Apply strip_document to BeautifulSoup Tag roots

[ryanduguid]

Summary

Apply strip_document to completed BeautifulSoup Tag conversion, addressing the boundary-newline asymmetry reported in #209. Reuse the existing newline-only policy after normal root processing when cached [document] conversion is available.

process the supplied root normally
if it is a non-document Tag and document conversion is available:
    apply the built-in strip_document policy to the result

The added finalisation does not invoke the resolved document callback. Normal tag dispatch and document callback ownership remain in place; the README describes the proposed Tag contract.

Evidence

• Before: Of 73 new parametrised cases, 43 fail and 30 pass against the original runtime, with no collection errors. The reported Tag root retains boundary newlines.
  After: All 156 tests pass on Linux Python 3.8.20 and Windows Python 3.14.8. Coverage includes exclusions, custom callbacks, cached None, method aliases, significant whitespace and original tree relationships.
• Configured Linux tox, python -m build -nwsx ., mypy . and mypy --strict tests/types.py pass. Windows pytest, configured flake8 and README lint pass.
• All 73 new cases pass at BeautifulSoup 4.9.0/six 1.15.0 and against the built wheel installed outside the source checkout. The build retains the existing version-of-None warning.

Merge danger

Door: two-way. Blast radius: fragment consumers.

This proposes a compatibility change: Tag boundaries, including custom converter output, now follow strip_document; an invalid mode raises after Tag processing when document conversion is available. The extra cached document lookup, including None, can change later document behaviour with stateful resolvers. strip_document=None preserves boundary text but still performs that lookup. Maintainer acceptance of these semantics remains necessary.

Unverified

The full dependency-floor suite cannot collect the existing test_escaping.py warning import on both the base and initial candidate. Other Python/Windows versions, macOS, alternative parsers, Windows packaging/types, independent execution of the verification producers, additional exception-precedence/custom-unavailable-invalid-mode probes, scanner parser coverage, hosted CI and downstream reliance remain unverified. No performance improvement is claimed.