documents: drawn text reads back from its end, decoded as written

Each piece records how its source decodes: Markdown's escapes and
references for text, HTML's for an HTML block's runs, none for code
spans, cut as CommonMark cuts their end spaces. Matching runs from the
piece's end, so quote marks and indents at a line's start are never
taken for text, and an alert's text needs no moved position: alert()
takes no source again.

Lines split on LF, CRLF and CR alike; LineMap matches each drawn line
to the written line it ends once, so an HTML block maps in linear time
and a tab cut to spaces lands on the tab. A write-back refuses the
press when the pressed block's kind, text or drawing changed.

The hostile HTML test measures growth: a fixed 2 s budget failed under
the full suite's load on the old code too.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

src/documents/document/point.rs

@@ -3,11 +3,10 @@ use std::rc::Rc;
 use gpui::{Context, MouseButton, Window};
 use markdown::mdast;
 
-use super::{DocumentEvent, MarkdownDocument, edit::Landing, edit::moved};
+use super::{DocumentEvent, MarkdownDocument, edit::Landing, edit::moved, parse::Block};
 use crate::documents::{
     render::{OnPoint, Place, Pointed},
-    tree::lines_at,
-    tree::{alert, is_toc},
+    tree::{Node, alert, is_toc, lines_at},
 };
 
 /// A top-level block's kind.
@@ -36,7 +35,7 @@ impl TargetBlock {
             mdast::Node::Paragraph(_) if is_toc(node, source) => Self::Toc,
             mdast::Node::Paragraph(_) => Self::Paragraph,
             mdast::Node::Heading(heading) => Self::Heading(heading.depth),
-            mdast::Node::Blockquote(quote) if alert(quote, source).is_some() => Self::Alert,
+            mdast::Node::Blockquote(quote) if alert(quote).is_some() => Self::Alert,
             mdast::Node::Blockquote(_) => Self::Quote,
             mdast::Node::List(_) => Self::List,
             mdast::Node::Code(_) => Self::Code,
@@ -95,6 +94,12 @@ impl MarkdownDocument {
         })
     }
 
+    /// A block's kind, text and what it draws.
+    fn drawn(&self, block: &Block) -> (TargetBlock, String, Option<Vec<Node>>) {
+        let text = self.parsed.text[block.range.clone()].to_string();
+        (block.kind, text, block.nodes.clone())
+    }
+
     /// Caret to `at`, else block `ix`; the owner hears.
     pub(super) fn point(
         &mut self,
@@ -109,7 +114,7 @@ impl MarkdownDocument {
             block => block.flatten(),
         };
         let kind = block.map(|block| block.kind);
-        let pressed = block.map(|block| self.parsed.text[block.range.clone()].to_string());
+        let pressed = block.map(|block| self.drawn(block));
         let at = at.or(block.map(|block| block.range.start));
         let written = self.commit(cx);
         if written.is_some() {
@@ -128,7 +133,7 @@ impl MarkdownDocument {
         // A write-back may change the pressed block.
         let now = (self.parsed.blocks.iter())
             .find(|block| block.range.start <= at && at <= block.range.end)
-            .map(|block| self.parsed.text[block.range.clone()].to_string());
+            .map(|block| self.drawn(block));
         if written.is_some() && pressed.is_some() && now != pressed {
             return log::warn!("markdown document: the write-back changed the pressed block");
         }

src/documents/document/tests/mod.rs

@@ -371,6 +371,7 @@ mod links;
 mod look;
 mod place;
 mod point;
+mod reading;
 mod toc;
 mod tools;
 

src/documents/document/tests/point.rs

@@ -6,11 +6,7 @@ use gpui::{Entity, Modifiers, MouseButton, Pixels, TestAppContext, VisualTestCon
 
 use super::{open, type_in};
 use crate::{
-    documents::{
-        DocumentEvent, DocumentTarget, MarkdownDocument, MarkdownMode, TargetBlock,
-        document::parse::parsed,
-        tree::{Inline, Node},
-    },
+    documents::{DocumentEvent, DocumentTarget, MarkdownDocument, MarkdownMode, TargetBlock},
     theme::{ActiveTheme, TextSize},
 };
 
@@ -191,42 +187,6 @@ fn press_lines(k: usize, line: usize, button: MouseButton, cx: &mut VisualTestCo
     press_at(x, y, button, cx);
 }
 
