rig-c 0.0.0-stage → 2.20.4

package/src/slots.ts

@@ -0,0 +1,892 @@
+/**
+ * Slot tracking — which part landed where, and when that cannot be said.
+ *
+ * ## Two matchers, in order, and a cap on both
+ *
+ * The cheap matcher labels the reference frame's connected components and asks
+ * which one each of the candidate's slots landed on. It is right whenever the
+ * parts of a shot are separate blobs, and it has three failure modes that honest
+ * ladder runs hit head-on (issues #34 and #37) — all three the same mistake, which
+ * is treating a blob as a part:
+ *
+ * - **Parts that touch label as one component.** Rung 4 is a disc with five chain
+ *   links hanging off it; they touch in every frame of every animation, so every
+ *   slot came back "ambiguous" and the run produced **no drift table at all**.
+ * - **Nearest-centroid has no notion of "too far to be the same thing".** Rung 5's
+ *   4 px ball, on the frames where it rests against the course and has no component
+ *   of its own, matched the floating girder 47 px away — and the summary line
+ *   reported **48.3 px of drift for a 4 px ball**, unflagged.
+ * - **A blob one part dominates passes for that part.** Rung 2's reference merges
+ *   the course, the water, the panel and both rings into one component in which the
+ *   course is 81 % of the ink, so it is 1.24x the course's own and no wider than its
+ *   box — both merge tests see nothing, and the run reported *"course drift
+ *   11.2 px"*, the distance to a five-part centroid (issue #37). `occupantsOf`
+ *   answers that one with the label map: anything else the candidate drew inside the
+ *   blob makes the blob's centroid nobody's position.
+ *
+ * So: components first, and when a component cannot be attributed to one slot, the
+ * slot's own rendered quad is **template-matched** against the reference in a
+ * window around where the candidate drew it. Touching parts stop being fatal —
+ * a link that overlaps its neighbour still correlates against its own pixels.
+ *
+ * ⛔ **Its own pixels means the ones you can see.** The template is the slot drawn
+ * alone and the reference is a composite, so a template pixel the candidate itself
+ * draws something over matches nothing at the true offset — and a residual that is
+ * there at every offset is not evidence about position. Left in, it *moves the
+ * answer*: issue #698 measured four of the seven shipped examples reporting
+ * 0.8–2.2 px against frames rendered from themselves, and on every one the score
+ * field's own whole-pixel minimum sat off the identity offset because the covered
+ * samples' gradient outvoted a shallow basin. `visibleField` drops them, and the
+ * minimum comes back to the origin on all seven.
+ *
+ * ⛔ And both matchers are capped by `searchRadius`: a part may be displaced by
+ * about its own size and still be the same part in the picture, and past that the
+ * honest report is **no match**. A number that is not a measurement of the slot it
+ * is printed beside is worse than a blank, because it is actionable and wrong.
+ */
+import { Plate, type RGBA } from '../tools/plate.ts';
+import { backgroundDistance, isContent } from './framing.ts';
+import {
+  frameGeometry,
+  pageFor,
+  projector,
+  rasterisePiece,
+  type Frame,
+  type Footprint,
+  type Viewport,
+} from './render_shared.ts';
+
+/** Components smaller than this are antialiasing crumbs, not parts. */
+const MIN_COMPONENT_PIXELS = 4;
+/** A second component this close to the nearest makes a component match a guess. */
+const AMBIGUITY_RATIO = 1.25;
+/**
+ * How much bigger than the slot a component may be and still be *that slot's*.
+ *
+ * Above it, the slot is inside something larger — the reference merged it with a
+ * neighbour, or drew it behind one — and the component's centroid is the merged
+ * blob's, not the part's.
+ *
+ * ⚠️ Pixel count alone is not enough, and rung 3's transcription is the proof: at
+ * `heavy/f0009` the pendulum touches the block, the two label as one 1227 px blob,
+ * and the pendulum's own 836 px makes that only 1.47x — under any ratio loose
+ * enough to tolerate antialiasing. The blob's **bounding box** gives it away (60 px
+ * wide against the slot's 39), so both tests have to pass. Erring towards "merged"
+ * is the safe direction: a slot wrongly called merged still gets a drift from the
+ * template matcher, where a merged blob wrongly called the slot's own reports the
+ * blob's centroid as the part's position.
+ */
+const MERGE_RATIO = 1.6;
+/** How much wider or taller than the slot a component may be, as a fraction. */
+const MERGE_MARGIN = 0.1;
+/** ...and never less than this, so antialiasing alone cannot trip it. */
+const MERGE_MARGIN_PIXELS = 2;
+/** Displacement bounds, in frame pixels: never search less, never search more. */
+const MIN_SEARCH_RADIUS = 4;
+const MAX_SEARCH_RADIUS = 32;
+/** Search radius as a fraction of the slot's own long side. */
+const SEARCH_SPAN = 0.75;
+/** How many template pixels a correlation samples, at most. */
+const MAX_SAMPLES = 256;
+/** A rival peak must be at least this far from the winner to count as a rival. */
+const RIVAL_GAP = 3;
+/**
+ * How distinctive a correlation peak must be before its offset is reported.
