leaf_core/

wysiwyg.rs

//! The WYSIWYG view: render the document with its markup *resolved*, not shown —
//! headings and code tagged with a typographic role (a frontend sizes or colours
//! them; see [`crate::style`]), `**bold**` as real bold, `# ` / `**` / `` ` ``
//! delimiters hidden — while keeping every visible glyph tied back to the source
//! byte it came from.
//!
//! That back-reference (`Glyph::src`) is what lets a caret still work: the caret
//! stays a source offset (shared with the source view), but the [`VisualMap`]
//! converts between an offset and a screen `(row, col)`, so cursor drawing,
//! mouse clicks, and vertical motion all operate in *visible* space.
//!
//! Left and Right instead walk the map's caret *stops* in document order. On
//! ordinary prose that's the same journey — the stops are laid out left to right
//! — and it steps over the hidden delimiters either way. They part company only
//! in a table, where the text is arranged in two dimensions and a cell wrapped
//! within its column continues *below* rather than to the right. Following the
//! document is what a caret means there.
//!
//! Text is walked from the AST (`str` nodes carry exact spans, and their text is
//! the verbatim source slice), so a Markdown and a Djot file that parse alike
//! render — and map — identically.

use std::cell::{Cell, RefCell};
use std::collections::HashMap;
use std::ops::Range;
use std::sync::OnceLock;

use twig::{Alignment, ContainerOrigin, DirectiveForm, Editor, FlatNode, Kind, QueryMatch};
use unicode_segmentation::UnicodeSegmentation;
use unicode_width::UnicodeWidthStr;

use crate::style::{
    Align, Baseline, FaceId, FaceRef, FaceTable, FontSize, LineHeight, MarkColor, Role, Style,
    TextColor, Token,
};

/// One rendered character plus the source byte offset it originates from.
/// Synthetic glyphs (a list bullet, a quote gutter) point at their block's
/// start, so clicking one lands the caret at the start of that block.
#[derive(Clone)]
pub struct Glyph {
    pub ch: char,
    pub style: Style,
    pub src: usize,
    /// Whether the caret may *rest* on this glyph. Decoration — a table border
    /// or a cell's alignment padding — is visible but isn't text, so the caret
    /// steps over it instead of into it. It also can't be a stop even in
    /// principle: a run of decoration shares one `src`, and a caret can only
    /// move by changing offset, so resting on it would pin horizontal motion.
    /// A click still maps through `src`, which is why decoration points at the
    /// text it decorates.
    ///
    /// Real text is a stop once per *grapheme cluster*, on the glyph that opens
    /// it: the continuation glyphs of an emoji or an accented letter are drawn,
    /// but standing between them is standing inside a character.
    pub stop: bool,
}

impl Glyph {
    /// The character to draw: [`ch`](Self::ch), except a list item's indent
    /// on its later rows ([`Role::ListIndent`]), which is spelled with the
    /// marker's characters for its width and drawn blank.
    pub fn drawn(&self) -> char {
        if self.style.role == Role::ListIndent {
            ' '
        } else {
            self.ch
        }
    }
}

/// One visual line. `end_src` is the source offset a caret sits at when placed
/// at the line's end (past its last glyph) — the anchor for end-of-line and
/// click-past-content.
///
/// `Clone` so a block's rows can be cached and re-emitted at a shifted offset
/// across an edit — see [`BlockCache`].
#[derive(Clone)]
pub struct VRow {
    pub glyphs: Vec<Glyph>,
    pub end_src: usize,
    /// A row that is drawn but holds no caret: a table's `├───┼───┤` rules, and
    /// the blank gap a block boundary is spelled with. Vertical motion steps
    /// over it, `pos_of_offset` never resolves onto it, and its stops (it has
    /// none) and `end_src` stay out of the map's stop table.
    ///
    /// Emptiness isn't the test — an empty paragraph is a blank row too, and a
    /// real caret stop. The test is whether the row is somewhere text can go.
    pub decoration: bool,
    /// This row is one line of a fenced or indented code block. Set on every row
    /// the `"code_block"` arm emits — including its blank lines, which carry no
    /// glyph to tell them apart otherwise. A frontend draws its own chrome (a
    /// border and a tinted background) around each maximal run of these, and
    /// scrolls them horizontally instead of wrapping; see
    /// [`VisualMap::code_blocks`]. Survives the row shuffling of [`BlockCache`]
    /// reuse and [`build_spliced`] because it rides on the row, not on a
    /// row-index span the way a table's picture does.
    pub code: bool,
    /// A fenced code block's info string (its language), carried on the *first*
    /// row of the block so it survives row reuse the way [`code`](Self::code)
    /// does. `None` on every other row, and on an indented block (which has no
    /// fence to label). A frontend paints it as a small label on the block's box
    /// and edits it through a prompt — see [`CodeBlockInfo::lang`]. It's a plain
    /// display string, not a source slice, so it needs no offset shifting; the
    /// label re-derives from twig on the next build.
    pub code_lang: Option<String>,
    /// This row belongs to a `:::name{.class}` directive container — twig's
    /// generic fenced-div block, whose meaning is entirely up to the host app
    /// (diaryx's `:::vis{.audience}` visibility blocks, say). Set on every row
    /// the `"directive"` arm emits, the same way [`code`](Self::code) marks a
    /// code block's rows, so a frontend can draw a tinted panel around each
    /// maximal run of these.
    pub directive: bool,
    /// A directive container's space-joined attrs — dot-prefixed classes
    /// (`.public .family` → `"public family"`) unioned with bare pandoc-style
    /// words (`public family`, no leading dot — diaryx's other `:::vis{...}`
    /// convention), carried on the block's *first* row only — the
    /// [`code_lang`](Self::code_lang) pattern. `None` on every other row, and
    /// when the directive carries no such attrs. A frontend paints it as a
    /// small label on the block's panel; it's a plain display string, not a
    /// source slice, so it rides row reuse untouched.
    pub directive_label: Option<String>,
    /// Set on the single placeholder row a block-level image renders to, carrying
    /// the image's destination and alt text; `None` on every other row. The row's
    /// glyphs are the default `🖼 alt` label (which a plain surface paints as-is);
    /// an image-capable frontend reads this to paint the real picture instead,
    /// skipping the row named by [`MediaInfo::rows_span`]. Like
    /// [`code_lang`](Self::code_lang) it's plain display strings, not source
    /// slices, so it rides row reuse and needs no offset shifting; the map's
    /// [`media`](VisualMap::media) side-table is derived from it once the rows
    /// are final, the same way [`code_blocks`](VisualMap::code_blocks) is.
    pub media: Option<MediaMark>,
    /// Set on the **first** row of a task list item, carrying whether its box is
    /// ticked; `None` on every other row, including a plain `list_item`'s. The
    /// row's glyphs already draw the box as `☐ `/`☑ ` in the marker's place, so a
    /// plain surface needs nothing further; a GUI reads this to paint a real
    /// checkbox widget and to know which way it is facing.
    ///
    /// A `bool` rather than a source span, for the reason
    /// [`code_lang`](Self::code_lang) is a plain string: it rides [`BlockCache`]
    /// reuse and [`build_spliced`] untouched, needing no offset shifting. To
    /// *toggle* the box, a frontend maps its click to a source offset the way it
    /// maps any other — the marker's glyphs carry the item's own `src` — and
    /// hands that to [`crate::Doc::toggle_task_at`].
    pub task: Option<bool>,
    /// Set on the single placeholder row a **leaf** directive (`::name{…}`)
    /// renders to, carrying its name and attributes; `None` on every other row.
    /// The container form isn't this — it wraps real blocks and marks each of
    /// them [`directive`](Self::directive) instead. Like [`media`](Self::media)
    /// it's plain display strings, so it rides row reuse untouched, and the map's
    /// [`directives`](VisualMap::directives) side-table is derived from it once
    /// the rows are final.
    pub leaf_directive: Option<DirectiveMark>,
    /// The heading level (1–6) of the block this row belongs to, on every row a
    /// `heading` emits (a long one wraps to several) and `None` everywhere else.
    ///
    /// A frontend that sizes a whole line — a proportional renderer giving the
    /// row a bigger line box — needs the level *per row*, and the glyphs can't
    /// always supply it: an empty heading (`# ` with nothing typed after it,
    /// which is what the toolbar's H1 leaves on a blank line) has no glyph to
    /// carry a [`Role::Heading`] at all, so a glyph scan called it body text and
    /// the line drew at body height until the first character landed. Riding the
    /// row says it once, for the empty case and the wrapped case alike.
    ///
    /// Per-*glyph* styling still comes from [`Role::Heading`] on the glyphs; this
    /// is the row-level fact, and the two agree wherever a heading has content —
    /// same `u8` level, clamped the same way [`heading_style`] clamps it.
    pub heading: Option<u8>,
    /// How this row's block is aligned across the measure — the author's
    /// `class="center"`, on every row the block emits and `None` for the
    /// theme's default, which is left.
    ///
    /// A *row* fact and not a glyph one for [`heading`](Self::heading)'s reason,
    /// and more sharply: alignment is a property of the *line*, not of the
    /// letters on it, so an empty paragraph the author has just centred has to
    /// carry it with no glyph to hang it on. It rides the row like a plain
    /// `Copy` flag, so [`BlockCache`] reuse and [`build_spliced`] carry it
    /// untouched.
    ///
    /// Read from the paragraph's or heading's own attributes and from those of
    /// every `div` around it, the nearest winning — so `<div class="center">`
    /// around three paragraphs centres all three, which is what the author of
    /// that HTML meant.
    pub align: Option<Align>,
    /// How far apart this row's block sets its lines, as a multiple of the
    /// theme's own line height — the author's `data-line-height`, on every row
    /// the block emits and `None` for the theme's spacing.
    ///
    /// A frontend that lays rows out in pixels scales the row's height by
    /// [`LineHeight::as_f32`]; one that draws a row per terminal line ignores
    /// it, the way it ignores a heading's size. Read at the same two levels
    /// [`align`](Self::align) is, and one of the menu's three names or the
    /// exact ratio the author asked for.
    pub line_height: Option<LineHeight>,
    /// What this row divides, on the blank rows a block boundary is *drawn* with
    /// and `None` on every other row — including the navigable blank lines of
    /// preserve-soft flow, which are somewhere text can go rather than a gap
    /// between blocks. So `boundary.is_some()` is exactly "this row is a drawn
    /// block boundary", the [`decoration`](Self::decoration) rows that come from
    /// [`Builder::emit_separators_before`].
    ///
    /// It exists because a boundary's *height* is a frontend decision but its
    /// *kind* is not. Typography spaces a boundary by what it separates — the
    /// margin above a heading is wider than the one between two paragraphs, so
    /// the heading groups with the text it introduces — and a frontend that has
    /// only rows to look at has to re-derive the structure by sniffing glyph
    /// roles. Three frontends sniffing separately is three chances to disagree
    /// about the same document. Core already knows, having just walked the AST
    /// to emit this row, so it says so once here and each frontend multiplies by
    /// its own spacing.
    pub boundary: Option<Boundary>,
    /// The offsets on this row where an inline mark's *content* ends under a
    /// hidden closing delimiter — the end of the `d` in `**bold**`, one byte
    /// before the `**` that draws nothing. Each is a caret stop with no glyph
    /// of its own: the caret standing there is drawn where the next glyph is,
    /// but typing there extends the mark, where typing past the delimiter
    /// leaves it. See [`VisualMap::mark_ends`] for the rule.
    ///
    /// Source offsets, so [`shift_row`] moves them with the glyphs; empty on
    /// decoration rows and on every row no mark closes on.
    pub mark_ends: Vec<usize>,
    /// The formulas this row stands in for, in glyph order: one [`MathMark`]
    /// per inline atom on the row, or the single block mark on the
    /// placeholder row a display formula renders to. Empty on every other
    /// row, and on the revealed line, where a formula is its TeX and no
    /// picture stands for it.
    ///
    /// Plain strings and a glyph *index* rather than a source offset, for
    /// [`media`](Self::media)'s reason: the mark rides [`BlockCache`] reuse and
    /// [`build_spliced`] untouched, and the glyph it names carries the offset.
    /// The map's [`math`](VisualMap::math) side-table is derived from these
    /// once the rows are final.
    pub math: Vec<MathMark>,
}

/// One formula a row stands in for — what a picture-capable frontend typesets
/// and draws in place of the glyph or the rows that hold its spot. See
/// [`VRow::math`] and, for the frontend's view of the same thing,
/// [`MathInfo`].
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct MathMark {
    /// The TeX between the delimiters, verbatim — a display block's keeps its
    /// newlines. What `leaf-math` typesets.
    pub tex: String,
    /// Display style (`$$…$$`) rather than text style (`$…$`). True for every
    /// block mark, and for a `$$…$$` written inside a line of prose, which is
    /// display style set inline.
    pub display: bool,
    /// The index into [`VRow::glyphs`] of the atom this mark stands behind —
    /// the one [`Role::Math`] glyph an inline formula renders to. `None` for
    /// a block mark, whose placeholder is the whole row.
    pub glyph: Option<usize>,
    /// How many rows a block formula reserves — the label row plus the blank
    /// fillers under it, the [`MediaMark::rows`] recipe, and from the same
    /// door: a terminal frontend that has typeset and measured the picture
    /// reports its height through [`crate::Doc::set_math_rows`]. `1` for an
    /// inline atom, which reserves nothing.
    pub rows: usize,
}

/// What a drawn block boundary separates: the kinds of the blocks it falls
/// between — the pair a frontend spaces by.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub struct Boundary {
    pub above: BlockClass,
    pub below: BlockClass,
}

/// The block kinds core tells apart when it walks a document — the vocabulary
/// [`Boundary`] is spelled in. A statement about *structure*, not about how any
/// of it should look: what a frontend does with "this gap sits above a heading"
/// is entirely the frontend's.
///
/// `Class` rather than `Kind` because [`twig::BlockKind`] already means
/// something else in this crate's public surface — the *command* vocabulary
/// (`Paragraph | Heading(n)`) a toolbar passes to [`Doc::set_block`](crate::Doc::set_block).
/// This is the reverse direction: what a block already *is*, read back off a
/// rendered row.
///
/// [`BlockClass::Other`] is the honest answer for a node kind core doesn't
/// separate out, so adding one here is additive for every frontend: nothing has
/// to change until it wants to space that kind differently.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub enum BlockClass {
    Paragraph,
    Heading,
    /// A whole list. Its *items* are [`BlockClass::ListItem`]; note that core
    /// draws no boundary row between two items of one list, tight or loose, so
    /// an item↔item pair never reaches a frontend.
    List,
    ListItem,
    Quote,
    Code,
    Table,
    /// A block-level image, video, or audio.
    ///
    /// Never reached through [`from_node_kind`](BlockClass::from_node_kind): a
    /// block picture is not a node of its own — [`Builder::media_only`] promotes
    /// the paragraph (or `<picture>`/`<video>` container) wrapping it — so the
    /// walk only ever sees the wrapper's kind. [`label_media_boundaries`] reads
    /// it back off the finished rows instead, after the fact.
    Media,
    /// A display formula on lines of its own — a paragraph holding nothing but
    /// a `$$…$$`. Never reached through [`from_node_kind`](BlockClass::from_node_kind)
    /// for [`Media`](BlockClass::Media)'s reason: the walk sees the wrapping
    /// paragraph, and [`label_media_boundaries`] relabels the gaps around the
    /// placeholder once the rows are final.
    Math,
    /// A `:::name{.class}` directive container.
    Directive,
    Rule,
    Footnote,
    Other,
}

impl BlockClass {
    /// Classify a twig node kind — the same vocabulary [`Builder::block`]
    /// matches on, so the two can't drift about what a block is. Both the
    /// whole-arena walk (which has [`FlatNode`]s) and the incremental top-level
    /// walk (which has only a query match's kind) reach it by this one door.
    pub fn from_node_kind(kind: &Kind) -> BlockClass {
        match kind {
            Kind::Para => BlockClass::Paragraph,
            Kind::Heading => BlockClass::Heading,
            Kind::BulletList | Kind::OrderedList | Kind::TaskList => BlockClass::List,
            Kind::ListItem | Kind::TaskListItem => BlockClass::ListItem,
            Kind::BlockQuote => BlockClass::Quote,
            Kind::CodeBlock => BlockClass::Code,
            Kind::Table => BlockClass::Table,
            Kind::Image => BlockClass::Media,
            // twig 2.8 folded `div`/`span`/`directive`/`element` into one
            // `container` kind, so a `:::note` panel and a promoted `<video>`
            // arrive here indistinguishable — telling them apart needs the
            // node's `origin`, and the incremental walk has only this kind.
            // `Directive` is the right answer for the case that motivates the
            // class (nothing else draws a tinted panel) and a harmless one for
            // the rest: `BlockClass` is descriptive and core never branches on
            // it. The one case where it was actively wrong — a promoted
            // `<video>`, which would have been handed to a frontend as something
            // to draw a fenced-div panel around — is corrected by
            // [`label_media_boundaries`] once the rows are final, along the same
            // door as a block image. Anything else that must be exact reads
            // [`container_is_directive`] off a real node.
            Kind::Container => BlockClass::Directive,
            Kind::ThematicBreak => BlockClass::Rule,
            Kind::Footnote => BlockClass::Footnote,
            _ => BlockClass::Other,
        }
    }
}

/// The name and attributes a leaf directive's placeholder row carries, so a
/// frontend that knows the host app's vocabulary can paint the real thing —
/// an embedded page for diaryx's `::embed{src=…}`, a generated table of
/// contents for a `::toc`, and the plain `⧉ name` label for one it doesn't
/// know. The peer of [`MediaMark`], and plain strings for the same reason: they
/// survive the row shuffling of [`BlockCache`] reuse and [`build_spliced`].
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct DirectiveMark {
    /// The directive's type — `embed`, `toc`, `vis` — with no leading colons.
    /// Core is agnostic of what it means: the vocabulary is the host app's.
    pub name: String,
    /// Its `{…}` attributes as `(key, value)` pairs in source order. A bare
    /// attribute (`{public}`) has a `None` value, the way twig reports it.
    pub attrs: Vec<(String, Option<String>)>,
    /// The directive's `[label]` text, flattened from its inline children, or
    /// empty when it has none. Also what the placeholder label shows.
    pub label: String,
    /// How many visual rows this directive reserves — the label row plus blank
    /// filler rows below it, so a frontend painting something real has the
    /// vertical room. `1` is the bare placeholder, and what every directive
    /// gets until a frontend that reserves rows — a terminal, which draws the
    /// host's lines over them — reports a height back through
    /// [`crate::Doc::set_directive_rows`], keyed by [`key`](Self::key). A
    /// pixel-laid-out GUI sets its own height regardless.
    pub rows: usize,
}

impl DirectiveMark {
    /// What [`crate::Doc::set_directive_rows`] knows this directive by.
    pub fn key(&self) -> DirectiveKey {
        DirectiveKey::new(&self.name, &self.label, &self.attrs)
    }
}

/// What a directive's reported height is keyed by: everything a host draws it
/// from — its name, its label and its attributes — and nothing about where it
/// stands. The peer of a picture's destination, which keys
/// [`crate::Doc::set_media_rows`]: two directives spelled alike draw alike and
/// so stand as tall, and the height survives an edit above them, which a key
/// by position would not.
///
/// An attribute's value is flattened, a bare attribute to `""`, because that is
/// the shape both bindings hand a host (`DirectiveAttr`, a string value) and so
/// the shape a host's report comes back in; the difference between `{wide}` and
/// `{wide=""}` is not one a drawing is expected to make.
#[derive(Clone, Debug, Default, PartialEq, Eq, Hash)]
pub struct DirectiveKey {
    pub name: String,
    pub label: String,
    pub attrs: Vec<(String, String)>,
}

impl DirectiveKey {
    /// The key for a directive named `name` with `label` (`""` for none) and
    /// `attrs` in source order.
    pub fn new(name: &str, label: &str, attrs: &[(String, Option<String>)]) -> Self {
        Self {
            name: name.to_string(),
            label: label.to_string(),
            attrs: attrs
                .iter()
                .map(|(k, v)| (k.clone(), v.clone().unwrap_or_default()))
                .collect(),
        }
    }
}

/// What a block-level media placeholder actually is, so a frontend knows which
/// widget to build over the reserved rows: a raster, a movie player, or a
/// transport with no picture at all. Core classifies and stops there — it opens
/// nothing, so this is a statement about the *markup*, not about a file it has
/// verified exists or can decode.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub enum MediaKind {
    /// A `![](…)` / `<img>` / `<picture>` — a still picture.
    Image,
    /// An HTML `<video>`. Markdown and Djot spell no video of their own, so this
    /// only ever arrives through `html_elements` promotion (or a `::video{…}`
    /// directive a host app maps itself, which core reports as a directive).
    Video,
    /// An HTML `<audio>` — a transport with no picture, so a frontend gives it a
    /// fixed control height rather than measuring an aspect ratio.
    Audio,
}

/// Which of the two caret homes a block media or a block leaf directive has —
/// see [`VisualMap::block_media_stop`] and [`VisualMap::block_directive_stop`].
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub enum MediaStop {
    /// The stop in front of the picture. What is typed here belongs above it.
    Before,
    /// The stop just past it. What is typed here belongs below it.
    After,
}

impl MediaKind {
    /// The emoji a plain surface prefixes the placeholder label with — the
    /// `🖼`/`🎬`/`🔊` that makes the row read as *a thing* rather than as text.
    fn sigil(self) -> char {
        match self {
            MediaKind::Image => '🖼',
            MediaKind::Video => '🎬',
            MediaKind::Audio => '🔊',
        }
    }
}

/// The destination and label a block-level media placeholder row carries, so a
/// capable frontend can resolve and paint the real thing. Plain strings (no
/// source offsets), so they survive the row shuffling of [`BlockCache`] reuse
/// and [`build_spliced`] untouched — see [`VRow::media`].
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct MediaMark {
    /// Whether this is a picture, a movie, or a sound — which widget the
    /// frontend builds over the reserved rows.
    pub kind: MediaKind,
    /// The media's link destination — a path, URL, or `data:` URI, verbatim from
    /// the AST. A frontend resolves a relative path against the document's
    /// directory itself; core holds no I/O.
    ///
    /// Empty is possible and legal for a `<video>`/`<audio>`, which may carry no
    /// `src` of its own and name its candidates in child `<source>`s instead —
    /// unlike an `<img>`, whose `src` *is* the picture. A frontend with an empty
    /// destination takes its URL from [`sources`](MediaMark::sources).
    pub destination: String,
    /// A `<picture>`'s theme/media alternatives, in document order, when this
    /// block image came from one; empty for a plain `![](…)` / bare `<img>`. Each
    /// is a `<source>`'s media query + candidate URL(s); a frontend that knows its
    /// theme picks the first whose media matches and falls back to [`destination`]
    /// (the `<img>`). Core keeps them verbatim and picks nothing — it has no theme.
    ///
    /// [`destination`]: MediaMark::destination
    pub sources: Vec<MediaSource>,
    /// The media's alt text (its rendered inline children, flattened), or empty
    /// when it has none. Also what the placeholder label shows. For a `<video>`/
    /// `<audio>` this is the element's own text content — the "your browser does
    /// not support…" fallback, which doubles as its accessible name.
    pub alt: String,
    /// A `<video poster="…">`'s still frame, verbatim, or empty when there is
    /// none (and always empty for an image or audio). It is an *image*
    /// destination, so a frontend already able to draw a picture can show it
    /// before the movie loads — or in place of one it can't play at all.
    pub poster: String,
    /// How many visual rows this media reserves — the placeholder label row plus
    /// the blank filler rows below it, so a frontend that paints a real raster has
    /// the vertical room to draw it. `1` is the bare placeholder (a frontend that
    /// can't draw pictures, or an image it couldn't resolve). A terminal frontend
    /// asks for as many rows as the fitted picture is tall; the pixel-laid-out GUI
    /// ignores this and sets its own row height, so it always leaves it `1`. The
    /// count comes from the frontend (via [`crate::Doc::set_media_rows`]) because
    /// core does no I/O and can't measure the image itself. See [`VRow::media`].
    pub rows: usize,
}

/// One `<source>` under a `<picture>`, `<video>`, or `<audio>`: a candidate URL
/// plus whichever of the two things HTML lets a `<source>` be chosen by — a
/// media query (`<picture>`) or a MIME type (`<video>`/`<audio>`). Verbatim from
/// the AST: core carries the alternatives and resolves none of them, having
/// neither a theme nor a codec list to judge them by.
///
/// The two spellings are normalised onto one field. `<picture>` writes
/// `srcset`, `<video>`/`<audio>` write `src`; both land in
/// [`srcset`](MediaSource::srcset), since a frontend wants the URL either way
/// and only `<picture>` ever uses the descriptor syntax.
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct MediaSource {
    /// The `<source media="…">` query, verbatim (`"(prefers-color-scheme: dark)"`),
    /// or empty for a `<source>` with no `media` (an unconditional override, and
    /// the norm for `<video>`/`<audio>`, which pick by codec rather than theme).
    pub media: String,
    /// The candidate URL(s): a `<picture>`'s `srcset` verbatim — one URL, or a
    /// comma-separated candidate list with `1x`/`2x`/width descriptors — or a
    /// `<video>`/`<audio>` `<source>`'s plain `src`. A frontend takes the first
    /// URL token; the theme and codec cases both only ever need that.
    pub srcset: String,
    /// The `<source type="…">` MIME type (`"video/webm"`), verbatim, or empty
    /// when the `<source>` declares none. How a `<video>`/`<audio>` frontend
    /// picks a candidate it can actually decode; a `<picture>`'s sources
    /// normally leave it empty and are chosen by [`media`](MediaSource::media).
    pub mime: String,
}

/// The rendered document plus the offset⇄position mapping the caret rides on.
#[derive(Clone, Default)]
pub struct VisualMap {
    /// The document's **default monospace rendering** — one [`VRow`] of glyphs
    /// per visual line, tables spelled with box-drawing borders (`│ ─ ┌┬┐…`) and
    /// cells padded to whole character-cell columns. Any monospace surface can
    /// draw these verbatim, so a consumer gets a working view for free: the TUI
    /// paints them as-is, and a five-line plain-text dump would too.
    ///
    /// It's a *default*, not the only truth. A frontend with its own geometry —
    /// a proportional GUI — lays text out in its own units, and for a table
    /// skips the box-drawn rows named by [`TableInfo::rows_span`] and draws from
    /// the structural [`TableInfo`] instead. The box glyphs live here rather than
    /// in a frontend precisely because they *are* a renderable default: unlike a
    /// colour (a role each surface must map to its own palette — see
    /// [`crate::style`]), `┌─┐` is finished text that needs no interpretation.
    pub rows: Vec<VRow>,
    /// The first source offset that is actually rendered — the caret floor for
    /// the WYSIWYG view. Non-zero when a leading `metadata` block (YAML/TOML
    /// frontmatter) is skipped: the frontmatter is preserved in the source and
    /// editable in the source view, but hidden and unreachable here, so the
    /// caret and selection can't wander into it (and copy won't grab it).
    pub content_start: usize,
    /// Every offset the caret may rest at, ascending and deduplicated: each
    /// row's stop glyphs plus the row's own end (the "after the last character"
    /// spot every line needs). Decoration contributes nothing.
    ///
    /// Left/Right read this instead of walking the grid, because the grid isn't
    /// laid out in offset order: a table with wrapped cells puts column 1's
    /// second line *below* column 2's first, so "the next stop rightward" and
    /// "the next stop in the document" part ways. Following the document is what
    /// a caret means — and on every row that *is* in order the two agree anyway,
    /// so nothing else has to change.
    stops: Vec<usize>,
    /// The caret's second home at the end of every hidden inline mark: the
    /// offset where the mark's content ends, one byte before its closing
    /// delimiter — ascending and deduplicated, from every row's
    /// [`VRow::mark_ends`].
    ///
    /// With delimiters hidden, `**bold** tail` draws one spot after the `d`
    /// and the source has two offsets for it: the content end (inside the
    /// mark, where typing extends the bold) and the byte past the `**` (where
    /// typing leaves it). Only the second is a glyph's offset, so only it was
    /// a stop, and a caret asked to rest at the first was snapped a whole
    /// character back onto the `d` — a drag over `bold` came back one letter
    /// short. The delete and backspace paths already settle the caret on the
    /// content end as its natural home there
    /// ([`crate::Doc::settle_inside_close_delims`]); this makes it one the
    /// caret can be placed at and step onto too.
    ///
    /// Kept apart from [`stops`](Self::stops) rather than merged in, because
    /// the two lists answer different questions. A stop with no glyph is
    /// invisible to a walk that pairs stops with characters — a system text
    /// input counting `position(from:offset:)` steps against the text it was
    /// shown would drift a character at every mark — and to word motion, which
    /// classifies a stop by the source byte under it (a `*`). So
    /// [`stop_after`](Self::stop_after) and its kin walk the glyph stops alone,
    /// and only the places a caret *rests* — snapping, resting checks, and
    /// Left/Right — read both.
    mark_ends: Vec<usize>,
    /// Every table in the document, in order, described structurally rather than
    /// drawn — see [`TableInfo`] for why both exist.
    pub tables: Vec<TableInfo>,
    /// Every fenced/indented code block, in order, as the range of [`rows`] it
    /// occupies — a frontend draws one bordered, tinted box around each and
    /// scrolls it horizontally rather than wrapping. Derived from the per-row
    /// [`VRow::code`] flag once the rows are final (so it survives incremental
    /// row reuse), the same way [`collect_stops`] derives the stop table.
    ///
    /// [`rows`]: VisualMap::rows
    pub code_blocks: Vec<CodeBlockInfo>,
    /// Every block-level image in the document, in order — one per placeholder
    /// row a frontend replaces with a real picture. Derived from the per-row
    /// [`VRow::media`] mark once the rows are final (so it survives incremental
    /// row reuse), the same way [`code_blocks`](VisualMap::code_blocks) is
    /// derived from [`VRow::code`].
    pub media: Vec<MediaInfo>,
    /// Every **leaf** directive in the document, in order — one per placeholder
    /// row a frontend may replace with whatever the host app's vocabulary makes
    /// of it. Derived from the per-row [`VRow::leaf_directive`] mark once the
    /// rows are final, exactly as [`media`](VisualMap::media) is.
    pub directives: Vec<DirectiveInfo>,
    /// Every formula standing as a picture in this map, in row order — each
    /// inline atom and each display block's placeholder. A frontend that
    /// paints pictures typesets each one's TeX and draws it over the glyph or
    /// the rows named. Derived from the per-row [`VRow::math`] marks once the
    /// rows are final, exactly as [`media`](VisualMap::media) is.
    ///
    /// A formula on the revealed line is not here: there it is its own TeX,
    /// drawn as code, and nothing stands in for it.
    pub math: Vec<MathInfo>,
    /// Every named font family this map's glyphs are set in, by the
    /// [`FaceId`] they carry — the side table that lets [`Style`] stay `Copy`
    /// while a family name stays a `String`.
    ///
    /// Not derived from the rows the way [`code_blocks`](Self::code_blocks) is,
    /// because the name is not on the rows: it is interned as the walker meets
    /// the attribute. So each of the three build paths assembles it from what
    /// it actually walked — a fresh build from its own walk, a cached build
    /// from each block's walk or the names its cache entry stored, a splice
    /// from the previous map's table plus the one block it re-rendered. A
    /// [`FaceId`] is derived from the name rather than being an index, which is
    /// what makes those three agree glyph for glyph; see the type's note.
    faces: FaceTable,
    /// The visible text tabulated over the stops — see [`Spelling`]. Derived
    /// on the first lookup that asks for it rather than by the build paths,
    /// since a map that is only ever drawn never needs it, and a splice
    /// would otherwise have to patch it the way it patches the stops.
    spelling: OnceLock<Spelling>,
}

/// The visible text ([`VisualMap::visible_text`]) as a table over the stops:
/// for stop `i`, the character the text spells it with — `None` for the
/// `'\n'` of a stop that draws no glyph — and the UTF-16 length of the text
/// before it. A UTF-16 index and a stop then find each other by binary
/// search, where they used to find each other by walking every character of
/// the document from the top: a frontend converts an `NSRange` end per
/// selection change and per misspelled word, and each conversion was a whole
/// document's work.
///
/// Built once per map and never maintained. The stops are private and fixed
/// for the map's life; the rows only lend the character each stop draws, and
/// a frontend that reshapes the rows after the build (the terminal inserts
/// blank filler rows) neither changes those characters nor asks this.
#[derive(Clone, Default)]
struct Spelling {
    /// Parallel to [`VisualMap::stops`].
    chars: Vec<Option<char>>,
    /// One longer than `chars`: `utf16[i]` is the UTF-16 length of the text
    /// before stop `i`, and the last entry the length of the whole text.
    utf16: Vec<usize>,
}

impl VisualMap {
    pub fn num_rows(&self) -> usize {
        self.rows.len()
    }

    /// The family name a glyph's [`FaceRef::Named`] stands for, or `None` for
    /// an id from another map — which a frontend draws in the theme's body
    /// face, as it draws a family it cannot resolve.
    pub fn face_name(&self, id: FaceId) -> Option<&str> {
        self.faces.name(id)
    }

    /// Every named family this map draws — how a frontend warms a font cache
    /// before it lays a frame out. See [`faces`](Self::faces).
    ///
    /// [`faces`]: VisualMap::faces
    pub fn faces(&self) -> &FaceTable {
        &self.faces
    }

    /// The width of `row` in display columns — the rightmost column its caret
    /// can occupy, and so what a goal column is clamped to on the way in.
    pub fn row_width(&self, row: usize) -> usize {
        self.rows.get(row).map_or(0, |r| r.width())
    }

    /// The screen `(row, col)` for a source offset — where to draw the caret:
    /// the *nearest* stop at or past `off`. Snaps a hidden offset (inside a
    /// delimiter) to the next visible glyph, and never resolves onto decoration
    /// (a table border, a cell's padding), which is drawn but holds no caret.
    ///
    /// "Nearest" rather than "the first one found" because a table's wrapped
    /// cells put rows slightly out of offset order: scanning top to bottom, the
    /// second line of column 1 comes *after* the first line of column 2 but
    /// holds smaller offsets. Where rows are in order the two rules agree.
    ///
    /// A soft wrap is the one place two rows want the same offset: the row above
    /// ends where the row below opens, the space the wrap ate being drawn on the
    /// row above and the offset past it being the row below's first character.
    /// It resolves *downstream*, to the row that character is on — the row
    /// above's last column is a phantom, a place the caret can be drawn but
    /// never sent, and resolving upstream into it is what pinned Down at the
    /// first wrap of a paragraph: it aimed at the row below's column 0, landed
    /// on the offset it already had, and read that back as the row above's end.
    ///
    /// The walk is over the rows, not their text: a row before `off` is
    /// dismissed by looking at where it opens and where it reaches — its first
    /// stop and its last — rather than at every glyph between, so the cost of
    /// a lookup mid-document is a few hundred rows' worth of two glyphs each,
    /// and the one row that holds the answer is the only one read through.
    /// It used to read every row through, and worse: the row's end candidate
    /// was built eagerly, and building it measured the row's width, which
    /// segments the row's whole text into grapheme clusters — so every row
    /// before `off` paid a full cluster walk to learn it held nothing, and a
    /// document of long paragraphs paid milliseconds per caret placement.
    ///
    /// Not a binary search, though the rows' opening stops ascend. A table's
    /// wrapped cells put several rows before the one an offset opens on that
    /// can still hold it, and `end_src` does not ascend across those rows at
    /// all, so a search would need a prefix table over the rows — and
    /// [`rows`](Self::rows) is public, reshaped by a frontend after the build
    /// (the terminal inserts filler rows under a heading), which is exactly
    /// what a table over it could not survive. The walk needs no table.
    pub fn pos_of_offset(&self, off: usize) -> (usize, usize) {
        // (src, row, the glyph — or `None` for the caret past the row's end)
        let mut best: Option<(usize, usize, Option<usize>)> = None;
        for (r, row) in self.rows.iter().enumerate() {
            if row.decoration {
                continue;
            }
            // A row's *first* stop never decreases from one row to the next —
            // true even across a table's wrapped cells, since a cell's lines run
            // downward. So once a row opens past the best found so far, no later
            // row can beat it and the scan stays proportional to `off`.
            let open = row
                .glyphs
                .iter()
                .find(|g| g.stop)
                .map_or(row.end_src, |g| g.src);
            if best.is_some_and(|b| open > b.0) {
                break;
            }
            // Offsets ascend *within* a row, so its last stop or its end is as
            // far as it reaches: a row that reaches short of `off` has nothing
            // to offer, and says so from its two ends.
            let reach = row
                .glyphs
                .iter()
                .rev()
                .find(|g| g.stop)
                .map_or(row.end_src, |g| g.src.max(row.end_src));
            if reach < off {
                continue;
            }
            // And so its first stop at or past `off` is the best it has.
            let cand = row
                .glyphs
                .iter()
                .enumerate()
                .find(|(_, g)| g.stop && g.src >= off)
                .map(|(i, g)| (g.src, r, Some(i)))
                .or_else(|| (row.end_src >= off).then_some((row.end_src, r, None)));
            // `<=`, so a tie goes to the later row: the only offset two rows
            // both hold is a wrap boundary, and it belongs to the row below.
            if let Some(c) = cand
                && best.is_none_or(|b| c.0 <= b.0)
            {
                best = Some(c);
            }
        }
        match best {
            Some((_, r, Some(i))) => (r, self.rows[r].col_of_glyph(i)),
            Some((_, r, None)) => (r, self.rows[r].width()),
            None => {
                let r = self.last_stop_row();
                (r, self.row_width(r))
            }
        }
    }

    /// The rows a source range occupies, inclusive: `(first, last)`.
    ///
    /// A *different question* from [`pos_of_offset`](Self::pos_of_offset), which
    /// is why it can't be spelled with two calls to it. That one answers "where
    /// does the caret go", and for a caret its forward snap is right — an offset
    /// inside a hidden delimiter has no column of its own, so the caret belongs
    /// at the next visible glyph, wherever that turns out to be. This one asks
    /// "which rows does this block cover", and there the snap is a trap: a
    /// footnote whose body *ends* in a link (`[^2]: [title](url)`) has a last
    /// byte inside the hidden destination, so `pos_of_offset(end - 1)` walked
    /// clean off the note's row and landed on the next note's — and a peek
    /// slicing `first..=last` out of the frame drew two notes where the reader
    /// asked for one. Every block ending in a link, an image, or any trailing
    /// hidden markup had the same fault; only a block ending in visible text
    /// (which is what the tests happened to use) did not.
    ///
    /// `row.end_src` is no help either: it is where the *rendered* text of a row
    /// ends, not how far into the source the block reaches, and redefining it
    /// would move every end-of-line caret.
    ///
    /// So the last row is found by asking which rows *open* before the range
    /// does, rather than by mapping its last byte: a row belongs to the range
    /// when its first caret stop lies before `range.end`. Decoration is skipped
    /// (a drawn gap between blocks is not part of either), and the answer is
    /// never shorter than one row — a range whose every byte is hidden still
    /// covers the row it started on.
    pub fn row_range_for(&self, range: Range<usize>) -> (usize, usize) {
        if self.rows.is_empty() {
            return (0, 0);
        }
        let first = self.pos_of_offset(range.start).0;
        let mut last = first;
        for (r, row) in self.rows.iter().enumerate().skip(first) {
            if row.decoration {
                continue;
            }
            let open = row
                .glyphs
                .iter()
                .find(|g| g.stop)
                .map_or(row.end_src, |g| g.src);
            if open >= range.end {
                // A row's first stop never decreases from one row to the next —
                // the invariant `pos_of_offset` breaks on, true even across a
                // table's wrapped cells — so nothing below can be in range.
                break;
            }
            last = r;
        }
        (first, last)
    }

    /// The source offset of the task checkbox drawn at `(row, col)`, or `None`
    /// when that cell holds no box — the hit-test a frontend runs on a click
    /// before treating it as a tick rather than a caret placement.
    ///
    /// Only the box's own cells answer. Clicking an item's *text* places the
    /// caret like any other click, so the box is a target aimed at rather than
    /// something tripped over while editing — which is also why this is a
    /// separate question from [`offset_of_pos`](Self::offset_of_pos) instead of
    /// a flag on the offset it returns.
    pub fn task_box_at(&self, row: usize, col: usize) -> Option<usize> {
        let r = self.rows.get(row)?;
        self.task_box_at_glyph(row, r.glyph_at_col(col)?)
    }

    /// [`task_box_at`](Self::task_box_at) keyed by glyph index rather than
    /// display column — for a frontend that shapes its own rows (the GUI) and so
    /// resolves a click to a glyph before it ever has a column.
    pub fn task_box_at_glyph(&self, row: usize, glyph: usize) -> Option<usize> {
        let r = self.rows.get(row)?;
        r.task?;
        let g = r.glyphs.get(glyph)?;
        (g.style.role == Role::ListMarker).then_some(g.src)
    }

    /// The source offset for a screen `(row, col)` — where a click or a
    /// visual-space move lands the caret. Clicking decoration maps through its
    /// `src`, which points at the text it decorates, so a click on a border or
    /// on a cell's padding lands in that cell.
    ///
    /// The inverse of [`pos_of_offset`](Self::pos_of_offset), which it has to
    /// agree with: `col` is a display column, and the one it names may be the
    /// far cell of a wide glyph — [`VRow::glyph_at_col`] is where that lands.
    pub fn offset_of_pos(&self, row: usize, col: usize) -> usize {
        let Some(r) = self.rows.get(row) else {
            // A click or drag below the last row — a short document with empty
            // space under it, dragged into to extend a selection. Land on the
            // document's last caret stop (its end), not offset 0: jumping the
            // caret to the top is the wrong direction, and 0 isn't even a stop
            // when the document opens on hidden frontmatter or a `# ` marker, so
            // returning it would leave the caret where it draws in one place and
            // types in another (`move_to` would then clamp it onto the unhomeable
            // frontmatter floor). `None` only for a document with no stops at all
            // (empty), where the caret has nowhere to be but 0.
            return self.stops.last().copied().unwrap_or(0);
        };
        match r.glyph_at_col(col).and_then(|i| r.glyphs.get(i)) {
            // A glyph that holds no caret is clickable, but where it points
            // isn't always somewhere the caret can be: the blank gap between two
            // paragraphs stands at an offset that belongs to neither of them,
            // and the tail of a grapheme cluster stands inside a character.
            // Land on the nearest real stop instead of handing back an offset
            // that looks like the gap but types into the paragraph above.
            Some(g) if !g.stop => self.nearest_stop(g.src),
            Some(g) => g.src,
            // A row's end is a stop by construction — unless the row is
            // decoration, which contributes none.
            None if r.decoration => self.nearest_stop(r.end_src),
            None => r.end_src,
        }
    }

    /// Which of a block media's two caret homes `off` is, or `None` for every
    /// other offset in the document.
    ///
    /// [`block_media`](Builder::block_media) gives a block-level image, video, or
    /// audio exactly two stops — one in front of it and one just past it — and
    /// nothing inside the markup. Both are ordinary offsets to everything else in
    /// core, but they are the two places where inserting text would *dissolve the
    /// picture*: `![](p.png)` with anything typed against it is no longer a block
    /// image but a paragraph with an inline one, and the frontend that was
    /// painting a photo there paints a text run instead. A caller that is about to
    /// insert asks this so it can open a paragraph first — see
    /// [`Doc::insert`](crate::Doc::insert).
    ///
    /// An *inline* image reports `None`: it has no placeholder row and no stops of
    /// its own, and typing beside one is ordinary editing.
    ///
    /// Answers with the media's own source span as well, since a caller that has
    /// to keep the picture whole usually has to address it — [`Doc::backspace`]
    /// takes the picture out in one piece rather than nibbling a byte off its
    /// markup, which is the same dissolution from the other side.
    ///
    /// [`Doc::backspace`]: crate::Doc::backspace
    pub fn block_media_stop(&self, off: usize) -> Option<(MediaStop, Range<usize>)> {
        self.media
            .iter()
            .find_map(|m| self.placeholder_stop(m.rows_span.start, off))
    }

    /// Which of a block leaf directive's two caret homes `off` is, or `None`
    /// for every other offset — [`block_media_stop`](Self::block_media_stop)
    /// for the [`directives`](Self::directives), for the same reason.
    ///
    /// A leaf directive is drawn on the picture's recipe (`⧉ label`, one stop
    /// in front of it and one just past it), and dissolves the same way: text
    /// typed against `::embed{src=x}` either end is a paragraph of raw source
    /// (`X::embed{src=x}`, `::embed{src=x}X`), and djot's empty `::: name`
    /// fence either stops opening or stops closing. `::page-break` is one of
    /// these. Answers with the directive's own span, as its peer does.
    pub fn block_directive_stop(&self, off: usize) -> Option<(MediaStop, Range<usize>)> {
        self.directives
            .iter()
            .find_map(|d| self.placeholder_stop(d.rows_span.start, off))
    }

    /// Whether `off` is one of the two stops of the placeholder row `row` —
    /// a block picture's or a leaf directive's, which share the recipe — and
    /// the span the placeholder stands for.
    fn placeholder_stop(&self, row: usize, off: usize) -> Option<(MediaStop, Range<usize>)> {
        let row = self.rows.get(row)?;
        // Every glyph of the `🖼 alt` / `⧉ label` label maps to the block's
        // start offset; the row's end is past its markup. Read the start off
        // the label rather than the first glyph, which on a quoted or listed
        // picture is the block prefix and points at the gutter.
        let start = row
            .glyphs
            .iter()
            .find(|g| g.style.role == Role::Image)
            .map(|g| g.src)?;
        if off == start {
            return Some((MediaStop::Before, start..row.end_src));
        }
        if off == row.end_src {
            return Some((MediaStop::After, start..row.end_src));
        }
        None
    }

    /// Whether `off` is a table's trailing caret stop — the one home past a
    /// table's last cell, at the block's own end ([`TableInfo::end_src`]).
    ///
    /// The table's peer of [`block_media_stop`](Self::block_media_stop)'s
    /// `After`: text inserted at that offset joins the table's last source
    /// line, and a line glued under a table is a row of it (`| 1 | 2 |x`), so
    /// a caller about to insert there opens a paragraph first — see
    /// [`Doc::insert`](crate::Doc::insert). Nothing else about the offset is
    /// special: it is where Down from the last row lands and where a click in
    /// the blank space under a trailing table lands.
    pub fn table_end_stop(&self, off: usize) -> bool {
        self.tables.iter().any(|t| t.end_src == off)
    }

    /// Snap `off` to the nearest caret stop — the funnel a frontend that
    /// hit-tests pixels straight to a source offset must run its result through.
    /// A click or drag can land in the blank gap a paragraph break is drawn with,
    /// or inside a hidden delimiter; both are offsets the caret can't rest at, so
    /// resting there would draw the caret in one place and type in another. This
    /// settles it on a real caret home instead. Idempotent on an offset that is
    /// already a stop — the `(row, col)` click path already snaps this way inside
    /// [`offset_of_pos`](Self::offset_of_pos), and this gives the pixel path the
    /// same guarantee. Returns `off` unchanged only for an empty document (no
    /// stops at all).
    pub fn snap_to_stop(&self, off: usize) -> usize {
        self.nearest_stop(off)
    }

    /// The caret stop nearest `off`, preferring the one before it when `off`
    /// falls exactly between two. Returns `off` unchanged if there are no stops
    /// at all (an empty document). A mark's content end counts: it is a place
    /// the caret rests, and the one a drag ending on a marked word means.
    fn nearest_stop(&self, off: usize) -> usize {
        let before = Self::last_at_or_before(&self.stops, off)
            .max(Self::last_at_or_before(&self.mark_ends, off));
        let after = match (
            Self::first_at_or_after(&self.stops, off),
            Self::first_at_or_after(&self.mark_ends, off),
        ) {
            (Some(a), Some(b)) => Some(a.min(b)),
            (a, b) => a.or(b),
        };
        match (before, after) {
            (Some(b), Some(a)) if off - b <= a - off => b,
            (_, Some(a)) => a,
            (Some(b), None) => b,
            (None, None) => off,
        }
    }

    /// The glyph stop nearest `off` — [`nearest_stop`](Self::nearest_stop)
    /// for a walk that pairs stops with characters, which a mark's content
    /// end has none of. A caret resting on one resolves to the glyph stop
    /// drawn at the same spot, the one just past the hidden delimiter, so the
    /// text a system input is shown from there and the steps it counts agree.
    pub fn snap_to_glyph_stop(&self, off: usize) -> usize {
        if self.mark_ends.binary_search(&off).is_ok()
            && let Some(next) = Self::first_at_or_after(&self.stops, off)
        {
            return next;
        }
        let before = Self::last_at_or_before(&self.stops, off);
        let after = Self::first_at_or_after(&self.stops, off);
        match (before, after) {
            (Some(b), Some(a)) if off - b <= a - off => b,
            (_, Some(a)) => a,
            (Some(b), None) => b,
            (None, None) => off,
        }
    }

    /// The last of `sorted` at or before `off`, if any.
    fn last_at_or_before(sorted: &[usize], off: usize) -> Option<usize> {
        let i = sorted.partition_point(|&s| s <= off);
        i.checked_sub(1).map(|i| sorted[i])
    }

    /// The first of `sorted` at or after `off`, if any.
    fn first_at_or_after(sorted: &[usize], off: usize) -> Option<usize> {
        let i = sorted.partition_point(|&s| s < off);
        sorted.get(i).copied()
    }

    /// The next place the caret rests past `off` — the next glyph stop or the
    /// next mark's content end, whichever comes first. What Right walks:
    /// leaving `**bold**` from the `d` is two presses, one onto the end of the
    /// bold (still bold, the toolbar lit) and one past its delimiter, at the
    /// same spot on screen. [`stop_after`](Self::stop_after) is the walk that
    /// skips the first, for every caller that pairs stops with characters.
    pub fn caret_stop_after(&self, off: usize) -> Option<usize> {
        match (
            self.stop_after(off),
            Self::first_at_or_after(&self.mark_ends, off + 1),
        ) {
            (Some(a), Some(b)) => Some(a.min(b)),
            (a, b) => a.or(b),
        }
    }

    /// The previous place the caret rests before `off` — the mirror of
    /// [`caret_stop_after`](Self::caret_stop_after), what Left walks.
    pub fn caret_stop_before(&self, off: usize) -> Option<usize> {
        self.stop_before(off).max(
            off.checked_sub(1)
                .and_then(|o| Self::last_at_or_before(&self.mark_ends, o)),
        )
    }

    /// Whether the caret can occupy `row` at all: decoration rows (a table's
    /// border rules) are stepped over by vertical motion.
    pub fn row_is_navigable(&self, row: usize) -> bool {
        self.rows.get(row).is_some_and(|r| !r.decoration)
    }

    /// The first offset the caret can rest at on `row` — its first stop, or the
    /// row's own end when it holds no text (an empty paragraph). `None` for a
    /// decoration row, which holds no caret at all.
    ///
    /// Not `offset_of_pos(row, 0)`: column 0 of a quoted or listed row is the
    /// gutter, and a gutter's `src` points at the *block* it opens, so the stop
    /// nearest it is the one on the block's first row rather than on this one.
    /// Which is right for a click — the gutter decorates the whole block — and
    /// wrong for Home, whose whole question is where *this* row starts.
    pub fn row_start(&self, row: usize) -> Option<usize> {
        let r = self.rows.get(row).filter(|r| !r.decoration)?;
        Some(
            r.glyphs
                .iter()
                .find(|g| g.stop)
                .map_or(r.end_src, |g| g.src),
        )
    }

    /// The last row the caret can rest on — the fallback when an offset is past
    /// everything rendered (a table's bottom border must not swallow the caret).
    fn last_stop_row(&self) -> usize {
        (0..self.rows.len())
            .rev()
            .find(|&r| self.row_is_navigable(r))
            .unwrap_or(0)
    }

    /// The nearest row above `row` the caret can occupy, skipping decoration.
    pub fn navigable_above(&self, row: usize) -> Option<usize> {
        (0..row.min(self.rows.len()))
            .rev()
            .find(|&r| self.row_is_navigable(r))
    }

    /// The nearest row below `row` the caret can occupy, skipping decoration.
    pub fn navigable_below(&self, row: usize) -> Option<usize> {
        ((row + 1)..self.rows.len()).find(|&r| self.row_is_navigable(r))
    }

    /// The caret stop just before `off` — one press of Left. `None` at the
    /// first stop in the document.
    ///
    /// Runs of decoration (a table border, a cell's alignment padding) are
    /// stepped over in a single press: they hold no stop, so they aren't in the
    /// table to land on.
    pub fn stop_before(&self, off: usize) -> Option<usize> {
        let i = self.stops.partition_point(|&s| s < off);
        i.checked_sub(1).map(|i| self.stops[i])
    }

    /// The caret stop just after `off` — one press of Right. `None` at the last
    /// stop in the document.
    pub fn stop_after(&self, off: usize) -> Option<usize> {
        let i = self.stops.partition_point(|&s| s <= off);
        self.stops.get(i).copied()
    }

    /// The first caret stop at or past `off` — where the caret at a hidden
    /// offset is *drawn*, and so where a rightward walk over the rendered text
    /// starts from.
    pub fn stop_at_or_after(&self, off: usize) -> Option<usize> {
        let i = self.stops.partition_point(|&s| s < off);
        self.stops.get(i).copied()
    }

    /// The last caret stop at or before `off` — where a leftward walk starts
    /// from. Snapping the way the walk is headed, rather than always forward,
    /// is what keeps a leftward motion from ever moving the caret right.
    pub fn stop_at_or_before(&self, off: usize) -> Option<usize> {
        let i = self.stops.partition_point(|&s| s <= off);
        i.checked_sub(1).map(|i| self.stops[i])
    }

    /// Whether the caret may rest at `off` — the invariant every motion in this
    /// view has to leave standing. A glyph stop, a row's end, or a hidden
    /// mark's content end ([`mark_ends`](Self::mark_ends)).
    pub fn is_stop(&self, off: usize) -> bool {
        self.stops.binary_search(&off).is_ok() || self.mark_ends.binary_search(&off).is_ok()
    }

    /// The visible text a caret crosses walking rightward from `from` up to
    /// (but not including) `to` — `UITextInput.text(in:)`'s `[from, to)` in
    /// *this* view. A hidden inline-mark delimiter (`**`, `` ` ``, `_`, an
    /// escape backslash) never got a glyph in the first place — see
    /// [`push_text`]/[`synth`] — so it contributes nothing; what's left is
    /// what's drawn on screen for that span, one character per caret stop.
    ///
    /// **Exactly one character per stop** is the contract, and it is the
    /// system text input's, not a nicety: `UITextInput`'s tokenizer reads a
    /// window of this text around a tap, indexes into it by the integer
    /// `offset(from:to:)` reports (`distance_offset` in `leaf-ffi`, a count of
    /// [`stop_after`](Self::stop_after) hops), finds a word boundary at some
    /// character index, and hands the delta back through
    /// `position(from:offset:)`, which hops stops again. If the text ever
    /// spends a character on something that is not a stop, or a stop on
    /// nothing, every index past that point is off by one and the word the
    /// reader double-tapped comes back shifted — into the header row of a
    /// table, or one letter short. So a stop that draws a glyph is spelled
    /// as that glyph, and a stop that draws none is spelled `'\n'`:
    ///
    /// - a row's own end stop ([`VRow::end_src`]) — the caret home past a
    ///   paragraph's, heading's, list item's, or code line's last glyph. This
    ///   is also what keeps two blocks' words apart: without it the last word
    ///   of one paragraph and the first of the next read as one run of
    ///   letters (`"…edb\n\nhello\n"` came back as `"edbhello"`), and the
    ///   tokenizer selected across the boundary. A list item's end is a
    ///   row end like any other, though no blank gap row follows it.
    /// - a table cell's end, which [`push_table_row`] draws as the gutter
    ///   space before the next `│` so the caret has somewhere to stand past
    ///   the cell's last character. To a reader of *this* text a cell ends a
    ///   line: spelled as a space, a touch surface that lands a tap at a
    ///   word's end past the space that follows it stepped into the next
    ///   cell — or the next row, from the last column.
    ///
    /// A hidden mark's content end ([`mark_ends`](Self::mark_ends)) is a place
    /// the caret rests but not a stop the walks above count, so it has no
    /// character here either; `from` is snapped to the glyph stop drawn at
    /// the same spot first, exactly as [`snap_to_glyph_stop`] does for those
    /// walks. `to` is left as given, so a stop landing exactly on it is still
    /// excluded — the same half-open range `distance_offset`'s loop counts.
    ///
    /// [`push_table_row`]: Builder::push_table_row
    /// [`snap_to_glyph_stop`]: Self::snap_to_glyph_stop
    pub fn visible_text(&self, from: usize, to: usize) -> String {
        self.visible_items(from, to)
            .into_iter()
            .map(|(_, ch)| ch.unwrap_or('\n'))
            .collect()
    }

