Read script properties from one private table: Georgian has no case, Burmese and Thai write no initials (#1256) - #1377

Implements the decision on #1256: one private table of script properties, read by every rule that asks, with the two leaves it settles.

Closes #1222
Closes #1234
Part of #1256

"Part of" rather than "Closes": three of #1256's sub-issues stay open after this PR.

• What does an initial look like in Devanagari, Bengali and the scripts like them? (एम. एस. वर्मा gives initials() 'ए. ए. व.') #1273: what an initial looks like in the Brahmic scripts. This is the brahmic row's own open question.
• Should case repair keep O'tkir and o'g'li lowercase after the apostrophe, as Uzbek writes them? (O'TKIR SOBIROV gives O'Tkir) #1274: the Uzbek apostrophe in case repair, a language question.
• Should capitalized() write the Dutch IJ digraph under a Dutch choice? (PIETER IJZERMAN gives Pieter Ijzerman) #1242: the Dutch IJ, a language question that the decision keeps out of the table.

decisions/script-properties lists all three under Open:. #1214 (the Cyrillic apostrophe) and #1224 (a particle opening a hyphenated word) are not rows: they are questions of a language's spelling, which no script column could hold, so they stay in decisions/R4.

What changed

• The table (_policy._SCRIPT_PROPERTIES). Each row is keyed by a private name, never by Script, and has spans, has_case, writes_initials, period_abbreviations, spells_initials, a reason and a source. It replaces _policy._NO_INITIALS, _vocab._INDIC_BRAHMIC and Python's own case properties at every name-level case contrast. Readers build matchers once at import from the spans a column selects, so no reader adds a frame. The first commit moves no reading: snapshot.jsonl is byte-identical, and the frame band and serial benchmarks pass.

• The split. _NO_INITIALS answered two questions: "writes initials" (T5, R7, the name match) and "writes period abbreviations" (H2's shape, S3's split and joined shapes). Thai answers them differently, so each is now its own column. One reader, period_joined_vocab (S3's dotted credential), was asking the wrong one. It moved in its own commit, with no reading moved.

• Georgian (Should Georgian script count as caseless? (ჯეკ MA gives family '' and suffix 'MA', where محمد MA gives family 'MA') #1222). Georgian, Mkhedruli and Mtavruli, has no case for the name-case contrast. Six readers drop its letters with str.translate:

  • the one-case fact
  • the credential lean
  • R5's gate
  • S8's two name-word-behind predicates
  • R4's Mtavruli repair, which now reads the row and changes the same 46 letters over every codepoint

  T7's and R3's lower-case run of initials read the row in one commit and went back to Unicode after the design-docs review. That run asks whether a dotted run of single letters is initials, which Georgian answers yes in its one case, so ნ.ბ. ბერიძე reads given ნ., middle ბ. as on master.
  Both of the following leaves rest on #1256's amendment of 2026-11-12. Each is its own commit.

• Burmese (Should initials() skip Burmese script, as it skips Han and Hangul? (ကြည် ကြည် ဌေး gives 'က. က. ဌ.') #1234). The Myanmar script writes no initials, decided from the writing convention. CLDR's my block is en's: all 42 patterns are equal once spaces and commas are removed, and it holds {surname-prefix} though Burmese names have no surname. (72c6766)

• Thai. Thai writes no initials, decided by the CLDR count: th uses an initial in 0 of 20 locale-authored patterns. It does write period abbreviations, so ดร., ร.ต.อ. and พ.บ. read as before. The accepted cost is a pen name like ส. ศรีสุข: it keeps its fields, but initials() gives ''. (28f73e9)

What moved (snapshot, master 161ebd2 → head)

Local corpora, snapshot.py --local … --base master (2026-11-12):

• jrc-names: 565 spellings moved, 561 of them in initials() only (548 Thai, 13 Burmese). The other four are Georgian and match their Arabic twins on master: ალექსი II მოსკოვის პატრიარქი and კარლ XVI შვედეთი now capitalize the numeral as Ii and Xvi, and პეტრე I რუსეთი and ფრანსუა I საფრანგეთი now report conjunction-or-initial (see Accepted)
• nepal-candidates: matches the base.
• opensanctions-parliaments: matches the base.

A test from #1223 moved, deliberately. test_georgian_in_mtavruli_repairs_to_mkhedruli pinned Ნინო Ბურჯანაძე (a Mtavruli first letter on Mkhedruli) as kept on the default capitalized() path. With R5's gate reading the row, Georgian is one case to the gate, so the name is written in Mkhedruli, as force=True already wrote it. A Mtavruli surname beside a Mkhedruli given name (ნინო ᲑᲣᲠᲯᲐᲜᲐᲫᲔ) moves the same way. That differs from Latin, where Jean DUPONT is kept: Latin's capitals are case, while Mtavruli is a style of a caseless script. This is recorded as DECIDED in decisions/R4 and as an Accepted bullet in rules/R5. The test keeps a row for each shape with the new reading.

