Measure and catch hallucinations in lease extraction

The harness is two prompt variants that differ in one sentence and one enum value, run over the same six leases three times each. Everything that decides the result is code with no model in it: a generator that writes the leases and their gold keys together, a scorer with five outcome classes, and a quote check that looks inside the cited clause. The decision that matters most is the scoring table. A claim on a topic the lease never mentions is counted as invented, separately from a wrong rule on a topic the lease does cover, because the two have different causes and different fixes.