When Safety Becomes Harm: Toward Subject-Preserving AI Safety in Long-Term Human–AI Interaction | Synapse