_gpu_dxg_alloc_gate.nx

_gpu_dxg_alloc_gate.nx -- SOVEREIGN-GPU ladder R4d (allocate REAL GPU memory on the 5080, UMD-less). R4c built the paging queue. R4d allocates actual memory through LX_DXCREATEALLOCATION via the UMD-less STANDARD-ALLOCATION / EXISTINGHEAP path -- no vendor (NVIDIA) UMD blob needed. The recipe was reverse- engineered: fault-probes pinned device@0/alloc_count@44/allocation_info@48; the WSL2-Linux-Kernel header (WebFetch, verified vs runtime) gave the field layout; and dxgkio_create_allocation's open-source validation gave the exact requirements: flags@56 = standard_allocation(bit16,0x10000) | existing_sysmem(bit5,0x20) = 0x10020 priv_drv_data_size@40 = 0 ; alloc_count@44 = 1 standard_allocation@32 -> d3dkmt_createstandardallocation{ type@0=EXISTINGHEAP(1); existing_heap.size@8 (!=0, page-aligned) } allocation_info@48 -> d3dddi_allocationinfo2{ sysmem@8 (!=NULL, page-aligned); allocation@0 (OUT handle); gpu_virtual_address@40 (OUT) } GREEN iff (author=organ, from REAL device returns): A) device-create chain ok; B) CREATEALLOCATION #1 ret==0 AND allocation handle != 0 (written into pre-zeroed entry@0); C) CREATEALLOCATION #2 (distinct sysmem) ret==0 AND handle != 0 AND != #1 (per-call REAL allocation, not a constant); D) TAMPER matrix, each from a real validation rule, all != the real ret==0: T1 flags w/o existing_sysmem (0x10000) -> STATUS_INVALID_PARAMETER; T2 priv_drv_data_size=16 -> -EINVAL; T3 existing_heap.size unaligned (65535) -> -EINVAL; T4 on a non-dxg fd -> -ENOTTY. fd close frees the allocations (per-fd dxgkrnl cleanup). Marker -> knowledge/status/gpu_dxg.log (DXGALLOCGATE). raw syscalls only. NO-WAVE: allocates memory (a real GPU object); gpu_va mapping=R4e, compute/throughput=R7. ZERO throughput. license_tier: ORIGINAL