@tryhamster/gerbil 1.11.4 → 1.13.0

package/dist/gpu-CU_Mldk0.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"gpu-CU_Mldk0.mjs","names":["tensors: Record<string, TensorDesc>","nodes: OpNode[]","executionOrder: string[]","tensors: Record<string, TensorDesc>","nodes: OpNode[]","executionOrder: string[]","DEFAULT_MAX_NEW_TOKENS","selectGraphWeights","nucleusSample","pick","softmaxInPlace","prompt: number[]","seq: number[]","audioTokens: number[]","_heapIndices: Uint32Array | null","_heapValues: Float32Array | null","argmax","penaltySet: Set<number> | null","cumulative","adapterReg: { id: number; deltas: LoRADelta[] } | null","messages: ChatMessage[]","req: Request","rows: Array<{ slot: number; req: Request }>","out: Record<string, ServeResult>","out: number[]","byIndex: number[]","config: ModelArchConfig","capabilities: ModelCapabilities","tensors: Record<string, TensorDesc>","nodes: OpNode[]","executionOrder: string[]","kvCacheDtype: DType","pendingNorm1Out: string | null","norm1Out: string","DEFAULT_MAX_FRAMES","selectGraphWeights","pick","frames: Uint32Array[]","graphOutputs: string[]","c1: number[]","c2: number[]","tensors: Record<string, TensorDesc>","nodes: OpNode[]","executionOrder: string[]","kvCacheDtype: DType","config: ModelArchConfig","out: AlignedWord[]","runs: SpeechRun[]","blocks: string[]","resp","selectGraphWeights","argmax","pick","siftDown","words: OuteSpeakerWord[]","seq: number[]","generated: number[]","scores","MAP_MODE_READ","allowed: Int32Array | null","pick","sizes: Array<[number, number, number]>","encK: Float32Array[]","encV: Float32Array[]","resolved: Record<string, number[]>","entries: Array<{ buffer: GPUBuffer }>","embedTokens: Float32Array[]","data: ArrayBufferView","sizes: Array<[number, number, number]>","batchEnc: GPUCommandEncoder | null","batchPass: GPUComputePassEncoder | null","rows","enc","p","resolved: Record<string, number[]>","entries: Array<{ buffer: GPUBuffer }>","GEMMA4_IMAGE_PROCESSOR: ImageProcessorConfig","QWEN3_5_IMAGE_PROCESSOR: ImageProcessorConfig","DEFAULT_PROBE: WebkitGroupProbe","next: WebkitGroupProbe","kvMode: KvMode","visionBundle: {\n      executor: VisionExecutor;\n      config: Record<string, unknown>;\n      posEmbedTable: Float32Array;\n    } | null","VisionExecutorImpl","maxSeqLen: number","webkitGroupSize: number | undefined","host","laneAdapters: Array<{ id: number; deltas: LoRADelta[] } | null>","messages: ChatMessage[]","inputIds: number[]","generatedIds: number[]","finishReason: GenerateResult[\"finishReason\"]","nextToken: number","steps: AgentStep[]","answer: AgentStep","callStep: AgentStep","resultText: string","resultStep: AgentStep","value: unknown","speakOpts: ParlerSpeakOptions","g4patches: Float32Array","g4gridHW: [number, number]","pre","gThw: [number, number, number]","numPatches","vision","imageTokenId","imageRun: number[]","imageRun","inputIds","patches: Float32Array","gridTHW: [number, number, number]","input","ids","cap","tokenQueue: string[]","resolve: (() => void) | null","checks: IntegrityCheckEntry[]","argmax","entry: IntegrityCheckEntry","e: any","reference: Record<string, { sum: number; first4: number[] }>","release!: () => void"],"sources":["../src/gpu/architectures/gemma4_vision.ts","../src/gpu/architectures/qwen3_5_vision.ts","../src/gpu/kani-tts.ts","../src/gpu/autocomplete-clean.ts","../src/gpu/sampler.ts","../src/gpu/batch/scheduler.ts","../src/gpu/batch/swarm.ts","../src/gpu/architectures/gepard_tts.ts","../src/gpu/gepard-tts.ts","../src/gpu/architectures/outetts.ts","../src/gpu/outetts-encode.ts","../src/gpu/outetts.ts","../src/gpu/parler-tts.ts","../src/gpu/vision-executor.ts","../src/gpu/vision-preprocess.ts","../src/browser/device-guards.ts","../src/gpu/index.ts"],"sourcesContent":["/**\n * Gemma 4 vision tower (`gemma4_vision`) graph generator.\n *\n * Builds the 16-layer SigLIP-style ViT that turns image patches into pooled,\n * projected image-embedding tokens of dim `text_hidden` (1536 for E2B), ready to\n * splice into the Gemma 4 text stream at `image_token_id`. This is the VISION\n * ENCODER ONLY — it does not splice tokens into the text stream (that is the LM\n * side, src/gpu/index.ts).\n *\n * FAMILY-GENERAL: every dimension is read from `vision_config` + `text_config`,\n * so the same generator serves Gemma 4 E2B / E4B / any sibling that ships a\n * `gemma4_vision` tower — only the dims differ.\n *\n * Architecture (verified against the live HF config.json `vision_config`, the\n * safetensors header of google/gemma-4-E2B-it + mlx-community/gemma-4-e2b-it-4bit,\n * and transformers `modeling_gemma4`):\n *\n *   patch_embedder:\n *     input_proj : Linear(3·patch² → hidden, NO bias)  [hidden, 3·patch²]=[768,768]\n *                  applied to pre-flattened 16×16×3 patches  → [N, 768]\n *     + position embeddings: TWO learned tables [pos_size, hidden]; per patch\n *       look up table[0][x] + table[1][y] (axial 2D), host-gathered → [N, 768]\n *   per encoder layer (Gemma sandwich norms, BIDIRECTIONAL attention):\n *     input_layernorm (RMSNorm) → q/k/v_proj (NO bias)\n *       → per-head q_norm / k_norm (RMSNorm over head_dim, NO v_norm)\n *       → 2D axial RoPE (θ=100) on Q,K  (host cos/sin, rotate_half)\n *       → Attention (NON-causal, 12 heads, head_dim 64)\n *       → o_proj (NO bias) → post_attention_layernorm (RMSNorm) → residual Add\n *     pre_feedforward_layernorm (RMSNorm)\n *       → GeGLU: down( gelu_tanh(gate) · up )   (gate/up/down, NO bias)\n *       → post_feedforward_layernorm (RMSNorm) → residual Add\n *   (no extra final norm — last layer's post_feedforward_layernorm is the tail)\n *   pooling:\n *     spatial k×k average-pool (k = pooling_kernel_size = 3) over the patch grid,\n *     expressed as a host-built sparse [Npool, N] pooling matrix · hidden → [Npool,768]\n *   multimodal projector (embed_vision.embedding_projection):\n *     Linear(hidden → text_hidden)  → [Npool, 1536]  spliced at image_token_id\n *\n * Gemma 4 RMSNorm stores the FULL gain directly (NO +1 bake — see model-loader),\n * matching the text tower. The 2D axial RoPE cos/sin and the axial position-embed\n * gather are functions of the patch grid only (not weights/pixels), so they are\n * computed on the host (vision-preprocess: buildGemma4VisionPositionTensors) and\n * fed as input activations — same discipline as the Qwen ViT.\n *\n * Mobile-safety: the only ops used are MatMul / MatMulBias / RMSNorm /\n * ApplyRotaryEmb / Attention(causal:false) / GELU / Mul / Add — all already\n * mobile-validated (≤16KB workgroup, no select(), clamped exp, no `enable f16`).\n * The pooling reuses MatMul (host pooling matrix) so NO new kernel is required.\n */\n\nimport type { ModelArchConfig, ModelCapabilities, ModelGraph, OpNode, TensorDesc } from \"../ir.js\";\nimport { GEMMA4_VIS_KEYS } from \"../ir.js\";\n\nexport interface Gemma4VisionGraphInfo {\n  hiddenSize: number;\n  numHeads: number;\n  headDim: number;\n  depth: number;\n  intermediateSize: number;\n  textHidden: number;\n  patchSize: number;\n  patchDim: number;\n  poolingKernelSize: number;\n  ropeTheta: number;\n  rmsNormEps: number;\n}\n\n/**\n * Resolve the Gemma 4 vision dims from a raw HF config. Accepts either the\n * top-level config (reads `.vision_config` + `.text_config.hidden_size`) or a\n * bare vision_config (then `textHidden` falls back to the projector row count if\n * present, else hidden). Family-general — no E2B constants.\n */\nexport function resolveGemma4VisionInfo(rawConfig: Record<string, unknown>): Gemma4VisionGraphInfo {\n  const vcfg = (rawConfig.vision_config as Record<string, unknown>) ?? rawConfig;\n  const tcfg = (rawConfig.text_config as Record<string, unknown>) ?? {};\n\n  const hidden_size = vcfg.hidden_size as number; // 768\n  const num_heads = vcfg.num_attention_heads as number; // 12\n  const head_dim = (vcfg.head_dim as number) ?? Math.floor(hidden_size / num_heads); // 64\n  const depth = vcfg.num_hidden_layers as number; // 16\n  const intermediate_size = vcfg.intermediate_size as number; // 3072\n  const patch_size = vcfg.patch_size as number; // 16\n  const in_channels = (vcfg.num_channels as number) ?? 3;\n  const patch_dim = in_channels * patch_size * patch_size; // 768\n  const pooling_kernel_size = (vcfg.pooling_kernel_size as number) ?? 1;\n  const rope = (vcfg.rope_parameters as Record<string, unknown>) ?? {};\n  const rope_theta = (rope.rope_theta as number) ?? 100.0;\n  const rms_norm_eps = (vcfg.rms_norm_eps as number) ?? 1e-6;\n  const text_hidden = (tcfg.hidden_size as number) ?? hidden_size;\n\n  return {\n    hiddenSize: hidden_size,\n    numHeads: num_heads,\n    headDim: head_dim,\n    depth,\n    intermediateSize: intermediate_size,\n    textHidden: text_hidden,\n    patchSize: patch_size,\n    patchDim: patch_dim,\n    poolingKernelSize: pooling_kernel_size,\n    ropeTheta: rope_theta,\n    rmsNormEps: rms_norm_eps,\n  };\n}\n\n/**\n * Dequantize an MLX affine-int4 weight to a plain f32 [rows, cols] matrix.\n * MLX packs 8 int4 values per u32 (low-nibble first); each group of `groupSize`\n * columns shares one scale + bias: w[r,c] = scale[r, c/gs] * q + bias[r, c/gs].\n * Used for the Gemma 4 multimodal projector (`embed_vision.embedding_projection`)\n * in MLX-4bit checkpoints, where (unlike the BF16 ViT body) the projector is int4.\n */\nexport function dequantizeMLXProjection(\n  packed: Uint32Array,\n  scales: Float32Array,\n  biases: Float32Array,\n  rows: number,\n  cols: number,\n  groupSize: number,\n): Float32Array {\n  const out = new Float32Array(rows * cols);\n  const numGroups = cols / groupSize;\n  const u32PerRow = cols / 8;\n  for (let r = 0; r < rows; r++) {\n    const rowPacked = r * u32PerRow;\n    const rowScale = r * numGroups;\n    for (let c = 0; c < cols; c++) {\n      const word = packed[rowPacked + (c >> 3)];\n      const q = (word >>> ((c & 7) * 4)) & 0xf;\n      const g = Math.floor(c / groupSize);\n      out[r * cols + c] = scales[rowScale + g] * q + biases[rowScale + g];\n    }\n  }\n  return out;\n}\n\n/**\n * If the Gemma 4 multimodal projector arrived as an MLX affine-int4 triplet\n * (`embed_vision.embedding_projection.{weight(U32), scales, biases}`), dequantize\n * it in-place to a plain f32 `embed_vision.embedding_projection.weight` and drop\n * the scales/biases, so the vision graph's plain MatMul on the projector works for\n * MLX-4bit checkpoints too. No-op for BF16 (HF) checkpoints (weight already f32).\n */\nexport function dequantizeGemma4VisionProjection(\n  weights: Map<string, { data: ArrayBufferView; shape: number[] }>,\n  groupSize: number,\n  rows: number,\n  cols: number,\n): void {\n  const wKey = GEMMA4_VIS_KEYS.projW;\n  const w = weights.get(wKey);\n  const sc = weights.get(\"embed_vision.embedding_projection.scales\");\n  const bi = weights.get(\"embed_vision.embedding_projection.biases\");\n  // Already f32 (BF16 source) or missing quant sidecars → nothing to do.\n  if (!w || !(w.data instanceof Uint32Array) || !sc || !bi) return;\n  const scales =\n    sc.data instanceof Float32Array\n      ? sc.data\n      : new Float32Array(sc.data.buffer, sc.data.byteOffset, sc.data.byteLength / 4);\n  const biases =\n    bi.data instanceof Float32Array\n      ? bi.data\n      : new Float32Array(bi.data.buffer, bi.data.byteOffset, bi.data.byteLength / 4);\n  const f32 = dequantizeMLXProjection(w.data, scales, biases, rows, cols, groupSize);\n  weights.set(wKey, { data: f32, shape: [rows, cols] });\n  weights.delete(\"embed_vision.embedding_projection.scales\");\n  weights.delete(\"embed_vision.embedding_projection.biases\");\n}\n\n/**\n * Ensure the Gemma3n multimodal-embedder norm weights exist in the weights map.\n *\n * `embed_vision.embedding_post_projection_norm` is a no-scale RMSNorm in HF\n * (Gemma3nRMSNorm(..., with_scale=False)), so the checkpoint ships NO weight for\n * it — we synthesize a ones gain (matching the kernel's `(x/rms)*weight`).\n *\n * `embed_vision.soft_embedding_norm` IS a learned RMSNorm; the full BF16 repo ships\n * its gain. Lean exports (e.g. some MLX-4bit conversions) omit it — we fall back to\n * a ones gain so the model still loads and produces coherent (if not bit-exact)\n * output. Call BEFORE VisionExecutor.uploadWeights().\n *\n * Gemma 4 vision RMSNorm stores the FULL gain (no +1 bake), so a ones gain is a\n * true identity scale.\n */\nexport function ensureGemma4VisionEmbedderNorms(\n  weights: Map<string, { data: ArrayBufferView; shape: number[] }>,\n  visionHidden: number,\n  textHidden: number,\n): void {\n  const ones = (n: number) => ({ data: new Float32Array(n).fill(1), shape: [n] });\n  if (!weights.has(GEMMA4_VIS_KEYS.postProjNormW)) {\n    weights.set(GEMMA4_VIS_KEYS.postProjNormW, ones(textHidden));\n  }\n  if (!weights.has(GEMMA4_VIS_KEYS.softEmbNormW)) {\n    weights.set(GEMMA4_VIS_KEYS.softEmbNormW, ones(visionHidden));\n  }\n}\n\n/**\n * Patch the ClippedMatMul nodes of a Gemma 4 vision graph with the calibrated clip\n * scalars from the checkpoint (Gemma4ClippableLinear's per-tensor input/output\n * min/max buffers), then drop those scalar tensors from the weights map so the\n * vision executor doesn't try to upload them as GPU buffers. Call BEFORE\n * VisionExecutor.uploadWeights(). Missing scalars default to ±inf (clip = identity),\n * so a checkpoint without calibration still loads.\n */\nexport function patchGemma4VisionClips(\n  graph: ModelGraph,\n  weights: Map<string, { data: ArrayBufferView; shape: number[] }>,\n): void {\n  const readScalar = (key: string | undefined): number | undefined => {\n    if (!key) return undefined;\n    const e = weights.get(key);\n    if (!e) return undefined;\n    const d =\n      e.data instanceof Float32Array\n        ? e.data\n        : new Float32Array(e.data.buffer, e.data.byteOffset, e.data.byteLength / 4);\n    return d.length > 0 ? d[0] : undefined;\n  };\n  for (const node of graph.nodes) {\n    if (node.opType !== \"ClippedMatMul\") continue;\n    const a = node.attributes;\n    const imin = readScalar(a.clip_input_min_key as string | undefined);\n    const imax = readScalar(a.clip_input_max_key as string | undefined);\n    const omin = readScalar(a.clip_output_min_key as string | undefined);\n    const omax = readScalar(a.clip_output_max_key as string | undefined);\n    if (imin !== undefined) a.imin = imin;\n    if (imax !== undefined) a.imax = imax;\n    if (omin !== undefined) a.omin = omin;\n    if (omax !== undefined) a.omax = omax;\n    for (const k of [\n      a.clip_input_min_key,\n      a.clip_input_max_key,\n      a.clip_output_min_key,\n      a.clip_output_max_key,\n    ] as Array<string | undefined>) {\n      if (k) weights.delete(k);\n    }\n  }\n}\n\n/**\n * Build the Gemma 4 ViT graph. Shaped by symbolic \"N\" (number of patches, runtime)\n * and \"Np\" (number of pooled tokens = ceil(grid_h/k)·ceil(grid_w/k), runtime),\n * resolved from input tensor dims — like the Qwen ViT's \"N\"/\"Nm\".\n */\nexport function generateGemma4VisionGraph(rawConfig: Record<string, unknown>): ModelGraph {\n  const info = resolveGemma4VisionInfo(rawConfig);\n  const {\n    hiddenSize: hidden_size,\n    numHeads: num_heads,\n    headDim: head_dim,\n    depth,\n    intermediateSize: intermediate_size,\n    textHidden: text_hidden,\n    patchDim: patch_dim,\n    rmsNormEps: eps,\n  } = info;\n\n  const qkv_dim = num_heads * head_dim; // 768 (== hidden for E2B)\n\n  const tensors: Record<string, TensorDesc> = {};\n  const nodes: OpNode[] = [];\n  const executionOrder: string[] = [];\n\n  function addTensor(desc: TensorDesc): void {\n    tensors[desc.name] = desc;\n  }\n  function addNode(node: OpNode): void {\n    nodes.push(node);\n    executionOrder.push(node.id);\n  }\n  function constant(name: string, shape: (number | string)[], safetensorsKey = name): void {\n    addTensor({ name, shape, dtype: \"f32\", storage: \"constant\", safetensorsKey });\n  }\n  function activation(name: string, shape: (number | string)[]): void {\n    addTensor({ name, shape, dtype: \"f32\", storage: \"activation\" });\n  }\n\n  /** MatMul (no bias): out[M,N] = in[M,K] @ W[N,K]^T. M from `mTensor`. */\n  function matmul(\n    id: string,\n    input: string,\n    weight: string,\n    output: string,\n    K: number,\n    N: number,\n    mTensor: string,\n  ): void {\n    addNode({\n      id,\n      opType: \"MatMul\",\n      inputs: [input, weight],\n      outputs: [output],\n      attributes: { M_tensor: mTensor, K, N },\n    });\n  }\n  /**\n   * Gemma4ClippableLinear: out = clamp(clamp(x,imin,imax) @ W^T, omin, omax).\n   * The four clip scalars live in the checkpoint as per-tensor tensors under\n   * `<projBase>.{input_min,input_max,output_min,output_max}`; the loader reads\n   * them and patches the op's imin/imax/omin/omax attributes (and drops the\n   * scalar tensors), like Gemma4's layer_scalar. `projBase` is the canonical\n   * key prefix of the projection (e.g. `vision_tower.encoder.layers.0.self_attn.q_proj`).\n   */\n  function clippedMatmul(\n    id: string,\n    input: string,\n    weight: string,\n    output: string,\n    K: number,\n    N: number,\n    mTensor: string,\n    projBase: string,\n  ): void {\n    addNode({\n      id,\n      opType: \"ClippedMatMul\",\n      inputs: [input, weight],\n      outputs: [output],\n      attributes: {\n        M_tensor: mTensor,\n        K,\n        N,\n        clip_input_min_key: `${projBase}.input_min`,\n        clip_input_max_key: `${projBase}.input_max`,\n        clip_output_min_key: `${projBase}.output_min`,\n        clip_output_max_key: `${projBase}.output_max`,\n      },\n    });\n  }\n  /** RMSNorm over `hidden` (per row), Gemma full-gain (no +1; baked elsewhere=none). */\n  function rmsnorm(\n    id: string,\n    input: string,\n    w: string,\n    output: string,\n    hidden: number,\n    seqTensor: string,\n  ): void {\n    addNode({\n      id,\n      opType: \"RMSNorm\",\n      inputs: [input, w],\n      outputs: [output],\n      attributes: { hidden_size: hidden, eps, seq_len_tensor: seqTensor },\n    });\n  }\n  function residualAdd(id: string, a: string, b: string, output: string): void {\n    addNode({\n      id,\n      opType: \"Add\",\n      inputs: [a, b],\n      outputs: [output],\n      attributes: { count_tensor: a, hidden_size },\n    });\n  }\n\n  // ── Inputs (host-written activations) ──\n  activation(\"g4v_patches\", [\"N\", patch_dim]); // [N, 768] pre-flattened patches\n  activation(\"g4v_pos_embeds\", [\"N\", hidden_size]); // [N, 768] axial pos-embed (host gather)\n  activation(\"g4v_cos\", [\"N\", head_dim]); // [N, 64] 2D-axial rotary cos\n  activation(\"g4v_sin\", [\"N\", head_dim]); // [N, 64] 2D-axial rotary sin\n  // Pooling matrix [Np, N], host-built (avg-pool weights). Multiplied as MatMul:\n  //   pooled[Np, hidden] = poolW[Np, N] @ hidden[N, hidden]  (K=N, output N=hidden)\n  // N is symbolic, so the pooling matrix is declared [Np, N] and resolved at run.\n  activation(\"g4v_pool_w\", [\"Np\", \"N\"]);\n\n  // ── Patch embed: input_proj (no bias) + axial position embeddings ──\n  // input_proj.weight is [hidden, patch_dim]=[768,768]. Linear, no bias.\n  constant(GEMMA4_VIS_KEYS.patchEmbedProjW, [hidden_size, patch_dim]);\n  activation(\"g4v_patch_mm\", [\"N\", hidden_size]);\n  activation(\"g4v_h0\", [\"N\", hidden_size]);\n\n  matmul(\n    \"g4v_patch_proj\",\n    \"g4v_patches\",\n    GEMMA4_VIS_KEYS.patchEmbedProjW,\n    \"g4v_patch_mm\",\n    patch_dim,\n    hidden_size,\n    \"g4v_patches\",\n  );\n  addNode({\n    id: \"g4v_add_pos\",\n    opType: \"Add\",\n    inputs: [\"g4v_patch_mm\", \"g4v_pos_embeds\"],\n    outputs: [\"g4v_h0\"],\n    attributes: { count_tensor: \"g4v_patch_mm\", hidden_size },\n  });\n\n  let prev = \"g4v_h0\";\n\n  for (let i = 0; i < depth; i++) {\n    const p = `g4v_b${i}`;\n\n    // ── input_layernorm (RMSNorm) ──\n    constant(GEMMA4_VIS_KEYS.inputNorm(i), [hidden_size]);\n    activation(`${p}_n1`, [\"N\", hidden_size]);\n    rmsnorm(`${p}_input_norm`, prev, GEMMA4_VIS_KEYS.inputNorm(i), `${p}_n1`, hidden_size, prev);\n\n    // ── q/k/v projections (NO bias) ──\n    constant(GEMMA4_VIS_KEYS.qProjW(i), [qkv_dim, hidden_size]);\n    constant(GEMMA4_VIS_KEYS.kProjW(i), [qkv_dim, hidden_size]);\n    constant(GEMMA4_VIS_KEYS.vProjW(i), [qkv_dim, hidden_size]);\n    activation(`${p}_q`, [\"N\", qkv_dim]);\n    activation(`${p}_k`, [\"N\", qkv_dim]);\n    activation(`${p}_v`, [\"N\", qkv_dim]);\n    const saBase = `vision_tower.encoder.layers.${i}.self_attn`;\n    clippedMatmul(\n      `${p}_q_proj`,\n      `${p}_n1`,\n      GEMMA4_VIS_KEYS.qProjW(i),\n      `${p}_q`,\n      hidden_size,\n      qkv_dim,\n      `${p}_n1`,\n      `${saBase}.q_proj`,\n    );\n    clippedMatmul(\n      `${p}_k_proj`,\n      `${p}_n1`,\n      GEMMA4_VIS_KEYS.kProjW(i),\n      `${p}_k`,\n      hidden_size,\n      qkv_dim,\n      `${p}_n1`,\n      `${saBase}.k_proj`,\n    );\n    clippedMatmul(\n      `${p}_v_proj`,\n      `${p}_n1`,\n      GEMMA4_VIS_KEYS.vProjW(i),\n      `${p}_v`,\n      hidden_size,\n      qkv_dim,\n      `${p}_n1`,\n      `${saBase}.v_proj`,\n    );\n\n    // ── per-head q_norm / k_norm (RMSNorm over head_dim) ──\n    constant(GEMMA4_VIS_KEYS.qNormW(i), [head_dim]);\n    constant(GEMMA4_VIS_KEYS.kNormW(i), [head_dim]);\n    activation(`${p}_qn`, [\"N\", qkv_dim]);\n    activation(`${p}_kn`, [\"N\", qkv_dim]);\n    rmsnorm(`${p}_q_norm`, `${p}_q`, GEMMA4_VIS_KEYS.qNormW(i), `${p}_qn`, head_dim, `${p}_q`);\n    rmsnorm(`${p}_k_norm`, `${p}_k`, GEMMA4_VIS_KEYS.kNormW(i), `${p}_kn`, head_dim, `${p}_k`);\n\n    // ── 2D axial RoPE on Q, K (host cos/sin, rotate_half) ──\n    activation(`${p}_qr`, [\"N\", qkv_dim]);\n    activation(`${p}_kr`, [\"N\", qkv_dim]);\n    addNode({\n      id: `${p}_rope_q`,\n      opType: \"ApplyRotaryEmb\",\n      inputs: [`${p}_qn`, \"g4v_cos\", \"g4v_sin\"],\n      outputs: [`${p}_qr`],\n      attributes: { num_heads, head_dim },\n    });\n    addNode({\n      id: `${p}_rope_k`,\n      opType: \"ApplyRotaryEmb\",\n      inputs: [`${p}_kn`, \"g4v_cos\", \"g4v_sin\"],\n      outputs: [`${p}_kr`],\n      attributes: { num_heads, head_dim },\n    });\n\n    // ── bidirectional attention ──\n    activation(`${p}_attn`, [\"N\", qkv_dim]);\n    addNode({\n      id: `${p}_attention`,\n      opType: \"Attention\",\n      inputs: [`${p}_qr`, `${p}_kr`, `${p}_v`],\n      outputs: [`${p}_attn`],\n      attributes: { num_q_heads: num_heads, num_kv_heads: num_heads, head_dim, causal: false },\n    });\n\n    // ── o_proj (NO bias) ──\n    constant(GEMMA4_VIS_KEYS.oProjW(i), [hidden_size, qkv_dim]);\n    activation(`${p}_o`, [\"N\", hidden_size]);\n    clippedMatmul(\n      `${p}_o_proj`,\n      `${p}_attn`,\n      GEMMA4_VIS_KEYS.oProjW(i),\n      `${p}_o`,\n      qkv_dim,\n      hidden_size,\n      `${p}_attn`,\n      `${saBase}.o_proj`,\n    );\n\n    // ── post_attention_layernorm (RMSNorm, applied to the attn-block output BEFORE\n    //     the residual add — Gemma \"post-norm\" sandwich) ──\n    constant(GEMMA4_VIS_KEYS.postAttnNorm(i), [hidden_size]);\n    activation(`${p}_post_attn`, [\"N\", hidden_size]);\n    rmsnorm(\n      `${p}_post_attn_norm`,\n      `${p}_o`,\n      GEMMA4_VIS_KEYS.postAttnNorm(i),\n      `${p}_post_attn`,\n      hidden_size,\n      `${p}_o`,\n    );\n\n    // residual 1\n    activation(`${p}_res1`, [\"N\", hidden_size]);\n    residualAdd(`${p}_residual1`, prev, `${p}_post_attn`, `${p}_res1`);\n\n    // ── pre_feedforward_layernorm (RMSNorm) ──\n    constant(GEMMA4_VIS_KEYS.preFfNorm(i), [hidden_size]);\n    activation(`${p}_n2`, [\"N\", hidden_size]);\n    rmsnorm(\n      `${p}_pre_ff_norm`,\n      `${p}_res1`,\n      GEMMA4_VIS_KEYS.preFfNorm(i),\n      `${p}_n2`,\n      hidden_size,\n      `${p}_res1`,\n    );\n\n    // ── GeGLU MLP: down( gelu_tanh(gate) · up )  (gate/up/down, NO bias) ──\n    constant(GEMMA4_VIS_KEYS.gateProjW(i), [intermediate_size, hidden_size]);\n    constant(GEMMA4_VIS_KEYS.upProjW(i), [intermediate_size, hidden_size]);\n    constant(GEMMA4_VIS_KEYS.downProjW(i), [hidden_size, intermediate_size]);\n    activation(`${p}_gate`, [\"N\", intermediate_size]);\n    activation(`${p}_up`, [\"N\", intermediate_size]);\n    activation(`${p}_gelu`, [\"N\", intermediate_size]);\n    activation(`${p}_geglu`, [\"N\", intermediate_size]);\n    activation(`${p}_mlp`, [\"N\", hidden_size]);\n\n    const mlpBase = `vision_tower.encoder.layers.${i}.mlp`;\n    clippedMatmul(\n      `${p}_gate_proj`,\n      `${p}_n2`,\n      GEMMA4_VIS_KEYS.gateProjW(i),\n      `${p}_gate`,\n      hidden_size,\n      intermediate_size,\n      `${p}_n2`,\n      `${mlpBase}.gate_proj`,\n    );\n    clippedMatmul(\n      `${p}_up_proj`,\n      `${p}_n2`,\n      GEMMA4_VIS_KEYS.upProjW(i),\n      `${p}_up`,\n      hidden_size,\n      intermediate_size,\n      `${p}_n2`,\n      `${mlpBase}.up_proj`,\n    );\n    addNode({\n      id: `${p}_gelu_op`,\n      opType: \"GELU\", // gelu_pytorch_tanh\n      inputs: [`${p}_gate`],\n      outputs: [`${p}_gelu`],\n      attributes: { count_tensor: `${p}_gate` },\n    });\n    addNode({\n      id: `${p}_geglu_mul`,\n      opType: \"Mul\",\n      inputs: [`${p}_gelu`, `${p}_up`],\n      outputs: [`${p}_geglu`],\n      attributes: { count_tensor: `${p}_gelu`, hidden_size: intermediate_size },\n    });\n    clippedMatmul(\n      `${p}_down_proj`,\n      `${p}_geglu`,\n      GEMMA4_VIS_KEYS.downProjW(i),\n      `${p}_mlp`,\n      intermediate_size,\n      hidden_size,\n      `${p}_geglu`,\n      `${mlpBase}.down_proj`,\n    );\n\n    // ── post_feedforward_layernorm (RMSNorm, before residual add) ──\n    constant(GEMMA4_VIS_KEYS.postFfNorm(i), [hidden_size]);\n    activation(`${p}_post_ff`, [\"N\", hidden_size]);\n    rmsnorm(\n      `${p}_post_ff_norm`,\n      `${p}_mlp`,\n      GEMMA4_VIS_KEYS.postFfNorm(i),\n      `${p}_post_ff`,\n      hidden_size,\n      `${p}_res1`,\n    );\n\n    // residual 2\n    activation(`${p}_res2`, [\"N\", hidden_size]);\n    residualAdd(`${p}_residual2`, `${p}_res1`, `${p}_post_ff`, `${p}_res2`);\n\n    prev = `${p}_res2`;\n  }\n\n  // ── Spatial average pooling → pooled tokens [Np, hidden] ──\n  // pooled = poolW[Np, N] @ hidden[N, hidden]. Reuse MatMul: input=poolW (the\n  // \"activation\" matrix), weight=encoder output viewed as [hidden, N]? No — MatMul\n  // is out[M,K]@W[Nout,K]^T. We instead treat poolW as the LHS [Np, N] and the\n  // encoder hidden as the RHS, but MatMul transposes the 2nd operand. To compute\n  // poolW @ hidden directly we use a dedicated AvgPoolMatMul that does\n  //   out[r, c] = sum_k poolW[r, k] * hidden[k, c]   (no transpose on hidden).\n  activation(\"g4v_pooled\", [\"Np\", hidden_size]);\n  addNode({\n    id: \"g4v_pool\",\n    opType: \"PoolMatMul\",\n    inputs: [\"g4v_pool_w\", prev],\n    outputs: [\"g4v_pooled\"],\n    attributes: { K_tensor: prev, hidden_size, np_tensor: \"g4v_pool_w\" },\n  });\n\n  // ── Multimodal embedder (Gemma3nMultimodalEmbedder), soft/image path ──\n  // HF forward for image features:\n  //   x = pooled * sqrt(vision_hidden)          # scale-invariant before RMSNorm → omitted\n  //   x = soft_embedding_norm(x)                # RMSNorm over vision hidden (WITH gain)\n  //   x = embedding_projection(x)               # Linear(vision_hidden → text_hidden), no bias\n  //   x = embedding_post_projection_norm(x)     # RMSNorm over text_hidden, with_scale=False (gain=1)\n  // The post-projection RMSNorm is what brings the image-soft tokens to the same\n  // ~unit-RMS scale as the ×sqrt(text_hidden) text embeddings they are spliced into;\n  // without it the projector output is ~sqrt(text_hidden)× too large and the LM\n  // ignores the image. The post-proj norm has no learned gain, so the loader\n  // synthesizes a ones weight; soft_embedding_norm's gain is a real weight (loader\n  // falls back to ones if a lean checkpoint omits it).\n  constant(GEMMA4_VIS_KEYS.softEmbNormW, [hidden_size]);\n  activation(\"g4v_soft_normed\", [\"Np\", hidden_size]);\n  rmsnorm(\n    \"g4v_soft_emb_norm\",\n    \"g4v_pooled\",\n    GEMMA4_VIS_KEYS.softEmbNormW,\n    \"g4v_soft_normed\",\n    hidden_size,\n    \"g4v_pooled\",\n  );\n\n  // embed_vision.embedding_projection.weight is [text_hidden, hidden]=[1536,768].\n  // No bias in the released checkpoint (weight/scales/biases are the int4 quant\n  // triplet for the WEIGHT, dequantized by dequantizeGemma4VisionProjection).\n  constant(GEMMA4_VIS_KEYS.projW, [text_hidden, hidden_size]);\n  activation(\"g4v_proj_out\", [\"Np\", text_hidden]);\n  matmul(\n    \"g4v_proj\",\n    \"g4v_soft_normed\",\n    GEMMA4_VIS_KEYS.projW,\n    \"g4v_proj_out\",\n    hidden_size,\n    text_hidden,\n    \"g4v_proj_out\",\n  );\n\n  constant(GEMMA4_VIS_KEYS.postProjNormW, [text_hidden]);\n  activation(\"g4v_image_embeds\", [\"Np\", text_hidden]);\n  rmsnorm(\n    \"g4v_post_proj_norm\",\n    \"g4v_proj_out\",\n    GEMMA4_VIS_KEYS.postProjNormW,\n    \"g4v_image_embeds\",\n    text_hidden,\n    \"g4v_proj_out\",\n  );\n\n  const config: ModelArchConfig = {\n    hidden_size,\n    num_layers: depth,\n    num_heads,\n    num_kv_heads: num_heads,\n    head_dim,\n    intermediate_size,\n    vocab_size: 0,\n    context_length:\n      ((rawConfig.vision_config as Record<string, unknown> | undefined)\n        ?.max_position_embeddings as number) ?? 0,\n    rms_norm_eps: eps,\n    norm_type: \"rmsnorm\",\n    rope_base: info.ropeTheta,\n    rope_dim: head_dim,\n    kv_layout: \"LHSd\",\n    is_moe: false,\n    has_vision_tower: true,\n    vision_architecture: \"gemma4_vision\",\n    vision_patch_size: info.patchSize,\n    vision_embed_dim: hidden_size,\n  };\n\n  const capabilities: ModelCapabilities = { text: true, vision: true, moe: false };\n\n  return {\n    architecture: \"Gemma4VisionModel\",\n    config,\n    capabilities,\n    tensors,\n    nodes,\n    executionOrder,\n    inputs: [\"g4v_patches\", \"g4v_pos_embeds\", \"g4v_cos\", \"g4v_sin\", \"g4v_pool_w\"],\n    outputs: [\"g4v_image_embeds\"],\n  };\n}\n","/**\n * Qwen3.5 vision tower (ViT) graph generator.\n *\n * Builds the 12-layer ViT that turns image patches into merged image-embedding\n * tokens of dim `out_hidden_size` (1024, matching the LM hidden size). This is\n * the VISION ENCODER ONLY — it does not splice tokens into the text stream.\n *\n * Architecture (verified against transformers 5.12 modeling_qwen3_5.py):\n *\n *  patch_embed:  MatMul(patches[N,1536], proj_w[768,1536]) + proj_b  → [N,768]\n *                then + pos_embeds (bilinear-interpolated, host-precomputed)\n *  per block (pre-norm, bidirectional):\n *    norm1 (LayerNorm) → qkv (MatMul[2304,768]+bias) → split Q/K/V [N,768]\n *      → ApplyRotaryEmb(Q), ApplyRotaryEmb(K)   (precomputed cos/sin, rotate_half)\n *      → Attention (NON-causal, 12 heads, head_dim 64)\n *      → proj (MatMul[768,768]+bias) → residual Add\n *    norm2 (LayerNorm) → fc1 (MatMul[3072,768]+bias) → GELU(tanh)\n *      → fc2 (MatMul[768,3072]+bias) → residual Add\n *  merger:\n *    norm (LayerNorm over 768) → reshape [N,768]→[N/4,3072] (free, row-major)\n *      → fc1 (MatMul[3072,3072]+bias) → GELU(exact erf) → fc2 (MatMul[1024,3072]+bias)\n *    → merged image tokens [N/4, 1024]\n *\n * The bilinear pos-embed interpolation and the 2D rotary cos/sin are position-\n * (grid-)dependent only — they're cheap and computed on the host (see\n * vision-preprocess.ts), then fed as input activations. This keeps the GPU graph\n * to the weight-dependent math while staying byte-identical to HF.\n *\n * Patches must already be ordered into spatial_merge_size×spatial_merge_size\n * groups (the HF image processor does this), so the merger's reshape picks up\n * correct 2×2 blocks with no gather.\n */\n\nimport type { ModelArchConfig, ModelCapabilities, ModelGraph, OpNode, TensorDesc } from \"../ir.js\";\nimport { CANONICAL_KEYS } from \"../ir.js\";\n\nconst VIS_NORM_EPS = 1e-6;\n\nexport interface VisionGraphInfo {\n  hiddenSize: number;\n  numHeads: number;\n  headDim: number;\n  depth: number;\n  intermediateSize: number;\n  outHiddenSize: number;\n  spatialMergeSize: number;\n  patchDim: number;\n  mergedDim: number;\n}\n\n/**\n * Build the ViT graph. The graph is shaped by symbolic \"N\" (number of patches),\n * resolved at run time from the input tensor's first dim — exactly like the LM's\n * symbolic \"T\".\n */\nexport function generateQwen3_5VisionGraph(rawConfig: Record<string, unknown>): ModelGraph {\n  const vcfg = (rawConfig.vision_config as Record<string, unknown>) ?? rawConfig;\n\n  const hidden_size = vcfg.hidden_size as number; // 768\n  const num_heads = vcfg.num_heads as number; // 12\n  const head_dim = Math.floor(hidden_size / num_heads); // 64\n  const depth = vcfg.depth as number; // 12\n  const intermediate_size = vcfg.intermediate_size as number; // 3072\n  const out_hidden_size = vcfg.out_hidden_size as number; // 1024\n  const spatial_merge_size = vcfg.spatial_merge_size as number; // 2\n  const in_channels = vcfg.in_channels as number; // 3\n  const temporal_patch_size = vcfg.temporal_patch_size as number; // 2\n  const patch_size = vcfg.patch_size as number; // 16\n  const patch_dim = in_channels * temporal_patch_size * patch_size * patch_size; // 1536\n  const merge_unit = spatial_merge_size * spatial_merge_size; // 4\n  const merged_in = hidden_size * merge_unit; // 3072\n\n  const tensors: Record<string, TensorDesc> = {};\n  const nodes: OpNode[] = [];\n  const executionOrder: string[] = [];\n\n  function addTensor(desc: TensorDesc): void {\n    tensors[desc.name] = desc;\n  }\n  function addNode(node: OpNode): void {\n    nodes.push(node);\n    executionOrder.push(node.id);\n  }\n  function constant(name: string, shape: (number | string)[]): void {\n    addTensor({ name, shape, dtype: \"f32\", storage: \"constant\", safetensorsKey: name });\n  }\n  function activation(name: string, shape: (number | string)[]): void {\n    addTensor({ name, shape, dtype: \"f32\", storage: \"activation\" });\n  }\n\n  /** MatMul (no bias) → out[N, Nout] = in[N,K] @ W[Nout,K]^T. M resolved from `mTensor`. */\n  function matmul(\n    id: string,\n    input: string,\n    weight: string,\n    output: string,\n    K: number,\n    N: number,\n    mTensor?: string,\n    mStatic?: number,\n  ): void {\n    addNode({\n      id,\n      opType: \"MatMul\",\n      inputs: [input, weight],\n      outputs: [output],\n      attributes: mTensor ? { M_tensor: mTensor, K, N } : { M: mStatic, K, N },\n    });\n  }\n  /** out[r,c] = in[r,c] + bias[c]. */\n  function addBias(id: string, input: string, bias: string, output: string, width: number): void {\n    addNode({\n      id,\n      opType: \"AddBias\",\n      inputs: [input, bias],\n      outputs: [output],\n      attributes: { width },\n    });\n  }\n  /**\n   * Fused MatMul + row-broadcast bias → out = in @ W^T + bias. Replaces a MatMul\n   * followed by a separate AddBias, removing a full read+write of the matmul\n   * output per ViT linear layer.\n   */\n  function matmulBias(\n    id: string,\n    input: string,\n    weight: string,\n    bias: string,\n    output: string,\n    K: number,\n    N: number,\n    mTensor: string,\n  ): void {\n    addNode({\n      id,\n      opType: \"MatMulBias\",\n      inputs: [input, weight, bias],\n      outputs: [output],\n      attributes: { M_tensor: mTensor, K, N },\n    });\n  }\n  function layernorm(\n    id: string,\n    input: string,\n    w: string,\n    b: string,\n    output: string,\n    hidden: number,\n    seqTensor: string,\n  ): void {\n    addNode({\n      id,\n      opType: \"LayerNorm\",\n      inputs: [input, w, b],\n      outputs: [output],\n      attributes: { hidden_size: hidden, eps: VIS_NORM_EPS, seq_len_tensor: seqTensor },\n    });\n  }\n\n  // ── Inputs (host-written activations) ──\n  activation(\"vis_patches\", [\"N\", patch_dim]); // [N, 1536]\n  activation(\"vis_pos_embeds\", [\"N\", hidden_size]); // [N, 768]\n  activation(\"vis_cos\", [\"N\", head_dim]); // [N, 64]\n  activation(\"vis_sin\", [\"N\", head_dim]); // [N, 64]\n\n  // ── Patch embed: MatMul + bias + pos_embeds ──\n  constant(CANONICAL_KEYS.visPatchEmbedWeight, [hidden_size, patch_dim]);\n  constant(CANONICAL_KEYS.visPatchEmbedBias, [hidden_size]);\n  activation(\"vis_patch_mm\", [\"N\", hidden_size]);\n  activation(\"vis_patch_bias\", [\"N\", hidden_size]);\n  activation(\"vis_h0\", [\"N\", hidden_size]);\n\n  matmulBias(\n    \"vis_patch_mm\",\n    \"vis_patches\",\n    CANONICAL_KEYS.visPatchEmbedWeight,\n    CANONICAL_KEYS.visPatchEmbedBias,\n    \"vis_patch_bias\",\n    patch_dim,\n    hidden_size,\n    \"vis_patches\",\n  );\n  addNode({\n    id: \"vis_add_pos\",\n    opType: \"Add\",\n    inputs: [\"vis_patch_bias\", \"vis_pos_embeds\"],\n    outputs: [\"vis_h0\"],\n    attributes: { count_tensor: \"vis_patch_bias\", hidden_size },\n  });\n\n  let prev = \"vis_h0\";\n  const qkv_dim = hidden_size * 3; // 2304\n\n  for (let i = 0; i < depth; i++) {\n    const p = `vis_b${i}`;\n\n    // norm1\n    constant(CANONICAL_KEYS.visBlockNorm1W(i), [hidden_size]);\n    constant(CANONICAL_KEYS.visBlockNorm1B(i), [hidden_size]);\n    activation(`${p}_n1`, [\"N\", hidden_size]);\n    layernorm(\n      `${p}_norm1`,\n      prev,\n      CANONICAL_KEYS.visBlockNorm1W(i),\n      CANONICAL_KEYS.visBlockNorm1B(i),\n      `${p}_n1`,\n      hidden_size,\n      prev,\n    );\n\n    // qkv projection + bias\n    constant(CANONICAL_KEYS.visBlockQkvW(i), [qkv_dim, hidden_size]);\n    constant(CANONICAL_KEYS.visBlockQkvB(i), [qkv_dim]);\n    activation(`${p}_qkv_mm`, [\"N\", qkv_dim]);\n    activation(`${p}_qkv`, [\"N\", qkv_dim]);\n    matmulBias(\n      `${p}_qkv_proj`,\n      `${p}_n1`,\n      CANONICAL_KEYS.visBlockQkvW(i),\n      CANONICAL_KEYS.visBlockQkvB(i),\n      `${p}_qkv`,\n      hidden_size,\n      qkv_dim,\n      `${p}_n1`,\n    );\n\n    // split Q/K/V [N,768] from fused [N,2304]\n    activation(`${p}_q`, [\"N\", hidden_size]);\n    activation(`${p}_k`, [\"N\", hidden_size]);\n    activation(`${p}_v`, [\"N\", hidden_size]);\n    const sliceCols = (id: string, output: string, offset: number) =>\n      addNode({\n        id,\n        opType: \"SliceCols\",\n        inputs: [`${p}_qkv`],\n        outputs: [output],\n        attributes: {\n          in_width: qkv_dim,\n          out_width: hidden_size,\n          col_offset: offset,\n          seq_len_tensor: `${p}_qkv`,\n        },\n      });\n    sliceCols(`${p}_split_q`, `${p}_q`, 0);\n    sliceCols(`${p}_split_k`, `${p}_k`, hidden_size);\n    sliceCols(`${p}_split_v`, `${p}_v`, 2 * hidden_size);\n\n    // rotary on Q, K\n    activation(`${p}_qr`, [\"N\", hidden_size]);\n    activation(`${p}_kr`, [\"N\", hidden_size]);\n    addNode({\n      id: `${p}_rope_q`,\n      opType: \"ApplyRotaryEmb\",\n      inputs: [`${p}_q`, \"vis_cos\", \"vis_sin\"],\n      outputs: [`${p}_qr`],\n      attributes: { num_heads, head_dim },\n    });\n    addNode({\n      id: `${p}_rope_k`,\n      opType: \"ApplyRotaryEmb\",\n      inputs: [`${p}_k`, \"vis_cos\", \"vis_sin\"],\n      outputs: [`${p}_kr`],\n      attributes: { num_heads, head_dim },\n    });\n\n    // bidirectional attention\n    activation(`${p}_attn`, [\"N\", hidden_size]);\n    addNode({\n      id: `${p}_attention`,\n      opType: \"Attention\",\n      inputs: [`${p}_qr`, `${p}_kr`, `${p}_v`],\n      outputs: [`${p}_attn`],\n      attributes: { num_q_heads: num_heads, num_kv_heads: num_heads, head_dim, causal: false },\n    });\n\n    // output projection + bias\n    constant(CANONICAL_KEYS.visBlockProjW(i), [hidden_size, hidden_size]);\n    constant(CANONICAL_KEYS.visBlockProjB(i), [hidden_size]);\n    activation(`${p}_proj_mm`, [\"N\", hidden_size]);\n    activation(`${p}_proj`, [\"N\", hidden_size]);\n    matmulBias(\n      `${p}_proj_op`,\n      `${p}_attn`,\n      CANONICAL_KEYS.visBlockProjW(i),\n      CANONICAL_KEYS.visBlockProjB(i),\n      `${p}_proj`,\n      hidden_size,\n      hidden_size,\n      `${p}_attn`,\n    );\n\n    // residual 1\n    activation(`${p}_res1`, [\"N\", hidden_size]);\n    addNode({\n      id: `${p}_residual1`,\n      opType: \"Add\",\n      inputs: [prev, `${p}_proj`],\n      outputs: [`${p}_res1`],\n      attributes: { count_tensor: prev, hidden_size },\n    });\n\n    // norm2 + MLP (GELU tanh)\n    constant(CANONICAL_KEYS.visBlockNorm2W(i), [hidden_size]);\n    constant(CANONICAL_KEYS.visBlockNorm2B(i), [hidden_size]);\n    constant(CANONICAL_KEYS.visBlockFc1W(i), [intermediate_size, hidden_size]);\n    constant(CANONICAL_KEYS.visBlockFc1B(i), [intermediate_size]);\n    constant(CANONICAL_KEYS.visBlockFc2W(i), [hidden_size, intermediate_size]);\n    constant(CANONICAL_KEYS.visBlockFc2B(i), [hidden_size]);\n    activation(`${p}_n2`, [\"N\", hidden_size]);\n    activation(`${p}_fc1_mm`, [\"N\", intermediate_size]);\n    activation(`${p}_fc1`, [\"N\", intermediate_size]);\n    activation(`${p}_gelu`, [\"N\", intermediate_size]);\n    activation(`${p}_fc2_mm`, [\"N\", hidden_size]);\n    activation(`${p}_fc2`, [\"N\", hidden_size]);\n    activation(`${p}_res2`, [\"N\", hidden_size]);\n\n    layernorm(\n      `${p}_norm2`,\n      `${p}_res1`,\n      CANONICAL_KEYS.visBlockNorm2W(i),\n      CANONICAL_KEYS.visBlockNorm2B(i),\n      `${p}_n2`,\n      hidden_size,\n      `${p}_res1`,\n    );\n    matmulBias(\n      `${p}_fc1_op`,\n      `${p}_n2`,\n      CANONICAL_KEYS.visBlockFc1W(i),\n      CANONICAL_KEYS.visBlockFc1B(i),\n      `${p}_fc1`,\n      hidden_size,\n      intermediate_size,\n      `${p}_n2`,\n    );\n    addNode({\n      id: `${p}_gelu_op`,\n      opType: \"GELU\",\n      inputs: [`${p}_fc1`],\n      outputs: [`${p}_gelu`],\n      attributes: { count_tensor: `${p}_fc1` },\n    });\n    matmulBias(\n      `${p}_fc2_op`,\n      `${p}_gelu`,\n      CANONICAL_KEYS.visBlockFc2W(i),\n      CANONICAL_KEYS.visBlockFc2B(i),\n      `${p}_fc2`,\n      intermediate_size,\n      hidden_size,\n      `${p}_gelu`,\n    );\n    addNode({\n      id: `${p}_residual2`,\n      opType: \"Add\",\n      inputs: [`${p}_res1`, `${p}_fc2`],\n      outputs: [`${p}_res2`],\n      attributes: { count_tensor: `${p}_res1`, hidden_size },\n    });\n\n    prev = `${p}_res2`;\n  }\n\n  // ── Merger ──\n  // norm over 768 (per-patch), then view [N,768]→[N/4,3072] (free), fc1 → erf-GELU → fc2.\n  // M for merger matmuls is N/4 (resolved from the [N/4,*] output tensor via M_tensor).\n  constant(CANONICAL_KEYS.visMergerNormW, [hidden_size]);\n  constant(CANONICAL_KEYS.visMergerNormB, [hidden_size]);\n  constant(CANONICAL_KEYS.visMergerFc1W, [merged_in, merged_in]);\n  constant(CANONICAL_KEYS.visMergerFc1B, [merged_in]);\n  constant(CANONICAL_KEYS.visMergerFc2W, [out_hidden_size, merged_in]);\n  constant(CANONICAL_KEYS.visMergerFc2B, [out_hidden_size]);\n\n  activation(\"vis_merge_norm\", [\"N\", hidden_size]);\n  // Reshaped views: these alias the same memory as the [N,768] norm output but\n  // are addressed as [N/4, 3072]. The executor allocates per-name buffers, so we\n  // declare them with the merged row count and copy is implicit (same byte layout).\n  activation(\"vis_merge_fc1_mm\", [\"Nm\", merged_in]);\n  activation(\"vis_merge_fc1\", [\"Nm\", merged_in]);\n  activation(\"vis_merge_gelu\", [\"Nm\", merged_in]);\n  activation(\"vis_merge_fc2_mm\", [\"Nm\", out_hidden_size]);\n  activation(\"vis_image_embeds\", [\"Nm\", out_hidden_size]);\n\n  layernorm(\n    \"vis_merger_norm\",\n    prev,\n    CANONICAL_KEYS.visMergerNormW,\n    CANONICAL_KEYS.visMergerNormB,\n    \"vis_merge_norm\",\n    hidden_size,\n    prev,\n  );\n  // fc1: input is vis_merge_norm viewed as [Nm, 3072]. K=3072, N=3072, M=Nm.\n  matmulBias(\n    \"vis_merger_fc1\",\n    \"vis_merge_norm\",\n    CANONICAL_KEYS.visMergerFc1W,\n    CANONICAL_KEYS.visMergerFc1B,\n    \"vis_merge_fc1\",\n    merged_in,\n    merged_in,\n    \"vis_merge_fc1\",\n  );\n  addNode({\n    id: \"vis_merger_gelu\",\n    opType: \"GeluErf\",\n    inputs: [\"vis_merge_fc1\"],\n    outputs: [\"vis_merge_gelu\"],\n    attributes: { count_tensor: \"vis_merge_fc1\" },\n  });\n  matmulBias(\n    \"vis_merger_fc2\",\n    \"vis_merge_gelu\",\n    CANONICAL_KEYS.visMergerFc2W,\n    CANONICAL_KEYS.visMergerFc2B,\n    \"vis_image_embeds\",\n    merged_in,\n    out_hidden_size,\n    \"vis_image_embeds\",\n  );\n\n  const config: ModelArchConfig = {\n    hidden_size,\n    num_layers: depth,\n    num_heads,\n    num_kv_heads: num_heads,\n    head_dim,\n    intermediate_size,\n    vocab_size: 0,\n    context_length: vcfg.num_position_embeddings as number,\n    rms_norm_eps: VIS_NORM_EPS,\n    norm_type: \"layernorm\",\n    rope_base: 10000.0,\n    rope_dim: head_dim,\n    kv_layout: \"LHSd\",\n    is_moe: false,\n    has_vision_tower: true,\n    vision_architecture: \"qwen3_5_vit\",\n    vision_patch_size: patch_size,\n    vision_embed_dim: hidden_size,\n  };\n\n  const capabilities: ModelCapabilities = { text: true, vision: true, moe: false };\n\n  return {\n    architecture: \"Qwen3_5VisionModel\",\n    config,\n    capabilities,\n    tensors,\n    nodes,\n    executionOrder,\n    inputs: [\"vis_patches\", \"vis_pos_embeds\", \"vis_cos\", \"vis_sin\"],\n    outputs: [\"vis_image_embeds\"],\n  };\n}\n","/**\n * KaniTTS — native text-to-speech engine for Gerbil's WebGPU backend.\n *\n * Kani-TTS-2 (nineninesix/kani-tts-2-en) is a two-stage TTS model that, like\n * Moonshine, needs more than one graph:\n *\n *   1. CODEC-LM BACKBONE (LFM2-350M body): autoregressively emits NanoCodec audio\n *      tokens (4 per frame) into the same vocab as text. Reuses LFM2's block math\n *      with two KaniTTS2 deltas — frame-level position IDs (the 4 audio tokens of a\n *      frame share a position) and learnable per-layer RoPE (α^(l)-scaled freqs) —\n *      both folded host-side into per-layer cos/sin tables fed to the MRoPE op.\n *   2. NANOCODEC DECODER (NVIDIA NeMo 22 kHz): FSQ dequant + causal HiFi-GAN conv\n *      decoder → 22 kHz PCM. Validated bit-exact vs MLX (test-nanocodec-decode.mjs).\n *\n * The AR loop runs on the host with full-logit readback so each frame's 4 audio\n * tokens are sampled per-codebook (constrained to the valid codebook window), then\n * the collected codes are decoded once through the NanoCodec graph.\n *\n * Validated on Dawn (desktop) via scripts/engine/test-kani-speak.mjs.\n */\n\nimport {\n  audioTokensToCodes,\n  buildKaniLayerCosSin,\n  computeKaniPositions,\n  generateKaniTtsGraph,\n  generateNanoCodecDecoderGraph,\n  KANI_END_OF_HUMAN,\n  KANI_START_OF_HUMAN,\n  type KaniConfig,\n  kaniAttentionLayerIndices,\n  kaniCosTensor,\n  kaniLayerAlpha,\n  kaniSinTensor,\n  parseKaniConfig,\n} from \"./architectures/kani_tts.js\";\nimport type { GPUContext } from \"./device.js\";\nimport { initGPU } from \"./device.js\";\nimport { Executor } from \"./executor.js\";\nimport { type LoadedKaniTTS, loadKaniTTS, quantizeKaniBackbone } from \"./model-loader.js\";\nimport type { Tokenizer } from \"./tokenizer.js\";\n\n/** Backbone weight precision for the codec-LM. */\nexport type KaniDType = \"q4\" | \"f32\";\n\nconst KANI_SAMPLE_RATE = 22_050;\nconst KANI_HOP = 1764; // PCM samples per frame (product of up-sample rates)\nconst KANI_END_OF_TEXT = 2; // tokenizer end-of-text marker spliced after the text\n/** NanoCodec decode chunking (keeps per-dispatch under WebGPU's 65535 cap). */\nconst KANI_CODEC_CHUNK_FRAMES = 32;\nconst KANI_CODEC_LOOKBACK_FRAMES = 32;\n/** Hard cap so a stuck decode cannot loop forever (reference uses 3000). */\nconst DEFAULT_MAX_NEW_TOKENS = 3000;\n\nexport interface KaniTTSOptions {\n  /** Backbone repo (default nineninesix/kani-tts-450m-0.2-ft, Apache-2.0). */\n  repo?: string;\n  /** NanoCodec repo (default the NeMo 22 kHz MLX checkpoint). */\n  codecRepo?: string;\n  revision?: string;\n  hfToken?: string;\n  cacheDir?: string;\n  /**\n   * Codec-LM backbone weight precision (default \"q4\"):\n   *   • \"q4\"  — group-wise INT4 (W4A16) on the LFM2 linear projections; embedding +\n   *             LM head stay f32. Backbone ~1.5 GB → ~0.37 GB. Recommended default.\n   *   • \"f32\" — full precision (largest, marginally highest fidelity).\n   * The NanoCodec decoder always runs f32 regardless.\n   */\n  dtype?: KaniDType;\n  /** Max self-attn KV-cache length (prompt + generated). Default 2048. */\n  maxSeqLen?: number;\n  onProgress?: (loaded: number, total: number, message: string) => void;\n}\n\nexport interface SpeakOptions {\n  /**\n   * Language/accent tag for KaniTTS2 multi-accent checkpoints (e.g. \"en_us\",\n   * \"en_nyork\", \"en_glasg\"). Applied as a `\"{tag}: {text}\"` prefix ONLY when the\n   * loaded checkpoint advertises accent tags (config.language_settings). On the\n   * plain LFM2 family (the Apache default kani-tts-450m-0.2-ft) there is a single\n   * trained voice and the raw text is tokenized verbatim — the tag is ignored.\n   * See `availableVoices()` for what the loaded model actually supports.\n   */\n  languageTag?: string;\n  /** Sampling temperature for the codec-LM AR decode (default 1.0). Higher = more varied prosody. */\n  temperature?: number;\n  /** Top-p nucleus threshold (default 0.95). */\n  topP?: number;\n  /** Top-k cutoff for the per-codebook sample (default 0 = disabled; keep all valid codes). */\n  topK?: number;\n  /** Repetition penalty (default 1.1). */\n  repetitionPenalty?: number;\n  /** Max audio frames to generate (caps duration). Default unbounded (maxSeqLen). */\n  maxFrames?: number;\n  /** Override max generated tokens (default 3000). */\n  maxNewTokens?: number;\n}\n\nexport interface SpeakResult {\n  /** Mono PCM in [-1, 1]. */\n  pcm: Float32Array;\n  /** Sample rate (22050). */\n  sampleRate: number;\n  /** Number of audio frames decoded. */\n  frames: number;\n  /** Audio duration in seconds. */\n  audioSeconds: number;\n}\n\n/** Build a fresh map of just the constant weights a graph references (see moonshine-stt). */\nfunction selectGraphWeights(\n  graph: { tensors: Record<string, { storage: string; safetensorsKey?: 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\") {\n      continue;\n    }\n    // Resolve by name, falling back to the tensor's safetensorsKey alias (handles\n    // LFM2's embedding_norm→norm rename — the aliasing loadModel() normally does,\n    // which this dedicated loader bypasses). The tied LM head shares the embed\n    // buffer directly (the head MatMul binds `embed_tokens.weight`), so there is no\n    // separate lm_head tensor to alias here.\n    const w =\n      weights.get(name) ?? (desc.safetensorsKey ? weights.get(desc.safetensorsKey) : undefined);\n    if (w) {\n      out.set(name, w);\n    }\n  }\n  return out;\n}\n\n/**\n * Top-p nucleus sample from a (already softmaxed) probability array over `ids`.\n * Sorts by prob desc, keeps the smallest prefix whose cumulative mass ≥ topP,\n * renormalizes, and samples one id.\n */\nfunction nucleusSample(probs: Float32Array, ids: Int32Array, topP: number): number {\n  const n = probs.length;\n  const order = Array.from({ length: n }, (_, i) => i).sort((a, b) => probs[b] - probs[a]);\n  let cum = 0;\n  let cutoff = n;\n  for (let r = 0; r < n; r++) {\n    cum += probs[order[r]];\n    if (cum >= topP) {\n      cutoff = r + 1;\n      break;\n    }\n  }\n  let keepSum = 0;\n  for (let r = 0; r < cutoff; r++) {\n    keepSum += probs[order[r]];\n  }\n  let pick = Math.random() * keepSum;\n  for (let r = 0; r < cutoff; r++) {\n    pick -= probs[order[r]];\n    if (pick <= 0) {\n      return ids[order[r]];\n    }\n  }\n  return ids[order[0]];\n}\n\n/**\n * In-place top-k filter over a probability array: keep only the `k` highest, zero\n * the rest, and renormalize. No-op when k <= 0 or k >= length (keep everything).\n */\nfunction applyTopK(scores: Float32Array, k: number): void {\n  const n = scores.length;\n  if (k <= 0 || k >= n) {\n    return;\n  }\n  const order = Array.from({ length: n }, (_, i) => i).sort((a, b) => scores[b] - scores[a]);\n  let keepSum = 0;\n  for (let r = 0; r < k; r++) {\n    keepSum += scores[order[r]];\n  }\n  for (let r = k; r < n; r++) {\n    scores[order[r]] = 0;\n  }\n  if (keepSum > 0) {\n    for (let i = 0; i < n; i++) {\n      scores[i] /= keepSum;\n    }\n  }\n}\n\n/** In-place softmax of `scores`. */\nfunction softmaxInPlace(scores: Float32Array): void {\n  let maxV = Number.NEGATIVE_INFINITY;\n  for (let i = 0; i < scores.length; i++) {\n    if (scores[i] > maxV) {\n      maxV = scores[i];\n    }\n  }\n  let sum = 0;\n  for (let i = 0; i < scores.length; i++) {\n    const e = Math.exp(scores[i] - maxV);\n    scores[i] = e;\n    sum += e;\n  }\n  for (let i = 0; i < scores.length; i++) {\n    scores[i] /= sum;\n  }\n}\n\nexport class KaniTTS {\n  private ctx: GPUContext;\n  private loaded: LoadedKaniTTS;\n  private tokenizer: Tokenizer;\n  private cfg: KaniConfig;\n  private rawConfig: Record<string, unknown>;\n  private maxSeqLen: number;\n  /** Backbone executor (built once; reused across speak() calls). */\n  private backboneExec: Executor;\n  /** The attention layer indices (carry learnable α) and their α values. */\n  private attnLayers: number[];\n  private layerAlpha: Map<number, number>;\n  private headDim: number;\n  private ropeBase: number;\n  private _destroyed = false;\n  /** Accent/voice tags the loaded checkpoint advertises (empty = single voice). */\n  private voiceTags: string[];\n\n  /** The loaded checkpoint's architecture string (\"KaniTTS2ForCausalLM\" or \"Lfm2ForCausalLM\"). */\n  readonly architecture: string;\n  /** The selected backbone precision (\"q4\" or \"f32\"). */\n  readonly dtype: KaniDType;\n\n  private constructor(ctx: GPUContext, loaded: LoadedKaniTTS, maxSeqLen: number, dtype: KaniDType) {\n    this.ctx = ctx;\n    this.loaded = loaded;\n    this.tokenizer = loaded.tokenizer;\n    this.rawConfig = loaded.rawConfig;\n    this.cfg = parseKaniConfig(loaded.rawConfig);\n    this.maxSeqLen = maxSeqLen;\n    // A gerbil-int4 checkpoint ships the linear projections already packed as\n    // .q/.scales/.zeros — it can ONLY be run as q4 (there are no f32 projection\n    // sources to feed an f32 MatMul graph). Force q4 regardless of the requested\n    // dtype so an offline-quantized R2 mirror loads directly without re-quantizing\n    // (quantizeKaniBackbone below is idempotent and becomes a no-op).\n    this.dtype = loaded.preQuantized ? \"q4\" : dtype;\n    const archList = (loaded.rawConfig.architectures as string[] | undefined) ?? [];\n    this.architecture = archList[0] ?? \"Lfm2ForCausalLM\";\n    // Accent tags live under config.language_settings.language_tags_list on the\n    // multi-accent KaniTTS2 checkpoints; the plain LFM2 family omits it (single voice).\n    const langSettings = loaded.rawConfig.language_settings as\n      | { language_tags_list?: unknown }\n      | undefined;\n    this.voiceTags = Array.isArray(langSettings?.language_tags_list)\n      ? (langSettings.language_tags_list as unknown[]).filter(\n          (v): v is string => typeof v === \"string\",\n        )\n      : [];\n\n    const hidden = this.rawConfig.hidden_size as number;\n    const heads = this.rawConfig.num_attention_heads as number;\n    this.headDim = (this.rawConfig.head_dim as number) ?? Math.floor(hidden / heads);\n    this.ropeBase = (this.rawConfig.rope_theta as number) ?? 1_000_000.0;\n\n    // ── Resolve per-layer learnable α from the alpha_weight scalars. ──\n    this.attnLayers = kaniAttentionLayerIndices(this.rawConfig);\n    this.layerAlpha = new Map<number, number>();\n    const useLearnable = (this.rawConfig.use_learnable_rope as boolean) ?? false;\n    for (const layer of this.attnLayers) {\n      let alpha = 1.0;\n      if (useLearnable) {\n        const w = loaded.backboneWeights.get(`learnable_rope_layers.${layer}.alpha_weight`);\n        const aw = w ? ((w.data as Float32Array)[0] ?? 0) : 0;\n        alpha = kaniLayerAlpha(aw, this.cfg);\n      }\n      this.layerAlpha.set(layer, alpha);\n    }\n\n    // ── Build + upload the backbone graph once. ──\n    this.backboneExec = this.buildBackbone(ctx, loaded, maxSeqLen);\n  }\n\n  /**\n   * Build the codec-LM backbone graph, quantize it when q4, and upload its weights\n   * into a fresh Executor. For a pre-quantized (gerbil-int4) checkpoint the linear\n   * projections already carry `.q`/`.scales`/`.zeros`, so quantizeKaniBackbone is an\n   * idempotent no-op (it skips any projection whose `.q` is already present) and the\n   * int4 weights are uploaded verbatim — no on-device re-quantization.\n   */\n  private buildBackbone(ctx: GPUContext, loaded: LoadedKaniTTS, maxSeqLen: number): Executor {\n    const useQ4 = this.dtype === \"q4\";\n    const graph = generateKaniTtsGraph(this.rawConfig, useQ4 ? \"q4\" : undefined);\n    if (useQ4) {\n      quantizeKaniBackbone(graph, loaded.backboneWeights);\n    }\n    const exec = new Executor(ctx, graph, { maxSeqLen, kvMode: \"f32\" });\n    exec.uploadWeightsMap(selectGraphWeights(graph, loaded.backboneWeights));\n    exec.initBindGroups();\n    return exec;\n  }\n\n  static async create(options: KaniTTSOptions = {}): Promise<KaniTTS> {\n    const ctx = await initGPU();\n    const loaded = await loadKaniTTS({\n      repo: options.repo,\n      codecRepo: options.codecRepo,\n      revision: options.revision,\n      hfToken: options.hfToken,\n      cacheDir: options.cacheDir,\n      onProgress: options.onProgress,\n    });\n    const ctxLen = (loaded.rawConfig.max_position_embeddings as number) ?? 4096;\n    const maxSeqLen = Math.min(options.maxSeqLen ?? 2048, ctxLen);\n    return new KaniTTS(ctx, loaded, maxSeqLen, options.dtype ?? \"q4\");\n  }\n\n  /** Write per-layer cos/sin for token rows [rowStart, rowStart+positions.length). */\n  private writeCosSin(positions: Float32Array, rowStart: number): void {\n    const rowBytes = this.headDim * 4; // rope_dim == head_dim, f32\n    for (const layer of this.attnLayers) {\n      const alpha = this.layerAlpha.get(layer) ?? 1.0;\n      const { cos, sin } = buildKaniLayerCosSin(positions, this.headDim, this.ropeBase, alpha);\n      if (rowStart === 0) {\n        this.backboneExec.writeInput(kaniCosTensor(layer), cos);\n        this.backboneExec.writeInput(kaniSinTensor(layer), sin);\n      } else {\n        this.backboneExec.writeInputAt(kaniCosTensor(layer), cos, rowStart * rowBytes);\n        this.backboneExec.writeInputAt(kaniSinTensor(layer), sin, rowStart * rowBytes);\n      }\n    }\n  }\n\n  /**\n   * Accent/voice tags the loaded checkpoint supports. Multi-accent KaniTTS2\n   * checkpoints (kani-tts-2-en) return e.g. [\"en_us\",\"en_nyork\",\"en_glasg\", …];\n   * the plain LFM2 family (the Apache default) returns [] (one trained voice,\n   * selectable only by fine-tuning the backbone — there is no runtime voice/\n   * speaker-prompt knob, and zero-shot cloning is not wired).\n   */\n  availableVoices(): string[] {\n    return [...this.voiceTags];\n  }\n\n  /** NanoCodec's fixed output sample rate (Hz). Not selectable — see SpeakResult. */\n  get sampleRate(): number {\n    return KANI_SAMPLE_RATE;\n  }\n\n  /**\n   * Synthesize speech for `text`. Returns 22 kHz mono PCM.\n   *\n   * Pipeline: build the [SOH]+text+[EOT,EOH] prompt → prefill the codec-LM →\n   * AR-decode 4-token frames (per-codebook constrained sampling) until end_of_speech\n   * → strip markers → codes → NanoCodec decode → PCM.\n   */\n  async speak(text: string, opts: SpeakOptions = {}): Promise<SpeakResult> {\n    if (this._destroyed) {\n      throw new Error(\"KaniTTS has been destroyed.\");\n    }\n    const temperature = opts.temperature ?? 1.0;\n    const topP = opts.topP ?? 0.95;\n    const topK = opts.topK ?? 0;\n    const repetitionPenalty = opts.repetitionPenalty ?? 1.1;\n    const maxNewTokens = Math.min(opts.maxNewTokens ?? DEFAULT_MAX_NEW_TOKENS, this.maxSeqLen);\n\n    // ── Build prompt: [SOH] + tokenizer(text) + [EOT, EOH]. ──\n    // Multi-accent KaniTTS2 checkpoints take a \"{tag}: {text}\" prefix; the plain\n    // LFM2 family tokenizes the raw text verbatim (matching the HF reference's\n    // prepare_input). The tokenizer prepends BOS (= start_of_text = 1) per its config.\n    const promptText =\n      this.voiceTags.length > 0\n        ? `${(opts.languageTag ?? this.voiceTags[0]).trim()}: ${text}`\n        : text;\n    const textIds = this.tokenizer.encode(promptText);\n    const prompt: number[] = [KANI_START_OF_HUMAN, ...textIds, KANI_END_OF_TEXT, KANI_END_OF_HUMAN];\n    if (prompt.length >= this.maxSeqLen) {\n      throw new Error(`Kani prompt is ${prompt.length} tokens >= maxSeqLen ${this.maxSeqLen}.`);\n    }\n\n    // ── Prefill. ──\n    this.backboneExec.reset();\n    const promptPos = computeKaniPositions(prompt, this.cfg);\n    this.writeCosSin(promptPos, 0);\n    const { logits } = await this.backboneExec.forward(new Uint32Array(prompt));\n\n    // ── AR decode loop. ──\n    const startTime = performance.now();\n    const { audioTokens, finished } = await this.runDecodeLoop(prompt, logits, {\n      temperature,\n      topP,\n      topK,\n      repetitionPenalty,\n      maxNewTokens,\n      maxFrames: opts.maxFrames ?? Number.POSITIVE_INFINITY,\n    });\n\n    // ── Strip to whole frames, convert to codes, NanoCodec decode. ──\n    const usable = audioTokens.length - (audioTokens.length % this.cfg.tokensPerFrame);\n    if (usable === 0) {\n      throw new Error(\n        `Kani produced no complete audio frames (finished=${finished}, ` +\n          `audioTokens=${audioTokens.length}). The codec-LM emitted no speech.`,\n      );\n    }\n    const { codes, numFrames } = audioTokensToCodes(audioTokens.slice(0, usable), this.cfg);\n    const pcm = await this.decodeCodes(codes, numFrames);\n\n    const totalTime = (performance.now() - startTime) / 1000;\n    const audioSeconds = (numFrames * KANI_HOP) / KANI_SAMPLE_RATE;\n    console.log(\n      `[kani] frames=${numFrames} audio=${audioSeconds.toFixed(2)}s wall=${totalTime.toFixed(\n        2,\n      )}s RTF=${(audioSeconds / totalTime).toFixed(2)}x`,\n    );\n\n    return { pcm, sampleRate: KANI_SAMPLE_RATE, frames: numFrames, audioSeconds };\n  }\n\n  /**\n   * Autoregressive decode: from the prefill logits, emit one token per step —\n   * greedy for the structural markers ([SOA][SOS]) and per-codebook constrained\n   * sampling once in speech — collecting the audio tokens between SOS and EOS.\n   * Writes each step's per-layer cos/sin row (frame-level logical position) before\n   * the forward. Returns the collected audio tokens and whether EOS/cap was hit.\n   */\n  private async runDecodeLoop(\n    prompt: number[],\n    prefillLogits: Float32Array,\n    p: {\n      temperature: number;\n      topP: number;\n      topK: number;\n      repetitionPenalty: number;\n      maxNewTokens: number;\n      maxFrames: number;\n    },\n  ): Promise<{ audioTokens: number[]; finished: boolean }> {\n    const seq: number[] = [...prompt];\n    const audioTokens: number[] = [];\n    const frameCap = p.maxFrames * this.cfg.tokensPerFrame;\n    let logits = prefillLogits;\n    let inSpeech = false;\n    let finished = false;\n\n    for (let step = 0; step < p.maxNewTokens; step++) {\n      const nextToken = inSpeech\n        ? this.sampleAudioToken(logits, audioTokens.length % this.cfg.tokensPerFrame, {\n            temperature: p.temperature,\n            topP: p.topP,\n            topK: p.topK,\n            repetitionPenalty: p.repetitionPenalty,\n            previous: seq,\n          })\n        : this.argmax(logits);\n\n      if (nextToken === this.cfg.startOfSpeech) {\n        inSpeech = true;\n      } else if (nextToken === this.cfg.endOfSpeech) {\n        seq.push(nextToken);\n        finished = true;\n        break;\n      } else if (inSpeech && nextToken >= this.cfg.audioTokensStart) {\n        audioTokens.push(nextToken);\n      }\n      seq.push(nextToken);\n\n      if (inSpeech && audioTokens.length >= frameCap) {\n        finished = true;\n        break;\n      }\n      if (seq.length >= this.maxSeqLen) {\n        break;\n      }\n\n      // cos/sin row for the next physical position = current seqPos, at the logical\n      // (frame-level) position of the token we are about to feed.\n      this.writeCosSin(\n        Float32Array.of(this.logicalPositionAt(seq)),\n        this.backboneExec.currentSeqPos,\n      );\n      const res = await this.backboneExec.forward(new Uint32Array([nextToken]));\n      logits = res.logits;\n    }\n    return { audioTokens, finished };\n  }\n\n  /** Logical (frame-level) position of the LAST token in `seq`. */\n  private logicalPositionAt(seq: number[]): number {\n    // computeKaniPositions returns positions BEFORE counting each token; the last\n    // entry is the position the final token should be rotated at.\n    const positions = computeKaniPositions(seq, this.cfg);\n    return positions[positions.length - 1];\n  }\n\n  /** Greedy argmax over a logits row. */\n  private argmax(logits: Float32Array): number {\n    let best = 0;\n    let bestVal = logits[0];\n    for (let i = 1; i < logits.length; i++) {\n      if (logits[i] > bestVal) {\n        bestVal = logits[i];\n        best = i;\n      }\n    }\n    return best;\n  }\n\n  /**\n   * Sample one audio token for codebook position `codebook`, constrained to that\n   * codebook's valid window [audio_tokens_start + 4032*c, +4032). Allows end_of_speech\n   * only at codebook 0 (frame boundary). Applies repetition penalty, temperature,\n   * optional top-k, then top-p nucleus sampling.\n   */\n  private sampleAudioToken(\n    logits: Float32Array,\n    codebook: number,\n    p: {\n      temperature: number;\n      topP: number;\n      topK: number;\n      repetitionPenalty: number;\n      previous: number[];\n    },\n  ): number {\n    const winStart = this.cfg.audioTokensStart + this.cfg.codebookSize * codebook;\n    const winLen = this.cfg.codebookSize;\n    // Candidate window: the codebook's valid tokens (+ end_of_speech at a frame\n    // boundary, codebook 0), so the sampled token is always a valid audio code.\n    const allowEos = codebook === 0;\n    const n = winLen + (allowEos ? 1 : 0);\n    const ids = new Int32Array(n);\n    for (let i = 0; i < winLen; i++) {\n      ids[i] = winStart + i;\n    }\n    if (allowEos) {\n      ids[winLen] = this.cfg.endOfSpeech;\n    }\n    // Repetition penalty + temperature → softmax → optional top-k → top-p sample.\n    const prev = p.repetitionPenalty !== 1.0 ? new Set(p.previous) : null;\n    const scores = new Float32Array(n);\n    for (let i = 0; i < n; i++) {\n      let s = logits[ids[i]];\n      if (prev?.has(ids[i])) {\n        s = s > 0 ? s / p.repetitionPenalty : s * p.repetitionPenalty;\n      }\n      scores[i] = s / p.temperature;\n    }\n    softmaxInPlace(scores);\n    applyTopK(scores, p.topK); // no-op when topK <= 0\n    return nucleusSample(scores, ids, p.topP);\n  }\n\n  /**\n   * Decode NanoCodec codes [groups, T] (group-major) → PCM.\n   *\n   * The decoder graph carries concrete lengths, and the upsampled conv activations\n   * for long clips overflow WebGPU's 65535 per-dimension dispatch cap. The decoder\n   * is fully causal with a small (≤ a few frames) receptive field, so we decode in\n   * frame chunks with a left-context lookback and keep only each chunk's own output\n   * samples — numerically identical to a single decode, but bounded per dispatch.\n   */\n  private async decodeCodes(codes: Uint32Array, numFrames: number): Promise<Float32Array> {\n    const groups = 4; // FSQ_NUM_GROUPS (codes are group-major [groups, T])\n    const lookback = KANI_CODEC_LOOKBACK_FRAMES;\n    const chunk = KANI_CODEC_CHUNK_FRAMES;\n    const pcm = new Float32Array(numFrames * KANI_HOP);\n\n    for (let start = 0; start < numFrames; start += chunk) {\n      const ctxStart = Math.max(0, start - lookback);\n      const ctxFrames = start - ctxStart; // discarded warmup frames\n      const end = Math.min(numFrames, start + chunk);\n      const winFrames = end - ctxStart;\n\n      // Slice the codes window [ctxStart, end) for all groups, group-major.\n      const winCodes = new Uint32Array(groups * winFrames);\n      for (let g = 0; g < groups; g++) {\n        for (let t = 0; t < winFrames; t++) {\n          winCodes[g * winFrames + t] = codes[g * numFrames + (ctxStart + t)];\n        }\n      }\n\n      const win = await this.decodeCodesWindow(winCodes, winFrames);\n      // Keep only this chunk's samples (drop the lookback warmup).\n      const dropSamples = ctxFrames * KANI_HOP;\n      const keepSamples = (end - start) * KANI_HOP;\n      pcm.set(win.subarray(dropSamples, dropSamples + keepSamples), start * KANI_HOP);\n    }\n\n    // Clamp to [-1, 1] (NanoCodec's output activation).\n    for (let i = 0; i < pcm.length; i++) {\n      pcm[i] = Math.min(1, Math.max(-1, pcm[i]));\n    }\n    return pcm;\n  }\n\n  /** Run the NanoCodec decoder graph for a single (bounded) code window → PCM. */\n  private async decodeCodesWindow(winCodes: Uint32Array, winFrames: number): Promise<Float32Array> {\n    const graph = generateNanoCodecDecoderGraph({ numFrames: winFrames });\n    const exec = new Executor(this.ctx, graph, { maxSeqLen: winFrames, kvMode: \"f32\" });\n    try {\n      exec.uploadWeightsMap(selectGraphWeights(graph, this.loaded.codecWeights));\n      exec.initBindGroups();\n      exec.reset();\n      exec.writeInput(\"audio_codes\", winCodes);\n      return await exec.runGraphOutput(graph.outputs[0], winFrames * KANI_HOP);\n    } finally {\n      exec.destroy();\n    }\n  }\n\n  destroy(): void {\n    if (this._destroyed) {\n      return;\n    }\n    this._destroyed = true;\n    this.backboneExec.destroy();\n    this.loaded.backboneWeights.clear();\n    this.loaded.codecWeights.clear();\n  }\n}\n","// Post-processing for on-device ghost-text autocomplete suggestions.\n//\n// Small quantized models (e.g. Qwen3.5-0.8B) love to (a) restate the text the\n// user already typed and (b) loop on a phrase. Naive cleanup that only drops an\n// output which echoes the WHOLE typed string verbatim lets a partial\n// tail-overlap (\"…the game is\" → \"the game is to…\") or a restatement of an\n// EARLIER phrase slip straight through, showing up as duplicated ghost text.\n// These pure helpers make the suggestion a clean, novel continuation regardless\n// of what the model emits. They are deliberately dependency-free and\n// side-effect-free so they are trivial to reason about and unit-test.\n\nconst DEFAULT_MAX_SUGGESTION_CHARS = 140;\nconst DEFAULT_RESTATEMENT_MIN_WORDS = 3;\nconst DEFAULT_INTERNAL_REPEAT_MIN_RUN = 6;\nconst LEADING_PUNCT = /^[.,;:!?)\\]}'\"”’%]/;\nconst WRAPPING_QUOTES_START = /^[\"'“”‘’`]+/;\nconst WRAPPING_QUOTES_END = /[\"'“”‘’`]+$/;\nconst TRAILING_WHITESPACE = /\\s+$/;\nconst WHITESPACE_RUN = /\\s+/g;\nconst AFTER_FIRST_NEWLINE = /\\r?\\n[\\s\\S]*$/;\nconst ENDS_WITH_SPACE = /\\s$/;\n\n/**\n * Longest span of characters that is both a suffix of `typed` and a prefix of\n * `suggestion`, dropped from the start of `suggestion`. Trailing whitespace on\n * `typed` is ignored so a duplicated word at the cursor (\"the \" + \"the store\")\n * is caught. Returns `suggestion` unchanged when there is no overlap.\n */\nexport function stripLeadingOverlap(typed: string, suggestion: string): string {\n  const t = typed.replace(TRAILING_WHITESPACE, \"\");\n  const max = Math.min(t.length, suggestion.length);\n  for (let k = max; k > 0; k--) {\n    if (t.slice(t.length - k) === suggestion.slice(0, k)) {\n      return suggestion.slice(k);\n    }\n  }\n  return suggestion;\n}\n\n/**\n * Drop a leading run of words from `suggestion` when that run already appears as\n * a contiguous phrase anywhere in `typed`. This catches the model restating an\n * EARLIER phrase (not just the immediate tail). Matching is case-insensitive,\n * whitespace-normalized, and word-boundary aware so it never strips on a\n * partial-word coincidence. Only strips runs of at least `minWords` words to\n * avoid nuking a legitimate short continuation.\n */\nexport function stripRestatement(\n  typed: string,\n  suggestion: string,\n  minWords = DEFAULT_RESTATEMENT_MIN_WORDS,\n): string {\n  const normTyped = ` ${typed.toLowerCase().replace(WHITESPACE_RUN, \" \").trim()} `;\n  const words = suggestion.split(WHITESPACE_RUN).filter(Boolean);\n  let matched = 0;\n  for (let n = 1; n <= words.length; n++) {\n    const phrase = ` ${words.slice(0, n).join(\" \").toLowerCase()} `;\n    if (normTyped.includes(phrase)) {\n      matched = n;\n    } else {\n      break;\n    }\n  }\n  if (matched >= minWords) {\n    return words.slice(matched).join(\" \");\n  }\n  return suggestion;\n}\n\n/**\n * Truncate `text` right before the first point where a span of `minRun` or more\n * consecutive words repeats a span seen earlier — the classic small-model loop\n * (\"…endless sands …endless sands…\"). Returns `text` unchanged (spacing\n * preserved) when no such repeat exists.\n */\nexport function truncateInternalRepeat(\n  text: string,\n  minRun = DEFAULT_INTERNAL_REPEAT_MIN_RUN,\n): string {\n  const words = text.split(WHITESPACE_RUN).filter(Boolean);\n  const seen = new Set<string>();\n  for (let j = 0; j + minRun <= words.length; j++) {\n    const gram = words\n      .slice(j, j + minRun)\n      .join(\" \")\n      .toLowerCase();\n    if (seen.has(gram)) {\n      return words.slice(0, j).join(\" \");\n    }\n    seen.add(gram);\n  }\n  return text;\n}\n\n/**\n * Cap `text` to at most `max` characters, cutting on a word boundary when a\n * reasonable one exists in the back half of the window.\n */\nexport function capLength(text: string, max: number): string {\n  if (text.length <= max) {\n    return text;\n  }\n  const cut = text.slice(0, max);\n  const lastSpace = cut.lastIndexOf(\" \");\n  return (lastSpace > max * 0.5 ? cut.slice(0, lastSpace) : cut).trimEnd();\n}\n\nexport interface CleanSuggestionOptions {\n  /** Keep only the first line of the raw output (default true). */\n  singleLine?: boolean;\n  /** Character cap for the returned continuation (default 140). */\n  maxChars?: number;\n}\n\n/**\n * Turn a raw model completion into a clean inline ghost continuation:\n * 1. keep only the first line (when `singleLine`),\n * 2. strip wrapping quotes,\n * 3. drop a full verbatim echo of the typed text,\n * 4. drop a leading character overlap with the typed tail,\n * 5. drop a leading word-run that restates an earlier typed phrase,\n * 6. truncate an internal phrase loop,\n * 7. cap the length,\n * 8. add a single smart leading space so the ghost joins the caret naturally.\n *\n * Returns \"\" when nothing novel is left — the ghost then simply shows nothing,\n * which is the correct behavior for a suggestion that only repeats the input.\n */\nexport function cleanSuggestion(\n  raw: string,\n  typed: string,\n  options: CleanSuggestionOptions = {},\n): string {\n  const { singleLine = true, maxChars = DEFAULT_MAX_SUGGESTION_CHARS } = options;\n\n  let s = singleLine ? raw.replace(AFTER_FIRST_NEWLINE, \"\") : raw;\n  s = s.replace(WRAPPING_QUOTES_START, \"\").replace(WRAPPING_QUOTES_END, \"\");\n  s = s.trim();\n  if (!s) {\n    return \"\";\n  }\n\n  const typedTrim = typed.trim();\n  if (typedTrim && s.startsWith(typedTrim)) {\n    s = s.slice(typedTrim.length).trimStart();\n  }\n  s = stripLeadingOverlap(typed, s).trimStart();\n  s = stripRestatement(typed, s);\n  s = truncateInternalRepeat(s);\n  s = capLength(s.trim(), maxChars).trim();\n  if (!s) {\n    return \"\";\n  }\n\n  const startsWithPunct = LEADING_PUNCT.test(s);\n  const typedEndsWithSpace = ENDS_WITH_SPACE.test(typed) || typed.length === 0;\n  return startsWithPunct || typedEndsWithSpace ? s : ` ${s}`;\n}\n","/**\n * CPU-side token sampling from logits.\n *\n * Applies temperature, top-k, and top-p (nucleus) filtering,\n * then samples from the resulting probability distribution.\n *\n * Uses typed arrays and min-heap for zero-allocation top-K selection.\n * For vocab_size ~152K with topK=50, this avoids creating 152K JS tuples.\n */\n\nexport interface SamplingParams {\n  temperature?: number; // Default: 0.7\n  topK?: number; // Default: 50\n  topP?: number; // Default: 0.9\n  repetitionPenalty?: number; // Default: 1.0 (disabled)\n  /**\n   * Seed for the sampling RNG. When set, sampled decoding is reproducible:\n   * the same params object draws a deterministic random stream (seeded on its\n   * first use — pass a fresh options object per generation for run-to-run\n   * reproducibility). Unset: Math.random() (non-deterministic, the default).\n   */\n  seed?: number;\n}\n\n/** mulberry32 — tiny deterministic PRNG for seeded sampling. */\nfunction mulberry32(seed: number): () => number {\n  let a = seed >>> 0;\n  return () => {\n    a = (a + 0x6d2b79f5) | 0;\n    let t = Math.imul(a ^ (a >>> 15), 1 | a);\n    t = (t + Math.imul(t ^ (t >>> 7), 61 | t)) ^ t;\n    return ((t ^ (t >>> 14)) >>> 0) / 4294967296;\n  };\n}\n\n/** Per-params seeded RNG streams (one stream per options object identity). */\nconst seededStreams = new WeakMap<SamplingParams, () => number>();\n\nfunction nextRandom(params: SamplingParams): number {\n  if (params.seed === undefined) return Math.random();\n  let rng = seededStreams.get(params);\n  if (!rng) {\n    rng = mulberry32(params.seed);\n    seededStreams.set(params, rng);\n  }\n  return rng();\n}\n\n// Pre-allocated buffers for top-K selection (avoid GC pressure in hot loop)\nlet _heapIndices: Uint32Array | null = null;\nlet _heapValues: Float32Array | null = null;\n\n/**\n * Sample a token ID from logits.\n *\n * Pipeline: repetition penalty → temperature → top-k (min-heap) → softmax → top-p → sample.\n */\nexport function sampleToken(\n  logits: Float32Array,\n  params: SamplingParams = {},\n  previousTokens?: number[],\n): number {\n  const temperature = params.temperature ?? 0.7;\n  const topK = params.topK ?? 50;\n  const topP = params.topP ?? 0.9;\n  const repetitionPenalty = params.repetitionPenalty ?? 1.0;\n\n  // Greedy decoding when temperature is ~0\n  if (temperature < 1e-6) {\n    return argmax(logits);\n  }\n\n  const N = logits.length;\n  const K = Math.min(topK > 0 ? topK : N, N);\n\n  // Ensure pre-allocated buffers are large enough\n  if (!_heapIndices || _heapIndices.length < K) {\n    _heapIndices = new Uint32Array(K);\n    _heapValues = new Float32Array(K);\n  }\n  const hIdx = _heapIndices;\n  const hVal = _heapValues!;\n\n  // Apply repetition penalty set (built once)\n  let penaltySet: Set<number> | null = null;\n  if (repetitionPenalty !== 1.0 && previousTokens?.length) {\n    penaltySet = new Set(previousTokens);\n  }\n\n  // Single pass: apply temperature + repetition penalty + top-K min-heap\n  let heapSize = 0;\n\n  for (let i = 0; i < N; i++) {\n    let s = logits[i];\n\n    // Repetition penalty\n    if (penaltySet?.has(i)) {\n      s = s > 0 ? s / repetitionPenalty : s * repetitionPenalty;\n    }\n\n    // Temperature\n    s /= temperature;\n\n    if (heapSize < K) {\n      // Fill the heap\n      hIdx[heapSize] = i;\n      hVal[heapSize] = s;\n      heapSize++;\n      if (heapSize === K) {\n        // Build min-heap\n        for (let j = (K >> 1) - 1; j >= 0; j--) {\n          siftDown(hIdx, hVal, j, K);\n        }\n      }\n    } else if (s > hVal[0]) {\n      // Replace min and sift down\n      hIdx[0] = i;\n      hVal[0] = s;\n      siftDown(hIdx, hVal, 0, K);\n    }\n  }\n\n  // Sort top-K by value descending (K is small, ~50)\n  // Simple insertion sort for small K\n  for (let i = 1; i < heapSize; i++) {\n    const vi = hVal[i];\n    const ii = hIdx[i];\n    let j = i - 1;\n    while (j >= 0 && hVal[j] < vi) {\n      hVal[j + 1] = hVal[j];\n      hIdx[j + 1] = hIdx[j];\n      j--;\n    }\n    hVal[j + 1] = vi;\n    hIdx[j + 1] = ii;\n  }\n\n  // Softmax over top-K candidates\n  const maxScore = hVal[0];\n  let sumExp = 0;\n  for (let i = 0; i < heapSize; i++) {\n    const p = Math.exp(hVal[i] - maxScore);\n    hVal[i] = p; // Reuse hVal for probabilities\n    sumExp += p;\n  }\n  const invSum = 1 / sumExp;\n  for (let i = 0; i < heapSize; i++) {\n    hVal[i] *= invSum;\n  }\n\n  // Apply top-p (nucleus): keep smallest set whose cumulative prob >= topP\n  let candidateCount = heapSize;\n  if (topP < 1.0) {\n    let cumulative = 0;\n    for (let i = 0; i < heapSize; i++) {\n      cumulative += hVal[i];\n      if (cumulative >= topP) {\n        candidateCount = i + 1;\n        break;\n      }\n    }\n    // Re-normalize\n    let sum = 0;\n    for (let i = 0; i < candidateCount; i++) sum += hVal[i];\n    const inv = 1 / sum;\n    for (let i = 0; i < candidateCount; i++) hVal[i] *= inv;\n  }\n\n  // Weighted random sample\n  const r = nextRandom(params);\n  let cumulative = 0;\n  for (let i = 0; i < candidateCount; i++) {\n    cumulative += hVal[i];\n    if (r <= cumulative) return hIdx[i];\n  }\n\n  return hIdx[candidateCount - 1];\n}\n\n/**\n * Return the index of the maximum value (greedy decoding).\n */\nexport function argmax(arr: Float32Array): number {\n  let maxIdx = 0;\n  let maxVal = arr[0];\n  for (let i = 1; i < arr.length; i++) {\n    if (arr[i] > maxVal) {\n      maxVal = arr[i];\n      maxIdx = i;\n    }\n  }\n  return maxIdx;\n}\n\n/** Min-heap sift down on parallel index/value typed arrays. */\nfunction siftDown(indices: Uint32Array, values: Float32Array, i: number, n: number): void {\n  while (true) {\n    let smallest = i;\n    const left = 2 * i + 1;\n    const right = 2 * i + 2;\n    if (left < n && values[left] < values[smallest]) smallest = left;\n    if (right < n && values[right] < values[smallest]) smallest = right;\n    if (smallest === i) break;\n    // Swap\n    const ti = indices[i];\n    indices[i] = indices[smallest];\n    indices[smallest] = ti;\n    const tv = values[i];\n    values[i] = values[smallest];\n    values[smallest] = tv;\n    i = smallest;\n  }\n}\n","/**\n * RequestScheduler — continuous-batching serving loop (Phase 5 of the\n * batching campaign, `GERBIL_BATCH=N`, Dawn only).\n *\n * Turns the Phase 2/4 static lockstep batch into a real serving loop:\n *\n *  - **Dynamic admission.