Same Luck, Different Choice: Twin-World Credit Assignment for Agentic Reinforcement Learning