engine: keep the compare test's inserted slides empty

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] = [
-        &|x, y| x.name.starts_with("!!") && x.name == y.name,
-        &|x, y| !x.name.is_empty() && x.name == y.name,
-        &|x, y| {
-            if x.kind_name() != y.kind_name() {
-                return false;
+    // Each shape's text, worked out once rather than inside every comparison.
+    let text_a: Vec<String> = a.shapes.iter().map(text_of).collect();
+    let text_b: Vec<String> = b.shapes.iter().map(text_of).collect();
+    let forced = |s: &Shape, _: &str| (s.name.starts_with("!!")).then(|| MorphKey::Name(s.name.clone()));
+    let named = |s: &Shape, _: &str| (!s.name.is_empty()).then(|| MorphKey::Name(s.name.clone()));
+    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);
-            !tx.trim().is_empty() && tx == text_of(y)
-        },
-        &|x, y| {
-            x.kind_name() == y.kind_name()
-                && x.geom.preset_name().is_some()
-                && x.geom.preset_name() == y.geom.preset_name()
-                && text_of(x).trim().is_empty()
-                && text_of(y).trim().is_empty()
-        },
-    ];
-    for pass in passes {
+            if let Some(k) = key(sb, text_b.get(j).map_or("", String::as_str)) {
+                index.entry(k).or_default().push_back(j);
+            }
+        }
+        if index.is_empty() {
+            continue;
+        }
         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));
-            if let Some((j, sb)) = found {
-                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));
+            let Some(k) = key(sa, text_a.get(i).map_or("", String::as_str)) else { continue };
+            let Some(j) = index.get_mut(&k).and_then(VecDeque::pop_front) else { continue };
+            let Some(sb) = b.shapes.get(j) else { continue };
+            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));
         }
     }
     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();
-        v.extend(self.groups.iter().filter(|g| g.0 == shape && v.contains(&g.2)).map(|g| g.1).collect::<Vec<_>>());
+        let mine = self.index.by_shape.get(&shape).map_or(&[][..], Vec::as_slice);
+        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();
-        for e in &self.effects {
-            if !v.contains(&e.timing.shape) {
-                v.push(e.timing.shape);
-            }
-        }
-        v
+        let mut seen: HashSet<ShapeId> = HashSet::new();
+        self.effects.iter().map(|e| e.timing.shape).filter(|s| seen.insert(*s)).collect()
     }
 
     /// 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 mine = self.effects.iter().filter(|e| e.timing.shape == shape && e.timing.paragraph == para);
-        let base_color = self.colors.iter().find(|(s, _)| *s == shape).map(|(_, c)| *c).unwrap_or(Rgba::rgb(0x15, 0x60, 0x82));
+        let para = para.map(|p| self.index.owner.get(&(shape, p)).copied().unwrap_or(p));
+        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.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/compare.rs

@@ -121,9 +121,10 @@ pub fn diff(cur: &Presentation, rev: &Presentation, by_position: bool) -> Vec<Ch
     let cur_matched: HashSet<usize> = pairs.iter().map(|p| p.0).collect();
     let rev_matched: HashSet<usize> = pairs.iter().map(|p| p.1).collect();
     let mut out = vec![];
+    let mut media = Media::new(cur, rev);
     for (i, j) in &pairs {
         let (Some(a), Some(b)) = (cur.slides.get(*i), rev.slides.get(*j)) else { continue };
-        let details = slide_details(cur, a, rev, b);
+        let details = slide_details(cur, a, rev, b, &mut media);
         if !details.is_empty() {
             out.push(Change { kind: ChangeKind::Changed, slide: *i, number: i.saturating_add(1), only_in: None, revised: Some(*j), details });
         }
@@ -140,10 +141,18 @@ pub fn diff(cur: &Presentation, rev: &Presentation, by_position: bool) -> Vec<Ch
             });
         }
     }
+    // Each revised slide's open match (the first pair naming it).
+    let mut rev_to_cur: HashMap<usize, usize> = HashMap::new();
+    for p in &pairs {
+        rev_to_cur.entry(p.1).or_insert(p.0);
+    }
+    // After the open slide that matches the nearest matched revised slide before it.
+    let mut at = 0;
     for (j, s) in rev.slides.iter().enumerate() {
+        if let Some(i) = rev_to_cur.get(&j) {
+            at = i + 1;
+        }
         if !rev_matched.contains(&j) {
-            // After the open slide that matches the revised slide before it.
-            let at = (0..j).rev().find_map(|k| pairs.iter().find(|p| p.1 == k).map(|p| p.0 + 1)).unwrap_or(0);
             out.push(Change {
                 kind: ChangeKind::Inserted,
                 slide: at,
@@ -174,8 +183,34 @@ fn shape_label(s: &Shape) -> String {
     if s.name.trim().is_empty() { format!("Shape {}", s.id.0) } else { s.name.clone() }
 }
 
