Fix a broken AR video pipeline

The supplied pipeline performs fast autoregressive video generation using an ODE-distilled checkpoint and a DMD checkpoint trained on top of it.

For a prompt describing a black Porsche racing forward, generated videos consistently show the car moving backward. The semantic and physical failure reproduces across prompts and seeds, which points to a systemic pipeline issue rather than an isolated bad sample.

Identify the root cause, propose a concrete fix, and implement it. Training from scratch is not required.