fix(cancel): narrow STOP critical-window so deploy-park cancel applies (ORCH-090)

Review P1: a STOP while a self-hosting task is PARKED on `deploy` awaiting the
manual `Confirm Deploy` was classified as a critical merge/deploy window solely
because the task still held the per-repo merge-lease (held from merge-gate through
deploy->done). That window is fully reversible — nothing is merged or deployed yet
(the irreversible merge_pr runs later in _handle_merge_verify, always under an
INITIATED marker). So the cancel was DEFERRED to run_deploy_finalizer, which only
runs after Phase B (Confirm Deploy) — the very step the operator pressed STOP to
avoid. Result: the deferred cancel was never applied, the task wedged non-terminal
holding the lease, blocking the repo's serial-gate (ORCH-088) and merges.

Fix: gate the merge-lease branch of cancel.in_critical_window on an actively
RUNNING actor (_task_has_running_actor). Lease held + running deploy/merge job ->
still deferred (genuine in-flight step). Lease held + no running actor (idle
deploy parking) -> NOT critical -> immediate full reset, which itself releases the
lease (step 3c) and drives the task terminal. INITIATED-marker deferral unchanged.

Also fixes review P2 (AC-6): set_task_cancel_requested now returns the first-stamp
fact (rowcount), and the deferred branch only notifies on the first transition —
a repeated STOP while still deferred no longer spams duplicate notifications.

Tests: test_d7_lease_held_idle_parking_is_not_critical,
test_d7_lease_held_with_running_actor_still_critical,
test_d7_stop_on_deploy_awaiting_confirm_full_resets,
test_d7_repeated_stop_in_critical_window_no_duplicate_notify. Full suite green (1349).

Refs: ORCH-090

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

src/cancel.py

@@ -3,8 +3,8 @@
 Leaf module mirroring ``src/serial_gate.py`` / ``src/labels.py``: pure,
 unit-testable, never-raise functions over config + the existing DB / deploy-state.
 Module-level imports are limited to ``config`` (and ``re``); the critical-window
-probe lazily imports ``self_deploy`` / ``merge_gate`` so a cycle can never form and
-an import failure degrades safely.
+probe lazily imports ``self_deploy`` / ``merge_gate`` / ``db`` so a cycle can never
+form and an import failure degrades safely.
 
 What it answers:
   * ``applies(repo)``           — is STOP-cancellation REAL for this repo?
@@ -78,16 +78,50 @@ def applies(repo: str) -> bool:
         return False
 
 
+def _task_has_running_actor(task_id) -> bool:
+    """True iff the task currently has a RUNNING job — an active merge/deploy actor.
+
+    Distinguishes a genuinely in-flight merge/deploy (a running deployer / deploy
+    finalizer job actually executing the irreversible step) from a task merely
+    PARKED on ``deploy`` awaiting the human ``Confirm Deploy`` (the merge-lease is
+    held across that wait, ORCH-036/043, but nothing is executing and nothing has
+    been merged/deployed). Lazily imports ``db``; raises on a db error so the caller
+    fails CLOSED (treat as critical) rather than silently mis-classifying on doubt.
+    """
+    if not task_id:
+        return False
+    from . import db
+    for job in db.get_active_jobs_for_task(task_id):
+        if job.get("status") == "running":
+            return True
+    return False
+
+
 def in_critical_window(task: dict) -> bool:
     """Is the task inside an irreversible merge/deploy step (ADR-001 D7 / AC-7)?
 
     A STOP that lands here must NOT tear the step apart (half-merge / detached prod
-    deploy / dead prod container, NFR-3). Two markers (existing, no new state):
+    deploy / dead prod container, NFR-3). Markers (existing, no new state):
       * self-deploy Phase B initiated — the ``INITIATED`` sentinel in
-        ``<repos_dir>/.deploy-state-<repo>/<wi>/`` (ORCH-036);
-      * the task currently HOLDS the per-repo merge-lease
-        ``<repos_dir>/.merge-lease-<repo>.json`` (ORCH-043), holder branch == task
-        branch.
+        ``<repos_dir>/.deploy-state-<repo>/<wi>/`` (ORCH-036) — the detached prod
+        deploy + the deterministic ``merge_pr`` (``_handle_merge_verify``, run later
+        under the SAME marker) are both covered here;
+      * the task HOLDS the per-repo merge-lease ``<repos_dir>/.merge-lease-<repo>.json``
+        (ORCH-043), holder branch == task branch, **AND** a merge/deploy actor is
+        actually RUNNING.
+
+    The merge-lease branch is gated on a running actor on purpose (ORCH-090 review
+    P1 fix). For the self-hosting repo the lease is HELD from the merge-gate PASS
+    (``deploy-staging -> deploy`` edge) right through to ``deploy -> done`` — including
+    the whole time the task sits PARKED on ``deploy`` awaiting a human ``Confirm
+    Deploy`` (Phase A). That wait is FULLY REVERSIBLE: nothing is merged or deployed
+    (the irreversible ``merge_pr`` only runs later in ``_handle_merge_verify``, always
+    under an ``INITIATED`` marker already caught above). Classifying that idle parking
+    as "critical" used to DEFER the cancel to a deploy finalizer that the operator —
+    having pressed STOP precisely to NOT confirm — never triggers, so the cancel was
+    never applied and the task wedged while still holding the lease (blocking the
+    repo's serial-gate / merges). Now idle parking (lease held, no running actor) is
+    NOT critical: the full reset runs immediately and itself releases the lease.
 
     fail-CLOSED (TR-3): any error/uncertainty -> True (DEFER cancellation). Outside
     the window -> False (apply the full reset immediately).
@@ -108,7 +142,16 @@ def in_critical_window(task: dict) -> bool:
         from . import merge_gate
         holder = merge_gate.current_lease_holder(repo)
         if holder and branch and holder == branch:
-            return True
+            # Lease held. Critical ONLY if an actor is actively merging/deploying;
+            # an idle task parked on `deploy` awaiting Confirm Deploy is reversible.
+            if _task_has_running_actor(task.get("id")):
+                return True
+            logger.info(
+                "cancel.in_critical_window: task %s holds the merge-lease but no "
+                "actor is running (idle deploy parking, awaiting Confirm Deploy) -> "
+                "NOT critical; full reset will release the lease", task.get("id"),
+            )
+            return False
     except Exception as e:  # noqa: BLE001 - fail-CLOSED on doubt
         logger.warning("cancel.in_critical_window merge-lease probe error: %s", e)
         return True