This is a list of things I definitely want to do.  I don't go into too many
specifics, because the specifics change often as I think about the problem
harder.

Core Parsing Algorithm (all changes should have test-cases):
* properly support EOF in lookahead (like s -> "A" | "A" "B"; should predict 1).
* properly support lookahead that predicts a return from a final RTN state
  like s -> a "X"; a -> "Y"*;
* try to detect when we are in an infinite loop (requires much thought).
* detect cases where some RTN alternatives have no GLA final state.
* deal with lexer-level ambiguity.  longest match will do for now, but
  we're not currently detecting s -> "A" | "AB";

Runtime:
* Richer callback specifiers.
* Bring back slotbufs: a cheap (stack only, no heap) way of saving parse_vals for
  the currently-open nodes of the parse tree.
* Provide a buffering layer.
* As the runtime starts to mature: language bindings.

Tests!  Everything should have tests, as much as possible:
* all aspects of compilation
* bitcode format (both reading and writing)
* JSON output from gzlparse.  Both well-formedness and accuracy.

Major design areas that exist only in my head:
* Embedding Lua to do things only an imperative language can do.
* Operator-precedence parsing using the shunting yard algorithm.
* Parallel parsers (for both embedded languages and things like whitespace/comments)
* Error recovery (basically: just yield to an imperative function).
* Generate imperative bytecode / JIT compile.
* AST-building.