Could use a release note, as it is a behavior change.

Added a release note and a test. The test doesn't really cover the motivating issue here, as rational does not inherit from np.number, but it's better than nothing. It may be worth adding some sort of test-only custom float type to better cover this sort of issue, but I'll punt on that for now.

There’s also the scaled float test dtype using the new dtype system. I want to add numpy-user-dtypes as a test-only dependency to make it easier to test stuff like this.

The behavior change seems fine to me and reads like what the code intended beyond that no one expected someone to write a user dtype set up like bfloat16.

Close/reopen. The test failure looks unrelated, but I would like to know what is going on.

@jakevdp Could you try rebasing on current master to see if that fixes the build failure on CircleCI?

rebased

@jakevdp Thanks, that fixed the build. There was still an undefined reference, which I have fixed.

Thanks @jakevdp .