gputex 0.3.1 → 0.3.3

package/dist/three.js

@@ -324,7 +324,7 @@ var Encoder = class {
 };
 
 // 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): 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        // 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).
@@ -449,8 +449,14 @@ 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);
@@ -632,7 +638,7 @@ 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): 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    // The refit is clamped to the block bbox: on multi-cluster blocks (a hard\n    // edge through two-colour noise) the unconstrained solve extrapolates far\n    // outside the block's colours and the per-channel [0,255] clamp then bends\n    // the hue \u2014 fringe pixels decode to colours that exist nowhere in the\n    // block. Constraining to the bbox also measures BETTER in plain SSE\n    // (+1.3 dB on the colour test card): the wild endpoints were losing more\n    // after quantisation + reassignment than the extrapolation ever bought.\n    let r = proj_fit(&pixels, lo, hi);\n    var ep0: Ep;\n    var ep1: Ep;\n    if (r.valid) { ep0 = pick_ep(clamp(r.e0, lo, hi)); ep1 = pick_ep(clamp(r.e1, lo, hi)); }\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";
 
 // src/bc7_fast_f16.wgsl
 var bc7_fast_f16_default = `// bc7 "fast" encoder \u2014 f16 variant (requires the shader-f16 feature).
@@ -753,12 +759,17 @@ fn encode(@builtin(global_invocation_id) gid: vec3<u32>) {
     pix[i] = px; lo = min(lo, px); hi = max(hi, px);
   }
 
-  // Seed fit from the raw bbox, then quantise the refit endpoints.
+  // Seed fit from the raw bbox, 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).
   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);   }
+  if (r.valid) { ep0 = pick_ep(clamp(r.e0, lo, hi)); ep1 = pick_ep(clamp(r.e1, lo, hi)); }
+  else         { ep0 = pick_ep(lo);                  ep1 = pick_ep(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).
@@ -829,7 +840,7 @@ 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): 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    // 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    let r = proj_fit(&pixels, lo, hi);\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).
@@ -925,10 +936,15 @@ fn encode(@builtin(global_invocation_id) gid: vec3<u32>) {
     pix[i] = px; lo = min(lo, px); hi = max(hi, px);
   }
 
+  // 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, lo, 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);
 
@@ -1204,7 +1220,7 @@ 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): 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    // 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 {
@@ -1242,7 +1258,7 @@ 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: 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    // 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 (+1.3 dB on the colour test card).\n    ep0 = pickEp(clamp(r.e0, lo, hi));\n    ep1 = pickEp(clamp(r.e1, lo, hi));\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";
 
 // src/webgl/BC7WebGLEncoder.ts
 var BC7WebGLEncoder = class extends WebGLBlockEncoder {
@@ -1261,7 +1277,7 @@ 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// 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    // 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 {
@@ -1446,19 +1462,153 @@ function padToBlockMultiple(level) {
   return { data: out, width: pw, height: ph };
 }
 
