address review: check the original Path for the dictionary write; isolate the confinement test

The shaper lets one call grow to 256 glyphs per input character, as an
embedded font's substitutions decide, and every glyph was then copied out.
Items longer than 1 KiB are now shaped in pieces (cut at a space where one
is near, never before a combining mark or after a joiner), each with the
rest of the paragraph as context and the item's script, and a piece that
expands past 32 glyphs per character falls back to one glyph per character
from the cmap.

crates/anim/src/morph.rs

@@ -1,5 +1,7 @@
 //! Morph: matching shapes between two slides, interpolating their boxes and building the frames.
 
+use std::collections::{HashMap, HashSet, VecDeque};
+
 use deckcraft_model::resolve::{self, Ctx};
 use deckcraft_model::{ColorRef, Fill, Geom, Presentation, Rgba, Shape, ShapeId, ShapeKind, Slide, Xfrm, walk};
 
@@ -15,48 +17,66 @@ fn text_of(s: &Shape) -> String {
 /// kind and preset geometry if neither carries text. A text box with neither a name nor its text on
 /// the other slide stays unpaired and fades in place, as in PowerPoint (#119). Each shape is used
 /// once; top-level shapes only (groups move whole).
+///
+/// Each pass compares a key of the two shapes, so the shapes of `b` are indexed by key once per
+/// pass: matching costs the sum of the two shape lists, not their product.
 pub fn morph_pairs(a: &Slide, b: &Slide) -> Vec<(ShapeId, ShapeId)> {
     let mut pairs: Vec<(ShapeId, ShapeId)> = Vec::new();
     let mut used_a = vec![false; a.shapes.len()];
     let mut used_b = vec![false; b.shapes.len()];
-    let passes: [&dyn Fn(&Shape, &Shape) -> bool; 4] = [
+    // Each shape's text, worked out once rather than inside every comparison.
-        &|x, y| x.name.starts_with("!!") && x.name == y.name,
+    let text_a: Vec<String> = a.shapes.iter().map(text_of).collect();
-        &|x, y| !x.name.is_empty() && x.name == y.name,
+    let text_b: Vec<String> = b.shapes.iter().map(text_of).collect();
-        &|x, y| {
+    let forced = |s: &Shape, _: &str| (s.name.starts_with("!!")).then(|| MorphKey::Name(s.name.clone()));
-            if x.kind_name() != y.kind_name() {
+    let named = |s: &Shape, _: &str| (!s.name.is_empty()).then(|| MorphKey::Name(s.name.clone()));
-                return false;
+    let same_text = |s: &Shape, t: &str| (!t.trim().is_empty()).then(|| MorphKey::Text(s.kind_name(), t.to_string()));
+    let same_geom = |s: &Shape, t: &str| match s.geom.preset_name() {
+        Some(g) if t.trim().is_empty() => Some(MorphKey::Geom(s.kind_name(), g.to_string())),
+        _ => None,
+    };
+    let passes: [&dyn Fn(&Shape, &str) -> Option<MorphKey>; 4] = [&forced, &named, &same_text, &same_geom];
+    for key in passes {
+        // Unused shapes of `b` by key, in slide order: the front of each queue is the first unused
+        // match, which is the one a search through `b` would find.
+        let mut index: HashMap<MorphKey, VecDeque<usize>> = HashMap::new();
+        for (j, sb) in b.shapes.iter().enumerate() {
+            if used_b.get(j).copied().unwrap_or(true) {
+                continue;
             }
-            let tx = text_of(x);
+            if let Some(k) = key(sb, text_b.get(j).map_or("", String::as_str)) {
-            !tx.trim().is_empty() && tx == text_of(y)
+                index.entry(k).or_default().push_back(j);
-        },
+            }
-        &|x, y| {
+        }
-            x.kind_name() == y.kind_name()
+        if index.is_empty() {
-                && x.geom.preset_name().is_some()
+            continue;
-                && x.geom.preset_name() == y.geom.preset_name()
+        }
-                && text_of(x).trim().is_empty()
-                && text_of(y).trim().is_empty()
-        },
-    ];
-    for pass in passes {
         for (i, sa) in a.shapes.iter().enumerate() {
             if used_a.get(i).copied().unwrap_or(true) {
                 continue;
             }
-            let found = b.shapes.iter().enumerate().find(|(j, sb)| !used_b.get(*j).copied().unwrap_or(true) && pass(sa, sb));
+            let Some(k) = key(sa, text_a.get(i).map_or("", String::as_str)) else { continue };
-            if let Some((j, sb)) = found {
+            let Some(j) = index.get_mut(&k).and_then(VecDeque::pop_front) else { continue };
-                if let Some(u) = used_a.get_mut(i) {
+            let Some(sb) = b.shapes.get(j) else { continue };
-                    *u = true;
+            if let Some(u) = used_a.get_mut(i) {
-                }
+                *u = true;
-                if let Some(u) = used_b.get_mut(j) {
-                    *u = true;
-                }
-                pairs.push((sa.id, sb.id));
             }
+            if let Some(u) = used_b.get_mut(j) {
+                *u = true;
+            }
+            pairs.push((sa.id, sb.id));
         }
     }
     pairs
 }
 
