The reason why np.where(condition) is discouraged is documented as follows:

Using nonzero directly should be preferred, as it behaves correctly for subclasses.

I don't recall any issue reported by users using scikit-learn with subclasses of NumPy arrays and having problems as a result. I am not even sure what "behaves correctly" means in that context: returning a NumPy array of integer indices or a subclass instance storing the same indices? For our use cases, I think returning a NumPy array (not subclassing) is perfectly fine, but any option might work.

Let's try:

>>> import numpy as np
>>> class my_array(np.ndarray):
...     pass
... 
>>> condition = np.asarray([True, False, False, True, True]).view(my_array)
>>> condition
my_array([ True, False, False,  True,  True])
>>> np.where(condition)[0]
array([0, 3, 4])
>>> condition.nonzero()[0]
array([0, 3, 4])
>>> np.nonzero(condition)[0]
array([0, 3, 4])

So I don't see any difference: neither returns an instance of the my_array subclass.

I traced this recommendation in the doc to this PR by @eric-wieser: numpy/numpy#11425

However, whenever we convert a scikit-learn estimator to add array API support, the xp.where prototype in the spec implies that the second and third argument are not optional (contrary to calling np.where with a single condition argument):

https://data-apis.org/array-api/latest/API_specification/generated/array_api.where.html#where

Since nonzero is also part of the array API, I think it would make sense to use xp.nonzero instead when we convert code to add array API support:

https://data-apis.org/array-api/latest/API_specification/generated/array_api.nonzero.html#nonzero

Apart from adding array API support, I would rather not change the code unless we identify a good reason to do it (in particular if we identify a user-facing impact). Personally, I am not convinced that one is particularly more readable than the other.

That being said, I would be curious to understand better why np.where(condition) is discouraged in the first place: it's not clear from the NumPy doc and my quick experiment above failed to clarify the issue.

Maybe the note about the issue with subclasses is actually a reference to using np.where with a masked array: numpy/numpy#11425 (comment).

@eric-wieser @seberg friendly ping. We would like to better understand the numpy recommendation of using np.asarray(x).nonzero() over np.where(x).

Personally, I would follow numpy‘s recommendation without questioning it too much: their realm, their expertise.

where skips the .nonzero() attribute, so if a subclass defines that it is different (maybe only then). Of course these days the subclass could define __array_wrap__ so it is less of a thing maybe.
I never loved the overload of where to have two meanings, although I bet "where" sounds nicer to many than nonzero.