** Requests queue at submit(); each iteration admits\n *    the next queued request into a FREE batch slot, bounded by the batch\n *    width AND KV-pool capacity (worst-case block need vs unreserved blocks).\n *    Admission control is the OOM defense — WebGPU has no memory\n *    introspection, so the scheduler never admits a request whose worst-case\n *    KV footprint could exhaust the block pool mid-flight.\n *  - **Lane retirement.** A sequence that hits EOS / max_tokens / a stop\n *    sequence is retired the moment its token is read: its KV blocks are\n *    freed, its slot is marked free for the next admit, and subsequent steps\n *    mask the lane (seq_lens = 0) — finished lanes stop being stepped instead\n *    of burning until the whole batch drains (the Phase 2 waste).\n *  - **Chunked prefill.** A newly admitted request's prompt is prefilled in\n *    bounded chunks (GERBIL_PREFILL_CHUNK tokens per iteration, default 32)\n *    on the single-sequence path INTO its slot's block-table row (tableBase),\n *    interleaved with masked decode steps for the lanes already decoding — a\n *    long prompt no longer stalls the batch. Only the final chunk reads\n *    logits back (the first sampled token); intermediate chunks are\n *    submit-only. Chunking never changes the math: same positions, same KV,\n *    same kernels — just split across iterations — so per-request greedy\n *    output is token-exact vs single-sequence generation.\n *\n * Decode steps are pipelined depth-2 like the single-sequence path; each\n * submitted step snapshots its active (slot → request) rows so tokens read\n * one step behind are delivered to the right request even when lanes were\n * retired or re-admitted in between. In-flight steps submitted before a\n * retirement still step the dead lane once — their stray KV append lands in\n * blocks that are either still owned by the lane or (if re-allocated) are\n * fully rewritten by the new owner's queue-ordered prefill before any read.\n *\n *  - **Per-request adapters** (`GERBIL_BATCH_LORA=1` on top of `GERBIL_BATCH`).\n *    `submit(prompt, { adapter: \"name\" })` decodes the request with a\n *    registered LoRA adapter (SGMV): its prefill runs under the single-sequence\n *    runtime overlay (like generateBatch's per-lane prefill) and every decode\n *    step re-derives the slot→adapter assignment from the CURRENTLY active\n *    lanes, so admissions, retirements, and slot reuse can never leak one\n *    request's adapter into another. Unregistered names fail at submit().\n *\n * Greedy-only (like generateBatch): sampling.temperature must be 0.\n */\n\nimport type { Executor } from \"../executor.js\";\nimport type { LoRADelta } from \"../lora.js\";\nimport { sampleToken } from \"../sampler.js\";\nimport type { ChatMessage } from \"../tokenizer.js\";\n\n/** Options for a single scheduled request. */\nexport interface ServeOptions {\n  /** Max new tokens (clamped to context; default 128). */\n  maxTokens?: number;\n  stopSequences?: string[];\n  systemPrompt?: string;\n  /** Greedy-only in Phase 5: temperature must be 0 (or omitted). */\n  sampling?: { temperature?: number };\n  /**\n   * Registered adapter name this request decodes with (SGMV — requires\n   * `GERBIL_BATCH_LORA=1` and a prior `registerAdapter` call on the engine);\n   * undefined = bare base. Requests with different adapters (or none) decode\n   * concurrently over the shared base, each token-exact vs a solo run with\n   * that adapter loaded.\n   */\n  adapter?: string;\n  /** Streamed per-token callback (decoded text piece). */\n  onToken?: (\n    token: string,\n    meta: { requestId: number; tokenId: number; generated: number },\n  ) => void;\n}\n\n/** Per-request result + serving timings (all performance.now() ms). */\nexport interface ServeResult {\n  requestId: number;\n  text: string;\n  tokenIds: number[];\n  tokensGenerated: number;\n  finishReason: \"eos\" | \"max_tokens\" | \"stop_sequence\";\n  /** The adapter name the request decoded with (echoed through), if any. */\n  adapter?: string;\n  /** submit() call time. */\n  submittedAt: number;\n  /** Admission into a batch slot (prefill start). */\n  admittedAt: number;\n  /** First token available (end of prefill). */\n  firstTokenAt: number;\n  finishedAt: number;\n  /** Time-to-first-token from submission (queue wait + prefill). */\n  ttftMs: number;\n  e2eMs: number;\n}\n\n/** Wiring the engine hands the scheduler (see WebGPUEngine.createScheduler). */\nexport interface SchedulerHost {\n  executor: Executor;\n  encodeChat(messages: ChatMessage[]): number[];\n  decodeToken(id: number): string;\n  eosTokenId: number | null;\n  eotTokenId: number | null;\n  /** Vocab-prune remap for CPU-sampled tokens (identity when unpruned). */\n  remapToken(id: number): number;\n  /**\n   * Resolve a registered batch-adapter name to its executor adapter id +\n   * parsed deltas. Must throw a clear error for unregistered names (surfaced\n   * at submit() time, before the request is queued).\n   */\n  resolveAdapter(name: string): { id: number; deltas: LoRADelta[] };\n  /** Serialize against generate()/embed() on the same engine. */\n  acquireLock(): Promise<() => void>;\n}\n\ninterface Request {\n  id: number;\n  promptIds: Uint32Array;\n  maxTokens: number;\n  stopSequences: string[];\n  /** Registered adapter name, or null for the bare base. */\n  adapterName: string | null;\n  /** Executor adapter id for decode steps (-1 = bare base). */\n  adapterId: number;\n  /** Adapter deltas applied as the runtime overlay during this request's prefill. */\n  adapterDeltas: LoRADelta[] | null;\n  onToken?: ServeOptions[\"onToken\"];\n  resolve: (r: ServeResult) => void;\n  reject: (e: Error) => void;\n  state: \"queued\" | \"prefilling\" | \"decoding\" | \"done\";\n  slot: number;\n  /** Prompt tokens prefilled so far (chunked prefill cursor). */\n  prefillPos: number;\n  /** Worst-case KV blocks reserved at admission. */\n  blocksReserved: number;\n  generatedIds: number[];\n  text: string;\n  finishReason: ServeResult[\"finishReason\"];\n  submittedAt: number;\n  admittedAt: number;\n  firstTokenAt: number;\n}\n\ninterface InFlightStep {\n  readbackSlot: number;\n  rows: Array<{ slot: number; req: Request }>;\n}\n\nconst PIPELINE_DEPTH = 2;\n\nexport interface SchedulerStats {\n  submitted: number;\n  completed: number;\n  stepsSubmitted: number;\n  prefillChunks: number;\n  tokensGenerated: number;\n  /** Wall-clock ms the serving loop was live (work outstanding). */\n  busyMs: number;\n  peakActiveLanes: number;\n}\n\nexport class RequestScheduler {\n  private readonly host: SchedulerHost;\n  private readonly batchSize: number;\n  private readonly maxSeqLen: number;\n  private readonly blockSize: number;\n  private readonly prefillChunk: number;\n  /** Free KV blocks kept unreserved as safety headroom. */\n  private readonly blockHeadroom: number;\n\n  private nextId = 1;\n  private queue: Request[] = [];\n  /** slot → occupying request (prefilling or decoding). */\n  private slots: Array<Request | null>;\n  /** Sum of blocksReserved across live (admitted, unfinished) requests. */\n  private reservedBlocks = 0;\n  private prefilling: Request | null = null;\n  private inFlight: InFlightStep[] = [];\n  /**\n   * The slot→adapter-id assignment last uploaded to the executor (null =\n   * nothing uploaded this batch session). Steps only re-upload when the\n   * assignment changes (admissions/retirements), and the very first upload is\n   * skipped entirely while every lane is bare base — an adapter-free serving\n   * session never touches the LoRA machinery.\n   */\n  private lastLaneAdapterIds: Int32Array | null = null;\n  private stepCounter = 0;\n  private loop: Promise<void> | null = null;\n  private closed = false;\n\n  private stats: SchedulerStats = {\n    submitted: 0,\n    completed: 0,\n    stepsSubmitted: 0,\n    prefillChunks: 0,\n    tokensGenerated: 0,\n    busyMs: 0,\n    peakActiveLanes: 0,\n  };\n\n  constructor(host: SchedulerHost, options: { prefillChunkTokens?: number } = {}) {\n    this.host = host;\n    this.batchSize = host.executor.batchSize;\n    if (this.batchSize <= 0) {\n      throw new Error(\"RequestScheduler requires GERBIL_BATCH=N (N >= 2) on the Dawn path\");\n    }\n    this.maxSeqLen = host.executor.maxSequenceLength;\n    this.blockSize = host.executor.kvBlockSizeTokens;\n    this.slots = new Array(this.batchSize).fill(null);\n    const envChunk =\n      typeof process !== \"undefined\" ? Number(process.env?.GERBIL_PREFILL_CHUNK) : Number.NaN;\n    // 128 balances per-forward fixed costs (small chunks tax prefill ~40% on\n    // ~100-token prompts — measured at load) against the decode stall a chunk\n    // inserts (~one 128-token prefill between decode steps, still bounded).\n    this.prefillChunk =\n      options.prefillChunkTokens ?? (Number.isFinite(envChunk) && envChunk >= 8 ? envChunk : 128);\n    this.blockHeadroom = 2;\n  }\n\n  /**\n   * Submit a request; resolves with the full result when it finishes.\n   * Tokens stream through options.onToken as they are decoded.\n   */\n  submit(prompt: string | ChatMessage[], options: ServeOptions = {}): Promise<ServeResult> {\n    if (this.closed) {\n      return Promise.reject(new Error(\"RequestScheduler is closed\"));\n    }\n    if ((options.sampling?.temperature ?? 0) >= 1e-6) {\n      return Promise.reject(\n        new Error(\"RequestScheduler: Phase 5 is greedy-only — pass sampling: { temperature: 0 }\"),\n      );\n    }\n    // Resolve the adapter NOW so an unregistered name fails the caller with a\n    // clear error instead of poisoning the serving loop later.\n    let adapterReg: { id: number; deltas: LoRADelta[] } | null = null;\n    if (options.adapter != null) {\n      if (!this.host.executor.batchLoraEnabled) {\n        return Promise.reject(\n          new Error(\n            \"RequestScheduler: per-request adapters require GERBIL_BATCH_LORA=1 (with GERBIL_BATCH=N)\",\n          ),\n        );\n      }\n      try {\n        adapterReg = this.host.resolveAdapter(options.adapter);\n      } catch (err) {\n        return Promise.reject(err instanceof Error ? err : new Error(String(err)));\n      }\n    }\n    const messages: ChatMessage[] =\n      typeof prompt === \"string\"\n        ? [\n            ...(options.systemPrompt\n              ? [{ role: \"system\" as const, content: options.systemPrompt }]\n              : []),\n            { role: \"user\" as const, content: prompt },\n          ]\n        : options.systemPrompt\n          ? [{ role: \"system\" as const, content: options.systemPrompt }, ...prompt]\n          : prompt;\n    const promptIds = new Uint32Array(this.host.encodeChat(messages));\n    // Clamp so the sequence (+ up to PIPELINE_DEPTH overshoot tokens from\n    // steps already in flight when the lane finishes) always fits the context.\n    const roomFor = this.maxSeqLen - promptIds.length - PIPELINE_DEPTH - 1;\n    if (roomFor < 1) {\n      return Promise.reject(\n        new Error(\n          `RequestScheduler: prompt (${promptIds.length} tokens) leaves no room in maxSeqLen ${this.maxSeqLen}`,\n        ),\n      );\n    }\n    const maxTokens = Math.min(options.maxTokens ?? 128, roomFor);\n\n    return new Promise<ServeResult>((resolve, reject) => {\n      const req: Request = {\n        id: this.nextId++,\n        promptIds,\n        maxTokens,\n        stopSequences: options.stopSequences ?? [],\n        adapterName: options.adapter ?? null,\n        adapterId: adapterReg?.id ?? -1,\n        adapterDeltas: adapterReg?.deltas ?? null,\n        onToken: options.onToken,\n        resolve,\n        reject,\n        state: \"queued\",\n        slot: -1,\n        prefillPos: 0,\n        blocksReserved: 0,\n        generatedIds: [],\n        text: \"\",\n        finishReason: \"max_tokens\",\n        submittedAt: performance.now(),\n        admittedAt: 0,\n        firstTokenAt: 0,\n      };\n      this.queue.push(req);\n      this.stats.submitted++;\n      this.pump();\n    });\n  }\n\n  /** Resolves when every submitted request has completed. */\n  async drain(): Promise<void> {\n    while (this.loop) await this.loop;\n  }\n\n  /** Stop accepting requests; resolves when in-flight work drains. */\n  async close(): Promise<void> {\n    this.closed = true;\n    await this.drain();\n  }\n\n  getStats(): SchedulerStats {\n    return { ...this.stats };\n  }\n\n  private pump(): void {\n    if (this.loop) return;\n    this.loop = this.run().finally(() => {\n      this.loop = null;\n      // Work may have arrived while the loop was tearing down.\n      if (this.hasWork()) this.pump();\n    });\n  }\n\n  private hasWork(): boolean {\n    return (\n      this.queue.length > 0 ||\n      this.prefilling !== null ||\n      this.inFlight.length > 0 ||\n      this.slots.some((s) => s !== null)\n    );\n  }\n\n  private activeMask(): { mask: Uint8Array; rows: Array<{ slot: number; req: Request }> } {\n    const mask = new Uint8Array(this.batchSize);\n    const rows: Array<{ slot: number; req: Request }> = [];\n    for (let b = 0; b < this.batchSize; b++) {\n      const req = this.slots[b];\n      if (req && req.state === \"decoding\") {\n        mask[b] = 1;\n        rows.push({ slot: b, req });\n      }\n    }\n    return { mask, rows };\n  }\n\n  /** The serving loop. Holds the engine generation lock while live. */\n  private async run(): Promise<void> {\n    const release = await this.host.acquireLock();\n    const ex = this.host.executor;\n    const loopStart = performance.now();\n    // Fresh batch session: every slot/block free (nothing can be live here —\n    // the loop only exits when fully drained).\n    ex.resetBatch();\n    this.stepCounter = 0;\n    this.lastLaneAdapterIds = null;\n\n    try {\n      while (this.hasWork()) {\n        this.admit();\n\n        // Submit a masked decode step for the decoding lanes (pipelined).\n        const { mask, rows } = this.activeMask();\n        if (rows.length > 0 && this.inFlight.length < PIPELINE_DEPTH) {\n          const readbackSlot = this.stepCounter % PIPELINE_DEPTH;\n          this.syncLaneAdapters();\n          ex.submitBatchDecodeStepMasked(mask, readbackSlot);\n          this.stepCounter++;\n          this.inFlight.push({ readbackSlot, rows });\n          this.stats.stepsSubmitted++;\n          if (rows.length > this.stats.peakActiveLanes) {\n            this.stats.peakActiveLanes = rows.length;\n          }\n        }\n\n        // One chunk of the pending prefill, overlapping the step above.\n        if (this.prefilling) await this.prefillChunkStep();\n\n        // Read the oldest in-flight step once the pipeline is full (depth-2:\n        // step k+1 runs on the GPU while k's tokens are processed), or when\n        // no step could be submitted and one is outstanding.\n        const pipelineFull = this.inFlight.length >= PIPELINE_DEPTH;\n        const idleLanes = rows.length === 0 && this.prefilling === null;\n        if (this.inFlight.length > 0 && (pipelineFull || idleLanes)) {\n          await this.readStep();\n        }\n      }\n    } catch (err) {\n      // Fail every live request rather than hanging their promises.\n      const error = err instanceof Error ? err : new Error(String(err));\n      if (this.prefilling?.adapterDeltas) {\n        try {\n          ex.clearRuntimeLoRA();\n        } catch {\n          // best-effort — the overlay must not outlive its request's prefill\n        }\n      }\n      for (const req of [...this.queue, ...this.slots.filter((s): s is Request => s !== null)]) {\n        if (req.state !== \"done\") {\n          req.state = \"done\";\n          req.reject(error);\n        }\n      }\n      this.queue = [];\n      this.slots.fill(null);\n      this.prefilling = null;\n      this.inFlight = [];\n      this.reservedBlocks = 0;\n      throw error;\n    } finally {\n      this.stats.busyMs += performance.now() - loopStart;\n      release();\n    }\n  }\n\n  /**\n   * Upload the slot→adapter assignment for the step about to be submitted,\n   * when it differs from the last uploaded one. Called immediately before\n   * every decode-step submit: `queue.writeBuffer` is queue-ordered, so steps\n   * already in flight keep the assignment they were submitted under, and a\n   * retiring lane's slot can never leak its adapter into the slot's next\n   * occupant — the next occupant's assignment is re-derived from `slots`\n   * before its first decode step. Masked (free/prefilling) lanes are pinned\n   * to bare base; their output is discarded regardless.\n   */\n  private syncLaneAdapters(): void {\n    if (!this.host.executor.batchLoraEnabled) return;\n    const ids = new Int32Array(this.batchSize).fill(-1);\n    let anyAdapter = false;\n    for (let b = 0; b < this.batchSize; b++) {\n      const req = this.slots[b];\n      if (req && req.state === \"decoding\" && req.adapterId >= 0) {\n        ids[b] = req.adapterId;\n        anyAdapter = true;\n      }\n    }\n    const last = this.lastLaneAdapterIds;\n    if (last === null) {\n      // Nothing uploaded yet this session: an all-base assignment needs no\n      // upload at all (the plain batch entry list is already in effect).\n      if (!anyAdapter) return;\n    } else if (ids.every((id, b) => id === last[b])) {\n      return;\n    }\n    this.host.executor.setBatchLaneAdapters(ids);\n    this.lastLaneAdapterIds = ids;\n  }\n\n  /**\n   * Admission: move the next queued request into a free slot when (a) a slot\n   * is free, (b) no other prefill is staged (the single-sequence prefill path\n   * stages SSM state in singleton buffers — one at a time), and (c) its\n   * worst-case block need fits the unreserved KV pool with headroom.\n   */\n  private admit(): void {\n    if (this.prefilling || this.queue.length === 0) return;\n    const slot = this.slots.indexOf(null);\n    if (slot === -1) return;\n\n    const req = this.queue[0];\n    const worstTokens = req.promptIds.length + req.maxTokens + PIPELINE_DEPTH;\n    const blocksNeeded = Math.ceil(worstTokens / this.blockSize);\n    const totalBlocks = this.host.executor.kvTotalBlocks;\n    if (this.reservedBlocks + blocksNeeded + this.blockHeadroom > totalBlocks) {\n      // Not enough KV capacity for this request's worst case — hold it (and\n      // everything behind it: FIFO) until retirements free blocks.\n      return;\n    }\n\n    this.queue.shift();\n    req.state = \"prefilling\";\n    req.slot = slot;\n    req.blocksReserved = blocksNeeded;\n    req.admittedAt = performance.now();\n    this.reservedBlocks += blocksNeeded;\n    this.slots[slot] = req;\n    this.prefilling = req;\n    // Adapter requests prefill with the proven single-sequence runtime\n    // overlay (exactly how generateBatch's per-lane prefill applies it), so\n    // the KV/SSM state and first sampled token match a solo run with that\n    // adapter loaded. The overlay only swaps the single-sequence entry lists;\n    // the interleaved batched decode steps use the batch entry lists and are\n    // untouched. Cleared after the final chunk (or on loop failure).\n    if (req.adapterDeltas) {\n      const { applied } = this.host.executor.applyRuntimeLoRA(req.adapterDeltas);\n      if (applied === 0) {\n        throw new Error(\n          `RequestScheduler: adapter \"${req.adapterName}\" resolved to 0 applicable prefill targets`,\n        );\n      }\n    }\n    this.host.executor.beginBatchSlot(slot);\n  }\n\n  /**\n   * Run one bounded prefill chunk for the staged request on the\n   * single-sequence path (into its slot's block-table row via tableBase).\n   * Intermediate chunks are submit-only; the final chunk reads logits, adopts\n   * the singleton SSM state into the slot's pool row, seeds the lane's first\n   * token, and flips the lane to decoding.\n   */\n  private async prefillChunkStep(): Promise<void> {\n    const req = this.prefilling;\n    if (!req) return;\n    const ex = this.host.executor;\n    const end = Math.min(req.prefillPos + this.prefillChunk, req.promptIds.length);\n    const isLast = end === req.promptIds.length;\n    const chunk = req.promptIds.subarray(req.prefillPos, end);\n    const { logits } = await ex.forward(chunk, { readLogits: isLast });\n    req.prefillPos = end;\n    this.stats.prefillChunks++;\n    if (!isLast) return;\n\n    ex.finishBatchSlot(req.slot);\n    if (req.adapterDeltas) ex.clearRuntimeLoRA();\n    this.prefilling = null;\n    const first = this.host.remapToken(sampleToken(logits, { temperature: 0 }));\n    ex.injectBatchToken(req.slot, first);\n    req.state = \"decoding\";\n    req.firstTokenAt = performance.now();\n    this.consumeToken(req, first);\n  }\n\n  /** Read the oldest in-flight step and deliver its tokens. */\n  private async readStep(): Promise<void> {\n    const step = this.inFlight.shift();\n    if (!step) return;\n    const tokens = await this.host.executor.readBatchTokens(step.readbackSlot);\n    for (const { slot, req } of step.rows) {\n      // The lane may have been retired (and even re-admitted) after this step\n      // was submitted — deliver only to the request that was stepped.\n      if (this.slots[slot] !== req || req.state !== \"decoding\") continue;\n      this.consumeToken(req, tokens[slot]);\n    }\n  }\n\n  /** Append one generated token; finish + retire the lane when terminal. */\n  private consumeToken(req: Request, tokenId: number): void {\n    req.generatedIds.push(tokenId);\n    this.stats.tokensGenerated++;\n    if (req.firstTokenAt === 0) req.firstTokenAt = performance.now();\n\n    const { eosTokenId, eotTokenId } = this.host;\n    if (\n      (eosTokenId !== null && tokenId === eosTokenId) ||\n      (eotTokenId !== null && tokenId === eotTokenId)\n    ) {\n      req.finishReason = \"eos\";\n      this.finish(req);\n      return;\n    }\n    const piece = this.host.decodeToken(tokenId);\n    req.text += piece;\n    if (req.stopSequences.some((s) => req.text.includes(s))) {\n      for (const s of req.stopSequences) {\n        const idx = req.text.indexOf(s);\n        if (idx !== -1) req.text = req.text.slice(0, idx);\n      }\n      req.finishReason = \"stop_sequence\";\n      this.finish(req);\n      return;\n    }\n    req.onToken?.(piece, {\n      requestId: req.id,\n      tokenId,\n      generated: req.generatedIds.length,\n    });\n    if (req.generatedIds.length >= req.maxTokens) {\n      req.finishReason = \"max_tokens\";\n      this.finish(req);\n    }\n  }\n\n  /** Retire the lane immediately: free KV blocks, open the slot, resolve. */\n  private finish(req: Request): void {\n    req.state = \"done\";\n    if (req.slot >= 0 && this.slots[req.slot] === req) {\n      this.host.executor.retireBatchSlot(req.slot);\n      this.slots[req.slot] = null;\n    }\n    this.reservedBlocks -= req.blocksReserved;\n    this.stats.completed++;\n    const finishedAt = performance.now();\n    req.resolve({\n      requestId: req.id,\n      text: req.text,\n      tokenIds: [...req.generatedIds],\n      tokensGenerated: req.generatedIds.length,\n      finishReason: req.finishReason,\n      adapter: req.adapterName ?? undefined,\n      submittedAt: req.submittedAt,\n      admittedAt: req.admittedAt,\n      firstTokenAt: req.firstTokenAt,\n      finishedAt,\n      ttftMs: req.firstTokenAt - req.submittedAt,\n      e2eMs: finishedAt - req.submittedAt,\n    });\n  }\n}\n","/**\n * Swarm — the MoTA (Mixture of Tuned Adapters) mode-B primitive\n * (docs/research/slm-swarms.md §8): a set of named members (LoRA adapters on\n * one shared base, or the bare base itself), a plan that fans one input out\n * into per-member sub-tasks, and a reduce that joins the results.\n *\n * `swarm.run(input)` submits every planned sub-task to the engine's\n * RequestScheduler with that member's per-request adapter — all sub-tasks\n * decode CONCURRENTLY in the same batched pass over the shared base, so the\n * fan-out costs roughly the slowest lane, not the sum of lanes. The\n * orchestration itself stays user-supplied and thin: default plan broadcasts\n * the input to every member; default reduce returns a member-keyed map of\n * results.\n *\n * Requires `GERBIL_BATCH=N` + `GERBIL_BATCH_LORA=1` (Node/Dawn) — created via\n * `WebGPUEngine.createSwarm()`, which registers every member's adapter first.\n * Greedy-only, like the scheduler underneath.\n */\n\nimport type { ChatMessage } from \"../tokenizer.js\";\nimport type { RequestScheduler, ServeOptions, ServeResult } from \"./scheduler.js\";\n\n/** Per-request options a swarm sub-task may set (the member picks the adapter). */\nexport type SwarmTaskOptions = Omit<ServeOptions, \"adapter\">;\n\n/** One planned sub-task: which member decodes which prompt. */\nexport interface SwarmTask {\n  /** Member name (a key of the swarm's `members`). */\n  member: string;\n  prompt: string | ChatMessage[];\n  /** Per-task overrides merged over the run's options. */\n  options?: SwarmTaskOptions;\n}\n\n/** One member's completed sub-task (plan order is preserved). */\nexport interface SwarmMemberResult {\n  member: string;\n  result: ServeResult;\n}\n\n/**\n * Swarm configuration.\n *\n * @template R The reduced result type (defaults to the member-keyed map the\n *   default reduce produces).\n */\nexport interface SwarmOptions<R = Record<string, ServeResult>> {\n  /**\n   * Member name → adapter source (`hf:owner/repo`, URL, or `file:./dir`), or\n   * null for a member that decodes on the bare base. Every non-null source is\n   * registered on the engine (factors packed into shared GPU buffers once).\n   */\n  members: Record<string, string | null>;\n  /**\n   * input → sub-tasks routed to members. Default: broadcast the input to\n   * every member (one sub-task each, declaration order).\n   */\n  plan?: (input: string, members: string[]) => SwarmTask[] | Promise<SwarmTask[]>;\n  /**\n   * Joined sub-results → final answer. Default: a member-keyed map of\n   * ServeResults (a duplicated member's later tasks get `name#2`, `name#3`…).\n   */\n  reduce?: (results: SwarmMemberResult[]) => R | Promise<R>;\n}\n\n/** Default reduce: member-keyed map, disambiguating duplicated members. */\nfunction reduceToKeyedMap(results: SwarmMemberResult[]): Record<string, ServeResult> {\n  const out: Record<string, ServeResult> = {};\n  const seen = new Map<string, number>();\n  for (const { member, result } of results) {\n    const n = (seen.get(member) ?? 0) + 1;\n    seen.set(member, n);\n    out[n === 1 ? member : `${member}#${n}`] = result;\n  }\n  return out;\n}\n\n/**\n * A swarm handle: registered members + plan/reduce over the engine's\n * scheduler. Create via {@link WebGPUEngine.createSwarm}.\n */\nexport class Swarm<R = Record<string, ServeResult>> {\n  private readonly scheduler: RequestScheduler;\n  private readonly memberSources: Record<string, string | null>;\n  private readonly plan: SwarmOptions<R>[\"plan\"];\n  private readonly reduce: SwarmOptions<R>[\"reduce\"];\n\n  constructor(scheduler: RequestScheduler, options: SwarmOptions<R>) {\n    if (Object.keys(options.members).length === 0) {\n      throw new Error(\"Swarm: members must contain at least one member\");\n    }\n    this.scheduler = scheduler;\n    this.memberSources = { ...options.members };\n    this.plan = options.plan;\n    this.reduce = options.reduce;\n  }\n\n  /** Member names, declaration order. */\n  get members(): string[] {\n    return Object.keys(this.memberSources);\n  }\n\n  /**\n   * Run one input through the swarm: plan → concurrent fan-out (one batched\n   * pass across the scheduler's lanes) → reduce.\n   *\n   * @param input The task input handed to the plan (default plan: broadcast\n   *   it to every member as the prompt).\n   * @param options Per-request options applied to every sub-task (a task's\n   *   own `options` win field-by-field).\n   */\n  async run(input: string, options: SwarmTaskOptions = {}): Promise<R> {\n    const tasks = this.plan\n      ? await this.plan(input, this.members)\n      : this.members.map((member): SwarmTask => ({ member, prompt: input }));\n    if (tasks.length === 0) {\n      throw new Error(\"Swarm: plan produced no sub-tasks\");\n    }\n    const results = await Promise.all(\n      tasks.map(async (task): Promise<SwarmMemberResult> => {\n        const result = await this.submitFor(task.member, task.prompt, {\n          ...options,\n          ...task.options,\n        });\n        return { member: task.member, result };\n      }),\n    );\n    if (this.reduce) return this.reduce(results);\n    // No custom reduce ⇒ R is the default map type.\n    return reduceToKeyedMap(results) as R;\n  }\n\n  /**\n   * Mode-A access: decode one prompt with a single member (no plan/reduce).\n   * Still rides the shared scheduler, so it batches with any concurrent work.\n   */\n  generate(\n    member: string,\n    prompt: string | ChatMessage[],\n    options: SwarmTaskOptions = {},\n  ): Promise<ServeResult> {\n    return this.submitFor(member, prompt, options);\n  }\n\n  private submitFor(\n    member: string,\n    prompt: string | ChatMessage[],\n    options: SwarmTaskOptions,\n  ): Promise<ServeResult> {\n    if (!Object.hasOwn(this.memberSources, member)) {\n      throw new Error(`Swarm: unknown member \"${member}\" (members: ${this.members.join(\", \")})`);\n    }\n    // Null source = bare-base member: submit without an adapter.\n    const adapter = this.memberSources[member] == null ? undefined : member;\n    return this.scheduler.submit(prompt, { ...options, adapter });\n  }\n}\n","/**\n * Gepard-1.0 — native streaming TTS architecture (nineninesix/gepard-1.0, Apache-2.0).\n *\n * Gepard is a frame-synchronous multihead codec-LM TTS:\n *   1. BACKBONE: a stock Qwen3.5 transformer body (Qwen3_5ForCausalLM export\n *      `nineninesix/qwen3_5-full-attn-only-14`) whose 14 layers are ALL\n *      `full_attention` — a strict subset of the hybrid graph qwen3_5.ts builds\n *      (same fused-q/gate attention, QK norms, sigmoid output gate, SwiGLU MLP).\n *      hidden 1024, 8 Q heads / 2 KV heads, head_dim 256, rope_theta 1e7,\n *      partial_rotary_factor 1.0 (full-head rotation — the nested rope_parameters\n *      copy is authoritative per the reference implementation), vocab 248320\n *      (tied embeddings; the lm_head is discarded — text generation is not a task).\n *   2. OVERLAY (host-side, tiny): 32 parallel classifier heads\n *      (`codebook_heads.{0..31}`, cardinalities [8,7,6,6] repeating → 216 logits\n *      total) emit ONE WHOLE AUDIO FRAME per forward pass, plus a binary\n *      `stop_head` (end-of-speech). The sampled 32 FSQ codes feed back through\n *      per-channel embeddings (`audio_embeddings.{i}` [L_i, 32]) → concat(1024) →\n *      Linear→GELU→Linear→affine-free LayerNorm → × audio_embed_scale into the\n *      next step's input embedding. All overlay math runs on the host (≈2 M MACs\n *      per frame — negligible next to the backbone forward).\n *   3. CODEC: NVIDIA NeMo NanoCodec 22 kHz 1.89 kbps 21.5 fps\n *      (nvidia/nemo-nano-codec-22khz-1.89kbps-21.5fps): GroupFSQ with 8 groups ×\n *      4 dims (levels [8,7,6,6], 2016 codes/group) + the SAME CausalHiFiGANDecoder\n *      family the validated Kani NanoCodec graph implements — only the geometry\n *      differs (up_sample_rates [8,8,4,2,2] → hop 1024, input_dim 32). The model\n *      emits per-dimension codes directly, so dequantization is a trivial host-side\n *      `(code − L//2)/(L//2)` and the decoder graph takes the latent as input.\n *\n * ── STEP-0 verification (live HF, July 2026) ─────────────────────────────────\n * config.json:  Qwen3_5ForCausalLM, 14×full_attention, hidden 1024, GQA 8/2,\n *   head_dim 256, intermediate 3584, attn_output_gate true, tie_word_embeddings,\n *   vocab 248320, rope_parameters { rope_theta 1e7, partial_rotary_factor 1.0,\n *   mrope_interleaved [11,11,10] } — with sequential text positions the three\n *   mRoPE position streams are equal, so it reduces exactly to 1-D RoPE.\n * gepard_config.json: audio_heads level_audio_0..31 = [8,7,6,6]×8,\n *   audio_embed_dim 32, special_tokens { start_of_text 248073, end_of_text 248074,\n *   start_of_speech 248070, tts_pad 248076 }, text_repetition { enabled, target 16,\n *   apply_below 13, max_repeats 8 }, codec { nvidia/nemo-nano-codec-22khz-1.89kbps-\n *   21.5fps, 8 layers, fsq_levels [8,7,6,6], 22050 Hz }.\n * model.safetensors: model.* = stock Qwen3.5 keys (q_proj fused [4096,1024] —\n *   per-head interleaved Q|gate, deinterleaved by the loader like the main Qwen3.5\n *   path; RMSNorm weights zero-centered → +1 offset applied at load). Overlay keys:\n *   audio_embeddings.{i}.weight [L_i,32], audio_embed_proj.{0,2}.{weight,bias},\n *   audio_embed_scale (scalar), codebook_heads.{i}.{weight,bias}, stop_head.*.\n *   Voice-cloning stack (ref_compressor.*, null_prefix, supcon_head.*) is present\n *   in the checkpoint but NOT wired here (deferred; default voice ships first).\n *\n * Generation (mirrors gepard-inference runner.py):\n *   text ids → adaptive repetition layout [ (SOT text EOT)×(R−1) | SOT text EOT SOS ]\n *   → prefill → hidden state at SOS seeds frame 0 → frame loop: embed previous\n *   frame → one cached decode step → stop head (sigmoid > threshold) → sample all\n *   32 heads independently (temperature/top-k/repetition penalty, fp32) → …\n *   → per-channel codes [32, T] → host dequant → NanoCodec decode → 22.05 kHz PCM.\n *   Positions are plain sequential (no frame-grouped position ids).\n *\n * The backbone graph takes a host-written `inputs_embeds` [T, 1024] activation\n * instead of `input_ids` (text rows are gathered host-side from the bf16 embed\n * table; audio rows come from the overlay MLP), so the ~1 GB embedding matrix\n * never touches the GPU. Its single output is the LAST position's final-norm\n * hidden state — exposed under the tensor name \"logits\" with config.vocab_size =\n * hidden_size so Executor.forward()'s standard logits readback returns it as-is.\n */\n\nimport type {\n  DType,\n  ModelArchConfig,\n  ModelCapabilities,\n  ModelGraph,\n  OpNode,\n  TensorDesc,\n} from \"../ir.js\";\nimport { CANONICAL_KEYS } from \"../ir.js\";\nimport { DEFAULT_GROUP_SIZE } from \"../quantize.js\";\nimport type { GraphDType, KVDType } from \"./index.js\";\n\n// ── Token / codec layout constants (verified against gepard_config.json) ──\nexport const GEPARD_START_OF_TEXT = 248_073;\nexport const GEPARD_END_OF_TEXT = 248_074;\nexport const GEPARD_START_OF_SPEECH = 248_070;\nexport const GEPARD_END_OF_SPEECH = 248_071;\nexport const GEPARD_TTS_PAD = 248_076;\n\n/** FSQ levels per dimension within each codec group (repeats per group). */\nexport const GEPARD_FSQ_LEVEL_PATTERN = [8, 7, 6, 6] as const;\n/** Number of codec groups (codebooks). */\nexport const GEPARD_NUM_GROUPS = 8;\n/** Total per-dimension channels per frame = groups × dims/group. */\nexport const GEPARD_NUM_CHANNELS = 32;\n/** NanoCodec 21.5 fps decoder geometry (codec_config.yaml, mlx-community mirror). */\nexport const GEPARD_CODEC_UP_SAMPLE_RATES = [8, 8, 4, 2, 2] as const;\nexport const GEPARD_CODEC_INPUT_DIM = 32;\n/** PCM samples per audio frame (product of the up-sample rates). */\nexport const GEPARD_HOP = 1024;\nexport const GEPARD_SAMPLE_RATE = 22_050;\n\n/** Per-channel FSQ cardinalities in head order (channel c = group⌊c/4⌋, dim c%4). */\nexport function gepardChannelLevels(): number[] {\n  const out: number[] = [];\n  for (let g = 0; g < GEPARD_NUM_GROUPS; g++) {\n    out.push(...GEPARD_FSQ_LEVEL_PATTERN);\n  }\n  return out;\n}\n\n/** Text-repetition layout config (stamped into gepard_config.json). */\nexport interface GepardTextRepetition {\n  enabled: boolean;\n  targetTextTokens: number;\n  applyBelow: number;\n  maxRepeats: number;\n}\n\n/** Parsed gepard_config.json — everything the host driver needs. */\nexport interface GepardConfig {\n  /** The nested Qwen3.5 backbone config (fed to generateGepardBackboneGraph). */\n  backboneConfig: Record<string, unknown>;\n  /** Per-head vocab sizes in channel order (level_audio_0..31). */\n  vocabSizes: number[];\n  audioEmbedDim: number;\n  startOfText: number;\n  endOfText: number;\n  startOfSpeech: number;\n  textRepetition: GepardTextRepetition;\n  /** Codec identity + geometry from the config's codec block. */\n  codecId: string;\n  fsqLevels: number[];\n  sampleRate: number;\n}\n\n/** Matches `level_audio_{i}` head names (channel index = numeric suffix). */\nconst LEVEL_AUDIO_RE = /^level_audio_(\\d+)$/;\n\n/**\n * Parse gepard_config.json. Head order is recovered from the numeric suffix of\n * each `level_audio_{i}` key (NOT the serialized key order), so the wiring is\n * immune to any JSON round-trip that re-sorts keys.\n */\nexport function parseGepardConfig(raw: Record<string, unknown>): GepardConfig {\n  const backboneConfig = raw.backbone_config as Record<string, unknown> | undefined;\n  if (!backboneConfig) {\n    throw new Error(\"gepard_config.json has no backbone_config — not a Gepard checkpoint.\");\n  }\n  const audioHeads = (raw.audio_heads as Record<string, number>) ?? {};\n  const byIndex: number[] = [];\n  for (const [name, size] of Object.entries(audioHeads)) {\n    const m = LEVEL_AUDIO_RE.exec(name);\n    if (!m) {\n      throw new Error(`Unexpected audio head name \"${name}\" in gepard_config.json.`);\n    }\n    byIndex[Number(m[1])] = size;\n  }\n  if (byIndex.length === 0) {\n    throw new Error(\"gepard_config.json has no audio_heads.\");\n  }\n  for (let i = 0; i < byIndex.length; i++) {\n    if (byIndex[i] == null) {\n      throw new Error(`gepard_config.json audio_heads is missing level_audio_${i}.`);\n    }\n  }\n  const special = (raw.special_tokens as Record<string, number>) ?? {};\n  const rep = (raw.text_repetition as Record<string, unknown>) ?? {};\n  const codec = (raw.codec as Record<string, unknown>) ?? {};\n  return {\n    backboneConfig,\n    vocabSizes: byIndex,\n    audioEmbedDim: (raw.audio_embed_dim as number) ?? 32,\n    startOfText: special.start_of_text ?? GEPARD_START_OF_TEXT,\n    endOfText: special.end_of_text ?? GEPARD_END_OF_TEXT,\n    startOfSpeech: special.start_of_speech ?? GEPARD_START_OF_SPEECH,\n    textRepetition: {\n      enabled: (rep.enabled as boolean) ?? false,\n      targetTextTokens: (rep.target_text_tokens as number) ?? 16,\n      applyBelow: (rep.apply_below as number) ?? 13,\n      maxRepeats: (rep.max_repeats as number) ?? 8,\n    },\n    codecId: (codec.codec_id as string) ?? \"nvidia/nemo-nano-codec-22khz-1.89kbps-21.5fps\",\n    fsqLevels: (codec.fsq_levels as number[]) ?? [...GEPARD_FSQ_LEVEL_PATTERN],\n    sampleRate: (codec.sample_rate as number) ?? GEPARD_SAMPLE_RATE,\n  };\n}\n\n/**\n * Build the (possibly repeated) text-id layout — mirrors the reference\n * TextRepeater byte-for-byte (text_repetition.py):\n *   R == 1 → [ SOT, *text, EOT, SOS ]\n *   R >  1 → [ SOT, *text, EOT ] × (R−1) + [ SOT, *text, EOT, SOS ]\n * Only the final (canonical) copy carries SOS. R is deterministic from the\n * text-token count: ceil(target / n) capped at maxRepeats, applied only when\n * n < applyBelow (short prompts derail without the extra text mass).\n */\nexport function buildGepardTextLayout(textTokenIds: number[], cfg: GepardConfig): number[] {\n  const rep = cfg.textRepetition;\n  const n = textTokenIds.length;\n  let repeats = 1;\n  if (rep.enabled && n > 0 && n < rep.applyBelow && n < rep.targetTextTokens) {\n    repeats = Math.max(1, Math.min(Math.ceil(rep.targetTextTokens / n), rep.maxRepeats));\n  }\n  const block = [cfg.startOfText, ...textTokenIds, cfg.endOfText];\n  const out: number[] = [];\n  for (let r = 0; r < repeats - 1; r++) {\n    out.push(...block);\n  }\n  out.push(...block, cfg.startOfSpeech);\n  return out;\n}\n\n/**\n * Dequantize per-channel FSQ codes [32, T] (channel-major) into the decoder\n * latent [32, T] f32: value = (code − L//2) / (L//2), matching\n * UnfoldedCodecModel.decode_from_codes / codec_ops.dequantize_codes.\n */\nexport function gepardCodesToLatent(codes: Uint32Array, numFrames: number): Float32Array {\n  const levels = gepardChannelLevels();\n  const latent = new Float32Array(GEPARD_NUM_CHANNELS * numFrames);\n  for (let c = 0; c < GEPARD_NUM_CHANNELS; c++) {\n    const scale = Math.max(1, Math.floor(levels[c] / 2));\n    const base = c * numFrames;\n    for (let t = 0; t < numFrames; t++) {\n      latent[base + t] = (codes[base + t] - scale) / scale;\n    }\n  }\n  return latent;\n}\n\n// ════════════════════════════════════════════════════════════════════════════\n// Backbone graph (stock Qwen3.5 full-attention body, embeds in → hidden out).\n//\n// Mirrors qwen3_5.ts's full-attention block exactly (fused Residual+RMSNorm\n// chaining included) with three deltas:\n//   • input is a host-written `inputs_embeds` [T, hidden] activation (the text\n//     rows are gathered from the bf16 embed table host-side; audio-frame rows\n//     come from the overlay MLP) — no Embedding op, no embed table on the GPU;\n//   • no Mamba branch (every Gepard layer is full_attention);\n//   • no lm_head — the output is SliceLastRow(final_norm) exposed as \"logits\"\n//     [1, hidden] with config.vocab_size = hidden_size, so Executor.forward()'s\n//     standard logits readback returns the final hidden state of the last\n//     position (which the host-side 32+1 heads consume).\n// ════════════════════════════════════════════════════════════════════════════\n\nexport function generateGepardBackboneGraph(\n  backboneConfig: Record<string, unknown>,\n  dtype?: GraphDType,\n  groupSizeOverride?: number,\n  kvDtype?: KVDType,\n): ModelGraph {\n  const hidden_size = backboneConfig.hidden_size as number;\n  const num_layers = backboneConfig.num_hidden_layers as number;\n  const num_heads = backboneConfig.num_attention_heads as number;\n  const num_kv_heads = (backboneConfig.num_key_value_heads as number) ?? num_heads;\n  const intermediate_size = backboneConfig.intermediate_size as number;\n  const context_length = (backboneConfig.max_position_embeddings as number) ?? 262_144;\n  const rms_norm_eps = (backboneConfig.rms_norm_eps as number) ?? 1e-6;\n  const head_dim = (backboneConfig.head_dim as number) ?? Math.floor(hidden_size / num_heads);\n\n  const q_dim = num_heads * head_dim;\n  const kv_dim = num_kv_heads * head_dim;\n\n  // RoPE: the nested rope_parameters copy is authoritative (the reference\n  // implementation forces partial_rotary_factor into it; Gepard ships 1.0 =\n  // full-head rotation). With sequential positions the interleaved mRoPE\n  // sections reduce exactly to 1-D RoPE, so the standard RoPE op applies.\n  const ropeParams = backboneConfig.rope_parameters as Record<string, unknown> | undefined;\n  const rope_base =\n    (ropeParams?.rope_theta as number) ?? (backboneConfig.rope_theta as number) ?? 10_000_000.0;\n  const partialRotaryFactor =\n    (ropeParams?.partial_rotary_factor as number) ??\n    (backboneConfig.partial_rotary_factor as number) ??\n    1.0;\n  const rope_dim = Math.floor(head_dim * partialRotaryFactor);\n\n  const layerTypes = (backboneConfig.layer_types as string[]) ?? [];\n  for (let i = 0; i < num_layers; i++) {\n    if (layerTypes.length > i && layerTypes[i] !== \"full_attention\") {\n      throw new Error(\n        `Gepard backbone expects all-full-attention layers; layer ${i} is \"${layerTypes[i]}\".`,\n      );\n    }\n  }\n\n  // vocab_size is set to hidden_size ON PURPOSE: the graph's \"logits\" output is\n  // the [1, hidden] final hidden state, and Executor.forward() reads back\n  // config.vocab_size floats from the tensor named \"logits\".\n  const config: ModelArchConfig = {\n    hidden_size,\n    num_layers,\n    num_heads,\n    num_kv_heads,\n    head_dim,\n    intermediate_size,\n    vocab_size: hidden_size,\n    context_length,\n    rms_norm_eps,\n    norm_type: \"rmsnorm\",\n    rope_base,\n    rope_dim,\n    kv_layout: \"LHSd\",\n    is_moe: false,\n    has_vision_tower: false,\n  };\n  const capabilities: ModelCapabilities = { text: true, vision: false, moe: false };\n\n  const tensors: Record<string, TensorDesc> = {};\n  const nodes: OpNode[] = [];\n  const executionOrder: string[] = [];\n  const addTensor = (desc: TensorDesc): void => {\n    tensors[desc.name] = desc;\n  };\n  const addNode = (node: OpNode): void => {\n    nodes.push(node);\n    executionOrder.push(node.id);\n  };\n\n  const useQ4 = dtype === \"q4\";\n  const groupSize = groupSizeOverride ?? DEFAULT_GROUP_SIZE;\n\n  const addLinearOp = (\n    id: string,\n    activationInput: string,\n    weightName: string,\n    weightShape: number[],\n    outputName: string,\n    K: number,\n    N: number,\n  ): void => {\n    if (useQ4) {\n      const wQ = `${weightName}.q`;\n      const wScales = `${weightName}.scales`;\n      const wZeros = `${weightName}.zeros`;\n      const totalElements = K * N;\n      addTensor({\n        name: wQ,\n        shape: [Math.ceil(totalElements / 8)],\n        dtype: \"u32\",\n        storage: \"constant\",\n      });\n      addTensor({\n        name: wScales,\n        shape: [Math.ceil(totalElements / groupSize)],\n        dtype: \"f32\",\n        storage: \"constant\",\n      });\n      addTensor({\n        name: wZeros,\n        shape: [Math.ceil(totalElements / groupSize)],\n        dtype: \"f32\",\n        storage: \"constant\",\n      });\n      addNode({\n        id,\n        opType: \"MatMulInt4\",\n        inputs: [activationInput, wQ, wScales, wZeros],\n        outputs: [outputName],\n        attributes: { M_tensor: activationInput, K, N, group_size: groupSize },\n      });\n      return;\n    }\n    addTensor({\n      name: weightName,\n      shape: weightShape,\n      dtype: \"f32\",\n      storage: \"constant\",\n      safetensorsKey: weightName,\n    });\n    addNode({\n      id,\n      opType: \"MatMul\",\n      inputs: [activationInput, weightName],\n      outputs: [outputName],\n      attributes: { M_tensor: activationInput, K, N },\n    });\n  };\n\n  // ── Input: host-written embeddings (read before written → persistent) ──\n  addTensor({\n    name: \"inputs_embeds\",\n    shape: [\"T\", hidden_size],\n    dtype: \"f32\",\n    storage: \"activation\",\n  });\n\n  // ── Full-attention block (identical math to qwen3_5.ts buildFullAttentionBlock) ──\n  const buildAttentionBlock = (i: number, prefix: string, norm1Out: string): string => {\n    const qOut = `${prefix}_q`;\n    const gateOut = `${prefix}_attn_gate`;\n    const kOut = `${prefix}_k`;\n    const vOut = `${prefix}_v`;\n    addTensor({ name: qOut, shape: [\"T\", q_dim], dtype: \"f32\", storage: \"activation\" });\n    addTensor({ name: gateOut, shape: [\"T\", q_dim], dtype: \"f32\", storage: \"activation\" });\n    addTensor({ name: kOut, shape: [\"T\", kv_dim], dtype: \"f32\", storage: \"activation\" });\n    addTensor({ name: vOut, shape: [\"T\", kv_dim], dtype: \"f32\", storage: \"activation\" });\n\n    addLinearOp(\n      `${prefix}_q_proj`,\n      norm1Out,\n      CANONICAL_KEYS.qProj(i),\n      [q_dim, hidden_size],\n      qOut,\n      hidden_size,\n      q_dim,\n    );\n    addLinearOp(\n      `${prefix}_gate_proj`,\n      norm1Out,\n      CANONICAL_KEYS.attnGate(i),\n      [q_dim, hidden_size],\n      gateOut,\n      hidden_size,\n      q_dim,\n    );\n    addLinearOp(\n      `${prefix}_k_proj`,\n      norm1Out,\n      CANONICAL_KEYS.kProj(i),\n      [kv_dim, hidden_size],\n      kOut,\n      hidden_size,\n      kv_dim,\n    );\n    addLinearOp(\n      `${prefix}_v_proj`,\n      norm1Out,\n      CANONICAL_KEYS.vProj(i),\n      [kv_dim, hidden_size],\n      vOut,\n      hidden_size,\n      kv_dim,\n    );\n\n    const qNormW = CANONICAL_KEYS.qNorm(i);\n    const kNormW = CANONICAL_KEYS.kNorm(i);\n    addTensor({\n      name: qNormW,\n      shape: [head_dim],\n      dtype: \"f32\",\n      storage: \"constant\",\n      safetensorsKey: qNormW,\n    });\n    addTensor({\n      name: kNormW,\n      shape: [head_dim],\n      dtype: \"f32\",\n      storage: \"constant\",\n      safetensorsKey: kNormW,\n    });\n    const qNormed = `${prefix}_q_normed`;\n    const kNormed = `${prefix}_k_normed`;\n    addTensor({ name: qNormed, shape: [\"T\", q_dim], dtype: \"f32\", storage: \"activation\" });\n    addTensor({ name: kNormed, shape: [\"T\", kv_dim], dtype: \"f32\", storage: \"activation\" });\n    addNode({\n      id: `${prefix}_q_norm`,\n      opType: \"RMSNorm\",\n      inputs: [qOut, qNormW],\n      outputs: [qNormed],\n      attributes: { hidden_size: head_dim, eps: rms_norm_eps, seq_len_tensor: qOut },\n    });\n    addNode({\n      id: `${prefix}_k_norm`,\n      opType: \"RMSNorm\",\n      inputs: [kOut, kNormW],\n      outputs: [kNormed],\n      attributes: { hidden_size: head_dim, eps: rms_norm_eps, seq_len_tensor: kOut },\n    });\n\n    addNode({\n      id: `${prefix}_rope`,\n      opType: \"RoPE\",\n      inputs: [qNormed, kNormed],\n      outputs: [qNormed, kNormed],\n      attributes: {\n        head_dim,\n        num_q_heads: num_heads,\n        num_kv_heads,\n        rope_base,\n        rope_dim,\n        partial_rotary: true,\n        seq_len_tensor: qNormed,\n      },\n    });\n\n    const kvCacheDtype: DType = kvDtype ?? \"f32\";\n    const kCache = `${prefix}_k_cache`;\n    const vCache = `${prefix}_v_cache`;\n    addTensor({ name: kCache, shape: [\"L_max\", kv_dim], dtype: kvCacheDtype, storage: \"kv_cache\" });\n    addTensor({ name: vCache, shape: [\"L_max\", kv_dim], dtype: kvCacheDtype, storage: \"kv_cache\" });\n    addNode({\n      id: `${prefix}_kv_cache_k`,\n      opType: \"KVCacheAppend\",\n      inputs: [kNormed],\n      outputs: [kCache],\n      attributes: { width: kv_dim, T_tensor: kNormed },\n    });\n    addNode({\n      id: `${prefix}_kv_cache_v`,\n      opType: \"KVCacheAppend\",\n      inputs: [vOut],\n      outputs: [vCache],\n      attributes: { width: kv_dim, T_tensor: vOut },\n    });\n\n    const attnOut = `${prefix}_attn_out`;\n    addTensor({ name: attnOut, shape: [\"T\", q_dim], dtype: \"f32\", storage: \"activation\" });\n    addNode({\n      id: `${prefix}_attn`,\n      opType: \"Attention\",\n      inputs: [qNormed, kCache, vCache],\n      outputs: [attnOut],\n      attributes: {\n        hidden_size: q_dim,\n        num_q_heads: num_heads,\n        num_kv_heads,\n        head_dim,\n        causal: true,\n        layer_index: i,\n      },\n    });\n\n    const gatedOut = `${prefix}_gated`;\n    addTensor({ name: gatedOut, shape: [\"T\", q_dim], dtype: \"f32\", storage: \"activation\" });\n    addNode({\n      id: `${prefix}_sigmoid_gate`,\n      opType: \"SigmoidGate\",\n      inputs: [attnOut, gateOut],\n      outputs: [gatedOut],\n      attributes: { count_tensor: attnOut },\n    });\n\n    const oProjOut = `${prefix}_o_proj_out`;\n    addTensor({ name: oProjOut, shape: [\"T\", hidden_size], dtype: \"f32\", storage: \"activation\" });\n    addLinearOp(\n      `${prefix}_o_proj`,\n      gatedOut,\n      CANONICAL_KEYS.oProj(i),\n      [hidden_size, q_dim],\n      oProjOut,\n      q_dim,\n      hidden_size,\n    );\n    return oProjOut;\n  };\n\n  // ── Transformer layers (fused ResidualRMSNorm chaining like qwen3_5.ts) ──\n  let prevOutput = \"inputs_embeds\";\n  let pendingNorm1Out: string | null = null;\n\n  for (let i = 0; i < num_layers; i++) {\n    const prefix = `layer${i}`;\n\n    let norm1Out: string;\n    if (pendingNorm1Out) {\n      norm1Out = pendingNorm1Out;\n    } else {\n      const normWeightName = CANONICAL_KEYS.layerInputNorm(i);\n      addTensor({\n        name: normWeightName,\n        shape: [hidden_size],\n        dtype: \"f32\",\n        storage: \"constant\",\n        safetensorsKey: normWeightName,\n      });\n      norm1Out = `${prefix}_norm1_out`;\n      addTensor({ name: norm1Out, shape: [\"T\", hidden_size], dtype: \"f32\", storage: \"activation\" });\n      addNode({\n        id: `${prefix}_norm1`,\n        opType: \"RMSNorm\",\n        inputs: [prevOutput, normWeightName],\n        outputs: [norm1Out],\n        attributes: { hidden_size, eps: rms_norm_eps, seq_len_tensor: prevOutput },\n      });\n    }\n\n    const attnOutput = buildAttentionBlock(i, prefix, norm1Out);\n\n    const resid1Out = `${prefix}_resid1`;\n    const postNormW = CANONICAL_KEYS.layerPostAttnNorm(i);\n    const norm2Out = `${prefix}_norm2_out`;\n    addTensor({ name: resid1Out, shape: [\"T\", hidden_size], dtype: \"f32\", storage: \"activation\" });\n    addTensor({\n      name: postNormW,\n      shape: [hidden_size],\n      dtype: \"f32\",\n      storage: \"constant\",\n      safetensorsKey: postNormW,\n    });\n    addTensor({ name: norm2Out, shape: [\"T\", hidden_size], dtype: \"f32\", storage: \"activation\" });\n    addNode({\n      id: `${prefix}_resid1_norm2`,\n      opType: \"ResidualRMSNorm\",\n      inputs: [prevOutput, attnOutput, postNormW],\n      outputs: [resid1Out, norm2Out],\n      attributes: { hidden_size, eps: rms_norm_eps, seq_len_tensor: prevOutput },\n    });\n\n    const gateW = CANONICAL_KEYS.gateProj(i);\n    const upW = CANONICAL_KEYS.upProj(i);\n    const downW = CANONICAL_KEYS.downProj(i);\n    const mlpGateOut = `${prefix}_gate_out`;\n    const upOut = `${prefix}_up_out`;\n    const swigluOut = `${prefix}_swiglu_out`;\n    const mlpOut = `${prefix}_mlp_out`;\n    addTensor({\n      name: mlpGateOut,\n      shape: [\"T\", intermediate_size],\n      dtype: \"f32\",\n      storage: \"activation\",\n    });\n    addTensor({\n      name: upOut,\n      shape: [\"T\", intermediate_size],\n      dtype: \"f32\",\n      storage: \"activation\",\n    });\n    addTensor({\n      name: swigluOut,\n      shape: [\"T\", intermediate_size],\n      dtype: \"f32\",\n      storage: \"activation\",\n    });\n    addTensor({ name: mlpOut, shape: [\"T\", hidden_size], dtype: \"f32\", storage: \"activation\" });\n    addLinearOp(\n      `${prefix}_gate`,\n      norm2Out,\n      gateW,\n      [intermediate_size, hidden_size],\n      mlpGateOut,\n      hidden_size,\n      intermediate_size,\n    );\n    addLinearOp(\n      `${prefix}_up`,\n      norm2Out,\n      upW,\n      [intermediate_size, hidden_size],\n      upOut,\n      hidden_size,\n      intermediate_size,\n    );\n    addNode({\n      id: `${prefix}_swiglu`,\n      opType: \"SwiGLU\",\n      inputs: [mlpGateOut, upOut],\n      outputs: [swigluOut],\n      attributes: { count_tensor: mlpGateOut },\n    });\n    addLinearOp(\n      `${prefix}_down`,\n      swigluOut,\n      downW,\n      [hidden_size, intermediate_size],\n      mlpOut,\n      intermediate_size,\n      hidden_size,\n    );\n\n    const resid2Out = `${prefix}_resid2`;\n    addTensor({ name: resid2Out, shape: [\"T\", hidden_size], dtype: \"f32\", storage: \"activation\" });\n    if (i < num_layers - 1) {\n      const nextNormW = CANONICAL_KEYS.layerInputNorm(i + 1);\n      addTensor({\n        name: nextNormW,\n        shape: [hidden_size],\n        dtype: \"f32\",\n        storage: \"constant\",\n        safetensorsKey: nextNormW,\n      });\n      const nextNorm1Out = `layer${i + 1}_norm1_out`;\n      addTensor({\n        name: nextNorm1Out,\n        shape: [\"T\", hidden_size],\n        dtype: \"f32\",\n        storage: \"activation\",\n      });\n      addNode({\n        id: `${prefix}_resid2_norm1`,\n        opType: \"ResidualRMSNorm\",\n        inputs: [resid1Out, mlpOut, nextNormW],\n        outputs: [resid2Out, nextNorm1Out],\n        attributes: { hidden_size, eps: rms_norm_eps, seq_len_tensor: resid1Out },\n      });\n      pendingNorm1Out = nextNorm1Out;\n    } else {\n      addNode({\n        id: `${prefix}_resid2`,\n        opType: \"Add\",\n        inputs: [resid1Out, mlpOut],\n        outputs: [resid2Out],\n        attributes: { count_tensor: resid1Out, hidden_size },\n      });\n      pendingNorm1Out = null;\n    }\n    prevOutput = resid2Out;\n  }\n\n  // ── Final norm → last row = the hidden state the host-side heads consume ──\n  addTensor({\n    name: CANONICAL_KEYS.FINAL_NORM,\n    shape: [hidden_size],\n    dtype: \"f32\",\n    storage: \"constant\",\n    safetensorsKey: CANONICAL_KEYS.FINAL_NORM,\n  });\n  addTensor({\n    name: \"final_norm_out\",\n    shape: [\"T\", hidden_size],\n    dtype: \"f32\",\n    storage: \"activation\",\n  });\n  addNode({\n    id: \"final_norm\",\n    opType: \"RMSNorm\",\n    inputs: [prevOutput, CANONICAL_KEYS.FINAL_NORM],\n    outputs: [\"final_norm_out\"],\n    attributes: { hidden_size, eps: rms_norm_eps, seq_len_tensor: prevOutput },\n  });\n\n  // \"logits\" is the LAST position's final-norm hidden state [1, hidden] — named\n  // so Executor.forward()'s standard readback (config.vocab_size == hidden_size\n  // floats from the \"logits\" buffer) returns it without any executor change.\n  addTensor({ name: \"logits\", shape: [1, hidden_size], dtype: \"f32\", storage: \"activation\" });\n  addNode({\n    id: \"slice_last_row\",\n    opType: \"SliceLastRow\",\n    inputs: [\"final_norm_out\"],\n    outputs: [\"logits\"],\n    attributes: { width: hidden_size },\n  });\n\n  return {\n    architecture: \"GepardForTTS\",\n    config,\n    capabilities,\n    tensors,\n    nodes,\n    executionOrder,\n    inputs: [\"inputs_embeds\"],\n    outputs: [\"logits\"],\n  };\n}\n","/**\n * Gepard-1.0 — native streaming text-to-speech engine for Gerbil's WebGPU backend.\n *\n * Gepard (nineninesix/gepard-1.0, Apache-2.0) is a frame-synchronous multihead\n * codec-LM TTS. Unlike the token-serial Kani/Oute engines (which emit one codec\n * token per forward), Gepard emits ONE WHOLE AUDIO FRAME per backbone forward:\n *\n *   1. BACKBONE (GPU): stock Qwen3.5 full-attention body (14 layers, hidden 1024)\n *      run through generateGepardBackboneGraph — input is a host-written\n *      `inputs_embeds` row, output is the last position's final-norm hidden state.\n *   2. OVERLAY (host): 32 parallel classifier heads (cardinalities [8,7,6,6]×8,\n *      216 logits total) + a binary stop head read the hidden state; the sampled\n *      32 FSQ codes feed back through per-channel embeddings → concat → GELU MLP\n *      → affine-free LayerNorm → × audio_embed_scale as the next step's input.\n *   3. CODEC (GPU): NanoCodec 21.5 fps decoder (same validated CausalHiFiGAN\n *      graph family as Kani's codec, rates [8,8,4,2,2], hop 1024) turns the\n *      per-channel codes into 22.05 kHz PCM — decoded incrementally in chunks\n *      with left-context lookback, so audio streams out while the model is\n *      still generating (≈1.5 s of audio per 32-frame chunk).\n *\n * Mirrors gepard-inference (runner.py) semantics: adaptive text repetition,\n * SOS-seeded first frame, stop-head termination, independent per-head sampling\n * (temperature / top-k / windowed repetition penalty) in fp32. Voice cloning\n * (the ref_compressor speaker prefix) is NOT wired yet — the default voice ships\n * first; cloning is a documented follow-up.\n */\n\nimport {\n  buildGepardTextLayout,\n  GEPARD_CODEC_INPUT_DIM,\n  GEPARD_CODEC_UP_SAMPLE_RATES,\n  GEPARD_HOP,\n  GEPARD_NUM_CHANNELS,\n  type GepardConfig,\n  generateGepardBackboneGraph,\n  gepardCodesToLatent,\n  parseGepardConfig,\n} from \"./architectures/gepard_tts.js\";\nimport { generateNanoCodecDecoderGraph } from \"./architectures/kani_tts.js\";\nimport type { GPUContext } from \"./device.js\";\nimport { initGPU } from \"./device.js\";\nimport { Executor } from \"./executor.js\";\nimport { type LoadedGepardTTS, loadGepardTTS, quantizeBackboneInt4 } from \"./model-loader.js\";\nimport type { Tokenizer } from \"./tokenizer.js\";\n\n/** Backbone weight precision for the codec-LM. */\nexport type GepardDType = \"q4\" | \"f32\";\n\n/** Codec decode chunking: 32 frames ≈ 1.49 s of audio per streamed chunk. */\nconst GEPARD_CODEC_CHUNK_FRAMES = 32;\n/**\n * Left-context warm-up frames re-decoded (and discarded) per chunk. The decoder\n * is fully causal with a bounded receptive field; 64 frames (≈65k samples) of\n * lookback keeps chunked output numerically identical to a monolithic decode\n * while staying under the WebGPU per-dimension dispatch cap.\n */\nconst GEPARD_CODEC_LOOKBACK_FRAMES = 64;\n/** Hard ceiling on generated frames (the reference runner's max_frames). */\nconst DEFAULT_MAX_FRAMES = 2000;\n/** PyTorch nn.LayerNorm default epsilon (audio_embed_proj's affine-free LN). */\nconst LAYERNORM_EPS = 1e-5;\n\nexport interface GepardTTSOptions {\n  /** Backbone repo (default nineninesix/gepard-1.0, Apache-2.0). */\n  repo?: string;\n  /** NanoCodec repo (default the mlx-community 21.5 fps MLX mirror). */\n  codecRepo?: string;\n  revision?: string;\n  hfToken?: string;\n  cacheDir?: string;\n  /**\n   * Backbone weight precision (default \"q4\"):\n   *   • \"q4\"  — group-wise INT4 (W4A16) on the Qwen3.5 linear projections\n   *             (backbone ~1.05 GB f32 → ~0.13 GB). Recommended default.\n   *   • \"f32\" — full precision (largest, highest fidelity).\n   * The overlay always runs f32 on the host; the codec decoder always runs f32.\n   */\n  dtype?: GepardDType;\n  /** Max KV-cache length: text layout + 1 position per audio frame. Default 2048. */\n  maxSeqLen?: number;\n  onProgress?: (loaded: number, total: number, message: string) => void;\n}\n\n/** One streamed audio chunk (decoded incrementally during generation). */\nexport interface GepardAudioChunk {\n  /** Mono PCM in [-1, 1] for this chunk only. */\n  pcm: Float32Array;\n  /** Sample rate (22050). */\n  sampleRate: number;\n  /** Index of the first audio frame this chunk covers. */\n  frameIndex: number;\n  /** True for the last chunk of the utterance. */\n  isFinal: boolean;\n}\n\nexport interface GepardSpeakOptions {\n  /** Sampling temperature (default 0.4 — the reference production operating point). */\n  temperature?: number;\n  /** Top-k cutoff per head (default 0 = disabled). */\n  topK?: number;\n  /** Stop-head sigmoid threshold (default 0.5). */\n  stopThreshold?: number;\n  /** Max audio frames to generate (caps duration; default 2000, ~93 s). */\n  maxFrames?: number;\n  /** Per-head repetition penalty over a recent-frame window (default 1.0 = off). */\n  repetitionPenalty?: number;\n  /** Number of recent frames tracked for repetitionPenalty (default 32). */\n  repetitionWindow?: number;\n  /**\n   * Streaming callback: called with each decoded audio chunk while generation is\n   * still running (frame-synchronous decode → real streaming, not post-hoc\n   * chunking). The final chunk carries `isFinal: true`. The returned SpeakResult\n   * still contains the full assembled PCM.\n   */\n  onChunk?: (chunk: GepardAudioChunk) => void;\n}\n\nexport interface GepardSpeakResult {\n  /** Mono PCM in [-1, 1]. */\n  pcm: Float32Array;\n  /** Sample rate (22050). */\n  sampleRate: number;\n  /** Number of audio frames decoded. */\n  frames: number;\n  /** Audio duration in seconds. */\n  audioSeconds: number;\n}\n\n/** Build a fresh map of just the constant weights a graph references. */\nfunction selectGraphWeights(\n  graph: { tensors: Record<string, { storage: string; safetensorsKey?: 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\") {\n      continue;\n    }\n    const w =\n      weights.get(name) ?? (desc.safetensorsKey ? weights.get(desc.safetensorsKey) : undefined);\n    if (w) {\n      out.set(name, w);\n    }\n  }\n  return out;\n}\n\n/** Abramowitz–Stegun 7.1.26 erf approximation (max |err| ≈ 1.5e-7). */\nfunction erf(x: number): number {\n  const sign = x < 0 ? -1 : 1;\n  const ax = Math.abs(x);\n  const t = 1 / (1 + 0.3275911 * ax);\n  const y =\n    1 -\n    ((((1.061405429 * t - 1.453152027) * t + 1.421413741) * t - 0.284496736) * t + 0.254829592) *\n      t *\n      Math.exp(-ax * ax);\n  return sign * y;\n}\n\n/** Exact (erf-based) GELU, matching PyTorch nn.GELU()'s default. */\nfunction gelu(x: number): number {\n  return 0.5 * x * (1 + erf(x / Math.SQRT2));\n}\n\n/** y = W·x + b for a row-major [out, in] weight. */\nfunction linear(\n  w: Float32Array,\n  b: Float32Array | null,\n  x: Float32Array,\n  outDim: number,\n  inDim: number,\n  out: Float32Array,\n): void {\n  for (let o = 0; o < outDim; o++) {\n    let acc = b ? b[o] : 0;\n    const base = o * inDim;\n    for (let k = 0; k < inDim; k++) {\n      acc += w[base + k] * x[k];\n    }\n    out[o] = acc;\n  }\n}\n\nexport class GepardTTS {\n  private ctx: GPUContext;\n  private loaded: LoadedGepardTTS;\n  private tokenizer: Tokenizer;\n  private cfg: GepardConfig;\n  private maxSeqLen: number;\n  private hidden: number;\n  /** Backbone executor (built once; reused across speak() calls). */\n  private backboneExec: Executor;\n  /** Cached codec executor for the steady-state window size (chunk + lookback). */\n  private codecExecCache = new Map<number, Executor>();\n  private _destroyed = false;\n\n  /** The selected backbone precision (\"q4\" or \"f32\"). */\n  readonly dtype: GepardDType;\n\n  // ── Overlay weights resolved once (f32) ──\n  private audioEmbTables: Float32Array[]; // [L_i, audio_embed_dim] per channel\n  private projW0: Float32Array;\n  private projB0: Float32Array;\n  private projW2: Float32Array;\n  private projB2: Float32Array;\n  private audioEmbedScale: number;\n  private headW: Float32Array[]; // [L_i, hidden] per channel\n  private headB: Float32Array[];\n  private stopW: Float32Array;\n  private stopB: number;\n\n  private constructor(\n    ctx: GPUContext,\n    loaded: LoadedGepardTTS,\n    maxSeqLen: number,\n    dtype: GepardDType,\n  ) {\n    this.ctx = ctx;\n    this.loaded = loaded;\n    this.tokenizer = loaded.tokenizer;\n    this.cfg = parseGepardConfig(loaded.gepardConfig);\n    this.maxSeqLen = maxSeqLen;\n    this.hidden = this.cfg.backboneConfig.hidden_size as number;\n    this.dtype = dtype;\n\n    // ── Resolve the host-side overlay weights once. ──\n    const overlay = loaded.overlayWeights;\n    const need = (key: string): Float32Array => {\n      const w = overlay.get(key);\n      if (!w) {\n        throw new Error(`Gepard overlay weight missing: ${key}`);\n      }\n      return w.data;\n    };\n    const n = this.cfg.vocabSizes.length;\n    if (n !== GEPARD_NUM_CHANNELS) {\n      throw new Error(`Gepard expects ${GEPARD_NUM_CHANNELS} audio heads, config has ${n}.`);\n    }\n    this.audioEmbTables = [];\n    this.headW = [];\n    this.headB = [];\n    for (let i = 0; i < n; i++) {\n      this.audioEmbTables.push(need(`audio_embeddings.${i}.weight`));\n      this.headW.push(need(`codebook_heads.${i}.weight`));\n      this.headB.push(need(`codebook_heads.${i}.bias`));\n    }\n    this.projW0 = need(\"audio_embed_proj.0.weight\");\n    this.projB0 = need(\"audio_embed_proj.0.bias\");\n    this.projW2 = need(\"audio_embed_proj.2.weight\");\n    this.projB2 = need(\"audio_embed_proj.2.bias\");\n    this.audioEmbedScale = need(\"audio_embed_scale\")[0];\n    this.stopW = need(\"stop_head.weight\");\n    this.stopB = need(\"stop_head.bias\")[0];\n\n    // ── Build + upload the backbone graph once. ──\n    const graph = generateGepardBackboneGraph(\n      this.cfg.backboneConfig,\n      this.dtype === \"q4\" ? \"q4\" : undefined,\n    );\n    if (this.dtype === \"q4\") {\n      quantizeBackboneInt4(graph, loaded.backboneWeights);\n    }\n    this.backboneExec = new Executor(ctx, graph, { maxSeqLen, kvMode: \"f32\" });\n    this.backboneExec.uploadWeightsMap(selectGraphWeights(graph, loaded.backboneWeights));\n    this.backboneExec.initBindGroups();\n  }\n\n  static async create(options: GepardTTSOptions = {}): Promise<GepardTTS> {\n    const ctx = await initGPU();\n    const loaded = await loadGepardTTS({\n      repo: options.repo,\n      codecRepo: options.codecRepo,\n      revision: options.revision,\n      hfToken: options.hfToken,\n      cacheDir: options.cacheDir,\n      onProgress: options.onProgress,\n    });\n    const ctxLen = (loaded.rawConfig.max_position_embeddings as number) ?? 262_144;\n    const maxSeqLen = Math.min(options.maxSeqLen ?? 2048, ctxLen);\n    return new GepardTTS(ctx, loaded, maxSeqLen, options.dtype ?? \"q4\");\n  }\n\n  /** NanoCodec's fixed output sample rate (Hz). */\n  get sampleRate(): number {\n    return this.cfg.sampleRate;\n  }\n\n  /** Preset voices. Gepard ships one default voice (cloning is a follow-up). */\n  availableVoices(): string[] {\n    return [];\n  }\n\n  /** Gather text-token rows from the raw BF16 embed table → f32 [T, hidden]. */\n  private embedTextTokens(ids: number[]): Float32Array {\n    const h = this.hidden;\n    const table = this.loaded.embedTokensBF16;\n    const out = new Float32Array(ids.length * h);\n    const view = new DataView(out.buffer);\n    for (let t = 0; t < ids.length; t++) {\n      const rowBase = ids[t] * h;\n      const dstBase = t * h;\n      for (let k = 0; k < h; k++) {\n        // bf16 → f32: the bf16 word is the top 16 bits of the f32 pattern.\n        view.setUint32((dstBase + k) * 4, table[rowBase + k] << 16, true);\n      }\n    }\n    return out;\n  }\n\n  /**\n   * Embed one audio frame (32 FSQ codes) through the overlay stack:\n   * per-channel lookup → concat → Linear→GELU→Linear → affine-free LayerNorm →\n   * × audio_embed_scale. Mirrors GepardModel._embed_audio.\n   */\n  private embedAudioFrame(codes: Uint32Array): Float32Array {\n    const h = this.hidden;\n    const dim = this.cfg.audioEmbedDim;\n    const x = new Float32Array(GEPARD_NUM_CHANNELS * dim);\n    for (let c = 0; c < GEPARD_NUM_CHANNELS; c++) {\n      const table = this.audioEmbTables[c];\n      const base = codes[c] * dim;\n      x.set(table.subarray(base, base + dim), c * dim);\n    }\n    const h1 = new Float32Array(h);\n    linear(this.projW0, this.projB0, x, h, x.length, h1);\n    for (let i = 0; i < h; i++) {\n      h1[i] = gelu(h1[i]);\n    }\n    const h2 = new Float32Array(h);\n    linear(this.projW2, this.projB2, h1, h, h, h2);\n    // Affine-free LayerNorm pins the frame-embedding scale.\n    let mean = 0;\n    for (let i = 0; i < h; i++) {\n      mean += h2[i];\n    }\n    mean /= h;\n    let variance = 0;\n    for (let i = 0; i < h; i++) {\n      const d = h2[i] - mean;\n      variance += d * d;\n    }\n    variance /= h;\n    const inv = this.audioEmbedScale / Math.sqrt(variance + LAYERNORM_EPS);\n    for (let i = 0; i < h; i++) {\n      h2[i] = (h2[i] - mean) * inv;\n    }\n    return h2;\n  }\n\n  /** Stop-head probability for a hidden state. */\n  private stopProbability(hidden: Float32Array): number {\n    let acc = this.stopB;\n    for (let k = 0; k < this.hidden; k++) {\n      acc += this.stopW[k] * hidden[k];\n    }\n    return 1 / (1 + Math.exp(-acc));\n  }\n\n  /**\n   * Sample one audio frame: all 32 heads independently in fp32 —\n   * windowed repetition penalty → temperature → top-k → softmax → multinomial.\n   * Mirrors GepardRunner._sample_frame / _sample_head_logits.\n   */\n  private sampleFrame(\n    hidden: Float32Array,\n    p: { temperature: number; topK: number; repetitionPenalty: number },\n    recentFrames: Uint32Array[] | null,\n  ): Uint32Array {\n    const frame = new Uint32Array(GEPARD_NUM_CHANNELS);\n    for (let c = 0; c < GEPARD_NUM_CHANNELS; c++) {\n      const L = this.cfg.vocabSizes[c];\n      const w = this.headW[c];\n      const b = this.headB[c];\n      const logits = new Float32Array(L);\n      for (let j = 0; j < L; j++) {\n        let acc = b[j];\n        const base = j * this.hidden;\n        for (let k = 0; k < this.hidden; k++) {\n          acc += w[base + k] * hidden[k];\n        }\n        logits[j] = acc;\n      }\n      if (p.repetitionPenalty !== 1.0 && recentFrames) {\n        for (const prev of recentFrames) {\n          const tok = prev[c];\n          logits[tok] =\n            logits[tok] > 0 ? logits[tok] / p.repetitionPenalty : logits[tok] * p.repetitionPenalty;\n        }\n      }\n      if (p.temperature !== 1.0) {\n        for (let j = 0; j < L; j++) {\n          logits[j] /= p.temperature;\n        }\n      }\n      if (p.topK > 0 && p.topK < L) {\n        const sorted = [...logits].sort((a, b2) => b2 - a);\n        const threshold = sorted[p.topK - 1];\n        for (let j = 0; j < L; j++) {\n          if (logits[j] < threshold) {\n            logits[j] = Number.NEGATIVE_INFINITY;\n          }\n        }\n      }\n      // Softmax + multinomial.\n      let maxV = Number.NEGATIVE_INFINITY;\n      for (let j = 0; j < L; j++) {\n        if (logits[j] > maxV) {\n          maxV = logits[j];\n        }\n      }\n      let sum = 0;\n      for (let j = 0; j < L; j++) {\n        const e = Math.exp(logits[j] - maxV);\n        logits[j] = e;\n        sum += e;\n      }\n      let pick = Math.random() * sum;\n      let chosen = L - 1;\n      for (let j = 0; j < L; j++) {\n        pick -= logits[j];\n        if (pick <= 0) {\n          chosen = j;\n          break;\n        }\n      }\n      frame[c] = chosen;\n    }\n    return frame;\n  }\n\n  /**\n   * Synthesize speech for `text`. Returns 22.05 kHz mono PCM; when\n   * `opts.onChunk` is set, audio streams out chunk-by-chunk during generation.\n   *\n   * Pipeline: adaptive-repetition text layout → prefill (host-gathered text\n   * embeddings) → frame-synchronous AR loop (one frame = 32 codes per forward,\n   * stop head decides termination) → incremental NanoCodec decode → PCM.\n   */\n  async speak(text: string, opts: GepardSpeakOptions = {}): Promise<GepardSpeakResult> {\n    if (this._destroyed) {\n      throw new Error(\"GepardTTS has been destroyed.\");\n    }\n    const temperature = opts.temperature ?? 0.4;\n    const topK = opts.topK ?? 0;\n    const stopThreshold = opts.stopThreshold ?? 0.5;\n    const repetitionPenalty = opts.repetitionPenalty ?? 1.0;\n    const repetitionWindow = opts.repetitionWindow ?? 32;\n\n    // ── Text layout: [ (SOT text EOT) × (R−1) | SOT text EOT SOS ]. ──\n    const textIds = this.tokenizer.encode(text);\n    const layout = buildGepardTextLayout(textIds, this.cfg);\n    if (layout.length + 2 >= this.maxSeqLen) {\n      throw new Error(`Gepard prompt is ${layout.length} tokens >= maxSeqLen ${this.maxSeqLen}.`);\n    }\n    const maxFrames = Math.min(\n      opts.maxFrames ?? DEFAULT_MAX_FRAMES,\n      this.maxSeqLen - layout.length - 1,\n    );\n\n    // ── Prefill: host-gathered text embeddings → hidden state at SOS. ──\n    const startTime = performance.now();\n    this.backboneExec.reset();\n    this.backboneExec.writeInput(\"inputs_embeds\", this.embedTextTokens(layout));\n    // input_ids content is unused (the graph reads inputs_embeds); only the\n    // length matters — it sets T for shape resolution and seqPos advancement.\n    let { logits: hidden } = await this.backboneExec.forward(new Uint32Array(layout.length));\n\n    // ── Frame-synchronous AR loop. ──\n    const frames: Uint32Array[] = [\n      this.sampleFrame(hidden, { temperature, topK, repetitionPenalty }, null),\n    ];\n    const streamer = new GepardChunkStreamer(this, opts.onChunk);\n    const sampleOpts = { temperature, topK, repetitionPenalty };\n    for (let step = 1; step < maxFrames; step++) {\n      const frameEmbed = this.embedAudioFrame(frames[frames.length - 1]);\n      this.backboneExec.writeInput(\"inputs_embeds\", frameEmbed);\n      const res = await this.backboneExec.forward(new Uint32Array(1));\n      hidden = res.logits;\n\n      // Stop decision precedes sampling (the reference runner's ordering).\n      if (this.stopProbability(hidden) > stopThreshold) {\n        break;\n      }\n      const window =\n        repetitionPenalty !== 1.0\n          ? frames.slice(repetitionWindow > 0 ? -repetitionWindow : 0)\n          : null;\n      frames.push(this.sampleFrame(hidden, sampleOpts, window));\n      await streamer.maybeEmit(frames, false);\n    }\n\n    // ── Final decode of any remaining frames + assembly. ──\n    await streamer.maybeEmit(frames, true);\n    const pcm = streamer.assembled(frames.length);\n\n    const totalTime = (performance.now() - startTime) / 1000;\n    const audioSeconds = (frames.length * GEPARD_HOP) / this.cfg.sampleRate;\n    console.log(\n      `[gepard] frames=${frames.length} audio=${audioSeconds.toFixed(2)}s ` +\n        `wall=${totalTime.toFixed(2)}s RTF=${(audioSeconds / totalTime).toFixed(2)}x`,\n    );\n    return { pcm, sampleRate: this.cfg.sampleRate, frames: frames.length, audioSeconds };\n  }\n\n  /**\n   * Decode frames [from, to) of the utterance to PCM, re-decoding up to\n   * `GEPARD_CODEC_LOOKBACK_FRAMES` of left context as warm-up (discarded).\n   * The decoder is causal with a bounded receptive field, so chunked output\n   * matches a monolithic decode.\n   */\n  async decodeFrameRange(frames: Uint32Array[], from: number, to: number): Promise<Float32Array> {\n    const ctxStart = Math.max(0, from - GEPARD_CODEC_LOOKBACK_FRAMES);\n    const winFrames = to - ctxStart;\n    // Channel-major codes [32, winFrames] → dequantized latent.\n    const codes = new Uint32Array(GEPARD_NUM_CHANNELS * winFrames);\n    for (let t = 0; t < winFrames; t++) {\n      const frame = frames[ctxStart + t];\n      for (let c = 0; c < GEPARD_NUM_CHANNELS; c++) {\n        codes[c * winFrames + t] = frame[c];\n      }\n    }\n    const latent = gepardCodesToLatent(codes, winFrames);\n    const win = await this.decodeLatentWindow(latent, winFrames);\n    const drop = (from - ctxStart) * GEPARD_HOP;\n    const keep = (to - from) * GEPARD_HOP;\n    const out = win.slice(drop, drop + keep);\n    for (let i = 0; i < out.length; i++) {\n      out[i] = Math.min(1, Math.max(-1, out[i]));\n    }\n    return out;\n  }\n\n  /** Run the codec graph for one latent window → raw PCM (pre-clamp). */\n  private async decodeLatentWindow(latent: Float32Array, winFrames: number): Promise<Float32Array> {\n    // The steady-state window size (chunk + lookback) recurs every chunk, so its\n    // executor (graph + uploaded weights) is cached; odd sizes are ephemeral.\n    const steady = winFrames === GEPARD_CODEC_CHUNK_FRAMES + GEPARD_CODEC_LOOKBACK_FRAMES;\n    let exec = steady ? this.codecExecCache.get(winFrames) : undefined;\n    let graphOutputs: string[];\n    if (exec) {\n      graphOutputs = exec.graphOutputs;\n    } else {\n      const graph = generateNanoCodecDecoderGraph({\n        numFrames: winFrames,\n        geometry: {\n          upSampleRates: GEPARD_CODEC_UP_SAMPLE_RATES,\n          inputDim: GEPARD_CODEC_INPUT_DIM,\n        },\n        latentInput: true,\n      });\n      exec = new Executor(this.ctx, graph, { maxSeqLen: winFrames, kvMode: \"f32\" });\n      exec.uploadWeightsMap(selectGraphWeights(graph, this.loaded.codecWeights));\n      exec.initBindGroups();\n      graphOutputs = graph.outputs;\n      if (steady) {\n        this.codecExecCache.set(winFrames, exec);\n      }\n    }\n    try {\n      exec.reset();\n      exec.writeInput(\"nc_latent\", latent);\n      return await exec.runGraphOutput(graphOutputs[0], winFrames * GEPARD_HOP);\n    } finally {\n      if (!steady) {\n        exec.destroy();\n      }\n    }\n  }\n\n  destroy(): void {\n    if (this._destroyed) {\n      return;\n    }\n    this._destroyed = true;\n    this.backboneExec.destroy();\n    for (const exec of this.codecExecCache.values()) {\n      exec.destroy();\n    }\n    this.codecExecCache.clear();\n    this.loaded.backboneWeights.clear();\n    this.loaded.overlayWeights.clear();\n    this.loaded.codecWeights.clear();\n  }\n}\n\n/**\n * Incremental chunk decoder: every GEPARD_CODEC_CHUNK_FRAMES completed frames\n * it decodes the new span (with lookback) and emits it via `onChunk`, keeping\n * all emitted PCM so the final result is assembled without a second decode.\n */\nclass GepardChunkStreamer {\n  private tts: GepardTTS;\n  private onChunk?: (chunk: GepardAudioChunk) => void;\n  private decodedUpTo = 0;\n  private chunks: Float32Array[] = [];\n\n  constructor(tts: GepardTTS, onChunk?: (chunk: GepardAudioChunk) => void) {\n    this.tts = tts;\n    this.onChunk = onChunk;\n  }\n\n  /** Decode + emit any complete pending chunk (or everything when `final`). */\n  async maybeEmit(frames: Uint32Array[], final: boolean): Promise<void> {\n    // Without a chunk consumer we still decode once at the end (final=true) —\n    // decoding lazily in one pass per chunk boundary would waste GPU work.\n    if (!(final || this.onChunk)) {\n      return;\n    }\n    while (frames.length - this.decodedUpTo >= GEPARD_CODEC_CHUNK_FRAMES || final) {\n      const from = this.decodedUpTo;\n      const to = final ? frames.length : Math.min(frames.length, from + GEPARD_CODEC_CHUNK_FRAMES);\n      if (to <= from) {\n        break;\n      }\n      const pcm = await this.tts.decodeFrameRange(frames, from, to);\n      this.chunks.push(pcm);\n      this.decodedUpTo = to;\n      this.onChunk?.({\n        pcm,\n        sampleRate: this.tts.sampleRate,\n        frameIndex: from,\n        isFinal: final && this.decodedUpTo >= frames.length,\n      });\n      if (final && this.decodedUpTo >= frames.length) {\n        break;\n      }\n    }\n  }\n\n  /** Concatenate all emitted chunks into the full utterance PCM. */\n  assembled(totalFrames: number): Float32Array {\n    const pcm = new Float32Array(totalFrames * GEPARD_HOP);\n    let offset = 0;\n    for (const chunk of this.chunks) {\n      pcm.set(chunk, offset);\n      offset += chunk.length;\n    }\n    return pcm;\n  }\n}\n","/**\n * OuteTTS-1.0-0.6B — native text-to-speech architecture (OuteAI/OuteTTS-1.0-0.6B).\n *\n * Two-stage codec-LM TTS, the same shape as Kani-TTS but with a plain Qwen3 backbone\n * and a descript-audio-codec (DAC) decoder instead of LFM2 + NanoCodec:\n *\n *   1. CODEC-LM BACKBONE = a plain Qwen3-0.6B `Qwen3ForCausalLM` (hidden 1024, 28\n *      layers, GQA 16/8 heads, head_dim 128, SwiGLU, QK-RMSNorm, full RoPE θ=1e6,\n *      tied embeddings, vocab 157760). The audio codes live in the SAME vocab as\n *      text: two interleaved codebooks `<|c1_i|>` (ids 151669..152693) and\n *      `<|c2_i|>` (ids 152694..153718). The backbone autoregressively emits one\n *      (c1,c2) pair per 512-sample frame between `<|audio_start|>` and `<|audio_end|>`.