@pinepaper.studio/mcp-server 1.6.7 → 1.6.9

package/dist/character/perform.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"perform.d.ts","sourceRoot":"","sources":["../../src/character/perform.ts"],"names":[],"mappings":"AA8CA,MAAM,WAAW,aAAa;IAC5B,EAAE,EAAE,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACrB,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,CAAC,EAAE,MAAM,CAAC;IACd;;;;;;;;;;;;;;;;OAgBG;IACH,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB;;;;;;;;;;;;;;;;;;;;;;;;OAwBG;IACH,SAAS,CAAC,EAAE,MAAM,CAAC;CACpB;AACD,MAAM,WAAW,kBAAkB;IACjC,OAAO,EAAE,MAAM,CAAC;IAChB,MAAM,EAAE,MAAM,CAAC;IACf,OAAO,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IAChC,KAAK,EAAE,MAAM,CAAC,MAAM,EAAE,aAAa,CAAC,CAAC;IACrC;;;;;;;;;;;;;OAaG;IACH,GAAG,CAAC,EAAE;QAAE,KAAK,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAC;QAAC,IAAI,CAAC,EAAE,MAAM,CAAC;QAAC,GAAG,CAAC,EAAE,MAAM,CAAC;QAAC,KAAK,CAAC,EAAE,OAAO,CAAA;KAAE,CAAC;CACtF;AACD,gFAAgF;AAChF,MAAM,WAAW,OAAO;IACtB,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,WAAW,GAAG,iBAAiB,GAAG,OAAO,GAAG,OAAO,GAAG,WAAW,GAAG,OAAO,CAAC;IAClF,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,mEAAmE;IACnE,MAAM,CAAC,EAAE,MAAM,EAAE,CAAC;IAClB,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,sEAAsE;IACtE,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,6EAA6E;IAC7E,IAAI,CAAC,EAAE,MAAM,GAAG,MAAM,EAAE,CAAC;IACzB,QAAQ,CAAC,EAAE,QAAQ,GAAG,QAAQ,GAAG,UAAU,GAAG,GAAG,GAAG,GAAG,GAAG,SAAS,CAAC;IACpE,+DAA+D;IAC/D,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,sDAAsD;IACtD,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,qFAAqF;IACrF,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB;;;;;;;;;OASG;IACH,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,kDAAkD;IAClD,UAAU,CAAC,EAAE,MAAM,CAAC;CACrB;AACD,yDAAyD;AACzD,MAAM,WAAW,IAAI;IACnB,EAAE,EAAE,MAAM,CAAC;IACX,OAAO,EAAE,MAAM,CAAC;IAChB,0EAA0E;IAC1E,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,8DAA8D;IAC9D,KAAK,CAAC,EAAE,MAAM,CAAC;CAChB;AACD,MAAM,WAAW,YAAY;IAC3B,EAAE,EAAE,MAAM,CAAC;IACX,SAAS,EAAE,kBAAkB,CAAC;IAC9B,EAAE,EAAE;QAAE,CAAC,EAAE,MAAM,CAAC;QAAC,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC;IAC7B,MAAM,EAAE,MAAM,CAAC;IACf,QAAQ,EAAE,OAAO,EAAE,CAAC;IACpB,KAAK,EAAE,IAAI,EAAE,CAAC;IACd,gEAAgE;IAChE,eAAe,EAAE,MAAM,CAAC;CACzB;AAED,KAAK,EAAE,GAAG;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,IAAI,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CAAE,CAAC;AAgC1D,wBAAgB,gBAAgB,CAAC,KAAK,EAAE,YAAY,GAAG;IAAE,GAAG,EAAE,EAAE,EAAE,CAAC;IAAC,KAAK,EAAE,MAAM,EAAE,CAAC;IAAC,MAAM,EAAE,MAAM,CAAC;IAAC,eAAe,EAAE,MAAM,EAAE,CAAA;CAAE,CAmhB/H"}

package/dist/character/perform.js

@@ -0,0 +1,626 @@
+/* VENDORED from mcp-cloud/src/services/character-perform.ts.
+ * sha256: 91737fa6410fbe24e7b969c9d1a97ca4c6694c117854ce3182606efa41e98db7
+ * Generator: scripts/sync-character-modules.mjs
+ *
+ * Do NOT edit here — change it in mcp-cloud and re-run the generator, or the
+ * two implementations drift and a figure performs differently depending on
+ * which surface a caller reached. Guarded by character.test.ts.
