This was previously not possible because of the bug in the
explicit/numdiff assembler. The large epsilon masked the error
in the storage derivative. Now it also works with small epsilon.