    /// The UTF-16 length of `visible_text(from, to)` — what an `NSRange`
    /// location into that text is, without building the string.
    ///
    /// AppKit's `NSTextInputClient` and `NSAccessibility` speak in UTF-16
    /// units of *the text as the system sees it*, which for leaf is the visible
    /// text — delimiters hidden. A frontend reporting its selection to the
    /// system converts each end with this and gets back an index into the
    /// string `visible_text(0, end)` returns, which is exactly what the system
    /// will index into.
    pub fn visible_utf16_len(&self, from: usize, to: usize) -> usize {
        let (lo, hi) = self.visible_span(from, to);
        let utf16 = &self.spelling().utf16;
        utf16[hi] - utf16[lo]
    }

    /// The inverse of `visible_utf16_len(0, ·)`: the source offset of the
    /// visible character a UTF-16 index into `visible_text(0, to)` lands on.
    ///
    /// An index inside a surrogate pair resolves to the character that owns
    /// it; one at or past the end of the text returns `None`, so a caller can
    /// substitute the document's end stop. The `\n` a row's or a cell's end
    /// is spelled with resolves to that end stop — a caret home, so a caller
    /// placing a caret there needs no snap.
    pub fn offset_at_visible_utf16(&self, to: usize, index: usize) -> Option<usize> {
        let (lo, hi) = self.visible_span(0, to);
        let utf16 = &self.spelling().utf16;
        // The stop whose text ends past the index is the one it lands on; a
        // stop whose text ends at or before it lies wholly before it.
        let target = utf16[lo] + index;
        let i = lo + utf16[lo + 1..=hi].partition_point(|&end| end <= target);
        (i < hi).then(|| self.stops[i])
    }

    /// The items `visible_text` spells, in order — one per caret stop in
    /// `[from, to)`, keyed by the stop's source offset: the glyph it draws
    /// (`Some`), or `None` for a stop with no character of its own, which the
    /// text spells `'\n'`. See [`visible_text`](Self::visible_text) for which
    /// stops those are and why.
    fn visible_items(&self, from: usize, to: usize) -> Vec<(usize, Option<char>)> {
        let (lo, hi) = self.visible_span(from, to);
        self.stops[lo..hi]
            .iter()
            .copied()
            .zip(self.spelling().chars[lo..hi].iter().copied())
            .collect()
    }

    /// The stops `visible_text(from, to)` spells, as a range into the stop
    /// table: from the glyph stop `from` snaps to, up to but excluding the
    /// first stop at or past `to`.
    fn visible_span(&self, from: usize, to: usize) -> (usize, usize) {
        let from = self.snap_to_glyph_stop(from);
        let lo = self.stops.partition_point(|&s| s < from);
        // The document's last stop is the end of the text, not a character in
        // it: `distance_offset` has no hop past it to pair one with.
        let last = self.stops.len().saturating_sub(1);
        let hi = self.stops.partition_point(|&s| s < to).min(last).max(lo);
        (lo, hi)
    }

    /// The [`Spelling`] of this map's stops, tabulated on first use.
    fn spelling(&self) -> &Spelling {
        self.spelling.get_or_init(|| {
            // The glyph each stop draws — the first at its offset in row
            // order, since a media row's label glyphs all share the media's
            // offset and a wrapped line's end is the next line's first glyph.
            // Row order only follows source order outside a table's wrapped
            // cells (see `pos_of_offset`), which is why each glyph looks its
            // stop up rather than the two being walked side by side.
            //
            // Within a row offsets ascend, so a cursor into the stop table
            // walks forward with the row's glyphs and the whole pass is a
            // read of the glyphs plus one search per row — a keystroke's
            // first conversion pays for this, so it is kept to that. The
            // search is repeated only if a row's offsets step back, which
            // none does; the walk is correct either way.
            let stops = &self.stops;
            let mut chars: Vec<Option<char>> = vec![None; stops.len()];
            for row in self.rows.iter().filter(|r| !r.decoration) {
                let mut i = 0;
                let mut prev = usize::MAX;
                for g in row.glyphs.iter().filter(|g| g.stop) {
                    if prev == usize::MAX || g.src < prev {
                        i = stops.partition_point(|&s| s < g.src);
                    } else {
                        while i < stops.len() && stops[i] < g.src {
                            i += 1;
                        }
                    }
                    prev = g.src;
                    if i < stops.len() && stops[i] == g.src && chars[i].is_none() {
                        chars[i] = Some(g.ch);
                    }
                }
            }
            // A cell's end stop has a glyph (the gutter space) but is spelled
            // as a line end; the structural grid is where the cells' offsets
            // live.
            for cell in self
                .tables
                .iter()
                .flat_map(|t| t.grid.iter())
                .flat_map(|r| r.cells.iter())
            {
                if let Ok(i) = self.stops.binary_search(&cell.end) {
                    chars[i] = None;
                }
            }
            let mut utf16 = Vec::with_capacity(chars.len() + 1);
            utf16.push(0);
            for ch in &chars {
                let before = *utf16.last().unwrap_or(&0);
                utf16.push(before + ch.map_or(1, char::len_utf16));
            }
            Spelling { chars, utf16 }
        })
    }
}

/// Collect the caret stops of a laid-out grid: every stop glyph's offset plus
/// every row's end, ascending and deduplicated. Duplicates are the norm rather
/// than the exception — a wrapped line's end is the same offset as the next
/// line's first glyph — and collapsing them is what makes one press of Left or
/// Right cross exactly one stop.
fn collect_stops(rows: &[VRow]) -> Vec<usize> {
    let mut stops: Vec<usize> = rows
        .iter()
        .filter(|r| !r.decoration)
        .flat_map(|r| {
            r.glyphs
                .iter()
                .filter(|g| g.stop)
                .map(|g| g.src)
                .chain(std::iter::once(r.end_src))
        })
        .collect();
    stops.sort_unstable();
    stops.dedup();
    stops
}

/// Collect every row's [`VRow::mark_ends`] into one ascending, deduplicated
/// table — the peer of [`collect_stops`] for the caret's second home at the
/// end of a hidden mark. A mark that closes at a row's end coincides with the
/// row's own end stop; that offset is in both tables, and harmlessly so.
fn collect_mark_ends(rows: &[VRow]) -> Vec<usize> {
    let mut ends: Vec<usize> = rows
        .iter()
        .filter(|r| !r.decoration)
        .flat_map(|r| r.mark_ends.iter().copied())
        .collect();
    ends.sort_unstable();
    ends.dedup();
    ends
}

/// Group the rows tagged [`VRow::code`] into one [`CodeBlockInfo`] per maximal
/// run — the block-level view a frontend needs to box and scroll each code
/// block. Two code blocks are always parted by the blank separator row a block
/// boundary is spelled with (never itself a code row), so a contiguous run is
/// exactly one block. Derived from the final rows rather than tracked through
/// the builder so it comes out right no matter how [`build_cached`] and
/// [`build_spliced`] shuffle rows around.
fn code_block_spans(rows: &[VRow]) -> Vec<CodeBlockInfo> {
    let mut blocks = Vec::new();
    let mut start: Option<usize> = None;
    for (i, row) in rows.iter().enumerate() {
        match (row.code, start) {
            (true, None) => start = Some(i),
            (false, Some(s)) => {
                blocks.push(CodeBlockInfo {
                    rows_span: s..i,
                    lang: rows[s].code_lang.clone(),
                });
                start = None;
            }
            _ => {}
        }
    }
    if let Some(s) = start {
        blocks.push(CodeBlockInfo {
            rows_span: s..rows.len(),
            lang: rows[s].code_lang.clone(),
        });
    }
    blocks
}

/// The value of `node`'s `key` attribute, if it carries one *with* a value. A
/// bare attribute (`controls`, `muted`) has a `None` value and so reads as
/// absent here — a caller wanting presence-not-value tests the list directly.
/// Shared by the media element and `<source>` readers.
fn attr_of(node: &FlatNode, key: &str) -> Option<String> {
    node.attrs
        .iter()
        .find(|(k, _)| k == key)
        .and_then(|(_, v)| v.clone())
}

/// Collect one [`MediaInfo`] per row carrying a [`VRow::media`] mark — the
/// block-level view a frontend needs to replace each placeholder row with a real
/// picture. The mark rides the block's *first* row and names how many rows the
/// media reserves ([`MediaMark::rows`]); the rows below it are blank
/// [`decoration`](VRow::decoration) fillers that hold the vertical space and no
/// caret. So the span runs from the marked row across those fillers. Derived from
/// the final rows rather than tracked through the builder so it survives however
/// [`build_cached`] and [`build_spliced`] shuffle rows around.
fn media_spans(rows: &[VRow]) -> Vec<MediaInfo> {
    rows.iter()
        .enumerate()
        .filter_map(|(i, row)| {
            row.media.as_ref().map(|m| MediaInfo {
                rows_span: i..i + m.rows.max(1),
                kind: m.kind,
                destination: m.destination.clone(),
                sources: m.sources.clone(),
                alt: m.alt.clone(),
                poster: m.poster.clone(),
            })
        })
        .collect()
}

/// Re-label the drawn block boundaries either side of a block-level media
/// placeholder, so the pair a frontend spaces by names the picture.
///
/// [`BlockClass::from_node_kind`] classifies the node the walk is standing on,
/// and a block image is never a node of its own: [`Builder::media_only`] promotes
/// its *wrapper* — a `paragraph`, or the `container` a `<video>`/`<audio>`
/// arrives as — so the gap above a picture reported [`BlockClass::Paragraph`] and
/// the gap above a movie reported [`BlockClass::Directive`], the class a frontend
/// paints a tinted panel for. [`BlockClass::Media`] was unreachable in
/// consequence: the vocabulary named a kind no frontend could ever be told about.
///
/// Done as a pass over the finished rows rather than inside the walk because
/// only the rows know. The incremental top-level walk carries no node arena at
/// all (`nodes: &[]`) and can classify by kind alone, so teaching the wrapper
/// promotion to the whole-arena walk would label the full and incremental builds
/// differently — the exact drift that walk's own comment forbids. Both builds
/// emit the same [`VRow::media`] marks, so both reach the same answer here. The
/// [`media_spans`] / [`code_block_spans`] pattern.
///
/// One gap can be spelled with several rows — [`Builder::emit_separators_before`]
/// draws the row that closes the block above and the row that opens the block
/// below, with any extra blank source lines navigable between them — and gives
/// every one of them the same [`Boundary`]. So the walk crosses those navigable
/// blanks and relabels the whole run, stopping at the first row that is neither.
fn label_media_boundaries(rows: &mut [VRow]) {
    // A display formula's placeholder is promoted from its paragraph exactly
    // as a picture is, and its gaps are relabelled the same way, as `Math`.
    let spans: Vec<(Range<usize>, BlockClass)> = rows
        .iter()
        .enumerate()
        .filter_map(|(i, row)| {
            if let Some(m) = &row.media {
                Some((i..i + m.rows.max(1), BlockClass::Media))
            } else {
                row.math
                    .iter()
                    .find(|m| m.glyph.is_none())
                    .map(|m| (i..i + m.rows.max(1), BlockClass::Math))
            }
        })
        .collect();
    // A row inside one gap: a drawn boundary to relabel, or one of the navigable
    // blank lines sitting between two drawn ones. Anything else ends the run.
    fn in_gap(row: &VRow) -> bool {
        row.boundary.is_some() || (!row.decoration && row.glyphs.is_empty())
    }
    for (span, class) in spans {
        for i in (0..span.start).rev() {
            if !in_gap(&rows[i]) {
                break;
            }
            if let Some(b) = rows[i].boundary.as_mut() {
                b.below = class;
            }
        }
        for row in rows.iter_mut().skip(span.end) {
            if !in_gap(row) {
                break;
            }
            if let Some(b) = row.boundary.as_mut() {
                b.above = class;
            }
        }
    }
}

/// Collect one [`DirectiveInfo`] per row carrying a [`VRow::leaf_directive`]
/// mark — the block-level view a frontend needs to replace each placeholder row
/// with whatever the directive means to it. The peer of [`media_spans`], derived
/// from the final rows for the same reason: it survives however [`build_cached`]
/// and [`build_spliced`] shuffle rows around.
fn directive_spans(rows: &[VRow]) -> Vec<DirectiveInfo> {
    rows.iter()
        .enumerate()
        .filter_map(|(i, row)| {
            row.leaf_directive.as_ref().map(|m| DirectiveInfo {
                rows_span: i..i + m.rows.max(1),
                name: m.name.clone(),
                attrs: m.attrs.clone(),
                label: m.label.clone(),
            })
        })
        .collect()
}

/// Collect one [`MathInfo`] per [`VRow::math`] mark — the peer of
/// [`media_spans`] and [`directive_spans`], derived from the final rows for the
/// same reason. A block mark spans its fillers; an atom spans its own row.
fn math_spans(rows: &[VRow]) -> Vec<MathInfo> {
    rows.iter()
        .enumerate()
        .flat_map(|(i, row)| {
            row.math.iter().map(move |m| {
                let src = match m.glyph {
                    Some(g) => row.glyphs.get(g).map_or(row.end_src, |g| g.src),
                    None => row
                        .glyphs
                        .iter()
                        .find(|g| g.style.role == Role::Math)
                        .map_or(row.end_src, |g| g.src),
                };
                MathInfo {
                    rows_span: i..i + m.rows.max(1),
                    row: i,
                    glyph: m.glyph,
                    tex: m.tex.clone(),
                    display: m.display,
                    src,
                }
            })
        })
        .collect()
}

/// The source range of a fenced code block's info string — everything on the
/// opening line past the fence (`` ```rust `` → the `rust`). `block_start` is the
/// code block node's `span.start`. `None` for an indented code block, which
/// opens with no fence to carry one. The range is empty for a fence written
/// bare (`` ``` `` alone), which is exactly where a language would be inserted.
///
/// A block inside a quote or a list item starts at its line's start, with the
/// container's marker in front of the fence — `> ```rust`, `- ```rust` — so
/// the marker is skipped first, and the fence's three-space indent allowance
/// is measured from where the container's content begins: past the marker on
/// the fence's own line, or, on a continuation line, from the content column
/// of the list item above. That is what keeps an *indented* code block inside
/// a list item answering `None`.
///
/// Shared by the WYSIWYG builder (to label the box) and [`crate::Doc`] (to edit
/// the label through a prompt), so the two agree on where the language lives.
pub fn code_info_span(source: &str, block_start: usize) -> Option<Range<usize>> {
    let rest = source.get(block_start..)?;
    let line_len = rest.find('\n').unwrap_or(rest.len());
    let line = &rest[..line_len];
    let (depth, quoted) = strip_quote_markers(line);
    let mut content = quoted;
    let mut listed = false;
    while let Some(n) = list_marker_len(&line[content..], 3) {
        content += n;
        listed = true;
    }
    let lead = leading_spaces(&line[content..]);
    // No marker of its own: a continuation line, whose leading spaces include
    // the enclosing item's indent. Only when the block owns its line — a span
    // that starts mid-line has had its container prefix measured off already.
    let at_line_start = block_start == 0 || source.as_bytes()[block_start - 1] == b'\n';
    let base = if listed || !at_line_start {
        0
    } else {
        item_content_column(&source[..block_start], depth, lead)
    };
    // A fence may be indented up to three spaces; past that it opens with a run
    // of the same fence character.
    if lead - base > 3 {
        return None;
    }
    let at = content + lead;
    let fence = line[at..].chars().next()?;
    if fence != '`' && fence != '~' {
        return None; // an indented block, not a fenced one
    }
    let fence_len = line[at..].chars().take_while(|&c| c == fence).count();
    let info_start = block_start + at + fence_len;
    Some(info_start..block_start + line_len)
}

/// The spaces a line opens with.
fn leading_spaces(s: &str) -> usize {
    s.bytes().take_while(|&b| b == b' ').count()
}

/// A line's block-quote markers — each `>` behind up to three spaces, with the
/// one space after it — as the quote depth and the byte offset past them.
fn strip_quote_markers(line: &str) -> (usize, usize) {
    let b = line.as_bytes();
    let (mut depth, mut i) = (0, 0);
    loop {
        let j = i + leading_spaces(&line[i..]);
        if j - i > 3 || b.get(j) != Some(&b'>') {
            return (depth, i);
        }
        depth += 1;
        i = j + 1 + usize::from(b.get(j + 1) == Some(&b' '));
    }
}

/// The length of a list marker opening `s` — its indent (at most `max_lead`
/// spaces), a bullet or an ordinal, and the one space the item's content
/// starts after — or `None` when `s` opens with none.
fn list_marker_len(s: &str, max_lead: usize) -> Option<usize> {
    let b = s.as_bytes();
    let lead = leading_spaces(s);
    if lead > max_lead {
        return None;
    }
    let mark = match b.get(lead)? {
        b'-' | b'*' | b'+' => 1,
        c if c.is_ascii_digit() => {
            let digits = b[lead..].iter().take_while(|c| c.is_ascii_digit()).count();
            if digits > 9 || !matches!(b.get(lead + digits), Some(b'.' | b')')) {
                return None;
            }
            digits + 1
        }
        _ => return None,
    };
    match b.get(lead + mark) {
        Some(b' ') => Some(lead + mark + 1),
        None => Some(lead + mark),
        _ => None,
    }
}

/// The content column of the list item a continuation line indented `lead`
/// spaces (past `depth` quote markers) sits in: the nearest item above whose
/// content starts at or before `lead`. `0` when a line at the margin, or a
/// change of quote depth, says there is no such item.
fn item_content_column(before: &str, depth: usize, lead: usize) -> usize {
    for line in before.strip_suffix('\n').unwrap_or(before).rsplit('\n') {
        let (d, q) = strip_quote_markers(line);
        let s = &line[q..];
        if s.trim().is_empty() {
            continue;
        }
        if d != depth {
            return 0;
        }
        let mut col = 0;
        while let Some(n) = list_marker_len(&s[col..], usize::MAX) {
            col += n;
        }
        if col > 0 && col <= lead {
            return col;
        }
        if col == 0 && leading_spaces(s) == 0 {
            return 0;
        }
    }
    0
}

/// A fenced code block's language for display: its info string, trimmed, or
/// `None` when there's no fence or the fence carries no language. The trimmed
/// text is what a frontend labels the box with; [`code_info_span`] is what an
/// edit replaces.
pub fn code_language(source: &str, block_start: usize) -> Option<String> {
    let span = code_info_span(source, block_start)?;
    let text = source.get(span)?.trim();
    (!text.is_empty()).then(|| text.to_string())
}

/// A horizontal rule's dash count when the map isn't wrapping to a column grid
/// (the GUI, which wraps at pixel width): a fixed, sane width the frontend can
/// paint or re-wrap, instead of a runaway count from an unbounded wrap width.
const UNWRAPPED_RULE_WIDTH: usize = 40;

/// The one glyph an inline formula renders to on a surface that paints
/// pictures in a line — what a plain surface handed such a map would show.
/// One column wide, so a column-wrapped build's arithmetic still holds; the
/// frontend that asked for atoms draws the picture as wide as it is.
const MATH_ATOM: char = '∑';

/// What the surface a map is built for can paint, and how tall its pictures
/// came out — the things about a frontend that change the rows core lays
/// down, gathered from the frontend by [`crate::Doc`] and threaded into every
/// build. The defaults are the plain surface: nothing painted in a line, every
/// picture a one-row placeholder, which is what every test that passes
/// `Surface::default()` gets and what every build got before there was one.
#[derive(Clone, Debug, Default, PartialEq, Eq)]
pub struct Surface {
    /// How many visual rows each block image reserves, keyed by its
    /// destination — the frontend's per-image height, set through
    /// [`crate::Doc::set_media_rows`] so [`Builder::block_media`] can size the
    /// placeholder without core doing any I/O. A destination absent from the
    /// map (or a `0`/`1` entry) reserves the bare one-row placeholder.
    pub media_rows: HashMap<String, usize>,
    /// The same for each display formula, keyed by its TeX verbatim (the
    /// [`MathMark::tex`] a frontend was handed) — set through
    /// [`crate::Doc::set_math_rows`] once the frontend has typeset and
    /// measured the picture.
    pub math_rows: HashMap<String, usize>,
    /// The same for each leaf directive, keyed by its [`DirectiveKey`] — set
    /// through [`crate::Doc::set_directive_rows`] by a frontend that draws a
    /// host's lines for the directive over reserved rows, as a terminal does.
    pub directive_rows: HashMap<DirectiveKey, usize>,
    /// Whether the surface can paint a picture *inside* a line of text, so an
    /// inline formula may render to one atom glyph it draws over
    /// ([`MathInfo`]). A pixel-laid-out frontend says yes; a terminal cannot
    /// composite an image over one cell, and leaves this off to keep an inline
    /// formula as the code-styled TeX it always showed. Off by default.
    pub inline_pictures: bool,
}

/// The one source line rendering its markup raw — the caret's, when the mode
/// or the content asks for it — and how much of the markup that means. See
/// [`crate::Doc::reveal_line`], which decides both.
///
/// Two grades because two things ask. Under
/// [`MarkupMode::Full`](crate::MarkupMode::Full) *every* delimiter on the line
/// comes back (`markup: true`). In the two hidden modes a formula still
/// reveals — its content is its TeX and not its picture, so hiding is not
/// enough — and the line is threaded through with `markup: false`: a math node
/// meeting it shows its source, and an emphasis meeting it hides its
/// asterisks as it always did.
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct Reveal {
    /// The source byte range of the line, newline excluded — empty but present
    /// on a blank line.
    pub line: Range<usize>,
    /// Whether ordinary inline markup reveals on it too, or only math.
    pub markup: bool,
}

impl Reveal {
    /// The caret's line under `MarkupMode::Full`: everything on it reveals.
    pub fn full(line: Range<usize>) -> Self {
        Reveal { line, markup: true }
    }

    /// The caret's line in a hidden mode, where only a formula reveals.
    pub fn math(line: Range<usize>) -> Self {
        Reveal {
            line,
            markup: false,
        }
    }

    /// Whether `span` meets this line — the test every arm that reveals runs.
    ///
    /// Touching at an endpoint counts: an emphasis ending exactly where the
    /// line does is on that line, and a zero-length reveal range (the caret
    /// alone on a blank line) still meets a node that starts there. The test
    /// is deliberately generous — the failure it avoids is revealing one
    /// delimiter of a pair while hiding the other, which looks like corruption
    /// rather than like markup.
    fn meets(&self, span: &Range<usize>) -> bool {
        span.start <= self.line.end && self.line.start <= span.end
    }
}

/// Render the document to a [`VisualMap`]. `wrap` is the column budget for
/// word-wrapping (`Some` for the monospace TUI), or `None` to emit one row per
/// block — the GUI does its own proportional pixel wrapping over these rows.
/// Text and offsets come from the AST (`str` nodes carry the verbatim source
/// slice and an exact span), so the original source string isn't needed here.
pub fn build(
    nodes: &[FlatNode],
    source: &str,
    wrap: Option<usize>,
    preserve_soft: bool,
    surface: &Surface,
    reveal: Option<Reveal>,
) -> VisualMap {
    let Some(doc) = nodes.iter().position(|n| n.kind == Kind::Doc) else {
        return VisualMap::default();
    };
    let top = top_level(nodes, doc);
    let mut b = Builder {
        nodes,
        source,
        wrap: wrap.map(|w| w.max(8)),
        rows: Vec::new(),
        tables: Vec::new(),
        last_off: 0,
        stepped_over: 0,
        surface,
        break_glyph: Cell::new(' '),
        preserve_soft,
        reveal: reveal.clone(),
        pending_mark_ends: RefCell::new(Vec::new()),
        pending_math: RefCell::new(Vec::new()),
        saw_math: Cell::new(false),
        presentation: Presentation::default(),
        faces: RefCell::new(FaceTable::default()),
    };
    // The hidden frontmatter's end is the baseline for the leading and trailing
    // blank rows and for the caret floor — see [`hidden_prefix_end`].
    // `top_level` has already dropped every `metadata` child, so read it off
    // the arena.
    let hidden_end = hidden_prefix_end(source, metadata_end_of(nodes, doc));
    let last_drawn = b.top_blocks(&top, hidden_end);
    b.emit_trailing_blank_lines(last_drawn.unwrap_or(BlockClass::Paragraph), hidden_end);
    let content_start = floor_of(
        &b.rows,
        top.first().map_or(hidden_end, |&i| nodes[i].span.start),
    );
    let stops = collect_stops(&b.rows);
    let mark_ends = collect_mark_ends(&b.rows);
    label_media_boundaries(&mut b.rows);
    let code_blocks = code_block_spans(&b.rows);
    let media = media_spans(&b.rows);
    let directives = directive_spans(&b.rows);
    let math = math_spans(&b.rows);
    VisualMap {
        rows: b.rows,
        content_start,
        stops,
        mark_ends,
        tables: b.tables,
        code_blocks,
        media,
        directives,
        math,
        faces: b.faces.into_inner(),
        spelling: OnceLock::new(),
    }
}

/// Like [`build`], but reuses a persistent [`BlockCache`] so an edit re-renders
/// only the top-level blocks whose source bytes changed *and* marshals only
/// those blocks from twig instead of the whole arena.
///
/// `top` is the document's top-level blocks — twig's `child_spans` of the doc
/// root: `(node_id, kind, span)` for each, in order. `fetch_subtree(node_id)`
/// marshals one block's subtree (local-indexed, root at 0) and is called *only*
/// for a block that missed the cache, i.e. one that actually changed. So a
/// keystroke marshals one small subtree, not ~20k nodes. The result is
/// byte-for-byte identical to [`build`] on the same document (the
/// `build_cached_matches_build` test pins this); [`build`] stays the cache-free,
/// whole-arena reference. This is the entry point [`crate::Doc`] uses.
// One builder, and every one of these is a distinct input to the same layout
// pass — a struct of them would be built at the one call site and unpacked
// here, which is the same arguments with an extra name in the way.
#[allow(clippy::too_many_arguments)]
pub fn build_cached(
    top: &[QueryMatch],
    source: &str,
    wrap: Option<usize>,
    preserve_soft: bool,
    surface: &Surface,
    reveal: Option<Reveal>,
    cache: &mut BlockCache,
    mut fetch_subtree: impl FnMut(u32) -> Vec<FlatNode>,
) -> VisualMap {
    let wrap = wrap.map(|w| w.max(8));

    // Wrapping is a function of the width, so a width change makes every cached
    // row's wrap wrong: start the cache over.
    if cache.wrap != Some(wrap) {
        cache.entries.clear();
        cache.wrap = Some(wrap);
    }
    cache.generation = cache.generation.wrapping_add(1);

    // Frontmatter (a leading `metadata` block) is document metadata, not prose:
    // hidden in the rich view exactly as [`Builder::blocks`] skips it.
    let blocks: Vec<&QueryMatch> = top.iter().filter(|m| m.kind != Kind::Metadata).collect();

    // The outer builder only accumulates rows/tables and spells block boundaries
    // — both a function of the source and `last_off`, never of a node array — so
    // it carries an empty `nodes`. Each changed block is rendered by a *fresh*
    // builder over that block's subtree.
    let mut b = Builder {
        nodes: &[],
        source,
        wrap,
        rows: Vec::new(),
        tables: Vec::new(),
        last_off: 0,
        stepped_over: 0,
        surface,
        break_glyph: Cell::new(' '),
        preserve_soft,
        reveal: reveal.clone(),
        pending_mark_ends: RefCell::new(Vec::new()),
        pending_math: RefCell::new(Vec::new()),
        saw_math: Cell::new(false),
        presentation: Presentation::default(),
        faces: RefCell::new(FaceTable::default()),
    };

    // Record the per-block row decomposition as we go, so a later
    // [`build_spliced`] can patch one block without rebuilding the map.
    let mut layout_blocks: Vec<BlockLayout> = Vec::with_capacity(blocks.len());
    // The document's face table, assembled block by block: from the walk on a
    // miss, from what the entry stored on a hit. The outer builder walks no
    // attributes of its own (it spells boundaries), so it never adds to it.
    let mut faces = FaceTable::default();
    let mut all_shift_safe = true;
    // The class of the last block that drew anything: what the next separator
    // closes, and what the trailing blank lines close at the end. A hidden block
    // (a comment) never becomes it — see [`Builder::block_or_hidden`], whose
    // step-over this loop repeats for the incremental walk.
    let mut above: Option<BlockClass> = None;
    let hidden_end = hidden_prefix_end(
        source,
        top.iter()
            .filter(|m| m.kind == Kind::Metadata)
            .map(|m| m.span.end)
            .next_back(),
    );
    for block in &blocks {
        let start = block.span.start;
        let before_sep = b.rows.len();
        // This walker has no node arena at all (see the `nodes: &[]` above),
        // but a top-level query match carries its kind — the same string
        // `BlockClass::from_node_kind` classifies for the whole-arena walk, so
        // the incremental and full builds label a boundary identically.
        let below = BlockClass::from_node_kind(&block.kind);
        if let Some(above) = above {
            b.emit_separators_before(start, &[], true, Boundary { above, below });
        } else {
            let from = b.leading_from(hidden_end);
            b.emit_leading_blank_lines(from, start, below);
        }
        let after_sep = b.rows.len();
        let bytes = block_bytes(source, &block.span);
        let hash = block_hash(bytes);
        // How this block meets the reveal line, if at all — part of its cache
        // key, since the same bytes render differently on the caret's line.
        let rkey = reveal_key(&reveal, &block.span);

        // Hit: clone the block's rows shifted to its current offset and restore
        // the (shifted) `last_off` so the next separator lands right — no marshal.
        // Only shift-safe blocks are ever cached, so a hit is safe by construction.
        let mut has_math = false;
        if let Some(hit) = cache.reuse(hash, bytes, &rkey) {
            faces.merge(&hit.faces);
            has_math = hit.has_math;
            let delta = start as isize - hit.built_start as isize;
            for row in &hit.rows {
                b.rows.push(shift_row(row, delta));
            }
            b.last_off = (hit.last_off as isize + delta) as usize;
            // `0` is a block that stepped over nothing, and no answer to shift.
            if hit.stepped_over > 0 {
                let stepped = (hit.stepped_over as isize + delta) as usize;
                b.stepped_over = b.stepped_over.max(stepped);
            }
        } else {
            // Miss: marshal just this block's subtree and render it. A subtree is
            // self-contained with local ids (root at 0) and absolute spans, so a
            // fresh builder over it produces the same rows the whole-arena path
            // would. An empty subtree (twig couldn't hand it back) renders nothing.
            let subtree = fetch_subtree(block.node_id);
            if !subtree.is_empty() {
                let mut sub = Builder {
                    nodes: &subtree,
                    source,
                    wrap,
                    rows: Vec::new(),
                    tables: Vec::new(),
                    last_off: 0,
                    stepped_over: 0,
                    surface,
                    break_glyph: Cell::new(' '),
                    preserve_soft,
                    reveal: reveal.clone(),
                    pending_mark_ends: RefCell::new(Vec::new()),
                    pending_math: RefCell::new(Vec::new()),
                    saw_math: Cell::new(false),
                    presentation: Presentation::default(),
                    faces: RefCell::new(FaceTable::default()),
                };
                sub.block(0, &[], &[]);
                // A block that drew nothing is stepped over, not stood on: its
                // `last_off` is its own end, so the separator after it counts
                // from there. The sub-builder started at 0 and never moved, and
                // 0 is where the next separator would otherwise count from —
                // every line of the document, as a blank row each.
                let last_off = if sub.rows.is_empty() {
                    block.span.end
                } else {
                    sub.last_off
                };
                let stepped_over = sub.stepped_over;
                b.stepped_over = b.stepped_over.max(stepped_over);
                // The names this block's glyph ids stand for. They go into the
                // document's table *and* into the cache entry, because a hit
                // re-emits these rows without walking an attribute again.
                let block_faces = sub.faces.into_inner();
                faces.merge(&block_faces);
                has_math = sub.saw_math.get();
                // Cache only a block that is table-free AND renders inside its own
                // span: those two are the conditions for reuse-by-shift to be
                // correct. A block failing either is re-rendered every build (a
                // fresh render always matches a fresh whole-document build).
                if sub.tables.is_empty() {
                    if rows_within(&sub.rows, &block.span) {
                        cache.store(
                            hash,
                            bytes,
                            start,
                            sub.rows.clone(),
                            last_off,
                            stepped_over,
                            rkey,
                            has_math,
                            block_faces,
                        );
                    }
                    b.rows.extend(sub.rows);
                } else {
                    // A table block is never cached; rebase its row-index
                    // bookkeeping onto the combined row vector and append.
                    let base = b.rows.len();
                    for t in &mut sub.tables {
                        t.rows_span = (t.rows_span.start + base)..(t.rows_span.end + base);
                    }
                    b.rows.extend(sub.rows);
                    b.tables.extend(sub.tables);
                }
                b.last_off = last_off;
            }
        }
        let content_rows = b.rows.len() - after_sep;
        let sep_rows = if content_rows == 0 {
            // Hidden: take back the separator drawn for it, so what stands
            // either side meets across one boundary. Its layout entry stays, at
            // no rows, so the splice arithmetic still counts one entry per block.
            b.rows.truncate(before_sep);
            // A cache hit restored the stored `last_off` above; an empty subtree
            // (twig couldn't hand it back) restored nothing. Either way the walk
            // stands past the block.
            b.last_off = b.last_off.max(block.span.end);
            b.stepped_over = b.stepped_over.max(block.span.end);
            0
        } else {
            above = Some(BlockClass::from_node_kind(&block.kind));
            all_shift_safe &= rows_within(&b.rows[after_sep..], &block.span);
            after_sep - before_sep
        };
        layout_blocks.push(BlockLayout {
            span: block.span.clone(),
            kind: block.kind.clone(),
            sep_rows,
            content_rows,
            has_math,
        });
    }

    let before_trailing = b.rows.len();
    b.emit_trailing_blank_lines(above.unwrap_or(BlockClass::Paragraph), hidden_end);
    let trailing_rows = b.rows.len() - before_trailing;

    // Evict every entry no block reused this build, so the cache tracks the
    // current document instead of growing without bound over a session.
    let g = cache.generation;
    cache.entries.retain(|_, bucket| {
        bucket.retain(|e| e.generation == g);
        !bucket.is_empty()
    });

    cache.layout = Layout {
        blocks: layout_blocks,
        trailing_rows,
        built_len: source.len(),
        has_tables: !b.tables.is_empty(),
        all_shift_safe,
        reveal: reveal.clone(),
    };

    // The first rendered offset is the first non-metadata block's start — the
    // analogue of [`first_content_offset`] for the top-level list. With nothing
    // but frontmatter it's the end of that frontmatter, and 0 for an empty
    // document ([`hidden_prefix_end`]).
    let content_start = floor_of(&b.rows, blocks.first().map_or(hidden_end, |m| m.span.start));
    let stops = collect_stops(&b.rows);
    let mark_ends = collect_mark_ends(&b.rows);
    label_media_boundaries(&mut b.rows);
    let code_blocks = code_block_spans(&b.rows);
    let media = media_spans(&b.rows);
    let directives = directive_spans(&b.rows);
    let math = math_spans(&b.rows);
    VisualMap {
        rows: b.rows,
        content_start,
        stops,
        mark_ends,
        tables: b.tables,
        code_blocks,
        media,
        directives,
        math,
        faces,
        spelling: OnceLock::new(),
    }
}

/// The fast path for a single-block edit: patch the previous [`VisualMap`] in
/// place rather than reassembling it. Returns `Some(new_map)` when it applies,
/// or `None` to tell the caller to fall back to [`build_cached`] (always
/// correct). Consumes `prev` either way — on `None` the caller rebuilds from
/// scratch and doesn't need it.
///
/// It applies only when `dirty` (twig's dirty byte range) falls inside exactly
/// one top-level block AND the block structure around it is unchanged — verified
/// by matching the new `top` list against the previous [`Layout`] block for
/// block: kinds unchanged, spans before the edit identical, spans after it
/// shifted by the byte delta, count unchanged. Any deviation — a block split or
/// merged, a fence opened to swallow later blocks, a table anywhere, a
/// multi-block edit — fails the match and returns `None`. That check is what
/// makes the byte-range trustworthy: twig's dirty range is exact about *bytes*
/// but silent about *reparse*, and the structural match catches the reparse
/// effects it can't see.
///
/// When it applies, the unchanged prefix rows move verbatim, the suffix rows
/// shift by the delta *in place* (integer adds, no glyph copy), and only the one
/// dirty block is re-marshalled and re-rendered; stops splice the same way by
/// offset. So the cost is O(rows after the edit), and nothing before the edit is
/// touched. The hash-keyed entry cache is left alone — a later [`build_cached`]
/// will miss on the changed block, re-render it, and evict the stale entry, so
/// chained splices neither corrupt nor grow it.
// One builder, and every one of these is a distinct input to the same layout
// pass — a struct of them would be built at the one call site and unpacked
// here, which is the same arguments with an extra name in the way.
#[allow(clippy::too_many_arguments)]
pub fn build_spliced(
    prev: VisualMap,
    source: &str,
    wrap: Option<usize>,
    preserve_soft: bool,
    top: &[QueryMatch],
    dirty: Range<usize>,
    surface: &Surface,
    reveal: Option<Reveal>,
    cache: &mut BlockCache,
    mut fetch_subtree: impl FnMut(u32) -> Vec<FlatNode>,
) -> Option<VisualMap> {
    let wrap = wrap.map(|w| w.max(8));
    // A width change invalidates every cached row — a full rebuild's job.
    if cache.wrap != Some(wrap) {
        return None;
    }
    // So does a moved reveal line, and for the same reason: this path reuses
    // every row outside the dirty block, and those rows encode which line was
    // showing its raw markup when they were built. Typing almost always moves
    // the caret, so under `MarkupMode::Full` this bails to `build_cached` on
    // most keystrokes — still block-cached, so only the edited block and the
    // revealed one actually re-render.
    if cache.layout.reveal != reveal {
        return None;
    }
    // Take the previous layout; on any bail below the caller rebuilds it (and the
    // map) via `build_cached`, so leaving it empty is fine. A table or a block
    // that renders outside its span (a degenerate inline span) makes shifting
    // unsound, so those force the full-rebuild path.
    let prev_layout = std::mem::take(&mut cache.layout);
    if prev_layout.built_len == 0 || prev_layout.has_tables || !prev_layout.all_shift_safe {
        return None;
    }
    // The layout addresses `prev` by row index, so it is only usable against the
    // map it was built from. A frontend is free to hold the map it was handed and
    // present it differently — leaf-ratatui splices blank filler rows under an
    // oversized heading so the raster has somewhere to stand — and if one of those
    // comes back here the row arithmetic below lands on the wrong rows: the
    // re-rendered block is laid over a filler and the rows it really occupied
    // survive into the suffix, stranding a stale copy of the edited line and
    // pushing everything after it one row down, once per keystroke. A row count
    // that doesn't match what this layout describes is the tell, and the honest
    // answer is the full rebuild.
    let described_rows = prev_layout
        .blocks
        .iter()
        .map(|pl| pl.sep_rows + pl.content_rows)
        .sum::<usize>()
        + prev_layout.trailing_rows;
    if described_rows != prev.rows.len() {
        return None;
    }

    let blocks: Vec<&QueryMatch> = top.iter().filter(|m| m.kind != Kind::Metadata).collect();
    if blocks.is_empty() || blocks.len() != prev_layout.blocks.len() {
        return None;
    }
    let delta = source.len() as isize - prev_layout.built_len as isize;

    // The single block whose NEW span contains the whole dirty range. A dirty
    // range straddling a block boundary (or a separator) finds none → bail.
    let k = blocks
        .iter()
        .position(|m| m.span.start <= dirty.start && dirty.end <= m.span.end)?;

    // Structural match: every OTHER block is unchanged — same kind throughout,
    // span identical before the edit and shifted by `delta` after it. A mismatch
    // means the reparse reshaped the block structure, which only a full rebuild
    // renders correctly.
    for (i, (m, pl)) in blocks.iter().zip(&prev_layout.blocks).enumerate() {
        if m.kind != pl.kind {
            return None;
        }
        if i == k {
            continue;
        }
        let want = if i < k {
            pl.span.clone()
        } else {
            (pl.span.start as isize + delta) as usize..(pl.span.end as isize + delta) as usize
        };
        if m.span != want {
            return None;
        }
    }
    // The dirty block itself: start unchanged (the edit is inside it, past its
    // start), end moved by exactly the delta.
    let pk_start = prev_layout.blocks[k].span.start;
    let pk_end = prev_layout.blocks[k].span.end;
    let pk_sep = prev_layout.blocks[k].sep_rows;
    let pk_content = prev_layout.blocks[k].content_rows;
    if blocks[k].span.start != pk_start || blocks[k].span.end != (pk_end as isize + delta) as usize
    {
        return None;
    }

    // Re-render the dirty block from its subtree. A table makes the splice
    // bookkeeping unsafe, so bail if one appears.
    let subtree = fetch_subtree(blocks[k].node_id);
    if subtree.is_empty() {
        return None;
    }
    let mut sub = Builder {
        nodes: &subtree,
        source,
        wrap,
        rows: Vec::new(),
        tables: Vec::new(),
        last_off: 0,
        stepped_over: 0,
        surface,
        break_glyph: Cell::new(' '),
        preserve_soft,
        reveal: reveal.clone(),
        pending_mark_ends: RefCell::new(Vec::new()),
        pending_math: RefCell::new(Vec::new()),
        saw_math: Cell::new(false),
        presentation: Presentation::default(),
        faces: RefCell::new(FaceTable::default()),
    };
    sub.block(0, &[], &[]);
    // A table, or content that renders outside the block's span (a degenerate
    // inline span), makes the shift bookkeeping unsound — fall back.
    if !sub.tables.is_empty() || !rows_within(&sub.rows, &blocks[k].span) {
        return None;
    }
    // The face table starts from the previous map's, because every row this
    // path keeps was built against it and its glyphs' ids still mean what they
    // meant. The re-rendered block adds whatever it met. An id the edit took
    // the last glyph of stays in the table, naming nothing — the price of not
    // walking the rows this path exists to avoid walking.
    let mut faces = prev.faces;
    faces.merge(&sub.faces.into_inner());
    let sub_saw_math = sub.saw_math.get();
    let new_content = sub.rows;
    let new_content_len = new_content.len();
    let new_stops = collect_stops(&new_content);
    let new_mark_ends = collect_mark_ends(&new_content);

    // Row span of the dirty block's CONTENT. Its leading separator stays in the
    // prefix: the gap before block k is unchanged, since k's start didn't move.
    let content_start_row: usize = prev_layout.blocks[..k]
        .iter()
        .map(|pl| pl.sep_rows + pl.content_rows)
        .sum::<usize>()
        + pk_sep;
    let content_end_row = content_start_row + pk_content;

    // Splice rows: [prefix | new content | suffix + delta]. The prefix moves
    // untouched; the suffix shifts in place — integer adds, no glyph copy.
    let mut rows = prev.rows;
    let mut suffix = rows.split_off(content_end_row);
    rows.truncate(content_start_row);
    for row in &mut suffix {
        shift_row_in_place(row, delta);
    }
    rows.reserve(new_content_len + suffix.len());
    rows.extend(new_content);
    rows.extend(suffix);

    // Splice stops by offset. The old dirty block covered `[pk_start, pk_end]`:
    // prefix stops fall below it, suffix stops above it (shift by delta), the new
    // content supplies the middle. The three ranges stay disjoint and ascending,
    // so the result needs no re-sort.
    let p1 = prev.stops.partition_point(|&s| s < pk_start);
    let p2 = prev.stops.partition_point(|&s| s <= pk_end);
    let mut stops = Vec::with_capacity(p1 + new_stops.len() + (prev.stops.len() - p2));
    stops.extend_from_slice(&prev.stops[..p1]);
    stops.extend(new_stops);
    for &s in &prev.stops[p2..] {
        stops.push((s as isize + delta) as usize);
    }
    // The mark ends splice the same way: they are offsets in the same
    // coordinates, cut at the same block.
    let m1 = prev.mark_ends.partition_point(|&s| s < pk_start);
    let m2 = prev.mark_ends.partition_point(|&s| s <= pk_end);
    let mut mark_ends = Vec::with_capacity(m1 + new_mark_ends.len() + (prev.mark_ends.len() - m2));
    mark_ends.extend_from_slice(&prev.mark_ends[..m1]);
    mark_ends.extend(new_mark_ends);
    for &s in &prev.mark_ends[m2..] {
        mark_ends.push((s as isize + delta) as usize);
    }

    // Record the patched layout for the next splice: spans move to the new
    // coordinates, and the dirty block takes its new content-row count.
    let mut new_blocks = prev_layout.blocks;
    for (pl, m) in new_blocks.iter_mut().zip(&blocks) {
        pl.span = m.span.clone();
    }
    new_blocks[k].content_rows = new_content_len;
    new_blocks[k].has_math = sub_saw_math;
    cache.layout = Layout {
        blocks: new_blocks,
        trailing_rows: prev_layout.trailing_rows,
        built_len: source.len(),
        has_tables: false,
        // Every prefix/suffix block was shift-safe last build (we bailed
        // otherwise) and the re-rendered block was just checked, so the patched
        // document is still entirely shift-safe.
        all_shift_safe: true,
        reveal,
    };

    label_media_boundaries(&mut rows);
    let code_blocks = code_block_spans(&rows);
    let media = media_spans(&rows);
    let directives = directive_spans(&rows);
    let math = math_spans(&rows);
    Some(VisualMap {
        rows,
        // The edit was inside one block, so no block moved its start and the
        // lines above the first are the ones the floor was taken from.
        content_start: prev.content_start,
        stops,
        mark_ends,
        tables: Vec::new(),
        code_blocks,
        media,
        directives,
        math,
        faces,
        spelling: OnceLock::new(),
    })
}

/// A persistent, content-keyed cache of the rows each top-level block renders
/// to — the [`VisualMap`] analogue of the GUI's ShapedLine cache, one level
/// down. Held by a [`crate::Doc`] and threaded into [`build_cached`], it is what
/// makes a rebuild after a keystroke cost "re-render the edited block + shift
/// the rest" instead of re-rendering the whole document.
///
/// A top-level block's rows are a pure function of its source bytes and the wrap
/// width, so an unchanged block's rows are cloned and their source offsets
/// shifted by the edit's byte delta rather than rebuilt glyph by glyph. Two
/// things make that purity hold: at the top level the render prefix is always
/// empty (nesting prefixes — a quote gutter, a list indent — exist only *inside*
/// a top-level block, within its cached unit), and a block's output never reads
/// the incoming `last_off` (it writes `last_off` from its own content before any
/// nested separator reads it). So the only thing that differs between two
/// positions of an unchanged block is a uniform offset shift. Keyed by a fast
/// hash of the block's bytes with the bytes kept for a verify-on-hit — exactly
/// the shape cache's weak-hash-then-compare, so a collision costs a re-render,
/// never a wrong row.
///
/// Tables are never cached (a block that emits any table row is always rebuilt):
/// their rows are cross-referenced from the map's `tables` side-table by row
/// index, which a blind offset-shift wouldn't fix up, and they are rare enough
/// that the simplicity beats the reuse.
#[derive(Default)]
pub struct BlockCache {
    /// The wrap width every entry was built at; a change invalidates all of
    /// them. `None` before the first build (distinct from `Some(None)`, the
    /// unwrapped GUI width).
    wrap: Option<Option<usize>>,
    /// Bumped once per [`build_cached`]. An entry reused or inserted this build
    /// carries the current value; stale entries are dropped at the end of it.
    generation: u64,
    /// `hash(bytes)` → the block(s) sharing that hash — a bucket because
    /// distinct blocks can collide, while two *identical* blocks share one entry
    /// (free dedup).
    entries: HashMap<u64, Vec<CachedBlock>>,
    /// The row/stop decomposition of the last build, which [`build_spliced`]
    /// patches in place for a single-block edit. Kept in step with whatever
    /// [`VisualMap`] was last produced; empty before the first build.
    layout: Layout,
}

/// How the last build's [`VisualMap`] decomposes into top-level blocks — the
/// bookkeeping [`build_spliced`] needs to splice one block's rows and stops
/// without rebuilding the whole map. Every field describes the *previous* build,
/// in that build's coordinates.
#[derive(Default)]
struct Layout {
    /// One entry per rendered (metadata-filtered) top-level block, in order.
    blocks: Vec<BlockLayout>,
    /// Trailing blank rows past the last block (from `emit_trailing_blank_lines`).
    trailing_rows: usize,
    /// The source length this layout was built at — the reference for the edit's
    /// byte delta.
    built_len: usize,
    /// Whether the last build drew any table. A table's cross-referenced row
    /// indices don't survive a blind splice, so their presence makes
    /// [`build_spliced`] bail to a full rebuild.
    has_tables: bool,
    /// Whether every block rendered strictly inside its own span (see
    /// [`rows_within`]). A block that doesn't — a malformed Markdown inline node
    /// that twig leaves with a degenerate `0..0` span renders at a fixed offset
    /// outside its block — can't be shifted correctly, so its presence makes
    /// [`build_spliced`] bail to a full rebuild.
    all_shift_safe: bool,
    /// The reveal line this layout was built under (see [`Builder::reveal`]).
    /// A splice reuses every row it isn't re-rendering, so a reveal line that
    /// has moved would leave the old line still showing its delimiters and the
    /// new one still hiding them — [`build_spliced`] bails when this changes.
    reveal: Option<Reveal>,
}

/// One top-level block's contribution to the last build: its span and kind (for
/// the structural match that proves only one block changed) and how many
/// separator and content rows it emitted (to locate its slice of the row
/// vector).
struct BlockLayout {
    span: Range<usize>,
    kind: Kind,
    sep_rows: usize,
    content_rows: usize,
    /// Whether the block holds a formula — see [`BlockCache::math_meets`].
    has_math: bool,
}

/// One cached block: the rows it rendered to, plus what a reuse at a new
/// position needs to shift them. Offsets are stored absolute (as built) and
/// shifted by `new_start - built_start` on reuse.
struct CachedBlock {
    /// The block's exact source bytes, compared on a hash hit so a collision
    /// can never hand back another block's rows.
    bytes: Box<[u8]>,
    /// The offset the rows were built at (the block's `span.start`).
    built_start: usize,
    /// The block's rows, offsets absolute as built.
    rows: Vec<VRow>,
    /// `last_off` after this block was emitted, absolute as built — restored
    /// (shifted) on reuse so the following separator lands correctly.
    last_off: usize,
    /// `stepped_over` after this block was emitted, absolute as built — the
    /// hidden tail a block ends with (a `</div>`), restored with `last_off`
    /// so the trailing blank lines are counted from past it on a hit too.
    stepped_over: usize,
    /// Where the reveal line fell *within this block* when the rows were built,
    /// as a block-relative byte range — see [`reveal_key`]. Compared alongside
    /// `bytes` on a hit, because identical source renders to different rows
    /// depending on whether the caret's line is inside it: the same `*em*`
    /// shows its asterisks on the revealed line and hides them everywhere else.
    ///
    /// Block-relative rather than absolute so an unaffected block still hits
    /// after an edit shifts it, and `None` for the overwhelmingly common
    /// no-reveal case — which is why an entry stored under `MarkupMode::None`
    /// keeps hitting for every block that isn't the caret's.
    reveal: Option<Range<usize>>,
    /// Whether the block holds a formula, so a hit can say so to the layout
    /// without walking the rows it is re-emitting.
    has_math: bool,
    /// The named families this block's glyphs are set in — see
    /// [`VisualMap::faces`]. Stored with the rows because a hit re-emits them
    /// without walking a `data-font` again, and the map still has to be able to
    /// say what the id on a reused glyph names. Empty for every block that
    /// names no family, which is nearly all of them.
    faces: FaceTable,
    /// The build that last reused or inserted this entry (see `generation`).
    generation: u64,
}

/// Where `reveal` falls inside a block, in block-relative bytes — the extra key
/// a cached block is stored and matched under.
///
/// `None` when the block doesn't meet the reveal line at all, which is every
/// block on every build in the two hidden modes, and all but one of them under
/// [`crate::MarkupMode::Full`]. So the cache keeps its hit rate as the caret
/// moves: only the line the caret leaves and the line it arrives at re-render.
fn reveal_key(reveal: &Option<Reveal>, span: &Range<usize>) -> Option<Range<usize>> {
    let r = reveal.as_ref()?;
    // The same generous intersection test `Builder::revealed` uses, so a block
    // is keyed as revealed exactly when its glyphs will be built that way.
    // Whether the line reveals markup or only math is not in the key: that is
    // a mode, and a mode change empties the cache.
    r.meets(span).then(|| {
        let start = r.line.start.max(span.start) - span.start;
        let end = r.line.end.min(span.end) - span.start;
        start..end
    })
}

impl BlockCache {
    /// Whether the caret's `line` meets a top-level block that holds a
    /// formula, as of the last build — the question [`crate::Doc::reveal_line`]
    /// asks in the hidden markup modes before it threads the line through, so
    /// that only a line with something to reveal costs a rebuild.
    ///
    /// Answered from the layout rather than from twig because the layout is
    /// free: every build records per block whether its walk met a formula
    /// ([`BlockLayout::has_math`]), and a query against the tree is
    /// O(document) on every frame. Coarse on purpose — a block, not a line —
    /// since a block with a formula in it is rare, small, and the one thing
    /// worth re-rendering as the caret moves through it.
    ///
    /// Exact between edits, when the spans are the document's. Across an edit
    /// the spans are the previous revision's, so the caller checks again once
    /// the new layout is in and rebuilds if the answer changed.
    pub(crate) fn math_meets(&self, line: &Range<usize>) -> bool {
        self.layout
            .blocks
            .iter()
            .any(|b| b.has_math && b.span.start <= line.end && line.start <= b.span.end)
    }

    /// Look up a block by hash, verify its bytes and reveal key, and on a hit
    /// stamp it used this build and hand back a borrow to shift-and-clone from.
    /// `None` on a miss (unknown hash, a collision whose bytes differ, or the
    /// same bytes built under a different reveal).
    fn reuse(
        &mut self,
        hash: u64,
        bytes: &[u8],
        reveal: &Option<Range<usize>>,
    ) -> Option<&CachedBlock> {
        let g = self.generation;
        let bucket = self.entries.get_mut(&hash)?;
        let e = bucket
            .iter_mut()
            .find(|e| &*e.bytes == bytes && &e.reveal == reveal)?;
        e.generation = g;
        Some(&*e)
    }

