Fixed (I hope!) in the trunk in r69459. I'll wait for buildbot results
(just in case) and then merge to 2.6, 3.1 and 3.0.
The same test_cmath failure can also be seen on OS X 10.5.6/Intel when
compiling with -fmpmath=387. Annoyingly, the fix above doesn't work here:
it seems that the OS X isinf is buggy. It doesn't seem worth working
around this bug though, since there's little sane reason to be compiling
with -fmpmath=387 in the first place, so it's unlikely that any regular
Python users will encounter this problem.