+// src/svg.ts
+function isSvgMarkup(source) {
+  return source.trimStart().startsWith("<");
+}
+function hasSvgExtension(url) {
+  return /\.svg$/i.test(url.split(/[?#]/, 1)[0]);
+}
+function isSvgBlob(blob) {
+  if (blob.type) {
+    return blob.type.split(";", 1)[0].trim().toLowerCase() === "image/svg+xml";
+  }
+  return typeof File !== "undefined" && blob instanceof File && hasSvgExtension(blob.name);
+}
+var ROOT_TAG_RE = /<svg(?=[\s/>])[^>]*>/;
+function getAttr(tag, name) {
+  const m = new RegExp(`\\s${name}\\s*=\\s*(?:"([^"]*)"|'([^']*)')`).exec(tag);
+  return m ? m[1] ?? m[2] ?? "" : null;
+}
+function removeAttr(tag, name) {
+  return tag.replace(new RegExp(`\\s${name}\\s*=\\s*(?:"[^"]*"|'[^']*')`, "g"), "");
+}
+function parseAbsoluteLength(value) {
+  if (value == null) return null;
+  const m = /^\s*\+?(\d+(?:\.\d+)?|\.\d+)(?:px)?\s*$/i.exec(value);
+  if (!m) return null;
+  const n = Number(m[1]);
+  return n > 0 ? n : null;
+}
+function parseSvgDimensions(svgText) {
+  const m = ROOT_TAG_RE.exec(svgText);
+  if (!m) return null;
+  const tag = m[0];
+  const dims = {
+    width: parseAbsoluteLength(getAttr(tag, "width")),
+    height: parseAbsoluteLength(getAttr(tag, "height")),
+    viewBoxWidth: null,
+    viewBoxHeight: null
+  };
+  const viewBox = getAttr(tag, "viewBox");
+  if (viewBox) {
+    const parts = viewBox.trim().split(/[\s,]+/).map(Number);
+    if (parts.length === 4 && parts.every(Number.isFinite) && parts[2] > 0 && parts[3] > 0) {
+      dims.viewBoxWidth = parts[2];
+      dims.viewBoxHeight = parts[3];
+    }
+  }
+  return dims;
+}
+function resolveSvgRasterSize(dims, size) {
+  if (size !== void 0 && typeof size === "object") {
+    const width = Math.round(size.width);
+    const height = Math.round(size.height);
+    if (!(width >= 1) || !(height >= 1)) {
+      throw new Error(`rasterizeSvg: svgSize must be \u22651\xD71 (got ${size.width}\xD7${size.height})`);
+    }
+    return { width, height };
+  }
+  let w = dims.width;
+  let h = dims.height;
+  if (dims.viewBoxWidth != null && dims.viewBoxHeight != null) {
+    if (w == null && h != null) w = h * dims.viewBoxWidth / dims.viewBoxHeight;
+    else if (h == null && w != null) h = w * dims.viewBoxHeight / dims.viewBoxWidth;
+    else if (w == null && h == null) {
+      w = dims.viewBoxWidth;
+      h = dims.viewBoxHeight;
+    }
+  }
+  if (typeof size === "number") {
+    if (!(size >= 1)) {
+      throw new Error(`rasterizeSvg: svgSize must be \u22651 (got ${size})`);
+    }
+    const aspect = w != null && h != null ? w / h : 1;
+    return aspect >= 1 ? { width: Math.round(size), height: Math.max(1, Math.round(size / aspect)) } : { width: Math.max(1, Math.round(size * aspect)), height: Math.round(size) };
+  }
+  if (w == null || h == null) {
+    throw new Error(
+      "rasterizeSvg: the SVG has no intrinsic size (no absolute width/height attributes and no viewBox) \u2014 pass svgSize to choose a rasterisation size"
+    );
+  }
+  return { width: Math.max(1, Math.round(w)), height: Math.max(1, Math.round(h)) };
+}
+function setSvgRootSize(svgText, width, height) {
+  const m = ROOT_TAG_RE.exec(svgText);
+  if (!m) {
+    throw new Error("rasterizeSvg: no <svg> root element found in source");
+  }
+  let tag = m[0];
+  const origWidth = parseAbsoluteLength(getAttr(tag, "width"));
+  const origHeight = parseAbsoluteLength(getAttr(tag, "height"));
+  const hasViewBox = getAttr(tag, "viewBox") != null;
+  tag = removeAttr(removeAttr(tag, "width"), "height");
+  let inject = ` width="${width}" height="${height}"`;
+  if (!hasViewBox && origWidth != null && origHeight != null) {
+    inject += ` viewBox="0 0 ${origWidth} ${origHeight}"`;
+  }
+  tag = `<svg${inject}${tag.slice("<svg".length)}`;
+  return svgText.slice(0, m.index) + tag + svgText.slice(m.index + m[0].length);
+}
+async function rasterizeSvg(source, options = {}) {
+  if (typeof Image === "undefined") {
+    throw new Error(
+      "rasterizeSvg: SVG rasterisation needs a DOM Image element and cannot run in this environment (e.g. a worker)"
+    );
+  }
+  const svgText = typeof source === "string" ? source : await source.text();
+  const dims = parseSvgDimensions(svgText);
+  if (!dims) {
+    throw new Error("rasterizeSvg: no <svg> root element found in source");
+  }
+  const { width, height } = resolveSvgRasterSize(dims, options.size);
+  const sized = setSvgRootSize(svgText, width, height);
+  const url = URL.createObjectURL(new Blob([sized], { type: "image/svg+xml;charset=utf-8" }));
+  try {
+    const img = new Image();
+    img.decoding = "async";
+    await new Promise((resolve, reject) => {
+      img.onload = () => resolve();
+      img.onerror = () => reject(new Error("rasterizeSvg: the browser failed to decode the SVG"));
+      img.src = url;
+    });
+    await img.decode().catch(() => {
+    });
+    const canvas = typeof OffscreenCanvas !== "undefined" ? new OffscreenCanvas(width, height) : Object.assign(document.createElement("canvas"), { width, height });
+    const ctx = canvas.getContext("2d");
+    if (!ctx) {
+      throw new Error("rasterizeSvg: no 2D context available");
+    }
+    ctx.drawImage(img, 0, 0, width, height);
+    return await createImageBitmap(canvas, { colorSpaceConversion: "none", premultiplyAlpha: "none" });
+  } finally {
+    URL.revokeObjectURL(url);
+  }
+}
+
 // src/three/compressTexture.ts
-import { LinearFilter as LinearFilter2, LinearSRGBColorSpace as LinearSRGBColorSpace2, RepeatWrapping as RepeatWrapping2, SRGBColorSpace as SRGBColorSpace2, Texture } from "three";
+import { ClampToEdgeWrapping as ClampToEdgeWrapping2, LinearFilter as LinearFilter2, LinearSRGBColorSpace as LinearSRGBColorSpace2, SRGBColorSpace as SRGBColorSpace2, Texture } from "three";
 
