@mcptoolshop/style-dataset-lab 3.4.0 → 3.5.0

package/lib/asset-masks.js

@@ -0,0 +1,323 @@
+/**
+ * Asset-lane color math — the palette-conformance gate ported
+ * formula-faithfully from facet's `tools/palette_gate.py` (read-only port,
+ * 2026-08-04), plus categorical class-share measurement for declared
+ * render-space channels.
+ *
+ * Port fidelity notes (checked against the source, not paraphrased):
+ *   - sRGB → linear → XYZ(D65) → Lab with the exact constants palette_gate.py
+ *     uses (which itself matches facet's e08_deltaE.py).
+ *   - C* = hypot(a*, b*); hue = atan2(b*, a*) in degrees, mod 360.
+ *   - A band with lo <= hi is [lo, hi]; a band with lo > hi WRAPS AROUND 0°.
+ *   - off-palette = (C* > min_chroma) AND (not in any band) AND inside the
+ *     EXACT silhouette (mask value > 0.5 of full scale — a keyed mask would
+ *     let background into the measurement; the E01 / E08-A2 failure).
+ *   - TWO numbers, because they are two different failures: total off-palette
+ *     (an antialiasing-spread diagnostic — its percentage bound is WITHDRAWN
+ *     per facet 2026-08-04 and gates nothing unless the palette declares one)
+ *     and the LARGEST 4-CONNECTED COMPONENT (an invented garment is one big
+ *     blob), which is the load-bearing gate.
+ *   - The chroma floor is load-bearing, not a tuning knob: hue is undefined
+ *     at low chroma (W3's steel greatsword sits at C* 1.6–2.8 inside the
+ *     blue band; without the floor it would flag on every view).
+ *
+ * Pure functions over decoded images (lib/png-meta.js output). No I/O, no
+ * channel semantics — the palette, bands, classes and tolerances all arrive
+ * from the asset's manifest. (Parnas boundary: this module knows color math;
+ * lib/asset-source.js knows the contract.)
+ */
+
+import { inputError } from './errors.js';
+
+// ─── sRGB → Lab (exact palette_gate.py constants) ────────────────────
+
+const XYZ_M = [
+  [0.4124564, 0.3575761, 0.1804375],
+  [0.2126729, 0.7151522, 0.0721750],
+  [0.0193339, 0.1191920, 0.9503041],
+];
+const WHITE = [0.95047, 1.0, 1.08883];
+const LAB_E = 216 / 24389;
+const LAB_K = 24389 / 27;
+
+function srgbToLinear(c) {
+  return c <= 0.04045 ? c / 12.92 : ((c + 0.055) / 1.055) ** 2.4;
+}
+
+function labF(t) {
+  return t > LAB_E ? Math.cbrt(t) : (LAB_K * t + 16) / 116;
+}
+
+/**
+ * Convert one sRGB pixel (0–255 ints) to Lab.
+ * @returns {[number, number, number]} [L, a, b]
+ */
+export function srgbToLab(r8, g8, b8) {
+  const r = srgbToLinear(r8 / 255);
+  const g = srgbToLinear(g8 / 255);
+  const b = srgbToLinear(b8 / 255);
+  const x = (XYZ_M[0][0] * r + XYZ_M[0][1] * g + XYZ_M[0][2] * b) / WHITE[0];
+  const y = (XYZ_M[1][0] * r + XYZ_M[1][1] * g + XYZ_M[1][2] * b) / WHITE[1];
+  const z = (XYZ_M[2][0] * r + XYZ_M[2][1] * g + XYZ_M[2][2] * b) / WHITE[2];
+  const fx = labF(x);
+  const fy = labF(y);
+  const fz = labF(z);
+  return [116 * fy - 16, 500 * (fx - fy), 200 * (fy - fz)];
+}
+
+// ─── Decoded-image pixel access ──────────────────────────────────────
+
+/**
+ * Read pixel i of a decoded PNG as [r, g, b] (0–255), independent of the
+ * source color type (grayscale, RGB, RGBA, or indexed via its PLTE).
+ */
+export function rgbAt(img, i) {
+  switch (img.colorType) {
+    case 0: {
+      const v = img.data[i];
+      return [v, v, v];
+    }
+    case 2: {
+      const o = i * 3;
+      return [img.data[o], img.data[o + 1], img.data[o + 2]];
+    }
+    case 6: {
+      const o = i * 4;
+      return [img.data[o], img.data[o + 1], img.data[o + 2]];
+    }
+    case 3: {
+      const entry = img.palette[img.data[i]];
+      return [entry[0], entry[1], entry[2]];
+    }
+    default:
+      throw inputError('PNG_UNSUPPORTED', `rgbAt: unsupported color type ${img.colorType}.`);
+  }
+}
+
+/** Alpha of pixel i (255 when the image has no alpha channel). */
+export function alphaAt(img, i) {
+  return img.colorType === 6 ? img.data[i * 4 + 3] : 255;
+}
+
+/**
+ * Binarize a decoded grayscale silhouette: value > 127 (i.e. > 0.5 of full
+ * scale, matching palette_gate.py's `> 0.5` on a /255 float) = figure.