    /// Cache the rows a freshly-rendered block produced (or refresh an existing
    /// entry for the same bytes and reveal — an identical block elsewhere, or a
    /// re-render).
    #[allow(clippy::too_many_arguments)]
    fn store(
        &mut self,
        hash: u64,
        bytes: &[u8],
        built_start: usize,
        rows: Vec<VRow>,
        last_off: usize,
        stepped_over: usize,
        reveal: Option<Range<usize>>,
        has_math: bool,
        faces: FaceTable,
    ) {
        let g = self.generation;
        let bucket = self.entries.entry(hash).or_default();
        if let Some(e) = bucket
            .iter_mut()
            .find(|e| &*e.bytes == bytes && e.reveal == reveal)
        {
            e.built_start = built_start;
            e.rows = rows;
            e.last_off = last_off;
            e.stepped_over = stepped_over;
            e.has_math = has_math;
            e.faces = faces;
            e.generation = g;
        } else {
            bucket.push(CachedBlock {
                bytes: bytes.into(),
                built_start,
                rows,
                last_off,
                stepped_over,
                reveal,
                has_math,
                faces,
                generation: g,
            });
        }
    }
}

/// The source bytes a top-level block covers — the block cache's key material.
///
/// Clamped to the source rather than sliced by the span as twig gives it,
/// because that span can end *past* the last byte: the final block of a document
/// with no trailing newline is closed on the virtual newline the parser supplies
/// at EOF, so its `span.end` is `source.len() + 1`. Slicing by such a range
/// yields `None`, and the obvious `unwrap_or(&[])` reads that as *this block has
/// no bytes* — the wrong answer twice over.
///
/// Two blocks whose spans both overrun then key alike, and the second is served
/// the first one's rows. That is not hypothetical: a footnote definition is a
/// root beside `doc` merged back into the top level by [`top_blocks`], while the
/// `section` above it spans the definition's bytes too, so both end at EOF —
/// and a document ending in `[^note]: …` renders that definition as a second
/// copy of the heading. Even alone, a block that keeps hashing empty as the user
/// types in it is served the stale rows built before the edit.
///
/// Clamping hands back the bytes the block really covers, which tells both cases
/// apart, and costs nothing for a span that was in range to begin with.
fn block_bytes<'a>(source: &'a str, span: &Range<usize>) -> &'a [u8] {
    let bytes = source.as_bytes();
    let start = span.start.min(bytes.len());
    &bytes[start..span.end.clamp(start, bytes.len())]
}

/// A fast, allocation-free content hash (FNV-1a) for a block's bytes. Weak by
/// design — the bytes are compared on a hit — so its only job is to spread
/// blocks across buckets cheaply. SipHash over every block's bytes on every
/// keystroke would cost more than it saves, the same lesson the shape cache
/// learned when it stopped hashing through the standard hasher.
fn block_hash(bytes: &[u8]) -> u64 {
    let mut h: u64 = 0xcbf2_9ce4_8422_2325;
    for &x in bytes {
        h ^= x as u64;
        h = h.wrapping_mul(0x0000_0100_0000_01b3);
    }
    h
}

/// Clone a cached row with every source offset advanced by `delta` — the whole
/// cost of reusing an unchanged block: integer adds where a rebuild would
/// re-shape every glyph.
fn shift_row(row: &VRow, delta: isize) -> VRow {
    let shift = |off: usize| (off as isize + delta) as usize;
    VRow {
        glyphs: row
            .glyphs
            .iter()
            .map(|g| Glyph {
                ch: g.ch,
                style: g.style,
                src: shift(g.src),
                stop: g.stop,
            })
            .collect(),
        end_src: shift(row.end_src),
        decoration: row.decoration,
        code: row.code,
        code_lang: row.code_lang.clone(),
        directive: row.directive,
        directive_label: row.directive_label.clone(),
        media: row.media.clone(),
        // A tick, not an offset — reuse carries it as-is, like `code_lang`.
        task: row.task,
        leaf_directive: row.leaf_directive.clone(),
        heading: row.heading,
        // Presentation, not offsets: names the author wrote, which a shifted
        // block still wears — like `code_lang`.
        align: row.align,
        line_height: row.line_height,
        // Structure, not offsets: a reused block's rows divide the same blocks
        // wherever the edit above moved them to.
        boundary: row.boundary,
        mark_ends: row.mark_ends.iter().map(|&o| shift(o)).collect(),
        // Strings and a glyph index: the glyph moved, the index into the row
        // did not.
        math: row.math.clone(),
    }
}

/// Advance a row's source offsets by `delta` in place — the suffix half of
/// [`build_spliced`], where the rows are already owned and only need shifting,
/// not copying.
fn shift_row_in_place(row: &mut VRow, delta: isize) {
    for g in &mut row.glyphs {
        g.src = (g.src as isize + delta) as usize;
    }
    row.end_src = (row.end_src as isize + delta) as usize;
    for o in &mut row.mark_ends {
        *o = (*o as isize + delta) as usize;
    }
}

/// Whether every source offset a block's rows carry falls inside the block's own
/// span — the precondition for reusing the block by a uniform offset shift. It
/// holds for well-formed blocks (their glyphs and row ends address bytes within
/// the block, synthetic glyphs point at the block start). It fails when a node
/// renders *outside* its block, which today means a malformed Markdown inline
/// node twig leaves with a degenerate `0..0` span: that content lands at a fixed
/// offset that doesn't move with the block. Such a block is re-rendered every
/// build instead of shifted, so the incremental map still matches a fresh one —
/// see [`build_cached`] and [`build_spliced`].
fn rows_within(rows: &[VRow], span: &Range<usize>) -> bool {
    rows.iter().all(|r| {
        r.end_src >= span.start
            && r.end_src <= span.end
            && r.glyphs
                .iter()
                .all(|g| g.src >= span.start && g.src <= span.end)
    })
}

/// Where the rendered document begins when a leading `metadata` block is all
/// there is — the end of that hidden frontmatter, past the newline that closes
/// its last line so the floor sits at the start of the (empty) body rather than
/// on the closing `---`.
///
/// With a real block after it the frontmatter's end is never needed: the floor
/// is that block's start, and the rows begin there. With nothing after it, both
/// the caret floor and the trailing-blank-line count would otherwise fall back
/// to offset 0 — inside the hidden frontmatter — which put the caret *before*
/// the metadata and made typing land ahead of the opening `---`.
fn hidden_prefix_end(source: &str, meta_end: Option<usize>) -> usize {
    let Some(end) = meta_end else { return 0 };
    let end = end.min(source.len());
    let rest = &source[end..];
    if rest.starts_with("\r\n") {
        end + 2
    } else if rest.starts_with('\n') {
        end + 1
    } else {
        end
    }
}

/// The caret floor: the first block's start, or the first blank line above it
/// when those lines draw rows ([`Builder::emit_leading_blank_lines`]).
fn floor_of(rows: &[VRow], first_start: usize) -> usize {
    rows.first()
        .map_or(first_start, |r| r.end_src.min(first_start))
}

/// The end of the document's hidden frontmatter: the last `metadata` child of
/// `doc`, which is what [`top_level`] and [`Builder::blocks`] drop. `None` when
/// there is none.
fn metadata_end_of(nodes: &[FlatNode], doc: usize) -> Option<usize> {
    let mut end = None;
    let mut child = nodes[doc].first_child;
    while let Some(cid) = child {
        let n = &nodes[cid.0 as usize];
        if n.kind == Kind::Metadata {
            end = Some(n.span.end);
        }
        child = n.next_sibling;
    }
    end
}

/// The document's rendered top-level blocks, as node indices in source order.
///
/// Not simply `doc`'s children, for two reasons. Frontmatter (a leading
/// `metadata` block) is document metadata rather than prose and is dropped, the
/// way [`Builder::blocks`] drops it. And a **footnote definition** (`[^1]: …`)
/// is not a child of `doc` at all: twig parses it as a root of its own, a
/// *sibling* of the document node with `parent == None`. A walk that starts at
/// `doc` therefore never reaches one, which is why a definition — and every
/// byte of its body — used to render as nothing at all. Merging the roots back
/// in by `span.start` puts each definition on screen exactly where it was
/// written, which is what keeps rows, stops, and offsets monotonic.
///
/// A **link reference definition** (`[foo]: /url`) is a root of the same kind,
/// and is merged for the opposite reason: it draws *nothing*, and the walk has
/// to know where it stands to step over it. A definition closing a README —
/// the `[links]: …` block under the prose — left no block over its lines, so
/// the separator logic read them as blank lines and drew an empty paragraph
/// per definition. Merged in, it is a hidden block like a comment, and
/// [`Builder::block_or_hidden`] moves the walk past it. One with no span
/// (`0..0`, what twig before 3.3.3 reported for every one) has nowhere to be
/// merged, and is left out as before.
///
/// Only those roots are merged. twig also leaves stray orphan `str` nodes
/// parented to nothing (the `*` of an emphasis run, for one); those are already
/// rendered as part of the subtree that owns their bytes, and re-emitting them
/// here would double them.
fn top_level(nodes: &[FlatNode], doc: usize) -> Vec<usize> {
    let mut out = Vec::new();
    let mut child = nodes[doc].first_child;
    while let Some(cid) = child {
        let n = &nodes[cid.0 as usize];
        if n.kind != Kind::Metadata {
            out.push(cid.0 as usize);
        }
        child = n.next_sibling;
    }
    out.extend(
        nodes
            .iter()
            .enumerate()
            .filter(|(_, n)| n.parent.is_none() && is_placed_definition(&n.kind, &n.span))
            .map(|(i, _)| i),
    );
    out.sort_by_key(|&i| nodes[i].span.start);
    out
}

/// Is a parentless node of `kind` at `span` a definition the top-level walk
/// merges in — a footnote definition, or a link reference definition that
/// knows where it stands? Shared by [`top_level`] and [`top_blocks`] so the
/// two walks cannot disagree about what the top-level blocks are.
fn is_placed_definition(kind: &Kind, span: &Range<usize>) -> bool {
    match *kind {
        Kind::Footnote => true,
        Kind::Reference => span.end > span.start,
        _ => false,
    }
}

/// The top-level blocks to hand [`build_cached`] / [`build_spliced`] — the
/// incremental path's twin of [`top_level`], which the two must agree with block
/// for block or the render paths diverge.
///
/// `child_spans(None)` gives `doc`'s children, which is all of them for an
/// ordinary document. A **footnote definition** is not one: twig parses `[^1]: …`
/// as a root beside `doc` with no parent, and indexes it at no offset either —
/// `node_at` inside its bytes answers `doc`, and a `query("footnote")` selector
/// finds nothing. Leaf used to discover them by marshalling the whole arena with
/// `nodes()` — the very cost the incremental path exists to avoid — behind a
/// byte-scan gate that gave documents with no `[^…]:` line a substring search
/// instead. twig 3.0's `definitions()` asks the library the question directly,
/// so both the marshal and the gate are gone.
///
/// The link reference definitions `definitions()` also reports are merged on
/// the same terms as [`top_level`] merges them — see [`is_placed_definition`].
///
/// This is the one part of the render that needs an [`Editor`] rather than a
/// marshalled node array. The builders themselves stay editor-free; this only
/// prepares their input.
pub(crate) fn top_blocks(editor: &mut Editor) -> Vec<QueryMatch> {
    let mut top = editor.child_spans(None).unwrap_or_default();
    let defs: Vec<QueryMatch> = definitions(editor)
        .into_iter()
        .filter(|m| is_placed_definition(&m.kind, &m.span))
        .collect();
    if defs.is_empty() {
        return top;
    }
    top.extend(defs);
    // Source order — what every offset-keyed thing downstream (rows, stops, the
    // splice path's block-for-block match) is built to assume.
    top.sort_by_key(|m| m.span.start);
    top
}

/// Every `[^label]: …` definition in the document, in whatever order twig
/// reports them.
///
/// Filtered to [`Kind::Footnote`]: `definitions()` also reports the *link*
/// reference definitions (`[foo]: /url`), which are [`top_blocks`]'s business
/// and not [`crate::Doc::footnote_at_caret`]'s.
///
/// Empty when the document can't be walked, which leaves [`top_blocks`] with
/// the ordinary top-level children and [`crate::Doc::footnote_at_caret`] with an
/// undefined reference — in both cases the same answer as a document that has
/// no definitions, which is the right way to degrade.
pub(crate) fn footnote_definitions(editor: &mut Editor) -> Vec<QueryMatch> {
    definitions(editor)
        .into_iter()
        .filter(|m| m.kind == Kind::Footnote)
        .collect()
}

/// Every definition twig resolves by label rather than by position — footnote
/// and link reference definitions both — or nothing when the document can't be
/// walked.
fn definitions(editor: &mut Editor) -> Vec<QueryMatch> {
    let Ok(mut doc) = editor.document() else {
        return Vec::new();
    };
    doc.definitions().unwrap_or_default()
}

/// The last line of a block whose span ends at `span_end` — a thematic break,
/// a page break — as the two offsets the rest of the crate means by it: where
/// the line's text ends, and the home past it — past the newline that ends the
/// line, the start of the line under it (or the text's end, with no newline to
/// pass). A rule's row ends at that home.
///
/// Read off the span less the newline djot's span takes in, since Markdown's
/// stops before it. Measured from djot's span end, the home after a rule landed
/// past the blank line under it, on the next block's first character.
pub(crate) fn block_line(source: &str, span_end: usize) -> (usize, usize) {
    let span_end = span_end.min(source.len());
    let text_end = source[..span_end]
        .strip_suffix('\n')
        .map_or(span_end, |s| s.strip_suffix('\r').unwrap_or(s).len());
    let rest = &source[text_end..];
    let newline = if rest.starts_with("\r\n") {
        2
    } else {
        usize::from(rest.starts_with('\n'))
    };
    (text_end, text_end + newline)
}

/// The label of the footnote definition starting at `start` — the `1` in
/// `[^1]: …`. twig gives the `footnote` node no label of its own (no `text`, no
/// `name`), and the bytes that spell it belong to no child node either — the
/// body `para` starts its *content* past them — so the source is the only place
/// to read it from. `None` when what's there isn't a definition after all.
pub(crate) fn footnote_label(source: &str, start: usize) -> Option<&str> {
    let rest = source.get(start..)?.strip_prefix("[^")?;
    let end = rest.find("]:")?;
    Some(&rest[..end])
}

/// Where the body of the footnote definition spanning `span` sits in `source` —
/// everything past the `[^1]:` marker, which is the part a reader actually wants
/// when they follow a reference.
///
/// Source bytes, verbatim but for the whitespace trimmed off each end: a note
/// that says `see *later*` answers with the asterisks in. Rendering that body is
/// a frontend's business the same way painting a [`Role`] is, and a caller that
/// wants it laid out already has the definition on screen where it was written.
///
/// The trim is what makes the common case read right — `[^1]: text` has a space
/// after the colon that belongs to the marker, not the note, and a definition's
/// span runs to the newline ending it.
///
/// The span is taken at its word, which it has only been safe to do since twig
/// 3.1: a djot definition's span used to run *past* its own last line, through
/// the blank line separating it from the next block and into that block's first
/// byte, so `[^2a]: a note.` came back as `"a note.\n\n["` and the offsets named
/// the following note's rows as well as this one's — a reader asking about one
/// footnote was shown two. leaf measured the body itself to get around that, and
/// paid for it: the scan stopped at the first blank line, so a note with a second
/// indented paragraph lost it. Both halves go away with the fix, since a blank
/// line *inside* a definition was always interior to the span and still is.
///
/// A range rather than a slice because "go to note" needs the *position* as much
/// as the text, and it needs the position of the body specifically: a
/// definition's `[^1]:` marker is decoration the caret can't occupy (the rich
/// view draws it as `[1] ` and gives it no stop), so aiming a caret at the
/// definition's first byte lands it on the nearest real stop instead — which is
/// up in the paragraph *above* the note. The body's first byte is a stop, and is
/// where a reader following a reference wants to arrive anyway.
pub(crate) fn footnote_body_span(source: &str, span: Range<usize>) -> Option<Range<usize>> {
    let rest = source.get(span.clone())?.strip_prefix("[^")?;
    let marker = rest.find("]:")?;
    // `span.start` + `[^` + the label + `]:`.
    let after_marker = span.start + 2 + marker + 2;
    let raw = source.get(after_marker..span.end)?;
    // Written as a start plus a length so an all-whitespace body lands on an
    // empty range at the end rather than an inverted one.
    let start = after_marker + (raw.len() - raw.trim_start().len());
    Some(start..start + raw.trim().len())
}

/// The label of the footnote *reference* spanning `span` — the `1` in `[^1]`.
///
/// The peer of [`footnote_label`] for the other half of the pair, and needed for
/// the same reason: a reference whose node carries neither a `content_span` nor
/// a `text` still spells its label plainly in the source. `None` when the bytes
/// aren't a reference after all.
pub(crate) fn footnote_reference_label(source: &str, span: Range<usize>) -> Option<&str> {
    let rest = source.get(span)?.strip_prefix("[^")?;
    let end = rest.find(']')?;
    Some(&rest[..end])
}

/// Where a heading's *content* starts — past the `#`s and the space the rich
/// view hides, for an ATX heading; the block's own start for a setext one (which
/// has no leading marker) and for a format that spells headings some other way.
///
/// Only an empty heading needs asking: with any content at all, the row ends on
/// its last glyph. Bounded to the heading's own first line so a marker-less
/// heading can't scan into the text under it.
fn heading_content_start(source: &str, span: &Range<usize>) -> usize {
    let end = span.end.min(source.len());
    let Some(line) = source.get(span.start..end) else {
        return span.start;
    };
    let line = line.split('\n').next().unwrap_or("");
    let hashes = line.len() - line.trim_start_matches('#').len();
    if hashes == 0 {
        return span.start;
    }
    let after = &line[hashes..];
    span.start + hashes + (after.len() - after.trim_start_matches([' ', '\t']).len())
}

struct Builder<'a> {
    nodes: &'a [FlatNode],
    /// The document source, consulted to place blank-line rows at the source
    /// offsets the caret should occupy on them (the AST drops blank lines).
    source: &'a str,
    /// The word-wrap column budget, or `None` to emit each block as a single
    /// unwrapped row (the frontend wraps).
    wrap: Option<usize>,
    rows: Vec<VRow>,
    /// Built alongside `rows`, never instead of them — see [`TableInfo`].
    tables: Vec<TableInfo>,
    /// The end offset of the last content emitted — the anchor for blank
    /// separator rows so the caret never snaps onto one.
    last_off: usize,
    /// The end of the last block the walk stepped over without drawing — a
    /// comment, which the rich view hides. `last_off` moves past it too, for the
    /// separators; this is kept apart so the trailing blank lines can be counted
    /// from it without also being counted from a code block's closing fence,
    /// which `last_off` likewise ends after. `0` until a hidden block is met.
    stepped_over: usize,
    /// How many rows each block image reserves, keyed by its destination — the
    /// frontend's per-image height, threaded in from [`crate::Doc::set_media_rows`]
    /// so [`Builder::block_media`] can size the placeholder without core doing any
    /// I/O. A destination absent from the map (or a `0`/`1` entry) reserves the
    /// bare one-row placeholder, which is the whole-document default and what
    /// every existing test — passing an empty map — still gets.
    surface: &'a Surface,
    /// The glyph a hard break renders as while the current inline run is built:
    /// a space in prose (a break folds into the flow the frontend wraps), but a
    /// newline (`\n`) inside a table cell, where a row is one source line and the
    /// only break it can carry is an explicit one that must show as a line of its
    /// own. Set around [`Builder::row_cells`] and otherwise left at `' '`.
    break_glyph: Cell<char>,
    /// Render a soft break (a bare newline inside a paragraph) as a line break
    /// where it was written, rather than folding it into the reflowed paragraph
    /// — the `LineFlow::Preserve` behaviour. A soft break emits a `'\n'` glyph
    /// (like a hard break in a cell), which [`Builder::emit_wrapped`] turns into
    /// a fresh visual row. `false` is the flowing-prose default. Inside a table
    /// cell (where `break_glyph` is already `'\n'`) it has no effect: a cell is
    /// one line and folds its own soft breaks regardless.
    preserve_soft: bool,
    /// The source byte range of the one line that should render its markup
    /// *raw* — the caret's line under `MarkupMode::Full` (see
    /// [`crate::Doc::reveal_line`]). `None` in every other mode and view, which
    /// is the delimiters-always-hidden behaviour every build had before the
    /// preference existed.
    ///
    /// Read only by [`Builder::revealed`], which every delimiter-bearing arm of
    /// [`Builder::inline`] consults. A range rather than a bare caret offset
    /// because the decision is per-*node*, not per-caret: a node is revealed
    /// when its span meets this line, so `*em*` shows both its asterisks even
    /// with the caret at one end of it.
    reveal: Option<Reveal>,
    /// The content ends of the hidden marks rendered since the last row was
    /// pushed — recorded as the inline walk meets each mark, and drained onto
    /// the rows as they are emitted (see [`Builder::take_mark_ends`]). A cell
    /// rather than a `&mut`, for the reason `break_glyph` is: the inline walk
    /// borrows the builder shared.
    pending_mark_ends: RefCell<Vec<usize>>,
    /// The inline atoms rendered since the last row was pushed, keyed by the
    /// formula's start offset — the offset the atom glyph carries, which is
    /// how [`Builder::take_math`] pairs each with its glyph once the wrap has
    /// decided which row it landed on. Drained the way `pending_mark_ends` is.
    pending_math: RefCell<Vec<(usize, MathMark)>>,
    /// Whether this walk met a formula at all, atom, block, or revealed — the
    /// fact [`BlockCache`] keeps per block so [`crate::Doc::reveal_line`] can
    /// tell whether the caret's line has anything to reveal without walking.
    saw_math: Cell<bool>,
    /// The presentation vocabulary in force at the block being walked — the
    /// keys the `div`s around it carry, folded together with the nearest
    /// winning, and [`Presentation::default`] at the top level.
    ///
    /// Saved and restored around each `div` in [`Builder::block`], so a block
    /// reads its own attributes over whatever its containers said and nothing
    /// leaks sideways to the block after it. It is per-*build* state rather
    /// than a parameter because every one of the dozen call sites of `block`
    /// would otherwise thread a value none of them care about.
    presentation: Presentation,
    /// The named families this walk has met, by the id its glyphs carry — see
    /// [`VisualMap::faces`]. A `RefCell` for [`pending_mark_ends`]'s reason:
    /// the inline walk borrows the builder shared, and a span's `data-font` is
    /// read from inside it.
    ///
    /// [`pending_mark_ends`]: Builder::pending_mark_ends
    faces: RefCell<FaceTable>,
}

/// The six presentation keys as the walker carries them down a block tree —
/// the two that are the block's ([`Align`], [`LineHeight`]) and the three that
/// are a run's but may be written on the block ([`FontSize`], [`FaceRef`],
/// [`TextColor`]).
///
/// `Copy` and five `Option`s, because folding is the whole of what it does:
/// [`under`](Presentation::under) reads a container's attributes over an
/// existing set and a key the container does not name keeps the value it had.
/// That is the "nearest wins" rule stated once, rather than at each of the
/// three levels a key can be written at. A name and a value fold alike: the
/// nearer node wins whichever of the two forms either of them wrote.
#[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]
struct Presentation {
    align: Option<Align>,
    line_height: Option<LineHeight>,
    size: Option<FontSize>,
    font: Option<FaceRef>,
    color: Option<TextColor>,
}

impl Presentation {
    /// This set with whatever `attrs` names written over it — the nearer node's
    /// answer where it has one, the outer node's where it hasn't.
    ///
    /// `faces` is the build's intern table, which a named family is recorded in
    /// on the way past: the glyph carries the id and the table carries the
    /// string. Shared rather than `&mut` because the inline walk this feeds
    /// borrows the builder shared, the way `pending_mark_ends` does.
    fn under(self, attrs: &[(String, Option<String>)], faces: &RefCell<FaceTable>) -> Self {
        Self {
            align: Align::from_attrs(attrs).or(self.align),
            line_height: LineHeight::from_attrs(attrs).or(self.line_height),
            size: FontSize::from_attrs(attrs).or(self.size),
            font: faces.borrow_mut().face_from_attrs(attrs).or(self.font),
            color: TextColor::from_attrs(attrs).or(self.color),
        }
    }

    /// `base` carrying the three run-level keys — the style a block's glyphs
    /// start from, which an attributed span inside it then writes over.
    fn over(self, base: Style) -> Style {
        base.size(self.size).font(self.font).color(self.color)
    }
}

impl Builder<'_> {
    /// Note that the mark `id` closes with a hidden delimiter, so its content
    /// end is a caret home — unless the mark is empty, where the end is the
    /// start and there is nothing to extend.
    fn note_mark_end(&self, id: usize) {
        let node = &self.nodes[id];
        if let Some(content) = &node.content_span
            && content.end < node.span.end
            && !content.is_empty()
        {
            self.pending_mark_ends.borrow_mut().push(content.end);
        }
    }

    /// The pending mark ends at or before `end_src`, for the row ending there
    /// — every mark rendered so far that closes on it. A mark's end never
    /// exceeds the end of the row its last glyph is on, so the leftovers are
    /// those of rows still to come.
    fn take_mark_ends(&self, end_src: usize) -> Vec<usize> {
        let mut pending = self.pending_mark_ends.borrow_mut();
        let (taken, kept): (Vec<usize>, Vec<usize>) =
            pending.drain(..).partition(|&o| o <= end_src);
        *pending = kept;
        taken
    }

    /// The pending inline atoms whose glyph is on the row `glyphs` is about
    /// to become — each paired with the index of the [`Role::Math`] glyph
    /// carrying its offset, in glyph order. An atom whose glyph landed on an
    /// earlier row was taken then; one on a later row is left for it. The
    /// peer of [`take_mark_ends`](Self::take_mark_ends), and why an atom is
    /// keyed by offset: the wrap decides the row, and the offset is what the
    /// glyph still carries once it has.
    fn take_math(&self, glyphs: &[Glyph]) -> Vec<MathMark> {
        let mut pending = self.pending_math.borrow_mut();
        if pending.is_empty() {
            return Vec::new();
        }
        let mut out = Vec::new();
        for (i, g) in glyphs.iter().enumerate() {
            if g.style.role != Role::Math || !g.stop {
                continue;
            }
            if let Some(at) = pending.iter().position(|(src, _)| *src == g.src) {
                let (_, mut mark) = pending.remove(at);
                mark.glyph = Some(i);
                out.push(mark);
            }
        }
        out
    }
    /// Whether `span` belongs to the line that is showing its raw markup. True
    /// only when a reveal line is set *for markup* (`MarkupMode::Full`) and the
    /// two ranges actually meet — see [`Reveal::meets`] for what meeting is.
    fn revealed(&self, span: &Range<usize>) -> bool {
        self.reveal
            .as_ref()
            .is_some_and(|r| r.markup && r.meets(span))
    }

    /// Whether a formula at `span` shows its TeX rather than its picture: it
    /// meets the reveal line, in *any* mode. A formula's content is its source
    /// and not its picture, so for it the choice is not between a clean
    /// surface and a raw one but between editable and not — which is why this
    /// does not read [`Reveal::markup`] the way [`revealed`](Self::revealed)
    /// does.
    fn math_revealed(&self, span: &Range<usize>) -> bool {
        self.reveal.as_ref().is_some_and(|r| r.meets(span))
    }

    /// The `(opening, closing)` source byte ranges of a node's delimiters — the
    /// bytes its `span` holds that its `content_span` doesn't.
    ///
    /// This is how *every* inline delimiter is recovered, rather than a table of
    /// spellings per kind: twig gives `*em*` a span of `13..17` and a content
    /// span of `14..16`, so the gaps at each end are the delimiters, whatever
    /// they happen to be. That matters because one kind has many spellings —
    /// `*em*` and `_em_` are both emphasis, `` `x` `` and ``` ``x`` ``` both
    /// verbatim — and re-deriving the text from the source is the only way to
    /// show back what the author actually typed. It also gets a link's
    /// asymmetric `[` / `](dest)` right for free.
    ///
    /// `None` when the node has no content span, or when content and span
    /// coincide (nothing was elided, so there is nothing to reveal).
    fn delims(&self, id: usize) -> Option<(Range<usize>, Range<usize>)> {
        let node = &self.nodes[id];
        let content = node.content_span.clone()?;
        let span = node.span.clone();
        // A content span that escapes its own node's span means the two are
        // describing different things; reveal nothing rather than slice wildly.
        if content.start < span.start || content.end > span.end {
            return None;
        }
        let (open, close) = (span.start..content.start, content.end..span.end);
        // A delimiter that spans a newline isn't this line's to reveal — a setext
        // heading's `\n=====` underline is the case that arises in practice. It
        // would also inject a `'\n'` glyph, which `emit_wrapped` reads as a hard
        // row break, so the row would split where the author wrote no break.
        let multiline =
            |r: &Range<usize>| self.source.get(r.clone()).is_some_and(|s| s.contains('\n'));
        if multiline(&open) || multiline(&close) {
            return None;
        }
        (!open.is_empty() || !close.is_empty()).then_some((open, close))
    }

    /// Emit the source bytes of `range` as revealed markup — real glyphs, each
    /// mapped to its own source byte and each a caret stop, so a delimiter shown
    /// is a delimiter that can be selected, edited and deleted like any other
    /// text. Styled [`Role::Delimiter`] on top of the run's own style, which is
    /// how a frontend tells scaffolding from prose and dims it.
    ///
    /// Deliberately *not* [`push_escaped_text`]: this is raw source, not parsed
    /// text, so there is no escape-driven drift between the two to correct.
    fn push_delim(&self, out: &mut Vec<Glyph>, range: &Range<usize>, base: Style) {
        let Some(text) = self.source.get(range.clone()) else {
            return;
        };
        push_text(out, text, range.start, base.role(Role::Delimiter));
    }

    /// Render an inline node's children wrapped in its raw delimiters when the
    /// node is on the revealed line, and bare (delimiters resolved away) when it
    /// isn't — the shared body of every delimiter-bearing arm of
    /// [`inline`](Self::inline).
    ///
    /// `style` is the resolved styling the content still gets in *both* modes:
    /// revealing `*em*` shows the asterisks *and* keeps the text italic, the
    /// live-preview behaviour. Showing the markup is not the same as turning the
    /// rendering off — that is what [`crate::View::Source`] is for.
    fn inline_delimited(&self, id: usize, style: Style, out: &mut Vec<Glyph>) {
        let show = self
            .revealed(&self.nodes[id].span)
            .then(|| self.delims(id))
            .flatten();
        if let Some((open, _)) = &show {
            self.push_delim(out, open, style);
        }
        self.recurse(id, style, out);
        match &show {
            Some((_, close)) => self.push_delim(out, close, style),
            // Hidden, so the content's end has no glyph after it: give the
            // caret its home there.
            None => self.note_mark_end(id),
        }
    }

    /// A verbatim span — or a formula drawn as its TeX — in the code style:
    /// its text at its content's offset, bracketed by its raw delimiters when
    /// `show` says the line is revealed and by nothing (plus the caret's home
    /// at the content's end) when it isn't.
    ///
    /// Not [`inline_delimited`](Self::inline_delimited): verbatim has no child
    /// nodes to recurse into — its content is its own `text` — so the fences
    /// bracket a [`push_text`] instead. The fences themselves keep
    /// `Role::Code`'s sibling treatment via [`push_delim`]'s role override.
    ///
    /// [`push_delim`]: Self::push_delim
    fn inline_verbatim(&self, id: usize, base: Style, out: &mut Vec<Glyph>, show: bool) {
        let node = &self.nodes[id];
        // The interior begins at `content_span.start` — past however many
        // backticks the fence used, which `span.start + 1` only guessed right
        // for a single one. Fall back to that guess if it's absent.
        let at = node
            .content_span
            .as_ref()
            .map_or(node.span.start + 1, |c| c.start);
        let style = base.role(Role::Code);
        let show = show.then(|| self.delims(id)).flatten();
        if let Some((open, _)) = &show {
            self.push_delim(out, open, style);
        }
        push_text(out, node.text.as_deref().unwrap_or(""), at, style);
        match &show {
            Some((_, close)) => self.push_delim(out, close, style),
            None => self.note_mark_end(id),
        }
    }

    fn children(&self, id: usize) -> Vec<usize> {
        let mut out = Vec::new();
        let mut c = self.nodes[id].first_child;
        while let Some(cid) = c {
            out.push(cid.0 as usize);
            c = self.nodes[cid.0 as usize].next_sibling;
        }
        out
    }

    /// Render a node's block children, a blank separator between each. `tight`
    /// suppresses the *fabricated* separator between adjacent children that share
    /// a source line boundary — a tight list item and the sub-list nested in it —
    /// while a real blank source line between them still opens a gap.
    fn blocks(&mut self, id: usize, pf: &[Glyph], pc: &[Glyph], tight: bool) {
        // Frontmatter (a leading `metadata` block) is document metadata, not
        // prose: hide it entirely in the rich-text view. Skipping it here means
        // no phantom blank rows for its lines and no separator before the first
        // real block — the document opens straight into its content.
        let kids: Vec<usize> = self
            .children(id)
            .into_iter()
            .filter(|&c| self.nodes[c].kind != Kind::Metadata)
            .collect();
        let mut above: Option<BlockClass> = None;
        for child in kids {
            let below = BlockClass::from_node_kind(&self.nodes[child].kind);
            let before_sep = self.rows.len();
            if let Some(above) = above {
                self.emit_separators_before(
                    self.nodes[child].span.start,
                    pc,
                    !tight,
                    Boundary { above, below },
                );
            }
            // The first *drawn* child wears the first-row prefix (a bullet, a
            // footnote label), not the first child: a comment opening a list
            // item draws nothing, and the bullet belongs to what follows it.
            let first = if above.is_none() { pf } else { pc };
            if self.block_or_hidden(child, before_sep, first, pc) {
                above = Some(below);
            }
        }
    }

    /// Render `child` after the separator [`Builder::emit_separators_before`]
    /// spelled for it from row `before_sep` on, and say whether it drew
    /// anything.
    ///
    /// A block that draws no rows — an HTML comment, which the rich view hides
    /// the way it hides frontmatter — is still *there* in the source, and the
    /// walk has to step over it: `last_off` moves past it so the next separator
    /// counts the blank lines from its end, not from wherever the last drawn
    /// block stopped. Left where it was, the separator counted every line of the
    /// comment as a blank row; and the cached path, whose per-block builder
    /// starts at offset 0, handed back a `last_off` of 0 and counted every line
    /// of the *document* — one phantom blank row per source line, once per
    /// comment. The separator drawn for it is taken back too, so a hidden block
    /// leaves no gap of its own: what stands either side of it meets across one
    /// boundary, as if the comment were not there.
    fn block_or_hidden(
        &mut self,
        child: usize,
        before_sep: usize,
        pf: &[Glyph],
        pc: &[Glyph],
    ) -> bool {
        let after_sep = self.rows.len();
        self.block(child, pf, pc);
        if self.rows.len() > after_sep {
            return true;
        }
        self.rows.truncate(before_sep);
        let end = self.nodes[child].span.end;
        self.last_off = self.last_off.max(end);
        self.stepped_over = self.stepped_over.max(end);
        false
    }

    /// Render an explicit, ordered list of top-level blocks — [`Builder::blocks`]
    /// for a walk that isn't "the children of one node". The document's top level
    /// no longer is: a footnote definition is a root beside `doc`, not under it,
    /// and [`top_level`] merges it into this list by source position.
    ///
    /// The separator between blocks is spelled by the same
    /// [`Builder::emit_separators_before`] the incremental top-level walk in
    /// [`build_cached`] uses, so the two paths can't drift on how a boundary
    /// looks.
    ///
    /// Returns the class of the last block that drew anything — what the
    /// trailing blank lines close — or `None` when nothing did.
    fn top_blocks(&mut self, ids: &[usize], hidden_end: usize) -> Option<BlockClass> {
        let mut above: Option<BlockClass> = None;
        for &child in ids {
            let below = BlockClass::from_node_kind(&self.nodes[child].kind);
            let before_sep = self.rows.len();
            if let Some(above) = above {
                self.emit_separators_before(
                    self.nodes[child].span.start,
                    &[],
                    true,
                    Boundary { above, below },
                );
            } else {
                let from = self.leading_from(hidden_end);
                self.emit_leading_blank_lines(from, self.nodes[child].span.start, below);
            }
            if self.block_or_hidden(child, before_sep, &[], &[]) {
                above = Some(below);
            }
        }
        above
    }

    /// Emit the blank separator row(s) that sit between a block ending at the
    /// current `last_off` and the next block starting at `next_start`, wearing
    /// the continuation prefix `pc`. Shared by [`Builder::blocks`] and the
    /// incremental top-level walk so the two can't drift on how a boundary is
    /// spelled.
    ///
    /// The blank line(s) between two blocks are real caret stops, each needing
    /// its *own* source offset — one strictly past the previous block's content,
    /// else it collides with that block's last row and `pos_of_offset`
    /// (first-match-wins) would resolve the caret onto the wrong row, pinning
    /// downward motion there.
    ///
    /// One row *per* blank source line, not a single collapsed separator: an
    /// empty paragraph opened between two blocks (Enter in the gap,
    /// `…\n\n\n\n…`) must be a navigable empty row, not vanish — else the caret
    /// in it snaps onto the *next* block's start and Enter looks like it did
    /// nothing.
    fn emit_separators_before(
        &mut self,
        next_start: usize,
        pc: &[Glyph],
        synthetic: bool,
        boundary: Boundary,
    ) {
        let mut offs = self.blank_rows_between(self.last_off, next_start);
        // Under a `</div>` the first blank line is the one Markdown needs to
        // end the div, not a line the author opened. Preserve flow drew it as
        // one, and Backspace there took it and glued the block below onto the
        // closing tag, as raw HTML.
        if self.preserve_soft && offs.first().is_some_and(|&o| self.closes_div_above(o)) {
            offs.remove(0);
            if offs.is_empty() {
                return;
            }
        }
        if offs.is_empty() {
            if !synthetic {
                // A tight list item's own text sits directly above the sub-list
                // nested in it — no fabricated gap. The "breathe" row belongs
                // between free-standing blocks, not between an item and its
                // child list, which the source writes on the very next line. A
                // real blank source line (a loose list) still lands a gap below,
                // because `blank_rows_between` found it and we never reach here.
                return;
            }
            // A tight gap with no blank line (e.g. a heading directly above its
            // text): keep the one conventional separator row so blocks still
            // breathe, as they always have.
            offs.push(self.blank_line_offset(self.last_off, next_start));
        }
        let last = offs.len() - 1;
        for (k, end_src) in offs.into_iter().enumerate() {
            // Only the drawn-only rows carry the boundary: the navigable blank
            // lines between them (and every blank line under preserve-soft flow)
            // are somewhere text can go, not a gap between blocks, and a frontend
            // that shrank one would be shrinking a line the author is typing on.
            let drawn = !self.preserve_soft && (k == 0 || k == last);
            // The blank line a boundary is *drawn* with isn't a place text can
            // go. The first one closes the block above and the last one opens the
            // block below — with a single blank line, the usual case, doing both
            // at once. Typing on either just continues the paragraph it abuts,
            // since the blank line it would need to be a paragraph of its own is
            // the very line being typed on. So they're a gap, like a table's
            // border: drawn, clickable, never a caret's home.
            //
            // The lines *between* them are the real ones. That's what Enter
            // opens: it inserts a paragraph break (`\n\n`), which leaves a blank
            // line spare on each side and the caret on the navigable line
            // between them.
            //
            // Preserve flow is the exception: there a bare `\n` is a visible line
            // break the author edits directly, so a lone blank line *is* a caret
            // home — typing on it makes the soft break the mode exists to show,
            // and Enter at a line's end lands the caret on exactly this row. So no
            // separator is drawn-only; every blank line is navigable.
            self.rows.push(VRow {
                glyphs: pc.to_vec(),
                end_src,
                decoration: drawn,
                code: false,
                code_lang: None,
                directive: false,
                directive_label: None,
                media: None,
                task: None,
                leaf_directive: None,
                heading: None,
                // A line the author can type on is a line of the block
                // around it, and keeps its spacing and alignment.
                align: (!drawn).then_some(self.presentation.align).flatten(),
                line_height: (!drawn).then_some(self.presentation.line_height).flatten(),
                boundary: drawn.then_some(boundary),
                mark_ends: Vec::new(),
                math: Vec::new(),
            });
        }
    }

    /// The blank lines above the first block that draws anything, from `from`
    /// — the first line past any hidden frontmatter or comment — to the line
    /// holding `next_start`. [`Builder::emit_trailing_blank_lines`] turned
    /// upside down: nothing above needs a gap, so every line is an empty
    /// paragraph but the last, which is the gap that opens the block below.
    ///
    /// One blank line is only that gap, and draws nothing, which keeps the
    /// usual line under frontmatter out of sight. Two are the empty paragraph
    /// Enter opens at the first block's start (`\n\nHello`), which drew
    /// nothing either: the caret stayed with the text and the text stayed put.
    /// Preserve flow draws every line, as it does between blocks, but the one
    /// under frontmatter, which is the frontmatter's and not the author's.
    fn emit_leading_blank_lines(&mut self, from: usize, next_start: usize, below: BlockClass) {
        let next_line_start = self.source[..next_start].rfind('\n').map_or(0, |p| p + 1);
        let mut offs = Vec::new();
        let mut start = from;
        while start < next_line_start {
            offs.push(start);
            match self.source[start..next_line_start].find('\n') {
                Some(k) => start += k + 1,
                None => break,
            }
        }
        if self.preserve_soft {
            if from > 0 && !offs.is_empty() {
                offs.remove(0);
            }
        } else if offs.len() < 2 {
            return;
        }
        let last = offs.len().saturating_sub(1);
        for (k, end_src) in offs.into_iter().enumerate() {
            let drawn = !self.preserve_soft && k == last;
            self.rows.push(VRow {
                glyphs: Vec::new(),
                end_src,
                decoration: drawn,
                code: false,
                code_lang: None,
                directive: false,
                directive_label: None,
                media: None,
                task: None,
                leaf_directive: None,
                heading: None,
                align: None,
                line_height: None,
                boundary: drawn.then_some(Boundary {
                    above: BlockClass::Paragraph,
                    below,
                }),
                mark_ends: Vec::new(),
                math: Vec::new(),
            });
        }
    }

    /// Whether the line above the one starting at `line` is a `</div>`.
    fn closes_div_above(&self, line: usize) -> bool {
        let Some(above) = self.source[..line].strip_suffix('\n') else {
            return false;
        };
        let above = above.strip_suffix('\r').unwrap_or(above);
        let start = above.rfind('\n').map_or(0, |p| p + 1);
        above[start..].trim() == "</div>"
    }

    /// The blank lines between a `<div>`'s last block and its closing tag,
    /// drawn inside the div, as the lines between two of its blocks are.
    ///
    /// One blank line is the one Markdown needs before `</div>`, and two are
    /// that and the gap, so neither draws (in Preserve flow, which has no
    /// gap, the second does). A third is the empty paragraph
    /// Enter opens at the end of the div's last block — the end of a line
    /// with a line height or an alignment on it — which drew nothing: the
    /// key looked dead, and each press left another line nobody could see.
    fn emit_div_trailing_lines(&mut self, id: usize, pc: &[Glyph]) {
        let Some(&last) = self.children(id).last() else {
            return;
        };
        if self.rows.is_empty() {
            return;
        }
        let end = self.nodes[id].span.end.min(self.source.len());
        let from = block_line(self.source, self.last_off).1;
        if from >= end {
            return;
        }
        let Some(close) = self.source[from..end].rfind("</div>") else {
            return;
        };
        let close_line = self.source[..from + close].rfind('\n').map_or(0, |p| p + 1);
        let mut offs = Vec::new();
        let mut start = from;
        while start < close_line {
            if !self.source[start..close_line]
                .lines()
                .next()
                .unwrap_or("")
                .trim()
                .is_empty()
            {
                return;
            }
            offs.push(start);
            match self.source[start..close_line].find('\n') {
                Some(k) => start += k + 1,
                None => break,
            }
        }
        // Preserve flow draws every blank line but the one `</div>` needs.
        if offs.len() < if self.preserve_soft { 2 } else { 3 } {
            return;
        }
        offs.pop();
        let above = BlockClass::from_node_kind(&self.nodes[last].kind);
        for (k, end_src) in offs.into_iter().enumerate() {
            let drawn = !self.preserve_soft && k == 0;
            self.rows.push(VRow {
                glyphs: pc.to_vec(),
                end_src,
                decoration: drawn,
                code: false,
                code_lang: None,
                directive: false,
                directive_label: None,
                media: None,
                task: None,
                leaf_directive: None,
                heading: None,
                align: (!drawn).then_some(self.presentation.align).flatten(),
                line_height: (!drawn).then_some(self.presentation.line_height).flatten(),
                boundary: drawn.then_some(Boundary {
                    above,
                    below: BlockClass::Paragraph,
                }),
                mark_ends: Vec::new(),
                math: Vec::new(),
            });
        }
    }

    /// Where the lines above the first drawn block begin: past the hidden
    /// frontmatter, or past the line of the last hidden block the walk
    /// stepped over, whichever is later.
    fn leading_from(&self, hidden_end: usize) -> usize {
        if self.last_off == 0 {
            return hidden_end;
        }
        block_line(self.source, self.last_off).1.max(hidden_end)
    }

    /// One block, drawn under whatever presentation the containers around it
    /// impose.
    ///
    /// A container named `div` with `Element` origin is transparent already —
    /// its children draw as themselves — and it now also *contributes* its
    /// vocabulary keys to every block it holds. That is the reading side of
    /// twig's own rule for where a Markdown block's attributes live: there is
    /// no attribute syntax to put on the paragraph, so `set_block_attrs` writes
    /// a `<div …>` around it, and reading one back has to look through the div.
    /// `<div class="center">` around three paragraphs centres all three, which
    /// is what the author of that HTML meant, and around one is the sole-child
    /// shape twig writes.
    ///
    /// Saved and restored rather than pushed onto a stack, so a nested div
    /// reads its own keys over its parent's and the block *after* the div is
    /// unaffected.
    fn block(&mut self, id: usize, pf: &[Glyph], pc: &[Glyph]) {
        if element_tag(&self.nodes[id]) == Some("div") {
            let saved = self.presentation;
            self.presentation = saved.under(&self.nodes[id].attrs, &self.faces);
            self.block_kind(id, pf, pc);
            self.emit_div_trailing_lines(id, pc);
            self.presentation = saved;
            // Step the walk past the closing `</div>`, as the fenced-div arm
            // below anchors past its `:::`. The tag sits on a line of its own
            // after the last child and the blank line under it, and the rich
            // view draws nothing for it — so left where the last child ended,
            // the separator logic read the tag's line as a blank line between
            // the div and the block below, and drew a navigable empty row there
            // that the author never opened and Backspace could not close; and
            // at the end of the document the trailing count read it as an empty
            // paragraph the author had left. It is hidden markup the walk steps
            // over, which is what `stepped_over` records, so both counts start
            // past it.
            let end = self.nodes[id].span.end;
            self.last_off = self.last_off.max(end);
            self.stepped_over = self.stepped_over.max(end);
            return;
        }
        self.block_kind(id, pf, pc);
    }

