Use generator pattern (cross-process yield via spaces res_queue) — ZeroGPU forks process so streamer queue isn't shared

Force streamer.end() in finally — transformers 5.x doesn't always auto-call it

Initial backend code: src/kpaa, runtime data, requirements