nx_nofloat_w12a12_gate.nx

nx_nofloat_w12a12_gate.nx -- the referee for the W12A12 decode matmul (search R0l attribution, 2026-09-16). SUBJECT: mm_pool_i8 in nx_nofloat_llm (the pooled SIMD matmul every decode projection goes through), driven IN-PROCESS on synthetic data whose shape is the real one (in_dim 8960 = the Qwen2.5-1.5B ffn_down, the widest projection, so the chunk bound is exercised at its limit) with activation OUTLIER channels planted, because a per-tensor scale is exactly the quantiser an outlier defeats. TEETH: the chunk bound is arithmetic (NF_CHUNK_K/8 * QMAX * QMAX <= i32 max) and 2048 is the planted bound- breaker; the kernel's max relative error against the exact i64 dot is BELOW the old 8-bit per-tensor quantiser's error on the same data (the old quantiser is re-implemented here as the CONTROL, so the improvement is measured, never asserted); the kernel is exact on a row the quantisers cannot lose (all values on the grid); and the chunked fold reproduces the unchunked sum on a short row (K < NF_CHUNK_K, one chunk). license_tier: ORIGINAL No hw writes (Rule 26). expect_exit: 0