How Well Can Reasoning Models Identify and Recover from Unhelpful Thoughts?