Again, you are confusing the resolution of the rules -- the pixel count if you will -- with a difference in kind. Step back a bit -- the GM has the party facing off against some set of foes. Here the normal loop applies -- the GM can determine if this is successful, or not successful, or if it's uncertain. Examples of autosuccess would be a 20th level party against a single CR0 rat -- no need to engage, GM perfectly okay to determine "rat dead, move on." Example of auto-failure would be 1st level party against Orcus -- don't roll, you die (or worse). The range for uncertainty is assumed wide here, but it need not be. The GM actually has the authority to determine auto-success/failure for any potential conflict. The social contract typically directs otherwise, but the rules do not.
So, then in the uncertainty case, we're still in the same place as for other things, but the detail level of the resolution process is higher. It's not a single check (or a few) but rather another process. In the end, the GM still determines the outcome, and can, in fact, exert control over the outcome at any point during the process. It's not necessary for the GM to run things to a specific end point -- they can interrupt at any point.
As for your 3) -- I will typically not lean on Rule Zero in discussion of games and how they play. If I do, I will be explicit about it. I find it best, when discussing rules, to not go for a "do whatever you want, because that's in the rules" explanation.