Since the behavior was changed, I think we need a versionchanged directive.

This function was added in 3.0. Prior to 3.3 it was always successful (if you pass an unicode object, that is required for most of PyUnicode API). Py_FatalError was added in 3.3, because not all unicode object are now valid. But I am not sure that it could be triggered in real code without using some broken extensions.

It may be safer to return 0 instead of returning -1 or crashing if PyUnicode_READY() fails. If return -1 and set an exception, it can lead to an unexpected behavior if the calling code expects only 0/1, and maybe even to crash in debug build due to set an exception and returning value.