+ */
+/**
+ * A CHARACTER IS DIRECTED, NOT DRAWN.
+ *
+ * This is the translation step the declarative graph exists for. A model says
+ * WHAT happens — this character, here, this tall, blinks at 1.2s and speaks
+ * from 2.0 to 4.5 — and the knowledge of HOW lives in the graph: which parts
+ * the figure has, which of them are alternates of one another, how long a blink
+ * lasts, how fast a mouth moves. Nothing here asks a model for geometry.
+ *
+ * WHY IT HAS TO BE THIS WAY. The alternative is what exists today: a model
+ * writes thirty-odd correct calls — create the skeleton, add each bone with the
+ * right parent and angle, create each shape, attach each to the right bone,
+ * then author poses — and one mistake anywhere leaves a broken figure. That is
+ * a lot of exactly-right output to ask of a language model, and it is the
+ * reported reason characters do not work. A symbol and a list of beats is
+ * something a small model can get right, and a small model getting it right is
+ * the difference between this running in the cloud and running on someone's
+ * laptop.
+ *
+ * THE THREE MOTIONS ARE THE ONES STORYTELLERS ASKED FOR — blink, speak, turn
+ * the head — and each is expressed with machinery both renderers were measured
+ * on this week: opacity keyframes and position keyframes. No shape keys, no IK,
+ * no deformers. A channel that needs a subsystem the container cannot draw is a
+ * channel that works in a preview and fails in the delivered video, which is
+ * the failure this whole layer is a response to.
+ *
+ * NOT LIP SYNC. `say` opens and closes the mouth for the length of a line; it
+ * does not match phonemes, because matching phonemes needs the word timings a
+ * voice track carries and no voice track exists yet. Called `say` rather than
+ * `lipSync` so nobody reads more into it than it does.
+ */
+import { DURATION_SCALE_MS } from "./duration-scale.js";
+import { pathBBox, scalePath, shiftPath } from "./path-bbox.js";
+/** Nearest duration token to `ms` — the same rounding the token repair uses. */
+const nearestOnScale = (ms) => DURATION_SCALE_MS.reduce((best, d) => (Math.abs(d - ms) < Math.abs(best - ms) ? d : best), DURATION_SCALE_MS[0]);
+/** Sorted, deduplicated keyframes — a track with two values at one instant is
+ *  ambiguous and the renderers resolve it differently. */
+function tidy(kfs) {
+    const byTime = new Map();
+    // LATER WINS AT THE SAME INSTANT, and the zero anchor is pushed last — so it
+    // is merged in first here, letting a real beat at t=0 override it.
+    for (const k of [...kfs].reverse()) {
+        const t = Math.round(k.time * 1000) / 1000;
+        byTime.set(t, { ...(byTime.get(t) ?? {}), ...k, time: t });
+    }
+    return [...byTime.values()].sort((a, b) => a.time - b.time);
+}
+/** Relative luminance of a #rgb/#rrggbb colour, 0..1. Unknown formats read as
+ *  light, so an unrecognised colour keeps the old stroked behaviour. */
+function luminance(colour) {
+    const h = colour.trim().replace(/^#/, "");
+    const full = h.length === 3 ? h.split("").map((c) => c + c).join("") : h;
+    if (!/^[0-9a-f]{6}$/i.test(full))
+        return 1;
+    const [r, g, b] = [0, 2, 4].map((i) => parseInt(full.slice(i, i + 2), 16) / 255);
+    return 0.2126 * r + 0.7152 * g + 0.0722 * b;
+}
+/** Is this fill already the drawing's own ink? Such a part must not be
+ *  outlined — see the note at the stroke below. */
+function isInkFill(colour) {
+    return luminance(colour) < 0.22;
+}
+export function performCharacter(input) {
+    const { id, depiction: dep, at, height, channels, beats, durationSeconds } = input;
+    const k = height / dep.height;
+    const byName = new Map(channels.map((c) => [c.name, c]));
+    /** Beats naming a channel this concept has no definition for — reported, not hidden. */
+    const unknownChannels = [];
+    const ops = [];
+    /**
+     * ONE CONTINUOUS CONTOUR, WHICH IS WHAT MAKES A FIGURE READ AS ONE OBJECT.
+     *
+     * Stroking each part separately draws the seams between primitives at exactly
+     * the weight that separates the figure from the sky. Rendered at 0.9s of
+     * `ledge-at-dusk`, the pigeon is a circle-with-an-outline sitting on an
+     * ellipse-with-an-outline and the head's contour cuts straight across the
+     * body — which is the "the head is not connected" complaint, drawn.