    fn block_kind(&mut self, id: usize, pf: &[Glyph], pc: &[Glyph]) {
        let node = &self.nodes[id];
        match node.kind.as_str() {
            "doc" | "section" => self.blocks(id, pf, pc, false),
            "heading" => {
                // A heading whose only visible content is a single image — a
                // banner set in an `<h1>` (`<h1><picture><img></picture></h1>`),
                // or `# ![](banner.png)` — is a block picture, not text. Render
                // it as one; anything with real heading text falls through.
                if let Some((m, kind)) = self.media_only(id) {
                    self.block_media(m, kind, id, pf);
                    return;
                }
                let level = node.level.unwrap_or(1);
                // A `data-size` on a heading scales the *heading's* ramp, not
                // the body's — the role and the step compose rather than
                // compete, which is the same thing a colour does to a link.
                let pres = self.presentation.under(&node.attrs, &self.faces);
                let style = pres.over(heading_style(level));
                let mut glyphs = Vec::new();
                // On the revealed line the `# ` comes back as real, editable
                // text in front of the heading. Only the opening marker: a
                // closing `#`-run (`## title ##`) is covered by the same
                // `delims` pair, and a setext underline is excluded there for
                // being on another line entirely.
                if let Some((open, close)) =
                    self.revealed(&node.span).then(|| self.delims(id)).flatten()
                {
                    self.push_delim(&mut glyphs, &open, style);
                    glyphs.extend(self.inline_children_with_trailing(id, style));
                    self.push_delim(&mut glyphs, &close, style);
                } else {
                    glyphs = self.inline_children_with_trailing(id, style);
                }
                // An *empty* heading — `# ` with nothing typed after it, which is
                // what the toolbar's H1 leaves on a blank line — has no glyph for
                // its row to end on, so the fallback below is the row's whole
                // extent: its only caret stop, and the offset every row after it
                // is measured from. The block's start is the wrong answer for
                // both, because it sits *in front of* the `# ` the rich view
                // hides: the caret drew (and typed) before the hashes, and the
                // rows below inherited an offset short by the marker's length,
                // which put the caret on one of them the moment the heading grew
                // text. Its content's start is where the caret belongs.
                let home = heading_content_start(self.source, &node.span);
                let first = self.rows.len();
                self.emit_wrapped(glyphs, home, pf, pc);
                // Stamp the level on every row the heading just emitted — a
                // wrapped heading's continuation rows as much as its first, and
                // an empty one's single glyphless row, which is the whole point
                // (see [`VRow::heading`]).
                for row in &mut self.rows[first..] {
                    row.heading = Some(level.min(255) as u8);
                    row.align = pres.align;
                    row.line_height = pres.line_height;
                }
            }
            "block_quote" => {
                let (start, end) = (node.span.start, node.span.end);
                let gutter = synth("│ ", Role::QuoteGutter, start);
                let f = concat(pf, &gutter);
                let c = concat(pc, &gutter);
                // A childless quote — a bare `> ` on an otherwise blank line,
                // which is what the toolbar's Quote button leaves there — has no
                // inner block to carry the gutter or a caret home, so `blocks`
                // emitted *nothing at all*: the quote didn't merely draw
                // unstyled, it disappeared, and a document that was only `> `
                // rendered zero rows with the caret nowhere to stand. Emit the
                // gutter row itself, ending just past the marker, exactly as an
                // empty `list_item` emits its bare bullet.
                if self.children(id).is_empty() {
                    self.push_row_at(f, end.min(self.source.len()));
                } else {
                    self.blocks(id, &f, &c, false);
                    self.emit_quote_trailing_lines(&c, end);
                }
            }
            // A generic `:::name{.class}` fenced-div container (twig's
            // `directive`, container form). Core is agnostic of `name` — it's
            // the host app's vocabulary (diaryx's `vis` for audience
            // visibility, say) and isn't available here regardless: twig only
            // threads an `element`'s tag name through `FlatNode::name`, not a
            // directive's own identifier. Every row gets marked `directive` (a
            // frontend draws a tinted panel around each maximal run, the
            // `code`/`code_block` recipe) and the first row carries a label —
            // the way a code fence's language rides only its first row.
            //
            // The label reads BOTH attribute conventions diaryx content
            // actually uses: twig's own dot-prefixed classes (`{.public
            // .family}`, one combined `class` attr) and bare pandoc-style
            // words with no leading dot (`{public family}` — the syntax
            // `diaryx_core::visibility`'s hand-rolled publish-time filter and
            // apps/web's directive serializer both write; twig parses each
            // bare word as its own attribute with an empty value, per
            // `languages/markdown/attributes.zig`). Reading only `.class`
            // would leave every *existing* diaryx `:::vis{...}` block
            // unlabeled.
            // Only the *container* form is the panel below. A `text` directive
            // is inline and never reaches the block walker (see `is_inline`); a
            // `leaf` one is a standalone block with no body, drawn as a
            // placeholder the way an image is.
            "container"
                if container_is_directive(node)
                    && node.directive_form == Some(DirectiveForm::Leaf) =>
            {
                self.block_directive(id, pf);
            }
            // HTML and AsciiDoc spell `insert_directive`'s page break in ways
            // of their own — `<page-break></page-break>`, an *element* with
            // no form at all, and `<<<`, a directive with none — so neither
            // meets the arm above. Both are named `page-break` and empty, and
            // both draw the same placeholder, for the same reason djot's
            // fence below does: a frontend that paginates on the mark must
            // not be able to tell which format the file is in. Narrow to the
            // one name: an arbitrary empty custom element is not a leaf
            // directive.
            "container"
                if node.name.as_deref() == Some(crate::doc::PAGE_BREAK)
                    && node.directive_form.is_none()
                    && node.origin.is_some()
                    && self.children(id).is_empty() =>
            {
                self.block_directive(id, pf);
            }
            // djot has no *leaf* directive form. `insert_directive` spells the
            // same document as an empty `::: page-break` fence — a container
            // with nothing in it — and the name comes back as the fence's one
            // class rather than as the node's name, because djot's div is
            // anonymous. Draw it as the placeholder Markdown's `::page-break`
            // gets, so a frontend that paginates on a `page-break`
            // [`DirectiveMark`] cannot tell which format the file is in.
            //
            // Narrow on purpose: only an *anonymous* empty fence. A Markdown
            // `:::note` with nothing in it keeps the reading it has, because
            // its name is its own and nothing about it says "a block with no
            // body" the way djot's spelling of a leaf directive does.
            "container"
                if container_is_directive(node)
                    && node.directive_form == Some(DirectiveForm::Container)
                    && node.name.as_deref().unwrap_or_default().is_empty()
                    && self.children(id).is_empty()
                    && !leaf_directive_identity(node, self.source).0.is_empty() =>
            {
                self.block_directive(id, pf);
            }
            "container" if container_is_directive(node) => {
                let label = directive_attr_label(&node.attrs);
                let start_row = self.rows.len();
                self.blocks(id, pf, pc, false);
                for (i, row) in self.rows[start_row..].iter_mut().enumerate() {
                    row.directive = true;
                    if i == 0 {
                        row.directive_label = label.clone();
                    }
                }
                // Anchor the block's end past its closing `:::` fence, exactly as
                // the code-block arm anchors past its ```` ``` ````. A container's
                // last content row ends at its last *child*, before the fence and
                // the blank line under it, so the separator logic counted the
                // fence line as a blank row of its own and drew a second boundary
                // — one gap's worth of margin twice, under every fenced div.
                self.last_off = node.span.end;
            }
            "bullet_list" | "ordered_list" | "task_list" => {
                let ordered = node.kind == Kind::OrderedList;
                let mut item_no = 0usize;
                let kids = self.children(id);
                for (i, child) in kids.iter().copied().enumerate() {
                    let kind = &self.nodes[child].kind;
                    if *kind == Kind::ListItem || *kind == Kind::TaskListItem {
                        let start = self.nodes[child].span.start;
                        item_no += 1;
                        // A task item's box replaces the bullet rather than
                        // joining it. The `[ ] ` that spells it is markup twig
                        // has already consumed — the item's paragraph *content*
                        // starts past it — so without a drawn box a task item
                        // was indistinguishable from a plain bullet, ticked or
                        // not. `☐`/`☑` is the marker for the same reason `•` is:
                        // it stands where the source's own marker stands. Which
                        // way it faces is `checked`, straight off the node.
                        let checked = self.nodes[child].checked;
                        let marker = match (checked, ordered) {
                            (Some(true), _) => "☑ ".to_string(),
                            (Some(false), _) => "☐ ".to_string(),
                            (None, true) => format!("{item_no}. "),
                            (None, false) => "• ".to_string(),
                        };
                        let bullet = synth(&marker, Role::ListMarker, start);
                        // The item's later rows wear the marker's own
                        // characters, drawn blank: the width a proportional
                        // face gives the marker, whatever the marker is.
                        let indent = synth(&marker, Role::ListIndent, start);
                        let first_row = self.rows.len();
                        self.block(child, &concat(pc, &bullet), &concat(pc, &indent));
                        // On the item's first row, the way `code_lang` rides the
                        // first row of its block.
                        if let (Some(c), Some(row)) = (checked, self.rows.get_mut(first_row)) {
                            row.task = Some(c);
                        }
                    } else {
                        // twig can nest a *following* top-level block as a direct
                        // child of the list rather than a sibling of it — e.g.
                        // `- item\n\n> quote` parses the block quote under the
                        // `bullet_list`. It isn't a list item, so render it de-nested:
                        // no bullet, at the list's own prefix, with the usual block
                        // separator — never `• │ quote`.
                        if i > 0 {
                            self.emit_separators_before(
                                self.nodes[child].span.start,
                                pc,
                                true,
                                Boundary {
                                    above: BlockClass::from_node_kind(
                                        &self.nodes[kids[i - 1]].kind,
                                    ),
                                    below: BlockClass::from_node_kind(&self.nodes[child].kind),
                                },
                            );
                        }
                        self.block(child, pc, pc);
                    }
                }
            }
            "list_item" | "task_list_item" => {
                // A childless item — the empty bullet you get the instant you
                // press Enter to open a new one — has no inner block to carry the
                // marker prefix or a caret home, so `blocks` would emit nothing
                // and the new bullet simply wouldn't appear until something was
                // typed into it. Emit the prefixed row itself, ending at a caret
                // stop just past the marker (the item's `span.end`), the way an
                // empty paragraph emits its one prefixed row via `emit_wrapped`.
                if self.children(id).is_empty() {
                    let home = self.nodes[id].span.end.min(self.source.len());
                    self.push_row_at(pf.to_vec(), home);
                } else {
                    // Tight: an item's text and the list nested under it butt
                    // together (`• a` / `  • b`), no fabricated blank row between —
                    // a loose item's real blank line still parts them.
                    self.blocks(id, pf, pc, true);
                }
            }
            // A footnote *definition* (`[^1]: the note`). It reaches this walker
            // only because [`top_level`] merges it back in — twig hangs it off no
            // parent at all, so a walk from `doc` never sees one and every byte
            // of its body used to render as nothing.
            //
            // Drawn as a hanging-indent item, the way a list item is: the marker
            // reads `[1] `, matching the `[1]` its references render as, so the
            // two can be paired by eye, and the body wraps under it. The marker
            // is synthetic decoration (one shared offset, never a caret stop) —
            // the `[^1]: ` that spells it in the source is markup, hidden like a
            // heading's `# `.
            "footnote" => {
                let (start, end) = (node.span.start, node.span.end);
                let source = self.source;
                let marker = format!("[{}] ", footnote_label(source, start).unwrap_or(""));
                let f = concat(pf, &synth(&marker, Role::ListMarker, start));
                let c = concat(pc, &synth(&marker, Role::ListIndent, start));
                if self.children(id).is_empty() {
                    // A definition with no body yet — the instant `[^1]: ` has
                    // been typed and nothing after it. `blocks` would emit
                    // nothing and the definition simply wouldn't appear, so emit
                    // the marker row itself with a caret home just past it,
                    // exactly as an empty list item does.
                    self.push_row_at(f, end.min(source.len()));
                } else {
                    self.blocks(id, &f, &c, false);
                }
            }
            // A link reference definition (`[foo]: /url`): resolved by label
            // into the links that use it, and drawn nowhere — the rich view has
            // no more use for its line than for a comment's. It is walked at all
            // (see [`top_level`]) so [`Builder::block_or_hidden`] can step the
            // walk past its bytes rather than count them as blank lines.
            "reference" => {}
            "table" => self.table(id, pf, pc),
            "code_block" => {
                let style = Style::default().role(Role::Code);
                let text = node.text.clone().unwrap_or_default();
                // Cut the block's *terminator*, not every trailing newline. A
                // block whose last line is empty spells that as a second `\n`,
                // and `trim_end_matches` ate it along with the terminator: the
                // Return that made the line got no row, so the caret placed on
                // it fell through to the paragraph below and typing landed
                // outside the block. twig's `content_span` is `text` less
                // exactly this one newline, so cutting one and no more is also
                // what keeps `code_line_offsets` lined up.
                let lines: Vec<&str> = text
                    .strip_suffix('\n')
                    .unwrap_or(text.as_str())
                    .split('\n')
                    .collect();
                // Each line at its own source offset, so the caret can walk the
                // code a character at a time like any other text. Where the
                // lines can't be lined up with the source there's no honest
                // offset to give, so the block maps coarsely to its start (and
                // stays a source-view job, as all of it once was).
                let offs = node
                    .content_span
                    .as_ref()
                    .and_then(|c| self.code_line_offsets(c, &lines));
                // The fence's info string, carried on the block's first row as
                // its language label (`None` for an indented block or a bare
                // fence). Kept on the row so it rides the block cache.
                let lang = code_language(self.source, node.span.start);
                // The block's syntax highlighting, a token per byte range of
                // each line — `None` unless the fence names a language the
                // grammars know (and unless the `syntax` feature is on). Done
                // here, once per build of the block, because the rows it
                // colours ride the block cache: an edit elsewhere in the
                // document reuses them, tokens and all.
                let tokens = lang.as_deref().and_then(|l| code_tokens(l, &lines));
                for (i, raw) in lines.iter().enumerate() {
                    let at = offs.as_ref().map_or(node.span.start, |o| o[i]);
                    // No gutter glyph: the block is set apart by the border and
                    // tint a frontend draws around the whole run of `code` rows,
                    // not by a per-line mark. Just the block prefix and the
                    // code text: the first-row prefix on the first line (a list
                    // item's marker) and the continuation on the rest (its
                    // indent), as a wrapped paragraph takes them.
                    let mut glyphs: Vec<Glyph> = if i == 0 { pf } else { pc }.to_vec();
                    match tokens.as_ref().and_then(|t| t.get(i)) {
                        Some(spans) => push_code_text(&mut glyphs, raw, at, style, spans),
                        None => push_text(&mut glyphs, raw, at, style),
                    }
                    // Explicitly past the line's *text*: a blank code line has no
                    // glyph, and any prefix's offset would put the row's end
                    // inside the next line.
                    self.push_row_at(glyphs, at + raw.len());
                    if let Some(row) = self.rows.last_mut() {
                        row.code = true;
                        if i == 0 {
                            row.code_lang = lang.clone();
                        }
                    }
                }
                // Anchor the block's end past its closing fence. Its last content
                // row ends at the last code line, before the ``` and the blank
                // line under it; without this the separator logic would count the
                // closing-fence line as its own blank row and open a phantom
                // second gap below the block.
                self.last_off = node.span.end;
            }
            "thematic_break" => {
                let full = self.wrap.unwrap_or(UNWRAPPED_RULE_WIDTH);
                let w = full.saturating_sub(prefix_width(pf)).max(4);
                let mut glyphs = pf.to_vec();
                for _ in 0..w {
                    glyphs.push(Glyph {
                        ch: '─',
                        style: Style::default().role(Role::Rule),
                        src: node.span.start,
                        // A rule is a block the caret can sit on, as it always
                        // has; it maps coarsely to the block's start.
                        stop: true,
                    });
                }
                // The dashes share one caret home in front of the atomic block,
                // while the row's end is the second home, past the newline that
                // ends the rule's line. Without that trailing stop a final rule
                // made the document end unreachable: Right could not cross it
                // and a click in the empty space below it snapped back before
                // the rule.
                let (text_end, after_line) = block_line(self.source, node.span.end);
                self.push_row_at(glyphs, after_line);
                // The walk stands at the rule itself, though, as it stands at
                // the end of any other block's last line: the lines under a
                // rule are counted from the newline that ends it. Counted from
                // the home past that newline, every gap under a rule came up a
                // line short, and the empty paragraph Enter opens beneath one
                // drew as nothing at all.
                self.last_off = text_end;
            }
            // A block-level image node with no wrapping paragraph — a promoted
            // top-level HTML `<img>` lands as a direct `doc` child like this
            // (a Markdown `![](…)` comes wrapped in a `para`, handled below).
            "image" => self.block_media(id, MediaKind::Image, id, pf),
            // The same case for a promoted top-level `<video>`/`<audio>`, which
            // arrives as a generic `container` rather than a node kind of its
            // own. It can't be found by the `media_only` scan below the way a
            // wrapped one is: that scan looks at a wrapper's *children*, and here
            // the media element is itself the block.
            "container" if matches!(element_tag(node), Some("video") | Some("audio")) => {
                let kind = match element_tag(node) {
                    Some("audio") => MediaKind::Audio,
                    _ => MediaKind::Video,
                };
                self.block_media(id, kind, id, pf);
            }
            _ => {
                // A container of blocks, or an inline-bearing paragraph.
                let kids = self.children(id);
                // A block-level image: a paragraph (or other wrapper — a
                // `<picture>`, an `<h1>` banner) whose only visible content is a
                // single `image` node. Render it as a placeholder row + record an
                // [`MediaInfo`] a capable frontend replaces. An image mixed with
                // real text or other images on the line isn't block-level and
                // falls through to the inline path below, still as its alt text.
                if let Some((m, kind)) = self.media_only(id) {
                    self.block_media(m, kind, id, pf);
                    return;
                }
                // A display formula on lines of its own — a paragraph whose
                // only visible content is one `display_math` — promotes the
                // same way, to a placeholder row and a [`MathMark`].
                if let Some(m) = self.math_only(id) {
                    self.block_math(m, pf, pc);
                    return;
                }
                let inline = !kids.is_empty() && kids.iter().all(|&c| is_inline(&self.nodes[c]));
                if inline || kids.is_empty() {
                    // The block's own attributes over its containers' — the
                    // three run-level keys become the style its glyphs start
                    // from, and the two line-level ones ride every row it
                    // emits, a wrapped paragraph's continuations included.
                    let pres = self.presentation.under(&node.attrs, &self.faces);
                    let glyphs =
                        self.inline_children_with_trailing(id, pres.over(Style::default()));
                    if !glyphs.is_empty() {
                        let first = self.rows.len();
                        self.emit_wrapped(glyphs, node.span.start, pf, pc);
                        for row in &mut self.rows[first..] {
                            row.align = pres.align;
                            row.line_height = pres.line_height;
                        }
                    }
                } else {
                    self.blocks(id, pf, pc, false);
                }
            }
        }
    }

    /// Render a table as a box-drawn grid: every column as wide as its widest
    /// cell, the header bold and ruled off, each cell padded to its column's
    /// alignment. This is the *default* monospace rendering (see
    /// [`VisualMap::rows`]); the same cells are also published structurally as
    /// [`TableInfo`], so a frontend that lays the grid out in its own units draws
    /// from there and skips the picture built here.
    ///
    /// The alignment comes from twig's `cell.alignment` — the delimiter row
    /// (`|:--|--:|`) that spells it out is consumed by the parser and leaves no
    /// node, so the snapshot is the only source for it.
    ///
    /// Borders and padding are *decoration*: they carry the source offset of the
    /// text they surround, so a click lands in that cell, but they're never
    /// caret stops — the caret steps cell-to-cell instead of into the box art.
    fn table(&mut self, id: usize, pf: &[Glyph], pc: &[Glyph]) {
        let node_end = self.nodes[id].span.end;
        // twig's shape is `[caption, row, row, …]`: the caption is always
        // present (usually empty in Markdown) and is not part of the grid.
        let row_ids: Vec<usize> = self
            .children(id)
            .into_iter()
            .filter(|&c| self.nodes[c].kind == Kind::Row)
            .collect();
        if row_ids.is_empty() {
            return;
        }
        // Lay every cell out first — the column widths depend on all of them.
        let grid: Vec<Vec<TableCell>> = row_ids.iter().map(|&r| self.row_cells(r)).collect();
        let heads: Vec<bool> = row_ids
            .iter()
            .map(|&r| self.nodes[r].head.unwrap_or(false))
            .collect();
        let cols = grid.iter().map(|r| r.len()).max().unwrap_or(0);
        if cols == 0 {
            return;
        }
        let mut widths = vec![0usize; cols];
        for row in &grid {
            for (c, cell) in row.iter().enumerate() {
                widths[c] = widths[c].max(cell_width(&cell.glyphs));
            }
        }
        // Every column at its widest cell is only the *wish*; a grid wider than
        // the surface has its far side hanging off the edge where no amount of
        // caret motion can reach it. Cut it down to what's actually there, and
        // let the cells wrap into the space they're given.
        if let Some(w) = self.wrap {
            fit_widths(&mut widths, w.saturating_sub(prefix_width(pc)));
        }

        // Where the picture starts, so a frontend drawing its own grid knows
        // which rows to skip. Recorded before the first border goes down.
        let rows_start = self.rows.len();

        let anchor = grid[0].first().map(|c| c.start).unwrap_or(node_end);
        self.push_rule(&rule_text(&widths, '┌', '┬', '┐'), anchor, pf);
        for (ri, row) in grid.iter().enumerate() {
            self.push_table_row(row, &widths, pc);
            // The rule under the header: only where the head actually ends.
            let ends_head = heads[ri] && heads.get(ri + 1) == Some(&false);
            if ends_head {
                let next = grid[ri + 1].first().map(|c| c.start).unwrap_or(node_end);
                self.push_rule(&rule_text(&widths, '├', '┼', '┤'), next, pc);
            }
        }
        // The bottom border is the one rule the caret can rest on: its end is
        // the table's trailing stop, the caret home just past the block — the
        // peer of a block picture's second stop, and of a rule's row end. Without
        // it a document ending in a table ended *inside* it: nothing after the
        // last cell was a stop, so Right could not leave the table, and a click
        // in the blank space under it snapped back into the last cell — or, on a
        // surface that resolved the click onto the border row, to the table's
        // first cell, since a decoration row's only stop is the nearest one.
        // Typing at the stop opens a paragraph first, as at a picture's — see
        // `Doc::open_paragraph_at_block_edge`. The glyphs stay non-stops at
        // `node_end`, so a click anywhere on the border lands past the table.
        self.push_rule_with_home(&rule_text(&widths, '└', '┴', '┘'), node_end, pc);

        // The same cells the picture above was drawn from, published unwrapped
        // and unpadded for a frontend that lays them out in pixels.
        self.tables.push(TableInfo {
            rows_span: rows_start..self.rows.len(),
            end_src: node_end,
            // The *continuation* prefix: `pf` opens the block and only its first
            // row wears it, but every row of a grid is a continuation of the
            // block the table sits in.
            prefix: pc.to_vec(),
            grid: grid
                .into_iter()
                .zip(heads)
                .map(|(cells, head)| TableRow { head, cells })
                .collect(),
        });
        // The table's own end anchors whatever separator follows it; the border
        // rows deliberately don't move `last_off` (they hold no content).
        self.last_off = node_end;
    }

    /// One row of laid-out cells, in column order.
    fn row_cells(&self, row: usize) -> Vec<TableCell> {
        // A cell is one source line, so a break within it is an explicit line
        // break (an inline `<br>`) that must render as a line of its own — not the
        // flow-folding space a break is in prose.
        self.break_glyph.set('\n');
        let cells = self
            .children(row)
            .into_iter()
            .filter(|&c| self.nodes[c].kind == Kind::Cell)
            .enumerate()
            .map(|(col, c)| {
                let n = &self.nodes[c];
                let style = if n.head.unwrap_or(false) {
                    Style::default().bold()
                } else {
                    Style::default()
                };
                // Only `content_span` bounds a cell's text, and an EMPTY cell
                // has none at all — twig records no interior for it — so both
                // offsets would fall back to the cell's `span.start`: on the
                // pipe that opens it, or (under a twig that gave every cell
                // the whole row's span) the row's start, where every empty
                // cell collapses onto one spot before the first `│` and a
                // caret there types *before* the table. Derive the interior
                // from the span's own pipes and this cell's column instead,
                // so each empty cell has a distinct, editable caret home.
                let span = n.content_span.clone().unwrap_or_else(|| {
                    let off = empty_cell_offset(
                        &self.source[n.span.start.min(self.source.len())
                            ..n.span.end.min(self.source.len())],
                        n.span.start,
                        col,
                    );
                    off..off
                });
                TableCell {
                    glyphs: self.inline_children(c, style),
                    start: span.start,
                    end: span.end,
                    align: n.alignment.unwrap_or(Alignment::Default),
                }
            })
            .collect();
        self.break_glyph.set(' ');
        cells
    }

    /// A horizontal rule between/around rows — entirely decoration.
    fn push_rule(&mut self, text: &str, src: usize, prefix: &[Glyph]) {
        self.push_rule_row(text, src, prefix, true);
    }

    /// A table's bottom border: drawn like the other rules, but a row the caret
    /// can rest on, its end (`src`) being the table's trailing stop.
    fn push_rule_with_home(&mut self, text: &str, src: usize, prefix: &[Glyph]) {
        self.push_rule_row(text, src, prefix, false);
    }

    fn push_rule_row(&mut self, text: &str, src: usize, prefix: &[Glyph], decoration: bool) {
        let glyphs = concat(prefix, &synth(text, Role::Rule, src));
        self.rows.push(VRow {
            glyphs,
            end_src: src,
            decoration,
            code: false,
            code_lang: None,
            directive: false,
            directive_label: None,
            media: None,
            task: None,
            leaf_directive: None,
            heading: None,
            align: None,
            line_height: None,
            boundary: None,
            mark_ends: Vec::new(),
            math: Vec::new(),
        });
    }

    /// One `│ a │ b │` row of the grid: real cell text between decoration.
    ///
    /// A row of cells is not a row of the screen — a cell wrapped to its column
    /// spans several, each one `│`-divided across the full width so the grid
    /// stays square. Cells in the same row are laid out independently and run
    /// out at their own heights; a column that has run dry pads out as
    /// decoration while its neighbours keep going.
    fn push_table_row(&mut self, cells: &[TableCell], widths: &[usize], prefix: &[Glyph]) {
        let fallback = cells.last().map(|c| c.end).unwrap_or(0);
        let laid: Vec<Vec<Vec<Glyph>>> = cells
            .iter()
            .enumerate()
            .map(|(ci, c)| wrap_glyphs(&c.glyphs, widths.get(ci).copied().unwrap_or(0)))
            .collect();
        let height = laid.iter().map(|l| l.len()).max().unwrap_or(1).max(1);

        for j in 0..height {
            let mut glyphs = prefix.to_vec();
            for (ci, &w) in widths.iter().enumerate() {
                let cell = cells.get(ci);
                let line = laid.get(ci).and_then(|l| l.get(j));
                // The divider before this column belongs to the cell it
                // introduces, so clicking it lands in that cell — on this line
                // of it, which is what's next to the divider being clicked.
                let at = line
                    .and_then(|l| l.first().map(|g| g.src))
                    .or_else(|| cell.map(|c| c.start))
                    .unwrap_or(fallback);
                glyphs.extend(synth("│", Role::Rule, at));
                match (cell, line) {
                    (Some(cell), Some(line)) => {
                        let pad = w.saturating_sub(glyphs_width(line));
                        let (lead, trail) = match cell.align {
                            Alignment::Right => (pad, 0),
                            Alignment::Center => (pad / 2, pad - pad / 2),
                            Alignment::Left | Alignment::Default => (0, pad),
                        };
                        // Every line renders at least one space after its text
                        // (the gutter before `│`), so there is always somewhere
                        // to put the "after the last character" caret a line
                        // needs. It's the one padding glyph that is a stop: on
                        // the cell's last line that's the cell's end, and on any
                        // other it's the space the wrap consumed.
                        let last = laid[ci].len() == j + 1;
                        let end = match last {
                            true => cell.end,
                            false => line
                                .last()
                                .map(|g| g.src + g.ch.len_utf8())
                                .unwrap_or(cell.end),
                        };
                        glyphs.extend(synth(&" ".repeat(lead + 1), Role::Body, at));
                        glyphs.extend(line.iter().cloned());
                        glyphs.push(Glyph {
                            ch: ' ',
                            style: Style::default(),
                            src: end,
                            stop: true,
                        });
                        glyphs.extend(synth(&" ".repeat(trail), Role::Body, end));
                    }
                    // A ragged row, or a column whose cell ended higher up: pad
                    // it out so the grid stays square.
                    _ => {
                        let at = cell.map(|c| c.end).unwrap_or(fallback);
                        glyphs.extend(synth(&" ".repeat(w + 2), Role::Body, at));
                    }
                }
            }
            glyphs.extend(synth("│", Role::Rule, fallback));
            // The row ends where its last stop does. A table row has no gap
            // between its final cell and the border, so inventing an end past
            // that would be a stop with nothing under it.
            let end_src = glyphs
                .iter()
                .rev()
                .find(|g| g.stop)
                .map_or(fallback, |g| g.src);
            let mark_ends = self.take_mark_ends(end_src);
            let math = self.take_math(&glyphs);
            self.rows.push(VRow {
                glyphs,
                end_src,
                decoration: false,
                code: false,
                code_lang: None,
                directive: false,
                directive_label: None,
                media: None,
                task: None,
                leaf_directive: None,
                heading: None,
                align: None,
                line_height: None,
                boundary: None,
                mark_ends,
                math,
            });
        }
    }

    /// Render a block-level image, video, or audio as one placeholder row: the
    /// `🖼 alt` / `🎬 alt` / `🔊 alt` label styled [`Role::Image`], every glyph
    /// mapped to the media's start offset and a caret stop there (they share the
    /// offset, so the stop table dedups them to a single home in front of it, as
    /// a rule's dashes do), and the row's end stop set past it so the caret can
    /// also rest after it. The row carries a [`MediaMark`] so [`media_spans`]
    /// publishes it as a [`MediaInfo`] a capable frontend replaces with the real
    /// picture or player; a plain surface paints the label as-is. `pf` is the
    /// block prefix (a list indent, a quote gutter) the row opens with, exactly
    /// as every other block honours it.
    fn block_media(&mut self, img: usize, kind: MediaKind, wrapper: usize, pf: &[Glyph]) {
        let node = &self.nodes[img];
        let start = node.span.start;
        let end = node.span.end;
        // An `image`'s URL is twig's `destination`; a `<video>`/`<audio>` is a
        // generic element, so its URL is the `src` attribute — and may be absent
        // entirely, the element naming its candidates in child `<source>`s.
        let destination = match kind {
            MediaKind::Image => node.destination.clone().unwrap_or_default(),
            MediaKind::Video | MediaKind::Audio => attr_of(node, "src").unwrap_or_default(),
        };
        let poster = match kind {
            MediaKind::Video => attr_of(node, "poster").unwrap_or_default(),
            MediaKind::Image | MediaKind::Audio => String::new(),
        };
        // The `<source>`s under the media element itself, not under `wrapper`: a
        // `<video>` is its own container, unlike an `<img>`, whose `<picture>`
        // alternatives are its *siblings* and so only reachable from the wrapper.
        let sources = match kind {
            MediaKind::Image => self.media_sources(wrapper),
            MediaKind::Video | MediaKind::Audio => self.media_sources(img),
        };
        let alt = self.image_alt(img);
        let sigil = kind.sigil();
        let label = if alt.is_empty() {
            // With no alt, name the file — but a `<video>` with neither `src` nor
            // alt has only its `<source>`s to be named by, so fall back to the
            // first candidate rather than labelling the row a bare sigil.
            let named = if destination.is_empty() {
                sources
                    .first()
                    .map(|s| s.srcset.as_str())
                    .unwrap_or_default()
            } else {
                &destination
            };
            format!("{sigil} {}", media_label(named))
        } else {
            format!("{sigil} {alt}")
        };
        let style = Style::default().role(Role::Image);
        let mut glyphs = pf.to_vec();
        for ch in label.chars() {
            glyphs.push(Glyph {
                ch,
                style,
                src: start,
                stop: true,
            });
        }
        // How many rows the frontend wants for this picture: the label row plus
        // the blank fillers below it. Absent (a GUI that lays images out in
        // pixels, an image that didn't resolve, or a plain surface) means the
        // bare one-row placeholder.
        let rows = self
            .surface
            .media_rows
            .get(&destination)
            .copied()
            .unwrap_or(1)
            .max(1);
        // End past the image so the caret has a stop after it: the last glyph's
        // offset is the image *start*, not its extent, so `push_row`'s
        // last-glyph rule would strand the end stop inside the markup.
        self.push_row_at(glyphs, end);
        if let Some(row) = self.rows.last_mut() {
            row.media = Some(MediaMark {
                kind,
                destination,
                sources,
                alt,
                poster,
                rows,
            });
        }
        // Reserve the picture's remaining height as blank `decoration` rows: drawn
        // (so the frontend has the vertical room to paint the raster over them),
        // but holding no caret and contributing no stops — vertical motion steps
        // over them and the caret's only homes stay the stop in front of the image
        // and the one just past it, both on the label row above. They anchor at the
        // image's end offset so a click on the picture's lower half lands after it,
        // the nearest caret home. Mirrors how a table's box-rule rows reserve space
        // without ever holding the caret.
        for _ in 1..rows {
            self.rows.push(VRow {
                glyphs: Vec::new(),
                end_src: end,
                decoration: true,
                code: false,
                code_lang: None,
                directive: false,
                directive_label: None,
                media: None,
                task: None,
                leaf_directive: None,
                heading: None,
                align: None,
                line_height: None,
                boundary: None,
                mark_ends: Vec::new(),
                math: Vec::new(),
            });
        }
        self.last_off = end;
    }

    /// The `<picture>` alternatives inside block-image `wrapper`, in document
    /// order — every `<source>` element in its subtree. Empty when there's no
    /// `<picture>`. Each is a `<source>`'s `media` + `srcset`; core keeps them
    /// verbatim and picks none (see [`MediaSource`]). A `<source>` with no
    /// `srcset` is dropped (nothing to load); its `media` may be empty (an
    /// unconditional override), which a frontend treats as always-matching.
    ///
    /// It scans the wrapper's whole subtree (via the forward `first_child` /
    /// `next_sibling` links, the reliable ones) rather than the `<img>`'s parent,
    /// for two reasons. A `<picture>` reaches core in two shapes: twig promotes a
    /// block `<picture>` to an `element(picture)` wrapping `[source, img]`, but
    /// leaves an inline one's tags as raw siblings — `[raw "<picture>", source,
    /// img, raw "</picture>"]` — so the `<source>`s sit at different depths in
    /// the two. And the editor's flat arena leaves a promoted inline node's
    /// `parent` back-pointer dangling on a phantom root, so only the wrapper
    /// (known at the call site) is a trustworthy anchor. A block image is the
    /// sole visible content of its wrapper, so every `<source>` under it is its
    /// picture's.
    fn media_sources(&self, wrapper: usize) -> Vec<MediaSource> {
        let mut out = Vec::new();
        self.collect_sources(wrapper, &mut out);
        out
    }

    fn collect_sources(&self, id: usize, out: &mut Vec<MediaSource>) {
        for c in self.children(id) {
            let node = &self.nodes[c];
            if node.name.as_deref() == Some("source") {
                // `<picture>` spells its candidate `srcset`, `<video>`/`<audio>`
                // spell it `src`. Both mean "the URL to load", so they normalise
                // onto one field; `srcset` wins where (illegally) both appear.
                let url = attr_of(node, "srcset").or_else(|| attr_of(node, "src"));
                if let Some(srcset) = url {
                    out.push(MediaSource {
                        media: attr_of(node, "media").unwrap_or_default(),
                        srcset,
                        mime: attr_of(node, "type").unwrap_or_default(),
                    });
                }
            }
            self.collect_sources(c, out);
        }
    }

    /// The single block-level media `id`'s subtree resolves to, or `None`.
    ///
    /// A wrapper is a block picture when the only *visible* thing under it is one
    /// image: whitespace-only text and structure-only elements (a `<picture>`'s
    /// `<source>`, which declares an alternate but paints nothing) don't count,
    /// and the search descends through wrapping elements (`<picture>`, a linking
    /// `<a>`). This is what makes `<p><img></p>`, a bare `<img>`, and
    /// `<h1><picture>…<img></picture></h1>` all render as one framed picture.
    /// Any real text, or a second image, means it isn't image-only — it falls
    /// back to inline rendering, where the image still shows as its alt text.
    ///
    /// [`FlatNode`]'s snapshot doesn't carry an element's tag name, so a
    /// `<source>` can't be skipped by name — but it needs no special case:
    /// contributing no image and no text, it's simply invisible to the scan.
    fn media_only(&self, id: usize) -> Option<(usize, MediaKind)> {
        let mut found = None;
        let mut count = 0usize;
        let mut has_text = false;
        self.scan_visual(id, &mut found, &mut count, &mut has_text);
        (count == 1 && !has_text).then(|| found.unwrap())
    }

    /// Walk `id`'s subtree tallying visible leaves for [`media_only`]: each
    /// image, `<video>`, or `<audio>` (remembering the last, counting the total)
    /// and whether any non-whitespace text appears. Media isn't descended into —
    /// an image's inline children are alt text, and a `<video>`'s are its
    /// no-support fallback and its `<source>` declarations, none of which is
    /// document content.
    ///
    /// [`media_only`]: Self::media_only
    fn scan_visual(
        &self,
        id: usize,
        found: &mut Option<(usize, MediaKind)>,
        count: &mut usize,
        has_text: &mut bool,
    ) {
        for c in self.children(id) {
            let node = &self.nodes[c];
            match node.kind.as_str() {
                "image" => {
                    *found = Some((c, MediaKind::Image));
                    *count += 1;
                }
                // A `<video>`/`<audio>` reaches core as a generic `container`
                // (twig gives neither a semantic node, so `html_elements`
                // promotion leaves the tag name on `name`). Counted as media and
                // *not* descended into, so its `<source>` children and its
                // "your browser does not support…" fallback text neither add a
                // second count nor make the block look like text.
                "container" if matches!(element_tag(node), Some("video") | Some("audio")) => {
                    let kind = match element_tag(node) {
                        Some("audio") => MediaKind::Audio,
                        _ => MediaKind::Video,
                    };
                    *found = Some((c, kind));
                    *count += 1;
                }
                // Text leaves: only non-whitespace counts as visible content.
                // (Twig keeps the whitespace `str`s between HTML tags — the
                // newlines and indentation inside a `<picture>` — as real nodes.)
                "str" | "smart_punctuation" | "verbatim" | "inline_math" | "display_math" => {
                    if node.text.as_deref().is_some_and(|t| !t.trim().is_empty()) {
                        *has_text = true;
                    }
                }
                // Structural breaks carry no visible glyph of their own.
                "soft_break" | "hard_break" | "non_breaking_space" => {}
                // Any other wrapper (emphasis, a link, a `<picture>`) is
                // transparent to the scan — descend into it.
                _ => self.scan_visual(c, found, count, has_text),
            }
        }
    }

    /// The single `display_math` that is all of `id`'s visible content, or
    /// `None` — [`media_only`](Self::media_only) for a formula. Whitespace-only
    /// text around it does not count (twig keeps the newlines either side of a
    /// `$$` on its own lines as `str`s); any other text, or a second formula,
    /// means the paragraph is prose with math in it, and falls through to the
    /// inline path.
    fn math_only(&self, id: usize) -> Option<usize> {
        let mut found = None;
        let mut count = 0usize;
        for c in self.children(id) {
            let node = &self.nodes[c];
            match node.kind.as_str() {
                "display_math" => {
                    found = Some(c);
                    count += 1;
                }
                "str" | "smart_punctuation" => {
                    if node.text.as_deref().is_some_and(|t| !t.trim().is_empty()) {
                        return None;
                    }
                }
                "soft_break" | "hard_break" | "non_breaking_space" => {}
                _ => return None,
            }
        }
        (count == 1).then(|| found.unwrap())
    }

    /// A display formula on lines of its own, drawn on
    /// [`block_media`](Self::block_media)'s recipe: one placeholder row —
    /// `∑` and the TeX on one line, every glyph [`Role::Math`] at the
    /// formula's start and a caret stop there, the row ending past the
    /// formula so the caret can rest after it too — carrying a [`MathMark`]
    /// for [`math_spans`] to publish, and under it as many blank decoration
    /// rows as the frontend said the picture is tall
    /// ([`Surface::math_rows`]).
    ///
    /// On the caret's line the block is its source instead: the `$$`
    /// delimiters in [`Role::Delimiter`] and the TeX between them in
    /// [`Role::Code`], line for line, the way a fence draws — so a formula is
    /// edited where it stands and folds back to its picture when the caret
    /// leaves. That is the same rule an inline formula follows, and it is
    /// what makes a block-level formula need no equivalent of
    /// [`VisualMap::block_media_stop`]: both of the placeholder's caret homes
    /// are on the formula's own lines, and standing on either reveals it.
    fn block_math(&mut self, math: usize, pf: &[Glyph], pc: &[Glyph]) {
        self.saw_math.set(true);
        let node = &self.nodes[math];
        let (start, end) = (node.span.start, node.span.end);
        let tex = node.text.clone().unwrap_or_default();
        if self.math_revealed(&node.span) {
            let mut glyphs = Vec::new();
            self.inline_verbatim(math, Style::default(), &mut glyphs, true);
            self.emit_wrapped(glyphs, start, pf, pc);
            self.last_off = end;
            return;
        }
        let style = Style::default().role(Role::Math);
        let mut glyphs = pf.to_vec();
        // One line of label: the TeX with its newlines folded, so the row
        // reads as one thing however the source laid it out.
        let label: String = tex.split_whitespace().collect::<Vec<_>>().join(" ");
        for ch in format!("{MATH_ATOM} {label}").chars() {
            glyphs.push(Glyph {
                ch,
                style,
                src: start,
                stop: true,
            });
        }
        let rows = self
            .surface
            .math_rows
            .get(&tex)
            .copied()
            .unwrap_or(1)
            .max(1);
        self.push_row_at(glyphs, end);
        if let Some(row) = self.rows.last_mut() {
            row.math = vec![MathMark {
                tex,
                display: true,
                glyph: None,
                rows,
            }];
        }
        for _ in 1..rows {
            self.rows.push(VRow {
                glyphs: Vec::new(),
                end_src: end,
                decoration: true,
                code: false,
                code_lang: None,
                directive: false,
                directive_label: None,
                media: None,
                task: None,
                leaf_directive: None,
                heading: None,
                align: None,
                line_height: None,
                boundary: None,
                mark_ends: Vec::new(),
                math: Vec::new(),
            });
        }
        self.last_off = end;
    }

    /// A leaf directive (`::name{…}`) as one placeholder row — the
    /// [`block_media`](Self::block_media) recipe, for the same reason: it is a
    /// block that renders as *a thing*, not as text, and the frontend paints
    /// whatever the host app's vocabulary makes of it.
    ///
    /// The row's glyphs are a `⧉ label` (or `⧉ name`) stand-in a plain surface
    /// paints as-is, every glyph anchored at the directive's start with a caret
    /// stop there, and the row ending past it so the caret can also rest after
    /// it. It carries a [`DirectiveMark`] for [`directive_spans`], and is marked
    /// [`directive`](VRow::directive) so a frontend already drawing the
    /// container form's panel frames this one identically for free.
    ///
    /// Before this, a leaf directive emitted no rows at all: it was invisible,
    /// held no caret, and vertical motion crossed a void where it stood.
    fn block_directive(&mut self, id: usize, pf: &[Glyph]) {
        let node = &self.nodes[id];
        let (start, end) = (node.span.start, node.span.end);
        let (name, attrs) = leaf_directive_identity(node, self.source);
        let label = self.image_alt(id); // its `[label]` children, flattened
        let shown = if label.is_empty() { &name } else { &label };
        // How many rows the frontend drew the directive in: the label row plus
        // the blank fillers below it. Absent means the bare placeholder.
        let rows = self
            .surface
            .directive_rows
            .get(&DirectiveKey::new(&name, &label, &attrs))
            .copied()
            .unwrap_or(1)
            .max(1);
        let style = Style::default().role(Role::Image);
        let mut glyphs = pf.to_vec();
        for ch in format!("⧉ {shown}").chars() {
            glyphs.push(Glyph {
                ch,
                style,
                src: start,
                stop: true,
            });
        }
        // End past the directive so the caret has a stop after it — the same
        // reason `block_media` anchors its row at the image's end.
        self.push_row_at(glyphs, end);
        if let Some(row) = self.rows.last_mut() {
            row.directive = true;
            row.leaf_directive = Some(DirectiveMark {
                name,
                attrs,
                label,
                rows,
            });
        }
        // The rest of the host's drawing as blank `decoration` rows, holding no
        // caret — `block_media`'s fillers, anchored at the directive's end for
        // the same reason: a click on the drawing's lower part lands after it.
        for _ in 1..rows {
            self.rows.push(VRow {
                glyphs: Vec::new(),
                end_src: end,
                decoration: true,
                code: false,
                code_lang: None,
                directive: false,
                directive_label: None,
                media: None,
                task: None,
                leaf_directive: None,
                heading: None,
                align: None,
                line_height: None,
                boundary: None,
                mark_ends: Vec::new(),
                math: Vec::new(),
            });
        }
        self.last_off = end;
    }

    /// An image's alt text: the flattened text of its inline descendants (an
    /// image's children *are* its alt content), empty when it has none. Also a
    /// leaf directive's `[label]`, which is the same shape — inline children
    /// standing for the block.
    fn image_alt(&self, id: usize) -> String {
        let mut out = String::new();
        self.collect_text(id, &mut out);
        out
    }

    /// Append every descendant's `text` to `out`, in document order. Inline text
    /// (`str`) nodes are leaves, so a node never contributes both its own text and
    /// a child's — no double counting.
    fn collect_text(&self, id: usize, out: &mut String) {
        for c in self.children(id) {
            if let Some(t) = &self.nodes[c].text {
                out.push_str(t);
            }
            self.collect_text(c, out);
        }
    }

    fn inline_children(&self, id: usize, base: Style) -> Vec<Glyph> {
        let mut out = Vec::new();
        for c in self.children(id) {
            self.inline(c, base, &mut out);
        }
        out
    }

    /// [`inline_children`](Self::inline_children) plus any trailing whitespace the
    /// block carries past its inline content (see [`trailing_ws_glyphs`]). Used
    /// for the leaf inline blocks — paragraphs and headings — whose own `span`
    /// bounds exactly one line of text, so the trailing gap is theirs. *Not* for
    /// a table cell, whose `span` is the whole row and would swallow the
    /// delimiters and neighbours between it and the row's end.
    ///
    /// [`trailing_ws_glyphs`]: Self::trailing_ws_glyphs
    fn inline_children_with_trailing(&self, id: usize, base: Style) -> Vec<Glyph> {
        let mut out = self.inline_children(id, base);
        out.extend(self.trailing_ws_glyphs(id, base));
        out
    }

    /// Glyphs for whatever trailing whitespace a block's source carries past its
    /// last inline node — the space(s) at the end of `hello ` that Markdown and
    /// Djot drop from the `str` node as insignificant. twig still records them:
    /// a block's `content_span` ends at its last meaningful character while its
    /// `span` runs to the end of the line's text (before the terminating
    /// newline), so the gap between the two *is* that trailing whitespace.
    ///
    /// Emitting it as real caret-stop glyphs is what lets the caret be drawn
    /// past the last visible character. Without it, typing a space at the end of
    /// a paragraph moved the caret in the source but not on screen — the caret
    /// stuck on the last glyph until the next visible character reparsed the
    /// space into an interior `str` node that finally carried it.
    ///
    /// Restricted to spaces: only they are safe to synthesize one-cell-per-byte,
    /// and only they are what the parser silently strips. Anything else in the
    /// gap means the span accounting isn't what this assumes, so it's left alone.
    fn trailing_ws_glyphs(&self, id: usize, style: Style) -> Vec<Glyph> {
        let node = &self.nodes[id];
        let Some(content) = &node.content_span else {
            return Vec::new();
        };
        let (from, to) = (content.end, node.span.end);
        let Some(slice) = (from < to).then(|| self.source.get(from..to)).flatten() else {
            return Vec::new();
        };
        if slice.is_empty() || slice.bytes().any(|b| b != b' ') {
            return Vec::new();
        }
        slice
            .bytes()
            .enumerate()
            .map(|(i, _)| Glyph {
                ch: ' ',
                style,
                src: from + i,
                stop: true,
            })
            .collect()
    }

    fn inline(&self, id: usize, base: Style, out: &mut Vec<Glyph>) {
        let node = &self.nodes[id];
        match node.kind.as_str() {
            "str" | "smart_punctuation" => push_escaped_text(
                out,
                node.text.as_deref().unwrap_or(""),
                node.span.clone(),
                self.source,
                base,
            ),
            "soft_break" | "hard_break" | "non_breaking_space" => {
                // A break renders as a real, caret-navigable glyph — but twig
                // gives it no span of its own (`0..0`), so the offset comes from
                // the text in front of it: one *past* the last glyph, which is
                // the newline the break stands for. Past, not on: sharing the
                // previous glyph's offset would put two stops on one byte, and a
                // caret that can't change offset can't move.
                let src = if node.span.start != 0 {
                    node.span.start
                } else {
                    out.last().map(|g| g.src + g.ch.len_utf8()).unwrap_or(0)
                };
                // A *hard* break renders as this run's break glyph — a newline
                // inside a table cell (its own line), the same space in prose the
                // frontend re-wraps. A soft break normally folds into a space;
                // under `LineFlow::Preserve` it renders as a `'\n'` too, so the
                // author's line break shows where it was written. Never inside a
                // cell (`break_glyph` is `'\n'` there): a cell is one line and
                // folds its own soft breaks regardless.
                let ch = if node.kind == Kind::HardBreak {
                    self.break_glyph.get()
                } else if node.kind == Kind::SoftBreak
                    && self.preserve_soft
                    && self.break_glyph.get() == ' '
                {
                    '\n'
                } else {
                    ' '
                };
                out.push(Glyph {
                    ch,
                    style: base,
                    src,
                    stop: true,
                });
            }
            // A cell's only spelling for an in-line break is a raw `<br>`; read it
            // back as one (outside a cell it stays the literal text it falls to
            // below). The tag's bytes carry no stop of their own — the line it
            // ends stops just before it, the next just after.
            "raw_inline" if self.break_glyph.get() == '\n' && is_br(node.text.as_deref()) => {
                out.push(Glyph {
                    ch: '\n',
                    style: base,
                    src: node.span.start,
                    stop: true,
                });
            }
            "emph" => self.inline_delimited(id, base.italic(), out),
            "strong" => self.inline_delimited(id, base.bold(), out),
            // A coloured highlight's emoji is spelling, not content: twig strips
            // it and records the colour on the node, so the glyphs are the
            // author's words and the colour rides the role. Revealed markup
            // still shows the emoji, because `delims` reads the source bytes
            // between the span and the content span — which is exactly the
            // `==🔴 ` the author typed.
            "mark" => {
                let color = MarkColor::from_attrs(&node.attrs);
                self.inline_delimited(id, base.role(Role::Mark(color)), out)
            }
            "insert" => self.inline_delimited(id, base.underline(), out),
            "delete" => self.inline_delimited(id, base.strikethrough(), out),
            // The one pair whose whole meaning is *where the glyphs sit*. Drawn
            // in the surrounding style otherwise, so `^**2**^` stays bold and a
            // superscript inside a heading keeps the heading's role — which is
            // exactly why this is a `Baseline` and not a `Role`.
            "superscript" => self.inline_delimited(id, base.baseline(Baseline::Super), out),
            "subscript" => self.inline_delimited(id, base.baseline(Baseline::Sub), out),
            "verbatim" => self.inline_verbatim(id, base, out, self.revealed(&node.span)),
            // A formula in a line. Three renderings, in order of preference:
            //
            // - On the caret's line it is its TeX in the code style with its
            //   delimiters shown — in *every* markup mode, because the
            //   content is not the picture and hiding the `$` alone would
            //   leave nothing to edit. `math_revealed` is `revealed` without
            //   the mode gate.
            // - On a surface that paints pictures in a line it is one atom
            //   glyph, `stop: true` at the formula's start, standing for the
            //   whole thing; the [`MathMark`] drained onto the row by
            //   [`take_math`](Self::take_math) says what the frontend draws
            //   there. The caret has the stop on the atom and the next glyph's
            //   past it, and nothing inside the markup.
            // - Elsewhere — a terminal — it is the code-styled TeX with the
            //   delimiters hidden, exactly the verbatim treatment, and what
            //   `inline_math` rendered as before there was anything else.
            //   `display_math` had no arm at all and fell to the default one,
            //   which pushed its text at the *node's* start, three bytes short
            //   of where the text sits.
            "inline_math" | "display_math" => {
                self.saw_math.set(true);
                if self.math_revealed(&node.span) {
                    self.inline_verbatim(id, base, out, true);
                } else if self.surface.inline_pictures {
                    let tex = node.text.clone().unwrap_or_default();
                    self.pending_math.borrow_mut().push((
                        node.span.start,
                        MathMark {
                            tex,
                            display: node.kind == Kind::DisplayMath,
                            glyph: None,
                            rows: 1,
                        },
                    ));
                    out.push(Glyph {
                        ch: MATH_ATOM,
                        style: base.role(Role::Math),
                        src: node.span.start,
                        stop: true,
                    });
                } else {
                    self.inline_verbatim(id, base, out, false);
                }
            }
            // An attributed span — the run-level half of the presentation
            // vocabulary. djot's `[text]{…}`, AsciiDoc's `[.a]#text#`, HTML's
            // and Markdown's `<span …>`: one node with a name twig hands back
            // for two of the four (see [`is_run_span`]), all four carrying the
            // author's `data-size`, `data-font` and `data-color` on the run
            // they cover.
            //
            // The keys are written over the surrounding style rather than
            // replacing it, so a span inside a block that names its own size
            // wins on size and keeps the block's face — the nearest-wins rule
            // the block walker applies through a `div`. A key the span does not
            // name is one the block still says.
            //
            // A `data-color` here is the text's *foreground*, where the same key
            // on a `mark` is a highlight's background: same vocabulary, same
            // enum, and no collision, because a `mark` is a `mark` and a span is
            // a span.
            //
            // Otherwise this is the plain `recurse` an anonymous container has
            // always had — no delimiters, because the `{…}` is markup and the
            // span's text is the author's words.
            "container" if is_run_span(node) && !self.children(id).is_empty() => {
                self.recurse(id, run_style(node, base, &self.faces), out)
            }
            // A text directive (`:name[label]{…}`) — the inline form of a generic
            // directive. Its `[label]` children are the visible text; the name and
            // the `{…}` attributes are the host app's vocabulary (diaryx's
            // `:vis[…]`) and stay hidden markup, exactly as a link's `](dest)` is.
            // Drawn in the surrounding style: a role of its own would need one
            // every frontend maps, and the bug this fixes is that the text was
            // invisible, not that it was unstyled.
            "container" if container_is_directive(node) && !self.children(id).is_empty() => {
                self.recurse(id, base, out)
            }
            // No `[label]`, so there are no children to render and recursing
            // emitted *nothing*: the directive's bytes vanished from the document
            // and left no caret stop behind. What to draw instead turns on
            // whether the syntax looks deliberate.
            //
            // Bare `:word` almost never is. twig matches a colon followed by any
            // letter-led word (`scanTextDirective`, deliberately matching remark),
            // so ordinary prose is full of them — `:see below`, a `:smile:`
            // shortcode, a stray colon before a word. Those are prose, and prose
            // renders as itself: every byte visible, every byte a caret stop, so a
            // colon typed by accident can be seen and deleted. Hiding them behind
            // a placeholder would be the invisible-and-unreachable failure this
            // arm exists to fix, just wearing a nicer glyph.
            "container" if container_is_directive(node) && node.attrs.is_empty() => {
                let span = node.span.clone();
                push_text(
                    out,
                    self.source.get(span.clone()).unwrap_or(""),
                    span.start,
                    base,
                );
            }
            // `{…}` attributes, though, are unmistakably deliberate — nobody
            // types `:vis{.family}` by accident, and diaryx writes exactly that
            // inline. So an attribute-bearing directive with no label draws as a
            // chip on `block_directive`'s recipe (`⧉ name attrs`, `Role::Image`),
            // the inline peer of the leaf form's placeholder row.
            //
            // Only the first glyph is a caret stop, and the whole chip shares the
            // directive's start offset: the caret treats it as one atomic thing
            // rather than walking hidden markup a byte at a time, and a paragraph
            // holding nothing but a chip still has a stop to be navigated to.
            "container" if container_is_directive(node) => {
                let start = node.span.start;
                let name = node.name.clone().unwrap_or_default();
                let shown = match directive_attr_label(&node.attrs) {
                    Some(attrs) if !name.is_empty() => format!("⧉ {name} {attrs}"),
                    Some(attrs) => format!("⧉ {attrs}"),
                    None => format!("⧉ {name}"),
                };
                let style = base.role(Role::Image);
                for (i, ch) in shown.chars().enumerate() {
                    out.push(Glyph {
                        ch,
                        style,
                        src: start,
                        stop: i == 0,
                    });
                }
            }
            // A footnote reference (`[^1]`). The label bracketed is what a reader
            // needs — bare, `note1` reads as a typo rather than a reference — so
            // the `^` is hidden as the spelling artefact it is (a link's
            // `](dest)` goes the same way) and the brackets are kept as
            // decoration: one shared offset, never a caret stop, like a table's
            // borders, so the caret walks the label alone.
            //
            // Styled `Role::Link`: a reference *is* a link to its definition, and
            // every frontend already paints that role. A role of its own would
            // need one in each of them, and what a frontend needs to tell the two
            // apart is not a paint colour but an answer to "what does clicking
            // here do" — which is [`Doc::footnote_at_caret`]'s job, not a glyph's.
            //
            // Raised, though, because that a reference is *set* differently from
            // the prose it interrupts is exactly what makes it read as a
            // reference. `[1]` at body size reads as bracketed text.
            "footnote_reference" => {
                let style = base.role(Role::Link);
                // Revealed, the reference is just its source bytes: the `^` that
                // is normally elided comes back and every byte becomes a real
                // stop, so the brackets stop being decoration and start being
                // text. That's the whole point of the mode, and it replaces the
                // hand-built chip below rather than decorating it — including the
                // raised baseline, since what's on screen there is source, and
                // source is set as prose.
                if self.revealed(&node.span) {
                    self.push_delim(out, &node.span, style);
                    return;
                }
                let style = style.baseline(Baseline::Super);
                // The label's own span, so its glyphs map to their true bytes.
                // Absent one, it starts past the `[^` that opens the reference.
                let (label, at) = match &node.content_span {
                    Some(c) => (self.source.get(c.clone()).unwrap_or(""), c.start),
                    None => (node.text.as_deref().unwrap_or(""), node.span.start + 2),
                };
                out.push(Glyph {
                    ch: '[',
                    style,
                    src: node.span.start,
                    stop: false,
                });
                push_text(out, label, at, style);
                out.push(Glyph {
                    ch: ']',
                    style,
                    src: node.span.end.saturating_sub(1),
                    stop: false,
                });
            }
            "link" | "url" | "email" => {
                let style = base.role(Role::Link);
                if self.children(id).is_empty() {
                    // A bare autolink (`<a@b.c>`, a naked URL): the destination
                    // *is* the visible text, so there is nothing elided to
                    // reveal and both modes draw the same thing.
                    push_text(
                        out,
                        node.destination
                            .as_deref()
                            .or(node.text.as_deref())
                            .unwrap_or("link"),
                        node.span.start,
                        style,
                    );
                } else {
                    // An inline link reveals asymmetrically — `[` before the
                    // label, `](dest)` after it — which the generic
                    // span-minus-content derivation already produces.
                    self.inline_delimited(id, style, out);
                }
            }
            _ => {
                if self.children(id).is_empty() {
                    if let Some(t) = &node.text {
                        push_text(out, t, node.span.start, base);
                    }
                } else {
                    self.recurse(id, base, out);
                }
            }
        }
    }

    fn recurse(&self, id: usize, style: Style, out: &mut Vec<Glyph>) {
        for c in self.children(id) {
            self.inline(c, style, out);
        }
    }

    /// Lay a block's inline `glyphs` into visual rows, prefixing the first with
    /// `pf` and the rest with `pc`. A preserved soft break arrives as a `'\n'`
    /// glyph (see the `soft_break` arm): a hard row boundary that splits the
    /// glyphs so each run lays out on its own and the author's line structure
    /// shows on screen. The `'\n'` is dropped from the row it closes and its
    /// source offset becomes that row's end stop — exactly how a table cell's
    /// in-line `<br>` is handled — so the caret can rest at the line's end
    /// without a zero-width control char leaking into what the frontends render.
    /// With no `'\n'` present (the folding default, and every build that isn't
    /// `LineFlow::Preserve`) there is one run and this is byte-identical to
    /// laying the glyphs out directly.
    fn emit_wrapped(&mut self, glyphs: Vec<Glyph>, block_start: usize, pf: &[Glyph], pc: &[Glyph]) {
        if !glyphs.iter().any(|g| g.ch == '\n') {
            self.emit_line(glyphs, block_start, pf, pc, None);
            return;
        }
        // Each run up to a '\n' is a line of its own: the first wears the block's
        // opening prefix, every later one the continuation prefix, and the break's
        // own offset ends the run's last row. The break glyph is dropped. A
        // trailing '\n' flushes its run and leaves nothing behind, so no spurious
        // blank row follows it.
        let mut run: Vec<Glyph> = Vec::new();
        let mut first = true;
        for g in glyphs {
            if g.ch == '\n' {
                let lead = if first { pf } else { pc };
                self.emit_line(std::mem::take(&mut run), block_start, lead, pc, Some(g.src));
                first = false;
            } else {
                run.push(g);
            }
        }
        if !run.is_empty() {
            let lead = if first { pf } else { pc };
            self.emit_line(run, block_start, lead, pc, None);
        }
    }

    /// Word-wrap a single line of `glyphs` (no interior line breaks) to the
    /// available width and push the visual rows, prefixing the first with `pf`
    /// and the rest with `pc`. `end`, when set, is the source offset that ends
    /// the line's final row — the offset of the break that terminated it, which
    /// the caller has already stripped from `glyphs`; when `None` the row ends
    /// just past its last glyph, as an unbroken block's does.
    fn emit_line(
        &mut self,
        glyphs: Vec<Glyph>,
        block_start: usize,
        pf: &[Glyph],
        pc: &[Glyph],
        end: Option<usize>,
    ) {
        // The line's final row ends at `end` when a break gave one, else just
        // past its last glyph (`push_row`'s default).
        let push_last = |b: &mut Self, row: Vec<Glyph>| match end {
            Some(e) => b.push_row_at(row, e),
            None => b.push_row(row, block_start),
        };

        // No column budget: emit the whole line as one row and let the frontend
        // wrap it at its own (pixel) width.
        let Some(width) = self.wrap else {
            let row = if glyphs.is_empty() {
                pf.to_vec()
            } else {
                concat(pf, &glyphs)
            };
            push_last(self, row);
            return;
        };

        // Split into words (maximal non-space runs), each carrying the space
        // glyph that followed it (so its source offset is preserved).
        let mut words: Vec<(Vec<Glyph>, Option<Glyph>)> = Vec::new();
        let mut word: Vec<Glyph> = Vec::new();
        for g in glyphs {
            if g.ch == ' ' {
                words.push((std::mem::take(&mut word), Some(g)));
            } else {
                word.push(g);
            }
        }
        if !word.is_empty() {
            words.push((word, None));
        }
        if words.is_empty() {
            // An empty block (or an empty preserved line) still occupies one
            // (prefixed) row.
            push_last(self, pf.to_vec());
            return;
        }

        let mut line: Vec<Glyph> = Vec::new();
        let mut used = 0usize;
        let mut first = true;
        for (w, space) in words {
            let avail = width
                .saturating_sub(prefix_width(if first { pf } else { pc }))
                .max(1);
            let cells = glyphs_width(&w);
            if used > 0 && used + cells > avail {
                let row = concat(if first { pf } else { pc }, &line);
                self.push_row(row, block_start);
                line = Vec::new();
                used = 0;
                first = false;
            }
            used += cells;
            line.extend(w);
            if let Some(sp) = space {
                used += 1;
                line.push(sp);
            }
        }
        let row = concat(if first { pf } else { pc }, &line);
        push_last(self, row);
    }

