_gpu_dxg_vidmem_gate.nx

_gpu_dxg_vidmem_gate.nx -- SOVEREIGN-GPU ladder R4b (read the REAL VRAM budget of the 5080). R4a created a GPU context. R4b reads the discrete adapter's REAL video-memory budget via LX_DXQUERYVIDEOMEMORYINFO -- an UNFAKEABLE number (~14.9 GiB on the 16GB 5080) proving we read actual GPU state off the silicon, not just opaque handle slots. This is the memory-sizing prerequisite for real GPU allocations (R4c) and the GEMM working set (R6). ABI (recon-pinned LIVE, _gpu_dxg_r4b_recon: no uapi header -> size-swept, only _IOC_SIZE=56 recognized; OUT field offsets located by buffer dump): LX_DXQUERYVIDEOMEMORYINFO = _IOWR(0x47, 0x0a, 56) = 0xC038470A struct (56B): process @0 (u64, 0=current, IN); adapter @8 (d3dkmthandle, IN); memory_segment_group @12 (u32, IN: 0=LOCAL/VRAM, 1=NON_LOCAL/shared-sysmem); budget @16 (u64, OUT); current_usage @24 (OUT); current_reservation @32 (OUT); available_for_reservation @40 (OUT); physical_adapter_index @48 (IN). GREEN iff (author=organ, from REAL device returns): A) enum ret==0 AND exactly ONE discrete hw adapter (R1 identity); B) query(discrete, LOCAL) ret==0 AND budget in [4 GiB, 32 GiB] (a real discrete-GPU-class VRAM budget, NOT hardcoded to the exact byte count -- bounds, so it stays valid across host-reservation drift); C) query(discrete, NON_LOCAL) ret==0 AND its budget != the LOCAL budget (DISCRIMINATION: the segment_group field is genuinely honored -> not a stuck constant); D) TAMPER: T1 bogus adapter(0xdeadbeef) -> ret!=0 AND budget stays 0 (no data w/o a real adapter); T2 wrong _IOC_SIZE(48) -> -ENOTTY; T3 on a non-dxg fd -> -ENOTTY. Marker -> knowledge/status/gpu_dxg.log (DXGVRAMGATE). raw syscalls only. NO-WAVE: reads memory STATE (a budget), NOT throughput -- zero speedup/exceed claimed (that is R7). license_tier: ORIGINAL