Decomposing the Entropy-Performance Exchange: The Missing Keys to Unlocking Effective Reinforcement Learning