    /// The source offset of each line of a code block's `text`.
    ///
    /// `content` is the block's `content_span` — where twig says the body lives
    /// in the source, fences already excluded. Its lines run 1:1 with the
    /// rendered `text` lines, so no search is needed; each is anchored at the
    /// *end* of its source line, which places it past whatever indent `text` had
    /// stripped (a fenced block's fences, an indented one's leading spaces)
    /// without having to know how much there was.
    ///
    /// `None` when the body and the rendered lines don't line up — a coarse
    /// fallback the caller turns into the block's start offset.
    fn code_line_offsets(&self, content: &Range<usize>, lines: &[&str]) -> Option<Vec<usize>> {
        let mut src_lines: Vec<(usize, &str)> = Vec::new();
        let mut at = content.start;
        for l in self.source.get(content.start..content.end)?.split('\n') {
            src_lines.push((at, l));
            at += l.len() + 1;
        }
        if src_lines.len() != lines.len() {
            return None;
        }
        Some(
            lines
                .iter()
                .zip(&src_lines)
                .map(|(l, (start, sl))| start + sl.len().saturating_sub(l.len()))
                .collect(),
        )
    }

    fn push_row(&mut self, glyphs: Vec<Glyph>, fallback: usize) {
        // Step past the character the *source* holds at the last glyph's offset,
        // not past the glyph's own `ch`. The two agree for ordinary text, but a
        // glyph is not always the character it stands on: `synth` decoration and
        // a substituted run (an image's `⧉ label`) share one offset by design.
        // Trusting `ch` there yields an offset inside a multi-byte character,
        // which every later slice of `source` panics on.
        let end_src = glyphs
            .last()
            .map(|g| {
                let at = g.src.min(self.source.len());
                at + self.source[at..].chars().next().map_or(0, char::len_utf8)
            })
            .unwrap_or(fallback);
        self.push_row_at(glyphs, end_src);
    }

    /// Push a row with an explicit end stop, for content that knows its own
    /// extent better than its last glyph does.
    fn push_row_at(&mut self, glyphs: Vec<Glyph>, end_src: usize) {
        self.last_off = end_src;
        let mark_ends = self.take_mark_ends(end_src);
        let math = self.take_math(&glyphs);
        self.rows.push(VRow {
            glyphs,
            end_src,
            decoration: false,
            code: false,
            code_lang: None,
            directive: false,
            directive_label: None,
            media: None,
            task: None,
            leaf_directive: None,
            heading: None,
            align: None,
            line_height: None,
            boundary: None,
            mark_ends,
            math,
        });
    }

    /// Where the last row's line ends — what the lines under it are counted
    /// from: the row's own end, or the walk's where that stands short of it.
    /// Only a thematic break's does, whose row ends at the caret's home past
    /// the newline under the rule while the walk stands at the rule itself
    /// (see its arm). Every other block leaves the walk at its last row's end
    /// or past it — past a closing fence, say, which the counts below reach
    /// by other means.
    fn last_line_end(&self) -> Option<usize> {
        self.rows.last().map(|r| r.end_src.min(self.last_off))
    }

    /// The quote's own trailing marker lines: the `>` / `> ` lines that lie past
    /// its last child but inside its span, one gutter row each.
    ///
    /// Pressing Enter at the end of `> a` writes `> a\n>\n> \n` — twig's
    /// spelling, and the right one. Those last two lines hold no block (a
    /// `block_quote`'s `content_span` still stops at its last child) so the
    /// children walk never reaches them, and they used to fall all the way to
    /// the document-level [`Builder::emit_trailing_blank_lines`], which knows no
    /// prefix: the gutter simply stopped, and a writer adding a line to a quote
    /// watched it draw as plain prose.
    ///
    /// This is only answerable since twig 3.2.0, where a Markdown `block_quote`'s
    /// span covers its own trailing marker lines (it reported `0..3` for that
    /// source and now reports `0..8`). Before that the lines belonged to no node
    /// at any level, and the only way to draw them was to sniff `>` off the raw
    /// source and re-derive the nesting depth by counting markers — format
    /// inference this crate exists to keep out of the render path.
    ///
    /// Each row is a real caret home rather than a decoration gap: the writer
    /// spelled every one of these lines with a marker of its own, so each is a
    /// line of the quote to stand on, not the spacing between two blocks (which
    /// is [`Builder::emit_separators_before`]'s, and falls *between* children
    /// where this never looks).
    fn emit_quote_trailing_lines(&mut self, pc: &[Glyph], end: usize) {
        let end = end.min(self.source.len());
        // From where the walk stands, not where the last row ends: a fenced
        // code block's last row is its last line of code, and its closing
        // fence (`> ```) is markup the block already stepped `last_off` past.
        // Counted from the row, the fence's line drew as an empty quoted line.
        let mut at = self.last_line_end().unwrap_or(0).max(self.last_off);
        // Walk line by line from the last child's end to the quote's, taking each
        // line's *end* as the row's offset — the caret home at the end of a line
        // is where one on an empty quoted line belongs, and it keeps every row's
        // offset distinct from its neighbours'.
        while at < end {
            let Some(k) = self.source[at..end].find('\n') else {
                break;
            };
            let line_start = at + k + 1;
            let line_end = self.source[line_start..end]
                .find('\n')
                .map_or(end, |i| line_start + i);
            self.push_row_at(pc.to_vec(), line_end);
            at = line_end;
        }
    }

    /// The source offset the caret rests at on the blank line separating a block
    /// that ends at `prev_end` from the next block starting at `next_start`:
    /// just past the newline that terminates the previous block, but kept
    /// strictly before the next block so the offset is unique to this row.
    fn blank_line_offset(&self, prev_end: usize, next_start: usize) -> usize {
        let after_nl = self.source[prev_end..]
            .find('\n')
            .map_or(prev_end, |p| prev_end + p + 1);
        after_nl.min(next_start.saturating_sub(1)).max(prev_end)
    }

    /// The source offset of each blank row between a block ending at `prev_end`
    /// and content starting at `next_start` — one per blank source line. The
    /// first newline terminates the previous block's line; every line it opens up
    /// to (but not including) the line that holds `next_start` is a blank row the
    /// caret can occupy. Offsets are unique and ascending so `pos_of_offset`
    /// resolves each to its own row. Empty when the two blocks are tight (no
    /// blank line between them).
    fn blank_rows_between(&self, prev_end: usize, next_start: usize) -> Vec<usize> {
        // Spans aren't always in tidy source order (e.g. a block after
        // frontmatter can start *before* the previous block's rendered content
        // ends). There's no blank line to place then — fall back to the clamped
        // single separator (an empty return) rather than slicing an inverted
        // range.
        if next_start <= prev_end {
            return Vec::new();
        }
        let gap = &self.source[prev_end..next_start];
        let Some(nl) = gap.find('\n') else {
            return Vec::new();
        };
        // The line holding `next_start` belongs to the next block; blank rows
        // stop before it.
        let next_line_start = self.source[..next_start].rfind('\n').map_or(0, |p| p + 1);
        let mut offs = Vec::new();
        let mut start = prev_end + nl + 1;
        while start < next_line_start {
            offs.push(start);
            match self.source[start..next_start].find('\n') {
                Some(k) => start += k + 1,
                None => break,
            }
        }
        offs
    }

    /// Blank lines the user typed past the end of the last block (e.g. two
    /// `Enter`s to open a fresh paragraph) leave no AST node, so nothing renders
    /// and the caret appears stuck on the old line. Reconstruct one empty row
    /// per extra trailing newline from the source, each at its own offset, so
    /// the caret rides down onto the new line the moment it's created.
    ///
    /// `above` is the class of the last block in the document — the one this gap
    /// closes. A document with no blocks at all has nothing above these rows, and
    /// [`BlockClass::Paragraph`] is the honest answer there too: what they are is
    /// empty paragraphs, on both sides of the gap.
    fn emit_trailing_blank_lines(&mut self, above: BlockClass, hidden_end: usize) {
        // With no rows at all the count starts past any hidden frontmatter, not
        // at 0: its newlines are not trailing blank lines, and counting them
        // opened phantom rows *inside* the metadata for a frontmatter-only file.
        //
        // Or past the last hidden block, if that is later: a closing comment
        // draws no row, and its lines are not blank lines the author opened.
        //
        // Or past the last byte of content, if *that* is later: a fenced code
        // block's last row ends at its last line of code, and the closing
        // fence under it is markup with no row of its own — so counted from
        // the row, the fence's own line and terminator read as two blank lines
        // and opened a phantom empty paragraph whose offset was *inside* the
        // fence. Typing on it broke the fence. (A setext heading's underline
        // was the same shape.) Trailing whitespace is not content, so a line
        // of spaces still counts as the blank line it looks like.
        let last_end = self
            .last_line_end()
            .unwrap_or(hidden_end)
            .max(self.stepped_over)
            .max(self.source.trim_end().len());
        if last_end >= self.source.len() {
            return;
        }
        // The first newline after the last content just terminates that line, so
        // a lone trailing `\n` (an ordinary file ending) opens no blank row. A
        // *second* newline opens an empty paragraph: render it the way a block
        // boundary is rendered — a blank spacer row, then the empty paragraph row
        // the caret rests on — so the just-pressed-Enter view already shows the
        // gap it will keep once text is typed, and typing doesn't shift the line
        // down. One row per trailing newline (each its own caret offset), the
        // last landing at the document end where the caret sits.
        let extra = self.source[last_end..].matches('\n').count();
        if extra < 2 {
            return;
        }
        for k in 1..=extra {
            self.rows.push(VRow {
                glyphs: Vec::new(),
                end_src: last_end + k,
                // As between two blocks: the first blank row is the gap that
                // closes the block above, not somewhere to type. Nothing follows
                // to need a gap of its own, though, so every row after it is a
                // real empty paragraph — the end of the document bounds the last
                // one the way a following block would. Preserve flow makes even
                // that first row navigable, as it does every blank line.
                decoration: !self.preserve_soft && k == 1,
                code: false,
                code_lang: None,
                directive: false,
                directive_label: None,
                media: None,
                task: None,
                leaf_directive: None,
                heading: None,
                align: None,
                line_height: None,
                // The one drawn row here is a block boundary like any other —
                // "rendered the way a block boundary is rendered" is the whole
                // point of it — so it says so, and a frontend spacing boundaries
                // spaces this one the same. The rows below it are navigable empty
                // paragraphs, not gaps.
                boundary: (!self.preserve_soft && k == 1).then_some(Boundary {
                    above,
                    below: BlockClass::Paragraph,
                }),
                mark_ends: Vec::new(),
                math: Vec::new(),
            });
        }
    }
}

// ── display width ────────────────────────────────────────────────────────────
//
// Two things a row can be counted in, and they are not the same number:
//
//   *glyphs*, one per codepoint — how the text is stored here, and what an
//   index into `VRow::glyphs` means; and
//   *columns*, one per terminal cell — where the text is drawn, and what every
//   `col` in this crate means.
//
// `你` is one glyph in two columns. Counting columns with `glyphs.len()` (or,
// in the source view, `chars().count()`) is the same number only for the ASCII
// that most fixtures are written in, and drifts one cell per wide character
// everywhere else — the caret drawn a column short of the text it types into.
// Everything below converts between the two; nothing else should have to.

/// The display width of `s` in terminal cells.
///
/// Measured per grapheme cluster, because that is the unit a surface advances
/// by: `👨‍👩‍👧` is five codepoints measuring 2 + 0 + 2 + 0 + 2 cells one at a
/// time, but the character they spell is drawn in 2. Both frontends already
/// measure it that way — ratatui asks `unicode-width` per cluster, and the GUI
/// asks its own text system — so the caret only lands where the text is if this
/// agrees with them.
pub fn text_width(s: &str) -> usize {
    UnicodeWidthStr::width(s)
}

/// One grapheme cluster of a laid-out row: the glyphs that spell it, and the
/// cells it is drawn in.
///
/// The cluster, not the glyph, is what has a width. A row's glyphs are one per
/// codepoint, so an accented letter or an emoji is several of them drawn in one
/// character's worth of cells — the glyph that opens the cluster claims those
/// cells, and the ones continuing it are drawn *inside* them rather than beside
/// them. It's the same cluster the stop table is built on: the opening glyph is
/// the one a caret can rest on, and so the only one whose column it can be
/// drawn at.
struct Cluster {
    /// Index of the glyph that opens it.
    glyph: usize,
    /// The display column it starts at.
    col: usize,
    /// How many cells it is drawn in. Zero for a cluster with no width of its
    /// own (a lone joiner), which therefore sits at no column at all.
    cells: usize,
}

/// Walk a row's glyphs as the clusters they spell, in column order.
fn clusters(glyphs: &[Glyph]) -> Vec<Cluster> {
    let text: String = glyphs.iter().map(|g| g.ch).collect();
    let mut out = Vec::new();
    let (mut glyph, mut col) = (0, 0);
    for cluster in text.graphemes(true) {
        let cells = text_width(cluster);
        out.push(Cluster { glyph, col, cells });
        // One glyph per codepoint, so a cluster spans exactly its own.
        glyph += cluster.chars().count();
        col += cells;
    }
    out
}

/// The display width of a run of glyphs.
fn glyphs_width(glyphs: &[Glyph]) -> usize {
    clusters(glyphs).last().map_or(0, |c| c.col + c.cells)
}

/// A cell's display width — the widest of its lines, since an in-cell `\n` break
/// splits it into several. Sizes the column that must hold every line.
fn cell_width(glyphs: &[Glyph]) -> usize {
    glyphs
        .split(|g| g.ch == '\n')
        .map(glyphs_width)
        .max()
        .unwrap_or(0)
}

/// Whether a raw inline HTML tag is a line break (`<br>`, `<br/>`, `<br />`,
/// case-insensitively) — the one tag a table cell reads as an in-cell break.
fn is_br(text: Option<&str>) -> bool {
    let Some(t) = text else { return false };
    matches!(
        t.trim().to_ascii_lowercase().replace(' ', "").as_str(),
        "<br>" | "<br/>"
    )
}

impl VRow {
    /// The row's width in display columns — and so the column of the caret
    /// placed past its last glyph, which is the rightmost column it can occupy.
    fn width(&self) -> usize {
        glyphs_width(&self.glyphs)
    }

    /// The display column glyph `i` is drawn at. Glyphs continuing a cluster
    /// report the column of the glyph that opened it, since that is where they
    /// are drawn; none of them is ever a stop, so no caret is placed by it.
    fn col_of_glyph(&self, i: usize) -> usize {
        clusters(&self.glyphs)
            .iter()
            .rev()
            .find(|c| c.glyph <= i)
            .map_or(0, |c| c.col)
    }

    /// The glyph drawn at display column `col`, or `None` past the row's last
    /// cell.
    ///
    /// A column landing on the *second* cell of a wide glyph resolves to that
    /// glyph: half a character is not a place to be, so clicking either cell of
    /// `你` means `你`, and the caret comes to rest at its start — the column it
    /// would be drawn at anyway. That rule is what makes the mapping invertible:
    /// every offset has one column, and every column has one offset.
    fn glyph_at_col(&self, col: usize) -> Option<usize> {
        clusters(&self.glyphs)
            .into_iter()
            .find(|c| col < c.col + c.cells)
            .map(|c| c.glyph)
    }
}

// ── helpers ──────────────────────────────────────────────────────────────────

/// The caret home inside an *empty* table cell (`col`, 0-based) whose node
/// `span` is `src` starting at byte `start`. twig gives an empty cell no
/// `content_span`, so its interior is read from the pipes: the home is one
/// space past the pipe that opens the cell — mimicking the `| ` padding a
/// filled cell has — and never at or past the pipe that closes it. So
/// `|  |  |` gives the two cells distinct, editable homes instead of both
/// collapsing onto the row's start.
///
/// Two shapes of span are read. A twig from 3.3.3 gave every cell of a row
/// the *row's* span, so the cell's own pipes are the `col`-th and
/// `col+1`-th unescaped `|` in it; a later twig spans a cell from the pipe
/// that opens it to the one that closes it, exclusive, so the span holds at
/// most that one pipe, at its start, and the closing one is the byte past
/// its end. The two are told apart by the pipes the span holds — a row's
/// span has several, or one that is not at its start.
fn empty_cell_offset(src: &str, start: usize, col: usize) -> usize {
    let bytes = src.as_bytes();
    let mut pipes = Vec::new();
    for (i, &b) in bytes.iter().enumerate() {
        if b == b'|' && (i == 0 || bytes[i - 1] != b'\\') {
            pipes.push(i);
        }
    }
    let whole_row = pipes.len() > 1 || pipes.first().is_some_and(|&i| i != 0);
    let (open, close) = if whole_row {
        (pipes.get(col).copied(), pipes.get(col + 1).copied())
    } else {
        (pipes.first().copied(), Some(src.len()))
    };
    match (open, close) {
        (Some(open), Some(close)) => {
            let lo = open + 1; // just inside the opening pipe
            let hi = close.saturating_sub(1); // just inside the closing pipe
            let inside = if hi < lo {
                lo
            } else {
                (open + 2).clamp(lo, hi)
            };
            start + inside
        }
        (Some(open), None) => start + open + 1,
        _ => start,
    }
}

/// One laid-out table cell: its rendered text, the source range that text
/// occupies (`start`/`end` are the caret anchors decoration points at), and the
/// column alignment its padding honours.
///
/// `glyphs` is the cell's inline content *unwrapped* — the box-drawn rows wrap
/// it to a column width, but a frontend laying the grid out itself needs the
/// text before that decision was made.
#[derive(Clone)]
pub struct TableCell {
    pub glyphs: Vec<Glyph>,
    pub start: usize,
    pub end: usize,
    pub align: Alignment,
}

/// One row of a table's grid, as the document spells it — not as it's drawn.
#[derive(Clone)]
pub struct TableRow {
    /// A header row: drawn bold, and ruled off from the body below it.
    pub head: bool,
    pub cells: Vec<TableCell>,
}

/// A table's structure, published alongside the box-drawn rows that spell it.
///
/// The rows in [`VisualMap::rows`] are the *default monospace* picture of a
/// table: every border a `│`, every column a whole number of character cells.
/// That picture is exactly right on any monospace surface, and unfixable off one
/// — in a proportional font the `│`s of two rows land at different x and the grid
/// shears. So a frontend that draws its own geometry reads this instead: the
/// cells, their alignment, and which rows are the head, with no opinion about
/// how wide a column is or what a border looks like.
///
/// Both are always built. The TUI paints `rows` and ignores this; the GUI skips
/// `rows` for the span in `rows_span` and draws from here. They describe the
/// same cells, so the caret lands on the same offsets either way.
#[derive(Clone)]
pub struct TableInfo {
    /// The `VisualMap::rows` this table's picture occupies, borders included —
    /// what a frontend drawing its own table skips over.
    pub rows_span: Range<usize>,
    /// The end of the table node's source span, and the offset its trailing
    /// caret stop sits at — the one caret home past the last cell, held by the
    /// bottom border row's end. Typing there opens a paragraph under the table
    /// rather than joining the block; see `Doc::open_paragraph_at_block_edge`.
    pub end_src: usize,
    /// The block prefix every row of this table carries — a blockquote's `│ `
    /// gutter, a list item's indent. Empty for a table at the top level.
    ///
    /// A frontend drawing its own grid has to render this and start the table
    /// past it, exactly as the picture does; a table nested in a quote that
    /// draws flush at the left margin has left the quote.
    pub prefix: Vec<Glyph>,
    pub grid: Vec<TableRow>,
}

/// A fenced or indented code block, named by the [`VisualMap::rows`] it occupies.
///
/// Unlike a table, the rows *are* the block's content — a frontend still paints
/// them, it just draws a border and a tinted background around the whole span
/// and lets the code inside scroll horizontally instead of wrapping. So this
/// carries only the row range; there's no structural alternative to the picture
/// the way [`TableInfo`] is one. Derived from [`VRow::code`] — see
/// [`code_block_spans`].
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct CodeBlockInfo {
    /// The contiguous run of [`VisualMap::rows`] this code block spans, blank
    /// code lines included.
    pub rows_span: Range<usize>,
    /// The block's language, from a fenced block's info string — what a frontend
    /// paints as a small label on the box (`` ```rust `` → `Some("rust")`).
    /// `None` for a fence written without one, or an indented block. Editing it
    /// goes through [`crate::Doc::set_code_language`], which re-finds the fence
    /// in the AST, so this stays a display string.
    pub lang: Option<String>,
}

/// A block-level image (`![alt](url)` on its own line), named by the single
/// [`VisualMap::rows`] row it occupies.
///
/// Like [`CodeBlockInfo`], the row *is* the block's default rendering — a plain
/// surface paints the `🖼 alt` placeholder glyphs as-is. An image-capable
/// frontend instead **skips the row in `rows_span`** and paints the resolved
/// picture there, exactly as it skips a [`TableInfo`]'s box-drawn rows. Derived
/// from [`VRow::media`] by [`media_spans`], so it survives the row reuse of
/// [`BlockCache`] and [`build_spliced`].
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct MediaInfo {
    /// The [`VisualMap::rows`] rows this media's placeholder occupies — what a
    /// capable frontend replaces with the picture or player.
    pub rows_span: Range<usize>,
    /// Whether this is a picture, a movie, or a sound — which widget the
    /// frontend builds over [`rows_span`](MediaInfo::rows_span). A frontend that
    /// handles only some kinds leaves the rest as core's placeholder rows, which
    /// already read sensibly on their own.
    pub kind: MediaKind,
    /// The media's link destination — a path, URL, or `data:` URI, verbatim from
    /// the AST. A frontend resolves a relative path against the document's own
    /// directory; core does no I/O. For a `<picture>` this is the `<img>`
    /// fallback — the source used when no [`sources`](MediaInfo::sources) media
    /// query matches (or the frontend has no theme). Empty when a `<video>`/
    /// `<audio>` carries no `src` and names its candidates in `<source>`s
    /// instead; [`resolve`](MediaInfo::resolve) already accounts for that.
    pub destination: String,
    /// The `<source>` alternatives in document order, or empty for a plain
    /// image. See [`MediaSource`]; a theme- or codec-aware frontend picks one and
    /// otherwise loads [`destination`](MediaInfo::destination).
    pub sources: Vec<MediaSource>,
    /// The media's alt text, flattened from its inline children (empty when it
    /// has none).
    pub alt: String,
    /// A `<video poster="…">`'s still frame, or empty when there is none — an
    /// image destination, resolved exactly as [`destination`] is.
    ///
    /// [`destination`]: MediaInfo::destination
    pub poster: String,
}

/// One leaf directive (`::name{…}`) as a frontend sees it: which rows its
/// placeholder occupies, its type, and its attributes. A plain surface paints
/// the `⧉ name` placeholder glyphs as-is; a frontend that knows the host app's
/// vocabulary **skips the rows in `rows_span`** and paints the real thing there,
/// exactly as an image-capable one does with [`MediaInfo`]. Derived from
/// [`VRow::leaf_directive`] by [`directive_spans`].
///
/// Core resolves nothing here — it has no idea what an `embed` or a `toc` is,
/// and deliberately so: the directive vocabulary belongs to the app on top.
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct DirectiveInfo {
    /// The [`VisualMap::rows`] rows this directive's placeholder occupies — the
    /// label row plus any blank fillers under it.
    pub rows_span: Range<usize>,
    /// The directive's type (`embed`, `toc`, `vis`), no leading colons.
    pub name: String,
    /// Its `{…}` attributes in source order; a bare one has a `None` value.
    pub attrs: Vec<(String, Option<String>)>,
    /// Its `[label]` text, flattened from its inline children (empty when it has
    /// none) — what the placeholder row shows.
    pub label: String,
}

/// One formula as a frontend sees it: where its picture goes, and the TeX to
/// typeset for it. Two shapes, told apart by [`glyph`](MathInfo::glyph):
///
/// - An **inline atom** — `$E = mc^2$` in a line of prose — is one
///   [`Role::Math`] glyph on `row` at index `glyph`, standing for the whole
///   formula. A frontend that paints pictures in a line typesets the TeX at
///   the run's font size and draws the picture in the glyph's place, as wide
///   as the picture is and with the text baseline through it at its height —
///   a run delegate on Apple, an inline element on the web. The glyph is a
///   caret stop at the formula's start; the stop after it is the next glyph's.
/// - A **display block** — a paragraph holding nothing but `$$…$$` — is the
///   placeholder row (`∑ tex`, every glyph at the formula's start) plus the
///   blank fillers `rows_span` reserves under it, exactly a [`MediaInfo`]'s
///   shape. A frontend skips those rows and paints the picture there, centred
///   on the measure.
///
/// Either way the frontend supplies the *width*: core lays out in glyphs and
/// cannot know how wide a typeset formula is, which is why an inline atom is
/// one glyph and not a run of them. Only in a **column-wrapped** build does
/// that matter to the wrap: a terminal never asks for atoms (see
/// [`Surface::inline_pictures`]), and the web re-fits a row the picture
/// overflows.
///
/// A plain surface paints the glyphs as they are — `∑` for an atom, the
/// labelled row for a block — and needs none of this. Derived from
/// [`VRow::math`] by [`math_spans`], so it survives the row reuse of
/// [`BlockCache`] and [`build_spliced`].
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct MathInfo {
    /// The [`VisualMap::rows`] rows this formula occupies: `row..row + 1` for
    /// an atom, the placeholder row and its fillers for a block.
    pub rows_span: Range<usize>,
    /// The row the atom glyph, or the block's placeholder label, is on.
    pub row: usize,
    /// The atom's index into that row's glyphs, or `None` for a block.
    pub glyph: Option<usize>,
    /// The TeX between the delimiters, verbatim.
    pub tex: String,
    /// Display style rather than text style — see [`MathMark::display`].
    pub display: bool,
    /// The formula's source start: where a click on its picture lands the
    /// caret, and the same offset every placeholder glyph carries.
    pub src: usize,
}

impl DirectiveInfo {
    /// The value of attribute `key`, if it has one with a value. The convenience
    /// a frontend reaches for first (`info.attr("src")`), since almost every
    /// directive that draws as something real is pointed at by one attribute.
    pub fn attr(&self, key: &str) -> Option<&str> {
        self.attrs
            .iter()
            .find(|(k, _)| k == key)
            .and_then(|(_, v)| v.as_deref())
    }

    /// What [`crate::Doc::set_directive_rows`] knows this directive by — the
    /// key a terminal reports the height of its drawing under.
    pub fn key(&self) -> DirectiveKey {
        DirectiveKey::new(&self.name, &self.label, &self.attrs)
    }
}

impl MediaInfo {
    /// The image URL to load under `scheme`: the first [`sources`] `<source>`
    /// whose media query matches, else the [`destination`] `<img>` fallback. The
    /// pick is a `<source>`'s first `srcset` URL or the destination — a frontend
    /// resolves whichever it gets against the document directory exactly as it
    /// resolves `destination`, and reserves/keys the picture under `destination`
    /// regardless, so a theme switch just re-picks without disturbing the layout.
    ///
    /// Only `prefers-color-scheme` is understood (that's what a light/dark banner
    /// uses); a `<source>` with any other media query is skipped, and one with no
    /// media at all always matches (an unconditional override). With no matching
    /// source — including every frontend that can't/doesn't theme and passes
    /// [`ColorScheme::Light`] to a dark-only picture — it's the plain `<img>`.
    ///
    /// [`sources`]: MediaInfo::sources
    /// [`destination`]: MediaInfo::destination
    pub fn resolve(&self, scheme: ColorScheme) -> &str {
        if let Some(url) = self
            .sources
            .iter()
            .find(|s| media_matches(&s.media, scheme))
            .and_then(|s| first_srcset_url(&s.srcset))
        {
            return url;
        }
        // A `<video>`/`<audio>` may carry no `src` of its own, naming its
        // candidates only in child `<source>`s — none of which matched above,
        // because a codec-typed `<source>` has no media query and core judges no
        // MIME types. Falling through to an empty destination would hand the
        // frontend nothing to load, so take the first candidate URL instead and
        // let the frontend reject it if it can't decode it. An `<img>` never
        // reaches this: its `src` is the picture.
        if self.destination.is_empty()
            && let Some(url) = self
                .sources
                .iter()
                .find_map(|s| first_srcset_url(&s.srcset))
        {
            return url;
        }
        &self.destination
    }

    /// The **still picture** that stands for this media under `scheme`, for a
    /// frontend that can rasterize an image but not play a movie — a terminal, or
    /// a GUI still growing its player. `None` when there is no picture to draw,
    /// which is the honest answer for audio and for a poster-less video: the
    /// caller leaves core's labelled placeholder row, which already reads as
    /// *a thing that isn't text*.
    ///
    /// This exists so those frontends never hand a `.mp4` to an image decoder.
    /// That fails harmlessly today (a failed decode falls back to the same
    /// placeholder), but it spends a file read and a decode attempt per frame to
    /// arrive where this gets in one match.
    pub fn still(&self, scheme: ColorScheme) -> Option<&str> {
        match self.kind {
            MediaKind::Image => Some(self.resolve(scheme)),
            // A `poster` is an image destination, so it resolves the same way —
            // but it is named directly and has no `<source>` alternatives of its
            // own, so it needs no theme matching.
            MediaKind::Video if !self.poster.is_empty() => Some(&self.poster),
            MediaKind::Video | MediaKind::Audio => None,
        }
    }
}

/// A frontend's active color scheme — what a `<picture>`'s `prefers-color-scheme`
/// `<source>`s are matched against by [`MediaInfo::resolve`]. A frontend with no
/// notion of theme passes [`Light`](ColorScheme::Light), the web's own default.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub enum ColorScheme {
    Light,
    Dark,
}

/// Whether a `<source media="…">` query applies under `scheme`. Empty media is
/// an unconditional `<source>` (always matches); otherwise only a
/// `prefers-color-scheme: dark|light` feature is understood — anything else
/// (a width query, `print`, …) doesn't match, so resolution falls through to the
/// next source or the `<img>`. Deliberately lax about the surrounding syntax
/// (`(prefers-color-scheme: dark)`, `screen and (prefers-color-scheme:dark)`):
/// it keys off the feature and its value, which is all the theme case needs.
fn media_matches(media: &str, scheme: ColorScheme) -> bool {
    let media = media.trim();
    if media.is_empty() {
        return true;
    }
    let lower = media.to_ascii_lowercase();
    let Some(after) = lower
        .split_once("prefers-color-scheme")
        .map(|(_, rest)| rest)
    else {
        return false;
    };
    // Skip the `:` and any spaces to reach the value word.
    let value = after.trim_start_matches([':', ' ', '\t']);
    let wanted = match scheme {
        ColorScheme::Light => "light",
        ColorScheme::Dark => "dark",
    };
    value.starts_with(wanted)
}

/// The first URL in a `srcset`: its first comma-separated candidate, before any
/// `1x`/`2x`/width descriptor. The theme case only ever puts one URL per
/// `<source>`, so the first candidate is the picture.
fn first_srcset_url(srcset: &str) -> Option<&str> {
    let first = srcset.split(',').next()?.trim();
    first.split_whitespace().next().filter(|u| !u.is_empty())
}

/// The narrowest a column may be squeezed. Below a few characters a column
/// stops carrying text and just shreds it one letter per line, which is worse
/// than letting the grid run wide.
const MIN_COL_WIDTH: usize = 3;

/// Shrink `widths` until the grid fits `avail` screen columns, taking from the
/// widest column each time so the loss is shared out rather than falling on
/// whichever column happens to be last. No column goes below
/// [`MIN_COL_WIDTH`]; a table with more columns than the surface has room for
/// still overflows, which is the honest outcome — there's nothing left to give.
fn fit_widths(widths: &mut [usize], avail: usize) {
    // Chrome: each column is its content plus a gutter either side, and every
    // column is closed by a `│` — with one more opening the row.
    let budget = avail.saturating_sub(3 * widths.len() + 1);
    while widths.iter().sum::<usize>() > budget {
        let Some(w) = widths.iter_mut().filter(|w| **w > MIN_COL_WIDTH).max() else {
            return;
        };
        *w -= 1;
    }
}

/// Word-wrap `glyphs` into lines of at most `width` columns, hard-breaking any
/// single word too long to fit.
///
/// Unlike a paragraph — where an overlong word just trails off the end of the
/// line — a table column is a hard boundary: a glyph past it lands on top of
/// the border, or on the next cell. So the width here is a promise, and a word
/// that won't keep it is broken.
///
/// The space at a break is dropped rather than hung past the edge. Its offset
/// isn't lost: the caller gives every line an end stop just past its last
/// glyph, which is exactly where that space was.
///
/// `width` is in display columns, and a break only ever falls between grapheme
/// clusters. Both matter to more than the picture: the caller anchors each
/// line's end stop just past its last glyph, so a line cut mid-cluster would
/// put a caret stop inside a character — reachable by Down or a click, and the
/// next Backspace would take the cluster apart from the middle.
///
/// An explicit in-cell break (a `\n` glyph, from a `<br>`) is a hard boundary:
/// each run between the breaks wraps on its own and the results stack. The break
/// glyphs are dropped — the caller's per-line end stop already sits exactly where
/// each break was, so no offset is lost.
fn wrap_glyphs(glyphs: &[Glyph], width: usize) -> Vec<Vec<Glyph>> {
    if glyphs.iter().any(|g| g.ch == '\n') {
        return glyphs
            .split(|g| g.ch == '\n')
            .flat_map(|seg| wrap_segment(seg, width))
            .collect();
    }
    wrap_segment(glyphs, width)
}

/// [`wrap_glyphs`] for a run with no explicit breaks — the word-wrap proper.
fn wrap_segment(glyphs: &[Glyph], width: usize) -> Vec<Vec<Glyph>> {
    let width = width.max(1);
    // Words are maximal non-space runs, each carrying the space that followed it
    // — which survives only if the next word joins it on this line.
    let mut words: Vec<(Vec<Glyph>, Option<Glyph>)> = Vec::new();
    let mut word: Vec<Glyph> = Vec::new();
    for g in glyphs {
        if g.ch == ' ' {
            words.push((std::mem::take(&mut word), Some(g.clone())));
        } else {
            word.push(g.clone());
        }
    }
    if !word.is_empty() {
        words.push((word, None));
    }

    let mut lines: Vec<Vec<Glyph>> = Vec::new();
    let mut line: Vec<Glyph> = Vec::new();
    let mut used = 0usize;
    let mut gap: Option<Glyph> = None;
    for (word, space) in words {
        for chunk in hard_break(&word, width) {
            let sep = gap.is_some() as usize;
            let cells = glyphs_width(chunk);
            if !line.is_empty() && used + sep + cells > width {
                lines.push(std::mem::take(&mut line));
                used = 0;
                gap = None; // the break swallows the space
            }
            if let Some(sp) = gap.take() {
                line.push(sp);
                used += 1;
            }
            line.extend_from_slice(chunk);
            used += cells;
        }
        gap = space;
    }
    // An empty cell is still one (empty) line — it has an end the caret can
    // sit at, which is how you type into it.
    if !line.is_empty() || lines.is_empty() {
        lines.push(line);
    }
    lines
}

/// Break a single word into pieces of at most `width` columns, cutting only
/// between grapheme clusters — the replacement for slicing it into fixed runs
/// of glyphs, which measures a wide character as one column and can cut an
/// emoji in half.
///
/// A cluster wider than the whole column still gets a piece to itself: there is
/// nowhere legal to cut it, and overflowing by a cell is better than splitting a
/// character. An empty word yields no pieces at all, which is what keeps a
/// double space from opening a line of its own.
fn hard_break(word: &[Glyph], width: usize) -> Vec<&[Glyph]> {
    let mut out = Vec::new();
    if word.is_empty() {
        return out;
    }
    let (mut start, mut used) = (0usize, 0usize);
    for c in clusters(word) {
        if used > 0 && used + c.cells > width {
            out.push(&word[start..c.glyph]);
            start = c.glyph;
            used = 0;
        }
        used += c.cells;
    }
    out.push(&word[start..]);
    out
}

/// A table rule spanning `widths`, e.g. `┌──────┬─────┐`. Each column is its
/// content width plus the one-space gutter on either side.
fn rule_text(widths: &[usize], left: char, mid: char, right: char) -> String {
    let mut s = String::new();
    s.push(left);
    for (i, w) in widths.iter().enumerate() {
        if i > 0 {
            s.push(mid);
        }
        for _ in 0..w + 2 {
            s.push('─');
        }
    }
    s.push(right);
    s
}

/// Push real document text: each glyph maps to its own source byte, and the one
/// that opens a grapheme cluster is the caret stop for the whole cluster.
///
/// Per cluster rather than per codepoint because a cluster is the character the
/// user sees, and it's the unit backspace and delete already step by. A stop
/// inside 👨‍👩‍👧 — five codepoints strung together with joiners — is a caret
/// parked in the middle of a character: one press of Right lands there, and the
/// next Backspace severs a joiner from what it joined, leaving a dangling ZWJ in
/// the source. The rest of the cluster still gets its glyph (it has to be
/// drawn); it just isn't somewhere to stand.
fn push_text(out: &mut Vec<Glyph>, text: &str, base_src: usize, style: Style) {
    for (gi, cluster) in text.grapheme_indices(true) {
        for (ci, ch) in cluster.char_indices() {
            out.push(Glyph {
                ch,
                style,
                src: base_src + gi + ci,
                stop: ci == 0,
            });
        }
    }
}

/// [`push_text`] for one line of a highlighted code block: the same glyphs at
/// the same offsets, each additionally carrying the [`Token`] of the span it
/// falls in — `spans` being the line's entry from [`code_tokens`], ascending
/// byte ranges *into `text`*. A byte between spans keeps `style` as it is.
///
/// Offsets are what matters here: a token changes how a glyph is painted and
/// nothing about where it is or which source byte it stands on, so a caret
/// walks a highlighted block exactly as it walks an unhighlighted one.
fn push_code_text(
    out: &mut Vec<Glyph>,
    text: &str,
    base_src: usize,
    style: Style,
    spans: &[(Range<usize>, Token)],
) {
    let mut spans = spans.iter().peekable();
    for (gi, cluster) in text.grapheme_indices(true) {
        // Spans are ascending, so the one covering this cluster's first byte
        // is at or after the one that covered the last; step past those ended.
        while spans.peek().is_some_and(|(r, _)| r.end <= gi) {
            spans.next();
        }
        let token = spans
            .peek()
            .filter(|(r, _)| r.contains(&gi))
            .map(|(_, t)| *t);
        // A cluster is classed whole, by its first byte: a grammar that split
        // an emoji's scalars between two tokens would otherwise split the
        // glyph, and no grammar means to.
        let style = style.token(token);
        for (ci, ch) in cluster.char_indices() {
            out.push(Glyph {
                ch,
                style,
                src: base_src + gi + ci,
                stop: ci == 0,
            });
        }
    }
}

/// One line's highlighting — `crate::syntax::LineTokens`, spelled here so the
/// shape exists whether or not the feature that fills it does.
type LineTokens = Vec<(Range<usize>, Token)>;

/// The syntax highlighting for a code block's lines, or `None` when the fence's
/// language is not one the grammars know. Without the `syntax` feature nothing
/// is known, and every code glyph draws in the plain code colour.
#[cfg(feature = "syntax")]
fn code_tokens(lang: &str, lines: &[&str]) -> Option<Vec<LineTokens>> {
    crate::syntax::highlight(lang, lines)
}

#[cfg(not(feature = "syntax"))]
fn code_tokens(_lang: &str, _lines: &[&str]) -> Option<Vec<LineTokens>> {
    None
}

/// Emit an inline `str`/`smart_punctuation` run, mapping every visible char back
/// to its *true* source byte even when the source carries backslash escapes the
/// parsed `text` dropped (`\*` → `*`). The naive `span.start + text_offset`
/// mapping [`push_text`] uses drifts by one byte after each escape, so a caret or
/// click past an escaped `*` would land on the wrong character; walking the text
/// against its source keeps them aligned, and the hidden escape backslash gets no
/// glyph of its own (it is a spelling artefact, not something the caret lands on).
fn push_escaped_text(
    out: &mut Vec<Glyph>,
    text: &str,
    span: Range<usize>,
    source: &str,
    style: Style,
) {
    let end = span.end.min(source.len());
    let src = source.get(span.start..end).unwrap_or("");
    // Fast path — no dropped bytes, so text and source align 1:1 (the common
    // case: prose with no escapes). Byte lengths equal ⇒ no backslash was eaten.
    if src.len() == text.len() {
        push_text(out, text, span.start, style);
        return;
    }
    // Slow path: some `\` was consumed. Walk char-by-char, skipping a backslash
    // in the source exactly when it escapes the next visible char (a real escape),
    // never when it is a literal backslash the parse kept (that case has equal
    // lengths and takes the fast path above).
    let sb = src.as_bytes();
    let mut si = 0usize;
    'text: for (_, cluster) in text.grapheme_indices(true) {
        for (ci, ch) in cluster.char_indices() {
            // The text outlasted the source it is being mapped onto. In a
            // consistent document that cannot happen on this path: the slow path
            // is only entered when the two lengths differ, and everything that
            // makes them differ makes the *source* the longer one — an escape
            // backslash the parse ate, or source folded into a neighbouring node.
            // A `smart_punctuation` node reports its canonical ASCII spelling
            // (`--`, `...`, `"`), which is never longer than what was written.
            //
            // So reaching here means `span` was measured against a document that
            // `source` is no longer, and there is no honest offset left to give
            // the remaining characters. Stop: the row comes out short, which is
            // a wrong picture of a document that is already inconsistent. The
            // alternative was `si` stepping past the end and the slice below
            // panicking — which is what it did, in a paint loop.
            if si >= sb.len() {
                break 'text;
            }
            // Advance to the source character this one came from, stepping over
            // whatever the parse dropped on the way. An escape backslash is the
            // common case, but not the only one: a span can cover source that
            // was folded into a neighbouring node (smart punctuation next to a
            // bracket gives `text: "]"` over a source span of `"…]"`). Advancing
            // by the *text* character's length assumed escapes were the only
            // divergence, so one dropped multi-byte character desynchronized
            // every glyph after it — placing `]` inside the `…` before it.
            while si < sb.len() && !src[si..].starts_with(ch) {
                si += src[si..].chars().next().map_or(1, char::len_utf8);
            }
            out.push(Glyph {
                ch,
                style,
                src: span.start + si.min(src.len()),
                stop: ci == 0,
            });
            si += src[si..]
                .chars()
                .next()
                .map_or(ch.len_utf8(), char::len_utf8);
        }
    }
}

/// Build synthetic decoration glyphs (a bullet, a gutter) all pointing at `src`,
/// each carrying `role` so the frontend can style it (`Role::Body` for plain
/// padding). Synthetic glyphs are never caret stops — they share one offset, so
/// the caret steps over them (a click still lands at `src`).
fn synth(text: &str, role: Role, src: usize) -> Vec<Glyph> {
    let style = Style::default().role(role);
    text.chars()
        .map(|ch| Glyph {
            ch,
            style,
            src,
            stop: false,
        })
        .collect()
}

fn concat(a: &[Glyph], b: &[Glyph]) -> Vec<Glyph> {
    let mut v = a.to_vec();
    v.extend_from_slice(b);
    v
}

/// The columns a row's prefix (a bullet, a quote gutter, an indent) takes up
/// before the text it introduces — what the wrap budget has left to spend.
fn prefix_width(prefix: &[Glyph]) -> usize {
    glyphs_width(prefix)
}

/// The label shown for an image with no alt text: the final path segment of its
/// destination (`img/cat.png` → `cat.png`), the whole destination when it has no
/// separator, and `"image"` when it's empty. A `data:` URI (which has no useful
/// tail) shows its scheme so the placeholder isn't a wall of base64.
fn media_label(dest: &str) -> String {
    if dest.is_empty() {
        return "image".to_string();
    }
    if dest.starts_with("data:") {
        return "data:…".to_string();
    }
    // Trim a query/fragment so a URL's `?v=2#frag` doesn't ride along.
    let clean = dest.split(['?', '#']).next().unwrap_or(dest);
    let tail = clean
        .trim_end_matches('/')
        .rsplit(['/', '\\'])
        .next()
        .unwrap_or(clean);
    if tail.is_empty() {
        dest.to_string()
    } else {
        tail.to_string()
    }
}

/// A directive's attributes read as a human label — what a frontend puts on a
/// container's tinted panel, and what an attribute-bearing inline directive
/// shows in its chip.
///
/// Reads BOTH conventions diaryx content actually uses: twig's own dot-prefixed
/// classes (`{.public .family}`, arriving as one combined `class` attr) and bare
/// pandoc-style words with no leading dot (`{public family}` — what
/// `diaryx_core::visibility`'s publish-time filter and apps/web's directive
/// serializer both write, and which twig parses as one valueless attribute
/// each). Reading only `.class` would leave every *existing* diaryx `:::vis{…}`
/// block unlabeled. A `key=value` attr is configuration rather than a name, so
/// it contributes nothing. `None` when nothing readable is left.
fn directive_attr_label(attrs: &[(String, Option<String>)]) -> Option<String> {
    let mut parts: Vec<String> = Vec::new();
    for (k, v) in attrs {
        if k == "class" {
            if let Some(v) = v
                && !v.is_empty()
            {
                parts.push(v.clone());
            }
        } else if v.as_deref().unwrap_or("").is_empty() {
            parts.push(k.clone());
        }
    }
    (!parts.is_empty()).then(|| parts.join(" "))
}

fn heading_style(level: u32) -> Style {
    // Just the role — a frontend decides how a heading of this level *looks*
    // (the terminal cycles a color and bolds it, the GUI scales the font). The
    // author wrote no emphasis here, so core records none. `level as u8` is safe:
    // Markdown/Djot cap headings at 6.
    Style::default().role(Role::Heading(level.min(255) as u8))
}

/// Is this `container` node a *directive* (`:::note{…}`, `::embed{…}`,
/// `:vis[…]`) rather than an HTML element (`<video>`, `<picture>`, `<div>`)?
///
/// twig 2.8 folded `div`/`span`/`directive`/`element` into one `container` kind,
/// and left nothing that separated them: `kind`, `name` and `directive_form` all
/// agree, field for field, on an HTML `<div>` and a Markdown `:::div`. Leaf
/// answered it by sniffing the span for whichever of `:` or `<` came first.
/// twig 3.0 records the answer at parse time as [`ContainerOrigin`], so this is
/// now the parser's own knowledge rather than a guess rebuilt from the bytes it
/// consumed.
pub(crate) fn container_is_directive(node: &FlatNode) -> bool {
    node.origin == Some(ContainerOrigin::Directive)
}

/// The tag a `container` node carries when it is an HTML element rather than a
/// directive — `Some("video")` for a promoted `<video>`, `None` for a `:::note`
/// or for any node that is not a container at all.
pub(crate) fn element_tag(node: &FlatNode) -> Option<&str> {
    (node.origin == Some(ContainerOrigin::Element))
        .then_some(node.name.as_deref())
        .flatten()
}

/// A leaf directive's name and whatever attributes are not part of spelling
/// it — the two things a [`DirectiveMark`] carries, which twig hands back
/// differently per format and which a frontend must not be able to tell apart.
///
/// Markdown's `::page-break` is a `Leaf`-form directive *named* `page-break`
/// with no attributes, and this returns it verbatim. Djot has no leaf form:
/// `insert_directive` writes the same document as an empty `::: page-break`
/// fence, whose container is anonymous (a djot div carries no name) and whose
/// name arrives as a class. Where the node has no name of its own, then, the
/// name is the word on the fence line when there is one — djot appends it
/// *after* the classes of an attribute line above (`{.wide}` over `:::
/// x-card` reads back as `class="wide x-card"`), so it is found by what the
/// fence says rather than by position — and otherwise, for a bare `:::` under
/// `{.page-break .wide}`, the first class. Whatever else the class said stays
/// an attribute.
fn leaf_directive_identity(
    node: &FlatNode,
    source: &str,
) -> (String, Vec<(String, Option<String>)>) {
    let named = node.name.clone().unwrap_or_default();
    if !named.is_empty() {
        return (named, node.attrs.clone());
    }
    let class = node
        .attrs
        .iter()
        .find(|(k, _)| k == "class")
        .and_then(|(_, v)| v.as_deref())
        .unwrap_or_default();
    let mut tokens = class.split_whitespace().collect::<Vec<_>>();
    let fence_word = source
        .get(node.span.start..)
        .and_then(|rest| rest.lines().next())
        .map(|line| line.trim_start_matches(['>', ' ', '\t', ':']).trim())
        .and_then(|word| word.split_whitespace().next());
    let at = fence_word
        .and_then(|word| tokens.iter().rposition(|t| *t == word))
        .unwrap_or(0);
    if tokens.is_empty() {
        return (named, node.attrs.clone());
    }
    let name = tokens.remove(at).to_string();
    let rest = tokens.join(" ");
    let attrs = node
        .attrs
        .iter()
        .filter_map(|(k, v)| {
            if k != "class" {
                return Some((k.clone(), v.clone()));
            }
            (!rest.is_empty()).then(|| (k.clone(), Some(rest.clone())))
        })
        .collect();
    (name, attrs)
}

/// Is this inline `container` an **attributed span** — the node leaf's run-level
/// vocabulary rides — rather than a named directive?
///
/// The four formats spell one span four ways and twig hands the name back for
/// two of them: HTML's and Markdown's `<span …>` arrive named `span` with
/// `Element` origin, while djot's `[text]{…}` and AsciiDoc's `[.a]#text#`
/// arrive anonymous (an empty name) with `Directive` origin. All four are the
/// same node to `wrap_range_attrs`, which is what writes them, so they are the
/// same node here.
///
/// A *named* directive is not one, whatever its name: a Markdown `:span[…]{…}`
/// is a directive the parser read as a directive, twig's own
/// `wrap_range_attrs` says so, and it keeps the handling it has.
///
/// **Anonymous is not enough**, and the form is what finishes the question:
/// a djot fenced div (`{.center}` / `:::` / … / `:::`) is anonymous too, with
/// the same `Directive` origin, and is a *block* — `Container` form against the
/// span's `Text`. Reading one as a span made every gesture and every query lie
/// about it: `set_text_color` over a word inside such a div copied the whole
/// div's attribute set — its `id` along with the rest — onto the new span, and
/// `alignment_at_caret` reported the div's `.center` as a *run's* answer while
/// the walker drew none. So the anonymous arm asks the form [`is_inline`] asks.
pub(crate) fn is_run_span(node: &FlatNode) -> bool {
    if node.kind != Kind::Container {
        return false;
    }
    match node.name.as_deref() {
        None | Some("") => node.directive_form == Some(DirectiveForm::Text),
        Some("span") => node.origin == Some(ContainerOrigin::Element),
        Some(_) => false,
    }
}

/// `base` with an attributed span's three run-level keys written over it — the
/// nearest-wins fold [`is_run_span`] describes, for one span. `faces` is the
/// build's intern table, as it is for [`Presentation::under`].
fn run_style(node: &FlatNode, base: Style, faces: &RefCell<FaceTable>) -> Style {
    Style {
        size: FontSize::from_attrs(&node.attrs).or(base.size),
        font: faces
            .borrow_mut()
            .face_from_attrs(&node.attrs)
            .or(base.font),
        color: TextColor::from_attrs(&node.attrs).or(base.color),
        ..base
    }
}

pub(crate) fn is_inline(node: &FlatNode) -> bool {
    // A directive is inline only in its `text` form (`:name[label]{…}`); the
    // `leaf` and `container` forms are blocks. All three report the same `kind`,
    // so the form is the only thing telling them apart — and getting it wrong
    // costs a whole paragraph: a text directive misread as a block makes its
    // paragraph fail the "all children inline" test in `block`, and the line is
    // then walked as a container of blocks, rendering as empty rows with no
    // caret home at all.
    //
    // An HTML element shares the `container` kind, and twig sets the same form
    // on the two tags the lightweight formats have a generic spelling for: a
    // `<span>` is `Text` and a `<div>` is `Container`, while a `<video>` or a
    // `<picture>` has no form at all. So the form answers for an element as it
    // answers for a directive, and the origin is not consulted — which is what
    // makes a `<span …>` inside a paragraph an inline node.
    //
    // It has to. `wrap_range_attrs` spells an attributed run as exactly that
    // span in Markdown and HTML, and a paragraph holding one whose kids were
    // not all inline failed the test below and was walked as a container of
    // blocks: the text either side of the span rendered as nothing at all.
    if node.kind == Kind::Container {
        return node.directive_form == Some(DirectiveForm::Text);
    }
    is_inline_kind(&node.kind)
}

/// [`is_inline`] by kind alone — for the ancestor walks, whose `QueryMatch`es
/// carry no `directive_form`. It answers `false` for every directive, which its
/// callers must (and do) reconcile: they pair it with `is_block_container`,
/// which claims every directive, so the pair's verdict is the same one a form
/// would have given. Anything looking at a *directive itself* wants [`is_inline`]
/// and a real node.
pub(crate) fn is_inline_kind(kind: &Kind) -> bool {
    matches!(
        kind,
        Kind::Str
            | Kind::SoftBreak
            | Kind::HardBreak
            | Kind::NonBreakingSpace
            | Kind::Emph
            | Kind::Strong
            | Kind::Mark
            | Kind::Insert
            | Kind::Delete
            | Kind::Verbatim
            | Kind::InlineMath
            | Kind::DisplayMath
            | Kind::Url
            | Kind::Email
            | Kind::Link
            | Kind::Image
            | Kind::SmartPunctuation
            | Kind::Superscript
            | Kind::Subscript
            | Kind::FootnoteReference
    )
}

/// Assert two maps are identical down to every glyph, stop, and table span — the
/// contract `build_cached` and `build_spliced` must hold against `build`. Lives
/// at module scope (not in `mod tests`) so the Doc-driven differential test in
/// `doc.rs` can reach it and the private `stops` field it compares.
#[cfg(test)]
pub(crate) fn assert_maps_eq(a: &VisualMap, b: &VisualMap, ctx: &str) {
    assert_eq!(a.rows.len(), b.rows.len(), "row count ({ctx})");
    for (i, (ra, rb)) in a.rows.iter().zip(&b.rows).enumerate() {
        assert_eq!(ra.end_src, rb.end_src, "row {i} end_src ({ctx})");
        assert_eq!(ra.decoration, rb.decoration, "row {i} decoration ({ctx})");
        // The incremental walk labels a boundary from a query match's kind
        // string and the whole-arena walk from a `FlatNode`'s; this is what says
        // the two doors reach the same answer.
        assert_eq!(ra.boundary, rb.boundary, "row {i} boundary ({ctx})");
        assert_eq!(ra.code, rb.code, "row {i} code ({ctx})");
        assert_eq!(ra.code_lang, rb.code_lang, "row {i} code_lang ({ctx})");
        assert_eq!(ra.align, rb.align, "row {i} align ({ctx})");
        assert_eq!(
            ra.line_height, rb.line_height,
            "row {i} line_height ({ctx})"
        );
        assert_eq!(
            ra.glyphs.len(),
            rb.glyphs.len(),
            "row {i} glyph count ({ctx})"
        );
        for (j, (ga, gb)) in ra.glyphs.iter().zip(&rb.glyphs).enumerate() {
            assert_eq!(
                (ga.ch, ga.src, ga.stop, ga.style),
                (gb.ch, gb.src, gb.stop, gb.style),
                "row {i} glyph {j} ({ctx})"
            );
        }
    }
    assert_eq!(a.content_start, b.content_start, "content_start ({ctx})");
    assert_eq!(a.stops, b.stops, "stops ({ctx})");
    assert_eq!(a.mark_ends, b.mark_ends, "mark_ends ({ctx})");
    assert_eq!(a.tables.len(), b.tables.len(), "table count ({ctx})");
    for (i, (ta, tb)) in a.tables.iter().zip(&b.tables).enumerate() {
        assert_eq!(ta.rows_span, tb.rows_span, "table {i} rows_span ({ctx})");
        assert_eq!(ta.end_src, tb.end_src, "table {i} end_src ({ctx})");
    }
    assert_eq!(a.code_blocks, b.code_blocks, "code_blocks ({ctx})");
    assert_eq!(a.media, b.media, "images ({ctx})");
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::style::{FontFamily, LineSpacing, SizeStep};
    use twig::{Editor, Format, NodeId};

    fn map(src: &str) -> VisualMap {
        let mut ed = Editor::new_str(src, Format::Markdown).unwrap();
        build_t(&ed.nodes().unwrap(), src, Some(80))
    }

