nx_beir_eval.nx

nx_beir_eval.nx -- SOVEREIGN BEIR harness: real external ground truth (BEIR/nfcorpus) scored by our own integer BM25 -> mean nDCG@10 (the standard BEIR metric). The HONEST external row the maturity ladder is gated on (F236): real 3633-doc corpus + 323 test queries + 12335 graded qrels, vs the published BEIR BM25 baseline (nfcorpus nDCG@10 ~0.32, Anserini). BM25 k1=0.9 b=0.4 (BEIR defaults). IDF=log2 (ranking-invariant vs ln -> identical nDCG). Integer/fixed-point throughout (no float) = bit-reproducible. Forward index (per-doc sorted term,tf). Tokenizer = lowercase [a-z0-9] runs -> PORTER STEM via nx_stem_lib (2026-08-13), NO stopwords (the stemming rung LANDED -- was the named gap to the stemmed 0.32 Anserini baseline). R0d be_bright (2026-09-14): `--set <root>` scores ANY BEIR-shaped set root (<root>/corpus.tsv, queries.tsv, qrels/test.tsv and, when present, excluded.tsv: BRIGHT lists each query's own source documents there and the benchmark removes them from that query's ranking, so this harness marks them unrankable for that query and counts them). Every capacity is now DERIVED from the files (sys_read_file sizes the buffer, newline counts size the tables, the longest line sizes the token scratch): the old 64 MiB file cap and 4096-document cap would have dropped 99 percent of BRIGHT's leetcode split in silence, which is the cap class this estate refuses. `--bm25-only` skips the rerank arms (their models were trained on the web corpus; BRIGHT's published floor is plain BM25, 14.5 average nDCG at 10 on the leaderboard page read 2026-09-14, search.refs bright). `--bright <root>` is be_bright: this binary runs itself once per split through tr_run_capture_to (one hour each, the largest split bounds it), prints one JSON row per split, the average over twelve of twelve (never over a partial set), and writes the estate's row into search.leaderboard beside the published rows with its rank DERIVED from their scores. The positional grammar is unchanged: [model.safetensors config.json vocab.txt [max_len [depth [workers]]]] after the flags; workers is the cross-encoder pool width (0 = serial). expect_exit: 0 license_tier: ORIGINAL