The observed faulty state transition is probably hard to test,
because we need to hook into internal states to observe them.
Namely this:

    07:30:16 | Build#ruby-2.2 enqueue: created -> pending
    07:30:16 | Pipeline#32    enqueue: created -> pending
    07:30:16 | Build#ruby-2.3 enqueue: created -> pending
    07:30:16 | Build#ruby-2.2     run: pending -> running
    07:30:16 | Pipeline#32        run: pending -> running
    07:30:29 | Build#ruby-2.2    drop: running -> failed
    07:30:29 | Pipeline#32        run: running -> running
    07:30:29 | Build#ruby-2.3     run: pending -> running
    07:30:30 | Pipeline#32        run: running -> running
    07:30:57 | Build#gem:build   skip: created -> skipped
    07:30:57 | Pipeline#32       drop: running -> failed
    07:30:57 | Build#gem:release skip: created -> skipped
    07:30:57 | Pipeline#32       drop:  failed -> failed
    07:30:57 | Build#ruby-2.3    drop: running -> failed
    07:30:57 | Pipeline#32       drop: running -> failed
                                       ^^^ Should be failed -> failed

However, the consequence of this, executing hooks twice would be
easy enough to observe. So we could at least test against this.
Keep in mind that if we ever changed how we execute the hooks
this won't be testing against faulty state transition.