✅ test(xpath): validate fn:* functions against W3C QT3 suite

[gaborbernat]

Feature #558 shipped the XPath 2.0 string functions (ends-with, string-join, lower-case, upper-case, matches, replace, and the rest of the ported subset), validated so far only by turbohtml's own tests and a differential against lxml. Coverage-green is not the same as spec-correct, so this adds the authoritative check: the W3C QT3 (XQuery/XPath 3.1) test suite, the standards body's own conformance oracle, run against turbohtml's engine.

The suite is vendored as a pinned shallow submodule under tests/conformance/qt3tests, following the tests/html5lib-tests convention. tests/conformance/test_xpath_qt3_conformance.py parses the QT3 catalog and, for each fn:* case, decides whether the query lives inside turbohtml's supported grammar. 🎯 An in-scope case runs against the case's asserted result (assert-true/assert-eq/assert-string-value/error) and fails on any mismatch. Everything turbohtml omits by design is skipped or xfailed with a per-case reason carried in the parametrize id, so the signal stays on what the engine claims to support rather than being diluted by XQuery and schema machinery.

The excluded families each map to a documented boundary: XQuery-only tests (gated on the catalog's own spec dependency), xs: schema constructors and casts, XPath 2.0 sequences, ranges, and value-comparison operators, XPath 3.x maps/arrays/arrows and higher-order functions, the XSD regex dialect (turbohtml's fn:matches/fn:replace run on Python's re, so \p/\P categories, class subtraction, the q flag, and the FORX error codes fall outside it), the typed error-code taxonomy (turbohtml raises Python exceptions), and XML namespace decomposition (turbohtml is HTML-first, so local-name equals name and namespace-uri is empty outside SVG/MathML).

Of the 2631 cases across the 33 supported families, 281 run in scope and all pass, 49 xpass where turbohtml still rejects invalid input, and the remaining 2301 are recorded with a per-case reason: skipped for unsupported grammar, xfailed for the error-code taxonomy. 🔍 No engine bug surfaced: every failure traced to one of the boundaries above, which the explanation doc now records. The one adjacent finding is that parse_xml rejects a leading BOM (auction.xml), so those context cases skip as unparseable environments; that sits in the XML tree builder, not the XPath engine.

The conformance module is omitted from the coverage gate like the existing *_differential.py oracles and skips when the submodule is absent, so it never drops the free-threaded, 3.15, or Windows coverage runs. Excluded cases use skip rather than xfail(run=False) because the latter is a pytest performance pathology at this scale.

validates #558

[codspeed-hq]

Merging this PR will not alter performance

✅ 83 untouched benchmarks
⏩ 18 skipped benchmarks¹

---

_{Comparing gaborbernat:test/558-xpath-qt3-conformance (62fc8ad) with main (95dfb19)}

Footnotes

1. 18 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