pdf_oxide 0.3.59

The fastest Rust PDF library with text extraction: 0.8ms mean, 100% pass rate on 3,830 PDFs. 5× faster than pdf_extract, 17× faster than oxidize_pdf. Extract, create, and edit PDFs.
Documentation
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
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
//! PDF → DOCX with **layout-preserving** text frames.
//!
//! Each PDF text span becomes a `<w:p>` paragraph wrapped in a
//! `<w:framePr>` with absolute coordinates anchored to the page. Word /
//! LibreOffice / OnlyOffice all honor this and render the result
//! visually similar to the source PDF — same fonts, sizes, colors, and
//! positions, but the text remains real selectable / editable text
//! (unlike the rasterization approach which produces pixel-only output).
//!
//! Trade-offs:
//!   - Works well for text-heavy documents with simple layouts
//!     (academic papers, forms, single/multi-column reports).
//!   - Doesn't reconstruct table grids, vector graphics, or embedded
//!     images (those would need additional work).
//!   - Word's frame engine handles positioning; large numbers of
//!     overlapping frames can be slow to lay out.
//!
//! The output DOCX is a flat zip containing exactly the parts Word needs:
//! `[Content_Types].xml`, `_rels/.rels`, `word/document.xml`, and
//! `word/_rels/document.xml.rels`.

use crate::error::{Error, Result};
use crate::layout::text_block::TextSpan;
use office_oxide::core::opc::{OpcWriter, PartName};
use office_oxide::core::relationships::rel_types;
use std::collections::HashMap;
use std::io::Cursor;

/// Twips per point. 1 pt = 20 twips (Word's basic unit).
const TWIPS_PER_PT: f32 = 20.0;
/// EMUs per point. 1 pt = 12 700 EMU.
const EMU_PER_PT: f64 = 12_700.0;

/// Convert a `PdfDocument` to layout-preserving DOCX bytes.
///
/// Iterates every page, extracts text spans with their bounding boxes,
/// and emits each as a positioned `w:framePr` paragraph in `word/document.xml`.
/// The DOCX page size is set to match the source PDF so frame coordinates
/// (in twips, 1/20 pt) map 1:1.
pub fn to_docx_bytes_layout(doc: &crate::document::PdfDocument) -> Result<Vec<u8>> {
    let n_pages = doc.page_count()?;
    if n_pages == 0 {
        return Err(Error::InvalidOperation("PDF has zero pages".into()));
    }

    // Build a per-page resource-name → real-face-name lookup so spans
    // referencing PDF font resource ids (e.g. "F75") can be emitted with
    // the actual font face that ships in the source PDF (e.g.
    // "TeXGyreTermesX-Regular"). `extract_spans` returns the resource id;
    // `page_font_face_lookups` resolves those to BaseFont names.
    let lookups = doc.page_font_face_lookups().unwrap_or_default();

    // Initialize OPC writer up-front so we can reserve image rIds in order
    // and use them in the body XML below.
    let cursor = Cursor::new(Vec::new());
    let mut opc =
        OpcWriter::new(cursor).map_err(|e| Error::InvalidOperation(format!("opc init: {e}")))?;
    let doc_part = PartName::new("/word/document.xml")
        .map_err(|e| Error::InvalidOperation(format!("part name: {e}")))?;
    opc.add_package_rel(rel_types::OFFICE_DOCUMENT, "word/document.xml");

    let mut pages: Vec<PageSpans> = Vec::with_capacity(n_pages);
    let mut media_idx = 0usize;
    for i in 0..n_pages {
        let (x1, y1, x2, y2) = doc.get_page_media_box(i)?;
        let w_pt = (x2 - x1).abs();
        let h_pt = (y2 - y1).abs();
        let mut spans = doc.extract_spans(i).unwrap_or_default();
        // Drop rotated text (page-edge watermarks like the arxiv
        // ID, rotated table headers). TextSpan carries no rotation
        // field but `extract_chars` does; identify rotated spans by
        // overlapping their bbox with rotated chars and drop them.
        // Mirrors the same filter in `pdf_to_ir`. Without this the
        // arxiv watermark renders as a horizontal text frame in the
        // wrong place.
        if let Ok(chars) = doc.extract_chars(i) {
            let chars_horizontal_dominant = if chars.is_empty() {
                true
            } else {
                let horiz = chars
                    .iter()
                    .filter(|c| c.rotation_degrees.abs() < 5.0)
                    .count();
                horiz * 4 >= chars.len() * 3
            };
            spans.retain(|s| {
                !crate::converters::pdf_to_ir::span_overlaps_rotated_chars(
                    s,
                    &chars,
                    chars_horizontal_dominant,
                )
            });
        }
        // Detect music-notation regions (hymnals, sheet music). When
        // present, drop spans whose centre falls inside a music
        // region — their music-font glyphs would otherwise emit as
        // garbled Latin letters because the recipient lacks the
        // Maestro / Bravura / Sonata face. The regions are
        // rasterised below and embedded as bitmap images instead.
        let music_regions = crate::converters::music_region_finder::find_music_regions(doc, i);
        if !music_regions.is_empty() {
            spans.retain(|s| {
                !music_regions
                    .iter()
                    .any(|r| crate::converters::music_region_finder::rect_contains_bbox(r, &s.bbox))
            });
        }
        merge_hyphenated_spans(&mut spans);
        // Stamp heading_level on spans whose font size is a heading-class
        // ratio above the page's body text. Tagged-PDF struct_role would
        // be more precise but isn't currently exposed on TextSpan; the
        // ratio heuristic matches what the existing markdown / pdf_to_ir
        // converters use (1.75 / 1.35 / 1.15× modal body size).
        annotate_heading_levels(&mut spans);
        let font_lookup = lookups.get(i).cloned().unwrap_or_default();

