Leaning the manifold is understanding. Reasoning, which takes place on the manifold, is applying that understanding.

I am not sure what you definition of "understanding" is that you apply here.

I mean understanding physics and the universe of natural possibilities; what can happen. Then comes why.

Fitting a manifold to a bunch of samples does not allow you to understand what can happen in the universe. For example, if you train a regular diffusion model on correct sudokus, it will produce sudokus with errors because it does not understand the rules.

You raise a good point for the diffusion case, which trains only on positive examples, but generally speaking negative examples will warp the manifold appropriately.

Sure, but you show a few correct examples to a human it will quickly pick up the correct rules. And this is understanding.