rustyfi_backend/

hbox.rs

use crate::font::FontKey;
use crate::graphics::{Color, GraphicsElem};
use crate::length::Length;
use crate::math::MathGlyph;
use crate::tabular::TabularBox;
use crate::vbox::VertBox;

/// Which font/size a string box is set in (`horz_string_info`).
#[derive(Clone, Debug, PartialEq)]
pub struct HorzStringInfo {
    pub font: FontKey,
    pub size: Length,
    /// A manual baseline raise (`ScriptFont::rising` scaled by the
    /// run's font size — `fontInfo.ml`'s `get_font_with_ratio`). Both PDF
    /// writers add it to the placed `ty` before `Tj`.
    pub rising: Length,
    /// `set-text-color`'s value at the time this run/glyph was built
    /// (`Context::text_color`). `Color::Gray(0.0)` (black) is the default and
    /// both writers emit NO color op for a black run.
    pub color: Color,
}

/// An index into `DocumentValue::images` (rustyfi-lang). Boxes carry this
/// rather than the decoded bytes, so cloning a box (routine during line
/// breaking) never copies image data.
#[derive(Clone, Copy, Debug, PartialEq, Eq, Hash)]
pub struct ImageId(pub usize);

/// An opaque index into a lang-side table of deferred page-break-hook
/// closures (`Interp::hooks`). `break_pages` places the box this token lives
/// in like any other content and never learns what the hook computes; a
/// lang-side post-pass (`fire_hooks`) reads the token back once geometry is
/// final. `DecoId`/`GraphicsFnId` below follow the same pattern.
#[derive(Clone, Copy, Debug, PartialEq, Eq, Hash)]
pub struct HookId(pub usize);

/// Deferred *decoration* closures (`Interp::decos`) for frames.
/// `fire_hooks` fires one with the frame's placed `(x, y, w, h, d)` and
/// accumulates the returned graphics onto the page.
#[derive(Clone, Copy, Debug, PartialEq, Eq, Hash)]
pub struct DecoId(pub usize);

/// Deferred `inline-graphics-outer` callbacks (`Interp::outer_graphics`),
/// read back by `resolve_outer_graphics_*` once line layout has resolved the
/// box's width.
#[derive(Clone, Copy, Debug, PartialEq, Eq, Hash)]
pub struct GraphicsFnId(pub usize);

/// A decoded raster image, referenced via its `ImageId`. Mirrors v0.0.6's
/// `ImageInfo` (`imageInfo.ml`).
///
/// Every source format is flattened to 8-bit `DeviceRGB` and any alpha
/// channel is dropped; transparency (`/SMask`) is unimplemented. `jpeg_dct`
/// additionally carries the source's original, still-DCT-encoded bytes for a
/// baseline JPEG, so `write_image_xobjects` can embed those directly instead
/// of re-encoding the flattened RGB8 samples.
#[derive(Clone, Debug, PartialEq)]
pub struct ImageResource {
    /// Row-major, top-to-bottom, 3-bytes-per-pixel RGB8 samples with no
    /// padding — exactly what `pdf_writer::Chunk::image_xobject` wants for a
    /// `DeviceRGB` image at `bits_per_component(8)`.
    pub samples: Vec<u8>,
    pub px_w: u32,
    pub px_h: u32,
    /// `Some` for a baseline (or extended-sequential) 8-bit JPEG with 1 or 3
    /// color components — see `sniff_baseline_jpeg_dct` for exactly which
    /// qualify. `None` for every non-JPEG source, and for a JPEG this port
    /// cannot map to a PDF colorspace without guessing (progressive, 12-bit,
    /// or 4-component CMYK/YCCK).
    pub jpeg_dct: Option<JpegDct>,
    /// `Some` when this resource is an imported page of an external PDF
    /// (`load-pdf-image`) rather than a decoded raster image;
    /// `samples`/`px_w`/`px_h` are then left empty, and PDF-page consumers
    /// branch on this field instead.
    pub pdf: Option<PdfPageResource>,
}

