From e04e559da359852a17c8fa55d42764d985262b98 Mon Sep 17 00:00:00 2001
From: vangel76 <vangel1976@googlemail.com>
Date: Mon, 12 Oct 2026 22:25:50 +0200
Subject: [PATCH 1/3] PDF: writer cleanup, q nesting limit, image dedupe, style
 only on change (#816)

Direct object writing, image dedupe, per-page style tracking, 28-level q limit,
explicit joins, /ProcSet dropped. See the PR description.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
---
 ROADMAP.md                        |   4 +-
 crates/pdf/src/lib.rs             | 404 +++++++++++++++++------
 crates/pdf/src/tests.rs           | 530 ++++++++++++++++++++++++++++++
 crates/pdf/src/tests/reference.rs | 349 ++++++++++++++++++++
 4 files changed, 1184 insertions(+), 103 deletions(-)
 create mode 100644 crates/pdf/src/tests/reference.rs

diff --git a/ROADMAP.md b/ROADMAP.md
index e8e1c06b3..74755ce60 100644
--- a/ROADMAP.md
+++ b/ROADMAP.md
@@ -2,7 +2,7 @@
 
 **Stage: alpha** · next: beta, ~25 points of "ready for real work" and ~160–240 h away
 
-> **Last reviewed:** 2026-10-11 · **Last updated:** 2026-10-12 · **Change:** minor (direct desktop printing through craft-print) · **Target:** Microsoft Excel (Microsoft 365)
+> **Last reviewed:** 2026-10-11 · **Last updated:** 2026-10-12 · **Change:** minor (PDF writer cleanup, #816) · **Target:** Microsoft Excel (Microsoft 365)
 
 GridCraft aims at full Microsoft Excel parity, and at being better in places: faster, open (XLSX
 is the native format), fully scriptable by agents (MCP, control channel, CLI), and on the web.
@@ -132,6 +132,7 @@ Ranked; estimates in Opus 5.5 agent-hours. Detail and the full beta-gate list in
 
 | Date | What landed |
 |---|---|
+| 2026-10-12 | PDF writer: objects are written straight to the output, identical pictures are stored once, style operators are emitted only when they change, `q` nesting is limited to 28 and polyline joins no longer leak into later strokes (#816) |
 | 2026-10-12 | File › Print selects a system printer, copies, duplex and grayscale, submits PDF jobs through craft-print on a worker and shows spooler errors; browser printing and PDF export remain available |
 | 2026-10-11 | Command-sequence fuzzing: random command sequences checked against undo/redo, XLSX round-trip and full recalculation; fixes for the table, legacy-array, spill, AutoSum, Watch Window and sheet-copy bugs it found |
 | 2026-10-11 | Password-protected workbooks open: GridCraft asks for the password and decrypts Agile and Standard encryption (#11) |
@@ -178,6 +179,7 @@ Ranked; estimates in Opus 5.5 agent-hours. Detail and the full beta-gate list in
 
 | Date | Change | Summary |
 |---|---|---|
+| 2026-10-12 | minor | Progress log: PDF writer cleanup (#816) |
 | 2026-10-12 | minor | Desktop File > Print now discovers printers and submits PDF jobs through pinned craft-print with printer choice, copies, duplex and grayscale; browser printing and PDF export remain available. Remaining print parity includes page ranges and driver-specific tray/options UI; physical printer fidelity needs hardware validation. |
 | 2026-10-11 | minor | Password-encrypted workbooks open (#11); progress log row |
 | 2026-10-11 | minor | Forecast Sheet (#322); progress log row |
diff --git a/crates/pdf/src/lib.rs b/crates/pdf/src/lib.rs
index 51c2df115..4af8810d4 100644
--- a/crates/pdf/src/lib.rs
+++ b/crates/pdf/src/lib.rs
@@ -13,7 +13,10 @@
 
 mod metrics;
 
+use std::collections::HashMap;
+use std::collections::hash_map::DefaultHasher;
 use std::fmt::Write as _;
+use std::hash::{Hash, Hasher};
 
 /// RGB colour, 0–255 per channel.
 pub type Rgb = [u8; 3];
@@ -221,16 +224,75 @@ fn n(v: f32) -> String {
     }
 }
 
-fn rgb_op(c: Rgb, stroke: bool) -> String {
-    let f = |v: u8| n(v as f32 / 255.0);
-    format!("{} {} {} {}", f(c[0]), f(c[1]), f(c[2]), if stroke { "RG" } else { "rg" })
+/// Appends `v / 255` with at most three decimals, exactly what `n(v as f32 / 255.0)` gives, but
+/// without a float round trip or an allocation per channel.
+fn push_unit(s: &mut String, v: u8) {
+    let m = (u32::from(v) * 1000 + 127) / 255;
+    match m {
+        0 => s.push('0'),
+        1000 => s.push('1'),
+        _ => {
+            let _ = write!(s, "0.{m:03}");
+            while s.ends_with('0') {
+                s.pop();
+            }
+        }
+    }
+}
+
+/// Appends a colour operator line without the trailing newline: `r g b rg` (fill) or `r g b RG`.
+fn push_rgb(s: &mut String, c: Rgb, stroke: bool) {
+    for v in c {
+        push_unit(s, v);
+        s.push(' ');
+    }
+    s.push_str(if stroke { "RG" } else { "rg" });
+}
+
+/// The dash array contents (`3 2`) of a dash pattern; empty for solid.
+fn dash_string(dash: &[f32]) -> String {
+    let d: Vec<String> = dash.iter().take(8).map(|v| n(v.max(0.0))).collect();
+    d.join(" ")
 }
 
 struct ImageData {
     width: u32,
     height: u32,
     rgb: Vec<u8>,
+    /// `None` when every pixel is fully opaque (no soft mask is written).
     alpha: Option<Vec<u8>>,
+    /// Hash of all of the above, to find duplicates quickly; equality still compares the bytes.
+    hash: u64,
+}
+
+impl ImageData {
+    fn new(width: u32, height: u32, rgb: Vec<u8>, alpha: Option<Vec<u8>>) -> ImageData {
+        let mut h = DefaultHasher::new();
+        width.hash(&mut h);
+        height.hash(&mut h);
+        rgb.hash(&mut h);
+        alpha.hash(&mut h);
+        ImageData { width, height, rgb, alpha, hash: h.finish() }
+    }
+
+    /// Same pixels (a hash match alone is never trusted).
+    fn same(&self, o: &ImageData) -> bool {
+        self.hash == o.hash && self.width == o.width && self.height == o.height && self.rgb == o.rgb && self.alpha == o.alpha
+    }
+}
+
+/// The part of the graphics state the writer tracks so that a style operator is only emitted
+/// when it changes. `None` = not set yet on this page (or since the last restore that undid it).
+#[derive(Clone, Debug, Default, PartialEq)]
+struct Style {
+    fill: Option<Rgb>,
+    stroke: Option<Rgb>,
+    /// Formatted line width.
+    width: Option<String>,
+    /// Formatted dash array contents.
+    dash: Option<String>,
+    /// Line join; PDF's initial value is 0 (miter).
+    join: u8,
 }
 
 /// A page under construction. Obtain one with [`PdfDoc::add_page`].
@@ -240,7 +302,12 @@ pub struct Page {
     content: String,
     fonts: [bool; 12],
     images: Vec<ImageData>,
-    depth: u32,
+    /// Style operators already in effect at the end of `content`.
+    style: Style,
+    /// One entry per accepted [`Page::save`], restored by [`Page::restore`].
+    saved: Vec<Style>,
+    /// `save` calls refused at the nesting limit; the same number of `restore` calls are ignored.
+    refused: u32,
 }
 
 /// Most bytes of content per page before further drawing is dropped (keeps hostile inputs from
@@ -248,6 +315,11 @@ pub struct Page {
 const MAX_PAGE_CONTENT: usize = 64 * 1024 * 1024;
 /// Largest image (pixels) embedded.
 const MAX_IMAGE_PIXELS: u64 = 40_000_000;
+/// Deepest `q` nesting a PDF consumer must support (the PDF specification's implementation limits).
+const MAX_SAVE_DEPTH: usize = 28;
+/// Line join styles, as the `j` operator takes them.
+const JOIN_MITER: u8 = 0;
+const JOIN_ROUND: u8 = 1;
 
 impl Page {
     pub fn width(&self) -> f32 {
@@ -269,15 +341,55 @@ impl Page {
         }
     }
 
