The moral of the story is always, always use at to schedule a reboot shortly in the future, no matter how harmless you think something is...
Today I stopped procrastinating about upgrading daedalus's kernel to 2.6.12 and did it. That worked fine. The reason for the upgrade was that I'd read somewhere that 2.6.12 fixed the problem with things like Dell's Baseboard Management Controller, where if you downed the interface in Linux, it downed it so far so that the BMC (which does some funky low-level physical interface sharing) also stopped being remotely accessible.
So I upgraded, hopeful in the fact that if I were to ever remotely shut down daedalus, I could turn it back on or reboot it remotely with ipmitool. I thought I'd test this, so I did this:
daedalus:~# ip l s dev eth0 down; sleep 10; ip l s dev eth0 up
and promptly kissed daedalus goodbye. Not just for 10 seconds, for until I could get the colo guys to kick it in the guts for me. Grrr.
At least I determined that the problem appears to persist. If I down an interface in Linux, it's down for the BMC as well. That sucks.