File size: 31,792 Bytes
2523bef
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
"""
Copyright (C) 2023 Artifex Software, Inc.

This file is part of PyMuPDF.

PyMuPDF is free software: you can redistribute it and/or modify it under the
terms of the GNU Affero General Public License as published by the Free
Software Foundation, either version 3 of the License, or (at your option)
any later version.

PyMuPDF is distributed in the hope that it will be useful, but WITHOUT ANY
WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
FOR A PARTICULAR PURPOSE. See the GNU Affero General Public License for more
details.

You should have received a copy of the GNU Affero General Public License
along with MuPDF. If not, see <https://www.gnu.org/licenses/agpl-3.0.en.html>

Alternative licensing terms are available from the licensor.
For commercial licensing, see <https://www.artifex.com/> or contact
Artifex Software, Inc., 39 Mesa Street, Suite 108A, San Francisco,
CA 94129, USA, for further information.

---------------------------------------------------------------------

PyMuPDF table grid refinement (opt-in extension).

Provides the refine_grid / refine_grid_structure / refine_grid_rows API that
refines a detected table's cell grid using page text and vector-graphics
geometry. Re-exported by pymupdf.table; never runs on the default
find_tables() path.
"""

import itertools
import re
import pymupdf


# Grid refinement runs three splitters, each taking (page, grid) and returning a
# grid without mutating the page:
#   * _split_shaded_rows            split rows the line grid merged but a cell
#                                   background-shading rectangle separates,
#   * _split_undersegmented_columns split a column that jams several values into
#                                   one cell,
#   * _split_overmerged_rows        split body rows that collapsed several
#                                   records into a single grid row.
# Word selection uses center-point cell membership with rotated/vertical span
# substitution; it is independent of the CHARS/extract_words path in
# pymupdf.table, so refinement needs no CHARS state.

_REFINE_LINE_GAP = 3.0  # center-y gap (points) that groups body words into lines


# --- word selection: center-point membership + rotated-span substitution -----
def _refine_rawdict_spans(page):
    """Flattened rawdict text spans carrying line-direction metadata.

    Only used to locate vertical/rotated text so those words can be replaced by
    span-level bboxes in _refine_page_words. Minimal fields (bbox/text/dir/wmode)
    -- the only ones the consumers read."""
    spans = []
    raw = page.get_text("rawdict")
    for block in raw.get("blocks", []):
        if block.get("type") != 0:
            continue
        for line in block.get("lines", []):
            direction = line.get("dir")
            wmode = line.get("wmode")
            for span in line.get("spans", []):
                text = "".join(str(char.get("c", "")) for char in span.get("chars", []))
                if not text.strip():
                    continue
                spans.append(
                    {
                        "bbox": [float(v) for v in span.get("bbox", [])],
                        "text": text,
                        "dir": direction,
                        "wmode": wmode,
                    }
                )
    return spans


def _refine_is_vertical_or_rotated(span):
    """True when a rawdict span's direction indicates non-horizontal text flow.

    wmode != 0 is vertical; otherwise compare |dir_x| vs |dir_y| (a missing or
    unparsable direction counts as horizontal)."""
    if span.get("wmode") not in (None, 0):
        return True
    direction = span.get("dir")
    if not isinstance(direction, (list, tuple)) or len(direction) < 2:
        return False
    try:
        dx = abs(float(direction[0]))
        dy = abs(float(direction[1]))
    except (TypeError, ValueError):
        return False
    return dy > dx


def _refine_page_words(page):
    """Center-point word list for grid refinement, cached on the page object.

