_gpu_dxg_context_gate.nx

_gpu_dxg_context_gate.nx -- SOVEREIGN-GPU ladder R4a (create a GPU CONTEXT on the 5080 WDDM device). R3 made a real WDDM device. The command-submission arc (R4) begins with a CONTEXT -- the queue GPU work is submitted on. The WSL2 paravirt path exposes LX_DXCREATECONTEXTVIRTUAL (nr=0x04), NOT plain CREATECONTEXT (nr=0x03 -> -ENOTTY on this kernel, recon-confirmed). This rung opens the context only; allocations (R4b) + command buffer/submit/fence (R4c) follow. ABI (recon-pinned LIVE, _gpu_dxg_r4_recon: no uapi header -> size-swept, only _IOC_SIZE=40 recognized; OUT context handle located by write-back scan): LX_DXCREATECONTEXTVIRTUAL = _IOWR(0x47, 0x04, 40) = 0xC0284704 struct d3dkmt_createcontextvirtual (40B): device @0 (in); node_ordinal @4; engine_affinity @8; client_hint @12; priv_drv_data @16 (u64 ptr, NULL ok); priv_drv_data_size @24; **context @32 (OUT)**. GREEN iff (author=organ, from REAL device returns): A) device-create chain ok (enum/discrete/open-adapter/create-device all ret==0); B) CREATECONTEXTVIRTUAL #1 ret==0 AND context1 != 0 (written into a pre-zeroed slot @32); C) CREATECONTEXTVIRTUAL #2 ret==0 AND context2 != 0 AND context2 != context1 (per-call allocation); D) TAMPER: T1 bogus device(0xdeadbeef) -> ret!=0 AND no context written; T2 wrong _IOC_SIZE(32) -> -ENOTTY; T3 on a non-dxg fd -> -ENOTTY. (Context handles share the per-process namespace with devices -> we never test cross-namespace numeric inequality; only write-back + per-call distinctness + tamper.) fd close releases all handles. Marker -> knowledge/status/gpu_dxg.log (DXGCTXGATE). raw syscalls only. NO-WAVE: context-create only, ZERO throughput. license_tier: ORIGINAL