>But I think there is already quite a few things possible via messages and metaclasses, even if one cannot do actual AST transformations.
That's true. In fact, there are things that messages and metaclasses can do that you can't do with AST transformations without implementing those abstractions.
>After all, the whole image is accessible, so you can dynamically ask any object for its definition, or even compiled code (bytecode or JIT) and change them.
I still miss this in Lisp. Some Lisps had that, once, but it's uncommon nowadays. it happens in the commercial CLs, but those are expensive. OS CL implementations rarely have good image support (in SBCL, image saving actually corrupts the RAM state to the point of nonrecoverability, and the docs recomend fork(2)ing if you want to save an image and continue your app).
Aside from the proprietary CLs, PicoLisp is the only modern lisp environment that has this kind of dynamic capability AFAICT (and while it does sort of have images, in the form of external symbols, the language doesn't encourage using them like this. Also, calling it modern is a stretch: it has more in common with LISP 1.5 than, say, CL). And the Schemes? Don't make me laugh. Scheme has many strengths, but reflection isn't one of them. It's something that I really wish the Lisps had.
One of the reasons I want to try my hand at implementing Lisp on the Spur VM at some point.