    Extract page words once (page.get_text("words")), then substitute vertical/
    rotated text: any horizontal word whose center falls inside a rotated span's
    bbox is dropped, and each rotated span is appended as a single word."""
    cache = getattr(page, "_refine_words_cache", None)
    if cache is not None:
        return cache
    words = [
        (float(w[0]), float(w[1]), float(w[2]), float(w[3]), str(w[4]))
        for w in page.get_text("words")
        if str(w[4]).strip()
    ]
    rotated_spans = [
        span
        for span in _refine_rawdict_spans(page)
        if _refine_is_vertical_or_rotated(span) and str(span.get("text") or "").strip()
    ]
    if rotated_spans:
        rotated_rects = [pymupdf.Rect(span.get("bbox") or []) for span in rotated_spans]
        kept = []
        for word in words:
            wx0, wy0, wx1, wy1, _ = word
            center = pymupdf.Point((wx0 + wx1) * 0.5, (wy0 + wy1) * 0.5)
            if any(rect.contains(center) for rect in rotated_rects if not rect.is_empty):
                continue
            kept.append(word)
        for span, rect in zip(rotated_spans, rotated_rects):
            if rect.is_empty:
                continue
            kept.append(
                (float(rect.x0), float(rect.y0), float(rect.x1), float(rect.y1), str(span["text"]))
            )
        words = sorted(kept, key=lambda item: (item[1], item[0]))
    try:
        setattr(page, "_refine_words_cache", words)
    except Exception:
        pass
    return words


def _refine_word_in_rect(wx0, wy0, wx1, wy1, x0, y0, x1, y1):
    """Word-to-cell membership by CENTER point (inclusive), not clip-overlap.

    A word straddling a column line is claimed by exactly one cell."""
    cx = (wx0 + wx1) * 0.5
    cy = (wy0 + wy1) * 0.5
    return x0 <= cx <= x1 and y0 <= cy <= y1


def _refine_words_in_rect(page, rect):
    """Page words whose center lies in rect, as (x0, y0, x1, y1, text) tuples."""
    x0, y0, x1, y1 = float(rect.x0), float(rect.y0), float(rect.x1), float(rect.y1)
    return [
        (wx0, wy0, wx1, wy1, text)
        for wx0, wy0, wx1, wy1, text in _refine_page_words(page)
        if _refine_word_in_rect(wx0, wy0, wx1, wy1, x0, y0, x1, y1)
    ]


def _refine_has_digit(text):
    """Loose value-like predicate: the text contains any digit."""
    return bool(re.search(r"\d", text.strip()))


def _refine_has_digit_no_alpha(text):
    """Strict value-like predicate: contains a digit and no letters."""
    stripped = text.strip()
    return bool(re.search(r"\d", stripped)) and not any(char.isalpha() for char in stripped)


# --- stage 1: split rows separated by cell background shading ----------------
def _refine_is_white(fill, *, threshold=0.95):
    if fill is None:
        return True
    try:
        r, g, b = fill[:3]
    except Exception:
        return True
    return float(r) > threshold and float(g) > threshold and float(b) > threshold


def _refine_table_rect(cells, table_bbox):
    if table_bbox is not None:
        rect = pymupdf.Rect(table_bbox)
        if not rect.is_empty:
            return rect
    rects = [
        pymupdf.Rect(cell)
        for row in cells
        for cell in row
        if cell is not None and not pymupdf.Rect(cell).is_empty
    ]
    if not rects:
        return None
    rect = pymupdf.Rect(rects[0])
    for item in rects[1:]:
        rect.include_rect(item)
    return rect


def _refine_raw_shaded_rects(page, table_rect, *, min_dim):
    out = []
    page_width = float(page.rect.width)
    for drawing in page.get_drawings():
        if _refine_is_white(drawing.get("fill")):
            continue
        for item in drawing.get("items", []):
            if item[0] != "re":
                continue
            rect = pymupdf.Rect(item[1])
            rect.normalize()
            if rect.width >= 0.95 * page_width:
                continue
            if rect.width < min_dim or rect.height < min_dim:
                continue
            if (rect & table_rect).is_empty:
                continue
            out.append(rect)
    return out


