Federated Learning (FL) enables collaborative model training across distributed devices while preserving data privacy. However, it faces critical security challenges, including centralization risks and poisoning attacks, which degrade robustness and scalability. Existing schemes struggle to simultaneously mitigate targeted and untargeted poisoning attacks, impose restrictive adversary ratio assumptions (poison ratio < 50%), and suffer from privacy-performance trade-offs. To address these limitations, we propose SharBipole, a decentralized FL scheme integrating sharding blockchain with a novel dual-metric defense mechanism, Bipole. SharBipole employs a Byzantine Fault Tolerant-enabled sharding architecture to eliminate single points of failure, reduce communication overhead, and enable parallel model aggregation. Meanwhile, the Bipole module defends against poisoning attacks using two adaptive similarity metrics to filter malicious updates dynamically. Reinforcement learning optimizes threshold adjustments, while noise-aware adaptive clipping balances privacy and model utility. Further, we give convergence analysis to prove the theoretical soundness and scalability of SharBipole. Lastly, extensive experimental evaluations demonstrate that SharBipole supports poison ratios exceeding 50% and improves throughput and latency. The model replacement attack with 60% adversaries is entirely ineffective against SharBipole, and the label-flipping attack achieves an attack success rate of only 2.344%. SharBipole establishes a scalable, secure, and privacy-preserving solution for distributed learning in massive environments.