gputex 0.3.2 → 0.3.4

package/dist/index.js

@@ -95,9 +95,12 @@ var Encoder = class {
   adapter;
   ownsDevice;
   disableF16;
-  // `!:` because these are set in `_buildPipeline()` which the constructor
-  // calls; TypeScript's flow analysis doesn't see through method calls.
-  _module;
+  // f32 module — created lazily by `_ensureModule()`: when the f16 fast
+  // module exists it serves the default path, so parsing/validating the
+  // (larger, dual-quality) f32 source is deferred until 'high' or the
+  // forced-f32 fallback is actually requested. Halves encoder construction
+  // cost on f16 hardware.
+  _module = null;
   // f16 'fast' module — built only when the device supports shader-f16 and the
   // subclass provides an f16 source. null otherwise (falls back to _module).
   _moduleF16 = null;
@@ -106,6 +109,26 @@ var Encoder = class {
   // cache below holds the specialised pipelines for encoders that support it.
   _pipeline;
   _pipelineCache = /* @__PURE__ */ new Map();
+  // -------------------------------------------------------------------- //
+  // Per-encoder GPU resource cache. Creating the source texture, output/
+  // staging buffers and bind group on every encode costs ~1ms of host time
+  // per call — for small/medium images that overhead dominates the encode
+  // (the compute pass itself is tens of µs at 512²). Sequential encodes
+  // (the common case: one texture after another, or a mip chain) reuse
+  // these; concurrent encodes on the same encoder see `_resourcesBusy` and
+  // fall back to transient resources, keeping the API contract unchanged.
+  // Buffers are grow-only, the texture is recreated on size change, and the
+  // bind group is cached per pipeline (with `layout: 'auto'` each pipeline
+  // has its own layout) until any bound resource is recreated.
+  _cachedSrcTex = null;
+  _cachedSrcW = 0;
+  _cachedSrcH = 0;
+  _cachedDst = null;
+  _cachedStaging = null;
+  _cachedParams = null;
+  _lastParams = null;
+  _bindGroupCache = /* @__PURE__ */ new Map();
+  _resourcesBusy = false;
   constructor({ device, adapter, ownsDevice = false, disableF16 = false }) {
     this.device = device;
     this.adapter = adapter;
@@ -115,16 +138,13 @@ var Encoder = class {
   }
   _buildPipeline() {
     const device = this.device;
-    const code = this.wgslSource();
-    this._module = device.createShaderModule({
-      label: `${this.label}-encoder`,
-      code
-    });
     if (this._useF16) {
       this._moduleF16 = device.createShaderModule({
         label: `${this.label}-encoder-f16`,
         code: this.wgslSourceFastF16()
       });
+    } else {
+      this._ensureModule();
     }
     if (this.supportsQuality) {
       this._pipeline = this._getPipeline("fast");
@@ -132,9 +152,19 @@ var Encoder = class {
       this._pipeline = device.createComputePipeline({
         label: `${this.label}-encoder-pipeline`,
         layout: "auto",
-        compute: { module: this._module, entryPoint: "encode" }
+        compute: { module: this._ensureModule(), entryPoint: "encode" }
+      });
+    }
+  }
+  /** The f32 module, parsed on first use (see `_module`). */
+  _ensureModule() {
+    if (!this._module) {
+      this._module = this.device.createShaderModule({
+        label: `${this.label}-encoder`,
+        code: this.wgslSource()
       });
     }
+    return this._module;
   }
   /**
    * Pipeline for a given quality level. Encoders that don't declare a
@@ -159,7 +189,7 @@ var Encoder = class {
       label: `${this.label}-encoder-pipeline-${quality}`,
       layout: "auto",
       compute: {
-        module: this._module,
+        module: this._ensureModule(),
         entryPoint: "encode",
         constants: { QUALITY_HIGH: quality === "high" ? 1 : 0 }
       }
@@ -168,6 +198,15 @@ var Encoder = class {
     return pipeline;
   }
   destroy() {
+    this._cachedSrcTex?.destroy();
+    this._cachedDst?.destroy();
+    this._cachedStaging?.destroy();
+    this._cachedParams?.destroy();
+    this._cachedSrcTex = null;
+    this._cachedDst = null;
+    this._cachedStaging = null;
+    this._cachedParams = null;
+    this._bindGroupCache.clear();
     if (this.ownsDevice) this.device.destroy();
   }
   /** WGSL `@workgroup_size` dimensions. Default 8×8×1. */
@@ -234,93 +273,158 @@ var Encoder = class {
     const blocksY = paddedHeight >> 2;
     const blockCount = blocksX * blocksY;
     const outByteLen = blockCount * this.bytesPerBlock;
-    const srcTex = device.createTexture({
-      label: `${this.label}-src`,
-      size: [paddedWidth, paddedHeight, 1],
-      format: "rgba8unorm",
-      // RENDER_ATTACHMENT is required by copyExternalImageToTexture
-      // (internally a blit) even though we never render into this texture.
-      usage: GPUTextureUsage.COPY_DST | GPUTextureUsage.TEXTURE_BINDING | GPUTextureUsage.RENDER_ATTACHMENT
-    });
-    uploadSourceTexture(device, srcTex, source, width, height, flipY, source instanceof ImageData);
-    const dstBuffer = device.createBuffer({
-      label: `${this.label}-dst`,
-      size: outByteLen,
-      usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC
-    });
-    const paramsBuffer = device.createBuffer({
-      label: `${this.label}-params`,
-      size: 16,
-      usage: GPUBufferUsage.UNIFORM | GPUBufferUsage.COPY_DST
-    });
-    device.queue.writeBuffer(paramsBuffer, 0, new Uint32Array([blocksX, blocksY, width, height]));
-    const pipeline = this._getPipeline(quality);
-    const bindGroup = device.createBindGroup({
-      label: `${this.label}-bg`,
-      layout: pipeline.getBindGroupLayout(0),
-      entries: [
-        { binding: 0, resource: srcTex.createView() },
-        { binding: 1, resource: { buffer: dstBuffer } },
-        { binding: 2, resource: { buffer: paramsBuffer } }
-      ]
-    });
-    const useTimestamps = withGpuTime && device.features.has("timestamp-query");
-    const querySet = useTimestamps ? device.createQuerySet({ type: "timestamp", count: 2 }) : null;
-    const queryBuffer = useTimestamps ? device.createBuffer({
-      label: `${this.label}-ts-resolve`,
-      size: 16,
-      usage: GPUBufferUsage.QUERY_RESOLVE | GPUBufferUsage.COPY_SRC
-    }) : null;
-    const [wgX, wgY] = this.workgroupSize;
-    const t0 = performance.now();
-    const enc = device.createCommandEncoder({ label: `${this.label}-encode` });
-    const pass = enc.beginComputePass(
-      querySet ? { timestampWrites: { querySet, beginningOfPassWriteIndex: 0, endOfPassWriteIndex: 1 } } : void 0
-    );
-    pass.setPipeline(pipeline);
-    pass.setBindGroup(0, bindGroup);
-    pass.dispatchWorkgroups(Math.ceil(blocksX / wgX), Math.ceil(blocksY / wgY), 1);
-    pass.end();
-    if (querySet && queryBuffer) enc.resolveQuerySet(querySet, 0, 2, queryBuffer, 0);
-    const staging = device.createBuffer({
-      label: `${this.label}-staging`,
-      size: outByteLen,
-      usage: GPUBufferUsage.COPY_DST | GPUBufferUsage.MAP_READ
-    });
-    enc.copyBufferToBuffer(dstBuffer, 0, staging, 0, outByteLen);
-    const tsStaging = querySet && queryBuffer ? device.createBuffer({
-      label: `${this.label}-ts-staging`,
-      size: 16,
-      usage: GPUBufferUsage.COPY_DST | GPUBufferUsage.MAP_READ
-    }) : null;
-    if (tsStaging && queryBuffer) enc.copyBufferToBuffer(queryBuffer, 0, tsStaging, 0, 16);
-    device.queue.submit([enc.finish()]);
-    await staging.mapAsync(GPUMapMode.READ);
-    const data = new Uint8Array(staging.getMappedRange().slice(0));
-    staging.unmap();
-    const encodeMs = performance.now() - t0;
-    let gpuMs;
-    if (tsStaging) {
-      await tsStaging.mapAsync(GPUMapMode.READ);
-      const [begin, end] = new BigUint64Array(tsStaging.getMappedRange().slice(0));
-      tsStaging.unmap();
-      tsStaging.destroy();
-      if (end !== void 0 && begin !== void 0 && end > begin) {
-        gpuMs = Number(end - begin) / 1e6;
+    const useCache = !this._resourcesBusy;
+    if (useCache) this._resourcesBusy = true;
+    let srcTex;
+    let dstBuffer;
+    let paramsBuffer;
+    let staging;
+    try {
+      let srcTexIsNew = true;
+      if (useCache && this._cachedSrcTex && this._cachedSrcW === paddedWidth && this._cachedSrcH === paddedHeight) {
+        srcTex = this._cachedSrcTex;
+        srcTexIsNew = false;
+      } else {
+        srcTex = device.createTexture({
+          label: `${this.label}-src`,
+          size: [paddedWidth, paddedHeight, 1],
+          format: "rgba8unorm",
+          // RENDER_ATTACHMENT is required by copyExternalImageToTexture
+          // (internally a blit) even though we never render into this texture.
+          usage: GPUTextureUsage.COPY_DST | GPUTextureUsage.TEXTURE_BINDING | GPUTextureUsage.RENDER_ATTACHMENT
+        });
+        if (useCache) {
+          this._cachedSrcTex?.destroy();
+          this._cachedSrcTex = srcTex;
+          this._cachedSrcW = paddedWidth;
+          this._cachedSrcH = paddedHeight;
+        }
+      }
+      uploadSourceTexture(device, srcTex, source, width, height, flipY, source instanceof ImageData);
+      let dstIsNew = true;
+      if (useCache && this._cachedDst && this._cachedDst.size >= outByteLen) {
+        dstBuffer = this._cachedDst;
+        dstIsNew = false;
+      } else {
+        dstBuffer = device.createBuffer({
+          label: `${this.label}-dst`,
+          size: outByteLen,
+          usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC
+        });
+        if (useCache) {
+          this._cachedDst?.destroy();
+          this._cachedDst = dstBuffer;
+        }
+      }
+      if (useCache && this._cachedStaging && this._cachedStaging.size >= outByteLen) {
+        staging = this._cachedStaging;
+      } else {
+        staging = device.createBuffer({
+          label: `${this.label}-staging`,
+          size: outByteLen,
+          usage: GPUBufferUsage.COPY_DST | GPUBufferUsage.MAP_READ
+        });
+        if (useCache) {
+          this._cachedStaging?.destroy();
+          this._cachedStaging = staging;
+        }
+      }
+      if (useCache) {
+        if (!this._cachedParams) {
+          this._cachedParams = device.createBuffer({
+            label: `${this.label}-params`,
+            size: 16,
+            usage: GPUBufferUsage.UNIFORM | GPUBufferUsage.COPY_DST
+          });
+          this._lastParams = null;
+        }
+        paramsBuffer = this._cachedParams;
+        const lp = this._lastParams;
+        if (!lp || lp[0] !== blocksX || lp[1] !== blocksY || lp[2] !== width || lp[3] !== height) {
+          device.queue.writeBuffer(paramsBuffer, 0, new Uint32Array([blocksX, blocksY, width, height]));
+          this._lastParams = [blocksX, blocksY, width, height];
+        }
+      } else {
+        paramsBuffer = device.createBuffer({
+          label: `${this.label}-params`,
+          size: 16,
+          usage: GPUBufferUsage.UNIFORM | GPUBufferUsage.COPY_DST
+        });
+        device.queue.writeBuffer(paramsBuffer, 0, new Uint32Array([blocksX, blocksY, width, height]));
+      }
+      const pipeline = this._getPipeline(quality);
+      if (useCache && (srcTexIsNew || dstIsNew)) this._bindGroupCache.clear();
+      let bindGroup = useCache ? this._bindGroupCache.get(pipeline) : void 0;
+      if (!bindGroup) {
+        bindGroup = device.createBindGroup({
+          label: `${this.label}-bg`,
+          layout: pipeline.getBindGroupLayout(0),
+          entries: [
+            { binding: 0, resource: srcTex.createView() },
+            { binding: 1, resource: { buffer: dstBuffer } },
+            { binding: 2, resource: { buffer: paramsBuffer } }
+          ]
+        });
+        if (useCache) this._bindGroupCache.set(pipeline, bindGroup);
+      }
+      const useTimestamps = withGpuTime && device.features.has("timestamp-query");
+      const querySet = useTimestamps ? device.createQuerySet({ type: "timestamp", count: 2 }) : null;
+      const queryBuffer = useTimestamps ? device.createBuffer({
+        label: `${this.label}-ts-resolve`,
+        size: 16,
+        usage: GPUBufferUsage.QUERY_RESOLVE | GPUBufferUsage.COPY_SRC
+      }) : null;
+      const [wgX, wgY] = this.workgroupSize;
+      const t0 = performance.now();
+      const enc = device.createCommandEncoder({ label: `${this.label}-encode` });
+      const pass = enc.beginComputePass(
+        querySet ? { timestampWrites: { querySet, beginningOfPassWriteIndex: 0, endOfPassWriteIndex: 1 } } : void 0
+      );
+      pass.setPipeline(pipeline);
+      pass.setBindGroup(0, bindGroup);
+      pass.dispatchWorkgroups(Math.ceil(blocksX / wgX), Math.ceil(blocksY / wgY), 1);
+      pass.end();
+      if (querySet && queryBuffer) enc.resolveQuerySet(querySet, 0, 2, queryBuffer, 0);
+      enc.copyBufferToBuffer(dstBuffer, 0, staging, 0, outByteLen);
+      const tsStaging = querySet && queryBuffer ? device.createBuffer({
+        label: `${this.label}-ts-staging`,
+        size: 16,
+        usage: GPUBufferUsage.COPY_DST | GPUBufferUsage.MAP_READ
+      }) : null;
+      if (tsStaging && queryBuffer) enc.copyBufferToBuffer(queryBuffer, 0, tsStaging, 0, 16);
+      device.queue.submit([enc.finish()]);
+      await staging.mapAsync(GPUMapMode.READ, 0, outByteLen);
+      const data = new Uint8Array(staging.getMappedRange(0, outByteLen).slice(0));
+      staging.unmap();
+      const encodeMs = performance.now() - t0;
+      let gpuMs;
+      if (tsStaging) {
+        await tsStaging.mapAsync(GPUMapMode.READ);
+        const [begin, end] = new BigUint64Array(tsStaging.getMappedRange().slice(0));
+        tsStaging.unmap();
+        tsStaging.destroy();
+        if (end !== void 0 && begin !== void 0 && end > begin) {
+          gpuMs = Number(end - begin) / 1e6;
+        }
+      }
+      querySet?.destroy();
+      queryBuffer?.destroy();
+      return { width, height, paddedWidth, paddedHeight, data, encodeMs, gpuMs };
+    } finally {
+      if (useCache) {
+        this._resourcesBusy = false;
+      } else {
+        srcTex?.destroy();
+        dstBuffer?.destroy();
+        staging?.destroy();
+        paramsBuffer?.destroy();
       }
     }
-    querySet?.destroy();
-    queryBuffer?.destroy();
-    srcTex.destroy();
-    dstBuffer.destroy();
-    staging.destroy();
-    paramsBuffer.destroy();
-    return { width, height, paddedWidth, paddedHeight, data, encodeMs, gpuMs };
   }
 };
 
