satteri_pulldown_cmark/

parse.rs

// Copyright 2017 Google Inc. All rights reserved.
//
// Permission is hereby granted, free of charge, to any person obtaining a copy
// of this software and associated documentation files (the "Software"), to deal
// in the Software without restriction, including without limitation the rights
// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
// copies of the Software, and to permit persons to whom the Software is
// furnished to do so, subject to the following conditions:
//
// The above copyright notice and this permission notice shall be included in
// all copies or substantial portions of the Software.
//
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
// THE SOFTWARE.

//! Tree-based two pass parser.

use alloc::{borrow::ToOwned, boxed::Box, collections::VecDeque, string::String, vec::Vec};
use core::{
    cell::Cell,
    cmp::{max, min},
    iter::FusedIterator,
    num::NonZeroU32,
    ops::{Index, Range},
};
use rustc_hash::FxHashMap;
use unicase::UniCase;

#[cfg(feature = "mdx")]
use crate::mdx::*;
use crate::{
    Alignment, BlockQuoteKind, CodeBlockKind, DirectiveKind, Event, HeadingLevel, LinkType,
    MetadataBlockKind, Options, Tag, TagEnd,
    firstpass::run_first_pass,
    linklabel::{FootnoteLabel, LinkLabel, ReferenceLabel, scan_link_label_rest},
    scanners::*,
    strings::CowStr,
    tree::{Tree, TreeIndex},
};

// Allowing arbitrary depth nested parentheses inside link destinations
// can create denial of service vulnerabilities if we're not careful.
// The simplest countermeasure is to limit their depth, which is
// explicitly allowed by the spec as long as the limit is at least 3:
// https://spec.commonmark.org/0.29/#link-destination
pub(crate) const LINK_MAX_NESTED_PARENS: usize = 32;

#[derive(Debug, Default, Clone, Copy)]
pub(crate) struct Item {
    pub start: usize,
    pub end: usize,
    pub body: ItemBody,
}

#[derive(Debug, PartialEq, Clone, Copy, Default)]
pub(crate) enum ItemBody {
    // These are possible inline items, need to be resolved in second pass.

    // repeats, can_open, can_close
    MaybeEmphasis(u32, bool, bool),
    /// Head of a run the first pass's escape handler would have swallowed,
    /// sitting on the byte after a deferred autolink candidate's last. Carries
    /// the flags for the run the link firing creates; blocked until it does.
    // repeats, can_open, can_close
    MaybeEmphasisEscaped(u32, bool, bool),
    // preceded_by_backslash, brace context
    MaybeMath(bool, u8),
    // quote byte, can_open, can_close
    MaybeSmartQuote(u8, bool, bool),
    MaybeCode(u32, bool), // number of backticks, preceded by backslash
    MaybeHtml(bool),      // preceded by backslash
    MaybeLinkOpen,
    // bool indicates whether or not the preceding section could be a reference
    MaybeLinkClose(bool),
    MaybeImage,
    /// Zero-width marker where a GFM autolink literal would start. Firing is
    /// decided in `handle_inline_pass1`, where bracket-resolution state is
    /// known; it never survives into `arena_build`.
    MaybeAutolink(AutolinkCandidateIndex),

    // These are inline items after resolution.
    Emphasis,
    Strong,
    Strikethrough,
    Superscript,
    Subscript,
    Math(CowIndex, bool), // true for display math
    Code(CowIndex),
    Link(LinkIndex),
    Image(LinkIndex),
    FootnoteReference(CowIndex),
    TaskListMarker(bool), // true for checked

    // These are also inline items.
    InlineHtml,
    OwnedInlineHtml(CowIndex),
    SynthesizeText(CowIndex),
    SynthesizeChar(char),
    Html,
    Text {
        backslash_escaped: bool,
    },
    SoftBreak,
    // true = is backlash
    HardBreak(bool),

    // Dummy node at the top of the tree - should not be used otherwise!
    #[default]
    Root,

    // These are block items.
    Paragraph,
    TightParagraph,
    Rule,
    Heading(HeadingLevel, Option<HeadingIndex>), // heading level
    FencedCodeBlock(FencedInfoIndex),
    MathBlock(CowIndex), // meta string (info after $$)
    // bool: true = lazy/no-extend (block was opened as a single-line
    // synthetic split, e.g. after an empty list item closed via blank
    // line); arena_build's trailing-indent extension must skip it.
    IndentCodeBlock(bool),
    HtmlBlock(bool), // true = trim trailing newline from value (type 6/7
    // always; type 1-5 only when their closer pattern was found, not when
    // the block ran out of input at EOF)
    BlockQuote(Option<BlockQuoteKind>),
    ContainerDirective(u8, DirectiveIndex), // (fence length, directive data)
    LeafDirective(DirectiveIndex),
    TextDirective(DirectiveIndex),
    // A container directive's `[label]`, holding inline content. Emitted as a
    // `paragraph` with `data.directiveLabel = true`. Its children are tokenized
    // by the normal inline pass, so emphasis/strong/links resolve naturally.
    DirectiveLabel,
    List(bool, u8, u32), // is_tight, list character, list start index
    ListItem(u32, bool), // indent level, spread (loose item)
    FootnoteDefinition(CowIndex),
    MetadataBlock(MetadataBlockKind),

    // Definition lists
    DefinitionList(bool), // is_tight
    // gets turned into either a paragraph or a definition list title,
    // depending on whether there's a definition after it
    MaybeDefinitionListTitle,
    DefinitionListTitle,
    DefinitionListDefinition(u32, bool), // indent level, spread (loose definition)

    // Tables
    Table(AlignmentIndex),
    TableHead,
    TableRow,
    TableCell,

    // MDX
    #[cfg(feature = "mdx")]
    MdxJsxFlowElement(JsxElementIndex),
    #[cfg(feature = "mdx")]
    MdxJsxTextElement(JsxElementIndex),
    #[cfg(feature = "mdx")]
    MdxFlowExpression(CowIndex),
    #[cfg(feature = "mdx")]
    MdxTextExpression(CowIndex),
    #[cfg(feature = "mdx")]
    MdxEsm(CowIndex),
}

impl ItemBody {
    pub(crate) fn is_maybe_inline(&self) -> bool {
        use ItemBody::*;
        matches!(
            *self,
            MaybeEmphasis(..)
                | MaybeEmphasisEscaped(..)
                | MaybeMath(..)
                | MaybeSmartQuote(..)
                | MaybeCode(..)
                | MaybeHtml(..)
                | MaybeLinkOpen
                | MaybeLinkClose(..)
                | MaybeImage
                | MaybeAutolink(..)
        )
    }
    pub(crate) fn is_block_level(&self) -> bool {
        !self.is_inline() && !matches!(self, ItemBody::Root)
    }
    fn is_inline(&self) -> bool {
        use ItemBody::*;
        matches!(
            *self,
            MaybeEmphasis(..)
                | MaybeEmphasisEscaped(..)
                | MaybeMath(..)
                | MaybeSmartQuote(..)
                | MaybeCode(..)
                | MaybeHtml(..)
                | MaybeLinkOpen
                | MaybeLinkClose(..)
                | MaybeImage
                | MaybeAutolink(..)
                | Emphasis
                | Strong
                | Strikethrough
                | Math(..)
                | Code(..)
                | Link(..)
                | Image(..)
                | FootnoteReference(..)
                | TaskListMarker(..)
                | InlineHtml
                | OwnedInlineHtml(..)
                | SynthesizeText(..)
                | SynthesizeChar(..)
                | Html
                | Text { .. }
                | SoftBreak
                | HardBreak(..)
        )
    }
}

#[derive(Debug)]
pub struct BrokenLink<'a> {
    pub span: core::ops::Range<usize>,
    pub link_type: LinkType,
    pub reference: CowStr<'a>,
}

/// Markdown event iterator.
pub struct Parser<'input, CB = DefaultParserCallbacks> {
    callbacks: CB,
    inner: ParserInner<'input>,
}

// Inner state for `Parser`, extracted so that it can remain generic over the callback without
// re-compiling complex logic for each instantiation of the generic type.
pub(crate) struct ParserInner<'input> {
    pub(crate) text: &'input str,
    pub(crate) options: Options,
    pub(crate) tree: Tree<Item>,
    pub(crate) allocs: Allocations<'input>,
    html_scan_guard: HtmlScanGuard,

    // https://github.com/pulldown-cmark/pulldown-cmark/issues/844
    // Consider this example:
    //
    //     [x]: xxx...
    //     [x]
    //     [x]
    //     [x]
    //
    // Which expands to this HTML:
    //
    //     <a href="xxx...">x</a>
    //     <a href="xxx...">x</a>
    //     <a href="xxx...">x</a>
    //
    // This is quadratic growth, because it's filling in the area of a square.
    // To prevent this, track how much it's expanded and limit it.
    link_ref_expansion_limit: usize,

    /// Earliest `(`-title start whose scan hit the block end with no `)`: later starts share that tail.
    unclosed_paren_title_floor: Cell<usize>,

    /// MDX validation errors collected during inline parsing.
    pub(crate) mdx_errors: Vec<(usize, String)>,

    // used by inline passes. store them here for reuse
    inline_stack: InlineStack,
    link_stack: LinkStack,
    wikilink_stack: LinkStack,
    code_delims: CodeDelims,
    math_delims: MathDelims,
}

impl<'input, CB> core::fmt::Debug for Parser<'input, CB> {
    fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result {
        // Only print the fields that have public types.
        f.debug_struct("Parser")
            .field("text", &self.inner.text)
            .field("options", &self.inner.options)
            .field("callbacks", &..)
            .finish()
    }
}

impl<'a> BrokenLink<'a> {
    /// Moves the link into version with a static lifetime.
    ///
    /// The `reference` member is cloned to a Boxed or Inline version.
    pub fn into_static(self) -> BrokenLink<'static> {
        BrokenLink {
            span: self.span.clone(),
            link_type: self.link_type,
            reference: self.reference.into_string().into(),
        }
    }
}

impl<'input> Parser<'input, DefaultParserCallbacks> {
    /// Creates a new event iterator for a markdown string without any options enabled.
    pub fn new(text: &'input str) -> Self {
        Self::new_ext(text, Options::empty())
    }

    /// Creates a new event iterator for a markdown string with given options.
    pub fn new_ext(text: &'input str, options: Options) -> Self {
        Self::new_with_callbacks(text, options, DefaultParserCallbacks)
    }
}

impl<'input, CB: ParserCallbacks<'input>> Parser<'input, CB> {
    /// Creates a new event iterator for markdown text with given options and callbacks.
    ///
    /// ```
    /// # use satteri_pulldown_cmark::{BrokenLink, CowStr, Event, Options, Parser, ParserCallbacks, Tag};
    /// struct CustomCallbacks;
    /// impl<'input> ParserCallbacks<'input> for CustomCallbacks {
    ///     fn handle_broken_link(
    ///         &mut self,
    ///         link: BrokenLink<'input>,
    ///     ) -> Option<(CowStr<'input>, CowStr<'input>)> {
    ///         Some(("https://target".into(), link.reference))
    ///     }
    /// }
    ///
    /// let mut parser =
    ///     Parser::new_with_callbacks("[broken]", Options::empty(), CustomCallbacks);
    ///
    /// assert!(matches!(
    ///     parser.nth(1),
    ///     Some(Event::Start(Tag::Link { .. }))
    /// ));
    /// ```
    ///
    /// See the [`ParserCallbacks`] trait for a list of callbacks that can be overridden.
    pub fn new_with_callbacks(text: &'input str, options: Options, callbacks: CB) -> Self {
        let text = crate::strip_leading_bom(text);
        let (mut tree, allocs, _firstpass_mdx_errors) = run_first_pass(text, options);
        tree.reset();
        let inline_stack = Default::default();
        let link_stack = Default::default();
        let wikilink_stack = Default::default();
        let html_scan_guard = Default::default();
        Parser {
            callbacks,

            inner: ParserInner {
                text,
                options,
                tree,
                allocs,
                inline_stack,
                link_stack,
                wikilink_stack,
                html_scan_guard,
                // always allow 100KiB
                link_ref_expansion_limit: text.len().max(100_000),
                unclosed_paren_title_floor: Cell::new(usize::MAX),
                mdx_errors: Vec::new(),
                code_delims: CodeDelims::new(),
                math_delims: MathDelims::new(),
            },
        }
    }

    /// Returns a reference to the internal `RefDefs` object, which provides access
    /// to the internal map of reference definitions.
    pub fn reference_definitions(&self) -> &RefDefs<'_> {
        &self.inner.allocs.refdefs
    }

    /// Returns MDX validation errors collected during parsing.
    /// Only populated when [`Options::ENABLE_MDX`] is active.
    pub fn mdx_errors(&self) -> &[(usize, String)] {
        &self.inner.mdx_errors
    }

    /// Consumes the event iterator and produces an iterator that produces
    /// `(Event, Range)` pairs, where the `Range` value maps to the corresponding
    /// range in the markdown source.
    pub fn into_offset_iter(self) -> OffsetIter<'input, CB> {
        OffsetIter { parser: self }
    }
}

impl<'input, F> Parser<'input, BrokenLinkCallback<F>> {
    /// In case the parser encounters any potential links that have a broken
    /// reference (e.g `[foo]` when there is no `[foo]: ` entry at the bottom)
    /// the provided callback will be called with the reference name,
    /// and the returned pair will be used as the link URL and title if it is not
    /// `None`.
    ///
    /// This constructor is provided for backwards compatibility.
    /// This and other callbacks can also be customized with [`Parser::new_with_callbacks`].
    pub fn new_with_broken_link_callback(
        text: &'input str,
        options: Options,
        broken_link_callback: Option<F>,
    ) -> Self
    where
        F: FnMut(BrokenLink<'input>) -> Option<(CowStr<'input>, CowStr<'input>)>,
    {
        Self::new_with_callbacks(text, options, BrokenLinkCallback(broken_link_callback))
    }
}

impl<'input> ParserInner<'input> {
    pub(crate) fn new(text: &'input str, options: Options) -> Self {
        let (mut tree, allocs, firstpass_mdx_errors) = run_first_pass(text, options);
        tree.reset();
        ParserInner {
            text,
            options,
            tree,
            allocs,
            inline_stack: Default::default(),
            link_stack: Default::default(),
            wikilink_stack: Default::default(),
            html_scan_guard: Default::default(),
            link_ref_expansion_limit: text.len().max(100_000),
            unclosed_paren_title_floor: Cell::new(usize::MAX),
            mdx_errors: firstpass_mdx_errors,
            code_delims: CodeDelims::new(),
            math_delims: MathDelims::new(),
        }
    }

    /// Use a link label to fetch a type, url, and title.
    ///
    /// This function enforces the [`link_ref_expansion_limit`].
    /// If it returns Some, it also consumes some of the fuel.
    /// If we're out of fuel, it immediately returns None.
    ///
    /// The URL and title are found in the [`RefDefs`] map.
    /// If they're not there, and a callback was provided by the user,
    /// `handle_broken_link` will be invoked and given the opportunity
    /// to provide a fallback.
    ///
    /// The link type (that's "link" or "image") depends on the usage site, and
    /// is provided by the caller of this function.
    /// This function returns a new one because, if it has to invoke a callback
    /// to find the information, the link type is [mapped to an unknown type].
    ///
    /// [mapped to an unknown type]: crate::LinkType::to_unknown
    /// [`link_ref_expansion_limit`]: Self::link_ref_expansion_limit
    fn fetch_link_type_url_title(
        &mut self,
        link_label: CowStr<'input>,
        span: Range<usize>,
        link_type: LinkType,
        callbacks: &mut dyn ParserCallbacks<'input>,
    ) -> Option<(LinkType, CowStr<'input>, CowStr<'input>)> {
        if self.link_ref_expansion_limit == 0 {
            return None;
        }

        let (link_type, url, title) = self
            .allocs
            .refdefs
            .get(link_label.as_ref())
            .map(|matching_def| {
                // found a matching definition!
                let title = matching_def
                    .title
                    .as_ref()
                    .cloned()
                    .unwrap_or_else(|| "".into());
                let url = matching_def.dest.clone();
                (link_type, url, title)
            })
            .or_else(|| {
                // Construct a BrokenLink struct, which will be passed to the callback
                let broken_link = BrokenLink {
                    span,
                    link_type,
                    reference: link_label,
                };

                callbacks
                    .handle_broken_link(broken_link)
                    .map(|(url, title)| (link_type.to_unknown(), url, title))
            })?;

        // Limit expansion from link references.
        // This isn't a problem for footnotes, because multiple references to the same one
        // reuse the same node, but links/images get their HREF/SRC copied.
        self.link_ref_expansion_limit = self
            .link_ref_expansion_limit
            .saturating_sub(url.len() + title.len());

        Some((link_type, url, title))
    }

    /// Handle inline markup.
    ///
    /// When the parser encounters any item indicating potential inline markup, all
    /// inline markup passes are run on the remainder of the chain.
    ///
    /// Note: there's some potential for optimization here, but that's future work.
    pub(crate) fn handle_inline(&mut self, callbacks: &mut dyn ParserCallbacks<'input>) {
        self.handle_inline_pass1(callbacks);
        // Resolve attention (emphasis/strong) and strikethrough/sub/sup. Two
        // delimiter families that can cross, so the resolve order matters,
        // mirroring micromark:
        //   * At the top level, the family whose marker is *tokenized first*
        //     resolves first (`resolveAllConstructs` registration order):
        //     `*~bar~*` → emphasis first; `~_~:_<` → strikethrough first.
        //     An inert leading marker still counts (it still tokenizes), so
        //     `_ ~a*b~*` resolves emphasis first even though `_` can't pair.
        //   * Inside any already-formed span (a link/image label, or the
        //     content between matched delimiters), micromark uses a fixed
        //     `insideSpan.null = [strikethrough, attention]`, i.e. *always
        //     strikethrough first*, independent of the top-level order. So
        //     `[_~a*b~*](/x)` resolves strikethrough first in the label even
        //     though `_ ~a*b~*` resolves emphasis first at the top level.
        let st_enabled = self.options.contains(Options::ENABLE_STRIKETHROUGH)
            || self.options.contains(Options::ENABLE_SUBSCRIPT)
            || self.options.contains(Options::ENABLE_SUPERSCRIPT);
        if !st_enabled {
            self.handle_emphasis_pass();
            return;
        }
        // The top-level order decision must see the whole inline scope from its
        // first child (which includes leading inert markers), not from `cur()`:
        // pass1 leaves `cur` at the first unresolved marker, past leading text
        // an inert marker might hide behind.
        let scope_first = self
            .tree
            .peek_up()
            .and_then(|p| self.tree[p].child)
            .or_else(|| self.tree.cur());
        let strikethrough_first = matches!(
            self.first_inline_marker_char(scope_first),
            Some(b'~') | Some(b'^')
        );
        self.resolve_inline_scope(self.tree.cur(), strikethrough_first);
    }

