fix(skills): measure seam-gate speed-match on the frame grid

.agents/skills/cut-the-curve/SKILL.md

@@ -105,6 +105,9 @@ Mechanics:
 
 - **Mirrored eases:** exit `power4.in` + entry `power4.out`, same distance and duration —
   the two halves of one `power4.inOut`, so velocity matches exactly at the cut.
+  `seam-gate` measures something slightly different: the average velocity over ~0.1s of
+  whole frames on each side of the cut. Mirrored halves read ≈1.0 there only when the cut
+  sits on the frame grid (a multiple of 1/fps).
 - **The fade trick:** exit opacity completes at ~25–30% of its travel (fade ≈ 0.18–0.3s
   vs motion 0.3–0.34s); entry ignites at ~0.35 opacity mid-path. Time the last fading
   element to die right at the cut — a gap where nothing moves reads as dead air.

.agents/skills/motion-doctrine/references/seam-gate.md

@@ -64,7 +64,10 @@ Lives at the project root. One row per seam; this is the vector ledger as data.
 }
 ```
 
-- `cut` — seconds on the master clock, the frame the incoming side ignites.
+- `cut` — seconds on the master clock, the frame the incoming side ignites. Keep it on
+  the frame grid (a multiple of 1/fps, with the ledger `fps` matching the composition):
+  the player floors every seek to a frame, so an off-grid cut is measured around the
+  nearest frame and reported as `cut-grid`.
 - `type` — `"cut"` (default; full vector checks), `"match-cut"` / `"morph"`
   (carrier-continuity + overlap only; motion may start AT the boundary).
 - `axis` — `"x"`, `"y"`, or `"z"` (z = scale). `dir` — sign of motion:
@@ -88,10 +91,18 @@ Lives at the project root. One row per seam; this is the vector ledger as data.
 | `zero-overlap`                       | rule 6 — one side visible per frame, never both                |
 | `z-sign-scan`                        | rule 7 — incoming scene's own entrances don't fight the Z sign |
 | `carrier-*`                          | rules 3/4 — carrier rect continuity, ancestor scale included   |
+| `cut-grid` (WARN)                    | `cut` sits between frames; checks ran around the nearest frame |
 
 Velocities are measured on `getBoundingClientRect` (center for x/y, width-ratio for z),
 so ancestor wrapper transforms are automatically included.
 
+`speed-match` compares **average** velocities over whole frames, not the instantaneous
+velocity at the cut: exit from cut−0.1s to cut−1f, entry from cut+1f to cut+0.1s. The two
+windows mirror each other, so a mirrored ease pair (same ease family, distance and
+duration) reads ≈1.0. The stamped defaults (x/y `power3.in` 12% · 0.34s →
+`power4.out` 10% · 0.42s; z 0.21s → `expo.out` 0.5s) read ≈0.8 on x/y and ≈1.1 on z.
+Lengthening an entry lowers its ratio.
+
 `__player.seek()` resolves asynchronously, so the gate awaits each seek before it reads. If every
 seam fails the same way (`exit-visible op 0.00`, `exit-moving 0 px/s`) and `probe` finds no movers,
 the reads landed before the seek did and the gate measured t=0, not the film.

.agents/skills/motion-doctrine/scripts/seam-gate.mjs

@@ -366,8 +366,7 @@ const fmtV = (v, axis) =>
 async function verify() {
   const ledgerPath = flag("ledger", "ledger.json");
   const ledger = JSON.parse(readFileSync(ledgerPath, "utf8"));
-  const fps = ledger.fps || FPS,
-    dt = 1 / fps;
+  const fps = ledger.fps || FPS;
   const compUrl = await ensureServer();
   const { evalJs } = await openPage(compUrl);
   const results = [];
@@ -377,10 +376,17 @@ async function verify() {
     const add = (check, status, detail) => rows.push({ check, status, detail });
     const cut = seam.cut;
     const type = seam.type || "cut";
-    const tA1 = Math.max(0, cut - 0.1),
-      tA2 = Math.max(0, cut - dt);
-    const tB1 = cut + dt,
-      tB2 = cut + 0.1;
+    // The player floors every seek to the frame grid, so sampling at cut ± k/fps for an
+    // off-grid cut shifts both exit samples up to a frame early and the first entry sample
+    // onto the cut itself, under-reading exit and over-reading entry. Sample whole frames
+    // placed symmetrically around the nearest frame instead.
+    const cutFrame = Math.round(cut * fps);
+    const offGrid = Math.abs(cut * fps - cutFrame) > 1e-3;
+    const win = Math.max(1, Math.round(0.1 * fps));
+    const tA1 = Math.max(0, cutFrame - win) / fps,
+      tA2 = Math.max(0, cutFrame - 1) / fps;
+    const tB1 = (cutFrame + 1) / fps,
+      tB2 = (cutFrame + win) / fps;
     const sels = [
       seam.exit?.selector,
       seam.entry?.selector,
@@ -390,6 +396,12 @@ async function verify() {
     const S = {};
     for (const t of [tA1, tA2, tB1, tB2])
       S[t] = await evalJs(`__seamGate.sample(${t}, ${JSON.stringify([...new Set(sels)])})`, true);
+    if (offGrid)
+      add(
+        "cut-grid",
+        "WARN",
+        `cut ${cut}s is between frames at ${fps} fps; measured around ${(cutFrame / fps).toFixed(3)}s — put cuts on the frame grid`,
+      );
 
     if (type === "cut") {
       const ex = seam.exit,

.claude/skills/cut-the-curve/SKILL.md

@@ -105,6 +105,9 @@ Mechanics:
 
 - **Mirrored eases:** exit `power4.in` + entry `power4.out`, same distance and duration —
   the two halves of one `power4.inOut`, so velocity matches exactly at the cut.
+  `seam-gate` measures something slightly different: the average velocity over ~0.1s of
+  whole frames on each side of the cut. Mirrored halves read ≈1.0 there only when the cut
+  sits on the frame grid (a multiple of 1/fps).
 - **The fade trick:** exit opacity completes at ~25–30% of its travel (fade ≈ 0.18–0.3s
   vs motion 0.3–0.34s); entry ignites at ~0.35 opacity mid-path. Time the last fading
   element to die right at the cut — a gap where nothing moves reads as dead air.

.claude/skills/motion-doctrine/references/seam-gate.md

@@ -64,7 +64,10 @@ Lives at the project root. One row per seam; this is the vector ledger as data.
 }
 ```
 
