pandoc 3.12

I'm pleased to announce the release of pandoc 3.12,
available in the usual places:

Binary packages & changelog: https://github.com/jgm/pandoc/releases/tag/3.12

Source & API documentation: http://hackage.haskell.org/package/pandoc-3.12

This release focuses on performance. Here are some benchmarks showing
the improvement since pandoc 3.10.2:

There shoud be even more dramatic gains in image-heavy documents, because
of optimizations in ImageSize.

In addition to performance improvements, there are many other small
improvements and bug fixes. See the changelog for full details.
Most of the performance gains and and many of the bugs were found
with the help of Claude Fable.

API changes:

• Text.Pandoc.ImageSize: ImageType now derives Eq.
• Text.Pandoc.Sources: new functions takeWhileP, takeWhile1P.
• Text.Pandoc.Shared: new functions stringifyInlines,
  compactifyTable. FromText constraint added to the signatures
  of htmlAttrs and tagWithAttrs.

Other changes of note:

• -t xml now respects --standalone, and produces a
  fragment (just the blocks) if it is not provided, like -t native.
• The default CSS now includes dark mode support.

Thanks to all who contributed, especially new contributors
Andonome, Clar Fon, Gaurav Vijay Jadhav, Robert Szarka,
Samuel Huang, Yusuf Efe, and zenor0.

Click to expand changelog

• Markdown reader:

  • Make alert keywords case-insensitive (#11836, Hendrik Erz). In addition, the class alert is now added to the produced Divs.
  • Fix stream position handling in base64DataURI.
  • Cheaply reject bareURL before trying uri/emailAddress.
  • Reset sourcepos in parseWithString' when parsing a block quote or list item. Otherwise it can happen that by the time parseWithString' is called, the position has already been set to the next file on the command line. Fixes an odd bug with rebase_relative_paths (#11888).
  • rebase_relative_paths: recognize URLs with unknown schemes (#11858).
  • Use takeWhile1P in the hot inline parsers str, code, enclosure, and mmdShortSubscript.
  • Reset sourcepos in parseWithString' when parsing a block quote or list item (#11888). This ensures that rebase_relative_paths will see the right source file.

• Typst reader:

  • Map form: "prose" citations to AuthorInText (#11846, Samuel Huang).
  • Make the handler maps monomorphic by wrapping the handlers in newtypes with polymorphic fields (BlockHandler, InlineHandler).
  • Store document labels in a Set instead of a list.
  • In pInline, only perform the label-target check for ref elements, not for every inline element.
  • Handle highlight as a mark span (#11879, Samuel Huang).
  • Collapse citations around soft break (#11897).
  • Handle block content in inline element bodies (#11881, Samuel Huang).
  • Handle #par (explicit paragraph element).

• LaTeX reader:

  • Support LaTeX3 (xparse) document commands (#7540). Support the features described in the LaTeX usrguide.
  • Fix \qed to produce U+00A0 (nbsp) instead of BEL.
  • Fix unescapeURL handling of escaped backslash.
  • Require that \newif names begin with “if”.
  • Fix position off-by-one after ## in tokenizer.
  • Fix doubled source positions in retokenizeComment.
  • Make raw token capture O(1) per token.
  • Make untokenize linear instead of quadratic.
  • Fail macroDef fast on non-macro-defining commands.
  • Peek at next token directly in peekTok instead of going through satisfyTok.
  • Skip doMacros state update for non-macro tokens.
  • Build command dispatch maps (inlineCommands, etc.) once per parse.
  • Keep ASCII quotes when ligatures are disabled.
  • Don’t discard state changes made in optional arguments.
  • Keep nested conditionals balanced in \iftrue etc.

• Docx reader:

  • Read ScreenTips as link titles (#11869, Robert Szarka).
  • Use comment-id instead of id in AST for comments.

• ODT reader:

  • Rewrite in monadic style. Net -994 lines. No changes to test output.
  • Handle textProperties on paragraph styles (#2623). Previously these just got ignored, not applied to the paragraph’s text.