\n *      Reuses the existing Qwen2/3 graph generator with a KV-cache attention path\n *      (generateOuteTtsBackboneGraph) — SIMPLER than Kani's LFM2 hybrid (no short\n *      conv, no learnable RoPE, no frame-level positions: standard sequential RoPE).\n *\n *   2. DAC.speech.v1.0 DECODER (ibm-research/DAC.speech.v1.0, 24 kHz, 2-codebook\n *      1.5 kbps): RVQ dequant (2× codebook lookup [1024,8] + out_proj Conv1d(8→1024,\n *      k1) + sum) → latent z[1024,T] → DacDecoder (Conv1d 1024→1536 k7 → 4 upsample\n *      blocks rates [8,5,4,2] → Snake → Conv1d 96→1 k7 → tanh) → 24 kHz PCM.\n *      hop_length 512 (token rate ≈ 46.875 Hz). FFT-free; every op maps to an\n *      EXISTING validated kernel (Conv1dFull / ConvTranspose1d / Snake1d / Tanh /\n *      Add) plus one new RVQDequant kernel (codebook lookup + out_proj sum).\n *\n * ── STEP-0 verification (live HF + .pth metadata, June 2026) ─────────────────\n * Backbone config.json (OuteAI/OuteTTS-1.0-0.6B):\n *   architectures [\"Qwen3ForCausalLM\"], model_type \"qwen3\",\n *   hidden_size 1024, num_hidden_layers 28, num_attention_heads 16,\n *   num_key_value_heads 8, head_dim 128, intermediate_size 3072, vocab_size 157760,\n *   rope_theta 1e6, rms_norm_eps 1e-6, tie_word_embeddings true, max_position 40960.\n *   added_tokens.json: <|c1_0..1024|> = ids 151669..152693 (1025 each), <|c2_0..1024|>\n *   = ids 152694..153718; <|text_start|>156722 <|text_end|>156723 <|audio_start|>156728\n *   <|audio_end|>156729 <|code|>156730 <|word_start|>156731 <|word_end|>156732\n *   <|features|>156733. (Index 1024 is a reserved vocab slot; real DAC codes are 0..1023.)\n *\n * DAC.speech.v1.0 weights_24khz_1.5kbps_v1.0.pth metadata.kwargs (descript-audio-codec):\n *   latent_dim 1024, decoder_dim 1536, decoder_rates [8,5,4,2], n_codebooks 2,\n *   codebook_size 1024, codebook_dim 8, sample_rate 24000 (hop_length 512).\n *   State-dict (decoder.model.*): plain descript layout, NOT the HF DacModel rename —\n *     decoder.model.0          Conv1d(1024→1536, k7, pad3)                  [weight_g/v + bias]\n *     decoder.model.{1..4}     DecoderBlock(rate r ∈ [8,5,4,2]):\n *       .block.0.alpha         Snake1d on Cin                               [1,Cin,1]\n *       .block.1               ConvTranspose1d(Cin→Cin/2, k=2r, stride r,\n *                                              pad=ceil(r/2), output_padding 0)\n *       .block.{2,3,4}         3× ResidualUnit(dilation d ∈ [1,3,9]):\n *         .block.0.alpha       Snake1d                                      [1,C,1]\n *         .block.1             Conv1d(C→C, k7, dilation d, pad=((7-1)*d)//2)\n *         .block.2.alpha       Snake1d\n *         .block.3             Conv1d(C→C, k1)                              (residual add)\n *     decoder.model.5.alpha    final Snake1d on 96                          [1,96,1]\n *     decoder.model.6          Conv1d(96→1, k7, pad3)\n *     + tanh (host-side after readback)\n *   Quantizer (quantizer.quantizers.{0,1}.*): codebook.weight [1024,8],\n *     out_proj.weight_g/v → [1024,8,1] (Cout 1024, Cin 8, K1), out_proj.bias [1024].\n *   Conv weights are weight-norm parametrizations (weight_g/weight_v); folded offline\n *   (W = g·v/‖v‖, L2 over dims except dim0) by the prep script into plain conv weights.\n *\n * ── DAC stage length arithmetic (verified against torch reference) ───────────\n *   ConvTranspose1d Lout = (L-1)*stride - 2*pad + (K-1) + output_padding + 1.\n *   With K=2r, pad=ceil(r/2), output_padding=0: even r → L*r exactly; odd r (rate 5)\n *   loses a sample (e.g. T=48 → 384 → 1919 → 7676 → 15352). The full decoded length is\n *   320·T − 8 (interior hop = ∏rates = 8·5·4·2 = 320, a constant 8-sample edge deficit),\n *   so DAC runs at 24000/320 = 75 fps — NOT the config's hop_length 512 (that is a\n *   different DAC variant's STFT hop). The residual convs are length-preserving\n *   (symmetric padding), so no center-crop is needed. The decoder is NON-causal and the\n *   odd-rate transpose makes interior windows position-dependent, so the whole clip is\n *   decoded in ONE pass — the codec conv/Snake/Add kernels use 2D dispatch to stay under\n *   WebGPU's 65535 per-dim workgroup cap (no chunking).\n *\n * ── STATUS ───────────────────────────────────────────────────────────────────\n * FOUNDATION (this file): RVQDequant + DAC decoder graph (generateDacSpeechDecoderGraph),\n *   config + token-layout parsing, and audioTokensToDacCodes. Backbone reuses the\n *   Qwen3 graph via the KV-cached generator. Validated bit-exact vs the torch DAC\n *   reference (scripts/engine/dac-speech-ref.py) by scripts/engine/test-outetts.mjs.\n * NOT in this phase: speaker/voice-clone prompt-prefix API, int4 backbone, the\n *   speak()/useTTS surface, R2 mirror, playground/docs.\n */\n\nimport type { DType, ModelArchConfig, ModelGraph, OpNode, TensorDesc } from \"../ir.js\";\nimport { CANONICAL_KEYS } from \"../ir.js\";\nimport { DEFAULT_GROUP_SIZE } from \"../quantize.js\";\nimport type { GraphDType, KVDType } from \"./index.js\";\n\n// ── OuteTTS-1.0 token layout (added_tokens.json, Interface Version 3) ──\nexport const OUTETTS_C1_BASE = 151_669; // <|c1_0|>\nexport const OUTETTS_C2_BASE = 152_694; // <|c2_0|>\nexport const OUTETTS_CODES_PER_CODEBOOK = 1025; // <|cN_0|>..<|cN_1024|> (1024 = reserved slot)\nexport const OUTETTS_TEXT_START = 156_722;\nexport const OUTETTS_TEXT_END = 156_723;\nexport const OUTETTS_AUDIO_START = 156_728;\nexport const OUTETTS_AUDIO_END = 156_729; // generation stop token\nexport const OUTETTS_CODE = 156_730;\nexport const OUTETTS_WORD_START = 156_731;\nexport const OUTETTS_WORD_END = 156_732;\nexport const OUTETTS_FEATURES = 156_733;\nexport const OUTETTS_IM_START = 151_644; // <|im_start|> (bos)\nexport const OUTETTS_IM_END = 151_645; // <|im_end|> (eos)\n\n// ── DAC.speech.v1.0 (24 kHz, 2-codebook 1.5 kbps) constants ──\nexport const DAC_SAMPLE_RATE = 24_000;\n// Decoder upsampling is the PRODUCT of decoder_rates (8·5·4·2 = 320), so the token\n// grid is 24000/320 = 75 fps. (config.json's hop_length 512 is the encoder/STFT hop\n// of a different DAC variant; the shipped .pth decoder_rates give the true 320× hop —\n// matching outetts' speaker-encoder tps=75.) The full decoded length is 320·T − 8\n// (a constant 8-sample edge deficit from the odd rate-5 padding); interior frames are\n// a clean 320 samples apart, which keeps chunked decode exactly placeable.\nexport const DAC_HOP = 320;\nexport const DAC_NUM_CODEBOOKS = 2;\nexport const DAC_CODEBOOK_SIZE = 1024;\nexport const DAC_CODEBOOK_DIM = 8;\nexport const DAC_LATENT_DIM = 1024; // quantizer out_proj output / decoder.conv1 input\nexport const DAC_DECODER_DIM = 1536; // decoder.conv1 output channels\nexport const DAC_DECODER_RATES = [8, 5, 4, 2] as const;\nexport const DAC_RESBLOCK_DILATIONS = [1, 3, 9] as const;\n// Encoder is the mirror of the decoder: downsampling strides [2,4,5,8] (the decoder\n// rates reversed). encoder_dim 64 → doubles per block (64→128→256→512→1024). The\n// final Conv1d(1024→1024, k3) emits the continuous latent the quantizer encodes.\nexport const DAC_ENCODER_DIM = 64;\nexport const DAC_ENCODER_RATES = [2, 4, 5, 8] as const;\n\n/** Parsed OuteTTS backbone config (the Qwen3 dims). Reuses the generic Qwen path. */\nexport interface OuteTtsConfig {\n  hidden_size: number;\n  num_layers: number;\n  num_heads: number;\n  num_kv_heads: number;\n  head_dim: number;\n  intermediate_size: number;\n  vocab_size: number;\n  rope_base: number;\n  rms_norm_eps: number;\n  max_position_embeddings: number;\n}\n\nexport function parseOuteTtsConfig(rawConfig: Record<string, unknown>): OuteTtsConfig {\n  const hidden_size = rawConfig.hidden_size as number;\n  const num_heads = rawConfig.num_attention_heads as number;\n  return {\n    hidden_size,\n    num_layers: rawConfig.num_hidden_layers as number,\n    num_heads,\n    num_kv_heads: (rawConfig.num_key_value_heads as number) ?? num_heads,\n    head_dim: (rawConfig.head_dim as number) ?? Math.floor(hidden_size / num_heads),\n    intermediate_size: rawConfig.intermediate_size as number,\n    vocab_size: rawConfig.vocab_size as number,\n    rope_base: (rawConfig.rope_theta as number) ?? 1_000_000,\n    rms_norm_eps: (rawConfig.rms_norm_eps as number) ?? 1e-6,\n    max_position_embeddings: (rawConfig.max_position_embeddings as number) ?? 40_960,\n  };\n}\n\n/**\n * Convert the flat generated token stream (the run BETWEEN <|audio_start|> and\n * <|audio_end|>) into a DAC code grid [NUM_CODEBOOKS, T] (codebook-major [c*T + t]).\n *\n * Mirrors outetts v3 `extract_audio_from_tokens`: filter the stream into two ordered\n * lists by membership of the c1 / c2 ranges, recover the raw code\n * (id − base), drop the reserved index 1024, and truncate to the shorter list. The\n * model emits c1,c2 interleaved per frame, but the parse is membership-based (it does\n * not assume strict pairing) so a stray token cannot desync the grid.\n */\nexport function audioTokensToDacCodes(tokenIds: number[]): {\n  codes: Uint32Array;\n  numFrames: number;\n} {\n  const c1: number[] = [];\n  const c2: number[] = [];\n  for (const id of tokenIds) {\n    if (id >= OUTETTS_C1_BASE && id < OUTETTS_C1_BASE + OUTETTS_CODES_PER_CODEBOOK) {\n      const code = id - OUTETTS_C1_BASE;\n      if (code < DAC_CODEBOOK_SIZE) {\n        c1.push(code);\n      }\n    } else if (id >= OUTETTS_C2_BASE && id < OUTETTS_C2_BASE + OUTETTS_CODES_PER_CODEBOOK) {\n      const code = id - OUTETTS_C2_BASE;\n      if (code < DAC_CODEBOOK_SIZE) {\n        c2.push(code);\n      }\n    }\n  }\n  const numFrames = Math.min(c1.length, c2.length);\n  const codes = new Uint32Array(DAC_NUM_CODEBOOKS * numFrames); // codebook-major [c*T + t]\n  for (let t = 0; t < numFrames; t++) {\n    codes[t] = c1[t];\n    codes[numFrames + t] = c2[t];\n  }\n  return { codes, numFrames };\n}\n\n/** ConvTranspose1d output length: (L-1)*stride - 2*pad + (K-1) + output_padding + 1. */\nfunction convTransposeLen(\n  L: number,\n  stride: number,\n  K: number,\n  pad: number,\n  outPad: number,\n): number {\n  return (L - 1) * stride - 2 * pad + (K - 1) + outPad + 1;\n}\n\n// ════════════════════════════════════════════════════════════════════════════\n// DAC.speech.v1.0 decoder graph (RVQ dequant + DacDecoder), codes → PCM.\n//\n// Input: audio codes [NUM_CODEBOOKS, T] (u32, codebook-major). Output: pcm [1, Lout]\n// (the raw conv output; the final tanh is applied host-side after readback). Weight\n// tensors carry the canonical `dac.*` names the prep script writes (weight-norm\n// already folded). Channels halve per stage (1536 → 768 → 384 → 192 → 96).\n// ════════════════════════════════════════════════════════════════════════════\n\n/** Canonical DAC decoder weight name. */\nfunction dacKey(name: string): string {\n  return `dac.${name}`;\n}\n\nexport interface DacSpeechGraphOptions {\n  /** Number of audio frames T (the code-grid width). */\n  numFrames: number;\n  /**\n   * Number of RVQ codebooks. Default 2 (DAC.speech.v1.0). Parler-TTS's\n   * dac_44khz_8kbps uses 9 — the RVQDequant kernel loops over NCB so the only\n   * change is the codebook/out_proj tensor first-dim and the codes grid height.\n   */\n  numCodebooks?: number;\n  /**\n   * Decoder upsampling rates (the DecoderBlock strides). Default [8,5,4,2]\n   * (DAC.speech.v1.0). Parler's dac_44khz uses [8,8,4,2] (hop 512). Channels\n   * still halve per stage from `decoderDim`.\n   */\n  decoderRates?: readonly number[];\n  /** First-conv output channels (decoder_dim). Default 1536. */\n  decoderDim?: number;\n  /** Quantizer out_proj output / first-conv input dim (latent_dim). Default 1024. */\n  latentDim?: number;\n}\n\nexport function generateDacSpeechDecoderGraph(opts: DacSpeechGraphOptions): ModelGraph {\n  const T = opts.numFrames;\n  const NCB = opts.numCodebooks ?? DAC_NUM_CODEBOOKS;\n  const rates = opts.decoderRates ?? DAC_DECODER_RATES;\n  const decoderDim = opts.decoderDim ?? DAC_DECODER_DIM;\n  const latentDim = opts.latentDim ?? DAC_LATENT_DIM;\n  const tensors: Record<string, TensorDesc> = {};\n  const nodes: OpNode[] = [];\n  const executionOrder: string[] = [];\n\n  const addTensor = (desc: TensorDesc): void => {\n    tensors[desc.name] = desc;\n  };\n  const addNode = (node: OpNode): void => {\n    nodes.push(node);\n    executionOrder.push(node.id);\n  };\n\n  // ── Input: audio codes [NCB, T] (u32, codebook-major) ──\n  addTensor({\n    name: \"audio_codes\",\n    shape: [NCB, T],\n    dtype: \"u32\",\n    storage: \"activation\",\n  });\n\n  // ── RVQ dequant → latent z[latentDim, T] ──\n  // codebook[NCB, codebook_size, dim], out_proj_w[NCB, Cout, dim], out_proj_b[NCB, Cout].\n  addTensor({\n    name: dacKey(\"quantizer.codebook\"),\n    shape: [NCB, DAC_CODEBOOK_SIZE, DAC_CODEBOOK_DIM],\n    dtype: \"f32\",\n    storage: \"constant\",\n    safetensorsKey: dacKey(\"quantizer.codebook\"),\n  });\n  addTensor({\n    name: dacKey(\"quantizer.out_proj.weight\"),\n    shape: [NCB, latentDim, DAC_CODEBOOK_DIM],\n    dtype: \"f32\",\n    storage: \"constant\",\n    safetensorsKey: dacKey(\"quantizer.out_proj.weight\"),\n  });\n  addTensor({\n    name: dacKey(\"quantizer.out_proj.bias\"),\n    shape: [NCB, latentDim],\n    dtype: \"f32\",\n    storage: \"constant\",\n    safetensorsKey: dacKey(\"quantizer.out_proj.bias\"),\n  });\n  addTensor({\n    name: \"dac_latent\",\n    shape: [latentDim, T],\n    dtype: \"f32\",\n    storage: \"activation\",\n  });\n  addNode({\n    id: \"dac_rvq\",\n    opType: \"RVQDequant\",\n    inputs: [\n      \"audio_codes\",\n      dacKey(\"quantizer.codebook\"),\n      dacKey(\"quantizer.out_proj.weight\"),\n      dacKey(\"quantizer.out_proj.bias\"),\n    ],\n    outputs: [\"dac_latent\"],\n    attributes: {\n      NCB,\n      Cout: latentDim,\n      T,\n      D: DAC_CODEBOOK_DIM,\n      codebook_size: DAC_CODEBOOK_SIZE,\n    },\n  });\n\n  /** Emit a full (non-causal) Conv1dFull with symmetric padding. */\n  function conv1dFull(\n    id: string,\n    input: string,\n    weightKey: string,\n    biasKey: string,\n    Cin: number,\n    Cout: number,\n    K: number,\n    L: number,\n    dilation: number,\n    padding: number,\n    output: string,\n  ): number {\n    const Lout = Math.floor((L + 2 * padding - dilation * (K - 1) - 1) / 1) + 1;\n    addTensor({\n      name: weightKey,\n      shape: [Cout, Cin, K],\n      dtype: \"f32\",\n      storage: \"constant\",\n      safetensorsKey: weightKey,\n    });\n    addTensor({\n      name: biasKey,\n      shape: [Cout],\n      dtype: \"f32\",\n      storage: \"constant\",\n      safetensorsKey: biasKey,\n    });\n    addTensor({ name: output, shape: [Cout, Lout], dtype: \"f32\", storage: \"activation\" });\n    addNode({\n      id,\n      opType: \"Conv1dFull\",\n      inputs: [input, weightKey, biasKey],\n      outputs: [output],\n      attributes: { Cin, Cout, L, Lout, K, stride: 1, padding, dilation },\n    });\n    return Lout;\n  }\n\n  /** Emit a full-channel Snake1d (alpha is [C], stored [1,C,1] → C). */\n  function snake(\n    id: string,\n    input: string,\n    C: number,\n    L: number,\n    alphaKey: string,\n    output: string,\n  ): void {\n    addTensor({\n      name: alphaKey,\n      shape: [C],\n      dtype: \"f32\",\n      storage: \"constant\",\n      safetensorsKey: alphaKey,\n    });\n    addTensor({ name: output, shape: [C, L], dtype: \"f32\", storage: \"activation\" });\n    addNode({\n      id,\n      opType: \"Snake1d\",\n      inputs: [input, alphaKey],\n      outputs: [output],\n      attributes: { C, L },\n    });\n  }\n\n  /** Emit a strided ConvTranspose1d (weight [Cin, Cout, K]). */\n  function convTranspose(\n    id: string,\n    input: string,\n    weightKey: string,\n    biasKey: string,\n    Cin: number,\n    Cout: number,\n    K: number,\n    L: number,\n    stride: number,\n    padding: number,\n    outPad: number,\n    output: string,\n  ): number {\n    const Lout = convTransposeLen(L, stride, K, padding, outPad);\n    addTensor({\n      name: weightKey,\n      shape: [Cin, Cout, K],\n      dtype: \"f32\",\n      storage: \"constant\",\n      safetensorsKey: weightKey,\n    });\n    addTensor({\n      name: biasKey,\n      shape: [Cout],\n      dtype: \"f32\",\n      storage: \"constant\",\n      safetensorsKey: biasKey,\n    });\n    addTensor({ name: output, shape: [Cout, Lout], dtype: \"f32\", storage: \"activation\" });\n    addNode({\n      id,\n      opType: \"ConvTranspose1d\",\n      inputs: [input, weightKey, biasKey],\n      outputs: [output],\n      attributes: { Cin, Cout, L, Lout, K, stride, padding, output_padding: outPad },\n    });\n    return Lout;\n  }\n\n  /** Element-wise residual add of two equal-shape [C, L] activations. */\n  function add(id: string, a: string, b: string, C: number, L: number, output: string): void {\n    addTensor({ name: output, shape: [C, L], dtype: \"f32\", storage: \"activation\" });\n    addNode({\n      id,\n      opType: \"Add\",\n      inputs: [a, b],\n      outputs: [output],\n      attributes: { count_tensor: a, hidden_size: C * L },\n    });\n  }\n\n  // ── decoder.model.0: Conv1d(latentDim → decoderDim, k7, pad3) ──\n  let cur = \"dac_conv1_out\";\n  let length = conv1dFull(\n    \"dac_conv1\",\n    \"dac_latent\",\n    dacKey(\"decoder.model.0.weight\"),\n    dacKey(\"decoder.model.0.bias\"),\n    latentDim,\n    decoderDim,\n    7,\n    T,\n    1,\n    3,\n    cur,\n  );\n  let channels = decoderDim;\n\n  // ── decoder.model.{1..N}: DecoderBlocks (N = rates.length) ──\n  for (let stage = 0; stage < rates.length; stage++) {\n    const m = stage + 1; // decoder.model index\n    const rate = rates[stage];\n    const outCh = channels / 2;\n    const upK = 2 * rate;\n    const upPad = Math.ceil(rate / 2);\n\n    // .block.0: Snake on `channels`\n    const snakeOut = `dac_s${stage}_snake`;\n    snake(\n      `dac_s${stage}_snake_op`,\n      cur,\n      channels,\n      length,\n      dacKey(`decoder.model.${m}.block.0.alpha`),\n      snakeOut,\n    );\n\n    // .block.1: ConvTranspose1d(channels → outCh, k=2*rate, stride=rate, pad=ceil/2, out_pad=0)\n    const upOut = `dac_s${stage}_up`;\n    length = convTranspose(\n      `dac_s${stage}_convt`,\n      snakeOut,\n      dacKey(`decoder.model.${m}.block.1.weight`),\n      dacKey(`decoder.model.${m}.block.1.bias`),\n      channels,\n      outCh,\n      upK,\n      length,\n      rate,\n      upPad,\n      0,\n      upOut,\n    );\n    channels = outCh;\n\n    // .block.{2,3,4}: 3× ResidualUnit (dilations [1,3,9]) — sequential.\n    let h = upOut;\n    for (let r = 0; r < DAC_RESBLOCK_DILATIONS.length; r++) {\n      const ru = r + 2; // .block index of this residual unit\n      const dil = DAC_RESBLOCK_DILATIONS[r];\n      const p = `decoder.model.${m}.block.${ru}.block`;\n      const pad = Math.floor(((7 - 1) * dil) / 2);\n      // .block.0 Snake → .block.1 Conv1d(C→C, k7, dilation) → .block.2 Snake → .block.3 Conv1d(k1)\n      const ia = `dac_s${stage}_r${r}_ia`;\n      snake(`${ia}_op`, h, channels, length, dacKey(`${p}.0.alpha`), ia);\n      const ic = `dac_s${stage}_r${r}_ic`;\n      conv1dFull(\n        `${ic}_op`,\n        ia,\n        dacKey(`${p}.1.weight`),\n        dacKey(`${p}.1.bias`),\n        channels,\n        channels,\n        7,\n        length,\n        dil,\n        pad,\n        ic,\n      );\n      const sa = `dac_s${stage}_r${r}_sa`;\n      snake(`${sa}_op`, ic, channels, length, dacKey(`${p}.2.alpha`), sa);\n      const sc = `dac_s${stage}_r${r}_sc`;\n      conv1dFull(\n        `${sc}_op`,\n        sa,\n        dacKey(`${p}.3.weight`),\n        dacKey(`${p}.3.bias`),\n        channels,\n        channels,\n        1,\n        length,\n        1,\n        0,\n        sc,\n      );\n      // residual: h = h + sc (length-preserving convs → no center-crop)\n      const resAdd = `dac_s${stage}_r${r}_res`;\n      add(`${resAdd}_op`, h, sc, channels, length, resAdd);\n      h = resAdd;\n    }\n    cur = h;\n  }\n\n  // ── decoder.model.{N+1}: final Snake on the last channel count ──\n  // After N DecoderBlocks (decoder.model.1..N), the final Snake is at index N+1 and\n  // the output conv at N+2 (DAC.speech N=4 → model.5 / model.6; same for dac_44khz).\n  const finalSnakeIdx = rates.length + 1;\n  const finalConvIdx = rates.length + 2;\n  const postSnake = \"dac_post_snake\";\n  snake(\n    \"dac_post_snake_op\",\n    cur,\n    channels,\n    length,\n    dacKey(`decoder.model.${finalSnakeIdx}.alpha`),\n    postSnake,\n  );\n\n  // ── decoder.model.{N+2}: Conv1d(channels → 1, k7, pad3) ──\n  const pcm = \"dac_pcm\";\n  const pcmLen = conv1dFull(\n    \"dac_conv2\",\n    postSnake,\n    dacKey(`decoder.model.${finalConvIdx}.weight`),\n    dacKey(`decoder.model.${finalConvIdx}.bias`),\n    channels,\n    1,\n    7,\n    length,\n    1,\n    3,\n    pcm,\n  );\n  // NOTE: the final tanh is applied host-side after readback (Math.tanh per sample),\n  // matching DAC's decoder.model.{N+3} Tanh. PCM = tanh(dac_pcm).\n\n  return {\n    architecture: \"DacSpeechDecoder\",\n    config: {\n      hidden_size: decoderDim,\n      num_layers: rates.length,\n      num_heads: 0,\n      num_kv_heads: 0,\n      head_dim: 0,\n      intermediate_size: 0,\n      vocab_size: 0,\n      context_length: pcmLen,\n      rms_norm_eps: 0,\n      norm_type: \"rmsnorm\",\n      rope_base: 0,\n      rope_dim: 0,\n      kv_layout: \"LHSd\",\n      is_moe: false,\n      has_vision_tower: false,\n    },\n    capabilities: { text: true, vision: false, moe: false },\n    tensors,\n    nodes,\n    executionOrder,\n    inputs: [\"audio_codes\"],\n    outputs: [pcm],\n  };\n}\n\n/**\n * Final decoded PCM length (samples) for a T-frame code grid, via the DAC stage chain.\n * `rates` defaults to DAC.speech.v1.0 ([8,5,4,2]); pass [8,8,4,2] for Parler's dac_44khz.\n */\nexport function dacOutputLength(\n  numFrames: number,\n  rates: readonly number[] = DAC_DECODER_RATES,\n): number {\n  let length = numFrames; // after conv1 (k7 pad3 → length-preserving)\n  for (const rate of rates) {\n    length = convTransposeLen(length, rate, 2 * rate, Math.ceil(rate / 2), 0);\n  }\n  return length; // conv2 (k7 pad3) is length-preserving\n}\n\n/** Strided/dilated Conv1d output length: floor((L + 2*pad - dilation*(K-1) - 1)/stride) + 1. */\nfunction conv1dLen(L: number, K: number, stride: number, pad: number, dilation: number): number {\n  return Math.floor((L + 2 * pad - dilation * (K - 1) - 1) / stride) + 1;\n}\n\n/** Number of latent frames the encoder produces for `numSamples` of padded 24 kHz PCM. */\nexport function dacEncoderFrames(numSamples: number): number {\n  let length = numSamples; // after block.0 (k7 pad3 → length-preserving)\n  for (const rate of DAC_ENCODER_RATES) {\n    // 3 residual units are length-preserving; the downsample conv is k=2r, stride r, pad=ceil(r/2).\n    length = conv1dLen(length, 2 * rate, rate, Math.ceil(rate / 2), 1);\n  }\n  return length; // block.6 (k3 pad1) is length-preserving\n}\n\n// ════════════════════════════════════════════════════════════════════════════\n// DAC.speech.v1.0 ENCODER + RVQ-encode graph (raw audio → c1/c2 codes), the voice-\n// cloning mirror of the decoder graph above. Input: pcm [1, L] (f32, 24 kHz mono,\n// already loudness-normalized + hop-padded host-side). Output: codes [NCB, T] (u32,\n// codebook-major). The encoder is descript-audio-codec's Encoder: Conv1d(1→64,k7) →\n// 4 EncoderBlocks (downsample strides [2,4,5,8], channels 64→128→256→512→1024) →\n// Snake → Conv1d(1024→1024,k3) → latent z[1024,T]; then a single RVQEncode op runs\n// the residual argmin over the 2 codebooks. Reuses Conv1dFull / Snake1d / Add. Every\n// weight carries its canonical `dac.encoder.*` / `dac.quantizer.*` name (weight-norm\n// already folded by prep-dac-speech.py). Validated <1e-4 vs torch DAC.encode.\n// ════════════════════════════════════════════════════════════════════════════\n\nexport interface DacEncoderGraphOptions {\n  /** Number of (hop-padded) input PCM samples L. */\n  numSamples: number;\n}\n\nexport function generateDacSpeechEncoderGraph(opts: DacEncoderGraphOptions): ModelGraph {\n  const L0 = opts.numSamples;\n  const tensors: Record<string, TensorDesc> = {};\n  const nodes: OpNode[] = [];\n  const executionOrder: string[] = [];\n\n  const addTensor = (desc: TensorDesc): void => {\n    tensors[desc.name] = desc;\n  };\n  const addNode = (node: OpNode): void => {\n    nodes.push(node);\n    executionOrder.push(node.id);\n  };\n\n  // ── Input: pcm [1, L] (f32, 24 kHz mono, host-padded to a hop multiple) ──\n  addTensor({ name: \"pcm_in\", shape: [1, L0], dtype: \"f32\", storage: \"activation\" });\n\n  /** Emit a Conv1dFull (symmetric padding) and return its output length. */\n  function conv1dFull(\n    id: string,\n    input: string,\n    weightKey: string,\n    biasKey: string,\n    Cin: number,\n    Cout: number,\n    K: number,\n    L: number,\n    stride: number,\n    dilation: number,\n    padding: number,\n    output: string,\n  ): number {\n    const Lout = conv1dLen(L, K, stride, padding, dilation);\n    addTensor({\n      name: weightKey,\n      shape: [Cout, Cin, K],\n      dtype: \"f32\",\n      storage: \"constant\",\n      safetensorsKey: weightKey,\n    });\n    addTensor({\n      name: biasKey,\n      shape: [Cout],\n      dtype: \"f32\",\n      storage: \"constant\",\n      safetensorsKey: biasKey,\n    });\n    addTensor({ name: output, shape: [Cout, Lout], dtype: \"f32\", storage: \"activation\" });\n    addNode({\n      id,\n      opType: \"Conv1dFull\",\n      inputs: [input, weightKey, biasKey],\n      outputs: [output],\n      attributes: { Cin, Cout, L, Lout, K, stride, padding, dilation },\n    });\n    return Lout;\n  }\n\n  /** Emit a full-channel Snake1d (alpha is [C], stored [1,C,1] → C). */\n  function snake(\n    id: string,\n    input: string,\n    C: number,\n    L: number,\n    alphaKey: string,\n    output: string,\n  ): void {\n    addTensor({\n      name: alphaKey,\n      shape: [C],\n      dtype: \"f32\",\n      storage: \"constant\",\n      safetensorsKey: alphaKey,\n    });\n    addTensor({ name: output, shape: [C, L], dtype: \"f32\", storage: \"activation\" });\n    addNode({\n      id,\n      opType: \"Snake1d\",\n      inputs: [input, alphaKey],\n      outputs: [output],\n      attributes: { C, L },\n    });\n  }\n\n  /** Element-wise residual add of two equal-shape [C, L] activations. */\n  function add(id: string, a: string, b: string, C: number, L: number, output: string): void {\n    addTensor({ name: output, shape: [C, L], dtype: \"f32\", storage: \"activation\" });\n    addNode({\n      id,\n      opType: \"Add\",\n      inputs: [a, b],\n      outputs: [output],\n      attributes: { count_tensor: a, hidden_size: C * L },\n    });\n  }\n\n  const encKey = (name: string): string => dacKey(`encoder.${name}`);\n\n  // ── block.0: Conv1d(1 → 64, k7, pad3) ──\n  let cur = \"dac_enc_conv0\";\n  let length = conv1dFull(\n    \"dac_enc_conv0_op\",\n    \"pcm_in\",\n    encKey(\"block.0.weight\"),\n    encKey(\"block.0.bias\"),\n    1,\n    DAC_ENCODER_DIM,\n    7,\n    L0,\n    1,\n    1,\n    3,\n    cur,\n  );\n  let channels = DAC_ENCODER_DIM;\n\n  // ── block.{1..4}: EncoderBlocks (3 residual units → Snake → strided downsample conv) ──\n  for (let stage = 0; stage < DAC_ENCODER_RATES.length; stage++) {\n    const m = stage + 1; // encoder.block index\n    const rate = DAC_ENCODER_RATES[stage];\n\n    // .block.{0,1,2}: 3× ResidualUnit (dilations [1,3,9]) — sequential, length-preserving.\n    let h = cur;\n    for (let r = 0; r < DAC_RESBLOCK_DILATIONS.length; r++) {\n      const dil = DAC_RESBLOCK_DILATIONS[r];\n      const p = `block.${m}.block.${r}.block`;\n      const pad = Math.floor(((7 - 1) * dil) / 2);\n      // .block.0 Snake → .block.1 Conv1d(C→C, k7, dilation) → .block.2 Snake → .block.3 Conv1d(k1)\n      const ia = `dac_enc_s${stage}_r${r}_ia`;\n      snake(`${ia}_op`, h, channels, length, encKey(`${p}.0.alpha`), ia);\n      const ic = `dac_enc_s${stage}_r${r}_ic`;\n      conv1dFull(\n        `${ic}_op`,\n        ia,\n        encKey(`${p}.1.weight`),\n        encKey(`${p}.1.bias`),\n        channels,\n        channels,\n        7,\n        length,\n        1,\n        dil,\n        pad,\n        ic,\n      );\n      const sa = `dac_enc_s${stage}_r${r}_sa`;\n      snake(`${sa}_op`, ic, channels, length, encKey(`${p}.2.alpha`), sa);\n      const sc = `dac_enc_s${stage}_r${r}_sc`;\n      conv1dFull(\n        `${sc}_op`,\n        sa,\n        encKey(`${p}.3.weight`),\n        encKey(`${p}.3.bias`),\n        channels,\n        channels,\n        1,\n        length,\n        1,\n        1,\n        0,\n        sc,\n      );\n      const resAdd = `dac_enc_s${stage}_r${r}_res`;\n      add(`${resAdd}_op`, h, sc, channels, length, resAdd);\n      h = resAdd;\n    }\n\n    // .block.3: Snake on `channels`\n    const snakeOut = `dac_enc_s${stage}_snake`;\n    snake(`${snakeOut}_op`, h, channels, length, encKey(`block.${m}.block.3.alpha`), snakeOut);\n\n    // .block.4: strided Conv1d(channels → 2*channels, k=2*rate, stride=rate, pad=ceil(rate/2))\n    const outCh = channels * 2;\n    const downOut = `dac_enc_s${stage}_down`;\n    length = conv1dFull(\n      `${downOut}_op`,\n      snakeOut,\n      encKey(`block.${m}.block.4.weight`),\n      encKey(`block.${m}.block.4.bias`),\n      channels,\n      outCh,\n      2 * rate,\n      length,\n      rate,\n      1,\n      Math.ceil(rate / 2),\n      downOut,\n    );\n    channels = outCh;\n    cur = downOut;\n  }\n\n  // ── block.5: final Snake on 1024 ──\n  const postSnake = \"dac_enc_post_snake\";\n  snake(\"dac_enc_post_snake_op\", cur, channels, length, encKey(\"block.5.alpha\"), postSnake);\n\n  // ── block.6: Conv1d(1024 → 1024, k3, pad1) → latent z[1024, T] ──\n  const latent = \"dac_enc_latent\";\n  const T = conv1dFull(\n    \"dac_enc_conv6_op\",\n    postSnake,\n    encKey(\"block.6.weight\"),\n    encKey(\"block.6.bias\"),\n    channels,\n    DAC_LATENT_DIM,\n    3,\n    length,\n    1,\n    1,\n    1,\n    latent,\n  );\n\n  // ── RVQ encode: latent z[1024, T] → codes[NCB, T] (residual argmin over codebooks) ──\n  addTensor({\n    name: dacKey(\"quantizer.in_proj.weight\"),\n    shape: [DAC_NUM_CODEBOOKS, DAC_CODEBOOK_DIM, DAC_LATENT_DIM],\n    dtype: \"f32\",\n    storage: \"constant\",\n    safetensorsKey: dacKey(\"quantizer.in_proj.weight\"),\n  });\n  addTensor({\n    name: dacKey(\"quantizer.in_proj.bias\"),\n    shape: [DAC_NUM_CODEBOOKS, DAC_CODEBOOK_DIM],\n    dtype: \"f32\",\n    storage: \"constant\",\n    safetensorsKey: dacKey(\"quantizer.in_proj.bias\"),\n  });\n  addTensor({\n    name: dacKey(\"quantizer.codebook\"),\n    shape: [DAC_NUM_CODEBOOKS, DAC_CODEBOOK_SIZE, DAC_CODEBOOK_DIM],\n    dtype: \"f32\",\n    storage: \"constant\",\n    safetensorsKey: dacKey(\"quantizer.codebook\"),\n  });\n  addTensor({\n    name: dacKey(\"quantizer.out_proj.weight\"),\n    shape: [DAC_NUM_CODEBOOKS, DAC_LATENT_DIM, DAC_CODEBOOK_DIM],\n    dtype: \"f32\",\n    storage: \"constant\",\n    safetensorsKey: dacKey(\"quantizer.out_proj.weight\"),\n  });\n  addTensor({\n    name: dacKey(\"quantizer.out_proj.bias\"),\n    shape: [DAC_NUM_CODEBOOKS, DAC_LATENT_DIM],\n    dtype: \"f32\",\n    storage: \"constant\",\n    safetensorsKey: dacKey(\"quantizer.out_proj.bias\"),\n  });\n  addTensor({\n    name: \"dac_codes\",\n    shape: [DAC_NUM_CODEBOOKS, T],\n    dtype: \"u32\",\n    storage: \"activation\",\n  });\n  addNode({\n    id: \"dac_rvq_encode\",\n    opType: \"RVQEncode\",\n    inputs: [\n      latent,\n      dacKey(\"quantizer.in_proj.weight\"),\n      dacKey(\"quantizer.in_proj.bias\"),\n      dacKey(\"quantizer.codebook\"),\n      dacKey(\"quantizer.out_proj.weight\"),\n      dacKey(\"quantizer.out_proj.bias\"),\n    ],\n    outputs: [\"dac_codes\"],\n    attributes: {\n      NCB: DAC_NUM_CODEBOOKS,\n      Cout: DAC_LATENT_DIM,\n      T,\n      D: DAC_CODEBOOK_DIM,\n      codebook_size: DAC_CODEBOOK_SIZE,\n    },\n  });\n\n  return {\n    architecture: \"DacSpeechEncoder\",\n    config: {\n      hidden_size: DAC_LATENT_DIM,\n      num_layers: DAC_ENCODER_RATES.length,\n      num_heads: 0,\n      num_kv_heads: 0,\n      head_dim: 0,\n      intermediate_size: 0,\n      vocab_size: 0,\n      context_length: T,\n      rms_norm_eps: 0,\n      norm_type: \"rmsnorm\",\n      rope_base: 0,\n      rope_dim: 0,\n      kv_layout: \"LHSd\",\n      is_moe: false,\n      has_vision_tower: false,\n    },\n    capabilities: { text: true, vision: false, moe: false },\n    tensors,\n    nodes,\n    executionOrder,\n    inputs: [\"pcm_in\"],\n    outputs: [\"dac_codes\"],\n  };\n}\n\n// ════════════════════════════════════════════════════════════════════════════\n// OuteTTS codec-LM backbone graph (plain Qwen3-0.6B + KV cache).\n//\n// A standard Qwen3ForCausalLM decoder with an INCREMENTAL KV-cache attention path\n// (KVCacheAppend + cached Attention), so the AR speak() loop decodes one token per\n// forward instead of re-prefilling. The generic Qwen2/3 generator (qwen2.ts) is\n// prefill-only; this is the same block math (RMSNorm → QKV → QK-RMSNorm → full RoPE\n// → GQA → SwiGLU) wired for cached decode, matching the Kani/Gemma backbone pattern.\n// Tied embeddings; full-vocab logits (the AR driver picks c1/c2 on the host). f32.\n// ════════════════════════════════════════════════════════════════════════════\n\nexport function generateOuteTtsBackboneGraph(\n  rawConfig: Record<string, unknown>,\n  dtype?: GraphDType,\n  groupSize?: number,\n  kvDtype?: KVDType,\n): ModelGraph {\n  const cfg = parseOuteTtsConfig(rawConfig);\n  const useQ4 = dtype === \"q4\";\n  const groupSz = groupSize ?? DEFAULT_GROUP_SIZE;\n  const hidden_size = cfg.hidden_size;\n  const num_layers = cfg.num_layers;\n  const num_heads = cfg.num_heads;\n  const num_kv_heads = cfg.num_kv_heads;\n  const head_dim = cfg.head_dim;\n  const rope_dim = head_dim; // Qwen3 uses full-head RoPE\n  const rope_base = cfg.rope_base;\n  const rms_norm_eps = cfg.rms_norm_eps;\n  const intermediate_size = cfg.intermediate_size;\n  const vocab_size = cfg.vocab_size;\n  const q_dim = num_heads * head_dim;\n  const kv_dim = num_kv_heads * head_dim;\n  const kvCacheDtype: DType = kvDtype ?? \"f32\";\n\n  const config: ModelArchConfig = {\n    hidden_size,\n    num_layers,\n    num_heads,\n    num_kv_heads,\n    head_dim,\n    intermediate_size,\n    vocab_size,\n    context_length: cfg.max_position_embeddings,\n    rms_norm_eps,\n    norm_type: \"rmsnorm\",\n    rope_base,\n    rope_dim,\n    kv_layout: \"LHSd\",\n    is_moe: false,\n    has_vision_tower: false,\n  };\n\n  const tensors: Record<string, TensorDesc> = {};\n  const nodes: OpNode[] = [];\n  const executionOrder: string[] = [];\n  const addTensor = (desc: TensorDesc): void => {\n    tensors[desc.name] = desc;\n  };\n  const addNode = (node: OpNode): void => {\n    nodes.push(node);\n    executionOrder.push(node.id);\n  };\n  // Linear projections: MatMulInt4 (weight-only W4A16) when q4, else plain MatMul. The\n  // q4 weight tensors (.q/.scales/.zeros) are produced by the loader from the f32 source\n  // under `weightName` (see model-loader.ts quantizeKaniBackbone, which is generic over\n  // MatMulInt4 nodes). The tied embedding + LM head are kept f32.\n  const addLinear = (\n    id: string,\n    activationInput: string,\n    weightName: string,\n    weightShape: number[],\n    outputName: string,\n    K: number,\n    N: number,\n  ): void => {\n    if (useQ4) {\n      const totalElements = K * N;\n      addTensor({\n        name: `${weightName}.q`,\n        shape: [Math.ceil(totalElements / 8)],\n        dtype: \"u32\",\n        storage: \"constant\",\n      });\n      addTensor({\n        name: `${weightName}.scales`,\n        shape: [Math.ceil(totalElements / groupSz)],\n        dtype: \"f32\",\n        storage: \"constant\",\n      });\n      addTensor({\n        name: `${weightName}.zeros`,\n        shape: [Math.ceil(totalElements / groupSz)],\n        dtype: \"f32\",\n        storage: \"constant\",\n      });\n      addNode({\n        id,\n        opType: \"MatMulInt4\",\n        inputs: [activationInput, `${weightName}.q`, `${weightName}.scales`, `${weightName}.zeros`],\n        outputs: [outputName],\n        attributes: { M_tensor: activationInput, K, N, group_size: groupSz },\n      });\n      return;\n    }\n    addTensor({\n      name: weightName,\n      shape: weightShape,\n      dtype: \"f32\",\n      storage: \"constant\",\n      safetensorsKey: weightName,\n    });\n    addNode({\n      id,\n      opType: \"MatMul\",\n      inputs: [activationInput, weightName],\n      outputs: [outputName],\n      attributes: { M_tensor: activationInput, K, N },\n    });\n  };\n\n  // ── Input + embedding ──\n  addTensor({ name: \"input_ids\", shape: [\"T\"], dtype: \"u32\", storage: \"activation\" });\n  addTensor({ name: \"embed_out\", shape: [\"T\", hidden_size], dtype: \"f32\", storage: \"activation\" });\n  addTensor({\n    name: CANONICAL_KEYS.EMBED,\n    shape: [vocab_size, hidden_size],\n    dtype: \"f32\",\n    storage: \"constant\",\n    safetensorsKey: CANONICAL_KEYS.EMBED,\n  });\n  addNode({\n    id: \"embed\",\n    opType: \"Embedding\",\n    inputs: [\"input_ids\", CANONICAL_KEYS.EMBED],\n    outputs: [\"embed_out\"],\n    attributes: { vocab_size, hidden_size },\n  });\n\n  let prevOutput = \"embed_out\";\n  for (let i = 0; i < num_layers; i++) {\n    const prefix = `layer${i}`;\n\n    // ── input_layernorm ──\n    const normW = CANONICAL_KEYS.layerInputNorm(i);\n    addTensor({\n      name: normW,\n      shape: [hidden_size],\n      dtype: \"f32\",\n      storage: \"constant\",\n      safetensorsKey: normW,\n    });\n    const norm1Out = `${prefix}_norm1_out`;\n    addTensor({ name: norm1Out, shape: [\"T\", hidden_size], dtype: \"f32\", storage: \"activation\" });\n    addNode({\n      id: `${prefix}_norm1`,\n      opType: \"RMSNorm\",\n      inputs: [prevOutput, normW],\n      outputs: [norm1Out],\n      attributes: { hidden_size, eps: rms_norm_eps, seq_len_tensor: prevOutput },\n    });\n\n    // ── Q/K/V projections ──\n    const qOut = `${prefix}_q`;\n    const kOut = `${prefix}_k`;\n    const vOut = `${prefix}_v`;\n    addTensor({ name: qOut, shape: [\"T\", q_dim], dtype: \"f32\", storage: \"activation\" });\n    addTensor({ name: kOut, shape: [\"T\", kv_dim], dtype: \"f32\", storage: \"activation\" });\n    addTensor({ name: vOut, shape: [\"T\", kv_dim], dtype: \"f32\", storage: \"activation\" });\n    addLinear(\n      `${prefix}_q_proj`,\n      norm1Out,\n      CANONICAL_KEYS.qProj(i),\n      [q_dim, hidden_size],\n      qOut,\n      hidden_size,\n      q_dim,\n    );\n    addLinear(\n      `${prefix}_k_proj`,\n      norm1Out,\n      CANONICAL_KEYS.kProj(i),\n      [kv_dim, hidden_size],\n      kOut,\n      hidden_size,\n      kv_dim,\n    );\n    addLinear(\n      `${prefix}_v_proj`,\n      norm1Out,\n      CANONICAL_KEYS.vProj(i),\n      [kv_dim, hidden_size],\n      vOut,\n      hidden_size,\n      kv_dim,\n    );\n\n    // ── QK per-head RMSNorm (Qwen3) ──\n    const qNormW = CANONICAL_KEYS.qNorm(i);\n    const kNormW = CANONICAL_KEYS.kNorm(i);\n    addTensor({\n      name: qNormW,\n      shape: [head_dim],\n      dtype: \"f32\",\n      storage: \"constant\",\n      safetensorsKey: qNormW,\n    });\n    addTensor({\n      name: kNormW,\n      shape: [head_dim],\n      dtype: \"f32\",\n      storage: \"constant\",\n      safetensorsKey: kNormW,\n    });\n    const qNormed = `${prefix}_q_normed`;\n    const kNormed = `${prefix}_k_normed`;\n    addTensor({ name: qNormed, shape: [\"T\", q_dim], dtype: \"f32\", storage: \"activation\" });\n    addTensor({ name: kNormed, shape: [\"T\", kv_dim], dtype: \"f32\", storage: \"activation\" });\n    addNode({\n      id: `${prefix}_q_norm`,\n      opType: \"RMSNorm\",\n      inputs: [qOut, qNormW],\n      outputs: [qNormed],\n      attributes: { hidden_size: head_dim, eps: rms_norm_eps, seq_len_tensor: qOut },\n    });\n    addNode({\n      id: `${prefix}_k_norm`,\n      opType: \"RMSNorm\",\n      inputs: [kOut, kNormW],\n      outputs: [kNormed],\n      attributes: { hidden_size: head_dim, eps: rms_norm_eps, seq_len_tensor: kOut },\n    });\n\n    // ── Full RoPE on Q and K (in place) ──\n    addNode({\n      id: `${prefix}_rope`,\n      opType: \"RoPE\",\n      inputs: [qNormed, kNormed],\n      outputs: [qNormed, kNormed],\n      attributes: {\n        head_dim,\n        num_q_heads: num_heads,\n        num_kv_heads,\n        rope_base,\n        rope_dim,\n        seq_len_tensor: qNormed,\n      },\n    });\n\n    // ── KV cache append (RoPE'd K, raw V) ──\n    const kCache = `${prefix}_k_cache`;\n    const vCache = `${prefix}_v_cache`;\n    addTensor({ name: kCache, shape: [\"L_max\", kv_dim], dtype: kvCacheDtype, storage: \"kv_cache\" });\n    addTensor({ name: vCache, shape: [\"L_max\", kv_dim], dtype: kvCacheDtype, storage: \"kv_cache\" });\n    addNode({\n      id: `${prefix}_kv_cache_k`,\n      opType: \"KVCacheAppend\",\n      inputs: [kNormed],\n      outputs: [kCache],\n      attributes: { width: kv_dim, T_tensor: kNormed },\n    });\n    addNode({\n      id: `${prefix}_kv_cache_v`,\n      opType: \"KVCacheAppend\",\n      inputs: [vOut],\n      outputs: [vCache],\n      attributes: { width: kv_dim, T_tensor: vOut },\n    });\n\n    // ── Causal GQA attention (reads KV cache) ──\n    const attnOut = `${prefix}_attn_out`;\n    addTensor({ name: attnOut, shape: [\"T\", q_dim], dtype: \"f32\", storage: \"activation\" });\n    addNode({\n      id: `${prefix}_attn`,\n      opType: \"Attention\",\n      inputs: [qNormed, kCache, vCache],\n      outputs: [attnOut],\n      attributes: {\n        hidden_size: q_dim,\n        num_q_heads: num_heads,\n        num_kv_heads,\n        head_dim,\n        causal: true,\n        layer_index: i,\n      },\n    });\n\n    // ── O projection ──\n    const oProjOut = `${prefix}_o_proj_out`;\n    addTensor({ name: oProjOut, shape: [\"T\", hidden_size], dtype: \"f32\", storage: \"activation\" });\n    addLinear(\n      `${prefix}_o_proj`,\n      attnOut,\n      CANONICAL_KEYS.oProj(i),\n      [hidden_size, q_dim],\n      oProjOut,\n      q_dim,\n      hidden_size,\n    );\n\n    // ── Residual 1 + post_attention_layernorm (fused) ──\n    const resid1Out = `${prefix}_resid1`;\n    const postNormW = CANONICAL_KEYS.layerPostAttnNorm(i);\n    const norm2Out = `${prefix}_norm2_out`;\n    addTensor({ name: resid1Out, shape: [\"T\", hidden_size], dtype: \"f32\", storage: \"activation\" });\n    addTensor({\n      name: postNormW,\n      shape: [hidden_size],\n      dtype: \"f32\",\n      storage: \"constant\",\n      safetensorsKey: postNormW,\n    });\n    addTensor({ name: norm2Out, shape: [\"T\", hidden_size], dtype: \"f32\", storage: \"activation\" });\n    addNode({\n      id: `${prefix}_resid1_norm2`,\n      opType: \"ResidualRMSNorm\",\n      inputs: [prevOutput, oProjOut, postNormW],\n      outputs: [resid1Out, norm2Out],\n      attributes: { hidden_size, eps: rms_norm_eps, seq_len_tensor: prevOutput },\n    });\n\n    // ── SwiGLU MLP ──\n    const gateOut = `${prefix}_gate_out`;\n    const upOut = `${prefix}_up_out`;\n    const swigluOut = `${prefix}_swiglu_out`;\n    const mlpOut = `${prefix}_mlp_out`;\n    addTensor({\n      name: gateOut,\n      shape: [\"T\", intermediate_size],\n      dtype: \"f32\",\n      storage: \"activation\",\n    });\n    addTensor({\n      name: upOut,\n      shape: [\"T\", intermediate_size],\n      dtype: \"f32\",\n      storage: \"activation\",\n    });\n    addTensor({\n      name: swigluOut,\n      shape: [\"T\", intermediate_size],\n      dtype: \"f32\",\n      storage: \"activation\",\n    });\n    addTensor({ name: mlpOut, shape: [\"T\", hidden_size], dtype: \"f32\", storage: \"activation\" });\n    addLinear(\n      `${prefix}_gate`,\n      norm2Out,\n      CANONICAL_KEYS.gateProj(i),\n      [intermediate_size, hidden_size],\n      gateOut,\n      hidden_size,\n      intermediate_size,\n    );\n    addLinear(\n      `${prefix}_up`,\n      norm2Out,\n      CANONICAL_KEYS.upProj(i),\n      [intermediate_size, hidden_size],\n      upOut,\n      hidden_size,\n      intermediate_size,\n    );\n    addNode({\n      id: `${prefix}_swiglu`,\n      opType: \"SwiGLU\",\n      inputs: [gateOut, upOut],\n      outputs: [swigluOut],\n      attributes: { count_tensor: gateOut },\n    });\n    addLinear(\n      `${prefix}_down`,\n      swigluOut,\n      CANONICAL_KEYS.downProj(i),\n      [hidden_size, intermediate_size],\n      mlpOut,\n      intermediate_size,\n      hidden_size,\n    );\n\n    // ── Residual 2 ──\n    const resid2Out = `${prefix}_resid2`;\n    addTensor({ name: resid2Out, shape: [\"T\", hidden_size], dtype: \"f32\", storage: \"activation\" });\n    addNode({\n      id: `${prefix}_resid2`,\n      opType: \"Add\",\n      inputs: [resid1Out, mlpOut],\n      outputs: [resid2Out],\n      attributes: { count_tensor: resid1Out, hidden_size },\n    });\n    prevOutput = resid2Out;\n  }\n\n  // ── Final norm ──\n  addTensor({\n    name: CANONICAL_KEYS.FINAL_NORM,\n    shape: [hidden_size],\n    dtype: \"f32\",\n    storage: \"constant\",\n    safetensorsKey: CANONICAL_KEYS.FINAL_NORM,\n  });\n  addTensor({\n    name: \"final_norm_out\",\n    shape: [\"T\", hidden_size],\n    dtype: \"f32\",\n    storage: \"activation\",\n  });\n  addNode({\n    id: \"final_norm\",\n    opType: \"RMSNorm\",\n    inputs: [prevOutput, CANONICAL_KEYS.FINAL_NORM],\n    outputs: [\"final_norm_out\"],\n    attributes: { hidden_size, eps: rms_norm_eps, seq_len_tensor: prevOutput },\n  });\n\n  // ── LM head (tied embeddings): last-row slice → full-vocab logits ──\n  addTensor({\n    name: \"final_norm_last\",\n    shape: [1, hidden_size],\n    dtype: \"f32\",\n    storage: \"activation\",\n  });\n  addNode({\n    id: \"slice_last_row\",\n    opType: \"SliceLastRow\",\n    inputs: [\"final_norm_out\"],\n    outputs: [\"final_norm_last\"],\n    attributes: { width: hidden_size },\n  });\n  addTensor({ name: \"logits\", shape: [1, vocab_size], dtype: \"f32\", storage: \"activation\" });\n  addNode({\n    id: \"lm_head\",\n    opType: \"MatMul\",\n    inputs: [\"final_norm_last\", CANONICAL_KEYS.EMBED],\n    outputs: [\"logits\"],\n    attributes: { M_tensor: \"final_norm_last\", K: hidden_size, N: vocab_size },\n  });\n\n  return {\n    architecture: \"Qwen3ForCausalLM\",\n    config,\n    capabilities: { text: true, vision: false, moe: false },\n    tensors,\n    nodes,\n    executionOrder,\n    inputs: [\"input_ids\"],\n    outputs: [\"logits\"],\n  };\n}\n","/**\n * OuteTTS voice-clone helpers — prosody feature extraction + word alignment for\n * `OuteTTS.encodeSpeaker` (raw audio → OuteSpeaker). These are a faithful port of\n * outetts v3 `audio_processor.Features` (energy / spectral_centroid / pitch) and the\n * `create_speaker_from_dict` word loop, so the per-word feature tokens the prompt\n * builder emits match the reference speaker JSONs the presets ship.\n *\n * The DAC encode (audio → c1/c2 codes) is done bit-exact on the GPU\n * (generateDacSpeechEncoderGraph + RVQEncode, validated 450/450 vs torch). The codes\n * are EXACT; only the word→frame boundaries are estimated. outetts uses Whisper\n * word-level timestamps to slice the code stream per word. We provide two on-device\n * estimators (both leave the codes untouched — they only decide which frames map to\n * which word, and which audio span each word's prosody is computed over):\n *\n *   1. alignWordsByEnergy (default) — a forced-segmentation aligner driven by the\n *      clip's OWN short-time energy envelope. It finds the silence gaps (low-energy\n *      runs) between utterances and snaps word boundaries onto them, distributing the\n *      transcript words across the detected speech runs (char-weighted within a run).\n *      This is audio-driven: boundaries land on real pauses instead of being guessed.\n *\n *   2. alignWordsProportional (fallback) — the original audio-BLIND estimator that\n *      spreads frames across words proportionally to char length. Used when the energy\n *      envelope yields no usable structure (one continuous speech run, or fewer gaps\n *      than word boundaries).\n *\n * On-device STT note: Gerbil's Moonshine STT (MoonshineSTT, src/gpu/moonshine-stt.ts)\n * is an encoder-decoder ASR that emits a transcript but does NOT expose Whisper-style\n * word timestamps — its decoder cross-attention scores are consumed inside the kernel\n * and never read back, so there is no per-token→frame map to slice on. Energy-based\n * forced segmentation is therefore the closest achievable on-device alignment; STT is\n * used (when wired by the caller) to obtain/confirm the WORD SEQUENCE, and the energy\n * aligner places the boundaries. See `OuteTTS.encodeSpeaker`.\n */\n\nconst DAC_TPS = 75; // DAC.speech latent frame rate (24000 / hop 320).\n\nconst WHITESPACE_RE = /\\s+/;\n\nconst PITCH_MIN_FREQ = 75.0;\nconst PITCH_MAX_FREQ = 600.0;\nconst PITCH_FRAME_LEN = 400;\nconst PITCH_HOP = 160;\nconst PITCH_THRESHOLD = 0.3;\nconst FEATURE_EPS = 1e-10;\n\n/** Round a [0,1] value to the [0,100] integer bucket the prompt builder expects. */\nfunction scaleValue(value: number): number {\n  return Math.round(Math.min(Math.max(value, 0), 1) * 100);\n}\n\n/** Naive DFT magnitude spectrum of a real signal (length N → ⌊N/2⌋+1 bins). */\nfunction rfftMag(x: Float32Array): Float32Array {\n  const N = x.length;\n  const bins = Math.floor(N / 2) + 1;\n  const mag = new Float32Array(bins);\n  for (let k = 0; k < bins; k++) {\n    let re = 0;\n    let im = 0;\n    const w = (-2 * Math.PI * k) / N;\n    for (let n = 0; n < N; n++) {\n      const a = w * n;\n      re += x[n] * Math.cos(a);\n      im += x[n] * Math.sin(a);\n    }\n    mag[k] = Math.hypot(re, im);\n  }\n  return mag;\n}\n\n/**\n * RMS energy of a clip, normalized to [0,1] (the reference clamps via scaleValue).\n * Matches `Features.extract_audio_features` energy = sqrt(mean(x^2)).\n */\nfunction rmsEnergy(audio: Float32Array): number {\n  let sum = 0;\n  for (let i = 0; i < audio.length; i++) {\n    sum += audio[i] * audio[i];\n  }\n  return audio.length ? Math.sqrt(sum / audio.length) : 0;\n}\n\n/**\n * Spectral centroid normalized to [0,1] of (sr/2). Matches the reference:\n * centroid = Σ f·|rfft(x)| / Σ|rfft(x)|, then / (sr/2).\n */\nfunction spectralCentroid(audio: Float32Array, sr: number): number {\n  if (audio.length === 0) {\n    return 0;\n  }\n  const spec = rfftMag(audio);\n  const nyquist = sr / 2;\n  const bins = spec.length;\n  let specSum = FEATURE_EPS;\n  let weighted = 0;\n  for (let k = 0; k < bins; k++) {\n    // freqs = linspace(0, sr/2, bins): freq[k] = k/(bins-1) * nyquist.\n    const freq = bins > 1 ? (k / (bins - 1)) * nyquist : 0;\n    weighted += freq * spec[k];\n    specSum += spec[k];\n  }\n  return weighted / specSum / nyquist;\n}\n\n/** Hann window of length N (periodic=false, matching torch.hann_window default). */\nfunction hannWindow(N: number): Float32Array {\n  const w = new Float32Array(N);\n  for (let n = 0; n < N; n++) {\n    w[n] = 0.5 - 0.5 * Math.cos((2 * Math.PI * n) / (N - 1));\n  }\n  return w;\n}\n\n/**\n * Autocorrelation-based average pitch normalized to [0,1] over [min,max] freq.\n * Faithful port of `calculate_pitch` + `extract_single_pitch_value`: framed Hann-\n * windowed autocorrelation, per-frame parabolic-interpolated peak in the valid lag\n * range, voicing threshold, then the mean over frames mapped to [0,1].\n */\n/** Hann-windowed autocorrelation of one frame into `autocorr` (lag 0..frameLen-1). */\nfunction frameAutocorr(\n  padded: Float32Array,\n  off: number,\n  window: Float32Array,\n  frame: Float32Array,\n  autocorr: Float32Array,\n): void {\n  for (let i = 0; i < PITCH_FRAME_LEN; i++) {\n    frame[i] = padded[off + i] * window[i];\n  }\n  for (let lag = 0; lag < PITCH_FRAME_LEN; lag++) {\n    let acc = 0;\n    for (let i = 0; i + lag < PITCH_FRAME_LEN; i++) {\n      acc += frame[i] * frame[i + lag];\n    }\n    autocorr[lag] = acc;\n  }\n}\n\n/** Voiced pitch (Hz) of one autocorrelation frame, or 0 if unvoiced. */\nfunction framePitch(autocorr: Float32Array, minIdx: number, maxIdx: number, sr: number): number {\n  let peakIdx = minIdx;\n  let peakVal = Number.NEGATIVE_INFINITY;\n  for (let lag = minIdx; lag < maxIdx; lag++) {\n    if (autocorr[lag] > peakVal) {\n      peakVal = autocorr[lag];\n      peakIdx = lag;\n    }\n  }\n  // Parabolic interpolation for sub-sample accuracy.\n  const idx = Math.min(Math.max(peakIdx, 1), PITCH_FRAME_LEN - 2);\n  const alpha = autocorr[idx - 1];\n  const beta = autocorr[idx];\n  const gamma = autocorr[idx + 1];\n  let delta = (0.5 * (alpha - gamma)) / (alpha - 2 * beta + gamma + 1e-8);\n  if (!(peakIdx > 0 && peakIdx < PITCH_FRAME_LEN - 1)) {\n    delta = 0;\n  }\n  const period = (peakIdx + delta) / sr;\n  let pitch = period > 0 ? 1 / period : 0;\n  const voiced = peakVal / (autocorr[0] + 1e-8) > PITCH_THRESHOLD;\n  if (!voiced) {\n    pitch = 0;\n  }\n  // Match torch: clamp AFTER the voicing mask, so unvoiced (0) clamps up to min_freq.\n  return Math.min(Math.max(pitch, PITCH_MIN_FREQ), PITCH_MAX_FREQ);\n}\n\nfunction averagePitch(audio: Float32Array, sr: number): number {\n  const N = audio.length;\n  if (N < PITCH_FRAME_LEN) {\n    return 0;\n  }\n  const padLen = (PITCH_FRAME_LEN - (N % PITCH_HOP)) % PITCH_HOP;\n  const padded = new Float32Array(N + padLen);\n  padded.set(audio);\n  const numFrames = Math.floor((padded.length - PITCH_FRAME_LEN) / PITCH_HOP) + 1;\n  if (numFrames <= 0) {\n    return 0;\n  }\n  const window = hannWindow(PITCH_FRAME_LEN);\n  const minIdx = Math.max(1, Math.floor(sr / PITCH_MAX_FREQ));\n  const maxIdx = Math.min(PITCH_FRAME_LEN, Math.floor(sr / PITCH_MIN_FREQ));\n\n  const frame = new Float32Array(PITCH_FRAME_LEN);\n  const autocorr = new Float32Array(PITCH_FRAME_LEN);\n  let pitchSum = 0;\n  for (let f = 0; f < numFrames; f++) {\n    frameAutocorr(padded, f * PITCH_HOP, window, frame, autocorr);\n    pitchSum += framePitch(autocorr, minIdx, maxIdx, sr);\n  }\n  const avg = pitchSum / numFrames;\n  return (avg - PITCH_MIN_FREQ) / (PITCH_MAX_FREQ - PITCH_MIN_FREQ);\n}\n\nexport interface OuteProsodyFeatures {\n  energy: number;\n  spectral_centroid: number;\n  pitch: number;\n}\n\n/**\n * Port of `Features.extract_audio_features`: energy (RMS), spectral_centroid (FFT\n * centroid), pitch (autocorrelation), each scaled to the [0,100] integer bucket.\n * Returns all-zero defaults for an empty/invalid clip (matching the reference).\n */\nexport function extractAudioFeatures(audio: Float32Array, sr: number): OuteProsodyFeatures {\n  if (audio.length === 0) {\n    return { energy: 0, spectral_centroid: 0, pitch: 0 };\n  }\n  return {\n    energy: scaleValue(rmsEnergy(audio)),\n    spectral_centroid: scaleValue(spectralCentroid(audio, sr)),\n    pitch: scaleValue(averagePitch(audio, sr)),\n  };\n}\n\nexport interface AlignedWord {\n  word: string;\n  startFrame: number;\n  endFrame: number;\n}\n\n/**\n * Pragmatic word→frame alignment (no on-device Whisper): split the transcript into\n * whitespace words and distribute the `numFrames` encoded frames across them\n * proportionally to each word's character length. Returns contiguous [start,end)\n * frame spans covering the whole clip. This stands in for outetts' Whisper word-\n * timestamp slicing — the codes themselves are exact; only the per-word boundaries\n * are approximate.\n */\nexport function alignWordsProportional(transcript: string, numFrames: number): AlignedWord[] {\n  const words = transcript.split(WHITESPACE_RE).filter((w) => w.length > 0);\n  if (words.length === 0) {\n    return [];\n  }\n  const weights = words.map((w) => w.length);\n  const totalWeight = weights.reduce((a, b) => a + b, 0);\n  const out: AlignedWord[] = [];\n  let acc = 0;\n  for (let i = 0; i < words.length; i++) {\n    const startFrame = i === 0 ? 0 : Math.round((acc / totalWeight) * numFrames);\n    acc += weights[i];\n    const endFrame =\n      i === words.length - 1 ? numFrames : Math.round((acc / totalWeight) * numFrames);\n    out.push({ word: words[i], startFrame, endFrame: Math.max(endFrame, startFrame) });\n  }\n  return out;\n}\n\n/**\n * Short-time RMS energy envelope sampled at the DAC frame rate (one value per code\n * frame). frame f covers audio samples [f·hop, f·hop + win). Used by the energy\n * aligner to locate silence gaps. `hop` is DAC_HOP (320 samples ≈ 13.3 ms @ 24 kHz),\n * matching the DAC frame stride so envelope index == code-frame index.\n */\nexport function frameEnergyEnvelope(\n  audio: Float32Array,\n  hop: number,\n  numFrames: number,\n): Float32Array {\n  const env = new Float32Array(numFrames);\n  const win = hop; // one-hop window keeps env index aligned 1:1 with code frames.\n  for (let f = 0; f < numFrames; f++) {\n    const start = f * hop;\n    const end = Math.min(start + win, audio.length);\n    let sum = 0;\n    let n = 0;\n    for (let i = start; i < end; i++) {\n      sum += audio[i] * audio[i];\n      n++;\n    }\n    env[f] = n > 0 ? Math.sqrt(sum / n) : 0;\n  }\n  return env;\n}\n\n/**\n * A contiguous run of speech frames [start, end) detected in the energy envelope.\n */\ninterface SpeechRun {\n  start: number;\n  end: number;\n}\n\n/**\n * Segment the energy envelope into speech runs separated by silence gaps. A frame is\n * \"silent\" when its energy is below `floor + (peak-floor)*silenceRatio`; a run of\n * silent frames at least `minGapFrames` long is treated as a word/phrase boundary.\n * Leading/trailing silence is trimmed. Returns the speech runs in order.\n */\nfunction detectSpeechRuns(\n  env: Float32Array,\n  silenceRatio: number,\n  minGapFrames: number,\n): SpeechRun[] {\n  const numFrames = env.length;\n  if (numFrames === 0) {\n    return [];\n  }\n  let peak = 0;\n  let floor = Number.POSITIVE_INFINITY;\n  for (let f = 0; f < numFrames; f++) {\n    peak = Math.max(peak, env[f]);\n    floor = Math.min(floor, env[f]);\n  }\n  if (peak <= 0) {\n    return [];\n  }\n  const threshold = floor + (peak - floor) * silenceRatio;\n  const runs: SpeechRun[] = [];\n  let runStart = -1;\n  let silentRun = 0;\n  for (let f = 0; f < numFrames; f++) {\n    const voiced = env[f] > threshold;\n    if (voiced) {\n      if (runStart < 0) {\n        runStart = f;\n      }\n      silentRun = 0;\n    } else {\n      silentRun++;\n      // Close the current run only once the silence is long enough to be a real gap.\n      if (runStart >= 0 && silentRun >= minGapFrames) {\n        runs.push({ start: runStart, end: f - silentRun + 1 });\n        runStart = -1;\n      }\n    }\n  }\n  if (runStart >= 0) {\n    runs.push({ start: runStart, end: numFrames });\n  }\n  return runs.filter((r) => r.end > r.start);\n}\n\nconst ENERGY_SILENCE_RATIO = 0.15; // frames under 15% of dynamic range are \"silence\".\nconst ENERGY_MIN_GAP_FRAMES = 4; // ≥4 DAC frames (~53 ms) of silence = a boundary.\n\n/**\n * Distribute `numWords` whole words across speech runs proportionally to each run's\n * frame length, giving every run ≥1 word and summing to exactly `numWords`. Surplus\n * (from the per-run floor) is handed to the longest runs first; any overshoot is taken\n * back from the shortest runs (never below 1). Returns the per-run word count.\n */\nfunction distributeWordCounts(runLens: number[], numWords: number): number[] {\n  const totalRunLen = runLens.reduce((a, b) => a + b, 0) || 1;\n  const targets = runLens.map((len) => Math.max(1, Math.floor((len / totalRunLen) * numWords)));\n  const order = runLens.map((_, i) => i).sort((a, b) => runLens[b] - runLens[a]);\n  let assigned = targets.reduce((a, b) => a + b, 0);\n  let oi = 0;\n  while (assigned < numWords) {\n    targets[order[oi % order.length]]++;\n    assigned++;\n    oi++;\n  }\n  while (assigned > numWords) {\n    const idx = order[order.length - 1 - (oi % order.length)];\n    if (targets[idx] > 1) {\n      targets[idx]--;\n      assigned--;\n    }\n    oi++;\n  }\n  return targets;\n}\n\n/**\n * Split a contiguous frame span [spanStart, spanEnd) among `groupWords` char-\n * proportionally, appending the resulting {@link AlignedWord}s to `out`. Boundaries\n * are contiguous within the span (the last word reaches spanEnd exactly).\n */\nfunction splitRunWords(\n  groupWords: string[],\n  spanStart: number,\n  spanEnd: number,\n  out: AlignedWord[],\n): void {\n  const span = Math.max(1, spanEnd - spanStart);\n  const chars = groupWords.map((w) => w.length || 1);\n  const groupTotal = chars.reduce((a, b) => a + b, 0) || groupWords.length;\n  let acc = 0;\n  for (let g = 0; g < groupWords.length; g++) {\n    const wStart = g === 0 ? spanStart : spanStart + Math.round((acc / groupTotal) * span);\n    acc += chars[g];\n    const wEnd =\n      g === groupWords.length - 1 ? spanEnd : spanStart + Math.round((acc / groupTotal) * span);\n    out.push({ word: groupWords[g], startFrame: wStart, endFrame: Math.max(wEnd, wStart) });\n  }\n}\n\n/**\n * Forced word→frame alignment driven by the clip's energy envelope. Detects the\n * silence gaps in `audio` (at DAC frame resolution) and distributes the transcript's\n * words across the detected speech runs so that word boundaries land on real pauses.\n * Within a run, words are split char-proportionally (same local rule as the fallback).\n * Run spans are made contiguous across the silence gaps (a run owns the frames up to\n * the next run's start) so the per-word c1/c2 slices still tile the full [0, numFrames)\n * code stream with no dropped frames.\n *\n * Falls back to {@link alignWordsProportional} when the envelope has no usable\n * structure (≤1 speech run). The DAC codes are never touched; only the boundaries move.\n */\nexport function alignWordsByEnergy(\n  transcript: string,\n  numFrames: number,\n  audio: Float32Array,\n  hop: number,\n): AlignedWord[] {\n  const words = transcript.split(WHITESPACE_RE).filter((w) => w.length > 0);\n  if (words.length === 0) {\n    return [];\n  }\n  if (words.length === 1 || numFrames <= 0) {\n    return alignWordsProportional(transcript, numFrames);\n  }\n\n  const env = frameEnergyEnvelope(audio, hop, numFrames);\n  const runs = detectSpeechRuns(env, ENERGY_SILENCE_RATIO, ENERGY_MIN_GAP_FRAMES);\n  // Need at least 2 runs to beat the audio-blind estimator; otherwise fall back.\n  if (runs.length < 2) {\n    return alignWordsProportional(transcript, numFrames);\n  }\n\n  const runLens = runs.map((r) => r.end - r.start);\n  const targets = distributeWordCounts(runLens, words.length);\n\n  // Walk the words run-by-run, splitting each run's contiguous frame span among its\n  // words. The span extends to the next run's start (or the clip end), so the silence\n  // frames are absorbed into the preceding word and the slices tile [0, numFrames).\n  const out: AlignedWord[] = [];\n  let wordIdx = 0;\n  for (let r = 0; r < runs.length; r++) {\n    const spanStart = r === 0 ? 0 : runs[r].start;\n    const spanEnd = r === runs.length - 1 ? numFrames : runs[r + 1].start;\n    splitRunWords(words.slice(wordIdx, wordIdx + targets[r]), spanStart, spanEnd, out);\n    wordIdx += targets[r];\n  }\n  return out;\n}\n\nexport { DAC_TPS };\n","/**\n * OuteTTS — native text-to-speech engine for Gerbil's WebGPU backend.\n *\n * OuteTTS-1.0-0.6B (OuteAI/OuteTTS-1.0-0.6B) is a two-stage codec-LM TTS, the same\n * shape as Kani-TTS but with a plain Qwen3-0.6B backbone and the descript-audio-codec\n * DAC.speech.v1.0 decoder (24 kHz, 2 codebooks):\n *\n *   1. CODEC-LM BACKBONE (Qwen3-0.6B): autoregressively emits interleaved audio codes\n *      <|c1_i|>/<|c2_i|> (one pair per 512-sample frame) into the same vocab as text,\n *      between <|audio_start|> and <|audio_end|>. Standard Qwen3 block math\n *      (RMSNorm/QK-RMSNorm/full RoPE/GQA/SwiGLU) with a KV-cache decode path\n *      (generateOuteTtsBackboneGraph) — simpler than Kani's LFM2 hybrid.\n *   2. DAC.speech.v1.0 DECODER: RVQ dequant (2× codebook lookup + out_proj) → latent\n *      [1024,T] → DacDecoder (Conv1d + 4 upsample blocks + Snake + tanh) → 24 kHz PCM.\n *      Validated bit-exact vs the torch DAC reference (scripts/engine/test-outetts-codec.mjs,\n *      latent max|err| 2.9e-6, PCM max|err| 1.3e-6, cos 1.0).\n *\n * The AR loop runs on the host with full-logit readback; each step samples the next\n * audio token constrained to the current codebook's valid window (c1 on even\n * positions, c2 on odd), until <|audio_end|>. The collected tokens are split into a\n * [2, T] code grid (audioTokensToDacCodes) and decoded once through the DAC graph.\n *\n * This is the FOUNDATION: a single preset voice (OuteTTS's bundled en-female-1-neutral\n * speaker, embedded as in-context reference codes). Cloning, int4, the speak()/useTTS\n * public surface, R2 mirror, and the playground are a later phase.\n */\n\nimport {\n  audioTokensToDacCodes,\n  DAC_HOP,\n  DAC_NUM_CODEBOOKS,\n  DAC_SAMPLE_RATE,\n  dacEncoderFrames,\n  dacOutputLength,\n  generateDacSpeechDecoderGraph,\n  generateDacSpeechEncoderGraph,\n  generateOuteTtsBackboneGraph,\n  OUTETTS_AUDIO_END,\n  OUTETTS_IM_END,\n} from \"./architectures/outetts.js\";\nimport type { GPUContext } from \"./device.js\";\nimport { initGPU } from \"./device.js\";\nimport { Executor } from \"./executor.js\";\nimport { type LoadedOuteTTS, loadOuteTTS, quantizeBackboneInt4 } from \"./model-loader.js\";\nimport { alignWordsByEnergy, DAC_TPS, extractAudioFeatures } from \"./outetts-encode.js\";\nimport type { Tokenizer } from \"./tokenizer.js\";\n\nconst DEFAULT_MAX_NEW_TOKENS = 4096; // 2 tokens/frame → up to ~2048 frames (~27 s)\n\n/** A single word's reference codes in a speaker prompt (the bundled default voice). */\nexport interface OuteSpeakerWord {\n  word: string;\n  duration?: number;\n  c1: number[];\n  c2: number[];\n  features?: { energy?: number; spectral_centroid?: number; pitch?: number };\n}\nexport interface OuteSpeaker {\n  text: string;\n  words: OuteSpeakerWord[];\n}\n\n// ── OuteTTS v3 prompt-string builders (verbatim port of outetts'\n//    PromptProcessor: text_normalizations / merge_speaker_text / create_codes /\n//    get_completion_prompt). The whole prompt is assembled as a STRING with the\n//    literal <|...|> special-token markers, then tokenized in ONE pass — Gerbil's\n//    tokenizer splits on every added token, mirroring HF AutoTokenizer.encode(\n//    prompt, add_special_tokens=False). This reproduces the reference token layout\n//    EXACTLY (time/energy/spectral_centroid/pitch features per word), which the\n//    earlier foundation omitted (the cause of the incoherent speech). ──\n\nconst CONTROL_CHARS_RE =\n  // biome-ignore lint/suspicious/noControlCharactersInRegex: matches the reference control-char set\n  /[\\x00-\\x1F\\x7F-\\x9F­​-‍]/g;\nconst HIRAGANA_RE = /[぀-ゟ]/;\nconst KATAKANA_RE = /[゠-ヿ]/;\nconst HAN_RE = /[一-鿿]/;\nconst HTTP_URL_RE = /^https?:\\/\\//;\nconst TRAILING_SLASH_RE = /\\/$/;\n\n/** Port of PromptProcessor.text_normalizations (the ftfy step is approximated). */\nfunction normalizeOuteText(input: string): string {\n  let text = input.normalize(\"NFKC\");\n  text = text.replace(/…/g, \"...\");\n  text = text.replace(/\\.{2,}/g, \"...\");\n  text = text.replace(/[“”„‟«»]/g, '\"');\n  text = text.replace(/[‘’‛‹›`´]/g, \"'\");\n  text = text.replace(/[–—―−‐]/g, \"-\");\n  text = text.replace(/-{2,}/g, \"-\");\n  text = text.replace(CONTROL_CHARS_RE, \"\");\n  text = text.replace(/\\s+/g, \" \");\n  text = text.replace(/\\s+([,.?!:;])/g, \"$1\");\n  text = text.replace(/([,.?!:;])(?=[^\\s,.?!:;])/g, \"$1 \");\n  text = text.replace(/(\\w)\\s+'\\s*(\\w)/g, \"$1'$2\");\n  text = text.replace(/(\\w)\\s+'\\s*([,.?!:;\\s]|$)/g, \"$1'$2\");\n  text = text.replace(/(\\w)\\s*'\\s*([tsdmreSNTDMRE])\\b/gi, \"$1'$2\");\n  text = text.replace(/([?!])\\1+/g, \"$1\");\n  text = text.replace(/\\s+/g, \" \");\n  text = text.replace(/\"/g, \"\").trim();\n  return text;\n}\n\n/** Port of PromptProcessor._get_separator (CJK-aware). */\nfunction outeSeparator(text: string): string {\n  const hasHiragana = HIRAGANA_RE.test(text);\n  const hasKatakana = KATAKANA_RE.test(text);\n  const hasHan = HAN_RE.test(text);\n  if (hasHiragana || hasKatakana || hasHan) {\n    return \"。\";\n  }\n  return \". \";\n}\n\n/**\n * Port of PromptProcessor.merge_speaker_text: prepend the (normalized) speaker\n * transcript in front of the target text, joined by the right separator, and return\n * the trailing-separator string that outetts appends to the speaker's LAST word.\n */\nfunction mergeSpeakerText(inputText: string, speakerText: string): { merged: string; sep: string } {\n  const sp = speakerText.trim();\n  const separator = outeSeparator(sp);\n  const allowedEnds = separator === \"。\" ? [\"。\", \"？\", \"！\", \"?\", \"!\"] : [\".\", \"?\", \"!\"];\n  let rs = \"\";\n  if (sp) {\n    const last = sp[sp.length - 1];\n    if (!allowedEnds.includes(last)) {\n      rs = separator;\n    } else if (separator !== \"。\") {\n      rs = \" \";\n    }\n  }\n  return { merged: sp + rs + inputText.trim(), sep: rs.trim() };\n}\n\n/** Format one word's feature tokens (energy / spectral_centroid / pitch). */\nfunction outeFeatureTokens(f?: {\n  energy?: number;\n  spectral_centroid?: number;\n  pitch?: number;\n}): string {\n  const energy = f?.energy ?? 0;\n  const sc = f?.spectral_centroid ?? 0;\n  const pitch = f?.pitch ?? 0;\n  return `<|energy_${energy}|><|spectral_centroid_${sc}|><|pitch_${pitch}|>`;\n}\n\n/** Port of PromptProcessor.create_codes: one <|word_start|>…<|word_end|> block per word, \"\\n\"-joined. */\nfunction createSpeakerCodes(words: OuteSpeakerWord[]): string {\n  const blocks: string[] = [];\n  for (const w of words) {\n    const duration = (w.duration ?? 0).toFixed(2);\n    let block = `${w.word}<|features|><|t_${duration}|>${outeFeatureTokens(w.features)}`;\n    const n = Math.min(w.c1.length, w.c2.length);\n    let pairs = \"\";\n    for (let i = 0; i < n; i++) {\n      pairs += `<|c1_${w.c1[i]}|><|c2_${w.c2[i]}|>`;\n    }\n    block += `<|code|>${pairs}`;\n    blocks.push(`<|word_start|>${block}<|word_end|>`);\n  }\n  return blocks.join(\"\\n\");\n}\n\n/**\n * Build the OuteTTS v3 completion prompt STRING (outetts get_completion_prompt). The\n * whole prompt — text section, speaker word blocks (word + time/energy/spectral/pitch\n * features + interleaved c1/c2 codes), and the dangling <|word_start|> cue — is\n * assembled with literal <|...|> markers; the caller tokenizes it in one pass. Exposed\n * for token-level validation against the Python reference (test-outetts-prompt.mjs).\n */\nexport function buildOutePromptString(text: string, speaker?: OuteSpeaker): string {\n  const normalized = normalizeOuteText(text);\n  let promptText = normalized;\n  let codes = \"\";\n  if (speaker) {\n    const { merged, sep } = mergeSpeakerText(normalized, speaker.text);\n    promptText = merged;\n    // outetts appends the trailing separator to the LAST speaker word's `word`.\n    const words = speaker.words.map((w, i) =>\n      i === speaker.words.length - 1 ? { ...w, word: w.word + sep } : w,\n    );\n    codes = createSpeakerCodes(words);\n  }\n  let prompt = `<|im_start|>\\n<|text_start|>${promptText}<|text_end|>\\n<|audio_start|>\\n`;\n  if (speaker) {\n    prompt += `${codes}\\n<|word_start|>`;\n  }\n  return prompt;\n}\n\n/**\n * Load an OuteTTS preset speaker (the in-context reference voice). Accepts a full URL\n * (mirror), a local file path (Node), or a bare preset name resolved against\n * `baseURL/{name}.json`. The returned object is the OuteSpeaker the prompt builder\n * embeds (transcript + per-word c1/c2 reference codes + prosody features).\n */\nexport async function loadOuteSpeaker(voiceOrUrl: string, baseURL?: string): Promise<OuteSpeaker> {\n  // Already a parsed speaker passed through? (defensive — callers pass strings)\n  if (HTTP_URL_RE.test(voiceOrUrl)) {\n    const resp = await fetch(voiceOrUrl);\n    if (!resp.ok) {\n      throw new Error(`Failed to fetch OuteTTS speaker ${voiceOrUrl}: ${resp.status}`);\n    }\n    return (await resp.json()) as OuteSpeaker;\n  }\n  // Local file path (Node): read + parse.\n  if (voiceOrUrl.includes(\"/\") || voiceOrUrl.endsWith(\".json\")) {\n    const fs = await import(\"node:fs\").catch(() => null);\n    if (fs) {\n      return JSON.parse(fs.readFileSync(voiceOrUrl, \"utf8\")) as OuteSpeaker;\n    }\n  }\n  // Bare preset name → resolve against the mirror base URL.\n  if (!baseURL) {\n    throw new Error(\n      `loadOuteSpeaker: \"${voiceOrUrl}\" is a preset name but no baseURL was provided.`,\n    );\n  }\n  const url = `${baseURL.replace(TRAILING_SLASH_RE, \"\")}/${voiceOrUrl}.json`;\n  const resp = await fetch(url);\n  if (!resp.ok) {\n    throw new Error(`Failed to fetch OuteTTS preset \"${voiceOrUrl}\" from ${url}: ${resp.status}`);\n  }\n  return (await resp.json()) as OuteSpeaker;\n}\n\nexport interface OuteTTSOptions {\n  /** Backbone repo (default OuteAI/OuteTTS-1.0-0.6B). */\n  repo?: string;\n  /**\n   * Folded DAC.speech codec weights: a local path to scripts/engine/prep-dac-speech.py's\n   * output, or a mirror URL. Required (the official DAC repo ships only `.pth`).\n   */\n  codecWeightsPath: string;\n  /**\n   * The preset speaker (in-context reference). Pass OuteTTS's bundled\n   * en-female-1-neutral.json (text + words[].c1/c2). When omitted, synthesis runs\n   * text-only (the model free-runs a less-stable voice).\n   */\n  speaker?: OuteSpeaker;\n  /**\n   * Backbone precision. \"q4\" (default) loads/runs the 1.2 GB f32 backbone as group-wise\n   * weight-only INT4 (~400 MB, tied embedding kept f32). \"f32\" keeps full precision.\n   */\n  dtype?: OuteDType;\n  revision?: string;\n  hfToken?: string;\n  cacheDir?: string;\n  maxSeqLen?: number;\n  onProgress?: (loaded: number, total: number, message: string) => void;\n}\n\nexport type OuteDType = \"q4\" | \"f32\";\n\nexport interface OuteSpeakOptions {\n  /** Sampling temperature (OuteTTS default 0.4). */\n  temperature?: number;\n  /** Top-p nucleus threshold (default 0.9). */\n  topP?: number;\n  /** Top-k cutoff (default 40). */\n  topK?: number;\n  /** Repetition penalty (default 1.1) over the last `repetitionRange` tokens. */\n  repetitionPenalty?: number;\n  /** Repetition-penalty window (default 64). */\n  repetitionRange?: number;\n  /** Override max generated tokens (default 4096). */\n  maxNewTokens?: number;\n  /**\n   * Greedy (argmax) decode over the full vocab, matching HF do_sample=False. Used for\n   * deterministic reference validation (test-outetts-greedy.mjs). Default false\n   * (sampled).\n   */\n  greedy?: boolean;\n  /**\n   * Override the in-context speaker for this call (voice cloning / preset swap). When\n   * set, this speaker's reference word blocks are embedded in the prompt instead of the\n   * one the engine was created with. Build it from a reference clip via\n   * {@link OuteTTS.encodeSpeaker} (transcript + DAC codes), or pass a preset JSON.\n   */\n  speaker?: OuteSpeaker;\n}\n\nexport interface OuteSpeakResult {\n  /** Mono PCM in [-1, 1]. */\n  pcm: Float32Array;\n  /** Sample rate (24000). */\n  sampleRate: number;\n  /** Number of audio frames decoded. */\n  frames: number;\n  /** Audio duration in seconds. */\n  audioSeconds: number;\n}\n\n/** Build a fresh map of just the constant weights a graph references (see kani-tts). */\nfunction selectGraphWeights(\n  graph: { tensors: Record<string, { storage: string; safetensorsKey?: 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\") {\n      continue;\n    }\n    const w =\n      weights.get(name) ?? (desc.safetensorsKey ? weights.get(desc.safetensorsKey) : undefined);\n    if (w) {\n      out.set(name, w);\n    }\n  }\n  return out;\n}\n\n/** Index of the maximum logit (greedy / argmax decode). */\nfunction argmax(logits: Float32Array): number {\n  let best = 0;\n  let bestV = logits[0];\n  for (let i = 1; i < logits.length; i++) {\n    if (logits[i] > bestV) {\n      bestV = logits[i];\n      best = i;\n    }\n  }\n  return best;\n}\n\n/** In-place softmax of `scores`. */\nfunction softmaxInPlace(scores: Float32Array): void {\n  let maxV = Number.NEGATIVE_INFINITY;\n  for (const v of scores) {\n    if (v > maxV) {\n      maxV = v;\n    }\n  }\n  let sum = 0;\n  for (let i = 0; i < scores.length; i++) {\n    const e = Math.exp(scores[i] - maxV);\n    scores[i] = e;\n    sum += e;\n  }\n  for (let i = 0; i < scores.length; i++) {\n    scores[i] /= sum;\n  }\n}\n\n/** Top-p nucleus sample from a softmaxed prob array over `ids`. */\nfunction nucleusSample(probs: Float32Array, ids: Int32Array, topP: number): number {\n  const n = probs.length;\n  const order = Array.from({ length: n }, (_, i) => i).sort((a, b) => probs[b] - probs[a]);\n  let cum = 0;\n  let cutoff = n;\n  for (let r = 0; r < n; r++) {\n    cum += probs[order[r]];\n    if (cum >= topP) {\n      cutoff = r + 1;\n      break;\n    }\n  }\n  let keepSum = 0;\n  for (let r = 0; r < cutoff; r++) {\n    keepSum += probs[order[r]];\n  }\n  let pick = Math.random() * keepSum;\n  for (let r = 0; r < cutoff; r++) {\n    pick -= probs[order[r]];\n    if (pick <= 0) {\n      return ids[order[r]];\n    }\n  }\n  return ids[order[0]];\n}\n\n/**\n * Select the `k` highest-scoring token ids in a single O(n·logk) pass (bounded\n * min-heap), applying the repetition penalty to `previous` tokens inline so a\n * penalized token can drop out of the shortlist. Returns the kept ids and their\n * (penalized, pre-temperature) logits, unsorted.\n *\n * This replaces the two full-vocab `Array.from(...).sort(...)` passes the sampler\n * ran PER TOKEN over OuteTTS's ~157k vocab (the top-k filter + the nucleus sort):\n * a codec-LM emits hundreds-to-thousands of tokens per utterance, so those two\n * O(n·logn) sorts — plus their two fresh 157k-element allocations every step —\n * dominated host-side decode and made the browser SPEECH tab look stalled\n * (\"SYNTHESIZING…\" for minutes). Restricting the softmax + nucleus to the kept\n * top-k is distribution-identical to the old full-vocab path (softmax is\n * monotonic and the discarded tail is renormalized away either way).\n */\nfunction selectTopKPenalized(\n  logits: Float32Array,\n  k: number,\n  penalized: Set<number> | null,\n  penalty: number,\n): { ids: Int32Array; vals: Float32Array } {\n  const n = logits.length;\n  const cap = Math.min(k, n);\n  const heapVal = new Float32Array(cap);\n  const heapId = new Int32Array(cap);\n  let size = 0;\n\n  const penalize = (i: number, v: number): number => {\n    if (penalized?.has(i)) {\n      return v > 0 ? v / penalty : v * penalty;\n    }\n    return v;\n  };\n  const siftDown = (): void => {\n    let root = 0;\n    for (;;) {\n      let smallest = root;\n      const l = 2 * root + 1;\n      const r = 2 * root + 2;\n      if (l < size && heapVal[l] < heapVal[smallest]) smallest = l;\n      if (r < size && heapVal[r] < heapVal[smallest]) smallest = r;\n      if (smallest === root) break;\n      const tv = heapVal[root];\n      const ti = heapId[root];\n      heapVal[root] = heapVal[smallest];\n      heapId[root] = heapId[smallest];\n      heapVal[smallest] = tv;\n      heapId[smallest] = ti;\n      root = smallest;\n    }\n  };\n  const siftUp = (node0: number): void => {\n    let node = node0;\n    while (node > 0) {\n      const parent = (node - 1) >> 1;\n      if (heapVal[parent] <= heapVal[node]) break;\n      const tv = heapVal[node];\n      const ti = heapId[node];\n      heapVal[node] = heapVal[parent];\n      heapId[node] = heapId[parent];\n      heapVal[parent] = tv;\n      heapId[parent] = ti;\n      node = parent;\n    }\n  };\n\n  for (let i = 0; i < n; i++) {\n    const v = penalize(i, logits[i]);\n    if (size < cap) {\n      heapVal[size] = v;\n      heapId[size] = i;\n      size++;\n      siftUp(size - 1);\n    } else if (v > heapVal[0]) {\n      heapVal[0] = v;\n      heapId[0] = i;\n      siftDown();\n    }\n  }\n  return { ids: heapId.subarray(0, size), vals: heapVal.subarray(0, size) };\n}\n\n/**\n * Resample interleaved-or-mono PCM to mono at `targetRate` via linear interpolation.\n * Audio is assumed already mono (Float32 [-1,1]); if `srcRate === targetRate` the\n * input is returned as-is. Linear resampling is sufficient for codec conditioning\n * (the DAC encoder is robust to the small interpolation error vs torchaudio's\n * sinc resampler — the cloned codes are stable).\n */\nfunction resampleMono(audio: Float32Array, srcRate: number, targetRate: number): Float32Array {\n  if (srcRate === targetRate) {\n    return audio;\n  }\n  const ratio = targetRate / srcRate;\n  const outLen = Math.max(1, Math.round(audio.length * ratio));\n  const out = new Float32Array(outLen);\n  for (let i = 0; i < outLen; i++) {\n    const srcPos = i / ratio;\n    const i0 = Math.floor(srcPos);\n    const i1 = Math.min(i0 + 1, audio.length - 1);\n    const frac = srcPos - i0;\n    out[i] = audio[i0] * (1 - frac) + audio[i1] * frac;\n  }\n  return out;\n}\n\nexport class OuteTTS {\n  private ctx: GPUContext;\n  private loaded: LoadedOuteTTS;\n  private tokenizer: Tokenizer;\n  private maxSeqLen: number;\n  private backboneExec: Executor;\n  private speaker?: OuteSpeaker;\n  private _destroyed = false;\n\n  readonly sampleRate = DAC_SAMPLE_RATE;\n  /** The selected backbone precision (\"q4\" or \"f32\"). */\n  readonly dtype: OuteDType;\n\n  private constructor(\n    ctx: GPUContext,\n    loaded: LoadedOuteTTS,\n    maxSeqLen: number,\n    dtype: OuteDType,\n    speaker?: OuteSpeaker,\n  ) {\n    this.ctx = ctx;\n    this.loaded = loaded;\n    this.tokenizer = loaded.tokenizer;\n    this.maxSeqLen = maxSeqLen;\n    this.speaker = speaker;\n    this.dtype = dtype;\n\n    const useQ4 = dtype === \"q4\";\n    const graph = generateOuteTtsBackboneGraph(loaded.rawConfig, useQ4 ? \"q4\" : undefined);\n    if (useQ4) {\n      // Quantize the Qwen3 linear projections to INT4 in place (tied embedding stays f32).