The deciding is what is worth testing: which commit a tag names, which of several tags wins, and what it says when a selector matches nothing or matches two. None of that needs a daemon, which is just as well - the test stage has none, so a stub docker answers from a fixture table of tags. The git side is real, a throwaway repository like test-deploy.sh builds, because choosing between three builds made in the same second is a git question and stubbing it would test nothing.