docs: update README with judge-validated MEDIUM/HEAVY benchmark results

fix: bump bench max_tokens from 512 to 1024

feat: add LLM judge to quality benchmark

feat: extend quality benchmark to MEDIUM/HEAVY tiers

docs: add benchmark results and context optimization roadmap to README

fix: support synonym matching in quality benchmark

fix: relax keyword criteria for false negatives

feat: expand quality benchmark to 3-model comparison (Haiku/Sonnet/Opus)

feat: add quality benchmark — Haiku vs Sonnet on LIGHT tasks

feat: add MuxConfig presets + dual-baseline benchmark comparison

feat: add routing benchmark — 50 prompts, dry-run cost comparison

feat: implement Phase 3 Agent SDK wrapper — mux_query, recommend_model

feat: implement Phase 2 compaction — truncation, dedup, summarization

feat: implement Phase 1 core MVP — config, classifier, router

init: project scaffold with CLAUDE.md