_gpu_matmul_gate.nx

_gpu_matmul_gate.nx -- GPU 2x2 int8 MATMUL bit-identical CPU-vs-GPU gate (V-RAM-005 / X-DRV-003, the CUDA-moat-escape matmul primitive on the sovereign GPU-class device). Proves: (1) GPU-MATMUL -- the canonical op-list emitter nx_drv_proto_emit, driven by gpu_matmul_virt.spec alone, authors a SPIR-V-gated MATMUL2 driver; the sovereign emu (+ rv64im_min_gpu) runs it: the device computes C=A*B for A=[[1,2],[3,4]] B=[[5,6],[7,8]] -> C=[[19,22],[43,50]], packs the four int8 outputs into RESULTPEEK=0x322b1613, the driver reads it back. Golden "GPUENRINGCMDMM". (2) BIT-IDENTICAL -- the gate computes the SAME 2x2 matmul on the CPU (independent), packs it, and asserts == the value the GPU round-trip verified. CPU lane == GPU lane, MEASURED. (3) INPUT-TAMPER -- bump A00 (1->2) in the spec, re-emit, re-run: the device computes a DIFFERENT C so the RESULTPEEK read-back verify fails -> golden drops (real matmul, not a stuck constant). (4) DISTINCT -- the SAME emitter binary on the DOT8 op-list yields a different golden. HONEST SCOPE: this proves CORRECTNESS (bit-identical) of a small int8 matmul tile on the sovereign emu. The throughput EXCEED vs cuBLAS is a SEPARATE real-RTX-5080 measurement, NOT claimed here. Verdict from a PRINTED marker off a REAL sovereign run; raw bytes read SOVEREIGNLY. NO qemu/gcc. Marker -> knowledge/status/driver_spec.log (GPUMATMULGATE row). license_tier: ORIGINAL