These are indeed tough but very good questions! 
Here are some of my thoughts in terms of the ideal scenario and the practical scenario.
Ideally, we expect the robots and computer-assisted human to be marked by OGS, or explicitly state their identity in the profile page or at the beginning of the game.
In reality, we cannot enforce the players to disclose their identities, so there’s no practical way to tell.
Ideally, yes.
In reality, some opponents might not be aware that they can be facing a computer-assisted player in the tournament, so disclosing it could cause some unnecessary upsets to them. Nevertheless, people joining the Alan Turing tournaments are supposed to understand that the opponent could be an AI or computer-assisted human (this is another reason why I need to highlight this condition in the next tournament). I think the decision is up to the participants for now.
Ideally, if the rank of the players are not consistent with their strength in the other games, the games should not be rated.
In reality, all tournaments on OGS seem to be rated and the TD cannot change that. Also, I assume most players are human and most game results still make sense for the ranking.
Ideally, if the strengths of the players are not consistent with their ranks, McMahon would not make sense in this tournament. Ideally, a random pairing format would be better.
In reality, the ranks still make sense for most players and since there isn’t a random pairing option but only rank/strength-based pairing currently for multiple-group tournaments, we still need a reduced McMahon system to offset the difference in strengths of the opponents in practice. Actually, assuming there’s a 8k player starting with -3 points but actually representing a 6d bot, the player can simply win all the games in all three rounds (~18 games) and obtain a score of 15, which would almost certainly be in the top-3 for this tournament. If we’d like to give more chance to the lower-ranked players, we probably can reduce the McMahon difference further in the next Alan Turing tournament. As an extreme case, Meijin Handicap 2016 gives every player the same starting score because we assume every handicap game to be even.
Finally, I think one ideal way to solve most of the problems is to make sure every account has a relatively stable and accurate strength. For example, PlayerA (1k), AI_owned_by_PlayerA (2d), PlayerA_with_AI_Help (3d) would be three accounts of the same player with different but inherently stable ranks. However, at current stage I think there isn’t a big problem if the strength of the player is not too far away from the strength of the AI.