>what an actual non-Markovian LLM would look like? Something which has an infinite context? It's not physically possible... At this point it sounds like "it's a Markov chain" is not a very useful statement.

Uh... You're moving the goalposts. I don't know the exact definition of "LLM", but let's suppose that it and the definition of "Markov chain" make us conclude that LLMs are necessarily Markov chains. "It's a Markov chain" is still a useful statement, because there are processes that are neither LLMs nor Markov chains.