✅ test(dom): validate source locations against parse5

[gaborbernat]

parse(source_locations=True) ships as turbohtml's take on parse5's sourceCodeLocationInfo, so the spans it stamps only mean something if they land where parse5's do. 🎯 This validates them against the reference implementation on parse5's own location-info corpus: the five real-world documents (CERN, a DX portal, the parse5 GitHub page, the WHATWG HTML spec, and a MediaWiki article) that parse5's suite parses to exercise its location code.

parse5 is JavaScript, so a committed Node runner (tools/bench/node/parse5_location_runner.js, reusing the existing parse5 competitor setup) emits every element's start-tag, end-tag, and per-attribute spans, and a pytest harness parses the same newline-normalized input through turbohtml and compares span-for-span. Both sides key elements by start-tag start offset and normalize to turbohtml's convention (parse5's columns are 1-based, its offsets and lines match), and turbohtml parses with scripting=True to line the trees up with parse5's default. The corpus is a pinned shallow submodule under tests/conformance/parse5; the harness skips when Node or the submodule is absent, the way the *_differential.py oracles do, so it stays out of the coverage gate. All 9767 elements agree on all three span kinds.

The corpus surfaced a real bug. 🐛 A body, html, or form element that the source closes with its own tag left source_location.end_tag at None: these three close through an insertion-mode switch or an out-of-order stack removal, never the normal pop that stamps the end-tag span, so an explicit </body>, </html>, or </form> went unrecorded. A record_end_tag_location helper now stamps them from those handlers; the start-tag and per-attribute spans were already byte-exact against parse5 before the fix.

Validates #548.

[codspeed-hq]

Merging this PR will not alter performance

✅ 83 untouched benchmarks
⏩ 18 skipped benchmarks¹

---

_{Comparing gaborbernat:test/548-parse5-location-conformance (d156093) with main (4ab5e35)}

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. ↩