fix(responses): preserve terminal errors through stream cleanup

[jbeckwith-oai]

Summary

This pull request fixes Responses streaming errors being masked by transport teardown failures. After yielding a provider failure event, the model now closes the stream explicitly and raises the saved ModelBehaviorError, preserving the provider's diagnostics. It also recognizes error as terminal in the request-ID stream wrapper.

Adding error to the wrapper alone fixes cleanup-callback failures but leaves real HTTP teardown failures exposed during iteration. The regression tests exercise both boundaries with synthetic payloads and the real OpenAI client, while retaining pre-terminal transport failures and the successful-terminal control.

Test plan

• Confirmed the callback-error case and real HTTP error/response.failed/response.incomplete cases fail before the fix.
• Responses model tests: 192 passed, including cleanup and cancellation coverage.
• Independent code review: no actionable findings; 11 focused checks passed.
• Full repository verification: formatting, lint, mypy, pyright, and parallel/serial test suites passed. Native macOS sandbox tests are delegated to the dedicated CI job per repository policy.

Issue number

Fixes #5181

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
• I've completed an independent code review before submitting this PR

[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 2026-09-27T16:47:21.540217Z	36d3776	PR opened
🔒 Security Review	✅ Completed 2026-09-27T16:49:06.742933Z	36d3776	PR opened

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.

[markstuart-oai]

Reviewed at 36d3776ff54c0fde32ad77059170f9f08074f51a. Stopping iteration after the failure event keeps transport teardown from replacing the saved provider error, while the existing finally path still closes the stream. Recognizing error as terminal also covers the wrapper cleanup path. The regressions distinguish context cleanup from HTTP/SSE teardown, preserve provider diagnostics and event delivery, and retain the pre-terminal close-failure control. No blocking findings.

Verified all 22 hosted checks passed on this commit. Source review only here; I did not execute the reported local suites or live-provider tests.