 // src/bc1.wgsl
-var bc1_default = "// BC1 (DXT1) compute shader encoder.\n//\n// Each invocation encodes one 4x4 pixel block into an 8-byte BC1 block\n// written as 2 x u32 into the destination storage buffer.\n//\n// BC1 block layout (little-endian):\n//   u32[0]: color0 (low 16) | color1 (high 16)   both in RGB565\n//   u32[1]: 16 x 2-bit indices, pixel 0 = bits 0..1, pixel 15 = bits 30..31\n//\n// We always force the 4-color mode (color0 > color1, numeric 16-bit):\n//   idx 0 -> color0\n//   idx 1 -> color1\n//   idx 2 -> (2*color0 +   color1) / 3\n//   idx 3 -> (  color0 + 2*color1) / 3\n//\n// QUALITY LEVELS (pipeline-overridable constant `QUALITY_HIGH`)\n//   fast (0, default): bounding-box endpoints inset by ~half a 565 cell, then\n//     ONE fused pass that projects every pixel onto the decoded-endpoint line\n//     (the 4 palette entries are colinear and evenly spaced, so the nearest\n//     entry is the rounded projection \u2014 no 4-entry search) while accumulating\n//     the least-squares refit sums; the refit endpoints are re-quantised and a\n//     final projection pass assigns the indices, packed on the fly.\n//   high (1): endpoints are seeded from the block's principal colour axis\n//     (covariance power-iteration) as well as the bbox diagonal, each refined by\n//     several least-squares passes with full 4-entry searches; the lower-error\n//     family wins. Mirrors bc1_ref.ts.\n\n// 0 = fast (default), 1 = high-quality. Set via pipeline constants.\noverride QUALITY_HIGH: u32 = 0u;\n\nstruct Params {\n  blocks_x: u32,\n  blocks_y: u32,\n  width:    u32,\n  height:   u32,\n};\n\n@group(0) @binding(0) var src_tex: texture_2d<f32>;\n@group(0) @binding(1) var<storage, read_write> dst: array<u32>;\n@group(0) @binding(2) var<uniform> params: Params;\n\nfn to565(c: vec3<f32>) -> u32 {\n  // Round-to-nearest quantization into 5-6-5.\n  let r = u32(clamp(floor(c.r * 31.0 + 0.5), 0.0, 31.0));\n  let g = u32(clamp(floor(c.g * 63.0 + 0.5), 0.0, 63.0));\n  let b = u32(clamp(floor(c.b * 31.0 + 0.5), 0.0, 31.0));\n  return (r << 11u) | (g << 5u) | b;\n}\n\nfn from565(c: u32) -> vec3<f32> {\n  let r = (c >> 11u) & 31u;\n  let g = (c >>  5u) & 63u;\n  let b =  c         & 31u;\n  // 5/6-bit -> 8-bit: (x*527+23)>>6 (6-bit: 259/33) \u2014 round-to-nearest\n  // scaling, matching bc1_ref.ts and typical hardware decoders (white ->\n  // 255). Integer u32 math is exact. NOTE: this is NOT plain bit-replication\n  // ((x<<3)|(x>>2)) \u2014 they differ for some codes (e.g. 5-bit 3 -> 25 vs 24).\n  // Selecting indices against this palette is what makes the encoder agree\n  // with what the GPU will actually sample.\n  let r8 = (r * 527u + 23u) >> 6u;\n  let g8 = (g * 259u + 33u) >> 6u;\n  let b8 = (b * 527u + 23u) >> 6u;\n  return vec3<f32>(vec3<u32>(r8, g8, b8)) / 255.0;\n}\n\n// 4-color-mode interpolation weights: palette[j] = wa(j)*c0 + wb(j)*c1.\nfn wa(j: u32) -> f32 {\n  switch j {\n    case 0u:  { return 1.0; }\n    case 1u:  { return 0.0; }\n    case 2u:  { return 2.0 / 3.0; }\n    default:  { return 1.0 / 3.0; }   // case 3u\n  }\n}\nfn wb(j: u32) -> f32 {\n  switch j {\n    case 0u:  { return 0.0; }\n    case 1u:  { return 1.0; }\n    case 2u:  { return 1.0 / 3.0; }\n    default:  { return 2.0 / 3.0; }   // case 3u\n  }\n}\n\nfn build_palette(c0: u32, c1: u32, pal: ptr<function, array<vec3<f32>, 4>>) {\n  let p0 = from565(c0);\n  let p1 = from565(c1);\n  for (var j: u32 = 0u; j < 4u; j = j + 1u) {\n    (*pal)[j] = wa(j) * p0 + wb(j) * p1;\n  }\n}\n\n// Assign each of the 16 pixels its nearest palette entry (full 4-entry L2),\n// writing indices into `out_idx` and returning the total squared error.\nfn assign_indices(\n  pixels:  ptr<function, array<vec3<f32>, 16>>,\n  pal:     ptr<function, array<vec3<f32>, 4>>,\n  out_idx: ptr<function, array<u32, 16>>,\n) -> f32 {\n  var err: f32 = 0.0;\n  for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n    let c = (*pixels)[k];\n    var best_j: u32 = 0u;\n    var best_d: f32 = 1e30;\n    for (var j: u32 = 0u; j < 4u; j = j + 1u) {\n      let d = (*pal)[j] - c;\n      let d2 = dot(d, d);\n      if (d2 < best_d) {\n        best_d = d2;\n        best_j = j;\n      }\n    }\n    (*out_idx)[k] = best_j;\n    err = err + best_d;\n  }\n  return err;\n}\n\n// One least-squares refit pass: solve the 2x2 normal equations for the endpoint\n// colours that minimise \u03A3\u2016wa\xB7e0 + wb\xB7e1 \u2212 c\u2016\xB2 under the current indices. The\n// three channels share the scalar sums, so it's one 2x2 solve with vec3 RHS.\nstruct RefitResult { e0: vec3<f32>, e1: vec3<f32>, valid: bool };\nfn refit(\n  pixels:  ptr<function, array<vec3<f32>, 16>>,\n  indices: ptr<function, array<u32, 16>>,\n) -> RefitResult {\n  var sAA: f32 = 0.0; var sBB: f32 = 0.0; var sAB: f32 = 0.0;\n  var sAV: vec3<f32> = vec3<f32>(0.0);\n  var sBV: vec3<f32> = vec3<f32>(0.0);\n  for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n    let a = wa((*indices)[k]);\n    let b = wb((*indices)[k]);\n    let v = (*pixels)[k];\n    sAA = sAA + a * a;\n    sBB = sBB + b * b;\n    sAB = sAB + a * b;\n    sAV = sAV + a * v;\n    sBV = sBV + b * v;\n  }\n  var out: RefitResult;\n  let det = sAA * sBB - sAB * sAB;\n  if (abs(det) < 1e-9) {\n    out.valid = false;\n    return out;\n  }\n  out.e0 = clamp((sBB * sAV - sAB * sBV) / det, vec3<f32>(0.0), vec3<f32>(1.0));\n  out.e1 = clamp((sAA * sBV - sAB * sAV) / det, vec3<f32>(0.0), vec3<f32>(1.0));\n  out.valid = true;\n  return out;\n}\n\n// Candidate solution tracked across endpoint seeds / refit passes.\nstruct Best { c0: u32, c1: u32, indices: array<u32, 16>, err: f32 };\n\n// Quantize (hi, lo) to 565, force 4-color mode, assign indices, then refine with\n// up to `max_refits` least-squares passes. Commits to `*best` only on strict\n// improvement.\nfn fit_from_endpoints(\n  pixels: ptr<function, array<vec3<f32>, 16>>,\n  hi: vec3<f32>,\n  lo: vec3<f32>,\n  max_refits: u32,\n  best: ptr<function, Best>,\n) {\n  var c0 = to565(hi);\n  var c1 = to565(lo);\n  // 4-color mode requires color0 > color1.\n  if (c0 == c1) {\n    if (c1 > 0u) { c1 = c1 - 1u; } else { c0 = c0 + 1u; }\n  } else if (c0 < c1) {\n    let t = c0; c0 = c1; c1 = t;\n  }\n\n  var pal: array<vec3<f32>, 4>;\n  var idx: array<u32, 16>;\n  build_palette(c0, c1, &pal);\n  var err = assign_indices(pixels, &pal, &idx);\n  if (err < (*best).err) {\n    (*best).c0 = c0; (*best).c1 = c1; (*best).indices = idx; (*best).err = err;\n  }\n\n  for (var rp: u32 = 0u; rp < max_refits; rp = rp + 1u) {\n    let r = refit(pixels, &idx);\n    if (!r.valid) { break; }\n    var nc0 = to565(r.e0);\n    var nc1 = to565(r.e1);\n    // A refit that flips/equalises the endpoints would change decode mode;\n    // keep 4-color mode, and stop once it stops moving.\n    if (nc0 < nc1) { let t = nc0; nc0 = nc1; nc1 = t; }\n    if (nc0 == nc1) { break; }\n    if (nc0 == c0 && nc1 == c1) { break; }\n    build_palette(nc0, nc1, &pal);\n    let nerr = assign_indices(pixels, &pal, &idx);\n    c0 = nc0; c1 = nc1; err = nerr;\n    if (nerr < (*best).err) {\n      (*best).c0 = nc0; (*best).c1 = nc1; (*best).indices = idx; (*best).err = nerr;\n    }\n  }\n}\n\n// Principal colour axis via covariance power-iteration, seeded with the bbox\n// diagonal. Returns a unit axis, or vec3(0) for a degenerate (constant) block.\nfn principal_axis(\n  pixels: ptr<function, array<vec3<f32>, 16>>,\n  mean: vec3<f32>,\n  seed: vec3<f32>,\n) -> vec3<f32> {\n  // Symmetric 3x3 covariance, stored as its three rows.\n  var c0v = vec3<f32>(0.0);\n  var c1v = vec3<f32>(0.0);\n  var c2v = vec3<f32>(0.0);\n  for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n    let d = (*pixels)[k] - mean;\n    c0v = c0v + d.x * d;\n    c1v = c1v + d.y * d;\n    c2v = c2v + d.z * d;\n  }\n  var v = seed;\n  var len = length(v);\n  if (len < 1e-9) { return vec3<f32>(0.0); }\n  v = v / len;\n  for (var iter: u32 = 0u; iter < 8u; iter = iter + 1u) {\n    let nv = vec3<f32>(dot(c0v, v), dot(c1v, v), dot(c2v, v));\n    len = length(nv);\n    if (len < 1e-12) { return vec3<f32>(0.0); }\n    v = nv / len;\n  }\n  return v;\n}\n\n@compute @workgroup_size(8, 8, 1)\nfn encode(@builtin(global_invocation_id) gid: vec3<u32>) {\n  if (gid.x >= params.blocks_x || gid.y >= params.blocks_y) {\n    return;\n  }\n\n  let bx = gid.x;\n  let by = gid.y;\n  let block_index = by * params.blocks_x + bx;\n\n  let base = vec2<i32>(i32(bx) * 4, i32(by) * 4);\n  let max_xy = vec2<i32>(i32(params.width) - 1, i32(params.height) - 1);\n\n  var pixels: array<vec3<f32>, 16>;\n  var bb_min = vec3<f32>(1.0, 1.0, 1.0);\n  var bb_max = vec3<f32>(0.0, 0.0, 0.0);\n  var mean = vec3<f32>(0.0);\n\n  for (var i: u32 = 0u; i < 16u; i = i + 1u) {\n    let lx = i32(i & 3u);\n    let ly = i32(i >> 2u);\n    // Clamp to edge for non-multiple-of-4 textures.\n    let p  = clamp(base + vec2<i32>(lx, ly), vec2<i32>(0, 0), max_xy);\n    let c  = textureLoad(src_tex, p, 0).rgb;\n    pixels[i] = c;\n    bb_min = min(bb_min, c);\n    bb_max = max(bb_max, c);\n    mean = mean + c;\n  }\n  mean = mean * (1.0 / 16.0);\n\n  // Inset the bounding box by ~half an RGB565 cell (1/16) so the quantized\n  // 4-color palette covers the real data range more tightly (stb_dxt heuristic).\n  let inset = (bb_max - bb_min) / 16.0;\n  let bbox_hi = clamp(bb_max - inset, vec3<f32>(0.0), vec3<f32>(1.0));\n  let bbox_lo = clamp(bb_min + inset, vec3<f32>(0.0), vec3<f32>(1.0));\n\n  if (QUALITY_HIGH == 0u) {\n    // -------- fast: projection + fused LSQ refit + reprojection --------\n    var c0 = to565(bbox_hi);\n    var c1 = to565(bbox_lo);\n    if (c0 == c1) {\n      if (c1 > 0u) { c1 = c1 - 1u; } else { c0 = c0 + 1u; }\n    } else if (c0 < c1) {\n      let t = c0; c0 = c1; c1 = t;\n    }\n    let p0 = from565(c0);\n    let p1 = from565(c1);\n\n    // Fused pass: projection assignment + LSQ sums + the seed solution's\n    // packed indices and squared error. Level \u2192 BC1 index: 0\u21920 (c0), 1\u21922,\n    // 2\u21923, 3\u21921 (c1); as a packed LUT: (0x78 >> 2L) & 3.\n    var idx_bits: u32 = 0u;\n    let dir = p1 - p0;\n    let dd = dot(dir, dir);\n    if (dd > 0.0) {\n      let inv = 3.0 / dd;\n      var sAA: f32 = 0.0; var sBB: f32 = 0.0; var sAB: f32 = 0.0;\n      var sAV = vec3<f32>(0.0); var sBV = vec3<f32>(0.0);\n      var s_min = 3.0; var s_max = 0.0;\n      var seed_err: f32 = 0.0;\n      for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n        let v = pixels[k];\n        let s = clamp(floor(dot(v - p0, dir) * inv + 0.5), 0.0, 3.0);\n        s_min = min(s_min, s); s_max = max(s_max, s);\n        let b = s * (1.0 / 3.0); let a = 1.0 - b;\n        sAA = sAA + a * a; sBB = sBB + b * b; sAB = sAB + a * b;\n        sAV = sAV + a * v; sBV = sBV + b * v;\n        let e = v - (p0 + b * dir);\n        seed_err = seed_err + dot(e, e);\n        idx_bits = idx_bits | (((0x78u >> (u32(s) * 2u)) & 3u) << (k * 2u));\n      }\n      let det = sAA * sBB - sAB * sAB;\n      // Refit only on a well-conditioned system: when every pixel lands on\n      // ONE level (flat blocks \u2014 the 4-colour nudge forces c0 \u2260 c1 even\n      // then) the system is rank-1 and det/numerators are pure float noise;\n      // the solve would return garbage endpoints. With \u22652 levels\n      // det = \u03A3_i<j (b_j \u2212 b_i)\xB2 \u2265 ~1.67, so 1e-3 is a safe guard.\n      if (s_min < s_max && abs(det) > 1e-3) {\n        let e0 = clamp((sBB * sAV - sAB * sBV) / det, vec3<f32>(0.0), vec3<f32>(1.0));\n        let e1 = clamp((sAA * sBV - sAB * sAV) / det, vec3<f32>(0.0), vec3<f32>(1.0));\n        var nc0 = to565(e0);\n        var nc1 = to565(e1);\n        if (nc0 == nc1) {\n          if (nc1 > 0u) { nc1 = nc1 - 1u; } else { nc0 = nc0 + 1u; }\n        } else if (nc0 < nc1) {\n          let t = nc0; nc0 = nc1; nc1 = t;\n        }\n        let np0 = from565(nc0);\n        let np1 = from565(nc1);\n        let ndir = np1 - np0;\n        let ndd = dot(ndir, ndir);\n        if (ndd > 0.0 && !(nc0 == c0 && nc1 == c1)) {\n          // Reproject against the refit endpoints and accept them only if\n          // the block error actually decreases (the refit minimises a\n          // continuous objective; after 565 quantisation it can lose).\n          let ninv = 3.0 / ndd;\n          var refit_err: f32 = 0.0;\n          var nidx_bits: u32 = 0u;\n          for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n            let v = pixels[k];\n            let s = clamp(floor(dot(v - np0, ndir) * ninv + 0.5), 0.0, 3.0);\n            let e = v - (np0 + s * (1.0 / 3.0) * ndir);\n            refit_err = refit_err + dot(e, e);\n            nidx_bits = nidx_bits | (((0x78u >> (u32(s) * 2u)) & 3u) << (k * 2u));\n          }\n          if (refit_err < seed_err) {\n            c0 = nc0; c1 = nc1;\n            idx_bits = nidx_bits;\n          }\n        }\n      }\n    }\n\n    let out = block_index * 2u;\n    dst[out]      = c0 | (c1 << 16u);\n    dst[out + 1u] = idx_bits;\n    return;\n  }\n\n  // ------------------------------ high --------------------------------- //\n  var best: Best;\n  best.err = 1e30;\n\n  // Seed from the principal colour axis: project all texels onto it, take the\n  // extreme projections as endpoints, inset along the axis. Then also try the\n  // bbox seed and keep whichever family yields the lower error.\n  let axis = principal_axis(&pixels, mean, bb_max - bb_min);\n  if (dot(axis, axis) > 0.0) {\n    var t_min: f32 = 1e30;\n    var t_max: f32 = -1e30;\n    for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n      let t = dot(pixels[k] - mean, axis);\n      t_min = min(t_min, t);\n      t_max = max(t_max, t);\n    }\n    let pad = (t_max - t_min) / 16.0;\n    let pca_hi = clamp(mean + (t_max - pad) * axis, vec3<f32>(0.0), vec3<f32>(1.0));\n    let pca_lo = clamp(mean + (t_min + pad) * axis, vec3<f32>(0.0), vec3<f32>(1.0));\n    fit_from_endpoints(&pixels, pca_hi, pca_lo, 3u, &best);\n  }\n  fit_from_endpoints(&pixels, bbox_hi, bbox_lo, 3u, &best);\n\n  var indices: u32 = 0u;\n  for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n    indices = indices | ((best.indices[k] & 3u) << (k * 2u));\n  }\n\n  let out = block_index * 2u;\n  dst[out]      = best.c0 | (best.c1 << 16u);\n  dst[out + 1u] = indices;\n}\n";
+var bc1_default = "// BC1 (DXT1) compute shader encoder.\n//\n// Each invocation encodes one 4x4 pixel block into an 8-byte BC1 block\n// written as 2 x u32 into the destination storage buffer.\n//\n// BC1 block layout (little-endian):\n//   u32[0]: color0 (low 16) | color1 (high 16)   both in RGB565\n//   u32[1]: 16 x 2-bit indices, pixel 0 = bits 0..1, pixel 15 = bits 30..31\n//\n// We always force the 4-color mode (color0 > color1, numeric 16-bit):\n//   idx 0 -> color0\n//   idx 1 -> color1\n//   idx 2 -> (2*color0 +   color1) / 3\n//   idx 3 -> (  color0 + 2*color1) / 3\n//\n// QUALITY LEVELS (pipeline-overridable constant `QUALITY_HIGH`)\n//   fast (0, default): principal-axis endpoint seed (the high path's\n//     covariance power-iteration; inset bbox on degenerate blocks), inset by\n//     ~half a 565 cell along the axis, then ONE fused pass that projects\n//     every pixel onto the decoded-endpoint line (the 4 palette entries are\n//     colinear and evenly spaced, so the nearest entry is the rounded\n//     projection \u2014 no 4-entry search) while accumulating the least-squares\n//     refit sums; the refit endpoints are re-quantised and a final projection\n//     pass assigns the indices, packed on the fly.\n//   high (1): endpoints are seeded from the block's principal colour axis\n//     (covariance power-iteration) as well as the bbox diagonal, each refined by\n//     several least-squares passes with full 4-entry searches; the lower-error\n//     family wins. Mirrors bc1_ref.ts.\n\n// 0 = fast (default), 1 = high-quality. Set via pipeline constants.\noverride QUALITY_HIGH: u32 = 0u;\n\nstruct Params {\n  blocks_x: u32,\n  blocks_y: u32,\n  width:    u32,\n  height:   u32,\n};\n\n@group(0) @binding(0) var src_tex: texture_2d<f32>;\n@group(0) @binding(1) var<storage, read_write> dst: array<u32>;\n@group(0) @binding(2) var<uniform> params: Params;\n\nfn to565(c: vec3<f32>) -> u32 {\n  // Round-to-nearest quantization into 5-6-5.\n  let r = u32(clamp(floor(c.r * 31.0 + 0.5), 0.0, 31.0));\n  let g = u32(clamp(floor(c.g * 63.0 + 0.5), 0.0, 63.0));\n  let b = u32(clamp(floor(c.b * 31.0 + 0.5), 0.0, 31.0));\n  return (r << 11u) | (g << 5u) | b;\n}\n\nfn from565(c: u32) -> vec3<f32> {\n  let r = (c >> 11u) & 31u;\n  let g = (c >>  5u) & 63u;\n  let b =  c         & 31u;\n  // 5/6-bit -> 8-bit: (x*527+23)>>6 (6-bit: 259/33) \u2014 round-to-nearest\n  // scaling, matching bc1_ref.ts and typical hardware decoders (white ->\n  // 255). Integer u32 math is exact. NOTE: this is NOT plain bit-replication\n  // ((x<<3)|(x>>2)) \u2014 they differ for some codes (e.g. 5-bit 3 -> 25 vs 24).\n  // Selecting indices against this palette is what makes the encoder agree\n  // with what the GPU will actually sample.\n  let r8 = (r * 527u + 23u) >> 6u;\n  let g8 = (g * 259u + 33u) >> 6u;\n  let b8 = (b * 527u + 23u) >> 6u;\n  return vec3<f32>(vec3<u32>(r8, g8, b8)) / 255.0;\n}\n\n// 4-color-mode interpolation weights: palette[j] = wa(j)*c0 + wb(j)*c1.\nfn wa(j: u32) -> f32 {\n  switch j {\n    case 0u:  { return 1.0; }\n    case 1u:  { return 0.0; }\n    case 2u:  { return 2.0 / 3.0; }\n    default:  { return 1.0 / 3.0; }   // case 3u\n  }\n}\nfn wb(j: u32) -> f32 {\n  switch j {\n    case 0u:  { return 0.0; }\n    case 1u:  { return 1.0; }\n    case 2u:  { return 1.0 / 3.0; }\n    default:  { return 2.0 / 3.0; }   // case 3u\n  }\n}\n\nfn build_palette(c0: u32, c1: u32, pal: ptr<function, array<vec3<f32>, 4>>) {\n  let p0 = from565(c0);\n  let p1 = from565(c1);\n  for (var j: u32 = 0u; j < 4u; j = j + 1u) {\n    (*pal)[j] = wa(j) * p0 + wb(j) * p1;\n  }\n}\n\n// Assign each of the 16 pixels its nearest palette entry (full 4-entry L2),\n// writing indices into `out_idx` and returning the total squared error.\nfn assign_indices(\n  pixels:  ptr<function, array<vec3<f32>, 16>>,\n  pal:     ptr<function, array<vec3<f32>, 4>>,\n  out_idx: ptr<function, array<u32, 16>>,\n) -> f32 {\n  var err: f32 = 0.0;\n  for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n    let c = (*pixels)[k];\n    var best_j: u32 = 0u;\n    var best_d: f32 = 1e30;\n    for (var j: u32 = 0u; j < 4u; j = j + 1u) {\n      let d = (*pal)[j] - c;\n      let d2 = dot(d, d);\n      if (d2 < best_d) {\n        best_d = d2;\n        best_j = j;\n      }\n    }\n    (*out_idx)[k] = best_j;\n    err = err + best_d;\n  }\n  return err;\n}\n\n// One least-squares refit pass: solve the 2x2 normal equations for the endpoint\n// colours that minimise \u03A3\u2016wa\xB7e0 + wb\xB7e1 \u2212 c\u2016\xB2 under the current indices. The\n// three channels share the scalar sums, so it's one 2x2 solve with vec3 RHS.\nstruct RefitResult { e0: vec3<f32>, e1: vec3<f32>, valid: bool };\nfn refit(\n  pixels:  ptr<function, array<vec3<f32>, 16>>,\n  indices: ptr<function, array<u32, 16>>,\n) -> RefitResult {\n  var sAA: f32 = 0.0; var sBB: f32 = 0.0; var sAB: f32 = 0.0;\n  var sAV: vec3<f32> = vec3<f32>(0.0);\n  var sBV: vec3<f32> = vec3<f32>(0.0);\n  for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n    let a = wa((*indices)[k]);\n    let b = wb((*indices)[k]);\n    let v = (*pixels)[k];\n    sAA = sAA + a * a;\n    sBB = sBB + b * b;\n    sAB = sAB + a * b;\n    sAV = sAV + a * v;\n    sBV = sBV + b * v;\n  }\n  var out: RefitResult;\n  let det = sAA * sBB - sAB * sAB;\n  if (abs(det) < 1e-9) {\n    out.valid = false;\n    return out;\n  }\n  out.e0 = clamp((sBB * sAV - sAB * sBV) / det, vec3<f32>(0.0), vec3<f32>(1.0));\n  out.e1 = clamp((sAA * sBV - sAB * sAV) / det, vec3<f32>(0.0), vec3<f32>(1.0));\n  out.valid = true;\n  return out;\n}\n\n// Candidate solution tracked across endpoint seeds / refit passes.\nstruct Best { c0: u32, c1: u32, indices: array<u32, 16>, err: f32 };\n\n// Quantize (hi, lo) to 565, force 4-color mode, assign indices, then refine with\n// up to `max_refits` least-squares passes. Commits to `*best` only on strict\n// improvement.\nfn fit_from_endpoints(\n  pixels: ptr<function, array<vec3<f32>, 16>>,\n  hi: vec3<f32>,\n  lo: vec3<f32>,\n  max_refits: u32,\n  best: ptr<function, Best>,\n) {\n  var c0 = to565(hi);\n  var c1 = to565(lo);\n  // 4-color mode requires color0 > color1.\n  if (c0 == c1) {\n    if (c1 > 0u) { c1 = c1 - 1u; } else { c0 = c0 + 1u; }\n  } else if (c0 < c1) {\n    let t = c0; c0 = c1; c1 = t;\n  }\n\n  var pal: array<vec3<f32>, 4>;\n  var idx: array<u32, 16>;\n  build_palette(c0, c1, &pal);\n  var err = assign_indices(pixels, &pal, &idx);\n  if (err < (*best).err) {\n    (*best).c0 = c0; (*best).c1 = c1; (*best).indices = idx; (*best).err = err;\n  }\n\n  for (var rp: u32 = 0u; rp < max_refits; rp = rp + 1u) {\n    let r = refit(pixels, &idx);\n    if (!r.valid) { break; }\n    var nc0 = to565(r.e0);\n    var nc1 = to565(r.e1);\n    // A refit that flips/equalises the endpoints would change decode mode;\n    // keep 4-color mode, and stop once it stops moving.\n    if (nc0 < nc1) { let t = nc0; nc0 = nc1; nc1 = t; }\n    if (nc0 == nc1) { break; }\n    if (nc0 == c0 && nc1 == c1) { break; }\n    build_palette(nc0, nc1, &pal);\n    let nerr = assign_indices(pixels, &pal, &idx);\n    c0 = nc0; c1 = nc1; err = nerr;\n    if (nerr < (*best).err) {\n      (*best).c0 = nc0; (*best).c1 = nc1; (*best).indices = idx; (*best).err = nerr;\n    }\n  }\n}\n\n// Principal colour axis via covariance power-iteration, seeded with the bbox\n// diagonal. Returns a unit axis, or vec3(0) for a degenerate (constant) block.\nfn principal_axis(\n  pixels: ptr<function, array<vec3<f32>, 16>>,\n  mean: vec3<f32>,\n  seed: vec3<f32>,\n) -> vec3<f32> {\n  // Symmetric 3x3 covariance, stored as its three rows.\n  var c0v = vec3<f32>(0.0);\n  var c1v = vec3<f32>(0.0);\n  var c2v = vec3<f32>(0.0);\n  for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n    let d = (*pixels)[k] - mean;\n    c0v = c0v + d.x * d;\n    c1v = c1v + d.y * d;\n    c2v = c2v + d.z * d;\n  }\n  var v = seed;\n  var len = length(v);\n  if (len < 1e-9) { return vec3<f32>(0.0); }\n  v = v / len;\n  for (var iter: u32 = 0u; iter < 8u; iter = iter + 1u) {\n    let nv = vec3<f32>(dot(c0v, v), dot(c1v, v), dot(c2v, v));\n    len = length(nv);\n    if (len < 1e-12) { return vec3<f32>(0.0); }\n    v = nv / len;\n  }\n  return v;\n}\n\n@compute @workgroup_size(8, 8, 1)\nfn encode(@builtin(global_invocation_id) gid: vec3<u32>) {\n  if (gid.x >= params.blocks_x || gid.y >= params.blocks_y) {\n    return;\n  }\n\n  let bx = gid.x;\n  let by = gid.y;\n  let block_index = by * params.blocks_x + bx;\n\n  let base = vec2<i32>(i32(bx) * 4, i32(by) * 4);\n  let max_xy = vec2<i32>(i32(params.width) - 1, i32(params.height) - 1);\n\n  var pixels: array<vec3<f32>, 16>;\n  var bb_min = vec3<f32>(1.0, 1.0, 1.0);\n  var bb_max = vec3<f32>(0.0, 0.0, 0.0);\n  var mean = vec3<f32>(0.0);\n\n  for (var i: u32 = 0u; i < 16u; i = i + 1u) {\n    let lx = i32(i & 3u);\n    let ly = i32(i >> 2u);\n    // Clamp to edge for non-multiple-of-4 textures.\n    let p  = clamp(base + vec2<i32>(lx, ly), vec2<i32>(0, 0), max_xy);\n    let c  = textureLoad(src_tex, p, 0).rgb;\n    pixels[i] = c;\n    bb_min = min(bb_min, c);\n    bb_max = max(bb_max, c);\n    mean = mean + c;\n  }\n  mean = mean * (1.0 / 16.0);\n\n  // Inset the bounding box by ~half an RGB565 cell (1/16) so the quantized\n  // 4-color palette covers the real data range more tightly (stb_dxt heuristic).\n  let inset = (bb_max - bb_min) / 16.0;\n  let bbox_hi = clamp(bb_max - inset, vec3<f32>(0.0), vec3<f32>(1.0));\n  let bbox_lo = clamp(bb_min + inset, vec3<f32>(0.0), vec3<f32>(1.0));\n\n  if (QUALITY_HIGH == 0u) {\n    // -------- fast: PCA seed + projection + fused LSQ refit + reprojection\n    // Seed endpoints from the block's principal colour axis (same\n    // power-iteration as the high path). The bbox diagonal is sign-blind: on\n    // anti-correlated channels (normal maps, hue edges) it points across the\n    // data instead of along it, and the LSQ refit \u2014 which fits endpoints\n    // GIVEN the projection indices \u2014 can't recover from a wrong axis.\n    // Degenerate (near-flat) blocks keep the inset-bbox seed.\n    var seed_hi = bbox_hi;\n    var seed_lo = bbox_lo;\n    let axis = principal_axis(&pixels, mean, bb_max - bb_min);\n    if (dot(axis, axis) > 0.0) {\n      var t_min: f32 = 1e30;\n      var t_max: f32 = -1e30;\n      for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n        let t = dot(pixels[k] - mean, axis);\n        t_min = min(t_min, t);\n        t_max = max(t_max, t);\n      }\n      // Inset along the axis by ~half a 565 cell (stb_dxt heuristic,\n      // matching the bbox inset).\n      let pad = (t_max - t_min) / 16.0;\n      seed_hi = clamp(mean + (t_max - pad) * axis, vec3<f32>(0.0), vec3<f32>(1.0));\n      seed_lo = clamp(mean + (t_min + pad) * axis, vec3<f32>(0.0), vec3<f32>(1.0));\n    }\n    var c0 = to565(seed_hi);\n    var c1 = to565(seed_lo);\n    if (c0 == c1) {\n      if (c1 > 0u) { c1 = c1 - 1u; } else { c0 = c0 + 1u; }\n    } else if (c0 < c1) {\n      let t = c0; c0 = c1; c1 = t;\n    }\n    let p0 = from565(c0);\n    let p1 = from565(c1);\n\n    // Fused pass: projection assignment + LSQ sums + the seed solution's\n    // packed indices and squared error. Level \u2192 BC1 index: 0\u21920 (c0), 1\u21922,\n    // 2\u21923, 3\u21921 (c1); as a packed LUT: (0x78 >> 2L) & 3.\n    var idx_bits: u32 = 0u;\n    let dir = p1 - p0;\n    let dd = dot(dir, dir);\n    if (dd > 0.0) {\n      let inv = 3.0 / dd;\n      var sAA: f32 = 0.0; var sBB: f32 = 0.0; var sAB: f32 = 0.0;\n      var sAV = vec3<f32>(0.0); var sBV = vec3<f32>(0.0);\n      var s_min = 3.0; var s_max = 0.0;\n      var seed_err: f32 = 0.0;\n      for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n        let v = pixels[k];\n        let s = clamp(floor(dot(v - p0, dir) * inv + 0.5), 0.0, 3.0);\n        s_min = min(s_min, s); s_max = max(s_max, s);\n        let b = s * (1.0 / 3.0); let a = 1.0 - b;\n        sAA = sAA + a * a; sBB = sBB + b * b; sAB = sAB + a * b;\n        sAV = sAV + a * v; sBV = sBV + b * v;\n        let e = v - (p0 + b * dir);\n        seed_err = seed_err + dot(e, e);\n        idx_bits = idx_bits | (((0x78u >> (u32(s) * 2u)) & 3u) << (k * 2u));\n      }\n      let det = sAA * sBB - sAB * sAB;\n      // Refit only on a well-conditioned system: when every pixel lands on\n      // ONE level (flat blocks \u2014 the 4-colour nudge forces c0 \u2260 c1 even\n      // then) the system is rank-1 and det/numerators are pure float noise;\n      // the solve would return garbage endpoints. With \u22652 levels\n      // det = \u03A3_i<j (b_j \u2212 b_i)\xB2 \u2265 ~1.67, so 1e-3 is a safe guard.\n      if (s_min < s_max && abs(det) > 1e-3) {\n        // Clamp the refit to the block bbox (not [0,1]): on multi-cluster\n        // blocks the unconstrained solve extrapolates far outside the block's\n        // colours and the per-channel clamp then bends the hue \u2014 fringe pixels\n        // decode to colours that exist nowhere in the block. Constraining to\n        // the bbox also measures better in plain SSE (+1.6 dB on the colour\n        // test card), so the accept-if-better guard below keeps more refits.\n        let e0 = clamp((sBB * sAV - sAB * sBV) / det, bb_min, bb_max);\n        let e1 = clamp((sAA * sBV - sAB * sAV) / det, bb_min, bb_max);\n        var nc0 = to565(e0);\n        var nc1 = to565(e1);\n        if (nc0 == nc1) {\n          if (nc1 > 0u) { nc1 = nc1 - 1u; } else { nc0 = nc0 + 1u; }\n        } else if (nc0 < nc1) {\n          let t = nc0; nc0 = nc1; nc1 = t;\n        }\n        let np0 = from565(nc0);\n        let np1 = from565(nc1);\n        let ndir = np1 - np0;\n        let ndd = dot(ndir, ndir);\n        if (ndd > 0.0 && !(nc0 == c0 && nc1 == c1)) {\n          // Reproject against the refit endpoints and accept them only if\n          // the block error actually decreases (the refit minimises a\n          // continuous objective; after 565 quantisation it can lose).\n          let ninv = 3.0 / ndd;\n          var refit_err: f32 = 0.0;\n          var nidx_bits: u32 = 0u;\n          for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n            let v = pixels[k];\n            let s = clamp(floor(dot(v - np0, ndir) * ninv + 0.5), 0.0, 3.0);\n            let e = v - (np0 + s * (1.0 / 3.0) * ndir);\n            refit_err = refit_err + dot(e, e);\n            nidx_bits = nidx_bits | (((0x78u >> (u32(s) * 2u)) & 3u) << (k * 2u));\n          }\n          if (refit_err < seed_err) {\n            c0 = nc0; c1 = nc1;\n            idx_bits = nidx_bits;\n          }\n        }\n      }\n    }\n\n    let out = block_index * 2u;\n    dst[out]      = c0 | (c1 << 16u);\n    dst[out + 1u] = idx_bits;\n    return;\n  }\n\n  // ------------------------------ high --------------------------------- //\n  var best: Best;\n  best.err = 1e30;\n\n  // Seed from the principal colour axis: project all texels onto it, take the\n  // extreme projections as endpoints, inset along the axis. Then also try the\n  // bbox seed and keep whichever family yields the lower error.\n  let axis = principal_axis(&pixels, mean, bb_max - bb_min);\n  if (dot(axis, axis) > 0.0) {\n    var t_min: f32 = 1e30;\n    var t_max: f32 = -1e30;\n    for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n      let t = dot(pixels[k] - mean, axis);\n      t_min = min(t_min, t);\n      t_max = max(t_max, t);\n    }\n    let pad = (t_max - t_min) / 16.0;\n    let pca_hi = clamp(mean + (t_max - pad) * axis, vec3<f32>(0.0), vec3<f32>(1.0));\n    let pca_lo = clamp(mean + (t_min + pad) * axis, vec3<f32>(0.0), vec3<f32>(1.0));\n    fit_from_endpoints(&pixels, pca_hi, pca_lo, 3u, &best);\n  }\n  fit_from_endpoints(&pixels, bbox_hi, bbox_lo, 3u, &best);\n\n  var indices: u32 = 0u;\n  for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n    indices = indices | ((best.indices[k] & 3u) << (k * 2u));\n  }\n\n  let out = block_index * 2u;\n  dst[out]      = best.c0 | (best.c1 << 16u);\n  dst[out + 1u] = indices;\n}\n";
 
 // src/bc1_fast_f16.wgsl
 var bc1_fast_f16_default = `// bc1 "fast" encoder \u2014 f16 variant (requires the shader-f16 feature).
@@ -330,7 +434,8 @@ var bc1_fast_f16_default = `// bc1 "fast" encoder \u2014 f16 variant (requires t
 // f16 ([0,1] domain). The algorithm is the same family as the BC7/ASTC fast
 // paths rather than a port of bc1.wgsl's fast branch:
 //
-//   1. bbox endpoints, inset by ~half a 565 cell (stb_dxt heuristic)
+//   1. principal-axis endpoint seed (covariance power-iteration; inset bbox
+//      on degenerate blocks), inset by ~half a 565 cell (stb_dxt heuristic)
 //   2. quantise to 565, force 4-colour mode (c0 > c1)
 //   3. ONE fused pass: project every pixel onto the decoded-endpoint line
 //      (the 4 palette entries are colinear and evenly spaced, so the nearest
@@ -397,15 +502,66 @@ fn encode(@builtin(global_invocation_id) gid: vec3<u32>) {
   var pix: array<h3, 16>;
   var mn = h3(1.0);
   var mxv = h3(0.0);
+  var mean = h3(0.0);
   for (var i: u32 = 0u; i < 16u; i = i + 1u) {
     let p = clamp(base + vec2<i32>(i32(i & 3u), i32(i >> 2u)), vec2<i32>(0), mx);
     let px = h3(textureLoad(src_tex, p, 0).rgb);
     pix[i] = px; mn = min(mn, px); mxv = max(mxv, px);
+    mean = mean + px;
   }
+  mean = mean * h(1.0 / 16.0);
 
-  // Inset bbox by ~half a 565 cell so the quantised palette hugs the data.
-  let inset = (mxv - mn) * h(1.0 / 16.0);
-  let seed = order565(to565(clamp(mxv - inset, h3(0.0), h3(1.0))), to565(clamp(mn + inset, h3(0.0), h3(1.0))));
+  // Seed endpoints from the block's principal colour axis (covariance
+  // power-iteration, seeded with the bbox diagonal \u2014 same family as the
+  // 'high' path). The bbox diagonal is sign-blind: on anti-correlated
+  // channels (normal maps, hue edges) it points across the data instead of
+  // along it, the projection indices come out garbage, and the LSQ refit \u2014
+  // which fits endpoints GIVEN those indices \u2014 can't recover. Deviations are
+  // pre-scaled \xD716 so covariance entries for shallow blocks stay in f16's
+  // normal range (span ~1/255 \u2192 d\xB2 \u2248 1e-3) while full-range sums stay \u22644096;
+  // the iteration renormalises by the max component (a plain length() of the
+  // matvec output could overflow f16), so only the direction survives \u2014 the
+  // \xD7256 covariance scale is irrelevant.
+  var seed_hi: h3;
+  var seed_lo: h3;
+  var c0v = h3(0.0);
+  var c1v = h3(0.0);
+  var c2v = h3(0.0);
+  for (var k: u32 = 0u; k < 16u; k = k + 1u) {
+    let d = (pix[k] - mean) * h(16.0);
+    c0v = c0v + d.x * d;
+    c1v = c1v + d.y * d;
+    c2v = c2v + d.z * d;
+  }
+  var axis = mxv - mn;
+  var axis_ok = true;
+  for (var it: u32 = 0u; it < 4u; it = it + 1u) {
+    let nv = h3(dot(c0v, axis), dot(c1v, axis), dot(c2v, axis));
+    let m = max(max(abs(nv.x), abs(nv.y)), abs(nv.z));
+    if (m < h(1e-4)) { axis_ok = false; break; }
+    axis = nv / m;
+  }
+  if (axis_ok) {
+    axis = axis / length(axis);
+    var t_min = h(4.0);
+    var t_max = h(-4.0);
+    for (var k: u32 = 0u; k < 16u; k = k + 1u) {
+      let t = dot(pix[k] - mean, axis);
+      t_min = min(t_min, t);
+      t_max = max(t_max, t);
+    }
+    // Inset along the axis by ~half a 565 cell (stb_dxt heuristic, matching
+    // the degenerate-case bbox inset below).
+    let pad = (t_max - t_min) * h(1.0 / 16.0);
+    seed_hi = clamp(mean + (t_max - pad) * axis, h3(0.0), h3(1.0));
+    seed_lo = clamp(mean + (t_min + pad) * axis, h3(0.0), h3(1.0));
+  } else {
+    // Degenerate (near-flat) block: inset bbox seed, as before.
+    let inset = (mxv - mn) * h(1.0 / 16.0);
+    seed_hi = clamp(mxv - inset, h3(0.0), h3(1.0));
+    seed_lo = clamp(mn + inset, h3(0.0), h3(1.0));
+  }
+  let seed = order565(to565(seed_hi), to565(seed_lo));
   var c0 = seed.x;
   var c1 = seed.y;
   let p0 = from565(c0);
@@ -445,8 +601,14 @@ fn encode(@builtin(global_invocation_id) gid: vec3<u32>) {
     // det = \u03A3_i<j (b_j \u2212 b_i)\xB2 \u2265 15\xB7(1/3)\xB2 \u2248 1.67, far above the ~0.05 f16
     // noise floor \u2014 0.5 separates the two regimes cleanly.
     if (s_min < s_max && abs(det) > h(0.5)) {
-      let e0 = clamp((sBB * sAV - sAB * sBV) / det, h3(0.0), h3(1.0));
-      let e1 = clamp((sAA * sBV - sAB * sAV) / det, h3(0.0), h3(1.0));
+      // Clamp the refit to the block bbox (not [0,1]): on multi-cluster
+      // blocks the unconstrained solve extrapolates far outside the block's
+      // colours and the per-channel clamp then bends the hue \u2014 fringe pixels
+      // decode to colours that exist nowhere in the block. Constraining to
+      // the bbox also measures better in plain SSE (+1.6 dB on the colour
+      // test card), so the accept-if-better guard below keeps more refits.
+      let e0 = clamp((sBB * sAV - sAB * sBV) / det, mn, mxv);
+      let e1 = clamp((sAA * sBV - sAB * sAV) / det, mn, mxv);
       let refit = order565(to565(e0), to565(e1));
       let np0 = from565(refit.x);
       let np1 = from565(refit.y);
@@ -508,12 +670,12 @@ var BC1Encoder = class extends Encoder {
 };
 
