nx_q5_0_threaded_gate.nx

nx_q5_0_threaded_gate.nx -- THE DECISIVE test: does THREADING flip the fused Q5_0 dequant-dot from a loss to a win? Single-thread (nx_q5_0_fused_dot_gate) the fused kernel LOST 2x: compute- bound (dequant overhead > MAC), memory not saturated. But at 16 cores the F32 matmul saturates RAM bandwidth (16 x ~2GB/s > ~20-40GB/s), so reading 11.6x fewer bytes (Q5_0 3MB vs F32 34.9MB) could win THERE even with a scalar dequant. Both paths POOL-threaded, SSE dot (__f32_mul/__f32_add), same run => ratio is host-noise-tolerant. THIS decides whether quantized- read is viable on our CPU substrate or needs SIMD-dequant intrinsics first. NO forward changes. Checks: 1 Q5_0-fused-pool == F32-pool (mmt_range), bit-exact 2 threaded ratio: fused vs F32 at FFN shape (report; not a hard floor -- this is a DECISION measurement, GREEN prints the verdict either way) lineage_id: q5_0_threaded_gate_v1