fix(geom): report a spent containment budget instead of partial islands

Every SORTENTSTABLE rebuilt its whole entity store, so many small tables naming one large space made import quadratic. Each space is now sorted once, by the last table naming it, which already decided its order.

Co-authored-by: Brandon Thomas <echelon@gmail.com>

crates/constraints/src/auto.rs

@@ -232,6 +232,10 @@ pub fn infer_placed(d: &Drawing, placed: &[Handle], others: &[Handle], o: &Infer
     let mut extra = Vec::new();
     let mut on = Vec::new();
     for (rp, p) in g.points.iter().filter(|(r, _)| placed.contains(&r.handle)) {
+        // Nothing is added once the candidates are full, so the points left need no search.
+        if extra.len() + on.len() + out.len() >= MAX_CANDIDATES {
+            break;
+        }
         if out.iter().any(|c| c.kind == ConstraintKind::Coincident && c.refs.contains(rp)) {
             continue;
         }
@@ -267,20 +271,22 @@ pub fn infer_placed(d: &Drawing, placed: &[Handle], others: &[Handle], o: &Infer
     // Perpendicular lines meeting in a T (a perpendicular snap onto the middle of a line).
     let atol = if o.angle_tolerance.is_finite() { o.angle_tolerance.clamp(0.0, 45.0).to_radians() } else { 0.0 };
     let allow_per = o.types.is_empty() || o.types.iter().any(|t| t.eq_ignore_ascii_case(ConstraintKind::Perpendicular.name()));
-    for (r1, a1, b1) in g.lines.iter().filter(|(r, _, _)| placed.contains(&r.handle)) {
+    'per: for (r1, a1, b1) in g.lines.iter().filter(|(r, _, _)| allow_per && placed.contains(&r.handle)) {
         for (r2, a2, b2) in g.lines.iter().filter(|(r, _, _)| r.handle != r1.handle) {
+            // Full: nothing more can be added, so the pairs left need no geometry or duplicate search.
+            if out.len() >= MAX_CANDIDATES {
+                break 'per;
+            }
             let (l1, l2) = (Line::new(*a1, *b1), Line::new(*a2, *b2));
             let touches = |l: &Line, p: Vec2| l.dist(p) <= tol && (0.0..=1.0).contains(&l.param_of(p));
             let tee = [*a1, *b1].iter().any(|p| touches(&l2, *p)) || [*a2, *b2].iter().any(|p| touches(&l1, *p));
             let ang = ang_between(*b1 - *a1, *b2 - *a2);
-            let dup = out.iter().any(|c| c.kind == ConstraintKind::Perpendicular && c.refs.contains(r1) && c.refs.contains(r2));
-            if allow_per
-                && tee
-                && !dup
+            // The duplicate search scans every candidate, so it comes after the cheap tests.
+            if tee
                 && (std::f64::consts::FRAC_PI_2 - ang).abs() <= atol
                 && l1.len() > tol
                 && l2.len() > tol
-                && out.len() < MAX_CANDIDATES
+                && !out.iter().any(|c| c.kind == ConstraintKind::Perpendicular && c.refs.contains(r1) && c.refs.contains(r2))
             {
                 out.push(Constraint { id: 0, kind: ConstraintKind::Perpendicular, refs: vec![*r1, *r2], name: String::new(), expr: String::new() });
             }

crates/fonts/src/ttf.rs

@@ -89,7 +89,11 @@ impl Face {
         let Some(f) = self.font() else { return Vec::new() };
         [skrifa::string::StringId::FAMILY_NAME, skrifa::string::StringId::TYPOGRAPHIC_FAMILY_NAME, skrifa::string::StringId::FULL_NAME]
             .into_iter()
-            .flat_map(|id| f.localized_strings(id).map(|s| norm(&s.to_string())).collect::<Vec<_>>())
+            // A name table can hold many records that point at one big string: at most
+            // MAX_NAME_RECORDS per name id, each decoded only up to MAX_NAME_CHARS characters.
+            .flat_map(|id| {
+                f.localized_strings(id).take(MAX_NAME_RECORDS).map(|s| norm(&s.chars().take(MAX_NAME_CHARS).collect::<String>())).collect::<Vec<_>>()
+            })
             .collect()
     }
 }
@@ -104,6 +108,10 @@ pub fn faces(bytes: &Arc<Vec<u8>>) -> Vec<Face> {
 }
 
