GRIEVANCES

 - opening two modules that introduce colliding names should be an error or 
   at least a warning

 - an operator/operand mismatch occurs between two record types that
   have N correctly unifying fields and M non-unifying fields should
   hilight the non-unifying fields in some fashion (currently you 
   have to eyeball it, which is painful on large fields)

 - error messages should demonstrate the source context by
   "ASCII-underlining" or ocaml-like highlighting

 - you should be able to declare a top-level function type, not just inline

 - type inference messages are nigh-useless (okay, fine, open research
   problem)

 - uncaught exceptions that propagate to the top level appear to silently
   disappear, requiring every incompete case statement to include
   a "catch-all" clause that prints out appropriate debugging information. 
   Java does this better, by printing out the full stack trace.

 - error messages at the REPL are sent to stdout, not stderr

 - there's no analog of perl -e to evaluate one-liners from the shell

 - heap image files have architecture-dependent filename suffixes, which
   makes it especially hard to create makefile targets for them

 - there's no way to inspect a type from the REPL

 - documentation is completely disorganized, out of date, and incomplete

 - no GNU readline at the REPL

 - no nested signatures (e.g., signatures declared inside of structures)

 - no polymorphic recursion! this has bitten us a couple times

 - no recursive modules