+ *
+ * ⭐ It rises with the displacement being claimed, and that is the whole point. A
+ * peak sitting where the candidate already drew the slot is only confirming a
+ * position, so a weak peak is enough; a peak claiming the part is most of a search
+ * radius away is claiming something big, and the bar for it is correspondingly
+ * high. Rung 4 is the case that fixed the constant: `chain1` — a 9x27 sliver in a
+ * chain of near-identical links — correlated 26 px away at confidence 0.16, and
+ * that number went straight into the summary line as the run's worst drift. Under
+ * this rule it needs 0.60 and comes back as **no match**, which is the honest
+ * answer for a repetitive structure.
+ */
+const MIN_CONFIDENCE = 0.15;
+/** How much more distinctive a peak must be at a full radius out than at zero. */
+const CONFIDENCE_SLOPE = 0.45;
+/** The residual must be under this fraction of the slot's own contrast. */
+const MAX_RESIDUAL_FRACTION = 0.5;
+
+export interface Component {
+  pixels: number;
+  cx: number;
+  cy: number;
+  minX: number;
+  minY: number;
+  maxX: number;
+  maxY: number;
+}
+
+/**
+ * The reference frame's components, and which one each pixel belongs to.
+ *
+ * The label map is what makes "is anything ELSE inside this blob?" a measurement
+ * rather than a guess from bounding boxes — see `occupantsOf`. It is indexed
+ * row-major on the frame's own grid, and `-1` is background or a component too
+ * small to be a part.
+ */
+export interface ComponentField {
+  components: Component[];
+  labels: Int32Array;
+  width: number;
+  height: number;
+}
+
+/**
+ * Connected components of "not the background colour", 8-connected.
+ *
+ * 8-connected rather than 4: a thin diagonal — a bar, a stick, a shadow's edge —
+ * breaks into a dotted line under 4-connectivity, and then one part reads as
+ * twenty and every match is ambiguous for a reason that is about the labeller.
+ */
+export function componentsOf(plate: Plate, background: RGBA): Component[] {
+  return componentField(plate, background).components;
+}
+
+/** The same labelling, with the map kept — see `ComponentField`. */
+export function componentField(plate: Plate, background: RGBA): ComponentField {
+  const { width, height } = plate;
+  const label = new Int32Array(width * height).fill(-1);
+  const out: Component[] = [];
+  const stack: number[] = [];
+  for (let y0 = 0; y0 < height; y0++) {
+    for (let x0 = 0; x0 < width; x0++) {
+      const seed = y0 * width + x0;
+      if (label[seed] !== -1 || !isContent(plate, x0, y0, background)) continue;
+      const id = out.length;
+      label[seed] = id;
+      stack.push(seed);
+      let pixels = 0;
+      let sx = 0;
+      let sy = 0;
+      let minX = width;
+      let minY = height;
+      let maxX = -1;
+      let maxY = -1;
+      while (stack.length > 0) {
+        const at = stack.pop() as number;
+        const x = at % width;
+        const y = (at - x) / width;
+        pixels++;
+        sx += x + 0.5;
+        sy += y + 0.5;
+        if (x < minX) minX = x;
+        if (x > maxX) maxX = x;
+        if (y < minY) minY = y;
+        if (y > maxY) maxY = y;
+        for (let dy = -1; dy <= 1; dy++) {
+          for (let dx = -1; dx <= 1; dx++) {
+            const nx = x + dx;
+            const ny = y + dy;
+            if (nx < 0 || ny < 0 || nx >= width || ny >= height) continue;
+            const n = ny * width + nx;
+            if (label[n] !== -1 || !isContent(plate, nx, ny, background)) continue;
+            label[n] = id;
+            stack.push(n);
+          }
+        }
+      }
+      out.push({ pixels, cx: sx / pixels, cy: sy / pixels, minX, minY, maxX: maxX + 1, maxY: maxY + 1 });
+    }
+  }
+  // Crumbs out, biggest first — and the label map carried through the reorder, so
+  // a label is always an index into the array the caller is handed. Renumbering
+  // rather than sorting the map is what keeps the two from drifting apart.
+  const keep = out.map((c, id) => ({ c, id })).filter(({ c }) => c.pixels >= MIN_COMPONENT_PIXELS);
+  keep.sort((a, b) => b.c.pixels - a.c.pixels);
+  const renumbered = new Int32Array(out.length).fill(-1);
+  keep.forEach(({ id }, index) => {
+    renumbered[id] = index;
+  });
+  for (let at = 0; at < label.length; at++) label[at] = label[at] < 0 ? -1 : renumbered[label[at]];
+  return { components: keep.map(({ c }) => c), labels: label, width, height };
+}
+
+/**
+ * Which drawn slots' ink sits inside each component, by centroid.
