Code quality and architecture still matter, because they also make it easier for LLMs to reason about the system.
That said, Opus 4.8 and Codex 5.5 both can write code that is higher quality than your average engineer. They are not quite there yet in terms of code re-use, but I think that's a solvable problem.
Running a couple of "scan for potential refactors"/"any duplicated code" prompt threads is already a long way there.
Regarding code quality, the largest issue I've run into is pollution that stems from committing too much unfiltered LLM code. They introduce some type of structures into the codebase that are hard to read for a human, then start reusing them or use them as example to create new ones, then when a human needs to quickly hop in and make changes, it's not as easy to do.