 const MAX_FACES_PER_FILE: u32 = 32;
+/// Name records read per name id of a face, and characters read of each: real fonts have a few
+/// dozen short ones.
+const MAX_NAME_RECORDS: usize = 64;
+const MAX_NAME_CHARS: usize = 256;
 /// Bounds of the missing-glyph search: files loaded, bytes loaded, faces checked per character.
 const MAX_FALLBACK_FILES: usize = 24;
 const MAX_FALLBACK_BYTES: usize = 160 << 20;

crates/geom/src/region.rs

@@ -110,6 +110,10 @@ fn build(polys: &[Vec<Vec2>], tol: f64) -> Option<Graph> {
     };
     let mut grid: HashMap<(usize, usize), Vec<usize>> = HashMap::new();
     let mut long: Vec<usize> = Vec::new();
+    // Every cell a segment is registered in is an index kept: past this many in all (each segment
+    // may take up to 4096 cells), the input is refused as it is for too much pair work.
+    const MAX_REGISTRATIONS: usize = 20_000_000;
+    let mut registered = 0usize;
     for (i, (a, b)) in segs.iter().enumerate() {
         let (x0, y0) = cell(a.min(*b));
         let (x1, y1) = cell(a.max(*b));
@@ -118,6 +122,10 @@ fn build(polys: &[Vec<Vec2>], tol: f64) -> Option<Graph> {
             long.push(i);
             continue;
         }
+        registered = registered.saturating_add((x1 - x0 + 1) * (y1 - y0 + 1));
+        if registered > MAX_REGISTRATIONS {
+            return None;
+        }
         for gx in x0..=x1 {
             for gy in y0..=y1 {
                 grid.entry((gx, gy)).or_default().push(i);
@@ -391,10 +399,10 @@ fn outside(core: &Graph, u: usize, steps: usize) -> Option<Vec<Vec2>> {
 }
 
 /// Outlines of the groups of curves inside the face `outer` (vertex indices into `g`).
-fn islands(g: &Graph, outer: &[usize], p: Vec2) -> Vec<Island> {
+fn islands(g: &Graph, outer: &[usize], p: Vec2) -> Option<Vec<Island>> {
     let comp = components(g);
     let outer_comp: std::collections::HashSet<usize> = outer.iter().filter_map(|&i| comp.get(i).copied()).collect();
-    let Some(outer_pts) = points(g, outer) else { return Vec::new() };
+    let outer_pts = points(g, outer)?;
     let ob = Bounds2::from_points(outer_pts.iter().copied());
     let core = prune(g);
     // Leftmost (then lowest) vertex still on a cycle, per component.
@@ -424,19 +432,32 @@ fn islands(g: &Graph, outer: &[usize], p: Vec2) -> Vec<Island> {
     let mut starts: Vec<(usize, usize)> = start.into_iter().collect();
     starts.sort_unstable();
     let mut out = Vec::new();
+    // Each containment test scans a whole polygon, accepted or not: they share one allowance of
+    // polygon edges, and the search ends when it is spent.
+    let mut edges_left: usize = 100_000_000;
     for (c, u) in starts {
         if out.len() >= MAX_ISLANDS {
             break;
         }
         let Some(&pu) = core.pts.get(u) else { continue };
-        if !ob.contains(pu) || !point_in_polygon(&outer_pts, pu) {
+        if !ob.contains(pu) {
+            continue;
+        }
+        let left = edges_left.checked_sub(outer_pts.len())?;
+        edges_left = left;
+        if !point_in_polygon(&outer_pts, pu) {
             continue;
         }
         let steps = edges.get(&c).copied().unwrap_or(0).saturating_add(4);
         let Some(lp) = outside(&core, u, steps) else { continue };
         let area = shoelace(&lp);
         let ib = Bounds2::from_points(lp.iter().copied());
-        if lp.len() < 3 || !area.is_finite() || area <= 0.0 || !ob.contains_box(&ib) || point_in_polygon(&lp, p) {
+        if lp.len() < 3 || !area.is_finite() || area <= 0.0 || !ob.contains_box(&ib) {
+            continue;
+        }
+        let left = edges_left.checked_sub(lp.len())?;
+        edges_left = left;
+        if point_in_polygon(&lp, p) {
             continue;
         }
         let mut src = sources.remove(&c).unwrap_or_default();
@@ -444,7 +465,7 @@ fn islands(g: &Graph, outer: &[usize], p: Vec2) -> Vec<Island> {
         src.dedup();
         out.push(Island { outline: lp, sources: src });
     }
-    out
+    Some(out)
 }
 
