Clean up orphaned VM data dirs from crashed VMs

[JAORMX]

Summary

Fixes #87 — orphaned COW rootfs clones from crashed VMs leak disk indefinitely.

When a VM crashes or bbox is killed (SIGKILL, OOM) before WithCleanDataDir() runs, the COW-cloned rootfs at ~/.config/broodbox/vms/<name>/data/rootfs-work/ survives — potentially hundreds of MB per orphan, accumulating unbounded across crashed runs.

CleanupStaleLogs already removes a whole VM directory when its bbox sentinel PID is dead, but no sentinel is written for runs that pass --log-file, so those data clones leak.

Approach

Add CleanupStaleVMData, keyed off go-microvm's per-VM state file (state.Manager.Load(), which reads go-microvm-state.json without locking) rather than the bbox sentinel:

• Scan ~/.config/broodbox/vms/*/data/.
• A data dir whose state is active: true but whose recorded runner PID is dead was orphaned → remove it.
• A live runner PID is left untouched, so it's safe to run at startup alongside concurrent VMs (each VM has its own uniquely named dir). A recycled PID now hosting an unrelated live process is conservatively treated as alive and skipped; the orphan is reclaimed on a later run.
• After removing the data dir, best-effort removes the now-empty parent (reclaiming sentinel-less --log-file runs) via a non-recursive Remove, which only succeeds on an empty dir — so it never deletes lingering logs or a sentinel that CleanupStaleLogs still owns.

Wired into the composition root next to CleanupStaleLogs / CleanupStaleSnapshots.

Note: the issue referenced state.json; the actual go-microvm state file is go-microvm-state.json. Reusing state.Manager.Load() means we don't hardcode the filename or schema.

Tests

New table of cases in cleanup_test.go (state fixtures written via the canonical state package):

• orphaned (active + dead PID) → data dir and empty parent removed
• live runner (active + our own PID) → preserved
• inactive state (clean shutdown) → preserved
• orphan with parent logs → only data dir removed, parent left for log cleanup
• no state file → preserved
• nonexistent vms dir → no panic

Verified: task lint (0 issues), full task test passes (incl. -race).

🤖 Generated with Claude Code