Merge pull request #1629 from ramox81/codex/mtext-rtl-layout

Text layout: right-to-left runs, TEXT direction and leading spaces.

src/entities/mtext.rs

@@ -83,6 +83,34 @@ fn drawing_dir_str(d: &DrawingDirection) -> &'static str {
 /// click-to-select preview. Uses the exact same layout opts as `to_render`
 /// so the boxes line up with the rendered glyphs.
 
+/// The text MTEXT lays out: its stored value, except that the spaces between
+/// a field's value and a right-to-left word right after it read with that
+/// word — the reference draws `שלום 0 עולם` with no gap after the field
+/// (the spaces end up after the word), while plain text keeps the gap.
+fn display_text(t: &MText, document: &codec::CadDocument) -> String {
+    use crate::entities::text_support::is_rtl_char;
+    let value = &t.value;
+    if !value.chars().any(is_rtl_char) {
+        return value.clone();
+    }
+    let Some(spans) = codec::fields::field_spans(document, t.common.handle, value) else {
+        return value.clone();
+    };
+    let mut out = value.clone();
+    for span in spans.iter().rev() {
+        let rest = &out[span.end..];
+        let gap = rest.len() - rest.trim_start_matches(' ').len();
+        let after = &rest[gap..];
+        if gap == 0 || !after.chars().next().is_some_and(is_rtl_char) {
+            continue;
+        }
+        let word = after.find(|c: char| c == ' ' || c == '\\').unwrap_or(after.len());
+        let moved = format!("{}{}{}", &after[..word], &rest[..gap], &after[word..]);
+        out.replace_range(span.end.., &moved);
+    }
+    out
+}
+
 pub fn glyph_boxes(t: &MText, document: &codec::CadDocument) -> Vec<GlyphBox> {
     let resolved_style = resolve_text_style(&t.style, document);
     let attach_h_anchor: f32 = match t.attachment_point {
@@ -112,7 +140,7 @@ pub fn glyph_boxes(t: &MText, document: &codec::CadDocument) -> Vec<GlyphBox> {
     };
     // A field shows its stored value: fields update on their evaluation
     // events (open, save, plot, regen, UPDATEFIELD), never on a redraw.
-    let display = t.value.clone();
+    let display = display_text(t, document);
     let layout = layout_mtext(&MTextRenderOpts {
         value: &display,
         insertion: [
@@ -170,7 +198,7 @@ fn to_render(t: &MText, document: &codec::CadDocument) -> RenderEntity {
     };
     // A field shows its stored value: fields update on their evaluation
     // events (open, save, plot, regen, UPDATEFIELD), never on a redraw.
-    let display = t.value.clone();
+    let display = display_text(t, document);
     let layout = layout_mtext(&MTextRenderOpts {
         value: &display,
         insertion: [

src/entities/text.rs

@@ -281,38 +281,13 @@ pub fn text_run_placement_at_scale(
         // box one way while the strokes run the other (the bounds advance is
         // always positive — it uses |width_factor|). Left keeps its 0 reference.
         let sign = width_factor.signum();
-        let is_rtl = {
-            let mut strong_rtl = false;
-            for ch in value_for_bounds.chars() {
-                match unicode_bidi::bidi_class(ch) {
-                    unicode_bidi::BidiClass::R | unicode_bidi::BidiClass::AL => {
-                        strong_rtl = true;
-                        break;
-                    }
-                    unicode_bidi::BidiClass::L => {
-                        break;
-                    }
-                    _ => {}
-                }
-            }
-            strong_rtl
-        };
+        // Left / right alignment do not depend on the text's script: the
+        // reference lays a line that starts with Hebrew or Arabic out from its
+        // left end like any other.
         let ax = match t.horizontal_alignment {
-            HA::Left => {
-                if is_rtl {
-                    b.advance * sign
-                } else {
-                    0.0
-                }
-            }
+            HA::Left => 0.0,
             HA::Center | HA::Middle => b.advance * 0.5 * sign,
-            HA::Right => {
-                if is_rtl {
-                    0.0
-                } else {
-                    b.advance * sign
-                }
-            }
+            HA::Right => b.advance * sign,
             HA::Aligned | HA::Fit => 0.0,
         };
         // Vertical anchor uses the inked extent (cap / baseline geometry).

