5 releases 5 tags

RSS feed

• v0.13.0 02437c224f

  Compare

  v0.13.0 Stable

  sicher released this 2026-10-22 15:13:15 +02:00 | 0 commits to main since this release

  Added

  • Continuing styled text across regions: WriteResult.resume_at (a
    TextPosition: word index plus byte offset, where writing stopped) and
    ParagraphOptions.start. Call writeStyled again with the same
    StyledText and start = resume_at to continue in the next column or page with
    every span and style intact. Before, the only continuation was
    overflow, which for styled text holds just the rest of one span -
    later spans were silently lost.
  • platen.Path: the path builder type returned by doc.createPath() is
    now exported, so helpers can take a *platen.Path.
  • Complete GSUB: every lookup type - single, multiple, alternate,
    ligature, context (formats 1-3), chained context (formats 1-3),
    extension, reverse chained - with lookup flags (ignore base glyphs /
    ligatures / marks, mark attachment classes, mark filtering sets, via
    GDEF), applied in lookup-list order over the whole run as HarfBuzz does,
    for any feature. Before, only single, ligature and chained context
    format 3 lookups were read, each wired to a few features. New default
    features, as in HarfBuzz: ccmp, locl, rlig, clig (turn off with
    setFeature). ccmp gives dotless i/j under accents (Gentium, Inter:
    j́, ị́, Lithuanian į́) and joins emoji ZWJ sequences, skin tones
    and flags into one glyph (Noto Color Emoji).
  • Default ignorables are removed: U+FE0F/U+FE0E, ZWJ, ZWNJ, zero-width
    space, soft hyphen, word joiner, BOM and the other Default_Ignorable code
    points no longer draw (U+FE0F was .notdef in Noto Color Emoji, a soft
    hyphen passed to writeAt a visible hyphen) and add no width. During
    GSUB they are skipped where HarfBuzz skips them - ZWJ always, ZWNJ only
    in context - so keycaps (1️⃣), ❤️ and 🏳️‍🌈 join, f+ZWJ+i still
    ligates and ZWNJ still blocks ligatures. Their code points stay out of
    extracted text.
  • Variation sequences (cmap format 14): a character followed by a
    variation selector draws as the glyph the font lists for the pair - CJK
    ideograph variants, text-style emoji with U+FE0E in Noto Emoji - as
    HarfBuzz does; unlisted selectors are dropped as before. font-info
    shows whether a font has them.
  • CLI shape <font> [+feat|-feat]...: glyph IDs for each stdin line, as
    rendering shapes it.