-    /// Saves the graphics state (pair with [`Page::restore`]).
+    /// Appends a fill colour operator to `s` unless that colour is already in effect.
+    fn set_fill(&mut self, s: &mut String, c: Rgb) {
+        if self.style.fill != Some(c) {
+            push_rgb(s, c, false);
+            s.push('\n');
+            self.style.fill = Some(c);
+        }
+    }
+
+    /// Appends the stroke colour, width, dash and join operators to `s` that differ from the
+    /// style already in effect.
+    fn set_stroke(&mut self, s: &mut String, c: Rgb, width: f32, dash: &str, join: u8) {
+        if self.style.stroke != Some(c) {
+            push_rgb(s, c, true);
+            s.push('\n');
+            self.style.stroke = Some(c);
+        }
+        let w = n(width);
+        if self.style.width.as_deref() != Some(w.as_str()) {
+            let _ = writeln!(s, "{w} w");
+            self.style.width = Some(w);
+        }
+        if self.style.dash.as_deref() != Some(dash) {
+            let _ = writeln!(s, "[{dash}] 0 d");
+            self.style.dash = Some(dash.to_string());
+        }
+        if self.style.join != join {
+            let _ = writeln!(s, "{join} j");
+            self.style.join = join;
+        }
+    }
+
+    /// Saves the graphics state (pair with [`Page::restore`]). Nesting is limited to 28 levels, the
+    /// depth every PDF consumer must support; a `save` beyond that is ignored, and so is the
+    /// `restore` that would have matched it.
     pub fn save(&mut self) {
-        self.depth = self.depth.saturating_add(1);
+        if self.saved.len() >= MAX_SAVE_DEPTH {
+            self.refused = self.refused.saturating_add(1);
+            return;
+        }
+        self.saved.push(self.style.clone());
         self.content.push_str("q\n");
     }
     /// Restores the graphics state saved by [`Page::save`]; extra calls are ignored.
     pub fn restore(&mut self) {
-        if self.depth > 0 {
-            self.depth -= 1;
+        if self.refused > 0 {
+            self.refused -= 1;
+        } else if let Some(style) = self.saved.pop() {
+            self.style = style;
             self.content.push_str("Q\n");
         }
     }
@@ -288,15 +400,22 @@ impl Page {
     }
 
     pub fn fill_rect(&mut self, x: f32, y: f32, w: f32, h: f32, color: Rgb) {
-        if w <= 0.0 || h <= 0.0 {
+        if w <= 0.0 || h <= 0.0 || self.full() {
             return;
         }
-        let s = format!("{}\n{} {} {} {} re f", rgb_op(color, false), n(x), n(self.y(y + h)), n(w), n(h));
+        let mut s = String::new();
+        self.set_fill(&mut s, color);
+        let _ = write!(s, "{} {} {} {} re f", n(x), n(self.y(y + h)), n(w), n(h));
         self.op(&s);
     }
 
     pub fn rect_stroke(&mut self, x: f32, y: f32, w: f32, h: f32, width: f32, color: Rgb) {
-        let s = format!("{}\n{} w [] 0 d\n{} {} {} {} re S", rgb_op(color, true), n(width), n(x), n(self.y(y + h)), n(w), n(h));
+        if self.full() {
+            return;
+        }
+        let mut s = String::new();
+        self.set_stroke(&mut s, color, width, "", JOIN_MITER);
+        let _ = write!(s, "{} {} {} {} re S", n(x), n(self.y(y + h)), n(w), n(h));
         self.op(&s);
     }
 
@@ -306,19 +425,22 @@ impl Page {
 
     /// A line with a dash pattern (`[on, off, …]` in points; empty = solid).
     pub fn stroke_line_dashed(&mut self, x0: f32, y0: f32, x1: f32, y1: f32, width: f32, color: Rgb, dash: &[f32]) {
-        let d: Vec<String> = dash.iter().take(8).map(|v| n(v.max(0.0))).collect();
-        let s =
-            format!("{}\n{} w [{}] 0 d\n{} {} m {} {} l S", rgb_op(color, true), n(width), d.join(" "), n(x0), n(self.y(y0)), n(x1), n(self.y(y1)));
+        if self.full() {
+            return;
+        }
+        let mut s = String::new();
+        self.set_stroke(&mut s, color, width, &dash_string(dash), JOIN_MITER);
+        let _ = write!(s, "{} {} m {} {} l S", n(x0), n(self.y(y0)), n(x1), n(self.y(y1)));
         self.op(&s);
     }
 
-    /// An open polyline.
+    /// An open polyline (round joins).
     pub fn polyline(&mut self, pts: &[[f32; 2]], width: f32, color: Rgb, dash: &[f32]) {
-        if pts.len() < 2 {
+        if pts.len() < 2 || self.full() {
             return;
         }
-        let d: Vec<String> = dash.iter().take(8).map(|v| n(v.max(0.0))).collect();
-        let mut s = format!("{}\n{} w [{}] 0 d 1 j\n", rgb_op(color, true), n(width), d.join(" "));
+        let mut s = String::new();
+        self.set_stroke(&mut s, color, width, &dash_string(dash), JOIN_ROUND);
         self.path(&mut s, pts, false);
         s.push_str(" S");
         self.op(&s);
@@ -333,18 +455,17 @@ impl Page {
         }
     }
 
-    /// A closed polygon, filled and/or stroked.
+    /// A closed polygon, filled and/or stroked (round joins).
     pub fn polygon(&mut self, pts: &[[f32; 2]], fill: Option<Rgb>, stroke: Option<(Rgb, f32)>) {
-        if pts.len() < 2 || (fill.is_none() && stroke.is_none()) {
+        if pts.len() < 2 || (fill.is_none() && stroke.is_none()) || self.full() {
             return;
         }
         let mut s = String::new();
         if let Some(f) = fill {
-            s.push_str(&rgb_op(f, false));
-            s.push('\n');
+            self.set_fill(&mut s, f);
         }
         if let Some((c, w)) = stroke {
-            let _ = writeln!(s, "{}\n{} w [] 0 d 1 j", rgb_op(c, true), n(w));
+            self.set_stroke(&mut s, c, w, "", JOIN_ROUND);
         }
         self.path(&mut s, pts, true);
         s.push_str(match (fill.is_some(), stroke.is_some()) {
@@ -357,7 +478,7 @@ impl Page {
 
     /// A circle approximated by four cubic Béziers.
     pub fn circle(&mut self, cx: f32, cy: f32, r: f32, fill: Option<Rgb>, stroke: Option<(Rgb, f32)>) {
-        if r <= 0.0 || (fill.is_none() && stroke.is_none()) {
+        if r <= 0.0 || (fill.is_none() && stroke.is_none()) || self.full() {
             return;
         }
         const K: f32 = 0.552_284_8;
@@ -365,11 +486,10 @@ impl Page {
         let y = self.y(cy);
         let mut s = String::new();
         if let Some(f) = fill {
-            s.push_str(&rgb_op(f, false));
-            s.push('\n');
+            self.set_fill(&mut s, f);
         }
         if let Some((c, w)) = stroke {
-            let _ = writeln!(s, "{}\n{} w [] 0 d", rgb_op(c, true), n(w));
+            self.set_stroke(&mut s, c, w, "", JOIN_MITER);
         }
         let _ = write!(
             s,
@@ -424,11 +544,16 @@ impl Page {
         if let Some(f) = self.fonts.get_mut(fi) {
             *f = true;
         }
+        if self.full() {
+            return;
+        }
         let (sn, cs) = if angle.is_finite() { angle.sin_cos() } else { (0.0, 1.0) };
+        // The colour goes in front of the text object, so it is tracked like any other style.
+        let mut colour = String::new();
+        self.set_fill(&mut colour, color);
         // Clockwise in y-down = counter-clockwise negated in PDF's y-up space.
         let s = format!(
-            "BT\n{}\n/F{} {} Tf\n{} {} {} {} {} {} Tm\n{} Tj\nET",
-            rgb_op(color, false),
+            "{colour}BT\n/F{} {} Tf\n{} {} {} {} {} {} Tm\n{} Tj\nET",
             fi + 1,
             n(size.min(10_000.0)),
             n(cs),
@@ -449,16 +574,36 @@ impl Page {
         if pw == 0 || ph == 0 || px > MAX_IMAGE_PIXELS || (rgba.len() as u64) < px * 4 || w <= 0.0 || h <= 0.0 {
             return;
         }
+        // The image is drawn inside its own q/Q, which counts towards the nesting limit.
+        if self.saved.len() >= MAX_SAVE_DEPTH {
+            return;
+        }
         let px = px as usize;
         let mut rgb = Vec::with_capacity(px * 3);
-        let mut alpha = Vec::with_capacity(px);
-        for p in rgba.as_chunks::<4>().0.iter().take(px) {
+        // The alpha plane is only allocated once a non-opaque pixel shows up.
+        let mut alpha: Option<Vec<u8>> = None;
+        for (i, p) in rgba.as_chunks::<4>().0.iter().take(px).enumerate() {
             rgb.extend_from_slice(&[p[0], p[1], p[2]]);
-            alpha.push(p[3]);
+            match &mut alpha {
+                Some(a) => a.push(p[3]),
+                None if p[3] != 255 => {
+                    let mut a = Vec::with_capacity(px);
+                    a.resize(i, 255);
+                    a.push(p[3]);
+                    alpha = Some(a);
+                }
+                None => {}
+            }
         }
-        let alpha = if alpha.iter().all(|a| *a == 255) { None } else { Some(alpha) };
-        let idx = self.images.len();
-        self.images.push(ImageData { width: pw, height: ph, rgb, alpha });
+        let img = ImageData::new(pw, ph, rgb, alpha);
+        // The same picture drawn again on this page shares one XObject.
+        let idx = match self.images.iter().position(|i| i.same(&img)) {
+            Some(i) => i,
+            None => {
+                self.images.push(img);
+                self.images.len() - 1
+            }
+        };
         let s = format!("q {} 0 0 {} {} {} cm /Im{} Do Q", n(w), n(h), n(x), n(self.y(y + h)), idx + 1);
         self.op(&s);
     }
@@ -520,6 +665,49 @@ fn pdf_text_string(s: &str) -> String {
     pdf_string(s)
 }
 
