WAL does stop corruption, but codex's lock handling is the shaky part, since it sets busy_timeout to 5000 and doesn't retry SQLITE_BUSY. issues 20213 and 35555 seem to show that under concurrent use, and 35555 was still open as of august. the journal_mode stall in 35704 only involves one process, so I'd treat it as a side note

unquoting the pattern inside excluded should make state_*.sqlite match and survive schema bumps, so test that on a scratch copy first. keep each DB with its wal and shm files in the same folder. and check with sqlite3 whether unified history is in state_5.sqlite, because if it is, excluding the DBs loses most of the benefit