-/// The source byte under drawn `shown` in block 0.
-fn read_back(source: &str, shown: &str) -> Option<usize> {
-    fn find<'a>(nodes: &'a [Node], shown: &str) -> Option<&'a Inline> {
-        nodes.iter().find_map(|node| match node {
-            Node::Paragraph(inline) | Node::Heading(_, inline) if inline.text.contains(shown) => {
-                Some(inline)
-            }
-            Node::Quote(body) | Node::Alert { body, .. } | Node::Footnote { body, .. } => {
-                find(body, shown)
-            }
-            _ => None,
-        })
-    }
-    let parsed = parsed(source);
-    let block = &parsed.blocks[0];
-    let inline = find(block.nodes.as_ref().expect("drawn"), shown).expect("shown");
-    let at = inline.text.find(shown).expect("shown");
-    let written = inline.source_at(&source[block.range.start..], at)?;
-    Some(block.range.start + written)
-}
-
-#[test]
-fn quote_marks_escapes_trims_and_alert_markers_read_back() {
-    let cases = [
-        ("> alpha\n> beta", "eta"),
-        ("> [!NOTE]\n> alpha beta", "beta"),
-        ("a \\*star\\* z", "*star"),
-        ("a ` xy ` z", "xy"),
-        ("<div>alpha beta</div>\n", "eta"),
-        ("> <div>\n> alpha beta</div>\n", "beta"),
-    ];
-    for (source, shown) in cases {
-        assert_eq!(read_back(source, shown), source.find(shown), "{source:?}");
-    }
-}
-
 #[gpui::test]
 fn a_secondary_press_in_a_quote_lands_past_its_mark(cx: &mut TestAppContext) {
     let source = "> alpha\n> beta";
@@ -328,3 +288,104 @@ fn a_write_back_that_changes_the_pressed_block_points_nowhere(cx: &mut TestAppCo
     );
     assert_ne!(caret(&document, cx), written.find("eta").unwrap());
 }
+
+#[gpui::test]
+fn crlf_lines_press_where_they_were_written(cx: &mut TestAppContext) {
+    let source = "```
+alpha
+beta
+```
+";
+    let (document, cx) = open(source, MarkdownMode::Visual, cx);
+    let events = heard(&document, cx);
+    press_lines(2, 0, MouseButton::Right, cx);
+    let at = source.find("pha").unwrap();
+    let expected = DocumentTarget {
+        in_code: true,
+        ..target(at, TargetBlock::Code)
+    };
+    assert_eq!(*events.borrow(), [expected]);
+    press_lines(2, 0, MouseButton::Left, cx);
+    assert_eq!(opened_caret(&document, cx), at);
+}
+
+#[gpui::test]
+fn crlf_front_matter_presses_where_written(cx: &mut TestAppContext) {
+    let source = "---
+title: value
+author: person
+---
+";
+    let (document, cx) = open(source, MarkdownMode::Visual, cx);
+    let events = heard(&document, cx);
+    press_lines(7, 0, MouseButton::Right, cx);
+    let at = source.find("value").unwrap();
+    assert_eq!(*events.borrow(), [target(at, TargetBlock::FrontMatter)]);
+    press_lines(7, 0, MouseButton::Left, cx);
+    assert_eq!(opened_caret(&document, cx), at);
+}
+
+#[gpui::test]
+fn a_crlf_definition_presses_where_written(cx: &mut TestAppContext) {
+    let source = "[ref]: https://ely.dev
+  \"title\"
+";
+    let (document, cx) = open(source, MarkdownMode::Visual, cx);
+    let events = heard(&document, cx);
+    press_on(0, 10, MouseButton::Right, cx);
+    assert_eq!(*events.borrow(), [target(10, TargetBlock::Definition)]);
+}
+
+#[gpui::test]
+fn a_changed_definition_points_nowhere(cx: &mut TestAppContext) {
+    let source = "[ref]: https://old.dev
+
+Go [there][ref]
+";
+    let (document, cx) = open(source, MarkdownMode::Visual, cx);
+    let events = heard(&document, cx);
+    type_in(&document, Some(0), Some("[ref]: https://new.dev"), cx);
+    press_on(1, 4, MouseButton::Right, cx);
+    assert_eq!(
+        super::text(&document, cx),
+        source.replace("old.dev", "new.dev")
+    );
+    assert!(
+        events.borrow().is_empty(),
+        "the link changed under the press"
+    );
+}
+
+#[gpui::test]
+fn a_quoted_greater_than_presses_on_itself(cx: &mut TestAppContext) {
+    let source = "> <div>
+> >alpha beta</div>";
+    let (document, cx) = open(source, MarkdownMode::Visual, cx);
+    let events = heard(&document, cx);
+    let bounds = cx.debug_bounds("markdown-block-0").expect("drawn");
+    let (body, _) = sizes(cx);
+    let x = bounds.left() + px(31.0) + body * 0.6 * 0.25;
+    press_at(x, bounds.top() + px(10.0), MouseButton::Right, cx);
+    let at = source.find(">alpha").unwrap();
+    assert_eq!(*events.borrow(), [target(at, TargetBlock::Quote)]);
+}
+
+#[gpui::test]
+fn an_escaped_backslash_presses_on_itself(cx: &mut TestAppContext) {
+    let source = r"a \\ z";
+    let (document, cx) = open(source, MarkdownMode::Visual, cx);
+    let events = heard(&document, cx);
+    press_on(0, 2, MouseButton::Right, cx);
+    let at = source.find(r"\ z").unwrap();
+    assert_eq!(*events.borrow(), [target(at, TargetBlock::Paragraph)]);
+}
+
+#[gpui::test]
+fn an_html_space_moved_by_a_tag_presses_on_its_word(cx: &mut TestAppContext) {
+    let source = "<div>alpha <b> beta</b></div>";
+    let (document, cx) = open(source, MarkdownMode::Visual, cx);
+    let events = heard(&document, cx);
+    press_on(0, 8, MouseButton::Right, cx);
+    let at = source.find("eta").unwrap();
+    assert_eq!(*events.borrow(), [target(at, TargetBlock::Html)]);
+}

