After writing the simplest possible test of open(2) in C to eliminate anything in Python, I finally established that the seemingly random behavior of the test is in fact caused by a rather odd bug in OS X 10.4.  The filename being used in the test is intended to be invalidly encoded.  On 10.4, open(2) of that filename returns errno 2 (ENOENT) as expected *if* the containing directory is empty.  But if there is at least one existing file in the directory, the open returns errno 22 (EINVAL) instead.  Go figure!  Knowing that, it's easy to reproduce the test failure by running the test directly.  So, the buildbot failures are intermittent depending on what files previous tests have left behind in the working directory. (I haven't checked intermediate versions but current OS X 10.8 does not have this odd behavior: ENOENT is always returned.)  The original proposed solution still seems to me to be the right one: simply change the test to pass if either EINVAL or ENOENT is returned.