    /// [`map`] over a Djot source. Djot is the format that spells superscript
    /// and subscript at all — Markdown has no syntax for either.
    fn map_djot(src: &str) -> VisualMap {
        let mut ed = Editor::new_str(src, Format::Djot).unwrap();
        build_t(&ed.nodes().unwrap(), src, Some(80))
    }

    /// The baseline every glyph spelling `ch` was built with, in row order —
    /// how a test reads a raised or lowered run off the map without caring
    /// which row it landed on.
    fn baselines_of(m: &VisualMap, ch: char) -> Vec<Baseline> {
        m.rows
            .iter()
            .flat_map(|r| r.glyphs.iter())
            .filter(|g| g.ch == ch)
            .map(|g| g.style.baseline)
            .collect()
    }

    /// [`map`] at a chosen wrap width.
    fn map_at(src: &str, wrap: Option<usize>) -> VisualMap {
        let mut ed = Editor::new_str(src, Format::Markdown).unwrap();
        build_t(&ed.nodes().unwrap(), src, wrap)
    }

    /// [`map`], but with twig's `directives` extension on (off by twig's own
    /// default) — the `:::name{.class}` fenced-div containers leaf-core's
    /// `"directive"` wysiwyg arm renders.
    fn map_directives(src: &str) -> VisualMap {
        let mut ed = Editor::new_ext(
            src.as_bytes(),
            Format::Markdown,
            twig::MarkdownExtensions {
                directives: true,
                ..Default::default()
            },
        )
        .unwrap();
        build_t(&ed.nodes().unwrap(), src, Some(80))
    }

    /// [`map`] in `format`, parsed the way every leaf document is — the
    /// extensions [`crate::doc::parse_extensions`] turns on, which is what
    /// pairs a Markdown `<div …>` with its `</div>` into a container and makes
    /// `::page-break` a directive rather than a paragraph of colons.
    fn map_leaf(src: &str, format: Format) -> VisualMap {
        let mut ed =
            Editor::new_ext(src.as_bytes(), format, crate::doc::parse_extensions()).unwrap();
        build_t(&ed.nodes().unwrap(), src, Some(80))
    }

    /// The alignment and line spacing of every row that draws text, in order —
    /// how a test reads a block property off the map.
    fn line_facts(m: &VisualMap) -> Vec<(Option<Align>, Option<LineHeight>)> {
        m.rows
            .iter()
            .filter(|r| r.glyphs.iter().any(|g| !g.ch.is_whitespace()))
            .map(|r| (r.align, r.line_height))
            .collect()
    }

    /// The style of the glyph spelling `ch`, first occurrence — how a test reads
    /// a run property off the map.
    fn style_of(m: &VisualMap, ch: char) -> Style {
        m.rows
            .iter()
            .flat_map(|r| r.glyphs.iter())
            .find(|g| g.ch == ch)
            .unwrap_or_else(|| panic!("no glyph {ch:?} in the map"))
            .style
    }

    /// [`map`] with soft breaks preserved (`LineFlow::Preserve`).
    fn map_preserve(src: &str, wrap: Option<usize>) -> VisualMap {
        let mut ed = Editor::new_str(src, Format::Markdown).unwrap();
        build(
            &ed.nodes().unwrap(),
            src,
            wrap,
            true,
            &Surface::default(),
            None,
        )
    }

    /// The cache-free reference [`build`], with no per-image height overrides —
    /// every block image stays its default one-row placeholder. The tests that
    /// need a taller image drive it through [`crate::Doc::set_media_rows`] instead.
    fn build_t(nodes: &[FlatNode], src: &str, wrap: Option<usize>) -> VisualMap {
        build(nodes, src, wrap, false, &Surface::default(), None)
    }

    /// An arena and a string that disagree — spans reaching past the source they
    /// are built against.
    ///
    /// [`crate::Doc`] keeps the two in step, so this is a "cannot happen" that
    /// nonetheless *did*: `examples/bench` timed a loop of `edit_range` and then
    /// went on handing the grown editor's spans to a builder holding the string
    /// from before it, and every run ended in a slice panic rather than a
    /// number. `push_escaped_text` was already written to survive the mismatch —
    /// it clamps the span's end and falls back to an empty slice — and this is
    /// the half of that intent it did not carry through.
    ///
    /// Rendering the wrong thing is the acceptable answer here; panicking in a
    /// paint loop is not.
    #[test]
    fn a_source_shorter_than_the_arena_built_over_it_renders_rather_than_panicking() {
        // An escape puts the run on `push_escaped_text`'s slow path — the fast
        // path is a length comparison that a truncated source fails anyway.
        let src = "alpha \\*beta\\* gamma delta epsilon\n";
        let mut ed = Editor::new_str(src, Format::Markdown).unwrap();
        let nodes = ed.nodes().unwrap();

        // Every truncation of it, so the cut lands before, inside and after the
        // escaped run rather than only where one hand-picked index put it.
        for cut in 0..=src.len() {
            if !src.is_char_boundary(cut) {
                continue;
            }
            let map = build_t(&nodes, &src[..cut], Some(80));
            for row in &map.rows {
                for g in &row.glyphs {
                    assert!(
                        g.src <= src.len(),
                        "cut {cut}: glyph {:?} points past the source at {}",
                        g.ch,
                        g.src
                    );
                }
            }
        }
    }

    fn rendered(m: &VisualMap) -> String {
        m.rows
            .iter()
            .map(|r| r.glyphs.iter().map(Glyph::drawn).collect::<String>())
            .collect::<Vec<_>>()
            .join("\n")
    }

    /// Render a source both ways: `build` over the whole marshalled arena (the
    /// reference), and `build_cached` driven the way [`crate::Doc`] drives it —
    /// top-level blocks from `child_spans`, per-block subtrees on a miss.
    fn render_both(
        ed: &mut Editor,
        src: &str,
        wrap: Option<usize>,
        cache: &mut BlockCache,
    ) -> (VisualMap, VisualMap) {
        let all = ed.nodes().unwrap();
        let surface = Surface::default();
        let plain = build(&all, src, wrap, false, &surface, None);
        let top = top_blocks(ed);
        let cached = build_cached(&top, src, wrap, false, &surface, None, cache, |id| {
            ed.subtree(NodeId(id)).unwrap_or_default()
        });
        (plain, cached)
    }

    /// The whole correctness claim of the block cache: `build_cached` produces a
    /// byte-identical map to `build`, on a fresh cache *and* — the case that
    /// actually exercises reuse-and-shift plus per-block subtree marshalling — on
    /// a warm cache after the source has been edited underneath it.
    /// **Every glyph must stand on the character it claims.** A row's source
    /// extent is computed from its last glyph's offset, so a glyph carrying an
    /// offset that is not its own character's start yields a row end inside a
    /// multi-byte character — and every later slice of the source panics on it.
    ///
    /// Reproduces a real crash from a journal entry: a bracketed elision inside
    /// a blockquote (`[…]`) gave the closing bracket a `text` of `"]"` over a
    /// source span covering `"…]"`, because the parse folded the ellipsis into a
    /// neighbouring node. `push_escaped_text` walked that span assuming a
    /// dropped backslash was the only way text and source could diverge, so the
    /// `]` landed on the `…`'s first byte:
    /// `byte index 1236 is not a char boundary; it is inside '…'`.
    #[test]
    fn a_glyph_never_lands_inside_the_character_before_it() {
        let src = "> engage with it rather than look away. […]\n>\n> The through-line\n";
        let vmap = map(src);
        for (r, row) in vmap.rows.iter().enumerate() {
            assert!(
                src.is_char_boundary(row.end_src.min(src.len())),
                "row {r} ends at {} — inside a character",
                row.end_src
            );
            for g in &row.glyphs {
                assert!(
                    src.is_char_boundary(g.src.min(src.len())),
                    "row {r} has {:?} at {}, which is inside a character",
                    g.ch,
                    g.src
                );
            }
        }
        // The elision survives, and its bracket sits on the real `]`.
        let text: String = vmap
            .rows
            .iter()
            .flat_map(|r| r.glyphs.iter().map(|g| g.ch))
            .collect();
        assert!(text.contains("[…]"), "the elision should render: {text:?}");
        let close = vmap
            .rows
            .iter()
            .flat_map(|r| r.glyphs.iter())
            .find(|g| g.ch == ']')
            .expect("a closing bracket");
        assert_eq!(
            src[close.src..].chars().next(),
            Some(']'),
            "the bracket glyph should stand on the source's own `]`"
        );
    }

    #[test]
    fn build_cached_matches_build() {
        let docs = [
            "# Title\n\nThe quick brown fox.\n\nAnother paragraph here.\n",
            "## H\n\n- one\n- two\n- three\n\n> a quote\n> continued\n",
            "para one\n\n```\ncode\nlines\n```\n\nafter code\n",
            "| a | b |\n|---|---|\n| 1 | 2 |\n\ntext after a table\n",
            "line\n- \nsetext?\n\nreal para\n\n\n\ntrailing blanks\n",
            "> quote with **bold** and a [link](https://x.dev)\n>\n> - item\n> - item2\n\ntail\n",
            "intro\n\n![a cat](img/cat.png)\n\nbetween\n\n![](https://x.dev/logo.svg)\n\nend\n",
            "- text item\n- ![alt](pic.png)\n- more text\n",
            // Footnotes: twig parses each definition as a root beside `doc`, so
            // these are the docs where the reference build and the incremental
            // one could disagree about what the top-level blocks even are.
            "A claim[^1] and another[^src].\n\n[^1]: First note.\n\n[^src]: Second.\n\ntail\n",
            "note[^a]\n\n[^a]: body **bold**\n    wrapped on\n    three lines\n\nafter\n",
            // No trailing newline. twig closes the document's last block on the
            // virtual newline it supplies at EOF, so that block's `span.end` is
            // `source.len() + 1` — a range that slices no bytes at all. Keying
            // the block cache off such a slice made every last block hash alike;
            // see [`block_bytes`].
            "# Title\n\nThe quick brown fox.\n\nA tail with no newline",
            "A claim[^1] and another[^src].\n\n[^1]: First note.\n[^src]: Second, ending the file.",
            // Comments draw nothing. The per-block builder the cached path
            // renders one with starts at offset 0 and, drawing nothing, never
            // moved — so the walk went on from 0 and spelled every line of the
            // document as a blank row. One at the start, one between blocks,
            // one at the end, so each position is covered.
            "<!-- lead -->\n\npara\n\n<!-- exec -->\n```\ncode\n```\n\nafter\n\n<!-- trail -->\n",
            // A Markdown `<div>` ends with a hidden `</div>` line the walk
            // steps over — between blocks and closing the file, so both the
            // separator after it and the trailing count are covered.
            "above\n\n<div class=\"center\">\n\nhello\n\n</div>\n\nbelow\n",
            "above\n\n<div class=\"center\">\n\nhello\n\nworld\n\n</div>\n",
            // Link reference definitions: roots beside `doc` like footnotes,
            // but drawing nothing. Alone between blocks, glued under a
            // paragraph, and closing the file under a comment — the README
            // shape.
            "see [a] and [b]\n\n[a]: /a\n\nmid\n[b]: /b \"bee\"\n\nend [c]\n\n<!-- links -->\n[c]: /c\n",
            // A rule's row ends past the newline under it while the walk
            // stands at the rule, and the rule's block is never cached for
            // it: an empty line under one mid-document and closing it.
            "para\n\n---\n\n\n\nbetween rules\n\n---\n\n",
            // Blank lines above the first block draw rows: bare, past
            // frontmatter, and past a comment that draws nothing.
            "\n\n\nfirst\n\nsecond\n",
            "---\ntitle: x\n---\n\n\nafter frontmatter\n",
            "<!-- lead -->\n\n\n\nafter a comment\n",
            // An empty paragraph at the end of a div draws inside it.
            "<div data-line-height=\"1.5\">\n\nI cry\n\nknees\n\n\n\n</div>\n\nafter\n",
        ];
        for wrap in [None, Some(80usize), Some(20)] {
            for src in docs {
                let ctx = format!("wrap={wrap:?} src={src:?}");
                let mut ed = Editor::new_str(src, Format::Markdown).unwrap();
                let mut cache = BlockCache::default();

                // 1) Fresh cache equals the cache-free build.
                let (plain, cached) = render_both(&mut ed, src, wrap, &mut cache);
                assert_maps_eq(&plain, &cached, &format!("fresh {ctx}"));

                // 2) Type a char mid-document, reparse, rebuild with the now-warm
                //    cache: the edited block is re-marshalled and re-rendered,
                //    every block below it is reused shifted, and the result must
                //    still match a from-scratch build.
                let at = (src.len() / 2..=src.len())
                    .find(|&i| src.is_char_boundary(i))
                    .unwrap();
                ed.edit_range(at, at, "Z").unwrap();
                let src2 = ed.source_str().unwrap();
                let (plain2, cached2) = render_both(&mut ed, &src2, wrap, &mut cache);
                assert_maps_eq(&plain2, &cached2, &format!("after insert {ctx}"));

                // 3) Delete it again: offsets shift back the other way, and the
                //    warm cache must not hand back stale shifted rows.
                ed.edit_range(at, at + 1, "").unwrap();
                let src3 = ed.source_str().unwrap();
                let (plain3, cached3) = render_both(&mut ed, &src3, wrap, &mut cache);
                assert_maps_eq(&plain3, &cached3, &format!("after delete {ctx}"));
            }
        }
    }

    /// A document that does not end in a newline is the one place twig hands
    /// leaf a top-level span that addresses no source: the last block is closed
    /// on the virtual newline the parser supplies at EOF, so its `span.end` is
    /// `source.len() + 1`. The block cache keys on the bytes under that span, and
    /// reading the out-of-range slice as *no bytes* broke it two ways at once —
    /// [`block_bytes`] has the full account. Both ways are checked here, because
    /// they fail independently.
    #[test]
    fn a_block_running_past_the_last_byte_still_keys_the_cache_by_its_own_bytes() {
        // One: two overrunning blocks collide. A footnote definition is a root
        // beside `doc` that [`top_blocks`] merges into the top level, while the
        // `section` above it spans the definition's bytes too — so when the
        // definition ends the file, both blocks end past it. The second was
        // served the first's rows, and the definition rendered as a copy of the
        // heading.
        let src = "A claim[^1] worth checking.\n\n# A heading with a reference[^1] in it\n\n[^1]: The first note.\n[^note]: A note with a word for a label.";
        let mut ed = Editor::new_str(src, Format::Djot).unwrap();
        let (plain, cached) = render_both(&mut ed, src, Some(80), &mut BlockCache::default());
        assert_maps_eq(&plain, &cached, "a definition ending the file");
        let text = rendered(&cached);
        assert!(
            text.ends_with("[note] A note with a word for a label."),
            "the last definition should render itself: {text:?}"
        );
        assert_eq!(
            text.matches("A heading with a reference").count(),
            1,
            "the heading should render exactly once: {text:?}"
        );

        // Two: one overrunning block goes stale. Its bytes are its cache key, so
        // a block that keeps hashing the same however it is edited is served the
        // rows built before the edit — the whole last line frozen as the user
        // types in it.
        let mut cache = BlockCache::default();
        let first = "first para\n\n# A heading\n\nlast para with no newline";
        let mut ed = Editor::new_str(first, Format::Djot).unwrap();
        let (_, warm) = render_both(&mut ed, first, Some(80), &mut cache);
        assert!(rendered(&warm).ends_with("last para with no newline"));

        let second = "first para\n\n# A heading\n\nDIFFERENT text without a newline";
        let mut ed = Editor::new_str(second, Format::Djot).unwrap();
        let (plain, cached) = render_both(&mut ed, second, Some(80), &mut cache);
        assert_maps_eq(&plain, &cached, "edited last block, warm cache");
        let text = rendered(&cached);
        assert!(
            text.ends_with("DIFFERENT text without a newline"),
            "the warm cache served the pre-edit rows: {text:?}"
        );
    }

    #[test]
    fn resolves_markup_to_plain_text() {
        let text = rendered(&map("# Title\n\na **bold** word\n"));
        assert!(!text.contains('#'), "heading marker shown: {text:?}");
        assert!(!text.contains("**"), "strong delimiters shown: {text:?}");
        assert!(text.contains("Title") && text.contains("bold word"));
    }

    #[test]
    fn every_glyph_points_at_its_source_byte() {
        let src = "a **bold** c\n";
        let m = map(src);
        for row in &m.rows {
            for g in &row.glyphs {
                // A real (non-synthetic) glyph's source byte is the glyph's char.
                if g.src < src.len()
                    && src.is_char_boundary(g.src)
                    && let Some(sc) = src[g.src..].chars().next()
                    && sc == g.ch
                {
                    continue;
                }
                // Synthetic prefixes (none here) would be the only exceptions.
                panic!("glyph {:?} at src {} doesn't match source", g.ch, g.src);
            }
        }
    }

    #[test]
    fn offset_and_position_round_trip_on_visible_text() {
        let m = map("hello world\n");
        let (r, c) = m.pos_of_offset(6); // the 'w'
        assert_eq!(m.offset_of_pos(r, c), 6);
    }

    #[test]
    fn visible_utf16_indices_count_the_text_the_system_sees() {
        // Hidden delimiters, a two-unit emoji, and a block gap — every way the
        // visible text's UTF-16 length parts company with a source byte count.
        let src = "a **b\u{1F600}** c\n\nd\n";
        let m = map(src);
        let end = m.snap_to_stop(src.len());
        let text = m.visible_text(0, end);
        assert_eq!(text, "a b\u{1F600} c\nd");

        // Forward: the index of each offset is where that character sits in
        // the visible string, in UTF-16 units.
        for (i, (src_off, _)) in m.visible_items(0, end).iter().enumerate() {
            let expect: usize = text.chars().take(i).map(char::len_utf16).sum();
            assert_eq!(
                m.visible_utf16_len(0, *src_off),
                expect,
                "utf16 index of source offset {src_off}"
            );
            // And back: the index resolves to the offset it came from.
            assert_eq!(m.offset_at_visible_utf16(end, expect), Some(*src_off));
        }
        // Inside the emoji's surrogate pair resolves to the emoji.
        let emoji_src = src.find('\u{1F600}').unwrap();
        let emoji_idx = m.visible_utf16_len(0, emoji_src);
        assert_eq!(
            m.offset_at_visible_utf16(end, emoji_idx + 1),
            Some(emoji_src)
        );
        // At or past the end is nobody's character.
        let total = m.visible_utf16_len(0, end);
        assert_eq!(total, text.encode_utf16().count());
        assert_eq!(m.offset_at_visible_utf16(end, total), None);
    }

    /// The documents the lookups are checked against their reference scans
    /// on: every shape that puts rows out of source order or a stop out of
    /// step with a glyph. A wrapped table, whose cells' second lines sit
    /// below the next column's first; a wide grapheme and an emoji outside
    /// the BMP; hidden delimiters and a link's hidden destination; a list
    /// with synthetic markers; a code block; an image row whose label glyphs
    /// all share one offset; an empty paragraph, a heading, and a wrapped
    /// paragraph — at a narrow width so the table and the prose both wrap,
    /// and unwrapped, which is how a GUI builds it.
    fn lookup_maps() -> Vec<(String, VisualMap)> {
        let table = "| left cell that wraps | right |\n|---|---|\n| a longer cell than the column can hold | b |\n| `k` | 你好 |\n\n";
        let srcs = [
            "hello world\n",
            "a **b\u{1F600}** c\n\nd\n",
            "- one *two*\n- three\n\n\n# Title\n\nend [link](https://e.org/x) tail\n",
            &format!("{table}```\nx\ny\n```\n\n![alt](a.png)\n\ntail 你好 **bold** here\n"),
            "one two three four five six seven eight nine ten eleven twelve thirteen\n\n| a | b |\n|-|-|\n| c d e f g h | i |\n",
        ];
        let mut out = Vec::new();
        for src in srcs {
            for wrap in [Some(14), None] {
                out.push((format!("{src:?} at {wrap:?}"), map_at(src, wrap)));
            }
        }
        out
    }

    /// `pos_of_offset` as it was written before the walk learnt to dismiss a
    /// row from its ends: every row read through, the nearest stop kept.
    fn pos_of_offset_by_scan(m: &VisualMap, off: usize) -> (usize, usize) {
        let mut best: Option<(usize, usize, usize)> = None;
        for (r, row) in m.rows.iter().enumerate() {
            if row.decoration {
                continue;
            }
            let cand = row
                .glyphs
                .iter()
                .enumerate()
                .find(|(_, g)| g.stop && g.src >= off)
                .map(|(i, g)| (g.src, r, row.col_of_glyph(i)))
                .or_else(|| (row.end_src >= off).then_some((row.end_src, r, row.width())));
            if let Some(c) = cand
                && best.is_none_or(|b| c.0 <= b.0)
            {
                best = Some(c);
            }
        }
        match best {
            Some((_, r, c)) => (r, c),
            None => {
                let r = m.last_stop_row();
                (r, m.row_width(r))
            }
        }
    }

    /// `visible_items` as it was written before the spelling was tabulated:
    /// every stop glyph in the range gathered, sorted, and paired up.
    fn visible_items_by_scan(m: &VisualMap, from: usize, to: usize) -> Vec<(usize, Option<char>)> {
        let (lo, hi) = m.visible_span(from, to);
        let from = m.snap_to_glyph_stop(from);
        let mut glyphs: Vec<(usize, char)> = m
            .rows
            .iter()
            .filter(|r| !r.decoration)
            .flat_map(|r| r.glyphs.iter())
            .filter(|g| g.stop && g.src >= from && g.src < to)
            .map(|g| (g.src, g.ch))
            .collect();
        glyphs.sort_by_key(|&(src, _)| src);
        glyphs.dedup_by_key(|&mut (src, _)| src);
        let cell_ends: Vec<usize> = m
            .tables
            .iter()
            .flat_map(|t| t.grid.iter())
            .flat_map(|r| r.cells.iter())
            .map(|c| c.end)
            .collect();
        m.stops[lo..hi]
            .iter()
            .map(|&s| {
                let ch = glyphs
                    .iter()
                    .find(|&&(src, _)| src == s)
                    .filter(|_| !cell_ends.contains(&s))
                    .map(|&(_, ch)| ch);
                (s, ch)
            })
            .collect()
    }

    #[test]
    fn pos_of_offset_agrees_with_reading_every_row_through() {
        for (name, m) in lookup_maps() {
            let len = m.rows.iter().map(|r| r.end_src).max().unwrap_or(0) + 2;
            for off in 0..=len {
                assert_eq!(
                    m.pos_of_offset(off),
                    pos_of_offset_by_scan(&m, off),
                    "offset {off} of {name}"
                );
            }
        }
    }

    #[test]
    fn the_tabulated_spelling_agrees_with_gathering_the_glyphs() {
        for (name, m) in lookup_maps() {
            let end = m.stops.last().copied().unwrap_or(0);
            // Whole text, and every window a frontend might ask for.
            let mut spans = vec![(0, end)];
            for &a in m.stops.iter().step_by(3) {
                spans.push((a, end));
                spans.push((0, a));
                spans.push((a, (a + 5).min(end)));
            }
            for (from, to) in spans {
                let items = visible_items_by_scan(&m, from, to);
                assert_eq!(
                    m.visible_items(from, to),
                    items,
                    "items {from}..{to} of {name}"
                );
                let utf16: usize = items
                    .iter()
                    .map(|(_, ch)| ch.map_or(1, char::len_utf16))
                    .sum();
                assert_eq!(
                    m.visible_utf16_len(from, to),
                    utf16,
                    "utf16 {from}..{to} of {name}"
                );
            }
            // And back: every UTF-16 index in the text resolves to the stop
            // that spells it, as the scan would have found it.
            let items = visible_items_by_scan(&m, 0, end);
            let mut seen = 0;
            for (src, ch) in &items {
                for u in seen..seen + ch.map_or(1, char::len_utf16) {
                    assert_eq!(
                        m.offset_at_visible_utf16(end, u),
                        Some(*src),
                        "index {u} of {name}"
                    );
                }
                seen += ch.map_or(1, char::len_utf16);
            }
            assert_eq!(
                m.offset_at_visible_utf16(end, seen),
                None,
                "past the end of {name}"
            );
        }
    }

    #[test]
    fn visible_text_spends_exactly_one_character_on_every_stop() {
        // A list (whose items' ends no gap row follows), a table (whose cells'
        // ends draw a gutter space), and a code block (one row per line):
        // every place the text used to part company with the stop count, in
        // both directions. `UITextInput`'s tokenizer indexes this text by
        // that count, so the two must agree exactly between any two stops.
        let src = "- one\n- two\n\n| a | b |\n| - | - |\n| c | d |\n\n```\nx\ny\n```\n\nend\n";
        let m = map(src);
        let end = m.snap_to_stop(src.len());
        // The table's trailing stop draws no glyph, so it is spelled as a line
        // end too: to the system the table ends on a blank line, which is
        // where the caret past it stands.
        assert_eq!(m.visible_text(0, end), "one\ntwo\na\nb\nc\nd\n\nx\ny\nend");
        // Between any two stops, one character per hop.
        let first = m.snap_to_glyph_stop(0);
        let stops: Vec<usize> = std::iter::successors(Some(first), |&o| m.stop_after(o)).collect();
        for (i, &a) in stops.iter().enumerate() {
            for (j, &b) in stops.iter().enumerate().skip(i) {
                assert_eq!(
                    m.visible_text(a, b).chars().count(),
                    j - i,
                    "text between stops {a} and {b}"
                );
            }
        }
        // A cell's end is spelled as a line end, not the space it draws, so a
        // tap landing past `a`'s last letter has nothing to step over into `b`.
        let a_end = src.find("a |").unwrap() + 1;
        assert_eq!(m.visible_text(a_end, a_end + 1), "\n");
    }

    #[test]
    fn unwrapped_mode_emits_one_row_per_paragraph() {
        // A long paragraph that would wrap under a column budget stays a single
        // row when wrap is None (the GUI wraps it at pixel width instead).
        let long = "one two three four five six seven eight nine ten eleven twelve\n";
        let mut ed = Editor::new_str(long, Format::Markdown).unwrap();
        let wrapped = build_t(&ed.nodes().unwrap(), long, Some(12));
        let unwrapped = build_t(&ed.nodes().unwrap(), long, None);
        assert!(wrapped.num_rows() > 1, "narrow column should wrap");
        assert_eq!(unwrapped.num_rows(), 1, "no budget should keep it one row");
        // Every glyph's source byte is preserved in the single row.
        let text: String = unwrapped.rows[0].glyphs.iter().map(|g| g.ch).collect();
        assert_eq!(text.trim_end(), long.trim_end());
    }

    fn line_texts(m: &VisualMap) -> Vec<String> {
        m.rows
            .iter()
            .map(|r| {
                // Trim the trailing whitespace a row may carry — the zero-width
                // '\n' that closes a preserved line, and any space glyph left at
                // a wrap boundary (both real caret stops, neither visible text).
                r.glyphs
                    .iter()
                    .map(|g| g.ch)
                    .collect::<String>()
                    .trim_end()
                    .to_string()
            })
            .collect()
    }

    #[test]
    fn preserve_lays_each_soft_break_on_its_own_row() {
        // A soft break (a bare newline inside a paragraph) folds into a space by
        // default — the whole paragraph is one reflowed row...
        let src = "one two\nthree four\n";
        let mut ed = Editor::new_str(src, Format::Markdown).unwrap();
        let folded = build_t(&ed.nodes().unwrap(), src, None);
        assert_eq!(folded.num_rows(), 1, "fold: one reflowed row");
        assert_eq!(
            line_texts(&folded),
            vec!["one two three four"],
            "break folded to a space"
        );

        // ...and under Preserve it renders where it was written, a row per line.
        let kept = map_preserve(src, None);
        assert_eq!(
            line_texts(&kept),
            vec!["one two", "three four"],
            "preserve: a row per line"
        );
    }

    #[test]
    fn a_preserved_break_keeps_the_newline_offset_as_a_caret_stop() {
        // The break must leave a caret stop at the newline byte, or the caret
        // could not rest at the end of the first line. The '\n' glyph is dropped
        // from the row (so nothing stray renders); its offset (7 here) becomes the
        // row's end stop instead — the same offset the folded space would carry.
        let src = "one two\nthree four\n";
        let m = map_preserve(src, None);
        assert!(
            !m.rows[0].glyphs.iter().any(|g| g.ch == '\n'),
            "the break glyph is dropped"
        );
        assert_eq!(
            m.rows[0].end_src, 7,
            "the first row ends at the newline byte"
        );
        assert!(m.is_stop(7), "the newline offset is a caret stop");
        // Row end offsets stay strictly ascending — no two rows pin one offset.
        let offs: Vec<usize> = m.rows.iter().map(|r| r.end_src).collect();
        assert!(
            offs.windows(2).all(|w| w[0] < w[1]),
            "offsets not unique: {offs:?}"
        );
    }

    #[test]
    fn preserved_lines_wrap_independently() {
        // Each preserved line wraps to the column on its own; the break between
        // them is hard, so a word never crosses it — "gamma" and "delta" could
        // share a row on width alone but the soft break keeps them apart.
        let src = "alpha beta gamma\ndelta epsilon\n";
        let m = map_preserve(src, Some(12));
        assert_eq!(
            line_texts(&m),
            vec!["alpha beta", "gamma", "delta", "epsilon"],
            "each source line wraps on its own"
        );
    }

    #[test]
    fn an_empty_paragraph_between_blocks_renders_its_own_rows() {
        // "A", then two blank lines (an empty paragraph opened with Enter), then
        // "B": the empty paragraph must be navigable rows, not collapsed onto B.
        // Rows: "A", spacer, empty-paragraph, spacer, "B" — each blank row a
        // distinct source offset.
        let m = map("A\n\n\n\nB\n");
        let text: Vec<String> = m
            .rows
            .iter()
            .map(|r| r.glyphs.iter().map(|g| g.ch).collect())
            .collect();
        assert_eq!(text, vec!["A", "", "", "", "B"], "got {text:?}");
        let offs: Vec<usize> = m.rows.iter().map(|r| r.end_src).collect();
        // Strictly ascending — no two rows share an offset (else the caret pins).
        assert!(
            offs.windows(2).all(|w| w[0] < w[1]),
            "offsets not unique: {offs:?}"
        );
    }

    #[test]
    fn a_tight_block_boundary_still_gets_one_separator() {
        // A heading directly above text (no blank line between) keeps the single
        // conventional separator row, as before.
        let m = map("# H\ntext\n");
        let text: Vec<String> = m
            .rows
            .iter()
            .map(|r| r.glyphs.iter().map(|g| g.ch).collect())
            .collect();
        assert_eq!(text, vec!["H", "", "text"], "got {text:?}");
    }

    #[test]
    fn an_escaped_delimiter_renders_without_its_backslash_and_maps_true_offsets() {
        // `a\*b` renders the three visible chars `a * b` — the escape backslash
        // is hidden — and every glyph points at its real source byte, so a caret
        // past the escape lands right (the `*` at source 2, `b` at source 3, not
        // the drifted 1/2 the naive text-offset mapping gave).
        let m = map("a\\*b\n");
        let row: Vec<(char, usize)> = m.rows[0].glyphs.iter().map(|g| (g.ch, g.src)).collect();
        assert_eq!(row, vec![('a', 0), ('*', 2), ('b', 3)], "got {row:?}");
    }

    #[test]
    fn an_escaped_hash_stays_a_paragraph_and_shows_the_hash() {
        // `\# hi` is a paragraph beginning with a literal `#`, not a heading —
        // the backslash is hidden, the `#` shown at its true offset.
        let m = map("\\# hi\n");
        let text: String = m.rows[0].glyphs.iter().map(|g| g.ch).collect();
        assert_eq!(text, "# hi");
        assert_eq!(
            m.rows[0].glyphs[0].src, 1,
            "the # is at source byte 1, past the \\"
        );
    }

    #[test]
    fn a_tight_nested_list_hangs_its_sublist_directly_under_the_item() {
        // A list item's own text and the sub-list nested under it are written on
        // adjacent source lines, so the rich view butts them together — no
        // fabricated blank row. Regression: the synthetic "breathe" separator
        // used to open a gap between `• a` and its `  • b`.
        assert_eq!(rendered(&map("- a\n  - b\n")), "• a\n  • b");
    }

    #[test]
    fn a_loose_nested_list_keeps_its_real_blank_line() {
        // A genuine blank source line (a loose list) still parts the item from
        // its sub-list — only the *fabricated* separator is suppressed, never a
        // real one the author typed. The gap row wears the item's continuation
        // prefix (the two-space indent), so it renders as "  ", not empty.
        assert_eq!(rendered(&map("- a\n\n  - b\n")), "• a\n  \n  • b");
    }

    #[test]
    fn a_code_block_in_a_list_item_wears_one_marker() {
        // The item's marker goes on the block's first line and its indent on
        // the rest, as a wrapped paragraph's rows do. Every line wore the
        // marker once, so a two-line block in an item read as two items.
        assert_eq!(
            rendered(&map("- ```\n  one\n  two\n  ```\n")),
            "• one\n  two"
        );
        assert_eq!(
            rendered(&map("1. ```\n   one\n   two\n   ```\n")),
            "1. one\n   two"
        );
        assert_eq!(
            rendered(&map("- [ ] ```\n  one\n  two\n  ```\n")),
            "☐ one\n  two"
        );
    }

    #[test]
    fn an_items_later_rows_wear_its_marker_as_a_blank_indent() {
        // Spelled with the marker's characters so a proportional face gives
        // the indent exactly the marker's width; drawn blank everywhere.
        for (src, marker) in [
            ("- a\n\n  b\n", "• "),
            ("1. a\n\n   b\n", "1. "),
            ("- [ ] a\n\n  b\n", "☐ "),
        ] {
            let m = map(src);
            let last = m.rows.last().unwrap();
            let indent: String = last
                .glyphs
                .iter()
                .take_while(|g| g.style.role == Role::ListIndent)
                .map(|g| g.ch)
                .collect();
            assert_eq!(indent, marker, "{src:?}");
            assert!(rendered(&m).ends_with(&format!("{}b", " ".repeat(marker.chars().count()))));
        }
    }

    #[test]
    fn a_code_block_in_a_quote_draws_no_row_for_its_closing_fence() {
        // The gutter is the same on every row. The closing fence is markup,
        // and drew an empty quoted line under the code as if the writer had
        // typed one, at the end of the document or before more.
        assert_eq!(
            rendered(&map("> ```\n> one\n> two\n> ```\n")),
            "│ one\n│ two"
        );
        assert_eq!(
            rendered(&map("> ```\n> one\n> ```\n\nafter\n")),
            "│ one\n\nafter"
        );
        // A quoted line the writer added under the fence is still one.
        assert_eq!(rendered(&map("> ```\n> one\n> ```\n>\n")), "│ one\n│ ");
    }

    #[test]
    fn frontmatter_is_hidden_and_the_document_opens_into_its_content() {
        // Leading YAML frontmatter renders nothing — no phantom blank rows for
        // its lines, no leading gap — and `content_start` points at the first
        // real block so the caret floor can keep out of the hidden metadata.
        let fm = "---\nconfig: prov.yaml\ncontents:\n- '[Sample](sample.md)'\n---\n";
        let src = format!("{fm}# leaf\n\nA line.\n");
        let m = map(&src);
        let text = rendered(&m);
        assert!(
            !text.contains("config"),
            "frontmatter body leaked: {text:?}"
        );
        assert!(!text.contains("prov"), "frontmatter body leaked: {text:?}");
        assert_eq!(
            m.rows[0].glyphs.iter().map(|g| g.ch).collect::<String>(),
            "leaf"
        );
        assert_eq!(
            m.content_start,
            fm.len(),
            "floor should be the first real block"
        );
    }

    #[test]
    fn a_frontmatter_only_document_puts_the_floor_after_the_frontmatter() {
        // Nothing to render, so the caret floor is the end of the hidden
        // frontmatter — not 0, which is *before* the opening `---` and made the
        // first keystroke in a fresh metadata-only note land ahead of it. And
        // the frontmatter's own newlines are not trailing blank lines: they used
        // to open phantom rows at offsets 1..4, inside the metadata.
        let src = "---\ntitle: 2026-08-29\nid: f8s32cd\n---\n";
        let m = map(src);
        assert_eq!(m.content_start, src.len(), "floor must clear the metadata");
        assert!(
            m.rows.is_empty(),
            "frontmatter must render no rows: {:?}",
            rendered(&m)
        );
        assert!(
            m.stops.is_empty(),
            "no stop may sit inside the metadata: {:?}",
            m.stops
        );
    }

    #[test]
    fn a_frontmatter_only_document_still_counts_its_real_blank_lines() {
        // Two blank lines after the frontmatter are the author's empty paragraph
        // and still render, counted from the metadata's end rather than from 0.
        let fm = "---\ntitle: n\n---\n";
        let m = map(&format!("{fm}\n\n"));
        assert_eq!(m.content_start, fm.len());
        assert_eq!(m.rows.len(), 2, "the two trailing newlines each open a row");
        assert!(
            m.rows.iter().all(|r| r.end_src > fm.len()),
            "rows must sit past the frontmatter"
        );
    }

    #[test]
    fn a_document_without_frontmatter_has_a_zero_floor() {
        let m = map("# leaf\n\nbody\n");
        assert_eq!(m.content_start, 0);
    }

    #[test]
    fn trailing_spaces_become_caret_stops_so_the_caret_can_be_drawn_past_them() {
        // Markdown/Djot drop the trailing space in `hello ` from the `str` node,
        // so without help the row would end at `hello` and the caret couldn't be
        // drawn past column 5 — typing a space at a line's end wouldn't move it
        // on screen until the next visible character reparsed the space into an
        // interior node. The builder recovers it from the block's span/content_span
        // gap and emits it as a real, caret-stoppable glyph.
        let m = map("hello \n");
        assert_eq!(
            m.rows[0].glyphs.iter().map(|g| g.ch).collect::<String>(),
            "hello "
        );
        assert_eq!(
            m.rows[0].end_src, 6,
            "the row now ends past the trailing space"
        );
        // The caret can rest both on and past the space.
        assert_eq!(m.pos_of_offset(5), (0, 5), "between 'o' and the space");
        assert_eq!(m.pos_of_offset(6), (0, 6), "past the space");
        // Two trailing spaces, both stops.
        let m = map("hello  \n");
        assert_eq!(
            m.rows[0].glyphs.iter().map(|g| g.ch).collect::<String>(),
            "hello  "
        );
        assert_eq!(m.pos_of_offset(7), (0, 7));
    }

    #[test]
    fn a_headings_trailing_space_is_a_caret_stop_too() {
        // The hidden `# ` marker means `# hi ` renders as `hi ` in three columns;
        // the caret past the trailing space lands on the third.
        let m = map("# hi \n");
        assert_eq!(
            m.rows[0].glyphs.iter().map(|g| g.ch).collect::<String>(),
            "hi "
        );
        assert_eq!(m.pos_of_offset(5), (0, 3));
    }

    #[test]
    fn a_table_cells_trailing_padding_is_not_mistaken_for_block_trailing_space() {
        // A cell's own `span` is the whole row, so the trailing-whitespace
        // recovery must not run for cells or it would swallow the `│` delimiters
        // and neighbours between the cell text and the row's end. The grid stays
        // exactly as before.
        let text = rendered(&map(TABLE));
        assert!(
            text.contains("│ Pear │   3 │"),
            "cell padding disturbed:\n{text}"
        );
    }

    #[test]
    fn a_click_below_the_last_row_lands_on_the_last_stop_not_offset_zero() {
        // A drag into the empty space under a short document used to resolve to
        // offset 0 — the wrong direction, and not even a caret stop when the
        // document opens on hidden frontmatter (its `content_start` floor is not
        // a stop), which crashed the caret invariant. It now lands on the last
        // stop: the end of the document, where dragging downward should reach.
        let fm = "---\ntitle: n\n---\n";
        let m = map(&format!("{fm}# Hi\n\nbody\n"));
        let below = m.num_rows() + 5;
        let off = m.offset_of_pos(below, 0);
        assert!(
            m.is_stop(off),
            "offset {off} from a below-content click is not a stop"
        );
        assert_eq!(
            off,
            m.stops.last().copied().unwrap(),
            "should be the document's last stop"
        );
        assert!(
            off > fm.len(),
            "must not fall onto the hidden frontmatter floor"
        );
    }

    #[test]
    fn offset_of_pos_is_a_stop_for_every_row_including_past_the_end() {
        // The invariant the caret motion asserts: whatever cell a click names,
        // the offset it resolves to is one the caret can actually rest at.
        for src in [
            "hello \n",
            "# A heading here \n\nbody text goes on \n",
            "---\nk: v\n---\n# Title\n\nprose here that wraps a bit \n",
        ] {
            let m = map(src);
            for row in 0..m.num_rows() + 3 {
                for col in 0..30 {
                    let off = m.offset_of_pos(row, col);
                    assert!(
                        m.is_stop(off),
                        "row {row} col {col} → {off} is not a stop in {src:?}"
                    );
                }
            }
        }
    }

    /// `| Name | Qty |` with Name left-aligned and Qty right-aligned.
    const TABLE: &str = "| Name | Qty |\n|:-----|----:|\n| Pear | 3 |\n| Fig | 12 |\n";

    #[test]
    fn a_table_renders_as_an_aligned_grid() {
        let text = rendered(&map(TABLE));
        assert_eq!(
            text,
            "┌──────┬─────┐\n\
             │ Name │ Qty │\n\
             ├──────┼─────┤\n\
             │ Pear │   3 │\n\
             │ Fig  │  12 │\n\
             └──────┴─────┘",
            "got:\n{text}"
        );
    }

    #[test]
    fn table_columns_honour_their_alignment() {
        // Centre and default(left) come straight from twig's cell.alignment —
        // the delimiter row it's spelled in is consumed and has no node.
        let text = rendered(&map("| A | Bee |\n| --- | :---: |\n| x | y |\n"));
        assert!(text.contains("│ x │  y  │"), "centred column: {text:?}");
    }

    #[test]
    fn table_borders_are_decoration_the_caret_never_lands_on() {
        let m = map(TABLE);
        // The top and header rules are whole decoration rows.
        for r in [0, 2] {
            assert!(m.rows[r].decoration, "row {r} should be a decoration rule");
            assert!(
                !m.rows[r].glyphs.iter().any(|g| g.stop),
                "row {r} has a stop"
            );
        }
        // The bottom border is the exception: no glyph of it is a stop, but
        // its end is the table's trailing caret home — the one place the caret
        // can stand past the last cell.
        let bottom = &m.rows[5];
        assert!(
            !bottom.decoration,
            "the bottom border holds the trailing stop"
        );
        assert!(
            !bottom.glyphs.iter().any(|g| g.stop),
            "the bottom border's glyphs are not stops"
        );
        assert!(m.is_stop(bottom.end_src), "the trailing stop is a stop");
        assert!(m.table_end_stop(bottom.end_src));
        assert_eq!(
            bottom.end_src,
            TABLE.trim_end_matches('\n').len(),
            "the trailing stop is the table's own end, before its newline"
        );
        assert!(
            !m.table_end_stop(TABLE.rfind("12").unwrap() + 2),
            "a cell's end is not the trailing stop"
        );
        // A content row's `│` and padding are decoration; only the cell text
        // and each cell's one end-stop are stops.
        let header = &m.rows[1];
        assert!(!header.decoration);
        for g in &header.glyphs {
            if g.ch == '│' {
                assert!(!g.stop, "a border is not a caret stop");
            }
        }
        let stops: String = header
            .glyphs
            .iter()
            .filter(|g| g.stop)
            .map(|g| g.ch)
            .collect();
        assert_eq!(stops, "Name Qty ", "cell text plus one end-stop space each");
    }

    #[test]
    fn a_cell_maps_to_its_own_source_text() {
        let m = map(TABLE);
        // "Pear" starts at byte 32 in TABLE; the caret there draws on the 'P'.
        let pear = TABLE.find("Pear").unwrap();
        let (r, c) = m.pos_of_offset(pear);
        assert_eq!(m.rows[r].glyphs[c].ch, 'P');
        assert_eq!(m.offset_of_pos(r, c), pear, "round trips");
    }

    #[test]
    fn a_wide_table_is_cut_to_fit_and_its_cells_wrap() {
        // Columns wider than the surface used to run off the right edge, where
        // nothing could reach them. They're cut to the budget instead, and the
        // text wraps down inside the column — the header rule stays put, and
        // an alignment holds on every line of a wrapped cell, not just the first.
        let src = "| Ingredient | Notes |\n|---|---:|\n\
                   | flour milled coarse | sift it twice |\n| salt | a pinch |\n";
        let mut ed = Editor::new_str(src, Format::Markdown).unwrap();
        let m = build_t(&ed.nodes().unwrap(), src, Some(30));
        let text = rendered(&m);
        assert_eq!(
            text,
            "┌──────────────┬─────────────┐\n\
             │ Ingredient   │       Notes │\n\
             ├──────────────┼─────────────┤\n\
             │ flour milled │     sift it │\n\
             │ coarse       │       twice │\n\
             │ salt         │     a pinch │\n\
             └──────────────┴─────────────┘",
            "got:\n{text}"
        );
        for (r, row) in m.rows.iter().enumerate() {
            assert!(
                row.glyphs.len() <= 30,
                "row {r} overflows: {}",
                row.glyphs.len()
            );
        }
    }

    #[test]
    fn a_column_too_narrow_for_a_word_breaks_it_rather_than_spilling() {
        // A paragraph lets an overlong word trail off the end of the line; a
        // table column can't — a glyph past the border lands on the border.
        let src = "| A | B |\n|---|---|\n| antidisestablishmentarianism | x |\n";
        let mut ed = Editor::new_str(src, Format::Markdown).unwrap();
        let m = build_t(&ed.nodes().unwrap(), src, Some(20));
        for (r, row) in m.rows.iter().enumerate() {
            assert!(
                row.glyphs.len() <= 20,
                "row {r} overflows: {}",
                row.glyphs.len()
            );
        }
        // Broken across lines, but whole: every letter is still drawn, at its
        // own source byte, where the caret can reach it.
        let word = "antidisestablishmentarianism";
        let at = src.find(word).unwrap();
        for (i, ch) in word.char_indices() {
            assert!(
                m.rows
                    .iter()
                    .flat_map(|r| r.glyphs.iter())
                    .any(|g| g.stop && g.src == at + i && g.ch == ch),
                "{ch:?} at {} was lost to the break",
                at + i
            );
        }
    }

    #[test]
    fn a_code_block_maps_each_line_to_its_own_source_text() {
        // Every glyph used to point at the block's start, which made the whole
        // block one offset — visible, but impossible to put a caret inside.
        let src = "```rust\nlet x = 1;\nfn f() {}\n```\n";
        let m = map(src);
        for row in &m.rows {
            for g in row.glyphs.iter().filter(|g| g.stop) {
                assert_eq!(
                    src[g.src..].chars().next(),
                    Some(g.ch),
                    "glyph {:?} at {} isn't the source byte it claims",
                    g.ch,
                    g.src
                );
            }
        }
    }

    #[test]
    fn an_indented_code_block_maps_past_its_stripped_indent() {
        // twig strips the four-space indent, so `text` isn't a source slice and
        // the lines have to be re-found. Offsets land on the code, not the indent.
        let src = "    indented\n    code\n";
        let m = map(src);
        let stops: Vec<(char, usize)> = m
            .rows
            .iter()
            .flat_map(|r| r.glyphs.iter().filter(|g| g.stop).map(|g| (g.ch, g.src)))
            .collect();
        assert_eq!(
            stops[0],
            ('i', 4),
            "first line should start past the indent"
        );
        assert!(
            stops.contains(&('c', 17)),
            "second line misplaced: {stops:?}"
        );
    }

    #[test]
    fn a_fenced_block_whose_code_echoes_its_info_string_maps_to_the_code() {
        // The one case that defeats a forward search: the opening fence
        // ```` ```rust ```` ends with the same text as the code under it.
        let src = "```rust\nrust\n```\n";
        let m = map(src);
        let first = m.rows[0].glyphs.iter().find(|g| g.stop).unwrap();
        assert_eq!(first.src, 8, "matched the info string, not the code");
    }

    #[test]
    fn a_code_block_carries_no_gutter_and_is_published_as_a_row_span() {
        // The old `▏ ` gutter is gone: a code row is the block prefix (none, at
        // the top level) plus the code text, and the whole run is named in
        // `code_blocks` so a frontend can box it.
        let src = "para\n\n```\ncode\nlines\n```\n\nafter\n";
        let m = map(src);
        assert_eq!(m.code_blocks.len(), 1, "one code block");
        let span = m.code_blocks[0].rows_span.clone();
        let rows: Vec<String> = m.rows[span.clone()]
            .iter()
            .map(|r| r.glyphs.iter().map(|g| g.ch).collect())
            .collect();
        assert_eq!(rows, vec!["code".to_string(), "lines".to_string()]);
        assert!(!rendered(&m).contains('▏'), "gutter still drawn");
        assert!(
            m.rows[span].iter().all(|r| r.code),
            "every row in the span is flagged code"
        );
    }

    #[test]
    fn an_empty_last_line_in_a_code_block_is_a_row_of_its_own() {
        // `trim_end_matches('\n')` cut the block's terminator *and* the newline
        // that spells a trailing empty line, so the row the Return had just made
        // never appeared and the caret on it fell through to the block below.
        // Every empty line is a row, wherever in the block it falls.
        let src = "prose\n\n```\nalpha\nbeta\n\n```\n\nafter\n";
        let m = map(src);
        let span = m.code_blocks[0].rows_span.clone();
        let rows: Vec<String> = m.rows[span.clone()]
            .iter()
            .map(|r| r.glyphs.iter().map(|g| g.ch).collect())
            .collect();
        assert_eq!(
            rows,
            vec!["alpha".to_string(), "beta".to_string(), String::new()],
            "the empty last line gets a row"
        );
        assert!(
            m.rows[span.clone()].iter().all(|r| r.code),
            "the empty row is flagged code like the rest of the block"
        );
        // And it is the *source's* empty line, not a coarse fallback to the
        // block start: the offset the caret resolves to is the one Return made.
        let empty = span.end - 1;
        assert_eq!(
            m.rows[empty].end_src,
            src.find("beta\n\n").unwrap() + "beta\n".len(),
            "the empty row maps to the line the Return opened"
        );

        // Nothing is invented where there is no empty line, and a second one is
        // a second row.
        assert_eq!(
            map("```\nalpha\nbeta\n```\n").code_blocks[0]
                .rows_span
                .len(),
            2,
            "a block that ends at its last code line keeps two rows"
        );
        assert_eq!(
            map("```\nalpha\n\n\n```\n").code_blocks[0].rows_span.len(),
            3,
            "two trailing empty lines are two rows"
        );
    }

    #[test]
    fn a_directive_container_is_tinted_and_labeled_on_its_first_row() {
        // diaryx's `:::vis{.public .family}` visibility block, and any other
        // `:::name{.class}` fenced div — core is agnostic of `name`.
        let src = ":::vis{.public .family}\nhello\n\nworld\n:::\nafter\n";
        let m = map_directives(src);

        let content_rows: Vec<usize> = (0..m.rows.len()).filter(|&i| m.rows[i].directive).collect();
        assert!(!content_rows.is_empty(), "some row is flagged directive");

        let after_rows: Vec<usize> = (0..m.rows.len())
            .filter(|&i| !content_rows.contains(&i) && !m.rows[i].glyphs.is_empty())
            .collect();
        assert!(
            after_rows.iter().all(|&i| !m.rows[i].directive),
            "content outside the fence isn't tinted"
        );

        let labels: Vec<&str> = content_rows
            .iter()
            .filter_map(|&i| m.rows[i].directive_label.as_deref())
            .collect();
        assert_eq!(
            labels,
            vec!["public family"],
            "only the first row carries the label"
        );

        assert_eq!(
            rendered(&m)
                .lines()
                .filter(|l| !l.is_empty())
                .collect::<Vec<_>>(),
            vec!["hello", "world", "after"],
            "fence markers don't leak into the rendered text"
        );
    }

    #[test]
    fn a_bare_word_directive_is_labeled_same_as_dot_classes() {
        // diaryx_core::visibility's own `:::vis{public family}` — no leading
        // dots — is what apps/web's directive serializer and the native
        // publish-time filter both actually write today, distinct from twig's
        // `.class` convention. Both must label the same way so every existing
        // diaryx `:::vis{...}` block reads, not just newly dot-authored ones.
        let src = ":::vis{public family}\nhello\n:::\n";
        let m = map_directives(src);
        let label = m.rows.iter().find_map(|r| r.directive_label.clone());
        assert_eq!(label.as_deref(), Some("public family"));
    }

    #[test]
    fn a_text_directive_keeps_its_paragraph_visible() {
        // Regression: an inline `:name[label]{…}` used to make its paragraph
        // fail the "all children inline" test, so the whole line was walked as
        // a container of blocks and rendered as empty rows with NO caret stops —
        // the text vanished from the editor and the caret couldn't enter it.
        // diaryx's inline `:vis[…]` is exactly this shape.
        let src = "Text with :abbr[HTML]{title=\"HyperText\"} inline.\n";
        let m = map_directives(src);
        assert_eq!(rendered(&m).trim_end(), "Text with HTML inline.");
        // Every character of the line is a caret home, markup excluded — the
        // label reads as ordinary text, the way a link's does.
        let stops: usize = m
            .rows
            .iter()
            .map(|r| r.glyphs.iter().filter(|g| g.stop).count())
            .sum();
        assert_eq!(stops, "Text with HTML inline.".chars().count());
        // It is inline, so it is not the container form's tinted panel.
        assert!(m.rows.iter().all(|r| !r.directive));
    }

    #[test]
    fn a_text_directives_label_maps_to_its_true_source_bytes() {
        // Regression (needs twig-doc >= 2.5.0): twig parses a `[label]` as a
        // detached slice, and until it rebased the enclosing scan's segments
        // onto it every node inside the label reported a span of `(0,0)`. Read
        // by anything that trusts a span that means "byte 0", so the label's
        // glyphs mapped to the START OF THE DOCUMENT — a click on the label put
        // the caret at the top of the file, its stops collided with the real
        // first line's, and an edit there landed on the wrong bytes entirely.
        //
        // The sibling test `a_text_directive_keeps_its_paragraph_visible` only
        // counts stops, which is exactly why this went unnoticed: the right
        // NUMBER of stops at completely wrong offsets.
        let src = "x :abbr[HTML]{title=\"y\"} z\n";
        let m = map_directives(src);
        let stops: Vec<(char, usize)> = m
            .rows
            .iter()
            .flat_map(|r| &r.glyphs)
            .filter(|g| g.stop)
            .map(|g| (g.ch, g.src))
            .collect();
        // `HTML` sits at 8..12. The name, brackets and `{…}` are hidden markup
        // the caret steps over, so the line's stops run 0, 1, 8..12, then 24.
        assert_eq!(
            stops,
            [
                ('x', 0),
                (' ', 1),
                ('H', 8),
                ('T', 9),
                ('M', 10),
                ('L', 11),
                (' ', 24),
                ('z', 25)
            ]
        );
    }

    #[test]
    fn every_glyph_in_a_directive_label_points_at_its_source_byte() {
        // The `every_glyph_points_at_its_source_byte` invariant, extended over
        // directive labels now that their offsets are real. Nested markup is
        // included: its delimiters are hidden, so the visible glyphs must skip
        // them and still name their own bytes.
        let src = "x :abbr[a *b* c] y and :vis[family only] z\n";
        let m = map_directives(src);
        for g in m.rows.iter().flat_map(|r| &r.glyphs).filter(|g| g.stop) {
            let at = src[g.src..].chars().next();
            assert_eq!(
                at,
                Some(g.ch),
                "glyph {:?} claims byte {}, which is {at:?}",
                g.ch,
                g.src
            );
        }
        assert_eq!(rendered(&m).trim_end(), "x a b c y and family only z");
    }