        // Extract images on this page; for each, write the PNG to
        // `word/media/imageN.png` and register an IMAGE relationship.
        // The returned rId is what the body XML's `r:embed="…"` references.
        let mut page_images: Vec<PageImage> = Vec::new();
        if let Ok(imgs) = doc.extract_images(i) {
            for img in imgs {
                let bbox = match img.bbox() {
                    Some(b) => *b,
                    None => continue,
                };
                let png = match img.to_png_bytes() {
                    Ok(b) if !b.is_empty() => b,
                    _ => continue,
                };
                media_idx += 1;
                let target = format!("/word/media/image{}.png", media_idx);
                let part = PartName::new(&target)
                    .map_err(|e| Error::InvalidOperation(format!("part name {target}: {e}")))?;
                opc.add_part(&part, "image/png", &png)
                    .map_err(|e| Error::InvalidOperation(format!("opc add_part {target}: {e}")))?;
                let rid = opc.add_part_rel(
                    &doc_part,
                    rel_types::IMAGE,
                    &format!("media/image{}.png", media_idx),
                );
                page_images.push(PageImage { bbox, rid });
            }
        }

        // Rasterise Form-XObject + inline-image regions (academic-
        // paper figures, agency logos, accessibility marks) into
        // `word/media/` so the round-trip preserves them. Without
        // this the layout-mode DOCX path drops every figure drawn as
        // a vector Form XObject. Mirrors the supplemental pass in
        // `pdf_to_ir::extract_page_images`.
        #[cfg(feature = "rendering")]
        {
            let existing: Vec<(f32, f32, f32, f32)> = page_images
                .iter()
                .map(|pi| (pi.bbox.x, pi.bbox.y, pi.bbox.width, pi.bbox.height))
                .collect();
            let regions = crate::converters::form_xobject_finder::rasterize_form_and_inline_regions(
                doc, i, h_pt, &existing,
            );
            for ((x_pdf, y_pdf, w, h), png) in regions {
                media_idx += 1;
                let target = format!("/word/media/image{}.png", media_idx);
                let part = PartName::new(&target)
                    .map_err(|e| Error::InvalidOperation(format!("part name {target}: {e}")))?;
                opc.add_part(&part, "image/png", &png)
                    .map_err(|e| Error::InvalidOperation(format!("opc add_part {target}: {e}")))?;
                let rid = opc.add_part_rel(
                    &doc_part,
                    rel_types::IMAGE,
                    &format!("media/image{}.png", media_idx),
                );
                page_images.push(PageImage {
                    bbox: crate::geometry::Rect::new(x_pdf, y_pdf, w, h),
                    rid,
                });
            }
        }

        // Extract drawing paths and reduce to simple shapes (lines / rects).
        // Table borders, underlines, dividers, page frames all live here.
        let mut shapes: Vec<SimpleShape> = Vec::new();
        if let Ok(paths) = doc.extract_paths(i) {
            for p in paths {
                shapes.extend(simplify_path(&p));
            }
        }
        // Drop shape artefacts from `extract_paths` that would
        // wipe the DOCX page:
        // (a) Bbox dramatically larger than the page (e.g. 538×2340pt
        //     on an 841pt page). 20% bleed margin handles legitimate
        //     page-frame rects.
        // (b) Solid-black filled rect covering >25% of the page —
        //     `extract_paths` sometimes reports a near-full-page
        //     filled rect for what's actually a clipping / frame
        //     path; the source slide has a white background, so the
        //     "black" fill is a parser artefact.
        // (c) Solid-white filled rect covering >50% of the page — a
        //     few source PDFs emit a full-page white rect as the
        //     page background BEFORE the text, but the docx round-
        //     trip emits shapes AFTER text so the white rect
        //     occludes the rendered text. The page is already white
        //     in the round-trip output, so dropping this is the
        //     cleanest fix.
        let page_w_max = w_pt * 1.2;
        let page_h_max = h_pt * 1.2;
        let page_area = w_pt * h_pt;
        shapes.retain(|s| match s {
            SimpleShape::Rect { bbox, fill_rgb, .. } => {
                if bbox.width > page_w_max || bbox.height > page_h_max {
                    return false;
                }
                let area = bbox.width * bbox.height;
                if page_area <= 0.0 {
                    return true;
                }
                let area_ratio = area / page_area;
                if matches!(fill_rgb, Some((0, 0, 0))) && area_ratio > 0.25 {
                    return false;
                }
                if matches!(fill_rgb, Some((255, 255, 255))) && area_ratio > 0.5 {
                    return false;
                }
                true
            },
            _ => true,
        });

        // Suppress shapes (e.g. stroked horizontal staff lines, slurs
        // rendered as Bezier paths, beam rects) whose centre falls
        // inside a music region — they'd otherwise visually conflict
        // with the rasterised music bitmap.
        if !music_regions.is_empty() {
            shapes.retain(|s| {
                let (cx, cy) = match s {
                    SimpleShape::Line {
                        x1_pt,
                        y1_pt,
                        x2_pt,
                        y2_pt,
                        ..
                    } => ((x1_pt + x2_pt) * 0.5, (y1_pt + y2_pt) * 0.5),
                    SimpleShape::Rect { bbox, .. } => {
                        (bbox.x + bbox.width * 0.5, bbox.y + bbox.height * 0.5)
                    },
                };
                !music_regions
                    .iter()
                    .any(|r| crate::converters::music_region_finder::rect_contains_point(r, cx, cy))
            });
        }

