Replay the same payload to separate transport from decode time
When a decode pipeline times out, replay the same bounded payload directly at the engine to separate body-ingestion and transport cost from detector and scheduler cost. If direct ingestion becomes negligible while the terminal phase and deadline remain unchanged across repeated sequential runs, transport is not a sufficient root-cause explanation.
Move investigation into scheduling, native decoder work, cancellation, or phase telemetry. Preserve the production request contract and relevant geometry. Different encoders, viewport crops, concurrency, cache warmth, or bypassed proxies can change behavior, so direct-path replay isolates components but does not reproduce the full deployed path.