There is no "entire corpus of StackOverflow-esque fix suggestions" about anything which is newer than a few years. I'm using cutting edge Android frameworks all the time. Yet, LLMs fix problems even when Google/Kagi has zero answers, which happens more often than not. We are way over this requirement.
I especially found that there is no difference between languages based on that. All generated code's architecture is terrible, if you don't actively manually maintain them all the time. If you don't have a few 10s of thousands of finely architected code already in your codebase, from which they can understand how it should be really done. And the reason, I think, is quite simple: the average code on the internet - regardless of market penetration of the given language - is simply bad.
Well, for the time since then, the LLM providers have a corpus of every code suggestion they made to users and whether it resulted in positive or negative sentiment afterwards - which is arguably even more powerful. There's still some level of RLHF that's more prominent for popular languages.
As you noted, of course, this doesn't apply to architecture. But that's also why I try to make sessions as turn-efficient as possible - you need every bit of context to get it to solve its own architectural rabbit holes.
I especially found that there is no difference between languages based on that. All generated code's architecture is terrible, if you don't actively manually maintain them all the time. If you don't have a few 10s of thousands of finely architected code already in your codebase, from which they can understand how it should be really done. And the reason, I think, is quite simple: the average code on the internet - regardless of market penetration of the given language - is simply bad.