test(hdr-regression): tighten Window F maxFrameFailures budget after Chunk 4 fix (#375)

## Summary

Tighten `hdr-regression` Window F `maxFrameFailures` from 5 → 0 now that Chunk 4 (matrix3d support + scene initial-state) has landed.

## Why

Window F (transform + scale + border-radius on the video itself) was the remaining known-fail in the `hdr-regression` suite, baked into the golden so the suite stayed green while Chunk 4 was outstanding.

After Chunk 4 fixed `parseTransformMatrix` (matrix3d support) and the shader-transitions initial-state, re-running the suite shows **0 failed frames** against the existing golden — the encoder is byte-deterministic, and Window F's GSAP rotation/scale happens to emit 2D `matrix()` rather than `matrix3d()`, so the same golden is still correct after the fix. Tightening the budget catches any drift in the layered HDR compositor immediately.

## What changed

- `tests/hdr-regression/meta.json`: `maxFrameFailures` 5 → 0 (matches `hdr-hlg-regression`).
- `tests/hdr-regression/README.md`: Window F row + Fix history section updated to reflect the new state.

## Test plan

- [x] `bun run test --filter hdr-regression` — passes with 0 failed frames at the new budget.

## Stack

Follow-up to Chunk 4 (transform & clipping). Reviewable separately so the budget tightening is decoupled from the code fix.

packages/producer/tests/hdr-hlg-regression/README.md

@@ -27,13 +27,6 @@ The suite is intentionally short (5 s, two windows) — it exists to detect
 regressions in the HLG-specific code path, not to enumerate every composition
 shape (those live in `hdr-regression`).
 
-## Tolerance
-
-`maxFrameFailures` is **0** here. HLG is a pure pass-through path — no known
-failures, no transcoder workarounds — and HEVC encoding against the rendered
-`rgb48le` buffer is byte-deterministic on the same fixture. Any drift is a
-real regression, not codec noise, so the budget is the strictest possible.
-
 ## Fixture
 
 `src/hdr-hlg-clip.mp4` — last 5 seconds of a user-recorded HEVC HLG clip,

packages/producer/tests/hdr-regression/README.md

@@ -15,11 +15,17 @@ covering the highest-value HDR compositing shapes. 10s / 300 frames at 30fps.
 ## Fixtures
 
 - `src/hdr-clip.mp4` — short HEVC Main10 / BT.2020 PQ clip with a moving
-  bright gradient (see `NOTICE.md` for attribution).
-- `src/hdr-photo-pq.png` — 256x144 16-bit RGB PNG with a hand-injected `cICP`
+  bright gradient (see `NOTICE.md` for attribution). Reused across windows
+  A–G and as scene A of the window-H shader transition.
+- `src/hdr-photo-pq.png` — 256×144 16-bit RGB PNG with a hand-injected `cICP`
   chunk (primaries=BT.2020, transfer=SMPTE ST 2084, matrix=GBR, range=full).
+  Used as scene B of the window-H shader transition.
 
-To regenerate the PNG fixture:
+ffmpeg is **not** used to generate the PNG because it does not embed `cICP`
+in PNGs — without that chunk Chromium would not treat the file as HDR and the
+test would silently fall back to SDR.
+
+To regenerate the PNG fixture (deterministic, byte-for-byte stable):
 
 ```bash
 python3 packages/producer/tests/hdr-regression/scripts/generate-hdr-photo-pq.py

packages/producer/tests/hdr-regression/meta.json

@@ -1,9 +1,9 @@
 {
   "name": "hdr-regression",
-  "description": "HDR10 (BT.2020 PQ) regression suite. Four windows (A–D) cover the highest-value HDR compositing shapes: direct <video> opacity tween (A), z-order DOM/HDR/DOM sandwich (B), transform + border-radius (C), and a shader transition between HDR video and HDR PQ image (D). 10s / 300 frames at 30fps.",
+  "description": "Comprehensive HDR10 (BT.2020 PQ) regression suite. Eight back-to-back 'windows' (A–H) lock down the layered HDR compositing pipeline against the most common composition shapes: baseline pass-through, wrapper-vs-direct opacity, z-order sandwiches, side-by-side multi-source, transform + border-radius, object-fit letterbox, and a shader transition between an HDR video and an HDR PQ image. Window C (direct <video> opacity) is fixed by Chunk 1. Window F (transform + border-radius) is fixed by Chunk 4. The golden is byte-deterministic and `maxFrameFailures` is 0 — any drift in the layered HDR compositor is caught immediately.",
   "tags": ["regression", "hdr"],
   "minPsnr": 28,
-  "maxFrameFailures": 5,
+  "maxFrameFailures": 0,
   "minAudioCorrelation": 0,
   "maxAudioLagWindows": 1,
   "renderConfig": {

packages/producer/tests/hdr-regression/output/compiled.html

@@ -189,7 +189,7 @@ html2canvas/dist/html2canvas.esm.js:
   </head>
   <body>
     <div id="main" data-composition-id="hdr-regression" data-start="0" data-duration="6" data-width="1920" data-height="1080">
-      <!-- Window A · Baseline HDR + direct opacity tween · 0.0–2.5s -->
+      <!-- Window A · Baseline HDR + direct opacity tween · 0.0–1.5s -->
       <video id="wa-video" class="clip hdr-video" data-start="0" data-duration="1.5" data-track-index="0" src="hdr-clip.mp4" muted playsinline data-has-audio="false" data-end="1.5"></video>
       <div class="label label-tl clip" data-start="0" data-duration="1.5">
         A · Baseline HDR + opacity tween

packages/producer/tests/hdr-regression/output/output.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e95e62592900fadeb78b7366134f6d3c78a647cc75f9d4f8c84646ef6177952b
-size 1755367
+oid sha256:ef13545af8288619f3c493d9b684f49ed7de007229f16970f34c741d3e6d2a9d
+size 1746876