I think if you had Claude Fable build a lightweight terminal like Ghostty, you'd get a good one.
An LLM may be fine to implement quickly an application that more or less works, but I have not seen any evidence that any of them is able to write well optimized code, whenever the application is not just a duplicate of something older.
Ghosstty and a few other newer terminal emulators are not only lightweight, but they also implement very fast GPU rendering, which is one of their important advantages.
So not only their set of features matters, but also the achieved performance.
Of course, one could use an LLM to write a very fast terminal emulator, but for that a competent programmer must be included in the feedback loop, which would require many iterations and running benchmarks and emulation conformity tests. It cannot be done by vibe coding with a prompt like "write a lightweight terminal emulator".
I've also given various models access to the tools necessary to measure performance and instructions to make things faster and they do pretty well. Hill climbing on a validatable measure is reasonably easy these days.
My direct experience here directly contradicts pretty much everything you've said can't be done. I.e. the human out of the loop stuff.
However, it is more focused on multiplexing.