Disentangling Mathematical Reasoning in LLMs: A Methodological Investigation of Internal Mechanisms