+     *
+     * The fix is the one flat-vector artwork has always used, and both Rive and
+     * Lottie files are built this way: lay the ink down FIRST, as an unfilled
+     * double-width stroke of every silhouette part, then paint the fills over it.
+     * A stroke is centred on its path, so each fill covers the inner half of its
+     * own ink and the inner half of every ink it overlaps. What survives is the
+     * outer boundary of the UNION — one contour, no internal seams — without any
+     * boolean path arithmetic.
+     *
+     * The ink copy is `carriedBy` its source, so it inherits that part's motion by
+     * construction. There is no second track to drift.
+     *
+     * Opt-in per style: a style that wants separate outlines keeps them.
+     */
+    const UNION_INK = dep.ink?.union === true;
+    if (UNION_INK) {
+        for (const [pid, part] of Object.entries({ ...dep.parts })) {
+            // Only the silhouette. Parts that are already ink, or carry reduced
+            // interior linework, are detail drawn ON the form and are not part of
+            // its boundary.
+            if ((part.inkWeight ?? 1) < 1)
+                continue;
+            if (dep.palette[part.fill ?? ""] ?? part.fill) {
+                const fill = dep.palette[part.fill ?? ""] ?? part.fill ?? "";
+                if (isInkFill(fill))
+                    continue;
+            }
+            dep.parts[`${pid}__ink`] = { at: part.at, path: part.path, carriedBy: pid, inkWeight: 2 };
+        }
+    }
+    /** Ink copies first: everything painted after them covers their inner half. */
+    const partIds = Object.keys(dep.parts).sort((a, b) => (a.endsWith("__ink") ? 0 : 1) - (b.endsWith("__ink") ? 0 : 1));
+    /**
+     * Every part this one carries, at any depth. A chain (skull → beak → nostril)
+     * has to propagate the whole way down or the deepest part is the one that
+     * tears off, which is the hardest kind to notice.
+     *
+     * Cycles in the data are ignored rather than thrown on: a malformed depiction
+     * should draw a stiff figure, not fail the render.
+     */
+    /**
+     * A BONE'S MOTION, EXPRESSED FOR THE PART IT CARRIES.
+     *
+     * Translation is the easy half and the only half the first version did: the
+     * carrier's dx is the child's dx, so the same offsets are handed straight on.
+     *
+     * ROTATION IS NOT. A bone that turns swings everything on it AROUND ITS OWN
+     * PIVOT — that is what distinguishes a head turning from a head being slid
+     * sideways, and it is the difference the eye reads as a figure being solid.
+     * Copying the carrier's rotation onto the child instead makes the child spin
+     * about its own centre: the beak would rotate on the spot while the skull
+     * turns beneath it, which is a worse artefact than the one the bone was
+     * built to remove.
+     *
+     * So a carried part gets BOTH — the same angle, and the displacement its rest
+     * position travels along the arc:
+     *
+     *     v  = childRest - pivot
+     *     v' = R(theta) * v
+     *     offset = v' - v
+     *
+     * Positions here are the part's placed centre in scene units, which is what
+     * `at` resolves to at the emission below, so the arc is computed in the same
+     * space it will be drawn in.
+     */
+    const carry = (acc, carrier, child, property) => {
+        if (property !== "rotation")
+            return acc;
+        const cp = dep.parts[carrier], ch = dep.parts[child];
+        if (!cp || !ch)
+            return acc;
+        const vx = (ch.at[0] - cp.at[0]) * k;
+        const vy = (ch.at[1] - cp.at[1]) * k;
+        return acc.map((kf) => {
+            const deg = typeof kf.rotation === "number" ? kf.rotation : 0;
+            const r = (deg * Math.PI) / 180;
+            const cos = Math.cos(r), sin = Math.sin(r);
+            return {
+                ...kf,
+                rotation: deg,
+                x: vx * cos - vy * sin - vx,
+                y: vx * sin + vy * cos - vy,
+            };
+        });
+    };
+    const carriedTransitively = (carrier) => {
+        const out = [];
+        const seen = new Set([carrier]);
+        let front = [carrier];
+        while (front.length) {
+            const next = [];
+            for (const id of front) {
+                for (const pid of partIds) {
+                    if (dep.parts[pid]?.carriedBy !== id || seen.has(pid))
+                        continue;
+                    seen.add(pid);
+                    out.push(pid);
+                    next.push(pid);
+                }
+            }
+            front = next;
+        }
+        return out;
+    };
+    // Which drawn parts belong to which alternate group. A part named as the
+    // `alternate` of any channel starts hidden; the base starts visible.