+/// Starts object `id` at the current end of `out` and records its offset for the xref table.
+fn begin_obj(out: &mut Vec<u8>, offsets: &mut Vec<usize>, id: u32) {
+    let i = id as usize;
+    if offsets.len() <= i {
+        offsets.resize(i + 1, 0);
+    }
+    if let Some(slot) = offsets.get_mut(i) {
+        *slot = out.len();
+    }
+    let _ = writeln!(Utf8(out), "{id} 0 obj");
+}
+
+/// Writes `id 0 obj <body> endobj`.
+fn dict_obj(out: &mut Vec<u8>, offsets: &mut Vec<usize>, id: u32, body: &str) {
+    begin_obj(out, offsets, id);
+    out.extend_from_slice(body.as_bytes());
+    out.extend_from_slice(b"\nendobj\n");
+}
+
+/// Writes a stream object: `dict` (the dictionary entries without `<<`, `>>` and `/Length`, may be
+/// empty) and `data` as the stream body.
+fn stream_obj(out: &mut Vec<u8>, offsets: &mut Vec<usize>, id: u32, dict: &str, data: &[u8]) {
+    begin_obj(out, offsets, id);
+    out.extend_from_slice(b"<<");
+    if !dict.is_empty() {
+        out.push(b' ');
+        out.extend_from_slice(dict.as_bytes());
+    }
+    let _ = write!(Utf8(out), " /Length {} >>\nstream\n", data.len());
+    out.extend_from_slice(data);
+    out.extend_from_slice(b"\nendstream\nendobj\n");
+}
+
+/// Lets `write!` append text to a byte buffer.
+struct Utf8<'a>(&'a mut Vec<u8>);
+
+impl std::fmt::Write for Utf8<'_> {
+    fn write_str(&mut self, s: &str) -> std::fmt::Result {
+        self.0.extend_from_slice(s.as_bytes());
+        Ok(())
+    }
+}
+
 impl PdfDoc {
     pub fn new() -> PdfDoc {
         PdfDoc::default()
@@ -529,7 +717,16 @@ impl PdfDoc {
     pub fn add_page(&mut self, width_pt: f32, height_pt: f32) -> &mut Page {
         let width = if width_pt.is_finite() { width_pt.clamp(3.0, 14_400.0) } else { 612.0 };
         let height = if height_pt.is_finite() { height_pt.clamp(3.0, 14_400.0) } else { 792.0 };
-        self.pages.push(Page { width, height, content: String::new(), fonts: [false; 12], images: Vec::new(), depth: 0 });
+        self.pages.push(Page {
+            width,
+            height,
+            content: String::new(),
+            fonts: [false; 12],
+            images: Vec::new(),
+            style: Style::default(),
+            saved: Vec::new(),
+            refused: 0,
+        });
         let i = self.pages.len() - 1;
         &mut self.pages[i]
     }
@@ -544,10 +741,22 @@ impl PdfDoc {
     }
 
     /// Serializes the document. A document without pages gets one blank Letter page.
+    ///
+    /// Every object is written straight into the output buffer (nothing is staged in an
+    /// intermediate list of bodies); only the xref table needs the offset of each object id, and
+    /// the ids are contiguous from 1. Objects are therefore in the file in the order they are
+    /// numbered during writing (fonts, then per page its new images, content stream and page
+    /// dictionary), with the catalog, page tree and info dictionary last.
     pub fn finish(mut self) -> Vec<u8> {
         if self.pages.is_empty() {
             self.add_page(612.0, 792.0);
         }
+        for page in &mut self.pages {
+            page.refused = 0;
+            while !page.saved.is_empty() {
+                page.restore();
+            }
+        }
         // Object numbering: 1 catalog, 2 page tree, 3 info, then fonts, then per page.
         let mut used = [false; 12];
         for p in &self.pages {
@@ -563,7 +772,12 @@ impl PdfDoc {
                 next += 1;
             }
         }
-        let mut objects: Vec<(u32, Vec<u8>)> = Vec::new();
+        // Page contents are moved out so each one is freed as soon as it is in the output.
+        let contents: Vec<String> = self.pages.iter_mut().map(|p| std::mem::take(&mut p.content)).collect();
+        let reserve = contents.iter().fold(4096usize, |acc, c| acc.saturating_add(c.len()).saturating_add(1024));
+        let mut out: Vec<u8> = Vec::with_capacity(reserve);
+        out.extend_from_slice(b"%PDF-1.7\n%\xE2\xE3\xCF\xD3\n");
+        let mut offsets: Vec<usize> = vec![0; 4];
         let mut kids = Vec::new();
         let mut font_dict = String::new();
         for (i, id) in font_ids.iter().enumerate() {
@@ -571,76 +785,75 @@ impl PdfDoc {
                 let _ = write!(font_dict, "/F{} {} 0 R ", i + 1, id);
                 let name = Font::ALL.get(i).copied().unwrap_or_default().base_name();
                 let enc = if name.starts_with("Symbol") { "" } else { " /Encoding /WinAnsiEncoding" };
-                objects.push((*id, format!("<< /Type /Font /Subtype /Type1 /BaseFont /{name}{enc} >>").into_bytes()));
+                dict_obj(&mut out, &mut offsets, *id, &format!("<< /Type /Font /Subtype /Type1 /BaseFont /{name}{enc} >>"));
             }
         }
-        for page in &mut self.pages {
-            while page.depth > 0 {
-                page.restore();
-            }
+        // Images already written, by content hash: (page, image, object id). A hash match is
+        // confirmed by comparing the pixels before an object is shared.
+        let mut written: HashMap<u64, Vec<(usize, usize, u32)>> = HashMap::new();
+        for (pi, (page, content)) in self.pages.iter().zip(contents).enumerate() {
             let mut xobjects = String::new();
             for (k, img) in page.images.iter().enumerate() {
-                let img_id = next;
-                next += 1;
-                let smask = match &img.alpha {
-                    Some(a) => {
-                        let id = next;
+                let known = written.get(&img.hash).and_then(|c| {
+                    c.iter().find_map(|(p, i, id)| {
+                        let other = self.pages.get(*p)?.images.get(*i)?;
+                        img.same(other).then_some(*id)
+                    })
+                });
+                let img_id = match known {
+                    Some(id) => id,
+                    None => {
+                        let img_id = next;
                         next += 1;
-                        let data = run_length(a);
-                        let mut o = format!(
-                            "<< /Type /XObject /Subtype /Image /Width {} /Height {} /ColorSpace /DeviceGray /BitsPerComponent 8 /Filter /RunLengthDecode /Length {} >>\nstream\n",
-                            img.width,
-                            img.height,
-                            data.len()
-                        )
-                        .into_bytes();
-                        o.extend_from_slice(&data);
-                        o.extend_from_slice(b"\nendstream");
-                        objects.push((id, o));
-                        format!(" /SMask {id} 0 R")
+                        let mut dict = format!("/Type /XObject /Subtype /Image /Width {} /Height {} /BitsPerComponent 8", img.width, img.height);
+                        let smask_id = img.alpha.as_ref().map(|_| {
+                            let id = next;
+                            next += 1;
+                            id
+                        });
+                        dict.push_str(" /ColorSpace /DeviceRGB");
+                        if let Some(id) = smask_id {
+                            let _ = write!(dict, " /SMask {id} 0 R");
+                        }
+                        dict.push_str(" /Filter /RunLengthDecode");
+                        stream_obj(&mut out, &mut offsets, img_id, &dict, &run_length(&img.rgb));
+                        if let (Some(id), Some(a)) = (smask_id, &img.alpha) {
+                            let dict = format!(
+                                "/Type /XObject /Subtype /Image /Width {} /Height {} /ColorSpace /DeviceGray /BitsPerComponent 8 /Filter /RunLengthDecode",
+                                img.width, img.height
+                            );
+                            stream_obj(&mut out, &mut offsets, id, &dict, &run_length(a));
+                        }
+                        written.entry(img.hash).or_default().push((pi, k, img_id));
+                        img_id
                     }
-                    None => String::new(),
                 };
-                let data = run_length(&img.rgb);
-                let mut o = format!(
-                    "<< /Type /XObject /Subtype /Image /Width {} /Height {} /ColorSpace /DeviceRGB /BitsPerComponent 8{} /Filter /RunLengthDecode /Length {} >>\nstream\n",
-                    img.width,
-                    img.height,
-                    smask,
-                    data.len()
-                )
-                .into_bytes();
-                o.extend_from_slice(&data);
-                o.extend_from_slice(b"\nendstream");
-                objects.push((img_id, o));
                 let _ = write!(xobjects, "/Im{} {} 0 R ", k + 1, img_id);
             }
             let content_id = next;
             let page_id = next + 1;
             next += 2;
-            let body = std::mem::take(&mut page.content);
-            let mut o = format!("<< /Length {} >>\nstream\n", body.len()).into_bytes();
-            o.extend_from_slice(body.as_bytes());
-            o.extend_from_slice(b"endstream");
-            objects.push((content_id, o));
+            stream_obj(&mut out, &mut offsets, content_id, "", content.as_bytes());
+            drop(content);
             let xo = if xobjects.is_empty() { String::new() } else { format!(" /XObject << {xobjects}>>") };
-            objects.push((
+            dict_obj(
+                &mut out,
+                &mut offsets,
                 page_id,
-                format!(
-                    "<< /Type /Page /Parent 2 0 R /MediaBox [0 0 {} {}] /Resources << /ProcSet [/PDF /Text /ImageB /ImageC] /Font << {}>>{} >> /Contents {} 0 R >>",
+                &format!(
+                    "<< /Type /Page /Parent 2 0 R /MediaBox [0 0 {} {}] /Resources << /Font << {}>>{} >> /Contents {} 0 R >>",
                     n(page.width),
                     n(page.height),
                     font_dict,
                     xo,
                     content_id
-                )
-                .into_bytes(),
-            ));
+                ),
+            );
             kids.push(page_id);
         }
-        objects.push((1, b"<< /Type /Catalog /Pages 2 0 R >>".to_vec()));
+        dict_obj(&mut out, &mut offsets, 1, "<< /Type /Catalog /Pages 2 0 R >>");
         let kids_s: Vec<String> = kids.iter().map(|k| format!("{k} 0 R")).collect();
-        objects.push((2, format!("<< /Type /Pages /Kids [{}] /Count {} >>", kids_s.join(" "), kids.len()).into_bytes()));
+        dict_obj(&mut out, &mut offsets, 2, &format!("<< /Type /Pages /Kids [{}] /Count {} >>", kids_s.join(" "), kids.len()));
         let creator = if self.info.creator.is_empty() { "GridCraft" } else { &self.info.creator };
         let mut info = format!("<< /Producer {} /Creator {}", pdf_text_string("GridCraft PDF"), pdf_text_string(creator));
         if !self.info.title.is_empty() {
@@ -650,27 +863,14 @@ impl PdfDoc {
             let _ = write!(info, " /Author {}", pdf_text_string(&self.info.author));
         }
         info.push_str(" >>");
-        objects.push((3, info.into_bytes()));
-        objects.sort_by_key(|(id, _)| *id);
+        dict_obj(&mut out, &mut offsets, 3, &info);
 