/// An embedded page of an external PDF (`load-pdf-image`), carrying just
/// enough of the source page's object graph to re-emit it as a PDF **Form
/// XObject**. Deliberately `lopdf`-free plain data so `rustyfi-backend` need
/// not depend on `lopdf`.
#[derive(Clone, Debug, PartialEq)]
pub struct PdfPageResource {
    /// The source page's `/MediaBox`, `(x0, y0, x1, y1)` in raw PDF points
    /// (upstream `loadPdf.ml`: MediaBox only, no `/CropBox` fallback).
    pub media_box: (f64, f64, f64, f64),
    /// The page's content stream(s), already inflated (`/FlateDecode`
    /// resolved) and concatenated (with a separating space per PDF rules
    /// when a page has more than one content stream) — ready to wrap
    /// verbatim in a Form XObject's stream body.
    pub content: Vec<u8>,
    /// The imported object subtree reachable from the page's `/Resources`,
    /// self-contained and keyed by source object number so the writer can
    /// remap references to freshly allocated output `Ref`s. Local id `0` is
    /// reserved (PDF object number 0 is always free/unused in a well-formed
    /// file) and holds the page's own `/Resources` dictionary itself
    /// (whether it was a direct or an indirect object in the source); every
    /// other entry is a real source object number.
    pub resources: ImportedObjects,
}

/// A serialized subtree of a *foreign* PDF's object graph. See
/// `PdfPageResource::resources` for the local-id convention.
#[derive(Clone, Debug, Default, PartialEq)]
pub struct ImportedObjects(pub Vec<(u32, ObjRepr)>);

/// A minimal sum type mirroring the PDF object grammar, just enough to
/// re-emit an imported object verbatim. `Ref(u32)` refers to another entry's
/// local id in the same `ImportedObjects` table; the writer treats an
/// unresolved `Ref` as an importer bug rather than fetching it.
#[derive(Clone, Debug, PartialEq)]
pub enum ObjRepr {
    Null,
    Bool(bool),
    Int(i64),
    Real(f64),
    /// A PDF name's raw bytes, without the leading `/` and with `#xx`
    /// escapes already decoded (mirrors `lopdf::Object::Name`).
    Name(Vec<u8>),
    /// A PDF string's raw bytes (mirrors `lopdf::Object::String`, literal or
    /// hex — both collapse to bytes here since we only ever re-emit them).
    String(Vec<u8>),
    Ref(u32),
    Array(Vec<ObjRepr>),
    Dict(Vec<(Vec<u8>, ObjRepr)>),
    /// A stream object: its dictionary entries (excluding `/Length`, which
    /// the writer derives) plus already-decompressed content bytes. The
    /// writer decides filtering/compression at write time; the imported
    /// content is kept in cleartext form here so no `lopdf`-specific codec
    /// state needs to cross the boundary.
    Stream(Vec<(Vec<u8>, ObjRepr)>, Vec<u8>),
}

/// The original, still-DCT-encoded bytes of a source JPEG, plus enough
/// metadata (`components`) for the writer to pick the matching
/// `/ColorSpace`. Never re-derived from the flattened `samples`, which have
/// already lost the JPEG's own subsampling/quantization.
#[derive(Clone, Debug, PartialEq)]
pub struct JpegDct {
    /// The complete original file contents, `FFD8` (SOI) to `FFD9` (EOI),
    /// byte-for-byte — exactly the stream a `/Filter /DCTDecode` XObject wants.
    pub bytes: Vec<u8>,
    /// Color components from the JPEG's SOF marker: `1` (grayscale ->
    /// `/DeviceGray`) or `3` (YCbCr/RGB -> `/DeviceRGB`).
    /// `sniff_baseline_jpeg_dct` never returns any other value.
    pub components: u8,
}

