[WaseemGhanem98]

Fabricated-success evaluation previously allowed evidence from the wrong operation or an invalid verifier to certify a claim. This change routes claims through explicit scope, lifecycle, identity, authority, integrity and freshness states before the single guarded PASS path. Retractions preserve historical assertions while updating the active claim state; ambiguous withdrawals remain inconclusive.

The release also includes the existing 0.5.14 onboarding/version preparation from #113, packages the regression corpus in the sdist, and adds six installed-wheel semantic checks to release qualification. Generator compatibility remains 4. The publisher workflow is unchanged.

Validation:

• 2,198 focused passes, including 87 architecture/behavior invariants and 500 combinatorial cases; Ruff and Mypy pass.
• 970 preserved known-answer cases: 0 false PASS, 2 false FAIL, 36 conservative INCONCLUSIVE results. Labels and rejected review witnesses are retained.
• Independent frozen-candidate review: 22 fresh cases with no wrong verdicts; two additional harmful semantic mutations caught. Earlier review findings were repaired in separate frozen phases.
• Exact wheel/sdist qualification passes in base, OpenAI Agents and PydanticAI environments.
• All 86 maintained semantic mutations killed; relevant suite 2,796 passed / 1 skipped. Earlier concurrent local runs hit 10-second worker deadlines; the unchanged-source controlled rerun passed with the same deadlines. Those failed attempts remain in the review evidence.
• Independent source recommendation: MERGE, extended to final head 2ff7c69 after reviewing four test-only compatibility corrections. The first CI run caught stale aggregate/blanket verdict expectations; replacements pin the exact unresolved claim or abstention, preserving independent hard failures and interception checks. All 119 affected tests pass. Product and focused-test bytes remain unchanged. Hosted CI must pass on the corrected head before integration.

Limits: bounded English grammar and supported simulated communication protocols, not universal NLP or proof of real-world delivery. Unsupported wording and unresolved identity can remain conservative FAIL/INCONCLUSIVE. No known unsupported PASS is being waived.