satteri-pulldown-cmark 0.6.0

//! The first pass resolves all block structure, generating an AST. Within a block, items
//! are in a linear chain with potential inline markup identified.

use alloc::{string::String, vec::Vec};
use core::{cmp::max, ops::Range};

use satteri_arena::line_ending_iter;
use unicase::UniCase;

#[cfg(feature = "mdx")]
use crate::mdx::*;
use crate::{
    HeadingLevel, LinkType, MetadataBlockKind, Options,
    linklabel::{LinkLabel, scan_link_label_rest},
    parse::{
        Allocations, AutolinkCandidate, DirectiveAttrData, FootnoteDef, HeadingAttributes, Item,
        ItemBody, LINK_MAX_NESTED_PARENS, LinkDef, scan_containers,
    },
    post_passes::{scan_autolink_literal, scan_email_autolink},
    scanners::*,
    strings::CowStr,
    tree::{Tree, TreeIndex},
};

pub(crate) fn run_first_pass(
    text: &str,
    options: Options,
) -> (Tree<Item>, Allocations<'_>, Vec<(usize, String)>) {
    // Measured: real-world Markdown yields ~1 tree item per 10 source bytes.
    let start_capacity = max(128, text.len() / 10);
    let lookup_table = &create_lut(&options);
    let first_pass = FirstPass {
        text,
        tree: Tree::with_capacity(start_capacity),
        begin_list_item: None,
        last_line_blank: false,
        list_interrupted_paragraph: false,
        refdef_interrupted_paragraph: false,
        allocs: Allocations::new(),
        options,
        lookup_table,
        brace_context_next: 0,
        brace_context_stack: Vec::new(),
        mdx_errors: Vec::new(),
        #[cfg(feature = "mdx")]
        mdx_expr_allocator: oxc_allocator::Allocator::default(),
        pending_lazy_blockquote_close: false,
    };
    first_pass.run()
}

// Each level of brace nesting adds another entry to a hash table.
// To limit the amount of memory consumed, do not create a new brace
// context beyond some amount deep.
//
// There are actually two limits at play here: this one,
// and the one where the maximum amount of distinct contexts passes
// the 255 item limit imposed by using `u8`. When over 255 distinct
// contexts are created, it wraps around, while this one instead makes it
// saturate, which is a better behavior.
const MATH_BRACE_CONTEXT_MAX_NESTING: usize = 25;

/// State for the first parsing pass.
pub(crate) struct FirstPass<'a, 'b> {
    pub(crate) text: &'a str,
    pub(crate) tree: Tree<Item>,
    begin_list_item: Option<usize>,
    last_line_blank: bool,
    list_interrupted_paragraph: bool,
    // Narrower variant set after a link reference definition: the def's
    // residual paragraph state interrupts only ordered-list markers
    // whose index ≠ 1. Unlike `list_interrupted_paragraph`, it must not
    // suppress blank-after-marker list items, which mdx-js/micromark
    // still accept after a refdef (`[r]: /x\n\n-\n- b` → list with
    // empty + filled items, not paragraph + list).
    refdef_interrupted_paragraph: bool,
    pub(crate) allocs: Allocations<'a>,
    pub(crate) options: Options,
    lookup_table: &'b LookupTable,
    /// Math environment brace nesting.
    brace_context_stack: Vec<u8>,
    brace_context_next: usize,
    /// MDX errors collected during first pass.
    pub(crate) mdx_errors: Vec<(usize, String)>,
    /// Mirrors micromark's `lazy[line]` for indented-code-after-blockquote-
    /// close. Set when `parse_block` pops a blockquote due to failed
    /// continuation; consumed (and cleared) by the next indented code block,
    /// which becomes one-line-only — matching `furtherStart`'s lazy-line
    /// rejection. Without this, `>\n    bar\n    baz` would merge `bar`
    /// and `baz` into one code block instead of splitting.
    pending_lazy_blockquote_close: bool,
    /// Reusable bump allocator for oxc parses (expression-body validation,
    /// ESM completeness checks). Avoids `Allocator::default()` heap alloc
    /// on every expression — the allocator is `reset()` between parses.
    #[cfg(feature = "mdx")]
    pub(crate) mdx_expr_allocator: oxc_allocator::Allocator,
}

impl<'a, 'b> FirstPass<'a, 'b> {
    fn run(mut self) -> (Tree<Item>, Allocations<'a>, Vec<(usize, String)>) {
        let mut ix = 0;
        while ix < self.text.len() {
            ix = self.parse_block(ix);
        }
        while self.tree.spine_len() > 0 {
            self.pop(ix);
        }
        (self.tree, self.allocs, self.mdx_errors)
    }

    /// Split a term paragraph's inline children into one detached `(head, tail)`
    /// chain per source line, breaking at (and dropping) the `SoftBreak`s that
    /// separate the lines. Each segment becomes its own `<dt>`, matching
    /// mdast-util-definition-list.
    fn collect_term_segments(&mut self, child: Option<TreeIndex>) -> Vec<(TreeIndex, TreeIndex)> {
        let mut segments = Vec::new();
        let mut seg: Option<(TreeIndex, TreeIndex)> = None;
        let mut cur = child;
        while let Some(node) = cur {
            let next = self.tree[node].next;
            if matches!(self.tree[node].item.body, ItemBody::SoftBreak) {
                if let Some(s) = seg.take() {
                    segments.push(s);
                }
            } else {
                self.tree[node].next = None;
                seg = Some(match seg {
                    None => (node, node),
                    Some((head, tail)) => {
                        self.tree[tail].next = Some(node);
                        (head, node)
                    }
                });
            }
            cur = next;
        }
        if let Some(s) = seg.take() {
            segments.push(s);
        }
        segments
    }

    /// Returns offset after block.
    fn parse_block(&mut self, mut start_ix: usize) -> usize {
        let bytes = self.text.as_bytes();
        let mut line_start = LineStart::new(&bytes[start_ix..]);

        // math spans and their braces are tracked only within a single block
        self.brace_context_stack.clear();
        self.brace_context_next = 0;

        let i = scan_containers(&self.tree, &mut line_start, self.options);
        if i < self.tree.spine_len() {
            self.list_interrupted_paragraph = false;
            self.refdef_interrupted_paragraph = false;
            // If any popped container was a blockquote AND the current line
            // is non-blank, the next indented code block on this line is
            // "lazy" wrt the blockquote — micromark would limit it to one
            // line. Skip the marker on blank lines: a blank line that pops
            // the blockquote is a *proper* close, after which subsequent
            // indented code blocks merge normally (e.g. `> Foo\n\n    bar
            // \n    baz` keeps `bar\nbaz` as one code block).
            let probe_ix = start_ix + line_start.bytes_scanned();
            let line_is_blank = scan_blank_line(&bytes[probe_ix..]).is_some();
            if !line_is_blank
                && !self.pending_lazy_blockquote_close
                && self.tree.walk_spine().skip(i).any(|&ix| {
                    matches!(
                        self.tree[ix].item.body,
                        ItemBody::BlockQuote(..) | ItemBody::ListItem(..)
                    )
                })
            {
                self.pending_lazy_blockquote_close = true;
            }
        }
        for _ in i..self.tree.spine_len() {
            self.pop(start_ix);
        }

        // Before processing new containers: if we arrive here with a
        // non-blank line pending, the previous line was a non-trailing blank
        // that belongs to the currently-open list item or definition. Mark it
        // as spread (loose). Must happen before the new-containers loop, which
        // could open a deeper nested list item and shift the spine tip.
        if self.last_line_blank {
            let content_probe = start_ix + line_start.bytes_scanned();
            let has_content =
                content_probe < bytes.len() && scan_blank_line(&bytes[content_probe..]).is_none();
            if has_content
                && let Some(up) = self.tree.peek_up()
                && self.tree[up].child.is_some()
            {
                match self.tree[up].item.body {
                    ItemBody::ListItem(_, _) => self.mark_enclosing_listitem_spread(),
                    ItemBody::DefinitionListDefinition(indent, _) => {
                        self.tree[up].item.body = ItemBody::DefinitionListDefinition(indent, true);
                    }
                    _ => {}
                }
            }
        }

        // Process new containers
        loop {
            let save = line_start.clone();
            let mut outer_indent = line_start.scan_space_upto(4);
            if outer_indent >= 4 {
                if self.options.contains(Options::ENABLE_MDX) {
                    // MDX has no indented code blocks.  Speculatively scan all
                    // remaining whitespace to look for a new container (list
                    // marker, blockquote, or — when enabled — a directive fence)
                    // at deeper indentation. If no new container is found we
                    // restore and break so that the indentation is preserved for
                    // content that belongs to an existing list item.
                    let extra = line_start.scan_space_upto(usize::MAX);
                    let mdx_ix = start_ix + line_start.bytes_scanned();
                    let has_directive = self.options.contains(Options::ENABLE_DIRECTIVE)
                        && scan_ch_repeat(&bytes[mdx_ix..], b':') > 1;
                    let has_container = scan_listitem(&bytes[mdx_ix..]).is_some()
                        || (mdx_ix < bytes.len() && bytes[mdx_ix] == b'>')
                        || has_directive;
                    if !has_container {
                        // Tables are detected later in parse_paragraph when
                        // the table opener (`|`) is the first non-whitespace
                        // char on the line. Don't break here if this line
                        // could be a table start — restore only the outer
                        // indent and let paragraph-level parsing continue
                        // so scan_paragraph_table sees it.
                        if self.options.contains(Options::ENABLE_TABLES)
                            && mdx_ix < bytes.len()
                            && bytes[mdx_ix] == b'|'
                        {
                            // keep the whitespace we consumed; break so we
                            // advance into paragraph parsing with the pipe
                            // as the first non-whitespace byte.
                            break;
                        }
                        line_start = save;
                        break;
                    }
                    // The deeper list marker's continuation indent must account
                    // for the whitespace we just consumed past the initial 4.
                    outer_indent += extra;
                } else {
                    line_start = save;
                    break;
                }
            }
            if self.options.contains(Options::ENABLE_FOOTNOTES) {
                // Footnote definitions
                let container_start = start_ix + line_start.bytes_scanned();
                if let Some(bytecount) = self.parse_footnote(container_start) {
                    start_ix = container_start + bytecount;
                    line_start = LineStart::new(&bytes[start_ix..]);
                    continue;
                }
            }
            let container_start = start_ix + line_start.bytes_scanned();
            if let Some((ch, index, indent)) = line_start.scan_list_marker_with_indent_and_clamp(
                outer_indent,
                !self.options.contains(Options::ENABLE_MDX),
            ) {
                let after_marker_index = start_ix + line_start.bytes_scanned();
                let already_in_list = self
                    .tree
                    .peek_up()
                    .is_some_and(|ix| matches!(self.tree[ix].item.body, ItemBody::List(_, _, _)));
                let after_marker_blank = {
                    let rest = &bytes[after_marker_index..];
                    rest.is_empty() || scan_blank_line(rest).is_some()
                };
                // An empty list marker can't interrupt a paragraph — including
                // the residual paragraph state left by a refdef. After
                // `[a]: u\n>*`, the `*` inside the new blockquote should be
                // paragraph text, not an empty list item, because micromark's
                // paragraph token is still considered "interrupted".
                if (self.list_interrupted_paragraph || self.refdef_interrupted_paragraph)
                    && !already_in_list
                    && after_marker_blank
                {
                    self.list_interrupted_paragraph = false;
                    self.refdef_interrupted_paragraph = false;
                    line_start = save;
                    break;
                }
                // micromark's list construct rejects start != 1 when
                // `self.interrupt` is set (mirrors `currentConstruct` lingering
                // after the previous block). After an indented code block,
                // `2. b` becomes a paragraph rather than starting a fresh
                // ordered list (`    code\n\n2. b` → `[code, paragraph]`).
                // Index-1 starts are always allowed.
                //
                // The check is TEXTUAL: only an exact `1.` or `1)` (single
                // digit, no leading zero) opens a list across a paragraph
                // boundary. `01.`, `001.`, `10.` all stay paragraph text —
                // micromark scans the literal marker bytes, not the numeric
                // value.
                //
                // This applies even when *nested* inside an existing list —
                // after `[ref]: /uri\n1. - 2. foo` the `2.` inside the inner
                // unordered list-item also fails the interrupt check, so it
                // stays as paragraph text instead of opening a new ordered
                // list. (Index-1 markers and any markers with a clear flag
                // are unaffected.)
                let is_textual_one = (ch == b'.' || ch == b')')
                    && bytes.get(container_start) == Some(&b'1')
                    && bytes
                        .get(container_start + 1)
                        .is_some_and(|&b| b == b'.' || b == b')');
                if (self.list_interrupted_paragraph || self.refdef_interrupted_paragraph)
                    && (ch == b'.' || ch == b')')
                    && !is_textual_one
                {
                    self.list_interrupted_paragraph = false;
                    self.refdef_interrupted_paragraph = false;
                    line_start = save;
                    break;
                }
                self.continue_list(container_start, ch, index);
                self.tree.append(Item {
                    start: container_start,
                    end: after_marker_index, // will get updated later if item not empty
                    body: ItemBody::ListItem(indent, false),
                });
                self.tree.push();
                if let Some(n) = scan_blank_line(&bytes[after_marker_index..]) {
                    self.begin_list_item = Some(after_marker_index + n);
                    return after_marker_index + n;
                }
                if self.options.contains(Options::ENABLE_TASKLISTS) {
                    let saved_line_start = line_start.clone();
                    let task_list_marker =
                        line_start.scan_task_list_marker().map(|is_checked| Item {
                            start: after_marker_index,
                            end: start_ix + line_start.bytes_scanned(),
                            body: ItemBody::TaskListMarker(is_checked),
                        });
                    if let Some(task_list_marker) = task_list_marker {
                        let rest = &bytes[task_list_marker.end..];
                        let marker_ate_newline = matches!(
                            bytes.get(task_list_marker.end.wrapping_sub(1)),
                            Some(b'\n' | b'\r')
                        );
                        // Skip spaces/tabs on the rest of the marker line to
                        // find the first real content (if any).
                        let trailing_ws = rest
                            .iter()
                            .position(|&b| b != b' ' && b != b'\t')
                            .unwrap_or(rest.len());
                        let after_ws = &rest[trailing_ws..];
                        let rest_of_line_blank =
                            after_ws.is_empty() || matches!(after_ws.first(), Some(b'\n' | b'\r'));
                        // When the rest of the marker line is blank, see if the
                        // next line has lazy-continuation content we should
                        // attach to the task item's paragraph.
                        //
                        // `marker_ate_newline=true` means the scanner already
                        // consumed the newline as the trailing whitespace, so
                        // the marker line ended right after `]` and `after_ws`
                        // is on the next line — handle it the same as a blank
                        // marker tail.
                        let lazy_continuation_start = if marker_ate_newline {
                            let start = task_list_marker.end + trailing_ws;
                            (start < bytes.len()
                                && scan_blank_line(&bytes[start..]).is_none()
                                && !scan_paragraph_interrupt_no_table(
                                    &bytes[start..],
                                    true,
                                    self.options.contains(Options::ENABLE_FOOTNOTES),
                                    self.options.contains(Options::ENABLE_DEFINITION_LIST),
                                    self.options.contains(Options::ENABLE_MDX),
                                    self.options.contains(Options::ENABLE_MATH_MULTI_DOLLAR),
                                    self.options.contains(Options::ENABLE_DIRECTIVE),
                                    &self.tree,
                                    self.tree.spine_len(),
                                ))
                            .then_some(start)
                        } else if rest_of_line_blank {
                            let newline_len = scan_eol(after_ws).unwrap_or(0);
                            let start = task_list_marker.end + trailing_ws + newline_len;
                            (newline_len > 0
                                && start < bytes.len()
                                && scan_blank_line(&bytes[start..]).is_none()
                                && !scan_paragraph_interrupt_no_table(
                                    &bytes[start..],
                                    true,
                                    self.options.contains(Options::ENABLE_FOOTNOTES),
                                    self.options.contains(Options::ENABLE_DEFINITION_LIST),
                                    self.options.contains(Options::ENABLE_MDX),
                                    self.options.contains(Options::ENABLE_MATH_MULTI_DOLLAR),
                                    self.options.contains(Options::ENABLE_DIRECTIVE),
                                    &self.tree,
                                    self.tree.spine_len(),
                                ))
                            .then_some(start)
                        } else {
                            None
                        };
                        if let Some(new_start) = lazy_continuation_start {
                            return self.parse_paragraph(new_start, Some(task_list_marker));
                        } else if rest_of_line_blank || marker_ate_newline {
                            // No paragraph content found for the task item:
                            // either the next line is a paragraph interrupt
                            // (so it can't lazily continue the task item) or
                            // the next line is blank.
                            line_start = saved_line_start;
                        } else {
                            return self
                                .parse_paragraph(task_list_marker.end, Some(task_list_marker));
                        }
                    }
                }
            } else if let Some((indent, child, item)) = self
                .options
                .contains(Options::ENABLE_DEFINITION_LIST)
                .then(|| {
                    self.tree
                        .cur()
                        .map(|cur| (self.tree[cur].child, &mut self.tree[cur].item))
                })
                .flatten()
                .filter(|(_, item)| {
                    match item.body {
                        // A term may sit at most one blank line from its
                        // definition (pandoc / mdast-util-definition-list); two+
                        // disconnect it. `item.end` includes the paragraph's
                        // trailing line ending, so the gap to the marker holds 0
                        // line endings when tight, 1 when loose, 2+ when disconnected.
                        ItemBody::Paragraph | ItemBody::TightParagraph => {
                            let gap_start = item.end.min(container_start);
                            line_ending_iter(&bytes[gap_start..container_start])
                                .take(2)
                                .count()
                                <= 1
                        }
                        ItemBody::MaybeDefinitionListTitle
                        | ItemBody::DefinitionListDefinition(..) => true,
                        _ => false,
                    }
                })
                .and_then(|item| {
                    Some((
                        line_start
                            .scan_definition_list_definition_marker_with_indent(outer_indent)?,
                        item.0,
                        item.1,
                    ))
                })
            {
                match item.body {
                    ItemBody::Paragraph | ItemBody::TightParagraph => {
                        // Repurpose the term paragraph as the `<dl>`, one `<dt>`
                        // per source line.
                        item.body = ItemBody::DefinitionList(true);
                        let Item { start, end, .. } = *item;
                        let segments = self.collect_term_segments(child);
                        // Drop the old inline chain (now re-parented under the
                        // titles) and append the titles as the list's children.
                        let list_idx = self.tree.cur().unwrap();
                        self.tree[list_idx].child = None;
                        self.tree.push();
                        if segments.is_empty() {
                            self.tree.append(Item {
                                start,
                                end,
                                body: ItemBody::DefinitionListTitle,
                            });
                        } else {
                            for (head, tail) in segments {
                                let t_start = self.tree[head].item.start;
                                let t_end = self.tree[tail].item.end;
                                let title = self.tree.append(Item {
                                    start: t_start,
                                    end: t_end,
                                    body: ItemBody::DefinitionListTitle,
                                });
                                self.tree[title].child = Some(head);
                            }
                        }
                    }
                    ItemBody::MaybeDefinitionListTitle => {
                        // A new term after a definition: reuse this paragraph as
                        // the first `<dt>`; any further source lines become
                        // sibling `<dt>`s, mirroring the initial-term split.
                        item.body = ItemBody::DefinitionListTitle;
                        let node_idx = self.tree.cur().unwrap();
                        let segments = self.collect_term_segments(child);
                        if let Some((head, tail)) = segments.first().copied() {
                            let t_start = self.tree[head].item.start;
                            let t_end = self.tree[tail].item.end;
                            self.tree[node_idx].child = Some(head);
                            self.tree[node_idx].item.start = t_start;
                            self.tree[node_idx].item.end = t_end;
                        }
                        for (head, tail) in segments.into_iter().skip(1) {
                            let t_start = self.tree[head].item.start;
                            let t_end = self.tree[tail].item.end;
                            let title = self.tree.append(Item {
                                start: t_start,
                                end: t_end,
                                body: ItemBody::DefinitionListTitle,
                            });
                            self.tree[title].child = Some(head);
                        }
                    }
                    ItemBody::DefinitionListDefinition(..) => {}
                    _ => unreachable!(),
                }
                let after_marker_index = start_ix + line_start.bytes_scanned();
                // Record looseness per-definition (blank line before the marker)
                // so mixed tight/loose lists render correctly. The list-wide
                // `is_tight` below still drives `surgerize_tight_list`, and stays
                // consistent because a fully tight list has no loose definition.
                let loose = self.last_line_blank;
                self.tree.append(Item {
                    start: container_start - outer_indent,
                    end: after_marker_index, // will get updated later if item not empty
                    body: ItemBody::DefinitionListDefinition(indent, loose),
                });
                if let Some(ItemBody::DefinitionList(is_tight)) =
                    self.tree.peek_up().map(|cur| &mut self.tree[cur].item.body)
                    && self.last_line_blank
                {
                    *is_tight = false;
                    self.last_line_blank = false;
                }
                self.tree.push();
                if let Some(n) = scan_blank_line(&bytes[after_marker_index..]) {
                    self.begin_list_item = Some(after_marker_index + n);
                    return after_marker_index + n;
                }
            } else if line_start.scan_blockquote_marker() {
                let kind = if self.options.contains(Options::ENABLE_GITHUB_ALERTS) {
                    line_start.scan_blockquote_tag()
                } else {
                    None
                };
                self.finish_list(start_ix);
                self.tree.append(Item {
                    start: container_start,
                    end: 0, // will get set later
                    body: ItemBody::BlockQuote(kind),
                });
                self.tree.push();
                if kind.is_some() {
                    // blockquote tag leaves us at the end of the line
                    // we need to scan through all the container syntax for the next line
                    // and break out if we can't re-scan all of them
                    let ix = start_ix + line_start.bytes_scanned();
                    let mut lazy_line_start = LineStart::new(&bytes[ix..]);
                    let tree_position =
                        scan_containers(&self.tree, &mut lazy_line_start, self.options);
                    let current_container = tree_position == self.tree.spine_len();
                    if !lazy_line_start.scan_space(4)
                        && self.scan_paragraph_interrupt(
                            &bytes[ix + lazy_line_start.bytes_scanned()..],
                            current_container,
                            tree_position,
                        )
                    {
                        return ix;
                    } else {
                        // blockquote tags act as if they were nested in a paragraph
                        // so you can lazily continue the imaginary paragraph off of them
                        line_start = lazy_line_start;
                        line_start.scan_all_space();
                        start_ix = ix;
                        break;
                    }
                }
            } else if self.options.contains(Options::ENABLE_DIRECTIVE)
                && scan_ch_repeat(&bytes[(start_ix + line_start.bytes_scanned())..], b':') > 1
            {
                let colon_start = start_ix + line_start.bytes_scanned();
                let colon_count = scan_ch_repeat(&bytes[colon_start..], b':');
                if colon_count >= 3 && self.tree.spine_len() <= u8::MAX as usize {
                    // Container directive (:::+)
                    let fence_length = core::cmp::min(colon_count, u8::MAX as usize);
                    let after_colons = colon_start + colon_count;
                    if let Some((mut dir_data, content_end)) =
                        parse_directive_after_colons(self.text, bytes, after_colons)
                    {
                        // initialSize per micromark-extension-directive: cols of
                        // leading whitespace before `:::` *after* outer-container
                        // prefix stripping. Drives the dedent for multi-line MDX
                        // expression attribute values nested in the body.
                        dir_data.initial_size = outer_indent.min(u8::MAX as usize) as u8;
                        // Close any open list before opening a sibling directive,
                        // matching how blockquote handles the same transition.
                        self.finish_list(start_ix);
                        // For block directives, advance to end of line
                        let after = &bytes[content_end..];
                        let ws = scan_space_or_tab(after);
                        let line_end = content_end + ws + scan_nextline(&after[ws..]);
                        // `[label]` offsets (0/0 when no brackets) — capture
                        // before `dir_data` is moved into the allocator.
                        let label_start = dir_data.label_start;
                        let label_end = dir_data.label_end;
                        let dir_ix = self.allocs.allocate_directive(dir_data);
                        self.tree.append(Item {
                            start: container_start,
                            end: 0,
                            body: ItemBody::ContainerDirective(fence_length as u8, dir_ix),
                        });
                        self.tree.push();
                        // Emit the label as a real inline-tokenized child so the
                        // normal inline pass resolves emphasis/strong/links/code.
                        if label_start != 0 || label_end != 0 {
                            self.append_container_directive_label(label_start, label_end);
                        }
                        return line_end;
                    } else {
                        break;
                    }
                } else if colon_count == 2 {
                    // Leaf directive (::)
                    let after_colons = colon_start + 2;
                    if let Some((dir_data, line_end)) =
                        parse_directive_after_colons(self.text, bytes, after_colons)
                    {
                        // Verify only whitespace follows on the line
                        let remaining = &bytes[line_end..];
                        let ws = scan_space_or_tab(remaining);
                        let at_eol = line_end + ws >= bytes.len()
                            || bytes[line_end + ws] == b'\n'
                            || bytes[line_end + ws] == b'\r';
                        if at_eol || line_end >= bytes.len() {
                            self.finish_list(start_ix);
                            let label_start = dir_data.label_start;
                            let label_end = dir_data.label_end;
                            let dir_ix = self.allocs.allocate_directive(dir_data);
                            self.tree.append(Item {
                                start: container_start,
                                end: line_end,
                                body: ItemBody::LeafDirective(dir_ix),
                            });
                            // The label is the directive's inline content: tokenize
                            // it directly as children (no label paragraph).
                            if label_start < label_end {
                                self.tree.push();
                                self.parse_line(
                                    label_start,
                                    Some(label_end),
                                    TableParseMode::Disabled,
                                    label_start,
                                );
                                self.tree.pop();
                            }
                            let next_line = line_end + scan_nextline(&bytes[line_end..]);
                            return next_line;
                        }
                    }
                    break;
                } else {
                    break;
                }
            } else {
                line_start = save;
                break;
            }
        }

        if self.options.contains(Options::ENABLE_DIRECTIVE) {
            let mut pop_count = None;
            let mut fence_line_end = start_ix;
            // Closing fence may be indented up to 3 spaces relative to the
            // current container cursor — matches remark-directive. Try
            // matching against each enclosing ContainerDirective; if the
            // speculative space consumption doesn't lead to a match, restore
            // line_start before continuing with normal block parsing.
            let fence_save = line_start.clone();
            let _ = line_start.scan_space_upto(3);
            let mut matched_length: Option<u8> = None;
            for (i, &node_ix) in self.tree.walk_spine().rev().enumerate() {
                match self.tree[node_ix].item.body {
                    ItemBody::ContainerDirective(length, ..) => {
                        let probe = line_start.clone();
                        if line_start.scan_closing_container_extensions_fence(length) {
                            let after_fence = start_ix + line_start.bytes_scanned();
                            fence_line_end = after_fence + scan_nextline(&bytes[after_fence..]);
                            pop_count = Some(i + 1);
                            matched_length = Some(length);
                            break;
                        }
                        line_start = probe;
                    }
                    ItemBody::List(..) | ItemBody::ListItem(..) => {}
                    _ => break,
                }
            }
            if pop_count.is_none() {
                line_start = fence_save;
            }

            if let Some(mut c) = pop_count {
                // remark-directive closes a contiguous run of same-length
                // container directives on a single closing `:::` — the fence
                // line is shared between the innermost open directive and
                // all of its ancestors that use the same fence length. Walk
                // further up the spine and bump the pop count accordingly.
                if let Some(length) = matched_length {
                    let spine: Vec<TreeIndex> = self.tree.walk_spine().copied().collect();
                    if c < spine.len() {
                        for &ancestor_ix in spine.iter().rev().skip(c) {
                            match self.tree[ancestor_ix].item.body {
                                ItemBody::ContainerDirective(other_length, ..)
                                    if other_length == length =>
                                {
                                    c += 1;
                                }
                                ItemBody::List(..) | ItemBody::ListItem(..) => {}
                                _ => break,
                            }
                        }
                    }
                }
                for _ in 0..c {
                    self.pop(fence_line_end);
                }
                return fence_line_end;
            }
        }
        let ix = start_ix + line_start.bytes_scanned();

        if let Some(n) = scan_blank_line(&bytes[ix..]) {
            // A blank line breaks the "lazy-immediately-after-close"
            // restriction: micromark's `furtherStart` only rejects lazy
            // lines that directly follow the implicit close, not ones with
            // a blank line in between. Clearing here lets a subsequent
            // indented code block parse as a normal multi-line block.
            self.pending_lazy_blockquote_close = false;
            if let Some(node_ix) = self.tree.peek_up() {
                match &mut self.tree[node_ix].item.body {
                    ItemBody::ContainerDirective(..) => {
                        // Blank lines inside a container directive propagate
                        // looseness to the enclosing list item — matching
                        // remark-directive's behavior where a directive with
                        // blank-line-separated content makes its ancestor list
                        // item spread.
                        self.mark_enclosing_listitem_spread();
                    }
                    ItemBody::BlockQuote(..) => {
                        // Blank `>` line inside a blockquote separates whatever
                        // closed the outer paragraph from the next block within
                        // the blockquote. Clear `list_interrupted_paragraph`
                        // so an empty `-` marker on the following blockquote
                        // line opens a fresh list (matches mdx-js/micromark:
                        // `_\n>\n>-` → paragraph + blockquote(list(empty))).
                        self.list_interrupted_paragraph = false;
                    }
                    ItemBody::ListItem(indent, _)
                    | ItemBody::DefinitionListDefinition(indent, _)
                        if self.begin_list_item.is_some() =>
                    {
                        self.last_line_blank = true;
                        if !self.options.contains(Options::ENABLE_MDX) {
                            // This is a blank list item.
                            // While the list itself can be continued no matter how many blank lines
                            // there are between this one and the next one, it cannot nest anything
                            // else, so its indentation should not be subtracted from the line start.
                            // In MDX mode we keep the indent because there are no indented code
                            // blocks, so continuation content at any depth should remain part of
                            // the list item.
                            *indent = 0;
                        }
                    }
                    _ => {
                        self.last_line_blank = true;
                    }
                }
            } else {
                self.last_line_blank = true;
            }
            return ix + n;
        }

