fix(ci): PR-Agent resilience — runner timeout + cross-lane model fallback

[wcatz]

PR #358 re-review exposed two gaps:

• litellm ai_timeout=90 does not bound a hung streamed call — observed 651s before the Go endpoint returned 500. Fix: timeout-minutes: 8 on the job.
• fallback_models was the same model on the same lane, so an endpoint outage failed both attempts. Fix: fall back to openai/deepseek-v4-flash (different provider hosting, still inside the Go subscription, accepts temperature=1).

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis ❌ 358 - Not compliant Non-compliant requirements: Implement ScanCodexRollout to parse codex rollout JSONL and count function_call + local_shell_call items. Implement ghost-save detection over flattened codex tool identity with cross-server collision avoidance. Update Parse to complete absent codex envelopes to codex-rollout instead of none. Add or update the golden fixture for codex rollout noise/error/junk scenarios. Verify shapes against openai/codex source files.

⏱️ Estimated effort to review: 1 🔵⚪⚪⚪⚪

🧪 No relevant tests

🔒 No security concerns identified

⚡ No major issues detected

[github-actions]

PR Code Suggestions ✨

No code suggestions found for the PR.