When Context Leads but Parametric Memory Follows in Large Language Models