docling 1.99.2

DocumentConverter and format backends for docling.rs (a Rust port of docling).
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
//! Shared Office-Open-XML (OOXML) helpers for the DOCX/PPTX/XLSX backends.
//!
//! An OOXML file is a ZIP of XML "parts" plus `.rels` files that wire parts
//! together by relationship id. This module wraps the ZIP, parses relationship
//! files, resolves the (possibly `../`-relative) part paths, and counts embedded
//! pictures in a drawing part.

use std::collections::HashMap;
use std::io::{Cursor, Read};

use docling_core::PictureImage;
use quick_xml::events::Event;
use quick_xml::Reader;
use zip::ZipArchive;

/// Hard cap on a single decompressed OOXML part (512 MiB). Real documents
/// stay far below this; the limit only exists to stop a decompression-bomb
/// part from exhausting memory. Override with `DOCLING_RS_MAX_PART_BYTES`.
pub(crate) fn max_part_bytes() -> u64 {
    docling_core::env::parse("DOCLING_RS_MAX_PART_BYTES").unwrap_or(512 * 1024 * 1024)
}

/// A read-only view over the parts of an OOXML package.
///
/// Cloning is cheap (the file bytes are shared behind an `Arc`, the ZIP
/// central directory is reference-counted by the `zip` crate), which gives
/// each rayon worker its own independent cursor over the same archive.
#[derive(Clone)]
pub struct Package {
    zip: ZipArchive<Cursor<std::sync::Arc<[u8]>>>,
}

impl Package {
    pub fn open(bytes: &[u8]) -> Option<Self> {
        ZipArchive::new(Cursor::new(std::sync::Arc::from(bytes)))
            .ok()
            .map(|zip| Self { zip })
    }

    /// The package's member paths, in central-directory order.
    pub fn names(&self) -> impl Iterator<Item = &str> {
        self.zip.file_names()
    }

    /// Whether any member cannot be read because it is encrypted — docling's
    /// `is_encrypted` for iWork packages. Standard ZIP encryption sets bit 0
    /// of the general-purpose flags; Pages does not use that: it leaves the
    /// flag clear and writes a compression method outside the set ZIP defines
    /// (stored, deflate, bzip2 = 12, lzma = 14), so both signals are needed.
    /// Reads only the central directory — no member is decompressed.
    #[allow(deprecated)] // `CompressionMethod::Unsupported` is how `zip` reports methods it lacks a codec for
    pub fn any_encrypted(&mut self) -> bool {
        use zip::CompressionMethod;
        (0..self.zip.len()).any(|i| match self.zip.by_index_raw(i) {
            Ok(file) => {
                let readable = matches!(
                    file.compression(),
                    CompressionMethod::Stored
                        | CompressionMethod::Deflated
                        | CompressionMethod::Unsupported(12)
                        | CompressionMethod::Unsupported(14)
                );
                file.encrypted() || !readable
            }
            Err(_) => false,
        })
    }

    /// Read a part to a string, or `None` if it is absent, not valid UTF-8, or
    /// nested deeper than the XML guard allows (see [`super::xml_depth`]; the
    /// part is reported and skipped, like an unreadable one).
    pub fn read(&mut self, path: &str) -> Option<String> {
        let bytes = self.read_bytes(path)?;
        let text = String::from_utf8(bytes).ok()?;
        if let Err(e) = super::xml_depth::check(&text, path) {
            eprintln!("docling: {e}; part skipped");
            return None;
        }
        Some(text)
    }

    /// Read a part's raw bytes (e.g. an embedded image), or `None` if absent.
    ///
    /// A single part is never allowed to inflate past [`max_part_bytes`]: an
    /// OOXML file is a ZIP, and a "zip bomb" part (a few KB deflating to many
    /// GB) would otherwise exhaust memory and abort the process. Reads stop at
    /// the cap and the oversized part is rejected (`None`) rather than
    /// truncated, so a partial part never reaches an XML parser.
    pub fn read_bytes(&mut self, path: &str) -> Option<Vec<u8>> {
        self.read_bytes_capped(path, max_part_bytes())
    }

    fn read_bytes_capped(&mut self, path: &str, cap: u64) -> Option<Vec<u8>> {
        let file = self.zip.by_name(path).ok()?;
        // Reject up front when the central directory already advertises an
        // oversized part; still cap the actual read in case the header lies.
        if file.size() > cap {
            return None;
        }
        let mut out = Vec::new();
        // read_to_end on a `.take(cap + 1)` reader: if it returns cap+1 bytes,
        // the part exceeded the cap and is rejected rather than truncated.
        file.take(cap + 1).read_to_end(&mut out).ok()?;
        if out.len() as u64 > cap {
            return None;
        }
        Some(out)
    }

