This isn't really right:

• You would have to use copy in the asarray call.
• The copy=copy in astype() doesn't have the same semantics.

I am actually not sure if astype should coerce from to an array at all? If this does a forced cast, then astype becomes asarray().

I also wonder if any coercion should be allowed here and I would agree if it shouldn't.

Then e.g. np.astype([1,2,3], np.float64) just wouldn't be supported.

I think I agree about no coercion here. Instead, it seems to me like this should verify that the input is already an array.

That would leave the question of how exactly to do that. The dispatcher for __array_function__ was already added, so I think it should work to check with isinstance(x, np.ndarray) (allows subclasses, but not duck arrays - those should be using __array_function__).

And, if we do that, what's the right exception type? I'd think TypeError?

I second TypeError. Its documentation says:

Raised when an operation or function is applied to an object of inappropriate type.

Maybe also a note that array-likes are explicitly not supported here?

Right, I updated it to x : ndarray to follow existing docs in other functions.
I also added the note.

Uh, I noticed I misspelled PR number - fix is in #25054.