🐛 FIX: build-aborting crashes on hostile input, and docutils 0.23 compat

[chrisjsewell]

Fixes #1152, fixes #1031.

Parser defects where bad (or merely unlucky) input aborted the whole build instead of degrading gracefully, plus docutils 0.23 support. Each fix has a regression test, and the branch has been through an extensive multi-angle review (adversarial input fuzzing, old-vs-new behavior comparison for previously-building documents, performance benchmarks, docutils 0.20→0.23 compatibility verification against real wheels) — the second commit contains the hardening that review produced.

Crash fixes

Deeply nested HTML → RecursionError (html_to_nodes.py): rendering the parsed HTML AST recurses once per nesting level (including re-entrantly, via nested directive parses), and the existing try/except only covered tokenization — so a few hundred nested <div>s inside an html_admonition killed the build with an uncaught RecursionError. The conversion is now guarded: it warns (myst.html) and falls back to passing through the raw HTML. Verified byte-identical output for all non-pathological HTML.

Hostile YAML front matter (config/main.py, base.py, directives.py): read_topmatter and render_front_matter caught only ParserError/ScannerError. Three further classes escaped and aborted the build: ConstructorError/ComposerError (e.g. title: !Ref X), the bare ValueError PyYAML leaks for out-of-range timestamps (date: 2021-99-99), and RecursionError from deeply nested YAML (PyYAML's composer recurses per nesting level). All sites now share a YAML_LOAD_ERRORS tuple and emit the intended myst.topmatter warning; the directive-options YAML path (only reachable with validate_options=False, e.g. myst-nb) gets the same treatment.

YAML alias-expansion bomb (base.py): yaml.safe_load is immune to "billion laughs" (aliases stay shared), but serializing the loaded structure for the docinfo field list expanded it — ~20 lines of front matter hung the build indefinitely in json.dumps, and a self-referential alias raised ValueError. Field values are now measured first (identity-memoized, so linear time even for exponential graphs) and skipped with a warning when oversized or unserializable.

Footnote label colliding with a heading name (base.py): the duplicate-definition check used document.nameids, which holds every target name, so # Note + [^note]: ... silently discarded the footnote definition with a misleading warning and a dangling reference. The check now gates on document.nametypes: collisions with implicit names (headings — the bug) keep the footnote (explicit-over-implicit is standard docutils behaviour, INFO-level only; verified no change to anchors/{ref}/slug resolution), while collisions with explicit targets ((name)=, figure :name:, eval-rst-created footnotes) keep the existing drop-with-warning behaviour byte-for-byte — letting those through would strip the name from both nodes and break previously working references. Also O(1) per definition, where a naive per-definition scan measured O(n²) (+9.7s on a 10k-footnote document).

docutils 0.23 (#1152)

As reported by @befeleme, docutils 0.23 also attaches info-level (severity 1) system messages to the doctree, where previously they only went to the report stream — breaking test_link_resolution[explicit>implicit], which runs with report_level=0. This was the only 0.23 failure, and it is test-side, so:

• the test now drops severity<2 system_message nodes from the tree on docutils ≥0.23 before comparing (they are still asserted via the captured report stream)
• the pin is extended to docutils>=0.20,<0.24 (supersedes ⬆️ Update docutils requirement from <0.23,>=0.20 to >=0.20,<0.24 #1144, which dependabot should auto-close)
• 0.23 is added to the check-myst-docutils CI matrix, so the combination is exercised for real (the pytest matrix is unaffected: Sphinx itself still pins docutils <0.23, so 0.23 currently only reaches docutils-standalone users; worth a sphinx-side matrix entry once Sphinx lifts its cap)

Also included

• Cherry-pick of add link to auto_mode in sphinx-autodoc2 #1032 (thanks @nerdoc): docs link to auto_mode in sphinx-autodoc2, closing auto_mode is never explained / missing? #1031. Please keep the Co-authored-by trailer when squashing so authorship is preserved.

Full test suite passes locally (1116 passed, 11 skipped) on docutils 0.22.4, and the docutils-only test files pass on a real docutils 0.23.