nx_f32_embed_gate.nx

nx_f32_embed_gate.nx -- KAT for the NEURAL EMBEDDING TAP (rung 1 of the retrieval ladder). WHY THIS GATE IS SHAPED THIS WAY. Our count-based PPMI reranker was measured on BEIR/nfcorpus at 255 permille against BM25's 305 -- a LOSS -- and the banked conclusion was: A GATE PROVING A MECHANISM ON A FIXTURE HAS NOT MEASURED THE MODEL. So this gate deliberately does NOT assert "an embedding came out non-zero"; that passes on noise. It asserts SEMANTIC ORDERING a count model cannot fake: the related pair shares almost NO surface tokens (cat/kitten, sat/rested, mat/rug), so a lexical or co-occurrence model has nothing to latch onto. Only learned semantics ranks it correctly. TEETH (both polarities -- a gate that can only pass is furniture): 1 NON-VACUITY : the embedding is not all-zero 2 SELF-SIM : cos(x,x) == 1000 permille -- the metric itself is sane 3 SEMANTIC : related pair scores ABOVE unrelated pair <- the real assertion 4 NEG-CONTROL : the unrelated pair must NOT also score near-1; if everything is similar to everything the space has COLLAPSED and tooth 3 passes for the wrong reason 5 DETERMINISM : same input twice = identical vector (integer f32, no hidden RNG) license_tier: ORIGINAL No hw writes (Rule 26). expect_exit: 0