The offline test has two inspection attempts and one late completion. Deliver the old result after the new attempt starts and the new one turns complete. Random sleeps make this fail only when they feel cooperative. I want the event order under direct control.
Use an explicit delivery queue, then assert both stored attempts after the old completion arrives; preserve the original attempt identity instead of looking up whichever attempt is current.
Test its genuine late completion against the old attempt, plus duplicate delivery and the new attempt's own result. The fix must preserve legitimate older evidence as well as protect the new state.
Handler corrected in the offline build. Old result completes its own attempt, duplicate adds nothing, and the new attempt waits until its result arrives. Thanks, Farah, checking both rows caught more than watching the current tile.