[derek73]

Builds #753 sub-phase 3j, file-strict, as merged in #985 (decisions/vocabulary-model's "2026-10-30 #753, file-strict" entry and the RFC's "Strict: the chosen files and nothing else").

What it does

parser_for(..., strict=True) compiles only the chosen languages' vocabulary files, plus every native-script reading (a word with a letter and no Latin letter). Each tag selects its own file and its ancestors' files, as a plain choice does. mul.toml and und.toml are not among them. Strict applies as a delta against the default compile, as a plain choice does, through _reading_index.choose and compile_readings over strict_pool.

The four API points (decided 2026-10-31)

• (a) A strict call that chooses no language raises ValueError. The message names the packs and offers a pack that chooses a language, e.g. Locale(code='zh', lexicon=Lexicon.empty(), languages=frozenset({'zh'})). A test runs the offered code.
• (b) There is no public locales.choose.
• (c) _reading_index.strict_ungated derives the words strict ungates. It compares each strict choice with the plain choice of the same tags, not with the default's markers. STRICT_UNGATED in tests/v2/test_reading_index.py holds the result: 16 words, 27 word-and-tag pairs over 13 tags, exactly the decisions entry's list.
• (d) A strict choice accepts a tag whose file holds no reading (hu, kk, kok, vi, zu). A plain choice is still refused, and its message now says strict accepts the tag. To know these files, the reading index now records every file's tag (FILES in nameparser/config/_index.py, _reading_index.files()).

Other changes

• UNCHOOSABLE's comment and selected's refusal now say that a plain choice never removes a reading of mul/und and that a strict one drops both.

• The build now refuses prefer and strict as TOML keys, with a message saying that a strength is the caller's choice. The RFC had already decided this message. The old message said both were "reserved for phase 3". The vocab README table is updated to match.

• CLI:

  • --strict is new; without --locale it is a usage error (exit 2).
  • --locale falls back to a bare choice when the code is a language tag but not a pack (--locale ko). A code that is neither exits 2 and lists both the pack codes and the tags.
  • Under --strict, a pack that chooses no language uses its own code as the language (--locale zh --strict). --locale tr_az --strict exits 2 and suggests a tag to give (--locale tr --strict).
  • A pack's language with a region or script subtag (zh-CN, zh-Hant, es-MX, ja-JP, pt-PT) exits 2 naming the pack (use --locale zh), since a bare choice of it would silently drop the pack's segmentation, policy, list and segmenter. A tag spelled for a pack's code is that pack: pt-BR, PT-BR and pt_br all resolve to locales.PT_BR, and tr-AZ to locales.TR_AZ.

  These CLI conveniences were chosen in this PR for the debug helper and are recorded as such.

Docs

• docs/design:
  • The RFC's status lines, API bullets and 3j row.
  • A new decisions entry "2026-10-31 Phase 3: language packs that prefer or restrict their language #753, phase 3j", plus a supersession note on the file-strict entry's head and updated Standing/Open lines.
  • mechanisms.md#LOCALE-PACKS-PURE-DATA and AGENTS.md.
• docs/locales.rst has a new section, "Reading one language only", next to "Choosing languages", where choices are documented. Its worked example is romanized Korean. docs/usage.rst documents the --strict flag and links to that section.
• Release-log bullet. The default reading of Kim Do Hyun was checked against the 1.4.0 and 2.8.0 wheels.

Verification

• ORCID table: re-run on the built strict=True with the decisions entry's recipe. Every cell matches: 14 countries × 4 columns.

• Worked examples: checked on the built code.

  Name	Default	Strict ko
  Kim Do Hyun	family Do Hyun	middle Do, family Hyun
  Kim Bin Hyun	family Bin Hyun	middle Bin
  Ra Min Jun	title Ra	given Ra
  Kim Min Ra	suffix Ra	family Ra (suffix again under ko + en)

  Also: Gen Hoshino reads given Gen under strict ja, and De Xu reads given De, family Xu under strict zh. Every example in the 2026-10-30 entry still holds.

• No default parse moves: tests/v2/test_snapshot.py passes unchanged.

• Checks run:

  • uv run pytest: 34,198 passed.
  • uv run pytest -n0 -m serial, uv run mypy, uv run ruff check and tools/vocab/build.py --check: all pass.
  • The full suite also passes on Python 3.15, without namedivider (this venv) and with it.
  • Sphinx doctest and HTML builds are clean.

• Review: the design-docs reviewer found nothing of high severity; its low findings are fixed. A later independent review's items (the CLI subtag refusal above, the tr_az --strict remedy, the supersession marks on the 2026-10-30 entry, a mixed-script is_native_script row) and Codecov's patch coverage (the CLI error tests now run in process) are fixed too, and a second design-docs pass over them found nothing further after its two fixes.

Not in this PR: Vietnamese readings in vi.toml, nl prepositions, the 16 ungated words.

🤖 Generated with Claude Code