nx_semcorpus_build.nx

nx_semcorpus_build.nx -- R3b-scale CORPUS-LEVEL distributional embeddings (the rung after 3 measured-negative per-question climbs; see nx_qabench MODE2 header). Builds ONE persisted hashed count-embedding matrix from ALL banked benchmark context prose (~4.5MB Wikipedia-register: qab_squad_big + qab_hotpot_big + qab_musique_big), then nx_qabench's MODE2 loads it and selects sentences by corpus-informed cosine. METHOD (sovereign, integer, no training loop -- Random-Indexing/feature-hashing lineage): text = every JSON string value >= 48 chars (paragraphs+questions; SHORT strings -- keys, titles, GOLD ANSWERS -- are structurally excluded = no label leakage; corpus is the inference-visible context pool, unsupervised). token row = djb2(token) % 16384; for each content token, each stop-filtered context token within +-4 hashes to one of 256 dims: M[row*256 + dim] += 1. Persisted -> knowledge/index/semcorpus_v1.bin (NXSEMC1 header). HONEST CAVEATS baked in: hash collisions (~2-4 words/row) + raw counts (no PPMI) = v1; probes below REPORT whether corpus stats order won~defeated > won~purple (the exact wall) -- reported, not asserted. TEETH: T1 strings>=3000 T2 token-events>=200000 T3 nonzero cells>=50000 T4 write+reload spot-check T5 NEG: identity cos=1000 and a gibberish token's empty row scores 0 vs everything. expect_exit: 0 license_tier: ORIGINAL