+ *
+ * ## Why this exists: a blob one part dominates is still a blob
+ *
+ * The size and bounding-box tests below catch a merge when the merged neighbour is
+ * a material fraction of the blob. They cannot catch the case issue #37 filed:
+ * rung 2's reference merges the course, the water, the panel and both rings into
+ * one component, and the **course is 81 % of it**, so the blob is only 1.24x the
+ * course's own ink and barely wider than the course's own box. It passed every
+ * test, and the summary line reported *"course drift 11.2 px"* — the distance from
+ * the course's centroid to the centroid of a blob holding four other parts, which
+ * is not a measurement of the course at all.
+ *
+ * One label lookup per drawn slot answers it exactly: if anything else the
+ * candidate drew lands on this component's own pixels, the component is more than
+ * one part and its centroid is nobody's position. That is the same judgement the
+ * two-claimants rule below already makes; what was missing is that a slot which
+ * never got as far as *claiming* the blob — because it was refused for being 13x
+ * too small, or diverted to the template matcher — still proves the blob is shared.
+ */
+function occupantsOf(field: ComponentField, footprints: Map<string, Footprint>): Map<number, string[]> {
+  const out = new Map<number, string[]>();
+  for (const [slot, foot] of footprints) {
+    if (foot.pixels === 0) continue;
+    const x = Math.floor(foot.cx);
+    const y = Math.floor(foot.cy);
+    if (x < 0 || y < 0 || x >= field.width || y >= field.height) continue;
+    const label = field.labels[y * field.width + x];
+    if (label < 0) continue;
+    const seen = out.get(label) ?? [];
+    seen.push(slot);
+    out.set(label, seen);
+  }
+  return out;
+}
+
+/** How the drift beside a slot was arrived at. `none` means it could not be. */
+export type MatchMethod = 'component' | 'template' | 'none';
+
+export interface SlotTrack {
+  slot: string;
+  candidate: { cx: number; cy: number; width: number; height: number; pixels: number } | null;
+  method: MatchMethod;
+  /** The reference component this slot was matched to — component matches only. */
+  reference: { cx: number; cy: number; width: number; height: number; pixels: number } | null;
+  /** Centroid distance in frame pixels, or the correlation offset's length. */
+  drift: number | null;
+  /** The same displacement with its direction, reference minus candidate. */
+  driftX: number | null;
+  driftY: number | null;
+  /** Bounding-box differences — component matches only, where a bbox is known. */
+  widthDrift: number | null;
+  heightDrift: number | null;
+  /** 0..1 for a template match: how much better the winner is than its best rival. */
+  confidence: number | null;
+  /** How far the match was allowed to look, in frame pixels. */
+  searchRadius: number | null;
+  /**
+   * The whole-pixel offset the sub-pixel step refined — template matches only.
+   *
+   * It is the half of the answer that is a *displacement*: `(0, 0)` says the
+   * correlation put the part where the candidate drew it and everything in
+   * `drift` is the parabola's own residue.
+   */
+  wholePixel: { dx: number; dy: number } | null;
+  /** Set when the drift is not a measurement of this slot, saying why. */
+  ambiguity: string | null;
+}
+
+/**
+ * The most this match could have read, given the winner it found.
+ *
+ * `hypot(|dx| + SUBPIXEL_CLAMP, |dy| + SUBPIXEL_CLAMP)`, because the sub-pixel
+ * step is clamped to half a pixel on each axis. A reader comparing a drift to a
+ * floor needs it beside the figure: 0.44 px under a bound of 0.71 px is the
+ * instrument at rest, and 0.44 px under a bound of 1.58 px is a part that moved a
+ * whole pixel and came most of the way back. `null` for a component match — a
+ * distance between two centroids is bounded by nothing but the search radius.
+ *
+ * 🔒 **A function and not a field**, and the red-first run is why. It was stored
+ * on the track first, and `C29`'s plant — one track's whole-pixel winner forced a
+ * pixel off the identity — printed *"bounded by 0.71 px — the correlation moved
+ * this slot by (1, 0) whole pixel(s)"*: the sentence and the figure beside it came
+ * from two fields that could disagree, which is this repository's own antipattern
+ * with a drift figure attached to it. One derivation, one place.
+ */
+export function driftBound(track: SlotTrack): number | null {
+  if (track.method !== 'template' || track.wholePixel === null) return null;
+  return Math.hypot(Math.abs(track.wholePixel.dx) + SUBPIXEL_CLAMP, Math.abs(track.wholePixel.dy) + SUBPIXEL_CLAMP);
+}
+
+/**
+ * How far this slot may be displaced and still be the same thing in the picture.
+ *
+ * Tied to the slot's own size, because that is what makes the bound mean
+ * something: past about its own long side a part no longer overlaps where it was,
+ * and a correlation peak out there is another object, not this one moved.
+ */
+export function searchRadius(width: number, height: number): number {
+  const span = Math.round(Math.max(width, height) * SEARCH_SPAN);
+  return Math.max(MIN_SEARCH_RADIUS, Math.min(MAX_SEARCH_RADIUS, span));
+}
+
+/** What the template matcher needs to draw one slot on its own. */
+export interface SlotSource {
+  frame: Frame;
+  pages: Map<string, Plate>;
+  viewport: Viewport;
+  background: RGBA;
+  reference: Plate;
+}
+
+// ---------------------------------------------------------------------------
+// the component pass
+// ---------------------------------------------------------------------------
+
+interface Pending {
+  track: SlotTrack;
+  foot: Footprint;
+  claimed: Component | null;
+}
+
+/**
+ * Match each drawn slot to a reference component, then template-match the rest.
