Are sparse NLU features passed to TEDPolicy?

The diagrams in this blog post suggest that features from CountVectorFeaturizer are fed into the TEDPolicy.

However, I added debug statements to the TEDPolicy code (Rasa version 3.1) and only saw basic sparse features (intent, action, entities, slots, active loop).

Does the TEDPolicy only make use of NLU features in the end-to-end mode described here?