A language model does not have to reason to be able to produce textual matter corresponding to code. For example, somewhere, n blogs were written about algorith

A language model does not have to reason to be able to produce textual matter corresponding to code. For example, somewhere, n blogs were written about algorithm x. Elsewhere, z archives in github have the algo implemented in various languages. Correlating that bit of text from say wiki and related code is precisely what it has been doing anyway. Remember: it has no sense of semantics - it is “tokens” all the way down. So, the fact that you see the code as code and the explanation as explanation is completely opaque to the LLM. All it has to do is match things up.