Why does yesterday's test result finish the attempt I just started?

At the moment a background thread sleeps and sends it, so the machine running the test gets a vote in the order

Replace that timing guess with an event queue the test advances explicitly.

Only the job label in my model, which is exactly what gets reused; the software owner is reviewing how the real interface can carry attempt identity

Keep the reused label in the test. It is part of the failure case.

Can you reproduce it now without any sleeps?

Yes, Liam, queued delivery fails the old handler every time I've run it, and Daniel's unfinished-state assertion catches the wrong completion directly

Give the test separate identities for separate attempts. Deliver duplicates as well as the delayed event, and check that a completed attempt cannot be credited twice. Keep the real interface work separate from this model change.

Added unique attempt identities in the model and the delayed case now passes, but a duplicate still adds another entry to the completion list

There's your next assertion. One accepted completion entry per attempt. I'd also want the ignored duplicate visible somewhere for diagnosis, because otherwise support will wonder why the raw message count differs.

The model now retains the duplicate as an event without crediting another completion; or rather, that's what the new handler does, and the old handler fails both checks

What happens if the application restarts before the duplicate arrives?

And does the test restore saved state or start empty?

Restore the state the application design actually persists. Include a crash after receipt but before saving the completion. Recovery has to handle that uncertainty rather than assuming the event never occurred.

Liam, it started empty, which was no help; I've added saved-state restoration and a crash point before the save, Elliot, with the unresolved case left for reconciliation

Also deliver the new attempt's valid completion before the old one. The old event must not overwrite the newer result or reopen a completed attempt.

Who will read the unresolved case screen? Your software owner should try the wording with them. 'Unknown' without the attempt or reason sends people straight back to asking you.

Quality's shift lead will review that screen, Daniel; model tests now cover both completion orders, duplicates and the save gap, but the live interface proposal is still with its owners