        // Save `remaining_space` here to avoid needing to backtrack `line_start` for HTML blocks
        let remaining_space = line_start.remaining_space();
        let content_start_ix = start_ix + line_start.bytes_scanned();

        let mut indent = line_start.scan_space_upto(4);
        if indent == 4 {
            if self.options.contains(Options::ENABLE_MDX) {
                // MDX does not support indented code blocks. Track the full
                // leading whitespace as the indent so that deeply-indented
                // fenced code blocks inside containers (lists, JSX, directives)
                // strip the right amount from their content lines.
                indent += line_start.scan_space_upto(usize::MAX);
            } else {
                // If finish_list will pop an EMPTY list item (marker
                // followed by blank, no content), the subsequent indented
                // code block becomes "lazy" — mirrors micromark's
                // `furtherStart` restriction. Without this,
                // `*\n\n      bar\n      baz` would merge into one code
                // block; remark separates them. The `begin_list_item`
                // flag distinguishes the empty case (set on blank-after-
                // marker open) from list items whose first content is
                // still being processed (`-             bbb` has 12+
                // spaces of post-marker indent that's an indented code
                // block, but the list item is NOT empty in source).
                // Only fire when a real blank line separates the marker
                // from the indented content (`*\n\n    foo` → split). For
                // the directly-following case (`-\n             bbb` →
                // indented code stays IN the list item), don't fire.
                // begin_list_item records the byte right after the
                // marker's line ending; if start_ix > that, a blank line
                // sat between.
                let empty_listitem_will_close =
                    self.begin_list_item.is_some_and(|bli| start_ix > bli)
                        && self.tree.peek_up().is_some_and(|ix| {
                            matches!(self.tree[ix].item.body, ItemBody::ListItem(..))
                                && self.tree[ix].child.is_none()
                        });
                self.finish_list(start_ix);
                if empty_listitem_will_close {
                    self.pending_lazy_blockquote_close = true;
                }
                let ix = start_ix + line_start.bytes_scanned();
                let remaining_space = line_start.remaining_space();
                return self.parse_indented_code_block(content_start_ix, ix, remaining_space);
            }
        }

        let ix = start_ix + line_start.bytes_scanned();

        // Metadata blocks cannot be indented and, matching remark-frontmatter,
        // only match at the very start of the document.
        if indent == 0
            && ix == 0
            && self.tree.spine_len() == 0
            && let Some((_n, metadata_block_ch)) = scan_metadata_block(
                &bytes[ix..],
                self.options
                    .contains(Options::ENABLE_YAML_STYLE_METADATA_BLOCKS),
                self.options
                    .contains(Options::ENABLE_PLUSES_DELIMITED_METADATA_BLOCKS),
            )
        {
            self.finish_list(start_ix);
            return self.parse_metadata_block(ix, metadata_block_ch);
        }

        // MDX blocks, must be checked before HTML blocks since JSX looks like HTML.
        #[cfg(feature = "mdx")]
        if self.options.contains(Options::ENABLE_MDX) {
            // MDX ESM: lines starting with `import` or `export`.
            // ESM is only valid at the document root — but a still-open
            // List (parent of a closed-empty ListItem) doesn't count as a
            // container that would forbid ESM. The list will close before
            // the ESM block.
            let at_root_for_esm = indent == 0
                && self
                    .tree
                    .walk_spine()
                    .all(|&ix| matches!(self.tree[ix].item.body, ItemBody::List(..)));
            if at_root_for_esm && let Some(end_ix) = scan_mdx_esm(&bytes[ix..]) {
                let mut final_end = end_ix;

                // If the scanned ESM block is incomplete (e.g. an export
                // spanning a blank line), retry across blank lines using
                // oxc, matching the reference mdxjs behavior.
                let candidate = self.text[ix..ix + final_end].trim_end();
                if !candidate.is_empty() {
                    use crate::mdx::EsmParseResult;
                    let mut allocator = oxc_allocator::Allocator::default();
                    match crate::mdx::try_parse_esm(candidate, &mut allocator) {
                        EsmParseResult::Complete => {}
                        EsmParseResult::Incomplete => {
                            let mut pos = ix + final_end;
                            loop {
                                let blank_start = pos;
                                while pos < bytes.len()
                                    && (bytes[pos] == b'\n'
                                        || bytes[pos] == b'\r'
                                        || bytes[pos] == b' '
                                        || bytes[pos] == b'\t')
                                {
                                    pos += 1;
                                }
                                if pos == blank_start || pos >= bytes.len() {
                                    break;
                                }
                                let chunk_start = pos;
                                while pos < bytes.len() {
                                    pos += scan_nextline(&bytes[pos..]);
                                    if pos < bytes.len()
                                        && (bytes[pos] == b'\n' || bytes[pos] == b'\r')
                                    {
                                        break;
                                    }
                                }
                                if pos == chunk_start {
                                    break;
                                }
                                final_end = pos - ix;
                                let candidate = self.text[ix..ix + final_end].trim_end();
                                match crate::mdx::try_parse_esm(candidate, &mut allocator) {
                                    EsmParseResult::Complete => break,
                                    EsmParseResult::Incomplete => continue,
                                    EsmParseResult::Error => break,
                                }
                            }
                        }
                        EsmParseResult::Error => {}
                    }
                }

                self.finish_list(start_ix);
                return self.parse_mdx_esm(ix, ix + final_end);
            }

            // MDX JSX flow: line starting with `<` followed by component name or fragment
            if bytes[ix] == b'<'
                && let Some(end_ix) =
                    self.scan_mdx_flow_in_container(ix, |b, c| scan_mdx_jsx_block(b, c))
            {
                self.finish_list(start_ix);
                let result = self.parse_mdx_jsx_flow(ix, ix + end_ix);
                // A blank line inside the consumed flow block means the
                // enclosing list item is "loose" (spread=true). remark
                // detects this naturally since its tokenizer reads line-
                // by-line; we consume the whole block atomically, so we
                // have to inspect the span here.
                if contains_blank_line(&bytes[ix..ix + end_ix]) {
                    self.last_line_blank = true;
                    self.mark_enclosing_listitem_spread();
                }
                return result;
            }

            // MDX expression flow: line starting with `{`
            if bytes[ix] == b'{' {
                if let Some(end_ix) =
                    self.scan_mdx_flow_in_container(ix, |b, c| scan_mdx_expression_block(b, c))
                {
                    self.finish_list(start_ix);
                    let result = self.parse_mdx_jsx_flow(ix, ix + end_ix);
                    if contains_blank_line(&bytes[ix..ix + end_ix]) {
                        self.last_line_blank = true;
                        self.mark_enclosing_listitem_spread();
                    }
                    return result;
                }
                // If the inline scanner also can't find a closing `}`, it's truly unclosed.
                // (If it CAN find one, the `{` will be handled as inline in a paragraph.)
                if scan_mdx_inline_expression(&bytes[ix..]).is_none() {
                    self.mdx_errors.push((
                        ix,
                        "Unexpected end of file in expression, expected a corresponding \
                         closing brace for `{`"
                            .to_string(),
                    ));
                }
            }
        }

        // HTML Blocks, completely disabled in MDX mode (all tags are JSX).
        if bytes[ix] == b'<' && !self.options.contains(Options::ENABLE_MDX) {
            // Only the part of `indent` consumed from leftover tab-expansion
            // spaces needs to be synthesized — those don't have raw bytes in
            // [content_start_ix..]. Indent consumed from actual space bytes is
            // already present in the source range. `scan_space_inner` drains
            // `spaces_remaining` first, so the tab-leftover portion is
            // `min(saved_remaining_space, indent)`.
            let synth = remaining_space.min(indent);

            // Types 1-5 are all detected by one function and all end with the same
            // pattern
            if let Some(html_end_tag) = get_html_end_tag(&bytes[(ix + 1)..]) {
                self.finish_list(start_ix);
                return self.parse_html_block_type_1_to_5(content_start_ix, html_end_tag, synth, 0);
            }

            // Detect type 6
            if starts_html_block_type_6(&bytes[(ix + 1)..]) {
                self.finish_list(start_ix);
                return self.parse_html_block_type_6_or_7(content_start_ix, synth, 0);
            }

            // Detect type 7
            if let Some(_html_bytes) = scan_html_type_7(&bytes[ix..]) {
                self.finish_list(start_ix);
                return self.parse_html_block_type_6_or_7(content_start_ix, synth, 0);
            }
        }

        if let Ok(n) = scan_hrule(&bytes[ix..]) {
            self.finish_list(start_ix);
            return self.parse_hrule(n, ix);
        }

        if let Some(atx_size) = scan_atx_heading(&bytes[ix..]) {
            self.finish_list(start_ix);
            return self.parse_atx_heading(ix, atx_size);
        }

        if let Some((n, fence_ch)) = scan_code_fence(&bytes[ix..]) {
            self.finish_list(start_ix);
            return self.parse_fenced_code_block(ix, indent, fence_ch, n);
        }

        if self.options.contains(Options::ENABLE_MATH_MULTI_DOLLAR)
            && let Some(n) = scan_math_fence(&bytes[ix..])
        {
            self.finish_list(start_ix);
            return self.parse_math_block(ix, indent, n);
        }

        // parse refdef
        while let Some((bytecount, label, link_def)) =
            self.parse_refdef_total(start_ix + line_start.bytes_scanned())
        {
            self.allocs
                .refdefs_all
                .push((label.clone(), link_def.clone()));
            self.allocs.refdefs.0.entry(label).or_insert(link_def);
            let container_start = start_ix + line_start.bytes_scanned();
            let mut ix = container_start + bytecount;
            // A blank line between the refdef and the next interrupting block
            // means micromark has already closed the refdef's residual paragraph
            // before that next block opens — so the new block isn't "interrupting"
            // an open paragraph and `refdef_interrupted_paragraph` must not fire.
            let refdef_terminated_by_blank;
            // Refdefs act as if they were contained within a paragraph, for purposes of lazy
            // continuations. For example:
            //
            // ```
            // > [foo]: http://example.com
            // bar
            // ```
            //
            // is equivalent to
            //
            // ```
            // > bar
            //
            // [foo]: http://example.com
            // ```
            if let Some(nl) = scan_blank_line(&bytes[ix..]) {
                // `scan_blank_line` happily consumes the refdef's own terminator
                // newline. A SECOND blank line at the new position is what
                // distinguishes "refdef\nnext-block" (still in paragraph)
                // from "refdef\n\nnext-block" (paragraph closed).
                let after_terminator = ix + nl;
                refdef_terminated_by_blank = scan_blank_line(&bytes[after_terminator..]).is_some();
                ix = after_terminator;
            } else {
                self.finish_list(start_ix);
                // Refdefs share micromark's paragraph token: the def's
                // residual paragraph state lingers after the def's bytes
                // are consumed. When the next line opens a new block,
                // ordered-list markers with start != 1 are rejected.
                // Use the narrower `refdef_interrupted_paragraph` flag so
                // blank-after-marker bullets like `-\n` are still
                // recognized as list items.
                self.refdef_interrupted_paragraph = true;
                return ix;
            }
            if let Some(lazy_line_start) = self.scan_next_line_or_lazy_continuation(&bytes[ix..]) {
                line_start = lazy_line_start;
                start_ix = ix;
            } else {
                self.finish_list(start_ix);
                // Only carry the interrupt state forward when the refdef wasn't
                // already closed by a blank line. `[a]: u\n\n>*` should open a
                // real list inside the blockquote; `[a]: u\n>*` should not.
                if !refdef_terminated_by_blank {
                    self.refdef_interrupted_paragraph = true;
                }
                return ix;
            }
        }

        let ix = start_ix + line_start.bytes_scanned();

        self.parse_paragraph(ix, None)
    }

    /// footnote definitions and GFM quote markers can be "interrupted"
    /// like paragraphs, but otherwise can't have other blocks after them.
    ///
    /// Call this at the end of the line to parse that. If it succeeeds,
    /// this returns the LineStart for the new line.
    fn scan_next_line_or_lazy_continuation<'input>(
        &mut self,
        bytes: &'input [u8],
    ) -> Option<LineStart<'input>> {
        let mut line_start = LineStart::new(bytes);
        let tree_position = scan_containers(&self.tree, &mut line_start, self.options);
        let current_container = tree_position == self.tree.spine_len();
        if self.options.contains(Options::ENABLE_MDX) {
            // MDX: indented code blocks are disabled, so consume all leading
            // whitespace and always check for paragraph interrupts.  This
            // ensures deeply-indented code fences interrupt paragraphs.
            line_start.scan_all_space();
            if self.scan_paragraph_interrupt(
                &bytes[line_start.bytes_scanned()..],
                current_container,
                tree_position,
            ) || scan_blank_line(&bytes[line_start.bytes_scanned()..]).is_some()
            {
                None
            } else {
                Some(line_start)
            }
        } else if (!line_start.scan_space(4)
            && self.scan_paragraph_interrupt(
                &bytes[line_start.bytes_scanned()..],
                current_container,
                tree_position,
            ))
            || scan_blank_line(&bytes[line_start.bytes_scanned()..]).is_some()
        {
            None
        } else {
            line_start.scan_all_space();
            Some(line_start)
        }
    }

    /// Returns the offset of the first line after the table.
    /// Assumptions: current focus is a table element and the table header
    /// matches the separator line (same number of columns).
    fn parse_table(
        &mut self,
        table_cols: usize,
        head_start: usize,
        body_start: usize,
    ) -> Option<usize> {
        // filled empty cells are limited to protect against quadratic growth
        // https://github.com/raphlinus/pulldown-cmark/issues/832
        let mut missing_empty_cells = 0;
        // parse header. this shouldn't fail because we made sure the table header is ok
        let (_sep_start, thead_ix) =
            self.parse_table_row_inner(head_start, table_cols, &mut missing_empty_cells)?;
        self.tree[thead_ix].item.body = ItemBody::TableHead;

        // parse body
        let mut ix = body_start;
        while let Some((next_ix, _row_ix)) =
            self.parse_table_row(ix, table_cols, &mut missing_empty_cells)
        {
            ix = next_ix;
        }

        self.pop(ix);
        Some(ix)
    }

    /// Call this when containers are taken care of.
    /// Returns bytes scanned, row_ix
    fn parse_table_row_inner(
        &mut self,
        mut ix: usize,
        row_cells: usize,
        missing_empty_cells: &mut usize,
    ) -> Option<(usize, TreeIndex)> {
        let bytes = self.text.as_bytes();
        let mut cells = 0;
        let mut final_cell_ix = None;

        let old_cur = self.tree.cur();
        let row_ix = self.tree.append(Item {
            start: ix,
            end: 0, // set at end of this function
            body: ItemBody::TableRow,
        });
        self.tree.push();

        let mut first_iter = true;
        let mut saw_opening_pipe = false;
        loop {
            let cell_start = ix;
            let pipe_consumed = scan_ch(&bytes[ix..], b'|');
            ix += pipe_consumed;
            if first_iter && pipe_consumed > 0 {
                saw_opening_pipe = true;
            }
            first_iter = false;
            let _start_ix = ix;
            ix += scan_space_or_tab(&bytes[ix..]);

            if let Some(eol_bytes) = scan_eol(&bytes[ix..]) {
                // A line with only an opening `|` and no content (e.g. stray
                // `              |` between table rows) emits a row with a
                // single empty cell in remark-gfm instead of terminating the
                // table. Mirror that here.
                if saw_opening_pipe && cells == 0 {
                    let empty_cell_ix = self.tree.append(Item {
                        start: cell_start,
                        end: ix,
                        body: ItemBody::TableCell,
                    });
                    final_cell_ix = Some(empty_cell_ix);
                    cells = 1;
                }
                ix += eol_bytes;
                break;
            }

            let cell_ix = self.tree.append(Item {
                start: cell_start,
                end: ix,
                body: ItemBody::TableCell,
            });
            self.tree.push();
            let (next_ix, _brk) = self.parse_line(ix, None, TableParseMode::Active, ix);

            self.tree[cell_ix].item.end = next_ix;
            self.tree.pop();

            ix = next_ix;
            cells += 1;

            if cells == row_cells {
                final_cell_ix = Some(cell_ix);
            }
        }

        if let (Some(cur), 0) = (old_cur, cells) {
            self.pop(ix);
            self.tree[cur].next = None;
            return None;
        }

        // `mdast-util-gfm-table` keeps the source cell count exactly — neither
        // truncated to the header width nor padded with empties. HAST padding
        // happens downstream in `mdast-util-to-hast`, and overflow cells are
        // discarded there too.
        let _ = row_cells;
        let _ = missing_empty_cells;

        // Extend the final cell's end to include any trailing `|` and
        // whitespace, matching remark's convention. With overflow preserved,
        // the final cell is the last parsed cell, not the header-column cell.
        let last_cell_ix = {
            let mut walker = self.tree[row_ix].child;
            let mut last = None;
            while let Some(c) = walker {
                last = Some(c);
                walker = self.tree[c].next;
            }
            last
        };
        if let Some(cell_ix) = last_cell_ix {
            let row_end = ix;
            let mut cell_end = self.tree[cell_ix].item.end;
            let bytes = self.text.as_bytes();
            while cell_end < row_end
                && cell_end < bytes.len()
                && bytes[cell_end] != b'\n'
                && bytes[cell_end] != b'\r'
            {
                cell_end += 1;
            }
            self.tree[cell_ix].item.end = cell_end;
        }
        let _ = final_cell_ix;

        self.pop(ix);

        Some((ix, row_ix))
    }

    /// Returns first offset after the row and the tree index of the row.
    fn parse_table_row(
        &mut self,
        mut ix: usize,
        row_cells: usize,
        missing_empty_cells: &mut usize,
    ) -> Option<(usize, TreeIndex)> {
        let bytes = self.text.as_bytes();
        let mut line_start = LineStart::new(&bytes[ix..]);
        let tree_position = scan_containers(&self.tree, &mut line_start, self.options);
        let current_container = tree_position == self.tree.spine_len();
        if !current_container {
            return None;
        }
        // A line with ≥4 leading spaces (after container prefixes) is an
        // indented code block, which interrupts table continuation. Only
        // accept up to 3 spaces of leading indentation before the row.
        // MDX disables indented code blocks, so a table nested inside JSX
        // flow can have body rows at any indent — match the head's MDX
        // allowance (see scan_table_head paragraph_interrupt).
        if self.options.contains(Options::ENABLE_MDX) {
            line_start.scan_all_space();
        } else {
            let _ = line_start.scan_space_upto(3);
            if !line_start.is_at_eol() && line_start.scan_space(1) {
                return None;
            }
        }
        ix += line_start.bytes_scanned();
        if scan_paragraph_interrupt_no_table(
            &bytes[ix..],
            current_container,
            self.options.contains(Options::ENABLE_FOOTNOTES),
            self.options.contains(Options::ENABLE_DEFINITION_LIST),
            self.options.contains(Options::ENABLE_MDX),
            self.options.contains(Options::ENABLE_MATH_MULTI_DOLLAR),
            self.options.contains(Options::ENABLE_DIRECTIVE),
            &self.tree,
            tree_position,
        ) {
            return None;
        }

        let (ix, row_ix) = self.parse_table_row_inner(ix, row_cells, missing_empty_cells)?;
        Some((ix, row_ix))
    }

    /// Returns offset of line start after paragraph.
    /// Append a container directive's `[label]` as a `DirectiveLabel` child
    /// holding inline-tokenized content. The label paragraph spans the brackets
    /// (`[` … `]`); its children come from the normal inline pass run over the
    /// inner content, so emphasis/strong/links/code resolve like anywhere else.
    /// Must be called with the directive itself on the spine.
    fn append_container_directive_label(&mut self, label_start: usize, label_end: usize) {
        let bracket_offset = label_start.saturating_sub(1);
        let bracket_end = label_end + 1;
        self.tree.append(Item {
            start: bracket_offset,
            end: bracket_end,
            body: ItemBody::DirectiveLabel,
        });
        self.tree.push();
        if label_start < label_end {
            self.parse_line(
                label_start,
                Some(label_end),
                TableParseMode::Disabled,
                label_start,
            );
        }
        self.tree.pop();
    }

    fn parse_paragraph(&mut self, start_ix: usize, tasklist_marker: Option<Item>) -> usize {
        self.list_interrupted_paragraph = false;
        let body = if let Some(ItemBody::DefinitionList(_)) =
            self.tree.peek_up().map(|idx| self.tree[idx].item.body)
        {
            if self.tree.cur().is_none_or(|idx| {
                matches!(
                    &self.tree[idx].item.body,
                    ItemBody::DefinitionListDefinition(..)
                )
            }) {
                // blank lines between the previous definition and this one don't count
                self.last_line_blank = false;
                ItemBody::MaybeDefinitionListTitle
            } else {
                self.finish_list(start_ix);
                ItemBody::Paragraph
            }
        } else {
            self.finish_list(start_ix);
            ItemBody::Paragraph
        };
        let node_ix = self.tree.append(Item {
            start: start_ix,
            end: 0, // will get set later
            body,
        });
        self.tree.push();

        if let Some(item) = tasklist_marker {
            self.tree.append(item);
        }

        let bytes = self.text.as_bytes();
        let mut ix = start_ix;
        loop {
            let scan_mode = if self.options.contains(Options::ENABLE_TABLES) && ix == start_ix {
                TableParseMode::Scan
            } else {
                TableParseMode::Disabled
            };
            let (next_ix, brk) = self.parse_line(ix, None, scan_mode, start_ix);

            // break out when we find a table
            if let Some(Item {
                body: ItemBody::Table(alignment_ix),
                ..
            }) = brk
            {
                let table_cols = self.allocs[alignment_ix].len();
                self.tree[node_ix].item.body = ItemBody::Table(alignment_ix);
                // this clears out any stuff we may have appended - but there may
                // be a cleaner way
                self.tree[node_ix].child = None;
                self.tree.pop();
                if body == ItemBody::MaybeDefinitionListTitle {
                    self.finish_list(ix);
                }
                self.tree.push();
                if let Some(ix) = self.parse_table(table_cols, ix, next_ix) {
                    return ix;
                }
            }

            ix = next_ix;
            let mut line_start = LineStart::new(&bytes[ix..]);
            let tree_position = scan_containers(&self.tree, &mut line_start, self.options);
            let current_container = tree_position == self.tree.spine_len();

            let trailing_backslash_pos = match brk {
                Some(Item {
                    start,
                    body: ItemBody::HardBreak(true),
                    ..
                }) if bytes[start] == b'\\' => Some(start),
                _ => None,
            };
            // In MDX mode, indented code blocks are disabled, so consume all
            // leading whitespace and always check for paragraph interrupts.
            let is_indented = if self.options.contains(Options::ENABLE_MDX) {
                line_start.scan_all_space();
                false
            } else {
                line_start.scan_space(4)
            };
            if !is_indented {
                let ix_new = ix + line_start.bytes_scanned();
                if current_container
                    && let Some(ix_setext) =
                        self.parse_setext_heading(ix_new, node_ix, trailing_backslash_pos.is_some())
                {
                    if let Some(pos) = trailing_backslash_pos {
                        self.tree.append_text(pos, pos + 1, false);
                    }
                    self.pop(ix_setext);
                    if body == ItemBody::MaybeDefinitionListTitle {
                        self.finish_list(ix);
                    }
                    return ix_setext;
                }
                // first check for non-empty lists, then for other interrupts
                let suffix = &bytes[ix_new..];
                if self.scan_paragraph_interrupt(suffix, current_container, tree_position) {
                    // MDX-only suppression: if the previous line opened an
                    // inline JSX tag whose body spans into this line, the
                    // tag is still open and the line's would-be interrupt
                    // is actually part of the JSX (typically the closing
                    // `>`). Don't break the paragraph here — let the inline
                    // resolver pick up the JSX.
                    if !(self.options.contains(Options::ENABLE_MDX)
                        && prev_line_has_open_inline_jsx(bytes, ix, self.options.has_math()))
                    {
                        if let Some(pos) = trailing_backslash_pos {
                            self.tree.append_text(pos, pos + 1, false);
                        }
                        self.list_interrupted_paragraph = scan_listitem(suffix).is_some()
                            || scan_blockquote_start(suffix).is_some();
                        // Type-7 HTML block on a lazy-continuation line is
                        // a special case in remark/micromark: the new HTML
                        // block opens INSIDE the still-open container (the
                        // line is fed through the container's child flow
                        // parser, where type-7 is allowed). Open it inline
                        // for BlockQuote and ListItem parents so the HTML
                        // becomes a sibling of the open paragraph instead
                        // of popping the container.
                        //
                        // Exception: if the html line is at EOF without a
                        // trailing newline, micromark's lazy handling treats
                        // it as an incomplete line and the html ends up at
                        // root (e.g. `- a\n<a>` → `[list, html]`, vs
                        // `- a\n<a>\n` → `[list[item[para,html]]]`).
                        if !self.options.contains(Options::ENABLE_MDX)
                            && !current_container
                            && suffix.starts_with(b"<")
                            && scan_html_type_7(suffix).is_some()
                            && !starts_html_block_type_6(&suffix[1..])
                            && get_html_end_tag(&suffix[1..]).is_none()
                        {
                            // Find end of the html line. If there's no
                            // newline anywhere in `suffix`, the line is at
                            // EOF without a trailing newline — fall through
                            // to the normal break path so containers pop.
                            let line_terminated = suffix.iter().any(|&b| b == b'\n' || b == b'\r');
                            // `tree_position` is the count of spine items
                            // that matched container prefixes. The container
                            // we're STILL inside is at spine index
                            // `tree_position` (the first one that did NOT
                            // match), since lazy continuation keeps the
                            // unmatched container open.
                            let parent_is_container =
                                self.tree.walk_spine().nth(tree_position).is_some_and(|ix| {
                                    matches!(
                                        self.tree[*ix].item.body,
                                        ItemBody::BlockQuote(..) | ItemBody::ListItem(..)
                                    )
                                });
                            if parent_is_container && line_terminated {
                                self.pop(ix);
                                return self.parse_html_block_type_6_or_7(
                                    ix_new,
                                    line_start.remaining_space(),
                                    0,
                                );
                            }
                        }
                        break;
                    }
                }
                if self.options.contains(Options::ENABLE_DIRECTIVE)
                    && !current_container
                    && line_start.scan_closing_container_extensions_fence(3)
                {
                    break;
                }
            }
            line_start.scan_all_space();
            if line_start.is_at_eol() {
                if let Some(pos) = trailing_backslash_pos {
                    self.tree.append_text(pos, pos + 1, false);
                }
                break;
            }

            if self.options.contains(Options::ENABLE_DIRECTIVE) {
                let mut closes = false;
                for &node_ix in self.tree.walk_spine().rev().skip(1) {
                    match self.tree[node_ix].item.body {
                        ItemBody::ContainerDirective(length, ..) => {
                            let probe = line_start.clone();
                            if line_start.scan_closing_container_extensions_fence(length) {
                                closes = true;
                                break;
                            }
                            line_start = probe;
                        }
                        // Walk past list/listItem — closing `:::` can sit at the
                        // list's indent level and still close an outer directive
                        // (matches remark-directive). A blockquote on the spine,
                        // by contrast, hides an inner `:::` as its own content.
                        ItemBody::List(..) | ItemBody::ListItem(..) => {}
                        _ => break,
                    }
                }

                if closes {
                    break;
                }
            }

            ix = next_ix + line_start.bytes_scanned();
            if let Some(item) = brk {
                self.tree.append(item);
            }
        }

        self.pop(ix);
        ix
    }

    /// Returns end ix of setext_heading on success.
    fn parse_setext_heading(
        &mut self,
        ix: usize,
        node_ix: TreeIndex,
        has_trailing_content: bool,
    ) -> Option<usize> {
        let bytes = self.text.as_bytes();
        let (n, level) = scan_setext_heading(&bytes[ix..])?;
        let mut attrs = None;

        if let Some(cur_ix) = self.tree.cur() {
            let parent_ix = self.tree.peek_up().unwrap();
            let header_start = self.tree[parent_ix].item.start;
            // Note that `self.tree[parent_ix].item.end` might be zero at this point.
            // Use the end position of the current node (i.e. the last known child
            // of the parent) instead.
            let header_end = self.tree[cur_ix].item.end;