    /// Resolve emphasis and strikethrough/sub/sup at this scope in the given
    /// order, then descend into each formed span. Nested content is always
    /// resolved strikethrough-first (micromark's `insideSpan.null`).
    fn resolve_inline_scope(&mut self, start: Option<TreeIndex>, strikethrough_first: bool) {
        if strikethrough_first {
            self.resolve_tildes_carets_in_scope(start, false);
            self.handle_emphasis_in_scope(start);
        } else {
            self.handle_emphasis_in_scope(start);
            self.resolve_tildes_carets_in_scope(start, false);
        }
        let mut cur = start;
        while let Some(cur_ix) = cur {
            let next = self.tree[cur_ix].next;
            if matches!(
                self.tree[cur_ix].item.body,
                ItemBody::Emphasis
                    | ItemBody::Strong
                    | ItemBody::Strikethrough
                    | ItemBody::Subscript
                    | ItemBody::Superscript
                    | ItemBody::Link(_)
                    | ItemBody::Image(_)
            ) {
                let child = self.tree[cur_ix].child;
                if self.scope_has_unresolved(child) {
                    self.resolve_inline_scope(child, true);
                }
            }
            cur = next;
        }
    }

    /// An all-`Text` scope is inert: resolvers and descent only touch other bodies.
    #[inline]
    fn scope_has_unresolved(&self, start: Option<TreeIndex>) -> bool {
        let mut cur = start;
        while let Some(cur_ix) = cur {
            if !matches!(self.tree[cur_ix].item.body, ItemBody::Text { .. }) {
                return true;
            }
            cur = self.tree[cur_ix].next;
        }
        false
    }

    /// Find the first emphasis/strikethrough marker (`*` `_` `~` `^`) in
    /// `start..`, in source order. Used to pick the resolve order between the
    /// two families: micromark registers a family's `resolveAll` when its
    /// marker is *tokenized*, so an inert marker (one that can neither open
    /// nor close, e.g. `_` before a space) still counts even though satteri
    /// leaves it as plain text rather than a `MaybeEmphasis`. Scan text nodes
    /// too, not just `MaybeEmphasis`, or `_ ~a*b~*` would wrongly resolve the
    /// `~` first (the inert leading `_` is the real first marker).
    fn first_inline_marker_char(&self, start: Option<TreeIndex>) -> Option<u8> {
        // Only count markers for *enabled* families, matching which delimiters
        // the reference actually tokenizes. `~` counts only with strikethrough
        // or subscript on; `^` only with superscript on. Otherwise a literal
        // `^` (or `~`) would wrongly skew the emphasis-vs-strikethrough order.
        let tilde = self.options.contains(Options::ENABLE_STRIKETHROUGH)
            || self.options.contains(Options::ENABLE_SUBSCRIPT);
        let caret = self.options.contains(Options::ENABLE_SUPERSCRIPT);
        let is_marker =
            |c: u8| matches!(c, b'*' | b'_') || (c == b'~' && tilde) || (c == b'^' && caret);
        let bytes = self.text.as_bytes();
        let mut cur = start;
        while let Some(cur_ix) = cur {
            match self.tree[cur_ix].item.body {
                ItemBody::MaybeEmphasis(..) => {
                    let c = bytes[self.tree[cur_ix].item.start];
                    if is_marker(c) {
                        return Some(c);
                    }
                }
                ItemBody::Text { backslash_escaped } => {
                    let item = &self.tree[cur_ix].item;
                    // A backslash-escaped leading byte (the `\` was stripped)
                    // is not a marker micromark would tokenize.
                    let from = item.start + usize::from(backslash_escaped);
                    if let Some(off) = bytes[from..item.end].iter().position(|&c| is_marker(c)) {
                        return Some(bytes[from + off]);
                    }
                }
                _ => {}
            }
            cur = self.tree[cur_ix].next;
        }
        None
    }

    /// Recursive emphasis pass. Processes `*`/`_` MaybeEmphasis at this
    /// scope, then descends into any inline containers (Emphasis,
    /// Strong, Strikethrough, Link, Image, etc.) to do the same in
    /// their children.
    fn handle_emphasis_pass(&mut self) {
        let start = self.tree.cur();
        self.resolve_emphasis_recursive(start);
    }

    fn resolve_emphasis_recursive(&mut self, start: Option<TreeIndex>) {
        self.handle_emphasis_in_scope(start);

        let mut cur = start;
        while let Some(cur_ix) = cur {
            let next = self.tree[cur_ix].next;
            match self.tree[cur_ix].item.body {
                ItemBody::Emphasis
                | ItemBody::Strong
                | ItemBody::Strikethrough
                | ItemBody::Subscript
                | ItemBody::Superscript
                | ItemBody::Link(_)
                | ItemBody::Image(_) => {
                    let child = self.tree[cur_ix].child;
                    if self.scope_has_unresolved(child) {
                        self.resolve_emphasis_recursive(child);
                    }
                }
                _ => {}
            }
            cur = next;
        }
    }

    /// Handle inline HTML, code spans, and links.
    ///
    /// This function handles both inline HTML and code spans, because they have
    /// the same precedence. It also handles links, even though they have lower
    /// precedence, because the URL of links must not be processed.
    fn handle_inline_pass1(&mut self, callbacks: &mut dyn ParserCallbacks<'input>) {
        let mut cur = self.tree.cur();
        let mut prev = None;

        let block_end = self.tree[self.tree.peek_up().unwrap()].item.end;
        let block_text = &self.text[..block_end];
        self.unclosed_paren_title_floor.set(usize::MAX);

        while let Some(mut cur_ix) = cur {
            match self.tree[cur_ix].item.body {
                ItemBody::MaybeHtml(preceded_by_backslash) => {
                    if preceded_by_backslash {
                        // No autolink claimed the `\`, so the `<` is literal.
                        self.tree[cur_ix].item.body = ItemBody::Text {
                            backslash_escaped: true,
                        };
                        prev = cur;
                        cur = self.tree[cur_ix].next;
                        continue;
                    }
                    // MDX inline JSX: check before HTML
                    #[cfg(feature = "mdx")]
                    if self.options.contains(Options::ENABLE_MDX) {
                        let start = self.tree[cur_ix].item.start;
                        let next_byte = block_text.as_bytes().get(start + 1).copied();

                        // In MDX, `<!` is not valid (no HTML comments).
                        if next_byte == Some(b'!') {
                            self.mdx_errors.push((
                                start,
                                "Unexpected character `!` (U+0021) before name, expected a \
                                 character that can start a name, such as a letter, `$`, or `_` \
                                 (note: to create a comment in MDX, use `{/* text */}`)"
                                    .to_string(),
                            ));
                            self.tree[cur_ix].item.body = ItemBody::Text {
                                backslash_escaped: false,
                            };
                            prev = cur;
                            cur = self.tree[cur_ix].next;
                            continue;
                        }

                        if let Some(total_len) =
                            scan_mdx_inline_jsx(&block_text.as_bytes()[start..])
                        {
                            let end = start + total_len;
                            let node = scan_nodes_to_ix(&self.tree, self.tree[cur_ix].next, end);
                            let raw = &block_text[start..end];
                            let col = crate::mdx::column_at(block_text.as_bytes(), start);
                            let jsx_data = crate::mdx::parse_jsx_tag_with_column(raw, col, 0);
                            let mut allocator = oxc_allocator::Allocator::default();
                            crate::mdx::validate_jsx_expressions(
                                raw,
                                &jsx_data.attrs,
                                |rel| start + rel,
                                &mut allocator,
                                &mut self.mdx_errors,
                            );
                            let jsx_ix = self.allocs.allocate_jsx_element(jsx_data);
                            self.tree[cur_ix].item.body = ItemBody::MdxJsxTextElement(jsx_ix);
                            self.tree[cur_ix].item.end = end;
                            self.tree[cur_ix].next = node;
                            prev = cur;
                            cur = node;
                            if let Some(node_ix) = cur {
                                self.tree[node_ix].item.start =
                                    max(self.tree[node_ix].item.start, end);
                            }
                            continue;
                        }

                        // mdx-js fallback rule:
                        //   `<` + space/tab → always literal `<` (text).
                        //   `<` + newline   → JSX tag may span lines; treat
                        //                      as text only if the next
                        //                      non-whitespace byte is benign
                        //                      (not `>`, not EOF/blank-line)
                        //                      AND the line containing it
                        //                      isn't a setext underline
                        //                      (`-`+ or `=`+), which would
                        //                      promote the `<` into a heading
                        //                      whose JSX validation fails.
                        //   `<` + anything else (incl. EOF) → parse error
                        //                      (`<\`, `<,`, `<{`, `<<`, `<.`,
                        //                       …).
                        let bytes_block = block_text.as_bytes();
                        let is_text_fallback = match next_byte {
                            Some(b' ' | b'\t') => true,
                            Some(b'\n' | b'\r') => {
                                // Skip whitespace + container prefixes when
                                // probing for the first significant byte
                                // after `\n`. A `>` at line start inside a
                                // blockquote is the container marker, not a
                                // JSX-like delimiter.
                                let bq_depth = self
                                    .tree
                                    .walk_spine()
                                    .filter(|&&ix| {
                                        matches!(self.tree[ix].item.body, ItemBody::BlockQuote(..))
                                    })
                                    .count();
                                let mut probe = start + 1;
                                loop {
                                    while probe < bytes_block.len()
                                        && matches!(
                                            bytes_block[probe],
                                            b' ' | b'\t' | b'\n' | b'\r'
                                        )
                                    {
                                        probe += 1;
                                    }
                                    if bq_depth == 0
                                        || probe >= bytes_block.len()
                                        || bytes_block[probe] != b'>'
                                    {
                                        break;
                                    }
                                    let mut consumed = 0;
                                    while consumed < bq_depth
                                        && probe < bytes_block.len()
                                        && bytes_block[probe] == b'>'
                                    {
                                        probe += 1;
                                        if probe < bytes_block.len() && bytes_block[probe] == b' ' {
                                            probe += 1;
                                        }
                                        consumed += 1;
                                    }
                                }
                                if probe >= bytes_block.len() || bytes_block[probe] == b'>' {
                                    false
                                } else {
                                    // Reject if `probe`'s line is a setext
                                    // underline (only `-` or only `=`, then
                                    // optional whitespace to EOL/EOF) AND
                                    // would actually promote the `<`-line
                                    // to a heading. Inside a blockquote
                                    // container the underline line is
                                    // typically a lazy continuation (no
                                    // `>` prefix) and doesn't promote, so
                                    // skip the rejection.
                                    let underline_char = bytes_block[probe];
                                    if !matches!(underline_char, b'-' | b'=') {
                                        true
                                    } else {
                                        let mut q = probe;
                                        while q < bytes_block.len()
                                            && bytes_block[q] == underline_char
                                        {
                                            q += 1;
                                        }
                                        while q < bytes_block.len()
                                            && matches!(bytes_block[q], b' ' | b'\t')
                                        {
                                            q += 1;
                                        }
                                        let at_eol = q >= bytes_block.len()
                                            || matches!(bytes_block[q], b'\n' | b'\r');
                                        if !at_eol {
                                            true
                                        } else {
                                            // Container check: a blockquote
                                            // `>` (possibly after up to 3
                                            // spaces) on the line opening
                                            // the `<` means the underline
                                            // line would need the same
                                            // prefix to actually promote a
                                            // setext heading. Without it,
                                            // the underline is lazy
                                            // paragraph continuation, so
                                            // accept as text.
                                            //
                                            // Same for listitems: if the
                                            // spine has a ListItem and the
                                            // underline line starts at a
                                            // column less than the listitem
                                            // content column, it's lazy
                                            // continuation and doesn't
                                            // promote — accept as text.
                                            let mut ls = start;
                                            while ls > 0
                                                && !matches!(bytes_block[ls - 1], b'\n' | b'\r')
                                            {
                                                ls -= 1;
                                            }
                                            let mut k = ls;
                                            let mut sp = 0;
                                            while k < start && bytes_block[k] == b' ' && sp < 3 {
                                                k += 1;
                                                sp += 1;
                                            }
                                            if k < start && bytes_block[k] == b'>' {
                                                true
                                            } else {
                                                // Underline line start.
                                                let mut us = probe;
                                                while us > 0
                                                    && !matches!(bytes_block[us - 1], b'\n' | b'\r')
                                                {
                                                    us -= 1;
                                                }
                                                let mut underline_col = 0;
                                                let mut uk = us;
                                                while uk < probe && bytes_block[uk] == b' ' {
                                                    uk += 1;
                                                    underline_col += 1;
                                                }
                                                let listitem_indent = self
                                                    .tree
                                                    .walk_spine()
                                                    .filter_map(|&ix| {
                                                        match self.tree[ix].item.body {
                                                            ItemBody::ListItem(indent, _) => {
                                                                Some(indent as usize)
                                                            }
                                                            _ => None,
                                                        }
                                                    })
                                                    .next();
                                                let in_blockquote =
                                                    self.tree.walk_spine().any(|&ix| {
                                                        matches!(
                                                            self.tree[ix].item.body,
                                                            ItemBody::BlockQuote(..)
                                                        )
                                                    });
                                                // BlockQuote container: an
                                                // underline line missing the
                                                // `>` prefix is lazy
                                                // continuation and doesn't
                                                // promote. Detect by checking
                                                // the underline line's source
                                                // (not block_text, which has
                                                // already stripped the
                                                // prefix).
                                                let bq_lazy = if in_blockquote {
                                                    underline_col < 1
                                                        || !bytes_block[us..probe].contains(&b'>')
                                                } else {
                                                    false
                                                };
                                                matches!(listitem_indent, Some(i) if underline_col < i)
                                                    || bq_lazy
                                            }
                                        }
                                    }
                                }
                            }
                            _ => false,
                        };
                        if !is_text_fallback {
                            self.mdx_errors.push((
                                start,
                                "Unexpected character after `<`, expected a valid JSX tag \
                                 (note: to create a link in MDX, use `[text](url)`)"
                                    .to_string(),
                            ));
                        }

                        self.tree[cur_ix].item.body = ItemBody::Text {
                            backslash_escaped: false,
                        };
                        prev = cur;
                        cur = self.tree[cur_ix].next;
                        continue;
                    }

                    let next = self.tree[cur_ix].next;
                    let autolink = if let Some(next_ix) = next {
                        scan_autolink(block_text, self.tree[next_ix].item.start)
                    } else {
                        None
                    };

                    if let Some((ix, uri, link_type)) = autolink {
                        let node = scan_nodes_to_ix(&self.tree, next, ix);
                        let text_node = self.tree.create_node(Item {
                            start: self.tree[cur_ix].item.start + 1,
                            end: ix - 1,
                            body: ItemBody::Text {
                                backslash_escaped: false,
                            },
                        });
                        let link_ix =
                            self.allocs
                                .allocate_link(link_type, uri, "".into(), "".into());
                        self.tree[cur_ix].item.body = ItemBody::Link(link_ix);
                        self.tree[cur_ix].item.end = ix;
                        self.tree[cur_ix].next = node;
                        self.tree[cur_ix].child = Some(text_node);
                        prev = cur;
                        cur = node;
                        if let Some(node_ix) = cur {
                            let orig_start = self.tree[node_ix].item.start;
                            let new_start = max(orig_start, ix);
                            self.tree[node_ix].item.start = new_start;
                            // When the autolink's closing `>` consumed the byte
                            // that was the target of a preceding `\` escape,
                            // the trailing text's `backslash_escaped` flag is
                            // stale — clear it so arena_build doesn't extend
                            // the text node's source span back over bytes the
                            // link now owns. Mirrors the inline-link fix.
                            if new_start > orig_start
                                && let ItemBody::Text { backslash_escaped } =
                                    &mut self.tree[node_ix].item.body
                            {
                                *backslash_escaped = false;
                            }
                        }
                        continue;
                    } else {
                        let inline_html = next.and_then(|next_ix| {
                            self.scan_inline_html(
                                block_text.as_bytes(),
                                self.tree[next_ix].item.start,
                            )
                        });
                        if let Some((span, ix)) = inline_html {
                            let node = scan_nodes_to_ix(&self.tree, next, ix);
                            self.tree[cur_ix].item.body = if !span.is_empty() {
                                let converted_string =
                                    String::from_utf8(span).expect("invalid utf8");
                                ItemBody::OwnedInlineHtml(
                                    self.allocs.allocate_cow(converted_string.into()),
                                )
                            } else {
                                ItemBody::InlineHtml
                            };
                            self.tree[cur_ix].item.end = ix;
                            self.tree[cur_ix].next = node;
                            prev = cur;
                            cur = node;
                            if let Some(node_ix) = cur {
                                let orig_start = self.tree[node_ix].item.start;
                                let new_start = max(orig_start, ix);
                                self.tree[node_ix].item.start = new_start;
                                // Inline HTML may consume bytes that a `\X`
                                // escape was attached to (e.g. `\*` inside
                                // an attribute value). Clear the stale flag
                                // so arena_build doesn't extend the trail
                                // back over bytes the HTML now owns.
                                if new_start > orig_start
                                    && let ItemBody::Text { backslash_escaped } =
                                        &mut self.tree[node_ix].item.body
                                {
                                    *backslash_escaped = false;
                                }
                            }
                            continue;
                        }
                    }
                    self.tree[cur_ix].item.body = ItemBody::Text {
                        backslash_escaped: false,
                    };
                }
                ItemBody::MaybeMath(preceded_by_backslash, _brace_context) => {
                    if preceded_by_backslash {
                        self.tree[cur_ix].item.body = ItemBody::Text {
                            backslash_escaped: true,
                        };
                        prev = cur;
                        cur = self.tree[cur_ix].next;
                        continue;
                    }
                    // Count consecutive $ from the opening position
                    let mut open_count = 1usize;
                    let mut open_end = cur_ix;
                    {
                        let mut peek = self.tree[cur_ix].next;
                        while let Some(peek_ix) = peek {
                            if matches!(self.tree[peek_ix].item.body, ItemBody::MaybeMath(..))
                                && self.tree[peek_ix].item.start == self.tree[open_end].item.end
                            {
                                open_count += 1;
                                open_end = peek_ix;
                                peek = self.tree[peek_ix].next;
                            } else {
                                break;
                            }
                        }
                    }

                    // Single- and multi-dollar math can be toggled
                    // independently (mirroring remark-math's
                    // `singleDollarTextMath`). When this run's length isn't
                    // an enabled delimiter, the `$` is literal text — so
                    // prose like `$50 to $100` never becomes a math span.
                    let count_enabled = if open_count == 1 {
                        self.options.contains(Options::ENABLE_MATH_SINGLE_DOLLAR)
                    } else {
                        self.options.contains(Options::ENABLE_MATH_MULTI_DOLLAR)
                    };
                    if !count_enabled {
                        let mut text_ix = cur_ix;
                        loop {
                            self.tree[text_ix].item.body = ItemBody::Text {
                                backslash_escaped: false,
                            };
                            if text_ix == open_end {
                                break;
                            }
                            match self.tree[text_ix].next {
                                Some(next) => text_ix = next,
                                None => break,
                            }
                        }
                        prev = cur;
                        cur = self.tree[cur_ix].next;
                        continue;
                    }