+ *
+ * Returns the tracks in slot-name order, plus how many reference components the
+ * candidate accounted for — a component nothing overlaps is something in the shot
+ * the candidate has not drawn.
+ */
+export function matchSlots(
+  footprints: Map<string, Footprint>,
+  field: ComponentField,
+  source: SlotSource | null,
+): { tracks: SlotTrack[]; matchedComponents: number } {
+  const components = field.components;
+  const pending: Pending[] = [];
+  const takenBy = new Map<Component, string[]>();
+  const occupants = occupantsOf(field, footprints);
+  /** Which component each one is, so an occupancy list can be looked up by it. */
+  const idOf = new Map<Component, number>();
+  components.forEach((component, id) => idOf.set(component, id));
+
+  for (const [slot, foot] of [...footprints].sort((a, b) => a[0].localeCompare(b[0]))) {
+    const track = blankTrack(slot);
+    if (foot.pixels === 0) {
+      track.ambiguity = 'the candidate draws nothing here — the slot is empty or entirely outside the frame';
+      pending.push({ track, foot, claimed: null });
+      continue;
+    }
+    track.candidate = {
+      cx: foot.cx,
+      cy: foot.cy,
+      width: foot.maxX - foot.minX,
+      height: foot.maxY - foot.minY,
+      pixels: Math.round(foot.pixels),
+    };
+    const radius = searchRadius(track.candidate.width, track.candidate.height);
+    track.searchRadius = radius;
+    if (components.length === 0) {
+      track.ambiguity = 'the reference frame is empty — nothing to match against';
+      pending.push({ track, foot, claimed: null });
+      continue;
+    }
+
+    // A component whose box holds this slot's centroid and is not much bigger than
+    // the slot is that slot's own blob. One much bigger than the slot is a merge:
+    // its centroid is the merged shape's and says nothing about this part.
+    const covering = components.filter(
+      (c) => foot.cx >= c.minX - 1 && foot.cx <= c.maxX + 1 && foot.cy >= c.minY - 1 && foot.cy <= c.maxY + 1,
+    );
+    const width = track.candidate.width;
+    const height = track.candidate.height;
+    const margin = Math.max(MERGE_MARGIN_PIXELS, MERGE_MARGIN * Math.max(width, height));
+    const own = covering
+      .filter(
+        (c) =>
+          c.pixels <= foot.pixels * MERGE_RATIO &&
+          c.maxX - c.minX <= width + margin &&
+          c.maxY - c.minY <= height + margin,
+      )
+      .sort((a, b) => Math.abs(a.pixels - foot.pixels) - Math.abs(b.pixels - foot.pixels))[0];
+
+    // Containment plus a compatible size is a far stronger claim than nearest
+    // centroid, so it is taken at face value and the runner-up test below — which
+    // exists to catch a *guess* — does not apply to it.
+    if (own) {
+      fillComponentMatch(track, foot, own);
+      claim(takenBy, own, slot);
+      pending.push({ track, foot, claimed: own });
+      continue;
+    }
+
+    if (covering.length > 0) {
+      const biggest = covering.sort((a, b) => b.pixels - a.pixels)[0];
+      track.ambiguity =
+        `this slot is inside a reference component ${(biggest.pixels / Math.max(1, foot.pixels)).toFixed(1)}x its ` +
+        `size and ${biggest.maxX - biggest.minX}x${biggest.maxY - biggest.minY} px against its ${width}x${height} — ` +
+        'the reference merged it with something it touches or is drawn behind';
+      pending.push({ track, foot, claimed: null });
+      continue;
+    }
+
+    // Nothing contains it: the slot landed in open background. Nearest centroid is
+    // a guess, so it is bounded by what this slot could plausibly have moved, and
+    // a rival about as near makes it a guess between two things.
+    const ranked = components
+      .map((c) => ({ c, d: Math.hypot(c.cx - foot.cx, c.cy - foot.cy) }))
+      .sort((a, b) => a.d - b.d);
+    const nearest = ranked[0];
+    if (nearest.d > radius) {
+      track.ambiguity =
+        `the nearest reference component is ${nearest.d.toFixed(1)} px away, past the ${radius} px this ` +
+        `${Math.round(Math.max(track.candidate.width, track.candidate.height))} px slot could have moved and still ` +
+        'be itself';
+      pending.push({ track, foot, claimed: null });
+      continue;
+    }
+    if (ranked[1] && ranked[1].d <= nearest.d * AMBIGUITY_RATIO) {
+      track.ambiguity =
+        `two reference components are about equally near (${nearest.d.toFixed(1)} px and ${ranked[1].d.toFixed(1)} px)`;
+      pending.push({ track, foot, claimed: null });
+      continue;
+    }
+    fillComponentMatch(track, foot, nearest.c);
+    claim(takenBy, nearest.c, slot);
+    pending.push({ track, foot, claimed: nearest.c });
+  }
+
+  // A component two slots both claim is one blob the reference merged. Neither
+  // claim is a measurement of its slot, so both drop to the template matcher.