src/documents/document/tests/reading.rs

@@ -0,0 +1,123 @@
+//! Drawn text reads back to its source byte (R4.5).
+
+use crate::documents::{
+    TargetBlock,
+    document::parse::parsed,
+    tree::{Inline, Node, lines_at},
+};
+
+/// The source byte under drawn `shown` in block 0.
+fn read_back(source: &str, shown: &str) -> Option<usize> {
+    fn find<'a>(nodes: &'a [Node], shown: &str) -> Option<&'a Inline> {
+        nodes.iter().find_map(|node| match node {
+            Node::Paragraph(inline) | Node::Heading(_, inline) if inline.text.contains(shown) => {
+                Some(inline)
+            }
+            Node::Quote(body) | Node::Alert { body, .. } | Node::Footnote { body, .. } => {
+                find(body, shown)
+            }
+            _ => None,
+        })
+    }
+    let parsed = parsed(source);
+    let block = &parsed.blocks[0];
+    let inline = find(block.nodes.as_ref().expect("drawn"), shown).expect("shown");
+    let at = inline.text.find(shown).expect("shown");
+    let written = inline.source_at(&source[block.range.start..], at)?;
+    Some(block.range.start + written)
+}
+
+/// The source byte under drawn `word` of block 0's code.
+fn code_back(source: &str, word: &str) -> Option<usize> {
+    fn code(nodes: &[Node]) -> (&str, usize) {
+        match &nodes[0] {
+            Node::Code { text, from, .. } => (text, *from),
+            Node::Quote(body) => code(body),
+            other => panic!("not code: {other:?}"),
+        }
+    }
+    let parsed = parsed(source);
+    let block = &parsed.blocks[0];
+    let (drawn, from) = code(block.nodes.as_ref().expect("drawn"));
+    let at = drawn.find(word).expect("drawn");
+    let written = lines_at(&source[block.range.start..], from, drawn, at)?;
+    Some(block.range.start + written)
+}
+
+#[test]
+fn text_reads_back_past_marks_escapes_trims_and_markers() {
+    let cases = [
+        ("> alpha\n> beta", "eta"),
+        ("> [!NOTE]\n> alpha beta", "beta"),
+        ("> [!NOTE]  \n> alpha beta", "beta"),
+        ("> [!NOTE]\\\n> alpha beta", "beta"),
+        ("> [!NOTE]\n> \\*beta\\*", "*beta"),
+        ("> [!NOTE]&#10;alpha beta", "beta"),
+        ("a \\*star\\* z", "*star"),
+        ("a \\\\ z", "\\ z"),
+        ("a ` xy ` z", "xy"),
+        ("a `  xy  ` z", " xy "),
+        ("a ` >xy ` z", ">xy"),
+        ("a `\\\\xy` z", "xy"),
+        ("<div>alpha beta</div>\n", "eta"),
+        ("> <div>\n> alpha beta</div>\n", "beta"),
+        ("> <div>\n> >alpha beta</div>", ">alpha"),
+        ("> <div>\n> >alpha beta</div>", "beta"),
+        ("<div>alpha <b>beta</b> gamma</div>", "gamma"),
+        ("<div>alpha <b>beta</b> gamma</div>", "beta"),
+        ("<div>alpha <b> beta</b></div>", "eta"),
+        ("<div>one</div>\n<div>alpha beta</div>", "beta"),
+    ];
+    for (source, shown) in cases {
+        assert_eq!(read_back(source, shown), source.find(shown), "{source:?}");
+    }
+}
+
+#[test]
+fn a_reference_reads_back_to_its_ampersand() {
+    let source = "a &amp; b";
+    assert_eq!(read_back(source, "& b"), source.find('&'));
+    assert_eq!(read_back(source, "b"), source.find('b'));
+}
+
+#[test]
+fn text_after_an_emoji_reads_back_and_before_it_lands_at_its_start() {
+    let source = "ab :smile: y";
+    assert_eq!(read_back(source, "y"), source.find('y'));
+    assert_eq!(read_back(source, "b"), Some(0));
+}
+
+#[test]
+fn code_reads_back_line_for_line_whatever_its_line_ends() {
+    let cases = [
+        "```\nalpha\nbeta\n```\n",
+        "```\r\nalpha\r\nbeta\r\n```\r\n",
+        "```\ralpha\rbeta\r```\r",
+        "    alpha\n    beta\n",
+        "> ```\n> alpha\n> beta\n> ```\n",
+    ];
+    for source in cases {
+        for word in ["pha", "eta"] {
+            assert_eq!(code_back(source, word), source.find(word), "{source:?}");
+        }
+    }
+}
+
+#[test]
+fn code_cut_from_a_tab_reads_back_past_it() {
+    for spaces in 0..=3 {
+        for prefix in ["\t", " \t", "  \t", "   \t", "\t\t"] {
+            let indent = " ".repeat(spaces);
+            let source = format!("{indent}```\n{indent}alpha\n{prefix}beta\n{indent}```\n");
+            assert_eq!(code_back(&source, "eta"), source.find("eta"), "{source:?}");
+        }
+    }
+}
+
+#[test]
+fn a_decoded_line_end_reads_without_a_written_one() {
+    let quote = parsed("> alpha&#10;beta");
+    assert_eq!(quote.blocks[0].kind, TargetBlock::Quote);
+    let alert = parsed("> [!NOTE]&#10;alpha beta");
+    assert_eq!(alert.blocks[0].kind, TargetBlock::Alert);
+}

