Having poked around the source as part of implementing full vim support, I believe that nodes.coffee [1] exercises every single coffeescript feature except splats and js escaping, so that might be interesting to people as well even if it is a less interesting part of the source.
Quick links to other interesting bits (also accessible from the "Jump to" link on that page):
The Lexer, which has some extra smarts to help keep the grammar cleaner than it otherwise would be:
http://jashkenas.github.com/coffee-script/documentation/docs...
The implementation of lexical scope:
http://jashkenas.github.com/coffee-script/documentation/docs...
The "coffee" command-line utility:
http://jashkenas.github.com/coffee-script/documentation/docs...
All of those pages are created by a little literate-programming-style documentation generator called "Docco". If you run it on itself, you get this:
http://jashkenas.github.com/docco/
Hopefully we'll get some good patches from having the source in the open like this.