                    // Scan forward for a matching run of the same count
                    let mut scan = self.tree[open_end].next;
                    let mut close_ix = None;
                    while let Some(scan_ix) = scan {
                        if matches!(self.tree[scan_ix].item.body, ItemBody::MaybeMath(..)) {
                            let mut run = 1usize;
                            let mut run_end = scan_ix;
                            let mut peek = self.tree[scan_ix].next;
                            while let Some(peek_ix) = peek {
                                if matches!(self.tree[peek_ix].item.body, ItemBody::MaybeMath(..))
                                    && self.tree[peek_ix].item.start == self.tree[run_end].item.end
                                {
                                    run += 1;
                                    run_end = peek_ix;
                                    peek = self.tree[peek_ix].next;
                                } else {
                                    break;
                                }
                            }
                            if run == open_count {
                                close_ix = Some(scan_ix);
                                break;
                            }
                            // Skip past this non-matching run
                            scan = self.tree[run_end].next;
                            continue;
                        }
                        scan = self.tree[scan_ix].next;
                    }

                    if let Some(scan_ix) = close_ix {
                        self.make_math_span(cur_ix, scan_ix);
                    } else {
                        let mut fail_ix = cur_ix;
                        loop {
                            self.tree[fail_ix].item.body = ItemBody::Text {
                                backslash_escaped: false,
                            };
                            if fail_ix == open_end {
                                break;
                            }
                            if let Some(next) = self.tree[fail_ix].next {
                                fail_ix = next;
                            } else {
                                break;
                            }
                        }
                    }
                }
                ItemBody::MaybeCode(search_count, preceded_by_backslash) => {
                    let mut search_count = search_count as usize;
                    if preceded_by_backslash {
                        search_count -= 1;
                        if search_count == 0 {
                            self.tree[cur_ix].item.body = ItemBody::Text {
                                backslash_escaped: true,
                            };
                            prev = cur;
                            cur = self.tree[cur_ix].next;
                            continue;
                        }
                    }

                    if self.code_delims.is_populated() {
                        // we have previously scanned all codeblock delimiters,
                        // so we can reuse that work
                        if let Some(scan_ix) = self.code_delims.find(cur_ix, search_count) {
                            self.make_code_span(cur_ix, scan_ix, preceded_by_backslash);
                        } else {
                            self.tree[cur_ix].item.body = ItemBody::Text {
                                backslash_escaped: preceded_by_backslash,
                            };
                        }
                    } else {
                        // we haven't previously scanned all codeblock delimiters,
                        // so walk the AST
                        let mut scan = if search_count > 0 {
                            self.tree[cur_ix].next
                        } else {
                            None
                        };
                        while let Some(scan_ix) = scan {
                            if let ItemBody::MaybeCode(delim_count, _) =
                                self.tree[scan_ix].item.body
                            {
                                let delim_count = delim_count as usize;
                                if search_count == delim_count {
                                    self.make_code_span(cur_ix, scan_ix, preceded_by_backslash);
                                    self.code_delims.clear();
                                    break;
                                } else {
                                    self.code_delims.insert(delim_count, scan_ix);
                                }
                            }
                            scan = self.tree[scan_ix].next;
                        }
                        if scan.is_none() {
                            self.tree[cur_ix].item.body = ItemBody::Text {
                                backslash_escaped: preceded_by_backslash,
                            };
                        }
                    }
                }
                ItemBody::MaybeAutolink(cand_ix) => {
                    // An unresolved bracket opener blocks the construct, and
                    // the stack holds exactly those.
                    let next = self.tree[cur_ix].next;
                    if !self.link_stack.is_empty() {
                        // Zero-width `Text` rather than an unlink: the emphasis
                        // resolver addresses nodes by arena index, so dropping
                        // one from the chain would hand it whatever came next.
                        self.tree[cur_ix].item.body = ItemBody::Text {
                            backslash_escaped: false,
                        };
                        prev = cur;
                        cur = next;
                        continue;
                    }
                    // Reusing the marker node as the `Link` keeps the preceding
                    // sibling's `next` pointer valid.
                    let cand = self.allocs[cand_ix];
                    let node_after = scan_nodes_to_ix(&self.tree, next, cand.end);
                    let text_child = self.tree.create_node(Item {
                        start: cand.start,
                        end: cand.end,
                        body: ItemBody::Text {
                            backslash_escaped: false,
                        },
                    });
                    self.tree[cur_ix].item = Item {
                        start: cand.start,
                        end: cand.end,
                        body: ItemBody::Link(cand.link),
                    };
                    self.tree[cur_ix].child = Some(text_child);
                    self.tree[cur_ix].next = node_after;
                    if let Some(node_after_ix) = node_after {
                        let orig_start = self.tree[node_after_ix].item.start;
                        let new_start = max(orig_start, cand.end);
                        // A `\` the first pass read as a hard-break marker
                        // turns out to be the URL's last byte.
                        if orig_start < cand.end
                            && matches!(
                                self.tree[node_after_ix].item.body,
                                ItemBody::HardBreak(true)
                            )
                        {
                            self.tree[node_after_ix].item.body = ItemBody::SoftBreak;
                        }
                        // The clamp below can't trim an item that carries its
                        // own content: a character reference the URL ends
                        // inside would still emit its decoded value.
                        if orig_start < cand.end
                            && matches!(
                                self.tree[node_after_ix].item.body,
                                ItemBody::SynthesizeText(..)
                            )
                        {
                            self.tree[node_after_ix].item.body = ItemBody::Text {
                                backslash_escaped: false,
                            };
                        }
                        self.tree[node_after_ix].item.start = new_start;
                        // The `\` a flag refers to sits one byte before its
                        // item, so the link owns it here. A surviving flag
                        // would stretch the successor's span back over it, and
                        // on a `<` keep the inline HTML from opening.
                        if orig_start <= cand.end {
                            match &mut self.tree[node_after_ix].item.body {
                                ItemBody::Text { backslash_escaped }
                                | ItemBody::MaybeHtml(backslash_escaped) => {
                                    *backslash_escaped = false;
                                }
                                _ => {}
                            }
                        }
                        self.repair_construct_after_url_end(cand.end, node_after_ix);
                    }
                }
                ItemBody::MaybeEmphasisEscaped(count, ..) => {
                    let count = count as usize;
                    // Nothing unblocked it, so the escape stands: the delimiter
                    // it hid is literal and the run after it is one shorter.
                    self.tree[cur_ix].item.body = ItemBody::Text {
                        backslash_escaped: true,
                    };
                    let c = self.text.as_bytes()[self.tree[cur_ix].item.start];
                    if !crate::firstpass::delim_run_is_valid(c, count - 1, self.options) {
                        let mut scan = self.tree[cur_ix].next;
                        for _ in 1..count {
                            let Some(next_ix) = scan else { break };
                            self.tree[next_ix].item.body = ItemBody::Text {
                                backslash_escaped: false,
                            };
                            scan = self.tree[next_ix].next;
                        }
                    }
                }
                ItemBody::MaybeLinkOpen => {
                    self.tree[cur_ix].item.body = ItemBody::Text {
                        backslash_escaped: false,
                    };
                    let link_open_doubled = self.tree[cur_ix]
                        .next
                        .map(|ix| self.tree[ix].item.body == ItemBody::MaybeLinkOpen)
                        .unwrap_or(false);
                    if self.options.contains(Options::ENABLE_WIKILINKS) && link_open_doubled {
                        self.wikilink_stack.push(LinkStackEl {
                            node: cur_ix,
                            ty: LinkStackTy::Link,
                        });
                    }
                    self.link_stack.push(LinkStackEl {
                        node: cur_ix,
                        ty: LinkStackTy::Link,
                    });
                }
                ItemBody::MaybeImage => {
                    self.tree[cur_ix].item.body = ItemBody::Text {
                        backslash_escaped: false,
                    };
                    let link_open_doubled = self.tree[cur_ix]
                        .next
                        .map(|ix| self.tree[ix].item.body == ItemBody::MaybeLinkOpen)
                        .unwrap_or(false);
                    if self.options.contains(Options::ENABLE_WIKILINKS) && link_open_doubled {
                        self.wikilink_stack.push(LinkStackEl {
                            node: cur_ix,
                            ty: LinkStackTy::Image,
                        });
                    }
                    self.link_stack.push(LinkStackEl {
                        node: cur_ix,
                        ty: LinkStackTy::Image,
                    });
                }
                ItemBody::MaybeLinkClose(could_be_ref) => {
                    self.tree[cur_ix].item.body = ItemBody::Text {
                        backslash_escaped: false,
                    };
                    let tos_link = self.link_stack.pop();
                    if self.options.contains(Options::ENABLE_WIKILINKS)
                        && self.tree[cur_ix]
                            .next
                            .map(|ix| {
                                matches!(self.tree[ix].item.body, ItemBody::MaybeLinkClose(..))
                            })
                            .unwrap_or(false)
                        && let Some(node) = self.handle_wikilink(block_text, cur_ix, prev)
                    {
                        cur = self.tree[node].next;
                        continue;
                    }
                    if let Some(tos) = tos_link {
                        // skip rendering if already in a link, unless its an
                        // image
                        if tos.ty != LinkStackTy::Image
                            && matches!(
                                self.tree[self.tree.peek_up().unwrap()].item.body,
                                ItemBody::Link(..)
                            )
                        {
                            continue;
                        }
                        if tos.ty == LinkStackTy::Disabled {
                            continue;
                        }
                        let next = self.tree[cur_ix].next;
                        // remark's image label start eats the `[`, so `![^x](y)` stays an image.
                        let footnote_first = tos.ty == LinkStackTy::Link
                            && self.defined_footnote_label(tos.node, cur_ix);
                        if !footnote_first
                            && let Some((next_ix, url, title)) =
                                self.scan_inline_link(block_text, self.tree[cur_ix].item.end, next)
                        {
                            let next_node = scan_nodes_to_ix(&self.tree, next, next_ix);
                            if let Some(prev_ix) = prev {
                                self.tree[prev_ix].next = None;
                            }
                            cur = Some(tos.node);
                            cur_ix = tos.node;
                            let link_ix =
                                self.allocs
                                    .allocate_link(LinkType::Inline, url, title, "".into());
                            self.tree[cur_ix].item.body = if tos.ty == LinkStackTy::Image {
                                ItemBody::Image(link_ix)
                            } else {
                                ItemBody::Link(link_ix)
                            };
                            self.tree[cur_ix].child = self.tree[cur_ix].next;
                            self.tree[cur_ix].next = next_node;
                            self.tree[cur_ix].item.end = next_ix;
                            if let Some(next_node_ix) = next_node {
                                let orig_start = self.tree[next_node_ix].item.start;
                                let new_start = max(orig_start, next_ix);
                                self.tree[next_node_ix].item.start = new_start;
                                // If the text node's start was advanced past
                                // its original position (the link's URL or
                                // title consumed the bytes the escape was
                                // attached to), the `backslash_escaped`
                                // flag no longer applies — clear it so the
                                // arena-build position fixup doesn't extend
                                // the text node's source span back over
                                // bytes already owned by the link.
                                if new_start > orig_start
                                    && let ItemBody::Text { backslash_escaped } =
                                        &mut self.tree[next_node_ix].item.body
                                {
                                    *backslash_escaped = false;
                                }
                            }

                            if tos.ty == LinkStackTy::Link {
                                self.disable_all_links();
                            }
                        } else {
                            // Footnote-first check: if the first bracket content is
                            // `[^X]` where `X` has a matching footnote definition,
                            // emit a FootnoteReference regardless of what follows.
                            // Otherwise `[^X][Y]` would be resolved as a link whose
                            // text happens to start with `^`, which diverges from
                            // remark-gfm's two-node parse (footnote + trailing ref).
                            let first_bracket_start = self.tree[tos.node].item.start;
                            let first_bracket_end = self.tree[cur_ix].item.end;
                            let first_bracket_text =
                                &self.text[first_bracket_start..first_bracket_end];
                            if let Some((label_len, ReferenceLabel::Footnote(footlabel))) =
                                scan_link_label(&self.tree, first_bracket_text, self.options)
                                // A code span can swallow the label's `]`, leaving `cur_ix` on a later one.
                                && label_len == first_bracket_text.len()
                                && self.allocs.footdefs.contains(&footlabel)
                            {
                                let footref = self.allocs.allocate_cow(footlabel);
                                if let Some(def) = self
                                    .allocs
                                    .footdefs
                                    .get_mut(self.allocs.cows[footref.0 as usize].to_owned())
                                {
                                    def.use_count += 1;
                                }
                                let footnote_ix = if tos.ty == LinkStackTy::Image {
                                    self.tree[tos.node].next = Some(cur_ix);
                                    self.tree[tos.node].child = None;
                                    self.tree[tos.node].item.body = ItemBody::SynthesizeChar('!');
                                    self.tree[cur_ix].item.start =
                                        self.tree[tos.node].item.start + 1;
                                    self.tree[tos.node].item.end =
                                        self.tree[tos.node].item.start + 1;
                                    cur_ix
                                } else {
                                    tos.node
                                };
                                self.tree[footnote_ix].next = next;
                                self.tree[footnote_ix].child = None;
                                self.tree[footnote_ix].item.body =
                                    ItemBody::FootnoteReference(footref);
                                self.tree[footnote_ix].item.end = first_bracket_end;
                                prev = Some(footnote_ix);
                                cur = next;
                                self.link_stack.clear();
                                continue;
                            }
                            // ok, so its not an inline link. maybe it is a reference
                            // to a defined link?
                            let scan_result =
                                scan_reference(&self.tree, block_text, next, self.options);
                            let (node_after_link, link_type) = match scan_result {
                                // [label][reference]
                                RefScan::LinkLabel(_, end_ix) => {
                                    // Toggle reference viability of the last closing bracket,
                                    // so that we can skip it on future iterations in case
                                    // it fails in this one. In particular, we won't call
                                    // the broken link callback twice on one reference.
                                    let reference_close_node = if let Some(node) =
                                        scan_nodes_to_ix(&self.tree, next, end_ix - 1)
                                    {
                                        node
                                    } else {
                                        continue;
                                    };
                                    self.tree[reference_close_node].item.body =
                                        ItemBody::MaybeLinkClose(false);
                                    // The label scan walks raw source, so it can
                                    // stop inside a wider item, e.g. an MDX
                                    // expression or directive holding a `]`. The
                                    // label owns up to `end_ix`; the rest is
                                    // literal text, and without this it would
                                    // belong to no node.
                                    let close_end = self.tree[reference_close_node].item.end;
                                    let next_node = if close_end > end_ix {
                                        self.tree[reference_close_node].item.end = end_ix;
                                        let tail = self.tree.create_node(Item {
                                            start: end_ix,
                                            end: close_end,
                                            body: ItemBody::Text {
                                                backslash_escaped: false,
                                            },
                                        });
                                        self.tree[tail].next = self.tree[reference_close_node].next;
                                        self.tree[reference_close_node].next = Some(tail);
                                        Some(tail)
                                    } else {
                                        self.tree[reference_close_node].next
                                    };

                                    (next_node, LinkType::Reference)
                                }
                                // [reference][]
                                RefScan::Collapsed(next_node) => {
                                    // This reference has already been tried, and it's not
                                    // valid. Skip it.
                                    if !could_be_ref {
                                        continue;
                                    }
                                    (next_node, LinkType::Collapsed)
                                }
                                // [X][^Y] — full-reference form with a footnote-shaped
                                // second label. Per CommonMark the full-ref has to
                                // resolve to a link definition, which `^Y` never will;
                                // shortcut fallback is NOT tried. Leave both brackets
                                // literal and let `[^Y]` be parsed as a footnote on
                                // its own MaybeLinkClose iteration.
                                RefScan::UnexpectedFootnote => continue,
                                // `[text][invalid_label]` — the `[` after `[text]`
                                // started a label slot but it wasn't a valid label
                                // (e.g. unescaped `[` inside). Spec: a shortcut link
                                // can't be followed by `[`, so don't fall back to
                                // shortcut. Leave both brackets literal.
                                RefScan::FailedInvalidLabel => continue,
                                // [shortcut]
                                //
                                // [shortcut]: /blah
                                RefScan::Failed => {
                                    if !could_be_ref {
                                        continue;
                                    }
                                    (next, LinkType::Shortcut)
                                }
                            };

                            // FIXME: references and labels are mixed in the naming of variables
                            // below. Disambiguate!

                            // (label, source_ix end)
                            let label: Option<(ReferenceLabel<'input>, usize)> = match scan_result {
                                RefScan::LinkLabel(l, end_ix) => {
                                    Some((ReferenceLabel::Link(l), end_ix))
                                }
                                RefScan::Collapsed(..)
                                | RefScan::Failed
                                | RefScan::FailedInvalidLabel
                                | RefScan::UnexpectedFootnote => {
                                    // No label? maybe it is a shortcut reference
                                    let label_start = self.tree[tos.node].item.end - 1;
                                    let label_end = self.tree[cur_ix].item.end;
                                    scan_link_label(
                                        &self.tree,
                                        &self.text[label_start..label_end],
                                        self.options,
                                    )
                                    .map(|(ix, label)| (label, label_start + ix))
                                    .filter(|(_, end)| *end == label_end)
                                }
                            };

                            let id = match &label {
                                Some(
                                    (ReferenceLabel::Link(l), _) | (ReferenceLabel::Footnote(l), _),
                                ) => l.clone(),
                                None => "".into(),
                            };

                            // see if it's a footnote reference
                            if let Some((ReferenceLabel::Footnote(l), end)) = label {
                                let footref = self.allocs.allocate_cow(l);
                                if let Some(def) = self
                                    .allocs
                                    .footdefs
                                    .get_mut(self.allocs.cows[footref.0 as usize].to_owned())
                                {
                                    def.use_count += 1;
                                }
                                if self
                                    .allocs
                                    .footdefs
                                    .contains(&self.allocs.cows[footref.0 as usize])
                                {
                                    // If this came from a MaybeImage, then the `!` prefix
                                    // isn't part of the footnote reference.
                                    let footnote_ix = if tos.ty == LinkStackTy::Image {
                                        self.tree[tos.node].next = Some(cur_ix);
                                        self.tree[tos.node].child = None;
                                        self.tree[tos.node].item.body =
                                            ItemBody::SynthesizeChar('!');
                                        self.tree[cur_ix].item.start =
                                            self.tree[tos.node].item.start + 1;
                                        self.tree[tos.node].item.end =
                                            self.tree[tos.node].item.start + 1;
                                        cur_ix
                                    } else {
                                        tos.node
                                    };
                                    // use `next` instead of `node_after_link` because
                                    // node_after_link is calculated for a [collapsed][] link,
                                    // which footnotes don't support.
                                    self.tree[footnote_ix].next = next;
                                    self.tree[footnote_ix].child = None;
                                    self.tree[footnote_ix].item.body =
                                        ItemBody::FootnoteReference(footref);
                                    self.tree[footnote_ix].item.end = end;
                                    prev = Some(footnote_ix);
                                    cur = next;
                                    self.link_stack.clear();
                                    continue;
                                }
                            } else if let Some((ReferenceLabel::Link(link_label), end)) = label
                                && let Some((def_link_type, url, title)) = self
                                    .fetch_link_type_url_title(
                                        link_label,
                                        (self.tree[tos.node].item.start)..end,
                                        link_type,
                                        callbacks,
                                    )
                            {
                                let link_ix =
                                    self.allocs.allocate_link(def_link_type, url, title, id);
                                self.tree[tos.node].item.body = if tos.ty == LinkStackTy::Image {
                                    ItemBody::Image(link_ix)
                                } else {
                                    ItemBody::Link(link_ix)
                                };
                                let label_node = self.tree[tos.node].next;

                                // lets do some tree surgery to add the link to the tree
                                // 1st: skip the label node and close node
                                self.tree[tos.node].next = node_after_link;

                                // then, if it exists, add the label node as a child to the link node
                                if label_node != cur {
                                    self.tree[tos.node].child = label_node;