-        let mut out: Vec<u8> = Vec::new();
-        out.extend_from_slice(b"%PDF-1.7\n%\xE2\xE3\xCF\xD3\n");
-        let mut offsets = vec![0usize; next as usize];
-        for (id, body) in &objects {
-            if let Some(slot) = offsets.get_mut(*id as usize) {
-                *slot = out.len();
-            }
-            out.extend_from_slice(format!("{id} 0 obj\n").as_bytes());
-            out.extend_from_slice(body);
-            out.extend_from_slice(b"\nendobj\n");
-        }
         let xref = out.len();
-        let mut x = format!("xref\n0 {}\n0000000000 65535 f \n", next);
+        let _ = write!(Utf8(&mut out), "xref\n0 {}\n0000000000 65535 f \n", next);
         for off in offsets.iter().skip(1) {
-            let _ = writeln!(x, "{off:010} 00000 n ");
+            let _ = writeln!(Utf8(&mut out), "{off:010} 00000 n ");
         }
-        let _ = write!(x, "trailer\n<< /Size {} /Root 1 0 R /Info 3 0 R >>\nstartxref\n{}\n%%EOF\n", next, xref);
-        out.extend_from_slice(x.as_bytes());
+        let _ = write!(Utf8(&mut out), "trailer\n<< /Size {} /Root 1 0 R /Info 3 0 R >>\nstartxref\n{}\n%%EOF\n", next, xref);
         out
     }
 }
diff --git a/crates/pdf/src/tests.rs b/crates/pdf/src/tests.rs
index e7b488a20..96a879656 100644
--- a/crates/pdf/src/tests.rs
+++ b/crates/pdf/src/tests.rs
@@ -1,5 +1,8 @@
 use super::*;
 
