fix(browser): preserve late-failure teardown evidence

[seonghobae]

Partial implementation of #28. Current head bb5e8f834c37f9ce35f84db8ed1146da3659d6aa ordinarily merges prior child a1341e9c8595c38446d64aaaadeab999062eab57 with current #144 3c0c5c363c2da0b4e8a621226e2f7766c735b743.

Buyer / reliability boundary

This Draft preserves observed browser-root and captured Chromium process-set termination outcomes when ordinary controlled browser work fails late. The forced-close lane accepts success only after root and sampled-set termination are both observed. Profile cleanup is separate evidence and never changes a failed task into a pass.

Test-first integration / current local proof

The predecessor passed 200 Python contracts but failed an expected-five assertion by collecting only three inherited failure-path contracts. Ordinary parent adoption carries the parent's real failure-path and release-record regressions without changing the production runner or either child-owned process-set/forced-close regression. It required no conflict resolution or new runtime abstraction.

Current integrated verification passes five focused inherited contracts, all 202 Python contracts, compileall, Rust 1.97.1 formatting, locked workspace/all-target check and tests, strict Clippy, rustdoc, and pinned-nightly owned-production coverage enforcement. All 415 functions, 3,555 lines, 4,444 regions and 476 branches are covered at 100%.

Exact-head CI 33956596014 and Manifest V3 Compatibility 33956596013 are queued, not GREEN. Predecessor runs 33916706051 and 33916706095 were cancelled. Local verification does not replace hosted compatibility or protected-main acceptance.

Informational review / explicit limits

Ten controlled probes execute the actual browser-pass and outer-trial flows, with browser/OS observations injected rather than launching Chromium. Cleanup ordering, root survival, sampled-set survival, observation PermissionError and incomplete/interrupted execution are checked in both lanes. Private exception messages are absent from returned evidence.

Late ordinary failure retains observed true/false root and captured-set results. Incomplete capture omits set evidence. The forced-close no-exception path observes both identities after cleanup and fails if either remains live. Completing session/driver cleanup does not prove that those processes are already dead. A forced-close interruption skips post-finally observers and leaves termination unproven; observer errors produce only bounded fallback failure evidence. #146 separately owns retained forced-close failure/driver-cleanup evidence and must adopt this current parent.

Both informational threads are resolved with these evidence limits. Thread resolution is not counted approval.

Historical provenance — not current acceptance

Original reconstruction RED 0dcd3882678a11114b73dcc9d71bbac12284c71e merged old child 312cd4a7f31fa43d42dadd8f43bafffae8f82e62 and then-parent 5ed8fb140f031fbc637175169fd7796692e6d6c3; CI 32610336594, job 97122222101, failed the missing child contracts. Production repair 8f3b8041c292967abab710d387baec140af942aa later passed CI 32610769453 and compatibility 32610769483, with coverage artifact 9485466766, ZIP SHA-256 09cfd1f202ac02c9efafdb159baa74fbaf1ed83ed6ae3b1dc5af9c50b02945ed. Those historical results, then-current descendant claims and old reviews do not transfer to the current head.

Authority / truth boundary

Keep Draft while dependencies remain unprotected and current checks/eligible approval remain unmet. This bounded Linux PID/start-time compatibility evidence does not establish cgroup/task ownership, processes appearing after sampling, cross-platform supervision, authenticated Chromium/ChromeDriver provenance, OS-wide orphan absence or a shipped product browser API. No protected merge, self-approval, bypass, force-push, workflow/ruleset/secret mutation, tag, release or publication is included.

Instrumentation limit: a fresh report from the pinned cargo-llvm-cov 0.8.6 branch data emits warning: --branch option is unstable. The repository's four-dimension numerical coverage verifier passes, but the instrumentation is experimental and this is not warning-free measurement or release acceptance. The warning is recorded, not suppressed or bypassed; Rust compiler/Clippy/rustdoc results remain separate.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[devin-ai-integration]

Devin Review found 2 potential issues.