Recorded as Accepted

• Word-level tests stay on Unicode: a Mtavruli word is a word in capitals. O9 still reads ᲑᲣᲠᲯᲐᲜᲐᲫᲔ ნინო under capitals_as_family as family ᲑᲣᲠᲯᲐᲜᲐᲫᲔ, so it loses no reading. Under that switch ჯეკ ბერიძე MA keeps family MA, unchanged from master.

• A mid-name roman numeral in a Georgian name now repairs as it already did in an Arabic one (ალექსი II … → Ii, as محمد II علي حسن gives on master). That is a defect of caseless names in general, not of this row.

• The Myanmar row's period_abbreviations stays true, there being no evidence either way: ကက. ဌေး still reads title ကက..

• A Georgian dotted run followed by a credential that S8 reads as a name word stays whole. The credential's Latin capitals then count among the name words, which turns off T7's all-lower-case reading. So ნ.ბ. ბერიძე Jr. ნინო gives given ნ.ბ., middle ბერიძე Jr., family ნინო, where master gave given ნ., middle ბ., suffix Jr. ნინო. ნ.ბ. MA keeps its run whole. Master already reads a run whole beside a Hebrew or Arabic word (ნ.ბ. בן ბერიძე). Recorded in rules/T7's Accepted bullet.

• A listed Georgian salutation behind a credential is the family name: ჯეკ ბერიძე MA ბატონი gives family ბატონი (master: title ბატონი, suffix MA); ქალბატონი and ქ-ნი read the same way. Its Hebrew and Latin Title-case twins already read this way on master (משה כהן MA מר, Jack Smith MA Mr), and behind Jr. it stays the title. Recorded in rules/S8's Accepted bullet.

Docs

• rules/S6, R5 (an Except bullet and an Accepted bullet), R7 and H2's Accepted clause.
• A new decisions/script-properties, with each row's evidence, the CLDR recipe with its copy check, and a dated table. Entries in decisions/R4, R5, R7, S6, H2, cjk-full-stops and initials-repertoire.
• mechanisms.md#SCRIPT-PROPERTY-TABLE.
• nameparser/AGENTS.md: a sweep for a split column.
• Release-log bullets for Georgian (with the capitalized() move), Burmese and Thai.

Swept against master: every comma shape, the three orders, every policy switch, every locale pack, NFC and NFD, Parser.revise, both sides of matching().check, and Parser.suggest. 11 mutations of the changed conditions are each caught by a test, and so are the code review's row-value mutations.

Moves under the packs and the suggester, all recorded in decisions/script-properties:

• Under the two-family-names packs (locales.CS, SK, ES, ES_AR), a Thai or Burmese letter with a period between two names is now the first family word rather than a middle initial: Nino ก. Beridze gives family ก. Beridze, as John 王. Smith already did on master. This goes through rules/O6.
• Under locales.NE (initials_as_family), such letters are no run, through rules/O8.
• A lone က. or ก. now reports given-or-family, as any lone word does.
• A Burmese or Thai word before a Latin letter drops the order report (က V, สมชาย V).
• Parser.suggest no longer suggests FAMILY_FIRST for a group of Thai or Burmese names, each a word followed by a letter. That is recorded in decisions/group-suggestion. suggest_measure.py --base master: 0 cells moved.

Commit e8e200b moved tags. It is "no reading moved" against master, not against its parent: 72c6766 had dropped shape:acronym and vocab:suffix-ambiguous from က.က. in က.က. က, and e8e200b gives them back. No field moved.

Mutations, on the final head (bb606fd)

Each changed condition was mutated alone, and the test-file set below was run against it (test_script_properties, test_render, test_match, pipeline/, test_cases, test_rules_doc). All 22 are killed:

Myanmar Extended-C (U+116D0–116FF) is the one span no test pins: it holds only digits as of Unicode 17. It stays in the row so that a letter assigned there later reads like the script's other letters, and it moves no reading today.

Checks

• uv run pytest, -n0 -m serial, mypy, ruff, sphinx-build -b doctest (874, 0 failures), python -m doctest README.rst, tools/vocab/build.py --check
• the suite on Python 3.15 in its own environment
• the suite without namedivider-python, and with it

Review aspects run

• code (pr-review-toolkit:code-reviewer): two findings, fixed in 1226317
• comments (comment-analyzer): eight findings, fixed in b212787
• design docs (design-docs-reviewer, all 12 axes): four findings, fixed in 1226317
• independent docs review (from the coordinator): A1–A3 and B1–B13, fixed in b9db338. This includes reverting the T7 and R3 run sites, which had lost a reading
• independent code review (from the coordinator): C1–C8, fixed in 3720c4a. C1 was already fixed by b9db338's revert, so no Arabic reading moves. The other fixes:
  • packs and the suggester swept and recorded
  • the table and the caseless map are read-only
  • has_case is typed Literal[False] | None
  • an empty column fails the import by name
  • row values and spans are pinned by readings and by Unicode letter names
  • the Script enum is bound to the table by span
• tests (pr-test-analyzer, with 9 single-site mutations): six findings, fixed in 840528e. These are a widened source scan with a planted control, R4's repair added to the per-row test, Thai guards for is_leading_title and split_pair, the Georgian comma shape, and an in-suite negative control (11 twin rows apart with the caseless table emptied)

🤖 Generated with Claude Code