Hmmm, until now we explicitly protected from it, so this might break someones __array_ufunc__ implementation in theory (I haven't looked if any of the big ones -- e.g. dask/astropy/cupy would notice).

But, I can't think of a different way to ensure the warning isn't given. But we could at least make it conditional on where to (very much) reduce the potential blast radius (since where= isn't used a lot).

Can you also check if the docs need to be updated about this? And maybe add a brief release note.

EDIT: Also clearly, the tests are failing, and need to be adapted to test whatever new behavior we have.

@ngoldbaum, do you know why the tests are failing? I thought where would be in normal_kwds, but the test failure seems to indicate that it's not, unless I'm making a silly mistake.

I think the reason for the failure is that out is passed by position and not kwarg, so it is never part of the dict from the start. I dunno if it matters much, I guess you would instead have to add out=None to the kwargs (but then if there are multiple out should it be (None, None)?
(I would be tempted to pass in the info that where is passed, as it is available, but more of an opinion. Just to avoid the Contains)

One thing that might also help is to debug a local build of numpy with gdb or lldb:

https://numpy.org/devdocs/dev/development_advanced_debugging.html#c-debuggers

I'll try to take a closer look to understand what's going wrong. Also ping @mattip - this is fixing some fallout from merging #29813 for projects that implement __array_ufunc__.

Thanks @ngoldbaum, I will get that set up.

I'm surprised wheremask_obj == NULL didn't work, I need to read the code further to determine how the override knows if where is passed,.