• HTML reader:

  • Don’t require a closing tag for checkbox inputs. <input> is a void element.
  • Allow omitted </tr> in tables. The </tr> closing tag is optional in HTML.
  • Don’t drop cells when there are too few <col> elements.
  • Let the first of duplicate attributes win. HTML specifies that the first occurrence wins.
  • Respect raw_html for inline <style> elements.
  • Use a Map for the note table.
  • Report the noteref position for unresolved notes. The ReferenceNotFound warning was logged after the whole document had been parsed, so it always pointed at the end of the input. Record the position of the first reference to each note and use it in the warning.
  • Find list style anywhere in the class attribute. Previously <ol class="fancy lower-roman"> got DefaultStyle, because the whole class attribute was compared against the known style names. Check each class individually. As a side effect, an unrecognized class no longer prevents falling back to the style attribute.
  • Limit iframe nesting depth.
  • htmlTag: Don’t copy the remaining input on each invocation. A space was appended to the remaining input to guarantee a TagPosition token after the parsed tag; since the input is a strict Text, this copied the entire remaining input every time htmlTag was called (e.g. for every inline HTML tag in a markdown document), giving quadratic behavior in tag-dense documents. Instead, handle the case where the tag is the final token by computing the end-of-input position directly.
  • Pass the tag name to pSpanLike. The inline dispatcher already knows which span-like element it is looking at, so there is no need for pSpanLike to try a parser.
  • Fix pre/code attribute precedence for first-wins dedup.
  • Implement pSatisfy as a single parsec primitive.
  • Add a fast path to pTagText: when the text contains no character that could parse as anything but Str, Space, or SoftBreak under the enabled extensions (and we are not in a pre element), return B.text directly.
  • Reorder pTagContents to try pStr and pSpace before the math, smart punctuation, and raw TeX parsers. This is safe because pStr cannot consume the special characters that start those parsers, and it avoids most guard checks on the slow path. This makes the html reader benchmark about 33% faster and halves its allocation.

• Muse reader:

  • Try str earlier in inline parser. This makes the reader 2x faster and reduces heap allocation by 60%.
  • Check raw line before parsing table rows. Table row parsers are speculatively tried at every paragraph line boundary via the terminator continuation. This change gives another 2x speed improvement.

• Org reader:

  • Fix hard parse failure on bare backslash.
  • Recognize .pdf as an image format (#11859). This matches the behavior of Emacs, which will render [[file:foo.pdf]] as an image.
  • Make #+OPTIONS: ^:nil disable all sub-/superscript parsing.
  • Allow - and _ in inline footnote labels. Org footnote labels may contain word-constituent characters, hyphens and underscores.
  • Fix swapped arguments in exportSettings parser.
  • Don’t lowercase meta keyword twice.
  • Avoid quadratic complexity when parsing table rows. Parsing a 32000-row table drops from 7.8s to 0.9s (and scales linearly instead of quadratically); maximum residency also improves.
  • Use a Set for anchor ids. Reading a document with 64000 anchors and as many internal links drops from 23.5s to 5.2s and now scales linearly.

• RST reader:

  • Use a predicate to test for for special characters.
  • Avoid double-parsing lines preceding a non-underline.
  • Fail fast when no link can start at this position.
  • Use lookupGE to find the next anonymous key. Parsing a document with 16000 anonymous links drops from 5.2s to 2.0s.
  • Replace association list with Map for resolving note references. Parsing a document with 16000 named notes drops from 5.8s to 3.5s, and scaling is now linear.
  • Fix a bug that led to an empty class for code blocks with no specified language.

• CommonMark reader:

  • Avoid nested walks in tex_math_gfm handling.
  • Avoid rebuilding the token list with map id in sourceToToks in the common case where the source starts at line 1.

• AsciiDoc reader:

  • Resolve footnotes, stem, and icons during conversion. Previously the reader made three separate passes over the parsed AsciiDoc AST (for footnotes, stem math types, and icons) before converting to the pandoc AST. Instead, resolve all three during the toPandoc conversion, which already traverses everything once, threading footnote state and document attributes through a StateT layer. Table headers are now converted before body rows so that footnote resolution follows document order. This makes the reader about twice as fast on typical documents.

• Djot reader:

  • Fix a bug that led to an empty class for code blocks with no specified language.

• Vimwiki reader:

  • Split class attribute and put it in the proper slot in pandoc’s Attr.

