This seems less like a failure of structured outputs and more like expecting LLMs to behave like deterministic parsers — or am I missing something?