Huh, good to know. I just looked at how it is done, and we are literally peeking at the characters surrounding tokens and running a bit of Haskell to guess the classification of the previous/next token. Seems a bit brittle but I guess this is not something that is likely to change. I am also surprised it didn't slow down the lexer, but maybe it's lost in the noise.
Anyway, I don't have a strong opinion on the proposal as this is an extension I never use, but the proposed behavior seems more intuitive for humans, so it seems reasonable to accept it.
Iavor