+  for (const [component, claimants] of takenBy) {
+    if (claimants.length < 2) continue;
+    for (const entry of pending) {
+      if (!claimants.includes(entry.track.slot)) continue;
+      const others = claimants.filter((s) => s !== entry.track.slot);
+      entry.track.ambiguity =
+        `shares one reference component with ${others.map((s) => JSON.stringify(s)).join(', ')} — they touch or ` +
+        'overlap in this frame';
+      entry.claimed = component;
+      clearMatch(entry.track);
+    }
+  }
+
+  // ...and a component ONE slot claimed while other ink of the candidate's sits
+  // inside it is the same blob by the other route — the one that reported a
+  // dominant part's distance to a five-part blob as that part's drift (#37). The
+  // claim goes to the template matcher for the same reason: a centroid shared by
+  // several parts is not this part's position.
+  for (const entry of pending) {
+    if (entry.claimed === null || entry.track.ambiguity !== null) continue;
+    const id = idOf.get(entry.claimed);
+    if (id === undefined) continue;
+    const others = (occupants.get(id) ?? []).filter((slot) => slot !== entry.track.slot);
+    if (others.length === 0) continue;
+    entry.track.ambiguity =
+      `this slot's reference component also holds ${others.map((s) => JSON.stringify(s)).join(', ')} — ` +
+      `${entry.claimed.pixels} px of blob against this slot's own ${Math.round(entry.foot.pixels)} px, so its ` +
+      "centroid is the merged shape's and not this part's";
+    clearMatch(entry.track);
+  }
+
+  // The fallback: anything the components could not attribute, correlated against
+  // the pixels of its own that the candidate's composite lets show. The owner mask
+  // is a property of the frame rather than of the slot, so it is taken once and
+  // only when something is actually going to correlate against it.
+  if (source) {
+    const wanted = pending.filter(
+      (entry) => entry.track.ambiguity !== null && entry.track.candidate !== null && entry.foot.pixels > 0,
+    );
+    if (wanted.length > 0) {
+      const visible = visibleField(source);
+      for (const entry of wanted) applyTemplateMatch(entry.track, entry.foot, source, visible);
+    }
+  }
+
+  const tracks = pending.map((p) => p.track);
+  return { tracks, matchedComponents: countExplained(footprints, components) };
+}
+
+function claim(takenBy: Map<Component, string[]>, component: Component, slot: string): void {
+  const claimants = takenBy.get(component) ?? [];
+  claimants.push(slot);
+  takenBy.set(component, claimants);
+}
+
+function blankTrack(slot: string): SlotTrack {
+  return {
+    slot,
+    candidate: null,
+    method: 'none',
+    reference: null,
+    drift: null,
+    driftX: null,
+    driftY: null,
+    widthDrift: null,
+    heightDrift: null,
+    confidence: null,
+    searchRadius: null,
+    wholePixel: null,
+    ambiguity: null,
+  };
+}
+
+function fillComponentMatch(track: SlotTrack, foot: Footprint, component: Component): void {
+  track.method = 'component';
+  track.reference = {
+    cx: component.cx,
+    cy: component.cy,
+    width: component.maxX - component.minX,
+    height: component.maxY - component.minY,
+    pixels: component.pixels,
+  };
+  track.driftX = component.cx - foot.cx;
+  track.driftY = component.cy - foot.cy;
+  track.drift = Math.hypot(track.driftX, track.driftY);
+  track.widthDrift = foot.maxX - foot.minX - track.reference.width;
+  track.heightDrift = foot.maxY - foot.minY - track.reference.height;
+}
+
+function clearMatch(track: SlotTrack): void {
+  track.method = 'none';
+  track.reference = null;
+  track.drift = null;
+  track.driftX = null;
+  track.driftY = null;
+  track.widthDrift = null;
+  track.heightDrift = null;
+  track.wholePixel = null;
+}
+
+/**
+ * How many reference components the candidate accounts for.
+ *
+ * Overlap rather than the drift match, and deliberately so: a slot whose drift
+ * could not be measured has still *drawn* over that blob, and counting it as
+ * unaccounted would report "something in the shot you have not drawn" about a part
+ * that is right there. What this number is for is the opposite case — a component
+ * no slot reaches at all.
