fix(m1-scorecard): resolve 3 blocking scrutiny issues

- scientific-method: replication qualifier now consistently false in JSON and MD (fresh replication pending M5)
- occams-razor: evidence cites existing evals/results/run1/swe-occams-razor-improved.json with exact mcnemar_p=0.505
- Coverage/verdict separation: 4 skills with objective evidence (circle-of-competence, margin-of-safety, occams-razor, second-order) now have proper verdicts instead of 'unmeasured'; counts yield 17 measured / 22 unmeasured
- JSON and MD mutually consistent on all 39 skills