v0.1.3

First tagged snapshot of the lab: an A2A-style back-office agent can propose a risky action, but nothing hits the ledger unless a second person approves it. Includes policy tiers (auto, needs-approval, forbidden), approval by a different person, a hash-chained log and a kill switch. Live samples are small and use a single model, so they are indicative only and not a model score. Lab only. Not a production release. See CHANGELOG.md.