            // Truncating to the block's `{` (below) drops any trailing directive
            // run with it, so reparse that run to keep its own `{...}` attributes.
            let (content_end, attrs_, tail_start) =
                match self.extract_and_parse_heading_attribute_block(header_start, header_end) {
                    Some(block) => (block.block_open, Some(block.attrs), block.tail_start),
                    None => (header_end, None, None),
                };
            attrs = attrs_;

            // The block already errored as an expression during the paragraph
            // pass, before the underline marked this a heading; drop that stale
            // error now that it's attributes (a directive tail reparses below).
            #[cfg(feature = "mdx")]
            if attrs.is_some() && self.options.contains(Options::ENABLE_MDX) {
                self.mdx_errors
                    .retain(|(offset, _)| *offset < content_end || *offset >= header_end);
            }

            // strip trailing whitespace
            let new_end = if has_trailing_content {
                content_end
            } else {
                let mut last_line_start = header_start;
                if attrs.is_some() {
                    loop {
                        let next_line_start =
                            last_line_start + scan_nextline(&bytes[last_line_start..content_end]);
                        if next_line_start >= content_end {
                            break;
                        }
                        let mut line_start = LineStart::new(&bytes[next_line_start..content_end]);
                        if scan_containers(&self.tree, &mut line_start, self.options)
                            != self.tree.spine_len()
                        {
                            break;
                        }
                        last_line_start = next_line_start + line_start.bytes_scanned();
                    }
                }
                let trailing_ws =
                    scan_rev_while(&bytes[last_line_start..content_end], is_space_or_tab);
                content_end - trailing_ws
            };

            if attrs.is_some() {
                // remove trailing block attributes
                self.tree.truncate_siblings(new_end);
            }

            if let Some(cur_ix) = self.tree.cur() {
                self.tree[cur_ix].item.end = new_end;
            }

            if let Some(tail_start) = tail_start {
                self.parse_line(
                    tail_start,
                    Some(header_end),
                    TableParseMode::Disabled,
                    tail_start,
                );
            }
        }

        self.tree[node_ix].item.body = ItemBody::Heading(
            level,
            attrs.map(|attrs| self.allocs.allocate_heading(attrs)),
        );

        // remark/micromark quirk: when a setext heading directly follows a
        // run of link reference definitions (no blank line separating
        // any of them), the heading's position extends back to the start
        // of the FIRST definition in that adjacent run. Micromark opens
        // one paragraph token spanning the whole chain; each definition
        // splits off as its own node but the heading inherits the
        // chain's original start.
        let bytes = self.text.as_bytes();
        let is_adjacent = |prev_end: usize, next_start: usize| -> bool {
            if next_start <= prev_end {
                return false;
            }
            // Allow exactly one line ending (the def's line terminator)
            // plus any amount of leading whitespace on the next line.
            // Indented-code can't interrupt the def's still-open
            // paragraph, so even 4+ space indents are valid paragraph
            // continuations and qualify for chain-back.
            let gap = &bytes[prev_end..next_start];
            let mut line_endings = 0;
            let mut i = 0;
            while i < gap.len() {
                let b = gap[i];
                if b == b'\n' || b == b'\r' {
                    if line_endings > 0 {
                        return false;
                    }
                    line_endings += 1;
                    // A CRLF terminates one line, not two.
                    i += if b == b'\r' && gap.get(i + 1) == Some(&b'\n') {
                        2
                    } else {
                        1
                    };
                } else if b == b' ' || b == b'\t' {
                    if line_endings == 0 {
                        return false;
                    }
                    // OK: whitespace after the line break is the
                    // next line's leading indent.
                    i += 1;
                } else {
                    return false;
                }
            }
            line_endings == 1
        };
        let original_start = self.tree[node_ix].item.start;
        // Don't chain back when the heading's *first* content line is
        // itself shaped like a setext underline (≤3 leading spaces, then
        // only `-` or `=`, then optional trailing whitespace). Micromark
        // tries to interpret such a line as an underline for the def's
        // residual paragraph, which fails and resets the paragraph
        // token — so the heading no longer inherits the def's start.
        // Other content-line shapes (plain text, `*`, `+`, …) leave the
        // paragraph token intact and the chain-back applies.
        let first_line_end = memchr::memchr2(b'\n', b'\r', &bytes[original_start..ix])
            .map(|p| original_start + p)
            .unwrap_or(ix);
        let first_line = &bytes[original_start..first_line_end];
        let is_setext_underline_shape = {
            let mut p = 0;
            while p < first_line.len() && first_line[p] == b' ' && p < 3 {
                p += 1;
            }
            if p < first_line.len() && (first_line[p] == b'-' || first_line[p] == b'=') {
                let c = first_line[p];
                let mut q = p + 1;
                while q < first_line.len() && first_line[q] == c {
                    q += 1;
                }
                while q < first_line.len() && (first_line[q] == b' ' || first_line[q] == b'\t') {
                    q += 1;
                }
                q == first_line.len()
            } else {
                false
            }
        };
        if !is_setext_underline_shape {
            let mut cur_start = original_start;
            for def in self.allocs.refdefs_all.iter().rev() {
                let def_end = def.1.span.end;
                let def_start = def.1.span.start;
                if is_adjacent(def_end, cur_start) {
                    cur_start = def_start;
                } else {
                    break;
                }
            }
            if cur_start < original_start {
                self.tree[node_ix].item.start = cur_start;
            }
        }

        Some(ix + n)
    }

    /// Commit a detected autolink. Only sound when nothing can still block it.
    fn append_autolink_link(&mut self, d: AutolinkDetection, begin_text: usize, escaped: bool) {
        let link_ix = self
            .allocs
            .allocate_link(d.link_type, d.url.into(), "".into(), "".into());
        self.tree.append_text(begin_text, d.start, escaped);
        let link_node_ix = self.tree.append(Item {
            start: d.start,
            end: d.end,
            body: ItemBody::Link(link_ix),
        });
        let text_child = self.tree.create_node(Item {
            start: d.start,
            end: d.end,
            body: ItemBody::Text {
                backslash_escaped: false,
            },
        });
        self.tree[link_node_ix].child = Some(text_child);
    }

    /// Record a detected autolink as a zero-width marker: no byte's
    /// tokenization changes, so the candidate can still be dropped.
    fn append_autolink_marker(&mut self, d: AutolinkDetection, begin_text: usize, escaped: bool) {
        let link = self
            .allocs
            .allocate_link(d.link_type, d.url.into(), "".into(), "".into());
        let cand = self.allocs.allocate_autolink_candidate(AutolinkCandidate {
            start: d.start,
            end: d.end,
            link,
        });
        self.tree.append_text(begin_text, d.start, escaped);
        self.tree.append(Item {
            start: d.start,
            end: d.start,
            body: ItemBody::MaybeAutolink(cand),
        });
    }

    /// Parse a line of input, appending text and items to tree.
    ///
    /// Returns: index after line and an item representing the break.
    fn parse_line(
        &mut self,
        start: usize,
        end: Option<usize>,
        mode: TableParseMode,
        scope_start: usize,
    ) -> (usize, Option<Item>) {
        let bytes = self.text.as_bytes();
        let bytes = match end {
            Some(end) => &bytes[..end],
            None => bytes,
        };
        let bytes_len = bytes.len();
        let mut pipes = 0;
        let mut last_pipe_ix = start;
        let mut begin_text = start;
        let mut backslash_escaped = false;
        // End offset of the most-recently-emitted non-text item whose
        // bytes can contain a `\` that is NOT a text-context backslash
        // (e.g. an inline GFM autolink Link whose URL ended in `\`).
        // Used by the `\n` hardbreak check to skip those `\`s.
        let mut last_inline_emission_end: usize = start;
        // Lowest start a further candidate may take: a committed candidate owns
        // its bytes, a deferred one only rules out a retry at the same start.
        let mut candidate_floor: usize = start;
        // Ends of deferred candidates still in play. Overlapping candidates are
        // possible, so this is a set; empty in every line without one.
        let mut deferred_ends: Vec<usize> = Vec::new();

        let (final_ix, brk) = iterate_special_bytes(self.lookup_table, bytes, start, |ix, byte| {
            match byte {
                b'\n' | b'\r' => {
                    if let TableParseMode::Active = mode {
                        return LoopInstruction::BreakAtWith(ix, None);
                    }

                    let mut i = ix;
                    let eol_bytes = scan_eol(&bytes[ix..]).unwrap();

                    let end_ix = ix + eol_bytes;
                    // CommonMark hardbreak: an odd number of trailing source
                    // `\` chars before `\n`. Bytes inside an inline-emitted
                    // Link (e.g. GFM literal autolink consumed a `\` as URL
                    // content) are NOT text-context backslashes — stop the
                    // scan at `last_inline_emission_end` so those don't
                    // count.
                    let trailing_backslashes = {
                        let mut p = ix;
                        while p > last_inline_emission_end && bytes[p - 1] == b'\\' {
                            p -= 1;
                        }
                        ix - p
                    };

                    // GFM table detection: check if the next line is a valid
                    // table delimiter. Runs before hard-break so that inputs
                    // like `foo\\\n|-` resolve to a table (keeping the trailing
                    // backslash in the header cell) rather than a paragraph
                    // with a hard break. Headers without pipes are allowed
                    // (`scan_table_head` still requires a pipe in the
                    // delimiter, so setext headings are unaffected). We skip
                    // delimiters that would also be a valid list-item marker,
                    // since block-level lists take precedence over tables.
                    if mode == TableParseMode::Scan {
                        let next_line_ix = ix + eol_bytes;
                        let mut line_start = LineStart::new(&bytes[next_line_ix..]);
                        if scan_containers(&self.tree, &mut line_start, self.options)
                            == self.tree.spine_len()
                        {
                            // In MDX, the delimiter row of a table nested inside
                            // a list item may have extra leading whitespace
                            // beyond the container continuation. No indented
                            // code blocks, so consume it before scan_table_head.
                            if self.options.contains(Options::ENABLE_MDX) {
                                line_start.scan_all_space();
                            }
                            let table_head_ix = next_line_ix + line_start.bytes_scanned();
                            let delim = &bytes[table_head_ix..];
                            // Allow up to 3 spaces of leading indent: a line
                            // like ` - …` is a valid bullet (and lists win
                            // over table delimiter recognition).
                            let leading_spaces =
                                delim.iter().take(3).take_while(|&&b| b == b' ').count();
                            let delim_is_list_item =
                                scan_listitem(&delim[leading_spaces..]).is_some();
                            let (table_head_bytes, alignment) = if delim_is_list_item {
                                (0, vec![])
                            } else {
                                scan_table_head(delim)
                            };

                            if table_head_bytes > 0 {
                                let header_count =
                                    count_header_cols(bytes, pipes, start, last_pipe_ix);

                                if alignment.len() == header_count {
                                    let alignment_ix = self.allocs.allocate_alignment(alignment);
                                    let end_ix = table_head_ix + table_head_bytes;
                                    return LoopInstruction::BreakAtWith(
                                        end_ix,
                                        Some(Item {
                                            start: i,
                                            end: end_ix, // must update later
                                            body: ItemBody::Table(alignment_ix),
                                        }),
                                    );
                                }
                            }
                        }
                    }

                    if trailing_backslashes % 2 == 1 && end_ix < bytes_len {
                        i -= 1;
                        self.tree.append_text(begin_text, i, backslash_escaped);
                        backslash_escaped = false;
                        return LoopInstruction::BreakAtWith(
                            end_ix,
                            Some(Item {
                                start: i,
                                end: end_ix,
                                body: ItemBody::HardBreak(true),
                            }),
                        );
                    }

                    let trailing_spaces = scan_rev_while(&bytes[..ix], |c| c == b' ');
                    let has_tab_before_spaces = trailing_spaces > 0
                        && ix > trailing_spaces
                        && bytes[ix - trailing_spaces - 1] == b'\t';
                    if trailing_spaces >= 2 && !has_tab_before_spaces {
                        i -= trailing_spaces;
                        self.tree.append_text(begin_text, i, backslash_escaped);
                        backslash_escaped = false;
                        return LoopInstruction::BreakAtWith(
                            end_ix,
                            Some(Item {
                                start: i,
                                end: end_ix,
                                body: ItemBody::HardBreak(false),
                            }),
                        );
                    }

                    let trailing_whitespace = scan_rev_while(&bytes[..ix], is_space_or_tab);
                    self.tree
                        .append_text(begin_text, ix - trailing_whitespace, backslash_escaped);
                    backslash_escaped = false;

                    LoopInstruction::BreakAtWith(
                        end_ix,
                        Some(Item {
                            start: i,
                            end: end_ix,
                            body: ItemBody::SoftBreak,
                        }),
                    )
                }
                b'\\' if bytes.get(ix + 1).copied().is_some_and(is_ascii_punctuation) => {
                    self.tree.append_text(begin_text, ix, backslash_escaped);
                    if bytes[ix + 1] == b'`' {
                        let count = 1 + scan_ch_repeat(&bytes[(ix + 2)..], b'`');
                        self.tree.append(Item {
                            start: ix + 1,
                            end: ix + count + 1,
                            body: ItemBody::MaybeCode(count, true),
                        });
                        begin_text = ix + 1 + count;
                        backslash_escaped = false;
                        LoopInstruction::ContinueAndSkip(count)
                    } else if bytes[ix + 1] == b'|' && TableParseMode::Active == mode {
                        // Yeah, it's super weird that backslash escaped pipes in tables aren't "real"
                        // backslash escapes.
                        //
                        // This tree structure is intended for the benefit of inline analysis, and it
                        // is supposed to operate as-if backslash escaped pipes were stripped out in a
                        // separate pass.
                        begin_text = ix + 1;
                        // The `\` isn't content, but the span still covers it.
                        backslash_escaped = true;
                        LoopInstruction::ContinueAndSkip(1)
                    } else if bytes[ix + 1] == b'<' {
                        // Still emit the marker: a deferred autolink may end on
                        // this `\`, in which case the link owns it and the
                        // inline HTML opens after all.
                        self.tree.append(Item {
                            start: ix + 1,
                            end: ix + 2,
                            body: ItemBody::MaybeHtml(true),
                        });
                        begin_text = ix + 2;
                        backslash_escaped = false;
                        LoopInstruction::ContinueAndSkip(1)
                    } else if bytes[ix + 1] == b'$' && self.options.has_math() {
                        // In math context, \$ should still produce a MaybeMath
                        // delimiter so it can close a math span. The backslash
                        // only prevents opening.
                        begin_text = ix + 1;
                        backslash_escaped = true;
                        LoopInstruction::ContinueAndSkip(0)
                    } else if let Some((count, fired)) = (deferred_ends.contains(&(ix + 1)))
                        .then(|| escaped_delim_run(self.text, start, ix, mode, self.options))
                        .flatten()
                    {
                        // A deferred candidate's URL ends on this `\`, so the
                        // run after it is the link's to unblock. Emitted in the
                        // shape the link firing wants; `MaybeEmphasisEscaped`
                        // holds it back until the splice says so.
                        let blocked = delim_run_flags(
                            self.text,
                            start,
                            ix + 2,
                            count - 1,
                            mode,
                            self.options,
                        );
                        self.tree.append(Item {
                            start: ix + 1,
                            end: ix + 2,
                            body: ItemBody::MaybeEmphasisEscaped(count, fired.0, fired.1),
                        });
                        for i in 1..count {
                            self.tree.append(Item {
                                start: ix + 1 + i,
                                end: ix + 2 + i,
                                body: ItemBody::MaybeEmphasis(count - i, blocked.0, blocked.1),
                            });
                        }
                        begin_text = ix + 1 + count;
                        backslash_escaped = false;
                        LoopInstruction::ContinueAndSkip(count)
                    } else {
                        begin_text = ix + 1;
                        backslash_escaped = true;
                        LoopInstruction::ContinueAndSkip(1)
                    }
                }
                c @ b'*' | c @ b'_' | c @ b'~' | c @ b'^' => {
                    // GFM precedence: an email literal starting at this `_`
                    // wins over the attention sequence; skipping MaybeEmphasis
                    // keeps `_-_@…` from forming a pair that hides the email.
                    if c == b'_' && self.options.contains(Options::ENABLE_GFM) {
                        let paragraph_floor = self
                            .tree
                            .peek_up()
                            .map(|nix| self.tree[nix].item.start)
                            .unwrap_or(start);
                        if let Some((email_start, email_end, full_url)) =
                            scan_email_forward_from_atext(bytes, ix, begin_text, paragraph_floor)
                            && email_start >= candidate_floor
                        {
                            let d = AutolinkDetection {
                                start: email_start,
                                end: email_end,
                                link_type: LinkType::Email,
                                url: full_url
                                    .strip_prefix("mailto:")
                                    .map(str::to_owned)
                                    .unwrap_or(full_url),
                            };
                            if defer_autolink_decision(bytes, paragraph_floor, ix, self.options) {
                                candidate_floor = email_start + 1;
                                // Fall through to attention handling: a
                                // marker that fires splices those
                                // delimiters away, and a blocked one
                                // wanted them.
                                self.append_autolink_marker(d, begin_text, backslash_escaped);
                                if email_start > begin_text {
                                    backslash_escaped = false;
                                }
                                begin_text = email_start;
                            } else {
                                candidate_floor = email_end;
                                self.append_autolink_link(d, begin_text, backslash_escaped);
                                backslash_escaped = false;
                                begin_text = email_end;
                                last_inline_emission_end = email_end;
                                let skip = email_end.saturating_sub(ix + 1);
                                return LoopInstruction::ContinueAndSkip(skip);
                            }
                        }
                    }
                    let string_suffix = &self.text[ix..];
                    let count = 1 + scan_ch_repeat(&string_suffix.as_bytes()[1..], c);
                    let can_open = delim_run_can_open(
                        &self.text[start..],
                        string_suffix,
                        count,
                        ix - start,
                        mode,
                        self.options,
                    );
                    let can_close = delim_run_can_close(
                        &self.text[start..],
                        string_suffix,
                        count,
                        ix - start,
                        mode,
                        self.options,
                    );
                    let is_valid_seq = delim_run_is_valid(c, count, self.options);

                    if (can_open || can_close) && is_valid_seq {
                        self.tree.append_text(begin_text, ix, backslash_escaped);
                        backslash_escaped = false;
                        for i in 0..count {
                            self.tree.append(Item {
                                start: ix + i,
                                end: ix + i + 1,
                                body: ItemBody::MaybeEmphasis(count - i, can_open, can_close),
                            });
                        }
                        begin_text = ix + count;
                    }
                    LoopInstruction::ContinueAndSkip(count - 1)
                }
                b'$' => {
                    let brace_context =
                        if self.brace_context_stack.len() > MATH_BRACE_CONTEXT_MAX_NESTING {
                            self.brace_context_next as u8
                        } else {
                            self.brace_context_stack.last().copied().unwrap_or_else(|| {
                                self.brace_context_stack.push(!0);
                                !0
                            })
                        };

                    // `backslash_escaped` applies to the `$` itself only when
                    // the escape sits directly before it (`\$`, pending text
                    // run empty). For `\\$` or `\X$` the escape is consumed by
                    // the earlier char and must not bleed into the delimiter.
                    let dollar_escaped = backslash_escaped && begin_text == ix;
                    self.tree.append_text(begin_text, ix, backslash_escaped);
                    self.tree.append(Item {
                        start: ix,
                        end: ix + 1,
                        body: ItemBody::MaybeMath(dollar_escaped, brace_context),
                    });
                    begin_text = ix + 1;
                    backslash_escaped = false;
                    LoopInstruction::ContinueAndSkip(0)
                }
                #[cfg(feature = "mdx")]
                b'{' if self.options.contains(Options::ENABLE_MDX) => {
                    // If `{` sits inside a pair of matching backtick runs on
                    // the current line, it's part of a code span's text —
                    // code spans take priority over MDX expressions in
                    // remark. Skip inline-expression detection so the `{` is
                    // consumed as literal text (the enclosing code span will
                    // pick it up when backtick pairing resolves).
                    //
                    // Same treatment for `{` inside a CommonMark link URL
                    // `[...](...)`: mdx-js does not evaluate expressions in
                    // URLs (`[a]({x})` round-trips with URL "{x}", literal),
                    // so treat the `{` as plain text and let the link
                    // resolver claim the bytes. This also avoids a hard
                    // parse error on unmatched `{` like `[a]({)`.
                    //
                    // Inline math `$...$` owns its content too: braces in
                    // LaTeX (`\frac{-b}{2a}`) are math text, not expressions —
                    // matching block `$$` and the autolink math-span check.
                    if is_inside_code_span(bytes, ix)
                        || is_inside_link_url_parens(bytes, ix, scope_start)
                        || is_inside_open_inline_jsx_tag(bytes, ix)
                        || (self.options.has_math() && is_inside_math_span(bytes, ix))
                    {
                        LoopInstruction::ContinueAndSkip(0)
                    } else {
                        // MDX inline expression: try to scan balanced braces.
                        // Lazy-paragraph continuation rules differ between
                        // text- and flow-position `{`. mdx-js's text
                        // tokenizer (`{` after content on a paragraph line)
                        // sets `allowLazy: true`, so body chars on a lazy
                        // line are kept. Its flow tokenizer (`{` first on a
                        // line in a container) sets `allowLazy: false` and
                        // errors. The block-level pass already tried flow;
                        // fall-through here means the flow scan failed, but
                        // we still need to reproduce its strict-lazy
                        // behavior for `{` at line start.
                        let scan_result = if self.tree.spine_len() > 0 {
                            let check = self.make_container_line_check();
                            let allow_lazy_body = !is_at_paragraph_line_start(bytes, ix);
                            scan_mdx_inline_expression_in_container(
                                &bytes[ix..],
                                &check,
                                allow_lazy_body,
                            )
                        } else {
                            scan_mdx_inline_expression(&bytes[ix..])
                        };
                        if let Some((content_start, content_end, total_len)) = scan_result {
                            self.tree.append_text(begin_text, ix, backslash_escaped);
                            backslash_escaped = false;
                            // Strip container prefixes (e.g. blockquote `>`)
                            // from continuation lines and apply the 2-col
                            // indent dedent. Combined in one walk so the
                            // tab-stop math sees the correct per-line
                            // starting column (lazy lines start at col 0;
                            // strict lines start at the post-prefix column).
                            let (normalized, offset_map) =
                                self.inline_expression_value(ix + content_start, ix + content_end);
                            // Validate the expression body as JS via oxc.
                            // Without this, `{h<}` etc. silently produce a
                            // phantom mdxTextExpression and only error at
                            // JS emit. Allocator is reused across calls.
                            if let Some((err_offset, detail)) =
                                crate::mdx::try_parse_expression_body(
                                    &normalized,
                                    &mut self.mdx_expr_allocator,
                                )
                            {
                                // For single-line bodies the map is empty and
                                // the normalized text is a verbatim slice, so a
                                // direct offset is exact; multi-line bodies
                                // resolve through the map.
                                let source_offset =
                                    satteri_arena::mdx_types::Location::relative_to_absolute(
                                        &offset_map,
                                        err_offset,
                                    )
                                    .unwrap_or(ix + content_start + err_offset);
                                self.mdx_errors.push((
                                    source_offset,
                                    format!("Could not parse expression with oxc: {detail}"),
                                ));
                            }
                            let cow_ix = self.allocs.allocate_cow(normalized.into());
                            self.tree.append(Item {
                                start: ix,
                                end: ix + total_len,
                                body: ItemBody::MdxTextExpression(cow_ix),
                            });
                            begin_text = ix + total_len;
                            LoopInstruction::ContinueAndSkip(total_len - 1)
                        } else {
                            // Unclosed expression.
                            self.mdx_errors.push((
                                ix,
                                "Unexpected end of file in expression, expected a corresponding \
                             closing brace for `{`"
                                    .to_string(),
                            ));
                            LoopInstruction::ContinueAndSkip(0)
                        }
                    }
                }
                b'{' => {
                    if self.brace_context_stack.len() == MATH_BRACE_CONTEXT_MAX_NESTING {
                        self.brace_context_stack.push(self.brace_context_next as u8);
                        self.brace_context_next = MATH_BRACE_CONTEXT_MAX_NESTING;
                    } else if self.brace_context_stack.len() > MATH_BRACE_CONTEXT_MAX_NESTING {
                        // When we reach the limit of nesting, switch from actually matching
                        // braces to just counting them.
                        self.brace_context_next += 1;
                    } else if !self.brace_context_stack.is_empty() {
                        // Store nothing if no math environment has been reached yet.
                        self.brace_context_stack.push(self.brace_context_next as u8);
                        self.brace_context_next += 1;
                    }
                    LoopInstruction::ContinueAndSkip(0)
                }
                b'}' => {
                    if let &mut [ref mut top_level_context] = &mut self.brace_context_stack[..] {
                        // Unbalanced Braces
                        //
                        // The initial, root top-level brace context is -1, but this is changed whenever an unbalanced
                        // close brace is encountered:
                        //
                        //     This is not a math environment: $}$
                        //     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^|^
                        //     -1                               |-2
                        //
                        // To ensure this can't get parsed as math, each side of the unbalanced
                        // brace is an irreversibly separate brace context. As long as the math
                        // environment itself contains balanced braces, they should share a top level context.
                        //
                        //     Math environment contains 2+2: $}$2+2$
                        //                                       ^^^ this is a math environment
                        *top_level_context = top_level_context.wrapping_sub(1);
                    } else if self.brace_context_stack.len() > MATH_BRACE_CONTEXT_MAX_NESTING {
                        // When we exceed 25 levels of nesting, switch from accurately balancing braces
                        // to just counting them. When we dip back below the limit, switch back.
                        if self.brace_context_next <= MATH_BRACE_CONTEXT_MAX_NESTING {
                            self.brace_context_stack.pop();
                        } else {
                            self.brace_context_next -= 1;
                        }
                    } else {
                        self.brace_context_stack.pop();
                    }
                    LoopInstruction::ContinueAndSkip(0)
                }
                b'`' => {
                    let count = 1 + scan_ch_repeat(&bytes[(ix + 1)..], b'`');
                    self.tree.append_text(begin_text, ix, backslash_escaped);
                    backslash_escaped = false;
                    self.tree.append(Item {
                        start: ix,
                        end: ix + count,
                        body: ItemBody::MaybeCode(count, false),
                    });
                    begin_text = ix + count;
                    LoopInstruction::ContinueAndSkip(count - 1)
                }
                b'<' if self.options.contains(Options::ENABLE_MDX)
                    || bytes.get(ix + 1) != Some(&b'\\') =>
                {
                    // In MDX mode `<\…` is not a CommonMark backslash escape;
                    // the MaybeHtml resolver below validates it (and rejects
                    // `<\>` as an invalid JSX tag start).
                    // Note: could detect some non-HTML cases and early escape here, but not
                    // clear that's a win.
                    self.tree.append_text(begin_text, ix, backslash_escaped);
                    backslash_escaped = false;
                    self.tree.append(Item {
                        start: ix,
                        end: ix + 1,
                        body: ItemBody::MaybeHtml(false),
                    });
                    begin_text = ix + 1;
                    LoopInstruction::ContinueAndSkip(0)
                }
                b'!' if bytes.get(ix + 1) == Some(&b'[') => {
                    self.tree.append_text(begin_text, ix, backslash_escaped);
                    backslash_escaped = false;
                    self.tree.append(Item {
                        start: ix,
                        end: ix + 2,
                        body: ItemBody::MaybeImage,
                    });
                    begin_text = ix + 2;
                    LoopInstruction::ContinueAndSkip(1)
                }
                b'[' => {
                    self.tree.append_text(begin_text, ix, backslash_escaped);
                    backslash_escaped = false;
                    self.tree.append(Item {
                        start: ix,
                        end: ix + 1,
                        body: ItemBody::MaybeLinkOpen,
                    });
                    begin_text = ix + 1;
                    LoopInstruction::ContinueAndSkip(0)
                }
                b']' => {
                    self.tree.append_text(begin_text, ix, backslash_escaped);
                    backslash_escaped = false;
                    self.tree.append(Item {
                        start: ix,
                        end: ix + 1,
                        body: ItemBody::MaybeLinkClose(true),
                    });
                    begin_text = ix + 1;
                    LoopInstruction::ContinueAndSkip(0)
                }
                b'&' => match scan_entity(&bytes[ix..]) {
                    (n, Some(value)) => {
                        self.tree.append_text(begin_text, ix, backslash_escaped);
                        backslash_escaped = false;
                        self.tree.append(Item {
                            start: ix,
                            end: ix + n,
                            body: ItemBody::SynthesizeText(self.allocs.allocate_cow(value)),
                        });
                        begin_text = ix + n;
                        LoopInstruction::ContinueAndSkip(n - 1)
                    }
                    _ => LoopInstruction::ContinueAndSkip(0),
                },
                b':' if self.options.contains(Options::ENABLE_DIRECTIVE) => {
                    // Text directive: :name[label]{attrs}
                    // Must not be preceded by another colon (to avoid ::, :::)
                    if ix > 0 && bytes[ix - 1] == b':' {
                        LoopInstruction::ContinueAndSkip(0)
                    } else if is_inside_code_span(bytes, ix) {
                        // Code spans bind tighter: a `:` inside one is literal, not
                        // a directive whose label scan would reach past the span's
                        // backticks and swallow them (issue #158).
                        LoopInstruction::ContinueAndSkip(0)
                    } else if let Some((dir_data, end_pos)) =
                        parse_directive_after_colons(self.text, bytes, ix + 1)
                    {
                        // :name: (followed by colon) is NOT a directive (emoji compat)
                        if end_pos < bytes.len() && bytes[end_pos] == b':' {
                            let name_end = ix + 1 + dir_data.name.len();
                            if name_end == end_pos {
                                // bare :name: with no label/attrs
                                return LoopInstruction::ContinueAndSkip(0);
                            }
                        }
                        self.tree.append_text(begin_text, ix, backslash_escaped);
                        backslash_escaped = false;
                        let label_start = dir_data.label_start;
                        let label_end = dir_data.label_end;
                        let dir_ix = self.allocs.allocate_directive(dir_data);
                        let consumed = end_pos - ix;
                        self.tree.append(Item {
                            start: ix,
                            end: end_pos,
                            body: ItemBody::TextDirective(dir_ix),
                        });
                        // Tokenize the label inline as the directive's children,
                        // re-entering the inline scanner over the `[…]` span.
                        if label_start < label_end {
                            self.tree.push();
                            self.parse_line(
                                label_start,
                                Some(label_end),
                                TableParseMode::Disabled,
                                label_start,
                            );
                            self.tree.pop();
                        }
                        begin_text = end_pos;
                        LoopInstruction::ContinueAndSkip(consumed - 1)
                    } else {
                        LoopInstruction::ContinueAndSkip(0)
                    }
                }
                b'|' => {
                    // Only escaped when an odd number of backslashes precedes
                    // the pipe. `\|` → escaped; `\\|` → literal `\` then
                    // separator pipe.
                    let preceding_backslashes = scan_rev_while(&bytes[..ix], |b| b == b'\\');
                    if preceding_backslashes % 2 == 1 {
                        LoopInstruction::ContinueAndSkip(0)
                    } else if let TableParseMode::Active = mode {
                        LoopInstruction::BreakAtWith(ix, None)
                    } else {
                        last_pipe_ix = ix;
                        pipes += 1;
                        LoopInstruction::ContinueAndSkip(0)
                    }
                }
                b'.' if matches!(bytes.get(ix + 1..), Some(&[b'.', b'.', ..])) => {
                    self.tree.append_text(begin_text, ix, backslash_escaped);
                    backslash_escaped = false;
                    self.tree.append(Item {
                        start: ix,
                        end: ix + 3,
                        body: ItemBody::SynthesizeChar('…'),
                    });
                    begin_text = ix + 3;
                    LoopInstruction::ContinueAndSkip(2)
                }
                b'-' => {
                    let count = 1 + scan_ch_repeat(&bytes[(ix + 1)..], b'-');
                    if count == 1 {
                        LoopInstruction::ContinueAndSkip(0)
                    } else {
                        let itembody = if count == 2 {
                            ItemBody::SynthesizeChar('–')
                        } else if count == 3 {
                            ItemBody::SynthesizeChar('—')
                        } else {
                            let buf = crate::post_passes::smart_dash_run(count);
                            ItemBody::SynthesizeText(self.allocs.allocate_cow(buf.into()))
                        };