+    const alternates = new Set();
+    for (const c of channels) {
+        if (c.alternate)
+            alternates.add(c.alternate);
+        // Every speech state except the first is an alternate of it: the closed
+        // mouth is what the face wears at rest.
+        for (const st of (c.states ?? []).slice(1))
+            alternates.add(st);
+    }
+    const partFill = (pid) => {
+        const part = dep.parts[pid];
+        return dep.palette[part?.fill ?? ""] ?? part?.fill ?? "#1f2937";
+    };
+    const hidden = (p) => [...alternates].some((a) => p === a || p.startsWith(`${a}_`));
+    /**
+     * A CONTOUR CANNOT BE WIDER THAN THE THING IT CONTAINS.
+     *
+     * The union layer strokes at double weight so that each fill covers the inner
+     * half and the outer half survives as the silhouette. On a large part that is
+     * exactly right. On a SMALL one the two halves meet in the middle and the
+     * stroke floods the shape: the pigeon's tail rendered as a black smear, having
+     * been a clean wedge before.
+     *
+     * So the weight is capped by the part's own short side. Half of the smaller
+     * dimension is the most that can be drawn before the strokes from opposite
+     * edges overlap, and a third leaves the shape still legible as a shape.
+     */
+    const inkWidthFor = (pid) => {
+        const nominal = dep.ink.width * k * 0.6 * (dep.parts[pid]?.inkWeight ?? 1);
+        const b = pathBBox(dep.parts[pid]?.path ?? "");
+        if (!b)
+            return Math.max(0.6, nominal);
+        const short = Math.min(b.maxX - b.minX, b.maxY - b.minY) * k;
+        if (!Number.isFinite(short) || short <= 0)
+            return Math.max(0.6, nominal);
+        return Math.max(0.6, Math.min(nominal, short / 3));
+    };
+    // A CHANNEL MAY NAME SEVERAL PARTS. A vehicle's suspension lifts the body,
+    // the glazing and the trim as one thing; declaring that as three identical
+    // channels (the pigeon's four bob channels) says the same thing three times
+    // and lets them drift apart. A bare string still works and means what it
+    // always did.
+    const matching = (spec) => {
+        const wanted = Array.isArray(spec) ? spec : [spec];
+        return partIds.filter((p) => wanted.some((w) => p === w || p.startsWith(`${w}_`)));
+    };
+    /** The value each transform channel says this part sits at when idle. */
+    const restingOf = (pid) => {
+        const out = {};
+        for (const c of channels) {
+            if (c.kind !== "transform" || !c.part || !c.property)
+                continue;
+            if (!matching(c.part).includes(pid))
+                continue;
+            const rest = c.rest ?? 1;
+            if (rest !== 1)
+                out[c.property] = rest;
+        }
+        return out;
+    };
+    // ── THE FIGURE ────────────────────────────────────────────────────────────
+    // Every part is a path placed in the frame's space. `at` is the part's centre
+    // in the figure's own space, so the figure moves as one by construction.
+    for (const pid of partIds) {
+        const p = dep.parts[pid];
+        ops.push({
+            tool: "pinepaper_create_item",
+            args: {
+                itemType: "path",
+                position: { x: 0, y: 0 },
+                properties: {
+                    id: `${id}_${pid}`,
+                    // ONE FIGURE, NOT EIGHT SHAPES. The parts are a single thing that
+                    // moves and reads together — that is what a rig IS — and saying so
+                    // is the graph's job, not the caller's. Without it every gate that
+                    // counts units sees eight competing elements, and
+                    // `compositionVerdict` (whose own comment says a rigged figure must
+                    // not be punished for being properly assembled) fails the very
+                    // piece the rig exists to make.
+                    groupId: id,
+                    pathData: shiftPath(scalePath(p.path, k), at.x + p.at[0] * k, at.y + p.at[1] * k),
+                    closed: true,
+                    fillColor: partFill(pid),
+                    // Scaled with the figure: an outline that stays 4px while the drawing
+                    // grows is a hairline on a large one and a smear on a small one.
+                    //
+                    // BUT INK IS NOT OUTLINED. The corpus is FLAT art: its linework is
+                    // already drawn, as filled dark shapes. A face arrives as ONE
+                    // compound path of six subpaths — two eyes, brows, nose, mouth — and
+                    // stroking it adds a 6px outline around every one of those features
+                    // at a 250px figure height, which merges them into dark smears. The
+                    // figure has no eyes and its face reads as damage.
