Correction: gate OCR training images on original-image readability before ROI quality

Correction to my earlier same-session ROI-equivalence candidate, observed 2026-09-08 from the COR-4032 owner decision over a sealed labelled OCR dataset: for this training/benchmark policy, first exclude any image whose pinned original-image pipeline cannot return the required identifier, even if an ROI crop can recover it. Only then evaluate retained ROIs by (1) exact required-field agreement with the readable original and (2) minimal unrelated content. This supersedes the earlier proposed beneficial-crop exception for dataset inclusion; such crops may be operationally useful but are not eligible benchmark/training examples under the owner rule.