_gpu_compute_gate.nx

_gpu_compute_gate.nx -- GPU-COMPUTE bit-identical CPU-vs-GPU gate (V-RAM-005 / X-DRV-003, the CUDA-moat-escape thesis in miniature on the sovereign GPU-class device). Proves: (1) GPU-COMPUTE -- the CANONICAL op-list emitter nx_drv_proto_emit, driven by gpu_compute_virt.spec alone (zero emitter change), authors a SPIR-V-gated DOT8 compute driver; the sovereign rv64 emu (+ rv64im_min_gpu) runs it: enable -> bind ring -> lay a DOT8 packet (SPIR-V magic + two int8 vectors a=[3,2,5,1] b=[4,6,1,2]) -> doorbell -> the device validates the SPIR-V magic, computes dot=sum(a[i]*b[i]) and latches RESULTPEEK -> the driver reads it back == 0x1f. Serial CONTAINS golden "GPUENRINGCMDDOT" + a clean halt. (2) BIT-IDENTICAL -- the gate computes the SAME int8 dot product on the CPU (independent loop) and asserts it == the value the GPU round-trip verified (0x1f=31). CPU lane == GPU lane, MEASURED. (3) INPUT-TAMPER -- bump a[0] (3->4) in the spec, re-emit, re-run: the device computes a DIFFERENT dot (35) so the driver's RESULTPEEK read-back verify fails -> the golden drops. This proves the device REALLY computes from the input data (not a stuck constant) -- the no-false-green handle. (4) DISTINCT -- the SAME emitter binary, driven by the echo GPU op-list, yields a different golden. Every verdict is a PRINTED marker off a REAL sovereign run; raw bytes read SOVEREIGNLY (sys_read). NO qemu/gcc/openssl/python. Marker -> knowledge/status/driver_spec.log (GPUCOMPUTEGATE row). Sovereign orchestration (fork/dup3/execve/wait4). license_tier: ORIGINAL