Test value comes from protecting meaningful boundaries, not mirroring implementation

A test earns its CI cost when it protects an externally meaningful boundary—untrusted input limits, failure classification, privacy redaction, bounded retries/deadlines, or domain acceptance—not merely when it repeats implementation text. Mock-heavy tests that only feed a desired model/provider response through and assert the same value should be consolidated into one boundary-contract test or moved to a labeled corpus/live evaluation; they cannot demonstrate model accuracy. Exact prompt wording and full telemetry-object equality are useful only for the few tokens/fields whose change would alter policy, privacy, or operability. This heuristic is strongest for integration adapters and AI extraction code; it does not justify removing cheap regression fixtures that encode previously observed parser failures.

Evidence/example — 2026-09-09

Repository-source audit of the Shield OCR/barcode/Rekognition unit slice found 23 passing files and 199 cases (local Vitest run: 13.57s wall, 404108 KB peak RSS). High-value coverage concentrated at trust and failure boundaries: bounded remote image reads, private endpoint/config validation, remote response schema rejection, local parser validation, cancellation/deadline behavior, privacy-safe telemetry, ROI/allocation bounds, and stable error taxonomy. The largest weak examples were a mocked license-specific classification whose mocked response already contained the expected classification, repeated model calls made only to inspect prompt substrings, a historical 7,493,879-byte allocation fixture below a general 10 MiB boundary, and a WASM failure test that invokes the same failing decode twice. Provenance: repository files under web/src/lib/{ocr,barcode,rekognition} and a focused local Vitest run. This is a static/unit-test audit; it does not measure production defect-prevention rates or E2E flakiness.