                        self.tree.append_text(begin_text, ix, backslash_escaped);
                        backslash_escaped = false;
                        self.tree.append(Item {
                            start: ix,
                            end: ix + count,
                            body: itembody,
                        });
                        begin_text = ix + count;
                        LoopInstruction::ContinueAndSkip(count - 1)
                    }
                }
                c @ b'\'' | c @ b'"' => {
                    let string_suffix = &self.text[ix..];
                    let can_open = delim_run_can_open(
                        &self.text[start..],
                        string_suffix,
                        1,
                        ix - start,
                        mode,
                        self.options,
                    );
                    let can_close = delim_run_can_close(
                        &self.text[start..],
                        string_suffix,
                        1,
                        ix - start,
                        mode,
                        self.options,
                    );

                    self.tree.append_text(begin_text, ix, backslash_escaped);
                    backslash_escaped = false;
                    self.tree.append(Item {
                        start: ix,
                        end: ix + 1,
                        body: ItemBody::MaybeSmartQuote(c, can_open, can_close),
                    });
                    begin_text = ix + 1;

                    LoopInstruction::ContinueAndSkip(0)
                }
                b'h' | b'H' | b'w' | b'W' | b'@' if self.options.contains(Options::ENABLE_GFM) => {
                    // GFM literal autolink. Only the construct path lands here;
                    // `gfm_autolink_literal_pass` is the backstop for the rest.
                    //
                    // `start` is the current *line* start, so take the floor
                    // from the Paragraph on the spine: a `[` on an earlier
                    // line still has to count.
                    let paragraph_floor = self
                        .tree
                        .peek_up()
                        .map(|ix| self.tree[ix].item.start)
                        .unwrap_or(start);
                    let detection =
                        detect_gfm_autolink(bytes, ix, byte, paragraph_floor, begin_text)
                            .filter(|d| d.start >= candidate_floor);
                    if let Some(d) = detection {
                        let (cand_start, cand_end) = (d.start, d.end);
                        if defer_autolink_decision(bytes, paragraph_floor, ix, self.options) {
                            candidate_floor = cand_start + 1;
                            // The scan only moves forward, so ends behind it can
                            // never be probed again; dropping them here keeps the
                            // probe in the escape arm off a growing list.
                            deferred_ends.retain(|&e| e > ix);
                            deferred_ends.push(cand_end);
                            // Leave `begin_text` at the candidate's start: its
                            // bytes stay ordinary text unless the marker fires.
                            self.append_autolink_marker(d, begin_text, backslash_escaped);
                            if cand_start > begin_text {
                                backslash_escaped = false;
                            }
                            begin_text = cand_start;
                            LoopInstruction::ContinueAndSkip(0)
                        } else {
                            candidate_floor = cand_end;
                            self.append_autolink_link(d, begin_text, backslash_escaped);
                            begin_text = cand_end;
                            last_inline_emission_end = cand_end;
                            backslash_escaped = false;
                            // Skip the URL bytes so later callbacks don't
                            // re-trigger inside it; ContinueAndSkip(N) advances
                            // by N then +1.
                            LoopInstruction::ContinueAndSkip(cand_end.saturating_sub(ix + 1))
                        }
                    } else {
                        LoopInstruction::ContinueAndSkip(0)
                    }
                }
                _ => LoopInstruction::ContinueAndSkip(0),
            }
        });

        if brk.is_none() {
            let trailing_whitespace = scan_rev_while(&bytes[begin_text..final_ix], is_space_or_tab);
            // need to close text at eof
            self.tree.append_text(
                begin_text,
                final_ix - trailing_whitespace,
                backslash_escaped,
            );
        }
        (final_ix, brk)
    }

    /// When start_ix is at the beginning of an HTML block of type 1 to 5,
    /// this will find the end of the block, adding the block itself to the
    /// tree and also keeping track of the lines of HTML within the block.
    ///
    /// The html_end_tag is the tag that must be found on a line to end the block.
    /// True when `line_start` (already advanced past container prefixes) sits
    /// on a *pure* closing `:::` fence for an enclosing container directive:
    /// colons only, indented at most 3 spaces, nothing but whitespace after,
    /// and at least as long as that directive's opening fence. HTML blocks
    /// consult this so they stop before the fence instead of swallowing it —
    /// remark-directive recognizes the close at the container level, before
    /// the block's continuation rule fires.
    fn at_closing_directive_fence(&self, line_start: &LineStart<'_>) -> bool {
        if !self.options.contains(Options::ENABLE_DIRECTIVE) {
            return false;
        }
        for &node_ix in self.tree.walk_spine().rev() {
            match self.tree[node_ix].item.body {
                ItemBody::ContainerDirective(length, ..) => {
                    // Keep walking outward on a miss: a shorter fence can still
                    // close an ancestor directive that opened with fewer colons.
                    let mut probe = line_start.clone();
                    let _ = probe.scan_space_upto(3);
                    if probe.scan_closing_container_extensions_fence(length) {
                        probe.scan_all_space();
                        if probe.is_at_eol() {
                            return true;
                        }
                    }
                }
                ItemBody::HtmlBlock(..) | ItemBody::List(..) | ItemBody::ListItem(..) => {}
                _ => break,
            }
        }
        false
    }

    fn parse_html_block_type_1_to_5(
        &mut self,
        start_ix: usize,
        html_end_tag: &str,
        mut remaining_space: usize,
        mut indent: usize,
    ) -> usize {
        // HTML block is a full block — clear lingering empty-list-marker
        // suppression from a prior block (see `parse_hrule` comment).
        self.list_interrupted_paragraph = false;
        // Tree-position of the just-appended HtmlBlock — patched below to
        // record whether the closer pattern was found (so the value gets its
        // trailing newline trimmed) or the block ran to EOF without one.
        let block_node = self.tree.append(Item {
            start: start_ix,
            end: 0, // set later
            body: ItemBody::HtmlBlock(false),
        });
        self.tree.push();

        let bytes = self.text.as_bytes();
        let mut ix = start_ix;
        let end_ix;
        // Two distinct close paths:
        //   * `closer_pattern_found`: the `html_end_tag` matched on a line —
        //     remark trims the trailing newline of the close line.
        //   * container exit / EOF: the block was terminated externally
        //     (parent list item ended, document ended). The trailing newline
        //     of the LAST html line is content per remark, so we don't trim.
        let mut closer_pattern_found = false;
        // Spine depth above the HTML block itself — if we drop below this,
        // a parent container has ended (e.g. list item closed by blank
        // line + outdent). The HTML block ends WITHOUT consuming the
        // boundary line (matches micromark, which has the parent close
        // before the HTML block's continuation rule fires).
        let parent_spine_len = self.tree.spine_len() - 1;
        // Whether any ancestor container is a blockquote. micromark treats
        // the line ending right before a blockquote-close (blank line +
        // outdent) as part of the blockquote's separator, not the HTML
        // block content — so `> <style\n\nfoo` emits `<style` (no trailing
        // `\n`). A bare list item keeps the `\n`. EOF or continuation-
        // marker close ALSO keeps the `\n` for both. The flag below is
        // only consulted on the blank-line exit path.
        let ancestor_has_blockquote = self
            .tree
            .walk_spine()
            .any(|&ix| matches!(self.tree[ix].item.body, ItemBody::BlockQuote(..)));
        // Block ended via two different external close paths:
        //   * lazy: a non-blank line breaks container indent (`* <!-- foo\nbar` —
        //     `bar` at col 0 doesn't satisfy list item indent). REF trims for
        //     ANY parent.
        //   * blank: a blank line ends the parent container. REF trims for
        //     blockquote but KEEPS for list item.
        let mut closed_via_blank = false;
        let mut closed_via_lazy = false;
        loop {
            // Before consuming the current line, peek for a container exit.
            // If we're inside a list/blockquote and the current line is
            // blank, look ahead to the next non-blank line to see if the
            // parent closes there. If yes, end the HTML block now (don't
            // append the blank line).
            if parent_spine_len > 0 && scan_blank_line(&bytes[ix..]).is_some() {
                let mut peek_ix = ix;
                while peek_ix < bytes.len() {
                    if let Some(adv) = scan_blank_line(&bytes[peek_ix..]) {
                        peek_ix += adv;
                    } else {
                        break;
                    }
                }
                if peek_ix == bytes.len() {
                    end_ix = ix;
                    closed_via_blank = true;
                    break;
                }
                let mut peek_line_start = LineStart::new(&bytes[peek_ix..]);
                let n_peek = scan_containers(&self.tree, &mut peek_line_start, self.options);
                if n_peek <= parent_spine_len {
                    end_ix = ix;
                    closed_via_blank = true;
                    break;
                }
            }

            let line_start_ix = ix;
            ix += scan_nextline(&bytes[ix..]);
            self.append_html_line(remaining_space.max(indent), line_start_ix, ix);

            let mut line_start = LineStart::new(&bytes[ix..]);
            let n_containers = scan_containers(&self.tree, &mut line_start, self.options);
            if n_containers < self.tree.spine_len() {
                end_ix = ix;
                if scan_blank_line(&bytes[ix..]).is_some() {
                    closed_via_blank = true;
                } else {
                    // When the outdent is caused by a NEW container marker
                    // (sibling list/bq) opening on this line, remark keeps
                    // the trailing newline of the html block. Only a real
                    // lazy line (paragraph/heading/etc.) trims it.
                    let next_line_ix = ix + line_start.bytes_scanned();
                    let rest = &bytes[next_line_ix..];
                    let opens_container =
                        scan_blockquote_start(rest).is_some() || scan_listitem(rest).is_some();
                    if !opens_container {
                        closed_via_lazy = true;
                    }
                }
                break;
            }

            if self.text[line_start_ix..ix].contains(html_end_tag) {
                end_ix = ix;
                closer_pattern_found = true;
                break;
            }

            let next_line_ix = ix + line_start.bytes_scanned();
            if next_line_ix == self.text.len() {
                end_ix = next_line_ix;
                break;
            }
            // A closing directive fence ends the block before swallowing it.
            if self.at_closing_directive_fence(&line_start) {
                end_ix = ix;
                break;
            }
            ix = next_line_ix;
            remaining_space = line_start.remaining_space();
            indent = 0;
        }
        // Trim trailing `\n` from the html value when:
        //   * the html_end_tag closed the block on its own line
        //   * a lazy line broke the parent container (any parent)
        //   * a blank line closed a blockquote-parented block
        // Don't trim when the block ended via EOF after a continuation
        // marker, or via blank inside a list item.
        let trim_trailing = closer_pattern_found
            || closed_via_lazy
            || (ancestor_has_blockquote && closed_via_blank);
        if trim_trailing {
            self.tree[block_node].item.body = ItemBody::HtmlBlock(true);
        }
        self.pop(end_ix);
        ix
    }

    /// When start_ix is at the beginning of an HTML block of type 6 or 7,
    /// this will consume lines until there is a blank line and keep track of
    /// the HTML within the block.
    fn parse_html_block_type_6_or_7(
        &mut self,
        start_ix: usize,
        mut remaining_space: usize,
        mut indent: usize,
    ) -> usize {
        // HTML block is a full block — clear lingering empty-list-marker
        // suppression from a prior block (see `parse_hrule` comment).
        self.list_interrupted_paragraph = false;
        self.tree.append(Item {
            start: start_ix,
            end: 0, // set later
            body: ItemBody::HtmlBlock(true),
        });
        self.tree.push();

        let bytes = self.text.as_bytes();
        let mut ix = start_ix;
        let end_ix;
        loop {
            let line_start_ix = ix;
            ix += scan_nextline(&bytes[ix..]);
            self.append_html_line(remaining_space.max(indent), line_start_ix, ix);

            let mut line_start = LineStart::new(&bytes[ix..]);
            let n_containers = scan_containers(&self.tree, &mut line_start, self.options);
            if n_containers < self.tree.spine_len() || line_start.is_at_eol() {
                end_ix = ix;
                break;
            }

            let next_line_ix = ix + line_start.bytes_scanned();
            if next_line_ix == self.text.len() || scan_blank_line(&bytes[next_line_ix..]).is_some()
            {
                end_ix = next_line_ix;
                break;
            }
            // A closing directive fence ends the block before swallowing it.
            if self.at_closing_directive_fence(&line_start) {
                end_ix = ix;
                break;
            }
            ix = next_line_ix;
            remaining_space = line_start.remaining_space();
            indent = 0;
        }
        self.pop(end_ix);
        ix
    }

    fn parse_indented_code_block(
        &mut self,
        line_start_ix: usize,
        start_ix: usize,
        mut remaining_space: usize,
    ) -> usize {
        // Consume the lazy-after-blockquote-close marker. When set, this
        // code block is one-line only — micromark's `furtherStart` for
        // `codeIndented` rejects lazy lines, so an indented line that
        // follows the implicitly-closed blockquote can't extend across
        // subsequent (also-lazy) lines. Cleared after consumption so a
        // following code block (e.g. after a blank line) parses normally.
        let lazy_one_line = self.pending_lazy_blockquote_close;
        self.pending_lazy_blockquote_close = false;

        self.tree.append(Item {
            start: line_start_ix,
            end: 0, // will get set later
            body: ItemBody::IndentCodeBlock(lazy_one_line),
        });
        self.tree.push();
        let bytes = self.text.as_bytes();
        let mut last_nonblank_child = None;
        let mut last_nonblank_ix = 0;
        let mut end_ix = 0;
        self.last_line_blank = false;

        let mut ix = start_ix;
        loop {
            let line_start_ix = ix;
            ix += scan_nextline(&bytes[ix..]);
            self.append_code_text(remaining_space, line_start_ix, ix);
            // TODO(spec clarification): should we synthesize newline at EOF?

            if !self.last_line_blank {
                last_nonblank_child = self.tree.cur();
                last_nonblank_ix = ix;
                end_ix = ix;
            }

            if lazy_one_line {
                break;
            }

            let mut line_start = LineStart::new(&bytes[ix..]);
            let n_containers = scan_containers(&self.tree, &mut line_start, self.options);
            if n_containers < self.tree.spine_len()
                || !(line_start.scan_space(4) || line_start.is_at_eol())
            {
                break;
            }
            let next_line_ix = ix + line_start.bytes_scanned();
            if next_line_ix == self.text.len() {
                break;
            }
            ix = next_line_ix;
            remaining_space = line_start.remaining_space();
            // Only treat the post-strip line as blank when nothing but EOL is
            // left. Lines whose whitespace survives the strip (e.g. `\t\t` →
            // `\t`) still belong to the code block; remark keeps them.
            self.last_line_blank = scan_eol(&bytes[ix..]).is_some();
        }

        // Trim trailing blank lines.
        if let Some(child) = last_nonblank_child {
            self.tree[child].next = None;
            self.tree[child].item.end = last_nonblank_ix;
        }
        self.pop(end_ix);
        // Set `interrupt`-equivalent state ONLY when this code block wasn't a
        // one-line lazy block opened immediately after a popped blockquote.
        // For `>\n\t9\n+` micromark's `interrupt` is false at `+` (the bq's
        // blank-line state propagates through the lazy code), so the empty
        // marker is allowed to open a list. For `\t9\n+` and `code\n\n2.b`,
        // the code wasn't lazy and the suppression should fire.
        if !lazy_one_line {
            self.list_interrupted_paragraph = true;
        }
        ix
    }

    fn parse_fenced_code_block(
        &mut self,
        start_ix: usize,
        indent: usize,
        fence_ch: u8,
        n_fence_char: usize,
    ) -> usize {
        // Fenced code is a full block — clear lingering empty-list-marker
        // suppression from a prior block (see `parse_hrule` comment).
        self.list_interrupted_paragraph = false;
        let bytes = self.text.as_bytes();
        let mut info_start = start_ix + n_fence_char;
        info_start += scan_space_or_tab(&bytes[info_start..]);
        // TODO: info strings are typically very short. wouldn't it be faster
        // to just do a forward scan here?
        let mut ix = info_start + scan_nextline(&bytes[info_start..]);
        // Strip only the line terminator (\n, \r, \r\n) — remark/mdast preserves
        // trailing spaces in the fence info string so they end up in `meta`.
        let info_end = ix - scan_rev_while(&bytes[info_start..ix], |b| b == b'\n' || b == b'\r');
        // The lang/meta split is taken on raw source, so a character reference
        // decoding to whitespace stays inside the language.
        let raw_info = &self.text[info_start..info_end];
        let raw_lang_end = raw_info.find([' ', '\t']).map_or(raw_info.len(), |ix| ix);
        let in_table = self.tree.is_in_table();
        let lang_len = unescape(&raw_info[..raw_lang_end], in_table).len() as u32;
        let info_string = unescape(raw_info, in_table);
        self.tree.append(Item {
            start: start_ix,
            end: 0, // will get set later
            body: ItemBody::FencedCodeBlock(self.allocs.allocate_cow(info_string), lang_len),
        });
        self.tree.push();
        loop {
            // EOF reached without a closing fence — pop and end the block.
            // arena_build.rs strips the final \n later; no extra trim here.
            if ix >= bytes.len() {
                self.pop(ix);
                return ix;
            }
            let mut line_start = LineStart::new(&bytes[ix..]);
            let n_containers = scan_containers(&self.tree, &mut line_start, self.options);
            if n_containers < self.tree.spine_len() {
                // Container outdent ends the code block. For a leaf-block
                // interrupt (paragraph/heading/fence-close), the blank line
                // immediately before the outdent acts as the container's
                // separator, so we strip one extra \n here. arena_build.rs
                // strips the last content line's \n afterward — together
                // that's 2 \n's stripped, matching `- ```\n  foo\n\noo` →
                // `foo`.
                //
                // BUT when the outdent is caused by a NEW container marker
                // on this line (sibling list-item, blockquote, etc.), remark
                // does NOT consume the preceding blank as a separator — the
                // marker itself is what terminates the previous item. So we
                // strip 0 extra here (arena_build still strips 1), keeping
                // the blank line(s) as code content.
                let next_line_ix = ix + line_start.bytes_scanned();
                let rest = &bytes[next_line_ix..];
                let line_starts_container =
                    scan_blockquote_start(rest).is_some() || scan_listitem(rest).is_some();
                let extra = if line_starts_container { 0 } else { 1 };
                if extra > 0 {
                    trim_trailing_newlines_from_code_block(&mut self.tree, bytes, extra);
                }
                self.pop(ix);
                return ix;
            }
            if self.options.contains(Options::ENABLE_MDX) {
                // MDX: check for closing fence after consuming all leading whitespace.
                // Use a separate scanner so the content line_start only strips the
                // fence indent (preserving code indentation).
                let mut close_line_start = line_start.clone();
                close_line_start.scan_all_space();
                let close_ix = ix + close_line_start.bytes_scanned();
                if let Some(n) = scan_closing_code_fence(&bytes[close_ix..], fence_ch, n_fence_char)
                {
                    ix = close_ix + n;
                    self.pop(ix);
                    return ix + scan_blank_line(&bytes[ix..]).unwrap_or(0);
                }
                // For content lines, only strip up to `indent` spaces (the fence's
                // own indentation), preserving any deeper indentation as code content.
                line_start.scan_space(indent);
            } else {
                line_start.scan_space(indent);
                let mut close_line_start = line_start.clone();
                if !close_line_start.scan_space(4 - indent) {
                    let close_ix = ix + close_line_start.bytes_scanned();
                    if let Some(n) =
                        scan_closing_code_fence(&bytes[close_ix..], fence_ch, n_fence_char)
                    {
                        ix = close_ix + n;
                        self.pop(ix);
                        // try to read trailing whitespace or it will register as a completely blank line
                        return ix + scan_blank_line(&bytes[ix..]).unwrap_or(0);
                    }
                }
            }
            let remaining_space = line_start.remaining_space();
            ix += line_start.bytes_scanned();
            let next_ix = ix + scan_nextline(&bytes[ix..]);
            self.append_code_text(remaining_space, ix, next_ix);
            ix = next_ix;
        }
    }

    fn parse_math_block(&mut self, start_ix: usize, indent: usize, n_fence_char: usize) -> usize {
        // Math fence is a full block — clear lingering empty-list-marker
        // suppression from a prior block (see `parse_hrule` comment).
        self.list_interrupted_paragraph = false;
        let bytes = self.text.as_bytes();
        let mut meta_start = start_ix + n_fence_char;
        meta_start += scan_space_or_tab(&bytes[meta_start..]);
        let mut ix = meta_start + scan_nextline(&bytes[meta_start..]);
        // Only strip the trailing newline; preserve any trailing spaces/tabs
        // in the meta string (remark keeps them verbatim).
        let meta_end = ix - scan_rev_while(&bytes[meta_start..ix], |c| c == b'\n' || c == b'\r');
        let meta_string = if meta_start < meta_end {
            unescape(&self.text[meta_start..meta_end], self.tree.is_in_table())
        } else {
            "".into()
        };
        self.tree.append(Item {
            start: start_ix,
            end: 0,
            body: ItemBody::MathBlock(self.allocs.allocate_cow(meta_string)),
        });
        self.tree.push();
        // mdast-util-math strips a leading AND trailing newline from the
        // accumulated content (matching micromark's `mathFlow`, which doesn't
        // emit the lineEnding past the last content chunk before a closing
        // fence or container exit). To replicate, peek a blank-within-container
        // line before appending it: if the NEXT line breaks the parent
        // container, the blank line's `\n` is given back to the surrounding
        // context and shouldn't enter the math value.
        loop {
            if ix >= bytes.len() {
                self.pop(ix);
                return ix;
            }
            let mut line_start = LineStart::new(&bytes[ix..]);
            let n_containers = scan_containers(&self.tree, &mut line_start, self.options);
            if n_containers < self.tree.spine_len() {
                self.pop(ix);
                return ix;
            }
            line_start.scan_space(indent);
            let mut close_line_start = line_start.clone();
            if !close_line_start.scan_space(4 - indent) {
                let close_ix = ix + close_line_start.bytes_scanned();
                if let Some(n) = scan_closing_math_fence(&bytes[close_ix..], n_fence_char) {
                    ix = close_ix + n;
                    self.pop(ix);
                    return ix + scan_blank_line(&bytes[ix..]).unwrap_or(0);
                }
            }
            let remaining_space = line_start.remaining_space();
            let content_start = ix + line_start.bytes_scanned();
            let next_ix = content_start + scan_nextline(&bytes[content_start..]);
            // Blank-within-container line lookahead: if this line's content is
            // empty (just the `\n`) and the next line breaks the parent
            // container, micromark exits the math construct without emitting
            // the lineEnding for this blank line. Skip it here so the math
            // value matches remark.
            let line_is_blank = content_start + 1 == next_ix
                && matches!(bytes[content_start], b'\n' | b'\r')
                || content_start == next_ix;
            if line_is_blank && next_ix < bytes.len() {
                let mut peek_ls = LineStart::new(&bytes[next_ix..]);
                let peek_n = scan_containers(&self.tree, &mut peek_ls, self.options);
                if peek_n < self.tree.spine_len() {
                    self.pop(ix);
                    return ix;
                }
            }
            self.append_code_text(remaining_space, content_start, next_ix);
            ix = next_ix;
        }
    }

    fn parse_metadata_block(&mut self, start_ix: usize, metadata_block_ch: u8) -> usize {
        let bytes = self.text.as_bytes();
        let metadata_block_kind = match metadata_block_ch {
            b'-' => MetadataBlockKind::YamlStyle,
            b'+' => MetadataBlockKind::PlusesStyle,
            _ => panic!("Erroneous metadata block character when parsing metadata block"),
        };
        // 3 delimiter characters
        let mut ix = start_ix + 3 + scan_nextline(&bytes[start_ix + 3..]);
        self.tree.append(Item {
            start: start_ix,
            end: 0, // will get set later
            body: ItemBody::MetadataBlock(metadata_block_kind),
        });
        self.tree.push();
        loop {
            let mut line_start = LineStart::new(&bytes[ix..]);
            let n_containers = scan_containers(&self.tree, &mut line_start, self.options);
            if n_containers < self.tree.spine_len() {
                break;
            }
            if let (_, 0) = calc_indent(&bytes[ix..], 4)
                && let Some(n) = scan_closing_metadata_block(&bytes[ix..], metadata_block_ch)
            {
                ix += n;
                break;
            }
            let remaining_space = line_start.remaining_space();
            ix += line_start.bytes_scanned();
            let next_ix = ix + scan_nextline(&bytes[ix..]);
            // Metadata blocks preserve CRLF — remark-frontmatter keeps the
            // original line endings in the yaml/toml value. (The `append_code_text`
            // path normalizes to LF, which is correct for code blocks.)
            if remaining_space > 0 {
                let cow_ix = self.allocs.allocate_cow("   "[..remaining_space].into());
                self.tree.append(Item {
                    start: ix,
                    end: ix,
                    body: ItemBody::SynthesizeText(cow_ix),
                });
            }
            self.tree.append_text(ix, next_ix, false);
            ix = next_ix;
        }

        self.pop(ix);

        // try to read trailing whitespace or it will register as a completely blank line
        ix + scan_blank_line(&bytes[ix..]).unwrap_or(0)
    }

    fn append_code_text(&mut self, remaining_space: usize, start: usize, end: usize) {
        if remaining_space > 0 {
            let cow_ix = self.allocs.allocate_cow("   "[..remaining_space].into());
            self.tree.append(Item {
                start,
                end: start,
                body: ItemBody::SynthesizeText(cow_ix),
            });
        }
        // remark preserves CRLF verbatim in code-block / html / yaml values,
        // so we don't normalize line endings here.
        self.tree.append_text(start, end, false);
    }