+                    //
+                    // Invisible at contact-sheet scale, unmissable at full size, which is
+                    // how it survived review: the eyes were four pixels wide there.
+                    //
+                    // So the stroke goes on the SILHOUETTE — the skin and garment shapes
+                    // an inked scene needs to sit around — and never on a part that is
+                    // already ink.
+                    // An ink copy carries NO fill — it exists to be half-covered. Its
+                    // source loses its own stroke, because the union layer is now the
+                    // only thing that draws the contour.
+                    ...(pid.endsWith("__ink") ? { fillColor: "transparent" } : {}),
+                    // An ink copy is ALWAYS stroked. It declares no fill, so partFill
+                    // falls back to the dark default and isInkFill reads it as linework
+                    // and refuses to outline it — which left the union layer invisible
+                    // and the bird with no contour at all.
+                    ...(dep.ink
+                        && (pid.endsWith("__ink") || !isInkFill(partFill(pid)))
+                        && (dep.parts[pid]?.inkWeight ?? 1) > 0
+                        && !(UNION_INK && !pid.endsWith("__ink")) ? {
+                        strokeColor: dep.ink.color,
+                        // The part's own share of the style's ink — see `inkWeight`. The
+                        // floor stays, so a light interior line is thin and never absent.
+                        strokeWidth: inkWidthFor(pid),
+                        ...(dep.ink.join ? { strokeJoin: dep.ink.join } : {}),
+                        ...(dep.ink.cap ? { strokeCap: dep.ink.cap } : {}),
+                    } : {}),
+                    ...(hidden(pid) ? { opacity: 0 } : {}),
+                    // A PART STARTS AT ITS RESTING VALUE WHETHER OR NOT IT IS ANIMATED.
+                    //
+                    // The resting transform was written only into the keyframe track, and
+                    // a part with no beats gets no track — so a channel declaring a rest
+                    // other than 1 was ignored entirely for any figure nobody asked to
+                    // move. Measured on the Open Peeps cast: an eyelid whose rest is
+                    // scaleY 0.02 rendered at full height and sat over the eye, so twelve
+                    // characters had their eyes shut until someone thought to blink them.
+                    ...restingOf(pid),
+                },
+            },
+        });
+    }
+    // ── THE PERFORMANCE ───────────────────────────────────────────────────────
+    // One track per part, built from every beat that touches it, then written
+    // once. Writing a track per beat would leave the last one winning and the
+    // rest silently discarded.
+    const opacity = new Map();
+    const shiftX = new Map();
+    const transform = new Map();
+    const cycle = new Map();
+    /** A part's drawn extent, so a rig amplitude given as a fraction means the
+     *  same thing at every size the figure is drawn. */
+    const partSize = (pid) => {
+        const b = pathBBox(dep.parts[pid]?.path ?? "");
+        return b ? { w: (b.maxX - b.minX) * k, h: (b.maxY - b.minY) * k } : { w: 10 * k, h: 10 * k };
+    };
+    const swap = (base, alt, from, to) => {
+        // A hard cut, not a dissolve: two drawings of one face cross-fading reads
+        // as a double exposure, not as a mouth opening.
+        const e = 0.001;
+        for (const p of matching(base))
+            opacity.set(p, [...(opacity.get(p) ?? []), { time: Math.max(0, from - e), opacity: 1 }, { time: from, opacity: 0 }, { time: to, opacity: 0 }, { time: to + e, opacity: 1 }]);
+        for (const p of matching(alt))
+            opacity.set(p, [...(opacity.get(p) ?? []), { time: Math.max(0, from - e), opacity: 0 }, { time: from, opacity: 1 }, { time: to, opacity: 1 }, { time: to + e, opacity: 0 }]);
+    };
+    for (const b of beats) {
+        const c = byName.get(b.channel);
+        /**
+         * A BEAT ON A CHANNEL THIS CONCEPT DOES NOT DECLARE WAS DROPPED IN SILENCE.
+         *
+         * `continue` alone, with no record. Measured across the shipped corpus
+         * 2026-09-18 by mutation — changing the channel name left the scene
+         * byte-identical — 28 beats were discarded this way: 14 `blink`,
+         * 10 `headTurn`, 4 `say`. The SAME channel names work on other characters,
+         * which is what makes it so easy to miss: an author writes `blink` for a
+         * cast of four, it animates on three, and the fourth just never blinks.