    /// Map each `/image` relationship id of `part` to its extracted
    /// [`PictureImage`] (`base_dir` is the part's directory for resolving
    /// targets, e.g. `word` / `ppt`). Unreadable or undecodable images are skipped.
    pub fn image_rels(&mut self, part: &str, base_dir: &str) -> HashMap<String, PictureImage> {
        let rels = self.rels_for(part);
        let mut out = HashMap::new();
        for r in &rels {
            if !r.rel_type.ends_with("/image") {
                continue;
            }
            let path = resolve(base_dir, &r.target);
            if let Some(bytes) = self.read_bytes(&path) {
                if let Some(img) = picture_image(&path, bytes) {
                    out.insert(r.id.clone(), img);
                }
            }
        }
        out
    }

    /// The `.rels` file governing `part` (e.g. `xl/worksheets/sheet1.xml` →
    /// `xl/worksheets/_rels/sheet1.xml.rels`), parsed into relationships.
    pub fn rels_for(&mut self, part: &str) -> Vec<Relationship> {
        let (dir, file) = split_path(part);
        let rels_path = if dir.is_empty() {
            format!("_rels/{file}.rels")
        } else {
            format!("{dir}/_rels/{file}.rels")
        };
        self.read(&rels_path)
            .map(|x| parse_rels(&x))
            .unwrap_or_default()
    }
}

/// A single `<Relationship Id Type Target>` entry from a `.rels` part.
pub struct Relationship {
    pub id: String,
    pub rel_type: String,
    pub target: String,
}

/// Parse a `.rels` XML document into its relationships.
pub fn parse_rels(xml: &str) -> Vec<Relationship> {
    let mut reader = Reader::from_str(xml);
    let mut buf = Vec::new();
    let mut out = Vec::new();
    loop {
        match reader.read_event_into(&mut buf) {
            Ok(Event::Empty(e)) | Ok(Event::Start(e)) if e.name().as_ref() == b"Relationship" => {
                let (mut id, mut rel_type, mut target) =
                    (String::new(), String::new(), String::new());
                for attr in e.attributes().flatten() {
                    let value = String::from_utf8_lossy(attr.value.as_ref()).into_owned();
                    match attr.key.as_ref() {
                        b"Id" => id = value,
                        b"Type" => rel_type = value,
                        b"Target" => target = value,
                        _ => {}
                    }
                }
                out.push(Relationship {
                    id,
                    rel_type,
                    target,
                });
            }
            Ok(Event::Eof) | Err(_) => break,
            _ => {}
        }
        buf.clear();
    }
    out
}

/// Split a part path into its directory and file name.
fn split_path(path: &str) -> (&str, &str) {
    match path.rfind('/') {
        Some(i) => (&path[..i], &path[i + 1..]),
        None => ("", path),
    }
}

/// Resolve a relationship `target` against the directory of the part that owns
/// the `.rels`, collapsing `.` / `..` segments. A leading `/` is package-absolute.
pub fn resolve(base_dir: &str, target: &str) -> String {
    if let Some(abs) = target.strip_prefix('/') {
        return abs.to_string();
    }
    let mut parts: Vec<&str> = base_dir.split('/').filter(|p| !p.is_empty()).collect();
    for seg in target.split('/') {
        match seg {
            "" | "." => {}
            ".." => {
                parts.pop();
            }
            s => parts.push(s),
        }
    }
    parts.join("/")
}

/// Build a [`PictureImage`] from image bytes, reading the pixel size from the
/// header (decode-free). A Windows metafile (EMF/WMF), which the `image`
/// crate cannot read, is rendered to PNG instead (#536); `None` for anything
/// else it can't read.
pub fn picture_image(path: &str, data: Vec<u8>) -> Option<PictureImage> {
    let dims = image::ImageReader::new(Cursor::new(&data))
        .with_guessed_format()
        .ok()
        .and_then(|r| r.into_dimensions().ok());
    let Some((width, height)) = dims else {
        return super::metafile::render(&data, None);
    };
    Some(PictureImage {
        dpi: PictureImage::DEFAULT_DPI,
        mimetype: mime_for(path).to_string(),
        width,
        height,
        data,
    })
}

