nx_gpu_dxg_probe.nx

nx_gpu_dxg_probe.nx -- SOVEREIGN-GPU foundation organ (R0): reach the REAL RTX 5080 through its actual device interface, /dev/dxg (the WSL dxgkrnl GPU-paravirt thunk), with RAW syscalls ONLY. NO libvulkan / libcuda / libdxcore / libd3d12 / libc -- the emitted elf links NOTHING. Every byte here is open(2) / ioctl(2) / close(2) / write(2) directly via __syscall. This is the productionized form of the recon _dxg_enum_probe: a reusable probe FUNCTION (dxg_enumadapters) that the gate (_gpu_dxg_gate.nx) calls, plus a standalone main() so the organ can be run on its own. It performs the MINIMAL real call -- LX_DXENUMADAPTERS2 -- and reads back the REAL adapter count + adapter[0] handle + LUID from the 5080. ABI (pinned from WSL2-Linux-Kernel linux-msft-wsl-6.6.y, include/uapi/misc/d3dkmthk.h; CONFIRMED live: the kernel accepted exactly 0xC0104714 and returned ret=0, num_adapters=2): #define LX_DXENUMADAPTERS2 _IOWR(0x47, 0x14, struct d3dkmt_enumadapters2) = (dir _IOWR=3 <<30) | (size 16 <<16) | (magic 0x47 <<8) | (nr 0x14) = 0xC0104714 struct d3dkmt_enumadapters2 { __u32 num_adapters; __u32 reserved; __u64 *adapters; } // 16B struct d3dkmt_adapterinfo { d3dkmthandle handle(4); winluid luid(8); u32 num_sources; u32 present_move_regions_preferred; } // 20B PROTOCOL: set num_adapters = capacity, adapters = ptr to an array of d3dkmt_adapterinfo; the kernel writes the REAL count back into num_adapters and fills the array. NO-FALSE-GREEN: the gate also issues a TAMPER (magic-0x99 ioctl) which the real device rejects with a real errno (-ENOTTY=-25) != the real ret=0; that distinguishes a real device response from a no-op. (The gate owns the verdict; this organ exposes the raw mechanism + a result struct it can inspect.) SCOPE (NO-WAVE): this rung proves ONLY "Nishi reaches the real RTX 5080 sovereignly via the /dev/dxg last-mile". It claims NO throughput / speedup / exceed -- that is R7 (a real cuBLAS head-to-head, many rungs up). Do NOT read a perf number out of this. license_tier: ORIGINAL