src/documents/tree/blocks.rs

@@ -69,7 +69,7 @@ impl Reader<'_> {
         let mdast::Node::Blockquote(quote) = node else {
             return None;
         };
-        let (kind, parts) = alert(quote, self.source)?;
+        let (kind, parts) = alert(quote)?;
         let mut body = Vec::new();
         self.blocks(&parts, &mut body);
         Some(Node::Alert { kind, body })
@@ -77,11 +77,11 @@ impl Reader<'_> {
 }
 
 /// A quote's alert kind and its blocks after the marker.
-pub fn alert(quote: &mdast::Blockquote, source: &str) -> Option<(AlertKind, Vec<mdast::Node>)> {
+pub fn alert(quote: &mdast::Blockquote) -> Option<(AlertKind, Vec<mdast::Node>)> {
     let [mdast::Node::Paragraph(paragraph), rest @ ..] = &quote.children[..] else {
         return None;
     };
-    let (kind, after) = marker(&paragraph.children, source)?;
+    let (kind, after) = marker(&paragraph.children)?;
     if after.is_empty() && rest.is_empty() {
         return None;
     }
@@ -101,7 +101,7 @@ pub fn is_toc(node: &mdast::Node, source: &str) -> bool {
 }
 
 /// A plain first line naming a kind, and what follows.
-fn marker(children: &[mdast::Node], source: &str) -> Option<(AlertKind, Vec<mdast::Node>)> {
+fn marker(children: &[mdast::Node]) -> Option<(AlertKind, Vec<mdast::Node>)> {
     let mut line = String::new();
     for (at, child) in children.iter().enumerate() {
         let head = match child {
@@ -116,7 +116,6 @@ fn marker(children: &[mdast::Node], source: &str) -> Option<(AlertKind, Vec<mdas
                     let after = after.strip_prefix("\r\n").unwrap_or(&after[1..]);
                     let mut kept = text.clone();
                     kept.value = after.to_string();
-                    past_line(&mut kept, source);
                     (!after.is_empty()).then_some(mdast::Node::Text(kept))
                 }
             },
@@ -129,25 +128,6 @@ fn marker(children: &[mdast::Node], source: &str) -> Option<(AlertKind, Vec<mdas
     Some((AlertKind::marked(&line)?, Vec::new()))
 }
 
-/// `text` starts past its first line, as its value does.
-fn past_line(text: &mut mdast::Text, source: &str) {
-    let start = &mut text
-        .position
-        .as_mut()
-        .expect("parsed text has a place")
-        .start;
-    let written = &source[start.offset..];
-    let end = written.find(['\n', '\r']).expect("its line end is written");
-    let width = if written[end..].starts_with("\r\n") {
-        2
-    } else {
-        1
-    };
-    start.offset += end + width;
-    start.line += 1;
-    start.column = 1;
-}
-
 #[cfg(test)]
 mod tests {
     use super::*;

src/documents/tree/html.rs

@@ -300,7 +300,7 @@ fn decoded(text: &str, attribute: bool) -> String {
 const LONGEST: usize = 32;
 
 /// A reference at `text`'s start, as HTML reads it.
-fn reference(text: &str, attribute: bool) -> Option<(String, usize)> {
+pub(super) fn reference(text: &str, attribute: bool) -> Option<(String, usize)> {
     let body = text.strip_prefix('&')?;
     if let Some(number) = body.strip_prefix('#') {
         let (digits, radix) = match number.strip_prefix(['x', 'X']) {

src/documents/tree/html_block.rs

@@ -1,15 +1,24 @@
 use std::ops::Range;
 
-use super::{Inline, Run, html, sources::Follow};
+use super::{Escapes, Inline, Run, html, sources::LineMap};
 
