n    return [cdiv(Cout * Lout, 64), 1, 1];\n  },\n  buildParams(op) {\n    return buildUniformBuffer([\n      op.attributes.Cin as number,\n      op.attributes.Cout as number,\n      op.attributes.L as number,\n      op.attributes.Lout as number,\n      op.attributes.K as number,\n      op.attributes.stride as number,\n      op.attributes.padding as number,\n      op.attributes.dilation as number,\n    ]);\n  },\n};\n\nconst convTranspose1dSpec: KernelSpec = {\n  shaderCode: WGSL_CONV_TRANSPOSE1D,\n  entryPoint: \"main\",\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n  getDispatchSize(op) {\n    const Cout = op.attributes.Cout as number;\n    const Lout = op.attributes.Lout as number;\n    return [cdiv(Cout * Lout, 64), 1, 1];\n  },\n  buildParams(op) {\n    return buildUniformBuffer([\n      op.attributes.Cin as number,\n      op.attributes.Cout as number,\n      op.attributes.L as number,\n      op.attributes.Lout as number,\n      op.attributes.K as number,\n      op.attributes.stride as number,\n      op.attributes.padding as number,\n      op.attributes.output_padding as number,\n    ]);\n  },\n};\n\nconst snake1dSpec: KernelSpec = {\n  shaderCode: WGSL_SNAKE1D,\n  entryPoint: \"main\",\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n  getDispatchSize(op) {\n    const C = op.attributes.C as number;\n    const L = op.attributes.L as number;\n    return [cdiv(C * L, 64), 1, 1];\n  },\n  buildParams(op) {\n    return buildUniformBuffer([op.attributes.C as number, op.attributes.L as number]);\n  },\n};\n\n// ════════════════════════════════════════════════════════════════════════\n// Kani-TTS-2 NanoCodec (NVIDIA NeMo, 22 kHz 0.6 kbps 12.5 fps) decoder kernels.\n// Validated bit-exact vs the real MLX NanoCodec via scripts/engine/test-nanocodec\n// -decode.mjs (max|err| 4.2e-6, f32 rounding only). Three new ops beyond the DAC\n// cluster above:\n//   - FSQDequant: code idx -> per-dim continuous latent (finite scalar quantization,\n//     4 groups x 4 dims, levels [9,8,8,7], mixed-radix base [1,9,72,576]).\n//   - HalfSnake1d: snake on the first C/2 channels, leaky-relu(0.01) on the rest.\n//   - ConvTranspose1dDepthwise: causal depthwise (groups=Cout) transposed conv —\n//     insert (stride-1) zeros, cross-correlate the FLIPPED kernel, right-trim.\n// The causal full conv reuses Conv1dFull with padding=(K-1)*dilation, Lout=L.\n// Mobile-safe: flat workgroup_size(64), no select(), no shared memory, no exp(),\n// 1 thread per output element, no `enable f16`.\n// ════════════════════════════════════════════════════════════════════════\n\nconst WGSL_FSQ_DEQUANT = `\\\n// Finite-scalar-quantization decode. codes[groups,T] (u32) -> latent[groups*dims,T]\n// (f32). Channel ch = g*dims + d; per-dim level/base passed as small constants.\nstruct Params {\n  groups: u32, dims: u32, T: u32, _pad: u32,\n  l0: u32, l1: u32, l2: u32, l3: u32,\n  b0: u32, b1: u32, b2: u32, b3: u32,\n}\n@group(0) @binding(0) var<storage, read> codes: array<u32>;        // [groups, T]\n@group(0) @binding(1) var<storage, read_write> latent: array<f32>; // [groups*dims, T]\n@group(0) @binding(2) var<storage, read> params: Params;\n\n@compute @workgroup_size(64)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let total = params.groups * params.dims * params.T;\n  if (gid.x >= total) { return; }\n  let ch = gid.x / params.T;\n  let t = gid.x % params.T;\n  let d = ch % params.dims;\n  let g = ch / params.dims;\n  var lvl: u32 = params.l0;\n  var base: u32 = params.b0;\n  if (d == 1u) { lvl = params.l1; base = params.b1; }\n  else if (d == 2u) { lvl = params.l2; base = params.b2; }\n  else if (d == 3u) { lvl = params.l3; base = params.b3; }\n  let idx = codes[g * params.T + t];\n  let nonneg = (idx / base) % lvl;\n  let scale = lvl / 2u;\n  latent[ch * params.T + t] = (f32(nonneg) - f32(scale)) / f32(scale);\n}\n`;\n\nconst WGSL_HALF_SNAKE1D = `\\\n// HalfSnake: first half channels -> snake(x; alpha); rest -> leaky_relu(x, 0.01).\nstruct Params { C: u32, L: u32, half: u32, _pad: u32 }\n@group(0) @binding(0) var<storage, read> x: array<f32>;       // [C, L]\n@group(0) @binding(1) var<storage, read> alpha: array<f32>;   // [half]\n@group(0) @binding(2) var<storage, read_write> y: array<f32>; // [C, L]\n@group(0) @binding(3) var<storage, read> params: Params;\n\n@compute @workgroup_size(64)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let total = params.C * params.L;\n  if (gid.x >= total) { return; }\n  let c = gid.x / params.L;\n  let v = x[gid.x];\n  if (c < params.half) {\n    let a = alpha[c];\n    let s = sin(a * v);\n    y[gid.x] = v + (1.0 / (a + 1e-9)) * s * s;\n  } else {\n    if (v >= 0.0) { y[gid.x] = v; } else { y[gid.x] = 0.01 * v; }\n  }\n}\n`;\n\nconst WGSL_CONV_TRANSPOSE1D_DW = `\\\n// Causal depthwise (groups = Cout) transposed conv. Weight in PyTorch\n// ConvTranspose1d layout [Cin, 1, K] (out/groups = 1); each output channel oc is fed\n// by cpg = Cin/Cout consecutive input channels (oc*cpg .. oc*cpg+cpg-1). Matches MLX\n// CausalConvTranspose1d: insert (stride-1) zeros, pad (K-1) both sides, cross-correlate\n// the FLIPPED kernel, then trim (K-stride) from the right. Output length Lout = L*stride.\n// Gather form (one thread per (oc, ot)): contributions where it*stride = ot+k-(K-1),\n// 0<=k<K, with flipped weight w[K-1-k].\nstruct Params { Cin: u32, Cout: u32, L: u32, Lout: u32, K: u32, stride: u32, cpg: u32, _pad: u32 }\n@group(0) @binding(0) var<storage, read> x: array<f32>;       // [Cin, L]\n@group(0) @binding(1) var<storage, read> w: array<f32>;       // [Cin, 1, K]\n@group(0) @binding(2) var<storage, read> bias: array<f32>;    // [Cout]\n@group(0) @binding(3) var<storage, read_write> y: array<f32>; // [Cout, Lout]\n@group(0) @binding(4) var<storage, read> params: Params;\n\n@compute @workgroup_size(64)\nfn main(@builtin(global_invocation_id) gid: vec3u) {\n  let total = params.Cout * params.Lout;\n  if (gid.x >= total) { return; }\n  let oc = gid.x / params.Lout;\n  let ot = gid.x % params.Lout;\n  var acc: f32 = 0.0;\n  for (var j: u32 = 0u; j < params.cpg; j = j + 1u) {\n    let ic = oc * params.cpg + j;\n    let x_row = ic * params.L;\n    let w_row = ic * params.K;\n    for (var k: u32 = 0u; k < params.K; k = k + 1u) {\n      let num = i32(ot) + i32(k) - (i32(params.K) - 1);\n      if (num >= 0 && (num % i32(params.stride)) == 0) {\n        let it = num / i32(params.stride);\n        if (it >= 0 && it < i32(params.L)) {\n          acc = acc + x[x_row + u32(it)] * w[w_row + (params.K - 1u - k)];\n        }\n      }\n    }\n  }\n  y[oc * params.Lout + ot] = acc + bias[oc];\n}\n`;\n\nconst fsqDequantSpec: KernelSpec = {\n  shaderCode: WGSL_FSQ_DEQUANT,\n  entryPoint: \"main\",\n  bindings: [{ type: \"storage-read\" }, { type: \"storage-read-write\" }, { type: \"uniform\" }],\n  getDispatchSize(op) {\n    const groups = op.attributes.groups as number;\n    const dims = op.attributes.dims as number;\n    const T = op.attributes.T as number;\n    return [cdiv(groups * dims * T, 64), 1, 1];\n  },\n  buildParams(op) {\n    const levels = op.attributes.levels as number[];\n    const base = op.attributes.base as number[];\n    return buildUniformBuffer([\n      op.attributes.groups as number,\n      op.attributes.dims as number,\n      op.attributes.T as number,\n      0,\n      levels[0],\n      levels[1],\n      levels[2],\n      levels[3],\n      base[0],\n      base[1],\n      base[2],\n      base[3],\n    ]);\n  },\n};\n\nconst halfSnake1dSpec: KernelSpec = {\n  shaderCode: WGSL_HALF_SNAKE1D,\n  entryPoint: \"main\",\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n  getDispatchSize(op) {\n    const C = op.attributes.C as number;\n    const L = op.attributes.L as number;\n    return [cdiv(C * L, 64), 1, 1];\n  },\n  buildParams(op) {\n    const C = op.attributes.C as number;\n    return buildUniformBuffer([C, op.attributes.L as number, Math.floor(C / 2), 0]);\n  },\n};\n\nconst convTranspose1dDepthwiseSpec: KernelSpec = {\n  shaderCode: WGSL_CONV_TRANSPOSE1D_DW,\n  entryPoint: \"main\",\n  bindings: [\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read\" },\n    { type: \"storage-read-write\" },\n    { type: \"uniform\" },\n  ],\n  getDispatchSize(op) {\n    const Cout = op.attributes.Cout as number;\n    const Lout = op.attributes.Lout as number;\n    return [cdiv(Cout * Lout, 64), 1, 1];\n  },\n  buildParams(op) {\n    const Cin = op.attributes.Cin as number;\n    const Cout = op.attributes.Cout as number;\n    return buildUniformBuffer([\n      Cin,\n      Cout,\n      op.attributes.L as number,\n      op.attributes.Lout as number,\n      op.attributes.K as number,\n      op.attributes.stride as number,\n      Math.floor(Cin / Cout),\n      0,\n    ]);\n  },\n};\n\nexport const KERNEL_REGISTRY: Partial<Record<OpType, KernelSpec>> = {\n  Embedding: embeddingSpec,\n  EmbeddingInt4: embeddingInt4Spec,\n  MatMul: matmulSpec,\n  MatMulBias: matmulBiasSpec,\n  MatMulInt4: matmulInt4Spec,\n  Add: addSpec,\n  Mul: mulSpec,\n  SwiGLU: swigluSpec,\n  SiLU: siluSpec,\n  ResidualRMSNorm: residualRmsnormSpec,\n  GELU: geluSpec,\n  GeluErf: geluErfSpec,\n  AddBias: addBiasSpec,\n  ApplyRotaryEmb: applyRotarySpec,\n  MRoPE: mropeSpec,\n  EmbedSplice: embedSpliceSpec,\n  SliceCols: sliceColsSpec,\n  MulCols: mulColsSpec,\n  PoolMatMul: poolMatMulSpec,\n  ClippedMatMul: clippedMatMulSpec,\n  RMSNorm: rmsnormSpec,\n  LayerNorm: layernormSpec,\n  RoPE: ropeSpec,\n  Attention: attentionSpec,\n  CrossAttention: crossAttentionSpec,\n  Softmax: softmaxSpec,\n  CausalConv1d: causalConv1dSpec,\n  CausalConv1dSiLU: causalConv1dSiluSpec,\n  CausalConv1dGated: causalConv1dGatedSpec,\n  SigmoidGate: sigmoidGateSpec,\n  MambaSSM: mambaSSMSpec,\n  KVCacheAppend: kvCacheAppendSpec,\n  ConvStateUpdate: convStateUpdateSpec,\n  SliceLastRow: sliceLastRowSpec,\n  MeanPool: meanPoolSpec,\n  Scale: scaleSpec,\n  Softcap: softcapSpec,\n  L2Norm: l2NormSpec,\n  // Audio codec-decoder (OmniVoice native TTS) — validated bit-exact, see above.\n  Conv1dFull: conv1dFullSpec,\n  ConvTranspose1d: convTranspose1dSpec,\n  Snake1d: snake1dSpec,\n  // Kani-TTS-2 NanoCodec decoder — validated bit-exact vs MLX, see above.\n  FSQDequant: fsqDequantSpec,\n  HalfSnake1d: halfSnake1dSpec,\n  ConvTranspose1dDepthwise: convTranspose1dDepthwiseSpec,\n  // Moonshine STT encoder frontend.\n  Tanh: tanhSpec,\n  Transpose: transposeSpec,\n  GroupNorm: groupnormSpec,\n};\n","/**\n * Graph executor — runs a ModelGraph on WebGPU.\n *\n * Optimized for minimal per-token overhead:\n * - Persistent bind groups (created once after weight upload, reused every forward)\n * - Pre-allocated input buffer (no per-forward buffer creation)\n * - Cached uniform buffers (contents updated via writeBuffer, no recreation)\n * - GPU-side argmax for greedy decode (4 bytes readback vs vocab_size*4)\n * - Dual dispatch paths: tiled matmul for prefill, K-parallel matvec for decode\n */\n\nimport type { KvMode } from \"./architectures/index.js\";\nimport type { GPUContext } from \"./device.js\";\nimport {\n  createBindGroup,\n  createReadbackBuffer,\n  createStorageBuffer,\n  createUniformBuffer,\n  destroyBuffers,\n  getOrCreatePipeline,\n} from \"./device.js\";\nimport type { ModelGraph, OpNode } from \"./ir.js\";\nimport { DTYPE_BYTES } from \"./ir.js\";\nimport {\n  ARGMAX_SPEC,\n  ATTENTION_F16_SPEC,\n  ATTENTION_PACKED_F16_SPEC,\n  DUAL_KV_CACHE_APPEND_F16_SPEC,\n  DUAL_KV_CACHE_APPEND_PACKED_F16_SPEC,\n  DUAL_KV_CACHE_APPEND_SPEC,\n  DUAL_MATVEC_INT4_SPEC,\n  DUAL_RMSNORM_SPEC,\n  GATED_MATVEC_INT4_SPEC,\n  KERNEL_REGISTRY,\n  type KernelSpec,\n  KV_CACHE_APPEND_F16_SPEC,\n  KV_CACHE_APPEND_PACKED_F16_SPEC,\n  LAYERNORM_NO_BIAS_SPEC,\n  MAMBA_SSM_F16_SPEC,\n  MATVEC_INT4_SPEC,\n  MATVEC_INT4_SUBGROUPS_SPEC,\n  MATVEC_SPEC,\n  MATVEC_SUBGROUPS_SPEC,\n  ROPE_INTERLEAVED_SPEC,\n  type RuntimeContext,\n  SWIGLU_GATED_MATVEC_INT4_SPEC,\n  SWIGLU_MATVEC_INT4_SPEC,\n  SWIGLU_MATVEC_SPEC,\n} from \"./kernels/registry.js\";\nimport type { WeightSource } from \"./weight-source.js\";\n\nconst MAP_MODE_READ = 0x0001;\n\n/**\n * Safari/Metal workaround: shader variant alternation.\n * Metal caches argument buffers per compiled function. When consecutive dispatches\n * use the same WGSL code (same Metal function), Metal reuses the previous dispatch's\n * argument buffer, ignoring setBindGroup(). We alternate between variant 0/1 of each\n * shader (prepending `const _MV: u32 = Xu;`) to force different Metal function\n * specializations, preventing argument buffer reuse.\n */\n\nexport interface ExecutorOptions {\n  maxSeqLen: number;\n  /** KV cache kernel strategy. Defaults to \"native-f16\" when not specified. */\n  kvMode?: KvMode;\n  /**\n   * WebKit only: dispatches per command buffer, with at most one command\n   * buffer in flight (awaited). 1 (default) is the proven-correct floor on\n   * iPad; larger values are faster if this WebKit version keeps storage\n   * writes visible across dispatches within one submission. Sweepable via\n   * the ?group=N URL param.\n   */\n  webkitGroupSize?: number;\n}\n\nexport interface ForwardResult {\n  logits: Float32Array;\n}\n\n/**\n * Pre-compiled dispatch entry for a single graph op.\n * Created once during initBindGroups(), reused every forward pass.\n */\ninterface DispatchEntry {\n  nodeId: string;\n  node: OpNode;\n  spec: KernelSpec;\n  pipeline: GPUComputePipeline;\n  bindGroup: GPUBindGroup;\n  uniformBuffer: GPUBuffer;\n  /** Cached last-written params bytes. Skip writeBuffer when unchanged. */\n  lastParamsBytes: Uint8Array | null;\n  /** Cached last dispatch size. */\n  lastDispatchSize: [number, number, number] | null;\n}\n\nexport class Executor {\n  private ctx: GPUContext;\n  private graph: ModelGraph;\n\n  private weightBuffers: Map<string, GPUBuffer> = new Map();\n  private activationBuffers: Map<string, GPUBuffer> = new Map();\n  private ssmStateBuffers: Map<string, GPUBuffer> = new Map();\n  private kvCacheBuffers: Map<string, GPUBuffer> = new Map();\n\n  /** Pre-allocated input_ids buffer (maxSeqLen * 4 bytes). */\n  private inputIdsBuffer: GPUBuffer;\n\n  /**\n   * CPU-resident Per-Layer-Embeddings (PLE) source for Gemma 4. The PLE table\n   * (`embed_tokens_per_layer`, [vocab, L*256]) is ~1.17GB at 4-bit and is kept\n   * OFF the GPU. Each forward step we gather + dequantize only the rows for the\n   * actual input tokens and upload a tiny [T, L*256] f32 buffer. See setPleSource.\n   */\n  private pleSource: {\n    packed: Uint32Array; // flat row-major INT4 nibbles, 8 per u32\n    scales: Float32Array; // per-group scale (Gerbil convention)\n    zeros: Float32Array; // per-group zero\n    width: number; // L*256 (row width)\n    groupSize: number; // dequant group size\n    targetTensor: string; // activation tensor the gathered rows are written to\n    cache?: {\n      cacheName: string;\n      packedKey: string;\n      scalesKey: string;\n      zerosKey: string;\n      packedLen: number;\n    };\n  } | null = null;\n  /** Reusable scratch for the dequantized PLE rows (resized on demand). */\n  private pleScratch: Float32Array | null = null;\n  /**\n   * One-time promise that materializes a cache-backed PLE table into heap. The\n   * table is read from CacheStorage on the FIRST forward — i.e. AFTER the GPU\n   * weight upload has completed, so the ~1.17 GB does not stack on top of the\n   * upload's transient allocations at the load-time memory high-water mark.\n   */\n  private pleReady: Promise<void> | null = null;\n\n  /**\n   * Dummy GPU buffer bound to an otherwise-aliasing storage-read-write slot.\n   * Used by RoPE-Q-only nodes (Gemma 4 KV-shared layers) whose node lists the\n   * same tensor as input and output: the RoPE kernel always declares two\n   * read_write bindings (q, k), but with num_kv_heads=0 the k slot is never\n   * touched. WebGPU still rejects two read_write bindings aliasing one buffer,\n   * so we bind this throwaway buffer to the unused k slot. Allocated lazily.\n   */\n  private bindingScratchBuffer: GPUBuffer | null = null;\n  /** Readback buffer for logits. */\n  private logitsReadback: GPUBuffer;\n  /** GPU buffer for argmax result (1 u32). */\n  private argmaxResultBuffer: GPUBuffer;\n  /** Readback buffer for argmax result (1 u32). */\n  private argmaxReadback: GPUBuffer;\n  /** Readback ring for pipelined greedy decode (created lazily). */\n  private decodeReadbacks: GPUBuffer[] = [];\n\n  /**\n   * Staging buffer for uniform param updates.\n   * Safari/Metal has weaker visibility guarantees for queue.writeBuffer() to\n   * UNIFORM buffers — early writes get dropped when hundreds are queued.\n   * Instead, we pack all params into this STORAGE staging buffer (1 writeBuffer),\n   * then use encoder.copyBufferToBuffer to distribute to each uniform buffer.\n   * Copies are GPU-sequenced and guaranteed to complete before compute passes.\n   */\n  private uniformStagingBuffer: GPUBuffer | null = null;\n  private uniformStagingCapacity: number = 0;\n\n  /** Dispatch entries for prefill (M>1): uses tiled matmul kernels. */\n  private dispatchEntries: DispatchEntry[] = [];\n  /** Dispatch entries for decode (M=1): uses K-parallel matvec kernels. */\n  private decodeEntries: DispatchEntry[] = [];\n  /** Argmax dispatch entry (created in initBindGroups). */\n  private argmaxEntry: {\n    pipeline: GPUComputePipeline;\n    bindGroup: GPUBindGroup;\n    uniformBuffer: GPUBuffer;\n  } | null = null;\n\n  /** True when running on Safari/WebKit (needs multi-encoder submit). */\n  readonly needsMultiEncoder: boolean;\n\n  private maxSeqLen: number;\n  private kvMode: KvMode;\n  private seqPos: number = 0;\n  private webkitGroupSize: number;\n\n  // ── Env-gated decode profiler (GERBIL_PROFILE) ──\n  // Times each dispatch with timestamp-query (one compute pass per dispatch) and\n  // accumulates GPU nanoseconds per opType. OFF on the normal inference path.\n  private profileEnabled = false;\n  private readonly profileData = new Map<string, { ns: number; count: number }>();\n  private querySet: GPUQuerySet | null = null;\n  private queryResolveBuf: GPUBuffer | null = null;\n  private queryReadbackBuf: GPUBuffer | null = null;\n\n  constructor(ctx: GPUContext, graph: ModelGraph, options: ExecutorOptions) {\n    this.ctx = ctx;\n    this.graph = graph;\n    this.maxSeqLen = options.maxSeqLen;\n    this.kvMode = options.kvMode ?? \"native-f16\";\n    this.webkitGroupSize = Math.max(1, options.webkitGroupSize ?? 1);\n    this.profileEnabled =\n      ctx.hasTimestamp && typeof process !== \"undefined\" && process.env?.GERBIL_PROFILE != null;\n\n    // WebKit's WebGPU fails to make storage writes visible across dispatches\n    // within one submission at production scale (see docs/mobile-failure-diagnosis.md);\n    // Dawn (Chrome, node) does not need the multi-encoder workaround.\n    this.needsMultiEncoder = ctx.isWebKitWebGPU;\n\n    // Pre-allocate input_ids buffer at max size\n    this.inputIdsBuffer = createStorageBuffer(ctx, \"input_ids\", options.maxSeqLen * 4);\n\n    this.allocateActivationBuffers();\n    this.allocateSSMStateBuffers();\n    this.allocateKVCacheBuffers();\n\n    // Guard against vocab_size 0 (non-LM graphs, e.g. the NanoCodec decoder run\n    // via runGraphOutput): a zero-size buffer is invalid in WebGPU.\n    this.logitsReadback = createReadbackBuffer(\n      ctx,\n      \"logits_readback\",\n      Math.max(4, graph.config.vocab_size * 4),\n    );\n    this.argmaxResultBuffer = createStorageBuffer(ctx, \"argmax_result\", 4);\n    this.argmaxReadback = createReadbackBuffer(ctx, \"argmax_readback\", 4);\n  }\n\n  /**\n   * Register the CPU-resident Gemma 4 PLE table. The quantized table is kept in\n   * JS memory (NOT uploaded to a GPU buffer), and {@link forward} gathers +\n   * dequantizes only the rows for the current input tokens each step, uploading a\n   * small [T, width] f32 buffer into `targetTensor`. This is what keeps Gemma 4\n   * mobile-viable: resident GPU memory is just the active transformer weights.\n   */\n  setPleSource(src: {\n    packed: Uint32Array;\n    scales: Float32Array;\n    zeros: Float32Array;\n    width: number;\n    groupSize: number;\n    targetTensor: string;\n    cache?: {\n      cacheName: string;\n      packedKey: string;\n      scalesKey: string;\n      zerosKey: string;\n      packedLen: number;\n    };\n  }): void {\n    this.pleSource = src;\n  }\n\n  /**\n   * Materialize a cache-backed PLE table into heap exactly once. Deferred to the\n   * first forward (after GPU upload) so the ~1.17 GB table is not co-resident\n   * with the upload's transient allocations during the load-time memory peak.\n   */\n  private ensurePleLoaded(): Promise<void> {\n    const src = this.pleSource;\n    if (!src || !src.cache) return Promise.resolve();\n    if (this.pleReady) return this.pleReady;\n    const { cacheName, packedKey, scalesKey, zerosKey } = src.cache;\n    this.pleReady = (async () => {\n      const cache = await caches.open(cacheName);\n      const read = async (key: string): Promise<ArrayBuffer> => {\n        const resp = await cache.match(new Request(key));\n        if (!resp) throw new Error(`PLE cache entry missing: ${key}`);\n        return resp.arrayBuffer();\n      };\n      const [pBuf, sBuf, zBuf] = await Promise.all([\n        read(packedKey),\n        read(scalesKey),\n        read(zerosKey),\n      ]);\n      src.packed = new Uint32Array(pBuf);\n      src.scales = new Float32Array(sBuf);\n      src.zeros = new Float32Array(zBuf);\n    })();\n    return this.pleReady;\n  }\n\n  /**\n   * Gather + dequantize the PLE rows for `inputIds` and upload them into the\n   * target activation buffer. Touches only T rows (T*width floats) — the full\n   * [vocab, width] quantized table never goes to the GPU.\n   */\n  private async streamPleRows(inputIds: Uint32Array): Promise<void> {\n    const src = this.pleSource;\n    if (!src) return;\n    if (src.cache) await this.ensurePleLoaded();\n    const T = inputIds.length;\n    const { packed, scales, zeros, width, groupSize } = src;\n    const needed = T * width;\n    if (!this.pleScratch || this.pleScratch.length < needed) {\n      this.pleScratch = new Float32Array(needed);\n    }\n    const out = this.pleScratch;\n    for (let t = 0; t < T; t++) {\n      const tokenId = inputIds[t];\n      const rowBase = tokenId * width; // flat element offset of this row\n      const outBase = t * width;\n      for (let d = 0; d < width; d++) {\n        const flat = rowBase + d;\n        const packedIdx = flat >>> 3; // flat / 8\n        const nibblePos = (flat & 7) * 4; // (flat % 8) * 4\n        const nibble = (packed[packedIdx] >>> nibblePos) & 0xf;\n        const groupIdx = (flat / groupSize) | 0;\n        out[outBase + d] = (nibble - zeros[groupIdx]) * scales[groupIdx];\n      }\n    }\n    const buffer = this.getBuffer(src.targetTensor);\n    if (!buffer) {\n      throw new Error(`PLE target tensor \"${src.targetTensor}\" has no GPU buffer`);\n    }\n    this.ctx.device.queue.writeBuffer(buffer, 0, out.buffer, out.byteOffset, needed * 4);\n  }\n\n  /**\n   * Stream weights to the GPU one tensor at a time, pulling each from the\n   * `WeightSource` only when it is about to be uploaded and dropping the\n   * reference immediately afterward.\n   *\n   * This is the property that bounds peak JS heap: with a cache-backed source\n   * (browser/mobile), only ONE tensor's bytes are materialized in heap at a time\n   * (read from CacheStorage in `source.get()`), uploaded to its GPU buffer, then\n   * released before the next tensor is fetched. The whole model is never co-\n   * resident in heap. With a heap-backed source (Node/desktop), behavior matches\n   * the old Map path (the source deletes each entry as it is consumed).\n   *\n   * Accepts either a `WeightSource` (new, streamed/async) or a plain Map\n   * (back-compat for callers that build a Map directly, e.g. Kani/Moonshine).\n   */\n  async uploadWeights(\n    source: WeightSource | Map<string, { data: ArrayBufferView; shape: number[] }>,\n  ): Promise<void> {\n    if (source instanceof Map) {\n      this.uploadWeightsMap(source);\n      return;\n    }\n    for (const name of source.keys()) {\n      const entry = await source.get(name);\n      if (!entry) continue;\n      const { data } = entry;\n      const buffer = createStorageBuffer(this.ctx, `weight_${name}`, data.byteLength, data);\n      this.weightBuffers.set(name, buffer);\n      // Each `entry` reference goes out of scope here; with a cache-backed source\n      // its bytes can be GC'd before the next tensor is pulled — bounding peak heap.\n    }\n  }\n\n  /**\n   * Synchronous Map upload (Node/desktop and the Kani/Moonshine sub-executors,\n   * which build small heap Maps). Deletes each entry as it is consumed to free\n   * the JS-side bytes once they are GPU-resident. Safe to call from a constructor.\n   */\n  uploadWeightsMap(weights: Map<string, { data: ArrayBufferView; shape: number[] }>): void {\n    const names = [...weights.keys()];\n    for (const name of names) {\n      const entry = weights.get(name);\n      if (!entry) continue;\n      const buffer = createStorageBuffer(\n        this.ctx,\n        `weight_${name}`,\n        entry.data.byteLength,\n        entry.data,\n      );\n      this.weightBuffers.set(name, buffer);\n      weights.delete(name);\n    }\n  }\n\n  /**\n   * Build all pipelines and bind groups. Call ONCE after uploadWeights().\n   *\n   * Creates two dispatch entry arrays:\n   * - dispatchEntries: tiled matmul for prefill (any M)\n   * - decodeEntries: K-parallel matvec for decode (M=1)\n   */\n  initBindGroups(): void {\n    for (const nodeId of this.graph.executionOrder) {\n      const node = this.graph.nodes.find((n) => n.id === nodeId)!;\n      let spec = KERNEL_REGISTRY[node.opType];\n      if (!spec) throw new Error(`No kernel for op type: ${node.opType}`);\n\n      // Use f16 kernel variants when KV cache tensors are f16.\n      // \"packed-f16\" uses pack2x16float (Safari-safe), \"native-f16\" uses enable f16.\n      if (node.opType === \"KVCacheAppend\") {\n        const hasF16KV = node.outputs.some((out) => this.graph.tensors[out]?.dtype === \"f16\");\n        if (hasF16KV) {\n          spec =\n            this.kvMode === \"packed-f16\"\n              ? KV_CACHE_APPEND_PACKED_F16_SPEC\n              : KV_CACHE_APPEND_F16_SPEC;\n        }\n      } else if (node.opType === \"Attention\") {\n        const hasF16KV = node.inputs.some((inp) => this.graph.tensors[inp]?.dtype === \"f16\");\n        if (hasF16KV) {\n          spec = this.kvMode === \"packed-f16\" ? ATTENTION_PACKED_F16_SPEC : ATTENTION_F16_SPEC;\n        }\n      } else if (node.opType === \"MambaSSM\" && this.ctx.hasF16 && !this.ctx.isWebKitWebGPU) {\n        // f16 SSM state storage halves the dominant state bandwidth.\n        // Dawn only — WebKit keeps the f32 kernel (unverified there).\n        spec = MAMBA_SSM_F16_SPEC;\n      } else if (node.opType === \"LayerNorm\" && node.inputs.length === 2) {\n        // Weight-only LayerNorm (Moonshine encoder/decoder norms have no bias):\n        // the node binds [input, weight] only, so use the no-bias variant whose\n        // binding layout matches.\n        spec = LAYERNORM_NO_BIAS_SPEC;\n      } else if (node.opType === \"RoPE\" && node.attributes.interleaved === true) {\n        // Moonshine rotates adjacent dim pairs (2p, 2p+1), not the [p, p+half]\n        // split the default kernel uses.\n        spec = ROPE_INTERLEAVED_SPEC;\n      }\n\n      const pipeline = getOrCreatePipeline(\n        this.ctx,\n        `kernel_${nodeId}`,\n        spec.shaderCode,\n        spec.entryPoint,\n      );\n\n      // Create uniform buffer with dummy params (will be overwritten before first dispatch)\n      const dummyShapes = this.resolveShapes(1);\n      const paramsData = spec.buildParams(node, dummyShapes, { seqPos: 0 });\n      const uniformBuffer = createUniformBuffer(this.ctx, `uniform_${nodeId}`, paramsData);\n\n      // Gather buffers for bind group\n      const bufferEntries = this.gatherBuffers(spec, node, uniformBuffer);\n      const bindGroup = createBindGroup(this.ctx, pipeline, bufferEntries, `bg_${nodeId}`);\n\n      const prefillEntry: DispatchEntry = {\n        nodeId,\n        node,\n        spec,\n        pipeline,\n        bindGroup,\n        uniformBuffer,\n        lastParamsBytes: null,\n        lastDispatchSize: null,\n      };\n      this.dispatchEntries.push(prefillEntry);\n\n      // For matmul ops, create a separate decode entry with the matvec kernel\n      const isMatmul = node.opType === \"MatMul\" || node.opType === \"MatMulInt4\";\n      if (isMatmul) {\n        // Subgroups-accelerated matvec reduction. GATED OFF by default: on Dawn\n        // (M-series desktop) it measured a large DECODE REGRESSION (~210→47 tok/s\n        // Qwen3.5-0.8B, ~593→146 tok/s LFM2.5-350M) — coherent output, but the\n        // extra workgroup barriers + per-column subgroup combine outweigh the\n        // tiny saved 16/32-element sequential reduction these matvecs do. The\n        // variant is kept registered and selectable behind an opt-in flag for\n        // future hardware/driver re-evaluation; default path is the portable\n        // kernel. WebKit is excluded regardless (mobile-safety rules).\n        const subgroupsOptIn =\n          (globalThis as { process?: { env?: Record<string, string | undefined> } }).process?.env\n            ?.GERBIL_MATVEC_SUBGROUPS === \"1\";\n        const useSubgroups = subgroupsOptIn && this.ctx.hasSubgroups && !this.ctx.isWebKitWebGPU;\n        const mvSpec =\n          node.opType === \"MatMulInt4\"\n            ? useSubgroups\n              ? MATVEC_INT4_SUBGROUPS_SPEC\n              : MATVEC_INT4_SPEC\n            : useSubgroups\n              ? MATVEC_SUBGROUPS_SPEC\n              : MATVEC_SPEC;\n\n        const mvPipeline = getOrCreatePipeline(\n          this.ctx,\n          `matvec_${nodeId}`,\n          mvSpec.shaderCode,\n          mvSpec.entryPoint,\n        );\n        const mvUniformData = mvSpec.buildParams(node, dummyShapes, { seqPos: 0 });\n        const mvUniform = createUniformBuffer(this.ctx, `mv_uniform_${nodeId}`, mvUniformData);\n        const mvBuffers = this.gatherBuffers(mvSpec, node, mvUniform);\n        const mvBindGroup = createBindGroup(this.ctx, mvPipeline, mvBuffers, `bg_mv_${nodeId}`);\n\n        this.decodeEntries.push({\n          nodeId,\n          node,\n          spec: mvSpec,\n          pipeline: mvPipeline,\n          bindGroup: mvBindGroup,\n          uniformBuffer: mvUniform,\n          lastParamsBytes: null,\n          lastDispatchSize: null,\n        });\n      } else {\n        // Non-matmul ops share the same entry for decode\n        this.decodeEntries.push(prefillEntry);\n      }\n    }\n\n    // ── Fuse gate_proj + up_proj + SwiGLU → single SwiGLUMatVec dispatch ──\n    // Scans decode entries for the pattern: MatMul(input→gate) + MatMul(input→up) + SwiGLU(gate,up→out)\n    // Replaces 3 dispatches with 1 fused kernel that reads input once.\n    this.fuseSwiGLUDecodeEntries();\n\n    // ── Fuse adjacent INT4 projections sharing the same input (q+gate, k+v) ──\n    // Full-attention decode runs q_proj/gate_proj/k_proj/v_proj all reading the\n    // same input_layernorm output. Merge the q+gate pair and the k+v pair into\n    // one DualMatVecInt4 dispatch each, removing one GPU round-trip per pair.\n    this.fuseDualMatVecDecodeEntries();\n\n    // ── Fuse adjacent K+V cache appends (full-attention layers) ──\n    this.fuseDualKVCacheAppendEntries();\n\n    // ── Fuse attention SigmoidGate into the INT4 o_proj that consumes it ──\n    this.fuseGatedOProjDecodeEntries();\n\n    // ── Fuse Mamba SwiGLU (silu(z)*norm) into the INT4 out_proj ──\n    this.fuseSwiGLUGatedProjDecodeEntries();\n\n    // ── Fuse adjacent per-head RMSNorms (q_norm + k_norm) ──\n    this.fuseDualRMSNormDecodeEntries();\n\n    // Decode dispatch count per token — the mobile-proportional metric. On WebKit\n    // each dispatch is a submit+drain round-trip, so decode tok/s ≈ 1/this. Logged\n    // so every fusion's dispatch reduction is measurable on desktop/node.\n    console.log(`[executor] decode: ${this.decodeEntries.length} dispatches/token`);\n\n    // Create argmax bind group\n    const logitsBuffer = this.getBuffer(\"logits\");\n    if (logitsBuffer) {\n      const argmaxPipeline = getOrCreatePipeline(\n        this.ctx,\n        \"argmax\",\n        ARGMAX_SPEC.shaderCode,\n        ARGMAX_SPEC.entryPoint,\n      );\n      const argmaxUniform = createUniformBuffer(\n        this.ctx,\n        \"uniform_argmax\",\n        new ArrayBuffer(16), // {count, offset} — will be overwritten\n      );\n      const argmaxBindGroup = createBindGroup(\n        this.ctx,\n        argmaxPipeline,\n        [{ buffer: logitsBuffer }, { buffer: this.argmaxResultBuffer }, { buffer: argmaxUniform }],\n        \"bg_argmax\",\n      );\n      this.argmaxEntry = {\n        pipeline: argmaxPipeline,\n        bindGroup: argmaxBindGroup,\n        uniformBuffer: argmaxUniform,\n      };\n    }\n\n    if (this.needsMultiEncoder) {\n      console.log(`[executor] WebKit WebGPU detected — using per-dispatch submit`);\n    }\n\n    // Create staging buffer for uniform param updates.\n    // Size = max(total prefill uniform bytes, total decode uniform bytes + argmax).\n    let prefillTotal = 0;\n    for (const entry of this.dispatchEntries) {\n      prefillTotal += Math.ceil(entry.uniformBuffer.size / 4) * 4;\n    }\n    let decodeTotal = 0;\n    for (const entry of this.decodeEntries) {\n      decodeTotal += Math.ceil(entry.uniformBuffer.size / 4) * 4;\n    }\n    decodeTotal += 16; // argmax params (8 bytes, padded)\n    // Add space for input_ids (routed through staging too, for Safari safety)\n    const inputIdBytes = this.maxSeqLen * 4;\n    this.uniformStagingCapacity = Math.max(prefillTotal, decodeTotal) + inputIdBytes;\n    this.uniformStagingBuffer = this.ctx.device.createBuffer({\n      label: \"uniform_staging\",\n      size: this.uniformStagingCapacity,\n      usage: 0x0004 | 0x0008, // COPY_SRC | COPY_DST\n    });\n  }\n\n  /**\n   * Run a forward pass. Uses matvec kernels for M=1 (decode), tiled for M>1 (prefill).\n   */\n  async forward(inputIds: Uint32Array): Promise<ForwardResult> {\n    const T = inputIds.length;\n\n    const resolvedShapes = this.resolveShapes(T);\n    const runtimeContext: RuntimeContext = { seqPos: this.seqPos };\n\n    // Use matvec entries for single-token decode, tiled for prefill\n    const entries = T === 1 ? this.decodeEntries : this.dispatchEntries;\n\n    // Write input_ids directly — no staging buffer.\n    this.ctx.device.queue.writeBuffer(this.inputIdsBuffer, 0, inputIds as BufferSource);\n\n    // Gemma 4: stream the CPU-resident PLE rows for these tokens (no-op otherwise).\n    await this.streamPleRows(inputIds);\n\n    // Build params and write directly to each uniform buffer.\n    // Also compute dispatch sizes up front.\n    const dispatchSizes: Array<[number, number, number]> = new Array(entries.length);\n    for (let i = 0; i < entries.length; i++) {\n      const entry = entries[i];\n      const paramsData = entry.spec.buildParams(entry.node, resolvedShapes, runtimeContext);\n      const paramsView = new Uint8Array(paramsData);\n\n      let changed = !entry.lastParamsBytes || entry.lastParamsBytes.length !== paramsView.length;\n      if (!changed) {\n        const cached = entry.lastParamsBytes!;\n        for (let j = 0; j < paramsView.length; j++) {\n          if (paramsView[j] !== cached[j]) {\n            changed = true;\n            break;\n          }\n        }\n      }\n\n      if (changed) {\n        this.ctx.device.queue.writeBuffer(entry.uniformBuffer, 0, paramsData);\n        entry.lastParamsBytes = new Uint8Array(paramsView.slice(0));\n      }\n\n      dispatchSizes[i] = entry.spec.getDispatchSize(entry.node, resolvedShapes, runtimeContext);\n    }\n\n    if (this.profileEnabled) {\n      await this.runProfiledDispatches(entries, dispatchSizes);\n    } else if (this.needsMultiEncoder) {\n      // WebKit: group dispatches into command buffers of webkitGroupSize\n      // (one compute pass per dispatch), with at most ONE command buffer in\n      // flight (awaited). Fire-and-forget per-dispatch submits queue ~400\n      // unbounded in-flight Metal command buffers per token — the documented\n      // WebKit resource-exhaustion anti-pattern (WebKit bug 311598). Batching\n      // many dispatches into one submission zeroed activations on iPad\n      // (docs/metal-safari-intel.md); sweep ?group=N to find this WebKit\n      // version's correct/fast granularity (llama.cpp runs ~64/CB on iOS 26.4).\n      const group = this.webkitGroupSize;\n      for (let start = 0; start < entries.length; start += group) {\n        const enc = this.ctx.device.createCommandEncoder();\n        const end = Math.min(start + group, entries.length);\n        for (let i = start; i < end; i++) {\n          const p = enc.beginComputePass();\n          p.setPipeline(entries[i].pipeline);\n          p.setBindGroup(0, entries[i].bindGroup);\n          p.dispatchWorkgroups(...dispatchSizes[i]);\n          p.end();\n        }\n        this.ctx.device.queue.submit([enc.finish()]);\n        await this.ctx.device.queue.onSubmittedWorkDone();\n      }\n      const logitsBuffer = this.getBuffer(\"logits\");\n      if (logitsBuffer) {\n        const copyEnc = this.ctx.device.createCommandEncoder();\n        const vocabSize = this.graph.config.vocab_size;\n        // logits is [1, vocab] — SliceLastRow feeds lm_head only the last position\n        copyEnc.copyBufferToBuffer(logitsBuffer, 0, this.logitsReadback, 0, vocabSize * 4);\n        this.ctx.device.queue.submit([copyEnc.finish()]);\n      }\n    } else {\n      // Desktop/Dawn: single encoder, single compute pass\n      const encoder = this.ctx.device.createCommandEncoder({ label: \"forward\" });\n      const pass = encoder.beginComputePass({ label: \"fwd_pass\" });\n      for (let i = 0; i < entries.length; i++) {\n        pass.setPipeline(entries[i].pipeline);\n        pass.setBindGroup(0, entries[i].bindGroup);\n        pass.dispatchWorkgroups(...dispatchSizes[i]);\n      }\n      pass.end();\n      const logitsBuffer = this.getBuffer(\"logits\");\n      if (logitsBuffer) {\n        const vocabSize = this.graph.config.vocab_size;\n        // logits is [1, vocab] — SliceLastRow feeds lm_head only the last position\n        encoder.copyBufferToBuffer(logitsBuffer, 0, this.logitsReadback, 0, vocabSize * 4);\n      }\n      this.ctx.device.queue.submit([encoder.finish()]);\n    }\n\n    const vocabSize = this.graph.config.vocab_size;\n    await this.logitsReadback.mapAsync(MAP_MODE_READ, 0, vocabSize * 4);\n    const mapped = this.logitsReadback.getMappedRange(0, vocabSize * 4);\n    const logits = new Float32Array(mapped.slice(0));\n    this.logitsReadback.unmap();\n\n    this.seqPos += T;\n    return { logits };\n  }\n\n  /**\n   * Profiling variant of the desktop dispatch path: one compute pass per dispatch,\n   * each bracketed by timestamp queries, so we get per-op GPU time. Accumulates\n   * into profileData by opType. Only runs under GERBIL_PROFILE with timestamp-query\n   * support — slower than the batched path (it measures relative cost, not tok/s).\n   */\n  private async runProfiledDispatches(\n    entries: DispatchEntry[],\n    dispatchSizes: number[][],\n    copyLogits = true,\n  ): Promise<void> {\n    const n = entries.length;\n    const queryCount = n * 2;\n    if (!this.querySet) {\n      this.querySet = this.ctx.device.createQuerySet({ type: \"timestamp\", count: queryCount });\n      // QUERY_RESOLVE (0x0200) | COPY_SRC (0x0004)\n      this.queryResolveBuf = this.ctx.device.createBuffer({\n        size: queryCount * 8,\n        usage: 0x0200 | 0x0004,\n      });\n      // MAP_READ (0x0001) | COPY_DST (0x0008)\n      this.queryReadbackBuf = this.ctx.device.createBuffer({\n        size: queryCount * 8,\n        usage: 0x0001 | 0x0008,\n      });\n    }\n    const querySet = this.querySet;\n    const resolveBuf = this.queryResolveBuf as GPUBuffer;\n    const readbackBuf = this.queryReadbackBuf as GPUBuffer;\n\n    const encoder = this.ctx.device.createCommandEncoder({ label: \"fwd_profiled\" });\n    for (let i = 0; i < n; i++) {\n      const pass = encoder.beginComputePass({\n        timestampWrites: {\n          querySet,\n          beginningOfPassWriteIndex: i * 2,\n          endOfPassWriteIndex: i * 2 + 1,\n        },\n      });\n      pass.setPipeline(entries[i].pipeline);\n      pass.setBindGroup(0, entries[i].bindGroup);\n      pass.dispatchWorkgroups(dispatchSizes[i][0], dispatchSizes[i][1], dispatchSizes[i][2]);\n      pass.end();\n    }\n    if (copyLogits) {\n      const logitsBuffer = this.getBuffer(\"logits\");\n      const vocabSize = this.graph.config.vocab_size;\n      if (logitsBuffer) {\n        encoder.copyBufferToBuffer(logitsBuffer, 0, this.logitsReadback, 0, vocabSize * 4);\n      }\n    }\n    encoder.resolveQuerySet(querySet, 0, queryCount, resolveBuf, 0);\n    encoder.copyBufferToBuffer(resolveBuf, 0, readbackBuf, 0, queryCount * 8);\n    this.ctx.device.queue.submit([encoder.finish()]);\n\n    await readbackBuf.mapAsync(MAP_MODE_READ, 0, queryCount * 8);\n    const ts = new BigUint64Array(readbackBuf.getMappedRange(0, queryCount * 8).slice(0));\n    readbackBuf.unmap();\n    for (let i = 0; i < n; i++) {\n      const ns = Number(ts[i * 2 + 1] - ts[i * 2]);\n      if (ns <= 0) {\n        continue;\n      }\n      const key = entries[i].node.opType;\n      const cur = this.profileData.get(key) ?? { ns: 0, count: 0 };\n      cur.ns += ns;\n      cur.count += 1;\n      this.profileData.set(key, cur);\n    }\n  }\n\n  /** Per-opType GPU time (ns) + dispatch count accumulated by GERBIL_PROFILE, hottest first. */\n  getProfile(): Array<{ opType: string; ns: number; count: number }> {\n    return [...this.profileData.entries()]\n      .map(([opType, v]) => ({ opType, ns: v.ns, count: v.count }))\n      .sort((a, b) => b.ns - a.ns);\n  }\n\n  /** Clear accumulated profiler data (e.g. to drop warm-up tokens). */\n  resetProfile(): void {\n    this.profileData.clear();\n  }\n\n  /** GPU dispatches per decode token (post-fusion). On mobile this drives the\n   * submit-group count = ceil(dispatchCount / webkitGroupSize). */\n  get decodeDispatchCount(): number {\n    return this.decodeEntries.length;\n  }\n\n  /** Device limit that gates the INT4 projection fusions (they need ≥9). If a\n   * device caps at 8 the dual/gated/swiglu-gated INT4 fusions silently fall back,\n   * inflating the decode dispatch count. */\n  get maxStorageBuffers(): number {\n    return this.ctx.limits.maxStorageBuffersPerShaderStage;\n  }\n\n  /**\n   * Profile ONE real decode step: times the actual `decodeEntries` (the kernels\n   * the pipelined greedy benchmark runs) with per-dispatch timestamps. Timing is\n   * token-independent, so pass any valid id; runs un-pipelined with a synchronous\n   * timestamp readback (measurement only, not for production decode). Argmax (one\n   * tiny dispatch) is intentionally excluded — it is not a hotspot target.\n   */\n  async profileDecodeStep(tokenId: number): Promise<void> {\n    const resolvedShapes = this.resolveShapes(1);\n    const runtimeContext: RuntimeContext = { seqPos: this.seqPos };\n    this.ctx.device.queue.writeBuffer(this.inputIdsBuffer, 0, new Uint32Array([tokenId]));\n\n    const dispatchSizes: number[][] = new Array(this.decodeEntries.length);\n    for (let i = 0; i < this.decodeEntries.length; i++) {\n      const entry = this.decodeEntries[i];\n      const paramsData = entry.spec.buildParams(entry.node, resolvedShapes, runtimeContext);\n      this.ctx.device.queue.writeBuffer(entry.uniformBuffer, 0, paramsData);\n      dispatchSizes[i] = entry.spec.getDispatchSize(entry.node, resolvedShapes, runtimeContext);\n    }\n    await this.runProfiledDispatches(this.decodeEntries, dispatchSizes, false);\n    this.seqPos += 1;\n  }\n\n  /**\n   * Run a single forward pass over `inputIds` and read back the L2-normalized\n   * embedding vector. Requires an embedding graph (one whose output tensor is\n   * \"embedding\", produced by the last-token-pool + L2-norm tail).\n   *\n   * Always runs in a fresh-state single pass (caller should reset() first):\n   * embeddings are non-autoregressive, so the whole sequence is one prefill.\n   */\n  async embed(inputIds: Uint32Array): Promise<Float32Array> {\n    const embeddingBuffer = this.getBuffer(\"embedding\");\n    if (!embeddingBuffer) {\n      throw new Error(\n        \"embed() requires an embedding graph (no 'embedding' output tensor found). \" +\n          \"Load the model with { embedding: true }.\",\n      );\n    }\n    const hiddenSize = this.graph.config.hidden_size;\n    const byteLen = hiddenSize * 4;\n\n    const T = inputIds.length;\n    const resolvedShapes = this.resolveShapes(T);\n    const runtimeContext: RuntimeContext = { seqPos: this.seqPos };\n\n    // Embeddings process the full sequence at once → always use prefill entries.\n    const entries = T === 1 ? this.decodeEntries : this.dispatchEntries;\n\n    this.ctx.device.queue.writeBuffer(this.inputIdsBuffer, 0, inputIds as BufferSource);\n\n    const dispatchSizes: Array<[number, number, number]> = new Array(entries.length);\n    for (let i = 0; i < entries.length; i++) {\n      const entry = entries[i];\n      const paramsData = entry.spec.buildParams(entry.node, resolvedShapes, runtimeContext);\n      const paramsView = new Uint8Array(paramsData);\n\n      let changed = !entry.lastParamsBytes || entry.lastParamsBytes.length !== paramsView.length;\n      if (!changed) {\n        const cached = entry.lastParamsBytes!;\n        for (let j = 0; j < paramsView.length; j++) {\n          if (paramsView[j] !== cached[j]) {\n            changed = true;\n            break;\n          }\n        }\n      }\n      if (changed) {\n        this.ctx.device.queue.writeBuffer(entry.uniformBuffer, 0, paramsData);\n        entry.lastParamsBytes = new Uint8Array(paramsView.slice(0));\n      }\n\n      dispatchSizes[i] = entry.spec.getDispatchSize(entry.node, resolvedShapes, runtimeContext);\n    }\n\n    // Dedicated readback buffer (COPY_DST | MAP_READ).\n    const readback = this.ctx.device.createBuffer({ size: byteLen, usage: 0x0001 | 0x0008 });\n\n    if (this.needsMultiEncoder) {\n      const group = this.webkitGroupSize;\n      for (let start = 0; start < entries.length; start += group) {\n        const enc = this.ctx.device.createCommandEncoder();\n        const end = Math.min(start + group, entries.length);\n        for (let i = start; i < end; i++) {\n          const p = enc.beginComputePass();\n          p.setPipeline(entries[i].pipeline);\n          p.setBindGroup(0, entries[i].bindGroup);\n          p.dispatchWorkgroups(...dispatchSizes[i]);\n          p.end();\n        }\n        this.ctx.device.queue.submit([enc.finish()]);\n        await this.ctx.device.queue.onSubmittedWorkDone();\n      }\n      const copyEnc = this.ctx.device.createCommandEncoder();\n      copyEnc.copyBufferToBuffer(embeddingBuffer, 0, readback, 0, byteLen);\n      this.ctx.device.queue.submit([copyEnc.finish()]);\n    } else {\n      const encoder = this.ctx.device.createCommandEncoder({ label: \"embed\" });\n      const pass = encoder.beginComputePass({ label: \"embed_pass\" });\n      for (let i = 0; i < entries.length; i++) {\n        pass.setPipeline(entries[i].pipeline);\n        pass.setBindGroup(0, entries[i].bindGroup);\n        pass.dispatchWorkgroups(...dispatchSizes[i]);\n      }\n      pass.end();\n      encoder.copyBufferToBuffer(embeddingBuffer, 0, readback, 0, byteLen);\n      this.ctx.device.queue.submit([encoder.finish()]);\n    }\n\n    await readback.mapAsync(MAP_MODE_READ, 0, byteLen);\n    const out = new Float32Array(readback.getMappedRange(0, byteLen).slice(0));\n    readback.unmap();\n    readback.destroy();\n\n    this.seqPos += T;\n    return out;\n  }\n\n  /**\n   * Generic one-shot dispatch for a non-LM graph: run every entry once over a\n   * single forward and read back `elemCount` f32 elements of the named output\n   * tensor. Used to execute the NanoCodec decoder graph (codes→PCM) — its ops use\n   * concrete lengths and it has no \"logits\" output, so the normal forward()\n   * logits-readback path does not apply. Caller writes any inputs (e.g. audio_codes)\n   * via writeInput() and reset()s first.\n   */\n  async runGraphOutput(outputName: string, elemCount: number): Promise<Float32Array> {\n    const outBuffer = this.getBuffer(outputName);\n    if (!outBuffer) {\n      throw new Error(`runGraphOutput: no GPU buffer for output tensor \"${outputName}\".`);\n    }\n    const byteLen = elemCount * 4;\n    // Codec graphs carry concrete lengths; dispatchEntries (M>1 path) builds the\n    // full pipeline. Pass T=1 so shape resolution keeps any symbolic dims minimal\n    // (the codec ops ignore input_ids entirely).\n    const resolvedShapes = this.resolveShapes(1);\n    const runtimeContext: RuntimeContext = { seqPos: this.seqPos };\n    const entries = this.dispatchEntries;\n\n    const dispatchSizes: Array<[number, number, number]> = new Array(entries.length);\n    for (let i = 0; i < entries.length; i++) {\n      const entry = entries[i];\n      const paramsData = entry.spec.buildParams(entry.node, resolvedShapes, runtimeContext);\n      this.ctx.device.queue.writeBuffer(entry.uniformBuffer, 0, paramsData);\n      entry.lastParamsBytes = new Uint8Array(new Uint8Array(paramsData).slice(0));\n      dispatchSizes[i] = entry.spec.getDispatchSize(entry.node, resolvedShapes, runtimeContext);\n    }\n\n    const readback = this.ctx.device.createBuffer({ size: byteLen, usage: 0x0001 | 0x0008 });\n\n    if (this.needsMultiEncoder) {\n      const group = this.webkitGroupSize;\n      for (let start = 0; start < entries.length; start += group) {\n        const enc = this.ctx.device.createCommandEncoder();\n        const end = Math.min(start + group, entries.length);\n        for (let i = start; i < end; i++) {\n          const pass = enc.beginComputePass();\n          pass.setPipeline(entries[i].pipeline);\n          pass.setBindGroup(0, entries[i].bindGroup);\n          pass.dispatchWorkgroups(...dispatchSizes[i]);\n          pass.end();\n        }\n        this.ctx.device.queue.submit([enc.finish()]);\n        await this.ctx.device.queue.onSubmittedWorkDone();\n      }\n      const copyEnc = this.ctx.device.createCommandEncoder();\n      copyEnc.copyBufferToBuffer(outBuffer, 0, readback, 0, byteLen);\n      this.ctx.device.queue.submit([copyEnc.finish()]);\n    } else {\n      const encoder = this.ctx.device.createCommandEncoder({ label: \"runGraphOutput\" });\n      const pass = encoder.beginComputePass({ label: \"run_pass\" });\n      for (let i = 0; i < entries.length; i++) {\n        pass.setPipeline(entries[i].pipeline);\n        pass.setBindGroup(0, entries[i].bindGroup);\n        pass.dispatchWorkgroups(...dispatchSizes[i]);\n      }\n      pass.end();\n      encoder.copyBufferToBuffer(outBuffer, 0, readback, 0, byteLen);\n      this.ctx.device.queue.submit([encoder.finish()]);\n    }\n\n    await readback.mapAsync(MAP_MODE_READ, 0, byteLen);\n    const out = new Float32Array(readback.getMappedRange(0, byteLen).slice(0));\n    readback.unmap();\n    readback.destroy();\n    return out;\n  }\n\n  /**\n   * Greedy decode step: forward + GPU argmax. Returns token ID directly.\n   * Always uses matvec kernels (M=1). Reads back 4 bytes instead of vocab_size*4.\n   */\n  async forwardArgmax(inputIds: Uint32Array): Promise<number> {\n    const T = inputIds.length;\n\n    const resolvedShapes = this.resolveShapes(T);\n    const runtimeContext: RuntimeContext = { seqPos: this.seqPos };\n\n    // Write input_ids directly\n    this.ctx.device.queue.writeBuffer(this.inputIdsBuffer, 0, inputIds as BufferSource);\n\n    // Write uniform params directly to each buffer\n    const argmaxDispatchSizes: Array<[number, number, number]> = new Array(\n      this.decodeEntries.length,\n    );\n    for (let i = 0; i < this.decodeEntries.length; i++) {\n      const entry = this.decodeEntries[i];\n      const paramsData = entry.spec.buildParams(entry.node, resolvedShapes, runtimeContext);\n      const paramsView = new Uint8Array(paramsData);\n\n      let changed = !entry.lastParamsBytes || entry.lastParamsBytes.length !== paramsView.length;\n      if (!changed) {\n        const cached = entry.lastParamsBytes!;\n        for (let j = 0; j < paramsView.length; j++) {\n          if (paramsView[j] !== cached[j]) {\n            changed = true;\n            break;\n          }\n        }\n      }\n\n      if (changed) {\n        this.ctx.device.queue.writeBuffer(entry.uniformBuffer, 0, paramsData);\n        entry.lastParamsBytes = new Uint8Array(paramsView.slice(0));\n      }\n\n      argmaxDispatchSizes[i] = entry.spec.getDispatchSize(\n        entry.node,\n        resolvedShapes,\n        runtimeContext,\n      );\n    }\n\n    // Write argmax params directly.\n    // logits is [1, vocab] (SliceLastRow feeds lm_head only the last position),\n    // so the argmax always reads row 0.\n    if (this.argmaxEntry) {\n      const vocabSize = this.graph.config.vocab_size;\n      const argmaxParams = new ArrayBuffer(8);\n      const view = new DataView(argmaxParams);\n      view.setUint32(0, vocabSize, true);\n      view.setUint32(4, 0, true);\n      this.ctx.device.queue.writeBuffer(this.argmaxEntry.uniformBuffer, 0, argmaxParams);\n    }\n\n    if (this.needsMultiEncoder) {\n      // WebKit: grouped submits with one command buffer in flight\n      // (see forward() comment for rationale)\n      const group = this.webkitGroupSize;\n      for (let start = 0; start < this.decodeEntries.length; start += group) {\n        const enc = this.ctx.device.createCommandEncoder();\n        const end = Math.min(start + group, this.decodeEntries.length);\n        for (let i = start; i < end; i++) {\n          const p = enc.beginComputePass();\n          p.setPipeline(this.decodeEntries[i].pipeline);\n          p.setBindGroup(0, this.decodeEntries[i].bindGroup);\n          p.dispatchWorkgroups(...argmaxDispatchSizes[i]);\n          p.end();\n        }\n        this.ctx.device.queue.submit([enc.finish()]);\n        await this.ctx.device.queue.onSubmittedWorkDone();\n      }\n      // Argmax dispatch and readback copy in SEPARATE submits, mirroring\n      // forward(): the copy reads argmaxResultBuffer written by the dispatch,\n      // so it must sit behind a command buffer boundary on WebKit.