src/entities/text_support.rs

@@ -867,6 +867,73 @@ pub fn is_rtl_char(c: char) -> bool {
 /// Reorder a line's atoms into visual order (left to right) using the Unicode
 /// Bidirectional Algorithm (UBA) Rule L2. If the line contains no RTL characters
 /// and is not an RTL paragraph, returns the atoms unchanged.
+/// Left-to-right mark: keeps the shaper's paragraph direction left to right
+/// and closes a right-to-left run inside a word.
+pub(crate) const LRM: char = '‎';
+
+/// Whether a word is written right to left as a whole: right-to-left
+/// letters (and their marks) only.
+fn is_rtl_word(w: &str) -> bool {
+    w.chars().any(is_rtl_char)
+        && w.chars().all(|c| {
+            matches!(
+                unicode_bidi::bidi_class(c),
+                unicode_bidi::BidiClass::R | unicode_bidi::BidiClass::AL | unicode_bidi::BidiClass::NSM
+            )
+        })
+}
+
+/// A word as the reference draws it inside MTEXT: each run of right-to-left
+/// letters reads backwards in place, everything else (digits, punctuation,
+/// Latin) keeps its order — `אב_123_גד` shows `בא_123_דג`. A mark after
+/// each right-to-left run keeps the shaper from carrying neutrals and
+/// numbers into it.
+pub(crate) fn rtl_display_word(word: String) -> String {
+    if !word.chars().any(is_rtl_char) || is_rtl_word(&word) {
+        return word;
+    }
+    let mut out = String::with_capacity(word.len() + 6);
+    let mut in_rtl = false;
+    for c in word.chars() {
+        let rtl = matches!(
+            unicode_bidi::bidi_class(c),
+            unicode_bidi::BidiClass::R | unicode_bidi::BidiClass::AL
+        ) || (in_rtl && unicode_bidi::bidi_class(c) == unicode_bidi::BidiClass::NSM);
+        if in_rtl && !rtl {
+            out.push(LRM);
+        }
+        out.push(c);
+        in_rtl = rtl;
+    }
+    out
+}
+
+/// The reference's MTEXT line order: words keep their order left to right,
+/// except that a run of right-to-left words — with the spaces between and
+/// after them — reads backwards as a whole (`אבג דה abc` shows
+/// ` הד גבאabc`). Each right-to-left word is itself shaped backwards.
+pub fn reorder_rtl_runs(atoms: Vec<LayoutAtom>) -> Vec<LayoutAtom> {
+    let rtl = |a: &LayoutAtom| matches!(&a.kind, AtomKind::Word(w) if is_rtl_word(w));
+    if !atoms.iter().any(rtl) {
+        return atoms;
+    }
+    let mut atoms = atoms;
+    let mut i = 0;
+    while i < atoms.len() {
+        if !rtl(&atoms[i]) {
+            i += 1;
+            continue;
+        }
+        let mut end = i + 1;
+        while end < atoms.len() && (rtl(&atoms[end]) || matches!(atoms[end].kind, AtomKind::Space)) {
+            end += 1;
+        }
+        atoms[i..end].reverse();
+        i = end;
+    }
+    atoms
+}
+
 pub fn reorder_line_atoms(atoms: Vec<LayoutAtom>, is_rtl: bool) -> Vec<LayoutAtom> {
     if atoms.len() <= 1 && !is_rtl {
         return atoms;
@@ -1005,6 +1072,69 @@ pub(crate) fn cjk_break_between(prev: char, next: char) -> bool {
     !CLOSING.contains(next) && !OPENING.contains(prev)
 }
 