+/// What one [`morph_pairs`] pass matches shapes on.
+#[derive(Clone, Debug, PartialEq, Eq, Hash)]
+enum MorphKey {
+    Name(String),
+    Text(&'static str, String),
+    Geom(&'static str, String),
+}
+
 fn lerp(a: f64, b: f64, t: f64) -> f64 {
     let v = a + (b - a) * t;
     if v.is_finite() {
@@ -256,13 +276,23 @@ pub fn morph_frame(pres: &Presentation, a: &Slide, b: &Slide, t: f64, hidden_b:
         f.shapes.push((sh, old));
         id
     };
+    // Lookups by id, built once (the first shape or pair with an id wins, as a search would).
+    let paired_a: HashSet<ShapeId> = pairs.iter().map(|(ia, _)| *ia).collect();
+    let mut partner: HashMap<ShapeId, ShapeId> = HashMap::with_capacity(pairs.len());
+    for (ia, ib) in &pairs {
+        partner.entry(*ib).or_insert(*ia);
+    }
+    let mut a_by_id: HashMap<ShapeId, &Shape> = HashMap::with_capacity(a.shapes.len());
+    for x in &a.shapes {
+        a_by_id.entry(x.id).or_insert(x);
+    }
     for sa in &a.shapes {
-        if !pairs.iter().any(|(ia, _)| *ia == sa.id) {
+        if !paired_a.contains(&sa.id) {
             push(sa.clone(), true, 1.0 - s);
         }
     }
     for sb in &b.shapes {
-        let pair = pairs.iter().find(|(_, ib)| *ib == sb.id).and_then(|(ia, _)| a.shapes.iter().find(|x| x.id == *ia));
+        let pair = partner.get(&sb.id).and_then(|ia| a_by_id.get(ia).copied());
         let Some(sa) = pair else {
             if !hidden_b(sb.id) {
                 push(sb.clone(), false, s);

