round() delegates to class.__round__, so what's happening here is that numpy's float implements __round__ in a way it returns a float.

There is no restriction on what the class can return:

>>> class A:
...     def __round__(self):
...         return "Oh, no!"
...
>>> round(A())
'Oh, no!'

Making additional restrictions is backwards incompatible. Maybe we should add something in the docs regarding the no-restriction point.