fix(dispatch): stop burning breaker strikes on dependency-held blocked tasks; collision notify rides the delegate transaction (#726)

* fix(dispatch): stop burning breaker strikes on dependency-held blocked tasks; collision notify rides the delegate transaction

Two follow-ups from the 2026-07-29 blocked-task wave:

- _dispatch_blocker_work now skips a blocked task whose dependencies are
  non-terminal BEFORE the respawn gate. spawn_agent's readiness gate was
  refusing these anyway, but each refused attempt burned a respawn-
  breaker strike and, once tripped at 4, a duplicate CEO escalation
  every tick (seen live on the eslint-audit task held behind a paused
  backend dependency). The dispatcher re-checks each tick and proceeds
  the moment the dependency goes terminal — same resume path as before,
  minus the noise.

- send_collision_sequencing_notification gains the db_session threading
  its sibling senders already have, and TaskService passes its own
  session. delegate wires collision edges inside a transaction whose
  held-back child row is not yet committed; the notification INSERT ran
  on a separate auto-commit connection and died on the related_task_id
  FK (ForeignKeyViolationError seen live in cell_pm/delegate), silently
  losing the coordination alert. Riding the caller's transaction makes
  the row visible and commits both atomically.

* test(notification): cast the fake session for the gate's tests-scope mypy

---------

Co-authored-by: Renn F <rennf93@users.noreply.github.com>

tests/unit/services/test_notification.py

@@ -544,6 +544,25 @@ async def test_send_collision_sequencing_notification(
     assert any("sequenced behind" in r.subject for r in rows)
 
 
+@pytest.mark.asyncio
+async def test_send_collision_sequencing_notification_rides_caller_session(
+    svc: NotificationService,
+) -> None:
+    """2026-07-29: delegate wires collision edges inside a transaction whose
+    held-back task row is not yet committed — the notification INSERT must
+    ride that same session, or its related_task_id FK fails on any other
+    connection. No _patch_db_context here: opening a separate connection
+    would hit the real engine and fail the test."""
+    db = _FakeDb(agent_uuid=uuid4())
+    await svc.send_collision_sequencing_notification(
+        held_back_task_id="t2",
+        blocking_task_id="t1",
+        held_back_assignee="be-dev-1",
+        db_session=cast("Any", db),
+    )
+    assert [r for r in db.added if r.related_task_id == "t2"]
+
+
 @pytest.mark.asyncio
 async def test_send_unblock_notification(svc: NotificationService) -> None:
     aid = uuid4()