The way XRay render text is more compositing than rendering.

They will render text to a bitmap using HTML canvas and then use that as a texture for GPU.

GPU will be compositing those bitmaps, it won't actually be rendering text.

Is it really so? Can OpenGL update only a small part of a framebuffer (for example, when a single letter is typed) or will it rerender the whole text?

Depends on how you render.

I've written a OpenGL engine for a university course that use partial frame updates to reduce the number of redraws per frame.

IIRC the partial rerender variant was about 3 times faster than the full render.

If done in a smart way, of course you can. Similar thing with canvas. If letter/word spacing depends on that character, then you'd probably need to rerender more stuff.