A very painful bug of mine turns out to originate in the different behaviors of np.copy(a) and a.copy() — specifically, the fact that the latter preserves subclass, whereas the former does not.

If I look through the actual code, it's clear why this happens, but it's certainly not clear from the documentation that it will happen. In fact, I consider the documentation for np.copy downright misleading because it addresses one difference between the two (namely, the default order argument), which lulled me into believing that it was the only difference. (And yes, the documentation technically refers to ndarray.copy, while my subclass was not ndarray, but I don't read documentation like a lawyer.)

I'll acknowledge that changing the default behavior is off the table, so I'll suggest two possibilities that would at least help avoid this situation:

1. Add a subok=False option to np.copy that is just passed through to np.array. The default behavior will remain the same, but the presence of this argument will alert the user to the fact that there might be an issue.

2. Add to the docstring an explicit statement of the fact that subclasses will not pass through.