    /// Appends a line of HTML to the tree.
    fn append_html_line(&mut self, remaining_space: usize, start: usize, end: usize) {
        if remaining_space > 0 {
            let cow_ix = self.allocs.allocate_cow("   "[..remaining_space].into());
            self.tree.append(Item {
                start,
                end: start,
                body: ItemBody::SynthesizeText(cow_ix),
            });
        }
        // remark preserves CRLF verbatim in html blocks — emit the raw range.
        self.tree.append(Item {
            start,
            end,
            body: ItemBody::Html,
        });
    }

    /// Pop a container, setting its end.
    fn pop(&mut self, ix: usize) {
        let cur_ix = self.tree.pop().unwrap();
        self.tree[cur_ix].item.end = ix;
        if let ItemBody::DefinitionList(_) = self.tree[cur_ix].item.body {
            fixup_end_of_definition_list(&mut self.tree, cur_ix);
            self.begin_list_item = None;
        }
        if let ItemBody::List(true, _, _) | ItemBody::DefinitionList(true) =
            self.tree[cur_ix].item.body
        {
            surgerize_tight_list(&mut self.tree, cur_ix);
            self.begin_list_item = None;
        }
    }

    /// Close a list if it's open. Also set loose if last line was blank
    /// and end current list if it's a lone, empty item
    fn finish_list(&mut self, ix: usize) {
        self.finish_empty_list_item();
        if let Some(node_ix) = self.tree.peek_up()
            && let ItemBody::List(_, _, _) | ItemBody::DefinitionList(_) =
                self.tree[node_ix].item.body
        {
            self.pop(ix);
        }
        if self.last_line_blank {
            if let Some(node_ix) = self.tree.peek_grandparent()
                && let ItemBody::List(ref mut is_tight, _, _)
                | ItemBody::DefinitionList(ref mut is_tight) = self.tree[node_ix].item.body
            {
                *is_tight = false;
            }
            self.last_line_blank = false;
        }
    }

    fn mark_enclosing_listitem_spread(&mut self) {
        let spine: Vec<TreeIndex> = self.tree.walk_spine().copied().collect();
        for node_ix in spine.into_iter().rev() {
            if let ItemBody::ListItem(indent, _) = self.tree[node_ix].item.body {
                self.tree[node_ix].item.body = ItemBody::ListItem(indent, true);
                return;
            }
        }
    }

    fn finish_empty_list_item(&mut self) {
        if let Some(begin_list_item) = self.begin_list_item
            && self.last_line_blank
        {
            // A list item can begin with at most one blank line.
            if let Some(node_ix) = self.tree.peek_up()
                && let ItemBody::ListItem(_, _) | ItemBody::DefinitionListDefinition(..) =
                    self.tree[node_ix].item.body
            {
                self.pop(begin_list_item);
            }
        }
        self.begin_list_item = None;
    }

    /// Continue an existing list or start a new one if there's not an open
    /// list that matches.
    fn continue_list(&mut self, start: usize, ch: u8, index: u64) {
        self.finish_empty_list_item();
        if let Some(node_ix) = self.tree.peek_up() {
            if let ItemBody::List(ref mut is_tight, existing_ch, _) = self.tree[node_ix].item.body
                && existing_ch == ch
            {
                if self.last_line_blank {
                    *is_tight = false;
                    self.last_line_blank = false;
                }
                return;
            }
            // TODO: this is not the best choice for end; maybe get end from last list item.
            self.finish_list(start);
        }
        self.tree.append(Item {
            start,
            end: 0, // will get set later
            body: ItemBody::List(true, ch, index),
        });
        self.tree.push();
        self.last_line_blank = false;
    }

    /// Parse a thematic break.
    ///
    /// Returns index of start of next line.
    fn parse_hrule(&mut self, hrule_size: usize, ix: usize) -> usize {
        self.tree.append(Item {
            start: ix,
            end: ix + hrule_size,
            body: ItemBody::Rule,
        });
        // A thematic break is a block-terminator. The "list interrupted
        // paragraph" suppression from the previous block (e.g. indented
        // code) no longer applies — a new list after the rule can be
        // empty and still open.
        self.list_interrupted_paragraph = false;
        ix + hrule_size
    }

    /// Parse an ATX heading.
    ///
    /// Returns index of start of next line.
    fn parse_atx_heading(&mut self, start: usize, atx_level: HeadingLevel) -> usize {
        // The heading is a full block; any "interrupting paragraph"
        // suppression from a prior block (e.g. indented code) no longer
        // applies — a new empty list after the heading can still open.
        // Mirrors the equivalent clear in `parse_hrule`.
        self.list_interrupted_paragraph = false;
        let mut ix = start;
        let heading_ix = self.tree.append(Item {
            start,
            end: 0,                    // set later
            body: ItemBody::default(), // set later
        });
        ix += atx_level as usize;
        // next char is space or eol (guaranteed by scan_atx_heading)
        let bytes = self.text.as_bytes();
        if let Some(eol_bytes) = scan_eol(&bytes[ix..]) {
            self.tree[heading_ix].item.end = ix + eol_bytes;
            self.tree[heading_ix].item.body = ItemBody::Heading(atx_level, None);
            return ix + eol_bytes;
        }
        // skip leading spaces
        let skip_spaces = scan_space_or_tab(&bytes[ix..]);
        ix += skip_spaces;

        // now handle the header text
        let header_start = ix;
        let header_node_idx = self.tree.push(); // so that we can set the endpoint later

        let attr_block = if self.options.contains(Options::ENABLE_HEADING_ATTRIBUTES) {
            let header_end = header_start + scan_nextline(&bytes[header_start..]);
            self.extract_and_parse_heading_attribute_block(header_start, header_end)
                .map(|block| (block, header_end))
        } else {
            None
        };
        let (end, content_end, attrs) = if let Some((block, header_end)) = attr_block {
            self.parse_line(ix, Some(block.block_open), TableParseMode::Disabled, ix);
            // Reparse the trailing directive run separately so it keeps its own `{...}`.
            let content_end = match block.tail_start {
                Some(tail_start) => {
                    self.parse_line(
                        tail_start,
                        Some(header_end),
                        TableParseMode::Disabled,
                        tail_start,
                    );
                    header_end
                }
                None => block.block_open,
            };
            (header_end, content_end, Some(block.attrs))
        } else {
            // ATX headings can't span lines. In MDX mode bound the inline
            // scan to the current line so an unclosed expression body like
            // `# {1 +\n2}` errors at end-of-line (matching mdx-js) instead
            // of consuming across the newline to find a closing `}`.
            let line_end = if self.options.contains(Options::ENABLE_MDX) {
                Some(header_start + scan_nextline(&bytes[header_start..]))
            } else {
                None
            };
            let (line_ix, line_brk) = self.parse_line(ix, line_end, TableParseMode::Disabled, ix);
            ix = line_ix;
            // Backslash at end is actually hard line break
            if let Some(Item {
                start,
                end,
                body: ItemBody::HardBreak(true),
            }) = line_brk
            {
                self.tree.append_text(start, end, false);
            }
            (ix, ix, None)
        };
        self.tree[header_node_idx].item.end = end;

        // remove trailing matter from header text
        let mut empty_text_node = false;
        if let Some(cur_ix) = self.tree.cur() {
            // remove closing of the ATX heading
            let header_text = &bytes[header_start..content_end];
            let mut limit = header_text
                .iter()
                .rposition(|&b| !(b == b'\n' || b == b'\r' || b == b' ' || b == b'\t'))
                .map_or(0, |i| i + 1);
            let closer = header_text[..limit]
                .iter()
                .rposition(|&b| b != b'#')
                .map_or(0, |i| i + 1);
            if closer == 0 {
                limit = closer;
            } else {
                let spaces = scan_rev_while(&header_text[..closer], |b| b == b' ' || b == b'\t');
                if spaces > 0 {
                    limit = closer - spaces;
                }
            }
            // if text is only spaces, then remove them
            self.tree[cur_ix].item.end = limit + header_start;

            // limit = 0 when text is empty after removing spaces
            if limit == 0 {
                empty_text_node = true;
            }
        }

        if empty_text_node {
            self.tree.remove_node();
        } else {
            self.tree.pop();
        }
        self.tree[heading_ix].item.body = ItemBody::Heading(
            atx_level,
            attrs.map(|attrs| self.allocs.allocate_heading(attrs)),
        );

        end
    }

    /// Returns the number of bytes scanned on success.
    fn parse_footnote(&mut self, start: usize) -> Option<usize> {
        let bytes = &self.text.as_bytes()[start..];
        if !bytes.starts_with(b"[^") {
            return None;
        }
        // GFM footnote labels can't be empty or contain whitespace.
        // `[^a b]` and `[^]` fall through to the regular refdef path.
        let (mut i, label) =
            scan_link_label_rest(&self.text[start + 2..], &|_| None, self.tree.is_in_table())?;
        // GFM/micromark's labelInside rejects whitespace CHARACTER-BY-CHARACTER.
        // Any space/tab/eol — including leading whitespace stripped by
        // scan_link_label_rest's trim — invalidates the footnote definition
        // (it falls through to a regular reference definition). Inspect the
        // raw source bytes between `[^` and `]`, not the trimmed `label`.
        let raw_label = &self.text.as_bytes()[start + 2..start + 2 + i.saturating_sub(1)];
        if raw_label.is_empty()
            || raw_label
                .iter()
                .any(|&b| b == b' ' || b == b'\t' || b == b'\r' || b == b'\n')
        {
            return None;
        }
        i += 2;
        if bytes.get(i) != Some(&b':') {
            return None;
        }
        i += 1;
        self.finish_list(start);
        if let Some(node_ix) = self.tree.peek_up()
            && let ItemBody::FootnoteDefinition(..) = self.tree[node_ix].item.body
        {
            // finish previous footnote if it's still open
            self.pop(start);
        }
        i += scan_space_or_tab(&bytes[i..]);
        self.allocs
            .footdefs
            .0
            .insert(UniCase::new(label.clone()), FootnoteDef { use_count: 0 });
        self.tree.append(Item {
            start,
            end: 0, // will get set later
            // TODO: check whether the label here is strictly necessary
            body: ItemBody::FootnoteDefinition(self.allocs.allocate_cow(label)),
        });
        self.tree.push();
        Some(i)
    }

    /// Tries to parse a reference label, which can be interrupted by new blocks.
    /// On success, returns the number of bytes of the label and the label itself.
    fn parse_refdef_label(&self, start: usize) -> Option<(usize, CowStr<'a>)> {
        scan_link_label_rest(
            &self.text[start..],
            &|bytes| {
                let mut line_start = LineStart::new(bytes);
                let tree_position = scan_containers(&self.tree, &mut line_start, self.options);
                let current_container = tree_position == self.tree.spine_len();
                if line_start.scan_space(4) {
                    return Some(line_start.bytes_scanned());
                }
                let bytes_scanned = line_start.bytes_scanned();
                let suffix = &bytes[bytes_scanned..];
                if self.scan_paragraph_interrupt(suffix, current_container, tree_position)
                    || (current_container && scan_setext_heading(suffix).is_some())
                {
                    None
                } else {
                    Some(bytes_scanned)
                }
            },
            self.tree.is_in_table(),
        )
    }

    /// Returns number of bytes scanned, label and definition on success.
    fn parse_refdef_total(&mut self, start: usize) -> Option<(usize, LinkLabel<'a>, LinkDef<'a>)> {
        let bytes = &self.text.as_bytes()[start..];
        if bytes.first() != Some(&b'[') {
            return None;
        }
        let (mut i, label) = self.parse_refdef_label(start + 1)?;
        i += 1;
        if bytes.get(i) != Some(&b':') {
            return None;
        }
        i += 1;
        let (bytecount, link_def) = self.scan_refdef(start, start + i)?;
        Some((bytecount + i, UniCase::new(label), link_def))
    }

    /// Returns number of bytes and number of newlines
    fn scan_refdef_space(&self, bytes: &[u8], mut i: usize) -> Option<(usize, usize)> {
        let mut newlines = 0;
        loop {
            let whitespaces = scan_space_or_tab(&bytes[i..]);
            i += whitespaces;
            if let Some(eol_bytes) = scan_eol(&bytes[i..]) {
                i += eol_bytes;
                newlines += 1;
                if newlines > 1 {
                    return None;
                }
            } else {
                break;
            }
            let mut line_start = LineStart::new(&bytes[i..]);
            let tree_position = scan_containers(&self.tree, &mut line_start, self.options);
            let current_container = tree_position == self.tree.spine_len();
            if !line_start.scan_space(4) {
                let suffix = &bytes[i + line_start.bytes_scanned()..];
                if self.scan_paragraph_interrupt(suffix, current_container, tree_position)
                    || scan_setext_heading(suffix).is_some()
                {
                    return None;
                }
            }
            i += line_start.bytes_scanned();
        }
        Some((i, newlines))
    }

    // returns (bytelength, title_str)
    fn scan_refdef_title<'t>(&self, text: &'t str) -> Option<(usize, CowStr<'t>)> {
        let bytes = text.as_bytes();
        // Unlike CommonMark, remark keeps an unescaped `(` inside a paren title as content.
        let closing_delim = match bytes.first()? {
            b'\'' => b'\'',
            b'"' => b'"',
            b'(' => b')',
            _ => return None,
        };
        let mut bytecount = 1;
        let mut linestart = 1;

        let mut linebuf = None;

        while let Some(&c) = bytes.get(bytecount) {
            match c {
                b'\n' | b'\r' => {
                    // push text to line buffer
                    // this is used to strip the block formatting:
                    //
                    // > [first]: http://example.com "
                    // > second"
                    //
                    // should get turned into `<a href="example.com" title=" second">first</a>`
                    let linebuf = if let Some(linebuf) = &mut linebuf {
                        linebuf
                    } else {
                        linebuf = Some(String::new());
                        linebuf.as_mut().unwrap()
                    };
                    linebuf.push_str(&text[linestart..bytecount]);
                    // The title's line endings are content, kept byte for byte.
                    linebuf.push(c as char);
                    bytecount += 1;
                    if c == b'\r' && bytes.get(bytecount) == Some(&b'\n') {
                        linebuf.push('\n');
                        bytecount += 1;
                    }
                    let mut line_start = LineStart::new(&bytes[bytecount..]);
                    let tree_position = scan_containers(&self.tree, &mut line_start, self.options);
                    let current_container = tree_position == self.tree.spine_len();
                    if !line_start.scan_space(4) {
                        let suffix = &bytes[bytecount + line_start.bytes_scanned()..];
                        if self.scan_paragraph_interrupt(suffix, current_container, tree_position)
                            || scan_setext_heading(suffix).is_some()
                        {
                            return None;
                        }
                    }
                    line_start.scan_all_space();
                    bytecount += line_start.bytes_scanned();
                    linestart = bytecount;
                    if scan_blank_line(&bytes[bytecount..]).is_some() {
                        // blank line - not allowed
                        return None;
                    }
                }
                b'\\' => {
                    bytecount += 1;
                    if let Some(c) = bytes.get(bytecount)
                        && c != &b'\r'
                        && c != &b'\n'
                    {
                        bytecount += 1;
                    }
                }
                c if c == closing_delim => {
                    let cow = if let Some(mut linebuf) = linebuf {
                        linebuf.push_str(&text[linestart..bytecount]);
                        CowStr::from(linebuf)
                    } else {
                        CowStr::from(&text[linestart..bytecount])
                    };
                    return Some((bytecount + 1, cow));
                }
                _ => {
                    bytecount += 1;
                }
            }
        }
        None
    }

    /// Returns # of bytes and definition.
    /// Assumes the label of the reference including colon has already been scanned.
    fn scan_refdef(&self, span_start: usize, start: usize) -> Option<(usize, LinkDef<'a>)> {
        let bytes = self.text.as_bytes();

        // whitespace between label and url (including up to one newline)
        let (mut i, _newlines) = self.scan_refdef_space(bytes, start)?;

        // scan link dest
        let (dest_length, dest) = scan_link_dest(self.text, i, LINK_MAX_NESTED_PARENS)?;
        if dest_length == 0 {
            return None;
        }
        let dest = unescape(dest, self.tree.is_in_table());
        i += dest_length;

        // remark folds trailing same-line whitespace after the URL into the
        // definition's source span even with no title. Extend the no-title
        // fallback to match.
        let span_end = i + scan_space_or_tab(&bytes[i..]);

        // no title
        let mut backup = (
            span_end - start,
            LinkDef {
                dest,
                title: None,
                span: span_start..span_end,
            },
        );

        // scan whitespace between dest and label
        let (mut i, newlines) =
            if let Some((new_i, mut newlines)) = self.scan_refdef_space(bytes, i) {
                if i == self.text.len() {
                    newlines += 1;
                }
                if new_i == i && newlines == 0 {
                    return None;
                }
                if newlines > 1 {
                    return Some(backup);
                };
                (new_i, newlines)
            } else {
                return Some(backup);
            };

        // scan title
        // if this fails but newline == 1, return also a refdef without title
        if let Some((title_length, title)) = self.scan_refdef_title(&self.text[i..]) {
            i += title_length;
            if scan_blank_line(&bytes[i..]).is_some() {
                // remark folds trailing same-line whitespace after the
                // title into the definition's source span (matches the
                // behavior already applied above in the no-title path).
                let span_end = i + scan_space_or_tab(&bytes[i..]);
                backup.0 = i - start;
                backup.1.span = span_start..span_end;
                backup.1.title = Some(unescape(title, self.tree.is_in_table()));
                return Some(backup);
            }
        }
        if newlines > 0 { Some(backup) } else { None }
    }

    /// Checks whether we should break a paragraph on the given input.
    fn scan_paragraph_interrupt(
        &self,
        bytes: &[u8],
        current_container: bool,
        tree_position: usize,
    ) -> bool {
        // The free `scan_paragraph_interrupt_no_table` helper checks MDX JSX
        // flow elements with a `None` container check, which misfires inside
        // a blockquote or similar container where the `>` continuation prefix
        // would be mis-read as a JSX close `>`. Re-check that specific case
        // with the spine's container prefix so multi-line JSX in blockquotes
        // correctly interrupts paragraphs.
        #[cfg(feature = "mdx")]
        if self.options.contains(Options::ENABLE_MDX)
            && bytes.starts_with(b"<")
            && self
                .scan_mdx_flow_in_container_bytes(bytes, |b, c| scan_mdx_jsx_block(b, c))
                .is_some()
        {
            return true;
        }
        if scan_paragraph_interrupt_no_table(
            bytes,
            current_container,
            self.options.contains(Options::ENABLE_FOOTNOTES),
            self.options.contains(Options::ENABLE_DEFINITION_LIST),
            self.options.contains(Options::ENABLE_MDX),
            self.options.contains(Options::ENABLE_MATH_MULTI_DOLLAR),
            self.options.contains(Options::ENABLE_DIRECTIVE),
            &self.tree,
            tree_position,
        ) {
            return true;
        }
        // pulldown-cmark traditionally only interrupted paragraphs on "heavy"
        // table headers (lines starting with `|`). remark-gfm also lets a
        // "light" table interrupt: any header line followed by a valid
        // delimiter row, provided the header isn't a lazy-continuation line
        // (lazy lines can't open new blocks).
        //
        // ```markdown
        // This is a table
        // | a | b | c |
        // |---|---|---|
        // | d | e | f |
        //
        // Also a table
        //  a | b | c
        // ---|---|---
        // ```
        if !self.options.contains(Options::ENABLE_TABLES) {
            return false;
        }
        if !bytes.starts_with(b"|") && !current_container {
            return false;
        }

        // Cheap pre-filter: a delimiter row must contain at least one `-`.
        // Container paragraphs hit this path on every continuation line, so
        // skipping `scan_table_head` when the next line obviously can't be a
        // delimiter row matters for parse perf.
        // Use SIMD-backed `memchr` rather than a `position` closure — this
        // path runs on every paragraph continuation line.
        let Some(eol_off) = memchr::memchr2(b'\n', b'\r', bytes) else {
            return false;
        };
        let next_line_ix = eol_off + scan_eol(&bytes[eol_off..]).unwrap();
        // One pass: a `-` reached before any line ending is a `-` on the next line.
        if memchr::memchr3(b'-', b'\n', b'\r', &bytes[next_line_ix..])
            .is_none_or(|p| bytes[next_line_ix + p] != b'-')
        {
            return false;
        }

        // Scan the table head. The part that looks like:
        //
        //     |---|---|---|
        //
        // Also scan any containing items, since it's on its own line, and
        // might be nested inside a block quote or something
        //
        //     > Table: First
        //     > | first col | second col |
        //     > |-----------|------------|
        //     ^
        //     | need to skip over the `>` when checking for the table
        let mut line_start = LineStart::new(&bytes[next_line_ix..]);
        if scan_containers(&self.tree, &mut line_start, self.options) != self.tree.spine_len() {
            return false;
        }
        // In MDX, a table nested inside a list item may be indented beyond the
        // list's continuation column (no indented code blocks to conflict
        // with). Consume any extra whitespace so scan_table_head sees the
        // delimiter row flush-left.
        if self.options.contains(Options::ENABLE_MDX) {
            line_start.scan_all_space();
        }
        let table_head_ix = next_line_ix + line_start.bytes_scanned();
        let (table_head_bytes, alignment) = scan_table_head(&bytes[table_head_ix..]);

        if table_head_bytes == 0 {
            return false;
        }

        // First line, count unescaped pipes. A run of consecutive backslashes
        // toggles the escape state: `\|` escapes the pipe, `\\|` is a literal
        // `\` followed by an unescaped separator pipe.
        let mut pipes = 0;
        let mut bsesc = false;
        let mut last_pipe_ix = 0;
        for (i, &byte) in bytes[..eol_off].iter().enumerate() {
            match byte {
                b'\\' => {
                    bsesc = !bsesc;
                    continue;
                }
                b'|' if !bsesc => {
                    pipes += 1;
                    last_pipe_ix = i;
                }
                _ => {}
            }
            bsesc = false;
        }

        // computing header count from number of pipes
        let header_count = count_header_cols(bytes, pipes, 0, last_pipe_ix);

        // make sure they match the number of columns we find in separator line
        alignment.len() == header_count
    }

    /// Extracts and parses a heading's trailing attribute block, with its span.
    fn extract_and_parse_heading_attribute_block(
        &mut self,
        header_start: usize,
        header_end: usize,
    ) -> Option<HeadingAttrBlock<'a>> {
        if !self.options.contains(Options::ENABLE_HEADING_ATTRIBUTES) {
            return None;
        }

        let bytes = self.text.as_bytes();
        let search_end = if self.options.contains(Options::ENABLE_DIRECTIVE) {
            self.heading_attr_block_search_end(header_start, header_end)
        } else {
            header_end
        };

        let header_bytes = &bytes[header_start..search_end];
        let (content_len, attr_block_range_rel) =
            extract_attribute_block_content_from_header_text(header_bytes);
        let range = attr_block_range_rel?;

        // Claim as attributes only when the body isn't valid JS; otherwise MDX
        // treats `{...}` as an expression. try_parse_expression_body returns None
        // when it parses, so `?` bails and leaves real expressions alone.
        #[cfg(feature = "mdx")]
        if self.options.contains(Options::ENABLE_MDX) {
            let inner_start = header_start + range.start;
            let inner_end = header_start + range.end;
            crate::mdx::try_parse_expression_body(
                &self.text[inner_start..inner_end],
                &mut self.mdx_expr_allocator,
            )?;
        }

        let attrs = parse_inside_attribute_block(
            &self.text[(header_start + range.start)..(header_start + range.end)],
        )?;

        let block_open = header_start + content_len;
        let block_close = header_start + range.end + 1;
        let tail_has_content = bytes[block_close..header_end]
            .iter()
            .any(|&b| !matches!(b, b' ' | b'\t' | b'\n' | b'\r'));
        Some(HeadingAttrBlock {
            attrs,
            block_open,
            tail_start: if tail_has_content {
                Some(block_close)
            } else {
                None
            },
        })
    }

    /// The offset up to which a heading's trailing attribute block is searched.
    /// A trailing run of text directives (each `:name[label]{attrs}`, separated
    /// only by whitespace) is heading content and owns any `{...}` it contains,
    /// so the search stops before that run.
    fn heading_attr_block_search_end(&self, header_start: usize, header_end: usize) -> usize {
        let bytes = self.text.as_bytes();
        let mut spans: Vec<(usize, usize)> = Vec::new();
        let mut i = header_start;
        while i < header_end {
            if bytes[i] == b':'
                && let Some(end) = scan_heading_text_directive(self.text, bytes, i, header_end)
            {
                spans.push((i, end));
                i = end;
                continue;
            }
            i += 1;
        }

        let mut pos = header_end;
        pos -= scan_rev_while(&bytes[header_start..pos], |b| {
            matches!(b, b' ' | b'\t' | b'\n' | b'\r')
        });
        for &(start, end) in spans.iter().rev() {
            if end != pos {
                break;
            }
            pos = start;
            pos -= scan_rev_while(&bytes[header_start..pos], |b| {
                matches!(b, b' ' | b'\t' | b'\n' | b'\r')
            });
        }
        pos
    }
}

/// Scanning modes for `Parser`'s `parse_line` method.
#[derive(PartialEq, Eq, Copy, Clone)]
enum TableParseMode {
    /// Inside a paragraph, scanning for table headers.
    Scan,
    /// Inside a table.
    Active,
    /// Inside a paragraph, not scanning for table headers.
    Disabled,
}

/// Computes the number of header columns in a table line by computing the number of dividing pipes
/// that aren't followed or preceded by whitespace.
fn count_header_cols(
    bytes: &[u8],
    mut pipes: usize,
    mut start: usize,
    last_pipe_ix: usize,
) -> usize {
    // A header with no pipes is a single-column "light" table row.
    if pipes == 0 {
        return 1;
    }
    // was first pipe preceded by whitespace? if so, subtract one
    start += scan_space_or_tab(&bytes[start..]);
    if bytes[start] == b'|' {
        pipes -= 1;
    }

    // was last pipe followed by whitespace? if so, sub one
    if scan_blank_line(&bytes[(last_pipe_ix + 1)..]).is_some() {
        pipes
    } else {
        pipes + 1
    }
}

/// Checks whether we should break a paragraph on the given input.
///
/// Use `FirstPass::scan_paragraph_interrupt` in any context that allows
/// tables to interrupt the paragraph.
/// Whether an MDX JSX flow element or `{…}` flow expression at the line start
/// interrupts a paragraph. In the lite (non-mdx) build this is always false —
/// the gated `mdx::*` scanners don't exist there.
#[cfg(feature = "mdx")]
fn mdx_block_interrupts(bytes: &[u8], mdx: bool) -> bool {
    (mdx && bytes.starts_with(b"<") && scan_mdx_jsx_block(bytes, None).is_some())
        || (mdx && bytes.starts_with(b"{") && scan_mdx_expression_block(bytes, None).is_some())
}

#[cfg(not(feature = "mdx"))]
fn mdx_block_interrupts(_bytes: &[u8], _mdx: bool) -> bool {
    false
}

#[allow(clippy::too_many_arguments)]
fn scan_paragraph_interrupt_no_table(
    bytes: &[u8],
    current_container: bool,
    has_footnote: bool,
    definition_list: bool,
    mdx: bool,
    math: bool,
    directive: bool,
    tree: &Tree<Item>,
    tree_position: usize,
) -> bool {
    scan_eol(bytes).is_some()
        || scan_hrule(bytes).is_ok()
        || scan_atx_heading(bytes).is_some()
        || scan_code_fence(bytes).is_some()
        || (math && scan_math_fence(bytes).is_some())
        || (directive && scan_interrupting_container_extensions_fence(bytes))
        || scan_blockquote_start(bytes).is_some()
        || scan_listitem(bytes).is_some_and(|(ix, delim, _index, _)| {
            ! current_container ||
            tree.is_in_table() ||
            // We don't allow interruption by either empty lists or numbered
            // lists whose marker isn't *textually* the single character `1`
            // (so `01.`, `001.`, `10.` all stay paragraph text — micromark
            // matches the literal marker, not the parsed integer value).
            (delim == b'*' || delim == b'-' || delim == b'+'
                || (bytes.first() == Some(&b'1')
                    && matches!(bytes.get(1), Some(b'.') | Some(b')'))))
                && (scan_blank_line(&bytes[ix..]).is_none())
        })
        // HTML types 1–6 interrupt paragraphs. MDX disables HTML blocks
        // entirely (everything `<…>` is JSX), so the type-1/6 gate
        // shouldn't fire — a line like `</div>after` is just paragraph
        // text in MDX, not a new HTML block.
        || (!mdx
            && bytes.starts_with(b"<")
            && (get_html_end_tag(&bytes[1..]).is_some() || starts_html_block_type_6(&bytes[1..])))
        // Type-7 HTML blocks normally can't interrupt a paragraph, but
        // micromark/remark close the paragraph when a type-7 start appears
        // on a *lazy-continuation* line of a container (typically a
        // blockquote without `>` on the next line). Mirror that quirk:
        // accept type-7 only when we're not inside the current container,
        // so plain-paragraph cases (`oo\n<a href="bar">`) keep the inline
        // HTML behavior the spec requires. MDX disables HTML blocks
        // entirely (everything `<...>` is JSX), so don't trigger the
        // quirk in MDX mode either.
        || (!mdx
            && !current_container
            && bytes.starts_with(b"<")
            && scan_html_type_7(bytes).is_some())
        // MDX JSX flow elements and `{ ... }` flow expressions also interrupt
        // paragraphs. Behind a cfg-switched helper so the gated `mdx::*` scan
        // functions aren't named in the lite build (where it returns false).
        || mdx_block_interrupts(bytes, mdx)
        || definition_list
            && ((current_container
                && tree.peek_up().is_some_and(|cur| {
                    matches!(
                        tree[cur].item.body,
                        ItemBody::Paragraph
                            | ItemBody::TightParagraph
                            | ItemBody::MaybeDefinitionListTitle
                    )
                }))
                || tree.walk_spine().nth(tree_position).is_some_and(|cur| {
                    matches!(tree[*cur].item.body, ItemBody::DefinitionListDefinition(..))
                }))
            && is_definition_list_marker(bytes)
        || (has_footnote
            && bytes.starts_with(b"[^")
            && scan_link_label_rest(
                core::str::from_utf8(&bytes[2..]).unwrap(),
                &|_| None,
                tree.is_in_table(),
            )
            .is_some_and(|(len, _)| bytes.get(2 + len) == Some(&b':')))
}