\n      quantizeBackboneInt4(graph, loaded.backboneWeights);\n    }\n    const exec = new Executor(ctx, graph, { maxSeqLen, kvMode: \"f32\" });\n    exec.uploadWeightsMap(selectGraphWeights(graph, loaded.backboneWeights));\n    exec.initBindGroups();\n    this.backboneExec = exec;\n  }\n\n  static async create(options: OuteTTSOptions): Promise<OuteTTS> {\n    const ctx = await initGPU();\n    const loaded = await loadOuteTTS({\n      repo: options.repo,\n      codecWeightsPath: options.codecWeightsPath,\n      revision: options.revision,\n      hfToken: options.hfToken,\n      cacheDir: options.cacheDir,\n      onProgress: options.onProgress,\n    });\n    const ctxLen = (loaded.rawConfig.max_position_embeddings as number) ?? 40_960;\n    const maxSeqLen = Math.min(options.maxSeqLen ?? 4096, ctxLen);\n    return new OuteTTS(ctx, loaded, maxSeqLen, options.dtype ?? \"q4\", options.speaker);\n  }\n\n  /**\n   * Build the synthesis prompt as a STRING (outetts v3 get_completion_prompt) and\n   * tokenize in one pass. With a speaker, the speaker transcript is merged in front of\n   * the target text (merge_speaker_text), the speaker's reference word blocks\n   * (word + <|features|> + <|t_dur|> + energy/spectral_centroid/pitch + <|code|> +\n   * c1/c2 pairs) are embedded in-context, and the prompt ends with a dangling\n   * <|word_start|> to cue generation. Without a speaker, a plain text-only prompt.\n   *\n   *   <|im_start|>\\n<|text_start|>{text}<|text_end|>\\n<|audio_start|>\\n\n   *     [ {speaker word blocks \"\\n\"-joined} \\n <|word_start|> ]\n   *\n   * This reproduces the reference token layout EXACTLY (validated to the byte vs\n   * outetts.PromptProcessor — see scripts/engine/outetts-ref-prompt.py and\n   * test-outetts-prompt.mjs). The per-word time/feature tokens were the gating fix:\n   * the earlier foundation dropped them and the backbone produced incoherent audio.\n   */\n  private buildPrompt(text: string, speaker?: OuteSpeaker): number[] {\n    return this.tokenizer.encode(buildOutePromptString(text, speaker ?? this.speaker));\n  }\n\n  /**\n   * Synthesize speech for `text`. Returns 24 kHz mono PCM.\n   *\n   * Pipeline: build prompt → prefill backbone → AR-decode interleaved c1/c2 audio\n   * tokens until <|audio_end|> → split to a [2, T] DAC code grid → DAC decode → tanh\n   * → PCM.\n   */\n  async speak(text: string, opts: OuteSpeakOptions = {}): Promise<OuteSpeakResult> {\n    if (this._destroyed) {\n      throw new Error(\"OuteTTS has been destroyed.\");\n    }\n    const startTime = performance.now();\n    const audioTokens = await this.generateTokens(text, opts);\n\n    const { codes, numFrames } = audioTokensToDacCodes(audioTokens);\n    if (numFrames === 0) {\n      throw new Error(\n        `OuteTTS produced no complete audio frames (audioTokens=${audioTokens.length}).`,\n      );\n    }\n    const pcm = await this.decodeCodes(codes, numFrames);\n\n    const totalTime = (performance.now() - startTime) / 1000;\n    // True duration uses the actual decoded sample count (320·T − 8), not the nominal hop.\n    const audioSeconds = pcm.length / DAC_SAMPLE_RATE;\n    console.log(\n      `[outetts] frames=${numFrames} audio=${audioSeconds.toFixed(2)}s wall=${totalTime.toFixed(\n        2,\n      )}s RTF=${(audioSeconds / totalTime).toFixed(2)}x`,\n    );\n\n    return { pcm, sampleRate: DAC_SAMPLE_RATE, frames: numFrames, audioSeconds };\n  }\n\n  /**\n   * Build an in-context speaker reference for voice cloning.\n   *\n   * Two inputs are supported:\n   *  1. A pre-encoded {@link OuteSpeaker} (transcript + per-word c1/c2 DAC codes +\n   *     prosody features — the preset-JSON shape). Returned verbatim.\n   *  2. Raw audio (Float32 PCM + transcript). The clip is resampled to 24 kHz mono,\n   *     hop-padded, run through the DAC.speech ENCODER + RVQ-encode on the GPU\n   *     (generateDacSpeechEncoderGraph; encoder latent validated bit-exact and codes\n   *     450/450 vs torch DAC.encode — scripts/engine/test-outetts-encode.mjs), then\n   *     split into per-word {@link OuteSpeakerWord} blocks with prosody features.\n   *\n   * Fidelity vs outetts `create_speaker`: the DAC-encode (audio → c1/c2 codes) and the\n   * energy/spectral_centroid/pitch feature math are a faithful port. Two documented\n   * approximations: (a) word→frame ALIGNMENT — outetts uses Whisper word timestamps to\n   * slice the code stream. Gerbil's on-device Moonshine STT does not expose word\n   * timestamps (its decoder cross-attention is never read back), so we instead run an\n   * energy-forced aligner ({@link alignWordsByEnergy}) that snaps word boundaries onto\n   * the clip's own silence gaps — audio-driven, tighter than the old char-proportional\n   * guess, and exact when each word is pause-separated; it falls back to proportional\n   * for continuous speech. (b) we skip outetts' pyloudnorm pre-normalization (it mainly\n   * rescales absolute level / the energy bucket). The cloned voice timbre and prosody\n   * come through; only the per-word boundaries remain an estimate.\n   */\n  async encodeSpeaker(\n    ref: OuteSpeaker | { audio: Float32Array; sampleRate: number; transcript: string },\n  ): Promise<OuteSpeaker> {\n    if (\"words\" in ref && \"text\" in ref) {\n      return ref;\n    }\n    if (this._destroyed) {\n      throw new Error(\"OuteTTS has been destroyed.\");\n    }\n    const { audio, sampleRate, transcript } = ref;\n    if (!audio || audio.length === 0) {\n      throw new Error(\"OuteTTS.encodeSpeaker: empty audio.\");\n    }\n    if (!transcript || transcript.trim().length === 0) {\n      throw new Error(\"OuteTTS.encodeSpeaker: a transcript of the clip is required.\");\n    }\n\n    // Resample to 24 kHz mono and hop-pad (DAC.preprocess: ceil(L/hop)*hop).\n    const mono = resampleMono(audio, sampleRate, DAC_SAMPLE_RATE);\n    const paddedLen = Math.ceil(mono.length / DAC_HOP) * DAC_HOP;\n    const pcm = new Float32Array(paddedLen);\n    pcm.set(mono);\n\n    // DAC.speech encode (Conv downsample stack + RVQ argmin) → codes[NCB, T].\n    const numFrames = dacEncoderFrames(paddedLen);\n    const codes = await this.encodeAudioToCodes(pcm, numFrames);\n    const c1All = codes.subarray(0, numFrames);\n    const c2All = codes.subarray(numFrames, 2 * numFrames);\n\n    // Per-word blocks: energy-forced frame alignment (snaps boundaries to the clip's\n    // own silence gaps; falls back to char-proportional when the envelope is flat) +\n    // per-word prosody features sliced from the matching audio span.\n    const aligned = alignWordsByEnergy(transcript, numFrames, mono, DAC_HOP);\n    const words: OuteSpeakerWord[] = aligned.map((a) => {\n      const startSample = Math.min(a.startFrame * DAC_HOP, mono.length);\n      const endSample = Math.min(a.endFrame * DAC_HOP, mono.length);\n      const wordAudio = mono.subarray(startSample, endSample);\n      const features = extractAudioFeatures(wordAudio, DAC_SAMPLE_RATE);\n      return {\n        word: a.word,\n        duration: Math.round(((a.endFrame - a.startFrame) / DAC_TPS) * 100) / 100,\n        c1: Array.from(c1All.subarray(a.startFrame, a.endFrame)),\n        c2: Array.from(c2All.subarray(a.startFrame, a.endFrame)),\n        features,\n      };\n    });\n\n    return { text: transcript.trim(), words };\n  }\n\n  /** Run the DAC.speech encoder + RVQ-encode graph → codebook-major codes[NCB·T] (u32). */\n  private async encodeAudioToCodes(pcm: Float32Array, numFrames: number): Promise<Uint32Array> {\n    const graph = generateDacSpeechEncoderGraph({ numSamples: pcm.length });\n    const exec = new Executor(this.ctx, graph, { maxSeqLen: numFrames, kvMode: \"f32\" });\n    try {\n      exec.uploadWeightsMap(selectGraphWeights(graph, this.loaded.codecWeights));\n      exec.initBindGroups();\n      exec.reset();\n      exec.writeInput(\"pcm_in\", pcm);\n      // The codes buffer is u32; runGraphOutput returns an f32 view of the same bytes.\n      const raw = await exec.runGraphOutput(\"dac_codes\", DAC_NUM_CODEBOOKS * numFrames);\n      return new Uint32Array(raw.buffer, raw.byteOffset, DAC_NUM_CODEBOOKS * numFrames).slice();\n    } finally {\n      exec.destroy();\n    }\n  }\n\n  /** Build the prompt, prefill, and AR-decode the raw generated token stream. */\n  private async generateTokens(text: string, opts: OuteSpeakOptions): Promise<number[]> {\n    const prompt = this.buildPrompt(text, opts.speaker);\n    if (prompt.length >= this.maxSeqLen) {\n      throw new Error(`OuteTTS prompt is ${prompt.length} tokens >= maxSeqLen ${this.maxSeqLen}.`);\n    }\n    this.backboneExec.reset();\n    const { logits } = await this.backboneExec.forward(new Uint32Array(prompt));\n    return this.runDecodeLoop(prompt, logits, {\n      temperature: opts.temperature ?? 0.4,\n      topP: opts.topP ?? 0.9,\n      topK: opts.topK ?? 40,\n      repetitionPenalty: opts.repetitionPenalty ?? 1.1,\n      repetitionRange: opts.repetitionRange ?? 64,\n      maxNewTokens: Math.min(opts.maxNewTokens ?? DEFAULT_MAX_NEW_TOKENS, this.maxSeqLen),\n      greedy: opts.greedy ?? false,\n    });\n  }\n\n  /** Test hook: return the raw generated token stream (no DAC decode). */\n  async generateTokensForTest(text: string, opts: OuteSpeakOptions = {}): Promise<number[]> {\n    return this.generateTokens(text, opts);\n  }\n\n  /**\n   * Autoregressive decode: from the prefill logits, sample one token per step over the\n   * full vocab (greedy argmax, or temperature/top-k/top-p with repetition penalty).\n   * Stops at <|audio_end|>/<|im_end|> or the token cap. Returns the raw token stream\n   * (word-block structure + interleaved c1/c2 codes); audioTokensToDacCodes extracts\n   * the codes.\n   */\n  private async runDecodeLoop(\n    prompt: number[],\n    prefillLogits: Float32Array,\n    p: {\n      temperature: number;\n      topP: number;\n      topK: number;\n      repetitionPenalty: number;\n      repetitionRange: number;\n      maxNewTokens: number;\n      greedy: boolean;\n    },\n  ): Promise<number[]> {\n    const seq: number[] = [...prompt];\n    const generated: number[] = [];\n    let logits = prefillLogits;\n\n    for (let step = 0; step < p.maxNewTokens; step++) {\n      // OuteTTS does NOT decode in a rigid c1/c2 window: after the cued\n      // <|word_start|>, the backbone emits a full word block\n      // (word-text <|features|> <|t_..|> energy/spectral/pitch <|code|> c1 c2 … <|word_end|>)\n      // per word, then <|audio_end|>. So the loop samples over the FULL vocab (greedy\n      // argmax, or temp/top-k/top-p with repetition penalty) and lets the model produce\n      // its native structure; audioTokensToDacCodes() filters the c1/c2 codes out of the\n      // stream. (Constraining to the codebook window — the foundation's bug — starved the\n      // model of its structural tokens and produced incoherent audio.)\n      const next = p.greedy\n        ? argmax(logits)\n        : this.sampleToken(logits, {\n            temperature: p.temperature,\n            topP: p.topP,\n            topK: p.topK,\n            repetitionPenalty: p.repetitionPenalty,\n            previous: seq.slice(-p.repetitionRange),\n          });\n\n      if (next === OUTETTS_AUDIO_END || next === OUTETTS_IM_END) {\n        break;\n      }\n      generated.push(next);\n      seq.push(next);\n\n      if (seq.length >= this.maxSeqLen) {\n        break;\n      }\n      const res = await this.backboneExec.forward(new Uint32Array([next]));\n      logits = res.logits;\n    }\n    return generated;\n  }\n\n  /**\n   * Sample one token over the FULL vocab: repetition penalty over the recent window,\n   * then temperature → top-k → top-p nucleus. (OuteTTS emits structural word-block\n   * tokens interleaved with audio codes, so the sampler must not be window-restricted.)\n   */\n  private sampleToken(\n    logits: Float32Array,\n    p: {\n      temperature: number;\n      topP: number;\n      topK: number;\n      repetitionPenalty: number;\n      previous: number[];\n    },\n  ): number {\n    const n = logits.length;\n    const prev = p.repetitionPenalty !== 1.0 ? new Set(p.previous) : null;\n\n    // Fast path (the common case — the SDK/site default is topK=40): shortlist the\n    // top-k candidates in one O(n·logk) pass instead of two full-vocab sorts, then\n    // softmax + nucleus over just those k. Distribution-identical to the old path.\n    if (p.topK > 0 && p.topK < n) {\n      const { ids: kIds, vals: kVals } = selectTopKPenalized(\n        logits,\n        p.topK,\n        prev,\n        p.repetitionPenalty,\n      );\n      const scores = new Float32Array(kVals.length);\n      for (let i = 0; i < kVals.length; i++) {\n        scores[i] = kVals[i] / p.temperature;\n      }\n      softmaxInPlace(scores);\n      return nucleusSample(scores, Int32Array.from(kIds), p.topP);\n    }\n\n    // Fallback (top-k disabled): full-vocab temperature → softmax → nucleus.\n    const ids = new Int32Array(n);\n    const scores = new Float32Array(n);\n    for (let i = 0; i < n; i++) {\n      ids[i] = i;\n      let s = logits[i];\n      if (prev?.has(i)) {\n        s = s > 0 ? s / p.repetitionPenalty : s * p.repetitionPenalty;\n      }\n      scores[i] = s / p.temperature;\n    }\n    softmaxInPlace(scores);\n    return nucleusSample(scores, ids, p.topP);\n  }\n\n  /**\n   * Decode the [2, T] DAC code grid (codebook-major) → PCM, applying the final tanh.\n   * The DAC decoder is non-causal (symmetric padding) and its odd rate-5 transpose\n   * makes interior windows position-dependent, so the whole clip is decoded in ONE\n   * pass (no chunking). The codec conv/Snake/Add kernels use 2D dispatch (the\n   * dispatch2D helper) so even multi-second clips stay under WebGPU's 65535 per-dim\n   * workgroup cap.\n   */\n  private async decodeCodes(codes: Uint32Array, numFrames: number): Promise<Float32Array> {\n    const graph = generateDacSpeechDecoderGraph({ numFrames });\n    const pcmLen = dacOutputLength(numFrames);\n    const exec = new Executor(this.ctx, graph, { maxSeqLen: numFrames, kvMode: \"f32\" });\n    try {\n      exec.uploadWeightsMap(selectGraphWeights(graph, this.loaded.codecWeights));\n      exec.initBindGroups();\n      exec.reset();\n      exec.writeInput(\"audio_codes\", codes);\n      const raw = await exec.runGraphOutput(graph.outputs[0], pcmLen);\n      const pcm = new Float32Array(raw.length);\n      for (let i = 0; i < raw.length; i++) {\n        pcm[i] = Math.tanh(raw[i]);\n      }\n      return pcm;\n    } finally {\n      exec.destroy();\n    }\n  }\n\n  destroy(): void {\n    if (this._destroyed) {\n      return;\n    }\n    this._destroyed = true;\n    this.backboneExec.destroy();\n    this.loaded.backboneWeights.clear();\n    this.loaded.codecWeights.clear();\n  }\n}\n","/**\n * ParlerTTS — native describe-your-voice text-to-speech for Gerbil's WebGPU backend.\n *\n * Parler-TTS (parler-tts/parler-tts-mini-v1) is an encoder-decoder codec-LM: you\n * describe the voice in natural language and a Flan-T5 encoder turns that DESCRIPTION\n * into cross-attention states, while a 24-layer decoder LM cross-attends to them and\n * autoregressively emits DAC audio codes for the TRANSCRIPT. Same dual-graph shape as\n * MoonshineSTT (encoder once → frozen cross-attn K/V → AR decode), and the same DAC\n * codec family as OuteTTS (the shared generateDacSpeechDecoderGraph, here with 9\n * codebooks + rates [8,8,4,2]).\n *\n * Pipeline (engine.speak(text, { describeVoice })):\n *   1. Tokenize the DESCRIPTION with the T5 tokenizer; build the relative-position\n *      bias; run the T5 encoder once → encoder_out + per-decoder-layer frozen K/V.\n *   2. Tokenize the TRANSCRIPT (embed_prompts); prefill its embeddings (+ sinusoidal\n *      positions) through the decoder so the self-attn KV-cache holds the prompt.\n *   3. AR-decode 9 interleaved codebooks under the delay pattern: each step the host\n *      sums the 9 per-codebook embeddings of the current (delayed) codes + the\n *      sinusoidal position, forwards one step, and samples each of the 9 lm_head logit\n *      rows. Stop on EOS in codebook 0 (or the frame cap).\n *   4. Revert the delay pattern → [9, frames] code grid → shared DAC decode → 44.1 kHz\n *      PCM (tanh applied host-side).\n *\n * STATUS — WIRED + RUNS END-TO-END. Typechecks + builds, and the full pipeline runs\n * on Dawn (node): parler-tts/parler-tts-mini-v1 loads, the T5 encoder runs, the prompt\n * prefills, the 9-codebook delay-pattern AR loop decodes, and the shared DAC decoder\n * produces finite, non-silent, in-range 44.1 kHz PCM (verified via\n * scripts/engine/test-parler-speak.mjs). Reached from the high-level engine through\n * `g.speak(text, { model: DEFAULT_MODELS.ttsParler, describeVoice })`, which routes\n * Parler repos here directly (bypassing the single-graph text loader, the same way\n * MoonshineSTT is reached for STT). A bit-exact reference diff vs transformers\n * ParlerTTSForConditionalGeneration (T5 encoder cosine + a decoder forward; see\n * scripts/engine/parler-*-ref.py stubs) remains as a follow-up for numerical parity.\n * The architecture, weight layout, tokenization, delay pattern, and DAC config are all\n * verified against the live config.json + safetensors header.\n */\n\nimport { dacOutputLength, generateDacSpeechDecoderGraph } from \"./architectures/outetts.js\";\nimport {\n  buildT5RelativeBias,\n  generateParlerDecoderGraph,\n  generateParlerEncoderGraph,\n  PARLER_BOS_TOKEN_ID,\n  PARLER_DAC_DECODER_DIM,\n  PARLER_DAC_LATENT_DIM,\n  PARLER_DECODER_RATES,\n  PARLER_EOS_TOKEN_ID,\n  PARLER_SAMPLE_RATE,\n  parseParlerConfig,\n  revertDelayPattern,\n} from \"./architectures/parler.js\";\nimport {\n  createBindGroup,\n  createStorageBuffer,\n  createUniformBuffer,\n  destroyBuffers,\n  type GPUContext,\n  getOrCreatePipeline,\n  initGPU,\n} from \"./device.js\";\nimport { Executor } from \"./executor.js\";\nimport { DTYPE_BYTES, type ModelGraph, type OpNode } from \"./ir.js\";\nimport { KERNEL_REGISTRY, type KernelSpec } from \"./kernels/registry.js\";\nimport { type LoadedParlerTTS, loadParlerTTS } from \"./model-loader.js\";\n\nconst MAP_MODE_READ = 0x0001;\n/** Default voice description when the caller doesn't supply one. */\nconst DEFAULT_VOICE_DESCRIPTION =\n  \"A clear, neutral voice speaking at a moderate pace with good audio quality.\";\n/** Hard cap on decoded audio frames (~86 fps → ~30 s at 2600). */\nconst DEFAULT_MAX_FRAMES = 2600;\n\nexport interface ParlerTTSOptions {\n  /** Repo (default parler-tts/parler-tts-mini-v1). */\n  repo?: string;\n  revision?: string;\n  hfToken?: string;\n  cacheDir?: string;\n  maxSeqLen?: number;\n  onProgress?: (loaded: number, total: number, message: string) => void;\n}\n\nexport interface ParlerSpeakOptions {\n  /** Natural-language voice description (the Parler hook). */\n  describeVoice?: string;\n  /** Sampling temperature (Parler default 1.0). */\n  temperature?: number;\n  /** Top-k cutoff (default 0 = off; Parler uses sampling with do_sample=True). */\n  topK?: number;\n  /** Greedy (argmax per codebook) for deterministic reference validation. */\n  greedy?: boolean;\n  /** Override the max decoded frames. */\n  maxFrames?: number;\n  /**\n   * Debug hook: receives the reverted [numCodebooks, frames] code grid (codebook-major)\n   * just before DAC decode. Used by the reference token-match test; no production effect.\n   */\n  onCodes?: (codes: Uint32Array, numCodebooks: number, frames: number) => void;\n}\n\nexport interface ParlerSpeakResult {\n  /** Mono PCM in [-1, 1]. */\n  pcm: Float32Array;\n  /** Sample rate (44100). */\n  sampleRate: number;\n  /** Number of acoustic frames decoded. */\n  frames: number;\n  /** Audio duration in seconds. */\n  audioSeconds: number;\n}\n\n/** Build a fresh map of just the constant weights a graph references. */\nfunction selectGraphWeights(\n  graph: { tensors: Record<string, { storage: string; safetensorsKey?: 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\") {\n      continue;\n    }\n    const w =\n      weights.get(name) ?? (desc.safetensorsKey ? weights.get(desc.safetensorsKey) : undefined);\n    if (w) {\n      out.set(name, w);\n    }\n  }\n  return out;\n}\n\n/** Argmax over a logit row. */\nfunction argmax(logits: Float32Array): number {\n  let best = 0;\n  let bestV = logits[0];\n  for (let i = 1; i < logits.length; i++) {\n    if (logits[i] > bestV) {\n      bestV = logits[i];\n      best = i;\n    }\n  }\n  return best;\n}\n\n/** Temperature + optional top-k sample over a logit row. */\nfunction sampleLogits(logits: Float32Array, temperature: number, topK: number): number {\n  const n = logits.length;\n  const scaled = new Float32Array(n);\n  let maxV = Number.NEGATIVE_INFINITY;\n  const invT = 1 / Math.max(temperature, 1e-5);\n  for (let i = 0; i < n; i++) {\n    scaled[i] = logits[i] * invT;\n    if (scaled[i] > maxV) {\n      maxV = scaled[i];\n    }\n  }\n  // top-k mask (keep indices of the k largest).\n  let allowed: Int32Array | null = null;\n  if (topK > 0 && topK < n) {\n    const order = Array.from({ length: n }, (_, i) => i).sort((a, b) => scaled[b] - scaled[a]);\n    allowed = new Int32Array(order.slice(0, topK));\n  }\n  let sum = 0;\n  const probs = new Float32Array(n);\n  if (allowed) {\n    for (const idx of allowed) {\n      const e = Math.exp(scaled[idx] - maxV);\n      probs[idx] = e;\n      sum += e;\n    }\n  } else {\n    for (let i = 0; i < n; i++) {\n      const e = Math.exp(scaled[i] - maxV);\n      probs[i] = e;\n      sum += e;\n    }\n  }\n  let pick = Math.random() * sum;\n  if (allowed) {\n    for (const idx of allowed) {\n      pick -= probs[idx];\n      if (pick <= 0) {\n        return idx;\n      }\n    }\n    return allowed[0];\n  }\n  for (let i = 0; i < n; i++) {\n    pick -= probs[i];\n    if (pick <= 0) {\n      return i;\n    }\n  }\n  return n - 1;\n}\n\n// ════════════════════════════════════════════════════════════════════════════\n// Encoder executor — runs the T5 encoder graph once and reads back encoder_out +\n// the per-decoder-layer frozen cross-attn K/V (modeled on MoonshineEncoderExecutor).\n// ════════════════════════════════════════════════════════════════════════════\n\ninterface EncDispatch {\n  node: OpNode;\n  spec: KernelSpec;\n  pipeline: GPUComputePipeline;\n  bindGroup: GPUBindGroup;\n  uniformBuffer: GPUBuffer;\n}\n\nclass ParlerEncoderExecutor {\n  private ctx: GPUContext;\n  private graph: ModelGraph;\n  private decLayers: number;\n  private hidden: number;\n  private sDesc: number;\n  private weightBuffers = new Map<string, GPUBuffer>();\n  private activationBuffers = new Map<string, GPUBuffer>();\n  private dispatches: EncDispatch[] = [];\n\n  constructor(ctx: GPUContext, graph: ModelGraph, decLayers: number, sDesc: number) {\n    this.ctx = ctx;\n    this.graph = graph;\n    this.decLayers = decLayers;\n    this.hidden = graph.config.hidden_size;\n    this.sDesc = sDesc;\n    for (const [name, desc] of Object.entries(graph.tensors)) {\n      if (desc.storage !== \"activation\") {\n        continue;\n      }\n      // Resolve symbolic dims (e.g. input_ids' \"T\") to the concrete description length.\n      const count = desc.shape.reduce(\n        (a: number, dim) => a * (typeof dim === \"number\" ? dim : sDesc),\n        1,\n      );\n      this.activationBuffers.set(\n        name,\n        createStorageBuffer(this.ctx, `penc_${name}`, count * DTYPE_BYTES[desc.dtype]),\n      );\n    }\n  }\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\") {\n        continue;\n      }\n      const w = weights.get(name);\n      if (!w) {\n        throw new Error(`ParlerEncoderExecutor: missing weight \"${name}\"`);\n      }\n      this.weightBuffers.set(\n        name,\n        createStorageBuffer(this.ctx, `pencw_${name}`, w.data.byteLength, w.data),\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      const spec = KERNEL_REGISTRY[node.opType];\n      if (!spec) {\n        throw new Error(`ParlerEncoderExecutor: no kernel for op \"${node.opType}\"`);\n      }\n      const pipeline = getOrCreatePipeline(\n        this.ctx,\n        `penck_${nodeId}`,\n        spec.shaderCode,\n        spec.entryPoint,\n      );\n      const paramsData = spec.buildParams(node, shapes, { seqPos: 0 });\n      const uniformBuffer = createUniformBuffer(this.ctx, `pencu_${nodeId}`, paramsData);\n      const bufferEntries = this.gatherBuffers(spec, node, uniformBuffer);\n      const bindGroup = createBindGroup(this.ctx, pipeline, bufferEntries, `pencbg_${nodeId}`);\n      this.dispatches.push({ node, spec, pipeline, bindGroup, uniformBuffer });\n    }\n  }\n\n  /** Run the encoder. `inputIds` u32 [S], `relBias` f32 [heads,S,S]. */\n  async encode(\n    inputIds: Uint32Array,\n    relBias: Float32Array,\n  ): Promise<{ encK: Float32Array[]; encV: Float32Array[] }> {\n    const q = this.ctx.device.queue;\n    const inputBuf = this.activationBuffers.get(\"input_ids\");\n    const biasBuf = this.activationBuffers.get(\"t5_rel_bias\");\n    if (!(inputBuf && biasBuf)) {\n      throw new Error(\"ParlerEncoderExecutor: input_ids / t5_rel_bias buffers missing.\");\n    }\n    q.writeBuffer(inputBuf, 0, inputIds as BufferSource);\n    q.writeBuffer(biasBuf, 0, relBias as BufferSource);\n\n    const shapes = this.resolveShapes();\n    const sizes: Array<[number, number, number]> = [];\n    for (const d of this.dispatches) {\n      const params = d.spec.buildParams(d.node, shapes, { seqPos: 0 });\n      q.writeBuffer(d.uniformBuffer, 0, params);\n      sizes.push(d.spec.getDispatchSize(d.node, shapes, { seqPos: 0 }));\n    }\n\n    if (this.ctx.isWebKitWebGPU) {\n      for (let i = 0; i < this.dispatches.length; i++) {\n        const enc = this.ctx.device.createCommandEncoder();\n        const p = enc.beginComputePass();\n        p.setPipeline(this.dispatches[i].pipeline);\n        p.setBindGroup(0, this.dispatches[i].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: \"parler_encode\" });\n      const pass = enc.beginComputePass({ label: \"parler_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 encK: Float32Array[] = [];\n    const encV: Float32Array[] = [];\n    for (let i = 0; i < this.decLayers; i++) {\n      encK.push(await this.readBack(`enc_k_layer${i}`, this.sDesc * this.hidden));\n      encV.push(await this.readBack(`enc_v_layer${i}`, this.sDesc * this.hidden));\n    }\n    return { encK, encV };\n  }\n\n  destroy(): void {\n    destroyBuffers([...this.weightBuffers.values()]);\n    destroyBuffers([...this.activationBuffers.values()]);\n    for (const d of this.dispatches) {\n      d.uniformBuffer.destroy();\n    }\n    this.weightBuffers.clear();\n    this.activationBuffers.clear();\n    this.dispatches = [];\n  }\n\n  private resolveShapes(): 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((dim) => (typeof dim === \"number\" ? dim : this.sDesc));\n    }\n    return resolved;\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(`ParlerEncoderExecutor: no buffer for \"${tensorName}\" in op ${node.id}`);\n        }\n        entries.push({ buffer });\n      }\n    }\n    return entries;\n  }\n\n  private async readBack(name: string, elements: number): Promise<Float32Array> {\n    const buffer = this.activationBuffers.get(name);\n    if (!buffer) {\n      throw new Error(`ParlerEncoderExecutor: no buffer \"${name}\"`);\n    }\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\n// ════════════════════════════════════════════════════════════════════════════\n// ParlerTTS engine\n// ════════════════════════════════════════════════════════════════════════════\n\nexport class ParlerTTS {\n  private ctx: GPUContext;\n  private loaded: LoadedParlerTTS;\n  private cfg: ReturnType<typeof parseParlerConfig>;\n  private maxSeqLen: number;\n  private _destroyed = false;\n\n  readonly sampleRate = PARLER_SAMPLE_RATE;\n\n  private constructor(ctx: GPUContext, loaded: LoadedParlerTTS, maxSeqLen: number) {\n    this.ctx = ctx;\n    this.loaded = loaded;\n    this.cfg = parseParlerConfig(loaded.rawConfig);\n    this.maxSeqLen = maxSeqLen;\n  }\n\n  static async create(options: ParlerTTSOptions = {}): Promise<ParlerTTS> {\n    const ctx = await initGPU();\n    const loaded = await loadParlerTTS({\n      repo: options.repo,\n      revision: options.revision,\n      hfToken: options.hfToken,\n      cacheDir: options.cacheDir,\n      onProgress: options.onProgress,\n    });\n    const cfg = parseParlerConfig(loaded.rawConfig);\n    const maxSeqLen = Math.min(options.maxSeqLen ?? 4096, cfg.max_position);\n    return new ParlerTTS(ctx, loaded, maxSeqLen);\n  }\n\n  /**\n   * Generate a random natural-language voice description using a host-side template\n   * bank (no model call). Pass the result as `describeVoice` for variety. A loaded\n   * text model can write richer ones — exposed on the Gerbil class as\n   * {@link inventVoiceDescription}.\n   */\n  static randomVoiceDescription(): string {\n    const gender = pick([\"male\", \"female\"]);\n    const tone = pick([\"warm\", \"bright\", \"calm\", \"energetic\", \"gentle\", \"confident\", \"soft\"]);\n    const pace = pick([\"slightly slow\", \"moderate\", \"slightly fast\"]);\n    const quality = pick([\n      \"very clear, studio-quality recording\",\n      \"clean recording with minimal background noise\",\n      \"close-sounding, high-quality audio\",\n    ]);\n    return `A ${tone} ${gender} voice speaking at a ${pace} pace, in a ${quality}.`;\n  }\n\n  /** Synthesize `text` in the described voice. Returns 44.1 kHz mono PCM. */\n  async speak(text: string, opts: ParlerSpeakOptions = {}): Promise<ParlerSpeakResult> {\n    if (this._destroyed) {\n      throw new Error(\"ParlerTTS has been destroyed.\");\n    }\n    const startTime = performance.now();\n    const description = (opts.describeVoice ?? DEFAULT_VOICE_DESCRIPTION).trim();\n\n    // ── 1. Encoder: tokenize description, build rel bias, run once. ──\n    const descIds = this.loaded.tokenizer.encode(description);\n    const sDesc = descIds.length;\n    const biasTable = this.loaded.modelWeights.get(\n      \"text_encoder.encoder.block.0.layer.0.SelfAttention.relative_attention_bias.weight\",\n    );\n    if (!biasTable) {\n      throw new Error(\"ParlerTTS: relative_attention_bias.weight missing from checkpoint.\");\n    }\n    const relBias = buildT5RelativeBias(\n      biasTable.data as Float32Array,\n      this.cfg.rel_num_buckets,\n      this.cfg.rel_max_distance,\n      this.cfg.enc_heads,\n      sDesc,\n    );\n\n    const encGraph = generateParlerEncoderGraph(this.loaded.rawConfig, sDesc);\n    const encExec = new ParlerEncoderExecutor(this.ctx, encGraph, this.cfg.dec_layers, sDesc);\n    let encK: Float32Array[];\n    let encV: Float32Array[];\n    try {\n      encExec.uploadWeights(selectGraphWeights(encGraph, this.loaded.modelWeights));\n      encExec.initBindGroups();\n      const enc = await encExec.encode(new Uint32Array(descIds), relBias);\n      encK = enc.encK;\n      encV = enc.encV;\n    } finally {\n      encExec.destroy();\n    }\n\n    // ── 2/3. Decoder: prompt prefill + AR delay-pattern decode. ──\n    const promptIds = this.loaded.tokenizer.encode(text);\n    const delayed = await this.decodeAudioTokens(promptIds, encK, encV, sDesc, opts);\n\n    // ── 4. Revert the delay pattern → [9, frames] → DAC decode. ──\n    const { codes, frames } = revertDelayPattern(\n      delayed.codes,\n      this.cfg.num_codebooks,\n      delayed.steps,\n    );\n    if (frames <= 0) {\n      throw new Error(`ParlerTTS produced no audio frames (steps=${delayed.steps}).`);\n    }\n    opts.onCodes?.(codes, this.cfg.num_codebooks, frames);\n    const pcm = await this.decodeCodes(codes, frames);\n\n    const totalTime = (performance.now() - startTime) / 1000;\n    const audioSeconds = pcm.length / PARLER_SAMPLE_RATE;\n    console.log(\n      `[parler] desc=${sDesc} frames=${frames} audio=${audioSeconds.toFixed(2)}s wall=${totalTime.toFixed(2)}s`,\n    );\n    return { pcm, sampleRate: PARLER_SAMPLE_RATE, frames, audioSeconds };\n  }\n\n  /**\n   * Run the decoder: prefill the transcript prompt rows, then AR-decode the 9-codebook\n   * delay-pattern stream. Returns the DELAYED code grid [9, steps] (codebook-major)\n   * plus the step count, for revertDelayPattern.\n   *\n   * Each step the host sums the 9 per-codebook embeddings of the current delayed codes\n   * + the sinusoidal position, writes that single [1,H] row to `dec_input_embed`,\n   * forwards one step (KV-cache grows over [prompt ++ audio]), and reads back 9 logit\n   * rows — one per codebook lm_head.\n   */\n  private async decodeAudioTokens(\n    promptIds: number[],\n    encK: Float32Array[],\n    encV: Float32Array[],\n    sEnc: number,\n    opts: ParlerSpeakOptions,\n  ): Promise<{ codes: Uint32Array; steps: number }> {\n    const cfg = this.cfg;\n    const H = cfg.dec_hidden;\n    const NCB = cfg.num_codebooks;\n    const graph = generateParlerDecoderGraph(this.loaded.rawConfig, sEnc);\n    const maxFrames = Math.min(\n      opts.maxFrames ?? DEFAULT_MAX_FRAMES,\n      this.maxSeqLen - promptIds.length,\n    );\n    const maxSteps = maxFrames + (NCB - 1);\n    const exec = new Executor(this.ctx, graph, {\n      maxSeqLen: this.maxSeqLen,\n      kvMode: \"f32\",\n    });\n\n    // Host-side embedding tables.\n    const embedTokens: Float32Array[] = [];\n    for (let cb = 0; cb < NCB; cb++) {\n      const t = this.loaded.modelWeights.get(`decoder.model.decoder.embed_tokens.${cb}.weight`);\n      if (!t) {\n        throw new Error(`ParlerTTS: embed_tokens.${cb} missing.`);\n      }\n      embedTokens.push(t.data as Float32Array);\n    }\n    const embedPositions = this.loaded.modelWeights.get(\n      \"decoder.model.decoder.embed_positions.weights\",\n    );\n    const embedPrompts = this.loaded.modelWeights.get(\"embed_prompts.weight\");\n    if (!(embedPositions && embedPrompts)) {\n      throw new Error(\"ParlerTTS: embed_positions / embed_prompts missing.\");\n    }\n    const posTable = embedPositions.data as Float32Array;\n    const promptTable = embedPrompts.data as Float32Array;\n\n    const grid = new Uint32Array(NCB * maxSteps); // codebook-major delayed codes\n    let steps = 0;\n    try {\n      exec.uploadWeightsMap(selectGraphWeights(graph, this.loaded.modelWeights));\n      exec.initBindGroups();\n      for (let i = 0; i < cfg.dec_layers; i++) {\n        exec.writeInput(`enc_k_layer${i}`, encK[i]);\n        exec.writeInput(`enc_v_layer${i}`, encV[i]);\n      }\n      exec.reset();\n\n      let pos = 0;\n      const row = new Float32Array(H);\n      const logitOutputs = graph.outputs.map((name) => ({ name, elemCount: cfg.audio_vocab }));\n\n      // ── Prompt prefill: feed each transcript-token embedding (+ position) one row at\n      //    a time so the self-attn KV-cache fills with the prompt span (positions\n      //    0..promptLen-1) before audio decode. (One row per step keeps the T=1 graph;\n      //    runGraphStep advances seqPos per call.) ──\n      for (const tokenId of promptIds) {\n        addRow(row, promptTable, tokenId, H, true);\n        addPosition(row, posTable, pos, H);\n        exec.writeInput(\"dec_input_embed\", row);\n        await exec.runGraphStep([logitOutputs[0]], 1);\n        pos++;\n      }\n\n      // ── AR audio decode with delay pattern. ──\n      // Codebook cb is \"active\" only once step >= cb; before that it stays at the bos\n      // fill so its stream is delayed by cb frames. Step 0 is seeded with bos across all\n      // codebooks (decoder_start_token_id).\n      //\n      // EOS propagation mirrors ParlerTTSLogitsProcessor: when the first not-yet-finished\n      // codebook emits EOS, the \"frontier\" advances one codebook; every codebook strictly\n      // beyond the frontier is forbidden from emitting EOS (its EOS logit is masked to\n      // -inf before argmax/sampling). Once the frontier reaches the last codebook and it\n      // emits EOS, all codebooks are finished and we stop. This decodes exactly the\n      // (NCB-1)-step delay tail the reference completes after codebook 0's EOS, so the\n      // de-delayed grid trims cleanly instead of trailing post-EOS junk.\n      let prevCodes = new Uint32Array(NCB).fill(PARLER_BOS_TOKEN_ID);\n      let frontier = 0; // index of the first codebook not yet finished (allowed to EOS).\n      let finished = false;\n      for (let step = 0; step < maxSteps && !finished; step++) {\n        // Sum the 9 per-codebook embeddings of the current step's input codes.\n        row.fill(0);\n        for (let cb = 0; cb < NCB; cb++) {\n          addRow(row, embedTokens[cb], prevCodes[cb], H, false);\n        }\n        addPosition(row, posTable, pos, H);\n        exec.writeInput(\"dec_input_embed\", row);\n        const logitRows = await exec.runGraphStep(logitOutputs, 1);\n        pos++;\n\n        const nextCodes = new Uint32Array(NCB);\n        for (let cb = 0; cb < NCB; cb++) {\n          if (step < cb) {\n            nextCodes[cb] = PARLER_BOS_TOKEN_ID;\n            continue;\n          }\n          // Forbid EOS for codebooks beyond the frontier (delay-pattern constraint).\n          if (cb > frontier) {\n            logitRows[cb][PARLER_EOS_TOKEN_ID] = Number.NEGATIVE_INFINITY;\n          }\n          const code = opts.greedy\n            ? argmax(logitRows[cb])\n            : sampleLogits(logitRows[cb], opts.temperature ?? 1.0, opts.topK ?? 0);\n          nextCodes[cb] = code;\n          grid[cb * maxSteps + step] = code < PARLER_CODEBOOK_SIZE ? code : 0;\n        }\n\n        // Advance the frontier if the current first-unfinished codebook emitted EOS.\n        if (frontier < NCB - 1 && nextCodes[frontier] === PARLER_EOS_TOKEN_ID) {\n          frontier++;\n        } else if (frontier === NCB - 1 && nextCodes[frontier] === PARLER_EOS_TOKEN_ID) {\n          finished = true;\n        }\n\n        steps = step + 1;\n        prevCodes = nextCodes;\n        if (pos >= this.maxSeqLen) {\n          break;\n        }\n      }\n    } finally {\n      exec.destroy();\n    }\n\n    // Re-pack the [NCB, maxSteps] grid to a tight [NCB, steps] codebook-major grid; the\n    // delay pattern is reverted by the caller (revertDelayPattern reads with stride steps).\n    const tight = new Uint32Array(NCB * steps);\n    for (let cb = 0; cb < NCB; cb++) {\n      for (let s = 0; s < steps; s++) {\n        tight[cb * steps + s] = grid[cb * maxSteps + s];\n      }\n    }\n    return { codes: tight, steps };\n  }\n\n  /** Decode the [9, frames] code grid → 44.1 kHz PCM (tanh host-side). */\n  private async decodeCodes(codes: Uint32Array, numFrames: number): Promise<Float32Array> {\n    const graph = generateDacSpeechDecoderGraph({\n      numFrames,\n      numCodebooks: this.cfg.num_codebooks,\n      decoderRates: PARLER_DECODER_RATES,\n      decoderDim: PARLER_DAC_DECODER_DIM,\n      latentDim: PARLER_DAC_LATENT_DIM,\n    });\n    const pcmLen = dacOutputLength(numFrames, PARLER_DECODER_RATES);\n    const exec = new Executor(this.ctx, graph, { maxSeqLen: numFrames, kvMode: \"f32\" });\n    try {\n      exec.uploadWeightsMap(selectGraphWeights(graph, this.loaded.codecWeights));\n      exec.initBindGroups();\n      exec.reset();\n      exec.writeInput(\"audio_codes\", codes);\n      const raw = await exec.runGraphOutput(graph.outputs[0], pcmLen);\n      const pcm = new Float32Array(raw.length);\n      for (let i = 0; i < raw.length; i++) {\n        pcm[i] = Math.tanh(raw[i]);\n      }\n      return pcm;\n    } finally {\n      exec.destroy();\n    }\n  }\n\n  destroy(): void {\n    if (this._destroyed) {\n      return;\n    }\n    this._destroyed = true;\n    this.loaded.modelWeights.clear();\n    this.loaded.codecWeights.clear();\n  }\n}\n\nconst PARLER_CODEBOOK_SIZE = 1024;\n\n/** Random pick from a list. */\nfunction pick<T>(arr: readonly T[]): T {\n  return arr[Math.floor(Math.random() * arr.length)];\n}\n\n/** Add (or set) the embedding row for `tokenId` into `row`. */\nfunction addRow(\n  row: Float32Array,\n  table: Float32Array,\n  tokenId: number,\n  H: number,\n  set: boolean,\n): void {\n  const base = tokenId * H;\n  if (set) {\n    for (let i = 0; i < H; i++) {\n      row[i] = table[base + i];\n    }\n  } else {\n    for (let i = 0; i < H; i++) {\n      row[i] += table[base + i];\n    }\n  }\n}\n\n/** Add the sinusoidal position row (from the learned embed_positions table). */\nfunction addPosition(row: Float32Array, posTable: Float32Array, pos: number, H: number): void {\n  const base = pos * H;\n  for (let i = 0; i < H; i++) {\n    row[i] += posTable[base + i];\n  }\n}\n","/**\n * VisionExecutor — runs the Qwen3.5 ViT graph on WebGPU.\n *\n * Kept separate from the text `Executor` (which is specialized for\n * input_ids / logits / argmax / KV-cache decode). The vision tower is a single\n * non-autoregressive prefill over `N` patches, so this executor is deliberately\n * simple: one buffer per tensor, dispatch in executionOrder, read back the\n * merged image embeddings. It reuses the shared kernel registry and device\n * helpers, so kernel math is identical to the text path.\n *\n * Symbolic dims:\n *   \"N\"  → number of patches (runtime)\n *   \"Nm\" → N / spatial_merge_unit (merged token count)\n */\n\nimport type { GPUContext } from \"./device.js\";\nimport {\n  createBindGroup,\n  createStorageBuffer,\n  createUniformBuffer,\n  destroyBuffers,\n  getOrCreatePipeline,\n} from \"./device.js\";\nimport { CANONICAL_KEYS, DTYPE_BYTES, type ModelGraph, type OpNode } from \"./ir.js\";\nimport { KERNEL_REGISTRY, type KernelSpec, MATMUL_BIAS_F16C_SPEC } from \"./kernels/registry.js\";\n\nconst MAP_MODE_READ = 0x0001;\n\n/** Pack an f32 array to IEEE half (f16) bits (round-to-nearest-even). */\nconst _f16scratch = new ArrayBuffer(4);\nconst _f16f32 = new Float32Array(_f16scratch);\nconst _f16u32 = new Uint32Array(_f16scratch);\nfunction packF16(src: Float32Array): Uint16Array {\n  const out = new Uint16Array(src.length);\n  for (let i = 0; i < src.length; i++) {\n    _f16f32[0] = src[i];\n    const f = _f16u32[0];\n    const sign = (f >>> 16) & 0x8000;\n    const exp = ((f >>> 23) & 0xff) - 112; // 127 - 15\n    const mant = f & 0x7fffff;\n    if (exp <= 0) {\n      out[i] = sign;\n    } else if (exp >= 0x1f) {\n      out[i] = sign | 0x7c00;\n    } else {\n      const round = (mant & 0x1000) !== 0;\n      let half = sign | (exp << 10) | (mant >>> 13);\n      if (round) half += 1;\n      out[i] = half;\n    }\n  }\n  return out;\n}\n\ninterface VisionDispatch {\n  node: OpNode;\n  spec: KernelSpec;\n  pipeline: GPUComputePipeline;\n  bindGroup: GPUBindGroup;\n  uniformBuffer: GPUBuffer;\n}\n\nexport interface VisionInputs {\n  /** Flattened patches [N, patch_dim]. */\n  patches: Float32Array;\n  /** Bilinear-interpolated learned pos embeddings [N, hidden_size]. */\n  posEmbeds: Float32Array;\n  /** Precomputed rotary cos [N, head_dim]. */\n  cos: Float32Array;\n  /** Precomputed rotary sin [N, head_dim]. */\n  sin: Float32Array;\n  /** Number of patches (rows). */\n  numPatches: number;\n}\n\n/**\n * Coarse stage callback for the ViT encode. Fires before each transformer layer\n * (and the pre/post stages) so a host harness can localize a mobile GPU-process\n * crash to a specific layer instead of seeing only \"crashed after load\". Kept\n * synchronous and cheap; throwing is the caller's responsibility.\n */\nexport type VisionStageCallback = (\n  stage: string,\n  info?: { layer?: number; total?: number },\n) => void;\n\n/**\n * WebKit only: the ViT bidirectional Attention op is a single dispatch whose\n * grid is [N query rows, heads], each row looping over all N keys — O(N²) work in\n * one submission. Past a few hundred rows that can exceed Metal's ~few-second GPU\n * watchdog and kill the web-content process (indistinguishable from OOM). We split\n * the attention grid into windows of this many query rows, each its own\n * submit + onSubmittedWorkDone() drain, so no single submission runs long enough\n * to trip the watchdog. Math is unchanged: every query row computes identically\n * regardless of how the grid is partitioned.\n */\nconst WEBKIT_ATTN_CHUNK_ROWS = 64;\n\nexport class VisionExecutor {\n  private ctx: GPUContext;\n  private graph: ModelGraph;\n  private mergeUnit: number;\n\n  private weightBuffers = new Map<string, GPUBuffer>();\n  private activationBuffers = new Map<string, GPUBuffer>();\n  private dispatches: VisionDispatch[] = [];\n  private maxPatches: number;\n  /** Weight (B) names of MatMulBias nodes stored as f16 (empty without shader-f16). */\n  private f16WeightNames = new Set<string>();\n\n  /** Runtime pooled-token count for the Gemma 4 ViT (\"Np\" dim); 0 for Qwen. */\n  private gemma4Np = 0;\n  /** True when this graph is the Gemma 4 vision tower (uses \"Np\" + PoolMatMul). */\n  private readonly isGemma4: boolean;\n\n  constructor(ctx: GPUContext, graph: ModelGraph, maxPatches: number) {\n    this.ctx = ctx;\n    this.graph = graph;\n    this.maxPatches = maxPatches;\n    this.isGemma4 = graph.architecture === \"Gemma4VisionModel\";\n    // merge unit = merger fc1 in-width / block hidden size (3072 / 768 = 4).\n    // Gemma 4 has no merger (it pools via a host matrix), so fall back to 1.\n    const mergedIn = graph.tensors[CANONICAL_KEYS.visMergerFc1W]?.shape[1] as number;\n    const hidden = graph.config.hidden_size;\n    this.mergeUnit = mergedIn && hidden ? mergedIn / hidden : 1;\n    if (ctx.hasF16) {\n      for (const node of graph.nodes) {\n        if (node.opType === \"MatMulBias\" && node.inputs[1]) {\n          this.f16WeightNames.add(node.inputs[1]);\n        }\n      }\n    }\n    this.allocateActivationBuffers();\n  }\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) throw new Error(`VisionExecutor: missing weight \"${name}\"`);\n      let data: ArrayBufferView = w.data;\n      if (this.f16WeightNames.has(name) && w.data instanceof Float32Array) {\n        data = packF16(w.data);\n      }\n      const buffer = createStorageBuffer(this.ctx, `vweight_${name}`, data.byteLength, data);\n      this.weightBuffers.set(name, buffer);\n    }\n  }\n\n  initBindGroups(): void {\n    const dummyShapes = this.resolveShapes(this.maxPatches);\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(`VisionExecutor: no kernel for op \"${node.opType}\"`);\n      if (node.opType === \"MatMulBias\" && this.f16WeightNames.has(node.inputs[1])) {\n        spec = MATMUL_BIAS_F16C_SPEC;\n      }\n\n      const pipeline = getOrCreatePipeline(\n        this.ctx,\n        `vkernel_${nodeId}`,\n        spec.shaderCode,\n        spec.entryPoint,\n      );\n      const paramsData = spec.buildParams(node, dummyShapes, { seqPos: 0 });\n      const uniformBuffer = createUniformBuffer(this.ctx, `vuniform_${nodeId}`, paramsData);\n      const bufferEntries = this.gatherBuffers(spec, node, uniformBuffer);\n      const bindGroup = createBindGroup(this.ctx, pipeline, bufferEntries, `vbg_${nodeId}`);\n      this.dispatches.push({ node, spec, pipeline, bindGroup, uniformBuffer });\n    }\n  }\n\n  /**\n   * Encode patches → merged image embeddings [Nm, out_hidden_size].\n   *\n   * `onStage` (optional) fires coarse phase breadcrumbs during the WebKit path so\n   * a host can localize a GPU-process crash to a specific layer.\n   */\n  async encode(\n    inputs: VisionInputs,\n    onStage?: VisionStageCallback,\n  ): Promise<{ embeds: Float32Array; rows: number; dim: number }> {\n    const N = inputs.numPatches;\n    if (N > this.maxPatches) {\n      throw new Error(`VisionExecutor: ${N} patches exceeds maxPatches=${this.maxPatches}`);\n    }\n    const q = this.ctx.device.queue;\n    q.writeBuffer(this.activationBuffers.get(\"vis_patches\")!, 0, inputs.patches as BufferSource);\n    q.writeBuffer(\n      this.activationBuffers.get(\"vis_pos_embeds\")!,\n      0,\n      inputs.posEmbeds as BufferSource,\n    );\n    q.writeBuffer(this.activationBuffers.get(\"vis_cos\")!, 0, inputs.cos as BufferSource);\n    q.writeBuffer(this.activationBuffers.get(\"vis_sin\")!, 0, inputs.sin as BufferSource);\n\n    const resolvedShapes = this.resolveShapes(N);\n    const ctxRt = { seqPos: 0 };\n\n    // Update uniforms + dispatch sizes. Attention params build with qOffset=0 here\n    // (correct for the Dawn single-pass path); the WebKit path re-writes the\n    // Attention uniform per query-row chunk with a nonzero q_offset below.\n    const sizes: Array<[number, number, number]> = [];\n    for (const d of this.dispatches) {\n      const params = d.spec.buildParams(d.node, resolvedShapes, ctxRt);\n      q.writeBuffer(d.uniformBuffer, 0, params);\n      sizes.push(d.spec.getDispatchSize(d.node, resolvedShapes, ctxRt));\n    }\n\n    const multiEncoder = this.ctx.isWebKitWebGPU;\n    // V1 (GERBIL_VIT_BATCH): batch consecutive non-attention dispatches into one\n    // command buffer per stretch, draining only at attention boundaries — the\n    // stretches are ~14 dispatches/layer, keeping each submission watchdog-safe\n    // while cutting the submit+drain count from ~366 to ~2/layer + attn chunks.\n    // Command order is identical either way (single queue), so output is\n    // bit-exact; the drains only cost latency. Flag-gated pending an on-device\n    // watchdog verdict.\n    const batchVit = (globalThis as { GERBIL_VIT_BATCH?: unknown }).GERBIL_VIT_BATCH != null;\n    let batchEnc: GPUCommandEncoder | null = null;\n    let batchPass: GPUComputePassEncoder | null = null;\n    const flushBatch = async () => {\n      if (!batchEnc || !batchPass) return;\n      batchPass.end();\n      q.submit([batchEnc.finish()]);\n      await q.onSubmittedWorkDone();\n      batchEnc = null;\n      batchPass = null;\n    };\n    if (multiEncoder) {\n      onStage?.(\"vit-encode-start\", { total: this.dispatches.length });\n      for (let i = 0; i < this.dispatches.length; i++) {\n        const d = this.dispatches[i];\n        if (d.node.opType === \"Attention\") {\n          await flushBatch();\n          // Split the O(N²) attention grid into watchdog-safe query-row windows.\n          // Each window: write the uniform with its q_offset, submit, drain. The\n          // drain guarantees the chunk completes before the next overwrites the\n          // shared uniform buffer.\n          const [gridX, gridY, gridZ] = sizes[i];\n          const layer = Number.parseInt(d.node.id.replace(/\\D+/g, \"\"), 10);\n          if (!Number.isNaN(layer)) onStage?.(\"vit-attn\", { layer });\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, resolvedShapes, {\n              seqPos: 0,\n              qOffset: start,\n            });\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        if (batchVit) {\n          if (!batchEnc) {\n            batchEnc = this.ctx.device.createCommandEncoder();\n            batchPass = batchEnc.beginComputePass();\n          }\n          batchPass!.setPipeline(d.pipeline);\n          batchPass!.setBindGroup(0, d.bindGroup);\n          batchPass!.dispatchWorkgroups(...sizes[i]);\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      await flushBatch();\n      onStage?.(\"vit-encode-done\");\n    } else {\n      const enc = this.ctx.device.createCommandEncoder({ label: \"vision_encode\" });\n      const pass = enc.beginComputePass({ label: \"vision_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    // Read back the merged embeddings.\n    const rows = N / this.mergeUnit;\n    const dim = this.graph.config.vision_embed_dim\n      ? (this.graph.tensors.vis_image_embeds.shape[1] as number)\n      : (this.graph.tensors.vis_image_embeds.shape[1] as number);\n    const byteLen = rows * dim * 4;\n    const out = this.activationBuffers.get(\"vis_image_embeds\")!;\n    const readback = this.ctx.device.createBuffer({ size: byteLen, usage: 0x0001 | 0x0008 });\n    const enc = this.ctx.device.createCommandEncoder();\n    enc.copyBufferToBuffer(out, 0, readback, 0, byteLen);\n    this.ctx.device.queue.submit([enc.finish()]);\n    await readback.mapAsync(MAP_MODE_READ, 0, byteLen);\n    const embeds = new Float32Array(readback.getMappedRange(0, byteLen).slice(0));\n    readback.unmap();\n    readback.destroy();\n    return { embeds, rows, dim };\n  }\n\n  /**\n   * Encode patches through the Gemma 4 ViT → projected image tokens [Np, text_hidden].\n   *\n   * Distinct from the Qwen `encode()`: the Gemma graph has 5 inputs (patches,\n   * axial pos-embeds, axial rotary cos/sin, and a host-built [Np,N] pooling matrix)\n   * and its output rows (Np) are the pooled soft-token count, resolved from the\n   * pooling matrix rather than an N/mergeUnit ratio. Reuses the same dispatch\n   * machinery + WebKit per-dispatch-drain discipline.\n   */\n  async encodeGemma4(\n    inputs: {\n      patches: Float32Array; // [N, patch_dim]\n      posEmbeds: Float32Array; // [N, hidden]\n      cos: Float32Array; // [N, head_dim]\n      sin: Float32Array; // [N, head_dim]\n      poolMatrix: Float32Array; // [Np, N]\n      numPatches: number;\n      numPooled: number;\n    },\n    onStage?: VisionStageCallback,\n  ): Promise<{ embeds: Float32Array; rows: number; dim: number }> {\n    const N = inputs.numPatches;\n    if (N > this.maxPatches) {\n      throw new Error(`VisionExecutor(Gemma4): ${N} patches exceeds maxPatches=${this.maxPatches}`);\n    }\n    this.gemma4Np = inputs.numPooled;\n    const q = this.ctx.device.queue;\n    q.writeBuffer(this.activationBuffers.get(\"g4v_patches\")!, 0, inputs.patches as BufferSource);\n    q.writeBuffer(\n      this.activationBuffers.get(\"g4v_pos_embeds\")!,\n      0,\n      inputs.posEmbeds as BufferSource,\n    );\n    q.writeBuffer(this.activationBuffers.get(\"g4v_cos\")!, 0, inputs.cos as BufferSource);\n    q.writeBuffer(this.activationBuffers.get(\"g4v_sin\")!, 0, inputs.sin as BufferSource);\n    q.writeBuffer(this.activationBuffers.get(\"g4v_pool_w\")!, 0, inputs.poolMatrix as BufferSource);\n\n    const resolvedShapes = this.resolveShapes(N);\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, resolvedShapes, ctxRt);\n      q.writeBuffer(d.uniformBuffer, 0, params);\n      sizes.push(d.spec.getDispatchSize(d.node, resolvedShapes, ctxRt));\n    }\n\n    const multiEncoder = this.ctx.isWebKitWebGPU;\n    if (multiEncoder) {\n      onStage?.(\"vit-encode-start\", { total: this.dispatches.length });\n      for (let i = 0; i < this.dispatches.length; i++) {\n        const d = this.dispatches[i];\n        if (d.node.opType === \"Attention\") {\n          const [gridX, gridY, gridZ] = sizes[i];\n          const layer = Number.parseInt(d.node.id.replace(/\\D+/g, \"\"), 10);\n          if (!Number.isNaN(layer)) onStage?.(\"vit-attn\", { layer });\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, resolvedShapes, {\n              seqPos: 0,\n              qOffset: start,\n            });\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      onStage?.(\"vit-encode-done\");\n    } else {\n      const enc = this.ctx.device.createCommandEncoder({ label: \"gemma4_vision_encode\" });\n      const pass = enc.beginComputePass({ label: \"gemma4_vision_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 rows = inputs.numPooled;\n    const dim = this.graph.tensors.g4v_image_embeds.shape[1] as number;\n    const byteLen = rows * dim * 4;\n    const out = this.activationBuffers.get(\"g4v_image_embeds\")!;\n    const readback = this.ctx.device.createBuffer({ size: byteLen, usage: 0x0001 | 0x0008 });\n    const enc = this.ctx.device.createCommandEncoder();\n    enc.copyBufferToBuffer(out, 0, readback, 0, byteLen);\n    this.ctx.device.queue.submit([enc.finish()]);\n    await readback.mapAsync(MAP_MODE_READ, 0, byteLen);\n    const embeds = new Float32Array(readback.getMappedRange(0, byteLen).slice(0));\n    readback.unmap();\n    readback.destroy();\n    return { embeds, rows, dim };\n  }\n\n  /** True if this executor is the Gemma 4 vision tower. */\n  get gemma4(): boolean {\n    return this.isGemma4;\n  }\n\n  /** Read back any named activation (debug). Must be called right after encode(). */\n  async debugReadBuffer(name: string, maxElements?: number): Promise<Float32Array> {\n    const buffer = this.activationBuffers.get(name) ?? this.weightBuffers.get(name);\n    if (!buffer) throw new Error(`VisionExecutor: no buffer \"${name}\"`);\n    const byteLen = maxElements ? Math.min(maxElements * 4, buffer.size) : 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  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  /** Max pooled tokens for buffer sizing: maxPatches with no merge/pool ratio applied. */\n  private maxPooled(): number {\n    // Gemma 4 pools by k² (k = pooling_kernel_size). With no merger key we cannot\n    // read k here, so reserve maxPatches rows (an upper bound — Np ≤ N always).\n    return this.maxPatches;\n  }\n\n  private resolveShapes(N: number): Record<string, number[]> {\n    const Nm = N / this.mergeUnit;\n    const Np = this.gemma4Np || N; // resolved per-encode for Gemma 4; ≤ N\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 === \"N\") return N;\n        if (d === \"Nm\") return Nm;\n        if (d === \"Np\") return Np;\n        return d as number;\n      });\n    }\n    return resolved;\n  }\n\n  private allocateActivationBuffers(): void {\n    // One dedicated buffer per activation tensor sized at maxPatches. Vision\n    // graphs are tiny (≤ a few thousand patches × ≤3072 floats), so no pooling.\n    // Only tensors actually referenced by a node get a buffer: the vision arch\n    // files declare orphan `*_mm` intermediates from a pre-fused design, and\n    // blindly allocating them costs ~331 MB on WebKit (~1.3 GB desktop) — a\n    // large slice of the iPhone vision OOM for zero work.\n    const referenced = new Set<string>();\n    for (const node of this.graph.nodes) {\n      for (const t of node.inputs) referenced.add(t);\n      for (const t of node.outputs) referenced.add(t);\n    }\n    let skippedBytes = 0;\n    for (const [name, desc] of Object.entries(this.graph.tensors)) {\n      if (desc.storage !== \"activation\") continue;\n      if (!referenced.has(name)) {\n        const shape = desc.shape.map((d) => {\n          if (d === \"N\") return this.maxPatches;\n          if (d === \"Nm\") return this.maxPatches / this.mergeUnit;\n          if (d === \"Np\") return this.maxPooled();\n          return d as number;\n        });\n        skippedBytes += shape.reduce((a, b) => a * b, 1) * DTYPE_BYTES[desc.dtype];\n        continue;\n      }\n      const shape = desc.shape.map((d) => {\n        if (d === \"N\") return this.maxPatches;\n        if (d === \"Nm\") return this.maxPatches / this.mergeUnit;\n        if (d === \"Np\") return this.maxPooled();\n        return d as number;\n      });\n      const bytes = shape.reduce((a, b) => a * b, 1) * DTYPE_BYTES[desc.dtype];\n      this.activationBuffers.set(name, createStorageBuffer(this.ctx, `vact_${name}`, bytes));\n    }\n    if (skippedBytes > 0) {\n      console.log(\n        `[vision] skipped ${(skippedBytes / 1e6).toFixed(0)} MB of unreferenced activation tensors`,\n      );\n    }\n  }\n\n  private gatherBuffers(\n    spec: KernelSpec,\n    node: OpNode,\n    uniformBuffer: GPUBuffer,\n  ): Array<{ buffer: GPUBuffer }> {\n    const entries: Array<{ buffer: GPUBuffer }> = [];\n    const tensorNames = [...node.inputs, ...node.outputs];\n    let tensorIdx = 0;\n    for (const binding of spec.bindings) {\n      if (binding.type === \"uniform\") {\n        entries.push({ buffer: uniformBuffer });\n      } else {\n        const tensorName = tensorNames[tensorIdx++];\n        const buffer = this.weightBuffers.get(tensorName) ?? this.activationBuffers.get(tensorName);\n        if (!buffer)\n          throw new Error(`VisionExecutor: no buffer for \"${tensorName}\" in op ${node.id}`);\n        entries.push({ buffer });\n      }\n    }\n    return entries;\n  }\n}\n","/**\n * Host-side vision preprocessing for the Qwen3.5 ViT.\n *\n * The learned position embeddings (bilinear-interpolated over the patch grid)\n * and the 2D rotary cos/sin tables are functions of the image grid (t, h, w)\n * ONLY — not of the model weights or pixel values. They are cheap to compute on\n * the CPU and fed to the GPU graph as input activations, keeping the graph to\n * weight-dependent math while staying byte-identical to HF transformers.\n *\n * Ports (verified against transformers 5.12 vision_utils.py + modeling_qwen3_5):\n *   - get_vision_bilinear_indices_and_weights  → buildPosEmbeds()\n *   - get_vision_position_ids                  → buildPositionIds()\n *   - Qwen3_5VisionRotaryEmbedding             → buildRotaryCosSin()\n *\n * Patch ordering: both the pos-embed gather and the position ids reorder patches\n * into spatial_merge_size×spatial_merge_size groups, matching the HF image\n * processor's output ordering, so the merger's [N,h]→[N/u,h*u] reshape lines up.\n */\n\nexport interface VisionGridConfig {\n  hiddenSize: number; // 768\n  numHeads: number; // 12\n  numPositionEmbeddings: number; // 2304 → num_grid_per_side = 48\n  spatialMergeSize: number; // 2\n  ropeTheta?: number; // 10000\n}\n\nexport interface VisionPositionTensors {\n  /** [N, hidden_size] bilinear-interpolated learned position embeddings. */\n  posEmbeds: Float32Array;\n  /** [N, head_dim] rotary cosines. */\n  cos: Float32Array;\n  /** [N, head_dim] rotary sines. */\n  sin: Float32Array;\n  numPatches: number;\n}\n\n/** linspace(0, n-1, count) matching torch.linspace. */\nfunction linspace(stop: number, count: number): Float64Array {\n  const out = new Float64Array(count);\n  if (count === 1) {\n    out[0] = 0;\n    return out;\n  }\n  const step = stop / (count - 1);\n  for (let i = 0; i < count; i++) out[i] = i * step;\n  return out;\n}\n\n/**\n * Compute the spatial-merge reorder index for a (h, w) grid:\n * reorder[k] gives the source patch index (row-major h*w) for output slot k,\n * grouping 2×2 (merge×merge) spatial blocks contiguously. Matches the `reorder`\n * in get_vision_bilinear_indices_and_weights and the position-id reshape.\n */\nfunction spatialMergeReorder(h: number, w: number, merge: number): Int32Array {\n  // HF: h_idx[h//m, m], w_idx[w//m, m];\n  // (h_idx[:,:,None,None]*w + w_idx[None,None,:,:]).transpose(1,2).flatten()\n  // Result dims (after transpose): [h//m, w//m, m(h), m(w)].\n  const hb = h / merge;\n  const wb = w / merge;\n  const out = new Int32Array(h * w);\n  let k = 0;\n  for (let bh = 0; bh < hb; bh++) {\n    for (let bw = 0; bw < wb; bw++) {\n      for (let ih = 0; ih < merge; ih++) {\n        for (let iw = 0; iw < merge; iw++) {\n          const hh = bh * merge + ih;\n          const ww = bw * merge + iw;\n          out[k++] = hh * w + ww;\n        }\n      }\n    }\n  }\n  return out;\n}\n\n/**\n * Build bilinear-interpolated learned position embeddings [N, hidden].\n * posEmbedTable is the raw pos_embed.weight [num_position_embeddings, hidden].\n */\nexport function buildPosEmbeds(\n  gridTHW: [number, number, number],\n  posEmbedTable: Float32Array,\n  cfg: VisionGridConfig,\n): Float32Array {\n  const [t, h, w] = gridTHW;\n  const hidden = cfg.hiddenSize;\n  const side = Math.round(Math.sqrt(cfg.numPositionEmbeddings)); // 48\n  const merge = cfg.spatialMergeSize;\n\n  const hGrid = linspace(side - 1, h);\n  const wGrid = linspace(side - 1, w);\n  const hFloor = new Int32Array(h);\n  const hCeil = new Int32Array(h);\n  const hFrac = new Float64Array(h);\n  for (let i = 0; i < h; i++) {\n    hFloor[i] = Math.trunc(hGrid[i]);\n    hCeil[i] = Math.min(hFloor[i] + 1, side - 1);\n    hFrac[i] = hGrid[i] - hFloor[i];\n  }\n  const wFloor = new Int32Array(w);\n  const wCeil = new Int32Array(w);\n  const wFrac = new Float64Array(w);\n  for (let j = 0; j < w; j++) {\n    wFloor[j] = Math.trunc(wGrid[j]);\n    wCeil[j] = Math.min(wFloor[j] + 1, side - 1);\n    wFrac[j] = wGrid[j] - wFloor[j];\n  }\n\n  // Per (row-major h*w) patch: 4 corner indices + weights.\n  const hw = h * w;\n  const cornerIdx = [\n    new Int32Array(hw),\n    new Int32Array(hw),\n    new Int32Array(hw),\n    new Int32Array(hw),\n  ];\n  const cornerW = [\n    new Float64Array(hw),\n    new Float64Array(hw),\n    new Float64Array(hw),\n    new Float64Array(hw),\n  ];\n  let p = 0;\n  for (let i = 0; i < h; i++) {\n    const hf = hFloor[i] * side;\n    const hc = hCeil[i] * side;\n    const hfr = hFrac[i];\n    for (let j = 0; j < w; j++) {\n      cornerIdx[0][p] = hf + wFloor[j];\n      cornerIdx[1][p] = hf + wCeil[j];\n      cornerIdx[2][p] = hc + wFloor[j];\n      cornerIdx[3][p] = hc + wCeil[j];\n      cornerW[0][p] = (1 - hfr) * (1 - wFrac[j]);\n      cornerW[1][p] = (1 - hfr) * wFrac[j];\n      cornerW[2][p] = hfr * (1 - wFrac[j]);\n      cornerW[3][p] = hfr * wFrac[j];\n      p++;\n    }\n  }\n\n  const reorder = spatialMergeReorder(h, w, merge);\n  const N = t * hw;\n  const out = new Float32Array(N * hidden);\n  // For each temporal repeat and reordered patch, gather the 4 weighted corners.\n  for (let tt = 0; tt < t; tt++) {\n    for (let k = 0; k < hw; k++) {\n      const src = reorder[k];\n      const dstRow = tt * hw + k;\n      const dstBase = dstRow * hidden;\n      for (let corner = 0; corner < 4; corner++) {\n        const tableRow = cornerIdx[corner][src] * hidden;\n        const wgt = cornerW[corner][src];\n        if (wgt === 0) continue;\n        for (let d = 0; d < hidden; d++) {\n          out[dstBase + d] += posEmbedTable[tableRow + d] * wgt;\n        }\n      }\n    }\n  }\n  return out;\n}\n\n/**\n * Build the reordered (row, col) position ids [N, 2] for rotary, matching\n * get_vision_position_ids.\n */\nexport function buildPositionIds(gridTHW: [number, number, number], merge: number): Int32Array {\n  const [t, h, w] = gridTHW;\n  const hb = h / merge;\n  const wb = w / merge;\n  const hw = h * w;\n  const perFrame = new Int32Array(hw * 2);\n  let k = 0;\n  // hpos/wpos reshaped [h//m, m, w//m, m].transpose(1,2).flatten() → block-major.\n  for (let bh = 0; bh < hb; bh++) {\n    for (let bw = 0; bw < wb; bw++) {\n      for (let ih = 0; ih < merge; ih++) {\n        for (let iw = 0; iw < merge; iw++) {\n          perFrame[k * 2] = bh * merge + ih; // row (h) position\n          perFrame[k * 2 + 1] = bw * merge + iw; // col (w) position\n          k++;\n        }\n      }\n    }\n  }\n  const out = new Int32Array(t * hw * 2);\n  for (let tt = 0; tt < t; tt++) out.set(perFrame, tt * hw * 2);\n  return out;\n}\n\n/**\n * Build rotary cos/sin tables [N, head_dim] from position ids, matching\n * Qwen3_5VisionRotaryEmbedding + the cat((rotary, rotary)) in VisionModel.forward.\n *\n * rotary_pos_emb(position_ids) = (position_ids[..,None] * inv_freq).flatten(1)\n *   where inv_freq has length (head_dim/2)/2 = head_dim/4, computed over dim=head_dim/2.\n * For each token the two position components (h, w) each produce head_dim/4 freqs,\n * concatenated → head_dim/2, then duplicated → head_dim for cos/sin.\n */\nexport function buildRotaryCosSin(\n  positionIds: Int32Array,\n  headDim: number,\n  theta = 10000.0,\n): { cos: Float32Array; sin: Float32Array; numPatches: number } {\n  // rotary embedding dim = head_dim / 2; inv_freq over arange(0, dim, 2)/dim.\n  const rotaryDim = headDim / 2; // 32\n  const half = rotaryDim / 2; // 16 freqs per position component\n  const invFreq = new Float64Array(half);\n  for (let i = 0; i < half; i++) {\n    invFreq[i] = 1.0 / theta ** ((2 * i) / rotaryDim);\n  }\n\n  const N = positionIds.length / 2;\n  // rotary_pos_emb per token = [h*invFreq (half), w*invFreq (half)] → length rotaryDim.\n  // emb = cat(rotary, rotary) → length head_dim. cos/sin = cos/sin(emb).\n  const cos = new Float32Array(N * headDim);\n  const sin = new Float32Array(N * headDim);\n  for (let n = 0; n < N; n++) {\n    const hp = positionIds[n * 2];\n    const wp = positionIds[n * 2 + 1];\n    const base = n * headDim;\n    for (let i = 0; i < half; i++) {\n      const fh = hp * invFreq[i];\n      const fw = wp * invFreq[i];\n      // emb layout: [h-freqs(16), w-freqs(16), h-freqs(16), w-freqs(16)] = head_dim(64)\n      const ch = Math.cos(fh);\n      const sh = Math.sin(fh);\n      const cw = Math.cos(fw);\n      const sw = Math.sin(fw);\n      cos[base + i] = ch;\n      sin[base + i] = sh;\n      cos[base + half + i] = cw;\n      sin[base + half + i] = sw;\n      cos[base + rotaryDim + i] = ch;\n      sin[base + rotaryDim + i] = sh;\n      cos[base + rotaryDim + half + i] = cw;\n      sin[base + rotaryDim + half + i] = sw;\n    }\n  }\n  return { cos, sin, numPatches: N };\n}\n\n/**\n * Build all host position tensors for a single image grid in one call.\n */\nexport function buildVisionPositionTensors(\n  gridTHW: [number, number, number],\n  posEmbedTable: Float32Array,\n  cfg: VisionGridConfig,\n): VisionPositionTensors {\n  const headDim = Math.floor(cfg.hiddenSize / cfg.numHeads);\n  const posEmbeds = buildPosEmbeds(gridTHW, posEmbedTable, cfg);\n  const positionIds = buildPositionIds(gridTHW, cfg.spatialMergeSize);\n  const { cos, sin, numPatches } = buildRotaryCosSin(\n    positionIds,\n    headDim,\n    cfg.ropeTheta ?? 10000.0,\n  );\n  return { posEmbeds, cos, sin, numPatches };\n}\n\n// ───────────────────────────────────────────────────────────────────────────\n// GEMMA 4 vision tower (`gemma4_vision`) host preprocessing.\n//\n// Distinct from the Qwen ViT above: Gemma 4 is a SigLIP-style encoder with\n//   - pre-flattened [N, 3·16·16] patches (Linear patch-embed, no Conv2d),\n//   - AXIAL learned position embeddings: per patch (x,y), pos = table[0][x]+table[1][y],\n//   - 2D axial RoPE (θ=100): spatial_dim = head_dim/2; inv_freq over arange(0,spatial_dim,2)/spatial_dim;\n//     per spatial dim emb=cat(f,f); final cos/sin = cat([emb_x, emb_y]) of length head_dim,\n//     applied with the standard global-half rotate_half (the existing ApplyRotaryEmb kernel),\n//   - k×k average pooling (k=pooling_kernel_size) → soft tokens, as a [Np, N] matrix.\n// Verified against transformers modeling_gemma4 + image_processing_gemma4 and the\n// live config/safetensors of google/gemma-4-E2B-it (+ MLX 4bit convert).\n// ───────────────────────────────────────────────────────────────────────────\n\nexport interface Gemma4VisionGridConfig {\n  hiddenSize: number; // 768\n  numHeads: number; // 12\n  headDim: number; // 64\n  ropeTheta: number; // 100\n  poolingKernelSize: number; // 3\n}\n\nexport interface Gemma4VisionPositionTensors {\n  /** [N, hidden] axial position embeddings (table[0][x] + table[1][y]). */\n  posEmbeds: Float32Array;\n  /** [N, headDim] axial rotary cosines. */\n  cos: Float32Array;\n  /** [N, headDim] axial rotary sines. */\n  sin: Float32Array;\n  /** [Np, N] average-pooling matrix (1/k² in-cell, 0 elsewhere). */\n  poolMatrix: Float32Array;\n  /** number of patches N (= gridH*gridW). */\n  numPatches: number;\n  /** number of pooled (soft) tokens Np (= ceil(gridH/k)*ceil(gridW/k)). */\n  numPooled: number;\n}\n\n/**\n * Per-patch (x, y) grid coordinates in row-major order, matching the Gemma4\n * image processor's meshgrid(arange(width), arange(height)) reshape: patch p at\n * row r (0..gridH-1), col c (0..gridW-1) → x=c, y=r, index = r*gridW + c.\n */\nfunction gemma4PatchXY(gridH: number, gridW: number): { x: Int32Array; y: Int32Array } {\n  const n = gridH * gridW;\n  const x = new Int32Array(n);\n  const y = new Int32Array(n);\n  let p = 0;\n  for (let r = 0; r < gridH; r++) {\n    for (let c = 0; c < gridW; c++) {\n      x[p] = c;\n      y[p] = r;\n      p++;\n    }\n  }\n  return { x, y };\n}\n\n/**\n * Build axial learned position embeddings [N, hidden] from the [2, posSize, hidden]\n * table: pos[p] = table[0][x_p] + table[1][y_p]. Direct lookup, no interpolation\n * (HF F.embedding on clamped positions).\n */\nexport function buildGemma4PosEmbeds(\n  gridH: number,\n  gridW: number,\n  posEmbedTable: Float32Array, // [2, posSize, hidden] flattened\n  hidden: number,\n  posSize: number,\n): Float32Array {\n  const { x, y } = gemma4PatchXY(gridH, gridW);\n  const n = gridH * gridW;\n  const out = new Float32Array(n * hidden);\n  const yPlane = posSize * hidden; // table[1] base offset\n  for (let p = 0; p < n; p++) {\n    const xi = Math.max(0, x[p]);\n    const yi = Math.max(0, y[p]);\n    const xBase = xi * hidden;\n    const yBase = yPlane + yi * hidden;\n    const dst = p * hidden;\n    for (let d = 0; d < hidden; d++) {\n      out[dst + d] = posEmbedTable[xBase + d] + posEmbedTable[yBase + d];\n    }\n  }\n  return out;\n}\n\n/**\n * Build the 2D axial rotary cos/sin tables [N, headDim].\n *   spatial_dim = headDim / 2;  inv_freq[j] = 1/theta^((2j)/spatial_dim), j in [0, spatial_dim/2)\n *   per spatial dim: f = pos * inv_freq (spatial_dim/2 values); emb = cat(f, f) (spatial_dim values)\n *   cos/sin = cat([emb_x, emb_y]) → headDim values, layout [fx,fx,fy,fy].\n * Applied with the global-half rotate_half kernel (ApplyRotaryEmb), which computes\n * out = x*cos + rotate_half(x)*sin element-wise — exact for this layout.\n */\nexport function buildGemma4RotaryCosSin(\n  gridH: number,\n  gridW: number,\n  headDim: number,\n  theta: number,\n): { cos: Float32Array; sin: Float32Array } {\n  const { x, y } = gemma4PatchXY(gridH, gridW);\n  const n = gridH * gridW;\n  const spatialDim = headDim / 2; // 32\n  const half = spatialDim / 2; // 16 freqs per spatial dim\n  const invFreq = new Float64Array(half);\n  for (let j = 0; j < half; j++) {\n    invFreq[j] = 1.0 / theta ** ((2 * j) / spatialDim);\n  }\n  const cos = new Float32Array(n * headDim);\n  const sin = new Float32Array(n * headDim);\n  for (let p = 0; p < n; p++) {\n    const xp = Math.max(0, x[p]);\n    const yp = Math.max(0, y[p]);\n    const base = p * headDim;\n    for (let j = 0; j < half; j++) {\n      const fx = xp * invFreq[j];\n      const fy = yp * invFreq[j];\n      const cx = Math.cos(fx);\n      const sx = Math.sin(fx);\n      const cy = Math.cos(fy);\n      const sy = Math.sin(fy);\n      // layout: [fx(0..15), fx(16..31), fy(32..47), fy(48..63)]\n      cos[base + j] = cx;\n      sin[base + j] = sx;\n      cos[base + half + j] = cx;\n      sin[base + half + j] = sx;\n      cos[base + spatialDim + j] = cy;\n      sin[base + spatialDim + j] = sy;\n      cos[base + spatialDim + half + j] = cy;\n      sin[base + spatialDim + half + j] = sy;\n    }\n  }\n  return { cos, sin };\n}\n\n/**\n * Build the [Np, N] average-pooling matrix for k×k spatial pooling over the real\n * (unpadded) grid, matching modeling_gemma4's kernel_idxs/one_hot pooling:\n *   cell(p) = floor(x_p/k) + ceil(gridW/k) * floor(y_p/k)\n *   poolMatrix[cell, p] = 1/k²  (so pooled = poolMatrix @ hidden = mean over the k×k block)\n * Np = ceil(gridH/k) * ceil(gridW/k). Each pooled cell averages exactly the patches\n * that fall in it (edge cells with fewer than k² patches still divide by k², matching\n * HF's fixed 1/k² normalization).\n */\nexport function buildGemma4PoolMatrix(\n  gridH: number,\n  gridW: number,\n  k: number,\n): { poolMatrix: Float32Array; numPooled: number } {\n  const n = gridH * gridW;\n  const cellsW = Math.ceil(gridW / k);\n  const cellsH = Math.ceil(gridH / k);\n  const np = cellsW * cellsH;\n  const inv = 1.0 / (k * k);\n  const poolMatrix = new Float32Array(np * n);\n  const { x, y } = gemma4PatchXY(gridH, gridW);\n  for (let p = 0; p < n; p++) {\n    const cx = Math.floor(x[p] / k);\n    const cy = Math.floor(y[p] / k);\n    const cell = cx + cellsW * cy;\n    poolMatrix[cell * n + p] = inv;\n  }\n  return { poolMatrix, numPooled: np };\n}\n\n/**\n * Build all Gemma 4 vision host tensors for one image grid in one call.\n * `posEmbedTable` is the raw [2, posSize, hidden] flattened table.\n */\nexport function buildGemma4VisionPositionTensors(\n  gridH: number,\n  gridW: number,\n  posEmbedTable: Float32Array,\n  posSize: number,\n  cfg: Gemma4VisionGridConfig,\n): Gemma4VisionPositionTensors {\n  const posEmbeds = buildGemma4PosEmbeds(gridH, gridW, posEmbedTable, cfg.hiddenSize, posSize);\n  const { cos, sin } = buildGemma4RotaryCosSin(gridH, gridW, cfg.headDim, cfg.ropeTheta);\n  const { poolMatrix, numPooled } = buildGemma4PoolMatrix(gridH, gridW, cfg.poolingKernelSize);\n  return { posEmbeds, cos, sin, poolMatrix, numPatches: gridH * gridW, numPooled };\n}\n\n/** Gemma 4 image processor config (from processor_config.json). */\nexport const GEMMA4_IMAGE_PROCESSOR: ImageProcessorConfig = {\n  patchSize: 16,\n  temporalPatchSize: 1, // single frame, not duplicated (Gemma is not temporal)\n  mergeSize: 1, // pooling (k=3) replaces Qwen's spatial-merge reshape\n  imageMean: [0.0, 0.0, 0.0],\n  imageStd: [1.0, 1.0, 1.0],\n  rescaleFactor: 1 / 255,\n  minPixels: 16 * 16, // a single 48×48-divisible tile minimum\n  maxPixels: 2520 * 16 * 16, // max_patches(280·9) · patch²\n};\n\nexport interface Gemma4PreprocessedImage {\n  /** Flattened patches [N, 3·patch²] row-major (row-major patch grid). */\n  patches: Float32Array;\n  /** Patch grid (gridH, gridW). */\n  gridHW: [number, number];\n}\n\n/**\n * Preprocess a decoded RGB image for the Gemma 4 ViT: aspect-preserving resize so\n * the patch grid is ≤ max_soft_tokens·k² patches and H,W divisible by k·patch,\n * rescale ×1/255 (no normalize), patchify row-major into [N, 3·16·16].\n *\n * @param pixels row-major HWC RGB (0..255), length width*height*3.\n */\nexport function preprocessImageGemma4(\n  pixels: Float32Array | Uint8ClampedArray | Uint8Array,\n  width: number,\n  height: number,\n  maxSoftTokens = 280,\n  poolingKernelSize = 3,\n  patchSize = 16,\n): Gemma4PreprocessedImage {\n  const factor = poolingKernelSize * patchSize; // 48\n  const maxPatches = maxSoftTokens * poolingKernelSize * poolingKernelSize; // 2520\n  // Largest aspect-preserving (H',W') divisible by `factor` with (H'/16)*(W'/16) ≤ maxPatches.\n  const floorByFactor = (n: number) => Math.max(factor, Math.floor(n / factor) * factor);\n  let outH = floorByFactor(height);\n  let outW = floorByFactor(width);\n  // Shrink (preserving aspect) until the patch budget is met.\n  while ((outH / patchSize) * (outW / patchSize) > maxPatches) {\n    const beta = Math.sqrt(((outH / patchSize) * (outW / patchSize)) / maxPatches);\n    outH = floorByFactor(outH / beta);\n    outW = floorByFactor(outW / beta);\n    if (outH <= factor && outW <= factor) break;\n  }\n  const resized = bilinearResize(pixels, height, width, outH, outW);\n\n  const gridH = outH / patchSize;\n  const gridW = outW / patchSize;\n  const ch = 3;\n  const ps = patchSize;\n  const patchDim = ch * ps * ps; // 768\n  const numP = gridH * gridW;\n  const patches = new Float32Array(numP * patchDim);\n  const rescale = 1 / 255;\n\n  // Row-major patch grid; per patch flatten layout [C, ps_h, ps_w] (channel-major)\n  // to match the processor's (num_channels, ph, pw) → flatten ordering.\n  let pIdx = 0;\n  for (let pr = 0; pr < gridH; pr++) {\n    for (let pc = 0; pc < gridW; pc++) {\n      const base = pIdx * patchDim;\n      let kk = 0;\n      for (let c = 0; c < ch; c++) {\n        for (let py = 0; py < ps; py++) {\n          const row = (pr * ps + py) * outW;\n          for (let px = 0; px < ps; px++) {\n            patches[base + kk] = resized[(row + pc * ps + px) * 3 + c] * rescale;\n            kk++;\n          }\n        }\n      }\n      pIdx++;\n    }\n  }\n  return { patches, gridHW: [gridH, gridW] };\n}\n\n// ───────────────────────────────────────────────────────────────────────────\n// LM-SIDE multimodal preprocessing: image → patches, M-RoPE position ids & cos/sin.\n// These feed the text model (not the ViT). Ported from transformers 5.12:\n//   - Qwen2VLImageProcessor.smart_resize / _preprocess  → preprocessImage()\n//   - Qwen3_5Model.get_rope_index / get_vision_position_ids → buildMRoPEPositionIds()\n//   - Qwen3_5TextRotaryEmbedding.forward (interleaved M-RoPE) → buildMRoPECosSin()\n// ───────────────────────────────────────────────────────────────────────────\n\nexport interface ImageProcessorConfig {\n  patchSize: number; // 16\n  temporalPatchSize: number; // 2\n  mergeSize: number; // 2\n  imageMean: [number, number, number]; // (0.5, 0.5, 0.5)\n  imageStd: [number, number, number]; // (0.5, 0.5, 0.5)\n  /** rescale factor applied to raw 0..255 pixels before normalization (1/255). */\n  rescaleFactor: number; // 1/255\n  /** min total pixels after resize (shortest_edge). */\n  minPixels: number; // 65536  (4*16*16*... HF default: 256*28*28 family → here shortest_edge)\n  /** max total pixels after resize (longest_edge). */\n  maxPixels: number; // 16777216\n}\n\nexport const QWEN3_5_IMAGE_PROCESSOR: ImageProcessorConfig = {\n  patchSize: 16,\n  temporalPatchSize: 2,\n  mergeSize: 2,\n  imageMean: [0.5, 0.5, 0.5],\n  imageStd: [0.5, 0.5, 0.5],\n  rescaleFactor: 1 / 255,\n  minPixels: 65_536,\n  maxPixels: 16_777_216,\n};\n\nexport interface PreprocessedImage {\n  /** Flattened patches [N, 1536] in the spatial-merge order encodeImage expects. */\n  patches: Float32Array;\n  /** (t, h, w) patch grid. t=1 for a single image, h/w in patch units. */\n  gridTHW: [number, number, number];\n}\n\n/**\n * Qwen2-VL smart-resize: round H and W to multiples of factor=patch*merge,\n * keeping aspect ratio and clamping the total pixel budget to [minPixels, maxPixels].\n * Matches transformers.models.qwen2_vl.image_processing.smart_resize.\n */\nexport function smartResize(\n  height: number,\n  width: number,\n  factor: number,\n  minPixels: number,\n  maxPixels: number,\n): [number, number] {\n  if (height < factor || width < factor) {\n    // HF raises; clamp up to one factor so tiny images still produce a grid.