 // src/bc5.wgsl
-var bc5_default = "// BC5 (RGTC2) compute shader encoder.\n//\n// Each invocation encodes one 4x4 pixel block into a 16-byte BC5 block\n// written as 4 x u32 into the destination storage buffer.\n//\n// BC5 = two BC4 blocks concatenated:\n//   block bytes  0..7  : BC4 of R channel (normal.x for tangent-space normals)\n//   block bytes  8..15 : BC4 of G channel (normal.y)\n//\n// Each BC4 half-block (8 bytes):\n//   byte 0     : red0 (8-bit endpoint)\n//   byte 1     : red1 (8-bit endpoint)\n//   bytes 2..7 : 16 \xD7 3-bit indices, LSB-first, pixel 0 at bit 0\n//\n// We always produce the 6-interpolation mode (red0 > red1). See\n// `bc4_ref.js` for the reasoning and the CPU reference this shader is\n// ported from \u2014 the algorithm and edge cases mirror it line-for-line.\n//\n// Pipeline per channel:\n//   1. Load 16 single-channel values, find min/max \u2192 initial endpoints.\n//   2. Quantize to 8-bit. Nudge apart if equal (forces 6-interp mode).\n//   3. Build palette, assign each texel its nearest entry (full L2).\n//   4. One-pass least-squares refinement: solve the 2\xD72 normal equations\n//      for the (r0, r1) that minimizes \u03A3(palette[i_k] \u2212 v_k)\xB2. Accept\n//      only if quantized endpoints still satisfy r0 > r1 AND total\n//      squared error decreased.\n//   5. Pack 2 endpoint bytes + 48 bits of indices into the 8-byte block.\n//\n// The candidate endpoints/indices/error are tracked in place \u2014 the refit\n// overwrites them only when accepted \u2014 so no 16-entry index array is ever\n// copied across a function return.\n//\n// QUALITY LEVELS (pipeline-overridable constant `QUALITY_HIGH`)\n//   fast (0, default): bbox endpoints + O(1) projection assignment per texel.\n//     The 8-entry palette in 6-interp mode is colinear and EVENLY spaced from\n//     r0 to r1 (levels 0..7 in palette order 0,2,3,4,5,6,7,1), so the nearest\n//     entry is the rounded projection onto the r0\u2192r1 axis \u2014 no 8-entry\n//     search, and the 3-bit indices are packed on the fly. The LSQ refit is\n//     skipped (buys only ~0.36 dB).\n//   high (1): full nearest search + refit, matches bc4_ref/bc5_ref.\n\n// 0 = fast (default), 1 = exhaustive/high-quality. Set via pipeline constants.\noverride QUALITY_HIGH: u32 = 0u;\n\nstruct Params {\n  blocks_x: u32,\n  blocks_y: u32,\n  width:    u32,\n  height:   u32,\n};\n\n@group(0) @binding(0) var src_tex: texture_2d<f32>;\n@group(0) @binding(1) var<storage, read_write> dst: array<u32>;\n@group(0) @binding(2) var<uniform> params: Params;\n\n// 6-interpolation-mode palette weights. palette[j] = W0_6[j]*r0 + W1_6[j]*r1.\n// Expressed as a switch so we don't rely on module-scope const arrays.\nfn w0_6(j: u32) -> f32 {\n  switch j {\n    case 0u:  { return 1.0; }\n    case 1u:  { return 0.0; }\n    case 2u:  { return 6.0 / 7.0; }\n    case 3u:  { return 5.0 / 7.0; }\n    case 4u:  { return 4.0 / 7.0; }\n    case 5u:  { return 3.0 / 7.0; }\n    case 6u:  { return 2.0 / 7.0; }\n    default:  { return 1.0 / 7.0; }   // case 7u\n  }\n}\n\nfn w1_6(j: u32) -> f32 {\n  switch j {\n    case 0u:  { return 0.0; }\n    case 1u:  { return 1.0; }\n    case 2u:  { return 1.0 / 7.0; }\n    case 3u:  { return 2.0 / 7.0; }\n    case 4u:  { return 3.0 / 7.0; }\n    case 5u:  { return 4.0 / 7.0; }\n    case 6u:  { return 5.0 / 7.0; }\n    default:  { return 6.0 / 7.0; }   // case 7u\n  }\n}\n\nfn quantize8(v: f32) -> u32 {\n  // Round-to-nearest, clamp to [0, 255]. floor(x + 0.5) is the same\n  // rounding rule the CPU reference uses.\n  return u32(clamp(floor(v * 255.0 + 0.5), 0.0, 255.0));\n}\n\n// Build the 8-entry palette for endpoints (r0f, r1f) in normalised space.\nfn build_pal(r0f: f32, r1f: f32, pal: ptr<function, array<f32, 8>>) {\n  for (var j: u32 = 0u; j < 8u; j = j + 1u) {\n    (*pal)[j] = w0_6(j) * r0f + w1_6(j) * r1f;\n  }\n}\n\n// Assign each of the 16 values its nearest palette entry (full 8-entry L2),\n// writing indices into `out_idx` and returning the total squared error.\nfn assign_all(\n  values:  ptr<function, array<f32, 16>>,\n  pal:     ptr<function, array<f32, 8>>,\n  out_idx: ptr<function, array<u32, 16>>,\n) -> f32 {\n  var err: f32 = 0.0;\n  for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n    let v = (*values)[k];\n    var best_j: u32 = 0u;\n    var best_d: f32 = 1e20;\n    for (var j: u32 = 0u; j < 8u; j = j + 1u) {\n      let d  = (*pal)[j] - v;\n      let d2 = d * d;\n      if (d2 < best_d) {\n        best_d = d2;\n        best_j = j;\n      }\n    }\n    (*out_idx)[k] = best_j;\n    err = err + best_d;\n  }\n  return err;\n}\n\n// Encode 16 single-channel values into an 8-byte BC4 block, packed as\n// two little-endian u32s (u32[0] = bytes 0..3, u32[1] = bytes 4..7).\nfn encode_bc4(values: ptr<function, array<f32, 16>>) -> vec2<u32> {\n  // ---------------- 1. Initial endpoints: bbox of input ----------------\n  var vmin: f32 = 1.0;\n  var vmax: f32 = 0.0;\n  for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n    vmin = min(vmin, (*values)[k]);\n    vmax = max(vmax, (*values)[k]);\n  }\n  var r0: u32 = quantize8(vmax);\n  var r1: u32 = quantize8(vmin);\n  // Force 6-interp mode: red0 > red1 strictly.\n  if (r0 == r1) {\n    if (r1 > 0u) { r1 = r1 - 1u; }\n    else         { r0 = r0 + 1u; }\n  }\n\n  if (QUALITY_HIGH == 0u) {\n    // -------- fast: projection assignment, indices packed on the fly ----\n    // level = round(7\xB7(v \u2212 r0)/(r1 \u2212 r0)); level \u2192 BC4 index LUT (0,2,3,4,\n    // 5,6,7,1) packed as 3-bit entries in 0x3F58D0. Pixel k's 3 bits start\n    // at bit 3k+16 of the (w0,w1) pair (bytes 0..1 are the endpoints).\n    let r0f = f32(r0) / 255.0;\n    let scale = 7.0 / (f32(r1) / 255.0 - r0f);\n    var w0 = r0 | (r1 << 8u);\n    var w1 = 0u;\n    for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n      let L = u32(clamp(floor(((*values)[k] - r0f) * scale + 0.5), 0.0, 7.0));\n      let idx = (0x3F58D0u >> (L * 3u)) & 7u;\n      let bit = 3u * k + 16u;\n      if (bit <= 29u) {\n        w0 = w0 | (idx << bit);\n      } else if (bit >= 32u) {\n        w1 = w1 | (idx << (bit - 32u));\n      } else {\n        // k = 5 straddles the word boundary (bits 31..33).\n        w0 = w0 | (idx << bit);\n        w1 = w1 | (idx >> (32u - bit));\n      }\n    }\n    return vec2<u32>(w0, w1);\n  }\n\n  // ---------------- 2. Initial palette + indices + error --------------\n  var pal: array<f32, 8>;\n  build_pal(f32(r0) / 255.0, f32(r1) / 255.0, &pal);\n  var indices: array<u32, 16>;\n  var err = assign_all(values, &pal, &indices);\n\n  // ---------------- 3. Refinement: least-squares on (r0, r1) ----------\n  // High-quality only \u2014 the refit is the bulk of the per-channel cost and the\n  // branch is resolved at pipeline-compile time, so the fast path skips all of\n  // it (the sums loop included), not just the acceptance test.\n  if (QUALITY_HIGH != 0u) {\n    // Normal equations for palette[j] = a_j * r0 + b_j * r1:\n    //   [\u03A3AA  \u03A3AB] [r0]   [\u03A3AV]\n    //   [\u03A3AB  \u03A3BB] [r1] = [\u03A3BV]\n    var sAA: f32 = 0.0; var sBB: f32 = 0.0; var sAB: f32 = 0.0;\n    var sAV: f32 = 0.0; var sBV: f32 = 0.0;\n    for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n      let a = w0_6(indices[k]);\n      let b = w1_6(indices[k]);\n      let v = (*values)[k];\n      sAA = sAA + a * a;\n      sBB = sBB + b * b;\n      sAB = sAB + a * b;\n      sAV = sAV + a * v;\n      sBV = sBV + b * v;\n    }\n    let det = sAA * sBB - sAB * sAB;\n    // Degenerate system \u2192 skip refinement.\n    if (abs(det) > 1e-9) {\n      let new_r0 = clamp((sBB * sAV - sAB * sBV) / det, 0.0, 1.0);\n      let new_r1 = clamp((sAA * sBV - sAB * sAV) / det, 0.0, 1.0);\n      let qR0 = quantize8(new_r0);\n      let qR1 = quantize8(new_r1);\n      // Only accept refinements that stay in 6-interp mode. A refinement\n      // that flips or equalizes the endpoints would change decode mode.\n      if (qR0 > qR1) {\n        build_pal(f32(qR0) / 255.0, f32(qR1) / 255.0, &pal);\n        var idx2: array<u32, 16>;\n        let err2 = assign_all(values, &pal, &idx2);\n        if (err2 < err) {\n          r0 = qR0;\n          r1 = qR1;\n          indices = idx2;\n          err = err2;\n        }\n      }\n    }\n  }\n\n  // ---------------- 4. Pack 48-bit index field + 2 endpoint bytes -----\n  // The 48-bit index field spans block bytes 2..7. Split into idx_lo\n  // (low 32 bits of the field) and idx_hi (high 16 bits). An index at\n  // bit position 3k straddles the 32-bit boundary iff 3k < 32 < 3k+3\n  // (only k = 10, 11 straddle: bits 30..32 and 33..35; actually k=10\n  // is bits 30..32, k=11 is 33..35 \u2014 so k=10 straddles). We handle\n  // straddles by writing to both halves.\n  var idx_lo: u32 = 0u;\n  var idx_hi: u32 = 0u;\n  for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n    let bit = 3u * k;\n    let v = indices[k] & 7u;\n    if (bit + 3u <= 32u) {\n      idx_lo = idx_lo | (v << bit);\n    } else if (bit >= 32u) {\n      idx_hi = idx_hi | (v << (bit - 32u));\n    } else {\n      // Straddle: low part into idx_lo's top, high part into idx_hi's bottom.\n      idx_lo = idx_lo | (v << bit);\n      idx_hi = idx_hi | (v >> (32u - bit));\n    }\n  }\n\n  // Final u32s, both little-endian:\n  //   u32[0] bytes = red0, red1, idx_lo[7:0], idx_lo[15:8]\n  //   u32[1] bytes = idx_lo[23:16], idx_lo[31:24], idx_hi[7:0], idx_hi[15:8]\n  let out_lo = r0 | (r1 << 8u) | ((idx_lo & 0xFFFFu) << 16u);\n  let out_hi = (idx_lo >> 16u) | (idx_hi << 16u);\n\n  return vec2<u32>(out_lo, out_hi);\n}\n\n@compute @workgroup_size(8, 8, 1)\nfn encode(@builtin(global_invocation_id) gid: vec3<u32>) {\n  if (gid.x >= params.blocks_x || gid.y >= params.blocks_y) {\n    return;\n  }\n\n  let bx = gid.x;\n  let by = gid.y;\n  let block_index = by * params.blocks_x + bx;\n\n  let base   = vec2<i32>(i32(bx) * 4, i32(by) * 4);\n  let max_xy = vec2<i32>(i32(params.width) - 1, i32(params.height) - 1);\n\n  // Load 4\xD74 RG values, splitting into per-channel arrays so each can\n  // be handed to encode_bc4 independently.\n  var r_values: array<f32, 16>;\n  var g_values: array<f32, 16>;\n  for (var i: u32 = 0u; i < 16u; i = i + 1u) {\n    let lx = i32(i & 3u);\n    let ly = i32(i >> 2u);\n    // Clamp to edge for non-multiple-of-4 input sizes.\n    let p  = clamp(base + vec2<i32>(lx, ly), vec2<i32>(0, 0), max_xy);\n    let c  = textureLoad(src_tex, p, 0);\n    r_values[i] = c.r;\n    g_values[i] = c.g;\n  }\n\n  let r_block = encode_bc4(&r_values);\n  let g_block = encode_bc4(&g_values);\n\n  // BC5 block = R half (bytes 0..7) || G half (bytes 8..15) = 4 u32s.\n  let out = block_index * 4u;\n  dst[out + 0u] = r_block.x;\n  dst[out + 1u] = r_block.y;\n  dst[out + 2u] = g_block.x;\n  dst[out + 3u] = g_block.y;\n}\n";
+var bc5_default = "// BC5 (RGTC2) compute shader encoder.\n//\n// Each invocation encodes one 4x4 pixel block into a 16-byte BC5 block\n// written as 4 x u32 into the destination storage buffer.\n//\n// BC5 = two BC4 blocks concatenated:\n//   block bytes  0..7  : BC4 of R channel (normal.x for tangent-space normals)\n//   block bytes  8..15 : BC4 of G channel (normal.y)\n//\n// Each BC4 half-block (8 bytes):\n//   byte 0     : red0 (8-bit endpoint)\n//   byte 1     : red1 (8-bit endpoint)\n//   bytes 2..7 : 16 \xD7 3-bit indices, LSB-first, pixel 0 at bit 0\n//\n// We always produce the 6-interpolation mode (red0 > red1). See\n// `bc4_ref.js` for the reasoning and the CPU reference this shader is\n// ported from \u2014 the algorithm and edge cases mirror it line-for-line.\n//\n// Pipeline per channel:\n//   1. Load 16 single-channel values, find min/max \u2192 initial endpoints.\n//   2. Quantize to 8-bit. Nudge apart if equal (forces 6-interp mode).\n//   3. Build palette, assign each texel its nearest entry (full L2).\n//   4. One-pass least-squares refinement: solve the 2\xD72 normal equations\n//      for the (r0, r1) that minimizes \u03A3(palette[i_k] \u2212 v_k)\xB2. Accept\n//      only if quantized endpoints still satisfy r0 > r1 AND total\n//      squared error decreased.\n//   5. Pack 2 endpoint bytes + 48 bits of indices into the 8-byte block.\n//\n// The candidate endpoints/indices/error are tracked in place \u2014 the refit\n// overwrites them only when accepted \u2014 so no 16-entry index array is ever\n// copied across a function return.\n//\n// QUALITY LEVELS (pipeline-overridable constant `QUALITY_HIGH`)\n//   fast (0, default): bbox endpoints + O(1) projection assignment per texel.\n//     The 8-entry palette in 6-interp mode is colinear and EVENLY spaced from\n//     r0 to r1 (levels 0..7 in palette order 0,2,3,4,5,6,7,1), so the nearest\n//     entry is the rounded projection onto the r0\u2192r1 axis \u2014 no 8-entry\n//     search, and the 3-bit indices are packed on the fly. The LSQ refit sums\n//     are accumulated in the same fused pass; the requantised refit is\n//     accepted only if it lowers the block error (worth ~1.3 dB on the\n//     normal-map card).\n//   high (1): full nearest search + refit, matches bc4_ref/bc5_ref.\n\n// 0 = fast (default), 1 = exhaustive/high-quality. Set via pipeline constants.\noverride QUALITY_HIGH: u32 = 0u;\n\nstruct Params {\n  blocks_x: u32,\n  blocks_y: u32,\n  width:    u32,\n  height:   u32,\n};\n\n@group(0) @binding(0) var src_tex: texture_2d<f32>;\n@group(0) @binding(1) var<storage, read_write> dst: array<u32>;\n@group(0) @binding(2) var<uniform> params: Params;\n\n// 6-interpolation-mode palette weights. palette[j] = W0_6[j]*r0 + W1_6[j]*r1.\n// Expressed as a switch so we don't rely on module-scope const arrays.\nfn w0_6(j: u32) -> f32 {\n  switch j {\n    case 0u:  { return 1.0; }\n    case 1u:  { return 0.0; }\n    case 2u:  { return 6.0 / 7.0; }\n    case 3u:  { return 5.0 / 7.0; }\n    case 4u:  { return 4.0 / 7.0; }\n    case 5u:  { return 3.0 / 7.0; }\n    case 6u:  { return 2.0 / 7.0; }\n    default:  { return 1.0 / 7.0; }   // case 7u\n  }\n}\n\nfn w1_6(j: u32) -> f32 {\n  switch j {\n    case 0u:  { return 0.0; }\n    case 1u:  { return 1.0; }\n    case 2u:  { return 1.0 / 7.0; }\n    case 3u:  { return 2.0 / 7.0; }\n    case 4u:  { return 3.0 / 7.0; }\n    case 5u:  { return 4.0 / 7.0; }\n    case 6u:  { return 5.0 / 7.0; }\n    default:  { return 6.0 / 7.0; }   // case 7u\n  }\n}\n\nfn quantize8(v: f32) -> u32 {\n  // Round-to-nearest, clamp to [0, 255]. floor(x + 0.5) is the same\n  // rounding rule the CPU reference uses.\n  return u32(clamp(floor(v * 255.0 + 0.5), 0.0, 255.0));\n}\n\n// Build the 8-entry palette for endpoints (r0f, r1f) in normalised space.\nfn build_pal(r0f: f32, r1f: f32, pal: ptr<function, array<f32, 8>>) {\n  for (var j: u32 = 0u; j < 8u; j = j + 1u) {\n    (*pal)[j] = w0_6(j) * r0f + w1_6(j) * r1f;\n  }\n}\n\n// Assign each of the 16 values its nearest palette entry (full 8-entry L2),\n// writing indices into `out_idx` and returning the total squared error.\nfn assign_all(\n  values:  ptr<function, array<f32, 16>>,\n  pal:     ptr<function, array<f32, 8>>,\n  out_idx: ptr<function, array<u32, 16>>,\n) -> f32 {\n  var err: f32 = 0.0;\n  for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n    let v = (*values)[k];\n    var best_j: u32 = 0u;\n    var best_d: f32 = 1e20;\n    for (var j: u32 = 0u; j < 8u; j = j + 1u) {\n      let d  = (*pal)[j] - v;\n      let d2 = d * d;\n      if (d2 < best_d) {\n        best_d = d2;\n        best_j = j;\n      }\n    }\n    (*out_idx)[k] = best_j;\n    err = err + best_d;\n  }\n  return err;\n}\n\n// Encode 16 single-channel values into an 8-byte BC4 block, packed as\n// two little-endian u32s (u32[0] = bytes 0..3, u32[1] = bytes 4..7).\n// vmin/vmax are the channel's min/max, computed in the caller's load loop \u2014\n// fusing that scan there saves a 16-value pass per channel.\nfn encode_bc4(values: ptr<function, array<f32, 16>>, vmin: f32, vmax: f32) -> vec2<u32> {\n  // ---------------- 1. Initial endpoints: bbox of input ----------------\n  var r0: u32 = quantize8(vmax);\n  var r1: u32 = quantize8(vmin);\n  // Force 6-interp mode: red0 > red1 strictly.\n  if (r0 == r1) {\n    if (r1 > 0u) { r1 = r1 - 1u; }\n    else         { r0 = r0 + 1u; }\n  }\n\n  if (QUALITY_HIGH == 0u) {\n    // -------- fast: fused projection + LSQ refit, accept-if-better -------\n    // ONE fused pass: level = round(7\xB7(v \u2212 r0)/(r1 \u2212 r0)) projection\n    // assignment (the 8-entry 6-interp palette is colinear and evenly\n    // spaced, so the rounded projection IS the nearest-entry search), the\n    // seed solution's packed indices and squared error, and the least-\n    // squares normal-equation sums. The refit endpoints are re-quantised,\n    // reprojected, and accepted only if the block error decreases \u2014 worth\n    // ~1.3 dB on the normal-map card over the refit-free seed.\n    // Level \u2192 BC4 index LUT (0,2,3,4,5,6,7,1) packed as 3-bit entries in\n    // 0x3F58D0. Pixel k's 3 bits start at bit 3k+16 of the (w0,w1) pair\n    // (bytes 0..1 are the endpoints); k = 5 straddles the word boundary.\n    let r0f = f32(r0) / 255.0;\n    let dir = f32(r1) / 255.0 - r0f;\n    let scale = 7.0 / dir;\n    var w0 = r0 | (r1 << 8u);\n    var w1 = 0u;\n    var sAA = 0.0; var sBB = 0.0; var sAB = 0.0;\n    var sAV = 0.0; var sBV = 0.0;\n    var s_min = 7.0; var s_max = 0.0;\n    var seed_err = 0.0;\n    for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n      let vr = (*values)[k] - r0f;\n      let L = clamp(floor(vr * scale + 0.5), 0.0, 7.0);\n      s_min = min(s_min, L); s_max = max(s_max, L);\n      let b = L * (1.0 / 7.0); let a = 1.0 - b;\n      sAA = sAA + a * a; sBB = sBB + b * b; sAB = sAB + a * b;\n      sAV = sAV + a * vr; sBV = sBV + b * vr;\n      let e = vr - b * dir;\n      seed_err = seed_err + e * e;\n      let idx = (0x3F58D0u >> (u32(L) * 3u)) & 7u;\n      let bit = 3u * k + 16u;\n      if (bit <= 29u) {\n        w0 = w0 | (idx << bit);\n      } else if (bit >= 32u) {\n        w1 = w1 | (idx << (bit - 32u));\n      } else {\n        w0 = w0 | (idx << bit);\n        w1 = w1 | (idx >> (32u - bit));\n      }\n    }\n\n    // Rank-1 guard: with every pixel on ONE level the system is singular\n    // (det is float rounding noise); with \u22652 distinct levels\n    // det = \u03A3_i<j (b_j \u2212 b_i)\xB2 \u2265 15/49 \u2248 0.306.\n    if (s_min < s_max) {\n      let det = sAA * sBB - sAB * sAB;\n      if (abs(det) > 1e-3) {\n        // Clamp the refit to the block's value range (a strict-SSE win vs\n        // clamping to [0,1], same as the other formats' fast paths).\n        let e0 = clamp(r0f + (sBB * sAV - sAB * sBV) / det, vmin, vmax);\n        let e1 = clamp(r0f + (sAA * sBV - sAB * sAV) / det, vmin, vmax);\n        let n0 = quantize8(e0);\n        let n1 = quantize8(e1);\n        // Keep 6-interp mode (r0 > r1 strictly); skip the no-op refit.\n        if (n0 > n1 && !(n0 == r0 && n1 == r1)) {\n          let n0f = f32(n0) / 255.0;\n          let ndir = f32(n1) / 255.0 - n0f;\n          let nscale = 7.0 / ndir;\n          var nw0 = n0 | (n1 << 8u);\n          var nw1 = 0u;\n          var refit_err = 0.0;\n          for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n            let vr = (*values)[k] - n0f;\n            let L = clamp(floor(vr * nscale + 0.5), 0.0, 7.0);\n            let e = vr - L * (1.0 / 7.0) * ndir;\n            refit_err = refit_err + e * e;\n            let idx = (0x3F58D0u >> (u32(L) * 3u)) & 7u;\n            let bit = 3u * k + 16u;\n            if (bit <= 29u) {\n              nw0 = nw0 | (idx << bit);\n            } else if (bit >= 32u) {\n              nw1 = nw1 | (idx << (bit - 32u));\n            } else {\n              nw0 = nw0 | (idx << bit);\n              nw1 = nw1 | (idx >> (32u - bit));\n            }\n          }\n          if (refit_err < seed_err) {\n            return vec2<u32>(nw0, nw1);\n          }\n        }\n      }\n    }\n    return vec2<u32>(w0, w1);\n  }\n\n  // ---------------- 2. Initial palette + indices + error --------------\n  var pal: array<f32, 8>;\n  build_pal(f32(r0) / 255.0, f32(r1) / 255.0, &pal);\n  var indices: array<u32, 16>;\n  var err = assign_all(values, &pal, &indices);\n\n  // ---------------- 3. Refinement: least-squares on (r0, r1) ----------\n  // High-quality only \u2014 the refit is the bulk of the per-channel cost and the\n  // branch is resolved at pipeline-compile time, so the fast path skips all of\n  // it (the sums loop included), not just the acceptance test.\n  if (QUALITY_HIGH != 0u) {\n    // Normal equations for palette[j] = a_j * r0 + b_j * r1:\n    //   [\u03A3AA  \u03A3AB] [r0]   [\u03A3AV]\n    //   [\u03A3AB  \u03A3BB] [r1] = [\u03A3BV]\n    var sAA: f32 = 0.0; var sBB: f32 = 0.0; var sAB: f32 = 0.0;\n    var sAV: f32 = 0.0; var sBV: f32 = 0.0;\n    for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n      let a = w0_6(indices[k]);\n      let b = w1_6(indices[k]);\n      let v = (*values)[k];\n      sAA = sAA + a * a;\n      sBB = sBB + b * b;\n      sAB = sAB + a * b;\n      sAV = sAV + a * v;\n      sBV = sBV + b * v;\n    }\n    let det = sAA * sBB - sAB * sAB;\n    // Degenerate system \u2192 skip refinement.\n    if (abs(det) > 1e-9) {\n      let new_r0 = clamp((sBB * sAV - sAB * sBV) / det, 0.0, 1.0);\n      let new_r1 = clamp((sAA * sBV - sAB * sAV) / det, 0.0, 1.0);\n      let qR0 = quantize8(new_r0);\n      let qR1 = quantize8(new_r1);\n      // Only accept refinements that stay in 6-interp mode. A refinement\n      // that flips or equalizes the endpoints would change decode mode.\n      if (qR0 > qR1) {\n        build_pal(f32(qR0) / 255.0, f32(qR1) / 255.0, &pal);\n        var idx2: array<u32, 16>;\n        let err2 = assign_all(values, &pal, &idx2);\n        if (err2 < err) {\n          r0 = qR0;\n          r1 = qR1;\n          indices = idx2;\n          err = err2;\n        }\n      }\n    }\n  }\n\n  // ---------------- 4. Pack 48-bit index field + 2 endpoint bytes -----\n  // The 48-bit index field spans block bytes 2..7. Split into idx_lo\n  // (low 32 bits of the field) and idx_hi (high 16 bits). An index at\n  // bit position 3k straddles the 32-bit boundary iff 3k < 32 < 3k+3\n  // (only k = 10, 11 straddle: bits 30..32 and 33..35; actually k=10\n  // is bits 30..32, k=11 is 33..35 \u2014 so k=10 straddles). We handle\n  // straddles by writing to both halves.\n  var idx_lo: u32 = 0u;\n  var idx_hi: u32 = 0u;\n  for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n    let bit = 3u * k;\n    let v = indices[k] & 7u;\n    if (bit + 3u <= 32u) {\n      idx_lo = idx_lo | (v << bit);\n    } else if (bit >= 32u) {\n      idx_hi = idx_hi | (v << (bit - 32u));\n    } else {\n      // Straddle: low part into idx_lo's top, high part into idx_hi's bottom.\n      idx_lo = idx_lo | (v << bit);\n      idx_hi = idx_hi | (v >> (32u - bit));\n    }\n  }\n\n  // Final u32s, both little-endian:\n  //   u32[0] bytes = red0, red1, idx_lo[7:0], idx_lo[15:8]\n  //   u32[1] bytes = idx_lo[23:16], idx_lo[31:24], idx_hi[7:0], idx_hi[15:8]\n  let out_lo = r0 | (r1 << 8u) | ((idx_lo & 0xFFFFu) << 16u);\n  let out_hi = (idx_lo >> 16u) | (idx_hi << 16u);\n\n  return vec2<u32>(out_lo, out_hi);\n}\n\n@compute @workgroup_size(8, 8, 1)\nfn encode(@builtin(global_invocation_id) gid: vec3<u32>) {\n  if (gid.x >= params.blocks_x || gid.y >= params.blocks_y) {\n    return;\n  }\n\n  let bx = gid.x;\n  let by = gid.y;\n  let block_index = by * params.blocks_x + bx;\n\n  let base   = vec2<i32>(i32(bx) * 4, i32(by) * 4);\n  let max_xy = vec2<i32>(i32(params.width) - 1, i32(params.height) - 1);\n\n  // Load 4\xD74 RG values, splitting into per-channel arrays so each can\n  // be handed to encode_bc4 independently; the per-channel min/max scan is\n  // fused into the same loop.\n  var r_values: array<f32, 16>;\n  var g_values: array<f32, 16>;\n  var r_min: f32 = 1.0; var r_max: f32 = 0.0;\n  var g_min: f32 = 1.0; var g_max: f32 = 0.0;\n  for (var i: u32 = 0u; i < 16u; i = i + 1u) {\n    let lx = i32(i & 3u);\n    let ly = i32(i >> 2u);\n    // Clamp to edge for non-multiple-of-4 input sizes.\n    let p  = clamp(base + vec2<i32>(lx, ly), vec2<i32>(0, 0), max_xy);\n    let c  = textureLoad(src_tex, p, 0);\n    r_values[i] = c.r;\n    g_values[i] = c.g;\n    r_min = min(r_min, c.r); r_max = max(r_max, c.r);\n    g_min = min(g_min, c.g); g_max = max(g_max, c.g);\n  }\n\n  let r_block = encode_bc4(&r_values, r_min, r_max);\n  let g_block = encode_bc4(&g_values, g_min, g_max);\n\n  // BC5 block = R half (bytes 0..7) || G half (bytes 8..15) = 4 u32s.\n  let out = block_index * 4u;\n  dst[out + 0u] = r_block.x;\n  dst[out + 1u] = r_block.y;\n  dst[out + 2u] = g_block.x;\n  dst[out + 3u] = g_block.y;\n}\n";
 
 // src/bc5_fast_f16.wgsl
 var bc5_fast_f16_default = `// bc5 "fast" encoder \u2014 f16 variant (requires the shader-f16 feature).
-// Two BC4 halves (R and G), no refit \u2014 same output family as bc5.wgsl's fast
-// branch, tuned for throughput:
+// Two BC4 halves (R and G) \u2014 same output family as bc5.wgsl's fast branch,
+// tuned for throughput:
 //
 //   \u2022 The 8-entry palette in 6-interpolation mode is COLINEAR and EVENLY
 //     spaced from r0 to r1 (levels 0..7 in palette order 0,2,3,4,5,6,7,1),
@@ -522,9 +684,21 @@ var bc5_fast_f16_default = `// bc5 "fast" encoder \u2014 f16 variant (requires t
 //   \u2022 Math runs in the exact-integer [0,255] f16 domain: endpoints and pixel
 //     values are whole numbers \u2264 255 (exact in f16), so the only rounding is
 //     the single 1/(r1\u2212r0) division.
+//   \u2022 ONE fused pass per channel: projection assignment + the least-squares
+//     refit sums + the seed solution's packed indices and squared error. The
+//     refit endpoints are re-quantised, reprojected, and accepted only if
+//     the block error decreases (same accept-if-better family as the BC1
+//     fast path) \u2014 worth ~1.3 dB on the normal-map card.
 //   \u2022 3-bit indices are packed into the 48-bit field on the fly \u2014 no
 //     array<u32,16> private array and no separate packing loop.
 //
+// f16 range notes: value sums accumulate v \u2212 r0 (the affine-basis shift trick
+// from the BC7/ASTC fast paths) scaled by 1/16, and error residuals are
+// scaled by 1/16 before squaring \u2014 worst-case magnitudes stay \u22724k, well
+// inside f16's 65504 max, with rounding a small fraction of a level. The
+// accept-if-better guard makes any residual f16 noise fail-safe (worst case:
+// the refit is rejected and the seed solution ships).
+//
 // Level \u2192 BC4 index (0\u2192r0 ... 7\u2192r1): 0,2,3,4,5,6,7,1 \u2014 packed 3-bit LUT
 // 0x3F58D0 = sum(idx[L] << 3L).
 //
@@ -537,14 +711,43 @@ struct Params { blocks_x: u32, blocks_y: u32, width: u32, height: u32, };
 @group(0) @binding(1) var<storage, read_write> dst: array<u32>;
 @group(0) @binding(2) var<uniform> params: Params;
 
-// Encode one channel (16 values in exact-integer [0,255] f16) to a BC4 half.
-fn encode_bc4(values: ptr<function, array<h, 16>>) -> vec2<u32> {
-  var vmin = h(255.0);
-  var vmax = h(0.0);
+// Project the 16 values onto the r0\u2192r1 axis, packing the 3-bit indices on the
+// fly (pixel k's index starts at bit 3k of the 48-bit field, i.e. bit 3k+16
+// of the (w0,w1) pair; k = 5 straddles the word boundary) and accumulating
+// the squared error (residuals scaled by 1/16 before squaring). Returns the
+// packed words with the endpoint bytes already in place.
+struct Proj { w0: u32, w1: u32, err: h };
+fn project_pack(values: ptr<function, array<h, 16>>, r0: u32, r1: u32) -> Proj {
+  let r0f = h(f32(r0));
+  let dir = h(f32(r1)) - r0f;
+  let scale = h(7.0) / dir;
+  var out: Proj;
+  out.w0 = r0 | (r1 << 8u);
+  out.w1 = 0u;
+  out.err = h(0.0);
   for (var k: u32 = 0u; k < 16u; k = k + 1u) {
-    vmin = min(vmin, (*values)[k]);
-    vmax = max(vmax, (*values)[k]);
+    let vr = (*values)[k] - r0f;
+    let L = clamp(floor(vr * scale + h(0.5)), h(0.0), h(7.0));
+    let e = (vr - L * h(1.0 / 7.0) * dir) * h(1.0 / 16.0);
+    out.err = out.err + e * e;
+    let idx = (0x3F58D0u >> (u32(L) * 3u)) & 7u;
+    let bit = 3u * k + 16u;
+    if (bit <= 29u) {
+      out.w0 = out.w0 | (idx << bit);
+    } else if (bit >= 32u) {
+      out.w1 = out.w1 | (idx << (bit - 32u));
+    } else {
+      out.w0 = out.w0 | (idx << bit);
+      out.w1 = out.w1 | (idx >> (32u - bit));
+    }
   }
+  return out;
+}
+
+// Encode one channel (16 values in exact-integer [0,255] f16) to a BC4 half.
+// vmin/vmax are the channel's min/max, computed in the caller's load loop \u2014
+// fusing that scan there saves a 16-value pass per channel.
+fn encode_bc4(values: ptr<function, array<h, 16>>, vmin: h, vmax: h) -> vec2<u32> {
   var r0 = u32(vmax); // values are exact integers \u2014 no rounding needed
   var r1 = u32(vmin);
   if (r0 == r1) {
@@ -552,31 +755,67 @@ fn encode_bc4(values: ptr<function, array<h, 16>>) -> vec2<u32> {
     if (r1 > 0u) { r1 = r1 - 1u; } else { r0 = r0 + 1u; }
   }
 
