fix: persist retry state, add API retry logic, raise on GraphQL errors

[chanakyav]

Summary

Addresses all three items in issue #74 — retry state persistence, GitHub API retry logic, and silent GraphQL error handling.

Changes

1. Persist _retry_counts for crash recovery (schema v8)

• persistence.py: Added retry_counts_json TEXT column. Schema v7 to v8, migration entry, _TASK_COLUMNS updated.
• orchestrator.py: __init__ loads _retry_counts from task["retry_counts_json"] (JSON deserialized). After incrementing a retry counter in _do_verify_push() or _do_verify_pr(), persists via update_task(task_id, retry_counts_json=json.dumps(...)).

Before: If the orchestrator crashed and resumed during VERIFY_PUSH, _retry_counts reset to {}, allowing infinite retry loops.

2. Exponential backoff retry in _run_gh()

• github_api.py: _run_gh() now retries up to 3 times with exponential backoff (1s, 2s, 4s) for transient errors.
• Transient errors detected by matching stderr against patterns: rate limit, abuse detection, server error, 502, 503, 504, timed out, connection refused, connection reset, network is unreachable.
• Non-transient errors (404, auth, etc.) fail immediately with no retry.
• check=False callers are unaffected — they still get empty stdout on failure.

3. Raise GitHubAPIError on GraphQL errors

• github_api.py: get_unresolved_review_comments() and get_latest_copilot_review_thread_ts() now raise GitHubAPIError when GraphQL returns errors with no usable data.
• Partial errors (errors + data present) are logged as warnings but data is still returned.
• The orchestrator's existing try/except around these calls will catch and handle the new exceptions.

Tests Added (10 new)

• test_retry_counts_json_persists + migration test updated (test_persistence.py)
• test_retries_on_rate_limit (test_github_api.py)
• test_no_retry_on_permanent_error (test_github_api.py)
• test_retries_on_server_error (test_github_api.py)
• test_gives_up_after_max_retries (test_github_api.py)
• test_check_false_no_raise (test_github_api.py)
• test_unresolved_comments_raises_on_graphql_error_no_data (test_github_api.py)
• test_unresolved_comments_partial_error_returns_data (test_github_api.py)
• test_latest_thread_ts_raises_on_graphql_error_no_data (test_github_api.py)
• test_latest_thread_ts_partial_error_returns_data (test_github_api.py)

Verification

• ruff check src/ tests/ — zero errors
• pytest tests/ — 293 tests passed, zero failures, no regressions

Closes #74

---

🤖 autopilot-loop