nx_recall_dense.nx

nx_recall_dense.nx -- the SOVEREIGN dense/semantic reranker over the REAL shipped model. Loads knowledge/index/semppmi_v1.bin (41MB, 102318-word PPMI, built by nx_semppmi_build, probes GREEN) and reranks a candidate shortlist by INTEGER LATE-INTERACTION maxsim soft- alignment: score(query,doc) = SUM over query terms q of MAX over doc terms w of cos_ppmi(q,w) (exact match = identity 1000). This is the ColBERT-class late-interaction pattern in pure fixed-point -- the R2 semantic reranker made REAL (replaces the fixture embeddings the nx_recall_rerank gate used with the actual 102K-vocab count model). EVIDENCE-FORK (no navel-gazing): a CONTROLLED keyword-tie fixture -- each candidate holds EXACTLY ONE query keyword, so first-stage lexical retrieval TIES all four (expected P@k = R/N, a coin-flip). Each relevant doc's OTHER term is a MEASURED synonym of the missing query term (won~defeated=86, city~stadium=62); each distractor's other term is MEASURED-unrelated (won~purple=25, city~january=35). The dense reranker must recover the relevant docs. Every deciding cosine is printed in the JSON = fully auditable, not asserted. RED + exit1 if the dense P@2 fails to beat the lexical coin-flip (an inversion cannot hide). This is a controlled UNIT proof of the mechanism on the real model; corpus-scale nDCG is the flagged BEIR row (F236), not claimed here. expect_exit: 0 license_tier: ORIGINAL