crates/anim/src/tests.rs

@@ -1155,3 +1155,58 @@ fn many_by_paragraph_animations_on_one_shape_share_paragraph_targets() {
     assert_eq!(tl.effects().len(), 6_000);
     assert_eq!(tl.paragraph_targets(ShapeId(1)), vec![0, 1, 2]);
 }
+
+#[test]
+fn morph_pairs_take_the_first_unused_match_in_slide_order() {
+    let mut a = Slide::default();
+    let mut b = Slide::default();
+    a.shapes = vec![shape(1, "Same"), shape(2, "Same"), shape(3, "Other")];
+    b.shapes = vec![shape(11, "Same"), shape(12, "Same"), shape(13, "Same")];
+    // By name first, in slide order; the shape of `a` left over pairs by kind and geometry with
+    // the shape of `b` left over.
+    assert_eq!(morph_pairs(&a, &b), [(ShapeId(1), ShapeId(11)), (ShapeId(2), ShapeId(12)), (ShapeId(3), ShapeId(13))]);
+}
+
+#[test]
+fn morph_of_large_slides_stays_linear() {
+    let n = 20_000u32;
+    let text = |i: u32, side: &str| Some(TextBody::from_text(&format!("{side} {i}")));
+    let mut a = Slide::default();
+    let mut b = Slide::default();
+    // Text no shape on the other slide shares: every pass looks for a partner and finds none.
+    a.shapes = (0..n).map(|i| Shape { text: text(i, "a"), ..shape(i + 1, "") }).collect();
+    b.shapes = (0..n).map(|i| Shape { text: text(i, "b"), ..shape(n + i + 1, "") }).collect();
+    // One named pair at opposite ends.
+    a.shapes[0].name = "Logo".into();
+    b.shapes[n as usize - 1].name = "Logo".into();
+    // No wall-clock bound (it would flake on a loaded runner): at this size a quadratic pairing
+    // takes minutes, so a regression shows as a test that hangs rather than a failed assertion.
+    assert_eq!(morph_pairs(&a, &b), [(ShapeId(1), ShapeId(2 * n))]);
+    let p = deckcraft_model::defaults::new_presentation(None);
+    let f = morph_frame(&p, &a, &b, 0.5, &|_| false);
+    // Every shape fades in place; the named pair differs in text, so it cross-fades as two.
+    assert_eq!(f.shapes.len(), 2 * n as usize);
+}
+
+#[test]
+fn paragraph_states_of_a_large_build_stay_linear() {
+    let n = 20_000usize;
+    let mut s = slide_with(vec![]);
+    let mut body = TextBody::from_text(&vec!["line"; n].join("\n"));
+    // One first-level paragraph owning every other paragraph.
+    for p in body.paragraphs.iter_mut().skip(1) {
+        p.level = 1;
+    }
+    s.shapes[0].text = Some(body);
+    let mut a = anim(1, AnimClass::Entrance, "fade", AnimStart::OnClick);
+    a.text_build = TextBuild::ByParagraph;
+    s.animations = vec![a];
+    let tl = Timeline::new(&s, W, H);
+    // No wall-clock bound, as above: a per-paragraph scan of the slide would take minutes here.
+    assert_eq!(tl.paragraph_targets(ShapeId(1)).len(), n);
+    for p in 0..n {
+        assert!(!tl.state(ShapeId(1), Some(p), 0, 0.0).visible);
+        assert!(tl.state(ShapeId(1), Some(p), 1, 0.0).visible);
+    }
+    assert_eq!(tl.animated_shapes(), [ShapeId(1)]);
+}

crates/anim/src/timeline.rs

@@ -103,6 +103,41 @@ pub struct Timeline {
     /// By-paragraph builds animate first-level paragraphs; deeper paragraphs follow their parent:
     /// (shape, member paragraph, owner paragraph).
     groups: Vec<(ShapeId, usize, usize)>,
+    /// Lookups built with the timeline, so evaluating a shape or paragraph costs its own effects
+    /// rather than a scan of every effect, group and colour on the slide.
+    index: TimelineIndex,
+}
+
+#[derive(Clone, Debug, Default)]
+struct TimelineIndex {
+    /// Effects (indices into `effects`, in order) by target shape and paragraph.
+    by_target: HashMap<(ShapeId, Option<usize>), Vec<usize>>,
+    /// Effects (indices into `effects`, in order) by target shape.
+    by_shape: HashMap<ShapeId, Vec<usize>>,
+    /// Owner paragraph of each member paragraph (the first group recorded wins).
+    owner: HashMap<(ShapeId, usize), usize>,
+    /// Groups (member, owner) by shape, in order.
+    groups_by_shape: HashMap<ShapeId, Vec<(usize, usize)>>,
+    /// Base colour by shape (the first recorded wins).
+    color: HashMap<ShapeId, Rgba>,
+}
+
+impl TimelineIndex {
+    fn build(effects: &[Effect], groups: &[(ShapeId, usize, usize)], colors: &[(ShapeId, Rgba)]) -> TimelineIndex {
+        let mut ix = TimelineIndex::default();
+        for (i, e) in effects.iter().enumerate() {
+            ix.by_target.entry((e.timing.shape, e.timing.paragraph)).or_default().push(i);
+            ix.by_shape.entry(e.timing.shape).or_default().push(i);
+        }
+        for &(shape, member, owner) in groups {
+            ix.owner.entry((shape, member)).or_insert(owner);
+            ix.groups_by_shape.entry(shape).or_default().push((member, owner));
+        }
+        for &(shape, c) in colors {
+            ix.color.entry(shape).or_insert(c);
+        }
+        ix
+    }
 }
 
 enum Status {
@@ -324,6 +359,7 @@ impl Timeline {
             };
             tl.colors.push((id, c));
         }
+        tl.index = TimelineIndex::build(&tl.effects, &tl.groups, &tl.colors);
         tl
     }
 
