Auditing Framing-Sensitive Behavioral Instability in Large Language Models for Mental Health Interactions