Budgeted batch flush under memory pressure must wait for the queue to drain

In a producer/worker/committer pipeline with a strict memory budget, a producer that holds an open batch of small jobs should, when a reservation fails, queue the batch and then wait until the queue drains before reserving again. If it reserves straight away, it takes the memory the batch's worker just freed, and the committer's downstream writer calls (whose growth is accounted when it happens, not reserved at admission) fail their own strict check. This failure depends on timing (about 1 in 7 runs here); waiting for the drain removes it. Test it deterministically: make every job small and send no barriers, so nothing is queued before the pressure point, then set the budget between the unbatched run's peak and the batched run's pre-flush accounting. Run it 5 times so a racy variant is caught.