-fn media_bytes(p: &Presentation, id: MediaId) -> Option<&[u8]> {
-    p.media(id).map(|m| m.data.as_slice())
+/// Both decks' media bytes by id, and which picture pairs were already compared: a deck can
+/// show one picture many times, and its bytes are compared once per pair.
+struct Media<'a> {
+    cur: HashMap<MediaId, &'a [u8]>,
+    rev: HashMap<MediaId, &'a [u8]>,
+    same: HashMap<(MediaId, MediaId), bool>,
+}
+
+impl<'a> Media<'a> {
+    fn new(cur: &'a Presentation, rev: &'a Presentation) -> Self {
+        // The first item with an id wins, as in `Presentation::media`.
+        let index = |p: &'a Presentation| {
+            let mut m = HashMap::new();
+            for x in &p.media {
+                m.entry(x.id).or_insert(x.data.as_slice());
+            }
+            m
+        };
+        Media { cur: index(cur), rev: index(rev), same: HashMap::new() }
+    }
+    /// Whether the pictures (by media id, `None` for a shape that is not a picture) differ.
+    fn differ(&mut self, a: Option<MediaId>, b: Option<MediaId>) -> bool {
+        let (ba, bb) = (a.and_then(|x| self.cur.get(&x)), b.and_then(|y| self.rev.get(&y)));
+        match (a, b) {
+            (Some(x), Some(y)) if ba.is_some() && bb.is_some() => !*self.same.entry((x, y)).or_insert_with(|| ba == bb),
+            _ => ba.is_some() || bb.is_some(),
+        }
+    }
 }
 
 /// A shape without its ids (shape and media), for comparing across decks.
@@ -204,7 +239,7 @@ fn turn(ctx: Option<&Ctx>, s: &Shape) -> (f64, bool, bool) {
     drawn_box(ctx, s).map_or((0.0, false, false), |x| (x.rot, x.flip_h, x.flip_v))
 }
 
-fn shape_details(cp: &Presentation, ca: Option<&Ctx>, a: &Shape, rp: &Presentation, cb: Option<&Ctx>, b: &Shape, out: &mut Vec<String>) {
+fn shape_details(ca: Option<&Ctx>, a: &Shape, cb: Option<&Ctx>, b: &Shape, media: &mut Media, out: &mut Vec<String>) {
     let label = shape_label(b);
     let text = |s: &Shape| s.text.as_ref().map(|t| t.text()).unwrap_or_default();
     let mut what = vec![];
@@ -234,11 +269,11 @@ fn shape_details(cp: &Presentation, ca: Option<&Ctx>, a: &Shape, rp: &Presentati
     if a.fill != b.fill || a.line != b.line || a.style != b.style || a.effects != b.effects {
         what.push("formatting");
     }
-    let pic = |p: &Presentation, s: &Shape| match &s.kind {
-        ShapeKind::Picture { fill } => media_bytes(p, fill.media).map(|b| b.to_vec()),
+    let pic = |s: &Shape| match &s.kind {
+        ShapeKind::Picture { fill } => Some(fill.media),
         _ => None,
     };
-    if pic(cp, a) != pic(rp, b) {
+    if media.differ(pic(a), pic(b)) {
         what.push("picture");
     }
     // Anything the checks above do not cover (alt text, hidden, a link…) is reported too, also
@@ -265,21 +300,32 @@ fn shape_details(cp: &Presentation, ca: Option<&Ctx>, a: &Shape, rp: &Presentati
     }
 }
 
