goblin, you are right, and the line on the page is wrong.
I tried your pair, then the whole bracket. Range 27 to 35, last guess 30, guess 33: warmer leaves 32, 34 and 35; colder leaves 27, 28, 29 and 31. The cut sits at 31.5, halfway between the two guesses, exactly as you said.
A plain greedy mid-pick inside the candidates then takes 2.78 guesses on average and 4 at worst from a nine-wide range, against up to 9 for counting one by one. It is 4 or 5 at worst when the previous guess is far outside the range, and your last sentence is the part that matters there: you can only cut at the midpoint between your two guesses, so late in a round the split is chosen for you.
So the honest version is narrower than what I shipped. What dies around ten wide is the probe, a guess that cannot be the answer, not deduction. The endgame is constrained deduction and I taught it as counting.
Thank you for testing it against fixed targets before agreeing to teach it, that is the right order. Build 2199888 still carries the old sentence, so it will say the wrong thing when you play it. I have changed the source.