-  // Projection assignment: level = round(7\xB7(v \u2212 r0)/(r1 \u2212 r0)), clamped.
-  // |v \u2212 r0| \u2264 r0 \u2212 r1 for every in-block value, so the product stays \u2264 7.
+  // Fused seed pass: projection assignment (level = round(7\xB7(v \u2212 r0)/
+  // (r1 \u2212 r0)), clamped \u2014 |v \u2212 r0| \u2264 r0 \u2212 r1 for in-block values) + packed
+  // indices + seed error + the LSQ normal-equation sums. Value sums
+  // accumulate (v \u2212 r0)/16: the shift keeps the accumulators proportional to
+  // the block's span, the exact power-of-two scale keeps products \u2264 4080.
   let r0f = h(f32(r0));
-  let scale = h(7.0) / (h(f32(r1)) - r0f);
-  var lo: u32 = 0u;
-  var hi = r0 | (r1 << 8u); // endpoint bytes live in the low 16 bits of u32[0]
-  // Pixel k's 3-bit index starts at bit 3k of the 48-bit field, i.e. bit
-  // 3k+16 of u32[0] for k \u2264 4, straddling into u32[1] from k = 5 (bit 31).
-  var w0 = hi;
-  var w1 = 0u;
+  let dir = h(f32(r1)) - r0f;
+  let scale = h(7.0) / dir;
+  var seed: Proj;
+  seed.w0 = r0 | (r1 << 8u);
+  seed.w1 = 0u;
+  seed.err = h(0.0);
+  var sAA = h(0.0); var sBB = h(0.0); var sAB = h(0.0);
+  var sAV = h(0.0); var sBV = h(0.0);
+  var s_min = h(7.0); var s_max = h(0.0);
   for (var k: u32 = 0u; k < 16u; k = k + 1u) {
-    let L = u32(clamp(floor(((*values)[k] - r0f) * scale + h(0.5)), h(0.0), h(7.0)));
-    let idx = (0x3F58D0u >> (L * 3u)) & 7u;
+    let vr = (*values)[k] - r0f;
+    let L = clamp(floor(vr * scale + h(0.5)), h(0.0), h(7.0));
+    s_min = min(s_min, L); s_max = max(s_max, L);
+    let b = L * h(1.0 / 7.0); let a = h(1.0) - b;
+    let vr16 = vr * h(1.0 / 16.0);
+    sAA = sAA + a * a; sBB = sBB + b * b; sAB = sAB + a * b;
+    sAV = sAV + a * vr16; sBV = sBV + b * vr16;
+    let e = vr16 - b * dir * h(1.0 / 16.0);
+    seed.err = seed.err + e * e;
+    let idx = (0x3F58D0u >> (u32(L) * 3u)) & 7u;
     let bit = 3u * k + 16u;
     if (bit <= 29u) {
-      w0 = w0 | (idx << bit);
+      seed.w0 = seed.w0 | (idx << bit);
     } else if (bit >= 32u) {
-      w1 = w1 | (idx << (bit - 32u));
+      seed.w1 = seed.w1 | (idx << (bit - 32u));
     } else {
       // k = 5 straddles the word boundary (bits 31..33).
-      w0 = w0 | (idx << bit);
-      w1 = w1 | (idx >> (32u - bit));
+      seed.w0 = seed.w0 | (idx << bit);
+      seed.w1 = seed.w1 | (idx >> (32u - bit));
+    }
+  }
+
+  // LSQ refit, accepted only if the requantised endpoints lower the block
+  // error. Rank-1 guard: with every pixel on ONE level the system is
+  // singular and det is pure f16 rounding noise (\u22720.03); with \u22652 distinct
+  // levels det = \u03A3_i<j (b_j \u2212 b_i)\xB2 \u2265 15/49 \u2248 0.306 \u2014 0.1 separates cleanly.
+  if (s_min < s_max) {
+    let det = sAA * sBB - sAB * sAB;
+    if (abs(det) > h(0.1)) {
+      // \xD716 undoes the accumulator scale; clamp to the block's value range
+      // (a strict-SSE win vs clamping to [0,255], same as the other formats).
+      let e0 = clamp(r0f + (sBB * sAV - sAB * sBV) * h(16.0) / det, vmin, vmax);
+      let e1 = clamp(r0f + (sAA * sBV - sAB * sAV) * h(16.0) / det, vmin, vmax);
+      let n0 = u32(floor(e0 + h(0.5)));
+      let n1 = u32(floor(e1 + h(0.5)));
+      // Keep 6-interp mode (r0 > r1 strictly); skip the no-op refit.
+      if (n0 > n1 && !(n0 == r0 && n1 == r1)) {
+        let refit = project_pack(values, n0, n1);
+        if (refit.err < seed.err) {
+          return vec2<u32>(refit.w0, refit.w1);
+        }
+      }
     }
   }
-  return vec2<u32>(w0, w1);
+  return vec2<u32>(seed.w0, seed.w1);
 }
 
 @compute @workgroup_size(8, 8, 1)
@@ -587,14 +826,18 @@ fn encode(@builtin(global_invocation_id) gid: vec3<u32>) {
   let mx = vec2<i32>(i32(params.width) - 1, i32(params.height) - 1);
   var rv: array<h, 16>;
   var gv: array<h, 16>;
+  var rmin = h(255.0); var rmax = h(0.0);
+  var gmin = h(255.0); var gmax = h(0.0);
   for (var i: u32 = 0u; i < 16u; i = i + 1u) {
     let p = clamp(base + vec2<i32>(i32(i & 3u), i32(i >> 2u)), vec2<i32>(0), mx);
     let c = textureLoad(src_tex, p, 0);
-    rv[i] = h(c.r * 255.0);
-    gv[i] = h(c.g * 255.0);
+    let r = h(c.r * 255.0);
+    let g = h(c.g * 255.0);
+    rv[i] = r; rmin = min(rmin, r); rmax = max(rmax, r);
+    gv[i] = g; gmin = min(gmin, g); gmax = max(gmax, g);
   }
-  let rb = encode_bc4(&rv);
-  let gb = encode_bc4(&gv);
+  let rb = encode_bc4(&rv, rmin, rmax);
+  let gb = encode_bc4(&gv, gmin, gmax);
   let o = bi * 4u;
   dst[o] = rb.x; dst[o + 1u] = rb.y; dst[o + 2u] = gb.x; dst[o + 3u] = gb.y;
 }
@@ -628,28 +871,31 @@ var BC5Encoder = class extends Encoder {
 };
 
 // src/bc7.wgsl
-var bc7_default = "// BC7 (BPTC) mode 6 compute shader encoder.\n//\n// One invocation per 4\xD74 block. Emits 16 bytes = 4 u32s into the storage\n// buffer at `dst[block_index * 4 .. + 3]`.\n//\n// QUALITY LEVELS (pipeline-overridable constant `QUALITY_HIGH`)\n//   fast (0, default): O(N) bounding-box seed \u2192 one fused pass that projects\n//     each pixel onto the endpoint line (the 16 palette entries are colinear,\n//     so the nearest index is the rounded projection \u2014 no palette build, no\n//     16-entry search) while accumulating the least-squares refit sums, then\n//     a reprojection against the quantised refit endpoints for the final\n//     indices, packed on the fly into two nibble words.\n//   high (1): farthest-pair seed, exhaustive p-bit search over all four\n//     (p0,p1) \u2208 {0,1}\xB2 combos, full 16-entry nearest search, one LSQ refit \u2014\n//     matches bc7_ref.ts up to FP tie-breaks.\n//\n// The fast/high branch is selected at pipeline-compile time, so the driver\n// eliminates the unused code entirely.\n//\n// MODE 6 LAYOUT (LSB-first, bit 0 = byte 0's bit 0)\n//   bits 0..6    mode field      (0b0000001 \u2014 only bit 6 is 1)\n//   bits 7..13   R0 (7-bit)   bits 14..20 R1   bits 21..27 G0   bits 28..34 G1\n//   bits 35..41  B0   bits 42..48 B1   bits 49..55 A0   bits 56..62 A1\n//   bit  63      P0   bit 64 P1\n//   bits 65..67  pixel 0 index (3 bits; anchor, MSB implicit 0)\n//   bits 68..71  pixel 1 index (4 bits) ... bits 124..127 pixel 15 index\n//\n// Effective 8-bit endpoint channel = (7_bit_value << 1) | p_bit.\n// Palette[i] = ((64 \u2212 W4[i]) \xD7 e0_8 + W4[i] \xD7 e1_8 + 32) >> 6, integer.\n//\n// The block is assembled with straight-line constant shifts (see the layout\n// summary in bc7_fast_f16.wgsl) \u2014 a generic write_bits() helper's dynamic\n// word indexing keeps the output array out of registers.\n\n// 0 = fast (default), 1 = exhaustive/high-quality. Set via pipeline constants.\noverride QUALITY_HIGH: u32 = 0u;\n\nstruct Params {\n  blocks_x: u32,\n  blocks_y: u32,\n  width:    u32,\n  height:   u32,\n};\n\n@group(0) @binding(0) var src_tex: texture_2d<f32>;\n@group(0) @binding(1) var<storage, read_write> dst: array<u32>;\n@group(0) @binding(2) var<uniform> params: Params;\n\n// Mode 6 interpolation weights (\xD7 1/64), fixed by the spec (`W4` in bc7_ref.ts).\nfn w4(i: u32) -> u32 {\n  switch i {\n    case 0u:  { return  0u; }\n    case 1u:  { return  4u; }\n    case 2u:  { return  9u; }\n    case 3u:  { return 13u; }\n    case 4u:  { return 17u; }\n    case 5u:  { return 21u; }\n    case 6u:  { return 26u; }\n    case 7u:  { return 30u; }\n    case 8u:  { return 34u; }\n    case 9u:  { return 38u; }\n    case 10u: { return 43u; }\n    case 11u: { return 47u; }\n    case 12u: { return 51u; }\n    case 13u: { return 55u; }\n    case 14u: { return 60u; }\n    default:  { return 64u; }   // case 15u\n  }\n}\n\nfn interp4(e0: vec4<i32>, e1: vec4<i32>, w: i32) -> vec4<i32> {\n  return ((64 - w) * e0 + w * e1 + vec4<i32>(32)) >> vec4<u32>(6u);\n}\n\nfn to8(v: vec4<f32>) -> vec4<i32> {\n  return vec4<i32>(clamp(floor(v * 255.0 + 0.5), vec4<f32>(0.0), vec4<f32>(255.0)));\n}\n\nfn dist2(a: vec4<i32>, b: vec4<i32>) -> i32 {\n  let d = a - b;\n  let e = d * d;\n  return e.x + e.y + e.z + e.w;\n}\n\n// Quantize an 8-bit ideal endpoint to (7-bit value, reconstructed 8-bit) under\n// a fixed p-bit, all four channels at once. q7 = round((ideal8 \u2212 p)/2); used by\n// both paths.\nstruct QuantPair { seven: vec4<i32>, eight: vec4<i32> };\nfn quantize_endpoint(ideal8: vec4<i32>, p: u32) -> QuantPair {\n  let q = vec4<i32>(clamp(\n    floor((vec4<f32>(ideal8) - f32(p)) / 2.0 + 0.5),\n    vec4<f32>(0.0), vec4<f32>(127.0),\n  ));\n  let eff = (q << vec4<u32>(1u)) | vec4<i32>(i32(p));\n  return QuantPair(q, eff);\n}\n\n// ============================ FAST PATH ================================ //\n\n// Endpoint with its chosen p-bit, picked by minimum quantisation error.\nstruct Ep { seven: vec4<i32>, eight: vec4<i32>, p: u32 };\nfn pick_ep(ideal: vec4<i32>) -> Ep {\n  let a = quantize_endpoint(ideal, 0u);\n  let b = quantize_endpoint(ideal, 1u);\n  if (dist2(b.eight, ideal) < dist2(a.eight, ideal)) { return Ep(b.seven, b.eight, 1u); }\n  return Ep(a.seven, a.eight, 0u);\n}\n\n// One pass over the block: project every pixel onto the e0\u2192e1 line and\n// accumulate the least-squares normal-equation sums; solve for the refit\n// endpoints (in 8-bit space). Indices are not produced here \u2014 the caller\n// reprojects against the quantised refit endpoints anyway.\nstruct Fit { e0: vec4<i32>, e1: vec4<i32>, valid: bool };\nfn proj_fit(pixels: ptr<function, array<vec4<i32>, 16>>, e0: vec4<i32>, e1: vec4<i32>) -> Fit {\n  var out: Fit;\n  out.valid = false;\n  let dir = vec4<f32>(e1 - e0);\n  let dd = dot(dir, dir);\n  if (dd == 0.0) { return out; }\n  let e0f = vec4<f32>(e0);\n  let inv = 15.0 / dd;\n  var sAA: f32 = 0.0; var sBB: f32 = 0.0; var sAB: f32 = 0.0;\n  var sAV: vec4<f32> = vec4<f32>(0.0); var sBV: vec4<f32> = vec4<f32>(0.0);\n  var s_min = 15.0; var s_max = 0.0;\n  for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n    let v = vec4<f32>((*pixels)[k]);\n    let s = clamp(floor(dot(v - e0f, dir) * inv + 0.5), 0.0, 15.0);\n    s_min = min(s_min, s); s_max = max(s_max, s);\n    let b = s * (1.0 / 15.0); let a = 1.0 - b;\n    sAA = sAA + a * a; sBB = sBB + b * b; sAB = sAB + a * b;\n    sAV = sAV + a * v; sBV = sBV + b * v;\n  }\n  // Rank-1 guard: if every pixel projects to ONE level the system is\n  // singular \u2014 det and the numerators are pure float rounding noise and the\n  // solve returns garbage endpoints. With \u22652 levels det \u2265 15/225 \u2248 0.067.\n  if (s_min == s_max) { return out; }\n  let det = sAA * sBB - sAB * sAB;\n  if (abs(det) < 1e-3) { return out; }\n  out.e0 = vec4<i32>(clamp(round((sBB * sAV - sAB * sBV) / det), vec4<f32>(0.0), vec4<f32>(255.0)));\n  out.e1 = vec4<i32>(clamp(round((sAA * sBV - sAB * sAV) / det), vec4<f32>(0.0), vec4<f32>(255.0)));\n  out.valid = true;\n  return out;\n}\n\n// ============================ HIGH PATH ================================ //\n\nstruct Pair { a: vec4<i32>, b: vec4<i32> };\nfn farthest_pair(pixels: ptr<function, array<vec4<i32>, 16>>) -> Pair {\n  var best_d: i32 = 0;\n  var pa = (*pixels)[0];\n  var pb = (*pixels)[1];\n  for (var i: u32 = 0u; i < 16u; i = i + 1u) {\n    let xi = (*pixels)[i];\n    for (var j: u32 = i + 1u; j < 16u; j = j + 1u) {\n      let d = dist2(xi, (*pixels)[j]);\n      if (d > best_d) { best_d = d; pa = xi; pb = (*pixels)[j]; }\n    }\n  }\n  return Pair(pa, pb);\n}\n\nfn build_palette_6(e0: vec4<i32>, e1: vec4<i32>, pal: ptr<function, array<vec4<i32>, 16>>) {\n  for (var i: u32 = 0u; i < 16u; i = i + 1u) {\n    (*pal)[i] = interp4(e0, e1, i32(w4(i)));\n  }\n}\n\nfn assign_all(\n  pixels:  ptr<function, array<vec4<i32>, 16>>,\n  pal:     ptr<function, array<vec4<i32>, 16>>,\n  out_idx: ptr<function, array<u32, 16>>,\n) -> i32 {\n  var err: i32 = 0;\n  for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n    let px = (*pixels)[k];\n    var best_i: u32 = 0u;\n    var best_d: i32 = 2147483647;\n    for (var i: u32 = 0u; i < 16u; i = i + 1u) {\n      let d = dist2(px, (*pal)[i]);\n      if (d < best_d) { best_d = d; best_i = i; }\n    }\n    (*out_idx)[k] = best_i;\n    err = err + best_d;\n  }\n  return err;\n}\n\nstruct BestMode6 {\n  e0_7: vec4<i32>, e1_7: vec4<i32>,\n  p0: u32,         p1: u32,\n  indices: array<u32, 16>,\n  err: i32,\n};\n\n// Exhaustive p-bit search (high path); commits to `*best` only on improvement.\nfn try_pbit_combos(\n  pixels: ptr<function, array<vec4<i32>, 16>>,\n  ideal0: vec4<i32>,\n  ideal1: vec4<i32>,\n  best:   ptr<function, BestMode6>,\n) {\n  var local_best = (*best).err;\n  var pal: array<vec4<i32>, 16>;\n  var tmp: array<u32, 16>;\n  for (var p0: u32 = 0u; p0 < 2u; p0 = p0 + 1u) {\n    let q0 = quantize_endpoint(ideal0, p0);\n    for (var p1: u32 = 0u; p1 < 2u; p1 = p1 + 1u) {\n      let q1 = quantize_endpoint(ideal1, p1);\n      build_palette_6(q0.eight, q1.eight, &pal);\n      let err = assign_all(pixels, &pal, &tmp);\n      if (err < local_best) {\n        local_best = err;\n        (*best).e0_7 = q0.seven;\n        (*best).e1_7 = q1.seven;\n        (*best).p0 = p0;\n        (*best).p1 = p1;\n        (*best).indices = tmp;\n        (*best).err = err;\n      }\n    }\n  }\n}\n\n// Exact-weight LSQ refit (high path); matches bc7_ref.ts `refitEndpointsMode6`.\nstruct RefitResult { e0: vec4<i32>, e1: vec4<i32>, valid: bool };\nfn refit_endpoints(\n  pixels: ptr<function, array<vec4<i32>, 16>>,\n  indices: ptr<function, array<u32, 16>>,\n) -> RefitResult {\n  var sAA: f32 = 0.0;\n  var sBB: f32 = 0.0;\n  var sAB: f32 = 0.0;\n  var sAV: vec4<f32> = vec4<f32>(0.0);\n  var sBV: vec4<f32> = vec4<f32>(0.0);\n  for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n    let i = (*indices)[k];\n    let a = f32(64u - w4(i)) / 64.0;\n    let b = f32(w4(i)) / 64.0;\n    let v = vec4<f32>((*pixels)[k]);\n    sAA = sAA + a * a;\n    sBB = sBB + b * b;\n    sAB = sAB + a * b;\n    sAV = sAV + a * v;\n    sBV = sBV + b * v;\n  }\n  let det = sAA * sBB - sAB * sAB;\n  var out: RefitResult;\n  if (abs(det) < 1e-9) {\n    out.valid = false;\n    return out;\n  }\n  let e0f = (sBB * sAV - sAB * sBV) / det;\n  let e1f = (sAA * sBV - sAB * sAV) / det;\n  out.e0 = vec4<i32>(clamp(round(e0f), vec4<f32>(0.0), vec4<f32>(255.0)));\n  out.e1 = vec4<i32>(clamp(round(e1f), vec4<f32>(0.0), vec4<f32>(255.0)));\n  out.valid = true;\n  return out;\n}\n\n// ------------------------------- Entry --------------------------------- //\n\n@compute @workgroup_size(8, 8, 1)\nfn encode(@builtin(global_invocation_id) gid: vec3<u32>) {\n  if (gid.x >= params.blocks_x || gid.y >= params.blocks_y) {\n    return;\n  }\n\n  let bx = gid.x;\n  let by = gid.y;\n  let block_index = by * params.blocks_x + bx;\n\n  let base   = vec2<i32>(i32(bx) * 4, i32(by) * 4);\n  let max_xy = vec2<i32>(i32(params.width) - 1, i32(params.height) - 1);\n\n  // Load 16 RGBA pixels (8-bit integer domain) and the per-channel bbox.\n  var pixels: array<vec4<i32>, 16>;\n  var lo = vec4<i32>(255);\n  var hi = vec4<i32>(0);\n  for (var i: u32 = 0u; i < 16u; i = i + 1u) {\n    let lx = i32(i & 3u);\n    let ly = i32(i >> 2u);\n    let p  = clamp(base + vec2<i32>(lx, ly), vec2<i32>(0, 0), max_xy);\n    let px = to8(textureLoad(src_tex, p, 0));\n    pixels[i] = px;\n    lo = min(lo, px);\n    hi = max(hi, px);\n  }\n\n  // Both branches produce: 7-bit endpoints + p-bits, and the 16 4-bit indices\n  // packed LSB-first into two nibble words (pixel k \u2192 bits 4k..4k+3).\n  var e0_7: vec4<i32>;\n  var e1_7: vec4<i32>;\n  var p0: u32;\n  var p1: u32;\n  var ilo: u32 = 0u;\n  var ihi: u32 = 0u;\n\n  if (QUALITY_HIGH != 0u) {\n    let fp = farthest_pair(&pixels);\n    var best: BestMode6;\n    best.err = 2147483647;\n    try_pbit_combos(&pixels, fp.a, fp.b, &best);\n    let refit = refit_endpoints(&pixels, &best.indices);\n    if (refit.valid) {\n      try_pbit_combos(&pixels, refit.e0, refit.e1, &best);\n    }\n    e0_7 = best.e0_7; e1_7 = best.e1_7; p0 = best.p0; p1 = best.p1;\n    for (var k: u32 = 0u; k < 8u; k = k + 1u) {\n      ilo = ilo | (best.indices[k] << (k * 4u));\n    }\n    for (var k: u32 = 8u; k < 16u; k = k + 1u) {\n      ihi = ihi | (best.indices[k] << ((k - 8u) * 4u));\n    }\n  } else {\n    // Seed the fused LSQ fit from the raw bbox, then quantise the refit\n    // endpoints and reproject for the final indices.\n    let r = proj_fit(&pixels, lo, hi);\n    var ep0: Ep;\n    var ep1: Ep;\n    if (r.valid) { ep0 = pick_ep(r.e0); ep1 = pick_ep(r.e1); }\n    else         { ep0 = pick_ep(lo);   ep1 = pick_ep(hi);   }\n    let dir = vec4<f32>(ep1.eight - ep0.eight);\n    let dd = dot(dir, dir);\n    if (dd > 0.0) {\n      let e0f = vec4<f32>(ep0.eight);\n      let inv = 15.0 / dd;\n      for (var k: u32 = 0u; k < 8u; k = k + 1u) {\n        let s = clamp(floor(dot(vec4<f32>(pixels[k]) - e0f, dir) * inv + 0.5), 0.0, 15.0);\n        ilo = ilo | (u32(s) << (k * 4u));\n      }\n      for (var k: u32 = 8u; k < 16u; k = k + 1u) {\n        let s = clamp(floor(dot(vec4<f32>(pixels[k]) - e0f, dir) * inv + 0.5), 0.0, 15.0);\n        ihi = ihi | (u32(s) << ((k - 8u) * 4u));\n      }\n    }\n    e0_7 = ep0.seven; e1_7 = ep1.seven; p0 = ep0.p; p1 = ep1.p;\n  }\n\n  // Anchor rule \u2014 pixel 0's index MSB must be 0. Swapping endpoints reflects\n  // every index (i \u2192 15\u2212i), which on packed nibbles is a bitwise NOT.\n  if ((ilo & 0x8u) != 0u) {\n    let t7 = e0_7; e0_7 = e1_7; e1_7 = t7;\n    let tp = p0;   p0   = p1;   p1   = tp;\n    ilo = ~ilo; ihi = ~ihi;\n  }\n\n  // Straight-line mode-6 packing (see layout at the top of the file).\n  let e0 = vec4<u32>(e0_7);\n  let e1 = vec4<u32>(e1_7);\n  let w0 = 0x40u | (e0.x << 7u) | (e1.x << 14u) | (e0.y << 21u) | (e1.y << 28u);\n  let w1 = (e1.y >> 4u) | (e0.z << 3u) | (e1.z << 10u) | (e0.w << 17u) | (e1.w << 24u) | (p0 << 31u);\n  let w2 = p1 | ((ilo & 0x7u) << 1u) | (ilo & 0xFFFFFFF0u);\n  let w3 = ihi;\n\n  let out = block_index * 4u;\n  dst[out + 0u] = w0;\n  dst[out + 1u] = w1;\n  dst[out + 2u] = w2;\n  dst[out + 3u] = w3;\n}\n";
+var bc7_default = "// BC7 (BPTC) mode 6 compute shader encoder.\n//\n// One invocation per 4\xD74 block. Emits 16 bytes = 4 u32s into the storage\n// buffer at `dst[block_index * 4 .. + 3]`.\n//\n// QUALITY LEVELS (pipeline-overridable constant `QUALITY_HIGH`)\n//   fast (0, default): principal-axis seed (covariance power-iteration; bbox\n//     on degenerate blocks) at the exact projection extents, quantised\n//     directly \u2014 no LSQ refit; with the seed on the principal axis, mode 6's\n//     16-level palette leaves the refit under 0.15 dB, unlike the 4-level\n//     BC1/ASTC fast paths which keep theirs \u2014 then one pass that projects\n//     each pixel onto the endpoint line (the 16 palette entries are\n//     colinear, so the nearest index is the rounded projection \u2014 no palette\n//     build, no 16-entry search), packed on the fly into two nibble words.\n//   high (1): farthest-pair seed, exhaustive p-bit search over all four\n//     (p0,p1) \u2208 {0,1}\xB2 combos, full 16-entry nearest search, one LSQ refit \u2014\n//     matches bc7_ref.ts up to FP tie-breaks.\n//\n// The fast/high branch is selected at pipeline-compile time, so the driver\n// eliminates the unused code entirely.\n//\n// MODE 6 LAYOUT (LSB-first, bit 0 = byte 0's bit 0)\n//   bits 0..6    mode field      (0b0000001 \u2014 only bit 6 is 1)\n//   bits 7..13   R0 (7-bit)   bits 14..20 R1   bits 21..27 G0   bits 28..34 G1\n//   bits 35..41  B0   bits 42..48 B1   bits 49..55 A0   bits 56..62 A1\n//   bit  63      P0   bit 64 P1\n//   bits 65..67  pixel 0 index (3 bits; anchor, MSB implicit 0)\n//   bits 68..71  pixel 1 index (4 bits) ... bits 124..127 pixel 15 index\n//\n// Effective 8-bit endpoint channel = (7_bit_value << 1) | p_bit.\n// Palette[i] = ((64 \u2212 W4[i]) \xD7 e0_8 + W4[i] \xD7 e1_8 + 32) >> 6, integer.\n//\n// The block is assembled with straight-line constant shifts (see the layout\n// summary in bc7_fast_f16.wgsl) \u2014 a generic write_bits() helper's dynamic\n// word indexing keeps the output array out of registers.\n\n// 0 = fast (default), 1 = exhaustive/high-quality. Set via pipeline constants.\noverride QUALITY_HIGH: u32 = 0u;\n\nstruct Params {\n  blocks_x: u32,\n  blocks_y: u32,\n  width:    u32,\n  height:   u32,\n};\n\n@group(0) @binding(0) var src_tex: texture_2d<f32>;\n@group(0) @binding(1) var<storage, read_write> dst: array<u32>;\n@group(0) @binding(2) var<uniform> params: Params;\n\n// Mode 6 interpolation weights (\xD7 1/64), fixed by the spec (`W4` in bc7_ref.ts).\nfn w4(i: u32) -> u32 {\n  switch i {\n    case 0u:  { return  0u; }\n    case 1u:  { return  4u; }\n    case 2u:  { return  9u; }\n    case 3u:  { return 13u; }\n    case 4u:  { return 17u; }\n    case 5u:  { return 21u; }\n    case 6u:  { return 26u; }\n    case 7u:  { return 30u; }\n    case 8u:  { return 34u; }\n    case 9u:  { return 38u; }\n    case 10u: { return 43u; }\n    case 11u: { return 47u; }\n    case 12u: { return 51u; }\n    case 13u: { return 55u; }\n    case 14u: { return 60u; }\n    default:  { return 64u; }   // case 15u\n  }\n}\n\nfn interp4(e0: vec4<i32>, e1: vec4<i32>, w: i32) -> vec4<i32> {\n  return ((64 - w) * e0 + w * e1 + vec4<i32>(32)) >> vec4<u32>(6u);\n}\n\nfn to8(v: vec4<f32>) -> vec4<i32> {\n  return vec4<i32>(clamp(floor(v * 255.0 + 0.5), vec4<f32>(0.0), vec4<f32>(255.0)));\n}\n\nfn dist2(a: vec4<i32>, b: vec4<i32>) -> i32 {\n  let d = a - b;\n  let e = d * d;\n  return e.x + e.y + e.z + e.w;\n}\n\n// Quantize an 8-bit ideal endpoint to (7-bit value, reconstructed 8-bit) under\n// a fixed p-bit, all four channels at once. q7 = round((ideal8 \u2212 p)/2); used by\n// both paths.\nstruct QuantPair { seven: vec4<i32>, eight: vec4<i32> };\nfn quantize_endpoint(ideal8: vec4<i32>, p: u32) -> QuantPair {\n  let q = vec4<i32>(clamp(\n    floor((vec4<f32>(ideal8) - f32(p)) / 2.0 + 0.5),\n    vec4<f32>(0.0), vec4<f32>(127.0),\n  ));\n  let eff = (q << vec4<u32>(1u)) | vec4<i32>(i32(p));\n  return QuantPair(q, eff);\n}\n\n// ============================ FAST PATH ================================ //\n\n// Endpoint with its chosen p-bit, picked by minimum quantisation error.\nstruct Ep { seven: vec4<i32>, eight: vec4<i32>, p: u32 };\nfn pick_ep(ideal: vec4<i32>) -> Ep {\n  let a = quantize_endpoint(ideal, 0u);\n  let b = quantize_endpoint(ideal, 1u);\n  if (dist2(b.eight, ideal) < dist2(a.eight, ideal)) { return Ep(b.seven, b.eight, 1u); }\n  return Ep(a.seven, a.eight, 0u);\n}\n\n// Principal colour axis via power-iteration over precomputed, mean-corrected\n// covariance rows (the moments are accumulated for free in the pixel-load\n// loop), seeded with the bbox diagonal. Returns a unit axis, or vec4(0) for\n// a degenerate (constant) block. Same family as bc1.wgsl's principal_axis;\n// used by the fast path to seed the LSQ fit \u2014 the bbox diagonal is\n// sign-blind and points across anti-correlated data (normal maps, hue edges)\n// instead of along it.\nfn principal_axis4(\n  c0v: vec4<f32>,\n  c1v: vec4<f32>,\n  c2v: vec4<f32>,\n  c3v: vec4<f32>,\n  seed: vec4<f32>,\n) -> vec4<f32> {\n  var v = seed;\n  var len = length(v);\n  if (len < 1e-9) { return vec4<f32>(0.0); }\n  v = v / len;\n  for (var iter: u32 = 0u; iter < 8u; iter = iter + 1u) {\n    let nv = vec4<f32>(dot(c0v, v), dot(c1v, v), dot(c2v, v), dot(c3v, v));\n    len = length(nv);\n    if (len < 1e-12) { return vec4<f32>(0.0); }\n    v = nv / len;\n  }\n  return v;\n}\n\n// ============================ HIGH PATH ================================ //\n\nstruct Pair { a: vec4<i32>, b: vec4<i32> };\nfn farthest_pair(pixels: ptr<function, array<vec4<i32>, 16>>) -> Pair {\n  var best_d: i32 = 0;\n  var pa = (*pixels)[0];\n  var pb = (*pixels)[1];\n  for (var i: u32 = 0u; i < 16u; i = i + 1u) {\n    let xi = (*pixels)[i];\n    for (var j: u32 = i + 1u; j < 16u; j = j + 1u) {\n      let d = dist2(xi, (*pixels)[j]);\n      if (d > best_d) { best_d = d; pa = xi; pb = (*pixels)[j]; }\n    }\n  }\n  return Pair(pa, pb);\n}\n\nfn build_palette_6(e0: vec4<i32>, e1: vec4<i32>, pal: ptr<function, array<vec4<i32>, 16>>) {\n  for (var i: u32 = 0u; i < 16u; i = i + 1u) {\n    (*pal)[i] = interp4(e0, e1, i32(w4(i)));\n  }\n}\n\nfn assign_all(\n  pixels:  ptr<function, array<vec4<i32>, 16>>,\n  pal:     ptr<function, array<vec4<i32>, 16>>,\n  out_idx: ptr<function, array<u32, 16>>,\n) -> i32 {\n  var err: i32 = 0;\n  for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n    let px = (*pixels)[k];\n    var best_i: u32 = 0u;\n    var best_d: i32 = 2147483647;\n    for (var i: u32 = 0u; i < 16u; i = i + 1u) {\n      let d = dist2(px, (*pal)[i]);\n      if (d < best_d) { best_d = d; best_i = i; }\n    }\n    (*out_idx)[k] = best_i;\n    err = err + best_d;\n  }\n  return err;\n}\n\nstruct BestMode6 {\n  e0_7: vec4<i32>, e1_7: vec4<i32>,\n  p0: u32,         p1: u32,\n  indices: array<u32, 16>,\n  err: i32,\n};\n\n// Exhaustive p-bit search (high path); commits to `*best` only on improvement.\nfn try_pbit_combos(\n  pixels: ptr<function, array<vec4<i32>, 16>>,\n  ideal0: vec4<i32>,\n  ideal1: vec4<i32>,\n  best:   ptr<function, BestMode6>,\n) {\n  var local_best = (*best).err;\n  var pal: array<vec4<i32>, 16>;\n  var tmp: array<u32, 16>;\n  for (var p0: u32 = 0u; p0 < 2u; p0 = p0 + 1u) {\n    let q0 = quantize_endpoint(ideal0, p0);\n    for (var p1: u32 = 0u; p1 < 2u; p1 = p1 + 1u) {\n      let q1 = quantize_endpoint(ideal1, p1);\n      build_palette_6(q0.eight, q1.eight, &pal);\n      let err = assign_all(pixels, &pal, &tmp);\n      if (err < local_best) {\n        local_best = err;\n        (*best).e0_7 = q0.seven;\n        (*best).e1_7 = q1.seven;\n        (*best).p0 = p0;\n        (*best).p1 = p1;\n        (*best).indices = tmp;\n        (*best).err = err;\n      }\n    }\n  }\n}\n\n// Exact-weight LSQ refit (high path); matches bc7_ref.ts `refitEndpointsMode6`.\nstruct RefitResult { e0: vec4<i32>, e1: vec4<i32>, valid: bool };\nfn refit_endpoints(\n  pixels: ptr<function, array<vec4<i32>, 16>>,\n  indices: ptr<function, array<u32, 16>>,\n) -> RefitResult {\n  var sAA: f32 = 0.0;\n  var sBB: f32 = 0.0;\n  var sAB: f32 = 0.0;\n  var sAV: vec4<f32> = vec4<f32>(0.0);\n  var sBV: vec4<f32> = vec4<f32>(0.0);\n  for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n    let i = (*indices)[k];\n    let a = f32(64u - w4(i)) / 64.0;\n    let b = f32(w4(i)) / 64.0;\n    let v = vec4<f32>((*pixels)[k]);\n    sAA = sAA + a * a;\n    sBB = sBB + b * b;\n    sAB = sAB + a * b;\n    sAV = sAV + a * v;\n    sBV = sBV + b * v;\n  }\n  let det = sAA * sBB - sAB * sAB;\n  var out: RefitResult;\n  if (abs(det) < 1e-9) {\n    out.valid = false;\n    return out;\n  }\n  let e0f = (sBB * sAV - sAB * sBV) / det;\n  let e1f = (sAA * sBV - sAB * sAV) / det;\n  out.e0 = vec4<i32>(clamp(round(e0f), vec4<f32>(0.0), vec4<f32>(255.0)));\n  out.e1 = vec4<i32>(clamp(round(e1f), vec4<f32>(0.0), vec4<f32>(255.0)));\n  out.valid = true;\n  return out;\n}\n\n// ------------------------------- Entry --------------------------------- //\n\n@compute @workgroup_size(8, 8, 1)\nfn encode(@builtin(global_invocation_id) gid: vec3<u32>) {\n  if (gid.x >= params.blocks_x || gid.y >= params.blocks_y) {\n    return;\n  }\n\n  let bx = gid.x;\n  let by = gid.y;\n  let block_index = by * params.blocks_x + bx;\n\n  let base   = vec2<i32>(i32(bx) * 4, i32(by) * 4);\n  let max_xy = vec2<i32>(i32(params.width) - 1, i32(params.height) - 1);\n\n  // Load 16 RGBA pixels (8-bit integer domain) and the per-channel bbox,\n  // with the covariance moments FUSED in: d = px \u2212 pixel0 (first-pixel-\n  // relative, so the sums scale with the block's span; d is integer-valued\n  // and \u2264255, exact in f32). Only the fast branch consumes the moment\n  // accumulators \u2014 the QUALITY_HIGH pipeline dead-code-eliminates them.\n  var pixels: array<vec4<i32>, 16>;\n  var lo = vec4<i32>(255);\n  var hi = vec4<i32>(0);\n  var p0f = vec4<f32>(0.0);\n  var sd = vec4<f32>(0.0);\n  var c0v = vec4<f32>(0.0);\n  var c1v = vec4<f32>(0.0);\n  var c2v = vec4<f32>(0.0);\n  var c3v = vec4<f32>(0.0);\n  for (var i: u32 = 0u; i < 16u; i = i + 1u) {\n    let lx = i32(i & 3u);\n    let ly = i32(i >> 2u);\n    let p  = clamp(base + vec2<i32>(lx, ly), vec2<i32>(0, 0), max_xy);\n    let px = to8(textureLoad(src_tex, p, 0));\n    pixels[i] = px;\n    lo = min(lo, px);\n    hi = max(hi, px);\n    if (i == 0u) { p0f = vec4<f32>(px); }\n    let d = vec4<f32>(px) - p0f;\n    sd = sd + d;\n    c0v = c0v + d.x * d;\n    c1v = c1v + d.y * d;\n    c2v = c2v + d.z * d;\n    c3v = c3v + d.w * d;\n  }\n  let mean = p0f + sd * (1.0 / 16.0);\n\n  // Both branches produce: 7-bit endpoints + p-bits, and the 16 4-bit indices\n  // packed LSB-first into two nibble words (pixel k \u2192 bits 4k..4k+3).\n  var e0_7: vec4<i32>;\n  var e1_7: vec4<i32>;\n  var p0: u32;\n  var p1: u32;\n  var ilo: u32 = 0u;\n  var ihi: u32 = 0u;\n\n  if (QUALITY_HIGH != 0u) {\n    let fp = farthest_pair(&pixels);\n    var best: BestMode6;\n    best.err = 2147483647;\n    try_pbit_combos(&pixels, fp.a, fp.b, &best);\n    let refit = refit_endpoints(&pixels, &best.indices);\n    if (refit.valid) {\n      try_pbit_combos(&pixels, refit.e0, refit.e1, &best);\n    }\n    e0_7 = best.e0_7; e1_7 = best.e1_7; p0 = best.p0; p1 = best.p1;\n    for (var k: u32 = 0u; k < 8u; k = k + 1u) {\n      ilo = ilo | (best.indices[k] << (k * 4u));\n    }\n    for (var k: u32 = 8u; k < 16u; k = k + 1u) {\n      ihi = ihi | (best.indices[k] << ((k - 8u) * 4u));\n    }\n  } else {\n    // Seed endpoints from the block's principal colour axis (covariance\n    // power-iteration; bbox on degenerate blocks) at the exact projection\n    // extents, quantise, and assign indices in one projection pass. No LSQ\n    // refit: with the seed already on the principal axis, mode 6's 16-level\n    // palette leaves the refit \u22640.05 dB on the colour card and \u22640.15 dB on\n    // the normal card \u2014 not worth its two extra 16-pixel passes (the coarse\n    // 4-level BC1/ASTC fast paths DO keep theirs).\n    // Mean-correct the fused moments: C = \u03A3dd\u1D40 \u2212 (\u03A3d)(\u03A3d)\u1D40/16.\n    let sd16 = sd * (1.0 / 16.0);\n    let r0v = c0v - sd.x * sd16;\n    let r1v = c1v - sd.y * sd16;\n    let r2v = c2v - sd.z * sd16;\n    let r3v = c3v - sd.w * sd16;\n    var seed0 = lo;\n    var seed1 = hi;\n    let axis = principal_axis4(r0v, r1v, r2v, r3v, vec4<f32>(hi - lo));\n    if (dot(axis, axis) > 0.0) {\n      // Exact projection extents along the axis. (A Rayleigh-quotient span\n      // estimate was tried in place of this pass \u2014 it saves 16 dots but\n      // costs 0.1\u20130.8 dB and 4\u201310\xD7 on the worst-easy-block gate: \u03C3\n      // misjudges two-cluster and outlier blocks. The pass stays.)\n      var t_min: f32 = 1e30;\n      var t_max: f32 = -1e30;\n      for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n        let t = dot(vec4<f32>(pixels[k]) - mean, axis);\n        t_min = min(t_min, t);\n        t_max = max(t_max, t);\n      }\n      seed0 = vec4<i32>(clamp(round(mean + t_min * axis), vec4<f32>(0.0), vec4<f32>(255.0)));\n      seed1 = vec4<i32>(clamp(round(mean + t_max * axis), vec4<f32>(0.0), vec4<f32>(255.0)));\n    }\n    let ep0 = pick_ep(seed0);\n    let ep1 = pick_ep(seed1);\n    let dir = vec4<f32>(ep1.eight - ep0.eight);\n    let dd = dot(dir, dir);\n    if (dd > 0.0) {\n      let e0f = vec4<f32>(ep0.eight);\n      let inv = 15.0 / dd;\n      for (var k: u32 = 0u; k < 8u; k = k + 1u) {\n        let s = clamp(floor(dot(vec4<f32>(pixels[k]) - e0f, dir) * inv + 0.5), 0.0, 15.0);\n        ilo = ilo | (u32(s) << (k * 4u));\n      }\n      for (var k: u32 = 8u; k < 16u; k = k + 1u) {\n        let s = clamp(floor(dot(vec4<f32>(pixels[k]) - e0f, dir) * inv + 0.5), 0.0, 15.0);\n        ihi = ihi | (u32(s) << ((k - 8u) * 4u));\n      }\n    }\n    e0_7 = ep0.seven; e1_7 = ep1.seven; p0 = ep0.p; p1 = ep1.p;\n  }\n\n  // Anchor rule \u2014 pixel 0's index MSB must be 0. Swapping endpoints reflects\n  // every index (i \u2192 15\u2212i), which on packed nibbles is a bitwise NOT.\n  if ((ilo & 0x8u) != 0u) {\n    let t7 = e0_7; e0_7 = e1_7; e1_7 = t7;\n    let tp = p0;   p0   = p1;   p1   = tp;\n    ilo = ~ilo; ihi = ~ihi;\n  }\n\n  // Straight-line mode-6 packing (see layout at the top of the file).\n  let e0 = vec4<u32>(e0_7);\n  let e1 = vec4<u32>(e1_7);\n  let w0 = 0x40u | (e0.x << 7u) | (e1.x << 14u) | (e0.y << 21u) | (e1.y << 28u);\n  let w1 = (e1.y >> 4u) | (e0.z << 3u) | (e1.z << 10u) | (e0.w << 17u) | (e1.w << 24u) | (p0 << 31u);\n  let w2 = p1 | ((ilo & 0x7u) << 1u) | (ilo & 0xFFFFFFF0u);\n  let w3 = ihi;\n\n  let out = block_index * 4u;\n  dst[out + 0u] = w0;\n  dst[out + 1u] = w1;\n  dst[out + 2u] = w2;\n  dst[out + 3u] = w3;\n}\n";
 