• Typst writer:

  • Omit blank line at end of block (#11844). This is just a cosmetic change; it is not semantically significant.
  • Emit label for bare table if present (#11849). Previously a label was only emitted if the table was figurized (and not .typst:no-figure).

• RST writer:

  • Apply nowrap to just footnote label, not body. This bug surfaced after nowrap was fixed in doclayout.

• Docx writer:

  • Fix sizing for images (#11838). Setting size to a percent now scales image to percent of page width. Previously a percent width or height would just provide a maximum bound rather than scaling.
  • Include default style even if paragraph has non-style props (#11867). Previously a paragraph that was, e.g. center-aligned would be missing its default Body Text style. Also ensure that any block sequence (list items, table cells, block quote, div), First Paragraph is set for the first paragraph in the item.
  • Use _ to start all bookmark names (#11845). This ensures that they are “hidden” and will not be read by screen readers.
  • Honor CSL hanging-indent and spacing hints (#11871, Samuel Huang).
  • Add a fast path to withDirection.
  • Avoid double withDirection for Space and SoftBreak.
  • Map link/image titles to ScreenTips.
  • Write link titles as ScreenTips (#11869, Robert Szarka).
  • Properly signal error if reference docx can’t be parsed. Previously this led to an incomprehensible error at a later phase.
  • Make convertSpace linear instead of quadratic. On an ad hoc benchmark (4 paragraphs of 80K words each), conversion time drops from 3.2s to 1.1s, and time no longer depends on paragraph length (16 x 20K words previously took 1.8s, now also 1.2s).
  • Cache token-type styles instead of rebuilding per Code. On an ad hoc benchmark with 100K inline code spans, conversion time drops from 2.4s to 1.5s.
  • Add FirstParagraph class after display math (#11900). This way the continuation text can be styled flush-left in a style where normal paragraphs are indented.
  • Treat a “mixed” widths table as all-default (#11899). Previously if some columns were ColWidthDefault and others ColWidth 0.x, we would get columns with a specified width of 0 for the default ones. Instead, treat all columns as default in this case.
  • Use comment-id instead of id in AST for comments. Note that the Docx writer will still interpret an id attribute for legacy compatibility, so if you use markdown files that specify id, they should still work.

• TEI writer:

  • Use rend, not rendition attribute, on milestone (#11842, Yusuf Efe) rendition takes pointers to rendition descriptions, while rend is the free-text attribute, which is what a plain “line” value needs.

• LaTeX writer:

  • Avoid nested walk in table cell line-break handling.
  • Don’t use footnotehyper for notes in longtable. Instead, generate them manually as we do for floating tables. This removes our dependency on footnotehyper, and resolves a compatibility problem with endfloat (#11857).
  • Remove unused stInternalLinks state field. The field has not been read since the writer switched to always adding hypertargets, but we were still doing a full-document query to populate it.
  • Avoid needless conversions in sectionHeader. The note-free and link-free variants of the heading text were rendered for every heading, even though the former is only needed when the heading contains a note, image, or identified span, and the latter only for unnumbered, listed headings.
  • Fuse preprocessing passes in inlineListToLaTeX. The strut-insertion and quote-kerning fixups were separate list traversals, allocating an intermediate list each, and this function is called at every level of inline nesting. Combine them into a single pass.
  • Avoid String round-trip for highlighted inline code.
  • Don’t unpack code string to choose lstinline delimiter.
  • Only compute PDF trailer ID when SOURCE_DATE_EPOCH is set.
  • Use showHex instead of (slower) printf in toLabel.

• RST writer:

  • Don’t re-transform stored labels and alt text. Reference labels, image substitution labels, and alt text are stored after the document-wide walk in inlineListToRST has already applied transformInlines (flattening, backslash-space insertion, etc.). Rendering them with inlineListToRST applied these non-idempotent transformations a second time, inserting duplicate \ markers. With --reference-links this could make the inline reference and its definition render differently, producing a broken RST reference.

