Self-Referential Induction Increases Response Instability Relative to Unresolvable and Verifiable Questions in Large Language Models