Kigi-CLI/crates/codegen/kigi-markdown/src/parse.rs

//! Markdown parser - transforms markdown text into styled highlight ranges.
//!
//! The parser processes markdown events and populates buffers with:
//! - Highlights: Style ranges for inline formatting
//! - Replaces: Syntax-highlighted code blocks
//! - Transforms: Character substitutions (bullets, etc.)
//! - Table replaces: Formatted table content

use std::ops::Range;

use anstyle::Style;
use pulldown_cmark::{CodeBlockKind, CowStr, Event, Tag, TagEnd, TextMergeWithOffset};
use ratatui::style::Stylize as RatatuiStylize;
use ratatui::text::{Line, Span};

use crate::buffers::{
    CodeBlockMeta, Highlight, LinkTarget, MarkdownBuffers, Replace, StyledCell, TableHyperlink,
    TableReplace, TableState, Transform, floor_char_boundary, unicode_display_width,
};
use crate::checkpoint::CheckpointKind;
use crate::latex;
use crate::open_code_highlighter::OpenCodeHighlighter;
use crate::style::{MarkdownStyle, TableBorders};
use crate::syntax::{Syntect, syntax_highlight_raw};

/// Trait for converting anstyle to ratatui style.
trait StyleInto<T> {
    fn style_into(self) -> T;
}

impl StyleInto<ratatui::style::Style> for Style {
    fn style_into(self) -> ratatui::style::Style {
        use ratatui::style::{Modifier, Style as RStyle};

        let mut style = RStyle::default();

        if let Some(fg) = self.get_fg_color() {
            style = style.fg(anstyle_to_ratatui_color(fg));
        }
        if let Some(bg) = self.get_bg_color() {
            style = style.bg(anstyle_to_ratatui_color(bg));
        }

        let effects = self.get_effects();
        let mut modifiers = Modifier::empty();
        if effects.contains(anstyle::Effects::BOLD) {
            modifiers |= Modifier::BOLD;
        }
        if effects.contains(anstyle::Effects::DIMMED) {
            modifiers |= Modifier::DIM;
        }
        if effects.contains(anstyle::Effects::ITALIC) {
            modifiers |= Modifier::ITALIC;
        }
        if effects.contains(anstyle::Effects::UNDERLINE) {
            modifiers |= Modifier::UNDERLINED;
        }
        if effects.contains(anstyle::Effects::STRIKETHROUGH) {
            modifiers |= Modifier::CROSSED_OUT;
        }
        if effects.contains(anstyle::Effects::HIDDEN) {
            modifiers |= Modifier::HIDDEN;
        }

        style.add_modifier(modifiers)
    }
}

fn anstyle_to_ratatui_color(color: anstyle::Color) -> ratatui::style::Color {
    use ratatui::style::Color;
    match color {
        anstyle::Color::Ansi(ansi) => match ansi {
            anstyle::AnsiColor::Black => Color::Black,
            anstyle::AnsiColor::Red => Color::Red,
            anstyle::AnsiColor::Green => Color::Green,
            anstyle::AnsiColor::Yellow => Color::Yellow,
            anstyle::AnsiColor::Blue => Color::Blue,
            anstyle::AnsiColor::Magenta => Color::Magenta,
            anstyle::AnsiColor::Cyan => Color::Cyan,
            anstyle::AnsiColor::White => Color::Gray,
            anstyle::AnsiColor::BrightBlack => Color::DarkGray,
            anstyle::AnsiColor::BrightRed => Color::LightRed,
            anstyle::AnsiColor::BrightGreen => Color::LightGreen,
            anstyle::AnsiColor::BrightYellow => Color::LightYellow,
            anstyle::AnsiColor::BrightBlue => Color::LightBlue,
            anstyle::AnsiColor::BrightMagenta => Color::LightMagenta,
            anstyle::AnsiColor::BrightCyan => Color::LightCyan,
            anstyle::AnsiColor::BrightWhite => Color::White,
        },
        anstyle::Color::Ansi256(idx) => Color::Indexed(idx.index()),
        anstyle::Color::Rgb(rgb) => Color::Rgb(rgb.0, rgb.1, rgb.2),
    }
}

/// Find a substring within a haystack, optionally searching outside or using rfind.
fn find_substring(
    haystack: &str,
    needle: &CowStr,
    allow_outside: bool,
    rfind: bool,
) -> Option<Range<usize>> {
    if needle.is_empty() || haystack.is_empty() || needle.len() > haystack.len() {
        return None;
    }
    if !allow_outside {
        if let CowStr::Borrowed(needle) = needle {
            let (hp, np) = (haystack.as_ptr(), needle.as_ptr());
            unsafe {
                let (he, ne) = (hp.add(haystack.len()), np.add(needle.len()));
                if np >= hp && ne <= he {
                    let offset = np.offset_from(hp) as usize;
                    let range = offset..(offset + needle.len());
                    if cfg!(debug_assertions) {
                        assert_eq!(&haystack.as_bytes()[range.clone()], needle.as_bytes());
                    }
                    return Some(range);
                }
            }
        }
        None
    } else {
        if rfind {
            haystack.rfind(needle.as_ref())
        } else {
            haystack.find(needle.as_ref())
        }
        .map(|pos| pos..(pos + needle.len()))
    }
}

/// Decode a single HTML character entity reference (`entity` includes the
/// leading `&` and trailing `;`) into its replacement string.
///
/// Delegates to [`html_escape`] for the full HTML5 named set plus numeric
/// references (decimal `&#NN;` and hexadecimal `&#xNN;`), matching what
/// pulldown-cmark decodes in table cells so prose and tables stay consistent.
///
/// Returns `None` when:
/// - the reference is unrecognized (`html_escape` leaves it unchanged), or
/// - it decodes to a control character (`&#27;`, `&#0;`, …). Substituting raw
///   control bytes would let untrusted markdown inject terminal escape
///   sequences, so the raw source is left literal instead.
fn decode_html_entity(entity: &str) -> Option<String> {
    let decoded = html_escape::decode_html_entities(entity);
    // Unchanged output means `html_escape` did not recognize the reference.
    if decoded.as_ref() == entity {
        return None;
    }
    if decoded.chars().any(char::is_control) {
        return None;
    }
    Some(decoded.into_owned())
}

/// Check if there's a blank line (empty line) after the given position.
fn has_blank_line_after(text: &str, pos: usize) -> bool {
    text.as_bytes()[pos..]
        .iter()
        .copied()
        .find(|&c| c != b' ' && c != b'\t')
        == Some(b'\n')
}

/// Transient state for the fenced code block currently being parsed.
///
/// Fenced blocks never nest (an inner fence closes the outer), so a single
/// `Option` suffices. Finalized in the `TagEnd::CodeBlock` arm, where the body
/// range and the block range together decide whether the fence was closed.
struct PendingCodeBlock {
    info: String,
    /// Body byte range in the raw source. Initialized to an empty range just
    /// past the opening fence line, then widened to the merged body text range
    /// as text events arrive (`body_seen` distinguishes the empty-body case).
    body_range: Range<usize>,
    /// De-prefixed body content: pulldown's merged text gives the logical code
    /// with container markers (blockquote `>`, list indent) stripped and CRLF
    /// normalized to `\n` — i.e. the clean diagram/code source.
    body_text: String,
    body_seen: bool,
}

/// Markdown parser that processes events and populates buffers.
///
/// After calling `parse()`, the transient state (tag_stack, table_state, depth)
/// is dropped and a `ParsedMarkdown` is returned for rendering.
pub struct MarkdownParser<'a, 'b, 'syn, 'oc> {
    text: &'a str,
    ms: MarkdownStyle,
    buffers: &'b mut MarkdownBuffers,
    syntect: Option<&'syn Syntect>,
    /// Incremental highlighter for the trailing still-open fenced code block.
    /// Only set by the streaming tail re-render; `None` for batch renders, in
    /// which case code blocks go through the from-scratch [`syntax_highlight_raw`].
    open_code: Option<&'oc mut OpenCodeHighlighter>,
    // Transient state (dropped after parse)
    tag_stack: Vec<Tag<'a>>,
    table_state: Option<TableState>,
    depth: usize,
    /// Current blockquote nesting depth (0 = not in any blockquote).
    /// Used to determine which `>` on a line belongs to the current level.
    bq_depth: usize,
    last_checkpoint: Option<(CheckpointKind, usize)>,
    /// Maximum width for rendered tables (in display columns).
    /// When `Some(w)`, column widths are shrunk proportionally so the table
    /// fits within `w` columns.  When `None`, columns use natural widths.
    max_table_width: Option<usize>,
    /// Monotonically increasing counter for assigning stable link IDs.
    /// Persisted across `rerender_tail` calls via the streaming renderer.
    link_id_counter: u32,
    /// When `true` (default), CommonMark soft breaks inside a paragraph
    /// collapse to a single space. When `false`, the source newline is
    /// preserved so each source line surfaces as its own visual line —
    /// required by the line-numbered plan preview, where rendered lines
    /// must map 1:1 to file lines.
    collapse_soft_breaks: bool,
    /// In-progress fenced code block, set between its start and end events.
    pending_code_block: Option<PendingCodeBlock>,
}

