I do understand the contention is that an LLM would be less thoughtful in editorializing which bits to make interactive, reasoning about the progression in understanding and delight by the user.

I'm not so sure it's that far out of reach, though. From what I've seen the reasoning models do, they're not too far away from being able to run a strategy of figuring out interesting increments of a problem, parameterizing them, making an interactive scene for those parameters, ... it feels within reach.