 /// An HTML block's text, as a browser shows it.
 pub(super) fn html_block_text(written: &str, source: &str, span: Range<usize>) -> Inline {
-    let (mut run, mut follow) = (Run::default(), Follow::new(written, source, span.clone()));
+    let (mut run, map) = (Run::default(), LineMap::new(source, span.start, written));
     let mut rest = written;
     // Each text run keeps where it was written.
-    let mut words = |run: &mut Run, rest: &str, text: &str| {
-        if let Some(at) = follow.to(written.len() - rest.len()) {
-            run.text.sources.push((run.text.text.len(), at..span.end));
+    let words = |run: &mut Run, rest: &str, text: &str| {
+        let from = written.len() - rest.len();
+        let at = |value| map.as_ref()?.at(value);
+        // Blank text draws nothing where it stands.
+        if text.chars().all(html::spacing) {
+            return run.words(text);
+        }
+        match at(from).zip(at(from + text.len())) {
+            Some((start, end)) => {
+                (run.text.sources).push((run.text.text.len(), start..end, Escapes::Html))
+            }
+            None => log::warn!("markdown: an HTML block's text lost its source"),
         }
         run.words(text);
     };
@@ -278,23 +287,32 @@ mod tests {
 
     #[test]
     fn hostile_html_reads_in_linear_time() {
-        let started = web_time::Instant::now();
-        for source in [
-            "<".repeat(1 << 17),
-            "<a".repeat(1 << 16),
-            "<!--".repeat(1 << 16),
-            "<b>".repeat(1 << 16),
-            "<b>".repeat(1 << 15) + &"</i>".repeat(1 << 15),
-            "<script></script>".repeat(1 << 13),
-            "<div>a ".repeat(1 << 16),
-            "<b> ".repeat(1 << 16),
-            "<li><b>x ".repeat(1 << 15),
-            "<p><i>x ".repeat(1 << 15),
-        ] {
-            let block = format!("<div>\n{source}");
-            parse(&block);
-        }
-        let spent = started.elapsed();
-        assert!(spent.as_secs() < 2, "took {spent:?}");
+        let sources = |n: usize| {
+            [
+                "<".repeat(2 * n),
+                "<a".repeat(n),
+                "<!--".repeat(n),
+                "<b>".repeat(n),
+                "<b>".repeat(n / 2) + &"</i>".repeat(n / 2),
+                "<script></script>".repeat(n / 8),
+                "<div>a ".repeat(n),
+                "<b> ".repeat(n),
+                "<li><b>x ".repeat(n / 2),
+                "<p><i>x ".repeat(n / 2),
+            ]
+        };
+        let time = |n| {
+            let started = web_time::Instant::now();
+            for source in sources(n) {
+                parse(&format!(
+                    "<div>
+{source}"
+                ));
+            }
+            started.elapsed()
+        };
+        let (small, large) = (time(1 << 14), time(1 << 16));
+        // Four times the input, at most eight the time.
+        assert!(large < small * 8, "{small:?} then {large:?}");
     }
 }

src/documents/tree/mod.rs

@@ -13,6 +13,7 @@ mod stack;
 
 pub use blocks::{AlertKind, DocumentHeading, alert, is_toc};
 use html_block::html_block_text;
+pub use sources::Escapes;
 pub(crate) use sources::lines_at;
 use stack::close;
 
@@ -38,7 +39,7 @@ pub struct Inline {
     pub styles: Vec<(Range<usize>, Style)>,
     pub links: Vec<(Range<usize>, String)>,
     /// Where each piece was written: drawn start, source span.
-    pub sources: Vec<(usize, Range<usize>)>,
+    pub sources: Vec<(usize, Range<usize>, Escapes)>,
 }
 
 /// A block of a markdown document.
@@ -342,8 +343,9 @@ impl<'a> Reader<'a> {
         if !matches!(node, mdast::Node::Text(_) | mdast::Node::Html(_)) {
             run.shortcodes.split();
         }
-        if let Some(written) = sources::written(node, self.source) {
-            run.text.sources.push((run.text.text.len(), written));
+        if let Some((written, escapes)) = sources::written(node, self.source) {
+            let at = run.text.text.len();
+            run.text.sources.push((at, written, escapes));
         }
         let text = &mut run.text;
         let (children, (name, style), url) = match node {

src/documents/tree/sources.rs

@@ -1,11 +1,21 @@
-use std::{iter::Peekable, ops::Range, str::CharIndices};
+use std::ops::Range;
 
-use markdown::mdast::Node;
+use markdown::{decode_named, mdast::Node};
 
-use super::{Inline, Node as Drawn, span};
+use super::{Inline, Node as Drawn, html, span};
 
