Follow the measured tradeoffs.

Explore recorded optimization runs across vLLM, SGLang, and TensorRT-LLM. Replay throughput and p95 latency observations, inspect configurations, and keep failed and pruned trials in view.