bench: all-Rust BEIR benchmark + README above-the-fold

.gitignore

@@ -39,3 +39,8 @@ venv/
 .DS_Store
 .idea/
 .vscode/
+
+# BEIR benchmark harness — embedding cache and result files.
+/.cache/ordvec-beir/
+/results/beir/*
+!/results/beir/.gitkeep

CHANGELOG.md

@@ -7,6 +7,21 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## Unreleased
 
+### Added
+
+- **Reproducible BEIR benchmark harness** (`make benchmark-beir`; dev-only,
+  excluded from the published crate). All latency is measured in a single Rust
+  process (`benchmarks/beir-bench`) — ordvec's rank/sign methods against an exact
+  inner-product baseline (`flat`, identical retrieval to FAISS `IndexFlatIP`, via
+  a pure-Rust SIMD GEMM) and a pure-Rust HNSW (`hnsw_rs`, M=32) — so the
+  comparison is apples-to-apples (same machine, batch, thread count, no
+  Python/FFI in the hot path). Covers single-query / batched / 32-thread regimes
+  and a corpus-size scaling sweep on public BEIR datasets, with the corpus
+  embedded by Harrier-Q8 (GGUF `Q8_0` via `llama-cpp-python`, CUDA). The README
+  now leads with the resulting scaling curve, latency bars, and nDCG@10 table;
+  every figure is regenerated by the harness (nothing hand-entered). Replaces the
+  previous private-arXiv real-embedding numbers in the README.
+
 ### Performance
 
 - **AVX-512 VPOPCNTDQ scan kernels now cover every `dim` (a multiple of 64), not