I am curious as to why. They should be represented like joins, right?

They’re slow as fuck and most of them can be rewritten to not be correlated, so we attempt to heavily rewrite them. There’s basically an entire subfield of cs research that focuses on strategies for rewriting CSQs and the papers have gotten progressively more and more complicated making them very difficult to reproduce without bugs. I’ve spent easily weeks just reading these papers let alone implementing them.

I'd love to read a deep dive on this. Database implementation internals are always very interesting.