@@ -413,8 +449,12 @@ impl Timeline {
 
     /// Paragraph indices of `shape` animated separately (by-paragraph builds).
     pub fn paragraph_targets(&self, shape: ShapeId) -> Vec<usize> {
-        let mut v: Vec<usize> = self.effects.iter().filter(|e| e.timing.shape == shape).filter_map(|e| e.timing.paragraph).collect();
+        let mine = self.index.by_shape.get(&shape).map_or(&[][..], Vec::as_slice);
-        v.extend(self.groups.iter().filter(|g| g.0 == shape && v.contains(&g.2)).map(|g| g.1).collect::<Vec<_>>());
+        let mut v: Vec<usize> = mine.iter().filter_map(|&i| self.effects.get(i)).filter_map(|e| e.timing.paragraph).collect();
+        let owners: HashSet<usize> = v.iter().copied().collect();
+        if let Some(groups) = self.index.groups_by_shape.get(&shape) {
+            v.extend(groups.iter().filter(|(_, owner)| owners.contains(owner)).map(|(member, _)| *member));
+        }
         v.sort_unstable();
         v.dedup();
         v
@@ -422,13 +462,8 @@ impl Timeline {
 
     /// Shapes with any effect.
     pub fn animated_shapes(&self) -> Vec<ShapeId> {
-        let mut v: Vec<ShapeId> = Vec::new();
+        let mut seen: HashSet<ShapeId> = HashSet::new();
-        for e in &self.effects {
+        self.effects.iter().map(|e| e.timing.shape).filter(|s| seen.insert(*s)).collect()
-            if !v.contains(&e.timing.shape) {
-                v.push(e.timing.shape);
-            }
-        }
-        v
     }
 
     /// State of `shape` (or of one of its paragraphs, `para`) when `step` main steps have been
@@ -456,9 +491,9 @@ impl Timeline {
         triggers: &[TriggerPlay],
         stamps: &[StepStamp],
     ) -> AnimState {
-        let para = para.map(|p| self.groups.iter().find(|g| g.0 == shape && g.1 == p).map(|g| g.2).unwrap_or(p));
+        let para = para.map(|p| self.index.owner.get(&(shape, p)).copied().unwrap_or(p));
-        let mine = self.effects.iter().filter(|e| e.timing.shape == shape && e.timing.paragraph == para);
+        let mine = self.index.by_target.get(&(shape, para)).map_or(&[][..], Vec::as_slice).iter().filter_map(|&i| self.effects.get(i));
-        let base_color = self.colors.iter().find(|(s, _)| *s == shape).map(|(_, c)| *c).unwrap_or(Rgba::rgb(0x15, 0x60, 0x82));
+        let base_color = self.index.color.get(&shape).copied().unwrap_or(Rgba::rgb(0x15, 0x60, 0x82));
         let mut st = AnimState::default();
         // Initial visibility comes from the first effect in the Animation Pane order.
         let mut ordered: Vec<&Effect> = mine.collect();

crates/engine/src/file_access.rs

@@ -85,14 +85,20 @@ impl AutomationRoots {
 
     /// May the file or folder at `path` be read (or written)? `Err` says why not.
     pub fn check(&self, path: &str, access: Access) -> Result<(), String> {
+        self.check_path(Path::new(path), access)
+    }
+
+    /// [`Self::check`] for a path that may not be UTF-8: the checked path is the one opened.
+    pub fn check_path(&self, path: &Path, access: Access) -> Result<(), String> {
+        let path_text = path.display();
         let roots = match access {
             Access::Read => &self.read,
             Access::Write => &self.write,
         };
         if roots.is_empty() {
             return match access {
-                Access::Read => Err(format!("automation filesystem access is not granted: read authority is absent: {path}")),
+                Access::Read => Err(format!("automation filesystem access is not granted: read authority is absent: {path_text}")),
-                Access::Write => Err(format!("automation filesystem access is not granted: write authority is absent: {path}")),
+                Access::Write => Err(format!("automation filesystem access is not granted: write authority is absent: {path_text}")),
             };
         }
         let resolved = resolve(path)?;
@@ -101,8 +107,8 @@ impl AutomationRoots {
         }
         let roots = roots.iter().map(|root| root.shown.display().to_string()).collect::<Vec<_>>().join(", ");
         match access {
-            Access::Read => Err(format!("automation path rejected: outside the read roots {roots}: {path}")),
+            Access::Read => Err(format!("automation path rejected: outside the read roots {roots}: {path_text}")),
-            Access::Write => Err(format!("automation path rejected: outside the write roots {roots}: {path}")),
+            Access::Write => Err(format!("automation path rejected: outside the write roots {roots}: {path_text}")),
         }
     }
 }
@@ -191,17 +197,23 @@ pub fn check_write(path: &str) -> Result<(), String> {
     current().map_or(Ok(()), |r| r.check(path, Access::Write))
 }
 
