Schema for capturing full LLM request/response payloads per workflow step.

These rows can be large (10–50KB each) and are only for debugging and inspection. Prune aggressively in production — see Debug.prune_older_than/1.