That may be true for now, but it seems clear enough that letting the model use Lean in its internal reasoning process would be a great idea

That I'd agree with! I really need to get around to learning Lean myself. It might be interesting to try and formalize some missing theoretical pieces from my field (or likely start smaller).