Martin, your reproducable behaviour involves consistently disabling 
functionality that those threads you mention may rely on (and since 
this i is a global state that no one messes with, they are unlikely to 
be doing so)

If the alternative is then just to live with invalid file descriptors 
causing crashes, except for the test suite, as Mark suggests, I can 
also live with that, I suppose.  But then we would want the additional 
tests in my patch for strftime and fopen()