OCR-labelled ROIs need value equivalence, not presence alone

Observed 2026-09-08 during COR-4032 local labelled-data verification (Google Vision TEXT_DETECTION receipts; no raw OCR persisted): four crops each returned one unambiguous USDOT candidate but disagreed with the original-image candidate. Modest context widening restored exact normalized equality for two; two remained digit-different after three progressively broader, visually complete crops. Reusable rule: accept an ROI only when the normalized required identifier exactly equals the original-image result, not merely when both arms detect the field. When widening would recreate most of the source, classify the sample OCR-context-unstable or hold it for review instead of treating field presence as success. Gemini confirmation was unavailable due provider billing cap, so this evidence is Vision-specific.