\n      if (this.argmaxEntry) {\n        const enc = this.ctx.device.createCommandEncoder();\n        const p = enc.beginComputePass();\n        p.setPipeline(this.argmaxEntry.pipeline);\n        p.setBindGroup(0, this.argmaxEntry.bindGroup);\n        p.dispatchWorkgroups(1, 1, 1);\n        p.end();\n        this.ctx.device.queue.submit([enc.finish()]);\n\n        const copyEnc = this.ctx.device.createCommandEncoder();\n        copyEnc.copyBufferToBuffer(this.argmaxResultBuffer, 0, this.argmaxReadback, 0, 4);\n        this.ctx.device.queue.submit([copyEnc.finish()]);\n      }\n    } else {\n      // Desktop/Dawn: single encoder, single compute pass\n      const encoder = this.ctx.device.createCommandEncoder({ label: \"argmax\" });\n      const pass = encoder.beginComputePass({ label: \"am_pass\" });\n      for (let i = 0; i < this.decodeEntries.length; i++) {\n        const entry = this.decodeEntries[i];\n        pass.setPipeline(entry.pipeline);\n        pass.setBindGroup(0, entry.bindGroup);\n        pass.dispatchWorkgroups(...argmaxDispatchSizes[i]);\n      }\n      if (this.argmaxEntry) {\n        pass.setPipeline(this.argmaxEntry.pipeline);\n        pass.setBindGroup(0, this.argmaxEntry.bindGroup);\n        pass.dispatchWorkgroups(1, 1, 1);\n      }\n      pass.end();\n      encoder.copyBufferToBuffer(this.argmaxResultBuffer, 0, this.argmaxReadback, 0, 4);\n      this.ctx.device.queue.submit([encoder.finish()]);\n    }\n\n    await this.argmaxReadback.mapAsync(MAP_MODE_READ, 0, 4);\n    const mapped = this.argmaxReadback.getMappedRange(0, 4);\n    const tokenId = new Uint32Array(mapped.slice(0))[0];\n    this.argmaxReadback.unmap();\n\n    this.seqPos += T;\n    return tokenId;\n  }\n\n  /** KV-cache positions still available for decode steps. */\n  decodeCapacityRemaining(): number {\n    return this.maxSeqLen - this.seqPos;\n  }\n\n  /** Number of decode steps that may be in flight in the pipelined path. */\n  static readonly PIPELINE_DEPTH = 2;\n\n  /**\n   * Pipelined greedy decode step (Dawn only — WebKit uses forwardArgmax).\n   *\n   * Encodes one full decode forward + argmax and submits WITHOUT awaiting\n   * completion. The input token is taken from `tokenId` for the first step\n   * after prefill; for subsequent steps (tokenId === null) the previous step's\n   * argmax result is copied into input_ids ON THE GPU, so the decode loop\n   * never blocks on a readback before submitting the next step. The argmax\n   * result is copied to a per-slot readback buffer, read later (one step\n   * behind) via readDecodeToken(slot).\n   *\n   * queue.writeBuffer is queue-ordered: uniform updates land after the\n   * previously submitted step's command buffer and before this one's, so\n   * shared uniform buffers are safe with multiple steps in flight.\n   */\n  submitGreedyDecodeStep(tokenId: number | null, slot: number): void {\n    const resolvedShapes = this.resolveShapes(1);\n    const runtimeContext: RuntimeContext = { seqPos: this.seqPos };\n\n    if (this.decodeReadbacks.length === 0) {\n      for (let i = 0; i < Executor.PIPELINE_DEPTH; i++) {\n        this.decodeReadbacks.push(createReadbackBuffer(this.ctx, `decode_readback_${i}`, 4));\n      }\n    }\n\n    if (tokenId !== null) {\n      this.ctx.device.queue.writeBuffer(this.inputIdsBuffer, 0, new Uint32Array([tokenId]));\n      // Argmax params are decode-invariant — write once per generation.\n      if (this.argmaxEntry) {\n        const argmaxParams = new ArrayBuffer(8);\n        const view = new DataView(argmaxParams);\n        view.setUint32(0, this.graph.config.vocab_size, true);\n        view.setUint32(4, 0, true);\n        this.ctx.device.queue.writeBuffer(this.argmaxEntry.uniformBuffer, 0, argmaxParams);\n      }\n    }\n\n    // Update uniforms (same skip-unchanged caching as forwardArgmax)\n    const dispatchSizes: Array<[number, number, number]> = new Array(this.decodeEntries.length);\n    for (let i = 0; i < this.decodeEntries.length; i++) {\n      const entry = this.decodeEntries[i];\n      const paramsData = entry.spec.buildParams(entry.node, resolvedShapes, runtimeContext);\n      const paramsView = new Uint8Array(paramsData);\n\n      let changed = !entry.lastParamsBytes || entry.lastParamsBytes.length !== paramsView.length;\n      if (!changed) {\n        const cached = entry.lastParamsBytes!;\n        for (let j = 0; j < paramsView.length; j++) {\n          if (paramsView[j] !== cached[j]) {\n            changed = true;\n            break;\n          }\n        }\n      }\n\n      if (changed) {\n        this.ctx.device.queue.writeBuffer(entry.uniformBuffer, 0, paramsData);\n        entry.lastParamsBytes = new Uint8Array(paramsView.slice(0));\n      }\n\n      dispatchSizes[i] = entry.spec.getDispatchSize(entry.node, resolvedShapes, runtimeContext);\n    }\n\n    const encoder = this.ctx.device.createCommandEncoder({ label: \"decode_step\" });\n    if (tokenId === null) {\n      // Chain: previous step's argmax result becomes this step's input token.\n      encoder.copyBufferToBuffer(this.argmaxResultBuffer, 0, this.inputIdsBuffer, 0, 4);\n    }\n    const pass = encoder.beginComputePass({ label: \"decode_pass\" });\n    for (let i = 0; i < this.decodeEntries.length; i++) {\n      const entry = this.decodeEntries[i];\n      pass.setPipeline(entry.pipeline);\n      pass.setBindGroup(0, entry.bindGroup);\n      pass.dispatchWorkgroups(...dispatchSizes[i]);\n    }\n    if (this.argmaxEntry) {\n      pass.setPipeline(this.argmaxEntry.pipeline);\n      pass.setBindGroup(0, this.argmaxEntry.bindGroup);\n      pass.dispatchWorkgroups(1, 1, 1);\n    }\n    pass.end();\n    encoder.copyBufferToBuffer(this.argmaxResultBuffer, 0, this.decodeReadbacks[slot], 0, 4);\n    this.ctx.device.queue.submit([encoder.finish()]);\n\n    this.seqPos += 1;\n  }\n\n  /** Read back the token produced by the pipelined step that used `slot`. */\n  async readDecodeToken(slot: number): Promise<number> {\n    const buf = this.decodeReadbacks[slot];\n    await buf.mapAsync(MAP_MODE_READ, 0, 4);\n    const mapped = buf.getMappedRange(0, 4);\n    const tokenId = new Uint32Array(mapped.slice(0))[0];\n    buf.unmap();\n    return tokenId;\n  }\n\n  reset(): void {\n    this.seqPos = 0;\n    // Zero SSM state buffers\n    for (const buf of this.ssmStateBuffers.values()) {\n      this.ctx.device.queue.writeBuffer(buf, 0, new Uint8Array(buf.size));\n    }\n    // Clear cached params — shapes change between prefill (T=prompt_len) and decode (T=1)\n    for (const entry of this.dispatchEntries) {\n      entry.lastParamsBytes = null;\n      entry.lastDispatchSize = null;\n    }\n    for (const entry of this.decodeEntries) {\n      entry.lastParamsBytes = null;\n      entry.lastDispatchSize = null;\n    }\n  }\n\n  /**\n   * Diagnostic: dispatch ONLY the first kernel (EmbeddingInt4) using the\n   * production bind group, pipeline, and buffers — but in isolation (1 dispatch,\n   * no staging, fresh encoder). Compares against full forward pass to isolate\n   * whether the issue is the bind group/pipeline or the multi-dispatch context.\n   */\n  async debugFirstDispatch(inputIds: Uint32Array): Promise<{\n    nodeId: string;\n    opType: string;\n    dispatchSize: [number, number, number];\n    output: Float32Array;\n  }> {\n    const T = inputIds.length;\n    const resolvedShapes = this.resolveShapes(T);\n    const runtimeContext: RuntimeContext = { seqPos: this.seqPos };\n    const entries = T === 1 ? this.decodeEntries : this.dispatchEntries;\n    const entry = entries[0];\n\n    // Write input_ids directly (no staging)\n    this.ctx.device.queue.writeBuffer(this.inputIdsBuffer, 0, inputIds as BufferSource);\n\n    // Write params directly to uniform buffer (no staging)\n    const paramsData = entry.spec.buildParams(entry.node, resolvedShapes, runtimeContext);\n    this.ctx.device.queue.writeBuffer(entry.uniformBuffer, 0, paramsData);\n\n    // Single dispatch with fresh encoder\n    const dispatchSize = entry.spec.getDispatchSize(entry.node, resolvedShapes, runtimeContext) as [\n      number,\n      number,\n      number,\n    ];\n    const encoder = this.ctx.device.createCommandEncoder({ label: \"debug_first\" });\n    const pass = encoder.beginComputePass({ label: \"debug_first_pass\" });\n    pass.setPipeline(entry.pipeline);\n    pass.setBindGroup(0, entry.bindGroup);\n    pass.dispatchWorkgroups(...dispatchSize);\n    pass.end();\n    this.ctx.device.queue.submit([encoder.finish()]);\n\n    // Read back the output tensor (first 16 elements)\n    const outputName = entry.node.outputs[0];\n    const output = await this.debugReadBuffer(outputName, 16);\n\n    return {\n      nodeId: entry.nodeId,\n      opType: entry.node.opType,\n      dispatchSize,\n      output,\n    };\n  }\n\n  /**\n   * Diagnostic: run a single dispatch entry by index, in isolation.\n   * Call after debugFirstDispatch() to test whether entry[1] (RMSNorm)\n   * can read embed_out written by entry[0] (EmbeddingInt4).\n   */\n  async debugDispatchEntry(\n    entryIndex: number,\n    T: number,\n  ): Promise<{ nodeId: string; opType: string; output: Float32Array }> {\n    const resolvedShapes = this.resolveShapes(T);\n    const runtimeContext: RuntimeContext = { seqPos: this.seqPos };\n    const entries = T === 1 ? this.decodeEntries : this.dispatchEntries;\n    const entry = entries[entryIndex];\n\n    // Write params directly\n    const paramsData = entry.spec.buildParams(entry.node, resolvedShapes, runtimeContext);\n    this.ctx.device.queue.writeBuffer(entry.uniformBuffer, 0, paramsData);\n\n    // Single dispatch\n    const dispatchSize = entry.spec.getDispatchSize(entry.node, resolvedShapes, runtimeContext) as [\n      number,\n      number,\n      number,\n    ];\n    const encoder = this.ctx.device.createCommandEncoder({ label: `debug_entry_${entryIndex}` });\n    const pass = encoder.beginComputePass({ label: `debug_pass_${entryIndex}` });\n    pass.setPipeline(entry.pipeline);\n    pass.setBindGroup(0, entry.bindGroup);\n    pass.dispatchWorkgroups(...dispatchSize);\n    pass.end();\n    this.ctx.device.queue.submit([encoder.finish()]);\n\n    const outputName = entry.node.outputs[0];\n    const output = await this.debugReadBuffer(outputName, 16);\n    return { nodeId: entry.nodeId, opType: entry.node.opType, output };\n  }\n\n  /**\n   * Diagnostic: compute the JS-side params for the first N decode entries\n   * WITHOUT dispatching. Shows what buildParams produces.\n   */\n  debugComputeParams(\n    T: number,\n    count = 5,\n  ): Array<{\n    idx: number;\n    nodeId: string;\n    opType: string;\n    paramsU32: number[];\n    dispatchSize: [number, number, number];\n  }> {\n    const resolvedShapes = this.resolveShapes(T);\n    const runtimeContext: RuntimeContext = { seqPos: this.seqPos };\n    const entries = T === 1 ? this.decodeEntries : this.dispatchEntries;\n    const results = [];\n    for (let i = 0; i < Math.min(count, entries.length); i++) {\n      const entry = entries[i];\n      const paramsData = entry.spec.buildParams(entry.node, resolvedShapes, runtimeContext);\n      const u32 = new Uint32Array(paramsData);\n      const dispatchSize = entry.spec.getDispatchSize(\n        entry.node,\n        resolvedShapes,\n        runtimeContext,\n      ) as [number, number, number];\n      results.push({\n        idx: i,\n        nodeId: entry.nodeId,\n        opType: entry.node.opType,\n        paramsU32: Array.from(u32),\n        dispatchSize,\n      });\n    }\n    return results;\n  }\n\n  /**\n   * Diagnostic: after a forward pass, read back output tensors at several points\n   * in the pipeline to find where data drops to zero.\n   */\n  async debugPipelineProbe(T: number): Promise<\n    Array<{\n      idx: number;\n      nodeId: string;\n      opType: string;\n      tensor: string;\n      sum: number;\n      first4: number[];\n      uniformParams?: number[];\n    }>\n  > {\n    const entries = T === 1 ? this.decodeEntries : this.dispatchEntries;\n    // Focus on the first few entries to find where zeros start\n    const totalEntries = entries.length;\n    const indices = [\n      0,\n      1,\n      2,\n      3,\n      Math.floor(totalEntries / 4),\n      Math.floor(totalEntries / 2),\n      Math.floor((totalEntries * 3) / 4),\n      totalEntries - 1,\n    ];\n    const unique = [...new Set(indices)].filter((i) => i >= 0 && i < totalEntries);\n\n    const results: Array<{\n      idx: number;\n      nodeId: string;\n      opType: string;\n      tensor: string;\n      sum: number;\n      first4: number[];\n      uniformParams?: number[];\n    }> = [];\n    for (const i of unique) {\n      const entry = entries[i];\n      const outputName = entry.node.outputs[0];\n      if (!outputName) continue;\n      try {\n        const data = await this.debugReadBuffer(outputName, 16);\n        let sum = 0;\n        for (let j = 0; j < data.length; j++) sum += data[j];\n\n        // Also read back the uniform buffer to verify params\n        let uniformParams: number[] | undefined;\n        try {\n          const ubSize = entry.uniformBuffer.size;\n          const readback = this.ctx.device.createBuffer({\n            size: ubSize,\n            usage: 0x0001 | 0x0008, // MAP_READ | COPY_DST\n          });\n          const enc = this.ctx.device.createCommandEncoder();\n          enc.copyBufferToBuffer(entry.uniformBuffer, 0, readback, 0, ubSize);\n          this.ctx.device.queue.submit([enc.finish()]);\n          await readback.mapAsync(MAP_MODE_READ, 0, ubSize);\n          const u32 = new Uint32Array(readback.getMappedRange(0, ubSize).slice(0));\n          uniformParams = Array.from(u32);\n          readback.unmap();\n          readback.destroy();\n        } catch {\n          /* skip if uniform readback fails */\n        }\n\n        results.push({\n          idx: i,\n          nodeId: entry.nodeId,\n          opType: entry.node.opType,\n          tensor: outputName,\n          sum: Math.round(sum * 1e6) / 1e6,\n          first4: [data[0], data[1], data[2], data[3]].map((v) => Math.round(v * 1e6) / 1e6),\n          uniformParams,\n        });\n      } catch {\n        results.push({\n          idx: i,\n          nodeId: entry.nodeId,\n          opType: entry.node.opType,\n          tensor: outputName,\n          sum: NaN,\n          first4: [NaN, NaN, NaN, NaN],\n        });\n      }\n    }\n    return results;\n  }\n\n  debugWriteBuffer(tensorName: string, data: ArrayBufferView): void {\n    const buffer = this.getBuffer(tensorName);\n    if (!buffer) throw new Error(`No buffer for \"${tensorName}\"`);\n    this.ctx.device.queue.writeBuffer(buffer, 0, data as BufferSource);\n  }\n\n  /**\n   * Write a host-supplied activation input buffer (e.g. multimodal M-RoPE\n   * cos/sin, spliced vision embeddings, image row-map). The buffer must be a\n   * persistent activation tensor in the graph. Call before forward().\n   */\n  writeInput(tensorName: string, data: ArrayBufferView): void {\n    const buffer = this.getBuffer(tensorName);\n    if (!buffer) throw new Error(`No input buffer for \"${tensorName}\"`);\n    this.ctx.device.queue.writeBuffer(buffer, 0, data as BufferSource);\n  }\n\n  /** Write a host activation buffer at a byte offset (for per-row decode updates). */\n  writeInputAt(tensorName: string, data: ArrayBufferView, byteOffset: number): void {\n    const buffer = this.getBuffer(tensorName);\n    if (!buffer) throw new Error(`No input buffer for \"${tensorName}\"`);\n    this.ctx.device.queue.writeBuffer(buffer, byteOffset, data as BufferSource);\n  }\n\n  /** True if the graph has a buffer with this name (multimodal-capability probe). */\n  hasBuffer(tensorName: string): boolean {\n    return this.getBuffer(tensorName) !== undefined;\n  }\n\n  /** Current sequence position (number of tokens processed since reset). */\n  get currentSeqPos(): number {\n    return this.seqPos;\n  }\n\n  async debugReadBuffer(\n    tensorName: string,\n    maxElements?: number,\n    byteOffset?: number,\n  ): Promise<Float32Array> {\n    const buffer = this.getBuffer(tensorName);\n    if (!buffer) throw new Error(`No buffer for \"${tensorName}\"`);\n    const srcOffset = byteOffset ?? 0;\n    const remaining = buffer.size - srcOffset;\n    const byteLen = maxElements ? Math.min(maxElements * 4, remaining) : remaining;\n    const readback = this.ctx.device.createBuffer({\n      size: byteLen,\n      usage: 0x0001 | 0x0008,\n    });\n    const encoder = this.ctx.device.createCommandEncoder();\n    encoder.copyBufferToBuffer(buffer, srcOffset, readback, 0, byteLen);\n    this.ctx.device.queue.submit([encoder.finish()]);\n    await readback.mapAsync(MAP_MODE_READ, 0, byteLen);\n    const data = new Float32Array(readback.getMappedRange(0, byteLen).slice(0));\n    readback.unmap();\n    readback.destroy();\n    return data;\n  }\n\n  destroy(): void {\n    destroyBuffers([...this.weightBuffers.values()]);\n    // Pooled activation buffers are shared by multiple tensor names — dedupe.\n    destroyBuffers([...new Set(this.activationBuffers.values())]);\n    destroyBuffers([...this.ssmStateBuffers.values()]);\n    destroyBuffers([...this.kvCacheBuffers.values()]);\n    for (const entry of this.dispatchEntries) {\n      entry.uniformBuffer.destroy();\n    }\n    // Only destroy decode-specific uniform buffers (matmul nodes have their own)\n    for (let i = 0; i < this.decodeEntries.length; i++) {\n      if (this.decodeEntries[i] !== this.dispatchEntries[i]) {\n        this.decodeEntries[i].uniformBuffer.destroy();\n      }\n    }\n    if (this.argmaxEntry) {\n      this.argmaxEntry.uniformBuffer.destroy();\n    }\n    if (this.uniformStagingBuffer) {\n      this.uniformStagingBuffer.destroy();\n      this.uniformStagingBuffer = null;\n    }\n    this.inputIdsBuffer.destroy();\n    this.logitsReadback.destroy();\n    this.argmaxResultBuffer.destroy();\n    this.argmaxReadback.destroy();\n    destroyBuffers(this.decodeReadbacks);\n    this.decodeReadbacks = [];\n    this.weightBuffers.clear();\n    this.activationBuffers.clear();\n    this.ssmStateBuffers.clear();\n    this.kvCacheBuffers.clear();\n    this.dispatchEntries = [];\n    this.decodeEntries = [];\n  }\n\n  // ── Private ──────────────────────────────────────────────────────\n\n  /**\n   * Allocate activation buffers with liveness-based reuse.\n   *\n   * One dedicated buffer per activation tensor at full maxSeqLen is ~2.3GB for\n   * Qwen3.5-0.8B at T=512 — over the iOS jetsam budget on its own. Instead,\n   * a buffer returns to a size-keyed pool once its tensor's last reader has\n   * executed, so concurrently-live tensors share a small working set.\n   *\n   * Graph outputs and tensors read before they are written (cross-forward\n   * state) keep dedicated buffers. Within a forward, dispatches execute in\n   * executionOrder on every path (single-pass Dawn, per-dispatch WebKit), and\n   * WebGPU synchronizes hazards between dispatches, so reuse is safe.\n   *\n   * Caveat: debugReadBuffer() on an intermediate tensor is only meaningful\n   * before a later op reuses its buffer (probes that stop mid-graph are fine).\n   */\n  private allocateActivationBuffers(): void {\n    const isActivation = (name: string) => this.graph.tensors[name]?.storage === \"activation\";\n    // Debug: give every activation its own buffer so post-hoc debugReadBuffer of\n    // intermediate tensors is reliable (pooling otherwise recycles buffers).\n    const noPool = typeof process !== \"undefined\" && process.env?.GERBIL_NO_ACT_POOL === \"1\";\n    if (noPool) {\n      for (const name of Object.keys(this.graph.tensors)) {\n        if (!isActivation(name)) continue;\n        const desc = this.graph.tensors[name];\n        const shape = desc.shape.map((d) =>\n          d === \"T\" || d === \"L_max\" ? this.maxSeqLen : (d as number),\n        );\n        const bytes = shape.reduce((a, b) => a * b, 1) * DTYPE_BYTES[desc.dtype];\n        this.activationBuffers.set(\n          name,\n          createStorageBuffer(this.ctx, `act_nopool_${name}`, bytes),\n        );\n      }\n      console.log(\n        `[executor] activations: NO-POOL debug mode → ${this.activationBuffers.size} dedicated buffers`,\n      );\n      return;\n    }\n    const resolveBytes = (name: string): number => {\n      const desc = this.graph.tensors[name];\n      const shape = desc.shape.map((d) => {\n        if (d === \"T\") return this.maxSeqLen;\n        if (d === \"L_max\") return this.maxSeqLen;\n        return d as number;\n      });\n      return shape.reduce((a, b) => a * b, 1) * DTYPE_BYTES[desc.dtype];\n    };\n\n    const nodeById = new Map(this.graph.nodes.map((n) => [n.id, n]));\n    const order = this.graph.executionOrder;\n    const persistent = new Set<string>(this.graph.outputs);\n    const firstDef = new Map<string, number>();\n    const lastUse = new Map<string, number>();\n    for (let i = 0; i < order.length; i++) {\n      const node = nodeById.get(order[i])!;\n      for (const inp of node.inputs) {\n        if (!isActivation(inp)) continue;\n        // Read before any write → carries state across forwards (or in-place op);\n        // never pool it.\n        if (!firstDef.has(inp)) persistent.add(inp);\n        lastUse.set(inp, i);\n      }\n      for (const out of node.outputs) {\n        if (isActivation(out) && !firstDef.has(out)) firstDef.set(out, i);\n      }\n    }\n\n    const freePool = new Map<number, GPUBuffer[]>();\n    let created = 0;\n    const released = new Set<string>();\n    for (let i = 0; i < order.length; i++) {\n      const node = nodeById.get(order[i])!;\n      // Acquire outputs first so a node never writes the buffer of a tensor\n      // it also reads.\n      for (const out of node.outputs) {\n        if (!isActivation(out) || persistent.has(out) || this.activationBuffers.has(out)) {\n          continue;\n        }\n        const byteSize = resolveBytes(out);\n        const pool = freePool.get(byteSize);\n        let buffer = pool?.pop();\n        if (!buffer) {\n          buffer = createStorageBuffer(this.ctx, `act_pool_${created}_${byteSize}b`, byteSize);\n          created++;\n        }\n        this.activationBuffers.set(out, buffer);\n      }\n      // Release buffers whose tensor will never be read again.\n      for (const name of [...node.inputs, ...node.outputs]) {\n        if (!isActivation(name) || persistent.has(name) || released.has(name)) continue;\n        const last = lastUse.get(name) ?? firstDef.get(name) ?? i;\n        if (last > i) continue;\n        const buffer = this.activationBuffers.get(name);\n        if (!buffer) continue;\n        released.add(name);\n        const byteSize = resolveBytes(name);\n        const pool = freePool.get(byteSize) ?? [];\n        pool.push(buffer);\n        freePool.set(byteSize, pool);\n      }\n    }\n\n    // Dedicated buffers for graph outputs, cross-forward state, and any\n    // activation tensor not touched by the execution order.\n    for (const name of Object.keys(this.graph.tensors)) {\n      if (!isActivation(name) || this.activationBuffers.has(name)) continue;\n      this.activationBuffers.set(\n        name,\n        createStorageBuffer(this.ctx, `act_${name}`, resolveBytes(name)),\n      );\n    }\n\n    const unique = new Set(this.activationBuffers.values());\n    let totalBytes = 0;\n    for (const b of unique) totalBytes += b.size;\n    console.log(\n      `[executor] activations: ${this.activationBuffers.size} tensors → ${unique.size} buffers (${(totalBytes / 1e6).toFixed(0)} MB)`,\n    );\n  }\n\n  private allocateSSMStateBuffers(): void {\n    for (const [name, desc] of Object.entries(this.graph.tensors)) {\n      if (desc.storage !== \"ssm_state\") continue;\n      const shape = desc.shape as number[];\n      const byteSize = shape.reduce((a, b) => a * b, 1) * DTYPE_BYTES[desc.dtype];\n      const buffer = createStorageBuffer(this.ctx, `ssm_${name}`, byteSize);\n      this.ssmStateBuffers.set(name, buffer);\n    }\n  }\n\n  private allocateKVCacheBuffers(): void {\n    for (const [name, desc] of Object.entries(this.graph.tensors)) {\n      if (desc.storage !== \"kv_cache\") continue;\n      const shape = desc.shape.map((d) => {\n        if (d === \"L_max\") return this.maxSeqLen;\n        return d as number;\n      });\n      const byteSize = shape.reduce((a, b) => a * b, 1) * DTYPE_BYTES[desc.dtype];\n      const buffer = createStorageBuffer(this.ctx, `kv_${name}`, byteSize);\n      this.kvCacheBuffers.set(name, buffer);\n    }\n  }\n\n  private resolveShapes(T: number): Record<string, number[]> {\n    const resolved: Record<string, number[]> = {};\n    for (const [name, desc] of Object.entries(this.graph.tensors)) {\n      resolved[name] = desc.shape.map((d) => {\n        if (d === \"T\") return T;\n        if (d === \"L_max\") return this.seqPos + T;\n        return d as number;\n      });\n    }\n    return resolved;\n  }\n\n  private getBuffer(tensorName: string): GPUBuffer | undefined {\n    return (\n      this.weightBuffers.get(tensorName) ??\n      this.activationBuffers.get(tensorName) ??\n      this.ssmStateBuffers.get(tensorName) ??\n      this.kvCacheBuffers.get(tensorName)\n    );\n  }\n\n  /**\n   * Detect gate_proj + up_proj + SwiGLU patterns in decode entries and replace\n   * with a single fused SwiGLUMatVec dispatch. Saves 2 dispatches per MLP block.\n   */\n  private fuseSwiGLUDecodeEntries(): void {\n    let fusionCount = 0;\n    const maxBindings = this.ctx.limits.maxStorageBuffersPerShaderStage;\n\n    for (let i = 0; i < this.decodeEntries.length - 2; i++) {\n      const e1 = this.decodeEntries[i]; // gate_proj MatVec\n      const e2 = this.decodeEntries[i + 1]; // up_proj MatVec\n      const e3 = this.decodeEntries[i + 2]; // SwiGLU\n\n      // Check op types: two MatMul/MatMulInt4 followed by SwiGLU\n      const isE1Mat = e1.node.opType === \"MatMul\" || e1.node.opType === \"MatMulInt4\";\n      const isE2Mat = e2.node.opType === \"MatMul\" || e2.node.opType === \"MatMulInt4\";\n      if (!isE1Mat || !isE2Mat || e3.node.opType !== \"SwiGLU\") continue;\n\n      // Both projections must use the same op type (both F32 or both INT4)\n      if (e1.node.opType !== e2.node.opType) continue;\n\n      // INT4 fused kernel needs 9 storage bindings — skip if device doesn't support it\n      const isInt4Fusion = e1.node.opType === \"MatMulInt4\";\n      if (isInt4Fusion && maxBindings < 9) continue;\n\n      // Both must share the same input activation (first input tensor)\n      if (e1.node.inputs[0] !== e2.node.inputs[0]) continue;\n\n      // SwiGLU must consume the outputs of both projections\n      const gateOut = e1.node.outputs[0];\n      const upOut = e2.node.outputs[0];\n      if (e3.node.inputs[0] !== gateOut || e3.node.inputs[1] !== upOut) continue;\n\n      // Same K and N dimensions\n      if (e1.node.attributes.K !== e2.node.attributes.K) continue;\n      if (e1.node.attributes.N !== e2.node.attributes.N) continue;\n\n      // Pooled activation aliasing: the fused kernel reads the shared input\n      // and writes the SwiGLU output in ONE dispatch — they must not share\n      // a pooled buffer (separate dispatches are hazard-synchronized; a\n      // single dispatch is not).\n      if (this.getBuffer(e1.node.inputs[0]) === this.getBuffer(e3.node.outputs[0])) continue;\n\n      // Pattern matched — create fused dispatch entry\n      const fusedSpec = isInt4Fusion ? SWIGLU_MATVEC_INT4_SPEC : SWIGLU_MATVEC_SPEC;\n      const fusedPipeline = getOrCreatePipeline(\n        this.ctx,\n        `swiglu_mv_${e1.nodeId}`,\n        fusedSpec.shaderCode,\n        fusedSpec.entryPoint,\n      );\n\n      const fusedUniformData = fusedSpec.buildParams(e1.node, {}, { seqPos: 0 });\n      const fusedUniform = createUniformBuffer(\n        this.ctx,\n        `uniform_swiglu_mv_${e1.nodeId}`,\n        fusedUniformData,\n      );\n\n      // Build bind group manually: input, gate weights, up weights, output, uniform\n      const inputBuf = this.getBuffer(e1.node.inputs[0])!;\n      const outputBuf = this.getBuffer(e3.node.outputs[0])!;\n\n      let bufferEntries: Array<{ buffer: GPUBuffer }>;\n      if (isInt4Fusion) {\n        // INT4: A, gate_q, gate_s, gate_z, up_q, up_s, up_z, output, params\n        const gateQ = this.getBuffer(e1.node.inputs[1])!;\n        const gateS = this.getBuffer(e1.node.inputs[2])!;\n        const gateZ = this.getBuffer(e1.node.inputs[3])!;\n        const upQ = this.getBuffer(e2.node.inputs[1])!;\n        const upS = this.getBuffer(e2.node.inputs[2])!;\n        const upZ = this.getBuffer(e2.node.inputs[3])!;\n        bufferEntries = [\n          { buffer: inputBuf },\n          { buffer: gateQ },\n          { buffer: gateS },\n          { buffer: gateZ },\n          { buffer: upQ },\n          { buffer: upS },\n          { buffer: upZ },\n          { buffer: outputBuf },\n          { buffer: fusedUniform },\n        ];\n      } else {\n        // F32: A, B_gate, B_up, output, params\n        const gateBuf = this.getBuffer(e1.node.inputs[1])!;\n        const upBuf = this.getBuffer(e2.node.inputs[1])!;\n        bufferEntries = [\n          { buffer: inputBuf },\n          { buffer: gateBuf },\n          { buffer: upBuf },\n          { buffer: outputBuf },\n          { buffer: fusedUniform },\n        ];\n      }\n\n      const fusedBindGroup = createBindGroup(\n        this.ctx,\n        fusedPipeline,\n        bufferEntries,\n        `bg_swiglu_mv_${e1.nodeId}`,\n      );\n\n      const fusedEntry: DispatchEntry = {\n        nodeId: `fused_swiglu_${e1.nodeId}`,\n        node: e1.node, // Use gate node for K/N attributes\n        spec: fusedSpec,\n        pipeline: fusedPipeline,\n        bindGroup: fusedBindGroup,\n        uniformBuffer: fusedUniform,\n        lastParamsBytes: null,\n        lastDispatchSize: null,\n      };\n\n      // Replace 3 entries with 1 fused entry\n      this.decodeEntries.splice(i, 3, fusedEntry);\n      fusionCount++;\n      // Don't increment i — check if next position is also fusable\n    }\n\n    if (fusionCount > 0) {\n      console.log(\n        `[executor] Fused ${fusionCount} SwiGLU MLP blocks (saved ${fusionCount * 2} dispatches)`,\n      );\n    }\n  }\n\n  /**\n   * Fuse two adjacent INT4 projections that share the same input activation and\n   * the same K/N (e.g. q_proj+gate_proj and k_proj+v_proj in full-attention\n   * decode) into a single DualMatVecInt4 dispatch. Reads the shared input vector\n   * once and writes both projection outputs — removing one GPU round-trip (one\n   * submit+drain on Safari/iOS) per fused pair.\n   *\n   * Numerics are identical to running the two MatVecInt4 kernels separately\n   * (same dequant, same K-parallel reduction), so this is WebKit-safe: it reuses\n   * the proven INT4 matvec math and only merges two writes into one dispatch.\n   *\n   * Must run AFTER fuseSwiGLUDecodeEntries so the MLP gate+up pair (which is\n   * consumed by a SwiGLU node) is already collapsed and won't be matched here.\n   */\n  private fuseDualMatVecDecodeEntries(): void {\n    let fusionCount = 0;\n    // DualMatVecInt4 needs 10 storage/uniform bindings (9 storage + 1 uniform).\n    const maxBindings = this.ctx.limits.maxStorageBuffersPerShaderStage;\n    if (maxBindings < 9) return;\n\n    for (let i = 0; i < this.decodeEntries.length - 1; i++) {\n      const e0 = this.decodeEntries[i];\n      const e1 = this.decodeEntries[i + 1];\n\n      // Both must be INT4 matmuls.\n      if (e0.node.opType !== \"MatMulInt4\" || e1.node.opType !== \"MatMulInt4\") continue;\n      // Same input activation (first input tensor).\n      if (e0.node.inputs[0] !== e1.node.inputs[0]) continue;\n      // Same K, N, group_size.\n      if (e0.node.attributes.K !== e1.node.attributes.K) continue;\n      if (e0.node.attributes.N !== e1.node.attributes.N) continue;\n      if (e0.node.attributes.group_size !== e1.node.attributes.group_size) continue;\n\n      const out0 = e0.node.outputs[0];\n      const out1 = e1.node.outputs[0];\n      const inputBuf = this.getBuffer(e0.node.inputs[0]);\n      const out0Buf = this.getBuffer(out0);\n      const out1Buf = this.getBuffer(out1);\n      if (!inputBuf || !out0Buf || !out1Buf) continue;\n\n      // Pooled-buffer aliasing guard: the single fused dispatch reads the shared\n      // input and writes both outputs. None of these may share a pooled buffer\n      // (separate dispatches are hazard-synchronized; one dispatch is not), and\n      // the two outputs must be distinct buffers.\n      if (inputBuf === out0Buf || inputBuf === out1Buf || out0Buf === out1Buf) continue;\n\n      const fusedSpec = DUAL_MATVEC_INT4_SPEC;\n      const fusedPipeline = getOrCreatePipeline(\n        this.ctx,\n        `dual_mv_${e0.nodeId}`,\n        fusedSpec.shaderCode,\n        fusedSpec.entryPoint,\n      );\n\n      const fusedUniformData = fusedSpec.buildParams(e0.node, {}, { seqPos: 0 });\n      const fusedUniform = createUniformBuffer(\n        this.ctx,\n        `uniform_dual_mv_${e0.nodeId}`,\n        fusedUniformData,\n      );\n\n      // Bind group: A, B0_q, B0_s, B0_z, B1_q, B1_s, B1_z, out0, out1, params\n      const b0q = this.getBuffer(e0.node.inputs[1])!;\n      const b0s = this.getBuffer(e0.node.inputs[2])!;\n      const b0z = this.getBuffer(e0.node.inputs[3])!;\n      const b1q = this.getBuffer(e1.node.inputs[1])!;\n      const b1s = this.getBuffer(e1.node.inputs[2])!;\n      const b1z = this.getBuffer(e1.node.inputs[3])!;\n\n      const bufferEntries = [\n        { buffer: inputBuf },\n        { buffer: b0q },\n        { buffer: b0s },\n        { buffer: b0z },\n        { buffer: b1q },\n        { buffer: b1s },\n        { buffer: b1z },\n        { buffer: out0Buf },\n        { buffer: out1Buf },\n        { buffer: fusedUniform },\n      ];\n\n      const fusedBindGroup = createBindGroup(\n        this.ctx,\n        fusedPipeline,\n        bufferEntries,\n        `bg_dual_mv_${e0.nodeId}`,\n      );\n\n      const fusedEntry: DispatchEntry = {\n        nodeId: `fused_dual_${e0.nodeId}`,\n        node: e0.node, // K/N/group_size attributes shared\n        spec: fusedSpec,\n        pipeline: fusedPipeline,\n        bindGroup: fusedBindGroup,\n        uniformBuffer: fusedUniform,\n        lastParamsBytes: null,\n        lastDispatchSize: null,\n      };\n\n      this.decodeEntries.splice(i, 2, fusedEntry);\n      fusionCount++;\n      // Don't increment i — the next pair may also be fusable.\n    }\n\n    if (fusionCount > 0) {\n      console.log(\n        `[executor] Fused ${fusionCount} dual INT4 projections (saved ${fusionCount} dispatches)`,\n      );\n    }\n  }\n\n  /**\n   * Fuse the adjacent K-cache and V-cache appends in each full-attention layer\n   * into a single DualKVCacheAppend dispatch. Both are pure memcpys into f32\n   * caches sharing the same width and dst_offset, so one dispatch with two\n   * src/dst buffers writes both — removing one GPU round-trip per layer.\n   *\n   * Supports f32, native-f16, and packed-f16 caches (the dual kernel mirrors the\n   * single-append kernel selected for the active kvMode). Numerically identical\n   * to the separate appends — WebKit-safe (the packed-f16 variant is the Safari\n   * path and uses pack2x16float, no `enable f16`).\n   */\n  private fuseDualKVCacheAppendEntries(): void {\n    let fusionCount = 0;\n\n    for (let i = 0; i < this.decodeEntries.length - 1; i++) {\n      const e0 = this.decodeEntries[i]; // K append\n      const e1 = this.decodeEntries[i + 1]; // V append\n\n      if (e0.node.opType !== \"KVCacheAppend\" || e1.node.opType !== \"KVCacheAppend\") continue;\n      // Same per-element width (kv_dim) → one param set drives both.\n      if (e0.node.attributes.width !== e1.node.attributes.width) continue;\n\n      // Select the dual kernel matching the cache dtype/mode. Both destinations\n      // must share the same dtype (they always do: K and V caches are symmetric).\n      const dstK = e0.node.outputs[0];\n      const dstV = e1.node.outputs[0];\n      const dtK = this.graph.tensors[dstK]?.dtype;\n      const dtV = this.graph.tensors[dstV]?.dtype;\n      if (dtK !== dtV) continue;\n      let fusedSpec: KernelSpec;\n      if (dtK === \"f32\") {\n        fusedSpec = DUAL_KV_CACHE_APPEND_SPEC;\n      } else if (dtK === \"f16\") {\n        fusedSpec =\n          this.kvMode === \"packed-f16\"\n            ? DUAL_KV_CACHE_APPEND_PACKED_F16_SPEC\n            : DUAL_KV_CACHE_APPEND_F16_SPEC;\n      } else {\n        continue;\n      }\n\n      const srcK = this.getBuffer(e0.node.inputs[0]);\n      const srcV = this.getBuffer(e1.node.inputs[0]);\n      const dstKBuf = this.getBuffer(dstK);\n      const dstVBuf = this.getBuffer(dstV);\n      if (!srcK || !srcV || !dstKBuf || !dstVBuf) continue;\n      // Distinct caches; sources distinct from caches.\n      if (dstKBuf === dstVBuf) continue;\n\n      const fusedPipeline = getOrCreatePipeline(\n        this.ctx,\n        `dual_kv_${e0.nodeId}`,\n        fusedSpec.shaderCode,\n        fusedSpec.entryPoint,\n      );\n\n      const dummyShapes = this.resolveShapes(1);\n      const fusedUniformData = fusedSpec.buildParams(e0.node, dummyShapes, { seqPos: 0 });\n      const fusedUniform = createUniformBuffer(\n        this.ctx,\n        `uniform_dual_kv_${e0.nodeId}`,\n        fusedUniformData,\n      );\n\n      const bufferEntries = [\n        { buffer: srcK },\n        { buffer: srcV },\n        { buffer: dstKBuf },\n        { buffer: dstVBuf },\n        { buffer: fusedUniform },\n      ];\n\n      const fusedBindGroup = createBindGroup(\n        this.ctx,\n        fusedPipeline,\n        bufferEntries,\n        `bg_dual_kv_${e0.nodeId}`,\n      );\n\n      const fusedEntry: DispatchEntry = {\n        nodeId: `fused_kv_${e0.nodeId}`,\n        node: e0.node, // width + T_tensor for per-token param rebuild\n        spec: fusedSpec,\n        pipeline: fusedPipeline,\n        bindGroup: fusedBindGroup,\n        uniformBuffer: fusedUniform,\n        lastParamsBytes: null,\n        lastDispatchSize: null,\n      };\n\n      this.decodeEntries.splice(i, 2, fusedEntry);\n      fusionCount++;\n    }\n\n    if (fusionCount > 0) {\n      console.log(\n        `[executor] Fused ${fusionCount} dual KV-cache appends (saved ${fusionCount} dispatches)`,\n      );\n    }\n  }\n\n  /**\n   * Fuse the attention SigmoidGate (attn_out * sigmoid(gate)) into the INT4\n   * o_proj that consumes it: a GatedMatVecInt4 reads attn_out and gate directly,\n   * applies the sigmoid gate to its input vector, and runs the projection in ONE\n   * dispatch — removing the standalone SigmoidGate (one round-trip per\n   * full-attention layer).\n   *\n   * Numerically identical to SigmoidGate→MatVecInt4 (same gate formula, same INT4\n   * dequant + reduction). Slight extra ALU: the gated input is recomputed per\n   * output column, but A reads hit L1 and the saved submit+drain dominates on\n   * mobile. WebKit risk: low — reuses the proven INT4 matvec, only the A vector\n   * is built from two reads + a sigmoid (no new reduction/barrier pattern).\n   *\n   * Runs after the dual fusions so it only sees the post-attention SigmoidGate.\n   */\n  private fuseGatedOProjDecodeEntries(): void {\n    let fusionCount = 0;\n\n    for (let i = 0; i < this.decodeEntries.length - 1; i++) {\n      const eg = this.decodeEntries[i]; // SigmoidGate\n      const eo = this.decodeEntries[i + 1]; // o_proj MatMulInt4\n\n      if (eg.node.opType !== \"SigmoidGate\") continue;\n      if (eo.node.opType !== \"MatMulInt4\") continue;\n\n      // o_proj must consume the gate's output as its activation input.\n      const gatedOut = eg.node.outputs[0];\n      if (eo.node.inputs[0] !== gatedOut) continue;\n\n      // SigmoidGate inputs: [x (attn_out), gate]\n      const attnName = eg.node.inputs[0];\n      const gateName = eg.node.inputs[1];\n      const attnBuf = this.getBuffer(attnName);\n      const gateBuf = this.getBuffer(gateName);\n      const wQ = this.getBuffer(eo.node.inputs[1]);\n      const wS = this.getBuffer(eo.node.inputs[2]);\n      const wZ = this.getBuffer(eo.node.inputs[3]);\n      const outBuf = this.getBuffer(eo.node.outputs[0]);\n      if (!attnBuf || !gateBuf || !wQ || !wS || !wZ || !outBuf) continue;\n\n      // Pooled aliasing: the single fused dispatch reads attn + gate and writes\n      // the projection output — none may share a pooled buffer.\n      if (outBuf === attnBuf || outBuf === gateBuf || attnBuf === gateBuf) continue;\n\n      const fusedSpec = GATED_MATVEC_INT4_SPEC;\n      const fusedPipeline = getOrCreatePipeline(\n        this.ctx,\n        `gated_o_${eo.nodeId}`,\n        fusedSpec.shaderCode,\n        fusedSpec.entryPoint,\n      );\n\n      const fusedUniformData = fusedSpec.buildParams(eo.node, {}, { seqPos: 0 });\n      const fusedUniform = createUniformBuffer(\n        this.ctx,\n        `uniform_gated_o_${eo.nodeId}`,\n        fusedUniformData,\n      );\n\n      const bufferEntries = [\n        { buffer: attnBuf },\n        { buffer: gateBuf },\n        { buffer: wQ },\n        { buffer: wS },\n        { buffer: wZ },\n        { buffer: outBuf },\n        { buffer: fusedUniform },\n      ];\n\n      const fusedBindGroup = createBindGroup(\n        this.ctx,\n        fusedPipeline,\n        bufferEntries,\n        `bg_gated_o_${eo.nodeId}`,\n      );\n\n      const fusedEntry: DispatchEntry = {\n        nodeId: `fused_gated_o_${eo.nodeId}`,\n        node: eo.node, // K/N/group_size attributes for per-token param rebuild\n        spec: fusedSpec,\n        pipeline: fusedPipeline,\n        bindGroup: fusedBindGroup,\n        uniformBuffer: fusedUniform,\n        lastParamsBytes: null,\n        lastDispatchSize: null,\n      };\n\n      this.decodeEntries.splice(i, 2, fusedEntry);\n      fusionCount++;\n    }\n\n    if (fusionCount > 0) {\n      console.log(\n        `[executor] Fused ${fusionCount} gated o_proj blocks (saved ${fusionCount} dispatches)`,\n      );\n    }\n  }\n\n  /**\n   * Fuse a standalone SwiGLU (silu(gate) * up) into the INT4 projection that\n   * consumes its output: a SwiGLUGatedMatVecInt4 reads gate and up directly,\n   * builds the gated input vector, and runs the projection in ONE dispatch.\n   * Targets the Mamba block's mamba_swiglu (silu(z) * norm_out) feeding out_proj\n   * — one round-trip saved per linear-attention layer.\n   *\n   * The MLP SwiGLU is already collapsed into a SwiGLUMatVec entry by\n   * fuseSwiGLUDecodeEntries (it has no surviving standalone SwiGLU node), so only\n   * the Mamba SwiGLU matches here. Numerically identical to SwiGLU→MatVecInt4;\n   * WebKit risk low (reuses the proven INT4 matvec, only the A vector changes).\n   */\n  private fuseSwiGLUGatedProjDecodeEntries(): void {\n    let fusionCount = 0;\n\n    for (let i = 0; i < this.decodeEntries.length - 1; i++) {\n      const es = this.decodeEntries[i]; // SwiGLU\n      const ep = this.decodeEntries[i + 1]; // projection MatMulInt4\n\n      if (es.node.opType !== \"SwiGLU\") continue;\n      if (ep.node.opType !== \"MatMulInt4\") continue;\n\n      const swigluOut = es.node.outputs[0];\n      if (ep.node.inputs[0] !== swigluOut) continue;\n\n      // SwiGLU inputs: [gate, up]\n      const gateName = es.node.inputs[0];\n      const upName = es.node.inputs[1];\n      const gateBuf = this.getBuffer(gateName);\n      const upBuf = this.getBuffer(upName);\n      const wQ = this.getBuffer(ep.node.inputs[1]);\n      const wS = this.getBuffer(ep.node.inputs[2]);\n      const wZ = this.getBuffer(ep.node.inputs[3]);\n      const outBuf = this.getBuffer(ep.node.outputs[0]);\n      if (!gateBuf || !upBuf || !wQ || !wS || !wZ || !outBuf) continue;\n\n      // Pooled aliasing: the single fused dispatch reads gate + up and writes the\n      // projection output — none may share a pooled buffer.\n      if (outBuf === gateBuf || outBuf === upBuf || gateBuf === upBuf) continue;\n\n      const fusedSpec = SWIGLU_GATED_MATVEC_INT4_SPEC;\n      const fusedPipeline = getOrCreatePipeline(\n        this.ctx,\n        `swiglu_gated_${ep.nodeId}`,\n        fusedSpec.shaderCode,\n        fusedSpec.entryPoint,\n      );\n\n      const fusedUniformData = fusedSpec.buildParams(ep.node, {}, { seqPos: 0 });\n      const fusedUniform = createUniformBuffer(\n        this.ctx,\n        `uniform_swiglu_gated_${ep.nodeId}`,\n        fusedUniformData,\n      );\n\n      const bufferEntries = [\n        { buffer: gateBuf },\n        { buffer: upBuf },\n        { buffer: wQ },\n        { buffer: wS },\n        { buffer: wZ },\n        { buffer: outBuf },\n        { buffer: fusedUniform },\n      ];\n\n      const fusedBindGroup = createBindGroup(\n        this.ctx,\n        fusedPipeline,\n        bufferEntries,\n        `bg_swiglu_gated_${ep.nodeId}`,\n      );\n\n      const fusedEntry: DispatchEntry = {\n        nodeId: `fused_swiglu_gated_${ep.nodeId}`,\n        node: ep.node,\n        spec: fusedSpec,\n        pipeline: fusedPipeline,\n        bindGroup: fusedBindGroup,\n        uniformBuffer: fusedUniform,\n        lastParamsBytes: null,\n        lastDispatchSize: null,\n      };\n\n      this.decodeEntries.splice(i, 2, fusedEntry);\n      fusionCount++;\n    }\n\n    if (fusionCount > 0) {\n      console.log(\n        `[executor] Fused ${fusionCount} SwiGLU-gated projections (saved ${fusionCount} dispatches)`,\n      );\n    }\n  }\n\n  /**\n   * Fuse two adjacent per-row RMSNorms sharing hidden_size + eps into a single\n   * DualRMSNorm dispatch (e.g. the per-head q_norm and k_norm in full-attention\n   * decode). One workgroup still handles one row; the fused grid just spans both\n   * inputs' rows, so each row's reduction is unchanged — numerically identical to\n   * two separate RMSNorm dispatches. One round-trip saved per fused pair.\n   *\n   * WebKit risk: low — same single-workgroup reduction as the proven RMSNorm\n   * kernel, only the row→input routing is added.\n   */\n  private fuseDualRMSNormDecodeEntries(): void {\n    let fusionCount = 0;\n\n    for (let i = 0; i < this.decodeEntries.length - 1; i++) {\n      const e0 = this.decodeEntries[i];\n      const e1 = this.decodeEntries[i + 1];\n\n      if (e0.node.opType !== \"RMSNorm\" || e1.node.opType !== \"RMSNorm\") continue;\n      // Must share per-row width and eps (the dual kernel uses one of each).\n      if (e0.node.attributes.hidden_size !== e1.node.attributes.hidden_size) continue;\n      if (e0.node.attributes.eps !== e1.node.attributes.eps) continue;\n\n      const in0 = e0.node.inputs[0];\n      const w0 = e0.node.inputs[1];\n      const out0 = e0.node.outputs[0];\n      const in1 = e1.node.inputs[0];\n      const w1 = e1.node.inputs[1];\n      const out1 = e1.node.outputs[0];\n\n      const in0Buf = this.getBuffer(in0);\n      const w0Buf = this.getBuffer(w0);\n      const out0Buf = this.getBuffer(out0);\n      const in1Buf = this.getBuffer(in1);\n      const w1Buf = this.getBuffer(w1);\n      const out1Buf = this.getBuffer(out1);\n      if (!in0Buf || !w0Buf || !out0Buf || !in1Buf || !w1Buf || !out1Buf) continue;\n\n      // The two normalizations must write distinct buffers. (In-place RMSNorm\n      // where input===output is fine: each row reads then writes its own row,\n      // and the fused kernel preserves per-row ordering.) Cross-aliasing — where\n      // one norm's input shares a pooled buffer with the OTHER norm's output —\n      // would create an unsynchronized read/write hazard within the single\n      // dispatch, so reject it.\n      if (out0Buf === out1Buf) continue;\n      if (in0Buf === out1Buf || in1Buf === out0Buf) continue;\n\n      // Build a synthetic node carrying both row sources so the spec can resolve\n      // per-token row counts each decode step.\n      const fusedNode: OpNode = {\n        id: `fused_dualnorm_${e0.nodeId}`,\n        opType: \"RMSNorm\",\n        inputs: [in0, w0, in1, w1],\n        outputs: [out0, out1],\n        attributes: {\n          hidden_size: e0.node.attributes.hidden_size,\n          eps: e0.node.attributes.eps,\n          seq_len_tensor0: e0.node.attributes.seq_len_tensor,\n          seq_len_tensor1: e1.node.attributes.seq_len_tensor,\n          seq_len0: e0.node.attributes.seq_len,\n          seq_len1: e1.node.attributes.seq_len,\n        },\n      };\n\n      const fusedSpec = DUAL_RMSNORM_SPEC;\n      const fusedPipeline = getOrCreatePipeline(\n        this.ctx,\n        `dual_norm_${e0.nodeId}`,\n        fusedSpec.shaderCode,\n        fusedSpec.entryPoint,\n      );\n\n      const dummyShapes = this.resolveShapes(1);\n      const fusedUniformData = fusedSpec.buildParams(fusedNode, dummyShapes, { seqPos: 0 });\n      const fusedUniform = createUniformBuffer(\n        this.ctx,\n        `uniform_dual_norm_${e0.nodeId}`,\n        fusedUniformData,\n      );\n\n      const bufferEntries = [\n        { buffer: in0Buf },\n        { buffer: w0Buf },\n        { buffer: in1Buf },\n        { buffer: w1Buf },\n        { buffer: out0Buf },\n        { buffer: out1Buf },\n        { buffer: fusedUniform },\n      ];\n\n      const fusedBindGroup = createBindGroup(\n        this.ctx,\n        fusedPipeline,\n        bufferEntries,\n        `bg_dual_norm_${e0.nodeId}`,\n      );\n\n      const fusedEntry: DispatchEntry = {\n        nodeId: fusedNode.id,\n        node: fusedNode,\n        spec: fusedSpec,\n        pipeline: fusedPipeline,\n        bindGroup: fusedBindGroup,\n        uniformBuffer: fusedUniform,\n        lastParamsBytes: null,\n        lastDispatchSize: null,\n      };\n\n      this.decodeEntries.splice(i, 2, fusedEntry);\n      fusionCount++;\n    }\n\n    if (fusionCount > 0) {\n      console.log(\n        `[executor] Fused ${fusionCount} dual RMSNorms (saved ${fusionCount} dispatches)`,\n      );\n    }\n  }\n\n  /**\n   * Gather buffer entries for a bind group, matching the kernel spec's binding layout.\n   * Uses the pre-allocated inputIdsBuffer for the \"input_ids\" tensor.\n   */\n  private gatherBuffers(\n    spec: KernelSpec,\n    node: OpNode,\n    uniformBuffer: GPUBuffer,\n  ): Array<{ buffer: GPUBuffer }> {\n    const entries: Array<{ buffer: GPUBuffer }> = [];\n    const allTensorNames = [...node.inputs, ...node.outputs];\n\n    // Track buffers already bound to a writable slot in this group so we can\n    // detect (and break) read_write aliasing, which WebGPU rejects.\n    const writableBound = new Set<GPUBuffer>();\n\n    for (let i = 0; i < spec.bindings.length; i++) {\n      const binding = spec.bindings[i];\n\n      if (binding.type === \"uniform\") {\n        entries.push({ buffer: uniformBuffer });\n      } else {\n        const tensorIdx = i - spec.bindings.filter((b, j) => j < i && b.type === \"uniform\").length;\n        const tensorName = allTensorNames[tensorIdx];\n        if (!tensorName) throw new Error(`Missing tensor for binding ${i} in op ${node.id}`);\n\n        let buffer: GPUBuffer | undefined;\n        if (tensorName === \"input_ids\") {\n          buffer = this.inputIdsBuffer;\n        } else {\n          buffer = this.getBuffer(tensorName);\n        }\n        if (!buffer) throw new Error(`No GPU buffer for tensor \"${tensorName}\" in op ${node.id}`);\n\n        // Break read_write aliasing: if a writable binding would reuse a buffer\n        // already bound writable in this group (e.g. RoPE-Q-only nodes that list\n        // the same tensor as input+output and leave the k slot unused), bind a\n        // throwaway scratch buffer instead. Safe because the aliasing slot is not\n        // actually accessed at dispatch time (num_kv_heads=0 → no k work).\n        if (binding.type === \"storage-read-write\") {\n          if (writableBound.has(buffer)) {\n            buffer = this.getBindingScratchBuffer(buffer.size);\n          } else {\n            writableBound.add(buffer);\n          }\n        }\n\n        entries.push({ buffer });\n      }\n    }\n\n    return entries;\n  }\n\n  /** Lazily allocate a scratch storage buffer at least `minBytes` large. */\n  private getBindingScratchBuffer(minBytes: number): GPUBuffer {\n    if (!this.bindingScratchBuffer || this.bindingScratchBuffer.size < minBytes) {\n      this.bindingScratchBuffer = createStorageBuffer(\n        this.ctx,\n        \"binding_scratch\",\n        Math.max(minBytes, 256),\n      );\n    }\n    return this.bindingScratchBuffer;\n  }\n}\n","/**\n * GPTQ → Gerbil INT4 format adapter.\n *\n * Repacks GPTQ-quantized weights (qweight/qzeros/scales) into the flat\n * row-major packing expected by Gerbil's MatMulInt4 WGSL kernel.\n *\n * GPTQ layout (AutoGPTQ, per linear layer):\n *   qweight: Int32Array, shape [K/8, N] — 8 nibbles per int32 along INPUT dim (K)\n *   scales:  Float32Array, shape [K/group_size, N] — per-group per-output-column\n *   qzeros:  Int32Array, shape [K/group_size, N/8] — packed 4-bit zero points along N\n *\n * Gerbil layout (per linear layer):\n *   packed:  Uint32Array, flat [N*K/8] — 8 nibbles per u32 in row-major (N,K) order\n *   scales:  Float32Array, flat [N * K/group_size] — sequential in (N,K) row-major\n *   zeros:   Float32Array, flat [N * K/group_size]\n *\n * The kernel indexes: flat_idx = col * K + k, group = flat_idx / group_size\n */\n\nexport interface GPTQTensors {\n  qweight: Int32Array; // [K/8, N]\n  scales: Float32Array; // [K/group_size, N]\n  qzeros: Int32Array; // [K/group_size, N/8]\n  K: number;\n  N: number;\n  groupSize: number;\n}\n\nexport interface GerbilINT4Tensors {\n  packed: Uint32Array;\n  scales: Float32Array;\n  zeros: Float32Array;\n}\n\n/**\n * Repack GPTQ tensors into Gerbil's flat INT4 format.\n *\n * This is pure data reshuffling — no quantization math. The 4-bit values,\n * scales, and zero points are preserved exactly.\n */\nexport function repackGPTQ(gptq: GPTQTensors): GerbilINT4Tensors {\n  const { qweight, scales: gptqScales, qzeros, K, N, groupSize } = gptq;\n  const totalElements = N * K;\n  const groupsPerCol = Math.ceil(K / groupSize);\n  const totalGroups = N * groupsPerCol;\n\n  const packed = new Uint32Array(Math.ceil(totalElements / 8));\n  const scales = new Float32Array(totalGroups);\n  const zeros = new Float32Array(totalGroups);\n\n  // Repack nibbles: GPTQ [K/8, N] → Gerbil flat [N*K/8]\n  // GPTQ: qweight[(k >> 3) * N + col], nibble at (k & 7) * 4\n  // Gerbil: packed[(col*K+k)/8], nibble at ((col*K+k) % 8) * 4\n  for (let k = 0; k < K; k++) {\n    const gptqRow = k >>> 3;\n    const gptqNibblePos = k & 7;\n    const gptqRowBase = gptqRow * N;\n    for (let col = 0; col < N; col++) {\n      // Extract nibble from GPTQ\n      const gptqWord = qweight[gptqRowBase + col];\n      const nibble = (gptqWord >>> (gptqNibblePos * 4)) & 0xf;\n\n      // Write nibble into Gerbil flat layout\n      const flatIdx = col * K + k;\n      const packedIdx = flatIdx >>> 3;\n      const nibblePos = flatIdx & 7;\n      packed[packedIdx] |= nibble << (nibblePos * 4);\n    }\n  }\n\n  // Repack scales: GPTQ [groups_per_col, N] → Gerbil flat [N * groups_per_col]\n  // GPTQ: scales[g * N + col]\n  // Gerbil: scales[col * groups_per_col + g]\n  for (let g = 0; g < groupsPerCol; g++) {\n    const srcRow = g * N;\n    for (let col = 0; col < N; col++) {\n      scales[col * groupsPerCol + g] = gptqScales[srcRow + col];\n    }\n  }\n\n  // Repack zeros: GPTQ [groups_per_col, N/8] packed int4 → Gerbil flat float32\n  // GPTQ: qzeros[g * (N/8) + (col >> 3)], nibble at (col & 7) * 4\n  // Gerbil: zeros[col * groups_per_col + g] as float32\n  //\n  // GPTQ convention: stored qzeros = actual_zero_point - 1.\n  // All standard GPTQ kernels (exllama, marlin) add +1 during dequant.