The parent process mutates shared batcher state (_span_buffer, _running_size, _active.flag, _flush_event, _running = False) before forking, but only releases the lock afterward — it never restores those fields. The at-fork hook only runs after_in_child, so the parent retains the mutated state. If another test later reuses the same client/batcher (or if the batcher's flusher thread observes _running = False), behavior may diverge from expectations and cause flaky failures.