impl ImageResource {
    /// Scan raw file bytes for a JPEG **SOF0** (baseline DCT) or **SOF1**
    /// (extended sequential DCT) marker — the two variants a `/DCTDecode`
    /// XObject can safely wrap verbatim, matching upstream's own JPEG
    /// special-case (`imageInfo.ml`'s bypass of decode/re-encode for `Jpeg`).
    /// Returns `None`, meaning "fall back to the flattened RGB8 embedding",
    /// for:
    ///
    /// - anything that isn't a JPEG (no `FFD8` SOI marker);
    /// - a malformed/truncated JPEG (a segment length running past the end of
    ///   the buffer, or scan data reached before any SOF marker);
    /// - progressive, lossless, arithmetic-coded, or hierarchical JPEGs (any
    ///   SOF other than `0xC0`/`0xC1`) — viewer `DCTDecode` support for these
    ///   is inconsistent, so only the two universally-supported variants are
    ///   trusted;
    /// - non-8-bit sample precision;
    /// - anything other than 1 or 3 color components — in particular
    ///   4-component CMYK/YCCK, whose correct embedding depends on an Adobe
    ///   APP14 transform flag (some store inverted samples) this port does not
    ///   interpret.
    ///
    /// `bytes` is consumed so the `Some` case hands the file contents into
    /// `JpegDct` with no copy.
    pub fn sniff_baseline_jpeg_dct(bytes: Vec<u8>) -> Option<JpegDct> {
        if bytes.len() < 4 || bytes[0] != 0xFF || bytes[1] != 0xD8 {
            return None; // no SOI marker: not a JPEG.
        }
        let mut i = 2usize;
        while i < bytes.len() {
            if bytes[i] != 0xFF {
                return None; // not aligned on a marker; bail rather than guess.
            }
            // Marker codes may be preceded by any number of 0xFF fill bytes.
            let mut j = i + 1;
            while j < bytes.len() && bytes[j] == 0xFF {
                j += 1;
            }
            if j >= bytes.len() {
                return None;
            }
            let marker = bytes[j];
            i = j + 1;
            // Standalone markers carry no length-prefixed segment: SOI
            // (stray, shouldn't recur but is harmless), EOI, TEM, RSTn.
            if marker == 0xD8 || marker == 0xD9 || marker == 0x01 || (0xD0..=0xD7).contains(&marker)
            {
                continue;
            }
            if i + 2 > bytes.len() {
                return None;
            }
            let seg_len = u16::from_be_bytes([bytes[i], bytes[i + 1]]) as usize;
            if seg_len < 2 || i + seg_len > bytes.len() {
                return None;
            }
            if marker == 0xDA {
                return None; // start-of-scan reached before any SOF: bail.
            }
            // SOF0..SOF15 (0xC0..0xCF) except 0xC4 (DHT), 0xC8 (JPG
            // extension, reserved), 0xCC (DAC) — those codes overlap the
            // SOF range but aren't frame headers.
            let is_sof = (0xC0..=0xCF).contains(&marker) && ![0xC4, 0xC8, 0xCC].contains(&marker);
            if is_sof {
                if marker != 0xC0 && marker != 0xC1 {
                    return None; // progressive/lossless/hierarchical: bail.
                }
                // Segment payload after the 2-byte length: precision(1),
                // height(2 BE), width(2 BE), num_components(1).
                let payload = &bytes[i + 2..i + seg_len];
                if payload.len() < 6 {
                    return None;
                }
                let precision = payload[0];
                let components = payload[5];
                return if precision == 8 && (components == 1 || components == 3) {
                    Some(JpegDct { bytes, components })
                } else {
                    None
                };
            }
            i += seg_len;
        }
        None
    }

