I think the build failures on the OpenBLAS CI are real. Might just be a matter of giving result in the convert_to_@name@ template a default value?

Fun that it suddenly shows up, it isn't new. Anyway, harmless but not wrong, easy enough to avoid.

I guess this is because isinstance(np.float64(1.), float) is True, right?

This is a subtlety that we might otherwise run into in Python, I guess. So a bit :).
I guess I explained my reasoning in the issue (and meetings more) only:

• I think NEP 50 is really about convenience of using Python literals in code. Subclasses are too niche to be very relevant there, IMO.
• Subclasses can have any odd meaning (including something like float64 that we don't even know about). Treating them like Python float is honestly probably wrong at least half the time. (Of course we do end up treating them like float64, which isn't really better).

Either way, in some future a custom DType could probably create custom weak scalars if that comes up. But that would have complicated things (and this was difficult enough). Once legacy promotion paths are gone, I suspect one could attack it. Although, I honestly doubt it is worthwhile to do it even if we can.

Thanks, that all makes sense!

Thanks @seberg