From previous experience we know that LLMs are great at confidently producing plausible output with problems that are hard to detect and fix without expert level knowledge and experience.

What experiences exactly? Whether the output is correct or just plausible and how many and how severe the problems are depends on the actual problem, and this statement is way too broad. There are also enough experiences where it works. It depends on how you use it.

I don't see too much of a difference between a LLM producing plausible output as the result of reverse engineering and a human engineer producing plausible output as the result of reverse engineering. Again, we're not talking about absolute experts doing this for the past 20+ years spending months to do this. For a "regular" engineer doing this, it's the same best guess.