    #[test]
    fn a_directive_labels_nested_emphasis_keeps_both_its_style_and_its_offsets() {
        let src = "x :abbr[a *b* c] y\n";
        let m = map_directives(src);
        let b = m
            .rows
            .iter()
            .flat_map(|r| &r.glyphs)
            .find(|g| g.ch == 'b')
            .expect("the emphasised char");
        assert!(b.style.italic, "the label's *b* lost its emphasis");
        assert_eq!(b.src, 11, "the label's *b* lost its source byte");
    }

    #[test]
    fn a_bare_colon_word_renders_as_the_prose_it_almost_always_is() {
        // Regression: twig matches a colon followed by any letter-led word, so
        // ordinary prose is full of "text directives" nobody meant to write.
        // With no `[label]` there are no children, and the arm recursed into
        // them — rendering *nothing*. The word vanished from the document with
        // no caret stop left behind, so it could not even be deleted.
        for src in ["a :word b\n", "note :see below\n", ":smile: hi\n"] {
            let m = map_directives(src);
            assert_eq!(
                rendered(&m).trim_end(),
                src.trim_end(),
                "prose was eaten: {src:?}"
            );
        }
    }

    #[test]
    fn a_bare_colon_word_keeps_every_byte_a_caret_stop() {
        let src = "a :word b\n";
        let m = map_directives(src);
        // Nothing here is markup, so nothing is hidden: each byte maps to
        // itself and can be stood on, which is what makes the colon deletable.
        let stops: Vec<(char, usize)> = m
            .rows
            .iter()
            .flat_map(|r| &r.glyphs)
            .filter(|g| g.stop)
            .map(|g| (g.ch, g.src))
            .collect();
        assert_eq!(
            stops,
            "a :word b"
                .chars()
                .enumerate()
                .map(|(i, c)| (c, i))
                .collect::<Vec<_>>()
        );
    }

    #[test]
    fn an_attribute_bearing_text_directive_draws_a_chip() {
        // `{…}` is deliberate in a way a bare colon is not — diaryx writes
        // `:vis{.family}` inline — so this one reads as an embed, on the same
        // `⧉ label` recipe the leaf form's placeholder row uses.
        // Both attribute conventions label it: twig's dot-prefixed classes and
        // the bare pandoc-style words diaryx also writes.
        for src in ["a :vis{.family} b\n", "a :vis{family} b\n"] {
            let m = map_directives(src);
            assert_eq!(rendered(&m).trim_end(), "a ⧉ vis family b", "{src:?}");
        }
        // A `key=value` attr is configuration, not a name, so it adds nothing.
        let m = map_directives("a :foo{title=\"x\"} b\n");
        assert_eq!(rendered(&m).trim_end(), "a ⧉ foo b");
    }

    #[test]
    fn a_directive_chip_is_one_atomic_caret_stop_at_its_own_offset() {
        let src = "a :vis{.family} b\n";
        let m = map_directives(src);
        let stops: Vec<usize> = m
            .rows
            .iter()
            .flat_map(|r| &r.glyphs)
            .filter(|g| g.stop)
            .map(|g| g.src)
            .collect();
        // The chip contributes exactly one stop, at the directive's start (2),
        // so the caret steps over it whole instead of walking hidden markup a
        // byte at a time. `{.family}`'s bytes (3..15) are never stood on.
        assert_eq!(stops, [0, 1, 2, 15, 16]);
    }

    #[test]
    fn a_paragraph_holding_only_a_chip_is_still_navigable() {
        // With no stop of its own the row would be unreachable — the caret
        // could never be put on the line to edit or delete the directive.
        let m = map_directives(":vis{.family}\n");
        assert!(
            m.row_is_navigable(0),
            "a chip-only paragraph has no caret home"
        );
        assert_eq!(
            m.offset_of_pos(0, 0),
            0,
            "its caret home isn't the directive's start"
        );
    }

    #[test]
    fn a_ratio_or_a_clock_time_is_never_a_directive() {
        // twig requires a letter after the colon, so these stay prose — the
        // verbatim arm must not be reached for them at all.
        let src = "ratio 3:4 and 10:30\n";
        assert_eq!(
            rendered(&map_directives(src)).trim_end(),
            "ratio 3:4 and 10:30"
        );
    }

    #[test]
    fn a_leaf_directive_is_a_placeholder_row_with_its_attrs_published() {
        // `::name{…}` is a standalone block with no body — an embed, a table of
        // contents. It used to emit no rows at all: invisible, no caret home,
        // vertical motion crossing a void. Now it draws the image recipe's
        // placeholder and publishes what the host app needs to paint the real
        // thing.
        let src = "before\n\n::embed{src=\"demo.html\" height=\"400\"}\n\nafter\n";
        let m = map_directives(src);

        let row = m
            .rows
            .iter()
            .position(|r| r.leaf_directive.is_some())
            .expect("a placeholder row");
        assert_eq!(
            m.rows[row].glyphs.iter().map(|g| g.ch).collect::<String>(),
            "⧉ embed"
        );
        assert!(
            m.rows[row].glyphs.iter().any(|g| g.stop),
            "the caret can land on it"
        );
        assert!(
            m.rows[row].directive,
            "a frontend frames it like the container form"
        );

        assert_eq!(m.directives.len(), 1);
        let info = &m.directives[0];
        assert_eq!(info.name, "embed");
        assert_eq!(info.rows_span, row..row + 1);
        assert_eq!(info.attr("src"), Some("demo.html"));
        assert_eq!(info.attr("height"), Some("400"));
        assert_eq!(info.attr("nope"), None);
        // The prose around it is untouched.
        assert!(rendered(&m).contains("before") && rendered(&m).contains("after"));
    }

    #[test]
    fn a_leaf_directive_shows_its_label_and_honours_its_prefix() {
        // A `[label]` names the placeholder (the way an image's alt does), and a
        // quoted directive keeps the quote's gutter — it is a block like any
        // other, not a special case that escapes its container.
        let m = map_directives("::embed[Audience demo]{src=\"demo.html\"}\n");
        assert_eq!(rendered(&m).trim_end(), "⧉ Audience demo");
        assert_eq!(m.directives[0].label, "Audience demo");

        let quoted = map_directives("> ::embed{src=\"x.html\"}\n");
        assert_eq!(rendered(&quoted).trim_end(), "│ ⧉ embed");
        assert_eq!(quoted.directives[0].name, "embed");
    }

    #[test]
    fn a_container_directive_is_still_a_panel_not_a_placeholder() {
        // The three forms must not bleed into each other: only the leaf form is
        // a placeholder, and only the container form tints the blocks it wraps.
        let m = map_directives(":::note{.warning}\nBody\n:::\n");
        assert!(
            m.directives.is_empty(),
            "a container publishes no placeholder"
        );
        assert!(m.rows.iter().all(|r| r.leaf_directive.is_none()));
        assert_eq!(rendered(&m).trim_end(), "Body");
        assert!(
            m.rows
                .iter()
                .any(|r| r.directive && r.directive_label.as_deref() == Some("warning"))
        );
    }

    /// A production-path build with both extensions on — the only way to put a
    /// promoted HTML element and a directive in one document, which is what the
    /// `container` kind made necessary to tell apart. Returns the whole `Doc`
    /// because [`VisualMap`] is not `Clone`; read `doc.vmap`.
    fn doc_built(src: &str) -> crate::Doc {
        let mut doc = crate::Doc::from_source(src.to_string(), Format::Markdown).unwrap();
        doc.build_visual(80);
        doc
    }

    /// Every `container` node in `src`, parsed the way production does (both
    /// extensions on), paired with what [`container_is_directive`] makes of it.
    fn containers(src: &str) -> Vec<(String, bool, Option<DirectiveForm>)> {
        let mut ed = Editor::new_ext(
            src.as_bytes(),
            Format::Markdown,
            twig::MarkdownExtensions {
                directives: true,
                html_elements: true,
                ..Default::default()
            },
        )
        .unwrap();
        ed.nodes()
            .unwrap()
            .iter()
            .filter(|n| n.kind == Kind::Container)
            .map(|n| {
                (
                    n.name.clone().unwrap_or_default(),
                    container_is_directive(n),
                    n.directive_form,
                )
            })
            .collect()
    }

    #[test]
    fn a_directive_and_an_html_element_are_told_apart_by_spelling_not_by_form() {
        // twig 2.8 folded `div`/`span`/`directive`/`element` into one `container`
        // kind. `directive_form` reads as though it separates them and does not:
        // a block-level `<div>` reports `Some(DirectiveForm::Container)` exactly
        // as a `:::note` does. Trusting it would draw directive chrome — a tinted
        // panel, a `.class` audience label — on every pasted Slack/Docs div.
        for (src, name, want) in [
            (":::note{.a}\nbody\n:::\n", "note", true),
            ("::embed{src=x}\n", "embed", true),
            ("a :vis[hi]{.b} b\n", "vis", true),
            ("<div class=\"x\">\nhi\n</div>\n", "div", false),
            ("<video src=\"v.mp4\" controls></video>\n", "video", false),
            ("<audio src=\"a.mp3\" controls></audio>\n", "audio", false),
            ("<figure>\n\nhi\n\n</figure>\n", "figure", false),
            // The `:` in an attribute must not read as a directive opener: the
            // `<` of the tag comes first, and first one wins.
            (
                "<video src=\"http://x.test/v.mp4\" controls></video>\n",
                "video",
                false,
            ),
            (
                "<source media=\"(prefers-color-scheme: dark)\" srcset=\"d.svg\">\n",
                "source",
                false,
            ),
        ] {
            let found = containers(src);
            let hit = found.iter().find(|(n, ..)| n == name);
            let Some((_, is_directive, form)) = hit else {
                panic!("no `{name}` container in {src:?} — found {found:?}");
            };
            assert_eq!(*is_directive, want, "{name} in {src:?} (form was {form:?})");
        }

        // And the reason this can't just read the field: for the one collision
        // that matters, the field says the same thing for both.
        let div = containers("<div class=\"x\">\nhi\n</div>\n");
        let note = containers(":::note{.a}\nbody\n:::\n");
        assert_eq!(
            div[0].2, note[0].2,
            "if these ever differ, `directive_form` became usable and this rule can go"
        );
    }

    #[test]
    fn a_directive_nested_in_a_quote_or_list_is_still_a_directive() {
        // A container's span opens with its *block prefix*, not its own markup —
        // `> ::embed{…}` starts at the `>`. Reading only the first byte to tell a
        // directive from an element (both `container` since 2.8) therefore misses
        // every nested one, and the placeholder silently renders as nothing.
        for (src, ctx) in [
            ("> ::embed{src=\"x\"}\n", "quoted"),
            ("- ::embed{src=\"x\"}\n", "listed"),
            (">> ::embed{src=\"x\"}\n", "twice quoted"),
        ] {
            let m = map_directives(src);
            assert_eq!(m.directives.len(), 1, "{ctx} directive was lost");
            assert_eq!(m.directives[0].name, "embed", "{ctx}");
        }
    }

    #[test]
    fn a_video_is_still_media_and_not_a_directive() {
        // The other side of the same coin: `<video>` is a `container` too, and
        // must reach `block_media` rather than the directive arms.
        let doc = doc_built("<video src=\"clip.mp4\" controls></video>\n");
        assert_eq!(doc.vmap.media.len(), 1, "the video is block media");
        assert!(
            doc.vmap.rows.iter().all(|r| !r.directive),
            "the video drew directive chrome"
        );
    }

    #[test]
    fn a_directive_needs_the_extension_flag() {
        // `map` (twig's default extensions) leaves `directives` off — the fence
        // renders as literal paragraph text, same as any other unrecognized
        // punctuation, never corrupting or panicking.
        let src = ":::vis{.public}\nhello\n:::\n";
        let m = map(src);
        assert!(m.rows.iter().all(|r| !r.directive));
        assert!(rendered(&m).contains(":::vis{.public}"));
    }

    #[test]
    fn a_footnote_reference_keeps_its_paragraph_visible() {
        // Regression: `footnote_reference` was in neither `is_inline_kind` nor
        // the inline walker, so a paragraph carrying one failed the "all children
        // inline" test, was walked as a container of blocks, and rendered as
        // empty rows with no caret stop anywhere — the whole line vanished.
        let src = "A claim[^1] and more.\n";
        let m = map(src);
        assert_eq!(rendered(&m).trim_end(), "A claim[1] and more.");
        // The `^` is spelling, not text: hidden the way a link's `](dest)` is.
        assert!(!rendered(&m).contains('^'));
    }

    #[test]
    fn a_footnote_reference_is_raised_and_the_prose_around_it_is_not() {
        // What makes `[1]` read as a reference rather than as bracketed text.
        // The brackets ride with the label: the chip is one raised mark.
        let m = map("A claim[^1] and more.\n");
        assert_eq!(baselines_of(&m, '1'), vec![Baseline::Super]);
        assert_eq!(baselines_of(&m, '['), vec![Baseline::Super]);
        assert_eq!(baselines_of(&m, ']'), vec![Baseline::Super]);
        assert_eq!(baselines_of(&m, 'A'), vec![Baseline::Normal]);
    }

    #[test]
    fn a_footnote_reference_keeps_the_link_role_it_had() {
        // The raised baseline is added to the role, not swapped for it: every
        // frontend already paints `Role::Link`, and a reference is one.
        let m = map("A claim[^1].\n");
        let label = m
            .rows
            .iter()
            .flat_map(|r| &r.glyphs)
            .find(|g| g.ch == '1')
            .unwrap();
        assert_eq!(label.style.role, Role::Link);
        assert_eq!(label.style.baseline, Baseline::Super);
    }

    /// The [`Role`] of the first glyph spelling `ch` — how a test reads one
    /// run's styling off a map without caring which row it landed on.
    fn role_of(m: &VisualMap, ch: char) -> Role {
        m.rows
            .iter()
            .flat_map(|r| r.glyphs.iter())
            .find(|g| g.ch == ch)
            .unwrap_or_else(|| panic!("no glyph spelling {ch:?}"))
            .style
            .role
    }

    #[test]
    fn a_markdown_highlight_is_a_mark_and_a_coloured_one_names_its_colour() {
        // twig 3.3's `highlight`/`highlight_colors`, which `parse_extensions`
        // turns on for every leaf document: `==text==` is a `mark` in Markdown
        // and not the literal `==` it used to be, and `==🔴 text==` is one
        // carrying a colour.
        //
        // `doc_built` rather than `map`, deliberately — the extensions are
        // leaf's choice, not twig's default, so a test that parsed bare
        // Markdown here would be testing a document leaf never builds.
        let doc = doc_built("Plain ==yes== and ==🔴 red== ok\n");
        assert_eq!(role_of(&doc.vmap, 'y'), Role::Mark(None));
        assert_eq!(
            role_of(&doc.vmap, 'r'),
            Role::Mark(Some(MarkColor::Red)),
            "the `data-color` twig stripped the emoji into"
        );
        assert_eq!(role_of(&doc.vmap, 'P'), Role::Body);
    }

    #[test]
    fn the_emoji_that_named_a_highlight_is_markup_and_never_drawn() {
        // The colour is *spelling*: twig strips the emoji out of the mark's
        // content, so the reader sees the words and the wash, never the circle.
        // Drawing it would put a character in the rendered text that the author
        // wrote as syntax — the same mistake as drawing an emphasis's `*`.
        let doc = doc_built("Plain ==yes== and ==🔴 red== ok\n");
        let drawn: String = doc.vmap.rows[0].glyphs.iter().map(|g| g.ch).collect();
        assert_eq!(drawn, "Plain yes and red ok");
    }

    #[test]
    fn a_superscript_and_a_subscript_sit_off_the_baseline() {
        // Regression: both rendered flat, so the toolbar's superscript button
        // produced markup that looked exactly like the text around it.
        let m = map_djot("H~2~O and x^2^\n");
        assert_eq!(baselines_of(&m, '2'), vec![Baseline::Sub, Baseline::Super]);
        assert_eq!(baselines_of(&m, 'H'), vec![Baseline::Normal]);
        assert_eq!(baselines_of(&m, 'O'), vec![Baseline::Normal]);
    }

    #[test]
    fn a_raised_glyph_keeps_the_style_it_was_raised_out_of() {
        // Why this is a `Baseline` and not a `Role`: raising a glyph says where
        // it sits, and must not cost it what it already was.
        let m = map_djot("# Heading x^2^\n");
        let two = m
            .rows
            .iter()
            .flat_map(|r| &r.glyphs)
            .find(|g| g.ch == '2')
            .unwrap();
        assert_eq!(two.style.baseline, Baseline::Super);
        assert_eq!(two.style.role, Role::Heading(1), "still heading text");
    }

    #[test]
    fn a_footnote_references_brackets_are_decoration_and_only_its_label_is_a_stop() {
        let src = "see[^note] here\n";
        let m = map(src);
        // `[^note]` spans 3..10, its label `note` 5..9. The caret walks the
        // label; the brackets are drawn but never stood on, as a table's are,
        // and the `[^`/`]` bytes are stepped over like any hidden delimiter.
        let stops: Vec<usize> = m
            .rows
            .iter()
            .flat_map(|r| &r.glyphs)
            .filter(|g| g.stop)
            .map(|g| g.src)
            .collect();
        for off in 5..9 {
            assert!(
                stops.contains(&off),
                "label byte {off} isn't a caret stop: {stops:?}"
            );
        }
        for off in [3usize, 4, 9] {
            assert!(
                !stops.contains(&off),
                "delimiter byte {off} is a caret stop: {stops:?}"
            );
        }
    }

    #[test]
    fn a_task_item_draws_its_box_where_the_bullet_would_be() {
        // Regression: the `[ ] ` is markup twig consumes — the item's paragraph
        // content starts past it — so a task item used to render as `• todo`,
        // identical to a plain bullet and with no way to see it was ticked.
        let m = map("- [ ] todo\n- [x] done\n- plain\n");
        assert_eq!(rendered(&m), "☐ todo\n☑ done\n• plain");

        // The tick rides the item's first row, for a GUI that paints its own box.
        let ticks: Vec<Option<bool>> = m.rows.iter().map(|r| r.task).collect();
        assert_eq!(ticks, [Some(false), Some(true), None]);
    }

    #[test]
    fn a_task_items_box_survives_a_wrap_and_marks_only_the_first_row() {
        let m = map_at(
            "- [x] a much longer task that has to wrap somewhere\n",
            Some(20),
        );
        assert!(m.rows.len() > 1, "the item should wrap: {:?}", rendered(&m));
        assert_eq!(m.rows[0].task, Some(true));
        assert!(
            m.rows[1..].iter().all(|r| r.task.is_none()),
            "only the first row"
        );
        // The continuation lines hang under the box, not under column zero.
        assert!(
            rendered(&m)
                .lines()
                .nth(1)
                .is_some_and(|l| l.starts_with("  "))
        );
    }

    #[test]
    fn a_bracket_in_an_items_prose_is_not_a_checkbox() {
        // `task_checked` finds the box past the list marker; a plain item whose
        // text merely contains a bracket has none, and must keep its bullet.
        let m = map("- see [1] below\n");
        assert_eq!(rendered(&m), "• see [1] below");
        assert_eq!(m.rows[0].task, None);
    }

    #[test]
    fn a_footnote_definition_renders_where_it_was_written() {
        // Regression: twig parses `[^1]: …` as a root *beside* `doc` — not a
        // child of it — so the walk from `doc` never reached one and every byte
        // of the note's body rendered as nothing at all.
        let src = "A claim[^1].\n\n[^1]: The note body.\n\nAfter.\n";
        let m = map(src);
        let text = rendered(&m);
        assert!(
            text.contains("The note body."),
            "the note body is invisible: {text:?}"
        );
        // In source order — between the paragraph that cites it and the one
        // after — not hoisted to the end, and marked to match its reference.
        let lines: Vec<&str> = text.lines().filter(|l| !l.trim().is_empty()).collect();
        assert_eq!(lines, ["A claim[1].", "[1] The note body.", "After."]);
    }

    #[test]
    fn a_footnote_definitions_body_maps_to_its_own_source_bytes() {
        let src = "x[^a].\n\n[^a]: body\n";
        let m = map(src);
        // `body` sits at 14..18. Its glyphs must map there — a marker that ate
        // the offsets would put the caret in the wrong place on every click.
        let body: Vec<(char, usize)> = m
            .rows
            .iter()
            .flat_map(|r| &r.glyphs)
            .filter(|g| g.stop && g.src >= 14)
            .map(|g| (g.ch, g.src))
            .collect();
        assert_eq!(body, [('b', 14), ('o', 15), ('d', 16), ('y', 17)]);
    }

    #[test]
    fn an_empty_footnote_definition_still_shows_its_marker() {
        // The instant `[^1]: ` has been typed and nothing after it. `blocks`
        // renders no child, so without the explicit marker row the definition
        // wouldn't appear at all until something was typed into it.
        let src = "x[^1]\n\n[^1]:\n";
        let m = map(src);
        assert!(
            rendered(&m).contains("[1] "),
            "no marker row: {:?}",
            rendered(&m)
        );
    }

    #[test]
    fn a_footnote_definition_wearing_a_long_label_indents_its_wrapped_body() {
        let src = "x[^src]\n\n[^src]: one two three four five six seven\n";
        let m = map_at(src, Some(24));
        let text = rendered(&m);
        let lines: Vec<&str> = text.lines().filter(|l| !l.trim().is_empty()).collect();
        // Continuation lines hang under the marker, as a list item's do — the
        // indent is the marker's own width, not a fixed one.
        assert_eq!(lines[1].trim_end(), "[src] one two three four");
        assert!(
            lines[2].starts_with("      "),
            "body doesn't hang: {:?}",
            lines[2]
        );
        assert_eq!(lines[2].trim(), "five six seven");
    }

    #[test]
    fn a_code_block_leaves_exactly_one_blank_row_below_it() {
        // The closing fence line used to be miscounted as a blank separator,
        // opening a phantom second gap under the block. One block boundary is
        // one blank row, code block or not.
        let src = "para\n\n```\ncode\n```\n\nafter\n";
        let m = map(src);
        let code_end = m.code_blocks[0].rows_span.end;
        let after = m
            .rows
            .iter()
            .position(|r| r.glyphs.iter().map(|g| g.ch).collect::<String>() == "after")
            .unwrap();
        assert_eq!(
            after - code_end,
            1,
            "exactly one row between code and 'after'"
        );
    }

    #[test]
    fn a_fenced_block_publishes_its_language_on_its_code_block() {
        // The info string becomes the block's label; a bare fence and an indented
        // block carry none.
        assert_eq!(
            map("```rust\nlet x = 1;\n```\n").code_blocks[0]
                .lang
                .as_deref(),
            Some("rust")
        );
        assert_eq!(map("```\nplain\n```\n").code_blocks[0].lang, None);
        assert_eq!(map("    indented\n").code_blocks[0].lang, None);
    }

    fn lang_of(src: &str) -> Option<String> {
        map(src).code_blocks[0].lang.clone()
    }

    #[test]
    fn a_fence_in_a_quote_has_its_language() {
        // The block's span starts at the line, `> ` and all; the fence is past
        // the quote marker, and so is its info string.
        let src = "> ```rust\n> let x = 1;\n> ```\n";
        assert_eq!(lang_of(src).as_deref(), Some("rust"));
        let info = code_info_span(src, 0).unwrap();
        assert_eq!(&src[info], "rust", "the info string's exact bytes");
        assert_eq!(
            lang_of("> > ~~~ py\n> > x\n> > ~~~\n").as_deref(),
            Some("py")
        );
    }

    #[test]
    fn a_fence_in_a_list_item_has_its_language() {
        // On the item's own line, past its marker…
        assert_eq!(
            lang_of("- ```rust\n  let x = 1;\n  ```\n").as_deref(),
            Some("rust")
        );
        assert_eq!(
            lang_of("10. ```rust\n    let x = 1;\n    ```\n").as_deref(),
            Some("rust")
        );
        // …and on a line of its own inside the item, where the indent is the
        // item's content column and not the fence's.
        assert_eq!(
            lang_of("10. a\n\n    ```rust\n    let x = 1;\n    ```\n").as_deref(),
            Some("rust")
        );
        assert_eq!(
            lang_of("- a\n  - b\n\n    ```rust\n    x\n    ```\n").as_deref(),
            Some("rust")
        );
        // In a list in a quote.
        assert_eq!(
            lang_of("> - a\n>\n>   ```rust\n>   x\n>   ```\n").as_deref(),
            Some("rust")
        );
    }

    #[test]
    fn an_indented_block_in_a_list_item_still_has_no_language() {
        // Four spaces past the item's content column is an indented code
        // block, whatever its text looks like — the allowance is measured from
        // the item, not dropped.
        let src = "- a\n\n      ```rust\n";
        assert_eq!(map(src).code_blocks.len(), 1);
        assert_eq!(lang_of(src), None);
        assert_eq!(lang_of("- a\n\n      indented\n"), None);
    }

    /// The token every glyph spelling `ch` carries, in row order — how a test
    /// reads a block's highlighting off the map.
    fn tokens_of(m: &VisualMap, ch: char) -> Vec<Option<Token>> {
        m.rows
            .iter()
            .flat_map(|r| r.glyphs.iter())
            .filter(|g| g.ch == ch)
            .map(|g| g.style.token)
            .collect()
    }

    #[cfg(feature = "syntax")]
    #[test]
    fn a_fenced_block_in_a_known_language_carries_tokens() {
        // `let` is a keyword, the string literal a string, and the plain
        // identifier `x` nothing at all — it draws in the code colour. Every
        // glyph is still `Role::Code`: a token is beside the role, not instead.
        let m = map("```rust\nlet x = \"s\";\n```\n");
        assert_eq!(tokens_of(&m, 'l'), vec![Some(Token::Keyword)]);
        assert_eq!(tokens_of(&m, 'x'), vec![None]);
        assert_eq!(tokens_of(&m, '"'), vec![Some(Token::String); 2]);
        assert!(
            m.rows
                .iter()
                .filter(|r| r.code)
                .flat_map(|r| r.glyphs.iter())
                .all(|g| g.style.role == Role::Code),
            "a token replaced the code role"
        );
    }

    #[cfg(feature = "syntax")]
    #[test]
    fn a_token_changes_nothing_about_where_a_glyph_is() {
        // The same block with and without a language it can be highlighted in
        // lays out identically: same rows, same offsets, same stops. Only the
        // token differs, so the caret walks a highlighted block as it walked an
        // unhighlighted one.
        let hl = map("```rust\nlet x = 1; // c\nfn f() {}\n```\n");
        let plain = map("```text\nlet x = 1; // c\nfn f() {}\n```\n");
        assert_eq!(hl.rows.len(), plain.rows.len());
        for (a, b) in hl.rows.iter().zip(&plain.rows) {
            assert_eq!(a.end_src, b.end_src);
            assert_eq!(a.glyphs.len(), b.glyphs.len());
            for (ga, gb) in a.glyphs.iter().zip(&b.glyphs) {
                assert_eq!((ga.ch, ga.src, ga.stop), (gb.ch, gb.src, gb.stop));
                assert_eq!(ga.style.token(None), gb.style);
            }
        }
        assert!(tokens_of(&hl, 'l').iter().any(Option::is_some));
        assert!(tokens_of(&plain, 'l').iter().all(Option::is_none));
    }

    #[test]
    fn a_block_with_no_language_to_highlight_in_carries_no_tokens() {
        // A bare fence, an indented block, a fence in a language no grammar
        // covers, and inline code all draw as plain code — and so does a
        // `rust` fence when the `syntax` feature is off.
        for src in [
            "```\nlet x = 1;\n```\n",
            "    let x = 1;\n",
            "```no-such-language\nlet x = 1;\n```\n",
            "a `let x` b\n",
        ] {
            assert!(
                tokens_of(&map(src), 'l').iter().all(Option::is_none),
                "{src:?} was highlighted"
            );
        }
        #[cfg(not(feature = "syntax"))]
        assert!(
            tokens_of(&map("```rust\nlet x = 1;\n```\n"), 'l')
                .iter()
                .all(Option::is_none)
        );
    }

    #[test]
    fn inline_code_is_not_a_code_block() {
        // A `code` span inside prose is styled by role, not boxed: it's part of a
        // normal paragraph row, so it names no `code_blocks` entry.
        let m = map("a `snippet` b\n");
        assert!(m.code_blocks.is_empty(), "inline code wrongly boxed");
        assert!(
            m.rows.iter().all(|r| !r.code),
            "inline code flagged a code row"
        );
    }

    #[test]
    fn caret_steps_over_hidden_delimiters() {
        // "a **bold** c": bytes 8,9 are the closing ** — no glyph. Moving right
        // from 'd' (src 7) lands on the space before 'c' (src 10), not inside **.
        let m = map("a **bold** c\n");
        let (r, c) = m.pos_of_offset(7);
        assert_eq!(m.offset_of_pos(r, c + 1), 10);
    }

    // ── the structural view of a table ───────────────────────────────────────

    #[test]
    fn a_table_is_published_structurally_beside_its_picture() {
        let m = map(TABLE);
        let t = &m.tables[0];
        let cell = |r: usize, c: usize| -> String {
            t.grid[r].cells[c].glyphs.iter().map(|g| g.ch).collect()
        };
        assert_eq!(t.grid.len(), 3, "head + two body rows");
        assert_eq!(
            (cell(0, 0), cell(0, 1), cell(1, 0), cell(2, 1)),
            ("Name".into(), "Qty".into(), "Pear".into(), "12".into())
        );
        assert_eq!(
            t.grid.iter().map(|r| r.head).collect::<Vec<_>>(),
            [true, false, false]
        );
        // The alignment the delimiter row spelled, carried per cell — the only
        // place it survives, since the parser consumes that row.
        assert!(matches!(t.grid[1].cells[0].align, Alignment::Left));
        assert!(matches!(t.grid[1].cells[1].align, Alignment::Right));
    }

    #[test]
    fn a_block_media_is_published_structurally_beside_its_placeholder() {
        let m = map("intro\n\n![a cat](img/cat.png)\n\nend\n");
        assert_eq!(m.media.len(), 1, "one block image");
        let img = &m.media[0];
        assert_eq!(img.destination, "img/cat.png");
        assert_eq!(img.alt, "a cat");
        // The placeholder row named by `rows_span` carries the label a plain
        // surface paints and a capable frontend replaces.
        let row_text = |r: usize| -> String { m.rows[r].glyphs.iter().map(|g| g.ch).collect() };
        assert_eq!(
            img.rows_span.end - img.rows_span.start,
            1,
            "one placeholder row"
        );
        assert_eq!(row_text(img.rows_span.start), "🖼 a cat");
        // The row carries the mark `media_spans` derives the side-table from.
        assert!(m.rows[img.rows_span.start].media.is_some());
    }

    #[test]
    fn an_image_without_alt_labels_itself_with_its_filename() {
        let m = map("![](photos/beach.jpg)\n");
        let row = &m.rows[m.media[0].rows_span.start];
        assert_eq!(
            row.glyphs.iter().map(|g| g.ch).collect::<String>(),
            "🖼 beach.jpg"
        );
        assert_eq!(m.media[0].alt, "");
    }

    #[test]
    fn an_empty_cells_home_is_read_from_either_shape_of_span() {
        // A whole-row span: the cell's pipes are the `col`-th and next.
        let row = "|  |  |";
        assert_eq!(empty_cell_offset(row, 10, 0), 12);
        assert_eq!(empty_cell_offset(row, 10, 1), 15);
        // A cell's own span, opening pipe to closing pipe exclusive: the same
        // homes, each read from its own span.
        assert_eq!(empty_cell_offset("|  ", 10, 0), 12);
        assert_eq!(empty_cell_offset("|  ", 13, 1), 15);
        // Nothing to stand in: just inside the pipe, never past the span.
        assert_eq!(empty_cell_offset("|", 10, 0), 11);
        assert_eq!(empty_cell_offset("", 10, 1), 10);
    }

    #[test]
    fn a_hidden_marks_content_end_is_a_caret_home_but_not_a_glyph_stop() {
        // `a **bold** b`: the `d` is at 7, the content ends at 8, the closing
        // `**` draws nothing, and the space after it is at 10. Two homes at one
        // spot on screen: 8 (inside the bold) and 10 (past it).
        let m = map("a **bold** b\n");
        assert!(
            !m.stops.contains(&8),
            "8 has no glyph, so it is no glyph stop"
        );
        assert_eq!(m.mark_ends, vec![8]);
        assert!(m.is_stop(8), "but the caret may rest there");
        assert_eq!(m.snap_to_stop(8), 8, "and is left there when placed there");
        // Left/Right take both homes; the character-pairing walk takes one.
        assert_eq!(m.caret_stop_after(7), Some(8));
        assert_eq!(m.caret_stop_after(8), Some(10));
        assert_eq!(m.caret_stop_before(10), Some(8));
        assert_eq!(m.caret_stop_before(8), Some(7));
        assert_eq!(m.stop_after(7), Some(10));
        assert_eq!(m.stop_before(10), Some(7));
        // Drawn where the next glyph is: after the `d`, not on it.
        assert_eq!(m.pos_of_offset(8), m.pos_of_offset(10));
    }

    #[test]
    fn every_hidden_inline_mark_gives_its_content_end_a_home() {
        // One end per mark, whatever it is spelled with; nested marks closing
        // together share the outer's end and the inner's alike.
        assert_eq!(
            map("*em* `code` [link](u) ~~del~~\n").mark_ends,
            vec![3, 10, 17, 27]
        );
        assert_eq!(map("***both***\n").mark_ends, vec![7]);
        // A mark that closes at its row's end coincides with the row's own end
        // stop — one offset, in both tables.
        let m = map("**bold**\n");
        assert_eq!(m.mark_ends, vec![6]);
        assert!(m.stops.contains(&6));
        // Revealed, the delimiter is glyphs of its own and the end is an
        // ordinary glyph stop: nothing to add.
        let mut ed = Editor::new_str("a **bold** b\n", Format::Markdown).unwrap();
        let src = "a **bold** b\n";
        let revealed = build(
            &ed.nodes().unwrap(),
            src,
            Some(80),
            false,
            &Surface::default(),
            Some(Reveal::full(0..src.len())),
        );
        assert!(revealed.mark_ends.is_empty());
        assert!(revealed.stops.contains(&8));
    }

    #[test]
    fn a_marks_content_end_is_a_home_inside_a_table_cell() {
        let src = "| A | B |\n| --- | --- |\n| **bold** | other |\n";
        let m = map(src);
        let end = src.find("bold").unwrap() + 4; // 32, before the closing `**`
        assert_eq!(m.mark_ends, vec![end]);
        assert_eq!(m.snap_to_stop(end), end);
        // Drawn after the `d`, in this cell — where the cell's own end stop is.
        assert_eq!(m.pos_of_offset(end), m.pos_of_offset(end + 2));
    }

    #[test]
    fn a_block_media_gives_the_caret_a_home_before_and_after_it() {
        // `![x](y)` on its own line: the caret can rest in front of the image
        // (its start) and just past it (the row end), and nowhere inside the
        // markup — the same coarse mapping a thematic break uses.
        let src = "![x](y.png)\n";
        let m = map(src);
        let img = &m.rows[m.media[0].rows_span.start];
        let start = 0; // the image opens the document
        let end = "![x](y.png)".len();
        // Every placeholder glyph maps to the image start and is a stop there.
        assert!(img.glyphs.iter().all(|g| g.src == start && g.stop));
        assert_eq!(img.end_src, end, "the row ends past the image");
        assert_eq!(m.stops.first(), Some(&start));
        assert!(m.stops.contains(&end), "a stop sits after the image");
        // Nothing inside the markup is a stop.
        assert!(!m.stops.iter().any(|&s| s > start && s < end));
    }

    #[test]
    fn an_inline_image_amid_text_is_not_a_block_media() {
        // An image sharing its line with prose isn't block-level: it stays in the
        // inline path (rendered as its alt text), and publishes no MediaInfo.
        let m = map("see ![a cat](cat.png) here\n");
        assert!(m.media.is_empty(), "not a block image");
        assert!(
            rendered(&m).contains("a cat"),
            "alt text still renders inline"
        );
    }

    /// The block images `Doc` publishes for `src`, driven through the real
    /// production build (`build_visual` → `build_cached`) with `html_elements`
    /// on — the path a `<picture>` actually travels. Not the raw `build` the
    /// other tests use: the editor's flat whole-arena snapshot tangles the links
    /// of inline-promoted HTML (phantom roots, dangling `parent`s), which only
    /// the per-block subtree walk `build_cached` does untangles.
    fn doc_media(src: &str) -> Vec<MediaInfo> {
        let mut doc = crate::Doc::from_source(src.to_string(), Format::Markdown).unwrap();
        doc.build_visual(80);
        doc.vmap.media.clone()
    }

    #[test]
    fn a_video_block_is_media_with_its_src_poster_and_kind() {
        // The load-bearing assumption of video support: twig has no `video` node
        // kind, so `html_elements` promotion must land a `<video>` as a generic
        // `element` whose tag name and attributes survive onto `FlatNode` — the
        // same treatment `<picture>` gets. If that ever stops holding, this is
        // the test that says so.
        let m = doc_media("<video src=\"clip.mp4\" poster=\"still.png\" controls>\n</video>\n");
        assert_eq!(m.len(), 1, "the video is one block media");
        assert_eq!(m[0].kind, MediaKind::Video);
        assert_eq!(m[0].destination, "clip.mp4");
        assert_eq!(m[0].poster, "still.png");
    }

    #[test]
    fn a_single_line_video_is_a_block_too() {
        // The spelling everyone actually writes. It used to parse as a paragraph
        // of raw inline HTML — CommonMark opens a block on a complete tag only
        // when the line ends there, and its fixed tag list predates `<video>` —
        // so the tags never reached core as an element at all. twig 2.5.1 widened
        // that list under `html_elements`; this is the test that would catch the
        // pin sliding back.
        let m = doc_media("<video src=\"clip.mp4\" controls></video>\n");
        assert_eq!(m.len(), 1, "single-line <video> is a block");
        assert_eq!(m[0].kind, MediaKind::Video);
        assert_eq!(m[0].destination, "clip.mp4");
    }

    #[test]
    fn a_single_line_picture_is_a_block_with_its_alternatives() {
        // `<picture>` had the identical gap and it went unnoticed because the
        // conventional spelling breaks the lines. Same twig fix covers it.
        let src = "<picture><source media=\"(prefers-color-scheme: dark)\" srcset=\"d.svg\">\
                   <img src=\"l.svg\" alt=\"banner\"></picture>\n";
        let m = doc_media(src);
        assert_eq!(m.len(), 1);
        assert_eq!(m[0].kind, MediaKind::Image);
        assert_eq!(m[0].destination, "l.svg");
        assert_eq!(m[0].resolve(ColorScheme::Dark), "d.svg");
    }

    #[test]
    fn an_audio_block_is_media_with_no_poster() {
        let m = doc_media("<audio src=\"take.mp3\" controls>\n</audio>\n");
        assert_eq!(m.len(), 1);
        assert_eq!(m[0].kind, MediaKind::Audio);
        assert_eq!(m[0].destination, "take.mp3");
        assert!(m[0].poster.is_empty(), "audio has no poster frame");
    }

    #[test]
    fn a_videos_source_children_are_its_candidates_typed_by_mime() {
        // A `<video>` with no `src` of its own — the common shape, since it's how
        // you offer more than one codec. The candidates come from `<source src>`
        // (not `srcset`, which is `<picture>`'s spelling) and carry their MIME.
        let src = "<video controls>\n\
                   <source src=\"a.webm\" type=\"video/webm\">\n\
                   <source src=\"a.mp4\" type=\"video/mp4\">\n\
                   fallback\n\
                   </video>\n";
        let m = doc_media(src);
        assert_eq!(m.len(), 1);
        assert!(
            m[0].destination.is_empty(),
            "no src attribute on the element"
        );
        assert_eq!(m[0].sources.len(), 2);
        assert_eq!(m[0].sources[0].srcset, "a.webm");
        assert_eq!(m[0].sources[0].mime, "video/webm");
        assert_eq!(m[0].sources[1].srcset, "a.mp4");
        // With an empty destination, `resolve` falls through to the first
        // candidate rather than handing the frontend nothing to load.
        assert_eq!(m[0].resolve(ColorScheme::Light), "a.webm");
    }

    #[test]
    fn a_video_placeholder_row_carries_its_own_sigil_and_mark() {
        // The placeholder contract images already hold, now for a video: the row
        // renders as a labelled stand-in a plain surface can paint as-is, and
        // carries the mark a capable frontend replaces it from.
        let src = "<video src=\"clip.mp4\" controls>\n</video>\n";
        let mut doc = crate::Doc::from_source(src.to_string(), Format::Markdown).unwrap();
        doc.build_visual(80);
        let row = &doc.vmap.rows[doc.vmap.media[0].rows_span.start];
        let text: String = row.glyphs.iter().map(|g| g.ch).collect();
        assert!(
            text.starts_with('🎬'),
            "video sigil, not the image one: {text:?}"
        );
        assert!(row.media.is_some(), "the mark rides the placeholder row");
    }

    #[test]
    fn a_picture_block_carries_its_source_alternatives() {
        // A `<picture>` with a dark-mode `<source>`: one block image, whose
        // fallback destination is the `<img>` and whose `sources` carry the
        // `<source>`'s media + srcset for a theme-aware frontend to pick.
        let src = "<picture><source media=\"(prefers-color-scheme: dark)\" srcset=\"dark.svg\"><img src=\"light.svg\" alt=\"banner\"></picture>\n";
        let images = doc_media(src);
        assert_eq!(images.len(), 1, "the picture is one block image");
        let img = &images[0];
        assert_eq!(img.destination, "light.svg", "fallback is the <img>");
        assert_eq!(img.alt, "banner");
        assert_eq!(
            img.sources,
            vec![MediaSource {
                media: "(prefers-color-scheme: dark)".into(),
                srcset: "dark.svg".into(),
                mime: String::new(),
            }],
        );
    }

    #[test]
    fn a_picture_inside_a_heading_is_still_a_block_media_with_sources() {
        // fig.md's shape: the banner is an `<h1>` wrapping the `<picture>`.
        let src = "<h1><picture><source media=\"(prefers-color-scheme: dark)\" srcset=\"d.svg\"><img src=\"l.svg\" alt=\"fig\"></picture></h1>\n";
        let images = doc_media(src);
        assert_eq!(images.len(), 1, "heading-wrapped picture is a block image");
        assert_eq!(images[0].destination, "l.svg");
        assert_eq!(images[0].sources.len(), 1);
        assert_eq!(images[0].sources[0].srcset, "d.svg");
    }

    #[test]
    fn a_plain_image_has_no_media_sources() {
        // A bare Markdown image carries an empty `sources` — nothing to pick from.
        let images = doc_media("![alt](p.png)\n");
        assert_eq!(images.len(), 1);
        assert!(
            images[0].sources.is_empty(),
            "no <picture>, no alternatives"
        );
    }

    #[test]
    fn resolve_picks_the_source_matching_the_scheme() {
        let src = "<picture><source media=\"(prefers-color-scheme: dark)\" srcset=\"dark.svg\"><img src=\"light.svg\" alt=\"b\"></picture>\n";
        let images = doc_media(src);
        let img = &images[0];
        // Dark theme takes the dark source; light falls through to the <img>.
        assert_eq!(img.resolve(ColorScheme::Dark), "dark.svg");
        assert_eq!(img.resolve(ColorScheme::Light), "light.svg");
    }

    #[test]
    fn resolve_falls_back_for_a_plain_image_and_unknown_media() {
        // A plain image ignores the scheme.
        let plain = doc_media("![a](p.png)\n");
        assert_eq!(plain[0].resolve(ColorScheme::Dark), "p.png");

        // A <source> with an unrecognized media query is skipped; a light source
        // is taken under a light theme.
        let m = doc_media(
            "<picture><source media=\"print\" srcset=\"p.svg\"><source media=\"(prefers-color-scheme: light)\" srcset=\"l.svg\"><img src=\"f.svg\" alt=\"x\"></picture>\n",
        );
        assert_eq!(m[0].resolve(ColorScheme::Light), "l.svg");
        assert_eq!(
            m[0].resolve(ColorScheme::Dark),
            "f.svg",
            "no dark source → <img>"
        );
    }

    #[test]
    fn resolve_reads_the_first_srcset_url_ignoring_descriptors() {
        // A comma/descriptor srcset resolves to its first URL.
        assert_eq!(first_srcset_url("a.png 1x, b.png 2x"), Some("a.png"));
        assert_eq!(first_srcset_url("  solo.svg  "), Some("solo.svg"));
        assert_eq!(first_srcset_url(""), None);
        // An empty (unconditional) media always matches.
        assert!(media_matches("", ColorScheme::Light));
        assert!(media_matches(
            "(prefers-color-scheme:dark)",
            ColorScheme::Dark
        ));
        assert!(!media_matches(
            "(prefers-color-scheme: dark)",
            ColorScheme::Light
        ));
    }

    #[test]
    fn a_block_media_carries_its_list_prefix() {
        // An image that is a list item's body opens past the bullet, like every
        // other block does.
        let m = map("- ![alt](p.png)\n");
        let row = &m.rows[m.media[0].rows_span.start];
        let text: String = row.glyphs.iter().map(|g| g.ch).collect();
        assert!(
            text.starts_with("• "),
            "the list marker prefixes the image row: {text:?}"
        );
        assert!(text.contains("🖼 alt"));
    }

    #[test]
    fn the_structural_table_spans_exactly_its_drawn_rows() {
        // A frontend drawing its own grid skips `rows_span` and renders from
        // `grid`. If the span were short the leftover border rows would be
        // painted as text under the real table; if long it would eat a
        // neighbouring paragraph. Both are silent, so pin it to the picture.
        let m = map(&format!("before\n\n{TABLE}\nafter\n"));
        let t = &m.tables[0];
        let row_text = |r: usize| -> String { m.rows[r].glyphs.iter().map(|g| g.ch).collect() };
        assert!(
            row_text(t.rows_span.start).starts_with('┌'),
            "opens on the top border"
        );
        assert!(
            row_text(t.rows_span.end - 1).starts_with('└'),
            "closes on the bottom border"
        );
        assert!(
            !row_text(t.rows_span.start - 1).contains('┌'),
            "the row before the span is not the table's"
        );
        assert_eq!(
            row_text(t.rows_span.end),
            "",
            "the span ends before the gap row"
        );
    }

    #[test]
    fn a_nested_tables_structure_carries_the_block_prefix() {
        // The picture puts the quote's gutter on every row of the grid. A
        // frontend drawing its own table has to draw that too and start past it,
        // so the prefix has to travel with the structure — without it a quoted
        // table renders flush at the margin and leaves the quote it's in.
        let m = map("> | a | b |\n> |---|---|\n> | c | d |\n");
        let t = &m.tables[0];
        let prefix: String = t.prefix.iter().map(|g| g.ch).collect();
        assert_eq!(prefix, "│ ", "the quote's gutter should ride the structure");
        // And it matches what the picture actually drew.
        let drawn: String = m.rows[t.rows_span.start]
            .glyphs
            .iter()
            .map(|g| g.ch)
            .collect();
        assert!(
            drawn.starts_with(&prefix),
            "picture and structure disagree: {drawn:?}"
        );
    }

    #[test]
    fn a_top_level_table_carries_no_prefix() {
        assert!(map(TABLE).tables[0].prefix.is_empty());
    }

    #[test]
    fn structural_cells_are_unwrapped_even_when_the_picture_wraps_them() {
        // The picture wraps a cell to its column; a frontend laying the grid out
        // in pixels needs the text as the document spells it, before that
        // decision. Narrow enough that the drawn cell must break.
        let src = "| Name |\n|------|\n| alpha beta gamma |\n";
        let mut ed = Editor::new_str(src, Format::Markdown).unwrap();
        let m = build_t(&ed.nodes().unwrap(), src, Some(12));
        let drawn = rendered(&m);
        let cell: String = m.tables[0].grid[1].cells[0]
            .glyphs
            .iter()
            .map(|g| g.ch)
            .collect();
        assert_eq!(
            cell, "alpha beta gamma",
            "structure must not carry the wrap"
        );
        assert!(
            drawn.lines().count() > 5,
            "the picture should have wrapped, else this proves nothing:\n{drawn}"
        );
    }

    // ── display columns ──────────────────────────────────────────────────────

    #[test]
    fn a_table_column_is_as_wide_as_its_cells_are_drawn() {
        // A column sized by counting characters is drawn narrower than the text
        // it has to hold — `你好` is two characters in four cells — and the cell
        // spills over the border it is supposed to sit inside, taking the whole
        // grid out of square with it. Squareness is the property: every row of a
        // grid is drawn to the same column, whatever its cells are spelled with.
        for src in [
            "| A | B |\n|---|---|\n| 你好 | y |\n",
            "| A | B |\n|---|---|\n| a👨‍👩‍👧b | y |\n",
            "| A | 漢字 |\n|---|---|\n| x | y |\n",
        ] {
            let m = map(src);
            let widths: Vec<usize> = m.rows.iter().map(|r| r.width()).collect();
            assert!(
                widths.windows(2).all(|w| w[0] == w[1]),
                "ragged grid {widths:?} for {src:?}:\n{}",
                rendered(&m)
            );
        }
    }

    #[test]
    fn a_cell_wrapped_narrow_never_breaks_inside_a_character() {
        // A column too narrow for its cell hard-breaks the text, and every line
        // of it is given an end stop just past its last glyph. Broken into runs
        // of four glyphs, the first line of this cell ends between `👨‍👩` and the
        // joiner holding `👧` on — so its end stop lands inside a character,
        // where a click or Down can reach it and the next Backspace takes the
        // cluster apart from the middle.
        let src = "| A |\n|---|\n| 👨‍👩‍👧👨‍👩‍👧 |\n";
        let mut ed = Editor::new_str(src, Format::Markdown).unwrap();
        let m = build_t(&ed.nodes().unwrap(), src, Some(8));
        let boundaries: Vec<usize> = src
            .grapheme_indices(true)
            .map(|(i, _)| i)
            .chain(std::iter::once(src.len()))
            .collect();
        for off in (0..=src.len()).filter(|&o| m.is_stop(o)) {
            assert!(
                boundaries.contains(&off),
                "stop at {off} is inside a character:\n{}",
                rendered(&m)
            );
        }
    }

    #[test]
    fn a_wrapped_cell_keeps_every_line_inside_its_column() {
        // The width is a promise in a table, where a glyph past the column lands
        // on the border or in the next cell — and it is a promise about cells,
        // which is not what a count of glyphs measures.
        let src = "| A |\n|---|\n| 你好世界漢字 |\n";
        let mut ed = Editor::new_str(src, Format::Markdown).unwrap();
        let m = build_t(&ed.nodes().unwrap(), src, Some(14));
        for r in &m.rows {
            assert_eq!(r.width(), 14, "{:?} is not drawn to the grid", rendered(&m));
        }
    }

    #[test]
    fn a_hard_break_falls_between_clusters_and_measures_in_cells() {
        let glyphs = |s: &str| {
            let mut out = Vec::new();
            push_text(&mut out, s, 0, Style::default());
            out
        };
        let piece = |p: &[Glyph]| p.iter().map(|g| g.ch).collect::<String>();

        // Six cells of CJK broken at four: two characters, then one — never
        // between the two cells of `好`.
        let w = glyphs("你好世");
        let pieces: Vec<String> = hard_break(&w, 4).iter().map(|p| piece(p)).collect();
        assert_eq!(pieces, ["你好", "世"]);

        // A character wider than the column has nowhere legal to break, so it
        // keeps its cells rather than being cut in half.
        let w = glyphs("你好");
        let pieces: Vec<String> = hard_break(&w, 1).iter().map(|p| piece(p)).collect();
        assert_eq!(pieces, ["你", "好"]);

        // An empty word yields no pieces at all — a double space stays a space.
        assert!(hard_break(&[], 4).is_empty());
    }

    #[test]
    fn an_empty_list_item_still_gets_a_bulleted_row_with_a_caret_home() {
        // Pressing Enter at the end of a list item opens a new, empty item —
        // a childless `list_item`. Without a row of its own the new bullet
        // wouldn't appear until something was typed into it (the caret would be
        // stranded on an offset no row draws). It now renders as one prefixed
        // row whose end is a caret stop, so the bullet shows and the caret lands
        // just past the marker.
        let m = map("- item\n- \n");
        assert_eq!(m.num_rows(), 2, "the empty second item needs its own row");
        assert_eq!(
            m.rows[1].glyphs.iter().map(|g| g.ch).collect::<String>(),
            "• ",
            "the empty item draws just its bullet",
        );
        // Its end is the caret home (past the `- ` marker), and it's a real stop.
        assert!(
            m.is_stop(m.rows[1].end_src),
            "the empty item's caret home is not a stop"
        );
        assert_eq!(
            m.pos_of_offset(m.rows[1].end_src),
            (1, 2),
            "caret sits after '• '"
        );
    }

    #[test]
    fn a_notes_row_range_stops_at_the_note_even_when_it_ends_in_a_link() {
        // The peek bug: a note whose body ends in a link has its last byte
        // inside the hidden destination, so mapping `end - 1` through
        // `pos_of_offset` snapped *forward* — past its own row, past the drawn
        // gap, and onto the next note's row. The popover then drew both notes.
        let src = "A[^1] B[^2].\n\n[^1]: bare text\n\n[^2]: [title](https://example.com/x)\n\n[^3]: last\n";
        let m = map(src);
        let body = src.find("[title]").unwrap();
        let end = src.find("\n\n[^3]").unwrap();

        let (first, last) = m.row_range_for(body..end);
        assert_eq!(
            first, last,
            "a one-block note is one row, not a span onto the next"
        );

        // The old arithmetic, kept here as the thing that must stay wrong: it
        // is what this method exists instead of.
        assert_ne!(
            m.pos_of_offset(end - 1).0,
            last,
            "the forward snap still leaves the note's row — that is the whole point",
        );

        // A note ending in *visible* text was never broken, and still isn't:
        // both readings agree there, which is why the original test missed it.
        let plain = src.find("bare text").unwrap();
        let plain_end = src.find("\n\n[^2]").unwrap();
        let (pf, pl) = m.row_range_for(plain..plain_end);
        assert_eq!(pf, pl);
        assert_eq!(m.pos_of_offset(plain_end - 1).0, pl);
    }

    #[test]
    fn a_row_range_covers_every_row_of_a_block_that_spans_several() {
        // The range is a span, not a point: a quote of two paragraphs covers its
        // gap row and both of its text rows, so a peek draws the whole thing.
        let src = "> one\n>\n> two\n\nafter\n";
        let m = map(src);
        let (first, last) = m.row_range_for(0..src.find("\n\nafter").unwrap());
        assert_eq!((first, last), (0, 2));

        // And a range with no visible byte at all still covers the row it opened
        // on, rather than collapsing to nothing.
        let (f, l) = m.row_range_for(0..1);
        assert_eq!((f, l), (0, 0));
    }

    #[test]
    fn an_empty_block_quote_still_gets_a_gutter_row_with_a_caret_home() {
        // The peer of the empty list item, and the case that made an empty line
        // in a quote draw as plain body text: a childless `block_quote` — a bare
        // `> `, which is what the toolbar's Quote button leaves on a blank line —
        // has no inner block to carry the gutter, so the whole quote used to
        // render as *nothing*. It didn't merely lose its bar; the row went away
        // and the caret had no home on it.
        let m = map("a\n\n> \n\nb\n");
        assert_eq!(
            m.rows[2].glyphs.iter().map(|g| g.ch).collect::<String>(),
            "│ ",
            "the empty quote draws just its gutter",
        );
        assert!(
            m.rows[2]
                .glyphs
                .iter()
                .all(|g| g.style.role == Role::QuoteGutter)
        );
        assert!(
            !m.rows[2].decoration,
            "it is a line text can go on, not a drawn gap"
        );
        assert!(
            m.is_stop(m.rows[2].end_src),
            "the empty quote's caret home is not a stop"
        );
        assert_eq!(
            m.pos_of_offset(m.rows[2].end_src),
            (2, 2),
            "caret sits after '│ '"
        );

