fix(opencode): retry nested OpenAI errors

[MagMueller]

Issue for this PR

No linked issue.

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

OpenAI Responses streams can report a transient failure as a nested response.failed event with the error at response.error. The existing parser only recognized top-level error events, so nested server_error and server_is_overloaded events became terminal unknown errors.

This keeps the parser change narrow: those two nested transient codes become retryable APIError values and flow through the existing SessionRetry policy. Other nested response.failed codes retain their current terminal behavior.

Follow-up risk intentionally left out of scope: SessionRetry caps each backoff wait, but generic retryable API errors do not currently have a total-attempt cap.

How did you verify your code works?

• bun test test/session/retry.test.ts --timeout 30000: 37 passed
• bun typecheck from packages/opencode: passed
• pre-push filtered workspace typecheck: 16/16 packages passed

Added regression coverage for both nested transient codes and for a nested non-retryable code remaining terminal.

Screenshots / recordings

Not applicable; no UI change.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

---

Summary by cubic

Makes nested OpenAI stream server errors retryable so transient failures don’t abort sessions. Detects server_error and server_is_overloaded inside response.failed events and routes them through the existing retry policy.

• Bug Fixes
  • Updated opencode parser to handle nested response.failed errors: server_error and server_is_overloaded now map to retryable API errors; other nested codes stay terminal.
  • Added regression tests for both retryable and non-retryable nested cases.

^{Written for commit 05f9985. Summary will update on new commits.}

[cubic-dev-ai]

No issues found across 2 files

_{Re-trigger cubic}