I agree that drastically subsetting the capabilities of the HTTP web (content negotiation, for example, or advanced HTML tactics) leaves a ton of interop and backwards compatibility on the table. It’s kinda my biggest regret in my own work to date: atp is kind of a private web or splinterweb, in w3c terms.
To your point about lexica being a hollow shell without some kind of micro reference implementation, I would point to the the tiles work, specifically lexicon-specific intents, which is admittedly early and more of a glint in the eye of a plan for a plan than a solution to what you’re describing, but hey, maybe something you would want to track or contribute to. It addresses your usecase but not your philosophical point, perhaps, because it is a splinterweb/private-web solution to the kind of openworld semantics issue that remains difficult to solve without, idunno, the Semantic Web?