def _refine_cell_background_rects(raw_rects, *, contain_frac=0.9):
    keep = []
    for index, rect in enumerate(raw_rects):
        area = rect.get_area()
        if area <= 0:
            continue
        nested = any(
            other_index != index
            and (rect & other).get_area() >= contain_frac * area
            and other.get_area() > area * 1.05
            for other_index, other in enumerate(raw_rects)
        )
        if not nested:
            keep.append(rect)
    return keep


def _refine_cluster(values, *, tolerance):
    clusters = []
    current = []
    for value in sorted(values):
        if current and value - current[-1] > tolerance:
            clusters.append(sum(current) / len(current))
            current = []
        current.append(value)
    if current:
        clusters.append(sum(current) / len(current))
    return clusters


def _refine_border_lines(page, table_rect):
    xs = set()
    ys = set()
    for drawing in page.get_drawings():
        stroked = drawing.get("type") in ("s", "fs")
        for item in drawing.get("items", []):
            kind = item[0]
            if kind == "l":
                p1, p2 = item[1], item[2]
                if (
                    abs(p1.y - p2.y) < 1.0
                    and abs(p1.x - p2.x) > 10.0
                    and table_rect.y0 - 2.0 < (p1.y + p2.y) * 0.5 < table_rect.y1 + 2.0
                ):
                    ys.add(round((p1.y + p2.y) * 0.5, 1))
                elif (
                    abs(p1.x - p2.x) < 1.0
                    and abs(p1.y - p2.y) > 10.0
                    and table_rect.x0 - 2.0 < (p1.x + p2.x) * 0.5 < table_rect.x1 + 2.0
                ):
                    xs.add(round((p1.x + p2.x) * 0.5, 1))
                continue
            if kind != "re":
                continue
            rect = pymupdf.Rect(item[1])
            rect.normalize()
            if (rect & table_rect).is_empty:
                continue
            if rect.height < 2.0 and rect.width > 10.0:
                ys.add(round((rect.y0 + rect.y1) * 0.5, 1))
            elif rect.width < 2.0 and rect.height > 10.0:
                xs.add(round((rect.x0 + rect.x1) * 0.5, 1))
            elif stroked:
                ys.add(round(rect.y0, 1))
                ys.add(round(rect.y1, 1))
                xs.add(round(rect.x0, 1))
                xs.add(round(rect.x1, 1))
    return sorted(xs), sorted(ys)


def _refine_drop_near(edges, borders, *, tolerance):
    return [edge for edge in edges if not any(abs(edge - border) <= tolerance for border in borders)]


def _refine_has_text(words, x0, y0, x1, y1, *, margin):
    for wx0, wy0, wx1, wy1, text in words:
        if not str(text).strip():
            continue
        cx = (wx0 + wx1) * 0.5
        cy = (wy0 + wy1) * 0.5
        if x0 + margin < cx < x1 - margin and y0 + margin < cy < y1 - margin:
            return True
    return False


def _refine_content_filter_row_edges(words, y0, y1, candidates, x0, x1, *, margin):
    edges = sorted(edge for edge in candidates if y0 + margin < edge < y1 - margin)
    while edges:
        bounds = [y0, *edges, y1]
        empty_index = None
        for index in range(len(bounds) - 1):
            if not _refine_has_text(words, x0, bounds[index], x1, bounds[index + 1], margin=margin):
                empty_index = index
                break
        if empty_index is None:
            break
        if empty_index < len(edges):
            edges.pop(empty_index)
        elif empty_index - 1 >= 0:
            edges.pop(empty_index - 1)
        else:
            break
    return edges


def _split_shaded_rows(
    page,
    cells,
    table_bbox=None,
    *,
    cluster_tolerance=3.0,
    margin=2.0,
    min_dim=6.0,
    border_tolerance=2.5,
):
    """Split rows the line grid merged but a cell background rectangle separates.

