Add health-based smart stopping

[AnandChowdhary]

Watch CI signal quality (e.g., N consecutive failed PRs, or repeated identical lint errors). Introduce a --stall-threshold that pauses the loop, summarizes the failure causes in SHARED_TASK_NOTES.md, and waits for human intervention. Complement existing completion signals with positive heuristics: use Claude’s own summary (“All scoped tasks complete”) plus repo diff size/CI cleanliness to auto-stop.