+/// May `path` be written here? Always, without roots in force. For a path that may not be UTF-8.
+pub fn check_write_path(path: &Path) -> Result<(), String> {
+    current().map_or(Ok(()), |r| r.check_path(path, Access::Write))
+}
+
 /// `path` made absolute with its links followed; for a path that doesn't exist (yet), its nearest
 /// existing folder's links followed and the rest appended.
-fn resolve(path: &str) -> Result<PathBuf, String> {
+fn resolve(path: &Path) -> Result<PathBuf, String> {
-    if path.is_empty() {
+    if path.as_os_str().is_empty() {
         return Err("automation path rejected: path is empty".into());
     }
+    let shown = path.display();
     // Checked as given: making `…\NUL` absolute gives the device path `\\.\NUL` on Windows.
-    if cfg!(windows) && Path::new(path).components().any(|c| matches!(c, Component::Normal(n) if is_device_name(&n.to_string_lossy()))) {
+    if cfg!(windows) && path.components().any(|c| matches!(c, Component::Normal(n) if is_device_name(&n.to_string_lossy()))) {
-        return Err(format!("automation path rejected: reserved device names are not allowed: {path}"));
+        return Err(format!("automation path rejected: reserved device names are not allowed: {shown}"));
     }
-    let absolute = std::path::absolute(path).map_err(|e| format!("automation path rejected: cannot resolve {path}: {e}"))?;
+    let absolute = std::path::absolute(path).map_err(|e| format!("automation path rejected: cannot resolve {shown}: {e}"))?;
     // The names below the nearest existing folder, last first.
     let mut missing: Vec<OsString> = vec![];
     let mut at = absolute.as_path();
@@ -212,11 +224,11 @@ fn resolve(path: &str) -> Result<PathBuf, String> {
                 return Ok(found);
             }
             Err(e) => match std::fs::symlink_metadata(at) {
-                Ok(m) if m.file_type().is_symlink() => return Err(format!("automation path rejected: a link that leads nowhere: {path}")),
+                Ok(m) if m.file_type().is_symlink() => return Err(format!("automation path rejected: a link that leads nowhere: {shown}")),
                 // There, yet its real path can't be had (no permission): nothing to compare.
-                Ok(_) => return Err(format!("automation path rejected: cannot resolve {path}: {e}")),
+                Ok(_) => return Err(format!("automation path rejected: cannot resolve {shown}: {e}")),
                 Err(e) if e.kind() == std::io::ErrorKind::NotFound => {}
-                Err(e) => return Err(format!("automation path rejected: cannot resolve {path}: {e}")),
+                Err(e) => return Err(format!("automation path rejected: cannot resolve {shown}: {e}")),
             },
         }
         let parent = at.parent();
@@ -226,12 +238,12 @@ fn resolve(path: &str) -> Result<PathBuf, String> {
                 at = parent;
             }
             (Some(Component::ParentDir), _) => {
-                return Err(format!("automation path rejected: parent traversal below a folder that doesn't exist: {path}"));
+                return Err(format!("automation path rejected: parent traversal below a folder that doesn't exist: {shown}"));
             }
             // Down to the root (or drive) and it doesn't exist either.
             _ => {
                 let e = std::io::Error::from(std::io::ErrorKind::NotFound);
-                return Err(format!("automation path rejected: cannot resolve {path}: {e}"));
+                return Err(format!("automation path rejected: cannot resolve {shown}: {e}"));
             }
         }
     }

crates/engine/src/spell.rs

@@ -668,6 +668,8 @@ pub fn add_user_word(word: &str) -> std::io::Result<()> {
         return Ok(());
     }
     let path = user_dictionary_path().ok_or_else(|| std::io::Error::new(std::io::ErrorKind::NotFound, "no folder for the user dictionary"))?;
+    // A confined automation client may write only inside its write roots, this file included.
+    crate::file_access::check_write_path(&path).map_err(|e| std::io::Error::new(std::io::ErrorKind::PermissionDenied, e))?;
     if let Some(d) = path.parent() {
         std::fs::create_dir_all(d)?;
     }
@@ -953,6 +955,20 @@ SFX D Y 1\nSFX D 0 ed/S .\n";
         assert_eq!(w, vec!["check", "done"]);
     }
 
