We can hold the earlier completion in a test queue and deliver it after creating the new attempt. That reproduces the failure without any sleep.

Does the handler ignore the carried identity or overwrite it?

Ignores it. It finds the current attempt by the friendly job name, which both attempts share.

Then keep those equal job names in the test. Giving everything beautifully unique labels could let a wrong lookup pass while the real application keeps reusing names.

Can the support view open the earlier attempt as well?

And distinguish a duplicate result from a different result for that earlier attempt.

Louis, yes. Each attempt has its own detail row. Caleb, the result identity is separate too; we will repeat the same result and supply a distinct conflicting one as separate cases.

Good. A conflict needs review, not silent deduplication.

Add reopening between deliveries. The original result may be saved while the in-memory list of processed events has vanished. That's a different route back to the same inflated display.

Offline repair now looks up the carried attempt identity. The original late result updates the earlier attempt only; its duplicate adds no row. A distinct conflicting result stays visible for review.

What does a result with an unknown attempt do?

Stays unmatched, with its identity and reason available in the support view. It doesn't attach itself to the current coupon.

Reopen-between-deliveries also passes. We then deliver the current attempt's own result and it completes normally. Maintainer has the deterministic sequences and expected rows; this offline handler fix is finished.

Thanks for retaining the conflicting-result case. That's the sort of inconvenient evidence a too-helpful deduplicator can quietly throw away.