feat(vrt): run visual regression on the Canvas2D backend in CI

[chiefcll]

Summary

Completes the canvas visual-regression rollout started in #125. CI now tests both rendering backends on every PR.

Certified canvas baselines

• 151 chromium-ci-canvas snapshots, captured in the Docker CI environment.
• Determinism check: captured twice in separate container runs and pixel-diffed — all 151 byte-identical. (Protocol adopted after a one-off texture-placeholder timing flake during the feat(vrt): per-test render-mode gating for Canvas2D visual regression #125 triage; recommend the same capture-twice for future baseline updates.)
• Reviewed against the WebGL counterparts: shaders (rounded, border, shadow, gradients, blur, hole-punch, edge-fade), clipping, text, resize-mode cover/clip, and scaling all render correctly; text differs only by expected canvas-vs-SDF rasterization.

Runner default → all

pnpm test:visual now runs webgl and canvas by default (the option's comment anticipated this flip once baselines existed). Single-backend runs remain available via --renderMode webgl|canvas.

CI workflow matrix

• Unit tests split into their own job.
• Visual regression runs as a renderMode: [webgl, canvas] matrix with fail-fast: false — a failing check names the backend that regressed, and both legs always report.
• Failure artifacts are per-mode (failed-results-webgl / failed-results-canvas), which also fixes the artifact-name collision the matrix would otherwise hit.

Docs

visual-regression/README.md and CLAUDE.md now describe the dual-backend workflow: new rendering features need snapshots for both backends, or a renderModes = ['webgl'] export with a stated reason.

Verification

CI-style Docker compare run (RUNTIME_ENV=ci, default all mode) over the committed baselines:

• webgl: 180/180 snapshots pass
• canvas: 151/151 snapshots pass
• zero page errors in both legs

🤖 Generated with Claude Code