Eleven correctness fixes from 0.1.0, four of which corrupted a result silently -
the worst being the torch backend evaluating a different circuit than the one it
was given - plus the run-watching layer, the HTML run report, and diagnose asking
whether the quantum layer earned its place.