        // Rasterise each music region as a 150 DPI PNG and emit it
        // as a positional image. This is what actually preserves
        // staves / noteheads / stems through the round-trip — the
        // span/shape suppression above just keeps garbage from
        // overlapping the bitmap.
        #[cfg(feature = "rendering")]
        if !music_regions.is_empty() {
            let regions =
                crate::converters::music_region_finder::rasterize_music_regions(doc, i, h_pt);
            for ((x_pdf, y_pdf, w, h), png) in regions {
                media_idx += 1;
                let target = format!("/word/media/image{}.png", media_idx);
                let part = PartName::new(&target)
                    .map_err(|e| Error::InvalidOperation(format!("part name {target}: {e}")))?;
                opc.add_part(&part, "image/png", &png)
                    .map_err(|e| Error::InvalidOperation(format!("opc add_part {target}: {e}")))?;
                let rid = opc.add_part_rel(
                    &doc_part,
                    rel_types::IMAGE,
                    &format!("media/image{}.png", media_idx),
                );
                page_images.push(PageImage {
                    bbox: crate::geometry::Rect::new(x_pdf, y_pdf, w, h),
                    rid,
                });
            }
        }

        pages.push(PageSpans {
            w_pt,
            h_pt,
            spans,
            font_lookup,
            images: page_images,
            shapes,
        });
    }

    let (page_w_pt, page_h_pt) = (pages[0].w_pt, pages[0].h_pt);

    let body = build_document_body(&pages, page_w_pt, page_h_pt);
    let document_xml = wrap_document(&body);

    opc.add_part(
        &doc_part,
        "application/vnd.openxmlformats-officedocument.wordprocessingml.document.main+xml",
        document_xml.as_bytes(),
    )
    .map_err(|e| Error::InvalidOperation(format!("opc add_part document.xml: {e}")))?;

    // styles.xml — minimal scaffold defining the heading styles the
    // body references via `<w:pStyle w:val="HeadingN"/>`. Without
    // this, Word silently swallows the pStyle reference and the
    // affected paragraphs render as plain text.
    let styles_part = PartName::new("/word/styles.xml")
        .map_err(|e| Error::InvalidOperation(format!("part name styles.xml: {e}")))?;
    opc.add_part_rel(&doc_part, rel_types::STYLES, "styles.xml");
    opc.add_part(
        &styles_part,
        "application/vnd.openxmlformats-officedocument.wordprocessingml.styles+xml",
        layout_styles_xml().as_bytes(),
    )
    .map_err(|e| Error::InvalidOperation(format!("opc add_part styles.xml: {e}")))?;

    // Embed the source PDF's font programs (typeface preservation),
    // and wire them up via fontTable.xml + rels so Word actually
    // picks them up. Without the manifest plumbing the TTFs ship
    // but Word silently substitutes Calibri.
    let embedded_fonts = doc.extract_embedded_fonts().unwrap_or_default();
    if !embedded_fonts.is_empty() {
        let font_table_part = PartName::new("/word/fontTable.xml")
            .map_err(|e| Error::InvalidOperation(format!("part name fontTable.xml: {e}")))?;
        opc.add_part_rel(&doc_part, rel_types::FONT_TABLE, "fontTable.xml");

        let mut font_entries: Vec<(String, String)> = Vec::with_capacity(embedded_fonts.len());
        for (idx, (name, data)) in embedded_fonts.iter().enumerate() {
            let n = idx + 1;
            let safe_name = sanitize_font_filename(name);
            let target_rel = format!("fonts/font_{n}_{safe_name}.ttf");
            let target_abs = format!("/word/fonts/font_{n}_{safe_name}.ttf");
            let part = PartName::new(&target_abs)
                .map_err(|e| Error::InvalidOperation(format!("part name {target_abs}: {e}")))?;
            opc.add_part(&part, "application/x-font-ttf", data)
                .map_err(|e| Error::InvalidOperation(format!("opc add_part {target_abs}: {e}")))?;
            let rid = opc.add_part_rel(&font_table_part, rel_types::FONT, &target_rel);
            font_entries.push((name.clone(), rid));
        }

        opc.add_part(
            &font_table_part,
            "application/vnd.openxmlformats-officedocument.wordprocessingml.fontTable+xml",
            layout_font_table_xml(&font_entries).as_bytes(),
        )
        .map_err(|e| Error::InvalidOperation(format!("opc add_part fontTable.xml: {e}")))?;
    }

    let result = opc
        .finish()
        .map_err(|e| Error::InvalidOperation(format!("opc finish: {e}")))?;
    Ok(result.into_inner())
}

/// Minimal `word/styles.xml` shipped by the layout writer. Defines
/// `Normal` and `Heading1..6` so the body's `<w:pStyle>` references
/// resolve. Run-property defaults are intentionally bare — the
/// layout writer already inlines `<w:rPr>` per run with explicit
/// font/size/colour, so heading styling is driven by those, not by
/// the stylesheet.
fn layout_styles_xml() -> String {
    let mut out = String::new();
    out.push_str(r#"<?xml version="1.0" encoding="UTF-8" standalone="yes"?>"#);
    out.push_str(
        r#"<w:styles xmlns:w="http://schemas.openxmlformats.org/wordprocessingml/2006/main">"#,
    );
    out.push_str(
        r#"<w:style w:type="paragraph" w:styleId="Normal"><w:name w:val="Normal"/></w:style>"#,
    );
    for level in 1..=6 {
        let outline = level - 1;
        out.push_str(&format!(
            concat!(
                r#"<w:style w:type="paragraph" w:styleId="Heading{lvl}">"#,
                r#"<w:name w:val="heading {lvl}"/>"#,
                r#"<w:basedOn w:val="Normal"/>"#,
                r#"<w:pPr><w:outlineLvl w:val="{ol}"/></w:pPr>"#,
                r#"</w:style>"#
            ),
            lvl = level,
            ol = outline,
        ));
    }
    out.push_str(r#"</w:styles>"#);
    out
}

