Benchmarks

Benchmarks — headline numbers, all measured

Raw methodology + artifacts stay in the repo: docs/benchmarks.md, docs/benchmarks-requests.md, semantic raw JSON next to them. Dated records live in the CHANGELOG.

What	Number	Where measured
Semantic retrieval	recall@8 = 1.0 on 16/16 labeled questions (requests + urllib3); hybrid MRR 0.854 / 0.781	docs/benchmarks-semantic-*.json
Caller resolution vs grep	1 cie call vs 3 grep+reads, still not guaranteed correct	docs/benchmarks-requests.md (includes the 3-of-6 under-resolution miss)
Index speed	110 files → 1,902 nodes / 6,581 edges in ~1.9s	the README demo take (uncut cast in docs/demo)
Surface health	135 tools, conformance: 0 crashes, honest-empty envelopes	tool-test-lab harness
TESTS edges after v0.1.4	1 → 562 on cie itself	CHANGELOG [0.1.4]

The losses and ties are published too — the benchmarks docs report a tie and a real loss, not just the wins.