Update benchmark.md with v1.1 results (83.3% token reduction)

[Copilot]

Benchmark suite rerun shows improved compression metrics. Updated documentation to reflect actual measured values.

Metrics Updated

• Global token reduction: 74.6% → 83.3%
• File coverage: 29 → 30 files (9,671 raw tokens measured)
• Precision targeting: 96.7% (refined from "100% retrieval accuracy")
• Cost efficiency: 4x → 6x reduction in API token costs

Heavyweight Module Rankings

Reordered top 5 modules by actual token count with updated compression rates:

Module	Original → Compressed	Rate
indexer	1,339 → 49	96.3%
parser	1,232 → 63	94.9%
query	1,224 → 108	91.2%
vector_store	1,013 → 147	85.5%
generator	839 → 23	97.3%

Average compression across top 5: 93.3% → 93.0%

Context Window Impact

Updated GPT-4-8K usage calculations based on measured 9,671 raw tokens:

• Traditional approach: 12-15% window usage
• AMDB approach: ~2% window usage (1,615 tokens)

Benchmark version bumped to 1.1, dated 2026-02-11.

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.