/// Build `word/fontTable.xml` for the layout writer. Lists each
/// embedded font with `<w:embedRegular r:id="rIdN"/>` so Word
/// recognises the package's font programs.
fn layout_font_table_xml(entries: &[(String, String)]) -> String {
    let mut out = String::new();
    out.push_str(r#"<?xml version="1.0" encoding="UTF-8" standalone="yes"?>"#);
    out.push_str(r#"<w:fonts xmlns:w="http://schemas.openxmlformats.org/wordprocessingml/2006/main" xmlns:r="http://schemas.openxmlformats.org/officeDocument/2006/relationships">"#);
    for (name, rid) in entries {
        out.push_str(&format!(
            r#"<w:font w:name="{name}"><w:embedRegular r:id="{rid}"/></w:font>"#,
            name = xml_escape(name),
            rid = rid,
        ));
    }
    out.push_str(r#"</w:fonts>"#);
    out
}

fn xml_escape(s: &str) -> String {
    s.replace('&', "&amp;")
        .replace('<', "&lt;")
        .replace('>', "&gt;")
        .replace('"', "&quot;")
}

/// Build the inner contents of `<w:body>` — one positioned paragraph per
/// span, plus a final `<w:sectPr>` for page geometry.
fn build_document_body(pages: &[PageSpans], page_w_pt: f32, page_h_pt: f32) -> String {
    let mut out = String::with_capacity(64 * 1024);
    for (pi, page) in pages.iter().enumerate() {
        // Page break before every page except the first. Word treats a
        // hard break inside the body as the start of a new page.
        if pi > 0 {
            out.push_str(r#"<w:p><w:r><w:br w:type="page"/></w:r></w:p>"#);
        }

        // Group spans into lines (by Y position). Each line becomes
        // ONE positioned `<w:framePr>` paragraph carrying all the
        // line's spans as separate runs. Without this grouping, Word
        // emits each span at its absolute frame and the slightly
        // wider fallback font (when the source TeXGyre/NewTXMI face
        // isn't installed) makes adjacent frames overflow into each
        // other — the visible "text on top of text" the user reported.
        let lines = crate::converters::layout_lines::group_spans_into_lines(page.spans.clone());
        for line in &lines {
            // PDF coords: origin bottom-left, y increases upward.
            // DOCX framePr coords: origin top-left, y increases downward.
            let x_pt = line.x_pt.max(0.0).min(page.w_pt);
            let y_pt = (page.h_pt - line.y_pt - line.height_pt)
                .max(0.0)
                .min(page.h_pt);

            let x_twip = (x_pt * TWIPS_PER_PT) as i32;
            let y_twip = (y_pt * TWIPS_PER_PT) as i32;
            // Pad line width 1.5× the source bbox so fallback fonts
            // (Helvetica is wider than TeXGyre) don't clip the
            // trailing run; clamp to remaining page width.
            let line_w_pt = line.width_pt;
            let max_pad_w = (page.w_pt - x_pt).max(line_w_pt + 4.0);
            let w_pt = (line_w_pt * 1.5).max(line_w_pt + 16.0).min(max_pad_w);
            let w_twip = ((w_pt * TWIPS_PER_PT) as i32).max(40);
            let max_font = line
                .spans
                .iter()
                .map(|s| s.font_size)
                .fold(0.0_f32, f32::max)
                .max(line.height_pt);
            let line_h_pt = max_font * 1.4;
            let h_twip = (line_h_pt * TWIPS_PER_PT) as i32;

            // First span's heading_level marks the whole line as a
            // heading (rare for a heading to span multiple distinct
            // <w:p>s on one line; usually the entire line is the
            // heading). Falls back to plain body when None.
            let pstyle_xml = match line.spans.first().and_then(|s| s.heading_level) {
                Some(n) if (1..=6).contains(&n) => {
                    format!("<w:pStyle w:val=\"Heading{}\"/>", n)
                },
                _ => String::new(),
            };

            let mut runs_xml = String::with_capacity(256);
            let mut prev_right_pt: Option<f32> = None;
            for span in &line.spans {
                let text = span.text.trim_matches('\u{0000}');
                if text.is_empty() {
                    continue;
                }

                // Insert a synthetic space when there's a visible
                // gap and neither side already carries whitespace.
                if let Some(prev_right) = prev_right_pt {
                    let gap = span.bbox.x - prev_right;
                    let needs_space = gap > span.font_size * 0.25
                        && !runs_xml.ends_with(' ')
                        && !runs_xml.ends_with("&#160;")
                        && !text.starts_with(|c: char| c.is_whitespace());
                    if needs_space {
                        runs_xml.push_str(r#"<w:r><w:t xml:space="preserve"> </w:t></w:r>"#);
                    }
                }

                let sz_half_pt = (span.font_size * 2.0).round() as i32;
                let color_hex = if span.color.r == 0.0 && span.color.g == 0.0 && span.color.b == 0.0
                {
                    String::new()
                } else {
                    format!(
                        "<w:color w:val=\"{:02X}{:02X}{:02X}\"/>",
                        (span.color.r * 255.0).round().clamp(0.0, 255.0) as u8,
                        (span.color.g * 255.0).round().clamp(0.0, 255.0) as u8,
                        (span.color.b * 255.0).round().clamp(0.0, 255.0) as u8,
                    )
                };
                let bold = if span.font_weight.is_bold() {
                    "<w:b/><w:bCs/>"
                } else {
                    ""
                };
                let italic = if span.is_italic { "<w:i/><w:iCs/>" } else { "" };