 // src/bc7_fast_f16.wgsl
 var bc7_fast_f16_default = `// bc7 "fast" encoder \u2014 f16 variant (requires the shader-f16 feature).
-// Same algorithm family as the f32 fast path in bc7.wgsl (bbox seed \u2192
-// projection-based index assignment with a fused least-squares refit \u2192
-// reproject), tuned for throughput:
+// Same algorithm family as the f32 fast path in bc7.wgsl (principal-axis
+// seed at the exact projection extents \u2192 quantise \u2192 one projection-based
+// index-assignment pass), tuned for throughput:
 //
-//   \u2022 All projection / refit math in f16 ([0,1] domain). ~2\xD7 ALU throughput
-//     on f16-capable GPUs. The projection direction is pre-scaled by 32:
+//   \u2022 All projection math in f16 ([0,1] domain). ~2\xD7 ALU throughput on
+//     f16-capable GPUs. The projection direction is pre-scaled by 32:
 //     a shallow block (endpoints ~1/255 apart) has dd = dot(dir,dir) \u2248 1.5e-5,
 //     where 15/dd \u2248 10\u2076 overflows f16 (max 65504) to +inf and the products
-//     inside the projection dot are subnormal \u2014 the indices and the LSQ refit
-//     feeding on them turn to garbage (visible as banding on smooth
-//     gradients). Scaling dir by 32 multiplies the dots by 32 and dd by 1024;
-//     s = dot\xB7(32\xB715/dd\u2083\u2082) is the same quantity with every intermediate in
-//     f16's normal range (worst case inv = 480/0.0157 \u2248 3.0e4 < 65504).
-//   \u2022 The LSQ seed pass projects against the RAW bbox endpoints \u2014 quantising
-//     the seed first (pick_ep) costs two extra quantisation searches and
-//     doesn't measurably change where the refit lands.
+//     inside the projection dot are subnormal \u2014 the indices turn to garbage
+//     (visible as banding on smooth gradients). Scaling dir by 32 multiplies
+//     the dots by 32 and dd by 1024; s = dot\xB7(32\xB715/dd\u2083\u2082) is the same
+//     quantity with every intermediate in f16's normal range (worst case
+//     inv = 480/0.0157 \u2248 3.0e4 < 65504).
+//   \u2022 NO least-squares refit, unlike the BC1/BC5/ASTC fast paths: with the
+//     seed already on the principal axis at the exact projection extents,
+//     mode 6's fine 16-level palette leaves the refit \u22640.05 dB on the colour
+//     card and \u22640.15 dB on the normal card \u2014 not worth its two extra
+//     16-pixel passes. The coarse 4-level formats DO need it (dropping it
+//     there costs 0.5\u20131.3 dB).
 //   \u2022 Indices are packed into two u32 nibble words ON THE FLY during the
-//     final projection pass \u2014 no array<u32,16> private array. The BC7 anchor
+//     projection pass \u2014 no array<u32,16> private array. The BC7 anchor
 //     reflection (i \u2192 15\u2212i) is then just a bitwise NOT of both words.
 //   \u2022 The 128-bit block is assembled with straight-line constant shifts
 //     instead of a generic write_bits() helper (whose dynamic word indexing
@@ -686,53 +932,6 @@ fn pick_ep(ideal01: h4) -> Ep {
   return Ep(vec4<u32>(q0), e0 * h(1.0 / 255.0), 0u);
 }
 
-// One pass over the block: project every pixel onto the e0\u2192e1 line and
-// accumulate the least-squares normal-equation sums; solve for the refit
-// endpoints. Indices are NOT produced here \u2014 the caller reprojects against
-// the quantised refit endpoints anyway.
-//
-// The value sums accumulate v \u2212 e0, not v: the basis is affine (a + b = 1),
-// so fitting the shifted data and adding e0 back is the same fit, but the
-// accumulators scale with the block's span instead of its absolute level \u2014
-// on a shallow dark block, f16 rounding of absolute sums (ulp \u2248 0.12 of an
-// 8-bit level per add) drifts the refit endpoints by \xB11 level.
-struct Fit { e0: h4, e1: h4, valid: bool };
-fn proj_fit(pix: ptr<function, array<h4, 16>>, e0: h4, e1: h4) -> Fit {
-  var out: Fit;
-  out.valid = false;
-  // dir pre-scaled by 32 to keep dd and the projection dots in f16's normal
-  // range (see header). Spans below ~0.7 of an 8-bit step (dd\u2083\u2082 < 0.008,
-  // possible only for non-8-bit sources) are treated as flat \u2014 encoding them
-  // flat is under half a level of error, while running the math on them risks
-  // inv overflowing to +inf.
-  let dir = (e1 - e0) * h(32.0);
-  let dd = dot(dir, dir);
-  if (dd < h(0.008)) { return out; }
-  let inv = h(480.0) / dd; // 32\xB715/dd\u2083\u2082 \u2261 15/dd
-  var sAA = h(0.0); var sBB = h(0.0); var sAB = h(0.0);
-  var sAV = h4(0.0); var sBV = h4(0.0);
-  var s_min = h(15.0); var s_max = h(0.0);
-  for (var k: u32 = 0u; k < 16u; k = k + 1u) {
-    let vr = (*pix)[k] - e0;
-    let s = clamp(floor(dot(vr, dir) * inv + h(0.5)), h(0.0), h(15.0));
-    s_min = min(s_min, s); s_max = max(s_max, s);
-    let b = s * h(1.0 / 15.0); let a = h(1.0) - b;
-    sAA = sAA + a * a; sBB = sBB + b * b; sAB = sAB + a * b;
-    sAV = sAV + a * vr; sBV = sBV + b * vr;
-  }
-  // Rank-1 guard: if every pixel projects to ONE level the system is
-  // singular \u2014 det/numerators are pure f16 rounding noise and the solve
-  // returns garbage endpoints. With \u22652 distinct levels
-  // det = \u03A3_i<j (b_j \u2212 b_i)\xB2 \u2265 15/225 \u2248 0.067, so 0.02 is a safe floor.
-  if (s_min == s_max) { return out; }
-  let det = sAA * sBB - sAB * sAB;
-  if (abs(det) < h(0.02)) { return out; }
-  out.e0 = clamp(e0 + (sBB * sAV - sAB * sBV) / det, h4(0.0), h4(1.0));
-  out.e1 = clamp(e0 + (sAA * sBV - sAB * sAV) / det, h4(0.0), h4(1.0));
-  out.valid = true;
-  return out;
-}
-
 @compute @workgroup_size(8, 8, 1)
 fn encode(@builtin(global_invocation_id) gid: vec3<u32>) {
   if (gid.x >= params.blocks_x || gid.y >= params.blocks_y) { return; }
@@ -740,21 +939,88 @@ fn encode(@builtin(global_invocation_id) gid: vec3<u32>) {
   let base = vec2<i32>(i32(gid.x) * 4, i32(gid.y) * 4);
   let mx = vec2<i32>(i32(params.width) - 1, i32(params.height) - 1);
 
+  // Load pass, with the covariance moments FUSED in (no separate 16-pixel
+  // pass): d = (px \u2212 pixel0)\xB716, relative to the block's first pixel so the
+  // accumulators scale with the block's span \u2014 raw \u03A3v\xB7v\u1D40 moments would
+  // cancel catastrophically in f16 \u2014 and pre-scaled \xD716 so shallow blocks
+  // (span ~1/255 \u2192 d\xB2 \u2248 1e-3) clear the subnormal floor while full-range
+  // sums stay \u22644096. C = \u03A3dd\u1D40 \u2212 (\u03A3d)(\u03A3d)\u1D40/16 is the \xD7256-scaled covariance.
   var pix: array<h4, 16>;
   var lo = h4(1.0);
   var hi = h4(0.0);
+  var p0v = h4(0.0);
+  var sd = h4(0.0);
+  var c0v = h4(0.0);
+  var c1v = h4(0.0);
+  var c2v = h4(0.0);
+  var c3v = h4(0.0);
   for (var i: u32 = 0u; i < 16u; i = i + 1u) {
     let p = clamp(base + vec2<i32>(i32(i & 3u), i32(i >> 2u)), vec2<i32>(0), mx);
     let px = h4(textureLoad(src_tex, p, 0));
     pix[i] = px; lo = min(lo, px); hi = max(hi, px);
+    if (i == 0u) { p0v = px; }
+    let d = (px - p0v) * h(16.0);
+    sd = sd + d;
+    c0v = c0v + d.x * d;
+    c1v = c1v + d.y * d;
+    c2v = c2v + d.z * d;
+    c3v = c3v + d.w * d;
+  }
+  let mean = p0v + sd * h(1.0 / 256.0);
+  // Mean-correction via sd4\xB7sd4\u1D40 with sd4 = \u03A3d/4: (\u03A3d)(\u03A3d)\u1D40/16 with every
+  // product \u22644096 (a direct \u03A3d\xB7\u03A3d\u1D40 could hit 65536 and overflow f16).
+  let sd4 = sd * h(0.25);
+  c0v = c0v - sd4.x * sd4;
+  c1v = c1v - sd4.y * sd4;
+  c2v = c2v - sd4.z * sd4;
+  c3v = c3v - sd4.w * sd4;
+
+  // Seed endpoints from the block's principal colour axis (covariance
+  // power-iteration, seeded with the bbox diagonal \u2014 same family as the BC1
+  // 'high' path). The bbox diagonal is sign-blind: on anti-correlated
+  // channels (normal maps, hue edges) it points across the data instead of
+  // along it, and the LSQ refit \u2014 which fits endpoints GIVEN the projection
+  // indices \u2014 can't recover from a wrong axis. The iteration renormalises by
+  // the max component (a plain length() of the matvec output could overflow
+  // f16), so only the direction survives.
+  var seed_lo = lo;
+  var seed_hi = hi;
+  var axis = hi - lo;
+  var axis_ok = true;
+  for (var it: u32 = 0u; it < 4u; it = it + 1u) {
+    let nv = h4(dot(c0v, axis), dot(c1v, axis), dot(c2v, axis), dot(c3v, axis));
+    let m = max(max(abs(nv.x), abs(nv.y)), max(abs(nv.z), abs(nv.w)));
+    if (m < h(1e-4)) { axis_ok = false; break; }
+    axis = nv / m;
+  }
+  if (axis_ok) {
+    axis = axis / length(axis);
+    // Exact projection extents along the axis. (A Rayleigh-quotient span
+    // estimate was tried in place of this pass \u2014 it saves 16 dots but costs
+    // 0.1\u20130.8 dB and 4\u201310\xD7 on the worst-easy-block gate: \u03C3 misjudges
+    // two-cluster and outlier blocks and the quantised weight grid can't
+    // recover. The pass stays.)
+    var t_min = h(4.0);
+    var t_max = h(-4.0);
+    for (var k: u32 = 0u; k < 16u; k = k + 1u) {
+      let t = dot(pix[k] - mean, axis);
+      t_min = min(t_min, t);
+      t_max = max(t_max, t);
+    }
+    seed_lo = clamp(mean + t_min * axis, h4(0.0), h4(1.0));
+    seed_hi = clamp(mean + t_max * axis, h4(0.0), h4(1.0));
   }
 
-  // Seed fit from the raw bbox, then quantise the refit endpoints.
-  let r = proj_fit(&pix, lo, hi);
-  var ep0: Ep;
-  var ep1: Ep;
-  if (r.valid) { ep0 = pick_ep(r.e0); ep1 = pick_ep(r.e1); }
-  else         { ep0 = pick_ep(lo);   ep1 = pick_ep(hi);   }
+  // Fit from the principal-axis seed (bbox on degenerate blocks), then
+  // quantise the refit endpoints. The refit is clamped to the block bbox: on
+  // multi-cluster blocks the unconstrained solve extrapolates far outside
+  // the block's colours and the per-channel [0,1] clamp then bends the hue \u2014
+  // fringe pixels decode to colours that exist nowhere in the block.
+  // Constraining to the bbox also measures better in plain SSE (+1.3 dB on
+  // the colour test card).
+  // Quantise the PCA-extents seed directly \u2014 no LSQ refit (see header).
+  var ep0 = pick_ep(seed_lo);
+  var ep1 = pick_ep(seed_hi);
 
   // Final projection against the decoded endpoints, packing the 4-bit indices
   // into two nibble words as we go (pixel k \u2192 bits 4k..4k+3 of ilo/ihi).
@@ -825,12 +1091,12 @@ var BC7Encoder = class extends Encoder {
 };
 