  Fixed

  • Shaping differed from HarfBuzz (found by the new differential
    test): Inter's <-> became <- plus >, case together with tnum
    gave tabular but not case forms (features ran in a fixed order, not the
    font's lookup order), and Satoshi's frac gave no numerators or
    denominators.
  • CFF fonts lost contextual alternates: a subset of a .otf font kept
    only the first glyph drawn for each character, so a contextual variant
    of a character already drawn plainly (Inter's raised hyphen in A-B
    after a-b) rendered as .notdef. Every drawn glyph is kept now, as for
    TrueType.
  • Translucent color-font gradients painted opaque: COLRv1 gradient
    stops carry an alpha, which was dropped, so overlays meant to fade out
    covered the glyph - Noto's rainbow flag 🏳️‍🌈 came out grey, and about
    2,000 Noto Color Emoji glyphs had such gradients. Stop alpha is now
    drawn through a soft mask (or /ca when it is constant).
  • A color glyph without an outline painted the whole page: Bungee
    Spice's space is a color glyph whose clip glyph has no contours; the
    empty clip was written as W n with no path, which sets no clip, so its
    gradient covered the page from that point on. Such paints are now
    skipped (an empty clip shows nothing).
  • Underlines and strikethroughs ran into the space a line wraps at
    (write()): the line was drawn with its trailing space, and the
    decoration with it. The space now hangs past the line end, as in CSS,
    also when it ends one write() call and the next call's first word
    wraps; spaces inside the line stay decorated.
  • write() lost text before a double space: line batching reset
    the line at every space it skipped, so write("aa bb") drew only
    "bb" (at aa's place), and a call starting with a space mid-line lost
    it (" over" after "jumps" gave "jumpsover"). Spaces inside a line
    are kept as typed now.
  • SVG hairlines were rounded: stroke widths were written with two
    decimals in SVG units (the tiger's 0.172 became 0.17); now four.
  • Bitmap-only fonts had an empty FontBBox: an OTB pixel font's head
    box is 0 wide, so every glyph fell outside its Type3 font's box
    (poppler warned on each). Such fonts now get [0 0 0 0], which tells
    viewers to make no assumptions about glyph sizes.
  • Color emoji ligatures extracted as U+FFFD: the Type3 ToUnicode of
    a color font mapped only single code points, so a joined emoji (ZWJ
    family, flag, keycap) copied as a replacement character. It maps to the
    whole sequence now.
  • Text extraction of substituted glyphs: ToUnicode now follows the
    shaped glyphs' clusters - a ligature maps to its characters, a glyph a
    multiple substitution produced to its own character of the
    decomposition - instead of re-guessing ligatures after shaping.
  • Words wider than the line lost their end at a region boundary: a
    word broken over lines (soft hyphens, regular hyphens, or between
    characters) that reached the bottom of the region was cut off and
    layout went on with the next word. Writing now stops mid-word and
    resume_at carries the byte offset, so the next region continues the
    word (lines are the unit, as in TeX and CSS fragmentation). Pieces are
    measured in the span's own style (bold words were broken with regular
    widths).
  • Knuth-Plass put a too-wide word on one endless line: with
    .line_breaking = .optimal, a word wider than the line made every
    break infeasible and the whole paragraph came out as a single line
    running past the region. Such a word now gets a line of its own and is
    broken into pieces, as with greedy breaking.
  • measure_only drew broken words: measuring a paragraph containing
    a word wider than the line drew that word's pieces on the page.
  • Combining marks compose to the font's precomposed glyphs: a base
    followed by combining marks (e.g. i + U+0308) was drawn as separate
    glyphs positioned by GPOS - or simply overprinted in fonts without mark
    positioning - giving a dotted i under a diaeresis. Like HarfBuzz,
    shaping now composes the sequence into the precomposed character
    (U+00EF) when the font has a glyph for it; otherwise it stays
    decomposed. Measurement and rendering share the decoding, and the
    ToUnicode text of such a glyph is the sequence as typed.
  • Unused fonts were embedded whole: a font loaded into a document but
    never drawn with (only measured, or not used at all) has no glyphs to
    subset, so output() wrote the complete original font file as orphan
    objects - 25 MB for Noto Color Emoji. Only fonts some page draws with
    are written now.
  • Justified text in embedded Unicode fonts: lines set in a single
    style were justified only through the PDF Tw operator, which applies
    to the single-byte code 32 and never to 2-byte Identity-H codes (nor to
    color-font runs), so those lines came out ragged in every
    .identity font. The extra word space is now emitted after each space
    glyph as a TJ adjustment (or Tm offset in the mark-positioning
    path), and Tw is 0 for these runs (poppler applies it to the 2-byte
    code 0x0020 anyway, e.g. a ligature subset to CID 32). New fpdf2
    reference justify_embedded.
  • char_spacing in runs with combining marks: runs with GPOS mark
    positioning place every glyph with Tm, so Tc never applied and the
    text came out unspaced while measureStyled included the spacing. The
    spacing is now added per glyph in that path.
  • char_spacing leaked into later text: a run with non-zero
    char_spacing set the PDF Tc operator but never reset it, and Tc
    survives ET, so every later text run on the page was letter-spaced
    too, while layout measured it without the spacing (overlapping spans,
    text overrunning its region). The run now resets Tc before ET.
  • Gradient stops are copied: addLinearGradient/addRadialGradient
    kept a pointer to the caller's stops slice until output(); stops in
    a stack frame that had returned were read after free (garbage colors or
    a panic). The document now keeps its own copy.

  Internal

  • GSUB differential test: test/shaping/harfbuzz.txt holds HarfBuzz's
    glyphs for Latin, diacritic, default-ignorable and emoji texts in Inter,
    Satoshi, Gentium, Bungee, Charter and Noto Emoji under several feature
    sets (HarfBuzz removing default ignorables, as platen does), plus a synthetic font
    with one scenario per lookup type and flag
    (scripts/gen_gsub_test_font.py); scripts/gen_shaping_reference.py
    regenerates it.
  • zig build test -Dtest-filter=NAME runs only matching tests.
  • The 19 references platen generates itself (color fonts, SVG,
    decorations, line breaking, transforms, diacritics) are now also
    compared rendered, pixel for pixel (compare_pdfs.py --compare-render,
    poppler): the text-and-position comparison missed drawing changes.
    Seven had drifted - three from bugs fixed here (color_v1_gradient had
    the page-filling space baked in), the rest from intended changes
    (svg_embedding predated per-path stroke widths: its tiger had 3x
    thick whisker outlines).

