Towards Faithful and Controllable Personalization via Critique-Post-Edit Reinforcement Learning