 // src/astc4x4.wgsl
-var astc4x4_default = "// ASTC 4\xD74 LDR compute shader encoder.\n//\n// One invocation per 4\xD74 block. Emits 16 bytes = 4 u32s into the storage\n// buffer at `dst[block_index * 4 .. + 3]`.\n//\n// QUALITY LEVELS (pipeline-overridable constant `QUALITY_HIGH`)\n//   fast (0, default): O(N) bounding-box seed \u2192 one fused pass that projects\n//     each pixel onto the endpoint line (the 4 palette entries are colinear,\n//     so the nearest is the rounded projection \u2014 no per-entry search) while\n//     accumulating the least-squares refit sums, then a reprojection against\n//     the quantised refit endpoints with the weights packed on the fly. The\n//     endpoint ordering rule is applied before the weight pass, so no\n//     reflection is needed.\n//   high (1): O(N\xB2) farthest-pair seed, full 4-entry nearest search, one LSQ\n//     refit \u2014 matches astc4x4_ref.ts up to FP tie-breaks.\n// The fast branch is selected at pipeline-compile time; the driver eliminates\n// the unused (high) code.\n//\n// RESTRICTED SUBSET: single partition, no dual-plane, CEM 12 (LDR RGBA direct),\n// 4\xD74 weight grid with 2-bit weights (QUANT_4), 8-bit endpoints (QUANT_256).\n//\n// BLOCK LAYOUT (128 bits, LSB-first)\n//   bits [10:0]   block mode = 0x042\n//   bits [12:11]  partition count \u2212 1 = 0\n//   bits [16:13]  CEM = 12\n//   bits [80:17]  endpoints: R0 R1 G0 G1 B0 B1 A0 A1 (8-bit each)\n//   bits [127:96] 16 \xD7 2-bit weights; weight k: bit(127\u22122k)=lsb, bit(126\u22122k)=msb\n//\n// ENDPOINT ORDERING: if sum(e0.rgb) > sum(e1.rgb) swap endpoints and reflect\n// indices (w' = 3 \u2212 w) to keep the decoder out of blue contraction.\n\n// 0 = fast (default), 1 = exhaustive/high-quality. Set via pipeline constants.\noverride QUALITY_HIGH: u32 = 0u;\n\nstruct Params {\n  blocks_x: u32,\n  blocks_y: u32,\n  width:    u32,\n  height:   u32,\n};\n\n@group(0) @binding(0) var src_tex: texture_2d<f32>;\n@group(0) @binding(1) var<storage, read_write> dst: array<u32>;\n@group(0) @binding(2) var<uniform> params: Params;\n\nfn weight_unq(i: u32) -> i32 {\n  switch i {\n    case 0u: { return  0; }\n    case 1u: { return 21; }\n    case 2u: { return 43; }\n    default: { return 64; }  // case 3u\n  }\n}\n\nfn interp4(e0: vec4<i32>, e1: vec4<i32>, w: i32) -> vec4<i32> {\n  return ((64 - w) * e0 + w * e1 + vec4<i32>(32)) >> vec4<u32>(6u);\n}\n\nfn to8(v: vec4<f32>) -> vec4<i32> {\n  return vec4<i32>(clamp(floor(v * 255.0 + 0.5), vec4<f32>(0.0), vec4<f32>(255.0)));\n}\n\nfn dist2(a: vec4<i32>, b: vec4<i32>) -> i32 {\n  let d = a - b;\n  let e = d * d;\n  return e.x + e.y + e.z + e.w;\n}\n\n// ============================ FAST PATH ================================ //\n\n// One pass over the block: project every pixel onto the e0\u2192e1 line (4 levels,\n// QUANT_4 \u2248 thirds) and accumulate the least-squares normal-equation sums;\n// solve for the refit endpoints. Weights are not produced here \u2014 the caller\n// reprojects against the quantised refit endpoints anyway.\nstruct Fit { e0: vec4<i32>, e1: vec4<i32>, valid: bool };\nfn proj_fit(pixels: ptr<function, array<vec4<i32>, 16>>, e0: vec4<i32>, e1: vec4<i32>) -> Fit {\n  var out: Fit;\n  out.valid = false;\n  let dir = vec4<f32>(e1 - e0);\n  let dd = dot(dir, dir);\n  if (dd == 0.0) { return out; }\n  let e0f = vec4<f32>(e0);\n  let inv = 3.0 / dd;\n  var sAA: f32 = 0.0; var sBB: f32 = 0.0; var sAB: f32 = 0.0;\n  var sAV: vec4<f32> = vec4<f32>(0.0); var sBV: vec4<f32> = vec4<f32>(0.0);\n  var s_min = 3.0; var s_max = 0.0;\n  for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n    let v = vec4<f32>((*pixels)[k]);\n    let s = clamp(floor(dot(v - e0f, dir) * inv + 0.5), 0.0, 3.0);\n    s_min = min(s_min, s); s_max = max(s_max, s);\n    let b = s * (1.0 / 3.0); let a = 1.0 - b;\n    sAA = sAA + a * a; sBB = sBB + b * b; sAB = sAB + a * b;\n    sAV = sAV + a * v; sBV = sBV + b * v;\n  }\n  // Rank-1 guard: if every pixel projects to ONE level the system is\n  // singular \u2014 det and the numerators are pure float rounding noise and the\n  // solve returns garbage endpoints. With \u22652 levels det \u2265 15\xB7(1/3)\xB2 \u2248 1.67.\n  if (s_min == s_max) { return out; }\n  let det = sAA * sBB - sAB * sAB;\n  if (abs(det) < 1e-3) { return out; }\n  out.e0 = vec4<i32>(clamp(round((sBB * sAV - sAB * sBV) / det), vec4<f32>(0.0), vec4<f32>(255.0)));\n  out.e1 = vec4<i32>(clamp(round((sAA * sBV - sAB * sAV) / det), vec4<f32>(0.0), vec4<f32>(255.0)));\n  out.valid = true;\n  return out;\n}\n\n// ============================ HIGH PATH ================================ //\n\nstruct Pair { a: vec4<i32>, b: vec4<i32> };\nfn farthest_pair(pixels: ptr<function, array<vec4<i32>, 16>>) -> Pair {\n  var best_d: i32 = 0;\n  var pa = (*pixels)[0];\n  var pb = (*pixels)[1];\n  for (var i: u32 = 0u; i < 16u; i = i + 1u) {\n    let xi = (*pixels)[i];\n    for (var j: u32 = i + 1u; j < 16u; j = j + 1u) {\n      let d = dist2(xi, (*pixels)[j]);\n      if (d > best_d) { best_d = d; pa = xi; pb = (*pixels)[j]; }\n    }\n  }\n  return Pair(pa, pb);\n}\n\nfn build_palette(e0: vec4<i32>, e1: vec4<i32>, pal: ptr<function, array<vec4<i32>, 4>>) {\n  for (var i: u32 = 0u; i < 4u; i = i + 1u) {\n    (*pal)[i] = interp4(e0, e1, weight_unq(i));\n  }\n}\n\nfn assign_all(\n  pixels:  ptr<function, array<vec4<i32>, 16>>,\n  pal:     ptr<function, array<vec4<i32>, 4>>,\n  out_idx: ptr<function, array<u32, 16>>,\n) -> i32 {\n  var err: i32 = 0;\n  for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n    let px = (*pixels)[k];\n    var best_i: u32 = 0u;\n    var best_d: i32 = 2147483647;\n    for (var i: u32 = 0u; i < 4u; i = i + 1u) {\n      let d = dist2(px, (*pal)[i]);\n      if (d < best_d) { best_d = d; best_i = i; }\n    }\n    (*out_idx)[k] = best_i;\n    err = err + best_d;\n  }\n  return err;\n}\n\nstruct RefitResult { e0: vec4<i32>, e1: vec4<i32>, valid: bool };\nfn refit_endpoints(\n  pixels:  ptr<function, array<vec4<i32>, 16>>,\n  indices: ptr<function, array<u32, 16>>,\n) -> RefitResult {\n  var sAA: f32 = 0.0;\n  var sBB: f32 = 0.0;\n  var sAB: f32 = 0.0;\n  var sAV: vec4<f32> = vec4<f32>(0.0);\n  var sBV: vec4<f32> = vec4<f32>(0.0);\n  for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n    let unq = weight_unq((*indices)[k]);\n    let a = f32(64 - unq) / 64.0;\n    let b = f32(unq)      / 64.0;\n    let v = vec4<f32>((*pixels)[k]);\n    sAA = sAA + a * a;\n    sBB = sBB + b * b;\n    sAB = sAB + a * b;\n    sAV = sAV + a * v;\n    sBV = sBV + b * v;\n  }\n  let det = sAA * sBB - sAB * sAB;\n  var out: RefitResult;\n  if (abs(det) < 1e-9) {\n    out.valid = false;\n    return out;\n  }\n  let e0f = (sBB * sAV - sAB * sBV) / det;\n  let e1f = (sAA * sBV - sAB * sAV) / det;\n  out.e0 = vec4<i32>(clamp(round(e0f), vec4<f32>(0.0), vec4<f32>(255.0)));\n  out.e1 = vec4<i32>(clamp(round(e1f), vec4<f32>(0.0), vec4<f32>(255.0)));\n  out.valid = true;\n  return out;\n}\n\n// ------------------------------- Entry ---------------------------------- //\n\n@compute @workgroup_size(8, 8, 1)\nfn encode(@builtin(global_invocation_id) gid: vec3<u32>) {\n  if (gid.x >= params.blocks_x || gid.y >= params.blocks_y) {\n    return;\n  }\n\n  let bx = gid.x;\n  let by = gid.y;\n  let block_index = by * params.blocks_x + bx;\n\n  let base   = vec2<i32>(i32(bx) * 4, i32(by) * 4);\n  let max_xy = vec2<i32>(i32(params.width) - 1, i32(params.height) - 1);\n\n  var pixels: array<vec4<i32>, 16>;\n  var lo = vec4<i32>(255);\n  var hi = vec4<i32>(0);\n  for (var i: u32 = 0u; i < 16u; i = i + 1u) {\n    let p = clamp(base + vec2<i32>(i32(i & 3u), i32(i >> 2u)), vec2<i32>(0, 0), max_xy);\n    let px = to8(textureLoad(src_tex, p, 0));\n    pixels[i] = px;\n    lo = min(lo, px);\n    hi = max(hi, px);\n  }\n\n  // Both branches produce the final endpoints (already ordered so the\n  // decoder doesn't apply blue contraction) and the packed weight word\n  // (weight k's lsb at bit 31\u22122k, msb at bit 30\u22122k).\n  var e0: vec4<i32>;\n  var e1: vec4<i32>;\n  var w3: u32 = 0u;\n\n  if (QUALITY_HIGH != 0u) {\n    let fp = farthest_pair(&pixels);\n    e0 = fp.a;\n    e1 = fp.b;\n    var indices: array<u32, 16>;\n    var pal: array<vec4<i32>, 4>;\n    build_palette(e0, e1, &pal);\n    var err = assign_all(&pixels, &pal, &indices);\n    let refit = refit_endpoints(&pixels, &indices);\n    if (refit.valid) {\n      build_palette(refit.e0, refit.e1, &pal);\n      var idx2: array<u32, 16>;\n      let err2 = assign_all(&pixels, &pal, &idx2);\n      if (err2 < err) {\n        e0 = refit.e0;\n        e1 = refit.e1;\n        indices = idx2;\n        err = err2;\n      }\n    }\n    // Endpoint ordering, reflecting the assigned weights.\n    if (e0.x + e0.y + e0.z > e1.x + e1.y + e1.z) {\n      let tmp = e0; e0 = e1; e1 = tmp;\n      for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n        indices[k] = 3u - indices[k];\n      }\n    }\n    for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n      let w = indices[k] & 0x3u;\n      w3 = w3 | ((w & 1u) << (31u - 2u * k)) | (((w >> 1u) & 1u) << (30u - 2u * k));\n    }\n  } else {\n    // Fused LSQ fit seeded from the raw bbox, quantised refit endpoints,\n    // ordering applied BEFORE the weight pass so no reflection is needed.\n    let r = proj_fit(&pixels, lo, hi);\n    e0 = lo;\n    e1 = hi;\n    if (r.valid) { e0 = r.e0; e1 = r.e1; }\n    if (e0.x + e0.y + e0.z > e1.x + e1.y + e1.z) {\n      let tmp = e0; e0 = e1; e1 = tmp;\n    }\n    let dir = vec4<f32>(e1 - e0);\n    let dd = dot(dir, dir);\n    if (dd > 0.0) {\n      let e0f = vec4<f32>(e0);\n      let inv = 3.0 / dd;\n      for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n        let s = u32(clamp(floor(dot(vec4<f32>(pixels[k]) - e0f, dir) * inv + 0.5), 0.0, 3.0));\n        w3 = w3 | ((s & 1u) << (31u - 2u * k)) | (((s >> 1u) & 1u) << (30u - 2u * k));\n      }\n    }\n  }\n\n  // Straight-line packing: block mode 0x042 @0, partitions\u22121=0 @11, CEM 12\n  // @13, endpoints R0 R1 G0 G1 B0 B1 A0 A1 (8 bits each) from bit 17,\n  // weights in the last word.\n  let E0 = vec4<u32>(e0);\n  let E1 = vec4<u32>(e1);\n  let w0 = 0x042u | (12u << 13u) | (E0.x << 17u) | (E1.x << 25u);\n  let w1 = (E1.x >> 7u) | (E0.y << 1u) | (E1.y << 9u) | (E0.z << 17u) | (E1.z << 25u);\n  let w2 = (E1.z >> 7u) | (E0.w << 1u) | (E1.w << 9u);\n\n  let out = block_index * 4u;\n  dst[out + 0u] = w0;\n  dst[out + 1u] = w1;\n  dst[out + 2u] = w2;\n  dst[out + 3u] = w3;\n}\n";
+var astc4x4_default = "// ASTC 4\xD74 LDR compute shader encoder.\n//\n// One invocation per 4\xD74 block. Emits 16 bytes = 4 u32s into the storage\n// buffer at `dst[block_index * 4 .. + 3]`.\n//\n// QUALITY LEVELS (pipeline-overridable constant `QUALITY_HIGH`)\n//   fast (0, default): principal-axis seed (covariance power-iteration; bbox\n//     on degenerate blocks) \u2192 one fused pass that projects each pixel onto\n//     the endpoint line (the 4 palette entries are colinear, so the nearest\n//     is the rounded projection \u2014 no per-entry search) while accumulating the\n//     least-squares refit sums, then a reprojection against the quantised\n//     refit endpoints with the weights packed on the fly. The endpoint\n//     ordering rule is applied before the weight pass, so no reflection is\n//     needed.\n//   high (1): O(N\xB2) farthest-pair seed, full 4-entry nearest search, one LSQ\n//     refit \u2014 matches astc4x4_ref.ts up to FP tie-breaks.\n// The fast branch is selected at pipeline-compile time; the driver eliminates\n// the unused (high) code.\n//\n// RESTRICTED SUBSET: single partition, no dual-plane, CEM 12 (LDR RGBA direct),\n// 4\xD74 weight grid with 2-bit weights (QUANT_4), 8-bit endpoints (QUANT_256).\n//\n// BLOCK LAYOUT (128 bits, LSB-first)\n//   bits [10:0]   block mode = 0x042\n//   bits [12:11]  partition count \u2212 1 = 0\n//   bits [16:13]  CEM = 12\n//   bits [80:17]  endpoints: R0 R1 G0 G1 B0 B1 A0 A1 (8-bit each)\n//   bits [127:96] 16 \xD7 2-bit weights; weight k: bit(127\u22122k)=lsb, bit(126\u22122k)=msb\n//\n// ENDPOINT ORDERING: if sum(e0.rgb) > sum(e1.rgb) swap endpoints and reflect\n// indices (w' = 3 \u2212 w) to keep the decoder out of blue contraction.\n\n// 0 = fast (default), 1 = exhaustive/high-quality. Set via pipeline constants.\noverride QUALITY_HIGH: u32 = 0u;\n\nstruct Params {\n  blocks_x: u32,\n  blocks_y: u32,\n  width:    u32,\n  height:   u32,\n};\n\n@group(0) @binding(0) var src_tex: texture_2d<f32>;\n@group(0) @binding(1) var<storage, read_write> dst: array<u32>;\n@group(0) @binding(2) var<uniform> params: Params;\n\nfn weight_unq(i: u32) -> i32 {\n  switch i {\n    case 0u: { return  0; }\n    case 1u: { return 21; }\n    case 2u: { return 43; }\n    default: { return 64; }  // case 3u\n  }\n}\n\nfn interp4(e0: vec4<i32>, e1: vec4<i32>, w: i32) -> vec4<i32> {\n  return ((64 - w) * e0 + w * e1 + vec4<i32>(32)) >> vec4<u32>(6u);\n}\n\nfn to8(v: vec4<f32>) -> vec4<i32> {\n  return vec4<i32>(clamp(floor(v * 255.0 + 0.5), vec4<f32>(0.0), vec4<f32>(255.0)));\n}\n\nfn dist2(a: vec4<i32>, b: vec4<i32>) -> i32 {\n  let d = a - b;\n  let e = d * d;\n  return e.x + e.y + e.z + e.w;\n}\n\n// ============================ FAST PATH ================================ //\n\n// One pass over the block: project every pixel onto the e0\u2192e1 line (4 levels,\n// QUANT_4 \u2248 thirds) and accumulate the least-squares normal-equation sums;\n// solve for the refit endpoints. Weights are not produced here \u2014 the caller\n// reprojects against the quantised refit endpoints anyway.\nstruct Fit { e0: vec4<i32>, e1: vec4<i32>, valid: bool };\nfn proj_fit(pixels: ptr<function, array<vec4<i32>, 16>>, e0: vec4<i32>, e1: vec4<i32>) -> Fit {\n  var out: Fit;\n  out.valid = false;\n  let dir = vec4<f32>(e1 - e0);\n  let dd = dot(dir, dir);\n  if (dd == 0.0) { return out; }\n  let e0f = vec4<f32>(e0);\n  let inv = 3.0 / dd;\n  var sAA: f32 = 0.0; var sBB: f32 = 0.0; var sAB: f32 = 0.0;\n  var sAV: vec4<f32> = vec4<f32>(0.0); var sBV: vec4<f32> = vec4<f32>(0.0);\n  var s_min = 3.0; var s_max = 0.0;\n  for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n    let v = vec4<f32>((*pixels)[k]);\n    let s = clamp(floor(dot(v - e0f, dir) * inv + 0.5), 0.0, 3.0);\n    s_min = min(s_min, s); s_max = max(s_max, s);\n    let b = s * (1.0 / 3.0); let a = 1.0 - b;\n    sAA = sAA + a * a; sBB = sBB + b * b; sAB = sAB + a * b;\n    sAV = sAV + a * v; sBV = sBV + b * v;\n  }\n  // Rank-1 guard: if every pixel projects to ONE level the system is\n  // singular \u2014 det and the numerators are pure float rounding noise and the\n  // solve returns garbage endpoints. With \u22652 levels det \u2265 15\xB7(1/3)\xB2 \u2248 1.67.\n  if (s_min == s_max) { return out; }\n  let det = sAA * sBB - sAB * sAB;\n  if (abs(det) < 1e-3) { return out; }\n  out.e0 = vec4<i32>(clamp(round((sBB * sAV - sAB * sBV) / det), vec4<f32>(0.0), vec4<f32>(255.0)));\n  out.e1 = vec4<i32>(clamp(round((sAA * sBV - sAB * sAV) / det), vec4<f32>(0.0), vec4<f32>(255.0)));\n  out.valid = true;\n  return out;\n}\n\n// Principal colour axis via covariance power-iteration (RGBA, 8-bit integer\n// pixel domain), seeded with the bbox diagonal. Returns a unit axis, or\n// vec4(0) for a degenerate (constant) block. Used by the fast path to seed\n// the LSQ fit \u2014 the bbox diagonal is sign-blind and points across\n// anti-correlated data (normal maps, hue edges) instead of along it.\nfn principal_axis4(\n  pixels: ptr<function, array<vec4<i32>, 16>>,\n  mean: vec4<f32>,\n  seed: vec4<f32>,\n) -> vec4<f32> {\n  var c0v = vec4<f32>(0.0);\n  var c1v = vec4<f32>(0.0);\n  var c2v = vec4<f32>(0.0);\n  var c3v = vec4<f32>(0.0);\n  for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n    let d = vec4<f32>((*pixels)[k]) - mean;\n    c0v = c0v + d.x * d;\n    c1v = c1v + d.y * d;\n    c2v = c2v + d.z * d;\n    c3v = c3v + d.w * d;\n  }\n  var v = seed;\n  var len = length(v);\n  if (len < 1e-9) { return vec4<f32>(0.0); }\n  v = v / len;\n  for (var iter: u32 = 0u; iter < 8u; iter = iter + 1u) {\n    let nv = vec4<f32>(dot(c0v, v), dot(c1v, v), dot(c2v, v), dot(c3v, v));\n    len = length(nv);\n    if (len < 1e-12) { return vec4<f32>(0.0); }\n    v = nv / len;\n  }\n  return v;\n}\n\n// ============================ HIGH PATH ================================ //\n\nstruct Pair { a: vec4<i32>, b: vec4<i32> };\nfn farthest_pair(pixels: ptr<function, array<vec4<i32>, 16>>) -> Pair {\n  var best_d: i32 = 0;\n  var pa = (*pixels)[0];\n  var pb = (*pixels)[1];\n  for (var i: u32 = 0u; i < 16u; i = i + 1u) {\n    let xi = (*pixels)[i];\n    for (var j: u32 = i + 1u; j < 16u; j = j + 1u) {\n      let d = dist2(xi, (*pixels)[j]);\n      if (d > best_d) { best_d = d; pa = xi; pb = (*pixels)[j]; }\n    }\n  }\n  return Pair(pa, pb);\n}\n\nfn build_palette(e0: vec4<i32>, e1: vec4<i32>, pal: ptr<function, array<vec4<i32>, 4>>) {\n  for (var i: u32 = 0u; i < 4u; i = i + 1u) {\n    (*pal)[i] = interp4(e0, e1, weight_unq(i));\n  }\n}\n\nfn assign_all(\n  pixels:  ptr<function, array<vec4<i32>, 16>>,\n  pal:     ptr<function, array<vec4<i32>, 4>>,\n  out_idx: ptr<function, array<u32, 16>>,\n) -> i32 {\n  var err: i32 = 0;\n  for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n    let px = (*pixels)[k];\n    var best_i: u32 = 0u;\n    var best_d: i32 = 2147483647;\n    for (var i: u32 = 0u; i < 4u; i = i + 1u) {\n      let d = dist2(px, (*pal)[i]);\n      if (d < best_d) { best_d = d; best_i = i; }\n    }\n    (*out_idx)[k] = best_i;\n    err = err + best_d;\n  }\n  return err;\n}\n\nstruct RefitResult { e0: vec4<i32>, e1: vec4<i32>, valid: bool };\nfn refit_endpoints(\n  pixels:  ptr<function, array<vec4<i32>, 16>>,\n  indices: ptr<function, array<u32, 16>>,\n) -> RefitResult {\n  var sAA: f32 = 0.0;\n  var sBB: f32 = 0.0;\n  var sAB: f32 = 0.0;\n  var sAV: vec4<f32> = vec4<f32>(0.0);\n  var sBV: vec4<f32> = vec4<f32>(0.0);\n  for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n    let unq = weight_unq((*indices)[k]);\n    let a = f32(64 - unq) / 64.0;\n    let b = f32(unq)      / 64.0;\n    let v = vec4<f32>((*pixels)[k]);\n    sAA = sAA + a * a;\n    sBB = sBB + b * b;\n    sAB = sAB + a * b;\n    sAV = sAV + a * v;\n    sBV = sBV + b * v;\n  }\n  let det = sAA * sBB - sAB * sAB;\n  var out: RefitResult;\n  if (abs(det) < 1e-9) {\n    out.valid = false;\n    return out;\n  }\n  let e0f = (sBB * sAV - sAB * sBV) / det;\n  let e1f = (sAA * sBV - sAB * sAV) / det;\n  out.e0 = vec4<i32>(clamp(round(e0f), vec4<f32>(0.0), vec4<f32>(255.0)));\n  out.e1 = vec4<i32>(clamp(round(e1f), vec4<f32>(0.0), vec4<f32>(255.0)));\n  out.valid = true;\n  return out;\n}\n\n// ------------------------------- Entry ---------------------------------- //\n\n@compute @workgroup_size(8, 8, 1)\nfn encode(@builtin(global_invocation_id) gid: vec3<u32>) {\n  if (gid.x >= params.blocks_x || gid.y >= params.blocks_y) {\n    return;\n  }\n\n  let bx = gid.x;\n  let by = gid.y;\n  let block_index = by * params.blocks_x + bx;\n\n  let base   = vec2<i32>(i32(bx) * 4, i32(by) * 4);\n  let max_xy = vec2<i32>(i32(params.width) - 1, i32(params.height) - 1);\n\n  var pixels: array<vec4<i32>, 16>;\n  var lo = vec4<i32>(255);\n  var hi = vec4<i32>(0);\n  var isum = vec4<i32>(0);\n  for (var i: u32 = 0u; i < 16u; i = i + 1u) {\n    let p = clamp(base + vec2<i32>(i32(i & 3u), i32(i >> 2u)), vec2<i32>(0, 0), max_xy);\n    let px = to8(textureLoad(src_tex, p, 0));\n    pixels[i] = px;\n    lo = min(lo, px);\n    hi = max(hi, px);\n    isum = isum + px;\n  }\n  let mean = vec4<f32>(isum) * (1.0 / 16.0);\n\n  // Both branches produce the final endpoints (already ordered so the\n  // decoder doesn't apply blue contraction) and the packed weight word\n  // (weight k's lsb at bit 31\u22122k, msb at bit 30\u22122k).\n  var e0: vec4<i32>;\n  var e1: vec4<i32>;\n  var w3: u32 = 0u;\n\n  if (QUALITY_HIGH != 0u) {\n    let fp = farthest_pair(&pixels);\n    e0 = fp.a;\n    e1 = fp.b;\n    var indices: array<u32, 16>;\n    var pal: array<vec4<i32>, 4>;\n    build_palette(e0, e1, &pal);\n    var err = assign_all(&pixels, &pal, &indices);\n    let refit = refit_endpoints(&pixels, &indices);\n    if (refit.valid) {\n      build_palette(refit.e0, refit.e1, &pal);\n      var idx2: array<u32, 16>;\n      let err2 = assign_all(&pixels, &pal, &idx2);\n      if (err2 < err) {\n        e0 = refit.e0;\n        e1 = refit.e1;\n        indices = idx2;\n        err = err2;\n      }\n    }\n    // Endpoint ordering, reflecting the assigned weights.\n    if (e0.x + e0.y + e0.z > e1.x + e1.y + e1.z) {\n      let tmp = e0; e0 = e1; e1 = tmp;\n      for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n        indices[k] = 3u - indices[k];\n      }\n    }\n    for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n      let w = indices[k] & 0x3u;\n      w3 = w3 | ((w & 1u) << (31u - 2u * k)) | (((w >> 1u) & 1u) << (30u - 2u * k));\n    }\n  } else {\n    // Fused LSQ fit seeded from the block's principal colour axis\n    // (covariance power-iteration; bbox on degenerate blocks), quantised\n    // refit endpoints, ordering applied BEFORE the weight pass so no\n    // reflection is needed.\n    // The refit is clamped to the block bbox: on multi-cluster blocks the\n    // unconstrained solve extrapolates far outside the block's colours and the\n    // per-channel [0,255] clamp then bends the hue \u2014 fringe pixels decode to\n    // colours that exist nowhere in the block. Constraining to the bbox also\n    // measures better in plain SSE (+1.8 dB on the colour test card).\n    var seed0 = lo;\n    var seed1 = hi;\n    let axis = principal_axis4(&pixels, mean, vec4<f32>(hi - lo));\n    if (dot(axis, axis) > 0.0) {\n      var t_min: f32 = 1e30;\n      var t_max: f32 = -1e30;\n      for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n        let t = dot(vec4<f32>(pixels[k]) - mean, axis);\n        t_min = min(t_min, t);\n        t_max = max(t_max, t);\n      }\n      seed0 = vec4<i32>(clamp(round(mean + t_min * axis), vec4<f32>(0.0), vec4<f32>(255.0)));\n      seed1 = vec4<i32>(clamp(round(mean + t_max * axis), vec4<f32>(0.0), vec4<f32>(255.0)));\n    }\n    let r = proj_fit(&pixels, seed0, seed1);\n    e0 = lo;\n    e1 = hi;\n    if (r.valid) { e0 = clamp(r.e0, lo, hi); e1 = clamp(r.e1, lo, hi); }\n    if (e0.x + e0.y + e0.z > e1.x + e1.y + e1.z) {\n      let tmp = e0; e0 = e1; e1 = tmp;\n    }\n    let dir = vec4<f32>(e1 - e0);\n    let dd = dot(dir, dir);\n    if (dd > 0.0) {\n      let e0f = vec4<f32>(e0);\n      let inv = 3.0 / dd;\n      for (var k: u32 = 0u; k < 16u; k = k + 1u) {\n        let s = u32(clamp(floor(dot(vec4<f32>(pixels[k]) - e0f, dir) * inv + 0.5), 0.0, 3.0));\n        w3 = w3 | ((s & 1u) << (31u - 2u * k)) | (((s >> 1u) & 1u) << (30u - 2u * k));\n      }\n    }\n  }\n\n  // Straight-line packing: block mode 0x042 @0, partitions\u22121=0 @11, CEM 12\n  // @13, endpoints R0 R1 G0 G1 B0 B1 A0 A1 (8 bits each) from bit 17,\n  // weights in the last word.\n  let E0 = vec4<u32>(e0);\n  let E1 = vec4<u32>(e1);\n  let w0 = 0x042u | (12u << 13u) | (E0.x << 17u) | (E1.x << 25u);\n  let w1 = (E1.x >> 7u) | (E0.y << 1u) | (E1.y << 9u) | (E0.z << 17u) | (E1.z << 25u);\n  let w2 = (E1.z >> 7u) | (E0.w << 1u) | (E1.w << 9u);\n\n  let out = block_index * 4u;\n  dst[out + 0u] = w0;\n  dst[out + 1u] = w1;\n  dst[out + 2u] = w2;\n  dst[out + 3u] = w3;\n}\n";
 
 // src/astc4x4_fast_f16.wgsl
 var astc4x4_fast_f16_default = `// astc4x4 "fast" encoder \u2014 f16 variant (requires the shader-f16 feature).
-// Same algorithm family as the f32 fast path in astc4x4.wgsl (bbox seed \u2192
-// projection weight assignment with a fused least-squares refit \u2192
+// Same algorithm family as the f32 fast path in astc4x4.wgsl (principal-axis
+// seed \u2192 projection weight assignment with a fused least-squares refit \u2192
 // reproject), tuned for throughput:
 //
 //   \u2022 All projection / refit math in f16 ([0,1] domain). The projection
@@ -915,16 +1181,68 @@ fn encode(@builtin(global_invocation_id) gid: vec3<u32>) {
   var pix: array<h4, 16>;
   var lo = h4(1.0);
   var hi = h4(0.0);
+  var mean = h4(0.0);
   for (var i: u32 = 0u; i < 16u; i = i + 1u) {
     let p = clamp(base + vec2<i32>(i32(i & 3u), i32(i >> 2u)), vec2<i32>(0), mx);
     let px = h4(textureLoad(src_tex, p, 0));
     pix[i] = px; lo = min(lo, px); hi = max(hi, px);
+    mean = mean + px;
+  }
+  mean = mean * h(1.0 / 16.0);
+
+  // Seed endpoints from the block's principal colour axis (covariance
+  // power-iteration, seeded with the bbox diagonal). The bbox diagonal is
+  // sign-blind: on anti-correlated channels (normal maps, hue edges) it
+  // points across the data instead of along it, and the LSQ refit \u2014 which
+  // fits endpoints GIVEN the projection weights \u2014 can't recover from a wrong
+  // axis. Deviations are pre-scaled \xD716 so covariance entries for shallow
+  // blocks stay in f16's normal range (span ~1/255 \u2192 d\xB2 \u2248 1e-3) while
+  // full-range sums stay \u22644096; the iteration renormalises by the max
+  // component (a plain length() of the matvec output could overflow f16), so
+  // only the direction survives.
+  var seed_lo = lo;
+  var seed_hi = hi;
+  var c0v = h4(0.0);
+  var c1v = h4(0.0);
+  var c2v = h4(0.0);
+  var c3v = h4(0.0);
+  for (var k: u32 = 0u; k < 16u; k = k + 1u) {
+    let d = (pix[k] - mean) * h(16.0);
+    c0v = c0v + d.x * d;
+    c1v = c1v + d.y * d;
+    c2v = c2v + d.z * d;
+    c3v = c3v + d.w * d;
+  }
+  var axis = hi - lo;
+  var axis_ok = true;
+  for (var it: u32 = 0u; it < 4u; it = it + 1u) {
+    let nv = h4(dot(c0v, axis), dot(c1v, axis), dot(c2v, axis), dot(c3v, axis));
+    let m = max(max(abs(nv.x), abs(nv.y)), max(abs(nv.z), abs(nv.w)));
+    if (m < h(1e-4)) { axis_ok = false; break; }
+    axis = nv / m;
+  }
+  if (axis_ok) {
+    axis = axis / length(axis);
+    var t_min = h(4.0);
+    var t_max = h(-4.0);
+    for (var k: u32 = 0u; k < 16u; k = k + 1u) {
+      let t = dot(pix[k] - mean, axis);
+      t_min = min(t_min, t);
+      t_max = max(t_max, t);
+    }
+    seed_lo = clamp(mean + t_min * axis, h4(0.0), h4(1.0));
+    seed_hi = clamp(mean + t_max * axis, h4(0.0), h4(1.0));
   }
 
-  let r = proj_fit(&pix, lo, hi);
+  // The refit is clamped to the block bbox: on multi-cluster blocks the
+  // unconstrained solve extrapolates far outside the block's colours and the
+  // per-channel [0,1] clamp then bends the hue \u2014 fringe pixels decode to
+  // colours that exist nowhere in the block. Constraining to the bbox also
+  // measures better in plain SSE (+1.8 dB on the colour test card).
+  let r = proj_fit(&pix, seed_lo, seed_hi);
   var e0 = lo;
   var e1 = hi;
-  if (r.valid) { e0 = r.e0; e1 = r.e1; }
+  if (r.valid) { e0 = clamp(r.e0, lo, hi); e1 = clamp(r.e1, lo, hi); }
   var E0 = q8(e0);
   var E1 = q8(e1);
 
@@ -1200,7 +1518,7 @@ var WebGLBlockEncoder = class {
 };
 