-/// A leaf's written span; code and math inside their marks.
-pub(super) fn written(node: &Node, source: &str) -> Option<Range<usize>> {
+/// How a piece's source turns into its text.
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub enum Escapes {
+    Plain,
+    /// Backslash escapes and references, as Markdown reads them.
+    Markdown,
+    /// References, as HTML reads them.
+    Html,
+}
+
+/// A leaf's written span and how it reads.
+pub(super) fn written(node: &Node, source: &str) -> Option<(Range<usize>, Escapes)> {
     let whole = match node {
         Node::Text(_)
         | Node::Break(_)
@@ -17,14 +27,31 @@ pub(super) fn written(node: &Node, source: &str) -> Option<Range<usize>> {
         | Node::InlineMath(_) => span(node),
         _ => return None,
     };
-    let marks = match node {
-        Node::InlineCode(_) | Node::InlineMath(_) => {
-            let text = &source[whole.clone()];
-            text.len() - text.trim_start_matches(['`', '$']).len()
-        }
-        _ => 0,
+    Some(match node {
+        Node::Text(_) => (whole, Escapes::Markdown),
+        Node::InlineCode(_) => (inside(source, whole, '`'), Escapes::Plain),
+        Node::InlineMath(_) => (inside(source, whole, '$'), Escapes::Plain),
+        _ => (whole, Escapes::Plain),
+    })
+}
+
+/// A code span's text, one end space cut each side.
+fn inside(source: &str, whole: Range<usize>, mark: char) -> Range<usize> {
+    let written = &source[whole.clone()];
+    let open = written.len() - written.trim_start_matches(mark).len();
+    let close = written.len() - written.trim_end_matches(mark).len();
+    let text = &written[open..written.len().saturating_sub(close).max(open)];
+    let blank = |ch: char| matches!(ch, ' ' | '\n' | '\r');
+    let cut = text.starts_with(blank) && text.ends_with(blank) && !text.chars().all(blank);
+    let width = |crlf: bool| if crlf { 2 } else { 1 };
+    let (head, tail) = match cut {
+        true => (
+            width(text.starts_with("\r\n")),
+            width(text.ends_with("\r\n")),
+        ),
+        false => (0, 0),
     };
-    Some(whole.start + marks..whole.end)
+    whole.start + open + head..whole.start + open + text.len() - tail
 }
 
 /// Where a block's text lines start: past a fence line.
@@ -36,27 +63,93 @@ pub(super) fn body(source: &str, node: &Node) -> usize {
             .iter()
             .any(|fence| written.starts_with(fence));
     match fenced {
-        true => written
-            .find('\n')
-            .map_or(whole.end, |end| whole.start + end + 1),
+        true => lines(written)
+            .nth(1)
+            .map_or(whole.end, |line| whole.start + line.start),
         false => whole.start,
     }
 }
 
+/// `text`'s lines without their endings, split as CommonMark splits.
+fn lines(text: &str) -> impl Iterator<Item = Range<usize>> + '_ {
+    let mut next = Some(0);
+    std::iter::from_fn(move || {
+        let start = next?;
+        let rest = &text[start..];
+        let Some(end) = rest.find(['\n', '\r']) else {
+            next = None;
+            return Some(start..text.len());
+        };
+        let width = if rest[end..].starts_with("\r\n") {
+            2
+        } else {
+            1
+        };
+        next = Some(start + end + width);
+        Some(start..start + end)
+    })
+}
+
 /// Source byte under drawn byte `at`; lines start at `from`.
 pub(crate) fn lines_at(source: &str, from: usize, drawn: &str, at: usize) -> Option<usize> {
-    let line = drawn[..at].matches('\n').count();
-    let column = at - drawn[..at].rfind('\n').map_or(0, |end| end + 1);
-    let shown = drawn.split('\n').nth(line)?;
-    let mut start = from;
-    for _ in 0..line {
-        start += source.get(start..)?.find('\n')? + 1;
+    LineMap::new(source, from, drawn)?.at(at)
+}
+
+/// Each drawn line set against the written line it ends.
+pub(super) struct LineMap(Vec<Line>);
+
+struct Line {
+    drawn: Range<usize>,
+    /// Where its written line starts, and its shared end.
+    written: usize,
+    common: usize,
+    kept: usize,
+    /// The char before the shared end, as a cut tab.
+    before: usize,
+}
+
+impl LineMap {
+    pub(super) fn new(source: &str, from: usize, drawn: &str) -> Option<Self> {
+        let written = source.get(from..)?;
+        let lines = lines(drawn).zip(lines(written)).map(|(line, place)| {
+            let (shown, text) = (&drawn[line.clone()], &written[place.clone()]);
+            let kept: usize = match text.ends_with(shown) {
+                true => shown.len(),
+                false => (shown.chars().rev())
+                    .zip(text.chars().rev())
+                    .take_while(|(a, b)| a == b)
+                    .map(|(a, _)| a.len_utf8())
+                    .sum(),
+            };
+            let common = text.len() - kept;
+            let before =
+                (text[..common].chars().next_back()).map_or(0, |ch| common - ch.len_utf8());
+            Line {
+                drawn: line,
+                written: from + place.start,
+                common,
+                kept,
+                before,
+            }
+        });
+        Some(Self(lines.collect()))
+    }
+
+    /// Source byte under drawn byte `at`.
+    pub(super) fn at(&self, at: usize) -> Option<usize> {
+        let line = self
+            .0
+            .get(self.0.partition_point(|line| line.drawn.end < at))?;
+        let column = at.saturating_sub(line.drawn.start);
+        let shared = line.drawn.len() - line.kept;
+        Some(
+            line.written
+                + match column.checked_sub(shared) {
+                    Some(into) => line.common + into,
+                    None => line.before,
+                },
+        )
     }
-    let written = source[start..].split('\n').next()?.trim_end_matches('\r');
-    // Each drawn line ends its written one.
-    written
-        .ends_with(shown)
-        .then(|| start + written.len() - shown.len() + column)
 }
 
