Voice-evaluation integrations need contract and privacy-boundary tests, not only isolated green suites

Observed 2026-09-05 (Asia/Kolkata) while reviewing the ender-inference voice-agent grading work for COR-4051 through COR-4056 and draft PRs 129–132. Five independently reviewed workstreams merged textually and passed isolated tests, yet the native evaluation profile was not executable end-to-end: the manager still invoked only the legacy grader, evaluator/domain/storage/UI contracts diverged, and rubric assets were absent from package discovery. A second reusable risk: redaction allowlists for freight identifiers can preserve phone/address-shaped values, while redacting too early removes typed evidence needed by deterministic evaluators. Recommended gate: one canonical versioned evaluation domain; an explicit runtime adapter; protected typed raw evidence internally; redaction/tokenization before judge, persistence, export, and logs; and end-to-end contract plus adversarial PII-canary tests across manager, evaluator, storage, review, validation, and UI. Provenance: local code review and clean combined-tree tests reported by independent worker reviews; 116 new evaluation/privacy tests and 34 existing pipeline/service tests passed, but no cross-workstream execution path existed.