The agent watched desired-state with bucket.watch() (DeliverPolicy::New),
which delivers only future Puts — never the current value of existing
keys. On any agent restart, including a device reboot, the in-memory
reconciler cache started empty and the 30s ground-truth tick iterated
nothing, so containers stopped by the reboot were never restarted (only
an operator KV rewrite would bring them back). Switch to
watch_with_history() (DeliverPolicy::LastPerSubject): the agent replays
the current desired-state for its device on startup, re-learns its
deployments, and the idempotent apply path restarts the stopped
containers. The user podman.socket is already enabled + lingered, so it
survives reboot; the agent is the intended ground-truth reconciler
(ROADMAP 5.6).