This work presents a practical architecture for large-scale last-mile route optimization under real-world constraints, including vehicle capacity, package volume, time windows, and route limits. The system is designed to handle routing problems ranging from small instances to millions of stops without requiring problem partitioning or specialized infrastructure. It combines parallel constraint-aware clustering, distributed rebalancing, and fast route-level optimization to produce globally coherent fleet plans. Evaluated on the Amazon Last Mile Routing Research Challenge dataset, the approach reduces total route distance by 23.3% and route count by 13.6%, with a mean depot-level improvement of 17.59%, while satisfying all operational constraints. In extended experiments, the system processes up to one million stops in approximately 20 minutes on commodity hardware, exhibiting near-linear empirical scaling. This work argues that large-scale routing is fundamentally a systems problem, and demonstrates that scalable, efficient solutions can be achieved through architectural design rather than relying solely on algorithmic advances.