                                    // finally: disconnect list of children
                                    if let Some(prev_ix) = prev {
                                        self.tree[prev_ix].next = None;
                                    }
                                }

                                self.tree[tos.node].item.end = end;
                                // No `max(orig_start, end)` clamp here,
                                // unlike the inline-link splice: the item at
                                // `end - 1` either ends the label or was
                                // truncated to it, and the zero-width items
                                // that could sit at `end` all sit on a URL
                                // or email byte, never a `]`.
                                debug_assert!(
                                    node_after_link.is_none_or(|node_after_ix| {
                                        self.tree[node_after_ix].item.start >= end
                                    }),
                                    "reference splice must not overrun its successor",
                                );

                                // set up cur so next node will be node_after_link
                                cur = Some(tos.node);
                                cur_ix = tos.node;

                                if tos.ty == LinkStackTy::Link {
                                    self.disable_all_links();
                                }
                            }
                        }
                    }
                }
                _ => {}
            }
            prev = cur;
            cur = self.tree[cur_ix].next;
        }
        self.link_stack.clear();
        self.wikilink_stack.clear();
        self.code_delims.clear();
        self.math_delims.clear();
    }

    /// The construct opening on the byte after a URL-ending `\`. The first
    /// pass's escape handler consumed that byte, so the construct either never
    /// got an item or got a blocked one; now that the link owns the `\`, give
    /// it back.
    fn repair_construct_after_url_end(&mut self, cand_end: usize, node_ix: TreeIndex) {
        let item = self.tree[node_ix].item;
        if item.start != cand_end {
            return;
        }
        if let ItemBody::MaybeEmphasisEscaped(count, can_open, can_close) = item.body {
            self.tree[node_ix].item.body = ItemBody::MaybeEmphasis(count, can_open, can_close);
            // The rest of the run was emitted for the shorter run the escape
            // would have left, so only its flanking needs restating.
            let mut scan = self.tree[node_ix].next;
            for _ in 1..count {
                let Some(next_ix) = scan else { break };
                if let ItemBody::MaybeEmphasis(_, open, close) = &mut self.tree[next_ix].item.body {
                    *open = can_open;
                    *close = can_close;
                }
                scan = self.tree[next_ix].next;
            }
            return;
        }
        if !matches!(item.body, ItemBody::Text { .. })
            || self.text.as_bytes().get(cand_end) != Some(&b'&')
        {
            return;
        }
        let (n, Some(value)) = scan_entity(&self.text.as_bytes()[cand_end..]) else {
            return;
        };
        if cand_end + n > item.end {
            return;
        }
        let cow_ix = self.allocs.allocate_cow(value);
        if cand_end + n < item.end {
            let tail = self.tree.create_node(Item {
                start: cand_end + n,
                end: item.end,
                body: ItemBody::Text {
                    backslash_escaped: false,
                },
            });
            self.tree[tail].next = self.tree[node_ix].next;
            self.tree[node_ix].next = Some(tail);
        }
        self.tree[node_ix].item.end = cand_end + n;
        self.tree[node_ix].item.body = ItemBody::SynthesizeText(cow_ix);
    }

    /// Handles a wikilink.
    ///
    /// This function may bail early in case the link is malformed, so this
    /// acts as a control flow guard. Returns the link node if a wikilink was
    /// found and created.
    fn handle_wikilink(
        &mut self,
        block_text: &'input str,
        cur_ix: TreeIndex,
        prev: Option<TreeIndex>,
    ) -> Option<TreeIndex> {
        let next_ix = self.tree[cur_ix].next.unwrap();
        // this is a wikilink closing delim, try popping from
        // the wikilink stack
        if let Some(tos) = self.wikilink_stack.pop() {
            if tos.ty == LinkStackTy::Disabled {
                return None;
            }
            // fetches the beginning of the wikilink body
            let Some(body_node) = self.tree[tos.node].next.and_then(|ix| self.tree[ix].next) else {
                // skip if no next node exists, like at end of input
                return None;
            };
            let start_ix = self.tree[body_node].item.start;
            let end_ix = self.tree[cur_ix].item.start;
            let wikilink = match scan_wikilink_pipe(
                block_text,
                start_ix, // bounded by closing tag
                end_ix - start_ix,
            ) {
                Some((rest, wikitext)) => {
                    // bail early if the wikiname would be empty
                    if wikitext.is_empty() {
                        return None;
                    }
                    // [[WikiName|rest]]
                    let body_node = scan_nodes_to_ix(&self.tree, Some(body_node), rest);
                    if let Some(body_node) = body_node {
                        // break node so passes can actually format
                        // the display text
                        self.tree[body_node].item.start = rest;
                        Some((true, body_node, wikitext))
                    } else {
                        None
                    }
                }
                None => {
                    let wikitext = &block_text[start_ix..end_ix];
                    // bail early if the wikiname would be empty
                    if wikitext.is_empty() {
                        return None;
                    }
                    let body_node = self.tree.create_node(Item {
                        start: start_ix,
                        end: end_ix,
                        body: ItemBody::Text {
                            backslash_escaped: false,
                        },
                    });
                    Some((false, body_node, wikitext))
                }
            };

            if let Some((has_pothole, body_node, wikiname)) = wikilink {
                let link_ix = self.allocs.allocate_link(
                    LinkType::WikiLink { has_pothole },
                    wikiname.into(),
                    "".into(),
                    "".into(),
                );
                if let Some(prev_ix) = prev {
                    self.tree[prev_ix].next = None;
                }
                if tos.ty == LinkStackTy::Image {
                    self.tree[tos.node].item.body = ItemBody::Image(link_ix);
                } else {
                    self.tree[tos.node].item.body = ItemBody::Link(link_ix);
                }
                self.tree[tos.node].child = Some(body_node);
                self.tree[tos.node].next = self.tree[next_ix].next;
                self.tree[tos.node].item.end = end_ix + 2;
                self.disable_all_links();
                return Some(tos.node);
            }
        }

        None
    }

    /// Leaves `inline_stack` reset, so the next scope reuses it and its capacity.
    fn handle_emphasis_in_scope(&mut self, start: Option<TreeIndex>) {
        debug_assert!(self.inline_stack.is_reset());
        let mut prev = None;
        let mut prev_ix: TreeIndex;
        let mut cur = start;

        let mut single_quote_open: Option<TreeIndex> = None;
        let mut double_quote_open: bool = false;

        while let Some(mut cur_ix) = cur {
            match self.tree[cur_ix].item.body {
                ItemBody::MaybeEmphasis(count, can_open, can_close) => {
                    let mut count = count as usize;
                    let run_length = count;
                    let c = self.text.as_bytes()[self.tree[cur_ix].item.start];
                    let both = can_open && can_close;
                    // Defer `~`/`^` resolution to the post-pass.
                    // Without lookahead, the single-pass can't tell whether an
                    // earlier `*`/`_` opener will pair (in which case the
                    // `~`/`^` should match inside the future emphasis) or
                    // remain unmatched (in which case `~`/`^` would cross the
                    // boundary). micromark handles this with a separate
                    // strikethrough resolve phase that runs after emphasis.
                    if c == b'~' || c == b'^' {
                        prev_ix = cur_ix + count - 1;
                        prev = Some(prev_ix);
                        cur = self.tree[prev_ix].next;
                        continue;
                    }
                    if can_close {
                        while let Some(el) =
                            self.inline_stack
                                .find_match(&mut self.tree, c, run_length, count, both)
                        {
                            // have a match!
                            if let Some(prev_ix) = prev {
                                self.tree[prev_ix].next = None;
                            }
                            // Consume at most two markers per inner-loop pass
                            // (one `<strong>`/`<em>` per match), matching
                            // micromark's `use = open>1 && close>1 ? 2 : 1`.
                            // The outer `while let` then drives nesting by
                            // re-running `find_match` with the leftover
                            // counts, which is how `***foo***` becomes
                            // `<em><strong>foo</strong></em>` instead of one
                            // flat match.
                            let match_count = min(2, min(count, el.count));
                            // start, end are tree node indices
                            let mut end = cur_ix - 1;
                            let mut start = el.start + el.count;

                            // work from the inside out
                            while start > el.start + el.count - match_count {
                                let inc = if start > el.start + el.count - match_count + 1 {
                                    2
                                } else {
                                    1
                                };
                                let ty = if c == b'~' {
                                    if inc == 2 {
                                        if self.options.contains(Options::ENABLE_STRIKETHROUGH) {
                                            ItemBody::Strikethrough
                                        } else {
                                            ItemBody::Text {
                                                backslash_escaped: false,
                                            }
                                        }
                                    } else if self.options.contains(Options::ENABLE_SUBSCRIPT) {
                                        ItemBody::Subscript
                                    } else if self.options.contains(Options::ENABLE_STRIKETHROUGH) {
                                        ItemBody::Strikethrough
                                    } else {
                                        ItemBody::Text {
                                            backslash_escaped: false,
                                        }
                                    }
                                } else if c == b'^' {
                                    if self.options.contains(Options::ENABLE_SUPERSCRIPT) {
                                        ItemBody::Superscript
                                    } else {
                                        ItemBody::Text {
                                            backslash_escaped: false,
                                        }
                                    }
                                } else if inc == 2 {
                                    ItemBody::Strong
                                } else {
                                    ItemBody::Emphasis
                                };

                                let root = start - inc;
                                end = end + inc;
                                self.tree[root].item.body = ty;
                                self.tree[root].item.end = self.tree[end].item.end;
                                self.tree[root].child = Some(start);
                                self.tree[root].next = None;
                                start = root;
                            }

                            // set next for top most emph level
                            prev_ix = el.start + el.count - match_count;
                            prev = Some(prev_ix);
                            cur = self.tree[cur_ix + match_count - 1].next;
                            self.tree[prev_ix].next = cur;

                            if el.count > match_count {
                                self.inline_stack.push(InlineEl {
                                    start: el.start,
                                    count: el.count - match_count,
                                    run_length: el.run_length,
                                    c: el.c,
                                    both: el.both,
                                })
                            }
                            count -= match_count;
                            if count > 0 {
                                cur_ix = cur.unwrap();
                            } else {
                                break;
                            }
                        }
                    }
                    if count > 0 {
                        if can_open {
                            self.inline_stack.push(InlineEl {
                                start: cur_ix,
                                run_length,
                                count,
                                c,
                                both,
                            });
                        } else {
                            for i in 0..count {
                                self.tree[cur_ix + i].item.body = ItemBody::Text {
                                    backslash_escaped: false,
                                };
                            }
                        }
                        prev_ix = cur_ix + count - 1;
                        prev = Some(prev_ix);
                        cur = self.tree[prev_ix].next;
                    }
                }
                ItemBody::MaybeSmartQuote(c, can_open, can_close) => {
                    self.tree[cur_ix].item.body = match c {
                        b'\'' => {
                            if let (Some(open_ix), true) = (single_quote_open, can_close) {
                                self.tree[open_ix].item.body = ItemBody::SynthesizeChar('‘');
                                single_quote_open = None;
                            } else if can_open {
                                single_quote_open = Some(cur_ix);
                            }
                            ItemBody::SynthesizeChar('’')
                        }
                        _ /* double quote */ => {
                            if can_close && double_quote_open {
                                double_quote_open = false;
                                ItemBody::SynthesizeChar('”')
                            } else if can_open {
                                double_quote_open = true;
                                ItemBody::SynthesizeChar('“')
                            } else if can_close {
                                // Unmatched close-flanking quote (e.g. the inch
                                // mark in `24"`): render a closing curly quote.
                                ItemBody::SynthesizeChar('”')
                            } else {
                                // Non-flanking quote: opening curly quote.
                                ItemBody::SynthesizeChar('“')
                            }
                        }
                    };
                    prev = cur;
                    cur = self.tree[cur_ix].next;
                }
                ItemBody::HardBreak(true) => {
                    if self.tree[cur_ix].next.is_none() {
                        self.tree[cur_ix].item.body = ItemBody::SynthesizeChar('\\');
                    }
                    prev = cur;
                    cur = self.tree[cur_ix].next;
                }
                _ => {
                    prev = cur;
                    cur = self.tree[cur_ix].next;
                }
            }
        }
        self.inline_stack.pop_all(&mut self.tree);
    }

    /// Second-pass strikethrough/sub/sup resolution. Walks the tree
    /// hierarchically and resolves `~`/`^` MaybeEmphasis tokens within
    /// each inline scope independently. This matches micromark's
    /// post-emphasis resolve phase: a `~..~` pair only forms when both
    /// ends lie within the same enclosing scope (root, emphasis, link,
    /// etc.). Multi-char `~~` strikethrough was already resolved in
    /// the main pass.
    /// Resolve `~`/`^` runs at this scope. With `descend`, also recurse into
    /// already-formed spans; the per-scope driver passes `false` and handles
    /// descent itself so it can flip the resolve order for nested content.
    fn resolve_tildes_carets_in_scope(&mut self, start: Option<TreeIndex>, descend: bool) {
        let mut stack: Vec<InlineEl> = Vec::new();
        let mut cur = start;
        let mut prev: Option<TreeIndex> = None;
        while let Some(mut cur_ix) = cur {
            match self.tree[cur_ix].item.body {
                ItemBody::MaybeEmphasis(count, can_open, can_close) => {
                    let count = count as usize;
                    let c = self.text.as_bytes()[self.tree[cur_ix].item.start];
                    if c != b'~' && c != b'^' {
                        prev = Some(cur_ix);
                        cur = self.tree[cur_ix].next;
                        continue;
                    }
                    let run_length = count;
                    let mut remaining = count;
                    if can_close {
                        while remaining > 0 {
                            let res = stack
                                .iter()
                                .enumerate()
                                .rfind(|(_, el)| el.c == c && el.run_length == run_length);
                            let Some((matching_ix, matching_el)) = res else {
                                break;
                            };
                            let matching_el = *matching_el;
                            if let Some(prev_ix) = prev {
                                self.tree[prev_ix].next = None;
                            }
                            // Convert intermediate `~`/`^` openers above the
                            // match to text — they failed to find a pair.
                            for el in &stack[(matching_ix + 1)..] {
                                for i in 0..el.count {
                                    self.tree[el.start + i].item.body = ItemBody::Text {
                                        backslash_escaped: false,
                                    };
                                }
                            }
                            stack.truncate(matching_ix);
                            let match_count =
                                core::cmp::min(2, core::cmp::min(remaining, matching_el.count));
                            let mut end = cur_ix - 1;
                            let mut sub_start = matching_el.start + matching_el.count;
                            while sub_start > matching_el.start + matching_el.count - match_count {
                                let inc = if sub_start
                                    > matching_el.start + matching_el.count - match_count + 1
                                {
                                    2
                                } else {
                                    1
                                };
                                let ty = if c == b'~' {
                                    if inc == 2 {
                                        if self.options.contains(Options::ENABLE_STRIKETHROUGH) {
                                            ItemBody::Strikethrough
                                        } else {
                                            ItemBody::Text {
                                                backslash_escaped: false,
                                            }
                                        }
                                    } else if self.options.contains(Options::ENABLE_SUBSCRIPT) {
                                        ItemBody::Subscript
                                    } else if self.options.contains(Options::ENABLE_STRIKETHROUGH) {
                                        ItemBody::Strikethrough
                                    } else {
                                        ItemBody::Text {
                                            backslash_escaped: false,
                                        }
                                    }
                                } else if self.options.contains(Options::ENABLE_SUPERSCRIPT) {
                                    ItemBody::Superscript
                                } else {
                                    ItemBody::Text {
                                        backslash_escaped: false,
                                    }
                                };
                                let root = sub_start - inc;
                                end = end + inc;
                                self.tree[root].item.body = ty;
                                self.tree[root].item.end = self.tree[end].item.end;
                                self.tree[root].child = Some(sub_start);
                                self.tree[root].next = None;
                                sub_start = root;
                            }
                            let new_prev_ix = matching_el.start + matching_el.count - match_count;
                            let new_cur = self.tree[cur_ix + match_count - 1].next;
                            self.tree[new_prev_ix].next = new_cur;
                            prev = Some(new_prev_ix);
                            if matching_el.count > match_count {
                                stack.push(InlineEl {
                                    start: matching_el.start,
                                    count: matching_el.count - match_count,
                                    run_length: matching_el.run_length,
                                    c: matching_el.c,
                                    both: matching_el.both,
                                });
                            }
                            remaining -= match_count;
                            if remaining > 0 {
                                let Some(next_cur) = new_cur else { break };
                                cur_ix = next_cur;
                            } else {
                                break;
                            }
                        }
                    }
                    if remaining > 0 {
                        if can_open {
                            stack.push(InlineEl {
                                start: cur_ix,
                                count: remaining,
                                run_length,
                                c,
                                both: can_open && can_close,
                            });
                        } else {
                            for i in 0..remaining {
                                self.tree[cur_ix + i].item.body = ItemBody::Text {
                                    backslash_escaped: false,
                                };
                            }
                        }
                        let prev_ix = cur_ix + remaining - 1;
                        prev = Some(prev_ix);
                        cur = self.tree[prev_ix].next;
                    } else {
                        cur = self.tree[prev.unwrap()].next;
                    }
                    continue;
                }
                ItemBody::Emphasis
                | ItemBody::Strong
                | ItemBody::Strikethrough
                | ItemBody::Subscript
                | ItemBody::Superscript
                | ItemBody::Link(_)
                | ItemBody::Image(_)
                    if descend =>
                {
                    let child = self.tree[cur_ix].child;
                    self.resolve_tildes_carets_in_scope(child, true);
                }
                _ => {}
            }
            prev = Some(cur_ix);
            cur = self.tree[cur_ix].next;
        }
        // End of scope: any remaining openers couldn't find a closer.
        for el in stack {
            for i in 0..el.count {
                self.tree[el.start + i].item.body = ItemBody::Text {
                    backslash_escaped: false,
                };
            }
        }
    }

    fn disable_all_links(&mut self) {
        self.link_stack.disable_all_links();
        self.wikilink_stack.disable_all_links();
    }

    fn defined_footnote_label(&self, open_ix: TreeIndex, close_ix: TreeIndex) -> bool {
        let start = self.tree[open_ix].item.start;
        if !self.options.contains(Options::ENABLE_FOOTNOTES)
            || self.text.as_bytes().get(start + 1) != Some(&b'^')
        {
            return false;
        }
        let label_text = &self.text[start..self.tree[close_ix].item.end];
        let Some((len, ReferenceLabel::Footnote(label))) =
            scan_link_label(&self.tree, label_text, self.options)
        else {
            return false;
        };
        // micromark ends a footnote call at whitespace, where the shared label scan collapses it.
        len == label_text.len()
            && !label_text.as_bytes()[..len]
                .iter()
                .any(|b| matches!(b, b' ' | b'\t' | b'\n' | b'\r'))
            && self.allocs.footdefs.contains(&label)
    }

