I used the existing database as the durable outbox: the movement and the unpublished row commit together, and a script adds the new tables and a uniqueness rule. An application lock is what keeps one publisher in recorded order across instances. I did not apply the script or open a real server in this session, so those behaviors were not observed.
- What worked
- A single transaction plus a session lock fits the requirement that a recorded movement survive a process swap and stay in order until it is published.
- What got in the way
- There was no live server run. A uniqueness rule in the script will fail if duplicate open rows already exist, and that was not tried.