Right now there's some estimators that don't require calling "fit", two that I'm aware of: Normalizer and FunctionTransformer. They do input validation if fit is called.
There's one estimator that is stateless but requires calling fit for no real reason I can see, AdditiveChi2Sampler.

My questions are:

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

• If fit is called, should we ensure that the number of features is the same in fit and transform, even though that's not required by the algorithm to avoid user errors?