/// Custom word separator for table cells.
///
/// Like `AsciiSpace`, but also treats punctuation and symbol characters as
/// break opportunities when followed by a letter.  This lets tables break
/// lines at e.g. `foo/bar` or `hello-world` without ever splitting mid-word.
///
/// For each break point, the punctuation character is attached to whichever
/// side produces the shorter maximum segment — e.g. `ABCD-EFG` becomes
/// `ABCD` + `-EFG` (max 4) rather than `ABCD-` + `EFG` (max 5).
///
/// Only `,` and `.` between digits suppress the break — these are number
/// formatting (e.g. `$145,000`, `3.14`).  All other punctuation can break
/// even between digits, so phone numbers (`555-0101`), dates (`2019-03-15`)
/// etc. become breakable.
///
/// Returns `true` for `<br>`, `<br/>`, `<br />`, etc. (case-insensitive).
fn is_br_tag(html: &str) -> bool {
    let Some(inner) = html
        .trim()
        .strip_prefix('<')
        .and_then(|s| s.strip_suffix('>'))
    else {
        return false;
    };
    let tag = inner.trim();
    let tag = tag.strip_suffix('/').map_or(tag, str::trim);
    tag.eq_ignore_ascii_case("br")
}

/// URLs (validated via `url::Url::parse`) are treated as unbreakable
/// words — no break points are placed within a URL so that terminal
/// hyperlink detection (Cmd+Click) continues to work when cells wrap.
pub(crate) fn cell_word_separator<'a>(
    line: &'a str,
) -> Box<dyn Iterator<Item = textwrap::core::Word<'a>> + 'a> {
    // Pass 1: find break-point byte positions.
    // A break point sits between a punctuation/symbol char and the alphabetic
    // char that follows it.  We record (break_byte_idx, punct_byte_start)
    // where break_byte_idx is where the next word would start if we attach
    // the punct char to the left, and punct_byte_start is where the punct
    // char begins (for attaching it to the right instead).
    let mut breaks: Vec<(usize, usize)> = Vec::new();
    {
        let mut in_whitespace = false;
        let mut after_break_char = false;
        let mut prev_is_digit = false; // was the *previous* char a digit?
        let mut digit_before_break = false; // was the char before the break char a digit?
        let mut last_break_ch: char = '\0';
        let mut break_char_start: usize = 0;
        for (idx, ch) in line.char_indices() {
            let is_space = ch == ' ';
            let is_break_char = !is_space && !ch.is_alphanumeric();

            // After a break char, decide if we should split here.
            //
            // Two cases allow a break:
            //  a) Followed by a letter → always break (new word boundary).
            //  b) Followed by a digit AND the char before the punct was
            //     also a digit → break, UNLESS the punct is `,` or `.`
            //     (number formatting like `$145,000` or `3.14`).
            //
            // This means:
            //  - `foo/bar` breaks (letter after punct)           ✓
            //  - `555-0101` breaks (digit-hyphen-digit)          ✓
            //  - `$145,000` stays (digit-comma-digit)            ✓
            //  - `$145` stays (no digit before `$`)              ✓
            //  - `EMP-1001` breaks at hyphen (letter before it)  ✓
            let should_break = if in_whitespace && !is_space {
                true
            } else if after_break_char {
                if ch.is_alphabetic() {
                    true
                } else if ch.is_ascii_digit() && digit_before_break {
                    // digit-punct-digit: only break for non-formatting punct
                    last_break_ch != ',' && last_break_ch != '.'
                } else {
                    false
                }
            } else {
                false
            };

            if should_break {
                if in_whitespace {
                    breaks.push((idx, idx));
                } else {
                    breaks.push((idx, break_char_start));
                }
            }

            if is_break_char {
                break_char_start = idx;
                last_break_ch = ch;
                digit_before_break = prev_is_digit;
            }
            prev_is_digit = ch.is_ascii_digit();
            in_whitespace = is_space;
            after_break_char = is_break_char;
        }
    }

    // Filter out break points that fall inside a URL.
    // Each whitespace-delimited token is tested with `url::Url::parse`;
    // tokens that parse as valid URLs are protected from splitting.
    let url_ranges: Vec<Range<usize>> = {
        let mut ranges = Vec::new();
        let mut pos = 0;
        for token in line.split_whitespace() {
            let start = line[pos..].find(token).unwrap() + pos;
            let end = start + token.len();
            if url::Url::parse(token).is_ok() {
                ranges.push(start..end);
            }
            pos = end;
        }
        ranges
    };
    breaks.retain(|&(break_pos, _)| {
        !url_ranges
            .iter()
            .any(|r| break_pos > r.start && break_pos < r.end)
    });

    // Pass 2: decide attachment for each break point.
    // For punct breaks, choose the side that minimizes max(left_len, right_len).
    let mut split_positions: Vec<usize> = Vec::with_capacity(breaks.len());
    {
        let len = line.len();
        for (i, &(attach_left, attach_right)) in breaks.iter().enumerate() {
            if attach_left == attach_right {
                // Whitespace break — no choice.
                split_positions.push(attach_left);
            } else {
                // Determine segment boundaries for this break.
                let seg_start = if i == 0 { 0 } else { split_positions[i - 1] };
                let seg_end = if i + 1 < breaks.len() {
                    // Use the leftward attachment of the next break as a
                    // conservative estimate of the right segment end.
                    breaks[i + 1].0
                } else {
                    len
                };

                let left_if_attach_left = unicode_display_width(&line[seg_start..attach_left]);
                let right_if_attach_left = unicode_display_width(&line[attach_left..seg_end]);
                let max_attach_left = left_if_attach_left.max(right_if_attach_left);

                let left_if_attach_right = unicode_display_width(&line[seg_start..attach_right]);
                let right_if_attach_right = unicode_display_width(&line[attach_right..seg_end]);
                let max_attach_right = left_if_attach_right.max(right_if_attach_right);

                if max_attach_right < max_attach_left {
                    split_positions.push(attach_right);
                } else {
                    split_positions.push(attach_left);
                }
            }
        }
    }

    // Pass 3: emit Words at the chosen split positions.
    let mut pos = 0usize;
    let mut idx = 0usize;
    Box::new(std::iter::from_fn(move || {
        if pos >= line.len() {
            return None;
        }
        let end = if idx < split_positions.len() {
            let e = split_positions[idx];
            idx += 1;
            e
        } else {
            line.len()
        };
        let word = textwrap::core::Word::from(&line[pos..end]);
        pos = end;
        Some(word)
    }))
}

/// Output of [`MarkdownParser::format_table`]: the rendered lines of a single table.
#[derive(Default)]
struct FormattedTable {
    /// Plain-text lines (for ANSI rendering).
    lines: Vec<String>,
    /// Styled lines (for ratatui rendering).
    styled_lines: Vec<Line<'static>>,
    /// Per-line source offset within the table (0 = header, 1 = separator, 2+ = body rows).
    line_source_offsets: Vec<usize>,
    /// Hyperlinks (in table-local line coordinates) for links inside cells.
    hyperlinks: Vec<TableHyperlink>,
}

impl<'a, 'b, 'syn, 'oc> MarkdownParser<'a, 'b, 'syn, 'oc> {
    pub fn new(
        text: &'a str,
        ms: MarkdownStyle,
        buffers: &'b mut MarkdownBuffers,
        syntect: Option<&'syn Syntect>,
    ) -> Self {
        Self {
            text,
            ms,
            buffers,
            syntect,
            open_code: None,
            tag_stack: Vec::new(),
            table_state: None,
            depth: 0,
            bq_depth: 0,
            last_checkpoint: None,
            max_table_width: None,
            link_id_counter: 0,
            collapse_soft_breaks: true,
            pending_code_block: None,
        }
    }

    /// Set whether CommonMark soft breaks collapse to a space.
    ///
    /// Defaults to `true`. Set `false` for source-faithful rendering (plan
    /// preview) where each source line must keep its own visual line and
    /// `line_source_map` entry.
    pub fn collapse_soft_breaks(mut self, collapse: bool) -> Self {
        self.collapse_soft_breaks = collapse;
        self
    }

    /// Set the maximum width for rendered tables.
    ///
    /// When set, column widths are shrunk proportionally so the table
    /// fits within the given number of display columns.
    pub fn max_table_width(mut self, width: Option<usize>) -> Self {
        self.max_table_width = width;
        self
    }

    /// Set the starting link ID counter (for streaming renderer continuity).
    ///
    /// Internal: only the in-crate streaming renderer needs to manage the
    /// link counter across `rerender_tail` calls.  Consumers should use
    /// `StreamingMarkdownRenderer` instead of touching the parser directly.
    pub(crate) fn link_id_start(mut self, id: u32) -> Self {
        self.link_id_counter = id;
        self
    }

