Group failure handling + crash recovery

[hsyntax]

Parent PRD

#76

What to build

Ensure groups handle failures gracefully and can recover from crashes mid-group.

Failure handling (orchestrator):

• If the implement phase fails for a sub-issue (timeout, agent error, etc.), retry using existing timeout/retry logic (agent_timeout_retries).
• Retries apply to the individual sub-issue, NOT the entire group. Other completed sub-issues are preserved.
• If all retries are exhausted, leave the group in its current state for manual intervention. The group state persists with completed_sub_issues intact, and the failed sub-issue remains the current_task.
• On the next rlph iteration: the in-progress group is detected and the failed sub-issue is retried again (or the user can manually intervene by removing the state).

Crash recovery (orchestrator + state):

• On startup, if StateManager has a current_group with incomplete sub-issues:
  • Load the group state (parent ID, completed sub-issues, worktree path, branch).
  • Verify the worktree still exists. If not, recreate it by checking out the existing branch.
  • Determine the next eligible sub-issue (first uncompleted one whose dependencies are satisfied).
  • Resume from that sub-issue.
• The worktree and branch are the source of truth for completed work. Group state tracks which sub-issues are done.

Edge cases:

• Crash after agent commits but before state is updated: on recovery, detect the commit in the worktree (by checking git log for the sub-issue reference) and update state accordingly.
• Crash after PR submission but before review: resume directly into the review pipeline.

Acceptance criteria

• Failed sub-issue is retried (existing retry logic) without affecting completed sub-issues
• After retry exhaustion, group state persists with completed sub-issues intact
• On next iteration, the in-progress group is detected and the failed sub-issue is retried
• Crash recovery: group resumes from the last completed sub-issue
• Crash recovery: worktree is verified/recreated if missing
• Crash recovery: commit detection reconciles state with actual git history
• Crash after PR submission resumes into review pipeline
• Integration test: sub-issue fails → retry succeeds → group continues
• Integration test: simulate crash mid-group → restart → group resumes correctly
• cargo clippy passes with zero warnings

Blocked by

• Blocked by End-to-end group orchestration (tracer bullet) #78

User stories addressed

• User story 19: Group state survives crashes and allows resumption
• User story 20: Failed sub-issues retried before manual intervention