        // And a document that is *only* an empty quote still renders a row — it
        // used to render none at all, leaving the caret nowhere to stand.
        let m = map("> \n");
        assert_eq!(m.num_rows(), 1);
        assert_eq!(
            m.rows[0].glyphs.iter().map(|g| g.ch).collect::<String>(),
            "│ "
        );
    }

    #[test]
    fn a_quotes_own_trailing_marker_lines_stay_inside_the_quote() {
        // Enter at the end of `> a` writes `> a\n>\n> \n`. Those last two lines
        // hold no block — a quote's `content_span` stops at its last child — so
        // the children walk never reaches them, and they used to fall through to
        // the document-level trailing pass, which knows no prefix: the gutter
        // stopped and the writer's new line drew as plain prose. Fixable only
        // since twig 3.2.0, where the quote's *span* covers its own marker lines
        // (`0..3` before, `0..8` now) and there is finally a node saying they
        // are the quote's.
        let m = map("> a\n>\n> \n");
        assert_eq!(m.num_rows(), 3, "one row per line the quote spells");
        for (i, row) in m.rows.iter().enumerate() {
            let text = row.glyphs.iter().map(|g| g.ch).collect::<String>();
            assert!(text.starts_with("│ "), "row {i} lost the gutter: {text:?}");
            assert!(
                !row.decoration,
                "row {i} is a line to type on, not a drawn gap"
            );
            assert!(m.is_stop(row.end_src), "row {i} has no caret home");
        }
        assert_eq!(
            m.rows[0].glyphs.iter().map(|g| g.ch).collect::<String>(),
            "│ a"
        );
        // Distinct offsets, so ↑/↓ between them moves the caret rather than
        // landing twice on the same byte.
        assert!(m.rows[0].end_src < m.rows[1].end_src);
        assert!(m.rows[1].end_src < m.rows[2].end_src);

        // A blank line *after* the quote is not the quote's: it is spelled with
        // no marker, so it stays an ordinary boundary and the gutter ends.
        let m = map("> a\n\nb\n");
        assert_eq!(m.num_rows(), 3);
        assert_eq!(
            m.rows[2].glyphs.iter().map(|g| g.ch).collect::<String>(),
            "b"
        );
        assert!(
            !m.rows[1]
                .glyphs
                .iter()
                .any(|g| g.style.role == Role::QuoteGutter)
        );

        // Nesting is the case this could get wrong, and the depth has to come
        // from which quote's span the line falls in rather than from the row
        // above it. A trailing `>` under `> > a` matches only the OUTER quote,
        // so it wears one gutter; spell it `> >` and it wears two.
        let m = map("> > a\n>\n");
        assert_eq!(
            m.rows[0].glyphs.iter().map(|g| g.ch).collect::<String>(),
            "│ │ a"
        );
        assert_eq!(
            m.rows[1].glyphs.iter().map(|g| g.ch).collect::<String>(),
            "│ "
        );
        let m = map("> > a\n> >\n");
        assert_eq!(
            m.rows[1].glyphs.iter().map(|g| g.ch).collect::<String>(),
            "│ │ "
        );

        // And a marker line BETWEEN two quoted paragraphs is untouched: that is
        // the boundary `emit_separators_before` spells, and it stays a drawn gap
        // rather than becoming a line to type on.
        let m = map("> a\n>\n> b\n");
        assert_eq!(m.num_rows(), 3);
        assert!(
            m.rows[1].decoration,
            "the gap between two quoted blocks is still a gap"
        );
    }

    #[test]
    fn an_empty_ordered_item_gets_its_number_and_a_caret_home() {
        let m = map("1. item\n2. \n");
        assert_eq!(m.num_rows(), 2);
        assert_eq!(
            m.rows[1].glyphs.iter().map(|g| g.ch).collect::<String>(),
            "2. "
        );
        assert!(m.is_stop(m.rows[1].end_src));
        assert_eq!(
            m.pos_of_offset(m.rows[1].end_src),
            (1, 3),
            "caret sits after '2. '"
        );
    }

    #[test]
    fn an_empty_headings_caret_home_is_past_its_hidden_marker() {
        // The toolbar's H1 on a blank line writes `# ` and nothing else. The row
        // it renders is empty (the marker is hidden), so its end *is* its only
        // caret stop — and it has to be the offset past the `# `, where typing
        // continues the heading. Anchored at the block's start instead, the caret
        // drew in front of the hashes and the first character typed there landed
        // before them (`x# `), which isn't a heading at all.
        let m = map("# \n");
        assert_eq!(m.num_rows(), 1);
        assert!(m.rows[0].glyphs.is_empty(), "the `# ` marker is hidden");
        assert_eq!(m.rows[0].end_src, 2, "the caret home is past the marker");
        assert!(m.is_stop(2), "the empty heading's caret home is not a stop");
    }

    #[test]
    fn a_headings_rows_carry_its_level_even_with_nothing_typed_in_it() {
        // The row-level fact a proportional frontend sizes a whole line by. An
        // empty heading has no glyph to read a `Role::Heading` off, so a renderer
        // scanning glyphs drew `# ` (and its caret) at body height until the
        // first character landed.
        let m = map("# \n");
        assert_eq!(
            m.rows[0].heading,
            Some(1),
            "the empty heading knows its level"
        );

        // Every row of one that wraps, not just the first — and nothing else.
        let m = map_at(
            "## a heading long enough to wrap over two rows\n\nbody\n",
            Some(20),
        );
        let heads: Vec<Option<u8>> = m.rows.iter().map(|r| r.heading).collect();
        assert!(
            heads.iter().filter(|h| **h == Some(2)).count() >= 2,
            "got {heads:?}"
        );
        assert_eq!(
            m.rows.last().and_then(|r| r.heading),
            None,
            "the paragraph under it is not a heading",
        );
    }

    #[test]
    fn an_empty_heading_leaves_the_rows_under_it_at_their_own_offsets() {
        // The row's end is also what the *next* row's separator is measured from,
        // so an empty heading that under-reported it shifted every offset below —
        // and the blank line under the heading then claimed the same offset as the
        // heading's own end. `pos_of_offset` resolves such a tie downstream (a
        // soft wrap belongs to the row below), so the caret at the end of the
        // heading was drawn two rows lower, on the blank line.
        // `text\n\n# \n\n`: the heading's content opens at 8, and the two rows
        // under it end at 9 and 10 — the blank line and the document's end.
        let m = map("text\n\n# \n\n");
        let end = m.rows.last().expect("a trailing blank row").end_src;
        assert_eq!(end, 10, "the trailing rows must end at their real offsets");
        // The heading's caret home is its own row's, not one shared with a row
        // below — the tie that drew the caret two rows down.
        assert_eq!(m.pos_of_offset(8), (2, 0), "the empty heading's own row");
        assert!(
            m.rows[3..].iter().all(|r| r.end_src > 8),
            "rows below own later offsets"
        );
    }

    // ── block boundaries ─────────────────────────────────────────────────────

    /// Every drawn boundary in `src`, in order, as `(above, below)`.
    fn boundaries(m: &VisualMap) -> Vec<(BlockClass, BlockClass)> {
        m.rows
            .iter()
            .filter_map(|r| r.boundary)
            .map(|b| (b.above, b.below))
            .collect()
    }

    #[test]
    fn a_boundary_says_which_blocks_it_divides() {
        use BlockClass::*;
        let m = map("one\n\ntwo\n\n# Head\n\ntail\n\n> quoted\n\n```\ncode\n```\n\n");
        assert_eq!(
            boundaries(&m),
            vec![
                (Paragraph, Paragraph),
                (Paragraph, Heading),
                (Heading, Paragraph),
                (Paragraph, Quote),
                (Quote, Code),
                // The blank line the document trails off with is a boundary too
                // — it closes the last block above the empty paragraph the caret
                // rests on. See `emit_trailing_blank_lines`.
                (Code, Paragraph),
            ],
            "each gap names the pair it falls between, in document order"
        );
    }

    #[test]
    fn a_closing_fence_at_the_end_of_the_document_opens_no_phantom_row() {
        // The code block's last row ends at its last line of code; the closing
        // fence under it has no row. Counted from the row, the fence's line read
        // as a trailing blank line and opened an empty paragraph whose offset
        // was inside the fence — typing on it wrote into the backticks.
        let m = map("para\n\n```\ncode\n```\n");
        let texts: Vec<String> = m
            .rows
            .iter()
            .map(|r| r.glyphs.iter().map(|g| g.ch).collect())
            .collect();
        assert_eq!(texts, vec!["para", "", "code"], "a row under the fence");
        assert!(
            m.rows.last().is_some_and(|r| r.code),
            "the last row is the code"
        );
        // One more newline is the real empty paragraph, past the fence.
        let m = map("para\n\n```\ncode\n```\n\n");
        let last = m.rows.last().expect("a trailing row");
        assert_eq!(last.end_src, 20, "the trailing row stands past the fence");
        assert!(!last.decoration, "the trailing row is somewhere to type");
        // A setext underline is the same shape: markup under the last row.
        let m = map("Head\n====\n");
        assert_eq!(
            m.rows.len(),
            1,
            "a row under the underline: {:?}",
            m.rows.len()
        );
    }

    /// Each row as whether it is drawn-only and where it ends — the shape of
    /// the blank lines around a block, which is what its text can't show.
    fn row_shape(m: &VisualMap) -> Vec<(bool, usize)> {
        m.rows.iter().map(|r| (r.decoration, r.end_src)).collect()
    }

    #[test]
    fn the_lines_above_the_first_block_draw_as_the_lines_under_the_last_do() {
        // No row was drawn above the first block, so an empty paragraph opened
        // there drew nothing. As at the document's end, one blank line is only
        // the gap, and every line above that gap is somewhere to type.
        assert_eq!(row_shape(&map("\nb\n")), [(false, 2)]);
        let m = map("\n\nb\n");
        assert_eq!(row_shape(&m), [(false, 0), (true, 1), (false, 3)]);
        assert_eq!(m.content_start, 0, "the caret can reach the empty line");
        assert_eq!(m.pos_of_offset(0), (0, 0));
        assert_eq!(map("b\n").content_start, 0);
        assert_eq!(
            row_shape(&map("\n\n\n# H\n")),
            [(false, 0), (false, 1), (true, 2), (false, 6)]
        );

        // Past frontmatter, the line under it is the frontmatter's.
        let fm = "---\nt: x\n---\n";
        let m = map(&format!("{fm}\nb\n"));
        assert_eq!(row_shape(&m), [(false, 15)]);
        assert_eq!(m.content_start, 14);
        let m = map(&format!("{fm}\n\nb\n"));
        assert_eq!(row_shape(&m), [(false, 13), (true, 14), (false, 16)]);
        assert_eq!(m.content_start, 13);

        // Preserve flow draws every line, but the frontmatter's.
        let preserve = |src: &str| {
            let mut ed = Editor::new_str(src, Format::Markdown).unwrap();
            build(
                &ed.nodes().unwrap(),
                src,
                Some(80),
                true,
                &Surface::default(),
                None,
            )
        };
        assert_eq!(row_shape(&preserve("\nb\n")), [(false, 0), (false, 2)]);
        assert_eq!(row_shape(&preserve(&format!("{fm}\nb\n"))), [(false, 15)]);
        assert_eq!(
            row_shape(&preserve(&format!("{fm}\n\nb\n"))),
            [(false, 14), (false, 16)]
        );

        // Past a comment, counted from the line after it.
        assert_eq!(row_shape(&map("<!-- c -->\n\nb\n")), [(false, 13)]);
        assert_eq!(
            row_shape(&map("<!-- c -->\n\n\nb\n")),
            [(false, 11), (true, 12), (false, 14)]
        );
    }

    #[test]
    fn the_lines_under_a_rule_are_counted_from_the_rule() {
        // A rule's row ends at the caret's home past the newline under it, and
        // the counts of the blank lines below used to start from there too, so
        // every gap under a rule came up a line short: the empty paragraph
        // Enter opens beneath a rule drew as two gaps and no line, and the
        // caret put on it was drawn on the next block. They count from the
        // rule itself, as under any other block's last line.
        let src = "a\n\n---\n\n\n\nb\n";
        let m = map(src);
        assert_eq!(
            row_shape(&m),
            [
                (false, 1),
                (true, 2),
                (false, 7), // the rule, its home past the newline under it
                (true, 7),
                (false, 8), // the empty paragraph
                (true, 9),
                (false, 11),
            ]
        );
        assert_eq!(m.pos_of_offset(8), (4, 0), "the caret on the empty line");
        // The one blank line under a rule is still the one gap, not two.
        assert_eq!(
            row_shape(&map("a\n\n---\n\nb\n")),
            [(false, 1), (true, 2), (false, 7), (true, 7), (false, 9)]
        );

        // Closing the document, the line under the gap is somewhere to type.
        let m = map("a\n\n---\n\n");
        assert_eq!(
            row_shape(&m),
            [(false, 1), (true, 2), (false, 7), (true, 7), (false, 8)]
        );
        assert_eq!(m.pos_of_offset(8), (4, 0));
        // And a lone newline after the rule is only the rule's own.
        assert_eq!(
            row_shape(&map("a\n\n---\n")),
            [(false, 1), (true, 2), (false, 7)]
        );

        // A quote's own trailing lines under a rule: each a row, the first too.
        let m = map("> ---\n>\n> ");
        assert_eq!(row_shape(&m), [(false, 6), (false, 7), (false, 10)]);
    }

    #[test]
    fn the_home_past_a_djot_rule_is_the_line_under_it() {
        // djot's span takes in the newline that ends the rule, and the home
        // past the rule was measured from that end — one newline further on,
        // across the blank line, to the next block's first character. A caret
        // left after the rule was drawn on that block.
        let m = map_djot("a\n\n* * *\n\nb\n");
        assert_eq!(
            row_shape(&m),
            [(false, 1), (true, 2), (false, 9), (true, 9), (false, 11)]
        );
        assert_eq!(m.pos_of_offset(9).0, 2, "beside the rule, not on `b`");
        let m = map_djot("a\n\n* * *\n\n\n\nb\n");
        assert_eq!(m.pos_of_offset(10), (4, 0), "the caret on the empty line");
    }

    // ── hidden blocks ────────────────────────────────────────────────────────

    /// The row texts of `m`, one string per row.
    fn row_texts(m: &VisualMap) -> Vec<String> {
        m.rows
            .iter()
            .map(|r| r.glyphs.iter().map(|g| g.ch).collect())
            .collect()
    }

    #[test]
    fn a_div_s_closing_tag_is_not_a_blank_row() {
        // The `</div>` sits on a line of its own under the div's last child and
        // draws nothing. Counting the separator from the child's end read that
        // line as a blank line between the div and the block below — a
        // navigable empty row the author never opened — and at the end of the
        // file, as an empty trailing paragraph.
        let m = map("above\n\n<div class=\"center\">\n\nhello\n\n</div>\n\nbelow\n");
        assert_eq!(row_texts(&m), ["above", "", "hello", "", "below"]);
        assert!(!m.is_stop(36), "the `</div>` line is not a caret home");
        assert_eq!(
            m.stop_after(34),
            Some(44),
            "from `hello` the next stop is `below`"
        );

        let m = map("above\n\n<div class=\"center\">\n\nhello\n\n</div>\n");
        assert_eq!(row_texts(&m), ["above", "", "hello"], "no trailing rows");
    }

    #[test]
    fn a_comment_between_two_blocks_is_stepped_over_not_drawn_as_a_gap() {
        // `<!-- exec -->` is a top-level block that draws no rows. The blocks
        // either side of it meet across the one boundary a paragraph and a code
        // block always meet across — not that boundary *plus* one blank row per
        // line of the comment, which is what counting the separator from the
        // paragraph's end used to spell.
        let m = map("para one\n\n<!-- exec -->\n```\ncode\n```\n\nafter\n");
        assert_eq!(row_texts(&m), ["para one", "", "code", "", "after"]);
        assert_eq!(
            boundaries(&m),
            vec![
                (BlockClass::Paragraph, BlockClass::Code),
                (BlockClass::Code, BlockClass::Paragraph),
            ],
            "the boundary names the drawn blocks either side, not the comment"
        );
        // The gap stands past the comment, so the caret's row lookup never
        // resolves inside it.
        assert_eq!(
            m.rows[1].end_src, 23,
            "the gap row ends at the comment's end"
        );
    }

    #[test]
    fn a_comment_opening_the_document_draws_no_leading_gap() {
        let m = map("<!-- lead -->\n\npara\n");
        assert_eq!(row_texts(&m), ["para"]);
        assert_eq!(m.content_start, 0, "the comment is still the first block");
    }

    #[test]
    fn a_comment_closing_the_document_is_not_trailing_blank_lines() {
        // Its lines are not blank lines the author opened with Enter, so no
        // gap-plus-empty-paragraph is fabricated under the last drawn block.
        let m = map("para\n\n<!-- trail -->\n");
        assert_eq!(row_texts(&m), ["para"]);
        // Enter at the end of the document still opens the empty paragraph the
        // caret rests on: the newlines *after* the comment count as they would
        // after any block.
        let m = map("para\n\n<!-- trail -->\n\n");
        assert_eq!(row_texts(&m), ["para", "", ""]);
    }

    #[test]
    fn a_comment_in_a_list_item_leaves_the_bullet_to_what_follows_it() {
        // The first *drawn* child wears the item's marker; a hidden first child
        // would otherwise take it and leave the text without one.
        let m = map("- <!-- note -->\n\n  text\n- two\n");
        let texts = row_texts(&m);
        assert!(
            texts.iter().any(|t| t == "• text"),
            "the text wears the bullet: {texts:?}"
        );
        assert!(
            !texts.iter().any(|t| t == "• "),
            "no empty bullet row for the comment: {texts:?}"
        );
    }

    #[test]
    fn the_cached_build_does_not_spell_the_document_out_as_blank_rows_after_a_comment() {
        // The bug as seen: a 200-line document with one comment in it rendered
        // ~200 blank rows after the comment, one per source line, because the
        // comment's per-block builder handed back a `last_off` of 0. Parity with
        // `build` alone would not catch a *shared* wrong answer, so the count is
        // pinned outright.
        let body = (0..200)
            .map(|i| format!("line {i}"))
            .collect::<Vec<_>>()
            .join("\n\n");
        let src = format!("intro\n\n<!-- exec -->\n{body}\n");
        let mut ed = Editor::new_str(&src, Format::Markdown).unwrap();
        let mut cache = BlockCache::default();
        let (plain, cached) = render_both(&mut ed, &src, Some(80), &mut cache);
        assert_maps_eq(&plain, &cached, "comment then 200 paragraphs");
        // intro, then 200 × (gap, paragraph): 401 rows and not a row more.
        assert_eq!(cached.rows.len(), 401);
    }

    #[test]
    fn a_link_reference_definition_is_stepped_over_like_a_comment() {
        // `[a]: /a` is a root beside `doc` with no rows of its own. Merged into
        // the walk it is a hidden block: the blocks either side meet across one
        // boundary, and its line is not a blank row.
        let m = map("see [a]\n\n[a]: /a\n\nafter\n");
        assert_eq!(row_texts(&m), ["see a", "", "after"]);
        assert_eq!(
            boundaries(&m),
            vec![(BlockClass::Paragraph, BlockClass::Paragraph)]
        );
    }

    #[test]
    fn link_reference_definitions_closing_the_document_are_not_trailing_blank_lines() {
        // The README shape: prose, then a `[links]` block nobody reads. Its
        // lines used to be counted as blank ones, an empty paragraph per
        // definition under the last real block.
        let m = map("see [a] and [b]\n\n<!-- links -->\n[a]: /a\n[b]: /b \"bee\"\n");
        assert_eq!(row_texts(&m), ["see a and b"]);
    }

    #[test]
    fn a_definition_glued_under_a_paragraph_stays_inside_it() {
        // `[a]: /a` at the front of a paragraph's lines is stripped from the
        // paragraph's text, but the paragraph's span still starts on its line.
        // Both blocks start at the same offset; the definition, sorted first,
        // is stepped over, and the paragraph draws as it always did — one gap
        // above it, none inside.
        let m = map("intro\n\n[a]: /a\ntext [a]\n");
        assert_eq!(row_texts(&m), ["intro", "", "text a"]);
    }

    #[test]
    fn a_definition_with_no_span_is_left_out_of_the_walk() {
        // twig before 3.3.3 reported `0..0` for every link reference
        // definition. One of those has nowhere to be merged: sorted first by
        // its zero start it would open the document with a phantom block, and
        // the walk would step back to offset 0. It is simply not a block. A
        // footnote definition is always placed; it has a body to draw.
        assert!(!is_placed_definition(&Kind::Reference, &(0..0)));
        assert!(is_placed_definition(&Kind::Reference, &(7..14)));
        assert!(is_placed_definition(&Kind::Footnote, &(0..0)));
        assert!(!is_placed_definition(&Kind::Str, &(7..14)));
    }

    #[test]
    fn the_trailing_gap_closes_the_last_block() {
        // Two Enters at the end of a document: a drawn gap, then the navigable
        // empty paragraph. Only the gap is labelled, so a frontend that shrinks
        // boundaries shrinks the spacer and leaves the row being typed on alone.
        let m = map("# Head\n\n\n");
        assert_eq!(
            boundaries(&m),
            vec![(BlockClass::Heading, BlockClass::Paragraph)]
        );
    }

    #[test]
    fn only_the_drawn_gap_rows_carry_a_boundary() {
        let m = map("one\n\ntwo\n");
        for row in &m.rows {
            assert_eq!(
                row.boundary.is_some(),
                row.decoration,
                "a boundary is exactly a drawn gap row: {:?}",
                row.glyphs.iter().map(|g| g.ch).collect::<String>()
            );
        }
    }

    #[test]
    fn preserve_flow_labels_no_boundary() {
        // Every blank line is a caret home there — somewhere text can go, not a
        // gap between blocks — so nothing is drawn-only and nothing is labelled.
        // A frontend keying its spacing off `boundary` can't shrink a row the
        // author is about to type on.
        let m = map_preserve("one\n\ntwo\n\n# Head\n", Some(80));
        assert!(boundaries(&m).is_empty());
    }

    #[test]
    fn a_list_draws_no_boundary_between_its_items() {
        // Tight or loose, core puts no gap row between two items of one list —
        // so an item↔item boundary is a shape no frontend will ever be handed,
        // and spacing one is spacing something that isn't there.
        for src in ["- one\n- two\n", "- one\n\n- two\n"] {
            let m = map(src);
            assert!(
                boundaries(&m).is_empty(),
                "no gap row inside the list of {src:?}"
            );
        }
        // Leaving the list is an ordinary boundary, and the list is named as
        // what sits above it.
        let m = map("- one\n- two\n\npara\n");
        assert_eq!(
            boundaries(&m),
            vec![(BlockClass::List, BlockClass::Paragraph)]
        );
    }

    #[test]
    fn a_nested_boundary_names_the_blocks_inside_the_container() {
        // Two paragraphs inside a blockquote are divided by a Paragraph↔Paragraph
        // boundary — the quote is the container they're both in, not what the gap
        // separates.
        let m = map("> one\n>\n> two\n");
        assert_eq!(
            boundaries(&m),
            vec![(BlockClass::Paragraph, BlockClass::Paragraph)]
        );
    }

    #[test]
    fn a_directive_container_draws_one_boundary_like_every_other_block() {
        // A container's rows stop at its last *child*, so without anchoring
        // `last_off` past the closing `:::` the separator logic counted the fence
        // line as a blank row of its own and drew the gap twice — one authored
        // blank line, two boundaries, and a frontend spacing each of them put
        // double margin under every fenced div. The code-block arm anchors past
        // its ``` for exactly this reason; compare the two here.
        let fenced = map_directives(":::note\nin\n:::\n\ntwo\n");
        assert_eq!(
            boundaries(&fenced),
            vec![(BlockClass::Directive, BlockClass::Paragraph)],
            "one authored gap, one boundary row"
        );
        let code = map("```\nc\n```\n\ntwo\n");
        assert_eq!(
            boundaries(&code).len(),
            boundaries(&fenced).len(),
            "a fenced div spaces like a fenced code block"
        );
        // Nesting closes several fences at once; still one gap.
        let nested = map_directives(":::a\n:::b\nin\n:::\n:::\n\ntwo\n");
        assert_eq!(
            boundaries(&nested),
            vec![(BlockClass::Directive, BlockClass::Paragraph)]
        );
    }

    #[test]
    fn a_block_media_names_itself_in_the_boundaries_either_side() {
        use BlockClass::*;
        // A block image is never a node of its own — `media_only` promotes the
        // *paragraph* wrapping it — so classifying the node the walk stands on
        // called the picture `Paragraph` and left `BlockClass::Media` unreachable:
        // a frontend could not give a photo more air than a line of prose.
        // `label_media_boundaries` reads it back off the finished rows instead.
        let m = map("one\n\n![alt](p.png)\n\ntwo\n");
        assert_eq!(boundaries(&m), vec![(Paragraph, Media), (Media, Paragraph)]);
        // At the edges of the document too: the leading gap has no boundary of
        // its own, and the trailing one is `emit_trailing_blank_lines`'.
        let edges = map("![a](p.png)\n\nmid\n\n![b](q.png)\n");
        assert_eq!(
            boundaries(&edges),
            vec![(Media, Paragraph), (Paragraph, Media)]
        );
        // One gap spelled with several rows — the row closing the block above and
        // the row opening the one below, with the author's spare blank line
        // navigable between them — carries the same pair on every drawn row.
        let roomy = map("one\n\n\n\n![alt](p.png)\n");
        assert_eq!(
            boundaries(&roomy),
            vec![(Paragraph, Media), (Paragraph, Media)]
        );
    }

    #[test]
    fn a_block_video_is_media_at_its_boundaries_not_a_directive_panel() {
        // Worse than the image case before `label_media_boundaries`: a `<video>`
        // arrives as twig's generic `container`, which classifies `Directive` —
        // the one class a frontend reads as "draw a tinted panel here". A movie
        // got the chrome of a fenced div.
        let mut doc = crate::Doc::from_source(
            "one\n\n<video src=\"v.mp4\"></video>\n\ntwo\n".to_string(),
            Format::Markdown,
        )
        .unwrap();
        doc.build_visual(80);
        assert_eq!(
            boundaries(&doc.vmap),
            vec![
                (BlockClass::Paragraph, BlockClass::Media),
                (BlockClass::Media, BlockClass::Paragraph),
            ]
        );
    }

    #[test]
    fn the_incremental_walk_labels_boundaries_like_the_full_one() {
        // `assert_maps_eq` compares boundaries too, so this pins the two doors
        // into `BlockClass::from_node_kind` — a `FlatNode`'s kind on the full
        // build, a query match's on the cached one — against a document with one
        // of every boundary in it.
        let src = "one\n\n# Head\n\ntwo\n\n- a\n- b\n\n> q\n\n```\nc\n```\n\npara\n";
        let mut ed = Editor::new_str(src, Format::Markdown).unwrap();
        let mut cache = BlockCache::default();
        let (full, cached) = render_both(&mut ed, src, Some(80), &mut cache);
        assert_maps_eq(&full, &cached, "boundary labelling");
        assert!(
            !boundaries(&full).is_empty(),
            "the fixture has boundaries to compare"
        );
    }

    #[test]
    fn every_caret_stop_opens_a_cluster_of_its_row() {
        // The two ways of finding a cluster have to agree. `push_text` marks the
        // stops by segmenting one run of text; the column mapping segments the
        // whole row, decoration and all. A stop that came out as the *middle* of
        // some row-level cluster would be a caret with no column of its own —
        // drawn at the column of whatever swallowed it.
        let src = "# 標題\n\na **bold** e\u{0301}mo👨‍👩‍👧ji `x` 你好\n\n\
                   - 項目 one\n- e\u{0301}dge\n\n> 引用 text\n\n\
                   | A | 值 |\n|---|---|\n| 你好 | 👩‍🚀 |\n";
        let m = map(src);
        for (r, row) in m.rows.iter().enumerate() {
            let openers: Vec<usize> = clusters(&row.glyphs).iter().map(|c| c.glyph).collect();
            for (i, g) in row.glyphs.iter().enumerate() {
                assert!(
                    !g.stop || openers.contains(&i),
                    "row {r}: the stop at glyph {i} ({:?}) is inside a cluster, \
                     so it is drawn at another glyph's column",
                    g.ch
                );
            }
        }
    }

    // ── the presentation vocabulary ─────────────────────────────────────────

    /// A block's own attributes, in the three formats that spell one on the
    /// block itself: HTML's tag, djot's `{…}` line, and — the odd one — a
    /// Markdown `<div>` around it, which is where twig has to put a Markdown
    /// block's attributes because the format has nowhere else.
    #[test]
    fn a_block_carries_its_alignment_on_every_row_it_draws() {
        // HTML, on the paragraph. `lead` is somebody else's class and is
        // neither read nor in the way.
        let html = map_leaf("<p class=\"lead center\">hi</p>\n", Format::Html);
        assert_eq!(line_facts(&html), vec![(Some(Align::Center), None)]);

        // djot's attribute line, on the block.
        let dj = map_leaf("{.right}\nhi\n", Format::Djot);
        assert_eq!(line_facts(&dj), vec![(Some(Align::Right), None)]);

        // A heading carries it too, and on every row a wrapped one draws.
        let h = map_leaf("{.center}\n# a heading\n", Format::Djot);
        assert_eq!(line_facts(&h), vec![(Some(Align::Center), None)]);
        assert_eq!(h.rows[0].heading, Some(1));

        // Both keys at once, and the line spacing is read the same way.
        let both = map_leaf("{.justify data-line-height=\"1.5\"}\nhi\n", Format::Djot);
        assert_eq!(
            line_facts(&both),
            vec![(
                Some(Align::Justify),
                Some(LineHeight::Step(LineSpacing::OneHalf))
            )]
        );

        // An unknown token is somebody else's and the block draws at the
        // theme's alignment; a ratio outside the menu's three is the author's
        // own and draws at exactly what they wrote.
        let other = map_leaf("{.lead data-line-height=\"1.3\"}\nhi\n", Format::Djot);
        assert_eq!(line_facts(&other), vec![(None, LineHeight::ratio(1.3))]);

        // A value the grammar does not cover is neither: carried by the
        // document, drawn at the theme's spacing, and read as nothing at all.
        let em = map_leaf("{data-line-height=\"1.3em\"}\nhi\n", Format::Djot);
        assert_eq!(line_facts(&em), vec![(None, None)]);
    }

    /// `<div class="center">` around three paragraphs centres all three, which
    /// is what the author of that HTML meant — and around one is the sole-child
    /// shape twig's `set_block_attrs` writes in Markdown.
    #[test]
    fn a_div_lends_its_alignment_to_every_block_inside_it() {
        let m = map_leaf(
            "<div class=\"center\" data-line-height=\"2\">\n\none\n\ntwo\n\n</div>\n",
            Format::Markdown,
        );
        assert_eq!(
            line_facts(&m),
            vec![
                (
                    Some(Align::Center),
                    Some(LineHeight::Step(LineSpacing::Double))
                ),
                (
                    Some(Align::Center),
                    Some(LineHeight::Step(LineSpacing::Double))
                ),
            ]
        );

        // The nearer node wins, and the block after the div is untouched — the
        // context is restored, not left running.
        let nested = map_leaf(
            "<div class=\"center\">\n\n<div class=\"right\">\n\ninner\n\n</div>\n\nouter\n\n</div>\n\nafter\n",
            Format::Markdown,
        );
        assert_eq!(
            line_facts(&nested),
            vec![
                (Some(Align::Right), None),
                (Some(Align::Center), None),
                (None, None),
            ]
        );
    }

    /// Size, face and colour are the run's, and the block's when the whole
    /// block is meant — read at both levels with the nearer winning.
    #[test]
    fn a_span_s_size_beats_its_block_s_and_its_face_falls_through() {
        // `<div data-font>` over `<p data-size>` over `<span data-size>`: the
        // span wins on size, the block is still what says the face.
        // The span is not first on its line: a `<span …>` opening one is an
        // HTML *block* to CommonMark, which is a fact about Markdown and not
        // about this.
        let m = map_leaf(
            "<div data-font=\"serif\">\n\nc <span data-size=\"small\">a</span> b\n\n</div>\n",
            Format::Markdown,
        );
        let a = style_of(&m, 'a');
        assert_eq!(a.size, Some(FontSize::Step(SizeStep::Small)));
        assert_eq!(a.font, Some(FaceRef::Generic(FontFamily::Serif)));
        // The text outside the span keeps the div's face and no size at all.
        let b = style_of(&m, 'b');
        assert_eq!(b.size, None);
        assert_eq!(b.font, Some(FaceRef::Generic(FontFamily::Serif)));

        // djot spells the same span anonymously and it reads identically.
        let dj = map_leaf(
            "{data-size=\"large\"}\nx [y]{data-size=\"xx-large\" data-color=\"blue\"} z\n",
            Format::Djot,
        );
        assert_eq!(
            style_of(&dj, 'x').size,
            Some(FontSize::Step(SizeStep::Large))
        );
        assert_eq!(
            style_of(&dj, 'y').size,
            Some(FontSize::Step(SizeStep::XxLarge))
        );
        assert_eq!(
            style_of(&dj, 'y').color,
            Some(TextColor::Named(MarkColor::Blue))
        );
        // The block's size is still the block's outside the span.
        assert_eq!(
            style_of(&dj, 'z').size,
            Some(FontSize::Step(SizeStep::Large))
        );
        assert_eq!(style_of(&dj, 'z').color, None);
    }

    /// The exact half of the vocabulary reaches a glyph and a row by the same
    /// doors the names do — the fold has one rule, not one per form. A named
    /// family is the one that cannot ride the glyph as itself: the walker
    /// interns it and the glyph carries the id.
    #[test]
    fn an_exact_size_face_and_colour_reach_the_glyph_and_the_row() {
        let m = map_leaf(
            "<div data-line-height=\"1.3\">\n\nc <span data-size=\"14pt\" \
             data-color=\"#c03030\" data-font=\"Garamond\">a</span> b\n\n</div>\n",
            Format::Markdown,
        );
        let a = style_of(&m, 'a');
        assert_eq!(a.size, FontSize::points(14.0));
        assert_eq!(
            a.color,
            Some(TextColor::Rgb {
                r: 0xc0,
                g: 0x30,
                b: 0x30
            })
        );
        assert_eq!(a.font, Some(FaceRef::Named(FaceId::of("Garamond"))));
        assert_eq!(m.face_name(FaceId::of("Garamond")), Some("Garamond"));
        // The div's ratio is the row's, on every row the block draws.
        assert_eq!(line_facts(&m), vec![(None, LineHeight::ratio(1.3))]);
        // And the text outside the span has none of the span's three.
        let b = style_of(&m, 'b');
        assert_eq!((b.size, b.font, b.color), (None, None, None));

        // One name, one entry, however many spans wear it — the table is what
        // keeps a `Style` `Copy` and it should not grow per run.
        let twice = map_leaf(
            "x <span data-font=\"Garamond\">a</span> y <span data-font=\"Garamond\">b</span>\n",
            Format::Markdown,
        );
        assert_eq!(twice.faces().len(), 1);
        assert_eq!(
            style_of(&twice, 'a').font,
            style_of(&twice, 'b').font,
            "one family, one id"
        );

        // A generic needs no entry at all: it names itself.
        let generic = map_leaf(
            "<div data-font=\"serif\">\n\nhi\n\n</div>\n",
            Format::Markdown,
        );
        assert_eq!(
            style_of(&generic, 'h').font,
            Some(FaceRef::Generic(FontFamily::Serif))
        );
        assert!(generic.faces().is_empty());
    }

    /// The one key two nodes share. `data-color` on a `mark` is the highlight's
    /// *background* and reaches a glyph through [`Role::Mark`]; the same key on
    /// an attributed span is the text's foreground. Same vocabulary, same enum,
    /// no collision — and a mark inside a coloured span wears both.
    #[test]
    fn a_mark_keeps_its_highlight_colour_and_a_span_colours_the_text() {
        let m = map_leaf("a ==\u{1f534} red== b\n", Format::Markdown);
        let r = style_of(&m, 'r');
        assert_eq!(r.role, Role::Mark(Some(MarkColor::Red)));
        assert_eq!(r.color, None, "a highlight is not a text colour");

        let both = map_leaf(
            "<span data-color=\"blue\">a ==\u{1f534} red== b</span>\n",
            Format::Markdown,
        );
        let r = style_of(&both, 'r');
        assert_eq!(r.role, Role::Mark(Some(MarkColor::Red)), "the highlight");
        assert_eq!(
            r.color,
            Some(TextColor::Named(MarkColor::Blue)),
            "the letters"
        );
    }

    /// A page break is the `::page-break` leaf directive, and djot spells the
    /// same document as an empty `::: page-break` fence whose name comes back
    /// as a class, HTML as a `<page-break>` element and AsciiDoc as `<<<`. All
    /// four draw the placeholder row every leaf directive gets and carry the
    /// same [`DirectiveMark`], because a frontend that opens a page at one
    /// must not be able to tell which format the file is in.
    #[test]
    fn a_page_break_reads_the_same_in_every_format() {
        for (fmt, src) in [
            (Format::Markdown, "a\n\n::page-break\n\nb\n"),
            (Format::Djot, "a\n\n::: page-break\n:::\n\nb\n"),
            (
                Format::Html,
                "<p>a</p>\n\n<page-break></page-break>\n\n<p>b</p>\n",
            ),
            (Format::Asciidoc, "a\n\n<<<\n\nb\n"),
        ] {
            let m = map_leaf(src, fmt);
            let marks: Vec<&DirectiveMark> = m
                .rows
                .iter()
                .filter_map(|r| r.leaf_directive.as_ref())
                .collect();
            assert_eq!(marks.len(), 1, "{fmt:?} draws one placeholder");
            assert_eq!(marks[0].name, "page-break", "{fmt:?}");
            assert!(marks[0].attrs.is_empty(), "{fmt:?}: {:?}", marks[0].attrs);
            assert!(
                m.rows
                    .iter()
                    .any(|r| r.glyphs.iter().map(|g| g.ch).collect::<String>()
                        == "\u{29c9} page-break"),
                "{fmt:?} draws the label, got {:?}",
                m.rows
                    .iter()
                    .map(|r| r.glyphs.iter().map(|g| g.ch).collect::<String>())
                    .collect::<Vec<_>>()
            );
        }

        // A Markdown `:::note` with nothing in it is *not* this: its name is
        // its own, and nothing about it says "a block with no body" the way
        // djot's spelling of a leaf directive does.
        let empty_fence = map_leaf("::: note\n:::\n", Format::Markdown);
        assert!(
            empty_fence.rows.iter().all(|r| r.leaf_directive.is_none()),
            "a named empty fence keeps the reading it has"
        );
    }

    /// A djot fence carrying more than its name keeps the rest as an attribute
    /// rather than folding it into the name: with a bare `:::` fence, which
    /// names nothing itself, the *first* class token is the name.
    #[test]
    fn a_djot_fence_s_first_class_is_the_directive_s_name_and_the_rest_is_attributes() {
        let m = map_leaf("{.page-break .wide}\n:::\n:::\n", Format::Djot);
        let mark = m
            .rows
            .iter()
            .find_map(|r| r.leaf_directive.as_ref())
            .expect("a placeholder");
        assert_eq!(mark.name, "page-break");
        assert_eq!(
            mark.attrs,
            vec![("class".to_string(), Some("wide".to_string()))]
        );
    }

    /// With the name on the fence line, djot appends it *after* the classes of
    /// the attribute line above, so the name is the fence's word wherever it
    /// lands in the class — here, and inside a quote.
    #[test]
    fn a_djot_fence_s_word_is_the_directive_s_name_behind_an_attribute_line() {
        for src in [
            "{.wide src=\"u\"}\n::: x-card\n:::\n",
            "> {.wide src=\"u\"}\n> ::: x-card\n> :::\n",
        ] {
            let m = map_leaf(src, Format::Djot);
            let mark = m
                .rows
                .iter()
                .find_map(|r| r.leaf_directive.as_ref())
                .unwrap_or_else(|| panic!("a placeholder for {src:?}"));
            assert_eq!(mark.name, "x-card", "{src:?}");
            assert_eq!(
                mark.attrs,
                vec![
                    ("class".to_string(), Some("wide".to_string())),
                    ("src".to_string(), Some("u".to_string())),
                ],
                "{src:?}"
            );
        }
    }

    // ── math ─────────────────────────────────────────────────────────────────

    /// [`map_leaf`] on a chosen surface and reveal — the whole of what a math
    /// rendering turns on.
    fn map_math(src: &str, format: Format, surface: &Surface, reveal: Option<Reveal>) -> VisualMap {
        let mut ed =
            Editor::new_ext(src.as_bytes(), format, crate::doc::parse_extensions()).unwrap();
        build(&ed.nodes().unwrap(), src, Some(80), false, surface, reveal)
    }

    fn pictures() -> Surface {
        Surface {
            inline_pictures: true,
            ..Default::default()
        }
    }

    #[test]
    fn markdown_reads_math_and_a_dollar_before_whitespace_stays_prose() {
        // The `math` extension is on for every leaf document; twig's own rule
        // keeps a price out of it.
        let m = map_leaf("Say $E = mc^2$ for $5 and $6.\n", Format::Markdown);
        assert_eq!(row_texts(&m), vec!["Say E = mc^2 for $5 and $6."]);
        let e = m.rows[0].glyphs.iter().find(|g| g.ch == 'E').unwrap();
        assert_eq!(
            e.style.role,
            Role::Code,
            "the formula's TeX, in the code style"
        );
        assert_eq!(e.src, 5, "at its own byte, past the `$`");
        let five = m.rows[0].glyphs.iter().find(|g| g.ch == '5').unwrap();
        assert_eq!(five.style.role, Role::Body);
        assert!(m.math.is_empty(), "no picture stands in on a plain surface");
    }

    #[test]
    fn a_plain_surface_draws_inline_math_as_code_with_the_delimiters_hidden() {
        // The verbatim treatment, in both languages — what `inline_math` always
        // rendered as — with the caret's home at the content's end.
        for (src, fmt) in [
            ("a $x+y$ b\n", Format::Markdown),
            ("a $`x+y` b\n", Format::Djot),
        ] {
            let m = map_leaf(src, fmt);
            assert_eq!(row_texts(&m), vec!["a x+y b"], "{src:?}");
            let x = m.rows[0].glyphs.iter().find(|g| g.ch == 'x').unwrap();
            assert_eq!(x.style.role, Role::Code);
            assert!(!m.mark_ends.is_empty(), "the content end is a caret home");
        }
    }

    #[test]
    fn display_math_in_a_line_of_prose_puts_its_text_at_the_text_s_own_offset() {
        // `display_math` had no arm and fell to the default one, which pushed
        // the text at the *node's* start: three bytes short in djot, past `$$`
        // and the backtick.
        let m = map_leaf("Before $$`x`$$ after\n", Format::Djot);
        let x = m.rows[0].glyphs.iter().find(|g| g.ch == 'x').unwrap();
        assert_eq!(x.src, "Before $$`".len());
        assert_eq!(x.style.role, Role::Code);
        let m = map_leaf("Before $$x$$ after\n", Format::Markdown);
        let x = m.rows[0].glyphs.iter().find(|g| g.ch == 'x').unwrap();
        assert_eq!(x.src, "Before $$".len());
    }

    #[test]
    fn a_surface_that_paints_in_a_line_gets_one_atom_per_inline_formula() {
        let src = "Say $E = mc^2$ and $a$.\n";
        let m = map_math(src, Format::Markdown, &pictures(), None);
        assert_eq!(row_texts(&m), vec!["Say ∑ and ∑."]);
        assert_eq!(m.math.len(), 2);
        let first = &m.math[0];
        assert_eq!(first.tex, "E = mc^2");
        assert!(!first.display);
        assert_eq!(first.row, 0);
        assert_eq!(first.rows_span, 0..1);
        assert_eq!(first.glyph, Some(4));
        assert_eq!(first.src, 4, "the formula's start, where a click lands");
        let atom = &m.rows[0].glyphs[4];
        assert_eq!(atom.ch, MATH_ATOM);
        assert_eq!(atom.style.role, Role::Math);
        assert!(atom.stop);
        assert_eq!(atom.src, 4);
        // The caret has a stop on the atom and the next glyph's past it, and
        // nothing inside the markup.
        assert_eq!(m.stop_after(4), Some("Say $E = mc^2$".len()));
        assert!(m.mark_ends.is_empty(), "no home inside the hidden markup");
        // The row carries the marks the side-table is derived from.
        assert_eq!(m.rows[0].math.len(), 2);
        assert_eq!(m.rows[0].math[1].glyph, Some(m.math[1].glyph.unwrap()));
        assert_eq!(m.math[1].tex, "a");
    }

    #[test]
    fn a_display_formula_written_inline_is_an_atom_in_display_style() {
        let m = map_math("Before $$x$$ after\n", Format::Markdown, &pictures(), None);
        assert_eq!(row_texts(&m), vec!["Before ∑ after"]);
        assert_eq!(m.math.len(), 1);
        assert!(m.math[0].display);
        assert_eq!(m.math[0].tex, "x");
    }

    #[test]
    fn an_atom_follows_its_glyph_across_a_wrap() {
        // Fifteen words, then a formula that wraps onto the second row: the
        // mark is drained onto the row the glyph landed on, at its index there.
        let src = format!("{}$x$ end\n", "word ".repeat(15));
        let mut ed = Editor::new_ext(
            src.as_bytes(),
            Format::Markdown,
            crate::doc::parse_extensions(),
        )
        .unwrap();
        let m = build(
            &ed.nodes().unwrap(),
            &src,
            Some(40),
            false,
            &pictures(),
            None,
        );
        assert!(m.rows.len() >= 2);
        assert_eq!(m.math.len(), 1);
        let info = &m.math[0];
        let g = &m.rows[info.row].glyphs[info.glyph.unwrap()];
        assert_eq!(g.ch, MATH_ATOM);
        assert_eq!(g.src, src.find("$x$").unwrap());
        assert!(m.rows[..info.row].iter().all(|r| r.math.is_empty()));
    }

    #[test]
    fn an_atom_in_a_table_cell_rides_the_cell_s_row() {
        let m = map_math(
            "| a | b |\n|---|---|\n| $x$ | c |\n",
            Format::Markdown,
            &pictures(),
            None,
        );
        assert_eq!(m.math.len(), 1);
        let info = &m.math[0];
        assert_eq!(m.rows[info.row].glyphs[info.glyph.unwrap()].ch, MATH_ATOM);
    }

    #[test]
    fn a_display_formula_on_its_own_lines_is_a_block_placeholder_on_every_surface() {
        for (src, fmt) in [
            (
                "intro\n\n$$\n\\int_0^1 x\\,dx\n$$\n\nend\n",
                Format::Markdown,
            ),
            ("intro\n\n$$`\\int_0^1 x\\,dx`\n\nend\n", Format::Djot),
        ] {
            for surface in [Surface::default(), pictures()] {
                let m = map_math(src, fmt, &surface, None);
                assert_eq!(
                    row_texts(&m),
                    vec!["intro", "", "∑ \\int_0^1 x\\,dx", "", "end"],
                    "{src:?}"
                );
                assert_eq!(m.math.len(), 1);
                let info = &m.math[0];
                assert!(info.display);
                assert_eq!(info.glyph, None, "a block, not an atom");
                assert_eq!(info.rows_span, 2..3);
                assert_eq!(info.tex.trim(), "\\int_0^1 x\\,dx");
                let start = src.find("$$").unwrap();
                let end = start + src[start..].find("\n\nend").unwrap();
                assert_eq!(info.src, start);
                // Every label glyph at the formula's start, a stop there and
                // one past the block, nothing inside — a picture's two homes.
                let row = &m.rows[2];
                assert!(
                    row.glyphs
                        .iter()
                        .all(|g| g.src == start && g.style.role == Role::Math)
                );
                assert_eq!(row.end_src, end);
                assert_eq!(m.stop_after(start), Some(end));
                assert_eq!(m.stop_before(end), Some(start));
                // The gaps either side are labelled as a formula's.
                assert_eq!(m.rows[1].boundary.map(|b| b.below), Some(BlockClass::Math));
                assert_eq!(m.rows[3].boundary.map(|b| b.above), Some(BlockClass::Math));
            }
        }
    }

    #[test]
    fn a_display_block_reserves_the_rows_the_frontend_measured() {
        let src = "$$\nx\n$$\n\nend\n";
        let surface = Surface {
            math_rows: HashMap::from([("\nx\n".to_string(), 4)]),
            ..Default::default()
        };
        let m = map_math(src, Format::Markdown, &surface, None);
        assert_eq!(m.math[0].rows_span, 0..4);
        assert_eq!(row_texts(&m)[..5], ["∑ x", "", "", "", ""]);
        // The fillers are decoration: drawn, no caret, anchored past the block.
        for r in &m.rows[1..4] {
            assert!(r.decoration);
            assert_eq!(r.end_src, 7);
        }
        assert_eq!(m.stop_after(0), Some(7));
        // A height keyed by TeX that does not match reserves nothing.
        let surface = Surface {
            math_rows: HashMap::from([("x".to_string(), 4)]),
            ..Default::default()
        };
        let m = map_math(src, Format::Markdown, &surface, None);
        assert_eq!(m.math[0].rows_span, 0..1);
    }

    #[test]
    fn a_paragraph_with_prose_beside_a_display_formula_is_not_a_block() {
        let m = map_math("see\n$$\nx\n$$\n", Format::Markdown, &pictures(), None);
        assert!(
            m.math.iter().all(|i| i.glyph.is_some()),
            "an atom, not a placeholder"
        );
        let m = map_math(
            "$$\nx\n$$\n$$\ny\n$$\n",
            Format::Markdown,
            &pictures(),
            None,
        );
        assert_eq!(m.math.len(), 2);
        assert!(
            m.math.iter().all(|i| i.glyph.is_some()),
            "two formulas is prose with math in it"
        );
    }

    #[test]
    fn a_formula_on_the_reveal_line_is_its_tex_in_every_mode() {
        let src = "Say $E = mc^2$ here.\n";
        // The hidden modes' reveal: only the formula shows its markup.
        let m = map_math(
            src,
            Format::Markdown,
            &pictures(),
            Some(Reveal::math(0..src.len() - 1)),
        );
        assert_eq!(row_texts(&m), vec!["Say $E = mc^2$ here."]);
        assert!(
            m.math.is_empty(),
            "nothing stands in for a revealed formula"
        );
        let dollar = m.rows[0].glyphs.iter().find(|g| g.ch == '$').unwrap();
        assert_eq!(dollar.style.role, Role::Delimiter);
        let e = m.rows[0].glyphs.iter().find(|g| g.ch == 'E').unwrap();
        assert_eq!(e.style.role, Role::Code);
        // Full's reveal draws the same, and an emphasis beside it reveals too
        // where the hidden modes' does not.
        let src = "*a* $x$\n";
        let hidden = map_math(
            src,
            Format::Markdown,
            &pictures(),
            Some(Reveal::math(0..src.len() - 1)),
        );
        assert_eq!(row_texts(&hidden), vec!["a $x$"]);
        let full = map_math(
            src,
            Format::Markdown,
            &pictures(),
            Some(Reveal::full(0..src.len() - 1)),
        );
        assert_eq!(row_texts(&full), vec!["*a* $x$"]);
        // A reveal line that does not meet the formula leaves the atom.
        let other = map_math(
            "$x$\n\ntext\n",
            Format::Markdown,
            &pictures(),
            Some(Reveal::math(5..9)),
        );
        assert_eq!(row_texts(&other)[0], "∑");
    }

    #[test]
    fn a_display_block_on_the_reveal_line_is_its_source_lines_in_the_code_style() {
        let src = "intro\n\n$$\n\\int_0^1 x\n$$\n\nend\n";
        // Any line of the block reveals the whole of it: here the middle one.
        let mid = src.find("\\int").unwrap();
        let line = mid..mid + "\\int_0^1 x".len();
        let m = map_math(src, Format::Markdown, &pictures(), Some(Reveal::math(line)));
        assert_eq!(
            row_texts(&m),
            vec!["intro", "", "$$", "\\int_0^1 x", "$$", "", "end"]
        );
        assert!(m.math.is_empty());
        let fence = m.rows[2].glyphs.iter().find(|g| g.ch == '$').unwrap();
        assert_eq!(fence.style.role, Role::Delimiter);
        assert_eq!(fence.src, 7);
        let x = m.rows[3].glyphs.iter().find(|g| g.ch == 'x').unwrap();
        assert_eq!(x.style.role, Role::Code);
        assert_eq!(x.src, src.find("x\n$$").unwrap());
        // Every byte of the source is reachable: a stop on each fence and each
        // character between.
        assert!(m.is_stop(7));
        assert!(m.is_stop(mid));
        // The closing fence's line too, and the same for djot.
        let close = src.rfind("$$").unwrap();
        let m = map_math(
            src,
            Format::Markdown,
            &pictures(),
            Some(Reveal::math(close..close + 2)),
        );
        assert_eq!(row_texts(&m)[2], "$$");
        let src = "$$`\n\\int\n`\n";
        let m = map_math(src, Format::Djot, &pictures(), Some(Reveal::math(4..8)));
        assert_eq!(row_texts(&m), vec!["$$`", "\\int", "`"]);
    }

    #[test]
    fn a_block_that_holds_a_formula_says_so_to_the_cache() {
        let src = "plain\n\nwith $x$ in it\n\n$$\ny\n$$\n";
        let mut ed = Editor::new_ext(
            src.as_bytes(),
            Format::Markdown,
            crate::doc::parse_extensions(),
        )
        .unwrap();
        let mut cache = BlockCache::default();
        let top = top_blocks(&mut ed);
        let _ = build_cached(
            &top,
            src,
            None,
            false,
            &pictures(),
            None,
            &mut cache,
            |id| ed.subtree(NodeId(id)).unwrap_or_default(),
        );
        assert!(!cache.math_meets(&(0..5)), "the plain paragraph");
        assert!(cache.math_meets(&(7..21)), "the one with an atom");
        assert!(
            cache.math_meets(&(26..27)),
            "the middle line of the display block"
        );
        // A hit carries the answer without a walk: build again from the cache.
        let _ = build_cached(
            &top,
            src,
            None,
            false,
            &pictures(),
            None,
            &mut cache,
            |_| Vec::new(),
        );
        assert!(cache.math_meets(&(7..21)));
        assert!(!cache.math_meets(&(0..5)));
    }

    #[test]
    fn incremental_builds_agree_with_the_reference_on_math() {
        for src in [
            "a $x$ b\n\n$$\ny\n$$\n\nc\n",
            "one\n\ntwo $\\frac{a}{b}$ three\n",
            "$$\n\\int\n$$\n",
        ] {
            for surface in [Surface::default(), pictures()] {
                let mut ed = Editor::new_ext(
                    src.as_bytes(),
                    Format::Markdown,
                    crate::doc::parse_extensions(),
                )
                .unwrap();
                let all = ed.nodes().unwrap();
                let plain = build(&all, src, Some(80), false, &surface, None);
                let mut cache = BlockCache::default();
                let top = top_blocks(&mut ed);
                let cached = build_cached(
                    &top,
                    src,
                    Some(80),
                    false,
                    &surface,
                    None,
                    &mut cache,
                    |id| ed.subtree(NodeId(id)).unwrap_or_default(),
                );
                assert_eq!(row_texts(&plain), row_texts(&cached), "{src:?}");
                assert_eq!(plain.math, cached.math, "{src:?}");
                assert_eq!(plain.stops, cached.stops, "{src:?}");
            }
        }
    }
}