Synthetic Error Injection Fails to Elicit Self-Correction In Language Models