foolish-bandit makes Gary admit when it can't cite a source

The original Mike left it to the user to decide whether to trust an AI answer. foolish-bandit moves that judgment into the interface itself. When a reply comes back with no sources behind it, the tool now flags it in amber: treat this as unverified. When sources are present, they're listed with a plain reminder to check them before relying on the answer.

The same discipline runs through the drafting, explain, and review screens - short notes telling the user that drafts need review and that findings should trace back to the underlying contract. It's a small, cheap change, but it puts the warning where the work happens instead of burying it in a help page nobody opens. Worth noting: the amber "unverified" flag also fires on plain factual questions where there was never a document to cite, so the signal can read louder than intended.