+mod reference;
+use reference::{OldPage, interpret};
+
 fn find(hay: &[u8], needle: &[u8]) -> Option<usize> {
     hay.windows(needle.len()).position(|w| w == needle)
 }
@@ -157,3 +160,530 @@ fn hostile_inputs_never_panic() {
     let pdf = d.finish();
     check_xref(&pdf);
 }
+
+// ---------------------------------------------------------------------------------------------
+// Writer cleanup: equivalence with the previous content writer, structure, and one test per
+// behaviour change.
+// ---------------------------------------------------------------------------------------------
+
+/// Offsets of the in-use objects, by id, from the xref table.
+fn xref_offsets(pdf: &[u8]) -> Vec<usize> {
+    let sx = rfind(pdf, b"startxref\n").unwrap();
+    let rest = std::str::from_utf8(&pdf[sx + 10..]).unwrap();
+    let off: usize = rest.lines().next().unwrap().trim().parse().unwrap();
+    let table = std::str::from_utf8(&pdf[off..sx]).unwrap();
+    table.lines().skip(2).take_while(|l| l.len() == 18 || l.len() == 19).map(|e| e[..10].parse().unwrap()).collect()
+}
+
+struct Obj {
+    dict: String,
+    stream: Option<Vec<u8>>,
+}
+
+/// Parses every object of a document written by this crate, checking that each stream's
+/// `/Length` matches its data and is followed by `endstream`/`endobj`.
+fn parse_objects(pdf: &[u8]) -> Vec<Obj> {
+    let offsets = xref_offsets(pdf);
+    let mut objs = Vec::new();
+    for (id, off) in offsets.iter().enumerate() {
+        if id == 0 {
+            objs.push(Obj { dict: String::new(), stream: None });
+            continue;
+        }
+        let tail = &pdf[*off..];
+        let head = format!("{id} 0 obj\n");
+        assert!(tail.starts_with(head.as_bytes()), "object {id}");
+        let body = &tail[head.len()..];
+        let end = find(body, b"\nendobj\n").unwrap();
+        match find(body, b"\nstream\n").filter(|s| *s < end) {
+            Some(s) => {
+                let dict = String::from_utf8(body[..s].to_vec()).unwrap();
+                let l = dict.find("/Length ").unwrap() + 8;
+                let len: usize = dict[l..].split(' ').next().unwrap().parse().unwrap();
+                let data = &body[s + 8..];
+                assert!(data[len..].starts_with(b"\nendstream\nendobj\n"), "object {id}: /Length {len} is wrong");
+                objs.push(Obj { dict, stream: Some(data[..len].to_vec()) });
+            }
+            None => objs.push(Obj { dict: String::from_utf8(body[..end].to_vec()).unwrap(), stream: None }),
+        }
+    }
+    objs
+}
+
+fn ref_of(dict: &str, key: &str) -> Option<usize> {
+    let i = dict.find(key)? + key.len();
+    dict[i..].trim_start().split(' ').next()?.parse().ok()
+}
+
+struct ParsedPage {
+    media: String,
+    content: String,
+    /// `(Im name, object id)`.
+    xobjects: Vec<(String, usize)>,
+}
+
+fn parse_pages(pdf: &[u8]) -> (Vec<Obj>, Vec<ParsedPage>) {
+    let objs = parse_objects(pdf);
+    let mut pages = Vec::new();
+    for o in &objs {
+        if !o.dict.contains("/Type /Page ") {
+            continue;
+        }
+        let m = o.dict.find("/MediaBox [").unwrap() + 10;
+        let media = o.dict[m..m + o.dict[m..].find(']').unwrap() + 1].to_string();
+        let content_id = ref_of(&o.dict, "/Contents").unwrap();
+        let content = String::from_utf8(objs[content_id].stream.clone().unwrap()).unwrap();
+        let mut xobjects = Vec::new();
+        if let Some(x) = o.dict.find("/XObject << ") {
+            let rest = &o.dict[x + 11..];
+            let rest = &rest[..rest.find(">>").unwrap()];
+            let toks: Vec<&str> = rest.split_whitespace().collect();
+            for c in toks.chunks(4) {
+                xobjects.push((c[0].to_string(), c[1].parse().unwrap()));
+            }
+        }
+        pages.push(ParsedPage { media, content, xobjects });
+    }
+    (objs, pages)
+}
+
+fn rle_decode(d: &[u8]) -> Vec<u8> {
+    let mut out = Vec::new();
+    let mut i = 0;
+    while i < d.len() {
+        let l = d[i];
+        i += 1;
+        if l == 128 {
+            break;
+        } else if l < 128 {
+            out.extend_from_slice(&d[i..i + l as usize + 1]);
+            i += l as usize + 1;
+        } else {
+            out.extend(std::iter::repeat_n(d[i], 257 - l as usize));
+            i += 1;
+        }
+    }
+    out
+}
+
+/// Decodes an image XObject: `(width, height, rgb, alpha)`.
+fn decode_image(objs: &[Obj], id: usize) -> (u32, u32, Vec<u8>, Option<Vec<u8>>) {
+    let o = &objs[id];
+    let w = ref_of(&o.dict, "/Width").unwrap() as u32;
+    let h = ref_of(&o.dict, "/Height").unwrap() as u32;
+    let rgb = rle_decode(o.stream.as_ref().unwrap());
+    let alpha = ref_of(&o.dict, "/SMask").map(|m| rle_decode(objs[m].stream.as_ref().unwrap()));
+    (w, h, rgb, alpha)
+}
+
+/// Deterministic pseudo-random numbers for the corpus.
+struct Lcg(u64);
+impl Lcg {
+    fn next(&mut self) -> u64 {
+        self.0 = self.0.wrapping_mul(6364136223846793005).wrapping_add(1442695040888963407);
+        self.0 >> 33
+    }
+    fn below(&mut self, n: u64) -> u64 {
+        self.next() % n
+    }
+    fn f(&mut self) -> f32 {
+        self.below(600) as f32 + self.below(4) as f32 * 0.25
+    }
+}
+
+const PALETTE: [Rgb; 4] = [[0, 0, 0], [255, 0, 0], [12, 200, 99], [255, 255, 255]];
+const WIDTHS: [f32; 3] = [0.5, 1.0, 2.5];
+const DASHES: [&[f32]; 3] = [&[], &[3.0, 2.0], &[1.0]];
+
+/// A picture used by the corpus: `(rgba, w, h)`.
+fn pictures() -> Vec<(Vec<u8>, u32, u32)> {
+    let mut with_alpha = Vec::new();
+    for i in 0..16u8 {
+        with_alpha.extend_from_slice(&[i * 10, 0, 7, if i < 8 { 255 } else { 128 }]);
+    }
+    let opaque: Vec<u8> = (0..4 * 9).map(|i| if i % 4 == 3 { 255 } else { (i * 7) as u8 }).collect();
+    let mut last_px = vec![200u8; 4 * 6];
+    for px in last_px.chunks_mut(4) {
+        px[3] = 255;
+    }
+    last_px[4 * 5 + 3] = 3; // only the very last pixel is translucent
+    vec![(with_alpha, 4, 4), (opaque, 3, 3), (last_px, 3, 2)]
+}
+
+/// Applies a random operation to both writers.
+fn random_op(rng: &mut Lcg, new: &mut Page, old: &mut OldPage, pics: &[(Vec<u8>, u32, u32)], depth: &mut usize) {
+    let color = PALETTE[rng.below(4) as usize];
+    let width = WIDTHS[rng.below(3) as usize];
+    let dash = DASHES[rng.below(3) as usize];
+    let (x, y, w, h) = (rng.f(), rng.f(), rng.f() / 4.0 + 1.0, rng.f() / 4.0 + 1.0);
+    let pts = [[x, y], [x + w, y], [x + w, y + h]];
+    let stroke = if rng.below(2) == 0 { Some((PALETTE[rng.below(4) as usize], WIDTHS[rng.below(3) as usize])) } else { None };
+    let fill = if rng.below(2) == 0 { Some(color) } else { None };
+    match rng.below(14) {
+        0 | 1 => {
+            new.fill_rect(x, y, w, h, color);
+            old.fill_rect(x, y, w, h, color);
+        }
+        2 => {
+            new.rect_stroke(x, y, w, h, width, color);
+            old.rect_stroke(x, y, w, h, width, color);
+        }
+        3 => {
+            new.stroke_line_dashed(x, y, x + w, y + h, width, color, dash);
+            old.stroke_line_dashed(x, y, x + w, y + h, width, color, dash);
+        }
+        4 => {
+            new.polyline(&pts, width, color, dash);
+            old.polyline(&pts, width, color, dash);
+        }
+        5 => {
+            new.polygon(&pts, fill, stroke);
+            old.polygon(&pts, fill, stroke);
+        }
+        6 => {
+            new.circle(x, y, w, fill, stroke);
+            old.circle(x, y, w, fill, stroke);
+        }
+        7 | 8 => {
+            let font = Font::ALL[rng.below(12) as usize];
+            let angle = if rng.below(3) == 0 { 0.5 } else { 0.0 };
+            new.text_rotated(x, y, 10.0, font, color, "Hi (there)", angle);
+            old.text_rotated(x, y, 10.0, font, color, "Hi (there)", angle);
+        }
+        9 => {
+            new.clip_rect(x, y, w, h);
+            old.clip_rect(x, y, w, h);
+        }
+        10 if *depth < 20 => {
+            new.save();
+            old.save();
+            *depth += 1;
+        }
+        11 if *depth > 0 => {
+            new.restore();
+            old.restore();
+            *depth -= 1;
+        }
+        12 => {
+            let (px, pw, ph) = &pics[rng.below(3) as usize];
+            new.image_rgba(x, y, w, h, px, *pw, *ph);
+            old.image(x, y, w, h);
+        }
+        _ => {
+            new.text(x, y, 9.0, Font::Helvetica, color, "Plain");
+            old.text_rotated(x, y, 9.0, Font::Helvetica, color, "Plain", 0.0);
+        }
+    }
+}
+
+/// Builds the same random multi-page document with the new writer and the old content writer.
+fn corpus_doc(seed: u64) -> (Vec<u8>, Vec<String>) {
+    let mut rng = Lcg(seed);
+    let pics = pictures();
+    let mut doc = PdfDoc::new();
+    let mut old_pages = Vec::new();
+    for _ in 0..(1 + rng.below(4)) {
+        let mut old = OldPage::new(792.0);
+        let mut depth = 0usize;
+        let page = doc.add_page(612.0, 792.0);
+        for _ in 0..120 {
+            random_op(&mut rng, page, &mut old, &pics, &mut depth);
+        }
+        while depth > 0 {
+            page.restore();
+            old.restore();
+            depth -= 1;
+        }
+        old_pages.push(old.content);
+    }
+    (doc.finish(), old_pages)
+}
+
+#[test]
+fn new_writer_is_equivalent_to_the_old_one() {
+    let mut style_new = 0;
+    let mut style_old = 0;
+    for seed in 1..=40u64 {
+        let (pdf, old_pages) = corpus_doc(seed);
+        check_xref(&pdf);
+        let (_, pages) = parse_pages(&pdf);
+        assert_eq!(pages.len(), old_pages.len());
+        for (p, old) in pages.iter().zip(&old_pages) {
+            assert_eq!(p.media, "[0 0 612 792]");
+            let new = interpret(&p.content);
+            let old = interpret(old);
+            assert!(new.balanced && old.balanced, "q/Q balanced (seed {seed})");
+            assert_eq!(new.paints.len(), old.paints.len(), "seed {seed}");
+            for (i, (a, b)) in new.paints.iter().zip(&old.paints).enumerate() {
+                // Round joins leaking from polylines into later rectangles is the old bug (B7),
+                // compared separately in `join_does_not_leak_from_polylines`.
+                let mut a = a.clone();
+                let mut b = b.clone();
+                a.join.clear();
+                b.join.clear();
+                assert_eq!(a, b, "seed {seed}, painting op {i}");
+            }
+            style_new += new.style_ops;
+            style_old += old.style_ops;
+        }
+    }
+    // Sanity: the corpus really contains style changes, and the new writer emits fewer.
+    assert!(style_old > 1000, "corpus too small: {style_old}");
+    assert!(style_new < style_old, "{style_new} vs {style_old}");
+}
+
+#[test]
+fn structure_is_valid_for_all_documents() {
