concur: without the restricted wordlist and in hard mode, it does feel like locking in a letter hurts a lot; many of the popularised starts can easily get trapped after the first guess:
* SIREN gets trapped by [B/C/D/F/K/L/M/N/P/S/T/V/W]INES
* AROSE/SOARE gets trapped by RA[C/G/J/K/L/P/T/V/X/Z]ES (and maybe RARES/RASES) as well as S[C/E/H/N/P/T/W]ARE
* SALET gets trapped by [D/F/H/M/P/R/V/W/Z]EALS (but not TEALS or SEALS)
* ORATE/ROATE gets trapped by TA[B/K/L/M/P/S/V/W/X]ER (and maybe TATER)
most, if not all, of the "best strategy" presentations depend on the reduced solution list to avoid corridor traps ([B/D/E/F/H/L/M/N/R/S/T/W]IGHT is reduced to [E/F/L/M/N/R/S/T/W]IGHT for example; also almost all S-terminal plurals and present verbs are not on the list). given that, by an exhaustive search posted by Alex Selby, the best word averages 3.42 guesses and the worst word averages 4.10 guesses (in normal mode), it's probably more important not to fall into any traps while discovering information on hard mode, since the worst average has 1.90 guesses to spare (noting though that the worst guess ends up falling back to the 'optimal' guess in half the cases, and fails to guess the word some of the time).
from the human side of things, digraphs probably do play a significant role in making guesses; if R, H, T, S, L, and N are all ruled out, for example, you can be fairly certain the word has at least two vowels, without having to guess the vowels and restricting further guesses.
it doesn't feel like a binary search is optimal in hardmode, since confirmation reduces the ability to get more information, unless the search space is sufficiently small (which is true with the restricted word list, but not as much with the full word list)
(personally, CRWTH is my starter of choice, followed with confirmed letters with S+L+P / S+L+N / S+N+D)