sigh, looks like I'm not quite done flailing yet.
it occured to me that I could check my theory on the history
of splpmap() and ctx_alloc() by looking at previous releases,
and it looks like I'm wrong.  splpmap() didn't appear until 1.2,
and before that ctx_alloc() didn't use any spl calls.
but it was such a nice theory!  oh well.

on the other hand, I still think the patch in my previous message
will do what we want.  since cpu_switch() does "spl0(); splclock();"
at the top (I typoed that in the previous message), most of cpu_switch
runs at splclock() anyway, so using that IPL also for the little bit
at the end where we call ctx_alloc() seems harmless.

of course, this still doesn't explain what caused the overruns
to appear in the first place.  maybe I should just get some sleep...

-Chuck