\n  // We bake the +1 into the stored value so the kernel formula stays simple:\n  //   dequant = (nibble - zero) * scale\n  const zerosNDiv8 = N >>> 3;\n  for (let g = 0; g < groupsPerCol; g++) {\n    const srcRow = g * zerosNDiv8;\n    for (let col = 0; col < N; col++) {\n      const zWord = qzeros[srcRow + (col >>> 3)];\n      const zNibblePos = col & 7;\n      const zeroVal = (zWord >>> (zNibblePos * 4)) & 0xf;\n      zeros[col * groupsPerCol + g] = zeroVal + 1; // +1: GPTQ packing convention\n    }\n  }\n\n  return { packed, scales, zeros };\n}\n","/**\n * IndexedDB-backed blob cache for model weights.\n *\n * Why IndexedDB and not the Cache API / OPFS we hand-rolled before:\n *   - The Cache API ignores URL fragments in `match()`, which silently collapsed\n *     our per-tensor `${url}#t:start-len` keys onto one entry → wrong bytes → the\n *     `RRRR` garbage and `wt.data` crashes. IDB keys on the exact string.\n *   - OPFS via Worker (`createSyncAccessHandle`) is finicky on iOS Safari (OOM /\n *     transient failures) and added a lot of moving parts.\n * IndexedDB is the boring, proven large-binary store every browser supports. It\n * keys on arbitrary strings (no fragment footgun), stores ArrayBuffers directly,\n * and is durable (subject to the same per-origin quota as Cache/OPFS — install to\n * Home Screen lifts that ceiling; IDB doesn't change it, it just makes caching\n * actually CORRECT).\n *\n * Inert on Node (no indexedDB) — every export resolves to a miss/no-op so the\n * heap load path is unchanged off the browser.\n */\n\nconst DB_NAME = \"gerbil-weights-v1\";\nconst STORE = \"tensors\";\n\nfunction idbAvailable(): boolean {\n  return typeof indexedDB !== \"undefined\";\n}\n\nlet _dbPromise: Promise<IDBDatabase | null> | null = null;\n\nfunction openDB(): Promise<IDBDatabase | null> {\n  if (_dbPromise) return _dbPromise;\n  _dbPromise = new Promise<IDBDatabase | null>((resolve) => {\n    if (!idbAvailable()) {\n      resolve(null);\n      return;\n    }\n    try {\n      const req = indexedDB.open(DB_NAME, 1);\n      req.onupgradeneeded = () => {\n        const db = req.result;\n        if (!db.objectStoreNames.contains(STORE)) db.createObjectStore(STORE);\n      };\n      req.onsuccess = () => resolve(req.result);\n      req.onerror = () => resolve(null);\n      req.onblocked = () => resolve(null);\n    } catch {\n      resolve(null);\n    }\n  });\n  return _dbPromise;\n}\n\n/** A single transaction operation, promisified. Resolves the fallback on any error. */\nfunction tx<T>(\n  mode: IDBTransactionMode,\n  run: (store: IDBObjectStore) => IDBRequest,\n  fallback: T,\n): Promise<T> {\n  return openDB().then(\n    (db) =>\n      new Promise<T>((resolve) => {\n        if (!db) {\n          resolve(fallback);\n          return;\n        }\n        try {\n          const t = db.transaction(STORE, mode);\n          const req = run(t.objectStore(STORE));\n          req.onsuccess = () => resolve((req.result as T) ?? fallback);\n          req.onerror = () => resolve(fallback);\n          t.onabort = () => resolve(fallback);\n        } catch {\n          resolve(fallback);\n        }\n      }),\n  );\n}\n\n/** Read a key's bytes, or null on miss / no IDB / any error. */\nexport async function idbGet(key: string): Promise<ArrayBuffer | null> {\n  const v = await tx<ArrayBuffer | null>(\"readonly\", (s) => s.get(key), null);\n  return v instanceof ArrayBuffer ? v : null;\n}\n\n/** Cheap presence check (reads only the key via getKey — no body materialized). */\nexport async function idbHas(key: string): Promise<boolean> {\n  const k = await tx<IDBValidKey | undefined>(\"readonly\", (s) => s.getKey(key), undefined);\n  return k !== undefined;\n}\n\n/**\n * All cached keys, as a Set, in ONE transaction. Probing presence per-tensor\n * (hundreds of parallel getKey transactions) storms Safari's IndexedDB and hangs\n * the load; getAllKeys() does it in a single transaction so callers can check\n * membership in memory. Empty set on miss / no IDB / error.\n */\nexport async function idbAllKeys(): Promise<Set<string>> {\n  const keys = await tx<IDBValidKey[]>(\"readonly\", (s) => s.getAllKeys(), []);\n  return new Set((keys ?? []).map((k) => String(k)));\n}\n\n/** Write a key's bytes. Returns true on success, false on quota/error (caller\n * falls back to network — never throws). */\nexport async function idbPut(key: string, buf: ArrayBuffer): Promise<boolean> {\n  // put returns the key on success; map to boolean.\n  const r = await tx<IDBValidKey | null>(\"readwrite\", (s) => s.put(buf, key), null);\n  return r !== null;\n}\n\n/** Delete a key (best-effort). */\nexport async function idbDelete(key: string): Promise<void> {\n  await tx<unknown>(\"readwrite\", (s) => s.delete(key), null);\n}\n\n/** True when IndexedDB is usable in this environment (browser, not Node). */\nexport function idbCacheAvailable(): boolean {\n  return idbAvailable();\n}\n","/**\n * MLX 4-bit → Gerbil INT4 format adapter.\n *\n * Converts MLX quantized weights (weight/scales/biases) into the flat\n * row-major packing expected by Gerbil's MatMulInt4 / EmbeddingInt4 kernels.\n *\n * MLX layout (per quantized layer):\n *   weight: Uint32Array, shape [N, K/8] — 8 nibbles per u32, packed along K (last axis)\n *   scales: Float32Array, shape [N, K/group_size] — per-group scale factors\n *   biases: Float32Array, shape [N, K/group_size] — per-group bias values\n *\n * MLX dequant formula (affine mode):\n *   w = scale * nibble + bias\n *\n * Gerbil dequant formula:\n *   w = (nibble - zero) * scale\n *\n * Conversion: scale_gerbil = scale_mlx, zero_gerbil = -bias_mlx / scale_mlx\n *\n * MLX [N, K/8] is already row-major with nibbles packed along K — same layout\n * as Gerbil flat [N*K/8]. The nibble data is directly usable; only scales/zeros\n * need conversion.\n */\n\nimport type { GerbilINT4Tensors } from \"./gptq-adapter.js\";\n\nexport interface MLXTensors {\n  weight: Uint32Array; // [N, K/8] packed 4-bit\n  scales: Float32Array; // [N, K/group_size]\n  biases: Float32Array; // [N, K/group_size]\n  N: number;\n  K: number;\n  groupSize: number;\n}\n\n/**\n * Convert MLX 4-bit tensors into Gerbil's flat INT4 format.\n *\n * The packed nibbles are already in the correct layout — MLX [N, K/8] row-major\n * matches Gerbil flat [N*K/8]. We just convert the affine scale/bias to\n * Gerbil's (nibble - zero) * scale convention.\n */\nexport function repackMLX(mlx: MLXTensors): GerbilINT4Tensors {\n  const { weight, scales: mlxScales, biases: mlxBiases, N, K, groupSize } = mlx;\n  const groupsPerRow = Math.ceil(K / groupSize);\n  const totalGroups = N * groupsPerRow;\n\n  // Nibbles: MLX [N, K/8] is already flat [N*K/8] in row-major order\n  // The packing is identical: 8 nibbles per u32, little-nibble-first\n  const packed = weight;\n\n  // Convert affine scale/bias to Gerbil scale/zero\n  // MLX: w = scale * nibble + bias\n  // Gerbil: w = (nibble - zero) * scale\n  // Therefore: zero = -bias / scale, scale stays the same\n  const scales = new Float32Array(totalGroups);\n  const zeros = new Float32Array(totalGroups);\n\n  // MLX stores [N, groups_per_row] which matches Gerbil flat [N * groups_per_row]\n  // Both are row-major with groups sequential per output row — same layout.\n  for (let i = 0; i < totalGroups; i++) {\n    const s = mlxScales[i];\n    const b = mlxBiases[i];\n    scales[i] = s;\n    // When scale is zero, bias should also be zero (constant group)\n    zeros[i] = s !== 0 ? -b / s : 0;\n  }\n\n  return { packed, scales, zeros };\n}\n","/**\n * OPFS-backed blob cache for durable model-weight persistence.\n *\n * WHY THIS EXISTS\n * ---------------\n * On iOS Safari the CacheStorage API (`caches.open(...)`) used for model weights\n * is evicted under quota pressure in a plain tab — `navigator.storage.persist()`\n * is not granted outside an installed PWA — so the ~404 MB model is re-downloaded\n * on every visit. The Origin Private File System (OPFS) is the durable path, but:\n *\n *  • Main-thread OPFS `createWritable` throws \"out of quota\" mid-write on iOS\n *    Safari and can leave orphaned, unclearable bytes.\n *  • The ONLY iOS-safe OPFS write path is `FileSystemSyncAccessHandle`\n *    (`createSyncAccessHandle`), which is available ONLY inside a Worker.\n *\n * So this module runs all OPFS I/O on a dedicated Worker created from an inlined\n * blob URL (no build wiring, zero extra deps). The main thread talks to it over\n * postMessage; payload ArrayBuffers are TRANSFERRED (zero-copy) so a write does\n * not double peak memory. Writes are chunked inside the worker so the sync handle\n * never has to hold the whole tensor at once.\n *\n * KEY → FILENAME\n * --------------\n * Cache keys are URLs with `/`, `#`, `?` etc. (e.g.\n * `https://huggingface.co/...model.safetensors#t:8388608-2359296`). OPFS file\n * names cannot contain those, so each key is mapped to a stable, collision-\n * resistant filename via a 53-bit FNV-style hash plus the key length. Files live\n * in a versioned OPFS subdirectory (see OPFS_DIR_NAME).\n *\n * DESKTOP / NODE\n * --------------\n * Everything here is gated by `opfsAvailable()`, which is false in Node (no\n * `navigator.storage.getDirectory`, no `Worker`, no `Blob`). Callers must short-\n * circuit on that before touching any export here. The Node weight path never\n * reaches this module.\n */\n\n/** Versioned OPFS directory. Bump when the on-disk layout/semantics change. */\nconst OPFS_DIR_NAME = \"gerbil-models-v4\";\n\n/** Per-message write chunk inside the worker (bounded peak memory). */\nconst WRITE_CHUNK_BYTES = 8 * 1024 * 1024;\n\n/**\n * True when the durable OPFS+Worker path is usable. False in Node (no navigator/\n * Worker/Blob) and in browsers without OPFS or Workers. Cheap and synchronous so\n * the seam functions can bail before any async work.\n */\nexport function opfsAvailable(): boolean {\n  return (\n    typeof navigator !== \"undefined\" &&\n    typeof (navigator as { storage?: { getDirectory?: unknown } }).storage?.getDirectory ===\n      \"function\" &&\n    typeof Worker !== \"undefined\" &&\n    typeof Blob !== \"undefined\" &&\n    typeof URL !== \"undefined\" &&\n    typeof URL.createObjectURL === \"function\"\n  );\n}\n\n/**\n * Map an arbitrary cache key to a safe OPFS filename. FNV-1a over UTF-16 code\n * units folded into a 53-bit safe integer, combined with the key length so two\n * different keys are astronomically unlikely to collide. Deterministic across\n * reloads (this is what makes a tensor written last visit re-resolve this visit).\n */\nfunction keyToFileName(key: string): string {\n  // FNV-1a (32-bit) mixed twice for a wider hash; emit two hex halves + length.\n  let h1 = 0x81_1c_9d_c5;\n  let h2 = 0x01_00_01_93;\n  for (let i = 0; i < key.length; i++) {\n    const c = key.charCodeAt(i);\n    h1 ^= c;\n    h1 = Math.imul(h1, 0x01_00_01_93);\n    h2 ^= c + i;\n    h2 = Math.imul(h2, 0x85_eb_ca_6b);\n  }\n  const a = (h1 >>> 0).toString(16).padStart(8, \"0\");\n  const b = (h2 >>> 0).toString(16).padStart(8, \"0\");\n  return `${a}${b}-${key.length.toString(16)}.bin`;\n}\n\n// ── Inline worker source ─────────────────────────────────────────────────────\n//\n// Self-contained worker that owns one OPFS directory handle and performs\n// has/read/write/delete via createSyncAccessHandle. Kept as a string so the\n// existing browser tsdown entry emits it inline (no separate worker bundle).\n// The worker has no imports — everything it needs is inlined here.\nconst WORKER_SOURCE = `\nconst DIR_NAME = ${JSON.stringify(OPFS_DIR_NAME)};\nlet dirPromise = null;\n\nfunction getDir() {\n  if (!dirPromise) {\n    dirPromise = navigator.storage\n      .getDirectory()\n      .then((root) => root.getDirectoryHandle(DIR_NAME, { create: true }));\n  }\n  return dirPromise;\n}\n\nasync function handleHas(name) {\n  const dir = await getDir();\n  try {\n    const fh = await dir.getFileHandle(name, { create: false });\n    const access = await fh.createSyncAccessHandle();\n    const size = access.getSize();\n    access.close();\n    // A zero-byte file is a half-written/aborted entry — treat as absent.\n    return size > 0;\n  } catch {\n    return false;\n  }\n}\n\nasync function handleRead(name) {\n  const dir = await getDir();\n  const fh = await dir.getFileHandle(name, { create: false });\n  const access = await fh.createSyncAccessHandle();\n  try {\n    const size = access.getSize();\n    if (size === 0) {\n      access.close();\n      return null;\n    }\n    const buf = new ArrayBuffer(size);\n    access.read(new Uint8Array(buf), { at: 0 });\n    access.close();\n    return buf;\n  } catch (e) {\n    try { access.close(); } catch {}\n    throw e;\n  }\n}\n\nasync function handleWrite(name, buf) {\n  const dir = await getDir();\n  const fh = await dir.getFileHandle(name, { create: true });\n  const access = await fh.createSyncAccessHandle();\n  try {\n    access.truncate(0);\n    const total = buf.byteLength;\n    const view = new Uint8Array(buf);\n    const CHUNK = ${WRITE_CHUNK_BYTES};\n    let off = 0;\n    while (off < total) {\n      const end = Math.min(off + CHUNK, total);\n      // Subarray is a zero-copy view; write returns bytes written.\n      access.write(view.subarray(off, end), { at: off });\n      off = end;\n    }\n    access.flush();\n    const finalSize = access.getSize();\n    access.close();\n    if (finalSize !== total) {\n      // Write came up short (quota) — remove the partial so it never reads back\n      // as a valid (wrong-size) entry.\n      try { await dir.removeEntry(name); } catch {}\n      throw new Error(\"OPFS short write: \" + finalSize + \"/\" + total);\n    }\n  } catch (e) {\n    try { access.close(); } catch {}\n    // Best-effort cleanup of any partial file.\n    try { await dir.removeEntry(name); } catch {}\n    throw e;\n  }\n}\n\nasync function handleDelete(name) {\n  const dir = await getDir();\n  try { await dir.removeEntry(name); } catch {}\n}\n\nself.onmessage = async (ev) => {\n  const { id, op, name, buf } = ev.data;\n  try {\n    if (op === \"has\") {\n      const present = await handleHas(name);\n      self.postMessage({ id, ok: true, present });\n    } else if (op === \"read\") {\n      const out = await handleRead(name);\n      if (out) self.postMessage({ id, ok: true, buf: out }, [out]);\n      else self.postMessage({ id, ok: true, buf: null });\n    } else if (op === \"write\") {\n      await handleWrite(name, buf);\n      self.postMessage({ id, ok: true });\n    } else if (op === \"delete\") {\n      await handleDelete(name);\n      self.postMessage({ id, ok: true });\n    } else {\n      self.postMessage({ id, ok: false, error: \"unknown op: \" + op });\n    }\n  } catch (e) {\n    self.postMessage({ id, ok: false, error: String((e && e.message) || e) });\n  }\n};\n`;\n\ninterface PendingResolver {\n  resolve: (value: { present?: boolean; buf?: ArrayBuffer | null }) => void;\n  reject: (err: Error) => void;\n}\n\n/**\n * Lazily-created singleton worker client. Created on first use; if creation or\n * any handshake fails the whole OPFS path is marked unavailable so callers fall\n * back to CacheStorage cleanly.\n */\nclass OpfsClient {\n  private worker: Worker | null = null;\n  private blobUrl: string | null = null;\n  private pending = new Map<number, PendingResolver>();\n  private nextId = 1;\n  private broken = false;\n\n  private ensureWorker(): Worker | null {\n    if (this.broken) return null;\n    if (this.worker) return this.worker;\n    try {\n      const blob = new Blob([WORKER_SOURCE], { type: \"application/javascript\" });\n      this.blobUrl = URL.createObjectURL(blob);\n      const worker = new Worker(this.blobUrl);\n      worker.onmessage = (ev: MessageEvent) => {\n        const { id, ok, present, buf, error } = ev.data as {\n          id: number;\n          ok: boolean;\n          present?: boolean;\n          buf?: ArrayBuffer | null;\n          error?: string;\n        };\n        const p = this.pending.get(id);\n        if (!p) return;\n        this.pending.delete(id);\n        if (ok) p.resolve({ present, buf });\n        else p.reject(new Error(error || \"OPFS worker error\"));\n      };\n      worker.onerror = () => {\n        // Fatal worker error — reject everything in flight and disable OPFS.\n        this.broken = true;\n        for (const p of this.pending.values()) {\n          p.reject(new Error(\"OPFS worker crashed\"));\n        }\n        this.pending.clear();\n      };\n      this.worker = worker;\n      return worker;\n    } catch {\n      this.broken = true;\n      return null;\n    }\n  }\n\n  private call(\n    op: string,\n    name: string,\n    buf?: ArrayBuffer,\n  ): Promise<{ present?: boolean; buf?: ArrayBuffer | null }> {\n    const worker = this.ensureWorker();\n    if (!worker) return Promise.reject(new Error(\"OPFS unavailable\"));\n    const id = this.nextId++;\n    return new Promise((resolve, reject) => {\n      this.pending.set(id, { resolve, reject });\n      try {\n        if (buf) {\n          // Transfer the buffer (zero-copy). The caller must not reuse it after.\n          worker.postMessage({ id, op, name, buf }, [buf]);\n        } else {\n          worker.postMessage({ id, op, name });\n        }\n      } catch (e) {\n        this.pending.delete(id);\n        reject(e instanceof Error ? e : new Error(String(e)));\n      }\n    });\n  }\n\n  async has(key: string): Promise<boolean> {\n    const { present } = await this.call(\"has\", keyToFileName(key));\n    return Boolean(present);\n  }\n\n  async read(key: string): Promise<ArrayBuffer | null> {\n    const { buf } = await this.call(\"read\", keyToFileName(key));\n    return buf ?? null;\n  }\n\n  async write(key: string, buf: ArrayBuffer): Promise<void> {\n    await this.call(\"write\", keyToFileName(key), buf);\n  }\n\n  async delete(key: string): Promise<void> {\n    await this.call(\"delete\", keyToFileName(key));\n  }\n}\n\nlet _client: OpfsClient | null = null;\nfunction client(): OpfsClient {\n  if (!_client) _client = new OpfsClient();\n  return _client;\n}\n\n/**\n * Presence check for a key in the OPFS store. Returns false (never throws) so the\n * caller can fall through to CacheStorage on any error.\n */\nexport async function opfsHas(key: string): Promise<boolean> {\n  if (!opfsAvailable()) return false;\n  try {\n    return await client().has(key);\n  } catch {\n    return false;\n  }\n}\n\n/**\n * Read a key's bytes from OPFS. Returns null on miss or any error (caller falls\n * back to CacheStorage / network).\n */\nexport async function opfsRead(key: string): Promise<ArrayBuffer | null> {\n  if (!opfsAvailable()) return null;\n  try {\n    return await client().read(key);\n  } catch {\n    return null;\n  }\n}\n\n/**\n * Write a key's bytes to OPFS durably. The buffer is TRANSFERRED to the worker\n * (do not reuse `buf` after calling — pass a copy if you still need it). Returns\n * true on success, false on any failure (quota, worker crash, …) so the caller\n * can fall back to CacheStorage without crashing the load.\n */\nexport async function opfsWrite(key: string, buf: ArrayBuffer): Promise<boolean> {\n  if (!opfsAvailable()) return false;\n  try {\n    await client().write(key, buf);\n    return true;\n  } catch {\n    return false;\n  }\n}\n\n/** Delete a key from OPFS (best-effort, never throws). */\nexport async function opfsDelete(key: string): Promise<void> {\n  if (!opfsAvailable()) return;\n  try {\n    await client().delete(key);\n  } catch {\n    /* best-effort */\n  }\n}\n\n/**\n * Delete superseded OPFS model directories (old namespaces, e.g. a previous\n * `gerbil-models-v3`) to reclaim quota. On iOS the per-origin quota is ~1 GB and\n * orphaned namespaces can fill it, blocking the current model from caching at all\n * (every write then fails with quota/OOM → redownload every load). Directory\n * removal works on the MAIN thread — only createSyncAccessHandle needs a Worker —\n * so this is cheap and best-effort. Inert on Node (no navigator.storage).\n */\nexport async function opfsEvictStale(): Promise<void> {\n  try {\n    const root = (await (\n      navigator as { storage?: { getDirectory?: () => Promise<unknown> } }\n    )?.storage?.getDirectory?.()) as\n      | {\n          entries?: () => AsyncIterable<[string, { kind: string }]>;\n          removeEntry: (n: string, o?: { recursive?: boolean }) => Promise<void>;\n        }\n      | undefined;\n    if (!root) return;\n    const stale: string[] = [];\n    for await (const [name, handle] of root.entries?.() ?? []) {\n      if (\n        handle.kind === \"directory\" &&\n        name.startsWith(\"gerbil-models-\") &&\n        name !== OPFS_DIR_NAME\n      ) {\n        stale.push(name);\n      }\n    }\n    for (const name of stale) {\n      await root.removeEntry(name, { recursive: true }).catch(() => {\n        /* best-effort */\n      });\n    }\n  } catch {\n    /* best-effort */\n  }\n}\n\n/**\n * One-time health probe: actually write a few bytes to OPFS, read them back,\n * verify they round-trip, then clean up. Memoized. iOS Safari can report OPFS as\n * \"available\" (getDirectory/Worker/Blob all present) yet have the Worker's\n * createSyncAccessHandle writes silently fail — in which case we must NOT trust\n * OPFS, because the per-tensor write path transfers (detaches) the buffer and a\n * failed OPFS write would then have no CacheStorage fallback (→ nothing cached →\n * redownload every load). Callers gate OPFS use on THIS, not the cheap synchronous\n * opfsAvailable(). Resolves false if OPFS can't actually persist a round-trip.\n */\nlet _healthy: Promise<boolean> | null = null;\nexport function opfsHealthy(): Promise<boolean> {\n  if (_healthy) return _healthy;\n  _healthy = (async () => {\n    if (!opfsAvailable()) return false;\n    try {\n      const key = \"__opfs_health_probe__\";\n      const ok = await opfsWrite(key, new Uint8Array([7, 13, 42, 99]).buffer);\n      if (!ok) return false;\n      const back = await opfsRead(key);\n      await opfsDelete(key);\n      if (!back || back.byteLength !== 4) return false;\n      const b = new Uint8Array(back);\n      return b[0] === 7 && b[1] === 13 && b[2] === 42 && b[3] === 99;\n    } catch {\n      return false;\n    }\n  })();\n  return _healthy;\n}\n","/**\n * Safetensors binary format parser.\n *\n * Parses the HuggingFace safetensors format into typed array views\n * over the raw buffer — zero-copy where possible.\n */\n\nexport interface SafetensorEntry {\n  /** Tensor name as it appears in the file (e.g. \"model.layers.0.self_attn.q_proj.weight\"). */\n  name: string;\n  /** Data type string from the header. */\n  dtype: SafetensorDType;\n  /** Shape dimensions. */\n  shape: number[];\n  /** Byte offset of this tensor's data relative to the start of the data section. */\n  dataOffset: number;\n  /** Byte length of this tensor's data. */\n  dataLength: number;\n}\n\nexport type SafetensorDType =\n  | \"F16\"\n  | \"F32\"\n  | \"BF16\"\n  | \"I32\"\n  | \"I8\"\n  | \"U8\"\n  | \"F64\"\n  | \"BOOL\"\n  | \"I16\"\n  | \"U16\"\n  | \"U32\"\n  | \"I64\"\n  | \"U64\";\n\nexport interface SafetensorsFile {\n  /** Header length in bytes. */\n  headerLength: number;\n  /** Byte offset where tensor data begins (8 + headerLength). */\n  dataStart: number;\n  /** All tensor entries (excludes __metadata__). */\n  entries: SafetensorEntry[];\n  /** Optional metadata from __metadata__ key. */\n  metadata: Record<string, string> | null;\n}\n\n/**\n * Parse safetensors header from an ArrayBuffer.\n *\n * Can be called with just the header bytes (for streaming — parse header first,\n * then fetch tensor data by offset) or with the entire file.\n */\nexport function parseSafetensorsHeader(buffer: ArrayBuffer): SafetensorsFile {\n  if (buffer.byteLength < 8) {\n    throw new Error(\n      `Buffer too small to contain safetensors header length (need 8 bytes, got ${buffer.byteLength}).`,\n    );\n  }\n\n  const view = new DataView(buffer);\n\n  // First 8 bytes: header length as little-endian uint64.\n  // JS doesn't have native uint64, but headers are always < 4 GB.\n  const headerLength = Number(view.getBigUint64(0, true));\n\n  if (headerLength > buffer.byteLength - 8) {\n    throw new Error(\n      `Safetensors header length (${headerLength}) exceeds buffer size (${buffer.byteLength - 8}). ` +\n        `Pass at least the first ${headerLength + 8} bytes.`,\n    );\n  }\n\n  // Decode the JSON header\n  const headerBytes = new Uint8Array(buffer, 8, headerLength);\n  const headerStr = new TextDecoder().decode(headerBytes);\n  const header: Record<string, unknown> = JSON.parse(headerStr);\n\n  const dataStart = 8 + headerLength;\n  const entries: SafetensorEntry[] = [];\n  let metadata: Record<string, string> | null = null;\n\n  for (const [name, info] of Object.entries(header)) {\n    if (name === \"__metadata__\") {\n      metadata = info as Record<string, string>;\n      continue;\n    }\n\n    const { dtype, shape, data_offsets } = info as {\n      dtype: SafetensorDType;\n      shape: number[];\n      data_offsets: [number, number];\n    };\n\n    entries.push({\n      name,\n      dtype,\n      shape,\n      dataOffset: data_offsets[0],\n      dataLength: data_offsets[1] - data_offsets[0],\n    });\n  }\n\n  // Sort by offset for sequential access patterns\n  entries.sort((a, b) => a.dataOffset - b.dataOffset);\n\n  return { headerLength, dataStart, entries, metadata };\n}\n\n/** Byte alignment required for each dtype. */\nfunction dtypeAlignment(dtype: SafetensorDType): number {\n  switch (dtype) {\n    case \"F64\":\n    case \"I64\":\n    case \"U64\":\n      return 8;\n    case \"F32\":\n    case \"I32\":\n    case \"U32\":\n      return 4;\n    case \"F16\":\n    case \"BF16\":\n    case \"I16\":\n    case \"U16\":\n      return 2;\n    case \"I8\":\n    case \"U8\":\n    case \"BOOL\":\n      return 1;\n    default:\n      return 1;\n  }\n}\n\n/**\n * Create a typed array for the given dtype.\n *\n * If `offset` is aligned to the element size, returns a zero-copy view\n * into `buffer`. Otherwise, copies the relevant slice into a new\n * properly-aligned ArrayBuffer.\n */\nfunction makeTypedView(\n  buffer: ArrayBuffer,\n  offset: number,\n  byteLength: number,\n  dtype: SafetensorDType,\n): ArrayBufferView {\n  const align = dtypeAlignment(dtype);\n  const aligned = offset % align === 0;\n\n  // For single-byte types, alignment is always satisfied.\n  const src = aligned ? buffer : buffer.slice(offset, offset + byteLength);\n  const base = aligned ? offset : 0;\n\n  switch (dtype) {\n    case \"F32\":\n      return new Float32Array(src, base, byteLength / 4);\n    case \"F16\":\n      // F16 has no native TypedArray — return Uint16Array (bitwise).\n      return new Uint16Array(src, base, byteLength / 2);\n    case \"BF16\":\n      // BF16 also has no native TypedArray — return Uint16Array (bitwise).\n      return new Uint16Array(src, base, byteLength / 2);\n    case \"I32\":\n      return new Int32Array(src, base, byteLength / 4);\n    case \"U32\":\n      return new Uint32Array(src, base, byteLength / 4);\n    case \"I8\":\n      return new Int8Array(src, base, byteLength);\n    case \"U8\":\n      return new Uint8Array(src, base, byteLength);\n    case \"I16\":\n      return new Int16Array(src, base, byteLength / 2);\n    case \"U16\":\n      return new Uint16Array(src, base, byteLength / 2);\n    case \"F64\":\n      return new Float64Array(src, base, byteLength / 8);\n    case \"I64\":\n      return new BigInt64Array(src, base, byteLength / 8);\n    case \"U64\":\n      return new BigUint64Array(src, base, byteLength / 8);\n    case \"BOOL\":\n      return new Uint8Array(src, base, byteLength);\n    default:\n      throw new Error(`Unsupported safetensors dtype: ${dtype}`);\n  }\n}\n\n/**\n * Get a typed array view for a tensor entry from a complete safetensors buffer.\n * Zero-copy when the byte offset is properly aligned; copies otherwise.\n */\nexport function getTensorData(\n  buffer: ArrayBuffer,\n  file: SafetensorsFile,\n  entry: SafetensorEntry,\n): ArrayBufferView {\n  const offset = file.dataStart + entry.dataOffset;\n  return makeTypedView(buffer, offset, entry.dataLength, entry.dtype);\n}\n\n/**\n * Find a tensor entry by name (exact match).\n */\nexport function findTensor(file: SafetensorsFile, name: string): SafetensorEntry | undefined {\n  return file.entries.find((e) => e.name === name);\n}\n\n/**\n * Parse just the header from a Response (streaming-friendly).\n *\n * Reads the full response into memory and parses the header.\n * Returns the parsed header and the full buffer for tensor access.\n */\nexport async function parseSafetensorsFromResponse(\n  response: Response,\n): Promise<{ file: SafetensorsFile; fullBuffer: ArrayBuffer }> {\n  const fullBuffer = await response.arrayBuffer();\n  const file = parseSafetensorsHeader(fullBuffer);\n  return { file, fullBuffer };\n}\n\n/**\n * Compute the total byte size of all tensor data in a safetensors file.\n */\nexport function totalTensorBytes(file: SafetensorsFile): number {\n  let total = 0;\n  for (const entry of file.entries) {\n    total += entry.dataLength;\n  }\n  return total;\n}\n\n/**\n * Convert BF16 (bfloat16) data to F32.\n * BF16 is the upper 16 bits of an IEEE 754 float32, so conversion is a simple left-shift.\n */\nexport function bf16ToF32(bf16Bytes: Uint8Array): Float32Array {\n  const bf16 = new Uint16Array(bf16Bytes.buffer, bf16Bytes.byteOffset, bf16Bytes.byteLength / 2);\n  const f32 = new Float32Array(bf16.length);\n  const u32 = new Uint32Array(f32.buffer);\n  for (let i = 0; i < bf16.length; i++) {\n    u32[i] = bf16[i] << 16;\n  }\n  return f32;\n}\n\n/**\n * Convert F16 (IEEE 754 half-precision) data to F32.\n */\nexport function f16ToF32(f16View: ArrayBufferView): Float32Array {\n  const u16 =\n    f16View instanceof Uint16Array\n      ? f16View\n      : new Uint16Array(f16View.buffer, f16View.byteOffset, f16View.byteLength / 2);\n  const f32 = new Float32Array(u16.length);\n  for (let i = 0; i < u16.length; i++) {\n    const h = u16[i];\n    const sign = (h >> 15) & 1;\n    const exp = (h >> 10) & 0x1f;\n    const frac = h & 0x3ff;\n    if (exp === 0) {\n      // Subnormal or zero\n      f32[i] = (sign ? -1 : 1) * 2 ** -14 * (frac / 1024);\n    } else if (exp === 31) {\n      // Inf or NaN\n      f32[i] = frac === 0 ? (sign ? -Infinity : Infinity) : NaN;\n    } else {\n      f32[i] = (sign ? -1 : 1) * 2 ** (exp - 15) * (1 + frac / 1024);\n    }\n  }\n  return f32;\n}\n","/**\n * Pure JavaScript BPE tokenizer.\n *\n * Reads HuggingFace tokenizer.json — no WASM, no external dependencies.\n * Supports encoding, decoding, and chat template application.\n */\n\n// Byte-level BPE unicode mapping (same as HuggingFace's bytes_to_unicode).\n// Maps each byte (0-255) to a unique unicode codepoint such that printable\n// ASCII characters map to themselves, and non-printable bytes are shifted\n// into the U+0100+ range.\nconst BYTE_TO_UNICODE = new Map<number, number>();\nconst UNICODE_TO_BYTE = new Map<number, number>();\n\n{\n  // Printable bytes that map to themselves: 33-126 (! through ~), 161-172, 174-255\n  const bs: number[] = [];\n  for (let i = 33; i <= 126; i++) bs.push(i); // ! through ~\n  for (let i = 161; i <= 172; i++) bs.push(i); // ¡ through ¬\n  for (let i = 174; i <= 255; i++) bs.push(i); // ® through ÿ\n\n  const cs = [...bs]; // unicode codepoints (same for these ranges)\n\n  // Non-printable bytes get shifted to U+0100+\n  let n = 0;\n  for (let b = 0; b < 256; b++) {\n    if (!bs.includes(b)) {\n      bs.push(b);\n      cs.push(256 + n);\n      n++;\n    }\n  }\n\n  for (let i = 0; i < bs.length; i++) {\n    BYTE_TO_UNICODE.set(bs[i], cs[i]);\n    UNICODE_TO_BYTE.set(cs[i], bs[i]);\n  }\n}\n\nexport interface TokenizerConfig {\n  bosToken: string | null;\n  eosToken: string | null;\n  bosTokenId: number | null;\n  eosTokenId: number | null;\n  chatTemplate: string | null;\n  addBosToken: boolean;\n  addEosToken: boolean;\n}\n\nexport interface ChatMessage {\n  role: \"system\" | \"user\" | \"assistant\";\n  content: string;\n}\n\nexport class Tokenizer {\n  private vocab: Map<string, number>; // token string → id\n  private vocabReverse: Map<number, string>; // id → token string\n  private merges: Map<string, number>; // \"tokenA tokenB\" → priority (lower = merge first)\n  private specialTokens: Map<string, number>; // special token string → id\n  private addedTokens: Map<string, number>; // ALL added tokens (special + non-special like <think>)\n  private byteFallback: Map<number, string>; // byte value → byte-level token\n  /**\n   * SentencePiece mode (Gemma/Llama-style). When true, vocab uses U+2581 (▁) for\n   * spaces and raw UTF-8 tokens (NOT the GPT-2 byte-to-unicode \"Ġ\" mapping), and\n   * raw bytes fall back to <0xHH> tokens. When false, GPT-2 byte-level BPE.\n   */\n  private spmMode: boolean;\n  readonly config: TokenizerConfig;\n  readonly vocabSize: number;\n\n  private constructor(\n    vocab: Map<string, number>,\n    vocabReverse: Map<number, string>,\n    merges: Map<string, number>,\n    specialTokens: Map<string, number>,\n    addedTokens: Map<string, number>,\n    config: TokenizerConfig,\n    vocabSize: number,\n    spmMode: boolean,\n  ) {\n    this.vocab = vocab;\n    this.vocabReverse = vocabReverse;\n    this.merges = merges;\n    this.specialTokens = specialTokens;\n    this.addedTokens = addedTokens;\n    this.byteFallback = new Map();\n    this.spmMode = spmMode;\n    this.config = config;\n    this.vocabSize = vocabSize;\n\n    // Build byte fallback map (for handling unknown characters)\n    // HF tokenizers use <0xHH> tokens for raw bytes\n    for (let b = 0; b < 256; b++) {\n      const hex = b.toString(16).toUpperCase().padStart(2, \"0\");\n      const key = `<0x${hex}>`;\n      if (vocab.has(key)) {\n        this.byteFallback.set(b, key);\n      }\n    }\n  }\n\n  /**\n   * Create a tokenizer from HuggingFace JSON files.\n   */\n  static fromJSON(tokenizerJSON: any, tokenizerConfigJSON?: any): Tokenizer {\n    const model = tokenizerJSON.model;\n    if (!model || model.type !== \"BPE\") {\n      throw new Error(\n        `Unsupported tokenizer type: ${model?.type ?? \"unknown\"}. Only BPE is supported.`,\n      );\n    }\n\n    // Build vocab\n    const vocab = new Map<string, number>();\n    const vocabReverse = new Map<number, string>();\n    for (const [token, id] of Object.entries(model.vocab as Record<string, number>)) {\n      vocab.set(token, id);\n      vocabReverse.set(id, token);\n    }\n\n    // Detect SentencePiece mode (Gemma/Llama): the normalizer replaces spaces\n    // with U+2581 (▁) and/or the vocab is dominated by ▁-prefixed tokens. In SPM\n    // mode tokens are raw UTF-8 with ▁ for spaces and bytes fall back to <0xHH>,\n    // rather than the GPT-2 byte-to-unicode (Ġ) mapping.\n    const normalizerReplacesSpace = (() => {\n      const norm = tokenizerJSON.normalizer;\n      const checkOne = (n: any) =>\n        n?.type === \"Replace\" && n?.pattern?.String === \" \" && n?.content === \"▁\";\n      if (!norm) return false;\n      if (checkOne(norm)) return true;\n      if (Array.isArray(norm.normalizers)) return norm.normalizers.some(checkOne);\n      return false;\n    })();\n    const byteFallbackFlag = model.byte_fallback === true;\n    const spmMode = normalizerReplacesSpace || (byteFallbackFlag && \"▁the\" in model.vocab);\n\n    // Build merge priority map. Merges are either \"a b\" strings (GPT-2 style) or\n    // [\"a\", \"b\"] arrays (newer/Gemma tokenizer.json). Normalize both to \"a b\".\n    const merges = new Map<string, number>();\n    const mergeList = (model.merges ?? []) as Array<string | [string, string]>;\n    for (let i = 0; i < mergeList.length; i++) {\n      const m = mergeList[i];\n      const key = Array.isArray(m) ? `${m[0]} ${m[1]}` : m;\n      merges.set(key, i);\n    }\n\n    // Register added tokens — ALL of them go into addedTokensMap for splitting,\n    // only special:true ones go into specialTokens for skip-during-decode\n    const specialTokens = new Map<string, number>();\n    const addedTokensMap = new Map<string, number>();\n    const addedTokensList = tokenizerJSON.added_tokens as\n      | Array<{ id: number; content: string; special?: boolean }>\n      | undefined;\n    if (addedTokensList) {\n      for (const t of addedTokensList) {\n        vocab.set(t.content, t.id);\n        vocabReverse.set(t.id, t.content);\n        addedTokensMap.set(t.content, t.id);\n        if (t.special) {\n          specialTokens.set(t.content, t.id);\n        }\n      }\n    }\n\n    // Parse config\n    const bosToken = tokenizerConfigJSON?.bos_token ?? null;\n    const eosToken = tokenizerConfigJSON?.eos_token ?? null;\n    const chatTemplate = tokenizerConfigJSON?.chat_template ?? null;\n    const addBosToken = tokenizerConfigJSON?.add_bos_token ?? false;\n    const addEosToken = tokenizerConfigJSON?.add_eos_token ?? false;\n\n    const config: TokenizerConfig = {\n      bosToken,\n      eosToken,\n      bosTokenId: bosToken ? (vocab.get(bosToken) ?? null) : null,\n      eosTokenId: eosToken ? (vocab.get(eosToken) ?? null) : null,\n      chatTemplate,\n      addBosToken,\n      addEosToken,\n    };\n\n    return new Tokenizer(\n      vocab,\n      vocabReverse,\n      merges,\n      specialTokens,\n      addedTokensMap,\n      config,\n      vocab.size,\n      spmMode,\n    );\n  }\n\n  /**\n   * Resolve a literal token string (e.g. \"<|endoftext|>\") to its vocab id,\n   * or null if it isn't in the vocabulary.\n   */\n  tokenToId(token: string): number | null {\n    return this.vocab.get(token) ?? null;\n  }\n\n  /**\n   * Encode text into token IDs.\n   */\n  encode(text: string): number[] {\n    if (!text) return [];\n\n    const ids: number[] = [];\n\n    // Add BOS token if configured\n    if (this.config.addBosToken && this.config.bosTokenId !== null) {\n      ids.push(this.config.bosTokenId);\n    }\n\n    // Check for special tokens first — split text around them\n    const parts = this.splitOnSpecialTokens(text);\n\n    for (const part of parts) {\n      if (part.special) {\n        const id = this.addedTokens.get(part.text);\n        if (id !== undefined) {\n          ids.push(id);\n        }\n      } else {\n        // Pre-tokenize: split into words/chunks\n        // Use a simple regex-based pre-tokenizer similar to GPT-style\n        const chunks = this.preTokenize(part.text);\n        for (const chunk of chunks) {\n          const chunkIds = this.bpeEncode(chunk);\n          ids.push(...chunkIds);\n        }\n      }\n    }\n\n    // Add EOS token if configured\n    if (this.config.addEosToken && this.config.eosTokenId !== null) {\n      ids.push(this.config.eosTokenId);\n    }\n\n    return ids;\n  }\n\n  /**\n   * Decode token IDs back to text.\n   */\n  decode(ids: number[], skipSpecialTokens: boolean = true): string {\n    const pieces: string[] = [];\n\n    for (const id of ids) {\n      const token = this.vocabReverse.get(id);\n      if (!token) continue;\n\n      // Skip special tokens if requested\n      if (skipSpecialTokens && this.specialTokens.has(token)) {\n        continue;\n      }\n\n      pieces.push(token);\n    }\n\n    if (this.spmMode) {\n      // SentencePiece (Gemma/Llama): tokens are raw UTF-8 with ▁ for spaces and\n      // <0xHH> byte-fallback pieces. Concatenate the <0xHH> bytes into proper\n      // UTF-8 runs (Fuse), decode, and turn ▁ back into spaces.\n      let out = \"\";\n      let byteRun: number[] = [];\n      const flushBytes = () => {\n        if (byteRun.length === 0) return;\n        out += new TextDecoder().decode(new Uint8Array(byteRun));\n        byteRun = [];\n      };\n      for (const piece of pieces) {\n        const m = /^<0x([0-9A-Fa-f]{2})>$/.exec(piece);\n        if (m) {\n          byteRun.push(Number.parseInt(m[1], 16));\n        } else {\n          flushBytes();\n          out += piece;\n        }\n      }\n      flushBytes();\n      return out.replace(/▁/g, \" \");\n    }\n\n    // Join and decode byte-level BPE tokens.\n    // GPT/Qwen tokenizers use a byte-to-unicode mapping where non-printable\n    // bytes are shifted into the U+0100+ range:\n    //   byte 0x00-0x20 → U+0100-0x0120 (e.g. 0x0A=\\n → U+010A=Ċ, 0x20=space → U+0120=Ġ)\n    //   byte 0x7F-0xA0 → U+0121-0x0142\n    // We decode by reversing this mapping back to raw bytes.\n    let text = pieces.join(\"\");\n\n    // Build the reverse byte-level mapping\n    const chars: number[] = [];\n    for (let i = 0; i < text.length; i++) {\n      const cp = text.codePointAt(i)!;\n      const byte = UNICODE_TO_BYTE.get(cp);\n      if (byte !== undefined) {\n        chars.push(byte);\n      } else {\n        // Multi-byte UTF-8: encode the codepoint as UTF-8 bytes\n        const encoded = new TextEncoder().encode(String.fromCodePoint(cp));\n        for (const b of encoded) chars.push(b);\n        if (cp > 0xffff) i++; // skip surrogate pair\n      }\n    }\n    text = new TextDecoder().decode(new Uint8Array(chars));\n\n    // Handle byte-level fallback tokens like <0x0A> → \\n\n    text = text.replace(/<0x([0-9A-Fa-f]{2})>/g, (_, hex) => {\n      return String.fromCharCode(parseInt(hex, 16));\n    });\n\n    return text;\n  }\n\n  /**\n   * Apply chat template to messages.\n   *\n   * For now, implements the common ChatML format used by Qwen models:\n   * <|im_start|>system\\n{content}<|im_end|>\\n\n   * <|im_start|>user\\n{content}<|im_end|>\\n\n   * <|im_start|>assistant\\n\n   *\n   * TODO: Parse Jinja2 templates from tokenizer_config.json for full generality.\n   */\n  /**\n   * Gemma 4 turn format: `<bos><|turn>user\\n{content}<turn|>\\n<|turn>model\\n`.\n   * Gemma has no \"system\" role, so a system message is folded into the next user\n   * turn (matching the reference chat template).\n   */\n  private applyGemmaTurnTemplate(messages: ChatMessage[], addGenPrompt: boolean): string {\n    const parts: string[] = [];\n    if (this.config.bosToken) parts.push(this.config.bosToken);\n    let systemText = \"\";\n    for (const msg of messages) {\n      if (msg.role === \"system\") {\n        systemText += `${msg.content}\\n\\n`;\n        continue;\n      }\n      const role = msg.role === \"assistant\" ? \"model\" : msg.role;\n      const content = role === \"user\" && systemText ? systemText + msg.content : msg.content;\n      if (role === \"user\") systemText = \"\";\n      parts.push(`<|turn>${role}\\n${content}<turn|>\\n`);\n    }\n    if (addGenPrompt) parts.push(\"<|turn>model\\n\");\n    return parts.join(\"\");\n  }\n\n  applyChatTemplate(messages: ChatMessage[], options?: { addGenerationPrompt?: boolean }): string {\n    const addGenPrompt = options?.addGenerationPrompt ?? true;\n    const parts: string[] = [];\n\n    // ── Gemma 4 turn format ──\n    // Detected by the turn-delimiter tokens `<|turn>`/`<turn|>` (unique to the\n    // Gemma 4 vocab, absent from the ChatML/Qwen vocabs handled below).\n    if (this.addedTokens.has(\"<|turn>\") && this.addedTokens.has(\"<turn|>\")) {\n      return this.applyGemmaTurnTemplate(messages, addGenPrompt);\n    }\n\n    // Models whose chat template prepends bos at the start (e.g. LFM2's\n    // `{{- bos_token -}}` ...). The hardcoded ChatML body is otherwise shared.\n    const template = this.config.chatTemplate;\n    if (template?.includes(\"bos_token\") && this.config.bosToken) {\n      parts.push(this.config.bosToken);\n    }\n\n    for (const msg of messages) {\n      parts.push(`<|im_start|>${msg.role}\\n${msg.content}<|im_end|>\\n`);\n    }\n\n    if (addGenPrompt) {\n      parts.push(`<|im_start|>assistant\\n`);\n      // Some reasoning models (e.g. Qwen3/3.5) pre-fill an (empty) think block\n      // after the generation prompt for non-thinking mode. Only inject it when\n      // the model's own chat template actually emits `<think>` at generation\n      // time — otherwise models that merely *have* a <think> token but don't\n      // pre-fill it (e.g. LFM2.5) would be corrupted into degenerate output.\n      const templateWantsThink = template\n        ? template.includes(\"<think>\")\n        : this.addedTokens.has(\"<think>\");\n      if (templateWantsThink) {\n        parts.push(`<think>\\n\\n</think>\\n\\n`);\n      }\n    }\n\n    return parts.join(\"\");\n  }\n\n  /**\n   * Encode a chat conversation into token IDs.\n   */\n  encodeChat(messages: ChatMessage[], options?: { addGenerationPrompt?: boolean }): number[] {\n    const text = this.applyChatTemplate(messages, options);\n    return this.encode(text);\n  }\n\n  // ── Internal BPE implementation ──────────────────────────────────\n\n  private splitOnSpecialTokens(text: string): Array<{ text: string; special: boolean }> {\n    if (this.addedTokens.size === 0) return [{ text, special: false }];\n\n    const parts: Array<{ text: string; special: boolean }> = [];\n    // Sort added tokens by length (longest first) for greedy matching\n    // Uses ALL added tokens (not just special:true) so <think>/<|tool_call|> etc. are split\n    const sortedSpecials = [...this.addedTokens.keys()].sort((a, b) => b.length - a.length);\n\n    // Build regex that matches any special token\n    const escaped = sortedSpecials.map((s) => s.replace(/[.*+?^${}()|[\\]\\\\]/g, \"\\\\$&\"));\n    const regex = new RegExp(`(${escaped.join(\"|\")})`, \"g\");\n\n    let lastIndex = 0;\n    let match: RegExpExecArray | null;\n    while ((match = regex.exec(text)) !== null) {\n      if (match.index > lastIndex) {\n        parts.push({\n          text: text.slice(lastIndex, match.index),\n          special: false,\n        });\n      }\n      parts.push({ text: match[0], special: true });\n      lastIndex = regex.lastIndex;\n    }\n    if (lastIndex < text.length) {\n      parts.push({ text: text.slice(lastIndex), special: false });\n    }\n\n    return parts;\n  }\n\n  private preTokenize(text: string): string[] {\n    if (!text) return [];\n\n    if (this.spmMode) {\n      // SentencePiece (Gemma/Llama): normalize spaces → ▁ (U+2581), then split on\n      // the ▁ boundary keeping it with the FOLLOWING piece (Split \" \"\n      // MergedWithPrevious on the original, equivalent to ▁-prefixed words).\n      // BPE then runs on the raw ▁-encoded text — no byte-to-unicode mapping.\n      const encoded = text.replace(/ /g, \"▁\");\n      const chunks: string[] = [];\n      let i = 0;\n      while (i < encoded.length) {\n        let j = i + 1;\n        while (j < encoded.length && encoded[j] !== \"▁\") j++;\n        chunks.push(encoded.slice(i, j));\n        i = j;\n      }\n      return chunks;\n    }\n\n    // GPT-style pre-tokenization regex\n    // Splits on whitespace boundaries, keeping the space as a prefix (Ġ convention)\n    // Also splits numbers and punctuation\n    const pattern = /'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+/gu;\n    const matches = text.match(pattern);\n    if (!matches) return [];\n\n    // Convert to the byte-level representation HF uses:\n    // space → Ġ, etc. (the GPT2 byte-to-unicode mapping)\n    return matches.map((chunk) => this.textToTokenRepr(chunk));\n  }\n\n  private textToTokenRepr(text: string): string {\n    // Convert text to the representation used in the vocab\n    // In HF BPE tokenizers, spaces are typically represented as Ġ (U+0120)\n    // and other bytes have specific unicode mappings\n    let result = \"\";\n    for (let i = 0; i < text.length; i++) {\n      const c = text.charCodeAt(i);\n      if (c === 32) {\n        result += \"\\u0120\"; // space → Ġ\n      } else if (c < 33 || (c > 126 && c < 161)) {\n        // Control chars and non-printable → byte-level tokens\n        // Use the GPT2 byte encoder mapping\n        result += String.fromCharCode(c + 256);\n      } else {\n        result += text[i];\n      }\n    }\n    return result;\n  }\n\n  private bpeEncode(word: string): number[] {\n    if (word.length === 0) return [];\n\n    // Check if the whole word is in vocab\n    if (this.vocab.has(word)) {\n      return [this.vocab.get(word)!];\n    }\n\n    // Split into individual characters\n    let symbols = [...word];\n\n    // If any symbol isn't in vocab, try byte fallback\n    // Convert to individual characters and check\n    if (symbols.length === 1) {\n      const id = this.vocab.get(symbols[0]);\n      if (id !== undefined) return [id];\n      // Byte fallback\n      return this.encodeByteFallback(word);\n    }\n\n    // Iteratively merge the most frequent pair\n    while (symbols.length > 1) {\n      // Find the pair with the lowest merge rank (highest priority)\n      let _bestPair = \"\";\n      let bestRank = Infinity;\n      let bestIdx = -1;\n\n      for (let i = 0; i < symbols.length - 1; i++) {\n        const pair = `${symbols[i]} ${symbols[i + 1]}`;\n        const rank = this.merges.get(pair);\n        if (rank !== undefined && rank < bestRank) {\n          bestRank = rank;\n          _bestPair = pair;\n          bestIdx = i;\n        }\n      }\n\n      // No more merges possible\n      if (bestIdx === -1) break;\n\n      // Apply the merge\n      const merged = symbols[bestIdx] + symbols[bestIdx + 1];\n      symbols = [...symbols.slice(0, bestIdx), merged, ...symbols.slice(bestIdx + 2)];\n    }\n\n    // Convert symbols to IDs\n    const ids: number[] = [];\n    for (const sym of symbols) {\n      const id = this.vocab.get(sym);\n      if (id !== undefined) {\n        ids.push(id);\n      } else {\n        // Byte fallback for unknown symbols\n        ids.push(...this.encodeByteFallback(sym));\n      }\n    }\n    return ids;\n  }\n\n  private encodeByteFallback(text: string): number[] {\n    const encoder = new TextEncoder();\n    const bytes = encoder.encode(text);\n    const ids: number[] = [];\n    for (const b of bytes) {\n      const tok = this.byteFallback.get(b);\n      if (tok) {\n        const id = this.vocab.get(tok);\n        if (id !== undefined) ids.push(id);\n      }\n    }\n    return ids;\n  }\n}\n","/**\n * WeightSource — a tensor store the loader/executor can pull from one tensor at a\n * time, so the *whole* model never has to sit in the JS heap simultaneously.\n *\n * Two backends share one async-shaped interface:\n *\n *  • HeapWeightStore (Node / desktop) — wraps a plain `Map`. Identical behavior\n *    to the original code path: tensors are held in heap, gets/sets are O(1) heap\n *    ops (the async methods resolve synchronously). Zero behavior change.\n *\n *  • CacheWeightStore (browser / mobile) — keeps every tensor's BYTES out of the\n *    heap. Downloaded tensors live in the per-tensor download cache (IndexedDB,\n *    populated by model-loader's browserCacheWrite); transform OUTPUTS written by\n *    `set()` stage into a CacheStorage bucket. The heap holds only lightweight\n *    descriptors ({ shape, dtype, cacheName, cacheKey }) plus a small set of tiny\n *    \"override\" tensors. `get()` reads one tensor's bytes back on demand —\n *    IndexedDB for downloads, CacheStorage for staged outputs — applying any\n *    BF16/F16→F32 conversion. This bounds peak heap to roughly ONE tensor at a\n *    time instead of the entire weight set, which is what lets a 2.5 GB model\n *    load on iOS Safari without the tab being jetsam-killed.\n *\n * The browser path is selected automatically when CacheStorage is available.\n */\n\nimport { idbCacheAvailable, idbGet } from \"./idb-cache.js\";\nimport { opfsAvailable, opfsRead } from \"./opfs-cache.js\";\nimport { bf16ToF32, f16ToF32, type SafetensorDType } from \"./safetensors.js\";\n\n/** A single tensor's data + shape (the unit the executor uploads to a GPU buffer). */\nexport interface WeightEntry {\n  data: ArrayBufferView;\n  shape: number[];\n}\n\n/**\n * Read-side view consumed by the executor's streaming `uploadWeights`. Async by\n * design so a cache-backed store can fetch one tensor's bytes at a time.\n */\nexport interface WeightSource {\n  has(name: string): boolean;\n  keys(): string[];\n  readonly size: number;\n  /** Pull a single tensor (bytes materialized + dtype-converted on demand). */\n  get(name: string): Promise<WeightEntry | undefined>;\n  /**\n   * Release any transient backing storage (e.g. the browser transform-staging\n   * cache) once the consumer has finished uploading. No-op for the heap backend.\n   */\n  dispose?(): Promise<void>;\n}\n\n/**\n * Mutable store used while loading. The loader's post-download transforms\n * (RMSNorm +1 bake, fused-Q split, GPTQ/MLX repack, INT4 quantize, PLE/vision\n * fixups) read tensors, produce new ones, and drop sources through this\n * interface. All ops are async so the cache backend can stream bytes; the heap\n * backend resolves them synchronously.\n */\nexport interface MutableWeightStore extends WeightSource {\n  set(name: string, entry: WeightEntry): Promise<void>;\n  delete(name: string): Promise<void>;\n}\n\n// ── Heap backend (Node / desktop) — unchanged behavior ──────────────────────\n\nexport class HeapWeightStore implements MutableWeightStore {\n  private map: Map<string, WeightEntry>;\n\n  constructor(map?: Map<string, WeightEntry>) {\n    this.map = map ?? new Map();\n  }\n\n  has(name: string): boolean {\n    return this.map.has(name);\n  }\n\n  keys(): string[] {\n    return [...this.map.keys()];\n  }\n\n  get size(): number {\n    return this.map.size;\n  }\n\n  get(name: string): Promise<WeightEntry | undefined> {\n    return Promise.resolve(this.map.get(name));\n  }\n\n  set(name: string, entry: WeightEntry): Promise<void> {\n    this.map.set(name, entry);\n    return Promise.resolve();\n  }\n\n  delete(name: string): Promise<void> {\n    this.map.delete(name);\n    return Promise.resolve();\n  }\n\n  /** Escape hatch for the few Node paths that still want the raw Map. */\n  asMap(): Map<string, WeightEntry> {\n    return this.map;\n  }\n}\n\n// ── Cache backend (browser / mobile) — bounded peak heap ─────────────────────\n\n/**\n * A descriptor of a tensor whose bytes live in a durable cache rather than the\n * heap: IndexedDB for downloaded tensors, CacheStorage for `set()` staging\n * outputs. `dtype` is the *source* safetensors dtype so `get()` can reproduce the\n * same BF16/F16→F32 conversion the heap path applies inline during download.\n */\ninterface CacheDescriptor {\n  shape: number[];\n  /** Source safetensors dtype, or \"RAW\" for already-converted/derived bytes. */\n  dtype: SafetensorDType | \"RAW\";\n  /** Logical bucket tag: download cache (IndexedDB by cacheKey) vs staging cache (CacheStorage). */\n  cacheName: string;\n  cacheKey: string;\n  /** TypedArray ctor tag for RAW (transform-produced) bytes. */\n  view?: \"f32\" | \"u32\" | \"i32\" | \"u16\" | \"u8\";\n}\n\n/**\n * Build a typed-array view over a freshly-read (offset-0) ArrayBuffer. A cache\n * read always returns a fresh, offset-0 buffer, so it is aligned for every dtype\n * — no copy needed (unlike the merged-range download path which can be unaligned).