 // src/webgl/glsl/bc1.frag.glsl
-var bc1_frag_default = "#version 300 es\n// BC1 (DXT1) fragment-shader encoder \u2014 WebGL2 port of bc1.wgsl (fast path).\n//\n// One fragment per 4\xD74 block. Output is the 8-byte BC1 block as 2 \xD7 u32 in\n// outColor.rg (outColor.ba unused); the encoder reads back RGBA32UI and keeps\n// the low two words per block. This is the *fast* path only (the WGSL\n// `QUALITY_HIGH == 0` branch): bbox endpoints, 1/16 inset, RGB565 quantisation,\n// forced 4-colour mode, full 4-entry L2 index search, then a single\n// least-squares endpoint refit accepted only when it lowers the block's error.\n// See bc1.wgsl for the full derivation.\n\nprecision highp float;\nprecision highp int;\n\nuniform sampler2D uSrc;\nuniform ivec2 uSrcSize; // original (unpadded) width, height\nuniform int uFlipY;     // 1 = sample bottom-up (matches Three.js flipY)\n\nlayout(location = 0) out uvec4 outColor;\n\n// 4-colour-mode interpolation weights: pal[j] = WA[j]*c0 + WB[j]*c1.\nconst float WA[4] = float[4](1.0, 0.0, 2.0 / 3.0, 1.0 / 3.0);\nconst float WB[4] = float[4](0.0, 1.0, 1.0 / 3.0, 2.0 / 3.0);\n\nuint to565(vec3 c) {\n  uint r = uint(clamp(floor(c.r * 31.0 + 0.5), 0.0, 31.0));\n  uint g = uint(clamp(floor(c.g * 63.0 + 0.5), 0.0, 63.0));\n  uint b = uint(clamp(floor(c.b * 31.0 + 0.5), 0.0, 31.0));\n  return (r << 11) | (g << 5) | b;\n}\n\nvec3 from565(uint c) {\n  float r = float((c >> 11) & 31u);\n  float g = float((c >> 5) & 63u);\n  float b = float(c & 31u);\n  // 5/6-bit \u2192 8-bit. floor((x*527+23)/64) == (x<<3)|(x>>2): exact hardware\n  // bit-replication (white \u2192 255), so index selection matches the GPU decode.\n  float r8 = floor((r * 527.0 + 23.0) / 64.0);\n  float g8 = floor((g * 259.0 + 33.0) / 64.0);\n  float b8 = floor((b * 527.0 + 23.0) / 64.0);\n  return vec3(r8, g8, b8) / 255.0;\n}\n\nvoid main() {\n  ivec2 base = ivec2(gl_FragCoord.xy) * 4;\n  ivec2 maxXY = uSrcSize - ivec2(1);\n\n  vec3 pixels[16];\n  vec3 bbMin = vec3(1.0);\n  vec3 bbMax = vec3(0.0);\n  for (int i = 0; i < 16; i++) {\n    ivec2 p = clamp(base + ivec2(i & 3, i >> 2), ivec2(0), maxXY);\n    int sy = (uFlipY != 0) ? (uSrcSize.y - 1 - p.y) : p.y;\n    vec3 c = texelFetch(uSrc, ivec2(p.x, sy), 0).rgb;\n    pixels[i] = c;\n    bbMin = min(bbMin, c);\n    bbMax = max(bbMax, c);\n  }\n\n  // Inset the bbox by ~half an RGB565 cell (1/16) to tighten the quantised\n  // 4-colour palette around the real data range.\n  vec3 inset = (bbMax - bbMin) / 16.0;\n  vec3 hi = clamp(bbMax - inset, vec3(0.0), vec3(1.0));\n  vec3 lo = clamp(bbMin + inset, vec3(0.0), vec3(1.0));\n\n  uint c0 = to565(hi);\n  uint c1 = to565(lo);\n  // 4-colour mode requires color0 > color1.\n  if (c0 == c1) {\n    if (c1 > 0u) { c1 = c1 - 1u; } else { c0 = c0 + 1u; }\n  } else if (c0 < c1) {\n    uint tmp = c0; c0 = c1; c1 = tmp;\n  }\n\n  // Build the palette in decoded space, assign each pixel its nearest entry.\n  vec3 pal[4];\n  vec3 p0 = from565(c0);\n  vec3 p1 = from565(c1);\n  for (int j = 0; j < 4; j++) pal[j] = WA[j] * p0 + WB[j] * p1;\n\n  uint idx[16];\n  float err = 0.0;\n  for (int k = 0; k < 16; k++) {\n    vec3 c = pixels[k];\n    uint bestJ = 0u;\n    float bestD = 1e30;\n    for (int j = 0; j < 4; j++) {\n      vec3 d = pal[j] - c;\n      float d2 = dot(d, d);\n      if (d2 < bestD) { bestD = d2; bestJ = uint(j); }\n    }\n    idx[k] = bestJ;\n    err += bestD;\n  }\n\n  // One least-squares refit: re-solve the endpoints for the current indices,\n  // re-quantise, re-assign; keep it only if the squared error drops.\n  float sAA = 0.0, sBB = 0.0, sAB = 0.0;\n  vec3 sAV = vec3(0.0), sBV = vec3(0.0);\n  for (int k = 0; k < 16; k++) {\n    float a = WA[int(idx[k])];\n    float b = WB[int(idx[k])];\n    vec3 v = pixels[k];\n    sAA += a * a; sBB += b * b; sAB += a * b; sAV += a * v; sBV += b * v;\n  }\n  float det = sAA * sBB - sAB * sAB;\n  if (abs(det) > 1e-9) {\n    vec3 e0 = clamp((sBB * sAV - sAB * sBV) / det, vec3(0.0), vec3(1.0));\n    vec3 e1 = clamp((sAA * sBV - sAB * sAV) / det, vec3(0.0), vec3(1.0));\n    uint nc0 = to565(e0);\n    uint nc1 = to565(e1);\n    if (nc0 < nc1) { uint t = nc0; nc0 = nc1; nc1 = t; }\n    if (nc0 != nc1 && !(nc0 == c0 && nc1 == c1)) {\n      vec3 q0 = from565(nc0);\n      vec3 q1 = from565(nc1);\n      vec3 pal2[4];\n      for (int j = 0; j < 4; j++) pal2[j] = WA[j] * q0 + WB[j] * q1;\n      uint idx2[16];\n      float nerr = 0.0;\n      for (int k = 0; k < 16; k++) {\n        vec3 c = pixels[k];\n        uint bestJ = 0u;\n        float bestD = 1e30;\n        for (int j = 0; j < 4; j++) {\n          vec3 d = pal2[j] - c;\n          float d2 = dot(d, d);\n          if (d2 < bestD) { bestD = d2; bestJ = uint(j); }\n        }\n        idx2[k] = bestJ;\n        nerr += bestD;\n      }\n      if (nerr < err) {\n        c0 = nc0; c1 = nc1;\n        for (int k = 0; k < 16; k++) idx[k] = idx2[k];\n      }\n    }\n  }\n\n  uint indices = 0u;\n  for (int k = 0; k < 16; k++) indices |= (idx[k] & 3u) << (uint(k) * 2u);\n\n  outColor = uvec4(c0 | (c1 << 16), indices, 0u, 0u);\n}\n";
+var bc1_frag_default = "#version 300 es\n// BC1 (DXT1) fragment-shader encoder \u2014 WebGL2 port of bc1.wgsl (fast path).\n//\n// One fragment per 4\xD74 block. Output is the 8-byte BC1 block as 2 \xD7 u32 in\n// outColor.rg (outColor.ba unused); the encoder reads back RGBA32UI and keeps\n// the low two words per block. This is the *fast* path only (the WGSL\n// `QUALITY_HIGH == 0` branch): principal-axis endpoint seed (covariance\n// power-iteration; inset bbox on degenerate blocks), RGB565 quantisation,\n// forced 4-colour mode, full 4-entry L2 index search, then a single\n// least-squares endpoint refit accepted only when it lowers the block's error.\n// See bc1.wgsl for the full derivation.\n\nprecision highp float;\nprecision highp int;\n\nuniform sampler2D uSrc;\nuniform ivec2 uSrcSize; // original (unpadded) width, height\nuniform int uFlipY;     // 1 = sample bottom-up (matches Three.js flipY)\n\nlayout(location = 0) out uvec4 outColor;\n\n// 4-colour-mode interpolation weights: pal[j] = WA[j]*c0 + WB[j]*c1.\nconst float WA[4] = float[4](1.0, 0.0, 2.0 / 3.0, 1.0 / 3.0);\nconst float WB[4] = float[4](0.0, 1.0, 1.0 / 3.0, 2.0 / 3.0);\n\nuint to565(vec3 c) {\n  uint r = uint(clamp(floor(c.r * 31.0 + 0.5), 0.0, 31.0));\n  uint g = uint(clamp(floor(c.g * 63.0 + 0.5), 0.0, 63.0));\n  uint b = uint(clamp(floor(c.b * 31.0 + 0.5), 0.0, 31.0));\n  return (r << 11) | (g << 5) | b;\n}\n\nvec3 from565(uint c) {\n  float r = float((c >> 11) & 31u);\n  float g = float((c >> 5) & 63u);\n  float b = float(c & 31u);\n  // 5/6-bit \u2192 8-bit. floor((x*527+23)/64) == (x<<3)|(x>>2): exact hardware\n  // bit-replication (white \u2192 255), so index selection matches the GPU decode.\n  float r8 = floor((r * 527.0 + 23.0) / 64.0);\n  float g8 = floor((g * 259.0 + 33.0) / 64.0);\n  float b8 = floor((b * 527.0 + 23.0) / 64.0);\n  return vec3(r8, g8, b8) / 255.0;\n}\n\nvoid main() {\n  ivec2 base = ivec2(gl_FragCoord.xy) * 4;\n  ivec2 maxXY = uSrcSize - ivec2(1);\n\n  vec3 pixels[16];\n  vec3 bbMin = vec3(1.0);\n  vec3 bbMax = vec3(0.0);\n  vec3 mean = vec3(0.0);\n  for (int i = 0; i < 16; i++) {\n    ivec2 p = clamp(base + ivec2(i & 3, i >> 2), ivec2(0), maxXY);\n    int sy = (uFlipY != 0) ? (uSrcSize.y - 1 - p.y) : p.y;\n    vec3 c = texelFetch(uSrc, ivec2(p.x, sy), 0).rgb;\n    pixels[i] = c;\n    bbMin = min(bbMin, c);\n    bbMax = max(bbMax, c);\n    mean += c;\n  }\n  mean /= 16.0;\n\n  // Seed endpoints from the block's principal colour axis (covariance\n  // power-iteration, seeded with the bbox diagonal \u2014 mirrors bc1.wgsl). The\n  // bbox diagonal is sign-blind: on anti-correlated channels (normal maps,\n  // hue edges) it points across the data instead of along it, and the LSQ\n  // refit below \u2014 which fits endpoints GIVEN the indices \u2014 can't recover.\n  // Degenerate (near-flat) blocks keep the inset-bbox seed. Both seeds inset\n  // by ~half an RGB565 cell (1/16) to tighten the quantised palette.\n  vec3 c0v = vec3(0.0);\n  vec3 c1v = vec3(0.0);\n  vec3 c2v = vec3(0.0);\n  for (int k = 0; k < 16; k++) {\n    vec3 d = pixels[k] - mean;\n    c0v += d.x * d;\n    c1v += d.y * d;\n    c2v += d.z * d;\n  }\n  vec3 hi;\n  vec3 lo;\n  vec3 axis = bbMax - bbMin;\n  float alen = length(axis);\n  bool axisOk = alen > 1e-9;\n  if (axisOk) {\n    axis /= alen;\n    for (int it = 0; it < 8; it++) {\n      vec3 nv = vec3(dot(c0v, axis), dot(c1v, axis), dot(c2v, axis));\n      float nlen = length(nv);\n      if (nlen < 1e-12) { axisOk = false; break; }\n      axis = nv / nlen;\n    }\n  }\n  if (axisOk) {\n    float tMin = 1e30;\n    float tMax = -1e30;\n    for (int k = 0; k < 16; k++) {\n      float t = dot(pixels[k] - mean, axis);\n      tMin = min(tMin, t);\n      tMax = max(tMax, t);\n    }\n    float pad = (tMax - tMin) / 16.0;\n    hi = clamp(mean + (tMax - pad) * axis, vec3(0.0), vec3(1.0));\n    lo = clamp(mean + (tMin + pad) * axis, vec3(0.0), vec3(1.0));\n  } else {\n    vec3 inset = (bbMax - bbMin) / 16.0;\n    hi = clamp(bbMax - inset, vec3(0.0), vec3(1.0));\n    lo = clamp(bbMin + inset, vec3(0.0), vec3(1.0));\n  }\n\n  uint c0 = to565(hi);\n  uint c1 = to565(lo);\n  // 4-colour mode requires color0 > color1.\n  if (c0 == c1) {\n    if (c1 > 0u) { c1 = c1 - 1u; } else { c0 = c0 + 1u; }\n  } else if (c0 < c1) {\n    uint tmp = c0; c0 = c1; c1 = tmp;\n  }\n\n  // Build the palette in decoded space, assign each pixel its nearest entry.\n  vec3 pal[4];\n  vec3 p0 = from565(c0);\n  vec3 p1 = from565(c1);\n  for (int j = 0; j < 4; j++) pal[j] = WA[j] * p0 + WB[j] * p1;\n\n  uint idx[16];\n  float err = 0.0;\n  for (int k = 0; k < 16; k++) {\n    vec3 c = pixels[k];\n    uint bestJ = 0u;\n    float bestD = 1e30;\n    for (int j = 0; j < 4; j++) {\n      vec3 d = pal[j] - c;\n      float d2 = dot(d, d);\n      if (d2 < bestD) { bestD = d2; bestJ = uint(j); }\n    }\n    idx[k] = bestJ;\n    err += bestD;\n  }\n\n  // One least-squares refit: re-solve the endpoints for the current indices,\n  // re-quantise, re-assign; keep it only if the squared error drops.\n  float sAA = 0.0, sBB = 0.0, sAB = 0.0;\n  vec3 sAV = vec3(0.0), sBV = vec3(0.0);\n  for (int k = 0; k < 16; k++) {\n    float a = WA[int(idx[k])];\n    float b = WB[int(idx[k])];\n    vec3 v = pixels[k];\n    sAA += a * a; sBB += b * b; sAB += a * b; sAV += a * v; sBV += b * v;\n  }\n  float det = sAA * sBB - sAB * sAB;\n  if (abs(det) > 1e-9) {\n    // Clamp the refit to the block bbox (not [0,1]): on multi-cluster blocks\n    // the unconstrained LSQ solve extrapolates far outside the block's colours\n    // and the per-channel clamp then bends the hue \u2014 fringe pixels decode to\n    // colours that exist nowhere in the block. Constraining to the bbox also\n    // measures better in plain SSE (+1.6 dB on the colour test card), so the\n    // accept-if-better guard below keeps more refits.\n    vec3 e0 = clamp((sBB * sAV - sAB * sBV) / det, bbMin, bbMax);\n    vec3 e1 = clamp((sAA * sBV - sAB * sAV) / det, bbMin, bbMax);\n    uint nc0 = to565(e0);\n    uint nc1 = to565(e1);\n    if (nc0 < nc1) { uint t = nc0; nc0 = nc1; nc1 = t; }\n    if (nc0 != nc1 && !(nc0 == c0 && nc1 == c1)) {\n      vec3 q0 = from565(nc0);\n      vec3 q1 = from565(nc1);\n      vec3 pal2[4];\n      for (int j = 0; j < 4; j++) pal2[j] = WA[j] * q0 + WB[j] * q1;\n      uint idx2[16];\n      float nerr = 0.0;\n      for (int k = 0; k < 16; k++) {\n        vec3 c = pixels[k];\n        uint bestJ = 0u;\n        float bestD = 1e30;\n        for (int j = 0; j < 4; j++) {\n          vec3 d = pal2[j] - c;\n          float d2 = dot(d, d);\n          if (d2 < bestD) { bestD = d2; bestJ = uint(j); }\n        }\n        idx2[k] = bestJ;\n        nerr += bestD;\n      }\n      if (nerr < err) {\n        c0 = nc0; c1 = nc1;\n        for (int k = 0; k < 16; k++) idx[k] = idx2[k];\n      }\n    }\n  }\n\n  uint indices = 0u;\n  for (int k = 0; k < 16; k++) indices |= (idx[k] & 3u) << (uint(k) * 2u);\n\n  outColor = uvec4(c0 | (c1 << 16), indices, 0u, 0u);\n}\n";
 
 // src/webgl/BC1WebGLEncoder.ts
 var BC1WebGLEncoder = class extends WebGLBlockEncoder {
@@ -1219,7 +1537,7 @@ var BC1WebGLEncoder = class extends WebGLBlockEncoder {
 };
 
 // src/webgl/glsl/bc5.frag.glsl
-var bc5_frag_default = "#version 300 es\n// BC5 (RGTC2) fragment-shader encoder \u2014 WebGL2 port of bc5.wgsl (fast path).\n//\n// One fragment per 4\xD74 block \u2192 16-byte BC5 block as 4 \xD7 u32 in outColor.\n// BC5 = two BC4 halves (R then G). This is the *fast* path only: bbox\n// endpoints + a single full-L2 index assignment per channel, no LSQ refit\n// (the WGSL `QUALITY_HIGH` branch). Always emits 6-interpolation mode\n// (red0 > red1). See bc5.wgsl for the full derivation.\n\nprecision highp float;\nprecision highp int;\n\nuniform sampler2D uSrc;\nuniform ivec2 uSrcSize;\nuniform int uFlipY;\n\nlayout(location = 0) out uvec4 outColor;\n\n// 6-interpolation-mode palette weights: pal[j] = W0_6[j]*r0 + W1_6[j]*r1.\nconst float W0_6[8] = float[8](1.0, 0.0, 6.0 / 7.0, 5.0 / 7.0, 4.0 / 7.0, 3.0 / 7.0, 2.0 / 7.0, 1.0 / 7.0);\nconst float W1_6[8] = float[8](0.0, 1.0, 1.0 / 7.0, 2.0 / 7.0, 3.0 / 7.0, 4.0 / 7.0, 5.0 / 7.0, 6.0 / 7.0);\n\nuint quantize8(float v) {\n  return uint(clamp(floor(v * 255.0 + 0.5), 0.0, 255.0));\n}\n\n// Encode 16 single-channel values into an 8-byte BC4 block (two little-endian\n// u32s). Mirrors encode_bc4() in bc5.wgsl with the refit pass omitted.\nuvec2 encodeBC4(float values[16]) {\n  float vmin = 1.0;\n  float vmax = 0.0;\n  for (int k = 0; k < 16; k++) {\n    vmin = min(vmin, values[k]);\n    vmax = max(vmax, values[k]);\n  }\n  uint r0 = quantize8(vmax);\n  uint r1 = quantize8(vmin);\n  if (r0 == r1) {\n    if (r1 > 0u) { r1 = r1 - 1u; } else { r0 = r0 + 1u; }\n  }\n\n  float pal[8];\n  float r0f = float(r0) / 255.0;\n  float r1f = float(r1) / 255.0;\n  for (int j = 0; j < 8; j++) {\n    pal[j] = W0_6[j] * r0f + W1_6[j] * r1f;\n  }\n\n  uint indices[16];\n  for (int k = 0; k < 16; k++) {\n    float v = values[k];\n    uint bestJ = 0u;\n    float bestD = 1e20;\n    for (int j = 0; j < 8; j++) {\n      float d = pal[j] - v;\n      float d2 = d * d;\n      if (d2 < bestD) { bestD = d2; bestJ = uint(j); }\n    }\n    indices[k] = bestJ;\n  }\n\n  // Pack the 48-bit index field (bytes 2..7) split across two u32 halves.\n  uint idxLo = 0u;\n  uint idxHi = 0u;\n  for (int k = 0; k < 16; k++) {\n    uint bit = 3u * uint(k);\n    uint v = indices[k] & 7u;\n    if (bit + 3u <= 32u) {\n      idxLo = idxLo | (v << bit);\n    } else if (bit >= 32u) {\n      idxHi = idxHi | (v << (bit - 32u));\n    } else {\n      idxLo = idxLo | (v << bit);\n      idxHi = idxHi | (v >> (32u - bit));\n    }\n  }\n\n  uint outLo = r0 | (r1 << 8) | ((idxLo & 0xFFFFu) << 16);\n  uint outHi = (idxLo >> 16) | (idxHi << 16);\n  return uvec2(outLo, outHi);\n}\n\nvoid main() {\n  ivec2 base = ivec2(gl_FragCoord.xy) * 4;\n  ivec2 maxXY = uSrcSize - ivec2(1);\n\n  float rValues[16];\n  float gValues[16];\n  for (int i = 0; i < 16; i++) {\n    ivec2 p = clamp(base + ivec2(i & 3, i >> 2), ivec2(0), maxXY);\n    int sy = (uFlipY != 0) ? (uSrcSize.y - 1 - p.y) : p.y;\n    vec4 c = texelFetch(uSrc, ivec2(p.x, sy), 0);\n    rValues[i] = c.r;\n    gValues[i] = c.g;\n  }\n\n  uvec2 rBlock = encodeBC4(rValues);\n  uvec2 gBlock = encodeBC4(gValues);\n  outColor = uvec4(rBlock.x, rBlock.y, gBlock.x, gBlock.y);\n}\n";
+var bc5_frag_default = "#version 300 es\n// BC5 (RGTC2) fragment-shader encoder \u2014 WebGL2 port of bc5.wgsl (fast path).\n//\n// One fragment per 4\xD74 block \u2192 16-byte BC5 block as 4 \xD7 u32 in outColor.\n// BC5 = two BC4 halves (R then G). This is the *fast* path only: bbox\n// endpoints + a full-L2 index assignment per channel with the least-squares\n// refit sums accumulated in the same pass, then one refit accepted only when\n// it lowers the block's error (mirrors bc5.wgsl's fast branch \u2014 worth\n// ~1.3 dB on the normal-map card). Always emits 6-interpolation mode\n// (red0 > red1). See bc5.wgsl for the full derivation.\n\nprecision highp float;\nprecision highp int;\n\nuniform sampler2D uSrc;\nuniform ivec2 uSrcSize;\nuniform int uFlipY;\n\nlayout(location = 0) out uvec4 outColor;\n\n// 6-interpolation-mode palette weights: pal[j] = W0_6[j]*r0 + W1_6[j]*r1.\nconst float W0_6[8] = float[8](1.0, 0.0, 6.0 / 7.0, 5.0 / 7.0, 4.0 / 7.0, 3.0 / 7.0, 2.0 / 7.0, 1.0 / 7.0);\nconst float W1_6[8] = float[8](0.0, 1.0, 1.0 / 7.0, 2.0 / 7.0, 3.0 / 7.0, 4.0 / 7.0, 5.0 / 7.0, 6.0 / 7.0);\n\nuint quantize8(float v) {\n  return uint(clamp(floor(v * 255.0 + 0.5), 0.0, 255.0));\n}\n\n// Nearest-palette assignment with the LSQ normal-equation sums and total\n// squared error accumulated in the same pass. Sums are only consumed by the\n// caller's refit; err drives the accept-if-better test.\nstruct Assign { float err; float sAA; float sBB; float sAB; float sAV; float sBV; };\nAssign assignAll(float values[16], float pal[8], out uint indices[16]) {\n  Assign r = Assign(0.0, 0.0, 0.0, 0.0, 0.0, 0.0);\n  for (int k = 0; k < 16; k++) {\n    float v = values[k];\n    uint bestJ = 0u;\n    float bestD = 1e20;\n    for (int j = 0; j < 8; j++) {\n      float d = pal[j] - v;\n      float d2 = d * d;\n      if (d2 < bestD) { bestD = d2; bestJ = uint(j); }\n    }\n    indices[k] = bestJ;\n    r.err += bestD;\n    float a = W0_6[int(bestJ)];\n    float b = W1_6[int(bestJ)];\n    r.sAA += a * a; r.sBB += b * b; r.sAB += a * b; r.sAV += a * v; r.sBV += b * v;\n  }\n  return r;\n}\n\n// Encode 16 single-channel values into an 8-byte BC4 block (two little-endian\n// u32s). Mirrors encode_bc4() in bc5.wgsl's fast branch: bbox seed, fused\n// assignment + LSQ sums, refit accepted only if the error drops. vmin/vmax\n// are the channel's min/max, computed in the caller's load loop \u2014 fusing\n// that scan there saves a 16-value pass per channel.\nuvec2 encodeBC4(float values[16], float vmin, float vmax) {\n  uint r0 = quantize8(vmax);\n  uint r1 = quantize8(vmin);\n  if (r0 == r1) {\n    if (r1 > 0u) { r1 = r1 - 1u; } else { r0 = r0 + 1u; }\n  }\n\n  float pal[8];\n  float r0f = float(r0) / 255.0;\n  float r1f = float(r1) / 255.0;\n  for (int j = 0; j < 8; j++) {\n    pal[j] = W0_6[j] * r0f + W1_6[j] * r1f;\n  }\n\n  uint indices[16];\n  Assign seed = assignAll(values, pal, indices);\n\n  // One least-squares refit, accepted only if the requantised endpoints lower\n  // the block error. Clamp to the block's value range (a strict-SSE win vs\n  // clamping to [0,1], same as the other formats' fast paths); keep 6-interp\n  // mode (r0 > r1 strictly).\n  float det = seed.sAA * seed.sBB - seed.sAB * seed.sAB;\n  if (abs(det) > 1e-9) {\n    float e0 = clamp((seed.sBB * seed.sAV - seed.sAB * seed.sBV) / det, vmin, vmax);\n    float e1 = clamp((seed.sAA * seed.sBV - seed.sAB * seed.sAV) / det, vmin, vmax);\n    uint n0 = quantize8(e0);\n    uint n1 = quantize8(e1);\n    if (n0 > n1 && !(n0 == r0 && n1 == r1)) {\n      float pal2[8];\n      float n0f = float(n0) / 255.0;\n      float n1f = float(n1) / 255.0;\n      for (int j = 0; j < 8; j++) {\n        pal2[j] = W0_6[j] * n0f + W1_6[j] * n1f;\n      }\n      uint idx2[16];\n      Assign refit = assignAll(values, pal2, idx2);\n      if (refit.err < seed.err) {\n        r0 = n0; r1 = n1;\n        for (int k = 0; k < 16; k++) indices[k] = idx2[k];\n      }\n    }\n  }\n\n  // Pack the 48-bit index field (bytes 2..7) split across two u32 halves.\n  uint idxLo = 0u;\n  uint idxHi = 0u;\n  for (int k = 0; k < 16; k++) {\n    uint bit = 3u * uint(k);\n    uint v = indices[k] & 7u;\n    if (bit + 3u <= 32u) {\n      idxLo = idxLo | (v << bit);\n    } else if (bit >= 32u) {\n      idxHi = idxHi | (v << (bit - 32u));\n    } else {\n      idxLo = idxLo | (v << bit);\n      idxHi = idxHi | (v >> (32u - bit));\n    }\n  }\n\n  uint outLo = r0 | (r1 << 8) | ((idxLo & 0xFFFFu) << 16);\n  uint outHi = (idxLo >> 16) | (idxHi << 16);\n  return uvec2(outLo, outHi);\n}\n\nvoid main() {\n  ivec2 base = ivec2(gl_FragCoord.xy) * 4;\n  ivec2 maxXY = uSrcSize - ivec2(1);\n\n  float rValues[16];\n  float gValues[16];\n  float rMin = 1.0; float rMax = 0.0;\n  float gMin = 1.0; float gMax = 0.0;\n  for (int i = 0; i < 16; i++) {\n    ivec2 p = clamp(base + ivec2(i & 3, i >> 2), ivec2(0), maxXY);\n    int sy = (uFlipY != 0) ? (uSrcSize.y - 1 - p.y) : p.y;\n    vec4 c = texelFetch(uSrc, ivec2(p.x, sy), 0);\n    rValues[i] = c.r;\n    gValues[i] = c.g;\n    rMin = min(rMin, c.r); rMax = max(rMax, c.r);\n    gMin = min(gMin, c.g); gMax = max(gMax, c.g);\n  }\n\n  uvec2 rBlock = encodeBC4(rValues, rMin, rMax);\n  uvec2 gBlock = encodeBC4(gValues, gMin, gMax);\n  outColor = uvec4(rBlock.x, rBlock.y, gBlock.x, gBlock.y);\n}\n";
 
 // src/webgl/BC5WebGLEncoder.ts
 var BC5WebGLEncoder = class extends WebGLBlockEncoder {
@@ -1238,7 +1556,7 @@ var BC5WebGLEncoder = class extends WebGLBlockEncoder {
 };
 
