A failed `submit_listen` now writes the listen to the SHARED
`ListenBrainzCache.listen_queue` (daemon-root `cache/listenbrainz_v2.db`, the
same schema the desktop uses) instead of being lost. The scrobble task's loop
becomes a biased `select!` between the CoreEvent bus and a 120s drain interval
(the interval also fires once at task start), so a queue built up while offline
is retried oldest-first and cleared on reconnect, stopping at the first failure
and resuming on the next tick. Mirrors `qbz::scrobble::flush_listenbrainz_queue`:
the rusqlite Connection is opened inside a `spawn_blocking` per op and never held
across an await. The daemon uses its own `qbzd` cache root — it never touches the
desktop's dirs.
cc
vicrodh
903acdb044