    /// Provide an incremental highlighter for the trailing still-open fenced
    /// code block (streaming tail re-render only).
    ///
    /// Internal: lets `rerender_tail` persist syntect's resumable per-line state
    /// across passes so an open code block is highlighted in O(N) total instead
    /// of O(N²). Batch/non-streaming callers leave this `None`.
    pub(crate) fn open_code(mut self, cache: Option<&'oc mut OpenCodeHighlighter>) -> Self {
        self.open_code = cache;
        self
    }

    /// Parse markdown and return a ParsedMarkdown ready for rendering.
    ///
    /// Consumes self, dropping transient parsing state.
    pub fn parse(mut self) -> ParsedMarkdown<'a, 'b> {
        self.tag_stack.clear();
        self.buffers.clear();
        self.table_state = None;
        self.depth = 0;
        self.last_checkpoint = None;
        self.pending_code_block = None;

        for (event, range) in TextMergeWithOffset::new(kigi_markdown_core::offset_events(self.text))
        {
            self.on_event(event, range);
        }

        ParsedMarkdown::new(
            self.text,
            self.ms,
            self.buffers,
            self.last_checkpoint,
            self.link_id_counter,
        )
    }

    fn push_highlight(&mut self, style: Option<Style>, range: &Range<usize>) {
        self.buffers.highlights.push(Highlight {
            style,
            range: range.clone(),
        });
    }

    fn on_event(&mut self, event: Event<'a>, range: Range<usize>) {
        let mut parent_code_block = None;

        // Apply ALL ancestors' inner styles to non-marker events.
        let skip_inner_style = matches!(
            event,
            Event::Start(_) | Event::End(_) | Event::Code(_) | Event::InlineMath(_)
        );

        // Collect ancestor styles first (to avoid borrow issues)
        let ancestor_styles: Vec<Option<Style>> = if !skip_inner_style {
            // Inside a link, inline-format ancestors (strong/emphasis/
            // strikethrough) must not recolor the link text: their inner
            // styles carry the theme's default text fg, and these highlights
            // land *after* the link_text highlight pushed at Tag::Link start
            // — merge_styles is last-wins on fg, so keeping the fg would
            // clobber the link color (e.g. `**[bold link](url)**`). Only the
            // fg competes with link_text today, so effects (and any bg) pass
            // through.
            let in_link = self
                .tag_stack
                .iter()
                .any(|t| matches!(t, Tag::Link { .. } | Tag::Image { .. }));
            let strip_fg_in_link =
                |style: Style| if in_link { style.fg_color(None) } else { style };
            self.tag_stack
                .iter()
                .filter_map(|ancestor| match ancestor {
                    Tag::Heading { level, .. } => {
                        Some(Some(self.ms.heading_inner[(*level as i32) as usize - 1]))
                    }
                    Tag::Emphasis => Some(Some(strip_fg_in_link(self.ms.emphasis_inner))),
                    Tag::Strong => Some(Some(strip_fg_in_link(self.ms.strong_inner))),
                    Tag::Strikethrough => Some(Some(strip_fg_in_link(self.ms.strikethrough_inner))),
                    // Link/Image already push their own inner-style highlight
                    // (link_text) during on_start.  We just need ancestor_styles
                    // to be non-empty so the Event::Text branch below skips
                    // pushing ms.text — which would otherwise override the
                    // link_text foreground color via merge_styles' last-wins
                    // ordering.
                    Tag::Link { .. } | Tag::Image { .. } => Some(None),
                    Tag::CodeBlock(block) => {
                        parent_code_block = Some(match block {
                            CodeBlockKind::Fenced(lang) if !lang.is_empty() => {
                                Some(lang.to_owned())
                            }
                            _ => None,
                        });
                        None
                    }
                    _ => None,
                })
                .collect()
        } else {
            if let Some(Tag::CodeBlock(block)) = self.tag_stack.last() {
                parent_code_block = Some(match block {
                    CodeBlockKind::Fenced(lang) if !lang.is_empty() => Some(lang.to_owned()),
                    _ => None,
                });
            }
            Vec::new()
        };

        for style in &ancestor_styles {
            self.push_highlight(*style, &range);
        }

        match event {
            Event::Start(tag) => self.on_start(tag, range),
            Event::End(tag_end) => self.on_end(tag_end, range),
            Event::Text(text) => {
                // Capture text into table cell if we're inside a table
                if let Some(ref mut state) = self.table_state {
                    state.push_text(&text);
                }

                // Record the enclosing fenced block's raw byte range and its
                // de-prefixed body content. pulldown merges the body into one
                // text event, but accumulate defensively in case it is split.
                if parent_code_block.is_some()
                    && let Some(pending) = self.pending_code_block.as_mut()
                {
                    if pending.body_seen {
                        pending.body_range.start = pending.body_range.start.min(range.start);
                        pending.body_range.end = pending.body_range.end.max(range.end);
                    } else {
                        pending.body_range = range.clone();
                        pending.body_seen = true;
                    }
                    pending.body_text.push_str(&text);
                }

                if let Some(parent_code_block) = parent_code_block {
                    // Closed mermaid fences render as a diagram; open ones fall
                    // through so the source shows while still streaming.
                    if let Some(lang) = parent_code_block.as_deref()
                        && lang
                            .split_whitespace()
                            .next()
                            .is_some_and(|t| t.eq_ignore_ascii_case("mermaid"))
                        && range.end < self.text.len()
                        && self.try_push_mermaid(&text, &range)
                    {
                        return;
                    }
                    let highlighted = match parent_code_block {
                        Some(lang) => {
                            if let Some(syn) = self.syntect
                                && let Some(cache) = self.open_code.as_deref_mut()
                            {
                                // Streaming tail: the cache routes between the
                                // incremental open-block path and the
                                // closed-fence memo.
                                cache.highlight_block(
                                    syn,
                                    &lang,
                                    range.start,
                                    range.end >= self.text.len(),
                                    &text,
                                )
                            } else {
                                // Batch render (no streaming caches attached).
                                syntax_highlight_raw(self.syntect, &lang, &text)
                            }
                        }
                        None => None,
                    };
                    if let Some(highlighted) = highlighted {
                        self.buffers.replaces.push(Replace {
                            highlighted,
                            range: range.clone(),
                        });
                    } else {
                        self.push_highlight(Some(self.ms.code_untagged), &range);
                        self.buffers.untagged_code_ranges.push(range.clone());
                    }
                } else {
                    if ancestor_styles.is_empty() {
                        // Apply the default text style only when no ancestor
                        // (heading, strong, emphasis, etc.) already provides a
                        // color — otherwise ms.text would override them.
                        self.push_highlight(Some(self.ms.text), &range);
                    } else {
                        self.push_highlight(None, &range);
                    }
                    if self.table_state.is_none() {
                        self.scan_inline_html_entities(&range);
                    }
                }
            }
            Event::Code(code) => {
                // Capture code content into table cell if we're inside a table
                if let Some(ref mut state) = self.table_state {
                    let prev_code = state.cell_code;
                    state.cell_code = true;
                    state.push_text(&code);
                    state.cell_code = prev_code;
                }
                self.style_inline_code_span(&code, &range);
            }
            Event::InlineMath(math) => {
                // `$...$` inline math: render the TeX to Unicode and swap it
                // in via a pretty-mode transform. Falls back to inline-code
                // presentation when conversion declines (oversized input) or
                // produces nothing visible.
                let rendered = latex::latex_to_unicode_inline(&math).filter(|r| !r.is_empty());

                if let Some(ref mut state) = self.table_state {
                    match &rendered {
                        Some(r) => {
                            let prev_italic = state.cell_italic;
                            state.cell_italic = true;
                            state.push_text(r);
                            state.cell_italic = prev_italic;
                        }
                        None => {
                            let prev_code = state.cell_code;
                            state.cell_code = true;
                            state.push_text(&math);
                            state.cell_code = prev_code;
                        }
                    }
                }