+    #[test]
+    fn a_confined_client_cannot_write_the_user_dictionary_outside_its_roots() {
+        let root = std::env::temp_dir().join(format!("deckcraft-dict-root-{}", std::process::id()));
+        std::fs::create_dir_all(&root).unwrap();
+        let roots = crate::file_access::AutomationRoots::new(std::slice::from_ref(&root), std::slice::from_ref(&root)).unwrap();
+        // The user dictionary is a file of its own outside `root`, not whatever HOME points at.
+        let outside = std::env::temp_dir().join(format!("deckcraft-dict-outside-{}", std::process::id()));
+        set_user_dictionary_path(outside.join("user.dic"));
+        let r = crate::file_access::confine(roots.as_ref(), || add_user_word("zqxconfinedword"));
+        assert_eq!(r.map_err(|e| e.kind()), Err(std::io::ErrorKind::PermissionDenied));
+        assert!(!outside.exists(), "the refused write created nothing");
+        let _ = std::fs::remove_dir_all(&root);
+    }
+
     #[test]
     fn long_and_numeric_flags_and_latin1() {
         let d = Dictionary::parse(b"FLAG long\nSFX Aa Y 1\nSFX Aa 0 s .\n", b"1\ncat/Aa\n");

crates/fonts/src/shape.rs

@@ -37,56 +37,146 @@ pub struct ShapeParams<'a> {
     pub language: Option<&'a str>,
 }
 
