Reduce false positives on narrative prose 'done'

[Yesol-Pilot]

Known limitation (see README "Honest limitations"): a stray done in ordinary prose is flagged.

Scope

• Tighten the claim regexes to require a stronger completion context, and/or add a --prose mode that is more conservative.
• Add before/after test cases so precision is measurable, not vibes.

This directly improves signal-to-noise, which is the whole value of the tool.