Couple of aspects you are not considering.

1. If majority of developers use LLMs for coding, and then if someone developed a better React, then no one can really use it because the models does not know how to write idiomatic code in it. It might be easier to make the LLM implement workaround for the issue with the current React, than making it use an unfamiliar library or paradigm.

3. Without some critical adoption rate, the new library won't have enough training material, and this becomes a chicken and egg problem. This can only be broken by AI companies by somehow training the models on the new React by code explicitly written, but again, without critical adoption, there is no incentive for them to do it.

Imagine we are living in the time of jquery and all the LLMs are proficient in it. I don't think in such a world, no one would want Angular/React, because LLMs will happily handle all the complexity that using jquery results in.

We would have stuck with JQuery forever..

New libraries can be used with LLMs directly without training using skills. If you work in a company with bespoke frameworks, you can see this happening. This + RL post training would be the new cycle of adoption.