Baixar Pandoc 3.12 per a Windows

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
ewif 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.