v1.1.2

Fixed

• (a) or (b) failed with Unexpected token FUNCTION (#9) - the tokenizer classified or, and, div and mod followed by ( as function calls before applying the XPath 1.0 operator disambiguation rule (section 3.7). The rules now run in the order the specification lists them.
• xsl:sort ordering - text keys were upper-cased and compared with localeCompare, so §112 sorted before 100-00 and case-order folded the whole string. Text now compares by Unicode code point like libxslt (the engine of Chrome's native XSLTProcessor); lang or case-order switch to an Intl.Collator, where case-order only breaks ties.
• xsl:sort data-type="number" - non-numeric keys were treated as 0 and parseFloat accepted 12abc. Keys now use XPath number() and NaN sorts before every number, as section 10 requires.
• xsl:sort attribute value templates - order, data-type, case-order and lang were read literally, so order="{$sortOrder}" never took effect.
• match="/" also matched the document element, so the classic catalog stylesheet (/ template wrapping apply-templates, plus match="catalog") rendered its wrapper twice. / now matches only the root node (XSLT 5.2).
• Multi-step patterns never matched: c/d, */d, r//d, /r/c, c/d[2], e/@a, id('x')/d, key('k','v')//d. Patterns are now compiled once and matched right to left (new src/xslt/patterns.js, src/xslt/patternCompiler.js).
• Template conflicts: with equal priority and import precedence the last template now wins, like libxslt (XSLT 5.5 recovery).
• Diamond imports (two imported stylesheets importing the same third one) were rejected as circular; only real cycles are errors now. xsl:call-template now honours import precedence.
• Default output method: a result whose root element is <html> is serialized as HTML when the stylesheet has no xsl:output method (XSLT 16).
• Simplified stylesheets (<html xsl:version="1.0">) dropped their literal root element.
• Keys: several xsl:key elements with the same name now all feed the index, key() with several values returns document order without duplicates, and the index is rebuilt for every transformation instead of going stale after DOM changes.
• != on node-sets is existential (@n != 1 is true when some @n differs), and a node-set compared with a boolean uses boolean() (XPath 3.4).
• XML whitespace is only space, tab, CR and LF: normalize-space() and number() no longer treat a non-breaking space as whitespace.
• number() follows the XPath Number grammar: 1e3, 0x10, +5 and Infinity are NaN; 5. and .5 are numbers.
• string() of numbers never uses exponent notation (1e21 gives 1000000000000000000000).
• Prefixed function calls such as exsl:node-set($rtf) failed to parse. EXSLT exsl:node-set() and msxsl:node-set() are now available.
• xml: prefix is predeclared, so @xml:lang works without a declaration.
• @* no longer includes xmlns namespace declarations.
• Adjacent text and CDATA form one text node in the XPath data model (<r>a<![CDATA[b]]>c</r>: one text() with value abc).
• string-length(), substring(), translate() count characters, not UTF-16 code units.
• Template parameters: a template's xsl:param default was overridden by any same-named parameter in scope (the caller's own params, global params, or undeclared xsl:with-params). A param now takes a same-named with-param from its direct caller, otherwise its own default (XSLT 11.6).
• Variable scoping was dynamic: a called template saw its caller's local variables, and variables declared inside xsl:if, xsl:choose or xsl:for-each leaked out. Scoping is now lexical (XSLT 11.5).
• Global variables and params may reference ones declared later (XSLT 11.4); circular definitions report a clear error.
• xsl:copy-of select="/" copied nothing, and copying attribute nodes produced an empty text node instead of the attribute. Copied elements and attributes keep their namespaces.
• Result namespaces: namespace prefixes declared on xsl:template or literal result elements were ignored in XPath expressions; xsl:element ignored the default namespace in scope (<p xmlns=""/> inside XHTML); xsl:attribute name="xl:href" and namespace= lost the namespace; literal result elements did not carry their in-scope namespace declarations.
•   in stylesheets was stripped as whitespace (<td>&#160;</td> became <td/>).
• Attribute value templates containing { or } inside a string literal ({concat('{', 'x')}) were split incorrectly.
• xsl:number: grouping-separator and grouping-size were ignored, and format was not an attribute value template.
• transformToDocument() with method="text" returned null; it now returns <html><head/><body><pre>…</pre></body></html> like Chrome.
• xsl:attribute after child nodes is ignored with a warning, as libxslt does, instead of being added.
• Serialization: comments containing -- and processing instructions containing ?> are made well-formed instead of producing broken XML or throwing.
• Deep recursion: a recursive named template overflowed the stack after about 700 levels; it now reaches about 1,200 levels (frames per level cut from 11 to 5) and reports Template recursion too deep instead of leaking a RangeError.
• Long XPath expressions such as a 120-term sum failed with Maximum recursion depth exceeded (100) inside stylesheets; transformations now allow 1000 levels (XSLT_MAX_EXPRESSION_DEPTH).
• Reverse axes in predicates - preceding-sibling::*[1] and preceding::*[1] selected the farthest node instead of the nearest one. Reverse axes now use proximity positions (XPath 1.0 section 2.4), while node-sets are still returned in document order.
• Axes from attribute nodes - parent::, ancestor::, following:: and preceding:: returned nothing for an attribute context node.
• Transformations of documents with more than 10,000 matching nodes returned null (Result set exceeds maximum size). The standalone XPath API keeps its 10,000 guard for untrusted expressions; XsltEngine now allows 5,000,000 nodes per step (XSLT_MAX_RESULT_SIZE, configurable with the maxResultSize engine option).
• The xslt command crashed for every npm user with ERR_MODULE_NOT_FOUND: jsdom: the CLI imported jsdom, which was only a devDependency. jsdom is now an optional peer dependency, loaded on demand; when it is missing the CLI prints how to install it. The library itself keeps zero runtime dependencies.
• The CLI used the private processor._engine field instead of the public engine getter.
• CLI xsl:include, xsl:import and document() did not work: no loaders and no base URI were set. They now resolve relative to the referencing stylesheet, confined to the base directory; http:/https: URIs are refused, and a document() that cannot be loaded yields an empty node-set with a one-line warning.
• CLI input decoding always assumed UTF-8, so an ISO-8859-1 document came out as caf�. Files are now decoded per XML 1.0 Appendix F: byte order mark, then the XML declaration's encoding, else UTF-8 (new bin/lib/decode.js).

