[Tokha233]

Decode and encode a single H3 MP4 concurrently through an optional Decode and Save Video node. Keep the R85 VAE/RGB8 arithmetic and codec contract: two pinned D2H slots feed a request-owned copy worker and a bounded CPU encoder. Existing Decode/Save nodes remain available.

On RTX 5090 D v2, two saved R85 latents measured 11.509 → 9.311 s and 12.474 → 9.965 s from video decode to completed MP4 (BAAB/ABBA, four formal repeats per arm/case). Every RGB, PCM and MP4 hash matches the historical baseline. These are output-stage latency improvements. No DiT kernel or precision changes.

The October 8 full-pipeline check adds fixed conditioning → DiT → VAE → completed local MP4 for two representative 10/11-second clips. Same-GPU subprocess-export controls give 461.182 → 457.122 s and 461.942 → 457.664 s in opposite orders. An additional resident in-process encoder control gives 461.569 → 458.662 s (2.907 s saved; +0.634% inverse two-request queue duration). Across six runs, all 12 formal requests and six warmups retain the historical latent/RGB/PCM/MP4 hashes. GPU completion spans are essentially unchanged: the benefit is predominantly the final file tail. No sustained serving-QPS or P95 improvement is claimed. The resident-codec result is one same-card pair, not a long stress test; reference/text encoding, network and cold checkpoint loading are excluded.

Validation: 16 CPU/codec tests on the CUDA host; five CUDA chunk layouts, two concurrent sinks, callback failure and cleanup; full helper and actual Comfy output node produce the historical MP4; interrupt removes partial output and joins workers. Includes an opt-in public benchmark CLI, source hashes and raw records. A stalled native codec/driver still requires external process supervision; no hard cancellation deadline is claimed.

Inspired by SGLang #41819. This integration keeps SpeedKit’s PyAV codec contract and moves pinned-slot draining off the GPU submission thread. It is a SpeedKit feature, not a duplicate SGLang/Kitchen submission.

See docs/R85-STREAM-EXPORT.zh-CN.md, evidence/r85-stream-1002/, and evidence/r85-pipeline-1008/ for protocol, limitations and evidence.