                match rendered {
                    Some(r) => {
                        // One highlight + one transform spanning the entire
                        // `$...$` range: pretty mode shows the rendered math,
                        // raw mode shows the TeX source in the math style.
                        self.push_highlight(Some(self.ms.math), &range);
                        self.buffers.transforms.push(Transform {
                            range: range.clone(),
                            to: r,
                            force: false,
                        });
                    }
                    None => self.style_inline_code_span(&math, &range),
                }
            }
            Event::SoftBreak => {
                // Collapse soft breaks to spaces unless the next source
                // byte is a list-item indent or blockquote `>` marker
                // (the byte immediately after pulldown's SoftBreak range),
                // in which case the line ending belongs to a block
                // continuation and the renderer surfaces it as its own
                // visual line. The transform spans the full range so CRLF
                // (`\r\n`, 2 bytes) preserves byte length.
                if let Some(ref mut state) = self.table_state {
                    state.push_text(" ");
                } else {
                    let next = self.text.as_bytes().get(range.end);
                    let is_continuation = matches!(next, Some(b' ' | b'\t' | b'>' | b'|'));
                    if self.collapse_soft_breaks && !is_continuation {
                        let span = range.end - range.start;
                        debug_assert!(span >= 1, "SoftBreak range must cover at least one byte");
                        self.buffers.transforms.push(Transform {
                            range: range.clone(),
                            to: " ".repeat(span),
                            force: true,
                        });
                    }
                }
                self.push_highlight(None, &range);
            }
            Event::HardBreak => {
                if let Some(ref mut state) = self.table_state {
                    state.push_text("\n");
                }
                self.push_highlight(None, &range);
            }
            Event::Html(_) => {
                // Render HTML block content as regular text (not code).
                // pulldown-cmark treats XML-like tags (e.g. <example>) as HTML
                // blocks, which previously got code-block styling via Replace.
                self.push_highlight(Some(self.ms.text), &range);
            }
            Event::InlineHtml(html) => {
                if is_br_tag(&html) {
                    if let Some(ref mut state) = self.table_state {
                        state.push_text("\n");
                        self.push_highlight(Some(self.ms.text), &range);
                    } else {
                        self.buffers.transforms.push(Transform {
                            range: range.clone(),
                            to: "\n".to_string(),
                            force: false,
                        });
                        self.push_highlight(None, &range);
                    }
                } else if let Some(ref mut state) = self.table_state {
                    state.push_text(&html);
                    self.push_highlight(Some(self.ms.text), &range);
                } else if let Some(highlighted) = syntax_highlight_raw(self.syntect, "html", &html)
                {
                    self.buffers.replaces.push(Replace {
                        highlighted,
                        range: range.clone(),
                    });
                }
            }
            Event::DisplayMath(math) => {
                // `$$...$$` display math: render to Unicode block lines.
                if let Some(ref mut state) = self.table_state {
                    // Inside a table cell there is no room for a block:
                    // render single-line (rows joined with `; `).
                    match latex::latex_to_unicode_inline(&math).filter(|r| !r.is_empty()) {
                        Some(r) => {
                            let prev_italic = state.cell_italic;
                            state.cell_italic = true;
                            state.push_text(&r);
                            state.cell_italic = prev_italic;
                        }
                        None => {
                            let prev_code = state.cell_code;
                            state.cell_code = true;
                            state.push_text(&math);
                            state.cell_code = prev_code;
                        }
                    }
                    self.push_highlight(Some(self.ms.math), &range);
                } else if self.push_display_math_block(range.clone(), &math) {
                    // Raw mode shows the TeX source in the math style; pretty
                    // mode consumes the range via the block replacement.
                    self.push_highlight(Some(self.ms.math), &range);
                } else {
                    // Fallback (conversion declined / nothing visible):
                    // legacy presentation — TeX source highlighted as code.
                    self.push_highlight(Some(self.ms.code_outer), &range);
                    let outer_text = &self.text[range.clone()];
                    if let Some(r) = find_substring(outer_text, &math, true, false) {
                        let inner_range = (range.start + r.start)..(range.start + r.end);
                        if let Some(highlighted) = syntax_highlight_raw(self.syntect, "tex", &math)
                        {
                            self.buffers.replaces.push(Replace {
                                highlighted,
                                range: inner_range,
                            });
                        } else {
                            self.push_highlight(Some(self.ms.code_untagged), &inner_range);
                        }
                    }
                }
            }
            Event::FootnoteReference(_) => {
                self.push_highlight(Some(self.ms.link_outer), &range);
            }
            Event::Rule => {
                // Style and transform "---" to "───" (horizontal rule)
                self.push_highlight(Some(self.ms.rule), &range);
                let rule_text = &self.text[range.clone()];
                if let Some(marker_end) = rule_text.find('\n') {
                    // Transform only up to the newline
                    self.buffers.transforms.push(Transform {
                        range: range.start..range.start + marker_end,
                        to: "───".to_string(),
                        force: false,
                    });
                } else {
                    // No trailing newline, transform the whole range
                    self.buffers.transforms.push(Transform {
                        range: range.clone(),
                        to: "───".to_string(),
                        force: false,
                    });
                }
                if self.depth == 0 {
                    self.last_checkpoint = Some((CheckpointKind::ThematicBreak, range.end));
                }
            }
            Event::TaskListMarker(checked) => {
                let style = if checked {
                    self.ms.task_checked
                } else {
                    self.ms.task_unchecked
                };
                self.push_highlight(Some(style), &range);
            }
        }
    }

    fn on_start(&mut self, tag: Tag<'a>, range: Range<usize>) {
        // Track nesting depth for checkpoint detection
        match &tag {
            Tag::BlockQuote(_) | Tag::List(_) | Tag::Item | Tag::Table(_) => {
                self.depth += 1;
            }
            _ => {}
        }
        if matches!(&tag, Tag::BlockQuote(_)) {
            self.bq_depth += 1;
        }

        let mut more = Vec::new();
        let style = match &tag {
            Tag::Paragraph => None,
            Tag::Heading { level, .. } => {
                let level_usize = (*level as usize).saturating_sub(1).min(5);
                let heading_text = &self.text[range.clone()];
                if let Some(marker_end) = heading_text.find(|c: char| c != '#' && c != ' ') {
                    let marker_range = range.start..range.start + marker_end;
                    more.push(Highlight {
                        style: Some(self.ms.heading_outer[level_usize]),
                        range: marker_range,
                    });
                    None
                } else {
                    Some(self.ms.heading_outer[level_usize])
                }
            }
            Tag::BlockQuote(_) => {
                // Transform the `>` belonging to THIS blockquote level to `│`.
                //
                // For nested blockquotes (`> > inner`), pulldown-cmark emits nested
                // BlockQuote events.  The outer event's range covers all lines,
                // so each line in the outer range has `>` at position 0.  The inner
                // event's range starts mid-line on the first line (after `> `) but
                // at column 0 on subsequent lines.  On those subsequent lines, the
                // outer `>` is included in the inner range, so we must skip it.
                //
                // Strategy: on each line, determine how many `>` characters belong
                // to outer blockquote levels (by checking if the line starts at a
                // real line boundary in the source).  If it does, skip (bq_depth-1)
                // `>`s.  If it starts mid-line (first fragment), skip none.
                let bq_text = &self.text[range.clone()];
                let mut pos = range.start;

                for line in bq_text.split_inclusive('\n') {
                    // Does this fragment start at a source line boundary?
                    let at_line_start =
                        pos == 0 || self.text.as_bytes().get(pos - 1) == Some(&b'\n');
                    // If at a line start, outer levels already have `>`s that
                    // we must skip.  If mid-line (first fragment of range),
                    // the outer `>`s are before the range so skip 0.
                    let skip = if at_line_start { self.bq_depth - 1 } else { 0 };

                    let mut found = 0usize;
                    for (byte_offset, ch) in line.char_indices() {
                        if ch == '>' {
                            if found == skip {
                                let gt_pos = pos + byte_offset;
                                self.buffers.transforms.push(Transform {
                                    range: gt_pos..gt_pos + 1,
                                    to: "│".to_string(),
                                    force: false,
                                });
                                more.push(Highlight {
                                    style: Some(self.ms.blockquote_outer),
                                    range: gt_pos..gt_pos + 1,
                                });
                                break;
                            }
                            found += 1;
                        }
                    }

                    pos += line.len();
                }

                // Return None - we've handled the styling via per-line highlights
                None
            }
            Tag::CodeBlock(code) => {
                // Track the fenced block so its body span can be reported once
                // the fence closes. The body starts just past the opening fence
                // line; an empty-body fence keeps this empty range. Indented
                // code blocks are not fences and report no span.
                self.pending_code_block = match code {
                    CodeBlockKind::Fenced(lang) => {
                        let body_start = self.text[range.start..]
                            .find('\n')
                            .map_or(range.end, |nl| range.start + nl + 1);
                        Some(PendingCodeBlock {
                            info: lang.to_string(),
                            body_range: body_start..body_start,
                            body_text: String::new(),
                            body_seen: false,
                        })
                    }
                    CodeBlockKind::Indented => None,
                };

