I think the regulatory model (looping etc) first is good with examples that are less extreme and unusual than fto or limb/shh. I think hmgcr and cholesterol is a proximal regulatory effect.

Yes, and also how it informs selection of training datasets. Selecting English language text for English language comprehension, augmenting image datasets by including rotated versions of the same images for image recognition.

I think it already has some physics (eg covalent bond lengths), just less than you might expect (no force field concept). Similar, llms have some knowledge of language (eg the concept of words), just less than you might expect (no verb/noun concept).