@Dan_Stillman I think that process makes sense, and I think experimenting with anything is a good idea.
But it might also be worth considering integrating citeproc-js (or a rewritten or refactored typescript alternative as mentioned here) into the process?
When a number of us were working on enhancements during the COVID shutdown, the main issue that held us back in my view was what had previously been an advantage: the super loose coupling of schema and code. We never really had a sense of if or when anyone would implement any changes we were making, so we never felt comfortable cutting a release.
And, I think, we were also held back by the lack of compatibility rules, per one of your other points; what must CSL processors do with unknown attributes and elements. I agree you need that.
It’s one of the reason I decided to take the complete opposite approach in Citum; where everything is defined in and generated from (Rust) code: schemas, docs, etc. Styles are developed alongside that code in the same repo, and the most important ones embedded in the engine, and styles can inherit.