I suspect that one of the keys to user-friendly structured data will be storing sparse data, in the sense of incompletely defined data (note that's real persistence, not just existing in the editor). It's also important to allow data that doesn't (yet) conform to whatever schema is relevant, but I think sparse data will generalize to that case.
One of the undersold features of plain text is being able to exist in "invalid" states while it's being edited. Structured data UI needs to have a least that level of malleability to have a chance over the current paradigm.
Definitely don't do it that way, then. Give comments first class status.
Comments should actually be independent of the underlying structure they're documenting. I'd like them to be an independent layer of annotations that can be applied to any data.
One of the undersold features of plain text is being able to exist in "invalid" states while it's being edited. Structured data UI needs to have a least that level of malleability to have a chance over the current paradigm.