nx_f32_ffn_path_gate.nx

nx_f32_ffn_path_gate.nx -- ISOLATED, host-noise-immune proof that the F32 packed range-dot cached path beats the scalar threaded matmul_t at the real FFN decode shape (m=1, k=896, n=4864 = Qwen W_gate). Both paths run in the SAME loaded conditions, so the RATIO holds even when the host is noisy (the trap that contaminated the full-forward profiler). Warms the cache first so the timed cached calls are pure SIMD dots (the fill is one-time, amortized across all tokens in the real forward). Exact-int regime (dense small-int f32) so bit-exact compare across the two different accumulation orders (scalar sequential vs 8-wide) is legitimate. Checks: 1 cached (range-dot) result == scalar matmul_t, bit-exact 2 cached path >= floor x faster than scalar over N reps (both timed threaded; ratio printed) lineage_id: f32_ffn_path_gate_v1