Performance

• Trimming XML whitespace in normalize-space() and number() used a regular expression that backtracks quadratically on long whitespace runs; it is now a linear scan (400,000 spaces: 4 ms).
• Template matching is linear: apply-templates over 8,000 children with match="item[@id]" dropped from 90 s to 0.1 s, Muenchian grouping over 8,000 items from 8.5 s to 0.3 s, key lookups from 7 s to 0.2 s.
• The transformation from issue #9 (3.4 MB of HTML output) dropped from 28 s to 8 s. Axes walk firstChild/nextSibling instead of indexing jsdom NodeLists (each index access crosses a Proxy), step results are merged without quadratic concat/unshift, sort keys are computed once per node instead of once per comparison, and name tests only read namespaceURI and the document content type when the result depends on them.

Added

• src/xpath/axes.js (axis traversal) and src/xslt/sort.js (xsl:sort), each with a full test suite, plus src/regressions.test.js for reported issues.
• XSLT_MAX_RESULT_SIZE and XSLT_MAX_EXPRESSION_DEPTH exports and the maxResultSize / maxRecursionDepth / documentLoader engine options in the TypeScript declarations.
• New modules: src/xpath/strings.js, src/xslt/{patterns,patternCompiler,variables,stylesheetNamespaces,resultNamespaces,copying,avt}.js, bin/lib/{decode,loaders,output}.js.

Changed

• CLI stdout no longer gets an extra trailing newline when piped or redirected, so it is byte-identical to -o; a newline is only added for an interactive terminal.
• Stylesheets without xsl:output method whose result root is <html> are now serialized as HTML (no XML declaration, <br>); engine.outputSettings.method is null unless declared.
• Behaviour that relied on the fixed bugs changes accordingly: / no longer matches the document element, equal-priority templates pick the last one, number('1e3') is NaN, @* skips xmlns declarations.
• Templates no longer see their caller's local variables, and variables declared inside xsl:if/xsl:choose/xsl:for-each are not visible after them; stylesheets that relied on this now fail with Undefined variable, as they do in libxslt.
• Namespace declarations in scope in the stylesheet now appear on result elements (list them in exclude-result-prefixes to suppress them), matching libxslt.
• XsltContext gained globals and xpathVariables; the internal processElement/processXsltElement methods were replaced by an instruction dispatch table.
• engine.keys[name] is now an array of { match, use } definitions.
• xsl:sort without lang/case-order orders B before a (code point order), matching Chrome's native XSLTProcessor; previously the order was case-insensitive. Add lang="en" to get locale collation.

Dependencies

• eslint 10.10.0 -> 10.11.0, prettier 3.9.6 -> 3.9.9, jsdom 29.1.1 -> 30.1.1 (dev; jsdom is also an optional peer dependency >=25.0.0 for the CLI).
• GitHub Actions: docker/setup-buildx-action 4.4.1; the Pages workflow moved from checkout@v4, configure-pages@v5, upload-pages-artifact@v3, deploy-pages@v4 to the current releases, all pinned to commit SHAs, with FORCE_JAVASCRIPT_ACTIONS_TO_NODE24.