fix(cli): respect transparent image pixels in occlusion audit

[miguel-heygen]

What

Make the layout occlusion audit sample the rendered alpha of <img> pixels at each text probe point instead of treating every image rectangle as fully opaque. Preserve the conservative opaque fallback for unloaded or cross-origin images whose pixels cannot be read.

Why

A full-screen transparent paper texture was reported as covering visibly unobstructed labels. The checker emitted text_occluded because <img> elements were classified as opaque based only on their bounding box.

Source report: https://heygen.slack.com/archives/C0BGC335AQY/p1783981645222299

How

• Cache a same-origin canvas per image source.
• Map viewport probe coordinates through object-fit and object-position into source pixels.
• Combine sampled alpha with the element opacity chain before classifying a point as occluded.
• Keep video/canvas behavior and unreadable-image fallback unchanged.

Test plan

• Unit tests added/updated
• Manual testing performed
• Documentation updated (not applicable)

Verified:

• bun run --filter @hyperframes/cli test -- src/commands/layout-audit.browser.test.ts (52/52)
• bun run --filter @hyperframes/cli typecheck
• bunx oxlint on changed files
• pre-commit lint, format, fallow, typecheck hooks

[miguel-heygen]

Consolidated with the related layout and contrast audit fixes in #2401, which preserves this PR’s behavior and regression coverage.