fix(ci): pin chrome-headless-shell to fix regression baseline drift

[jrusso1020]

What

Pin chrome-headless-shell to 148.0.7778.167 in Dockerfile.test (was @stable) and regenerate the style-12-prod golden baseline under the pin.

Why

Dockerfile.test:56 installed chrome-headless-shell@stable via @puppeteer/browsers. @stable is a moving tag — every Chrome stable bump shifts pixel output enough to fail PSNR on the golden baselines, so the regression suite silently broke whenever the test image was rebuilt against a freshly-promoted stable. Pinning to a numeric version makes the test environment reproducible and ties Chrome bumps to explicit baseline regenerations.

While here, style-12-prod is also regenerated. PR #918 regenerated it once at b9bdc80, but that commit landed before the refactor: extract shared inlineSubCompositions from bundler and producer (581e7a7) and the linkedom-fragment fix (754b0ed) in the same stack. The refactor changes __hfRootSelector from null to a scoped [data-composition-id="..."] selector in inlined sub-compositions, which affects the render. style-12-prod was the one fixture in that stack that didn't get a second regen pass after the refactor, so it has been failing on plain origin/main from frame 8.26s onward (PSNR ~13 — the mondrian-colors blocks no longer match expected).

How

• Pin chrome-headless-shell@148.0.7778.167. This is what @stable currently resolves to per googlechromelabs.github.io/chrome-for-testing/last-known-good-versions.json and matches the version Miguel's earlier baseline regens were captured against.
• Regenerate packages/producer/tests/style-12-prod/output/{compiled.html,output.mp4} via docker:test:update. The compiled.html diff is small (~18 lines): scoped __hfRootSelector values now match what the post-refactor compiler emits.
• Added a comment on the pin: future Chrome bumps must be paired with docker:test:update.

Test plan

bun run --cwd packages/producer docker:build:test
bun run --cwd packages/producer docker:test style-12-prod      # ✅ PSNR 62-102 dB, 0 failed frames
bun run --cwd packages/producer docker:test style-3-prod       # ✅ unchanged baselines still pass under pin
bun run --cwd packages/producer docker:test style-5-prod       # ✅ unchanged baselines still pass under pin
bun run --cwd packages/producer docker:test sub-composition-video  # ✅ unchanged baselines still pass under pin

• Manual regression run (style-12-prod) green inside Dockerfile.test
• Spot-checked 3 other recently-regenerated baselines under the pin
• Unit tests added/updated — N/A (CI/baseline change)
• Documentation updated — N/A

[jrusso1020]

Closing in favor of two focused, parallel PRs:

• test: regenerate style-1-prod and style-12-prod baselines #925 (Miguel) — regenerates the style-1-prod and style-12-prod baselines that drifted after the sub-composition compiler refactor in fix: hold external sub-compositions in render mode + regenerate baseline #918.
• fix(ci): pin chrome-headless-shell + clamp PSNR checkpoint to a valid frame #926 — keeps the structural fixes from this PR: the chrome-headless-shell@148.0.7778.167 pin in Dockerfile.test, and the regression-harness.ts PSNR-parse fix that unblocks many-cuts (a pre-existing failure on main from fix: hold external sub-compositions in render mode + regenerate baseline #918 admin-merging through shard-2).

Splitting avoids the merge conflict that would happen on output.mp4 if both this PR and #925 land — same Chrome version, but ffmpeg encode isn't bit-deterministic across runs, so the LFS oids diverge.