Going through np.mat also fails on the meta-goal, which is to remove reasons for people to prefer np.matrix to np.ndarray, so that eventually we can deprecate the former without harm.

As far as this goal goes, it's all very well for some of us to say that users should toughen up or whatever, but it's useless: they'll just ignore you and use np.mat because it's easier. And then we have even more of a mess to clean up later.