+/// A word's font runs, in logical order: `true` for a stretch the shaper
+/// draws, `false` for one the stroke font draws. With a stroke font the
+/// reference draws a mixed word run by run — Latin, digits and punctuation
+/// in the style's font, the letters of a script that needs shaping (with the
+/// marks that close them) in a TrueType face — so `שלום|12` keeps a thin
+/// bar and digits. A TrueType style shapes the whole word.
+pub(crate) fn word_runs<'t>(text: &'t str, face: &Face) -> Vec<(&'t str, bool)> {
+    use crate::scene::text::web_font::requires_shaping;
+    if face.ttf_family().is_some() {
+        return vec![(text, true)];
+    }
+    if !requires_shaping(text) {
+        return vec![(text, false)];
+    }
+    let mut runs: Vec<(&str, bool)> = Vec::new();
+    let mut start = 0;
+    let mut shaped_run = None;
+    for (i, c) in text.char_indices() {
+        let mut buf = [0u8; 4];
+        let shaped = requires_shaping(c.encode_utf8(&mut buf))
+            || (shaped_run == Some(true)
+                && (c == LRM || unicode_bidi::bidi_class(c) == unicode_bidi::BidiClass::NSM));
+        match shaped_run {
+            Some(prev) if prev != shaped => {
+                runs.push((&text[start..i], prev));
+                start = i;
+            }
+            _ => {}
+        }
+        shaped_run = Some(shaped);
+    }
+    if let Some(shaped) = shaped_run {
+        runs.push((&text[start..], shaped));
+    }
+    runs
+}
+
+/// The shaper's face for a word run.
+fn run_family(face: &Face) -> &str {
+    face.ttf_family()
+        .unwrap_or_else(|| crate::scene::text::web_font::primary_script().family())
+}
+
+/// Advance of a stroke-font run (glyph units × `scale`).
+fn stroke_run_width(face: &Face, text: &str, tracking: f32, scale: f32) -> f32 {
+    text.chars()
+        .map(|ch| match face.glyph(ch) {
+            Some(g) => (g.advance + face.spacing_after(ch) * tracking) * scale,
+            None => (6.0 + face.letter_spacing() * tracking) * scale,
+        })
+        .sum()
+}
+
+/// Width of one word run, as it is drawn.
+pub(crate) fn run_width(face: &Face, text: &str, shaped: bool, tracking: f32, scale: f32) -> f32 {
+    if shaped {
+        if let Some(run) = crate::scene::text::ttf_glyph::shape_run(run_family(face), text) {
+            return run.advance * scale;
+        }
+    }
+    stroke_run_width(face, text, tracking, scale)
+}
+
 pub fn measure_word(
     text: &str,
     state: &RunState,
@@ -1015,23 +1145,10 @@ pub fn measure_word(
     let scale = run_scale(state, entity_h, base_wf);
     let font_name = resolve_font(state, base_font);
     let face = Face::resolve(&font_name);
-    let shaping_family = face.ttf_family().or_else(|| {
-        crate::scene::text::web_font::requires_shaping(text)
-            .then(|| crate::scene::text::web_font::primary_script().family())
-    });
-    if let Some(fam) = shaping_family {
-        if let Some(run) = crate::scene::text::ttf_glyph::shape_run(fam, text) {
-            return run.advance * scale;
-        }
-    }
-    let mut w = 0.0_f32;
-    for ch in text.chars() {
-        w += match face.glyph(ch) {
-            Some(g) => (g.advance + face.spacing_after(ch) * state.tracking) * scale,
-            None => (6.0 + face.letter_spacing() * state.tracking) * scale,
-        };
-    }
-    w
+    word_runs(text, &face)
+        .into_iter()
+        .map(|(run, shaped)| run_width(&face, run, shaped, state.tracking, scale))
+        .sum()
 }
 
