I've recently learned of Parsing Expression Grammars (PEGs) and Packrat Parsing. I think this could be well suited to an extensible language. This is a linear time algorithm which can easily take unfactored BNF style productions as its language definition. It is a memory hog, a reasonable size module might take 50M to parse, but I think the simplicity of syntax specification will more than out weigh the memory foot print. (That image processing language ran on a Vax11-750 with 8M of RAM shared with 10 people. This is a technique which is only now becoming feasible with larger memory sizes.)
I am currently prototyping a packrat parser for an extensible language. The prototype is in Javascript and runs inside a browser with fields for the grammar and the input string which makes interaction trivial.
It is not on a publicly accessible machine. I may put a snapshot of it in a public place if anyone asks.