nx_incumbent_lib.nx

nx_incumbent_lib.nx -- THE ONE RULER FOR A RUNG CONTRACT'S INCUMBENTS AND THE RECORD'S REJECTIONS (engineshift ES28). Operator 2026-09-14, on the search R0 review: "this is the stuff we need the search and nishi estate to do proactively not rely on you treasure hunting." A prior seat rebuilt the R0 rerank arm as a learned fusion over PPMI while the estate already held the trained dense embedding (knowledge/index/embed_v1.bin), the lib that scores with it (nx_dr_densejudge) and, in that lib's own header, the MEASURED REJECTION of exactly that fusion (DR-13). The reviewing seat found all three by glob and header reads: a retrieval failure, not a discovery. nx_capsearch ranks registered tools and libs by name, title and grammar; it cannot see a header's prose, a data asset, or a rejection. This lib answers a rung's contract from the record itself, and nx_compare_rank prints the answer on every open rung so no seat starts a build from its own memory. THE INDEX (a derived artefact, stamped, refused when stale): one row per organ source in the two runtime roots org|<name>|<root>|<header words, lowercase, unique, first 12 bytes each>|<knowledge/ paths the header names>|<rejection lines> and one per data asset under knowledge/index: asset|<name>|<name words> headed by # NX-DERIVED incumbent index generated_at=<epoch> files=<n> head_truncated=<n> assets=<n> corpus_complete=<0|1> The HEADER of an organ is its leading comment block (the // lines before the first code line); its words, its basename parts and its column-0 func names are one bag. A header longer than the read window is COUNTED (head_truncated), never silently cut. A bag word is kept to 12 bytes for display; no key or match reads past 8. THE MATCH: the contract terms are the rung's title, symbol and note tokenised to lowercase runs of letters and digits at least min_tok long, deduplicated by their MATCH PREFIX (rerank and reranker are one term, beat and beats are one term); an all-digit term is DROPPED and counted (a number is a measurement, never a capability word: measured 2026-09-14, the 321 in "321 permille" matched three throwaway test organs at a rare-token weight). A term matches a bag word when their leading bytes agree -- the first morph_long bytes for a term at least that long, else the WHOLE term (mine~miner is the prefix rule nx_capsearch measured on its gold set at 4 bytes; here the 4-byte key only narrows the posting chain, because measured on 2026-09-14 the 4-byte cut made dense match density and wired match wire). Each matched term weighs idf_scale / (idf_base + df), df being the number of organs carrying it (the weighting nx_capsearch ships), and a term carried by more than df_max_permil of the organs is DROPPED as non-discriminating and COUNTED, so no stop-word list is hand-written: the corpus decides which words say nothing. An organ's score is the sum over its distinct matched terms, and a term that also hits the organ's own NAME parts adds its idf a second time: the name is the organ's own declaration of what it is, a second piece of evidence weighed like the first (the rule nx_capsearch's known-good tooltop applies, a stem hit on the name counted beside a hit in the title). MEASURED 2026-09-14 on the live tree (20,820 organs): ranking by a name TIER first put an organ named for two process words (wired, exceed) second for R0 at a tiny weight, so the name hit is weighed, never sorted on. The rung's own organ is excluded. THE LIST IS THE ANSWER, NEVER A THRESHOLD: the top topn organs print with their name hits, score and matched terms, then the data assets those headers name plus assets whose own name matches a term, then the rejection lines those headers carry. PRIOR-REJECTIONS is its own axis: beyond the top organs, every organ whose NAME carries a contract term and whose header records a rejection is listed (capped by rej_orgs, overflow counted), because a rejection on an organ named for the contract's own word is the record answering the contract -- that is how R0