When a node transitions to RS_ROLLBACK, outstanding write-concern waiters in _replicationWaiterList (and the other three waiter lists) are never resolved – setErrorAll is only called during shutdown. If a thread wakes from markKilled with its future still unresolved, it attempts to acquire _mutex to remove itself from the waiter list. If _mutex is held by the rollback path at that moment, the thread deadlocks, retains its IntentRegistry token, and the rollback drain loops forever.