for C++ we have a marshal function for each class...
  which calls start_, then marshal for each thing in the class, then finish_.

Just use overloading to identify marshal functions (first argument? or should
the output_proc and closure be first and then the other stuff?)

unmarshal likewise takes an input proc and a closure, and a pointer (or
reference?) to storage for the type...

should we use stream classes instead of input procs? no: we want to be able to
use it conveniently from C too.