Generate a balanced 24-task dataset

The design decisions live in two small YAML files you write yourself: the dimensions, and twelve pairs of cases. A generator turns each case into three synthetic documents and the reference enquiries a careful conveyancer would draft, so every quote in a reference is copied from the rendered text rather than typed by hand. The decision that matters most is making each negative case explicit: expected_issues: [] says nothing about what the agent must not do, so every task also gets forbidden_issues, and a negative case is one that forbids the very topic its positive partner requires.