/// Whether an embedded picture is a Windows metafile — a placeable WMF (its
/// `9AC6CDD7` key) or an EMF (`" EMF"` at byte 40) — docling's `_is_metafile`.
/// Pillow cannot decode either off Windows, and upstream keeps such a picture
/// (payload-less, unless LibreOffice renders it) where any other undecodable
/// image is dropped.
pub fn is_metafile(data: &[u8]) -> bool {
    data.starts_with(&[0xd7, 0xcd, 0xc6, 0x9a]) || data.get(40..44) == Some(b" EMF")
}

/// PIL's `Image.info["dpi"]` (its horizontal value) for an embedded picture,
/// read the way Pillow reads it: a PNG `pHYs` chunk in pixels per metre, a
/// JPEG's JFIF density (per inch, or per centimetre × 2.54) else — when the
/// JFIF segment carries no unit — its EXIF `XResolution` / `ResolutionUnit`
/// (72 when the EXIF lacks them, as Pillow falls back), a BMP's pixels per
/// metre (÷ 39.3701), a TIFF's resolution tags. `None` when the format
/// records none (GIF, WebP, a PNG without `pHYs`): python-pptx's `Image.dpi`
/// then reports 72, and docling's `ImageRef.dpi` follows it.
pub fn image_dpi(data: &[u8]) -> Option<f64> {
    if data.starts_with(b"\x89PNG\r\n\x1a\n") {
        return png_dpi(data);
    }
    if data.starts_with(&[0xff, 0xd8]) {
        return jpeg_dpi(data);
    }
    if data.starts_with(b"BM") && data.len() >= 42 {
        // BITMAPINFOHEADER (40 bytes) and later carry `biXPelsPerMeter`.
        let header_size = u32::from_le_bytes([data[14], data[15], data[16], data[17]]);
        if header_size >= 40 {
            let ppm = i32::from_le_bytes([data[38], data[39], data[40], data[41]]);
            return Some(f64::from(ppm) / 39.3701);
        }
        return None;
    }
    if data.starts_with(b"II*\0") || data.starts_with(b"MM\0*") {
        // TiffImagePlugin: unit 2 → inches, 3 → centimetres, absent → inches
        // (backward compatibility), 1 → `resolution`, not `dpi`.
        let (xres, unit) = tiff_resolution(data)?;
        return match unit {
            Some(2) | None => Some(xres),
            Some(3) => Some(xres * 2.54),
            _ => None,
        };
    }
    None
}

fn png_dpi(data: &[u8]) -> Option<f64> {
    let mut pos = 8;
    while pos + 8 <= data.len() {
        let len =
            u32::from_be_bytes([data[pos], data[pos + 1], data[pos + 2], data[pos + 3]]) as usize;
        let kind = &data[pos + 4..pos + 8];
        if kind == b"pHYs" && len >= 9 && pos + 8 + len <= data.len() {
            let px =
                u32::from_be_bytes([data[pos + 8], data[pos + 9], data[pos + 10], data[pos + 11]]);
            // Unit 1 = metre; unit 0 records an aspect ratio only.
            return (data[pos + 16] == 1).then(|| f64::from(px) * 0.0254);
        }
        if kind == b"IDAT" || kind == b"IEND" {
            return None;
        }
        pos += 12 + len;
    }
    None
}

fn jpeg_dpi(data: &[u8]) -> Option<f64> {
    let mut dpi: Option<f64> = None;
    let mut pos = 2;
    while pos + 4 <= data.len() && data[pos] == 0xff {
        let marker = data[pos + 1];
        // Stand-alone markers (padding, RSTn, SOI) carry no length.
        if marker == 0xff || marker == 0x01 || (0xd0..=0xd8).contains(&marker) {
            pos += if marker == 0xff { 1 } else { 2 };
            continue;
        }
        // Start of scan / end of image: no more headers.
        if marker == 0xda || marker == 0xd9 {
            break;
        }
        let len = usize::from(u16::from_be_bytes([data[pos + 2], data[pos + 3]]));
        let seg = data.get(pos + 4..(pos + 2 + len).min(data.len()))?;
        if marker == 0xe0 && seg.starts_with(b"JFIF\0") && seg.len() >= 12 {
            let density = f64::from(u16::from_be_bytes([seg[8], seg[9]]));
            match seg[7] {
                1 => dpi = Some(density),
                2 => dpi = Some(density * 2.54),
                _ => {}
            }
        } else if marker == 0xe1 && seg.starts_with(b"Exif\0\0") && dpi.is_none() {
            // Pillow reads `XResolution` / `ResolutionUnit` from the EXIF only
            // when the JFIF gave no dpi, and falls back to 72 when either tag
            // is missing or unreadable.
            dpi = Some(match tiff_resolution(&seg[6..]) {
                Some((xres, Some(3))) => xres * 2.54,
                Some((xres, Some(_))) => xres,
                _ => 72.0,
            });
        }
        pos += 2 + len;
    }
    dpi
}

