nx_recall_rerank_gate.nx

nx_recall_rerank_gate.nx -- KAT + HONEST BENCH for the sovereign semantic reranker. exit 0 = pass; N = assertion N failed. Sovereign lane (nx_cc->nxasm). Reproduces the live failure: a relevant doc buried by lexical/fusion ranking but semantically parallel to the query -- the sovereign cosine reranker recovers it to rank 1. Kept HONEST: an irrelevant doc with a HIGH cosine (a semantic false- positive) is included, so the reranked nDCG lands ~0.96, NOT a rigged 1.0. The point is the measured LIFT over fusion-alone on the R0 ruler, not perfection.