nx_semppmi_build.nx

nx_semppmi_build.nx -- EXPLICIT-VOCAB SPARSE PPMI embeddings = the CANONICAL count-based method at full strength (Levy&Goldberg: full sparse PPMI rows rival trained embeddings on word similarity). This is the last count-method rung before "trained-only" is airtight: v1 (nx_semcorpus_build) failed its probes with HASHED rows (collisions) + a 256-dim PROJECTION (lossy) -- this build removes BOTH: every word gets its own row; contexts are the full vocabulary; cells are PPMI (positive pointwise mutual information, log2 x16 fixed-point); rows stay SPARSE (sorted triples). Corpus = same banked benchmark prose (no gold leakage: only JSON strings >= 48 chars). PIPELINE (one corpus walk): tokens -> 61-bit djb2x hash stream (+ -1 markers at string ends) -> vocab = sort(copy)+unique (explicit; 61-bit space => collisions ~0 at 60k words) -> window +-4 pair emission as packed keys (aid<<20|bid, both directions) -> sort -> RLE -> triples(cnt>=2) -> marginals -> PPMI16 -> row index + norms -> PROBES -> persist knowledge/index/semppmi_v1.bin. PROBES decide (same discipline as v1): if won~defeated fails to order again, count methods are DEAD at canonical strength and the R1 trained thread is the proven only door. TEETH structural + neg-controls. expect_exit: 0 license_tier: ORIGINAL