 impl Drawn {
@@ -64,7 +157,7 @@ impl Drawn {
     pub(crate) fn rebase(&mut self, start: usize) {
         match self {
             Drawn::Paragraph(inline) | Drawn::Heading(_, inline) => {
-                for (_, span) in &mut inline.sources {
+                for (_, span, _) in &mut inline.sources {
                     *span = span.start - start..span.end - start;
                 }
             }
@@ -90,109 +183,140 @@ impl Drawn {
 impl Inline {
     /// Source byte under drawn byte `at`, in some piece.
     pub(crate) fn source_at(&self, source: &str, at: usize) -> Option<usize> {
-        let piece = self.sources.partition_point(|(start, _)| *start <= at);
-        let (start, span) = self.sources.get(piece.checked_sub(1)?)?;
+        let piece = self.sources.partition_point(|(start, ..)| *start <= at);
+        let (start, span, escapes) = self.sources.get(piece.checked_sub(1)?)?;
         let end = self
             .sources
             .get(piece)
-            .map_or(self.text.len(), |(next, _)| *next);
+            .map_or(self.text.len(), |(next, ..)| *next);
         let drawn = self.text.get(*start..end.max(*start))?;
         let written = source.get(span.clone())?;
         // Unreadable text lands at its start.
-        Some(span.start + align(drawn, written, at - start).unwrap_or(0))
+        Some(span.start + align(drawn, written, *escapes, at - start).unwrap_or(0))
     }
 }
 