    Cell background-shading rectangles give row boundaries the border lines do
    not; candidate y-edges from those rectangles (minus edges that coincide with
    real borders, minus edges that would cut an empty band) subdivide each row."""
    table_rect = _refine_table_rect(cells, table_bbox)
    if table_rect is None or not cells:
        return cells

    raw_rects = _refine_raw_shaded_rects(page, table_rect, min_dim=min_dim)
    background_rects = _refine_cell_background_rects(raw_rects)
    if not background_rects:
        return cells

    y_edges = _refine_cluster(
        [edge for rect in background_rects for edge in (float(rect.y0), float(rect.y1))],
        tolerance=cluster_tolerance,
    )
    _border_xs, border_ys = _refine_border_lines(page, table_rect)
    y_edges = _refine_drop_near(y_edges, border_ys, tolerance=border_tolerance)
    if not y_edges:
        return cells

    words = _refine_page_words(page)
    new_cells = []
    added_rows = 0
    for row in cells:
        live = [pymupdf.Rect(cell) for cell in row if cell is not None]
        if not live:
            new_cells.append(row)
            continue
        ry0 = min(float(cell.y0) for cell in live)
        ry1 = max(float(cell.y1) for cell in live)
        rx0 = min(float(cell.x0) for cell in live)
        rx1 = max(float(cell.x1) for cell in live)
        cuts = _refine_content_filter_row_edges(words, ry0, ry1, y_edges, rx0, rx1, margin=margin)
        if not cuts:
            new_cells.append(row)
            continue
        bounds = [ry0, *cuts, ry1]
        for band_index in range(len(bounds) - 1):
            band = []
            for cell in row:
                if cell is None:
                    band.append(None)
                else:
                    rect = pymupdf.Rect(cell)
                    band.append(
                        [float(rect.x0), bounds[band_index], float(rect.x1), bounds[band_index + 1]]
                    )
            new_cells.append(band)
        added_rows += len(bounds) - 2

    if added_rows <= 0:
        return cells
    return new_cells


# --- stage 2: split under-segmented columns ----------------------------------
def _refine_value_like(text):
    return _refine_has_digit_no_alpha(text)


def _refine_span_value_like(text):
    return _refine_has_digit(text)


def _refine_cell_span_signal(page, cell, *, gap, line_tolerance):
    """True if a cell holds two value-like groups separated by a wide gap on one
    text line -- the signal that a single grid column packs several columns."""
    rect = pymupdf.Rect(cell)
    words = []
    for x0, y0, x1, y1, text in _refine_words_in_rect(page, rect):
        stripped = str(text).strip()
        if not stripped:
            continue
        cx = (float(x0) + float(x1)) / 2.0
        cy = (float(y0) + float(y1)) / 2.0
        if rect.x0 <= cx <= rect.x1 and rect.y0 <= cy <= rect.y1:
            words.append((float(x0), float(x1), cy, stripped))
    if len(words) < 2:
        return False

    words.sort(key=lambda item: item[2])
    lines = []
    current = [words[0]]
    for word in words[1:]:
        if word[2] - current[-1][2] > line_tolerance:
            lines.append(current)
            current = [word]
        else:
            current.append(word)
    lines.append(current)

    for line in lines:
        line.sort(key=lambda item: item[0])
        for index in range(len(line) - 1):
            if line[index + 1][0] - line[index][1] <= gap:
                continue
            left = " ".join(word[3] for word in line[: index + 1])
            right = " ".join(word[3] for word in line[index + 1 :])
            if _refine_span_value_like(left) and _refine_span_value_like(right):
                return True
    return False


def _refine_detect_span_columns(
    page, cells, *, gap=12.0, line_tolerance=4.0, support_ratio=0.3, min_support=2
):
    """Return the column indices that fire the under-segmentation gate: columns
    where enough rows carry the two-value-group signal (a support threshold)."""
    rows = len(cells)
    support_threshold = max(min_support, int(support_ratio * rows))
    col_hits = {}
    for row in cells:
        for col_index, cell in enumerate(row):
            if cell is None:
                continue
            if _refine_cell_span_signal(page, cell, gap=gap, line_tolerance=line_tolerance):
                col_hits[col_index] = col_hits.get(col_index, 0) + 1
    return sorted(col for col, count in col_hits.items() if count >= support_threshold)


