“Reward hacking is typically easy to spot during the iterative rubric development process because the model will quickly and repeatedly exhibit the rewarded but undesirable behavior.”