I guess I'll put in a few more cents.
There is a lot of crappy RMGUIs out there, so I understand peoples frustration and need for something new.
But RMGUIs doesn't have to be that crappy, and IMGUIs, while solving some problems, overlook others.
There is no reason an RMGUI would have to have complicated observers or invalidation logic.
In the RMGUI I'm currently writing, if you change something, vertex buffers will be invalidated and rebuilt before the next draw.
If not, well, then we reuse the vertex buffers from last frame. (normal case for 99% of the frames)
Its not hard to know if a change will mean the vertex buffer it belongs to need update.
I don't care at all about regions. That was relevant in software rendering. Not so much in HW.
You do decouple creation from update/reaction, but you do it for a reason. It might not even be the same person defining it.
But even if it is, I like to have my widgets initial position and layout defined in an xml. Nice and neat and does not pollute my code with layout, and I can even edit it in runtime and reload it for quick ui tweaks. (or complete ui changes as long as the same actions exists)
Since I use lambdas, there is no problem with extreme decoupling, all my callbacks are defined right then and there, without unnecessary code.
Code should be concerned with actions, and never care about exact sizes or positions.
I don't want to rewrite logic for when and how to draw my buttons for every game I do.
It makes sense for game engines where you have complex needs for culling and handling of transparency and effects.
Not that much for UIs.
To have higher class objects like views and popups make code brief and easy to read and modify the relevant parts of.
In my RMGUI, my UI is just another layer/pass in our graphics engine, I call "update" on it when it should be updated, and I call "draw" on it when it should be drawn. (so in a way IM, but on a higher level)
But I wouldn't want to call the equivalent of "draw" for each and every ui item each frame.
Not because I think it would be slow, but because I have no need to see and change that code, it can be completely generic for every UI.
Then of course, RMGUIs can be made overly complex... like most out there.
But it doesn't have to be like that, and I'm not sure IMGUIs really solve the problem, just moves the responsibilty
I don't want to discourage anyone from using an IMGUI though, I'm sure it can be perfect for a lot of projects.
I'll treat each paragraph as a bullet point, all are addressed:
1) Every RMGUI i've worked on (in games or otherwise) get overly complicated. It's inevitable. The most important thing in game development is iteration time, IMGUI tend to be quicker for this.
2) You can rebuild vertex buffers on invalidation sure, you should do that whether RM or IM. The topic is really about "what it takes to get my game gui done", both options have similar optimisations. I'm not sure what you mean by regions.
3) There isn't much practical need to decouple display and logic. The games ive worked on have had heavy XML driven RMGUI front ends and the artists and designers barely go near it. A well thought out IMGUI interface can produce code that is as easy to read as XML. If artists want to change the size of something they'll just look for the name and numbers wherever theyre declared.
4) Not sure about this but in my experience having callbacks registered to events etc gets very messy and it hard to track what is listening to what.
5) Surely when you call draw on your RMGUI it is fast enough that it doesn't hiccup the framerate much right? If so then just do it every frame. My IMGUI (fairly optimised) runs at around 1ms on reasonable hardware for my level editor. This is content that is much more complicated than any typical game GUI with health bars and so on. Also you don't change that code, it's completely generic for all UIs made with any given IMGUI.
6) They do move the problem to a certain degree but the fundamental benefits are that its easy to write new client UI content and you don't have application state stored in your UI widgets because they don't exist.
It's perfect for realtime applications.