But in this case, they are using poll as the time-honored way to sleep. Yes Apple should follow the expected POSIX behavior but at the same time if sleep is what is desired, there's usleep(3) and nanosleep(2) which are found in libc's time.h
Also, according to the developer, this same behavior was present from 10.3 to 10.9 ( Apple was certified to be compliant from 10.5 onwards which tells you that this isn't non conformant) and OS X did not even ship with a poll(2) implementation before 10.3 .
It's also possible that the test suite just doesn't check for this. In any case, the behavior is necessary for any application that polls against a set of fds of varying size, as the fd set may shrink to zero before expanding again, and and punishing the developer for not checking as special case that all other unixes handle automatically is pretty unfair.