    /// Intrinsic dimensions for aspect-ratio math (`use-image-by-width`):
    /// pixel extents for a raster resource, MediaBox point extents for an
    /// imported PDF page. Both ratios are dimensionless (px/px or pt/pt), so
    /// callers apply the same `height = width * ih/iw` regardless of kind —
    /// only the placement CTM needs to know which units these are.
    pub fn intrinsic_dims_pt(&self) -> (f64, f64) {
        if let Some(pdf) = &self.pdf {
            let (x0, y0, x1, y1) = pdf.media_box;
            (x1 - x0, y1 - y0)
        } else {
            (self.px_w as f64, self.px_h as f64)
        }
    }
}

/// A milestone-1 subset of `pure_horz_box` from horzBox.ml, keeping its
/// vocabulary so the full port extends rather than replaces it.
///
/// The `#[subast]` list names every *other* box type reachable from a field
/// of this enum; self-recursion (`Discretionary`, `Frame`) is implicit. It is
/// what [`crate::visit`]'s generated traversal descends through, and it is
/// **unchecked on stable Rust** — see that module's "The one trap" note and
/// the `tests/visit_reachability.rs` test standing in for the missing check.
#[derive(Clone, Debug, PartialEq, syan::visit::Ast)]
#[subast(
    crate::hbox::PureHorzBox,
    crate::vbox::VertBox,
    crate::graphics::GraphicsElem,
    crate::tabular::TabularBox
)]
pub enum PureHorzBox {
    /// Fixed text with pre-measured dimensions.
    InnerString {
        info: HorzStringInfo,
        text: String,
        width: Length,
        height: Length,
        depth: Length,
    },
    /// Interword glue.
    OuterEmpty {
        natural: Length,
        shrinkable: Length,
        stretchable: Length,
    },
    /// Infinitely stretchable glue (`inline-fil`).
    OuterFil,
    /// A fixed-width empty box with no stretch/shrink (`inline-skip`;
    /// v0.0.6: `PHSFixedEmpty`). Unlike `OuterEmpty` this is never a legal
    /// line-break point (see `is_glue`).
    FixedEmpty { width: Length },
    /// A raster image placed at a fixed on-page size (`use-image-by-width`).
    /// `width`/`height` are already-computed on-page dimensions (v0.0.6
    /// `ImageInfo.get_height_from_width`). Like `FixedEmpty`, never a legal
    /// line-break point (`is_glue`).
    Image {
        width: Length,
        height: Length,
        image: ImageId,
    },
    /// A break point that may or may not be taken (v0.0.6's
    /// `LBDiscretionary(penalty, id, pre, post_nobreak, post_break)`,
    /// `ref:src/backend/lineBreakBox.ml:22-27`). If the paragraph breaker
    /// chooses to break here, `pre_break` renders at the end of the closed
    /// line and `post_break` at the start of the next; otherwise `no_break`
    /// renders in its place. UAX#14 only needs zero-width inter-chunk
    /// break points with all three slots empty. Unlike `OuterEmpty`/`OuterFil`
    /// this is not "glue" (see `is_glue`) — it is scored separately via
    /// `is_break_point`/`break_penalty`.
    Discretionary {
        penalty: i32,
        pre_break: Vec<PureHorzBox>,
        post_break: Vec<PureHorzBox>,
        no_break: Vec<PureHorzBox>,
    },
    /// A box carrying resolved `graphics` elements (`inline-graphics`;
    /// v0.0.6: `PHGFixedGraphics`), coordinates already relative to the
    /// box's baseline-left origin. Carries a real depth (graphics can extend
    /// below the baseline), so both `height` and `depth` feed line metrics.
    /// Never a legal line-break point (see `is_glue`).
    Graphics {
        width: Length,
        height: Length,
        depth: Length,
        elems: Vec<GraphicsElem>,
        /// True when the callback ignored its placed-point argument, so its
        /// `elems` are PAGE-ABSOLUTE (e.g. a slydifi frame background /
        /// full-page decoration built with `fun _ -> …`). Such graphics must
        /// be emitted with an IDENTITY `cm` — NOT translated by the box's
        /// placed position — otherwise the whole decoration shifts off the
        /// page (the box is often placed at a negative text-origin). For an
        /// ordinary position-relative callback this is false and the writer's
        /// per-box `cm` translate applies as usual.
        origin_independent: bool,
    },
    /// `inline-graphics-outer` (v0.0.6 `PHGOuterFilGraphics`,
    /// vminst.ml:1891): a graphics box whose WIDTH stretches like
    /// `inline-fil` (upstream widinfo `{natural = 0; stretchable = Fils(1)}`,
    /// lineBreak.ml:40-48). `width` starts at ZERO and is written by
    /// `justify_line` with the box's per-fil slack share; the box is then
    /// replaced by a resolved `Graphics` in a lang-side post-pass that fires
    /// `fn_id`'s callback with that width. NOT glue (upstream's box is pure
    /// content, never a break point), but counted as a fil by
    /// `measure`/`justify_line`.
    GraphicsOuter {
        height: Length,
        depth: Length,
        width: Length,
        fn_id: GraphicsFnId,
    },
    /// A laid-out inline math run (`${…}`): one box carrying its own
    /// pre-shifted sub-glyphs, each with a vertical offset relative to this
    /// box's baseline (`MathGlyph::dy`) — the line model has only a
    /// horizontal `dx` per box and a single `baseline_y` per line, so a
    /// superscript can't be a separate box. `width`/`height`/`depth` are the
    /// run's outer metrics (computed by `read_math`), so the line breaker
    /// never re-enters the math engine. Never a legal line-break point (see
    /// `is_glue`) — a math run is laid out and flowed atomically.
    ///
    /// `rules`: filled paths the run needs alongside its glyphs — the
    /// fraction bar and radical sign/overbar are `Fill`s, not glyphs, since
    /// neither is drawable through a font's `Tj`. Box-local, y-**up**
    /// coordinates relative to this box's own baseline-left origin, exactly
    /// `PureHorzBox::Graphics::elems`' convention. `natural_width` is
    /// unaffected — a bar/radical-sign always sits within `glyphs`'
    /// already-measured span. Empty for the `read_math` path and for every
    /// atom `layout_math_atom` doesn't specially handle.
    Math {
        width: Length,
        height: Length,
        depth: Length,
        glyphs: Vec<MathGlyph>,
        rules: Vec<GraphicsElem>,
    },
    /// A deferred page-break hook (`hook-page-break`; v0.0.6's
    /// `PHGHookPageBreak`). Zero-width, renders nothing.
    HookPageBreak { id: HookId },
    /// A ruled grid box (`tabular`; v0.0.6's `PHGFixedTabular`), carrying
