feat(vrt): per-test render-mode gating for Canvas2D visual regression

[chiefcll]

Summary

First step toward running the visual regression suite against both backends (the runner already supports --renderMode canvas|all; what was missing is test-level compatibility). This PR makes the full example suite run clean under the Canvas2D backend: 151 snapshots, zero throws, zero page errors.

Gating mechanism

• Test modules can export renderModes: ('webgl' | 'canvas')[] (default: all modes). runAutomation skips non-matching tests and logs the skip so excluded coverage is visible in runner output.
• ExampleSettings gains a renderMode field so tests can branch per backend.

WebGL-only marks (3 tests)

• rtt-dimension, rtt-spritemap — renderToTexture is a documented noop on CanvasRenderer; the tests render blank.
• text-ssdf — SSDF atlas fonts exist only for the SDF/WebGL text renderer (confirmed blank under canvas).

Made mode-aware (17 tests)

The text/stress tests that deliberately render sdf and canvas text-renderer variants side by side threw under the canvas backend (No compatible text renderer found — the SDF engine is not registered there). They now skip the sdf variants when renderMode === 'canvas' and keep both under webgl. Notable cases:

• text-offscreen-move had a subheader hardcoded to 'sdf' even on its canvas pages.
• stress-animation / stress-tv use SDF text as HUD/content rather than as a comparison variant — they substitute the canvas text renderer wholesale under canvas.
• render-settings used the SDF-only Ubuntu-ssdf font family for every label, leaving all text as blank/placeholder boxes under canvas; it now falls back to the canvas-loaded Ubuntu face.

Verification

• WebGL output is unchanged: full visual regression compare run against pre-existing local baselines passes for every edited test. (5 unrelated failures are stale local baselines: 3 tests newer than the local capture, 2 gradient tests whose shaders changed in refactor(shaders): make all WebGL fragment shaders branchless for Mali 400 #118 after that capture.)
• Canvas suite runs clean: pnpm test:visual --capture --renderMode canvas completes 151 snapshots with no page errors and no automation throws, including spot-review of shader, clipping, scaling, and text output against webgl counterparts.

Next steps (separate PRs)

1. Capture and certify chromium-ci-canvas baselines via the Docker flow (capture twice + pixel-diff to guard against a texture-placeholder timing flake observed once during triage).
2. Flip the runner's default renderMode to all and split the CI visual job into a webgl/canvas matrix.

🤖 Generated with Claude Code