                let raw_font = &span.font_name;
                let resolved: String = if let Some(real) = page.font_lookup.get(raw_font) {
                    real.clone()
                } else {
                    let stripped = raw_font
                        .split_once('+')
                        .map(|(_, rest)| rest)
                        .unwrap_or(raw_font);
                    let looks_synthetic = stripped.len() < 12
                        && (stripped.bytes().next() == Some(b'F'))
                        && stripped.bytes().skip(1).all(|b| b.is_ascii_digit());
                    if looks_synthetic {
                        if span.is_monospace {
                            "Courier New"
                        } else {
                            "Times New Roman"
                        }
                        .to_string()
                    } else {
                        stripped.to_string()
                    }
                };
                let font = escape_xml(&resolved);

                runs_xml.push_str(&format!(
                    r#"<w:r><w:rPr><w:rFonts w:ascii="{font}" w:hAnsi="{font}" w:cs="{font}"/>{bold}{italic}<w:sz w:val="{sz}"/><w:szCs w:val="{sz}"/>{color}</w:rPr><w:t xml:space="preserve">{text}</w:t></w:r>"#,
                    font = font,
                    bold = bold, italic = italic,
                    sz = sz_half_pt, color = color_hex,
                    text = escape_xml(text),
                ));
                prev_right_pt = Some(span.bbox.x + span.bbox.width);
            }

            if runs_xml.is_empty() {
                continue;
            }

            out.push_str(&format!(
                r#"<w:p><w:pPr>{pstyle}<w:framePr w:w="{w}" w:h="{h}" w:hRule="exact" w:hSpace="0" w:vSpace="0" w:wrap="none" w:vAnchor="page" w:hAnchor="page" w:x="{x}" w:y="{y}"/><w:spacing w:before="0" w:after="0" w:line="240" w:lineRule="auto"/></w:pPr>{runs}</w:p>"#,
                pstyle = pstyle_xml,
                w = w_twip, h = h_twip, x = x_twip, y = y_twip,
                runs = runs_xml,
            ));
        }

        // Emit each vector shape (table border, divider, frame) as a
        // DrawingML wsp shape anchored to the page. Lines use prstGeom
        // line; rects use prstGeom rect with optional fill/stroke.
        for (sh_idx, shape) in page.shapes.iter().enumerate() {
            let id = pi * 10_000 + 5_000 + sh_idx + 1;
            match shape {
                SimpleShape::Line {
                    x1_pt,
                    y1_pt,
                    x2_pt,
                    y2_pt,
                    stroke_rgb,
                    stroke_w_pt,
                } => {
                    let xa = (*x1_pt as f64).min(*x2_pt as f64);
                    let xb = (*x1_pt as f64).max(*x2_pt as f64);
                    let ya = (*y1_pt as f64).min(*y2_pt as f64);
                    let yb = (*y1_pt as f64).max(*y2_pt as f64);
                    // Y flip
                    let top_y = page.h_pt as f64 - yb;
                    let off_x = (xa * EMU_PER_PT) as i64;
                    let off_y = (top_y * EMU_PER_PT) as i64;
                    let cx = (((xb - xa).max(1.0)) * EMU_PER_PT) as i64;
                    let cy = (((yb - ya).max(1.0)) * EMU_PER_PT) as i64;
                    let stroke_w_emu = ((*stroke_w_pt as f64).max(0.25) * EMU_PER_PT) as i64;
                    let (r, g, b) = stroke_rgb;
                    out.push_str(&format!(
                        r#"<w:p><w:r><w:drawing><wp:anchor xmlns:wp="http://schemas.openxmlformats.org/drawingml/2006/wordprocessingDrawing" xmlns:a="http://schemas.openxmlformats.org/drawingml/2006/main" xmlns:wps="http://schemas.microsoft.com/office/word/2010/wordprocessingShape" behindDoc="0" distT="0" distB="0" distL="0" distR="0" simplePos="0" locked="0" layoutInCell="1" allowOverlap="1" relativeHeight="0"><wp:simplePos x="0" y="0"/><wp:positionH relativeFrom="page"><wp:posOffset>{ox}</wp:posOffset></wp:positionH><wp:positionV relativeFrom="page"><wp:posOffset>{oy}</wp:posOffset></wp:positionV><wp:extent cx="{cx}" cy="{cy}"/><wp:effectExtent l="0" t="0" r="0" b="0"/><wp:wrapNone/><wp:docPr id="{id}" name="Line{id}"/><wp:cNvGraphicFramePr/><a:graphic><a:graphicData uri="http://schemas.microsoft.com/office/word/2010/wordprocessingShape"><wps:wsp><wps:cNvSpPr/><wps:spPr><a:xfrm><a:off x="0" y="0"/><a:ext cx="{cx}" cy="{cy}"/></a:xfrm><a:prstGeom prst="line"><a:avLst/></a:prstGeom><a:noFill/><a:ln w="{sw}"><a:solidFill><a:srgbClr val="{r:02X}{g:02X}{b:02X}"/></a:solidFill></a:ln></wps:spPr><wps:bodyPr/></wps:wsp></a:graphicData></a:graphic></wp:anchor></w:drawing></w:r></w:p>"#,
                        ox = off_x, oy = off_y, cx = cx, cy = cy,
                        id = id, sw = stroke_w_emu, r = r, g = g, b = b,
                    ));
                },
                SimpleShape::Rect {
                    bbox,
                    stroke_rgb,
                    fill_rgb,
                    stroke_w_pt,
                } => {
                    let x_pt = (bbox.x as f64).max(0.0);
                    let y_pt = (page.h_pt as f64 - bbox.y as f64 - bbox.height as f64).max(0.0);
                    let off_x = (x_pt * EMU_PER_PT) as i64;
                    let off_y = (y_pt * EMU_PER_PT) as i64;
                    let cx = ((bbox.width as f64).max(1.0) * EMU_PER_PT) as i64;
                    let cy = ((bbox.height as f64).max(1.0) * EMU_PER_PT) as i64;
                    let fill_xml = match fill_rgb {
                        Some((r, g, b)) => format!(
                            "<a:solidFill><a:srgbClr val=\"{:02X}{:02X}{:02X}\"/></a:solidFill>",
                            r, g, b
                        ),
                        None => "<a:noFill/>".to_string(),
                    };
                    let stroke_xml = match stroke_rgb {
                        Some((r, g, b)) => {
                            let stroke_w_emu =
                                ((*stroke_w_pt as f64).max(0.25) * EMU_PER_PT) as i64;
                            format!(
                                "<a:ln w=\"{}\"><a:solidFill><a:srgbClr val=\"{:02X}{:02X}{:02X}\"/></a:solidFill></a:ln>",
                                stroke_w_emu, r, g, b
                            )
                        },
                        None => "<a:ln><a:noFill/></a:ln>".to_string(),
                    };
                    out.push_str(&format!(
                        r#"<w:p><w:r><w:drawing><wp:anchor xmlns:wp="http://schemas.openxmlformats.org/drawingml/2006/wordprocessingDrawing" xmlns:a="http://schemas.openxmlformats.org/drawingml/2006/main" xmlns:wps="http://schemas.microsoft.com/office/word/2010/wordprocessingShape" behindDoc="1" distT="0" distB="0" distL="0" distR="0" simplePos="0" locked="0" layoutInCell="1" allowOverlap="1" relativeHeight="0"><wp:simplePos x="0" y="0"/><wp:positionH relativeFrom="page"><wp:posOffset>{ox}</wp:posOffset></wp:positionH><wp:positionV relativeFrom="page"><wp:posOffset>{oy}</wp:posOffset></wp:positionV><wp:extent cx="{cx}" cy="{cy}"/><wp:effectExtent l="0" t="0" r="0" b="0"/><wp:wrapNone/><wp:docPr id="{id}" name="Rect{id}"/><wp:cNvGraphicFramePr/><a:graphic><a:graphicData uri="http://schemas.microsoft.com/office/word/2010/wordprocessingShape"><wps:wsp><wps:cNvSpPr/><wps:spPr><a:xfrm><a:off x="0" y="0"/><a:ext cx="{cx}" cy="{cy}"/></a:xfrm><a:prstGeom prst="rect"><a:avLst/></a:prstGeom>{fill}{stroke}</wps:spPr><wps:bodyPr/></wps:wsp></a:graphicData></a:graphic></wp:anchor></w:drawing></w:r></w:p>"#,
                        ox = off_x, oy = off_y, cx = cx, cy = cy,
                        id = id, fill = fill_xml, stroke = stroke_xml,
                    ));
                },
            }
        }

