It does make sense, because the keepass entropy estimate presumably (like the excellent zxcvbn) tries to approximate the empirical distribution, not the theoretical uniform one.
In theory, "68703649" and "12345678" are equally likely to be pulled from the hat, but in practice one is a much better password than the other. You can reduce the search space by trying the passwords with higher (empirical) probability first.
Thanks. I've looked at the code, and it does not seem to try to estimate the empirical distribution (doesn't appear to be using dictionaries, for examples).
Then the discrepancy maybe comes from the number of glyphs within certain categories, or their repetition?
In theory, "68703649" and "12345678" are equally likely to be pulled from the hat, but in practice one is a much better password than the other. You can reduce the search space by trying the passwords with higher (empirical) probability first.