    /// each cell's already-laid-out inline run (`tabular::TabularCellBox`)
    /// plus the resolved rule graphics from the user's callback. The PDF
    /// writers recurse into it in `emit_box`, which is also where its three
    /// coordinate frames are reconciled.
    Tabular(TabularBox),
    /// An inline box carrying a whole block (`embed-block-top`; upstream's
    /// `PHGEmbeddedVert`/`HorzEmbeddedVertBreakable`). `block` is already
    /// broken into `VertBox` lines, which the writer stacks from the box's
    /// placed origin. ATOMIC — it does not split across a page boundary.
    EmbeddedBlock {
        width: Length,
        height: Length,
        depth: Length,
        block: Vec<VertBox>,
        /// Which of the block's lines sits on the surrounding text baseline:
        /// `false` = the FIRST line (`embed-block-top`, `adjust_to_first_line`),
        /// `true` = the LAST line (`embed-block-bottom`, `adjust_to_last_line`).
        /// Governs both this box's height/depth split and where the writers
        /// anchor the block's inner lines (`place_embedded_block`).
        anchor_last: bool,
        /// Built by `embed-block-BREAKABLE` (upstream
        /// `HorzEmbeddedVertBreakable`) rather than `embed-block-top`/`-bottom`
        /// (`HorzEmbeddedVert`). The breakable one is not laid out as inline
        /// content at all: the line breaker flushes the current line, splices
        /// the block's own vertical boxes straight into the vertical list
        /// (`AlreadyVert`, `lineBreak.ml:809-818`), and starts a fresh line.
        /// See `break_into_lines`.
        breakable: bool,
    },
    /// An UNBREAKABLE inline frame (`inline-frame-outer`/`-inner`; upstream
    /// `PHGOuterFrame`/`PHGInnerFrame`) — ATOMIC: contents are pre-fit at
    /// their natural width (`fit_cell`) and the frame never splits across a
    /// line break, matching upstream's model. `inline-frame-breakable` is NOT
    /// this variant — see [`PureHorzBox::InlineFrameMarker`].
    ///
    /// `width`/`height`/`depth` are the OUTER dims (padding included, baseline
    /// unshifted — padding grows the box, upstream lineBreak.ml's frame
    /// metrics). `contents` carry x-offsets from the frame's left edge (pad-L
    /// already applied), all on the frame's own baseline. `deco` is fired
    /// lang-side after placement; the writers draw nothing for it here.
    Frame {
        width: Length,
        height: Length,
        depth: Length,
        deco: DecoId,
        contents: Vec<(Length, PureHorzBox)>,
    },
    /// A placed block-frame marker (`VertBox::FrameStart`/`FrameEnd` after
    /// page breaking) — zero-width, renders nothing (writers' wildcard arm),
    /// read back by `fire_hooks` only.
    FrameMarker { id: DecoId, end: bool },
    /// One boundary of an inline BREAKABLE frame (`inline-frame-breakable`;
    /// upstream `HorzFrameBreakable`) — the horizontal twin of
    /// [`PureHorzBox::FrameMarker`], and for the same reason.
    ///
    /// Upstream keeps a breakable frame TRANSPARENT to the paragraph breaker:
    /// `LBFrameBreakable` threads the enclosing width map straight through its
    /// contents (`lineBreak.ml:1094`), so glue and discretionaries *inside* the
    /// frame are ordinary DP nodes of the enclosing paragraph, and `cut`
    /// (`:824`) re-frames the resulting fragments one line at a time
    /// (`append_framed_lines`), firing `decoS` for an unbroken frame and
    /// `decoH`/`decoM`/`decoT` per fragment for a broken one.
    ///
    /// This port's breaker is a flat index DP over one `Vec<PureHorzBox>`, so
    /// the same transparency is spelled the other way round: the primitive
    /// SPLICES the frame's contents into the paragraph stream (padding-L and
    /// -R as `FixedEmpty`, upstream's `append_horz_padding`) and brackets them
    /// with this zero-width marker pair, so the inner boxes simply *are* the
    /// paragraph's boxes. `fire_hooks` reassembles the fragments by walking
    /// the markers.
    ///
    /// `height`/`depth` are the WHOLE frame's padded content extent
    /// (`content ± pad`), carried on BOTH markers so that any line holding
    /// either one reserves the frame's full vertical extent. Upstream instead
    /// sizes each fragment from its own contents; for an unbroken frame (every
    /// bundled caller: `\ref`, `\href`, TOC entries — all zero-padding) the two
    /// agree exactly, and for a broken one this over-reserves a fragment by the
    /// difference between the frame's tallest content and that fragment's,
    /// which is zero for uniform text.
    InlineFrameMarker {
        id: DecoId,
        end: bool,
        height: Length,
        depth: Length,
    },
    /// `add-footnote`'s marker (v0.0.6 `PHGFootnote(imvblst)`,
    /// `horzBox.ml:283` → `ImHorzFootnote`, `:306`): a zero-width/height/
    /// depth inline box carrying the footnote's already-assembled block.
    /// Rides the paragraph like `HookPageBreak` (writers skip it via their
    /// wildcard arm); `chop_page` (pagebreak.rs) extracts it when the line
    /// carrying it is COMMITTED to a page, reserves the block's stacked
    /// height at the page bottom, and bottom-places the block in the same
    /// column (upstream `pageBreak.ml:131-142` + `handlePdf.ml:400-403`).
    /// The marker itself stays in the placed line's contents (render-inert)
    /// — extraction is a read-only scan, unlike upstream's removing
    /// `embed_page_info` (`pageInfo.ml:47`). Consequence: the block payload
    /// appears both (inert) inside its referencing line and (rendered) as
    /// bottom-placed lines — any future exhaustive consumer of a placed
    /// line's contents must treat this variant as inert or it will
    /// double-count the body.
    Footnote { block: Vec<VertBox> },
    /// An INERT reflow marker for emphasis runs (`\emph`/`\bold` in the
    /// repo-controlled stdlibs that opt in) and list-bullet fencing,
    /// emitted by the `inline-mark` primitive. Zero width/height/depth and
    /// renders nothing, so it contributes zero advance wherever it rides in a
    /// placed line's `contents`. Read only by the reflow HTML walker (the
    /// `html-support` branch's `reflow/inline.rs`), which uses
    /// `EmphStart`/`EmphEnd` to wrap `<em>`/`<strong>` and
    /// `BulletStart`/`BulletEnd` to suppress the drawn bullet/number glyph run
    /// (the real marker comes from the `<ul>`/`<ol>` itself).
    InlineMark(InlineMarkKind),
}