/// Return `true` if `pos` sits between two backtick runs of matching length
/// within the current paragraph. Used to give code spans priority over MDX
/// inline expressions — `` `{foo}` `` is a code span containing the text
/// `{foo}`, not an inline expression. Code spans can cross newlines, so we
/// scan the full paragraph (bounded by blank lines AND by changes in the
/// leading container prefix — `paragraph + blockquote` are different blocks
/// and their backticks must not pair).
/// Same shape as `is_inside_code_span` but for math `$…$` runs. Used to
/// keep GFM literal-autolink emission from cutting across a math span
/// boundary (e.g. `$<https://x$` — the `$` pair owns those bytes).
/// Narrow `[para_start, para_end)` so it doesn't span a display-math fence
/// line (`$$` / `$$ info`). Such a line is a block boundary the parser splits
/// on, so inline `$` runs must not pair across it. Without this, the trailing
/// `$$` in `a$$\n\frac{1}{2}\n$$` pairs with the inline `a$$` and the `{` is
/// misread as math text instead of an MDX expression. The fence line holding
/// `pos` is never itself a fence (it holds the `{`), so it's left intact.
fn clamp_scope_to_math_fences(
    bytes: &[u8],
    pos: usize,
    para_start: usize,
    para_end: usize,
) -> (usize, usize) {
    let mut lo = para_start;
    let mut hi = para_end;
    let mut line_start = para_start;
    while line_start < para_end {
        let mut line_end = line_start;
        while line_end < para_end && !matches!(bytes[line_end], b'\n' | b'\r') {
            line_end += 1;
        }
        let mut content = line_start;
        while content < line_end && bytes[content] == b' ' && content - line_start < 3 {
            content += 1;
        }
        let mut next_line = line_end;
        if bytes.get(next_line) == Some(&b'\r') {
            next_line += 1;
        }
        if bytes.get(next_line) == Some(&b'\n') {
            next_line += 1;
        }
        if scan_math_fence(&bytes[content..line_end]).is_some() {
            if line_start > pos {
                hi = line_start;
                break;
            }
            if line_end <= pos {
                lo = next_line;
            }
        }
        line_start = next_line;
    }
    (lo, hi)
}

fn is_inside_math_span(bytes: &[u8], pos: usize) -> bool {
    let (para_start, para_end) = scope_for_inline(bytes, pos);
    // Fast reject: need at least one `$` on each side of `pos` to form a
    // span. Skips the per-byte walk when the paragraph has no `$`.
    if pos <= para_start
        || pos >= para_end
        || memchr::memchr(b'$', &bytes[para_start..pos]).is_none()
        || memchr::memchr(b'$', &bytes[pos..para_end]).is_none()
    {
        return false;
    }
    // Only after the cheap reject: keep `$` runs from pairing across a `$$`
    // display-math fence line (a block boundary).
    let (para_start, para_end) = clamp_scope_to_math_fences(bytes, pos, para_start, para_end);
    // Collect `$` runs as `(start, len, escaped)`. A run preceded by an odd
    // number of backslashes is escaped (`\$`): matching the `MaybeMath`
    // resolver, the backslash only prevents *opening* a span; an escaped `$`
    // can still *close* one. So `$x\$y$` is a single span `x\` that closes at
    // the escaped `$`, and a `{` between the dollars is math text, not an
    // expression.
    let mut runs: Vec<(usize, usize, bool)> = Vec::with_capacity(4);
    let mut i = para_start;
    while i < para_end {
        if bytes[i] == b'$' {
            let start = i;
            while i < para_end && bytes[i] == b'$' {
                i += 1;
            }
            let mut backslashes = 0usize;
            let mut k = start;
            while k > para_start && bytes[k - 1] == b'\\' {
                backslashes += 1;
                k -= 1;
            }
            runs.push((start, i - start, backslashes % 2 == 1));
        } else {
            i += 1;
        }
    }
    if runs.len() < 2 {
        return false;
    }
    let mut paired = vec![false; runs.len()];
    for a in 0..runs.len() {
        // An escaped `$` run can't open a span (only close one).
        if paired[a] || runs[a].2 {
            continue;
        }
        for b in (a + 1)..runs.len() {
            if paired[b] {
                continue;
            }
            if runs[b].1 == runs[a].1 {
                paired[a] = true;
                paired[b] = true;
                let span_start = runs[a].0 + runs[a].1;
                let span_end = runs[b].0;
                if pos >= span_start && pos < span_end {
                    return true;
                }
                break;
            }
        }
    }
    false
}

fn is_inside_code_span(bytes: &[u8], pos: usize) -> bool {
    let (para_start, para_end) = scope_for_inline(bytes, pos);
    // Fast reject: a code span needs at least two backtick runs around
    // `pos`. If there's no backtick on either side, we're not in one.
    // memchr is SIMD-accelerated and a lot cheaper than the bracket-aware
    // walk below.
    if pos <= para_start
        || pos >= para_end
        || memchr::memchr(b'`', &bytes[para_start..pos]).is_none()
        || memchr::memchr(b'`', &bytes[pos..para_end]).is_none()
    {
        return false;
    }
    // Pre-scan for matched `[`/`]` pairs. Only a bracket that actually closes
    // scopes code-span pairing; an *unclosed* `[` (as in the code span
    // `` `a[b` ``) must not shift the depth, since code spans bind tighter
    // than links and a stray `[` is just content.
    let mut bracket_matched = vec![false; para_end - para_start];
    {
        let mut stack: Vec<usize> = Vec::new();
        let mut k = para_start;
        while k < para_end {
            if bytes[k] == b'\\' && k + 1 < para_end {
                k += 2;
                continue;
            }
            match bytes[k] {
                b'[' => stack.push(k),
                b']' => {
                    if let Some(open) = stack.pop() {
                        bracket_matched[open - para_start] = true;
                        bracket_matched[k - para_start] = true;
                    }
                }
                _ => {}
            }
            k += 1;
        }
    }
    // Collect backtick runs in the paragraph, skipping backslash-escaped
    // ones. Each run records its `[`-bracket nesting depth at the opening
    // byte so that backticks inside a `[label]` (e.g. a directive label
    // or a link's body) only pair with backticks at the same depth — not
    // with backticks in the surrounding paragraph text.
    let mut runs: Vec<(usize, usize, i32)> = Vec::with_capacity(8);
    let mut bracket_depth: i32 = 0;
    let mut i = para_start;
    while i < para_end {
        if bytes[i] == b'\\' && i + 1 < para_end {
            i += 2;
            continue;
        }
        match bytes[i] {
            b'[' if bracket_matched[i - para_start] => {
                bracket_depth += 1;
                i += 1;
            }
            b']' if bracket_depth > 0 && bracket_matched[i - para_start] => {
                bracket_depth -= 1;
                i += 1;
            }
            b'`' => {
                let start = i;
                while i < para_end && bytes[i] == b'`' {
                    i += 1;
                }
                runs.push((start, i - start, bracket_depth));
            }
            _ => i += 1,
        }
    }
    if runs.len() < 2 {
        return false;
    }
    // First-fit pair the runs: each open matches the next run of the same
    // length AND same bracket depth.
    let mut paired = vec![false; runs.len()];
    for a in 0..runs.len() {
        if paired[a] {
            continue;
        }
        for b in (a + 1)..runs.len() {
            if paired[b] {
                continue;
            }
            if runs[b].1 == runs[a].1 && runs[b].2 == runs[a].2 {
                paired[a] = true;
                paired[b] = true;
                let span_start = runs[a].0 + runs[a].1;
                let span_end = runs[b].0;
                if pos >= span_start && pos < span_end {
                    return true;
                }
                break;
            }
        }
    }
    false
}

/// Bound the inline-content scope around `pos` for code-span pairing.
/// Walks backward and forward line-by-line, stopping at blank lines AND at
/// any line whose container prefix differs from the current line's. This
/// prevents a backtick in a root paragraph from pairing with one inside a
/// blockquote (or vice versa) and incorrectly suppressing an MDX expression
/// scan in between.
fn scope_for_inline(bytes: &[u8], pos: usize) -> (usize, usize) {
    let cur_line_start = {
        let mut j = pos;
        while j > 0 && !matches!(bytes[j - 1], b'\n' | b'\r') {
            j -= 1;
        }
        j
    };
    let cur_prefix = leading_container_prefix(&bytes[cur_line_start..]);
    let start = scope_start_with_prefix(bytes, cur_line_start, cur_prefix);
    let end = scope_end_with_prefix(bytes, cur_line_start, cur_prefix);
    (start, end)
}

/// Count the blockquote (`>`) depth at the start of a line. Two lines belong
/// to the same inline-content scope only if their blockquote depths match.
/// Plain leading whitespace on a continuation line within a paragraph is not
/// a container boundary — only `>` markers are.
fn leading_container_prefix(line: &[u8]) -> usize {
    let mut depth = 0;
    let mut i = 0;
    loop {
        // Up to 3 spaces of indentation before each `>`.
        let mut spaces = 0;
        while i < line.len() && line[i] == b' ' && spaces < 3 {
            i += 1;
            spaces += 1;
        }
        if i < line.len() && line[i] == b'>' {
            depth += 1;
            i += 1;
            // Optional single space/tab after `>`.
            if i < line.len() && (line[i] == b' ' || line[i] == b'\t') {
                i += 1;
            }
        } else {
            break;
        }
    }
    depth
}

/// True when the line at `line_start` opens a list item (bullet or ordered
/// marker after ≤3 spaces of indent). Such a line begins a new leaf paragraph,
/// so it bounds an inline scope: `is_inside_code_span` / math must not scan
/// across it into a sibling item. Without this a tight list of N items
/// collapses into one scope and each per-item scan runs over the whole list.
fn line_opens_list_item(bytes: &[u8], line_start: usize) -> bool {
    let mut i = line_start;
    while i < bytes.len() && bytes[i] == b' ' && i - line_start < 3 {
        i += 1;
    }
    scan_listitem(&bytes[i..]).is_some()
}

fn scope_start_with_prefix(bytes: &[u8], cur_line_start: usize, prefix: usize) -> usize {
    if line_opens_list_item(bytes, cur_line_start) {
        return cur_line_start;
    }
    let mut line_start = cur_line_start;
    while line_start > 0 {
        let mut prev_end = line_start - 1;
        if prev_end > 0 && bytes[prev_end] == b'\n' && bytes[prev_end - 1] == b'\r' {
            prev_end -= 1;
        }
        let mut prev_start = prev_end;
        while prev_start > 0 && !matches!(bytes[prev_start - 1], b'\n' | b'\r') {
            prev_start -= 1;
        }
        let prev_line = &bytes[prev_start..prev_end];
        if prev_line.iter().all(|&b| b == b' ' || b == b'\t') {
            return line_start;
        }
        if leading_container_prefix(prev_line) != prefix {
            return line_start;
        }
        line_start = prev_start;
        // The item's first line is the paragraph start: include it, stop here.
        if line_opens_list_item(bytes, line_start) {
            return line_start;
        }
    }
    line_start
}

fn scope_end_with_prefix(bytes: &[u8], cur_line_start: usize, prefix: usize) -> usize {
    let mut i = cur_line_start;
    while i < bytes.len() && !matches!(bytes[i], b'\n' | b'\r') {
        i += 1;
    }
    while i < bytes.len() {
        let after_eol = if bytes[i] == b'\r' && bytes.get(i + 1) == Some(&b'\n') {
            i + 2
        } else {
            i + 1
        };
        let mut next_eol = after_eol;
        while next_eol < bytes.len() && !matches!(bytes[next_eol], b'\n' | b'\r') {
            next_eol += 1;
        }
        let next_line = &bytes[after_eol..next_eol];
        if next_line.iter().all(|&b| b == b' ' || b == b'\t') {
            return i;
        }
        if leading_container_prefix(next_line) != prefix {
            return i;
        }
        // A line opening a new list item starts a new paragraph: stop before it.
        if line_opens_list_item(bytes, after_eol) {
            return i;
        }
        i = next_eol;
    }
    i
}

/// Deeper label starts are taken as links, which only ever withholds a tail.
#[cfg(feature = "mdx")]
const LINK_LABEL_MAX_TRACKED_DEPTH: u32 = 64;

/// True if the byte at `pos` sits inside a CommonMark link tail `[...](...)`,
/// that is, some `](` earlier on the line opens a tail that parses as
/// destination, optional title, `)` and encloses `pos`. mdx-js does not
/// evaluate expressions in link URLs (URL `{x}` is literal text, URL-encoded
/// at render time), so we treat `{` here as plain text rather than an
/// expression start. This both avoids a hard error on unmatched `{` inside a
/// valid URL (e.g. `[a]({)`) and removes a brittle dance where the link
/// resolver previously had to reabsorb a stray `MdxTextExpression` token.
///
/// When no such tail encloses `pos` the link won't form. mdx-js then falls
/// back to expression scanning and errors on a dangling `{`. We mirror that
/// by returning false so the caller can run the expression scanner.
///
/// A resource may span lines, since `resourceBefore`,
/// `resourceDestinationAfter` and `resourceTitleAfter` all run
/// `factoryWhitespace` and `factoryTitle` accepts line endings; only the
/// destination is line-bounded. `scope_start` is where the enclosing block's
/// inline content begins, so a tail opening on an earlier line of the same
/// block still counts, as long as it is the line directly above: a resource
/// spanning two or more line endings, as in `[a](\n/u\n"ti{tle")`, stays a
/// known divergence. So do lazy and list-item continuations, since the forward
/// scan stops at a line that opens a block rather than tracking the enclosing
/// container.
#[cfg(feature = "mdx")]
fn is_inside_link_url_parens(bytes: &[u8], pos: usize, scope_start: usize) -> bool {
    let line_start = memchr::memrchr2(b'\n', b'\r', &bytes[..pos])
        .map(|i| i + 1)
        .unwrap_or(0);
    let opens_tail = |from: usize, to: usize| {
        memchr::memchr_iter(b']', &bytes[from..to])
            .any(|rel| bytes.get(from + rel + 1) == Some(&b'('))
    };
    let on_line = opens_tail(line_start, pos);
    let earlier = scope_start < line_start && opens_tail(scope_start, line_start);
    if !on_line && !earlier {
        return false;
    }
    let line_end = line_end_at(bytes, pos);
    let prefix = leading_container_prefix(&bytes[line_start..]);
    // Two reasons to widen: a `](` on this line wanted a label start it never
    // saw, or a tail on the line above is still open here. Only the second is
    // worth paying the block replay to guess at.
    let widen = if on_line {
        match replay_link_tails(
            bytes,
            pos,
            line_start,
            line_end,
            prefix,
            scope_start >= line_start,
        ) {
            TailReplay::Inside => return true,
            TailReplay::NeedsScope => true,
            TailReplay::Outside => {
                earlier && tail_reaches_from_line_above(bytes, line_start, pos, prefix)
            }
        }
    } else {
        tail_reaches_from_line_above(bytes, line_start, pos, prefix)
    };
    // `scope_start` is the block's own start, so this cannot cross into one.
    widen
        && scope_start < line_start
        && !earlier_lines_are_ambiguous(bytes, scope_start, line_start)
        && matches!(
            replay_link_tails(bytes, pos, scope_start, line_end, prefix, true),
            TailReplay::Inside
        )
}

/// Whether replaying the whole block could pay off: a `](` on the line
/// directly above `pos` opens a tail closing past it. Only that line is
/// scanned, so a tail whose `](` is two or more lines up, which a title
/// spanning both endings could still carry down, is refused and leaves the `{`
/// an expression. Cheaper than the block replay and only ever withholds it.
#[cfg(feature = "mdx")]
fn tail_reaches_from_line_above(
    bytes: &[u8],
    line_start: usize,
    pos: usize,
    prefix: usize,
) -> bool {
    if line_start == 0 {
        return false;
    }
    let mut eol = line_start - 1;
    if eol > 0 && bytes[eol] == b'\n' && bytes[eol - 1] == b'\r' {
        eol -= 1;
    }
    let above = memchr::memrchr2(b'\n', b'\r', &bytes[..eol])
        .map(|i| i + 1)
        .unwrap_or(0);
    let mut closes = TitleCloses::default();
    memchr::memchr_iter(b']', &bytes[above..eol]).any(|rel| {
        let rbracket = above + rel;
        bytes.get(rbracket + 1) == Some(&b'(')
            && link_tail_close(bytes, rbracket + 1, eol, prefix, &mut closes)
                .is_some_and(|rparen| rparen > pos)
    })
}

#[cfg(feature = "mdx")]
enum TailReplay {
    Inside,
    Outside,
    /// The line alone cannot answer: a `](` wanted a label start it never saw,
    /// or a backtick run whose pair may live on an earlier line.
    NeedsScope,
}

/// Replays `from..pos` the way micromark tokenizes inline content, tracking the
/// constructs that can own a bracket, and reports where `pos` landed.
#[cfg(feature = "mdx")]
fn replay_link_tails(
    bytes: &[u8],
    pos: usize,
    from: usize,
    scope_end: usize,
    prefix: usize,
    at_block_start: bool,
) -> TailReplay {
    let mut depth: u32 = 0;
    let mut image_kinds: u64 = 0;
    let mut inactive_below: u32 = 0;
    let mut closes = TitleCloses::default();
    let mut needs_scope = false;
    let mut i = from;
    // Refreshed only when `i` crosses a newline; rescanning per candidate makes
    // an ordinary line of links quadratic.
    let mut tail_line_end = line_end_at(bytes, from);
    while i < pos {
        match bytes[i] {
            b'\\' if i + 1 < scope_end && is_ascii_punctuation(bytes[i + 1]) => i += 2,
            b'`' => {
                needs_scope = true;
                let run = 1 + scan_ch_repeat(&bytes[i + 1..scope_end], b'`');
                match code_span_end(bytes, i + run, scope_end, run) {
                    // Only a replay from the block start can vouch for a pairing.
                    Some(end) if end > pos => {
                        return if at_block_start {
                            TailReplay::Inside
                        } else {
                            TailReplay::NeedsScope
                        };
                    }
                    Some(end) => i = end,
                    None => i += run,
                }
            }
            b'!' if bytes.get(i + 1) == Some(&b'[') => {
                depth += 1;
                if depth <= LINK_LABEL_MAX_TRACKED_DEPTH {
                    image_kinds |= 1 << (depth - 1);
                }
                i += 2;
            }
            b'[' => {
                depth += 1;
                if depth <= LINK_LABEL_MAX_TRACKED_DEPTH {
                    image_kinds &= !(1 << (depth - 1));
                }
                i += 1;
            }
            b']' => {
                if depth == 0 {
                    needs_scope |= bytes.get(i + 1) == Some(&b'(');
                    i += 1;
                    continue;
                }
                if i >= tail_line_end {
                    tail_line_end = line_end_at(bytes, i);
                }
                let is_image =
                    depth <= LINK_LABEL_MAX_TRACKED_DEPTH && image_kinds & (1 << (depth - 1)) != 0;
                let active = is_image || depth > inactive_below;
                depth -= 1;
                inactive_below = inactive_below.min(depth);
                if active
                    && bytes.get(i + 1) == Some(&b'(')
                    && let Some(rparen) =
                        link_tail_close(bytes, i + 1, tail_line_end, prefix, &mut closes)
                {
                    if rparen > pos {
                        return TailReplay::Inside;
                    }
                    // A link, unlike an image, deactivates the starts around it.
                    if !is_image {
                        inactive_below = depth;
                    }
                    i = rparen + 1;
                    continue;
                }
                i += 1;
            }
            _ => i += 1,
        }
    }
    if needs_scope {
        TailReplay::NeedsScope
    } else {
        TailReplay::Outside
    }
}

/// True if an earlier line of the block could hand `line_start` a construct
/// mdx-js resolves against the code span or tag satteri would keep.
#[cfg(feature = "mdx")]
fn earlier_lines_are_ambiguous(bytes: &[u8], scope_start: usize, line_start: usize) -> bool {
    if matches!(bytes.get(line_start), Some(b'{') | Some(b'<')) {
        return true;
    }
    // Only a `<` that can open a tag competes with the label; `5 < 6` cannot.
    memchr::memchr_iter(b'<', &bytes[scope_start..line_start]).any(|rel| {
        matches!(
            bytes.get(scope_start + rel + 1),
            Some(c) if c.is_ascii_alphabetic() || matches!(c, b'/' | b'>' | b'!' | b'?' | b'_' | b'$')
        )
    })
}

/// End of the line holding `i`.
#[cfg(feature = "mdx")]
fn line_end_at(bytes: &[u8], i: usize) -> usize {
    memchr::memchr2(b'\n', b'\r', &bytes[i..]).map_or(bytes.len(), |k| i + k)
}

/// End of the code span opened by a `run`-long backtick run, per CommonMark's
/// equal-length pairing, within a single line.
#[cfg(feature = "mdx")]
fn code_span_end(bytes: &[u8], from: usize, line_end: usize, run: usize) -> Option<usize> {
    let mut i = from;
    while let Some(rel) = memchr::memchr(b'`', &bytes[i..line_end]) {
        let start = i + rel;
        let len = 1 + scan_ch_repeat(&bytes[start + 1..line_end], b'`');
        if len == run {
            return Some(start + len);
        }
        i = start + len;
    }
    None
}

/// Index of the `)` closing the link tail opened at `lparen`, if it parses as
/// `(`, destination, optional whitespace-separated title, `)` on this line.
#[cfg(feature = "mdx")]
fn link_tail_close(
    bytes: &[u8],
    lparen: usize,
    line_end: usize,
    prefix: usize,
    closes: &mut TitleCloses,
) -> Option<usize> {
    let (i, line_end) = skip_resource_space(bytes, lparen + 1, line_end, prefix)?;
    let dest_end = scan_link_tail_dest(bytes, i, line_end)?;
    let (mut i, mut line_end) = skip_resource_space(bytes, dest_end, line_end, prefix)?;
    // Without separating whitespace the delimiter was part of the destination.
    if i > dest_end
        && let Some((title_end, title_line_end)) =
            scan_link_tail_title(bytes, i, line_end, prefix, closes)
    {
        (i, line_end) = skip_resource_space(bytes, title_end, title_line_end, prefix)?;
    }

    (i < line_end && bytes[i] == b')').then_some(i)
}

/// Spaces, tabs and line endings between the parts of a resource, following
/// paragraph continuation lines. Returns the position and its line's end.
#[cfg(feature = "mdx")]
fn skip_resource_space(
    bytes: &[u8],
    mut at: usize,
    mut line_end: usize,
    prefix: usize,
) -> Option<(usize, usize)> {
    loop {
        at += scan_while(&bytes[at..line_end], is_space_or_tab);
        if at < line_end {
            return Some((at, line_end));
        }
        at = resource_continuation_line(bytes, line_end, prefix)?;
        line_end = line_end_at(bytes, at);
    }
}

/// Where the next line's content starts when a paragraph continues past the
/// line ending at `eol`.
#[cfg(feature = "mdx")]
fn resource_continuation_line(bytes: &[u8], eol: usize, prefix: usize) -> Option<usize> {
    if eol >= bytes.len() {
        return None;
    }
    let mut next = eol + 1;
    if bytes[eol] == b'\r' && bytes.get(next) == Some(&b'\n') {
        next += 1;
    }
    let mut i = next;
    for _ in 0..prefix {
        i += scan_while_max(&bytes[i..], |b| b == b' ', 3);
        if bytes.get(i) != Some(&b'>') {
            return None;
        }
        i += 1;
        if bytes.get(i).is_some_and(|&b| is_space_or_tab(b)) {
            i += 1;
        }
    }
    let content = i + scan_while(&bytes[i..], is_space_or_tab);
    match bytes.get(content) {
        None | Some(b'\n' | b'\r') => None,
        Some(_) if opens_block(&bytes[content..]) => None,
        Some(_) => Some(i),
    }
}

/// Every branch below fires only on one of these bytes, so the match is a fast
/// reject for ordinary prose rather than a duplicate of their conditions.
#[cfg(feature = "mdx")]
fn opens_block(data: &[u8]) -> bool {
    let Some(&first) = data.first() else {
        return true;
    };
    if !matches!(
        first,
        b'<' | b'|'
            | b'{'
            | b':'
            | b'['
            | b'#'
            | b'>'
            | b'*'
            | b'-'
            | b'_'
            | b'+'
            | b'='
            | b'`'
            | b'~'
            | b'$'
            | b'0'..=b'9'
    ) {
        return false;
    }
    // `:` opens a directive fence or definition-list item, `[^` a footnote
    // definition; all interrupt a paragraph.
    matches!(first, b'<' | b'|' | b'{' | b':')
        || data.starts_with(b"[^")
        || scan_hrule(data).is_ok()
        || scan_atx_heading(data).is_some()
        || scan_code_fence(data).is_some()
        || scan_setext_heading(data).is_some()
        || scan_blockquote_start(data).is_some()
        || scan_listitem(data).is_some()
        || scan_math_fence(data).is_some()
        || scan_interrupting_container_extensions_fence(data)
}

/// End of the link destination starting at `start`, mirroring
/// [`scan_link_dest`] within a single line.
#[cfg(feature = "mdx")]
fn scan_link_tail_dest(bytes: &[u8], start: usize, line_end: usize) -> Option<usize> {
    let mut i = start;
    if bytes.get(i) == Some(&b'<') {
        i += 1;
        while i < line_end {
            match bytes[i] {
                b'<' => return None,
                b'>' => return Some(i + 1),
                b'\\' if i + 1 < line_end && is_ascii_punctuation(bytes[i + 1]) => i += 1,
                _ => {}
            }
            i += 1;
        }
        return None;
    }

    let mut nest = 0;
    while i < line_end {
        match bytes[i] {
            0x0..=0x20 => break,
            b'(' => {
                if nest > LINK_MAX_NESTED_PARENS {
                    return None;
                }
                nest += 1;
            }
            b')' => {
                if nest == 0 {
                    break;
                }
                nest -= 1;
            }
            b'\\' if i + 1 < line_end && is_ascii_punctuation(bytes[i + 1]) => i += 1,
            _ => {}
        }
        i += 1;
    }
    (nest == 0).then_some(i)
}

/// Running answer to "where does the next unescaped `\"`, `'` or `)` start".
/// Tail attempts move forward, so each byte of a line is examined once per
/// delimiter however many malformed tails ask.
#[cfg(feature = "mdx")]
#[derive(Default)]
struct TitleCloses {
    line_end: usize,
    asked_from: [usize; 3],
    found_at: [usize; 3],
}

#[cfg(feature = "mdx")]
impl TitleCloses {
    fn find(
        &mut self,
        bytes: &[u8],
        kind: usize,
        close: u8,
        start: usize,
        line_end: usize,
    ) -> Option<usize> {
        if self.line_end != line_end {
            *self = Self {
                line_end,
                asked_from: [usize::MAX; 3],
                found_at: [0; 3],
            };
        }
        let cached = self.asked_from[kind];
        if cached != usize::MAX && start >= cached && start <= self.found_at[kind] {
            return (self.found_at[kind] < line_end).then_some(self.found_at[kind]);
        }
        // An escaped close is still a close byte, so absence settles it outright.
        let mut i = if memchr::memchr(close, &bytes[start..line_end]).is_some() {
            let mut j = start;
            while j < line_end {
                if bytes[j] == close {
                    break;
                }
                if bytes[j] == b'\\' && j + 1 < line_end && is_ascii_punctuation(bytes[j + 1]) {
                    j += 1;
                }
                j += 1;
            }
            j
        } else {
            line_end
        };
        i = i.min(line_end);
        self.asked_from[kind] = start;
        self.found_at[kind] = i;
        (i < line_end).then_some(i)
    }
}

