> You would have a plugin to the Weston (or similar) compositor which understands how to render windows and GTK+ widgets.
You've just described server-side rendering. It's what we've spent 10 years trying to abandon in the X-world with direct rendering, and what we're finally abandoning completely with Wayland.
At the end of the day, maintaining a drawing API in the compositor is just a maintainer's nightmare, hence why we're leaving X (this is basically what HTML and a HTML rendering engine do. Use these, they are tailored for the task.
Just spend a few minutes pondering that "XML format" you suggest. I know a certain XLM-format that fits the features you seek, it's called HTML. Just imagine having to implement WebKit inside a compositing window manager. Now that sounds like a fun project :).