    /// Returns next byte index, url and title.
    fn scan_inline_link(
        &self,
        underlying: &'input str,
        mut ix: usize,
        node: Option<TreeIndex>,
    ) -> Option<(usize, CowStr<'input>, CowStr<'input>)> {
        if underlying.as_bytes().get(ix) != Some(&b'(') {
            return None;
        }
        ix += 1;

        let scan_separator = |ix: &mut usize| {
            *ix += scan_while(&underlying.as_bytes()[*ix..], is_space_or_tab);
            if let Some(bl) = scan_eol(&underlying.as_bytes()[*ix..]) {
                *ix += bl;
                *ix += skip_container_prefixes(
                    &self.tree,
                    &underlying.as_bytes()[*ix..],
                    self.options,
                );
            }
            *ix += scan_while(&underlying.as_bytes()[*ix..], is_space_or_tab);
        };

        scan_separator(&mut ix);

        let (dest_length, dest) = scan_link_dest(underlying, ix, LINK_MAX_NESTED_PARENS)?;
        let dest = unescape(dest, self.tree.is_in_table());
        ix += dest_length;

        let dest_end = ix;
        scan_separator(&mut ix);

        // A title is only reachable through whitespace after the destination.
        let title = if ix > dest_end
            && let Some((bytes_scanned, t)) = self.scan_link_title(underlying, ix, node)
        {
            ix += bytes_scanned;
            scan_separator(&mut ix);
            t
        } else {
            "".into()
        };
        if underlying.as_bytes().get(ix) != Some(&b')') {
            return None;
        }
        ix += 1;

        Some((ix, dest, title))
    }

    // returns (bytes scanned, title cow)
    fn scan_link_title(
        &self,
        text: &'input str,
        start_ix: usize,
        node: Option<TreeIndex>,
    ) -> Option<(usize, CowStr<'input>)> {
        let bytes = text.as_bytes();
        let open = match bytes.get(start_ix) {
            Some(b @ b'\'') | Some(b @ b'\"') | Some(b @ b'(') => *b,
            _ => return None,
        };
        if open == b'(' && start_ix >= self.unclosed_paren_title_floor.get() {
            return None;
        }
        // Unlike CommonMark, remark keeps an unescaped `(` inside a paren title as content.
        let close = if open == b'(' { b')' } else { open };

        let mut title = String::new();
        let mut mark = start_ix + 1;
        let mut i = start_ix + 1;

        while i < bytes.len() {
            let c = bytes[i];

            if c == close {
                let cow = if title.is_empty() {
                    (i - start_ix + 1, text[mark..i].into())
                } else {
                    title.push_str(&text[mark..i]);
                    (i - start_ix + 1, title.into())
                };

                return Some(cow);
            }

            if (c == b'\n' || c == b'\r')
                && let Some(node_ix) = scan_nodes_to_ix(&self.tree, node, i + 1)
                && self.tree[node_ix].item.start > i
            {
                title.push_str(&text[mark..i]);
                // The title's line endings are content, kept byte for byte.
                title.push(c as char);
                if c == b'\r' && bytes.get(i + 1) == Some(&b'\n') {
                    title.push('\n');
                }
                i = self.tree[node_ix].item.start;
                mark = i;
                continue;
            }
            if c == b'&'
                && let (n, Some(value)) = scan_entity(&bytes[i..])
            {
                title.push_str(&text[mark..i]);
                title.push_str(&value);
                i += n;
                mark = i;
                continue;
            }
            if self.tree.is_in_table()
                && c == b'\\'
                && i + 2 < bytes.len()
                && bytes[i + 1] == b'\\'
                && bytes[i + 2] == b'|'
            {
                // this runs if there are an even number of pipes in a table
                // if it's odd, then it gets parsed as normal
                title.push_str(&text[mark..i]);
                i += 2;
                mark = i;
            }
            if c == b'\\' && i + 1 < bytes.len() && is_ascii_punctuation(bytes[i + 1]) {
                title.push_str(&text[mark..i]);
                i += 1;
                mark = i;
            }

            i += 1;
        }

        if open == b'(' {
            let floor = self.unclosed_paren_title_floor.get();
            self.unclosed_paren_title_floor.set(floor.min(start_ix));
        }
        None
    }

    fn make_math_span(&mut self, open: TreeIndex, close: TreeIndex) {
        // Find the end of the opening run of consecutive $ tokens
        let mut open_end = open;
        {
            let mut peek = self.tree[open].next;
            while let Some(peek_ix) = peek {
                if matches!(self.tree[peek_ix].item.body, ItemBody::MaybeMath(..))
                    && self.tree[peek_ix].item.start == self.tree[open_end].item.end
                    && peek_ix != close
                {
                    open_end = peek_ix;
                    peek = self.tree[peek_ix].next;
                } else {
                    break;
                }
            }
        }
        // Find the end of the closing run
        let mut close_end = close;
        {
            let mut peek = self.tree[close].next;
            while let Some(peek_ix) = peek {
                if matches!(self.tree[peek_ix].item.body, ItemBody::MaybeMath(..))
                    && self.tree[peek_ix].item.start == self.tree[close_end].item.end
                {
                    close_end = peek_ix;
                    peek = self.tree[peek_ix].next;
                } else {
                    break;
                }
            }
        }

        let span_start = self.tree[open_end].item.end;
        let span_end = self.tree[close].item.start;

        if span_start > span_end {
            self.tree[open].item.body = ItemBody::Text {
                backslash_escaped: false,
            };
            return;
        }

        let spanned_text = &self.text[span_start..span_end];
        let spanned_bytes = spanned_text.as_bytes();
        let mut buf: Option<String> = None;

        let mut start_ix = 0;
        let mut ix = 0;
        while ix < spanned_bytes.len() {
            let c = spanned_bytes[ix];
            if c == b'\r' || c == b'\n' {
                ix += 1;
                let buf = buf.get_or_insert_with(|| String::with_capacity(spanned_bytes.len()));
                buf.push_str(&spanned_text[start_ix..ix]);
                // Use the full source bytes from this position (not just
                // the span slice) so scan_containers can see the real
                // line content past the closing backtick. With only the
                // span slice, a partial-indent line followed by buffer
                // end (e.g. `    ` + closing) was misread as EOL by
                // is_at_eol — letting the ListItem container "match" the
                // 4 spaces of a 5-indent item and over-strip the code
                // span's trailing whitespace.
                let from = span_start + ix;
                let (scanned, leftover) = skip_container_prefixes_with_remaining(
                    &self.tree,
                    &self.text.as_bytes()[from..],
                    self.options,
                );
                let scanned = scanned.min(spanned_bytes.len() - ix);
                ix += scanned;
                start_ix = ix;
                // Preserve leftover virtual columns from a tab the
                // container only partially consumed (e.g. `\t` in a 2-col
                // listitem leaves 2 spaces of content).
                for _ in 0..leftover {
                    buf.push(' ');
                }
            } else if c == b'\\'
                && spanned_bytes.get(ix + 1) == Some(&b'|')
                && self.tree.is_in_table()
            {
                let buf = buf.get_or_insert_with(|| String::with_capacity(spanned_bytes.len()));
                buf.push_str(&spanned_text[start_ix..ix]);
                buf.push('|');
                ix += 2;
                start_ix = ix;
            } else {
                ix += 1;
            }
        }

        if let Some(buf) = &mut buf {
            buf.push_str(&spanned_text[start_ix..]);
        }
        let cow: CowStr<'input> = strip_span_padding(buf, spanned_text);

        self.tree[open].item.body = ItemBody::Math(self.allocs.allocate_cow(cow), false);
        self.tree[open].item.end = self.tree[close_end].item.end;
        self.tree[open].next = self.tree[close_end].next;
    }

    /// Make a code span.
    ///
    /// Both `open` and `close` are matching MaybeCode items.
    fn make_code_span(&mut self, open: TreeIndex, close: TreeIndex, preceding_backslash: bool) {
        let span_start = self.tree[open].item.end;
        let span_end = self.tree[close].item.start;
        let mut buf: Option<String> = None;

        let spanned_text = &self.text[span_start..span_end];
        let spanned_bytes = spanned_text.as_bytes();
        let mut start_ix = 0;
        let mut ix = 0;
        while ix < spanned_bytes.len() {
            let c = spanned_bytes[ix];
            if c == b'\r' || c == b'\n' {
                let buf = buf.get_or_insert_with(|| String::with_capacity(spanned_bytes.len()));
                // The span's own line endings are content and are kept byte for
                // byte; only the container prefix below is stripped.
                buf.push_str(&spanned_text[start_ix..ix]);
                buf.push(c as char);
                ix += 1;
                if c == b'\r' && spanned_bytes.get(ix) == Some(&b'\n') {
                    buf.push('\n');
                    ix += 1;
                }
                // Use the full source bytes from this position (not just
                // the span slice) so scan_containers can see the real
                // line content past the closing backtick. With only the
                // span slice, a partial-indent line followed by buffer
                // end (e.g. `    ` + closing) was misread as EOL by
                // is_at_eol — letting the ListItem container "match" the
                // 4 spaces of a 5-indent item and over-strip the code
                // span's trailing whitespace.
                let from = span_start + ix;
                let (scanned, leftover) = skip_container_prefixes_with_remaining(
                    &self.tree,
                    &self.text.as_bytes()[from..],
                    self.options,
                );
                let scanned = scanned.min(spanned_bytes.len() - ix);
                ix += scanned;
                start_ix = ix;
                // Preserve leftover virtual columns from a tab the
                // container only partially consumed (e.g. `\t` in a 2-col
                // listitem leaves 2 spaces of content).
                for _ in 0..leftover {
                    buf.push(' ');
                }
            } else if c == b'\\'
                && spanned_bytes.get(ix + 1) == Some(&b'|')
                && self.tree.is_in_table()
            {
                let buf = buf.get_or_insert_with(|| String::with_capacity(spanned_bytes.len()));
                buf.push_str(&spanned_text[start_ix..ix]);
                buf.push('|');
                ix += 2;
                start_ix = ix;
            } else {
                ix += 1;
            }
        }

        if let Some(buf) = &mut buf {
            buf.push_str(&spanned_text[start_ix..]);
        }
        let cow: CowStr<'input> = strip_span_padding(buf, spanned_text);

        if preceding_backslash {
            self.tree[open].item.body = ItemBody::Text {
                backslash_escaped: true,
            };
            self.tree[open].item.end = self.tree[open].item.start + 1;
            self.tree[open].next = Some(close);
            self.tree[close].item.body = ItemBody::Code(self.allocs.allocate_cow(cow));
            self.tree[close].item.start = self.tree[open].item.start + 1;
        } else {
            self.tree[open].item.body = ItemBody::Code(self.allocs.allocate_cow(cow));
            self.tree[open].item.end = self.tree[close].item.end;
            self.tree[open].next = self.tree[close].next;
        }

        // MDX: errors recorded in pass 1 for `{` inside what turned out to be a
        // code span are false positives — the `{` is literal text.
        if !self.mdx_errors.is_empty() {
            self.mdx_errors
                .retain(|(offset, _)| *offset < span_start || *offset >= span_end);
        }
    }

    /// On success, returns a buffer containing the inline html and byte offset.
    /// When no bytes were skipped, the buffer will be empty and the html can be
    /// represented as a subslice of the input string.
    fn scan_inline_html(&mut self, bytes: &[u8], ix: usize) -> Option<(Vec<u8>, usize)> {
        let c = *bytes.get(ix)?;
        if c == b'!' {
            Some((
                vec![],
                scan_inline_html_comment(bytes, ix + 1, &mut self.html_scan_guard)?,
            ))
        } else if c == b'?' {
            Some((
                vec![],
                scan_inline_html_processing(bytes, ix + 1, &mut self.html_scan_guard)?,
            ))
        } else {
            let (span, i) = scan_html_block_inner(
                // Subtract 1 to include the < character
                &bytes[(ix - 1)..],
                Some(&|bytes| skip_container_prefixes(&self.tree, bytes, self.options)),
            )?;
            Some((span, i + ix - 1))
        }
    }
}

/// Returns number of containers scanned.
pub(crate) fn scan_containers(
    tree: &Tree<Item>,
    line_start: &mut LineStart<'_>,
    options: Options,
) -> usize {
    let mut i = 0;
    for &node_ix in tree.walk_spine() {
        match tree[node_ix].item.body {
            ItemBody::BlockQuote(..) => {
                let save = line_start.save_cursor();
                // In MDX mode indented code blocks are disabled, so the
                // ≤3-space cap on blockquote prefix indent doesn't apply —
                // tab- or 4+-space-indented `>` should still continue the
                // blockquote (matches micromark + remark-mdx).
                if options.contains(Options::ENABLE_MDX) {
                    line_start.scan_all_space();
                } else {
                    let _ = line_start.scan_space(3);
                }
                if !line_start.scan_blockquote_marker() {
                    line_start.restore_cursor(save);
                    break;
                }
            }
            ItemBody::ListItem(indent, _) => {
                let save = line_start.save_cursor();
                if !line_start.scan_space(indent as usize) && !line_start.is_at_eol() {
                    line_start.restore_cursor(save);
                    break;
                }
            }
            ItemBody::DefinitionListDefinition(indent, _) => {
                let save = line_start.save_cursor();
                if !line_start.scan_space(indent as usize) && !line_start.is_at_eol() {
                    line_start.restore_cursor(save);
                    break;
                }
            }
            ItemBody::FootnoteDefinition(..) if options.contains(Options::ENABLE_FOOTNOTES) => {
                let save = line_start.save_cursor();
                if !line_start.scan_space(4) && !line_start.is_at_eol() {
                    line_start.restore_cursor(save);
                    break;
                }
            }
            _ => (),
        }
        i += 1;
    }
    i
}

/// Strip one leading and one trailing space or line ending from a code or math
/// span, but only when both are present and the span isn't all whitespace. A
/// CRLF is a single line ending, so it goes as a pair.
///
/// `buf` is the span's rewritten content when a container prefix or `\|` escape
/// forced an allocation; otherwise the span is `spanned_text` verbatim.
fn strip_span_padding<'input>(buf: Option<String>, spanned_text: &'input str) -> CowStr<'input> {
    let s = buf.as_deref().unwrap_or(spanned_text);
    let lead = if s.starts_with("\r\n") {
        2
    } else {
        usize::from(matches!(s.as_bytes().first(), Some(b' ' | b'\n' | b'\r')))
    };
    let trail = if s.ends_with("\r\n") {
        2
    } else {
        usize::from(matches!(s.as_bytes().last(), Some(b' ' | b'\n' | b'\r')))
    };
    let all_spaces = s.bytes().all(|b| matches!(b, b' ' | b'\n' | b'\r'));

    if !all_spaces && lead > 0 && trail > 0 {
        if let Some(mut buf) = buf {
            if !buf.is_empty() {
                buf.truncate(buf.len() - trail);
                buf.replace_range(..lead, "");
            }
            buf.into()
        } else {
            spanned_text[lead..(spanned_text.len() - trail).max(lead)].into()
        }
    } else if let Some(buf) = buf {
        buf.into()
    } else {
        spanned_text.into()
    }
}

pub(crate) fn skip_container_prefixes(tree: &Tree<Item>, bytes: &[u8], options: Options) -> usize {
    let mut line_start = LineStart::new(bytes);
    let _ = scan_containers(tree, &mut line_start, options);
    line_start.bytes_scanned()
}

/// Like `skip_container_prefixes`, but also returns the leftover virtual
/// space columns from tab-stop expansion past the last consumed container
/// prefix. Used by math-span content extraction to faithfully reproduce
/// indentation that the container "ate" only partially — e.g. a single
/// `\t` (4 cols) in a list item with 2-col content indent leaves 2
/// trailing spaces of content.
fn skip_container_prefixes_with_remaining(
    tree: &Tree<Item>,
    bytes: &[u8],
    options: Options,
) -> (usize, usize) {
    let mut line_start = LineStart::new(bytes);
    let _ = scan_containers(tree, &mut line_start, options);
    (line_start.bytes_scanned(), line_start.remaining_space())
}

impl Tree<Item> {
    pub(crate) fn append_text(&mut self, start: usize, end: usize, backslash_escaped: bool) {
        if end > start {
            if let Some(ix) = self.cur()
                && matches!(self[ix].item.body, ItemBody::Text { .. })
                && self[ix].item.end == start
            {
                self[ix].item.end = end;
                return;
            }
            self.append(Item {
                start,
                end,
                body: ItemBody::Text { backslash_escaped },
            });
        }
    }
    /// Returns true if the current node is inside a table.
    ///
    /// If `cur` is an ItemBody::Table, it would return false,
    /// but since the `TableRow` and `TableHead` and `TableCell`
    /// are children of the table, anything doing inline parsing
    /// doesn't need to care about that.
    pub(crate) fn is_in_table(&self) -> bool {
        fn might_be_in_table(item: &Item) -> bool {
            item.body.is_inline()
                || matches!(item.body, |ItemBody::TableHead| ItemBody::TableRow
                    | ItemBody::TableCell)
        }
        for &ix in self.walk_spine().rev() {
            if matches!(self[ix].item.body, ItemBody::Table(_)) {
                return true;
            }
            if !might_be_in_table(&self[ix].item) {
                return false;
            }
        }
        false
    }
}

#[derive(Copy, Clone, Debug)]
struct InlineEl {
    /// offset of tree node
    start: TreeIndex,
    /// number of delimiters available for matching
    count: usize,
    /// length of the run that these delimiters came from
    run_length: usize,
    /// b'*', b'_', or b'~'
    c: u8,
    /// can both open and close
    both: bool,
}

#[derive(Debug, Clone, Default)]
struct InlineStack {
    stack: Vec<InlineEl>,
    // Lower bounds for matching indices in the stack. For example
    // a strikethrough delimiter will never match with any element
    // in the stack with index smaller than
    // `lower_bounds[InlineStack::TILDES]`.
    lower_bounds: [usize; 10],
}

impl InlineStack {
    /// These are indices into the lower bounds array.
    /// Not both refers to the property that the delimiter can not both
    /// be opener as a closer.
    const UNDERSCORE_NOT_BOTH: usize = 0;
    const ASTERISK_NOT_BOTH: usize = 1;
    const ASTERISK_BASE: usize = 2;
    const TILDES: usize = 5;
    const UNDERSCORE_BASE: usize = 6;
    const CIRCUMFLEXES: usize = 9;

    fn is_reset(&self) -> bool {
        self.stack.is_empty() && self.lower_bounds == [0; 10]
    }

    fn pop_all(&mut self, tree: &mut Tree<Item>) {
        for el in self.stack.drain(..) {
            for i in 0..el.count {
                tree[el.start + i].item.body = ItemBody::Text {
                    backslash_escaped: false,
                };
            }
        }
        self.lower_bounds = [0; 10];
    }

    fn get_lowerbound(&self, c: u8, count: usize, both: bool) -> usize {
        if c == b'_' {
            let mod3_lower = self.lower_bounds[InlineStack::UNDERSCORE_BASE + count % 3];
            if both {
                mod3_lower
            } else {
                min(
                    mod3_lower,
                    self.lower_bounds[InlineStack::UNDERSCORE_NOT_BOTH],
                )
            }
        } else if c == b'*' {
            let mod3_lower = self.lower_bounds[InlineStack::ASTERISK_BASE + count % 3];
            if both {
                mod3_lower
            } else {
                min(
                    mod3_lower,
                    self.lower_bounds[InlineStack::ASTERISK_NOT_BOTH],
                )
            }
        } else if c == b'^' {
            self.lower_bounds[InlineStack::CIRCUMFLEXES]
        } else {
            self.lower_bounds[InlineStack::TILDES]
        }
    }

    fn set_lowerbound(&mut self, c: u8, count: usize, both: bool, new_bound: usize) {
        if c == b'_' {
            if both {
                self.lower_bounds[InlineStack::UNDERSCORE_BASE + count % 3] = new_bound;
            } else {
                self.lower_bounds[InlineStack::UNDERSCORE_NOT_BOTH] = new_bound;
            }
        } else if c == b'*' {
            self.lower_bounds[InlineStack::ASTERISK_BASE + count % 3] = new_bound;
            if !both {
                self.lower_bounds[InlineStack::ASTERISK_NOT_BOTH] = new_bound;
            }
        } else if c == b'^' {
            self.lower_bounds[InlineStack::CIRCUMFLEXES] = new_bound;
        } else {
            self.lower_bounds[InlineStack::TILDES] = new_bound;
        }
    }

    fn truncate(&mut self, new_bound: usize) {
        self.stack.truncate(new_bound);
        for lower_bound in &mut self.lower_bounds {
            if *lower_bound > new_bound {
                *lower_bound = new_bound;
            }
        }
    }

