When max_batch_size is N (N > 0), and a received batch is > N, then we only read N messages from the table to which the batch was added. If there's no other incoming thread, it may take a few ms until the remaining messages are removed and delivered.
This is not usually an issue because the 'last seqno resender' should reduce the risk of this happening.

Solution: in NAKACK2/UNICAST3.removeAndDeliver(), do this in a loop until no messages have been read into the batch.