Extensibility by itself doesn't solve the terrarium problem because it doesn't define any standard, so there isn't anything to build on. When presented with extensibility, you still build the terrarium, it's just a customized one, and if you do it ground-up like Forth, you can potentially build it smaller and simpler. But in the end you still have a terrarium with assumed boundaries.
This has led me away from Forth-the-language in the last few months to explore the terrarium problem further, and I hit on the idea of treating this as an organizational issue solved at the level of the core UX. If the problem is defining boundaries in software, we should have better ways of doing that. At first I considered this in terms of selection - selection being one of the pillars of structured programming, and many of our improvements taking the form of easier selection metaphors. This gradually led me to explore the idea of document editing with a binder and sticky notes metaphor, with a supplemental compiler process taking the resulting complex, layered documents and processing them into a linear form for consumption. The presumption is that if I present a rich set of tools for defining types of divisions and groupings defined skeumorphically - pages, bins, tape, wire, guides, stickies, overlays - and then add a bit of labelling and indirection on top, a powerful "thinking tool" should emerge where the organization is easy and the compilation system makes it easy to query and traverse it in a customized way.
If I can finish designing it.