Learning to Bridge the Gap: Efficient Novelty Recovery with Planning and Reinforcement Learning