• Native writer:

  • Render directly instead of using pretty-show. Previously writeNative used pretty-show’s ppDoc, which shows the document, tokenizes and re-parses the result into a generic Value, and lays that out via Text.PrettyPrint.HughesPJ. The layout step dominated the cost of the writer (and of any pipeline producing native output). We now build a width-cached layout tree directly from the AST and render it with a small renderer that reproduces HughesPJ’s layout algorithm exactly (including the ribbon computation with ribbonsPerLine = 1.2 and the treatment of glued closing delimiters), so the output is byte-for-byte identical to before. Verified against the old binary on all golden .native files and the markdown test corpus at many column widths, in both standalone and plain modes. The writer is about 4x faster. Also remove the pretty and pretty-show dependencies.

• XML writer:

  • Respect --standalone. When standalone is selected, we get a full Pandoc element with xml header and metadata. When not, we get a fragment – just the blocks.
  • Use the pretty-printer instead of manual newlines. Render the document with ppcElement, using a configuration that treats elements with inline content as inline tags, so that no significant whitespace is added inside them. Consecutive text nodes are merged, SoftBreak is written as a literal newline, and whitespace runs that would not survive a roundtrip (e.g. " \n" or "\n\n") are encoded as Space and SoftBreak elements.
  • Encode attribute names that are not valid XML names. Pandoc attribute names may contain characters that are not allowed in XML attribute names, such as colons (typst:property). Use the common convention of encoding such characters as _xHHHH_, where HHHH is the hexadecimal code of the character: the writer encodes attribute names (foo:bar becomes foo_x003A_bar) and the reader decodes them.
  • Don’t drop attributes with empty values. The writer dropped any attribute with an empty value, so user key-value attributes like ("k","") disappeared and did not round trip.
  • Improve performance: the XML writer is now on par with the JSON writer.

• EPUB writer:

  • Render TOC item titles in the host monad. Previously each TOC item title in the nav entry was rendered with a separate runPure (writeHtmlStringForEPUB ...). Since every one of these invocations starts with a fresh CommonState, the translations YAML file was re-read and re-parsed for every TOC item, which accounted for a significant part of the EPUB writer’s run time on documents with many sections.
  • Speed up MathML/SVG detection for the manifest.

• Powerpoint writer:

  • Add archive entries in a single pass.
  • Cache the parsed slide master in WriterEnv.
  • Skip speaker-notes walk when there are no notes.

• ANSI writer:

  • Fix missing bar on blockquote’s first line (#11804, Gaurav Vijay Jadhav D.).

• HTML writer:

  • Fix typo in intrinsicEventsHTML4. The list had onmouseout twice and was missing onmousemove.
  • Don’t emit <p></p> for paragraphs with no rendered content.
  • Improve email obfuscation. Preserve formatting (e.g. emphasis) in the link text of obfuscated mailto links. Preserve link attributes in reference- and JavaScript-obfuscated links. Avoid double-escaping the already-rendered link text in the fallback branch for unparseable mailto URLs.
  • Respect --id-prefix in EPUB3 footnote section id.
  • Respect incremental/nonincremental classes inside columns.
  • Make KaTeX CSS URL handling consistent with the JS URL.
  • Use truncate consistently for table width percentages. Previously the table width could exceed the sum of column widths.
  • Avoid walking section contents when not producing slides.
  • Hoist attribute set unions to top level.
  • Make strToHtml more efficient. Replace the T.groupBy-based implementation, which allocated a list of Text fragments and round-tripped through String, with a simple T.break scanner.

• Org writer:

  • Don’t render table body rows twice. Table rows were converted once to compute column widths and then again to produce the output. Since blockListToOrg is stateful, any footnote in a table cell was registered twice, yielding duplicate footnote definitions and skewed numbering; it also doubled the rendering work. Reuse the first conversion.
  • Use #+begin_export html for raw HTML blocks. #+begin_html was removed in Org 9.0 (2016).
  • Don’t treat bare punctuation as a list marker. The check that keeps ordered list markers from ending up at the beginning of a line matched Str "." and Str ")", since T.all isDigit "" is True. Require at least one digit.
  • Don’t emit <<>> for spans with no id.
  • Avoid = delimiter for inline code containing =. Org has no escape mechanism inside verbatim text, so =code with == did not parse as verbatim. Fall back to the equivalent ~...~ delimiter when the content contains = (and no ~).
  • Escape square brackets in links. Link targets containing square brackets broke the bracket link syntax. Escape targets the way Emacs’ org-link-escape does: backslash-escape brackets and double backslash runs occurring before a bracket or at the end of the target. Link descriptions cannot contain escapes; instead, like org-link-make-string, insert a zero-width space between consecutive closing brackets and before a closing bracket at the end of the description.
  • Build escaped strings from chunks, not characters. escapeString allocated one Doc node per character for any string containing a non-alphanumeric character. Split on the (rare) special characters instead and emit intervening text as single literals. No change in output.
  • Emit definitions for footnotes nested in footnotes.
  • Use a counter for footnote numbers. Computing the reference number as length stNotes + 1 walked the accumulated note list for every footnote, making note numbering quadratic in the number of notes.