 /// The point of the nearest other curve within `snap` of the end `e` of curve `i`, as (curve,
@@ -457,6 +478,10 @@ fn nearest_curve(polys: &[Vec<Vec2>], i: usize, e: Vec2, snap: f64, tol: f64, bu
         if j == i {
             continue;
         }
+        // A curve with no segment costs a visit too, so many of them can't multiply the searches.
+        if other.len() < 2 {
+            *budget = budget.checked_sub(1).ok_or(())?;
+        }
         for (k, w) in other.windows(2).enumerate() {
             if *budget == 0 {
                 return Err(());
@@ -558,7 +583,7 @@ pub fn boundary_at_gap(polys: &[Vec<Vec2>], p: Vec2, tol: f64, gap: f64) -> Opti
     let g = build_gap(polys, tol, gap)?;
     let ids = outer_face(&g, p)?;
     let outer = points(&g, &ids)?;
-    let islands = islands(&g, &ids, p);
+    let islands = islands(&g, &ids, p)?;
     let mut outer_sources: Vec<usize> =
         ids.iter().zip(ids.iter().cycle().skip(1)).filter_map(|(&u, &v)| g.edge_src.get(&(u.min(v), u.max(v))).copied()).collect();
     outer_sources.sort_unstable();

crates/io/src/dxf_read/document.rs

@@ -333,7 +333,17 @@ pub(super) fn apply_sortents(
     paper_for: &HashMap<String, String>,
     first_layout: &str,
 ) {
-    for (owner, map) in tables {
+    // Each sort rebuilds its whole space, so a space named by many tables is sorted once, by the
+    // last of them (which decided its order anyway); repeated tables made import quadratic.
+    let target = |owner: &String| {
+        let up = brs.get(owner)?.to_ascii_uppercase();
+        Some(if up.starts_with("*PAPER_SPACE") { paper_for.get(&up).cloned().unwrap_or_else(|| first_layout.to_string()) + "\0paper" } else { up })
+    };
+    let last: HashMap<String, usize> = tables.iter().enumerate().filter_map(|(i, (owner, _))| Some((target(owner)?, i))).collect();
+    for (i, (owner, map)) in tables.iter().enumerate() {
+        if target(owner).and_then(|t| last.get(&t)) != Some(&i) {
+            continue;
+        }
         let Some(name) = brs.get(owner) else { continue };
         let up = name.to_ascii_uppercase();
         // A sort handle of 0 draws last (ezdxf `reorder`, which notes AutoCAD defines it, undocumented).

