Federated Learning (FL) enables privacy-preserving, collaborative model training across massive populations of edge devices. However, its practical deployment is impeded by device-network heterogeneity, Byzantine adversaries, and the scalability bottleneck of centralized coordination. We propose ByzTierFL, a fully decentralized, Byzantine-robust, tier-aware FL framework that addresses these challenges in a unified manner. ByzTierFL dynamically stratifies clients into tiers based on latency and data quality. ByzTierFL then performs inter-tier aggregation using ByzShield, a novel consensus mechanism that combines multi-model screening and reputation-driven filtering to suppress malicious updates. Built on a hybrid IPFS Hyperledger Fabric substrate, ByzTierFL supports auditable provenance and scalable storage, reducing on-chain communication overhead by 85%. Experiments on non-IID image and text benchmarks show that ByzTierFL reduces per round latency by 65.8% relative to PBFL baselines and maintains over 90% accuracy even when 20% of clients are Byzantine. The framework demonstrates strong scalability and resilience, providing robust and efficient federated learning for real-world heterogeneous and adversarial environments.