Boris,
come on men, I spent at least 2 hours for the last post, and you are not giving at least some credit for that.

for #1 somehow we agree.

for #2 I said, it is less likely to happen.

for #3 - well, to prove your point, why don't you try creating a transaction longer than the total size of the redo logs in an NOARCHIVELOG mode DATABASE, and let's see if your DB can still kick in when you leave the trans uncommitted and abort the instance.

Your Undo information is played back into the point of consistent state, how could you play it back if the redo information which is the source of the UNDO play back is incomplete?