Why bad scientific code beats code following "best practices