While fast charging is critical for reducing EV user waiting time, its high-current operation raises significant safety concerns. This paper proposes a constrained multi-stage fast-charging framework that integrates electrical, thermal, and aging constraints through reinforcement learning, formulating the problem as a multi-objective optimization with dynamic inconsistency tolerance boundaries. Leveraging a TD3 (Twin delayed deep deterministic policy gradient) algorithm enhanced by prioritized experience replay, the method autonomously adapts charging profiles to real-time pack heterogeneity while efficiently exploring constrained solution spaces. The framework incorporates comprehensive constraints including voltage, temperature, and action boundaries during model training to ensure operational safety. Experimental validation that the proposed method achieves an optimal trade-off between charging speed and battery degradation while ensuring operational safety through a penalty-based reward. Meanwhile, the approach demonstrates near-offline-optimization performance, with merely a 1.4% increase in charging time and a 1.9% compromise in reward value compared to ideal offline benchmarks.