_gpu_dxg_r3_recon.nx

_gpu_dxg_r3_recon.nx -- SOVEREIGN-GPU ladder R3 RECON (create a WDDM device on the RTX 5080). R2 opened a per-adapter kernel handle from the discrete LUID via LX_DXOPENADAPTERFROMLUID. R3 feeds that adapter handle to LX_DXCREATEDEVICE (_IOWR(0x47,0x02,...)) to obtain a real WDDM DEVICE handle -- the object that GPU command submission (R4+) hangs off of. The WSL2 kernel ships no uapi header, so this RECON discovers the ABI live (rule: trust runtime): (1) SIZE: _IOC_SIZE is encoded in the ioctl code. dxgkrnl dispatches on _IOC_NR=0x02 but rejects a wrong _IOC_SIZE (-EINVAL) vs an unknown nr (-ENOTTY). So we sweep candidate sizes for nr=0x02 with the real adapter handle at offset 0 and watch for ret==0 -> that size is the real sizeof. (2) DEVICE-HANDLE OFFSET: on the successful call, dxgkrnl writes the OUT device handle back into the struct. dxgkrnl hands out handles in the 0x4000xxxx space (R0/R2 adapter handles were 0x40000000/0x80/0xc0), so we scan the returned buffer for a fresh 0x4000xxxx value != the adapter handle -> that 4-byte slot is `device`. Prints size + offset + value so the gate can pin them. raw syscalls only (no libvulkan/cuda/dxcore/libc). Writes NOTHING to the status log. license_tier: ORIGINAL