PR #72 reports bad underlines in editors using the language server and
proposes removing `nonword: true`. The diagnosis is right, the change
isn't sufficient: the token consumes the leading space, so mid-sentence
"Yesterday I walked" matches " I" and the alert covers one character too
many, reading "such as ' I'". Removing `nonword` leaves that unchanged.

Using lookaround makes the match the pronoun itself. The two `I` tokens
collapse into one, since they differed only in the delimiter they
required.

`nonword` goes too, which the PR was right about for a different reason:
it prevents a project Vocab from applying to the rule.

No fixture covered this. Every `I` in the fixtures sits at the start of a
line, where `^` matches zero-width and no space is consumed, so the
existing expectations were identical either way. Added a mid-sentence
case, which moves from column 10 with ' I' to column 11 with 'I'.