When LLMs Stop Following Steps: A Diagnostic Study of Procedural Execution in Language Models