        // Emit each PDF image as a DrawingML floating picture anchored to the
        // page at the source coordinates. Each picture lives inside its own
        // host paragraph (Word requires drawings to sit inside a `<w:p>`).
        for (img_idx, img) in page.images.iter().enumerate() {
            let x_pt = img.bbox.x.max(0.0).min(page.w_pt) as f64;
            let y_pt = (page.h_pt - img.bbox.y - img.bbox.height).max(0.0) as f64;
            let y_pt = y_pt.min(page.h_pt as f64);
            let cx = (img.bbox.width.max(1.0) as f64 * EMU_PER_PT) as i64;
            let cy = (img.bbox.height.max(1.0) as f64 * EMU_PER_PT) as i64;
            let off_x = (x_pt * EMU_PER_PT) as i64;
            let off_y = (y_pt * EMU_PER_PT) as i64;
            let id = pi * 10_000 + img_idx + 1;
            out.push_str(&format!(
                r#"<w:p><w:r><w:drawing><wp:anchor xmlns:wp="http://schemas.openxmlformats.org/drawingml/2006/wordprocessingDrawing" xmlns:a="http://schemas.openxmlformats.org/drawingml/2006/main" xmlns:pic="http://schemas.openxmlformats.org/drawingml/2006/picture" behindDoc="1" distT="0" distB="0" distL="0" distR="0" simplePos="0" locked="0" layoutInCell="1" allowOverlap="1" relativeHeight="0"><wp:simplePos x="0" y="0"/><wp:positionH relativeFrom="page"><wp:posOffset>{ox}</wp:posOffset></wp:positionH><wp:positionV relativeFrom="page"><wp:posOffset>{oy}</wp:posOffset></wp:positionV><wp:extent cx="{cx}" cy="{cy}"/><wp:effectExtent l="0" t="0" r="0" b="0"/><wp:wrapNone/><wp:docPr id="{id}" name="Image{id}"/><wp:cNvGraphicFramePr/><a:graphic><a:graphicData uri="http://schemas.openxmlformats.org/drawingml/2006/picture"><pic:pic><pic:nvPicPr><pic:cNvPr id="{id}" name="Image{id}"/><pic:cNvPicPr/></pic:nvPicPr><pic:blipFill><a:blip xmlns:r="http://schemas.openxmlformats.org/officeDocument/2006/relationships" r:embed="{rid}"/><a:stretch><a:fillRect/></a:stretch></pic:blipFill><pic:spPr><a:xfrm><a:off x="0" y="0"/><a:ext cx="{cx}" cy="{cy}"/></a:xfrm><a:prstGeom prst="rect"><a:avLst/></a:prstGeom></pic:spPr></pic:pic></a:graphicData></a:graphic></wp:anchor></w:drawing></w:r></w:p>"#,
                ox = off_x, oy = off_y, cx = cx, cy = cy, id = id, rid = img.rid,
            ));
        }
    }

