✅ test(clean): validate sanitizer against DOMPurify suite

[gaborbernat]

The sanitizer's shipped DOMPurify-parity features (SAFE_FOR_TEMPLATES, SANITIZE_NAMED_PROPS, CUSTOM_ELEMENT_HANDLING, USE_PROFILES) were covered only by turbohtml's own hand-written expectations. Green coverage is not the same as spec-correctness, so this validates them against the authority: DOMPurify's own test corpus and a live DOMPurify build.

DOMPurify is vendored as a pinned shallow submodule at tests/conformance/DOMPurify, and tests/conformance/test_sanitizer_dompurify_conformance.py runs every payload in its test/fixtures/expect.mjs (~219 XSS vectors and config cases) through turbohtml.clean.sanitize. 🔐 The oracle is DOMPurify's actual output: a batch Node runner (tools/bench/node/dompurify_conformance_runner.js, reusing the existing isomorphic-dompurify dependency) sanitizes the same inputs under the config that mirrors each Policy feature, so the comparison is head-to-head rather than against a frozen string. The absolute security invariant is asserted hard over the whole corpus under every config: turbohtml's output reparses to an inert tree, and it never keeps an executable construct DOMPurify strips. Because DOMPurify keeps nearly all of HTML by default while a Policy is an explicit allowlist, byte-parity over arbitrary payloads is dominated by allowlist scope rather than the feature under test, so each feature is also validated for exact canonical parity on curated inputs whose tags both allowlists share, where the only variable is the config flag.

Two divergences are documented and xfailed rather than hidden, both cases where turbohtml is stricter and so never less safe: the template stripper collapses only marker runs that can open a template evaluation ({{/${/<%), leaving a bare unmatched close delimiter as inert text, while DOMPurify's greedy regex also eats the text before it; and stripping a disallowed HTML wrapper drops a nested foreign (SVG/MathML) subtree instead of hoisting it. ✨ No sanitize.c change was needed: the corpus found no case where turbohtml is less safe or diverges on a feature it claims to support. The harness carries no coverage (it is in run.omit, like the other differential suites) and skips cleanly when the submodule or Node build is absent.

validates #527 #531 #533 #534 #536

[codspeed-hq]

Merging this PR will not alter performance

✅ 83 untouched benchmarks
⏩ 18 skipped benchmarks¹

---

_{Comparing gaborbernat:test/sanitizer-dompurify-conformance (69e22ea) with main (db0a1e6)}

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. ↩