• Org reader and writer:

  • Make code line comma-escaping match Emacs’ behavior. Org’s escaping rule for code in src/example blocks adds a comma to lines matching ^[ \t]*,*(\*|#\+), i.e. lines already starting with commas before * or #+ get an additional comma; unescaping removes one comma from such lines. The writer previously left a literal ,#+foo line unescaped, and the reader then stripped its comma when reading the result back, corrupting the code on round trips. Writer and reader now both handle runs of commas, matching Emacs.

• Text.Pandoc.ImageSize:

  • ImageType now derives Eq [API change].
  • Fix typo in bare JPEG signature detection.
  • Fix size detection for lossless (VP8L) WebP.
  • Make EMF parsing more robust.
  • Don’t let zlib errors escape as exceptions in pdfSize. Treat malformed streams as a parse scanner (so we keep scanning the rest for a /MediaBox).
  • Fix AVIF detection and parsing.
  • Take bounding box origin into account for EPS. The size was computed from the upper corner alone, giving wrong dimensions for EPS files whose bounding box has a non-zero origin.
  • Handle commas and fractional numbers in SVG viewBox.
  • Add tests for image type and size detection (Tests.ImageSize).
  • Determine JPEG size without decoding the image (which can allocate huge amounts of memory). If the header scan fails, we still fall back to the full decoder.
  • Scan PDF object streams in chunks, not byte by byte.
  • Speed up findSvgTag by using a single pass. Up to 60X faster on files with few < characters.
  • Determine PNG size without decoding the image. If the header scan fails, we still fall back to the full decoder.
  • Use writerDpi for AVIF images instead of hardcoding 72. With the default options this changes the assumed resolution from 72 to 96 dpi.
  • Allow whitespace between number and unit in numUnit. So e.g. width="3 cm" is now recognized.
  • Handle largesize and size-0 boxes in AVIF parser.
  • Make checkDpi default to 72 for negative dpi values, not 0; a negative dpi would produce negative dimensions in sizeInPoints.

• Text.Pandoc.SelfContained:

  • Fix inverted charset condition in makeDataURI.
  • Keep semicolon in @import fallback output.
  • Make </script check case-insensitive.
  • Prefix all url(#...) occurrences in SVG attributes. Previously only the first got prefixed rewritten.
  • Handle gzip decompression errors gracefully.
  • Remove unused isHtml5 field from ConvertState.
  • Cache fetched resources. Previously every occurrence of a resource was fetched, decompressed, and CSS-rewritten independently, so a document referencing the same image N times triggered N network requests. Failed fetches are cached too, so each missing resource is now reported only once.
  • Only add role and aria-label when inlining SVGs. Do not add them to other elements with src attributes.
  • Escape only what is needed in textual data URIs.

• Text.Pandoc.UTF8:

  • Avoid copying input when it contains no CRs. toText and toTextLazyunconditionally ran a CR-removing filter over the input, allocating a full copy of the document even in the common case where no CRs are present. Check for a CR first (B.elem, a fast memchr) and reuse the input buffer unchanged if none is found; for the lazy variant, do this chunk-wise to preserve laziness. On a 10 MB LF-only input this makes toText over 4x faster; when CRs are present the extra scan is not measurable.
  • Make readFile exception-safe. Use withFile instead of openFile so the handle is closed even if reading throws.