+ * @returns {Uint8Array} 0/1 per pixel
+ */
+export function silhouetteToMask(img, label = 'silhouette') {
+  if (img.colorType !== 0) {
+    throw inputError(
+      'ASSET_ENCODING_MISMATCH',
+      `${label}: silhouette masks must be grayscale PNGs (color type 0), got ${img.colorType}.`
+    );
+  }
+  const mask = new Uint8Array(img.width * img.height);
+  for (let i = 0; i < mask.length; i++) mask[i] = img.data[i] > 127 ? 1 : 0;
+  return mask;
+}
+
+// ─── Connected components (4-connectivity, matching scipy.ndimage.label
+//     with its default cross-shaped structure) ────────────────────────
+
+/**
+ * Size in pixels of the largest 4-connected component of a 0/1 mask.
+ */
+export function largestComponent4(mask, width, height) {
+  const visited = new Uint8Array(mask.length);
+  const stack = new Int32Array(mask.length);
+  let largest = 0;
+  for (let start = 0; start < mask.length; start++) {
+    if (!mask[start] || visited[start]) continue;
+    let top = 0;
+    stack[top++] = start;
+    visited[start] = 1;
+    let size = 0;
+    while (top > 0) {
+      const i = stack[--top];
+      size++;
+      const x = i % width;
+      const y = (i - x) / width;
+      if (x > 0 && mask[i - 1] && !visited[i - 1]) { visited[i - 1] = 1; stack[top++] = i - 1; }
+      if (x < width - 1 && mask[i + 1] && !visited[i + 1]) { visited[i + 1] = 1; stack[top++] = i + 1; }
+      if (y > 0 && mask[i - width] && !visited[i - width]) { visited[i - width] = 1; stack[top++] = i - width; }
+      if (y < height - 1 && mask[i + width] && !visited[i + width]) { visited[i + width] = 1; stack[top++] = i + width; }
+    }
+    if (size > largest) largest = size;
+  }
+  return largest;
+}
+
+// ─── The palette gate ────────────────────────────────────────────────
+
+function median(values) {
+  if (values.length === 0) return 0;
+  const sorted = Float64Array.from(values).sort();
+  const mid = sorted.length >> 1;
+  return sorted.length % 2 ? sorted[mid] : (sorted[mid - 1] + sorted[mid]) / 2;
+}
+
+function round(v, places) {
+  const f = 10 ** places;
+  return Math.round(v * f) / f;
+}
+
+/**
+ * Run the ported palette-conformance gate on one render.
+ *
+ * @param {Object} renderImg — decoded render (lib/png-meta.js decodePng)
+ * @param {Object} silhouetteImg — decoded grayscale silhouette, same dims
+ * @param {Object} palette — the manifest's palette block:
+ *   {min_chroma, allowed_bands: [{name, hue_deg: [lo, hi]}],
+ *    gate: {max_offpalette_pct: number|null, max_offpalette_blob_px: number}}
+ * @param {string} [label] — for error messages
+ * @returns {{figure_px:number, offpalette_px:number, offpalette_pct:number,
+ *            largest_blob_px:number, pass:boolean, withdrawn_pct_bound:boolean,
+ *            dominant:Object|null}}
+ */
+export function paletteGate(renderImg, silhouetteImg, palette, label = 'render') {
+  if (renderImg.width !== silhouetteImg.width || renderImg.height !== silhouetteImg.height) {
+    // The parallel-mask ANDON from palette_gate.py: a mask that does not
+    // match its image measures the wrong pixels silently.
+    throw inputError(
+      'ASSET_MASK_MISMATCH',
+      `${label}: silhouette is ${silhouetteImg.width}x${silhouetteImg.height} but the render is ${renderImg.width}x${renderImg.height}.`
+    );
+  }
+
+  const mask = silhouetteToMask(silhouetteImg, `${label} silhouette`);
+  const bands = palette.allowed_bands.map((b) => ({
+    lo: Number(b.hue_deg[0]),
+    hi: Number(b.hue_deg[1]),
+  }));
+  const cmin = Number(palette.min_chroma);
+  const maxPct = palette.gate.max_offpalette_pct == null ? null : Number(palette.gate.max_offpalette_pct);
+  const maxBlob = Number(palette.gate.max_offpalette_blob_px);
+
+  const n = renderImg.width * renderImg.height;
+  const off = new Uint8Array(n);
+  let figurePx = 0;
+  let offPx = 0;
+  const offChroma = [];
+  const offHue = [];
+  const offRgb = [];
+
+  for (let i = 0; i < n; i++) {
+    if (!mask[i]) continue;
+    figurePx++;
+    const [r, g, b] = rgbAt(renderImg, i);
+    const [, a, bb] = srgbToLab(r, g, b);
+    const chroma = Math.hypot(a, bb);
+    if (chroma <= cmin) continue; // hue undefined below the floor
+    let hue = (Math.atan2(bb, a) * 180) / Math.PI;
+    hue = ((hue % 360) + 360) % 360;
+    let inBand = false;
+    for (const { lo, hi } of bands) {
+      if (lo <= hi ? hue >= lo && hue <= hi : hue >= lo || hue <= hi) { inBand = true; break; }
+    }
+    if (inBand) continue;
+    off[i] = 1;
+    offPx++;
+    offChroma.push(chroma);
+    offHue.push(hue);
+    offRgb.push([r, g, b]);
+  }
+
+  const pct = (100 * offPx) / Math.max(figurePx, 1);
+  const blob = offPx > 0 ? largestComponent4(off, renderImg.width, renderImg.height) : 0;
+  const pass = blob <= maxBlob && (maxPct === null || pct <= maxPct);
+
+  // Dominant off-palette mass, so a failure is diagnosable, not just flagged
+  // (same 36-bin histogram + medians as the source).
