Compositional Hardness of Code in Large Language Models -- A Probabilistic Perspective