Add vecdb (PyPI: vecdbc)

[navid72m]

vecdb is a from-scratch vector database in C11: HNSW graph with hand-written SIMD distance kernels (AVX-512 on x86, NEON on ARM). Published on PyPI as vecdbc; the Dockerfile installs it from there.

Adapter: ann_benchmarks/algorithms/vecdbc/ (module.py, config.yml, Dockerfile)
Handles euclidean directly and angular via L2-normalization
Single-threaded per the framework's convention
Tested locally: passes random-xs-20-euclidean (recall 1.0, exit 0 across all parameter groups) and runs cleanly on sift-128-euclidean
Added to CI test list for random-xs dataset

[navid72m]

Verification: recall/QPS on sift-128-euclidean

Ran the adapter through the standard harness. Recall@10 vs the provided ground truth, single-thread, M=16 / efConstruction=200, sweeping ef:

ef	recall@10	QPS
10	0.740	20,262
20	0.858	18,177
50	0.954	9,041
100	0.985	5,075
200	0.996	2,812
400	0.999	1,538

(Numbers above measured on Apple Silicon; the Docker harness on x86 will use the AVX2 build path. Posting these to show the adapter produces a correct, monotonic precision/recall frontier — recall climbs smoothly with ef and matches a reference HNSW implementation to ~3 decimals on this dataset.)

Build time on SIFT1M stays well under the "several hours" guideline (~hundreds of seconds at these parameters). The config sweeps M ∈ {12,16,24,32,48} so the framework can pick the precision/performance frontier rather than a single fixed point.

Also verified the angular path: vecdb uses squared-L2, so angular datasets are handled by L2-normalizing both base and query vectors (cosine ranking == L2 ranking on unit vectors), giving correct recall against cosine ground truth.

Happy to adjust parameters, the sweep, or anything else if it would make the benchmark more useful.