nx_lawsearch.nx

nx_lawsearch.nx -- DENSE SEMANTIC RETRIEVAL over the learnings corpus (the SOTA stage-2 unlock). WHY THIS EXISTS. The ecosystem owns a real 41MB PPMI model (knowledge/index/semppmi_v1.bin, vocab 102318, 2558014 triples) and measured dense p@1 1000 permil vs lexical 500 -- but NOTHING could query it: nx_recall_dense is a fixed-fixture proof with no argv, so the single largest retrieval asset we own was unreachable (debt 1785560370). This organ gives it a query interface, composing the CANONICAL nx_ppmi_lib rather than hand-rolling a third copy of the loader (D001 discipline -- the lib header records that two copies had already silently diverged). Sparse BM25 cannot match a synonym: "reversible" never retrieves "undoable". Our own bench sizes the gap -- BM25 ndcg@5 494 permil, fused 659, reranked 964 -- and 2026 agent-memory SOTA is hybrid sparse+dense merged by RRF then cross-encoder reranked. learn_query.ps1 already ships the sparse+RRF half; this is the dense view it fuses with. dense <query> <corpus.jsonl> <model.bin> [k] -> ranked JSON lines {score,rank,slug,text} selftest <model.bin> -> gate teeth incl. a DENSE-BEATS-LEXICAL tooth license_tier: ORIGINAL