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

tests/unit/test_ocr_layout.py

@@ -0,0 +1,75 @@
+"""Tests for OCR layout reordering.
+
+PaddleOCR emits text boxes in detection order, not visual reading order.
+On dense table layouts (Polda Kalbar Akpol-panitia regression) this
+interleaves columns within a row and breaks every downstream extractor
+that assumes top-to-bottom row order. ``sort_lines_by_layout`` rebuilds
+reading order from the bounding-box geometry.
+"""
+
+from __future__ import annotations
+
+from ocr_sprint.pipeline.ocr import OCRLine, OCRPage, sort_lines_by_layout
+
+
+def _box(x: float, y: float, w: float = 30, h: float = 15):
+    return ((x, y), (x + w, y), (x + w, y + h), (x, y + h))
+
+
+def _make(text: str, x: float, y: float) -> OCRLine:
+    return OCRLine(text=text, confidence=1.0, box=_box(x, y))
+
+
+class TestSortLinesByLayout:
+    def test_empty_returns_empty(self) -> None:
+        assert sort_lines_by_layout([]) == []
+
+    def test_already_sorted_is_stable(self) -> None:
+        lines = [_make("A", 10, 10), _make("B", 50, 10), _make("C", 10, 30)]
+        assert [ln.text for ln in sort_lines_by_layout(lines)] == ["A", "B", "C"]
+
+    def test_reorders_column_first_detection_to_row_first(self) -> None:
+        # Simulate a 2-row, 3-col table where Paddle returned cells
+        # column-first instead of row-first.
+        lines = [
+            _make("B1", 50, 10),
+            _make("B2", 50, 30),
+            _make("A1", 10, 10),
+            _make("A2", 10, 30),
+            _make("C1", 90, 10),
+            _make("C2", 90, 30),
+        ]
+        result = [ln.text for ln in sort_lines_by_layout(lines)]
+        assert result == ["A1", "B1", "C1", "A2", "B2", "C2"]
+
+    def test_groups_slightly_misaligned_cells_into_one_band(self) -> None:
+        # Real OCR boxes for a single visual row are rarely perfectly
+        # y-aligned; we still want them grouped.
+        lines = [
+            _make("LEFT", 10, 10),
+            _make("MID", 50, 12),  # 2px below LEFT — same row visually
+            _make("RIGHT", 90, 11),
+        ]
+        result = [ln.text for ln in sort_lines_by_layout(lines)]
+        assert result == ["LEFT", "MID", "RIGHT"]
+
+    def test_separates_rows_when_y_gap_exceeds_threshold(self) -> None:
+        # Lines with a y gap larger than ~½ line-height must NOT collapse
+        # into the same band.
+        lines = [
+            _make("ROW1A", 10, 10),
+            _make("ROW1B", 50, 10),
+            _make("ROW2A", 10, 30),  # gap of 20 vs height 15 → new band
+            _make("ROW2B", 50, 30),
+        ]
+        result = [ln.text for ln in sort_lines_by_layout(lines)]
+        assert result == ["ROW1A", "ROW1B", "ROW2A", "ROW2B"]
+
+    def test_ocrpage_text_uses_sorted_order(self) -> None:
+        lines = [
+            _make("RIGHT", 90, 10),
+            _make("LEFT", 10, 10),
+            _make("BOTTOM", 10, 30),
+        ]
+        page = OCRPage(lines=lines)
+        assert page.text == "LEFT\nRIGHT\nBOTTOM"