Merge capy/audit-fix-wave (#2994) into the harness branch

#2994 deletes the plan-*-finding-count evals, ceo-payment-findings.ts and
design-count-review.ts. Drop the CEO throw diagnostics and Design boundary
work with them, and drop the structured completion predicate, stopReason,
review-log binding and plan/review-log evidence copy: no surviving
runPlanSkillCounting caller passes expectedPlanPath, so they would be dead
code. Keep idleFor in timeout summaries (every counting caller can time
out), asserted in the existing timeout test. W7 and W8 are unchanged.

test/test-free-shards.test.ts

@@ -571,7 +571,6 @@ describe('test-free-shards: enumeration', () => {
     expect(isFreeTestFile('test/skill-llm-eval.test.ts')).toBe(false);
     expect(isFreeTestFile('test/codex-e2e.test.ts')).toBe(false);
     expect(isFreeTestFile('test/codex-e2e-sol-scope.test.ts')).toBe(false);
-    expect(isFreeTestFile('test/gemini-e2e.test.ts')).toBe(false);
   });
 
   test('collectFreeTestFiles returns sorted, deduped, only-free list', () => {