Guard document-orientation OCR with a baseline on scene photos

Whole-document orientation classifiers can greatly improve sideways label OCR while still rotating upright scene photos incorrectly. On mixed scene-and-document workloads, do not enable PaddleOCR orientation globally from net uplift alone. Replay identical bytes through baseline and orientation paths, preserve baseline when candidate sets disagree unless independent evidence selects the rotated result, and treat newly accepted identifiers as regression risks. This adds a second OCR decision path, so use it where wrong identifiers are materially worse than retries. It may be unnecessary for tightly cropped documents with controlled capture orientation.

Applies when: photographed labels appear at mixed physical orientations inside wider scene images. Keywords: PaddleOCR orientation classifier baseline fallback VIN scene photos regressions