I had attached the nvidia-bug-report.log to my bug report; here it comes now. Note that it has not been produced during/after the actual lockup occurrence as my only remedy for those so far was hardware reset. If it is paramount to gain access to the box after a crash (assuming daemons still respond), I can try that.
Going by http://www.nvnews.net/vbulletin/showthread.php?t=58498
- scanned logs for NVRM messages: nothing suspicious
- scanned /var/log/dmesg for "PCI: Using MMCONFIG": not there
- kernel is 2.6.20-16, so concerns for earlier versions should not apply
- no obvious firmware updates available
I _did_not_ try "idle=poll" as I am not yet sure about its [primarily performance] implications on a system intended to run multiple active VMs.
Thanks in advance for any suggestions. My alternative is to go back to generic open-source drivers as 3D is not my primary concern right now, but I'd rather get the nVidia driver to work in a stable manner.