feat: implement robust personnel data extraction pipeline with text-based fallback and coordinate-aware processing

src/ocr_sprint/pipeline/orchestrator.py

@@ -20,6 +20,7 @@ from ocr_sprint.pipeline.confidence import compute_confidence, route
 from ocr_sprint.pipeline.document_detect import DocumentDetectConfig, detect_and_correct
 from ocr_sprint.pipeline.extract.personnel import extract_personnel
 from ocr_sprint.pipeline.extract.personnel_text import (
+    extract_personnel_from_ocr_lines,
     extract_personnel_from_text,
     is_low_quality,
 )
@@ -144,12 +145,37 @@ def run_pipeline(content: bytes) -> PipelineOutput:
     # through the preferred path.
     if is_low_quality(personel):
         fallback_rows = extract_personnel_from_text(full_text)
+        # If text-based fallback produced rows but they all lack NRP
+        # (Pass 3 territory), retry with the column-aware extractor that
+        # uses OCR bounding boxes. On dense tables (e.g. Polda Kalbar
+        # Akpol-panitia), text-only Pass 3 bleeds adjacent columns into
+        # nama/jabatan because lines are interleaved within each Y-band;
+        # the columnar variant restricts each field to its visual column.
+        text_only_no_nrp = bool(fallback_rows) and all(
+            r.nrp is None for r in fallback_rows
+        )
+        if (not fallback_rows) or text_only_no_nrp:
+            ocr_lines = [ln for page in ocr_pages for ln in page.lines]
+            columnar_rows = extract_personnel_from_ocr_lines(ocr_lines)
+            if columnar_rows and (
+                not fallback_rows or len(columnar_rows) >= len(fallback_rows)
+            ):
+                fallback_rows = columnar_rows
         if fallback_rows:
             personel = fallback_rows
-            table_flags.append(ReviewFlag.PERSONNEL_TEXT_FALLBACK)
+            # Pass 3 / columnar emit rows with nrp=None for sprint
+            # templates without an NRP column. Surface that with a
+            # distinct flag so operators know to expect missing NRPs by
+            # design rather than by OCR failure.
+            no_nrp = all(r.nrp is None for r in fallback_rows)
+            if no_nrp:
+                table_flags.append(ReviewFlag.PERSONNEL_TEXT_FALLBACK_NO_NRP)
+            else:
+                table_flags.append(ReviewFlag.PERSONNEL_TEXT_FALLBACK)
             _logger.info(
                 "pipeline.personnel_text_fallback",
                 fallback_rows=len(fallback_rows),
+                no_nrp=no_nrp,
             )
 
     untuk_items = find_untuk_list(full_text)