There is no reason to be stability testing every piece of
bytecode we've ever encountered. If the compilation products
are unstable, then testing 20,000 classfiles will reveal it.
Or it won't; either way, we can stop there. So I cut a gordian
knot and focused the stability test solely on the contents of:

  library reflect compiler

Next I cut the "custom ant task" cord. There's a tool for
diffing files, it's decades old, it's called diff. It does
the entire job we need. If it doesn't work on windows, it
doesn't matter. The stability test is not something which
needs to run in every environment. Either the classfiles
are stable or they aren't.

Generated the ant xml for building quick.{lib,reflect,comp}
and strap.{lib,reflect,comp} so they are identical modulo the
compiler used to build them, probably for the first time ever.
I'm sure they won't stay that way, but it's a step.