 // src/three/buildTexture.ts
 import { RED_GREEN_RGTC2_Format, RGBA_ASTC_4x4_Format, RGBA_BPTC_Format, RGBA_S3TC_DXT1_Format } from "three";
 
 // src/three/textureAssembly.ts
 import {
+  ClampToEdgeWrapping,
   CompressedTexture,
   LinearFilter,
   LinearMipmapLinearFilter,
   LinearSRGBColorSpace,
-  RepeatWrapping,
   SRGBColorSpace
 } from "three";
 function assembleCompressedTexture(levels, threeFormat, effectiveSrgb) {
@@ -1476,7 +1626,7 @@ function assembleCompressedTexture(levels, threeFormat, effectiveSrgb) {
   texture.magFilter = LinearFilter;
   texture.minFilter = levels.length > 1 ? LinearMipmapLinearFilter : LinearFilter;
   texture.generateMipmaps = false;
-  texture.wrapS = texture.wrapT = RepeatWrapping;
+  texture.wrapS = texture.wrapT = ClampToEdgeWrapping;
   texture.needsUpdate = true;
   texture.userData.logicalWidth = base.width;
   texture.userData.logicalHeight = base.height;
@@ -1518,27 +1668,49 @@ async function encodeToTexture(encoder, source, { colorSpace = "srgb", quality =
 }
 