\n    const scale = factor / Math.min(height, width);\n    height = Math.max(factor, Math.round(height * scale));\n    width = Math.max(factor, Math.round(width * scale));\n  }\n  const roundByFactor = (n: number) => Math.round(n / factor) * factor;\n  const floorByFactor = (n: number) => Math.floor(n / factor) * factor;\n  const ceilByFactor = (n: number) => Math.ceil(n / factor) * factor;\n\n  let hBar = Math.max(factor, roundByFactor(height));\n  let wBar = Math.max(factor, roundByFactor(width));\n  if (hBar * wBar > maxPixels) {\n    const beta = Math.sqrt((height * width) / maxPixels);\n    hBar = Math.max(factor, floorByFactor(height / beta));\n    wBar = Math.max(factor, floorByFactor(width / beta));\n  } else if (hBar * wBar < minPixels) {\n    const beta = Math.sqrt(minPixels / (height * width));\n    hBar = ceilByFactor(height * beta);\n    wBar = ceilByFactor(width * beta);\n  }\n  return [hBar, wBar];\n}\n\n/**\n * Bilinear-resample an RGB image (HWC, 0..255 or 0..1) to (outH, outW).\n * Half-pixel-centered sampling to match PIL/torchvision BILINEAR closely.\n * `pixels` is row-major [H, W, 3].\n */\nfunction bilinearResize(\n  pixels: Float32Array | Uint8ClampedArray | Uint8Array,\n  inH: number,\n  inW: number,\n  outH: number,\n  outW: number,\n): Float32Array {\n  const out = new Float32Array(outH * outW * 3);\n  const scaleH = inH / outH;\n  const scaleW = inW / outW;\n  for (let oy = 0; oy < outH; oy++) {\n    let sy = (oy + 0.5) * scaleH - 0.5;\n    if (sy < 0) sy = 0;\n    if (sy > inH - 1) sy = inH - 1;\n    const y0 = Math.floor(sy);\n    const y1 = Math.min(y0 + 1, inH - 1);\n    const fy = sy - y0;\n    for (let ox = 0; ox < outW; ox++) {\n      let sx = (ox + 0.5) * scaleW - 0.5;\n      if (sx < 0) sx = 0;\n      if (sx > inW - 1) sx = inW - 1;\n      const x0 = Math.floor(sx);\n      const x1 = Math.min(x0 + 1, inW - 1);\n      const fx = sx - x0;\n      const i00 = (y0 * inW + x0) * 3;\n      const i01 = (y0 * inW + x1) * 3;\n      const i10 = (y1 * inW + x0) * 3;\n      const i11 = (y1 * inW + x1) * 3;\n      const od = (oy * outW + ox) * 3;\n      for (let c = 0; c < 3; c++) {\n        const top = pixels[i00 + c] * (1 - fx) + pixels[i01 + c] * fx;\n        const bot = pixels[i10 + c] * (1 - fx) + pixels[i11 + c] * fx;\n        out[od + c] = top * (1 - fy) + bot * fy;\n      }\n    }\n  }\n  return out;\n}\n\n/**\n * Preprocess a decoded RGB image into the [N, 1536] patch tensor + grid_thw that\n * `encodeImage()` expects, matching the HF Qwen2-VL image processor:\n *   smart_resize → rescale (×1/255) → normalize → temporal-pair (×temporal_patch_size)\n *   → patchify into spatial_merge×spatial_merge blocks → flatten to [N, C·T·P·P].\n *\n * @param pixels row-major HWC RGB, length width*height*3. Values 0..255 (default)\n *   or already 0..1 if `rescaled` is true.\n * @param width  source pixel width\n * @param height source pixel height\n */\nexport function preprocessImage(\n  pixels: Float32Array | Uint8ClampedArray | Uint8Array,\n  width: number,\n  height: number,\n  cfg: ImageProcessorConfig = QWEN3_5_IMAGE_PROCESSOR,\n  rescaled = false,\n): PreprocessedImage {\n  const factor = cfg.patchSize * cfg.mergeSize; // 32\n  const [outH, outW] = smartResize(height, width, factor, cfg.minPixels, cfg.maxPixels);\n  const resized = bilinearResize(pixels, height, width, outH, outW);\n\n  // rescale + normalize per channel → CHW float\n  const [mr, mg, mb] = cfg.imageMean;\n  const [sr, sg, sb] = cfg.imageStd;\n  const mean = [mr, mg, mb];\n  const std = [sr, sg, sb];\n  const rescale = rescaled ? 1 : cfg.rescaleFactor;\n  const chw = new Float32Array(3 * outH * outW);\n  const plane = outH * outW;\n  for (let y = 0; y < outH; y++) {\n    for (let x = 0; x < outW; x++) {\n      const src = (y * outW + x) * 3;\n      const dst = y * outW + x;\n      for (let c = 0; c < 3; c++) {\n        chw[c * plane + dst] = (resized[src + c] * rescale - mean[c]) / std[c];\n      }\n    }\n  }\n\n  // Patchify exactly like the HF processor's reshape/transpose chain.\n  // grid: t=1 (single frame, duplicated to temporal_patch_size), h/w in patches.\n  const t = 1;\n  const gridH = outH / cfg.patchSize;\n  const gridW = outW / cfg.patchSize;\n  const ps = cfg.patchSize;\n  const tps = cfg.temporalPatchSize;\n  const ch = 3;\n  const merge = cfg.mergeSize;\n  const gridHm = gridH / merge;\n  const gridWm = gridW / merge;\n  const patchDim = ch * tps * ps * ps; // 1536\n  const numPatches = t * gridH * gridW;\n  const patches = new Float32Array(numPatches * patchDim);\n\n  // HF flatten layout (per patch row, contiguous): [C, T, ps_h, ps_w].\n  // Output patch ordering: blocks of merge×merge, matching get_vision_*: see\n  // qwen2_vl reshape (grid_t, gridHm, merge, gridWm, merge, C, tps, ps, ps).\n  let pIdx = 0;\n  for (let gt = 0; gt < t; gt++) {\n    for (let bh = 0; bh < gridHm; bh++) {\n      for (let bw = 0; bw < gridWm; bw++) {\n        for (let mh = 0; mh < merge; mh++) {\n          for (let mw = 0; mw < merge; mw++) {\n            const patchRow = (bh * merge + mh) * ps;\n            const patchCol = (bw * merge + mw) * ps;\n            const base = pIdx * patchDim;\n            let k = 0;\n            for (let c = 0; c < ch; c++) {\n              for (let tt = 0; tt < tps; tt++) {\n                // single frame duplicated across the temporal dim\n                for (let py = 0; py < ps; py++) {\n                  const row = (patchRow + py) * outW;\n                  for (let px = 0; px < ps; px++) {\n                    patches[base + k] = chw[c * plane + row + (patchCol + px)];\n                    k++;\n                  }\n                }\n                void tt;\n              }\n            }\n            pIdx++;\n          }\n        }\n      }\n    }\n  }\n\n  return { patches, gridTHW: [t, gridH, gridW] };\n}\n\n/**\n * Build the 3D M-RoPE position ids [3, seq] for a text sequence that contains\n * `image_token_id` runs, matching Qwen3_5Model.get_rope_index for a single image.\n *\n * Text tokens advance all three (t,h,w) components together (standard 1D RoPE).\n * Each image run uses the vision grid: temporal=start (t=1), height=start+h_block,\n * width=start+w_block, in row-major (h outer, w inner) order; after the image the\n * cursor jumps by max(h_merged, w_merged). Returns Int32Array of length 3*seq\n * laid out as [T-row(seq), H-row(seq), W-row(seq)].\n */\n/**\n * Fill the 3D position rows for one image run starting at sequence index `at`\n * and logical `start`. Returns the next sequence index and the post-image cursor.\n */\nfunction fillImagePositions(\n  rows: { t: Int32Array; h: Int32Array; w: Int32Array },\n  at: number,\n  start: number,\n  grid: [number, number, number],\n  mergeSize: number,\n): { next: number; cursor: number } {\n  const [gt, gh, gw] = grid;\n  const hm = gh / mergeSize;\n  const wm = gw / mergeSize;\n  let i = at;\n  for (let tt = 0; tt < gt; tt++) {\n    for (let h = 0; h < hm; h++) {\n      for (let w = 0; w < wm; w++) {\n        rows.t[i] = start + tt;\n        rows.h[i] = start + h;\n        rows.w[i] = start + w;\n        i++;\n      }\n    }\n  }\n  return { next: i, cursor: start + Math.max(hm, wm) };\n}\n\nexport function buildMRoPEPositionIds(\n  inputIds: Int32Array | Uint32Array | number[],\n  imageGrids: Array<[number, number, number]>,\n  imageTokenId: number,\n  mergeSize: number,\n): Int32Array {\n  const seq = inputIds.length;\n  const rows = { t: new Int32Array(seq), h: new Int32Array(seq), w: new Int32Array(seq) };\n\n  let i = 0;\n  let cursor = 0;\n  let imageIdx = 0;\n  while (i < seq) {\n    if (inputIds[i] === imageTokenId) {\n      const res = fillImagePositions(rows, i, cursor, imageGrids[imageIdx++], mergeSize);\n      i = res.next;\n      cursor = res.cursor;\n    } else {\n      rows.t[i] = cursor;\n      rows.h[i] = cursor;\n      rows.w[i] = cursor;\n      cursor++;\n      i++;\n    }\n  }\n\n  const out = new Int32Array(3 * seq);\n  out.set(rows.t, 0);\n  out.set(rows.h, seq);\n  out.set(rows.w, 2 * seq);\n  return out;\n}\n\n/**\n * Per-pair frequency→dimension assignment for interleaved M-RoPE, matching\n * Qwen3_5TextRotaryEmbedding.apply_interleaved_mrope. For pair index i in\n * [0, sum(section)) the position component is section-cyclic: T,H,W,T,H,W,...\n * but each component capped at its section count. Returns an array of length\n * (rope_dim/2) with values 0=T, 1=H, 2=W.\n */\nexport function mropeFreqDims(mropeSection: [number, number, number]): Int32Array {\n  const total = mropeSection[0] + mropeSection[1] + mropeSection[2];\n  const dims = new Int32Array(total); // default 0 (T)\n  for (let d = 1; d <= 2; d++) {\n    const length = mropeSection[d] * 3;\n    for (let idx = d; idx < length && idx < total; idx += 3) {\n      dims[idx] = d;\n    }\n  }\n  return dims;\n}\n\n/**\n * Build the interleaved-M-RoPE cos/sin tables [seq, rope_dim] from 3D position\n * ids, matching Qwen3_5TextRotaryEmbedding.forward:\n *   freqs[d][i] = pos[d] * inv_freq[i],  inv_freq[i] = 1/theta^(2i/rope_dim)\n *   freq[i] picks component mropeFreqDims[i]; emb = cat(freqs, freqs).\n * cos/sin have length seq*rope_dim. For text-only (all 3 pos rows equal) this\n * reduces exactly to standard 1D partial RoPE.\n *\n * @param positionIds3 [3, seq] as produced by buildMRoPEPositionIds.\n * @param ropeDim      number of rotated dims per head (head_dim * partial_factor).\n */\nexport function buildMRoPECosSin(\n  positionIds3: Int32Array,\n  seq: number,\n  ropeDim: number,\n  theta: number,\n  mropeSection: [number, number, number],\n): { cos: Float32Array; sin: Float32Array } {\n  const half = ropeDim / 2; // == sum(mropeSection)\n  const freqDims = mropeFreqDims(mropeSection);\n  const invFreq = new Float64Array(half);\n  for (let i = 0; i < half; i++) {\n    invFreq[i] = 1.0 / theta ** ((2 * i) / ropeDim);\n  }\n  const cos = new Float32Array(seq * ropeDim);\n  const sin = new Float32Array(seq * ropeDim);\n  for (let n = 0; n < seq; n++) {\n    const base = n * ropeDim;\n    for (let i = 0; i < half; i++) {\n      const dim = freqDims[i]; // 0=T,1=H,2=W\n      const pos = positionIds3[dim * seq + n];\n      const angle = pos * invFreq[i];\n      const c = Math.cos(angle);\n      const s = Math.sin(angle);\n      // emb = cat(freqs, freqs): duplicate into [0,half) and [half,ropeDim)\n      cos[base + i] = c;\n      sin[base + i] = s;\n      cos[base + half + i] = c;\n      sin[base + half + i] = s;\n    }\n  }\n  return { cos, sin };\n}\n","// ============================================\n// iOS Model Guards & Device Capability Detection\n// ============================================\n\n// ============================================\n// Real native-engine model ids\n// ============================================\n//\n// The website loads models by their actual repo id (MLX 4-bit on-device builds,\n// or the upstream Qwen/Liquid repos on desktop). The guard matches on those ids\n// — NOT the old ONNX-era shorthands — so it must stay in sync with the real\n// checkpoints the engine can run. Matching is substring-based on a normalized\n// id so both `mlx-community/Qwen3.5-0.8B-4bit` and `Qwen/Qwen3.5-0.8B` resolve.\n\n/** Recommended safe model ids per modality (used as fallbacks on mobile). */\nconst SAFE_MOBILE_CHAT = \"mlx-community/Qwen3.5-0.8B-4bit\";\n\n/**\n * Approximate on-device (INT4) memory footprint in MB for the models the native\n * engine actually ships. Used for memory-aware selection and messaging.\n */\nexport const MODEL_SIZES: Record<string, number> = {\n  // Chat models (INT4, on-device)\n  \"qwen3.5-0.8b\": 650, // ~0.65GB\n  \"qwen3.5-2b\": 1700, // ~1.7GB\n  \"gemma-4-e2b\": 3600, // ~3.6GB\n  \"lfm2.5-350m\": 300, // ~0.3GB\n  // TTS models\n  \"kokoro-82m\": 350,\n  \"supertonic-66m\": 300,\n  // STT models\n  \"whisper-tiny\": 150,\n  \"whisper-tiny.en\": 150,\n  \"whisper-small\": 500,\n  // Embedding models\n  \"all-minilm-l6-v2\": 100,\n};\n\n/**\n * Normalize a repo/model id to a lowercase token stream for substring matching\n * (strips org prefixes' punctuation while preserving the model name tokens).\n */\nfunction normalizeId(modelId: string): string {\n  return modelId.toLowerCase().replace(/[^a-z0-9]/g, \"-\");\n}\n\n/**\n * iOS (WKWebView) model classification keyed off the REAL native-engine ids.\n *\n * - blocked: too large for the WKWebView memory ceiling on iPhone — will crash.\n *   gemma-4-e2b (~3.6GB) plus any vision checkpoint (the vision encoder pushes\n *   the working set well past what an iPhone can hold).\n * - risky: Qwen3.5-2B (~1.7GB) OOM-crashes iPhone WKWebView in practice — now\n *   treated as unsafe to auto-load (flagged `risky` so a UI can offer to force it).\n * - everything else (Qwen3.5-0.8B ~0.65GB, LFM2.5-350M) is allowed everywhere.\n */\nconst IOS_MODEL_LIMITS = {\n  /** Substrings (normalized) of ids that HARD-BLOCK on iPhone. */\n  blocked: [\"gemma-4-e2b\", \"gemma-4-e4b\"],\n  /** Substrings that mark a vision checkpoint (blocked on iPhone). */\n  visionMarkers: [\"vision\", \"-vl-\", \"-vl\", \"vlm\", \"image-text\", \"-it-vision\"],\n  /** Substrings of ids that WARN on iPhone (Qwen3.5-2B class). */\n  risky: [\"qwen3.5-2b\", \"qwen3-5-2b\"],\n  /** Maximum total memory budget in MB for iOS WKWebView. */\n  maxBudgetMB: 1800,\n} as const;\n\n/**\n * Check if a model is safe to load on the current device.\n * Returns guidance specific to iOS memory constraints. Matches on the real\n * native-engine repo ids (MLX 4-bit / upstream Qwen / Liquid).\n */\nexport function isModelSafeForDevice(modelId: string): {\n  safe: boolean;\n  /**\n   * Borderline: may run on the newest hardware but is prone to OOM. Reported\n   * unsafe (`safe: false`) so callers block by default; a UI can use `risky` to\n   * offer an explicit \"load anyway\" on capable devices.\n   */\n  risky: boolean;\n  reason: string;\n  recommendation?: string;\n  maxSafeModel?: string;\n} {\n  const ua = typeof navigator !== \"undefined\" ? navigator.userAgent : \"\";\n  const isIPhone = /iPhone|iPod/.test(ua);\n  const isIPad = /iPad/.test(ua);\n  const normalizedId = normalizeId(modelId);\n\n  const isVision = IOS_MODEL_LIMITS.visionMarkers.some((m) =>\n    normalizedId.includes(normalizeId(m)),\n  );\n  const isBlocked =\n    IOS_MODEL_LIMITS.blocked.some((m) => normalizedId.includes(normalizeId(m))) || isVision;\n  const isRisky = IOS_MODEL_LIMITS.risky.some((m) => normalizedId.includes(normalizeId(m)));\n\n  // iPhone is the hard constraint (smallest WKWebView budget). iPad has more\n  // headroom, so only the genuinely huge / vision checkpoints are blocked there.\n  if (isIPhone) {\n    if (isBlocked) {\n      if (isVision) {\n        return {\n          safe: false,\n          risky: false,\n          reason:\n            \"This model can also read images, which needs more memory than an iPhone gives a web page — it would run out of memory and crash.\",\n          recommendation:\n            \"Use the standard model for chat on your phone, or open this on a Mac or PC to work with images. On newer iPhones, adding this site to your Home Screen can unlock extra memory and may let image models run.\",\n          maxSafeModel: SAFE_MOBILE_CHAT,\n        };\n      }\n      return {\n        safe: false,\n        risky: false,\n        reason: \"This model is too large to fit in an iPhone's memory and would crash.\",\n        recommendation:\n          \"Use the standard model on your phone — it's fast and fully supported. The larger models need a Mac or PC. Adding this site to your Home Screen can also unlock extra memory on newer iPhones.\",\n        maxSafeModel: SAFE_MOBILE_CHAT,\n      };\n    }\n    if (isRisky) {\n      return {\n        safe: false,\n        risky: true,\n        reason: \"This model is on the large side for an iPhone and may run out of memory.\",\n        recommendation:\n          \"The standard model runs smoothly on iPhone. You can use this larger one on an iPad or computer.\",\n        maxSafeModel: SAFE_MOBILE_CHAT,\n      };\n    }\n    return { safe: true, risky: false, reason: \"Model is within iPhone memory limits.\" };\n  }\n\n  if (isIPad) {\n    // iPad tolerates the 2B class; only block the huge / vision checkpoints.\n    if (isBlocked) {\n      const reason = isVision\n        ? \"This model can also read images, which needs a lot of memory — it may run out and crash on iPad.\"\n        : \"This model is very large and may run out of memory on iPad.\";\n      return {\n        safe: false,\n        risky: false,\n        reason,\n        recommendation: \"Use the standard model or Qwen3.5 2B on iPad.\",\n        maxSafeModel: \"mlx-community/Qwen3.5-2B-4bit\",\n      };\n    }\n    return { safe: true, risky: false, reason: \"Model is within iPad memory limits.\" };\n  }\n\n  // Android - block the huge / vision checkpoints, allow the rest.\n  const isAndroid = /Android/.test(ua);\n  if (isAndroid && isBlocked) {\n    return {\n      safe: false,\n      risky: false,\n      reason: \"This model is very large and may run out of memory on this phone.\",\n      recommendation: \"Use the standard model or Qwen3.5 2B instead.\",\n      maxSafeModel: \"mlx-community/Qwen3.5-2B-4bit\",\n    };\n  }\n\n  // Desktop - all models are safe.\n  return { safe: true, risky: false, reason: \"Desktop browser has sufficient memory.\" };\n}\n\n/**\n * Get recommended models based on device memory and capabilities.\n * Helps prevent OOM crashes on low-memory mobile devices.\n */\nexport function getRecommendedModels(): {\n  chat: string;\n  tts: string;\n  stt: string;\n  embedding: string;\n  reason: string;\n  deviceMemory: number | null;\n  isMobile: boolean;\n} {\n  const ua = typeof navigator !== \"undefined\" ? navigator.userAgent : \"\";\n  const deviceMemory = typeof navigator !== \"undefined\" ? (navigator as any).deviceMemory : null;\n  const isMobile = /iPhone|iPad|iPod|Android|Mobile/.test(ua);\n\n  // Estimate available memory (deviceMemory reports total GB, not available)\n  // Mobile devices typically have less free memory due to OS overhead\n  const effectiveMemory = deviceMemory ? (isMobile ? deviceMemory * 0.4 : deviceMemory * 0.6) : 4;\n  const availableMB = effectiveMemory * 1024;\n\n  let chat: string;\n  let reason: string;\n\n  if (availableMB < 600) {\n    chat = \"LiquidAI/LFM2.5-350M\";\n    reason = \"Very low memory device - using smallest model (LFM2.5-350M)\";\n  } else if (isMobile && availableMB < 2200) {\n    chat = SAFE_MOBILE_CHAT;\n    reason = \"Mobile device - using Qwen3.5-0.8B to stay within the WKWebView memory limit\";\n  } else if (availableMB < 2200) {\n    chat = SAFE_MOBILE_CHAT;\n    reason = \"Standard model for moderate memory (Qwen3.5-0.8B)\";\n  } else {\n    chat = \"mlx-community/Qwen3.5-2B-4bit\";\n    reason = \"High memory available - using Qwen3.5-2B for better quality\";\n  }\n\n  return {\n    chat,\n    tts: \"kokoro-82m\",\n    stt: \"whisper-tiny.en\",\n    embedding: \"all-MiniLM-L6-v2\",\n    reason,\n    deviceMemory,\n    isMobile,\n  };\n}\n\n// ============================================\n// Session Phase Tracking (Reload Detection)\n// ============================================\n\ntype DownloadPhase = \"idle\" | \"downloading\" | \"caching\" | \"initializing\" | \"ready\" | \"error\";\n\nexport const SESSION_STORAGE_KEY = \"gerbil_session_phase\";\n\nexport type SessionState = {\n  phase: DownloadPhase;\n  modelId: string | null;\n  sessionId: string;\n  timestamp: number;\n  bytesDownloaded?: number;\n  totalBytes?: number;\n};\n\n/**\n * Generate a unique session ID for tracking across reloads.\n */\nfunction generateSessionId(): string {\n  return `${Date.now()}-${Math.random().toString(36).slice(2, 9)}`;\n}\n\n/**\n * Get or create the current session ID.\n */\nfunction getSessionId(): string {\n  if (typeof localStorage === \"undefined\") return generateSessionId();\n\n  let sessionId = sessionStorage.getItem(\"gerbil_session_id\");\n  if (!sessionId) {\n    sessionId = generateSessionId();\n    sessionStorage.setItem(\"gerbil_session_id\", sessionId);\n  }\n  return sessionId;\n}\n\n/**\n * Set the current download/initialization phase.\n * Used to detect if a reload happened during a critical operation.\n */\nexport function setDownloadPhase(\n  phase: DownloadPhase,\n  modelId?: string,\n  progress?: { bytesDownloaded: number; totalBytes: number },\n): void {\n  if (typeof localStorage === \"undefined\") return;\n\n  const state: SessionState = {\n    phase,\n    modelId: modelId || null,\n    sessionId: getSessionId(),\n    timestamp: Date.now(),\n    bytesDownloaded: progress?.bytesDownloaded,\n    totalBytes: progress?.totalBytes,\n  };\n\n  localStorage.setItem(SESSION_STORAGE_KEY, JSON.stringify(state));\n}\n\n/**\n * Get the last known download phase from storage.\n */\nexport function getDownloadPhase(): SessionState | null {\n  if (typeof localStorage === \"undefined\") return null;\n\n  try {\n    const raw = localStorage.getItem(SESSION_STORAGE_KEY);\n    if (!raw) return null;\n    return JSON.parse(raw) as SessionState;\n  } catch {\n    return null;\n  }\n}\n\n/**\n * Detect if the page reloaded during a model download/initialization.\n * This typically indicates an iOS memory crash.\n *\n * @returns Detection result with recommended action\n */\nexport function detectMemoryCrash(): {\n  crashed: boolean;\n  phase?: DownloadPhase;\n  modelId?: string;\n  timeSinceCrash?: number;\n  recommendation?: string;\n} {\n  const lastState = getDownloadPhase();\n  const currentSessionId = getSessionId();\n\n  if (!lastState) {\n    return { crashed: false };\n  }\n\n  // If session ID changed and we were in a critical phase, it's likely a crash/reload\n  const criticalPhases: DownloadPhase[] = [\"downloading\", \"caching\", \"initializing\"];\n  const wasInCriticalPhase = criticalPhases.includes(lastState.phase);\n  const sessionChanged = lastState.sessionId !== currentSessionId;\n  const timeSinceCrash = Date.now() - lastState.timestamp;\n\n  // Only count as crash if it happened recently (within 5 minutes) and session changed\n  const recentCrash = timeSinceCrash < 5 * 60 * 1000;\n\n  if (wasInCriticalPhase && sessionChanged && recentCrash) {\n    // Clear the state to prevent repeated detection\n    localStorage.removeItem(SESSION_STORAGE_KEY);\n\n    return {\n      crashed: true,\n      phase: lastState.phase,\n      modelId: lastState.modelId || undefined,\n      timeSinceCrash,\n      recommendation:\n        lastState.modelId && /2b|gemma|vision/i.test(lastState.modelId)\n          ? \"The model was too large for your device. Try Qwen3.5-0.8B instead.\"\n          : \"Your device ran out of memory. Try a smaller model or use a desktop browser.\",\n    };\n  }\n\n  return { crashed: false };\n}\n\n/**\n * Clear session phase (call when model loads successfully).\n */\nexport function clearDownloadPhase(): void {\n  if (typeof localStorage === \"undefined\") return;\n  localStorage.removeItem(SESSION_STORAGE_KEY);\n}\n\n// ============================================\n// WebKit Submit-Granularity (group size) Probe\n// ============================================\n//\n// On WebKit/iOS WebGPU, decode speed is bound by GPU round-trips: the executor\n// groups `webkitGroupSize` dispatches per command buffer, then submits + drains\n// (queue.onSubmittedWorkDone) per group. Larger groups amortize the drains and\n// can be ~5x faster on iPad. BUT the safe ceiling is device-dependent in three\n// classes:\n//   1. Newer WebKit (iPad 26.5): high group is correct + fast.\n//   2. Older WebKit: high group yields zero/garbage logits (wrong output).\n//   3. iPhone (iOS 18.7): high group HARD-CRASHES the GPU process (page dies).\n//\n// The crash class is the hard one: a bad group kills the page before any code\n// can record the result. So we use a localStorage BREADCRUMB that OUTLIVES the\n// page kill: we persist the candidate we're about to try BEFORE running any\n// GPU-heavy work. If the page survives, a later promotion clears `trying` and\n// records it as known-good. If the page crashed, `trying` is still set on the\n// next load, so we KNOW that rung crashed and cap below it.\n//\n// The probe escalates UP from the safe floor (never starts optimistic — a\n// first-visit crash is a terrible UX) and persists per-device in localStorage.\n\nexport const WEBKIT_GROUP_PROBE_KEY = \"gerbil-webkit-group-v1\";\n\n/**\n * Candidate group-size ladder. We escalate UP one rung per page load. 128 is\n * treated as effectively \"batch-all\" for this model — kept simple on purpose.\n */\nexport const WEBKIT_GROUP_LADDER = [1, 8, 32, 64, 128] as const;\n\nexport type WebkitGroupProbe = {\n  /** Largest group size proven crash-free AND correct on this device. Starts at 1. */\n  knownGood: number;\n  /** Candidate being attempted this page-load (null when not probing). */\n  trying: number | null;\n  /** True once we hit a crash/incorrect ceiling and should stop escalating. */\n  capped: boolean;\n};\n\nconst DEFAULT_PROBE: WebkitGroupProbe = { knownGood: 1, trying: null, capped: false };\n\n/** Read the persisted WebKit group probe record (guarded; safe on node). */\nexport function readGroupProbe(): WebkitGroupProbe {\n  if (typeof localStorage === \"undefined\") return { ...DEFAULT_PROBE };\n  try {\n    const raw = localStorage.getItem(WEBKIT_GROUP_PROBE_KEY);\n    if (!raw) return { ...DEFAULT_PROBE };\n    const parsed = JSON.parse(raw) as Partial<WebkitGroupProbe>;\n    return {\n      knownGood:\n        typeof parsed.knownGood === \"number\" && parsed.knownGood >= 1 ? parsed.knownGood : 1,\n      trying: typeof parsed.trying === \"number\" ? parsed.trying : null,\n      capped: parsed.capped === true,\n    };\n  } catch {\n    return { ...DEFAULT_PROBE };\n  }\n}\n\n/** Persist the WebKit group probe record (guarded; no-op on node). */\nexport function writeGroupProbe(rec: WebkitGroupProbe): void {\n  if (typeof localStorage === \"undefined\") return;\n  try {\n    localStorage.setItem(WEBKIT_GROUP_PROBE_KEY, JSON.stringify(rec));\n  } catch {\n    // Best-effort; private mode / quota can throw.\n  }\n}\n\n/**\n * The validated batching target. Originally 128 from an iPad sweep on a much\n * slower engine build (peak 26.9 tok/s at 128, 256 within 2.6%); the 2026-07\n * kernel campaign made per-dispatch GPU time far smaller, which moves the\n * amortization optimum UP. An on-device iPhone sweep (bench harness, engine\n * 1.6.1) measured Qwen3.5-0.8B sampled decode: g32=19, g128≈33, g256=43,\n * g1024=45, g4096=44 — a +36% win at 1024 vs the old 128 target — while\n * LFM2.5 is flat from g32 upward (no cost). The crash breadcrumb still caps\n * any device that can't sustain it.\n */\nconst NONPHONE_TARGET_GROUP = 1024;\n\n/**\n * Resolve the WebKit group size to use this session, recording `trying` as a\n * side effect so a crash this load is detectable on the next load.\n *\n * Algorithm (only meaningful on WebKit; inert otherwise):\n *  1. Read the record (default {knownGood:1, trying:null, capped:false}).\n *  2. If `trying !== null` on entry → the previous load set it but never cleared\n *     it → that load CRASHED at `trying`. Cap there, keep `knownGood`, clear\n *     `trying`. Use `knownGood` this session.\n *  3. Else if !capped and there is a rung above `knownGood` → set `trying = next`,\n *     persist BEFORE any GPU work, and use it (we're escalating).\n *  4. Else → use `knownGood`.\n *\n * @returns the group size to use this session.\n */\nexport function resolveWebkitGroupSize(args: {\n  override?: number;\n  isWebKit: boolean;\n  /**\n   * Memory-constrained devices (phones) are the \"crash class\": batching crashes\n   * the GPU process, and discovering that costs one user-visible crash before the\n   * breadcrumb caps it. When `conservative` is set we never auto-escalate — the\n   * device stays at its proven floor (group=1 on first run) and never risks that\n   * calibration crash. `?group=N` still lets such a device opt in explicitly.\n   */\n  conservative?: boolean;\n}): number {\n  // Explicit ?group=N override wins and skips the probe entirely.\n  if (args.override && args.override > 0) return args.override;\n  // Non-WebKit (desktop / Dawn / node) never groups — floor of 1, no bookkeeping.\n  if (!args.isWebKit) return 1;\n\n  const rec = readGroupProbe();\n\n  // Step 2: a stale `trying` means last load crashed at that rung.\n  if (rec.trying !== null) {\n    const crashedAt = rec.trying;\n    const next: WebkitGroupProbe = {\n      knownGood: rec.knownGood,\n      trying: null,\n      capped: true,\n    };\n    writeGroupProbe(next);\n    console.log(\n      `[engine] webkit group probe: previous load crashed at group=${crashedAt} → capping at knownGood=${next.knownGood}`,\n    );\n    return next.knownGood;\n  }\n\n  // Step 3: non-phone WebKit (iPad / Mac Safari) jumps straight to the validated\n  // sweet spot (group=128) instead of climbing rung-by-rung — these devices\n  // tolerate batching (iPad ran every rung to 256 coherently), so there's no\n  // reason to crawl. The breadcrumb (Step 2) caps it down if a particular device\n  // can't sustain it. Phones (conservative) NEVER escalate — they stay at the\n  // proven floor (group=1) to avoid the one calibration crash batching costs on\n  // the crash class.\n  if (!rec.capped && !args.conservative && rec.knownGood < NONPHONE_TARGET_GROUP) {\n    const next: WebkitGroupProbe = {\n      knownGood: rec.knownGood,\n      trying: NONPHONE_TARGET_GROUP,\n      capped: false,\n    };\n    writeGroupProbe(next); // PERSIST BEFORE any GPU-heavy work runs\n    console.log(\n      `[engine] webkit group probe: trying target group=${NONPHONE_TARGET_GROUP} (knownGood=${rec.knownGood})`,\n    );\n    return NONPHONE_TARGET_GROUP;\n  }\n\n  // Step 4: nothing to escalate (capped or at top rung) — use the proven floor.\n  console.log(\n    `[engine] webkit group probe: knownGood=${rec.knownGood} trying=null capped=${rec.capped} → using ${rec.knownGood}`,\n  );\n  return rec.knownGood;\n}\n\n/**\n * Promote (or cap) the WebKit group probe after the first successful forward.\n *\n * Call this once per page-load, after the model has loaded AND a first forward\n * completed without the page dying. The breadcrumb already handles the crash\n * class (the page death leaves `trying` set for the next load); this handles the\n * wrong-output class and records success.\n *\n * @param correct  true if the first forward produced non-corrupt output.\n *   - correct  → promote: knownGood = trying, trying = null.\n *   - incorrect → cap: keep knownGood at the prior rung, trying = null, capped.\n */\nexport function promoteGroupProbe(correct: boolean): void {\n  if (typeof localStorage === \"undefined\") return;\n  const rec = readGroupProbe();\n  // Nothing was being attempted this load — nothing to promote.\n  if (rec.trying === null) return;\n\n  if (correct) {\n    writeGroupProbe({ knownGood: rec.trying, trying: null, capped: rec.capped });\n    console.log(`[engine] webkit group probe: PROMOTED group=${rec.trying} to known-good`);\n  } else {\n    // Wrong-output class: this rung is incorrect (but did not crash). Keep the\n    // prior knownGood and stop escalating.\n    writeGroupProbe({ knownGood: rec.knownGood, trying: null, capped: true });\n    console.log(\n      `[engine] webkit group probe: group=${rec.trying} produced INCORRECT output → capping at knownGood=${rec.knownGood}`,\n    );\n  }\n}\n","/**\n * WebGPUEngine -- gerbil's native WebGPU inference engine.\n *\n * Public API:\n *   const engine = await WebGPUEngine.create({ repo: \"Qwen/Qwen3.5-0.8B\" });\n *   const result = await engine.generate(\"What is 2+2?\");\n *   engine.destroy();\n */\n\nimport {\n  dequantizeGemma4VisionProjection,\n  ensureGemma4VisionEmbedderNorms,\n  generateGemma4VisionGraph,\n  patchGemma4VisionClips,\n  resolveGemma4VisionInfo,\n} from \"./architectures/gemma4_vision.js\";\nimport type { KvMode } from \"./architectures/index.js\";\nimport { generateQwen3_5VisionGraph } from \"./architectures/qwen3_5_vision.js\";\nimport type { GPUContext, GPUDiagnosticResult } from \"./device.js\";\nimport { clearPipelineCache, initGPU, verifyGPU } from \"./device.js\";\nimport { Executor } from \"./executor.js\";\nimport type { ModelArchConfig, ModelCapabilities, ModelGraph } from \"./ir.js\";\nimport { type KaniDType, KaniTTS } from \"./kani-tts.js\";\nimport { buildLoRADeltas, fetchAdapter, type LoRADelta } from \"./lora.js\";\nimport { createKeyMapperForArch, type LoadModelOptions, loadModel } from \"./model-loader.js\";\n\nexport { createDefaultHFKeyMapper } from \"./ir.js\";\n// LoRA adapter (\"flavor\") loading — public engine capability. High-level users go\n// through Gerbil.loadAdapter; these are the primitives (also used by tests + the\n// serving side to parse/merge adapters directly).\nexport {\n  type AdapterConfig,\n  type AdapterSource,\n  applyLoRAToStore,\n  buildLoRADeltas,\n  fetchAdapter,\n  type LoRADelta,\n  splitFusedQGateDelta,\n} from \"./lora.js\";\n\nimport { cleanSuggestion } from \"./autocomplete-clean.js\";\nimport {\n  RequestScheduler,\n  type SchedulerHost,\n  type ServeOptions,\n  type ServeResult,\n} from \"./batch/scheduler.js\";\nimport { Swarm, type SwarmOptions } from \"./batch/swarm.js\";\nimport { type GepardDType, GepardTTS } from \"./gepard-tts.js\";\nimport { loadOuteSpeaker, type OuteDType, type OuteSpeaker, OuteTTS } from \"./outetts.js\";\nimport { type ParlerSpeakOptions, ParlerTTS } from \"./parler-tts.js\";\nimport { type SamplingParams, sampleToken } from \"./sampler.js\";\nimport type { ChatMessage, Tokenizer } from \"./tokenizer.js\";\nimport type { VisionExecutor } from \"./vision-executor.js\";\nimport { VisionExecutor as VisionExecutorImpl } from \"./vision-executor.js\";\nimport {\n  buildGemma4VisionPositionTensors,\n  buildMRoPECosSin,\n  buildMRoPEPositionIds,\n  buildVisionPositionTensors,\n  type ImageProcessorConfig,\n  preprocessImage,\n  preprocessImageGemma4,\n  QWEN3_5_IMAGE_PROCESSOR,\n} from \"./vision-preprocess.js\";\nimport { remapPrunedToken } from \"./vocab-prune.js\";\n\n// Per-capability default models live in a standalone module so the React hooks\n// can resolve defaults without statically importing this (heavy) engine module.\nexport {\n  DEFAULT_MODELS,\n  isTTSRepo,\n  OUTETTS_ASSETS,\n  OUTETTS_PRESET_VOICES,\n  resolveDefaultRepo,\n} from \"./defaults.js\";\n\nimport { promoteGroupProbe, resolveWebkitGroupSize } from \"../browser/device-guards.js\";\nimport {\n  DEFAULT_MODELS,\n  isTTSRepo,\n  OUTETTS_ASSETS,\n  OUTETTS_PRESET_VOICES,\n  resolveDefaultRepo,\n} from \"./defaults.js\";\n\nexport interface WebGPUEngineOptions extends Omit<LoadModelOptions, \"repo\"> {\n  /**\n   * HuggingFace repo ID (e.g. \"mlx-community/Qwen3.5-0.8B-4bit\") or full URL.\n   * Optional — when omitted, a sensible default is chosen for the requested\n   * capability (text, vision, or embeddings). See {@link DEFAULT_MODELS}.\n   */\n  repo?: string;\n  /** Max sequence length (default: from model config, capped at 4096). */\n  maxSeqLen?: number;\n  /** Override KV mode: \"f32\", \"native-f16\", or \"packed-f16\". Auto-detected if omitted. */\n  kvMode?: KvMode;\n  /**\n   * Build the vision encoder (Qwen3.5 ViT) alongside the text model so\n   * `encodeImage()` can turn image patches into merged image-embedding tokens.\n   * Only valid for vision-capable checkpoints (Qwen3.5). Downloads the ~192MB\n   * vision tower. Default: false.\n   */\n  enableVision?: boolean;\n  /** Max patches the vision encoder can process in one call (default 4096). */\n  maxVisionPatches?: number;\n}\n\nexport interface EncodeImageResult {\n  /** Merged image-embedding tokens, row-major [rows * dim]. */\n  embeds: Float32Array;\n  /** Number of merged tokens (numPatches / spatial_merge_size^2). */\n  rows: number;\n  /** Embedding dimension (out_hidden_size, 1024 for Qwen3.5). */\n  dim: number;\n}\n\nexport interface EmbedOptions {\n  /**\n   * Instruction prefix for query embeddings (Qwen3-Embedding convention:\n   * \"Instruct: {task}\\nQuery:{text}\"). Omit for document embeddings.\n   */\n  instruction?: string;\n  /**\n   * EmbeddingGemma task prefix. The model is asymmetric: queries and documents\n   * use different prefixes (`task: search result | query: ` vs `title: none |\n   * text: `). Pass \"query\" for search queries and \"document\" for the corpus\n   * being searched. Defaults to \"query\" for EmbeddingGemma when omitted. Ignored\n   * by non-Gemma embedding models (use `instruction` for Qwen3-Embedding).\n   */\n  taskType?: \"query\" | \"document\";\n  /**\n   * Override the raw task prefix prepended to the text (EmbeddingGemma). When\n   * set, takes precedence over `taskType`. Use for non-retrieval tasks, e.g.\n   * \"task: clustering | query: \" or \"task: classification | query: \".\n   */\n  taskPrompt?: string;\n  /** Max tokens to encode (longer inputs are truncated). Default: model context, capped at maxSeqLen. */\n  maxTokens?: number;\n}\n\n/** EmbeddingGemma task prefixes (from config_sentence_transformers.json prompts). */\nconst EMBEDDING_GEMMA_PROMPTS = {\n  query: \"task: search result | query: \",\n  document: \"title: none | text: \",\n} as const;\n\nexport interface GenerateOptions {\n  /** Max tokens to generate (default: 512). */\n  maxTokens?: number;\n  /** Stop generation on these strings. */\n  stopSequences?: string[];\n  /** Sampling parameters. */\n  sampling?: SamplingParams;\n  /** System prompt to prepend. */\n  systemPrompt?: string;\n  /** Callback for each generated token (for streaming). */\n  onToken?: (token: string, meta?: { tokenIndex: number; tps: number; elapsedMs: number }) => void;\n}\n\nexport interface GenerateResult {\n  /** Generated text. */\n  text: string;\n  /** Number of tokens generated. */\n  tokensGenerated: number;\n  /** Tokens per second. */\n  tokensPerSecond: number;\n  /** Total generation time in ms. */\n  totalTime: number;\n  /** Why generation stopped. */\n  finishReason: \"eos\" | \"max_tokens\" | \"stop_sequence\";\n  /**\n   * Decode-only wall time in ms (excludes prefill). Set by the batched path\n   * (generateBatch) so aggregate decode throughput is measurable directly.\n   */\n  decodeTime?: number;\n  /**\n   * Raw generated token ids, in order. Used by parity harnesses (e.g. the\n   * paged-KV token-exactness gate) to compare runs without decode ambiguity.\n   */\n  tokenIds?: number[];\n  /** Thinking content if model produced it (future). */\n  thinking?: string;\n}\n\n/**\n * A minimal JSON-schema-ish shape used by {@link WebGPUEngine.generateObject} to\n * validate generated output without pulling in a schema library. Only `required`\n * and `properties` are inspected (presence of required keys). Pass a predicate\n * function instead for arbitrary validation.\n */\nexport interface ObjectSchema {\n  /** Keys that must be present on the parsed object. */\n  required?: string[];\n  /** Property descriptors (only the key set is used for validation). */\n  properties?: Record<string, unknown>;\n  /** Allow extra schema fields (type, etc.) without TS complaints. */\n  [key: string]: unknown;\n}\n\n/** Validator for {@link WebGPUEngine.generateObject}: a schema object or predicate. */\nexport type ObjectValidator<T = unknown> = ObjectSchema | ((o: T) => boolean);\n\nexport interface GenerateObjectOptions extends GenerateOptions {\n  /**\n   * Validation target. Either a predicate `(o) => boolean` or a minimal\n   * JSON-schema-ish object with `required`/`properties` (required keys must\n   * exist). Omit to only require syntactically valid JSON.\n   */\n  schema?: ObjectValidator;\n  /**\n   * Max RETRIES after the first attempt (so up to `maxRetries + 1` generations).\n   * Default: 4.\n   */\n  maxRetries?: number;\n}\n\nexport interface GenerateObjectResult<T = unknown> {\n  /** The parsed + validated object (or array). */\n  object: T;\n  /** The raw model text the object was extracted from. */\n  text: string;\n  /** How many generation attempts it took (1 = first try). */\n  attempts: number;\n}\n\nexport interface IntegrityCheckEntry {\n  label: string;\n  length: number;\n  sum: number;\n  first4: number[];\n  argmax: number;\n  maxVal: number;\n  match?: \"PASS\" | \"FAIL\";\n  refSum?: number;\n  refArgmax?: number;\n  error?: string;\n  note?: string;\n}\n\nexport interface IntegrityCheckResult {\n  checks: IntegrityCheckEntry[];\n  allPass: boolean;\n}\n\n/**\n * Extract the first complete JSON object or array from a model's text output.\n *\n * Tolerant of prose, markdown, and ```json code fences: it scans for the first\n * `{` or `[`, then walks forward tracking string/escape state and brace/bracket\n * depth to find the matching close, returning the balanced substring. Returns\n * `undefined` when no JSON-looking span is present. Used by\n * {@link WebGPUEngine.generateObject}.\n */\nfunction extractJson(text: string): string | undefined {\n  // Strip code-fence markers; the brace scan below ignores surrounding prose.\n  const cleaned = text.replace(/```(?:json)?/gi, \"\");\n  const start = cleaned.search(/[{[]/);\n  if (start === -1) return undefined;\n  const open = cleaned[start];\n  const close = open === \"{\" ? \"}\" : \"]\";\n  let depth = 0;\n  let inString = false;\n  let escaped = false;\n  for (let i = start; i < cleaned.length; i++) {\n    const ch = cleaned[i];\n    if (inString) {\n      if (escaped) {\n        escaped = false;\n      } else if (ch === \"\\\\\") {\n        escaped = true;\n      } else if (ch === '\"') {\n        inString = false;\n      }\n      continue;\n    }\n    if (ch === '\"') {\n      inString = true;\n    } else if (ch === open) {\n      depth++;\n    } else if (ch === close) {\n      depth--;\n      if (depth === 0) return cleaned.slice(start, i + 1);\n    }\n  }\n  return undefined;\n}\n\n/**\n * The main WebGPU inference engine.\n *\n * Usage:\n *   const engine = await WebGPUEngine.create({ repo: \"Qwen/Qwen3.5-0.8B\" });\n *   const result = await engine.generate(\"Hello!\");\n *   console.log(result.text);\n *   engine.destroy();\n */\n\n/** System prompt that locks the model into \"continue the text\" autocomplete mode. */\nconst AUTOCOMPLETE_SYSTEM = [\n  \"You are an inline autocomplete engine.\",\n  \"Continue the user's text with a brief, natural continuation of the SAME sentence or thought.\",\n  \"Output ONLY the continuation text — no preamble, no quotes, no explanations, no assistant voice.\",\n  \"Do not answer questions; just continue the writing.\",\n  'Example — input: \"The quick brown fox\" → continuation: \" jumps over the lazy dog.\"',\n].join(\" \");\n\n/** A tool the agent loop can call. Framework-agnostic; pass `defineTool` output or a plain object. */\nexport interface AgentTool {\n  name: string;\n  description: string;\n  /** Optional JSON-schema-ish parameter description shown to the model. */\n  parameters?: unknown;\n  execute: (args: any) => Promise<string> | string;\n}\n\n/** One step of an agentic tool-calling run, for trace UIs. */\nexport type AgentStep =\n  | { kind: \"tool_call\"; tool: string; args: unknown }\n  | { kind: \"tool_result\"; tool: string; result: string }\n  | { kind: \"answer\"; text: string };\n\nfunction formatAgentToolsPrompt(tools: AgentTool[]): string {\n  const defs = tools\n    .map(\n      (t) =>\n        `## ${t.name}\\nDescription: ${t.description}\\nParameters: ${JSON.stringify(\n          t.parameters ?? { type: \"object\", properties: {} },\n        )}`,\n    )\n    .join(\"\\n\\n\");\n  return `You are a helpful assistant with access to tools.\n\n# Tools\n\n${defs}\n\n## How to call a tool\n\nReply with ONLY:\n<tool_call>\n{\"name\": \"tool_name\", \"arguments\": {\"param\": \"value\"}}\n</tool_call>\n\nWhen you have the final answer, reply normally with no tool_call.`;\n}\n\nfunction parseAgentToolCall(text: string): { name: string; args: unknown } | null {\n  const tagged = text.match(/<tool_call>\\s*([\\s\\S]*?)\\s*<\\/tool_call>/);\n  const json = tagged ? tagged[1] : (text.match(/\\{\\s*\"name\"\\s*:[\\s\\S]*\\}/)?.[0] ?? null);\n  if (!json) {\n    return null;\n  }\n  try {\n    const parsed = JSON.parse(json) as {\n      name?: string;\n      arguments?: unknown;\n      args?: unknown;\n    };\n    if (typeof parsed.name === \"string\") {\n      return { name: parsed.name, args: parsed.arguments ?? parsed.args ?? {} };\n    }\n  } catch {\n    // not a valid tool call\n  }\n  return null;\n}\n\nexport class WebGPUEngine {\n  // ── Base text-model resources ──\n  // These are null on a TTS-only engine (created via `create({ repo: <tts> })`),\n  // which skips GPU init + the single-graph base load and lazily builds the\n  // dedicated native TTS engine on the first speak(). The private getters below\n  // throw a clear error if a base-model op is attempted on such an engine.\n  private _ctx: GPUContext | null;\n  private _executor: Executor | null;\n  private _tokenizer: Tokenizer | null;\n  private get ctx(): GPUContext {\n    if (!this._ctx) throw new Error(WebGPUEngine.TTS_LITE_ERROR);\n    return this._ctx;\n  }\n  private get executor(): Executor {\n    if (!this._executor) throw new Error(WebGPUEngine.TTS_LITE_ERROR);\n    return this._executor;\n  }\n  private get tokenizer(): Tokenizer {\n    if (!this._tokenizer) throw new Error(WebGPUEngine.TTS_LITE_ERROR);\n    return this._tokenizer;\n  }\n  private static readonly TTS_LITE_ERROR =\n    \"This engine was created for a TTS checkpoint; base-model ops (generate/embed/describeImage) are unavailable. Use speak() / encodeSpeaker().\";\n  private _destroyed = false;\n  private _isEmbedding: boolean;\n  /** HF architecture string (e.g. \"Gemma3TextModel\", \"Qwen3ForCausalLM\"). */\n  private _architecture: string;\n  /** Vision encoder (built only when enableVision and the model is vision-capable). */\n  private visionExecutor: VisionExecutor | null;\n  /** Raw vision_config (for host preprocessing of grids). */\n  private visionConfig: Record<string, unknown> | null;\n  /** Raw pos_embed.weight table for bilinear interpolation. */\n  private visionPosEmbedTable: Float32Array | null;\n  /** True when the LM graph was built with the multimodal (M-RoPE + splice) path. */\n  private _multimodalGraph: boolean;\n  /** Raw config.json (for M-RoPE params: mrope_section, rope_theta, partial factor). */\n  private rawConfig: Record<string, unknown> | null;\n  /** Effective max sequence length (cos/sin table coverage). */\n  private maxSeqLen: number;\n  /** Original create() options (used to lazily spin up the Kani-TTS engine for speak()). */\n  private _createOptions: WebGPUEngineOptions;\n  /** Lazily-created Kani-TTS engine (codec-LM + NanoCodec) backing speak(). */\n  private _kaniTTS: KaniTTS | null = null;\n  /** Lazily-created OuteTTS engine (Qwen3-0.6B codec-LM + DAC.speech) backing speak(). */\n  private _outeTTS: OuteTTS | null = null;\n  /** The preset voice the lazily-built OuteTTS engine was created with. */\n  private _outeVoice: string | null = null;\n  /** Source of the runtime LoRA adapter currently applied on the static base, if any. */\n  private _currentAdapter: string | null = null;\n  /** Named adapters registered for per-lane batched decode (GERBIL_BATCH_LORA). */\n  private _batchAdapters = new Map<string, { id: number; deltas: LoRADelta[] }>();\n  /** Lazily-created Parler-TTS engine (Flan-T5 encoder + decoder LM + dac_44khz). */\n  private _parlerTTS: ParlerTTS | null = null;\n  /** Lazily-created Gepard-1.0 engine (Qwen3.5 multihead codec-LM + NanoCodec 21.5 fps). */\n  private _gepardTTS: GepardTTS | null = null;\n  /** Lazily-created continuous-batching scheduler (GERBIL_BATCH, Phase 5). */\n  private _scheduler: RequestScheduler | null = null;\n\n  /**\n   * WebKit group-size probe state. When true, a candidate group size is being\n   * tried this page-load and must be promoted (or capped) after the FIRST\n   * successful forward produces non-corrupt logits. Goes false once handled so\n   * promotion runs at most once per session. Always false on Dawn/node.\n   */\n  private _groupProbePending = false;\n\n  /**\n   * Tail of the per-engine generation queue. The shared {@link Executor} owns a\n   * single set of readback/staging buffers (logits + argmax) and one KV cache, so\n   * two overlapping generate()/embed()/describeImage() calls on the SAME engine\n   * instance would call `mapAsync` on the logits/argmax readback while a prior map\n   * is still pending (\"Buffer already has an outstanding map pending\") and corrupt\n   * each other's KV state. The browser SDK shares one engine per model across\n   * components (see SHARED_ENGINES in browser/use-engine.ts), so this overlap is\n   * real — e.g. a playground's chat handle and its example-generator on the same\n   * instance. We serialize whole generations through this promise chain: a second\n   * caller awaits the first. Per-engine and only engaged on genuine overlap, so it\n   * adds no per-token overhead to the decode hot path.\n   */\n  private _genQueueTail: Promise<void> = Promise.resolve();\n\n  /** Model capabilities (text, vision, moe). */\n  readonly capabilities: ModelCapabilities;\n  /** Model architecture config (null on a TTS-only lite engine). */\n  private readonly _config: ModelArchConfig | null;\n  /** Model architecture config. Throws on a TTS-only lite engine. */\n  get config(): ModelArchConfig {\n    if (!this._config) throw new Error(WebGPUEngine.TTS_LITE_ERROR);\n    return this._config;\n  }\n\n  private constructor(\n    ctx: GPUContext | null,\n    executor: Executor | null,\n    tokenizer: Tokenizer | null,\n    graph: ModelGraph | null,\n    opts: {\n      multimodalGraph: boolean;\n      rawConfig: Record<string, unknown>;\n      maxSeqLen: number;\n      createOptions: WebGPUEngineOptions;\n      /** True if a WebKit group-size candidate is being probed this session. */\n      groupProbePending?: boolean;\n    },\n    vision?: {\n      executor: VisionExecutor;\n      config: Record<string, unknown>;\n      posEmbedTable: Float32Array;\n    } | null,\n  ) {\n    this._ctx = ctx;\n    this._executor = executor;\n    this._tokenizer = tokenizer;\n    // A TTS-only engine has no base text graph; capabilities/config/architecture\n    // are placeholders (speak() routes purely off the repo). Everything else\n    // reads through the private getters, which throw if a base op is attempted.\n    this.capabilities = graph?.capabilities ?? { text: true, vision: false, moe: false };\n    this._config = graph?.config ?? null;\n    this._isEmbedding = graph?.outputs.includes(\"embedding\") ?? false;\n    this._architecture = graph?.architecture ?? \"\";\n    this.visionExecutor = vision?.executor ?? null;\n    this.visionConfig = vision?.config ?? null;\n    this.visionPosEmbedTable = vision?.posEmbedTable ?? null;\n    this._multimodalGraph = opts.multimodalGraph;\n    this.rawConfig = opts.rawConfig;\n    this.maxSeqLen = opts.maxSeqLen;\n    this._createOptions = opts.createOptions;\n    this._groupProbePending = opts.groupProbePending ?? false;\n  }\n\n  /** True if this engine has a vision encoder built (use encodeImage()). */\n  get hasVision(): boolean {\n    return this.visionExecutor !== null;\n  }\n\n  /** Per-opType decode GPU-time breakdown (only populated under GERBIL_PROFILE). */\n  getDecodeProfile(): Array<{ opType: string; ns: number; count: number }> {\n    return this.executor.getProfile();\n  }\n\n  /** Clear accumulated decode profiler data (e.g. to drop warm-up tokens). */\n  resetDecodeProfile(): void {\n    this.executor.resetProfile();\n  }\n\n  /** Profile ONE real decode step (the pipelined-greedy kernels). Token-independent\n   * timing — pass any valid id. Only meaningful under GERBIL_PROFILE. */\n  async profileDecodeStep(tokenId: number): Promise<void> {\n    await this.executor.profileDecodeStep(tokenId);\n  }\n\n  /** Decode dispatch count + the device's storage-buffer limit (which gates the\n   * INT4 projection fusions). Lets the iPad runner report whether fusions applied\n   * on-device or silently fell back (8 < 9 ⇒ more dispatches ⇒ more mobile drains). */\n  getDecodeStats(): { dispatches: number; maxStorageBuffers: number } {\n    return {\n      dispatches: this.executor.decodeDispatchCount,\n      maxStorageBuffers: this.executor.maxStorageBuffers,\n    };\n  }\n\n  // ── Runtime LoRA adapters (\"flavors\") ─────────────────────────────────\n  //\n  // Apply — or hot-swap — a LoRA adapter on the ALREADY-LOADED base at runtime,\n  // as low-rank correction kernels layered on the static (INT4 or float) base.\n  // The base weights are never re-quantized or re-uploaded: a tuned model = the\n  // resident base + a few-MB adapter. Different adapters swap on the same warm\n  // engine (the A/B factor buffers are rebuilt; the base is untouched), and\n  // `loadAdapter(null)` removes the correction entirely.\n\n  /**\n   * Load, swap, or remove a runtime LoRA adapter on the loaded base.\n   *\n   * @param source Adapter spec (`hf:owner/repo`, a URL, or `file:./dir` holding\n   *   `adapter_config.json` + `adapter_model.safetensors`), or `null` to drop it.\n   */\n  async loadAdapter(source: string | null): Promise<void> {\n    if (source === null) {\n      this.executor.clearRuntimeLoRA();\n      this._currentAdapter = null;\n      return;\n    }\n    const adapter = await fetchAdapter(source, {\n      hfToken: this._createOptions.hfToken,\n      revision: this._createOptions.revision,\n    });\n    if (!adapter) {\n      throw new Error(`No adapter found at ${source}.`);\n    }\n    const keyMapper = createKeyMapperForArch(this._architecture);\n    const deltas = buildLoRADeltas(adapter, keyMapper);\n    const { applied, skipped } = this.executor.applyRuntimeLoRA(deltas);\n    if (applied === 0) {\n      throw new Error(\n        `Adapter ${source} resolved to 0 applicable runtime targets on ${this._architecture} (skipped ${skipped}).`,\n      );\n    }\n    console.log(\n      `[engine] runtime adapter ${source}: applied ${applied} target(s)${skipped ? `, skipped ${skipped}` : \"\"}.`,\n    );\n    this._currentAdapter = source;\n  }\n\n  /** The runtime LoRA adapter currently applied on the base, or null. */\n  getAdapter(): string | null {\n    return this._currentAdapter;\n  }\n\n  /**\n   * Register a named LoRA adapter for per-lane batched decode (SGMV phase 1,\n   * `GERBIL_BATCH=N` + `GERBIL_BATCH_LORA=1`). The adapter's factors are\n   * fetched and packed into the executor's shared factor buffers ONCE; any\n   * number of batch lanes can then run it concurrently by passing its name in\n   * `generateBatch`'s per-lane `adapters` option. Registering the same name\n   * twice is a no-op.\n   */\n  async registerAdapter(name: string, source: string): Promise<void> {\n    this.checkDestroyed();\n    if (!this.executor.batchLoraEnabled) {\n      throw new Error(\n        \"registerAdapter requires GERBIL_BATCH=N (N >= 2) and GERBIL_BATCH_LORA=1 on the Dawn path\",\n      );\n    }\n    if (this._batchAdapters.has(name)) return;\n    const adapter = await fetchAdapter(source, {\n      hfToken: this._createOptions.hfToken,\n      revision: this._createOptions.revision,\n    });\n    if (!adapter) {\n      throw new Error(`No adapter found at ${source}.`);\n    }\n    const keyMapper = createKeyMapperForArch(this._architecture);\n    const deltas = buildLoRADeltas(adapter, keyMapper);\n    const id = this.executor.registerBatchAdapter(deltas);\n    this._batchAdapters.set(name, { id, deltas });\n    console.log(`[engine] batch adapter \"${name}\" (${source}) registered as id ${id}.`);\n  }\n\n  /** Names of the adapters registered for per-lane batched decode. */\n  getRegisteredAdapters(): string[] {\n    return [...this._batchAdapters.keys()];\n  }\n\n  /**\n   * Write a coarse crash-phase breadcrumb that survives a GPU-process kill / page\n   * reload. The iPad harness reads `localStorage[\"gerbil-crash-phase\"]` after a\n   * crash; without these, a describe-time crash only shows the last load phase\n   * (\"engine:ready\"). The describe path tags vit-encode / splice / text-decode so\n   * the next run shows WHERE it died, not just \"crashed after load\".\n   */\n  private setPhase(phase: string): void {\n    try {\n      if (typeof localStorage !== \"undefined\") {\n        localStorage.setItem(\"gerbil-crash-phase\", phase);\n      }\n    } catch {\n      // localStorage can throw in private mode / cross-origin; telemetry is best-effort.\n    }\n  }\n\n  /** True if this engine was loaded as an embedding model (use embed(), not generate()). */\n  get isEmbedding(): boolean {\n    return this._isEmbedding;\n  }\n\n  /**\n   * WebKit group-size probe promotion hook. Runs at most once per session, after\n   * the FIRST forward completes without the page dying. If the page had crashed\n   * at this group size, this code never runs and the localStorage breadcrumb\n   * (left by the resolver) caps the device on the next load — that is what makes\n   * the probe survive the crash class. Here we additionally handle the\n   * wrong-output class by inspecting the first forward's logits for corruption\n   * (NaN / Inf / all-zero / all-same), reusing the same signals as integrityCheck().\n   */\n  private maybePromoteGroupProbe(logits: Float32Array): void {\n    if (!this._groupProbePending) return;\n    this._groupProbePending = false; // promote at most once per session\n\n    // Sample a prefix (full-vocab scans are wasteful; corruption shows up early).\n    const n = Math.min(logits.length, 256);\n    let allZero = true;\n    let allSame = true;\n    let finite = true;\n    const first = logits[0];\n    for (let i = 0; i < n; i++) {\n      const v = logits[i];\n      if (!Number.isFinite(v)) {\n        finite = false;\n        break;\n      }\n      if (v !== 0) allZero = false;\n      if (v !== first) allSame = false;\n    }\n    const correct = finite && !allZero && !allSame;\n    promoteGroupProbe(correct);\n  }\n\n  /**\n   * Create and initialize a WebGPUEngine.\n   *\n   * Downloads the model from HuggingFace, compiles shaders, uploads weights.\n   */\n  static async create(options: WebGPUEngineOptions = {}): Promise<WebGPUEngine> {\n    // Fill in the default model for the requested capability when no repo is\n    // given, so callers can `create()` / `create({ embedding: true })` and it\n    // just works, while any explicit `repo` is still honored everywhere below.\n    options = { ...options, repo: resolveDefaultRepo(options) };\n\n    // ── TTS checkpoints: defer to the dedicated native engine ──\n    // Kani (conv-backbone codec-LM), OuteTTS (Qwen3 codec-LM), and Parler\n    // (encoder-decoder) are NOT loadable as a single text graph — feeding them\n    // through the base loader errors at load (e.g. Kani's \"No GPU buffer for\n    // tensor layers.0.conv.in_proj.weight\", Parler's \"not loadable via the\n    // single-graph generate() path\"). Return a lightweight TTS-only engine that\n    // stores these options and builds the correct native TTS engine\n    // (_kaniTTS / _outeTTS / _parlerTTS) lazily on the first speak(). This is the\n    // low-level path used by the browser `useTTS` hook.\n    if (isTTSRepo(options.repo)) {\n      return WebGPUEngine._buildTTSLite(options);\n    }\n\n    // Initialize WebGPU\n    const ctx = await initGPU();\n\n    // A failed build — most commonly an iOS OOM mid weight-upload — would\n    // otherwise leak the just-allocated GPUDevice and its ~400MB of buffers for\n    // the page's lifetime, and a retry would stack another ~400MB on top → a\n    // death spiral that crashes the tab (and reads as \"crashes even after a\n    // clean cache / reboot\"). Tear the device down on ANY failure.\n    try {\n      return await WebGPUEngine._buildFromContext(ctx, options);\n    } catch (err) {\n      try {\n        ctx.device.destroy();\n      } catch {\n        /* device may already be lost on a hard OOM */\n      }\n      throw err;\n    }\n  }\n\n  /**\n   * Build a lightweight TTS-only engine: no GPU device is initialized and no base\n   * text graph is loaded. It stores the create() options; the first `speak()`\n   * (or `encodeSpeaker()`) lazily builds the dedicated native TTS engine —\n   * {@link KaniTTS}, {@link OuteTTS}, or {@link ParlerTTS} — which brings its own\n   * GPU device and weights. Base-model ops throw {@link WebGPUEngine.TTS_LITE_ERROR}.\n   */\n  private static _buildTTSLite(options: WebGPUEngineOptions): WebGPUEngine {\n    return new WebGPUEngine(\n      null,\n      null,\n      null,\n      null,\n      {\n        multimodalGraph: false,\n        rawConfig: {},\n        // Sequence bound handed to the native TTS engine on first speak(); TTS\n        // clips are short, so a modest default is ample (callers can override).\n        maxSeqLen: options.maxSeqLen ?? 2048,\n        createOptions: options,\n      },\n      null,\n    );\n  }\n\n  private static async _buildFromContext(\n    ctx: Awaited<ReturnType<typeof initGPU>>,\n    options: WebGPUEngineOptions,\n  ): Promise<WebGPUEngine> {\n    // ── Detect environment ──\n    const isBrowser = typeof navigator !== \"undefined\" && typeof location !== \"undefined\";\n    const isSafari =\n      isBrowser && /Safari/.test(navigator.userAgent) && !/Chrome/.test(navigator.userAgent);\n\n    // ── URL parameter overrides (browser only) ──\n    const params = isBrowser ? new URLSearchParams(location.search) : null;\n    const forceKvF32 = params?.has(\"kvf32\");\n    const maxSeqOverride = params?.get(\"maxseq\");\n    const groupOverride = params?.get(\"group\");\n\n    // ── Determine KV mode ──\n    // Priority: explicit option > URL param > auto-detect\n    let kvMode: KvMode;\n    if (options.kvMode) {\n      kvMode = options.kvMode;\n    } else if (forceKvF32) {\n      kvMode = \"f32\";\n    } else if (isSafari && ctx.hasF16) {\n      // Safari/WebKit miscompiles native f16 types — use packed u32 + unpack2x16float\n      kvMode = \"packed-f16\";\n    } else if (ctx.hasF16) {\n      kvMode = \"native-f16\";\n    } else {\n      kvMode = \"f32\";\n    }\n\n    const kvDtype = kvMode === \"f32\" ? \"f32\" : \"f16\";\n\n    console.log(\n      `[engine] kvMode: ${kvMode}, kvDtype: ${kvDtype}, f16 supported: ${ctx.hasF16}, safari: ${isSafari}`,\n    );\n\n    // Crash-phase tracking for mobile debugging (survives tab crash)\n    const setPhase = (phase: string) => {\n      try {\n        if (typeof localStorage !== \"undefined\") localStorage.setItem(\"gerbil-crash-phase\", phase);\n      } catch {}\n    };\n\n    // Load model (config, tokenizer, weights). When vision is enabled, build the\n    // multimodal LM graph variant (M-RoPE + vision-embedding splice); text-only\n    // generation through it is numerically identical to the plain graph.\n    setPhase(\"engine:loading-model\");\n    // Mobile WebKit has a ~few-second GPU watchdog and a tight (~1.5-2GB) jetsam\n    // budget. The ViT Attention op is a single O(N²) dispatch over N patches, so a\n    // large N both over-reserves memory and risks a watchdog kill. The host caps\n    // the longest image side (≈448px → a few hundred patches), so reserve far\n    // fewer patches on WebKit. Desktop/Dawn keeps the roomy 4096 ceiling.\n    const maxVisionPatches = options.maxVisionPatches ?? (ctx.isWebKitWebGPU ? 1024 : 4096);\n    const multimodal = options.enableVision\n      ? { maxVisionTokens: Math.ceil(maxVisionPatches / 4) }\n      : undefined;\n    const { graph, tokenizer, weights, rawConfig, pleSource } = await loadModel({\n      ...options,\n      repo: resolveDefaultRepo(options), // resolves to a string (already filled above)\n      kvDtype,\n      multimodal,\n    });\n    setPhase(`engine:model-loaded:${weights.size}-weights`);\n\n    // ── Vision tower (opt-in) ──\n    // Build BEFORE the text executor's uploadWeights() consumes/deletes the\n    // weight entries. The vision tower needs its own f32 copy of visual.* and the\n    // raw pos_embed table for host bilinear interpolation.\n    let visionBundle: {\n      executor: VisionExecutor;\n      config: Record<string, unknown>;\n      posEmbedTable: Float32Array;\n    } | null = null;\n    // Gate on the checkpoint actually shipping a vision tower (vision_config +\n    // tower weights), not the text graph's capability flag (which describes\n    // the text graph only). Two families are supported:\n    //   • Qwen3.5 ViT  — \"visual.*\" weights, \"visual.pos_embed.weight\" table\n    //   • Gemma 4 ViT  — \"vision_tower.*\" weights, axial pos-embed table\n    //                    \"vision_tower.patch_embedder.position_embedding_table\"\n    const hasVisionConfig = rawConfig.vision_config != null;\n    const isGemma4Vision =\n      ((rawConfig.vision_config as Record<string, unknown> | undefined)?.model_type as string) ===\n        \"gemma4_vision\" || [...weights.keys()].some((k) => k.startsWith(\"vision_tower.\"));\n    const visPosKey = isGemma4Vision\n      ? \"vision_tower.patch_embedder.position_embedding_table\"\n      : \"visual.pos_embed.weight\";\n    const towerPrefix = isGemma4Vision ? \"vision_tower.\" : \"visual.\";\n    const hasVisualWeights = weights.keys().some((k) => k.startsWith(towerPrefix));\n    const wantVision = Boolean(options.enableVision && hasVisionConfig);\n    if (wantVision && hasVisualWeights) {\n      const visGraph = isGemma4Vision\n        ? generateGemma4VisionGraph(rawConfig)\n        : generateQwen3_5VisionGraph(rawConfig);\n      const maxPatches = maxVisionPatches;\n      // Drain the vision-tower tensors (+ projector / pos-embed) out of the\n      // WeightSource into a heap Map so the existing (synchronous, Map-based)\n      // vision-prep helpers and VisionExecutor.uploadWeights keep working\n      // unchanged. The tower is ~201 MB BF16 and only materialized when vision is\n      // explicitly enabled — bounded, and disjoint from the text weights the text\n      // executor streams below.\n      const visWeights = new Map<string, { data: ArrayBufferView; shape: number[] }>();\n      for (const k of weights.keys()) {\n        if (k.startsWith(towerPrefix) || k.startsWith(\"embed_vision.\") || k === visPosKey) {\n          const w = await weights.get(k);\n          if (w) visWeights.set(k, w);\n        }\n      }\n      // Patch Gemma 4 ClippableLinear clip scalars into the graph + drop them,\n      // and dequantize the MLX-int4 projector (no-op for BF16 checkpoints).\n      if (isGemma4Vision) {\n        patchGemma4VisionClips(visGraph, visWeights);\n        const gi = resolveGemma4VisionInfo(rawConfig);\n        const qc = (rawConfig.quantization_config ?? rawConfig.quantization) as\n          | Record<string, unknown>\n          | undefined;\n        const projGroupSize = (qc?.group_size as number) ?? 64;\n        dequantizeGemma4VisionProjection(visWeights, projGroupSize, gi.textHidden, gi.hiddenSize);\n        // Synthesize the multimodal-embedder norm gains (post-projection norm is\n        // no-scale → ones; soft_embedding_norm falls back to ones if a lean export\n        // omits it). The post-projection RMSNorm rescales the image soft tokens to\n        // the text-embedding magnitude so the LM actually attends to the image.\n        ensureGemma4VisionEmbedderNorms(visWeights, gi.hiddenSize, gi.textHidden);\n      }\n      const visExec = new VisionExecutorImpl(ctx, visGraph, maxPatches);\n      // Snapshot the pos-embed table before uploadWeights consumes it.\n      const posW = visWeights.get(visPosKey);\n      const posTable =\n        posW && posW.data instanceof Float32Array\n          ? new Float32Array(posW.data) // copy — weights map entry will be uploaded+freed\n          : posW\n            ? new Float32Array(\n                posW.data.buffer,\n                posW.data.byteOffset,\n                posW.data.byteLength / 4,\n              ).slice()\n            : null;\n      await visExec.uploadWeights(visWeights); // reads (does not delete) tower constants\n      visExec.initBindGroups();\n      if (posTable) {\n        visionBundle = {\n          executor: visExec,\n          config: (rawConfig.vision_config as Record<string, unknown>) ?? rawConfig,\n          posEmbedTable: posTable,\n        };\n      }\n    }\n\n    // ── Determine max sequence length ──\n    let maxSeqLen: number;\n    if (maxSeqOverride) {\n      maxSeqLen = Math.min(Number.parseInt(maxSeqOverride, 10), graph.config.context_length);\n    } else if (forceKvF32 && isSafari) {\n      // f32 KV on Safari = 2x memory; cap at 1024 to avoid OOM\n      maxSeqLen = Math.min(options.maxSeqLen ?? 1024, graph.config.context_length, 1024);\n    } else if (ctx.isWebKitWebGPU) {\n      // iOS/Safari: jetsam kills the web content process around 1.5-2GB.\n      // Default conservatively; honor explicit requests up to 2048.\n      maxSeqLen = Math.min(options.maxSeqLen ?? 512, graph.config.context_length, 2048);\n    } else {\n      maxSeqLen = Math.min(\n        options.maxSeqLen ?? graph.config.context_length,\n        graph.config.context_length,\n        4096,\n      );\n    }\n\n    console.log(`[engine] maxSeqLen: ${maxSeqLen}, architecture: ${graph.architecture}`);\n\n    // Log device limits for debugging Safari issues\n    if (ctx.isWebKitWebGPU) {\n      console.log(\n        `[engine] device limits: maxBufferSize=${ctx.limits.maxBufferSize}, ` +\n          `maxStorageBufferBindingSize=${ctx.limits.maxStorageBufferBindingSize}, ` +\n          `maxComputeWorkgroupStorageSize=${ctx.limits.maxComputeWorkgroupStorageSize}`,\n      );\n    }\n\n    // Create executor and upload weights inside error scopes — Safari reports\n    // failed allocations and invalid bind groups asynchronously, and without\n    // scopes the affected dispatches silently no-op.\n    setPhase(\"engine:allocating-buffers\");\n    ctx.device.pushErrorScope(\"out-of-memory\");\n    ctx.device.pushErrorScope(\"validation\");\n    // ── WebKit submit-granularity (group size) selection ──\n    // On WebKit, the executor groups `webkitGroupSize` dispatches per command\n    // buffer (one submit + GPU drain per group); larger groups amortize the\n    // round-trip and can be ~5x faster on iPad. The safe ceiling is\n    // device-dependent (some devices crash or produce garbage at high group),\n    // so we run a crash-surviving, self-calibrating probe that escalates UP from\n    // group=1 and persists the result per-device in localStorage. An explicit\n    // ?group=N overrides everything and skips the probe.\n    //   - groupOverride present → use it verbatim.\n    //   - WebKit, no override   → resolver picks the rung + records the breadcrumb.\n    //   - Dawn / node           → undefined (executor floors at 1; probe inert).\n    const groupOverrideNum = groupOverride ? Number.parseInt(groupOverride, 10) : undefined;\n    let webkitGroupSize: number | undefined;\n    // True only when we escalated to a new candidate this load and must promote\n    // (or cap) it after the first successful forward.\n    let probingGroup = false;\n    if (groupOverrideNum) {\n      webkitGroupSize = groupOverrideNum;\n      console.log(\n        `[engine] webkitGroupSize override: ${webkitGroupSize} dispatches/command buffer`,\n      );\n    } else if (ctx.isWebKitWebGPU) {\n      // The resolver picks the rung and, when escalating, writes `trying` to the\n      // localStorage breadcrumb BEFORE we touch the GPU. We then attempt a\n      // promotion after the first forward; promoteGroupProbe() is itself a no-op\n      // when `trying === null` (capped / at the floor), so it is always safe to\n      // flag this as pending — no need to re-derive whether we actually escalated.\n      //\n      // Phones are NO LONGER treated as a crash class. We thought iPhone batching\n      // crashed the GPU process — but that was memory pressure from the old broken\n      // cache (404 MB redownloaded into RAM every load + orphaned/corrupt entries\n      // filling the quota). With the IndexedDB cache that pressure is gone, and an\n      // iPhone (iOS 18.7) ran Qwen3.5-0.8B at group=4096 / 38.4 tok/s, coherent.\n      // So phones escalate like tablets; the crash-surviving breadcrumb still caps\n      // any genuinely memory-tight device down after a single bad load.\n      webkitGroupSize = resolveWebkitGroupSize({ isWebKit: true });\n      probingGroup = true;\n    }\n    const executor = new Executor(ctx, graph, { maxSeqLen, kvMode, webkitGroupSize });\n    // Gemma 4: register the CPU-resident PLE table (streamed per-step, never\n    // uploaded to a GPU buffer) before uploadWeights consumes the rest.\n    if (pleSource) {\n      executor.setPleSource(pleSource);\n    }\n    setPhase(\"engine:uploading-weights\");\n    await executor.uploadWeights(weights);\n    // Free the browser transform-staging cache now that every weight is GPU-\n    // resident. (The PLE table lives in a separate cache and is NOT evicted —\n    // it is lazily streamed on first forward.)\n    await weights.dispose?.();\n    setPhase(\"engine:compiling-shaders\");\n    executor.initBindGroups();\n    const validationError = await ctx.device.popErrorScope();\n    const oomError = await ctx.device.popErrorScope();\n    if (oomError || validationError) {\n      const detail = [\n        oomError ? `out-of-memory: ${oomError.message}` : null,\n        validationError ? `validation: ${validationError.message}` : null,\n      ]\n        .filter(Boolean)\n        .join(\"; \");\n      setPhase(`engine:gpu-error:${detail.slice(0, 120)}`);\n      throw new Error(\n        `GPU setup failed (${detail}). The model likely exceeds this device's memory budget — try a smaller maxSeqLen or a q4-quantized model.`,\n      );\n    }\n    setPhase(\"engine:ready\");\n\n    return new WebGPUEngine(\n      ctx,\n      executor,\n      tokenizer,\n      graph,\n      {\n        multimodalGraph: Boolean(multimodal),\n        rawConfig,\n        maxSeqLen,\n        createOptions: options,\n        groupProbePending: probingGroup,\n      },\n      visionBundle,\n    );\n  }\n\n  /**\n   * Encode an image (already preprocessed into patches) into merged\n   * image-embedding tokens of dim `out_hidden_size` (1024 for Qwen3.5).\n   *\n   * This is the VISION ENCODER ONLY — it returns the image tokens; it does not\n   * splice them into a text sequence or apply M-RoPE (that is the LM-side\n   * integration phase). Requires `enableVision: true` at create() on a\n   * vision-capable checkpoint.\n   *\n   * @param patches Flattened patches, row-major [numPatches, patch_dim].\n   *   patch_dim = in_channels * temporal_patch_size * patch_size^2 (1536 for Qwen3.5).\n   *   Patches must already be ordered in spatial_merge_size×spatial_merge_size\n   *   groups (as the HF image processor emits them).\n   * @param gridTHW The (temporal, height, width) patch-grid dims for the image.\n   *   numPatches must equal t*h*w.\n   */\n  async encodeImage(\n    patches: Float32Array,\n    gridTHW: [number, number, number],\n    onStage?: (stage: string, info?: { layer?: number; total?: number }) => void,\n  ): Promise<EncodeImageResult> {\n    this.checkDestroyed();\n    if (!this.visionExecutor || !this.visionConfig || !this.visionPosEmbedTable) {\n      throw new Error(\n        \"encodeImage() requires a vision encoder. Load with { enableVision: true } \" +\n          \"on a vision-capable checkpoint (e.g. Qwen/Qwen3.5-0.8B).\",\n      );\n    }\n    const vcfg = this.visionConfig;\n\n    // ── Gemma 4 ViT path ──\n    if (this.visionExecutor.gemma4) {\n      const info = resolveGemma4VisionInfo({ vision_config: vcfg });\n      const gridH = gridTHW[1];\n      const gridW = gridTHW[2];\n      // The pos-embed table is [2, posSize, hidden]; posSize = total/(2*hidden).\n      const posSize = Math.floor(this.visionPosEmbedTable.length / (2 * info.hiddenSize));\n      const host = buildGemma4VisionPositionTensors(\n        gridH,\n        gridW,\n        this.visionPosEmbedTable,\n        posSize,\n        {\n          hiddenSize: info.hiddenSize,\n          numHeads: info.numHeads,\n          headDim: info.headDim,\n          ropeTheta: info.ropeTheta,\n          poolingKernelSize: info.poolingKernelSize,\n        },\n      );\n      return this.visionExecutor.encodeGemma4(\n        {\n          patches,\n          posEmbeds: host.posEmbeds,\n          cos: host.cos,\n          sin: host.sin,\n          poolMatrix: host.poolMatrix,\n          numPatches: host.numPatches,\n          numPooled: host.numPooled,\n        },\n        onStage,\n      );\n    }\n\n    // ── Qwen3.5 ViT path ──\n    const hiddenSize = vcfg.hidden_size as number;\n    const numHeads = vcfg.num_heads as number;\n    const host = buildVisionPositionTensors(gridTHW, this.visionPosEmbedTable, {\n      hiddenSize,\n      numHeads,\n      numPositionEmbeddings: vcfg.num_position_embeddings as number,\n      spatialMergeSize: vcfg.spatial_merge_size as number,\n      ropeTheta: 10_000.0,\n    });\n    const numPatches = gridTHW[0] * gridTHW[1] * gridTHW[2];\n    return this.visionExecutor.encode(\n      {\n        patches,\n        posEmbeds: host.posEmbeds,\n        cos: host.cos,\n        sin: host.sin,\n        numPatches,\n      },\n      onStage,\n    );\n  }\n\n  // ── Multimodal M-RoPE helpers ────────────────────────────────────────────\n\n  /** Resolve M-RoPE params from rawConfig: rope_dim, theta, mrope_section. */\n  private mropeParams(): {\n    ropeDim: number;\n    theta: number;\n    section: [number, number, number];\n    imageTokenId: number;\n    mergeSize: number;\n  } {\n    const cfg = this.rawConfig ?? {};\n    const textCfg = (cfg.text_config as Record<string, unknown>) ?? cfg;\n    const rope = (textCfg.rope_parameters as Record<string, unknown>) ?? {};\n    const headDim = this.config.head_dim;\n    const partial = (rope.partial_rotary_factor as number) ?? 0.25;\n    const ropeDim = Math.floor(headDim * partial);\n    const theta = (rope.rope_theta as number) ?? this.config.rope_base ?? 10_000;\n    const section = ((rope.mrope_section as number[]) ?? [11, 11, 10]) as [number, number, number];\n    const visCfg = (cfg.vision_config as Record<string, unknown>) ?? {};\n    return {\n      ropeDim,\n      theta,\n      section,\n      imageTokenId: (cfg.image_token_id as number) ?? 248_056,\n      mergeSize: (visCfg.spatial_merge_size as number) ?? 2,\n    };\n  }\n\n  /**\n   * Write the M-RoPE cos/sin (token order) + image row-map for a prefill of\n   * `positionIds3` ([3, seq]). `rowMap[i]` = vision-buffer row for image tokens,\n   * -1 for text. Returns the logical position of the last token (for decode).\n   */\n  private writeMRoPEPrefill(positionIds3: Int32Array, seq: number, rowMap: Int32Array): number {\n    const { ropeDim, theta, section } = this.mropeParams();\n    const { cos, sin } = buildMRoPECosSin(positionIds3, seq, ropeDim, theta, section);\n    this.executor.writeInput(\"mrope_cos\", cos);\n    this.executor.writeInput(\"mrope_sin\", sin);\n    this.executor.writeInput(\"vision_row_map\", rowMap);\n    // last logical position = max over the 3 components of the final token.\n    const last = seq - 1;\n    return Math.max(positionIds3[last], positionIds3[seq + last], positionIds3[2 * seq + last]);\n  }\n\n  /**\n   * Write a single decode-step M-RoPE cos/sin row at table slot `seqPos` for a\n   * text token at logical position `logicalPos`, plus a -1 row-map entry.\n   */\n  private writeMRoPEDecodeStep(seqPos: number, logicalPos: number): void {\n    const { ropeDim, theta, section } = this.mropeParams();\n    // Text token: all 3 position components equal → 1D RoPE for this row.\n    const pid = new Int32Array(3);\n    pid[0] = logicalPos;\n    pid[1] = logicalPos;\n    pid[2] = logicalPos;\n    const { cos, sin } = buildMRoPECosSin(pid, 1, ropeDim, theta, section);\n    const rowBytes = ropeDim * 4;\n    this.executor.writeInputAt(\"mrope_cos\", cos, seqPos * rowBytes);\n    this.executor.writeInputAt(\"mrope_sin\", sin, seqPos * rowBytes);\n  }\n\n  /** Write linear-position M-RoPE inputs for a pure-text forward (no image). */\n  private writeMRoPELinearText(seq: number): void {\n    const { ropeDim, theta, section } = this.mropeParams();\n    // Positions 0..seq-1, all three components equal (1D RoPE).\n    const pid = new Int32Array(3 * seq);\n    for (let i = 0; i < seq; i++) {\n      pid[i] = i;\n      pid[seq + i] = i;\n      pid[2 * seq + i] = i;\n    }\n    const { cos, sin } = buildMRoPECosSin(pid, seq, ropeDim, theta, section);\n    this.executor.writeInput(\"mrope_cos\", cos);\n    this.executor.writeInput(\"mrope_sin\", sin);\n    const rowMap = new Int32Array(seq).fill(-1);\n    this.executor.writeInput(\"vision_row_map\", rowMap);\n  }\n\n  /**\n   * Generate text from a prompt.\n   *\n   * Serialized per engine: the shared executor's readback/staging buffers and KV\n   * cache cannot be shared by two overlapping generations, so a concurrent call on\n   * the same instance awaits this one (see {@link _genQueueTail}).\n   */\n  async generate(\n    prompt: string | ChatMessage[],\n    options: GenerateOptions = {},\n  ): Promise<GenerateResult> {\n    this.checkDestroyed();\n    const release = await this._acquireGenLock();\n    try {\n      return await this._generateLocked(prompt, options);\n    } finally {\n      release();\n    }\n  }\n\n  /**\n   * Batched lockstep generation (GERBIL_BATCH=N — Phase 2 of the\n   * continuous-batching campaign, Dawn only).\n   *\n   * Prefills each prompt sequentially into its batch slot (KV into the slot's\n   * block-table row, SSM/conv state adopted into the slot's state-pool rows),\n   * then decodes all N sequences in lockstep: one batched dispatch stream per\n   * step produces one token per row. Greedy-only in Phase 2 (per-row GPU\n   * argmax); rows that hit EOS/stop keep decoding in lockstep with their\n   * output discarded until every row has finished — rows never interact, so\n   * per-row streams are token-exact vs single-sequence generation.\n   *\n   * Requires exactly `executor.batchSize` prompts (fixed lockstep batch; the\n   * Phase 3 scheduler will lift this).