                // pulldown-cmark reports the code-block range starting at the
                // fence marker (```), excluding any leading indentation on the
                // opening-fence line. That indentation is present whenever the
                // block is indented at the top level or nested inside a list.
                // Extend the hidden `code_outer` highlight back over it so the
                // whole fence line is hidden in pretty mode. Without this, the
                // indentation leaks onto the first rendered code line, and the
                // renderer's fence-start detection (which checks that the byte
                // before the fence is a newline) misfires — mistaking the
                // closing fence for an opening one and emitting a spurious
                // blank line. Only extend when the prefix is pure whitespace so
                // structural prefixes (e.g. a blockquote `> `) are left intact.
                let line_start = self.text[..range.start].rfind('\n').map_or(0, |p| p + 1);
                let fence_start = if self.text[line_start..range.start]
                    .bytes()
                    .all(|b| b == b' ' || b == b'\t')
                {
                    line_start
                } else {
                    range.start
                };
                more.push(Highlight {
                    style: Some(self.ms.code_outer),
                    range: fence_start..range.end,
                });
                match code {
                    CodeBlockKind::Fenced(lang) if !lang.is_empty() => {
                        if let Some(r) =
                            find_substring(&self.text[range.clone()], lang, true, false)
                        {
                            let range = (r.start + range.start)..(r.end + range.start);
                            more.push(Highlight {
                                style: Some(self.ms.code_language),
                                range,
                            });
                        }
                    }
                    _ => (),
                }
                None
            }
            Tag::HtmlBlock => {
                // Don't syntax-highlight HTML blocks as code. In LLM output,
                // these are typically XML-like structural tags (e.g. <example>)
                // from system prompts, not actual HTML. Treating them as code
                // blocks (with background styling) causes visual inconsistency
                // because pulldown-cmark ends HTML blocks at blank lines,
                // making the first part look like code and the rest like text.
                None
            }
            Tag::List(_) => None,
            Tag::Item => {
                let item_text = &self.text[range.clone()];
                let trimmed = item_text.trim_start();
                let leading_ws = item_text.len() - trimmed.len();

                let marker_len = if trimmed.starts_with("- ") || trimmed.starts_with("* ") {
                    2
                } else if let Some(pos) = trimmed.find(". ") {
                    if pos > 0 && trimmed[..pos].chars().all(|c| c.is_ascii_digit()) {
                        pos + 2
                    } else {
                        0
                    }
                } else if let Some(pos) = trimmed.find(") ") {
                    if pos > 0 && trimmed[..pos].chars().all(|c| c.is_ascii_digit()) {
                        pos + 2
                    } else {
                        0
                    }
                } else {
                    0
                };

                if marker_len > 0 {
                    let marker_start = range.start + leading_ws;
                    let marker_end = marker_start + marker_len;
                    more.push(Highlight {
                        style: Some(self.ms.list_item),
                        range: marker_start..marker_end,
                    });

                    if trimmed.starts_with("- ") || trimmed.starts_with("* ") {
                        self.buffers.transforms.push(Transform {
                            range: marker_start..marker_start + 1,
                            to: "•".to_string(),
                            force: false,
                        });
                    }
                }
                None
            }
            Tag::Table(alignments) => {
                self.table_state = Some(TableState::new(alignments.to_vec(), range.start));
                Some(self.ms.table_outer)
            }
            Tag::TableHead => {
                if let Some(ref mut state) = self.table_state {
                    state.in_header = true;
                }
                Some(self.ms.table_outer)
            }
            Tag::TableRow => {
                if let Some(ref mut state) = self.table_state {
                    state.current_row.clear();
                }
                Some(self.ms.table_outer)
            }
            Tag::TableCell => {
                if let Some(ref mut state) = self.table_state {
                    state.current_cell.clear();
                }
                self.push_highlight(None, &range);
                None
            }
            Tag::Emphasis => {
                if let Some(ref mut state) = self.table_state {
                    state.cell_italic = true;
                }
                Some(self.ms.emphasis_outer)
            }
            Tag::Strong => {
                if let Some(ref mut state) = self.table_state {
                    state.cell_bold = true;
                }
                Some(self.ms.strong_outer)
            }
            Tag::Strikethrough => Some(self.ms.strikethrough_outer),
            Tag::Link {
                dest_url, title, ..
            }
            | Tag::Image {
                dest_url, title, ..
            } => {
                // Links inside a table cell go through the table renderer's
                // own hyperlink path (TableHyperlink in TableReplace).  The
                // paragraph link path (LinkTarget + chunk_link_offsets) can't
                // project links onto rendered table cells because the table
                // replace consumes the entire source range — no text chunk
                // ever covers the cell's link text.
                //
                // We still want a stable link `id` so terminal UIs can group
                // wrapped link fragments; assign one from the same counter
                // used by paragraph links and stash it in `cell_link` so the
                // following `Event::Text`s tag their CellSpans with it.
                if let Some(ref mut state) = self.table_state {
                    let id = self.link_id_counter;
                    self.link_id_counter += 1;
                    state.cell_link = Some((dest_url.to_string(), id));
                    self.tag_stack.push(tag);
                    return;
                }

                let tag_str = &self.text[range.clone()];

                if !title.is_empty() {
                    for t in [format!("\"{title}\""), format!("'{title}'")].map(CowStr::from) {
                        if let Some(r) = find_substring(tag_str, &t, true, true) {
                            let title_range = (r.start + range.start)..(r.end + range.start);
                            more.push(Highlight {
                                style: Some(self.ms.link_title),
                                range: title_range,
                            });
                        }
                    }
                }

                // We intentionally use allow_outside=true here (instead of the previous
                // pointer-based allow_outside=false) and then do an rfind on the prefix
                // before the (last) dest_url occurrence. This is required because dest_url
                // may be a CowStr::Owned (after percent-decoding or HTML entity expansion)
                // and therefore may not be a sub-slice of tag_str. The rfind on the strict
                // prefix guarantees we find the *structural* `](` closer even when the link
                // text, title, or the dest literal itself contains the byte sequence `](`.
                let url_rel_opt = find_substring(tag_str, dest_url, true, true);
                if let Some(r) = &url_rel_opt {
                    let url_range = (r.start + range.start)..(r.end + range.start);
                    more.push(Highlight {
                        style: Some(self.ms.link_url),
                        range: url_range,
                    });
                }

                let bracket_pos_opt = url_rel_opt
                    .as_ref()
                    .and_then(|r| tag_str[..r.start].rfind("](").map(|p| p..p + 2));
                if let Some(bracket_pos) = bracket_pos_opt {
                    let open_bracket = if tag_str.starts_with("![") { 1 } else { 0 };
                    if open_bracket > 0 {
                        more.push(Highlight {
                            style: Some(self.ms.link_outer),
                            range: range.start..range.start + 1,
                        });
                    }
                    more.push(Highlight {
                        style: Some(self.ms.link_outer),
                        range: range.start + open_bracket..range.start + open_bracket + 1,
                    });
                    let text_start = range.start + open_bracket + 1;
                    let text_end = bracket_pos.start + range.start;
                    if text_end > text_start {
                        more.push(Highlight {
                            style: Some(self.ms.link_text),
                            range: text_start..text_end,
                        });
                    }
                    let bracket_abs = bracket_pos.start + range.start;
                    more.push(Highlight {
                        style: Some(self.ms.link_outer),
                        range: bracket_abs..bracket_abs + 2,
                    });
                    more.push(Highlight {
                        style: Some(self.ms.link_outer),
                        range: range.end - 1..range.end,
                    });
                    self.buffers.transforms.push(Transform {
                        range: range.start + open_bracket..range.start + open_bracket + 1,
                        to: "".to_string(),
                        force: false,
                    });
                    self.buffers.transforms.push(Transform {
                        range: bracket_abs..bracket_abs + 2,
                        to: " (".to_string(),
                        force: false,
                    });
                    if text_end > text_start {
                        self.buffers.link_targets.push(LinkTarget {
                            source_range: text_start..text_end,
                            url: dest_url.to_string(),
                            id: self.link_id_counter,
                        });
                        self.link_id_counter += 1;
                    }
                    None
                } else {
                    self.buffers.link_targets.push(LinkTarget {
                        source_range: range.clone(),
                        url: dest_url.to_string(),
                        id: self.link_id_counter,
                    });
                    self.link_id_counter += 1;
                    Some(self.ms.link_outer)
                }
            }
            _ => None,
        };

        if let Some(style) = style {
            self.push_highlight(Some(style), &range);
        }
        for hl in more {
            self.buffers.highlights.push(hl);
        }

        self.tag_stack.push(tag);
    }

    fn on_end(&mut self, tag_end: TagEnd, range: Range<usize>) {
        self.tag_stack.pop();

        // Handle tag-specific end logic and determine if we need to push a style
        let style = match &tag_end {
            TagEnd::Emphasis => {
                // Reset italic for table cells (no highlight pushed)
                if let Some(ref mut state) = self.table_state {
                    state.cell_italic = false;
                }
                None
            }
            TagEnd::Strong => {
                // Reset bold for table cells (no highlight pushed)
                if let Some(ref mut state) = self.table_state {
                    state.cell_bold = false;
                }
                None
            }
            TagEnd::Strikethrough => None, // No highlight pushed
            TagEnd::CodeBlock => {
                // pulldown synthesizes a block end at end-of-input even for an
                // unterminated fence, so the end event alone does not prove
                // closure. A closing fence always sits after the body, so the
                // block range extends past the body exactly when the fence
                // closed. `take` clears the pending block in either case.
                if let Some(pending) = self.pending_code_block.take()
                    && pending.body_range.end < range.end
                {
                    self.buffers.code_blocks.push(CodeBlockMeta {
                        info: pending.info,
                        body: pending.body_text,
                        body_source_range: pending.body_range,
                    });
                }
                None
            }
            TagEnd::Link | TagEnd::Image => {
                // Clear link state for table cells so subsequent text in
                // the same cell isn't tagged as part of this link.
                if let Some(ref mut state) = self.table_state {
                    state.cell_link = None;
                }
                None
            }
            TagEnd::TableCell => {
                // Finish current cell
                if let Some(ref mut state) = self.table_state {
                    state
                        .current_row
                        .push(std::mem::take(&mut state.current_cell));
                    // Reset cell styles
                    state.cell_bold = false;
                    state.cell_italic = false;
                    state.cell_code = false;
                    state.cell_link = None;
                }
                None
            }
            TagEnd::TableRow => {
                // Finish body row (if not in header)
                if let Some(ref mut state) = self.table_state
                    && !state.in_header
                {
                    let row = std::mem::take(&mut state.current_row);
                    state.rows.push(row);
                }
                None
            }
            TagEnd::TableHead => {
                // Finish header row
                if let Some(ref mut state) = self.table_state {
                    state.header = std::mem::take(&mut state.current_row);
                    state.in_header = false;
                }
                None
            }
            TagEnd::Table => {
                // Finish table: format and store the replacement
                if let Some(mut state) = self.table_state.take() {
                    state.range.end = range.end;
                    let FormattedTable {
                        lines,
                        styled_lines,
                        line_source_offsets,
                        hyperlinks,
                    } = self.format_table(&state);
                    self.buffers.table_replaces.push(TableReplace {
                        lines,
                        styled_lines,
                        range: state.range,
                        line_source_offsets,
                        hyperlinks,
                    });
                }
                None
            }
            _ => None,
        };

        if let Some(style) = style {
            self.push_highlight(Some(style), &range);
        }