    /// Find an opener that can match `c` of original `run_length`.
    ///
    /// `current_count` is the **remaining** length of the closer being
    /// processed (chars not yet consumed by earlier inner-loop matches).
    /// We use it for CommonMark rule 9 (the "mod 3" both-side rule) so
    /// that after a partial consumption like `3*foo *bar**` the outer `*`
    /// can pair with what's left of the `**` — micromark re-evaluates the
    /// rule using only the *current* run lengths on each side.
    ///
    /// `run_length` is the original closer length; it stays stable across
    /// inner-loop iterations and is what the lower-bounds optimisation and
    /// the strict tilde/caret length check key off.
    fn find_match(
        &mut self,
        tree: &mut Tree<Item>,
        c: u8,
        run_length: usize,
        current_count: usize,
        both: bool,
    ) -> Option<InlineEl> {
        // Use current_count (the post-partial-consumption remaining length)
        // for the rule-9 mod-3 lowerbound key, not run_length. After an
        // inner-loop pass consumes part of the closer, the remaining
        // length sits in a different mod-3 bucket and may now satisfy
        // rule 9 with openers the earlier (longer) attempt failed
        // against. Keying on run_length would carry over the earlier
        // failure into the new bucket and block valid matches like the
        // outer `*` in `cz*x` `*foo***bar***baz` (closer `***` partial
        // remainder 1 should still reach the opener at offset 2).
        let lowerbound = min(
            self.stack.len(),
            self.get_lowerbound(c, current_count, both),
        );
        let res = self.stack[lowerbound..]
            .iter()
            .cloned()
            .enumerate()
            .rfind(|(_, el)| {
                if (c == b'~' || c == b'^') && run_length != el.run_length {
                    return false;
                }
                // Rule 9 (mod-3): for `*`/`_`, the openers on the stack are
                // checked against the *current* lengths — `el.count` reflects
                // remaining-after-partial-consumption when an opener has been
                // re-pushed, and `current_count` is the remaining closer.
                el.c == c
                    && (!both && !el.both
                        || !(current_count + el.count).is_multiple_of(3)
                        || current_count.is_multiple_of(3))
            });

        if let Some((matching_ix, matching_el)) = res {
            let matching_ix = matching_ix + lowerbound;
            for el in &self.stack[(matching_ix + 1)..] {
                for i in 0..el.count {
                    tree[el.start + i].item.body = ItemBody::Text {
                        backslash_escaped: false,
                    };
                }
            }
            self.truncate(matching_ix);
            Some(matching_el)
        } else {
            // For `*`/`_`, the lower-bound optimisation is safe because their
            // matching rule (CM "rule of three") is monotonic across future
            // closers with the same count. Tildes/carets match strictly by
            // equal run-length, so a failure at run-length 2 must not close
            // the door on a later run-length 1 closer matching an earlier
            // run-length 1 opener still on the stack. Key the bound by
            // `current_count` (the post-partial-consumption length) so it
            // applies only to closers whose remaining bucket actually
            // shares this failure mode.
            if c != b'~' && c != b'^' {
                self.set_lowerbound(c, current_count, both, self.stack.len());
            }
            None
        }
    }

    fn trim_lower_bound(&mut self, ix: usize) {
        self.lower_bounds[ix] = self.lower_bounds[ix].min(self.stack.len());
    }

    fn push(&mut self, el: InlineEl) {
        if el.c == b'~' {
            self.trim_lower_bound(InlineStack::TILDES);
        } else if el.c == b'^' {
            self.trim_lower_bound(InlineStack::CIRCUMFLEXES);
        }
        self.stack.push(el)
    }
}

#[derive(Debug, Clone)]
enum RefScan<'a> {
    // label, source ix of label end
    LinkLabel(CowStr<'a>, usize),
    // contains next node index
    Collapsed(Option<TreeIndex>),
    UnexpectedFootnote,
    Failed,
    // `[text][...]` where `[...]` started but is an invalid label
    // (e.g. contains unescaped `[`). The shortcut form for `[text]` is
    // suppressed because the spec says a shortcut link must NOT be
    // followed by `[` — even if that `[` doesn't form a valid label.
    FailedInvalidLabel,
}

/// Skips forward within a block to a node which spans (ends inclusive) the given
/// index into the source.
fn scan_nodes_to_ix(
    tree: &Tree<Item>,
    mut node: Option<TreeIndex>,
    ix: usize,
) -> Option<TreeIndex> {
    while let Some(node_ix) = node {
        let item = tree[node_ix].item;
        // A zero-width node *at* `ix` must not be skipped: callers splice the
        // result back in as the tail, so skipping drops it from the tree.
        if item.end <= ix && item.start < ix {
            node = tree[node_ix].next;
        } else {
            break;
        }
    }
    node
}

/// Scans an inline link label, which cannot be interrupted.
/// Returns number of bytes (including brackets) and label on success.
fn scan_link_label<'text>(
    tree: &Tree<Item>,
    text: &'text str,
    options: Options,
) -> Option<(usize, ReferenceLabel<'text>)> {
    let bytes = text.as_bytes();
    if bytes.len() < 2 || bytes[0] != b'[' {
        return None;
    }
    let linebreak_handler = |bytes: &[u8]| Some(skip_container_prefixes(tree, bytes, options));
    if options.contains(Options::ENABLE_FOOTNOTES)
        && b'^' == bytes[1]
        && bytes.get(2) != Some(&b']')
    {
        // GFM footnote labels don't wrap across line breaks.
        let linebreak_handler: &dyn Fn(&[u8]) -> Option<usize> = &|_| None;
        if let Some((byte_index, cow)) =
            scan_link_label_rest(&text[2..], linebreak_handler, tree.is_in_table())
        {
            return Some((byte_index + 2, ReferenceLabel::Footnote(cow)));
        }
    }
    let (byte_index, cow) =
        scan_link_label_rest(&text[1..], &linebreak_handler, tree.is_in_table())?;
    Some((byte_index + 1, ReferenceLabel::Link(cow)))
}

fn scan_reference<'b>(
    tree: &Tree<Item>,
    text: &'b str,
    cur: Option<TreeIndex>,
    options: Options,
) -> RefScan<'b> {
    let cur_ix = match cur {
        None => return RefScan::Failed,
        Some(cur_ix) => cur_ix,
    };
    let start = tree[cur_ix].item.start;
    let tail = &text.as_bytes()[start..];

    // If the `[` opening the candidate label was escaped in source
    // (preceded by an odd run of backslashes), it's a literal `[` and
    // can't start a reference label. Without this check the label
    // scanner walks raw source, which doesn't know that pulldown-cmark
    // already absorbed the `\` into a backslash-escape token, and it
    // would falsely consume `\[foo]` as `[foo]`.
    if tail.first() == Some(&b'[') && start > 0 {
        let src = text.as_bytes();
        let mut backslashes = 0usize;
        let mut j = start;
        while j > 0 && src[j - 1] == b'\\' {
            backslashes += 1;
            j -= 1;
        }
        if backslashes % 2 == 1 {
            return RefScan::Failed;
        }
    }

    if tail.starts_with(b"[]") {
        // The trailing `]` of the collapsed reference must already exist as a
        // tree node — pulldown-cmark emits each bracket as its own item, and
        // we only reach here when `tail` already contains `]`. Defensive
        // fallback to `Failed` if that invariant is somehow broken.
        let Some(closing_node) = tree[cur_ix].next else {
            return RefScan::Failed;
        };
        RefScan::Collapsed(tree[closing_node].next)
    } else {
        let label = scan_link_label(tree, &text[start..], options);
        match label {
            Some((ix, ReferenceLabel::Link(label))) => RefScan::LinkLabel(label, start + ix),
            Some((_ix, ReferenceLabel::Footnote(_label))) => RefScan::UnexpectedFootnote,
            None => {
                // If `[text]` is followed by `[` that looked like a label
                // opener, the shortcut form is suppressed even though the
                // label parse failed (CommonMark requires shortcut links
                // not be followed by `[`).
                if tail.starts_with(b"[") {
                    RefScan::FailedInvalidLabel
                } else {
                    RefScan::Failed
                }
            }
        }
    }
}

#[derive(Clone, Default)]
struct LinkStack {
    inner: Vec<LinkStackEl>,
    disabled_ix: usize,
}

impl LinkStack {
    fn is_empty(&self) -> bool {
        self.inner.is_empty()
    }

    fn push(&mut self, el: LinkStackEl) {
        self.inner.push(el);
    }

    fn pop(&mut self) -> Option<LinkStackEl> {
        let el = self.inner.pop();
        self.disabled_ix = core::cmp::min(self.disabled_ix, self.inner.len());
        el
    }

    fn clear(&mut self) {
        self.inner.clear();
        self.disabled_ix = 0;
    }

    fn disable_all_links(&mut self) {
        for el in &mut self.inner[self.disabled_ix..] {
            if el.ty == LinkStackTy::Link {
                el.ty = LinkStackTy::Disabled;
            }
        }
        self.disabled_ix = self.inner.len();
    }
}

#[derive(Clone, Debug)]
struct LinkStackEl {
    node: TreeIndex,
    ty: LinkStackTy,
}

#[derive(PartialEq, Clone, Debug)]
enum LinkStackTy {
    Link,
    Image,
    Disabled,
}

/// Contains the destination URL, title and source span of a reference definition.
#[derive(Clone, Debug)]
pub struct LinkDef<'a> {
    pub dest: CowStr<'a>,
    pub title: Option<CowStr<'a>>,
    pub span: Range<usize>,
}

impl<'a> LinkDef<'a> {
    pub fn into_static(self) -> LinkDef<'static> {
        LinkDef {
            dest: self.dest.into_static(),
            title: self.title.map(|s| s.into_static()),
            span: self.span,
        }
    }
}

/// Contains the destination URL, title and source span of a reference definition.
#[derive(Clone, Debug)]
pub struct FootnoteDef {
    pub use_count: usize,
}

/// Tracks tree indices of code span delimiters of each length. It should prevent
/// quadratic scanning behaviours by providing (amortized) constant time lookups.
struct CodeDelims {
    inner: FxHashMap<usize, VecDeque<TreeIndex>>,
    seen_first: bool,
}

impl CodeDelims {
    fn new() -> Self {
        Self {
            inner: Default::default(),
            seen_first: false,
        }
    }

    fn insert(&mut self, count: usize, ix: TreeIndex) {
        if self.seen_first {
            self.inner.entry(count).or_default().push_back(ix);
        } else {
            // Skip the first insert, since that delimiter will always
            // be an opener and not a closer.
            self.seen_first = true;
        }
    }

    fn is_populated(&self) -> bool {
        !self.inner.is_empty()
    }

    fn find(&mut self, open_ix: TreeIndex, count: usize) -> Option<TreeIndex> {
        while let Some(ix) = self.inner.get_mut(&count)?.pop_front() {
            if ix > open_ix {
                return Some(ix);
            }
        }
        None
    }

    fn clear(&mut self) {
        self.inner.clear();
        self.seen_first = false;
    }
}

/// Tracks brace contexts and delimiter length for math delimiters.
/// Provides amortized constant-time lookups.
struct MathDelims {
    inner: FxHashMap<u8, VecDeque<(TreeIndex, bool, bool)>>,
}

impl MathDelims {
    fn new() -> Self {
        Self {
            inner: Default::default(),
        }
    }

    fn clear(&mut self) {
        self.inner.clear();
    }
}

#[derive(Copy, Clone, PartialEq, Eq, Debug)]
pub(crate) struct LinkIndex(u32);

#[derive(Copy, Clone, PartialEq, Eq, Debug)]
pub(crate) struct CowIndex(u32);

#[derive(Copy, Clone, PartialEq, Eq, Debug)]
pub(crate) struct AlignmentIndex(u32);

#[derive(Copy, Clone, PartialEq, Eq, Debug)]
pub(crate) struct HeadingIndex(NonZeroU32);

#[cfg(feature = "mdx")]
#[derive(Copy, Clone, PartialEq, Eq, Debug)]
pub(crate) struct JsxElementIndex(u32);

#[derive(Copy, Clone, PartialEq, Eq, Debug)]
pub(crate) struct DirectiveIndex(u32);

#[derive(Copy, Clone, PartialEq, Eq, Debug)]
pub(crate) struct AutolinkCandidateIndex(u32);

#[derive(Copy, Clone, PartialEq, Eq, Debug)]
pub(crate) struct FencedInfoIndex(u32);

/// A GFM autolink literal the first pass found but did not commit to. The
/// `Link` is allocated up front so firing is only a body swap.
#[derive(Copy, Clone, Debug)]
pub(crate) struct AutolinkCandidate {
    /// Precedes the trigger byte when the email scan walks back over the local
    /// part.
    pub start: usize,
    pub end: usize,
    pub link: LinkIndex,
}

/// A parsed JSX attribute.
#[cfg(feature = "mdx")]
#[derive(Debug, Clone)]
pub(crate) enum JsxAttr<'a> {
    Boolean(CowStr<'a>),
    Literal(CowStr<'a>, CowStr<'a>),
    /// `name={value}`. The two `usize`s are the byte range of `value` within
    /// the opening tag, so a parse error can be validated against the verbatim
    /// source slice and resolved to an exact source position.
    Expression(CowStr<'a>, CowStr<'a>, usize, usize),
    /// `{...value}`. The two `usize`s are the byte range of `value` (including
    /// the leading `...`) within the opening tag.
    Spread(CowStr<'a>, usize, usize),
}

#[cfg(feature = "mdx")]
impl<'a> JsxAttr<'a> {
    pub fn into_static(self) -> JsxAttr<'static> {
        match self {
            JsxAttr::Boolean(n) => JsxAttr::Boolean(n.into_static()),
            JsxAttr::Literal(n, v) => JsxAttr::Literal(n.into_static(), v.into_static()),
            JsxAttr::Expression(n, v, start, end) => {
                JsxAttr::Expression(n.into_static(), v.into_static(), start, end)
            }
            JsxAttr::Spread(v, start, end) => JsxAttr::Spread(v.into_static(), start, end),
        }
    }
}

/// Pre-parsed JSX element data (name + attributes + tag classification).
#[cfg(feature = "mdx")]
#[derive(Debug, Clone)]
pub(crate) struct JsxElementData<'a> {
    pub name: CowStr<'a>,
    pub attrs: Vec<JsxAttr<'a>>,
    pub raw: CowStr<'a>,
    pub is_closing: bool,
    pub is_self_closing: bool,
}

#[cfg(feature = "mdx")]
impl<'a> JsxElementData<'a> {
    pub fn into_static(self) -> JsxElementData<'static> {
        JsxElementData {
            name: self.name.into_static(),
            attrs: self.attrs.into_iter().map(|a| a.into_static()).collect(),
            raw: self.raw.into_static(),
            is_closing: self.is_closing,
            is_self_closing: self.is_self_closing,
        }
    }
}

#[derive(Debug, Clone)]
pub(crate) struct DirectiveAttrData<'a> {
    pub name: CowStr<'a>,
    pub attributes: Vec<(CowStr<'a>, CowStr<'a>)>,
    pub label_start: usize,
    pub label_end: usize,
    /// Cols of leading whitespace before `:::` on the opening line, after
    /// outer-container prefix stripping. Mirrors micromark-extension-directive's
    /// `initialSize`, which controls how much the directive body's per-line
    /// linePrefix is stripped (up to `initialSize + 1` cols). Only meaningful
    /// for container directives — leaf/text directives leave this 0.
    pub initial_size: u8,
}

#[derive(Clone)]
pub(crate) struct Allocations<'a> {
    pub refdefs: RefDefs<'a>,
    /// Every refdef occurrence in source order, including duplicates that
    /// `refdefs` drops (it's a map and only keeps the first per label, since
    /// resolution picks the first match per CommonMark). Used to emit every
    /// definition as its own mdast `definition` node.
    pub refdefs_all: Vec<(LinkLabel<'a>, LinkDef<'a>)>,
    pub footdefs: FootnoteDefs<'a>,
    links: Vec<(LinkType, CowStr<'a>, CowStr<'a>, CowStr<'a>)>,
    cows: Vec<CowStr<'a>>,
    alignments: Vec<Vec<Alignment>>,
    headings: Vec<HeadingAttributes<'a>>,
    #[cfg(feature = "mdx")]
    jsx_elements: Vec<JsxElementData<'a>>,
    directives: Vec<DirectiveAttrData<'a>>,
    autolink_candidates: Vec<AutolinkCandidate>,
    /// Fence info string paired with the byte length of its `lang` part within
    /// the decoded string. The split is taken on raw source, so a character
    /// reference decoding to whitespace does not move it.
    fenced_infos: Vec<(CowStr<'a>, u32)>,
}

/// Used by the heading attributes extension.
#[derive(Clone)]
pub(crate) struct HeadingAttributes<'a> {
    pub id: Option<CowStr<'a>>,
    pub classes: Vec<CowStr<'a>>,
    pub attrs: Vec<(CowStr<'a>, Option<CowStr<'a>>)>,
}

/// Keeps track of the reference definitions defined in the document.
#[derive(Clone, Default, Debug)]
pub struct RefDefs<'input>(pub(crate) FxHashMap<LinkLabel<'input>, LinkDef<'input>>);

/// Keeps track of the footnote definitions defined in the document.
#[derive(Clone, Default, Debug)]
pub struct FootnoteDefs<'input>(pub(crate) FxHashMap<FootnoteLabel<'input>, FootnoteDef>);

impl<'input, 'b, 's> RefDefs<'input>
where
    's: 'b,
{
    /// Performs a lookup on reference label using unicode case folding.
    pub fn get(&'s self, key: &'b str) -> Option<&'b LinkDef<'input>> {
        self.0.get(&UniCase::new(key.into()))
    }

