Preserve Vietnamese Unicode in relevance and anchor matching

[bongsunflower01-boop]

Summary

ASCII-only term extraction splits Vietnamese words at accented letters and treats canonically equivalent Unicode differently. Preserve Unicode terms and NFC/NFD equivalence in relevance and anchor matching while retaining source text and technical identifiers.

Evidence / behavior

• Normalize comparison text with NFC and casefold; keep Vietnamese tone distinctions, attached combining marks and meaningful non-ASCII single-character syllables.
• Add a conservative Vietnamese stopword list while preserving negation and the existing English/technical-token behavior.
• Match equivalent NFC/NFD anchors at the original source offsets without applying English plural/stemming rules to accented words or product identifiers.
• Document remaining lexical-analysis and shared-stopword limitations. Fallback placement copy remains unchanged.

Testing

• Add 15 deterministic Unicode regression tests, including Vietnamese tones, NFC/NFD, combining marks, English compatibility, negation and technical IDs.
• python -m unittest discover -s tests -v passes: 81/81 tests on Windows / Python 3.12.14.
• Package compile, CLI version/status smoke checks and an offline Unicode reproducer pass on the same checkout.
• Python 3.11/3.13 were not run locally; the existing upstream CI matrix remains responsible for those environments.
• No live-site dependency was added to deterministic tests.

Safety and methodology

• No URL fetching or network behavior changed.
• Evidence fields and confidence workflow remain intact.
• Documentation explains behavioral changes and remaining limits without making search-ranking claims.
• No secrets or confidential datasets are included.