 /// Per-character cells of a word laid out by the shaper — the visual
@@ -1049,17 +1166,54 @@ pub fn word_cells(
     let scale = run_scale(state, entity_h, base_wf);
     let font_name = resolve_font(state, base_font);
     let face = Face::resolve(&font_name);
-    let family = face.ttf_family().or_else(|| {
-        crate::scene::text::web_font::requires_shaping(text)
-            .then(|| crate::scene::text::web_font::primary_script().family())
-    })?;
-    let run = crate::scene::text::ttf_glyph::shape_run(family, text)?;
-    Some(
-        crate::scene::text::ttf_glyph::char_cells(&run, text)
-            .into_iter()
-            .map(|(x0, x1)| (x0 * scale, x1 * scale))
-            .collect(),
-    )
+    let runs = word_runs(text, &face);
+    if !runs.iter().any(|(_, shaped)| *shaped) {
+        return None;
+    }
+    let mut cells = Vec::new();
+    let mut x = 0.0_f32;
+    for (run, shaped) in runs {
+        let shaped_run = shaped
+            .then(|| crate::scene::text::ttf_glyph::shape_run(run_family(&face), run))
+            .flatten();
+        match shaped_run {
+            Some(shaped_run) => {
+                cells.extend(
+                    crate::scene::text::ttf_glyph::char_cells(&shaped_run, run)
+                        .into_iter()
+                        .zip(run.chars())
+                        .filter(|(_, c)| *c != LRM)
+                        .map(|((x0, x1), _)| (x + x0 * scale, x + x1 * scale)),
+                );
+                x += shaped_run.advance * scale;
+            }
+            None => {
+                for ch in run.chars().filter(|c| *c != LRM) {
+                    let w = stroke_run_width(&face, ch.encode_utf8(&mut [0u8; 4]), state.tracking, scale);
+                    cells.push((x, x + w));
+                    x += w;
+                }
+            }
+        }
+    }
+    Some(cells)
+}
+
+/// The letter spacing a stroke-font word leaves after its last glyph (0 for
+/// shaped words, whose advance carries no extra spacing).
+fn trailing_letter_gap(atom: &LayoutAtom, entity_h: f32, base_wf: f32, base_font: &str) -> f32 {
+    let AtomKind::Word(text) = &atom.kind else {
+        return 0.0;
+    };
+    let font_name = resolve_font(&atom.state, base_font);
+    let face = Face::resolve(&font_name);
+    let Some((text, false)) = word_runs(text, &face).last().copied() else {
+        return 0.0;
+    };
+    text.chars()
+        .rev()
+        .find(|c| *c != LRM)
+        .map_or(0.0, |c| face.spacing_after(c) * atom.state.tracking * run_scale(&atom.state, entity_h, base_wf))
 }
 
 pub fn measure_space(state: &RunState, entity_h: f32, base_wf: f32, base_font: &str) -> f32 {
@@ -1141,7 +1295,11 @@ pub fn wrap_paragraph(
             AtomKind::Word(_) | AtomKind::Stack { .. } => {
                 let w = atom_width(&atom, entity_h, base_wf, base_font);
                 let max_w = line_max_w(subline_idx);
-                if !cur.is_empty() && cur_w + w > max_w && !after_align_tab {
+                let has_content = cur.iter().any(|a| !matches!(a.kind, AtomKind::Space));
+                // A word fits when its ink does: the letter spacing after its
+                // last glyph may run past the edge.
+                let ink_w = w - trailing_letter_gap(&atom, entity_h, base_wf, base_font);
+                if has_content && cur_w + ink_w > max_w && !after_align_tab {
                     while matches!(cur.last().map(|a| &a.kind), Some(AtomKind::Space)) {
                         cur.pop();
                     }
@@ -1155,7 +1313,9 @@ pub fn wrap_paragraph(
             }
             AtomKind::Space => {
                 after_align_tab = false;
-                if cur.is_empty() {
+                // A wrapped line does not start with the spaces it broke at;
+                // the paragraph's own leading spaces keep their width.
+                if cur.is_empty() && subline_idx > 0 {
                     continue;
                 }
                 cur_w += atom_width(&atom, entity_h, base_wf, base_font);
@@ -1526,7 +1686,7 @@ pub fn layout_mtext(opts: &MTextRenderOpts) -> MTextLayout {
                         } else if ch == ' ' || ch == '\t' {
                             if !word.is_empty() {
                                 atoms.push(LayoutAtom {
-                                    kind: AtomKind::Word(std::mem::take(&mut word)),
+                                    kind: AtomKind::Word(rtl_display_word(std::mem::take(&mut word))),
                                     state: run.state.clone(),
                                     char_offset: word_start,
                                 });
@@ -1554,7 +1714,7 @@ pub fn layout_mtext(opts: &MTextRenderOpts) -> MTextLayout {
                             if let Some(prev) = word.chars().last() {
                                 if cjk_break_between(prev, ch) {
                                     atoms.push(LayoutAtom {
-                                        kind: AtomKind::Word(std::mem::take(&mut word)),
+                                        kind: AtomKind::Word(rtl_display_word(std::mem::take(&mut word))),
                                         state: run.state.clone(),
                                         char_offset: word_start,
                                     });
@@ -1567,7 +1727,7 @@ pub fn layout_mtext(opts: &MTextRenderOpts) -> MTextLayout {
                     }
                     if !word.is_empty() {
                         atoms.push(LayoutAtom {
-                            kind: AtomKind::Word(word),
+                            kind: AtomKind::Word(rtl_display_word(word)),
                             state: run.state.clone(),
                             char_offset: word_start,
                         });
@@ -1603,52 +1763,24 @@ pub fn layout_mtext(opts: &MTextRenderOpts) -> MTextLayout {
             }
         }
 