    /// Provides an iterator over all the document's reference definitions.
    pub fn iter(
        &'s self,
    ) -> impl Iterator<Item = (&'s str, &'s LinkDef<'input>)> + use<'s, 'input> {
        self.0.iter().map(|(k, v)| (k.as_ref(), v))
    }
}

impl<'input, 'b, 's> FootnoteDefs<'input>
where
    's: 'b,
{
    /// Performs a lookup on reference label using unicode case folding.
    pub fn contains(&'s self, key: &'b str) -> bool {
        self.0.contains_key(&UniCase::new(key.into()))
    }
    /// Performs a lookup on reference label using unicode case folding.
    pub fn get_mut(&'s mut self, key: CowStr<'input>) -> Option<&'s mut FootnoteDef> {
        self.0.get_mut(&UniCase::new(key))
    }
}

impl<'a> Allocations<'a> {
    pub fn new() -> Self {
        Self {
            refdefs: RefDefs::default(),
            refdefs_all: Vec::new(),
            footdefs: FootnoteDefs::default(),
            links: Vec::with_capacity(128),
            cows: Vec::new(),
            alignments: Vec::new(),
            headings: Vec::new(),
            #[cfg(feature = "mdx")]
            jsx_elements: Vec::new(),
            directives: Vec::new(),
            autolink_candidates: Vec::new(),
            fenced_infos: Vec::new(),
        }
    }

    pub fn allocate_fenced_info(&mut self, info: CowStr<'a>, lang_len: u32) -> FencedInfoIndex {
        let ix = self.fenced_infos.len() as u32;
        self.fenced_infos.push((info, lang_len));
        FencedInfoIndex(ix)
    }

    pub fn take_fenced_info(&mut self, ix: FencedInfoIndex) -> (CowStr<'a>, u32) {
        core::mem::replace(&mut self.fenced_infos[ix.0 as usize], ("".into(), 0))
    }

    pub fn allocate_autolink_candidate(
        &mut self,
        candidate: AutolinkCandidate,
    ) -> AutolinkCandidateIndex {
        let ix = self.autolink_candidates.len() as u32;
        self.autolink_candidates.push(candidate);
        AutolinkCandidateIndex(ix)
    }

    pub fn allocate_cow(&mut self, cow: CowStr<'a>) -> CowIndex {
        let ix = self.cows.len() as u32;
        self.cows.push(cow);
        CowIndex(ix)
    }

    pub fn allocate_link(
        &mut self,
        ty: LinkType,
        url: CowStr<'a>,
        title: CowStr<'a>,
        id: CowStr<'a>,
    ) -> LinkIndex {
        let ix = self.links.len() as u32;
        self.links.push((ty, url, title, id));
        LinkIndex(ix)
    }

    pub fn allocate_alignment(&mut self, alignment: Vec<Alignment>) -> AlignmentIndex {
        let ix = self.alignments.len() as u32;
        self.alignments.push(alignment);
        AlignmentIndex(ix)
    }

    pub fn allocate_heading(&mut self, attrs: HeadingAttributes<'a>) -> HeadingIndex {
        let ix = self.headings.len() as u32;
        self.headings.push(attrs);
        // This won't panic. `self.headings.len()` can't be `u32::MAX` since
        // such a long Vec cannot fit in memory.
        let ix_nonzero = NonZeroU32::new(ix.wrapping_add(1)).expect("too many headings");
        HeadingIndex(ix_nonzero)
    }

    pub fn take_cow(&mut self, ix: CowIndex) -> CowStr<'a> {
        core::mem::replace(&mut self.cows[ix.0 as usize], "".into())
    }

    pub fn take_link(&mut self, ix: LinkIndex) -> (LinkType, CowStr<'a>, CowStr<'a>, CowStr<'a>) {
        let default_link = (LinkType::ShortcutUnknown, "".into(), "".into(), "".into());
        core::mem::replace(&mut self.links[ix.0 as usize], default_link)
    }

    pub fn take_alignment(&mut self, ix: AlignmentIndex) -> Vec<Alignment> {
        core::mem::take(&mut self.alignments[ix.0 as usize])
    }

    #[cfg(feature = "mdx")]
    pub fn allocate_jsx_element(&mut self, data: JsxElementData<'a>) -> JsxElementIndex {
        let ix = self.jsx_elements.len() as u32;
        self.jsx_elements.push(data);
        JsxElementIndex(ix)
    }

    pub fn allocate_directive(&mut self, data: DirectiveAttrData<'a>) -> DirectiveIndex {
        let ix = self.directives.len() as u32;
        self.directives.push(data);
        DirectiveIndex(ix)
    }

    pub fn take_directive(&mut self, ix: DirectiveIndex) -> DirectiveAttrData<'a> {
        core::mem::replace(
            &mut self.directives[ix.0 as usize],
            DirectiveAttrData {
                name: "".into(),
                attributes: Vec::new(),
                label_start: 0,
                label_end: 0,
                initial_size: 0,
            },
        )
    }

    pub fn directive_ref(&self, ix: DirectiveIndex) -> &DirectiveAttrData<'a> {
        &self.directives[ix.0 as usize]
    }

    #[cfg(feature = "mdx")]
    pub fn take_jsx_element(&mut self, ix: JsxElementIndex) -> JsxElementData<'a> {
        core::mem::replace(
            &mut self.jsx_elements[ix.0 as usize],
            JsxElementData {
                name: "".into(),
                attrs: Vec::new(),
                raw: "".into(),
                is_closing: false,
                is_self_closing: false,
            },
        )
    }
}

impl<'a> Index<CowIndex> for Allocations<'a> {
    type Output = CowStr<'a>;

    fn index(&self, ix: CowIndex) -> &Self::Output {
        self.cows.index(ix.0 as usize)
    }
}

impl<'a> Index<LinkIndex> for Allocations<'a> {
    type Output = (LinkType, CowStr<'a>, CowStr<'a>, CowStr<'a>);

    fn index(&self, ix: LinkIndex) -> &Self::Output {
        self.links.index(ix.0 as usize)
    }
}

impl<'a> Index<AutolinkCandidateIndex> for Allocations<'a> {
    type Output = AutolinkCandidate;

    fn index(&self, ix: AutolinkCandidateIndex) -> &Self::Output {
        self.autolink_candidates.index(ix.0 as usize)
    }
}

impl<'a> Index<AlignmentIndex> for Allocations<'a> {
    type Output = Vec<Alignment>;

    fn index(&self, ix: AlignmentIndex) -> &Self::Output {
        self.alignments.index(ix.0 as usize)
    }
}

impl<'a> Index<HeadingIndex> for Allocations<'a> {
    type Output = HeadingAttributes<'a>;

    fn index(&self, ix: HeadingIndex) -> &Self::Output {
        self.headings.index(ix.0.get() as usize - 1)
    }
}

/// A struct containing information on the reachability of certain inline HTML
/// elements. In particular, for cdata elements (`<![CDATA[`), processing
/// elements (`<?`) and declarations (`<!DECLARATION`). The respectives usizes
/// represent the indices before which a scan will always fail and can hence
/// be skipped.
#[derive(Clone, Default)]
pub(crate) struct HtmlScanGuard {
    pub cdata: usize,
    pub processing: usize,
    pub declaration: usize,
    pub comment: usize,
}

/// Trait to customize [`Parser`] behavior with callbacks. See [`Parser::new_with_callbacks`].
///
/// All methods have a default implementation, so you can choose which ones to override.
pub trait ParserCallbacks<'input> {
    /// Potentially provide a custom definition for a broken link.
    ///
    /// In case the parser encounters any potential links that have a broken
    /// reference (e.g `[foo]` when there is no `[foo]: ` entry at the bottom)
    /// this callback will be called with information about the reference,
    /// and the returned pair will be used as the link URL and title if it is not
    /// `None`.
    fn handle_broken_link(
        &mut self,
        #[allow(unused_variables)] link: BrokenLink<'input>,
    ) -> Option<(CowStr<'input>, CowStr<'input>)> {
        None
    }
}

/// Wrapper to implement [`ParserCallbacks::handle_broken_link`] with a closure.
///
/// Used internally by [`Parser::new_with_broken_link_callback`].
#[allow(missing_debug_implementations)]
pub struct BrokenLinkCallback<F>(Option<F>);

impl<'input, F> ParserCallbacks<'input> for BrokenLinkCallback<F>
where
    F: FnMut(BrokenLink<'input>) -> Option<(CowStr<'input>, CowStr<'input>)>,
{
    fn handle_broken_link(
        &mut self,
        link: BrokenLink<'input>,
    ) -> Option<(CowStr<'input>, CowStr<'input>)> {
        self.0.as_mut().and_then(|cb| cb(link))
    }
}

impl<'input> ParserCallbacks<'input> for Box<dyn ParserCallbacks<'input>> {
    fn handle_broken_link(
        &mut self,
        link: BrokenLink<'input>,
    ) -> Option<(CowStr<'input>, CowStr<'input>)> {
        (**self).handle_broken_link(link)
    }
}

/// [Parser] callbacks that do nothing.
///
/// Used when no custom callbacks are provided.
#[allow(missing_debug_implementations)]
pub struct DefaultParserCallbacks;

impl<'input> ParserCallbacks<'input> for DefaultParserCallbacks {}

/// Markdown event and source range iterator.
///
/// Generates tuples where the first element is the markdown event and the second
/// is a the corresponding range in the source string.
///
/// Constructed from a `Parser` using its
/// [`into_offset_iter`](struct.Parser.html#method.into_offset_iter) method.
#[derive(Debug)]
pub struct OffsetIter<'a, CB> {
    parser: Parser<'a, CB>,
}

impl<'a, CB: ParserCallbacks<'a>> OffsetIter<'a, CB> {
    /// Returns a reference to the internal reference definition tracker.
    pub fn reference_definitions(&self) -> &RefDefs<'_> {
        self.parser.reference_definitions()
    }

    /// Returns MDX validation errors collected during parsing.
    pub fn mdx_errors(&self) -> &[(usize, String)] {
        self.parser.mdx_errors()
    }
}

impl<'a, CB: ParserCallbacks<'a>> Iterator for OffsetIter<'a, CB> {
    type Item = (Event<'a>, Range<usize>);

    fn next(&mut self) -> Option<Self::Item> {
        self.parser
            .inner
            .next_event_range(&mut self.parser.callbacks)
    }
}

impl<'a, CB: ParserCallbacks<'a>> Iterator for Parser<'a, CB> {
    type Item = Event<'a>;

    fn next(&mut self) -> Option<Event<'a>> {
        self.inner
            .next_event_range(&mut self.callbacks)
            .map(|(event, _range)| event)
    }
}

impl<'a, CB: ParserCallbacks<'a>> FusedIterator for Parser<'a, CB> {}

impl<'input> ParserInner<'input> {
    fn next_event_range(
        &mut self,
        callbacks: &mut dyn ParserCallbacks<'input>,
    ) -> Option<(Event<'input>, Range<usize>)> {
        match self.tree.cur() {
            None => {
                let ix = self.tree.pop()?;
                let ix = if matches!(self.tree[ix].item.body, ItemBody::TightParagraph) {
                    // tight paragraphs emit nothing
                    self.tree.next_sibling(ix);
                    return self.next_event_range(callbacks);
                } else {
                    ix
                };
                let tag_end = body_to_tag_end(&self.tree[ix].item.body);
                self.tree.next_sibling(ix);
                let span = self.tree[ix].item.start..self.tree[ix].item.end;
                debug_assert!(span.start <= span.end);
                Some((Event::End(tag_end), span))
            }
            Some(cur_ix) => {
                let cur_ix = if matches!(self.tree[cur_ix].item.body, ItemBody::TightParagraph) {
                    // tight paragraphs emit nothing
                    self.tree.push();
                    self.tree.cur().unwrap()
                } else {
                    cur_ix
                };
                if self.tree[cur_ix].item.body.is_maybe_inline() {
                    self.handle_inline(callbacks);
                }

                let node = self.tree[cur_ix];
                let item = node.item;
                let event = item_to_event(item, self.text, &mut self.allocs);
                if let Event::Start(..) = event {
                    self.tree.push();
                } else {
                    self.tree.next_sibling(cur_ix);
                }
                debug_assert!(item.start <= item.end);
                Some((event, item.start..item.end))
            }
        }
    }
}

fn body_to_tag_end(body: &ItemBody) -> TagEnd {
    match *body {
        ItemBody::Paragraph => TagEnd::Paragraph,
        ItemBody::Emphasis => TagEnd::Emphasis,
        ItemBody::Superscript => TagEnd::Superscript,
        ItemBody::Subscript => TagEnd::Subscript,
        ItemBody::Strong => TagEnd::Strong,
        ItemBody::Strikethrough => TagEnd::Strikethrough,
        ItemBody::Link(..) => TagEnd::Link,
        ItemBody::Image(..) => TagEnd::Image,
        ItemBody::Heading(level, _) => TagEnd::Heading(level),
        ItemBody::IndentCodeBlock(..) | ItemBody::FencedCodeBlock(..) | ItemBody::MathBlock(..) => {
            TagEnd::CodeBlock
        }
        ItemBody::ContainerDirective(..) => TagEnd::Directive(DirectiveKind::Container),
        ItemBody::LeafDirective(..) => TagEnd::Directive(DirectiveKind::Leaf),
        ItemBody::TextDirective(..) => TagEnd::Directive(DirectiveKind::Text),
        ItemBody::BlockQuote(kind) => TagEnd::BlockQuote(kind),
        ItemBody::HtmlBlock(_) => TagEnd::HtmlBlock,
        ItemBody::List(_, c, _) => {
            let is_ordered = c == b'.' || c == b')';
            TagEnd::List(is_ordered)
        }
        ItemBody::ListItem(_, _) => TagEnd::Item,
        ItemBody::TableHead => TagEnd::TableHead,
        ItemBody::TableCell => TagEnd::TableCell,
        ItemBody::TableRow => TagEnd::TableRow,
        ItemBody::Table(..) => TagEnd::Table,
        ItemBody::FootnoteDefinition(..) => TagEnd::FootnoteDefinition,
        ItemBody::MetadataBlock(kind) => TagEnd::MetadataBlock(kind),
        ItemBody::DefinitionList(_) => TagEnd::DefinitionList,
        ItemBody::DefinitionListTitle => TagEnd::DefinitionListTitle,
        ItemBody::DefinitionListDefinition(..) => TagEnd::DefinitionListDefinition,
        #[cfg(feature = "mdx")]
        ItemBody::MdxJsxFlowElement(..) => TagEnd::MdxJsxFlowElement,
        #[cfg(feature = "mdx")]
        ItemBody::MdxJsxTextElement(..) => TagEnd::MdxJsxTextElement,
        _ => panic!("unexpected item body {:?}", body),
    }
}

fn item_to_event<'a>(item: Item, text: &'a str, allocs: &mut Allocations<'a>) -> Event<'a> {
    let tag = match item.body {
        ItemBody::Text { .. } => return Event::Text(text[item.start..item.end].into()),
        ItemBody::Code(cow_ix) => return Event::Code(allocs.take_cow(cow_ix)),
        ItemBody::SynthesizeText(cow_ix) => return Event::Text(allocs.take_cow(cow_ix)),
        ItemBody::SynthesizeChar(c) => return Event::Text(c.into()),
        ItemBody::HtmlBlock(_) => Tag::HtmlBlock,
        ItemBody::Html => return Event::Html(text[item.start..item.end].into()),
        ItemBody::InlineHtml => return Event::InlineHtml(text[item.start..item.end].into()),
        ItemBody::OwnedInlineHtml(cow_ix) => return Event::InlineHtml(allocs.take_cow(cow_ix)),
        ItemBody::SoftBreak => return Event::SoftBreak,
        ItemBody::HardBreak(_) => return Event::HardBreak,
        ItemBody::FootnoteReference(cow_ix) => {
            return Event::FootnoteReference(allocs.take_cow(cow_ix));
        }
        ItemBody::TaskListMarker(checked) => return Event::TaskListMarker(checked),
        ItemBody::Rule => return Event::Rule,
        ItemBody::Paragraph => Tag::Paragraph,
        ItemBody::Emphasis => Tag::Emphasis,
        ItemBody::Superscript => Tag::Superscript,
        ItemBody::Subscript => Tag::Subscript,
        ItemBody::Strong => Tag::Strong,
        ItemBody::Strikethrough => Tag::Strikethrough,
        ItemBody::Link(link_ix) => {
            let (link_type, dest_url, title, id) = allocs.take_link(link_ix);
            Tag::Link {
                link_type,
                dest_url,
                title,
                id,
            }
        }
        ItemBody::Image(link_ix) => {
            let (link_type, dest_url, title, id) = allocs.take_link(link_ix);
            Tag::Image {
                link_type,
                dest_url,
                title,
                id,
            }
        }
        ItemBody::Heading(level, Some(heading_ix)) => {
            let HeadingAttributes { id, classes, attrs } = allocs.index(heading_ix);
            Tag::Heading {
                level,
                id: id.clone(),
                classes: classes.clone(),
                attrs: attrs.clone(),
            }
        }
        ItemBody::Heading(level, None) => Tag::Heading {
            level,
            id: None,
            classes: Vec::new(),
            attrs: Vec::new(),
        },
        ItemBody::MathBlock(cow_ix) => {
            Tag::CodeBlock(CodeBlockKind::Fenced(allocs.take_cow(cow_ix)))
        }
        ItemBody::FencedCodeBlock(info_ix) => {
            Tag::CodeBlock(CodeBlockKind::Fenced(allocs.take_fenced_info(info_ix).0))
        }
        ItemBody::IndentCodeBlock(..) => Tag::CodeBlock(CodeBlockKind::Indented),
        ItemBody::ContainerDirective(_, dir_ix)
        | ItemBody::LeafDirective(dir_ix)
        | ItemBody::TextDirective(dir_ix) => {
            let kind = match item.body {
                ItemBody::ContainerDirective(..) => DirectiveKind::Container,
                ItemBody::LeafDirective(..) => DirectiveKind::Leaf,
                _ => DirectiveKind::Text,
            };
            let dir = allocs.take_directive(dir_ix);
            Tag::Directive {
                kind,
                name: dir.name,
                attributes: dir.attributes,
            }
        }
        ItemBody::BlockQuote(kind) => Tag::BlockQuote(kind),
        ItemBody::List(is_tight, c, listitem_start) => {
            if c == b'.' || c == b')' {
                Tag::List(Some(listitem_start as u64), is_tight)
            } else {
                Tag::List(None, is_tight)
            }
        }
        ItemBody::ListItem(_, _) => Tag::Item,
        ItemBody::TableHead => Tag::TableHead,
        ItemBody::TableCell => Tag::TableCell,
        ItemBody::TableRow => Tag::TableRow,
        ItemBody::Table(alignment_ix) => Tag::Table(allocs.take_alignment(alignment_ix)),
        ItemBody::FootnoteDefinition(cow_ix) => Tag::FootnoteDefinition(allocs.take_cow(cow_ix)),
        ItemBody::MetadataBlock(kind) => Tag::MetadataBlock(kind),
        ItemBody::Math(cow_ix, is_display) => {
            return if is_display {
                Event::DisplayMath(allocs.take_cow(cow_ix))
            } else {
                Event::InlineMath(allocs.take_cow(cow_ix))
            };
        }
        ItemBody::DefinitionList(_) => Tag::DefinitionList,
        ItemBody::DefinitionListTitle => Tag::DefinitionListTitle,
        ItemBody::DefinitionListDefinition(..) => Tag::DefinitionListDefinition,
        #[cfg(feature = "mdx")]
        ItemBody::MdxJsxFlowElement(jsx_ix) => {
            let jsx = allocs.take_jsx_element(jsx_ix);
            Tag::MdxJsxFlowElement(jsx.raw)
        }
        #[cfg(feature = "mdx")]
        ItemBody::MdxJsxTextElement(jsx_ix) => {
            let jsx = allocs.take_jsx_element(jsx_ix);
            Tag::MdxJsxTextElement(jsx.raw)
        }
        #[cfg(feature = "mdx")]
        ItemBody::MdxFlowExpression(cow_ix) => {
            return Event::MdxFlowExpression(allocs.take_cow(cow_ix));
        }
        #[cfg(feature = "mdx")]
        ItemBody::MdxTextExpression(cow_ix) => {
            return Event::MdxTextExpression(allocs.take_cow(cow_ix));
        }
        #[cfg(feature = "mdx")]
        ItemBody::MdxEsm(cow_ix) => return Event::MdxEsm(allocs.take_cow(cow_ix)),
        _ => panic!("unexpected item body {:?}", item.body),
    };

