Checkpoint: data access

Checkpoint: data access

One store, one state machine, one budget. The store enforces optimistic concurrency (version check or conflict); the retry loop re-reads the current version on conflict and converges in rounds; the pool drill keeps the occupancy arithmetic honest.

If your first SaveAll threw forever, the usual cause is reusing the stale expected version — the fix is re-read on conflict, exactly as in production retry loops. If VersionOf lied after a conflict, check that a failed Put mutates nothing.