\n */\nfunction viewForSourceDtype(buf: ArrayBuffer, dtype: SafetensorDType): ArrayBufferView {\n  switch (dtype) {\n    case \"F32\":\n      return new Float32Array(buf, 0, buf.byteLength / 4);\n    case \"I32\":\n      return new Int32Array(buf, 0, buf.byteLength / 4);\n    case \"U32\":\n      return new Uint32Array(buf, 0, buf.byteLength / 4);\n    case \"F16\":\n    case \"BF16\":\n      return new Uint16Array(buf, 0, buf.byteLength / 2);\n    default:\n      return new Uint8Array(buf, 0, buf.byteLength);\n  }\n}\n\nfunction viewForRaw(buf: ArrayBuffer, tag: CacheDescriptor[\"view\"]): ArrayBufferView {\n  switch (tag) {\n    case \"u32\":\n      return new Uint32Array(buf, 0, buf.byteLength / 4);\n    case \"i32\":\n      return new Int32Array(buf, 0, buf.byteLength / 4);\n    case \"u16\":\n      return new Uint16Array(buf, 0, buf.byteLength / 2);\n    case \"u8\":\n      return new Uint8Array(buf, 0, buf.byteLength);\n    default:\n      return new Float32Array(buf, 0, buf.byteLength / 4);\n  }\n}\n\nfunction rawTag(data: ArrayBufferView): CacheDescriptor[\"view\"] {\n  if (data instanceof Float32Array) return \"f32\";\n  if (data instanceof Uint32Array) return \"u32\";\n  if (data instanceof Int32Array) return \"i32\";\n  if (data instanceof Uint16Array) return \"u16\";\n  return \"u8\";\n}\n\nconst CACHE_BACKED_NAME = \"gerbil-weights-staging-v2\";\n\nexport class CacheWeightStore implements MutableWeightStore {\n  private descriptors = new Map<string, CacheDescriptor>();\n  /**\n   * Small heap-resident tensors that are NOT worth a cache round-trip: synthetic\n   * fill constants, sub-MB norm/scalar tensors, and anything a caller explicitly\n   * pins (e.g. the vision pos-embed table that index.ts snapshots before upload).\n   * Keeping these in heap is fine — they are individually tiny.\n   */\n  private overrides = new Map<string, WeightEntry>();\n  private cacheName: string;\n\n  /** Tensors at or below this size stay in heap rather than round-tripping cache. */\n  private static readonly HEAP_THRESHOLD = 512 * 1024; // 512 KB\n\n  constructor(cacheName: string = CACHE_BACKED_NAME) {\n    this.cacheName = cacheName;\n  }\n\n  private static stagingKey(name: string): string {\n    return `gerbil-staging:${name}`;\n  }\n\n  /**\n   * Register a tensor whose bytes already live in CacheStorage under\n   * `downloadCacheName`/`cacheKey` (the per-tensor download cache). No bytes are\n   * retained in heap. The descriptor records which cache bucket to read from so\n   * `get()` opens the right one (download cache vs this store's staging cache).\n   */\n  registerCached(\n    name: string,\n    shape: number[],\n    dtype: SafetensorDType,\n    cacheKey: string,\n    downloadCacheName: string,\n  ): void {\n    this.overrides.delete(name);\n    this.descriptors.set(name, { shape, dtype, cacheName: downloadCacheName, cacheKey });\n  }\n\n  has(name: string): boolean {\n    return this.overrides.has(name) || this.descriptors.has(name);\n  }\n\n  keys(): string[] {\n    const set = new Set<string>(this.descriptors.keys());\n    for (const k of this.overrides.keys()) set.add(k);\n    return [...set];\n  }\n\n  get size(): number {\n    return this.keys().length;\n  }\n\n  async get(name: string): Promise<WeightEntry | undefined> {\n    const pinned = this.overrides.get(name);\n    if (pinned) return pinned;\n    const desc = this.descriptors.get(name);\n    if (!desc) return undefined;\n\n    // RAW entries are transform outputs written by set() straight into this\n    // store's CacheStorage staging bucket — read them back from there.\n    if (desc.dtype === \"RAW\") {\n      const cache = await caches.open(desc.cacheName);\n      const resp = await cache.match(new Request(desc.cacheKey));\n      if (!resp) return undefined;\n      const rawBuf = await resp.arrayBuffer();\n      return { data: viewForRaw(rawBuf, desc.view), shape: desc.shape };\n    }\n\n    // Download tensors (registerCached) live in the per-tensor download cache,\n    // which is now IndexedDB — the SAME backend model-loader's browserCacheWrite\n    // writes to. Read from IDB first; fall back to OPFS / CacheStorage so any\n    // tensor cached by an older build still resolves. (Reading only OPFS +\n    // CacheStorage — the pre-IDB backends — is what made every streamed get()\n    // miss and crash model-loader's `(await store.get(...))!.data`.)\n    let buf: ArrayBuffer | null = null;\n    if (idbCacheAvailable()) {\n      buf = await idbGet(desc.cacheKey);\n    }\n    if (!buf && opfsAvailable()) {\n      buf = await opfsRead(desc.cacheKey);\n    }\n    if (!buf) {\n      const cache = await caches.open(desc.cacheName);\n      const resp = await cache.match(new Request(desc.cacheKey));\n      if (!resp) return undefined;\n      buf = await resp.arrayBuffer();\n    }\n    let data = viewForSourceDtype(buf, desc.dtype);\n    // Mirror the heap path's inline conversion (model-loader.ts ~860-933).\n    if (desc.dtype === \"BF16\") {\n      data = bf16ToF32(new Uint8Array(data.buffer, data.byteOffset, data.byteLength));\n    } else if (desc.dtype === \"F16\") {\n      data = f16ToF32(data);\n    }\n    return { data, shape: desc.shape };\n  }\n\n  async set(name: string, entry: WeightEntry): Promise<void> {\n    this.descriptors.delete(name);\n    // Tiny tensors stay in heap; large ones write through to cache so their bytes\n    // do not accumulate in the heap as transforms run layer-by-layer.\n    if (entry.data.byteLength <= CacheWeightStore.HEAP_THRESHOLD) {\n      this.overrides.set(name, entry);\n      return;\n    }\n    this.overrides.delete(name);\n    const cacheKey = CacheWeightStore.stagingKey(name);\n    const bytes = entry.data.buffer.slice(\n      entry.data.byteOffset,\n      entry.data.byteOffset + entry.data.byteLength,\n    ) as ArrayBuffer;\n    const cache = await caches.open(this.cacheName);\n    await cache.put(new Request(cacheKey), new Response(bytes));\n    this.descriptors.set(name, {\n      shape: entry.shape,\n      dtype: \"RAW\",\n      cacheName: this.cacheName,\n      cacheKey,\n      view: rawTag(entry.data),\n    });\n  }\n\n  async delete(name: string): Promise<void> {\n    this.overrides.delete(name);\n    const desc = this.descriptors.get(name);\n    this.descriptors.delete(name);\n    // Only evict staging entries we created; never evict the durable per-tensor\n    // download cache (a future load reuses those by stable byte-range key).\n    if (desc && desc.dtype === \"RAW\" && desc.cacheName === this.cacheName) {\n      try {\n        const cache = await caches.open(this.cacheName);\n        await cache.delete(new Request(desc.cacheKey));\n      } catch {\n        /* best-effort */\n      }\n    }\n  }\n\n  /** Drop all transform-staging cache entries (call after upload completes). */\n  async dispose(): Promise<void> {\n    try {\n      await caches.delete(this.cacheName);\n    } catch {\n      /* best-effort */\n    }\n  }\n}\n\n/** True when the browser CacheStorage API is available (i.e. not Node). */\nexport function cacheStorageAvailable(): boolean {\n  return typeof caches !== \"undefined\";\n}\n","/**\n * Model loader -- fetches model files from HuggingFace Hub.\n *\n * Handles:\n * - Fetching config.json, tokenizer.json, tokenizer_config.json\n * - Downloading safetensors weight files (with progress callbacks)\n * - HF key -> canonical key mapping\n * - Uploading weight data to GPU buffers via the Executor\n */\n\nimport { type GraphDType, generateGraph, type KVDType } from \"./architectures/index.js\";\nimport { foldNanoCodecWeightNorm, nanoCodecWeightMap } from \"./architectures/kani_tts.js\";\nimport { repackGPTQ } from \"./gptq-adapter.js\";\nimport { idbAllKeys, idbCacheAvailable, idbGet, idbHas, idbPut } from \"./idb-cache.js\";\nimport type { ModelGraph } from \"./ir.js\";\nimport { CANONICAL_KEYS, createDefaultHFKeyMapper, type HFKeyMapper } from \"./ir.js\";\nimport { repackMLX } from \"./mlx-adapter.js\";\nimport { opfsEvictStale } from \"./opfs-cache.js\";\nimport { DEFAULT_GROUP_SIZE, quantizeInt4 } from \"./quantize.js\";\nimport {\n  bf16ToF32,\n  f16ToF32,\n  getTensorData,\n  parseSafetensorsHeader,\n  type SafetensorEntry,\n  type SafetensorsFile,\n} from \"./safetensors.js\";\nimport { Tokenizer } from \"./tokenizer.js\";\nimport {\n  CacheWeightStore,\n  cacheStorageAvailable,\n  HeapWeightStore,\n  type MutableWeightStore,\n  type WeightSource,\n} from \"./weight-source.js\";\n\nexport interface LoadModelOptions {\n  /** HF repo ID (e.g. \"Qwen/Qwen3.5-0.8B\") or full URL. */\n  repo: string;\n  /** Progress callback: (loaded, total, message) */\n  onProgress?: (loaded: number, total: number, message: string) => void;\n  /** Custom HF key mapper (defaults to stripping \"model.\" prefix). */\n  keyMapper?: HFKeyMapper;\n  /** HuggingFace API token for gated models. */\n  hfToken?: string;\n  /** Revision/branch (default: \"main\"). */\n  revision?: string;\n  /** Local cache directory for downloaded files (Node.js only). */\n  cacheDir?: string;\n  /**\n   * Weight dtype:\n   *  - \"f32\"  full precision (or the repo's native quantization, e.g. MLX/GPTQ q4)\n   *  - \"q4\"   on-the-fly INT4 quantization (~4× smaller)\n   *  - \"auto\" (recommended) picks q4 on mobile (iOS/Android) to fit in device\n   *           memory and f32/native on desktop. Already-quantized repos\n   *           (MLX/GPTQ 4-bit) stay q4 regardless.\n   */\n  dtype?: GraphDType | \"auto\";\n  /** KV cache dtype: \"f16\" halves memory traffic during attention. Requires GPU f16 support. */\n  kvDtype?: KVDType;\n  /**\n   * Build an embedding graph (last-token pool + L2 norm) instead of an LM head.\n   * Only valid for Qwen2/Qwen3 CausalLM architectures (e.g. Qwen3-Embedding).\n   */\n  embedding?: boolean;\n  /**\n   * Build the multimodal LM graph variant (M-RoPE + image-embedding splice) so\n   * the text model can consume spliced image tokens. Only meaningful for\n   * Qwen3_5ForConditionalGeneration. Reserves `maxVisionTokens` rows for the\n   * vision-embedding buffer. Text-only generation through this graph is\n   * numerically identical to the non-multimodal graph (M-RoPE fed linear\n   * positions == standard 1D RoPE).\n   */\n  multimodal?: { maxVisionTokens: number };\n  /**\n   * Force-download and key-map the vision tower even without the multimodal LM\n   * graph. `enableVision` (via `multimodal`) already implies this; this flag is\n   * for callers that load weights directly (e.g. the vision-encoder validation\n   * scripts) and build the vision graph/executor themselves. When neither this\n   * nor `multimodal` is set, the ~201MB ViT is excluded from the download.\n   */\n  loadVisionTower?: boolean;\n}\n\n/**\n * Gemma 4 Per-Layer-Embeddings (PLE) source, kept CPU-resident.\n *\n * The PLE table (`embed_tokens_per_layer`, [vocab, num_layers*256]) is ~1.17GB\n * at 4-bit. Uploading it to a GPU buffer would make the model non-mobile-viable\n * and would hit the per-binding size cap. Instead the loader hands the quantized\n * table to the executor in JS memory; the executor gathers + dequantizes only the\n * rows for the current input tokens each forward step (a tiny [T, width] upload).\n */\nexport interface PleSource {\n  /**\n   * Flat row-major INT4 nibbles (Gerbil packing, 8 per u32). HEAP-RESIDENT path\n   * (Node/desktop). In the browser this is empty and `cache` is set instead so\n   * the ~1.17 GB table never sits in the JS heap during load.\n   */\n  packed: Uint32Array;\n  /** Per-group scales (Gerbil (nibble - zero) * scale convention). */\n  scales: Float32Array;\n  /** Per-group zero points. */\n  zeros: Float32Array;\n  /** Row width = num_layers * hidden_size_per_layer_input (E2B: 35*256 = 8960). */\n  width: number;\n  /** Dequant group size (MLX: 64). */\n  groupSize: number;\n  /** Activation tensor the per-step gathered rows are written into. */\n  targetTensor: string;\n  /**\n   * Browser only: when set, the quantized PLE table's bytes live in CacheStorage\n   * (not the heap). The executor reads the slice of nibbles/scales/zeros it needs\n   * for the current tokens on demand. Keeps peak load heap bounded.\n   */\n  cache?: {\n    cacheName: string;\n    packedKey: string;\n    scalesKey: string;\n    zerosKey: string;\n    /** packed.length (u32 count) — for bounds/Range math. */\n    packedLen: number;\n  };\n}\n\nexport interface LoadedModel {\n  /** The generated computation graph (IR). */\n  graph: ModelGraph;\n  /** The tokenizer. */\n  tokenizer: Tokenizer;\n  /**\n   * Weight tensors mapped to canonical names. A `WeightSource` so the executor\n   * can pull one tensor at a time (cache-backed in the browser, heap-backed on\n   * Node) instead of requiring the whole model to sit in heap at once. Use\n   * `get(name)` (async) to materialize a tensor's bytes on demand.\n   */\n  weights: WeightSource;\n  /** Raw config.json for reference. */\n  rawConfig: Record<string, unknown>;\n  /**\n   * CPU-resident Gemma 4 PLE table (set only for Gemma 4). Pass to\n   * `executor.setPleSource()` so the big table never becomes GPU-resident.\n   */\n  pleSource?: PleSource;\n}\n\n/**\n * Resolve a repo string to a base URL for HuggingFace Hub API.\n */\nfunction resolveHFBaseURL(repo: string, revision: string): string {\n  // If it's already a full URL, use it\n  if (repo.startsWith(\"http://\") || repo.startsWith(\"https://\")) {\n    return repo;\n  }\n  // Standard HF Hub URL\n  return `https://huggingface.co/${repo}/resolve/${revision}`;\n}\n\n// ── File cache helpers (Node.js only) ──────────────────────────────────\n\nlet _fs: typeof import(\"node:fs\") | null = null;\nlet _path: typeof import(\"node:path\") | null = null;\nlet _fsInitialized = false;\n\nasync function initFs(): Promise<void> {\n  if (_fsInitialized) return;\n  _fsInitialized = true;\n  try {\n    _fs = await import(\"node:fs\");\n    _path = await import(\"node:path\");\n  } catch {\n    // Not in Node.js (browser) — caching disabled\n  }\n}\n\nfunction getCachePath(cacheDir: string, filename: string): string {\n  return _path!.join(cacheDir, filename.replace(/\\//g, \"_\"));\n}\n\nfunction readCachedFile(cacheDir: string, filename: string): Buffer | null {\n  if (!_fs || !_path) return null;\n  const p = getCachePath(cacheDir, filename);\n  try {\n    return _fs.readFileSync(p);\n  } catch {\n    return null;\n  }\n}\n\nfunction writeCacheFile(\n  cacheDir: string,\n  filename: string,\n  data: ArrayBuffer | Buffer | Uint8Array,\n): void {\n  if (!_fs || !_path) return;\n  _fs.mkdirSync(cacheDir, { recursive: true });\n  const p = getCachePath(cacheDir, filename);\n  let buf: Buffer;\n  if (Buffer.isBuffer(data)) {\n    buf = data;\n  } else if (data instanceof Uint8Array) {\n    buf = Buffer.from(data);\n  } else {\n    buf = Buffer.from(data);\n  }\n  _fs.writeFileSync(p, buf);\n}\n\n// ── Browser cache helpers (OPFS-durable, CacheStorage fallback) ─────────\n//\n// Two-tier durable cache behind a single read/write/has seam:\n//\n//  1. OPFS via a dedicated Worker (`createSyncAccessHandle`, chunked writes) —\n//     the ONLY iOS-safe durable path. On iOS Safari the CacheStorage API is\n//     evicted under quota pressure in a plain tab (`navigator.storage.persist()`\n//     is not granted outside an installed PWA), so weights cached there are\n//     re-downloaded every visit. OPFS survives a full tab close + reopen. Main-\n//     thread OPFS `createWritable` throws \"out of quota\" mid-write on iOS, so all\n//     OPFS I/O runs on a blob-URL worker with sync access handles. See\n//     src/gpu/opfs-cache.ts.\n//\n//  2. CacheStorage (`caches.open(...)`) — fallback when OPFS/Worker is\n//     unavailable (older browsers, some embedded webviews). Best-effort durable;\n//     fine on desktop where quota pressure is rare.\n//\n// Reads prefer OPFS, then fall back to CacheStorage. Writes go to OPFS when\n// available and only to CacheStorage otherwise (no double-write — OPFS is the\n// durable winner and double-storing would waste the ~404 MB twice against quota).\n// All of this is inert on Node: `opfsAvailable()` and `typeof caches` are both\n// false there, so the helpers short-circuit and the Node weight path is\n// unaffected.\n\n// v4: storage layout changed (OPFS subdirectory `gerbil-models-v4` + CacheStorage\n// bucket of the same name). A fresh namespace also abandons any entries poisoned\n// during earlier buggy builds (the reload-loop era wrote right-size-wrong-content\n// tensors that size validation can't detect). The per-tensor cache logic itself\n// is verified correct by scripts/engine/test-cache-roundtrip.mjs (load → cache →\n// reload-from-cache → identical coherent output).\nconst BROWSER_CACHE_NAME = \"gerbil-models-v4\";\n\n// Per-tensor weight caching. ON by default — the round-trip test\n// (scripts/engine/test-cache-roundtrip.mjs) proves the cache read path is\n// correct, and the v3 namespace abandons entries poisoned by earlier buggy\n// builds. Can be force-disabled via GERBIL_TENSOR_CACHE=0 / globalThis flag if a\n// corruption issue ever resurfaces.\nconst TENSOR_CACHE_ENABLED =\n  (typeof process === \"undefined\" || process.env?.GERBIL_TENSOR_CACHE !== \"0\") &&\n  (globalThis as { GERBIL_TENSOR_CACHE?: boolean }).GERBIL_TENSOR_CACHE !== false;\n\nlet _persistRequested = false;\nasync function requestPersistence(): Promise<void> {\n  if (_persistRequested) return;\n  _persistRequested = true;\n  try {\n    // Only granted in an installed PWA on iOS; harmless best-effort elsewhere.\n    await (navigator as any)?.storage?.persist?.();\n  } catch {\n    /* best-effort */\n  }\n}\n\n// Model weights are cached in IndexedDB (see idb-cache.ts for why: the Cache API's\n// fragment-stripping collapsed our per-tensor keys → wrong bytes/garbage, and OPFS\n// via Worker was fragile on iOS). IDB keys on the exact string — no footguns. All\n// three helpers are inert on Node (no indexedDB → miss/no-op), so the heap load\n// path is unchanged off the browser.\n\n/** Read a cache key's bytes from IndexedDB, or null on miss. */\nasync function browserCacheRead(cacheKey: string): Promise<ArrayBuffer | null> {\n  return idbGet(cacheKey);\n}\n\n/** Cheap presence check — true without materializing the body. */\nasync function browserCacheHas(cacheKey: string): Promise<boolean> {\n  return idbHas(cacheKey);\n}\n\n/**\n * Write a cache key's bytes to IndexedDB (best-effort; never throws). The\n * `transferable` flag is accepted for call-site compatibility but unused — IDB\n * structured-clones the value, so the caller's buffer is never detached.\n */\nasync function browserCacheWrite(\n  cacheKey: string,\n  data: ArrayBuffer,\n  _transferable = false,\n): Promise<void> {\n  await requestPersistence();\n  await idbPut(cacheKey, data);\n}\n\n/**\n * Reclaim quota from the LEGACY caches, ONCE per session. The model cache is now\n * IndexedDB; the old CacheStorage (`gerbil-models-*`, fragment-keyed → the `RRRR`\n * corruption) and OPFS-worker stores are dead. iOS Safari's per-origin quota is\n * ~1 GB and that orphaned junk could consume nearly all of it, so we purge ALL of\n * it to free room for the IDB cache. Best-effort; inert on Node.\n */\nlet _evictedStale = false;\nasync function evictStaleCaches(): Promise<void> {\n  if (_evictedStale) return;\n  _evictedStale = true;\n  try {\n    if (typeof caches !== \"undefined\") {\n      const keys = await caches.keys();\n      await Promise.all(\n        keys.filter((k) => k.startsWith(\"gerbil-models-\")).map((k) => caches.delete(k)),\n      );\n    }\n  } catch {\n    /* best-effort */\n  }\n  await opfsEvictStale();\n}\n\n// ── Fetch helpers ──────────────────────────────────────────────────────\n\n/**\n * Fetch a JSON file from HF with optional auth and caching.\n */\nasync function fetchJSON(\n  baseURL: string,\n  filename: string,\n  hfToken?: string,\n  cacheDir?: string,\n): Promise<any> {\n  // Check Node.js cache first\n  if (cacheDir) {\n    const cached = readCachedFile(cacheDir, filename);\n    if (cached) return JSON.parse(cached.toString(\"utf-8\"));\n  }\n\n  // Check browser cache\n  const browserKey = `${baseURL}/${filename}`;\n  const browserCached = await browserCacheRead(browserKey);\n  if (browserCached) {\n    return JSON.parse(new TextDecoder().decode(browserCached));\n  }\n\n  const url = `${baseURL}/${filename}`;\n  const headers: Record<string, string> = {};\n  if (hfToken) {\n    headers.Authorization = `Bearer ${hfToken}`;\n  }\n  const response = await fetch(url, { headers });\n  if (!response.ok) {\n    throw new Error(`Failed to fetch ${url}: ${response.status} ${response.statusText}`);\n  }\n  const text = await response.text();\n\n  // Write to Node.js cache\n  if (cacheDir) {\n    writeCacheFile(cacheDir, filename, Buffer.from(text, \"utf-8\"));\n  }\n\n  // Write to browser cache\n  await browserCacheWrite(browserKey, new TextEncoder().encode(text).buffer);\n\n  return JSON.parse(text);\n}\n\n/**\n * Fetch a UTF-8 text file from HF (with caching). Returns null on 404 / error.\n * Used for sidecar files like `chat_template.jinja` that some models ship\n * instead of an inline `chat_template` in tokenizer_config.json.\n */\nasync function fetchText(\n  baseURL: string,\n  filename: string,\n  hfToken?: string,\n  cacheDir?: string,\n): Promise<string | null> {\n  if (cacheDir) {\n    const cached = readCachedFile(cacheDir, filename);\n    if (cached) return cached.toString(\"utf-8\");\n  }\n  const browserKey = `${baseURL}/${filename}`;\n  const browserCached = await browserCacheRead(browserKey);\n  if (browserCached) {\n    return new TextDecoder().decode(browserCached);\n  }\n  const url = `${baseURL}/${filename}`;\n  const headers: Record<string, string> = {};\n  if (hfToken) {\n    headers.Authorization = `Bearer ${hfToken}`;\n  }\n  const response = await fetch(url, { headers });\n  if (!response.ok) return null;\n  const text = await response.text();\n  if (cacheDir) {\n    writeCacheFile(cacheDir, filename, Buffer.from(text, \"utf-8\"));\n  }\n  await browserCacheWrite(browserKey, new TextEncoder().encode(text).buffer);\n  return text;\n}\n\n/**\n * Fetch a binary file from HF with progress tracking and caching.\n */\nasync function fetchBinary(\n  baseURL: string,\n  filename: string,\n  hfToken?: string,\n  onProgress?: (loaded: number, total: number) => void,\n  cacheDir?: string,\n): Promise<ArrayBuffer> {\n  // Check cache first\n  if (cacheDir) {\n    const cached = readCachedFile(cacheDir, filename);\n    if (cached) {\n      onProgress?.(cached.byteLength, cached.byteLength);\n      return cached.buffer.slice(\n        cached.byteOffset,\n        cached.byteOffset + cached.byteLength,\n      ) as ArrayBuffer;\n    }\n  }\n\n  const url = `${baseURL}/${filename}`;\n  const headers: Record<string, string> = {};\n  if (hfToken) {\n    headers.Authorization = `Bearer ${hfToken}`;\n  }\n\n  const response = await fetch(url, { headers });\n  if (!response.ok) {\n    throw new Error(`Failed to fetch ${url}: ${response.status} ${response.statusText}`);\n  }\n\n  const contentLength = Number(response.headers.get(\"content-length\") || 0);\n\n  if (!response.body || !onProgress) {\n    const buf = await response.arrayBuffer();\n    if (cacheDir) writeCacheFile(cacheDir, filename, buf);\n    return buf;\n  }\n\n  // Stream with progress\n  const reader = response.body.getReader();\n  const chunks: Uint8Array[] = [];\n  let loaded = 0;\n\n  while (true) {\n    const { done, value } = await reader.read();\n    if (done) break;\n    chunks.push(value);\n    loaded += value.byteLength;\n    onProgress(loaded, contentLength);\n  }\n\n  // Combine chunks into single ArrayBuffer\n  const result = new Uint8Array(loaded);\n  let offset = 0;\n  for (const chunk of chunks) {\n    result.set(chunk, offset);\n    offset += chunk.byteLength;\n  }\n\n  if (cacheDir) writeCacheFile(cacheDir, filename, result);\n  return result.buffer;\n}\n\n/**\n * Discover safetensors files in the repo.\n *\n * Models can have a single `model.safetensors` or multiple shards:\n * `model-00001-of-00002.safetensors`, `model-00002-of-00002.safetensors`\n *\n * We check for the index file first, then fall back to single file.\n */\nasync function discoverSafetensorsFiles(\n  baseURL: string,\n  hfToken?: string,\n  cacheDir?: string,\n): Promise<string[]> {\n  // Try the index file first (for sharded models)\n  try {\n    const index = await fetchJSON(baseURL, \"model.safetensors.index.json\", hfToken, cacheDir);\n    if (index.weight_map) {\n      // Get unique filenames from the weight map\n      const files = [...new Set(Object.values(index.weight_map))] as string[];\n      return files.sort();\n    }\n  } catch {\n    // No index file -- try single file\n  }\n\n  return [\"model.safetensors\"];\n}\n\n/**\n * Fetch a byte range from a URL over the network (no caching).\n */\nasync function fetchRangeRaw(\n  url: string,\n  start: number,\n  end: number,\n  hfToken?: string,\n): Promise<ArrayBuffer> {\n  const headers: Record<string, string> = {\n    Range: `bytes=${start}-${end - 1}`,\n  };\n  if (hfToken) {\n    headers.Authorization = `Bearer ${hfToken}`;\n  }\n  // Retry transient failures with backoff. HuggingFace rate-limits unauthenticated\n  // downloads (403/429), and CDN/network hiccups happen; a few backed-off retries\n  // turn those into a slightly slower load instead of a hard failure. (Once the\n  // model is cached in IDB these requests don't happen again.)\n  let lastStatus = 0;\n  for (let attempt = 0; attempt < 4; attempt++) {\n    let resp: Response;\n    try {\n      resp = await fetch(url, { headers });\n    } catch {\n      // Network error — back off and retry.\n      if (attempt < 3) {\n        await new Promise((r) => setTimeout(r, 600 * 2 ** attempt));\n        continue;\n      }\n      throw new Error(`Range request failed: network error for ${url}`);\n    }\n    if (resp.ok || resp.status === 206) return resp.arrayBuffer();\n    lastStatus = resp.status;\n    // 403/429 = rate limit (HF), 5xx = transient server error → retry; else fail.\n    const retriable = resp.status === 403 || resp.status === 429 || resp.status >= 500;\n    if (retriable && attempt < 3) {\n      await new Promise((r) => setTimeout(r, 600 * 2 ** attempt));\n      continue;\n    }\n    break;\n  }\n  throw new Error(\n    `Range request failed: ${lastStatus}${lastStatus === 403 || lastStatus === 429 ? \" (HuggingFace rate limit — wait a bit and retry; once cached this won't recur)\" : \"\"}`,\n  );\n}\n\n/**\n * Fetch a byte range from a URL, caching the exact range.\n *\n * Used for small, stable ranges (e.g. safetensors headers) whose boundaries\n * don't depend on which tensors a particular load needs. Tensor *bodies* are\n * NOT fetched through here — they're cached per-tensor by stable byte-range in\n * fetchSelectiveTensors so a vision load can reuse a text load's tensors.\n */\nasync function fetchRange(\n  url: string,\n  start: number,\n  end: number,\n  hfToken?: string,\n): Promise<ArrayBuffer> {\n  // Check browser cache for this range\n  const browserKey = `${url}?_range=${start}-${end}`;\n  const browserCached = await browserCacheRead(browserKey);\n  if (browserCached) return browserCached;\n\n  const buf = await fetchRangeRaw(url, start, end, hfToken);\n\n  // Cache for next time\n  await browserCacheWrite(browserKey, buf);\n\n  return buf;\n}\n\n/**\n * Fetch the safetensors header only (two range requests: 8 bytes + header).\n */\nasync function fetchSafetensorsHeader(url: string, hfToken?: string): Promise<SafetensorsFile> {\n  // Fetch first 8 bytes to get header length\n  const lengthBuf = await fetchRange(url, 0, 8, hfToken);\n  const headerLength = Number(new DataView(lengthBuf).getBigUint64(0, true));\n\n  // Fetch full header (8 + headerLength bytes)\n  const headerBuf = await fetchRange(url, 0, 8 + headerLength, hfToken);\n  return parseSafetensorsHeader(headerBuf);\n}\n\n/** A contiguous byte range to download. */\ninterface ByteRange {\n  start: number;\n  end: number;\n  entries: SafetensorEntry[];\n}\n\n/**\n * Merge close byte ranges to reduce HTTP requests.\n *\n * - `gapThreshold`: merge two ranges if the gap between them is smaller than this\n *   (avoids an extra request for a tiny skipped gap).\n * - `maxRangeBytes`: hard cap on a single merged range. Without this, a\n *   contiguously-packed safetensors merges into ONE range = the whole model, so\n *   the loader buffers the entire file in a single fetch — no streaming progress\n *   (the bar sits at 0/N until done) and a transient allocation big enough to\n *   crash mobile Safari. Capping splits the download into bounded segments that\n *   report progress as they land and keep peak memory low.\n */\nfunction mergeRanges(\n  entries: SafetensorEntry[],\n  dataStart: number,\n  gapThreshold: number = 1024 * 1024,\n  maxRangeBytes: number = Number.POSITIVE_INFINITY,\n): ByteRange[] {\n  if (entries.length === 0) return [];\n\n  // Sort by offset\n  const sorted = [...entries].sort((a, b) => a.dataOffset - b.dataOffset);\n\n  const ranges: ByteRange[] = [];\n  let current: ByteRange = {\n    start: dataStart + sorted[0].dataOffset,\n    end: dataStart + sorted[0].dataOffset + sorted[0].dataLength,\n    entries: [sorted[0]],\n  };\n\n  for (let i = 1; i < sorted.length; i++) {\n    const entryStart = dataStart + sorted[i].dataOffset;\n    const entryEnd = entryStart + sorted[i].dataLength;\n\n    const withinGap = entryStart - current.end <= gapThreshold;\n    const withinCap = entryEnd - current.start <= maxRangeBytes;\n    if (withinGap && withinCap) {\n      // Merge: extend the range (may include some skipped bytes, but avoids extra request)\n      current.end = Math.max(current.end, entryEnd);\n      current.entries.push(sorted[i]);\n    } else {\n      ranges.push(current);\n      current = { start: entryStart, end: entryEnd, entries: [sorted[i]] };\n    }\n  }\n  ranges.push(current);\n  return ranges;\n}\n\n/**\n * Selectively download needed tensors from a safetensors file using Range requests.\n * Returns a map of tensor name → { data, shape }.\n */\n/** Byte alignment required for a safetensors dtype's typed-array view. */\nfunction dtypeAlign(dtype: string): number {\n  if (dtype === \"F64\" || dtype === \"I64\" || dtype === \"U64\") return 8;\n  if (dtype === \"F32\" || dtype === \"I32\" || dtype === \"U32\") return 4;\n  if (dtype === \"F16\" || dtype === \"BF16\" || dtype === \"I16\" || dtype === \"U16\") return 2;\n  return 1;\n}\n\n/** Build the appropriate typed-array view of a tensor at `byteOffset` in `buf`. */\nfunction makeTensorView(\n  buf: ArrayBuffer,\n  byteOffset: number,\n  entry: SafetensorEntry,\n): ArrayBufferView {\n  // Aligned views can alias the buffer directly; unaligned ones must be copied\n  // into a fresh (offset-0) buffer so the typed-array constructor accepts them.\n  const aligned = byteOffset % dtypeAlign(entry.dtype) === 0;\n  const src = aligned ? buf : buf.slice(byteOffset, byteOffset + entry.dataLength);\n  const off = aligned ? byteOffset : 0;\n  switch (entry.dtype) {\n    case \"F32\":\n      return new Float32Array(src, off, entry.dataLength / 4);\n    case \"I32\":\n      return new Int32Array(src, off, entry.dataLength / 4);\n    case \"U32\":\n      return new Uint32Array(src, off, entry.dataLength / 4);\n    case \"F16\":\n    case \"BF16\":\n      return new Uint16Array(src, off, entry.dataLength / 2);\n    default:\n      return new Uint8Array(src, off, entry.dataLength);\n  }\n}\n\n/**\n * Stable per-tensor browser-cache key. Keyed by the tensor's *absolute* byte\n * range in the file (offset + length), which is fixed regardless of which other\n * tensors a given load requests. This is what lets a vision load reuse the\n * tensors a prior text load already downloaded — they hit the same keys —\n * instead of re-downloading because the merged Range boundaries differ.\n */\nfunction tensorCacheKey(url: string, file: SafetensorsFile, entry: SafetensorEntry): string {\n  const absStart = file.dataStart + entry.dataOffset;\n  return `${url}#t:${absStart}-${entry.dataLength}`;\n}\n\n/**\n * Selectively download needed tensors from a safetensors file using Range\n * requests.\n *\n * Two modes, gated by `cacheStore`:\n *\n *  • `cacheStore` supplied (browser): each tensor's bytes are written to the\n *    per-tensor download cache (as before) and the tensor is REGISTERED in the\n *    store as a cache-backed descriptor under its canonical name — its bytes are\n *    NOT retained in heap. The returned Map is empty. This is the change that\n *    bounds peak heap during download: instead of accumulating the whole model\n *    in the returned Map, the heap only ever holds the current range segment\n *    (≤ maxRangeBytes, 24 MB on mobile) which is released each iteration.\n *\n *  • `cacheStore` omitted (used by direct-load helpers): returns a Map of\n *    name → { data, shape } exactly as before.\n */\nasync function fetchSelectiveTensors(\n  url: string,\n  file: SafetensorsFile,\n  neededEntries: SafetensorEntry[],\n  hfToken?: string,\n  onProgress?: (loaded: number, total: number, message?: string) => void,\n  cacheStore?: { store: CacheWeightStore; keyMapper: HFKeyMapper },\n): Promise<Map<string, { data: ArrayBufferView; shape: number[] }>> {\n  const results = new Map<string, { data: ArrayBufferView; shape: number[] }>();\n  const totalBytes = neededEntries.reduce((sum, e) => sum + e.dataLength, 0);\n  let loadedBytes = 0;\n\n  const isMobile =\n    typeof navigator !== \"undefined\" && /iPhone|iPad|iPod|Android/i.test(navigator.userAgent);\n  // Cap merged-range size so the download streams in bounded segments: keeps\n  // peak transient memory low (critical on iOS) and lets the progress bar move.\n  const maxRangeBytes = isMobile ? 24 * 1024 * 1024 : 256 * 1024 * 1024;\n\n  // Register a downloaded tensor: either retain it in the result Map (Node-ish\n  // direct loads) or — in browser cache mode — record only a cache-backed\n  // descriptor under the canonical name so no bytes stay in heap.\n  const record = (entry: SafetensorEntry, key: string): void => {\n    if (cacheStore) {\n      const canonical = cacheStore.keyMapper(entry.name);\n      if (canonical) {\n        // Bytes live in the per-tensor download cache (BROWSER_CACHE_NAME).\n        cacheStore.store.registerCached(\n          canonical,\n          entry.shape,\n          entry.dtype,\n          key,\n          BROWSER_CACHE_NAME,\n        );\n      }\n      return;\n    }\n    // Fallback (no cache backend): caller still wants bytes in the Map.\n    // (This branch is only reachable when CacheStorage is unavailable.)\n  };\n\n  // Reclaim quota from the dead legacy CacheStorage/OPFS caches before we\n  // probe/write IDB — runs on the default (heap) browser path too, since that's\n  // where per-tensor IDB caching now happens.\n  // Fire-and-forget: legacy-cache cleanup must NEVER block (or hang) the load —\n  // OPFS dir iteration can stall on Safari, and it's only reclaiming dead junk.\n  if (TENSOR_CACHE_ENABLED) void evictStaleCaches();\n\n  // Pass 1: serve any tensor already in the per-tensor browser cache. A tensor\n  // cached by an earlier load (e.g. text-only) is reused here even if this load\n  // (e.g. enableVision) needs a different overall set — the key is per-tensor.\n  // Probe in parallel: hundreds of sequential CacheStorage.match() calls are slow\n  // on mobile and would make a fresh load look frozen before any download starts.\n  // In cache mode we only need to know WHICH tensors are present (HEAD-style), not\n  // their bytes — so probe with `caches.match` but discard bodies immediately.\n  const uncached: SafetensorEntry[] = [];\n  // Probe presence CHEAPLY (getKey, no bodies) in parallel. Reading the actual\n  // bytes for hundreds of tensors at once (the old heap-path Promise.all over\n  // browserCacheRead) stormed IndexedDB with ~696 concurrent transactions AND\n  // pulled ~404 MB into heap in one burst → Safari hung on a warm cache (the lean\n  // harness only ever hit a cold cache, so it never showed). Presence is cheap and\n  // parallel-safe; the heap path then reads the HIT bytes one at a time below.\n  // ONE getAllKeys() transaction, then membership in memory — NOT hundreds of\n  // parallel getKey transactions (that storm hangs Safari on a warm cache).\n  const cachedKeys = TENSOR_CACHE_ENABLED ? await idbAllKeys() : new Set<string>();\n  const presence = neededEntries.map((entry) => cachedKeys.has(tensorCacheKey(url, file, entry)));\n  for (let i = 0; i < neededEntries.length; i++) {\n    const entry = neededEntries[i];\n    const key = tensorCacheKey(url, file, entry);\n    if (!presence[i]) {\n      uncached.push(entry);\n      continue;\n    }\n    if (cacheStore) {\n      // Presence-only (the store reads bytes lazily later).\n      record(entry, key);\n      loadedBytes += entry.dataLength;\n      onProgress?.(loadedBytes, totalBytes);\n      continue;\n    }\n    // Heap path: read the cached bytes ONE AT A TIME (never all in parallel — that\n    // storms IDB and OOMs Safari). VALIDATE size; a partial/corrupt entry is\n    // treated as a miss so pass 2 re-downloads and overwrites it (self-healing).\n    const buf = await browserCacheRead(key);\n    if (buf && buf.byteLength >= entry.dataLength) {\n      results.set(entry.name, {\n        data: makeTensorView(buf, 0, entry),\n        shape: entry.shape,\n      });\n      loadedBytes += entry.dataLength;\n      onProgress?.(loadedBytes, totalBytes);\n    } else {\n      uncached.push(entry);\n    }\n  }\n\n  // Cache visibility: report how many tensors were reused vs must be downloaded,\n  // and which backend served them — so a silent \"redownloads everything\" cache\n  // miss is diagnosable on-device (most users have no Safari console). Surfaces\n  // through onProgress, which the iPad runner streams back to the dashboard.\n  if (TENSOR_CACHE_ENABLED) {\n    const hits = neededEntries.length - uncached.length;\n    const backend = idbCacheAvailable() ? \"idb\" : \"none\";\n    const msg = `cache[${backend}]: ${hits}/${neededEntries.length} tensors reused, ${uncached.length} to download`;\n    console.log(`[gerbil] ${msg}`);\n    onProgress?.(loadedBytes, totalBytes, msg);\n  }\n\n  // Pass 2: download only the uncached tensors (merged into bounded contiguous\n  // ranges), then cache each one individually by its stable key.\n  const ranges = mergeRanges(uncached, file.dataStart, 1024 * 1024, maxRangeBytes);\n  for (const range of ranges) {\n    const rangeBuf = await fetchRangeRaw(url, range.start, range.end, hfToken);\n    // A truncated range response (e.g. an aborted/partial download) would yield\n    // short tensor slices — bad views now and a poisoned cache later. Reject it\n    // loudly instead of silently caching garbage.\n    const expectedRangeBytes = range.end - range.start;\n    if (rangeBuf.byteLength < expectedRangeBytes) {\n      throw new Error(\n        `Incomplete weight download: got ${rangeBuf.byteLength} of ${expectedRangeBytes} bytes for ${url}. Retry.`,\n      );\n    }\n\n    for (const entry of range.entries) {\n      const localOffset = file.dataStart + entry.dataOffset - range.start;\n      const key = tensorCacheKey(url, file, entry);\n      // Cache the exact tensor bytes (a standalone copy) for cross-load reuse.\n      const tensorBytes = rangeBuf.slice(localOffset, localOffset + entry.dataLength);\n      // Only cache a full-size slice so we never write a short (poisoning) entry.\n      if (TENSOR_CACHE_ENABLED && tensorBytes.byteLength === entry.dataLength) {\n        // `tensorBytes` is a standalone slice not reused below (the heap path\n        // views `rangeBuf`, a different buffer), so it can be transferred to the\n        // OPFS worker zero-copy — this is the big-memory path, so avoiding a copy\n        // matters most here.\n        await browserCacheWrite(key, tensorBytes, true);\n      }\n      if (cacheStore) {\n        // Bytes are now durably in cache; do NOT retain them in heap.\n        record(entry, key);\n      } else {\n        results.set(entry.name, {\n          data: makeTensorView(rangeBuf, localOffset, entry),\n          shape: entry.shape,\n        });\n      }\n      loadedBytes += entry.dataLength;\n      onProgress?.(loadedBytes, totalBytes);\n    }\n    // `rangeBuf` (and its slices for the non-cache path) go out of scope here;\n    // in cache mode nothing references it, so the segment is freed before the\n    // next range is fetched — peak heap stays ≈ one range segment.\n  }\n\n  return results;\n}\n\n/**\n * Create the appropriate HF key mapper for a given architecture.\n *\n * Qwen3.5 conditional generation models use \"model.language_model.\" prefix for\n * text model weights and include visual encoder + MTP weights that we skip.\n * Standard models just use \"model.\" prefix.\n *\n * `wantVision` controls whether the vision tower is kept. When false (text-only\n * loads) the ViT is dropped from the selective download so the ~201MB tower is\n * never fetched. When true (enableVision) the tower is mapped to the canonical\n * \"visual.*\" keys the vision graph expects, regardless of whether the source\n * checkpoint names it \"model.visual.*\" (HF BF16) or \"vision_tower.*\" (MLX).\n */\nfunction createKeyMapperForArch(architectureName: string, wantVision = false): HFKeyMapper {\n  // KaniTTS2 ships the LFM2 backbone verbatim — the only structural rename it needs\n  // is the same self_attn.out_proj → o_proj as LFM2. Its extra tensors\n  // (learnable_rope_layers.*, speaker_emb_projection) pass through after stripping\n  // the \"model.\" prefix and are picked up by the codec-LM graph when it lands.\n  if (architectureName === \"Lfm2ForCausalLM\" || architectureName === \"KaniTTS2ForCausalLM\") {\n    // LFM2 weights live under \"model.\" like the others, but the attention\n    // output projection is named \"out_proj\" where the canonical IR (shared with\n    // Qwen) expects \"o_proj\". Rename it so the o_proj graph tensors resolve.\n    // All other LFM2 keys (operator_norm, ffn_norm, embedding_norm, conv.*,\n    // feed_forward.w1/w2/w3, q_layernorm/k_layernorm) are referenced verbatim\n    // by the generator via per-tensor safetensorsKey, so they pass through.\n    return (hfKey: string): string | null => {\n      let key = hfKey;\n      if (key.startsWith(\"model.\")) {\n        key = key.slice(6);\n      }\n      key = key.replace(/\\.self_attn\\.out_proj\\./, \".self_attn.o_proj.\");\n      return key;\n    };\n  }\n  if (architectureName === \"Qwen3_5ForConditionalGeneration\") {\n    return (hfKey: string): string | null => {\n      // Vision tower lives under one of two prefixes depending on the converter:\n      //   • \"model.visual.*\"  — HF BF16 checkpoint (Qwen/Qwen3.5-0.8B)\n      //   • \"vision_tower.*\"   — mlx-vlm 4-bit convert (mlx-community/…-4bit).\n      //     In MLX-4bit repos the text tower is quantized (U32 + scales/biases)\n      //     but the ViT is left BF16, so the suffix structure is identical to\n      //     the HF tower — only the prefix differs. Both map to the canonical\n      //     \"visual.*\" keys the vision graph generator expects.\n      // When vision is not requested we drop the tower so the selective\n      // downloader never fetches its ~201MB of weights.\n      if (hfKey.startsWith(\"model.visual.\")) {\n        return wantVision ? hfKey.slice(6) : null; // \"model.visual.*\" → \"visual.*\"\n      }\n      if (hfKey.startsWith(\"vision_tower.\")) {\n        // \"vision_tower.\".length === 13 → replace with \"visual.\"\n        return wantVision ? `visual.${hfKey.slice(13)}` : null;\n      }\n      // Always skip the multi-token-prediction head (an inference-speed\n      // optimization, not a modality).\n      if (hfKey.startsWith(\"mtp.\")) {\n        return null;\n      }\n      let key = hfKey;\n      // HF/GPTQ uses \"model.language_model.\" prefix, MLX uses \"language_model.model.\"\n      if (key.startsWith(\"model.language_model.\")) {\n        key = key.slice(21); // \"model.language_model.\".length\n      } else if (key.startsWith(\"language_model.model.\")) {\n        key = key.slice(21); // \"language_model.model.\".length (also 21 chars)\n      } else if (key.startsWith(\"model.\")) {\n        key = key.slice(6); // \"model.\".length\n      }\n      return key;\n    };\n  }\n  if (architectureName === \"Gemma4ForConditionalGeneration\") {\n    // Gemma 4 nests the text tower under \"language_model.\" and carries vision +\n    // audio towers we drop for text-only loads. Prefix shapes observed:\n    //   HF BF16:  \"model.language_model.*\"  (+ \"model.vision_tower.*\", \"model.audio_tower.*\",\n    //             \"model.embed_vision.*\", \"model.embed_audio.*\")\n    //   MLX-4bit: \"language_model.model.*\"  (+ \"vision_tower.*\", \"audio_tower.*\")\n    // The text tower keys after stripping match the canonical Gemma layout used\n    // by gemma4.ts (embed_tokens, embed_tokens_per_layer, per_layer_*, layers.N.*,\n    // norm). PLE/projection tensors pass through verbatim.\n    return (hfKey: string): string | null => {\n      // Vision tower: keep + canonicalize when wantVision; otherwise drop so the\n      // selective downloader never fetches it. Both prefixes (\"model.vision_tower.*\"\n      // HF BF16, \"vision_tower.*\" MLX) and the multimodal projector\n      // (\"model.embed_vision.*\" / \"embed_vision.*\") map to the canonical\n      // \"vision_tower.*\" / \"embed_vision.*\" keys gemma4_vision.ts expects.\n      if (hfKey.includes(\"vision_tower.\")) {\n        if (!wantVision) return null;\n        const idx = hfKey.indexOf(\"vision_tower.\");\n        return hfKey.slice(idx); // strip any \"model.\" prefix before vision_tower.\n      }\n      if (hfKey.includes(\"embed_vision.\")) {\n        if (!wantVision) return null;\n        const idx = hfKey.indexOf(\"embed_vision.\");\n        return hfKey.slice(idx);\n      }\n      // Always drop the audio modality (not supported in the vision path).\n      if (hfKey.includes(\"audio_tower.\") || hfKey.includes(\"embed_audio.\")) {\n        return null;\n      }\n      let key = hfKey;\n      if (key.startsWith(\"model.language_model.\")) {\n        key = key.slice(21); // \"model.language_model.\".length\n      } else if (key.startsWith(\"language_model.model.\")) {\n        key = key.slice(21); // \"language_model.model.\".length (also 21)\n      } else if (key.startsWith(\"language_model.\")) {\n        key = key.slice(15); // \"language_model.\".length (lm_head etc.)\n      } else if (key.startsWith(\"model.\")) {\n        key = key.slice(6);\n      }\n      return key;\n    };\n  }\n  if (architectureName === \"MoonshineForConditionalGeneration\") {\n    // Moonshine's HF keys already match the canonical Moonshine graph names once\n    // the \"model.\" prefix is stripped, e.g.\n    //   model.encoder.conv1.weight                       → encoder.conv1.weight\n    //   model.encoder.layers.0.self_attn.q_proj.weight   → encoder.layers.0.self_attn.q_proj.weight\n    //   model.decoder.layers.0.encoder_attn.q_proj.weight→ decoder.layers.0.encoder_attn.q_proj.weight\n    //   model.decoder.embed_tokens.weight                → decoder.embed_tokens.weight (tied lm_head)\n    // The default \"strip model.\" mapper is therefore exactly correct. (Moonshine\n    // weight loading runs through the dual-graph STT loader — remaining work —\n    // not the single-graph generate() path, which throws by design.)\n    return createDefaultHFKeyMapper();\n  }\n  return createDefaultHFKeyMapper();\n}\n\nconst PLE_CACHE_NAME = \"gerbil-ple-v2\";\n\n/**\n * Build a `PleSource` from the quantized PLE table.\n *\n * Node/desktop (`cacheStore` null): keep the arrays heap-resident — unchanged\n * behavior, the executor streams rows directly from them.\n *\n * Browser (`cacheStore` set): write the (already-compact INT4) packed/scales/\n * zeros to CacheStorage and return EMPTY heap arrays plus a `cache` descriptor.\n * This keeps the ~1.17 GB table out of the JS heap during the load/upload peak;\n * the executor lazily materializes it from cache after the GPU upload has\n * completed (past the memory high-water mark). See Executor.setPleSource.\n */\nasync function buildPleSource(\n  packed: Uint32Array,\n  scales: Float32Array,\n  zeros: Float32Array,\n  width: number,\n  groupSize: number,\n  cacheStore: CacheWeightStore | null,\n): Promise<PleSource> {\n  if (!cacheStore) {\n    return { packed, scales, zeros, width, groupSize, targetTensor: \"ple_embed_out\" };\n  }\n  const cache = await caches.open(PLE_CACHE_NAME);\n  const packedKey = \"ple:packed\";\n  const scalesKey = \"ple:scales\";\n  const zerosKey = \"ple:zeros\";\n  const copy = (a: ArrayBufferView): ArrayBuffer =>\n    a.buffer.slice(a.byteOffset, a.byteOffset + a.byteLength) as ArrayBuffer;\n  await cache.put(new Request(packedKey), new Response(copy(packed)));\n  await cache.put(new Request(scalesKey), new Response(copy(scales)));\n  await cache.put(new Request(zerosKey), new Response(copy(zeros)));\n  return {\n    // Empty heap arrays — the bytes live in cache now.\n    packed: new Uint32Array(0),\n    scales: new Float32Array(0),\n    zeros: new Float32Array(0),\n    width,\n    groupSize,\n    targetTensor: \"ple_embed_out\",\n    cache: {\n      cacheName: PLE_CACHE_NAME,\n      packedKey,\n      scalesKey,\n      zerosKey,\n      packedLen: packed.length,\n    },\n  };\n}\n\n/**\n * Load a model from HuggingFace Hub.\n *\n * 1. Fetch config.json -> determine architecture -> generate IR graph\n * 2. Fetch tokenizer.json + tokenizer_config.json -> build tokenizer\n * 3. Download safetensors -> parse headers -> extract weight data\n * 4. Map HF tensor keys -> canonical names\n */\nexport async function loadModel(options: LoadModelOptions): Promise<LoadedModel> {\n  const { repo, onProgress, hfToken, revision = \"main\", cacheDir, dtype } = options;\n\n  // Initialize Node.js fs for caching (no-op in browser)\n  await initFs();\n\n  // Resolve cache directory: use provided path or default to ~/.cache/gerbil/<repo>\n  let resolvedCacheDir = cacheDir;\n  if (!resolvedCacheDir && _fs && _path) {\n    const home = process.env.HOME || process.env.USERPROFILE || \"\";\n    if (home) {\n      resolvedCacheDir = _path.join(home, \".cache\", \"gerbil\", repo.replace(/\\//g, \"_\"), revision);\n    }\n  }\n\n  const baseURL = resolveHFBaseURL(repo, revision);\n\n  // -- Step 1: Fetch config.json --\n  onProgress?.(0, 100, \"Fetching model config...\");\n  const rawConfig = await fetchJSON(baseURL, \"config.json\", hfToken, resolvedCacheDir);\n\n  const architectureName = rawConfig.architectures?.[0];\n  if (!architectureName) {\n    throw new Error(\"config.json missing 'architectures' field\");\n  }\n\n  // Detect pre-quantized formats — force q4 graph when loading quantized models.\n  // MLX-lm writes the quant block under EITHER \"quantization_config\" (newer\n  // converts, often with mode:\"affine\") OR \"quantization\" (standard mlx-lm,\n  // typically just {bits, group_size} with no mode field). Read whichever exists.\n  const quantConfig = (rawConfig.quantization_config ?? rawConfig.quantization) as\n    | Record<string, unknown>\n    | undefined;\n  const isGPTQ = quantConfig?.quant_method === \"gptq\";\n\n  // MLX 4-bit detection. Two shapes both indicate a standard affine MLX convert:\n  //   1. {bits:4, mode:\"affine\", ...}                (newer converts)\n  //   2. {bits:4, group_size:N}  (no mode field)     (standard mlx-lm)\n  // CAUTION: DWQ converts are config-INDISTINGUISHABLE from (2) — both write only\n  // {bits:4, group_size}. The repack treats the weights as affine MLX, which\n  // silently produces garbage for DWQ. So a no-mode MLX config is only accepted\n  // for repos we've verified are standard (non-DWQ) affine converts. A verified\n  // load must still be sanity-checked at runtime (cosine reference) — see the\n  // embedding validation scripts.\n  const quantMode = quantConfig?.mode as string | undefined;\n  const hasMlxShape =\n    !isGPTQ &&\n    quantConfig?.bits === 4 &&\n    (quantMode === \"affine\" ||\n      (quantMode === undefined && typeof quantConfig?.group_size === \"number\"));\n\n  // Allowlist of repos verified to be standard (non-DWQ) MLX-4bit converts.\n  // Match is substring-insensitive so revision-pinned / mirrored repo strings\n  // still resolve. DWQ repos (…-4bit-DWQ) must NOT appear here.\n  const VERIFIED_MLX_REPOS = [\"mlx-community/embeddinggemma-300m-4bit\"];\n  const repoLower = repo.toLowerCase();\n  const isDWQ = repoLower.includes(\"dwq\");\n  const isVerifiedMlxRepo =\n    !isDWQ && VERIFIED_MLX_REPOS.some((r) => repoLower.includes(r.toLowerCase()));\n\n  // Accept MLX when the config explicitly says affine, OR when the no-mode shape\n  // is present AND the repo is on the verified allowlist (DWQ-trap guard).\n  const isMLX = hasMlxShape && (quantMode === \"affine\" || isVerifiedMlxRepo);\n  // Resolve \"auto\": on mobile (iOS/Android) prefer on-the-fly q4 so large f16\n  // checkpoints (e.g. LFM2.5-350M ~676 MB f16 → ~190 MB q4) fit in device memory;\n  // on desktop keep the repo's native precision. Already-quantized repos (MLX/\n  // GPTQ) are always q4.\n  const isMobileUA =\n    typeof navigator !== \"undefined\" && /iPhone|iPad|iPod|Android/i.test(navigator.userAgent);\n  const requestedDtype: GraphDType | undefined =\n    dtype === \"auto\" ? (isMobileUA ? \"q4\" : undefined) : dtype;\n  const resolvedDtype: GraphDType | undefined = isGPTQ || isMLX ? \"q4\" : requestedDtype;\n\n  // MLX uses a different group_size (typically 64 vs Gerbil default 128)\n  const quantGroupSize = isMLX ? (quantConfig?.group_size as number) : undefined;\n\n  if (isGPTQ) {\n    onProgress?.(2, 100, \"Detected GPTQ model — will repack to native INT4...\");\n  } else if (isMLX) {\n    onProgress?.(\n      2,\n      100,\n      `Detected MLX 4-bit model (group_size=${quantGroupSize}) — will repack...`,\n    );\n  }\n\n  // Select key mapper: Qwen3.5 conditional generation models have weights\n  // under \"model.language_model.\" prefix instead of the standard \"model.\" prefix.\n  // Vision is wanted when the caller requested the multimodal graph variant\n  // (enableVision) or explicitly asked for the tower (loadVisionTower — used by\n  // the direct-load vision validation scripts). Gate the ViT key-mapping on this\n  // so text-only loads stay lean (the ~201MB tower is excluded from download).\n  const wantVision = Boolean(options.multimodal || options.loadVisionTower);\n  const keyMapper = options.keyMapper ?? createKeyMapperForArch(architectureName, wantVision);\n\n  // Generate the computation graph (with INT4 ops if dtype is \"q4\" or GPTQ/MLX detected)\n  const graph = generateGraph(\n    architectureName,\n    rawConfig,\n    resolvedDtype,\n    quantGroupSize,\n    options.kvDtype,\n    options.embedding,\n    options.multimodal,\n  );\n\n  // -- Step 2: Fetch tokenizer --\n  onProgress?.(5, 100, \"Fetching tokenizer...\");\n  const [tokenizerJSON, tokenizerConfigJSON, chatTemplateJinja] = await Promise.all([\n    fetchJSON(baseURL, \"tokenizer.json\", hfToken, resolvedCacheDir),\n    fetchJSON(baseURL, \"tokenizer_config.json\", hfToken, resolvedCacheDir).catch(() => null),\n    // Some models (e.g. LFM2) ship the chat template as a sidecar `.jinja` file\n    // instead of inline in tokenizer_config.json. Fetch it as a fallback signal.\n    fetchText(baseURL, \"chat_template.jinja\", hfToken, resolvedCacheDir).catch(() => null),\n  ]);\n\n  // Prefer the inline template; otherwise fall back to the sidecar .jinja.