    // Final section: page size + zero margins (frames are already absolutely
    // positioned). DOCX requires a sectPr at the end of the body to set page
    // geometry for the whole document.
    let pw_twip = (page_w_pt * TWIPS_PER_PT) as i32;
    let ph_twip = (page_h_pt * TWIPS_PER_PT) as i32;
    out.push_str(&format!(
        r#"<w:sectPr><w:pgSz w:w="{pw}" w:h="{ph}"/><w:pgMar w:top="0" w:right="0" w:bottom="0" w:left="0" w:header="0" w:footer="0" w:gutter="0"/></w:sectPr>"#,
        pw = pw_twip, ph = ph_twip,
    ));

    out
}

struct PageSpans {
    w_pt: f32,
    h_pt: f32,
    spans: Vec<TextSpan>,
    /// PDF font-resource-name → real face name (subset prefix stripped).
    /// e.g. `"F75"` → `"TeXGyreTermesX-Regular"`.
    font_lookup: HashMap<String, String>,
    images: Vec<PageImage>,
    shapes: Vec<SimpleShape>,
}

struct PageImage {
    bbox: crate::geometry::Rect,
    rid: String,
}

/// Simplified vector primitive — just lines and rectangles. Anything more
/// complex (Bezier curves, multi-segment paths) is approximated by its
/// bounding box rectangle. Sufficient for table borders, underlines,
/// frames, and other rule-style decorations.
enum SimpleShape {
    Line {
        x1_pt: f32,
        y1_pt: f32,
        x2_pt: f32,
        y2_pt: f32,
        stroke_rgb: (u8, u8, u8),
        stroke_w_pt: f32,
    },
    Rect {
        bbox: crate::geometry::Rect,
        stroke_rgb: Option<(u8, u8, u8)>,
        fill_rgb: Option<(u8, u8, u8)>,
        stroke_w_pt: f32,
    },
}

fn wrap_document(body_xml: &str) -> String {
    // Note: w:document MUST include the namespace declarations Word checks
    // for. Missing one → "Word found unreadable content" dialog.
    // Image emission also requires `wp` (DrawingML wordprocessing),
    // `a` (DrawingML core), and `pic` (DrawingML picture).
    format!(
        r#"<?xml version="1.0" encoding="UTF-8" standalone="yes"?>
<w:document
    xmlns:w="http://schemas.openxmlformats.org/wordprocessingml/2006/main"
    xmlns:r="http://schemas.openxmlformats.org/officeDocument/2006/relationships"
    xmlns:wp="http://schemas.openxmlformats.org/drawingml/2006/wordprocessingDrawing"
    xmlns:a="http://schemas.openxmlformats.org/drawingml/2006/main"
    xmlns:pic="http://schemas.openxmlformats.org/drawingml/2006/picture">
<w:body>{}</w:body>
</w:document>"#,
        body_xml
    )
}

fn sanitize_font_filename(name: &str) -> String {
    name.chars()
        .map(|c| {
            if c.is_ascii_alphanumeric() || c == '-' || c == '_' {
                c
            } else {
                '_'
            }
        })
        .take(40)
        .collect()
}

/// Stamp `heading_level` on spans that look like a heading by font-size
/// ratio. Matches the thresholds the existing markdown / pdf_to_ir
/// converters use (1.75 / 1.35 / 1.15 × modal body size). Only spans
/// passing simple guards (≥2 alphabetic chars, ≤120 chars, no trailing
/// period for body text) are accepted.
fn annotate_heading_levels(spans: &mut [crate::layout::text_block::TextSpan]) {
    if spans.is_empty() {
        return;
    }
    // Modal body size: most-common font_size, rounded to 0.1 pt buckets.
    let mut buckets: std::collections::HashMap<i32, usize> = Default::default();
    for s in spans.iter() {
        let key = (s.font_size * 10.0).round() as i32;
        *buckets.entry(key).or_default() += 1;
    }
    let body_size_tenth = buckets
        .iter()
        .max_by_key(|(_, c)| *c)
        .map(|(k, _)| *k)
        .unwrap_or(120);
    let body = (body_size_tenth as f32) / 10.0;
    for s in spans.iter_mut() {
        if !looks_like_heading_text(&s.text) {
            continue;
        }
        let ratio = s.font_size / body.max(1.0);
        let level = if ratio >= 1.75 {
            Some(1)
        } else if ratio >= 1.35 {
            Some(2)
        } else if ratio >= 1.15 {
            Some(3)
        } else {
            None
        };
        if level.is_some() {
            s.heading_level = level;
        }
    }
}

/// Reject candidates that obviously aren't headings: too long, no
/// letters, ends in body-paragraph period.
fn looks_like_heading_text(text: &str) -> bool {
    let trimmed = text.trim();
    if trimmed.len() < 2 || trimmed.len() > 120 {
        return false;
    }
    if !trimmed.chars().any(|c| c.is_alphabetic()) {
        return false;
    }
    // A long sentence ending in `.` (and not a numbered heading like
    // "1.2.") is body text, not a heading.
    if trimmed.len() > 60 && trimmed.ends_with('.') {
        return false;
    }
    true
}