+         *
+         * Still dropped, because a beat cannot drive a channel that does not
+         * exist — only the silence is removed. The available names go with it,
+         * since "blink is not a channel here" is useless without "these are".
+         */
+        if (!c) {
+            unknownChannels.push(`character "${id}": beat at ${b.at}s names channel "${b.channel}", which this concept does not declare. ` +
+                `Available: ${channels.map((x) => x.name).join(", ") || "(none)"}.`);
+            continue;
+        }
+        if (c.kind === "alternate" && c.base && c.alternate) {
+            swap(c.base, c.alternate, b.at, b.at + (c.holdMs ?? 120) / 1000);
+        }
+        else if (c.kind === "speak" && c.states && c.states.length >= 2) {
+            /**
+             * SPEECH IS NOT A METRONOME, AND IT IS NOT TWO FACES.
+             *
+             * The first version alternated two WHOLE expressions six times a second.
+             * Both of those are wrong and the result was reported, correctly, as
+             * flashing: swapping whole faces changes the eyes and brows on every
+             * syllable, so the character strobes between two people, and an even 6Hz
+             * toggle is a machine rather than a mouth.
+             *
+             * So the states share their eyes — only the mouth differs between them —
+             * and a syllable OPENS AND CLOSES THROUGH THE MIDDLE STATE rather than
+             * cutting between extremes. That middle is the transition that was
+             * missing; with two states there is nowhere for the mouth to be on the
+             * way.
+             *
+             * The rhythm varies because real speech does. The golden ratio spreads
+             * the variation evenly without repeating, and it is deterministic, which
+             * the content-addressed render cache requires.
+             */
+            const until = b.until ?? b.at + 1;
+            const closed = c.states[0], open = c.states[c.states.length - 1];
+            const mid = c.states.length > 2 ? c.states[1] : open;
+            const rate = c.syllablesPerSecond ?? 3.6;
+            let t = b.at, k = 0;
+            while (t < until) {
+                // 0.72x to 1.28x of the nominal syllable, never the same twice running.
+                const jitter = 0.72 + 0.56 * (((k + 1) * 0.6180339887) % 1);
+                const dur = Math.min((1 / rate) * jitter, until - t);
+                if (dur < 0.05)
+                    break;
+                // open on the way up, hold, close on the way down, then a beat shut.
+                swap(closed, mid, t, t + dur * 0.18);
+                swap(closed, open, t + dur * 0.18, t + dur * 0.58);
+                swap(closed, mid, t + dur * 0.58, t + dur * 0.76);
+                t += dur;
+                k++;
+            }
+        }
+        else if (c.kind === "alternate-cycle" && c.base && c.alternate) {
+            const until = b.until ?? b.at + 1;
+            const period = 1 / (c.rateHz ?? 6);
+            // Open on the first half of each period, closed on the second — the
+            // mouth has to CLOSE between syllables or the character gapes.
+            for (let t = b.at; t < until; t += period)
+                swap(c.base, c.alternate, t, Math.min(until, t + period / 2));
+        }
+        else if (c.kind === "transform" && c.part && c.property) {
+            /**
+             * A GESTURE IS A DEFORMATION, NOT A DIFFERENT DRAWING.
+             *
+             * The swap kinds above cut between two shapes, which is right when the
+             * shapes genuinely differ — an open eye and a closed lid are drawn, not
+             * squashed. It is wrong for everything that merely MOVES: a mouth
+             * opening is one mouth changing, and expressing it as a cut throws away
+             * the whole middle of the gesture. That missing middle is what reads as
+             * flashing however well the two ends are chosen.
+             *
+             * As a transform it has an attack, a peak and a release, it eases, and
+             * the same channel works on any part of any concept — a wing, a sail, a
+             * lid. Nothing here is specific to a face.
+             */
+            const rest = c.rest ?? 1, peak = c.peak ?? 1.6;
+            const attack = (c.attackMs ?? 90) / 1000, release = (c.releaseMs ?? 130) / 1000;
+            const hold = Math.max(0, (b.until ?? b.at + attack + release) - b.at - attack - release);
+            const ease = c.easing ?? "easeInOut";
+            const key = (t, v) => ({ time: Math.max(0, t), [c.property]: v, easing: ease });
+            for (const p of matching(c.part)) {
+                transform.set(p, [...(transform.get(p) ?? []),
+                    key(b.at, rest), key(b.at + attack, peak),
+                    key(b.at + attack + hold, peak), key(b.at + attack + hold + release, rest)]);
+            }
+        }
+        else if (c.kind === "cycle" && c.part && c.property) {
+            /**
+             * A CYCLE IS THE GRAPH'S OWN RIG, FINALLY CONSUMED.