/// A TIFF header's first IFD: (`XResolution`, `ResolutionUnit`), the former
/// required — `None` when it is absent or malformed.
fn tiff_resolution(t: &[u8]) -> Option<(f64, Option<u16>)> {
    let le = match t.get(0..2)? {
        b"II" => true,
        b"MM" => false,
        _ => return None,
    };
    let u16_at = |o: usize| -> Option<u16> {
        let b = [*t.get(o)?, *t.get(o + 1)?];
        Some(if le {
            u16::from_le_bytes(b)
        } else {
            u16::from_be_bytes(b)
        })
    };
    let u32_at = |o: usize| -> Option<u32> {
        let b = [*t.get(o)?, *t.get(o + 1)?, *t.get(o + 2)?, *t.get(o + 3)?];
        Some(if le {
            u32::from_le_bytes(b)
        } else {
            u32::from_be_bytes(b)
        })
    };
    let ifd = u32_at(4)? as usize;
    let count = usize::from(u16_at(ifd)?);
    let (mut xres, mut unit) = (None, None);
    for i in 0..count {
        let e = ifd + 2 + 12 * i;
        let (tag, typ) = (u16_at(e)?, u16_at(e + 2)?);
        match (tag, typ) {
            // RATIONAL: two u32 at the value offset.
            (0x011a, 5) => {
                let off = u32_at(e + 8)? as usize;
                let (num, den) = (u32_at(off)?, u32_at(off + 4)?);
                if den == 0 {
                    return None;
                }
                xres = Some(f64::from(num) / f64::from(den));
            }
            // SHORT: inline in the value field.
            (0x0128, 3) => unit = Some(u16_at(e + 8)?),
            _ => {}
        }
    }
    Some((xres?, unit))
}

fn mime_for(path: &str) -> &'static str {
    match path
        .rsplit('.')
        .next()
        .unwrap_or("")
        .to_ascii_lowercase()
        .as_str()
    {
        "jpg" | "jpeg" => "image/jpeg",
        "gif" => "image/gif",
        "bmp" => "image/bmp",
        "tif" | "tiff" => "image/tiff",
        "webp" => "image/webp",
        _ => "image/png",
    }
}

/// The content type of a package part, resolved from `[Content_Types].xml`
/// (an exact `<Override PartName>` wins over a `<Default Extension>`).
pub fn content_type(content_types_xml: &str, part: &str) -> Option<String> {
    let mut reader = Reader::from_str(content_types_xml);
    let mut buf = Vec::new();
    let want_part = format!("/{}", part.trim_start_matches('/'));
    let ext = part.rsplit('.').next().unwrap_or("").to_ascii_lowercase();
    let mut default: Option<String> = None;
    loop {
        match reader.read_event_into(&mut buf) {
            Ok(Event::Empty(e)) | Ok(Event::Start(e)) => {
                let tag = e.name();
                let attr = |key: &[u8]| -> Option<String> {
                    e.attributes()
                        .flatten()
                        .find(|a| a.key.as_ref() == key)
                        .map(|a| String::from_utf8_lossy(a.value.as_ref()).into_owned())
                };
                match tag.as_ref() {
                    b"Override" => {
                        if attr(b"PartName").as_deref() == Some(want_part.as_str()) {
                            return attr(b"ContentType");
                        }
                    }
                    b"Default"
                        if attr(b"Extension")
                            .map(|x| x.to_ascii_lowercase())
                            .as_deref()
                            == Some(ext.as_str()) =>
                    {
                        default = attr(b"ContentType");
                    }
                    _ => {}
                }
            }
            Ok(Event::Eof) | Err(_) => break,
            _ => {}
        }
        buf.clear();
    }
    default
}