+    for seed in 1..=40u64 {
+        let (pdf, _) = corpus_doc(seed);
+        check_xref(&pdf);
+        parse_objects(&pdf); // every /Length matches its data
+    }
+    let empty = PdfDoc::new().finish();
+    check_xref(&empty);
+    parse_objects(&empty);
+}
+
+#[test]
+fn images_decode_to_their_pixels() {
+    let pics = pictures();
+    let mut d = PdfDoc::new();
+    for _ in 0..2 {
+        let p = d.add_page(300.0, 300.0);
+        for (px, w, h) in &pics {
+            p.image_rgba(1.0, 1.0, 20.0, 20.0, px, *w, *h);
+        }
+    }
+    let pdf = d.finish();
+    let (objs, pages) = parse_pages(&pdf);
+    assert_eq!(pages.len(), 2);
+    for page in &pages {
+        assert_eq!(page.xobjects.len(), 3);
+        for ((px, w, h), (_, id)) in pics.iter().zip(&page.xobjects) {
+            let (dw, dh, rgb, alpha) = decode_image(&objs, *id);
+            assert_eq!((dw, dh), (*w, *h));
+            let want_rgb: Vec<u8> = px.chunks(4).flat_map(|c| [c[0], c[1], c[2]]).collect();
+            assert_eq!(rgb, want_rgb);
+            let want_alpha: Vec<u8> = px.chunks(4).map(|c| c[3]).collect();
+            match alpha {
+                Some(a) => assert_eq!(a, want_alpha),
+                None => assert!(want_alpha.iter().all(|a| *a == 255)),
+            }
+        }
+    }
+}
+
+/// O6: no alpha plane (and no SMask object) for an opaque image; a single translucent pixel
+/// anywhere (first, middle, last) produces the full plane.
+#[test]
+fn opaque_images_have_no_soft_mask() {
+    let mut d = PdfDoc::new();
+    let p = d.add_page(100.0, 100.0);
+    p.image_rgba(0.0, 0.0, 10.0, 10.0, &[255; 4 * 12], 4, 3);
+    assert!(p.images[0].alpha.is_none());
+    let pdf = d.finish();
+    assert!(find(&pdf, b"/SMask").is_none());
+    assert_eq!(pdf.windows(15).filter(|w| *w == b"/Subtype /Image").count(), 1);
+
+    for translucent in [0usize, 5, 11] {
+        let mut px = vec![255u8; 4 * 12];
+        px[translucent * 4 + 3] = 7;
+        let mut d = PdfDoc::new();
+        d.add_page(100.0, 100.0).image_rgba(0.0, 0.0, 10.0, 10.0, &px, 4, 3);
+        let (objs, pages) = parse_pages(&d.finish());
+        let (_, _, _, alpha) = decode_image(&objs, pages[0].xobjects[0].1);
+        let mut want = vec![255u8; 12];
+        want[translucent] = 7;
+        assert_eq!(alpha, Some(want));
+    }
+}
+
+/// O4: the same picture on five pages is stored once (image + mask), by every page.
+#[test]
+fn identical_images_are_stored_once() {
+    let pics = pictures();
+    let (px, w, h) = &pics[0]; // has alpha: image object + mask object
+    let mut d = PdfDoc::new();
+    for _ in 0..5 {
+        let p = d.add_page(200.0, 200.0);
+        p.image_rgba(1.0, 1.0, 20.0, 20.0, px, *w, *h);
+        p.image_rgba(30.0, 1.0, 20.0, 20.0, px, *w, *h); // twice on one page
+    }
+    let pdf = d.finish();
+    check_xref(&pdf);
+    assert_eq!(pdf.windows(15).filter(|w| *w == b"/Subtype /Image").count(), 2, "one image + one soft mask");
+    let (_, pages) = parse_pages(&pdf);
+    let ids: Vec<usize> = pages.iter().flat_map(|p| p.xobjects.iter().map(|x| x.1)).collect();
+    assert_eq!(ids.len(), 5, "one XObject entry per page");
+    assert!(ids.iter().all(|i| *i == ids[0]));
+    for p in &pages {
+        assert_eq!(p.content.matches(" Do").count(), 2);
+    }
+}
+
+/// O4: a picture that differs in one byte, or only in size, is not merged.
+#[test]
+fn different_images_are_not_merged() {
+    let a = vec![50u8; 4 * 4];
+    let mut b = a.clone();
+    b[5] = 51;
+    let mut d = PdfDoc::new();
+    let p = d.add_page(100.0, 100.0);
+    p.image_rgba(0.0, 0.0, 5.0, 5.0, &a, 4, 1);
+    p.image_rgba(0.0, 0.0, 5.0, 5.0, &b, 4, 1);
+    p.image_rgba(0.0, 0.0, 5.0, 5.0, &a, 2, 2);
+    d.add_page(100.0, 100.0).image_rgba(0.0, 0.0, 5.0, 5.0, &a, 4, 1);
+    let pdf = d.finish();
+    assert_eq!(pdf.windows(15).filter(|w| *w == b"/Subtype /Image").count(), 6, "three distinct pictures, each with a soft mask (alpha 50)");
+}
+
+/// O4: equal hashes alone never merge images; the pixels are compared.
+#[test]
+fn hash_collisions_do_not_merge_images() {
+    let mut d = PdfDoc::new();
+    for v in [10u8, 20] {
+        d.add_page(100.0, 100.0).image_rgba(0.0, 0.0, 5.0, 5.0, &[v; 16], 2, 2);
+    }
+    for p in 0..2 {
+        d.page_mut(p).unwrap().images[0].hash = 42; // forced collision
+    }
+    let (objs, pages) = parse_pages(&d.finish());
+    let ids: Vec<usize> = pages.iter().map(|p| p.xobjects[0].1).collect();
+    assert_ne!(ids[0], ids[1]);
+    assert_eq!(decode_image(&objs, ids[0]).2, vec![10; 12]);
+    assert_eq!(decode_image(&objs, ids[1]).2, vec![20; 12]);
+}
+
+/// S5: the colour component table equals the old float formatting for every byte.
+#[test]
+fn colour_components_match_the_float_formatting() {
+    for v in 0..=255u8 {
+        let mut s = String::new();
+        push_unit(&mut s, v);
+        assert_eq!(s, n(v as f32 / 255.0), "byte {v}");
+    }
+    let mut s = String::new();
+    push_rgb(&mut s, [255, 128, 0], true);
+    assert_eq!(s, "1 0.502 0 RG");
+}
+
+/// S3: the obsolete /ProcSet entry is gone.
+#[test]
+fn no_procset() {
+    let (pdf, _) = corpus_doc(3);
+    assert!(find(&pdf, b"/ProcSet").is_none());
+}
+
+/// O3: a style that is already in effect is not emitted again.
+#[test]
+fn unchanged_style_is_emitted_once() {
+    let mut d = PdfDoc::new();
+    let p = d.add_page(200.0, 200.0);
+    for i in 0..50 {
+        p.rect_stroke(i as f32, 0.0, 5.0, 5.0, 1.5, [10, 20, 30]);
+    }
+    for i in 0..50 {
+        p.stroke_line_dashed(0.0, i as f32, 9.0, 9.0, 1.5, [10, 20, 30], &[2.0, 1.0]);
+    }
+    for _ in 0..50 {
+        p.fill_rect(0.0, 0.0, 4.0, 4.0, [1, 2, 3]);
+        p.text(5.0, 5.0, 9.0, Font::Helvetica, [1, 2, 3], "x");
+    }
+    let s = interpret(&p.content);
+    assert_eq!(p.content.matches(" RG\n").count(), 1);
+    assert_eq!(p.content.matches(" rg\n").count(), 1);
+    assert_eq!(p.content.matches(" w\n").count(), 1);
+    assert_eq!(p.content.matches(" d\n").count(), 2, "solid, then [2 1]");
+    assert_eq!(s.style_ops, 5);
+    assert_eq!(s.paints.len(), 50 * 4);
+}
+
+/// O3: the tracked style follows q/Q, and a new page starts from scratch.
+#[test]
+fn style_tracking_follows_save_restore_and_pages() {
+    let red = [255, 0, 0];
+    let blue = [0, 0, 255];
+    let mut d = PdfDoc::new();
+    let p = d.add_page(100.0, 100.0);
+    p.fill_rect(0.0, 0.0, 5.0, 5.0, red);
+    p.save();
+    p.fill_rect(0.0, 0.0, 5.0, 5.0, red); // still red inside: nothing to emit
+    p.fill_rect(0.0, 0.0, 5.0, 5.0, blue);
+    p.restore();
+    p.fill_rect(0.0, 0.0, 5.0, 5.0, red); // red again after Q without a new operator
+    p.fill_rect(0.0, 0.0, 5.0, 5.0, blue); // but blue must be re-emitted
+    let fills: Vec<String> = interpret(&p.content).paints.iter().map(|p| p.fill.clone()).collect();
+    assert_eq!(fills, ["1 0 0", "1 0 0", "0 0 1", "1 0 0", "0 0 1"]);
+    assert_eq!(p.content.matches(" rg\n").count(), 3, "red once, blue twice");
+    let p2 = d.add_page(100.0, 100.0);
+    p2.fill_rect(0.0, 0.0, 5.0, 5.0, red);
+    assert_eq!(p2.content.matches(" rg\n").count(), 1, "a new page re-emits the colour");
+}
+
+/// B6: only 28 nested saves are written, the matching extra restores are no-ops.
+#[test]
+fn save_nesting_is_limited_to_28() {
+    let mut d = PdfDoc::new();
+    let p = d.add_page(100.0, 100.0);
+    for _ in 0..40 {
+        p.save();
+    }
+    p.fill_rect(0.0, 0.0, 5.0, 5.0, [1, 1, 1]);
+    assert_eq!(p.content.matches("q\n").count(), 28);
+    // The 12 restores that match refused saves do nothing...
+    for _ in 0..12 {
+        p.restore();
+    }
+    assert_eq!(p.content.matches("Q\n").count(), 0);
+    assert_eq!(interpret(&p.content).max_depth, 28);
+    // ...the next 28 undo the accepted ones, and further restores are ignored.
+    for _ in 0..40 {
+        p.restore();
+    }
+    assert_eq!(p.content.matches("Q\n").count(), 28);
+    let i = interpret(&p.content);
+    assert!(i.balanced);
+    // An image needs its own q: dropped at the limit, drawn otherwise.
+    for _ in 0..28 {
+        p.save();
+    }
+    p.image_rgba(0.0, 0.0, 5.0, 5.0, &[9; 16], 2, 2);
+    assert!(!p.content.contains(" Do"));
+    p.restore();
+    p.image_rgba(0.0, 0.0, 5.0, 5.0, &[9; 16], 2, 2);
+    assert!(p.content.contains(" Do"));
+    let pdf = d.finish();
+    check_xref(&pdf);
+}
+
+/// B6: `finish()` closes what is still open, without exceeding the limit.
+#[test]
+fn finish_closes_open_saves() {
+    let mut d = PdfDoc::new();
+    let p = d.add_page(100.0, 100.0);
+    for _ in 0..33 {
+        p.save();
+    }
+    let (_, pages) = parse_pages(&d.finish());
+    let i = interpret(&pages[0].content);
+    assert!(i.balanced);
+    assert_eq!(i.qs, 28);
+}
+
+/// B7: round joins set for a polyline do not leak into later rectangles, lines and circles.
+#[test]
+fn join_does_not_leak_from_polylines() {
+    let mut d = PdfDoc::new();
+    let p = d.add_page(200.0, 200.0);
+    let pts = [[0.0, 0.0], [10.0, 0.0], [10.0, 10.0]];
+    p.rect_stroke(0.0, 0.0, 5.0, 5.0, 1.0, [0, 0, 0]);
+    p.polyline(&pts, 1.0, [0, 0, 0], &[]);
+    p.polyline(&pts, 1.0, [0, 0, 0], &[]);
+    p.rect_stroke(0.0, 0.0, 5.0, 5.0, 1.0, [0, 0, 0]);
+    p.stroke_line(0.0, 0.0, 9.0, 9.0, 1.0, [0, 0, 0]);
+    p.polygon(&pts, None, Some(([0, 0, 0], 1.0)));
+    p.circle(50.0, 50.0, 5.0, None, Some(([0, 0, 0], 1.0)));
+    let joins: Vec<String> = interpret(&p.content).paints.iter().map(|p| p.join.clone()).collect();
+    assert_eq!(joins, ["0", "1", "1", "0", "0", "1", "0"]);
+    assert_eq!(p.content.matches(" j\n").count(), 4, "1, 0, 1, 0: only on change");
+}
diff --git a/crates/pdf/src/tests/reference.rs b/crates/pdf/src/tests/reference.rs
new file mode 100644
index 000000000..d3a46dc21
--- /dev/null
+++ b/crates/pdf/src/tests/reference.rs
@@ -0,0 +1,349 @@
+//! The content-stream emission of the writer before the cleanup (a verbatim copy of the old
+//! per-operator formatting, which re-emitted the full style for every operator, had no `q` limit
+//! and left `1 j` set after polylines), plus an interpreter that reduces a content stream to the
+//! effective graphics state of every painting operator.
+
+use super::super::*;
+
+fn old_rgb_op(c: Rgb, stroke: bool) -> String {
+    let f = |v: u8| n(v as f32 / 255.0);
+    format!("{} {} {} {}", f(c[0]), f(c[1]), f(c[2]), if stroke { "RG" } else { "rg" })
+}
+
+/// The old content writer.
+#[derive(Default)]
+pub struct OldPage {
+    pub height: f32,
+    pub content: String,
+    depth: u32,
+    images: usize,
+}
+
+impl OldPage {
+    pub fn new(height: f32) -> OldPage {
+        OldPage { height, ..Default::default() }
+    }
+    fn y(&self, y: f32) -> f32 {
+        self.height - y
+    }
+    fn op(&mut self, s: &str) {
+        self.content.push_str(s);
