Block stale and unstable merge states

[coygeek]

Summary

• classify GitHub BEHIND merge states as blocked with an out-of-date head ref reason
• classify GitHub UNSTABLE merge states as blocked with a non-passing commit status reason
• keep CLEAN and HAS_HOOKS ready-state coverage in separate regression tests

Closes #18

Validation

• PYTHONPATH=src python3 -m pytest tests/test_cli.py -q
• PYTHONPATH=src python3 -m pytest -q
• ruff check src tests
• autoreview --mode local: clean, no accepted/actionable findings

[mberman84]

Requesting changes because the blanket UNSTABLE classification regresses the exact-main integration path. Merging this head onto current main fails test_integration_snapshot_uses_exact_commit_check_fallback (blocked instead of ready). Preserve the useful BEHIND handling, but derive failed vs waiting from configured required-check states and do not override successful exact-commit evidence. GitHub CI is still action_required; re-request review after updating the head.

[mberman84]

UNSTABLE means a non-passing commit status, which can be pending or an optional failure. Treating it unconditionally as blocked turns ordinary CI into a repair escalation and also blocks current main's integration fallback after exact required CI has passed. Use the reduced required-check states to distinguish failed from waiting, and allow exact successful evidence to remain ready.

[coygeek]

Addressed the review feedback in the updated head. BEHIND remains blocked as an out-of-date head ref, while UNSTABLE no longer blocks unconditionally: it waits only when DeployBot cannot prove the configured required checks have passed, preserving the exact-check integration fallback.

Verification: PYTHONPATH=src uv run --with pytest --with ruff python -m pytest tests/test_cli.py -q and uv run --with ruff ruff check src tests both pass.

[mberman84]

Re-reviewed on the latest main. BEHIND fails closed; UNSTABLE waits only without exact required-check proof and preserves exact-success integration behavior. Full suite: 227 passed, 3 skipped, 202 subtests; Ruff and structured autoreview clean.