-        // Trim leading + trailing Space atoms so line_w / cursor_start agree
-        // on the paragraph's visible content. Without this a stray trailing
-        // space measures wider than it draws and centring / right-alignment
-        // is off by half a space-width.
+        // Trailing Space atoms draw nothing; the line width that centring /
+        // right-alignment use leaves them out (see `line_w` below), while
+        // leading spaces keep their width, as in the reference application.
         //
-        // Skipped when emitting glyph boxes (the MText editor) so a space the
+        // Kept when emitting glyph boxes (the MText editor) so a space the
         // user just typed at the end keeps a selectable box and the caret can
         // sit after it.
         if !opts.want_glyph_boxes {
-            let first_word = atoms
-                .iter()
-                .position(|a| !matches!(a.kind, AtomKind::Space))
-                .unwrap_or(atoms.len());
-            atoms.drain(..first_word);
             while matches!(atoms.last().map(|a| &a.kind), Some(AtomKind::Space)) {
                 atoms.pop();
             }
         }
 
-        let is_rtl_para = {
-            let mut strong_rtl = None;
-            for atom in &atoms {
-                match &atom.kind {
-                    AtomKind::Word(w) => {
-                        for ch in w.chars() {
-                            match unicode_bidi::bidi_class(ch) {
-                                unicode_bidi::BidiClass::R | unicode_bidi::BidiClass::AL => {
-                                    strong_rtl = Some(true);
-                                    break;
-                                }
-                                unicode_bidi::BidiClass::L => {
-                                    strong_rtl = Some(false);
-                                    break;
-                                }
-                                _ => {}
-                            }
-                        }
-                        if strong_rtl.is_some() {
-                            break;
-                        }
-                    }
-                    _ => {}
-                }
-            }
-            strong_rtl.unwrap_or(false)
-        };
+        // The reference lays every MTEXT paragraph out left to right, also
+        // one that starts with Hebrew or Arabic: alignment follows the
+        // attachment, and only right-to-left runs read backwards in place
+        // (`reorder_rtl_runs`).
+        let is_rtl_para = false;
 
         // Wrap to the column the text actually flows down, not to the block:
         // measuring against the full width would let a line run across the
@@ -1678,7 +1810,7 @@ pub fn layout_mtext(opts: &MTextRenderOpts) -> MTextLayout {
             let atoms = if opts.vertical_text {
                 atoms
             } else {
-                reorder_line_atoms(atoms, is_rtl_para)
+                reorder_rtl_runs(atoms)
             };
             sub_lines.push(SubLine {
                 atoms,
@@ -2123,15 +2255,26 @@ pub fn layout_mtext(opts: &MTextRenderOpts) -> MTextLayout {
             }
         };
 