def _refine_column_words(page, cells, col):
    rows = []
    for row in cells:
        words = []
        if col < len(row) and row[col] is not None:
            rect = pymupdf.Rect(row[col])
            for x0, y0, x1, y1, text in _refine_words_in_rect(page, rect):
                if not str(text).strip():
                    continue
                cx = (float(x0) + float(x1)) / 2.0
                cy = (float(y0) + float(y1)) / 2.0
                if rect.x0 <= cx <= rect.x1 and rect.y0 <= cy <= rect.y1:
                    words.append((float(x0), float(x1), str(text)))
        rows.append(sorted(words))
    return rows


def _refine_cut_xs(rows_words, *, bridge_tolerance=0):
    intervals = [(x0, x1, row_index) for row_index, words in enumerate(rows_words) for (x0, x1, _) in words]
    if len(intervals) < 2:
        return []

    lo = min(x0 for x0, _, _ in intervals)
    hi = max(x1 for _, x1, _ in intervals)
    if hi - lo < 1:
        return []

    channels = []
    current = None
    x = lo
    while x <= hi:
        crossing_rows = {row_index for x0, x1, row_index in intervals if x0 < x < x1}
        if len(crossing_rows) > bridge_tolerance:
            if current is not None:
                channels.append((current[0], current[1]))
                current = None
        else:
            current = [x, x] if current is None else [current[0], x]
        x += 1.0
    if current is not None:
        channels.append((current[0], current[1]))

    cuts = []
    for left_edge, right_edge in channels:
        if left_edge <= lo + 1 or right_edge >= hi - 1:
            continue
        cut_x = (left_edge + right_edge) / 2.0
        crossing_count = sum(1 for words in rows_words if any(x0 < cut_x < x1 for x0, x1, _ in words))
        # Evaluate the symmetry/value guards on non-bridging rows only, so a
        # merged label spanning value subcolumns does not block column recovery.
        body_rows = [words for words in rows_words if not any(x0 < cut_x < x1 for x0, x1, _ in words)]
        left_rows = right_rows = 0
        left_ok = right_ok = 0
        left_n = right_n = 0
        left_values = []
        right_values = []
        for words in body_rows:
            left = [text for x0, x1, text in words if x1 <= cut_x]
            right = [text for x0, x1, text in words if x0 >= cut_x]
            if left:
                left_rows += 1
                left_n += 1
                left_ok += _refine_value_like(" ".join(left))
                left_values.extend(left)
            if right:
                right_rows += 1
                right_n += 1
                right_ok += _refine_value_like(" ".join(right))
                right_values.extend(right)

        if left_rows < 2 or right_rows < 2:
            continue
        if crossing_count > 0:
            if not _refine_value_like(" ".join(left_values)) or not _refine_value_like(" ".join(right_values)):
                continue
        elif (left_n and left_ok / left_n < 0.5) or (right_n and right_ok / right_n < 0.5):
            continue
        cuts.append(cut_x)

    return sorted(cuts)


def _refine_split_columns(page, cells, *, bridge_tolerance=0, allowed_cols=None):
    ncols = max((len(row) for row in cells), default=0)
    active_cols = None if allowed_cols is None else set(allowed_cols)
    col_cuts = {}
    for col in range(ncols):
        if active_cols is not None and col not in active_cols:
            continue
        cuts = _refine_cut_xs(_refine_column_words(page, cells, col), bridge_tolerance=bridge_tolerance)
        if cuts:
            col_cuts[col] = cuts

