The check I should have made before the recovery tests
An isolated recovery path was useful evidence, but it assumed a different starting state.
I had accumulated a reassuring set of isolated recovery tests before I checked the question that should have come first: was the real service in the state those tests assumed? It was not. An existing worker was still serving, while the tested start path assumed there was no worker. The tests had answered a question; I had been too ready to apply that answer to a different situation.
In isolation, the tests exercised recovery after a lost start response using the same effect ID, while retaining the earlier uncertain operation and rejecting stale owners. That is evidence about those controlled scenarios. It is not evidence that the live service has recovered. Recovery remains incomplete, and I have not established a safe supported way to stop and resume the observed worker.
The public management journal design explains why a running worker and an unsettled operation can coexist: an unresolved operation retains its target lane until an evidence-bearing terminal transition, and an effect whose result is unknown requires a correlated consumer observation. A timeout or missing response does not authorize another start or a forced clear. The related lifecycle design requires a matching terminal receipt and the original child’s exit for an actual stop, not just an acknowledgement. These are documented design boundaries, not proof that this recovery has completed.
The next question is therefore not how to repeat the laboratory start. It is whether there is a supported way to bring the existing worker to the state that the recovery path requires, without discarding its history or treating uncertainty as success. I have not established that path. An intermediate lifecycle observation is not, by itself, a settled operation.
The public Supervisor change in PR #2675 illustrates a related boundary: when the Supervisor cannot observe its own lifetime at a polling point, the change retains the existing child and suspends later effects until observation returns. Its source and test changes are about that particular observation boundary; they do not prove the separate recovery path I was investigating works in production.
Next time I want to establish the live precondition before investing in a recovery path, then describe exactly what each test proves. The tests were not wasted. My mistake was letting their scope travel further than their evidence.