/// End of the link title starting at `start`, mirroring `scan_link_title`
/// within a single line: an unescaped `(` inside a paren title is content.
#[cfg(feature = "mdx")]
fn scan_link_tail_title(
    bytes: &[u8],
    start: usize,
    line_end: usize,
    prefix: usize,
    closes: &mut TitleCloses,
) -> Option<(usize, usize)> {
    if start >= line_end {
        return None;
    }
    let (kind, close) = match bytes[start] {
        b'\'' => (0, b'\''),
        b'"' => (1, b'"'),
        b'(' => (2, b')'),
        _ => return None,
    };
    // `TitleCloses` is keyed on the line, so crossing is driven from out here
    // and each line keeps its own memo.
    let mut from = start + 1;
    let mut line_end = line_end;
    loop {
        if let Some(i) = closes.find(bytes, kind, close, from, line_end) {
            return Some((i + 1, line_end));
        }
        from = resource_continuation_line(bytes, line_end, prefix)?;
        line_end = line_end_at(bytes, from);
    }
}

/// True if the byte at `pos` is the first content char of its source line,
/// allowing only whitespace, blockquote markers (`>`), and at most one
/// leading list marker (`-`, `+`, `*`, or `N.`/`N)`) before it. Used to
/// distinguish flow-position `{` (which follows mdx-js's strict-no-lazy
/// flow rule) from text-position `{` after the block-level pass already
/// failed to take it as flow.
#[cfg(feature = "mdx")]
fn is_at_paragraph_line_start(bytes: &[u8], pos: usize) -> bool {
    let mut j = pos;
    while j > 0 && bytes[j - 1] != b'\n' && bytes[j - 1] != b'\r' {
        j -= 1;
    }
    // j..pos is the slice from line start (after newline) up to `{`.
    let line = &bytes[j..pos];
    let mut k = 0;
    while k < line.len() && (line[k] == b' ' || line[k] == b'\t') {
        k += 1;
    }
    while k < line.len() && line[k] == b'>' {
        k += 1;
        if k < line.len() && (line[k] == b' ' || line[k] == b'\t') {
            k += 1;
        }
    }
    if k < line.len() {
        let b = line[k];
        let consumed = if matches!(b, b'-' | b'+' | b'*')
            && line.get(k + 1).is_some_and(|c| *c == b' ' || *c == b'\t')
        {
            Some(k + 2)
        } else if b.is_ascii_digit() {
            let mut m = k + 1;
            while m < line.len() && line[m].is_ascii_digit() {
                m += 1;
            }
            if m < line.len()
                && (line[m] == b'.' || line[m] == b')')
                && line.get(m + 1).is_some_and(|c| *c == b' ' || *c == b'\t')
            {
                Some(m + 2)
            } else {
                None
            }
        } else {
            None
        };
        if let Some(after_marker) = consumed {
            k = after_marker;
            while k < line.len() && (line[k] == b' ' || line[k] == b'\t') {
                k += 1;
            }
        }
    }
    k == line.len()
}

/// True if the line ending at `ix - 1` contains an `<` that opens an inline
/// MDX JSX tag whose scan extends past `ix`. Used to suppress paragraph
/// interrupts on the line beginning at `ix` when a multi-line JSX tag from
/// the previous line is still open — the would-be interrupt char (e.g. the
/// closing `>`) actually belongs to the JSX tag.
///
/// In the lite (non-mdx) build the body is a `false` stub so the `&&`-guard
/// call sites compile unchanged without naming the gated `mdx::*` scanners;
/// `ENABLE_MDX` is never set there, so the real body would have returned
/// without effect anyway.
#[cfg(not(feature = "mdx"))]
fn prev_line_has_open_inline_jsx(_bytes: &[u8], _ix: usize, _has_math: bool) -> bool {
    false
}

#[cfg(feature = "mdx")]
fn prev_line_has_open_inline_jsx(bytes: &[u8], ix: usize, has_math: bool) -> bool {
    if ix == 0 || ix > bytes.len() {
        return false;
    }
    let mut prev_line_end = ix - 1;
    if !matches!(bytes[prev_line_end], b'\n' | b'\r') {
        return false;
    }
    if prev_line_end > 0 && bytes[prev_line_end] == b'\n' && bytes[prev_line_end - 1] == b'\r' {
        prev_line_end -= 1;
    }
    let prev_line_start =
        memchr::memrchr2(b'\n', b'\r', &bytes[..prev_line_end]).map_or(0, |nl| nl + 1);
    let line = &bytes[prev_line_start..prev_line_end];
    // Lines without `<` (or `\`) can't open a JSX tag — skip them outright.
    // For lines that do contain `<`, memchr to each candidate; a per-byte
    // walk dominated the profile for long-line documents.
    if memchr::memchr2(b'<', b'\\', line).is_none() {
        return false;
    }
    let mut offset = 0;
    while offset < line.len() {
        // memchr to the next interesting byte (`<` or `\`). `\<Foo` is a
        // literal `<` (backslash neutralizes the next byte), so we treat
        // `\` as a skip-2 escape.
        let Some(rel) = memchr::memchr2(b'<', b'\\', &line[offset..]) else {
            return false;
        };
        let i = offset + rel;
        if line[i] == b'\\' {
            offset = i + 2;
            continue;
        }
        let pos = prev_line_start + i;
        // A `<` inside a code span or math span is literal content, not a JSX
        // tag opener. `$<$` is math, so a following `>` line is a real
        // blockquote, not the continuation of a phantom `<$…>` tag. Mirrors
        // the same guards on the inline `{` handler.
        if is_inside_code_span(bytes, pos) || (has_math && is_inside_math_span(bytes, pos)) {
            offset = i + 1;
            continue;
        }
        if let Some(len) = crate::mdx::scan_mdx_inline_jsx(&bytes[pos..])
            && pos + len > ix
        {
            return true;
        }
        offset = i + 1;
    }
    false
}

/// True if `pos` falls inside the opening or closing tag of an inline MDX
/// JSX element on the same source line (back to the previous newline). Used
/// to suppress the first-pass `MdxTextExpression` handler for `{` that
/// actually belongs to a JSX attribute spread or attribute value — those
/// braces will be consumed by the inline JSX scanner in the second pass.
///
/// Lite (non-mdx) build: `false` stub, see [`prev_line_has_open_inline_jsx`].
#[cfg(not(feature = "mdx"))]
fn is_inside_open_inline_jsx_tag(_bytes: &[u8], _pos: usize) -> bool {
    false
}

#[cfg(feature = "mdx")]
fn is_inside_open_inline_jsx_tag(bytes: &[u8], pos: usize) -> bool {
    if pos == 0 || pos > bytes.len() {
        return false;
    }
    // Walk back to the start of the enclosing block: a blank line (two
    // consecutive line endings) or the start of input. Inline JSX tags
    // can span multiple paragraph-continuation lines, so stopping at a
    // single newline mis-classifies `{...}` on a continuation line as
    // text-position when it's actually inside an open JSX attribute.
    let mut line_start = pos;
    let mut seen_newline = false;
    while line_start > 0 {
        let b = bytes[line_start - 1];
        if matches!(b, b'\n' | b'\r') {
            if seen_newline {
                break;
            }
            seen_newline = true;
        } else if !matches!(b, b' ' | b'\t') {
            seen_newline = false;
        }
        line_start -= 1;
    }
    let mut i = line_start;
    while i < pos {
        // memchr to the next `<` or `\` candidate. `\<` is a literal `<`
        // (backslash neutralizes the next byte), so we skip-2 on `\`.
        let Some(rel) = memchr::memchr2(b'<', b'\\', &bytes[i..pos]) else {
            return false;
        };
        let j = i + rel;
        if bytes[j] == b'\\' {
            i = j + 2;
            continue;
        }
        if let Some(len) = crate::mdx::scan_mdx_inline_jsx(&bytes[j..])
            && j + len > pos
        {
            return true;
        }
        i = j + 1;
    }
    false
}

/// Walk forward from `start_ix` (an atext-class char like `_`) through
/// `+`/`-`/`.`/`_`/alphanumeric to find an `@`, then check whether an
/// email autolink tokenizes exactly at `start_ix..`. Returns `(start, end,
/// "mailto:..")` on success.
fn scan_email_forward_from_atext(
    bytes: &[u8],
    underscore_ix: usize,
    begin_text: usize,
    paragraph_start: usize,
) -> Option<(usize, usize, String)> {
    // Walk forward from the `_` through local-part chars to the `@`.
    let mut at_ix = underscore_ix;
    while at_ix < bytes.len() && is_email_local_char(bytes[at_ix]) {
        at_ix += 1;
    }
    if at_ix >= bytes.len() || bytes[at_ix] != b'@' {
        return None;
    }
    // `scan_email_autolink` walks back to the real local-part start, which can
    // sit *before* the `_` (e.g. `1a+-_@…`, where the `_` is mid-local-part).
    // The email construct wins over the emphasis as long as that start is at
    // or after the pending text boundary; otherwise an already-emitted
    // Maybe* token covers it and we must defer to the post-pass.
    let (sc_start, sc_end, full_url, retry_needed) = scan_email_autolink(bytes, at_ix, true)?;
    if retry_needed
        || sc_start > underscore_ix
        || sc_start < begin_text
        || sc_start < paragraph_start
    {
        return None;
    }
    Some((sc_start, sc_end, full_url))
}

fn is_email_local_char(b: u8) -> bool {
    b.is_ascii_alphanumeric() || matches!(b, b'+' | b'-' | b'.' | b'_')
}

/// True when a candidate at `pos` must go through `handle_inline_pass1`
/// instead of being committed on the spot.
///
/// True whenever an earlier byte in the block could open a construct that ends
/// up owning the trigger's bytes: `[` a bracket opener, `<` a pointed autolink
/// or inline HTML, `` ` `` a code span, `$` a math span.
fn defer_autolink_decision(bytes: &[u8], block_start: usize, pos: usize, options: Options) -> bool {
    let before = &bytes[block_start..pos];
    memchr::memchr3(b'[', b'<', b'`', before).is_some()
        || (options.has_math() && memchr::memchr(b'$', before).is_some())
}

/// A GFM autolink literal the scanner accepted, before anything is committed
/// to the tree.
struct AutolinkDetection {
    start: usize,
    end: usize,
    link_type: LinkType,
    url: String,
}

/// `(can_open, can_close)` for the run of `run_len` delimiters at `at`.
fn delim_run_flags(
    text: &str,
    start: usize,
    at: usize,
    run_len: usize,
    mode: TableParseMode,
    options: Options,
) -> (bool, bool) {
    if run_len == 0 || at >= text.len() {
        return (false, false);
    }
    let s = &text[start..];
    let suffix = &text[at..];
    (
        delim_run_can_open(s, suffix, run_len, at - start, mode, options),
        delim_run_can_close(s, suffix, run_len, at - start, mode, options),
    )
}

/// Only GFM strikethrough ever adds `~` to micromark's attention markers.
fn is_attention_marker(c: char, options: Options) -> bool {
    c == '*' || c == '_' || (c == '~' && options.contains(Options::ENABLE_STRIKETHROUGH))
}

fn tilde_is_delimiter(options: Options) -> bool {
    options.contains(Options::ENABLE_STRIKETHROUGH) || options.contains(Options::ENABLE_SUBSCRIPT)
}

/// A run of `~` only means anything at the lengths its extensions define.
pub(crate) fn delim_run_is_valid(c: u8, count: usize, options: Options) -> bool {
    c != b'~' || count == 2 || (count == 1 && tilde_is_delimiter(options))
}

/// The delimiter run a `\` at `ix` would swallow, when the byte after it ends
/// a deferred autolink candidate. `None` when nothing there could open or
/// close, in which case the escape is left to stand as usual.
fn escaped_delim_run(
    text: &str,
    start: usize,
    ix: usize,
    mode: TableParseMode,
    options: Options,
) -> Option<(usize, (bool, bool))> {
    let bytes = text.as_bytes();
    let c = *bytes.get(ix + 1)?;
    if !matches!(c, b'*' | b'_' | b'~' | b'^') {
        return None;
    }
    let count = 1 + scan_ch_repeat(&bytes[(ix + 2)..], c);
    let fired = delim_run_flags(text, start, ix + 1, count, mode, options);
    if (!fired.0 && !fired.1) || !delim_run_is_valid(c, count, options) {
        return None;
    }
    Some((count, fired))
}

/// An email literal starting at the same offset as a `www.` literal, which
/// GFM resolves in the email construct's favour. `www_end` bounds the search to
/// the www span, which the committed path skips outright and the deferred path
/// was already rescanning.
fn detect_email_inside_www(
    bytes: &[u8],
    ix: usize,
    www_end: usize,
    paragraph_start: usize,
    begin_text: usize,
) -> Option<AutolinkDetection> {
    // `_` is the one atext byte that can precede a www literal, and the
    // attention arm's own email hook already owns that case.
    if ix > 0 && bytes[ix - 1] == b'_' {
        return None;
    }
    let at_ix = ix + memchr::memchr(b'@', &bytes[ix..www_end])?;
    let (email_start, email_end, full_url, retry_needed) =
        crate::post_passes::scan_email_autolink(bytes, at_ix, true)?;
    // Opening past the trigger is the `@` hook's; opening before it needs an
    // atext predecessor, and `_` is the only one a www literal takes.
    if email_start != ix {
        return None;
    }
    debug_assert!(
        !retry_needed,
        "a `/` predecessor would have failed the www literal first"
    );
    debug_assert!(
        email_start >= begin_text && email_start >= paragraph_start,
        "the trigger is inside the current text run"
    );
    Some(AutolinkDetection {
        start: email_start,
        end: email_end,
        link_type: LinkType::Email,
        url: email_addr(full_url),
    })
}

/// `scan_email_autolink` returns `mailto:<addr>`; arena_build's Email-link path
/// prepends `mailto:` again, so strip it here.
fn email_addr(full_url: String) -> String {
    full_url
        .strip_prefix("mailto:")
        .map(str::to_owned)
        .unwrap_or(full_url)
}

/// Detect a GFM autolink literal at a `h`/`H`/`w`/`W`/`@` trigger. Detection
/// only: committing or deferring is the caller's call, since it turns on
/// state this function cannot see.
fn detect_gfm_autolink(
    bytes: &[u8],
    ix: usize,
    byte: u8,
    paragraph_start: usize,
    begin_text: usize,
) -> Option<AutolinkDetection> {
    // Cheap reject first: most triggers in prose can't start an autolink.
    let is_www = match byte {
        b'h' | b'H' | b'w' | b'W' => crate::post_passes::match_autolink_scheme(bytes, ix)?.1,
        b'@' => {
            // Email requires at least one atext char immediately before @.
            if ix == 0 || !is_email_local_char(bytes[ix - 1]) {
                return None;
            }
            false
        }
        _ => return None,
    };

    match byte {
        b'h' | b'H' | b'w' | b'W' => {
            // `fnr_only`: only the find-and-replace fallback would accept,
            // which is the post-pass's job.
            let (start, _raw_end, end, full_url, fnr_only) =
                scan_autolink_literal(bytes, ix, ix == paragraph_start)?;
            if fnr_only {
                return None;
            }
            // GFM registers the email construct ahead of www at the same
            // offset, so an `@` the www span would swallow wins instead.
            if is_www
                && let Some(email) =
                    detect_email_inside_www(bytes, ix, end, paragraph_start, begin_text)
            {
                return Some(email);
            }
            Some(AutolinkDetection {
                start,
                end,
                link_type: LinkType::Autolink,
                url: full_url,
            })
        }
        b'@' => {
            // The local-part walkback can start the link before `ix`. If it
            // would cross an already-emitted Maybe* item, that construct owns
            // the bytes, so leave the email to the post-pass.
            let (email_start, email_end, full_url, retry_needed) =
                scan_email_autolink(bytes, ix, true)?;
            if retry_needed {
                return None;
            }
            if email_start < begin_text {
                return None;
            }
            if email_start < paragraph_start {
                return None;
            }
            // A live backslash escape directly before the local part blocks the
            // construct: the escape and its char are one token, so there is no
            // boundary. `scan_email_autolink` only rejects a preceding `/`.
            if email_start > 0
                && bytes[email_start - 1] == b'\\'
                && is_ascii_punctuation(bytes[email_start])
            {
                return None;
            }
            Some(AutolinkDetection {
                start: email_start,
                end: email_end,
                link_type: LinkType::Email,
                url: email_addr(full_url),
            })
        }
        _ => None,
    }
}

#[cfg(feature = "mdx")]
fn contains_blank_line(bytes: &[u8]) -> bool {
    let mut i = 0;
    let mut at_line_start = true;
    let mut line_has_non_ws = false;
    while i < bytes.len() {
        let b = bytes[i];
        if b == b'\n' || b == b'\r' {
            if !line_has_non_ws && !at_line_start {
                // A line that contained only whitespace.
                return true;
            }
            // Skip \r\n as one terminator.
            if b == b'\r' && i + 1 < bytes.len() && bytes[i + 1] == b'\n' {
                i += 2;
            } else {
                i += 1;
            }
            // If the next line is *also* a terminator, it's a blank line.
            if i < bytes.len() && (bytes[i] == b'\n' || bytes[i] == b'\r') {
                return true;
            }
            at_line_start = true;
            line_has_non_ws = false;
            continue;
        }
        at_line_start = false;
        if b != b' ' && b != b'\t' {
            line_has_non_ws = true;
        }
        i += 1;
    }
    false
}

fn is_directive_name_start_ascii(c: u8) -> bool {
    c.is_ascii_alphanumeric()
}

fn is_directive_name_char_ascii(c: u8) -> bool {
    c.is_ascii_alphanumeric() || c == b'-' || c == b'_'
}

/// A heading's trailing attribute block, with where it sits in the line.
struct HeadingAttrBlock<'a> {
    attrs: HeadingAttributes<'a>,
    /// Absolute offset of the opening `{`.
    block_open: usize,
    /// Absolute offset just past the closing `}`, set only when a trailing run
    /// of text directives follows the block and must be parsed as content.
    tail_start: Option<usize>,
}

/// Extent of a text directive starting at `colon` (`bytes[colon] == b':'`), or
/// `None`. Mirrors the inline text-directive scanner's guards so a heading's
/// trailing directive run is recognized the same way it will later be parsed.
fn scan_heading_text_directive(
    text: &str,
    bytes: &[u8],
    colon: usize,
    limit: usize,
) -> Option<usize> {
    if colon > 0 && bytes[colon - 1] == b':' {
        return None;
    }
    let (dir, end_pos) = parse_directive_after_colons(text, bytes, colon + 1)?;
    if end_pos > limit {
        return None;
    }
    // `:name:` (bare name closed by a colon) is an emoji, not a directive.
    if end_pos < limit && bytes[end_pos] == b':' && colon + 1 + dir.name.len() == end_pos {
        return None;
    }
    Some(end_pos)
}

/// True when the char at `ix` is a valid directive-name character.
/// Non-ASCII characters are accepted when they are ID_Continue (letters,
/// marks, digits), which matches micromark-extension-directive's behavior of
/// terminating names on unicode punctuation (CJK full-stop, etc.) while
/// accepting CJK letters.
fn scan_directive_name_char(bytes: &[u8], ix: usize) -> Option<usize> {
    if ix >= bytes.len() {
        return None;
    }
    let b = bytes[ix];
    if b < 0x80 {
        return if is_directive_name_char_ascii(b) {
            Some(1)
        } else {
            None
        };
    }
    let rest = core::str::from_utf8(&bytes[ix..]).ok()?;
    let ch = rest.chars().next()?;
    if unicode_id_start::is_id_continue(ch) {
        Some(ch.len_utf8())
    } else {
        None
    }
}

fn scan_directive_name_start(bytes: &[u8]) -> Option<usize> {
    if bytes.is_empty() {
        return None;
    }
    let b = bytes[0];
    if b < 0x80 {
        return if is_directive_name_start_ascii(b) {
            Some(1)
        } else {
            None
        };
    }
    let rest = core::str::from_utf8(bytes).ok()?;
    let ch = rest.chars().next()?;
    if unicode_id_start::is_id_start(ch) {
        Some(ch.len_utf8())
    } else {
        None
    }
}

/// Parse a directive name per remark-directive spec.
/// Returns (name, bytes_consumed) or None.
fn scan_directive_name(bytes: &[u8]) -> Option<(usize, usize)> {
    let first = scan_directive_name_start(bytes)?;
    let mut len = first;
    while let Some(n) = scan_directive_name_char(bytes, len) {
        len += n;
    }
    if len == 0 {
        return None;
    }
    let last = bytes[len - 1];
    if last == b'-' || last == b'_' {
        return None;
    }
    Some((0, len))
}

/// Parse a directive label `[content]`. Returns (label_start, label_end, total_consumed).
/// label_start/label_end are byte offsets within `bytes` of the inner content.
fn scan_directive_label(bytes: &[u8]) -> Option<(usize, usize, usize)> {
    if bytes.is_empty() || bytes[0] != b'[' {
        return None;
    }
    let mut depth = 1i32;
    let mut i = 1;
    let label_start = 1;
    while i < bytes.len() {
        match bytes[i] {
            b'\\' if i + 1 < bytes.len() => {
                i += 2;
                continue;
            }
            b'[' => depth += 1,
            b']' => {
                depth -= 1;
                if depth == 0 {
                    return Some((label_start, i, i + 1));
                }
            }
            b'\n' | b'\r' => return None,
            _ => {}
        }
        i += 1;
    }
    None
}

/// Parse directive attributes `{...}`. Returns (attrs, total_consumed).
fn scan_directive_attributes(bytes: &[u8]) -> Option<(Vec<(CowStr<'_>, CowStr<'_>)>, usize)> {
    if bytes.is_empty() || bytes[0] != b'{' {
        return None;
    }
    let mut i = 1;
    let end = loop {
        if i >= bytes.len() {
            return None;
        }
        match bytes[i] {
            b'}' => break i,
            b'\n' | b'\r' => return None,
            b'\\' if i + 1 < bytes.len() => i += 2,
            _ => i += 1,
        }
    };
    let inner = &bytes[1..end];
    let attrs = parse_directive_attrs_inner(inner);
    Some((attrs, end + 1))
}

fn parse_directive_attrs_inner(bytes: &[u8]) -> Vec<(CowStr<'_>, CowStr<'_>)> {
    let mut attrs: Vec<(CowStr<'_>, CowStr<'_>)> = Vec::new();
    let mut classes: Vec<&str> = Vec::new();
    let mut id: Option<&str> = None;
    let text = core::str::from_utf8(bytes).unwrap_or("");
    let mut i = 0;

    while i < bytes.len() {
        match bytes[i] {
            b' ' | b'\t' => {
                i += 1;
            }
            b'#' => {
                i += 1;
                let start = i;
                while i < bytes.len() && !is_attr_shortcut_terminator(bytes[i]) {
                    i += 1;
                }
                if i > start {
                    id = Some(&text[start..i]);
                }
            }
            b'.' => {
                i += 1;
                let start = i;
                while i < bytes.len() && !is_attr_shortcut_terminator(bytes[i]) {
                    i += 1;
                }
                if i > start {
                    classes.push(&text[start..i]);
                }
            }
            _ => {
                let name_start = i;
                while i < bytes.len()
                    && (bytes[i].is_ascii_alphanumeric()
                        || bytes[i] == b'-'
                        || bytes[i] == b'.'
                        || bytes[i] == b':'
                        || bytes[i] == b'_')
                {
                    i += 1;
                }
                if i == name_start {
                    i += 1;
                    continue;
                }
                let name = &text[name_start..i];
                if i < bytes.len() && bytes[i] == b'=' {
                    i += 1;
                    if i < bytes.len() && (bytes[i] == b'"' || bytes[i] == b'\'') {
                        let quote = bytes[i];
                        i += 1;
                        let val_start = i;
                        while i < bytes.len() && bytes[i] != quote {
                            if bytes[i] == b'\\' && i + 1 < bytes.len() {
                                i += 1;
                            }
                            i += 1;
                        }
                        let val = &text[val_start..i];
                        if i < bytes.len() {
                            i += 1; // skip closing quote
                        }
                        attrs.push((name.into(), val.into()));
                    } else {
                        let val_start = i;
                        while i < bytes.len()
                            && bytes[i] != b' '
                            && bytes[i] != b'\t'
                            && bytes[i] != b'}'
                        {
                            i += 1;
                        }
                        attrs.push((name.into(), text[val_start..i].into()));
                    }
                } else {
                    attrs.push((name.into(), "".into()));
                }
            }
        }
    }

    if let Some(id_val) = id {
        attrs.push(("id".into(), id_val.into()));
    }
    if !classes.is_empty() {
        attrs.push(("class".into(), classes.join(" ").into()));
    }
    attrs
}

fn is_attr_shortcut_terminator(c: u8) -> bool {
    matches!(c, b'#' | b'.' | b'}' | b' ' | b'\t')
}

/// Parse name[label]{attrs} after the colon(s). Returns (DirectiveAttrData, end_position).
/// The end_position is right after the last parsed component (name, label, or attrs).
fn parse_directive_after_colons<'a>(
    text: &'a str,
    bytes: &'a [u8],
    start: usize,
) -> Option<(DirectiveAttrData<'a>, usize)> {
    let remaining = &bytes[start..];
    let (_, name_len) = scan_directive_name(remaining)?;
    let name: CowStr<'a> = text[start..start + name_len].into();
    let mut pos = start + name_len;

    let mut label_start = 0usize;
    let mut label_end = 0usize;

    // Label (no space allowed between name and [)
    if pos < bytes.len()
        && bytes[pos] == b'['
        && let Some((ls, le, consumed)) = scan_directive_label(&bytes[pos..])
    {
        label_start = pos + ls;
        label_end = pos + le;
        pos += consumed;
    }

    // Attributes (no space allowed between label/name and {)
    let mut attributes = Vec::new();
    if pos < bytes.len()
        && bytes[pos] == b'{'
        && let Some((attrs, consumed)) = scan_directive_attributes(&bytes[pos..])
    {
        attributes = attrs;
        pos += consumed;
    }

    Some((
        DirectiveAttrData {
            name,
            attributes,
            label_start,
            label_end,
            initial_size: 0,
        },
        pos,
    ))
}

/// Assumes `text_bytes` is preceded by `<`.
fn get_html_end_tag(text_bytes: &[u8]) -> Option<&'static str> {
    static BEGIN_TAGS: &[&[u8]; 4] = &[b"pre", b"style", b"script", b"textarea"];
    static ST_BEGIN_TAGS: &[&[u8]; 3] = &[b"!--", b"?", b"![CDATA["];

    for (beg_tag, end_tag) in BEGIN_TAGS
        .iter()
        .zip(["</pre>", "</style>", "</script>", "</textarea>"].iter())
    {
        let tag_len = beg_tag.len();

        if text_bytes.len() < tag_len {
            // begin tags are increasing in size
            break;
        }

        if !text_bytes[..tag_len].eq_ignore_ascii_case(beg_tag) {
            continue;
        }

        // Must either be the end of the line...
        if text_bytes.len() == tag_len {
            return Some(end_tag);
        }

        // ...or be followed by whitespace, newline, or '>'.
        let s = text_bytes[tag_len];
        if is_ascii_whitespace(s) || s == b'>' {
            return Some(end_tag);
        }
    }

    for (beg_tag, end_tag) in ST_BEGIN_TAGS.iter().zip(["-->", "?>", "]]>"].iter()) {
        if text_bytes.starts_with(beg_tag) {
            return Some(end_tag);
        }
    }

    if text_bytes.len() > 1 && text_bytes[0] == b'!' && text_bytes[1].is_ascii_alphabetic() {
        Some(">")
    } else {
        None
    }
}

/// Strip up to `max_newlines` trailing `\n` (or `\r`) chars from the
/// current fenced-code-block's content. The current tree position is
/// inside the code block (the loop hasn't popped yet); we shrink the
/// last child text node's range to drop the trailing line endings.
/// `bytes` is the full source byte slice — needed to peek at the last
/// byte before shrinking.
fn trim_trailing_newlines_from_code_block(
    tree: &mut Tree<Item>,
    bytes: &[u8],
    max_newlines: usize,
) {
    let Some(last_child) = tree.cur() else { return };
    if !matches!(tree[last_child].item.body, ItemBody::Text { .. }) {
        return;
    }
    let mut stripped = 0;
    while stripped < max_newlines {
        let start = tree[last_child].item.start;
        let end = tree[last_child].item.end;
        if end <= start {
            break;
        }
        let last = bytes[end - 1];
        if last == b'\n' {
            tree[last_child].item.end = end - 1;
            // Treat `\r\n` as one newline.
            if end >= 2 && bytes[end - 2] == b'\r' {
                tree[last_child].item.end = end - 2;
            }
            stripped += 1;
        } else if last == b'\r' {
            tree[last_child].item.end = end - 1;
            stripped += 1;
        } else {
            break;
        }
    }
}

fn surgerize_tight_list(tree: &mut Tree<Item>, list_ix: TreeIndex) {
    let mut list_item = tree[list_ix].child;
    while let Some(listitem_ix) = list_item {
        let mut node_ix = tree[listitem_ix].child;
        while let Some(node) = node_ix {
            if let ItemBody::Paragraph = tree[node].item.body {
                tree[node].item.body = ItemBody::TightParagraph;
            }
            node_ix = tree[node].next;
        }

        list_item = tree[listitem_ix].next;
    }
}

fn fixup_end_of_definition_list(tree: &mut Tree<Item>, list_ix: TreeIndex) {
    let mut list_item = tree[list_ix].child;
    let mut previous_list_item = None;
    while let Some(listitem_ix) = list_item {
        match &mut tree[listitem_ix].item.body {
            ItemBody::DefinitionListTitle | ItemBody::DefinitionListDefinition(..) => {
                previous_list_item = list_item;
                list_item = tree[listitem_ix].next;
            }
            body @ ItemBody::MaybeDefinitionListTitle => {
                *body = ItemBody::Paragraph;
                break;
            }
            _ => break,
        }
    }
    if let Some(previous_list_item) = previous_list_item {
        tree.truncate_to_parent(previous_list_item);
    }
}

