This issue is a continuation of the disussion with @seberg in #3501, along with #3502:

Lots of places (such as 1, 2) use PyArray_Type->tp_as_number->nb_multiply to call the multiply ufunc (etc.). Because of __array_priority__, the semantics of nb_multiply are no longer the same as the ufunc. Specifically, nb_multiply is required to to return NotImplemented in cases of priority inversion, whereas the ufunc has a job to do and just needs to do it, without asking questions about priorities.

EDIT (seberg): This is still the case and probably cause of a few bugs (I am very sure reading an issue recently). It shouldn't be hard to fix, we have the nb_ops or mapping that is used for the array object: That needs to used throughout for scalartypes.c.src and scalarmath.c.src would do so indirectly then, although directly might solve additional unnecessary dances.