nx_nofloat_w4_gemm_gate.nx

nx_nofloat_w4_gemm_gate.nx -- THE DECISIVE W4A16 EXPERIMENT (de-risks the inference-speed arc before any core-engine change). Decode is memory-bound (baseline: i8 3x faster than i32 = 4x fewer weight bytes). W4A16 stores weights as 4-bit (0.5 B/weight = HALF of i8) -> less memory read, but needs extra 4-bit unpack. Q: does the halved weight-read beat the unpack cost on the REAL decode shape (matvec: 1 x K activation * N x K weights)? This gate builds both paths (i8-weight and 4-bit-weight, SAME __i16x16_madd inner product), checks correctness (both approximate the int16 reference within quant error), and TIMES them. Verdict decides whether to build the full W4 kernel (and whether it needs SIMD-unpack primitives). Isolated -- touches NO serve code. license_tier: ORIGINAL expect_exit: 0