fix(engine): retry capture screenshot timeouts

[miguel-heygen]

What

Screenshot protocol timeouts now enter the existing adaptive capture retry, allowing the renderer to reduce worker concurrency instead of hard-failing a valid long render.

Why

Puppeteer's Page.captureScreenshot timed out signature fell through to authoring, so parallel capture treated a recoverable CDP timeout as fatal.

How

Classify the exact screenshot method timeout as protocol_timeout. Keep pressure-aware auto worker sizing as separate follow-up scope.

Test plan

• Unit tests added/updated: engine classifier RED-to-GREEN (14/14)
• Manual testing performed: exact wrapped parallel-worker runtime probe returned recoverable=true
• Documentation updated (not applicable; internal failure classification)

The full producer suite could not start in the task worktree because its existing dependency payload omitted generated and package dist artifacts; the exact producer recovery assertion is included for CI.

[miga-heygen]

Checked the new /Page\.captureScreenshot timed out/i pattern against the other classifier buckets (transient_browser, memory_exhaustion, verification) in captureFailure.ts — no overlap, so this can't misclassify into a different bucket. Small, additive, well-targeted: routes the exact Puppeteer CDP screenshot-timeout signature into the existing protocol_timeout recoverable path instead of falling through to fatal authoring. Tests cover both the classifier and the parallel-capture recovery check. LGTM.