nx_dr_semcorpus.nx

nx_dr_semcorpus.nx -- BATCH semantic judge (the at-scale enabler; eats the model-reload debt). nx_dr_semjudge_cli reloads the 41MB PPMI model per PROCESS: a 16-doc x 81-insight run reloaded it 16x = ~21 min, ALL of it model-load. This loads the model ONCE and scores MANY doc FILES against a SHARED insight set, so a full-corpus judge is no longer reload-bound. Behaviour-preserving BY COMPOSITION: it calls the SAME sj_load/sj_tokenize_ids/sj_maxsim (=> nx_ppmi_lib) as nx_dr_semjudge, so per (doc,insight) semantic_permil is bit-identical to the per-call judge -- the gate proves that against the live incumbent, not by assertion. nx_dr_semcorpus <insights_file (one insight per line)> <doc_file_1> [doc_file_2 ...] -> {"tool":"nx_dr_semcorpus","vocab":V,"n_insights":K,"docs":[ {"doc":0,"readable":1,"tokens":N,"scores":[s0,..,sK-1]}, ...]} An unreadable doc reports readable:0 (fail-loud), never a silent all-zero row. Every func <=6 params (NAS nx_cc >6-arg skew, seq239). No hardware writes (Rule 26). license_tier: ORIGINAL expect_exit: 0 module: nishi-core.research.dr_semcorpus depends: nx_dr_semjudge.nx v3 (OO tranche 1): local sc_puts/sc_q/sc_num DELETED -> the canonical nx_json_emit lib. Migration is BYTE-EXACT gated (same inputs, identical output bytes) -- zero lost function.