+ */
+function countExplained(footprints: Map<string, Footprint>, components: Component[]): number {
+  let explained = 0;
+  for (const component of components) {
+    for (const foot of footprints.values()) {
+      if (foot.pixels === 0) continue;
+      if (foot.minX >= component.maxX || component.minX >= foot.maxX) continue;
+      if (foot.minY >= component.maxY || component.minY >= foot.maxY) continue;
+      explained++;
+      break;
+    }
+  }
+  return explained;
+}
+
+// ---------------------------------------------------------------------------
+// the template pass
+// ---------------------------------------------------------------------------
+
+interface Template {
+  /** Patch origin in frame pixels. */
+  ox: number;
+  oy: number;
+  width: number;
+  height: number;
+  /** The slot alone, composited over the background. */
+  patch: Plate;
+  /** Offsets into the patch that carry the slot's own VISIBLE pixels. */
+  samples: Int32Array;
+  /** Mean distance from the background over those samples: how visible it is. */
+  contrast: number;
+  /** How much of the slot's own ink reaches the picture, and how much does not. */
+  ink: number;
+  covered: number;
+}
+
+/**
+ * A template to correlate, or why this slot has none.
+ *
+ * `covered` is a *reported* outcome and not a quiet miss: a slot every pixel of
+ * which the candidate draws over has no position to measure, and saying so is a
+ * different fact from a correlation that searched and found nothing.
+ */
+type TemplateOutcome =
+  | { kind: 'template'; template: Template }
+  | { kind: 'undrawn' }
+  | { kind: 'covered'; ink: number; width: number; height: number };
+
+/**
+ * Which slot you would SEE at each pixel of the CANDIDATE's own frame.
+ *
+ * The composite's own rule, last writer wins, over the candidate's pieces in draw
+ * order — `frameGeometry`'s owner mask, which already computes exactly this for
+ * the chain split. It is entirely candidate-side: what it answers is *which of my
+ * template's pixels do I myself cover*, which is a fact about the thing being
+ * looked for and never a reading of the reference.
+ *
+ * ⚠️ So it is one-sided by construction, and that is the honest half to have. A
+ * pixel the candidate leaves visible may be covered in the reference, and nothing
+ * here can know it; what it removes is the half that is knowable, which is the
+ * half that was moving the answer.
+ */
+interface VisibleField {
+  owner: Int32Array;
+  idOf: Map<string, number>;
+  width: number;
+  height: number;
+}
+
+function visibleField(source: SlotSource): VisibleField {
+  const idOf = new Map<string, number>();
+  for (const piece of source.frame.pieces) if (!idOf.has(piece.slot)) idOf.set(piece.slot, idOf.size);
+  const { owner } = frameGeometry(source.frame, source.pages, source.viewport, idOf);
+  return {
+    owner: owner ?? new Int32Array(source.viewport.width * source.viewport.height).fill(-1),
+    idOf,
+    width: source.viewport.width,
+    height: source.viewport.height,
+  };
+}
+
+/**
+ * The slot on its own, over the background, at the size it drew — sampled only
+ * where the candidate's own composite lets it show.
+ *
+ * Its own pieces and nothing else — the point of the fallback is that the
+ * reference merged this part with its neighbours, so the thing being looked for
+ * has to be the part rather than the blob.
+ *
+ * 🚨 And only the pixels of it that reach the picture. A template pixel the
+ * candidate draws something over cannot match the reference at the true offset,
+ * so it contributes a residual at *every* offset and the correlation is reading
+ * the occluder rather than the part — issue #698's whole finding. ⚠️ It is not a
+ * flat penalty that cancels: sliding the template moves those samples onto other
+ * pixels, and where the visible basin is shallow their gradient is what decides
+ * the winner. `contrast` is taken over the same retained samples, so the residual
+ * test below compares two figures measured on one population.
+ */
+function templateFor(slot: string, foot: Footprint, source: SlotSource, visible: VisibleField): TemplateOutcome {
+  const ox = Math.floor(foot.minX);
+  const oy = Math.floor(foot.minY);
+  const width = Math.ceil(foot.maxX) - ox;
+  const height = Math.ceil(foot.maxY) - oy;
+  if (width <= 0 || height <= 0) return { kind: 'undrawn' };
+  const patch = new Plate(width, height);
+  const bg = source.background;
+  for (let y = 0; y < height; y++) for (let x = 0; x < width; x++) patch.set(x, y, bg);
+  const project = projector(source.viewport);
+  const shifted = (wx: number, wy: number): [number, number] => {
+    const [px, py] = project(wx, wy);
+    return [px - ox, py - oy];
+  };
+  let drew = false;
+  for (const piece of source.frame.pieces) {
+    if (piece.slot !== slot) continue;
+    rasterisePiece(pageFor(source.pages, piece), piece, shifted, { width, height }, (px, py, r, g, b, a) => {
+      patch.blend(px, py, [r, g, b, a]);
+      drew = true;
+    });
+  }
+  if (!drew) return { kind: 'undrawn' };
+
+  const own = visible.idOf.get(slot) ?? -1;
+  const hits: number[] = [];
+  let ink = 0;
+  let sum = 0;
+  for (let y = 0; y < height; y++) {
+    for (let x = 0; x < width; x++) {
+      const d = backgroundDistance(patch, x, y, bg);
+      if (d <= 0) continue;
+      ink++;
+      const fx = ox + x;
+      const fy = oy + y;
+      // Off the frame counts as covered: a pixel outside the reference compares
+      // against the background at every offset, which is the same constant.
