eval(wave-d): fermi rework re-measure (-5pp -> +7.5pp); Wave C validation; isolation log fix

- fermi-estimation reworked: jeggers/fermi numeric OOM n=40, placebo 40% -> skill
  47.5%, +7.5pp p=0.371 (ns). The agent-native rework FLIPPED the sign (old -5pp ->
  +7.5pp, ~+12.5pp swing) — echoes the scientific-method rescue. KILL -> REWORK-RESCUED
  (confirm at n=150). Recorded exp-eval-numeric-fermi-reworked.
- Wave C adversarial validation (gemini-3.1-pro): claims 1,2,3,5 TRUE — results real,
  recorded faithfully, replications genuinely fail, no fabrication. Lone defect was
  cosmetic (run-correctness didn't log isolation status) -> fixed.
- dual-process eval deferred (no clean objective dataset); rework validated by 62->0
  anthropomorphizing-reference removal.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>