[bhope]

Summary

• Added streaming / LLM inference section explaining TTFH vs TTFT signal distinction and how hedge handles prefill disaggregated architectures
• Updated benchmark table with concrete numbers from 50k-request runs (p99: 64ms → 17ms, 74% reduction, ~9% overhead)
• Added no-hedging baseline to streaming latency chart for meaningful comparison
• Regenerated both eval.png and new eval_streaming.png charts using matplotlib with consistent styling
• Minor fixes: observability API, streaming table numbers, chart title