+        let visible_atoms = sub
+            .atoms
+            .iter()
+            .rposition(|a| !matches!(a.kind, AtomKind::Space))
+            .map_or(0, |i| i + 1);
         let line_w = line_total_width(
-            &sub.atoms,
+            &sub.atoms[..visible_atoms],
             entity_h,
             base_wf,
             &base_font_name,
             0.0,
             sub.indent_left,
             &sub.tab_stops,
-        );
+        )
+        // Aligned on the last glyph's ink: the letter spacing after it is not
+        // part of the width centring / right alignment use (the reference
+        // puts a right-aligned line's last stroke on the edge).
+        - sub.atoms[..visible_atoms]
+            .last()
+            .map_or(0.0, |atom| trailing_letter_gap(atom, entity_h, base_wf, &base_font_name));
         line_widths.push(line_w);
 
         let cursor_start = if rect_w > 0.0 {
@@ -2324,41 +2467,60 @@ pub fn layout_mtext(opts: &MTextRenderOpts) -> MTextLayout {
                         ins_y + (line_base_y + world_dy) as f64,
                     ];
                     // Glyphs from the plain word — the SDF run and the stroke
-                    // fallback both draw clean characters.
-                    let (strokes, fill_tris) = lff::tessellate_text_run(
-                        [0.0, 0.0],
-                        run_h,
-                        rot,
-                        signed_wf,
-                        oblique,
-                        tracking,
-                        &font_name,
-                        text,
-                    );
-                    let glyph_n = strokes.len();
-                    all_strokes.push(TextStroke {
-                        strokes,
-                        origin,
-                        color,
-                        fill_tris,
-                        plane: None,
-                        run: Some(GlyphRun {
-                            text: text.clone(),
-                            font: font_name.to_string(),
-                            height: run_h,
-                            rotation: rot,
-                            width_factor: signed_wf,
+                    // fallback both draw clean characters. A mixed word draws
+                    // run by run (see `word_runs`), each at its own pen.
+                    let face = Face::resolve(&font_name);
+                    let run_scale_x = run_h / 9.0 * signed_wf.abs();
+                    let mut run_x = 0.0_f32;
+                    for (run_text, shaped) in word_runs(text, &face) {
+                        let (strokes, fill_tris) = lff::tessellate_text_run(
+                            [0.0, 0.0],
+                            run_h,
+                            rot,
+                            signed_wf,
                             oblique,
                             tracking,
-                            bold: atom.state.bold,
-                        }),
-                    });
+                            &font_name,
+                            run_text,
+                        );
+                        let (dx, dy) = (run_x * cos_r, run_x * sin_r);
+                        all_strokes.push(TextStroke {
+                            strokes,
+                            origin: [origin[0] + dx as f64, origin[1] + dy as f64],
+                            color,
+                            fill_tris,
+                            plane: None,
+                            run: Some(GlyphRun {
+                                text: run_text.to_string(),
+                                font: font_name.to_string(),
+                                height: run_h,
+                                rotation: rot,
+                                width_factor: signed_wf,
+                                oblique,
+                                tracking,
+                                bold: atom.state.bold,
+                            }),
+                        });
+                        run_x += run_width(&face, run_text, shaped, tracking, run_scale_x);
+                    }
                     // Underline / overline / strike are lines, not glyphs; a
                     // run-group's strokes are suppressed by the SDF path, so
                     // emit the decorations in their own RUN-LESS group. Reuse
                     // lff's exact positions by tessellating the decorated word
                     // and taking the strokes it appends after the glyphs.
                     if atom.state.underline || atom.state.overline || atom.state.strike {
+                        let glyph_n = lff::tessellate_text_run(
+                            [0.0, 0.0],
+                            run_h,
+                            rot,
+                            signed_wf,
+                            oblique,
+                            tracking,
+                            &font_name,
+                            text,
+                        )
+                        .0
+                        .len();
                         let body = decorated(text, &atom.state);
                         let (deco, _) = lff::tessellate_text_run(
                             [0.0, 0.0],
@@ -2383,7 +2545,7 @@ pub fn layout_mtext(opts: &MTextRenderOpts) -> MTextLayout {
                     }
                     if opts.want_glyph_boxes {
                         let run_h = atom.state.height_mul * entity_h;
-                        let count = text.chars().count();
+                        let count = text.chars().filter(|c| *c != LRM).count();
                         let is_rtl = text.chars().any(is_rtl_char);
                         if is_rtl && count > 0 {
                             let word_w = if tracking != atom.state.tracking {
@@ -2429,7 +2591,7 @@ pub fn layout_mtext(opts: &MTextRenderOpts) -> MTextLayout {
                             let face = Face::resolve(&font_name);
                             let shaped = word_cells(text, &atom.state, entity_h, base_wf, &base_font_name);
                             let mut cx = cursor_x;
-                            for (ci, ch) in text.chars().enumerate() {
+                            for (ci, ch) in text.chars().filter(|c| *c != LRM).enumerate() {
                                 let adv = match face.glyph(ch) {
                                     Some(g) => {
                                         (g.advance + face.spacing_after(ch) * tracking) * scale

src/scene/text/ttf_glyph.rs

@@ -603,6 +603,9 @@ pub fn char_cells(run: &ShapedRun, text: &str) -> Vec<(f32, f32)> {
         .collect()
 }
 
