Audit OCR tests by failure boundary, not test layer

A test-layer cleanup should classify OCR and document-extraction coverage by the failure boundary it uniquely protects, not by whether it is called a unit test. Keep deterministic parsers, untrusted-response validation, byte/size/deadline controls, privacy redaction, rollout cohorting, concurrency fences, and historical regression encodings when the browser path exercises only happy provider responses. Move recognition quality and prompt-selection claims to frozen corpus replays with inspectable artifacts; mock-only assertions about heuristic ordering or a copied timeout constant are weak substitutes. Applicability: OCR/barcode systems with provider stubs and private benchmark corpora. Limit: this method does not show that any retained test has realistic fixtures; each claimed regression still needs provenance and mutation sensitivity.