 // src/three/compressTexture.ts
-async function sourceToBitmap(source) {
+async function sourceToBitmap(source, svgSize) {
   const opts = {
     colorSpaceConversion: "none",
     premultiplyAlpha: "none"
   };
   if (typeof source === "string") {
+    if (isSvgMarkup(source)) {
+      return rasterizeSvg(source, { size: svgSize });
+    }
     const resp = await fetch(source);
     if (!resp.ok) {
       throw new Error(`compressTexture: fetch ${source} failed (${resp.status})`);
     }
     const blob = await resp.blob();
+    if (isSvgBlob(blob) || !isImageMimeType(blob.type) && hasSvgExtension(source)) {
+      return rasterizeSvg(blob, { size: svgSize });
+    }
     return createImageBitmap(blob, opts);
   }
   if (source instanceof Blob) {
+    if (isSvgBlob(source)) {
+      return rasterizeSvg(source, { size: svgSize });
+    }
     return createImageBitmap(source, opts);
   }
   if (source instanceof ImageBitmap) {
     return source;
   }
+  if (typeof HTMLImageElement !== "undefined" && source instanceof HTMLImageElement) {
+    const src = source.currentSrc || source.src;
+    if (src && (hasSvgExtension(src) || /^data:image\/svg\+xml/i.test(src))) {
+      const resp = await fetch(src);
+      if (!resp.ok) {
+        throw new Error(`compressTexture: fetch ${src} failed (${resp.status})`);
+      }
+      return rasterizeSvg(await resp.blob(), { size: svgSize });
+    }
+  }
   return createImageBitmap(source, opts);
 }
+function isImageMimeType(type) {
+  return /^image\//i.test(type) && !/svg/i.test(type);
+}
 function bitmapToMipLevel(bitmap, flipY) {
   const w = bitmap.width, h = bitmap.height;
   const canvas = typeof OffscreenCanvas !== "undefined" ? new OffscreenCanvas(w, h) : Object.assign(document.createElement("canvas"), { width: w, height: h });
@@ -1562,7 +1734,7 @@ function wrapUncompressed(bitmap, srgb, flipY) {
   tex.colorSpace = srgb ? SRGBColorSpace2 : LinearSRGBColorSpace2;
   tex.magFilter = LinearFilter2;
   tex.minFilter = LinearFilter2;
-  tex.wrapS = tex.wrapT = RepeatWrapping2;
+  tex.wrapS = tex.wrapT = ClampToEdgeWrapping2;
   tex.generateMipmaps = false;
   tex.flipY = flipY;
   tex.needsUpdate = true;
@@ -1573,6 +1745,7 @@ async function compressTexture(source, options = {}) {
     hint = "color",
     preferredFormat,
     colorSpace = "srgb",
+    svgSize,
     flipY = true,
     mipmaps = false,
     quality = "fast",
@@ -1580,7 +1753,7 @@ async function compressTexture(source, options = {}) {
     adapter: providedAdapter
   } = options;
   const srgb = colorSpace === "srgb";
-  const bitmap = await sourceToBitmap(source);
+  const bitmap = await sourceToBitmap(source, svgSize);
   const viaWebGPU = await encodeViaWebGPU();
   if (viaWebGPU) return viaWebGPU;
   const viaWebGL = encodeViaWebGL();
@@ -1750,6 +1923,12 @@ var GputexLoader = class extends Loader {
   preferredFormat;
   /** Pick the sRGB or linear variant of the chosen format. Default 'srgb'. */
   colorSpace = "srgb";
+  /**
+   * Rasterisation size for SVG URLs — a number (longest side, aspect
+   * preserved) or exact `{ width, height }`. Default: the SVG's intrinsic
+   * size. See `CompressOptions.svgSize`.
+   */
+  svgSize;
   /** Flip the image vertically before encoding. Default true (matches Three.js convention). */
   flipY = true;
   /** Generate + encode a full mip chain. Default false. */
@@ -1781,6 +1960,7 @@ var GputexLoader = class extends Loader {
       hint: this.hint,
       preferredFormat: this.preferredFormat,
       colorSpace: this.colorSpace,
+      svgSize: this.svgSize,
       flipY: this.flipY,
       mipmaps: this.mipmaps,
       quality: this.quality,
@@ -1836,6 +2016,7 @@ export {
   getSharedWebGLContext,
   isWebGLAvailable,
   padToBlockMultiple,
+  rasterizeSvg,
   selectFormat,
   selectWebGLFormat,
   threeFormatFor