+             *
+             * `concept_rig` has described exactly this since the class layer was
+             * built — a part, a property, an amplitude and a duration — and nothing
+             * has ever read it. It is not a gesture with an end: a pigeon bobs while
+             * it walks, a tail flicks the whole time it is on screen, so this repeats
+             * for as long as the beat asks rather than resolving.
+             *
+             * THE THROW IS ASYMMETRIC. A sine reads as a metronome and a pigeon does
+             * not bob like one: the head goes forward fast and the body catches up
+             * slowly. Forward in the first third, recover across the rest.
+             *
+             * Amplitude is a FRACTION OF THE PART for x and y, degrees for rotation,
+             * which is how the graph states it — a pixel amplitude would mean
+             * something different at every size the figure is drawn.
+             */
+            const until = b.until ?? durationSeconds;
+            const period = (c.durationMs ?? 800) / 1000;
+            const ease = c.easing ?? "easeInOut";
+            /**
+             * THE THROW, STATED IN THE SYSTEM'S VOCABULARY.
+             *
+             * A cycle is asymmetric on purpose — a pigeon throws its head forward
+             * and the body catches up — and the asymmetry used to be a bare 0.33 of
+             * the period. That produced intervals belonging to no scale: an 800ms
+             * bob split into 264 and 536, and a piece made mostly of cycles came out
+             * at 20% on the duration scale, which is the definition of not having a
+             * timing system.
+             *
+             * Rounding the ATTACK onto the scale and giving the remainder to the
+             * release keeps the shape and puts both halves on it: 800 becomes
+             * 300/500, 1200 becomes 400/800. The asymmetry survives; the arbitrary
+             * number does not.
+             */
+            const attack = nearestOnScale(period * 0.33 * 1000) / 1000;
+            for (const pid of matching(c.part)) {
+                // A carried part measures against what carries it — see `spanPart`.
+                const ref = c.spanPart ? (matching(c.spanPart)[0] ?? pid) : pid;
+                const span = partSize(ref);
+                const amp = c.property === "rotation" ? (c.amplitude ?? 10)
+                    : (c.amplitude ?? 0.3) * (c.property === "x" ? span.w : span.h);
+                const acc = [];
+                for (let t = b.at; t < until; t += period) {
+                    acc.push({ time: t, [c.property]: 0, easing: ease });
+                    acc.push({ time: Math.min(until, t + attack), [c.property]: amp, easing: ease });
+                }
+                acc.push({ time: until, [c.property]: 0, easing: ease });
+                cycle.set(pid, [...(cycle.get(pid) ?? []), ...acc]);
+                /**
+                 * CARRIED PARTS MOVE WITH THEIR CARRIER, not alongside it.
+                 *
+                 * The same offsets, the same object, propagated down the carrier chain
+                 * — so there is no second computation that could round differently,
+                 * ease differently, or start a beat a frame late. This is the whole
+                 * fix: a figure cannot come apart at a joint it has no independent
+                 * motion across.
+                 */
+                for (const child of carriedTransitively(pid)) {
+                    cycle.set(child, [...(cycle.get(child) ?? []), ...carry(acc, pid, child, c.property)]);
+                }
+            }
+        }
+        else if (c.kind === "shift") {
+            /**
+             * A SHIFT MOVES WHAT IT NAMES — and this moved everything.
+             *
+             * `part` was read for every other kind of channel and ignored here, so a
+             * shift always displaced the ENTIRE figure. `pp:Character.headTurn` got
+             * away with it because that depiction is a portrait: head, face and
+             * features, no body, so "every part" and "the head" are the same set.
+             * `pp:Peep` has body_0..2, and the same channel would slide the whole
+             * person sideways to turn their head.
+             *
+             * Declaring a part now narrows it, and carried parts come along — so a
+             * head turn takes the face, the hair and the eyelids with it, which is
+             * the entire reason the bone exists. A channel that names nothing still
+             * moves everything, because that is what the portrait relies on.