\n  const configWithTemplate =\n    tokenizerConfigJSON && !tokenizerConfigJSON.chat_template && chatTemplateJinja\n      ? { ...tokenizerConfigJSON, chat_template: chatTemplateJinja }\n      : (tokenizerConfigJSON ?? (chatTemplateJinja ? { chat_template: chatTemplateJinja } : null));\n\n  const tokenizer = Tokenizer.fromJSON(tokenizerJSON, configWithTemplate);\n\n  // -- Step 3: Download safetensors --\n  onProgress?.(10, 100, \"Discovering weight files...\");\n  const safetensorsFiles = await discoverSafetensorsFiles(baseURL, hfToken, resolvedCacheDir);\n\n  // Cache-backed streaming store keeps the browser heap bounded (tensor bytes\n  // live in CacheStorage; the heap holds only descriptors) so very large models\n  // load on mobile without OOM. It is the DEFAULT on mobile — the heap-backed\n  // path keeps every tensor resident in the JS heap on top of the GPU buffers\n  // (and expands f16/bf16 -> f32 in-heap), doubling peak memory and OOM-crashing\n  // iOS WebKit right after the safetensors header read. Desktop/Node keep the\n  // proven heap path. Opt in elsewhere with `?stream=1` /\n  // `globalThis.GERBIL_STREAM_WEIGHTS = true`; force the heap path anywhere with\n  // `globalThis.GERBIL_STREAM_WEIGHTS = false`.\n  const streamFlag = (globalThis as { GERBIL_STREAM_WEIGHTS?: boolean }).GERBIL_STREAM_WEIGHTS;\n  const streamOptIn =\n    streamFlag === true ||\n    (typeof location !== \"undefined\" &&\n      typeof URLSearchParams !== \"undefined\" &&\n      new URLSearchParams(location.search).has(\"stream\"));\n  const useCacheBacked =\n    streamFlag !== false && (streamOptIn || isMobileUA) && cacheStorageAvailable();\n  const cacheStore = useCacheBacked ? new CacheWeightStore() : null;\n  const store: MutableWeightStore = cacheStore ?? new HeapWeightStore();\n\n  let filesLoaded = 0;\n  const totalFiles = safetensorsFiles.length;\n\n  for (const filename of safetensorsFiles) {\n    const fileUrl = `${baseURL}/${filename}`;\n\n    // Check cache first — if cached, use the old full-buffer path\n    let cachedBuffer: ArrayBuffer | null = null;\n    if (resolvedCacheDir) {\n      const cached = readCachedFile(resolvedCacheDir, filename);\n      if (cached) {\n        cachedBuffer = cached.buffer.slice(\n          cached.byteOffset,\n          cached.byteOffset + cached.byteLength,\n        ) as ArrayBuffer;\n      }\n    }\n\n    if (cachedBuffer) {\n      // Cached (Node disk cache) — extract all needed tensors from full buffer.\n      const file = parseSafetensorsHeader(cachedBuffer);\n      for (const entry of file.entries) {\n        const canonicalName = keyMapper(entry.name);\n        if (!canonicalName) continue;\n        let data: ArrayBufferView = getTensorData(cachedBuffer, file, entry);\n        if (entry.dtype === \"BF16\") {\n          data = bf16ToF32(new Uint8Array(data.buffer, data.byteOffset, data.byteLength));\n        }\n        if (entry.dtype === \"F16\") {\n          data = f16ToF32(data);\n        }\n        await store.set(canonicalName, { data, shape: entry.shape });\n      }\n    } else {\n      // Not cached — use selective Range-request downloading\n      const fileWeight = 85 / totalFiles;\n      const base = 10 + filesLoaded * fileWeight;\n\n      // Step 1: Fetch header only. This + the index probe above can take a few\n      // seconds with no bytes yet — without this emit the bar looks frozen at 10%\n      // (\"stuck at discovering weight files\") until the first chunk lands.\n      onProgress?.(base, 100, `Reading ${filename} header (${filesLoaded + 1}/${totalFiles})...`);\n      const file = await fetchSafetensorsHeader(fileUrl, hfToken);\n\n      // Step 2: Filter to needed entries\n      const neededEntries = file.entries.filter((e) => keyMapper(e.name) !== null);\n      const skippedBytes = file.entries\n        .filter((e) => keyMapper(e.name) === null)\n        .reduce((sum, e) => sum + e.dataLength, 0);\n      const neededBytes = neededEntries.reduce((sum, e) => sum + e.dataLength, 0);\n\n      const totalMB = (neededBytes / 1048576).toFixed(0);\n      if (skippedBytes > 0) {\n        const savedMB = (skippedBytes / 1048576).toFixed(0);\n        onProgress?.(\n          base,\n          100,\n          `Selective download: ${totalMB} MB needed, skipping ${savedMB} MB (vision/MTP)`,\n        );\n      }\n      // Show the size up front so the first (latency-heavy) chunk doesn't read as a freeze.\n      onProgress?.(base, 100, `Downloading ${filename} (0/${totalMB} MB)`);\n\n      // Step 3: Download needed tensors via Range requests\n      let lastReportedPct = -1;\n\n      const rawTensors = await fetchSelectiveTensors(\n        fileUrl,\n        file,\n        neededEntries,\n        hfToken,\n        (loaded, total) => {\n          const filePct = total > 0 ? (loaded / total) * fileWeight : 0;\n          const overallPct = Math.round(base + filePct);\n          if (overallPct <= lastReportedPct) return;\n          lastReportedPct = overallPct;\n          const sizeMB =\n            total > 0\n              ? ` (${(loaded / 1048576).toFixed(0)}/${(total / 1048576).toFixed(0)} MB)`\n              : \"\";\n          onProgress?.(base + filePct, 100, `Downloading ${filename}${sizeMB}`);\n        },\n        // Browser: register tensors as cache-backed descriptors under their\n        // canonical names — no bytes retained in heap. Node: omitted, so the\n        // returned Map carries the bytes (handled below).\n        cacheStore ? { store: cacheStore, keyMapper } : undefined,\n      );\n\n      // Step 4 (Node / non-cache only): map to canonical names with dtype\n      // conversion. In cache-backed mode `rawTensors` is empty — tensors were\n      // already registered into the store with the canonical name + source dtype,\n      // and `store.get()` applies BF16/F16→F32 on demand.\n      for (const [hfName, { data: rawData, shape }] of rawTensors) {\n        const canonicalName = keyMapper(hfName);\n        if (!canonicalName) continue;\n\n        let data: ArrayBufferView = rawData;\n        // Find the entry to check dtype\n        const entry = neededEntries.find((e) => e.name === hfName);\n        if (entry) {\n          if (entry.dtype === \"BF16\") {\n            data = bf16ToF32(new Uint8Array(data.buffer, data.byteOffset, data.byteLength));\n          }\n          if (entry.dtype === \"F16\") {\n            data = f16ToF32(data);\n          }\n        }\n\n        await store.set(canonicalName, { data, shape });\n      }\n    }\n\n    filesLoaded++;\n  }\n\n  onProgress?.(95, 100, \"Weights loaded.\");\n\n  // -- Split fused Q projection for Qwen3.5 full attention layers --\n  // HF stores q_proj.weight as [num_heads * 2 * head_dim, hidden_size].\n  // The layout is per-head interleaved: for each head h, rows\n  //   [h*2*head_dim .. h*2*head_dim + head_dim) are Q weights, and\n  //   [h*2*head_dim + head_dim .. (h+1)*2*head_dim) are gate weights.\n  // We deinterleave into q_proj.weight [num_heads*head_dim, hidden_size]\n  // and attn_gate.weight [num_heads*head_dim, hidden_size].\n  if (architectureName === \"Qwen3_5ForConditionalGeneration\") {\n    const textCfg = (rawConfig.text_config as Record<string, unknown>) ?? rawConfig;\n    const numLayers = textCfg.num_hidden_layers as number;\n    const layerTypesArr = (textCfg.layer_types as string[]) ?? [];\n    const fullAttnInterval = (textCfg.full_attention_interval as number) ?? 4;\n    const numHeads = textCfg.num_attention_heads as number; // 8\n    const headDim =\n      (textCfg.head_dim as number) ?? Math.floor((textCfg.hidden_size as number) / numHeads);\n\n    for (let i = 0; i < numLayers; i++) {\n      // Check if this is a full attention layer\n      let isFullAttn: boolean;\n      if (layerTypesArr.length > i) {\n        isFullAttn = layerTypesArr[i] === \"full_attention\";\n      } else {\n        isFullAttn = i % fullAttnInterval === fullAttnInterval - 1;\n      }\n      if (!isFullAttn) continue;\n\n      const qProjKey = `layers.${i}.self_attn.q_proj.weight`;\n      const attnGateKey = `layers.${i}.self_attn.attn_gate.weight`;\n      const fusedWeight = await store.get(qProjKey);\n\n      if (fusedWeight && !store.has(attnGateKey) && !isMLX) {\n        const fusedData = fusedWeight.data;\n        const [fusedRows, cols] = fusedWeight.shape;\n        const halfRows = fusedRows / 2; // num_heads * head_dim\n\n        const qData = new Float32Array(halfRows * cols);\n        const gateData = new Float32Array(halfRows * cols);\n\n        const src =\n          fusedData instanceof Float32Array\n            ? fusedData\n            : new Float32Array(fusedData.buffer, fusedData.byteOffset, fusedRows * cols);\n\n        // Deinterleave per-head: each head has a block of 2*head_dim rows\n        // First head_dim rows = Q, second head_dim rows = gate\n        const headBlock = 2 * headDim; // rows per head in fused weight\n        for (let h = 0; h < numHeads; h++) {\n          const srcBase = h * headBlock * cols;\n          const dstBase = h * headDim * cols;\n          qData.set(src.subarray(srcBase, srcBase + headDim * cols), dstBase);\n          gateData.set(src.subarray(srcBase + headDim * cols, srcBase + headBlock * cols), dstBase);\n        }\n\n        await store.set(qProjKey, { data: qData, shape: [halfRows, cols] });\n        await store.set(attnGateKey, { data: gateData, shape: [halfRows, cols] });\n      }\n\n      // GPTQ: split fused q_proj GPTQ tensors into q_proj + attn_gate\n      // GPTQ q_proj has N=4096 (fused), graph expects N=2048 each.\n      // The deinterleave is per-head: each head's block of 2*head_dim output columns\n      // has head_dim Q columns followed by head_dim gate columns.\n      const qProjQweight = `layers.${i}.self_attn.q_proj.qweight`;\n      const attnGateQweight = `layers.${i}.self_attn.attn_gate.qweight`;\n      if (store.has(qProjQweight) && !store.has(attnGateQweight)) {\n        const fusedN = numHeads * 2 * headDim; // 4096\n        const halfN = numHeads * headDim; // 2048\n        const K_val = textCfg.hidden_size as number; // 1024\n\n        // Split qweight [K/8, fusedN] → two [K/8, halfN] with per-head deinterleave\n        const qw = (await store.get(qProjQweight))!;\n        const qwSrc =\n          qw.data instanceof Int32Array\n            ? qw.data\n            : new Int32Array(qw.data.buffer, qw.data.byteOffset, qw.data.byteLength / 4);\n        const kDiv8 = K_val >>> 3;\n        const qwQ = new Int32Array(kDiv8 * halfN);\n        const qwG = new Int32Array(kDiv8 * halfN);\n        const headBlock2 = 2 * headDim; // output columns per head in fused tensor\n        for (let row = 0; row < kDiv8; row++) {\n          for (let h = 0; h < numHeads; h++) {\n            const srcOff = row * fusedN + h * headBlock2;\n            const dstOff = row * halfN + h * headDim;\n            for (let d = 0; d < headDim; d++) {\n              qwQ[dstOff + d] = qwSrc[srcOff + d];\n              qwG[dstOff + d] = qwSrc[srcOff + headDim + d];\n            }\n          }\n        }\n        await store.set(qProjQweight, { data: qwQ, shape: [kDiv8, halfN] });\n        await store.set(attnGateQweight, { data: qwG, shape: [kDiv8, halfN] });\n\n        // Split scales [groups, fusedN] → two [groups, halfN]\n        const qProjScales = `layers.${i}.self_attn.q_proj.scales`;\n        const attnGateScales = `layers.${i}.self_attn.attn_gate.scales`;\n        const sc = (await store.get(qProjScales))!;\n        const scSrc =\n          sc.data instanceof Float32Array\n            ? sc.data\n            : new Float32Array(sc.data.buffer, sc.data.byteOffset, sc.data.byteLength / 4);\n        const groups = sc.shape[0];\n        const scQ = new Float32Array(groups * halfN);\n        const scG = new Float32Array(groups * halfN);\n        for (let g = 0; g < groups; g++) {\n          for (let h = 0; h < numHeads; h++) {\n            const srcOff = g * fusedN + h * headBlock2;\n            const dstOff = g * halfN + h * headDim;\n            for (let d = 0; d < headDim; d++) {\n              scQ[dstOff + d] = scSrc[srcOff + d];\n              scG[dstOff + d] = scSrc[srcOff + headDim + d];\n            }\n          }\n        }\n        await store.set(qProjScales, { data: scQ, shape: [groups, halfN] });\n        await store.set(attnGateScales, { data: scG, shape: [groups, halfN] });\n\n        // Split qzeros [groups, fusedN/8] → two [groups, halfN/8]\n        // qzeros packs 8 output columns per int32. Since headDim (256) is divisible by 8,\n        // the deinterleave works on int32 words directly: headDim/8 words per head.\n        const qProjQzeros = `layers.${i}.self_attn.q_proj.qzeros`;\n        const attnGateQzeros = `layers.${i}.self_attn.attn_gate.qzeros`;\n        const qz = (await store.get(qProjQzeros))!;\n        const qzSrc =\n          qz.data instanceof Int32Array\n            ? qz.data\n            : new Int32Array(qz.data.buffer, qz.data.byteOffset, qz.data.byteLength / 4);\n        const fusedNDiv8 = fusedN >>> 3;\n        const halfNDiv8 = halfN >>> 3;\n        const headBlockDiv8 = headBlock2 >>> 3;\n        const headDimDiv8 = headDim >>> 3;\n        const qzQ = new Int32Array(groups * halfNDiv8);\n        const qzG = new Int32Array(groups * halfNDiv8);\n        for (let g = 0; g < groups; g++) {\n          for (let h = 0; h < numHeads; h++) {\n            const srcOff = g * fusedNDiv8 + h * headBlockDiv8;\n            const dstOff = g * halfNDiv8 + h * headDimDiv8;\n            for (let d = 0; d < headDimDiv8; d++) {\n              qzQ[dstOff + d] = qzSrc[srcOff + d];\n              qzG[dstOff + d] = qzSrc[srcOff + headDimDiv8 + d];\n            }\n          }\n        }\n        await store.set(qProjQzeros, { data: qzQ, shape: [groups, halfNDiv8] });\n        await store.set(attnGateQzeros, { data: qzG, shape: [groups, halfNDiv8] });\n\n        // Clean up g_idx tensors if present (we assume sequential grouping)\n        await store.delete(`layers.${i}.self_attn.q_proj.g_idx`);\n      }\n\n      // MLX: split fused q_proj MLX tensors into q_proj + attn_gate\n      // MLX q_proj.weight is packed [fusedN, K/8], scales/biases are [fusedN, K/gs]\n      // The split is along the N (output/row) dimension — same deinterleave pattern.\n      const mlxQProjWeight = `layers.${i}.self_attn.q_proj.weight`;\n      const mlxAttnGateWeight = `layers.${i}.self_attn.attn_gate.weight`;\n      if (isMLX && store.has(mlxQProjWeight) && !store.has(mlxAttnGateWeight)) {\n        const K_val = textCfg.hidden_size as number; // 1024\n        const fusedN = numHeads * 2 * headDim; // 4096\n        const halfN = numHeads * headDim; // 2048\n        const headBlock2 = 2 * headDim;\n        const mlxGS = (quantConfig?.group_size as number) ?? 64;\n\n        // Deinterleave [fusedN, cols] → two [halfN, cols] with per-head split\n        const deinterleaveMLX = <T extends Float32Array | Uint32Array>(\n          src: T,\n          fusedRows: number,\n          cols: number,\n          Ctor: { new (len: number): T },\n        ): [T, T] => {\n          const halfRows = fusedRows / 2;\n          const qArr = new Ctor(halfRows * cols);\n          const gArr = new Ctor(halfRows * cols);\n          for (let h = 0; h < numHeads; h++) {\n            const srcBase = h * headBlock2 * cols;\n            const dstBase = h * headDim * cols;\n            for (let r = 0; r < headDim; r++) {\n              const s0 = srcBase + r * cols;\n              const s1 = srcBase + (headDim + r) * cols;\n              const d0 = dstBase + r * cols;\n              for (let c = 0; c < cols; c++) {\n                qArr[d0 + c] = src[s0 + c];\n                gArr[d0 + c] = src[s1 + c];\n              }\n            }\n          }\n          return [qArr, gArr];\n        };\n\n        // Split packed weight [fusedN, K/8]\n        const wt = (await store.get(mlxQProjWeight))!;\n        const wtSrc =\n          wt.data instanceof Uint32Array\n            ? wt.data\n            : new Uint32Array(wt.data.buffer, wt.data.byteOffset, wt.data.byteLength / 4);\n        const colsW = K_val >>> 3; // K/8\n        const [wtQ, wtG] = deinterleaveMLX(wtSrc, fusedN, colsW, Uint32Array);\n        await store.set(mlxQProjWeight, { data: wtQ, shape: [halfN, colsW] });\n        await store.set(mlxAttnGateWeight, { data: wtG, shape: [halfN, colsW] });\n\n        // Split scales [fusedN, K/gs]\n        const mlxQProjScales = `layers.${i}.self_attn.q_proj.scales`;\n        const mlxAttnGateScales = `layers.${i}.self_attn.attn_gate.scales`;\n        const sc = (await store.get(mlxQProjScales))!;\n        const scSrc =\n          sc.data instanceof Float32Array\n            ? sc.data\n            : new Float32Array(sc.data.buffer, sc.data.byteOffset, sc.data.byteLength / 4);\n        const colsS = Math.ceil(K_val / mlxGS);\n        const [scQ, scG] = deinterleaveMLX(scSrc, fusedN, colsS, Float32Array);\n        await store.set(mlxQProjScales, { data: scQ, shape: [halfN, colsS] });\n        await store.set(mlxAttnGateScales, { data: scG, shape: [halfN, colsS] });\n\n        // Split biases [fusedN, K/gs]\n        const mlxQProjBiases = `layers.${i}.self_attn.q_proj.biases`;\n        const mlxAttnGateBiases = `layers.${i}.self_attn.attn_gate.biases`;\n        const bi = (await store.get(mlxQProjBiases))!;\n        const biSrc =\n          bi.data instanceof Float32Array\n            ? bi.data\n            : new Float32Array(bi.data.buffer, bi.data.byteOffset, bi.data.byteLength / 4);\n        const [biQ, biG] = deinterleaveMLX(biSrc, fusedN, colsS, Float32Array);\n        await store.set(mlxQProjBiases, { data: biQ, shape: [halfN, colsS] });\n        await store.set(mlxAttnGateBiases, { data: biG, shape: [halfN, colsS] });\n      }\n    }\n  }\n\n  // -- Apply (1 + weight) offset for ALL Qwen3.5 RMSNorm weights --\n  // Qwen3_5RMSNorm initializes weight=0 and applies (1 + weight) * normalized.\n  // ALL norms in the model use this class: input_layernorm, post_attention_layernorm,\n  // final_norm, q_norm, k_norm, and linear_attn.norm. We bake +1 into weights so\n  // the standard RMSNorm kernel (which does weight * normalized) produces correct results.\n  //\n  // EXCEPTION: MLX Community models already absorb the +1 during conversion —\n  // their RMSNorm uses `weight * normalized` (no +1), so the stored weights\n  // are already `1 + original_weight`. Applying +1 again would double-offset.\n  if (architectureName === \"Qwen3_5ForConditionalGeneration\" && !isMLX) {\n    const textCfg = (rawConfig.text_config as Record<string, unknown>) ?? rawConfig;\n    const numLayers = textCfg.num_hidden_layers as number;\n    const layerTypesArr = (textCfg.layer_types as string[]) ?? [];\n    const fullAttnInterval = (textCfg.full_attention_interval as number) ?? 4;\n\n    // Collect all norm weight keys\n    const normKeys: string[] = [];\n    for (let i = 0; i < numLayers; i++) {\n      // All layers have input/post-attention layernorms\n      normKeys.push(`layers.${i}.input_layernorm.weight`);\n      normKeys.push(`layers.${i}.post_attention_layernorm.weight`);\n\n      let isFullAttn: boolean;\n      if (layerTypesArr.length > i) {\n        isFullAttn = layerTypesArr[i] === \"full_attention\";\n      } else {\n        isFullAttn = i % fullAttnInterval === fullAttnInterval - 1;\n      }\n\n      if (isFullAttn) {\n        normKeys.push(`layers.${i}.self_attn.q_norm.weight`);\n        normKeys.push(`layers.${i}.self_attn.k_norm.weight`);\n      }\n      // NOTE: linear_attn.norm.weight uses Qwen3_5RMSNormGated which\n      // initializes weight=1 (not 0), so it does NOT need the +1 offset.\n    }\n    // Final norm\n    normKeys.push(\"norm.weight\");\n\n    for (const normKey of normKeys) {\n      const entry = await store.get(normKey);\n      if (entry) {\n        const data =\n          entry.data instanceof Float32Array\n            ? entry.data\n            : new Float32Array(entry.data.buffer, entry.data.byteOffset, entry.data.byteLength / 4);\n        const adjusted = new Float32Array(data.length);\n        for (let j = 0; j < data.length; j++) {\n          adjusted[j] = 1.0 + data[j];\n        }\n        await store.set(normKey, { data: adjusted, shape: entry.shape });\n      }\n    }\n  }\n\n  // -- Apply (1 + weight) offset for ALL Gemma3 RMSNorm weights --\n  // Gemma3RMSNorm computes `(1.0 + weight) * normalized` (mlx-lm:\n  // mx.fast.rms_norm(x, 1.0 + self.weight, eps)). The stored weight is the raw\n  // trained value centered near 0; we bake +1 so the standard RMSNorm kernel\n  // (weight * normalized) is correct. Unlike Qwen3.5, the MLX convert does NOT\n  // absorb the +1 for Gemma, so this runs for MLX too. Every per-layer norm uses\n  // this class: input_layernorm, post_attention_layernorm, pre_feedforward_\n  // layernorm, post_feedforward_layernorm, q_norm, k_norm, plus the final norm.\n  if (architectureName === \"Gemma3TextModel\" || architectureName === \"Gemma3Model\") {\n    const gemmaNumLayers = rawConfig.num_hidden_layers as number;\n    const gemmaNormKeys: string[] = [];\n    for (let i = 0; i < gemmaNumLayers; i++) {\n      gemmaNormKeys.push(`layers.${i}.input_layernorm.weight`);\n      gemmaNormKeys.push(`layers.${i}.post_attention_layernorm.weight`);\n      gemmaNormKeys.push(`layers.${i}.pre_feedforward_layernorm.weight`);\n      gemmaNormKeys.push(`layers.${i}.post_feedforward_layernorm.weight`);\n      gemmaNormKeys.push(`layers.${i}.self_attn.q_norm.weight`);\n      gemmaNormKeys.push(`layers.${i}.self_attn.k_norm.weight`);\n    }\n    gemmaNormKeys.push(\"norm.weight\");\n\n    for (const normKey of gemmaNormKeys) {\n      const entry = await store.get(normKey);\n      if (entry) {\n        const data =\n          entry.data instanceof Float32Array\n            ? entry.data\n            : new Float32Array(entry.data.buffer, entry.data.byteOffset, entry.data.byteLength / 4);\n        const adjusted = new Float32Array(data.length);\n        for (let j = 0; j < data.length; j++) {\n          adjusted[j] = 1.0 + data[j];\n        }\n        await store.set(normKey, { data: adjusted, shape: entry.shape });\n      }\n    }\n  }\n\n  // NOTE on Gemma 4 RMSNorm: unlike Gemma 3 (weights centered near 0, runtime\n  // `(1+weight)`), the released Gemma 4 / Gemma3n checkpoints store the FULL norm\n  // gain directly (raw input_layernorm values are ~7–68, rms ~12). So NO +1 bake\n  // is applied here — the standard `weight * normalized` kernel is already correct,\n  // and the large residual-stream magnitudes are expected for this model.\n\n  // -- Gemma 4 per-layer output scalar (`layer_scalar`) --\n  // Gemma4TextDecoderLayer ends with `hidden_states *= self.layer_scalar` — a single\n  // learned scalar per layer (shape [1], values ~0.018–0.5, NOT ones). The graph\n  // emits a placeholder Scale node (id `layer{i}_layer_scalar`); we read the scalar\n  // value here and patch the node's `scale` attribute, then drop the weight (it is a\n  // baked constant, not a GPU tensor). Without this the residual stream is unscaled\n  // and generation is incoherent.\n  if (architectureName === \"Gemma4ForConditionalGeneration\") {\n    for (const node of graph.nodes) {\n      const key = node.attributes?.layer_scalar_key as string | undefined;\n      if (!key) continue;\n      const entry = await store.get(key);\n      if (!entry) continue;\n      const data =\n        entry.data instanceof Float32Array\n          ? entry.data\n          : new Float32Array(entry.data.buffer, entry.data.byteOffset, entry.data.byteLength / 4);\n      node.attributes.scale = data[0];\n      await store.delete(key);\n    }\n  }\n\n  // -- Handle tied embeddings --\n  // When safetensorsKey differs from the tensor name (e.g. lm_head.weight\n  // pointing to embed_tokens.weight for tied embeddings), copy the data.\n  for (const [name, desc] of Object.entries(graph.tensors)) {\n    if (\n      desc.storage === \"constant\" &&\n      desc.safetensorsKey &&\n      desc.safetensorsKey !== name &&\n      !store.has(name)\n    ) {\n      const sourceData = await store.get(desc.safetensorsKey);\n      if (sourceData) {\n        await store.set(name, sourceData);\n      }\n    }\n  }\n\n  // -- Repack GPTQ weights into Gerbil INT4 format --\n  if (isGPTQ) {\n    onProgress?.(96, 100, \"Repacking GPTQ weights...\");\n    const gptqGroupSize = (quantConfig?.group_size as number) ?? 128;\n\n    for (const node of graph.nodes) {\n      if (node.opType !== \"MatMulInt4\") continue;\n\n      const K = node.attributes.K as number;\n      const N = node.attributes.N as number;\n\n      const qTensorName = node.inputs[1]; // e.g. layers.0.self_attn.q_proj.weight.q\n      const scalesTensorName = node.inputs[2]; // e.g. layers.0.self_attn.q_proj.weight.scales\n      const zerosTensorName = node.inputs[3]; // e.g. layers.0.self_attn.q_proj.weight.zeros\n\n      // Derive GPTQ tensor names from graph names:\n      // Graph: \"layers.0.self_attn.q_proj.weight.q\" → strip \".q\" → strip \".weight\" → base\n      const sourceName = qTensorName.slice(0, -2); // \"layers.0.self_attn.q_proj.weight\"\n      const gptqBase = sourceName.slice(0, -7); // \"layers.0.self_attn.q_proj\"\n\n      const gptqQweight = await store.get(`${gptqBase}.qweight`);\n      const gptqScales = await store.get(`${gptqBase}.scales`);\n      const gptqQzeros = await store.get(`${gptqBase}.qzeros`);\n\n      // Skip layers not quantized in GPTQ — they'll be handled by on-the-fly quantization below\n      if (!gptqQweight || !gptqScales || !gptqQzeros) continue;\n\n      // Ensure correct typed arrays for the adapter\n      const qweightData =\n        gptqQweight.data instanceof Int32Array\n          ? gptqQweight.data\n          : new Int32Array(\n              gptqQweight.data.buffer,\n              gptqQweight.data.byteOffset,\n              gptqQweight.data.byteLength / 4,\n            );\n      const scalesData =\n        gptqScales.data instanceof Float32Array\n          ? gptqScales.data\n          : new Float32Array(\n              gptqScales.data.buffer,\n              gptqScales.data.byteOffset,\n              gptqScales.data.byteLength / 4,\n            );\n      const qzerosData =\n        gptqQzeros.data instanceof Int32Array\n          ? gptqQzeros.data\n          : new Int32Array(\n              gptqQzeros.data.buffer,\n              gptqQzeros.data.byteOffset,\n              gptqQzeros.data.byteLength / 4,\n            );\n\n      const repacked = repackGPTQ({\n        qweight: qweightData,\n        scales: scalesData,\n        qzeros: qzerosData,\n        K,\n        N,\n        groupSize: gptqGroupSize,\n      });\n\n      await store.set(qTensorName, { data: repacked.packed, shape: [repacked.packed.length] });\n      await store.set(scalesTensorName, { data: repacked.scales, shape: [repacked.scales.length] });\n      await store.set(zerosTensorName, { data: repacked.zeros, shape: [repacked.zeros.length] });\n\n      // Clean up GPTQ source tensors to free memory\n      await store.delete(`${gptqBase}.qweight`);\n      await store.delete(`${gptqBase}.scales`);\n      await store.delete(`${gptqBase}.qzeros`);\n    }\n  }\n\n  // -- Repack MLX 4-bit weights into Gerbil INT4 format --\n  if (isMLX) {\n    onProgress?.(96, 100, \"Repacking MLX weights...\");\n    const mlxGS = (quantConfig?.group_size as number) ?? 64;\n\n    for (const node of graph.nodes) {\n      if (node.opType !== \"MatMulInt4\" && node.opType !== \"EmbeddingInt4\") continue;\n\n      // Derive N and K for this layer\n      let N: number, K: number;\n      if (node.opType === \"EmbeddingInt4\") {\n        N = node.attributes.vocab_size as number;\n        K = node.attributes.hidden_size as number;\n      } else {\n        N = node.attributes.N as number;\n        K = node.attributes.K as number;\n      }\n\n      const qTensorName = node.inputs[1]; // e.g. layers.0.self_attn.q_proj.weight.q\n      const scalesTensorName = node.inputs[2]; // e.g. layers.0.self_attn.q_proj.weight.scales\n      const zerosTensorName = node.inputs[3]; // e.g. layers.0.self_attn.q_proj.weight.zeros\n\n      // Derive MLX source tensor names:\n      // Graph: \"base.weight.q\" → strip \".q\" → \"base.weight\" (MLX packed weight)\n      // MLX scales: \"base.scales\", MLX biases: \"base.biases\"\n      const sourceName = qTensorName.slice(0, -2); // \"base.weight\"\n      const baseName = sourceName.slice(0, -7); // \"base\" (strip \".weight\")\n\n      const mlxWeight = await store.get(sourceName); // packed U32\n      const mlxScales = await store.get(`${baseName}.scales`); // F32 (already converted from BF16)\n      const mlxBiases = await store.get(`${baseName}.biases`); // F32 (already converted from BF16)\n\n      if (!mlxWeight || !mlxScales || !mlxBiases) continue;\n\n      const weightData =\n        mlxWeight.data instanceof Uint32Array\n          ? mlxWeight.data\n          : new Uint32Array(\n              mlxWeight.data.buffer,\n              mlxWeight.data.byteOffset,\n              mlxWeight.data.byteLength / 4,\n            );\n      const scalesData =\n        mlxScales.data instanceof Float32Array\n          ? mlxScales.data\n          : new Float32Array(\n              mlxScales.data.buffer,\n              mlxScales.data.byteOffset,\n              mlxScales.data.byteLength / 4,\n            );\n      const biasesData =\n        mlxBiases.data instanceof Float32Array\n          ? mlxBiases.data\n          : new Float32Array(\n              mlxBiases.data.buffer,\n              mlxBiases.data.byteOffset,\n              mlxBiases.data.byteLength / 4,\n            );\n\n      const repacked = repackMLX({\n        weight: weightData,\n        scales: scalesData,\n        biases: biasesData,\n        N,\n        K,\n        groupSize: mlxGS,\n      });\n\n      await store.set(qTensorName, { data: repacked.packed, shape: [repacked.packed.length] });\n      await store.set(scalesTensorName, { data: repacked.scales, shape: [repacked.scales.length] });\n      await store.set(zerosTensorName, { data: repacked.zeros, shape: [repacked.zeros.length] });\n\n      // Clean up MLX source tensors to free memory\n      await store.delete(sourceName);\n      await store.delete(`${baseName}.scales`);\n      await store.delete(`${baseName}.biases`);\n    }\n  }\n\n  // -- Quantize weights for INT4 if requested (or for non-GPTQ/MLX layers) --\n  if (resolvedDtype === \"q4\") {\n    onProgress?.(96, 100, \"Quantizing weights to INT4...\");\n    for (const node of graph.nodes) {\n      if (node.opType !== \"MatMulInt4\" && node.opType !== \"EmbeddingInt4\") continue;\n\n      const groupSz = (node.attributes.group_size as number) ?? DEFAULT_GROUP_SIZE;\n\n      const qTensorName = node.inputs[1]; // weight.q\n      const scalesTensorName = node.inputs[2]; // weight.scales\n      const zerosTensorName = node.inputs[3]; // weight.zeros\n\n      // Skip nodes already handled by GPTQ or MLX repacking\n      if (store.has(qTensorName)) continue;\n\n      // Derive source weight name by stripping \".q\" suffix\n      const sourceName = qTensorName.slice(0, -2);\n      let sourceWeight = await store.get(sourceName);\n      // Tied embeddings fallback: lm_head.weight may not exist in safetensors\n      // when tie_word_embeddings=true — use embed_tokens.weight instead.\n      if (!sourceWeight && sourceName === \"lm_head.weight\") {\n        sourceWeight = await store.get(\"embed_tokens.weight\");\n      }\n\n      if (sourceWeight) {\n        const f32Data =\n          sourceWeight.data instanceof Float32Array\n            ? sourceWeight.data\n            : new Float32Array(\n                sourceWeight.data.buffer,\n                sourceWeight.data.byteOffset,\n                sourceWeight.data.byteLength / 4,\n              );\n\n        const { packed, scales, zeros } = quantizeInt4(f32Data, groupSz);\n\n        await store.set(qTensorName, { data: packed, shape: [packed.length] });\n        await store.set(scalesTensorName, { data: scales, shape: [scales.length] });\n        await store.set(zerosTensorName, { data: zeros, shape: [zeros.length] });\n\n        // Free the original F32 weight to save memory\n        await store.delete(sourceName);\n      }\n    }\n  }\n\n  // -- MLX vision patch_embed axis fixup --\n  // The ViT patch_embed conv weight is a 5D tensor whose flattening must match\n  // the host patch vector's [C, T, ph, pw] order. The HF BF16 checkpoint stores\n  // it channels-first as [out, C, T, ph, pw] → flattens correctly as-is. The\n  // mlx-vlm convert stores it channels-LAST as [out, T, ph, pw, C], so the\n  // flat [out, 1536] inner order is permuted and the first matmul would be fed\n  // mismatched weights (manifests as low cosine, not NaN). Detect the\n  // channels-last 5D shape and permute back to [out, C, T, ph, pw] so the\n  // flattened weight matches the host patch layout on both checkpoints.\n  if (isMLX) {\n    const pe = await store.get(CANONICAL_KEYS.visPatchEmbedWeight);\n    const s = pe?.shape;\n    // channels-last conv weight: [out, T, ph, pw, C] with C === in_channels (3)\n    if (pe && s && s.length === 5 && s[4] === 3) {\n      const [outC, T, ph, pw, C] = s;\n      const src =\n        pe.data instanceof Float32Array\n          ? pe.data\n          : new Float32Array(pe.data.buffer, pe.data.byteOffset, pe.data.byteLength / 4);\n      const dst = new Float32Array(src.length);\n      // src index (channels-last):  ((((o*T + t)*ph + y)*pw + x)*C + c)\n      // dst index (channels-first): ((((o*C + c)*T + t)*ph + y)*pw + x)\n      for (let o = 0; o < outC; o++) {\n        for (let t = 0; t < T; t++) {\n          for (let y = 0; y < ph; y++) {\n            for (let x = 0; x < pw; x++) {\n              for (let c = 0; c < C; c++) {\n                const si = (((o * T + t) * ph + y) * pw + x) * C + c;\n                const di = (((o * C + c) * T + t) * ph + y) * pw + x;\n                dst[di] = src[si];\n              }\n            }\n          }\n        }\n      }\n      // Store flattened [out, C*T*ph*pw] = [out, 1536] in channels-first order.\n      await store.set(CANONICAL_KEYS.visPatchEmbedWeight, {\n        data: dst,\n        shape: [outC, C * T * ph * pw],\n      });\n    }\n  }\n\n  // -- Gemma 4: extract the PLE table to CPU-resident storage (NOT GPU) --\n  // The per-layer embedding table `embed_tokens_per_layer` is ~1.17GB at 4-bit.\n  // We keep it in JS memory and stream per-token rows at inference, so it never\n  // becomes a GPU buffer. Repack MLX 4-bit → Gerbil INT4 here; for an unquantized\n  // checkpoint, encode the f32 rows with a unit scale (zero=0) so the executor's\n  // single dequant path handles both. The PLE EmbeddingInt4 node was intentionally\n  // omitted from the graph (gemma4.ts), so the standard repack loop skips this\n  // table — it is the loader's job to divert it out of `weights`.\n  let pleSource: PleSource | undefined;\n  const isGemma4 = String(architectureName).startsWith(\"Gemma4\");\n  if (isGemma4) {\n    const pleBase = CANONICAL_KEYS.GEMMA4_PLE_EMBED.slice(0, -7); // \"embed_tokens_per_layer\"\n    const textCfg = (rawConfig.text_config as Record<string, unknown>) ?? rawConfig;\n    const pleDim = (textCfg.hidden_size_per_layer_input as number) ?? 256;\n    const numLayers = textCfg.num_hidden_layers as number;\n    const pleWidth = numLayers * pleDim; // L*256\n    const pleVocab =\n      (textCfg.vocab_size_per_layer_input as number) ?? (textCfg.vocab_size as number);\n\n    const packedEntry = await store.get(`${pleBase}.weight`);\n    const scalesEntry = await store.get(`${pleBase}.scales`);\n    const biasesEntry = await store.get(`${pleBase}.biases`);\n\n    if (isMLX && packedEntry && scalesEntry && biasesEntry) {\n      const mlxGS = (quantConfig?.group_size as number) ?? 64;\n      const packedU32 =\n        packedEntry.data instanceof Uint32Array\n          ? packedEntry.data\n          : new Uint32Array(\n              packedEntry.data.buffer,\n              packedEntry.data.byteOffset,\n              packedEntry.data.byteLength / 4,\n            );\n      const scalesF32 =\n        scalesEntry.data instanceof Float32Array\n          ? scalesEntry.data\n          : new Float32Array(\n              scalesEntry.data.buffer,\n              scalesEntry.data.byteOffset,\n              scalesEntry.data.byteLength / 4,\n            );\n      const biasesF32 =\n        biasesEntry.data instanceof Float32Array\n          ? biasesEntry.data\n          : new Float32Array(\n              biasesEntry.data.buffer,\n              biasesEntry.data.byteOffset,\n              biasesEntry.data.byteLength / 4,\n            );\n      const repacked = repackMLX({\n        weight: packedU32,\n        scales: scalesF32,\n        biases: biasesF32,\n        N: pleVocab,\n        K: pleWidth,\n        groupSize: mlxGS,\n      });\n      pleSource = await buildPleSource(\n        repacked.packed,\n        repacked.scales,\n        repacked.zeros,\n        pleWidth,\n        mlxGS,\n        cacheStore,\n      );\n      // Divert out of the store so uploadWeights() never sends it to the GPU.\n      await store.delete(`${pleBase}.weight`);\n      await store.delete(`${pleBase}.scales`);\n      await store.delete(`${pleBase}.biases`);\n      console.log(\n        `[Gerbil] Gemma4 PLE table kept ${cacheStore ? \"cache-resident\" : \"CPU-resident\"}: ` +\n          `[${pleVocab}, ${pleWidth}] int4 ` +\n          `(~${((repacked.packed.byteLength + repacked.scales.byteLength + repacked.zeros.byteLength) / 1e6).toFixed(0)} MB, 0 MB GPU)`,\n      );\n    } else if (packedEntry?.data instanceof Float32Array) {\n      // Unquantized f32 table: quantize to INT4 for compact storage and a\n      // single shared dequant path. (Quantization-sensitive precision concerns\n      // only apply when forced GPU-resident; here we still dequantize per-row.)\n      const { packed, scales, zeros } = quantizeInt4(packedEntry.data, DEFAULT_GROUP_SIZE);\n      pleSource = await buildPleSource(\n        packed,\n        scales,\n        zeros,\n        pleWidth,\n        DEFAULT_GROUP_SIZE,\n        cacheStore,\n      );\n      await store.delete(`${pleBase}.weight`);\n    } else {\n      console.warn(\n        \"[Gerbil] Gemma4: PLE table not found in weights — per-layer embeddings \" +\n          \"will be zero. Check the checkpoint ships embed_tokens_per_layer.\",\n      );\n    }\n  }\n\n  // -- Materialize synthetic fill-value constants (e.g. Gemma 4 v_norm ones) --\n  // Tensors declared with `fillValue` carry no checkpoint weight; the loader fills\n  // them with a constant at the declared shape. Used for parameter-free norms.\n  for (const [name, desc] of Object.entries(graph.tensors)) {\n    if (desc.storage === \"constant\" && desc.fillValue !== undefined && !store.has(name)) {\n      const length = desc.shape.reduce(\n        (acc: number, dim) => acc * (typeof dim === \"number\" ? dim : 1),\n        1,\n      );\n      const data = new Float32Array(length).fill(desc.fillValue);\n      await store.set(name, {\n        data,\n        shape: desc.shape.map((d) => (typeof d === \"number\" ? d : 1)),\n      });\n    }\n  }\n\n  // -- Verify required tensors --\n  const missingTensors: string[] = [];\n  for (const [name, desc] of Object.entries(graph.tensors)) {\n    if (desc.storage === \"constant\" && !store.has(name)) {\n      missingTensors.push(name);\n    }\n  }\n\n  if (missingTensors.length > 0) {\n    console.warn(\n      `[Gerbil] Missing ${missingTensors.length} weight tensors:`,\n      missingTensors.slice(0, 10),\n    );\n  }\n\n  onProgress?.(100, 100, \"Model loaded.\");\n\n  return { graph, tokenizer, weights: store, rawConfig, pleSource };\n}\n\n// ── Moonshine STT loader ───────────────────────────────────────────────────\n//\n// Moonshine is an encoder-decoder ASR model that cannot be expressed as a single\n// causal-LM graph (the standard loadModel path). It needs two graphs (a conv\n// encoder run once, an AR decoder with cross-attention) sharing one set of f32\n// weights. This loader downloads config + tokenizer + the single model.safetensors\n// and returns the canonical-named f32 weights for both graphs. Keys are mapped by\n// stripping the \"model.\" prefix the checkpoint uses (e.g. \"model.encoder.conv1.\n// weight\" → \"encoder.conv1.weight\"), which is exactly the safetensorsKey the\n// Moonshine graph generators reference.\n\nexport interface LoadedMoonshine {\n  /** Canonical-named f32 weights (data + shape), shared by encoder + decoder graphs. */\n  weights: Map<string, { data: ArrayBufferView; shape: number[] }>;\n  /** The (decode-capable) tokenizer. */\n  tokenizer: Tokenizer;\n  /** Raw config.json. */\n  rawConfig: Record<string, unknown>;\n}\n\nexport async function loadMoonshine(options: {\n  repo: string;\n  revision?: string;\n  hfToken?: string;\n  cacheDir?: string;\n  onProgress?: (loaded: number, total: number, message: string) => void;\n}): Promise<LoadedMoonshine> {\n  const revision = options.revision ?? \"main\";\n  const baseURL = resolveHFBaseURL(options.repo, revision);\n  await initFs();\n\n  options.onProgress?.(0, 100, \"Fetching config + tokenizer…\");\n  const [rawConfig, tokenizerJSON] = await Promise.all([\n    fetchJSON(baseURL, \"config.json\", options.hfToken, options.cacheDir),\n    fetchJSON(baseURL, \"tokenizer.json\", options.hfToken, options.cacheDir),\n  ]);\n  const tokenizerConfig = await fetchJSON(\n    baseURL,\n    \"tokenizer_config.json\",\n    options.hfToken,\n    options.cacheDir,\n  ).catch(() => null);\n  const tokenizer = Tokenizer.fromJSON(tokenizerJSON, tokenizerConfig);\n\n  options.onProgress?.(10, 100, \"Downloading weights…\");\n  const buf = await fetchBinary(\n    baseURL,\n    \"model.safetensors\",\n    options.hfToken,\n    (loaded, total) => options.onProgress?.(10 + (loaded / (total || 1)) * 85, 100, \"weights\"),\n    options.cacheDir,\n  );\n\n  const file = parseSafetensorsHeader(buf);\n  const weights = new Map<string, { data: ArrayBufferView; shape: number[] }>();\n  for (const entry of file.entries) {\n    // Strip the \"model.\" prefix the Moonshine checkpoint uses on every tensor.\n    const canonical = entry.name.startsWith(\"model.\") ? entry.name.slice(6) : entry.name;\n    let data: ArrayBufferView = getTensorData(buf, file, entry);\n    if (entry.dtype === \"BF16\") {\n      data = bf16ToF32(new Uint8Array(data.buffer, data.byteOffset, data.byteLength));\n    } else if (entry.dtype === \"F16\") {\n      data = f16ToF32(data);\n    }\n    weights.set(canonical, { data, shape: entry.shape });\n  }\n\n  options.onProgress?.(100, 100, \"Moonshine loaded.\");\n  return { weights, tokenizer, rawConfig };\n}\n\n// ════════════════════════════════════════════════════════════════════════════\n// Kani-TTS-2 dual-checkpoint loader (LFM2-350M codec-LM + NVIDIA NeMo NanoCodec).\n//\n// Kani-TTS-2 is a two-stage TTS model that, like Moonshine, cannot be one graph:\n//   1. the codec-LM backbone (LFM2-350M body) lives in the main repo\n//      (nineninesix/kani-tts-2-en, model.safetensors, \"model.\" prefix), and\n//   2. the NanoCodec decoder weights live in a SEPARATE repo\n//      (nineninesix/nemo-nano-codec-22khz-0.6kbps-12.5fps-MLX, model.safetensors)\n//      stored as weight-norm parametrizations under verbose NeMo key paths.\n//\n// This loader downloads both, folds the codec's weight-norm (W = g·v/‖v‖) and\n// renames its keys to the canonical `nanocodec.*` names the decoder graph expects,\n// and returns one merged f32 weight map (+ tokenizer + config) for the KaniTTS\n// consumer (src/gpu/kani-tts.ts), which builds the backbone + decoder graphs.\n// ════════════════════════════════════════════════════════════════════════════\n\n/** Default NanoCodec checkpoint Kani-TTS-2 uses (NeMo 22 kHz, 0.6 kbps, 12.5 fps). */\nexport const KANI_NANOCODEC_REPO = \"nineninesix/nemo-nano-codec-22khz-0.6kbps-12.5fps-MLX\";\n\nexport interface LoadedKaniTTS {\n  /** Canonical-named f32 weights for the codec-LM backbone (LFM2 keys). */\n  backboneWeights: Map<string, { data: ArrayBufferView; shape: number[] }>;\n  /** Folded NanoCodec decoder weights under canonical `nanocodec.*` names. */\n  codecWeights: Map<string, { data: ArrayBufferView; shape: number[] }>;\n  /** The text tokenizer. */\n  tokenizer: Tokenizer;\n  /** Raw backbone config.json (LFM2 dims + KaniTTS2 fields). */\n  rawConfig: Record<string, unknown>;\n}\n\n/** Matches the LFM2 attention out_proj key (renamed to o_proj for the canonical map). */\nconst KANI_OUT_PROJ_RE = /\\.self_attn\\.out_proj\\./;\n\n/** Coerce a safetensors tensor view (bf16/f16→f32) to a contiguous Float32Array. */\nfunction toF32(entry: SafetensorEntry, buf: ArrayBuffer, file: SafetensorsFile): Float32Array {\n  let data: ArrayBufferView = getTensorData(buf, file, entry);\n  if (entry.dtype === \"BF16\") {\n    data = bf16ToF32(new Uint8Array(data.buffer, data.byteOffset, data.byteLength));\n  } else if (entry.dtype === \"F16\") {\n    data = f16ToF32(data);\n  }\n  return data instanceof Float32Array\n    ? data\n    : new Float32Array(data.buffer, data.byteOffset, data.byteLength / 4);\n}\n\n/** Fold + rename the raw NanoCodec tensors into the canonical `nanocodec.*` map. */\nfunction buildKaniCodecWeights(\n  codecRaw: Map<string, { data: Float32Array; shape: number[] }>,\n): Map<string, { data: ArrayBufferView; shape: number[] }> {\n  const out = new Map<string, { data: ArrayBufferView; shape: number[] }>();\n  const { convs, plains } = nanoCodecWeightMap();\n  for (const c of convs) {\n    const g = codecRaw.get(c.gKey);\n    const v = codecRaw.get(c.vKey);\n    if (!(g && v)) {\n      throw new Error(\n        `NanoCodec weight-norm pair missing for ${c.graphKey}: ${c.gKey} / ${c.vKey}. ` +\n          \"Check the codec checkpoint (expected weight-norm parametrizations).\",\n      );\n    }\n    out.set(c.graphKey, {\n      data: foldNanoCodecWeightNorm(g.data, v.data, v.shape),\n      shape: v.shape,\n    });\n  }\n  for (const pl of plains) {\n    const t = codecRaw.get(pl.mlxKey);\n    if (!t) {\n      throw new Error(`NanoCodec tensor missing: ${pl.mlxKey} (for ${pl.graphKey}).`);\n    }\n    out.set(pl.graphKey, { data: t.data, shape: t.shape });\n  }\n  return out;\n}\n\nexport async function loadKaniTTS(options: {\n  /** Backbone repo (default nineninesix/kani-tts-2-en). */\n  repo?: string;\n  /** NanoCodec repo (default KANI_NANOCODEC_REPO). */\n  codecRepo?: string;\n  revision?: string;\n  hfToken?: string;\n  cacheDir?: string;\n  onProgress?: (loaded: number, total: number, message: string) => void;\n}): Promise<LoadedKaniTTS> {\n  const revision = options.revision ?? \"main\";\n  const repo = options.repo ?? \"nineninesix/kani-tts-2-en\";\n  const codecRepo = options.codecRepo ?? KANI_NANOCODEC_REPO;\n  const baseURL = resolveHFBaseURL(repo, revision);\n  const codecURL = resolveHFBaseURL(codecRepo, \"main\");\n  await initFs();\n\n  options.onProgress?.(0, 100, \"Fetching config + tokenizer…\");\n  const [rawConfig, tokenizerJSON] = await Promise.all([\n    fetchJSON(baseURL, \"config.json\", options.hfToken, options.cacheDir),\n    fetchJSON(baseURL, \"tokenizer.json\", options.hfToken, options.cacheDir),\n  ]);\n  const tokenizerConfig = await fetchJSON(\n    baseURL,\n    \"tokenizer_config.json\",\n    options.hfToken,\n    options.cacheDir,\n  ).catch(() => null);\n  const tokenizer = Tokenizer.fromJSON(tokenizerJSON, tokenizerConfig);\n\n  // ── Backbone weights (LFM2-350M): strip \"model.\" and rename out_proj→o_proj. ──\n  options.onProgress?.(8, 100, \"Downloading codec-LM backbone…\");\n  const backboneBuf = await fetchBinary(\n    baseURL,\n    \"model.safetensors\",\n    options.hfToken,\n    (loaded, total) =>\n      options.onProgress?.(8 + (loaded / (total || 1)) * 45, 100, \"backbone weights\"),\n    options.cacheDir,\n  );\n  const backboneFile = parseSafetensorsHeader(backboneBuf);\n  const backboneWeights = new Map<string, { data: ArrayBufferView; shape: number[] }>();\n  for (const entry of backboneFile.entries) {\n    const key = (entry.name.startsWith(\"model.\") ? entry.name.slice(6) : entry.name).replace(\n      KANI_OUT_PROJ_RE,\n      \".self_attn.o_proj.\",\n    );\n    backboneWeights.set(key, {\n      data: toF32(entry, backboneBuf, backboneFile),\n      shape: entry.shape,\n    });\n  }\n\n  // ── NanoCodec weights: fold weight-norm + rename to canonical nanocodec.* ──\n  options.onProgress?.(55, 100, \"Downloading NanoCodec decoder…\");\n  // The codec's model.safetensors shares its name with the backbone's, and\n  // fetchBinary caches by filename only — give the codec its own cache subdir so\n  // the two downloads do not collide.\n  const codecCacheDir = options.cacheDir ? `${options.cacheDir}/nanocodec` : undefined;\n  const codecBuf = await fetchBinary(\n    codecURL,\n    \"model.safetensors\",\n    options.hfToken,\n    (loaded, total) =>\n      options.onProgress?.(55 + (loaded / (total || 1)) * 40, 100, \"codec weights\"),\n    codecCacheDir,\n  );\n  const codecFile = parseSafetensorsHeader(codecBuf);\n  const codecRaw = new Map<string, { data: Float32Array; shape: number[] }>();\n  for (const entry of codecFile.entries) {\n    codecRaw.set(entry.name, { data: toF32(entry, codecBuf, codecFile), shape: entry.shape });\n  }\n  const codecWeights = buildKaniCodecWeights(codecRaw);\n\n  options.onProgress?.(100, 100, \"Kani-TTS-2 loaded.\");\n  return { backboneWeights, codecWeights, tokenizer, rawConfig };\n}\n","/**\n * MoonshineEncoderExecutor — runs the Moonshine STT encoder graph on WebGPU.\n *\n * Modeled on VisionExecutor: a single non-autoregressive prefill (conv frontend +\n * bidirectional transformer) run ONCE per utterance. The graph is length-static\n * (the Conv1dFull ops carry concrete L/Lout), so it is regenerated per utterance\n * from the input sample count (moonshineEncoderFrames) and a fresh executor is\n * built for that length.\n *\n * Beyond the encoder output, the graph also projects the frozen encoder hidden\n * state through each DECODER layer's encoder_attn.k_proj / v_proj, producing\n * enc_k_layer{i} / enc_v_layer{i}. The host reads these back and binds them frozen\n * into the decoder's CrossAttention across every AR decode step.\n *\n * WebKit discipline mirrors VisionExecutor: per-dispatch submit + drain, and the\n * bidirectional Attention op is windowed by query-row so no single dispatch runs\n * long enough to trip Metal's GPU watchdog. Dawn uses one submission.\n */\n\nimport type { GPUContext } from \"./device.js\";\nimport {\n  createBindGroup,\n  createStorageBuffer,\n  createUniformBuffer,\n  destroyBuffers,\n  getOrCreatePipeline,\n} from \"./device.js\";\nimport { DTYPE_BYTES, type ModelGraph, type OpNode } from \"./ir.js\";\nimport {\n  KERNEL_REGISTRY,\n  type KernelSpec,\n  LAYERNORM_NO_BIAS_SPEC,\n  ROPE_INTERLEAVED_SPEC,\n} from \"./kernels/registry.js\";\n\nconst MAP_MODE_READ = 0x0001;\n\n/**\n * WebKit only: window the bidirectional encoder Attention grid into chunks of this\n * many query rows, each its own submit + drain (same watchdog mitigation the vision\n * tower uses). Encoder frames are far fewer than ViT patches, but a long clip can\n * still produce a few hundred frames, so we keep the guard.\n */\nconst WEBKIT_ATTN_CHUNK_ROWS = 64;\n\ninterface MoonshineDispatch {\n  node: OpNode;\n  spec: KernelSpec;\n  pipeline: GPUComputePipeline;\n  bindGroup: GPUBindGroup;\n  uniformBuffer: GPUBuffer;\n}\n\nexport interface EncoderResult {\n  /** encoder_out [S_enc, hidden] (debug / parity). */\n  encoderOut: Float32Array;\n  /** Per-decoder-layer frozen K, indexed by layer. */\n  encK: Float32Array[];\n  /** Per-decoder-layer frozen V, indexed by layer. */\n  encV: Float32Array[];\n  /** Encoder length (frames). */\n  sEnc: number;\n}\n\nexport class MoonshineEncoderExecutor {\n  private ctx: GPUContext;\n  private graph: ModelGraph;\n  private decLayers: number;\n  private hidden: number;\n\n  private weightBuffers = new Map<string, GPUBuffer>();\n  private activationBuffers = new Map<string, GPUBuffer>();\n  private dispatches: MoonshineDispatch[] = [];\n\n  constructor(ctx: GPUContext, graph: ModelGraph, decLayers: number) {\n    this.ctx = ctx;\n    this.graph = graph;\n    this.decLayers = decLayers;\n    this.hidden = graph.config.hidden_size;\n    this.allocateActivationBuffers();\n  }\n\n  /**\n   * Upload the encoder constants. `weights` holds canonical-named f32 tensors;\n   * only those referenced by the graph are uploaded (the decoder weights are\n   * uploaded into the decoder Executor separately).\n   */\n  uploadWeights(weights: Map<string, { data: ArrayBufferView; shape: number[] }>): void {\n    for (const [name, desc] of Object.entries(this.graph.tensors)) {\n      if (desc.storage !== \"constant\") continue;\n      const w = weights.get(name);\n      if (w) {\n        const buffer = createStorageBuffer(this.ctx, `menc_${name}`, w.data.byteLength, w.data);\n        this.weightBuffers.set(name, buffer);\n      } else if (!desc.safetensorsKey) {\n        // Synthetic constant with no checkpoint source (e.g. conv1's zero bias —\n        // Conv1dFull requires a bias binding but Moonshine's conv1 has none).\n        // Allocate it zero-filled.\n        const count = (desc.shape as number[]).reduce((a, b) => a * (b as number), 1);\n        const zeros = new Float32Array(count);\n        const buffer = createStorageBuffer(this.ctx, `menc_${name}`, zeros.byteLength, zeros);\n        this.weightBuffers.set(name, buffer);\n      } else {\n        throw new Error(`MoonshineEncoderExecutor: missing weight \"${name}\"`);\n      }\n    }\n  }\n\n  initBindGroups(): void {\n    const shapes = this.resolveShapes();\n    for (const nodeId of this.graph.executionOrder) {\n      const node = this.graph.nodes.find((n) => n.id === nodeId)!;\n      let spec = KERNEL_REGISTRY[node.opType];\n      if (!spec) throw new Error(`MoonshineEncoderExecutor: no kernel for op \"${node.opType}\"`);\n      // Moonshine norms are weight-only — pick the no-bias LayerNorm variant.\n      if (node.opType === \"LayerNorm\" && node.inputs.length === 2) {\n        spec = LAYERNORM_NO_BIAS_SPEC;\n      } else if (node.opType === \"RoPE\" && node.attributes.interleaved === true) {\n        spec = ROPE_INTERLEAVED_SPEC;\n      }\n      const pipeline = getOrCreatePipeline(\n        this.ctx,\n        `mkernel_${nodeId}`,\n        spec.shaderCode,\n        spec.entryPoint,\n      );\n      const paramsData = spec.buildParams(node, shapes, { seqPos: 0 });\n      const uniformBuffer = createUniformBuffer(this.ctx, `muniform_${nodeId}`, paramsData);\n      const bufferEntries = this.gatherBuffers(spec, node, uniformBuffer);\n      const bindGroup = createBindGroup(this.ctx, pipeline, bufferEntries, `mbg_${nodeId}`);\n      this.dispatches.push({ node, spec, pipeline, bindGroup, uniformBuffer });\n    }\n  }\n\n  /** Run the conv frontend + encoder + K/V projection. `pcm` is raw 16kHz mono. */\n  async encode(pcm: Float32Array): Promise<EncoderResult> {\n    const q = this.ctx.device.queue;\n    q.writeBuffer(this.activationBuffers.get(\"pcm\")!, 0, pcm as BufferSource);\n\n    const shapes = this.resolveShapes();\n    const ctxRt = { seqPos: 0 };\n    const sizes: Array<[number, number, number]> = [];\n    for (const d of this.dispatches) {\n      const params = d.spec.buildParams(d.node, shapes, ctxRt);\n      q.writeBuffer(d.uniformBuffer, 0, params);\n      sizes.push(d.spec.getDispatchSize(d.node, shapes, ctxRt));\n    }\n\n    if (this.ctx.isWebKitWebGPU) {\n      for (let i = 0; i < this.dispatches.length; i++) {\n        const d = this.dispatches[i];\n        if (d.node.opType === \"Attention\") {\n          // Window the O(S²) bidirectional attention by query-row (watchdog guard).