Real query embeddings and a fixed query set for recall and the CI gate #23

Today the queries are random document vectors held out from the base (own data and DBpedia; GloVe and SIFT use their published query sets). Real user queries are often short and sit in a different part of the space, so a sampled-vector test can look healthy while recall on real queries drops.

1. Real query embeddings. Add --query-file (.npy / .parquet) and, for Postgres, --query-table / --query-column. When given, report recall on those queries and on the held-out sample as separate columns, and let --min-recall apply to either.

2. Sampling noise vs. a real regression. Query sampling is seeded (--seed), so the CI gate already uses a fixed sample while the data doesn't change. But there's no estimate of how much recall moves with a different sample. Add an option to resample the queries N times and report the spread, so a CI drop can be compared against it. A fixed query file (1) also gives a stable set to gate on.

Docs: say clearly in the README and docs/method.md which queries the numbers use.

Suggested in a Reddit comment on the launch post.