#[cfg(test)]
mod zip_bomb_tests {
    use super::Package;
    use std::io::Write;

    /// A minimal in-memory OOXML-style zip with one part of `part_len` bytes.
    fn zip_with_part(name: &str, part_len: usize) -> Vec<u8> {
        let mut buf = std::io::Cursor::new(Vec::new());
        {
            let mut zw = zip::ZipWriter::new(&mut buf);
            let opts: zip::write::FileOptions<'_, ()> = zip::write::FileOptions::default()
                .compression_method(zip::CompressionMethod::Deflated);
            zw.start_file(name, opts).unwrap();
            zw.write_all(&vec![b'a'; part_len]).unwrap();
            zw.finish().unwrap();
        }
        buf.into_inner()
    }

    #[test]
    fn oversized_part_is_rejected_not_truncated() {
        // A highly compressible 1 MiB part in a tiny zip (deflates to ~1 KB):
        // the decompression-bomb shape. With the cap below its size, the read
        // must return None rather than a truncated buffer.
        let bytes = zip_with_part("word/document.xml", 1024 * 1024);
        assert!(bytes.len() < 64 * 1024, "part should compress tiny");
        let mut pkg = Package::open(&bytes).unwrap();
        assert!(
            pkg.read_bytes_capped("word/document.xml", 4096).is_none(),
            "a part over the cap must be rejected"
        );
        // Under a generous cap it reads back in full.
        let out = pkg
            .read_bytes_capped("word/document.xml", 8 * 1024 * 1024)
            .expect("part under the cap reads");
        assert_eq!(out.len(), 1024 * 1024);
    }
}

#[cfg(test)]
pub(crate) mod image_dpi_tests {
    use super::{image_dpi, is_metafile};

    /// A PNG of `chunks` (length + type + data + a zero CRC, which the reader
    /// never checks).
    pub(crate) fn png(chunks: &[(&[u8; 4], Vec<u8>)]) -> Vec<u8> {
        let mut out = b"\x89PNG\r\n\x1a\n".to_vec();
        for (kind, data) in chunks {
            out.extend((data.len() as u32).to_be_bytes());
            out.extend(*kind);
            out.extend(data);
            out.extend([0; 4]);
        }
        out
    }

    fn phys(ppu: u32, unit: u8) -> Vec<u8> {
        let mut d = ppu.to_be_bytes().to_vec();
        d.extend(ppu.to_be_bytes());
        d.push(unit);
        d
    }

    /// Pillow: a `pHYs` in metres is `px * 0.0254` dpi; an aspect-only one
    /// (unit 0) or no chunk at all records no dpi.
    #[test]
    fn png_phys_in_metres_is_dpi() {
        let ihdr = (b"IHDR", vec![0; 13]);
        let dpi = image_dpi(&png(&[
            ihdr.clone(),
            (b"pHYs", phys(11811, 1)),
            (b"IEND", vec![]),
        ]));
        assert!((dpi.unwrap() - 299.9994).abs() < 1e-4, "{dpi:?}");
        assert_eq!(
            image_dpi(&png(&[ihdr.clone(), (b"pHYs", phys(1, 0))])),
            None
        );
        assert_eq!(
            image_dpi(&png(&[ihdr, (b"IDAT", vec![0]), (b"IEND", vec![])])),
            None
        );
    }

    fn jpeg(segments: &[(u8, Vec<u8>)]) -> Vec<u8> {
        let mut out = vec![0xff, 0xd8];
        for (marker, payload) in segments {
            out.extend([0xff, *marker]);
            out.extend((payload.len() as u16 + 2).to_be_bytes());
            out.extend(payload);
        }
        out.extend([0xff, 0xda, 0, 2]);
        out
    }

    fn jfif(unit: u8, density: u16) -> Vec<u8> {
        let mut d = b"JFIF\0\x01\x01".to_vec();
        d.push(unit);
        d.extend(density.to_be_bytes());
        d.extend(density.to_be_bytes());
        d.extend([0, 0]);
        d
    }

