test(studio): measure render drift against a producer frame in the edit accuracy bench

[miguel-heygen]

What

Adds render drift to the edit accuracy benchmark, its fifth metric. After each case's edit and a fresh Studio load, the producer renders the saved file at the same playhead, and the benchmark compares where the element lands in that frame with where the reloaded preview shows it. baseline.json gains the render column. It comes from the same full run on main 14e8700 as #4760's baseline: render passes 618/660.

Stacked on the edit accuracy benchmark PR (#4760); review that one first.

Why

A preview can show one thing while the export shows another. The other metrics only look at Studio, so they cannot catch a case where the preview and the rendered video disagree about the same file.

Related work

Refs #4760 (the benchmark this extends). The CI gate follows in its own PR.

How

• The producer frame comes from the engine's own capture session (createCaptureSession + captureFrameToBuffer), the path scripts/preview-capture.ts uses. It runs in the same pinned headless shell Studio is driven in, with JPEG output. PNG is the engine's transparent mode.
• The element's box is found by its flat fixture colour over the whole frame:
  • coverage per pixel comes from luminance, which JPEG keeps at full resolution;
  • area moments give the centre and both sides, corrected by 1/12 px² for pixel coverage;
  • so a rotated or cropped rectangle is measured the same way as an axis-aligned one.
• The reloaded preview is measured once it has held still for 1 s, as every snapshot in test(studio): an edit accuracy benchmark for manual editing #4760 is. If the reload never settles, render fails with it, flagged unsettled.
• If the producer cannot render the case, only render fails: render is null and renderError holds the message. The case's other metrics still count, and the table counts these cases.
• Rotated elements compare axis-aligned boxes. The producer page's DOM rect is recorded as a diagnostic only, because it ignores clip-path.
• A composition with no GSAP never registers a timeline, so the capture session would wait its full 45 s playerReadyTimeout. The bench passes 10 s instead. The frame is identical: the measured box matches to every printed digit.

Test plan

• A producer failure injected into two cases fails render alone (renderError). Undo, drop and reload still score, and there is no harness error.
• Pixel box against the producer's own DOM rect on an unrotated, uncropped case: within 0.01 px.
• A 1 px shift injected into the producer page reads as render drift 1.02 (clean 0.02). A 3 px shift injected into the reloaded preview reads as 3.01.
• Restacked on the settle fix: a 4-case run gives render 0.03 px on the clean cases and 50.0 px on resize-hold-center-r0-root-z100, with no harness errors.
• Full grid on main (before the settle fix): render passes 618/660. Worst is 110.05 px (resize-hold-pct-r30-root-z100): Studio's preview does not show the size the producer renders from the same file. The resize writes tl.set(... {width, height}, 0) into a zero-length timeline.
• oxlint, oxfmt, fallow audit --base origin/main and the comment ratchet pass.
• Unit tests added/updated. None: the injections above prove the metric.
• Comments follow CONTRIBUTING.md "Comments".

Before

Without render drift, the benchmark saw only Studio. This is the reloaded Studio preview for the worst case, which the Studio-only metrics accept.

After

The producer frame for the same saved file, at the same playhead, shows the element resized. Render drift now reports the 110 px gap, and the table has the render column.