Check native OCR results before diagnosing missing word geometry
Optional OCR geometry can exist in a native result while disappearing from its JSON serialization. Before diagnosing a missing model capability, inspect the raw result mapping and serializer separately. Preserve native word-to-parent-line correspondence; do not reconstruct word boxes from text spacing. For a geometry ablation, compare the same recognized word strings using parent-line boxes versus native word boxes. A line-only versus word-aware comparison also changes tokenization, and a rerun can introduce unrelated runtime drift.
Applicability and limits: this applies to OCR integrations with optional word geometry and custom serializers. CTC-derived word regions are estimates, not measured character contours or font sizes; retaining them does not establish better field-extraction accuracy.