There are only two roads and one village, and the problem is to find out which of the roads lead to the village. Forget about "guarding" and the "truth village".
If you ignore the random answerer for a second, then you ask "if I were to ask you if this road leads to the village, what would you say?". Let's call this the original question. The answer will be trustable, because the truth teller will tell the truth, and the liar will lie about a lie, making the result the truth as well.
Now we introduce a new question, and a third answerer -- random. Since we know we can solve the problem with only the original question and original two natives, the goal of this question is just to eliminate the random answerer.
The solution given in the article uses this new question in two ways. One, by asking it of one native and not ever asking that native the original question, if the native we happen to ask is the random answerer, we know we won't ask the original question of that answerer, so what they answer doesn't matter -- either way, we'll end up with the liar or truth teller to ask the original question of.
So we only need to design the new question assuming the one we ask it of is not the random answerer. So now we're back to asking a "double question", resulting in a truth about a truth or a lie about a lie. You ask the question you stated, knowing that the native identified as "least likely to tell the truth" will never be the random one, due to truth-about-truth, lie-about-lie, or due to the random answerer being neither B nor C.
If you ignore the random answerer for a second, then you ask "if I were to ask you if this road leads to the village, what would you say?". Let's call this the original question. The answer will be trustable, because the truth teller will tell the truth, and the liar will lie about a lie, making the result the truth as well.
Now we introduce a new question, and a third answerer -- random. Since we know we can solve the problem with only the original question and original two natives, the goal of this question is just to eliminate the random answerer.
The solution given in the article uses this new question in two ways. One, by asking it of one native and not ever asking that native the original question, if the native we happen to ask is the random answerer, we know we won't ask the original question of that answerer, so what they answer doesn't matter -- either way, we'll end up with the liar or truth teller to ask the original question of.
So we only need to design the new question assuming the one we ask it of is not the random answerer. So now we're back to asking a "double question", resulting in a truth about a truth or a lie about a lie. You ask the question you stated, knowing that the native identified as "least likely to tell the truth" will never be the random one, due to truth-about-truth, lie-about-lie, or due to the random answerer being neither B nor C.