After example fixtures are regenerated, make tests read ids/values from the example instead of hard-coding them
21 tests failed when the golden examples were regenerated from real analysis output: they hard-coded the house question id/value (q_service_amps, answer 100.0), a dollar figure in the example narrative ($9,302), and a bundle-only decimal. Fixes that generalise: load the first question from the example file and answer with its first non-null choice; find a dollar figure that appears once in the narrative and bump it; format a bundle-only value at the precision the checker compares (two decimals); for 'summary is stale after replan', do a first replan to a new budget, append a sentence stating it, then replan again so the sentence stops matching. Also remove stubs that replay example evaluations for unsupported measures: they put stale cards (a roof card on a 3rd-floor suite) into live results.