Rewrite the searches summariser and measure it

The experiment holds everything fixed except the system prompt: the same packs, the same user turn (documents first, request last), the same JSON schema, the same model and three trials each. That way any difference in the table can be traced to prompt lines. The decision that matters most is the gold set: each finding has a severity a fee earner would give it and a regex the scorer uses to match it, and the flood-risk item is "action" only at moderate or higher. Without that threshold in the gold, "always flag flood risk" would look like good recall instead of noise.