-fn slide_details(cp: &Presentation, a: &Slide, rp: &Presentation, b: &Slide) -> Vec<String> {
+fn slide_details(cp: &Presentation, a: &Slide, rp: &Presentation, b: &Slide, media: &mut Media) -> Vec<String> {
     let mut out = vec![];
     let mut used: HashSet<usize> = HashSet::new();
     let (ca, cb) = (Ctx::for_slide(cp, a), Ctx::for_slide(rp, b));
+    // The revised shapes by id (first wins) and by name (in order, with a cursor past used ones).
+    let mut by_id: HashMap<ShapeId, usize> = HashMap::new();
+    let mut by_name: HashMap<&str, (Vec<usize>, usize)> = HashMap::new();
+    for (k, x) in b.shapes.iter().enumerate() {
+        by_id.entry(x.id).or_insert(k);
+        if !x.name.is_empty() {
+            by_name.entry(x.name.as_str()).or_default().0.push(k);
+        }
+    }
     for sa in &a.shapes {
-        let k = b
-            .shapes
-            .iter()
-            .position(|x| x.id == sa.id)
-            .or_else(|| b.shapes.iter().enumerate().position(|(k, x)| !used.contains(&k) && x.name == sa.name && !x.name.is_empty()));
+        let k = by_id.get(&sa.id).copied().or_else(|| {
+            let (ks, next) = by_name.get_mut(sa.name.as_str())?;
+            while ks.get(*next).is_some_and(|k| used.contains(k)) {
+                *next += 1;
+            }
+            ks.get(*next).copied()
+        });
         match k {
             Some(k) if !used.contains(&k) => {
                 used.insert(k);
                 if let Some(sb) = b.shapes.get(k) {
-                    shape_details(cp, ca.as_ref(), sa, rp, cb.as_ref(), sb, &mut out);
+                    shape_details(ca.as_ref(), sa, cb.as_ref(), sb, media, &mut out);
                 }
             }
             _ => out.push(format!("{} deleted (only in the deck)", shape_label(sa))),
@@ -334,19 +380,34 @@ pub fn adopt(doc: &mut Presentation, rev: &Presentation, slide: &Slide, replacin
     let mut sl = slide.clone();
     // Media.
     let mut media_map: HashMap<MediaId, MediaId> = HashMap::new();
+    // Both decks' media by id (the first item with an id wins, as in `Presentation::media`),
+    // looked up once per reference instead of searched.
+    let mut rev_at: HashMap<MediaId, usize> = HashMap::new();
+    for (i, m) in rev.media.iter().enumerate() {
+        rev_at.entry(m.id).or_insert(i);
+    }
+    let mut doc_at: HashMap<MediaId, usize> = HashMap::new();
+    for (i, m) in doc.media.iter().enumerate() {
+        doc_at.entry(m.id).or_insert(i);
+    }
     for id in crate::media_refs::slide_media(&sl) {
-        let Some(m) = rev.media(id) else { continue };
-        match doc.media(id) {
+        let Some(m) = rev_at.get(&id).and_then(|&i| rev.media.get(i)) else { continue };
+        match doc_at.get(&id).and_then(|&i| doc.media.get(i)) {
             Some(x) if x.data == m.data => {}
             None => {
                 // Keeps the revision's id, so later ids must start past it.
+                doc_at.insert(id, doc.media.len());
                 doc.media.push(m.clone());
                 doc.next_id = doc.next_id.max(id.0);
             }
             Some(_) => {
+                let before = doc.media.len();
                 let new = doc
                     .add_media(&m.name, &m.content_type, m.data.as_ref().clone())
                     .ok_or_else(|| crate::cmd::bad("review.accept", crate::cmd::NO_ID_SPACE))?;
+                if doc.media.len() > before {
+                    doc_at.entry(new).or_insert(before);
+                }
                 media_map.insert(id, new);
             }
         }
@@ -519,6 +580,39 @@ mod tests {
         assert_eq!(s.doc().unwrap().doc.slides.len(), n, "comparing alone changes nothing");
     }
 