• Text.Pandoc.Class:

  • Avoid copying input in toTextM. Skip the CR-filtering copy when the input contains no CRs, as already done in Text.Pandoc.UTF8.toText.
  • Make runSilently error-safe. Previously, if the action passed to runSilently threw an error that was later caught, the verbosity remained pinned at ERROR and all previously accumulated log messages were lost. Now the original log and verbosity are restored even when the action fails.
  • Fix isRelativeToParentDir. Compare the first path component rather than just looking at a prefix, to correctly handle paths like ..foo/bar.yaml.
  • Fix base64 detection in extractURIData. The base64 indicator in a data URI is the final parameter of the media type and may follow other parameters, e.g. charset. Previously, the code expected ;base64 to be the only parameter.
  • Fix percent-decoding in extractURIData.
  • Report accurate offset in toTextM errors. Scan for the first invalid UTF-8 sequence and report its actual position and byte.
  • Reset HTTP manager in setNoCheckCertificate. The HTTP manager is created lazily with TLS settings based on stNoCheckCertificate and then cached in CommonState, so changing the option after the first request had no effect. Discard the cached manager when the option’s value changes.
  • Don’t follow symlink cycles in addToFileTree.
  • Finish factoring openURL into Text.Pandoc.Class.IO.HTTP. Commit 455bea9 added the new module but did not register it in pandoc.cabal or remove the original definitions.
  • logOutput: avoid multiple hPutStrLn, which can cause confusing interleaving.

• Text.Pandoc.MediaBag:

  • Use hashlazy to avoid copying media contents.
  • Treat data: and file: URI schemes case-insensitively.
  • Make MediaItem mime type and path strict. mediaContents is left lazy so contents need not be forced at insert time.
  • Use Text.Pandoc.URI.isURI in canonicalize. Network.URI.isURI treats Windows drive-letter paths like c:/foo.png as URIs.
  • Only reject .. as a path component. The insertMedia check used isInfixOf, so a harmless name like foo..bar.png was silently renamed to its content hash.
  • Collapse . and .. components in canonicalize. normalise does not remove redundant path components, so img/../a.png and a.png were distinct keys. Use makeCanonical (as PandocPure’s FileTree already does for its path-indexed map), which also handles duplicate and trailing slashes, replacing backslashes with slashes first.
  • Prevent mediaPath collisions between keys. The friendly mediaPath was derived by percent-unescaping the key, so distinct keys like a%20b.png and a b.png produced the same mediaPath (“a b.png”) and silently clobbered each other on extraction (and inside docx/epub archives). Now the original name is only kept if the key contains no percent sign, so mediaPath equals the key and distinct keys yield distinct paths; anything percent-encoded gets a content-hash name. Hashed names can only coincide for identical contents, which is harmless.

• Text.Pandoc.XML.Light:

  • Fix escaping of repeated ]]> in CDATA.

  • Make escStr more efficient.

  • Avoid round-trips in ppCDataS prettify path. This makes showCData and ppcCData unused, so they are removed.

  • Parse XML fragments from the event stream instead of using xml-conduit’s document parser, which requires a single root element. Our earlier woraround with a wrapper element was fragile. Behavior changes:

    • Content fragments with an XML declaration or DOCTYPE followed by multiple root elements, and text-only or empty input, now parse instead of erroring.
    • Attributes now preserve document order instead of being sorted alphabetically (the document parser stored them in a Map).
    • Errors for unresolved entities and mismatched tags now report source positions.

• Use text-builder package for rendering, instead of text’s lazy Text builder. On a large document this cuts docx conversion time by about 13%; output is byte-for-byte identical.

• Expose ppcTopElement from Output.

• ConfigPP now has a field inlineTag that checks for inline tags. Inline tags are printed on one line and not indented, by default. Export useInlineTags, prettyConfigPP.

• Text.Pandoc.Sources:

  • Add bulk takeWhileP/takeWhile1P combinators [API change]. Character streams over Sources previously had to be consumed one character at a time via satisfy, at a cost of several allocations and monadic binds per character. The new combinators scan a whole run of matching characters with a single parser invocation using T.span, while replicating the exact semantics of T.pack <$> many/many1 (satisfy f), including empty-chunk handling, position updates at chunk boundaries, and parsec’s error messages.
  • Use the new bulk takeWhileP/takeWhile1P combinators across readers.

• Text.Pandoc.Translations:

  • setTranslations now keeps an already-loaded translation table when the language is unchanged, instead of unconditionally clearing the cache and forcing a re-read and re-parse of the translations YAML file on the next translateTerm.
  • Term names in translation files are now parsed with a precomputed Map lookup instead of the derived Read instance, which is very slow for a 22-constructor enum.