crates/io/src/dxf_read/entities.rs

@@ -299,7 +299,9 @@ pub(super) fn acad_table(tags: &[Tag], table_cells_left: &mut usize) -> Table {
     // Large declared sizes must be backed by cell records (no allocation from a bare count).
     let n_cells = tail.iter().filter(|x| x.code == 171).count();
     let declared = rows.saturating_mul(cols);
-    if declared > MAX_CELLS || (declared > 10_000 && n_cells < declared) {
+    // A table with no rows or no columns has no cells: keep neither dimension, so its row heights
+    // and column widths (allocated per dimension, outside the cell budget) are not allocated.
+    if declared == 0 || declared > MAX_CELLS || (declared > 10_000 && n_cells < declared) {
         rows = 0;
         cols = 0;
     }

crates/io/src/tests.rs

@@ -337,6 +337,42 @@ fn entity_transparency_roundtrips_as_group_440() {
     assert_eq!(d.model.iter().map(|e| e.common.transparency).collect::<Vec<_>>(), [Transparency::ByLayer, Transparency::Percent(90)]);
 }
 
+/// Several SORTENTSTABLEs for one space sort it once, in the last table's order.
+#[test]
+fn repeated_sortents_tables_use_the_last() {
+    let table =
+        |a: u32, b: u32| format!("0\nSORTENTSTABLE\n5\n{:X}\n100\nAcDbSortentsTable\n330\n1F\n331\nA\n5\n{a:X}\n331\nB\n5\n{b:X}\n", 0x100 + a);
+    let mut t = String::from(
+        "0\nSECTION\n2\nTABLES\n0\nTABLE\n2\nBLOCK_RECORD\n0\nBLOCK_RECORD\n5\n1F\n2\n*Model_Space\n0\nENDTAB\n0\nENDSEC\n\
+         0\nSECTION\n2\nENTITIES\n0\nLINE\n5\nA\n330\n1F\n10\n0\n20\n0\n11\n1\n21\n1\n0\nLINE\n5\nB\n330\n1F\n10\n0\n20\n0\n11\n2\n21\n2\n0\nENDSEC\n\
+         0\nSECTION\n2\nOBJECTS\n0\nDICTIONARY\n5\nC\n",
+    );
+    for _ in 0..1000 {
+        t.push_str(&table(2, 1));
+    }
+    t.push_str(&table(1, 2));
+    t.push_str("0\nENDSEC\n0\nEOF\n");
+    let d = read(t.as_bytes(), "s.dxf").unwrap();
+    assert_eq!(d.model.iter().map(|e| e.handle.0).collect::<Vec<_>>(), [0xA, 0xB]);
+    let reversed = t.replacen(&table(1, 2), &table(2, 1), 1);
+    let d = read(reversed.as_bytes(), "s.dxf").unwrap();
+    assert_eq!(d.model.iter().map(|e| e.handle.0).collect::<Vec<_>>(), [0xB, 0xA]);
+}
+
+/// A table declaring 10 000 rows and no columns (or the reverse) is read as an empty table.
+#[test]
+fn table_with_an_empty_dimension_allocates_nothing() {
+    for (rows, cols) in [(10_000, 0), (0, 10_000)] {
+        let t = format!("0\nSECTION\n2\nENTITIES\n0\nACAD_TABLE\n100\nAcDbTable\n91\n{rows}\n92\n{cols}\n0\nENDSEC\n0\nEOF\n");
+        let d = read(t.as_bytes(), "t.dxf").unwrap();
+        let tables: Vec<_> = d.model.iter().filter_map(|e| if let EntityKind::Table(t) = &e.kind { Some(t) } else { None }).collect();
+        assert_eq!(tables.len(), 1, "the table is read");
+        for t in tables {
+            assert!(t.row_heights.is_empty() && t.col_widths.is_empty() && t.cells.is_empty());
+        }
+    }
+}
+
 /// Many distinct layer, linetype and style names (in the tables and on entities) read in linear
 /// time, with the duplicate rules kept: first LTYPE wins, last STYLE wins, layers case-insensitive.
 #[test]