"The second, more fundamental reason is that the relevant features are not given in advance. A large part of what the expert learns through experience is which features of the environment matter."

Yes.

"This is the deepest reason why experience cannot be compressed."

No. Finding relevant features and compressing their valuation is what ML systems do. Especially in vision systems. Early attempts at machine vision had human-chosen features and detectors for them. There used to be papers with suggested features - horizontal lines, vertical lines, diagonal lines, patterns of dots, colors, and such. That's where things were in the 1990s. Those have now been replaced by learning-generated low level feature detectors, which work better.

This does need a lot of training content. The training process is inefficient. It definitely compresses, though.