+             */
+            const dx = (b.value ?? 0) * (c.unitsPerValue ?? 9) * k;
+            const settle = b.until ?? b.at + 0.6;
+            const named = c.part ? matching(c.part) : partIds;
+            const moved = new Set(named);
+            for (const pid of named)
+                for (const child of carriedTransitively(pid))
+                    moved.add(child);
+            for (const p of moved) {
+                shiftX.set(p, [...(shiftX.get(p) ?? []), { time: Math.max(0, b.at - 0.001), dx: 0 }, { time: settle, dx }]);
+            }
+        }
+    }
+    // Position tracks need ABSOLUTE x, and a part's x is its own placed centre.
+    let tracks = 0;
+    for (const pid of partIds) {
+        const p = dep.parts[pid];
+        const baseX = at.x + p.at[0] * k;
+        const kfs = [];
+        for (const s of shiftX.get(pid) ?? [])
+            kfs.push({ time: s.time, x: baseX + s.dx });
+        for (const o of opacity.get(pid) ?? [])
+            kfs.push(o);
+        for (const tr of transform.get(pid) ?? [])
+            kfs.push(tr);
+        // A cycle's x/y are OFFSETS from where the part sits; everything else in a
+        // keyframe track is absolute, so they are added here rather than earlier.
+        for (const cy of cycle.get(pid) ?? []) {
+            const out = { time: cy.time, ...(cy.easing ? { easing: cy.easing } : {}) };
+            if (typeof cy.x === "number")
+                out.x = baseX + cy.x;
+            if (typeof cy.y === "number")
+                out.y = at.y + p.at[1] * k + cy.y;
+            if (typeof cy.rotation === "number")
+                out.rotation = cy.rotation;
+            if (typeof cy.scaleY === "number")
+                out.scaleY = cy.scaleY;
+            kfs.push(out);
+        }
+        if (!kfs.length)
+            continue;
+        // EVERY TRACK STARTS AT ZERO.
+        //
+        // An item whose first keyframe is at t=5 is INVISIBLE for the five seconds
+        // before it — measured: the head appeared exactly when its shift track
+        // began and the face exactly when its blink track began, and the first
+        // second of the film was an empty frame. The engine's own player treats
+        // "before the track" as "not yet", and this clock runs before it, so the
+        // player has the last word.
+        //
+        // Anchoring at zero with the baseline removes the state entirely rather
+        // than arguing with it: there is no "before the track" if the track covers
+        // the whole piece.
+        const restingTransform = {};
+        for (const c of channels) {
+            if (c.kind !== "transform" || !c.part || !c.property)
+                continue;
+            if (matching(c.part).includes(pid))
+                restingTransform[c.property] = c.rest ?? 1;
+        }
+        kfs.push({ time: 0, x: baseX, opacity: hidden(pid) ? 0 : 1, ...restingTransform });
+        // NOTHING ORGANIC MOVES AT A CONSTANT SPEED, and a rig is organic by
+        // definition. The transform and cycle channels set their own easing; the
+        // head-turn shift, the resting anchor and the closing hold did not, so a
+        // turn slid across the frame at a dead constant rate — the one tell that
+        // reads as "animated by a machine" no matter how good the drawing is.
+        // An explicit easing always wins; this only fills the silence.
+        const track = tidy(kfs).map((kf) => (kf.easing ? kf : { ...kf, easing: "easeInOut" }));
+        /**
+         * A track that never returns leaves the figure mid-gesture at the cut — and
+         * the hold must carry EVERY property the track animates, not whichever one
+         * the last keyframe happened to set.
+         *
+         * Cloning the final keyframe looked right and was not: the eye's last
+         * keyframe is a blink, which sets scaleY alone, so the closing hold dropped
+         * x. Every other part of the head ended with an explicit x at the duration
+         * and the eye did not — one keyframe's difference between a bound part and
+         * its carrier, which is exactly the kind of divergence the bone exists to
+         * make impossible. It was harmless here only because the held value and the
+         * last value agreed; that is luck, not a guarantee.
+         */
+        const last = track[track.length - 1];
+        if (last.time < durationSeconds) {
+            const held = { time: durationSeconds };
+            for (const kf of track) {
+                for (const [k, v] of Object.entries(kf)) {
+                    if (k === "time" || k === "easing")
+                        continue;
+                    held[k] = v;
+                }
+            }
+            if (last.easing)
+                held.easing = last.easing;
+            track.push(held);
+        }
+        ops.push({
+            tool: "pinepaper_keyframe_animate",
+            args: { itemId: `${id}_${pid}`, duration: durationSeconds, keyframes: track },
+        });
+        tracks++;
+    }
+    return { ops, parts: partIds.map((p) => `${id}_${p}`), tracks, unknownChannels };
+}
+//# sourceMappingURL=perform.js.map