Fiction-derived LLM evaluation datasets focused on clinical communication, AI trust rupture, emotionally charged conflict for human edge cases