+    #[test]
+    fn many_inserted_slides_and_same_named_pictures_compare_by_index() {
+        use std::sync::Arc;
+
+        use deckcraft_model::style::PictureFill;
+        use deckcraft_model::{Shape, ShapeId, ShapeKind, SlideId, defaults};
+
+        // One slide of 3,000 pictures of one image, all named "Box"; the revision has the same
+        // pictures under other ids (matched by name, in order) and 3,000 slides inserted after it.
+        let mut cur = defaults::new_presentation(None);
+        let img = cur.add_media("a.png", "image/png", vec![7; 4096]).unwrap();
+        let mut rev = cur.clone();
+        let pic = |id: u32| Shape {
+            id: ShapeId(id),
+            name: "Box".into(),
+            xfrm: Some(deckcraft_model::Xfrm::new(10.0, 10.0, 50.0, 50.0)),
+            kind: ShapeKind::Picture { fill: PictureFill { media: img, ..Default::default() } },
+            ..Default::default()
+        };
+        Arc::make_mut(&mut cur.slides[0]).shapes = (0..3000).map(|k| pic(100_000 + k)).collect();
+        Arc::make_mut(&mut rev.slides[0]).shapes = (0..3000).map(|k| pic(200_000 + k)).collect();
+        let mut base = (*rev.slides[0]).clone();
+        base.shapes.clear();
+        for k in 0..3000 {
+            let mut s = base.clone();
+            s.id = SlideId(300_000 + k);
+            rev.slides.push(Arc::new(s));
+        }
+        let changes = super::diff(&cur, &rev, false);
+        assert_eq!(changes.len(), 3000, "only the inserted slides differ");
+        assert!(changes.iter().all(|c| c.kind == super::ChangeKind::Inserted && c.slide == 1));
+    }
+
     /// A deck with one slide: a table whose cell has a picture fill, a picture background and a
     /// transition sound. `tag` starts the bytes of each.
     fn media_deck(tag: &str) -> deckcraft_model::Presentation {

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();
-        buf.set_pre_context(pre);
-        buf.set_post_context(post);
-        for (i, c) in text.char_indices() {
-            buf.add(map(c), (range.start + i) as u32);
-        }
-        buf.guess_segment_properties();
-        if let Some(s) = params.script.and_then(|t| Script::from_iso15924_tag(Tag::new(&t))) {
-            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));
-        for (info, pos) in gb.glyph_infos().iter().zip(gb.glyph_positions()) {
-            out.push(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,
-            });
+        // Pieces of one item share the script the whole item would be guessed to have.
+        let item_script = (chunks.len() > 1 && params.script.is_none()).then(|| {
+            let mut probe = UnicodeBuffer::new();
+            for c in text.chars() {
+                probe.add(map(c), 0);
+            }
+            probe.guess_segment_properties();
+            probe.script()
+        });
+        for c in &chunks {
+            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 };
+            let mut buf = UnicodeBuffer::new();
+            buf.set_pre_context(pre);
+            buf.set_post_context(post);
+            for (i, c) in piece.char_indices() {
+                buf.add(map(c), (r.start + i) as u32);
+            }
+            buf.guess_segment_properties();
+            if let Some(s) = params.script.and_then(|t| Script::from_iso15924_tag(Tag::new(&t))) {
+                buf.set_script(s);
+            } else if let Some(s) = item_script {
+                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()
-        && let Some(f) = face.skrifa()
-    {
+    if shaped.is_none() {
         // No shaper for this face: one glyph per character from the cmap, in visual order.
-        let cmap = f.charmap();
-        let gm = f.glyph_metrics(Size::unscaled(), face.location());
-        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 params.rtl {
-            out.reverse();
+        return cmap_glyphs(face, para, range, params.rtl, &map);
+    }
+    // Each piece is in visual order; right to left, the later pieces come first.
+    if params.rtl {
+        pieces.reverse();
+    }
+    let mut out = Vec::with_capacity(pieces.iter().map(Vec::len).sum());
+    for p in pieces {
+        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));
+    }
 }