Catch Me If You Can: How Smaller Reasoning Models Pretend to Reason with Mathematical Fidelity