As a follow up to my previous e-mail, I set highres=off on the kernel command line on a hunch, and the machine has been up for 22 days now. It looks like that made the problem go away, but that may just be because the offending code is called less.