fix(m5-full-sample): harden run-swe.js with retry+persistence, bring 4 short results to full n, reconcile scorecard

- Harden evals/run-swe.js: retry null per-item pairs (3 retries, 4 total attempts),
  persist per-item outcomes (items:[{id,skill_correct,placebo_correct,scored}]),
  exit non-zero with unscored ids when scored_count < dataset_count
- Re-run map-territory on FROZEN 224-item decisive split (was n=222):
  full-n=224, 57.1% vs 54.9% placebo, +2.2pp p=0.074, provenance preserved
- occams-razor and kepner-tregoe already at n=224 from prior hardened run
- scientific-method already at n=150 from prior hardened run
- Update scorecard JSON+MD evidence tuples to full-n numbers:
  occams-razor n=224 (-0.9pp p=0.724), kepner-tregoe n=224 (-1.8pp p=0.289),
  map-territory n=224 (+2.2pp p=0.074), scientific-method n=150 (+5.3pp p=0.061)
- Re-evaluate scientific-method: full-n=150 primary passes >=5pp but fails
  p<0.05 (0.061); downgrade ELEVATE→DIRECTIONAL-NOT-REPLICATED per rules
- Reconcile coverage counts: measured=19, unmeasured=20 (JSON+MD agree)
- Debugging verdicts remain NO-LIFT; exploratory tags preserved
- Exclude scripts/quality-report.json from commit