- RunOnce purged failed deliveries (a full scan) on every batch of 50: the Worker now
does it at most once every 5 minutes, by the store clock; a failed purge is retried
on the next run.
- Every recipient re-read the whole message from SQLite: a per-run map msgID -> body,
filled once per message even under concurrency, is dropped when the run ends.
- Tests count purges across batches and after the interval, and body reads per message
(including a missing body) through unexported hooks.