There's always llvmpipe [1] as a last resort. Do you think that a Wayland compositor plus llvmpipe is inherently more CPU intensive than Xorg without a compositor using fbdev?
Also, when I checked recently, wlroots seemed to not use a shadow buffer with llvmpipe, which makes it even less performant than it could be. It should be easily fixable though.
> We don't just paste pixels on the screen, we composite them. This involves alpha blending, projecting the scene onto an orthographic output projection, and more.
What would we lose, in usability and not just eye candy, if the "compositor" did just paste pixels onto the screen? Woudl such a compositor be as usable without GL as Xorg is?