+  let dominant = null;
+  if (offPx > 0) {
+    const hist = new Array(36).fill(0);
+    for (const h of offHue) hist[Math.min(35, Math.floor(h / 10))]++;
+    let top = 0;
+    for (let i = 1; i < 36; i++) if (hist[i] > hist[top]) top = i;
+    dominant = {
+      hue_lo: top * 10,
+      hue_hi: (top + 1) * 10,
+      share_of_offpalette: round((hist[top] / offPx) * 100, 1),
+      median_chroma: round(median(offChroma), 1),
+      median_rgb: [
+        Math.round(median(offRgb.map((p) => p[0]))),
+        Math.round(median(offRgb.map((p) => p[1]))),
+        Math.round(median(offRgb.map((p) => p[2]))),
+      ],
+    };
+  }
+
+  return {
+    figure_px: figurePx,
+    offpalette_px: offPx,
+    offpalette_pct: round(pct, 4),
+    largest_blob_px: blob,
+    pass,
+    withdrawn_pct_bound: maxPct === null,
+    dominant,
+  };
+}
+
+// ─── Categorical class shares ────────────────────────────────────────
+
+/**
+ * Measure per-class coverage of a categorical render-space channel inside
+ * the silhouette. Shares are computed for EVERY declared class — this module
+ * never knows which class matters; the manifest's consumers do.
+ *
+ * `filter: "nearest"` ⇒ exact color match (tolerance 0).
+ * `filter: "linear"`  ⇒ nearest declared class within
+ *                       `classification_tolerance` (RGB euclidean); pixels
+ *                       beyond it are honestly `unclassified`.
+ * Fully transparent pixels (alpha 0) are excluded — background, not class.
+ *
+ * @param {Object} channelImg — decoded channel instance
+ * @param {Object} silhouetteImg — decoded grayscale silhouette, same dims
+ * @param {Object} decl — the channel declaration:
+ *   {palette: [{name, rgb: [r,g,b]}], filter, classification_tolerance?}
+ * @param {string} [label]
+ * @returns {{figure_px:number,
+ *            classes:Object<string,{px:number, share:number}>,
+ *            unclassified:{px:number, share:number}}}
+ */
+export function classShares(channelImg, silhouetteImg, decl, label = 'channel') {
+  if (channelImg.width !== silhouetteImg.width || channelImg.height !== silhouetteImg.height) {
+    throw inputError(
+      'ASSET_MASK_MISMATCH',
+      `${label}: channel is ${channelImg.width}x${channelImg.height} but the silhouette is ${silhouetteImg.width}x${silhouetteImg.height}.`
+    );
+  }
+  const mask = silhouetteToMask(silhouetteImg, `${label} silhouette`);
+  const tolerance = decl.filter === 'nearest' ? 0 : Number(decl.classification_tolerance ?? 0);
+  const tol2 = tolerance * tolerance;
+  const classes = decl.palette.map((c) => ({ name: c.name, rgb: c.rgb, px: 0 }));
+
+  let figurePx = 0;
+  let unclassified = 0;
+  const n = channelImg.width * channelImg.height;
+
+  for (let i = 0; i < n; i++) {
+    if (!mask[i]) continue;
+    if (alphaAt(channelImg, i) === 0) continue; // transparent = background
+    figurePx++;
+    const [r, g, b] = rgbAt(channelImg, i);
+    let best = -1;
+    let bestD = Infinity;
+    for (let c = 0; c < classes.length; c++) {
+      const [cr, cg, cb] = classes[c].rgb;
+      const d = (r - cr) * (r - cr) + (g - cg) * (g - cg) + (b - cb) * (b - cb);
+      if (d < bestD) { bestD = d; best = c; }
+    }
+    if (best >= 0 && bestD <= tol2) classes[best].px++;
+    else unclassified++;
+  }
+
+  const share = (px) => round((100 * px) / Math.max(figurePx, 1), 4);
+  const out = {};
+  for (const c of classes) out[c.name] = { px: c.px, share: share(c.px) };
+  return {
+    figure_px: figurePx,
+    classes: out,
+    unclassified: { px: unclassified, share: share(unclassified) },
+  };
+}