  Downloads

  • Source code (ZIP)
    0 downloads
  • Source code (TAR.GZ)
    0 downloads
• v0.12.0 b6f2978ef2

  Compare

  v0.12.0 Stable

  sicher released this 2026-10-17 19:26:40 +02:00 | 34 commits to main since this release

  Added

  • Font collections (.ttc/.otc): doc.loadFontFromCollection(data, index, encoding) loads one face of a collection (the Document keeps its
    own copy). platen.extractFace rebuilds face N as a standalone font
    (tables byte-identical, 4-byte aligned, checksums kept);
    platen.listFaces / font_collection.describeFace give each face's
    family, subfamily, PostScript name, weight, width and bold/italic.
    A plain .ttf/.otf counts as a one-face collection.
  • System font lookup: platen.findSystemFont(io, allocator, family, .{ .bold, .italic }, opts) and platen.findSystemFontFamily search the
    platform font folders recursively and return caller-owned standalone
    sfnt bytes (extracted from collections) with family, subfamily,
    PostScript name, path and face index; FoundFamily.memoryOptions()
    feeds loadFontFamilyFromMemory. Matching uses the name table
    (typographic family, case-insensitive) and the subfamily; the per-user
    folders come from opts.home, so the library stays stateless.
  • CLI: font-faces, extract-face, find-font.
  • Legacy kern table kerning: fonts that kern only through a TrueType
    kern table (format 0; Microsoft and Apple layouts) are now kerned in
    rendering and in measureStyled; before, only GPOS kerning was read and
    such fonts (e.g. Bitstream Vera) came out unkerned.
    GPOS kerning still wins when present, as in HarfBuzz.

  Fixed

  • .win_ansi TrueType fonts dropped composite glyphs: the WinAnsi
    subsetter copied a composite (e.g. Inter's i, j, colon = base + dot)
    without its components and kept the original component IDs, so those
    characters rendered blank. Components are now included and remapped, as
    the .identity path already did.
  • PDF/A dates: /Info and XMP each read the clock, so a second boundary
    between them made CreationDate and xmp:CreateDate differ and
    veraPDF failed the file (6.7.3). output() now reads the clock once.
    Explicit metadata.creation_date/mod_date reached /Info only, so
    PDF/A with explicit dates was never compliant; they are now converted
    to ISO 8601 for XMP (same precision and time zone). Under PDF/A an
    explicit date that is not a valid PDF date fails with
    error.InvalidDate.
  • File IDs came from the millisecond clock: the trailer /ID (PDF/A and
    encrypted output) and the XMP DocumentID/InstanceID (one shared,
    non-UUID string). Outputs in the same millisecond, e.g. concurrent
    documents, got identical IDs, and doc.rng did not make output
    reproducible as documented. All three now come from doc.rng (default:
    system CSPRNG); the XMP IDs are distinct version-4 UUIDs. With doc.rng
    set and explicit dates, PDF/A and encrypted output is byte-identical.

  Downloads

  • Source code (ZIP)
    0 downloads
  • Source code (TAR.GZ)
    0 downloads
• v0.11.0 eeaca8f861

  Compare

  v0.11.0 Stable

  sicher released this 2026-10-06 10:02:58 +02:00 | 51 commits to main since this release

  Added

  • Bold and italic per span: StyledText spans take .bold and
    .italic (st.append("word", .{ .bold = true })), resolved through the
    base font's family like TextStyle.bold/.italic; widths use each
    span's variant, so wrapping and justification stay exact. A variant the
    family lacks is error.FontVariantNotLoaded, before anything is drawn.

  Fixed

  • writeBlock, and writeStyled with plain spans, ignored the base
    style's color and decoration (text came out black, without
    underline). TextStyleOptions fields are now optional overrides: null
    inherits from the base style. Existing calls such as
    .{ .color = platen.Color.blue } compile unchanged.

  Internal

  • The PDF comparator behind the reference tests compares fonts whenever
    either page has any (a page that lost all its fonts used to pass) and
    treats qpdf failures as errors; it has its own tests in the suite.
  • zig build test re-runs when a reference PDF, fixture or comparator
    script changes (they are declared inputs); before, a regenerated or even
    corrupted reference could be served a cached pass.

