The "can stop the container from child_exit" test was racy. All three workers called exit(1) immediately, so multiple child_exit callbacks could fire before container.stop took effect. The policy guards with unless container.stopping?, but that's a check-then-act race — several children can each observe stopping? == false and increment stop_count, making the stop_count == 1 assertion flaky.

Fix

Only the first worker exits, and only after all workers are ready (coordinated via a pipe). The others sleep. This guarantees exactly one child_exit drives the stop, so stop_count is deterministically 1.

Test-only, no library change. Verified stable across repeated runs.