I think it's worth clarifying that we're not talking about return_*, which isn't really applicable to the routing mechanism, because it's not about inputs..?

I'm okay with this at first thought... what are the risks if we have to then split them up later???

I think having to distinguish between return_* kwarg and "input metadata" will complicated the API for third party estimator developers.

Also for classifiers, partial_fit accepts classes that we do not usually accept in fit. We would need to introduce classes to fit if we want fit* to accept the same metadata.

Hmmm… I suppose there are a couple of needs for the return parameters: either it is needed directly by a wrapping estimator, or it’s needed for scoring or similar and requested by the user. I had been thinking of the former case, but I suppose you are right about the latter

I think having to distinguish between return_* kwarg and "input metadata" will complicated the API for third party estimator developers.

If developers would want to remove an arg from the routing mechanism, the API as is allows us to do so:

__metadata_request__fit = {'return_blah': RequestType.UNUSED}

The above code would remove return_blah from the whole request/routing mechanism.

So to summarize: fit and fit_transform have different input args and therefore we would have both methods: set_fit_request and set_partial_fit_request.

Do we also want to keep the other predict* group separate? I think it'd be foreseeable that decision_function accepts different input args than predict* at least.

So to summarize: fit and fit_transform have different input args and therefore we would have both methods: set_fit_request and set_partial_fit_request.

There's a typo here, I think?

Do we also want to keep the other predict* group separate? I think it'd be foreseeable that decision_function accepts different input args than predict* at least.

I'm really unsure about this one. I think these methods are likely to have different return_ flags, but otherwise should receive the same params??

So the conclusion seems to be that we're gonna keep them separate at least for now. I'll open a PR to reflect this.

Fixed in #23342