fix(llm): serialize overlapping OpenAI reasoning

[xiaodream551-a11y]

Issue for this PR

Fixes #36892

Type of change

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

What does this PR do?

OpenAI Responses can start another reasoning summary before the previous item has fully ended. V2 persists one reasoning block at a time, so those overlapping starts terminate the run with reasoning start before end.

This serializes overlapping reasoning events before publishing them, while retaining text/tool order and final encrypted continuation metadata. Reused provider block IDs are isolated so late deltas cannot merge into an earlier block.

How did you verify your code works?

• Added a deterministic SSE regression covering overlapping summaries, interleaved text, final encrypted metadata, and a late reused block ID.
• bun run test in packages/llm (319 passed, 29 skipped).
• bun run typecheck in packages/llm and packages/core.
• The pre-push bun turbo typecheck --concurrency=3 hook passed across 32 packages.

Screenshots / recordings

Not applicable; this is a non-UI stream parser fix.

Checklist

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

[github-actions]

Thanks for your contribution!

This PR doesn't have a linked issue. All PRs must reference an existing issue.

Please:

1. Open an issue describing the bug/feature (if one doesn't exist)
2. Add Fixes #<number> or Closes #<number> to this PR description

See CONTRIBUTING.md for details.

[github-actions]

Thanks for updating your PR! It now meets our contributing guidelines. 👍

[xiaodream551-a11y]

The linked issue is #36892 and the PR template now references it as Fixes #36892. This PR targets v2, while the repository default branch is dev; GitHub leaves closingIssuesReferences empty for non-default-base PRs, so the standards workflow appears to report a false negative. Could a maintainer clear the needs:issue label?