nx_queries_merge.nx

nx_queries_merge.nx -- REWRITE ENSEMBLE AS TERM UNION (search R0l+, 2026-09-16): join K BEIR query files row by row into ONE queries file whose text is the space-joined union of every file's text for that query id. WHY. Every reasoning rewrite of a query is one sample of the vocabulary the answer lives in; two rewrites of the same query from different engines or seeds cover different terms. With the BM25Q query-term saturation already in nx_beir_eval (S5), the union of K rewrites scores as one long query whose repeated terms saturate instead of piling up, so a K-rewrite ensemble needs NO evaluator change: it is a queries file. This organ makes that file, and nothing else (rank fusion of the K rankings is the next arm, and it lives in the evaluator when it comes). CONTRACT. Every input carries the same query ids in the same order (the rewrite driver preserves the split's order), and the organ REFUSES BY NAME on the first row whose id differs or on a row-count mismatch -- a silently misaligned merge would score one query's rewrite against another's judgments and look like a real number. The output is truncate-written to <out>.tmp and renamed, so a killed run leaves no half file. nx_queries_merge <out.tsv> <in1.tsv> <in2.tsv> [<in3.tsv> ...] exit 0 MERGED / 2 usage / 3 REFUSED (mismatch named) / 4 io license_tier: ORIGINAL No hw writes (Rule 26).