Match OCR fine-tuning and validation units to the regions served in production

Match OCR fine-tuning and validation units to the regions served in production. A recognizer trained and validated only on isolated words can improve its word metric without establishing that it still recognizes multiword detector crops, spaces, punctuation, or long strings. Before changing weights, compare label length, whitespace frequency, crop aspect ratio, and serving-region distributions. Evaluate the selected checkpoint through the unchanged detector and downstream extraction path, and reject an end-to-end regression even when the crop metric improves. If spaces disappear, check both the actual runtime decoder vocabulary/defaults and the training distribution; absence of an explicit space in exported YAML alone does not prove a decoder bug. Applicability: detector→recognizer pipelines where training crops are generated by a different teacher or granularity. Limit: these checks identify a distribution mismatch hypothesis, not its causal contribution; a matched-region ablation is still needed.