    if not col_cuts:
        return cells

    new_cells = []
    for row in cells:
        new_row = []
        for col, cell in enumerate(row):
            cuts = col_cuts.get(col)
            if not cuts:
                new_row.append(cell)
            elif cell is None:
                new_row.extend([None] * (len(cuts) + 1))
            else:
                x0, y0, x1, y1 = pymupdf.Rect(cell)
                xs = [x0] + [cut_x for cut_x in cuts if x0 < cut_x < x1] + [x1]
                for index in range(len(xs) - 1):
                    new_row.append([xs[index], y0, xs[index + 1], y1])
        new_cells.append(new_row)

    return new_cells


def _split_undersegmented_columns(
    page,
    cells,
    *,
    gap=12.0,
    line_tolerance=4.0,
    support_ratio=0.3,
    min_support=2,
    bridge_tolerance=0,
):
    """Split under-segmented columns, but only those that fire the value-group
    gate -- so ordinary single-value columns are never disturbed."""
    fired_cols = _refine_detect_span_columns(
        page,
        cells,
        gap=gap,
        line_tolerance=line_tolerance,
        support_ratio=support_ratio,
        min_support=min_support,
    )
    if not fired_cols:
        return cells
    return _refine_split_columns(
        page, cells, bridge_tolerance=bridge_tolerance, allowed_cols=set(fired_cols)
    )


# --- stage 3: split over-merged body rows ------------------------------------
def _refine_rects_from_cells(cells):
    return [pymupdf.Rect(cell) for row in cells for cell in row if cell is not None]


def _refine_union_rect(rects):
    if not rects:
        return None
    rect = pymupdf.Rect(rects[0])
    for item in rects[1:]:
        rect.include_rect(item)
    return rect


def _refine_best_column_bounds(cells):
    best_row = max(
        cells,
        key=lambda row: (sum(1 for cell in row if cell is not None), len(row)),
        default=[],
    )
    bounds = sorted(
        (float(rect.x0), float(rect.x1))
        for cell in best_row
        if cell is not None
        for rect in [pymupdf.Rect(cell)]
        if rect.x1 > rect.x0
    )
    return bounds


def _refine_cluster_words_by_y(words):
    """Group center-point words into body "lines" by a fixed center-y gap,
    returning each line's union bbox and joined text."""
    if not words:
        return []
    words = sorted(words, key=lambda item: ((item[1] + item[3]) / 2.0, item[0]))
    clusters = [[words[0]]]
    last_center = (words[0][1] + words[0][3]) / 2.0
    for word in words[1:]:
        center = (word[1] + word[3]) / 2.0
        if center - last_center > _REFINE_LINE_GAP:
            clusters.append([word])
        else:
            clusters[-1].append(word)
        last_center = center

    lines = []
    for cluster in clusters:
        lines.append(
            {
                "x0": min(word[0] for word in cluster),
                "y0": min(word[1] for word in cluster),
                "x1": max(word[2] for word in cluster),
                "y1": max(word[3] for word in cluster),
                "text": " ".join(word[4] for word in sorted(cluster, key=lambda item: item[0])),
            }
        )
    return lines


def _refine_line_columns(line, col_bounds):
    cols = set()
    line_x0 = float(line["x0"])
    line_x1 = float(line["x1"])
    center = (line_x0 + line_x1) / 2.0
    for index, (x0, x1) in enumerate(col_bounds):
        overlap = min(line_x1, x1) - max(line_x0, x0)
        if overlap > 0 or x0 <= center <= x1:
            cols.add(index)
    return cols


def _refine_page_words_in_rect(page, rect):
    words = []
    rx0, ry0, rx1, ry1 = float(rect.x0), float(rect.y0), float(rect.x1), float(rect.y1)
    for x0, y0, x1, y1, text in _refine_page_words(page):
        cx = (x0 + x1) * 0.5
        cy = (y0 + y1) * 0.5
        if rx0 <= cx <= rx1 and ry0 <= cy <= ry1:
            text = str(text).strip()
            if text:
                words.append((float(x0), float(y0), float(x1), float(y1), text))
    return words