• Text.Pandoc.Chunks:

  • Remove vestigial nav-path attribute and rmNavAttrs walk. This no longer did anything; output is unaffected.
  • Use compactifyTable for tables produced by all readers. Remove old ad hoc paraToPlain at the table cell level. This should ensure that we don’t get tables that mix Plain and Para (#11864). Such tables tend to look funny when rendered in docx and other formats.

• Text.Pandoc.Data:

  • Fix getDataFileNames with -embed_data_files. Previously it was not looking in the right directory and not recursing.

• Text.Pandoc.Shared:

  • taskListItemFromAscii: Fix incorrect treatment of [ ] as checked.
  • Add stringifyInlines, a single-pass stringify for inlines [API change]. This is about 6x faster than stringify for long inline sequences
  • Speed up stringify by making it accumulate [Text] and concatenate once at the end, instead of mappending at every node.
  • Use stringifyInlines instead of stringify where possible.
  • Add new function compactifyTable. [API change] This converts cells that consist in a single Para block to a Plain, provided the table contains only such cells (or empty cells).
  • Drop expired entries in decrementTrailingRowSpans. Entries whose RowSpan fell to 0 were left in the map, relying on every consumer to guard against them. Delete them instead.
  • Recognize empty task list items in toTaskListItem.
  • Make endsWithPlain look inside DefinitionList. endsWithPlain recursed into the last item of BulletList and OrderedList but ignored DefinitionList, so list items ending with a compact definition list were treated as loose by the RST, Org, and Haddock writers.
  • Avoid Text -> String round-trips in htmlAttrs. This adds a FromText constraint to htmlAttrs and tagWithAttrs [API change].
  • Fuse traversals in ensureValidXmlIdentifiers.

• Text.Pandoc.Writers.Shared:

  • Fix logic bug in splitSentences.
  • lookupMetaBool: treat empty block or inline list as False.
  • htmlAttrs: escape id and class attributes, like the others.
  • stripLeadingTrailingSpace - strip multiple Space, if present.
  • toSubscript: handle minus sign.
  • Fix ensureValidXmlIdentifiers for Figure and table sub-elements, resolving a bug that produced broken internal links in the HTML4/XHTML, EPUB, DocBook, TEI, ICML, FB2, and ODT writers.

• Text.Pandoc.Parsing:

  • Improve performance of uriScheme by using a trie. Up to 15% faster in URL-heavy documents with autolink_bare_uris.
  • Minor code cleanup.
  • Remove $ checks in math when delim is \( or \\(.
  • Make anyOrderedListMarker more efficient.

• HTML template:

  • Dark mode support (#11831, Clar Fon).

• reveal.js template:

  • Fix plugin paths (#11907).

• flake.nix: parse allow-newer and allow-newer-deps in stack.yaml.

• Make embed_data_files flag default to True. Remove flag settings from cabal.project. This makes it possible to override it on the command line.

• Depend on commonmark 0.3.1, commonmark-extensions 0.2.7.3, commonmark-pandoc 0.3.0.2 (major performance improvements).

• Depend on released asciidoc 0.1.1 (major performance improvements).

• Use released texmath 0.13.3 (major performance improvements).

• Depend on released djot 0.1.4.3 (major performance improvements).

• Use released citeproc 0.14 (major performance improvements).

• Use released doclayout 0.6.

• Use released zip-archive 0.5 (major performance improvements).

• Use released skylighting-0.15 (major performance improvements).

• Depend on released doctemplates 0.11.1.

• Depend on released typst 0.12.

• Require text >= 2.0.

• Bump upper bound for unicode-data.

• Allow crypton 2.0.x.

• Allow Diff 2.0.

• Add tools/diff-golden-tests.sh.

• Fix tools/diff-zip.sh on non-Darwin.

• Add tools/benchplot.js. This creates a nice graph comparing two benchmarks.

• typst-properties.md: fix fill syntax in Typst property examples (#11855, zenor0).

• Remove tested-with from cabal file. We tend not to keep it up to date.

• Fix typo in Lua filter example (#11875, Andonome).

• Fix a bug in jats-reader.xml (duplicate attribute)