+/// Most bytes of text given to the shaper at once. The shaper lets one call grow its buffers to
+/// 256 glyphs per character (a font's substitutions decide how many), so a long item is shaped in
+/// pieces of this size, each with the rest of the paragraph as context.
+const MAX_SHAPE_CHUNK: usize = 1024;
+/// Most glyphs kept per character of a piece: a piece the font expands further is laid out one
+/// glyph per character from the cmap instead.
+const MAX_GLYPHS_PER_CHAR: usize = 32;
+
 /// Shape `para[range]` with `face`, the rest of `para` as context. Glyphs are in visual order
 /// (left to right) and clusters are byte offsets into `para`. `map` substitutes characters (All
 /// Caps) without moving clusters. An empty or out-of-bounds range shapes nothing.
 pub fn shape_range(face: &FontFace, para: &str, range: Range<usize>, params: &ShapeParams, map: impl Fn(char) -> char) -> Vec<ShapedGlyph> {
-    let (Some(pre), Some(text), Some(post)) = (para.get(..range.start), para.get(range.clone()), para.get(range.end..)) else {
+    let (Some(_), Some(text), Some(_)) = (para.get(..range.start), para.get(range.clone()), para.get(range.end..)) else {
         return Vec::new();
     };
-    let mut out = Vec::with_capacity(text.len());
+    let chunks = chunks(text, MAX_SHAPE_CHUNK);
+    let mut pieces: Vec<Vec<ShapedGlyph>> = Vec::with_capacity(chunks.len());
     let shaped = face.hb().map(|hb| {
         let shaper = face.shaper.shaper(&hb).instance(face.instance.as_ref()).build();
-        let mut buf = UnicodeBuffer::new();
+        // Pieces of one item share the script the whole item would be guessed to have.
-        buf.set_pre_context(pre);
+        let item_script = (chunks.len() > 1 && params.script.is_none()).then(|| {
-        buf.set_post_context(post);
+            let mut probe = UnicodeBuffer::new();
-        for (i, c) in text.char_indices() {
+            for c in text.chars() {
-            buf.add(map(c), (range.start + i) as u32);
+                probe.add(map(c), 0);
-        }
+            }
-        buf.guess_segment_properties();
+            probe.guess_segment_properties();
-        if let Some(s) = params.script.and_then(|t| Script::from_iso15924_tag(Tag::new(&t))) {
+            probe.script()
-            buf.set_script(s);
+        });
-        }
+        for c in &chunks {
-        if let Some(l) = params.language.or_else(|| params.script.and_then(crate::script::default_language)).and_then(|l| l.parse::<Language>().ok())
+            let r = (range.start + c.start)..(range.start + c.end);
-        {
+            let (Some(pre), Some(piece), Some(post)) = (para.get(..r.start), para.get(r.clone()), para.get(r.end..)) else { continue };
-            buf.set_language(l);
+            let mut buf = UnicodeBuffer::new();
-        }
+            buf.set_pre_context(pre);
-        buf.set_direction(if params.rtl { Direction::RightToLeft } else { Direction::LeftToRight });
+            buf.set_post_context(post);
-        let gb = shaper.shape(buf, ShapeOptions::new().features(params.features));
+            for (i, c) in piece.char_indices() {
-        for (info, pos) in gb.glyph_infos().iter().zip(gb.glyph_positions()) {
+                buf.add(map(c), (r.start + i) as u32);
-            out.push(ShapedGlyph {
+            }
-                gid: info.glyph_id,
+            buf.guess_segment_properties();
-                cluster: info.cluster as usize,
+            if let Some(s) = params.script.and_then(|t| Script::from_iso15924_tag(Tag::new(&t))) {
-                x_advance: pos.x_advance,
+                buf.set_script(s);
-                x_offset: pos.x_offset,
+            } else if let Some(s) = item_script {
-                y_offset: pos.y_offset,
+                buf.set_script(s);
-            });
+            }
+            if let Some(l) =
+                params.language.or_else(|| params.script.and_then(crate::script::default_language)).and_then(|l| l.parse::<Language>().ok())
+            {
+                buf.set_language(l);
+            }
+            buf.set_direction(if params.rtl { Direction::RightToLeft } else { Direction::LeftToRight });
+            let gb = shaper.shape(buf, ShapeOptions::new().features(params.features));
+            let budget = piece.chars().count().saturating_mul(MAX_GLYPHS_PER_CHAR);
+            if gb.len() > budget {
+                // Font-controlled expansion past anything real text needs: keep it bounded.
+                pieces.push(cmap_glyphs(face, para, r, params.rtl, &map));
+                continue;
+            }
+            pieces.push(
+                gb.glyph_infos()
+                    .iter()
+                    .zip(gb.glyph_positions())
+                    .map(|(info, pos)| ShapedGlyph {
+                        gid: info.glyph_id,
+                        cluster: info.cluster as usize,
+                        x_advance: pos.x_advance,
+                        x_offset: pos.x_offset,
+                        y_offset: pos.y_offset,
+                    })
+                    .collect(),
+            );
         }
     });
-    if shaped.is_none()
+    if shaped.is_none() {
-        && let Some(f) = face.skrifa()
-    {
