They do, but it has to check whether the photos are 'found' and I've seen that be v slow in the past, but again, I think it's unlikely to be the cause on local storage. I've also seen it with ultra-large smart collections.
As a great believer in data, I ran a process monitor and recorded every file access. This applies only to windows, but it does not.
The short answer is we are both right. Sort of.
I shut down lightroom first with the grid display up, a small folder with a half dozen or so images, in it in the grid display, and the folder panel shown. Started a monitor of every file accessed in my folders.
It accessed the folder directory (itself) of each folder visible, it did not access any folders that were not visible (e.g. I have folders by year, and 2017 was expanded -- it showed each year once, but no dates within any year other than 2017, where it showed one access for each folder (day) inside of it).
It did not access any images inside of any folder except the one shown in the grid display. At all, not even an existence check. So very large numbers of images out there cause extremely little IO, they heavy hit is all done against the catalog. A large number of folders visible in the folder panel causes some IO, but in the grand scheme of things not all that much as you cannot fit that many folders on one screen.
The images in the grid display were only accessed if visible. I tried this again with a folder with 200+ images, and it only touched the dozen or so shown on the screen when it opened the grid. So the total images within a folder have little effect on open speed (assuming previews already built, I do not know if it builds only visible or all).
But... it is clearly using multiple queries against SQL for the folder counts, and it clearly takes quite a while. I'm not at all sure why it does not do this as one big query, but this means there is an additional reason for having the catalog on fast disks, and with respect to startup, having it minimally expanded.
Again, windows, and LR CC 2015.10. Though I would heavily doubt that Mac would differ in this regard, that would imply a level of functional difference that would be surprising.
Incidentally, this was done with Process Monitor from SysInternals (now a Microsoft company). It's free and if you are a bit technically savvy you can learn a whole lot about how programs behave. Since it can measure ALL access including its own, it is real easy to have it run amok and lock up a system, so use it carefully with good filters. But it's really helpful. I don't know if there's a similar program on Mac.