It's a gamble. 20% 3-guess win/80% 4-guess to FAIL or 100% 4-guess win.
Maximizing information from each guess until the answer is down to two candidates minimizes the number of guesses needed to win.
Sure, my point is that the metric "minimize mean number of guesses to win" isn't axiomatically the objective function: the 6 threshold already suggests that many people would rather have a "guaranteed 6 and average 5" rather than "average 4 but fail 5% percent of the time".
If someone considers a 4 to be a soft-failure then they will prefer (3,6,6) to (4,4,4) even though the latter is a lower mean number of guesses: it's a higher count of 4+ 'failures'. The strategy that considers (3,6,6) a better result than (4,4,4) wouldn't guess "dream" in the above example.
If you want to maximize your number of three-guess-wins even if your 6-guess-failure rate is higher, it is obviously can't be a correct strategy to use your third guess on a word that you know can't be correct.