  Downloads

  • Source code (ZIP)
    0 downloads
  • Source code (TAR.GZ)
    0 downloads
• v0.10.1 c881e08f87

  Compare

  v0.10.1 Stable

  sicher released this 2026-10-06 08:41:06 +02:00 | 65 commits to main since this release

  Fixed

  • Package size: dependents fetch ~0.4 MB instead of 45 MB and cache
    ~1.8 MB instead of ~100 MB. build.zig.zon .paths lists only the
    library (build files, src/, README, LICENSE, CHANGELOG), and
    .gitattributes keeps tests, fixtures, examples and dev scripts out of
    the forge's source archives. Clones are unchanged. The package hash
    differs from 0.10.0.

  Internal

  • CI prints the test summary, so runs show test counts and skips.

  Downloads

  • Source code (ZIP)
    0 downloads
  • Source code (TAR.GZ)
    0 downloads
• v0.10.0 b6367d78be

  Compare

  v0.10.0 Stable

  sicher released this 2026-10-06 01:22:39 +02:00 | 70 commits to main since this release

  First tagged release; earlier history is in git. Requires Zig 0.17.

  Behavior changes — callers may need to adapt

  Invalid input is now an error, never silently clamped or converted.

  • Numbers: every Document drawing/text/layout call rejects NaN,
    infinities and magnitudes above platen.max_value (1,000,000) with
    error.InvalidNumber, before changing the document. (A huge
    drawCircle radius used to crash.)
  • Fonts: unitsPerEm outside 16–16,384, or metrics that do not fit
    once scaled, are error.InvalidFont at load. Glyph IDs a font does not
    have are left out of subsets (they render as .notdef).
  • SVG: numbers that are not finite or exceed ±1e9 are invalid (treated
    like any unparseable value); transforms that compose past that bound are
    error.InvalidSvg.
  • CFF fonts: subroutine nesting over 10, expanded glyphs over 65,535
    bytes, or runaway expansion are error.InvalidCFF when subsetting.
  • PNG: chunk CRCs are verified (error.PngChecksumMismatch); bit depths
    invalid for the color type, zero sizes, and sizes the data cannot hold
    are rejected before any allocation.
  • PDF/A: standard (non-embedded) fonts, and SVGs containing <text>,
    are error.FontNotEmbedded at every level. Under PDF/A-1b, images with
    transparency are error.TransparencyNotAllowed and 16-bit images
    error.BitDepthNotAllowed (downsample, or use PDF/A-2b).
  • API types: opentype.MarkOffset fields are i32;
    CoverageTable.getGlyphCount returns usize and getGlyphAt takes one.
  • Link annotations carry /F 4 (print), as PDF/A requires.
  • The zig build docs step (broken since the generated docs were
    removed) is gone.

  Added

  • PNG: Adam7 interlacing, 16-bit images with alpha (16-bit soft mask),
    and tRNS color-key transparency on grayscale and RGB images. All 161
    valid PngSuite images now decode to PngSuite's own expected pixels.

  Fixed

  • Documents with more than 32 images, embedded fonts, SVGs or opacity
    states (or more than 16 gradients) produced broken PDFs: resource names
    collided.
  • output() could fail with WriteFailed on large page dictionaries.
  • Crashes on hostile input, found by fuzzing and an audit of ~50 sites:
    integer overflows and out-of-bounds reads in the TrueType, OpenType
    (GSUB/GPOS), COLR, sbix, CBLC and SVG-table parsers, the PNG/JPEG
    decoders and the TrueType/CFF subsetters.
  • CFF offset sizes were off by one: a 255-byte (or 65,535-byte) item
    crashed subsetting — reachable with ordinary fonts.
  • SVG: non-finite numbers were written into the PDF as inf; elements
    leaked on errors, in <defs> without an id, and on duplicate ids.
  • PDF/A output failed veraPDF: invalid dates (a Zig 0.17 formatting
    change), link annotations without the print flag, and standard fonts.
  • README examples that no longer compiled.

  Internal

  • Zig 0.17 migration; the version is single-sourced from build.zig.zon.
  • Tests: exact pixel comparison of images (none before), PngSuite against
    its own pixel truth, veraPDF validation of PDF/A output, and silent
    passing runs. 508 tests.
  • Releases follow the urverk devkit main/dev model (.devkit.conf).

  Downloads

  • Source code (ZIP)
    0 downloads
  • Source code (TAR.GZ)
    2 downloads