         // No shaper for this face: one glyph per character from the cmap, in visual order.
-        let cmap = f.charmap();
+        return cmap_glyphs(face, para, range, params.rtl, &map);
-        let gm = f.glyph_metrics(Size::unscaled(), face.location());
+    }
-        for (i, c) in text.char_indices() {
+    // Each piece is in visual order; right to left, the later pieces come first.
-            let g = cmap.map(map(c)).unwrap_or_default();
+    if params.rtl {
-            let adv = gm.advance_width(g).unwrap_or(face.upem as f32 * 0.5);
+        pieces.reverse();
-            out.push(ShapedGlyph { gid: g.to_u32(), cluster: range.start + i, x_advance: adv.round() as i32, x_offset: 0, y_offset: 0 });
+    }
-        }
+    let mut out = Vec::with_capacity(pieces.iter().map(Vec::len).sum());
-        if params.rtl {
+    for p in pieces {
-            out.reverse();
+        out.extend(p);
+    }
+    out
+}
+
+/// One glyph per character of `para[range]` from the cmap, in visual order.
+fn cmap_glyphs(face: &FontFace, para: &str, range: Range<usize>, rtl: bool, map: &impl Fn(char) -> char) -> Vec<ShapedGlyph> {
+    let (Some(text), Some(f)) = (para.get(range.clone()), face.skrifa()) else { return Vec::new() };
+    let cmap = f.charmap();
+    let gm = f.glyph_metrics(Size::unscaled(), face.location());
+    let mut out = Vec::with_capacity(text.len());
+    for (i, c) in text.char_indices() {
+        let g = cmap.map(map(c)).unwrap_or_default();
+        let adv = gm.advance_width(g).unwrap_or(face.upem as f32 * 0.5);
+        out.push(ShapedGlyph { gid: g.to_u32(), cluster: range.start + i, x_advance: adv.round() as i32, x_offset: 0, y_offset: 0 });
+    }
+    if rtl {
+        out.reverse();
+    }
+    out
+}
+
+/// `text` cut into byte ranges of at most `max` bytes (or one character, if longer), at the last
+/// space of a piece where there is one, else before a character that starts a cluster (not a
+/// combining mark or joiner, not after a joiner). Text that fits is one piece.
+fn chunks(text: &str, max: usize) -> Vec<Range<usize>> {
+    use unicode_bidi::BidiClass::{BN, NSM};
+    let mut out = Vec::new();
+    let mut start = 0;
+    while text.len() - start > max {
+        let window = text.get(start..).unwrap_or_default();
+        let mut space = None;
+        let mut boundary = None;
+        let mut any = None;
+        let mut prev: Option<char> = None;
+        for (i, c) in window.char_indices() {
+            if i > max {
+                break;
+            }
+            if i > 0 {
+                any = Some(i);
+                let joined = prev.is_some_and(|p| matches!(unicode_bidi::bidi_class(p), BN));
+                if !joined && !matches!(unicode_bidi::bidi_class(c), NSM | BN) {
+                    boundary = Some(i);
+                    if prev.is_some_and(char::is_whitespace) {
+                        space = Some(i);
+                    }
+                }
+            }
+            prev = Some(c);
         }
+        // A piece always moves forward by at least one character.
+        let first = window.chars().next().map_or(window.len(), char::len_utf8);
+        let cut = space.filter(|&i| i >= max / 2).or(boundary).or(any).unwrap_or(first).max(first);
+        out.push(start..start + cut);
+        start += cut;
+    }
+    if start < text.len() || out.is_empty() {
+        out.push(start..text.len());
     }
     out
 }
@@ -121,4 +211,37 @@ mod tests {
         // Not a char boundary.
         assert!(shape_range(&face, "سنة", 1..2, &ShapeParams::default(), |c| c).is_empty());
     }
+
+    #[test]
+    fn long_items_are_shaped_in_bounded_pieces() {
+        for text in ["word ".repeat(2_000), "e\u{301}".repeat(2_000), "x".repeat(5_000), "\u{1f469}\u{200d}\u{1f4bb} ".repeat(500)] {
+            let pieces = chunks(&text, MAX_SHAPE_CHUNK);
+            assert_eq!(pieces.first().map(|r| r.start), Some(0));
+            assert_eq!(pieces.last().map(|r| r.end), Some(text.len()));
+            for w in pieces.windows(2) {
+                assert_eq!(w[0].end, w[1].start);
+            }
+            for r in &pieces {
+                assert!(!r.is_empty() && r.len() <= MAX_SHAPE_CHUNK);
+                // Never before a combining mark, never after a joiner.
+                let rest = &text[r.end..];
+                assert!(!rest.starts_with('\u{301}'));
+                assert!(!text[..r.end].ends_with('\u{200d}') || rest.is_empty());
+            }
+        }
+        assert_eq!(chunks("short", MAX_SHAPE_CHUNK), vec![0..5]);
+        assert_eq!(chunks("", MAX_SHAPE_CHUNK), vec![0..0]);
+    }
+
+    #[test]
+    fn long_items_keep_every_cluster_in_visual_order() {
+        let face = FontDb::global().face(DEFAULT_FAMILY, "Regular");
+        let text = "ab cd ".repeat(1_000);
+        let g = shape_range(&face, &text, 0..text.len(), &ShapeParams::default(), |c| c);
+        assert_eq!(g.len(), text.len());
+        assert!(g.windows(2).all(|w| w[0].cluster < w[1].cluster));
+        let rtl = shape_range(&face, &text, 0..text.len(), &ShapeParams { rtl: true, ..Default::default() }, |c| c);
+        assert_eq!(rtl.len(), text.len());
+        assert!(rtl.windows(2).all(|w| w[0].cluster > w[1].cluster));
+    }
 }