        // Track depth and checkpoints
        match &tag_end {
            TagEnd::BlockQuote(_) | TagEnd::List(_) | TagEnd::Item | TagEnd::Table => {
                self.depth = self.depth.saturating_sub(1);
            }
            _ => {}
        }
        if matches!(&tag_end, TagEnd::BlockQuote(_)) {
            self.bq_depth = self.bq_depth.saturating_sub(1);
        }

        // Record checkpoint at depth=0 block boundaries
        if self.depth == 0 {
            let kind = match &tag_end {
                TagEnd::Paragraph => Some(CheckpointKind::Paragraph),
                TagEnd::Heading(_) => Some(CheckpointKind::Heading),
                TagEnd::CodeBlock => Some(CheckpointKind::CodeBlock),
                TagEnd::BlockQuote(_) => Some(CheckpointKind::BlockQuote),
                TagEnd::List(_) => Some(CheckpointKind::List),
                TagEnd::Table => Some(CheckpointKind::Table),
                TagEnd::HtmlBlock => Some(CheckpointKind::HtmlBlock),
                _ => None,
            };

            if let Some(kind) = kind {
                let has_blank = has_blank_line_after(self.text, range.end);
                let is_code_block = matches!(kind, CheckpointKind::CodeBlock);
                let at_eof = range.end >= self.text.len();
                let code_block_properly_closed = is_code_block && !at_eof;

                if has_blank || code_block_properly_closed {
                    // For code blocks, include one newline to properly close the block.
                    // For other blocks (paragraphs, headings, blockquotes, lists),
                    // DON'T include the trailing newline so that when the next chunk
                    // is added, the blank line separator is re-rendered.
                    let checkpoint_pos = if is_code_block && has_blank {
                        range.end + 1
                    } else {
                        range.end
                    };
                    self.last_checkpoint = Some((kind, checkpoint_pos));
                }
            }
        }
    }

    /// Render a mermaid code block into a [`MermaidReplace`]; `true` if drawn.
    fn try_push_mermaid(&mut self, text: &str, range: &Range<usize>) -> bool {
        let line_style = self.ms.rule.style_into();
        let styles = crate::mermaid::MermaidStyles {
            border: line_style,
            node_text: self.ms.text.style_into(),
            edge: line_style,
            edge_label: self.ms.emphasis_inner.style_into(),
            title: self.ms.strong_inner.style_into(),
        };
        match crate::mermaid::render(text, &styles, self.max_table_width) {
            Some(art) => {
                self.buffers
                    .mermaid_replaces
                    .push(crate::buffers::MermaidReplace {
                        lines: art.plain_lines,
                        styled_lines: art.styled_lines,
                        range: range.clone(),
                    });
                true
            }
            None => false,
        }
    }

    /// Apply inline-code styling to a code/math span: dim the delimiters,
    /// style the content. Shared by `Event::Code` and the inline-math
    /// fallback path.
    fn style_inline_code_span(&mut self, code: &CowStr<'_>, range: &Range<usize>) {
        // Find the actual content range (excluding the delimiters).
        let outer_text = &self.text[range.clone()];
        if let Some(inner_range) = find_substring(outer_text, code, false, false)
            .or_else(|| find_substring(outer_text, code, true, false))
        {
            let absolute_inner = (range.start + inner_range.start)..(range.start + inner_range.end);

            // Left delimiter
            if inner_range.start > 0 {
                self.push_highlight(
                    Some(self.ms.inline_code_outer),
                    &(range.start..absolute_inner.start),
                );
            }
            // Inner code
            self.push_highlight(Some(self.ms.inline_code_inner), &absolute_inner);
            // Right delimiter
            if range.end > absolute_inner.end {
                self.push_highlight(
                    Some(self.ms.inline_code_outer),
                    &(absolute_inner.end..range.end),
                );
            }
        } else {
            self.push_highlight(Some(self.ms.inline_code_inner), range);
        }
    }

    /// Scan a prose `Event::Text` source range for HTML character entity
    /// references (`&lt;`, `&gt;`, `&amp;`, numeric, …) and decode each via a
    /// pretty-mode transform, so e.g. `&lt;` displays as `<`.
    ///
    /// The source-faithful renderer renders the raw source bytes for prose,
    /// which would otherwise leave entities undecoded (table cells already
    /// decode through the cell-text path at `Event::Text` → `push_text`). The
    /// transform is non-`force`, so raw mode still shows the verbatim source.
    ///
    /// A `None`-style highlight is pushed over each entity's byte range so the
    /// renderer splits a chunk exactly there: this keeps the substitution from
    /// straddling a chunk boundary (which would emit the replacement twice)
    /// while leaving the surrounding text/ancestor styling untouched. Code
    /// spans and fenced blocks never reach here, so entities inside code stay
    /// literal.
    ///
    /// Panic-safety: pulldown-cmark guarantees `range` is a valid sub-slice
    /// of `self.text`; even so, the access goes through `str::get` and
    /// `slice::get` so a future invariant violation degrades to a no-op rather
    /// than panicking. The inner loop only advances over ASCII bytes
    /// (`#`/`a-z`/`A-Z`/`0-9`/`;`), guaranteeing `i` and `end` stay on UTF-8
    /// char boundaries.
    fn scan_inline_html_entities(&mut self, range: &Range<usize>) {
        let Some(slice) = self.text.get(range.clone()) else {
            debug_assert!(false, "pulldown-cmark text range out of bounds");
            return;
        };
        if !slice.contains('&') {
            return;
        }
        // Longest HTML5 named entity reference (`&CounterClockwiseContourIntegral;`)
        // is 33 bytes including the leading `&` and trailing `;`. Bounding the
        // scan keeps a run of bare `&` characters from degrading to O(n²).
        const MAX_ENTITY_LEN: usize = 33;
        let bytes = slice.as_bytes();
        let mut i = 0;
        while let Some(&b) = bytes.get(i) {
            if b != b'&' {
                i += 1;
                continue;
            }
            // An entity reference contains only ASCII name/numeric characters
            // and no internal `;`, so the first `;` reached while consuming
            // valid characters closes it. Stopping on any other byte avoids
            // both quadratic scans and slicing through a multi-byte char.
            let max = (i + MAX_ENTITY_LEN).min(bytes.len());
            let mut j = i + 1;
            let end = loop {
                if j >= max {
                    break None;
                }
                match bytes.get(j) {
                    Some(b';') => break Some(j),
                    Some(b'#' | b'a'..=b'z' | b'A'..=b'Z' | b'0'..=b'9') => j += 1,
                    _ => break None,
                }
            };
            if let Some(end) = end
                && let Some(entity) = slice.get(i..=end)
                && let Some(decoded) = decode_html_entity(entity)
            {
                let abs = (range.start + i)..(range.start + end + 1);
                // An earlier scan (e.g. `\(...\)` math) may have already claimed
                // these bytes with its own transform. Overlapping transforms
                // would each emit their replacement, so leave the entity to the
                // existing transform rather than double-substituting.
                let overlaps = self
                    .buffers
                    .transforms
                    .iter()
                    .any(|t| t.range.start < abs.end && abs.start < t.range.end);
                if !overlaps {
                    self.push_highlight(None, &abs);
                    self.buffers.transforms.push(Transform {
                        range: abs,
                        to: decoded,
                        force: false,
                    });
                }
                i = end + 1;
                continue;
            }
            i += 1;
        }
    }

