each VM installs BOINC from scratch at each run. The VM itself has identical configs EXCEPT with a different host name on each one.
I'll try changing that and letting it go back to automatic host names.

Not how I would do it, but at least, if not already done, have it set to report results immediately and abort any unfinished tasks, then communicate one more time to clear the queue before dismounting.

Somewhere there's a description of Deep Freeze as a methodology and certainly read of people restoring their backups of images so as to continue where they were whenever one or the other resource is available again for a crunch resume. Saw someone who does that with his rendering farm. Each image pegged for a specific device, quasi lossless, but for maybe one or the other task that goes overdue when the crunching breaks take too long, running bufferless probably best to reduce that risk to the minimum.

Presume with 'priority zero' you mean the Resource Share aka Project Weight which is default 100. Leaving it at that and setting the "Store at least 0 of work" and "Store up to an additional 0 days of work" would achieve the same thing, with the advantage that BOINC would fetch a new task several minutes before the most imminent one is about to finish i.e. a little anticipation. With zero resource you'd see a little idling of 1 thread as there is a delay in fetching a new task.

Never mind, I'll just put them to work on a GPU project and leave my physical rigs running here.