nx_gpu_mem.nx

nx_gpu_mem.nx -- HEAPS AND RESIDENCY over the dxg door (gpu GP6): allocate, place, make resident, EVICT, with the budget READ from the device, never assumed. One heap = one EXISTINGHEAP allocation of a page-aligned size, placed at a GPU virtual address through the paging queue and made resident; eviction is the device's own EVICT, and the proof that it happened is the device's own paging state (a make-resident after an evict answers DXG_MKRES_PENDING instead of 0). SCOPE, stated once: these heaps are SYSTEM-memory backed (the UMD-less EXISTINGHEAP path R4d measured), so they live in the NON-LOCAL segment; a VRAM (local-segment) heap needs the vendor UMD's private allocation data, the same blocker class as the command ring (gpu.plan GP5 lesson). Exhausting the non-local budget would exhaust the host's RAM, so this lib walks a BOUNDED window under the budget and never crosses it; the exhaust-to-evict test on the local segment is named on the board as owed, not claimed. license_tier: ORIGINAL No hw writes (Rule 26): per-process allocations, released with the process.