\n          const [gridX, gridY, gridZ] = sizes[i];\n          for (let start = 0; start < gridX; start += WEBKIT_ATTN_CHUNK_ROWS) {\n            const rows = Math.min(WEBKIT_ATTN_CHUNK_ROWS, gridX - start);\n            const params = d.spec.buildParams(d.node, shapes, { seqPos: 0, qOffset: start });\n            q.writeBuffer(d.uniformBuffer, 0, params);\n            const enc = this.ctx.device.createCommandEncoder();\n            const p = enc.beginComputePass();\n            p.setPipeline(d.pipeline);\n            p.setBindGroup(0, d.bindGroup);\n            p.dispatchWorkgroups(rows, gridY, gridZ);\n            p.end();\n            q.submit([enc.finish()]);\n            await q.onSubmittedWorkDone();\n          }\n          continue;\n        }\n        const enc = this.ctx.device.createCommandEncoder();\n        const p = enc.beginComputePass();\n        p.setPipeline(d.pipeline);\n        p.setBindGroup(0, d.bindGroup);\n        p.dispatchWorkgroups(...sizes[i]);\n        p.end();\n        q.submit([enc.finish()]);\n        await q.onSubmittedWorkDone();\n      }\n    } else {\n      const enc = this.ctx.device.createCommandEncoder({ label: \"moonshine_encode\" });\n      const pass = enc.beginComputePass({ label: \"moonshine_enc_pass\" });\n      for (let i = 0; i < this.dispatches.length; i++) {\n        pass.setPipeline(this.dispatches[i].pipeline);\n        pass.setBindGroup(0, this.dispatches[i].bindGroup);\n        pass.dispatchWorkgroups(...sizes[i]);\n      }\n      pass.end();\n      q.submit([enc.finish()]);\n    }\n\n    const sEnc = this.encoderFrames();\n    const encoderOut = await this.readBack(\"encoder_out\", sEnc * this.hidden);\n    const encK: Float32Array[] = [];\n    const encV: Float32Array[] = [];\n    for (let i = 0; i < this.decLayers; i++) {\n      encK.push(await this.readBack(`enc_k_layer${i}`, sEnc * this.hidden));\n      encV.push(await this.readBack(`enc_v_layer${i}`, sEnc * this.hidden));\n    }\n    return { encoderOut, encK, encV, sEnc };\n  }\n\n  /** Read back a named activation buffer after encode() (debug / parity checks). */\n  async readActivation(name: string, maxElements = 4096): Promise<Float32Array> {\n    return this.readBack(name, maxElements);\n  }\n\n  destroy(): void {\n    destroyBuffers([...this.weightBuffers.values()]);\n    destroyBuffers([...this.activationBuffers.values()]);\n    for (const d of this.dispatches) d.uniformBuffer.destroy();\n    this.weightBuffers.clear();\n    this.activationBuffers.clear();\n    this.dispatches = [];\n  }\n\n  // ── Private ──\n\n  /** Encoder frame count = first dim of encoder_out (resolved, numeric). */\n  private encoderFrames(): number {\n    return this.graph.tensors.encoder_out.shape[0] as number;\n  }\n\n  private resolveShapes(): Record<string, number[]> {\n    // The Moonshine encoder graph is fully concrete (no symbolic dims) — every\n    // tensor shape is numeric for the chosen sample count.\n    const resolved: Record<string, number[]> = {};\n    for (const [name, desc] of Object.entries(this.graph.tensors)) {\n      resolved[name] = desc.shape.map((dim) => dim as number);\n    }\n    return resolved;\n  }\n\n  private allocateActivationBuffers(): void {\n    for (const [name, desc] of Object.entries(this.graph.tensors)) {\n      if (desc.storage !== \"activation\") continue;\n      const bytes =\n        (desc.shape as number[]).reduce((a, b) => a * (b as number), 1) * DTYPE_BYTES[desc.dtype];\n      this.activationBuffers.set(name, createStorageBuffer(this.ctx, `mact_${name}`, bytes));\n    }\n  }\n\n  private async readBack(name: string, elements: number): Promise<Float32Array> {\n    const buffer = this.activationBuffers.get(name);\n    if (!buffer) throw new Error(`MoonshineEncoderExecutor: no buffer \"${name}\"`);\n    const byteLen = Math.min(elements * 4, buffer.size);\n    const readback = this.ctx.device.createBuffer({ size: byteLen, usage: 0x0001 | 0x0008 });\n    const enc = this.ctx.device.createCommandEncoder();\n    enc.copyBufferToBuffer(buffer, 0, readback, 0, byteLen);\n    this.ctx.device.queue.submit([enc.finish()]);\n    await readback.mapAsync(MAP_MODE_READ, 0, byteLen);\n    const data = new Float32Array(readback.getMappedRange(0, byteLen).slice(0));\n    readback.unmap();\n    readback.destroy();\n    return data;\n  }\n\n  private gatherBuffers(\n    spec: KernelSpec,\n    node: OpNode,\n    uniformBuffer: GPUBuffer,\n  ): Array<{ buffer: GPUBuffer }> {\n    const entries: Array<{ buffer: GPUBuffer }> = [];\n    const tensorNames = [...node.inputs, ...node.outputs];\n    let tensorIdx = 0;\n    for (const binding of spec.bindings) {\n      if (binding.type === \"uniform\") {\n        entries.push({ buffer: uniformBuffer });\n      } else {\n        const tensorName = tensorNames[tensorIdx++];\n        const buffer = this.weightBuffers.get(tensorName) ?? this.activationBuffers.get(tensorName);\n        if (!buffer) {\n          throw new Error(\n            `MoonshineEncoderExecutor: no buffer for \"${tensorName}\" in op ${node.id}`,\n          );\n        }\n        entries.push({ buffer });\n      }\n    }\n    return entries;\n  }\n}\n","/**\n * MoonshineSTT — native speech-to-text engine for Gerbil's WebGPU backend.\n *\n * Moonshine is an encoder-decoder ASR model with a raw-waveform conv frontend\n * (no FFT / mel spectrogram). Unlike the causal-LM path, it needs two graphs:\n *\n *   1. ENCODER (run once per utterance): conv frontend → bidirectional transformer\n *      → encoder hidden state, which is then projected through every decoder\n *      layer's cross-attention k_proj/v_proj into FROZEN K/V buffers.\n *   2. DECODER (autoregressive): causal self-attention with a growing KV-cache,\n *      plus cross-attention into the frozen encoder K/V at every step.\n *\n * The conv frontend is length-static (Conv1dFull carries concrete L/Lout), so the\n * encoder graph is regenerated per utterance from the input sample count, and the\n * decoder graph is regenerated with the resulting encoder frame count (S_enc).\n * Weights are uploaded once and reused across utterances via per-call executors.\n *\n * Validated on Dawn (desktop) via scripts/engine/test-moonshine-transcribe.mjs.\n * The kernels are mobile-safe (≤16KB workgroup memory, clamped exp/tanh, no\n * select(), no `enable f16`) and the executors use the WebKit submit/drain\n * discipline, so the same path runs on iPad.\n */\n\nimport {\n  generateMoonshineDecoderGraph,\n  generateMoonshineEncoderGraph,\n  moonshineEncoderFrames,\n  parseMoonshineConfig,\n} from \"./architectures/moonshine.js\";\nimport type { GPUContext } from \"./device.js\";\nimport { initGPU } from \"./device.js\";\nimport { Executor } from \"./executor.js\";\nimport { loadMoonshine } from \"./model-loader.js\";\nimport { MoonshineEncoderExecutor } from \"./moonshine-executor.js\";\nimport type { Tokenizer } from \"./tokenizer.js\";\n\nconst MOONSHINE_SAMPLE_RATE = 16_000;\n/** Hard cap so a stuck decode can't loop forever (config max_position_embeddings). */\nconst DEFAULT_MAX_NEW_TOKENS = 194;\n\nexport interface MoonshineSTTOptions {\n  /** HF repo (default UsefulSensors/moonshine-base). */\n  repo?: string;\n  revision?: string;\n  hfToken?: string;\n  cacheDir?: string;\n  onProgress?: (loaded: number, total: number, message: string) => void;\n}\n\nexport interface TranscribeOptions {\n  /** Stop after this many decoded tokens (default 194). */\n  maxNewTokens?: number;\n  /**\n   * RMS energy below which a clip is treated as silence/noise (`noSpeech`).\n   * Default 0.01 (normalized 16 kHz float PCM).\n   */\n  minRms?: number;\n  /**\n   * Clips shorter than this are treated as no-speech — a sub-threshold clip is\n   * almost always silence, and any transcript on it is a hallucination.\n   * Default 0.35 seconds.\n   */\n  minSpeechSeconds?: number;\n}\n\nexport interface TranscribeResult {\n  text: string;\n  /** Decoded token ids (excluding the start token, including the trailing EOS). */\n  tokens: number[];\n  /** Number of encoder frames produced by the conv frontend. */\n  encoderFrames: number;\n  /** Audio duration in seconds (samples / 16000). */\n  audioSeconds: number;\n  /** RMS energy of the input PCM (~0..1). */\n  speechRms: number;\n  /**\n   * True when the clip almost certainly contains no speech — too quiet, too\n   * short, an empty transcript, or a non-speech marker like \"[BLANK_AUDIO]\".\n   * Skip empty/hallucinated transcripts when this is set.\n   */\n  noSpeech: boolean;\n}\n\n/** Non-speech markers Moonshine can emit on silence, e.g. \"[BLANK_AUDIO]\", \"(silence)\". */\nconst NON_SPEECH_RE = /^\\s*[[(][^)\\]]*[)\\]]\\s*$/;\n\nfunction computeRms(pcm: Float32Array): number {\n  if (pcm.length === 0) {\n    return 0;\n  }\n  let sum = 0;\n  for (let i = 0; i < pcm.length; i++) {\n    sum += pcm[i] * pcm[i];\n  }\n  return Math.sqrt(sum / pcm.length);\n}\n\nexport class MoonshineSTT {\n  private ctx: GPUContext;\n  private weights: Map<string, { data: ArrayBufferView; shape: number[] }>;\n  private tokenizer: Tokenizer;\n  private rawConfig: Record<string, unknown>;\n  private bosTokenId: number;\n  private eosTokenId: number;\n  private decoderStartTokenId: number;\n  private _destroyed = false;\n\n  /** HF architecture string, for parity with WebGPUEngine. */\n  readonly architecture = \"MoonshineForConditionalGeneration\";\n\n  private constructor(\n    ctx: GPUContext,\n    weights: Map<string, { data: ArrayBufferView; shape: number[] }>,\n    tokenizer: Tokenizer,\n    rawConfig: Record<string, unknown>,\n  ) {\n    this.ctx = ctx;\n    this.weights = weights;\n    this.tokenizer = tokenizer;\n    this.rawConfig = rawConfig;\n    this.bosTokenId = (rawConfig.bos_token_id as number) ?? 1;\n    this.eosTokenId = (rawConfig.eos_token_id as number) ?? 2;\n    // Moonshine seeds the decoder with decoder_start_token_id (== bos == 1).\n    this.decoderStartTokenId = (rawConfig.decoder_start_token_id as number) ?? this.bosTokenId;\n  }\n\n  /** Download + initialize a Moonshine STT engine. */\n  static async create(options: MoonshineSTTOptions = {}): Promise<MoonshineSTT> {\n    const ctx = await initGPU();\n    const repo = options.repo ?? \"UsefulSensors/moonshine-base\";\n    const { weights, tokenizer, rawConfig } = await loadMoonshine({\n      repo,\n      revision: options.revision,\n      hfToken: options.hfToken,\n      cacheDir: options.cacheDir,\n      onProgress: options.onProgress,\n    });\n    return new MoonshineSTT(ctx, weights, tokenizer, rawConfig);\n  }\n\n  /**\n   * Transcribe raw 16 kHz mono PCM. Runs the conv frontend + encoder once, then\n   * greedily AR-decodes with cross-attention into the frozen encoder K/V, stopping\n   * on EOS. Returns the detokenized transcript.\n   */\n  async transcribe(pcm: Float32Array, opts: TranscribeOptions = {}): Promise<TranscribeResult> {\n    if (this._destroyed) throw new Error(\"MoonshineSTT has been destroyed.\");\n    if (pcm.length < 127) {\n      throw new Error(`PCM too short: ${pcm.length} samples (need >= conv1 kernel size 127).`);\n    }\n    const d = parseMoonshineConfig(this.rawConfig);\n    const sEnc = moonshineEncoderFrames(pcm.length);\n    if (sEnc < 1) {\n      throw new Error(`Audio too short: produces ${sEnc} encoder frames.`);\n    }\n\n    const audioSeconds = pcm.length / MOONSHINE_SAMPLE_RATE;\n    const speechRms = computeRms(pcm);\n    const minRms = opts.minRms ?? 0.01;\n    const minSpeechSeconds = opts.minSpeechSeconds ?? 0.35;\n\n    // Cheap VAD: skip the encode/decode entirely for clips too quiet or too short\n    // to contain speech (a sub-threshold clip is almost always silence; any\n    // transcript on it is a hallucination).\n    if (speechRms < minRms || audioSeconds < minSpeechSeconds) {\n      return {\n        text: \"\",\n        tokens: [],\n        encoderFrames: 0,\n        audioSeconds,\n        speechRms,\n        noSpeech: true,\n      };\n    }\n\n    // ── Encoder (once) ──\n    const encoderGraph = generateMoonshineEncoderGraph(this.rawConfig, pcm.length);\n    const encExec = new MoonshineEncoderExecutor(this.ctx, encoderGraph, d.dec_layers);\n    let encK: Float32Array[];\n    let encV: Float32Array[];\n    let frames: number;\n    try {\n      encExec.uploadWeights(this.weights);\n      encExec.initBindGroups();\n      const enc = await encExec.encode(pcm);\n      encK = enc.encK;\n      encV = enc.encV;\n      frames = enc.sEnc;\n    } finally {\n      encExec.destroy();\n    }\n\n    // ── Decoder (autoregressive, greedy) ──\n    const decoderGraph = generateMoonshineDecoderGraph(this.rawConfig, frames);\n    // maxSeqLen bounds the self-attn KV cache. The decode can never exceed the\n    // model's max_position_embeddings, so cap there.\n    const maxSeqLen = Math.min(opts.maxNewTokens ?? DEFAULT_MAX_NEW_TOKENS, d.context_length);\n    const decExec = new Executor(this.ctx, decoderGraph, { maxSeqLen, kvMode: \"f32\" });\n    const tokens: number[] = [];\n    try {\n      decExec.uploadWeightsMap(selectGraphWeights(decoderGraph, this.weights));\n      decExec.initBindGroups();\n      // Bind the frozen encoder K/V into each decoder layer's cross-attention.\n      for (let i = 0; i < d.dec_layers; i++) {\n        decExec.writeInput(`enc_k_layer${i}`, encK[i]);\n        decExec.writeInput(`enc_v_layer${i}`, encV[i]);\n      }\n      decExec.reset();\n\n      // Greedy decode: seed with the decoder start token, feed argmax back in.\n      let nextToken = this.decoderStartTokenId;\n      const limit = maxSeqLen;\n      for (let step = 0; step < limit; step++) {\n        const tok = await decExec.forwardArgmax(new Uint32Array([nextToken]));\n        tokens.push(tok);\n        if (tok === this.eosTokenId) break;\n        nextToken = tok;\n      }\n    } finally {\n      decExec.destroy();\n    }\n\n    const text = this.tokenizer.decode(tokens, /* skipSpecialTokens */ true).trim();\n    const noSpeech = text.length === 0 || NON_SPEECH_RE.test(text);\n    return {\n      text,\n      tokens,\n      encoderFrames: frames,\n      audioSeconds,\n      speechRms,\n      noSpeech,\n    };\n  }\n\n  destroy(): void {\n    this._destroyed = true;\n    this.weights.clear();\n  }\n}\n\n/**\n * Build a fresh map of just the constant weights a graph references.\n *\n * Executor.uploadWeights() DELETES entries from the map it is given (to free JS\n * memory as each weight reaches the GPU). Moonshine reuses the same weights across\n * utterances and across the encoder+decoder, so we hand each executor its OWN map\n * (deletes only touch the copy; the shared ArrayBufferViews are uploaded, never\n * mutated). Scoping to graph-referenced constants also avoids uploading the\n * encoder's weights into the decoder executor.\n */\nfunction selectGraphWeights(\n  graph: { tensors: Record<string, { storage: string }> },\n  weights: Map<string, { data: ArrayBufferView; shape: number[] }>,\n): Map<string, { data: ArrayBufferView; shape: number[] }> {\n  const out = new Map<string, { data: ArrayBufferView; shape: number[] }>();\n  for (const [name, desc] of Object.entries(graph.tensors)) {\n    if (desc.storage !== \"constant\") continue;\n    const w = weights.get(name);\n    if (w) out.set(name, w);\n  }\n  return out;\n}\n"],"mappings":";;;AASA,MAAM,WAAW;AACjB,MAAM,WAAW;AAEjB,MAAM,WAAW;AACjB,MAAMA,kBAAgB;;;;;;;;;AAsCtB,eAAsB,QAAQ,SAA+C;AAE3E,MAAK,OAAO,cAAc,eAAe,CAAC,UAAU,QAAQ,OAAO,YAAY,YAC7E,KAAI;EAGF,MAAM,EAAE,WADa,MADC,IAAI,SAAS,aAAa,2BAA2B,CAClC,SAAS;AAElD,MAAI,CAAE,WAAmB,UACvB,CAAC,WAAmB,YAAY,EAAE;AAEpC,EAAC,WAAmB,UAAU,MAAM,OAAO,EAAE,CAAC;SACxC;AAKV,KAAI,OAAO,cAAc,eAAe,CAAC,UAAU,IACjD,OAAM,IAAI,MACR,4GAED;CAIH,MAAM,WACJ,OAAO,cAAc,eAAe,4BAA4B,KAAK,UAAU,UAAU;CAE3F,MAAM,UAAU,MAAM,UAAU,IAAI,eAAe,EACjD,iBAAiB,WAAW,cAAc,oBAC3C,CAAC;AAEF,KAAI,CAAC,QACH,OAAM,IAAI,MACR,oHAED;CAIH,MAAM,SAAS,QAAQ,SAAS,IAAI,aAAa;CAIjD,MAAM,eAAe,QAAQ,SAAS,IAAI,YAA8B;CAKxE,MAAM,eADgB,OAAO,YAAY,eAAe,QAAQ,KAAK,kBAAkB,QACjD,QAAQ,SAAS,IAAI,kBAAkB;CAG7E,MAAMC,mBAAqC,EAAE;AAC7C,KAAI,OACF,kBAAiB,KAAK,aAAa;AAErC,KAAI,aACF,kBAAiB,KAAK,YAA8B;AAEtD,KAAI,aACF,kBAAiB,KAAK,kBAAkB;CAI1C,MAAMC,gBAAwC;EAC5C,eAAe,QAAQ,OAAO;EAC9B,6BAA6B,QAAQ,OAAO;EAC5C,0BAA0B,KAAK,IAAI,KAAK,QAAQ,OAAO,yBAAyB;EAChF,0BAA0B,KAAK,IAAI,KAAK,QAAQ,OAAO,yBAAyB;EAChF,gCAAgC,QAAQ,OAAO;EAC/C,iCAAiC,QAAQ,OAAO;EACjD;CAED,IAAIC;AACJ,KAAI;AACF,WAAS,MAAM,QAAQ,cAAc;GACnC;GACA,gBAAgB;GACjB,CAAC;UACK,GAAG;AAEV,MAAI;AACF,YAAS,MAAM,QAAQ,cAAc,EAAE,kBAAkB,CAAC;UACpD;AACN,SAAM,IAAI,MACR,mCAAmC,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,CAAC,mDAE/E;;;AAKL,QAAO,KAAK,MAAM,SAAS;AACzB,UAAQ,MAAM,gCAAgC,KAAK,UAAU,KAAK,QAAQ;AAC1E,WAAS,eAAe,KAAK,QAAQ,KAAK,QAAQ;GAClD;AAKF,KAAI;AACF,SAAO,qBAAqB,OAAgC;AAC1D,WAAQ,MAAM,kCAAkC,GAAG,OAAO,WAAW,GAAG,QAAQ;;SAE5E;CASR,IAAI,qBAAqB;AACzB,KAAI;EACF,MAAM,OAAQ,QAAgB,QAAS,QAAgB,WAAW;AAClE,MAAI,KACF,sBAAqB,UAAU,KAAK,UAAU,GAAG,QAAQ,KAAK,gBAAgB,GAAG,UAAU,KAAK,UAAU,GAAG,QAAQ,KAAK,eAAe;SAErI;CAMR,MAAM,KAAK,OAAO,cAAc,eAAe,UAAU,YAAY,UAAU,YAAY;CAC3F,MAAM,gBAAgB,GAAG,SAAS,cAAc;CAChD,MAAM,gBACJ,mBAAmB,KAAK,GAAG,IAC1B,iBAAiB,OAAO,cAAc,gBAAgB,UAAU,kBAAkB,KAAK;CAC1F,MAAM,cAAc,iBAAiB,CAAC,GAAG,SAAS,UAAU,IAAI,CAAC;CACjE,MAAM,iBAAiB,iBAAiB;AACxC,KAAI,CAAC,mBACH,sBAAqB,gBAAgB,GAAG,MAAM,GAAG,IAAI;AAGvD,QAAO;EACL;EACA,QAAQ,OAAO;EACf;EACA;EACA;EACA;EACA;EACD;;;;;AAQH,SAAgB,oBACd,KACA,OACA,WACA,MACW;CACX,MAAM,QAAkB;CAGxB,MAAM,cAAc,KAAK,KAAK,YAAY,EAAE,GAAG;CAE/C,MAAM,SAAS,IAAI,OAAO,aAAa;EACrC;EACA,MAAM;EACN;EACD,CAAC;AAIF,KAAI,KACF,KAAI,gBAAgB,YAClB,KAAI,OAAO,MAAM,YAAY,QAAQ,GAAG,KAAK;KAE7C,KAAI,OAAO,MAAM,YAAY,QAAQ,GAAG,KAAK,QAAQ,KAAK,YAAY,KAAK,WAAW;AAI1F,QAAO;;;;;AAMT,SAAgB,oBACd,KACA,OACA,MACW;CACX,MAAM,aAAa,gBAAgB,cAAc,KAAK,aAAa,KAAK;CAKxE,MAAM,cAAc,KAAK,KAAK,aAAa,GAAG,GAAG;CAEjD,MAAM,SAAS,IAAI,OAAO,aAAa;EACrC;EACA,MAAM;EACN,OAAiB;EAClB,CAAC;AAEF,KAAI,gBAAgB,YAClB,KAAI,OAAO,MAAM,YAAY,QAAQ,GAAG,KAAK;KAE7C,KAAI,OAAO,MAAM,YAAY,QAAQ,GAAG,KAAK,QAAQ,KAAK,YAAY,KAAK,WAAW;AAGxF,QAAO;;;;;AAMT,SAAgB,qBAAqB,KAAiB,OAAe,WAA8B;CACjG,MAAM,cAAc,KAAK,KAAK,YAAY,EAAE,GAAG;AAC/C,QAAO,IAAI,OAAO,aAAa;EAC7B;EACA,MAAM;EACN,OAAO,WAAW;EACnB,CAAC;;;;;;;AAyBJ,MAAM,iCAAiB,IAAI,SAAqD;;AAGhF,SAAS,iBAAiB,QAAoD;CAC5E,IAAI,QAAQ,eAAe,IAAI,OAAO;AACtC,KAAI,CAAC,OAAO;AACV,0BAAQ,IAAI,KAAK;AACjB,iBAAe,IAAI,QAAQ,MAAM;;AAEnC,QAAO;;;;;AAMT,SAAgB,oBACd,KACA,OACA,UACA,aAAqB,QACrB,wBACoB;CACpB,MAAM,gBAAgB,iBAAiB,IAAI,OAAO;CAClD,MAAM,WAAW,GAAG,SAAS,IAAI;CACjC,MAAM,SAAS,cAAc,IAAI,SAAS;AAC1C,KAAI,OAAQ,QAAO;CAEnB,IAAI,OAAO;CAKX,MAAM,UAAU,KAAK,SAAS,QAAQ,IAAI,KAAK,SAAS,QAAQ;AAChE,KAAI,IAAI,UAAU,WAAW,CAAC,KAAK,SAAS,aAAa,CACvD,QAAO,kBAAkB;CAG3B,MAAM,eAAe,IAAI,OAAO,mBAAmB;EACjD,OAAO,GAAG,MAAM;EAChB;EACD,CAAC;CAEF,IAAIC,SAAqC;AACzC,KAAI,wBAAwB;EAC1B,MAAM,kBAAkB,IAAI,OAAO,sBAAsB;GACvD,OAAO,GAAG,MAAM;GAChB,SAAS;GACV,CAAC;AACF,WAAS,IAAI,OAAO,qBAAqB;GACvC,OAAO,GAAG,MAAM;GAChB,kBAAkB,CAAC,gBAAgB;GACpC,CAAC;;CAGJ,MAAM,WAAW,IAAI,OAAO,sBAAsB;EAChD;EACA;EACA,SAAS;GACP,QAAQ;GACR;GACD;EACF,CAAC;AAEF,eAAc,IAAI,UAAU,SAAS;AACrC,QAAO;;;;;;AAOT,SAAgB,mBAAmB,QAA0B;AAC3D,KAAI,OACF,gBAAe,OAAO,OAAO;;;;;AAejC,SAAgB,gBACd,KACA,UACA,SACA,OACc;AACd,QAAO,IAAI,OAAO,gBAAgB;EAChC;EACA,QAAQ,SAAS,mBAAmB,EAAE;EACtC,SAAS,QAAQ,KAAK,OAAO,OAAO;GAClC,SAAS;GACT,UAAU;IACR,QAAQ,MAAM;IACd,QAAQ,MAAM,UAAU;IACxB,MAAM,MAAM,QAAQ,MAAM,OAAO;IAClC;GACF,EAAE;EACJ,CAAC;;;;;;AASJ,SAAgB,aACd,KACA,UACA,WACA,iBACA,kBAA0B,GAC1B,kBAA0B,GACpB;CACN,MAAM,UAAU,IAAI,OAAO,sBAAsB;CACjD,MAAM,OAAO,QAAQ,kBAAkB;AACvC,MAAK,YAAY,SAAS;AAC1B,MAAK,aAAa,GAAG,UAAU;AAC/B,MAAK,mBAAmB,iBAAiB,iBAAiB,gBAAgB;AAC1E,MAAK,KAAK;AACV,KAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;;;;;;;;AAW7C,eAAsB,eACpB,KACA,QACA,UACA,aAAqB,GACrB,YACuB;CACvB,MAAM,SAAS,cAAc,OAAO,OAAO;CAE3C,MAAM,UAAU,IAAI,OAAO,sBAAsB;AACjD,SAAQ,mBAAmB,QAAQ,YAAY,UAAU,GAAG,OAAO;AACnE,KAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;AAE3C,OAAM,SAAS,SAASJ,iBAAe,GAAG,OAAO;CACjD,MAAM,SAAS,SAAS,eAAe,GAAG,OAAO;CACjD,MAAM,SAAS,IAAI,aAAa,OAAO,MAAM,EAAE,CAAC;AAChD,UAAS,OAAO;AAEhB,QAAO;;;;;AAgCT,SAAgB,eAAe,SAA4B;AACzD,MAAK,MAAM,OAAO,QAChB,KAAI,SAAS;;;;;;AAqBjB,eAAsB,UAAU,KAA+C;CAC7E,MAAMK,UAAoB,EAAE;CAC5B,IAAI,kBAAkB;CACtB,IAAI,eAAe;CACnB,IAAI,oBAAoB;AAGxB,KAAI;AACF,UAAQ,KACN,kBAAkB,IAAI,OAAO,yBAAyB,cACvC,IAAI,OAAO,+BAA+B,YAC5C,IAAI,OAAO,gBAAgB,OAAO,MAAM,QAAQ,EAAE,CAAC,UACvD,IAAI,OAAO,cAAc,IAAI,eACvC;SACK;AAGR,KAAI;EACF,MAAM,WAAW,IAAI,aAAa;GAAC;GAAK;GAAK;GAAK;GAAK;GAAM;GAAK;GAAO;GAAM,CAAC;EAChF,MAAM,SAAS,oBAAoB,KAAK,YAAY,SAAS,YAAY,SAAS;EAClF,MAAM,UAAU,qBAAqB,KAAK,aAAa,SAAS,WAAW;EAC3E,MAAM,SAAS,MAAM,eAAe,KAAK,QAAQ,SAAS,GAAG,SAAS,WAAW;EAEjF,IAAI,QAAQ;AACZ,OAAK,IAAI,IAAI,GAAG,IAAI,SAAS,QAAQ,IACnC,KAAI,KAAK,IAAI,OAAO,KAAK,SAAS,GAAG,GAAG,MAAM;AAC5C,WAAQ,KAAK,uBAAuB,EAAE,cAAc,SAAS,GAAG,QAAQ,OAAO,KAAK;AACpF,WAAQ;;AAGZ,oBAAkB;AAClB,UAAQ,KACN,kBAAkB,gCAAgC,kCACnD;AACD,SAAO,SAAS;AAChB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,wBAAwB,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;;CAgBpF,MAAM,mBAAmB,OACvB,OACA,MACA,UACA,SACA,QACA,WACA,aAAqB,MACK;EAC1B,MAAM,WAAW,oBAAoB,KAAK,OAAO,MAAM,OAAO;EAC9D,MAAM,KAAK,gBACT,KACA,UACA,SAAS,KAAK,OAAO,EAAE,QAAQ,GAAG,EAAE,EACpC,GAAG,MAAM,KACV;EACD,MAAM,UAAU,IAAI,OAAO,sBAAsB;EACjD,MAAM,OAAO,QAAQ,kBAAkB;AACvC,OAAK,YAAY,SAAS;AAC1B,OAAK,aAAa,GAAG,GAAG;AACxB,OAAK,mBAAmB,WAAW;AACnC,OAAK,KAAK;AACV,UAAQ,mBAAmB,QAAQ,GAAG,SAAS,GAAG,UAAU;AAC5D,MAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;AAC3C,MAAI,IAAI,OAAO,MAAM,oBACnB,OAAM,IAAI,OAAO,MAAM,qBAAqB;AAE9C,QAAM,QAAQ,SAASL,iBAAe,GAAG,UAAU;EACnD,MAAM,SAAS,QAAQ,eAAe,GAAG,UAAU;EACnD,MAAM,SAAS,IAAI,aAAa,OAAO,MAAM,EAAE,CAAC;AAChD,UAAQ,OAAO;AACf,SAAO;;;CAIT,MAAM,sBAAsB,OAC1B,OACA,MACA,UACA,SACA,QACA,WACA,aAAqB,MACK;EAC1B,MAAM,WAAW,oBAAoB,KAAK,OAAO,MAAM,OAAO;AAQ9D,eAAa,KAAK,UAPP,gBACT,KACA,UACA,SAAS,KAAK,OAAO,EAAE,QAAQ,GAAG,EAAE,EACpC,GAAG,MAAM,KACV,EAE+B,WAAW;AAE3C,SAAO,eAAe,KAAK,QAAQ,SAAS,GAAG,UAAU;;AAI3D,KAAI;EACF,MAAM,SAAS;;;;;EAKf,MAAM,SAAS,oBAAoB,KAAK,UAAU,EAAE;EACpD,MAAM,UAAU,qBAAqB,KAAK,WAAW,EAAE;EACvD,MAAM,SAAS,MAAM,oBAAoB,MAAM,QAAQ,CAAC,OAAO,EAAE,SAAS,QAAQ,EAAE;EACpF,MAAM,KAAK,KAAK,IAAI,OAAO,KAAK,GAAK,GAAG;AACxC,UAAQ,KACN,KACI,8BAA8B,OAAO,GAAG,KACxC,sCAAsC,OAAO,GAAG,GACrD;AACD,SAAO,SAAS;AAChB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAI1E,KAAI;EACF,MAAM,SAAS;;;;;;EAOf,MAAM,QAAQ,oBAAoB,KAAK,SAAS,GADlC,IAAI,aAAa,CAAC,EAAI,CAAC,CACoB;EACzD,MAAM,SAAS,oBAAoB,KAAK,UAAU,EAAE;EACpD,MAAM,UAAU,qBAAqB,KAAK,WAAW,EAAE;EACvD,MAAM,SAAS,MAAM,iBAAiB,MAAM,QAAQ,CAAC,OAAO,OAAO,EAAE,SAAS,QAAQ,EAAE;EACxF,MAAM,KAAK,KAAK,IAAI,OAAO,KAAK,EAAI,GAAG;AACvC,UAAQ,KACN,KAAK,0BAA0B,OAAO,GAAG,KAAK,kCAAkC,OAAO,GAAG,GAC3F;AACD,iBAAe;AACf,QAAM,SAAS;AACf,SAAO,SAAS;AAChB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAI1E,KAAI;EACF,MAAM,SAAS;;;;;;;EAOf,MAAM,SAAS,oBAAoB,KAAK,UAAU,GAAG;EACrD,MAAM,UAAU,qBAAqB,KAAK,WAAW,GAAG;EACxD,MAAM,SAAS,MAAM,iBAAiB,MAAM,QAAQ,CAAC,OAAO,EAAE,SAAS,QAAQ,GAAG;EAClF,MAAM,KAAK,KAAK,IAAI,OAAO,KAAK,EAAI,GAAG,QAAQ,KAAK,IAAI,OAAO,KAAK,EAAI,GAAG;AAC3E,UAAQ,KACN,KACI,iBAAiB,MAAM,KAAK,OAAO,CAAC,KAAK,IAAI,CAAC,KAC9C,qBAAqB,MAAM,KAAK,OAAO,CAAC,KAAK,IAAI,CAAC,GACvD;AACD,SAAO,SAAS;AAChB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAM1E,KAAI;EACF,MAAM,SAAS;;;;;;;;;;;;EAYf,MAAM,SAAS,oBAAoB,KAAK,UAAU,EAAE;EACpD,MAAM,UAAU,qBAAqB,KAAK,WAAW,EAAE;EACvD,MAAM,SAAS,MAAM,iBAAiB,MAAM,QAAQ,CAAC,OAAO,EAAE,SAAS,QAAQ,EAAE;AACjF,sBAAoB,KAAK,IAAI,OAAO,KAAK,GAAK,GAAG;AACjD,UAAQ,KACN,oBACI,uBAAuB,OAAO,GAAG,KACjC,+BAA+B,OAAO,GAAG,GAC9C;AACD,SAAO,SAAS;AAChB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAI1E,KAAI;EACF,MAAM,SAAS;;;;;;;;EASf,MAAM,QAAQ,oBAAoB,KAAK,SAAS,IADlC,IAAI,aAAa;GAAC;GAAK;GAAK;GAAM;GAAI,CAAC,CACK;EAC1D,MAAM,SAAS,oBAAoB,KAAK,UAAU,GAAG;EACrD,MAAM,UAAU,qBAAqB,KAAK,WAAW,GAAG;EACxD,MAAM,SAAS,MAAM,oBAAoB,MAAM,QAAQ,CAAC,OAAO,OAAO,EAAE,SAAS,QAAQ,GAAG;EAC5F,MAAM,KAAK,KAAK,IAAI,OAAO,KAAK,EAAI,GAAG,QAAQ,KAAK,IAAI,OAAO,KAAK,GAAK,GAAG;AAC5E,UAAQ,KACN,KAAK,uBAAuB,2BAA2B,MAAM,KAAK,OAAO,CAAC,KAAK,IAAI,CAAC,GACrF;AACD,QAAM,SAAS;AACf,SAAO,SAAS;AAChB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAM1E,KAAI;EACF,MAAM,SAAS;;;;;;;;;;;;;;;;;;;;;;EAuBf,MAAM,QAAQ,IAAI,aAAa;GAAC;GAAK;GAAM;GAAK;GAAM;GAAO;GAAM;GAAO;GAAK,CAAC;EAChF,MAAM,QAAQ,oBAAoB,KAAK,SAAS,MAAM,YAAY,MAAM;EACxE,MAAM,SAAS,oBAAoB,KAAK,UAAU,MAAM,WAAW;EACnE,MAAM,UAAU,qBAAqB,KAAK,WAAW,MAAM,WAAW;EACtE,MAAM,SAAS,MAAM,iBACnB,MACA,QACA,CAAC,OAAO,OAAO,EACf,SACA,QACA,MAAM,WACP;EAED,IAAI,KAAK;EACT,MAAMM,aAAuB,EAAE;AAC/B,OAAK,IAAI,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;GACrC,MAAM,MAAM,KAAK,IAAI,KAAK,IAAI,MAAM,GAAG,GAAG,KAAM,KAAM;AACtD,OAAI,KAAK,IAAI,OAAO,KAAK,MAAM,GAAG,GAAG,KAAK;AACxC,SAAK;AACL,eAAW,KAAK,IAAI,EAAE,IAAI,MAAM,GAAG,GAAG,OAAO,KAAK;;;AAGtD,UAAQ,KACN,KACI,iCACA,qCAAqC,WAAW,KAAK,KAAK,CAAC,GAChE;AACD,QAAM,SAAS;AACf,SAAO,SAAS;AAChB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAK1E,KAAI;EACF,MAAM,SAAS;;;;;;;;;;;;;;;;;;;;;;;EAuBf,MAAM,QAAQ,IAAI,aAAa;GAAC;GAAK;GAAO;GAAK;GAAK;GAAM;GAAK;GAAK;GAAK,CAAC;EAC5E,MAAM,QAAQ,oBAAoB,KAAK,SAAS,MAAM,YAAY,MAAM;EACxE,MAAM,YAAY,oBAAoB,KAAK,aAAa,GAAG;EAC3D,MAAM,SAAS,oBAAoB,KAAK,UAAU,MAAM,WAAW;EACnE,MAAM,UAAU,qBAAqB,KAAK,WAAW,MAAM,WAAW;EACtE,MAAM,SAAS,MAAM,iBACnB,MACA,QACA;GAAC;GAAO;GAAW;GAAO,EAC1B,SACA,QACA,MAAM,WACP;EACD,IAAI,KAAK;EACT,MAAMA,aAAuB,EAAE;AAC/B,OAAK,IAAI,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;GACrC,MAAM,MAAM,KAAK,IAAI,KAAK,IAAI,MAAM,GAAG,GAAG,KAAM,KAAM;AACtD,OAAI,KAAK,IAAI,OAAO,KAAK,MAAM,GAAG,GAAG,KAAK;AACxC,SAAK;AACL,eAAW,KAAK,IAAI,EAAE,IAAI,MAAM,GAAG,GAAG,OAAO,KAAK;;;AAGtD,UAAQ,KACN,KACI,iCACA,qCAAqC,WAAW,KAAK,KAAK,CAAC,GAChE;AACD,QAAM,SAAS;AACf,YAAU,SAAS;AACnB,SAAO,SAAS;AAChB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAM1E,KAAI;EACF,MAAM,SAAS;;;;;;;;;;;;;;;;;;;;;;EAuBf,MAAM,QAAQ,IAAI,aAAa,IAAI;AACnC,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,IAAK,OAAM,KAAK,IAAI;EAC7C,MAAM,cAAe,MAAM,MAAM,MAAO;EACxC,MAAM,QAAQ,oBAAoB,KAAK,SAAS,MAAM,YAAY,MAAM;EACxE,MAAM,SAAS,oBAAoB,KAAK,UAAU,MAAM,EAAE;EAC1D,MAAM,UAAU,qBAAqB,KAAK,WAAW,MAAM,EAAE;EAE7D,MAAM,OADS,MAAM,iBAAiB,MAAM,QAAQ,CAAC,OAAO,OAAO,EAAE,SAAS,QAAQ,MAAM,EAAE,EAC3E;EACnB,MAAM,KAAK,KAAK,IAAI,MAAM,YAAY,GAAG,cAAc;AACvD,UAAQ,KACN,KACI,mCAAmC,IAAI,KACvC,2CAA2C,IAAI,aAAa,YAAY,GAC7E;AACD,QAAM,SAAS;AACf,SAAO,SAAS;AAChB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAK1E,KAAI;EACF,MAAM,SAAS;;;;;;;;;;;;;;;;;EAkBf,MAAM,aAAa,IAAI,YAAY,CAAC,WAAW,CAAC;EAChD,MAAM,YAAY,IAAI,aAAa,CAAC,GAAI,CAAC;EACzC,MAAM,WAAW,IAAI,aAAa,CAAC,EAAI,CAAC;EAExC,MAAM,WAAW;GAAC;GAAG;GAAG;GAAG;GAAG;GAAG;GAAG;GAAG;GAAE,CAAC,KAAK,OAAO,IAAI,KAAO,GAAI;EAErE,MAAM,YAAY,oBAAoB,KAAK,aAAa,GAAG,WAAW;EACtE,MAAM,WAAW,oBAAoB,KAAK,YAAY,GAAG,UAAU;EACnE,MAAM,UAAU,oBAAoB,KAAK,WAAW,GAAG,SAAS;EAChE,MAAM,SAAS,oBAAoB,KAAK,UAAU,GAAG;EACrD,MAAM,UAAU,qBAAqB,KAAK,WAAW,GAAG;EACxD,MAAM,SAAS,MAAM,iBACnB,MACA,QACA;GAAC;GAAW;GAAU;GAAS;GAAO,EACtC,SACA,QACA,GACD;EACD,IAAI,KAAK;EACT,MAAMA,aAAuB,EAAE;AAC/B,OAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IACrB,KAAI,KAAK,IAAI,OAAO,KAAK,SAAS,GAAG,GAAG,MAAO;AAC7C,QAAK;AACL,cAAW,KAAK,IAAI,EAAE,IAAI,SAAS,GAAG,GAAG,OAAO,KAAK;;AAGzD,UAAQ,KACN,KAAK,yBAAyB,6BAA6B,WAAW,KAAK,KAAK,CAAC,GAClF;AACD,YAAU,SAAS;AACnB,WAAS,SAAS;AAClB,UAAQ,SAAS;AACjB,SAAO,SAAS;AAChB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAM1E,KAAI;EACF,MAAM,YAAY;;;;;;;;EAQlB,MAAM,YAAY;;;;;;;EAclB,MAAM,SAAS,oBAAoB,KAAK,UAAU,IADlC,IAAI,aAAa;GAAC;GAAG;GAAG;GAAG;GAAE,CAAC,CACgB;EAC9D,MAAM,SAAS,oBAAoB,KAAK,UAAU,IAAI,IAAI,aAAa,EAAE,CAAC;EAC1E,MAAM,UAAU,qBAAqB,KAAK,WAAW,GAAG;EAExD,MAAM,cAAc,oBAAoB,KAAK,UAAU,WAAW,OAAO;EACzE,MAAM,cAAc,oBAAoB,KAAK,UAAU,WAAW,OAAO;EAEzE,MAAM,QAAQ,gBACZ,KACA,aACA,CAAC,EAAE,QAAQ,QAAQ,EAAE,EAAE,QAAQ,QAAQ,CAAC,EACxC,YACD;EACD,MAAM,QAAQ,gBAAgB,KAAK,aAAa,CAAC,EAAE,QAAQ,QAAQ,CAAC,EAAE,YAAY;EAElF,MAAM,UAAU,IAAI,OAAO,sBAAsB;EACjD,MAAM,OAAO,QAAQ,kBAAkB;AACvC,OAAK,YAAY,YAAY;AAC7B,OAAK,aAAa,GAAG,MAAM;AAC3B,OAAK,mBAAmB,EAAE;AAC1B,OAAK,YAAY,YAAY;AAC7B,OAAK,aAAa,GAAG,MAAM;AAC3B,OAAK,mBAAmB,EAAE;AAC1B,OAAK,YAAY,YAAY;AAC7B,OAAK,aAAa,GAAG,MAAM;AAC3B,OAAK,mBAAmB,EAAE;AAC1B,OAAK,KAAK;AACV,UAAQ,mBAAmB,QAAQ,GAAG,SAAS,GAAG,GAAG;AACrD,MAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;AAC3C,MAAI,IAAI,OAAO,MAAM,oBACnB,OAAM,IAAI,OAAO,MAAM,qBAAqB;AAE9C,QAAM,QAAQ,SAASN,iBAAe,GAAG,GAAG;EAC5C,MAAM,SAAS,QAAQ,eAAe,GAAG,GAAG;EAC5C,MAAM,SAAS,IAAI,aAAa,OAAO,MAAM,EAAE,CAAC;AAChD,UAAQ,OAAO;EAEf,MAAM,WAAW;GAAC;GAAG;GAAG;GAAG;GAAG;EAC9B,IAAI,KAAK;AACT,OAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IACrB,KAAI,KAAK,IAAI,OAAO,KAAK,SAAS,GAAG,GAAG,IAAM,MAAK;AAErD,UAAQ,KACN,KACI,iCAAiC,MAAM,KAAK,OAAO,CAAC,KAAK,IAAI,CAAC,KAC9D,yCAAyC,MAAM,KAAK,OAAO,CAAC,KAAK,IAAI,CAAC,aAAa,SAAS,KAAK,IAAI,CAAC,GAC3G;AACD,SAAO,SAAS;AAChB,SAAO,SAAS;AAChB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAK1E,KAAI;EACF,MAAM,SAAS;;;;;;;;;;;;;;;;;EAiBf,MAAM,SAAS,oBAAoB,KAAK,UAAU,GAAG;EACrD,MAAM,UAAU,qBAAqB,KAAK,WAAW,GAAG;EACxD,MAAM,SAAS,MAAM,iBAAiB,MAAM,QAAQ,CAAC,OAAO,EAAE,SAAS,QAAQ,GAAG;EAClF,MAAM,YAAY,OAAO,KAAK,KAAK,OAAO,KAAK,SAAS,OAAO,KAAK,KAAK,OAAO,KAAK;EACrF,MAAM,WAAW,KAAK,IAAI,OAAO,KAAK,KAAK,IAAI,IAAI,CAAC,GAAG;AACrC,OAAK,IAAI,OAAO,KAAK,EAAI;AAEzC,SAAO,OAAO,QAAQ,OAAO,MAAM,OAAO,GAAG,IAAI,OAAO,MAAM,OAAO,GAAG,IAAI,OAAO;EAErF,MAAM,SACJ,WAAW,OAAO,GAAG,cAAc,EAAE,CAAC,GAAG,OAAO,GAAG,cAAc,EAAE,CAAC,YACxD,OAAO,GAAG,cAAc,EAAE,CAAC,OAAO,OAAO,GAAG,QAAQ,EAAE,CAAC,kBACjD,OAAO,GAAG,cAAc,OAAO,GAAG,cAAc,OAAO,GAAG,QACpE,OAAO;AAEjB,UAAQ,KACN,aAAa,WACT,wBAAwB,OAAO,KAC/B,4BAA4B,OAAO,GACxC;AACD,SAAO,SAAS;AAChB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAM1E,KAAI;EACF,MAAM,KAAK,MAAM;EACjB,MAAM,WAAW,IAAI,aAAa,GAAG;AAErC,OAAK,IAAI,IAAI,GAAG,IAAI,IAAI,IACtB,UAAS,KAAK,KAAK,IAAI,IAAI,KAAM,IAAI,IAAI,OAAQ;EAEnD,MAAM,SAAS,oBAAoB,KAAK,UAAU,SAAS,YAAY,SAAS;EAChF,MAAM,UAAU,qBAAqB,KAAK,WAAW,SAAS,WAAW;EACzE,MAAM,SAAS,MAAM,eAAe,KAAK,QAAQ,SAAS,GAAG,SAAS,WAAW;EAEjF,IAAI,aAAa;EACjB,IAAI,mBAAmB;AACvB,OAAK,IAAI,IAAI,GAAG,IAAI,IAAI,IACtB,KAAI,KAAK,IAAI,OAAO,KAAK,SAAS,GAAG,GAAG,MAAM;AAC5C,OAAI,qBAAqB,GAAI,oBAAmB;AAChD;;EAGJ,MAAM,KAAK,eAAe;AAC1B,UAAQ,KACN,KACI,iCAAiC,GAAG,YACpC,2BAA2B,WAAW,yBAAyB,iBAAiB,cAAc,SAAS,kBAAkB,QAAQ,OAAO,kBAAkB,GAC/J;AACD,SAAO,SAAS;AAChB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAM1E,KAAI;EAGF,MAAM,UAAU,IAAI,aADA,MAAM,KACmB;EAC7C,MAAM,aAAa,MAAM;EACzB,MAAM,WAAW,MAAM;AACvB,OAAK,IAAI,IAAI,GAAG,IAAI,UAAU,IAC5B,SAAQ,aAAa,KAAK,KAAK,IAAI,IAAI,KAAM,IAAI,IAAI,OAAO;EAG9D,MAAM,OAAO,IAAI,aAAa,QAAQ,QAAQ,aAAa,GAAG,SAAS;EAEvE,MAAM,SAAS,oBAAoB,KAAK,UAAU,KAAK,YAAY,KAAK;EACxE,MAAM,UAAU,qBAAqB,KAAK,WAAW,KAAK,WAAW;EACrE,MAAM,SAAS,MAAM,eAAe,KAAK,QAAQ,SAAS,GAAG,KAAK,WAAW;EAE7E,IAAI,aAAa;EACjB,IAAI,mBAAmB;AACvB,OAAK,IAAI,IAAI,GAAG,IAAI,UAAU,IAC5B,KAAI,KAAK,IAAI,OAAO,KAAK,KAAK,GAAG,GAAG,MAAM;AACxC,OAAI,qBAAqB,GAAI,oBAAmB;AAChD;;EAGJ,MAAM,KAAK,eAAe;AAC1B,UAAQ,KACN,KACI,0DACA,8BAA8B,WAAW,yBAAyB,iBAAiB,cAAc,KAAK,kBAAkB,QAAQ,OAAO,kBAAkB,GAC9J;AACD,SAAO,SAAS;AAChB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAM1E,KAAI;EACF,MAAM,gBAAgB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;EA0DtB,MAAM,aAAa;EACnB,MAAM,YAAY,IAAI,aAAa,WAAW;EAC9C,MAAM,aAAa,IAAI,aAAa,WAAW;AAC/C,OAAK,IAAI,IAAI,GAAG,IAAI,YAAY,KAAK;AACnC,aAAU,KAAK,KAAK,IAAI,IAAI,IAAK,GAAG;AACpC,cAAW,KAAK,IAAM,KAAK,IAAI,IAAI,KAAM,GAAG;;EAI9C,IAAI,QAAQ;AACZ,OAAK,IAAI,IAAI,GAAG,IAAI,YAAY,IAAK,UAAS,UAAU,KAAK,UAAU;EACvE,MAAM,MAAM,KAAK,KAAK,QAAQ,aAAa,KAAK;EAChD,MAAM,iBAAiB,IAAI,aAAa,WAAW;AACnD,OAAK,IAAI,IAAI,GAAG,IAAI,YAAY,IAC9B,gBAAe,KAAM,UAAU,KAAK,MAAO,WAAW;EAIxD,MAAM,SAAS,IAAI,aAAa,EAAE;AAClC,SAAO,KAAK;EACZ,MAAM,UAAU,IAAI,YAAY,OAAO,OAAO,CAAC;EAG/C,MAAM,4BAAY,IAAI,YAAY,GAAG;EACrC,MAAM,aAAa,IAAI,SAAS,UAAU;AAC1C,aAAW,UAAU,GAAG,GAAG,KAAK;AAChC,aAAW,UAAU,GAAG,YAAY,KAAK;AACzC,aAAW,UAAU,GAAG,SAAS,KAAK;AACtC,aAAW,UAAU,IAAI,GAAG,KAAK;EAEjC,MAAM,QAAQ,oBAAoB,KAAK,SAAS,UAAU,YAAY,UAAU;EAChF,MAAM,OAAO,oBAAoB,KAAK,QAAQ,WAAW,YAAY,WAAW;EAChF,MAAM,SAAS,oBAAoB,KAAK,UAAU,aAAa,EAAE;EACjE,MAAM,aAAa,oBAAoB,KAAK,aAAa,UAAU;EACnE,MAAM,UAAU,qBAAqB,KAAK,WAAW,aAAa,EAAE;EAEpE,MAAM,WAAW,oBAAoB,KAAK,cAAc,eAAe,OAAO;EAC9E,MAAM,KAAK,gBACT,KACA,UACA;GAAC,EAAE,QAAQ,OAAO;GAAE,EAAE,QAAQ,MAAM;GAAE,EAAE,QAAQ,QAAQ;GAAE,EAAE,QAAQ,YAAY;GAAC,EACjF,QACD;EAED,MAAM,UAAU,IAAI,OAAO,sBAAsB;EACjD,MAAM,OAAO,QAAQ,kBAAkB;AACvC,OAAK,YAAY,SAAS;AAC1B,OAAK,aAAa,GAAG,GAAG;AACxB,OAAK,mBAAmB,EAAE;AAC1B,OAAK,KAAK;AACV,UAAQ,mBAAmB,QAAQ,GAAG,SAAS,GAAG,aAAa,EAAE;AACjE,MAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;AAC3C,MAAI,IAAI,OAAO,MAAM,oBACnB,OAAM,IAAI,OAAO,MAAM,qBAAqB;AAE9C,QAAM,QAAQ,SAASA,iBAAe,GAAG,aAAa,EAAE;EACxD,MAAM,SAAS,QAAQ,eAAe,GAAG,aAAa,EAAE;EACxD,MAAM,SAAS,IAAI,aAAa,OAAO,MAAM,EAAE,CAAC;AAChD,UAAQ,OAAO;EAEf,IAAI,SAAS;EACb,IAAI,YAAY;AAChB,OAAK,IAAI,IAAI,GAAG,IAAI,YAAY,KAAK;GACnC,MAAM,MAAM,KAAK,IAAI,OAAO,KAAK,eAAe,GAAG;AACnD,OAAI,MAAM,QAAQ;AAChB,aAAS;AACT,gBAAY;;;EAIhB,MAAM,KAAK,SAAS;AACpB,UAAQ,KACN,KACI,uCAAuC,OAAO,cAAc,EAAE,CAAC,OAAO,UAAU,MAChF,2CAA2C,OAAO,cAAc,EAAE,CAAC,OAAO,UAAU,cAAc,eAAe,WAAW,QAAQ,EAAE,CAAC,QAAQ,OAAO,WAAW,QAAQ,EAAE,CAAC,GACjL;AACD,QAAM,SAAS;AACf,OAAK,SAAS;AACd,SAAO,SAAS;AAChB,aAAW,SAAS;AACpB,UAAQ,SAAS;UACV,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAK1E,KAAI;EACF,MAAM,eAAe;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;EAsDrB,MAAM,QAAQ;EACd,MAAM,QAAQ;EACd,MAAM,QAAQ,IAAI,aAAa,MAAM;EACrC,MAAM,QAAQ,IAAI,aAAa,QAAQ,MAAM;AAC7C,OAAK,IAAI,IAAI,GAAG,IAAI,OAAO,IAAK,OAAM,KAAK,KAAK,IAAI,IAAI,IAAK,GAAG;AAChE,OAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,OAAO,IAAK,OAAM,KAAK,KAAK,IAAI,IAAI,KAAM,GAAG;EAGzE,MAAM,YAAY,IAAI,aAAa,MAAM;AACzC,OAAK,IAAI,IAAI,GAAG,IAAI,OAAO,KAAK;GAC9B,IAAI,IAAI;AACR,QAAK,IAAI,IAAI,GAAG,IAAI,OAAO,IAAK,MAAK,MAAM,KAAK,MAAM,IAAI,QAAQ;AAClE,aAAU,KAAK;;EAGjB,MAAM,4BAAY,IAAI,YAAY,EAAE;EACpC,MAAM,MAAM,IAAI,SAAS,UAAU;AACnC,MAAI,UAAU,GAAG,OAAO,KAAK;AAC7B,MAAI,UAAU,GAAG,OAAO,KAAK;EAE7B,MAAM,OAAO,oBAAoB,KAAK,QAAQ,MAAM,YAAY,MAAM;EACtE,MAAM,OAAO,oBAAoB,KAAK,QAAQ,MAAM,YAAY,MAAM;EACtE,MAAM,OAAO,oBAAoB,KAAK,QAAQ,QAAQ,EAAE;EACxD,MAAM,OAAO,oBAAoB,KAAK,aAAa,UAAU;EAC7D,MAAM,OAAO,qBAAqB,KAAK,WAAW,QAAQ,EAAE;EAE5D,MAAM,aAAa,oBAAoB,KAAK,aAAa,cAAc,OAAO;EAC9E,MAAM,OAAO,gBACX,KACA,YACA;GAAC,EAAE,QAAQ,MAAM;GAAE,EAAE,QAAQ,MAAM;GAAE,EAAE,QAAQ,MAAM;GAAE,EAAE,QAAQ,MAAM;GAAC,EACxE,QACD;EAGD,MAAM,WAAW,IAAI,OAAO,sBAAsB;EAClD,MAAM,QAAQ,SAAS,kBAAkB;AACzC,QAAM,YAAY,WAAW;AAC7B,QAAM,aAAa,GAAG,KAAK;AAC3B,QAAM,mBAAmB,KAAK,KAAK,QAAQ,EAAE,CAAC;AAC9C,QAAM,KAAK;AACX,WAAS,mBAAmB,MAAM,GAAG,MAAM,GAAG,QAAQ,EAAE;AACxD,MAAI,OAAO,MAAM,OAAO,CAAC,SAAS,QAAQ,CAAC,CAAC;AAC5C,MAAI,IAAI,OAAO,MAAM,oBACnB,OAAM,IAAI,OAAO,MAAM,qBAAqB;AAE9C,QAAM,KAAK,SAASA,iBAAe,GAAG,QAAQ,EAAE;EAChD,MAAM,UAAU,KAAK,eAAe,GAAG,QAAQ,EAAE;EACjD,MAAM,UAAU,IAAI,aAAa,QAAQ,MAAM,EAAE,CAAC;AAClD,OAAK,OAAO;EAEZ,IAAI,UAAU;EACd,IAAI,aAAa;AACjB,OAAK,IAAI,IAAI,GAAG,IAAI,OAAO,KAAK;GAC9B,MAAM,MAAM,KAAK,IAAI,QAAQ,KAAK,UAAU,GAAG;AAC/C,OAAI,MAAM,SAAS;AACjB,cAAU;AACV,iBAAa;;;EAGjB,MAAM,MAAM,UAAU;AACtB,UAAQ,KACN,MACI,sCAAsC,QAAQ,cAAc,EAAE,CAAC,OAAO,WAAW,MACjF,0CAA0C,QAAQ,cAAc,EAAE,CAAC,OAAO,WAAW,cAAc,UAAU,YAAY,QAAQ,EAAE,CAAC,QAAQ,QAAQ,YAAY,QAAQ,EAAE,CAAC,GAChL;AACD,OAAK,SAAS;AACd,OAAK,SAAS;AACd,OAAK,SAAS;AACd,OAAK,SAAS;AACd,OAAK,SAAS;UACP,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAO1E,KAAI;EACF,MAAM,aAAa;;;;;;;EAOnB,MAAM,iBAAiB;EACvB,MAAM,UAAU,oBAAoB,KAAK,WAAW,IAAI,IAAI,aAAa,EAAE,CAAC;EAC5E,MAAM,cAAc,qBAAqB,KAAK,WAAW,GAAG;EAC5D,MAAM,eAAe,oBAAoB,KAAK,MAAM,YAAY,OAAO;EACvE,MAAM,SAAS,gBAAgB,KAAK,cAAc,CAAC,EAAE,QAAQ,SAAS,CAAC,EAAE,QAAQ;EAEjF,MAAM,MAAM,IAAI,OAAO,sBAAsB;EAC7C,MAAM,QAAQ,IAAI,kBAAkB;AACpC,OAAK,IAAI,IAAI,GAAG,IAAI,gBAAgB,KAAK;AACvC,SAAM,YAAY,aAAa;AAC/B,SAAM,aAAa,GAAG,OAAO;AAC7B,SAAM,mBAAmB,EAAE;;AAE7B,QAAM,KAAK;AACX,MAAI,mBAAmB,SAAS,GAAG,aAAa,GAAG,GAAG;AACtD,MAAI,OAAO,MAAM,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;AACvC,MAAI,IAAI,OAAO,MAAM,oBACnB,OAAM,IAAI,OAAO,MAAM,qBAAqB;AAE9C,QAAM,YAAY,SAASA,iBAAe,GAAG,GAAG;EAChD,MAAM,aAAa,YAAY,eAAe,GAAG,GAAG;EACpD,MAAM,aAAa,IAAI,aAAa,WAAW,MAAM,EAAE,CAAC;AACxD,cAAY,OAAO;EAEnB,MAAM,aAAa,WAAW,OAAO,MAAM,KAAK,IAAI,IAAI,eAAe,GAAG,GAAI;AAC9E,UAAQ,KACN,aACI,QAAQ,eAAe,kBAAkB,WAAW,GAAG,KACvD,QAAQ,eAAe,2BAA2B,MAAM,KAAK,WAAW,CAAC,KAAK,IAAI,CAAC,cAAc,eAAe,GACrH;AACD,UAAQ,SAAS;AACjB,cAAY,SAAS;UACd,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAQ1E,KAAI;EACF,MAAM,UAAU;;;;;;;;EAQhB,MAAM,OAAO,oBAAoB,KAAK,UAAU,IAAI,IAAI,aAAa;GAAC;GAAG;GAAG;GAAG;GAAE,CAAC,CAAC;EACnF,MAAM,QAAQ,oBAAoB,KAAK,WAAW,IAAI,IAAI,aAAa,EAAE,CAAC;EAC1E,MAAM,QAAQ,oBAAoB,KAAK,WAAW,IAAI,IAAI,aAAa,EAAE,CAAC;EAC1E,MAAM,SAAS,qBAAqB,KAAK,YAAY,GAAG;EACxD,MAAM,SAAS,qBAAqB,KAAK,YAAY,GAAG;EAGxD,MAAM,YAAY,oBAAoB,KAAK,MAAM,SAAS,OAAO;EACjE,MAAM,OAAO,gBAAgB,KAAK,WAAW,CAAC,EAAE,QAAQ,MAAM,EAAE,EAAE,QAAQ,OAAO,CAAC,EAAE,SAAS;EAC7F,MAAM,OAAO,gBAAgB,KAAK,WAAW,CAAC,EAAE,QAAQ,MAAM,EAAE,EAAE,QAAQ,OAAO,CAAC,EAAE,SAAS;EAE7F,MAAM,OAAO,IAAI,OAAO,sBAAsB;EAC9C,MAAM,QAAQ,KAAK,kBAAkB;AACrC,QAAM,YAAY,UAAU;AAC5B,QAAM,aAAa,GAAG,KAAK;AAC3B,QAAM,mBAAmB,EAAE;AAC3B,QAAM,YAAY,UAAU;AAC5B,QAAM,aAAa,GAAG,KAAK;AAC3B,QAAM,mBAAmB,EAAE;AAC3B,QAAM,KAAK;AACX,OAAK,mBAAmB,OAAO,GAAG,QAAQ,GAAG,GAAG;AAChD,OAAK,mBAAmB,OAAO,GAAG,QAAQ,GAAG,GAAG;AAChD,MAAI,OAAO,MAAM,OAAO,CAAC,KAAK,QAAQ,CAAC,CAAC;AACxC,MAAI,IAAI,OAAO,MAAM,oBACnB,OAAM,IAAI,OAAO,MAAM,qBAAqB;AAE9C,QAAM,OAAO,SAASA,iBAAe,GAAG,GAAG;EAC3C,MAAM,MAAM,IAAI,aAAa,OAAO,eAAe,GAAG,GAAG,CAAC,MAAM,EAAE,CAAC;AACnE,SAAO,OAAO;AACd,QAAM,OAAO,SAASA,iBAAe,GAAG,GAAG;EAC3C,MAAM,MAAM,IAAI,aAAa,OAAO,eAAe,GAAG,GAAG,CAAC,MAAM,EAAE,CAAC;AACnE,SAAO,OAAO;EAEd,MAAM,YAAY;GAAC;GAAG;GAAG;GAAG;GAAE;EAC9B,MAAM,SAAS,UAAU,OAAO,GAAG,MAAM,KAAK,IAAI,IAAI,KAAK,EAAE,GAAG,IAAK;EACrE,MAAM,SAAS,UAAU,OAAO,GAAG,MAAM,KAAK,IAAI,IAAI,KAAK,EAAE,GAAG,IAAK;AACrE,UAAQ,KACN,UAAU,SACN,6CAA6C,MAAM,KAAK,IAAI,CAAC,UAAU,MAAM,KAAK,IAAI,CAAC,MACvF,iDAAiD,MAAM,KAAK,IAAI,CAAC,UAAU,MAAM,KAAK,IAAI,CAAC,eAAe,UAAU,IACzH;AACD,OAAK,SAAS;AACd,QAAM,SAAS;AACf,QAAM,SAAS;AACf,SAAO,SAAS;AAChB,SAAO,SAAS;UACT,GAAG;AACV,UAAQ,KAAK,cAAc,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE,GAAG;;AAG1E,QAAO;EAAE;EAAiB;EAAc;EAAmB;EAAS;;;;;;ACp+CtE,SAAS,KAAK,GAAW,GAAmB;AAC1C,QAAO,KAAK,KAAK,IAAI,EAAE;;;AAIzB,SAAS,aAAa,OAAuB;CAC3C,MAAM,IAAI,IAAI,aAAa,EAAE;AAC7B,GAAE,KAAK;AACP,QAAO,IAAI,YAAY,EAAE,OAAO,CAAC;;;;;;AAOnC,SAAS,mBAAmB,QAA+B;CACzD,MAAM,sBAAM,IAAI,YAAY,OAAO,SAAS,EAAE;CAC9C,MAAM,OAAO,IAAI,SAAS,IAAI;AAC9B,MAAK,IAAI,IAAI,GAAG,IAAI,OAAO,QAAQ,IACjC,MAAK,UAAU,IAAI,GAAG,OAAO,IAAI,KAAwB;AAE3D,QAAO;;;;;;;;;;AAwCT,SAAS,cACP,IACA,gBACA,QACQ;CAER,MAAM,MAAM,GAAG,WAAW;AAC1B,KAAI,OAAO,eAAe,KAGxB,QAFc,eAAe,KACT,QAAQ,GAAW,MAAc,IAAI,GAAG,EAAE,GAC/C;CAGjB,MAAM,WAAW,eAAe,GAAG,QAAQ;AAC3C,KAAI,YAAY,SAAS,UAAU,EACjC,QAAO,SAAS;AAElB,QAAO,GAAG,WAAW;;AAKvB,MAAM,iBAAiB;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA6BvB,MAAM,sBAAsB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAsD5B,MAAM,cAAc;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA4HpB,MAAM,mBAAmB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAmHzB,MAAM,0BAA0B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAkHhC,MAAM,mBAAmB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAiFzB,MAAM,cAAc;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAsFpB,MAAM,wBAAwB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA4E9B,MAAM,6BAA6B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA0FnC,MAAM,mBAAmB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAsGzB,MAAM,yBAAyB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAiG/B,MAAM,gCAAgC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA0FtC,MAAM,cAAc;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAoEpB,MAAM,eAAe;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA2DrB,MAAM,oBAAoB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA4E1B,MAAM,iBAAiB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA+EvB,MAAM,YAAY;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA4GlB,MAAM,wBAAwB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA8F9B,MAAM,iBAAiB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAyVvB,MAAM,uBAAuB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuP7B,MAAM,eAAe;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAyErB,MAAM,YAAY;;;;;;;;;;;;;;;;;;;;AAqBlB,MAAM,YAAY;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA6BlB,MAAM,WAAW;;;;;;;;;;;;;;;;;;;AAoBjB,MAAM,WAAW;;;;;;;;;;;;;;;;;;;AAoBjB,MAAM,sBAAsB;;;;;;;;;;;;;;;;;;;;;AAsB5B,MAAM,iBAAiB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAkCvB,MAAM,aAAa;;;;;;;;;;;;;;;;;;;;;;AA4BnB,MAAM,eAAe;;;;;;;;;;;;;;;;;;AAsBrB,MAAM,cAAc;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA8DpB,MAAM,gBAAgB;;;;;;;;;;;;;;;;;;;;;;;AAwBtB,MAAM,gBAAgB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAoCtB,MAAM,kBAAkB;;;;;;;;;;;;;;;;;;;;;;;;;;AA2BxB,MAAM,gBAAgB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAiCtB,MAAM,oBAAoB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA+C1B,MAAM,aAAa;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA8EnB,MAAM,oBAAoB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAiC1B,MAAM,cAAc;;;;;;;;;;;;;;;;;;;;;;AAuBpB,MAAM,qBAAqB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAkE3B,MAAM,0BAA0B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAoHhC,MAAM,wBAAwB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAsG9B,MAAM,wBAAwB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAmE9B,MAAM,qBAAqB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAkD3B,MAAM,0BAA0B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAkDhC,MAAM,2BAA2B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAkDjC,MAAM,oBAAoB;;;;;;;;;;;;;;;;;;;;;;AAuB1B,MAAM,iBAAiB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAkIvB,MAAM,qBAAqB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqI3B,MAAM,uBAAuB;;;;;;;;;;;;;;;;;;;;AAqB7B,MAAM,4BAA4B;;;;;;;;;;;;;;;;;;;;;;;;;AA0BlC,MAAM,gCAAgC;;;;;;;;;;;;;;;;;;;;;;;AAwBtC,MAAM,uCAAuC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA8B7C,MAAM,2BAA2B;;;;;;;;;;;;;;;;;;;;;AAsBjC,MAAM,kCAAkC;;;;;;;;;;;;;;;;;;;;;;;;;;;AA4BxC,MAAM,qBAAqB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA8Q3B,MAAM,4BAA4B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAySlC,MAAM,yBAAyB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAkE/B,MAAM,mBAAmB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAsCzB,MAAM,sBAAsB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAwH5B,SAAS,mBAAmB,IAAY,gBAAkD;CACxF,MAAM,WAAW,eAAe,GAAG,QAAQ;CAC3C,IAAI,QAAQ;AACZ,MAAK,MAAM,KAAK,SAAU,UAAS;AACnC,QAAO;;AAKT,MAAMO,UAAsB;CAC1B,YAAY;CACZ,YAAY;CAMZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAGjC,YAAY,IAAI,gBAAgB;AAG9B,SAAO,mBAAmB,CAFZ,mBAAmB,IAAI,eAAe,CAEnB,CAAC;;CAErC;AAID,MAAMC,UAAsB;CAC1B,YAAY;CACZ,YAAY;CAMZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAGjC,YAAY,IAAI,gBAAgB;AAG9B,SAAO,mBAAmB,CAFZ,mBAAmB,IAAI,eAAe,CAEnB,CAAC;;CAErC;AAID,MAAMC,mBAA+B;CACnC,YAAY;CACZ,YAAY;CAKZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CAEzF,gBAAgB,IAAI;EAClB,MAAM,QAAQ,GAAG,WAAW;AAC5B,SAAO;GAAC,KAAK,OAAO,IAAI;GAAE;GAAG;GAAE;;CAGjC,YAAY,IAAI,gBAAgB;EAC9B,MAAM,QAAQ,GAAG,WAAW;AAG5B,SAAO,mBAAmB,CAAC,SAFjB,eAAe,GAAG,OAAO,MAAM,MAAM,KAER,KAAK,MAAM,CAAC;;CAEtD;AAID,MAAMC,eAA2B;CAC/B,YAAY;CACZ,YAAY;CAKZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CAEzF,gBAAgB,IAAI;EAClB,MAAM,QAAQ,GAAG,WAAW;AAC5B,SAAO;GAAC,KAAK,OAAO,IAAI;GAAE;GAAG;GAAE;;CAGjC,YAAY,IAAI,gBAAgB;EAC9B,MAAM,QAAQ,GAAG,WAAW;AAI5B,SAAO,mBAAmB,CAFhB,eAAe,GAAG,OAAO,MAAM,MAAM,GAEjB,MAAM,CAAC;;CAExC;AAID,MAAMC,YAAwB;CAC5B,YAAY;CACZ,YAAY;CAKZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CAEzF,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAGjC,YAAY,IAAI,gBAAgB;EAC9B,MAAM,QAAQ,mBAAmB,IAAI,eAAe;EACpD,MAAM,QAAQ,GAAG,WAAW;AAE5B,SAAO,mBAAmB,CAAC,OAAO,aAAa,MAAM,CAAC,CAAC;;CAE1D;AAID,MAAMC,cAA0B;CAC9B,YAAY;CACZ,YAAY;CAKZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CAEzF,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAGjC,YAAY,IAAI,gBAAgB;EAC9B,MAAM,QAAQ,mBAAmB,IAAI,eAAe;EACpD,MAAM,MAAM,GAAG,WAAW;AAE1B,SAAO,mBAAmB,CAAC,OAAO,aAAa,IAAI,CAAC,CAAC;;CAExD;AAID,MAAMC,aAAyB;CAC7B,YAAY;CACZ,YAAY;CAKZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CAEzF,gBAAgB,IAAI,gBAAgB;AAGlC,SAAO;GADM,eAAe,GAAG,QAAQ,MAAM,MAAM;GACrC;GAAG;GAAE;;CAGrB,YAAY,IAAI,gBAAgB;EAC9B,MAAM,QAAQ,GAAG,WAAW;AAG5B,SAAO,mBAAmB,CAFb,eAAe,GAAG,QAAQ,MAAM,MAAM,GAElB,MAAM,CAAC;;CAE3C;AAID,MAAMC,aAAyB;CAC7B,YAAY;CACZ,YAAY;CAMZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAGjC,YAAY,IAAI,gBAAgB;AAE9B,SAAO,mBAAmB,CADZ,mBAAmB,IAAI,eAAe,CACnB,CAAC;;CAErC;AAID,MAAMC,sBAAkC;CACtC,YAAY;CACZ,YAAY;CAQZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,SAAS,GAAG,WAAW;AAE7B,SAAO;GADQ,cAAc,IAAI,gBAAgB,OAAO;GACxC;GAAG;GAAE;;CAGvB,YAAY,IAAI,gBAAgB;EAC9B,MAAM,SAAS,GAAG,WAAW;AAG7B,SAAO,mBAAmB;GAFX,cAAc,IAAI,gBAAgB,OAAO;GAErB;GAAQ,aAD9B,GAAG,WAAW,OAAkB,KACe;GAAE;GAAE,CAAC;;CAEpE;AAID,MAAMC,WAAuB;CAC3B,YAAY;CACZ,YAAY;CAKZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CAEzF,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAGjC,YAAY,IAAI,gBAAgB;AAG9B,SAAO,mBAAmB,CAFZ,mBAAmB,IAAI,eAAe,CAEnB,CAAC;;CAErC;AAID,MAAMC,WAAuB;CAC3B,YAAY;CACZ,YAAY;CAKZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CAEzF,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAGjC,YAAY,IAAI,gBAAgB;AAG9B,SAAO,mBAAmB,CAFZ,mBAAmB,IAAI,eAAe,CAEnB,CAAC;;CAErC;AAID,MAAMC,cAA0B;CAC9B,YAAY;CACZ,YAAY;CACZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CACzF,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAEjC,YAAY,IAAI,gBAAgB;AAE9B,SAAO,mBAAmB,CADZ,mBAAmB,IAAI,eAAe,CACnB,CAAC;;CAErC;AAID,MAAMC,cAA0B;CAC9B,YAAY;CACZ,YAAY;CAGZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CACD,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAEjC,YAAY,IAAI,gBAAgB;EAC9B,MAAM,QAAQ,mBAAmB,IAAI,eAAe;EACpD,MAAM,QAAQ,GAAG,WAAW;AAC5B,SAAO,mBAAmB,CAAC,OAAO,MAAM,CAAC;;CAE5C;AAID,MAAMC,gBAA4B;CAChC,YAAY;CACZ,YAAY;CAGZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CACzF,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAEjC,YAAY,IAAI,gBAAgB;EAC9B,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,YAAY,GAAG,WAAW;EAChC,MAAM,aAAa,GAAG,WAAW;EAEjC,MAAM,OADW,eAAe,GAAG,QAAQ,IACrB;AACtB,SAAO,mBAAmB;GAAC;GAAM;GAAU;GAAW;GAAW,CAAC;;CAErE;AAID,MAAMC,cAA0B;CAC9B,YAAY;CACZ,YAAY;CAGZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CACzF,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAEjC,YAAY,IAAI,gBAAgB;EAC9B,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,YAAY,GAAG,WAAW;EAChC,MAAM,QAAQ,GAAG,WAAW;EAC5B,MAAM,QAAQ,GAAG,WAAW;EAE5B,MAAM,OADW,eAAe,GAAG,QAAQ,IACrB;AACtB,SAAO,mBAAmB;GAAC;GAAM;GAAU;GAAW;GAAO;GAAO;GAAG;GAAG;GAAE,CAAC;;CAEhF;AAID,MAAMC,kBAA8B;CAClC,YAAY;CACZ,YAAY;CAGZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CACD,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAEjC,YAAY,IAAI,gBAAgB;EAC9B,MAAM,YAAY,GAAG,WAAW;EAChC,MAAM,WAAW,GAAG,WAAW;AAI/B,SAAO,mBAAmB;GAHT,eAAe,GAAG,QAAQ,IACpB,QAAQ,GAAG,MAAM,IAAI,GAAG,EAAE,IACxB,YAAY;GACD;GAAW;GAAU;GAAE,CAAC;;CAE/D;AAID,MAAMC,aAAyB;CAC7B,YAAY;CACZ,YAAY;CAMZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAIlC,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,UAAU,GAAG,WAAW;EAC9B,IAAIC;AACJ,MAAI,WAAW,eAAe,SAC5B,KAAI,eAAe,SAAS;MAE5B,KAAI,GAAG,WAAW;AAEpB,SAAO;GAAC,KAAK,GAAG,GAAG;GAAE,KAAK,GAAG,GAAG;GAAE;GAAE;;CAGtC,YAAY,IAAI,gBAAgB;EAE9B,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,IAAI,GAAG,WAAW;EAExB,MAAM,UAAU,GAAG,WAAW;EAC9B,IAAIA;AACJ,MAAI,WAAW,eAAe,SAC5B,KAAI,eAAe,SAAS;MAE5B,KAAI,GAAG,WAAW;AAEpB,SAAO,mBAAmB;GAAC;GAAG;GAAG;GAAE,CAAC;;CAEvC;AAID,MAAMC,iBAA6B;CACjC,YAAY;CACZ,YAAY;CAGZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,UAAU,GAAG,WAAW;EAC9B,IAAID;AACJ,MAAI,WAAW,eAAe,SAC5B,KAAI,eAAe,SAAS;MAE5B,KAAI,GAAG,WAAW;AAEpB,SAAO;GAAC,KAAK,GAAG,GAAG;GAAE,KAAK,GAAG,GAAG;GAAE;GAAE;;CAGtC,YAAY,IAAI,gBAAgB;EAC9B,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,UAAU,GAAG,WAAW;EAC9B,IAAIA;AACJ,MAAI,WAAW,eAAe,SAC5B,KAAI,eAAe,SAAS;MAE5B,KAAI,GAAG,WAAW;AAEpB,SAAO,mBAAmB;GAAC;GAAG;GAAG;GAAE,CAAC;;CAEvC;AAID,MAAME,iBAA6B;CACjC,YAAY;CACZ,YAAY;CAGZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,MAAM,eAAe,GAAG,QAAQ;EACtC,MAAM,KAAK,IAAI;EACf,MAAM,IAAI,IAAI;AACd,SAAO;GAAC,KAAK,KAAK,GAAG,GAAG;GAAE;GAAG;GAAE;;CAGjC,YAAY,IAAI,gBAAgB;EAC9B,MAAM,MAAM,eAAe,GAAG,QAAQ;EACtC,MAAM,KAAK,IAAI;EACf,MAAM,IAAI,IAAI;AAId,SAAO,mBAAmB;GAAC;GAFjB,eAAe,GAAG,OAAO,MAAM,MAAO,GAAG,WAAW;GAE5B;GAAE,CAAC;;CAExC;AAID,MAAMC,oBAAgC;CACpC,YAAY;CACZ,YAAY;CAGZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,UAAU,GAAG,WAAW;EAC9B,MAAM,IACJ,WAAW,eAAe,WAAW,eAAe,SAAS,KAAM,GAAG,WAAW;AACnF,SAAO;GAAC,KAAK,GAAG,GAAG;GAAE,KAAK,GAAG,GAAG;GAAE;GAAE;;CAGtC,YAAY,IAAI,gBAAgB;EAC9B,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,UAAU,GAAG,WAAW;EAC9B,MAAM,IACJ,WAAW,eAAe,WAAW,eAAe,SAAS,KAAM,GAAG,WAAW;EAEnF,MAAM,OAAQ,GAAG,WAAW,QAAmB,OAAO;EACtD,MAAM,OAAQ,GAAG,WAAW,QAAmB,OAAO;EACtD,MAAM,OAAQ,GAAG,WAAW,QAAmB,OAAO;EACtD,MAAM,OAAQ,GAAG,WAAW,QAAmB,OAAO;AAEtD,SAAO,mBAAmB;GACxB;GACA;GACA;GACA,aAAa,KAAK;GAClB,aAAa,KAAK;GAClB,aAAa,KAAK;GAClB,aAAa,KAAK;GACnB,CAAC;;CAEL;AAKD,MAAaC,wBAAoC;CAC/C,GAAG;CACH,YAAY;CACb;AAID,MAAMC,iBAA6B;CACjC,YAAY;CACZ,YAAY;CAQZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAElC,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,UAAU,GAAG,WAAW;EAC9B,IAAIL;AACJ,MAAI,WAAW,eAAe,SAC5B,KAAI,eAAe,SAAS;MAE5B,KAAI,GAAG,WAAW;AAEpB,SAAO;GAAC,KAAK,GAAG,GAAG;GAAE,KAAK,GAAG,GAAG;GAAE;GAAE;;CAGtC,YAAY,IAAI,gBAAgB;EAE9B,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,aAAa,GAAG,WAAW;EACjC,MAAM,UAAU,GAAG,WAAW;EAC9B,IAAIA;AACJ,MAAI,WAAW,eAAe,SAC5B,KAAI,eAAe,SAAS;MAE5B,KAAI,GAAG,WAAW;AAEpB,SAAO,mBAAmB;GAAC;GAAG;GAAG;GAAG;GAAW,CAAC;;CAEnD;AAID,MAAaM,cAA0B;CACrC,YAAY;CACZ,YAAY;CAGZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI;EAClB,MAAM,IAAI,GAAG,WAAW;AAExB,SAAO;GAAC,KAAK,GADE,EACQ;GAAE;GAAG;GAAE;;CAGhC,YAAY,IAAI;EACd,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,IAAI,GAAG,WAAW;AACxB,SAAO,mBAAmB,CAAC,GAAG,EAAE,CAAC;;CAEpC;AAID,MAAaC,mBAA+B;CAC1C,YAAY;CACZ,YAAY;CAGZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI;EAClB,MAAM,IAAI,GAAG,WAAW;AAGxB,SAAO;GAAC,KAAK,GADE,GACQ;GAAE;GAAG;GAAE;;CAGhC,YAAY,IAAI;EACd,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,aAAa,GAAG,WAAW;AACjC,SAAO,mBAAmB;GAAC;GAAG;GAAG;GAAY;GAAE,CAAC;;CAEnD;AAID,MAAaC,yBAAqC;CAChD,YAAY;CACZ,YAAY;CAGZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI;EAClB,MAAM,IAAI,GAAG,WAAW;AAExB,SAAO;GAAC,KAAK,GADE,EACQ;GAAE;GAAG;GAAE;;CAGhC,YAAY,IAAI;EACd,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,aAAa,GAAG,WAAW;AACjC,SAAO,mBAAmB;GAAC;GAAG;GAAG;GAAY;GAAE,CAAC;;CAEnD;AAID,MAAaC,gCAA4C;CACvD,GAAG;CACH,YAAY;CACb;AASD,MAAaC,wBAAoC;CAC/C,GAAG;CACH,YAAY;CACb;AAED,MAAaC,6BAAyC;CACpD,GAAG;CACH,YAAY;CACb;AAID,MAAaC,qBAAiC;CAC5C,YAAY;CACZ,YAAY;CAOZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI;EAClB,MAAM,IAAI,GAAG,WAAW;AAExB,SAAO;GAAC,KAAK,GADE,EACQ;GAAE;GAAG;GAAE;;CAGhC,YAAY,IAAI;EACd,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,IAAI,GAAG,WAAW;AACxB,SAAO,mBAAmB,CAAC,GAAG,EAAE,CAAC;;CAEpC;AAID,MAAaC,0BAAsC;CACjD,YAAY;CACZ,YAAY;CAWZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI;EAClB,MAAM,IAAI,GAAG,WAAW;AAExB,SAAO;GAAC,KAAK,GADE,EACQ;GAAE;GAAG;GAAE;;CAGhC,YAAY,IAAI;EACd,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,aAAa,GAAG,WAAW;AACjC,SAAO,mBAAmB;GAAC;GAAG;GAAG;GAAY;GAAE,CAAC;;CAEnD;AAID,MAAaC,wBAAoC;CAC/C,YAAY;CACZ,YAAY;CAMZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI;EAClB,MAAM,IAAI,GAAG,WAAW;AAExB,SAAO;GAAC,KAAK,GADE,EACQ;GAAE;GAAG;GAAE;;CAGhC,YAAY,IAAI;EACd,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,aAAa,GAAG,WAAW;AACjC,SAAO,mBAAmB;GAAC;GAAG;GAAG;GAAY;GAAE,CAAC;;CAEnD;AAID,MAAaC,cAA0B;CACrC,YAAY;CACZ,YAAY;CAEZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CAEzF,kBAAkB;AAChB,SAAO;GAAC;GAAG;GAAG;GAAE;;CAGlB,YAAY,KAAK,iBAAiB;AAEhC,SAAO,mBAAmB,CAAC,GAAG,EAAE,CAAC;;CAEpC;AAID,MAAMC,cAA0B;CAC9B,YAAY;CACZ,YAAY;CAMZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAElC,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,MAAM,GAAG,WAAW;EAC1B,IAAIC;AACJ,MAAI,OAAO,eAAe,KAGxB,WAFc,eAAe,KACT,QAAQ,GAAW,MAAc,IAAI,GAAG,EAAE,GAC5C;MAElB,WAAU,GAAG,WAAW;AAE1B,SAAO;GAAC;GAAS;GAAG;GAAE;;CAGxB,YAAY,IAAI,gBAAgB;EAE9B,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,MAAM,GAAG,WAAW;EAC1B,MAAM,MAAM,GAAG,WAAW;EAC1B,IAAIA;AACJ,MAAI,OAAO,eAAe,KAGxB,WAFc,eAAe,KACT,QAAQ,GAAW,MAAc,IAAI,GAAG,EAAE,GAC5C;MAElB,WAAU,GAAG,WAAW;AAE1B,SAAO,mBAAmB;GACxB;GACA;GACA,aAAa,IAAI;GACjB;GACD,CAAC;;CAEL;AAID,SAAS,gBACP,IACA,gBACA,QACA,aACQ;CACR,MAAM,cAAc,GAAG,WAAW;CAClC,MAAM,MAAM,GAAG,WAAW;AAC1B,KAAI,OAAO,eAAe,KAExB,QADc,eAAe,KAAK,QAAQ,GAAW,MAAc,IAAI,GAAG,EAAE,GAC7D;AAEjB,QAAO,GAAG,WAAW;;AAGvB,MAAaC,oBAAgC;CAC3C,YAAY;CACZ,YAAY;CAGZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;AAGlC,SAAO;GAFO,gBAAgB,IAAI,gBAAgB,mBAAmB,WAAW,GAClE,gBAAgB,IAAI,gBAAgB,mBAAmB,WAAW;GACzD;GAAG;GAAE;;CAG9B,YAAY,IAAI,gBAAgB;EAC9B,MAAM,QAAQ,gBAAgB,IAAI,gBAAgB,mBAAmB,WAAW;EAChF,MAAM,QAAQ,gBAAgB,IAAI,gBAAgB,mBAAmB,WAAW;EAChF,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,MAAM,GAAG,WAAW;AAC1B,SAAO,mBAAmB;GAAC;GAAO;GAAO;GAAa,aAAa,IAAI;GAAC,CAAC;;CAE5E;AAID,MAAMC,gBAA4B;CAChC,YAAY;CACZ,YAAY;CAOZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAElC,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,MAAM,GAAG,WAAW;EAC1B,IAAIF;AACJ,MAAI,OAAO,eAAe,KAGxB,WAFc,eAAe,KACT,QAAQ,GAAW,MAAc,IAAI,GAAG,EAAE,GAC5C;MAElB,WAAU,GAAG,WAAW;AAE1B,SAAO;GAAC;GAAS;GAAG;GAAE;;CAGxB,YAAY,IAAI,gBAAgB;EAE9B,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,MAAM,GAAG,WAAW;EAC1B,MAAM,MAAM,GAAG,WAAW;EAC1B,IAAIA;AACJ,MAAI,OAAO,eAAe,KAGxB,WAFc,eAAe,KACT,QAAQ,GAAW,MAAc,IAAI,GAAG,EAAE,GAC5C;MAElB,WAAU,GAAG,WAAW;AAE1B,SAAO,mBAAmB;GACxB;GACA;GACA,aAAa,IAAI;GACjB;GACD,CAAC;;CAEL;AAID,MAAMG,gBAA4B;CAChC,YAAY;CACZ,YAAY;CAMZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,WAAW,eAAe,GAAG,QAAQ;AAE3C,SAAO;GAAC,MADQ,WAAW,SAAS,KAAM,GAAG,WAAW,WACjC,aAAa,IAAI;GAAE;GAAG;GAAE;;CAGjD,YAAY,IAAI,gBAAgB;EAE9B,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,WAAW,eAAe,GAAG,QAAQ;AAE3C,SAAO,mBAAmB,CADV,WAAW,SAAS,KAAM,GAAG,WAAW,SACpB,YAAY,CAAC;;CAEpD;AAID,MAAMC,oBAAgC;CACpC,YAAY;CACZ,YAAY;CAQZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,WAAW,eAAe,GAAG,QAAQ;AAE3C,SAAO;GAAC,MADQ,WAAW,SAAS,KAAM,GAAG,WAAW,WACjC,aAAa,IAAI;GAAE;GAAG;GAAE;;CAGjD,YAAY,IAAI,gBAAgB;EAE9B,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,aAAa,GAAG,WAAW;EACjC,MAAM,WAAW,eAAe,GAAG,QAAQ;AAE3C,SAAO,mBAAmB;GADV,WAAW,SAAS,KAAM,GAAG,WAAW;GACpB;GAAa;GAAY;GAAE,CAAC;;CAEnE;AAID,MAAMC,WAAuB;CAC3B,YAAY;CACZ,YAAY;CAKZ,UAAU;EAAC,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CAE/F,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,eAAe,GAAG,WAAW;EACnC,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,WAAY,GAAG,WAAW,YAAuB;EAGvD,MAAM,YAAa,GAAG,WAAW,aAAwB,WAAW;EAEpE,MAAM,SAAS,GAAG,WAAW;EAC7B,IAAIL;AACJ,MAAI,UAAU,eAAe,QAC3B,WAAU,eAAe,QAAQ;MAEjC,WAAU,GAAG,WAAW;EAG1B,MAAM,gBAAgB,UAAU,cAAc;EAC9C,MAAM,gBAAgB,UAAU,eAAe;AAC/C,SAAO;GAAC,KAAK,KAAK,IAAI,eAAe,cAAc,EAAE,IAAI;GAAE;GAAG;GAAE;;CAGlE,YAAY,IAAI,gBAAgB,SAAS;EAGvC,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,eAAe,GAAG,WAAW;EACnC,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,YAAY,GAAG,WAAW;EAChC,MAAM,WAAY,GAAG,WAAW,YAAuB;EAKvD,MAAM,YAAa,GAAG,WAAW,aAAwB,WAAW;EACpE,MAAM,aAAc,GAAG,WAAW,cAAyB;EAC3D,MAAM,oBAAqB,GAAG,WAAW,qBAAgC,WAAW;EAEpF,MAAM,kBAAkB,SAAS,UAAW,GAAG,WAAW,mBAA8B;EAExF,MAAM,SAAS,GAAG,WAAW;EAC7B,IAAIA;AACJ,MAAI,UAAU,eAAe,QAC3B,WAAU,eAAe,QAAQ;MAEjC,WAAU,GAAG,WAAW;AAG1B,SAAO,mBAAmB;GACxB;GACA;GACA;GACA;GACA,aAAa,UAAU;GACvB;GACA;GACA;GACA;GACA;GACA;GACA;GACD,CAAC;;CAEL;;;;;;AAOD,MAAaM,wBAAoC;CAC/C,GAAG;CACH,YAAY;CACb;AAID,MAAMC,YAAwB;CAC5B,YAAY;CACZ,YAAY;CAEZ,UAAU;EACR,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,WAAW;EACpB;CACD,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,eAAe,GAAG,WAAW;EAEnC,MAAM,OADW,GAAG,WAAW,WACP;EACxB,MAAM,SAAS,GAAG,WAAW;AAM7B,SAAO;GAAC,MAJN,UAAU,eAAe,UACrB,eAAe,QAAQ,KACtB,GAAG,WAAW,WACG,KAAK,IAAI,aAAa,aAAa,GAAG,MAC1C,IAAI;GAAE;GAAG;GAAE;;CAEjC,YAAY,IAAI,gBAAgB,SAAS;EACvC,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,eAAe,GAAG,WAAW;EACnC,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,SAAS,GAAG,WAAW;EAC7B,MAAM,UACJ,UAAU,eAAe,UACrB,eAAe,QAAQ,KACtB,GAAG,WAAW;AAGrB,SAAO,mBAAmB;GACxB;GACA;GACA;GACA;GACA;GANiB,YAAY,IAAK,SAAS,UAAU,IAAK;GAQ1D;GACA;GACD,CAAC;;CAEL;AAID,MAAMC,kBAA8B;CAClC,YAAY;CACZ,YAAY;CAEZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CACD,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,SAAS,GAAG,WAAW;EAE7B,MAAM,UAAU,eADD,GAAG,WAAW,gBACU;AACvC,SAAO;GAAC,KAAK,UAAU,QAAQ,IAAI;GAAE;GAAG;GAAE;;CAE5C,YAAY,IAAI,gBAAgB;EAC9B,MAAM,SAAS,GAAG,WAAW;EAE7B,MAAM,UAAU,eADD,GAAG,WAAW,gBACU;AACvC,SAAO,mBAAmB,CAAC,SAAS,OAAO,CAAC;;CAE/C;AAID,MAAMC,gBAA4B;CAChC,YAAY;CACZ,YAAY;CAOZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAElC,MAAM,cAAc,GAAG,WAAW;EAElC,MAAM,WAAW,eAAe,GAAG,QAAQ;AAE3C,SAAO;GADG,WAAW,SAAS,KAAM,GAAG,WAAW;GACvC;GAAa;GAAE;;CAG5B,YAAY,IAAI,gBAAgB,SAAS;EAEvC,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,eAAe,GAAG,WAAW;EACnC,MAAM,WAAW,GAAG,WAAW;EAE/B,MAAM,kBAAkB,SAAS,UAAW,GAAG,WAAW,mBAA8B;EAGxF,MAAM,WAAW,eAAe,GAAG,QAAQ;EAC3C,MAAM,IAAI,WAAW,SAAS,KAAM,GAAG,WAAW;EAGlD,MAAM,UAAU,eAAe,GAAG,OAAO;AAgBzC,SAAO,mBAAmB;GACxB;GAhBQ,UAAU,QAAQ,KAAK;GAkB/B;GACA;GACA;GACA;GAlBiB,GAAG,WAAW,WAAuB,QAAQ,IAAI;GAKnD,SAAS,WAAW;GAgBnC,aAXkB,GAAG,WAAW,cAAyB,IAAI,KAAK,KAAK,SAAS,CAWxD;GACzB,CAAC;;CAEL;AAOD,MAAMC,qBAAiC;CACrC,YAAY;CACZ,YAAY;CAOZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,cAAc,GAAG,WAAW;EAElC,MAAM,WAAW,eAAe,GAAG,QAAQ;AAE3C,SAAO;GADG,WAAW,SAAS,KAAM,GAAG,WAAW;GACvC;GAAa;GAAE;;CAG5B,YAAY,IAAI,gBAAgB;EAE9B,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,eAAe,GAAG,WAAW;EACnC,MAAM,WAAW,GAAG,WAAW;EAG/B,MAAM,WAAW,eAAe,GAAG,QAAQ;EAC3C,MAAM,IAAI,WAAW,SAAS,KAAM,GAAG,WAAW;EAGlD,MAAM,SAAS,eAAe,GAAG,OAAO;AAGxC,SAAO,mBAAmB;GAAC;GAFjB,SAAS,OAAO,KAAM,GAAG,WAAW;GAEb;GAAa;GAAc;GAAS,CAAC;;CAEzE;AAID,MAAaC,kCAA8C;CACzD,YAAY;CACZ,YAAY;CAKZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CAEzF,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,QAAQ,GAAG,WAAW;EAC5B,MAAM,SAAS,GAAG,WAAW;EAC7B,IAAIC;AACJ,MAAI,UAAU,eAAe,QAC3B,KAAI,eAAe,QAAQ;MAE3B,KAAI,GAAG,WAAW;AAGpB,SAAO;GAAC,KAAM,IAAI,QAAS,GAAG,IAAI;GAAE;GAAG;GAAE;;CAG3C,YAAY,IAAI,gBAAgB,SAAS;EACvC,MAAM,QAAQ,GAAG,WAAW;EAC5B,MAAM,SAAS,SAAS,UAAU;EAClC,MAAM,SAAS,GAAG,WAAW;EAC7B,IAAIA;AACJ,MAAI,UAAU,eAAe,QAC3B,KAAI,eAAe,QAAQ;MAE3B,KAAI,GAAG,WAAW;AAGpB,SAAO,mBAAmB,CAAC,IAAI,OAAO,SAAS,MAAM,CAAC;;CAEzD;AAID,MAAaC,4BAAwC;CACnD,YAAY;CACZ,YAAY;CAOZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,WAAW,eAAe,GAAG,QAAQ;AAE3C,SAAO;GADG,WAAW,SAAS,KAAM,GAAG,WAAW;GACvC;GAAa;GAAE;;CAG5B,YAAY,IAAI,gBAAgB,SAAS;EACvC,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,eAAe,GAAG,WAAW;EACnC,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,kBAAkB,SAAS,UAAW,GAAG,WAAW,mBAA8B;EAExF,MAAM,WAAW,eAAe,GAAG,QAAQ;EAC3C,MAAM,IAAI,WAAW,SAAS,KAAM,GAAG,WAAW;EAElD,MAAM,UAAU,eAAe,GAAG,OAAO;AAMzC,SAAO,mBAAmB;GACxB;GANQ,UAAU,QAAQ,KAAK;GAQ/B;GACA;GACA;GACA;GACA,aATkB,GAAG,WAAW,cAAyB,IAAI,KAAK,KAAK,SAAS,CASxD;GACzB,CAAC;;CAEL;AAID,MAAMC,cAA0B;CAC9B,YAAY;CACZ,YAAY;CAKZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CAEzF,gBAAgB,IAAI;AAGlB,SAAO;GADU,GAAG,WAAW;GACb;GAAG;GAAE;;CAGzB,YAAY,IAAI;EAEd,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,WAAW,GAAG,WAAW;AAC/B,SAAO,mBAAmB,CAAC,UAAU,SAAS,CAAC;;CAElD;AAID,MAAMC,mBAA+B;CACnC,YAAY;CACZ,YAAY;CAOZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,MAAM,GAAG,WAAW;EAC1B,IAAIf;AACJ,MAAI,OAAO,eAAe,KACxB,WAAU,eAAe,KAAK;MAE9B,WAAU,GAAG,WAAW;AAE1B,SAAO;GAAC,KAAK,UAAU,UAAU,IAAI;GAAE;GAAG;GAAE;;CAG9C,YAAY,IAAI,gBAAgB;EAE9B,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,MAAM,GAAG,WAAW;EAC1B,IAAIA;AACJ,MAAI,OAAO,eAAe,KACxB,WAAU,eAAe,KAAK;MAE9B,WAAU,GAAG,WAAW;AAE1B,SAAO,mBAAmB;GAAC;GAAS;GAAU;GAAa;GAAE,CAAC;;CAEjE;AAID,MAAMgB,uBAAmC;CACvC,YAAY;CACZ,YAAY;CAEZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,MAAM,GAAG,WAAW;EAC1B,IAAIhB;AACJ,MAAI,OAAO,eAAe,KACxB,WAAU,eAAe,KAAK;MAE9B,WAAU,GAAG,WAAW;AAE1B,SAAO;GAAC,KAAK,UAAU,UAAU,IAAI;GAAE;GAAG;GAAE;;CAG9C,YAAY,IAAI,gBAAgB;EAC9B,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,MAAM,GAAG,WAAW;EAC1B,IAAIA;AACJ,MAAI,OAAO,eAAe,KACxB,WAAU,eAAe,KAAK;MAE9B,WAAU,GAAG,WAAW;AAE1B,SAAO,mBAAmB;GAAC;GAAS;GAAU;GAAa;GAAE,CAAC;;CAEjE;AAID,MAAMiB,wBAAoC;CACxC,YAAY;CACZ,YAAY;CAIZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,MAAM,GAAG,WAAW;EAC1B,IAAIjB;AACJ,MAAI,OAAO,eAAe,KACxB,WAAU,eAAe,KAAK;MAE9B,WAAU,GAAG,WAAW;AAE1B,SAAO;GAAC,KAAK,UAAU,UAAU,IAAI;GAAE;GAAG;GAAE;;CAG9C,YAAY,IAAI,gBAAgB;EAC9B,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,MAAM,GAAG,WAAW;EAC1B,IAAIA;AACJ,MAAI,OAAO,eAAe,KACxB,WAAU,eAAe,KAAK;MAE9B,WAAU,GAAG,WAAW;AAE1B,SAAO,mBAAmB;GAAC;GAAS;GAAU;GAAa;GAAE,CAAC;;CAEjE;AAID,MAAMkB,kBAA8B;CAClC,YAAY;CACZ,YAAY;CAMZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAGjC,YAAY,IAAI,gBAAgB;AAG9B,SAAO,mBAAmB,CAFZ,mBAAmB,IAAI,eAAe,CAEnB,CAAC;;CAErC;AAID,MAAMC,eAA2B;CAC/B,YAAY;CACZ,YAAY;CAUZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI;AAElB,SAAO;GADW,GAAG,WAAW;GACb;GAAG;GAAE;;CAG1B,YAAY,IAAI,gBAAgB;EAC9B,MAAM,YAAY,GAAG,WAAW;EAChC,MAAM,UAAU,GAAG,WAAW;EAC9B,MAAM,UAAU,GAAG,WAAW;EAC9B,MAAM,UAAU,GAAG,WAAW;EAE9B,MAAM,SAAS,GAAG,WAAW;EAC7B,IAAIP;AACJ,MAAI,UAAU,eAAe,QAC3B,KAAI,eAAe,QAAQ;MAE3B,KAAI,GAAG,WAAW;AAEpB,SAAO,mBAAmB;GAAC;GAAG;GAAW;GAAS;GAAS;GAAS;GAAG;GAAG;GAAE,CAAC;;CAEhF;AAID,MAAaQ,qBAAiC;CAC5C,GAAG;CACH,YAAY;CACb;AAID,MAAMC,oBAAgC;CACpC,YAAY;CACZ,YAAY;CAKZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CAEzF,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,QAAQ,GAAG,WAAW;EAC5B,MAAM,SAAS,GAAG,WAAW;EAC7B,IAAIT;AACJ,MAAI,UAAU,eAAe,QAC3B,KAAI,eAAe,QAAQ;MAE3B,KAAI,GAAG,WAAW;AAEpB,SAAO;GAAC,KAAK,IAAI,OAAO,IAAI;GAAE;GAAG;GAAE;;CAGrC,YAAY,IAAI,gBAAgB,SAAS;EACvC,MAAM,QAAQ,GAAG,WAAW;EAC5B,MAAM,SAAS,SAAS,UAAU;EAClC,MAAM,SAAS,GAAG,WAAW;EAC7B,IAAIA;AACJ,MAAI,UAAU,eAAe,QAC3B,KAAI,eAAe,QAAQ;MAE3B,KAAI,GAAG,WAAW;AAEpB,SAAO,mBAAmB,CAAC,IAAI,OAAO,SAAS,MAAM,CAAC;;CAEzD;AAID,MAAaU,4BAAwC;CACnD,YAAY;CACZ,YAAY;CAGZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,QAAQ,GAAG,WAAW;EAC5B,MAAM,SAAS,GAAG,WAAW;AAG7B,SAAO;GAAC,MADN,UAAU,eAAe,UAAU,eAAe,QAAQ,KAAM,GAAG,WAAW,KAC/D,OAAO,IAAI;GAAE;GAAG;GAAE;;CAGrC,YAAY,IAAI,gBAAgB,SAAS;EACvC,MAAM,QAAQ,GAAG,WAAW;EAC5B,MAAM,SAAS,SAAS,UAAU;EAClC,MAAM,SAAS,GAAG,WAAW;AAG7B,SAAO,mBAAmB,EADxB,UAAU,eAAe,UAAU,eAAe,QAAQ,KAAM,GAAG,WAAW,KACjD,OAAO,SAAS,MAAM,CAAC;;CAEzD;AAID,MAAaC,gCAA4C;CACvD,GAAG;CACH,YAAY;CACb;AAID,MAAaC,uCAAmD;CAC9D,GAAG;CACH,YAAY;CAEZ,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,QAAQ,GAAG,WAAW;EAC5B,MAAM,SAAS,GAAG,WAAW;AAG7B,SAAO;GAAC,MADN,UAAU,eAAe,UAAU,eAAe,QAAQ,KAAM,GAAG,WAAW,KAC9D,SAAU,GAAG,IAAI;GAAE;GAAG;GAAE;;CAE7C;AAID,MAAaC,2BAAuC;CAClD,YAAY;CACZ,YAAY;CAKZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CAEzF,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,QAAQ,GAAG,WAAW;EAC5B,MAAM,SAAS,GAAG,WAAW;EAC7B,IAAIb;AACJ,MAAI,UAAU,eAAe,QAC3B,KAAI,eAAe,QAAQ;MAE3B,KAAI,GAAG,WAAW;AAEpB,SAAO;GAAC,KAAK,IAAI,OAAO,IAAI;GAAE;GAAG;GAAE;;CAGrC,YAAY,IAAI,gBAAgB,SAAS;EACvC,MAAM,QAAQ,GAAG,WAAW;EAC5B,MAAM,SAAS,SAAS,UAAU;EAClC,MAAM,SAAS,GAAG,WAAW;EAC7B,IAAIA;AACJ,MAAI,UAAU,eAAe,QAC3B,KAAI,eAAe,QAAQ;MAE3B,KAAI,GAAG,WAAW;AAEpB,SAAO,mBAAmB,CAAC,IAAI,OAAO,SAAS,MAAM,CAAC;;CAEzD;AAID,MAAac,qBAAiC;CAC5C,YAAY;CACZ,YAAY;CAOZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,WAAW,eAAe,GAAG,QAAQ;AAE3C,SAAO;GADG,WAAW,SAAS,KAAM,GAAG,WAAW;GACvC;GAAa;GAAE;;CAG5B,YAAY,IAAI,gBAAgB,SAAS;EACvC,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,eAAe,GAAG,WAAW;EACnC,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,kBAAkB,SAAS,UAAW,GAAG,WAAW,mBAA8B;EAExF,MAAM,WAAW,eAAe,GAAG,QAAQ;EAC3C,MAAM,IAAI,WAAW,SAAS,KAAM,GAAG,WAAW;EAElD,MAAM,UAAU,eAAe,GAAG,OAAO;AAMzC,SAAO,mBAAmB;GACxB;GANQ,UAAU,QAAQ,KAAK;GAQ/B;GACA;GACA;GACA;GACA,aATkB,GAAG,WAAW,cAAyB,IAAI,KAAK,KAAK,SAAS,CASxD;GACzB,CAAC;;CAEL;AAID,MAAMC,sBAAkC;CACtC,YAAY;CACZ,YAAY;CAKZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CAEzF,gBAAgB,IAAI;EAClB,MAAM,WAAW,GAAG,WAAW;AAC/B,SAAO;GAAC,KAAK,UAAU,IAAI;GAAE;GAAG;GAAE;;CAGpC,YAAY,IAAI,gBAAgB;EAC9B,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,aAAa,GAAG,WAAW;EACjC,MAAM,SAAS,GAAG,WAAW;EAC7B,IAAI3B;AACJ,MAAI,UAAU,eAAe,QAC3B,WAAU,eAAe,QAAQ;MAEjC,WAAU,GAAG,WAAW;AAE1B,SAAO,mBAAmB;GAAC;GAAS;GAAU;GAAY;GAAE,CAAC;;CAEhE;;;;;;;;AAwBD,MAAM,mBAAmB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqCzB,MAAM,wBAAwB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA0C9B,MAAM,eAAe;;;;;;;;;;;;;;;;;;;AAsBrB,MAAM,YAAY;;;;;;;;;;;;;;;;AAiBlB,MAAM,iBAAiB;;;;;;;;;;;;;;;;;;;;AAqBvB,MAAM,iBAAiB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA4DvB,MAAM,yBAAyB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA+E/B,MAAa4B,yBAAqC;CAChD,YAAY;CACZ,YAAY;CACZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CACD,gBAAgB,IAAI,gBAAgB;EAClC,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,MAAM,GAAG,WAAW;EAC1B,IAAI5B;AACJ,MAAI,OAAO,eAAe,KAExB,WADc,eAAe,KACb,QAAQ,GAAG,MAAM,IAAI,GAAG,EAAE,GAAG;MAE7C,WAAU,GAAG,WAAW;AAE1B,SAAO;GAAC;GAAS;GAAG;GAAE;;CAExB,YAAY,IAAI,gBAAgB;EAC9B,MAAM,cAAc,GAAG,WAAW;EAClC,MAAM,MAAM,GAAG,WAAW;EAC1B,MAAM,MAAM,GAAG,WAAW;EAC1B,IAAIA;AACJ,MAAI,OAAO,eAAe,KAExB,WADc,eAAe,KACb,QAAQ,GAAG,MAAM,IAAI,GAAG,EAAE,GAAG;MAE7C,WAAU,GAAG,WAAW;AAE1B,SAAO,mBAAmB;GAAC;GAAS;GAAa,aAAa,IAAI;GAAE;GAAE,CAAC;;CAE1E;AAED,MAAM6B,WAAuB;CAC3B,YAAY;CACZ,YAAY;CACZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CACzF,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAEjC,YAAY,IAAI,gBAAgB;AAE9B,SAAO,mBAAmB,CADZ,mBAAmB,IAAI,eAAe,CACnB,CAAC;;CAErC;AAED,MAAMC,gBAA4B;CAChC,YAAY;CACZ,YAAY;CACZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CACzF,gBAAgB,IAAI,gBAAgB;AAElC,SAAO;GAAC,KADM,mBAAmB,IAAI,eAAe,EAChC,IAAI;GAAE;GAAG;GAAE;;CAEjC,YAAY,IAAI,gBAAgB;EAE9B,MAAM,UAAU,eAAe,GAAG,OAAO;AAGzC,SAAO,mBAAmB,CAFb,UAAU,QAAQ,KAAM,GAAG,WAAW,MACtC,UAAU,QAAQ,KAAM,GAAG,WAAW,KACb,CAAC;;CAE1C;AAED,MAAMC,gBAA4B;CAChC,YAAY;CACZ,YAAY;CACZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CAED,kBAAkB;AAChB,SAAO;GAAC;GAAG;GAAG;GAAE;;CAElB,YAAY,IAAI,gBAAgB;EAC9B,MAAM,WAAW,GAAG,WAAW;EAC/B,MAAM,UAAU,eAAe,GAAG,OAAO;AAGzC,SAAO,mBAAmB;GAAC;GAFZ,UAAU,QAAQ,KAAM,GAAG,WAAW;GAER,aADhC,GAAG,WAAW,OAAkB,KACiB;GAAE;GAAE,CAAC;;CAEtE;AAED,MAAMC,iBAA6B;CACjC,YAAY;CACZ,YAAY;CACZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CACD,gBAAgB,IAAI;EAClB,MAAM,OAAO,GAAG,WAAW;EAC3B,MAAM,OAAO,GAAG,WAAW;AAC3B,SAAO;GAAC,KAAK,OAAO,MAAM,GAAG;GAAE;GAAG;GAAE;;CAEtC,YAAY,IAAI;AACd,SAAO,mBAAmB;GACxB,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACf,CAAC;;CAEL;AAED,MAAMC,sBAAkC;CACtC,YAAY;CACZ,YAAY;CACZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CACD,gBAAgB,IAAI;EAClB,MAAM,OAAO,GAAG,WAAW;EAC3B,MAAM,OAAO,GAAG,WAAW;AAC3B,SAAO;GAAC,KAAK,OAAO,MAAM,GAAG;GAAE;GAAG;GAAE;;CAEtC,YAAY,IAAI;AACd,SAAO,mBAAmB;GACxB,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACf,CAAC;;CAEL;AAED,MAAMC,cAA0B;CAC9B,YAAY;CACZ,YAAY;CACZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CACD,gBAAgB,IAAI;EAClB,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,IAAI,GAAG,WAAW;AACxB,SAAO;GAAC,KAAK,IAAI,GAAG,GAAG;GAAE;GAAG;GAAE;;CAEhC,YAAY,IAAI;AACd,SAAO,mBAAmB,CAAC,GAAG,WAAW,GAAa,GAAG,WAAW,EAAY,CAAC;;CAEpF;AAiBD,MAAM,mBAAmB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAgCzB,MAAM,oBAAoB;;;;;;;;;;;;;;;;;;;;;;;AAwB1B,MAAM,2BAA2B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAwCjC,MAAMC,iBAA6B;CACjC,YAAY;CACZ,YAAY;CACZ,UAAU;EAAC,EAAE,MAAM,gBAAgB;EAAE,EAAE,MAAM,sBAAsB;EAAE,EAAE,MAAM,WAAW;EAAC;CACzF,gBAAgB,IAAI;EAClB,MAAM,SAAS,GAAG,WAAW;EAC7B,MAAM,OAAO,GAAG,WAAW;EAC3B,MAAM,IAAI,GAAG,WAAW;AACxB,SAAO;GAAC,KAAK,SAAS,OAAO,GAAG,GAAG;GAAE;GAAG;GAAE;;CAE5C,YAAY,IAAI;EACd,MAAM,SAAS,GAAG,WAAW;EAC7B,MAAM,OAAO,GAAG,WAAW;AAC3B,SAAO,mBAAmB;GACxB,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACd;GACA,OAAO;GACP,OAAO;GACP,OAAO;GACP,OAAO;GACP,KAAK;GACL,KAAK;GACL,KAAK;GACL,KAAK;GACN,CAAC;;CAEL;AAED,MAAMC,kBAA8B;CAClC,YAAY;CACZ,YAAY;CACZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CACD,gBAAgB,IAAI;EAClB,MAAM,IAAI,GAAG,WAAW;EACxB,MAAM,IAAI,GAAG,WAAW;AACxB,SAAO;GAAC,KAAK,IAAI,GAAG,GAAG;GAAE;GAAG;GAAE;;CAEhC,YAAY,IAAI;EACd,MAAM,IAAI,GAAG,WAAW;AACxB,SAAO,mBAAmB;GAAC;GAAG,GAAG,WAAW;GAAa,KAAK,MAAM,IAAI,EAAE;GAAE;GAAE,CAAC;;CAElF;AAED,MAAMC,+BAA2C;CAC/C,YAAY;CACZ,YAAY;CACZ,UAAU;EACR,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,gBAAgB;EACxB,EAAE,MAAM,sBAAsB;EAC9B,EAAE,MAAM,WAAW;EACpB;CACD,gBAAgB,IAAI;EAClB,MAAM,OAAO,GAAG,WAAW;EAC3B,MAAM,OAAO,GAAG,WAAW;AAC3B,SAAO;GAAC,KAAK,OAAO,MAAM,GAAG;GAAE;GAAG;GAAE;;CAEtC,YAAY,IAAI;EACd,MAAM,MAAM,GAAG,WAAW;EAC1B,MAAM,OAAO,GAAG,WAAW;AAC3B,SAAO,mBAAmB;GACxB;GACA;GACA,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACd,GAAG,WAAW;GACd,KAAK,MAAM,MAAM,KAAK;GACtB;GACD,CAAC;;CAEL;AAED,MAAaC,kBAAuD;CAClE,WAAW;CACX,eAAe;CACf,QAAQ;CACR,YAAY;CACZ,YAAY;CACZ,KAAK;CACL,KAAK;CACL,QAAQ;CACR,MAAM;CACN,iBAAiB;CACjB,MAAM;CACN,SAAS;CACT,SAAS;CACT,gBAAgB;CAChB,OAAO;CACP,aAAa;CACb,WAAW;CACX,SAAS;CACT,YAAY;CACZ,eAAe;CACf,SAAS;CACT,WAAW;CACX,MAAM;CACN,WAAW;CACX,gBAAgB;CAChB,SAAS;CACT,cAAc;CACd,kBAAkB;CAClB,mBAAmB;CACnB,aAAa;CACb,UAAU;CACV,eAAe;CACf,iBAAiB;CACjB,cAAc;CACd,UAAU;CACV,OAAO;CACP,SAAS;CACT,QAAQ;CAER,YAAY;CACZ,iBAAiB;CACjB,SAAS;CAET,YAAY;CACZ,aAAa;CACb,0BAA0B;CAE1B,MAAM;CACN,WAAW;CACX,WAAW;CACZ;;;;ACtjOD,MAAMC,kBAAgB;AA8CtB,IAAa,WAAb,MAAa,SAAS;CACpB,AAAQ;CACR,AAAQ;CAER,AAAQ,gCAAwC,IAAI,KAAK;CACzD,AAAQ,oCAA4C,IAAI,KAAK;CAC7D,AAAQ,kCAA0C,IAAI,KAAK;CAC3D,AAAQ,iCAAyC,IAAI,KAAK;;CAG1D,AAAQ;;;;;;;CAQR,AAAQ,YAcG;;CAEX,AAAQ,aAAkC;;;;;;;CAO1C,AAAQ,WAAiC;;;;;;;;;CAUzC,AAAQ,uBAAyC;;CAEjD,AAAQ;;CAER,AAAQ;;CAER,AAAQ;;CAER,AAAQ,kBAA+B,EAAE;;;;;;;;;CAUzC,AAAQ,uBAAyC;CACjD,AAAQ,yBAAiC;;CAGzC,AAAQ,kBAAmC,EAAE;;CAE7C,AAAQ,gBAAiC,EAAE;;CAE3C,AAAQ,cAIG;;CAGX,AAAS;CAET,AAAQ;CACR,AAAQ;CACR,AAAQ,SAAiB;CACzB,AAAQ;CAKR,AAAQ,iBAAiB;CACzB,AAAiB,8BAAc,IAAI,KAA4C;CAC/E,AAAQ,WAA+B;CACvC,AAAQ,kBAAoC;CAC5C,AAAQ,mBAAqC;CAE7C,YAAY,KAAiB,OAAmB,SAA0B;AACxE,OAAK,MAAM;AACX,OAAK,QAAQ;AACb,OAAK,YAAY,QAAQ;AACzB,OAAK,SAAS,QAAQ,UAAU;AAChC,OAAK,kBAAkB,KAAK,IAAI,GAAG,QAAQ,mBAAmB,EAAE;AAChE,OAAK,iBACH,IAAI,gBAAgB,OAAO,YAAY,eAAe,QAAQ,KAAK,kBAAkB;AAKvF,OAAK,oBAAoB,IAAI;AAG7B,OAAK,iBAAiB,oBAAoB,KAAK,aAAa,QAAQ,YAAY,EAAE;AAElF,OAAK,2BAA2B;AAChC,OAAK,yBAAyB;AAC9B,OAAK,wBAAwB;AAI7B,OAAK,iBAAiB,qBACpB,KACA,mBACA,KAAK,IAAI,GAAG,MAAM,OAAO,aAAa,EAAE,CACzC;AACD,OAAK,qBAAqB,oBAAoB,KAAK,iBAAiB,EAAE;AACtE,OAAK,iBAAiB,qBAAqB,KAAK,mBAAmB,EAAE;;;;;;;;;CAUvE,aAAa,KAcJ;AACP,OAAK,YAAY;;;;;;;CAQnB,AAAQ,kBAAiC;EACvC,MAAM,MAAM,KAAK;AACjB,MAAI,CAAC,OAAO,CAAC,IAAI,MAAO,QAAO,QAAQ,SAAS;AAChD,MAAI,KAAK,SAAU,QAAO,KAAK;EAC/B,MAAM,EAAE,WAAW,WAAW,WAAW,aAAa,IAAI;AAC1D,OAAK,YAAY,YAAY;GAC3B,MAAM,QAAQ,MAAM,OAAO,KAAK,UAAU;GAC1C,MAAM,OAAO,OAAO,QAAsC;IACxD,MAAM,OAAO,MAAM,MAAM,MAAM,IAAI,QAAQ,IAAI,CAAC;AAChD,QAAI,CAAC,KAAM,OAAM,IAAI,MAAM,4BAA4B,MAAM;AAC7D,WAAO,KAAK,aAAa;;GAE3B,MAAM,CAAC,MAAM,MAAM,QAAQ,MAAM,QAAQ,IAAI;IAC3C,KAAK,UAAU;IACf,KAAK,UAAU;IACf,KAAK,SAAS;IACf,CAAC;AACF,OAAI,SAAS,IAAI,YAAY,KAAK;AAClC,OAAI,SAAS,IAAI,aAAa,KAAK;AACnC,OAAI,QAAQ,IAAI,aAAa,KAAK;MAChC;AACJ,SAAO,KAAK;;;;;;;CAQd,MAAc,cAAc,UAAsC;EAChE,MAAM,MAAM,KAAK;AACjB,MAAI,CAAC,IAAK;AACV,MAAI,IAAI,MAAO,OAAM,KAAK,iBAAiB;EAC3C,MAAM,IAAI,SAAS;EACnB,MAAM,EAAE,QAAQ,QAAQ,OAAO,OAAO,cAAc;EACpD,MAAM,SAAS,IAAI;AACnB,MAAI,CAAC,KAAK,cAAc,KAAK,WAAW,SAAS,OAC/C,MAAK,aAAa,IAAI,aAAa,OAAO;EAE5C,MAAM,MAAM,KAAK;AACjB,OAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;GAE1B,MAAM,UADU,SAAS,KACC;GAC1B,MAAM,UAAU,IAAI;AACpB,QAAK,IAAI,IAAI,GAAG,IAAI,OAAO,KAAK;IAC9B,MAAM,OAAO,UAAU;IACvB,MAAM,YAAY,SAAS;IAC3B,MAAM,aAAa,OAAO,KAAK;IAC/B,MAAM,SAAU,OAAO,eAAe,YAAa;IACnD,MAAM,WAAY,OAAO,YAAa;AACtC,QAAI,UAAU,MAAM,SAAS,MAAM,aAAa,OAAO;;;EAG3D,MAAM,SAAS,KAAK,UAAU,IAAI,aAAa;AAC/C,MAAI,CAAC,OACH,OAAM,IAAI,MAAM,sBAAsB,IAAI,aAAa,qBAAqB;AAE9E,OAAK,IAAI,OAAO,MAAM,YAAY,QAAQ,GAAG,IAAI,QAAQ,IAAI,YAAY,SAAS,EAAE;;;;;;;;;;;;;;;;;CAkBtF,MAAM,cACJ,QACe;AACf,MAAI,kBAAkB,KAAK;AACzB,QAAK,iBAAiB,OAAO;AAC7B;;AAEF,OAAK,MAAM,QAAQ,OAAO,MAAM,EAAE;GAChC,MAAM,QAAQ,MAAM,OAAO,IAAI,KAAK;AACpC,OAAI,CAAC,MAAO;GACZ,MAAM,EAAE,SAAS;GACjB,MAAM,SAAS,oBAAoB,KAAK,KAAK,UAAU,QAAQ,KAAK,YAAY,KAAK;AACrF,QAAK,cAAc,IAAI,MAAM,OAAO;;;;;;;;CAWxC,iBAAiB,SAAwE;EACvF,MAAM,QAAQ,CAAC,GAAG,QAAQ,MAAM,CAAC;AACjC,OAAK,MAAM,QAAQ,OAAO;GACxB,MAAM,QAAQ,QAAQ,IAAI,KAAK;AAC/B,OAAI,CAAC,MAAO;GACZ,MAAM,SAAS,oBACb,KAAK,KACL,UAAU,QACV,MAAM,KAAK,YACX,MAAM,KACP;AACD,QAAK,cAAc,IAAI,MAAM,OAAO;AACpC,WAAQ,OAAO,KAAK;;;;;;;;;;CAWxB,iBAAuB;AACrB,OAAK,MAAM,UAAU,KAAK,MAAM,gBAAgB;GAC9C,MAAM,OAAO,KAAK,MAAM,MAAM,MAAM,MAAM,EAAE,OAAO,OAAO;GAC1D,IAAI,OAAO,gBAAgB,KAAK;AAChC,OAAI,CAAC,KAAM,OAAM,IAAI,MAAM,0BAA0B,KAAK,SAAS;AAInE,OAAI,KAAK,WAAW,iBAElB;QADiB,KAAK,QAAQ,MAAM,QAAQ,KAAK,MAAM,QAAQ,MAAM,UAAU,MAAM,CAEnF,QACE,KAAK,WAAW,eACZ,kCACA;cAEC,KAAK,WAAW,aAEzB;QADiB,KAAK,OAAO,MAAM,QAAQ,KAAK,MAAM,QAAQ,MAAM,UAAU,MAAM,CAElF,QAAO,KAAK,WAAW,eAAe,4BAA4B;cAE3D,KAAK,WAAW,cAAc,KAAK,IAAI,UAAU,CAAC,KAAK,IAAI,eAGpE,QAAO;YACE,KAAK,WAAW,eAAe,KAAK,OAAO,WAAW,EAI/D,QAAO;YACE,KAAK,WAAW,UAAU,KAAK,WAAW,gBAAgB,KAGnE,QAAO;GAGT,MAAM,WAAW,oBACf,KAAK,KACL,UAAU,UACV,KAAK,YACL,KAAK,WACN;GAGD,MAAM,cAAc,KAAK,cAAc,EAAE;GACzC,MAAM,aAAa,KAAK,YAAY,MAAM,aAAa,EAAE,QAAQ,GAAG,CAAC;GACrE,MAAM,gBAAgB,oBAAoB,KAAK,KAAK,WAAW,UAAU,WAAW;GAGpF,MAAM,gBAAgB,KAAK,cAAc,MAAM,MAAM,cAAc;GACnE,MAAM,YAAY,gBAAgB,KAAK,KAAK,UAAU,eAAe,MAAM,SAAS;GAEpF,MAAMC,eAA8B;IAClC;IACA;IACA;IACA;IACA;IACA;IACA,iBAAiB;IACjB,kBAAkB;IACnB;AACD,QAAK,gBAAgB,KAAK,aAAa;AAIvC,OADiB,KAAK,WAAW,YAAY,KAAK,WAAW,cAC/C;IAYZ,MAAM,eAFH,WAA0E,SAAS,KAChF,4BAA4B,OACK,KAAK,IAAI,gBAAgB,CAAC,KAAK,IAAI;IAC1E,MAAM,SACJ,KAAK,WAAW,eACZ,eACE,6BACA,mBACF,eACE,wBACA;IAER,MAAM,aAAa,oBACjB,KAAK,KACL,UAAU,UACV,OAAO,YACP,OAAO,WACR;IACD,MAAM,gBAAgB,OAAO,YAAY,MAAM,aAAa,EAAE,QAAQ,GAAG,CAAC;IAC1E,MAAM,YAAY,oBAAoB,KAAK,KAAK,cAAc,UAAU,cAAc;IACtF,MAAM,YAAY,KAAK,cAAc,QAAQ,MAAM,UAAU;IAC7D,MAAM,cAAc,gBAAgB,KAAK,KAAK,YAAY,WAAW,SAAS,SAAS;AAEvF,SAAK,cAAc,KAAK;KACtB;KACA;KACA,MAAM;KACN,UAAU;KACV,WAAW;KACX,eAAe;KACf,iBAAiB;KACjB,kBAAkB;KACnB,CAAC;SAGF,MAAK,cAAc,KAAK,aAAa;;AAOzC,OAAK,yBAAyB;AAM9B,OAAK,6BAA6B;AAGlC,OAAK,8BAA8B;AAGnC,OAAK,6BAA6B;AAGlC,OAAK,kCAAkC;AAGvC,OAAK,8BAA8B;AAKnC,UAAQ,IAAI,sBAAsB,KAAK,cAAc,OAAO,mBAAmB;EAG/E,MAAM,eAAe,KAAK,UAAU,SAAS;AAC7C,MAAI,cAAc;GAChB,MAAM,iBAAiB,oBACrB,KAAK,KACL,UACA,YAAY,YACZ,YAAY,WACb;GACD,MAAM,gBAAgB,oBACpB,KAAK,KACL,kCACA,IAAI,YAAY,GAAG,CACpB;AAOD,QAAK,cAAc;IACjB,UAAU;IACV,WARsB,gBACtB,KAAK,KACL,gBACA;KAAC,EAAE,QAAQ,cAAc;KAAE,EAAE,QAAQ,KAAK,oBAAoB;KAAE,EAAE,QAAQ,eAAe;KAAC,EAC1F,YACD;IAIC,eAAe;IAChB;;AAGH,MAAI,KAAK,kBACP,SAAQ,IAAI,gEAAgE;EAK9E,IAAI,eAAe;AACnB,OAAK,MAAM,SAAS,KAAK,gBACvB,iBAAgB,KAAK,KAAK,MAAM,cAAc,OAAO,EAAE,GAAG;EAE5D,IAAI,cAAc;AAClB,OAAK,MAAM,SAAS,KAAK,cACvB,gBAAe,KAAK,KAAK,MAAM,cAAc,OAAO,EAAE,GAAG;AAE3D,iBAAe;EAEf,MAAM,eAAe,KAAK,YAAY;AACtC,OAAK,yBAAyB,KAAK,IAAI,cAAc,YAAY,GAAG;AACpE,OAAK,uBAAuB,KAAK,IAAI,OAAO,aAAa;GACvD,OAAO;GACP,MAAM,KAAK;GACX,OAAO;GACR,CAAC;;;;;CAMJ,MAAM,QAAQ,UAA+C;EAC3D,MAAM,IAAI,SAAS;EAEnB,MAAM,iBAAiB,KAAK,cAAc,EAAE;EAC5C,MAAMC,iBAAiC,EAAE,QAAQ,KAAK,QAAQ;EAG9D,MAAM,UAAU,MAAM,IAAI,KAAK,gBAAgB,KAAK;AAGpD,OAAK,IAAI,OAAO,MAAM,YAAY,KAAK,gBAAgB,GAAG,SAAyB;AAGnF,QAAM,KAAK,cAAc,SAAS;EAIlC,MAAMC,gBAAiD,IAAI,MAAM,QAAQ,OAAO;AAChF,OAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,QAAQ,KAAK;GACvC,MAAM,QAAQ,QAAQ;GACtB,MAAM,aAAa,MAAM,KAAK,YAAY,MAAM,MAAM,gBAAgB,eAAe;GACrF,MAAM,aAAa,IAAI,WAAW,WAAW;GAE7C,IAAI,UAAU,CAAC,MAAM,mBAAmB,MAAM,gBAAgB,WAAW,WAAW;AACpF,OAAI,CAAC,SAAS;IACZ,MAAM,SAAS,MAAM;AACrB,SAAK,IAAI,IAAI,GAAG,IAAI,WAAW,QAAQ,IACrC,KAAI,WAAW,OAAO,OAAO,IAAI;AAC/B,eAAU;AACV;;;AAKN,OAAI,SAAS;AACX,SAAK,IAAI,OAAO,MAAM,YAAY,MAAM,eAAe,GAAG,WAAW;AACrE,UAAM,kBAAkB,IAAI,WAAW,WAAW,MAAM,EAAE,CAAC;;AAG7D,iBAAc,KAAK,MAAM,KAAK,gBAAgB,MAAM,MAAM,gBAAgB,eAAe;;AAG3F,MAAI,KAAK,eACP,OAAM,KAAK,sBAAsB,SAAS,cAAc;WAC/C,KAAK,mBAAmB;GASjC,MAAM,QAAQ,KAAK;AACnB,QAAK,IAAI,QAAQ,GAAG,QAAQ,QAAQ,QAAQ,SAAS,OAAO;IAC1D,MAAM,MAAM,KAAK,IAAI,OAAO,sBAAsB;IAClD,MAAM,MAAM,KAAK,IAAI,QAAQ,OAAO,QAAQ,OAAO;AACnD,SAAK,IAAI,IAAI,OAAO,IAAI,KAAK,KAAK;KAChC,MAAM,IAAI,IAAI,kBAAkB;AAChC,OAAE,YAAY,QAAQ,GAAG,SAAS;AAClC,OAAE,aAAa,GAAG,QAAQ,GAAG,UAAU;AACvC,OAAE,mBAAmB,GAAG,cAAc,GAAG;AACzC,OAAE,KAAK;;AAET,SAAK,IAAI,OAAO,MAAM,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;AAC5C,UAAM,KAAK,IAAI,OAAO,MAAM,qBAAqB;;GAEnD,MAAM,eAAe,KAAK,UAAU,SAAS;AAC7C,OAAI,cAAc;IAChB,MAAM,UAAU,KAAK,IAAI,OAAO,sBAAsB;IACtD,MAAMC,cAAY,KAAK,MAAM,OAAO;AAEpC,YAAQ,mBAAmB,cAAc,GAAG,KAAK,gBAAgB,GAAGA,cAAY,EAAE;AAClF,SAAK,IAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;;SAE7C;GAEL,MAAM,UAAU,KAAK,IAAI,OAAO,qBAAqB,EAAE,OAAO,WAAW,CAAC;GAC1E,MAAM,OAAO,QAAQ,iBAAiB,EAAE,OAAO,YAAY,CAAC;AAC5D,QAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,QAAQ,KAAK;AACvC,SAAK,YAAY,QAAQ,GAAG,SAAS;AACrC,SAAK,aAAa,GAAG,QAAQ,GAAG,UAAU;AAC1C,SAAK,mBAAmB,GAAG,cAAc,GAAG;;AAE9C,QAAK,KAAK;GACV,MAAM,eAAe,KAAK,UAAU,SAAS;AAC7C,OAAI,cAAc;IAChB,MAAMA,cAAY,KAAK,MAAM,OAAO;AAEpC,YAAQ,mBAAmB,cAAc,GAAG,KAAK,gBAAgB,GAAGA,cAAY,EAAE;;AAEpF,QAAK,IAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;;EAGlD,MAAM,YAAY,KAAK,MAAM,OAAO;AACpC,QAAM,KAAK,eAAe,SAASJ,iBAAe,GAAG,YAAY,EAAE;EACnE,MAAM,SAAS,KAAK,eAAe,eAAe,GAAG,YAAY,EAAE;EACnE,MAAM,SAAS,IAAI,aAAa,OAAO,MAAM,EAAE,CAAC;AAChD,OAAK,eAAe,OAAO;AAE3B,OAAK,UAAU;AACf,SAAO,EAAE,QAAQ;;;;;;;;CASnB,MAAc,sBACZ,SACA,eACA,aAAa,MACE;EACf,MAAM,IAAI,QAAQ;EAClB,MAAM,aAAa,IAAI;AACvB,MAAI,CAAC,KAAK,UAAU;AAClB,QAAK,WAAW,KAAK,IAAI,OAAO,eAAe;IAAE,MAAM;IAAa,OAAO;IAAY,CAAC;AAExF,QAAK,kBAAkB,KAAK,IAAI,OAAO,aAAa;IAClD,MAAM,aAAa;IACnB,OAAO;IACR,CAAC;AAEF,QAAK,mBAAmB,KAAK,IAAI,OAAO,aAAa;IACnD,MAAM,aAAa;IACnB,OAAO;IACR,CAAC;;EAEJ,MAAM,WAAW,KAAK;EACtB,MAAM,aAAa,KAAK;EACxB,MAAM,cAAc,KAAK;EAEzB,MAAM,UAAU,KAAK,IAAI,OAAO,qBAAqB,EAAE,OAAO,gBAAgB,CAAC;AAC/E,OAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;GAC1B,MAAM,OAAO,QAAQ,iBAAiB,EACpC,iBAAiB;IACf;IACA,2BAA2B,IAAI;IAC/B,qBAAqB,IAAI,IAAI;IAC9B,EACF,CAAC;AACF,QAAK,YAAY,QAAQ,GAAG,SAAS;AACrC,QAAK,aAAa,GAAG,QAAQ,GAAG,UAAU;AAC1C,QAAK,mBAAmB,cAAc,GAAG,IAAI,cAAc,GAAG,IAAI,cAAc,GAAG,GAAG;AACtF,QAAK,KAAK;;AAEZ,MAAI,YAAY;GACd,MAAM,eAAe,KAAK,UAAU,SAAS;GAC7C,MAAM,YAAY,KAAK,MAAM,OAAO;AACpC,OAAI,aACF,SAAQ,mBAAmB,cAAc,GAAG,KAAK,gBAAgB,GAAG,YAAY,EAAE;;AAGtF,UAAQ,gBAAgB,UAAU,GAAG,YAAY,YAAY,EAAE;AAC/D,UAAQ,mBAAmB,YAAY,GAAG,aAAa,GAAG,aAAa,EAAE;AACzE,OAAK,IAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;AAEhD,QAAM,YAAY,SAASA,iBAAe,GAAG,aAAa,EAAE;EAC5D,MAAM,KAAK,IAAI,eAAe,YAAY,eAAe,GAAG,aAAa,EAAE,CAAC,MAAM,EAAE,CAAC;AACrF,cAAY,OAAO;AACnB,OAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;GAC1B,MAAM,KAAK,OAAO,GAAG,IAAI,IAAI,KAAK,GAAG,IAAI,GAAG;AAC5C,OAAI,MAAM,EACR;GAEF,MAAM,MAAM,QAAQ,GAAG,KAAK;GAC5B,MAAM,MAAM,KAAK,YAAY,IAAI,IAAI,IAAI;IAAE,IAAI;IAAG,OAAO;IAAG;AAC5D,OAAI,MAAM;AACV,OAAI,SAAS;AACb,QAAK,YAAY,IAAI,KAAK,IAAI;;;;CAKlC,aAAmE;AACjE,SAAO,CAAC,GAAG,KAAK,YAAY,SAAS,CAAC,CACnC,KAAK,CAAC,QAAQ,QAAQ;GAAE;GAAQ,IAAI,EAAE;GAAI,OAAO,EAAE;GAAO,EAAE,CAC5D,MAAM,GAAG,MAAM,EAAE,KAAK,EAAE,GAAG;;;CAIhC,eAAqB;AACnB,OAAK,YAAY,OAAO;;;;CAK1B,IAAI,sBAA8B;AAChC,SAAO,KAAK,cAAc;;;;;CAM5B,IAAI,oBAA4B;AAC9B,SAAO,KAAK,IAAI,OAAO;;;;;;;;;CAUzB,MAAM,kBAAkB,SAAgC;EACtD,MAAM,iBAAiB,KAAK,cAAc,EAAE;EAC5C,MAAME,iBAAiC,EAAE,QAAQ,KAAK,QAAQ;AAC9D,OAAK,IAAI,OAAO,MAAM,YAAY,KAAK,gBAAgB,GAAG,IAAI,YAAY,CAAC,QAAQ,CAAC,CAAC;EAErF,MAAMG,gBAA4B,IAAI,MAAM,KAAK,cAAc,OAAO;AACtE,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,cAAc,QAAQ,KAAK;GAClD,MAAM,QAAQ,KAAK,cAAc;GACjC,MAAM,aAAa,MAAM,KAAK,YAAY,MAAM,MAAM,gBAAgB,eAAe;AACrF,QAAK,IAAI,OAAO,MAAM,YAAY,MAAM,eAAe,GAAG,WAAW;AACrE,iBAAc,KAAK,MAAM,KAAK,gBAAgB,MAAM,MAAM,gBAAgB,eAAe;;AAE3F,QAAM,KAAK,sBAAsB,KAAK,eAAe,eAAe,MAAM;AAC1E,OAAK,UAAU;;;;;;;;;;CAWjB,MAAM,MAAM,UAA8C;EACxD,MAAM,kBAAkB,KAAK,UAAU,YAAY;AACnD,MAAI,CAAC,gBACH,OAAM,IAAI,MACR,qHAED;EAGH,MAAM,UADa,KAAK,MAAM,OAAO,cACR;EAE7B,MAAM,IAAI,SAAS;EACnB,MAAM,iBAAiB,KAAK,cAAc,EAAE;EAC5C,MAAMH,iBAAiC,EAAE,QAAQ,KAAK,QAAQ;EAG9D,MAAM,UAAU,MAAM,IAAI,KAAK,gBAAgB,KAAK;AAEpD,OAAK,IAAI,OAAO,MAAM,YAAY,KAAK,gBAAgB,GAAG,SAAyB;EAEnF,MAAMC,gBAAiD,IAAI,MAAM,QAAQ,OAAO;AAChF,OAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,QAAQ,KAAK;GACvC,MAAM,QAAQ,QAAQ;GACtB,MAAM,aAAa,MAAM,KAAK,YAAY,MAAM,MAAM,gBAAgB,eAAe;GACrF,MAAM,aAAa,IAAI,WAAW,WAAW;GAE7C,IAAI,UAAU,CAAC,MAAM,mBAAmB,MAAM,gBAAgB,WAAW,WAAW;AACpF,OAAI,CAAC,SAAS;IACZ,MAAM,SAAS,MAAM;AACrB,SAAK,IAAI,IAAI,GAAG,IAAI,WAAW,QAAQ,IACrC,KAAI,WAAW,OAAO,OAAO,IAAI;AAC/B,eAAU;AACV;;;AAIN,OAAI,SAAS;AACX,SAAK,IAAI,OAAO,MAAM,YAAY,MAAM,eAAe,GAAG,WAAW;AACrE,UAAM,kBAAkB,IAAI,WAAW,WAAW,MAAM,EAAE,CAAC;;AAG7D,iBAAc,KAAK,MAAM,KAAK,gBAAgB,MAAM,MAAM,gBAAgB,eAAe;;EAI3F,MAAM,WAAW,KAAK,IAAI,OAAO,aAAa;GAAE,MAAM;GAAS,OAAO;GAAiB,CAAC;AAExF,MAAI,KAAK,mBAAmB;GAC1B,MAAM,QAAQ,KAAK;AACnB,QAAK,IAAI,QAAQ,GAAG,QAAQ,QAAQ,QAAQ,SAAS,OAAO;IAC1D,MAAM,MAAM,KAAK,IAAI,OAAO,sBAAsB;IAClD,MAAM,MAAM,KAAK,IAAI,QAAQ,OAAO,QAAQ,OAAO;AACnD,SAAK,IAAI,IAAI,OAAO,IAAI,KAAK,KAAK;KAChC,MAAM,IAAI,IAAI,kBAAkB;AAChC,OAAE,YAAY,QAAQ,GAAG,SAAS;AAClC,OAAE,aAAa,GAAG,QAAQ,GAAG,UAAU;AACvC,OAAE,mBAAmB,GAAG,cAAc,GAAG;AACzC,OAAE,KAAK;;AAET,SAAK,IAAI,OAAO,MAAM,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;AAC5C,UAAM,KAAK,IAAI,OAAO,MAAM,qBAAqB;;GAEnD,MAAM,UAAU,KAAK,IAAI,OAAO,sBAAsB;AACtD,WAAQ,mBAAmB,iBAAiB,GAAG,UAAU,GAAG,QAAQ;AACpE,QAAK,IAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;SAC3C;GACL,MAAM,UAAU,KAAK,IAAI,OAAO,qBAAqB,EAAE,OAAO,SAAS,CAAC;GACxE,MAAM,OAAO,QAAQ,iBAAiB,EAAE,OAAO,cAAc,CAAC;AAC9D,QAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,QAAQ,KAAK;AACvC,SAAK,YAAY,QAAQ,GAAG,SAAS;AACrC,SAAK,aAAa,GAAG,QAAQ,GAAG,UAAU;AAC1C,SAAK,mBAAmB,GAAG,cAAc,GAAG;;AAE9C,QAAK,KAAK;AACV,WAAQ,mBAAmB,iBAAiB,GAAG,UAAU,GAAG,QAAQ;AACpE,QAAK,IAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;;AAGlD,QAAM,SAAS,SAASH,iBAAe,GAAG,QAAQ;EAClD,MAAM,MAAM,IAAI,aAAa,SAAS,eAAe,GAAG,QAAQ,CAAC,MAAM,EAAE,CAAC;AAC1E,WAAS,OAAO;AAChB,WAAS,SAAS;AAElB,OAAK,UAAU;AACf,SAAO;;;;;;;;;;CAWT,MAAM,eAAe,YAAoB,WAA0C;EACjF,MAAM,YAAY,KAAK,UAAU,WAAW;AAC5C,MAAI,CAAC,UACH,OAAM,IAAI,MAAM,oDAAoD,WAAW,IAAI;EAErF,MAAM,UAAU,YAAY;EAI5B,MAAM,iBAAiB,KAAK,cAAc,EAAE;EAC5C,MAAME,iBAAiC,EAAE,QAAQ,KAAK,QAAQ;EAC9D,MAAM,UAAU,KAAK;EAErB,MAAMC,gBAAiD,IAAI,MAAM,QAAQ,OAAO;AAChF,OAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,QAAQ,KAAK;GACvC,MAAM,QAAQ,QAAQ;GACtB,MAAM,aAAa,MAAM,KAAK,YAAY,MAAM,MAAM,gBAAgB,eAAe;AACrF,QAAK,IAAI,OAAO,MAAM,YAAY,MAAM,eAAe,GAAG,WAAW;AACrE,SAAM,kBAAkB,IAAI,WAAW,IAAI,WAAW,WAAW,CAAC,MAAM,EAAE,CAAC;AAC3E,iBAAc,KAAK,MAAM,KAAK,gBAAgB,MAAM,MAAM,gBAAgB,eAAe;;EAG3F,MAAM,WAAW,KAAK,IAAI,OAAO,aAAa;GAAE,MAAM;GAAS,OAAO;GAAiB,CAAC;AAExF,MAAI,KAAK,mBAAmB;GAC1B,MAAM,QAAQ,KAAK;AACnB,QAAK,IAAI,QAAQ,GAAG,QAAQ,QAAQ,QAAQ,SAAS,OAAO;IAC1D,MAAM,MAAM,KAAK,IAAI,OAAO,sBAAsB;IAClD,MAAM,MAAM,KAAK,IAAI,QAAQ,OAAO,QAAQ,OAAO;AACnD,SAAK,IAAI,IAAI,OAAO,IAAI,KAAK,KAAK;KAChC,MAAM,OAAO,IAAI,kBAAkB;AACnC,UAAK,YAAY,QAAQ,GAAG,SAAS;AACrC,UAAK,aAAa,GAAG,QAAQ,GAAG,UAAU;AAC1C,UAAK,mBAAmB,GAAG,cAAc,GAAG;AAC5C,UAAK,KAAK;;AAEZ,SAAK,IAAI,OAAO,MAAM,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;AAC5C,UAAM,KAAK,IAAI,OAAO,MAAM,qBAAqB;;GAEnD,MAAM,UAAU,KAAK,IAAI,OAAO,sBAAsB;AACtD,WAAQ,mBAAmB,WAAW,GAAG,UAAU,GAAG,QAAQ;AAC9D,QAAK,IAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;SAC3C;GACL,MAAM,UAAU,KAAK,IAAI,OAAO,qBAAqB,EAAE,OAAO,kBAAkB,CAAC;GACjF,MAAM,OAAO,QAAQ,iBAAiB,EAAE,OAAO,YAAY,CAAC;AAC5D,QAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,QAAQ,KAAK;AACvC,SAAK,YAAY,QAAQ,GAAG,SAAS;AACrC,SAAK,aAAa,GAAG,QAAQ,GAAG,UAAU;AAC1C,SAAK,mBAAmB,GAAG,cAAc,GAAG;;AAE9C,QAAK,KAAK;AACV,WAAQ,mBAAmB,WAAW,GAAG,UAAU,GAAG,QAAQ;AAC9D,QAAK,IAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;;AAGlD,QAAM,SAAS,SAASH,iBAAe,GAAG,QAAQ;EAClD,MAAM,MAAM,IAAI,aAAa,SAAS,eAAe,GAAG,QAAQ,CAAC,MAAM,EAAE,CAAC;AAC1E,WAAS,OAAO;AAChB,WAAS,SAAS;AAClB,SAAO;;;;;;CAOT,MAAM,cAAc,UAAwC;EAC1D,MAAM,IAAI,SAAS;EAEnB,MAAM,iBAAiB,KAAK,cAAc,EAAE;EAC5C,MAAME,iBAAiC,EAAE,QAAQ,KAAK,QAAQ;AAG9D,OAAK,IAAI,OAAO,MAAM,YAAY,KAAK,gBAAgB,GAAG,SAAyB;EAGnF,MAAMI,sBAAuD,IAAI,MAC/D,KAAK,cAAc,OACpB;AACD,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,cAAc,QAAQ,KAAK;GAClD,MAAM,QAAQ,KAAK,cAAc;GACjC,MAAM,aAAa,MAAM,KAAK,YAAY,MAAM,MAAM,gBAAgB,eAAe;GACrF,MAAM,aAAa,IAAI,WAAW,WAAW;GAE7C,IAAI,UAAU,CAAC,MAAM,mBAAmB,MAAM,gBAAgB,WAAW,WAAW;AACpF,OAAI,CAAC,SAAS;IACZ,MAAM,SAAS,MAAM;AACrB,SAAK,IAAI,IAAI,GAAG,IAAI,WAAW,QAAQ,IACrC,KAAI,WAAW,OAAO,OAAO,IAAI;AAC/B,eAAU;AACV;;;AAKN,OAAI,SAAS;AACX,SAAK,IAAI,OAAO,MAAM,YAAY,MAAM,eAAe,GAAG,WAAW;AACrE,UAAM,kBAAkB,IAAI,WAAW,WAAW,MAAM,EAAE,CAAC;;AAG7D,uBAAoB,KAAK,MAAM,KAAK,gBAClC,MAAM,MACN,gBACA,eACD;;AAMH,MAAI,KAAK,aAAa;GACpB,MAAM,YAAY,KAAK,MAAM,OAAO;GACpC,MAAM,+BAAe,IAAI,YAAY,EAAE;GACvC,MAAM,OAAO,IAAI,SAAS,aAAa;AACvC,QAAK,UAAU,GAAG,WAAW,KAAK;AAClC,QAAK,UAAU,GAAG,GAAG,KAAK;AAC1B,QAAK,IAAI,OAAO,MAAM,YAAY,KAAK,YAAY,eAAe,GAAG,aAAa;;AAGpF,MAAI,KAAK,mBAAmB;GAG1B,MAAM,QAAQ,KAAK;AACnB,QAAK,IAAI,QAAQ,GAAG,QAAQ,KAAK,cAAc,QAAQ,SAAS,OAAO;IACrE,MAAM,MAAM,KAAK,IAAI,OAAO,sBAAsB;IAClD,MAAM,MAAM,KAAK,IAAI,QAAQ,OAAO,KAAK,cAAc,OAAO;AAC9D,SAAK,IAAI,IAAI,OAAO,IAAI,KAAK,KAAK;KAChC,MAAM,IAAI,IAAI,kBAAkB;AAChC,OAAE,YAAY,KAAK,cAAc,GAAG,SAAS;AAC7C,OAAE,aAAa,GAAG,KAAK,cAAc,GAAG,UAAU;AAClD,OAAE,mBAAmB,GAAG,oBAAoB,GAAG;AAC/C,OAAE,KAAK;;AAET,SAAK,IAAI,OAAO,MAAM,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;AAC5C,UAAM,KAAK,IAAI,OAAO,MAAM,qBAAqB;;AAKnD,OAAI,KAAK,aAAa;IACpB,MAAM,MAAM,KAAK,IAAI,OAAO,sBAAsB;IAClD,MAAM,IAAI,IAAI,kBAAkB;AAChC,MAAE,YAAY,KAAK,YAAY,SAAS;AACxC,MAAE,aAAa,GAAG,KAAK,YAAY,UAAU;AAC7C,MAAE,mBAAmB,GAAG,GAAG,EAAE;AAC7B,MAAE,KAAK;AACP,SAAK,IAAI,OAAO,MAAM,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;IAE5C,MAAM,UAAU,KAAK,IAAI,OAAO,sBAAsB;AACtD,YAAQ,mBAAmB,KAAK,oBAAoB,GAAG,KAAK,gBAAgB,GAAG,EAAE;AACjF,SAAK,IAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;;SAE7C;GAEL,MAAM,UAAU,KAAK,IAAI,OAAO,qBAAqB,EAAE,OAAO,UAAU,CAAC;GACzE,MAAM,OAAO,QAAQ,iBAAiB,EAAE,OAAO,WAAW,CAAC;AAC3D,QAAK,IAAI,IAAI,GAAG,IAAI,KAAK,cAAc,QAAQ,KAAK;IAClD,MAAM,QAAQ,KAAK,cAAc;AACjC,SAAK,YAAY,MAAM,SAAS;AAChC,SAAK,aAAa,GAAG,MAAM,UAAU;AACrC,SAAK,mBAAmB,GAAG,oBAAoB,GAAG;;AAEpD,OAAI,KAAK,aAAa;AACpB,SAAK,YAAY,KAAK,YAAY,SAAS;AAC3C,SAAK,aAAa,GAAG,KAAK,YAAY,UAAU;AAChD,SAAK,mBAAmB,GAAG,GAAG,EAAE;;AAElC,QAAK,KAAK;AACV,WAAQ,mBAAmB,KAAK,oBAAoB,GAAG,KAAK,gBAAgB,GAAG,EAAE;AACjF,QAAK,IAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;;AAGlD,QAAM,KAAK,eAAe,SAASN,iBAAe,GAAG,EAAE;EACvD,MAAM,SAAS,KAAK,eAAe,eAAe,GAAG,EAAE;EACvD,MAAM,UAAU,IAAI,YAAY,OAAO,MAAM,EAAE,CAAC,CAAC;AACjD,OAAK,eAAe,OAAO;AAE3B,OAAK,UAAU;AACf,SAAO;;;CAIT,0BAAkC;AAChC,SAAO,KAAK,YAAY,KAAK;;;CAI/B,OAAgB,iBAAiB;;;;;;;;;;;;;;;;CAiBjC,uBAAuB,SAAwB,MAAoB;EACjE,MAAM,iBAAiB,KAAK,cAAc,EAAE;EAC5C,MAAME,iBAAiC,EAAE,QAAQ,KAAK,QAAQ;AAE9D,MAAI,KAAK,gBAAgB,WAAW,EAClC,MAAK,IAAI,IAAI,GAAG,IAAI,SAAS,gBAAgB,IAC3C,MAAK,gBAAgB,KAAK,qBAAqB,KAAK,KAAK,mBAAmB,KAAK,EAAE,CAAC;AAIxF,MAAI,YAAY,MAAM;AACpB,QAAK,IAAI,OAAO,MAAM,YAAY,KAAK,gBAAgB,GAAG,IAAI,YAAY,CAAC,QAAQ,CAAC,CAAC;AAErF,OAAI,KAAK,aAAa;IACpB,MAAM,+BAAe,IAAI,YAAY,EAAE;IACvC,MAAM,OAAO,IAAI,SAAS,aAAa;AACvC,SAAK,UAAU,GAAG,KAAK,MAAM,OAAO,YAAY,KAAK;AACrD,SAAK,UAAU,GAAG,GAAG,KAAK;AAC1B,SAAK,IAAI,OAAO,MAAM,YAAY,KAAK,YAAY,eAAe,GAAG,aAAa;;;EAKtF,MAAMC,gBAAiD,IAAI,MAAM,KAAK,cAAc,OAAO;AAC3F,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,cAAc,QAAQ,KAAK;GAClD,MAAM,QAAQ,KAAK,cAAc;GACjC,MAAM,aAAa,MAAM,KAAK,YAAY,MAAM,MAAM,gBAAgB,eAAe;GACrF,MAAM,aAAa,IAAI,WAAW,WAAW;GAE7C,IAAI,UAAU,CAAC,MAAM,mBAAmB,MAAM,gBAAgB,WAAW,WAAW;AACpF,OAAI,CAAC,SAAS;IACZ,MAAM,SAAS,MAAM;AACrB,SAAK,IAAI,IAAI,GAAG,IAAI,WAAW,QAAQ,IACrC,KAAI,WAAW,OAAO,OAAO,IAAI;AAC/B,eAAU;AACV;;;AAKN,OAAI,SAAS;AACX,SAAK,IAAI,OAAO,MAAM,YAAY,MAAM,eAAe,GAAG,WAAW;AACrE,UAAM,kBAAkB,IAAI,WAAW,WAAW,MAAM,EAAE,CAAC;;AAG7D,iBAAc,KAAK,MAAM,KAAK,gBAAgB,MAAM,MAAM,gBAAgB,eAAe;;EAG3F,MAAM,UAAU,KAAK,IAAI,OAAO,qBAAqB,EAAE,OAAO,eAAe,CAAC;AAC9E,MAAI,YAAY,KAEd,SAAQ,mBAAmB,KAAK,oBAAoB,GAAG,KAAK,gBAAgB,GAAG,EAAE;EAEnF,MAAM,OAAO,QAAQ,iBAAiB,EAAE,OAAO,eAAe,CAAC;AAC/D,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,cAAc,QAAQ,KAAK;GAClD,MAAM,QAAQ,KAAK,cAAc;AACjC,QAAK,YAAY,MAAM,SAAS;AAChC,QAAK,aAAa,GAAG,MAAM,UAAU;AACrC,QAAK,mBAAmB,GAAG,cAAc,GAAG;;AAE9C,MAAI,KAAK,aAAa;AACpB,QAAK,YAAY,KAAK,YAAY,SAAS;AAC3C,QAAK,aAAa,GAAG,KAAK,YAAY,UAAU;AAChD,QAAK,mBAAmB,GAAG,GAAG,EAAE;;AAElC,OAAK,KAAK;AACV,UAAQ,mBAAmB,KAAK,oBAAoB,GAAG,KAAK,gBAAgB,OAAO,GAAG,EAAE;AACxF,OAAK,IAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;AAEhD,OAAK,UAAU;;;CAIjB,MAAM,gBAAgB,MAA+B;EACnD,MAAM,MAAM,KAAK,gBAAgB;AACjC,QAAM,IAAI,SAASH,iBAAe,GAAG,EAAE;EACvC,MAAM,SAAS,IAAI,eAAe,GAAG,EAAE;EACvC,MAAM,UAAU,IAAI,YAAY,OAAO,MAAM,EAAE,CAAC,CAAC;AACjD,MAAI,OAAO;AACX,SAAO;;CAGT,QAAc;AACZ,OAAK,SAAS;AAEd,OAAK,MAAM,OAAO,KAAK,gBAAgB,QAAQ,CAC7C,MAAK,IAAI,OAAO,MAAM,YAAY,KAAK,GAAG,IAAI,WAAW,IAAI,KAAK,CAAC;AAGrE,OAAK,MAAM,SAAS,KAAK,iBAAiB;AACxC,SAAM,kBAAkB;AACxB,SAAM,mBAAmB;;AAE3B,OAAK,MAAM,SAAS,KAAK,eAAe;AACtC,SAAM,kBAAkB;AACxB,SAAM,mBAAmB;;;;;;;;;CAU7B,MAAM,mBAAmB,UAKtB;EACD,MAAM,IAAI,SAAS;EACnB,MAAM,iBAAiB,KAAK,cAAc,EAAE;EAC5C,MAAME,iBAAiC,EAAE,QAAQ,KAAK,QAAQ;EAE9D,MAAM,SADU,MAAM,IAAI,KAAK,gBAAgB,KAAK,iBAC9B;AAGtB,OAAK,IAAI,OAAO,MAAM,YAAY,KAAK,gBAAgB,GAAG,SAAyB;EAGnF,MAAM,aAAa,MAAM,KAAK,YAAY,MAAM,MAAM,gBAAgB,eAAe;AACrF,OAAK,IAAI,OAAO,MAAM,YAAY,MAAM,eAAe,GAAG,WAAW;EAGrE,MAAM,eAAe,MAAM,KAAK,gBAAgB,MAAM,MAAM,gBAAgB,eAAe;EAK3F,MAAM,UAAU,KAAK,IAAI,OAAO,qBAAqB,EAAE,OAAO,eAAe,CAAC;EAC9E,MAAM,OAAO,QAAQ,iBAAiB,EAAE,OAAO,oBAAoB,CAAC;AACpE,OAAK,YAAY,MAAM,SAAS;AAChC,OAAK,aAAa,GAAG,MAAM,UAAU;AACrC,OAAK,mBAAmB,GAAG,aAAa;AACxC,OAAK,KAAK;AACV,OAAK,IAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;EAGhD,MAAM,aAAa,MAAM,KAAK,QAAQ;EACtC,MAAM,SAAS,MAAM,KAAK,gBAAgB,YAAY,GAAG;AAEzD,SAAO;GACL,QAAQ,MAAM;GACd,QAAQ,MAAM,KAAK;GACnB;GACA;GACD;;;;;;;CAQH,MAAM,mBACJ,YACA,GACmE;EACnE,MAAM,iBAAiB,KAAK,cAAc,EAAE;EAC5C,MAAMA,iBAAiC,EAAE,QAAQ,KAAK,QAAQ;EAE9D,MAAM,SADU,MAAM,IAAI,KAAK,gBAAgB,KAAK,iBAC9B;EAGtB,MAAM,aAAa,MAAM,KAAK,YAAY,MAAM,MAAM,gBAAgB,eAAe;AACrF,OAAK,IAAI,OAAO,MAAM,YAAY,MAAM,eAAe,GAAG,WAAW;EAGrE,MAAM,eAAe,MAAM,KAAK,gBAAgB,MAAM,MAAM,gBAAgB,eAAe;EAK3F,MAAM,UAAU,KAAK,IAAI,OAAO,qBAAqB,EAAE,OAAO,eAAe,cAAc,CAAC;EAC5F,MAAM,OAAO,QAAQ,iBAAiB,EAAE,OAAO,cAAc,cAAc,CAAC;AAC5E,OAAK,YAAY,MAAM,SAAS;AAChC,OAAK,aAAa,GAAG,MAAM,UAAU;AACrC,OAAK,mBAAmB,GAAG,aAAa;AACxC,OAAK,KAAK;AACV,OAAK,IAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;EAEhD,MAAM,aAAa,MAAM,KAAK,QAAQ;EACtC,MAAM,SAAS,MAAM,KAAK,gBAAgB,YAAY,GAAG;AACzD,SAAO;GAAE,QAAQ,MAAM;GAAQ,QAAQ,MAAM,KAAK;GAAQ;GAAQ;;;;;;CAOpE,mBACE,GACA,QAAQ,GAOP;EACD,MAAM,iBAAiB,KAAK,cAAc,EAAE;EAC5C,MAAMA,iBAAiC,EAAE,QAAQ,KAAK,QAAQ;EAC9D,MAAM,UAAU,MAAM,IAAI,KAAK,gBAAgB,KAAK;EACpD,MAAM,UAAU,EAAE;AAClB,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,IAAI,OAAO,QAAQ,OAAO,EAAE,KAAK;GACxD,MAAM,QAAQ,QAAQ;GACtB,MAAM,aAAa,MAAM,KAAK,YAAY,MAAM,MAAM,gBAAgB,eAAe;GACrF,MAAM,MAAM,IAAI,YAAY,WAAW;GACvC,MAAM,eAAe,MAAM,KAAK,gBAC9B,MAAM,MACN,gBACA,eACD;AACD,WAAQ,KAAK;IACX,KAAK;IACL,QAAQ,MAAM;IACd,QAAQ,MAAM,KAAK;IACnB,WAAW,MAAM,KAAK,IAAI;IAC1B;IACD,CAAC;;AAEJ,SAAO;;;;;;CAOT,MAAM,mBAAmB,GAUvB;EACA,MAAM,UAAU,MAAM,IAAI,KAAK,gBAAgB,KAAK;EAEpD,MAAM,eAAe,QAAQ;EAC7B,MAAM,UAAU;GACd;GACA;GACA;GACA;GACA,KAAK,MAAM,eAAe,EAAE;GAC5B,KAAK,MAAM,eAAe,EAAE;GAC5B,KAAK,MAAO,eAAe,IAAK,EAAE;GAClC,eAAe;GAChB;EACD,MAAM,SAAS,CAAC,GAAG,IAAI,IAAI,QAAQ,CAAC,CAAC,QAAQ,MAAM,KAAK,KAAK,IAAI,aAAa;EAE9E,MAAMK,UAQD,EAAE;AACP,OAAK,MAAM,KAAK,QAAQ;GACtB,MAAM,QAAQ,QAAQ;GACtB,MAAM,aAAa,MAAM,KAAK,QAAQ;AACtC,OAAI,CAAC,WAAY;AACjB,OAAI;IACF,MAAM,OAAO,MAAM,KAAK,gBAAgB,YAAY,GAAG;IACvD,IAAI,MAAM;AACV,SAAK,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,IAAK,QAAO,KAAK;IAGlD,IAAIC;AACJ,QAAI;KACF,MAAM,SAAS,MAAM,cAAc;KACnC,MAAM,WAAW,KAAK,IAAI,OAAO,aAAa;MAC5C,MAAM;MACN,OAAO;MACR,CAAC;KACF,MAAM,MAAM,KAAK,IAAI,OAAO,sBAAsB;AAClD,SAAI,mBAAmB,MAAM,eAAe,GAAG,UAAU,GAAG,OAAO;AACnE,UAAK,IAAI,OAAO,MAAM,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;AAC5C,WAAM,SAAS,SAASR,iBAAe,GAAG,OAAO;KACjD,MAAM,MAAM,IAAI,YAAY,SAAS,eAAe,GAAG,OAAO,CAAC,MAAM,EAAE,CAAC;AACxE,qBAAgB,MAAM,KAAK,IAAI;AAC/B,cAAS,OAAO;AAChB,cAAS,SAAS;YACZ;AAIR,YAAQ,KAAK;KACX,KAAK;KACL,QAAQ,MAAM;KACd,QAAQ,MAAM,KAAK;KACnB,QAAQ;KACR,KAAK,KAAK,MAAM,MAAM,IAAI,GAAG;KAC7B,QAAQ;MAAC,KAAK;MAAI,KAAK;MAAI,KAAK;MAAI,KAAK;MAAG,CAAC,KAAK,MAAM,KAAK,MAAM,IAAI,IAAI,GAAG,IAAI;KAClF;KACD,CAAC;WACI;AACN,YAAQ,KAAK;KACX,KAAK;KACL,QAAQ,MAAM;KACd,QAAQ,MAAM,KAAK;KACnB,QAAQ;KACR,KAAK;KACL,QAAQ;MAAC;MAAK;MAAK;MAAK;MAAI;KAC7B,CAAC;;;AAGN,SAAO;;CAGT,iBAAiB,YAAoB,MAA6B;EAChE,MAAM,SAAS,KAAK,UAAU,WAAW;AACzC,MAAI,CAAC,OAAQ,OAAM,IAAI,MAAM,kBAAkB,WAAW,GAAG;AAC7D,OAAK,IAAI,OAAO,MAAM,YAAY,QAAQ,GAAG,KAAqB;;;;;;;CAQpE,WAAW,YAAoB,MAA6B;EAC1D,MAAM,SAAS,KAAK,UAAU,WAAW;AACzC,MAAI,CAAC,OAAQ,OAAM,IAAI,MAAM,wBAAwB,WAAW,GAAG;AACnE,OAAK,IAAI,OAAO,MAAM,YAAY,QAAQ,GAAG,KAAqB;;;CAIpE,aAAa,YAAoB,MAAuB,YAA0B;EAChF,MAAM,SAAS,KAAK,UAAU,WAAW;AACzC,MAAI,CAAC,OAAQ,OAAM,IAAI,MAAM,wBAAwB,WAAW,GAAG;AACnE,OAAK,IAAI,OAAO,MAAM,YAAY,QAAQ,YAAY,KAAqB;;;CAI7E,UAAU,YAA6B;AACrC,SAAO,KAAK,UAAU,WAAW,KAAK;;;CAIxC,IAAI,gBAAwB;AAC1B,SAAO,KAAK;;CAGd,MAAM,gBACJ,YACA,aACA,YACuB;EACvB,MAAM,SAAS,KAAK,UAAU,WAAW;AACzC,MAAI,CAAC,OAAQ,OAAM,IAAI,MAAM,kBAAkB,WAAW,GAAG;EAC7D,MAAM,YAAY,cAAc;EAChC,MAAM,YAAY,OAAO,OAAO;EAChC,MAAM,UAAU,cAAc,KAAK,IAAI,cAAc,GAAG,UAAU,GAAG;EACrE,MAAM,WAAW,KAAK,IAAI,OAAO,aAAa;GAC5C,MAAM;GACN,OAAO;GACR,CAAC;EACF,MAAM,UAAU,KAAK,IAAI,OAAO,sBAAsB;AACtD,UAAQ,mBAAmB,QAAQ,WAAW,UAAU,GAAG,QAAQ;AACnE,OAAK,IAAI,OAAO,MAAM,OAAO,CAAC,QAAQ,QAAQ,CAAC,CAAC;AAChD,QAAM,SAAS,SAASA,iBAAe,GAAG,QAAQ;EAClD,MAAM,OAAO,IAAI,aAAa,SAAS,eAAe,GAAG,QAAQ,CAAC,MAAM,EAAE,CAAC;AAC3E,WAAS,OAAO;AAChB,WAAS,SAAS;AAClB,SAAO;;CAGT,UAAgB;AACd,iBAAe,CAAC,GAAG,KAAK,cAAc,QAAQ,CAAC,CAAC;AAEhD,iBAAe,CAAC,GAAG,IAAI,IAAI,KAAK,kBAAkB,QAAQ,CAAC,CAAC,CAAC;AAC7D,iBAAe,CAAC,GAAG,KAAK,gBAAgB,QAAQ,CAAC,CAAC;AAClD,iBAAe,CAAC,GAAG,KAAK,eAAe,QAAQ,CAAC,CAAC;AACjD,OAAK,MAAM,SAAS,KAAK,gBACvB,OAAM,cAAc,SAAS;AAG/B,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,cAAc,QAAQ,IAC7C,KAAI,KAAK,cAAc,OAAO,KAAK,gBAAgB,GACjD,MAAK,cAAc,GAAG,cAAc,SAAS;AAGjD,MAAI,KAAK,YACP,MAAK,YAAY,cAAc,SAAS;AAE1C,MAAI,KAAK,sBAAsB;AAC7B,QAAK,qBAAqB,SAAS;AACnC,QAAK,uBAAuB;;AAE9B,OAAK,eAAe,SAAS;AAC7B,OAAK,eAAe,SAAS;AAC7B,OAAK,mBAAmB,SAAS;AACjC,OAAK,eAAe,SAAS;AAC7B,iBAAe,KAAK,gBAAgB;AACpC,OAAK,kBAAkB,EAAE;AACzB,OAAK,cAAc,OAAO;AAC1B,OAAK,kBAAkB,OAAO;AAC9B,OAAK,gBAAgB,OAAO;AAC5B,OAAK,eAAe,OAAO;AAC3B,OAAK,kBAAkB,EAAE;AACzB,OAAK,gBAAgB,EAAE;;;;;;;;;;;;;;;;;;CAqBzB,AAAQ,4BAAkC;EACxC,MAAM,gBAAgB,SAAiB,KAAK,MAAM,QAAQ,OAAO,YAAY;AAI7E,MADe,OAAO,YAAY,eAAe,QAAQ,KAAK,uBAAuB,KACzE;AACV,QAAK,MAAM,QAAQ,OAAO,KAAK,KAAK,MAAM,QAAQ,EAAE;AAClD,QAAI,CAAC,aAAa,KAAK,CAAE;IACzB,MAAM,OAAO,KAAK,MAAM,QAAQ;IAIhC,MAAM,QAHQ,KAAK,MAAM,KAAK,MAC5B,MAAM,OAAO,MAAM,UAAU,KAAK,YAAa,EAChD,CACmB,QAAQ,GAAG,MAAM,IAAI,GAAG,EAAE,GAAG,YAAY,KAAK;AAClE,SAAK,kBAAkB,IACrB,MACA,oBAAoB,KAAK,KAAK,cAAc,QAAQ,MAAM,CAC3D;;AAEH,WAAQ,IACN,gDAAgD,KAAK,kBAAkB,KAAK,oBAC7E;AACD;;EAEF,MAAM,gBAAgB,SAAyB;GAC7C,MAAM,OAAO,KAAK,MAAM,QAAQ;AAMhC,UALc,KAAK,MAAM,KAAK,MAAM;AAClC,QAAI,MAAM,IAAK,QAAO,KAAK;AAC3B,QAAI,MAAM,QAAS,QAAO,KAAK;AAC/B,WAAO;KACP,CACW,QAAQ,GAAG,MAAM,IAAI,GAAG,EAAE,GAAG,YAAY,KAAK;;EAG7D,MAAM,WAAW,IAAI,IAAI,KAAK,MAAM,MAAM,KAAK,MAAM,CAAC,EAAE,IAAI,EAAE,CAAC,CAAC;EAChE,MAAM,QAAQ,KAAK,MAAM;EACzB,MAAM,aAAa,IAAI,IAAY,KAAK,MAAM,QAAQ;EACtD,MAAM,2BAAW,IAAI,KAAqB;EAC1C,MAAM,0BAAU,IAAI,KAAqB;AACzC,OAAK,IAAI,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;GACrC,MAAM,OAAO,SAAS,IAAI,MAAM,GAAG;AACnC,QAAK,MAAM,OAAO,KAAK,QAAQ;AAC7B,QAAI,CAAC,aAAa,IAAI,CAAE;AAGxB,QAAI,CAAC,SAAS,IAAI,IAAI,CAAE,YAAW,IAAI,IAAI;AAC3C,YAAQ,IAAI,KAAK,EAAE;;AAErB,QAAK,MAAM,OAAO,KAAK,QACrB,KAAI,aAAa,IAAI,IAAI,CAAC,SAAS,IAAI,IAAI,CAAE,UAAS,IAAI,KAAK,EAAE;;EAIrE,MAAM,2BAAW,IAAI,KAA0B;EAC/C,IAAI,UAAU;EACd,MAAM,2BAAW,IAAI,KAAa;AAClC,OAAK,IAAI,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;GACrC,MAAM,OAAO,SAAS,IAAI,MAAM,GAAG;AAGnC,QAAK,MAAM,OAAO,KAAK,SAAS;AAC9B,QAAI,CAAC,aAAa,IAAI,IAAI,WAAW,IAAI,IAAI,IAAI,KAAK,kBAAkB,IAAI,IAAI,CAC9E;IAEF,MAAM,WAAW,aAAa,IAAI;IAElC,IAAI,SADS,SAAS,IAAI,SAAS,EAChB,KAAK;AACxB,QAAI,CAAC,QAAQ;AACX,cAAS,oBAAoB,KAAK,KAAK,YAAY,QAAQ,GAAG,SAAS,IAAI,SAAS;AACpF;;AAEF,SAAK,kBAAkB,IAAI,KAAK,OAAO;;AAGzC,QAAK,MAAM,QAAQ,CAAC,GAAG,KAAK,QAAQ,GAAG,KAAK,QAAQ,EAAE;AACpD,QAAI,CAAC,aAAa,KAAK,IAAI,WAAW,IAAI,KAAK,IAAI,SAAS,IAAI,KAAK,CAAE;AAEvE,SADa,QAAQ,IAAI,KAAK,IAAI,SAAS,IAAI,KAAK,IAAI,KAC7C,EAAG;IACd,MAAM,SAAS,KAAK,kBAAkB,IAAI,KAAK;AAC/C,QAAI,CAAC,OAAQ;AACb,aAAS,IAAI,KAAK;IAClB,MAAM,WAAW,aAAa,KAAK;IACnC,MAAM,OAAO,SAAS,IAAI,SAAS,IAAI,EAAE;AACzC,SAAK,KAAK,OAAO;AACjB,aAAS,IAAI,UAAU,KAAK;;;AAMhC,OAAK,MAAM,QAAQ,OAAO,KAAK,KAAK,MAAM,QAAQ,EAAE;AAClD,OAAI,CAAC,aAAa,KAAK,IAAI,KAAK,kBAAkB,IAAI,KAAK,CAAE;AAC7D,QAAK,kBAAkB,IACrB,MACA,oBAAoB,KAAK,KAAK,OAAO,QAAQ,aAAa,KAAK,CAAC,CACjE;;EAGH,MAAM,SAAS,IAAI,IAAI,KAAK,kBAAkB,QAAQ,CAAC;EACvD,IAAI,aAAa;AACjB,OAAK,MAAM,KAAK,OAAQ,eAAc,EAAE;AACxC,UAAQ,IACN,2BAA2B,KAAK,kBAAkB,KAAK,aAAa,OAAO,KAAK,aAAa,aAAa,KAAK,QAAQ,EAAE,CAAC,MAC3H;;CAGH,AAAQ,0BAAgC;AACtC,OAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,KAAK,MAAM,QAAQ,EAAE;AAC7D,OAAI,KAAK,YAAY,YAAa;GAElC,MAAM,WADQ,KAAK,MACI,QAAQ,GAAG,MAAM,IAAI,GAAG,EAAE,GAAG,YAAY,KAAK;GACrE,MAAM,SAAS,oBAAoB,KAAK,KAAK,OAAO,QAAQ,SAAS;AACrE,QAAK,gBAAgB,IAAI,MAAM,OAAO;;;CAI1C,AAAQ,yBAA+B;AACrC,OAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,KAAK,MAAM,QAAQ,EAAE;AAC7D,OAAI,KAAK,YAAY,WAAY;GAKjC,MAAM,WAJQ,KAAK,MAAM,KAAK,MAAM;AAClC,QAAI,MAAM,QAAS,QAAO,KAAK;AAC/B,WAAO;KACP,CACqB,QAAQ,GAAG,MAAM,IAAI,GAAG,EAAE,GAAG,YAAY,KAAK;GACrE,MAAM,SAAS,oBAAoB,KAAK,KAAK,MAAM,QAAQ,SAAS;AACpE,QAAK,eAAe,IAAI,MAAM,OAAO;;;CAIzC,AAAQ,cAAc,GAAqC;EACzD,MAAMS,WAAqC,EAAE;AAC7C,OAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,KAAK,MAAM,QAAQ,CAC3D,UAAS,QAAQ,KAAK,MAAM,KAAK,MAAM;AACrC,OAAI,MAAM,IAAK,QAAO;AACtB,OAAI,MAAM,QAAS,QAAO,KAAK,SAAS;AACxC,UAAO;IACP;AAEJ,SAAO;;CAGT,AAAQ,UAAU,YAA2C;AAC3D,SACE,KAAK,cAAc,IAAI,WAAW,IAClC,KAAK,kBAAkB,IAAI,WAAW,IACtC,KAAK,gBAAgB,IAAI,WAAW,IACpC,KAAK,eAAe,IAAI,WAAW;;;;;;CAQvC,AAAQ,0BAAgC;EACtC,IAAI,cAAc;EAClB,MAAM,cAAc,KAAK,IAAI,OAAO;AAEpC,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,cAAc,SAAS,GAAG,KAAK;GACtD,MAAM,KAAK,KAAK,cAAc;GAC9B,MAAM,KAAK,KAAK,cAAc,IAAI;GAClC,MAAM,KAAK,KAAK,cAAc,IAAI;GAGlC,MAAM,UAAU,GAAG,KAAK,WAAW,YAAY,GAAG,KAAK,WAAW;GAClE,MAAM,UAAU,GAAG,KAAK,WAAW,YAAY,GAAG,KAAK,WAAW;AAClE,OAAI,CAAC,WAAW,CAAC,WAAW,GAAG,KAAK,WAAW,SAAU;AAGzD,OAAI,GAAG,KAAK,WAAW,GAAG,KAAK,OAAQ;GAGvC,MAAM,eAAe,GAAG,KAAK,WAAW;AACxC,OAAI,gBAAgB,cAAc,EAAG;AAGrC,OAAI,GAAG,KAAK,OAAO,OAAO,GAAG,KAAK,OAAO,GAAI;GAG7C,MAAM,UAAU,GAAG,KAAK,QAAQ;GAChC,MAAM,QAAQ,GAAG,KAAK,QAAQ;AAC9B,OAAI,GAAG,KAAK,OAAO,OAAO,WAAW,GAAG,KAAK,OAAO,OAAO,MAAO;AAGlE,OAAI,GAAG,KAAK,WAAW,MAAM,GAAG,KAAK,WAAW,EAAG;AACnD,OAAI,GAAG,KAAK,WAAW,MAAM,GAAG,KAAK,WAAW,EAAG;AAMnD,OAAI,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG,KAAK,KAAK,UAAU,GAAG,KAAK,QAAQ,GAAG,CAAE;GAG9E,MAAM,YAAY,eAAe,0BAA0B;GAC3D,MAAM,gBAAgB,oBACpB,KAAK,KACL,aAAa,GAAG,UAChB,UAAU,YACV,UAAU,WACX;GAED,MAAM,mBAAmB,UAAU,YAAY,GAAG,MAAM,EAAE,EAAE,EAAE,QAAQ,GAAG,CAAC;GAC1E,MAAM,eAAe,oBACnB,KAAK,KACL,qBAAqB,GAAG,UACxB,iBACD;GAGD,MAAM,WAAW,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAClD,MAAM,YAAY,KAAK,UAAU,GAAG,KAAK,QAAQ,GAAG;GAEpD,IAAIC;AACJ,OAAI,cAAc;IAEhB,MAAM,QAAQ,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;IAC/C,MAAM,QAAQ,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;IAC/C,MAAM,QAAQ,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;IAC/C,MAAM,MAAM,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;IAC7C,MAAM,MAAM,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;IAC7C,MAAM,MAAM,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;AAC7C,oBAAgB;KACd,EAAE,QAAQ,UAAU;KACpB,EAAE,QAAQ,OAAO;KACjB,EAAE,QAAQ,OAAO;KACjB,EAAE,QAAQ,OAAO;KACjB,EAAE,QAAQ,KAAK;KACf,EAAE,QAAQ,KAAK;KACf,EAAE,QAAQ,KAAK;KACf,EAAE,QAAQ,WAAW;KACrB,EAAE,QAAQ,cAAc;KACzB;UACI;IAEL,MAAM,UAAU,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;IACjD,MAAM,QAAQ,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;AAC/C,oBAAgB;KACd,EAAE,QAAQ,UAAU;KACpB,EAAE,QAAQ,SAAS;KACnB,EAAE,QAAQ,OAAO;KACjB,EAAE,QAAQ,WAAW;KACrB,EAAE,QAAQ,cAAc;KACzB;;GAGH,MAAM,iBAAiB,gBACrB,KAAK,KACL,eACA,eACA,gBAAgB,GAAG,SACpB;GAED,MAAMC,aAA4B;IAChC,QAAQ,gBAAgB,GAAG;IAC3B,MAAM,GAAG;IACT,MAAM;IACN,UAAU;IACV,WAAW;IACX,eAAe;IACf,iBAAiB;IACjB,kBAAkB;IACnB;AAGD,QAAK,cAAc,OAAO,GAAG,GAAG,WAAW;AAC3C;;AAIF,MAAI,cAAc,EAChB,SAAQ,IACN,oBAAoB,YAAY,4BAA4B,cAAc,EAAE,cAC7E;;;;;;;;;;;;;;;;CAkBL,AAAQ,8BAAoC;EAC1C,IAAI,cAAc;AAGlB,MADoB,KAAK,IAAI,OAAO,kCAClB,EAAG;AAErB,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,cAAc,SAAS,GAAG,KAAK;GACtD,MAAM,KAAK,KAAK,cAAc;GAC9B,MAAM,KAAK,KAAK,cAAc,IAAI;AAGlC,OAAI,GAAG,KAAK,WAAW,gBAAgB,GAAG,KAAK,WAAW,aAAc;AAExE,OAAI,GAAG,KAAK,OAAO,OAAO,GAAG,KAAK,OAAO,GAAI;AAE7C,OAAI,GAAG,KAAK,WAAW,MAAM,GAAG,KAAK,WAAW,EAAG;AACnD,OAAI,GAAG,KAAK,WAAW,MAAM,GAAG,KAAK,WAAW,EAAG;AACnD,OAAI,GAAG,KAAK,WAAW,eAAe,GAAG,KAAK,WAAW,WAAY;GAErE,MAAM,OAAO,GAAG,KAAK,QAAQ;GAC7B,MAAM,OAAO,GAAG,KAAK,QAAQ;GAC7B,MAAM,WAAW,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAClD,MAAM,UAAU,KAAK,UAAU,KAAK;GACpC,MAAM,UAAU,KAAK,UAAU,KAAK;AACpC,OAAI,CAAC,YAAY,CAAC,WAAW,CAAC,QAAS;AAMvC,OAAI,aAAa,WAAW,aAAa,WAAW,YAAY,QAAS;GAEzE,MAAM,YAAY;GAClB,MAAM,gBAAgB,oBACpB,KAAK,KACL,WAAW,GAAG,UACd,UAAU,YACV,UAAU,WACX;GAED,MAAM,mBAAmB,UAAU,YAAY,GAAG,MAAM,EAAE,EAAE,EAAE,QAAQ,GAAG,CAAC;GAC1E,MAAM,eAAe,oBACnB,KAAK,KACL,mBAAmB,GAAG,UACtB,iBACD;GAGD,MAAM,MAAM,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAC7C,MAAM,MAAM,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAC7C,MAAM,MAAM,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAC7C,MAAM,MAAM,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAC7C,MAAM,MAAM,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAC7C,MAAM,MAAM,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAE7C,MAAM,gBAAgB;IACpB,EAAE,QAAQ,UAAU;IACpB,EAAE,QAAQ,KAAK;IACf,EAAE,QAAQ,KAAK;IACf,EAAE,QAAQ,KAAK;IACf,EAAE,QAAQ,KAAK;IACf,EAAE,QAAQ,KAAK;IACf,EAAE,QAAQ,KAAK;IACf,EAAE,QAAQ,SAAS;IACnB,EAAE,QAAQ,SAAS;IACnB,EAAE,QAAQ,cAAc;IACzB;GAED,MAAM,iBAAiB,gBACrB,KAAK,KACL,eACA,eACA,cAAc,GAAG,SAClB;GAED,MAAMA,aAA4B;IAChC,QAAQ,cAAc,GAAG;IACzB,MAAM,GAAG;IACT,MAAM;IACN,UAAU;IACV,WAAW;IACX,eAAe;IACf,iBAAiB;IACjB,kBAAkB;IACnB;AAED,QAAK,cAAc,OAAO,GAAG,GAAG,WAAW;AAC3C;;AAIF,MAAI,cAAc,EAChB,SAAQ,IACN,oBAAoB,YAAY,gCAAgC,YAAY,cAC7E;;;;;;;;;;;;;CAeL,AAAQ,+BAAqC;EAC3C,IAAI,cAAc;AAElB,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,cAAc,SAAS,GAAG,KAAK;GACtD,MAAM,KAAK,KAAK,cAAc;GAC9B,MAAM,KAAK,KAAK,cAAc,IAAI;AAElC,OAAI,GAAG,KAAK,WAAW,mBAAmB,GAAG,KAAK,WAAW,gBAAiB;AAE9E,OAAI,GAAG,KAAK,WAAW,UAAU,GAAG,KAAK,WAAW,MAAO;GAI3D,MAAM,OAAO,GAAG,KAAK,QAAQ;GAC7B,MAAM,OAAO,GAAG,KAAK,QAAQ;GAC7B,MAAM,MAAM,KAAK,MAAM,QAAQ,OAAO;AAEtC,OAAI,QADQ,KAAK,MAAM,QAAQ,OAAO,MACrB;GACjB,IAAIC;AACJ,OAAI,QAAQ,MACV,aAAY;YACH,QAAQ,MACjB,aACE,KAAK,WAAW,eACZ,uCACA;OAEN;GAGF,MAAM,OAAO,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAC9C,MAAM,OAAO,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAC9C,MAAM,UAAU,KAAK,UAAU,KAAK;GACpC,MAAM,UAAU,KAAK,UAAU,KAAK;AACpC,OAAI,CAAC,QAAQ,CAAC,QAAQ,CAAC,WAAW,CAAC,QAAS;AAE5C,OAAI,YAAY,QAAS;GAEzB,MAAM,gBAAgB,oBACpB,KAAK,KACL,WAAW,GAAG,UACd,UAAU,YACV,UAAU,WACX;GAED,MAAM,cAAc,KAAK,cAAc,EAAE;GACzC,MAAM,mBAAmB,UAAU,YAAY,GAAG,MAAM,aAAa,EAAE,QAAQ,GAAG,CAAC;GACnF,MAAM,eAAe,oBACnB,KAAK,KACL,mBAAmB,GAAG,UACtB,iBACD;GAED,MAAM,gBAAgB;IACpB,EAAE,QAAQ,MAAM;IAChB,EAAE,QAAQ,MAAM;IAChB,EAAE,QAAQ,SAAS;IACnB,EAAE,QAAQ,SAAS;IACnB,EAAE,QAAQ,cAAc;IACzB;GAED,MAAM,iBAAiB,gBACrB,KAAK,KACL,eACA,eACA,cAAc,GAAG,SAClB;GAED,MAAMD,aAA4B;IAChC,QAAQ,YAAY,GAAG;IACvB,MAAM,GAAG;IACT,MAAM;IACN,UAAU;IACV,WAAW;IACX,eAAe;IACf,iBAAiB;IACjB,kBAAkB;IACnB;AAED,QAAK,cAAc,OAAO,GAAG,GAAG,WAAW;AAC3C;;AAGF,MAAI,cAAc,EAChB,SAAQ,IACN,oBAAoB,YAAY,gCAAgC,YAAY,cAC7E;;;;;;;;;;;;;;;;;CAmBL,AAAQ,8BAAoC;EAC1C,IAAI,cAAc;AAElB,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,cAAc,SAAS,GAAG,KAAK;GACtD,MAAM,KAAK,KAAK,cAAc;GAC9B,MAAM,KAAK,KAAK,cAAc,IAAI;AAElC,OAAI,GAAG,KAAK,WAAW,cAAe;AACtC,OAAI,GAAG,KAAK,WAAW,aAAc;GAGrC,MAAM,WAAW,GAAG,KAAK,QAAQ;AACjC,OAAI,GAAG,KAAK,OAAO,OAAO,SAAU;GAGpC,MAAM,WAAW,GAAG,KAAK,OAAO;GAChC,MAAM,WAAW,GAAG,KAAK,OAAO;GAChC,MAAM,UAAU,KAAK,UAAU,SAAS;GACxC,MAAM,UAAU,KAAK,UAAU,SAAS;GACxC,MAAM,KAAK,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAC5C,MAAM,KAAK,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAC5C,MAAM,KAAK,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAC5C,MAAM,SAAS,KAAK,UAAU,GAAG,KAAK,QAAQ,GAAG;AACjD,OAAI,CAAC,WAAW,CAAC,WAAW,CAAC,MAAM,CAAC,MAAM,CAAC,MAAM,CAAC,OAAQ;AAI1D,OAAI,WAAW,WAAW,WAAW,WAAW,YAAY,QAAS;GAErE,MAAM,YAAY;GAClB,MAAM,gBAAgB,oBACpB,KAAK,KACL,WAAW,GAAG,UACd,UAAU,YACV,UAAU,WACX;GAED,MAAM,mBAAmB,UAAU,YAAY,GAAG,MAAM,EAAE,EAAE,EAAE,QAAQ,GAAG,CAAC;GAC1E,MAAM,eAAe,oBACnB,KAAK,KACL,mBAAmB,GAAG,UACtB,iBACD;GAED,MAAM,gBAAgB;IACpB,EAAE,QAAQ,SAAS;IACnB,EAAE,QAAQ,SAAS;IACnB,EAAE,QAAQ,IAAI;IACd,EAAE,QAAQ,IAAI;IACd,EAAE,QAAQ,IAAI;IACd,EAAE,QAAQ,QAAQ;IAClB,EAAE,QAAQ,cAAc;IACzB;GAED,MAAM,iBAAiB,gBACrB,KAAK,KACL,eACA,eACA,cAAc,GAAG,SAClB;GAED,MAAMA,aAA4B;IAChC,QAAQ,iBAAiB,GAAG;IAC5B,MAAM,GAAG;IACT,MAAM;IACN,UAAU;IACV,WAAW;IACX,eAAe;IACf,iBAAiB;IACjB,kBAAkB;IACnB;AAED,QAAK,cAAc,OAAO,GAAG,GAAG,WAAW;AAC3C;;AAGF,MAAI,cAAc,EAChB,SAAQ,IACN,oBAAoB,YAAY,8BAA8B,YAAY,cAC3E;;;;;;;;;;;;;;CAgBL,AAAQ,mCAAyC;EAC/C,IAAI,cAAc;AAElB,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,cAAc,SAAS,GAAG,KAAK;GACtD,MAAM,KAAK,KAAK,cAAc;GAC9B,MAAM,KAAK,KAAK,cAAc,IAAI;AAElC,OAAI,GAAG,KAAK,WAAW,SAAU;AACjC,OAAI,GAAG,KAAK,WAAW,aAAc;GAErC,MAAM,YAAY,GAAG,KAAK,QAAQ;AAClC,OAAI,GAAG,KAAK,OAAO,OAAO,UAAW;GAGrC,MAAM,WAAW,GAAG,KAAK,OAAO;GAChC,MAAM,SAAS,GAAG,KAAK,OAAO;GAC9B,MAAM,UAAU,KAAK,UAAU,SAAS;GACxC,MAAM,QAAQ,KAAK,UAAU,OAAO;GACpC,MAAM,KAAK,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAC5C,MAAM,KAAK,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAC5C,MAAM,KAAK,KAAK,UAAU,GAAG,KAAK,OAAO,GAAG;GAC5C,MAAM,SAAS,KAAK,UAAU,GAAG,KAAK,QAAQ,GAAG;AACjD,OAAI,CAAC,WAAW,CAAC,SAAS,CAAC,MAAM,CAAC,MAAM,CAAC,MAAM,CAAC,OAAQ;AAIxD,OAAI,WAAW,WAAW,WAAW,SAAS,YAAY,MAAO;GAEjE,MAAM,YAAY;GAClB,MAAM,gBAAgB,oBACpB,KAAK,KACL,gBAAgB,GAAG,UACnB,UAAU,YACV,UAAU,WACX;GAED,MAAM,mBAAmB,UAAU,YAAY,GAAG,MAAM,EAAE,EAAE,EAAE,QAAQ,GAAG,CAAC;GAC1E,MAAM,eAAe,oBACnB,KAAK,KACL,wBAAwB,GAAG,UAC3B,iBACD;GAED,MAAM,gBAAgB;IACpB,EAAE,QAAQ,SAAS;IACnB,EAAE,QAAQ,OAAO;IACjB,EAAE,QAAQ,IAAI;IACd,EAAE,QAAQ,IAAI;IACd,EAAE,QAAQ,IAAI;IACd,EAAE,QAAQ,QAAQ;IAClB,EAAE,QAAQ,cAAc;IACzB;GAED,MAAM,iBAAiB,gBACrB,KAAK,KACL,eACA,eACA,mBAAmB,GAAG,SACvB;GAED,MAAMA,aAA4B;IAChC,QAAQ,sBAAsB,GAAG;IACjC,MAAM,GAAG;IACT,MAAM;IACN,UAAU;IACV,WAAW;IACX,eAAe;IACf,iBAAiB;IACjB,kBAAkB;IACnB;AAED,QAAK,cAAc,OAAO,GAAG,GAAG,WAAW;AAC3C;;AAGF,MAAI,cAAc,EAChB,SAAQ,IACN,oBAAoB,YAAY,mCAAmC,YAAY,cAChF;;;;;;;;;;;;CAcL,AAAQ,+BAAqC;EAC3C,IAAI,cAAc;AAElB,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,cAAc,SAAS,GAAG,KAAK;GACtD,MAAM,KAAK,KAAK,cAAc;GAC9B,MAAM,KAAK,KAAK,cAAc,IAAI;AAElC,OAAI,GAAG,KAAK,WAAW,aAAa,GAAG,KAAK,WAAW,UAAW;AAElE,OAAI,GAAG,KAAK,WAAW,gBAAgB,GAAG,KAAK,WAAW,YAAa;AACvE,OAAI,GAAG,KAAK,WAAW,QAAQ,GAAG,KAAK,WAAW,IAAK;GAEvD,MAAM,MAAM,GAAG,KAAK,OAAO;GAC3B,MAAM,KAAK,GAAG,KAAK,OAAO;GAC1B,MAAM,OAAO,GAAG,KAAK,QAAQ;GAC7B,MAAM,MAAM,GAAG,KAAK,OAAO;GAC3B,MAAM,KAAK,GAAG,KAAK,OAAO;GAC1B,MAAM,OAAO,GAAG,KAAK,QAAQ;GAE7B,MAAM,SAAS,KAAK,UAAU,IAAI;GAClC,MAAM,QAAQ,KAAK,UAAU,GAAG;GAChC,MAAM,UAAU,KAAK,UAAU,KAAK;GACpC,MAAM,SAAS,KAAK,UAAU,IAAI;GAClC,MAAM,QAAQ,KAAK,UAAU,GAAG;GAChC,MAAM,UAAU,KAAK,UAAU,KAAK;AACpC,OAAI,CAAC,UAAU,CAAC,SAAS,CAAC,WAAW,CAAC,UAAU,CAAC,SAAS,CAAC,QAAS;AAQpE,OAAI,YAAY,QAAS;AACzB,OAAI,WAAW,WAAW,WAAW,QAAS;GAI9C,MAAME,YAAoB;IACxB,IAAI,kBAAkB,GAAG;IACzB,QAAQ;IACR,QAAQ;KAAC;KAAK;KAAI;KAAK;KAAG;IAC1B,SAAS,CAAC,MAAM,KAAK;IACrB,YAAY;KACV,aAAa,GAAG,KAAK,WAAW;KAChC,KAAK,GAAG,KAAK,WAAW;KACxB,iBAAiB,GAAG,KAAK,WAAW;KACpC,iBAAiB,GAAG,KAAK,WAAW;KACpC,UAAU,GAAG,KAAK,WAAW;KAC7B,UAAU,GAAG,KAAK,WAAW;KAC9B;IACF;GAED,MAAM,YAAY;GAClB,MAAM,gBAAgB,oBACpB,KAAK,KACL,aAAa,GAAG,UAChB,UAAU,YACV,UAAU,WACX;GAED,MAAM,cAAc,KAAK,cAAc,EAAE;GACzC,MAAM,mBAAmB,UAAU,YAAY,WAAW,aAAa,EAAE,QAAQ,GAAG,CAAC;GACrF,MAAM,eAAe,oBACnB,KAAK,KACL,qBAAqB,GAAG,UACxB,iBACD;GAED,MAAM,gBAAgB;IACpB,EAAE,QAAQ,QAAQ;IAClB,EAAE,QAAQ,OAAO;IACjB,EAAE,QAAQ,QAAQ;IAClB,EAAE,QAAQ,OAAO;IACjB,EAAE,QAAQ,SAAS;IACnB,EAAE,QAAQ,SAAS;IACnB,EAAE,QAAQ,cAAc;IACzB;GAED,MAAM,iBAAiB,gBACrB,KAAK,KACL,eACA,eACA,gBAAgB,GAAG,SACpB;GAED,MAAMF,aAA4B;IAChC,QAAQ,UAAU;IAClB,MAAM;IACN,MAAM;IACN,UAAU;IACV,WAAW;IACX,eAAe;IACf,iBAAiB;IACjB,kBAAkB;IACnB;AAED,QAAK,cAAc,OAAO,GAAG,GAAG,WAAW;AAC3C;;AAGF,MAAI,cAAc,EAChB,SAAQ,IACN,oBAAoB,YAAY,wBAAwB,YAAY,cACrE;;;;;;CAQL,AAAQ,cACN,MACA,MACA,eAC8B;EAC9B,MAAMG,UAAwC,EAAE;EAChD,MAAM,iBAAiB,CAAC,GAAG,KAAK,QAAQ,GAAG,KAAK,QAAQ;EAIxD,MAAM,gCAAgB,IAAI,KAAgB;AAE1C,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,SAAS,QAAQ,KAAK;GAC7C,MAAM,UAAU,KAAK,SAAS;AAE9B,OAAI,QAAQ,SAAS,UACnB,SAAQ,KAAK,EAAE,QAAQ,eAAe,CAAC;QAClC;IAEL,MAAM,aAAa,eADD,IAAI,KAAK,SAAS,QAAQ,GAAG,MAAM,IAAI,KAAK,EAAE,SAAS,UAAU,CAAC;AAEpF,QAAI,CAAC,WAAY,OAAM,IAAI,MAAM,8BAA8B,EAAE,SAAS,KAAK,KAAK;IAEpF,IAAIC;AACJ,QAAI,eAAe,YACjB,UAAS,KAAK;QAEd,UAAS,KAAK,UAAU,WAAW;AAErC,QAAI,CAAC,OAAQ,OAAM,IAAI,MAAM,6BAA6B,WAAW,UAAU,KAAK,KAAK;AAOzF,QAAI,QAAQ,SAAS,qBACnB,KAAI,cAAc,IAAI,OAAO,CAC3B,UAAS,KAAK,wBAAwB,OAAO,KAAK;QAElD,eAAc,IAAI,OAAO;AAI7B,YAAQ,KAAK,EAAE,QAAQ,CAAC;;;AAI5B,SAAO;;;CAIT,AAAQ,wBAAwB,UAA6B;AAC3D,MAAI,CAAC,KAAK,wBAAwB,KAAK,qBAAqB,OAAO,SACjE,MAAK,uBAAuB,oBAC1B,KAAK,KACL,mBACA,KAAK,IAAI,UAAU,IAAI,CACxB;AAEH,SAAO,KAAK;;;;;;;;;;;;ACtzEhB,SAAgB,WAAW,MAAsC;CAC/D,MAAM,EAAE,SAAS,QAAQ,YAAY,QAAQ,GAAG,GAAG,cAAc;CACjE,MAAM,gBAAgB,IAAI;CAC1B,MAAM,eAAe,KAAK,KAAK,IAAI,UAAU;CAC7C,MAAM,cAAc,IAAI;CAExB,MAAM,SAAS,IAAI,YAAY,KAAK,KAAK,gBAAgB,EAAE,CAAC;CAC5D,MAAM,SAAS,IAAI,aAAa,YAAY;CAC5C,MAAM,QAAQ,IAAI,aAAa,YAAY;AAK3C,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,MAAM,UAAU,MAAM;EACtB,MAAM,gBAAgB,IAAI;EAC1B,MAAM,cAAc,UAAU;AAC9B,OAAK,IAAI,MAAM,GAAG,MAAM,GAAG,OAAO;GAGhC,MAAM,SADW,QAAQ,cAAc,SACT,gBAAgB,IAAM;GAGpD,MAAM,UAAU,MAAM,IAAI;GAC1B,MAAM,YAAY,YAAY;GAC9B,MAAM,YAAY,UAAU;AAC5B,UAAO,cAAc,UAAW,YAAY;;;AAOhD,MAAK,IAAI,IAAI,GAAG,IAAI,cAAc,KAAK;EACrC,MAAM,SAAS,IAAI;AACnB,OAAK,IAAI,MAAM,GAAG,MAAM,GAAG,MACzB,QAAO,MAAM,eAAe,KAAK,WAAW,SAAS;;CAYzD,MAAM,aAAa,MAAM;AACzB,MAAK,IAAI,IAAI,GAAG,IAAI,cAAc,KAAK;EACrC,MAAM,SAAS,IAAI;AACnB,OAAK,IAAI,MAAM,GAAG,MAAM,GAAG,OAAO;GAGhC,MAAM,UAFQ,OAAO,UAAU,QAAQ,SACpB,MAAM,KACgB,IAAM;AAC/C,SAAM,MAAM,eAAe,KAAK,UAAU;;;AAI9C,QAAO;EAAE;EAAQ;EAAQ;EAAO;;;;;;;;;;;;;;;;;;;;;;;AChFlC,MAAM,UAAU;AAChB,MAAM,QAAQ;AAEd,SAAS,eAAwB;AAC/B,QAAO,OAAO,cAAc;;AAG9B,IAAIC,aAAiD;AAErD,SAAS,SAAsC;AAC7C,KAAI,WAAY,QAAO;AACvB,cAAa,IAAI,SAA6B,YAAY;AACxD,MAAI,CAAC,cAAc,EAAE;AACnB,WAAQ,KAAK;AACb;;AAEF,MAAI;GACF,MAAM,MAAM,UAAU,KAAK,SAAS,EAAE;AACtC,OAAI,wBAAwB;IAC1B,MAAM,KAAK,IAAI;AACf,QAAI,CAAC,GAAG,iBAAiB,SAAS,MAAM,CAAE,IAAG,kBAAkB,MAAM;;AAEvE,OAAI,kBAAkB,QAAQ,IAAI,OAAO;AACzC,OAAI,gBAAgB,QAAQ,KAAK;AACjC,OAAI,kBAAkB,QAAQ,KAAK;UAC7B;AACN,WAAQ,KAAK;;GAEf;AACF,QAAO;;;AAIT,SAAS,GACP,MACA,KACA,UACY;AACZ,QAAO,QAAQ,CAAC,MACb,OACC,IAAI,SAAY,YAAY;AAC1B,MAAI,CAAC,IAAI;AACP,WAAQ,SAAS;AACjB;;AAEF,MAAI;GACF,MAAM,IAAI,GAAG,YAAY,OAAO,KAAK;GACrC,MAAM,MAAM,IAAI,EAAE,YAAY,MAAM,CAAC;AACrC,OAAI,kBAAkB,QAAS,IAAI,UAAgB,SAAS;AAC5D,OAAI,gBAAgB,QAAQ,SAAS;AACrC,KAAE,gBAAgB,QAAQ,SAAS;UAC7B;AACN,WAAQ,SAAS;;GAEnB,CACL;;;AAIH,eAAsB,OAAO,KAA0C;CACrE,MAAM,IAAI,MAAM,GAAuB,aAAa,MAAM,EAAE,IAAI,IAAI,EAAE,KAAK;AAC3E,QAAO,aAAa,cAAc,IAAI;;;;;;;;AAexC,eAAsB,aAAmC;CACvD,MAAM,OAAO,MAAM,GAAkB,aAAa,MAAM,EAAE,YAAY,EAAE,EAAE,CAAC;AAC3E,QAAO,IAAI,KAAK,QAAQ,EAAE,EAAE,KAAK,MAAM,OAAO,EAAE,CAAC,CAAC;;;;AAKpD,eAAsB,OAAO,KAAa,KAAoC;AAG5E,QADU,MAAM,GAAuB,cAAc,MAAM,EAAE,IAAI,KAAK,IAAI,EAAE,KAAK,KACpE;;;AASf,SAAgB,oBAA6B;AAC3C,QAAO,cAAc;;;;;;;;;;;;ACzEvB,SAAgB,UAAU,KAAoC;CAC5D,MAAM,EAAE,QAAQ,QAAQ,WAAW,QAAQ,WAAW,GAAG,GAAG,cAAc;CAE1E,MAAM,cAAc,IADC,KAAK,KAAK,IAAI,UAAU;CAK7C,MAAM,SAAS;CAMf,MAAM,SAAS,IAAI,aAAa,YAAY;CAC5C,MAAM,QAAQ,IAAI,aAAa,YAAY;AAI3C,MAAK,IAAI,IAAI,GAAG,IAAI,aAAa,KAAK;EACpC,MAAM,IAAI,UAAU;EACpB,MAAM,IAAI,UAAU;AACpB,SAAO,KAAK;AAEZ,QAAM,KAAK,MAAM,IAAI,CAAC,IAAI,IAAI;;AAGhC,QAAO;EAAE;EAAQ;EAAQ;EAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AC9BlC,MAAM,gBAAgB;;AAGtB,MAAM,oBAAoB,IAAI,OAAO;;;;;;AAOrC,SAAgB,gBAAyB;AACvC,QACE,OAAO,cAAc,eACrB,OAAQ,UAAuD,SAAS,iBACtE,cACF,OAAO,WAAW,eAClB,OAAO,SAAS,eAChB,OAAO,QAAQ,eACf,OAAO,IAAI,oBAAoB;;;;;;;;AAUnC,SAAS,cAAc,KAAqB;CAE1C,IAAI,KAAK;CACT,IAAI,KAAK;AACT,MAAK,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;EACnC,MAAM,IAAI,IAAI,WAAW,EAAE;AAC3B,QAAM;AACN,OAAK,KAAK,KAAK,IAAI,SAAc;AACjC,QAAM,IAAI;AACV,OAAK,KAAK,KAAK,IAAI,WAAc;;AAInC,QAAO,IAFI,OAAO,GAAG,SAAS,GAAG,CAAC,SAAS,GAAG,IAAI,IACvC,OAAO,GAAG,SAAS,GAAG,CAAC,SAAS,GAAG,IAAI,CAClC,GAAG,IAAI,OAAO,SAAS,GAAG,CAAC;;AAS7C,MAAM,gBAAgB;mBACH,KAAK,UAAU,cAAc,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;oBAsD7B,kBAAkB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAiEtC,IAAM,aAAN,MAAiB;CACf,AAAQ,SAAwB;CAChC,AAAQ,UAAyB;CACjC,AAAQ,0BAAU,IAAI,KAA8B;CACpD,AAAQ,SAAS;CACjB,AAAQ,SAAS;CAEjB,AAAQ,eAA8B;AACpC,MAAI,KAAK,OAAQ,QAAO;AACxB,MAAI,KAAK,OAAQ,QAAO,KAAK;AAC7B,MAAI;GACF,MAAM,OAAO,IAAI,KAAK,CAAC,cAAc,EAAE,EAAE,MAAM,0BAA0B,CAAC;AAC1E,QAAK,UAAU,IAAI,gBAAgB,KAAK;GACxC,MAAM,SAAS,IAAI,OAAO,KAAK,QAAQ;AACvC,UAAO,aAAa,OAAqB;IACvC,MAAM,EAAE,IAAI,IAAI,SAAS,KAAK,UAAU,GAAG;IAO3C,MAAM,IAAI,KAAK,QAAQ,IAAI,GAAG;AAC9B,QAAI,CAAC,EAAG;AACR,SAAK,QAAQ,OAAO,GAAG;AACvB,QAAI,GAAI,GAAE,QAAQ;KAAE;KAAS;KAAK,CAAC;QAC9B,GAAE,OAAO,IAAI,MAAM,SAAS,oBAAoB,CAAC;;AAExD,UAAO,gBAAgB;AAErB,SAAK,SAAS;AACd,SAAK,MAAM,KAAK,KAAK,QAAQ,QAAQ,CACnC,GAAE,uBAAO,IAAI,MAAM,sBAAsB,CAAC;AAE5C,SAAK,QAAQ,OAAO;;AAEtB,QAAK,SAAS;AACd,UAAO;UACD;AACN,QAAK,SAAS;AACd,UAAO;;;CAIX,AAAQ,KACN,IACA,MACA,KAC0D;EAC1D,MAAM,SAAS,KAAK,cAAc;AAClC,MAAI,CAAC,OAAQ,QAAO,QAAQ,uBAAO,IAAI,MAAM,mBAAmB,CAAC;EACjE,MAAM,KAAK,KAAK;AAChB,SAAO,IAAI,SAAS,SAAS,WAAW;AACtC,QAAK,QAAQ,IAAI,IAAI;IAAE;IAAS;IAAQ,CAAC;AACzC,OAAI;AACF,QAAI,IAEF,QAAO,YAAY;KAAE;KAAI;KAAI;KAAM;KAAK,EAAE,CAAC,IAAI,CAAC;QAEhD,QAAO,YAAY;KAAE;KAAI;KAAI;KAAM,CAAC;YAE/B,GAAG;AACV,SAAK,QAAQ,OAAO,GAAG;AACvB,WAAO,aAAa,QAAQ,IAAI,IAAI,MAAM,OAAO,EAAE,CAAC,CAAC;;IAEvD;;CAGJ,MAAM,IAAI,KAA+B;EACvC,MAAM,EAAE,YAAY,MAAM,KAAK,KAAK,OAAO,cAAc,IAAI,CAAC;AAC9D,SAAO,QAAQ,QAAQ;;CAGzB,MAAM,KAAK,KAA0C;EACnD,MAAM,EAAE,QAAQ,MAAM,KAAK,KAAK,QAAQ,cAAc,IAAI,CAAC;AAC3D,SAAO,OAAO;;CAGhB,MAAM,MAAM,KAAa,KAAiC;AACxD,QAAM,KAAK,KAAK,SAAS,cAAc,IAAI,EAAE,IAAI;;CAGnD,MAAM,OAAO,KAA4B;AACvC,QAAM,KAAK,KAAK,UAAU,cAAc,IAAI,CAAC;;;AAIjD,IAAIC,UAA6B;AACjC,SAAS,SAAqB;AAC5B,KAAI,CAAC,QAAS,WAAU,IAAI,YAAY;AACxC,QAAO;;;;;;AAoBT,eAAsB,SAAS,KAA0C;AACvE,KAAI,CAAC,eAAe,CAAE,QAAO;AAC7B,KAAI;AACF,SAAO,MAAM,QAAQ,CAAC,KAAK,IAAI;SACzB;AACN,SAAO;;;;;;;;;;;AAsCX,eAAsB,iBAAgC;AACpD,KAAI;EACF,MAAM,OAAQ,MACZ,WACC,SAAS,gBAAgB;AAM5B,MAAI,CAAC,KAAM;EACX,MAAMC,QAAkB,EAAE;AAC1B,aAAW,MAAM,CAAC,MAAM,WAAW,KAAK,WAAW,IAAI,EAAE,CACvD,KACE,OAAO,SAAS,eAChB,KAAK,WAAW,iBAAiB,IACjC,SAAS,cAET,OAAM,KAAK,KAAK;AAGpB,OAAK,MAAM,QAAQ,MACjB,OAAM,KAAK,YAAY,MAAM,EAAE,WAAW,MAAM,CAAC,CAAC,YAAY,GAE5D;SAEE;;;;;;;;;;;AC/UV,SAAgB,uBAAuB,QAAsC;AAC3E,KAAI,OAAO,aAAa,EACtB,OAAM,IAAI,MACR,4EAA4E,OAAO,WAAW,IAC/F;CAGH,MAAM,OAAO,IAAI,SAAS,OAAO;CAIjC,MAAM,eAAe,OAAO,KAAK,aAAa,GAAG,KAAK,CAAC;AAEvD,KAAI,eAAe,OAAO,aAAa,EACrC,OAAM,IAAI,MACR,8BAA8B,aAAa,yBAAyB,OAAO,aAAa,EAAE,6BAC7D,eAAe,EAAE,SAC/C;CAIH,MAAM,cAAc,IAAI,WAAW,QAAQ,GAAG,aAAa;CAC3D,MAAM,YAAY,IAAI,aAAa,CAAC,OAAO,YAAY;CACvD,MAAMC,SAAkC,KAAK,MAAM,UAAU;CAE7D,MAAM,YAAY,IAAI;CACtB,MAAMC,UAA6B,EAAE;CACrC,IAAIC,WAA0C;AAE9C,MAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,OAAO,EAAE;AACjD,MAAI,SAAS,gBAAgB;AAC3B,cAAW;AACX;;EAGF,MAAM,EAAE,OAAO,OAAO,iBAAiB;AAMvC,UAAQ,KAAK;GACX;GACA;GACA;GACA,YAAY,aAAa;GACzB,YAAY,aAAa,KAAK,aAAa;GAC5C,CAAC;;AAIJ,SAAQ,MAAM,GAAG,MAAM,EAAE,aAAa,EAAE,WAAW;AAEnD,QAAO;EAAE;EAAc;EAAW;EAAS;EAAU;;;AAIvD,SAAS,eAAe,OAAgC;AACtD,SAAQ,OAAR;EACE,KAAK;EACL,KAAK;EACL,KAAK,MACH,QAAO;EACT,KAAK;EACL,KAAK;EACL,KAAK,MACH,QAAO;EACT,KAAK;EACL,KAAK;EACL,KAAK;EACL,KAAK,MACH,QAAO;EACT,KAAK;EACL,KAAK;EACL,KAAK,OACH,QAAO;EACT,QACE,QAAO;;;;;;;;;;AAWb,SAAS,cACP,QACA,QACA,YACA,OACiB;CAEjB,MAAM,UAAU,SADF,eAAe,MAAM,KACA;CAGnC,MAAM,MAAM,UAAU,SAAS,OAAO,MAAM,QAAQ,SAAS,WAAW;CACxE,MAAM,OAAO,UAAU,SAAS;AAEhC,SAAQ,OAAR;EACE,KAAK,MACH,QAAO,IAAI,aAAa,KAAK,MAAM,aAAa,EAAE;EACpD,KAAK,MAEH,QAAO,IAAI,YAAY,KAAK,MAAM,aAAa,EAAE;EACnD,KAAK,OAEH,QAAO,IAAI,YAAY,KAAK,MAAM,aAAa,EAAE;EACnD,KAAK,MACH,QAAO,IAAI,WAAW,KAAK,MAAM,aAAa,EAAE;EAClD,KAAK,MACH,QAAO,IAAI,YAAY,KAAK,MAAM,aAAa,EAAE;EACnD,KAAK,KACH,QAAO,IAAI,UAAU,KAAK,MAAM,WAAW;EAC7C,KAAK,KACH,QAAO,IAAI,WAAW,KAAK,MAAM,WAAW;EAC9C,KAAK,MACH,QAAO,IAAI,WAAW,KAAK,MAAM,aAAa,EAAE;EAClD,KAAK,MACH,QAAO,IAAI,YAAY,KAAK,MAAM,aAAa,EAAE;EACnD,KAAK,MACH,QAAO,IAAI,aAAa,KAAK,MAAM,aAAa,EAAE;EACpD,KAAK,MACH,QAAO,IAAI,cAAc,KAAK,MAAM,aAAa,EAAE;EACrD,KAAK,MACH,QAAO,IAAI,eAAe,KAAK,MAAM,aAAa,EAAE;EACtD,KAAK,OACH,QAAO,IAAI,WAAW,KAAK,MAAM,WAAW;EAC9C,QACE,OAAM,IAAI,MAAM,kCAAkC,QAAQ;;;;;;;AAQhE,SAAgB,cACd,QACA,MACA,OACiB;AAEjB,QAAO,cAAc,QADN,KAAK,YAAY,MAAM,YACD,MAAM,YAAY,MAAM,MAAM;;;;;;AAuCrE,SAAgB,UAAU,WAAqC;CAC7D,MAAM,OAAO,IAAI,YAAY,UAAU,QAAQ,UAAU,YAAY,UAAU,aAAa,EAAE;CAC9F,MAAM,MAAM,IAAI,aAAa,KAAK,OAAO;CACzC,MAAM,MAAM,IAAI,YAAY,IAAI,OAAO;AACvC,MAAK,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,IAC/B,KAAI,KAAK,KAAK,MAAM;AAEtB,QAAO;;;;;AAMT,SAAgB,SAAS,SAAwC;CAC/D,MAAM,MACJ,mBAAmB,cACf,UACA,IAAI,YAAY,QAAQ,QAAQ,QAAQ,YAAY,QAAQ,aAAa,EAAE;CACjF,MAAM,MAAM,IAAI,aAAa,IAAI,OAAO;AACxC,MAAK,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;EACnC,MAAM,IAAI,IAAI;EACd,MAAM,OAAQ,KAAK,KAAM;EACzB,MAAM,MAAO,KAAK,KAAM;EACxB,MAAM,OAAO,IAAI;AACjB,MAAI,QAAQ,EAEV,KAAI,MAAM,OAAO,KAAK,KAAK,KAAK,OAAO,OAAO;WACrC,QAAQ,GAEjB,KAAI,KAAK,SAAS,IAAK,OAAO,YAAY,WAAY;MAEtD,KAAI,MAAM,OAAO,KAAK,KAAK,MAAM,MAAM,OAAO,IAAI,OAAO;;AAG7D,QAAO;;;;;;;;;;;ACnQT,MAAM,kCAAkB,IAAI,KAAqB;AACjD,MAAM,kCAAkB,IAAI,KAAqB;AAEjD;CAEE,MAAMC,KAAe,EAAE;AACvB,MAAK,IAAI,IAAI,IAAI,KAAK,KAAK,IAAK,IAAG,KAAK,EAAE;AAC1C,MAAK,IAAI,IAAI,KAAK,KAAK,KAAK,IAAK,IAAG,KAAK,EAAE;AAC3C,MAAK,IAAI,IAAI,KAAK,KAAK,KAAK,IAAK,IAAG,KAAK,EAAE;CAE3C,MAAM,KAAK,CAAC,GAAG,GAAG;CAGlB,IAAI,IAAI;AACR,MAAK,IAAI,IAAI,GAAG,IAAI,KAAK,IACvB,KAAI,CAAC,GAAG,SAAS,EAAE,EAAE;AACnB,KAAG,KAAK,EAAE;AACV,KAAG,KAAK,MAAM,EAAE;AAChB;;AAIJ,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,QAAQ,KAAK;AAClC,kBAAgB,IAAI,GAAG,IAAI,GAAG,GAAG;AACjC,kBAAgB,IAAI,GAAG,IAAI,GAAG,GAAG;;;AAmBrC,IAAa,YAAb,MAAa,UAAU;CACrB,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;;;;;;CAMR,AAAQ;CACR,AAAS;CACT,AAAS;CAET,AAAQ,YACN,OACA,cACA,QACA,eACA,aACA,QACA,WACA,SACA;AACA,OAAK,QAAQ;AACb,OAAK,eAAe;AACpB,OAAK,SAAS;AACd,OAAK,gBAAgB;AACrB,OAAK,cAAc;AACnB,OAAK,+BAAe,IAAI,KAAK;AAC7B,OAAK,UAAU;AACf,OAAK,SAAS;AACd,OAAK,YAAY;AAIjB,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,KAAK;GAE5B,MAAM,MAAM,MADA,EAAE,SAAS,GAAG,CAAC,aAAa,CAAC,SAAS,GAAG,IAAI,CACnC;AACtB,OAAI,MAAM,IAAI,IAAI,CAChB,MAAK,aAAa,IAAI,GAAG,IAAI;;;;;;CAQnC,OAAO,SAAS,eAAoB,qBAAsC;EACxE,MAAM,QAAQ,cAAc;AAC5B,MAAI,CAAC,SAAS,MAAM,SAAS,MAC3B,OAAM,IAAI,MACR,+BAA+B,OAAO,QAAQ,UAAU,0BACzD;EAIH,MAAM,wBAAQ,IAAI,KAAqB;EACvC,MAAM,+BAAe,IAAI,KAAqB;AAC9C,OAAK,MAAM,CAAC,OAAO,OAAO,OAAO,QAAQ,MAAM,MAAgC,EAAE;AAC/E,SAAM,IAAI,OAAO,GAAG;AACpB,gBAAa,IAAI,IAAI,MAAM;;EAO7B,MAAM,iCAAiC;GACrC,MAAM,OAAO,cAAc;GAC3B,MAAM,YAAY,MAChB,GAAG,SAAS,aAAa,GAAG,SAAS,WAAW,OAAO,GAAG,YAAY;AACxE,OAAI,CAAC,KAAM,QAAO;AAClB,OAAI,SAAS,KAAK,CAAE,QAAO;AAC3B,OAAI,MAAM,QAAQ,KAAK,YAAY,CAAE,QAAO,KAAK,YAAY,KAAK,SAAS;AAC3E,UAAO;MACL;EACJ,MAAM,mBAAmB,MAAM,kBAAkB;EACjD,MAAM,UAAU,2BAA4B,oBAAoB,UAAU,MAAM;EAIhF,MAAM,yBAAS,IAAI,KAAqB;EACxC,MAAM,YAAa,MAAM,UAAU,EAAE;AACrC,OAAK,IAAI,IAAI,GAAG,IAAI,UAAU,QAAQ,KAAK;GACzC,MAAM,IAAI,UAAU;GACpB,MAAM,MAAM,MAAM,QAAQ,EAAE,GAAG,GAAG,EAAE,GAAG,GAAG,EAAE,OAAO;AACnD,UAAO,IAAI,KAAK,EAAE;;EAKpB,MAAM,gCAAgB,IAAI,KAAqB;EAC/C,MAAM,iCAAiB,IAAI,KAAqB;EAChD,MAAM,kBAAkB,cAAc;AAGtC,MAAI,gBACF,MAAK,MAAM,KAAK,iBAAiB;AAC/B,SAAM,IAAI,EAAE,SAAS,EAAE,GAAG;AAC1B,gBAAa,IAAI,EAAE,IAAI,EAAE,QAAQ;AACjC,kBAAe,IAAI,EAAE,SAAS,EAAE,GAAG;AACnC,OAAI,EAAE,QACJ,eAAc,IAAI,EAAE,SAAS,EAAE,GAAG;;EAMxC,MAAM,WAAW,qBAAqB,aAAa;EACnD,MAAM,WAAW,qBAAqB,aAAa;EACnD,MAAM,eAAe,qBAAqB,iBAAiB;EAC3D,MAAM,cAAc,qBAAqB,iBAAiB;EAC1D,MAAM,cAAc,qBAAqB,iBAAiB;AAY1D,SAAO,IAAI,UACT,OACA,cACA,QACA,eACA,gBAf8B;GAC9B;GACA;GACA,YAAY,WAAY,MAAM,IAAI,SAAS,IAAI,OAAQ;GACvD,YAAY,WAAY,MAAM,IAAI,SAAS,IAAI,OAAQ;GACvD;GACA;GACA;GACD,EASC,MAAM,MACN,QACD;;;;;;CAOH,UAAU,OAA8B;AACtC,SAAO,KAAK,MAAM,IAAI,MAAM,IAAI;;;;;CAMlC,OAAO,MAAwB;AAC7B,MAAI,CAAC,KAAM,QAAO,EAAE;EAEpB,MAAMC,MAAgB,EAAE;AAGxB,MAAI,KAAK,OAAO,eAAe,KAAK,OAAO,eAAe,KACxD,KAAI,KAAK,KAAK,OAAO,WAAW;EAIlC,MAAM,QAAQ,KAAK,qBAAqB,KAAK;AAE7C,OAAK,MAAM,QAAQ,MACjB,KAAI,KAAK,SAAS;GAChB,MAAM,KAAK,KAAK,YAAY,IAAI,KAAK,KAAK;AAC1C,OAAI,OAAO,OACT,KAAI,KAAK,GAAG;SAET;GAGL,MAAM,SAAS,KAAK,YAAY,KAAK,KAAK;AAC1C,QAAK,MAAM,SAAS,QAAQ;IAC1B,MAAM,WAAW,KAAK,UAAU,MAAM;AACtC,QAAI,KAAK,GAAG,SAAS;;;AAM3B,MAAI,KAAK,OAAO,eAAe,KAAK,OAAO,eAAe,KACxD,KAAI,KAAK,KAAK,OAAO,WAAW;AAGlC,SAAO;;;;;CAMT,OAAO,KAAe,oBAA6B,MAAc;EAC/D,MAAMC,SAAmB,EAAE;AAE3B,OAAK,MAAM,MAAM,KAAK;GACpB,MAAM,QAAQ,KAAK,aAAa,IAAI,GAAG;AACvC,OAAI,CAAC,MAAO;AAGZ,OAAI,qBAAqB,KAAK,cAAc,IAAI,MAAM,CACpD;AAGF,UAAO,KAAK,MAAM;;AAGpB,MAAI,KAAK,SAAS;GAIhB,IAAI,MAAM;GACV,IAAIC,UAAoB,EAAE;GAC1B,MAAM,mBAAmB;AACvB,QAAI,QAAQ,WAAW,EAAG;AAC1B,WAAO,IAAI,aAAa,CAAC,OAAO,IAAI,WAAW,QAAQ,CAAC;AACxD,cAAU,EAAE;;AAEd,QAAK,MAAM,SAAS,QAAQ;IAC1B,MAAM,IAAI,yBAAyB,KAAK,MAAM;AAC9C,QAAI,EACF,SAAQ,KAAK,OAAO,SAAS,EAAE,IAAI,GAAG,CAAC;SAClC;AACL,iBAAY;AACZ,YAAO;;;AAGX,eAAY;AACZ,UAAO,IAAI,QAAQ,MAAM,IAAI;;EAS/B,IAAI,OAAO,OAAO,KAAK,GAAG;EAG1B,MAAMC,QAAkB,EAAE;AAC1B,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;GACpC,MAAM,KAAK,KAAK,YAAY,EAAE;GAC9B,MAAM,OAAO,gBAAgB,IAAI,GAAG;AACpC,OAAI,SAAS,OACX,OAAM,KAAK,KAAK;QACX;IAEL,MAAM,UAAU,IAAI,aAAa,CAAC,OAAO,OAAO,cAAc,GAAG,CAAC;AAClE,SAAK,MAAM,KAAK,QAAS,OAAM,KAAK,EAAE;AACtC,QAAI,KAAK,MAAQ;;;AAGrB,SAAO,IAAI,aAAa,CAAC,OAAO,IAAI,WAAW,MAAM,CAAC;AAGtD,SAAO,KAAK,QAAQ,0BAA0B,GAAG,QAAQ;AACvD,UAAO,OAAO,aAAa,SAAS,KAAK,GAAG,CAAC;IAC7C;AAEF,SAAO;;;;;;;;;;;;;;;;;CAkBT,AAAQ,uBAAuB,UAAyB,cAA+B;EACrF,MAAMC,QAAkB,EAAE;AAC1B,MAAI,KAAK,OAAO,SAAU,OAAM,KAAK,KAAK,OAAO,SAAS;EAC1D,IAAI,aAAa;AACjB,OAAK,MAAM,OAAO,UAAU;AAC1B,OAAI,IAAI,SAAS,UAAU;AACzB,kBAAc,GAAG,IAAI,QAAQ;AAC7B;;GAEF,MAAM,OAAO,IAAI,SAAS,cAAc,UAAU,IAAI;GACtD,MAAM,UAAU,SAAS,UAAU,aAAa,aAAa,IAAI,UAAU,IAAI;AAC/E,OAAI,SAAS,OAAQ,cAAa;AAClC,SAAM,KAAK,UAAU,KAAK,IAAI,QAAQ,WAAW;;AAEnD,MAAI,aAAc,OAAM,KAAK,iBAAiB;AAC9C,SAAO,MAAM,KAAK,GAAG;;CAGvB,kBAAkB,UAAyB,SAAqD;EAC9F,MAAM,eAAe,SAAS,uBAAuB;EACrD,MAAMA,QAAkB,EAAE;AAK1B,MAAI,KAAK,YAAY,IAAI,UAAU,IAAI,KAAK,YAAY,IAAI,UAAU,CACpE,QAAO,KAAK,uBAAuB,UAAU,aAAa;EAK5D,MAAM,WAAW,KAAK,OAAO;AAC7B,MAAI,UAAU,SAAS,YAAY,IAAI,KAAK,OAAO,SACjD,OAAM,KAAK,KAAK,OAAO,SAAS;AAGlC,OAAK,MAAM,OAAO,SAChB,OAAM,KAAK,eAAe,IAAI,KAAK,IAAI,IAAI,QAAQ,cAAc;AAGnE,MAAI,cAAc;AAChB,SAAM,KAAK,0BAA0B;AASrC,OAH2B,WACvB,SAAS,SAAS,UAAU,GAC5B,KAAK,YAAY,IAAI,UAAU,CAEjC,OAAM,KAAK,0BAA0B;;AAIzC,SAAO,MAAM,KAAK,GAAG;;;;;CAMvB,WAAW,UAAyB,SAAuD;EACzF,MAAM,OAAO,KAAK,kBAAkB,UAAU,QAAQ;AACtD,SAAO,KAAK,OAAO,KAAK;;CAK1B,AAAQ,qBAAqB,MAAyD;AACpF,MAAI,KAAK,YAAY,SAAS,EAAG,QAAO,CAAC;GAAE;GAAM,SAAS;GAAO,CAAC;EAElE,MAAMC,QAAmD,EAAE;EAM3D,MAAM,UAHiB,CAAC,GAAG,KAAK,YAAY,MAAM,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,SAAS,EAAE,OAAO,CAGxD,KAAK,MAAM,EAAE,QAAQ,uBAAuB,OAAO,CAAC;EACnF,MAAM,QAAQ,IAAI,OAAO,IAAI,QAAQ,KAAK,IAAI,CAAC,IAAI,IAAI;EAEvD,IAAI,YAAY;EAChB,IAAIC;AACJ,UAAQ,QAAQ,MAAM,KAAK,KAAK,MAAM,MAAM;AAC1C,OAAI,MAAM,QAAQ,UAChB,OAAM,KAAK;IACT,MAAM,KAAK,MAAM,WAAW,MAAM,MAAM;IACxC,SAAS;IACV,CAAC;AAEJ,SAAM,KAAK;IAAE,MAAM,MAAM;IAAI,SAAS;IAAM,CAAC;AAC7C,eAAY,MAAM;;AAEpB,MAAI,YAAY,KAAK,OACnB,OAAM,KAAK;GAAE,MAAM,KAAK,MAAM,UAAU;GAAE,SAAS;GAAO,CAAC;AAG7D,SAAO;;CAGT,AAAQ,YAAY,MAAwB;AAC1C,MAAI,CAAC,KAAM,QAAO,EAAE;AAEpB,MAAI,KAAK,SAAS;GAKhB,MAAM,UAAU,KAAK,QAAQ,MAAM,IAAI;GACvC,MAAMC,SAAmB,EAAE;GAC3B,IAAI,IAAI;AACR,UAAO,IAAI,QAAQ,QAAQ;IACzB,IAAI,IAAI,IAAI;AACZ,WAAO,IAAI,QAAQ,UAAU,QAAQ,OAAO,IAAK;AACjD,WAAO,KAAK,QAAQ,MAAM,GAAG,EAAE,CAAC;AAChC,QAAI;;AAEN,UAAO;;EAOT,MAAM,UAAU,KAAK,MADL,qEACmB;AACnC,MAAI,CAAC,QAAS,QAAO,EAAE;AAIvB,SAAO,QAAQ,KAAK,UAAU,KAAK,gBAAgB,MAAM,CAAC;;CAG5D,AAAQ,gBAAgB,MAAsB;EAI5C,IAAI,SAAS;AACb,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;GACpC,MAAM,IAAI,KAAK,WAAW,EAAE;AAC5B,OAAI,MAAM,GACR,WAAU;YACD,IAAI,MAAO,IAAI,OAAO,IAAI,IAGnC,WAAU,OAAO,aAAa,IAAI,IAAI;OAEtC,WAAU,KAAK;;AAGnB,SAAO;;CAGT,AAAQ,UAAU,MAAwB;AACxC,MAAI,KAAK,WAAW,EAAG,QAAO,EAAE;AAGhC,MAAI,KAAK,MAAM,IAAI,KAAK,CACtB,QAAO,CAAC,KAAK,MAAM,IAAI,KAAK,CAAE;EAIhC,IAAI,UAAU,CAAC,GAAG,KAAK;AAIvB,MAAI,QAAQ,WAAW,GAAG;GACxB,MAAM,KAAK,KAAK,MAAM,IAAI,QAAQ,GAAG;AACrC,OAAI,OAAO,OAAW,QAAO,CAAC,GAAG;AAEjC,UAAO,KAAK,mBAAmB,KAAK;;AAItC,SAAO,QAAQ,SAAS,GAAG;GAGzB,IAAI,WAAW;GACf,IAAI,UAAU;AAEd,QAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,SAAS,GAAG,KAAK;IAC3C,MAAM,OAAO,GAAG,QAAQ,GAAG,GAAG,QAAQ,IAAI;IAC1C,MAAM,OAAO,KAAK,OAAO,IAAI,KAAK;AAClC,QAAI,SAAS,UAAa,OAAO,UAAU;AACzC,gBAAW;AAEX,eAAU;;;AAKd,OAAI,YAAY,GAAI;GAGpB,MAAM,SAAS,QAAQ,WAAW,QAAQ,UAAU;AACpD,aAAU;IAAC,GAAG,QAAQ,MAAM,GAAG,QAAQ;IAAE;IAAQ,GAAG,QAAQ,MAAM,UAAU,EAAE;IAAC;;EAIjF,MAAMP,MAAgB,EAAE;AACxB,OAAK,MAAM,OAAO,SAAS;GACzB,MAAM,KAAK,KAAK,MAAM,IAAI,IAAI;AAC9B,OAAI,OAAO,OACT,KAAI,KAAK,GAAG;OAGZ,KAAI,KAAK,GAAG,KAAK,mBAAmB,IAAI,CAAC;;AAG7C,SAAO;;CAGT,AAAQ,mBAAmB,MAAwB;EAEjD,MAAM,QADU,IAAI,aAAa,CACX,OAAO,KAAK;EAClC,MAAMA,MAAgB,EAAE;AACxB,OAAK,MAAM,KAAK,OAAO;GACrB,MAAM,MAAM,KAAK,aAAa,IAAI,EAAE;AACpC,OAAI,KAAK;IACP,MAAM,KAAK,KAAK,MAAM,IAAI,IAAI;AAC9B,QAAI,OAAO,OAAW,KAAI,KAAK,GAAG;;;AAGtC,SAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACveX,IAAa,kBAAb,MAA2D;CACzD,AAAQ;CAER,YAAY,KAAgC;AAC1C,OAAK,MAAM,uBAAO,IAAI,KAAK;;CAG7B,IAAI,MAAuB;AACzB,SAAO,KAAK,IAAI,IAAI,KAAK;;CAG3B,OAAiB;AACf,SAAO,CAAC,GAAG,KAAK,IAAI,MAAM,CAAC;;CAG7B,IAAI,OAAe;AACjB,SAAO,KAAK,IAAI;;CAGlB,IAAI,MAAgD;AAClD,SAAO,QAAQ,QAAQ,KAAK,IAAI,IAAI,KAAK,CAAC;;CAG5C,IAAI,MAAc,OAAmC;AACnD,OAAK,IAAI,IAAI,MAAM,MAAM;AACzB,SAAO,QAAQ,SAAS;;CAG1B,OAAO,MAA6B;AAClC,OAAK,IAAI,OAAO,KAAK;AACrB,SAAO,QAAQ,SAAS;;;CAI1B,QAAkC;AAChC,SAAO,KAAK;;;;;;;;AA4BhB,SAAS,mBAAmB,KAAkB,OAAyC;AACrF,SAAQ,OAAR;EACE,KAAK,MACH,QAAO,IAAI,aAAa,KAAK,GAAG,IAAI,aAAa,EAAE;EACrD,KAAK,MACH,QAAO,IAAI,WAAW,KAAK,GAAG,IAAI,aAAa,EAAE;EACnD,KAAK,MACH,QAAO,IAAI,YAAY,KAAK,GAAG,IAAI,aAAa,EAAE;EACpD,KAAK;EACL,KAAK,OACH,QAAO,IAAI,YAAY,KAAK,GAAG,IAAI,aAAa,EAAE;EACpD,QACE,QAAO,IAAI,WAAW,KAAK,GAAG,IAAI,WAAW;;;AAInD,SAAS,WAAW,KAAkB,KAA+C;AACnF,SAAQ,KAAR;EACE,KAAK,MACH,QAAO,IAAI,YAAY,KAAK,GAAG,IAAI,aAAa,EAAE;EACpD,KAAK,MACH,QAAO,IAAI,WAAW,KAAK,GAAG,IAAI,aAAa,EAAE;EACnD,KAAK,MACH,QAAO,IAAI,YAAY,KAAK,GAAG,IAAI,aAAa,EAAE;EACpD,KAAK,KACH,QAAO,IAAI,WAAW,KAAK,GAAG,IAAI,WAAW;EAC/C,QACE,QAAO,IAAI,aAAa,KAAK,GAAG,IAAI,aAAa,EAAE;;;AAIzD,SAAS,OAAO,MAAgD;AAC9D,KAAI,gBAAgB,aAAc,QAAO;AACzC,KAAI,gBAAgB,YAAa,QAAO;AACxC,KAAI,gBAAgB,WAAY,QAAO;AACvC,KAAI,gBAAgB,YAAa,QAAO;AACxC,QAAO;;AAGT,MAAM,oBAAoB;AAE1B,IAAa,mBAAb,MAAa,iBAA+C;CAC1D,AAAQ,8BAAc,IAAI,KAA8B;;;;;;;CAOxD,AAAQ,4BAAY,IAAI,KAA0B;CAClD,AAAQ;;CAGR,OAAwB,iBAAiB,MAAM;CAE/C,YAAY,YAAoB,mBAAmB;AACjD,OAAK,YAAY;;CAGnB,OAAe,WAAW,MAAsB;AAC9C,SAAO,kBAAkB;;;;;;;;CAS3B,eACE,MACA,OACA,OACA,UACA,mBACM;AACN,OAAK,UAAU,OAAO,KAAK;AAC3B,OAAK,YAAY,IAAI,MAAM;GAAE;GAAO;GAAO,WAAW;GAAmB;GAAU,CAAC;;CAGtF,IAAI,MAAuB;AACzB,SAAO,KAAK,UAAU,IAAI,KAAK,IAAI,KAAK,YAAY,IAAI,KAAK;;CAG/D,OAAiB;EACf,MAAM,MAAM,IAAI,IAAY,KAAK,YAAY,MAAM,CAAC;AACpD,OAAK,MAAM,KAAK,KAAK,UAAU,MAAM,CAAE,KAAI,IAAI,EAAE;AACjD,SAAO,CAAC,GAAG,IAAI;;CAGjB,IAAI,OAAe;AACjB,SAAO,KAAK,MAAM,CAAC;;CAGrB,MAAM,IAAI,MAAgD;EACxD,MAAM,SAAS,KAAK,UAAU,IAAI,KAAK;AACvC,MAAI,OAAQ,QAAO;EACnB,MAAM,OAAO,KAAK,YAAY,IAAI,KAAK;AACvC,MAAI,CAAC,KAAM,QAAO;AAIlB,MAAI,KAAK,UAAU,OAAO;GAExB,MAAM,OAAO,OADC,MAAM,OAAO,KAAK,KAAK,UAAU,EACtB,MAAM,IAAI,QAAQ,KAAK,SAAS,CAAC;AAC1D,OAAI,CAAC,KAAM,QAAO;AAElB,UAAO;IAAE,MAAM,WADA,MAAM,KAAK,aAAa,EACL,KAAK,KAAK;IAAE,OAAO,KAAK;IAAO;;EASnE,IAAIQ,MAA0B;AAC9B,MAAI,mBAAmB,CACrB,OAAM,MAAM,OAAO,KAAK,SAAS;AAEnC,MAAI,CAAC,OAAO,eAAe,CACzB,OAAM,MAAM,SAAS,KAAK,SAAS;AAErC,MAAI,CAAC,KAAK;GAER,MAAM,OAAO,OADC,MAAM,OAAO,KAAK,KAAK,UAAU,EACtB,MAAM,IAAI,QAAQ,KAAK,SAAS,CAAC;AAC1D,OAAI,CAAC,KAAM,QAAO;AAClB,SAAM,MAAM,KAAK,aAAa;;EAEhC,IAAI,OAAO,mBAAmB,KAAK,KAAK,MAAM;AAE9C,MAAI,KAAK,UAAU,OACjB,QAAO,UAAU,IAAI,WAAW,KAAK,QAAQ,KAAK,YAAY,KAAK,WAAW,CAAC;WACtE,KAAK,UAAU,MACxB,QAAO,SAAS,KAAK;AAEvB,SAAO;GAAE;GAAM,OAAO,KAAK;GAAO;;CAGpC,MAAM,IAAI,MAAc,OAAmC;AACzD,OAAK,YAAY,OAAO,KAAK;AAG7B,MAAI,MAAM,KAAK,cAAc,iBAAiB,gBAAgB;AAC5D,QAAK,UAAU,IAAI,MAAM,MAAM;AAC/B;;AAEF,OAAK,UAAU,OAAO,KAAK;EAC3B,MAAM,WAAW,iBAAiB,WAAW,KAAK;EAClD,MAAM,QAAQ,MAAM,KAAK,OAAO,MAC9B,MAAM,KAAK,YACX,MAAM,KAAK,aAAa,MAAM,KAAK,WACpC;AAED,SADc,MAAM,OAAO,KAAK,KAAK,UAAU,EACnC,IAAI,IAAI,QAAQ,SAAS,EAAE,IAAI,SAAS,MAAM,CAAC;AAC3D,OAAK,YAAY,IAAI,MAAM;GACzB,OAAO,MAAM;GACb,OAAO;GACP,WAAW,KAAK;GAChB;GACA,MAAM,OAAO,MAAM,KAAK;GACzB,CAAC;;CAGJ,MAAM,OAAO,MAA6B;AACxC,OAAK,UAAU,OAAO,KAAK;EAC3B,MAAM,OAAO,KAAK,YAAY,IAAI,KAAK;AACvC,OAAK,YAAY,OAAO,KAAK;AAG7B,MAAI,QAAQ,KAAK,UAAU,SAAS,KAAK,cAAc,KAAK,UAC1D,KAAI;AAEF,UADc,MAAM,OAAO,KAAK,KAAK,UAAU,EACnC,OAAO,IAAI,QAAQ,KAAK,SAAS,CAAC;UACxC;;;CAOZ,MAAM,UAAyB;AAC7B,MAAI;AACF,SAAM,OAAO,OAAO,KAAK,UAAU;UAC7B;;;;AAOZ,SAAgB,wBAAiC;AAC/C,QAAO,OAAO,WAAW;;;;;;;;;;;;;;;;;AC3K3B,SAAS,iBAAiB,MAAc,UAA0B;AAEhE,KAAI,KAAK,WAAW,UAAU,IAAI,KAAK,WAAW,WAAW,CAC3D,QAAO;AAGT,QAAO,0BAA0B,KAAK,WAAW;;AAKnD,IAAIC,MAAuC;AAC3C,IAAIC,QAA2C;AAC/C,IAAI,iBAAiB;AAErB,eAAe,SAAwB;AACrC,KAAI,eAAgB;AACpB,kBAAiB;AACjB,KAAI;AACF,QAAM,MAAM,OAAO;AACnB,UAAQ,MAAM,OAAO;SACf;;AAKV,SAAS,aAAa,UAAkB,UAA0B;AAChE,QAAO,MAAO,KAAK,UAAU,SAAS,QAAQ,OAAO,IAAI,CAAC;;AAG5D,SAAS,eAAe,UAAkB,UAAiC;AACzE,KAAI,CAAC,OAAO,CAAC,MAAO,QAAO;CAC3B,MAAM,IAAI,aAAa,UAAU,SAAS;AAC1C,KAAI;AACF,SAAO,IAAI,aAAa,EAAE;SACpB;AACN,SAAO;;;AAIX,SAAS,eACP,UACA,UACA,MACM;AACN,KAAI,CAAC,OAAO,CAAC,MAAO;AACpB,KAAI,UAAU,UAAU,EAAE,WAAW,MAAM,CAAC;CAC5C,MAAM,IAAI,aAAa,UAAU,SAAS;CAC1C,IAAIC;AACJ,KAAI,OAAO,SAAS,KAAK,CACvB,OAAM;UACG,gBAAgB,WACzB,OAAM,OAAO,KAAK,KAAK;KAEvB,OAAM,OAAO,KAAK,KAAK;AAEzB,KAAI,cAAc,GAAG,IAAI;;AAiC3B,MAAM,qBAAqB;AAO3B,MAAM,wBACH,OAAO,YAAY,eAAe,QAAQ,KAAK,wBAAwB,QACvE,WAAiD,wBAAwB;AAE5E,IAAI,oBAAoB;AACxB,eAAe,qBAAoC;AACjD,KAAI,kBAAmB;AACvB,qBAAoB;AACpB,KAAI;AAEF,QAAO,WAAmB,SAAS,WAAW;SACxC;;;AAYV,eAAe,iBAAiB,UAA+C;AAC7E,QAAO,OAAO,SAAS;;;;;;;AAazB,eAAe,kBACb,UACA,MACA,gBAAgB,OACD;AACf,OAAM,oBAAoB;AAC1B,OAAM,OAAO,UAAU,KAAK;;;;;;;;;AAU9B,IAAI,gBAAgB;AACpB,eAAe,mBAAkC;AAC/C,KAAI,cAAe;AACnB,iBAAgB;AAChB,KAAI;AACF,MAAI,OAAO,WAAW,aAAa;GACjC,MAAM,OAAO,MAAM,OAAO,MAAM;AAChC,SAAM,QAAQ,IACZ,KAAK,QAAQ,MAAM,EAAE,WAAW,iBAAiB,CAAC,CAAC,KAAK,MAAM,OAAO,OAAO,EAAE,CAAC,CAChF;;SAEG;AAGR,OAAM,gBAAgB;;;;;AAQxB,eAAe,UACb,SACA,UACA,SACA,UACc;AAEd,KAAI,UAAU;EACZ,MAAM,SAAS,eAAe,UAAU,SAAS;AACjD,MAAI,OAAQ,QAAO,KAAK,MAAM,OAAO,SAAS,QAAQ,CAAC;;CAIzD,MAAM,aAAa,GAAG,QAAQ,GAAG;CACjC,MAAM,gBAAgB,MAAM,iBAAiB,WAAW;AACxD,KAAI,cACF,QAAO,KAAK,MAAM,IAAI,aAAa,CAAC,OAAO,cAAc,CAAC;CAG5D,MAAM,MAAM,GAAG,QAAQ,GAAG;CAC1B,MAAMC,UAAkC,EAAE;AAC1C,KAAI,QACF,SAAQ,gBAAgB,UAAU;CAEpC,MAAM,WAAW,MAAM,MAAM,KAAK,EAAE,SAAS,CAAC;AAC9C,KAAI,CAAC,SAAS,GACZ,OAAM,IAAI,MAAM,mBAAmB,IAAI,IAAI,SAAS,OAAO,GAAG,SAAS,aAAa;CAEtF,MAAM,OAAO,MAAM,SAAS,MAAM;AAGlC,KAAI,SACF,gBAAe,UAAU,UAAU,OAAO,KAAK,MAAM,QAAQ,CAAC;AAIhE,OAAM,kBAAkB,YAAY,IAAI,aAAa,CAAC,OAAO,KAAK,CAAC,OAAO;AAE1E,QAAO,KAAK,MAAM,KAAK;;;;;;;AAQzB,eAAe,UACb,SACA,UACA,SACA,UACwB;AACxB,KAAI,UAAU;EACZ,MAAM,SAAS,eAAe,UAAU,SAAS;AACjD,MAAI,OAAQ,QAAO,OAAO,SAAS,QAAQ;;CAE7C,MAAM,aAAa,GAAG,QAAQ,GAAG;CACjC,MAAM,gBAAgB,MAAM,iBAAiB,WAAW;AACxD,KAAI,cACF,QAAO,IAAI,aAAa,CAAC,OAAO,cAAc;CAEhD,MAAM,MAAM,GAAG,QAAQ,GAAG;CAC1B,MAAMA,UAAkC,EAAE;AAC1C,KAAI,QACF,SAAQ,gBAAgB,UAAU;CAEpC,MAAM,WAAW,MAAM,MAAM,KAAK,EAAE,SAAS,CAAC;AAC9C,KAAI,CAAC,SAAS,GAAI,QAAO;CACzB,MAAM,OAAO,MAAM,SAAS,MAAM;AAClC,KAAI,SACF,gBAAe,UAAU,UAAU,OAAO,KAAK,MAAM,QAAQ,CAAC;AAEhE,OAAM,kBAAkB,YAAY,IAAI,aAAa,CAAC,OAAO,KAAK,CAAC,OAAO;AAC1E,QAAO;;;;;AAMT,eAAe,YACb,SACA,UACA,SACA,YACA,UACsB;AAEtB,KAAI,UAAU;EACZ,MAAM,SAAS,eAAe,UAAU,SAAS;AACjD,MAAI,QAAQ;AACV,gBAAa,OAAO,YAAY,OAAO,WAAW;AAClD,UAAO,OAAO,OAAO,MACnB,OAAO,YACP,OAAO,aAAa,OAAO,WAC5B;;;CAIL,MAAM,MAAM,GAAG,QAAQ,GAAG;CAC1B,MAAMA,UAAkC,EAAE;AAC1C,KAAI,QACF,SAAQ,gBAAgB,UAAU;CAGpC,MAAM,WAAW,MAAM,MAAM,KAAK,EAAE,SAAS,CAAC;AAC9C,KAAI,CAAC,SAAS,GACZ,OAAM,IAAI,MAAM,mBAAmB,IAAI,IAAI,SAAS,OAAO,GAAG,SAAS,aAAa;CAGtF,MAAM,gBAAgB,OAAO,SAAS,QAAQ,IAAI,iBAAiB,IAAI,EAAE;AAEzE,KAAI,CAAC,SAAS,QAAQ,CAAC,YAAY;EACjC,MAAM,MAAM,MAAM,SAAS,aAAa;AACxC,MAAI,SAAU,gBAAe,UAAU,UAAU,IAAI;AACrD,SAAO;;CAIT,MAAM,SAAS,SAAS,KAAK,WAAW;CACxC,MAAMC,SAAuB,EAAE;CAC/B,IAAI,SAAS;AAEb,QAAO,MAAM;EACX,MAAM,EAAE,MAAM,UAAU,MAAM,OAAO,MAAM;AAC3C,MAAI,KAAM;AACV,SAAO,KAAK,MAAM;AAClB,YAAU,MAAM;AAChB,aAAW,QAAQ,cAAc;;CAInC,MAAM,SAAS,IAAI,WAAW,OAAO;CACrC,IAAI,SAAS;AACb,MAAK,MAAM,SAAS,QAAQ;AAC1B,SAAO,IAAI,OAAO,OAAO;AACzB,YAAU,MAAM;;AAGlB,KAAI,SAAU,gBAAe,UAAU,UAAU,OAAO;AACxD,QAAO,OAAO;;;;;;;;;;AAWhB,eAAe,yBACb,SACA,SACA,UACmB;AAEnB,KAAI;EACF,MAAM,QAAQ,MAAM,UAAU,SAAS,gCAAgC,SAAS,SAAS;AACzF,MAAI,MAAM,WAGR,QADc,CAAC,GAAG,IAAI,IAAI,OAAO,OAAO,MAAM,WAAW,CAAC,CAAC,CAC9C,MAAM;SAEf;AAIR,QAAO,CAAC,oBAAoB;;;;;AAM9B,eAAe,cACb,KACA,OACA,KACA,SACsB;CACtB,MAAMD,UAAkC,EACtC,OAAO,SAAS,MAAM,GAAG,MAAM,KAChC;AACD,KAAI,QACF,SAAQ,gBAAgB,UAAU;CAMpC,IAAI,aAAa;AACjB,MAAK,IAAI,UAAU,GAAG,UAAU,GAAG,WAAW;EAC5C,IAAIE;AACJ,MAAI;AACF,UAAO,MAAM,MAAM,KAAK,EAAE,SAAS,CAAC;UAC9B;AAEN,OAAI,UAAU,GAAG;AACf,UAAM,IAAI,SAAS,MAAM,WAAW,GAAG,MAAM,KAAK,QAAQ,CAAC;AAC3D;;AAEF,SAAM,IAAI,MAAM,2CAA2C,MAAM;;AAEnE,MAAI,KAAK,MAAM,KAAK,WAAW,IAAK,QAAO,KAAK,aAAa;AAC7D,eAAa,KAAK;AAGlB,OADkB,KAAK,WAAW,OAAO,KAAK,WAAW,OAAO,KAAK,UAAU,QAC9D,UAAU,GAAG;AAC5B,SAAM,IAAI,SAAS,MAAM,WAAW,GAAG,MAAM,KAAK,QAAQ,CAAC;AAC3D;;AAEF;;AAEF,OAAM,IAAI,MACR,yBAAyB,aAAa,eAAe,OAAO,eAAe,MAAM,mFAAmF,KACrK;;;;;;;;;;AAWH,eAAe,WACb,KACA,OACA,KACA,SACsB;CAEtB,MAAM,aAAa,GAAG,IAAI,UAAU,MAAM,GAAG;CAC7C,MAAM,gBAAgB,MAAM,iBAAiB,WAAW;AACxD,KAAI,cAAe,QAAO;CAE1B,MAAM,MAAM,MAAM,cAAc,KAAK,OAAO,KAAK,QAAQ;AAGzD,OAAM,kBAAkB,YAAY,IAAI;AAExC,QAAO;;;;;AAMT,eAAe,uBAAuB,KAAa,SAA4C;CAE7F,MAAM,YAAY,MAAM,WAAW,KAAK,GAAG,GAAG,QAAQ;AAKtD,QAAO,uBADW,MAAM,WAAW,KAAK,GAAG,IAHtB,OAAO,IAAI,SAAS,UAAU,CAAC,aAAa,GAAG,KAAK,CAAC,EAGb,QAAQ,CAC7B;;;;;;;;;;;;;;AAsB1C,SAAS,YACP,SACA,WACA,eAAuB,OAAO,MAC9B,gBAAwB,OAAO,mBAClB;AACb,KAAI,QAAQ,WAAW,EAAG,QAAO,EAAE;CAGnC,MAAM,SAAS,CAAC,GAAG,QAAQ,CAAC,MAAM,GAAG,MAAM,EAAE,aAAa,EAAE,WAAW;CAEvE,MAAMC,SAAsB,EAAE;CAC9B,IAAIC,UAAqB;EACvB,OAAO,YAAY,OAAO,GAAG;EAC7B,KAAK,YAAY,OAAO,GAAG,aAAa,OAAO,GAAG;EAClD,SAAS,CAAC,OAAO,GAAG;EACrB;AAED,MAAK,IAAI,IAAI,GAAG,IAAI,OAAO,QAAQ,KAAK;EACtC,MAAM,aAAa,YAAY,OAAO,GAAG;EACzC,MAAM,WAAW,aAAa,OAAO,GAAG;EAExC,MAAM,YAAY,aAAa,QAAQ,OAAO;EAC9C,MAAM,YAAY,WAAW,QAAQ,SAAS;AAC9C,MAAI,aAAa,WAAW;AAE1B,WAAQ,MAAM,KAAK,IAAI,QAAQ,KAAK,SAAS;AAC7C,WAAQ,QAAQ,KAAK,OAAO,GAAG;SAC1B;AACL,UAAO,KAAK,QAAQ;AACpB,aAAU;IAAE,OAAO;IAAY,KAAK;IAAU,SAAS,CAAC,OAAO,GAAG;IAAE;;;AAGxE,QAAO,KAAK,QAAQ;AACpB,QAAO;;;;;;;AAQT,SAAS,WAAW,OAAuB;AACzC,KAAI,UAAU,SAAS,UAAU,SAAS,UAAU,MAAO,QAAO;AAClE,KAAI,UAAU,SAAS,UAAU,SAAS,UAAU,MAAO,QAAO;AAClE,KAAI,UAAU,SAAS,UAAU,UAAU,UAAU,SAAS,UAAU,MAAO,QAAO;AACtF,QAAO;;;AAIT,SAAS,eACP,KACA,YACA,OACiB;CAGjB,MAAM,UAAU,aAAa,WAAW,MAAM,MAAM,KAAK;CACzD,MAAM,MAAM,UAAU,MAAM,IAAI,MAAM,YAAY,aAAa,MAAM,WAAW;CAChF,MAAM,MAAM,UAAU,aAAa;AACnC,SAAQ,MAAM,OAAd;EACE,KAAK,MACH,QAAO,IAAI,aAAa,KAAK,KAAK,MAAM,aAAa,EAAE;EACzD,KAAK,MACH,QAAO,IAAI,WAAW,KAAK,KAAK,MAAM,aAAa,EAAE;EACvD,KAAK,MACH,QAAO,IAAI,YAAY,KAAK,KAAK,MAAM,aAAa,EAAE;EACxD,KAAK;EACL,KAAK,OACH,QAAO,IAAI,YAAY,KAAK,KAAK,MAAM,aAAa,EAAE;EACxD,QACE,QAAO,IAAI,WAAW,KAAK,KAAK,MAAM,WAAW;;;;;;;;;;AAWvD,SAAS,eAAe,KAAa,MAAuB,OAAgC;AAE1F,QAAO,GAAG,IAAI,KADG,KAAK,YAAY,MAAM,WACZ,GAAG,MAAM;;;;;;;;;;;;;;;;;;;AAoBvC,eAAe,sBACb,KACA,MACA,eACA,SACA,YACA,YACkE;CAClE,MAAM,0BAAU,IAAI,KAAyD;CAC7E,MAAM,aAAa,cAAc,QAAQ,KAAK,MAAM,MAAM,EAAE,YAAY,EAAE;CAC1E,IAAI,cAAc;CAMlB,MAAM,gBAHJ,OAAO,cAAc,eAAe,4BAA4B,KAAK,UAAU,UAAU,GAG1D,KAAK,OAAO,OAAO,MAAM,OAAO;CAKjE,MAAM,UAAU,OAAwB,QAAsB;AAC5D,MAAI,YAAY;GACd,MAAM,YAAY,WAAW,UAAU,MAAM,KAAK;AAClD,OAAI,UAEF,YAAW,MAAM,eACf,WACA,MAAM,OACN,MAAM,OACN,KACA,mBACD;AAEH;;;AAWJ,KAAI,qBAAsB,CAAK,kBAAkB;CASjD,MAAMC,WAA8B,EAAE;CAStC,MAAM,aAAa,uBAAuB,MAAM,YAAY,mBAAG,IAAI,KAAa;CAChF,MAAM,WAAW,cAAc,KAAK,UAAU,WAAW,IAAI,eAAe,KAAK,MAAM,MAAM,CAAC,CAAC;AAC/F,MAAK,IAAI,IAAI,GAAG,IAAI,cAAc,QAAQ,KAAK;EAC7C,MAAM,QAAQ,cAAc;EAC5B,MAAM,MAAM,eAAe,KAAK,MAAM,MAAM;AAC5C,MAAI,CAAC,SAAS,IAAI;AAChB,YAAS,KAAK,MAAM;AACpB;;AAEF,MAAI,YAAY;AAEd,UAAO,OAAO,IAAI;AAClB,kBAAe,MAAM;AACrB,gBAAa,aAAa,WAAW;AACrC;;EAKF,MAAM,MAAM,MAAM,iBAAiB,IAAI;AACvC,MAAI,OAAO,IAAI,cAAc,MAAM,YAAY;AAC7C,WAAQ,IAAI,MAAM,MAAM;IACtB,MAAM,eAAe,KAAK,GAAG,MAAM;IACnC,OAAO,MAAM;IACd,CAAC;AACF,kBAAe,MAAM;AACrB,gBAAa,aAAa,WAAW;QAErC,UAAS,KAAK,MAAM;;AAQxB,KAAI,sBAAsB;EACxB,MAAM,OAAO,cAAc,SAAS,SAAS;EAE7C,MAAM,MAAM,SADI,mBAAmB,GAAG,QAAQ,OACjB,KAAK,KAAK,GAAG,cAAc,OAAO,mBAAmB,SAAS,OAAO;AAClG,UAAQ,IAAI,YAAY,MAAM;AAC9B,eAAa,aAAa,YAAY,IAAI;;CAK5C,MAAM,SAAS,YAAY,UAAU,KAAK,WAAW,OAAO,MAAM,cAAc;AAChF,MAAK,MAAM,SAAS,QAAQ;EAC1B,MAAM,WAAW,MAAM,cAAc,KAAK,MAAM,OAAO,MAAM,KAAK,QAAQ;EAI1E,MAAM,qBAAqB,MAAM,MAAM,MAAM;AAC7C,MAAI,SAAS,aAAa,mBACxB,OAAM,IAAI,MACR,mCAAmC,SAAS,WAAW,MAAM,mBAAmB,aAAa,IAAI,UAClG;AAGH,OAAK,MAAM,SAAS,MAAM,SAAS;GACjC,MAAM,cAAc,KAAK,YAAY,MAAM,aAAa,MAAM;GAC9D,MAAM,MAAM,eAAe,KAAK,MAAM,MAAM;GAE5C,MAAM,cAAc,SAAS,MAAM,aAAa,cAAc,MAAM,WAAW;AAE/E,OAAI,wBAAwB,YAAY,eAAe,MAAM,WAK3D,OAAM,kBAAkB,KAAK,aAAa,KAAK;AAEjD,OAAI,WAEF,QAAO,OAAO,IAAI;OAElB,SAAQ,IAAI,MAAM,MAAM;IACtB,MAAM,eAAe,UAAU,aAAa,MAAM;IAClD,OAAO,MAAM;IACd,CAAC;AAEJ,kBAAe,MAAM;AACrB,gBAAa,aAAa,WAAW;;;AAOzC,QAAO;;;;;;;;;;;;;;;AAgBT,SAAS,uBAAuB,kBAA0B,aAAa,OAAoB;AAKzF,KAAI,qBAAqB,qBAAqB,qBAAqB,sBAOjE,SAAQ,UAAiC;EACvC,IAAI,MAAM;AACV,MAAI,IAAI,WAAW,SAAS,CAC1B,OAAM,IAAI,MAAM,EAAE;AAEpB,QAAM,IAAI,QAAQ,2BAA2B,qBAAqB;AAClE,SAAO;;AAGX,KAAI,qBAAqB,kCACvB,SAAQ,UAAiC;AAUvC,MAAI,MAAM,WAAW,gBAAgB,CACnC,QAAO,aAAa,MAAM,MAAM,EAAE,GAAG;AAEvC,MAAI,MAAM,WAAW,gBAAgB,CAEnC,QAAO,aAAa,UAAU,MAAM,MAAM,GAAG,KAAK;AAIpD,MAAI,MAAM,WAAW,OAAO,CAC1B,QAAO;EAET,IAAI,MAAM;AAEV,MAAI,IAAI,WAAW,wBAAwB,CACzC,OAAM,IAAI,MAAM,GAAG;WACV,IAAI,WAAW,wBAAwB,CAChD,OAAM,IAAI,MAAM,GAAG;WACV,IAAI,WAAW,SAAS,CACjC,OAAM,IAAI,MAAM,EAAE;AAEpB,SAAO;;AAGX,KAAI,qBAAqB,iCASvB,SAAQ,UAAiC;AAMvC,MAAI,MAAM,SAAS,gBAAgB,EAAE;AACnC,OAAI,CAAC,WAAY,QAAO;GACxB,MAAM,MAAM,MAAM,QAAQ,gBAAgB;AAC1C,UAAO,MAAM,MAAM,IAAI;;AAEzB,MAAI,MAAM,SAAS,gBAAgB,EAAE;AACnC,OAAI,CAAC,WAAY,QAAO;GACxB,MAAM,MAAM,MAAM,QAAQ,gBAAgB;AAC1C,UAAO,MAAM,MAAM,IAAI;;AAGzB,MAAI,MAAM,SAAS,eAAe,IAAI,MAAM,SAAS,eAAe,CAClE,QAAO;EAET,IAAI,MAAM;AACV,MAAI,IAAI,WAAW,wBAAwB,CACzC,OAAM,IAAI,MAAM,GAAG;WACV,IAAI,WAAW,wBAAwB,CAChD,OAAM,IAAI,MAAM,GAAG;WACV,IAAI,WAAW,kBAAkB,CAC1C,OAAM,IAAI,MAAM,GAAG;WACV,IAAI,WAAW,SAAS,CACjC,OAAM,IAAI,MAAM,EAAE;AAEpB,SAAO;;AAGX,KAAI,qBAAqB,oCAUvB,QAAO,0BAA0B;AAEnC,QAAO,0BAA0B;;AAGnC,MAAM,iBAAiB;;;;;;;;;;;;;AAcvB,eAAe,eACb,QACA,QACA,OACA,OACA,WACA,YACoB;AACpB,KAAI,CAAC,WACH,QAAO;EAAE;EAAQ;EAAQ;EAAO;EAAO;EAAW,cAAc;EAAiB;CAEnF,MAAM,QAAQ,MAAM,OAAO,KAAK,eAAe;CAC/C,MAAM,YAAY;CAClB,MAAM,YAAY;CAClB,MAAM,WAAW;CACjB,MAAM,QAAQ,MACZ,EAAE,OAAO,MAAM,EAAE,YAAY,EAAE,aAAa,EAAE,WAAW;AAC3D,OAAM,MAAM,IAAI,IAAI,QAAQ,UAAU,EAAE,IAAI,SAAS,KAAK,OAAO,CAAC,CAAC;AACnE,OAAM,MAAM,IAAI,IAAI,QAAQ,UAAU,EAAE,IAAI,SAAS,KAAK,OAAO,CAAC,CAAC;AACnE,OAAM,MAAM,IAAI,IAAI,QAAQ,SAAS,EAAE,IAAI,SAAS,KAAK,MAAM,CAAC,CAAC;AACjE,QAAO;EAEL,QAAQ,IAAI,YAAY,EAAE;EAC1B,QAAQ,IAAI,aAAa,EAAE;EAC3B,OAAO,IAAI,aAAa,EAAE;EAC1B;EACA;EACA,cAAc;EACd,OAAO;GACL,WAAW;GACX;GACA;GACA;GACA,WAAW,OAAO;GACnB;EACF;;;;;;;;;;AAWH,eAAsB,UAAU,SAAiD;CAC/E,MAAM,EAAE,MAAM,YAAY,SAAS,WAAW,QAAQ,UAAU,UAAU;AAG1E,OAAM,QAAQ;CAGd,IAAI,mBAAmB;AACvB,KAAI,CAAC,oBAAoB,OAAO,OAAO;EACrC,MAAM,OAAO,QAAQ,IAAI,QAAQ,QAAQ,IAAI,eAAe;AAC5D,MAAI,KACF,oBAAmB,MAAM,KAAK,MAAM,UAAU,UAAU,KAAK,QAAQ,OAAO,IAAI,EAAE,SAAS;;CAI/F,MAAM,UAAU,iBAAiB,MAAM,SAAS;AAGhD,cAAa,GAAG,KAAK,2BAA2B;CAChD,MAAM,YAAY,MAAM,UAAU,SAAS,eAAe,SAAS,iBAAiB;CAEpF,MAAM,mBAAmB,UAAU,gBAAgB;AACnD,KAAI,CAAC,iBACH,OAAM,IAAI,MAAM,4CAA4C;CAO9D,MAAM,cAAe,UAAU,uBAAuB,UAAU;CAGhE,MAAM,SAAS,aAAa,iBAAiB;CAW7C,MAAM,YAAY,aAAa;CAC/B,MAAM,cACJ,CAAC,UACD,aAAa,SAAS,MACrB,cAAc,YACZ,cAAc,UAAa,OAAO,aAAa,eAAe;CAKnE,MAAM,qBAAqB,CAAC,yCAAyC;CACrE,MAAM,YAAY,KAAK,aAAa;CAEpC,MAAM,oBACJ,CAFY,UAAU,SAAS,MAAM,IAE3B,mBAAmB,MAAM,MAAM,UAAU,SAAS,EAAE,aAAa,CAAC,CAAC;CAI/E,MAAM,QAAQ,gBAAgB,cAAc,YAAY;CAKxD,MAAM,aACJ,OAAO,cAAc,eAAe,4BAA4B,KAAK,UAAU,UAAU;CAG3F,MAAMC,gBAAwC,UAAU,QAAQ,OAD9D,UAAU,SAAU,aAAa,OAAO,SAAa;CAIvD,MAAM,iBAAiB,QAAS,aAAa,aAAwB;AAErE,KAAI,OACF,cAAa,GAAG,KAAK,sDAAsD;UAClE,MACT,cACE,GACA,KACA,wCAAwC,eAAe,oBACxD;CASH,MAAM,aAAa,QAAQ,QAAQ,cAAc,QAAQ,gBAAgB;CACzE,MAAM,YAAY,QAAQ,aAAa,uBAAuB,kBAAkB,WAAW;CAG3F,MAAM,QAAQ,cACZ,kBACA,WACA,eACA,gBACA,QAAQ,SACR,QAAQ,WACR,QAAQ,WACT;AAGD,cAAa,GAAG,KAAK,wBAAwB;CAC7C,MAAM,CAAC,eAAe,qBAAqB,qBAAqB,MAAM,QAAQ,IAAI;EAChF,UAAU,SAAS,kBAAkB,SAAS,iBAAiB;EAC/D,UAAU,SAAS,yBAAyB,SAAS,iBAAiB,CAAC,YAAY,KAAK;EAGxF,UAAU,SAAS,uBAAuB,SAAS,iBAAiB,CAAC,YAAY,KAAK;EACvF,CAAC;CAGF,MAAM,qBACJ,uBAAuB,CAAC,oBAAoB,iBAAiB,oBACzD;EAAE,GAAG;EAAqB,eAAe;EAAmB,GAC3D,wBAAwB,oBAAoB,EAAE,eAAe,mBAAmB,GAAG;CAE1F,MAAM,YAAY,UAAU,SAAS,eAAe,mBAAmB;AAGvE,cAAa,IAAI,KAAK,8BAA8B;CACpD,MAAM,mBAAmB,MAAM,yBAAyB,SAAS,SAAS,iBAAiB;CAW3F,MAAM,aAAc,WAAmD;CACvE,MAAM,cACJ,eAAe,QACd,OAAO,aAAa,eACnB,OAAO,oBAAoB,eAC3B,IAAI,gBAAgB,SAAS,OAAO,CAAC,IAAI,SAAS;CAGtD,MAAM,aADJ,eAAe,UAAU,eAAe,eAAe,uBAAuB,GAC5C,IAAI,kBAAkB,GAAG;CAC7D,MAAMC,QAA4B,cAAc,IAAI,iBAAiB;CAErE,IAAI,cAAc;CAClB,MAAM,aAAa,iBAAiB;AAEpC,MAAK,MAAM,YAAY,kBAAkB;EACvC,MAAM,UAAU,GAAG,QAAQ,GAAG;EAG9B,IAAIC,eAAmC;AACvC,MAAI,kBAAkB;GACpB,MAAM,SAAS,eAAe,kBAAkB,SAAS;AACzD,OAAI,OACF,gBAAe,OAAO,OAAO,MAC3B,OAAO,YACP,OAAO,aAAa,OAAO,WAC5B;;AAIL,MAAI,cAAc;GAEhB,MAAM,OAAO,uBAAuB,aAAa;AACjD,QAAK,MAAM,SAAS,KAAK,SAAS;IAChC,MAAM,gBAAgB,UAAU,MAAM,KAAK;AAC3C,QAAI,CAAC,cAAe;IACpB,IAAIC,OAAwB,cAAc,cAAc,MAAM,MAAM;AACpE,QAAI,MAAM,UAAU,OAClB,QAAO,UAAU,IAAI,WAAW,KAAK,QAAQ,KAAK,YAAY,KAAK,WAAW,CAAC;AAEjF,QAAI,MAAM,UAAU,MAClB,QAAO,SAAS,KAAK;AAEvB,UAAM,MAAM,IAAI,eAAe;KAAE;KAAM,OAAO,MAAM;KAAO,CAAC;;SAEzD;GAEL,MAAM,aAAa,KAAK;GACxB,MAAM,OAAO,KAAK,cAAc;AAKhC,gBAAa,MAAM,KAAK,WAAW,SAAS,WAAW,cAAc,EAAE,GAAG,WAAW,MAAM;GAC3F,MAAM,OAAO,MAAM,uBAAuB,SAAS,QAAQ;GAG3D,MAAM,gBAAgB,KAAK,QAAQ,QAAQ,MAAM,UAAU,EAAE,KAAK,KAAK,KAAK;GAC5E,MAAM,eAAe,KAAK,QACvB,QAAQ,MAAM,UAAU,EAAE,KAAK,KAAK,KAAK,CACzC,QAAQ,KAAK,MAAM,MAAM,EAAE,YAAY,EAAE;GAG5C,MAAM,WAFc,cAAc,QAAQ,KAAK,MAAM,MAAM,EAAE,YAAY,EAAE,GAE5C,SAAS,QAAQ,EAAE;AAClD,OAAI,eAAe,GAAG;IACpB,MAAM,WAAW,eAAe,SAAS,QAAQ,EAAE;AACnD,iBACE,MACA,KACA,uBAAuB,QAAQ,uBAAuB,QAAQ,kBAC/D;;AAGH,gBAAa,MAAM,KAAK,eAAe,SAAS,MAAM,QAAQ,MAAM;GAGpE,IAAI,kBAAkB;GAEtB,MAAM,aAAa,MAAM,sBACvB,SACA,MACA,eACA,UACC,QAAQ,UAAU;IACjB,MAAM,UAAU,QAAQ,IAAK,SAAS,QAAS,aAAa;IAC5D,MAAM,aAAa,KAAK,MAAM,OAAO,QAAQ;AAC7C,QAAI,cAAc,gBAAiB;AACnC,sBAAkB;IAClB,MAAM,SACJ,QAAQ,IACJ,MAAM,SAAS,SAAS,QAAQ,EAAE,CAAC,IAAI,QAAQ,SAAS,QAAQ,EAAE,CAAC,QACnE;AACN,iBAAa,OAAO,SAAS,KAAK,eAAe,WAAW,SAAS;MAKvE,aAAa;IAAE,OAAO;IAAY;IAAW,GAAG,OACjD;AAMD,QAAK,MAAM,CAAC,QAAQ,EAAE,MAAM,SAAS,YAAY,YAAY;IAC3D,MAAM,gBAAgB,UAAU,OAAO;AACvC,QAAI,CAAC,cAAe;IAEpB,IAAIA,OAAwB;IAE5B,MAAM,QAAQ,cAAc,MAAM,MAAM,EAAE,SAAS,OAAO;AAC1D,QAAI,OAAO;AACT,SAAI,MAAM,UAAU,OAClB,QAAO,UAAU,IAAI,WAAW,KAAK,QAAQ,KAAK,YAAY,KAAK,WAAW,CAAC;AAEjF,SAAI,MAAM,UAAU,MAClB,QAAO,SAAS,KAAK;;AAIzB,UAAM,MAAM,IAAI,eAAe;KAAE;KAAM;KAAO,CAAC;;;AAInD;;AAGF,cAAa,IAAI,KAAK,kBAAkB;AASxC,KAAI,qBAAqB,mCAAmC;EAC1D,MAAM,UAAW,UAAU,eAA2C;EACtE,MAAM,YAAY,QAAQ;EAC1B,MAAM,gBAAiB,QAAQ,eAA4B,EAAE;EAC7D,MAAM,mBAAoB,QAAQ,2BAAsC;EACxE,MAAM,WAAW,QAAQ;EACzB,MAAM,UACH,QAAQ,YAAuB,KAAK,MAAO,QAAQ,cAAyB,SAAS;AAExF,OAAK,IAAI,IAAI,GAAG,IAAI,WAAW,KAAK;GAElC,IAAIC;AACJ,OAAI,cAAc,SAAS,EACzB,cAAa,cAAc,OAAO;OAElC,cAAa,IAAI,qBAAqB,mBAAmB;AAE3D,OAAI,CAAC,WAAY;GAEjB,MAAM,WAAW,UAAU,EAAE;GAC7B,MAAM,cAAc,UAAU,EAAE;GAChC,MAAM,cAAc,MAAM,MAAM,IAAI,SAAS;AAE7C,OAAI,eAAe,CAAC,MAAM,IAAI,YAAY,IAAI,CAAC,OAAO;IACpD,MAAM,YAAY,YAAY;IAC9B,MAAM,CAAC,WAAW,QAAQ,YAAY;IACtC,MAAM,WAAW,YAAY;IAE7B,MAAM,QAAQ,IAAI,aAAa,WAAW,KAAK;IAC/C,MAAM,WAAW,IAAI,aAAa,WAAW,KAAK;IAElD,MAAM,MACJ,qBAAqB,eACjB,YACA,IAAI,aAAa,UAAU,QAAQ,UAAU,YAAY,YAAY,KAAK;IAIhF,MAAM,YAAY,IAAI;AACtB,SAAK,IAAI,IAAI,GAAG,IAAI,UAAU,KAAK;KACjC,MAAM,UAAU,IAAI,YAAY;KAChC,MAAM,UAAU,IAAI,UAAU;AAC9B,WAAM,IAAI,IAAI,SAAS,SAAS,UAAU,UAAU,KAAK,EAAE,QAAQ;AACnE,cAAS,IAAI,IAAI,SAAS,UAAU,UAAU,MAAM,UAAU,YAAY,KAAK,EAAE,QAAQ;;AAG3F,UAAM,MAAM,IAAI,UAAU;KAAE,MAAM;KAAO,OAAO,CAAC,UAAU,KAAK;KAAE,CAAC;AACnE,UAAM,MAAM,IAAI,aAAa;KAAE,MAAM;KAAU,OAAO,CAAC,UAAU,KAAK;KAAE,CAAC;;GAO3E,MAAM,eAAe,UAAU,EAAE;GACjC,MAAM,kBAAkB,UAAU,EAAE;AACpC,OAAI,MAAM,IAAI,aAAa,IAAI,CAAC,MAAM,IAAI,gBAAgB,EAAE;IAC1D,MAAM,SAAS,WAAW,IAAI;IAC9B,MAAM,QAAQ,WAAW;IACzB,MAAM,QAAQ,QAAQ;IAGtB,MAAM,KAAM,MAAM,MAAM,IAAI,aAAa;IACzC,MAAM,QACJ,GAAG,gBAAgB,aACf,GAAG,OACH,IAAI,WAAW,GAAG,KAAK,QAAQ,GAAG,KAAK,YAAY,GAAG,KAAK,aAAa,EAAE;IAChF,MAAM,QAAQ,UAAU;IACxB,MAAM,MAAM,IAAI,WAAW,QAAQ,MAAM;IACzC,MAAM,MAAM,IAAI,WAAW,QAAQ,MAAM;IACzC,MAAM,aAAa,IAAI;AACvB,SAAK,IAAI,MAAM,GAAG,MAAM,OAAO,MAC7B,MAAK,IAAI,IAAI,GAAG,IAAI,UAAU,KAAK;KACjC,MAAM,SAAS,MAAM,SAAS,IAAI;KAClC,MAAM,SAAS,MAAM,QAAQ,IAAI;AACjC,UAAK,IAAI,IAAI,GAAG,IAAI,SAAS,KAAK;AAChC,UAAI,SAAS,KAAK,MAAM,SAAS;AACjC,UAAI,SAAS,KAAK,MAAM,SAAS,UAAU;;;AAIjD,UAAM,MAAM,IAAI,cAAc;KAAE,MAAM;KAAK,OAAO,CAAC,OAAO,MAAM;KAAE,CAAC;AACnE,UAAM,MAAM,IAAI,iBAAiB;KAAE,MAAM;KAAK,OAAO,CAAC,OAAO,MAAM;KAAE,CAAC;IAGtE,MAAM,cAAc,UAAU,EAAE;IAChC,MAAM,iBAAiB,UAAU,EAAE;IACnC,MAAM,KAAM,MAAM,MAAM,IAAI,YAAY;IACxC,MAAM,QACJ,GAAG,gBAAgB,eACf,GAAG,OACH,IAAI,aAAa,GAAG,KAAK,QAAQ,GAAG,KAAK,YAAY,GAAG,KAAK,aAAa,EAAE;IAClF,MAAM,SAAS,GAAG,MAAM;IACxB,MAAM,MAAM,IAAI,aAAa,SAAS,MAAM;IAC5C,MAAM,MAAM,IAAI,aAAa,SAAS,MAAM;AAC5C,SAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,IAC1B,MAAK,IAAI,IAAI,GAAG,IAAI,UAAU,KAAK;KACjC,MAAM,SAAS,IAAI,SAAS,IAAI;KAChC,MAAM,SAAS,IAAI,QAAQ,IAAI;AAC/B,UAAK,IAAI,IAAI,GAAG,IAAI,SAAS,KAAK;AAChC,UAAI,SAAS,KAAK,MAAM,SAAS;AACjC,UAAI,SAAS,KAAK,MAAM,SAAS,UAAU;;;AAIjD,UAAM,MAAM,IAAI,aAAa;KAAE,MAAM;KAAK,OAAO,CAAC,QAAQ,MAAM;KAAE,CAAC;AACnE,UAAM,MAAM,IAAI,gBAAgB;KAAE,MAAM;KAAK,OAAO,CAAC,QAAQ,MAAM;KAAE,CAAC;IAKtE,MAAM,cAAc,UAAU,EAAE;IAChC,MAAM,iBAAiB,UAAU,EAAE;IACnC,MAAM,KAAM,MAAM,MAAM,IAAI,YAAY;IACxC,MAAM,QACJ,GAAG,gBAAgB,aACf,GAAG,OACH,IAAI,WAAW,GAAG,KAAK,QAAQ,GAAG,KAAK,YAAY,GAAG,KAAK,aAAa,EAAE;IAChF,MAAM,aAAa,WAAW;IAC9B,MAAM,YAAY,UAAU;IAC5B,MAAM,gBAAgB,eAAe;IACrC,MAAM,cAAc,YAAY;IAChC,MAAM,MAAM,IAAI,WAAW,SAAS,UAAU;IAC9C,MAAM,MAAM,IAAI,WAAW,SAAS,UAAU;AAC9C,SAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,IAC1B,MAAK,IAAI,IAAI,GAAG,IAAI,UAAU,KAAK;KACjC,MAAM,SAAS,IAAI,aAAa,IAAI;KACpC,MAAM,SAAS,IAAI,YAAY,IAAI;AACnC,UAAK,IAAI,IAAI,GAAG,IAAI,aAAa,KAAK;AACpC,UAAI,SAAS,KAAK,MAAM,SAAS;AACjC,UAAI,SAAS,KAAK,MAAM,SAAS,cAAc;;;AAIrD,UAAM,MAAM,IAAI,aAAa;KAAE,MAAM;KAAK,OAAO,CAAC,QAAQ,UAAU;KAAE,CAAC;AACvE,UAAM,MAAM,IAAI,gBAAgB;KAAE,MAAM;KAAK,OAAO,CAAC,QAAQ,UAAU;KAAE,CAAC;AAG1E,UAAM,MAAM,OAAO,UAAU,EAAE,yBAAyB;;GAM1D,MAAM,iBAAiB,UAAU,EAAE;GACnC,MAAM,oBAAoB,UAAU,EAAE;AACtC,OAAI,SAAS,MAAM,IAAI,eAAe,IAAI,CAAC,MAAM,IAAI,kBAAkB,EAAE;IACvE,MAAM,QAAQ,QAAQ;IACtB,MAAM,SAAS,WAAW,IAAI;IAC9B,MAAM,QAAQ,WAAW;IACzB,MAAM,aAAa,IAAI;IACvB,MAAM,QAAS,aAAa,cAAyB;IAGrD,MAAM,mBACJ,KACA,WACA,MACA,SACW;KACX,MAAM,WAAW,YAAY;KAC7B,MAAM,OAAO,IAAI,KAAK,WAAW,KAAK;KACtC,MAAM,OAAO,IAAI,KAAK,WAAW,KAAK;AACtC,UAAK,IAAI,IAAI,GAAG,IAAI,UAAU,KAAK;MACjC,MAAM,UAAU,IAAI,aAAa;MACjC,MAAM,UAAU,IAAI,UAAU;AAC9B,WAAK,IAAI,IAAI,GAAG,IAAI,SAAS,KAAK;OAChC,MAAM,KAAK,UAAU,IAAI;OACzB,MAAM,KAAK,WAAW,UAAU,KAAK;OACrC,MAAM,KAAK,UAAU,IAAI;AACzB,YAAK,IAAI,IAAI,GAAG,IAAI,MAAM,KAAK;AAC7B,aAAK,KAAK,KAAK,IAAI,KAAK;AACxB,aAAK,KAAK,KAAK,IAAI,KAAK;;;;AAI9B,YAAO,CAAC,MAAM,KAAK;;IAIrB,MAAM,KAAM,MAAM,MAAM,IAAI,eAAe;IAC3C,MAAM,QACJ,GAAG,gBAAgB,cACf,GAAG,OACH,IAAI,YAAY,GAAG,KAAK,QAAQ,GAAG,KAAK,YAAY,GAAG,KAAK,aAAa,EAAE;IACjF,MAAM,QAAQ,UAAU;IACxB,MAAM,CAAC,KAAK,OAAO,gBAAgB,OAAO,QAAQ,OAAO,YAAY;AACrE,UAAM,MAAM,IAAI,gBAAgB;KAAE,MAAM;KAAK,OAAO,CAAC,OAAO,MAAM;KAAE,CAAC;AACrE,UAAM,MAAM,IAAI,mBAAmB;KAAE,MAAM;KAAK,OAAO,CAAC,OAAO,MAAM;KAAE,CAAC;IAGxE,MAAM,iBAAiB,UAAU,EAAE;IACnC,MAAM,oBAAoB,UAAU,EAAE;IACtC,MAAM,KAAM,MAAM,MAAM,IAAI,eAAe;IAC3C,MAAM,QACJ,GAAG,gBAAgB,eACf,GAAG,OACH,IAAI,aAAa,GAAG,KAAK,QAAQ,GAAG,KAAK,YAAY,GAAG,KAAK,aAAa,EAAE;IAClF,MAAM,QAAQ,KAAK,KAAK,QAAQ,MAAM;IACtC,MAAM,CAAC,KAAK,OAAO,gBAAgB,OAAO,QAAQ,OAAO,aAAa;AACtE,UAAM,MAAM,IAAI,gBAAgB;KAAE,MAAM;KAAK,OAAO,CAAC,OAAO,MAAM;KAAE,CAAC;AACrE,UAAM,MAAM,IAAI,mBAAmB;KAAE,MAAM;KAAK,OAAO,CAAC,OAAO,MAAM;KAAE,CAAC;IAGxE,MAAM,iBAAiB,UAAU,EAAE;IACnC,MAAM,oBAAoB,UAAU,EAAE;IACtC,MAAM,KAAM,MAAM,MAAM,IAAI,eAAe;IAK3C,MAAM,CAAC,KAAK,OAAO,gBAHjB,GAAG,gBAAgB,eACf,GAAG,OACH,IAAI,aAAa,GAAG,KAAK,QAAQ,GAAG,KAAK,YAAY,GAAG,KAAK,aAAa,EAAE,EACxC,QAAQ,OAAO,aAAa;AACtE,UAAM,MAAM,IAAI,gBAAgB;KAAE,MAAM;KAAK,OAAO,CAAC,OAAO,MAAM;KAAE,CAAC;AACrE,UAAM,MAAM,IAAI,mBAAmB;KAAE,MAAM;KAAK,OAAO,CAAC,OAAO,MAAM;KAAE,CAAC;;;;AAc9E,KAAI,qBAAqB,qCAAqC,CAAC,OAAO;EACpE,MAAM,UAAW,UAAU,eAA2C;EACtE,MAAM,YAAY,QAAQ;EAC1B,MAAM,gBAAiB,QAAQ,eAA4B,EAAE;EAC7D,MAAM,mBAAoB,QAAQ,2BAAsC;EAGxE,MAAMC,WAAqB,EAAE;AAC7B,OAAK,IAAI,IAAI,GAAG,IAAI,WAAW,KAAK;AAElC,YAAS,KAAK,UAAU,EAAE,yBAAyB;AACnD,YAAS,KAAK,UAAU,EAAE,kCAAkC;GAE5D,IAAID;AACJ,OAAI,cAAc,SAAS,EACzB,cAAa,cAAc,OAAO;OAElC,cAAa,IAAI,qBAAqB,mBAAmB;AAG3D,OAAI,YAAY;AACd,aAAS,KAAK,UAAU,EAAE,0BAA0B;AACpD,aAAS,KAAK,UAAU,EAAE,0BAA0B;;;AAMxD,WAAS,KAAK,cAAc;AAE5B,OAAK,MAAM,WAAW,UAAU;GAC9B,MAAM,QAAQ,MAAM,MAAM,IAAI,QAAQ;AACtC,OAAI,OAAO;IACT,MAAM,OACJ,MAAM,gBAAgB,eAClB,MAAM,OACN,IAAI,aAAa,MAAM,KAAK,QAAQ,MAAM,KAAK,YAAY,MAAM,KAAK,aAAa,EAAE;IAC3F,MAAM,WAAW,IAAI,aAAa,KAAK,OAAO;AAC9C,SAAK,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,IAC/B,UAAS,KAAK,IAAM,KAAK;AAE3B,UAAM,MAAM,IAAI,SAAS;KAAE,MAAM;KAAU,OAAO,MAAM;KAAO,CAAC;;;;AAatE,KAAI,qBAAqB,qBAAqB,qBAAqB,eAAe;EAChF,MAAM,iBAAiB,UAAU;EACjC,MAAME,gBAA0B,EAAE;AAClC,OAAK,IAAI,IAAI,GAAG,IAAI,gBAAgB,KAAK;AACvC,iBAAc,KAAK,UAAU,EAAE,yBAAyB;AACxD,iBAAc,KAAK,UAAU,EAAE,kCAAkC;AACjE,iBAAc,KAAK,UAAU,EAAE,mCAAmC;AAClE,iBAAc,KAAK,UAAU,EAAE,oCAAoC;AACnE,iBAAc,KAAK,UAAU,EAAE,0BAA0B;AACzD,iBAAc,KAAK,UAAU,EAAE,0BAA0B;;AAE3D,gBAAc,KAAK,cAAc;AAEjC,OAAK,MAAM,WAAW,eAAe;GACnC,MAAM,QAAQ,MAAM,MAAM,IAAI,QAAQ;AACtC,OAAI,OAAO;IACT,MAAM,OACJ,MAAM,gBAAgB,eAClB,MAAM,OACN,IAAI,aAAa,MAAM,KAAK,QAAQ,MAAM,KAAK,YAAY,MAAM,KAAK,aAAa,EAAE;IAC3F,MAAM,WAAW,IAAI,aAAa,KAAK,OAAO;AAC9C,SAAK,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,IAC/B,UAAS,KAAK,IAAM,KAAK;AAE3B,UAAM,MAAM,IAAI,SAAS;KAAE,MAAM;KAAU,OAAO,MAAM;KAAO,CAAC;;;;AAkBtE,KAAI,qBAAqB,iCACvB,MAAK,MAAM,QAAQ,MAAM,OAAO;EAC9B,MAAM,MAAM,KAAK,YAAY;AAC7B,MAAI,CAAC,IAAK;EACV,MAAM,QAAQ,MAAM,MAAM,IAAI,IAAI;AAClC,MAAI,CAAC,MAAO;EACZ,MAAM,OACJ,MAAM,gBAAgB,eAClB,MAAM,OACN,IAAI,aAAa,MAAM,KAAK,QAAQ,MAAM,KAAK,YAAY,MAAM,KAAK,aAAa,EAAE;AAC3F,OAAK,WAAW,QAAQ,KAAK;AAC7B,QAAM,MAAM,OAAO,IAAI;;AAO3B,MAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,MAAM,QAAQ,CACtD,KACE,KAAK,YAAY,cACjB,KAAK,kBACL,KAAK,mBAAmB,QACxB,CAAC,MAAM,IAAI,KAAK,EAChB;EACA,MAAM,aAAa,MAAM,MAAM,IAAI,KAAK,eAAe;AACvD,MAAI,WACF,OAAM,MAAM,IAAI,MAAM,WAAW;;AAMvC,KAAI,QAAQ;AACV,eAAa,IAAI,KAAK,4BAA4B;EAClD,MAAM,gBAAiB,aAAa,cAAyB;AAE7D,OAAK,MAAM,QAAQ,MAAM,OAAO;AAC9B,OAAI,KAAK,WAAW,aAAc;GAElC,MAAM,IAAI,KAAK,WAAW;GAC1B,MAAM,IAAI,KAAK,WAAW;GAE1B,MAAM,cAAc,KAAK,OAAO;GAChC,MAAM,mBAAmB,KAAK,OAAO;GACrC,MAAM,kBAAkB,KAAK,OAAO;GAKpC,MAAM,WADa,YAAY,MAAM,GAAG,GAAG,CACf,MAAM,GAAG,GAAG;GAExC,MAAM,cAAc,MAAM,MAAM,IAAI,GAAG,SAAS,UAAU;GAC1D,MAAM,aAAa,MAAM,MAAM,IAAI,GAAG,SAAS,SAAS;GACxD,MAAM,aAAa,MAAM,MAAM,IAAI,GAAG,SAAS,SAAS;AAGxD,OAAI,CAAC,eAAe,CAAC,cAAc,CAAC,WAAY;GA4BhD,MAAM,WAAW,WAAW;IAC1B,SAzBA,YAAY,gBAAgB,aACxB,YAAY,OACZ,IAAI,WACF,YAAY,KAAK,QACjB,YAAY,KAAK,YACjB,YAAY,KAAK,aAAa,EAC/B;IAoBL,QAlBA,WAAW,gBAAgB,eACvB,WAAW,OACX,IAAI,aACF,WAAW,KAAK,QAChB,WAAW,KAAK,YAChB,WAAW,KAAK,aAAa,EAC9B;IAaL,QAXA,WAAW,gBAAgB,aACvB,WAAW,OACX,IAAI,WACF,WAAW,KAAK,QAChB,WAAW,KAAK,YAChB,WAAW,KAAK,aAAa,EAC9B;IAML;IACA;IACA,WAAW;IACZ,CAAC;AAEF,SAAM,MAAM,IAAI,aAAa;IAAE,MAAM,SAAS;IAAQ,OAAO,CAAC,SAAS,OAAO,OAAO;IAAE,CAAC;AACxF,SAAM,MAAM,IAAI,kBAAkB;IAAE,MAAM,SAAS;IAAQ,OAAO,CAAC,SAAS,OAAO,OAAO;IAAE,CAAC;AAC7F,SAAM,MAAM,IAAI,iBAAiB;IAAE,MAAM,SAAS;IAAO,OAAO,CAAC,SAAS,MAAM,OAAO;IAAE,CAAC;AAG1F,SAAM,MAAM,OAAO,GAAG,SAAS,UAAU;AACzC,SAAM,MAAM,OAAO,GAAG,SAAS,SAAS;AACxC,SAAM,MAAM,OAAO,GAAG,SAAS,SAAS;;;AAK5C,KAAI,OAAO;AACT,eAAa,IAAI,KAAK,2BAA2B;EACjD,MAAM,QAAS,aAAa,cAAyB;AAErD,OAAK,MAAM,QAAQ,MAAM,OAAO;AAC9B,OAAI,KAAK,WAAW,gBAAgB,KAAK,WAAW,gBAAiB;GAGrE,IAAIC,GAAWC;AACf,OAAI,KAAK,WAAW,iBAAiB;AACnC,QAAI,KAAK,WAAW;AACpB,QAAI,KAAK,WAAW;UACf;AACL,QAAI,KAAK,WAAW;AACpB,QAAI,KAAK,WAAW;;GAGtB,MAAM,cAAc,KAAK,OAAO;GAChC,MAAM,mBAAmB,KAAK,OAAO;GACrC,MAAM,kBAAkB,KAAK,OAAO;GAKpC,MAAM,aAAa,YAAY,MAAM,GAAG,GAAG;GAC3C,MAAM,WAAW,WAAW,MAAM,GAAG,GAAG;GAExC,MAAM,YAAY,MAAM,MAAM,IAAI,WAAW;GAC7C,MAAM,YAAY,MAAM,MAAM,IAAI,GAAG,SAAS,SAAS;GACvD,MAAM,YAAY,MAAM,MAAM,IAAI,GAAG,SAAS,SAAS;AAEvD,OAAI,CAAC,aAAa,CAAC,aAAa,CAAC,UAAW;GA2B5C,MAAM,WAAW,UAAU;IACzB,QAzBA,UAAU,gBAAgB,cACtB,UAAU,OACV,IAAI,YACF,UAAU,KAAK,QACf,UAAU,KAAK,YACf,UAAU,KAAK,aAAa,EAC7B;IAoBL,QAlBA,UAAU,gBAAgB,eACtB,UAAU,OACV,IAAI,aACF,UAAU,KAAK,QACf,UAAU,KAAK,YACf,UAAU,KAAK,aAAa,EAC7B;IAaL,QAXA,UAAU,gBAAgB,eACtB,UAAU,OACV,IAAI,aACF,UAAU,KAAK,QACf,UAAU,KAAK,YACf,UAAU,KAAK,aAAa,EAC7B;IAML;IACA;IACA,WAAW;IACZ,CAAC;AAEF,SAAM,MAAM,IAAI,aAAa;IAAE,MAAM,SAAS;IAAQ,OAAO,CAAC,SAAS,OAAO,OAAO;IAAE,CAAC;AACxF,SAAM,MAAM,IAAI,kBAAkB;IAAE,MAAM,SAAS;IAAQ,OAAO,CAAC,SAAS,OAAO,OAAO;IAAE,CAAC;AAC7F,SAAM,MAAM,IAAI,iBAAiB;IAAE,MAAM,SAAS;IAAO,OAAO,CAAC,SAAS,MAAM,OAAO;IAAE,CAAC;AAG1F,SAAM,MAAM,OAAO,WAAW;AAC9B,SAAM,MAAM,OAAO,GAAG,SAAS,SAAS;AACxC,SAAM,MAAM,OAAO,GAAG,SAAS,SAAS;;;AAK5C,KAAI,kBAAkB,MAAM;AAC1B,eAAa,IAAI,KAAK,gCAAgC;AACtD,OAAK,MAAM,QAAQ,MAAM,OAAO;AAC9B,OAAI,KAAK,WAAW,gBAAgB,KAAK,WAAW,gBAAiB;GAErE,MAAM,UAAW,KAAK,WAAW,cAAyB;GAE1D,MAAM,cAAc,KAAK,OAAO;GAChC,MAAM,mBAAmB,KAAK,OAAO;GACrC,MAAM,kBAAkB,KAAK,OAAO;AAGpC,OAAI,MAAM,IAAI,YAAY,CAAE;GAG5B,MAAM,aAAa,YAAY,MAAM,GAAG,GAAG;GAC3C,IAAI,eAAe,MAAM,MAAM,IAAI,WAAW;AAG9C,OAAI,CAAC,gBAAgB,eAAe,iBAClC,gBAAe,MAAM,MAAM,IAAI,sBAAsB;AAGvD,OAAI,cAAc;IAUhB,MAAM,EAAE,QAAQ,QAAQ,UAAU,aARhC,aAAa,gBAAgB,eACzB,aAAa,OACb,IAAI,aACF,aAAa,KAAK,QAClB,aAAa,KAAK,YAClB,aAAa,KAAK,aAAa,EAChC,EAEiD,QAAQ;AAEhE,UAAM,MAAM,IAAI,aAAa;KAAE,MAAM;KAAQ,OAAO,CAAC,OAAO,OAAO;KAAE,CAAC;AACtE,UAAM,MAAM,IAAI,kBAAkB;KAAE,MAAM;KAAQ,OAAO,CAAC,OAAO,OAAO;KAAE,CAAC;AAC3E,UAAM,MAAM,IAAI,iBAAiB;KAAE,MAAM;KAAO,OAAO,CAAC,MAAM,OAAO;KAAE,CAAC;AAGxE,UAAM,MAAM,OAAO,WAAW;;;;AAcpC,KAAI,OAAO;EACT,MAAM,KAAK,MAAM,MAAM,IAAI,eAAe,oBAAoB;EAC9D,MAAM,IAAI,IAAI;AAEd,MAAI,MAAM,KAAK,EAAE,WAAW,KAAK,EAAE,OAAO,GAAG;GAC3C,MAAM,CAAC,MAAM,GAAG,IAAI,IAAI,KAAK;GAC7B,MAAM,MACJ,GAAG,gBAAgB,eACf,GAAG,OACH,IAAI,aAAa,GAAG,KAAK,QAAQ,GAAG,KAAK,YAAY,GAAG,KAAK,aAAa,EAAE;GAClF,MAAM,MAAM,IAAI,aAAa,IAAI,OAAO;AAGxC,QAAK,IAAI,IAAI,GAAG,IAAI,MAAM,IACxB,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IACrB,MAAK,IAAI,IAAI,GAAG,IAAI,IAAI,IACtB,MAAK,IAAI,IAAI,GAAG,IAAI,IAAI,IACtB,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;IAC1B,MAAM,QAAQ,IAAI,IAAI,KAAK,KAAK,KAAK,KAAK,KAAK,IAAI;IACnD,MAAM,QAAQ,IAAI,IAAI,KAAK,IAAI,KAAK,KAAK,KAAK,KAAK;AACnD,QAAI,MAAM,IAAI;;AAOxB,SAAM,MAAM,IAAI,eAAe,qBAAqB;IAClD,MAAM;IACN,OAAO,CAAC,MAAM,IAAI,IAAI,KAAK,GAAG;IAC/B,CAAC;;;CAYN,IAAIC;AAEJ,KADiB,OAAO,iBAAiB,CAAC,WAAW,SAAS,EAChD;EACZ,MAAM,UAAU,eAAe,iBAAiB,MAAM,GAAG,GAAG;EAC5D,MAAM,UAAW,UAAU,eAA2C;EACtE,MAAM,SAAU,QAAQ,+BAA0C;EAElE,MAAM,WADY,QAAQ,oBACG;EAC7B,MAAM,WACH,QAAQ,8BAA0C,QAAQ;EAE7D,MAAM,cAAc,MAAM,MAAM,IAAI,GAAG,QAAQ,SAAS;EACxD,MAAM,cAAc,MAAM,MAAM,IAAI,GAAG,QAAQ,SAAS;EACxD,MAAM,cAAc,MAAM,MAAM,IAAI,GAAG,QAAQ,SAAS;AAExD,MAAI,SAAS,eAAe,eAAe,aAAa;GACtD,MAAM,QAAS,aAAa,cAAyB;GAyBrD,MAAM,WAAW,UAAU;IACzB,QAxBA,YAAY,gBAAgB,cACxB,YAAY,OACZ,IAAI,YACF,YAAY,KAAK,QACjB,YAAY,KAAK,YACjB,YAAY,KAAK,aAAa,EAC/B;IAmBL,QAjBA,YAAY,gBAAgB,eACxB,YAAY,OACZ,IAAI,aACF,YAAY,KAAK,QACjB,YAAY,KAAK,YACjB,YAAY,KAAK,aAAa,EAC/B;IAYL,QAVA,YAAY,gBAAgB,eACxB,YAAY,OACZ,IAAI,aACF,YAAY,KAAK,QACjB,YAAY,KAAK,YACjB,YAAY,KAAK,aAAa,EAC/B;IAKL,GAAG;IACH,GAAG;IACH,WAAW;IACZ,CAAC;AACF,eAAY,MAAM,eAChB,SAAS,QACT,SAAS,QACT,SAAS,OACT,UACA,OACA,WACD;AAED,SAAM,MAAM,OAAO,GAAG,QAAQ,SAAS;AACvC,SAAM,MAAM,OAAO,GAAG,QAAQ,SAAS;AACvC,SAAM,MAAM,OAAO,GAAG,QAAQ,SAAS;AACvC,WAAQ,IACN,kCAAkC,aAAa,mBAAmB,eAAe,KAC3E,SAAS,IAAI,SAAS,aACnB,SAAS,OAAO,aAAa,SAAS,OAAO,aAAa,SAAS,MAAM,cAAc,KAAK,QAAQ,EAAE,CAAC,gBACjH;aACQ,aAAa,gBAAgB,cAAc;GAIpD,MAAM,EAAE,QAAQ,QAAQ,UAAU,aAAa,YAAY,MAAM,mBAAmB;AACpF,eAAY,MAAM,eAChB,QACA,QACA,OACA,UACA,oBACA,WACD;AACD,SAAM,MAAM,OAAO,GAAG,QAAQ,SAAS;QAEvC,SAAQ,KACN,0IAED;;AAOL,MAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,MAAM,QAAQ,CACtD,KAAI,KAAK,YAAY,cAAc,KAAK,cAAc,UAAa,CAAC,MAAM,IAAI,KAAK,EAAE;EACnF,MAAM,SAAS,KAAK,MAAM,QACvB,KAAa,QAAQ,OAAO,OAAO,QAAQ,WAAW,MAAM,IAC7D,EACD;EACD,MAAM,OAAO,IAAI,aAAa,OAAO,CAAC,KAAK,KAAK,UAAU;AAC1D,QAAM,MAAM,IAAI,MAAM;GACpB;GACA,OAAO,KAAK,MAAM,KAAK,MAAO,OAAO,MAAM,WAAW,IAAI,EAAG;GAC9D,CAAC;;CAKN,MAAMC,iBAA2B,EAAE;AACnC,MAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,MAAM,QAAQ,CACtD,KAAI,KAAK,YAAY,cAAc,CAAC,MAAM,IAAI,KAAK,CACjD,gBAAe,KAAK,KAAK;AAI7B,KAAI,eAAe,SAAS,EAC1B,SAAQ,KACN,oBAAoB,eAAe,OAAO,mBAC1C,eAAe,MAAM,GAAG,GAAG,CAC5B;AAGH,cAAa,KAAK,KAAK,gBAAgB;AAEvC,QAAO;EAAE;EAAO;EAAW,SAAS;EAAO;EAAW;EAAW;;AAuBnE,eAAsB,cAAc,SAMP;CAC3B,MAAM,WAAW,QAAQ,YAAY;CACrC,MAAM,UAAU,iBAAiB,QAAQ,MAAM,SAAS;AACxD,OAAM,QAAQ;AAEd,SAAQ,aAAa,GAAG,KAAK,+BAA+B;CAC5D,MAAM,CAAC,WAAW,iBAAiB,MAAM,QAAQ,IAAI,CACnD,UAAU,SAAS,eAAe,QAAQ,SAAS,QAAQ,SAAS,EACpE,UAAU,SAAS,kBAAkB,QAAQ,SAAS,QAAQ,SAAS,CACxE,CAAC;CACF,MAAM,kBAAkB,MAAM,UAC5B,SACA,yBACA,QAAQ,SACR,QAAQ,SACT,CAAC,YAAY,KAAK;CACnB,MAAM,YAAY,UAAU,SAAS,eAAe,gBAAgB;AAEpE,SAAQ,aAAa,IAAI,KAAK,uBAAuB;CACrD,MAAM,MAAM,MAAM,YAChB,SACA,qBACA,QAAQ,UACP,QAAQ,UAAU,QAAQ,aAAa,KAAM,UAAU,SAAS,KAAM,IAAI,KAAK,UAAU,EAC1F,QAAQ,SACT;CAED,MAAM,OAAO,uBAAuB,IAAI;CACxC,MAAM,0BAAU,IAAI,KAAyD;AAC7E,MAAK,MAAM,SAAS,KAAK,SAAS;EAEhC,MAAM,YAAY,MAAM,KAAK,WAAW,SAAS,GAAG,MAAM,KAAK,MAAM,EAAE,GAAG,MAAM;EAChF,IAAIP,OAAwB,cAAc,KAAK,MAAM,MAAM;AAC3D,MAAI,MAAM,UAAU,OAClB,QAAO,UAAU,IAAI,WAAW,KAAK,QAAQ,KAAK,YAAY,KAAK,WAAW,CAAC;WACtE,MAAM,UAAU,MACzB,QAAO,SAAS,KAAK;AAEvB,UAAQ,IAAI,WAAW;GAAE;GAAM,OAAO,MAAM;GAAO,CAAC;;AAGtD,SAAQ,aAAa,KAAK,KAAK,oBAAoB;AACnD,QAAO;EAAE;EAAS;EAAW;EAAW;;;AAoB1C,MAAa,sBAAsB;;AAcnC,MAAM,mBAAmB;;AAGzB,SAAS,MAAM,OAAwB,KAAkB,MAAqC;CAC5F,IAAIA,OAAwB,cAAc,KAAK,MAAM,MAAM;AAC3D,KAAI,MAAM,UAAU,OAClB,QAAO,UAAU,IAAI,WAAW,KAAK,QAAQ,KAAK,YAAY,KAAK,WAAW,CAAC;UACtE,MAAM,UAAU,MACzB,QAAO,SAAS,KAAK;AAEvB,QAAO,gBAAgB,eACnB,OACA,IAAI,aAAa,KAAK,QAAQ,KAAK,YAAY,KAAK,aAAa,EAAE;;;AAIzE,SAAS,sBACP,UACyD;CACzD,MAAM,sBAAM,IAAI,KAAyD;CACzE,MAAM,EAAE,OAAO,WAAW,oBAAoB;AAC9C,MAAK,MAAM,KAAK,OAAO;EACrB,MAAM,IAAI,SAAS,IAAI,EAAE,KAAK;EAC9B,MAAM,IAAI,SAAS,IAAI,EAAE,KAAK;AAC9B,MAAI,EAAE,KAAK,GACT,OAAM,IAAI,MACR,0CAA0C,EAAE,SAAS,IAAI,EAAE,KAAK,KAAK,EAAE,KAAK,uEAE7E;AAEH,MAAI,IAAI,EAAE,UAAU;GAClB,MAAM,wBAAwB,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM;GACtD,OAAO,EAAE;GACV,CAAC;;AAEJ,MAAK,MAAM,MAAM,QAAQ;EACvB,MAAM,IAAI,SAAS,IAAI,GAAG,OAAO;AACjC,MAAI,CAAC,EACH,OAAM,IAAI,MAAM,6BAA6B,GAAG,OAAO,QAAQ,GAAG,SAAS,IAAI;AAEjF,MAAI,IAAI,GAAG,UAAU;GAAE,MAAM,EAAE;GAAM,OAAO,EAAE;GAAO,CAAC;;AAExD,QAAO;;AAGT,eAAsB,YAAY,SASP;CACzB,MAAM,WAAW,QAAQ,YAAY;CACrC,MAAM,OAAO,QAAQ,QAAQ;CAC7B,MAAM,YAAY,QAAQ,aAAa;CACvC,MAAM,UAAU,iBAAiB,MAAM,SAAS;CAChD,MAAM,WAAW,iBAAiB,WAAW,OAAO;AACpD,OAAM,QAAQ;AAEd,SAAQ,aAAa,GAAG,KAAK,+BAA+B;CAC5D,MAAM,CAAC,WAAW,iBAAiB,MAAM,QAAQ,IAAI,CACnD,UAAU,SAAS,eAAe,QAAQ,SAAS,QAAQ,SAAS,EACpE,UAAU,SAAS,kBAAkB,QAAQ,SAAS,QAAQ,SAAS,CACxE,CAAC;CACF,MAAM,kBAAkB,MAAM,UAC5B,SACA,yBACA,QAAQ,SACR,QAAQ,SACT,CAAC,YAAY,KAAK;CACnB,MAAM,YAAY,UAAU,SAAS,eAAe,gBAAgB;AAGpE,SAAQ,aAAa,GAAG,KAAK,iCAAiC;CAC9D,MAAM,cAAc,MAAM,YACxB,SACA,qBACA,QAAQ,UACP,QAAQ,UACP,QAAQ,aAAa,IAAK,UAAU,SAAS,KAAM,IAAI,KAAK,mBAAmB,EACjF,QAAQ,SACT;CACD,MAAM,eAAe,uBAAuB,YAAY;CACxD,MAAM,kCAAkB,IAAI,KAAyD;AACrF,MAAK,MAAM,SAAS,aAAa,SAAS;EACxC,MAAM,OAAO,MAAM,KAAK,WAAW,SAAS,GAAG,MAAM,KAAK,MAAM,EAAE,GAAG,MAAM,MAAM,QAC/E,kBACA,qBACD;AACD,kBAAgB,IAAI,KAAK;GACvB,MAAM,MAAM,OAAO,aAAa,aAAa;GAC7C,OAAO,MAAM;GACd,CAAC;;AAIJ,SAAQ,aAAa,IAAI,KAAK,iCAAiC;CAI/D,MAAM,gBAAgB,QAAQ,WAAW,GAAG,QAAQ,SAAS,cAAc;CAC3E,MAAM,WAAW,MAAM,YACrB,UACA,qBACA,QAAQ,UACP,QAAQ,UACP,QAAQ,aAAa,KAAM,UAAU,SAAS,KAAM,IAAI,KAAK,gBAAgB,EAC/E,cACD;CACD,MAAM,YAAY,uBAAuB,SAAS;CAClD,MAAM,2BAAW,IAAI,KAAsD;AAC3E,MAAK,MAAM,SAAS,UAAU,QAC5B,UAAS,IAAI,MAAM,MAAM;EAAE,MAAM,MAAM,OAAO,UAAU,UAAU;EAAE,OAAO,MAAM;EAAO,CAAC;CAE3F,MAAM,eAAe,sBAAsB,SAAS;AAEpD,SAAQ,aAAa,KAAK,KAAK,qBAAqB;AACpD,QAAO;EAAE;EAAiB;EAAc;EAAW;EAAW;;;;;ACrrEhE,MAAM,gBAAgB;;;;;;;AAQtB,MAAM,yBAAyB;AAqB/B,IAAa,2BAAb,MAAsC;CACpC,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CAER,AAAQ,gCAAgB,IAAI,KAAwB;CACpD,AAAQ,oCAAoB,IAAI,KAAwB;CACxD,AAAQ,aAAkC,EAAE;CAE5C,YAAY,KAAiB,OAAmB,WAAmB;AACjE,OAAK,MAAM;AACX,OAAK,QAAQ;AACb,OAAK,YAAY;AACjB,OAAK,SAAS,MAAM,OAAO;AAC3B,OAAK,2BAA2B;;;;;;;CAQlC,cAAc,SAAwE;AACpF,OAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,KAAK,MAAM,QAAQ,EAAE;AAC7D,OAAI,KAAK,YAAY,WAAY;GACjC,MAAM,IAAI,QAAQ,IAAI,KAAK;AAC3B,OAAI,GAAG;IACL,MAAM,SAAS,oBAAoB,KAAK,KAAK,QAAQ,QAAQ,EAAE,KAAK,YAAY,EAAE,KAAK;AACvF,SAAK,cAAc,IAAI,MAAM,OAAO;cAC3B,CAAC,KAAK,gBAAgB;IAI/B,MAAM,QAAS,KAAK,MAAmB,QAAQ,GAAG,MAAM,IAAK,GAAc,EAAE;IAC7E,MAAM,QAAQ,IAAI,aAAa,MAAM;IACrC,MAAM,SAAS,oBAAoB,KAAK,KAAK,QAAQ,QAAQ,MAAM,YAAY,MAAM;AACrF,SAAK,cAAc,IAAI,MAAM,OAAO;SAEpC,OAAM,IAAI,MAAM,6CAA6C,KAAK,GAAG;;;CAK3E,iBAAuB;EACrB,MAAM,SAAS,KAAK,eAAe;AACnC,OAAK,MAAM,UAAU,KAAK,MAAM,gBAAgB;GAC9C,MAAM,OAAO,KAAK,MAAM,MAAM,MAAM,MAAM,EAAE,OAAO,OAAO;GAC1D,IAAI,OAAO,gBAAgB,KAAK;AAChC,OAAI,CAAC,KAAM,OAAM,IAAI,MAAM,+CAA+C,KAAK,OAAO,GAAG;AAEzF,OAAI,KAAK,WAAW,eAAe,KAAK,OAAO,WAAW,EACxD,QAAO;YACE,KAAK,WAAW,UAAU,KAAK,WAAW,gBAAgB,KACnE,QAAO;GAET,MAAM,WAAW,oBACf,KAAK,KACL,WAAW,UACX,KAAK,YACL,KAAK,WACN;GACD,MAAM,aAAa,KAAK,YAAY,MAAM,QAAQ,EAAE,QAAQ,GAAG,CAAC;GAChE,MAAM,gBAAgB,oBAAoB,KAAK,KAAK,YAAY,UAAU,WAAW;GACrF,MAAM,gBAAgB,KAAK,cAAc,MAAM,MAAM,cAAc;GACnE,MAAM,YAAY,gBAAgB,KAAK,KAAK,UAAU,eAAe,OAAO,SAAS;AACrF,QAAK,WAAW,KAAK;IAAE;IAAM;IAAM;IAAU;IAAW;IAAe,CAAC;;;;CAK5E,MAAM,OAAO,KAA2C;EACtD,MAAM,IAAI,KAAK,IAAI,OAAO;AAC1B,IAAE,YAAY,KAAK,kBAAkB,IAAI,MAAM,EAAG,GAAG,IAAoB;EAEzE,MAAM,SAAS,KAAK,eAAe;EACnC,MAAM,QAAQ,EAAE,QAAQ,GAAG;EAC3B,MAAMQ,QAAyC,EAAE;AACjD,OAAK,MAAM,KAAK,KAAK,YAAY;GAC/B,MAAM,SAAS,EAAE,KAAK,YAAY,EAAE,MAAM,QAAQ,MAAM;AACxD,KAAE,YAAY,EAAE,eAAe,GAAG,OAAO;AACzC,SAAM,KAAK,EAAE,KAAK,gBAAgB,EAAE,MAAM,QAAQ,MAAM,CAAC;;AAG3D,MAAI,KAAK,IAAI,eACX,MAAK,IAAI,IAAI,GAAG,IAAI,KAAK,WAAW,QAAQ,KAAK;GAC/C,MAAM,IAAI,KAAK,WAAW;AAC1B,OAAI,EAAE,KAAK,WAAW,aAAa;IAEjC,MAAM,CAAC,OAAO,OAAO,SAAS,MAAM;AACpC,SAAK,IAAI,QAAQ,GAAG,QAAQ,OAAO,SAAS,wBAAwB;KAClE,MAAM,OAAO,KAAK,IAAI,wBAAwB,QAAQ,MAAM;KAC5D,MAAM,SAAS,EAAE,KAAK,YAAY,EAAE,MAAM,QAAQ;MAAE,QAAQ;MAAG,SAAS;MAAO,CAAC;AAChF,OAAE,YAAY,EAAE,eAAe,GAAG,OAAO;KACzC,MAAMC,QAAM,KAAK,IAAI,OAAO,sBAAsB;KAClD,MAAMC,MAAID,MAAI,kBAAkB;AAChC,SAAE,YAAY,EAAE,SAAS;AACzB,SAAE,aAAa,GAAG,EAAE,UAAU;AAC9B,SAAE,mBAAmB,MAAM,OAAO,MAAM;AACxC,SAAE,KAAK;AACP,OAAE,OAAO,CAACA,MAAI,QAAQ,CAAC,CAAC;AACxB,WAAM,EAAE,qBAAqB;;AAE/B;;GAEF,MAAM,MAAM,KAAK,IAAI,OAAO,sBAAsB;GAClD,MAAM,IAAI,IAAI,kBAAkB;AAChC,KAAE,YAAY,EAAE,SAAS;AACzB,KAAE,aAAa,GAAG,EAAE,UAAU;AAC9B,KAAE,mBAAmB,GAAG,MAAM,GAAG;AACjC,KAAE,KAAK;AACP,KAAE,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;AACxB,SAAM,EAAE,qBAAqB;;OAE1B;GACL,MAAM,MAAM,KAAK,IAAI,OAAO,qBAAqB,EAAE,OAAO,oBAAoB,CAAC;GAC/E,MAAM,OAAO,IAAI,iBAAiB,EAAE,OAAO,sBAAsB,CAAC;AAClE,QAAK,IAAI,IAAI,GAAG,IAAI,KAAK,WAAW,QAAQ,KAAK;AAC/C,SAAK,YAAY,KAAK,WAAW,GAAG,SAAS;AAC7C,SAAK,aAAa,GAAG,KAAK,WAAW,GAAG,UAAU;AAClD,SAAK,mBAAmB,GAAG,MAAM,GAAG;;AAEtC,QAAK,KAAK;AACV,KAAE,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;;EAG1B,MAAM,OAAO,KAAK,eAAe;EACjC,MAAM,aAAa,MAAM,KAAK,SAAS,eAAe,OAAO,KAAK,OAAO;EACzE,MAAME,OAAuB,EAAE;EAC/B,MAAMC,OAAuB,EAAE;AAC/B,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,WAAW,KAAK;AACvC,QAAK,KAAK,MAAM,KAAK,SAAS,cAAc,KAAK,OAAO,KAAK,OAAO,CAAC;AACrE,QAAK,KAAK,MAAM,KAAK,SAAS,cAAc,KAAK,OAAO,KAAK,OAAO,CAAC;;AAEvE,SAAO;GAAE;GAAY;GAAM;GAAM;GAAM;;;CAIzC,MAAM,eAAe,MAAc,cAAc,MAA6B;AAC5E,SAAO,KAAK,SAAS,MAAM,YAAY;;CAGzC,UAAgB;AACd,iBAAe,CAAC,GAAG,KAAK,cAAc,QAAQ,CAAC,CAAC;AAChD,iBAAe,CAAC,GAAG,KAAK,kBAAkB,QAAQ,CAAC,CAAC;AACpD,OAAK,MAAM,KAAK,KAAK,WAAY,GAAE,cAAc,SAAS;AAC1D,OAAK,cAAc,OAAO;AAC1B,OAAK,kBAAkB,OAAO;AAC9B,OAAK,aAAa,EAAE;;;CAMtB,AAAQ,gBAAwB;AAC9B,SAAO,KAAK,MAAM,QAAQ,YAAY,MAAM;;CAG9C,AAAQ,gBAA0C;EAGhD,MAAMC,WAAqC,EAAE;AAC7C,OAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,KAAK,MAAM,QAAQ,CAC3D,UAAS,QAAQ,KAAK,MAAM,KAAK,QAAQ,IAAc;AAEzD,SAAO;;CAGT,AAAQ,4BAAkC;AACxC,OAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,KAAK,MAAM,QAAQ,EAAE;AAC7D,OAAI,KAAK,YAAY,aAAc;GACnC,MAAM,QACH,KAAK,MAAmB,QAAQ,GAAG,MAAM,IAAK,GAAc,EAAE,GAAG,YAAY,KAAK;AACrF,QAAK,kBAAkB,IAAI,MAAM,oBAAoB,KAAK,KAAK,QAAQ,QAAQ,MAAM,CAAC;;;CAI1F,MAAc,SAAS,MAAc,UAAyC;EAC5E,MAAM,SAAS,KAAK,kBAAkB,IAAI,KAAK;AAC/C,MAAI,CAAC,OAAQ,OAAM,IAAI,MAAM,wCAAwC,KAAK,GAAG;EAC7E,MAAM,UAAU,KAAK,IAAI,WAAW,GAAG,OAAO,KAAK;EACnD,MAAM,WAAW,KAAK,IAAI,OAAO,aAAa;GAAE,MAAM;GAAS,OAAO;GAAiB,CAAC;EACxF,MAAM,MAAM,KAAK,IAAI,OAAO,sBAAsB;AAClD,MAAI,mBAAmB,QAAQ,GAAG,UAAU,GAAG,QAAQ;AACvD,OAAK,IAAI,OAAO,MAAM,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;AAC5C,QAAM,SAAS,SAAS,eAAe,GAAG,QAAQ;EAClD,MAAM,OAAO,IAAI,aAAa,SAAS,eAAe,GAAG,QAAQ,CAAC,MAAM,EAAE,CAAC;AAC3E,WAAS,OAAO;AAChB,WAAS,SAAS;AAClB,SAAO;;CAGT,AAAQ,cACN,MACA,MACA,eAC8B;EAC9B,MAAMC,UAAwC,EAAE;EAChD,MAAM,cAAc,CAAC,GAAG,KAAK,QAAQ,GAAG,KAAK,QAAQ;EACrD,IAAI,YAAY;AAChB,OAAK,MAAM,WAAW,KAAK,SACzB,KAAI,QAAQ,SAAS,UACnB,SAAQ,KAAK,EAAE,QAAQ,eAAe,CAAC;OAClC;GACL,MAAM,aAAa,YAAY;GAC/B,MAAM,SAAS,KAAK,cAAc,IAAI,WAAW,IAAI,KAAK,kBAAkB,IAAI,WAAW;AAC3F,OAAI,CAAC,OACH,OAAM,IAAI,MACR,4CAA4C,WAAW,UAAU,KAAK,KACvE;AAEH,WAAQ,KAAK,EAAE,QAAQ,CAAC;;AAG5B,SAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;AClPX,MAAM,wBAAwB;;AAE9B,MAAM,yBAAyB;;AA8C/B,MAAM,gBAAgB;AAEtB,SAAS,WAAW,KAA2B;AAC7C,KAAI,IAAI,WAAW,EACjB,QAAO;CAET,IAAI,MAAM;AACV,MAAK,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,IAC9B,QAAO,IAAI,KAAK,IAAI;AAEtB,QAAO,KAAK,KAAK,MAAM,IAAI,OAAO;;AAGpC,IAAa,eAAb,MAAa,aAAa;CACxB,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ,aAAa;;CAGrB,AAAS,eAAe;CAExB,AAAQ,YACN,KACA,SACA,WACA,WACA;AACA,OAAK,MAAM;AACX,OAAK,UAAU;AACf,OAAK,YAAY;AACjB,OAAK,YAAY;AACjB,OAAK,aAAc,UAAU,gBAA2B;AACxD,OAAK,aAAc,UAAU,gBAA2B;AAExD,OAAK,sBAAuB,UAAU,0BAAqC,KAAK;;;CAIlF,aAAa,OAAO,UAA+B,EAAE,EAAyB;EAC5E,MAAM,MAAM,MAAM,SAAS;EAE3B,MAAM,EAAE,SAAS,WAAW,cAAc,MAAM,cAAc;GAC5D,MAFW,QAAQ,QAAQ;GAG3B,UAAU,QAAQ;GAClB,SAAS,QAAQ;GACjB,UAAU,QAAQ;GAClB,YAAY,QAAQ;GACrB,CAAC;AACF,SAAO,IAAI,aAAa,KAAK,SAAS,WAAW,UAAU;;;;;;;CAQ7D,MAAM,WAAW,KAAmB,OAA0B,EAAE,EAA6B;AAC3F,MAAI,KAAK,WAAY,OAAM,IAAI,MAAM,mCAAmC;AACxE,MAAI,IAAI,SAAS,IACf,OAAM,IAAI,MAAM,kBAAkB,IAAI,OAAO,2CAA2C;EAE1F,MAAM,IAAI,qBAAqB,KAAK,UAAU;EAC9C,MAAM,OAAO,uBAAuB,IAAI,OAAO;AAC/C,MAAI,OAAO,EACT,OAAM,IAAI,MAAM,6BAA6B,KAAK,kBAAkB;EAGtE,MAAM,eAAe,IAAI,SAAS;EAClC,MAAM,YAAY,WAAW,IAAI;EACjC,MAAM,SAAS,KAAK,UAAU;EAC9B,MAAM,mBAAmB,KAAK,oBAAoB;AAKlD,MAAI,YAAY,UAAU,eAAe,iBACvC,QAAO;GACL,MAAM;GACN,QAAQ,EAAE;GACV,eAAe;GACf;GACA;GACA,UAAU;GACX;EAIH,MAAM,eAAe,8BAA8B,KAAK,WAAW,IAAI,OAAO;EAC9E,MAAM,UAAU,IAAI,yBAAyB,KAAK,KAAK,cAAc,EAAE,WAAW;EAClF,IAAIC;EACJ,IAAIC;EACJ,IAAIC;AACJ,MAAI;AACF,WAAQ,cAAc,KAAK,QAAQ;AACnC,WAAQ,gBAAgB;GACxB,MAAM,MAAM,MAAM,QAAQ,OAAO,IAAI;AACrC,UAAO,IAAI;AACX,UAAO,IAAI;AACX,YAAS,IAAI;YACL;AACR,WAAQ,SAAS;;EAInB,MAAM,eAAe,8BAA8B,KAAK,WAAW,OAAO;EAG1E,MAAM,YAAY,KAAK,IAAI,KAAK,gBAAgB,wBAAwB,EAAE,eAAe;EACzF,MAAM,UAAU,IAAI,SAAS,KAAK,KAAK,cAAc;GAAE;GAAW,QAAQ;GAAO,CAAC;EAClF,MAAMC,SAAmB,EAAE;AAC3B,MAAI;AACF,WAAQ,iBAAiB,mBAAmB,cAAc,KAAK,QAAQ,CAAC;AACxE,WAAQ,gBAAgB;AAExB,QAAK,IAAI,IAAI,GAAG,IAAI,EAAE,YAAY,KAAK;AACrC,YAAQ,WAAW,cAAc,KAAK,KAAK,GAAG;AAC9C,YAAQ,WAAW,cAAc,KAAK,KAAK,GAAG;;AAEhD,WAAQ,OAAO;GAGf,IAAI,YAAY,KAAK;GACrB,MAAM,QAAQ;AACd,QAAK,IAAI,OAAO,GAAG,OAAO,OAAO,QAAQ;IACvC,MAAM,MAAM,MAAM,QAAQ,cAAc,IAAI,YAAY,CAAC,UAAU,CAAC,CAAC;AACrE,WAAO,KAAK,IAAI;AAChB,QAAI,QAAQ,KAAK,WAAY;AAC7B,gBAAY;;YAEN;AACR,WAAQ,SAAS;;EAGnB,MAAM,OAAO,KAAK,UAAU,OAAO,QAAgC,KAAK,CAAC,MAAM;EAC/E,MAAM,WAAW,KAAK,WAAW,KAAK,cAAc,KAAK,KAAK;AAC9D,SAAO;GACL;GACA;GACA,eAAe;GACf;GACA;GACA;GACD;;CAGH,UAAgB;AACd,OAAK,aAAa;AAClB,OAAK,QAAQ,OAAO;;;;;;;;;;;;;AAcxB,SAAS,mBACP,OACA,SACyD;CACzD,MAAM,sBAAM,IAAI,KAAyD;AACzE,MAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,MAAM,QAAQ,EAAE;AACxD,MAAI,KAAK,YAAY,WAAY;EACjC,MAAM,IAAI,QAAQ,IAAI,KAAK;AAC3B,MAAI,EAAG,KAAI,IAAI,MAAM,EAAE;;AAEzB,QAAO"}