ec287fd12e
Each module is under 1050 lines: - ocr_pipeline_common.py (354) - shared state, cache, models, helpers - ocr_pipeline_sessions.py (483) - session CRUD, image serving, doc-type - ocr_pipeline_geometry.py (1025) - deskew, dewarp, structure, columns - ocr_pipeline_rows.py (348) - row detection, box-overlay helper - ocr_pipeline_words.py (876) - word detection (SSE), paddle-direct - ocr_pipeline_ocr_merge.py (615) - merge helpers, kombi endpoints - ocr_pipeline_postprocess.py (929) - LLM review, reconstruction, export - ocr_pipeline_auto.py (705) - auto-mode orchestrator, reprocess ocr_pipeline_api.py is now a 61-line thin wrapper that re-exports router, _cache, and test-imported symbols for backward compatibility. No changes needed in main.py or tests. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>