BUG: incoming incompatibility with numpy-dev

[neutrinoceros]

example logs

The impacted test isn't released yet (source #593)
This is another blocker for our next patch release.

ping @jzuhone and @kyleaoman

Keep in mind it's possible that the actual bug is upstream in numpy. I'll attempt a bisection to narrow it down.

[kyleaoman]

Actually this is not a problem - when I implemented this we decided that the difference in units was acceptable, and the test is written this way. In the "failure" the function now returns the "more correct" units, if you see what I mean. I would suggest amending the test and calling it a win if some upstream change in numpy has miraculously fixed this inconsistency...

[neutrinoceros]

Gotcha. I'm running a bisection now so hopefully we can pin point this miracle and take an informed decision.

[neutrinoceros]

bisected to numpy/numpy@a3a2f2f
(numpy/numpy#29951)

[neutrinoceros]

@kyleaoman does this make sense to you somehow ? I don't understand what happened off the bat, nor if we can reasonably take assume the new behavior is guaranteed to be stable going forward.

[kyleaoman]

At a glance I'd say that it's that some calls to multiply have been replaced with square. Easy to see how this would affect std, and since the special handling is around multiply specifically it's a different call stack entirely.

Not trying to be too lazy but could just update the test to handle before/after this numpy commit as both passing? Or one of the cases could be an xfail... like try: assert except assertionerror: if numpy version < xyz: xfail. Sorry for the deplorable formatting, typing on phone...

[neutrinoceros]

Apparently it's still not clear wether the new behavior can be relied on: numpy/numpy#29951 (comment)

so I propose we wait this one out.