output, not the normalised output.  The latter may have newlines
removed, making the diff unreadable.  I broke this recently, but it
was only working by accident, now it is working by design (and is
commented, to boot).

Caused failures when run in parallel

Tests can now specify that they require the interpreter (i.e. that
they use one of ghci, annotations, TH, etc).

Sometimes, "prog 2> err" won't actually create err on Windows.

I think the changes are just due to changes in the random library.