Harness-native RL for coding agents: the trained policy and the training task index.