Currently it is hard to find which features are selected by model-driven feature selectors.
It would be easy to refactor _LearntSelectorMixin to have a mask_ property that stores the boolean mask of the selected features, and then use that in transform.
Possibly related to #4242.