Sebastian Berg

Seems OK to me, I do wonder a bit if there isn't a shorter solution (and ideally `:g` formatting would probably just work). But these are testing utils, and I...

Ah, if you care a lot, I guess we could add a `.. versionchanged::` but with this being in testing, I don't think it is particularly important and adds visual...

There are multiple approaches to this. For 2.0, i.e. the "major version" C-extensions had to be recompiled and a good amount of python libraries broke for simple things like `NaN`...

There is no good reason to pin to exactly NumPy `2.*`/`~2`/`>=2,

I'll note that indepndent of your own relase cadence, there are sometimes misgivings about declaring a narrow version support, because of transient pinning: If your users tend to pin your...

We just brought it up, and I think the main reason this might fail I think is that before there was never any casts, `out` was basically ignored (i.e. replaced)...

Yeah, we should move on, it's just would be nicer to know how much and how badly it might break things. It seemed like this isn't used much (thus why...

I would prefer the strict version since that will shout at you in a useful way, while the other one could in principle just do the wrong thing silently. (The...

Ref https://github.com/numpy/numpy/pull/24818, does anyone know what the argument/reasoning for this was? * Avoid variable number of result dimensions? * Deprecated `cross([0, 1], [1, 2, 3])` from working?

OK, since I missed it there. The issue is the first thing and the reasoning is in https://github.com/numpy/numpy/issues/13718 which suggested a `cross2d`, though.