fix: preserve streamed cancellation and pending guardrail cleanup

[jbeckwith-oai]

This pull request fixes consumer cancellation being swallowed while stream_events() waits for a pending input guardrail or run-loop finalization. It also prevents cancellation of a pending parallel input guardrail from skipping computer disposal and trace/span completion.

Terminal waits distinguish consumer cancellation from child cancellation, settle the cancelled child, and await registered provider/sandbox cleanup before propagating cancellation. The guardrail-verdict helper tolerates cancelled children only during final cleanup, preserving cancellation before turn-output persistence. Explicit result.cancel() behavior and existing late-error handling remain intact.

Five event-gated regressions exercise the public runner's terminal waits, disposal, trace closure, registered cleanup, and session persistence. Cancellation during already-running asynchronous cleanup and repeated cancellation remain separate work; this change does not guarantee that interrupted cleanup completes.

Test plan

• The original three regressions fail against the original runtime; the two feedback regressions fail against the initial PR runtime. All pass with this fix.
• Focused cancellation, guardrail, computer, provider, and sandbox tests: 309 passed, 6 native sandbox skips.
• Streamed-runner tests, including finalizer error logging: 206 passed.
• Two independent reviews of the final diff: clean.
• Full verification: formatting, lint, mypy, pyright, and tests passed (11,510 parallel tests and 89 serial tests; 69 skipped).
• Native macOS sandbox tests were excluded locally under the repository's Codex policy; the dedicated CI job covers them.

Issue number

Closes #5213

Checks

• I've added new tests, if relevant
• I've run .agents/skills/code-change-verification/scripts/run.sh
• I've confirmed all verification steps pass
• If using Codex, I've run /review before submitting this PR

The implementation-final-review workflow completed with two independent clean reviews; the separate /review command was not invoked.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-28T17:15:42.666839Z	846e167	New commits
🔒 Security Review	✅ Completed 2026-09-28T17:16:06.321086Z	846e167	New commits

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 211fe74d44

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[markstuart-oai]

Reviewed the full diff at 846e167. The terminal waits now distinguish consumer cancellation from child cancellation, preserve guardrail cancellation before output persistence, and reach the existing registered cleanup owners. The event-gated regressions cover the intended lifecycle boundaries. No actionable correctness or structural findings within the stated scope; interruption of cleanup already in progress remains outside this fix.

Validation: pinned-source review and all 22 successful checks on this commit, including Windows and native macOS sandbox coverage. I did not run tests locally.

[gomission]

@jbeckwith-oai, thanks for landing the fix. The run_loop regression covers the additional no-guardrail/active-disposal checkpoint we documented in #5213. The PR also keeps the distinction between propagating cancellation and completing already-running cleanup clear.

If our reproduction informed that coverage or scope, could you add a short acknowledgment of @gomission in the PR description, linked to that comment? The attribution would be for the additional reproduction and boundary check; @hsusul supplied the original report and proposed fix, and you implemented and landed the merged change.