\n   */\n  async generateBatch(\n    prompts: Array<string | ChatMessage[]>,\n    options: GenerateOptions & {\n      maxTokensPerRow?: number[];\n      /**\n       * Per-lane adapter names (SGMV phase 1 — requires GERBIL_BATCH_LORA=1\n       * and prior {@link registerAdapter} calls). `adapters[i]` selects the\n       * adapter lane i decodes with; null/undefined = bare base. Lanes with\n       * different adapters (or none) decode concurrently over the shared base.\n       */\n      adapters?: Array<string | null | undefined>;\n    } = {},\n  ): Promise<GenerateResult[]> {\n    this.checkDestroyed();\n    const B = this.executor.batchSize;\n    if (B <= 0) {\n      throw new Error(\"generateBatch requires GERBIL_BATCH=N (N >= 2) on the Dawn path\");\n    }\n    if (prompts.length !== B) {\n      throw new Error(\n        `generateBatch: got ${prompts.length} prompts for a fixed batch of ${B} ` +\n          `(set GERBIL_BATCH=${prompts.length})`,\n      );\n    }\n    if (this._multimodalGraph || this.executor.hasPleSource() || this.executor.hasRuntimeAdapter) {\n      throw new Error(\"generateBatch: multimodal, PLE, and runtime-LoRA models are Phase 2+ work\");\n    }\n    const { maxTokens = 512, stopSequences = [], sampling = {}, systemPrompt } = options;\n    if ((sampling.temperature ?? 0.7) >= 1e-6) {\n      throw new Error(\"generateBatch: Phase 2 is greedy-only — pass sampling: { temperature: 0 }\");\n    }\n\n    // ── Per-lane adapters (GERBIL_BATCH_LORA) ──\n    const laneAdapterNames = options.adapters;\n    if (laneAdapterNames && laneAdapterNames.length !== B) {\n      throw new Error(\n        `generateBatch: got ${laneAdapterNames.length} lane adapters for a batch of ${B}`,\n      );\n    }\n    const wantLora = laneAdapterNames?.some((a) => a != null) ?? false;\n    if (wantLora && !this.executor.batchLoraEnabled) {\n      throw new Error(\"generateBatch: per-lane adapters require GERBIL_BATCH_LORA=1\");\n    }\n    const laneAdapters: Array<{ id: number; deltas: LoRADelta[] } | null> = new Array(B).fill(null);\n    if (laneAdapterNames) {\n      for (let i = 0; i < B; i++) {\n        const name = laneAdapterNames[i];\n        if (name == null) continue;\n        const reg = this._batchAdapters.get(name);\n        if (!reg) {\n          throw new Error(\n            `generateBatch: adapter \"${name}\" not registered (call registerAdapter first)`,\n          );\n        }\n        laneAdapters[i] = reg;\n      }\n    }\n\n    const release = await this._acquireGenLock();\n    try {\n      const startTime = performance.now();\n      this.executor.resetBatch();\n\n      const eosId = this.tokenizer.config.eosTokenId;\n      const eotId = this.resolveEndOfTurnId();\n      const keepPlan = this.config.vocabKeepPlan;\n      const remapTok = keepPlan\n        ? (i: number): number => remapPrunedToken(i, keepPlan)\n        : (i: number): number => i;\n\n      const rows = prompts.map((_p, i) => ({\n        generatedIds: [] as number[],\n        text: \"\",\n        finished: false,\n        finishReason: \"max_tokens\" as GenerateResult[\"finishReason\"],\n        // Per-row token budget (serving benches assign each request its own).\n        limit: options.maxTokensPerRow?.[i] ?? maxTokens,\n      }));\n\n      // ── Sequential per-slot prefill (batched prefill is Phase 3) ──\n      const firstTokens = new Uint32Array(B);\n      for (let i = 0; i < B; i++) {\n        const p = prompts[i];\n        const messages: ChatMessage[] =\n          typeof p === \"string\"\n            ? [\n                ...(systemPrompt ? [{ role: \"system\" as const, content: systemPrompt }] : []),\n                { role: \"user\" as const, content: p },\n              ]\n            : systemPrompt\n              ? [{ role: \"system\" as const, content: systemPrompt }, ...p]\n              : p;\n        const inputIds = this.tokenizer.encodeChat(messages, { addGenerationPrompt: true });\n        // Lane adapter: prefill runs on the single-sequence path, so apply the\n        // lane's adapter as the proven runtime overlay for exactly this\n        // prompt's prefill (KV/SSM state and the first sampled token then\n        // match a solo run with that adapter loaded), then drop it.\n        const laneReg = laneAdapters[i];\n        if (laneReg) this.executor.applyRuntimeLoRA(laneReg.deltas);\n        try {\n          this.executor.beginBatchSlot(i);\n          const { logits } = await this.executor.forward(new Uint32Array(inputIds));\n          this.executor.finishBatchSlot(i);\n          firstTokens[i] = remapTok(sampleToken(logits, { temperature: 0 }));\n        } finally {\n          if (laneReg) this.executor.clearRuntimeLoRA();\n        }\n      }\n\n      // Per-lane adapter selection for the batched decode steps: the shared\n      // base GEMMs are untouched; lanes with an adapter get gathered low-rank\n      // corrections, bare-base lanes cost nothing extra.\n      if (this.executor.batchLoraEnabled) {\n        this.executor.setBatchLaneAdapters(laneAdapters.map((a) => (a ? a.id : -1)));\n      }\n\n      const consume = (row: (typeof rows)[number], tok: number): void => {\n        if (row.finished) return;\n        row.generatedIds.push(tok);\n        if ((eosId !== null && tok === eosId) || (eotId !== null && tok === eotId)) {\n          row.finished = true;\n          row.finishReason = \"eos\";\n          return;\n        }\n        row.text += this.tokenizer.decode([tok], true);\n        if (stopSequences.some((s) => row.text.includes(s))) {\n          for (const s of stopSequences) {\n            const idx = row.text.indexOf(s);\n            if (idx !== -1) row.text = row.text.slice(0, idx);\n          }\n          row.finished = true;\n          row.finishReason = \"stop_sequence\";\n          return;\n        }\n        if (row.generatedIds.length >= row.limit) row.finished = true;\n      };\n\n      const decodeStart = performance.now();\n      for (let i = 0; i < B; i++) consume(rows[i], firstTokens[i]);\n\n      // ── Lockstep batched decode, pipelined depth-2 like the single path ──\n      const depth = Executor.PIPELINE_DEPTH;\n      const lens = this.executor.batchSlotLengths;\n      let capacity = Number.POSITIVE_INFINITY;\n      for (let i = 0; i < B; i++) {\n        capacity = Math.min(capacity, this.maxSeqLen - lens[i] - 1);\n      }\n      const maxRowLimit = rows.reduce((a, r) => Math.max(a, r.limit), 0);\n      const stepsNeeded = Math.max(0, Math.min(maxRowLimit - 1, capacity));\n      let submitted = 0;\n      let consumedSteps = 0;\n      while (consumedSteps < stepsNeeded && !rows.every((r) => r.finished)) {\n        while (submitted < stepsNeeded && submitted < consumedSteps + depth) {\n          this.executor.submitBatchDecodeStep(\n            submitted === 0 ? firstTokens : null,\n            submitted % depth,\n          );\n          submitted++;\n        }\n        const toks = await this.executor.readBatchTokens(consumedSteps % depth);\n        consumedSteps++;\n        for (let i = 0; i < B; i++) consume(rows[i], toks[i]);\n      }\n\n      const decodeTime = performance.now() - decodeStart;\n      const totalTime = performance.now() - startTime;\n      return rows.map((r) => ({\n        text: r.text,\n        tokensGenerated: r.generatedIds.length,\n        tokensPerSecond: r.generatedIds.length / (totalTime / 1000),\n        totalTime,\n        decodeTime,\n        finishReason: r.finishReason,\n        tokenIds: [...r.generatedIds],\n      }));\n    } finally {\n      release();\n    }\n  }\n\n  /**\n   * Continuous-batching scheduler (GERBIL_BATCH=N — Phase 5 of the batching\n   * campaign, Dawn only). Returns the engine's request scheduler (created on\n   * first call): submit many requests with independent prompts/limits and the\n   * scheduler multiplexes them over the batch lanes with dynamic admission,\n   * immediate lane retirement, and chunked prefill. Greedy-only; per-request\n   * output is token-exact vs single-sequence generate().\n   */\n  createScheduler(): RequestScheduler {\n    this.checkDestroyed();\n    if (this._scheduler) return this._scheduler;\n    if (this.executor.batchSize <= 0) {\n      throw new Error(\"createScheduler requires GERBIL_BATCH=N (N >= 2) on the Dawn path\");\n    }\n    if (this._multimodalGraph || this.executor.hasPleSource() || this.executor.hasRuntimeAdapter) {\n      throw new Error(\"createScheduler: multimodal, PLE, and runtime-LoRA models are unsupported\");\n    }\n    const keepPlan = this.config.vocabKeepPlan;\n    const host: SchedulerHost = {\n      executor: this.executor,\n      encodeChat: (messages) => this.tokenizer.encodeChat(messages, { addGenerationPrompt: true }),\n      decodeToken: (id) => this.tokenizer.decode([id], true),\n      eosTokenId: this.tokenizer.config.eosTokenId,\n      eotTokenId: this.resolveEndOfTurnId(),\n      remapToken: keepPlan ? (i) => remapPrunedToken(i, keepPlan) : (i) => i,\n      resolveAdapter: (name) => {\n        const reg = this._batchAdapters.get(name);\n        if (!reg) {\n          throw new Error(\n            `scheduler: adapter \"${name}\" not registered (call registerAdapter first)`,\n          );\n        }\n        return reg;\n      },\n      acquireLock: () => this._acquireGenLock(),\n    };\n    this._scheduler = new RequestScheduler(host);\n    return this._scheduler;\n  }\n\n  /**\n   * Create a swarm — the MoTA mode-B primitive (docs/research/slm-swarms.md):\n   * named members (LoRA adapters on this engine's shared base, or null for the\n   * bare base), an optional plan that fans an input out into per-member\n   * sub-tasks, and an optional reduce that joins the results. `swarm.run()`\n   * decodes every sub-task CONCURRENTLY through the continuous-batching\n   * scheduler with per-request adapters — one batched pass, so the fan-out\n   * costs roughly the slowest member, not the sum.\n   *\n   * Every non-null member source is registered via {@link registerAdapter}\n   * (downloads in parallel; factors packed into shared GPU buffers once).\n   * Requires `GERBIL_BATCH=N` (N >= 2) + `GERBIL_BATCH_LORA=1` on the Dawn\n   * path; greedy-only like the scheduler underneath.\n   */\n  async createSwarm<R = Record<string, ServeResult>>(options: SwarmOptions<R>): Promise<Swarm<R>> {\n    this.checkDestroyed();\n    if (!this.executor.batchLoraEnabled) {\n      throw new Error(\n        \"createSwarm requires GERBIL_BATCH=N (N >= 2) and GERBIL_BATCH_LORA=1 on the Dawn path\",\n      );\n    }\n    const scheduler = this.createScheduler();\n    await Promise.all(\n      Object.entries(options.members).map(([name, source]) =>\n        source == null ? Promise.resolve() : this.registerAdapter(name, source),\n      ),\n    );\n    return new Swarm(scheduler, options);\n  }\n\n  /**\n   * Resolve the end-of-turn stop token id, or null if the model has none.\n   *\n   * Chat models like Gemma 4 end an assistant turn with a dedicated end-of-turn\n   * token (`<turn|>`, id 106) rather than the generic `<eos>`. Every text-decode\n   * loop must treat it as a stop token or the model runs on past its answer and\n   * repeats until it hits `maxTokens`. Resolved from the tokenizer's added tokens\n   * so it is null for models without a turn delimiter (Qwen, LFM2, …), leaving\n   * their decode behaviour unchanged.\n   */\n  private resolveEndOfTurnId(): number | null {\n    return this.tokenizer.tokenToId(\"<turn|>\") ?? this.tokenizer.tokenToId(\"<end_of_turn>\");\n  }\n\n  private async _generateLocked(\n    prompt: string | ChatMessage[],\n    options: GenerateOptions = {},\n  ): Promise<GenerateResult> {\n    this.checkDestroyed();\n\n    const { maxTokens = 512, stopSequences = [], sampling = {}, systemPrompt, onToken } = options;\n\n    // Reset executor for new generation\n    this.executor.reset();\n\n    // Tokenize\n    let inputIds: number[];\n    if (typeof prompt === \"string\") {\n      // Wrap in chat format\n      const messages: ChatMessage[] = [];\n      if (systemPrompt) {\n        messages.push({ role: \"system\", content: systemPrompt });\n      }\n      messages.push({ role: \"user\", content: prompt });\n      inputIds = this.tokenizer.encodeChat(messages, { addGenerationPrompt: true });\n    } else {\n      // Already chat messages\n      const messages = systemPrompt\n        ? [{ role: \"system\" as const, content: systemPrompt }, ...prompt]\n        : prompt;\n      inputIds = this.tokenizer.encodeChat(messages, { addGenerationPrompt: true });\n    }\n\n    const startTime = performance.now();\n    const isGreedy = (sampling.temperature ?? 0.7) < 1e-6;\n\n    // Multimodal graph, text-only prompt. Two flavors:\n    //  - Qwen3.5 (M-RoPE): feed linear-position cos/sin (reduces to 1D RoPE) + an\n    //    all-text row map; decode steps write one cos/sin row each.\n    //  - Gemma 4 (standard 1D RoPE, self-computed): no host cos/sin. Just clear the\n    //    vision row-map to all -1 so EmbedSplice passes the text embeds through.\n    const hasMRoPE = this._multimodalGraph && this.executor.hasBuffer(\"mrope_cos\");\n    if (this._multimodalGraph) {\n      if (hasMRoPE) {\n        this.writeMRoPELinearText(inputIds.length);\n      } else if (this.executor.hasBuffer(\"vision_row_map\")) {\n        this.executor.writeInput(\"vision_row_map\", new Int32Array(inputIds.length).fill(-1));\n      }\n    }\n\n    // Prefill: process all input tokens at once\n    let { logits } = await this.executor.forward(new Uint32Array(inputIds));\n\n    // WebKit group-size probe: the first forward survived (page didn't die), so\n    // promote the candidate group size if its output is non-corrupt, else cap.\n    // No-op on Dawn/node and after the first call. Done before decode so a wrong\n    // group is recorded even if the user stops the generation early.\n    this.maybePromoteGroupProbe(logits);\n\n    // Decode loop\n    const generatedIds: number[] = [];\n    let finishReason: GenerateResult[\"finishReason\"] = \"max_tokens\";\n    let generatedText = \"\";\n\n    const eosId = this.tokenizer.config.eosTokenId;\n    const eotId = this.resolveEndOfTurnId();\n    // Decode-only timing (excludes prefill) for live tok/s in the onToken meta.\n    const decodeStart = performance.now();\n\n    // Emit one token; returns true when generation should stop.\n    const consumeToken = (nextToken: number): boolean => {\n      generatedIds.push(nextToken);\n\n      // Check EOS / end-of-turn (Gemma ends turns on <turn|>, not <eos>)\n      if ((eosId !== null && nextToken === eosId) || (eotId !== null && nextToken === eotId)) {\n        finishReason = \"eos\";\n        return true;\n      }\n\n      // Decode token and check stop sequences\n      const tokenText = this.tokenizer.decode([nextToken], true);\n      generatedText += tokenText;\n      if (onToken) {\n        const elapsedMs = performance.now() - decodeStart;\n        const tokenIndex = generatedIds.length;\n        onToken(tokenText, {\n          tokenIndex,\n          elapsedMs,\n          tps: elapsedMs > 0 ? (tokenIndex / elapsedMs) * 1000 : 0,\n        });\n      }\n\n      if (stopSequences.some((s) => generatedText.includes(s))) {\n        // Trim the stop sequence from the end\n        for (const s of stopSequences) {\n          const idx = generatedText.indexOf(s);\n          if (idx !== -1) {\n            generatedText = generatedText.slice(0, idx);\n          }\n        }\n        finishReason = \"stop_sequence\";\n        return true;\n      }\n      return false;\n    };\n\n    // M-RoPE decode positions are logical (text continues after the image at the\n    // post-image cursor), so each step needs a fresh cos/sin row — incompatible\n    // with the GPU-chained pipelined path. mmDecode tracks the logical position.\n    // Gemma 4's multimodal graph uses self-computed sequential RoPE (no host cos/sin),\n    // so it is NOT an mmDecode case and can use the fast pipelined path.\n    const mmDecode = hasMRoPE;\n    let mmLogicalPos = inputIds.length; // text-only: logical == token index\n\n    // Gemma 4 streams a CPU-resident Per-Layer-Embedding table: each decode step\n    // must gather+upload the PLE rows for the ACTUAL next token id on the CPU\n    // (streamPleRows, inside forward()) before the forward runs. The GPU-chained\n    // pipelined path and forwardArgmax both pick the next token ON THE GPU and\n    // never call back into the CPU, so they leave `ple_embed_out` stale — the\n    // model produces one good token then collapses into repetition. Force such\n    // models onto the per-step forward() + CPU-argmax loop (proven exact), which\n    // streams PLE for every token. Costs the pipelined-decode speedup, but PLE\n    // already requires per-step CPU work so there is no faster correct path.\n    const streamsPle = this.executor.hasPleSource();\n\n    // Vocab-pruned head (Lever B): the head emits logits in the PRUNED index space.\n    // The argmax shader remaps its winning index back to the real token id on the\n    // GPU (ARGMAX Params prefix_len/tail_start), so forwardArgmax AND the GPU-chained\n    // pipelined path both yield real token ids — pruning works on every decode path,\n    // including the fast Dawn pipeline. Only host-side sampleToken (which runs over\n    // the pruned logits array) needs the CPU remap below.\n    const keepPlan = this.config.vocabKeepPlan;\n    const remapTok = keepPlan\n      ? (i: number): number => remapPrunedToken(i, keepPlan)\n      : (i: number): number => i;\n\n    if (isGreedy && !this.executor.needsMultiEncoder && !mmDecode && !streamsPle) {\n      // ── Pipelined greedy decode (Dawn) ──\n      // Overlay-safe: submitGreedyDecodeStep encodes activeDecodeEntries, so a\n      // runtime LoRA overlay (selectively-unfused base + spliced low-rank\n      // dispatches) runs inside the same pass and its correction lands before\n      // the argmax reads logits — identical entries to the forwardArgmax path.\n      // The argmax result is chained into input_ids on the GPU, so up to\n      // PIPELINE_DEPTH decode steps stay in flight while tokens are read\n      // back one step behind. Removes per-token CPU prep and map latency\n      // from the critical path.\n      const firstToken = remapTok(sampleToken(logits, sampling, [...inputIds, ...generatedIds]));\n      if (!consumeToken(firstToken)) {\n        const stepsNeeded = Math.min(maxTokens - 1, this.executor.decodeCapacityRemaining());\n        const windowK = this.executor.decodeWindowK;\n        if (windowK > 0) {\n          // ── K-step decode windows (GERBIL_DECODE_WINDOW, Dawn only) ──\n          // Each window records K full decode steps + argmax into ONE command\n          // buffer / ONE submit, with GPU-side token chaining between steps\n          // and a single 4·K-byte readback at window end. EOS is detected at\n          // window granularity, so up to K-1 tokens past EOS are computed and\n          // discarded here. Scaffold: windows run sequentially (no depth-2\n          // window pipelining yet). See\n          // docs/design/dispatch-collapse-campaign.md §3 Lever 1.\n          let produced = 0;\n          let stopped = false;\n          while (!stopped && produced < stepsNeeded) {\n            const steps = Math.min(windowK, stepsNeeded - produced);\n            this.executor.submitGreedyDecodeWindow(produced === 0 ? firstToken : null, steps, 0);\n            const tokens = await this.executor.readDecodeWindow(0, steps);\n            produced += steps;\n            for (const tok of tokens) {\n              if (consumeToken(tok)) {\n                stopped = true;\n                break;\n              }\n            }\n          }\n        } else {\n          const depth = Executor.PIPELINE_DEPTH;\n          let submitted = 0;\n          let consumed = 0;\n          while (consumed < stepsNeeded) {\n            while (submitted < stepsNeeded && submitted < consumed + depth) {\n              this.executor.submitGreedyDecodeStep(\n                submitted === 0 ? firstToken : null,\n                submitted % depth,\n              );\n              submitted++;\n            }\n            const tok = await this.executor.readDecodeToken(consumed % depth);\n            consumed++;\n            if (consumeToken(tok)) break;\n          }\n        }\n      }\n    } else {\n      // PLE-streaming models (Gemma 4) must run the CPU forward() path every step\n      // (forwardArgmax bypasses the per-token PLE gather). Sample greedily on the\n      // CPU from the forward() logits instead of the GPU argmax shader.\n      const useCpuForward = !isGreedy || streamsPle;\n      for (let step = 0; step < maxTokens; step++) {\n        let nextToken: number;\n        if (step === 0 || useCpuForward) {\n          nextToken = remapTok(sampleToken(logits, sampling, [...inputIds, ...generatedIds]));\n        } else {\n          // Greedy multimodal/desktop fallback: write this step's M-RoPE row\n          // (logical position) before the GPU forward reads it.\n          if (mmDecode) this.writeMRoPEDecodeStep(this.executor.currentSeqPos, mmLogicalPos);\n          // The argmax shader already remapped the pruned index → real token id.\n          nextToken = await this.executor.forwardArgmax(\n            new Uint32Array([generatedIds[generatedIds.length - 1]]),\n          );\n          mmLogicalPos++;\n        }\n\n        if (consumeToken(nextToken)) break;\n\n        if (useCpuForward) {\n          if (mmDecode) this.writeMRoPEDecodeStep(this.executor.currentSeqPos, mmLogicalPos);\n          const result = await this.executor.forward(new Uint32Array([nextToken]));\n          logits = result.logits;\n          mmLogicalPos++;\n        }\n      }\n    }\n\n    const totalTime = performance.now() - startTime;\n    const tokensGenerated = generatedIds.length;\n    const tokensPerSecond = tokensGenerated / (totalTime / 1000);\n\n    return {\n      text: generatedText,\n      tokensGenerated,\n      tokensPerSecond,\n      totalTime,\n      finishReason,\n      tokenIds: [...generatedIds],\n    };\n  }\n\n  /**\n   * Inline autocomplete: continue `prefix` with a brief, single-line continuation.\n   * Wraps `generate` with low-latency defaults (16 tokens, temp 0.3, a repetition\n   * penalty of 1.25, stop at the first newline) + a continuation system prompt,\n   * then cleans the output: strip after newline, dequote, drop an echoed prefix,\n   * drop a leading overlap with the typed tail, drop a leading word-run that\n   * restates an earlier typed phrase, truncate an internal phrase loop, cap the\n   * length, and add a smart leading space.\n   *\n   * The repetition penalty plus the overlap/restatement/loop cleanup keep small\n   * quantized models (e.g. Qwen3.5-0.8B) from echoing the typed text or looping\n   * on a phrase. Both are on by default; `repetitionPenalty` and `maxChars` are\n   * overridable.\n   *\n   * ```ts\n   * const suggestion = await engine.autocomplete(\"The quick brown fox\");\n   * // \" jumps over the lazy dog.\"\n   * ```\n   */\n  async autocomplete(\n    prefix: string,\n    opts: {\n      maxTokens?: number;\n      temperature?: number;\n      /** Repetition penalty applied during sampling (default 1.25; 1.0 disables). */\n      repetitionPenalty?: number;\n      stop?: string[];\n      singleLine?: boolean;\n      /** Character cap for the cleaned continuation (default 140). */\n      maxChars?: number;\n    } = {},\n  ): Promise<string> {\n    const out = await this.generate(prefix, {\n      systemPrompt: AUTOCOMPLETE_SYSTEM,\n      maxTokens: opts.maxTokens ?? 16,\n      sampling: {\n        temperature: opts.temperature ?? 0.3,\n        repetitionPenalty: opts.repetitionPenalty ?? 1.25,\n      },\n      stopSequences: opts.stop ?? [\"\\n\"],\n    });\n    return cleanSuggestion(out.text, prefix, {\n      singleLine: opts.singleLine ?? true,\n      maxChars: opts.maxChars,\n    });\n  }\n\n  /**\n   * Rewrite `text` in a target tone (e.g. \"professional\", \"friendly\", \"concise\",\n   * \"playful\", \"pirate\") or with free-form `instructions`. Returns only the\n   * rewritten text.\n   *\n   * ```ts\n   * await engine.rewrite(\"hey can u send the file\", { tone: \"professional\" });\n   * ```\n   */\n  async rewrite(\n    text: string,\n    opts: {\n      tone?: string;\n      instructions?: string;\n      maxTokens?: number;\n      temperature?: number;\n    } = {},\n  ): Promise<string> {\n    const system =\n      opts.instructions ??\n      `Rewrite the user's text in a ${opts.tone ?? \"professional\"} tone. ` +\n        \"Output ONLY the rewritten text — no preamble, no quotes, no commentary.\";\n    const out = await this.generate(text, {\n      systemPrompt: system,\n      maxTokens: opts.maxTokens ?? 256,\n      sampling: { temperature: opts.temperature ?? 0.7 },\n    });\n    return out.text.trim();\n  }\n\n  /**\n   * Agentic tool-calling loop: generate, parse a `<tool_call>`, run the matching\n   * tool's `execute`, feed the result back, and repeat up to `maxSteps` until the\n   * model answers without calling a tool. Returns the final text + a step trace.\n   *\n   * ```ts\n   * const { text, steps } = await engine.generateWithTools(\"Weather in Paris?\", {\n   *   tools: [weatherTool],\n   * });\n   * ```\n   */\n  async generateWithTools(\n    prompt: string | ChatMessage[],\n    opts: {\n      tools: AgentTool[];\n      maxSteps?: number;\n      onStep?: (step: AgentStep) => void;\n      maxTokens?: number;\n      sampling?: { temperature?: number; topK?: number; topP?: number };\n    },\n  ): Promise<{ text: string; steps: AgentStep[] }> {\n    const { tools, maxSteps = 5, onStep, maxTokens, sampling } = opts;\n    const systemPrompt = formatAgentToolsPrompt(tools);\n    const messages: ChatMessage[] =\n      typeof prompt === \"string\" ? [{ role: \"user\", content: prompt }] : [...prompt];\n    const steps: AgentStep[] = [];\n    let finalText = \"\";\n\n    for (let i = 0; i < maxSteps; i++) {\n      const result = await this.generate(messages, { systemPrompt, maxTokens, sampling });\n      const call = parseAgentToolCall(result.text);\n      if (!call) {\n        finalText = result.text;\n        const answer: AgentStep = { kind: \"answer\", text: result.text };\n        steps.push(answer);\n        onStep?.(answer);\n        break;\n      }\n\n      const callStep: AgentStep = { kind: \"tool_call\", tool: call.name, args: call.args };\n      steps.push(callStep);\n      onStep?.(callStep);\n\n      const tool = tools.find((t) => t.name === call.name);\n      let resultText: string;\n      if (tool) {\n        try {\n          resultText = String(await tool.execute(call.args));\n        } catch (e) {\n          resultText = `Error executing ${call.name}: ${e}`;\n        }\n      } else {\n        resultText = `Error: unknown tool \"${call.name}\"`;\n      }\n\n      const resultStep: AgentStep = { kind: \"tool_result\", tool: call.name, result: resultText };\n      steps.push(resultStep);\n      onStep?.(resultStep);\n\n      messages.push({ role: \"assistant\", content: result.text });\n      messages.push({ role: \"user\", content: `Tool ${call.name} returned:\\n${resultText}` });\n      finalText = resultText;\n    }\n\n    return { text: finalText, steps };\n  }\n\n  /**\n   * Generate a STRUCTURED object: generate text, extract the first JSON\n   * object/array, parse it, validate it, and RETRY until it is valid (on-device\n   * tokens are free, so re-rolling a malformed JSON is cheap).\n   *\n   * Extraction is tolerant: prose, markdown, and ```json code fences are\n   * stripped, then the outermost balanced `{...}` or `[...]` is matched and\n   * `JSON.parse`d. Validation is one of:\n   *  - a predicate `(o) => boolean` (return false to reject),\n   *  - a minimal JSON-schema-ish object with `required` (those keys must exist),\n   *  - nothing (only valid JSON is required).\n   *\n   * On each retry the prompt is nudged with a terse \"return ONLY valid JSON…\"\n   * instruction (including the required-key shape when known). Throws a clear\n   * error if it never validates within `maxRetries + 1` attempts.\n   *\n   * ```ts\n   * const { object } = await engine.generateObject(\n   *   'Extract {name, age} from: \"I am Sarah, 28\"',\n   *   { schema: { required: [\"name\", \"age\"] } },\n   * );\n   * // object === { name: \"Sarah\", age: 28 }\n   * ```\n   *\n   * @typeParam T Expected object type (not enforced at runtime — validate via schema).\n   */\n  async generateObject<T = unknown>(\n    prompt: string,\n    options: GenerateObjectOptions = {},\n  ): Promise<GenerateObjectResult<T>> {\n    this.checkDestroyed();\n    const { schema, maxRetries = 4, ...generateOpts } = options;\n\n    const validate = (value: unknown): boolean => {\n      if (typeof schema === \"function\") {\n        return (schema as (o: unknown) => boolean)(value);\n      }\n      if (schema && typeof schema === \"object\" && Array.isArray(schema.required)) {\n        if (value === null || typeof value !== \"object\") return false;\n        const obj = value as Record<string, unknown>;\n        return schema.required.every((key) => key in obj);\n      }\n      // No schema → any valid JSON value is acceptable.\n      return true;\n    };\n\n    // Build the retry nudge once: include the required-key shape when known so\n    // the model gets a concrete target to match.\n    const shape =\n      schema && typeof schema === \"object\" && Array.isArray(schema.required)\n        ? `{ ${schema.required.join(\", \")} }`\n        : \"the requested shape\";\n    const nudge = `\\n\\nReturn ONLY valid JSON matching ${shape}. No prose, no markdown, no code fences.`;\n\n    const attemptsMax = Math.max(0, maxRetries) + 1;\n    let lastText = \"\";\n    let lastError = \"\";\n    for (let attempt = 1; attempt <= attemptsMax; attempt++) {\n      const promptForAttempt = attempt === 1 ? prompt : prompt + nudge;\n      const result = await this.generate(promptForAttempt, generateOpts);\n      lastText = result.text;\n      const parsed = extractJson(result.text);\n      if (parsed === undefined) {\n        lastError = \"no JSON object/array found in output\";\n        continue;\n      }\n      let value: unknown;\n      try {\n        value = JSON.parse(parsed);\n      } catch (e) {\n        lastError = `JSON.parse failed: ${e instanceof Error ? e.message : String(e)}`;\n        continue;\n      }\n      if (!validate(value)) {\n        lastError = \"parsed JSON failed schema validation\";\n        continue;\n      }\n      return { object: value as T, text: result.text, attempts: attempt };\n    }\n\n    throw new Error(\n      `generateObject failed after ${attemptsMax} attempt(s): ${lastError}. ` +\n        `Last output: ${JSON.stringify(lastText.slice(0, 200))}`,\n    );\n  }\n\n  /**\n   * Text-to-speech: text → 22 kHz PCM via KaniTTS (LFM2-350M codec-LM + NVIDIA\n   * NeMo NanoCodec). Returns `{ pcm: Float32Array, sampleRate: 22050 }`.\n   *\n   * Runs the full pipeline: the codec-LM backbone autoregressively emits NanoCodec\n   * audio tokens (4 per frame), then the bit-exact NanoCodec decoder (FSQ + causal\n   * HiFi-GAN) turns the codes into PCM. The heavy lifting lives in {@link KaniTTS}\n   * (src/gpu/kani-tts.ts); this lazily constructs that engine on first use\n   * (downloading the NanoCodec codec checkpoint alongside the backbone).\n   *\n   * The default checkpoint is the Apache-2.0 `nineninesix/kani-tts-450m-0.2-ft`\n   * (plain `Lfm2ForCausalLM` backbone); the earlier `KaniTTS2ForCausalLM`\n   * checkpoints (multi-accent) also work. The codec-LM backbone defaults to INT4\n   * (`dtype: \"q4\"`) — pass `dtype: \"f32\"` for full precision.\n   */\n  async speak(\n    text: string,\n    options: {\n      languageTag?: string;\n      /** Kani accent tag, OR an OuteTTS preset voice name (e.g. \"en-female-1-neutral\"). */\n      voice?: string;\n      temperature?: number;\n      topP?: number;\n      topK?: number;\n      repetitionPenalty?: number;\n      maxFrames?: number;\n      /** Backbone precision for the lazily-built TTS engine (default \"q4\"). */\n      dtype?: KaniDType | OuteDType;\n      /** Override the OuteTTS folded DAC codec weights URL/path (else the mirror). */\n      codecWeightsPath?: string;\n      /**\n       * OuteTTS voice cloning: a pre-encoded speaker reference (transcript + per-word\n       * DAC c1/c2 codes + prosody features, the preset-JSON shape). When set, this voice\n       * is used in-context for this call instead of the preset. Build it from a raw\n       * reference clip via `OuteTTS.encodeSpeaker({ audio, sampleRate, transcript })`\n       * (DAC.speech encode + prosody), or pass a preset/pre-encoded speaker JSON.\n       */\n      cloneSpeaker?: OuteSpeaker;\n      /**\n       * Parler-TTS describe-your-voice hook: a natural-language voice description\n       * (\"a warm female voice, slightly fast, clean studio recording\"). Only used by\n       * the Parler checkpoint (DEFAULT_MODELS.ttsParler). When omitted, a neutral\n       * default description is used.\n       */\n      describeVoice?: string;\n      /**\n       * Stop-head sigmoid threshold (Gepard only, default 0.5). Higher = later stop.\n       */\n      stopThreshold?: number;\n      /**\n       * Streaming audio callback (Gepard only). Gepard is frame-synchronous, so\n       * audio is decoded and delivered chunk-by-chunk WHILE generation runs — the\n       * callback fires with each ~1.5 s PCM chunk (the final one has isFinal:true).\n       */\n      onChunk?: (chunk: {\n        pcm: Float32Array;\n        sampleRate: number;\n        frameIndex: number;\n        isFinal: boolean;\n      }) => void;\n    } = {},\n  ): Promise<{ pcm: Float32Array; sampleRate: number; frames: number; audioSeconds: number }> {\n    this.checkDestroyed();\n    const repo = this._createOptions.repo ?? \"\";\n\n    // ── Parler-TTS (selectable describe-your-voice engine): Flan-T5 + decoder + dac_44khz. ──\n    if (/parler/i.test(repo) || /parler/i.test(this._architecture)) {\n      return this.speakParler(text, options);\n    }\n\n    // ── OuteTTS (selectable many-voices engine): Qwen3-0.6B codec-LM + DAC.speech. ──\n    if (/outetts/i.test(repo) || /oute/i.test(this._architecture)) {\n      return this.speakOute(text, options);\n    }\n\n    // ── Gepard-1.0 (selectable streaming engine): Qwen3.5 multihead codec-LM + NanoCodec. ──\n    if (/gepard/i.test(repo) || /gepard/i.test(this._architecture)) {\n      return this.speakGepard(text, options);\n    }\n\n    // ── Kani-TTS (default). ──\n    // TTS capability is gated on the loaded repo being a Kani-TTS checkpoint (the\n    // architecture string alone no longer identifies it — the Apache default is a\n    // plain Lfm2ForCausalLM). Both KaniTTS2 and the plain-LFM2 family route here;\n    // KaniTTS.create() auto-detects the per-checkpoint position/RoPE scheme.\n    const isTTS =\n      this._architecture === \"KaniTTS2ForCausalLM\" ||\n      /kani-tts/i.test(repo) ||\n      /\\btts\\b/i.test(repo);\n    if (!isTTS) {\n      throw new Error(\n        `speak() requires a Kani-TTS or OuteTTS checkpoint (e.g. \"${DEFAULT_MODELS.tts}\" ` +\n          `or \"${DEFAULT_MODELS.ttsOute}\"), loaded engine is \"${this._architecture}\" (repo \"${repo}\").`,\n      );\n    }\n    // Lazily build the dual-graph Kani engine (codec-LM + NanoCodec). It loads its\n    // own weights (the backbone repo + the separate NanoCodec checkpoint), so it is\n    // independent of this engine's text executor.\n    if (!this._kaniTTS) {\n      this._kaniTTS = await KaniTTS.create({\n        repo: this._createOptions.repo,\n        revision: this._createOptions.revision,\n        hfToken: this._createOptions.hfToken,\n        cacheDir: this._createOptions.cacheDir,\n        maxSeqLen: this.maxSeqLen,\n        dtype: options.dtype as KaniDType | undefined,\n      });\n    }\n    return this._kaniTTS.speak(text, {\n      languageTag: options.voice ?? options.languageTag,\n      temperature: options.temperature,\n      topP: options.topP,\n      topK: options.topK,\n      repetitionPenalty: options.repetitionPenalty,\n      maxFrames: options.maxFrames,\n    });\n  }\n\n  /**\n   * Parler-TTS speak path: lazily build the Parler engine (Flan-T5 description encoder\n   * + cross-attending decoder LM + dac_44khz) and synthesize `text` in the voice given\n   * by `describeVoice`. Single-checkpoint (all weights, incl. the DAC, in one file).\n   */\n  private async speakParler(\n    text: string,\n    options: { describeVoice?: string; temperature?: number; topK?: number; maxFrames?: number },\n  ): Promise<{ pcm: Float32Array; sampleRate: number; frames: number; audioSeconds: number }> {\n    if (!this._parlerTTS) {\n      this._parlerTTS = await ParlerTTS.create({\n        repo: this._createOptions.repo,\n        revision: this._createOptions.revision,\n        hfToken: this._createOptions.hfToken,\n        cacheDir: this._createOptions.cacheDir,\n        maxSeqLen: this.maxSeqLen,\n      });\n    }\n    const speakOpts: ParlerSpeakOptions = {\n      describeVoice: options.describeVoice,\n      temperature: options.temperature,\n      topK: options.topK,\n      maxFrames: options.maxFrames,\n    };\n    return this._parlerTTS.speak(text, speakOpts);\n  }\n\n  /** A random natural-language voice description for Parler-TTS (host template bank). */\n  randomVoiceDescription(): string {\n    return ParlerTTS.randomVoiceDescription();\n  }\n\n  /**\n   * Gepard-1.0 speak path: lazily build the Gepard engine (stock Qwen3.5\n   * full-attention codec-LM + 32 host-side FSQ classifier heads + NanoCodec\n   * 21.5 fps decoder) and synthesize `text`. Frame-synchronous: when `onChunk`\n   * is set, audio streams out during generation. Single default voice for now\n   * (voice cloning via the ref-compressor is a documented follow-up).\n   */\n  private async speakGepard(\n    text: string,\n    options: {\n      temperature?: number;\n      topK?: number;\n      repetitionPenalty?: number;\n      maxFrames?: number;\n      stopThreshold?: number;\n      dtype?: KaniDType | OuteDType | GepardDType;\n      onChunk?: (chunk: {\n        pcm: Float32Array;\n        sampleRate: number;\n        frameIndex: number;\n        isFinal: boolean;\n      }) => void;\n    },\n  ): Promise<{ pcm: Float32Array; sampleRate: number; frames: number; audioSeconds: number }> {\n    if (!this._gepardTTS) {\n      this._gepardTTS = await GepardTTS.create({\n        repo: this._createOptions.repo,\n        revision: this._createOptions.revision,\n        hfToken: this._createOptions.hfToken,\n        cacheDir: this._createOptions.cacheDir,\n        maxSeqLen: this.maxSeqLen,\n        dtype: options.dtype as GepardDType | undefined,\n      });\n    }\n    return this._gepardTTS.speak(text, {\n      temperature: options.temperature,\n      topK: options.topK,\n      repetitionPenalty: options.repetitionPenalty,\n      maxFrames: options.maxFrames,\n      stopThreshold: options.stopThreshold,\n      onChunk: options.onChunk,\n    });\n  }\n\n  /**\n   * OuteTTS speak path: lazily build the OuteTTS engine (loading the chosen preset\n   * voice's speaker JSON + the folded DAC codec weights from the mirror), rebuilding\n   * if the requested voice changed. Mirrors the Kani public shape (temperature/topP/topK).\n   */\n  private async speakOute(\n    text: string,\n    options: {\n      voice?: string;\n      temperature?: number;\n      topP?: number;\n      topK?: number;\n      repetitionPenalty?: number;\n      dtype?: KaniDType | OuteDType;\n      codecWeightsPath?: string;\n      cloneSpeaker?: OuteSpeaker;\n    },\n  ): Promise<{ pcm: Float32Array; sampleRate: number; frames: number; audioSeconds: number }> {\n    await this.ensureOute(options);\n    return this._outeTTS!.speak(text, {\n      temperature: options.temperature,\n      topP: options.topP,\n      topK: options.topK,\n      repetitionPenalty: options.repetitionPenalty,\n      // Per-call cloned voice override (else the preset the engine was built with).\n      speaker: options.cloneSpeaker,\n    });\n  }\n\n  /**\n   * Lazily build (or reuse) the OuteTTS engine for the requested preset voice,\n   * rebuilding if the voice changed. Loads the preset speaker JSON + folded DAC\n   * codec weights from the mirror (or a locally-cached `{cacheDir}/{voice}.json`).\n   * Shared by {@link speakOute} and {@link encodeSpeaker}.\n   */\n  private async ensureOute(options: {\n    voice?: string;\n    dtype?: KaniDType | OuteDType;\n    codecWeightsPath?: string;\n  }): Promise<void> {\n    const voice = options.voice ?? OUTETTS_ASSETS.defaultVoice;\n    if (this._outeTTS && this._outeVoice !== voice) {\n      this._outeTTS.destroy();\n      this._outeTTS = null;\n    }\n    if (this._outeTTS) {\n      return;\n    }\n    // Prefer a locally-cached speaker JSON ({cacheDir}/{voice}.json), else the mirror.\n    const cacheDir = this._createOptions.cacheDir;\n    let speakerRef = voice;\n    if (cacheDir) {\n      const local = `${cacheDir.replace(/\\/$/, \"\")}/${voice}.json`;\n      const fs = await import(\"node:fs\").catch(() => null);\n      if (fs?.existsSync(local)) {\n        speakerRef = local;\n      }\n    }\n    const speaker = await loadOuteSpeaker(speakerRef, OUTETTS_ASSETS.speakersBaseURL);\n    this._outeTTS = await OuteTTS.create({\n      repo: this._createOptions.repo,\n      revision: this._createOptions.revision,\n      hfToken: this._createOptions.hfToken,\n      cacheDir: this._createOptions.cacheDir,\n      maxSeqLen: this.maxSeqLen,\n      dtype: (options.dtype as OuteDType | undefined) ?? \"q4\",\n      codecWeightsPath: options.codecWeightsPath ?? OUTETTS_ASSETS.codecWeightsURL,\n      speaker,\n    });\n    this._outeVoice = voice;\n  }\n\n  /**\n   * OuteTTS voice cloning: encode a raw reference clip (mono PCM + its transcript)\n   * into a reusable speaker (DAC.speech encode → per-word c1/c2 codes + prosody).\n   * Pass the result to `speak(text, { cloneSpeaker })` to synthesize new text in that\n   * voice. Requires an OuteTTS checkpoint (the folded codec carries the encoder).\n   */\n  async encodeSpeaker(ref: {\n    audio: Float32Array;\n    sampleRate: number;\n    transcript: string;\n  }): Promise<OuteSpeaker> {\n    this.checkDestroyed();\n    const repo = this._createOptions.repo ?? \"\";\n    if (!(/outetts/i.test(repo) || /oute/i.test(this._architecture))) {\n      throw new Error(\n        `encodeSpeaker() requires an OuteTTS checkpoint (e.g. \"${DEFAULT_MODELS.ttsOute}\"), ` +\n          `loaded engine is \"${this._architecture}\" (repo \"${repo}\").`,\n      );\n    }\n    await this.ensureOute({});\n    return this._outeTTS!.encodeSpeaker(ref);\n  }\n\n  /** Accent/voice tags the lazily-built TTS engine supports ([] = single voice). */\n  ttsVoices(): string[] {\n    if (/outetts/i.test(this._createOptions.repo ?? \"\") || /oute/i.test(this._architecture)) {\n      return [...OUTETTS_PRESET_VOICES];\n    }\n    return this._kaniTTS?.availableVoices() ?? [];\n  }\n\n  /**\n   * Describe an image: image-in → text-out. Runs the vision encoder, splices the\n   * merged image tokens into a text prompt, applies multimodal M-RoPE, and\n   * generates a description. Requires `enableVision: true` at create().\n   *\n   * Image input forms:\n   *  - `{ pixels, width, height }` — decoded RGB (HWC, 0..255), host-preprocessed\n   *    (smart-resize/normalize/patchify) to match the HF image processor.\n   *  - `{ patches, gridTHW }` — already-built [N,1536] patch tensor + grid (e.g.\n   *    HF-exact pixel_values from a reference; skips host preprocessing).\n   */\n  async describeImage(\n    image:\n      | { pixels: Float32Array | Uint8ClampedArray | Uint8Array; width: number; height: number }\n      | { patches: Float32Array; gridTHW: [number, number, number] },\n    prompt = \"Describe this image.\",\n    options: GenerateOptions & { imageProcessor?: ImageProcessorConfig } = {},\n  ): Promise<GenerateResult> {\n    this.checkDestroyed();\n    const release = await this._acquireGenLock();\n    try {\n      return await this._describeImageLocked(image, prompt, options);\n    } finally {\n      release();\n    }\n  }\n\n  private async _describeImageLocked(\n    image:\n      | { pixels: Float32Array | Uint8ClampedArray | Uint8Array; width: number; height: number }\n      | { patches: Float32Array; gridTHW: [number, number, number] },\n    prompt = \"Describe this image.\",\n    options: GenerateOptions & { imageProcessor?: ImageProcessorConfig } = {},\n  ): Promise<GenerateResult> {\n    this.checkDestroyed();\n    if (!this.visionExecutor) {\n      throw new Error(\n        \"describeImage() requires a vision encoder. Load with { enableVision: true } \" +\n          \"on a vision-capable checkpoint (Qwen3.5 or Gemma 4).\",\n      );\n    }\n\n    // ── Gemma 4 family path ──\n    // ViT encoder + projector are native + validated (cosine 1.0 vs reference).\n    // Gemma 4 uses STANDARD sequential 1D RoPE (positions 0..N-1 across text+image),\n    // computed inside each layer from the KV write-position — so, unlike Qwen3.5's\n    // M-RoPE path, there are no host cos/sin inputs. We splice the merged vision\n    // tokens into the `image_token_id` rows (EmbedSplice in the multimodal graph)\n    // and run a plain self-RoPE prefill + greedy/sampled decode.\n    if (this.visionExecutor.gemma4) {\n      if (!this._multimodalGraph) {\n        throw new Error(\n          \"Gemma 4 describeImage() requires the multimodal text graph. Load with \" +\n            \"{ enableVision: true } so generateGemma4Graph builds the vision_embeds + \" +\n            \"EmbedSplice variant.\",\n        );\n      }\n      let g4patches: Float32Array;\n      let g4gridHW: [number, number];\n      if (\"patches\" in image) {\n        g4patches = image.patches;\n        g4gridHW = [image.gridTHW[1], image.gridTHW[2]];\n      } else {\n        const pre = preprocessImageGemma4(image.pixels, image.width, image.height);\n        g4patches = pre.patches;\n        g4gridHW = pre.gridHW;\n      }\n      const gThw: [number, number, number] = [1, g4gridHW[0], g4gridHW[1]];\n      const numPatches = gThw[0] * gThw[1] * gThw[2];\n      this.setPhase(`describe:vit-encode:N=${numPatches}`);\n      const vision = await this.encodeImage(g4patches, gThw, (stage, info) => {\n        const suffix = info?.layer != null ? `:L${info.layer}` : \"\";\n        this.setPhase(`describe:${stage}${suffix}`);\n      });\n\n      // Build the Gemma 4 multimodal token sequence. The HF Gemma processor expands\n      // the `<|image|>` chat placeholder into: boi(`<|image>`) + image_token×N +\n      // eoi(`<image|>`), bracketed by the \"\\n\\n\" the image template adds, all inside\n      // the user turn. We mirror that exactly so token positions match HF.\n      const cfg = this.rawConfig ?? {};\n      const imageTokenId = (cfg.image_token_id as number) ?? 258_880;\n      const boiId = this.tokenizer.tokenToId(\"<|image>\") ?? (cfg.boi_token_id as number) ?? 255_999;\n      const eoiId = this.tokenizer.tokenToId(\"<image|>\") ?? (cfg.eoi_token_id as number) ?? 258_882;\n      const bos = this.tokenizer.config.bosToken\n        ? (this.tokenizer.tokenToId(this.tokenizer.config.bosToken) ?? [])\n        : [];\n      const bosIds = Array.isArray(bos) ? bos : [bos];\n      // User-turn prefix/suffix around the image+prompt, matching applyGemmaTurnTemplate\n      // (`<|turn>user\\n … <turn|>\\n<|turn>model\\n`). The image block is \"\\n\\n\" + boi +\n      // soft tokens + eoi + \"\\n\\n\", then the text prompt.\n      const turnUserOpen = this.tokenizer.encode(\"<|turn>user\\n\");\n      const imgLead = this.tokenizer.encode(\"\\n\\n\");\n      const imgTrail = this.tokenizer.encode(\"\\n\\n\");\n      const promptIds = this.tokenizer.encode(prompt);\n      const turnClose = this.tokenizer.encode(\"<turn|>\\n<|turn>model\\n\");\n      const imageRun: number[] = new Array(vision.rows).fill(imageTokenId);\n      const inputIds: number[] = [\n        ...bosIds,\n        ...turnUserOpen,\n        ...imgLead,\n        boiId,\n        ...imageRun,\n        eoiId,\n        ...imgTrail,\n        ...promptIds,\n        ...turnClose,\n      ];\n\n      return this.runMultimodalGemma4(vision.embeds, inputIds, imageTokenId, options);\n    }\n\n    if (!this._multimodalGraph) {\n      throw new Error(\n        \"describeImage() requires a multimodal engine. Load with { enableVision: true } \" +\n          \"on a vision-capable checkpoint (e.g. Qwen/Qwen3.5-0.8B).\",\n      );\n    }\n\n    // 1. Image → patches → merged vision tokens [N, 1024].\n    const procCfg = options.imageProcessor ?? QWEN3_5_IMAGE_PROCESSOR;\n    let patches: Float32Array;\n    let gridTHW: [number, number, number];\n    if (\"patches\" in image) {\n      patches = image.patches;\n      gridTHW = image.gridTHW;\n    } else {\n      const pre = preprocessImage(image.pixels, image.width, image.height, procCfg);\n      patches = pre.patches;\n      gridTHW = pre.gridTHW;\n    }\n    // Phase breadcrumbs survive a GPU-process kill so the harness can attribute a\n    // describe-time crash to vit-encode vs splice vs text-decode.\n    const numPatches = gridTHW[0] * gridTHW[1] * gridTHW[2];\n    this.setPhase(`describe:vit-encode:N=${numPatches}`);\n    const vision = await this.encodeImage(patches, gridTHW, (stage, info) => {\n      const suffix = info?.layer != null ? `:L${info.layer}` : \"\";\n      this.setPhase(`describe:${stage}${suffix}`);\n    });\n\n    // 2. Build the multimodal token sequence. Match the HF chat template:\n    //    <|im_start|>user\\n<vision_start><image_pad × N><vision_end>{prompt}\n    //    <|im_end|>\\n<|im_start|>assistant\\n<think>\\n\\n</think>\\n\\n\n    const visStart = this.tokenizer.tokenToId(\"<|vision_start|>\") ?? 248_053;\n    const visEnd = this.tokenizer.tokenToId(\"<|vision_end|>\") ?? 248_054;\n    const { imageTokenId } = this.mropeParams();\n    const numImageTokens = vision.rows;\n    const pre = this.tokenizer.encode(\"<|im_start|>user\\n\");\n    const post = this.tokenizer.encode(\n      `${prompt}<|im_end|>\\n<|im_start|>assistant\\n<think>\\n\\n</think>\\n\\n`,\n    );\n    const imageRun: number[] = new Array(numImageTokens).fill(imageTokenId);\n    const inputIds: number[] = [...pre, visStart, ...imageRun, visEnd, ...post];\n\n    return this.runMultimodal(vision.embeds, gridTHW, inputIds, options);\n  }\n\n  /**\n   * Prepare the multimodal prefill: upload vision embeds, build the image row-map\n   * and 3D M-RoPE cos/sin, reset state, and write all host inputs. Returns the\n   * input ids and the post-image logical cursor for decode. Does NOT run forward.\n   */\n  private prepareMultimodalPrefill(\n    visionEmbeds: Float32Array,\n    gridTHW: [number, number, number],\n    inputIds: number[],\n  ): { lastLogicalPos: number } {\n    const { imageTokenId, mergeSize } = this.mropeParams();\n    const seq = inputIds.length;\n    if (seq > this.maxSeqLen) {\n      throw new Error(\n        `describeImage: prompt+image is ${seq} tokens > maxSeqLen ${this.maxSeqLen}. ` +\n          \"Increase maxSeqLen or use a smaller image.\",\n      );\n    }\n    this.executor.reset();\n    this.executor.writeInput(\"vision_embeds\", visionEmbeds);\n\n    const rowMap = new Int32Array(seq).fill(-1);\n    let v = 0;\n    for (let i = 0; i < seq; i++) {\n      if (inputIds[i] === imageTokenId) rowMap[i] = v++;\n    }\n    const positionIds3 = buildMRoPEPositionIds(inputIds, [gridTHW], imageTokenId, mergeSize);\n    const lastLogicalPos = this.writeMRoPEPrefill(positionIds3, seq, rowMap);\n    return { lastLogicalPos };\n  }\n\n  /**\n   * Gemma 4 multimodal prefill + decode. Unlike Qwen3.5 (M-RoPE), Gemma 4 uses\n   * STANDARD sequential 1D RoPE computed inside each layer from the KV write\n   * position, so there are no host cos/sin inputs and decode positions are simply\n   * the running seqPos — identical to plain text generation. We only upload the\n   * merged vision embeds + an image-token row-map (EmbedSplice scatters them into\n   * the image_token rows) before the forward pass.\n   */\n  private async runMultimodalGemma4(\n    visionEmbeds: Float32Array,\n    inputIds: number[],\n    imageTokenId: number,\n    options: GenerateOptions,\n  ): Promise<GenerateResult> {\n    const seq = inputIds.length;\n    if (seq > this.maxSeqLen) {\n      throw new Error(\n        `describeImage: prompt+image is ${seq} tokens > maxSeqLen ${this.maxSeqLen}. ` +\n          \"Increase maxSeqLen or use a smaller image.\",\n      );\n    }\n    this.setPhase(`describe:splice:seq=${seq}`);\n    this.executor.reset();\n    this.executor.writeInput(\"vision_embeds\", visionEmbeds);\n    // row_map[i] = vision-buffer row for the i-th image token, -1 for text rows.\n    const rowMap = new Int32Array(seq).fill(-1);\n    let v = 0;\n    for (let i = 0; i < seq; i++) {\n      if (inputIds[i] === imageTokenId) rowMap[i] = v++;\n    }\n    this.executor.writeInput(\"vision_row_map\", rowMap);\n\n    this.setPhase(\"describe:text-decode\");\n    const { maxTokens = 512, stopSequences = [], sampling = {}, onToken } = options;\n    const startTime = performance.now();\n    const isGreedy = (sampling.temperature ?? 0.7) < 1e-6;\n\n    let { logits } = await this.executor.forward(new Uint32Array(inputIds));\n\n    const generatedIds: number[] = [];\n    let finishReason: GenerateResult[\"finishReason\"] = \"max_tokens\";\n    let generatedText = \"\";\n    const eosId = this.tokenizer.config.eosTokenId;\n    // Gemma model turns end with the end-of-turn token `<turn|>` (id 106), NOT the\n    // generic `<eos>` — treat it as a stop so chat decode doesn't run on into\n    // repetition past the answer.\n    const eotId = this.resolveEndOfTurnId();\n\n    const consumeToken = (nextToken: number): boolean => {\n      generatedIds.push(nextToken);\n      if ((eosId !== null && nextToken === eosId) || (eotId !== null && nextToken === eotId)) {\n        finishReason = \"eos\";\n        return true;\n      }\n      const tokenText = this.tokenizer.decode([nextToken], true);\n      generatedText += tokenText;\n      onToken?.(tokenText);\n      if (stopSequences.some((s) => generatedText.includes(s))) {\n        for (const s of stopSequences) {\n          const idx = generatedText.indexOf(s);\n          if (idx !== -1) generatedText = generatedText.slice(0, idx);\n        }\n        finishReason = \"stop_sequence\";\n        return true;\n      }\n      return false;\n    };\n\n    for (let step = 0; step < maxTokens; step++) {\n      let nextToken: number;\n      if (step === 0) {\n        nextToken = sampleToken(logits, sampling, [...inputIds, ...generatedIds]);\n      } else if (isGreedy) {\n        nextToken = await this.executor.forwardArgmax(\n          new Uint32Array([generatedIds[generatedIds.length - 1]]),\n        );\n      } else {\n        nextToken = sampleToken(logits, sampling, [...inputIds, ...generatedIds]);\n      }\n\n      if (consumeToken(nextToken)) break;\n\n      if (!isGreedy) {\n        const result = await this.executor.forward(new Uint32Array([nextToken]));\n        logits = result.logits;\n      }\n    }\n\n    const totalTime = performance.now() - startTime;\n    const tokensGenerated = generatedIds.length;\n    return {\n      text: generatedText,\n      tokensGenerated,\n      tokensPerSecond: tokensGenerated / (totalTime / 1000),\n      totalTime,\n      finishReason,\n      tokenIds: [...generatedIds],\n    };\n  }\n\n  /** Prepare + prefill + decode for a fully-specified multimodal token sequence. */\n  private async runMultimodal(\n    visionEmbeds: Float32Array,\n    gridTHW: [number, number, number],\n    inputIds: number[],\n    options: GenerateOptions,\n  ): Promise<GenerateResult> {\n    this.setPhase(`describe:splice:seq=${inputIds.length}`);\n    const { lastLogicalPos } = this.prepareMultimodalPrefill(visionEmbeds, gridTHW, inputIds);\n    this.setPhase(\"describe:text-decode\");\n    return this.generateFromPrepared(inputIds, lastLogicalPos + 1, options);\n  }\n\n  /**\n   * Debug: run ONLY the multimodal prefill for an explicit token sequence and\n   * return the spliced input embeddings [seq, hidden] + first-token logits. Lets\n   * tests compare the fused text+vision stream and M-RoPE numerically vs HF\n   * without the decode loop overwriting intermediate buffers.\n   */\n  async debugMultimodalPrefill(\n    patches: Float32Array,\n    gridTHW: [number, number, number],\n    inputIds: number[],\n  ): Promise<{ splicedEmbeds: Float32Array; logits: Float32Array; seq: number }> {\n    this.checkDestroyed();\n    if (!this._multimodalGraph || !this.visionExecutor) {\n      throw new Error(\"debugMultimodalPrefill requires a multimodal engine (enableVision: true).\");\n    }\n    const vision = await this.encodeImage(patches, gridTHW);\n    this.prepareMultimodalPrefill(vision.embeds, gridTHW, inputIds);\n    const { logits } = await this.executor.forward(new Uint32Array(inputIds));\n    const splicedEmbeds = await this.executor.debugReadBuffer(\n      \"embed_spliced\",\n      inputIds.length * this.config.hidden_size,\n    );\n    return { splicedEmbeds, logits, seq: inputIds.length };\n  }\n\n  /**\n   * Internal: run prefill (assumes M-RoPE/splice inputs already written) + decode,\n   * with decode logical positions starting at `decodeStartPos`. Used by\n   * describeImage so the post-image cursor is honored.\n   */\n  private async generateFromPrepared(\n    inputIds: number[],\n    decodeStartPos: number,\n    options: GenerateOptions,\n  ): Promise<GenerateResult> {\n    const { maxTokens = 512, stopSequences = [], sampling = {}, onToken } = options;\n    const startTime = performance.now();\n    const isGreedy = (sampling.temperature ?? 0.7) < 1e-6;\n\n    let { logits } = await this.executor.forward(new Uint32Array(inputIds));\n\n    const generatedIds: number[] = [];\n    let finishReason: GenerateResult[\"finishReason\"] = \"max_tokens\";\n    let generatedText = \"\";\n    const eosId = this.tokenizer.config.eosTokenId;\n    const eotId = this.resolveEndOfTurnId();\n    let mmLogicalPos = decodeStartPos;\n\n    const consumeToken = (nextToken: number): boolean => {\n      generatedIds.push(nextToken);\n      if ((eosId !== null && nextToken === eosId) || (eotId !== null && nextToken === eotId)) {\n        finishReason = \"eos\";\n        return true;\n      }\n      const tokenText = this.tokenizer.decode([nextToken], true);\n      generatedText += tokenText;\n      onToken?.(tokenText);\n      if (stopSequences.some((s) => generatedText.includes(s))) {\n        for (const s of stopSequences) {\n          const idx = generatedText.indexOf(s);\n          if (idx !== -1) generatedText = generatedText.slice(0, idx);\n        }\n        finishReason = \"stop_sequence\";\n        return true;\n      }\n      return false;\n    };\n\n    for (let step = 0; step < maxTokens; step++) {\n      let nextToken: number;\n      if (step === 0) {\n        nextToken = sampleToken(logits, sampling, [...inputIds, ...generatedIds]);\n      } else if (isGreedy) {\n        this.writeMRoPEDecodeStep(this.executor.currentSeqPos, mmLogicalPos);\n        nextToken = await this.executor.forwardArgmax(\n          new Uint32Array([generatedIds[generatedIds.length - 1]]),\n        );\n        mmLogicalPos++;\n      } else {\n        nextToken = sampleToken(logits, sampling, [...inputIds, ...generatedIds]);\n      }\n\n      if (consumeToken(nextToken)) break;\n\n      if (!isGreedy) {\n        this.writeMRoPEDecodeStep(this.executor.currentSeqPos, mmLogicalPos);\n        const result = await this.executor.forward(new Uint32Array([nextToken]));\n        logits = result.logits;\n        mmLogicalPos++;\n      }\n    }\n\n    const totalTime = performance.now() - startTime;\n    const tokensGenerated = generatedIds.length;\n    return {\n      text: generatedText,\n      tokensGenerated,\n      tokensPerSecond: tokensGenerated / (totalTime / 1000),\n      totalTime,\n      finishReason,\n      tokenIds: [...generatedIds],\n    };\n  }\n\n  /**\n   * Embed text into an L2-normalized vector. The pooling strategy depends on the\n   * model: Qwen3-Embedding uses last-token (EOS-position) pooling, while\n   * EmbeddingGemma (Gemma3 encoder) uses mean pooling over all tokens followed by\n   * a 2-layer Dense head. Requires an embedding model (loaded with\n   * { embedding: true }).\n   *\n   * The returned Float32Array has unit L2 norm, so cosine similarity reduces to a\n   * dot product. Length is the model's embedding dim (768 for EmbeddingGemma;\n   * config.hidden_size for Qwen3-Embedding).\n   *\n   * EmbeddingGemma is asymmetric — pass `{ taskType: \"query\" }` for search\n   * queries and `{ taskType: \"document\" }` for the corpus, or a raw\n   * `{ taskPrompt }` for other tasks (clustering/classification/STS).\n   */\n  async embed(text: string, options: EmbedOptions = {}): Promise<Float32Array> {\n    this.checkDestroyed();\n    const release = await this._acquireGenLock();\n    try {\n      return await this._embedLocked(text, options);\n    } finally {\n      release();\n    }\n  }\n\n  private async _embedLocked(text: string, options: EmbedOptions = {}): Promise<Float32Array> {\n    this.checkDestroyed();\n    if (!this._isEmbedding) {\n      throw new Error(\n        \"embed() requires an embedding model. Load with { embedding: true } \" +\n          \"(e.g. repo: 'Qwen/Qwen3-Embedding-0.6B' or \" +\n          \"'mlx-community/embeddinggemma-300m-4bit').\",\n      );\n    }\n\n    // Reset non-autoregressive state before each embed call.\n    this.executor.reset();\n\n    const { instruction, taskType, taskPrompt, maxTokens } = options;\n    const isGemmaEncoder =\n      this._architecture === \"Gemma3TextModel\" || this._architecture === \"Gemma3Model\";\n\n    if (isGemmaEncoder) {\n      // EmbeddingGemma: apply the task prefix, mean-pool over all tokens (no EOS\n      // pooling token appended — every token participates in the mean).\n      const prefix =\n        taskPrompt ?? EMBEDDING_GEMMA_PROMPTS[taskType ?? \"query\"] ?? EMBEDDING_GEMMA_PROMPTS.query;\n      const input = `${prefix}${text}`;\n\n      // The Gemma tokenizer prepends <bos> per its config (add_bos_token: true).\n      let ids = this.tokenizer.encode(input);\n\n      const cap = maxTokens ?? this.config.context_length;\n      if (ids.length > cap) {\n        ids = ids.slice(0, cap);\n      }\n\n      return this.executor.embed(new Uint32Array(ids));\n    }\n\n    // Qwen3-Embedding: optional instruction prefix + last-token (EOS) pooling.\n    const input = instruction ? `Instruct: ${instruction}\\nQuery:${text}` : text;\n\n    let ids = this.tokenizer.encode(input);\n\n    // Last-token pooling reads the final position; Qwen3-Embedding pools the\n    // hidden state at a trailing <|endoftext|> (id 151643 — the model config's\n    // eos_token_id), NOT the chat <|im_end|> token the tokenizer reports as\n    // eos_token. Resolve the literal token and append it if absent.\n    const padId = this.tokenizer.tokenToId(\"<|endoftext|>\") ?? this.tokenizer.config.eosTokenId;\n    if (padId !== null && ids[ids.length - 1] !== padId) {\n      ids.push(padId);\n    }\n\n    // Truncate to budget, keeping the trailing pooling token as the last token.\n    const cap = maxTokens ?? this.config.context_length;\n    if (ids.length > cap) {\n      const tail = padId !== null && ids[ids.length - 1] === padId ? [padId] : [];\n      ids = [...ids.slice(0, cap - tail.length), ...tail];\n    }\n\n    return this.executor.embed(new Uint32Array(ids));\n  }\n\n  /**\n   * Generate text as an async iterator (streaming).\n   *\n   * Uses the onToken callback from generate() to push tokens into a queue\n   * that the async generator yields from. The generator returns the full\n   * GenerateResult when generation completes.\n   *\n   * Usage:\n   *   const gen = engine.stream(\"Hello!\");\n   *   for await (const token of gen) {\n   *     process.stdout.write(token);\n   *   }\n   *   const result = gen.next(); // { done: true, value: GenerateResult }\n   */\n  async *stream(\n    prompt: string | ChatMessage[],\n    options: GenerateOptions = {},\n  ): AsyncGenerator<string, GenerateResult, undefined> {\n    this.checkDestroyed();\n\n    // Token queue: onToken callback pushes here, generator yields from here.\n    const tokenQueue: string[] = [];\n    let resolve: (() => void) | null = null;\n    let done = false;\n\n    // Notify the generator that a new token is available\n    const pushToken = (token: string): void => {\n      tokenQueue.push(token);\n      if (resolve) {\n        const r = resolve;\n        resolve = null;\n        r();\n      }\n    };\n\n    // Wait until at least one token is available or generation is done\n    const waitForToken = (): Promise<void> => {\n      if (tokenQueue.length > 0 || done) {\n        return Promise.resolve();\n      }\n      return new Promise<void>((r) => {\n        resolve = r;\n      });\n    };\n\n    // Start generation in the background with our onToken callback\n    const genPromise = this.generate(prompt, {\n      ...options,\n      onToken: (token) => {\n        options.onToken?.(token); // Forward to user's callback too\n        pushToken(token);\n      },\n    }).then((result) => {\n      done = true;\n      // Wake up the generator if it's waiting\n      if (resolve) {\n        const r = resolve;\n        resolve = null;\n        r();\n      }\n      return result;\n    });\n\n    // Yield tokens as they arrive\n    let yielded = 0;\n    while (true) {\n      await waitForToken();\n\n      // Yield all queued tokens\n      while (yielded < tokenQueue.length) {\n        yield tokenQueue[yielded++];\n      }\n\n      // If generation is done and we've yielded everything, return the result\n      if (done && yielded >= tokenQueue.length) {\n        break;\n      }\n    }\n\n    return await genPromise;\n  }\n\n  /**\n   * Debug: read back a named GPU buffer (weight or activation).\n   * Call after forward() to inspect intermediate values.\n   */\n  async debugReadBuffer(tensorName: string, maxElements?: number): Promise<Float32Array> {\n    return this.executor.debugReadBuffer(tensorName, maxElements);\n  }\n\n  /**\n   * Run GPU diagnostics (buffer integrity, compute, shared memory).\n   * Useful for isolating Safari/WebKit-specific WebGPU issues.\n   */\n  async diagnose(): Promise<GPUDiagnosticResult> {\n    return verifyGPU(this.ctx);\n  }\n\n  /**\n   * Run GPU diagnostics without loading a model.\n   * Quick way to check if WebGPU is working correctly on this device.\n   */\n  static async quickDiagnose(): Promise<GPUDiagnosticResult> {\n    const ctx = await initGPU();\n    const result = await verifyGPU(ctx);\n    clearPipelineCache(ctx.device);\n    ctx.device.destroy();\n    return result;\n  }\n\n  /**\n   * Run a raw forward pass (no tokenization/chat template).\n   * Returns logits for the last token.\n   */\n  async rawForward(inputIds: Uint32Array): Promise<{ logits: Float32Array }> {\n    return this.executor.forward(inputIds);\n  }\n\n  /**\n   * Reset executor state (SSM, positions, etc.)\n   */\n  resetState(): void {\n    this.executor.reset();\n  }\n\n  /**\n   * Encode text to token IDs (useful for debugging / token counting).\n   */\n  encode(text: string): number[] {\n    return this.tokenizer.encode(text);\n  }\n\n  /**\n   * Decode token IDs to text.\n   */\n  decode(ids: number[], skipSpecialTokens?: boolean): string {\n    return this.tokenizer.decode(ids, skipSpecialTokens);\n  }\n\n  /**\n   * Integrity check: reads back key weight tensors and runs a single forward pass,\n   * returning checksums for comparison against a known-good reference (Dawn/Node.js).\n   *\n   * Use this to isolate Safari/iPad corruption:\n   * - If weights mismatch → fetch/download pipeline is corrupt\n   * - If weights match but logits mismatch → kernel computation bug on Metal\n   *\n   * Resets executor state before and after (safe to call anytime).\n   */\n  async integrityCheck(): Promise<IntegrityCheckResult> {\n    this.checkDestroyed();\n    const checks: IntegrityCheckEntry[] = [];\n    const log = (label: string, data: Float32Array | Uint32Array) => {\n      const f = data instanceof Float32Array ? data : new Float32Array(data.buffer);\n      let sum = 0;\n      let maxVal = -Infinity;\n      let argmax = 0;\n      for (let i = 0; i < f.length; i++) {\n        sum += f[i];\n        if (f[i] > maxVal) {\n          maxVal = f[i];\n          argmax = i;\n        }\n      }\n      const entry: IntegrityCheckEntry = {\n        label,\n        length: f.length,\n        sum: Math.round(sum * 1e6) / 1e6,\n        first4: [\n          Math.round(f[0] * 1e6) / 1e6,\n          Math.round(f[1] * 1e6) / 1e6,\n          Math.round(f[2] * 1e6) / 1e6,\n          Math.round(f[3] * 1e6) / 1e6,\n        ],\n        argmax,\n        maxVal: Math.round(maxVal * 1e4) / 1e4,\n      };\n      checks.push(entry);\n      return entry;\n    };\n\n    // 0. Report Metal detection and adapter info\n    checks.push({\n      label: `webkit_detection: isWebKitWebGPU=${this.ctx.isWebKitWebGPU} needsMultiEncoder=${this.executor.needsMultiEncoder}`,\n      length: 0,\n      sum: 0,\n      first4: [0, 0, 0, 0],\n      argmax: 0,\n      maxVal: 0,\n      note: this.ctx.adapterDescription,\n    });\n\n    // 1. Read back weight tensors (no forward pass needed)\n    const weightNames = [\"norm.weight\", \"layers.0.input_layernorm.weight\"];\n\n    for (const name of weightNames) {\n      try {\n        const data = await this.debugReadBuffer(name);\n        log(name, data);\n      } catch {\n        checks.push({\n          label: name,\n          length: 0,\n          sum: NaN,\n          first4: [NaN, NaN, NaN, NaN],\n          argmax: -1,\n          maxVal: NaN,\n          error: \"buffer not found\",\n        });\n      }\n    }\n\n    // 2. Try to read INT4 packed weights (gate_proj AND embed_tokens)\n    for (const qName of [\n      \"layers.0.mlp.gate_proj.weight.q\",\n      \"embed_tokens.weight.q\",\n      \"embed_tokens.weight.scales\",\n      \"embed_tokens.weight.zeros\",\n    ]) {\n      try {\n        const q = await this.debugReadBuffer(qName, 16);\n        log(`${qName} (reinterpret)`, q);\n      } catch {\n        checks.push({\n          label: qName,\n          length: 0,\n          sum: NaN,\n          first4: [NaN, NaN, NaN, NaN],\n          argmax: -1,\n          maxVal: NaN,\n          error: \"not found\",\n        });\n      }\n    }\n\n    // 3a. Single-dispatch diagnostic: run ONLY EmbeddingInt4 with production bind group\n    this.executor.reset();\n    try {\n      const result = await this.executor.debugFirstDispatch(new Uint32Array([1]));\n      const entry = log(`single_dispatch(${result.opType})`, result.output);\n      entry.note = `dispatch=${result.dispatchSize.join(\",\")} node=${result.nodeId}`;\n    } catch (e: any) {\n      checks.push({\n        label: \"single_dispatch\",\n        length: 0,\n        sum: NaN,\n        first4: [NaN, NaN, NaN, NaN],\n        argmax: -1,\n        maxVal: NaN,\n        error: e.message,\n      });\n    }\n\n    // 3a2. Run entry[1] (RMSNorm) in isolation — embed_out still has data from debugFirstDispatch.\n    // If RMSNorm produces non-zero output, the bind group is correct and it's a barrier issue.\n    // If zeros, the bind group is pointing to the wrong buffer.\n    try {\n      const result = await this.executor.debugDispatchEntry(1, 1);\n      const entry = log(`isolated_entry1(${result.opType})`, result.output);\n      entry.note = `node=${result.nodeId}`;\n    } catch (e: any) {\n      checks.push({\n        label: \"isolated_entry1\",\n        length: 0,\n        sum: NaN,\n        first4: [NaN, NaN, NaN, NaN],\n        argmax: -1,\n        maxVal: NaN,\n        error: e.message,\n      });\n    }\n\n    // 3a3. Run entry[2] (MatVecInt4) in isolation — layer0_norm1_out has data from isolated_entry1.\n    // If non-zero: MatVecInt4 kernel works on this GPU.\n    // If zeros: the kernel itself is broken on this GPU (not a barrier issue).\n    try {\n      const result = await this.executor.debugDispatchEntry(2, 1);\n      const entry = log(`isolated_entry2(${result.opType})`, result.output);\n      entry.note = `node=${result.nodeId}`;\n    } catch (e: any) {\n      checks.push({\n        label: \"isolated_entry2\",\n        length: 0,\n        sum: NaN,\n        first4: [NaN, NaN, NaN, NaN],\n        argmax: -1,\n        maxVal: NaN,\n        error: e.message,\n      });\n    }\n\n    // 3b. Show JS-computed params for first 5 entries (no GPU dispatch needed)\n    this.executor.reset();\n    const jsParams = this.executor.debugComputeParams(1, 5);\n    for (const p of jsParams) {\n      checks.push({\n        label: `jsParams[${p.idx}] ${p.opType} dispatch=[${p.dispatchSize}] u32=[${p.paramsU32}]`,\n        length: p.paramsU32.length,\n        sum: p.paramsU32.reduce((a, b) => a + b, 0),\n        first4: p.paramsU32.slice(0, 4).map(Number),\n        argmax: 0,\n        maxVal: 0,\n      });\n    }\n\n    // 3c. Full forward pass with token 1 — compare logits\n    this.executor.reset();\n    try {\n      const { logits } = await this.executor.forward(new Uint32Array([1]));\n      log(\"logits(token=1)\", logits);\n\n      // Also read embed_out (embedding layer output)\n      try {\n        const embed = await this.debugReadBuffer(\"embed_out\", 16);\n        log(\"embed_out\", embed);\n      } catch {\n        // embed_out may not exist in all architectures\n      }\n\n      // 3c. Pipeline probe: read activations at ~8 points to find where data drops to zero\n      try {\n        const probes = await this.executor.debugPipelineProbe(1);\n        for (const p of probes) {\n          const paramsStr = p.uniformParams ? ` params=[${p.uniformParams.join(\",\")}]` : \"\";\n          const entry: IntegrityCheckEntry = {\n            label: `probe[${p.idx}] ${p.opType} → ${p.tensor}${paramsStr}`,\n            length: 16,\n            sum: p.sum,\n            first4: p.first4,\n            argmax: 0,\n            maxVal: Math.max(...p.first4.map(Math.abs)),\n          };\n          // Flag if sum is exactly 0 (likely dead activation)\n          if (p.sum === 0 && p.first4.every((v) => v === 0)) {\n            entry.match = \"FAIL\";\n            entry.note = \"all zeros — activation dead\";\n          }\n          checks.push(entry);\n        }\n      } catch (e: any) {\n        checks.push({\n          label: \"pipeline_probe\",\n          length: 0,\n          sum: NaN,\n          first4: [NaN, NaN, NaN, NaN],\n          argmax: -1,\n          maxVal: NaN,\n          error: e.message,\n        });\n      }\n    } catch (e: any) {\n      checks.push({\n        label: \"logits(token=1)\",\n        length: 0,\n        sum: NaN,\n        first4: [NaN, NaN, NaN, NaN],\n        argmax: -1,\n        maxVal: NaN,\n        error: e.message,\n      });\n    }\n\n    // Reset state so the engine is clean for actual use\n    this.executor.reset();\n\n    // Dawn reference values for Qwen3.5-0.8B Q4 (updated per current build).\n    // Weight checksums are stable across builds; logits may drift with code changes.\n    // Key diagnostic: if weights match but logits differ → kernel bug on Metal.\n    // If weights don't match → data transfer/download corruption on Safari.\n    const reference: Record<string, { sum: number; first4: number[] }> = {\n      \"norm.weight\": {\n        sum: 4412.571289,\n        first4: [0.222656, 3.75, 3.640625, 4.0625],\n      },\n      \"layers.0.input_layernorm.weight\": {\n        sum: 1267.312683,\n        first4: [2.0, 1.294922, 1.535156, 1.746094],\n      },\n      embed_out: {\n        sum: -0.067839,\n        first4: [0.010059, 0.0264, 0.001888, -0.030794],\n      },\n    };\n\n    // Compare weights against reference and flag obvious corruption\n    for (const check of checks) {\n      const ref = reference[check.label];\n      if (ref) {\n        const sumDelta = Math.abs(check.sum - ref.sum);\n        const sumMatch = sumDelta < Math.max(Math.abs(ref.sum) * 0.001, 1);\n        const first4Match = check.first4.every((v, i) => Math.abs(v - ref.first4[i]) < 0.01);\n        check.match = sumMatch && first4Match ? \"PASS\" : \"FAIL\";\n        check.refSum = ref.sum;\n      } else if (check.label === \"logits(token=1)\") {\n        // For logits: check for corruption signals (NaN, Inf, all-same, argmax in range)\n        const hasNaN = check.first4.some((v) => Number.isNaN(v));\n        const hasInf = check.first4.some((v) => !Number.isFinite(v));\n        const allSame = check.first4.every((v) => v === check.first4[0]);\n        const argmaxValid = check.argmax >= 0 && check.argmax < check.length;\n        check.match = !hasNaN && !hasInf && !allSame && argmaxValid ? \"PASS\" : \"FAIL\";\n      }\n    }\n\n    // Console output\n    console.log(\"\\n=== INTEGRITY CHECK ===\");\n    for (const c of checks) {\n      const status = c.error ? \"ERR\" : (c.match ?? \"---\");\n      console.log(\n        `[${status}] ${c.label}: sum=${c.sum}, first4=[${c.first4.join(\", \")}], ` +\n          `argmax=${c.argmax}, len=${c.length}` +\n          (c.refSum !== undefined ? ` (ref sum=${c.refSum})` : \"\") +\n          (c.error ? ` ERROR: ${c.error}` : \"\") +\n          (c.note ? ` | ${c.note}` : \"\"),\n      );\n    }\n    console.log(\"=== END INTEGRITY CHECK ===\\n\");\n\n    const allPass = checks.every((c) => c.match === \"PASS\" || c.match === undefined);\n    return { checks, allPass };\n  }\n\n  /**\n   * Destroy the engine and free all GPU resources.\n   */\n  destroy(): void {\n    if (this._destroyed) return;\n    this._destroyed = true;\n    // A TTS-only lite engine has no base executor/device (both null); only its\n    // lazily-built native TTS engines need tearing down.\n    this._executor?.destroy();\n    this._kaniTTS?.destroy();\n    this._outeTTS?.destroy();\n    this._parlerTTS?.destroy();\n    this._gepardTTS?.destroy();\n    this.visionExecutor?.destroy();\n    if (this._ctx) {\n      clearPipelineCache(this._ctx.device);\n      this._ctx.device.destroy();\n    }\n  }\n\n  private checkDestroyed(): void {\n    if (this._destroyed) {\n      throw new Error(\"WebGPUEngine has been destroyed\");\n    }\n  }\n\n  /**\n   * Acquire the per-engine generation lock (see {@link _genQueueTail}). Returns a\n   * release fn that MUST be called in a `finally` so the chain never wedges. This\n   * is a plain promise-chain mutex: each caller links onto the current tail and\n   * only proceeds once the previous holder releases.\n   *\n   * Non-reentrant: only the top-level generation entry points that directly drive\n   * the shared executor readback — generate(), embed(), describeImage() — acquire\n   * it. Helpers that delegate to generate() (autocomplete, rewrite,\n   * generateWithTools, generateObject, stream) inherit serialization through that\n   * single call and must NOT acquire it themselves, or they would self-deadlock.\n   */\n  private async _acquireGenLock(): Promise<() => void> {\n    let release!: () => void;\n    const next = new Promise<void>((resolve) => {\n      release = resolve;\n    });\n    const prev = this._genQueueTail;\n    this._genQueueTail = next;\n    await prev;\n    return release;\n  }\n}\n\nexport {\n  dequantizeGemma4VisionProjection,\n  dequantizeMLXProjection,\n  type Gemma4VisionGraphInfo,\n  generateGemma4VisionGraph,\n  patchGemma4VisionClips,\n  resolveGemma4VisionInfo,\n} from \"./architectures/gemma4_vision.js\";\n// Gepard-1.0 native streaming TTS (stock Qwen3.5 full-attention body + 32 parallel\n// FSQ classifier heads emitting a whole frame per forward + NanoCodec 21.5 fps\n// decoder). End-to-end: GepardTTS.create() then engine.speak(text, { model:\n// DEFAULT_MODELS.ttsGepard }) → 22 kHz PCM, streamed chunk-by-chunk via onChunk.\nexport {\n  buildGepardTextLayout,\n  type GepardConfig,\n  generateGepardBackboneGraph,\n  gepardChannelLevels,\n  gepardCodesToLatent,\n  parseGepardConfig,\n} from \"./architectures/gepard_tts.js\";\nexport type { GraphDType, KVDType, KvMode } from \"./architectures/index.js\";\n// Kani-TTS-2 native TTS (LFM2-350M codec-LM + NVIDIA NeMo NanoCodec). End-to-end:\n// KaniTTS.create() then engine.speak(text) → 22 kHz PCM. The NanoCodec decoder is\n// validated bit-exact; the codec-LM AR driver (frame positions + learnable RoPE)\n// lives in kani-tts.ts.\nexport {\n  audioTokensToCodes,\n  generateKaniTtsGraph,\n  generateNanoCodecDecoderGraph,\n  parseKaniConfig,\n} from \"./architectures/kani_tts.js\";\n// Moonshine native STT (encoder-decoder). End-to-end: MoonshineSTT.create() then\n// engine.transcribe(pcm16k). CrossAttention + conv-frontend kernels validated;\n// dual-graph executor wired (encoder once → frozen K/V → AR decode w/ cross-attn).\nexport {\n  generateMoonshineDecoderGraph,\n  generateMoonshineEncoderGraph,\n  MOONSHINE_REMAINING_WORK,\n  moonshineEncoderFrames,\n  parseMoonshineConfig,\n} from \"./architectures/moonshine.js\";\n// OuteTTS-1.0-0.6B native TTS (plain Qwen3-0.6B codec-LM + descript DAC.speech\n// decoder, 24 kHz / 2-codebook). FOUNDATION: OuteTTS.create({ codecWeightsPath })\n// then .speak(text) → 24 kHz PCM. The DAC decoder (RVQDequant + Conv/Snake stack)\n// is validated bit-exact vs the torch reference (scripts/engine/test-outetts-codec.mjs);\n// the AR backbone loop lives in outetts.ts. Cloning / int4 / engine.speak surface\n// are the next phase.