test_curses currently fails on 3.1 when run under regrtest.  It passes
if run in verbose mode or standalone.  The cause turns out to be that
when not run in verbose mode regrtest substitutes an io.StringIO
instance for sys.stdout, and that object does not have a .fileno
attribute.  The line that fails is this:

curses.setupterm(fd=sys.stdout.fileno())

This test fails only on 3.1 for a couple of different reasons.  On trunk
and py3k, r73072 and the r73678 merge removed the regrtest check that
made sure the tests produced no output on stdout, and therefore it no
longer replaces stdout with a StringIO instance .  On 2.6, on the other
hand, the above line references sys.__stdout__.fileno() (presumably for
this very reason!)  r74181 was the commit that changed this to the above
on the py3k branch, citing it as the cause of regrtest "duplicating some
output".

It seems to me that not checking for output on stdout in regrtest is a
regression in the quality of the test runner.  It is also not clear to
me how test_curses writing to the real stdout would cause regrtest to
duplicate output.  So I'm putting the authors of the two patches
involved on the nosy list so they can comment.  FYI I reverted the
stdout change in test_curses and regrtest seems to run fine, so I'm not
sure what output was being duplicated.