opencode-qa serve-wake-split-probe --expect fixed verdict depends on runner speed (wake count flips on unmodified code) #8958

Summary

.agents/skills/opencode-qa/scripts/serve-wake-split-probe.sh --expect fixed returns INCONCLUSIVE (exit 1) on unmodified code whenever the parent session takes more than ~5 s to go idle after the parent wake's noReply admission.

Evidence

Six runs on a busy machine (base and a branch that does not touch this path):

In the INCONCLUSIVE runs the omo log shows Recorded admit-only parent wake ..., then Requeued dispatched parent wake after no assistant output (retryCount 1) about 5 s later, then a second promptAsync dispatching that the fake LLM counts as branch=wake. In the FIXED runs the harness reached FINAL_COUNTS about 2 s after the admission, before that window.

Expected

The --expect fixed verdict is deterministic. It should either wait for the retained reply-required wake to settle (consumed or re-dispatched) before counting, or treat a liveness re-dispatch after a completed parent turn as the designed outcome rather than a split. As written, the verdict depends on how fast the runner is.

Acceptance criteria

• Ten consecutive --expect fixed runs on the same code give the same verdict, including under CPU load.

Found while validating #8956.