Yes. The handler currently completes the active attempt, regardless of the event's identifier.

Then reproduce that exact error first. Also record what should happen to the old attempt; leaving the new one alone is only half the expected result.

Does the real completion contain the same attempt identity that your test will supply?

Priya's question protects the test from inventing a better interface than the real one. If the identity is absent or ambiguous, the expected behaviour should preserve that uncertainty rather than let the test helper assign it using knowledge the application cannot have.

The captured events do contain attempt identifiers. The application discards them before choosing the active state.

Jack, has the deterministic sequence reproduced the original failure now? Interested in that before the duplicate and restart cases make the test harder to read.

It reproduces every time. Corrected handler matches by attempt identity; old completion no longer finishes the new one.

Now deliver a matched completion twice and check the report as well as the attempt state. Also deliver the old result after the newer attempt legitimately finishes. A state machine can look correct while a summary counter still treats every delivered event as another completed coupon.

Duplicate replay leaves the attempt unchanged but increments the report twice. Separate reporting defect confirmed. Restart persistence still untested.

Worth finding. Keep the two reproduced failures separate so the reporting fix cannot hide a tracker regression. You have a useful test sequence, not a completed reconciliation fix yet.