The escalation hands over the crime scene

The small model handles the refactor until it doesn't: three attempts, same failing test. Most escalation paths throw the transcript at a bigger model and let it start over from a description.

This one ships the state. At the stuck point, the orchestrator snapshots the container: half-finished refactor, failing test, every intermediate file exactly as attempt three left it. Then it dispatches a new headless run on the same container with a stronger model in the model field, and the new run opens its eyes inside the machine where the problem actually lives.

If the strong model also fails, the snapshot is still there for a human, who inherits the identical scene: nothing paraphrased, nothing lost in the retelling.

The escalation doesn't retell the problem: it hands over the crime scene.