    /// Push a pretty-mode block replacement rendering `latex_src` as display
    /// math over `range`. Returns `false` when conversion declines
    /// (oversized input) or produces nothing visible; callers then fall back
    /// to a raw presentation.
    ///
    /// Reuses the table block-replacement machinery: pre-rendered styled
    /// lines that substitute the source range in pretty mode only, so raw
    /// mode keeps showing the TeX source.
    fn push_display_math_block(&mut self, range: Range<usize>, latex_src: &str) -> bool {
        let Some(rendered) = latex::latex_to_unicode_display(latex_src) else {
            return false;
        };
        if rendered.is_empty() {
            return false;
        }
        // Consume the line ending right after the closing delimiter, like
        // table ranges do. Without this, a batch render emits an extra blank
        // line after the block (the source newline) that the streaming
        // checkpoint+tail path does not, breaking render convergence.
        let mut range = range;
        if self.text[range.end..].starts_with("\r\n") {
            range.end += 2;
        } else if self.text[range.end..].starts_with('\n') {
            range.end += 1;
        }
        let style: ratatui::style::Style = self.ms.math.style_into();
        let src_newlines = self.text[range.clone()]
            .bytes()
            .filter(|&b| b == b'\n')
            .count();
        let mut lines = Vec::with_capacity(rendered.len());
        let mut styled_lines = Vec::with_capacity(rendered.len());
        let mut line_source_offsets = Vec::with_capacity(rendered.len());
        for (i, line) in rendered.iter().enumerate() {
            let text = format!("  {line}");
            styled_lines.push(Line::from(Span::styled(text.clone(), style)));
            lines.push(text);
            // Best-effort scroll mapping: the i-th rendered line maps to the
            // i-th content line of the block (clamped to its source lines).
            line_source_offsets.push((i + 1).min(src_newlines));
        }
        self.buffers.table_replaces.push(TableReplace {
            lines,
            styled_lines,
            range,
            line_source_offsets,
            hyperlinks: Vec::new(),
        });
        true
    }

    /// Format a buffered table into lines with box-drawing borders.
    fn format_table(&self, state: &TableState) -> FormattedTable {
        let borders = TableBorders::BOX;
        let padding = 1;

        // Style already adapted - no need to call adapt_style again
        let border_style: ratatui::style::Style = self.ms.rule.style_into().dim();

        let all_rows: Vec<&Vec<StyledCell>> = std::iter::once(&state.header)
            .chain(state.rows.iter())
            .filter(|r| !r.is_empty())
            .collect();

        if all_rows.is_empty() {
            return FormattedTable::default();
        }

        let num_cols = all_rows.iter().map(|r| r.len()).max().unwrap_or(0);
        if num_cols == 0 {
            return FormattedTable::default();
        }

        let mut col_widths: Vec<usize> = vec![0; num_cols];
        for row in &all_rows {
            for (col_idx, cell) in row.iter().enumerate() {
                let text = cell.plain_text();
                let cell_width = text
                    .split('\n')
                    .map(unicode_display_width)
                    .max()
                    .unwrap_or(0);
                if col_idx < col_widths.len() {
                    col_widths[col_idx] = col_widths[col_idx].max(cell_width);
                }
            }
        }

        // Constrain column widths to fit within max_table_width if set.
        // Table width = 1 (left border) + sum(col_width + 2*padding) + (num_cols-1) separators + 1 (right border)
        //             = 1 + sum(col_width) + num_cols * 2 * padding + (num_cols - 1) + 1
        //             = num_cols * (2 * padding + 1) + sum(col_width) + 2 - 1
        if let Some(max_width) = self.max_table_width {
            let overhead = num_cols * (2 * padding + 1) + 1; // borders + padding
            let content_budget = max_width.saturating_sub(overhead);
            let total_content: usize = col_widths.iter().sum();

            if total_content > content_budget && total_content > 0 {
                // Compute per-column minimum widths: the longest unbreakable
                // word across all cells in each column.  The word separator
                // determines what counts as unbreakable (e.g. "Catherine",
                // "$145,000", "EMP-1001").
                let mut min_col_widths: Vec<usize> = vec![1; num_cols];
                for row in &all_rows {
                    for (col, cell) in row.iter().enumerate() {
                        if col >= num_cols {
                            break;
                        }
                        let text = cell.plain_text();
                        for word in cell_word_separator(&text) {
                            let w = unicode_display_width(word.word);
                            min_col_widths[col] = min_col_widths[col].max(w);
                        }
                    }
                }

                // Start every column at its minimum, then distribute the
                // remaining budget proportionally to how much extra width
                // each column wants (natural − minimum).  This guarantees
                // the total never exceeds the budget while respecting mins.
                let min_total: usize = min_col_widths.iter().sum();
                let extra_budget = content_budget.saturating_sub(min_total);

                // How much each column *wants* above its minimum.
                let extra_wants: Vec<usize> = col_widths
                    .iter()
                    .enumerate()
                    .map(|(i, &w)| w.saturating_sub(min_col_widths[i]))
                    .collect();
                let total_extra_want: usize = extra_wants.iter().sum();

                let mut new_widths = min_col_widths.clone();
                if total_extra_want > 0 && extra_budget > 0 {
                    // Distribute proportionally.
                    for (i, &want) in extra_wants.iter().enumerate() {
                        let share = (want as f64 * extra_budget as f64 / total_extra_want as f64)
                            .floor() as usize;
                        new_widths[i] += share;
                    }

                    // Hand out any remaining columns (from floor rounding)
                    // to columns with the most unmet want, one at a time.
                    let used: usize = new_widths.iter().sum();
                    let mut remaining = content_budget.saturating_sub(used);
                    if remaining > 0 {
                        let mut indices: Vec<usize> = (0..num_cols).collect();
                        // Sort by unmet want descending.
                        indices.sort_by(|&a, &b| {
                            let unmet_a = col_widths[a].saturating_sub(new_widths[a]);
                            let unmet_b = col_widths[b].saturating_sub(new_widths[b]);
                            unmet_b.cmp(&unmet_a)
                        });
                        for &idx in &indices {
                            if remaining == 0 {
                                break;
                            }
                            // Don't grow beyond original natural width.
                            if new_widths[idx] < col_widths[idx] {
                                new_widths[idx] += 1;
                                remaining -= 1;
                            }
                        }
                    }
                }

                col_widths = new_widths;
            }
        }

        let alignments: Vec<_> = (0..num_cols)
            .map(|i| {
                state
                    .alignments
                    .get(i)
                    .copied()
                    .unwrap_or(pulldown_cmark::Alignment::None)
            })
            .collect();

        let mut lines = Vec::new();
        let mut styled_lines = Vec::new();
        let mut line_source_offsets: Vec<usize> = Vec::new();
        let mut hyperlinks: Vec<TableHyperlink> = Vec::new();

        // Source line layout within a table:
        //   offset 0: header row   (| Col A | Col B |)
        //   offset 1: separator    (|-------|-------|)
        //   offset 2+: body rows   (| val1  | val2  |)
        let header_offset = 0usize;
        let separator_offset = 1usize;

        // Top border — belongs to the header line
        let top_border = self.format_border_line(
            &col_widths,
            padding,
            borders.c_tl(),
            borders.t_t(),
            borders.c_tr(),
            borders.h(),
        );
        styled_lines.push(Line::styled(top_border.clone(), border_style));
        lines.push(top_border);
        line_source_offsets.push(header_offset);

        // Header row
        if !state.header.is_empty() {
            let (row_plains, row_styleds, row_links) = self.format_styled_content_lines(
                &state.header,
                &col_widths,
                &alignments,
                padding,
                borders.v(),
                border_style,
                true,
            );
            let base_line = styled_lines.len();
            for (p, s) in row_plains.into_iter().zip(row_styleds) {
                lines.push(p);
                styled_lines.push(s);
                line_source_offsets.push(header_offset);
            }
            for mut link in row_links {
                link.line_offset += base_line;
                hyperlinks.push(link);
            }

            // Header separator
            let sep = self.format_border_line(
                &col_widths,
                padding,
                borders.t_l(),
                borders.x(),
                borders.t_r(),
                borders.h(),
            );
            styled_lines.push(Line::styled(sep.clone(), border_style));
            lines.push(sep);
            line_source_offsets.push(separator_offset);
        }

        // Body rows
        for (i, row) in state.rows.iter().enumerate() {
            let row_offset = separator_offset + 1 + i; // offset 2, 3, ...

            let (row_plains, row_styleds, row_links) = self.format_styled_content_lines(
                row,
                &col_widths,
                &alignments,
                padding,
                borders.v(),
                border_style,
                false,
            );
            let base_line = styled_lines.len();
            for (p, s) in row_plains.into_iter().zip(row_styleds) {
                lines.push(p);
                styled_lines.push(s);
                line_source_offsets.push(row_offset);
            }
            for mut link in row_links {
                link.line_offset += base_line;
                hyperlinks.push(link);
            }

            // Row divider between body rows (not after last row)
            if i < state.rows.len().saturating_sub(1) {
                let row_sep = self.format_border_line(
                    &col_widths,
                    padding,
                    borders.t_l(),
                    borders.x(),
                    borders.t_r(),
                    borders.h(),
                );
                styled_lines.push(Line::styled(row_sep.clone(), border_style));
                lines.push(row_sep);
                line_source_offsets.push(row_offset);
            }
        }

        // Bottom border — belongs to the last body row
        let last_row_offset = separator_offset + state.rows.len();
        let bottom_border = self.format_border_line(
            &col_widths,
            padding,
            borders.c_bl(),
            borders.t_b(),
            borders.c_br(),
            borders.h(),
        );
        styled_lines.push(Line::styled(bottom_border.clone(), border_style));
        lines.push(bottom_border);
        line_source_offsets.push(last_row_offset);

