[derek73]

Summary

The package's licence expression read LGPL-2.1-or-later alone, while the wheel ships data derived from five third-party datasets under other licences. Two of those licences ask their text or notice to travel with copies, and the tree held neither text.

• Licence expression: LGPL-2.1-or-later AND Apache-2.0 AND GFDL-1.2-no-invariants-or-later AND CC-BY-4.0 AND OGDL-Taiwan-1.0 AND Unicode-3.0. packaging canonicalizes it unchanged, and the wheel and sdist carry it as License-Expression (Metadata-Version 2.4).

• LICENSE-nam_dict (new): nam_dict's copyright and licence notices as the file states them, then the GFDL 1.2 text. It covers the 3.1 familiar forms and the suggester's common given names.

• LICENSE-unicode (new): the Unicode License v3 notice. It covers locales/zh.py's simplified surname forms, which have shipped since 2.5.0: 19 of the block's 324 forms come only from Unihan's variant table. It also covers the 3.1 suggester's pinyin surnames. zh.py's docstring now credits Unihan.

• README.rst's License section lists every source and where its terms are met. This is the PyPI page.

• tests/v2/test_licensing.py holds the licence files, license-files, the expression, the README and the modules crediting the attribution-only sources (CC BY 4.0, OGDL) to one another. Checked against deliberate breakages: dropping LICENSE-unicode from license-files, CC-BY-4.0 from the expression, and zh.py's README credit each fail it.

• decisions/shipped-data-licences (new key) records:

  • each source, its first release, what its licence asks and how that is met;
  • what is left out and why: CC0 sources, corpora used only for counts, and census lists with no stated terms;
  • an open question: SUFFIX_ACRONYMS' Wikipedia origin;
  • relicensing to LGPL-3.0, weighed and not taken.

  AGENTS.md's Extension Patterns now says what a new data source needs.

The code's licence is unchanged. Releases 2.5.0 through 3.0.1 shipped OGDL and Unihan data, and 3.0.0 onward CC BY data, behind the LGPL-only expression. Debian's debian/copyright will need a paragraph per source when 3.1 is packaged.

The entry also records the rule for future sources (FACTS WE MEASURE): data the project derives by measuring (its own counts, cutoffs and filters) is its own, with sources credited as a courtesy. A curated list copied whole, one compiler's judgements of which entries belong, keeps its licence, and if that licence is a burden it goes in an optional extra package instead of the core. Whether the attribution-only terms stay in the expression under that rule is recorded as Open; they stay for now.

Follow-ups, each its own PR:

• replace Unihan with Wikidata's pinyin;
• derive the suggester's common given names from ORCID instead of nam_dict;
• drop nam_dict's familiar-form pairs before 3.1, keeping the ODU list's;
• rebuild the non-English familiar forms from CC0 data (ORCID, GND), reviewed by hand.

Each removal shrinks the expression and the files again, and the test flags any of them left behind.

Review

• design-docs-reviewer on the decisions entry, the AGENTS.md paragraph and the licence-file headers. It found:

  • Unihan's 2.5.0 shipping in zh.py;
  • the census lists and the Wikipedia import missing from the entry;
  • a test claim wider than the test;
  • a per-release overstatement;
  • the "rest of nameparser" sentences in the two licence files.

  All are fixed in this commit. The missing PR number is added in a follow-up commit.

Test plan

• uv run pytest: 46,722 passed, 38 skipped, 3 xfailed
• uv run ruff check, and uv run mypy on the new test
• sphinx-build -b doctest docs, python -m doctest README.rst, and readme_renderer on README.rst
• uv build: the wheel's dist-info/licenses/ and the sdist carry LICENSE, AUTHORS, LICENSE-nicknames, LICENSE-nam_dict and LICENSE-unicode

🤖 Generated with Claude Code