Should a zero-width space separate words, so ​王小明 reads as 王小明 does? #851

A zero-width space (U+200B) carries no name content, but the tokenizer keeps it inside the word it touches. So an invisible character changes the reading.

>>> parse("王小明")            # family='王小明'
>>> parse("​王小明")      # given='​王小明'
>>> parse("​Dr. John Smith")
# given='​Dr.', middle='John', family='Smith'   (the title is missed)
>>> parse("John​Smith")
# given='John​Smith'                            (one word)

Under parser_for(locales.ZH), ​王小明 stays whole in given too: the name is no longer wholly Han, so neither the family-first script order nor segmentation applies.

Where it was found

The tw-donors local corpus (#844, fetched 2026-10-22, data dated 2026-10-05): 520 of its 308,667 distinct names open with U+200B, and each reads given where the same name without it reads family. 409 of those 520 also appear in the corpus without the character, so the same donors read two ways. None of the other local corpora (ctgov-officials, fec-individuals, ukec-individuals, psc-individuals, tse-donors) and none of the committed corpora holds a U+200B.

What the tokenizer does today

rules/T1 (separators, not joiners) makes "an ignorable character, one that carries no name content", a separator. The set it implements (_BIDI and _EMOJI_RANGES in nameparser/_pipeline/_tokenize.py) is the bidi controls (U+061C, U+200E, U+200F, U+202A to U+202E, U+2066 to U+2069) and emoji. U+200B is not in it, and Python's str.isspace() is False for it, so it reaches the token as a letter would.

Proposal, in rule shape

When a name holds a zero-width space (U+200B),

Then the name splits there, as at whitespace: the character is in no token, field or rendered view, and the original string keeps it.

Because U+200B carries no name content, and Unicode defines it as a word boundary with no width. Thai, Khmer and Burmese text write it between words where there is no space. That is T1's case, so it belongs in T1's set.

"​王小明"         →  family="王小明"
"​Dr. John Smith"  →  title="Dr."  given="John"  family="Smith"
"John​Smith"       →  given="John"  family="Smith"
"王​小明"          →  given="小明"  family="王"   (two wholly-Han words, read family-first, as "王‏小明" does today)

Excluded (stays inside the word, as today):

• U+200C ZERO WIDTH NON-JOINER and U+200D ZERO WIDTH JOINER carry content: they decide how the letters around them join. The committed corpora hold them inside Malayalam, Kannada, Sinhala and Arabic-script (Persian) words (corpus_cldr.jsonl and corpus_wikidata_native.jsonl, 2026-10-22). Splitting there would cut real names in two.
• U+2060 WORD JOINER means "no break here". Splitting at it reverses its meaning.

Open questions

1. Under which switch? Unconditional, like whitespace and the nakaguro (rules/T2's precedent), or behind a Policy opt-out like strip_bidi? strip_bidi itself is the wrong switch, since U+200B is not a direction control.
2. U+FEFF, the byte-order mark, also called ZERO WIDTH NO-BREAK SPACE: as a no-break space it is a joiner, but in names it nearly always arrives as a stray BOM at the start of a CSV field. No corpus held one on 2026-10-22. Strip it at a word's edge only, or leave it out?
3. Separator or transparent at the interpunct? T3's flank guard reads through the stripped bidi marks (_flank) but not through whitespace. Should 威廉·​莎士比亚 behave like the bidi case or like a space? Today 威廉·‏莎士比亚 reads given 威廉, family 莎士比亚, as 威廉·莎士比亚 does, while 威廉· 莎士比亚 reads given 威廉·, family 莎士比亚.
4. U+00AD SOFT HYPHEN is out of scope. It sits inside words, so separating there would split them, and T1 never joins. 20 names in the tse-donors local corpus held one inside a word on 2026-10-22, read today as one word with the hyphen kept, which may already be right.

What it would move

No committed corpus holds a U+200B, so tests/v2/test_snapshot.py and the CLDR/GND agreement records would not move. The local tw-donors snapshot would move its 520 names, from given to family.