+      if (fx < 0 || fy < 0 || fx >= visible.width || fy >= visible.height) continue;
+      if (visible.owner[fy * visible.width + fx] !== own) continue;
+      hits.push(y * width + x);
+      sum += d;
+    }
+  }
+  if (ink === 0) return { kind: 'undrawn' };
+  if (hits.length === 0) return { kind: 'covered', ink, width, height };
+  const stride = Math.max(1, Math.ceil(hits.length / MAX_SAMPLES));
+  const samples: number[] = [];
+  for (let i = 0; i < hits.length; i += stride) samples.push(hits[i]);
+  return {
+    kind: 'template',
+    template: {
+      ox,
+      oy,
+      width,
+      height,
+      patch,
+      samples: Int32Array.from(samples),
+      contrast: sum / hits.length,
+      ink,
+      covered: ink - hits.length,
+    },
+  };
+}
+
+/** Mean absolute RGB difference between the template and the reference at an offset. */
+function scoreAt(template: Template, source: SlotSource, dx: number, dy: number): number {
+  const { patch, samples, width, ox, oy } = template;
+  const reference = source.reference;
+  const bg = source.background;
+  let sum = 0;
+  for (let i = 0; i < samples.length; i++) {
+    const at = samples[i];
+    const x = at % width;
+    const y = (at - x) / width;
+    const a = patch.get(x, y);
+    const rx = ox + x + dx;
+    const ry = oy + y + dy;
+    const b =
+      rx < 0 || ry < 0 || rx >= reference.width || ry >= reference.height ? bg : reference.get(rx, ry);
+    sum += (Math.abs(a[0] - b[0]) + Math.abs(a[1] - b[1]) + Math.abs(a[2] - b[2])) / 3;
+  }
+  return sum / samples.length;
+}
+
+/**
+ * The offsets one axis of the coarse sweep visits, in ascending order.
+ *
+ * ⭐ Anchored on **zero** and symmetric about it, and both halves of that are
+ * correctness rather than taste. Zero is the offset that says *the part landed
+ * where the candidate drew it*, which is the answer on every frame of a correct
+ * rig — so a lattice that does not carry it cannot report a correct rig as
+ * correct. And the window this searches is `±radius`, so a lattice reaching one
+ * stride further one way than the other is searching a different window from the
+ * one the radius names.
+ *
+ * 🚨 The sweep used to run `for (dx = -radius; dx <= radius; dx += coarse)`,
+ * which puts the origin on the lattice only when `coarse` divides `radius`.
+ * Issue #678: `gallery/squash` checked against frames rendered from **itself**
+ * reported `slot drift worst 3.7 px "ear_r"`. The ear's radius is 26 and its
+ * stride 3, so the lattice ran `-26, -23, … -2, 1, 4 …` — the origin missing,
+ * and the nearest lattice point below it two whole pixels out. The coarse winner
+ * landed at `(1, -2)` scoring 37.33, the halving refinement walked it to
+ * `(2, -3)` at 37.28, and the exhaustive field over the same radius has its
+ * minimum at `(0, 0)` scoring **35.88**. The refinement cannot recover it: it
+ * searches `±1` around the coarse winner, and `dy = -2` is two away from zero.
+ */
+function sweepOffsets(radius: number, coarse: number): number[] {
+  const out = [0];
+  for (let d = coarse; d <= radius; d += coarse) {
+    out.unshift(-d);
+    out.push(d);
+  }
+  return out;
+}
+
+/**
+ * Correlate one slot against the reference inside its own search radius.
+ *
+ * Multi-resolution: a full sweep at a coarse stride, then halving steps around the
+ * winner, then a parabolic refinement for the sub-pixel part. That keeps the cost
+ * near-constant in the radius — a big part gets a big window without paying its
+ * square — and the coarse sweep doubles as the rival field the confidence is read
+ * from.
+ *
+ * ⚠️ What is left over on an identity run is the **sub-pixel step**, and it
+ * cannot be zero: the parabola is fitted through three whole-pixel residuals that
+ * antialiasing alone makes slightly uneven, so its vertex sits a fraction off the
+ * winner. `parabolic` clamps that to `SUBPIXEL_CLAMP` on each axis, which bounds
+ * the identity floor **provided the whole-pixel winner is the identity offset** —
+ * and `templateFor` masking the occluded samples is what makes that second half
+ * true. `track.wholePixel` publishes the winner and `driftBound` turns it into the
+ * figure, so a reader never has to assume it. See `docs/AUTHORING.md` §9.2.
