Dharmesh Joshi

Thanks Ahmet, that's a fair catch. The snippet in the post is deliberately simplified: it fixes the hardcoded-model problem but, as you say, _loaded_pipelines only ever grows, so an episode with several overrides can bring the same memory pressure back. An explicit residency budget with eviction of inactive pipelines is the natural next step, and the lease idea matters because a pipeline shouldn't be unloaded halfway through a scene. The three-checkpoint alternating test, recording resident memory after each switch, is a great regression case too. I'll fold that into a follow-up. Appreciate the detailed review.