/// Merge end-of-line hyphenated spans. PDF text wraps long words across
/// lines as `"agree-"` followed by `"laws"` on the next baseline. The
/// layout writer would otherwise emit them as two adjacent floating
/// frames whose extracted text concatenates to `"agree-laws"` or
/// `"agreelaws"` — a word that doesn't exist. We detect:
///   - prev span text ends in `-`
///   - next span starts with a lowercase letter
///   - next span sits on a different baseline (lower y in PDF coords)
///   - next span begins horizontally near where prev would have
///
/// and merge: drop the trailing `-`, splice the next span's text onto
/// the prev span's text, drop the next span. Keeps the original bbox /
/// font of the prev span so downstream positioning is unchanged.
fn merge_hyphenated_spans(spans: &mut Vec<crate::layout::text_block::TextSpan>) {
    if spans.len() < 2 {
        return;
    }
    let mut i = 0;
    while i + 1 < spans.len() {
        let merge = {
            let cur = &spans[i];
            let next = &spans[i + 1];
            let cur_text = cur.text.trim_end();
            // Hyphen at end of current span, lowercase letter starting the next.
            let ends_hyphen = cur_text.ends_with('-')
                && cur_text.len() >= 2
                && cur_text
                    .chars()
                    .rev()
                    .nth(1)
                    .is_some_and(|c| c.is_alphabetic());
            let starts_lower = next
                .text
                .trim_start()
                .chars()
                .next()
                .is_some_and(|c| c.is_lowercase());
            // Different baseline (PDF y decreases downward in line stacking;
            // a hyphenation continuation lives on a lower line).
            let cur_cy = cur.bbox.y + cur.bbox.height * 0.5;
            let next_cy = next.bbox.y + next.bbox.height * 0.5;
            let line_h = cur.font_size.max(next.font_size).max(1.0);
            let new_line = (cur_cy - next_cy).abs() > line_h * 0.5;
            ends_hyphen && starts_lower && new_line
        };
        if merge {
            // Drop trailing '-' and any whitespace, splice next text.
            let next_text = spans[i + 1].text.trim_start().to_string();
            let mut cur_text = spans[i].text.trim_end().to_string();
            cur_text.pop(); // remove '-'
            cur_text.push_str(&next_text);
            spans[i].text = cur_text;
            // Extend the bbox to cover both fragments so the frame holds
            // the merged word visually. Use union of original boxes.
            let next_bbox = spans[i + 1].bbox;
            let cur_bbox = spans[i].bbox;
            let min_x = cur_bbox.x.min(next_bbox.x);
            let min_y = cur_bbox.y.min(next_bbox.y);
            let max_x = (cur_bbox.x + cur_bbox.width).max(next_bbox.x + next_bbox.width);
            let max_y = (cur_bbox.y + cur_bbox.height).max(next_bbox.y + next_bbox.height);
            spans[i].bbox = crate::geometry::Rect::new(min_x, min_y, max_x - min_x, max_y - min_y);
            spans.remove(i + 1);
            // Don't advance i — the merged span might chain to the next one.
        } else {
            i += 1;
        }
    }
}

/// Reduce a `PathContent` to one or more `SimpleShape`s.
///
/// Recognises:
/// - A single MoveTo + LineTo pair → a single line segment.
/// - A single Rectangle operation → a rect.
/// - Anything else → bounding-box rectangle (a coarse approximation,
///   but better than dropping the path entirely).
fn simplify_path(path: &crate::elements::PathContent) -> Vec<SimpleShape> {
    use crate::elements::PathOperation::*;
    let stroke_rgb = path.stroke_color.as_ref().map(|c| {
        (
            (c.r * 255.0).round().clamp(0.0, 255.0) as u8,
            (c.g * 255.0).round().clamp(0.0, 255.0) as u8,
            (c.b * 255.0).round().clamp(0.0, 255.0) as u8,
        )
    });
    let fill_rgb = path.fill_color.as_ref().map(|c| {
        (
            (c.r * 255.0).round().clamp(0.0, 255.0) as u8,
            (c.g * 255.0).round().clamp(0.0, 255.0) as u8,
            (c.b * 255.0).round().clamp(0.0, 255.0) as u8,
        )
    });

    // Most lines come as `MoveTo(x1,y1), LineTo(x2,y2)`.
    if let [MoveTo(x1, y1), LineTo(x2, y2)] = path.operations.as_slice() {
        if let Some(stroke) = stroke_rgb {
            return vec![SimpleShape::Line {
                x1_pt: *x1,
                y1_pt: *y1,
                x2_pt: *x2,
                y2_pt: *y2,
                stroke_rgb: stroke,
                stroke_w_pt: path.stroke_width.max(0.25),
            }];
        }
    }
    // Rectangle primitive.
    if let [Rectangle(x, y, w, h), ..] = path.operations.as_slice() {
        return vec![SimpleShape::Rect {
            bbox: crate::geometry::Rect::new(*x, *y, *w, *h),
            stroke_rgb,
            fill_rgb,
            stroke_w_pt: path.stroke_width.max(0.25),
        }];
    }
    // Fallback: render the path's bounding box. Loses curvature but
    // preserves position and footprint for things like rounded boxes.
    if path.bbox.width > 0.5 && path.bbox.height > 0.5 {
        vec![SimpleShape::Rect {
            bbox: path.bbox,
            stroke_rgb,
            fill_rgb,
            stroke_w_pt: path.stroke_width.max(0.25),
        }]
    } else {
        vec![]
    }
}

fn escape_xml(s: &str) -> String {
    let mut out = String::with_capacity(s.len() + 8);
    for c in s.chars() {
        // Collapse Mathematical Alphanumeric Symbols (U+1D400-1D7FF) to
        // their plain Latin/Greek base — `𝑥` → `x`, `𝛽` → `β`. Word can't
        // typeset the math-italic codepoints from the standard fonts, and
        // downstream extraction normalises them anyway. Doing it here keeps
        // the layout DOCX clean.
        let c = crate::fonts::encoding::math_alphanumeric_base(c as u32)
            .and_then(char::from_u32)
            .unwrap_or(c);
        match c {
            '<' => out.push_str("&lt;"),
            '>' => out.push_str("&gt;"),
            '&' => out.push_str("&amp;"),
            '"' => out.push_str("&quot;"),
            '\'' => out.push_str("&apos;"),
            // Strip control chars (XML 1.0 forbids most of them).
            c if (c as u32) < 0x20 && c != '\t' && c != '\n' && c != '\r' => {},
            c => out.push(c),
        }
    }
    out
}