test(search): raise golden-eval floors after ranking fix (MRR 0.7, top-3)

Board job cc269365 follow-through: with the definition boost + test-file
demotion landed (bc86c719), the measured mean MRR is 0.889 with worst rank
3. Gate now enforces mean MRR >= 0.7 and every judged answer in the top 3
(was 0.3 / top 10 under the pre-fix baseline).