which Leon repaired so I understood what the problem was.

unit, integration, and regression tests can now share code as they are
compiled together.

Unit tests are now in package leon.unit and shared code lives in
leon.test.

It used to be extremely bad: if the interesting failing test is the last
one (of 300), CEGISLike would test 1000*300 times before ordering it to
make it first.

Now we assume that interesting tests are discovered early.