Should we remove the requirement to calling fit if it can be avoided?

+1. This can avoid awkward things like #12514 and maybe more user-friendly.

This issue is still open. Any workaround when we have "known" categories in OneHotEncoder?

We are dealing with this issue now with the parameter validation: some stateless estimators would like to validate the validity of some input parameters, usually done at fit even though we will not learn anything on X.

I would be in favour of always having to call fit to validate those parameters and keep the stateless meaning for the estimator not extracting information from the training X useful to transform any X.

At least, having this behaviour will not make these estimators different from others but would still have the tag to ensure some mathematical consistency regarding their stateless aspect.

This PR can be closed since #25190 solve the issue and define the behaviour of the stateless estimator. #24230 also defines a new stateless transformer.

I see that I forgot to mention that in one of the meetings we propose to always make parameter validation in fit but we don't want to request calling it. We added a common test to ensure that this behavior is consistent across scikit-learn.