Architecting a Fault-Tolerant LiteLLM Router: Cost vs. Throughput Across 232 Inference Providers