That's a good question.
The abstraction I have in my head is that Doenet is a problem type that can be embedded in different systems, such as LMSs. The situation with the PreTeXt and Runestone is more confusing, so let me first just talk abstractly in terms of an LMS. Then maybe you can help be figure out how that translates into the PreTeXt and Runestone embedding. I'm hoping this isn't particular to Doenet.
When Doenet is embedded in an LMS, it relies on the LMS to tell it when to get a new attempt and what random variant to select. It doesn't expose any UI for choosing a new attempt or changing the random variant, as those decisions seem to belong in the domain of the LMS. (If a variant isn't provided, it happily chooses a random variant.) I would imagine that details around which variant to be used would be an instructor decision (such as whether all students get the same variant, how many attempts allowed).
I guess my question is whether this separation of concerns is maintained for a PreTeXt textbook. From your response, it sounds like it is not. I agree that it could get complex with many different types of interactives (though I don't think this a function of an interactive per se, but instead of a problem/exercise, which is what Doenet aspires to become). I'm wondering it is worth having some type of abstraction around this concept if it is a feature users want.
Duane