Harden stuck-job retry accounting and retry config validation

tests/test_transitions.py

@@ -169,3 +169,49 @@ def test_requeue_allows_one_retry_then_fails_on_second_timeout() -> None:
         assert second[0].failed is True
         failed = session.execute(select(ReviewJob).where(ReviewJob.id == job.id)).scalar_one()
         assert failed.status == JobStatus.failed
+
+
+def test_recovery_persists_timeout_retry_even_if_latest_run_not_running() -> None:
+    session_factory = get_session_factory()
+    with session_factory() as session:
+        job = enqueue_job(
+            session,
+            repo="acme/repo",
+            pr_number=5,
+            head_sha="33334444",
+            trigger_comment_id=1005,
+            trigger_comment_body="@codex review",
+            requested_by="alice",
+            command=ParsedCommand(name="review", raw="@codex review"),
+        )
+        claimed = claim_next_job(session)
+        assert claimed is not None
+        assert claimed.id == job.id
+
+    with session_factory() as session:
+        latest_run = session.execute(select(ReviewRun).where(ReviewRun.job_id == job.id).order_by(ReviewRun.id.desc()).limit(1)).scalar_one()
+        latest_run.status = RunStatus.succeeded
+        latest_run.finished_at = datetime.now(timezone.utc)
+        stale = session.execute(select(ReviewJob).where(ReviewJob.id == job.id)).scalar_one()
+        stale.started_at = datetime.now(timezone.utc) - timedelta(seconds=601)
+        session.commit()
+
+    with session_factory() as session:
+        first = recover_stuck_jobs(session, lease_timeout_seconds=300, action="requeue", max_retries=1)
+        assert len(first) == 1
+        assert first[0].failed is False
+        claimed_again = claim_next_job(session)
+        assert claimed_again is not None
+        assert claimed_again.id == job.id
+
+    with session_factory() as session:
+        stale_again = session.execute(select(ReviewJob).where(ReviewJob.id == job.id)).scalar_one()
+        stale_again.started_at = datetime.now(timezone.utc) - timedelta(seconds=601)
+        session.commit()
+
+    with session_factory() as session:
+        second = recover_stuck_jobs(session, lease_timeout_seconds=300, action="requeue", max_retries=1)
+        assert len(second) == 1
+        assert second[0].failed is True
+        final = session.execute(select(ReviewJob).where(ReviewJob.id == job.id)).scalar_one()
+        assert final.status == JobStatus.failed