No, I had not tried that. Yes, that works. Thanks.
However, we still need to avoid it, and we do need the output/error files (just in case things go amiss, we want to be able to do a post-mortem analysis).
This seems to work. The exiting-state still takes abnormally long (about a minute or so), but it does finish, so the nodes get cleared for further jobs.
Kudos to you both. Thanks!