0003 — The row lock is held for the entire processing duration

Context. With claim, processing, and delete inside one transaction, the FOR UPDATE lock — and the DB connection carrying the transaction — stays held for as long as the handler runs.

Decision. Accept that a slow handler holds its row lock and a connection for the whole processing time. At small scale this is fine, and it is what makes crash recovery free: rollback releases the lock.

Consequences. A long-running job pins a transaction and a connection for its full duration, which does not scale under real concurrency. Superseded by moving the claim out of the DB lock and into row data (status='processing' plus locked_at), shrinking the lock to the millisecond claim statement.