/// The marker kind a `PureHorzBox::InlineMark` carries. `strong` is chosen AT
/// THE WRAP SITE (which stdlib command calls `inline-mark` with which tag),
/// not recovered from the box tree: `\emph` -> `strong: false` (`<em>`),
/// `\bold`/`\strong` -> `strong: true` (`<strong>`).
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub enum InlineMarkKind {
    /// Opens `<em>` (`strong = false`) or `<strong>` (`strong = true`).
    EmphStart { strong: bool },
    /// Closes the innermost open emphasis span.
    EmphEnd,
    /// Opens a fence around a drawn bullet/number glyph run (`itemize.satyh`'s
    /// `make-bullet`/`enumerate-item`'s numeral) — the reflow walker drops
    /// everything between this and the matching `BulletEnd`, since the real
    /// `<ul>`/`<ol>` marker replaces it.
    BulletStart,
    /// Closes the bullet fence.
    BulletEnd,
    /// Immediately precedes the `pre_break` slot the LINE BREAKER spliced onto
    /// a line it chose to end here — the hyphen that prints before the break.
    ///
    /// It exists so a reflowing renderer can undo the hyphenation exactly. The
    /// splice produces an ordinary `InnerString`, indistinguishable from a
    /// hyphen the author typed, so rejoining the lines had to guess from the
    /// shape of the text — and the guess deleted real hyphens: a paragraph
    /// wrapping at `code-printer` came out as `codeprinter`. Emitted by
    /// `linebreak::line_content`, zero-width and rendered by nothing, exactly
    /// like the other marks here.
    BreakHyphen,
}

