Raised in #22986 (comment) the question is whether these two sets of methods should always accept the same metadata:

• fit and partial_fit
• predict, predict_proba, predict_log_proba, and decision_function.

I think they should, and I suggest we add a common test for them accepting the same metadata, and only have set_fit_request and set_predict_request exposed to users.

cc @jnothman