fix(capture): retry crashed renderers in fresh contexts

[chubes4]

Summary

Refs #567.

• Track the capture page's real crash event and retry that viewport once in a fresh context while the browser is still connected. Reuse the existing shared browser replacement when the entire browser disconnects.
• Stop subsequent best-effort capture stages after a detected crash, including chrome-fidelity after failed analysis.
• Retain failed retry diagnostics with attempt: 2; a persistent crash remains failed while later routes can succeed.
• Add a child-process regression using real Chromium, a neutral local HTTP fixture, CDP Page.crash, pending response-body and evaluation work, and healthy later-route HTML/screenshots. Existing Page doubles gain the real event-listener contract.

Evidence

Pinned Linux reproduction: Node 22.23.2, Playwright 1.62.1, mcr.microsoft.com/playwright:v1.62.1-noble, four CPUs, 2 GiB memory, 256 MiB shared memory.

• Untouched release 0.18.7 / 4ddeafe: both new regressions fail; capture keeps issuing chrome-fidelity work on the crashed page and does not retry the connected renderer.
• Candidate: 2/2 regressions pass, child processes exit 0. Transient crash: two attempts, one crash, two captured routes, zero failures/restarts. Persistent crash: two attempts, two crashes, one captured later route, one final source renderer crashed failure at attempt two, zero browser restarts.
• Pending fetch/evaluate/body/CDP operations reject and settle, contexts are zero after each route, and the browser is disconnected at completion.
• Linux cgroup memory.events: max=0, oom=0, oom_kill=0.
• Local real-browser regression also passed on the production Node/Playwright versions. Complete repository gates passed with locked dependencies on Node 22.23.2:
  • npm test -- --maxWorkers=4: 150 files / 1,676 tests passed.
  • npm run build: passed; both bundles regenerated.
  • npm run test:package: passed (installed/relocated browserless and browser-backed workflows).
  • git diff --check: passed.

Generated dist/ was restored after verification per the repository's feature-PR policy.

Evidence boundary

This fixes the reproduced DLA page-retry gap. It does not establish that DLA can contain the production _CRSession._onMessage fatal assertion: ordinary neutral crashes did not reproduce that assertion, including 20 Linux crash cycles with pending work and a 100-evaluation burst. OOM and the production crash's initiating operation remain unproven.

Playwright's owning callback fix is microsoft/playwright#42936, commit 132be89d216e50964a564bbe0f63c6847e8d1338: ignore late response IDs on an already-crashed session after its pending callbacks have been rejected. That guard is absent in released 1.62.1 and 1.63.0, and present on upstream main. The host-owned follow-up is to select a toolchain containing that commit, prove its late-reply regression on Linux, and rerun the live capture. This PR makes no dependency-upgrade or fatal-assertion containment claim.

Execution provenance / AI assistance

Direct finalization outside Homeboy was explicitly authorized on 2026-10-06 after original Cook agent-task-55d0b962-b052-4b97-aebe-5a5912a144ea failed. OpenAI gpt-6.1-sol via OpenCode performed scoped source investigation, neutral browser reproductions, implementation, verification, and drafted this PR. Homeboy SSH/file transport was used for the isolated Linux evidence workload; Cook did not own finalization. Chris Huber authorized the task and draft PR. Live site capture and independent binding/popup changes remain owned by the parent task.