\nexport {\n  audioTokensToDacCodes,\n  dacOutputLength,\n  generateDacSpeechDecoderGraph,\n  generateOuteTtsBackboneGraph,\n  parseOuteTtsConfig,\n} from \"./architectures/outetts.js\";\nexport { generateQwen3_5VisionGraph } from \"./architectures/qwen3_5_vision.js\";\nexport {\n  RequestScheduler,\n  type SchedulerStats,\n  type ServeOptions,\n  type ServeResult,\n} from \"./batch/scheduler.js\";\nexport {\n  Swarm,\n  type SwarmMemberResult,\n  type SwarmOptions,\n  type SwarmTask,\n  type SwarmTaskOptions,\n} from \"./batch/swarm.js\";\nexport type { GPUDiagnosticResult } from \"./device.js\";\nexport { initGPU } from \"./device.js\";\nexport { Executor } from \"./executor.js\";\nexport {\n  type GepardAudioChunk,\n  type GepardDType,\n  type GepardSpeakOptions,\n  type GepardSpeakResult,\n  GepardTTS,\n  type GepardTTSOptions,\n} from \"./gepard-tts.js\";\nexport type { ModelArchConfig, ModelCapabilities } from \"./ir.js\";\nexport {\n  type KaniDType,\n  KaniTTS,\n  type KaniTTSOptions,\n  type SpeakOptions,\n  type SpeakResult,\n} from \"./kani-tts.js\";\nexport {\n  type LoadedKaniTTS,\n  type LoadedMoonshine,\n  loadKaniTTS,\n  loadModel,\n  loadMoonshine,\n  quantizeKaniBackbone,\n} from \"./model-loader.js\";\nexport { MoonshineEncoderExecutor } from \"./moonshine-executor.js\";\nexport {\n  MoonshineSTT,\n  type MoonshineSTTOptions,\n  type TranscribeOptions,\n  type TranscribeResult,\n} from \"./moonshine-stt.js\";\nexport {\n  buildOutePromptString,\n  loadOuteSpeaker,\n  type OuteDType,\n  type OuteSpeaker,\n  type OuteSpeakerWord,\n  type OuteSpeakOptions,\n  type OuteSpeakResult,\n  OuteTTS,\n  type OuteTTSOptions,\n} from \"./outetts.js\";\nexport {\n  type ParlerSpeakOptions,\n  type ParlerSpeakResult,\n  ParlerTTS,\n  type ParlerTTSOptions,\n} from \"./parler-tts.js\";\nexport type { SamplingParams } from \"./sampler.js\";\n// Re-export useful types\nexport type { ChatMessage } from \"./tokenizer.js\";\nexport { Tokenizer } from \"./tokenizer.js\";\nexport type { VisionInputs } from \"./vision-executor.js\";\nexport { VisionExecutor } from \"./vision-executor.js\";\nexport type {\n  Gemma4VisionGridConfig,\n  Gemma4VisionPositionTensors,\n  ImageProcessorConfig,\n  PreprocessedImage,\n  VisionGridConfig,\n  VisionPositionTensors,\n} from \"./vision-preprocess.js\";\nexport {\n  buildGemma4PoolMatrix,\n  buildGemma4PosEmbeds,\n  buildGemma4RotaryCosSin,\n  buildGemma4VisionPositionTensors,\n  buildMRoPECosSin,\n  buildMRoPEPositionIds,\n  buildPosEmbeds,\n  buildPositionIds,\n  buildRotaryCosSin,\n  buildVisionPositionTensors,\n  GEMMA4_IMAGE_PROCESSOR,\n  mropeFreqDims,\n  preprocessImage,\n  preprocessImageGemma4,\n  QWEN3_5_IMAGE_PROCESSOR,\n  smartResize,\n} from \"./vision-preprocess.js\";\n"],"mappings":";;;;;;;;;;;AAyEA,SAAgB,wBAAwB,WAA2D;CACjG,MAAM,OAAQ,UAAU,iBAA6C;CACrE,MAAM,OAAQ,UAAU,eAA2C,EAAE;CAErE,MAAM,cAAc,KAAK;CACzB,MAAM,YAAY,KAAK;CACvB,MAAM,WAAY,KAAK,YAAuB,KAAK,MAAM,cAAc,UAAU;CACjF,MAAM,QAAQ,KAAK;CACnB,MAAM,oBAAoB,KAAK;CAC/B,MAAM,aAAa,KAAK;CAExB,MAAM,aADe,KAAK,gBAA2B,KACrB,aAAa;CAC7C,MAAM,sBAAuB,KAAK,uBAAkC;CAEpE,MAAM,cADQ,KAAK,mBAA+C,EAAE,EAC3C,cAAyB;CAClD,MAAM,eAAgB,KAAK,gBAA2B;AAGtD,QAAO;EACL,YAAY;EACZ,UAAU;EACV,SAAS;EACT;EACA,kBAAkB;EAClB,YARmB,KAAK,eAA0B;EASlD,WAAW;EACX,UAAU;EACV,mBAAmB;EACnB,WAAW;EACX,YAAY;EACb;;;;;;;;;AAUH,SAAgB,wBACd,QACA,QACA,QACA,MACA,MACA,WACc;CACd,MAAM,MAAM,IAAI,aAAa,OAAO,KAAK;CACzC,MAAM,YAAY,OAAO;CACzB,MAAM,YAAY,OAAO;AACzB,MAAK,IAAI,IAAI,GAAG,IAAI,MAAM,KAAK;EAC7B,MAAM,YAAY,IAAI;EACtB,MAAM,WAAW,IAAI;AACrB,OAAK,IAAI,IAAI,GAAG,IAAI,MAAM,KAAK;GAE7B,MAAM,IADO,OAAO,aAAa,KAAK,SAChB,IAAI,KAAK,IAAM;GACrC,MAAM,IAAI,KAAK,MAAM,IAAI,UAAU;AACnC,OAAI,IAAI,OAAO,KAAK,OAAO,WAAW,KAAK,IAAI,OAAO,WAAW;;;AAGrE,QAAO;;;;;;;;;AAUT,SAAgB,iCACd,SACA,WACA,MACA,MACM;CACN,MAAM,OAAO,gBAAgB;CAC7B,MAAM,IAAI,QAAQ,IAAI,KAAK;CAC3B,MAAM,KAAK,QAAQ,IAAI,2CAA2C;CAClE,MAAM,KAAK,QAAQ,IAAI,2CAA2C;AAElE,KAAI,CAAC,KAAK,EAAE,EAAE,gBAAgB,gBAAgB,CAAC,MAAM,CAAC,GAAI;CAC1D,MAAM,SACJ,GAAG,gBAAgB,eACf,GAAG,OACH,IAAI,aAAa,GAAG,KAAK,QAAQ,GAAG,KAAK,YAAY,GAAG,KAAK,aAAa,EAAE;CAClF,MAAM,SACJ,GAAG,gBAAgB,eACf,GAAG,OACH,IAAI,aAAa,GAAG,KAAK,QAAQ,GAAG,KAAK,YAAY,GAAG,KAAK,aAAa,EAAE;CAClF,MAAM,MAAM,wBAAwB,EAAE,MAAM,QAAQ,QAAQ,MAAM,MAAM,UAAU;AAClF,SAAQ,IAAI,MAAM;EAAE,MAAM;EAAK,OAAO,CAAC,MAAM,KAAK;EAAE,CAAC;AACrD,SAAQ,OAAO,2CAA2C;AAC1D,SAAQ,OAAO,2CAA2C;;;;;;;;;;;;;;;;;AAkB5D,SAAgB,gCACd,SACA,cACA,YACM;CACN,MAAM,QAAQ,OAAe;EAAE,MAAM,IAAI,aAAa,EAAE,CAAC,KAAK,EAAE;EAAE,OAAO,CAAC,EAAE;EAAE;AAC9E,KAAI,CAAC,QAAQ,IAAI,gBAAgB,cAAc,CAC7C,SAAQ,IAAI,gBAAgB,eAAe,KAAK,WAAW,CAAC;AAE9D,KAAI,CAAC,QAAQ,IAAI,gBAAgB,aAAa,CAC5C,SAAQ,IAAI,gBAAgB,cAAc,KAAK,aAAa,CAAC;;;;;;;;;;AAYjE,SAAgB,uBACd,OACA,SACM;CACN,MAAM,cAAc,QAAgD;AAClE,MAAI,CAAC,IAAK,QAAO;EACjB,MAAM,IAAI,QAAQ,IAAI,IAAI;AAC1B,MAAI,CAAC,EAAG,QAAO;EACf,MAAM,IACJ,EAAE,gBAAgB,eACd,EAAE,OACF,IAAI,aAAa,EAAE,KAAK,QAAQ,EAAE,KAAK,YAAY,EAAE,KAAK,aAAa,EAAE;AAC/E,SAAO,EAAE,SAAS,IAAI,EAAE,KAAK;;AAE/B,MAAK,MAAM,QAAQ,MAAM,OAAO;AAC9B,MAAI,KAAK,WAAW,gBAAiB;EACrC,MAAM,IAAI,KAAK;EACf,MAAM,OAAO,WAAW,EAAE,mBAAyC;EACnE,MAAM,OAAO,WAAW,EAAE,mBAAyC;EACnE,MAAM,OAAO,WAAW,EAAE,oBAA0C;EACpE,MAAM,OAAO,WAAW,EAAE,oBAA0C;AACpE,MAAI,SAAS,OAAW,GAAE,OAAO;AACjC,MAAI,SAAS,OAAW,GAAE,OAAO;AACjC,MAAI,SAAS,OAAW,GAAE,OAAO;AACjC,MAAI,SAAS,OAAW,GAAE,OAAO;AACjC,OAAK,MAAM,KAAK;GACd,EAAE;GACF,EAAE;GACF,EAAE;GACF,EAAE;GACH,CACC,KAAI,EAAG,SAAQ,OAAO,EAAE;;;;;;;;AAU9B,SAAgB,0BAA0B,WAAgD;CACxF,MAAM,OAAO,wBAAwB,UAAU;CAC/C,MAAM,EACJ,YAAY,aACZ,UAAU,WACV,SAAS,UACT,OACA,kBAAkB,mBAClB,YAAY,aACZ,UAAU,WACV,YAAY,QACV;CAEJ,MAAM,UAAU,YAAY;CAE5B,MAAMA,UAAsC,EAAE;CAC9C,MAAMC,QAAkB,EAAE;CAC1B,MAAMC,iBAA2B,EAAE;CAEnC,SAAS,UAAU,MAAwB;AACzC,UAAQ,KAAK,QAAQ;;CAEvB,SAAS,QAAQ,MAAoB;AACnC,QAAM,KAAK,KAAK;AAChB,iBAAe,KAAK,KAAK,GAAG;;CAE9B,SAAS,SAAS,MAAc,OAA4B,iBAAiB,MAAY;AACvF,YAAU;GAAE;GAAM;GAAO,OAAO;GAAO,SAAS;GAAY;GAAgB,CAAC;;CAE/E,SAAS,WAAW,MAAc,OAAkC;AAClE,YAAU;GAAE;GAAM;GAAO,OAAO;GAAO,SAAS;GAAc,CAAC;;;CAIjE,SAAS,OACP,IACA,OACA,QACA,QACA,GACA,GACA,SACM;AACN,UAAQ;GACN;GACA,QAAQ;GACR,QAAQ,CAAC,OAAO,OAAO;GACvB,SAAS,CAAC,OAAO;GACjB,YAAY;IAAE,UAAU;IAAS;IAAG;IAAG;GACxC,CAAC;;;;;;;;;;CAUJ,SAAS,cACP,IACA,OACA,QACA,QACA,GACA,GACA,SACA,UACM;AACN,UAAQ;GACN;GACA,QAAQ;GACR,QAAQ,CAAC,OAAO,OAAO;GACvB,SAAS,CAAC,OAAO;GACjB,YAAY;IACV,UAAU;IACV;IACA;IACA,oBAAoB,GAAG,SAAS;IAChC,oBAAoB,GAAG,SAAS;IAChC,qBAAqB,GAAG,SAAS;IACjC,qBAAqB,GAAG,SAAS;IAClC;GACF,CAAC;;;CAGJ,SAAS,QACP,IACA,OACA,GACA,QACA,QACA,WACM;AACN,UAAQ;GACN;GACA,QAAQ;GACR,QAAQ,CAAC,OAAO,EAAE;GAClB,SAAS,CAAC,OAAO;GACjB,YAAY;IAAE,aAAa;IAAQ;IAAK,gBAAgB;IAAW;GACpE,CAAC;;CAEJ,SAAS,YAAY,IAAY,GAAW,GAAW,QAAsB;AAC3E,UAAQ;GACN;GACA,QAAQ;GACR,QAAQ,CAAC,GAAG,EAAE;GACd,SAAS,CAAC,OAAO;GACjB,YAAY;IAAE,cAAc;IAAG;IAAa;GAC7C,CAAC;;AAIJ,YAAW,eAAe,CAAC,KAAK,UAAU,CAAC;AAC3C,YAAW,kBAAkB,CAAC,KAAK,YAAY,CAAC;AAChD,YAAW,WAAW,CAAC,KAAK,SAAS,CAAC;AACtC,YAAW,WAAW,CAAC,KAAK,SAAS,CAAC;AAItC,YAAW,cAAc,CAAC,MAAM,IAAI,CAAC;AAIrC,UAAS,gBAAgB,iBAAiB,CAAC,aAAa,UAAU,CAAC;AACnE,YAAW,gBAAgB,CAAC,KAAK,YAAY,CAAC;AAC9C,YAAW,UAAU,CAAC,KAAK,YAAY,CAAC;AAExC,QACE,kBACA,eACA,gBAAgB,iBAChB,gBACA,WACA,aACA,cACD;AACD,SAAQ;EACN,IAAI;EACJ,QAAQ;EACR,QAAQ,CAAC,gBAAgB,iBAAiB;EAC1C,SAAS,CAAC,SAAS;EACnB,YAAY;GAAE,cAAc;GAAgB;GAAa;EAC1D,CAAC;CAEF,IAAI,OAAO;AAEX,MAAK,IAAI,IAAI,GAAG,IAAI,OAAO,KAAK;EAC9B,MAAM,IAAI,QAAQ;AAGlB,WAAS,gBAAgB,UAAU,EAAE,EAAE,CAAC,YAAY,CAAC;AACrD,aAAW,GAAG,EAAE,MAAM,CAAC,KAAK,YAAY,CAAC;AACzC,UAAQ,GAAG,EAAE,cAAc,MAAM,gBAAgB,UAAU,EAAE,EAAE,GAAG,EAAE,MAAM,aAAa,KAAK;AAG5F,WAAS,gBAAgB,OAAO,EAAE,EAAE,CAAC,SAAS,YAAY,CAAC;AAC3D,WAAS,gBAAgB,OAAO,EAAE,EAAE,CAAC,SAAS,YAAY,CAAC;AAC3D,WAAS,gBAAgB,OAAO,EAAE,EAAE,CAAC,SAAS,YAAY,CAAC;AAC3D,aAAW,GAAG,EAAE,KAAK,CAAC,KAAK,QAAQ,CAAC;AACpC,aAAW,GAAG,EAAE,KAAK,CAAC,KAAK,QAAQ,CAAC;AACpC,aAAW,GAAG,EAAE,KAAK,CAAC,KAAK,QAAQ,CAAC;EACpC,MAAM,SAAS,+BAA+B,EAAE;AAChD,gBACE,GAAG,EAAE,UACL,GAAG,EAAE,MACL,gBAAgB,OAAO,EAAE,EACzB,GAAG,EAAE,KACL,aACA,SACA,GAAG,EAAE,MACL,GAAG,OAAO,SACX;AACD,gBACE,GAAG,EAAE,UACL,GAAG,EAAE,MACL,gBAAgB,OAAO,EAAE,EACzB,GAAG,EAAE,KACL,aACA,SACA,GAAG,EAAE,MACL,GAAG,OAAO,SACX;AACD,gBACE,GAAG,EAAE,UACL,GAAG,EAAE,MACL,gBAAgB,OAAO,EAAE,EACzB,GAAG,EAAE,KACL,aACA,SACA,GAAG,EAAE,MACL,GAAG,OAAO,SACX;AAGD,WAAS,gBAAgB,OAAO,EAAE,EAAE,CAAC,SAAS,CAAC;AAC/C,WAAS,gBAAgB,OAAO,EAAE,EAAE,CAAC,SAAS,CAAC;AAC/C,aAAW,GAAG,EAAE,MAAM,CAAC,KAAK,QAAQ,CAAC;AACrC,aAAW,GAAG,EAAE,MAAM,CAAC,KAAK,QAAQ,CAAC;AACrC,UAAQ,GAAG,EAAE,UAAU,GAAG,EAAE,KAAK,gBAAgB,OAAO,EAAE,EAAE,GAAG,EAAE,MAAM,UAAU,GAAG,EAAE,IAAI;AAC1F,UAAQ,GAAG,EAAE,UAAU,GAAG,EAAE,KAAK,gBAAgB,OAAO,EAAE,EAAE,GAAG,EAAE,MAAM,UAAU,GAAG,EAAE,IAAI;AAG1F,aAAW,GAAG,EAAE,MAAM,CAAC,KAAK,QAAQ,CAAC;AACrC,aAAW,GAAG,EAAE,MAAM,CAAC,KAAK,QAAQ,CAAC;AACrC,UAAQ;GACN,IAAI,GAAG,EAAE;GACT,QAAQ;GACR,QAAQ;IAAC,GAAG,EAAE;IAAM;IAAW;IAAU;GACzC,SAAS,CAAC,GAAG,EAAE,KAAK;GACpB,YAAY;IAAE;IAAW;IAAU;GACpC,CAAC;AACF,UAAQ;GACN,IAAI,GAAG,EAAE;GACT,QAAQ;GACR,QAAQ;IAAC,GAAG,EAAE;IAAM;IAAW;IAAU;GACzC,SAAS,CAAC,GAAG,EAAE,KAAK;GACpB,YAAY;IAAE;IAAW;IAAU;GACpC,CAAC;AAGF,aAAW,GAAG,EAAE,QAAQ,CAAC,KAAK,QAAQ,CAAC;AACvC,UAAQ;GACN,IAAI,GAAG,EAAE;GACT,QAAQ;GACR,QAAQ;IAAC,GAAG,EAAE;IAAM,GAAG,EAAE;IAAM,GAAG,EAAE;IAAI;GACxC,SAAS,CAAC,GAAG,EAAE,OAAO;GACtB,YAAY;IAAE,aAAa;IAAW,cAAc;IAAW;IAAU,QAAQ;IAAO;GACzF,CAAC;AAGF,WAAS,gBAAgB,OAAO,EAAE,EAAE,CAAC,aAAa,QAAQ,CAAC;AAC3D,aAAW,GAAG,EAAE,KAAK,CAAC,KAAK,YAAY,CAAC;AACxC,gBACE,GAAG,EAAE,UACL,GAAG,EAAE,QACL,gBAAgB,OAAO,EAAE,EACzB,GAAG,EAAE,KACL,SACA,aACA,GAAG,EAAE,QACL,GAAG,OAAO,SACX;AAID,WAAS,gBAAgB,aAAa,EAAE,EAAE,CAAC,YAAY,CAAC;AACxD,aAAW,GAAG,EAAE,aAAa,CAAC,KAAK,YAAY,CAAC;AAChD,UACE,GAAG,EAAE,kBACL,GAAG,EAAE,KACL,gBAAgB,aAAa,EAAE,EAC/B,GAAG,EAAE,aACL,aACA,GAAG,EAAE,IACN;AAGD,aAAW,GAAG,EAAE,QAAQ,CAAC,KAAK,YAAY,CAAC;AAC3C,cAAY,GAAG,EAAE,aAAa,MAAM,GAAG,EAAE,aAAa,GAAG,EAAE,OAAO;AAGlE,WAAS,gBAAgB,UAAU,EAAE,EAAE,CAAC,YAAY,CAAC;AACrD,aAAW,GAAG,EAAE,MAAM,CAAC,KAAK,YAAY,CAAC;AACzC,UACE,GAAG,EAAE,eACL,GAAG,EAAE,QACL,gBAAgB,UAAU,EAAE,EAC5B,GAAG,EAAE,MACL,aACA,GAAG,EAAE,OACN;AAGD,WAAS,gBAAgB,UAAU,EAAE,EAAE,CAAC,mBAAmB,YAAY,CAAC;AACxE,WAAS,gBAAgB,QAAQ,EAAE,EAAE,CAAC,mBAAmB,YAAY,CAAC;AACtE,WAAS,gBAAgB,UAAU,EAAE,EAAE,CAAC,aAAa,kBAAkB,CAAC;AACxE,aAAW,GAAG,EAAE,QAAQ,CAAC,KAAK,kBAAkB,CAAC;AACjD,aAAW,GAAG,EAAE,MAAM,CAAC,KAAK,kBAAkB,CAAC;AAC/C,aAAW,GAAG,EAAE,QAAQ,CAAC,KAAK,kBAAkB,CAAC;AACjD,aAAW,GAAG,EAAE,SAAS,CAAC,KAAK,kBAAkB,CAAC;AAClD,aAAW,GAAG,EAAE,OAAO,CAAC,KAAK,YAAY,CAAC;EAE1C,MAAM,UAAU,+BAA+B,EAAE;AACjD,gBACE,GAAG,EAAE,aACL,GAAG,EAAE,MACL,gBAAgB,UAAU,EAAE,EAC5B,GAAG,EAAE,QACL,aACA,mBACA,GAAG,EAAE,MACL,GAAG,QAAQ,YACZ;AACD,gBACE,GAAG,EAAE,WACL,GAAG,EAAE,MACL,gBAAgB,QAAQ,EAAE,EAC1B,GAAG,EAAE,MACL,aACA,mBACA,GAAG,EAAE,MACL,GAAG,QAAQ,UACZ;AACD,UAAQ;GACN,IAAI,GAAG,EAAE;GACT,QAAQ;GACR,QAAQ,CAAC,GAAG,EAAE,OAAO;GACrB,SAAS,CAAC,GAAG,EAAE,OAAO;GACtB,YAAY,EAAE,cAAc,GAAG,EAAE,QAAQ;GAC1C,CAAC;AACF,UAAQ;GACN,IAAI,GAAG,EAAE;GACT,QAAQ;GACR,QAAQ,CAAC,GAAG,EAAE,QAAQ,GAAG,EAAE,KAAK;GAChC,SAAS,CAAC,GAAG,EAAE,QAAQ;GACvB,YAAY;IAAE,cAAc,GAAG,EAAE;IAAQ,aAAa;IAAmB;GAC1E,CAAC;AACF,gBACE,GAAG,EAAE,aACL,GAAG,EAAE,SACL,gBAAgB,UAAU,EAAE,EAC5B,GAAG,EAAE,OACL,mBACA,aACA,GAAG,EAAE,SACL,GAAG,QAAQ,YACZ;AAGD,WAAS,gBAAgB,WAAW,EAAE,EAAE,CAAC,YAAY,CAAC;AACtD,aAAW,GAAG,EAAE,WAAW,CAAC,KAAK,YAAY,CAAC;AAC9C,UACE,GAAG,EAAE,gBACL,GAAG,EAAE,OACL,gBAAgB,WAAW,EAAE,EAC7B,GAAG,EAAE,WACL,aACA,GAAG,EAAE,OACN;AAGD,aAAW,GAAG,EAAE,QAAQ,CAAC,KAAK,YAAY,CAAC;AAC3C,cAAY,GAAG,EAAE,aAAa,GAAG,EAAE,QAAQ,GAAG,EAAE,WAAW,GAAG,EAAE,OAAO;AAEvE,SAAO,GAAG,EAAE;;AAUd,YAAW,cAAc,CAAC,MAAM,YAAY,CAAC;AAC7C,SAAQ;EACN,IAAI;EACJ,QAAQ;EACR,QAAQ,CAAC,cAAc,KAAK;EAC5B,SAAS,CAAC,aAAa;EACvB,YAAY;GAAE,UAAU;GAAM;GAAa,WAAW;GAAc;EACrE,CAAC;AAcF,UAAS,gBAAgB,cAAc,CAAC,YAAY,CAAC;AACrD,YAAW,mBAAmB,CAAC,MAAM,YAAY,CAAC;AAClD,SACE,qBACA,cACA,gBAAgB,cAChB,mBACA,aACA,aACD;AAKD,UAAS,gBAAgB,OAAO,CAAC,aAAa,YAAY,CAAC;AAC3D,YAAW,gBAAgB,CAAC,MAAM,YAAY,CAAC;AAC/C,QACE,YACA,mBACA,gBAAgB,OAChB,gBACA,aACA,aACA,eACD;AAED,UAAS,gBAAgB,eAAe,CAAC,YAAY,CAAC;AACtD,YAAW,oBAAoB,CAAC,MAAM,YAAY,CAAC;AACnD,SACE,sBACA,gBACA,gBAAgB,eAChB,oBACA,aACA,eACD;AA2BD,QAAO;EACL,cAAc;EACd,QA3B8B;GAC9B;GACA,YAAY;GACZ;GACA,cAAc;GACd;GACA;GACA,YAAY;GACZ,gBACI,UAAU,eACR,2BAAsC;GAC5C,cAAc;GACd,WAAW;GACX,WAAW,KAAK;GAChB,UAAU;GACV,WAAW;GACX,QAAQ;GACR,kBAAkB;GAClB,qBAAqB;GACrB,mBAAmB,KAAK;GACxB,kBAAkB;GACnB;EAOC,cALsC;GAAE,MAAM;GAAM,QAAQ;GAAM,KAAK;GAAO;EAM9E;EACA;EACA;EACA,QAAQ;GAAC;GAAe;GAAkB;GAAW;GAAW;GAAa;EAC7E,SAAS,CAAC,mBAAmB;EAC9B;;;;;ACnpBH,MAAM,eAAe;;;;;;AAmBrB,SAAgB,2BAA2B,WAAgD;CACzF,MAAM,OAAQ,UAAU,iBAA6C;CAErE,MAAM,cAAc,KAAK;CACzB,MAAM,YAAY,KAAK;CACvB,MAAM,WAAW,KAAK,MAAM,cAAc,UAAU;CACpD,MAAM,QAAQ,KAAK;CACnB,MAAM,oBAAoB,KAAK;CAC/B,MAAM,kBAAkB,KAAK;CAC7B,MAAM,qBAAqB,KAAK;CAChC,MAAM,cAAc,KAAK;CACzB,MAAM,sBAAsB,KAAK;CACjC,MAAM,aAAa,KAAK;CACxB,MAAM,YAAY,cAAc,sBAAsB,aAAa;CAEnE,MAAM,YAAY,eADC,qBAAqB;CAGxC,MAAMC,UAAsC,EAAE;CAC9C,MAAMC,QAAkB,EAAE;CAC1B,MAAMC,iBAA2B,EAAE;CAEnC,SAAS,UAAU,MAAwB;AACzC,UAAQ,KAAK,QAAQ;;CAEvB,SAAS,QAAQ,MAAoB;AACnC,QAAM,KAAK,KAAK;AAChB,iBAAe,KAAK,KAAK,GAAG;;CAE9B,SAAS,SAAS,MAAc,OAAkC;AAChE,YAAU;GAAE;GAAM;GAAO,OAAO;GAAO,SAAS;GAAY,gBAAgB;GAAM,CAAC;;CAErF,SAAS,WAAW,MAAc,OAAkC;AAClE,YAAU;GAAE;GAAM;GAAO,OAAO;GAAO,SAAS;GAAc,CAAC;;;;;;;CAqCjE,SAAS,WACP,IACA,OACA,QACA,MACA,QACA,GACA,GACA,SACM;AACN,UAAQ;GACN;GACA,QAAQ;GACR,QAAQ;IAAC;IAAO;IAAQ;IAAK;GAC7B,SAAS,CAAC,OAAO;GACjB,YAAY;IAAE,UAAU;IAAS;IAAG;IAAG;GACxC,CAAC;;CAEJ,SAAS,UACP,IACA,OACA,GACA,GACA,QACA,QACA,WACM;AACN,UAAQ;GACN;GACA,QAAQ;GACR,QAAQ;IAAC;IAAO;IAAG;IAAE;GACrB,SAAS,CAAC,OAAO;GACjB,YAAY;IAAE,aAAa;IAAQ,KAAK;IAAc,gBAAgB;IAAW;GAClF,CAAC;;AAIJ,YAAW,eAAe,CAAC,KAAK,UAAU,CAAC;AAC3C,YAAW,kBAAkB,CAAC,KAAK,YAAY,CAAC;AAChD,YAAW,WAAW,CAAC,KAAK,SAAS,CAAC;AACtC,YAAW,WAAW,CAAC,KAAK,SAAS,CAAC;AAGtC,UAAS,eAAe,qBAAqB,CAAC,aAAa,UAAU,CAAC;AACtE,UAAS,eAAe,mBAAmB,CAAC,YAAY,CAAC;AACzD,YAAW,gBAAgB,CAAC,KAAK,YAAY,CAAC;AAC9C,YAAW,kBAAkB,CAAC,KAAK,YAAY,CAAC;AAChD,YAAW,UAAU,CAAC,KAAK,YAAY,CAAC;AAExC,YACE,gBACA,eACA,eAAe,qBACf,eAAe,mBACf,kBACA,WACA,aACA,cACD;AACD,SAAQ;EACN,IAAI;EACJ,QAAQ;EACR,QAAQ,CAAC,kBAAkB,iBAAiB;EAC5C,SAAS,CAAC,SAAS;EACnB,YAAY;GAAE,cAAc;GAAkB;GAAa;EAC5D,CAAC;CAEF,IAAI,OAAO;CACX,MAAM,UAAU,cAAc;AAE9B,MAAK,IAAI,IAAI,GAAG,IAAI,OAAO,KAAK;EAC9B,MAAM,IAAI,QAAQ;AAGlB,WAAS,eAAe,eAAe,EAAE,EAAE,CAAC,YAAY,CAAC;AACzD,WAAS,eAAe,eAAe,EAAE,EAAE,CAAC,YAAY,CAAC;AACzD,aAAW,GAAG,EAAE,MAAM,CAAC,KAAK,YAAY,CAAC;AACzC,YACE,GAAG,EAAE,SACL,MACA,eAAe,eAAe,EAAE,EAChC,eAAe,eAAe,EAAE,EAChC,GAAG,EAAE,MACL,aACA,KACD;AAGD,WAAS,eAAe,aAAa,EAAE,EAAE,CAAC,SAAS,YAAY,CAAC;AAChE,WAAS,eAAe,aAAa,EAAE,EAAE,CAAC,QAAQ,CAAC;AACnD,aAAW,GAAG,EAAE,UAAU,CAAC,KAAK,QAAQ,CAAC;AACzC,aAAW,GAAG,EAAE,OAAO,CAAC,KAAK,QAAQ,CAAC;AACtC,aACE,GAAG,EAAE,YACL,GAAG,EAAE,MACL,eAAe,aAAa,EAAE,EAC9B,eAAe,aAAa,EAAE,EAC9B,GAAG,EAAE,OACL,aACA,SACA,GAAG,EAAE,KACN;AAGD,aAAW,GAAG,EAAE,KAAK,CAAC,KAAK,YAAY,CAAC;AACxC,aAAW,GAAG,EAAE,KAAK,CAAC,KAAK,YAAY,CAAC;AACxC,aAAW,GAAG,EAAE,KAAK,CAAC,KAAK,YAAY,CAAC;EACxC,MAAM,aAAa,IAAY,QAAgB,WAC7C,QAAQ;GACN;GACA,QAAQ;GACR,QAAQ,CAAC,GAAG,EAAE,MAAM;GACpB,SAAS,CAAC,OAAO;GACjB,YAAY;IACV,UAAU;IACV,WAAW;IACX,YAAY;IACZ,gBAAgB,GAAG,EAAE;IACtB;GACF,CAAC;AACJ,YAAU,GAAG,EAAE,WAAW,GAAG,EAAE,KAAK,EAAE;AACtC,YAAU,GAAG,EAAE,WAAW,GAAG,EAAE,KAAK,YAAY;AAChD,YAAU,GAAG,EAAE,WAAW,GAAG,EAAE,KAAK,IAAI,YAAY;AAGpD,aAAW,GAAG,EAAE,MAAM,CAAC,KAAK,YAAY,CAAC;AACzC,aAAW,GAAG,EAAE,MAAM,CAAC,KAAK,YAAY,CAAC;AACzC,UAAQ;GACN,IAAI,GAAG,EAAE;GACT,QAAQ;GACR,QAAQ;IAAC,GAAG,EAAE;IAAK;IAAW;IAAU;GACxC,SAAS,CAAC,GAAG,EAAE,KAAK;GACpB,YAAY;IAAE;IAAW;IAAU;GACpC,CAAC;AACF,UAAQ;GACN,IAAI,GAAG,EAAE;GACT,QAAQ;GACR,QAAQ;IAAC,GAAG,EAAE;IAAK;IAAW;IAAU;GACxC,SAAS,CAAC,GAAG,EAAE,KAAK;GACpB,YAAY;IAAE;IAAW;IAAU;GACpC,CAAC;AAGF,aAAW,GAAG,EAAE,QAAQ,CAAC,KAAK,YAAY,CAAC;AAC3C,UAAQ;GACN,IAAI,GAAG,EAAE;GACT,QAAQ;GACR,QAAQ;IAAC,GAAG,EAAE;IAAM,GAAG,EAAE;IAAM,GAAG,EAAE;IAAI;GACxC,SAAS,CAAC,GAAG,EAAE,OAAO;GACtB,YAAY;IAAE,aAAa;IAAW,cAAc;IAAW;IAAU,QAAQ;IAAO;GACzF,CAAC;AAGF,WAAS,eAAe,cAAc,EAAE,EAAE,CAAC,aAAa,YAAY,CAAC;AACrE,WAAS,eAAe,cAAc,EAAE,EAAE,CAAC,YAAY,CAAC;AACxD,aAAW,GAAG,EAAE,WAAW,CAAC,KAAK,YAAY,CAAC;AAC9C,aAAW,GAAG,EAAE,QAAQ,CAAC,KAAK,YAAY,CAAC;AAC3C,aACE,GAAG,EAAE,WACL,GAAG,EAAE,QACL,eAAe,cAAc,EAAE,EAC/B,eAAe,cAAc,EAAE,EAC/B,GAAG,EAAE,QACL,aACA,aACA,GAAG,EAAE,OACN;AAGD,aAAW,GAAG,EAAE,QAAQ,CAAC,KAAK,YAAY,CAAC;AAC3C,UAAQ;GACN,IAAI,GAAG,EAAE;GACT,QAAQ;GACR,QAAQ,CAAC,MAAM,GAAG,EAAE,OAAO;GAC3B,SAAS,CAAC,GAAG,EAAE,OAAO;GACtB,YAAY;IAAE,cAAc;IAAM;IAAa;GAChD,CAAC;AAGF,WAAS,eAAe,eAAe,EAAE,EAAE,CAAC,YAAY,CAAC;AACzD,WAAS,eAAe,eAAe,EAAE,EAAE,CAAC,YAAY,CAAC;AACzD,WAAS,eAAe,aAAa,EAAE,EAAE,CAAC,mBAAmB,YAAY,CAAC;AAC1E,WAAS,eAAe,aAAa,EAAE,EAAE,CAAC,kBAAkB,CAAC;AAC7D,WAAS,eAAe,aAAa,EAAE,EAAE,CAAC,aAAa,kBAAkB,CAAC;AAC1E,WAAS,eAAe,aAAa,EAAE,EAAE,CAAC,YAAY,CAAC;AACvD,aAAW,GAAG,EAAE,MAAM,CAAC,KAAK,YAAY,CAAC;AACzC,aAAW,GAAG,EAAE,UAAU,CAAC,KAAK,kBAAkB,CAAC;AACnD,aAAW,GAAG,EAAE,OAAO,CAAC,KAAK,kBAAkB,CAAC;AAChD,aAAW,GAAG,EAAE,QAAQ,CAAC,KAAK,kBAAkB,CAAC;AACjD,aAAW,GAAG,EAAE,UAAU,CAAC,KAAK,YAAY,CAAC;AAC7C,aAAW,GAAG,EAAE,OAAO,CAAC,KAAK,YAAY,CAAC;AAC1C,aAAW,GAAG,EAAE,QAAQ,CAAC,KAAK,YAAY,CAAC;AAE3C,YACE,GAAG,EAAE,SACL,GAAG,EAAE,QACL,eAAe,eAAe,EAAE,EAChC,eAAe,eAAe,EAAE,EAChC,GAAG,EAAE,MACL,aACA,GAAG,EAAE,OACN;AACD,aACE,GAAG,EAAE,UACL,GAAG,EAAE,MACL,eAAe,aAAa,EAAE,EAC9B,eAAe,aAAa,EAAE,EAC9B,GAAG,EAAE,OACL,aACA,mBACA,GAAG,EAAE,KACN;AACD,UAAQ;GACN,IAAI,GAAG,EAAE;GACT,QAAQ;GACR,QAAQ,CAAC,GAAG,EAAE,MAAM;GACpB,SAAS,CAAC,GAAG,EAAE,OAAO;GACtB,YAAY,EAAE,cAAc,GAAG,EAAE,OAAO;GACzC,CAAC;AACF,aACE,GAAG,EAAE,UACL,GAAG,EAAE,QACL,eAAe,aAAa,EAAE,EAC9B,eAAe,aAAa,EAAE,EAC9B,GAAG,EAAE,OACL,mBACA,aACA,GAAG,EAAE,OACN;AACD,UAAQ;GACN,IAAI,GAAG,EAAE;GACT,QAAQ;GACR,QAAQ,CAAC,GAAG,EAAE,QAAQ,GAAG,EAAE,MAAM;GACjC,SAAS,CAAC,GAAG,EAAE,OAAO;GACtB,YAAY;IAAE,cAAc,GAAG,EAAE;IAAQ;IAAa;GACvD,CAAC;AAEF,SAAO,GAAG,EAAE;;AAMd,UAAS,eAAe,gBAAgB,CAAC,YAAY,CAAC;AACtD,UAAS,eAAe,gBAAgB,CAAC,YAAY,CAAC;AACtD,UAAS,eAAe,eAAe,CAAC,WAAW,UAAU,CAAC;AAC9D,UAAS,eAAe,eAAe,CAAC,UAAU,CAAC;AACnD,UAAS,eAAe,eAAe,CAAC,iBAAiB,UAAU,CAAC;AACpE,UAAS,eAAe,eAAe,CAAC,gBAAgB,CAAC;AAEzD,YAAW,kBAAkB,CAAC,KAAK,YAAY,CAAC;AAIhD,YAAW,oBAAoB,CAAC,MAAM,UAAU,CAAC;AACjD,YAAW,iBAAiB,CAAC,MAAM,UAAU,CAAC;AAC9C,YAAW,kBAAkB,CAAC,MAAM,UAAU,CAAC;AAC/C,YAAW,oBAAoB,CAAC,MAAM,gBAAgB,CAAC;AACvD,YAAW,oBAAoB,CAAC,MAAM,gBAAgB,CAAC;AAEvD,WACE,mBACA,MACA,eAAe,gBACf,eAAe,gBACf,kBACA,aACA,KACD;AAED,YACE,kBACA,kBACA,eAAe,eACf,eAAe,eACf,iBACA,WACA,WACA,gBACD;AACD,SAAQ;EACN,IAAI;EACJ,QAAQ;EACR,QAAQ,CAAC,gBAAgB;EACzB,SAAS,CAAC,iBAAiB;EAC3B,YAAY,EAAE,cAAc,iBAAiB;EAC9C,CAAC;AACF,YACE,kBACA,kBACA,eAAe,eACf,eAAe,eACf,oBACA,WACA,iBACA,mBACD;AAyBD,QAAO;EACL,cAAc;EACd,QAzB8B;GAC9B;GACA,YAAY;GACZ;GACA,cAAc;GACd;GACA;GACA,YAAY;GACZ,gBAAgB,KAAK;GACrB,cAAc;GACd,WAAW;GACX,WAAW;GACX,UAAU;GACV,WAAW;GACX,QAAQ;GACR,kBAAkB;GAClB,qBAAqB;GACrB,mBAAmB;GACnB,kBAAkB;GACnB;EAOC,cALsC;GAAE,MAAM;GAAM,QAAQ;GAAM,KAAK;GAAO;EAM9E;EACA;EACA;EACA,QAAQ;GAAC;GAAe;GAAkB;GAAW;GAAU;EAC/D,SAAS,CAAC,mBAAmB;EAC9B;;;;;;;;;;;;;;;;;;;;;;;;;ACzZH,MAAM,mBAAmB;AACzB,MAAM,WAAW;AACjB,MAAM,mBAAmB;;AAEzB,MAAM,0BAA0B;AAChC,MAAM,6BAA6B;;AAEnC,MAAMC,2BAAyB;;AA2D/B,SAASC,qBACP,OACA,SACyD;CACzD,MAAM,sBAAM,IAAI,KAAyD;AACzE,MAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,MAAM,QAAQ,EAAE;AACxD,MAAI,KAAK,YAAY,WACnB;EAOF,MAAM,IACJ,QAAQ,IAAI,KAAK,KAAK,KAAK,iBAAiB,QAAQ,IAAI,KAAK,eAAe,GAAG;AACjF,MAAI,EACF,KAAI,IAAI,MAAM,EAAE;;AAGpB,QAAO;;;;;;;AAQT,SAASC,gBAAc,OAAqB,KAAiB,MAAsB;CACjF,MAAM,IAAI,MAAM;CAChB,MAAM,QAAQ,MAAM,KAAK,EAAE,QAAQ,GAAG,GAAG,GAAG,MAAM,EAAE,CAAC,MAAM,GAAG,MAAM,MAAM,KAAK,MAAM,GAAG;CACxF,IAAI,MAAM;CACV,IAAI,SAAS;AACb,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,SAAO,MAAM,MAAM;AACnB,MAAI,OAAO,MAAM;AACf,YAAS,IAAI;AACb;;;CAGJ,IAAI,UAAU;AACd,MAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,IAC1B,YAAW,MAAM,MAAM;CAEzB,IAAIC,SAAO,KAAK,QAAQ,GAAG;AAC3B,MAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,KAAK;AAC/B,YAAQ,MAAM,MAAM;AACpB,MAAIA,UAAQ,EACV,QAAO,IAAI,MAAM;;AAGrB,QAAO,IAAI,MAAM;;;;;;AAOnB,SAAS,UAAU,QAAsB,GAAiB;CACxD,MAAM,IAAI,OAAO;AACjB,KAAI,KAAK,KAAK,KAAK,EACjB;CAEF,MAAM,QAAQ,MAAM,KAAK,EAAE,QAAQ,GAAG,GAAG,GAAG,MAAM,EAAE,CAAC,MAAM,GAAG,MAAM,OAAO,KAAK,OAAO,GAAG;CAC1F,IAAI,UAAU;AACd,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IACrB,YAAW,OAAO,MAAM;AAE1B,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IACrB,QAAO,MAAM,MAAM;AAErB,KAAI,UAAU,EACZ,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IACrB,QAAO,MAAM;;;AAMnB,SAASC,iBAAe,QAA4B;CAClD,IAAI,OAAO,OAAO;AAClB,MAAK,IAAI,IAAI,GAAG,IAAI,OAAO,QAAQ,IACjC,KAAI,OAAO,KAAK,KACd,QAAO,OAAO;CAGlB,IAAI,MAAM;AACV,MAAK,IAAI,IAAI,GAAG,IAAI,OAAO,QAAQ,KAAK;EACtC,MAAM,IAAI,KAAK,IAAI,OAAO,KAAK,KAAK;AACpC,SAAO,KAAK;AACZ,SAAO;;AAET,MAAK,IAAI,IAAI,GAAG,IAAI,OAAO,QAAQ,IACjC,QAAO,MAAM;;AAIjB,IAAa,UAAb,MAAa,QAAQ;CACnB,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;;CAER,AAAQ;;CAER,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ,aAAa;;CAErB,AAAQ;;CAGR,AAAS;;CAET,AAAS;CAET,AAAQ,YAAY,KAAiB,QAAuB,WAAmB,OAAkB;AAC/F,OAAK,MAAM;AACX,OAAK,SAAS;AACd,OAAK,YAAY,OAAO;AACxB,OAAK,YAAY,OAAO;AACxB,OAAK,MAAM,gBAAgB,OAAO,UAAU;AAC5C,OAAK,YAAY;AAMjB,OAAK,QAAQ,OAAO,eAAe,OAAO;AAE1C,OAAK,gBADa,OAAO,UAAU,iBAA0C,EAAE,EAClD,MAAM;EAGnC,MAAM,eAAe,OAAO,UAAU;AAGtC,OAAK,YAAY,MAAM,QAAQ,cAAc,mBAAmB,GAC3D,aAAa,mBAAiC,QAC5C,MAAmB,OAAO,MAAM,SAClC,GACD,EAAE;EAEN,MAAM,SAAS,KAAK,UAAU;EAC9B,MAAM,QAAQ,KAAK,UAAU;AAC7B,OAAK,UAAW,KAAK,UAAU,YAAuB,KAAK,MAAM,SAAS,MAAM;AAChF,OAAK,WAAY,KAAK,UAAU,cAAyB;AAGzD,OAAK,aAAa,0BAA0B,KAAK,UAAU;AAC3D,OAAK,6BAAa,IAAI,KAAqB;EAC3C,MAAM,eAAgB,KAAK,UAAU,sBAAkC;AACvE,OAAK,MAAM,SAAS,KAAK,YAAY;GACnC,IAAI,QAAQ;AACZ,OAAI,cAAc;IAChB,MAAM,IAAI,OAAO,gBAAgB,IAAI,yBAAyB,MAAM,eAAe;AAEnF,YAAQ,eADG,IAAM,EAAE,KAAsB,MAAM,IAAK,GACzB,KAAK,IAAI;;AAEtC,QAAK,WAAW,IAAI,OAAO,MAAM;;AAInC,OAAK,eAAe,KAAK,cAAc,KAAK,QAAQ,UAAU;;;;;;;;;CAUhE,AAAQ,cAAc,KAAiB,QAAuB,WAA6B;EACzF,MAAM,QAAQ,KAAK,UAAU;EAC7B,MAAM,QAAQ,qBAAqB,KAAK,WAAW,QAAQ,OAAO,OAAU;AAC5E,MAAI,MACF,sBAAqB,OAAO,OAAO,gBAAgB;EAErD,MAAM,OAAO,IAAI,SAAS,KAAK,OAAO;GAAE;GAAW,QAAQ;GAAO,CAAC;AACnE,OAAK,iBAAiBH,qBAAmB,OAAO,OAAO,gBAAgB,CAAC;AACxE,OAAK,gBAAgB;AACrB,SAAO;;CAGT,aAAa,OAAO,UAA0B,EAAE,EAAoB;EAClE,MAAM,MAAM,MAAM,SAAS;EAC3B,MAAM,SAAS,MAAM,YAAY;GAC/B,MAAM,QAAQ;GACd,WAAW,QAAQ;GACnB,UAAU,QAAQ;GAClB,SAAS,QAAQ;GACjB,UAAU,QAAQ;GAClB,YAAY,QAAQ;GACrB,CAAC;EACF,MAAM,SAAU,OAAO,UAAU,2BAAsC;AAEvE,SAAO,IAAI,QAAQ,KAAK,QADN,KAAK,IAAI,QAAQ,aAAa,MAAM,OAAO,EAClB,QAAQ,SAAS,KAAK;;;CAInE,AAAQ,YAAY,WAAyB,UAAwB;EACnE,MAAM,WAAW,KAAK,UAAU;AAChC,OAAK,MAAM,SAAS,KAAK,YAAY;GACnC,MAAM,QAAQ,KAAK,WAAW,IAAI,MAAM,IAAI;GAC5C,MAAM,EAAE,KAAK,QAAQ,qBAAqB,WAAW,KAAK,SAAS,KAAK,UAAU,MAAM;AACxF,OAAI,aAAa,GAAG;AAClB,SAAK,aAAa,WAAW,cAAc,MAAM,EAAE,IAAI;AACvD,SAAK,aAAa,WAAW,cAAc,MAAM,EAAE,IAAI;UAClD;AACL,SAAK,aAAa,aAAa,cAAc,MAAM,EAAE,KAAK,WAAW,SAAS;AAC9E,SAAK,aAAa,aAAa,cAAc,MAAM,EAAE,KAAK,WAAW,SAAS;;;;;;;;;;;CAYpF,kBAA4B;AAC1B,SAAO,CAAC,GAAG,KAAK,UAAU;;;CAI5B,IAAI,aAAqB;AACvB,SAAO;;;;;;;;;CAUT,MAAM,MAAM,MAAc,OAAqB,EAAE,EAAwB;AACvE,MAAI,KAAK,WACP,OAAM,IAAI,MAAM,8BAA8B;EAEhD,MAAM,cAAc,KAAK,eAAe;EACxC,MAAM,OAAO,KAAK,QAAQ;EAC1B,MAAM,OAAO,KAAK,QAAQ;EAC1B,MAAM,oBAAoB,KAAK,qBAAqB;EACpD,MAAM,eAAe,KAAK,IAAI,KAAK,gBAAgBD,0BAAwB,KAAK,UAAU;EAM1F,MAAM,aACJ,KAAK,UAAU,SAAS,IACpB,IAAI,KAAK,eAAe,KAAK,UAAU,IAAI,MAAM,CAAC,IAAI,SACtD;EAEN,MAAMK,SAAmB;GAAC;GAAqB,GAD/B,KAAK,UAAU,OAAO,WAAW;GACU;GAAkB;GAAkB;AAC/F,MAAI,OAAO,UAAU,KAAK,UACxB,OAAM,IAAI,MAAM,kBAAkB,OAAO,OAAO,uBAAuB,KAAK,UAAU,GAAG;AAI3F,OAAK,aAAa,OAAO;EACzB,MAAM,YAAY,qBAAqB,QAAQ,KAAK,IAAI;AACxD,OAAK,YAAY,WAAW,EAAE;EAC9B,MAAM,EAAE,WAAW,MAAM,KAAK,aAAa,QAAQ,IAAI,YAAY,OAAO,CAAC;EAG3E,MAAM,YAAY,YAAY,KAAK;EACnC,MAAM,EAAE,aAAa,aAAa,MAAM,KAAK,cAAc,QAAQ,QAAQ;GACzE;GACA;GACA;GACA;GACA;GACA,WAAW,KAAK,aAAa,OAAO;GACrC,CAAC;EAGF,MAAM,SAAS,YAAY,SAAU,YAAY,SAAS,KAAK,IAAI;AACnE,MAAI,WAAW,EACb,OAAM,IAAI,MACR,oDAAoD,SAAS,gBAC5C,YAAY,OAAO,oCACrC;EAEH,MAAM,EAAE,OAAO,cAAc,mBAAmB,YAAY,MAAM,GAAG,OAAO,EAAE,KAAK,IAAI;EACvF,MAAM,MAAM,MAAM,KAAK,YAAY,OAAO,UAAU;EAEpD,MAAM,aAAa,YAAY,KAAK,GAAG,aAAa;EACpD,MAAM,eAAgB,YAAY,WAAY;AAC9C,UAAQ,IACN,iBAAiB,UAAU,SAAS,aAAa,QAAQ,EAAE,CAAC,SAAS,UAAU,QAC7E,EACD,CAAC,SAAS,eAAe,WAAW,QAAQ,EAAE,CAAC,GACjD;AAED,SAAO;GAAE;GAAK,YAAY;GAAkB,QAAQ;GAAW;GAAc;;;;;;;;;CAU/E,MAAc,cACZ,QACA,eACA,GAQuD;EACvD,MAAMC,MAAgB,CAAC,GAAG,OAAO;EACjC,MAAMC,cAAwB,EAAE;EAChC,MAAM,WAAW,EAAE,YAAY,KAAK,IAAI;EACxC,IAAI,SAAS;EACb,IAAI,WAAW;EACf,IAAI,WAAW;AAEf,OAAK,IAAI,OAAO,GAAG,OAAO,EAAE,cAAc,QAAQ;GAChD,MAAM,YAAY,WACd,KAAK,iBAAiB,QAAQ,YAAY,SAAS,KAAK,IAAI,gBAAgB;IAC1E,aAAa,EAAE;IACf,MAAM,EAAE;IACR,MAAM,EAAE;IACR,mBAAmB,EAAE;IACrB,UAAU;IACX,CAAC,GACF,KAAK,OAAO,OAAO;AAEvB,OAAI,cAAc,KAAK,IAAI,cACzB,YAAW;YACF,cAAc,KAAK,IAAI,aAAa;AAC7C,QAAI,KAAK,UAAU;AACnB,eAAW;AACX;cACS,YAAY,aAAa,KAAK,IAAI,iBAC3C,aAAY,KAAK,UAAU;AAE7B,OAAI,KAAK,UAAU;AAEnB,OAAI,YAAY,YAAY,UAAU,UAAU;AAC9C,eAAW;AACX;;AAEF,OAAI,IAAI,UAAU,KAAK,UACrB;AAKF,QAAK,YACH,aAAa,GAAG,KAAK,kBAAkB,IAAI,CAAC,EAC5C,KAAK,aAAa,cACnB;AAED,aADY,MAAM,KAAK,aAAa,QAAQ,IAAI,YAAY,CAAC,UAAU,CAAC,CAAC,EAC5D;;AAEf,SAAO;GAAE;GAAa;GAAU;;;CAIlC,AAAQ,kBAAkB,KAAuB;EAG/C,MAAM,YAAY,qBAAqB,KAAK,KAAK,IAAI;AACrD,SAAO,UAAU,UAAU,SAAS;;;CAItC,AAAQ,OAAO,QAA8B;EAC3C,IAAI,OAAO;EACX,IAAI,UAAU,OAAO;AACrB,OAAK,IAAI,IAAI,GAAG,IAAI,OAAO,QAAQ,IACjC,KAAI,OAAO,KAAK,SAAS;AACvB,aAAU,OAAO;AACjB,UAAO;;AAGX,SAAO;;;;;;;;CAST,AAAQ,iBACN,QACA,UACA,GAOQ;EACR,MAAM,WAAW,KAAK,IAAI,mBAAmB,KAAK,IAAI,eAAe;EACrE,MAAM,SAAS,KAAK,IAAI;EAGxB,MAAM,WAAW,aAAa;EAC9B,MAAM,IAAI,UAAU,WAAW,IAAI;EACnC,MAAM,MAAM,IAAI,WAAW,EAAE;AAC7B,OAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,IAC1B,KAAI,KAAK,WAAW;AAEtB,MAAI,SACF,KAAI,UAAU,KAAK,IAAI;EAGzB,MAAM,OAAO,EAAE,sBAAsB,IAAM,IAAI,IAAI,EAAE,SAAS,GAAG;EACjE,MAAM,SAAS,IAAI,aAAa,EAAE;AAClC,OAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;GAC1B,IAAI,IAAI,OAAO,IAAI;AACnB,OAAI,MAAM,IAAI,IAAI,GAAG,CACnB,KAAI,IAAI,IAAI,IAAI,EAAE,oBAAoB,IAAI,EAAE;AAE9C,UAAO,KAAK,IAAI,EAAE;;AAEpB,mBAAe,OAAO;AACtB,YAAU,QAAQ,EAAE,KAAK;AACzB,SAAOL,gBAAc,QAAQ,KAAK,EAAE,KAAK;;;;;;;;;;;CAY3C,MAAc,YAAY,OAAoB,WAA0C;EACtF,MAAM,SAAS;EACf,MAAM,WAAW;EACjB,MAAM,QAAQ;EACd,MAAM,MAAM,IAAI,aAAa,YAAY,SAAS;AAElD,OAAK,IAAI,QAAQ,GAAG,QAAQ,WAAW,SAAS,OAAO;GACrD,MAAM,WAAW,KAAK,IAAI,GAAG,QAAQ,SAAS;GAC9C,MAAM,YAAY,QAAQ;GAC1B,MAAM,MAAM,KAAK,IAAI,WAAW,QAAQ,MAAM;GAC9C,MAAM,YAAY,MAAM;GAGxB,MAAM,WAAW,IAAI,YAAY,SAAS,UAAU;AACpD,QAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,IAC1B,MAAK,IAAI,IAAI,GAAG,IAAI,WAAW,IAC7B,UAAS,IAAI,YAAY,KAAK,MAAM,IAAI,aAAa,WAAW;GAIpE,MAAM,MAAM,MAAM,KAAK,kBAAkB,UAAU,UAAU;GAE7D,MAAM,cAAc,YAAY;GAChC,MAAM,eAAe,MAAM,SAAS;AACpC,OAAI,IAAI,IAAI,SAAS,aAAa,cAAc,YAAY,EAAE,QAAQ,SAAS;;AAIjF,OAAK,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,IAC9B,KAAI,KAAK,KAAK,IAAI,GAAG,KAAK,IAAI,IAAI,IAAI,GAAG,CAAC;AAE5C,SAAO;;;CAIT,MAAc,kBAAkB,UAAuB,WAA0C;EAC/F,MAAM,QAAQ,8BAA8B,EAAE,WAAW,WAAW,CAAC;EACrE,MAAM,OAAO,IAAI,SAAS,KAAK,KAAK,OAAO;GAAE,WAAW;GAAW,QAAQ;GAAO,CAAC;AACnF,MAAI;AACF,QAAK,iBAAiBD,qBAAmB,OAAO,KAAK,OAAO,aAAa,CAAC;AAC1E,QAAK,gBAAgB;AACrB,QAAK,OAAO;AACZ,QAAK,WAAW,eAAe,SAAS;AACxC,UAAO,MAAM,KAAK,eAAe,MAAM,QAAQ,IAAI,YAAY,SAAS;YAChE;AACR,QAAK,SAAS;;;CAIlB,UAAgB;AACd,MAAI,KAAK,WACP;AAEF,OAAK,aAAa;AAClB,OAAK,aAAa,SAAS;AAC3B,OAAK,OAAO,gBAAgB,OAAO;AACnC,OAAK,OAAO,aAAa,OAAO;;;;;;AC5lBpC,MAAM,+BAA+B;AACrC,MAAM,gCAAgC;AACtC,MAAM,kCAAkC;AACxC,MAAM,gBAAgB;AACtB,MAAM,wBAAwB;AAC9B,MAAM,sBAAsB;AAC5B,MAAM,sBAAsB;AAC5B,MAAM,iBAAiB;AACvB,MAAM,sBAAsB;AAC5B,MAAM,kBAAkB;;;;;;;AAQxB,SAAgB,oBAAoB,OAAe,YAA4B;CAC7E,MAAM,IAAI,MAAM,QAAQ,qBAAqB,GAAG;CAChD,MAAM,MAAM,KAAK,IAAI,EAAE,QAAQ,WAAW,OAAO;AACjD,MAAK,IAAI,IAAI,KAAK,IAAI,GAAG,IACvB,KAAI,EAAE,MAAM,EAAE,SAAS,EAAE,KAAK,WAAW,MAAM,GAAG,EAAE,CAClD,QAAO,WAAW,MAAM,EAAE;AAG9B,QAAO;;;;;;;;;;AAWT,SAAgB,iBACd,OACA,YACA,WAAW,+BACH;CACR,MAAM,YAAY,IAAI,MAAM,aAAa,CAAC,QAAQ,gBAAgB,IAAI,CAAC,MAAM,CAAC;CAC9E,MAAM,QAAQ,WAAW,MAAM,eAAe,CAAC,OAAO,QAAQ;CAC9D,IAAI,UAAU;AACd,MAAK,IAAI,IAAI,GAAG,KAAK,MAAM,QAAQ,KAAK;EACtC,MAAM,SAAS,IAAI,MAAM,MAAM,GAAG,EAAE,CAAC,KAAK,IAAI,CAAC,aAAa,CAAC;AAC7D,MAAI,UAAU,SAAS,OAAO,CAC5B,WAAU;MAEV;;AAGJ,KAAI,WAAW,SACb,QAAO,MAAM,MAAM,QAAQ,CAAC,KAAK,IAAI;AAEvC,QAAO;;;;;;;;AAST,SAAgB,uBACd,MACA,SAAS,iCACD;CACR,MAAM,QAAQ,KAAK,MAAM,eAAe,CAAC,OAAO,QAAQ;CACxD,MAAM,uBAAO,IAAI,KAAa;AAC9B,MAAK,IAAI,IAAI,GAAG,IAAI,UAAU,MAAM,QAAQ,KAAK;EAC/C,MAAM,OAAO,MACV,MAAM,GAAG,IAAI,OAAO,CACpB,KAAK,IAAI,CACT,aAAa;AAChB,MAAI,KAAK,IAAI,KAAK,CAChB,QAAO,MAAM,MAAM,GAAG,EAAE,CAAC,KAAK,IAAI;AAEpC,OAAK,IAAI,KAAK;;AAEhB,QAAO;;;;;;AAOT,SAAgB,UAAU,MAAc,KAAqB;AAC3D,KAAI,KAAK,UAAU,IACjB,QAAO;CAET,MAAM,MAAM,KAAK,MAAM,GAAG,IAAI;CAC9B,MAAM,YAAY,IAAI,YAAY,IAAI;AACtC,SAAQ,YAAY,MAAM,KAAM,IAAI,MAAM,GAAG,UAAU,GAAG,KAAK,SAAS;;;;;;;;;;;;;;;;AAwB1E,SAAgB,gBACd,KACA,OACA,UAAkC,EAAE,EAC5B;CACR,MAAM,EAAE,aAAa,MAAM,WAAW,iCAAiC;CAEvE,IAAI,IAAI,aAAa,IAAI,QAAQ,qBAAqB,GAAG,GAAG;AAC5D,KAAI,EAAE,QAAQ,uBAAuB,GAAG,CAAC,QAAQ,qBAAqB,GAAG;AACzE,KAAI,EAAE,MAAM;AACZ,KAAI,CAAC,EACH,QAAO;CAGT,MAAM,YAAY,MAAM,MAAM;AAC9B,KAAI,aAAa,EAAE,WAAW,UAAU,CACtC,KAAI,EAAE,MAAM,UAAU,OAAO,CAAC,WAAW;AAE3C,KAAI,oBAAoB,OAAO,EAAE,CAAC,WAAW;AAC7C,KAAI,iBAAiB,OAAO,EAAE;AAC9B,KAAI,uBAAuB,EAAE;AAC7B,KAAI,UAAU,EAAE,MAAM,EAAE,SAAS,CAAC,MAAM;AACxC,KAAI,CAAC,EACH,QAAO;CAGT,MAAM,kBAAkB,cAAc,KAAK,EAAE;CAC7C,MAAM,qBAAqB,gBAAgB,KAAK,MAAM,IAAI,MAAM,WAAW;AAC3E,QAAO,mBAAmB,qBAAqB,IAAI,IAAI;;;;;;ACnIzD,SAAS,WAAW,MAA4B;CAC9C,IAAI,IAAI,SAAS;AACjB,cAAa;AACX,MAAK,IAAI,aAAc;EACvB,IAAI,IAAI,KAAK,KAAK,IAAK,MAAM,IAAK,IAAI,EAAE;AACxC,MAAK,IAAI,KAAK,KAAK,IAAK,MAAM,GAAI,KAAK,EAAE,GAAI;AAC7C,WAAS,IAAK,MAAM,QAAS,KAAK;;;;AAKtC,MAAM,gCAAgB,IAAI,SAAuC;AAEjE,SAAS,WAAW,QAAgC;AAClD,KAAI,OAAO,SAAS,OAAW,QAAO,KAAK,QAAQ;CACnD,IAAI,MAAM,cAAc,IAAI,OAAO;AACnC,KAAI,CAAC,KAAK;AACR,QAAM,WAAW,OAAO,KAAK;AAC7B,gBAAc,IAAI,QAAQ,IAAI;;AAEhC,QAAO,KAAK;;AAId,IAAIO,eAAmC;AACvC,IAAIC,cAAmC;;;;;;AAOvC,SAAgB,YACd,QACA,SAAyB,EAAE,EAC3B,gBACQ;CACR,MAAM,cAAc,OAAO,eAAe;CAC1C,MAAM,OAAO,OAAO,QAAQ;CAC5B,MAAM,OAAO,OAAO,QAAQ;CAC5B,MAAM,oBAAoB,OAAO,qBAAqB;AAGtD,KAAI,cAAc,KAChB,QAAOC,SAAO,OAAO;CAGvB,MAAM,IAAI,OAAO;CACjB,MAAM,IAAI,KAAK,IAAI,OAAO,IAAI,OAAO,GAAG,EAAE;AAG1C,KAAI,CAAC,gBAAgB,aAAa,SAAS,GAAG;AAC5C,iBAAe,IAAI,YAAY,EAAE;AACjC,gBAAc,IAAI,aAAa,EAAE;;CAEnC,MAAM,OAAO;CACb,MAAM,OAAO;CAGb,IAAIC,aAAiC;AACrC,KAAI,sBAAsB,KAAO,gBAAgB,OAC/C,cAAa,IAAI,IAAI,eAAe;CAItC,IAAI,WAAW;AAEf,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,IAAI,IAAI,OAAO;AAGf,MAAI,YAAY,IAAI,EAAE,CACpB,KAAI,IAAI,IAAI,IAAI,oBAAoB,IAAI;AAI1C,OAAK;AAEL,MAAI,WAAW,GAAG;AAEhB,QAAK,YAAY;AACjB,QAAK,YAAY;AACjB;AACA,OAAI,aAAa,EAEf,MAAK,IAAI,KAAK,KAAK,KAAK,GAAG,KAAK,GAAG,IACjC,UAAS,MAAM,MAAM,GAAG,EAAE;aAGrB,IAAI,KAAK,IAAI;AAEtB,QAAK,KAAK;AACV,QAAK,KAAK;AACV,YAAS,MAAM,MAAM,GAAG,EAAE;;;AAM9B,MAAK,IAAI,IAAI,GAAG,IAAI,UAAU,KAAK;EACjC,MAAM,KAAK,KAAK;EAChB,MAAM,KAAK,KAAK;EAChB,IAAI,IAAI,IAAI;AACZ,SAAO,KAAK,KAAK,KAAK,KAAK,IAAI;AAC7B,QAAK,IAAI,KAAK,KAAK;AACnB,QAAK,IAAI,KAAK,KAAK;AACnB;;AAEF,OAAK,IAAI,KAAK;AACd,OAAK,IAAI,KAAK;;CAIhB,MAAM,WAAW,KAAK;CACtB,IAAI,SAAS;AACb,MAAK,IAAI,IAAI,GAAG,IAAI,UAAU,KAAK;EACjC,MAAM,IAAI,KAAK,IAAI,KAAK,KAAK,SAAS;AACtC,OAAK,KAAK;AACV,YAAU;;CAEZ,MAAM,SAAS,IAAI;AACnB,MAAK,IAAI,IAAI,GAAG,IAAI,UAAU,IAC5B,MAAK,MAAM;CAIb,IAAI,iBAAiB;AACrB,KAAI,OAAO,GAAK;EACd,IAAIC,eAAa;AACjB,OAAK,IAAI,IAAI,GAAG,IAAI,UAAU,KAAK;AACjC,mBAAc,KAAK;AACnB,OAAIA,gBAAc,MAAM;AACtB,qBAAiB,IAAI;AACrB;;;EAIJ,IAAI,MAAM;AACV,OAAK,IAAI,IAAI,GAAG,IAAI,gBAAgB,IAAK,QAAO,KAAK;EACrD,MAAM,MAAM,IAAI;AAChB,OAAK,IAAI,IAAI,GAAG,IAAI,gBAAgB,IAAK,MAAK,MAAM;;CAItD,MAAM,IAAI,WAAW,OAAO;CAC5B,IAAI,aAAa;AACjB,MAAK,IAAI,IAAI,GAAG,IAAI,gBAAgB,KAAK;AACvC,gBAAc,KAAK;AACnB,MAAI,KAAK,WAAY,QAAO,KAAK;;AAGnC,QAAO,KAAK,iBAAiB;;;;;AAM/B,SAAgBF,SAAO,KAA2B;CAChD,IAAI,SAAS;CACb,IAAI,SAAS,IAAI;AACjB,MAAK,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,IAC9B,KAAI,IAAI,KAAK,QAAQ;AACnB,WAAS,IAAI;AACb,WAAS;;AAGb,QAAO;;;AAIT,SAAS,SAAS,SAAsB,QAAsB,GAAW,GAAiB;AACxF,QAAO,MAAM;EACX,IAAI,WAAW;EACf,MAAM,OAAO,IAAI,IAAI;EACrB,MAAM,QAAQ,IAAI,IAAI;AACtB,MAAI,OAAO,KAAK,OAAO,QAAQ,OAAO,UAAW,YAAW;AAC5D,MAAI,QAAQ,KAAK,OAAO,SAAS,OAAO,UAAW,YAAW;AAC9D,MAAI,aAAa,EAAG;EAEpB,MAAM,KAAK,QAAQ;AACnB,UAAQ,KAAK,QAAQ;AACrB,UAAQ,YAAY;EACpB,MAAM,KAAK,OAAO;AAClB,SAAO,KAAK,OAAO;AACnB,SAAO,YAAY;AACnB,MAAI;;;;;;AC/DR,MAAM,iBAAiB;AAavB,IAAa,mBAAb,MAA8B;CAC5B,AAAiB;CACjB,AAAiB;CACjB,AAAiB;CACjB,AAAiB;CACjB,AAAiB;;CAEjB,AAAiB;CAEjB,AAAQ,SAAS;CACjB,AAAQ,QAAmB,EAAE;;CAE7B,AAAQ;;CAER,AAAQ,iBAAiB;CACzB,AAAQ,aAA6B;CACrC,AAAQ,WAA2B,EAAE;;;;;;;;CAQrC,AAAQ,qBAAwC;CAChD,AAAQ,cAAc;CACtB,AAAQ,OAA6B;CACrC,AAAQ,SAAS;CAEjB,AAAQ,QAAwB;EAC9B,WAAW;EACX,WAAW;EACX,gBAAgB;EAChB,eAAe;EACf,iBAAiB;EACjB,QAAQ;EACR,iBAAiB;EAClB;CAED,YAAY,MAAqB,UAA2C,EAAE,EAAE;AAC9E,OAAK,OAAO;AACZ,OAAK,YAAY,KAAK,SAAS;AAC/B,MAAI,KAAK,aAAa,EACpB,OAAM,IAAI,MAAM,qEAAqE;AAEvF,OAAK,YAAY,KAAK,SAAS;AAC/B,OAAK,YAAY,KAAK,SAAS;AAC/B,OAAK,QAAQ,IAAI,MAAM,KAAK,UAAU,CAAC,KAAK,KAAK;EACjD,MAAM,WACJ,OAAO,YAAY,cAAc,OAAO,QAAQ,KAAK,qBAAqB,GAAG;AAI/E,OAAK,eACH,QAAQ,uBAAuB,OAAO,SAAS,SAAS,IAAI,YAAY,IAAI,WAAW;AACzF,OAAK,gBAAgB;;;;;;CAOvB,OAAO,QAAgC,UAAwB,EAAE,EAAwB;AACvF,MAAI,KAAK,OACP,QAAO,QAAQ,uBAAO,IAAI,MAAM,6BAA6B,CAAC;AAEhE,OAAK,QAAQ,UAAU,eAAe,MAAM,KAC1C,QAAO,QAAQ,uBACb,IAAI,MAAM,+EAA+E,CAC1F;EAIH,IAAIG,aAAyD;AAC7D,MAAI,QAAQ,WAAW,MAAM;AAC3B,OAAI,CAAC,KAAK,KAAK,SAAS,iBACtB,QAAO,QAAQ,uBACb,IAAI,MACF,2FACD,CACF;AAEH,OAAI;AACF,iBAAa,KAAK,KAAK,eAAe,QAAQ,QAAQ;YAC/C,KAAK;AACZ,WAAO,QAAQ,OAAO,eAAe,QAAQ,MAAM,IAAI,MAAM,OAAO,IAAI,CAAC,CAAC;;;EAG9E,MAAMC,WACJ,OAAO,WAAW,WACd,CACE,GAAI,QAAQ,eACR,CAAC;GAAE,MAAM;GAAmB,SAAS,QAAQ;GAAc,CAAC,GAC5D,EAAE,EACN;GAAE,MAAM;GAAiB,SAAS;GAAQ,CAC3C,GACD,QAAQ,eACN,CAAC;GAAE,MAAM;GAAmB,SAAS,QAAQ;GAAc,EAAE,GAAG,OAAO,GACvE;EACR,MAAM,YAAY,IAAI,YAAY,KAAK,KAAK,WAAW,SAAS,CAAC;EAGjE,MAAM,UAAU,KAAK,YAAY,UAAU,SAAS,iBAAiB;AACrE,MAAI,UAAU,EACZ,QAAO,QAAQ,uBACb,IAAI,MACF,6BAA6B,UAAU,OAAO,uCAAuC,KAAK,YAC3F,CACF;EAEH,MAAM,YAAY,KAAK,IAAI,QAAQ,aAAa,KAAK,QAAQ;AAE7D,SAAO,IAAI,SAAsB,SAAS,WAAW;GACnD,MAAMC,MAAe;IACnB,IAAI,KAAK;IACT;IACA;IACA,eAAe,QAAQ,iBAAiB,EAAE;IAC1C,aAAa,QAAQ,WAAW;IAChC,WAAW,YAAY,MAAM;IAC7B,eAAe,YAAY,UAAU;IACrC,SAAS,QAAQ;IACjB;IACA;IACA,OAAO;IACP,MAAM;IACN,YAAY;IACZ,gBAAgB;IAChB,cAAc,EAAE;IAChB,MAAM;IACN,cAAc;IACd,aAAa,YAAY,KAAK;IAC9B,YAAY;IACZ,cAAc;IACf;AACD,QAAK,MAAM,KAAK,IAAI;AACpB,QAAK,MAAM;AACX,QAAK,MAAM;IACX;;;CAIJ,MAAM,QAAuB;AAC3B,SAAO,KAAK,KAAM,OAAM,KAAK;;;CAI/B,MAAM,QAAuB;AAC3B,OAAK,SAAS;AACd,QAAM,KAAK,OAAO;;CAGpB,WAA2B;AACzB,SAAO,EAAE,GAAG,KAAK,OAAO;;CAG1B,AAAQ,OAAa;AACnB,MAAI,KAAK,KAAM;AACf,OAAK,OAAO,KAAK,KAAK,CAAC,cAAc;AACnC,QAAK,OAAO;AAEZ,OAAI,KAAK,SAAS,CAAE,MAAK,MAAM;IAC/B;;CAGJ,AAAQ,UAAmB;AACzB,SACE,KAAK,MAAM,SAAS,KACpB,KAAK,eAAe,QACpB,KAAK,SAAS,SAAS,KACvB,KAAK,MAAM,MAAM,MAAM,MAAM,KAAK;;CAItC,AAAQ,aAAgF;EACtF,MAAM,OAAO,IAAI,WAAW,KAAK,UAAU;EAC3C,MAAMC,OAA8C,EAAE;AACtD,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,WAAW,KAAK;GACvC,MAAM,MAAM,KAAK,MAAM;AACvB,OAAI,OAAO,IAAI,UAAU,YAAY;AACnC,SAAK,KAAK;AACV,SAAK,KAAK;KAAE,MAAM;KAAG;KAAK,CAAC;;;AAG/B,SAAO;GAAE;GAAM;GAAM;;;CAIvB,MAAc,MAAqB;EACjC,MAAM,UAAU,MAAM,KAAK,KAAK,aAAa;EAC7C,MAAM,KAAK,KAAK,KAAK;EACrB,MAAM,YAAY,YAAY,KAAK;AAGnC,KAAG,YAAY;AACf,OAAK,cAAc;AACnB,OAAK,qBAAqB;AAE1B,MAAI;AACF,UAAO,KAAK,SAAS,EAAE;AACrB,SAAK,OAAO;IAGZ,MAAM,EAAE,MAAM,SAAS,KAAK,YAAY;AACxC,QAAI,KAAK,SAAS,KAAK,KAAK,SAAS,SAAS,gBAAgB;KAC5D,MAAM,eAAe,KAAK,cAAc;AACxC,UAAK,kBAAkB;AACvB,QAAG,4BAA4B,MAAM,aAAa;AAClD,UAAK;AACL,UAAK,SAAS,KAAK;MAAE;MAAc;MAAM,CAAC;AAC1C,UAAK,MAAM;AACX,SAAI,KAAK,SAAS,KAAK,MAAM,gBAC3B,MAAK,MAAM,kBAAkB,KAAK;;AAKtC,QAAI,KAAK,WAAY,OAAM,KAAK,kBAAkB;IAKlD,MAAM,eAAe,KAAK,SAAS,UAAU;IAC7C,MAAM,YAAY,KAAK,WAAW,KAAK,KAAK,eAAe;AAC3D,QAAI,KAAK,SAAS,SAAS,MAAM,gBAAgB,WAC/C,OAAM,KAAK,UAAU;;WAGlB,KAAK;GAEZ,MAAM,QAAQ,eAAe,QAAQ,MAAM,IAAI,MAAM,OAAO,IAAI,CAAC;AACjE,OAAI,KAAK,YAAY,cACnB,KAAI;AACF,OAAG,kBAAkB;WACf;AAIV,QAAK,MAAM,OAAO,CAAC,GAAG,KAAK,OAAO,GAAG,KAAK,MAAM,QAAQ,MAAoB,MAAM,KAAK,CAAC,CACtF,KAAI,IAAI,UAAU,QAAQ;AACxB,QAAI,QAAQ;AACZ,QAAI,OAAO,MAAM;;AAGrB,QAAK,QAAQ,EAAE;AACf,QAAK,MAAM,KAAK,KAAK;AACrB,QAAK,aAAa;AAClB,QAAK,WAAW,EAAE;AAClB,QAAK,iBAAiB;AACtB,SAAM;YACE;AACR,QAAK,MAAM,UAAU,YAAY,KAAK,GAAG;AACzC,YAAS;;;;;;;;;;;;;CAcb,AAAQ,mBAAyB;AAC/B,MAAI,CAAC,KAAK,KAAK,SAAS,iBAAkB;EAC1C,MAAM,MAAM,IAAI,WAAW,KAAK,UAAU,CAAC,KAAK,GAAG;EACnD,IAAI,aAAa;AACjB,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,WAAW,KAAK;GACvC,MAAM,MAAM,KAAK,MAAM;AACvB,OAAI,OAAO,IAAI,UAAU,cAAc,IAAI,aAAa,GAAG;AACzD,QAAI,KAAK,IAAI;AACb,iBAAa;;;EAGjB,MAAM,OAAO,KAAK;AAClB,MAAI,SAAS,MAGX;OAAI,CAAC,WAAY;aACR,IAAI,OAAO,IAAI,MAAM,OAAO,KAAK,GAAG,CAC7C;AAEF,OAAK,KAAK,SAAS,qBAAqB,IAAI;AAC5C,OAAK,qBAAqB;;;;;;;;CAS5B,AAAQ,QAAc;AACpB,MAAI,KAAK,cAAc,KAAK,MAAM,WAAW,EAAG;EAChD,MAAM,OAAO,KAAK,MAAM,QAAQ,KAAK;AACrC,MAAI,SAAS,GAAI;EAEjB,MAAM,MAAM,KAAK,MAAM;EACvB,MAAM,cAAc,IAAI,UAAU,SAAS,IAAI,YAAY;EAC3D,MAAM,eAAe,KAAK,KAAK,cAAc,KAAK,UAAU;EAC5D,MAAM,cAAc,KAAK,KAAK,SAAS;AACvC,MAAI,KAAK,iBAAiB,eAAe,KAAK,gBAAgB,YAG5D;AAGF,OAAK,MAAM,OAAO;AAClB,MAAI,QAAQ;AACZ,MAAI,OAAO;AACX,MAAI,iBAAiB;AACrB,MAAI,aAAa,YAAY,KAAK;AAClC,OAAK,kBAAkB;AACvB,OAAK,MAAM,QAAQ;AACnB,OAAK,aAAa;AAOlB,MAAI,IAAI,eAAe;GACrB,MAAM,EAAE,YAAY,KAAK,KAAK,SAAS,iBAAiB,IAAI,cAAc;AAC1E,OAAI,YAAY,EACd,OAAM,IAAI,MACR,8BAA8B,IAAI,YAAY,4CAC/C;;AAGL,OAAK,KAAK,SAAS,eAAe,KAAK;;;;;;;;;CAUzC,MAAc,mBAAkC;EAC9C,MAAM,MAAM,KAAK;AACjB,MAAI,CAAC,IAAK;EACV,MAAM,KAAK,KAAK,KAAK;EACrB,MAAM,MAAM,KAAK,IAAI,IAAI,aAAa,KAAK,cAAc,IAAI,UAAU,OAAO;EAC9E,MAAM,SAAS,QAAQ,IAAI,UAAU;EACrC,MAAM,QAAQ,IAAI,UAAU,SAAS,IAAI,YAAY,IAAI;EACzD,MAAM,EAAE,WAAW,MAAM,GAAG,QAAQ,OAAO,EAAE,YAAY,QAAQ,CAAC;AAClE,MAAI,aAAa;AACjB,OAAK,MAAM;AACX,MAAI,CAAC,OAAQ;AAEb,KAAG,gBAAgB,IAAI,KAAK;AAC5B,MAAI,IAAI,cAAe,IAAG,kBAAkB;AAC5C,OAAK,aAAa;EAClB,MAAM,QAAQ,KAAK,KAAK,WAAW,YAAY,QAAQ,EAAE,aAAa,GAAG,CAAC,CAAC;AAC3E,KAAG,iBAAiB,IAAI,MAAM,MAAM;AACpC,MAAI,QAAQ;AACZ,MAAI,eAAe,YAAY,KAAK;AACpC,OAAK,aAAa,KAAK,MAAM;;;CAI/B,MAAc,WAA0B;EACtC,MAAM,OAAO,KAAK,SAAS,OAAO;AAClC,MAAI,CAAC,KAAM;EACX,MAAM,SAAS,MAAM,KAAK,KAAK,SAAS,gBAAgB,KAAK,aAAa;AAC1E,OAAK,MAAM,EAAE,MAAM,SAAS,KAAK,MAAM;AAGrC,OAAI,KAAK,MAAM,UAAU,OAAO,IAAI,UAAU,WAAY;AAC1D,QAAK,aAAa,KAAK,OAAO,MAAM;;;;CAKxC,AAAQ,aAAa,KAAc,SAAuB;AACxD,MAAI,aAAa,KAAK,QAAQ;AAC9B,OAAK,MAAM;AACX,MAAI,IAAI,iBAAiB,EAAG,KAAI,eAAe,YAAY,KAAK;EAEhE,MAAM,EAAE,YAAY,eAAe,KAAK;AACxC,MACG,eAAe,QAAQ,YAAY,cACnC,eAAe,QAAQ,YAAY,YACpC;AACA,OAAI,eAAe;AACnB,QAAK,OAAO,IAAI;AAChB;;EAEF,MAAM,QAAQ,KAAK,KAAK,YAAY,QAAQ;AAC5C,MAAI,QAAQ;AACZ,MAAI,IAAI,cAAc,MAAM,MAAM,IAAI,KAAK,SAAS,EAAE,CAAC,EAAE;AACvD,QAAK,MAAM,KAAK,IAAI,eAAe;IACjC,MAAM,MAAM,IAAI,KAAK,QAAQ,EAAE;AAC/B,QAAI,QAAQ,GAAI,KAAI,OAAO,IAAI,KAAK,MAAM,GAAG,IAAI;;AAEnD,OAAI,eAAe;AACnB,QAAK,OAAO,IAAI;AAChB;;AAEF,MAAI,UAAU,OAAO;GACnB,WAAW,IAAI;GACf;GACA,WAAW,IAAI,aAAa;GAC7B,CAAC;AACF,MAAI,IAAI,aAAa,UAAU,IAAI,WAAW;AAC5C,OAAI,eAAe;AACnB,QAAK,OAAO,IAAI;;;;CAKpB,AAAQ,OAAO,KAAoB;AACjC,MAAI,QAAQ;AACZ,MAAI,IAAI,QAAQ,KAAK,KAAK,MAAM,IAAI,UAAU,KAAK;AACjD,QAAK,KAAK,SAAS,gBAAgB,IAAI,KAAK;AAC5C,QAAK,MAAM,IAAI,QAAQ;;AAEzB,OAAK,kBAAkB,IAAI;AAC3B,OAAK,MAAM;EACX,MAAM,aAAa,YAAY,KAAK;AACpC,MAAI,QAAQ;GACV,WAAW,IAAI;GACf,MAAM,IAAI;GACV,UAAU,CAAC,GAAG,IAAI,aAAa;GAC/B,iBAAiB,IAAI,aAAa;GAClC,cAAc,IAAI;GAClB,SAAS,IAAI,eAAe;GAC5B,aAAa,IAAI;GACjB,YAAY,IAAI;GAChB,cAAc,IAAI;GAClB;GACA,QAAQ,IAAI,eAAe,IAAI;GAC/B,OAAO,aAAa,IAAI;GACzB,CAAC;;;;;;;ACnhBN,SAAS,iBAAiB,SAA2D;CACnF,MAAMC,MAAmC,EAAE;CAC3C,MAAM,uBAAO,IAAI,KAAqB;AACtC,MAAK,MAAM,EAAE,QAAQ,YAAY,SAAS;EACxC,MAAM,KAAK,KAAK,IAAI,OAAO,IAAI,KAAK;AACpC,OAAK,IAAI,QAAQ,EAAE;AACnB,MAAI,MAAM,IAAI,SAAS,GAAG,OAAO,GAAG,OAAO;;AAE7C,QAAO;;;;;;AAOT,IAAa,QAAb,MAAoD;CAClD,AAAiB;CACjB,AAAiB;CACjB,AAAiB;CACjB,AAAiB;CAEjB,YAAY,WAA6B,SAA0B;AACjE,MAAI,OAAO,KAAK,QAAQ,QAAQ,CAAC,WAAW,EAC1C,OAAM,IAAI,MAAM,kDAAkD;AAEpE,OAAK,YAAY;AACjB,OAAK,gBAAgB,EAAE,GAAG,QAAQ,SAAS;AAC3C,OAAK,OAAO,QAAQ;AACpB,OAAK,SAAS,QAAQ;;;CAIxB,IAAI,UAAoB;AACtB,SAAO,OAAO,KAAK,KAAK,cAAc;;;;;;;;;;;CAYxC,MAAM,IAAI,OAAe,UAA4B,EAAE,EAAc;EACnE,MAAM,QAAQ,KAAK,OACf,MAAM,KAAK,KAAK,OAAO,KAAK,QAAQ,GACpC,KAAK,QAAQ,KAAK,YAAuB;GAAE;GAAQ,QAAQ;GAAO,EAAE;AACxE,MAAI,MAAM,WAAW,EACnB,OAAM,IAAI,MAAM,oCAAoC;EAEtD,MAAM,UAAU,MAAM,QAAQ,IAC5B,MAAM,IAAI,OAAO,SAAqC;GACpD,MAAM,SAAS,MAAM,KAAK,UAAU,KAAK,QAAQ,KAAK,QAAQ;IAC5D,GAAG;IACH,GAAG,KAAK;IACT,CAAC;AACF,UAAO;IAAE,QAAQ,KAAK;IAAQ;IAAQ;IACtC,CACH;AACD,MAAI,KAAK,OAAQ,QAAO,KAAK,OAAO,QAAQ;AAE5C,SAAO,iBAAiB,QAAQ;;;;;;CAOlC,SACE,QACA,QACA,UAA4B,EAAE,EACR;AACtB,SAAO,KAAK,UAAU,QAAQ,QAAQ,QAAQ;;CAGhD,AAAQ,UACN,QACA,QACA,SACsB;AACtB,MAAI,CAAC,OAAO,OAAO,KAAK,eAAe,OAAO,CAC5C,OAAM,IAAI,MAAM,0BAA0B,OAAO,cAAc,KAAK,QAAQ,KAAK,KAAK,CAAC,GAAG;EAG5F,MAAM,UAAU,KAAK,cAAc,WAAW,OAAO,SAAY;AACjE,SAAO,KAAK,UAAU,OAAO,QAAQ;GAAE,GAAG;GAAS;GAAS,CAAC;;;;;;AC9EjE,MAAa,uBAAuB;AACpC,MAAa,qBAAqB;AAClC,MAAa,yBAAyB;;AAKtC,MAAa,2BAA2B;CAAC;CAAG;CAAG;CAAG;CAAE;;AAEpD,MAAa,oBAAoB;;AAEjC,MAAa,sBAAsB;;AAEnC,MAAa,+BAA+B;CAAC;CAAG;CAAG;CAAG;CAAG;CAAE;AAC3D,MAAa,yBAAyB;;AAEtC,MAAa,aAAa;AAC1B,MAAa,qBAAqB;;AAGlC,SAAgB,sBAAgC;CAC9C,MAAMC,MAAgB,EAAE;AACxB,MAAK,IAAI,IAAI,GAAG,IAAI,mBAAmB,IACrC,KAAI,KAAK,GAAG,yBAAyB;AAEvC,QAAO;;;AA6BT,MAAM,iBAAiB;;;;;;AAOvB,SAAgB,kBAAkB,KAA4C;CAC5E,MAAM,iBAAiB,IAAI;AAC3B,KAAI,CAAC,eACH,OAAM,IAAI,MAAM,uEAAuE;CAEzF,MAAM,aAAc,IAAI,eAA0C,EAAE;CACpE,MAAMC,UAAoB,EAAE;AAC5B,MAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,WAAW,EAAE;EACrD,MAAM,IAAI,eAAe,KAAK,KAAK;AACnC,MAAI,CAAC,EACH,OAAM,IAAI,MAAM,+BAA+B,KAAK,0BAA0B;AAEhF,UAAQ,OAAO,EAAE,GAAG,IAAI;;AAE1B,KAAI,QAAQ,WAAW,EACrB,OAAM,IAAI,MAAM,yCAAyC;AAE3D,MAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,QAAQ,IAClC,KAAI,QAAQ,MAAM,KAChB,OAAM,IAAI,MAAM,yDAAyD,EAAE,GAAG;CAGlF,MAAM,UAAW,IAAI,kBAA6C,EAAE;CACpE,MAAM,MAAO,IAAI,mBAA+C,EAAE;CAClE,MAAM,QAAS,IAAI,SAAqC,EAAE;AAC1D,QAAO;EACL;EACA,YAAY;EACZ,eAAgB,IAAI,mBAA8B;EAClD,aAAa,QAAQ,iBAAiB;EACtC,WAAW,QAAQ,eAAe;EAClC,eAAe,QAAQ,mBAAmB;EAC1C,gBAAgB;GACd,SAAU,IAAI,WAAuB;GACrC,kBAAmB,IAAI,sBAAiC;GACxD,YAAa,IAAI,eAA0B;GAC3C,YAAa,IAAI,eAA0B;GAC5C;EACD,SAAU,MAAM,YAAuB;EACvC,WAAY,MAAM,cAA2B,CAAC,GAAG,yBAAyB;EAC1E,YAAa,MAAM,eAA0B;EAC9C;;;;;;;;;;;AAYH,SAAgB,sBAAsB,cAAwB,KAA6B;CACzF,MAAM,MAAM,IAAI;CAChB,MAAM,IAAI,aAAa;CACvB,IAAI,UAAU;AACd,KAAI,IAAI,WAAW,IAAI,KAAK,IAAI,IAAI,cAAc,IAAI,IAAI,iBACxD,WAAU,KAAK,IAAI,GAAG,KAAK,IAAI,KAAK,KAAK,IAAI,mBAAmB,EAAE,EAAE,IAAI,WAAW,CAAC;CAEtF,MAAM,QAAQ;EAAC,IAAI;EAAa,GAAG;EAAc,IAAI;EAAU;CAC/D,MAAMD,MAAgB,EAAE;AACxB,MAAK,IAAI,IAAI,GAAG,IAAI,UAAU,GAAG,IAC/B,KAAI,KAAK,GAAG,MAAM;AAEpB,KAAI,KAAK,GAAG,OAAO,IAAI,cAAc;AACrC,QAAO;;;;;;;AAQT,SAAgB,oBAAoB,OAAoB,WAAiC;CACvF,MAAM,SAAS,qBAAqB;CACpC,MAAM,SAAS,IAAI,aAAa,sBAAsB,UAAU;AAChE,MAAK,IAAI,IAAI,GAAG,IAAI,qBAAqB,KAAK;EAC5C,MAAM,QAAQ,KAAK,IAAI,GAAG,KAAK,MAAM,OAAO,KAAK,EAAE,CAAC;EACpD,MAAM,OAAO,IAAI;AACjB,OAAK,IAAI,IAAI,GAAG,IAAI,WAAW,IAC7B,QAAO,OAAO,MAAM,MAAM,OAAO,KAAK,SAAS;;AAGnD,QAAO;;AAkBT,SAAgB,4BACd,gBACA,OACA,mBACA,SACY;CACZ,MAAM,cAAc,eAAe;CACnC,MAAM,aAAa,eAAe;CAClC,MAAM,YAAY,eAAe;CACjC,MAAM,eAAgB,eAAe,uBAAkC;CACvE,MAAM,oBAAoB,eAAe;CACzC,MAAM,iBAAkB,eAAe,2BAAsC;CAC7E,MAAM,eAAgB,eAAe,gBAA2B;CAChE,MAAM,WAAY,eAAe,YAAuB,KAAK,MAAM,cAAc,UAAU;CAE3F,MAAM,QAAQ,YAAY;CAC1B,MAAM,SAAS,eAAe;CAM9B,MAAM,aAAa,eAAe;CAClC,MAAM,YACH,YAAY,cAA0B,eAAe,cAAyB;CACjF,MAAM,sBACH,YAAY,yBACZ,eAAe,yBAChB;CACF,MAAM,WAAW,KAAK,MAAM,WAAW,oBAAoB;CAE3D,MAAM,aAAc,eAAe,eAA4B,EAAE;AACjE,MAAK,IAAI,IAAI,GAAG,IAAI,YAAY,IAC9B,KAAI,WAAW,SAAS,KAAK,WAAW,OAAO,iBAC7C,OAAM,IAAI,MACR,4DAA4D,EAAE,OAAO,WAAW,GAAG,IACpF;CAOL,MAAME,SAA0B;EAC9B;EACA;EACA;EACA;EACA;EACA;EACA,YAAY;EACZ;EACA;EACA,WAAW;EACX;EACA;EACA,WAAW;EACX,QAAQ;EACR,kBAAkB;EACnB;CACD,MAAMC,eAAkC;EAAE,MAAM;EAAM,QAAQ;EAAO,KAAK;EAAO;CAEjF,MAAMC,UAAsC,EAAE;CAC9C,MAAMC,QAAkB,EAAE;CAC1B,MAAMC,iBAA2B,EAAE;CACnC,MAAM,aAAa,SAA2B;AAC5C,UAAQ,KAAK,QAAQ;;CAEvB,MAAM,WAAW,SAAuB;AACtC,QAAM,KAAK,KAAK;AAChB,iBAAe,KAAK,KAAK,GAAG;;CAG9B,MAAM,QAAQ,UAAU;CACxB,MAAM,YAAY,qBAAqB;CAEvC,MAAM,eACJ,IACA,iBACA,YACA,aACA,YACA,GACA,MACS;AACT,MAAI,OAAO;GACT,MAAM,KAAK,GAAG,WAAW;GACzB,MAAM,UAAU,GAAG,WAAW;GAC9B,MAAM,SAAS,GAAG,WAAW;GAC7B,MAAM,gBAAgB,IAAI;AAC1B,aAAU;IACR,MAAM;IACN,OAAO,CAAC,KAAK,KAAK,gBAAgB,EAAE,CAAC;IACrC,OAAO;IACP,SAAS;IACV,CAAC;AACF,aAAU;IACR,MAAM;IACN,OAAO,CAAC,KAAK,KAAK,gBAAgB,UAAU,CAAC;IAC7C,OAAO;IACP,SAAS;IACV,CAAC;AACF,aAAU;IACR,MAAM;IACN,OAAO,CAAC,KAAK,KAAK,gBAAgB,UAAU,CAAC;IAC7C,OAAO;IACP,SAAS;IACV,CAAC;AACF,WAAQ;IACN;IACA,QAAQ;IACR,QAAQ;KAAC;KAAiB;KAAI;KAAS;KAAO;IAC9C,SAAS,CAAC,WAAW;IACrB,YAAY;KAAE,UAAU;KAAiB;KAAG;KAAG,YAAY;KAAW;IACvE,CAAC;AACF;;AAEF,YAAU;GACR,MAAM;GACN,OAAO;GACP,OAAO;GACP,SAAS;GACT,gBAAgB;GACjB,CAAC;AACF,UAAQ;GACN;GACA,QAAQ;GACR,QAAQ,CAAC,iBAAiB,WAAW;GACrC,SAAS,CAAC,WAAW;GACrB,YAAY;IAAE,UAAU;IAAiB;IAAG;IAAG;GAChD,CAAC;;AAIJ,WAAU;EACR,MAAM;EACN,OAAO,CAAC,KAAK,YAAY;EACzB,OAAO;EACP,SAAS;EACV,CAAC;CAGF,MAAM,uBAAuB,GAAW,QAAgB,aAA6B;EACnF,MAAM,OAAO,GAAG,OAAO;EACvB,MAAM,UAAU,GAAG,OAAO;EAC1B,MAAM,OAAO,GAAG,OAAO;EACvB,MAAM,OAAO,GAAG,OAAO;AACvB,YAAU;GAAE,MAAM;GAAM,OAAO,CAAC,KAAK,MAAM;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AACnF,YAAU;GAAE,MAAM;GAAS,OAAO,CAAC,KAAK,MAAM;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AACtF,YAAU;GAAE,MAAM;GAAM,OAAO,CAAC,KAAK,OAAO;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AACpF,YAAU;GAAE,MAAM;GAAM,OAAO,CAAC,KAAK,OAAO;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AAEpF,cACE,GAAG,OAAO,UACV,UACA,eAAe,MAAM,EAAE,EACvB,CAAC,OAAO,YAAY,EACpB,MACA,aACA,MACD;AACD,cACE,GAAG,OAAO,aACV,UACA,eAAe,SAAS,EAAE,EAC1B,CAAC,OAAO,YAAY,EACpB,SACA,aACA,MACD;AACD,cACE,GAAG,OAAO,UACV,UACA,eAAe,MAAM,EAAE,EACvB,CAAC,QAAQ,YAAY,EACrB,MACA,aACA,OACD;AACD,cACE,GAAG,OAAO,UACV,UACA,eAAe,MAAM,EAAE,EACvB,CAAC,QAAQ,YAAY,EACrB,MACA,aACA,OACD;EAED,MAAM,SAAS,eAAe,MAAM,EAAE;EACtC,MAAM,SAAS,eAAe,MAAM,EAAE;AACtC,YAAU;GACR,MAAM;GACN,OAAO,CAAC,SAAS;GACjB,OAAO;GACP,SAAS;GACT,gBAAgB;GACjB,CAAC;AACF,YAAU;GACR,MAAM;GACN,OAAO,CAAC,SAAS;GACjB,OAAO;GACP,SAAS;GACT,gBAAgB;GACjB,CAAC;EACF,MAAM,UAAU,GAAG,OAAO;EAC1B,MAAM,UAAU,GAAG,OAAO;AAC1B,YAAU;GAAE,MAAM;GAAS,OAAO,CAAC,KAAK,MAAM;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AACtF,YAAU;GAAE,MAAM;GAAS,OAAO,CAAC,KAAK,OAAO;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AACvF,UAAQ;GACN,IAAI,GAAG,OAAO;GACd,QAAQ;GACR,QAAQ,CAAC,MAAM,OAAO;GACtB,SAAS,CAAC,QAAQ;GAClB,YAAY;IAAE,aAAa;IAAU,KAAK;IAAc,gBAAgB;IAAM;GAC/E,CAAC;AACF,UAAQ;GACN,IAAI,GAAG,OAAO;GACd,QAAQ;GACR,QAAQ,CAAC,MAAM,OAAO;GACtB,SAAS,CAAC,QAAQ;GAClB,YAAY;IAAE,aAAa;IAAU,KAAK;IAAc,gBAAgB;IAAM;GAC/E,CAAC;AAEF,UAAQ;GACN,IAAI,GAAG,OAAO;GACd,QAAQ;GACR,QAAQ,CAAC,SAAS,QAAQ;GAC1B,SAAS,CAAC,SAAS,QAAQ;GAC3B,YAAY;IACV;IACA,aAAa;IACb;IACA;IACA;IACA,gBAAgB;IAChB,gBAAgB;IACjB;GACF,CAAC;EAEF,MAAMC,eAAsB,WAAW;EACvC,MAAM,SAAS,GAAG,OAAO;EACzB,MAAM,SAAS,GAAG,OAAO;AACzB,YAAU;GAAE,MAAM;GAAQ,OAAO,CAAC,SAAS,OAAO;GAAE,OAAO;GAAc,SAAS;GAAY,CAAC;AAC/F,YAAU;GAAE,MAAM;GAAQ,OAAO,CAAC,SAAS,OAAO;GAAE,OAAO;GAAc,SAAS;GAAY,CAAC;AAC/F,UAAQ;GACN,IAAI,GAAG,OAAO;GACd,QAAQ;GACR,QAAQ,CAAC,QAAQ;GACjB,SAAS,CAAC,OAAO;GACjB,YAAY;IAAE,OAAO;IAAQ,UAAU;IAAS;GACjD,CAAC;AACF,UAAQ;GACN,IAAI,GAAG,OAAO;GACd,QAAQ;GACR,QAAQ,CAAC,KAAK;GACd,SAAS,CAAC,OAAO;GACjB,YAAY;IAAE,OAAO;IAAQ,UAAU;IAAM;GAC9C,CAAC;EAEF,MAAM,UAAU,GAAG,OAAO;AAC1B,YAAU;GAAE,MAAM;GAAS,OAAO,CAAC,KAAK,MAAM;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AACtF,UAAQ;GACN,IAAI,GAAG,OAAO;GACd,QAAQ;GACR,QAAQ;IAAC;IAAS;IAAQ;IAAO;GACjC,SAAS,CAAC,QAAQ;GAClB,YAAY;IACV,aAAa;IACb,aAAa;IACb;IACA;IACA,QAAQ;IACR,aAAa;IACd;GACF,CAAC;EAEF,MAAM,WAAW,GAAG,OAAO;AAC3B,YAAU;GAAE,MAAM;GAAU,OAAO,CAAC,KAAK,MAAM;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AACvF,UAAQ;GACN,IAAI,GAAG,OAAO;GACd,QAAQ;GACR,QAAQ,CAAC,SAAS,QAAQ;GAC1B,SAAS,CAAC,SAAS;GACnB,YAAY,EAAE,cAAc,SAAS;GACtC,CAAC;EAEF,MAAM,WAAW,GAAG,OAAO;AAC3B,YAAU;GAAE,MAAM;GAAU,OAAO,CAAC,KAAK,YAAY;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AAC7F,cACE,GAAG,OAAO,UACV,UACA,eAAe,MAAM,EAAE,EACvB,CAAC,aAAa,MAAM,EACpB,UACA,OACA,YACD;AACD,SAAO;;CAIT,IAAI,aAAa;CACjB,IAAIC,kBAAiC;AAErC,MAAK,IAAI,IAAI,GAAG,IAAI,YAAY,KAAK;EACnC,MAAM,SAAS,QAAQ;EAEvB,IAAIC;AACJ,MAAI,gBACF,YAAW;OACN;GACL,MAAM,iBAAiB,eAAe,eAAe,EAAE;AACvD,aAAU;IACR,MAAM;IACN,OAAO,CAAC,YAAY;IACpB,OAAO;IACP,SAAS;IACT,gBAAgB;IACjB,CAAC;AACF,cAAW,GAAG,OAAO;AACrB,aAAU;IAAE,MAAM;IAAU,OAAO,CAAC,KAAK,YAAY;IAAE,OAAO;IAAO,SAAS;IAAc,CAAC;AAC7F,WAAQ;IACN,IAAI,GAAG,OAAO;IACd,QAAQ;IACR,QAAQ,CAAC,YAAY,eAAe;IACpC,SAAS,CAAC,SAAS;IACnB,YAAY;KAAE;KAAa,KAAK;KAAc,gBAAgB;KAAY;IAC3E,CAAC;;EAGJ,MAAM,aAAa,oBAAoB,GAAG,QAAQ,SAAS;EAE3D,MAAM,YAAY,GAAG,OAAO;EAC5B,MAAM,YAAY,eAAe,kBAAkB,EAAE;EACrD,MAAM,WAAW,GAAG,OAAO;AAC3B,YAAU;GAAE,MAAM;GAAW,OAAO,CAAC,KAAK,YAAY;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AAC9F,YAAU;GACR,MAAM;GACN,OAAO,CAAC,YAAY;GACpB,OAAO;GACP,SAAS;GACT,gBAAgB;GACjB,CAAC;AACF,YAAU;GAAE,MAAM;GAAU,OAAO,CAAC,KAAK,YAAY;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AAC7F,UAAQ;GACN,IAAI,GAAG,OAAO;GACd,QAAQ;GACR,QAAQ;IAAC;IAAY;IAAY;IAAU;GAC3C,SAAS,CAAC,WAAW,SAAS;GAC9B,YAAY;IAAE;IAAa,KAAK;IAAc,gBAAgB;IAAY;GAC3E,CAAC;EAEF,MAAM,QAAQ,eAAe,SAAS,EAAE;EACxC,MAAM,MAAM,eAAe,OAAO,EAAE;EACpC,MAAM,QAAQ,eAAe,SAAS,EAAE;EACxC,MAAM,aAAa,GAAG,OAAO;EAC7B,MAAM,QAAQ,GAAG,OAAO;EACxB,MAAM,YAAY,GAAG,OAAO;EAC5B,MAAM,SAAS,GAAG,OAAO;AACzB,YAAU;GACR,MAAM;GACN,OAAO,CAAC,KAAK,kBAAkB;GAC/B,OAAO;GACP,SAAS;GACV,CAAC;AACF,YAAU;GACR,MAAM;GACN,OAAO,CAAC,KAAK,kBAAkB;GAC/B,OAAO;GACP,SAAS;GACV,CAAC;AACF,YAAU;GACR,MAAM;GACN,OAAO,CAAC,KAAK,kBAAkB;GAC/B,OAAO;GACP,SAAS;GACV,CAAC;AACF,YAAU;GAAE,MAAM;GAAQ,OAAO,CAAC,KAAK,YAAY;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AAC3F,cACE,GAAG,OAAO,QACV,UACA,OACA,CAAC,mBAAmB,YAAY,EAChC,YACA,aACA,kBACD;AACD,cACE,GAAG,OAAO,MACV,UACA,KACA,CAAC,mBAAmB,YAAY,EAChC,OACA,aACA,kBACD;AACD,UAAQ;GACN,IAAI,GAAG,OAAO;GACd,QAAQ;GACR,QAAQ,CAAC,YAAY,MAAM;GAC3B,SAAS,CAAC,UAAU;GACpB,YAAY,EAAE,cAAc,YAAY;GACzC,CAAC;AACF,cACE,GAAG,OAAO,QACV,WACA,OACA,CAAC,aAAa,kBAAkB,EAChC,QACA,mBACA,YACD;EAED,MAAM,YAAY,GAAG,OAAO;AAC5B,YAAU;GAAE,MAAM;GAAW,OAAO,CAAC,KAAK,YAAY;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AAC9F,MAAI,IAAI,aAAa,GAAG;GACtB,MAAM,YAAY,eAAe,eAAe,IAAI,EAAE;AACtD,aAAU;IACR,MAAM;IACN,OAAO,CAAC,YAAY;IACpB,OAAO;IACP,SAAS;IACT,gBAAgB;IACjB,CAAC;GACF,MAAM,eAAe,QAAQ,IAAI,EAAE;AACnC,aAAU;IACR,MAAM;IACN,OAAO,CAAC,KAAK,YAAY;IACzB,OAAO;IACP,SAAS;IACV,CAAC;AACF,WAAQ;IACN,IAAI,GAAG,OAAO;IACd,QAAQ;IACR,QAAQ;KAAC;KAAW;KAAQ;KAAU;IACtC,SAAS,CAAC,WAAW,aAAa;IAClC,YAAY;KAAE;KAAa,KAAK;KAAc,gBAAgB;KAAW;IAC1E,CAAC;AACF,qBAAkB;SACb;AACL,WAAQ;IACN,IAAI,GAAG,OAAO;IACd,QAAQ;IACR,QAAQ,CAAC,WAAW,OAAO;IAC3B,SAAS,CAAC,UAAU;IACpB,YAAY;KAAE,cAAc;KAAW;KAAa;IACrD,CAAC;AACF,qBAAkB;;AAEpB,eAAa;;AAIf,WAAU;EACR,MAAM,eAAe;EACrB,OAAO,CAAC,YAAY;EACpB,OAAO;EACP,SAAS;EACT,gBAAgB,eAAe;EAChC,CAAC;AACF,WAAU;EACR,MAAM;EACN,OAAO,CAAC,KAAK,YAAY;EACzB,OAAO;EACP,SAAS;EACV,CAAC;AACF,SAAQ;EACN,IAAI;EACJ,QAAQ;EACR,QAAQ,CAAC,YAAY,eAAe,WAAW;EAC/C,SAAS,CAAC,iBAAiB;EAC3B,YAAY;GAAE;GAAa,KAAK;GAAc,gBAAgB;GAAY;EAC3E,CAAC;AAKF,WAAU;EAAE,MAAM;EAAU,OAAO,CAAC,GAAG,YAAY;EAAE,OAAO;EAAO,SAAS;EAAc,CAAC;AAC3F,SAAQ;EACN,IAAI;EACJ,QAAQ;EACR,QAAQ,CAAC,iBAAiB;EAC1B,SAAS,CAAC,SAAS;EACnB,YAAY,EAAE,OAAO,aAAa;EACnC,CAAC;AAEF,QAAO;EACL,cAAc;EACd;EACA;EACA;EACA;EACA;EACA,QAAQ,CAAC,gBAAgB;EACzB,SAAS,CAAC,SAAS;EACpB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AC9qBH,MAAM,4BAA4B;;;;;;;AAOlC,MAAM,+BAA+B;;AAErC,MAAMC,uBAAqB;;AAE3B,MAAM,gBAAgB;;AAqEtB,SAASC,qBACP,OACA,SACyD;CACzD,MAAM,sBAAM,IAAI,KAAyD;AACzE,MAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,MAAM,QAAQ,EAAE;AACxD,MAAI,KAAK,YAAY,WACnB;EAEF,MAAM,IACJ,QAAQ,IAAI,KAAK,KAAK,KAAK,iBAAiB,QAAQ,IAAI,KAAK,eAAe,GAAG;AACjF,MAAI,EACF,KAAI,IAAI,MAAM,EAAE;;AAGpB,QAAO;;;AAIT,SAAS,IAAI,GAAmB;CAC9B,MAAM,OAAO,IAAI,IAAI,KAAK;CAC1B,MAAM,KAAK,KAAK,IAAI,EAAE;CACtB,MAAM,IAAI,KAAK,IAAI,WAAY;AAM/B,QAAO,QAJL,QACI,cAAc,IAAI,eAAe,IAAI,eAAe,IAAI,cAAe,IAAI,cAC7E,IACA,KAAK,IAAI,CAAC,KAAK,GAAG;;;AAKxB,SAAS,KAAK,GAAmB;AAC/B,QAAO,KAAM,KAAK,IAAI,IAAI,IAAI,KAAK,MAAM;;;AAI3C,SAAS,OACP,GACA,GACA,GACA,QACA,OACA,KACM;AACN,MAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,KAAK;EAC/B,IAAI,MAAM,IAAI,EAAE,KAAK;EACrB,MAAM,OAAO,IAAI;AACjB,OAAK,IAAI,IAAI,GAAG,IAAI,OAAO,IACzB,QAAO,EAAE,OAAO,KAAK,EAAE;AAEzB,MAAI,KAAK;;;AAIb,IAAa,YAAb,MAAa,UAAU;CACrB,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;;CAER,AAAQ;;CAER,AAAQ,iCAAiB,IAAI,KAAuB;CACpD,AAAQ,aAAa;;CAGrB,AAAS;CAGT,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CAER,AAAQ,YACN,KACA,QACA,WACA,OACA;AACA,OAAK,MAAM;AACX,OAAK,SAAS;AACd,OAAK,YAAY,OAAO;AACxB,OAAK,MAAM,kBAAkB,OAAO,aAAa;AACjD,OAAK,YAAY;AACjB,OAAK,SAAS,KAAK,IAAI,eAAe;AACtC,OAAK,QAAQ;EAGb,MAAM,UAAU,OAAO;EACvB,MAAM,QAAQ,QAA8B;GAC1C,MAAM,IAAI,QAAQ,IAAI,IAAI;AAC1B,OAAI,CAAC,EACH,OAAM,IAAI,MAAM,kCAAkC,MAAM;AAE1D,UAAO,EAAE;;EAEX,MAAM,IAAI,KAAK,IAAI,WAAW;AAC9B,MAAI,MAAM,oBACR,OAAM,IAAI,MAAM,kBAAkB,oBAAoB,2BAA2B,EAAE,GAAG;AAExF,OAAK,iBAAiB,EAAE;AACxB,OAAK,QAAQ,EAAE;AACf,OAAK,QAAQ,EAAE;AACf,OAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,QAAK,eAAe,KAAK,KAAK,oBAAoB,EAAE,SAAS,CAAC;AAC9D,QAAK,MAAM,KAAK,KAAK,kBAAkB,EAAE,SAAS,CAAC;AACnD,QAAK,MAAM,KAAK,KAAK,kBAAkB,EAAE,OAAO,CAAC;;AAEnD,OAAK,SAAS,KAAK,4BAA4B;AAC/C,OAAK,SAAS,KAAK,0BAA0B;AAC7C,OAAK,SAAS,KAAK,4BAA4B;AAC/C,OAAK,SAAS,KAAK,0BAA0B;AAC7C,OAAK,kBAAkB,KAAK,oBAAoB,CAAC;AACjD,OAAK,QAAQ,KAAK,mBAAmB;AACrC,OAAK,QAAQ,KAAK,iBAAiB,CAAC;EAGpC,MAAM,QAAQ,4BACZ,KAAK,IAAI,gBACT,KAAK,UAAU,OAAO,OAAO,OAC9B;AACD,MAAI,KAAK,UAAU,KACjB,sBAAqB,OAAO,OAAO,gBAAgB;AAErD,OAAK,eAAe,IAAI,SAAS,KAAK,OAAO;GAAE;GAAW,QAAQ;GAAO,CAAC;AAC1E,OAAK,aAAa,iBAAiBA,qBAAmB,OAAO,OAAO,gBAAgB,CAAC;AACrF,OAAK,aAAa,gBAAgB;;CAGpC,aAAa,OAAO,UAA4B,EAAE,EAAsB;EACtE,MAAM,MAAM,MAAM,SAAS;EAC3B,MAAM,SAAS,MAAM,cAAc;GACjC,MAAM,QAAQ;GACd,WAAW,QAAQ;GACnB,UAAU,QAAQ;GAClB,SAAS,QAAQ;GACjB,UAAU,QAAQ;GAClB,YAAY,QAAQ;GACrB,CAAC;EACF,MAAM,SAAU,OAAO,UAAU,2BAAsC;AAEvE,SAAO,IAAI,UAAU,KAAK,QADR,KAAK,IAAI,QAAQ,aAAa,MAAM,OAAO,EAChB,QAAQ,SAAS,KAAK;;;CAIrE,IAAI,aAAqB;AACvB,SAAO,KAAK,IAAI;;;CAIlB,kBAA4B;AAC1B,SAAO,EAAE;;;CAIX,AAAQ,gBAAgB,KAA6B;EACnD,MAAM,IAAI,KAAK;EACf,MAAM,QAAQ,KAAK,OAAO;EAC1B,MAAM,MAAM,IAAI,aAAa,IAAI,SAAS,EAAE;EAC5C,MAAM,OAAO,IAAI,SAAS,IAAI,OAAO;AACrC,OAAK,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;GACnC,MAAM,UAAU,IAAI,KAAK;GACzB,MAAM,UAAU,IAAI;AACpB,QAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IAErB,MAAK,WAAW,UAAU,KAAK,GAAG,MAAM,UAAU,MAAM,IAAI,KAAK;;AAGrE,SAAO;;;;;;;CAQT,AAAQ,gBAAgB,OAAkC;EACxD,MAAM,IAAI,KAAK;EACf,MAAM,MAAM,KAAK,IAAI;EACrB,MAAM,IAAI,IAAI,aAAa,sBAAsB,IAAI;AACrD,OAAK,IAAI,IAAI,GAAG,IAAI,qBAAqB,KAAK;GAC5C,MAAM,QAAQ,KAAK,eAAe;GAClC,MAAM,OAAO,MAAM,KAAK;AACxB,KAAE,IAAI,MAAM,SAAS,MAAM,OAAO,IAAI,EAAE,IAAI,IAAI;;EAElD,MAAM,KAAK,IAAI,aAAa,EAAE;AAC9B,SAAO,KAAK,QAAQ,KAAK,QAAQ,GAAG,GAAG,EAAE,QAAQ,GAAG;AACpD,OAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IACrB,IAAG,KAAK,KAAK,GAAG,GAAG;EAErB,MAAM,KAAK,IAAI,aAAa,EAAE;AAC9B,SAAO,KAAK,QAAQ,KAAK,QAAQ,IAAI,GAAG,GAAG,GAAG;EAE9C,IAAI,OAAO;AACX,OAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IACrB,SAAQ,GAAG;AAEb,UAAQ;EACR,IAAI,WAAW;AACf,OAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;GAC1B,MAAM,IAAI,GAAG,KAAK;AAClB,eAAY,IAAI;;AAElB,cAAY;EACZ,MAAM,MAAM,KAAK,kBAAkB,KAAK,KAAK,WAAW,cAAc;AACtE,OAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IACrB,IAAG,MAAM,GAAG,KAAK,QAAQ;AAE3B,SAAO;;;CAIT,AAAQ,gBAAgB,QAA8B;EACpD,IAAI,MAAM,KAAK;AACf,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,IAC/B,QAAO,KAAK,MAAM,KAAK,OAAO;AAEhC,SAAO,KAAK,IAAI,KAAK,IAAI,CAAC,IAAI;;;;;;;CAQhC,AAAQ,YACN,QACA,GACA,cACa;EACb,MAAM,QAAQ,IAAI,YAAY,oBAAoB;AAClD,OAAK,IAAI,IAAI,GAAG,IAAI,qBAAqB,KAAK;GAC5C,MAAM,IAAI,KAAK,IAAI,WAAW;GAC9B,MAAM,IAAI,KAAK,MAAM;GACrB,MAAM,IAAI,KAAK,MAAM;GACrB,MAAM,SAAS,IAAI,aAAa,EAAE;AAClC,QAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;IAC1B,IAAI,MAAM,EAAE;IACZ,MAAM,OAAO,IAAI,KAAK;AACtB,SAAK,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,IAC/B,QAAO,EAAE,OAAO,KAAK,OAAO;AAE9B,WAAO,KAAK;;AAEd,OAAI,EAAE,sBAAsB,KAAO,aACjC,MAAK,MAAM,QAAQ,cAAc;IAC/B,MAAM,MAAM,KAAK;AACjB,WAAO,OACL,OAAO,OAAO,IAAI,OAAO,OAAO,EAAE,oBAAoB,OAAO,OAAO,EAAE;;AAG5E,OAAI,EAAE,gBAAgB,EACpB,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IACrB,QAAO,MAAM,EAAE;AAGnB,OAAI,EAAE,OAAO,KAAK,EAAE,OAAO,GAAG;IAE5B,MAAM,YADS,CAAC,GAAG,OAAO,CAAC,MAAM,GAAG,OAAO,KAAK,EAAE,CACzB,EAAE,OAAO;AAClC,SAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IACrB,KAAI,OAAO,KAAK,UACd,QAAO,KAAK,OAAO;;GAKzB,IAAI,OAAO,OAAO;AAClB,QAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IACrB,KAAI,OAAO,KAAK,KACd,QAAO,OAAO;GAGlB,IAAI,MAAM;AACV,QAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;IAC1B,MAAM,IAAI,KAAK,IAAI,OAAO,KAAK,KAAK;AACpC,WAAO,KAAK;AACZ,WAAO;;GAET,IAAIC,SAAO,KAAK,QAAQ,GAAG;GAC3B,IAAI,SAAS,IAAI;AACjB,QAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,cAAQ,OAAO;AACf,QAAIA,UAAQ,GAAG;AACb,cAAS;AACT;;;AAGJ,SAAM,KAAK;;AAEb,SAAO;;;;;;;;;;CAWT,MAAM,MAAM,MAAc,OAA2B,EAAE,EAA8B;AACnF,MAAI,KAAK,WACP,OAAM,IAAI,MAAM,gCAAgC;EAElD,MAAM,cAAc,KAAK,eAAe;EACxC,MAAM,OAAO,KAAK,QAAQ;EAC1B,MAAM,gBAAgB,KAAK,iBAAiB;EAC5C,MAAM,oBAAoB,KAAK,qBAAqB;EACpD,MAAM,mBAAmB,KAAK,oBAAoB;EAIlD,MAAM,SAAS,sBADC,KAAK,UAAU,OAAO,KAAK,EACG,KAAK,IAAI;AACvD,MAAI,OAAO,SAAS,KAAK,KAAK,UAC5B,OAAM,IAAI,MAAM,oBAAoB,OAAO,OAAO,uBAAuB,KAAK,UAAU,GAAG;EAE7F,MAAM,YAAY,KAAK,IACrB,KAAK,aAAaF,sBAClB,KAAK,YAAY,OAAO,SAAS,EAClC;EAGD,MAAM,YAAY,YAAY,KAAK;AACnC,OAAK,aAAa,OAAO;AACzB,OAAK,aAAa,WAAW,iBAAiB,KAAK,gBAAgB,OAAO,CAAC;EAG3E,IAAI,EAAE,QAAQ,WAAW,MAAM,KAAK,aAAa,QAAQ,IAAI,YAAY,OAAO,OAAO,CAAC;EAGxF,MAAMG,SAAwB,CAC5B,KAAK,YAAY,QAAQ;GAAE;GAAa;GAAM;GAAmB,EAAE,KAAK,CACzE;EACD,MAAM,WAAW,IAAI,oBAAoB,MAAM,KAAK,QAAQ;EAC5D,MAAM,aAAa;GAAE;GAAa;GAAM;GAAmB;AAC3D,OAAK,IAAI,OAAO,GAAG,OAAO,WAAW,QAAQ;GAC3C,MAAM,aAAa,KAAK,gBAAgB,OAAO,OAAO,SAAS,GAAG;AAClE,QAAK,aAAa,WAAW,iBAAiB,WAAW;AAEzD,aADY,MAAM,KAAK,aAAa,QAAQ,IAAI,YAAY,EAAE,CAAC,EAClD;AAGb,OAAI,KAAK,gBAAgB,OAAO,GAAG,cACjC;GAEF,MAAM,SACJ,sBAAsB,IAClB,OAAO,MAAM,mBAAmB,IAAI,CAAC,mBAAmB,EAAE,GAC1D;AACN,UAAO,KAAK,KAAK,YAAY,QAAQ,YAAY,OAAO,CAAC;AACzD,SAAM,SAAS,UAAU,QAAQ,MAAM;;AAIzC,QAAM,SAAS,UAAU,QAAQ,KAAK;EACtC,MAAM,MAAM,SAAS,UAAU,OAAO,OAAO;EAE7C,MAAM,aAAa,YAAY,KAAK,GAAG,aAAa;EACpD,MAAM,eAAgB,OAAO,SAAS,aAAc,KAAK,IAAI;AAC7D,UAAQ,IACN,mBAAmB,OAAO,OAAO,SAAS,aAAa,QAAQ,EAAE,CAAC,SACxD,UAAU,QAAQ,EAAE,CAAC,SAAS,eAAe,WAAW,QAAQ,EAAE,CAAC,GAC9E;AACD,SAAO;GAAE;GAAK,YAAY,KAAK,IAAI;GAAY,QAAQ,OAAO;GAAQ;GAAc;;;;;;;;CAStF,MAAM,iBAAiB,QAAuB,MAAc,IAAmC;EAC7F,MAAM,WAAW,KAAK,IAAI,GAAG,OAAO,6BAA6B;EACjE,MAAM,YAAY,KAAK;EAEvB,MAAM,QAAQ,IAAI,YAAY,sBAAsB,UAAU;AAC9D,OAAK,IAAI,IAAI,GAAG,IAAI,WAAW,KAAK;GAClC,MAAM,QAAQ,OAAO,WAAW;AAChC,QAAK,IAAI,IAAI,GAAG,IAAI,qBAAqB,IACvC,OAAM,IAAI,YAAY,KAAK,MAAM;;EAGrC,MAAM,SAAS,oBAAoB,OAAO,UAAU;EACpD,MAAM,MAAM,MAAM,KAAK,mBAAmB,QAAQ,UAAU;EAC5D,MAAM,QAAQ,OAAO,YAAY;EACjC,MAAM,QAAQ,KAAK,QAAQ;EAC3B,MAAM,MAAM,IAAI,MAAM,MAAM,OAAO,KAAK;AACxC,OAAK,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,IAC9B,KAAI,KAAK,KAAK,IAAI,GAAG,KAAK,IAAI,IAAI,IAAI,GAAG,CAAC;AAE5C,SAAO;;;CAIT,MAAc,mBAAmB,QAAsB,WAA0C;EAG/F,MAAM,SAAS,cAAc,4BAA4B;EACzD,IAAI,OAAO,SAAS,KAAK,eAAe,IAAI,UAAU,GAAG;EACzD,IAAIC;AACJ,MAAI,KACF,gBAAe,KAAK;OACf;GACL,MAAM,QAAQ,8BAA8B;IAC1C,WAAW;IACX,UAAU;KACR,eAAe;KACf,UAAU;KACX;IACD,aAAa;IACd,CAAC;AACF,UAAO,IAAI,SAAS,KAAK,KAAK,OAAO;IAAE,WAAW;IAAW,QAAQ;IAAO,CAAC;AAC7E,QAAK,iBAAiBH,qBAAmB,OAAO,KAAK,OAAO,aAAa,CAAC;AAC1E,QAAK,gBAAgB;AACrB,kBAAe,MAAM;AACrB,OAAI,OACF,MAAK,eAAe,IAAI,WAAW,KAAK;;AAG5C,MAAI;AACF,QAAK,OAAO;AACZ,QAAK,WAAW,aAAa,OAAO;AACpC,UAAO,MAAM,KAAK,eAAe,aAAa,IAAI,YAAY,WAAW;YACjE;AACR,OAAI,CAAC,OACH,MAAK,SAAS;;;CAKpB,UAAgB;AACd,MAAI,KAAK,WACP;AAEF,OAAK,aAAa;AAClB,OAAK,aAAa,SAAS;AAC3B,OAAK,MAAM,QAAQ,KAAK,eAAe,QAAQ,CAC7C,MAAK,SAAS;AAEhB,OAAK,eAAe,OAAO;AAC3B,OAAK,OAAO,gBAAgB,OAAO;AACnC,OAAK,OAAO,eAAe,OAAO;AAClC,OAAK,OAAO,aAAa,OAAO;;;;;;;;AASpC,IAAM,sBAAN,MAA0B;CACxB,AAAQ;CACR,AAAQ;CACR,AAAQ,cAAc;CACtB,AAAQ,SAAyB,EAAE;CAEnC,YAAY,KAAgB,SAA6C;AACvE,OAAK,MAAM;AACX,OAAK,UAAU;;;CAIjB,MAAM,UAAU,QAAuB,OAA+B;AAGpE,MAAI,EAAE,SAAS,KAAK,SAClB;AAEF,SAAO,OAAO,SAAS,KAAK,eAAe,6BAA6B,OAAO;GAC7E,MAAM,OAAO,KAAK;GAClB,MAAM,KAAK,QAAQ,OAAO,SAAS,KAAK,IAAI,OAAO,QAAQ,OAAO,0BAA0B;AAC5F,OAAI,MAAM,KACR;GAEF,MAAM,MAAM,MAAM,KAAK,IAAI,iBAAiB,QAAQ,MAAM,GAAG;AAC7D,QAAK,OAAO,KAAK,IAAI;AACrB,QAAK,cAAc;AACnB,QAAK,UAAU;IACb;IACA,YAAY,KAAK,IAAI;IACrB,YAAY;IACZ,SAAS,SAAS,KAAK,eAAe,OAAO;IAC9C,CAAC;AACF,OAAI,SAAS,KAAK,eAAe,OAAO,OACtC;;;;CAMN,UAAU,aAAmC;EAC3C,MAAM,MAAM,IAAI,aAAa,cAAc,WAAW;EACtD,IAAI,SAAS;AACb,OAAK,MAAM,SAAS,KAAK,QAAQ;AAC/B,OAAI,IAAI,OAAO,OAAO;AACtB,aAAU,MAAM;;AAElB,SAAO;;;;;;AC1iBX,MAAa,kBAAkB;AAC/B,MAAa,kBAAkB;AAC/B,MAAa,6BAA6B;AAI1C,MAAa,oBAAoB;AAMjC,MAAa,iBAAiB;AAG9B,MAAa,kBAAkB;AAO/B,MAAa,UAAU;AACvB,MAAa,oBAAoB;AACjC,MAAa,oBAAoB;AACjC,MAAa,mBAAmB;AAChC,MAAa,iBAAiB;AAC9B,MAAa,kBAAkB;AAC/B,MAAa,oBAAoB;CAAC;CAAG;CAAG;CAAG;CAAE;AAC7C,MAAa,yBAAyB;CAAC;CAAG;CAAG;CAAE;AAI/C,MAAa,kBAAkB;AAC/B,MAAa,oBAAoB;CAAC;CAAG;CAAG;CAAG;CAAE;AAgB7C,SAAgB,mBAAmB,WAAmD;CACpF,MAAM,cAAc,UAAU;CAC9B,MAAM,YAAY,UAAU;AAC5B,QAAO;EACL;EACA,YAAY,UAAU;EACtB;EACA,cAAe,UAAU,uBAAkC;EAC3D,UAAW,UAAU,YAAuB,KAAK,MAAM,cAAc,UAAU;EAC/E,mBAAmB,UAAU;EAC7B,YAAY,UAAU;EACtB,WAAY,UAAU,cAAyB;EAC/C,cAAe,UAAU,gBAA2B;EACpD,yBAA0B,UAAU,2BAAsC;EAC3E;;;;;;;;;;;;AAaH,SAAgB,sBAAsB,UAGpC;CACA,MAAMI,KAAe,EAAE;CACvB,MAAMC,KAAe,EAAE;AACvB,MAAK,MAAM,MAAM,SACf,KAAI,MAAM,mBAAmB,KAAK,kBAAkB,4BAA4B;EAC9E,MAAM,OAAO,KAAK;AAClB,MAAI,OAAO,kBACT,IAAG,KAAK,KAAK;YAEN,MAAM,mBAAmB,KAAK,kBAAkB,4BAA4B;EACrF,MAAM,OAAO,KAAK;AAClB,MAAI,OAAO,kBACT,IAAG,KAAK,KAAK;;CAInB,MAAM,YAAY,KAAK,IAAI,GAAG,QAAQ,GAAG,OAAO;CAChD,MAAM,QAAQ,IAAI,YAAY,oBAAoB,UAAU;AAC5D,MAAK,IAAI,IAAI,GAAG,IAAI,WAAW,KAAK;AAClC,QAAM,KAAK,GAAG;AACd,QAAM,YAAY,KAAK,GAAG;;AAE5B,QAAO;EAAE;EAAO;EAAW;;;AAI7B,SAAS,iBACP,GACA,QACA,GACA,KACA,QACQ;AACR,SAAQ,IAAI,KAAK,SAAS,IAAI,OAAO,IAAI,KAAK,SAAS;;;AAazD,SAAS,OAAO,MAAsB;AACpC,QAAO,OAAO;;AAwBhB,SAAgB,8BAA8B,MAAyC;CACrF,MAAM,IAAI,KAAK;CACf,MAAM,MAAM,KAAK,gBAAgB;CACjC,MAAM,QAAQ,KAAK,gBAAgB;CACnC,MAAM,aAAa,KAAK,cAAc;CACtC,MAAM,YAAY,KAAK,aAAa;CACpC,MAAMC,UAAsC,EAAE;CAC9C,MAAMC,QAAkB,EAAE;CAC1B,MAAMC,iBAA2B,EAAE;CAEnC,MAAM,aAAa,SAA2B;AAC5C,UAAQ,KAAK,QAAQ;;CAEvB,MAAM,WAAW,SAAuB;AACtC,QAAM,KAAK,KAAK;AAChB,iBAAe,KAAK,KAAK,GAAG;;AAI9B,WAAU;EACR,MAAM;EACN,OAAO,CAAC,KAAK,EAAE;EACf,OAAO;EACP,SAAS;EACV,CAAC;AAIF,WAAU;EACR,MAAM,OAAO,qBAAqB;EAClC,OAAO;GAAC;GAAK;GAAmB;GAAiB;EACjD,OAAO;EACP,SAAS;EACT,gBAAgB,OAAO,qBAAqB;EAC7C,CAAC;AACF,WAAU;EACR,MAAM,OAAO,4BAA4B;EACzC,OAAO;GAAC;GAAK;GAAW;GAAiB;EACzC,OAAO;EACP,SAAS;EACT,gBAAgB,OAAO,4BAA4B;EACpD,CAAC;AACF,WAAU;EACR,MAAM,OAAO,0BAA0B;EACvC,OAAO,CAAC,KAAK,UAAU;EACvB,OAAO;EACP,SAAS;EACT,gBAAgB,OAAO,0BAA0B;EAClD,CAAC;AACF,WAAU;EACR,MAAM;EACN,OAAO,CAAC,WAAW,EAAE;EACrB,OAAO;EACP,SAAS;EACV,CAAC;AACF,SAAQ;EACN,IAAI;EACJ,QAAQ;EACR,QAAQ;GACN;GACA,OAAO,qBAAqB;GAC5B,OAAO,4BAA4B;GACnC,OAAO,0BAA0B;GAClC;EACD,SAAS,CAAC,aAAa;EACvB,YAAY;GACV;GACA,MAAM;GACN;GACA,GAAG;GACH,eAAe;GAChB;EACF,CAAC;;CAGF,SAAS,WACP,IACA,OACA,WACA,SACA,KACA,MACA,GACA,GACA,UACA,SACA,QACQ;EACR,MAAM,OAAO,KAAK,OAAO,IAAI,IAAI,UAAU,YAAY,IAAI,KAAK,KAAK,EAAE,GAAG;AAC1E,YAAU;GACR,MAAM;GACN,OAAO;IAAC;IAAM;IAAK;IAAE;GACrB,OAAO;GACP,SAAS;GACT,gBAAgB;GACjB,CAAC;AACF,YAAU;GACR,MAAM;GACN,OAAO,CAAC,KAAK;GACb,OAAO;GACP,SAAS;GACT,gBAAgB;GACjB,CAAC;AACF,YAAU;GAAE,MAAM;GAAQ,OAAO,CAAC,MAAM,KAAK;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AACrF,UAAQ;GACN;GACA,QAAQ;GACR,QAAQ;IAAC;IAAO;IAAW;IAAQ;GACnC,SAAS,CAAC,OAAO;GACjB,YAAY;IAAE;IAAK;IAAM;IAAG;IAAM;IAAG,QAAQ;IAAG;IAAS;IAAU;GACpE,CAAC;AACF,SAAO;;;CAIT,SAAS,MACP,IACA,OACA,GACA,GACA,UACA,QACM;AACN,YAAU;GACR,MAAM;GACN,OAAO,CAAC,EAAE;GACV,OAAO;GACP,SAAS;GACT,gBAAgB;GACjB,CAAC;AACF,YAAU;GAAE,MAAM;GAAQ,OAAO,CAAC,GAAG,EAAE;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AAC/E,UAAQ;GACN;GACA,QAAQ;GACR,QAAQ,CAAC,OAAO,SAAS;GACzB,SAAS,CAAC,OAAO;GACjB,YAAY;IAAE;IAAG;IAAG;GACrB,CAAC;;;CAIJ,SAAS,cACP,IACA,OACA,WACA,SACA,KACA,MACA,GACA,GACA,QACA,SACA,QACA,QACQ;EACR,MAAM,OAAO,iBAAiB,GAAG,QAAQ,GAAG,SAAS,OAAO;AAC5D,YAAU;GACR,MAAM;GACN,OAAO;IAAC;IAAK;IAAM;IAAE;GACrB,OAAO;GACP,SAAS;GACT,gBAAgB;GACjB,CAAC;AACF,YAAU;GACR,MAAM;GACN,OAAO,CAAC,KAAK;GACb,OAAO;GACP,SAAS;GACT,gBAAgB;GACjB,CAAC;AACF,YAAU;GAAE,MAAM;GAAQ,OAAO,CAAC,MAAM,KAAK;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AACrF,UAAQ;GACN;GACA,QAAQ;GACR,QAAQ;IAAC;IAAO;IAAW;IAAQ;GACnC,SAAS,CAAC,OAAO;GACjB,YAAY;IAAE;IAAK;IAAM;IAAG;IAAM;IAAG;IAAQ;IAAS,gBAAgB;IAAQ;GAC/E,CAAC;AACF,SAAO;;;CAIT,SAAS,IAAI,IAAY,GAAW,GAAW,GAAW,GAAW,QAAsB;AACzF,YAAU;GAAE,MAAM;GAAQ,OAAO,CAAC,GAAG,EAAE;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AAC/E,UAAQ;GACN;GACA,QAAQ;GACR,QAAQ,CAAC,GAAG,EAAE;GACd,SAAS,CAAC,OAAO;GACjB,YAAY;IAAE,cAAc;IAAG,aAAa,IAAI;IAAG;GACpD,CAAC;;CAIJ,IAAI,MAAM;CACV,IAAI,SAAS,WACX,aACA,cACA,OAAO,yBAAyB,EAChC,OAAO,uBAAuB,EAC9B,WACA,YACA,GACA,GACA,GACA,GACA,IACD;CACD,IAAI,WAAW;AAGf,MAAK,IAAI,QAAQ,GAAG,QAAQ,MAAM,QAAQ,SAAS;EACjD,MAAM,IAAI,QAAQ;EAClB,MAAM,OAAO,MAAM;EACnB,MAAM,QAAQ,WAAW;EACzB,MAAM,MAAM,IAAI;EAChB,MAAM,QAAQ,KAAK,KAAK,OAAO,EAAE;EAGjC,MAAM,WAAW,QAAQ,MAAM;AAC/B,QACE,QAAQ,MAAM,YACd,KACA,UACA,QACA,OAAO,iBAAiB,EAAE,gBAAgB,EAC1C,SACD;EAGD,MAAM,QAAQ,QAAQ,MAAM;AAC5B,WAAS,cACP,QAAQ,MAAM,SACd,UACA,OAAO,iBAAiB,EAAE,iBAAiB,EAC3C,OAAO,iBAAiB,EAAE,eAAe,EACzC,UACA,OACA,KACA,QACA,MACA,OACA,GACA,MACD;AACD,aAAW;EAGX,IAAI,IAAI;AACR,OAAK,IAAI,IAAI,GAAG,IAAI,uBAAuB,QAAQ,KAAK;GACtD,MAAM,KAAK,IAAI;GACf,MAAM,MAAM,uBAAuB;GACnC,MAAM,IAAI,iBAAiB,EAAE,SAAS,GAAG;GACzC,MAAM,MAAM,KAAK,MAAQ,IAAS,MAAO,EAAE;GAE3C,MAAM,KAAK,QAAQ,MAAM,IAAI,EAAE;AAC/B,SAAM,GAAG,GAAG,MAAM,GAAG,UAAU,QAAQ,OAAO,GAAG,EAAE,UAAU,EAAE,GAAG;GAClE,MAAM,KAAK,QAAQ,MAAM,IAAI,EAAE;AAC/B,cACE,GAAG,GAAG,MACN,IACA,OAAO,GAAG,EAAE,WAAW,EACvB,OAAO,GAAG,EAAE,SAAS,EACrB,UACA,UACA,GACA,QACA,KACA,KACA,GACD;GACD,MAAM,KAAK,QAAQ,MAAM,IAAI,EAAE;AAC/B,SAAM,GAAG,GAAG,MAAM,IAAI,UAAU,QAAQ,OAAO,GAAG,EAAE,UAAU,EAAE,GAAG;GACnE,MAAM,KAAK,QAAQ,MAAM,IAAI,EAAE;AAC/B,cACE,GAAG,GAAG,MACN,IACA,OAAO,GAAG,EAAE,WAAW,EACvB,OAAO,GAAG,EAAE,SAAS,EACrB,UACA,UACA,GACA,QACA,GACA,GACA,GACD;GAED,MAAM,SAAS,QAAQ,MAAM,IAAI,EAAE;AACnC,OAAI,GAAG,OAAO,MAAM,GAAG,IAAI,UAAU,QAAQ,OAAO;AACpD,OAAI;;AAEN,QAAM;;CAMR,MAAM,gBAAgB,MAAM,SAAS;CACrC,MAAM,eAAe,MAAM,SAAS;CACpC,MAAM,YAAY;AAClB,OACE,qBACA,KACA,UACA,QACA,OAAO,iBAAiB,cAAc,QAAQ,EAC9C,UACD;CAGD,MAAM,MAAM;CACZ,MAAM,SAAS,WACb,aACA,WACA,OAAO,iBAAiB,aAAa,SAAS,EAC9C,OAAO,iBAAiB,aAAa,OAAO,EAC5C,UACA,GACA,GACA,QACA,GACA,GACA,IACD;AAID,QAAO;EACL,cAAc;EACd,QAAQ;GACN,aAAa;GACb,YAAY,MAAM;GAClB,WAAW;GACX,cAAc;GACd,UAAU;GACV,mBAAmB;GACnB,YAAY;GACZ,gBAAgB;GAChB,cAAc;GACd,WAAW;GACX,WAAW;GACX,UAAU;GACV,WAAW;GACX,QAAQ;GACR,kBAAkB;GACnB;EACD,cAAc;GAAE,MAAM;GAAM,QAAQ;GAAO,KAAK;GAAO;EACvD;EACA;EACA;EACA,QAAQ,CAAC,cAAc;EACvB,SAAS,CAAC,IAAI;EACf;;;;;;AAOH,SAAgB,gBACd,WACA,QAA2B,mBACnB;CACR,IAAI,SAAS;AACb,MAAK,MAAM,QAAQ,MACjB,UAAS,iBAAiB,QAAQ,MAAM,IAAI,MAAM,KAAK,KAAK,OAAO,EAAE,EAAE,EAAE;AAE3E,QAAO;;;AAIT,SAAS,UAAU,GAAW,GAAW,QAAgB,KAAa,UAA0B;AAC9F,QAAO,KAAK,OAAO,IAAI,IAAI,MAAM,YAAY,IAAI,KAAK,KAAK,OAAO,GAAG;;;AAIvE,SAAgB,iBAAiB,YAA4B;CAC3D,IAAI,SAAS;AACb,MAAK,MAAM,QAAQ,kBAEjB,UAAS,UAAU,QAAQ,IAAI,MAAM,MAAM,KAAK,KAAK,OAAO,EAAE,EAAE,EAAE;AAEpE,QAAO;;AAoBT,SAAgB,8BAA8B,MAA0C;CACtF,MAAM,KAAK,KAAK;CAChB,MAAMF,UAAsC,EAAE;CAC9C,MAAMC,QAAkB,EAAE;CAC1B,MAAMC,iBAA2B,EAAE;CAEnC,MAAM,aAAa,SAA2B;AAC5C,UAAQ,KAAK,QAAQ;;CAEvB,MAAM,WAAW,SAAuB;AACtC,QAAM,KAAK,KAAK;AAChB,iBAAe,KAAK,KAAK,GAAG;;AAI9B,WAAU;EAAE,MAAM;EAAU,OAAO,CAAC,GAAG,GAAG;EAAE,OAAO;EAAO,SAAS;EAAc,CAAC;;CAGlF,SAAS,WACP,IACA,OACA,WACA,SACA,KACA,MACA,GACA,GACA,QACA,UACA,SACA,QACQ;EACR,MAAM,OAAO,UAAU,GAAG,GAAG,QAAQ,SAAS,SAAS;AACvD,YAAU;GACR,MAAM;GACN,OAAO;IAAC;IAAM;IAAK;IAAE;GACrB,OAAO;GACP,SAAS;GACT,gBAAgB;GACjB,CAAC;AACF,YAAU;GACR,MAAM;GACN,OAAO,CAAC,KAAK;GACb,OAAO;GACP,SAAS;GACT,gBAAgB;GACjB,CAAC;AACF,YAAU;GAAE,MAAM;GAAQ,OAAO,CAAC,MAAM,KAAK;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AACrF,UAAQ;GACN;GACA,QAAQ;GACR,QAAQ;IAAC;IAAO;IAAW;IAAQ;GACnC,SAAS,CAAC,OAAO;GACjB,YAAY;IAAE;IAAK;IAAM;IAAG;IAAM;IAAG;IAAQ;IAAS;IAAU;GACjE,CAAC;AACF,SAAO;;;CAIT,SAAS,MACP,IACA,OACA,GACA,GACA,UACA,QACM;AACN,YAAU;GACR,MAAM;GACN,OAAO,CAAC,EAAE;GACV,OAAO;GACP,SAAS;GACT,gBAAgB;GACjB,CAAC;AACF,YAAU;GAAE,MAAM;GAAQ,OAAO,CAAC,GAAG,EAAE;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AAC/E,UAAQ;GACN;GACA,QAAQ;GACR,QAAQ,CAAC,OAAO,SAAS;GACzB,SAAS,CAAC,OAAO;GACjB,YAAY;IAAE;IAAG;IAAG;GACrB,CAAC;;;CAIJ,SAAS,IAAI,IAAY,GAAW,GAAW,GAAW,GAAW,QAAsB;AACzF,YAAU;GAAE,MAAM;GAAQ,OAAO,CAAC,GAAG,EAAE;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AAC/E,UAAQ;GACN;GACA,QAAQ;GACR,QAAQ,CAAC,GAAG,EAAE;GACd,SAAS,CAAC,OAAO;GACjB,YAAY;IAAE,cAAc;IAAG,aAAa,IAAI;IAAG;GACpD,CAAC;;CAGJ,MAAM,UAAU,SAAyB,OAAO,WAAW,OAAO;CAGlE,IAAI,MAAM;CACV,IAAI,SAAS,WACX,oBACA,UACA,OAAO,iBAAiB,EACxB,OAAO,eAAe,EACtB,GACA,iBACA,GACA,IACA,GACA,GACA,GACA,IACD;CACD,IAAI,WAAW;AAGf,MAAK,IAAI,QAAQ,GAAG,QAAQ,kBAAkB,QAAQ,SAAS;EAC7D,MAAM,IAAI,QAAQ;EAClB,MAAM,OAAO,kBAAkB;EAG/B,IAAI,IAAI;AACR,OAAK,IAAI,IAAI,GAAG,IAAI,uBAAuB,QAAQ,KAAK;GACtD,MAAM,MAAM,uBAAuB;GACnC,MAAM,IAAI,SAAS,EAAE,SAAS,EAAE;GAChC,MAAM,MAAM,KAAK,MAAQ,IAAS,MAAO,EAAE;GAE3C,MAAM,KAAK,YAAY,MAAM,IAAI,EAAE;AACnC,SAAM,GAAG,GAAG,MAAM,GAAG,UAAU,QAAQ,OAAO,GAAG,EAAE,UAAU,EAAE,GAAG;GAClE,MAAM,KAAK,YAAY,MAAM,IAAI,EAAE;AACnC,cACE,GAAG,GAAG,MACN,IACA,OAAO,GAAG,EAAE,WAAW,EACvB,OAAO,GAAG,EAAE,SAAS,EACrB,UACA,UACA,GACA,QACA,GACA,KACA,KACA,GACD;GACD,MAAM,KAAK,YAAY,MAAM,IAAI,EAAE;AACnC,SAAM,GAAG,GAAG,MAAM,IAAI,UAAU,QAAQ,OAAO,GAAG,EAAE,UAAU,EAAE,GAAG;GACnE,MAAM,KAAK,YAAY,MAAM,IAAI,EAAE;AACnC,cACE,GAAG,GAAG,MACN,IACA,OAAO,GAAG,EAAE,WAAW,EACvB,OAAO,GAAG,EAAE,SAAS,EACrB,UACA,UACA,GACA,QACA,GACA,GACA,GACA,GACD;GACD,MAAM,SAAS,YAAY,MAAM,IAAI,EAAE;AACvC,OAAI,GAAG,OAAO,MAAM,GAAG,IAAI,UAAU,QAAQ,OAAO;AACpD,OAAI;;EAIN,MAAM,WAAW,YAAY,MAAM;AACnC,QAAM,GAAG,SAAS,MAAM,GAAG,UAAU,QAAQ,OAAO,SAAS,EAAE,gBAAgB,EAAE,SAAS;EAG1F,MAAM,QAAQ,WAAW;EACzB,MAAM,UAAU,YAAY,MAAM;AAClC,WAAS,WACP,GAAG,QAAQ,MACX,UACA,OAAO,SAAS,EAAE,iBAAiB,EACnC,OAAO,SAAS,EAAE,eAAe,EACjC,UACA,OACA,IAAI,MACJ,QACA,MACA,GACA,KAAK,KAAK,OAAO,EAAE,EACnB,QACD;AACD,aAAW;AACX,QAAM;;CAIR,MAAM,YAAY;AAClB,OAAM,yBAAyB,KAAK,UAAU,QAAQ,OAAO,gBAAgB,EAAE,UAAU;CAGzF,MAAM,SAAS;CACf,MAAM,IAAI,WACR,oBACA,WACA,OAAO,iBAAiB,EACxB,OAAO,eAAe,EACtB,UACA,gBACA,GACA,QACA,GACA,GACA,GACA,OACD;AAGD,WAAU;EACR,MAAM,OAAO,2BAA2B;EACxC,OAAO;GAAC;GAAmB;GAAkB;GAAe;EAC5D,OAAO;EACP,SAAS;EACT,gBAAgB,OAAO,2BAA2B;EACnD,CAAC;AACF,WAAU;EACR,MAAM,OAAO,yBAAyB;EACtC,OAAO,CAAC,mBAAmB,iBAAiB;EAC5C,OAAO;EACP,SAAS;EACT,gBAAgB,OAAO,yBAAyB;EACjD,CAAC;AACF,WAAU;EACR,MAAM,OAAO,qBAAqB;EAClC,OAAO;GAAC;GAAmB;GAAmB;GAAiB;EAC/D,OAAO;EACP,SAAS;EACT,gBAAgB,OAAO,qBAAqB;EAC7C,CAAC;AACF,WAAU;EACR,MAAM,OAAO,4BAA4B;EACzC,OAAO;GAAC;GAAmB;GAAgB;GAAiB;EAC5D,OAAO;EACP,SAAS;EACT,gBAAgB,OAAO,4BAA4B;EACpD,CAAC;AACF,WAAU;EACR,MAAM,OAAO,0BAA0B;EACvC,OAAO,CAAC,mBAAmB,eAAe;EAC1C,OAAO;EACP,SAAS;EACT,gBAAgB,OAAO,0BAA0B;EAClD,CAAC;AACF,WAAU;EACR,MAAM;EACN,OAAO,CAAC,mBAAmB,EAAE;EAC7B,OAAO;EACP,SAAS;EACV,CAAC;AACF,SAAQ;EACN,IAAI;EACJ,QAAQ;EACR,QAAQ;GACN;GACA,OAAO,2BAA2B;GAClC,OAAO,yBAAyB;GAChC,OAAO,qBAAqB;GAC5B,OAAO,4BAA4B;GACnC,OAAO,0BAA0B;GAClC;EACD,SAAS,CAAC,YAAY;EACtB,YAAY;GACV,KAAK;GACL,MAAM;GACN;GACA,GAAG;GACH,eAAe;GAChB;EACF,CAAC;AAEF,QAAO;EACL,cAAc;EACd,QAAQ;GACN,aAAa;GACb,YAAY,kBAAkB;GAC9B,WAAW;GACX,cAAc;GACd,UAAU;GACV,mBAAmB;GACnB,YAAY;GACZ,gBAAgB;GAChB,cAAc;GACd,WAAW;GACX,WAAW;GACX,UAAU;GACV,WAAW;GACX,QAAQ;GACR,kBAAkB;GACnB;EACD,cAAc;GAAE,MAAM;GAAM,QAAQ;GAAO,KAAK;GAAO;EACvD;EACA;EACA;EACA,QAAQ,CAAC,SAAS;EAClB,SAAS,CAAC,YAAY;EACvB;;AAcH,SAAgB,6BACd,WACA,OACA,WACA,SACY;CACZ,MAAM,MAAM,mBAAmB,UAAU;CACzC,MAAM,QAAQ,UAAU;CACxB,MAAM,UAAU,aAAa;CAC7B,MAAM,cAAc,IAAI;CACxB,MAAM,aAAa,IAAI;CACvB,MAAM,YAAY,IAAI;CACtB,MAAM,eAAe,IAAI;CACzB,MAAM,WAAW,IAAI;CACrB,MAAM,WAAW;CACjB,MAAM,YAAY,IAAI;CACtB,MAAM,eAAe,IAAI;CACzB,MAAM,oBAAoB,IAAI;CAC9B,MAAM,aAAa,IAAI;CACvB,MAAM,QAAQ,YAAY;CAC1B,MAAM,SAAS,eAAe;CAC9B,MAAMC,eAAsB,WAAW;CAEvC,MAAMC,SAA0B;EAC9B;EACA;EACA;EACA;EACA;EACA;EACA;EACA,gBAAgB,IAAI;EACpB;EACA,WAAW;EACX;EACA;EACA,WAAW;EACX,QAAQ;EACR,kBAAkB;EACnB;CAED,MAAMJ,UAAsC,EAAE;CAC9C,MAAMC,QAAkB,EAAE;CAC1B,MAAMC,iBAA2B,EAAE;CACnC,MAAM,aAAa,SAA2B;AAC5C,UAAQ,KAAK,QAAQ;;CAEvB,MAAM,WAAW,SAAuB;AACtC,QAAM,KAAK,KAAK;AAChB,iBAAe,KAAK,KAAK,GAAG;;CAM9B,MAAM,aACJ,IACA,iBACA,YACA,aACA,YACA,GACA,MACS;AACT,MAAI,OAAO;GACT,MAAM,gBAAgB,IAAI;AAC1B,aAAU;IACR,MAAM,GAAG,WAAW;IACpB,OAAO,CAAC,KAAK,KAAK,gBAAgB,EAAE,CAAC;IACrC,OAAO;IACP,SAAS;IACV,CAAC;AACF,aAAU;IACR,MAAM,GAAG,WAAW;IACpB,OAAO,CAAC,KAAK,KAAK,gBAAgB,QAAQ,CAAC;IAC3C,OAAO;IACP,SAAS;IACV,CAAC;AACF,aAAU;IACR,MAAM,GAAG,WAAW;IACpB,OAAO,CAAC,KAAK,KAAK,gBAAgB,QAAQ,CAAC;IAC3C,OAAO;IACP,SAAS;IACV,CAAC;AACF,WAAQ;IACN;IACA,QAAQ;IACR,QAAQ;KAAC;KAAiB,GAAG,WAAW;KAAK,GAAG,WAAW;KAAU,GAAG,WAAW;KAAQ;IAC3F,SAAS,CAAC,WAAW;IACrB,YAAY;KAAE,UAAU;KAAiB;KAAG;KAAG,YAAY;KAAS;IACrE,CAAC;AACF;;AAEF,YAAU;GACR,MAAM;GACN,OAAO;GACP,OAAO;GACP,SAAS;GACT,gBAAgB;GACjB,CAAC;AACF,UAAQ;GACN;GACA,QAAQ;GACR,QAAQ,CAAC,iBAAiB,WAAW;GACrC,SAAS,CAAC,WAAW;GACrB,YAAY;IAAE,UAAU;IAAiB;IAAG;IAAG;GAChD,CAAC;;AAIJ,WAAU;EAAE,MAAM;EAAa,OAAO,CAAC,IAAI;EAAE,OAAO;EAAO,SAAS;EAAc,CAAC;AACnF,WAAU;EAAE,MAAM;EAAa,OAAO,CAAC,KAAK,YAAY;EAAE,OAAO;EAAO,SAAS;EAAc,CAAC;AAChG,WAAU;EACR,MAAM,eAAe;EACrB,OAAO,CAAC,YAAY,YAAY;EAChC,OAAO;EACP,SAAS;EACT,gBAAgB,eAAe;EAChC,CAAC;AACF,SAAQ;EACN,IAAI;EACJ,QAAQ;EACR,QAAQ,CAAC,aAAa,eAAe,MAAM;EAC3C,SAAS,CAAC,YAAY;EACtB,YAAY;GAAE;GAAY;GAAa;EACxC,CAAC;CAEF,IAAI,aAAa;AACjB,MAAK,IAAI,IAAI,GAAG,IAAI,YAAY,KAAK;EACnC,MAAM,SAAS,QAAQ;EAGvB,MAAM,QAAQ,eAAe,eAAe,EAAE;AAC9C,YAAU;GACR,MAAM;GACN,OAAO,CAAC,YAAY;GACpB,OAAO;GACP,SAAS;GACT,gBAAgB;GACjB,CAAC;EACF,MAAM,WAAW,GAAG,OAAO;AAC3B,YAAU;GAAE,MAAM;GAAU,OAAO,CAAC,KAAK,YAAY;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AAC7F,UAAQ;GACN,IAAI,GAAG,OAAO;GACd,QAAQ;GACR,QAAQ,CAAC,YAAY,MAAM;GAC3B,SAAS,CAAC,SAAS;GACnB,YAAY;IAAE;IAAa,KAAK;IAAc,gBAAgB;IAAY;GAC3E,CAAC;EAGF,MAAM,OAAO,GAAG,OAAO;EACvB,MAAM,OAAO,GAAG,OAAO;EACvB,MAAM,OAAO,GAAG,OAAO;AACvB,YAAU;GAAE,MAAM;GAAM,OAAO,CAAC,KAAK,MAAM;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AACnF,YAAU;GAAE,MAAM;GAAM,OAAO,CAAC,KAAK,OAAO;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AACpF,YAAU;GAAE,MAAM;GAAM,OAAO,CAAC,KAAK,OAAO;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AACpF,YACE,GAAG,OAAO,UACV,UACA,eAAe,MAAM,EAAE,EACvB,CAAC,OAAO,YAAY,EACpB,MACA,aACA,MACD;AACD,YACE,GAAG,OAAO,UACV,UACA,eAAe,MAAM,EAAE,EACvB,CAAC,QAAQ,YAAY,EACrB,MACA,aACA,OACD;AACD,YACE,GAAG,OAAO,UACV,UACA,eAAe,MAAM,EAAE,EACvB,CAAC,QAAQ,YAAY,EACrB,MACA,aACA,OACD;EAGD,MAAM,SAAS,eAAe,MAAM,EAAE;EACtC,MAAM,SAAS,eAAe,MAAM,EAAE;AACtC,YAAU;GACR,MAAM;GACN,OAAO,CAAC,SAAS;GACjB,OAAO;GACP,SAAS;GACT,gBAAgB;GACjB,CAAC;AACF,YAAU;GACR,MAAM;GACN,OAAO,CAAC,SAAS;GACjB,OAAO;GACP,SAAS;GACT,gBAAgB;GACjB,CAAC;EACF,MAAM,UAAU,GAAG,OAAO;EAC1B,MAAM,UAAU,GAAG,OAAO;AAC1B,YAAU;GAAE,MAAM;GAAS,OAAO,CAAC,KAAK,MAAM;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AACtF,YAAU;GAAE,MAAM;GAAS,OAAO,CAAC,KAAK,OAAO;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AACvF,UAAQ;GACN,IAAI,GAAG,OAAO;GACd,QAAQ;GACR,QAAQ,CAAC,MAAM,OAAO;GACtB,SAAS,CAAC,QAAQ;GAClB,YAAY;IAAE,aAAa;IAAU,KAAK;IAAc,gBAAgB;IAAM;GAC/E,CAAC;AACF,UAAQ;GACN,IAAI,GAAG,OAAO;GACd,QAAQ;GACR,QAAQ,CAAC,MAAM,OAAO;GACtB,SAAS,CAAC,QAAQ;GAClB,YAAY;IAAE,aAAa;IAAU,KAAK;IAAc,gBAAgB;IAAM;GAC/E,CAAC;AAGF,UAAQ;GACN,IAAI,GAAG,OAAO;GACd,QAAQ;GACR,QAAQ,CAAC,SAAS,QAAQ;GAC1B,SAAS,CAAC,SAAS,QAAQ;GAC3B,YAAY;IACV;IACA,aAAa;IACb;IACA;IACA;IACA,gBAAgB;IACjB;GACF,CAAC;EAGF,MAAM,SAAS,GAAG,OAAO;EACzB,MAAM,SAAS,GAAG,OAAO;AACzB,YAAU;GAAE,MAAM;GAAQ,OAAO,CAAC,SAAS,OAAO;GAAE,OAAO;GAAc,SAAS;GAAY,CAAC;AAC/F,YAAU;GAAE,MAAM;GAAQ,OAAO,CAAC,SAAS,OAAO;GAAE,OAAO;GAAc,SAAS;GAAY,CAAC;AAC/F,UAAQ;GACN,IAAI,GAAG,OAAO;GACd,QAAQ;GACR,QAAQ,CAAC,QAAQ;GACjB,SAAS,CAAC,OAAO;GACjB,YAAY;IAAE,OAAO;IAAQ,UAAU;IAAS;GACjD,CAAC;AACF,UAAQ;GACN,IAAI,GAAG,OAAO;GACd,QAAQ;GACR,QAAQ,CAAC,KAAK;GACd,SAAS,CAAC,OAAO;GACjB,YAAY;IAAE,OAAO;IAAQ,UAAU;IAAM;GAC9C,CAAC;EAGF,MAAM,UAAU,GAAG,OAAO;AAC1B,YAAU;GAAE,MAAM;GAAS,OAAO,CAAC,KAAK,MAAM;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AACtF,UAAQ;GACN,IAAI,GAAG,OAAO;GACd,QAAQ;GACR,QAAQ;IAAC;IAAS;IAAQ;IAAO;GACjC,SAAS,CAAC,QAAQ;GAClB,YAAY;IACV,aAAa;IACb,aAAa;IACb;IACA;IACA,QAAQ;IACR,aAAa;IACd;GACF,CAAC;EAGF,MAAM,WAAW,GAAG,OAAO;AAC3B,YAAU;GAAE,MAAM;GAAU,OAAO,CAAC,KAAK,YAAY;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AAC7F,YACE,GAAG,OAAO,UACV,SACA,eAAe,MAAM,EAAE,EACvB,CAAC,aAAa,MAAM,EACpB,UACA,OACA,YACD;EAGD,MAAM,YAAY,GAAG,OAAO;EAC5B,MAAM,YAAY,eAAe,kBAAkB,EAAE;EACrD,MAAM,WAAW,GAAG,OAAO;AAC3B,YAAU;GAAE,MAAM;GAAW,OAAO,CAAC,KAAK,YAAY;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AAC9F,YAAU;GACR,MAAM;GACN,OAAO,CAAC,YAAY;GACpB,OAAO;GACP,SAAS;GACT,gBAAgB;GACjB,CAAC;AACF,YAAU;GAAE,MAAM;GAAU,OAAO,CAAC,KAAK,YAAY;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AAC7F,UAAQ;GACN,IAAI,GAAG,OAAO;GACd,QAAQ;GACR,QAAQ;IAAC;IAAY;IAAU;IAAU;GACzC,SAAS,CAAC,WAAW,SAAS;GAC9B,YAAY;IAAE;IAAa,KAAK;IAAc,gBAAgB;IAAY;GAC3E,CAAC;EAGF,MAAM,UAAU,GAAG,OAAO;EAC1B,MAAM,QAAQ,GAAG,OAAO;EACxB,MAAM,YAAY,GAAG,OAAO;EAC5B,MAAM,SAAS,GAAG,OAAO;AACzB,YAAU;GACR,MAAM;GACN,OAAO,CAAC,KAAK,kBAAkB;GAC/B,OAAO;GACP,SAAS;GACV,CAAC;AACF,YAAU;GACR,MAAM;GACN,OAAO,CAAC,KAAK,kBAAkB;GAC/B,OAAO;GACP,SAAS;GACV,CAAC;AACF,YAAU;GACR,MAAM;GACN,OAAO,CAAC,KAAK,kBAAkB;GAC/B,OAAO;GACP,SAAS;GACV,CAAC;AACF,YAAU;GAAE,MAAM;GAAQ,OAAO,CAAC,KAAK,YAAY;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AAC3F,YACE,GAAG,OAAO,QACV,UACA,eAAe,SAAS,EAAE,EAC1B,CAAC,mBAAmB,YAAY,EAChC,SACA,aACA,kBACD;AACD,YACE,GAAG,OAAO,MACV,UACA,eAAe,OAAO,EAAE,EACxB,CAAC,mBAAmB,YAAY,EAChC,OACA,aACA,kBACD;AACD,UAAQ;GACN,IAAI,GAAG,OAAO;GACd,QAAQ;GACR,QAAQ,CAAC,SAAS,MAAM;GACxB,SAAS,CAAC,UAAU;GACpB,YAAY,EAAE,cAAc,SAAS;GACtC,CAAC;AACF,YACE,GAAG,OAAO,QACV,WACA,eAAe,SAAS,EAAE,EAC1B,CAAC,aAAa,kBAAkB,EAChC,QACA,mBACA,YACD;EAGD,MAAM,YAAY,GAAG,OAAO;AAC5B,YAAU;GAAE,MAAM;GAAW,OAAO,CAAC,KAAK,YAAY;GAAE,OAAO;GAAO,SAAS;GAAc,CAAC;AAC9F,UAAQ;GACN,IAAI,GAAG,OAAO;GACd,QAAQ;GACR,QAAQ,CAAC,WAAW,OAAO;GAC3B,SAAS,CAAC,UAAU;GACpB,YAAY;IAAE,cAAc;IAAW;IAAa;GACrD,CAAC;AACF,eAAa;;AAIf,WAAU;EACR,MAAM,eAAe;EACrB,OAAO,CAAC,YAAY;EACpB,OAAO;EACP,SAAS;EACT,gBAAgB,eAAe;EAChC,CAAC;AACF,WAAU;EACR,MAAM;EACN,OAAO,CAAC,KAAK,YAAY;EACzB,OAAO;EACP,SAAS;EACV,CAAC;AACF,SAAQ;EACN,IAAI;EACJ,QAAQ;EACR,QAAQ,CAAC,YAAY,eAAe,WAAW;EAC/C,SAAS,CAAC,iBAAiB;EAC3B,YAAY;GAAE;GAAa,KAAK;GAAc,gBAAgB;GAAY;EAC3E,CAAC;AAGF,WAAU;EACR,MAAM;EACN,OAAO,CAAC,GAAG,YAAY;EACvB,OAAO;EACP,SAAS;EACV,CAAC;AACF,SAAQ;EACN,IAAI;EACJ,QAAQ;EACR,QAAQ,CAAC,iBAAiB;EAC1B,SAAS,CAAC,kBAAkB;EAC5B,YAAY,EAAE,OAAO,aAAa;EACnC,CAAC;AACF,WAAU;EAAE,MAAM;EAAU,OAAO,CAAC,GAAG,WAAW;EAAE,OAAO;EAAO,SAAS;EAAc,CAAC;AAC1F,SAAQ;EACN,IAAI;EACJ,QAAQ;EACR,QAAQ,CAAC,mBAAmB,eAAe,MAAM;EACjD,SAAS,CAAC,SAAS;EACnB,YAAY;GAAE,UAAU;GAAmB,GAAG;GAAa,GAAG;GAAY;EAC3E,CAAC;AAEF,QAAO;EACL,cAAc;EACd;EACA,cAAc;GAAE,MAAM;GAAM,QAAQ;GAAO,KAAK;GAAO;EACvD;EACA;EACA;EACA,QAAQ,CAAC,YAAY;EACrB,SAAS,CAAC,SAAS;EACpB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AC10CH,MAAM,UAAU;AAEhB,MAAM,gBAAgB;AAEtB,MAAM,iBAAiB;AACvB,MAAM,iBAAiB;AACvB,MAAM,kBAAkB;AACxB,MAAM,YAAY;AAClB,MAAM,kBAAkB;AACxB,MAAM,cAAc;;AAGpB,SAAS,WAAW,OAAuB;AACzC,QAAO,KAAK,MAAM,KAAK,IAAI,KAAK,IAAI,OAAO,EAAE,EAAE,EAAE,GAAG,IAAI;;;AAI1D,SAAS,QAAQ,GAA+B;CAC9C,MAAM,IAAI,EAAE;CACZ,MAAM,OAAO,KAAK,MAAM,IAAI,EAAE,GAAG;CACjC,MAAM,MAAM,IAAI,aAAa,KAAK;AAClC,MAAK,IAAI,IAAI,GAAG,IAAI,MAAM,KAAK;EAC7B,IAAI,KAAK;EACT,IAAI,KAAK;EACT,MAAM,IAAK,KAAK,KAAK,KAAK,IAAK;AAC/B,OAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;GAC1B,MAAM,IAAI,IAAI;AACd,SAAM,EAAE,KAAK,KAAK,IAAI,EAAE;AACxB,SAAM,EAAE,KAAK,KAAK,IAAI,EAAE;;AAE1B,MAAI,KAAK,KAAK,MAAM,IAAI,GAAG;;AAE7B,QAAO;;;;;;AAOT,SAAS,UAAU,OAA6B;CAC9C,IAAI,MAAM;AACV,MAAK,IAAI,IAAI,GAAG,IAAI,MAAM,QAAQ,IAChC,QAAO,MAAM,KAAK,MAAM;AAE1B,QAAO,MAAM,SAAS,KAAK,KAAK,MAAM,MAAM,OAAO,GAAG;;;;;;AAOxD,SAAS,iBAAiB,OAAqB,IAAoB;AACjE,KAAI,MAAM,WAAW,EACnB,QAAO;CAET,MAAM,OAAO,QAAQ,MAAM;CAC3B,MAAM,UAAU,KAAK;CACrB,MAAM,OAAO,KAAK;CAClB,IAAI,UAAU;CACd,IAAI,WAAW;AACf,MAAK,IAAI,IAAI,GAAG,IAAI,MAAM,KAAK;EAE7B,MAAM,OAAO,OAAO,IAAK,KAAK,OAAO,KAAM,UAAU;AACrD,cAAY,OAAO,KAAK;AACxB,aAAW,KAAK;;AAElB,QAAO,WAAW,UAAU;;;AAI9B,SAAS,WAAW,GAAyB;CAC3C,MAAM,IAAI,IAAI,aAAa,EAAE;AAC7B,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IACrB,GAAE,KAAK,KAAM,KAAM,KAAK,IAAK,IAAI,KAAK,KAAK,KAAM,IAAI,GAAG;AAE1D,QAAO;;;;;;;;;AAUT,SAAS,cACP,QACA,KACA,QACA,OACA,UACM;AACN,MAAK,IAAI,IAAI,GAAG,IAAI,iBAAiB,IACnC,OAAM,KAAK,OAAO,MAAM,KAAK,OAAO;AAEtC,MAAK,IAAI,MAAM,GAAG,MAAM,iBAAiB,OAAO;EAC9C,IAAI,MAAM;AACV,OAAK,IAAI,IAAI,GAAG,IAAI,MAAM,iBAAiB,IACzC,QAAO,MAAM,KAAK,MAAM,IAAI;AAE9B,WAAS,OAAO;;;;AAKpB,SAAS,WAAW,UAAwB,QAAgB,QAAgB,IAAoB;CAC9F,IAAI,UAAU;CACd,IAAI,UAAU,OAAO;AACrB,MAAK,IAAI,MAAM,QAAQ,MAAM,QAAQ,MACnC,KAAI,SAAS,OAAO,SAAS;AAC3B,YAAU,SAAS;AACnB,YAAU;;CAId,MAAM,MAAM,KAAK,IAAI,KAAK,IAAI,SAAS,EAAE,EAAE,kBAAkB,EAAE;CAC/D,MAAM,QAAQ,SAAS,MAAM;CAC7B,MAAM,OAAO,SAAS;CACtB,MAAM,QAAQ,SAAS,MAAM;CAC7B,IAAI,QAAS,MAAO,QAAQ,UAAW,QAAQ,IAAI,OAAO,QAAQ;AAClE,KAAI,EAAE,UAAU,KAAK,UAAU,kBAAkB,GAC/C,SAAQ;CAEV,MAAM,UAAU,UAAU,SAAS;CACnC,IAAI,QAAQ,SAAS,IAAI,IAAI,SAAS;AAEtC,KAAI,EADW,WAAW,SAAS,KAAK,QAAQ,iBAE9C,SAAQ;AAGV,QAAO,KAAK,IAAI,KAAK,IAAI,OAAO,eAAe,EAAE,eAAe;;AAGlE,SAAS,aAAa,OAAqB,IAAoB;CAC7D,MAAM,IAAI,MAAM;AAChB,KAAI,IAAI,gBACN,QAAO;CAET,MAAM,UAAU,kBAAmB,IAAI,aAAc;CACrD,MAAM,SAAS,IAAI,aAAa,IAAI,OAAO;AAC3C,QAAO,IAAI,MAAM;CACjB,MAAM,YAAY,KAAK,OAAO,OAAO,SAAS,mBAAmB,UAAU,GAAG;AAC9E,KAAI,aAAa,EACf,QAAO;CAET,MAAM,SAAS,WAAW,gBAAgB;CAC1C,MAAM,SAAS,KAAK,IAAI,GAAG,KAAK,MAAM,KAAK,eAAe,CAAC;CAC3D,MAAM,SAAS,KAAK,IAAI,iBAAiB,KAAK,MAAM,KAAK,eAAe,CAAC;CAEzE,MAAM,QAAQ,IAAI,aAAa,gBAAgB;CAC/C,MAAM,WAAW,IAAI,aAAa,gBAAgB;CAClD,IAAI,WAAW;AACf,MAAK,IAAI,IAAI,GAAG,IAAI,WAAW,KAAK;AAClC,gBAAc,QAAQ,IAAI,WAAW,QAAQ,OAAO,SAAS;AAC7D,cAAY,WAAW,UAAU,QAAQ,QAAQ,GAAG;;AAGtD,SADY,WAAW,YACT,mBAAmB,iBAAiB;;;;;;;AAcpD,SAAgB,qBAAqB,OAAqB,IAAiC;AACzF,KAAI,MAAM,WAAW,EACnB,QAAO;EAAE,QAAQ;EAAG,mBAAmB;EAAG,OAAO;EAAG;AAEtD,QAAO;EACL,QAAQ,WAAW,UAAU,MAAM,CAAC;EACpC,mBAAmB,WAAW,iBAAiB,OAAO,GAAG,CAAC;EAC1D,OAAO,WAAW,aAAa,OAAO,GAAG,CAAC;EAC3C;;;;;;;;;;AAiBH,SAAgB,uBAAuB,YAAoB,WAAkC;CAC3F,MAAM,QAAQ,WAAW,MAAM,cAAc,CAAC,QAAQ,MAAM,EAAE,SAAS,EAAE;AACzE,KAAI,MAAM,WAAW,EACnB,QAAO,EAAE;CAEX,MAAM,UAAU,MAAM,KAAK,MAAM,EAAE,OAAO;CAC1C,MAAM,cAAc,QAAQ,QAAQ,GAAG,MAAM,IAAI,GAAG,EAAE;CACtD,MAAMG,MAAqB,EAAE;CAC7B,IAAI,MAAM;AACV,MAAK,IAAI,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;EACrC,MAAM,aAAa,MAAM,IAAI,IAAI,KAAK,MAAO,MAAM,cAAe,UAAU;AAC5E,SAAO,QAAQ;EACf,MAAM,WACJ,MAAM,MAAM,SAAS,IAAI,YAAY,KAAK,MAAO,MAAM,cAAe,UAAU;AAClF,MAAI,KAAK;GAAE,MAAM,MAAM;GAAI;GAAY,UAAU,KAAK,IAAI,UAAU,WAAW;GAAE,CAAC;;AAEpF,QAAO;;;;;;;;AAST,SAAgB,oBACd,OACA,KACA,WACc;CACd,MAAM,MAAM,IAAI,aAAa,UAAU;CACvC,MAAM,MAAM;AACZ,MAAK,IAAI,IAAI,GAAG,IAAI,WAAW,KAAK;EAClC,MAAM,QAAQ,IAAI;EAClB,MAAM,MAAM,KAAK,IAAI,QAAQ,KAAK,MAAM,OAAO;EAC/C,IAAI,MAAM;EACV,IAAI,IAAI;AACR,OAAK,IAAI,IAAI,OAAO,IAAI,KAAK,KAAK;AAChC,UAAO,MAAM,KAAK,MAAM;AACxB;;AAEF,MAAI,KAAK,IAAI,IAAI,KAAK,KAAK,MAAM,EAAE,GAAG;;AAExC,QAAO;;;;;;;;AAiBT,SAAS,iBACP,KACA,cACA,cACa;CACb,MAAM,YAAY,IAAI;AACtB,KAAI,cAAc,EAChB,QAAO,EAAE;CAEX,IAAI,OAAO;CACX,IAAI,QAAQ,OAAO;AACnB,MAAK,IAAI,IAAI,GAAG,IAAI,WAAW,KAAK;AAClC,SAAO,KAAK,IAAI,MAAM,IAAI,GAAG;AAC7B,UAAQ,KAAK,IAAI,OAAO,IAAI,GAAG;;AAEjC,KAAI,QAAQ,EACV,QAAO,EAAE;CAEX,MAAM,YAAY,SAAS,OAAO,SAAS;CAC3C,MAAMC,OAAoB,EAAE;CAC5B,IAAI,WAAW;CACf,IAAI,YAAY;AAChB,MAAK,IAAI,IAAI,GAAG,IAAI,WAAW,IAE7B,KADe,IAAI,KAAK,WACZ;AACV,MAAI,WAAW,EACb,YAAW;AAEb,cAAY;QACP;AACL;AAEA,MAAI,YAAY,KAAK,aAAa,cAAc;AAC9C,QAAK,KAAK;IAAE,OAAO;IAAU,KAAK,IAAI,YAAY;IAAG,CAAC;AACtD,cAAW;;;AAIjB,KAAI,YAAY,EACd,MAAK,KAAK;EAAE,OAAO;EAAU,KAAK;EAAW,CAAC;AAEhD,QAAO,KAAK,QAAQ,MAAM,EAAE,MAAM,EAAE,MAAM;;AAG5C,MAAM,uBAAuB;AAC7B,MAAM,wBAAwB;;;;;;;AAQ9B,SAAS,qBAAqB,SAAmB,UAA4B;CAC3E,MAAM,cAAc,QAAQ,QAAQ,GAAG,MAAM,IAAI,GAAG,EAAE,IAAI;CAC1D,MAAM,UAAU,QAAQ,KAAK,QAAQ,KAAK,IAAI,GAAG,KAAK,MAAO,MAAM,cAAe,SAAS,CAAC,CAAC;CAC7F,MAAM,QAAQ,QAAQ,KAAK,GAAG,MAAM,EAAE,CAAC,MAAM,GAAG,MAAM,QAAQ,KAAK,QAAQ,GAAG;CAC9E,IAAI,WAAW,QAAQ,QAAQ,GAAG,MAAM,IAAI,GAAG,EAAE;CACjD,IAAI,KAAK;AACT,QAAO,WAAW,UAAU;AAC1B,UAAQ,MAAM,KAAK,MAAM;AACzB;AACA;;AAEF,QAAO,WAAW,UAAU;EAC1B,MAAM,MAAM,MAAM,MAAM,SAAS,IAAK,KAAK,MAAM;AACjD,MAAI,QAAQ,OAAO,GAAG;AACpB,WAAQ;AACR;;AAEF;;AAEF,QAAO;;;;;;;AAQT,SAAS,cACP,YACA,WACA,SACA,KACM;CACN,MAAM,OAAO,KAAK,IAAI,GAAG,UAAU,UAAU;CAC7C,MAAM,QAAQ,WAAW,KAAK,MAAM,EAAE,UAAU,EAAE;CAClD,MAAM,aAAa,MAAM,QAAQ,GAAG,MAAM,IAAI,GAAG,EAAE,IAAI,WAAW;CAClE,IAAI,MAAM;AACV,MAAK,IAAI,IAAI,GAAG,IAAI,WAAW,QAAQ,KAAK;EAC1C,MAAM,SAAS,MAAM,IAAI,YAAY,YAAY,KAAK,MAAO,MAAM,aAAc,KAAK;AACtF,SAAO,MAAM;EACb,MAAM,OACJ,MAAM,WAAW,SAAS,IAAI,UAAU,YAAY,KAAK,MAAO,MAAM,aAAc,KAAK;AAC3F,MAAI,KAAK;GAAE,MAAM,WAAW;GAAI,YAAY;GAAQ,UAAU,KAAK,IAAI,MAAM,OAAO;GAAE,CAAC;;;;;;;;;;;;;;;AAgB3F,SAAgB,mBACd,YACA,WACA,OACA,KACe;CACf,MAAM,QAAQ,WAAW,MAAM,cAAc,CAAC,QAAQ,MAAM,EAAE,SAAS,EAAE;AACzE,KAAI,MAAM,WAAW,EACnB,QAAO,EAAE;AAEX,KAAI,MAAM,WAAW,KAAK,aAAa,EACrC,QAAO,uBAAuB,YAAY,UAAU;CAItD,MAAM,OAAO,iBADD,oBAAoB,OAAO,KAAK,UAAU,EACnB,sBAAsB,sBAAsB;AAE/E,KAAI,KAAK,SAAS,EAChB,QAAO,uBAAuB,YAAY,UAAU;CAItD,MAAM,UAAU,qBADA,KAAK,KAAK,MAAM,EAAE,MAAM,EAAE,MAAM,EACF,MAAM,OAAO;CAK3D,MAAMD,MAAqB,EAAE;CAC7B,IAAI,UAAU;AACd,MAAK,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;EACpC,MAAM,YAAY,MAAM,IAAI,IAAI,KAAK,GAAG;EACxC,MAAM,UAAU,MAAM,KAAK,SAAS,IAAI,YAAY,KAAK,IAAI,GAAG;AAChE,gBAAc,MAAM,MAAM,SAAS,UAAU,QAAQ,GAAG,EAAE,WAAW,SAAS,IAAI;AAClF,aAAW,QAAQ;;AAErB,QAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACrYT,MAAM,yBAAyB;AAwB/B,MAAM,mBAEJ;AACF,MAAM,cAAc;AACpB,MAAM,cAAc;AACpB,MAAM,SAAS;AACf,MAAM,cAAc;AACpB,MAAM,oBAAoB;;AAG1B,SAAS,kBAAkB,OAAuB;CAChD,IAAI,OAAO,MAAM,UAAU,OAAO;AAClC,QAAO,KAAK,QAAQ,MAAM,MAAM;AAChC,QAAO,KAAK,QAAQ,WAAW,MAAM;AACrC,QAAO,KAAK,QAAQ,aAAa,KAAI;AACrC,QAAO,KAAK,QAAQ,cAAc,IAAI;AACtC,QAAO,KAAK,QAAQ,YAAY,IAAI;AACpC,QAAO,KAAK,QAAQ,UAAU,IAAI;AAClC,QAAO,KAAK,QAAQ,kBAAkB,GAAG;AACzC,QAAO,KAAK,QAAQ,QAAQ,IAAI;AAChC,QAAO,KAAK,QAAQ,kBAAkB,KAAK;AAC3C,QAAO,KAAK,QAAQ,8BAA8B,MAAM;AACxD,QAAO,KAAK,QAAQ,oBAAoB,QAAQ;AAChD,QAAO,KAAK,QAAQ,8BAA8B,QAAQ;AAC1D,QAAO,KAAK,QAAQ,oCAAoC,QAAQ;AAChE,QAAO,KAAK,QAAQ,cAAc,KAAK;AACvC,QAAO,KAAK,QAAQ,QAAQ,IAAI;AAChC,QAAO,KAAK,QAAQ,MAAM,GAAG,CAAC,MAAM;AACpC,QAAO;;;AAIT,SAAS,cAAc,MAAsB;CAC3C,MAAM,cAAc,YAAY,KAAK,KAAK;CAC1C,MAAM,cAAc,YAAY,KAAK,KAAK;CAC1C,MAAM,SAAS,OAAO,KAAK,KAAK;AAChC,KAAI,eAAe,eAAe,OAChC,QAAO;AAET,QAAO;;;;;;;AAQT,SAAS,iBAAiB,WAAmB,aAAsD;CACjG,MAAM,KAAK,YAAY,MAAM;CAC7B,MAAM,YAAY,cAAc,GAAG;CACnC,MAAM,cAAc,cAAc,MAAM;EAAC;EAAK;EAAK;EAAK;EAAK;EAAI,GAAG;EAAC;EAAK;EAAK;EAAI;CACnF,IAAI,KAAK;AACT,KAAI,IAAI;EACN,MAAM,OAAO,GAAG,GAAG,SAAS;AAC5B,MAAI,CAAC,YAAY,SAAS,KAAK,CAC7B,MAAK;WACI,cAAc,IACvB,MAAK;;AAGT,QAAO;EAAE,QAAQ,KAAK,KAAK,UAAU,MAAM;EAAE,KAAK,GAAG,MAAM;EAAE;;;AAI/D,SAAS,kBAAkB,GAIhB;AAIT,QAAO,YAHQ,GAAG,UAAU,EAGF,wBAFf,GAAG,qBAAqB,EAEkB,YADvC,GAAG,SAAS,EAC6C;;;AAIzE,SAAS,mBAAmB,OAAkC;CAC5D,MAAME,SAAmB,EAAE;AAC3B,MAAK,MAAM,KAAK,OAAO;EACrB,MAAM,YAAY,EAAE,YAAY,GAAG,QAAQ,EAAE;EAC7C,IAAI,QAAQ,GAAG,EAAE,KAAK,kBAAkB,SAAS,IAAI,kBAAkB,EAAE,SAAS;EAClF,MAAM,IAAI,KAAK,IAAI,EAAE,GAAG,QAAQ,EAAE,GAAG,OAAO;EAC5C,IAAI,QAAQ;AACZ,OAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IACrB,UAAS,QAAQ,EAAE,GAAG,GAAG,SAAS,EAAE,GAAG,GAAG;AAE5C,WAAS,WAAW;AACpB,SAAO,KAAK,iBAAiB,MAAM,cAAc;;AAEnD,QAAO,OAAO,KAAK,KAAK;;;;;;;;;AAU1B,SAAgB,sBAAsB,MAAc,SAA+B;CACjF,MAAM,aAAa,kBAAkB,KAAK;CAC1C,IAAI,aAAa;CACjB,IAAI,QAAQ;AACZ,KAAI,SAAS;EACX,MAAM,EAAE,QAAQ,QAAQ,iBAAiB,YAAY,QAAQ,KAAK;AAClE,eAAa;AAKb,UAAQ,mBAHM,QAAQ,MAAM,KAAK,GAAG,MAClC,MAAM,QAAQ,MAAM,SAAS,IAAI;GAAE,GAAG;GAAG,MAAM,EAAE,OAAO;GAAK,GAAG,EACjE,CACgC;;CAEnC,IAAI,SAAS,+BAA+B,WAAW;AACvD,KAAI,QACF,WAAU,GAAG,MAAM;AAErB,QAAO;;;;;;;;AAST,eAAsB,gBAAgB,YAAoB,SAAwC;AAEhG,KAAI,YAAY,KAAK,WAAW,EAAE;EAChC,MAAMC,SAAO,MAAM,MAAM,WAAW;AACpC,MAAI,CAACA,OAAK,GACR,OAAM,IAAI,MAAM,mCAAmC,WAAW,IAAIA,OAAK,SAAS;AAElF,SAAQ,MAAMA,OAAK,MAAM;;AAG3B,KAAI,WAAW,SAAS,IAAI,IAAI,WAAW,SAAS,QAAQ,EAAE;EAC5D,MAAM,KAAK,MAAM,OAAO,WAAW,YAAY,KAAK;AACpD,MAAI,GACF,QAAO,KAAK,MAAM,GAAG,aAAa,YAAY,OAAO,CAAC;;AAI1D,KAAI,CAAC,QACH,OAAM,IAAI,MACR,qBAAqB,WAAW,iDACjC;CAEH,MAAM,MAAM,GAAG,QAAQ,QAAQ,mBAAmB,GAAG,CAAC,GAAG,WAAW;CACpE,MAAM,OAAO,MAAM,MAAM,IAAI;AAC7B,KAAI,CAAC,KAAK,GACR,OAAM,IAAI,MAAM,mCAAmC,WAAW,SAAS,IAAI,IAAI,KAAK,SAAS;AAE/F,QAAQ,MAAM,KAAK,MAAM;;;AAuE3B,SAASC,qBACP,OACA,SACyD;CACzD,MAAM,sBAAM,IAAI,KAAyD;AACzE,MAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,MAAM,QAAQ,EAAE;AACxD,MAAI,KAAK,YAAY,WACnB;EAEF,MAAM,IACJ,QAAQ,IAAI,KAAK,KAAK,KAAK,iBAAiB,QAAQ,IAAI,KAAK,eAAe,GAAG;AACjF,MAAI,EACF,KAAI,IAAI,MAAM,EAAE;;AAGpB,QAAO;;;AAIT,SAASC,SAAO,QAA8B;CAC5C,IAAI,OAAO;CACX,IAAI,QAAQ,OAAO;AACnB,MAAK,IAAI,IAAI,GAAG,IAAI,OAAO,QAAQ,IACjC,KAAI,OAAO,KAAK,OAAO;AACrB,UAAQ,OAAO;AACf,SAAO;;AAGX,QAAO;;;AAIT,SAAS,eAAe,QAA4B;CAClD,IAAI,OAAO,OAAO;AAClB,MAAK,MAAM,KAAK,OACd,KAAI,IAAI,KACN,QAAO;CAGX,IAAI,MAAM;AACV,MAAK,IAAI,IAAI,GAAG,IAAI,OAAO,QAAQ,KAAK;EACtC,MAAM,IAAI,KAAK,IAAI,OAAO,KAAK,KAAK;AACpC,SAAO,KAAK;AACZ,SAAO;;AAET,MAAK,IAAI,IAAI,GAAG,IAAI,OAAO,QAAQ,IACjC,QAAO,MAAM;;;AAKjB,SAAS,cAAc,OAAqB,KAAiB,MAAsB;CACjF,MAAM,IAAI,MAAM;CAChB,MAAM,QAAQ,MAAM,KAAK,EAAE,QAAQ,GAAG,GAAG,GAAG,MAAM,EAAE,CAAC,MAAM,GAAG,MAAM,MAAM,KAAK,MAAM,GAAG;CACxF,IAAI,MAAM;CACV,IAAI,SAAS;AACb,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,SAAO,MAAM,MAAM;AACnB,MAAI,OAAO,MAAM;AACf,YAAS,IAAI;AACb;;;CAGJ,IAAI,UAAU;AACd,MAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,IAC1B,YAAW,MAAM,MAAM;CAEzB,IAAIC,SAAO,KAAK,QAAQ,GAAG;AAC3B,MAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,KAAK;AAC/B,YAAQ,MAAM,MAAM;AACpB,MAAIA,UAAQ,EACV,QAAO,IAAI,MAAM;;AAGrB,QAAO,IAAI,MAAM;;;;;;;;;;;;;;;;;AAkBnB,SAAS,oBACP,QACA,GACA,WACA,SACyC;CACzC,MAAM,IAAI,OAAO;CACjB,MAAM,MAAM,KAAK,IAAI,GAAG,EAAE;CAC1B,MAAM,UAAU,IAAI,aAAa,IAAI;CACrC,MAAM,SAAS,IAAI,WAAW,IAAI;CAClC,IAAI,OAAO;CAEX,MAAM,YAAY,GAAW,MAAsB;AACjD,MAAI,WAAW,IAAI,EAAE,CACnB,QAAO,IAAI,IAAI,IAAI,UAAU,IAAI;AAEnC,SAAO;;CAET,MAAMC,mBAAuB;EAC3B,IAAI,OAAO;AACX,WAAS;GACP,IAAI,WAAW;GACf,MAAM,IAAI,IAAI,OAAO;GACrB,MAAM,IAAI,IAAI,OAAO;AACrB,OAAI,IAAI,QAAQ,QAAQ,KAAK,QAAQ,UAAW,YAAW;AAC3D,OAAI,IAAI,QAAQ,QAAQ,KAAK,QAAQ,UAAW,YAAW;AAC3D,OAAI,aAAa,KAAM;GACvB,MAAM,KAAK,QAAQ;GACnB,MAAM,KAAK,OAAO;AAClB,WAAQ,QAAQ,QAAQ;AACxB,UAAO,QAAQ,OAAO;AACtB,WAAQ,YAAY;AACpB,UAAO,YAAY;AACnB,UAAO;;;CAGX,MAAM,UAAU,UAAwB;EACtC,IAAI,OAAO;AACX,SAAO,OAAO,GAAG;GACf,MAAM,SAAU,OAAO,KAAM;AAC7B,OAAI,QAAQ,WAAW,QAAQ,MAAO;GACtC,MAAM,KAAK,QAAQ;GACnB,MAAM,KAAK,OAAO;AAClB,WAAQ,QAAQ,QAAQ;AACxB,UAAO,QAAQ,OAAO;AACtB,WAAQ,UAAU;AAClB,UAAO,UAAU;AACjB,UAAO;;;AAIX,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,MAAM,IAAI,SAAS,GAAG,OAAO,GAAG;AAChC,MAAI,OAAO,KAAK;AACd,WAAQ,QAAQ;AAChB,UAAO,QAAQ;AACf;AACA,UAAO,OAAO,EAAE;aACP,IAAI,QAAQ,IAAI;AACzB,WAAQ,KAAK;AACb,UAAO,KAAK;AACZ,eAAU;;;AAGd,QAAO;EAAE,KAAK,OAAO,SAAS,GAAG,KAAK;EAAE,MAAM,QAAQ,SAAS,GAAG,KAAK;EAAE;;;;;;;;;AAU3E,SAAS,aAAa,OAAqB,SAAiB,YAAkC;AAC5F,KAAI,YAAY,WACd,QAAO;CAET,MAAM,QAAQ,aAAa;CAC3B,MAAM,SAAS,KAAK,IAAI,GAAG,KAAK,MAAM,MAAM,SAAS,MAAM,CAAC;CAC5D,MAAM,MAAM,IAAI,aAAa,OAAO;AACpC,MAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,KAAK;EAC/B,MAAM,SAAS,IAAI;EACnB,MAAM,KAAK,KAAK,MAAM,OAAO;EAC7B,MAAM,KAAK,KAAK,IAAI,KAAK,GAAG,MAAM,SAAS,EAAE;EAC7C,MAAM,OAAO,SAAS;AACtB,MAAI,KAAK,MAAM,OAAO,IAAI,QAAQ,MAAM,MAAM;;AAEhD,QAAO;;AAGT,IAAa,UAAb,MAAa,QAAQ;CACnB,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ,aAAa;CAErB,AAAS,aAAa;;CAEtB,AAAS;CAET,AAAQ,YACN,KACA,QACA,WACA,OACA,SACA;AACA,OAAK,MAAM;AACX,OAAK,SAAS;AACd,OAAK,YAAY,OAAO;AACxB,OAAK,YAAY;AACjB,OAAK,UAAU;AACf,OAAK,QAAQ;EAEb,MAAM,QAAQ,UAAU;EACxB,MAAM,QAAQ,6BAA6B,OAAO,WAAW,QAAQ,OAAO,OAAU;AACtF,MAAI,MAEF,sBAAqB,OAAO,OAAO,gBAAgB;EAErD,MAAM,OAAO,IAAI,SAAS,KAAK,OAAO;GAAE;GAAW,QAAQ;GAAO,CAAC;AACnE,OAAK,iBAAiBH,qBAAmB,OAAO,OAAO,gBAAgB,CAAC;AACxE,OAAK,gBAAgB;AACrB,OAAK,eAAe;;CAGtB,aAAa,OAAO,SAA2C;EAC7D,MAAM,MAAM,MAAM,SAAS;EAC3B,MAAM,SAAS,MAAM,YAAY;GAC/B,MAAM,QAAQ;GACd,kBAAkB,QAAQ;GAC1B,UAAU,QAAQ;GAClB,SAAS,QAAQ;GACjB,UAAU,QAAQ;GAClB,YAAY,QAAQ;GACrB,CAAC;EACF,MAAM,SAAU,OAAO,UAAU,2BAAsC;AAEvE,SAAO,IAAI,QAAQ,KAAK,QADN,KAAK,IAAI,QAAQ,aAAa,MAAM,OAAO,EAClB,QAAQ,SAAS,MAAM,QAAQ,QAAQ;;;;;;;;;;;;;;;;;;CAmBpF,AAAQ,YAAY,MAAc,SAAiC;AACjE,SAAO,KAAK,UAAU,OAAO,sBAAsB,MAAM,WAAW,KAAK,QAAQ,CAAC;;;;;;;;;CAUpF,MAAM,MAAM,MAAc,OAAyB,EAAE,EAA4B;AAC/E,MAAI,KAAK,WACP,OAAM,IAAI,MAAM,8BAA8B;EAEhD,MAAM,YAAY,YAAY,KAAK;EACnC,MAAM,cAAc,MAAM,KAAK,eAAe,MAAM,KAAK;EAEzD,MAAM,EAAE,OAAO,cAAc,sBAAsB,YAAY;AAC/D,MAAI,cAAc,EAChB,OAAM,IAAI,MACR,0DAA0D,YAAY,OAAO,IAC9E;EAEH,MAAM,MAAM,MAAM,KAAK,YAAY,OAAO,UAAU;EAEpD,MAAM,aAAa,YAAY,KAAK,GAAG,aAAa;EAEpD,MAAM,eAAe,IAAI,SAAS;AAClC,UAAQ,IACN,oBAAoB,UAAU,SAAS,aAAa,QAAQ,EAAE,CAAC,SAAS,UAAU,QAChF,EACD,CAAC,SAAS,eAAe,WAAW,QAAQ,EAAE,CAAC,GACjD;AAED,SAAO;GAAE;GAAK,YAAY;GAAiB,QAAQ;GAAW;GAAc;;;;;;;;;;;;;;;;;;;;;;;;;;CA2B9E,MAAM,cACJ,KACsB;AACtB,MAAI,WAAW,OAAO,UAAU,IAC9B,QAAO;AAET,MAAI,KAAK,WACP,OAAM,IAAI,MAAM,8BAA8B;EAEhD,MAAM,EAAE,OAAO,YAAY,eAAe;AAC1C,MAAI,CAAC,SAAS,MAAM,WAAW,EAC7B,OAAM,IAAI,MAAM,sCAAsC;AAExD,MAAI,CAAC,cAAc,WAAW,MAAM,CAAC,WAAW,EAC9C,OAAM,IAAI,MAAM,+DAA+D;EAIjF,MAAM,OAAO,aAAa,OAAO,YAAY,gBAAgB;EAC7D,MAAM,YAAY,KAAK,KAAK,KAAK,SAAS,QAAQ,GAAG;EACrD,MAAM,MAAM,IAAI,aAAa,UAAU;AACvC,MAAI,IAAI,KAAK;EAGb,MAAM,YAAY,iBAAiB,UAAU;EAC7C,MAAM,QAAQ,MAAM,KAAK,mBAAmB,KAAK,UAAU;EAC3D,MAAM,QAAQ,MAAM,SAAS,GAAG,UAAU;EAC1C,MAAM,QAAQ,MAAM,SAAS,WAAW,IAAI,UAAU;EAMtD,MAAMI,QADU,mBAAmB,YAAY,WAAW,MAAM,QAAQ,CAC/B,KAAK,MAAM;GAClD,MAAM,cAAc,KAAK,IAAI,EAAE,aAAa,SAAS,KAAK,OAAO;GACjE,MAAM,YAAY,KAAK,IAAI,EAAE,WAAW,SAAS,KAAK,OAAO;GAE7D,MAAM,WAAW,qBADC,KAAK,SAAS,aAAa,UAAU,EACN,gBAAgB;AACjE,UAAO;IACL,MAAM,EAAE;IACR,UAAU,KAAK,OAAQ,EAAE,WAAW,EAAE,cAAc,UAAW,IAAI,GAAG;IACtE,IAAI,MAAM,KAAK,MAAM,SAAS,EAAE,YAAY,EAAE,SAAS,CAAC;IACxD,IAAI,MAAM,KAAK,MAAM,SAAS,EAAE,YAAY,EAAE,SAAS,CAAC;IACxD;IACD;IACD;AAEF,SAAO;GAAE,MAAM,WAAW,MAAM;GAAE;GAAO;;;CAI3C,MAAc,mBAAmB,KAAmB,WAAyC;EAC3F,MAAM,QAAQ,8BAA8B,EAAE,YAAY,IAAI,QAAQ,CAAC;EACvE,MAAM,OAAO,IAAI,SAAS,KAAK,KAAK,OAAO;GAAE,WAAW;GAAW,QAAQ;GAAO,CAAC;AACnF,MAAI;AACF,QAAK,iBAAiBJ,qBAAmB,OAAO,KAAK,OAAO,aAAa,CAAC;AAC1E,QAAK,gBAAgB;AACrB,QAAK,OAAO;AACZ,QAAK,WAAW,UAAU,IAAI;GAE9B,MAAM,MAAM,MAAM,KAAK,eAAe,aAAa,oBAAoB,UAAU;AACjF,UAAO,IAAI,YAAY,IAAI,QAAQ,IAAI,YAAY,oBAAoB,UAAU,CAAC,OAAO;YACjF;AACR,QAAK,SAAS;;;;CAKlB,MAAc,eAAe,MAAc,MAA2C;EACpF,MAAM,SAAS,KAAK,YAAY,MAAM,KAAK,QAAQ;AACnD,MAAI,OAAO,UAAU,KAAK,UACxB,OAAM,IAAI,MAAM,qBAAqB,OAAO,OAAO,uBAAuB,KAAK,UAAU,GAAG;AAE9F,OAAK,aAAa,OAAO;EACzB,MAAM,EAAE,WAAW,MAAM,KAAK,aAAa,QAAQ,IAAI,YAAY,OAAO,CAAC;AAC3E,SAAO,KAAK,cAAc,QAAQ,QAAQ;GACxC,aAAa,KAAK,eAAe;GACjC,MAAM,KAAK,QAAQ;GACnB,MAAM,KAAK,QAAQ;GACnB,mBAAmB,KAAK,qBAAqB;GAC7C,iBAAiB,KAAK,mBAAmB;GACzC,cAAc,KAAK,IAAI,KAAK,gBAAgB,wBAAwB,KAAK,UAAU;GACnF,QAAQ,KAAK,UAAU;GACxB,CAAC;;;CAIJ,MAAM,sBAAsB,MAAc,OAAyB,EAAE,EAAqB;AACxF,SAAO,KAAK,eAAe,MAAM,KAAK;;;;;;;;;CAUxC,MAAc,cACZ,QACA,eACA,GASmB;EACnB,MAAMK,MAAgB,CAAC,GAAG,OAAO;EACjC,MAAMC,YAAsB,EAAE;EAC9B,IAAI,SAAS;AAEb,OAAK,IAAI,OAAO,GAAG,OAAO,EAAE,cAAc,QAAQ;GAShD,MAAM,OAAO,EAAE,SACXL,SAAO,OAAO,GACd,KAAK,YAAY,QAAQ;IACvB,aAAa,EAAE;IACf,MAAM,EAAE;IACR,MAAM,EAAE;IACR,mBAAmB,EAAE;IACrB,UAAU,IAAI,MAAM,CAAC,EAAE,gBAAgB;IACxC,CAAC;AAEN,OAAI,SAAS,qBAAqB,SAAS,eACzC;AAEF,aAAU,KAAK,KAAK;AACpB,OAAI,KAAK,KAAK;AAEd,OAAI,IAAI,UAAU,KAAK,UACrB;AAGF,aADY,MAAM,KAAK,aAAa,QAAQ,IAAI,YAAY,CAAC,KAAK,CAAC,CAAC,EACvD;;AAEf,SAAO;;;;;;;CAQT,AAAQ,YACN,QACA,GAOQ;EACR,MAAM,IAAI,OAAO;EACjB,MAAM,OAAO,EAAE,sBAAsB,IAAM,IAAI,IAAI,EAAE,SAAS,GAAG;AAKjE,MAAI,EAAE,OAAO,KAAK,EAAE,OAAO,GAAG;GAC5B,MAAM,EAAE,KAAK,MAAM,MAAM,UAAU,oBACjC,QACA,EAAE,MACF,MACA,EAAE,kBACH;GACD,MAAMM,WAAS,IAAI,aAAa,MAAM,OAAO;AAC7C,QAAK,IAAI,IAAI,GAAG,IAAI,MAAM,QAAQ,IAChC,UAAO,KAAK,MAAM,KAAK,EAAE;AAE3B,kBAAeA,SAAO;AACtB,UAAO,cAAcA,UAAQ,WAAW,KAAK,KAAK,EAAE,EAAE,KAAK;;EAI7D,MAAM,MAAM,IAAI,WAAW,EAAE;EAC7B,MAAM,SAAS,IAAI,aAAa,EAAE;AAClC,OAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,OAAI,KAAK;GACT,IAAI,IAAI,OAAO;AACf,OAAI,MAAM,IAAI,EAAE,CACd,KAAI,IAAI,IAAI,IAAI,EAAE,oBAAoB,IAAI,EAAE;AAE9C,UAAO,KAAK,IAAI,EAAE;;AAEpB,iBAAe,OAAO;AACtB,SAAO,cAAc,QAAQ,KAAK,EAAE,KAAK;;;;;;;;;;CAW3C,MAAc,YAAY,OAAoB,WAA0C;EACtF,MAAM,QAAQ,8BAA8B,EAAE,WAAW,CAAC;EAC1D,MAAM,SAAS,gBAAgB,UAAU;EACzC,MAAM,OAAO,IAAI,SAAS,KAAK,KAAK,OAAO;GAAE,WAAW;GAAW,QAAQ;GAAO,CAAC;AACnF,MAAI;AACF,QAAK,iBAAiBP,qBAAmB,OAAO,KAAK,OAAO,aAAa,CAAC;AAC1E,QAAK,gBAAgB;AACrB,QAAK,OAAO;AACZ,QAAK,WAAW,eAAe,MAAM;GACrC,MAAM,MAAM,MAAM,KAAK,eAAe,MAAM,QAAQ,IAAI,OAAO;GAC/D,MAAM,MAAM,IAAI,aAAa,IAAI,OAAO;AACxC,QAAK,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,IAC9B,KAAI,KAAK,KAAK,KAAK,IAAI,GAAG;AAE5B,UAAO;YACC;AACR,QAAK,SAAS;;;CAIlB,UAAgB;AACd,MAAI,KAAK,WACP;AAEF,OAAK,aAAa;AAClB,OAAK,aAAa,SAAS;AAC3B,OAAK,OAAO,gBAAgB,OAAO;AACnC,OAAK,OAAO,aAAa,OAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AC3wBpC,MAAMQ,kBAAgB;;AAEtB,MAAM,4BACJ;;AAEF,MAAM,qBAAqB;;AA0C3B,SAAS,mBACP,OACA,SACyD;CACzD,MAAM,sBAAM,IAAI,KAAyD;AACzE,MAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,MAAM,QAAQ,EAAE;AACxD,MAAI,KAAK,YAAY,WACnB;EAEF,MAAM,IACJ,QAAQ,IAAI,KAAK,KAAK,KAAK,iBAAiB,QAAQ,IAAI,KAAK,eAAe,GAAG;AACjF,MAAI,EACF,KAAI,IAAI,MAAM,EAAE;;AAGpB,QAAO;;;AAIT,SAAS,OAAO,QAA8B;CAC5C,IAAI,OAAO;CACX,IAAI,QAAQ,OAAO;AACnB,MAAK,IAAI,IAAI,GAAG,IAAI,OAAO,QAAQ,IACjC,KAAI,OAAO,KAAK,OAAO;AACrB,UAAQ,OAAO;AACf,SAAO;;AAGX,QAAO;;;AAIT,SAAS,aAAa,QAAsB,aAAqB,MAAsB;CACrF,MAAM,IAAI,OAAO;CACjB,MAAM,SAAS,IAAI,aAAa,EAAE;CAClC,IAAI,OAAO,OAAO;CAClB,MAAM,OAAO,IAAI,KAAK,IAAI,aAAa,KAAK;AAC5C,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,SAAO,KAAK,OAAO,KAAK;AACxB,MAAI,OAAO,KAAK,KACd,QAAO,OAAO;;CAIlB,IAAIC,UAA6B;AACjC,KAAI,OAAO,KAAK,OAAO,GAAG;EACxB,MAAM,QAAQ,MAAM,KAAK,EAAE,QAAQ,GAAG,GAAG,GAAG,MAAM,EAAE,CAAC,MAAM,GAAG,MAAM,OAAO,KAAK,OAAO,GAAG;AAC1F,YAAU,IAAI,WAAW,MAAM,MAAM,GAAG,KAAK,CAAC;;CAEhD,IAAI,MAAM;CACV,MAAM,QAAQ,IAAI,aAAa,EAAE;AACjC,KAAI,QACF,MAAK,MAAM,OAAO,SAAS;EACzB,MAAM,IAAI,KAAK,IAAI,OAAO,OAAO,KAAK;AACtC,QAAM,OAAO;AACb,SAAO;;KAGT,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,MAAM,IAAI,KAAK,IAAI,OAAO,KAAK,KAAK;AACpC,QAAM,KAAK;AACX,SAAO;;CAGX,IAAIC,SAAO,KAAK,QAAQ,GAAG;AAC3B,KAAI,SAAS;AACX,OAAK,MAAM,OAAO,SAAS;AACzB,aAAQ,MAAM;AACd,OAAIA,UAAQ,EACV,QAAO;;AAGX,SAAO,QAAQ;;AAEjB,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,YAAQ,MAAM;AACd,MAAIA,UAAQ,EACV,QAAO;;AAGX,QAAO,IAAI;;AAgBb,IAAM,wBAAN,MAA4B;CAC1B,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ,gCAAgB,IAAI,KAAwB;CACpD,AAAQ,oCAAoB,IAAI,KAAwB;CACxD,AAAQ,aAA4B,EAAE;CAEtC,YAAY,KAAiB,OAAmB,WAAmB,OAAe;AAChF,OAAK,MAAM;AACX,OAAK,QAAQ;AACb,OAAK,YAAY;AACjB,OAAK,SAAS,MAAM,OAAO;AAC3B,OAAK,QAAQ;AACb,OAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,MAAM,QAAQ,EAAE;AACxD,OAAI,KAAK,YAAY,aACnB;GAGF,MAAM,QAAQ,KAAK,MAAM,QACtB,GAAW,QAAQ,KAAK,OAAO,QAAQ,WAAW,MAAM,QACzD,EACD;AACD,QAAK,kBAAkB,IACrB,MACA,oBAAoB,KAAK,KAAK,QAAQ,QAAQ,QAAQ,YAAY,KAAK,OAAO,CAC/E;;;CAIL,cAAc,SAAwE;AACpF,OAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,KAAK,MAAM,QAAQ,EAAE;AAC7D,OAAI,KAAK,YAAY,WACnB;GAEF,MAAM,IAAI,QAAQ,IAAI,KAAK;AAC3B,OAAI,CAAC,EACH,OAAM,IAAI,MAAM,0CAA0C,KAAK,GAAG;AAEpE,QAAK,cAAc,IACjB,MACA,oBAAoB,KAAK,KAAK,SAAS,QAAQ,EAAE,KAAK,YAAY,EAAE,KAAK,CAC1E;;;CAIL,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,MAAM,OAAO,gBAAgB,KAAK;AAClC,OAAI,CAAC,KACH,OAAM,IAAI,MAAM,4CAA4C,KAAK,OAAO,GAAG;GAE7E,MAAM,WAAW,oBACf,KAAK,KACL,SAAS,UACT,KAAK,YACL,KAAK,WACN;GACD,MAAM,aAAa,KAAK,YAAY,MAAM,QAAQ,EAAE,QAAQ,GAAG,CAAC;GAChE,MAAM,gBAAgB,oBAAoB,KAAK,KAAK,SAAS,UAAU,WAAW;GAClF,MAAM,gBAAgB,KAAK,cAAc,MAAM,MAAM,cAAc;GACnE,MAAM,YAAY,gBAAgB,KAAK,KAAK,UAAU,eAAe,UAAU,SAAS;AACxF,QAAK,WAAW,KAAK;IAAE;IAAM;IAAM;IAAU;IAAW;IAAe,CAAC;;;;CAK5E,MAAM,OACJ,UACA,SACyD;EACzD,MAAM,IAAI,KAAK,IAAI,OAAO;EAC1B,MAAM,WAAW,KAAK,kBAAkB,IAAI,YAAY;EACxD,MAAM,UAAU,KAAK,kBAAkB,IAAI,cAAc;AACzD,MAAI,EAAE,YAAY,SAChB,OAAM,IAAI,MAAM,kEAAkE;AAEpF,IAAE,YAAY,UAAU,GAAG,SAAyB;AACpD,IAAE,YAAY,SAAS,GAAG,QAAwB;EAElD,MAAM,SAAS,KAAK,eAAe;EACnC,MAAMC,QAAyC,EAAE;AACjD,OAAK,MAAM,KAAK,KAAK,YAAY;GAC/B,MAAM,SAAS,EAAE,KAAK,YAAY,EAAE,MAAM,QAAQ,EAAE,QAAQ,GAAG,CAAC;AAChE,KAAE,YAAY,EAAE,eAAe,GAAG,OAAO;AACzC,SAAM,KAAK,EAAE,KAAK,gBAAgB,EAAE,MAAM,QAAQ,EAAE,QAAQ,GAAG,CAAC,CAAC;;AAGnE,MAAI,KAAK,IAAI,eACX,MAAK,IAAI,IAAI,GAAG,IAAI,KAAK,WAAW,QAAQ,KAAK;GAC/C,MAAM,MAAM,KAAK,IAAI,OAAO,sBAAsB;GAClD,MAAM,IAAI,IAAI,kBAAkB;AAChC,KAAE,YAAY,KAAK,WAAW,GAAG,SAAS;AAC1C,KAAE,aAAa,GAAG,KAAK,WAAW,GAAG,UAAU;AAC/C,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,iBAAiB,CAAC;GAC5E,MAAM,OAAO,IAAI,iBAAiB,EAAE,OAAO,mBAAmB,CAAC;AAC/D,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,MAAMC,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,KAAK,QAAQ,KAAK,OAAO,CAAC;AAC3E,QAAK,KAAK,MAAM,KAAK,SAAS,cAAc,KAAK,KAAK,QAAQ,KAAK,OAAO,CAAC;;AAE7E,SAAO;GAAE;GAAM;GAAM;;CAGvB,UAAgB;AACd,iBAAe,CAAC,GAAG,KAAK,cAAc,QAAQ,CAAC,CAAC;AAChD,iBAAe,CAAC,GAAG,KAAK,kBAAkB,QAAQ,CAAC,CAAC;AACpD,OAAK,MAAM,KAAK,KAAK,WACnB,GAAE,cAAc,SAAS;AAE3B,OAAK,cAAc,OAAO;AAC1B,OAAK,kBAAkB,OAAO;AAC9B,OAAK,aAAa,EAAE;;CAGtB,AAAQ,gBAA0C;EAChD,MAAMC,WAAqC,EAAE;AAC7C,OAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,KAAK,MAAM,QAAQ,CAC3D,UAAS,QAAQ,KAAK,MAAM,KAAK,QAAS,OAAO,QAAQ,WAAW,MAAM,KAAK,MAAO;AAExF,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,MAAM,yCAAyC,WAAW,UAAU,KAAK,KAAK;AAE1F,WAAQ,KAAK,EAAE,QAAQ,CAAC;;AAG5B,SAAO;;CAGT,MAAc,SAAS,MAAc,UAAyC;EAC5E,MAAM,SAAS,KAAK,kBAAkB,IAAI,KAAK;AAC/C,MAAI,CAAC,OACH,OAAM,IAAI,MAAM,qCAAqC,KAAK,GAAG;EAE/D,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,SAASP,iBAAe,GAAG,QAAQ;EAClD,MAAM,OAAO,IAAI,aAAa,SAAS,eAAe,GAAG,QAAQ,CAAC,MAAM,EAAE,CAAC;AAC3E,WAAS,OAAO;AAChB,WAAS,SAAS;AAClB,SAAO;;;AAQX,IAAa,YAAb,MAAa,UAAU;CACrB,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,AAAQ,aAAa;CAErB,AAAS,aAAa;CAEtB,AAAQ,YAAY,KAAiB,QAAyB,WAAmB;AAC/E,OAAK,MAAM;AACX,OAAK,SAAS;AACd,OAAK,MAAM,kBAAkB,OAAO,UAAU;AAC9C,OAAK,YAAY;;CAGnB,aAAa,OAAO,UAA4B,EAAE,EAAsB;EACtE,MAAM,MAAM,MAAM,SAAS;EAC3B,MAAM,SAAS,MAAM,cAAc;GACjC,MAAM,QAAQ;GACd,UAAU,QAAQ;GAClB,SAAS,QAAQ;GACjB,UAAU,QAAQ;GAClB,YAAY,QAAQ;GACrB,CAAC;EACF,MAAM,MAAM,kBAAkB,OAAO,UAAU;AAE/C,SAAO,IAAI,UAAU,KAAK,QADR,KAAK,IAAI,QAAQ,aAAa,MAAM,IAAI,aAAa,CAC3B;;;;;;;;CAS9C,OAAO,yBAAiC;EACtC,MAAM,SAAS,KAAK,CAAC,QAAQ,SAAS,CAAC;AAQvC,SAAO,KAPM,KAAK;GAAC;GAAQ;GAAU;GAAQ;GAAa;GAAU;GAAa;GAAO,CAAC,CAOxE,GAAG,OAAO,uBANd,KAAK;GAAC;GAAiB;GAAY;GAAgB,CAAC,CAMV,cALvC,KAAK;GACnB;GACA;GACA;GACD,CAAC,CAC2E;;;CAI/E,MAAM,MAAM,MAAc,OAA2B,EAAE,EAA8B;AACnF,MAAI,KAAK,WACP,OAAM,IAAI,MAAM,gCAAgC;EAElD,MAAM,YAAY,YAAY,KAAK;EACnC,MAAM,eAAe,KAAK,iBAAiB,2BAA2B,MAAM;EAG5E,MAAM,UAAU,KAAK,OAAO,UAAU,OAAO,YAAY;EACzD,MAAM,QAAQ,QAAQ;EACtB,MAAM,YAAY,KAAK,OAAO,aAAa,IACzC,oFACD;AACD,MAAI,CAAC,UACH,OAAM,IAAI,MAAM,qEAAqE;EAEvF,MAAM,UAAU,oBACd,UAAU,MACV,KAAK,IAAI,iBACT,KAAK,IAAI,kBACT,KAAK,IAAI,WACT,MACD;EAED,MAAM,WAAW,2BAA2B,KAAK,OAAO,WAAW,MAAM;EACzE,MAAM,UAAU,IAAI,sBAAsB,KAAK,KAAK,UAAU,KAAK,IAAI,YAAY,MAAM;EACzF,IAAII;EACJ,IAAIC;AACJ,MAAI;AACF,WAAQ,cAAc,mBAAmB,UAAU,KAAK,OAAO,aAAa,CAAC;AAC7E,WAAQ,gBAAgB;GACxB,MAAM,MAAM,MAAM,QAAQ,OAAO,IAAI,YAAY,QAAQ,EAAE,QAAQ;AACnE,UAAO,IAAI;AACX,UAAO,IAAI;YACH;AACR,WAAQ,SAAS;;EAInB,MAAM,YAAY,KAAK,OAAO,UAAU,OAAO,KAAK;EACpD,MAAM,UAAU,MAAM,KAAK,kBAAkB,WAAW,MAAM,MAAM,OAAO,KAAK;EAGhF,MAAM,EAAE,OAAO,WAAW,mBACxB,QAAQ,OACR,KAAK,IAAI,eACT,QAAQ,MACT;AACD,MAAI,UAAU,EACZ,OAAM,IAAI,MAAM,6CAA6C,QAAQ,MAAM,IAAI;AAEjF,OAAK,UAAU,OAAO,KAAK,IAAI,eAAe,OAAO;EACrD,MAAM,MAAM,MAAM,KAAK,YAAY,OAAO,OAAO;EAEjD,MAAM,aAAa,YAAY,KAAK,GAAG,aAAa;EACpD,MAAM,eAAe,IAAI,SAAS;AAClC,UAAQ,IACN,iBAAiB,MAAM,UAAU,OAAO,SAAS,aAAa,QAAQ,EAAE,CAAC,SAAS,UAAU,QAAQ,EAAE,CAAC,GACxG;AACD,SAAO;GAAE;GAAK,YAAY;GAAoB;GAAQ;GAAc;;;;;;;;;;;;CAatE,MAAc,kBACZ,WACA,MACA,MACA,MACA,MACgD;EAChD,MAAM,MAAM,KAAK;EACjB,MAAM,IAAI,IAAI;EACd,MAAM,MAAM,IAAI;EAChB,MAAM,QAAQ,2BAA2B,KAAK,OAAO,WAAW,KAAK;EAKrE,MAAM,WAJY,KAAK,IACrB,KAAK,aAAa,oBAClB,KAAK,YAAY,UAAU,OAC5B,IAC6B,MAAM;EACpC,MAAM,OAAO,IAAI,SAAS,KAAK,KAAK,OAAO;GACzC,WAAW,KAAK;GAChB,QAAQ;GACT,CAAC;EAGF,MAAMG,cAA8B,EAAE;AACtC,OAAK,IAAI,KAAK,GAAG,KAAK,KAAK,MAAM;GAC/B,MAAM,IAAI,KAAK,OAAO,aAAa,IAAI,sCAAsC,GAAG,SAAS;AACzF,OAAI,CAAC,EACH,OAAM,IAAI,MAAM,2BAA2B,GAAG,WAAW;AAE3D,eAAY,KAAK,EAAE,KAAqB;;EAE1C,MAAM,iBAAiB,KAAK,OAAO,aAAa,IAC9C,gDACD;EACD,MAAM,eAAe,KAAK,OAAO,aAAa,IAAI,uBAAuB;AACzE,MAAI,EAAE,kBAAkB,cACtB,OAAM,IAAI,MAAM,sDAAsD;EAExE,MAAM,WAAW,eAAe;EAChC,MAAM,cAAc,aAAa;EAEjC,MAAM,OAAO,IAAI,YAAY,MAAM,SAAS;EAC5C,IAAI,QAAQ;AACZ,MAAI;AACF,QAAK,iBAAiB,mBAAmB,OAAO,KAAK,OAAO,aAAa,CAAC;AAC1E,QAAK,gBAAgB;AACrB,QAAK,IAAI,IAAI,GAAG,IAAI,IAAI,YAAY,KAAK;AACvC,SAAK,WAAW,cAAc,KAAK,KAAK,GAAG;AAC3C,SAAK,WAAW,cAAc,KAAK,KAAK,GAAG;;AAE7C,QAAK,OAAO;GAEZ,IAAI,MAAM;GACV,MAAM,MAAM,IAAI,aAAa,EAAE;GAC/B,MAAM,eAAe,MAAM,QAAQ,KAAK,UAAU;IAAE;IAAM,WAAW,IAAI;IAAa,EAAE;AAMxF,QAAK,MAAM,WAAW,WAAW;AAC/B,WAAO,KAAK,aAAa,SAAS,GAAG,KAAK;AAC1C,gBAAY,KAAK,UAAU,KAAK,EAAE;AAClC,SAAK,WAAW,mBAAmB,IAAI;AACvC,UAAM,KAAK,aAAa,CAAC,aAAa,GAAG,EAAE,EAAE;AAC7C;;GAeF,IAAI,YAAY,IAAI,YAAY,IAAI,CAAC,KAAK,oBAAoB;GAC9D,IAAI,WAAW;GACf,IAAI,WAAW;AACf,QAAK,IAAI,OAAO,GAAG,OAAO,YAAY,CAAC,UAAU,QAAQ;AAEvD,QAAI,KAAK,EAAE;AACX,SAAK,IAAI,KAAK,GAAG,KAAK,KAAK,KACzB,QAAO,KAAK,YAAY,KAAK,UAAU,KAAK,GAAG,MAAM;AAEvD,gBAAY,KAAK,UAAU,KAAK,EAAE;AAClC,SAAK,WAAW,mBAAmB,IAAI;IACvC,MAAM,YAAY,MAAM,KAAK,aAAa,cAAc,EAAE;AAC1D;IAEA,MAAM,YAAY,IAAI,YAAY,IAAI;AACtC,SAAK,IAAI,KAAK,GAAG,KAAK,KAAK,MAAM;AAC/B,SAAI,OAAO,IAAI;AACb,gBAAU,MAAM;AAChB;;AAGF,SAAI,KAAK,SACP,WAAU,IAAI,uBAAuB,OAAO;KAE9C,MAAM,OAAO,KAAK,SACd,OAAO,UAAU,IAAI,GACrB,aAAa,UAAU,KAAK,KAAK,eAAe,GAAK,KAAK,QAAQ,EAAE;AACxE,eAAU,MAAM;AAChB,UAAK,KAAK,WAAW,QAAQ,OAAO,uBAAuB,OAAO;;AAIpE,QAAI,WAAW,MAAM,KAAK,UAAU,cAAc,oBAChD;aACS,aAAa,MAAM,KAAK,UAAU,cAAc,oBACzD,YAAW;AAGb,YAAQ,OAAO;AACf,gBAAY;AACZ,QAAI,OAAO,KAAK,UACd;;YAGI;AACR,QAAK,SAAS;;EAKhB,MAAM,QAAQ,IAAI,YAAY,MAAM,MAAM;AAC1C,OAAK,IAAI,KAAK,GAAG,KAAK,KAAK,KACzB,MAAK,IAAI,IAAI,GAAG,IAAI,OAAO,IACzB,OAAM,KAAK,QAAQ,KAAK,KAAK,KAAK,WAAW;AAGjD,SAAO;GAAE,OAAO;GAAO;GAAO;;;CAIhC,MAAc,YAAY,OAAoB,WAA0C;EACtF,MAAM,QAAQ,8BAA8B;GAC1C;GACA,cAAc,KAAK,IAAI;GACvB,cAAc;GACd,YAAY;GACZ,WAAW;GACZ,CAAC;EACF,MAAM,SAAS,gBAAgB,WAAW,qBAAqB;EAC/D,MAAM,OAAO,IAAI,SAAS,KAAK,KAAK,OAAO;GAAE,WAAW;GAAW,QAAQ;GAAO,CAAC;AACnF,MAAI;AACF,QAAK,iBAAiB,mBAAmB,OAAO,KAAK,OAAO,aAAa,CAAC;AAC1E,QAAK,gBAAgB;AACrB,QAAK,OAAO;AACZ,QAAK,WAAW,eAAe,MAAM;GACrC,MAAM,MAAM,MAAM,KAAK,eAAe,MAAM,QAAQ,IAAI,OAAO;GAC/D,MAAM,MAAM,IAAI,aAAa,IAAI,OAAO;AACxC,QAAK,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,IAC9B,KAAI,KAAK,KAAK,KAAK,IAAI,GAAG;AAE5B,UAAO;YACC;AACR,QAAK,SAAS;;;CAIlB,UAAgB;AACd,MAAI,KAAK,WACP;AAEF,OAAK,aAAa;AAClB,OAAK,OAAO,aAAa,OAAO;AAChC,OAAK,OAAO,aAAa,OAAO;;;AAIpC,MAAM,uBAAuB;;AAG7B,SAAS,KAAQ,KAAsB;AACrC,QAAO,IAAI,KAAK,MAAM,KAAK,QAAQ,GAAG,IAAI,OAAO;;;AAInD,SAAS,OACP,KACA,OACA,SACA,GACA,KACM;CACN,MAAM,OAAO,UAAU;AACvB,KAAI,IACF,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IACrB,KAAI,KAAK,MAAM,OAAO;KAGxB,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IACrB,KAAI,MAAM,MAAM,OAAO;;;AAM7B,SAAS,YAAY,KAAmB,UAAwB,KAAa,GAAiB;CAC5F,MAAM,OAAO,MAAM;AACnB,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IACrB,KAAI,MAAM,SAAS,OAAO;;;;;ACzrB9B,MAAM,gBAAgB;;AAGtB,MAAM,8BAAc,IAAI,YAAY,EAAE;AACtC,MAAM,UAAU,IAAI,aAAa,YAAY;AAC7C,MAAM,UAAU,IAAI,YAAY,YAAY;AAC5C,SAAS,QAAQ,KAAgC;CAC/C,MAAM,MAAM,IAAI,YAAY,IAAI,OAAO;AACvC,MAAK,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;AACnC,UAAQ,KAAK,IAAI;EACjB,MAAM,IAAI,QAAQ;EAClB,MAAM,OAAQ,MAAM,KAAM;EAC1B,MAAM,OAAQ,MAAM,KAAM,OAAQ;EAClC,MAAM,OAAO,IAAI;AACjB,MAAI,OAAO,EACT,KAAI,KAAK;WACA,OAAO,GAChB,KAAI,KAAK,OAAO;OACX;GACL,MAAM,SAAS,OAAO,UAAY;GAClC,IAAI,OAAO,OAAQ,OAAO,KAAO,SAAS;AAC1C,OAAI,MAAO,SAAQ;AACnB,OAAI,KAAK;;;AAGb,QAAO;;;;;;;;;;;;AA6CT,MAAM,yBAAyB;AAE/B,IAAa,iBAAb,MAA4B;CAC1B,AAAQ;CACR,AAAQ;CACR,AAAQ;CAER,AAAQ,gCAAgB,IAAI,KAAwB;CACpD,AAAQ,oCAAoB,IAAI,KAAwB;CACxD,AAAQ,aAA+B,EAAE;CACzC,AAAQ;;CAER,AAAQ,iCAAiB,IAAI,KAAa;;CAG1C,AAAQ,WAAW;;CAEnB,AAAiB;CAEjB,YAAY,KAAiB,OAAmB,YAAoB;AAClE,OAAK,MAAM;AACX,OAAK,QAAQ;AACb,OAAK,aAAa;AAClB,OAAK,WAAW,MAAM,iBAAiB;EAGvC,MAAM,WAAW,MAAM,QAAQ,eAAe,gBAAgB,MAAM;EACpE,MAAM,SAAS,MAAM,OAAO;AAC5B,OAAK,YAAY,YAAY,SAAS,WAAW,SAAS;AAC1D,MAAI,IAAI,QACN;QAAK,MAAM,QAAQ,MAAM,MACvB,KAAI,KAAK,WAAW,gBAAgB,KAAK,OAAO,GAC9C,MAAK,eAAe,IAAI,KAAK,OAAO,GAAG;;AAI7C,OAAK,2BAA2B;;CAGlC,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,CAAC,EAAG,OAAM,IAAI,MAAM,mCAAmC,KAAK,GAAG;GACnE,IAAIC,OAAwB,EAAE;AAC9B,OAAI,KAAK,eAAe,IAAI,KAAK,IAAI,EAAE,gBAAgB,aACrD,QAAO,QAAQ,EAAE,KAAK;GAExB,MAAM,SAAS,oBAAoB,KAAK,KAAK,WAAW,QAAQ,KAAK,YAAY,KAAK;AACtF,QAAK,cAAc,IAAI,MAAM,OAAO;;;CAIxC,iBAAuB;EACrB,MAAM,cAAc,KAAK,cAAc,KAAK,WAAW;AACvD,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,qCAAqC,KAAK,OAAO,GAAG;AAC/E,OAAI,KAAK,WAAW,gBAAgB,KAAK,eAAe,IAAI,KAAK,OAAO,GAAG,CACzE,QAAO;GAGT,MAAM,WAAW,oBACf,KAAK,KACL,WAAW,UACX,KAAK,YACL,KAAK,WACN;GACD,MAAM,aAAa,KAAK,YAAY,MAAM,aAAa,EAAE,QAAQ,GAAG,CAAC;GACrE,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;;;;;;;;;CAU5E,MAAM,OACJ,QACA,SAC8D;EAC9D,MAAM,IAAI,OAAO;AACjB,MAAI,IAAI,KAAK,WACX,OAAM,IAAI,MAAM,mBAAmB,EAAE,8BAA8B,KAAK,aAAa;EAEvF,MAAM,IAAI,KAAK,IAAI,OAAO;AAC1B,IAAE,YAAY,KAAK,kBAAkB,IAAI,cAAc,EAAG,GAAG,OAAO,QAAwB;AAC5F,IAAE,YACA,KAAK,kBAAkB,IAAI,iBAAiB,EAC5C,GACA,OAAO,UACR;AACD,IAAE,YAAY,KAAK,kBAAkB,IAAI,UAAU,EAAG,GAAG,OAAO,IAAoB;AACpF,IAAE,YAAY,KAAK,kBAAkB,IAAI,UAAU,EAAG,GAAG,OAAO,IAAoB;EAEpF,MAAM,iBAAiB,KAAK,cAAc,EAAE;EAC5C,MAAM,QAAQ,EAAE,QAAQ,GAAG;EAK3B,MAAMC,QAAyC,EAAE;AACjD,OAAK,MAAM,KAAK,KAAK,YAAY;GAC/B,MAAM,SAAS,EAAE,KAAK,YAAY,EAAE,MAAM,gBAAgB,MAAM;AAChE,KAAE,YAAY,EAAE,eAAe,GAAG,OAAO;AACzC,SAAM,KAAK,EAAE,KAAK,gBAAgB,EAAE,MAAM,gBAAgB,MAAM,CAAC;;EAGnE,MAAM,eAAe,KAAK,IAAI;EAQ9B,MAAM,WAAY,WAA8C,oBAAoB;EACpF,IAAIC,WAAqC;EACzC,IAAIC,YAA0C;EAC9C,MAAM,aAAa,YAAY;AAC7B,OAAI,CAAC,YAAY,CAAC,UAAW;AAC7B,aAAU,KAAK;AACf,KAAE,OAAO,CAAC,SAAS,QAAQ,CAAC,CAAC;AAC7B,SAAM,EAAE,qBAAqB;AAC7B,cAAW;AACX,eAAY;;AAEd,MAAI,cAAc;AAChB,aAAU,oBAAoB,EAAE,OAAO,KAAK,WAAW,QAAQ,CAAC;AAChE,QAAK,IAAI,IAAI,GAAG,IAAI,KAAK,WAAW,QAAQ,KAAK;IAC/C,MAAM,IAAI,KAAK,WAAW;AAC1B,QAAI,EAAE,KAAK,WAAW,aAAa;AACjC,WAAM,YAAY;KAKlB,MAAM,CAAC,OAAO,OAAO,SAAS,MAAM;KACpC,MAAM,QAAQ,OAAO,SAAS,EAAE,KAAK,GAAG,QAAQ,QAAQ,GAAG,EAAE,GAAG;AAChE,SAAI,CAAC,OAAO,MAAM,MAAM,CAAE,WAAU,YAAY,EAAE,OAAO,CAAC;AAC1D,UAAK,IAAI,QAAQ,GAAG,QAAQ,OAAO,SAAS,wBAAwB;MAClE,MAAMC,SAAO,KAAK,IAAI,wBAAwB,QAAQ,MAAM;MAC5D,MAAM,SAAS,EAAE,KAAK,YAAY,EAAE,MAAM,gBAAgB;OACxD,QAAQ;OACR,SAAS;OACV,CAAC;AACF,QAAE,YAAY,EAAE,eAAe,GAAG,OAAO;MACzC,MAAMC,QAAM,KAAK,IAAI,OAAO,sBAAsB;MAClD,MAAMC,MAAID,MAAI,kBAAkB;AAChC,UAAE,YAAY,EAAE,SAAS;AACzB,UAAE,aAAa,GAAG,EAAE,UAAU;AAC9B,UAAE,mBAAmBD,QAAM,OAAO,MAAM;AACxC,UAAE,KAAK;AACP,QAAE,OAAO,CAACC,MAAI,QAAQ,CAAC,CAAC;AACxB,YAAM,EAAE,qBAAqB;;AAE/B;;AAEF,QAAI,UAAU;AACZ,SAAI,CAAC,UAAU;AACb,iBAAW,KAAK,IAAI,OAAO,sBAAsB;AACjD,kBAAY,SAAS,kBAAkB;;AAEzC,eAAW,YAAY,EAAE,SAAS;AAClC,eAAW,aAAa,GAAG,EAAE,UAAU;AACvC,eAAW,mBAAmB,GAAG,MAAM,GAAG;AAC1C;;IAEF,MAAMA,QAAM,KAAK,IAAI,OAAO,sBAAsB;IAClD,MAAM,IAAIA,MAAI,kBAAkB;AAChC,MAAE,YAAY,EAAE,SAAS;AACzB,MAAE,aAAa,GAAG,EAAE,UAAU;AAC9B,MAAE,mBAAmB,GAAG,MAAM,GAAG;AACjC,MAAE,KAAK;AACP,MAAE,OAAO,CAACA,MAAI,QAAQ,CAAC,CAAC;AACxB,UAAM,EAAE,qBAAqB;;AAE/B,SAAM,YAAY;AAClB,aAAU,kBAAkB;SACvB;GACL,MAAMA,QAAM,KAAK,IAAI,OAAO,qBAAqB,EAAE,OAAO,iBAAiB,CAAC;GAC5E,MAAM,OAAOA,MAAI,iBAAiB,EAAE,OAAO,eAAe,CAAC;AAC3D,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,CAACA,MAAI,QAAQ,CAAC,CAAC;;EAI1B,MAAM,OAAO,IAAI,KAAK;EACtB,MAAM,MAAM,KAAK,MAAM,OAAO,mBACzB,KAAK,MAAM,QAAQ,iBAAiB,MAAM,KAC1C,KAAK,MAAM,QAAQ,iBAAiB,MAAM;EAC/C,MAAM,UAAU,OAAO,MAAM;EAC7B,MAAM,MAAM,KAAK,kBAAkB,IAAI,mBAAmB;EAC1D,MAAM,WAAW,KAAK,IAAI,OAAO,aAAa;GAAE,MAAM;GAAS,OAAO;GAAiB,CAAC;EACxF,MAAM,MAAM,KAAK,IAAI,OAAO,sBAAsB;AAClD,MAAI,mBAAmB,KAAK,GAAG,UAAU,GAAG,QAAQ;AACpD,OAAK,IAAI,OAAO,MAAM,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;AAC5C,QAAM,SAAS,SAAS,eAAe,GAAG,QAAQ;EAClD,MAAM,SAAS,IAAI,aAAa,SAAS,eAAe,GAAG,QAAQ,CAAC,MAAM,EAAE,CAAC;AAC7E,WAAS,OAAO;AAChB,WAAS,SAAS;AAClB,SAAO;GAAE;GAAQ;GAAM;GAAK;;;;;;;;;;;CAY9B,MAAM,aACJ,QASA,SAC8D;EAC9D,MAAM,IAAI,OAAO;AACjB,MAAI,IAAI,KAAK,WACX,OAAM,IAAI,MAAM,2BAA2B,EAAE,8BAA8B,KAAK,aAAa;AAE/F,OAAK,WAAW,OAAO;EACvB,MAAM,IAAI,KAAK,IAAI,OAAO;AAC1B,IAAE,YAAY,KAAK,kBAAkB,IAAI,cAAc,EAAG,GAAG,OAAO,QAAwB;AAC5F,IAAE,YACA,KAAK,kBAAkB,IAAI,iBAAiB,EAC5C,GACA,OAAO,UACR;AACD,IAAE,YAAY,KAAK,kBAAkB,IAAI,UAAU,EAAG,GAAG,OAAO,IAAoB;AACpF,IAAE,YAAY,KAAK,kBAAkB,IAAI,UAAU,EAAG,GAAG,OAAO,IAAoB;AACpF,IAAE,YAAY,KAAK,kBAAkB,IAAI,aAAa,EAAG,GAAG,OAAO,WAA2B;EAE9F,MAAM,iBAAiB,KAAK,cAAc,EAAE;EAC5C,MAAM,QAAQ,EAAE,QAAQ,GAAG;EAC3B,MAAMJ,QAAyC,EAAE;AACjD,OAAK,MAAM,KAAK,KAAK,YAAY;GAC/B,MAAM,SAAS,EAAE,KAAK,YAAY,EAAE,MAAM,gBAAgB,MAAM;AAChE,KAAE,YAAY,EAAE,eAAe,GAAG,OAAO;AACzC,SAAM,KAAK,EAAE,KAAK,gBAAgB,EAAE,MAAM,gBAAgB,MAAM,CAAC;;AAInE,MADqB,KAAK,IAAI,gBACZ;AAChB,aAAU,oBAAoB,EAAE,OAAO,KAAK,WAAW,QAAQ,CAAC;AAChE,QAAK,IAAI,IAAI,GAAG,IAAI,KAAK,WAAW,QAAQ,KAAK;IAC/C,MAAM,IAAI,KAAK,WAAW;AAC1B,QAAI,EAAE,KAAK,WAAW,aAAa;KACjC,MAAM,CAAC,OAAO,OAAO,SAAS,MAAM;KACpC,MAAM,QAAQ,OAAO,SAAS,EAAE,KAAK,GAAG,QAAQ,QAAQ,GAAG,EAAE,GAAG;AAChE,SAAI,CAAC,OAAO,MAAM,MAAM,CAAE,WAAU,YAAY,EAAE,OAAO,CAAC;AAC1D,UAAK,IAAI,QAAQ,GAAG,QAAQ,OAAO,SAAS,wBAAwB;MAClE,MAAMG,SAAO,KAAK,IAAI,wBAAwB,QAAQ,MAAM;MAC5D,MAAM,SAAS,EAAE,KAAK,YAAY,EAAE,MAAM,gBAAgB;OACxD,QAAQ;OACR,SAAS;OACV,CAAC;AACF,QAAE,YAAY,EAAE,eAAe,GAAG,OAAO;MACzC,MAAMC,QAAM,KAAK,IAAI,OAAO,sBAAsB;MAClD,MAAMC,MAAID,MAAI,kBAAkB;AAChC,UAAE,YAAY,EAAE,SAAS;AACzB,UAAE,aAAa,GAAG,EAAE,UAAU;AAC9B,UAAE,mBAAmBD,QAAM,OAAO,MAAM;AACxC,UAAE,KAAK;AACP,QAAE,OAAO,CAACC,MAAI,QAAQ,CAAC,CAAC;AACxB,YAAM,EAAE,qBAAqB;;AAE/B;;IAEF,MAAMA,QAAM,KAAK,IAAI,OAAO,sBAAsB;IAClD,MAAM,IAAIA,MAAI,kBAAkB;AAChC,MAAE,YAAY,EAAE,SAAS;AACzB,MAAE,aAAa,GAAG,EAAE,UAAU;AAC9B,MAAE,mBAAmB,GAAG,MAAM,GAAG;AACjC,MAAE,KAAK;AACP,MAAE,OAAO,CAACA,MAAI,QAAQ,CAAC,CAAC;AACxB,UAAM,EAAE,qBAAqB;;AAE/B,aAAU,kBAAkB;SACvB;GACL,MAAMA,QAAM,KAAK,IAAI,OAAO,qBAAqB,EAAE,OAAO,wBAAwB,CAAC;GACnF,MAAM,OAAOA,MAAI,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,CAACA,MAAI,QAAQ,CAAC,CAAC;;EAG1B,MAAM,OAAO,OAAO;EACpB,MAAM,MAAM,KAAK,MAAM,QAAQ,iBAAiB,MAAM;EACtD,MAAM,UAAU,OAAO,MAAM;EAC7B,MAAM,MAAM,KAAK,kBAAkB,IAAI,mBAAmB;EAC1D,MAAM,WAAW,KAAK,IAAI,OAAO,aAAa;GAAE,MAAM;GAAS,OAAO;GAAiB,CAAC;EACxF,MAAM,MAAM,KAAK,IAAI,OAAO,sBAAsB;AAClD,MAAI,mBAAmB,KAAK,GAAG,UAAU,GAAG,QAAQ;AACpD,OAAK,IAAI,OAAO,MAAM,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;AAC5C,QAAM,SAAS,SAAS,eAAe,GAAG,QAAQ;EAClD,MAAM,SAAS,IAAI,aAAa,SAAS,eAAe,GAAG,QAAQ,CAAC,MAAM,EAAE,CAAC;AAC7E,WAAS,OAAO;AAChB,WAAS,SAAS;AAClB,SAAO;GAAE;GAAQ;GAAM;GAAK;;;CAI9B,IAAI,SAAkB;AACpB,SAAO,KAAK;;;CAId,MAAM,gBAAgB,MAAc,aAA6C;EAC/E,MAAM,SAAS,KAAK,kBAAkB,IAAI,KAAK,IAAI,KAAK,cAAc,IAAI,KAAK;AAC/E,MAAI,CAAC,OAAQ,OAAM,IAAI,MAAM,8BAA8B,KAAK,GAAG;EACnE,MAAM,UAAU,cAAc,KAAK,IAAI,cAAc,GAAG,OAAO,KAAK,GAAG,OAAO;EAC9E,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,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,YAAoB;AAG1B,SAAO,KAAK;;CAGd,AAAQ,cAAc,GAAqC;EACzD,MAAM,KAAK,IAAI,KAAK;EACpB,MAAM,KAAK,KAAK,YAAY;EAC5B,MAAME,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,KAAM,QAAO;AACvB,OAAI,MAAM,KAAM,QAAO;AACvB,UAAO;IACP;AAEJ,SAAO;;CAGT,AAAQ,4BAAkC;EAOxC,MAAM,6BAAa,IAAI,KAAa;AACpC,OAAK,MAAM,QAAQ,KAAK,MAAM,OAAO;AACnC,QAAK,MAAM,KAAK,KAAK,OAAQ,YAAW,IAAI,EAAE;AAC9C,QAAK,MAAM,KAAK,KAAK,QAAS,YAAW,IAAI,EAAE;;EAEjD,IAAI,eAAe;AACnB,OAAK,MAAM,CAAC,MAAM,SAAS,OAAO,QAAQ,KAAK,MAAM,QAAQ,EAAE;AAC7D,OAAI,KAAK,YAAY,aAAc;AACnC,OAAI,CAAC,WAAW,IAAI,KAAK,EAAE;IACzB,MAAM,QAAQ,KAAK,MAAM,KAAK,MAAM;AAClC,SAAI,MAAM,IAAK,QAAO,KAAK;AAC3B,SAAI,MAAM,KAAM,QAAO,KAAK,aAAa,KAAK;AAC9C,SAAI,MAAM,KAAM,QAAO,KAAK,WAAW;AACvC,YAAO;MACP;AACF,oBAAgB,MAAM,QAAQ,GAAG,MAAM,IAAI,GAAG,EAAE,GAAG,YAAY,KAAK;AACpE;;GAQF,MAAM,QANQ,KAAK,MAAM,KAAK,MAAM;AAClC,QAAI,MAAM,IAAK,QAAO,KAAK;AAC3B,QAAI,MAAM,KAAM,QAAO,KAAK,aAAa,KAAK;AAC9C,QAAI,MAAM,KAAM,QAAO,KAAK,WAAW;AACvC,WAAO;KACP,CACkB,QAAQ,GAAG,MAAM,IAAI,GAAG,EAAE,GAAG,YAAY,KAAK;AAClE,QAAK,kBAAkB,IAAI,MAAM,oBAAoB,KAAK,KAAK,QAAQ,QAAQ,MAAM,CAAC;;AAExF,MAAI,eAAe,EACjB,SAAQ,IACN,qBAAqB,eAAe,KAAK,QAAQ,EAAE,CAAC,wCACrD;;CAIL,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,MAAM,kCAAkC,WAAW,UAAU,KAAK,KAAK;AACnF,WAAQ,KAAK,EAAE,QAAQ,CAAC;;AAG5B,SAAO;;;;;;;ACjfX,SAAS,SAAS,MAAc,OAA6B;CAC3D,MAAM,MAAM,IAAI,aAAa,MAAM;AACnC,KAAI,UAAU,GAAG;AACf,MAAI,KAAK;AACT,SAAO;;CAET,MAAM,OAAO,QAAQ,QAAQ;AAC7B,MAAK,IAAI,IAAI,GAAG,IAAI,OAAO,IAAK,KAAI,KAAK,IAAI;AAC7C,QAAO;;;;;;;;AAST,SAAS,oBAAoB,GAAW,GAAW,OAA2B;CAI5E,MAAM,KAAK,IAAI;CACf,MAAM,KAAK,IAAI;CACf,MAAM,MAAM,IAAI,WAAW,IAAI,EAAE;CACjC,IAAI,IAAI;AACR,MAAK,IAAI,KAAK,GAAG,KAAK,IAAI,KACxB,MAAK,IAAI,KAAK,GAAG,KAAK,IAAI,KACxB,MAAK,IAAI,KAAK,GAAG,KAAK,OAAO,KAC3B,MAAK,IAAI,KAAK,GAAG,KAAK,OAAO,MAAM;EACjC,MAAM,KAAK,KAAK,QAAQ;EACxB,MAAM,KAAK,KAAK,QAAQ;AACxB,MAAI,OAAO,KAAK,IAAI;;AAK5B,QAAO;;;;;;AAOT,SAAgB,eACd,SACA,eACA,KACc;CACd,MAAM,CAAC,GAAG,GAAG,KAAK;CAClB,MAAM,SAAS,IAAI;CACnB,MAAM,OAAO,KAAK,MAAM,KAAK,KAAK,IAAI,sBAAsB,CAAC;CAC7D,MAAM,QAAQ,IAAI;CAElB,MAAM,QAAQ,SAAS,OAAO,GAAG,EAAE;CACnC,MAAM,QAAQ,SAAS,OAAO,GAAG,EAAE;CACnC,MAAM,SAAS,IAAI,WAAW,EAAE;CAChC,MAAM,QAAQ,IAAI,WAAW,EAAE;CAC/B,MAAM,QAAQ,IAAI,aAAa,EAAE;AACjC,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,SAAO,KAAK,KAAK,MAAM,MAAM,GAAG;AAChC,QAAM,KAAK,KAAK,IAAI,OAAO,KAAK,GAAG,OAAO,EAAE;AAC5C,QAAM,KAAK,MAAM,KAAK,OAAO;;CAE/B,MAAM,SAAS,IAAI,WAAW,EAAE;CAChC,MAAM,QAAQ,IAAI,WAAW,EAAE;CAC/B,MAAM,QAAQ,IAAI,aAAa,EAAE;AACjC,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,SAAO,KAAK,KAAK,MAAM,MAAM,GAAG;AAChC,QAAM,KAAK,KAAK,IAAI,OAAO,KAAK,GAAG,OAAO,EAAE;AAC5C,QAAM,KAAK,MAAM,KAAK,OAAO;;CAI/B,MAAM,KAAK,IAAI;CACf,MAAM,YAAY;EAChB,IAAI,WAAW,GAAG;EAClB,IAAI,WAAW,GAAG;EAClB,IAAI,WAAW,GAAG;EAClB,IAAI,WAAW,GAAG;EACnB;CACD,MAAM,UAAU;EACd,IAAI,aAAa,GAAG;EACpB,IAAI,aAAa,GAAG;EACpB,IAAI,aAAa,GAAG;EACpB,IAAI,aAAa,GAAG;EACrB;CACD,IAAI,IAAI;AACR,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,MAAM,KAAK,OAAO,KAAK;EACvB,MAAM,KAAK,MAAM,KAAK;EACtB,MAAM,MAAM,MAAM;AAClB,OAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,aAAU,GAAG,KAAK,KAAK,OAAO;AAC9B,aAAU,GAAG,KAAK,KAAK,MAAM;AAC7B,aAAU,GAAG,KAAK,KAAK,OAAO;AAC9B,aAAU,GAAG,KAAK,KAAK,MAAM;AAC7B,WAAQ,GAAG,MAAM,IAAI,QAAQ,IAAI,MAAM;AACvC,WAAQ,GAAG,MAAM,IAAI,OAAO,MAAM;AAClC,WAAQ,GAAG,KAAK,OAAO,IAAI,MAAM;AACjC,WAAQ,GAAG,KAAK,MAAM,MAAM;AAC5B;;;CAIJ,MAAM,UAAU,oBAAoB,GAAG,GAAG,MAAM;CAChD,MAAM,IAAI,IAAI;CACd,MAAM,MAAM,IAAI,aAAa,IAAI,OAAO;AAExC,MAAK,IAAI,KAAK,GAAG,KAAK,GAAG,KACvB,MAAK,IAAI,IAAI,GAAG,IAAI,IAAI,KAAK;EAC3B,MAAM,MAAM,QAAQ;EAEpB,MAAM,WADS,KAAK,KAAK,KACA;AACzB,OAAK,IAAI,SAAS,GAAG,SAAS,GAAG,UAAU;GACzC,MAAM,WAAW,UAAU,QAAQ,OAAO;GAC1C,MAAM,MAAM,QAAQ,QAAQ;AAC5B,OAAI,QAAQ,EAAG;AACf,QAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,IAC1B,KAAI,UAAU,MAAM,cAAc,WAAW,KAAK;;;AAK1D,QAAO;;;;;;AAOT,SAAgB,iBAAiB,SAAmC,OAA2B;CAC7F,MAAM,CAAC,GAAG,GAAG,KAAK;CAClB,MAAM,KAAK,IAAI;CACf,MAAM,KAAK,IAAI;CACf,MAAM,KAAK,IAAI;CACf,MAAM,WAAW,IAAI,WAAW,KAAK,EAAE;CACvC,IAAI,IAAI;AAER,MAAK,IAAI,KAAK,GAAG,KAAK,IAAI,KACxB,MAAK,IAAI,KAAK,GAAG,KAAK,IAAI,KACxB,MAAK,IAAI,KAAK,GAAG,KAAK,OAAO,KAC3B,MAAK,IAAI,KAAK,GAAG,KAAK,OAAO,MAAM;AACjC,WAAS,IAAI,KAAK,KAAK,QAAQ;AAC/B,WAAS,IAAI,IAAI,KAAK,KAAK,QAAQ;AACnC;;CAKR,MAAM,MAAM,IAAI,WAAW,IAAI,KAAK,EAAE;AACtC,MAAK,IAAI,KAAK,GAAG,KAAK,GAAG,KAAM,KAAI,IAAI,UAAU,KAAK,KAAK,EAAE;AAC7D,QAAO;;;;;;;;;;;AAYT,SAAgB,kBACd,aACA,SACA,QAAQ,KACsD;CAE9D,MAAM,YAAY,UAAU;CAC5B,MAAM,OAAO,YAAY;CACzB,MAAM,UAAU,IAAI,aAAa,KAAK;AACtC,MAAK,IAAI,IAAI,GAAG,IAAI,MAAM,IACxB,SAAQ,KAAK,IAAM,UAAW,IAAI,IAAK;CAGzC,MAAM,IAAI,YAAY,SAAS;CAG/B,MAAM,MAAM,IAAI,aAAa,IAAI,QAAQ;CACzC,MAAM,MAAM,IAAI,aAAa,IAAI,QAAQ;AACzC,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,MAAM,KAAK,YAAY,IAAI;EAC3B,MAAM,KAAK,YAAY,IAAI,IAAI;EAC/B,MAAM,OAAO,IAAI;AACjB,OAAK,IAAI,IAAI,GAAG,IAAI,MAAM,KAAK;GAC7B,MAAM,KAAK,KAAK,QAAQ;GACxB,MAAM,KAAK,KAAK,QAAQ;GAExB,MAAM,KAAK,KAAK,IAAI,GAAG;GACvB,MAAM,KAAK,KAAK,IAAI,GAAG;GACvB,MAAM,KAAK,KAAK,IAAI,GAAG;GACvB,MAAM,KAAK,KAAK,IAAI,GAAG;AACvB,OAAI,OAAO,KAAK;AAChB,OAAI,OAAO,KAAK;AAChB,OAAI,OAAO,OAAO,KAAK;AACvB,OAAI,OAAO,OAAO,KAAK;AACvB,OAAI,OAAO,YAAY,KAAK;AAC5B,OAAI,OAAO,YAAY,KAAK;AAC5B,OAAI,OAAO,YAAY,OAAO,KAAK;AACnC,OAAI,OAAO,YAAY,OAAO,KAAK;;;AAGvC,QAAO;EAAE;EAAK;EAAK,YAAY;EAAG;;;;;AAMpC,SAAgB,2BACd,SACA,eACA,KACuB;CACvB,MAAM,UAAU,KAAK,MAAM,IAAI,aAAa,IAAI,SAAS;CACzD,MAAM,YAAY,eAAe,SAAS,eAAe,IAAI;CAE7D,MAAM,EAAE,KAAK,KAAK,eAAe,kBADb,iBAAiB,SAAS,IAAI,iBAAiB,EAGjE,SACA,IAAI,aAAa,IAClB;AACD,QAAO;EAAE;EAAW;EAAK;EAAK;EAAY;;;;;;;AA6C5C,SAAS,cAAc,OAAe,OAAiD;CACrF,MAAM,IAAI,QAAQ;CAClB,MAAM,IAAI,IAAI,WAAW,EAAE;CAC3B,MAAM,IAAI,IAAI,WAAW,EAAE;CAC3B,IAAI,IAAI;AACR,MAAK,IAAI,IAAI,GAAG,IAAI,OAAO,IACzB,MAAK,IAAI,IAAI,GAAG,IAAI,OAAO,KAAK;AAC9B,IAAE,KAAK;AACP,IAAE,KAAK;AACP;;AAGJ,QAAO;EAAE;EAAG;EAAG;;;;;;;AAQjB,SAAgB,qBACd,OACA,OACA,eACA,QACA,SACc;CACd,MAAM,EAAE,GAAG,MAAM,cAAc,OAAO,MAAM;CAC5C,MAAM,IAAI,QAAQ;CAClB,MAAM,MAAM,IAAI,aAAa,IAAI,OAAO;CACxC,MAAM,SAAS,UAAU;AACzB,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,MAAM,KAAK,KAAK,IAAI,GAAG,EAAE,GAAG;EAC5B,MAAM,KAAK,KAAK,IAAI,GAAG,EAAE,GAAG;EAC5B,MAAM,QAAQ,KAAK;EACnB,MAAM,QAAQ,SAAS,KAAK;EAC5B,MAAM,MAAM,IAAI;AAChB,OAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,IAC1B,KAAI,MAAM,KAAK,cAAc,QAAQ,KAAK,cAAc,QAAQ;;AAGpE,QAAO;;;;;;;;;;AAWT,SAAgB,wBACd,OACA,OACA,SACA,OAC0C;CAC1C,MAAM,EAAE,GAAG,MAAM,cAAc,OAAO,MAAM;CAC5C,MAAM,IAAI,QAAQ;CAClB,MAAM,aAAa,UAAU;CAC7B,MAAM,OAAO,aAAa;CAC1B,MAAM,UAAU,IAAI,aAAa,KAAK;AACtC,MAAK,IAAI,IAAI,GAAG,IAAI,MAAM,IACxB,SAAQ,KAAK,IAAM,UAAW,IAAI,IAAK;CAEzC,MAAM,MAAM,IAAI,aAAa,IAAI,QAAQ;CACzC,MAAM,MAAM,IAAI,aAAa,IAAI,QAAQ;AACzC,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,MAAM,KAAK,KAAK,IAAI,GAAG,EAAE,GAAG;EAC5B,MAAM,KAAK,KAAK,IAAI,GAAG,EAAE,GAAG;EAC5B,MAAM,OAAO,IAAI;AACjB,OAAK,IAAI,IAAI,GAAG,IAAI,MAAM,KAAK;GAC7B,MAAM,KAAK,KAAK,QAAQ;GACxB,MAAM,KAAK,KAAK,QAAQ;GACxB,MAAM,KAAK,KAAK,IAAI,GAAG;GACvB,MAAM,KAAK,KAAK,IAAI,GAAG;GACvB,MAAM,KAAK,KAAK,IAAI,GAAG;GACvB,MAAM,KAAK,KAAK,IAAI,GAAG;AAEvB,OAAI,OAAO,KAAK;AAChB,OAAI,OAAO,KAAK;AAChB,OAAI,OAAO,OAAO,KAAK;AACvB,OAAI,OAAO,OAAO,KAAK;AACvB,OAAI,OAAO,aAAa,KAAK;AAC7B,OAAI,OAAO,aAAa,KAAK;AAC7B,OAAI,OAAO,aAAa,OAAO,KAAK;AACpC,OAAI,OAAO,aAAa,OAAO,KAAK;;;AAGxC,QAAO;EAAE;EAAK;EAAK;;;;;;;;;;;AAYrB,SAAgB,sBACd,OACA,OACA,GACiD;CACjD,MAAM,IAAI,QAAQ;CAClB,MAAM,SAAS,KAAK,KAAK,QAAQ,EAAE;CAEnC,MAAM,KAAK,SADI,KAAK,KAAK,QAAQ,EAAE;CAEnC,MAAM,MAAM,KAAO,IAAI;CACvB,MAAM,aAAa,IAAI,aAAa,KAAK,EAAE;CAC3C,MAAM,EAAE,GAAG,MAAM,cAAc,OAAO,MAAM;AAC5C,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAG1B,MAAM,OAFK,KAAK,MAAM,EAAE,KAAK,EAAE,GAEb,SADP,KAAK,MAAM,EAAE,KAAK,EAAE;AAE/B,aAAW,OAAO,IAAI,KAAK;;AAE7B,QAAO;EAAE;EAAY,WAAW;EAAI;;;;;;AAOtC,SAAgB,iCACd,OACA,OACA,eACA,SACA,KAC6B;CAC7B,MAAM,YAAY,qBAAqB,OAAO,OAAO,eAAe,IAAI,YAAY,QAAQ;CAC5F,MAAM,EAAE,KAAK,QAAQ,wBAAwB,OAAO,OAAO,IAAI,SAAS,IAAI,UAAU;CACtF,MAAM,EAAE,YAAY,cAAc,sBAAsB,OAAO,OAAO,IAAI,kBAAkB;AAC5F,QAAO;EAAE;EAAW;EAAK;EAAK;EAAY,YAAY,QAAQ;EAAO;EAAW;;;AAIlF,MAAaC,yBAA+C;CAC1D,WAAW;CACX,mBAAmB;CACnB,WAAW;CACX,WAAW;EAAC;EAAK;EAAK;EAAI;CAC1B,UAAU;EAAC;EAAK;EAAK;EAAI;CACzB,eAAe,IAAI;CACnB,WAAW;CACX,WAAW,OAAO,KAAK;CACxB;;;;;;;;AAgBD,SAAgB,sBACd,QACA,OACA,QACA,gBAAgB,KAChB,oBAAoB,GACpB,YAAY,IACa;CACzB,MAAM,SAAS,oBAAoB;CACnC,MAAM,aAAa,gBAAgB,oBAAoB;CAEvD,MAAM,iBAAiB,MAAc,KAAK,IAAI,QAAQ,KAAK,MAAM,IAAI,OAAO,GAAG,OAAO;CACtF,IAAI,OAAO,cAAc,OAAO;CAChC,IAAI,OAAO,cAAc,MAAM;AAE/B,QAAQ,OAAO,aAAc,OAAO,aAAa,YAAY;EAC3D,MAAM,OAAO,KAAK,KAAO,OAAO,aAAc,OAAO,aAAc,WAAW;AAC9E,SAAO,cAAc,OAAO,KAAK;AACjC,SAAO,cAAc,OAAO,KAAK;AACjC,MAAI,QAAQ,UAAU,QAAQ,OAAQ;;CAExC,MAAM,UAAU,eAAe,QAAQ,QAAQ,OAAO,MAAM,KAAK;CAEjE,MAAM,QAAQ,OAAO;CACrB,MAAM,QAAQ,OAAO;CACrB,MAAM,KAAK;CACX,MAAM,KAAK;CACX,MAAM,WAAW,KAAK,KAAK;CAC3B,MAAM,OAAO,QAAQ;CACrB,MAAM,UAAU,IAAI,aAAa,OAAO,SAAS;CACjD,MAAM,UAAU,IAAI;CAIpB,IAAI,OAAO;AACX,MAAK,IAAI,KAAK,GAAG,KAAK,OAAO,KAC3B,MAAK,IAAI,KAAK,GAAG,KAAK,OAAO,MAAM;EACjC,MAAM,OAAO,OAAO;EACpB,IAAI,KAAK;AACT,OAAK,IAAI,IAAI,GAAG,IAAI,IAAI,IACtB,MAAK,IAAI,KAAK,GAAG,KAAK,IAAI,MAAM;GAC9B,MAAM,OAAO,KAAK,KAAK,MAAM;AAC7B,QAAK,IAAI,KAAK,GAAG,KAAK,IAAI,MAAM;AAC9B,YAAQ,OAAO,MAAM,SAAS,MAAM,KAAK,KAAK,MAAM,IAAI,KAAK;AAC7D;;;AAIN;;AAGJ,QAAO;EAAE;EAAS,QAAQ,CAAC,OAAO,MAAM;EAAE;;AAyB5C,MAAaC,0BAAgD;CAC3D,WAAW;CACX,mBAAmB;CACnB,WAAW;CACX,WAAW;EAAC;EAAK;EAAK;EAAI;CAC1B,UAAU;EAAC;EAAK;EAAK;EAAI;CACzB,eAAe,IAAI;CACnB,WAAW;CACX,WAAW;CACZ;;;;;;AAcD,SAAgB,YACd,QACA,OACA,QACA,WACA,WACkB;AAClB,KAAI,SAAS,UAAU,QAAQ,QAAQ;EAErC,MAAM,QAAQ,SAAS,KAAK,IAAI,QAAQ,MAAM;AAC9C,WAAS,KAAK,IAAI,QAAQ,KAAK,MAAM,SAAS,MAAM,CAAC;AACrD,UAAQ,KAAK,IAAI,QAAQ,KAAK,MAAM,QAAQ,MAAM,CAAC;;CAErD,MAAM,iBAAiB,MAAc,KAAK,MAAM,IAAI,OAAO,GAAG;CAC9D,MAAM,iBAAiB,MAAc,KAAK,MAAM,IAAI,OAAO,GAAG;CAC9D,MAAM,gBAAgB,MAAc,KAAK,KAAK,IAAI,OAAO,GAAG;CAE5D,IAAI,OAAO,KAAK,IAAI,QAAQ,cAAc,OAAO,CAAC;CAClD,IAAI,OAAO,KAAK,IAAI,QAAQ,cAAc,MAAM,CAAC;AACjD,KAAI,OAAO,OAAO,WAAW;EAC3B,MAAM,OAAO,KAAK,KAAM,SAAS,QAAS,UAAU;AACpD,SAAO,KAAK,IAAI,QAAQ,cAAc,SAAS,KAAK,CAAC;AACrD,SAAO,KAAK,IAAI,QAAQ,cAAc,QAAQ,KAAK,CAAC;YAC3C,OAAO,OAAO,WAAW;EAClC,MAAM,OAAO,KAAK,KAAK,aAAa,SAAS,OAAO;AACpD,SAAO,aAAa,SAAS,KAAK;AAClC,SAAO,aAAa,QAAQ,KAAK;;AAEnC,QAAO,CAAC,MAAM,KAAK;;;;;;;AAQrB,SAAS,eACP,QACA,KACA,KACA,MACA,MACc;CACd,MAAM,MAAM,IAAI,aAAa,OAAO,OAAO,EAAE;CAC7C,MAAM,SAAS,MAAM;CACrB,MAAM,SAAS,MAAM;AACrB,MAAK,IAAI,KAAK,GAAG,KAAK,MAAM,MAAM;EAChC,IAAI,MAAM,KAAK,MAAO,SAAS;AAC/B,MAAI,KAAK,EAAG,MAAK;AACjB,MAAI,KAAK,MAAM,EAAG,MAAK,MAAM;EAC7B,MAAM,KAAK,KAAK,MAAM,GAAG;EACzB,MAAM,KAAK,KAAK,IAAI,KAAK,GAAG,MAAM,EAAE;EACpC,MAAM,KAAK,KAAK;AAChB,OAAK,IAAI,KAAK,GAAG,KAAK,MAAM,MAAM;GAChC,IAAI,MAAM,KAAK,MAAO,SAAS;AAC/B,OAAI,KAAK,EAAG,MAAK;AACjB,OAAI,KAAK,MAAM,EAAG,MAAK,MAAM;GAC7B,MAAM,KAAK,KAAK,MAAM,GAAG;GACzB,MAAM,KAAK,KAAK,IAAI,KAAK,GAAG,MAAM,EAAE;GACpC,MAAM,KAAK,KAAK;GAChB,MAAM,OAAO,KAAK,MAAM,MAAM;GAC9B,MAAM,OAAO,KAAK,MAAM,MAAM;GAC9B,MAAM,OAAO,KAAK,MAAM,MAAM;GAC9B,MAAM,OAAO,KAAK,MAAM,MAAM;GAC9B,MAAM,MAAM,KAAK,OAAO,MAAM;AAC9B,QAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;IAC1B,MAAM,MAAM,OAAO,MAAM,MAAM,IAAI,MAAM,OAAO,MAAM,KAAK;IAC3D,MAAM,MAAM,OAAO,MAAM,MAAM,IAAI,MAAM,OAAO,MAAM,KAAK;AAC3D,QAAI,KAAK,KAAK,OAAO,IAAI,MAAM,MAAM;;;;AAI3C,QAAO;;;;;;;;;;;;;AAcT,SAAgB,gBACd,QACA,OACA,QACA,MAA4B,yBAC5B,WAAW,OACQ;CAEnB,MAAM,CAAC,MAAM,QAAQ,YAAY,QAAQ,OAD1B,IAAI,YAAY,IAAI,WACqB,IAAI,WAAW,IAAI,UAAU;CACrF,MAAM,UAAU,eAAe,QAAQ,QAAQ,OAAO,MAAM,KAAK;CAGjE,MAAM,CAAC,IAAI,IAAI,MAAM,IAAI;CACzB,MAAM,CAAC,IAAI,IAAI,MAAM,IAAI;CACzB,MAAM,OAAO;EAAC;EAAI;EAAI;EAAG;CACzB,MAAM,MAAM;EAAC;EAAI;EAAI;EAAG;CACxB,MAAM,UAAU,WAAW,IAAI,IAAI;CACnC,MAAM,MAAM,IAAI,aAAa,IAAI,OAAO,KAAK;CAC7C,MAAM,QAAQ,OAAO;AACrB,MAAK,IAAI,IAAI,GAAG,IAAI,MAAM,IACxB,MAAK,IAAI,IAAI,GAAG,IAAI,MAAM,KAAK;EAC7B,MAAM,OAAO,IAAI,OAAO,KAAK;EAC7B,MAAM,MAAM,IAAI,OAAO;AACvB,OAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IACrB,KAAI,IAAI,QAAQ,QAAQ,QAAQ,MAAM,KAAK,UAAU,KAAK,MAAM,IAAI;;CAO1E,MAAM,IAAI;CACV,MAAM,QAAQ,OAAO,IAAI;CACzB,MAAM,QAAQ,OAAO,IAAI;CACzB,MAAM,KAAK,IAAI;CACf,MAAM,MAAM,IAAI;CAChB,MAAM,KAAK;CACX,MAAM,QAAQ,IAAI;CAClB,MAAM,SAAS,QAAQ;CACvB,MAAM,SAAS,QAAQ;CACvB,MAAM,WAAW,KAAK,MAAM,KAAK;CACjC,MAAM,aAAa,IAAI,QAAQ;CAC/B,MAAM,UAAU,IAAI,aAAa,aAAa,SAAS;CAKvD,IAAI,OAAO;AACX,MAAK,IAAI,KAAK,GAAG,KAAK,GAAG,KACvB,MAAK,IAAI,KAAK,GAAG,KAAK,QAAQ,KAC5B,MAAK,IAAI,KAAK,GAAG,KAAK,QAAQ,KAC5B,MAAK,IAAI,KAAK,GAAG,KAAK,OAAO,KAC3B,MAAK,IAAI,KAAK,GAAG,KAAK,OAAO,MAAM;EACjC,MAAM,YAAY,KAAK,QAAQ,MAAM;EACrC,MAAM,YAAY,KAAK,QAAQ,MAAM;EACrC,MAAM,OAAO,OAAO;EACpB,IAAI,IAAI;AACR,OAAK,IAAI,IAAI,GAAG,IAAI,IAAI,IACtB,MAAK,IAAI,KAAK,GAAG,KAAK,KAAK,KAEzB,MAAK,IAAI,KAAK,GAAG,KAAK,IAAI,MAAM;GAC9B,MAAM,OAAO,WAAW,MAAM;AAC9B,QAAK,IAAI,KAAK,GAAG,KAAK,IAAI,MAAM;AAC9B,YAAQ,OAAO,KAAK,IAAI,IAAI,QAAQ,OAAO,WAAW;AACtD;;;AAMR;;AAOV,QAAO;EAAE;EAAS,SAAS;GAAC;GAAG;GAAO;GAAM;EAAE;;;;;;;;;;;;;;;;AAiBhD,SAAS,mBACP,MACA,IACA,OACA,MACA,WACkC;CAClC,MAAM,CAAC,IAAI,IAAI,MAAM;CACrB,MAAM,KAAK,KAAK;CAChB,MAAM,KAAK,KAAK;CAChB,IAAI,IAAI;AACR,MAAK,IAAI,KAAK,GAAG,KAAK,IAAI,KACxB,MAAK,IAAI,IAAI,GAAG,IAAI,IAAI,IACtB,MAAK,IAAI,IAAI,GAAG,IAAI,IAAI,KAAK;AAC3B,OAAK,EAAE,KAAK,QAAQ;AACpB,OAAK,EAAE,KAAK,QAAQ;AACpB,OAAK,EAAE,KAAK,QAAQ;AACpB;;AAIN,QAAO;EAAE,MAAM;EAAG,QAAQ,QAAQ,KAAK,IAAI,IAAI,GAAG;EAAE;;AAGtD,SAAgB,sBACd,UACA,YACA,cACA,WACY;CACZ,MAAM,MAAM,SAAS;CACrB,MAAM,OAAO;EAAE,GAAG,IAAI,WAAW,IAAI;EAAE,GAAG,IAAI,WAAW,IAAI;EAAE,GAAG,IAAI,WAAW,IAAI;EAAE;CAEvF,IAAI,IAAI;CACR,IAAI,SAAS;CACb,IAAI,WAAW;AACf,QAAO,IAAI,IACT,KAAI,SAAS,OAAO,cAAc;EAChC,MAAM,MAAM,mBAAmB,MAAM,GAAG,QAAQ,WAAW,aAAa,UAAU;AAClF,MAAI,IAAI;AACR,WAAS,IAAI;QACR;AACL,OAAK,EAAE,KAAK;AACZ,OAAK,EAAE,KAAK;AACZ,OAAK,EAAE,KAAK;AACZ;AACA;;CAIJ,MAAM,MAAM,IAAI,WAAW,IAAI,IAAI;AACnC,KAAI,IAAI,KAAK,GAAG,EAAE;AAClB,KAAI,IAAI,KAAK,GAAG,IAAI;AACpB,KAAI,IAAI,KAAK,GAAG,IAAI,IAAI;AACxB,QAAO;;;;;;;;;AAUT,SAAgB,cAAc,cAAoD;CAChF,MAAM,QAAQ,aAAa,KAAK,aAAa,KAAK,aAAa;CAC/D,MAAM,OAAO,IAAI,WAAW,MAAM;AAClC,MAAK,IAAI,IAAI,GAAG,KAAK,GAAG,KAAK;EAC3B,MAAM,SAAS,aAAa,KAAK;AACjC,OAAK,IAAI,MAAM,GAAG,MAAM,UAAU,MAAM,OAAO,OAAO,EACpD,MAAK,OAAO;;AAGhB,QAAO;;;;;;;;;;;;;AAcT,SAAgB,iBACd,cACA,KACA,SACA,OACA,cAC0C;CAC1C,MAAM,OAAO,UAAU;CACvB,MAAM,WAAW,cAAc,aAAa;CAC5C,MAAM,UAAU,IAAI,aAAa,KAAK;AACtC,MAAK,IAAI,IAAI,GAAG,IAAI,MAAM,IACxB,SAAQ,KAAK,IAAM,UAAW,IAAI,IAAK;CAEzC,MAAM,MAAM,IAAI,aAAa,MAAM,QAAQ;CAC3C,MAAM,MAAM,IAAI,aAAa,MAAM,QAAQ;AAC3C,MAAK,IAAI,IAAI,GAAG,IAAI,KAAK,KAAK;EAC5B,MAAM,OAAO,IAAI;AACjB,OAAK,IAAI,IAAI,GAAG,IAAI,MAAM,KAAK;GAG7B,MAAM,QADM,aADA,SAAS,KACU,MAAM,KACjB,QAAQ;GAC5B,MAAM,IAAI,KAAK,IAAI,MAAM;GACzB,MAAM,IAAI,KAAK,IAAI,MAAM;AAEzB,OAAI,OAAO,KAAK;AAChB,OAAI,OAAO,KAAK;AAChB,OAAI,OAAO,OAAO,KAAK;AACvB,OAAI,OAAO,OAAO,KAAK;;;AAG3B,QAAO;EAAE;EAAK;EAAK;;;;;AChfrB,MAAa,yBAAyB;AAiBtC,MAAMC,gBAAkC;CAAE,WAAW;CAAG,QAAQ;CAAM,QAAQ;CAAO;;AAGrF,SAAgB,iBAAmC;AACjD,KAAI,OAAO,iBAAiB,YAAa,QAAO,EAAE,GAAG,eAAe;AACpE,KAAI;EACF,MAAM,MAAM,aAAa,QAAQ,uBAAuB;AACxD,MAAI,CAAC,IAAK,QAAO,EAAE,GAAG,eAAe;EACrC,MAAM,SAAS,KAAK,MAAM,IAAI;AAC9B,SAAO;GACL,WACE,OAAO,OAAO,cAAc,YAAY,OAAO,aAAa,IAAI,OAAO,YAAY;GACrF,QAAQ,OAAO,OAAO,WAAW,WAAW,OAAO,SAAS;GAC5D,QAAQ,OAAO,WAAW;GAC3B;SACK;AACN,SAAO,EAAE,GAAG,eAAe;;;;AAK/B,SAAgB,gBAAgB,KAA6B;AAC3D,KAAI,OAAO,iBAAiB,YAAa;AACzC,KAAI;AACF,eAAa,QAAQ,wBAAwB,KAAK,UAAU,IAAI,CAAC;SAC3D;;;;;;;;;;;;AAeV,MAAM,wBAAwB;;;;;;;;;;;;;;;;AAiB9B,SAAgB,uBAAuB,MAW5B;AAET,KAAI,KAAK,YAAY,KAAK,WAAW,EAAG,QAAO,KAAK;AAEpD,KAAI,CAAC,KAAK,SAAU,QAAO;CAE3B,MAAM,MAAM,gBAAgB;AAG5B,KAAI,IAAI,WAAW,MAAM;EACvB,MAAM,YAAY,IAAI;EACtB,MAAMC,OAAyB;GAC7B,WAAW,IAAI;GACf,QAAQ;GACR,QAAQ;GACT;AACD,kBAAgB,KAAK;AACrB,UAAQ,IACN,+DAA+D,UAAU,0BAA0B,KAAK,YACzG;AACD,SAAO,KAAK;;AAUd,KAAI,CAAC,IAAI,UAAU,CAAC,KAAK,gBAAgB,IAAI,YAAY,uBAAuB;AAM9E,kBAL+B;GAC7B,WAAW,IAAI;GACf,QAAQ;GACR,QAAQ;GACT,CACoB;AACrB,UAAQ,IACN,oDAAoD,sBAAsB,cAAc,IAAI,UAAU,GACvG;AACD,SAAO;;AAIT,SAAQ,IACN,0CAA0C,IAAI,UAAU,sBAAsB,IAAI,OAAO,WAAW,IAAI,YACzG;AACD,QAAO,IAAI;;;;;;;;;;;;;;AAeb,SAAgB,kBAAkB,SAAwB;AACxD,KAAI,OAAO,iBAAiB,YAAa;CACzC,MAAM,MAAM,gBAAgB;AAE5B,KAAI,IAAI,WAAW,KAAM;AAEzB,KAAI,SAAS;AACX,kBAAgB;GAAE,WAAW,IAAI;GAAQ,QAAQ;GAAM,QAAQ,IAAI;GAAQ,CAAC;AAC5E,UAAQ,IAAI,+CAA+C,IAAI,OAAO,gBAAgB;QACjF;AAGL,kBAAgB;GAAE,WAAW,IAAI;GAAW,QAAQ;GAAM,QAAQ;GAAM,CAAC;AACzE,UAAQ,IACN,sCAAsC,IAAI,OAAO,oDAAoD,IAAI,YAC1G;;;;;;;;;;;;;;;ACvYL,MAAM,0BAA0B;CAC9B,OAAO;CACP,UAAU;CACX;;;;;;;;;;AA6GD,SAAS,YAAY,MAAkC;CAErD,MAAM,UAAU,KAAK,QAAQ,kBAAkB,GAAG;CAClD,MAAM,QAAQ,QAAQ,OAAO,OAAO;AACpC,KAAI,UAAU,GAAI,QAAO;CACzB,MAAM,OAAO,QAAQ;CACrB,MAAM,QAAQ,SAAS,MAAM,MAAM;CACnC,IAAI,QAAQ;CACZ,IAAI,WAAW;CACf,IAAI,UAAU;AACd,MAAK,IAAI,IAAI,OAAO,IAAI,QAAQ,QAAQ,KAAK;EAC3C,MAAM,KAAK,QAAQ;AACnB,MAAI,UAAU;AACZ,OAAI,QACF,WAAU;YACD,OAAO,KAChB,WAAU;YACD,OAAO,KAChB,YAAW;AAEb;;AAEF,MAAI,OAAO,KACT,YAAW;WACF,OAAO,KAChB;WACS,OAAO,OAAO;AACvB;AACA,OAAI,UAAU,EAAG,QAAO,QAAQ,MAAM,OAAO,IAAI,EAAE;;;;;;;;;;;;;;AAiBzD,MAAM,sBAAsB;CAC1B;CACA;CACA;CACA;CACA;CACD,CAAC,KAAK,IAAI;AAiBX,SAAS,uBAAuB,OAA4B;AAS1D,QAAO;;;;EARM,MACV,KACE,MACC,MAAM,EAAE,KAAK,iBAAiB,EAAE,YAAY,gBAAgB,KAAK,UAC/D,EAAE,cAAc;EAAE,MAAM;EAAU,YAAY,EAAE;EAAE,CACnD,GACJ,CACA,KAAK,OAAO,CAKV;;;;;;;;;;;AAYP,SAAS,mBAAmB,MAAsD;CAChF,MAAM,SAAS,KAAK,MAAM,2CAA2C;CACrE,MAAM,OAAO,SAAS,OAAO,KAAM,KAAK,MAAM,2BAA2B,GAAG,MAAM;AAClF,KAAI,CAAC,KACH,QAAO;AAET,KAAI;EACF,MAAM,SAAS,KAAK,MAAM,KAAK;AAK/B,MAAI,OAAO,OAAO,SAAS,SACzB,QAAO;GAAE,MAAM,OAAO;GAAM,MAAM,OAAO,aAAa,OAAO,QAAQ,EAAE;GAAE;SAErE;AAGR,QAAO;;AAGT,IAAa,eAAb,MAAa,aAAa;CAMxB,AAAQ;CACR,AAAQ;CACR,AAAQ;CACR,IAAY,MAAkB;AAC5B,MAAI,CAAC,KAAK,KAAM,OAAM,IAAI,MAAM,aAAa,eAAe;AAC5D,SAAO,KAAK;;CAEd,IAAY,WAAqB;AAC/B,MAAI,CAAC,KAAK,UAAW,OAAM,IAAI,MAAM,aAAa,eAAe;AACjE,SAAO,KAAK;;CAEd,IAAY,YAAuB;AACjC,MAAI,CAAC,KAAK,WAAY,OAAM,IAAI,MAAM,aAAa,eAAe;AAClE,SAAO,KAAK;;CAEd,OAAwB,iBACtB;CACF,AAAQ,aAAa;CACrB,AAAQ;;CAER,AAAQ;;CAER,AAAQ;;CAER,AAAQ;;CAER,AAAQ;;CAER,AAAQ;;CAER,AAAQ;;CAER,AAAQ;;CAER,AAAQ;;CAER,AAAQ,WAA2B;;CAEnC,AAAQ,WAA2B;;CAEnC,AAAQ,aAA4B;;CAEpC,AAAQ,kBAAiC;;CAEzC,AAAQ,iCAAiB,IAAI,KAAkD;;CAE/E,AAAQ,aAA+B;;CAEvC,AAAQ,aAA+B;;CAEvC,AAAQ,aAAsC;;;;;;;CAQ9C,AAAQ,qBAAqB;;;;;;;;;;;;;;CAe7B,AAAQ,gBAA+B,QAAQ,SAAS;;CAGxD,AAAS;;CAET,AAAiB;;CAEjB,IAAI,SAA0B;AAC5B,MAAI,CAAC,KAAK,QAAS,OAAM,IAAI,MAAM,aAAa,eAAe;AAC/D,SAAO,KAAK;;CAGd,AAAQ,YACN,KACA,UACA,WACA,OACA,MAQA,QAKA;AACA,OAAK,OAAO;AACZ,OAAK,YAAY;AACjB,OAAK,aAAa;AAIlB,OAAK,eAAe,OAAO,gBAAgB;GAAE,MAAM;GAAM,QAAQ;GAAO,KAAK;GAAO;AACpF,OAAK,UAAU,OAAO,UAAU;AAChC,OAAK,eAAe,OAAO,QAAQ,SAAS,YAAY,IAAI;AAC5D,OAAK,gBAAgB,OAAO,gBAAgB;AAC5C,OAAK,iBAAiB,QAAQ,YAAY;AAC1C,OAAK,eAAe,QAAQ,UAAU;AACtC,OAAK,sBAAsB,QAAQ,iBAAiB;AACpD,OAAK,mBAAmB,KAAK;AAC7B,OAAK,YAAY,KAAK;AACtB,OAAK,YAAY,KAAK;AACtB,OAAK,iBAAiB,KAAK;AAC3B,OAAK,qBAAqB,KAAK,qBAAqB;;;CAItD,IAAI,YAAqB;AACvB,SAAO,KAAK,mBAAmB;;;CAIjC,mBAAyE;AACvE,SAAO,KAAK,SAAS,YAAY;;;CAInC,qBAA2B;AACzB,OAAK,SAAS,cAAc;;;;CAK9B,MAAM,kBAAkB,SAAgC;AACtD,QAAM,KAAK,SAAS,kBAAkB,QAAQ;;;;;CAMhD,iBAAoE;AAClE,SAAO;GACL,YAAY,KAAK,SAAS;GAC1B,mBAAmB,KAAK,SAAS;GAClC;;;;;;;;CAkBH,MAAM,YAAY,QAAsC;AACtD,MAAI,WAAW,MAAM;AACnB,QAAK,SAAS,kBAAkB;AAChC,QAAK,kBAAkB;AACvB;;EAEF,MAAM,UAAU,MAAM,aAAa,QAAQ;GACzC,SAAS,KAAK,eAAe;GAC7B,UAAU,KAAK,eAAe;GAC/B,CAAC;AACF,MAAI,CAAC,QACH,OAAM,IAAI,MAAM,uBAAuB,OAAO,GAAG;EAGnD,MAAM,SAAS,gBAAgB,SADb,uBAAuB,KAAK,cAAc,CACV;EAClD,MAAM,EAAE,SAAS,YAAY,KAAK,SAAS,iBAAiB,OAAO;AACnE,MAAI,YAAY,EACd,OAAM,IAAI,MACR,WAAW,OAAO,+CAA+C,KAAK,cAAc,YAAY,QAAQ,IACzG;AAEH,UAAQ,IACN,4BAA4B,OAAO,YAAY,QAAQ,YAAY,UAAU,aAAa,YAAY,GAAG,GAC1G;AACD,OAAK,kBAAkB;;;CAIzB,aAA4B;AAC1B,SAAO,KAAK;;;;;;;;;;CAWd,MAAM,gBAAgB,MAAc,QAA+B;AACjE,OAAK,gBAAgB;AACrB,MAAI,CAAC,KAAK,SAAS,iBACjB,OAAM,IAAI,MACR,4FACD;AAEH,MAAI,KAAK,eAAe,IAAI,KAAK,CAAE;EACnC,MAAM,UAAU,MAAM,aAAa,QAAQ;GACzC,SAAS,KAAK,eAAe;GAC7B,UAAU,KAAK,eAAe;GAC/B,CAAC;AACF,MAAI,CAAC,QACH,OAAM,IAAI,MAAM,uBAAuB,OAAO,GAAG;EAGnD,MAAM,SAAS,gBAAgB,SADb,uBAAuB,KAAK,cAAc,CACV;EAClD,MAAM,KAAK,KAAK,SAAS,qBAAqB,OAAO;AACrD,OAAK,eAAe,IAAI,MAAM;GAAE;GAAI;GAAQ,CAAC;AAC7C,UAAQ,IAAI,2BAA2B,KAAK,KAAK,OAAO,qBAAqB,GAAG,GAAG;;;CAIrF,wBAAkC;AAChC,SAAO,CAAC,GAAG,KAAK,eAAe,MAAM,CAAC;;;;;;;;;CAUxC,AAAQ,SAAS,OAAqB;AACpC,MAAI;AACF,OAAI,OAAO,iBAAiB,YAC1B,cAAa,QAAQ,sBAAsB,MAAM;UAE7C;;;CAMV,IAAI,cAAuB;AACzB,SAAO,KAAK;;;;;;;;;;;CAYd,AAAQ,uBAAuB,QAA4B;AACzD,MAAI,CAAC,KAAK,mBAAoB;AAC9B,OAAK,qBAAqB;EAG1B,MAAM,IAAI,KAAK,IAAI,OAAO,QAAQ,IAAI;EACtC,IAAI,UAAU;EACd,IAAI,UAAU;EACd,IAAI,SAAS;EACb,MAAM,QAAQ,OAAO;AACrB,OAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;GAC1B,MAAM,IAAI,OAAO;AACjB,OAAI,CAAC,OAAO,SAAS,EAAE,EAAE;AACvB,aAAS;AACT;;AAEF,OAAI,MAAM,EAAG,WAAU;AACvB,OAAI,MAAM,MAAO,WAAU;;AAG7B,oBADgB,UAAU,CAAC,WAAW,CAAC,QACb;;;;;;;CAQ5B,aAAa,OAAO,UAA+B,EAAE,EAAyB;AAI5E,YAAU;GAAE,GAAG;GAAS,MAAM,mBAAmB,QAAQ;GAAE;AAW3D,MAAI,UAAU,QAAQ,KAAK,CACzB,QAAO,aAAa,cAAc,QAAQ;EAI5C,MAAM,MAAM,MAAM,SAAS;AAO3B,MAAI;AACF,UAAO,MAAM,aAAa,kBAAkB,KAAK,QAAQ;WAClD,KAAK;AACZ,OAAI;AACF,QAAI,OAAO,SAAS;WACd;AAGR,SAAM;;;;;;;;;;CAWV,OAAe,cAAc,SAA4C;AACvE,SAAO,IAAI,aACT,MACA,MACA,MACA,MACA;GACE,iBAAiB;GACjB,WAAW,EAAE;GAGb,WAAW,QAAQ,aAAa;GAChC,eAAe;GAChB,EACD,KACD;;CAGH,aAAqB,kBACnB,KACA,SACuB;EAEvB,MAAM,YAAY,OAAO,cAAc,eAAe,OAAO,aAAa;EAC1E,MAAM,WACJ,aAAa,SAAS,KAAK,UAAU,UAAU,IAAI,CAAC,SAAS,KAAK,UAAU,UAAU;EAGxF,MAAM,SAAS,YAAY,IAAI,gBAAgB,SAAS,OAAO,GAAG;EAClE,MAAM,aAAa,QAAQ,IAAI,QAAQ;EACvC,MAAM,iBAAiB,QAAQ,IAAI,SAAS;EAC5C,MAAM,gBAAgB,QAAQ,IAAI,QAAQ;EAI1C,IAAIC;AACJ,MAAI,QAAQ,OACV,UAAS,QAAQ;WACR,WACT,UAAS;WACA,YAAY,IAAI,OAEzB,UAAS;WACA,IAAI,OACb,UAAS;MAET,UAAS;EAGX,MAAM,UAAU,WAAW,QAAQ,QAAQ;AAE3C,UAAQ,IACN,oBAAoB,OAAO,aAAa,QAAQ,mBAAmB,IAAI,OAAO,YAAY,WAC3F;EAGD,MAAM,YAAY,UAAkB;AAClC,OAAI;AACF,QAAI,OAAO,iBAAiB,YAAa,cAAa,QAAQ,sBAAsB,MAAM;WACpF;;AAMV,WAAS,uBAAuB;EAMhC,MAAM,mBAAmB,QAAQ,qBAAqB,IAAI,iBAAiB,OAAO;EAClF,MAAM,aAAa,QAAQ,eACvB,EAAE,iBAAiB,KAAK,KAAK,mBAAmB,EAAE,EAAE,GACpD;EACJ,MAAM,EAAE,OAAO,WAAW,SAAS,WAAW,cAAc,MAAM,UAAU;GAC1E,GAAG;GACH,MAAM,mBAAmB,QAAQ;GACjC;GACA;GACD,CAAC;AACF,WAAS,uBAAuB,QAAQ,KAAK,UAAU;EAMvD,IAAIC,eAIO;EAOX,MAAM,kBAAkB,UAAU,iBAAiB;EACnD,MAAM,iBACF,UAAU,eAAuD,eACjE,mBAAmB,CAAC,GAAG,QAAQ,MAAM,CAAC,CAAC,MAAM,MAAM,EAAE,WAAW,gBAAgB,CAAC;EACrF,MAAM,YAAY,iBACd,yDACA;EACJ,MAAM,cAAc,iBAAiB,kBAAkB;EACvD,MAAM,mBAAmB,QAAQ,MAAM,CAAC,MAAM,MAAM,EAAE,WAAW,YAAY,CAAC;AAE9E,MADmB,QAAQ,QAAQ,gBAAgB,gBAAgB,IACjD,kBAAkB;GAClC,MAAM,WAAW,iBACb,0BAA0B,UAAU,GACpC,2BAA2B,UAAU;GACzC,MAAM,aAAa;GAOnB,MAAM,6BAAa,IAAI,KAAyD;AAChF,QAAK,MAAM,KAAK,QAAQ,MAAM,CAC5B,KAAI,EAAE,WAAW,YAAY,IAAI,EAAE,WAAW,gBAAgB,IAAI,MAAM,WAAW;IACjF,MAAM,IAAI,MAAM,QAAQ,IAAI,EAAE;AAC9B,QAAI,EAAG,YAAW,IAAI,GAAG,EAAE;;AAK/B,OAAI,gBAAgB;AAClB,2BAAuB,UAAU,WAAW;IAC5C,MAAM,KAAK,wBAAwB,UAAU;AAK7C,qCAAiC,aAJrB,UAAU,uBAAuB,UAAU,eAG5B,cAAyB,IACQ,GAAG,YAAY,GAAG,WAAW;AAKzF,oCAAgC,YAAY,GAAG,YAAY,GAAG,WAAW;;GAE3E,MAAM,UAAU,IAAIC,eAAmB,KAAK,UAAU,WAAW;GAEjE,MAAM,OAAO,WAAW,IAAI,UAAU;GACtC,MAAM,WACJ,QAAQ,KAAK,gBAAgB,eACzB,IAAI,aAAa,KAAK,KAAK,GAC3B,OACE,IAAI,aACF,KAAK,KAAK,QACV,KAAK,KAAK,YACV,KAAK,KAAK,aAAa,EACxB,CAAC,OAAO,GACT;AACR,SAAM,QAAQ,cAAc,WAAW;AACvC,WAAQ,gBAAgB;AACxB,OAAI,SACF,gBAAe;IACb,UAAU;IACV,QAAS,UAAU,iBAA6C;IAChE,eAAe;IAChB;;EAKL,IAAIC;AACJ,MAAI,eACF,aAAY,KAAK,IAAI,OAAO,SAAS,gBAAgB,GAAG,EAAE,MAAM,OAAO,eAAe;WAC7E,cAAc,SAEvB,aAAY,KAAK,IAAI,QAAQ,aAAa,MAAM,MAAM,OAAO,gBAAgB,KAAK;WACzE,IAAI,eAGb,aAAY,KAAK,IAAI,QAAQ,aAAa,KAAK,MAAM,OAAO,gBAAgB,KAAK;MAEjF,aAAY,KAAK,IACf,QAAQ,aAAa,MAAM,OAAO,gBAClC,MAAM,OAAO,gBACb,KACD;AAGH,UAAQ,IAAI,uBAAuB,UAAU,kBAAkB,MAAM,eAAe;AAGpF,MAAI,IAAI,eACN,SAAQ,IACN,yCAAyC,IAAI,OAAO,cAAc,gCACjC,IAAI,OAAO,4BAA4B,mCACpC,IAAI,OAAO,iCAChD;AAMH,WAAS,4BAA4B;AACrC,MAAI,OAAO,eAAe,gBAAgB;AAC1C,MAAI,OAAO,eAAe,aAAa;EAYvC,MAAM,mBAAmB,gBAAgB,OAAO,SAAS,eAAe,GAAG,GAAG;EAC9E,IAAIC;EAGJ,IAAI,eAAe;AACnB,MAAI,kBAAkB;AACpB,qBAAkB;AAClB,WAAQ,IACN,sCAAsC,gBAAgB,4BACvD;aACQ,IAAI,gBAAgB;AAc7B,qBAAkB,uBAAuB,EAAE,UAAU,MAAM,CAAC;AAC5D,kBAAe;;EAEjB,MAAM,WAAW,IAAI,SAAS,KAAK,OAAO;GAAE;GAAW;GAAQ;GAAiB,CAAC;AAGjF,MAAI,UACF,UAAS,aAAa,UAAU;AAElC,WAAS,2BAA2B;AACpC,QAAM,SAAS,cAAc,QAAQ;AAIrC,QAAM,QAAQ,WAAW;AACzB,WAAS,2BAA2B;AACpC,WAAS,gBAAgB;EACzB,MAAM,kBAAkB,MAAM,IAAI,OAAO,eAAe;EACxD,MAAM,WAAW,MAAM,IAAI,OAAO,eAAe;AACjD,MAAI,YAAY,iBAAiB;GAC/B,MAAM,SAAS,CACb,WAAW,kBAAkB,SAAS,YAAY,MAClD,kBAAkB,eAAe,gBAAgB,YAAY,KAC9D,CACE,OAAO,QAAQ,CACf,KAAK,KAAK;AACb,YAAS,oBAAoB,OAAO,MAAM,GAAG,IAAI,GAAG;AACpD,SAAM,IAAI,MACR,qBAAqB,OAAO,4GAC7B;;AAEH,WAAS,eAAe;AAExB,SAAO,IAAI,aACT,KACA,UACA,WACA,OACA;GACE,iBAAiB,QAAQ,WAAW;GACpC;GACA;GACA,eAAe;GACf,mBAAmB;GACpB,EACD,aACD;;;;;;;;;;;;;;;;;;CAmBH,MAAM,YACJ,SACA,SACA,SAC4B;AAC5B,OAAK,gBAAgB;AACrB,MAAI,CAAC,KAAK,kBAAkB,CAAC,KAAK,gBAAgB,CAAC,KAAK,oBACtD,OAAM,IAAI,MACR,qIAED;EAEH,MAAM,OAAO,KAAK;AAGlB,MAAI,KAAK,eAAe,QAAQ;GAC9B,MAAM,OAAO,wBAAwB,EAAE,eAAe,MAAM,CAAC;GAC7D,MAAM,QAAQ,QAAQ;GACtB,MAAM,QAAQ,QAAQ;GAEtB,MAAM,UAAU,KAAK,MAAM,KAAK,oBAAoB,UAAU,IAAI,KAAK,YAAY;GACnF,MAAMC,SAAO,iCACX,OACA,OACA,KAAK,qBACL,SACA;IACE,YAAY,KAAK;IACjB,UAAU,KAAK;IACf,SAAS,KAAK;IACd,WAAW,KAAK;IAChB,mBAAmB,KAAK;IACzB,CACF;AACD,UAAO,KAAK,eAAe,aACzB;IACE;IACA,WAAWA,OAAK;IAChB,KAAKA,OAAK;IACV,KAAKA,OAAK;IACV,YAAYA,OAAK;IACjB,YAAYA,OAAK;IACjB,WAAWA,OAAK;IACjB,EACD,QACD;;EAIH,MAAM,aAAa,KAAK;EACxB,MAAM,WAAW,KAAK;EACtB,MAAM,OAAO,2BAA2B,SAAS,KAAK,qBAAqB;GACzE;GACA;GACA,uBAAuB,KAAK;GAC5B,kBAAkB,KAAK;GACvB,WAAW;GACZ,CAAC;EACF,MAAM,aAAa,QAAQ,KAAK,QAAQ,KAAK,QAAQ;AACrD,SAAO,KAAK,eAAe,OACzB;GACE;GACA,WAAW,KAAK;GAChB,KAAK,KAAK;GACV,KAAK,KAAK;GACV;GACD,EACD,QACD;;;CAMH,AAAQ,cAMN;EACA,MAAM,MAAM,KAAK,aAAa,EAAE;EAEhC,MAAM,QADW,IAAI,eAA2C,KAC1C,mBAA+C,EAAE;EACvE,MAAM,UAAU,KAAK,OAAO;EAC5B,MAAM,UAAW,KAAK,yBAAoC;EAC1D,MAAM,UAAU,KAAK,MAAM,UAAU,QAAQ;EAC7C,MAAM,QAAS,KAAK,cAAyB,KAAK,OAAO,aAAa;EACtE,MAAM,UAAY,KAAK,iBAA8B;GAAC;GAAI;GAAI;GAAG;EACjE,MAAM,SAAU,IAAI,iBAA6C,EAAE;AACnE,SAAO;GACL;GACA;GACA;GACA,cAAe,IAAI,kBAA6B;GAChD,WAAY,OAAO,sBAAiC;GACrD;;;;;;;CAQH,AAAQ,kBAAkB,cAA0B,KAAa,QAA4B;EAC3F,MAAM,EAAE,SAAS,OAAO,YAAY,KAAK,aAAa;EACtD,MAAM,EAAE,KAAK,QAAQ,iBAAiB,cAAc,KAAK,SAAS,OAAO,QAAQ;AACjF,OAAK,SAAS,WAAW,aAAa,IAAI;AAC1C,OAAK,SAAS,WAAW,aAAa,IAAI;AAC1C,OAAK,SAAS,WAAW,kBAAkB,OAAO;EAElD,MAAM,OAAO,MAAM;AACnB,SAAO,KAAK,IAAI,aAAa,OAAO,aAAa,MAAM,OAAO,aAAa,IAAI,MAAM,MAAM;;;;;;CAO7F,AAAQ,qBAAqB,QAAgB,YAA0B;EACrE,MAAM,EAAE,SAAS,OAAO,YAAY,KAAK,aAAa;EAEtD,MAAM,MAAM,IAAI,WAAW,EAAE;AAC7B,MAAI,KAAK;AACT,MAAI,KAAK;AACT,MAAI,KAAK;EACT,MAAM,EAAE,KAAK,QAAQ,iBAAiB,KAAK,GAAG,SAAS,OAAO,QAAQ;EACtE,MAAM,WAAW,UAAU;AAC3B,OAAK,SAAS,aAAa,aAAa,KAAK,SAAS,SAAS;AAC/D,OAAK,SAAS,aAAa,aAAa,KAAK,SAAS,SAAS;;;CAIjE,AAAQ,qBAAqB,KAAmB;EAC9C,MAAM,EAAE,SAAS,OAAO,YAAY,KAAK,aAAa;EAEtD,MAAM,MAAM,IAAI,WAAW,IAAI,IAAI;AACnC,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,KAAK;AAC5B,OAAI,KAAK;AACT,OAAI,MAAM,KAAK;AACf,OAAI,IAAI,MAAM,KAAK;;EAErB,MAAM,EAAE,KAAK,QAAQ,iBAAiB,KAAK,KAAK,SAAS,OAAO,QAAQ;AACxE,OAAK,SAAS,WAAW,aAAa,IAAI;AAC1C,OAAK,SAAS,WAAW,aAAa,IAAI;EAC1C,MAAM,SAAS,IAAI,WAAW,IAAI,CAAC,KAAK,GAAG;AAC3C,OAAK,SAAS,WAAW,kBAAkB,OAAO;;;;;;;;;CAUpD,MAAM,SACJ,QACA,UAA2B,EAAE,EACJ;AACzB,OAAK,gBAAgB;EACrB,MAAM,UAAU,MAAM,KAAK,iBAAiB;AAC5C,MAAI;AACF,UAAO,MAAM,KAAK,gBAAgB,QAAQ,QAAQ;YAC1C;AACR,YAAS;;;;;;;;;;;;;;;;;;CAmBb,MAAM,cACJ,SACA,UASI,EAAE,EACqB;AAC3B,OAAK,gBAAgB;EACrB,MAAM,IAAI,KAAK,SAAS;AACxB,MAAI,KAAK,EACP,OAAM,IAAI,MAAM,kEAAkE;AAEpF,MAAI,QAAQ,WAAW,EACrB,OAAM,IAAI,MACR,sBAAsB,QAAQ,OAAO,gCAAgC,EAAE,qBAChD,QAAQ,OAAO,GACvC;AAEH,MAAI,KAAK,oBAAoB,KAAK,SAAS,cAAc,IAAI,KAAK,SAAS,kBACzE,OAAM,IAAI,MAAM,4EAA4E;EAE9F,MAAM,EAAE,YAAY,KAAK,gBAAgB,EAAE,EAAE,WAAW,EAAE,EAAE,iBAAiB;AAC7E,OAAK,SAAS,eAAe,OAAQ,KACnC,OAAM,IAAI,MAAM,4EAA4E;EAI9F,MAAM,mBAAmB,QAAQ;AACjC,MAAI,oBAAoB,iBAAiB,WAAW,EAClD,OAAM,IAAI,MACR,sBAAsB,iBAAiB,OAAO,gCAAgC,IAC/E;AAGH,OADiB,kBAAkB,MAAM,MAAM,KAAK,KAAK,IAAI,UAC7C,CAAC,KAAK,SAAS,iBAC7B,OAAM,IAAI,MAAM,+DAA+D;EAEjF,MAAMC,eAAkE,IAAI,MAAM,EAAE,CAAC,KAAK,KAAK;AAC/F,MAAI,iBACF,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;GAC1B,MAAM,OAAO,iBAAiB;AAC9B,OAAI,QAAQ,KAAM;GAClB,MAAM,MAAM,KAAK,eAAe,IAAI,KAAK;AACzC,OAAI,CAAC,IACH,OAAM,IAAI,MACR,2BAA2B,KAAK,+CACjC;AAEH,gBAAa,KAAK;;EAItB,MAAM,UAAU,MAAM,KAAK,iBAAiB;AAC5C,MAAI;GACF,MAAM,YAAY,YAAY,KAAK;AACnC,QAAK,SAAS,YAAY;GAE1B,MAAM,QAAQ,KAAK,UAAU,OAAO;GACpC,MAAM,QAAQ,KAAK,oBAAoB;GACvC,MAAM,WAAW,KAAK,OAAO;GAC7B,MAAM,WAAW,YACZ,MAAsB,iBAAiB,GAAG,SAAS,IACnD,MAAsB;GAE3B,MAAM,OAAO,QAAQ,KAAK,IAAI,OAAO;IACnC,cAAc,EAAE;IAChB,MAAM;IACN,UAAU;IACV,cAAc;IAEd,OAAO,QAAQ,kBAAkB,MAAM;IACxC,EAAE;GAGH,MAAM,cAAc,IAAI,YAAY,EAAE;AACtC,QAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;IAC1B,MAAM,IAAI,QAAQ;IAClB,MAAMC,WACJ,OAAO,MAAM,WACT,CACE,GAAI,eAAe,CAAC;KAAE,MAAM;KAAmB,SAAS;KAAc,CAAC,GAAG,EAAE,EAC5E;KAAE,MAAM;KAAiB,SAAS;KAAG,CACtC,GACD,eACE,CAAC;KAAE,MAAM;KAAmB,SAAS;KAAc,EAAE,GAAG,EAAE,GAC1D;IACR,MAAM,WAAW,KAAK,UAAU,WAAW,UAAU,EAAE,qBAAqB,MAAM,CAAC;IAKnF,MAAM,UAAU,aAAa;AAC7B,QAAI,QAAS,MAAK,SAAS,iBAAiB,QAAQ,OAAO;AAC3D,QAAI;AACF,UAAK,SAAS,eAAe,EAAE;KAC/B,MAAM,EAAE,WAAW,MAAM,KAAK,SAAS,QAAQ,IAAI,YAAY,SAAS,CAAC;AACzE,UAAK,SAAS,gBAAgB,EAAE;AAChC,iBAAY,KAAK,SAAS,YAAY,QAAQ,EAAE,aAAa,GAAG,CAAC,CAAC;cAC1D;AACR,SAAI,QAAS,MAAK,SAAS,kBAAkB;;;AAOjD,OAAI,KAAK,SAAS,iBAChB,MAAK,SAAS,qBAAqB,aAAa,KAAK,MAAO,IAAI,EAAE,KAAK,GAAI,CAAC;GAG9E,MAAM,WAAW,KAA4B,QAAsB;AACjE,QAAI,IAAI,SAAU;AAClB,QAAI,aAAa,KAAK,IAAI;AAC1B,QAAK,UAAU,QAAQ,QAAQ,SAAW,UAAU,QAAQ,QAAQ,OAAQ;AAC1E,SAAI,WAAW;AACf,SAAI,eAAe;AACnB;;AAEF,QAAI,QAAQ,KAAK,UAAU,OAAO,CAAC,IAAI,EAAE,KAAK;AAC9C,QAAI,cAAc,MAAM,MAAM,IAAI,KAAK,SAAS,EAAE,CAAC,EAAE;AACnD,UAAK,MAAM,KAAK,eAAe;MAC7B,MAAM,MAAM,IAAI,KAAK,QAAQ,EAAE;AAC/B,UAAI,QAAQ,GAAI,KAAI,OAAO,IAAI,KAAK,MAAM,GAAG,IAAI;;AAEnD,SAAI,WAAW;AACf,SAAI,eAAe;AACnB;;AAEF,QAAI,IAAI,aAAa,UAAU,IAAI,MAAO,KAAI,WAAW;;GAG3D,MAAM,cAAc,YAAY,KAAK;AACrC,QAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IAAK,SAAQ,KAAK,IAAI,YAAY,GAAG;GAG5D,MAAM,QAAQ,SAAS;GACvB,MAAM,OAAO,KAAK,SAAS;GAC3B,IAAI,WAAW,OAAO;AACtB,QAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IACrB,YAAW,KAAK,IAAI,UAAU,KAAK,YAAY,KAAK,KAAK,EAAE;GAE7D,MAAM,cAAc,KAAK,QAAQ,GAAG,MAAM,KAAK,IAAI,GAAG,EAAE,MAAM,EAAE,EAAE;GAClE,MAAM,cAAc,KAAK,IAAI,GAAG,KAAK,IAAI,cAAc,GAAG,SAAS,CAAC;GACpE,IAAI,YAAY;GAChB,IAAI,gBAAgB;AACpB,UAAO,gBAAgB,eAAe,CAAC,KAAK,OAAO,MAAM,EAAE,SAAS,EAAE;AACpE,WAAO,YAAY,eAAe,YAAY,gBAAgB,OAAO;AACnE,UAAK,SAAS,sBACZ,cAAc,IAAI,cAAc,MAChC,YAAY,MACb;AACD;;IAEF,MAAM,OAAO,MAAM,KAAK,SAAS,gBAAgB,gBAAgB,MAAM;AACvE;AACA,SAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IAAK,SAAQ,KAAK,IAAI,KAAK,GAAG;;GAGvD,MAAM,aAAa,YAAY,KAAK,GAAG;GACvC,MAAM,YAAY,YAAY,KAAK,GAAG;AACtC,UAAO,KAAK,KAAK,OAAO;IACtB,MAAM,EAAE;IACR,iBAAiB,EAAE,aAAa;IAChC,iBAAiB,EAAE,aAAa,UAAU,YAAY;IACtD;IACA;IACA,cAAc,EAAE;IAChB,UAAU,CAAC,GAAG,EAAE,aAAa;IAC9B,EAAE;YACK;AACR,YAAS;;;;;;;;;;;CAYb,kBAAoC;AAClC,OAAK,gBAAgB;AACrB,MAAI,KAAK,WAAY,QAAO,KAAK;AACjC,MAAI,KAAK,SAAS,aAAa,EAC7B,OAAM,IAAI,MAAM,oEAAoE;AAEtF,MAAI,KAAK,oBAAoB,KAAK,SAAS,cAAc,IAAI,KAAK,SAAS,kBACzE,OAAM,IAAI,MAAM,4EAA4E;EAE9F,MAAM,WAAW,KAAK,OAAO;AAmB7B,OAAK,aAAa,IAAI,iBAlBM;GAC1B,UAAU,KAAK;GACf,aAAa,aAAa,KAAK,UAAU,WAAW,UAAU,EAAE,qBAAqB,MAAM,CAAC;GAC5F,cAAc,OAAO,KAAK,UAAU,OAAO,CAAC,GAAG,EAAE,KAAK;GACtD,YAAY,KAAK,UAAU,OAAO;GAClC,YAAY,KAAK,oBAAoB;GACrC,YAAY,YAAY,MAAM,iBAAiB,GAAG,SAAS,IAAI,MAAM;GACrE,iBAAiB,SAAS;IACxB,MAAM,MAAM,KAAK,eAAe,IAAI,KAAK;AACzC,QAAI,CAAC,IACH,OAAM,IAAI,MACR,uBAAuB,KAAK,+CAC7B;AAEH,WAAO;;GAET,mBAAmB,KAAK,iBAAiB;GAC1C,CAC2C;AAC5C,SAAO,KAAK;;;;;;;;;;;;;;;;CAiBd,MAAM,YAA6C,SAA6C;AAC9F,OAAK,gBAAgB;AACrB,MAAI,CAAC,KAAK,SAAS,iBACjB,OAAM,IAAI,MACR,wFACD;EAEH,MAAM,YAAY,KAAK,iBAAiB;AACxC,QAAM,QAAQ,IACZ,OAAO,QAAQ,QAAQ,QAAQ,CAAC,KAAK,CAAC,MAAM,YAC1C,UAAU,OAAO,QAAQ,SAAS,GAAG,KAAK,gBAAgB,MAAM,OAAO,CACxE,CACF;AACD,SAAO,IAAI,MAAM,WAAW,QAAQ;;;;;;;;;;;;CAatC,AAAQ,qBAAoC;AAC1C,SAAO,KAAK,UAAU,UAAU,UAAU,IAAI,KAAK,UAAU,UAAU,gBAAgB;;CAGzF,MAAc,gBACZ,QACA,UAA2B,EAAE,EACJ;AACzB,OAAK,gBAAgB;EAErB,MAAM,EAAE,YAAY,KAAK,gBAAgB,EAAE,EAAE,WAAW,EAAE,EAAE,cAAc,YAAY;AAGtF,OAAK,SAAS,OAAO;EAGrB,IAAIC;AACJ,MAAI,OAAO,WAAW,UAAU;GAE9B,MAAMD,WAA0B,EAAE;AAClC,OAAI,aACF,UAAS,KAAK;IAAE,MAAM;IAAU,SAAS;IAAc,CAAC;AAE1D,YAAS,KAAK;IAAE,MAAM;IAAQ,SAAS;IAAQ,CAAC;AAChD,cAAW,KAAK,UAAU,WAAW,UAAU,EAAE,qBAAqB,MAAM,CAAC;SACxE;GAEL,MAAM,WAAW,eACb,CAAC;IAAE,MAAM;IAAmB,SAAS;IAAc,EAAE,GAAG,OAAO,GAC/D;AACJ,cAAW,KAAK,UAAU,WAAW,UAAU,EAAE,qBAAqB,MAAM,CAAC;;EAG/E,MAAM,YAAY,YAAY,KAAK;EACnC,MAAM,YAAY,SAAS,eAAe,MAAO;EAOjD,MAAM,WAAW,KAAK,oBAAoB,KAAK,SAAS,UAAU,YAAY;AAC9E,MAAI,KAAK,kBACP;OAAI,SACF,MAAK,qBAAqB,SAAS,OAAO;YACjC,KAAK,SAAS,UAAU,iBAAiB,CAClD,MAAK,SAAS,WAAW,kBAAkB,IAAI,WAAW,SAAS,OAAO,CAAC,KAAK,GAAG,CAAC;;EAKxF,IAAI,EAAE,WAAW,MAAM,KAAK,SAAS,QAAQ,IAAI,YAAY,SAAS,CAAC;AAMvE,OAAK,uBAAuB,OAAO;EAGnC,MAAME,eAAyB,EAAE;EACjC,IAAIC,eAA+C;EACnD,IAAI,gBAAgB;EAEpB,MAAM,QAAQ,KAAK,UAAU,OAAO;EACpC,MAAM,QAAQ,KAAK,oBAAoB;EAEvC,MAAM,cAAc,YAAY,KAAK;EAGrC,MAAM,gBAAgB,cAA+B;AACnD,gBAAa,KAAK,UAAU;AAG5B,OAAK,UAAU,QAAQ,cAAc,SAAW,UAAU,QAAQ,cAAc,OAAQ;AACtF,mBAAe;AACf,WAAO;;GAIT,MAAM,YAAY,KAAK,UAAU,OAAO,CAAC,UAAU,EAAE,KAAK;AAC1D,oBAAiB;AACjB,OAAI,SAAS;IACX,MAAM,YAAY,YAAY,KAAK,GAAG;IACtC,MAAM,aAAa,aAAa;AAChC,YAAQ,WAAW;KACjB;KACA;KACA,KAAK,YAAY,IAAK,aAAa,YAAa,MAAO;KACxD,CAAC;;AAGJ,OAAI,cAAc,MAAM,MAAM,cAAc,SAAS,EAAE,CAAC,EAAE;AAExD,SAAK,MAAM,KAAK,eAAe;KAC7B,MAAM,MAAM,cAAc,QAAQ,EAAE;AACpC,SAAI,QAAQ,GACV,iBAAgB,cAAc,MAAM,GAAG,IAAI;;AAG/C,mBAAe;AACf,WAAO;;AAET,UAAO;;EAQT,MAAM,WAAW;EACjB,IAAI,eAAe,SAAS;EAW5B,MAAM,aAAa,KAAK,SAAS,cAAc;EAQ/C,MAAM,WAAW,KAAK,OAAO;EAC7B,MAAM,WAAW,YACZ,MAAsB,iBAAiB,GAAG,SAAS,IACnD,MAAsB;AAE3B,MAAI,YAAY,CAAC,KAAK,SAAS,qBAAqB,CAAC,YAAY,CAAC,YAAY;GAU5E,MAAM,aAAa,SAAS,YAAY,QAAQ,UAAU,CAAC,GAAG,UAAU,GAAG,aAAa,CAAC,CAAC;AAC1F,OAAI,CAAC,aAAa,WAAW,EAAE;IAC7B,MAAM,cAAc,KAAK,IAAI,YAAY,GAAG,KAAK,SAAS,yBAAyB,CAAC;IACpF,MAAM,UAAU,KAAK,SAAS;AAC9B,QAAI,UAAU,GAAG;KASf,IAAI,WAAW;KACf,IAAI,UAAU;AACd,YAAO,CAAC,WAAW,WAAW,aAAa;MACzC,MAAM,QAAQ,KAAK,IAAI,SAAS,cAAc,SAAS;AACvD,WAAK,SAAS,yBAAyB,aAAa,IAAI,aAAa,MAAM,OAAO,EAAE;MACpF,MAAM,SAAS,MAAM,KAAK,SAAS,iBAAiB,GAAG,MAAM;AAC7D,kBAAY;AACZ,WAAK,MAAM,OAAO,OAChB,KAAI,aAAa,IAAI,EAAE;AACrB,iBAAU;AACV;;;WAID;KACL,MAAM,QAAQ,SAAS;KACvB,IAAI,YAAY;KAChB,IAAI,WAAW;AACf,YAAO,WAAW,aAAa;AAC7B,aAAO,YAAY,eAAe,YAAY,WAAW,OAAO;AAC9D,YAAK,SAAS,uBACZ,cAAc,IAAI,aAAa,MAC/B,YAAY,MACb;AACD;;MAEF,MAAM,MAAM,MAAM,KAAK,SAAS,gBAAgB,WAAW,MAAM;AACjE;AACA,UAAI,aAAa,IAAI,CAAE;;;;SAIxB;GAIL,MAAM,gBAAgB,CAAC,YAAY;AACnC,QAAK,IAAI,OAAO,GAAG,OAAO,WAAW,QAAQ;IAC3C,IAAIC;AACJ,QAAI,SAAS,KAAK,cAChB,aAAY,SAAS,YAAY,QAAQ,UAAU,CAAC,GAAG,UAAU,GAAG,aAAa,CAAC,CAAC;SAC9E;AAGL,SAAI,SAAU,MAAK,qBAAqB,KAAK,SAAS,eAAe,aAAa;AAElF,iBAAY,MAAM,KAAK,SAAS,cAC9B,IAAI,YAAY,CAAC,aAAa,aAAa,SAAS,GAAG,CAAC,CACzD;AACD;;AAGF,QAAI,aAAa,UAAU,CAAE;AAE7B,QAAI,eAAe;AACjB,SAAI,SAAU,MAAK,qBAAqB,KAAK,SAAS,eAAe,aAAa;AAElF,eADe,MAAM,KAAK,SAAS,QAAQ,IAAI,YAAY,CAAC,UAAU,CAAC,CAAC,EACxD;AAChB;;;;EAKN,MAAM,YAAY,YAAY,KAAK,GAAG;EACtC,MAAM,kBAAkB,aAAa;EACrC,MAAM,kBAAkB,mBAAmB,YAAY;AAEvD,SAAO;GACL,MAAM;GACN;GACA;GACA;GACA;GACA,UAAU,CAAC,GAAG,aAAa;GAC5B;;;;;;;;;;;;;;;;;;;;;CAsBH,MAAM,aACJ,QACA,OASI,EAAE,EACW;AAUjB,SAAO,iBATK,MAAM,KAAK,SAAS,QAAQ;GACtC,cAAc;GACd,WAAW,KAAK,aAAa;GAC7B,UAAU;IACR,aAAa,KAAK,eAAe;IACjC,mBAAmB,KAAK,qBAAqB;IAC9C;GACD,eAAe,KAAK,QAAQ,CAAC,KAAK;GACnC,CAAC,EACyB,MAAM,QAAQ;GACvC,YAAY,KAAK,cAAc;GAC/B,UAAU,KAAK;GAChB,CAAC;;;;;;;;;;;CAYJ,MAAM,QACJ,MACA,OAKI,EAAE,EACW;EACjB,MAAM,SACJ,KAAK,gBACL,gCAAgC,KAAK,QAAQ,eAAe;AAO9D,UALY,MAAM,KAAK,SAAS,MAAM;GACpC,cAAc;GACd,WAAW,KAAK,aAAa;GAC7B,UAAU,EAAE,aAAa,KAAK,eAAe,IAAK;GACnD,CAAC,EACS,KAAK,MAAM;;;;;;;;;;;;;CAcxB,MAAM,kBACJ,QACA,MAO+C;EAC/C,MAAM,EAAE,OAAO,WAAW,GAAG,QAAQ,WAAW,aAAa;EAC7D,MAAM,eAAe,uBAAuB,MAAM;EAClD,MAAMJ,WACJ,OAAO,WAAW,WAAW,CAAC;GAAE,MAAM;GAAQ,SAAS;GAAQ,CAAC,GAAG,CAAC,GAAG,OAAO;EAChF,MAAMK,QAAqB,EAAE;EAC7B,IAAI,YAAY;AAEhB,OAAK,IAAI,IAAI,GAAG,IAAI,UAAU,KAAK;GACjC,MAAM,SAAS,MAAM,KAAK,SAAS,UAAU;IAAE;IAAc;IAAW;IAAU,CAAC;GACnF,MAAM,OAAO,mBAAmB,OAAO,KAAK;AAC5C,OAAI,CAAC,MAAM;AACT,gBAAY,OAAO;IACnB,MAAMC,SAAoB;KAAE,MAAM;KAAU,MAAM,OAAO;KAAM;AAC/D,UAAM,KAAK,OAAO;AAClB,aAAS,OAAO;AAChB;;GAGF,MAAMC,WAAsB;IAAE,MAAM;IAAa,MAAM,KAAK;IAAM,MAAM,KAAK;IAAM;AACnF,SAAM,KAAK,SAAS;AACpB,YAAS,SAAS;GAElB,MAAM,OAAO,MAAM,MAAM,MAAM,EAAE,SAAS,KAAK,KAAK;GACpD,IAAIC;AACJ,OAAI,KACF,KAAI;AACF,iBAAa,OAAO,MAAM,KAAK,QAAQ,KAAK,KAAK,CAAC;YAC3C,GAAG;AACV,iBAAa,mBAAmB,KAAK,KAAK,IAAI;;OAGhD,cAAa,wBAAwB,KAAK,KAAK;GAGjD,MAAMC,aAAwB;IAAE,MAAM;IAAe,MAAM,KAAK;IAAM,QAAQ;IAAY;AAC1F,SAAM,KAAK,WAAW;AACtB,YAAS,WAAW;AAEpB,YAAS,KAAK;IAAE,MAAM;IAAa,SAAS,OAAO;IAAM,CAAC;AAC1D,YAAS,KAAK;IAAE,MAAM;IAAQ,SAAS,QAAQ,KAAK,KAAK,cAAc;IAAc,CAAC;AACtF,eAAY;;AAGd,SAAO;GAAE,MAAM;GAAW;GAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA6BnC,MAAM,eACJ,QACA,UAAiC,EAAE,EACD;AAClC,OAAK,gBAAgB;EACrB,MAAM,EAAE,QAAQ,aAAa,GAAG,GAAG,iBAAiB;EAEpD,MAAM,YAAY,UAA4B;AAC5C,OAAI,OAAO,WAAW,WACpB,QAAQ,OAAmC,MAAM;AAEnD,OAAI,UAAU,OAAO,WAAW,YAAY,MAAM,QAAQ,OAAO,SAAS,EAAE;AAC1E,QAAI,UAAU,QAAQ,OAAO,UAAU,SAAU,QAAO;IACxD,MAAM,MAAM;AACZ,WAAO,OAAO,SAAS,OAAO,QAAQ,OAAO,IAAI;;AAGnD,UAAO;;EAST,MAAM,QAAQ,uCAHZ,UAAU,OAAO,WAAW,YAAY,MAAM,QAAQ,OAAO,SAAS,GAClE,KAAK,OAAO,SAAS,KAAK,KAAK,CAAC,MAChC,sBACqD;EAE3D,MAAM,cAAc,KAAK,IAAI,GAAG,WAAW,GAAG;EAC9C,IAAI,WAAW;EACf,IAAI,YAAY;AAChB,OAAK,IAAI,UAAU,GAAG,WAAW,aAAa,WAAW;GACvD,MAAM,mBAAmB,YAAY,IAAI,SAAS,SAAS;GAC3D,MAAM,SAAS,MAAM,KAAK,SAAS,kBAAkB,aAAa;AAClE,cAAW,OAAO;GAClB,MAAM,SAAS,YAAY,OAAO,KAAK;AACvC,OAAI,WAAW,QAAW;AACxB,gBAAY;AACZ;;GAEF,IAAIC;AACJ,OAAI;AACF,YAAQ,KAAK,MAAM,OAAO;YACnB,GAAG;AACV,gBAAY,sBAAsB,aAAa,QAAQ,EAAE,UAAU,OAAO,EAAE;AAC5E;;AAEF,OAAI,CAAC,SAAS,MAAM,EAAE;AACpB,gBAAY;AACZ;;AAEF,UAAO;IAAE,QAAQ;IAAY,MAAM,OAAO;IAAM,UAAU;IAAS;;AAGrE,QAAM,IAAI,MACR,+BAA+B,YAAY,eAAe,UAAU,iBAClD,KAAK,UAAU,SAAS,MAAM,GAAG,IAAI,CAAC,GACzD;;;;;;;;;;;;;;;;;CAkBH,MAAM,MACJ,MACA,UA2CI,EAAE,EACoF;AAC1F,OAAK,gBAAgB;EACrB,MAAM,OAAO,KAAK,eAAe,QAAQ;AAGzC,MAAI,UAAU,KAAK,KAAK,IAAI,UAAU,KAAK,KAAK,cAAc,CAC5D,QAAO,KAAK,YAAY,MAAM,QAAQ;AAIxC,MAAI,WAAW,KAAK,KAAK,IAAI,QAAQ,KAAK,KAAK,cAAc,CAC3D,QAAO,KAAK,UAAU,MAAM,QAAQ;AAItC,MAAI,UAAU,KAAK,KAAK,IAAI,UAAU,KAAK,KAAK,cAAc,CAC5D,QAAO,KAAK,YAAY,MAAM,QAAQ;AAYxC,MAAI,EAHF,KAAK,kBAAkB,yBACvB,YAAY,KAAK,KAAK,IACtB,WAAW,KAAK,KAAK,EAErB,OAAM,IAAI,MACR,4DAA4D,eAAe,IAAI,QACtE,eAAe,QAAQ,wBAAwB,KAAK,cAAc,WAAW,KAAK,KAC5F;AAKH,MAAI,CAAC,KAAK,SACR,MAAK,WAAW,MAAM,QAAQ,OAAO;GACnC,MAAM,KAAK,eAAe;GAC1B,UAAU,KAAK,eAAe;GAC9B,SAAS,KAAK,eAAe;GAC7B,UAAU,KAAK,eAAe;GAC9B,WAAW,KAAK;GAChB,OAAO,QAAQ;GAChB,CAAC;AAEJ,SAAO,KAAK,SAAS,MAAM,MAAM;GAC/B,aAAa,QAAQ,SAAS,QAAQ;GACtC,aAAa,QAAQ;GACrB,MAAM,QAAQ;GACd,MAAM,QAAQ;GACd,mBAAmB,QAAQ;GAC3B,WAAW,QAAQ;GACpB,CAAC;;;;;;;CAQJ,MAAc,YACZ,MACA,SAC0F;AAC1F,MAAI,CAAC,KAAK,WACR,MAAK,aAAa,MAAM,UAAU,OAAO;GACvC,MAAM,KAAK,eAAe;GAC1B,UAAU,KAAK,eAAe;GAC9B,SAAS,KAAK,eAAe;GAC7B,UAAU,KAAK,eAAe;GAC9B,WAAW,KAAK;GACjB,CAAC;EAEJ,MAAMC,YAAgC;GACpC,eAAe,QAAQ;GACvB,aAAa,QAAQ;GACrB,MAAM,QAAQ;GACd,WAAW,QAAQ;GACpB;AACD,SAAO,KAAK,WAAW,MAAM,MAAM,UAAU;;;CAI/C,yBAAiC;AAC/B,SAAO,UAAU,wBAAwB;;;;;;;;;CAU3C,MAAc,YACZ,MACA,SAc0F;AAC1F,MAAI,CAAC,KAAK,WACR,MAAK,aAAa,MAAM,UAAU,OAAO;GACvC,MAAM,KAAK,eAAe;GAC1B,UAAU,KAAK,eAAe;GAC9B,SAAS,KAAK,eAAe;GAC7B,UAAU,KAAK,eAAe;GAC9B,WAAW,KAAK;GAChB,OAAO,QAAQ;GAChB,CAAC;AAEJ,SAAO,KAAK,WAAW,MAAM,MAAM;GACjC,aAAa,QAAQ;GACrB,MAAM,QAAQ;GACd,mBAAmB,QAAQ;GAC3B,WAAW,QAAQ;GACnB,eAAe,QAAQ;GACvB,SAAS,QAAQ;GAClB,CAAC;;;;;;;CAQJ,MAAc,UACZ,MACA,SAU0F;AAC1F,QAAM,KAAK,WAAW,QAAQ;AAC9B,SAAO,KAAK,SAAU,MAAM,MAAM;GAChC,aAAa,QAAQ;GACrB,MAAM,QAAQ;GACd,MAAM,QAAQ;GACd,mBAAmB,QAAQ;GAE3B,SAAS,QAAQ;GAClB,CAAC;;;;;;;;CASJ,MAAc,WAAW,SAIP;EAChB,MAAM,QAAQ,QAAQ,SAAS,eAAe;AAC9C,MAAI,KAAK,YAAY,KAAK,eAAe,OAAO;AAC9C,QAAK,SAAS,SAAS;AACvB,QAAK,WAAW;;AAElB,MAAI,KAAK,SACP;EAGF,MAAM,WAAW,KAAK,eAAe;EACrC,IAAI,aAAa;AACjB,MAAI,UAAU;GACZ,MAAM,QAAQ,GAAG,SAAS,QAAQ,OAAO,GAAG,CAAC,GAAG,MAAM;AAEtD,QADW,MAAM,OAAO,WAAW,YAAY,KAAK,GAC5C,WAAW,MAAM,CACvB,cAAa;;EAGjB,MAAM,UAAU,MAAM,gBAAgB,YAAY,eAAe,gBAAgB;AACjF,OAAK,WAAW,MAAM,QAAQ,OAAO;GACnC,MAAM,KAAK,eAAe;GAC1B,UAAU,KAAK,eAAe;GAC9B,SAAS,KAAK,eAAe;GAC7B,UAAU,KAAK,eAAe;GAC9B,WAAW,KAAK;GAChB,OAAQ,QAAQ,SAAmC;GACnD,kBAAkB,QAAQ,oBAAoB,eAAe;GAC7D;GACD,CAAC;AACF,OAAK,aAAa;;;;;;;;CASpB,MAAM,cAAc,KAIK;AACvB,OAAK,gBAAgB;EACrB,MAAM,OAAO,KAAK,eAAe,QAAQ;AACzC,MAAI,EAAE,WAAW,KAAK,KAAK,IAAI,QAAQ,KAAK,KAAK,cAAc,EAC7D,OAAM,IAAI,MACR,yDAAyD,eAAe,QAAQ,wBACzD,KAAK,cAAc,WAAW,KAAK,KAC3D;AAEH,QAAM,KAAK,WAAW,EAAE,CAAC;AACzB,SAAO,KAAK,SAAU,cAAc,IAAI;;;CAI1C,YAAsB;AACpB,MAAI,WAAW,KAAK,KAAK,eAAe,QAAQ,GAAG,IAAI,QAAQ,KAAK,KAAK,cAAc,CACrF,QAAO,CAAC,GAAG,sBAAsB;AAEnC,SAAO,KAAK,UAAU,iBAAiB,IAAI,EAAE;;;;;;;;;;;;;CAc/C,MAAM,cACJ,OAGA,SAAS,wBACT,UAAuE,EAAE,EAChD;AACzB,OAAK,gBAAgB;EACrB,MAAM,UAAU,MAAM,KAAK,iBAAiB;AAC5C,MAAI;AACF,UAAO,MAAM,KAAK,qBAAqB,OAAO,QAAQ,QAAQ;YACtD;AACR,YAAS;;;CAIb,MAAc,qBACZ,OAGA,SAAS,wBACT,UAAuE,EAAE,EAChD;AACzB,OAAK,gBAAgB;AACrB,MAAI,CAAC,KAAK,eACR,OAAM,IAAI,MACR,mIAED;AAUH,MAAI,KAAK,eAAe,QAAQ;AAC9B,OAAI,CAAC,KAAK,iBACR,OAAM,IAAI,MACR,sKAGD;GAEH,IAAIC;GACJ,IAAIC;AACJ,OAAI,aAAa,OAAO;AACtB,gBAAY,MAAM;AAClB,eAAW,CAAC,MAAM,QAAQ,IAAI,MAAM,QAAQ,GAAG;UAC1C;IACL,MAAMC,QAAM,sBAAsB,MAAM,QAAQ,MAAM,OAAO,MAAM,OAAO;AAC1E,gBAAYA,MAAI;AAChB,eAAWA,MAAI;;GAEjB,MAAMC,OAAiC;IAAC;IAAG,SAAS;IAAI,SAAS;IAAG;GACpE,MAAMC,eAAa,KAAK,KAAK,KAAK,KAAK,KAAK;AAC5C,QAAK,SAAS,yBAAyBA,eAAa;GACpD,MAAMC,WAAS,MAAM,KAAK,YAAY,WAAW,OAAO,OAAO,SAAS;IACtE,MAAM,SAAS,MAAM,SAAS,OAAO,KAAK,KAAK,UAAU;AACzD,SAAK,SAAS,YAAY,QAAQ,SAAS;KAC3C;GAMF,MAAM,MAAM,KAAK,aAAa,EAAE;GAChC,MAAMC,iBAAgB,IAAI,kBAA6B;GACvD,MAAM,QAAQ,KAAK,UAAU,UAAU,WAAW,IAAK,IAAI,gBAA2B;GACtF,MAAM,QAAQ,KAAK,UAAU,UAAU,WAAW,IAAK,IAAI,gBAA2B;GACtF,MAAM,MAAM,KAAK,UAAU,OAAO,WAC7B,KAAK,UAAU,UAAU,KAAK,UAAU,OAAO,SAAS,IAAI,EAAE,GAC/D,EAAE;GACN,MAAM,SAAS,MAAM,QAAQ,IAAI,GAAG,MAAM,CAAC,IAAI;GAI/C,MAAM,eAAe,KAAK,UAAU,OAAO,gBAAgB;GAC3D,MAAM,UAAU,KAAK,UAAU,OAAO,OAAO;GAC7C,MAAM,WAAW,KAAK,UAAU,OAAO,OAAO;GAC9C,MAAM,YAAY,KAAK,UAAU,OAAO,OAAO;GAC/C,MAAM,YAAY,KAAK,UAAU,OAAO,0BAA0B;GAClE,MAAMC,aAAqB,IAAI,MAAMF,SAAO,KAAK,CAAC,KAAKC,eAAa;GACpE,MAAMjB,aAAqB;IACzB,GAAG;IACH,GAAG;IACH,GAAG;IACH;IACA,GAAGmB;IACH;IACA,GAAG;IACH,GAAG;IACH,GAAG;IACJ;AAED,UAAO,KAAK,oBAAoBH,SAAO,QAAQI,YAAUH,gBAAc,QAAQ;;AAGjF,MAAI,CAAC,KAAK,iBACR,OAAM,IAAI,MACR,0IAED;EAIH,MAAM,UAAU,QAAQ,kBAAkB;EAC1C,IAAII;EACJ,IAAIC;AACJ,MAAI,aAAa,OAAO;AACtB,aAAU,MAAM;AAChB,aAAU,MAAM;SACX;GACL,MAAMT,QAAM,gBAAgB,MAAM,QAAQ,MAAM,OAAO,MAAM,QAAQ,QAAQ;AAC7E,aAAUA,MAAI;AACd,aAAUA,MAAI;;EAIhB,MAAM,aAAa,QAAQ,KAAK,QAAQ,KAAK,QAAQ;AACrD,OAAK,SAAS,yBAAyB,aAAa;EACpD,MAAM,SAAS,MAAM,KAAK,YAAY,SAAS,UAAU,OAAO,SAAS;GACvE,MAAM,SAAS,MAAM,SAAS,OAAO,KAAK,KAAK,UAAU;AACzD,QAAK,SAAS,YAAY,QAAQ,SAAS;IAC3C;EAKF,MAAM,WAAW,KAAK,UAAU,UAAU,mBAAmB,IAAI;EACjE,MAAM,SAAS,KAAK,UAAU,UAAU,iBAAiB,IAAI;EAC7D,MAAM,EAAE,iBAAiB,KAAK,aAAa;EAC3C,MAAM,iBAAiB,OAAO;EAC9B,MAAM,MAAM,KAAK,UAAU,OAAO,qBAAqB;EACvD,MAAM,OAAO,KAAK,UAAU,OAC1B,GAAG,OAAO,4DACX;EACD,MAAMK,WAAqB,IAAI,MAAM,eAAe,CAAC,KAAK,aAAa;EACvE,MAAMlB,WAAqB;GAAC,GAAG;GAAK;GAAU,GAAG;GAAU;GAAQ,GAAG;GAAK;AAE3E,SAAO,KAAK,cAAc,OAAO,QAAQ,SAAS,UAAU,QAAQ;;;;;;;CAQtE,AAAQ,yBACN,cACA,SACA,UAC4B;EAC5B,MAAM,EAAE,cAAc,cAAc,KAAK,aAAa;EACtD,MAAM,MAAM,SAAS;AACrB,MAAI,MAAM,KAAK,UACb,OAAM,IAAI,MACR,kCAAkC,IAAI,sBAAsB,KAAK,UAAU,8CAE5E;AAEH,OAAK,SAAS,OAAO;AACrB,OAAK,SAAS,WAAW,iBAAiB,aAAa;EAEvD,MAAM,SAAS,IAAI,WAAW,IAAI,CAAC,KAAK,GAAG;EAC3C,IAAI,IAAI;AACR,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,IACvB,KAAI,SAAS,OAAO,aAAc,QAAO,KAAK;EAEhD,MAAM,eAAe,sBAAsB,UAAU,CAAC,QAAQ,EAAE,cAAc,UAAU;AAExF,SAAO,EAAE,gBADc,KAAK,kBAAkB,cAAc,KAAK,OAAO,EAC/C;;;;;;;;;;CAW3B,MAAc,oBACZ,cACA,UACA,cACA,SACyB;EACzB,MAAM,MAAM,SAAS;AACrB,MAAI,MAAM,KAAK,UACb,OAAM,IAAI,MACR,kCAAkC,IAAI,sBAAsB,KAAK,UAAU,8CAE5E;AAEH,OAAK,SAAS,uBAAuB,MAAM;AAC3C,OAAK,SAAS,OAAO;AACrB,OAAK,SAAS,WAAW,iBAAiB,aAAa;EAEvD,MAAM,SAAS,IAAI,WAAW,IAAI,CAAC,KAAK,GAAG;EAC3C,IAAI,IAAI;AACR,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,IACvB,KAAI,SAAS,OAAO,aAAc,QAAO,KAAK;AAEhD,OAAK,SAAS,WAAW,kBAAkB,OAAO;AAElD,OAAK,SAAS,uBAAuB;EACrC,MAAM,EAAE,YAAY,KAAK,gBAAgB,EAAE,EAAE,WAAW,EAAE,EAAE,YAAY;EACxE,MAAM,YAAY,YAAY,KAAK;EACnC,MAAM,YAAY,SAAS,eAAe,MAAO;EAEjD,IAAI,EAAE,WAAW,MAAM,KAAK,SAAS,QAAQ,IAAI,YAAY,SAAS,CAAC;EAEvE,MAAMC,eAAyB,EAAE;EACjC,IAAIC,eAA+C;EACnD,IAAI,gBAAgB;EACpB,MAAM,QAAQ,KAAK,UAAU,OAAO;EAIpC,MAAM,QAAQ,KAAK,oBAAoB;EAEvC,MAAM,gBAAgB,cAA+B;AACnD,gBAAa,KAAK,UAAU;AAC5B,OAAK,UAAU,QAAQ,cAAc,SAAW,UAAU,QAAQ,cAAc,OAAQ;AACtF,mBAAe;AACf,WAAO;;GAET,MAAM,YAAY,KAAK,UAAU,OAAO,CAAC,UAAU,EAAE,KAAK;AAC1D,oBAAiB;AACjB,aAAU,UAAU;AACpB,OAAI,cAAc,MAAM,MAAM,cAAc,SAAS,EAAE,CAAC,EAAE;AACxD,SAAK,MAAM,KAAK,eAAe;KAC7B,MAAM,MAAM,cAAc,QAAQ,EAAE;AACpC,SAAI,QAAQ,GAAI,iBAAgB,cAAc,MAAM,GAAG,IAAI;;AAE7D,mBAAe;AACf,WAAO;;AAET,UAAO;;AAGT,OAAK,IAAI,OAAO,GAAG,OAAO,WAAW,QAAQ;GAC3C,IAAIC;AACJ,OAAI,SAAS,EACX,aAAY,YAAY,QAAQ,UAAU,CAAC,GAAG,UAAU,GAAG,aAAa,CAAC;YAChE,SACT,aAAY,MAAM,KAAK,SAAS,cAC9B,IAAI,YAAY,CAAC,aAAa,aAAa,SAAS,GAAG,CAAC,CACzD;OAED,aAAY,YAAY,QAAQ,UAAU,CAAC,GAAG,UAAU,GAAG,aAAa,CAAC;AAG3E,OAAI,aAAa,UAAU,CAAE;AAE7B,OAAI,CAAC,SAEH,WADe,MAAM,KAAK,SAAS,QAAQ,IAAI,YAAY,CAAC,UAAU,CAAC,CAAC,EACxD;;EAIpB,MAAM,YAAY,YAAY,KAAK,GAAG;EACtC,MAAM,kBAAkB,aAAa;AACrC,SAAO;GACL,MAAM;GACN;GACA,iBAAiB,mBAAmB,YAAY;GAChD;GACA;GACA,UAAU,CAAC,GAAG,aAAa;GAC5B;;;CAIH,MAAc,cACZ,cACA,SACA,UACA,SACyB;AACzB,OAAK,SAAS,uBAAuB,SAAS,SAAS;EACvD,MAAM,EAAE,mBAAmB,KAAK,yBAAyB,cAAc,SAAS,SAAS;AACzF,OAAK,SAAS,uBAAuB;AACrC,SAAO,KAAK,qBAAqB,UAAU,iBAAiB,GAAG,QAAQ;;;;;;;;CASzE,MAAM,uBACJ,SACA,SACA,UAC6E;AAC7E,OAAK,gBAAgB;AACrB,MAAI,CAAC,KAAK,oBAAoB,CAAC,KAAK,eAClC,OAAM,IAAI,MAAM,4EAA4E;EAE9F,MAAM,SAAS,MAAM,KAAK,YAAY,SAAS,QAAQ;AACvD,OAAK,yBAAyB,OAAO,QAAQ,SAAS,SAAS;EAC/D,MAAM,EAAE,WAAW,MAAM,KAAK,SAAS,QAAQ,IAAI,YAAY,SAAS,CAAC;AAKzE,SAAO;GAAE,eAJa,MAAM,KAAK,SAAS,gBACxC,iBACA,SAAS,SAAS,KAAK,OAAO,YAC/B;GACuB;GAAQ,KAAK,SAAS;GAAQ;;;;;;;CAQxD,MAAc,qBACZ,UACA,gBACA,SACyB;EACzB,MAAM,EAAE,YAAY,KAAK,gBAAgB,EAAE,EAAE,WAAW,EAAE,EAAE,YAAY;EACxE,MAAM,YAAY,YAAY,KAAK;EACnC,MAAM,YAAY,SAAS,eAAe,MAAO;EAEjD,IAAI,EAAE,WAAW,MAAM,KAAK,SAAS,QAAQ,IAAI,YAAY,SAAS,CAAC;EAEvE,MAAMF,eAAyB,EAAE;EACjC,IAAIC,eAA+C;EACnD,IAAI,gBAAgB;EACpB,MAAM,QAAQ,KAAK,UAAU,OAAO;EACpC,MAAM,QAAQ,KAAK,oBAAoB;EACvC,IAAI,eAAe;EAEnB,MAAM,gBAAgB,cAA+B;AACnD,gBAAa,KAAK,UAAU;AAC5B,OAAK,UAAU,QAAQ,cAAc,SAAW,UAAU,QAAQ,cAAc,OAAQ;AACtF,mBAAe;AACf,WAAO;;GAET,MAAM,YAAY,KAAK,UAAU,OAAO,CAAC,UAAU,EAAE,KAAK;AAC1D,oBAAiB;AACjB,aAAU,UAAU;AACpB,OAAI,cAAc,MAAM,MAAM,cAAc,SAAS,EAAE,CAAC,EAAE;AACxD,SAAK,MAAM,KAAK,eAAe;KAC7B,MAAM,MAAM,cAAc,QAAQ,EAAE;AACpC,SAAI,QAAQ,GAAI,iBAAgB,cAAc,MAAM,GAAG,IAAI;;AAE7D,mBAAe;AACf,WAAO;;AAET,UAAO;;AAGT,OAAK,IAAI,OAAO,GAAG,OAAO,WAAW,QAAQ;GAC3C,IAAIC;AACJ,OAAI,SAAS,EACX,aAAY,YAAY,QAAQ,UAAU,CAAC,GAAG,UAAU,GAAG,aAAa,CAAC;YAChE,UAAU;AACnB,SAAK,qBAAqB,KAAK,SAAS,eAAe,aAAa;AACpE,gBAAY,MAAM,KAAK,SAAS,cAC9B,IAAI,YAAY,CAAC,aAAa,aAAa,SAAS,GAAG,CAAC,CACzD;AACD;SAEA,aAAY,YAAY,QAAQ,UAAU,CAAC,GAAG,UAAU,GAAG,aAAa,CAAC;AAG3E,OAAI,aAAa,UAAU,CAAE;AAE7B,OAAI,CAAC,UAAU;AACb,SAAK,qBAAqB,KAAK,SAAS,eAAe,aAAa;AAEpE,cADe,MAAM,KAAK,SAAS,QAAQ,IAAI,YAAY,CAAC,UAAU,CAAC,CAAC,EACxD;AAChB;;;EAIJ,MAAM,YAAY,YAAY,KAAK,GAAG;EACtC,MAAM,kBAAkB,aAAa;AACrC,SAAO;GACL,MAAM;GACN;GACA,iBAAiB,mBAAmB,YAAY;GAChD;GACA;GACA,UAAU,CAAC,GAAG,aAAa;GAC5B;;;;;;;;;;;;;;;;;CAkBH,MAAM,MAAM,MAAc,UAAwB,EAAE,EAAyB;AAC3E,OAAK,gBAAgB;EACrB,MAAM,UAAU,MAAM,KAAK,iBAAiB;AAC5C,MAAI;AACF,UAAO,MAAM,KAAK,aAAa,MAAM,QAAQ;YACrC;AACR,YAAS;;;CAIb,MAAc,aAAa,MAAc,UAAwB,EAAE,EAAyB;AAC1F,OAAK,gBAAgB;AACrB,MAAI,CAAC,KAAK,aACR,OAAM,IAAI,MACR,2JAGD;AAIH,OAAK,SAAS,OAAO;EAErB,MAAM,EAAE,aAAa,UAAU,YAAY,cAAc;AAIzD,MAFE,KAAK,kBAAkB,qBAAqB,KAAK,kBAAkB,eAEjD;GAKlB,MAAMoB,UAAQ,GADZ,cAAc,wBAAwB,YAAY,YAAY,wBAAwB,QAC9D;GAG1B,IAAIC,QAAM,KAAK,UAAU,OAAOD,QAAM;GAEtC,MAAME,QAAM,aAAa,KAAK,OAAO;AACrC,OAAID,MAAI,SAASC,MACf,SAAMD,MAAI,MAAM,GAAGC,MAAI;AAGzB,UAAO,KAAK,SAAS,MAAM,IAAI,YAAYD,MAAI,CAAC;;EAIlD,MAAM,QAAQ,cAAc,aAAa,YAAY,UAAU,SAAS;EAExE,IAAI,MAAM,KAAK,UAAU,OAAO,MAAM;EAMtC,MAAM,QAAQ,KAAK,UAAU,UAAU,gBAAgB,IAAI,KAAK,UAAU,OAAO;AACjF,MAAI,UAAU,QAAQ,IAAI,IAAI,SAAS,OAAO,MAC5C,KAAI,KAAK,MAAM;EAIjB,MAAM,MAAM,aAAa,KAAK,OAAO;AACrC,MAAI,IAAI,SAAS,KAAK;GACpB,MAAM,OAAO,UAAU,QAAQ,IAAI,IAAI,SAAS,OAAO,QAAQ,CAAC,MAAM,GAAG,EAAE;AAC3E,SAAM,CAAC,GAAG,IAAI,MAAM,GAAG,MAAM,KAAK,OAAO,EAAE,GAAG,KAAK;;AAGrD,SAAO,KAAK,SAAS,MAAM,IAAI,YAAY,IAAI,CAAC;;;;;;;;;;;;;;;;CAiBlD,OAAO,OACL,QACA,UAA2B,EAAE,EACsB;AACnD,OAAK,gBAAgB;EAGrB,MAAME,aAAuB,EAAE;EAC/B,IAAIC,UAA+B;EACnC,IAAI,OAAO;EAGX,MAAM,aAAa,UAAwB;AACzC,cAAW,KAAK,MAAM;AACtB,OAAI,SAAS;IACX,MAAM,IAAI;AACV,cAAU;AACV,OAAG;;;EAKP,MAAM,qBAAoC;AACxC,OAAI,WAAW,SAAS,KAAK,KAC3B,QAAO,QAAQ,SAAS;AAE1B,UAAO,IAAI,SAAe,MAAM;AAC9B,cAAU;KACV;;EAIJ,MAAM,aAAa,KAAK,SAAS,QAAQ;GACvC,GAAG;GACH,UAAU,UAAU;AAClB,YAAQ,UAAU,MAAM;AACxB,cAAU,MAAM;;GAEnB,CAAC,CAAC,MAAM,WAAW;AAClB,UAAO;AAEP,OAAI,SAAS;IACX,MAAM,IAAI;AACV,cAAU;AACV,OAAG;;AAEL,UAAO;IACP;EAGF,IAAI,UAAU;AACd,SAAO,MAAM;AACX,SAAM,cAAc;AAGpB,UAAO,UAAU,WAAW,OAC1B,OAAM,WAAW;AAInB,OAAI,QAAQ,WAAW,WAAW,OAChC;;AAIJ,SAAO,MAAM;;;;;;CAOf,MAAM,gBAAgB,YAAoB,aAA6C;AACrF,SAAO,KAAK,SAAS,gBAAgB,YAAY,YAAY;;;;;;CAO/D,MAAM,WAAyC;AAC7C,SAAO,UAAU,KAAK,IAAI;;;;;;CAO5B,aAAa,gBAA8C;EACzD,MAAM,MAAM,MAAM,SAAS;EAC3B,MAAM,SAAS,MAAM,UAAU,IAAI;AACnC,qBAAmB,IAAI,OAAO;AAC9B,MAAI,OAAO,SAAS;AACpB,SAAO;;;;;;CAOT,MAAM,WAAW,UAA0D;AACzE,SAAO,KAAK,SAAS,QAAQ,SAAS;;;;;CAMxC,aAAmB;AACjB,OAAK,SAAS,OAAO;;;;;CAMvB,OAAO,MAAwB;AAC7B,SAAO,KAAK,UAAU,OAAO,KAAK;;;;;CAMpC,OAAO,KAAe,mBAAqC;AACzD,SAAO,KAAK,UAAU,OAAO,KAAK,kBAAkB;;;;;;;;;;;;CAatD,MAAM,iBAAgD;AACpD,OAAK,gBAAgB;EACrB,MAAMC,SAAgC,EAAE;EACxC,MAAM,OAAO,OAAe,SAAqC;GAC/D,MAAM,IAAI,gBAAgB,eAAe,OAAO,IAAI,aAAa,KAAK,OAAO;GAC7E,IAAI,MAAM;GACV,IAAI,SAAS;GACb,IAAIC,WAAS;AACb,QAAK,IAAI,IAAI,GAAG,IAAI,EAAE,QAAQ,KAAK;AACjC,WAAO,EAAE;AACT,QAAI,EAAE,KAAK,QAAQ;AACjB,cAAS,EAAE;AACX,gBAAS;;;GAGb,MAAMC,QAA6B;IACjC;IACA,QAAQ,EAAE;IACV,KAAK,KAAK,MAAM,MAAM,IAAI,GAAG;IAC7B,QAAQ;KACN,KAAK,MAAM,EAAE,KAAK,IAAI,GAAG;KACzB,KAAK,MAAM,EAAE,KAAK,IAAI,GAAG;KACzB,KAAK,MAAM,EAAE,KAAK,IAAI,GAAG;KACzB,KAAK,MAAM,EAAE,KAAK,IAAI,GAAG;KAC1B;IACD;IACA,QAAQ,KAAK,MAAM,SAAS,IAAI,GAAG;IACpC;AACD,UAAO,KAAK,MAAM;AAClB,UAAO;;AAIT,SAAO,KAAK;GACV,OAAO,oCAAoC,KAAK,IAAI,eAAe,qBAAqB,KAAK,SAAS;GACtG,QAAQ;GACR,KAAK;GACL,QAAQ;IAAC;IAAG;IAAG;IAAG;IAAE;GACpB,QAAQ;GACR,QAAQ;GACR,MAAM,KAAK,IAAI;GAChB,CAAC;AAKF,OAAK,MAAM,QAFS,CAAC,eAAe,kCAAkC,CAGpE,KAAI;AAEF,OAAI,MADS,MAAM,KAAK,gBAAgB,KAAK,CAC9B;UACT;AACN,UAAO,KAAK;IACV,OAAO;IACP,QAAQ;IACR,KAAK;IACL,QAAQ;KAAC;KAAK;KAAK;KAAK;KAAI;IAC5B,QAAQ;IACR,QAAQ;IACR,OAAO;IACR,CAAC;;AAKN,OAAK,MAAM,SAAS;GAClB;GACA;GACA;GACA;GACD,CACC,KAAI;GACF,MAAM,IAAI,MAAM,KAAK,gBAAgB,OAAO,GAAG;AAC/C,OAAI,GAAG,MAAM,iBAAiB,EAAE;UAC1B;AACN,UAAO,KAAK;IACV,OAAO;IACP,QAAQ;IACR,KAAK;IACL,QAAQ;KAAC;KAAK;KAAK;KAAK;KAAI;IAC5B,QAAQ;IACR,QAAQ;IACR,OAAO;IACR,CAAC;;AAKN,OAAK,SAAS,OAAO;AACrB,MAAI;GACF,MAAM,SAAS,MAAM,KAAK,SAAS,mBAAmB,IAAI,YAAY,CAAC,EAAE,CAAC,CAAC;GAC3E,MAAM,QAAQ,IAAI,mBAAmB,OAAO,OAAO,IAAI,OAAO,OAAO;AACrE,SAAM,OAAO,YAAY,OAAO,aAAa,KAAK,IAAI,CAAC,QAAQ,OAAO;WAC/DC,GAAQ;AACf,UAAO,KAAK;IACV,OAAO;IACP,QAAQ;IACR,KAAK;IACL,QAAQ;KAAC;KAAK;KAAK;KAAK;KAAI;IAC5B,QAAQ;IACR,QAAQ;IACR,OAAO,EAAE;IACV,CAAC;;AAMJ,MAAI;GACF,MAAM,SAAS,MAAM,KAAK,SAAS,mBAAmB,GAAG,EAAE;GAC3D,MAAM,QAAQ,IAAI,mBAAmB,OAAO,OAAO,IAAI,OAAO,OAAO;AACrE,SAAM,OAAO,QAAQ,OAAO;WACrBA,GAAQ;AACf,UAAO,KAAK;IACV,OAAO;IACP,QAAQ;IACR,KAAK;IACL,QAAQ;KAAC;KAAK;KAAK;KAAK;KAAI;IAC5B,QAAQ;IACR,QAAQ;IACR,OAAO,EAAE;IACV,CAAC;;AAMJ,MAAI;GACF,MAAM,SAAS,MAAM,KAAK,SAAS,mBAAmB,GAAG,EAAE;GAC3D,MAAM,QAAQ,IAAI,mBAAmB,OAAO,OAAO,IAAI,OAAO,OAAO;AACrE,SAAM,OAAO,QAAQ,OAAO;WACrBA,GAAQ;AACf,UAAO,KAAK;IACV,OAAO;IACP,QAAQ;IACR,KAAK;IACL,QAAQ;KAAC;KAAK;KAAK;KAAK;KAAI;IAC5B,QAAQ;IACR,QAAQ;IACR,OAAO,EAAE;IACV,CAAC;;AAIJ,OAAK,SAAS,OAAO;EACrB,MAAM,WAAW,KAAK,SAAS,mBAAmB,GAAG,EAAE;AACvD,OAAK,MAAM,KAAK,SACd,QAAO,KAAK;GACV,OAAO,YAAY,EAAE,IAAI,IAAI,EAAE,OAAO,aAAa,EAAE,aAAa,SAAS,EAAE,UAAU;GACvF,QAAQ,EAAE,UAAU;GACpB,KAAK,EAAE,UAAU,QAAQ,GAAG,MAAM,IAAI,GAAG,EAAE;GAC3C,QAAQ,EAAE,UAAU,MAAM,GAAG,EAAE,CAAC,IAAI,OAAO;GAC3C,QAAQ;GACR,QAAQ;GACT,CAAC;AAIJ,OAAK,SAAS,OAAO;AACrB,MAAI;GACF,MAAM,EAAE,WAAW,MAAM,KAAK,SAAS,QAAQ,IAAI,YAAY,CAAC,EAAE,CAAC,CAAC;AACpE,OAAI,mBAAmB,OAAO;AAG9B,OAAI;AAEF,QAAI,aADU,MAAM,KAAK,gBAAgB,aAAa,GAAG,CAClC;WACjB;AAKR,OAAI;IACF,MAAM,SAAS,MAAM,KAAK,SAAS,mBAAmB,EAAE;AACxD,SAAK,MAAM,KAAK,QAAQ;KACtB,MAAM,YAAY,EAAE,gBAAgB,YAAY,EAAE,cAAc,KAAK,IAAI,CAAC,KAAK;KAC/E,MAAMD,QAA6B;MACjC,OAAO,SAAS,EAAE,IAAI,IAAI,EAAE,OAAO,KAAK,EAAE,SAAS;MACnD,QAAQ;MACR,KAAK,EAAE;MACP,QAAQ,EAAE;MACV,QAAQ;MACR,QAAQ,KAAK,IAAI,GAAG,EAAE,OAAO,IAAI,KAAK,IAAI,CAAC;MAC5C;AAED,SAAI,EAAE,QAAQ,KAAK,EAAE,OAAO,OAAO,MAAM,MAAM,EAAE,EAAE;AACjD,YAAM,QAAQ;AACd,YAAM,OAAO;;AAEf,YAAO,KAAK,MAAM;;YAEbC,GAAQ;AACf,WAAO,KAAK;KACV,OAAO;KACP,QAAQ;KACR,KAAK;KACL,QAAQ;MAAC;MAAK;MAAK;MAAK;MAAI;KAC5B,QAAQ;KACR,QAAQ;KACR,OAAO,EAAE;KACV,CAAC;;WAEGA,GAAQ;AACf,UAAO,KAAK;IACV,OAAO;IACP,QAAQ;IACR,KAAK;IACL,QAAQ;KAAC;KAAK;KAAK;KAAK;KAAI;IAC5B,QAAQ;IACR,QAAQ;IACR,OAAO,EAAE;IACV,CAAC;;AAIJ,OAAK,SAAS,OAAO;EAMrB,MAAMC,YAA+D;GACnE,eAAe;IACb,KAAK;IACL,QAAQ;KAAC;KAAU;KAAM;KAAU;KAAO;IAC3C;GACD,mCAAmC;IACjC,KAAK;IACL,QAAQ;KAAC;KAAK;KAAU;KAAU;KAAS;IAC5C;GACD,WAAW;IACT,KAAK;IACL,QAAQ;KAAC;KAAU;KAAQ;KAAU;KAAU;IAChD;GACF;AAGD,OAAK,MAAM,SAAS,QAAQ;GAC1B,MAAM,MAAM,UAAU,MAAM;AAC5B,OAAI,KAAK;IAEP,MAAM,WADW,KAAK,IAAI,MAAM,MAAM,IAAI,IAAI,GAClB,KAAK,IAAI,KAAK,IAAI,IAAI,IAAI,GAAG,MAAO,EAAE;IAClE,MAAM,cAAc,MAAM,OAAO,OAAO,GAAG,MAAM,KAAK,IAAI,IAAI,IAAI,OAAO,GAAG,GAAG,IAAK;AACpF,UAAM,QAAQ,YAAY,cAAc,SAAS;AACjD,UAAM,SAAS,IAAI;cACV,MAAM,UAAU,mBAAmB;IAE5C,MAAM,SAAS,MAAM,OAAO,MAAM,MAAM,OAAO,MAAM,EAAE,CAAC;IACxD,MAAM,SAAS,MAAM,OAAO,MAAM,MAAM,CAAC,OAAO,SAAS,EAAE,CAAC;IAC5D,MAAM,UAAU,MAAM,OAAO,OAAO,MAAM,MAAM,MAAM,OAAO,GAAG;IAChE,MAAM,cAAc,MAAM,UAAU,KAAK,MAAM,SAAS,MAAM;AAC9D,UAAM,QAAQ,CAAC,UAAU,CAAC,UAAU,CAAC,WAAW,cAAc,SAAS;;;AAK3E,UAAQ,IAAI,4BAA4B;AACxC,OAAK,MAAM,KAAK,QAAQ;GACtB,MAAM,SAAS,EAAE,QAAQ,QAAS,EAAE,SAAS;AAC7C,WAAQ,IACN,IAAI,OAAO,IAAI,EAAE,MAAM,QAAQ,EAAE,IAAI,YAAY,EAAE,OAAO,KAAK,KAAK,CAAC,YACzD,EAAE,OAAO,QAAQ,EAAE,YAC5B,EAAE,WAAW,SAAY,aAAa,EAAE,OAAO,KAAK,OACpD,EAAE,QAAQ,WAAW,EAAE,UAAU,OACjC,EAAE,OAAO,MAAM,EAAE,SAAS,IAC9B;;AAEH,UAAQ,IAAI,gCAAgC;AAG5C,SAAO;GAAE;GAAQ,SADD,OAAO,OAAO,MAAM,EAAE,UAAU,UAAU,EAAE,UAAU,OAAU;GACtD;;;;;CAM5B,UAAgB;AACd,MAAI,KAAK,WAAY;AACrB,OAAK,aAAa;AAGlB,OAAK,WAAW,SAAS;AACzB,OAAK,UAAU,SAAS;AACxB,OAAK,UAAU,SAAS;AACxB,OAAK,YAAY,SAAS;AAC1B,OAAK,YAAY,SAAS;AAC1B,OAAK,gBAAgB,SAAS;AAC9B,MAAI,KAAK,MAAM;AACb,sBAAmB,KAAK,KAAK,OAAO;AACpC,QAAK,KAAK,OAAO,SAAS;;;CAI9B,AAAQ,iBAAuB;AAC7B,MAAI,KAAK,WACP,OAAM,IAAI,MAAM,kCAAkC;;;;;;;;;;;;;;CAgBtD,MAAc,kBAAuC;EACnD,IAAIC;EACJ,MAAM,OAAO,IAAI,SAAe,YAAY;AAC1C,aAAU;IACV;EACF,MAAM,OAAO,KAAK;AAClB,OAAK,gBAAgB;AACrB,QAAM;AACN,SAAO"}