+        self.content.push('\n');
+    }
+    pub fn save(&mut self) {
+        self.depth += 1;
+        self.content.push_str("q\n");
+    }
+    pub fn restore(&mut self) {
+        if self.depth > 0 {
+            self.depth -= 1;
+            self.content.push_str("Q\n");
+        }
+    }
+    pub fn clip_rect(&mut self, x: f32, y: f32, w: f32, h: f32) {
+        let s = format!("{} {} {} {} re W n", n(x), n(self.y(y + h)), n(w.max(0.0)), n(h.max(0.0)));
+        self.op(&s);
+    }
+    pub fn fill_rect(&mut self, x: f32, y: f32, w: f32, h: f32, color: Rgb) {
+        if w <= 0.0 || h <= 0.0 {
+            return;
+        }
+        let s = format!("{}\n{} {} {} {} re f", old_rgb_op(color, false), n(x), n(self.y(y + h)), n(w), n(h));
+        self.op(&s);
+    }
+    pub fn rect_stroke(&mut self, x: f32, y: f32, w: f32, h: f32, width: f32, color: Rgb) {
+        let s = format!("{}\n{} w [] 0 d\n{} {} {} {} re S", old_rgb_op(color, true), n(width), n(x), n(self.y(y + h)), n(w), n(h));
+        self.op(&s);
+    }
+    pub fn stroke_line_dashed(&mut self, x0: f32, y0: f32, x1: f32, y1: f32, width: f32, color: Rgb, dash: &[f32]) {
+        let d: Vec<String> = dash.iter().take(8).map(|v| n(v.max(0.0))).collect();
+        let s = format!(
+            "{}\n{} w [{}] 0 d\n{} {} m {} {} l S",
+            old_rgb_op(color, true),
+            n(width),
+            d.join(" "),
+            n(x0),
+            n(self.y(y0)),
+            n(x1),
+            n(self.y(y1))
+        );
+        self.op(&s);
+    }
+    pub fn polyline(&mut self, pts: &[[f32; 2]], width: f32, color: Rgb, dash: &[f32]) {
+        if pts.len() < 2 {
+            return;
+        }
+        let d: Vec<String> = dash.iter().take(8).map(|v| n(v.max(0.0))).collect();
+        let mut s = format!("{}\n{} w [{}] 0 d 1 j\n", old_rgb_op(color, true), n(width), d.join(" "));
+        self.path(&mut s, pts, false);
+        s.push_str(" S");
+        self.op(&s);
+    }
+    fn path(&self, s: &mut String, pts: &[[f32; 2]], close: bool) {
+        for (i, p) in pts.iter().enumerate() {
+            let _ = write!(s, "{} {} {} ", n(p[0]), n(self.y(p[1])), if i == 0 { "m" } else { "l" });
+        }
+        if close {
+            s.push('h');
+        }
+    }
+    pub fn polygon(&mut self, pts: &[[f32; 2]], fill: Option<Rgb>, stroke: Option<(Rgb, f32)>) {
+        if pts.len() < 2 || (fill.is_none() && stroke.is_none()) {
+            return;
+        }
+        let mut s = String::new();
+        if let Some(f) = fill {
+            s.push_str(&old_rgb_op(f, false));
+            s.push('\n');
+        }
+        if let Some((c, w)) = stroke {
+            let _ = writeln!(s, "{}\n{} w [] 0 d 1 j", old_rgb_op(c, true), n(w));
+        }
+        self.path(&mut s, pts, true);
+        s.push_str(match (fill.is_some(), stroke.is_some()) {
+            (true, true) => " B",
+            (true, false) => " f",
+            _ => " S",
+        });
+        self.op(&s);
+    }
+    pub fn circle(&mut self, cx: f32, cy: f32, r: f32, fill: Option<Rgb>, stroke: Option<(Rgb, f32)>) {
+        if r <= 0.0 || (fill.is_none() && stroke.is_none()) {
+            return;
+        }
+        let y = self.y(cy);
+        let mut s = String::new();
+        if let Some(f) = fill {
+            s.push_str(&old_rgb_op(f, false));
+            s.push('\n');
+        }
+        if let Some((c, w)) = stroke {
+            let _ = writeln!(s, "{}\n{} w [] 0 d", old_rgb_op(c, true), n(w));
+        }
+        const K: f32 = 0.552_284_8;
+        let k = r * K;
+        let _ = write!(
+            s,
+            "{} {} m {} {} {} {} {} {} c {} {} {} {} {} {} c {} {} {} {} {} {} c {} {} {} {} {} {} c h",
+            n(cx + r),
+            n(y),
+            n(cx + r),
+            n(y + k),
+            n(cx + k),
+            n(y + r),
+            n(cx),
+            n(y + r),
+            n(cx - k),
+            n(y + r),
+            n(cx - r),
+            n(y + k),
+            n(cx - r),
+            n(y),
+            n(cx - r),
+            n(y - k),
+            n(cx - k),
+            n(y - r),
+            n(cx),
+            n(y - r),
+            n(cx + k),
+            n(y - r),
+            n(cx + r),
+            n(y - k),
+            n(cx + r),
+            n(y)
+        );
+        s.push_str(match (fill.is_some(), stroke.is_some()) {
+            (true, true) => " B",
+            (true, false) => " f",
+            _ => " S",
+        });
+        self.op(&s);
+    }
+    pub fn text_rotated(&mut self, x: f32, y: f32, size: f32, font: Font, color: Rgb, text: &str, angle: f32) {
+        let fi = font.index();
+        let (sn, cs) = angle.sin_cos();
+        let s = format!(
+            "BT\n{}\n/F{} {} Tf\n{} {} {} {} {} {} Tm\n{} Tj\nET",
+            old_rgb_op(color, false),
+            fi + 1,
+            n(size),
+            n(cs),
+            n(-sn),
+            n(sn),
+            n(cs),
+            n(x),
+            n(self.y(y)),
+            pdf_string(text)
+        );
+        self.op(&s);
+    }
+    pub fn image(&mut self, x: f32, y: f32, w: f32, h: f32) {
+        self.images += 1;
+        let s = format!("q {} 0 0 {} {} {} cm /Im{} Do Q", n(w), n(h), n(x), n(self.y(y + h)), self.images);
+        self.op(&s);
+    }
+}
+
+/// The effective state at one painting operator. Fields that cannot influence the operator
+/// (the stroke state of a fill, the fill of a stroke) are left empty.
+#[derive(Clone, Debug, PartialEq, Default)]
+pub struct Paint {
+    pub op: String,
+    /// Path construction / text operands since the previous painting operator.
+    pub geometry: String,
+    pub fill: String,
+    pub stroke: String,
+    pub width: String,
+    pub dash: String,
+    pub join: String,
+    /// Clip paths in effect (innermost last).
+    pub clip: Vec<String>,
+}
+
+#[derive(Default, Debug)]
+pub struct Interp {
+    pub paints: Vec<Paint>,
+    pub max_depth: usize,
+    /// Number of `q` operators.
+    pub qs: usize,
+    /// Number of colour/width/dash/join operators.
+    pub style_ops: usize,
+    pub balanced: bool,
+}
+
+fn tokens(content: &str) -> Vec<String> {
+    let mut out = Vec::new();
+    let mut cur = String::new();
+    let mut chars = content.chars();
+    while let Some(c) = chars.next() {
+        match c {
+            '(' => {
+                if !cur.is_empty() {
+                    out.push(std::mem::take(&mut cur));
+                }
+                let mut s = String::from("(");
+                while let Some(d) = chars.next() {
+                    s.push(d);
+                    if d == '\\' {
+                        if let Some(e) = chars.next() {
+                            s.push(e);
+                        }
+                    } else if d == ')' {
+                        break;
+                    }
+                }
+                out.push(s);
+            }
+            '[' | ']' => {
+                if !cur.is_empty() {
+                    out.push(std::mem::take(&mut cur));
+                }
+                out.push(c.to_string());
+            }
+            c if c.is_whitespace() => {
+                if !cur.is_empty() {
+                    out.push(std::mem::take(&mut cur));
+                }
+            }
+            c => cur.push(c),
+        }
+    }
+    if !cur.is_empty() {
+        out.push(cur);
+    }
+    out
+}
+
+#[derive(Clone)]
+struct State {
+    fill: String,
+    stroke: String,
+    width: String,
+    dash: String,
+    join: String,
+    clip: Vec<String>,
+}
+
+/// Interprets a content stream: tracks q/Q, colours, width, dash, join and clips.
+pub fn interpret(content: &str) -> Interp {
+    let mut st = State { fill: "0 0 0".into(), stroke: "0 0 0".into(), width: "1".into(), dash: String::new(), join: "0".into(), clip: Vec::new() };
+    let mut stack: Vec<State> = Vec::new();
+    let mut r = Interp { balanced: true, ..Default::default() };
+    let mut operands: Vec<String> = Vec::new();
+    let mut path: Vec<String> = Vec::new();
+    let mut pending_clip = false;
+    for t in tokens(content) {
+        let is_op = t.chars().next().is_some_and(|c| c.is_ascii_alphabetic() || c == '*') && !t.starts_with('/');
+        if !is_op {
+            operands.push(t);
+            continue;
+        }
+        let ops = std::mem::take(&mut operands);
+        match t.as_str() {
+            "q" => {
+                stack.push(st.clone());
+                r.qs += 1;
+                r.max_depth = r.max_depth.max(stack.len());
+            }
+            "Q" => match stack.pop() {
+                Some(s) => st = s,
+                None => r.balanced = false,
+            },
+            "rg" => {
+                st.fill = ops.join(" ");
+                r.style_ops += 1;
+            }
+            "RG" => {
+                st.stroke = ops.join(" ");
+                r.style_ops += 1;
+            }
+            "w" => {
+                st.width = ops.join(" ");
+                r.style_ops += 1;
+            }
+            "d" => {
+                // `[a b] 0` -> `a b`
+                st.dash = ops.iter().take_while(|o| *o != "]").skip(1).cloned().collect::<Vec<_>>().join(" ");
+                r.style_ops += 1;
+            }
+            "j" => {
+                st.join = ops.join(" ");
+                r.style_ops += 1;
+            }
+            "re" | "m" | "l" | "c" | "h" | "cm" | "Tf" | "Tm" => {
+                path.push(format!("{} {}", ops.join(" "), t));
+            }
+            "BT" | "ET" => {}
+            "W" => pending_clip = true,
+            "n" => {
+                if pending_clip {
+                    st.clip.push(path.join(" "));
+                    pending_clip = false;
+                }
+                path.clear();
+            }
+            "Tj" => {
+                path.push(format!("{} {}", ops.join(" "), t));
+                r.paints.push(Paint { op: t.clone(), geometry: path.join(" "), fill: st.fill.clone(), clip: st.clip.clone(), ..Default::default() });
+                path.clear();
+            }
+            "f" | "S" | "B" | "Do" => {
+                let mut p = Paint { op: t.clone(), geometry: path.join(" "), clip: st.clip.clone(), ..Default::default() };
+                if t == "f" || t == "B" {
+                    p.fill = st.fill.clone();
+                }
+                if t == "S" || t == "B" {
+                    p.stroke = st.stroke.clone();
+                    p.width = st.width.clone();
+                    p.dash = st.dash.clone();
+                    p.join = st.join.clone();
+                }
+                r.paints.push(p);
+                path.clear();
+            }
+            _ => {}
+        }
+    }
+    r.balanced &= stack.is_empty();
+    r
+}

