_gpu_bm_doorbell_gate.nx

_gpu_bm_doorbell_gate.nx -- BARE-METAL SOVEREIGN-GPU rung BM-GPU-5: the FULL doorbell->execute path. Composes BM-GPU-4 (GPFIFO channel + entry) + BM-GPU-1 (pushbuffer execution): the driver writes pushbuffers + GPFIFO entries, advances GP_PUT, and RINGS THE DOORBELL; the modeled host FIFO then walks the ring GP_GET..GP_PUT, decodes each entry, fetches+executes each pushbuffer, and the semaphores land in GPU memory. THIS IS THE EXACT SOVEREIGN SUBMIT->EXECUTE LOOP WSL'S HOST-TRANSLATION BLOCKED -- closed end-to-end in a spec-faithful model. NISHI-ECOSYSTEM-ONLY: our channel + our ring + our pushbuffer + our FIFO model. HONEST SCOPE: full path LOGIC vs spec; real doorbell MMIO + real channel in VRAM = BM-GPU-6 on native Linux. GREEN iff: before the doorbell both semaphores read 0; after the doorbell BOTH land their distinct payloads and GP_GET advances to GP_PUT (FIFO consumed the ring); AND tampers (no doorbell => nothing executes; corrupt pushbuffer => its semaphore doesn't land) hold. Marker -> knowledge/status/gpu_baremetal.log (BMDOORBELLGATE). license_tier: ORIGINAL