"Initiate canary," she said, though no one else was in the room to hear it.

She watched logs stitch back into pattern: no more HOT flags, no more orphaned PIDs. And then a line she had been waiting for: ALL CLEAR.

Outside, sunlight moved over the edge of the server room window. The drones, freed from their paused limbo, traced clean arcs against the sky. In the logs, the word HOT no longer appeared, but the memory of it stayed with Mira—the kind of small, heated failure that teaches the system how to be cooler next time.

Mira scrolled to the top of the config, then to the comment line. She changed it—not the contents of the config, but the process: she added a small, defensive watchdog to Locksmith's startup sequence that checked for stale locks on boot and scheduled more aggressive garbage collection. She pushed the change and wrote a terse commit message: fix: reclaim stale locks on boot; reduce GC interval.

Mira pulled up the config file. Its contents were tidy: settings for aim sensitivity, safety thresholds, and a single comment line scrawled in a careless hand: # last touched by node-7 @ 03:12. Node-7 was offline. The system insisted the lock was active, though no process owned it.

She deployed to the three drones. Telemetry flooded in: stable heart rates, smooth trajectory corrections, and then, bleakly, one drone reported "lock mismatch: aim_lock_config.conf HOT". The canary refused the shadow config—the lock check happened locally before accepting any override.

"Design for ghosts," Mira said. "State loves to linger. Make it easy to be explicit about ownership, and always have a safe bypass."

ERROR: aim_lock_config.conf: HOT

It was an absurd word to see in a machine log, yet the machines felt it. Drones paused mid-patrol, loading arms stalled in the factory, and the research cluster throttled itself into an awkward limbo. "Hot" meant a file the lock manager refused to open—an in-memory semaphore indicating someone else had it. Only problem: nothing else should have been holding it. The lock should have released when the orchestrator completed its update cycle thirty minutes prior.

She ran the kernel toggle: echo 0 > /sys/locks/aim_lock_config/conf_locked. The system replied with a terse OK. The lock bit cleared. For a moment nothing else happened, as if the cluster checked its pulse. Then Locksmith's watchdog thread reanimated, reacquiring the file in a clean state. Node-7's ghost in the machine vanished.

"Stale lock," she whispered. The phrase clanged differently in production: stale locks meant machines held against change, and when machines refuse change, humans lose control.

Back to the kernel. Mira dumped the lock table, inspected kernel logs, saw a kernel panic thread that had restarted the lock manager with an incomplete cleanup. The restart sequence left the lock bit set but with no owner. The fix was delicate: unset the kernel lock bit manually, but only after ensuring no process would try to regrab it mid-op. That meant stopping the aim orchestrator—a bolder move.

She paged the on-call network: "Going to stop-orchestrator for 90s to clear stale lock." Silence. Then a terse reply: "Acknowledge. Hold point." It arrived with the authority to proceed.

"Lesson?" the junior asked.

She traced the lock's metadata to a zippy little microservice nicknamed Locksmith—a lightweight guardian intended to prevent concurrent configuration writes. Locksmith's metrics showed a heartbeat frozen at 03:12. Its PID was gone, but the kernel still held the inode as taken. That was impossible; file locks shouldn't survive process death.