    Event::Start(tag)
}

#[cfg(test)]
mod test {
    use alloc::{borrow::ToOwned, string::ToString, vec::Vec};

    use super::*;
    use crate::tree::Node;

    // TODO: move these tests to tests/html.rs?

    fn parser_with_extensions(text: &str) -> Parser<'_> {
        let mut opts = Options::empty();
        opts.insert(Options::ENABLE_TABLES);
        opts.insert(Options::ENABLE_FOOTNOTES);
        opts.insert(Options::ENABLE_STRIKETHROUGH);
        opts.insert(Options::ENABLE_SUPERSCRIPT);
        opts.insert(Options::ENABLE_SUBSCRIPT);
        opts.insert(Options::ENABLE_TASKLISTS);

        Parser::new_ext(text, opts)
    }

    #[test]
    #[cfg(target_pointer_width = "64")]
    fn node_size() {
        let node_size = core::mem::size_of::<Node<Item>>();
        assert_eq!(32, node_size);
    }

    #[test]
    #[cfg(target_pointer_width = "64")]
    fn body_size() {
        let body_size = core::mem::size_of::<ItemBody>();
        assert_eq!(8, body_size);
    }

    #[test]
    fn single_open_fish_bracket() {
        // dont crash
        assert_eq!(3, Parser::new("<").count());
    }

    #[test]
    fn lone_hashtag() {
        // dont crash
        assert_eq!(2, Parser::new("#").count());
    }

    #[test]
    fn lots_of_backslashes() {
        // dont crash
        Parser::new("\\\\\r\r").count();
        Parser::new("\\\r\r\\.\\\\\r\r\\.\\").count();
    }

    #[test]
    fn issue_1030() {
        let mut opts = Options::empty();
        opts.insert(Options::ENABLE_WIKILINKS);

        let parser = Parser::new_ext("For a new ferrari, [[Wikientry|click here]]!", opts);

        let offsets = parser
            .into_offset_iter()
            .map(|(_ev, range)| range)
            .collect::<Vec<_>>();
        let expected_offsets = vec![
            (0..44),  // Paragraph START
            (0..19),  // `For a new ferrari, `
            (19..43), // Wikilink START
            (31..41), // `click here`
            (19..43), // Wikilink END
            (43..44), // `!`
            (0..44),  // Paragraph END
        ];
        assert_eq!(offsets, expected_offsets);
    }

    #[test]
    fn issue_320() {
        // dont crash
        parser_with_extensions(":\r\t> |\r:\r\t> |\r").count();
    }

    #[test]
    fn issue_319() {
        // dont crash
        parser_with_extensions("|\r-]([^|\r-]([^").count();
        parser_with_extensions("|\r\r=][^|\r\r=][^car").count();
    }

    #[test]
    fn issue_303() {
        // dont crash
        parser_with_extensions("[^\r\ra]").count();
        parser_with_extensions("\r\r]Z[^\x00\r\r]Z[^\x00").count();
    }

    #[test]
    fn issue_313() {
        // dont crash
        parser_with_extensions("*]0[^\r\r*]0[^").count();
        parser_with_extensions("[^\r> `][^\r> `][^\r> `][").count();
    }

    #[test]
    fn issue_311() {
        // dont crash
        parser_with_extensions("\\\u{0d}-\u{09}\\\u{0d}-\u{09}").count();
    }

    #[test]
    fn issue_283() {
        let input = core::str::from_utf8(b"\xf0\x9b\xb2\x9f<td:^\xf0\x9b\xb2\x9f").unwrap();
        // dont crash
        parser_with_extensions(input).count();
    }

    #[test]
    fn issue_289() {
        // dont crash
        parser_with_extensions("> - \\\n> - ").count();
        parser_with_extensions("- \n\n").count();
    }

    #[test]
    fn issue_306() {
        // dont crash
        parser_with_extensions("*\r_<__*\r_<__*\r_<__*\r_<__").count();
    }

    #[test]
    fn issue_305() {
        // dont crash
        parser_with_extensions("_6**6*_*").count();
    }

    #[test]
    fn another_emphasis_panic() {
        parser_with_extensions("*__#_#__*").count();
    }

    #[test]
    fn offset_iter() {
        let event_offsets: Vec<_> = Parser::new("*hello* world")
            .into_offset_iter()
            .map(|(_ev, range)| range)
            .collect();
        let expected_offsets = vec![(0..13), (0..7), (1..6), (0..7), (7..13), (0..13)];
        assert_eq!(expected_offsets, event_offsets);
    }

    #[test]
    fn reference_link_offsets() {
        let range =
            Parser::new("# H1\n[testing][Some reference]\n\n[Some reference]: https://github.com")
                .into_offset_iter()
                .filter_map(|(ev, range)| match ev {
                    Event::Start(
                        Tag::Link {
                            link_type: LinkType::Reference,
                            ..
                        },
                        ..,
                    ) => Some(range),
                    _ => None,
                })
                .next()
                .unwrap();
        assert_eq!(5..30, range);
    }

    #[test]
    fn footnote_offsets() {
        let range = parser_with_extensions("Testing this[^1] out.\n\n[^1]: Footnote.")
            .into_offset_iter()
            .filter_map(|(ev, range)| match ev {
                Event::FootnoteReference(..) => Some(range),
                _ => None,
            })
            .next()
            .unwrap();
        assert_eq!(12..16, range);
    }

    #[test]
    fn footnote_offsets_exclamation() {
        let mut immediately_before_footnote = None;
        let range = parser_with_extensions("Testing this![^1] out.\n\n[^1]: Footnote.")
            .into_offset_iter()
            .filter_map(|(ev, range)| match ev {
                Event::FootnoteReference(..) => Some(range),
                _ => {
                    immediately_before_footnote = Some((ev, range));
                    None
                }
            })
            .next()
            .unwrap();
        assert_eq!(13..17, range);
        if let (Event::Text(exclamation), range_exclamation) =
            immediately_before_footnote.as_ref().unwrap()
        {
            assert_eq!("!", &exclamation[..]);
            assert_eq!(&(12..13), range_exclamation);
        } else {
            panic!("what came first, then? {immediately_before_footnote:?}");
        }
    }

    #[test]
    fn table_offset() {
        let markdown = "a\n\nTesting|This|Outtt\n--|:--:|--:\nSome Data|Other data|asdf";
        let event_offset = parser_with_extensions(markdown)
            .into_offset_iter()
            .map(|(_ev, range)| range)
            .nth(3)
            .unwrap();
        let expected_offset = 3..59;
        assert_eq!(expected_offset, event_offset);
    }

    #[test]
    fn table_cell_span() {
        let markdown = "a|b|c\n--|--|--\na|  |c";
        let event_offset = parser_with_extensions(markdown)
            .into_offset_iter()
            .filter_map(|(ev, span)| match ev {
                Event::Start(Tag::TableCell) => Some(span),
                _ => None,
            })
            .nth(4)
            .unwrap();
        // Cell span includes the leading `|` delimiter (matching remark).
        let expected_offset_start = "a|b|c\n--|--|--\na".len();
        assert_eq!(
            expected_offset_start..(expected_offset_start + 3),
            event_offset
        );
    }

    #[test]
    fn offset_iter_issue_378() {
        let event_offsets: Vec<_> = Parser::new("a [b](c) d")
            .into_offset_iter()
            .map(|(_ev, range)| range)
            .collect();
        let expected_offsets = vec![(0..10), (0..2), (2..8), (3..4), (2..8), (8..10), (0..10)];
        assert_eq!(expected_offsets, event_offsets);
    }

    #[test]
    fn offset_iter_issue_404() {
        let event_offsets: Vec<_> = Parser::new("###\n")
            .into_offset_iter()
            .map(|(_ev, range)| range)
            .collect();
        let expected_offsets = vec![(0..4), (0..4)];
        assert_eq!(expected_offsets, event_offsets);
    }

    #[test]
    fn broken_links_called_only_once() {
        for &(markdown, expected) in &[
            ("See also [`g()`][crate::g].", 1),
            ("See also [`g()`][crate::g][].", 1),
            ("[brokenlink1] some other node [brokenlink2]", 2),
        ] {
            let mut times_called = 0;
            let callback = &mut |_broken_link: BrokenLink| {
                times_called += 1;
                None
            };
            let parser =
                Parser::new_with_broken_link_callback(markdown, Options::empty(), Some(callback));
            for _ in parser {}
            assert_eq!(times_called, expected);
        }
    }

    #[test]
    fn simple_broken_link_callback() {
        let test_str = "This is a link w/o def: [hello][world]";
        let mut callback = |broken_link: BrokenLink| {
            assert_eq!("world", broken_link.reference.as_ref());
            assert_eq!(&test_str[broken_link.span], "[hello][world]");
            let url = "YOLO".into();
            let title = "SWAG".to_owned().into();
            Some((url, title))
        };
        let parser =
            Parser::new_with_broken_link_callback(test_str, Options::empty(), Some(&mut callback));
        let mut link_tag_count = 0;
        for (typ, url, title, id) in parser.filter_map(|event| match event {
            Event::Start(Tag::Link {
                link_type,
                dest_url,
                title,
                id,
            }) => Some((link_type, dest_url, title, id)),
            _ => None,
        }) {
            link_tag_count += 1;
            assert_eq!(typ, LinkType::ReferenceUnknown);
            assert_eq!(url.as_ref(), "YOLO");
            assert_eq!(title.as_ref(), "SWAG");
            assert_eq!(id.as_ref(), "world");
        }
        assert!(link_tag_count > 0);
    }

    #[test]
    fn code_block_kind_check_fenced() {
        let parser = Parser::new("hello\n```test\ntadam\n```");
        let mut found = 0;
        for (ev, _range) in parser.into_offset_iter() {
            if let Event::Start(Tag::CodeBlock(CodeBlockKind::Fenced(syntax))) = ev {
                assert_eq!(syntax.as_ref(), "test");
                found += 1;
            }
        }
        assert_eq!(found, 1);
    }

    #[test]
    fn code_block_kind_check_indented() {
        let parser = Parser::new("hello\n\n    ```test\n    tadam\nhello");
        let mut found = 0;
        for (ev, _range) in parser.into_offset_iter() {
            if let Event::Start(Tag::CodeBlock(CodeBlockKind::Indented)) = ev {
                found += 1;
            }
        }
        assert_eq!(found, 1);
    }

    #[test]
    fn ref_defs() {
        let input = r###"[a B c]: http://example.com
[another]: https://google.com

text

[final ONE]: http://wikipedia.org
"###;
        let mut parser = Parser::new(input);

        assert!(parser.reference_definitions().get("a b c").is_some());
        assert!(parser.reference_definitions().get("nope").is_none());

        if let Some(_event) = parser.next() {
            // testing keys with shorter lifetimes than parser and its input
            let s = "final one".to_owned();
            let link_def = parser.reference_definitions().get(&s).unwrap();
            let span = &input[link_def.span.clone()];
            assert_eq!(span, "[final ONE]: http://wikipedia.org");
        }
    }

    #[test]
    #[allow(clippy::extra_unused_lifetimes)]
    fn common_lifetime_patterns_allowed<'b>() {
        let temporary_str = String::from("xyz");

        // NOTE: this is a limitation of Rust, it doesn't allow putting lifetime parameters on the closure itself.
        // Hack it by attaching the lifetime to the test function instead.
        // TODO: why is the `'b` lifetime required at all? Changing it to `'_` breaks things :(
        let mut closure = |link: BrokenLink<'b>| Some(("#".into(), link.reference));

        fn function(link: BrokenLink<'_>) -> Option<(CowStr<'_>, CowStr<'_>)> {
            Some(("#".into(), link.reference))
        }

        for _ in Parser::new_with_broken_link_callback(
            "static lifetime",
            Options::empty(),
            Some(&mut closure),
        ) {}
        /* This fails to compile. Because the closure can't say `for <'a> fn(BrokenLink<'a>) ->
         * CowStr<'a>` and has to use the enclosing `'b` lifetime parameter, `temporary_str` lives
         * shorter than `'b`. I think this is unlikely to occur in real life, and if it does, the
         * fix is simple: move it out to a function that allows annotating the lifetimes.
         */
        //for _ in Parser::new_with_broken_link_callback(&temporary_str, Options::empty(), Some(&mut callback)) {
        //}

        for _ in Parser::new_with_broken_link_callback(
            "static lifetime",
            Options::empty(),
            Some(&mut function),
        ) {}
        for _ in Parser::new_with_broken_link_callback(
            &temporary_str,
            Options::empty(),
            Some(&mut function),
        ) {}
    }

    #[test]
    fn inline_html_inside_blockquote() {
        // Regression for #960
        let input = "> <foo\n> bar>";
        let events: Vec<_> = Parser::new(input).collect();
        let expected = [
            Event::Start(Tag::BlockQuote(None)),
            Event::Start(Tag::Paragraph),
            Event::InlineHtml(CowStr::Boxed("<foo\nbar>".to_string().into())),
            Event::End(TagEnd::Paragraph),
            Event::End(TagEnd::BlockQuote(None)),
        ];
        assert_eq!(&events, &expected);
    }

    #[test]
    fn wikilink_has_pothole() {
        let input = "[[foo]] [[bar|baz]]";
        let events: Vec<_> = Parser::new_ext(input, Options::ENABLE_WIKILINKS).collect();
        let expected = [
            Event::Start(Tag::Paragraph),
            Event::Start(Tag::Link {
                link_type: LinkType::WikiLink { has_pothole: false },
                dest_url: CowStr::Borrowed("foo"),
                title: CowStr::Borrowed(""),
                id: CowStr::Borrowed(""),
            }),
            Event::Text(CowStr::Borrowed("foo")),
            Event::End(TagEnd::Link),
            Event::Text(CowStr::Borrowed(" ")),
            Event::Start(Tag::Link {
                link_type: LinkType::WikiLink { has_pothole: true },
                dest_url: CowStr::Borrowed("bar"),
                title: CowStr::Borrowed(""),
                id: CowStr::Borrowed(""),
            }),
            Event::Text(CowStr::Borrowed("baz")),
            Event::End(TagEnd::Link),
            Event::End(TagEnd::Paragraph),
        ];
        assert_eq!(&events, &expected);
    }

    #[cfg(feature = "mdx")]
    fn mdx_parser(text: &str) -> Parser<'_> {
        Parser::new_ext(text, Options::ENABLE_MDX)
    }

    #[cfg(feature = "mdx")]
    #[test]
    fn mdx_esm_import() {
        let events: Vec<_> = mdx_parser("import {Chart} from './chart.js'\n").collect();
        assert_eq!(events.len(), 1);
        assert!(matches!(&events[0], Event::MdxEsm(s) if s.contains("import")));
    }

    #[cfg(feature = "mdx")]
    #[test]
    fn mdx_esm_export() {
        let events: Vec<_> = mdx_parser("export const meta = {}\n").collect();
        assert_eq!(events.len(), 1);
        assert!(matches!(&events[0], Event::MdxEsm(s) if s.contains("export")));
    }

    #[cfg(feature = "mdx")]
    #[test]
    fn mdx_flow_expression() {
        let events: Vec<_> = mdx_parser("{1 + 1}\n").collect();
        assert_eq!(events.len(), 1);
        assert!(matches!(&events[0], Event::MdxFlowExpression(s) if s.as_ref() == "1 + 1"));
    }

    #[cfg(feature = "mdx")]
    #[test]
    fn mdx_jsx_flow_self_closing() {
        let events: Vec<_> = mdx_parser("<Chart values={[1,2,3]} />\n").collect();
        assert!(!events.is_empty());
        assert!(
            matches!(&events[0], Event::Start(Tag::MdxJsxFlowElement(s)) if s.contains("Chart"))
        );
    }

    #[cfg(feature = "mdx")]
    #[test]
    fn mdx_jsx_flow_fragment() {
        let events: Vec<_> = mdx_parser("<>\n").collect();
        assert!(!events.is_empty());
        assert!(matches!(
            &events[0],
            Event::Start(Tag::MdxJsxFlowElement(_))
        ));
    }

    #[cfg(feature = "mdx")]
    #[test]
    fn mdx_inline_expression() {
        let events: Vec<_> = mdx_parser("hello {name} world\n").collect();
        let has_expr = events
            .iter()
            .any(|e| matches!(e, Event::MdxTextExpression(s) if s.as_ref() == "name"));
        assert!(
            has_expr,
            "Expected inline MDX expression, got: {:?}",
            events
        );
    }

    #[cfg(feature = "mdx")]
    #[test]
    fn mdx_inline_jsx() {
        let events: Vec<_> = mdx_parser("hello <Badge /> world\n").collect();
        let has_jsx = events
            .iter()
            .any(|e| matches!(e, Event::Start(Tag::MdxJsxTextElement(s)) if s.contains("Badge")));
        assert!(has_jsx, "Expected inline MDX JSX, got: {:?}", events);
    }

    #[cfg(feature = "mdx")]
    #[test]
    fn mdx_all_tags_are_jsx() {
        // In MDX mode, all tags (including lowercase) are JSX, not HTML.
        let events: Vec<_> = mdx_parser("hello <em>world</em>\n").collect();
        let has_jsx = events
            .iter()
            .any(|e| matches!(e, Event::Start(Tag::MdxJsxTextElement(_))));
        assert!(has_jsx, "In MDX mode, <em> should be JSX: {:?}", events);
    }

    #[test]
    fn mdx_does_not_interfere_without_flag() {
        // Without ENABLE_MDX, none of this should be parsed as MDX.
        let events: Vec<_> = Parser::new("import foo from 'bar'\n").collect();
        // Should be a regular paragraph.
        assert!(
            events
                .iter()
                .any(|e| matches!(e, Event::Start(Tag::Paragraph)))
        );
    }

    #[cfg(feature = "mdx")]
    #[test]
    fn mdx_expression_in_heading() {
        let events: Vec<_> = mdx_parser("# {title}\n").collect();
        let has_heading = events
            .iter()
            .any(|e| matches!(e, Event::Start(Tag::Heading { .. })));
        assert!(has_heading, "Should have a heading");
        let has_expr = events
            .iter()
            .any(|e| matches!(e, Event::MdxTextExpression(s) if s.as_ref() == "title"));
        assert!(
            has_expr,
            "Heading should contain MdxTextExpression, got: {:?}",
            events
        );
    }

    #[cfg(feature = "mdx")]
    #[test]
    fn mdx_expression_mixed_text_in_heading() {
        let events: Vec<_> = mdx_parser("## Hello {name}\n").collect();
        let has_text = events
            .iter()
            .any(|e| matches!(e, Event::Text(s) if s.contains("Hello")));
        let has_expr = events
            .iter()
            .any(|e| matches!(e, Event::MdxTextExpression(s) if s.as_ref() == "name"));
        assert!(has_text, "Should have text, got: {:?}", events);
        assert!(has_expr, "Should have expression, got: {:?}", events);
    }
}