One occurrence, PR #348 shard-3 (https://github.com/conciv-dev/conciv/actions/runs/31411476740): T9 'two sends released into the same tick never overlap as two live runs' failed with the two turn outputs in reverse order ('same tick second' before 'same tick first'). Serialization held (no overlap); only arrival order flipped. 3/3 green in local isolation on the same commit. The test releases two sends in one tick and asserts completion order matches release order — under a loaded runner the scheduler can dequeue them reversed if the release path has no explicit ordering guarantee. Fix owed: either assert order-insensitively (set equality + no-overlap invariant, which is what T9 actually protects) or make the release deterministic in the test harness.
One occurrence, PR #348 shard-3 (https://github.com/conciv-dev/conciv/actions/runs/31411476740): T9 'two sends released into the same tick never overlap as two live runs' failed with the two turn outputs in reverse order ('same tick second' before 'same tick first'). Serialization held (no overlap); only arrival order flipped. 3/3 green in local isolation on the same commit. The test releases two sends in one tick and asserts completion order matches release order — under a loaded runner the scheduler can dequeue them reversed if the release path has no explicit ordering guarantee. Fix owed: either assert order-insensitively (set equality + no-overlap invariant, which is what T9 actually protects) or make the release deterministic in the test harness.