From cca09cc8f49fc716aff0f34be9e79ec901173c78 Mon Sep 17 00:00:00 2001
From: vangel76 <vangel1976@googlemail.com>
Date: Sat, 7 Nov 2026 17:19:10 +0100
Subject: [PATCH 2/3] PDF: track the first text run with a flag

text_rotated counted the Tf operators written so far on every run, which is
quadratic for text that alternates between WinAnsi and CID characters.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
---
 crates/pdf/src/lib.rs   |  3 ++-
 crates/pdf/src/tests.rs | 12 ++++++++++++
 2 files changed, 14 insertions(+), 1 deletion(-)

diff --git a/crates/pdf/src/lib.rs b/crates/pdf/src/lib.rs
index 2b5a92679..9fcb6bd75 100644
--- a/crates/pdf/src/lib.rs
+++ b/crates/pdf/src/lib.rs
@@ -743,8 +743,9 @@ impl Page {
             if let Some(f) = self.fonts.get_mut(slot) {
                 *f = true;
             }
+            let first_run = runs.is_empty();
             let _ = writeln!(runs, "/F{} {} Tf", slot + 1, size);
-            if runs.matches("Tf\n").count() == 1 {
+            if first_run {
                 runs.push_str(&tm);
             }
             let _ = writeln!(runs, "{string} Tj");
diff --git a/crates/pdf/src/tests.rs b/crates/pdf/src/tests.rs
index 1ffa5b842..d5054610c 100644
--- a/crates/pdf/src/tests.rs
+++ b/crates/pdf/src/tests.rs
@@ -873,6 +873,18 @@ fn cjk_text_is_not_replaced_by_question_marks() -> Result<(), Box<dyn Error>> {
     Ok(())
 }
 
+#[test]
+fn many_alternating_runs_share_one_text_matrix() {
+    let mut doc = PdfDoc::new();
+    let p = doc.add_page(612.0, 792.0);
+    p.text(10.0, 20.0, 10.0, Font::Helvetica, [0, 0, 0], &"a\u{4e2d}".repeat(2000));
+    let pdf = doc.finish();
+    let count = |needle: &[u8]| pdf.windows(needle.len()).filter(|w| *w == needle).count();
+    assert_eq!(count(b" Tm\n"), 1, "the position is set once, before the first run");
+    assert_eq!(count(b" Tf\n"), 4000, "every run selects its font");
+    assert_eq!(count(b" Tj\n"), 4000);
+}
+
 #[test]
 fn symbol_font_draws_maths_in_its_own_encoding() {
     let mut doc = PdfDoc::new();

From 01fe67010d4dc5fdb2187a837d39b9830775e321 Mon Sep 17 00:00:00 2001
From: vangel76 <vangel1976@googlemail.com>
Date: Sun, 8 Nov 2026 12:19:46 +0100
Subject: [PATCH 3/3] PDF: look up page images by hash instead of scanning
 every earlier image

Page::image_rgba compared each new picture with all earlier pictures on
the page, so N distinct pictures cost N(N-1)/2 comparisons. Keep a
page-local hash-to-index map (pixels still compared with ImageData::same)
and add a regression test with 3,000 distinct pictures.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---
 crates/pdf/src/lib.rs   | 21 ++++++++++++++++++++-
 crates/pdf/src/tests.rs | 20 ++++++++++++++++++++
 2 files changed, 40 insertions(+), 1 deletion(-)

diff --git a/crates/pdf/src/lib.rs b/crates/pdf/src/lib.rs
index 9fcb6bd75..b94aaf6a7 100644
--- a/crates/pdf/src/lib.rs
+++ b/crates/pdf/src/lib.rs
@@ -434,6 +434,11 @@ pub struct Page {
     content: String,
     fonts: [bool; FONT_SLOTS],
     images: Vec<ImageData>,
+    /// Indexes into `images`, by content hash; a hash match is confirmed with [`ImageData::same`].
+    image_index: HashMap<u64, Vec<usize>>,
+    /// Pixel comparisons made while looking up duplicate pictures (test instrumentation).
+    #[cfg(test)]
+    image_comparisons: usize,
     /// Style operators already in effect at the end of `content`.
     style: Style,
     /// One entry per accepted [`Page::save`], restored by [`Page::restore`].
@@ -793,9 +798,20 @@ impl Page {
         }
         let img = ImageData::new(pw, ph, rgb, alpha);
         // The same picture drawn again on this page shares one XObject.
-        let idx = match self.images.iter().position(|i| i.same(&img)) {
+        let candidates = self.image_index.entry(img.hash).or_default();
+        #[cfg(test)]
+        let comparisons = &mut self.image_comparisons;
+        let known = candidates.iter().copied().find(|&i| {
+            #[cfg(test)]
+            {
+                *comparisons += 1;
+            }
+            self.images.get(i).is_some_and(|o| o.same(&img))
+        });
+        let idx = match known {
             Some(i) => i,
             None => {
+                candidates.push(self.images.len());
                 self.images.push(img);
                 self.images.len() - 1
             }
@@ -934,6 +950,9 @@ impl PdfDoc {
             content: String::new(),
             fonts: [false; FONT_SLOTS],
             images: Vec::new(),
+            image_index: HashMap::new(),
+            #[cfg(test)]
+            image_comparisons: 0,
             links: Vec::new(),
             style: Style::default(),
             saved: Vec::new(),
diff --git a/crates/pdf/src/tests.rs b/crates/pdf/src/tests.rs
index d5054610c..b86300ff0 100644
--- a/crates/pdf/src/tests.rs
+++ b/crates/pdf/src/tests.rs
@@ -571,6 +571,26 @@ fn different_images_are_not_merged() {
     assert_eq!(pdf.windows(15).filter(|w| *w == b"/Subtype /Image").count(), 6, "three distinct pictures, each with a soft mask (alpha 50)");
 }
 
+/// O4: many distinct pictures on one page are looked up by hash, not by comparing each new
+/// picture with every earlier one (which made dense picture exports quadratic).
+#[test]
+fn many_distinct_images_on_one_page_are_not_compared_pairwise() {
+    const N: u32 = 3000;
+    let mut d = PdfDoc::new();
+    let p = d.add_page(612.0, 792.0);
+    for i in 0..N {
+        let [a, b, c, _] = i.to_le_bytes();
+        p.image_rgba(Rect::new(0.0, 0.0, 1.0, 1.0), &[a, b, c, 255], [1, 1]);
+    }
+    // A repeat of the first picture still shares its XObject.
+    p.image_rgba(Rect::new(5.0, 5.0, 1.0, 1.0), &[0, 0, 0, 255], [1, 1]);
+    assert_eq!(p.images.len(), N as usize);
+    assert!(p.content.ends_with("/Im1 Do Q\n"), "the repeated picture reuses /Im1");
+    assert!(p.image_comparisons <= 1, "{} pixel comparisons for {N} distinct pictures", p.image_comparisons);
+    let pdf = d.finish();
+    assert_eq!(pdf.windows(15).filter(|w| *w == b"/Subtype /Image").count(), N as usize);
+}
+
 /// O4: equal hashes alone never merge images; the pixels are compared.
 #[test]
 fn hash_collisions_do_not_merge_images() {