-/// Where drawn byte `at` was written, past marks.
-fn align(drawn: &str, written: &str, at: usize) -> Option<usize> {
-    let mut reading = Reading::new(written, 0);
-    for (byte, shown) in drawn.char_indices() {
-        let found = reading.take(shown)?;
-        if byte >= at {
+/// A source char as read, and where it was written.
+struct Read {
+    ch: char,
+    at: usize,
+    /// Written as itself, not escaped or referenced.
+    raw: bool,
+    /// A container mark or indent at a line's start.
+    lead: bool,
+}
+
+impl Read {
+    fn ends_line(&self) -> bool {
+        self.raw && matches!(self.ch, '\n' | '\r')
+    }
+}
+
+/// `written`'s chars, escapes and references read.
+fn read(written: &str, escapes: Escapes) -> Vec<Read> {
+    let mut out = Vec::with_capacity(written.len());
+    let (mut at, mut line_start) = (0, false);
+    while let Some(ch) = written[at..].chars().next() {
+        let rest = &written[at..];
+        let decoded = match (escapes, ch) {
+            (Escapes::Markdown, '\\') => (rest[1..].chars().next())
+                .filter(char::is_ascii_punctuation)
+                .map(|punct| (punct.to_string(), 2, at + 1)),
+            (Escapes::Markdown, '&') => reference(rest).map(|(value, used)| (value, used, at)),
+            (Escapes::Html, '&') => {
+                html::reference(rest, false).map(|(value, used)| (value, used, at))
+            }
+            _ => None,
+        };
+        if let Some((value, used, place)) = decoded {
+            out.extend(value.chars().map(|ch| Read {
+                ch,
+                at: place,
+                raw: false,
+                lead: false,
+            }));
+            (at, line_start) = (at + used, false);
+            continue;
+        }
+        let ends = matches!(ch, '\n' | '\r');
+        let lead = line_start && !ends && matches!(ch, ' ' | '\t' | '>');
+        out.push(Read {
+            ch,
+            at,
+            raw: true,
+            lead,
+        });
+        (at, line_start) = (at + ch.len_utf8(), ends || lead);
+    }
+    out
+}
+
+/// A reference at `text`'s start, as Markdown reads it.
+fn reference(text: &str) -> Option<(String, usize)> {
+    let body = text.strip_prefix('&')?;
+    let (end, _) = body.char_indices().take(34).find(|(_, ch)| *ch == ';')?;
+    let name = &body[..end];
+    let code = |digits: &str, radix: u32, most: usize| {
+        let fits =
+            (1..=most).contains(&digits.len()) && digits.chars().all(|ch| ch.is_digit(radix));
+        let code = u32::from_str_radix(digits, radix).ok().filter(|_| fits)?;
+        let ch = char::from_u32(code).filter(|ch| *ch != '\0');
+        Some(ch.unwrap_or(char::REPLACEMENT_CHARACTER).to_string())
+    };
+    let value = match name.strip_prefix('#') {
+        Some(hex) if hex.starts_with(['x', 'X']) => code(&hex[1..], 16, 6)?,
+        Some(digits) => code(digits, 10, 7)?,
+        None if name.chars().all(|ch| ch.is_ascii_alphanumeric()) => decode_named(name, true)?,
+        None => return None,
+    };
+    Some((value, end + 2))
+}
+
+/// Where drawn byte `at` was written, matched from the end.
+fn align(drawn: &str, written: &str, escapes: Escapes, at: usize) -> Option<usize> {
+    if at >= drawn.len() {
+        return Some(written.len());
+    }
+    let source = read(written, escapes);
+    let mut j = source.len();
+    let mut shown = drawn.char_indices().rev().peekable();
+    while let Some((byte, d)) = shown.next() {
+        let found = loop {
+            let s = source[..j].last()?;
+            // A join: a line's end and the marks after it.
+            if d.is_whitespace() && (s.ends_line() || s.lead) {
+                while source[..j].last().is_some_and(|s| s.lead) {
+                    j -= 1;
+                }
+                j -= 1;
+                let crlf = source[j].ch == '\n' && j > 0;
+                if crlf && source[j - 1].ends_line() && source[j - 1].ch == '\r' {
+                    j -= 1;
+                }
+                if d == '\n' && shown.peek().is_some_and(|(_, ch)| *ch == '\r') {
+                    let (cr, _) = shown.next().expect("peeked");
+                    if cr == at {
+                        return Some(source[j].at);
+                    }
+                }
+                break source[j].at;
+            }
+            if d == s.ch {
+                j -= 1;
+                break s.at;
+            }
+            // White space, line ends and marks the text dropped.
+            if s.raw && (matches!(s.ch, ' ' | '\t') || s.ends_line() || s.lead) {
+                j -= 1;
+                continue;
+            }
+            return None;
+        };
+        if byte == at {
             return Some(found);
         }
     }
-    Some(reading.end)
-}
-
-/// Finds drawn chars, front to back, in written text.
-struct Reading<'a> {
-    written: Peekable<CharIndices<'a>>,
-    base: usize,
-    /// Past the last char found.
-    end: usize,
-}
-
-impl<'a> Reading<'a> {
-    fn new(written: &'a str, base: usize) -> Self {
-        Self {
-            written: written.char_indices().peekable(),
-            base,
-            end: base,
-        }
-    }
-
-    /// Where `shown` was written; true if a char holds it.
-    fn at(&mut self, shown: char) -> Option<(usize, bool)> {
-        loop {
-            let Some(&(byte, written)) = self.written.peek() else {
-                return shown.is_whitespace().then_some((self.end, false));
-            };
-            let byte = self.base + byte;
-            match (shown, written) {
-                (a, b) if a == b => return Some((byte, true)),
-                (a, b) if a.is_whitespace() => return Some((byte, b.is_whitespace())),
-                // Joins, indents, escapes and quote marks.
-                (_, ' ' | '\t' | '\r' | '\n' | '\\' | '>') => {
-                    self.written.next();
-                }
-                _ => return None,
-            }
-        }
-    }
-
-    fn take(&mut self, shown: char) -> Option<usize> {
-        let (byte, held) = self.at(shown)?;
-        if held {
-            let (_, written) = self.written.next().expect("a held char");
-            self.end = byte + written.len_utf8();
-        }
-        Some(byte)
-    }
-}
-
-/// A node's value bytes mapped to source, front to back.
-pub(super) struct Follow<'a> {
-    value: &'a str,
-    done: usize,
-    reading: Option<Reading<'a>>,
-}
-
-impl<'a> Follow<'a> {
-    pub(super) fn new(value: &'a str, source: &'a str, span: Range<usize>) -> Self {
-        let reading = Reading::new(&source[span.clone()], span.start);
-        Self {
-            value,
-            done: 0,
-            reading: Some(reading),
-        }
-    }
-
-    /// Where value byte `at` was written.
-    pub(super) fn to(&mut self, at: usize) -> Option<usize> {
-        let reading = self.reading.as_mut()?;
-        let lost = self.value[self.done..at]
-            .chars()
-            .any(|shown| reading.take(shown).is_none());
-        self.done = at;
-        if lost {
-            log::warn!("markdown: a block's text lost its source");
-            self.reading = None;
-            return None;
-        }
-        let next = self.value[at..].chars().next()?;
-        self.reading.as_mut()?.at(next).map(|(byte, _)| byte)
-    }
+    None
 }
 
 #[cfg(test)]
@@ -226,10 +350,11 @@ mod tests {
     }
 
     #[test]
-    fn text_that_does_not_read_back_lands_at_its_start() {
+    fn references_read_back_and_emoji_end_the_reading() {
         let source = "a &amp; b";
-        assert_eq!(under(source, "b"), 0, "the reference ends the reading");
-        let source = "x :smile: y";
-        assert_eq!(under(source, "y"), 0);
+        assert_eq!(under(source, "b"), source.find('b').unwrap());
+        let source = "ab :smile: y";
+        assert_eq!(under(source, "y"), source.find('y').unwrap());
+        assert_eq!(under(source, "b"), 0, "before the emoji: its start");
     }
 }