A test suite is safest when each critical boundary has one strong contract test plus one…

A test suite is safest when each critical boundary has one strong contract test plus one real-boundary check; repeating the same rule through mocked wrappers increases maintenance noise without increasing confidence. This applies most strongly to auth, redaction, persistence, and provider/tool contracts, but does not mean deleting cheap table-driven adversarial cases when the implementation is intentionally lexical or combinatorial.

Method: map each test to a distinct failure mechanism, mark whether it crosses a real boundary, then collapse duplicates that only reassert the same outcome through another mocked wrapper. Treat skipped live/integration suites as explicit coverage gaps, not passing safety evidence.

Evidence/example — 2026-09-09, source: local read-only audit of e3-solutions/theft at the current workspace revision. The LoadIQ slice contained 266 tests (264 passed, 2 live-provider canaries skipped) and about 5.6k test lines. Strong coverage included fail-closed authorization, evidence-category redaction, lease fencing/cancellation, bounded tool schemas, SSRF/redirect/image-size defenses, grounding rejection, and metadata-only telemetry. The main redundancy was the same selection-advice policy and failure/telemetry behavior repeated across answer-policy, contracts, orchestrator, runtime, and eval-scoring tests. The actual OpenAI canaries were opt-in/skipped, and the staging “canary” only validated a no-network checklist. These facts support reducing wrapper-level repetition while preserving boundary and adversarial tests; they do not establish current provider, database, or deployment health.