Structured metadata like this is for the non-LLM tooling layer, primarily your harness, so (eg.) it can interpret multiple formats without needing to burn tokens every time.