+/// Left-to-right mark shaped ahead of every run.
+const BASE_MARK: char = '‎';
+
 type ShapeCache = HashMap<(String, String), Option<Arc<ShapedRun>>>;
 
 fn shape_cache() -> &'static Mutex<ShapeCache> {
@@ -901,7 +904,11 @@ fn build_shaped(_family: &str, text: &str) -> Option<ShapedRun> {
     let attrs = Attrs::new().family(Family::SansSerif);
     let mut buffer = Buffer::new(&mut font_system, Metrics::new(SHAPE_FS, SHAPE_FS));
     buffer.set_size(&mut font_system, None, None);
-    buffer.set_text(&mut font_system, text, &attrs, Shaping::Advanced, None);
+    // A leading left-to-right mark keeps the paragraph direction left to
+    // right, as the reference lays text out (right-to-left runs still read
+    // backwards); `cells` stay relative to `text`.
+    let marked = format!("{BASE_MARK}{text}");
+    buffer.set_text(&mut font_system, &marked, &attrs, Shaping::Advanced, None);
     buffer.shape_until_scroll(&mut font_system, false);
 
     let mut glyphs = Vec::new();
@@ -910,7 +917,10 @@ fn build_shaped(_family: &str, text: &str) -> Option<ShapedRun> {
     for run in buffer.layout_runs() {
         advance = advance.max(run.line_w * px_to_9);
         for glyph in run.glyphs.iter() {
-            cells.push((glyph.start, glyph.end, glyph.x * px_to_9, (glyph.x + glyph.w) * px_to_9));
+            if glyph.start >= BASE_MARK.len_utf8() {
+                let m = BASE_MARK.len_utf8();
+                cells.push((glyph.start - m, glyph.end - m, glyph.x * px_to_9, (glyph.x + glyph.w) * px_to_9));
+            }
             let face_index = font_system
                 .db_mut()
                 .face(glyph.font_id)
@@ -978,7 +988,11 @@ fn build_shaped(family: &str, text: &str) -> Option<ShapedRun> {
     let mut buf = Buffer::new(&mut fs, Metrics::new(SHAPE_FS, SHAPE_FS));
     // No wrapping: a run is a single line.
     buf.set_size(&mut fs, None, None);
-    buf.set_text(&mut fs, text, &attrs, Shaping::Advanced, None);
+    // A leading left-to-right mark keeps the paragraph direction left to
+    // right, as the reference lays text out (right-to-left runs still read
+    // backwards); `cells` stay relative to `text`.
+    let marked = format!("{BASE_MARK}{text}");
+    buf.set_text(&mut fs, &marked, &attrs, Shaping::Advanced, None);
     buf.shape_until_scroll(&mut fs, false);
 
     if primary_metrics.is_none() {
@@ -1017,7 +1031,10 @@ fn build_shaped(family: &str, text: &str) -> Option<ShapedRun> {
     for run in buf.layout_runs() {
         advance = advance.max(run.line_w * px_to_9);
         for g in run.glyphs.iter() {
-            cells.push((g.start, g.end, g.x * px_to_9, (g.x + g.w) * px_to_9));
+            if g.start >= BASE_MARK.len_utf8() {
+                let m = BASE_MARK.len_utf8();
+                cells.push((g.start - m, g.end - m, g.x * px_to_9, (g.x + g.w) * px_to_9));
+            }
             let face_index = fs.db_mut().face(g.font_id).map(|f| f.index).unwrap_or(0);
             let Some(font) = fs.get_font(g.font_id, g.font_weight) else {
                 continue;