fix streamed response corruption in generation logs

[laceyp99]

Summary

• keep streamed delta accumulation authoritative for final_response so UI decorations cannot corrupt generation logs
• render provider errors, interrupted-generation errors, and reasoning warnings as distinct completed status messages with metadata
• cover partial-response failures, warnings, and concurrent background stream failures with regression tests

Closes #24

Testing

• py -3.13 -m ruff format --check . - passed
• py -3.13 -m ruff check . - passed
• py -3.13 -m pytest -q - passed (113 tests)

[github-actions]

Agent PR review

Validation

Validation passed: python -m ruff format --check ., python -m ruff check ., and python -m pytest all completed successfully.

Summary

The PR changes streamed generation logging so final_response is accumulated only from streamed deltas, while warnings/errors are rendered as separate completed status messages. It also updates serialization and adds regression coverage for partial-response failures, unsupported reasoning warnings, and background stream failures.

Issues and risks

No blocking issues found.

I reviewed README.md, confirmed no AGENTS.md exists, read codex-review-context.md, and compared origin/pr/39 against origin/main. The changed paths in arena/app.py and arena/ui/display.py appear consistent with the stated intent: UI status messages no longer overwrite or contaminate the logged model response.

Test coverage

The PR adds focused coverage for:

• preserving partial streamed output after an error
• storing warning/error messages as metadata-bearing completed messages
• preserving accumulated final_response
• handling concurrent background stream failures

The review context reports these checks passed: python -m ruff format --check ., python -m ruff check ., and python -m pytest. I did not rerun them locally because the workspace is read-only.

Remaining non-blocking gap: there does not appear to be a browser-level Gradio smoke test proving the new gr.ChatMessage metadata renders exactly as intended after serialization/reload.

Recommendation

Looks safe to continue review.

---

This is an automated review. Please treat it as advisory; human review is still required.