Self-Improvement with Auxiliary Rewards: Learning from Verified Traces and Evaluating Transfer