Creating a CPU node (cpu-2 / cpu-8) and waking a stopped one with its files intact. Measured by a real cpu-2 node driven through the same handlers a customer's calls ride: create, write a marker file, stop, move the disk to durable storage, wake through the restore path, read the marker back. Its own bar because CPU nodes run on separate machines from GPU nodes - either fleet can fail while the other works. A sold-out fleet stays green: a full queue is healthy, not an incident.