v0.4.26

fix: keep restoring merges until their sources appear

CheckMergeMount previously ran once at boot, so a merge whose source
disks were not mounted yet (or whose mount point was occupied) was
never restored for the rest of the session. Reconcile every 30 seconds;
the pass is a silent no-op while all merges are mounted.

Fail loud on restore: a non-empty merge mount point now reports the
offending entries, and the merge_init_status endpoint returns the
"error" state with the last restore failure instead of the stale
"initialized" state.

Serialize merge lifecycle operations with a service mutex so the
periodic reconciler cannot race UI create/update/remove calls, and
re-check the database before restoring a merge that may have been
removed while the pass was in flight.