/// Determines whether the delimiter run starting at given index is
/// left-flanking, as defined by the commonmark spec (and isn't intraword
/// for _ delims).
/// suffix is &s[ix..], which is passed in as an optimization, since taking
/// a string subslice is O(n).
fn delim_run_can_open(
    s: &str,
    suffix: &str,
    run_len: usize,
    ix: usize,
    mode: TableParseMode,
    options: Options,
) -> bool {
    let next_char = if let Some(c) = suffix[run_len..].chars().next() {
        c
    } else {
        return false;
    };
    if next_char.is_whitespace() {
        return false;
    }
    if ix == 0 {
        return true;
    }
    if mode == TableParseMode::Active {
        if s.as_bytes()[..ix].ends_with(b"|") && !s.as_bytes()[..ix].ends_with(br"\|") {
            return true;
        }
        if next_char == '|' {
            return false;
        }
    }
    let delim = suffix.bytes().next().unwrap();
    if delim == b'*' && is_attention_marker(next_char, options) {
        return true;
    }
    if (delim == b'*' || delim == b'^') && !is_punctuation(next_char) {
        return true;
    }
    // GFM holds `~~` to the same flanking rules as `**`, so `a~~/foo~~` must not open.
    if delim == b'~' && run_len > 1 && !is_punctuation(next_char) {
        return true;
    }
    let prev_char = s[..ix].chars().last().unwrap();
    // Subscript is intraword by design (`H~2~O`), unlike GFM strikethrough.
    if delim == b'~' && options.contains(Options::ENABLE_SUBSCRIPT) && !is_punctuation(next_char) {
        return true;
    }
    if delim == b'~' && options.contains(Options::ENABLE_STRIKETHROUGH) && run_len == 1 {
        return !is_punctuation(next_char)
            || (is_punctuation(next_char)
                && (prev_char.is_whitespace() || is_punctuation(prev_char)));
    }

    // Double quotes can open after a non-space word character. For example, `에"About Me"` has
    // quoted text attached directly after Korean text. Digits are excluded: a
    // quote after a digit is an inch mark (`24"x36"`), not an opening quote.
    if delim == b'"' {
        return (!is_punctuation(next_char) && !prev_char.is_ascii_digit())
            || prev_char.is_whitespace()
            || is_punctuation(prev_char);
    }

    prev_char.is_whitespace()
        || is_punctuation(prev_char) && (delim != b'\'' || ![']', ')'].contains(&prev_char))
}

fn delim_run_can_close(
    s: &str,
    suffix: &str,
    run_len: usize,
    ix: usize,
    mode: TableParseMode,
    options: Options,
) -> bool {
    if ix == 0 {
        return false;
    }
    let prev_char = s[..ix].chars().last().unwrap();
    if prev_char.is_whitespace() {
        return false;
    }
    let next_char = if let Some(c) = suffix[run_len..].chars().next() {
        c
    } else {
        return true;
    };
    if mode == TableParseMode::Active {
        if s.as_bytes()[..ix].ends_with(b"|") && !s.as_bytes()[..ix].ends_with(br"\|") {
            return false;
        }
        if next_char == '|' {
            return true;
        }
    }
    let delim = suffix.bytes().next().unwrap();
    if delim == b'*' && is_attention_marker(prev_char, options) {
        return true;
    }
    if (delim == b'*' || delim == b'^') && !is_punctuation(prev_char) {
        return true;
    }
    if delim == b'~' && run_len > 1 && !is_punctuation(prev_char) {
        return true;
    }
    // Subscript is intraword by design (`H~2~O`), unlike GFM strikethrough.
    if delim == b'~' && options.contains(Options::ENABLE_SUBSCRIPT) {
        return true;
    }
    if delim == b'~' && options.contains(Options::ENABLE_STRIKETHROUGH) && run_len == 1 {
        return !is_punctuation(prev_char)
            || (is_punctuation(prev_char)
                && (next_char.is_whitespace() || is_punctuation(next_char)));
    }

    // Double quotes can close before a non-space word character. For example, `"About Me"로` has
    // Korean text attached directly after the quoted phrase.
    if delim == b'"' {
        return !is_punctuation(prev_char)
            || next_char.is_whitespace()
            || is_punctuation(next_char);
    }

    next_char.is_whitespace() || is_punctuation(next_char)
}

fn create_lut(options: &Options) -> LookupTable {
    special_bytes(options)
}

fn special_bytes(options: &Options) -> [bool; 256] {
    let mut bytes = [false; 256];
    let standard_bytes = [
        b'\n', b'\r', b'*', b'_', b'&', b'\\', b'[', b']', b'<', b'!', b'`',
    ];

    for &byte in &standard_bytes {
        bytes[byte as usize] = true;
    }
    if options.contains(Options::ENABLE_TABLES) {
        bytes[b'|' as usize] = true;
    }
    if tilde_is_delimiter(*options) {
        bytes[b'~' as usize] = true;
    }
    if options.contains(Options::ENABLE_SUPERSCRIPT) {
        bytes[b'^' as usize] = true;
    }
    if options.has_math() {
        bytes[b'$' as usize] = true;
        bytes[b'{' as usize] = true;
        bytes[b'}' as usize] = true;
    }
    if options.contains(Options::ENABLE_MDX) {
        bytes[b'{' as usize] = true;
        bytes[b'}' as usize] = true;
    }
    if options.has_smart_ellipses() {
        bytes[b'.' as usize] = true;
    }
    if options.has_smart_dashes() {
        bytes[b'-' as usize] = true;
    }
    if options.has_smart_quotes() {
        bytes[b'"' as usize] = true;
        bytes[b'\'' as usize] = true;
    }
    if options.contains(Options::ENABLE_DIRECTIVE) {
        bytes[b':' as usize] = true;
    }
    if options.contains(Options::ENABLE_GFM) {
        // GFM literal autolinks: protocol (h/H), www (w/W), email (@).
        // Fires during inline tokenization so URL bytes are consumed
        // before the bracket/emphasis/code-span resolvers see them.
        bytes[b'h' as usize] = true;
        bytes[b'H' as usize] = true;
        bytes[b'w' as usize] = true;
        bytes[b'W' as usize] = true;
        bytes[b'@' as usize] = true;
    }

    bytes
}

enum LoopInstruction<T> {
    /// Continue looking for more special bytes, but skip next few bytes.
    ContinueAndSkip(usize),
    /// Break looping immediately, returning with the given index and value.
    BreakAtWith(usize, T),
}

/// Result of `extend_indented_code_block`: the position end the block
/// should advertise and how many blank lines (line terminators) to
/// append to its `value` to match remark's behavior.
pub(crate) struct IndentCodeExtension {
    pub end_offset: u32,
    pub extra_blank_lines: usize,
}

/// Walk forward from `item.end` past blank/whitespace lines that meet
/// the indented-code-block indent threshold, returning the extended
/// end and the number of blank lines to fold back into the body.
///
/// Returns `None` when:
///   * the item isn't an indented code block, or
///   * the block was opened as a lazy continuation
///     (`IndentCodeBlock(true)`), or
///   * the parent is a blockquote — the blank lines downstream would
///     need to carry the `>` marker to extend, and firstpass already
///     produces the correct end for that case.
pub(crate) fn extend_indented_code_block(
    item: &Item,
    source: &[u8],
    parent_body: Option<&ItemBody>,
    start_column: u32,
    // Where to begin scanning forward — pass the mdast-trimmed end
    // (typically `mdast_position_end(item, source, parent_body)`) so
    // the initial `\n` between the code body and the first blank line
    // is correctly counted toward `extra_blank_lines`.
    start_from: u32,
) -> Option<IndentCodeExtension> {
    if !matches!(item.body, ItemBody::IndentCodeBlock(_)) {
        return None;
    }
    if matches!(item.body, ItemBody::IndentCodeBlock(true)) {
        return None;
    }
    if matches!(parent_body, Some(ItemBody::BlockQuote(_))) {
        return None;
    }

    // A blank-or-content line continues *this* code block only when
    // its tab-expanded indent reaches the chunk-content column
    // (start_column + 3): at top level col 4 of indent, inside a
    // list-item with content-col 3 it's col 6, etc.
    let required_indent_cols = (start_column as usize).saturating_add(3);
    let line_indent_cols = |bytes: &[u8], from: usize| -> usize {
        let mut p = from;
        let mut cols = 0usize;
        while p < bytes.len() {
            match bytes[p] {
                b' ' => cols += 1,
                b'\t' => cols += 4 - (cols % 4),
                _ => break,
            }
            p += 1;
        }
        cols
    };

    let mut pos = start_from as usize;
    let mut last_indented_end: Option<usize> = None;
    let mut last_indented_newlines = 0usize;
    let mut newlines_skipped = 0usize;
    loop {
        while pos < source.len() && (source[pos] == b'\r' || source[pos] == b'\n') {
            if source[pos] == b'\r' {
                pos += 1;
                if pos < source.len() && source[pos] == b'\n' {
                    pos += 1;
                }
            } else {
                pos += 1;
            }
            newlines_skipped += 1;
        }
        if pos >= source.len() {
            break;
        }
        let is_indented = line_indent_cols(source, pos) >= required_indent_cols;
        if !is_indented {
            // Whitespace-only line that doesn't meet the threshold
            // still sits between chunks — its newline contributes to
            // `newlines_skipped`. Anything else ends the extension.
            let mut p = pos;
            while p < source.len() && (source[p] == b' ' || source[p] == b'\t') {
                p += 1;
            }
            if p < source.len() && (source[p] == b'\r' || source[p] == b'\n') {
                pos = p;
                continue;
            }
            break;
        }
        let line_start = pos;
        while pos < source.len() && source[pos] != b'\n' && source[pos] != b'\r' {
            pos += 1;
        }
        let line_content = &source[line_start..pos];
        if !line_content.iter().all(|&b| b == b' ' || b == b'\t') {
            break;
        }
        last_indented_end = Some(pos);
        last_indented_newlines = newlines_skipped;
    }

    last_indented_end.map(|end| IndentCodeExtension {
        end_offset: end as u32,
        // The first walked newline followed the last chunk and is
        // already accounted for by the body content. Any additional
        // newlines are blank inter-chunk lines that need to be folded
        // back into the value.
        extra_blank_lines: last_indented_newlines.saturating_sub(1),
    })
}

/// remark quirk: when a list-item's `cont_end` (the end derived from its
/// children) lands on a line terminator AND a sibling list-item follows,
/// the list-item's position extends through the next item's marker to
/// that item's first content column.
///
/// Example: `- \`\`\`\n- d` — listItem 1 ends at the content column of
/// listItem 2 (start of `d`), not after its own fenced-code body.
///
/// Returns the extended end offset when the rule applies; otherwise
/// `None`.
pub(crate) fn extend_list_item_to_next_sibling_content(
    tree: &Tree<Item>,
    ix: TreeIndex,
    source: &[u8],
    cont_end: u32,
) -> Option<u32> {
    if cont_end == 0 {
        return None;
    }
    if !matches!(source.get(cont_end as usize - 1), Some(b'\n' | b'\r')) {
        return None;
    }
    let next_ix = tree[ix].next?;
    if !matches!(tree[next_ix].item.body, ItemBody::ListItem(..)) {
        return None;
    }
    let child_ix = tree[next_ix].child?;
    let child_start = tree[child_ix].item.start as u32;
    if child_start > cont_end {
        Some(child_start)
    } else {
        None
    }
}

/// Nested-blockquote extension: when an inner BlockQuote sits inside
/// outer BlockQuote(s) and the next line(s) carry the outer `>`
/// marker(s) but NOT this inner one's, the inner bq's end extends
/// through the outer markers. Matches micromark's lazy-bq position:
/// the inner bq absorbs trailing marker-only lines that its parent
/// still claims.
///
/// Skipped when this bq has a next sibling — those marker-only lines
/// then belong to the outer bq's space *between* siblings, not to
/// this inner one.
///
/// Returns the extended `cont_end` when the rule fires, else `None`.
pub(crate) fn extend_inner_blockquote_through_outer_markers(
    tree: &Tree<Item>,
    ix: TreeIndex,
    source: &[u8],
    cont_end: u32,
) -> Option<u32> {
    if !matches!(tree[ix].item.body, ItemBody::BlockQuote(..)) {
        return None;
    }
    if tree[ix].next.is_some() {
        return None;
    }
    let outer_bq_count = tree
        .walk_spine()
        .filter(|&&i| matches!(tree[i].item.body, ItemBody::BlockQuote(_)))
        .count();
    if outer_bq_count == 0 {
        return None;
    }
    let next_start = tree[ix]
        .next
        .map(|n| tree[n].item.start)
        .unwrap_or(source.len());
    let mut pos = cont_end as usize;
    let mut new_end: Option<u32> = None;
    while pos < next_start {
        if pos < source.len() && source[pos] == b'\r' {
            pos += 1;
            if pos < source.len() && source[pos] == b'\n' {
                pos += 1;
            }
        } else if pos < source.len() && source[pos] == b'\n' {
            pos += 1;
        }
        if pos >= next_start || pos >= source.len() {
            break;
        }
        let mut scan = pos;
        let mut markers = 0usize;
        while markers < outer_bq_count && scan < source.len() {
            while scan < source.len() && matches!(source[scan], b' ' | b'\t') {
                scan += 1;
            }
            if scan < source.len() && source[scan] == b'>' {
                scan += 1;
                markers += 1;
            } else {
                break;
            }
        }
        if markers < outer_bq_count {
            break;
        }
        // After the outer markers we must NOT see this inner bq's `>` —
        // and we must NOT see content. If we do, the inner bq either
        // continues normally (let firstpass handle it) or the next
        // block belongs to the outer bq, not to this inner one.
        let mut p = scan;
        while p < source.len() && matches!(source[p], b' ' | b'\t') {
            p += 1;
        }
        if p < source.len() && source[p] == b'>' {
            break;
        }
        if p < source.len() && !matches!(source[p], b'\n' | b'\r') {
            break;
        }
        new_end = Some(scan as u32);
        pos = p;
    }
    new_end
}

/// remark/micromark quirk: when a list inside a blockquote is followed
/// by more blockquote content (e.g. `>>- one\n>>\n  >  > two`), the
/// list "absorbs" the blank `>>` continuation lines between it and the
/// next sibling block. The list's end extends to right after the
/// parent blockquote markers on the last blank continuation line.
///
/// Only fires when the spine above this list is *pure* blockquotes
/// (no enclosing list-item) — see comments inline.
///
/// Returns the extended `cont_end` when the rule fires, else `None`.
pub(crate) fn extend_list_in_blockquote_through_marker_lines(
    tree: &Tree<Item>,
    ix: TreeIndex,
    source: &[u8],
    cont_end: u32,
) -> Option<u32> {
    if !matches!(tree[ix].item.body, ItemBody::List(..)) {
        return None;
    }
    let bq_count = tree
        .walk_spine()
        .filter(|&&i| matches!(tree[i].item.body, ItemBody::BlockQuote(_)))
        .count();
    if bq_count == 0 {
        return None;
    }
    let any_outer_listitem = tree
        .walk_spine()
        .any(|&i| matches!(tree[i].item.body, ItemBody::ListItem(..)));
    if any_outer_listitem {
        return None;
    }
    let next_start = tree[ix]
        .next
        .map(|n| tree[n].item.start)
        .unwrap_or(source.len());
    if next_start <= cont_end as usize {
        return None;
    }

    let mut pos = cont_end as usize;
    let mut new_end = cont_end;
    let mut saw_blank = false;
    let mut stop = false;
    while pos < next_start && !stop {
        if pos < source.len() && source[pos] == b'\r' {
            pos += 1;
            if pos < source.len() && source[pos] == b'\n' {
                pos += 1;
            }
        } else if pos < source.len() && source[pos] == b'\n' {
            pos += 1;
        }
        if pos >= next_start {
            break;
        }
        // CommonMark: a blockquote continuation line allows 0–3 spaces
        // of leading indent. 4+ spaces would make the line an indented
        // code block at the outer scope.
        {
            let mut leading_cols = 0usize;
            let mut ws = pos;
            while ws < source.len() && matches!(source[ws], b' ' | b'\t') && leading_cols < 4 {
                leading_cols += if source[ws] == b'\t' {
                    4 - (leading_cols % 4)
                } else {
                    1
                };
                ws += 1;
            }
            if leading_cols >= 4 {
                break;
            }
        }
        let mut markers_found = 0usize;
        let mut scan = pos;
        while markers_found < bq_count && scan < source.len() {
            while scan < source.len() && matches!(source[scan], b' ' | b'\t') {
                scan += 1;
            }
            if scan < source.len() && source[scan] == b'>' {
                scan += 1;
                markers_found += 1;
            } else {
                break;
            }
        }
        if markers_found < bq_count {
            break;
        }
        let after_markers = scan;
        let mut p = after_markers;
        while p < source.len() && matches!(source[p], b' ' | b'\t') {
            p += 1;
        }
        let blank = p >= source.len() || matches!(source[p], b'\n' | b'\r');
        if blank {
            // Extend through trailing whitespace on the marker line.
            new_end = p as u32;
            saw_blank = true;
            pos = p;
        } else if !saw_blank && source.get(p) == Some(&b'>') {
            // First line, deeper-nested bq that doesn't continue this
            // list. Include the outer markers in the list span, stop.
            new_end = after_markers as u32;
            stop = true;
        } else if !saw_blank {
            // First line, regular text — sibling paragraph in same bq.
            stop = true;
        } else {
            // Blank lines preceded content — don't include its markers.
            stop = true;
        }
    }

    if new_end > cont_end {
        Some(new_end)
    } else {
        None
    }
}

/// Compute the MDAST `position.end` byte offset for a block item.
///
/// The first-pass tree records each block's `item.end` as the offset just
/// past the block's bytes — which for most blocks includes the trailing
/// line terminator. The mdast representation, however, follows remark's
/// rules: most blocks strip all trailing line terminators, but math and
/// fenced-code at EOF preserve them; in-blockquote fenced code at EOF
/// trims; an unclosed list-item fenced code only keeps its trailing `\n`
/// when the outdent line opens a new container marker.
///
/// `parent_body` is the body of the spine parent (the block one level up)
/// at the time the block closes — needed for the blockquote-at-EOF and
/// list-item-fenced-code carve-outs.
pub(crate) fn mdast_position_end(
    item: &Item,
    source: &[u8],
    parent_body: Option<&ItemBody>,
) -> u32 {
    let end = item.end as u32;
    if !item.body.is_block_level() {
        return end;
    }
    let is_math = matches!(item.body, ItemBody::MathBlock(_));
    let math_at_eof = is_math && end as usize >= source.len();
    let is_fenced = matches!(item.body, ItemBody::FencedCodeBlock(..));
    let fenced_at_eof = is_fenced && end as usize >= source.len();
    let parent_is_bq = matches!(parent_body, Some(ItemBody::BlockQuote(_)));
    let parent_is_listitem = matches!(parent_body, Some(ItemBody::ListItem(..)));
    // Blockquote-parented fenced/math at EOF: the blockquote closed
    // because the next non-existent line carries no `>` marker, so
    // the trailing `\n` belongs to neither the inner block nor the
    // blockquote (`>$$\\\n` / `>\`\`\`\n` — both end before the `\n`).
    let fenced_at_eof_in_bq = fenced_at_eof && parent_is_bq;
    let math_at_eof_in_bq = math_at_eof && parent_is_bq;
    // Unclosed fenced/math in a list-item ended by container outdent:
    // the trailing `\n` at end-1 belongs to the inner block ONLY when
    // the outdent line opens a new container (list marker or `>`).
    // When the next block is a leaf (paragraph, heading, indented
    // code), remark drops the `\n`.
    let next_line_opens_container = end > 1
        && matches!(source.get(end as usize - 1), Some(b'\n' | b'\r'))
        && !matches!(source.get(end as usize - 2), Some(b'\n' | b'\r'))
        && parent_is_listitem
        && {
            match source.get(end as usize).copied() {
                Some(b'-' | b'*' | b'+' | b'>') => true,
                Some(c) if c.is_ascii_digit() => {
                    let mut p = end as usize + 1;
                    while p < source.len() && source[p].is_ascii_digit() {
                        p += 1;
                    }
                    matches!(source.get(p), Some(b'.' | b')'))
                }
                _ => false,
            }
        };
    let fenced_unclosed_in_listitem = is_fenced && !fenced_at_eof && next_line_opens_container;
    let math_unclosed_in_listitem = is_math && !math_at_eof && next_line_opens_container;
    let skip_trim =
        (math_at_eof || fenced_at_eof || fenced_unclosed_in_listitem || math_unclosed_in_listitem)
            && !fenced_at_eof_in_bq
            && !math_at_eof_in_bq;
    if skip_trim {
        return end;
    }
    let mut e = end;
    while e > item.start as u32 && matches!(source.get(e as usize - 1), Some(b'\n' | b'\r')) {
        e -= 1;
        // Math/fenced code: only strip a single line terminator.
        if is_math || is_fenced {
            break;
        }
    }
    e
}

type LookupTable = [bool; 256];

/// This function walks the byte slices from the given index and
/// calls the callback function on all bytes (and their indices) that are in the following set:
/// `` ` ``, `\`, `&`, `*`, `_`, `~`, `!`, `<`, `[`, `]`, `|`, `\r`, `\n`
/// It is guaranteed not call the callback on other bytes.
/// Whenever `callback(ix, byte)` returns a `ContinueAndSkip(n)` value, the callback
/// will not be called with an index that is less than `ix + n + 1`.
/// When the callback returns a `BreakAtWith(end_ix, opt+val)`, no more callbacks will be
/// called and the function returns immediately with the return value `(end_ix, opt_val)`.
/// If `BreakAtWith(..)` is never returned, this function will return the first
/// index that is outside the byteslice bound and a `None` value.
fn iterate_special_bytes<F, T>(
    lut: &LookupTable,
    bytes: &[u8],
    ix: usize,
    callback: F,
) -> (usize, Option<T>)
where
    F: FnMut(usize, u8) -> LoopInstruction<Option<T>>,
{
    scalar_iterate_special_bytes(lut, bytes, ix, callback)
}

fn scalar_iterate_special_bytes<F, T>(
    lut: &[bool; 256],
    bytes: &[u8],
    mut ix: usize,
    mut callback: F,
) -> (usize, Option<T>)
where
    F: FnMut(usize, u8) -> LoopInstruction<Option<T>>,
{
    while ix < bytes.len() {
        let b = bytes[ix];
        if lut[b as usize] {
            match callback(ix, b) {
                LoopInstruction::ContinueAndSkip(skip) => {
                    ix += skip;
                }
                LoopInstruction::BreakAtWith(ix, val) => {
                    return (ix, val);
                }
            }
        }
        ix += 1;
    }

    (ix, None)
}

/// Split the usual heading content range and the content inside the trailing attribute block.
///
/// Returns `(leading_content_len, Option<trailing_attr_block_range>)`.
///
/// Note that `trailing_attr_block_range` will be empty range when the block
/// is `{}`, since the range is content inside the wrapping `{` and `}`.
///
/// The closing `}` of an attribute block can have trailing whitespaces.
/// They are automatically trimmed when the attribute block is being searched.
///
/// However, this method does not trim the trailing whitespaces of heading content.
/// It is callers' responsibility to trim them if necessary.
fn extract_attribute_block_content_from_header_text(
    heading: &[u8],
) -> (usize, Option<Range<usize>>) {
    let heading_len = heading.len();
    let mut ix = heading_len;
    ix -= scan_rev_while(heading, |b| {
        b == b'\n' || b == b'\r' || b == b' ' || b == b'\t'
    });
    if ix == 0 {
        return (heading_len, None);
    }

    let attr_block_close = ix - 1;
    if heading.get(attr_block_close) != Some(&b'}') {
        // The last character is not `}`. No attribute blocks found.
        return (heading_len, None);
    }
    // move cursor before the closing right brace (`}`)
    ix -= 1;

    ix -= scan_rev_while(&heading[..ix], |b| {
        // Characters to be excluded:
        //  * `{` and `}`: special characters to open and close an attribute block.
        //  * `\\`: a special character to escape many characters and disable some syntaxes.
        //      + Handling of this escape character differs among markdown processors.
        //      + Escaped characters will be separate text node from neighbors, so
        //        it is not easy to handle unescaped string and trim the trailing block.
        //  * `<` and `>`: special characters to start and end HTML tag.
        //      + No known processors converts `{#<i>foo</i>}` into
        //        `id="&lt;i&gt;foo&lt;/&gt;"` as of this writing, so hopefully
        //        this restriction won't cause compatibility issues.
        //  * `\n` and `\r`: a newline character.
        //      + Setext heading can have multiple lines. However it is hard to support
        //        attribute blocks that have newline inside, since the parsing proceeds line by
        //        line and lines will be separate nodes even they are logically a single text.
        !matches!(b, b'{' | b'}' | b'<' | b'>' | b'\\' | b'\n' | b'\r')
    });
    if ix == 0 {
        // `{` is not found. No attribute blocks available.
        return (heading_len, None);
    }
    let attr_block_open = ix - 1;
    if heading[attr_block_open] != b'{' {
        // `{` is not found. No attribute blocks available.
        return (heading_len, None);
    }

    (attr_block_open, Some(ix..attr_block_close))
}

/// Parses an attribute block content, such as `.class1 #id .class2`.
///
/// Returns `(id, classes)`.
///
/// It is callers' responsibility to find opening and closing characters of the attribute
/// block. Usually [`extract_attribute_block_content_from_header_text`] function does it for you.
///
/// Note that this parsing requires explicit whitespace separators between
/// attributes. This is intentional design with the reasons below:
///
/// * to keep conversion simple and easy to understand for any possible input,
/// * to avoid adding less obvious conversion rule that can reduce compatibility
///   with other implementations more, and
/// * to follow the major design of implementations with the support for the
///   attribute blocks extension (as of this writing).
///
/// See also: [`Options::ENABLE_HEADING_ATTRIBUTES`].
///
/// [`Options::ENABLE_HEADING_ATTRIBUTES`]: `crate::Options::ENABLE_HEADING_ATTRIBUTES`
fn parse_inside_attribute_block(inside_attr_block: &str) -> Option<HeadingAttributes<'_>> {
    let mut id = None;
    let mut classes = Vec::new();
    let mut attrs = Vec::new();

    let bytes = inside_attr_block.as_bytes();
    let len = bytes.len();
    let mut i = 0;
    while i < len {
        if bytes[i].is_ascii_whitespace() {
            i += 1;
            continue;
        }
        // A value introduced by `=` may be wrapped in matching quotes, letting it
        // contain spaces; the token then runs to the closing quote rather than the
        // next whitespace. Backslashes can't reach this far (the block extractor
        // excludes them), so the quoted span needs no escape handling.
        let start = i;
        let mut quote = None;
        while i < len {
            let b = bytes[i];
            if let Some(q) = quote {
                if b == q {
                    quote = None;
                }
            } else if (b == b'"' || b == b'\'') && i > start && bytes[i - 1] == b'=' {
                quote = Some(b);
            } else if b.is_ascii_whitespace() {
                break;
            }
            i += 1;
        }

        let attr = &inside_attr_block[start..i];
        if attr.len() > 1 {
            let first_byte = attr.as_bytes()[0];
            if first_byte == b'#' {
                id = Some(attr[1..].into());
            } else if first_byte == b'.' {
                classes.push(attr[1..].into());
            } else if let Some((key, value)) = attr.split_once('=') {
                // `id=`/`class=` fold into the `#`/`.` channels so a heading
                // mixing shorthand and explicit forms emits a single id and
                // one merged class list instead of duplicate attributes.
                let value = unquote_attribute_value(value);
                match key {
                    "id" => id = Some(value.into()),
                    "class" => classes.push(value.into()),
                    _ => attrs.push((key.into(), Some(value.into()))),
                }
            } else {
                attrs.push((attr.into(), None));
            }
        }
    }

    if id.is_none() && classes.is_empty() && attrs.is_empty() {
        return None;
    }
    Some(HeadingAttributes { id, classes, attrs })
}

/// Strips a matching pair of surrounding `"` or `'` from an attribute value.
/// An unbalanced quote is kept verbatim so malformed input round-trips literally.
fn unquote_attribute_value(value: &str) -> &str {
    let bytes = value.as_bytes();
    if bytes.len() >= 2 {
        let quote = bytes[0];
        if (quote == b'"' || quote == b'\'') && bytes[bytes.len() - 1] == quote {
            return &value[1..value.len() - 1];
        }
    }
    value
}