Thanks!

One complication to consider: if the presence of __numpy_ufunc__ disables __array_wrap__ and friends (as it currently does), then carelessly adding __numpy_ufunc__ to ndarray might cause existing subclasses to break.

I guess the most practical way to implement ndarray.__numpy_ufunc__ in the short term would be to teach the ufunc machinery to (as a special case) ignore the presence of this particular __numpy_ufunc__ method when seen, and then its implementation could be

def __numpy_ufunc__(self, ...):
    if <there are any objects with __numpy_ufunc__ methods>:
        return NotImplemented
    else:
        return getattr(ufunc, method)(...)

and this would both avoid infinite recursion and allow subclasses to still take advantage of the __array_wrap__ stuff (since while they would inherit a __numpy_ufunc__ implementation, it would be ignored).

Kinda ugly all around though...

It seems ndarray.__array_ufunc__ exists nowadays, so closing.