feat(review-suite)!: remove the unproven verification-sufficiency pass

[shaug]

Summary

Removes the verification-sufficiency pass and its mandatory
verification_sufficiency_evidence field from review-correctness and the
shared review-result contract, advancing schema_version 1.3 → 1.4. The
traversal (consumer/impact) pass and consumer_impact_evidence — the other
mechanism #52/#53 added — are unchanged, byte-for-byte.

• skills/review-correctness/SKILL.md: removed the verification-sufficiency
  pass instruction. The underlying question (does a claimed test actually
  exercise the triggering condition it addresses?) remains part of ordinary
  correctness review, just without a separately mandated pass or required
  evidence field. The traversal pass instruction is preserved exactly as
  shipped.
• review-suite/contracts/review-result.schema.json,
  review-suite/CONTRACT.md, review-suite/scripts/validate.py:
  verification_sufficiency_evidence removed entirely (full removal, not an
  optional field — see the decision record for why); schema_version
1.3 → 1.4; a stale 1.3 result now fails with a clear migration error.
• Every canonical fixture, standalone expected result, and orchestration case
  across all six bundled review skills migrated to 1.4 atomically; all six
  bundled references/review-suite/ copies refreshed via just sync-contracts and confirmed byte-identical.
• verification-sufficiency-guard fixture retired (its sole purpose was the
  removed field's clean-pairing behavior). standalone-verification- sufficiency-gap adapted, not deleted: its underlying defect pattern is
  a genuine correctness catch independent of the removed pass (see
  review-suite/fixtures/missing-test/ for the same pattern caught by
  ordinary review), so its changes_required verdict and blocking finding
  are preserved.
• stale-claim-release-guard and audit-log-flush-keyword-probe corpus
  cases untouched — still general correctness cases.
• New decision record: review-suite/evals/v2/VERIFICATION-SUFFICIENCY-REMOVAL.md,
  citing S1-ABLATION-MATRIX.md and discriminating-case-validation.md
  directly.
• README.md and CHANGELOG.md updated to describe the removal and its
  evidentiary basis.

Why

Two independent, honestly-sourced measurements found no demonstrated value
for this pass:

• stale-claim-release-guard (Run the preregistered review v2 ablation and integration closeout #57's ablation matrix): resolves identically
  at mean_combined_recall 1.0 whether the pass is enabled or disabled.
• audit-log-flush-keyword-probe (Source harder discriminating cases for #52/#53's traversal and verification-sufficiency passes #89's harder-case validation): also
  resolves identically at mean_combined_recall 1.0 in both configurations.

On top of the absence of demonstrated value, the pass carries a confirmed,
twice-reproduced cost: run in isolation (traversal pass disabled), it
regresses session-continuation-summary — a previously clean control — to
3/5 false positives.

By contrast, the traversal pass does show a real, reproducible gap in
the same evidence (artifact-promotion-environment-shortcut,
mean_combined_recall 1.0 → 0.4) and is out of scope for this removal.

This ticket executes a removal decision already approved by the repository
owner on 2026-07-29; it does not re-litigate the evidence.

Non-goals honored

• Traversal pass / consumer_impact_evidence: untouched (verified
  byte-identical against base, independently confirmed by a fresh
  review-code-change cycle).
• stale-claim-release-guard / audit-log-flush-keyword-probe corpus cases:
  untouched as corpus content.
• review-suite/evals/baseline/v1/, gate-manifest.json,
  DECISION-RECORD.md, FAILURE-TAXONOMY.md, audits/, and every existing
  review-suite/evals/v2/ measurement document (S1-ABLATION-MATRIX.md,
  FROZEN-V2-CONFIGURATION.md, CLOSEOUT-REPORT.md,
  discriminating-case-validation.md, etc.): untouched.
• Epic Epic: Make review-clean verdicts evidence-backed and empirically evaluated #49 closure: not performed (separate, later action).
• carve-changesets: not invoked — this candidate fit an ordinary single PR.

Test plan

• just format — all checks passed, no reformatting needed
• just lint — skills-ref validate passed for all 8 skills; plugin
  packaging validation passed
• just test — 432 tests across review-suite/scripts/tests (310) and
  all 6 skill-level scripts/tests suites (122), all passing
• Fresh, isolated review-code-change cycle at the final head — clean
  aggregate verdict, zero findings across all three lenses
  (solution-simplicity, correctness, code-simplicity)

Fixes #93