        FormattedTable {
            lines,
            styled_lines,
            line_source_offsets,
            hyperlinks,
        }
    }

    /// Word-wrap a cell's plain text into lines of at most `width` display columns.
    /// Returns a Vec of Strings, one per visual line.
    ///
    /// Delegates to `textwrap::wrap` with a custom word separator that allows
    /// line breaks after spaces, punctuation, and symbol characters — but never
    /// mid-word.  If a single word is wider than `width` it overflows rather
    /// than being chopped.
    fn wrap_cell_text(text: &str, width: usize) -> Vec<String> {
        if width == 0 {
            return vec![String::new()];
        }
        let opts = textwrap::Options::new(width)
            .wrap_algorithm(textwrap::WrapAlgorithm::FirstFit)
            .word_separator(textwrap::WordSeparator::Custom(cell_word_separator))
            .break_words(false);
        let wrapped = textwrap::wrap(text, opts);
        if wrapped.is_empty() {
            vec![String::new()]
        } else {
            wrapped.into_iter().map(|cow| cow.into_owned()).collect()
        }
    }

    /// Format a table row that may span multiple visual lines (when cells wrap).
    ///
    /// Returns `(plain_lines, styled_lines, hyperlinks)` — one plain + styled
    /// entry per visual line, plus any hyperlinks discovered in cell spans.
    /// Hyperlink `line_offset`s are relative to the first visual line of
    /// this row (caller adds the absolute base to embed in the table).
    #[allow(clippy::too_many_arguments)]
    fn format_styled_content_lines(
        &self,
        cells: &[StyledCell],
        col_widths: &[usize],
        alignments: &[pulldown_cmark::Alignment],
        padding: usize,
        v: char,
        border_style: ratatui::style::Style,
        is_header: bool,
    ) -> (Vec<String>, Vec<Line<'static>>, Vec<TableHyperlink>) {
        // 1. Wrap each cell's text into lines constrained to col_widths[i].
        let wrapped_cells: Vec<Vec<String>> = (0..col_widths.len())
            .map(|i| {
                let text = cells.get(i).map(|c| c.plain_text()).unwrap_or_default();
                Self::wrap_cell_text(&text, col_widths[i])
            })
            .collect();

        // 2. Determine the number of visual lines (max wrapped lines across cells).
        let num_visual_lines = wrapped_cells.iter().map(|c| c.len()).max().unwrap_or(1);

        // 3. Build each visual line.
        let mut all_plains = Vec::with_capacity(num_visual_lines);
        let mut all_styled = Vec::with_capacity(num_visual_lines);
        let mut all_links: Vec<TableHyperlink> = Vec::new();

        for vis_line in 0..num_visual_lines {
            let mut plain = String::new();
            let mut spans: Vec<Span<'static>> = Vec::new();
            // Running display column on this visual line; used to record
            // hyperlink column ranges in the table-local coordinate system.
            let mut display_col: usize = 0;

            plain.push(v);
            spans.push(Span::styled(v.to_string(), border_style));
            display_col += unicode_display_width(&v.to_string());

            for (i, width) in col_widths.iter().enumerate() {
                let cell_line_text = wrapped_cells[i]
                    .get(vis_line)
                    .map(|s| s.as_str())
                    .unwrap_or("");
                let cell_line_width = unicode_display_width(cell_line_text);
                let total_padding = width.saturating_sub(cell_line_width);

                let alignment = alignments
                    .get(i)
                    .copied()
                    .unwrap_or(pulldown_cmark::Alignment::None);
                let (left_pad, right_pad) = match alignment {
                    pulldown_cmark::Alignment::Left | pulldown_cmark::Alignment::None => {
                        (0, total_padding)
                    }
                    pulldown_cmark::Alignment::Right => (total_padding, 0),
                    pulldown_cmark::Alignment::Center => {
                        let left = total_padding / 2;
                        (left, total_padding - left)
                    }
                };

                // Left padding
                let left_space = " ".repeat(padding + left_pad);
                let left_space_width = unicode_display_width(&left_space);
                plain.push_str(&left_space);
                spans.push(Span::raw(left_space));
                display_col += left_space_width;

                // Cell text — slice original styled spans to match this
                // visual line's character range, preserving per-span formatting
                // (bold, italic, code, link) across wrap boundaries.
                if !cell_line_text.is_empty() {
                    if let Some(cell) = cells.get(i) {
                        // Find the byte offset of this visual line within the full
                        // cell plain text, then emit styled spans covering that range.
                        let full_text = cell.plain_text();
                        // Sum of all previous visual lines' lengths + spaces between
                        // them gives our start offset in the full plain text.
                        let prev_len: usize = (0..vis_line)
                            .map(|vl| wrapped_cells[i].get(vl).map(|s| s.len()).unwrap_or(0))
                            .sum();
                        // Wrapped-line byte lengths may not land on a char boundary.
                        let prev_len = floor_char_boundary(&full_text, prev_len);
                        // Account for spaces consumed by textwrap between lines
                        let line_start = full_text[prev_len..]
                            .find(cell_line_text)
                            .map(|off| prev_len + off)
                            .unwrap_or(prev_len);
                        let line_end = (line_start + cell_line_text.len()).min(full_text.len());

                        // Walk the cell's spans, emitting the slice that overlaps
                        // [line_start..line_end].
                        let mut offset = 0usize;
                        for cell_span in &cell.spans {
                            let span_start = offset;
                            let span_end = offset + cell_span.text.len();
                            offset = span_end;

                            // Intersect [span_start..span_end] with [line_start..line_end]
                            let start = span_start.max(line_start);
                            let end = span_end.min(line_end);
                            if start >= end {
                                continue;
                            }

                            let slice = &full_text[start..end];
                            if slice.is_empty() {
                                continue;
                            }

                            let mut style: ratatui::style::Style = self.ms.text.style_into();
                            if is_header || cell_span.bold {
                                style = style.bold();
                            }
                            if cell_span.italic {
                                style = style.italic();
                            }
                            if cell_span.code {
                                style = self.ms.inline_code_inner.style_into();
                            }
                            if let Some((url, id)) = &cell_span.link {
                                // Apply link styling additively (preserves
                                // bold/italic if combined).  link_text style
                                // typically adds underline + accent color so
                                // the cell visually matches paragraph link
                                // rendering.
                                let link_style: ratatui::style::Style =
                                    self.ms.link_text.style_into();
                                style = style.patch(link_style);

                                let slice_width = unicode_display_width(slice);
                                all_links.push(TableHyperlink {
                                    line_offset: vis_line,
                                    column_range: display_col..(display_col + slice_width),
                                    url: url.clone(),
                                    id: *id,
                                });
                            }
                            let slice_width = unicode_display_width(slice);
                            plain.push_str(slice);
                            spans.push(Span::styled(slice.to_string(), style));
                            display_col += slice_width;
                        }
                    } else {
                        plain.push_str(cell_line_text);
                        spans.push(Span::raw(cell_line_text.to_string()));
                        display_col += cell_line_width;
                    }
                }

                // Right padding
                let right_space = " ".repeat(right_pad + padding);
                let right_space_width = unicode_display_width(&right_space);
                plain.push_str(&right_space);
                spans.push(Span::raw(right_space));
                display_col += right_space_width;

                // Column separator
                plain.push(v);
                spans.push(Span::styled(v.to_string(), border_style));
                display_col += unicode_display_width(&v.to_string());
            }

            all_plains.push(plain);
            all_styled.push(Line::from(spans));
        }

        (all_plains, all_styled, all_links)
    }

    fn format_border_line(
        &self,
        col_widths: &[usize],
        padding: usize,
        left: char,
        mid: char,
        right: char,
        h: char,
    ) -> String {
        let mut line = String::new();
        line.push(left);
        for (i, &width) in col_widths.iter().enumerate() {
            let total_width = width + padding * 2;
            for _ in 0..total_width {
                line.push(h);
            }
            if i < col_widths.len() - 1 {
                line.push(mid);
            }
        }
        line.push(right);
        line
    }
}

/// Parsed markdown ready for rendering.
///
/// Created by `MarkdownParser::parse()`. Contains the source text, style,
/// and a reference to the populated buffers. Transient parsing state has
/// been dropped at this point.
pub struct ParsedMarkdown<'a, 'b> {
    pub(crate) text: &'a str,
    pub(crate) ms: MarkdownStyle,
    pub(crate) buffers: &'b mut MarkdownBuffers,
    pub(crate) last_checkpoint: Option<(CheckpointKind, usize)>,
    pub(crate) next_link_id: u32,
}

impl<'a, 'b> ParsedMarkdown<'a, 'b> {
    pub fn new(
        text: &'a str,
        ms: MarkdownStyle,
        buffers: &'b mut MarkdownBuffers,
        last_checkpoint: Option<(CheckpointKind, usize)>,
        next_link_id: u32,
    ) -> Self {
        Self {
            text,
            ms,
            buffers,
            last_checkpoint,
            next_link_id,
        }
    }
}