fix(ci): pin chrome-headless-shell + clamp PSNR checkpoint to a valid frame

[jrusso1020]

What

Two narrow fixes to keep the regression suite green and reproducible. Stale baselines from the sub-comp refactor are being regenerated separately in #925; this PR is just the structural fixes that #925 can't make on its own.

1. Pin chrome-headless-shell in Dockerfile.test to 148.0.7778.167 instead of @stable.
2. Clamp the last PSNR checkpoint to a frame the video stream actually contains so the harness stops crashing on many-cuts.

Why

Chrome pin

@stable is a moving tag. Every Chrome stable promotion shifts pixel output enough to fail PSNR on the golden baselines, so the regression suite silently broke whenever Docker.test rebuilt against a freshly-promoted stable. Pinning to the version @stable currently resolves to (matching what main's regenerated baselines were captured under) makes Chrome bumps an explicit, batched-with-baseline-regen action. The comment on the RUN line spells out the bump procedure.

PSNR-parse crash on many-cuts

runTestSuite samples 100 checkpoints across min(rendered, snapshot) container duration. Container duration includes audio padding past the last video frame — many-cuts is 5.654s container vs 5.6s of video at 30fps = 168 frames. At i=99 the raw container duration mapped to time 5.59746s → frame index 168 (round(5.59746 × 30)), one past the last frame the stream contains. ffmpeg's psnr filter emits no average: line for a non-existent frame, so the harness crashed with Unable to parse PSNR output at 5.59746s — pre-existing on plain origin/main (#918 admin-merged through this same failure on shard-2). Miguel's regen via --update doesn't catch it because --update only writes the snapshot; it doesn't validate.

Subtracting one frame interval from the sampling duration guarantees the last checkpoint always lands on a real frame.

How

• Dockerfile.test: chrome-headless-shell@stable → chrome-headless-shell@148.0.7778.167 (+ a comment documenting the bump procedure).
• packages/producer/src/regression-harness.ts: introduce sampleDuration = max(0, videoDuration - 1/fps) and use it in place of videoDuration when computing the per-checkpoint time. Also reuses the already-resolved fps variable inside the loop (was being recomputed via fpsToNumber(...) on every call to psnrAtCheckpoint).

Test plan

Local Docker reproduction:

bun run --cwd packages/producer docker:build:test
bun run --cwd packages/producer docker:test many-cuts                    # ✅ green
bun run --cwd packages/producer docker:test style-3-prod style-5-prod \
                                            sub-composition-video        # ✅ green

• many-cuts no longer crashes with Unable to parse PSNR output
• Three sub-composition-using baselines that regenerated cleanly in fix: hold external sub-compositions in render mode + regenerate baseline #918 still pass under the pin
• Unit tests added/updated — N/A (CI/harness)
• Documentation updated — N/A

Coordination with #925

#925 (Miguel) regenerates the style-1-prod and style-12-prod baselines that drifted after #918's compiler refactor. That's content-level regen; this PR is structural. They're independent and either can land first — the other will then merge cleanly. Closes #919 (which had both changes plus baselines that would conflict with #925).