Thanks to the ReleaseReason tip, I was able to determine that the unexpectedly released job was being condor_released, and so I wrote a wrapper script to log the process tree and environment each time that condor_release was run, which led back to another housekeeping job that was releasing held jobs if a given mountpoint was operational, apparently assuming that any held job was held because of a mountpoint problem..
I'm sufficiently impressed by this sleuthing that I wonder if this sort of audit log ought to be something that HTCondor offers as a knob for its admins....
- ToddM