What I meant was that it's a regression in the parity of behavior between normal and TEST_DIRECT cases. In the TEST_DIRECT case, there is no fork. The only way that files get cleaned up is if they are in entered into the temp_name_list in the PREPARE phase.