def _refine_merge_overlapping_lines(lines, *, frac):
    if not lines:
        return []
    items = sorted(lines)
    merged = [[float(items[0][0]), float(items[0][1])]]
    for y0, y1 in items[1:]:
        current = merged[-1]
        overlap = current[1] - float(y0)
        min_height = min(current[1] - current[0], float(y1) - float(y0))
        if min_height > 0 and overlap / min_height >= frac:
            current[0] = min(current[0], float(y0))
            current[1] = max(current[1], float(y1))
        else:
            merged.append([float(y0), float(y1)])
    return [(y0, y1) for y0, y1 in merged]


def _refine_detect_row_overmerge(page, cells, *, clean_threshold=0.85, header_row_count=1):
    """Decide whether the body rows are over-merged. Returns the geometry needed
    to re-cut them (body bbox, column bounds, per-record y-bounds) or None.

    Fires only when the body text clusters into clean multi-column "record" lines
    (clean_ratio >= clean_threshold) and there are more records than existing body
    rows -- i.e. several records collapsed into one grid row."""
    col_bounds = _refine_best_column_bounds(cells)
    header_row_count = max(1, min(int(header_row_count), len(cells))) if cells else 0
    existing_body_rows = max(0, len(cells) - header_row_count)
    if len(cells) <= header_row_count:
        return None
    if len(col_bounds) < 2:
        return None

    body_rect = _refine_union_rect(_refine_rects_from_cells(cells[header_row_count:]))
    if body_rect is None or body_rect.is_empty:
        return None

    words = _refine_page_words_in_rect(page, body_rect)
    lines = _refine_cluster_words_by_y(words)
    if not lines:
        return None

    min_cols = max(2, (len(col_bounds) + 1) // 2)  # == max(2, ceil(len/2))
    record_lines = []
    for line in lines:
        cols = sorted(_refine_line_columns(line, col_bounds))
        if len(cols) >= min_cols:
            record_lines.append(line)

    clean_ratio = len(record_lines) / len(lines)
    if clean_ratio < clean_threshold:
        return None
    if len(record_lines) <= existing_body_rows:
        return None

    return {
        "body_bbox": [body_rect.x0, body_rect.y0, body_rect.x1, body_rect.y1],
        "col_bounds": [[round(x0, 2), round(x1, 2)] for x0, x1 in col_bounds],
        "record_line_bounds": [[float(line["y0"]), float(line["y1"])] for line in record_lines],
        "existing_body_rows": existing_body_rows,
        "header_row_count": header_row_count,
    }


def _split_overmerged_rows(
    page, cells, *, clean_threshold=0.85, merge_overlap_frac=0.35, header_row_count=1
):
    """Re-cut over-merged body rows into one grid row per detected record.

    Keeps the header rows intact, then rebuilds the body as evenly-bounded rows
    (cut at the midpoints between consecutive record centers) across the detected
    column bounds. header_row_count is the number of leading header rows to keep."""
    meta = _refine_detect_row_overmerge(
        page, cells, clean_threshold=clean_threshold, header_row_count=header_row_count
    )
    if meta is None:
        return cells

    body_bbox = pymupdf.Rect(meta["body_bbox"])
    col_bounds = [(float(x0), float(x1)) for x0, x1 in meta["col_bounds"]]
    raw_line_bounds = [(float(y0), float(y1)) for y0, y1 in meta["record_line_bounds"]]
    line_bounds = _refine_merge_overlapping_lines(raw_line_bounds, frac=merge_overlap_frac)
    if len(line_bounds) <= int(meta["existing_body_rows"]):
        return cells