    /// Little-endian TIFF with `XResolution = num/den` and, optionally,
    /// `ResolutionUnit`.
    fn tiff(num: u32, den: u32, unit: Option<u16>) -> Vec<u8> {
        let mut t = b"II*\0".to_vec();
        t.extend(8u32.to_le_bytes());
        let n: u16 = if unit.is_some() { 2 } else { 1 };
        t.extend(n.to_le_bytes());
        let rational_at = 8 + 2 + 12 * u32::from(n) + 4;
        t.extend(0x011au16.to_le_bytes());
        t.extend(5u16.to_le_bytes());
        t.extend(1u32.to_le_bytes());
        t.extend(rational_at.to_le_bytes());
        if let Some(u) = unit {
            t.extend(0x0128u16.to_le_bytes());
            t.extend(3u16.to_le_bytes());
            t.extend(1u32.to_le_bytes());
            t.extend(u.to_le_bytes());
            t.extend([0, 0]);
        }
        t.extend(0u32.to_le_bytes()); // next IFD
        t.extend(num.to_le_bytes());
        t.extend(den.to_le_bytes());
        t
    }

    /// Pillow: the JFIF density in inches or centimetres; with no JFIF unit,
    /// the EXIF `XResolution` (centimetres × 2.54), 72 when the EXIF lacks
    /// the tags; nothing at all when neither segment says.
    #[test]
    fn jpeg_density_from_jfif_then_exif() {
        assert_eq!(image_dpi(&jpeg(&[(0xe0, jfif(1, 96))])), Some(96.0));
        let dpcm = image_dpi(&jpeg(&[(0xe0, jfif(2, 118))])).unwrap();
        assert!((dpcm - 299.72).abs() < 1e-9);
        let mut exif = b"Exif\0\0".to_vec();
        exif.extend(tiff(300, 1, Some(2)));
        assert_eq!(
            image_dpi(&jpeg(&[(0xe0, jfif(0, 1)), (0xe1, exif)])),
            Some(300.0)
        );
        let mut exif_cm = b"Exif\0\0".to_vec();
        exif_cm.extend(tiff(100, 1, Some(3)));
        assert_eq!(image_dpi(&jpeg(&[(0xe1, exif_cm)])), Some(254.0));
        let mut exif_bare = b"Exif\0\0".to_vec();
        exif_bare.extend(tiff(300, 1, None));
        assert_eq!(
            image_dpi(&jpeg(&[(0xe1, exif_bare)])),
            Some(72.0),
            "no unit tag → Pillow's 72"
        );
        assert_eq!(image_dpi(&jpeg(&[(0xe0, jfif(0, 1))])), None);
        // A JFIF density beats a later EXIF, an earlier EXIF yields to it.
        let mut exif = b"Exif\0\0".to_vec();
        exif.extend(tiff(300, 1, Some(2)));
        assert_eq!(
            image_dpi(&jpeg(&[(0xe1, exif.clone()), (0xe0, jfif(1, 96))])),
            Some(96.0)
        );
    }

    /// A BMP's pixels per metre ÷ 39.3701; a TIFF's resolution in inches (unit
    /// 2 or none), centimetres (3), or none at all (1).
    #[test]
    fn bmp_and_tiff_resolutions() {
        let mut bmp = vec![0u8; 54];
        bmp[..2].copy_from_slice(b"BM");
        bmp[14..18].copy_from_slice(&40u32.to_le_bytes());
        bmp[38..42].copy_from_slice(&11811i32.to_le_bytes());
        assert!((image_dpi(&bmp).unwrap() - 299.9992).abs() < 1e-3);
        assert_eq!(image_dpi(&tiff(200, 1, Some(2))), Some(200.0));
        assert_eq!(image_dpi(&tiff(200, 1, None)), Some(200.0));
        assert_eq!(image_dpi(&tiff(100, 1, Some(3))), Some(254.0));
        assert_eq!(image_dpi(&tiff(100, 1, Some(1))), None);
        assert_eq!(image_dpi(b"GIF89a"), None);
    }

    /// docling's `_is_metafile`: a placeable WMF key or an EMF signature at 40.
    #[test]
    fn metafiles_by_signature() {
        assert!(is_metafile(&[0xd7, 0xcd, 0xc6, 0x9a, 0, 0]));
        let mut emf = vec![1u8; 44];
        emf[40..44].copy_from_slice(b" EMF");
        assert!(is_metafile(&emf));
        assert!(!is_metafile(b"\x89PNG\r\n\x1a\n"));
        assert!(!is_metafile(&[]));
    }
}