Bases: OmniARAsyncScheduler
VoxCPM2 scheduler variant for full unified decode graph serving.
VoxCPM2's full unified decode graph only applies to pure decode batches. When a decode-ready request is already running, this scheduler defers new waiting admissions for the current tick so the decode batch can stay on the VoxCPM2 graph path. This is a model-local serving policy, not a generic AR scheduler rule.