    centers = [(y0 + y1) / 2.0 for y0, y1 in line_bounds]
    edges = [float(body_bbox.y0)]
    for prev, nxt in zip(centers, centers[1:]):
        edges.append((prev + nxt) / 2.0)
    edges.append(float(body_bbox.y1))

    header_rows = cells[: int(meta["header_row_count"])]
    new_body = []
    for row_index in range(len(line_bounds)):
        y0 = edges[row_index]
        y1 = edges[row_index + 1]
        new_row = []
        for x0, x1 in col_bounds:
            rect = pymupdf.Rect(x0, y0, x1, y1)
            new_row.append([float(rect.x0), float(rect.y0), float(rect.x1), float(rect.y1)])
        new_body.append(new_row)

    return [*header_rows, *new_body]


# --- public seam -------------------------------------------------------------
def refine_grid_structure(page, cells, *, table_bbox=None):
    """Structural half of refine_grid: split shaded rows, then under-segmented
    columns. Exposed separately so a caller that must know the header/body
    boundary of the post-structure grid can insert that computation between the
    two phases. table_bbox, when given, bounds the shaded-rectangle search;
    otherwise the cells' union is used."""
    cells = _split_shaded_rows(page, cells, table_bbox)
    cells = _split_undersegmented_columns(page, cells)
    return cells


def refine_grid_rows(page, cells, *, header_row_count=1, clean_threshold=0.85, merge_overlap_frac=0.35):
    """Row half of refine_grid: split body rows that collapsed several records
    into one grid row. header_row_count is the number of leading header rows to
    keep intact; callers that resolve a header region pass it in, and the default
    of 1 is a conservative single-header assumption."""
    return _split_overmerged_rows(
        page,
        cells,
        clean_threshold=clean_threshold,
        merge_overlap_frac=merge_overlap_frac,
        header_row_count=header_row_count,
    )


def refine_grid(page, cells, *, table_bbox=None, header_row_count=1):
    """Refine a detected table's cell grid and return the refined grid.

    ``cells`` is a row-major grid: a list of rows, each a list of ``[x0, y0, x1,
    y1]`` cell rectangles (``None`` for a gap). The grid is refined in three
    stages -- split rows that cell background shading separates, split columns
    that jam several values into one cell, and split body rows that merged
    several records -- using the page's words and vector graphics. The result is
    a new grid in the same format (rows may be added and rows may widen).

    ``table_bbox`` optionally bounds the table (else the cells' union is used).
    ``header_row_count`` is the number of leading header rows to preserve when
    re-cutting body rows (default 1). This is the all-in-one convenience wrapper;
    a caller needing the header boundary of the intermediate grid can instead
    call :func:`refine_grid_structure` then :func:`refine_grid_rows`.
    """
    cells = refine_grid_structure(page, cells, table_bbox=table_bbox)
    cells = refine_grid_rows(page, cells, header_row_count=header_row_count)
    return cells


def _refine_cells_to_grid(cells):
    """Convert a Table's flat cell list into the row-major grid refine_grid wants.

    Mirrors Table.rows: columns are the sorted distinct left edges, rows are the
    cells grouped by top edge, each padded with None where a column has no cell."""
    if not cells:
        return []
    xs = sorted({c[0] for c in cells})
    col_index = {x: i for i, x in enumerate(xs)}
    grid = []
    ordered = sorted(cells, key=lambda c: (c[1], c[0]))
    for _y, row_cells in itertools.groupby(ordered, key=lambda c: c[1]):
        row = [None] * len(xs)
        for cell in row_cells:
            row[col_index[cell[0]]] = [float(cell[0]), float(cell[1]), float(cell[2]), float(cell[3])]
        grid.append(row)
    return grid


def _refine_grid_to_cells(grid):
    """Flatten a refined grid back to a Table's flat (x0, y0, x1, y1) cell list."""
    return [
        (float(cell[0]), float(cell[1]), float(cell[2]), float(cell[3]))
        for row in grid
        for cell in row
        if cell is not None
    ]