/// TeX's forced-break convention: a discretionary penalty this low or
/// lower means the paragraph breaker must end a line there. `text_to_boxes`
/// uses this to turn a UAX#14 `Mandatory` boundary (e.g. a literal newline)
/// into a break the DP cannot skip over (see `linebreak.rs`'s `floor`).
pub const FORCED_BREAK_PENALTY: i32 = -10_000;

/// The mirror convention at the other extreme: a `Discretionary` carrying this
/// penalty offers NO break at all — it exists only to render its `no_break`
/// slot, unbreakably.
///
/// This is upstream's `LBPure(lphb)`, the `PreventBreak` arm of
/// `discretionary_if_breakable` (`convertText.ml:189-190`). A pure glue box is
/// not a breakpoint but its stretch/shrink still count toward the line's
/// elasticity (`lineBreak.ml`'s `add_width_all`).
///
/// The port's box model has no separate "pure elastic box": a bare
/// `OuterEmpty` IS a breakpoint (`is_glue`). So `PreventBreak` is modelled as a
/// `Discretionary` with every break slot empty, its content in `no_break`, and
/// this penalty — which `is_break_point` reads as "not a candidate", keeping
/// one code path in `text_to_boxes` for both arms.
pub const NO_BREAK_PENALTY: i32 = i32::MAX;