 // src/webgl/glsl/bc7.frag.glsl
-var bc7_frag_default = "#version 300 es\n// BC7 (BPTC) mode-6 fragment-shader encoder \u2014 WebGL2 port of bc7.wgsl (fast).\n//\n// One fragment per 4\xD74 block \u2192 16-byte block as 4 \xD7 u32 in outColor. Fast path\n// only: O(N) bbox seed \u2192 endpoints fitted by a single least-squares pass whose\n// normal-equation sums are accumulated during a projection-based index\n// assignment (palette is colinear, so the nearest entry is found by projecting\n// onto the endpoint line \u2014 O(1) per pixel). Mirrors the `QUALITY_HIGH == 0`\n// branch of bc7.wgsl; see that file for the mode-6 bit layout and rationale.\n//\n// Determinism note: the WGSL refit uses round() (half-to-even); here we use\n// floor(x + 0.5) for portability. The two differ only at exact .5 ties, a\n// sub-LSB endpoint nudge that is visually identical.\n\nprecision highp float;\nprecision highp int;\n\nuniform sampler2D uSrc;\nuniform ivec2 uSrcSize;\nuniform int uFlipY;\n\nlayout(location = 0) out uvec4 outColor;\n\n// Per-invocation scratch (mirrors the WGSL function-scope arrays passed by ptr).\nivec4 gPixels[16];\nuint gIdx[16];\n\nstruct QuantPair { ivec4 seven; ivec4 eight; };\nstruct Ep { ivec4 seven; ivec4 eight; uint p; };\nstruct Fit { ivec4 e0; ivec4 e1; bool valid; };\n\nivec4 to8(vec4 v) {\n  return ivec4(clamp(floor(v * 255.0 + 0.5), vec4(0.0), vec4(255.0)));\n}\n\nint dist2(ivec4 a, ivec4 b) {\n  ivec4 d = a - b;\n  ivec4 e = d * d;\n  return e.x + e.y + e.z + e.w;\n}\n\n// Quantize an 8-bit ideal endpoint to (7-bit value, reconstructed 8-bit) under\n// a fixed p-bit, all four channels at once.\nQuantPair quantizeEndpoint(ivec4 ideal8, uint p) {\n  ivec4 q = ivec4(clamp(floor((vec4(ideal8) - float(p)) / 2.0 + 0.5), vec4(0.0), vec4(127.0)));\n  // eff = (q << 1) | p. q*2 is even and p \u2208 {0,1}, so q*2 + p is identical and\n  // avoids any vector-shift-by-scalar portability question.\n  ivec4 eff = q * 2 + ivec4(int(p));\n  return QuantPair(q, eff);\n}\n\n// Endpoint with its chosen p-bit, picked by minimum quantisation error.\nEp pickEp(ivec4 ideal) {\n  QuantPair a = quantizeEndpoint(ideal, 0u);\n  QuantPair b = quantizeEndpoint(ideal, 1u);\n  if (dist2(b.eight, ideal) < dist2(a.eight, ideal)) {\n    return Ep(b.seven, b.eight, 1u);\n  }\n  return Ep(a.seven, a.eight, 0u);\n}\n\n// Projection index assignment over gPixels \u2192 gIdx. When `fit`, accumulate the\n// LSQ normal-equation sums in the same pass and return refitted endpoints.\nFit projAssign(ivec4 pe0, ivec4 pe1, bool fit) {\n  Fit res;\n  res.e0 = ivec4(0);\n  res.e1 = ivec4(0);\n  res.valid = false;\n  ivec4 dir = pe1 - pe0;\n  int dd = dir.x * dir.x + dir.y * dir.y + dir.z * dir.z + dir.w * dir.w;\n  if (dd == 0) {\n    for (int k = 0; k < 16; k++) { gIdx[k] = 0u; }\n    return res;\n  }\n  float inv = 15.0 / float(dd);\n  float sAA = 0.0, sBB = 0.0, sAB = 0.0;\n  vec4 sAV = vec4(0.0), sBV = vec4(0.0);\n  for (int k = 0; k < 16; k++) {\n    ivec4 q = gPixels[k] - pe0;\n    float proj = float(q.x * dir.x + q.y * dir.y + q.z * dir.z + q.w * dir.w) * inv;\n    float s = clamp(floor(proj + 0.5), 0.0, 15.0);\n    gIdx[k] = uint(s);\n    if (fit) {\n      vec4 v = vec4(gPixels[k]);\n      float b = s / 15.0;\n      float a = 1.0 - b;\n      sAA += a * a; sBB += b * b; sAB += a * b; sAV += a * v; sBV += b * v;\n    }\n  }\n  if (!fit) { return res; }\n  float det = sAA * sBB - sAB * sAB;\n  if (abs(det) < 1e-9) { return res; }\n  res.e0 = ivec4(clamp(floor((sBB * sAV - sAB * sBV) / det + 0.5), vec4(0.0), vec4(255.0)));\n  res.e1 = ivec4(clamp(floor((sAA * sBV - sAB * sAV) / det + 0.5), vec4(0.0), vec4(255.0)));\n  res.valid = true;\n  return res;\n}\n\nvoid writeBits(inout uint block[4], uint pos, uint nbits, uint value) {\n  uint v = value & ((1u << nbits) - 1u);\n  uint wordLo = pos / 32u;\n  uint bitLo = pos % 32u;\n  uint bitsInLo = min(nbits, 32u - bitLo);\n  uint maskLo = ((1u << bitsInLo) - 1u) << bitLo;\n  block[wordLo] = (block[wordLo] & ~maskLo) | ((v << bitLo) & maskLo);\n  if (bitsInLo < nbits) {\n    uint bitsInHi = nbits - bitsInLo;\n    uint maskHi = (1u << bitsInHi) - 1u;\n    uint valHi = v >> bitsInLo;\n    block[wordLo + 1u] = (block[wordLo + 1u] & ~maskHi) | (valHi & maskHi);\n  }\n}\n\nvoid main() {\n  ivec2 base = ivec2(gl_FragCoord.xy) * 4;\n  ivec2 maxXY = uSrcSize - ivec2(1);\n\n  ivec4 lo = ivec4(255);\n  ivec4 hi = ivec4(0);\n  for (int i = 0; i < 16; i++) {\n    ivec2 p = clamp(base + ivec2(i & 3, i >> 2), ivec2(0), maxXY);\n    int sy = (uFlipY != 0) ? (uSrcSize.y - 1 - p.y) : p.y;\n    ivec4 px = to8(texelFetch(uSrc, ivec2(p.x, sy), 0));\n    gPixels[i] = px;\n    lo = min(lo, px);\n    hi = max(hi, px);\n  }\n\n  Ep ep0 = pickEp(lo);\n  Ep ep1 = pickEp(hi);\n  Fit r = projAssign(ep0.eight, ep1.eight, true);\n  if (r.valid) {\n    ep0 = pickEp(r.e0);\n    ep1 = pickEp(r.e1);\n    projAssign(ep0.eight, ep1.eight, false);\n  }\n  ivec4 e0_7 = ep0.seven;\n  ivec4 e1_7 = ep1.seven;\n  uint p0 = ep0.p;\n  uint p1 = ep1.p;\n\n  // Anchor rule \u2014 pixel 0's index MSB must be 0; otherwise swap endpoints and\n  // reflect every index (decoded image unchanged).\n  if ((gIdx[0] & 0x8u) != 0u) {\n    ivec4 t = e0_7; e0_7 = e1_7; e1_7 = t;\n    uint tp = p0; p0 = p1; p1 = tp;\n    for (int k = 0; k < 16; k++) { gIdx[k] = 15u - gIdx[k]; }\n  }\n\n  uint block[4];\n  block[0] = 0u; block[1] = 0u; block[2] = 0u; block[3] = 0u;\n  uint pos = 0u;\n  writeBits(block, pos, 7u, 0x40u); pos += 7u;\n  writeBits(block, pos, 7u, uint(e0_7.x)); pos += 7u;\n  writeBits(block, pos, 7u, uint(e1_7.x)); pos += 7u;\n  writeBits(block, pos, 7u, uint(e0_7.y)); pos += 7u;\n  writeBits(block, pos, 7u, uint(e1_7.y)); pos += 7u;\n  writeBits(block, pos, 7u, uint(e0_7.z)); pos += 7u;\n  writeBits(block, pos, 7u, uint(e1_7.z)); pos += 7u;\n  writeBits(block, pos, 7u, uint(e0_7.w)); pos += 7u;\n  writeBits(block, pos, 7u, uint(e1_7.w)); pos += 7u;\n  writeBits(block, pos, 1u, p0); pos += 1u;\n  writeBits(block, pos, 1u, p1); pos += 1u;\n  writeBits(block, pos, 3u, gIdx[0] & 0x7u); pos += 3u;\n  for (int k = 1; k < 16; k++) {\n    writeBits(block, pos, 4u, gIdx[k] & 0xFu);\n    pos += 4u;\n  }\n\n  outColor = uvec4(block[0], block[1], block[2], block[3]);\n}\n";
+var bc7_frag_default = "#version 300 es\n// BC7 (BPTC) mode-6 fragment-shader encoder \u2014 WebGL2 port of bc7.wgsl (fast).\n//\n// One fragment per 4\xD74 block \u2192 16-byte block as 4 \xD7 u32 in outColor. Fast path\n// only: principal-axis seed (covariance power-iteration; bbox on degenerate\n// blocks) at the exact projection extents, quantised directly (no LSQ refit \u2014\n// mode 6's 16-level palette leaves it under 0.15 dB) \u2192 one projection-based\n// index-assignment pass (palette is colinear, so the nearest entry is found\n// by projecting onto the endpoint line \u2014 O(1) per pixel). Mirrors the\n// `QUALITY_HIGH == 0` branch of bc7.wgsl; see that file for the mode-6 bit\n// layout and rationale.\n//\n// Determinism note: the WGSL refit uses round() (half-to-even); here we use\n// floor(x + 0.5) for portability. The two differ only at exact .5 ties, a\n// sub-LSB endpoint nudge that is visually identical.\n\nprecision highp float;\nprecision highp int;\n\nuniform sampler2D uSrc;\nuniform ivec2 uSrcSize;\nuniform int uFlipY;\n\nlayout(location = 0) out uvec4 outColor;\n\n// Per-invocation scratch (mirrors the WGSL function-scope arrays passed by ptr).\nivec4 gPixels[16];\nuint gIdx[16];\n\nstruct QuantPair { ivec4 seven; ivec4 eight; };\nstruct Ep { ivec4 seven; ivec4 eight; uint p; };\nstruct Fit { ivec4 e0; ivec4 e1; bool valid; };\n\nivec4 to8(vec4 v) {\n  return ivec4(clamp(floor(v * 255.0 + 0.5), vec4(0.0), vec4(255.0)));\n}\n\nint dist2(ivec4 a, ivec4 b) {\n  ivec4 d = a - b;\n  ivec4 e = d * d;\n  return e.x + e.y + e.z + e.w;\n}\n\n// Quantize an 8-bit ideal endpoint to (7-bit value, reconstructed 8-bit) under\n// a fixed p-bit, all four channels at once.\nQuantPair quantizeEndpoint(ivec4 ideal8, uint p) {\n  ivec4 q = ivec4(clamp(floor((vec4(ideal8) - float(p)) / 2.0 + 0.5), vec4(0.0), vec4(127.0)));\n  // eff = (q << 1) | p. q*2 is even and p \u2208 {0,1}, so q*2 + p is identical and\n  // avoids any vector-shift-by-scalar portability question.\n  ivec4 eff = q * 2 + ivec4(int(p));\n  return QuantPair(q, eff);\n}\n\n// Endpoint with its chosen p-bit, picked by minimum quantisation error.\nEp pickEp(ivec4 ideal) {\n  QuantPair a = quantizeEndpoint(ideal, 0u);\n  QuantPair b = quantizeEndpoint(ideal, 1u);\n  if (dist2(b.eight, ideal) < dist2(a.eight, ideal)) {\n    return Ep(b.seven, b.eight, 1u);\n  }\n  return Ep(a.seven, a.eight, 0u);\n}\n\n// Principal colour axis via power-iteration over precomputed, mean-corrected\n// covariance rows (the moments are accumulated for free in the pixel-load\n// loop), seeded with the bbox diagonal. Returns a unit axis, or vec4(0.0)\n// for a degenerate (constant) block. The bbox diagonal alone is sign-blind\n// and points across anti-correlated data (normal maps, hue edges) instead of\n// along it.\nvec4 principalAxis(vec4 c0v, vec4 c1v, vec4 c2v, vec4 c3v, vec4 seed) {\n  vec4 v = seed;\n  float len = length(v);\n  if (len < 1e-9) { return vec4(0.0); }\n  v /= len;\n  for (int it = 0; it < 8; it++) {\n    vec4 nv = vec4(dot(c0v, v), dot(c1v, v), dot(c2v, v), dot(c3v, v));\n    len = length(nv);\n    if (len < 1e-12) { return vec4(0.0); }\n    v = nv / len;\n  }\n  return v;\n}\n\n// Projection index assignment over gPixels \u2192 gIdx. When `fit`, accumulate the\n// LSQ normal-equation sums in the same pass and return refitted endpoints.\nFit projAssign(ivec4 pe0, ivec4 pe1, bool fit) {\n  Fit res;\n  res.e0 = ivec4(0);\n  res.e1 = ivec4(0);\n  res.valid = false;\n  ivec4 dir = pe1 - pe0;\n  int dd = dir.x * dir.x + dir.y * dir.y + dir.z * dir.z + dir.w * dir.w;\n  if (dd == 0) {\n    for (int k = 0; k < 16; k++) { gIdx[k] = 0u; }\n    return res;\n  }\n  float inv = 15.0 / float(dd);\n  float sAA = 0.0, sBB = 0.0, sAB = 0.0;\n  vec4 sAV = vec4(0.0), sBV = vec4(0.0);\n  for (int k = 0; k < 16; k++) {\n    ivec4 q = gPixels[k] - pe0;\n    float proj = float(q.x * dir.x + q.y * dir.y + q.z * dir.z + q.w * dir.w) * inv;\n    float s = clamp(floor(proj + 0.5), 0.0, 15.0);\n    gIdx[k] = uint(s);\n    if (fit) {\n      vec4 v = vec4(gPixels[k]);\n      float b = s / 15.0;\n      float a = 1.0 - b;\n      sAA += a * a; sBB += b * b; sAB += a * b; sAV += a * v; sBV += b * v;\n    }\n  }\n  if (!fit) { return res; }\n  float det = sAA * sBB - sAB * sAB;\n  if (abs(det) < 1e-9) { return res; }\n  res.e0 = ivec4(clamp(floor((sBB * sAV - sAB * sBV) / det + 0.5), vec4(0.0), vec4(255.0)));\n  res.e1 = ivec4(clamp(floor((sAA * sBV - sAB * sAV) / det + 0.5), vec4(0.0), vec4(255.0)));\n  res.valid = true;\n  return res;\n}\n\nvoid writeBits(inout uint block[4], uint pos, uint nbits, uint value) {\n  uint v = value & ((1u << nbits) - 1u);\n  uint wordLo = pos / 32u;\n  uint bitLo = pos % 32u;\n  uint bitsInLo = min(nbits, 32u - bitLo);\n  uint maskLo = ((1u << bitsInLo) - 1u) << bitLo;\n  block[wordLo] = (block[wordLo] & ~maskLo) | ((v << bitLo) & maskLo);\n  if (bitsInLo < nbits) {\n    uint bitsInHi = nbits - bitsInLo;\n    uint maskHi = (1u << bitsInHi) - 1u;\n    uint valHi = v >> bitsInLo;\n    block[wordLo + 1u] = (block[wordLo + 1u] & ~maskHi) | (valHi & maskHi);\n  }\n}\n\nvoid main() {\n  ivec2 base = ivec2(gl_FragCoord.xy) * 4;\n  ivec2 maxXY = uSrcSize - ivec2(1);\n\n  // Load pass with the covariance moments FUSED in: d = px \u2212 pixel0\n  // (first-pixel-relative, so the sums scale with the block's span).\n  ivec4 lo = ivec4(255);\n  ivec4 hi = ivec4(0);\n  vec4 p0f = vec4(0.0);\n  vec4 sd = vec4(0.0);\n  vec4 c0v = vec4(0.0);\n  vec4 c1v = vec4(0.0);\n  vec4 c2v = vec4(0.0);\n  vec4 c3v = vec4(0.0);\n  for (int i = 0; i < 16; i++) {\n    ivec2 p = clamp(base + ivec2(i & 3, i >> 2), ivec2(0), maxXY);\n    int sy = (uFlipY != 0) ? (uSrcSize.y - 1 - p.y) : p.y;\n    ivec4 px = to8(texelFetch(uSrc, ivec2(p.x, sy), 0));\n    gPixels[i] = px;\n    lo = min(lo, px);\n    hi = max(hi, px);\n    if (i == 0) { p0f = vec4(px); }\n    vec4 d = vec4(px) - p0f;\n    sd += d;\n    c0v += d.x * d;\n    c1v += d.y * d;\n    c2v += d.z * d;\n    c3v += d.w * d;\n  }\n  vec4 mean = p0f + sd / 16.0;\n  // Mean-correct the fused moments: C = \u03A3dd\u1D40 \u2212 (\u03A3d)(\u03A3d)\u1D40/16.\n  vec4 sd16 = sd / 16.0;\n  c0v -= sd.x * sd16;\n  c1v -= sd.y * sd16;\n  c2v -= sd.z * sd16;\n  c3v -= sd.w * sd16;\n\n  ivec4 seed0 = lo;\n  ivec4 seed1 = hi;\n  vec4 axis = principalAxis(c0v, c1v, c2v, c3v, vec4(hi - lo));\n  if (dot(axis, axis) > 0.0) {\n    // Exact projection extents along the axis. (A Rayleigh-quotient span\n    // estimate was tried in place of this pass \u2014 it saves 16 dots but costs\n    // 0.1\u20130.8 dB and 4\u201310\xD7 on the worst-easy-block gate: \u03C3 misjudges\n    // two-cluster and outlier blocks. The pass stays.)\n    float tMin = 1e30;\n    float tMax = -1e30;\n    for (int k = 0; k < 16; k++) {\n      float t = dot(vec4(gPixels[k]) - mean, axis);\n      tMin = min(tMin, t);\n      tMax = max(tMax, t);\n    }\n    seed0 = ivec4(clamp(floor(mean + tMin * axis + 0.5), vec4(0.0), vec4(255.0)));\n    seed1 = ivec4(clamp(floor(mean + tMax * axis + 0.5), vec4(0.0), vec4(255.0)));\n  }\n\n  // Quantise the PCA-extents seed directly and assign indices in one\n  // projection pass \u2014 no LSQ refit: with the seed already on the principal\n  // axis, mode 6's 16-level palette leaves the refit under 0.15 dB (the\n  // coarse 4-level BC1/ASTC fast paths DO keep theirs).\n  Ep ep0 = pickEp(seed0);\n  Ep ep1 = pickEp(seed1);\n  projAssign(ep0.eight, ep1.eight, false);\n  ivec4 e0_7 = ep0.seven;\n  ivec4 e1_7 = ep1.seven;\n  uint p0 = ep0.p;\n  uint p1 = ep1.p;\n\n  // Anchor rule \u2014 pixel 0's index MSB must be 0; otherwise swap endpoints and\n  // reflect every index (decoded image unchanged).\n  if ((gIdx[0] & 0x8u) != 0u) {\n    ivec4 t = e0_7; e0_7 = e1_7; e1_7 = t;\n    uint tp = p0; p0 = p1; p1 = tp;\n    for (int k = 0; k < 16; k++) { gIdx[k] = 15u - gIdx[k]; }\n  }\n\n  uint block[4];\n  block[0] = 0u; block[1] = 0u; block[2] = 0u; block[3] = 0u;\n  uint pos = 0u;\n  writeBits(block, pos, 7u, 0x40u); pos += 7u;\n  writeBits(block, pos, 7u, uint(e0_7.x)); pos += 7u;\n  writeBits(block, pos, 7u, uint(e1_7.x)); pos += 7u;\n  writeBits(block, pos, 7u, uint(e0_7.y)); pos += 7u;\n  writeBits(block, pos, 7u, uint(e1_7.y)); pos += 7u;\n  writeBits(block, pos, 7u, uint(e0_7.z)); pos += 7u;\n  writeBits(block, pos, 7u, uint(e1_7.z)); pos += 7u;\n  writeBits(block, pos, 7u, uint(e0_7.w)); pos += 7u;\n  writeBits(block, pos, 7u, uint(e1_7.w)); pos += 7u;\n  writeBits(block, pos, 1u, p0); pos += 1u;\n  writeBits(block, pos, 1u, p1); pos += 1u;\n  writeBits(block, pos, 3u, gIdx[0] & 0x7u); pos += 3u;\n  for (int k = 1; k < 16; k++) {\n    writeBits(block, pos, 4u, gIdx[k] & 0xFu);\n    pos += 4u;\n  }\n\n  outColor = uvec4(block[0], block[1], block[2], block[3]);\n}\n";
 
 // src/webgl/BC7WebGLEncoder.ts
 var BC7WebGLEncoder = class extends WebGLBlockEncoder {
@@ -1257,7 +1575,7 @@ var BC7WebGLEncoder = class extends WebGLBlockEncoder {
 };
 
 // src/webgl/glsl/astc4x4.frag.glsl
-var astc4x4_frag_default = "#version 300 es\n// ASTC 4\xD74 LDR fragment-shader encoder \u2014 WebGL2 port of astc4x4.wgsl (fast).\n//\n// One fragment per 4\xD74 block \u2192 16-byte block as 4 \xD7 u32 in outColor. Restricted\n// subset: single partition, no dual-plane, CEM 12 (LDR RGBA direct), 4\xD74 weight\n// grid with 2-bit weights (QUANT_4), 8-bit endpoints (QUANT_256). Fast path:\n// bbox seed \u2192 one LSQ refit fused into a projection weight assignment (4 colinear\n// levels). Mirrors the `QUALITY_HIGH == 0` branch of astc4x4.wgsl; see that file\n// for the 128-bit block layout.\n//\n// Determinism note: floor(x + 0.5) replaces WGSL round() for the refit endpoints\n// (sub-LSB difference at exact .5 ties only).\n\nprecision highp float;\nprecision highp int;\n\nuniform sampler2D uSrc;\nuniform ivec2 uSrcSize;\nuniform int uFlipY;\n\nlayout(location = 0) out uvec4 outColor;\n\nivec4 gPixels[16];\nuint gIdx[16];\n\nstruct Fit { ivec4 e0; ivec4 e1; bool valid; };\n\nivec4 to8(vec4 v) {\n  return ivec4(clamp(floor(v * 255.0 + 0.5), vec4(0.0), vec4(255.0)));\n}\n\n// Projection weight assignment over 4 levels (QUANT_4 \u2248 thirds), with the LSQ\n// normal-equation sums accumulated in the same pass for a fused refit.\nFit projAssign(ivec4 pe0, ivec4 pe1, bool fit) {\n  Fit res;\n  res.e0 = ivec4(0);\n  res.e1 = ivec4(0);\n  res.valid = false;\n  ivec4 dir = pe1 - pe0;\n  int dd = dir.x * dir.x + dir.y * dir.y + dir.z * dir.z + dir.w * dir.w;\n  if (dd == 0) {\n    for (int k = 0; k < 16; k++) { gIdx[k] = 0u; }\n    return res;\n  }\n  float inv = 3.0 / float(dd);\n  float sAA = 0.0, sBB = 0.0, sAB = 0.0;\n  vec4 sAV = vec4(0.0), sBV = vec4(0.0);\n  for (int k = 0; k < 16; k++) {\n    ivec4 q = gPixels[k] - pe0;\n    float proj = float(q.x * dir.x + q.y * dir.y + q.z * dir.z + q.w * dir.w) * inv;\n    float s = clamp(floor(proj + 0.5), 0.0, 3.0);\n    gIdx[k] = uint(s);\n    if (fit) {\n      vec4 v = vec4(gPixels[k]);\n      float b = s / 3.0;\n      float a = 1.0 - b;\n      sAA += a * a; sBB += b * b; sAB += a * b; sAV += a * v; sBV += b * v;\n    }\n  }\n  if (!fit) { return res; }\n  float det = sAA * sBB - sAB * sAB;\n  if (abs(det) < 1e-9) { return res; }\n  res.e0 = ivec4(clamp(floor((sBB * sAV - sAB * sBV) / det + 0.5), vec4(0.0), vec4(255.0)));\n  res.e1 = ivec4(clamp(floor((sAA * sBV - sAB * sAV) / det + 0.5), vec4(0.0), vec4(255.0)));\n  res.valid = true;\n  return res;\n}\n\nvoid writeBits(inout uint block[4], uint pos, uint nbits, uint value) {\n  uint v = value & ((1u << nbits) - 1u);\n  uint wordLo = pos / 32u;\n  uint bitLo = pos % 32u;\n  uint bitsInLo = min(nbits, 32u - bitLo);\n  uint maskLo = ((1u << bitsInLo) - 1u) << bitLo;\n  block[wordLo] = (block[wordLo] & ~maskLo) | ((v << bitLo) & maskLo);\n  if (bitsInLo < nbits) {\n    uint bitsInHi = nbits - bitsInLo;\n    uint maskHi = (1u << bitsInHi) - 1u;\n    uint valHi = v >> bitsInLo;\n    block[wordLo + 1u] = (block[wordLo + 1u] & ~maskHi) | (valHi & maskHi);\n  }\n}\n\nvoid main() {\n  ivec2 base = ivec2(gl_FragCoord.xy) * 4;\n  ivec2 maxXY = uSrcSize - ivec2(1);\n\n  ivec4 lo = ivec4(255);\n  ivec4 hi = ivec4(0);\n  for (int i = 0; i < 16; i++) {\n    ivec2 p = clamp(base + ivec2(i & 3, i >> 2), ivec2(0), maxXY);\n    int sy = (uFlipY != 0) ? (uSrcSize.y - 1 - p.y) : p.y;\n    ivec4 px = to8(texelFetch(uSrc, ivec2(p.x, sy), 0));\n    gPixels[i] = px;\n    lo = min(lo, px);\n    hi = max(hi, px);\n  }\n\n  ivec4 e0 = lo;\n  ivec4 e1 = hi;\n  Fit r = projAssign(e0, e1, true);\n  if (r.valid) {\n    e0 = r.e0;\n    e1 = r.e1;\n    projAssign(e0, e1, false);\n  }\n\n  // Endpoint ordering so the decoder doesn't apply blue contraction.\n  if (e0.x + e0.y + e0.z > e1.x + e1.y + e1.z) {\n    ivec4 t = e0; e0 = e1; e1 = t;\n    for (int k = 0; k < 16; k++) { gIdx[k] = 3u - gIdx[k]; }\n  }\n\n  uint block[4];\n  block[0] = 0u; block[1] = 0u; block[2] = 0u; block[3] = 0u;\n  writeBits(block, 0u, 11u, 0x042u);\n  writeBits(block, 11u, 2u, 0u);\n  writeBits(block, 13u, 4u, 12u);\n  writeBits(block, 17u + 0u * 8u, 8u, uint(e0.x));\n  writeBits(block, 17u + 1u * 8u, 8u, uint(e1.x));\n  writeBits(block, 17u + 2u * 8u, 8u, uint(e0.y));\n  writeBits(block, 17u + 3u * 8u, 8u, uint(e1.y));\n  writeBits(block, 17u + 4u * 8u, 8u, uint(e0.z));\n  writeBits(block, 17u + 5u * 8u, 8u, uint(e1.z));\n  writeBits(block, 17u + 6u * 8u, 8u, uint(e0.w));\n  writeBits(block, 17u + 7u * 8u, 8u, uint(e1.w));\n\n  uint w3 = 0u;\n  for (int k = 0; k < 16; k++) {\n    uint w = gIdx[k] & 0x3u;\n    w3 = w3 | ((w & 1u) << (31u - 2u * uint(k))) | (((w >> 1u) & 1u) << (30u - 2u * uint(k)));\n  }\n  block[3] = w3;\n\n  outColor = uvec4(block[0], block[1], block[2], block[3]);\n}\n";
+var astc4x4_frag_default = "#version 300 es\n// ASTC 4\xD74 LDR fragment-shader encoder \u2014 WebGL2 port of astc4x4.wgsl (fast).\n//\n// One fragment per 4\xD74 block \u2192 16-byte block as 4 \xD7 u32 in outColor. Restricted\n// subset: single partition, no dual-plane, CEM 12 (LDR RGBA direct), 4\xD74 weight\n// grid with 2-bit weights (QUANT_4), 8-bit endpoints (QUANT_256). Fast path:\n// principal-axis seed (covariance power-iteration; bbox on degenerate blocks)\n// \u2192 one LSQ refit fused into a projection weight assignment (4 colinear\n// levels). Mirrors the `QUALITY_HIGH == 0` branch of astc4x4.wgsl; see that file\n// for the 128-bit block layout.\n//\n// Determinism note: floor(x + 0.5) replaces WGSL round() for the refit endpoints\n// (sub-LSB difference at exact .5 ties only).\n\nprecision highp float;\nprecision highp int;\n\nuniform sampler2D uSrc;\nuniform ivec2 uSrcSize;\nuniform int uFlipY;\n\nlayout(location = 0) out uvec4 outColor;\n\nivec4 gPixels[16];\nuint gIdx[16];\n\nstruct Fit { ivec4 e0; ivec4 e1; bool valid; };\n\nivec4 to8(vec4 v) {\n  return ivec4(clamp(floor(v * 255.0 + 0.5), vec4(0.0), vec4(255.0)));\n}\n\n// Principal colour axis of gPixels via covariance power-iteration, seeded\n// with the bbox diagonal. Returns a unit axis, or vec4(0.0) for a degenerate\n// (constant) block. The bbox diagonal alone is sign-blind and points across\n// anti-correlated data (normal maps, hue edges) instead of along it.\nvec4 principalAxis(vec4 mean, vec4 seed) {\n  vec4 c0v = vec4(0.0);\n  vec4 c1v = vec4(0.0);\n  vec4 c2v = vec4(0.0);\n  vec4 c3v = vec4(0.0);\n  for (int k = 0; k < 16; k++) {\n    vec4 d = vec4(gPixels[k]) - mean;\n    c0v += d.x * d;\n    c1v += d.y * d;\n    c2v += d.z * d;\n    c3v += d.w * d;\n  }\n  vec4 v = seed;\n  float len = length(v);\n  if (len < 1e-9) { return vec4(0.0); }\n  v /= len;\n  for (int it = 0; it < 8; it++) {\n    vec4 nv = vec4(dot(c0v, v), dot(c1v, v), dot(c2v, v), dot(c3v, v));\n    len = length(nv);\n    if (len < 1e-12) { return vec4(0.0); }\n    v = nv / len;\n  }\n  return v;\n}\n\n// Projection weight assignment over 4 levels (QUANT_4 \u2248 thirds), with the LSQ\n// normal-equation sums accumulated in the same pass for a fused refit.\nFit projAssign(ivec4 pe0, ivec4 pe1, bool fit) {\n  Fit res;\n  res.e0 = ivec4(0);\n  res.e1 = ivec4(0);\n  res.valid = false;\n  ivec4 dir = pe1 - pe0;\n  int dd = dir.x * dir.x + dir.y * dir.y + dir.z * dir.z + dir.w * dir.w;\n  if (dd == 0) {\n    for (int k = 0; k < 16; k++) { gIdx[k] = 0u; }\n    return res;\n  }\n  float inv = 3.0 / float(dd);\n  float sAA = 0.0, sBB = 0.0, sAB = 0.0;\n  vec4 sAV = vec4(0.0), sBV = vec4(0.0);\n  for (int k = 0; k < 16; k++) {\n    ivec4 q = gPixels[k] - pe0;\n    float proj = float(q.x * dir.x + q.y * dir.y + q.z * dir.z + q.w * dir.w) * inv;\n    float s = clamp(floor(proj + 0.5), 0.0, 3.0);\n    gIdx[k] = uint(s);\n    if (fit) {\n      vec4 v = vec4(gPixels[k]);\n      float b = s / 3.0;\n      float a = 1.0 - b;\n      sAA += a * a; sBB += b * b; sAB += a * b; sAV += a * v; sBV += b * v;\n    }\n  }\n  if (!fit) { return res; }\n  float det = sAA * sBB - sAB * sAB;\n  if (abs(det) < 1e-9) { return res; }\n  res.e0 = ivec4(clamp(floor((sBB * sAV - sAB * sBV) / det + 0.5), vec4(0.0), vec4(255.0)));\n  res.e1 = ivec4(clamp(floor((sAA * sBV - sAB * sAV) / det + 0.5), vec4(0.0), vec4(255.0)));\n  res.valid = true;\n  return res;\n}\n\nvoid writeBits(inout uint block[4], uint pos, uint nbits, uint value) {\n  uint v = value & ((1u << nbits) - 1u);\n  uint wordLo = pos / 32u;\n  uint bitLo = pos % 32u;\n  uint bitsInLo = min(nbits, 32u - bitLo);\n  uint maskLo = ((1u << bitsInLo) - 1u) << bitLo;\n  block[wordLo] = (block[wordLo] & ~maskLo) | ((v << bitLo) & maskLo);\n  if (bitsInLo < nbits) {\n    uint bitsInHi = nbits - bitsInLo;\n    uint maskHi = (1u << bitsInHi) - 1u;\n    uint valHi = v >> bitsInLo;\n    block[wordLo + 1u] = (block[wordLo + 1u] & ~maskHi) | (valHi & maskHi);\n  }\n}\n\nvoid main() {\n  ivec2 base = ivec2(gl_FragCoord.xy) * 4;\n  ivec2 maxXY = uSrcSize - ivec2(1);\n\n  ivec4 lo = ivec4(255);\n  ivec4 hi = ivec4(0);\n  ivec4 isum = ivec4(0);\n  for (int i = 0; i < 16; i++) {\n    ivec2 p = clamp(base + ivec2(i & 3, i >> 2), ivec2(0), maxXY);\n    int sy = (uFlipY != 0) ? (uSrcSize.y - 1 - p.y) : p.y;\n    ivec4 px = to8(texelFetch(uSrc, ivec2(p.x, sy), 0));\n    gPixels[i] = px;\n    lo = min(lo, px);\n    hi = max(hi, px);\n    isum += px;\n  }\n  vec4 mean = vec4(isum) / 16.0;\n\n  ivec4 e0 = lo;\n  ivec4 e1 = hi;\n  vec4 axis = principalAxis(mean, vec4(hi - lo));\n  if (dot(axis, axis) > 0.0) {\n    float tMin = 1e30;\n    float tMax = -1e30;\n    for (int k = 0; k < 16; k++) {\n      float t = dot(vec4(gPixels[k]) - mean, axis);\n      tMin = min(tMin, t);\n      tMax = max(tMax, t);\n    }\n    e0 = ivec4(clamp(floor(mean + tMin * axis + 0.5), vec4(0.0), vec4(255.0)));\n    e1 = ivec4(clamp(floor(mean + tMax * axis + 0.5), vec4(0.0), vec4(255.0)));\n  }\n  Fit r = projAssign(e0, e1, true);\n  if (r.valid) {\n    // Clamp the refit to the block bbox: on multi-cluster blocks the\n    // unconstrained LSQ solve extrapolates far outside the block's colours and\n    // the per-channel [0,255] clamp then bends the hue \u2014 fringe pixels decode\n    // to colours that exist nowhere in the block. Constraining to the bbox\n    // also measures better in plain SSE (+1.8 dB on the colour test card).\n    e0 = clamp(r.e0, lo, hi);\n    e1 = clamp(r.e1, lo, hi);\n    projAssign(e0, e1, false);\n  }\n\n  // Endpoint ordering so the decoder doesn't apply blue contraction.\n  if (e0.x + e0.y + e0.z > e1.x + e1.y + e1.z) {\n    ivec4 t = e0; e0 = e1; e1 = t;\n    for (int k = 0; k < 16; k++) { gIdx[k] = 3u - gIdx[k]; }\n  }\n\n  uint block[4];\n  block[0] = 0u; block[1] = 0u; block[2] = 0u; block[3] = 0u;\n  writeBits(block, 0u, 11u, 0x042u);\n  writeBits(block, 11u, 2u, 0u);\n  writeBits(block, 13u, 4u, 12u);\n  writeBits(block, 17u + 0u * 8u, 8u, uint(e0.x));\n  writeBits(block, 17u + 1u * 8u, 8u, uint(e1.x));\n  writeBits(block, 17u + 2u * 8u, 8u, uint(e0.y));\n  writeBits(block, 17u + 3u * 8u, 8u, uint(e1.y));\n  writeBits(block, 17u + 4u * 8u, 8u, uint(e0.z));\n  writeBits(block, 17u + 5u * 8u, 8u, uint(e1.z));\n  writeBits(block, 17u + 6u * 8u, 8u, uint(e0.w));\n  writeBits(block, 17u + 7u * 8u, 8u, uint(e1.w));\n\n  uint w3 = 0u;\n  for (int k = 0; k < 16; k++) {\n    uint w = gIdx[k] & 0x3u;\n    w3 = w3 | ((w & 1u) << (31u - 2u * uint(k))) | (((w >> 1u) & 1u) << (30u - 2u * uint(k)));\n  }\n  block[3] = w3;\n\n  outColor = uvec4(block[0], block[1], block[2], block[3]);\n}\n";
 
 // src/webgl/ASTC4x4WebGLEncoder.ts
 var ASTC4x4WebGLEncoder = class extends WebGLBlockEncoder {