It's a bit funny, I wonder if just flagging it as parametric would actually work :). To some degree, it is well defined (as in, it is identical to the user passing no dtype).
OTOH, in practice it should typically be a user error here. I do wonder if there is an earlier place to reject this nicely...

But, overall, I am also happy with just doing this, I don't think we would have to fret over ever changing it in the future anyway.

I wonder if just flagging it as parametric would actually work :)

It turns out this makes the test I added create afloat64 array, which I guess makes sense, there must be a default cast tofloat64 somewhere.

I do wonder if there is an earlier place to reject this nicely...

I ended up moving the check intoPyArray_DTypeOrDescrConverterRequired so the error happens a little earlier.

Yeah, I think returning float64 there is actually completely fine after all that is whatnp.dtype() gives. Although I am not confident we end up there the right path.

Making this an error for now here seems fine then. If anyone ever feals to relax it to return the float64, I would not be opposed at all, though!

Thanks Nathan, let's go with this and possibly relax it if we ever find a reason for it.