+ */
+function applyTemplateMatch(track: SlotTrack, foot: Footprint, source: SlotSource, visible: VisibleField): void {
+  if (!track.candidate || track.searchRadius === null) return;
+  const outcome = templateFor(track.slot, foot, source, visible);
+  if (outcome.kind === 'undrawn') return;
+  if (outcome.kind === 'covered') {
+    track.method = 'none';
+    track.ambiguity =
+      `${track.ambiguity ?? 'no component of its own'}; its drift is not measurable — every one of the ` +
+      `${outcome.ink} px this ${outcome.width}x${outcome.height} px slot draws is covered by something the ` +
+      'candidate draws over it, so none of its own ink reaches the picture to be correlated against';
+    return;
+  }
+  const template = outcome.template;
+  if (template.contrast <= 0) return;
+  const radius = track.searchRadius;
+  const cache = new Map<number, number>();
+  const score = (dx: number, dy: number): number => {
+    const key = (dy + MAX_SEARCH_RADIUS * 2) * 1024 + (dx + MAX_SEARCH_RADIUS * 2);
+    const seen = cache.get(key);
+    if (seen !== undefined) return seen;
+    const value = scoreAt(template, source, dx, dy);
+    cache.set(key, value);
+    return value;
+  };
+
+  const coarse = Math.max(1, Math.round(radius / 8));
+  let bestX = 0;
+  let bestY = 0;
+  let best = Infinity;
+  const sweep: Array<{ dx: number; dy: number; s: number }> = [];
+  const axis = sweepOffsets(radius, coarse);
+  for (const dy of axis) {
+    for (const dx of axis) {
+      const s = score(dx, dy);
+      sweep.push({ dx, dy, s });
+      if (s < best) {
+        best = s;
+        bestX = dx;
+        bestY = dy;
+      }
+    }
+  }
+  for (let step = coarse; step > 1; ) {
+    step = Math.max(1, Math.floor(step / 2));
+    for (let dy = bestY - step; dy <= bestY + step; dy += step) {
+      for (let dx = bestX - step; dx <= bestX + step; dx += step) {
+        if (Math.abs(dx) > radius || Math.abs(dy) > radius) continue;
+        const s = score(dx, dy);
+        if (s < best) {
+          best = s;
+          bestX = dx;
+          bestY = dy;
+        }
+      }
+    }
+  }
+
+  // A winner nobody else came close to is a located part. A winner its neighbours
+  // match just as well is a featureless blob, and no offset is evidence.
+  const gap = Math.max(RIVAL_GAP, coarse);
+  let rival = Infinity;
+  for (const { dx, dy, s } of sweep) {
+    if (Math.hypot(dx - bestX, dy - bestY) < gap) continue;
+    if (s < rival) rival = s;
+  }
+  const confidence = Number.isFinite(rival) && rival > 0 ? Math.max(0, Math.min(1, 1 - best / rival)) : 0;
+
+  const reach = Math.min(1, Math.hypot(bestX, bestY) / radius);
+  const required = MIN_CONFIDENCE + reach * CONFIDENCE_SLOPE;
+  if (best > template.contrast * MAX_RESIDUAL_FRACTION || confidence < required) {
+    track.method = 'none';
+    track.confidence = Number.isFinite(confidence) ? confidence : 0;
+    track.ambiguity =
+      `${track.ambiguity ?? 'no component of its own'}; correlating the slot's own pixels found no match within ` +
+      `${radius} px either (best residual ${best.toFixed(1)} against its own ${template.contrast.toFixed(1)} of ` +
+      `contrast over the ${template.ink - template.covered} of its ${template.ink} px that reach the picture; ` +
+      `confidence ${(Number.isFinite(confidence) ? confidence : 0).toFixed(2)} where ` +
+      `${required.toFixed(2)} is needed ${Math.hypot(bestX, bestY).toFixed(1)} px out)`;
+    return;
+  }
+
+  const dx = bestX + parabolic(score(bestX - 1, bestY), best, score(bestX + 1, bestY));
+  const dy = bestY + parabolic(score(bestX, bestY - 1), best, score(bestX, bestY + 1));
+  track.method = 'template';
+  track.reference = null;
+  track.driftX = dx;
+  track.driftY = dy;
+  track.drift = Math.hypot(dx, dy);
+  track.widthDrift = null;
+  track.heightDrift = null;
+  track.confidence = confidence;
+  track.wholePixel = { dx: bestX, dy: bestY };
+  track.ambiguity = null;
+}
+
+/**
+ * How far off the whole-pixel winner the sub-pixel step may place the answer.
+ *
+ * Half a pixel on each axis, because past that the neighbouring whole pixel is
+ * the better winner and the search would have found it. It is what bounds the
+ * instrument's identity floor at `hypot(0.5, 0.5)` px, which `docs/AUTHORING.md`
+ * §9.2 states as the floor an author reads a drift against.
+ */
+export const SUBPIXEL_CLAMP = 0.5;
+
+/** Sub-pixel minimum of the parabola through three samples one pixel apart. */
+function parabolic(before: number, at: number, after: number): number {
+  const denominator = before - 2 * at + after;
+  if (!Number.isFinite(denominator) || Math.abs(denominator) < 1e-9) return 0;
+  return Math.max(-SUBPIXEL_CLAMP, Math.min(SUBPIXEL_CLAMP, (before - after) / (2 * denominator)));
+}
+
+/** Is this track's drift a measurement of this slot? */
+export function isAttributable(track: SlotTrack): boolean {
+  return track.drift !== null && track.ambiguity === null;
+}