-- `cut` — seconds on the master clock, the frame the incoming side ignites.
+- `cut` — seconds on the master clock, the frame the incoming side ignites. Keep it on
+  the frame grid (a multiple of 1/fps, with the ledger `fps` matching the composition):
+  the player floors every seek to a frame, so an off-grid cut is measured around the
+  nearest frame and reported as `cut-grid`.
 - `type` — `"cut"` (default; full vector checks), `"match-cut"` / `"morph"`
   (carrier-continuity + overlap only; motion may start AT the boundary).
 - `axis` — `"x"`, `"y"`, or `"z"` (z = scale). `dir` — sign of motion:
@@ -88,10 +91,18 @@ Lives at the project root. One row per seam; this is the vector ledger as data.
 | `zero-overlap`                       | rule 6 — one side visible per frame, never both                |
 | `z-sign-scan`                        | rule 7 — incoming scene's own entrances don't fight the Z sign |
 | `carrier-*`                          | rules 3/4 — carrier rect continuity, ancestor scale included   |
+| `cut-grid` (WARN)                    | `cut` sits between frames; checks ran around the nearest frame |
 
 Velocities are measured on `getBoundingClientRect` (center for x/y, width-ratio for z),
 so ancestor wrapper transforms are automatically included.
 
+`speed-match` compares **average** velocities over whole frames, not the instantaneous
+velocity at the cut: exit from cut−0.1s to cut−1f, entry from cut+1f to cut+0.1s. The two
+windows mirror each other, so a mirrored ease pair (same ease family, distance and
+duration) reads ≈1.0. The stamped defaults (x/y `power3.in` 12% · 0.34s →
+`power4.out` 10% · 0.42s; z 0.21s → `expo.out` 0.5s) read ≈0.8 on x/y and ≈1.1 on z.
+Lengthening an entry lowers its ratio.
+
 `__player.seek()` resolves asynchronously, so the gate awaits each seek before it reads. If every
 seam fails the same way (`exit-visible op 0.00`, `exit-moving 0 px/s`) and `probe` finds no movers,
 the reads landed before the seek did and the gate measured t=0, not the film.

.claude/skills/motion-doctrine/scripts/seam-gate.mjs

@@ -366,8 +366,7 @@ const fmtV = (v, axis) =>
 async function verify() {
   const ledgerPath = flag("ledger", "ledger.json");
   const ledger = JSON.parse(readFileSync(ledgerPath, "utf8"));
-  const fps = ledger.fps || FPS,
-    dt = 1 / fps;
+  const fps = ledger.fps || FPS;
   const compUrl = await ensureServer();
   const { evalJs } = await openPage(compUrl);
   const results = [];
@@ -377,10 +376,17 @@ async function verify() {
     const add = (check, status, detail) => rows.push({ check, status, detail });
     const cut = seam.cut;
     const type = seam.type || "cut";
-    const tA1 = Math.max(0, cut - 0.1),
-      tA2 = Math.max(0, cut - dt);
-    const tB1 = cut + dt,
-      tB2 = cut + 0.1;
+    // The player floors every seek to the frame grid, so sampling at cut ± k/fps for an
+    // off-grid cut shifts both exit samples up to a frame early and the first entry sample
+    // onto the cut itself, under-reading exit and over-reading entry. Sample whole frames
+    // placed symmetrically around the nearest frame instead.
+    const cutFrame = Math.round(cut * fps);
+    const offGrid = Math.abs(cut * fps - cutFrame) > 1e-3;
+    const win = Math.max(1, Math.round(0.1 * fps));
+    const tA1 = Math.max(0, cutFrame - win) / fps,
+      tA2 = Math.max(0, cutFrame - 1) / fps;
+    const tB1 = (cutFrame + 1) / fps,
+      tB2 = (cutFrame + win) / fps;
     const sels = [
       seam.exit?.selector,
       seam.entry?.selector,
@@ -390,6 +396,12 @@ async function verify() {
     const S = {};
     for (const t of [tA1, tA2, tB1, tB2])
       S[t] = await evalJs(`__seamGate.sample(${t}, ${JSON.stringify([...new Set(sels)])})`, true);
+    if (offGrid)
+      add(
+        "cut-grid",
+        "WARN",
+        `cut ${cut}s is between frames at ${fps} fps; measured around ${(cutFrame / fps).toFixed(3)}s — put cuts on the frame grid`,
+      );
 
     if (type === "cut") {
       const ex = seam.exit,