Text this: How Code Representation Shapes False-Positive Dynamics in Cross-Language LLM Vulnerability Detection