impl PureHorzBox {
    pub fn natural_width(&self) -> Length {
        match self {
            PureHorzBox::InnerString { width, .. } => *width,
            PureHorzBox::OuterEmpty { natural, .. } => *natural,
            PureHorzBox::OuterFil => Length::ZERO,
            PureHorzBox::FixedEmpty { width } => *width,
            PureHorzBox::Image { width, .. } => *width,
            // Un-taken discretionary: renders as `no_break` (hyphenation —
            // `linebreak.rs`'s `line_content` handles the taken case, which
            // never reaches this generic accessor). Empty for
            // UAX#14-only discretionaries, hence zero then.
            PureHorzBox::Discretionary { no_break, .. } => no_break
                .iter()
                .map(PureHorzBox::natural_width)
                .fold(Length::ZERO, |acc, w| acc + w),
            PureHorzBox::Graphics { width, .. } => *width,
            // Fil semantics: zero natural width, like `OuterFil`.
            PureHorzBox::GraphicsOuter { .. } => Length::ZERO,
            PureHorzBox::Math { width, .. } => *width,
            // `EvHorzHookPageBreak` has width `Length.zero` (pageInfo.ml:42).
            PureHorzBox::HookPageBreak { .. } => Length::ZERO,
            PureHorzBox::Tabular(tab) => tab.width,
            PureHorzBox::EmbeddedBlock { width, .. } => *width,
            PureHorzBox::Frame { width, .. } => *width,
            PureHorzBox::FrameMarker { .. } => Length::ZERO,
            PureHorzBox::InlineFrameMarker { .. } => Length::ZERO,
            // `ImHorzFootnote` is skipped by every width scan upstream,
            // lineBreak.ml:1200/1254.
            PureHorzBox::Footnote { .. } => Length::ZERO,
            PureHorzBox::InlineMark(_) => Length::ZERO,
        }
    }

    /// `false` for every variant except the two glue kinds.
    pub fn is_glue(&self) -> bool {
        matches!(
            self,
            PureHorzBox::OuterEmpty { .. } | PureHorzBox::OuterFil
        )
    }

    /// A legal paragraph-break candidate: glue (today's only breakpoints)
    /// or a discretionary (UAX#14/hyphenation break points). CJK text has
    /// no glue at all, so discretionaries are its *only* break candidates.
    ///
    /// EXCEPT a `NO_BREAK_PENALTY` discretionary, which is upstream's
    /// `LBPure(glue)` in disguise (see that constant): it renders its
    /// `no_break` slot and offers no edge.
    pub fn is_break_point(&self) -> bool {
        match self {
            PureHorzBox::Discretionary { penalty, .. } => *penalty != NO_BREAK_PENALTY,
            _ => self.is_glue(),
        }
    }

    /// The break's own penalty (TeX's discretionary/`\penalty`
    /// convention): 0 for glue (no preference either way), a
    /// discretionary's own `penalty` otherwise.
    pub(crate) fn break_penalty(&self) -> i32 {
        match self {
            PureHorzBox::Discretionary { penalty, .. } => *penalty,
            _ => 0,
        }
    }

    /// Whether breaking here is not just legal but mandatory
    /// (`penalty <= FORCED_BREAK_PENALTY`).
    pub(crate) fn is_forced_break(&self) -> bool {
        self.break_penalty() <= FORCED_BREAK_PENALTY
    }
}

/// `horz_box`: the wrapper stays even though `Pure` is its only variant so
/// far, keeping line-break input the shape lineBreak.ml expects.
#[derive(Clone, Debug, PartialEq)]
pub enum HorzBox {
    Pure(PureHorzBox),
}