1.6.0

Security

• Long runs of single-letter roman suffix tokens no longer cost quadratic CPU. A comma, salutation, comma-credential, or surname-first name followed by many I, V, or X tokens (Smith, V V V ...) rescanned every earlier token for each candidate, so a 16 KB row took about 20 s of CPU and a row near the 1 MiB limit could take many minutes. The suffix scan now finds the preceding name token once per pass, and a 16 KB row takes about 0.03 s. Runs of parenthesized (MA) credentials were also quadratic, at a lower constant (about 1.5 s for 16 KB), and are fixed too. Earlier notes that described hostile-input parsing as linear did not hold for these shapes.
• setWhitespace() now throws InvalidArgumentException for a whitespace set that is not valid UTF-8, instead of falling back to byte-level collapsing. With such a set, such as " \xA0" written for a no-break space, collapsing split multibyte characters and control stripping was skipped, so escape, bell, and bidi override characters and invalid UTF-8 reached name fields and exports. parse(), Confidence::assess(), and the Name constructor apply the same check, and control stripping now always runs on scrubbed input, so output is always valid UTF-8.
• Normalization strips every Unicode format character (category Cf), not only a fixed list of bidi and zero-width characters, so the soft hyphen, tag characters (which can carry hidden text), and interlinear annotation marks no longer reach name fields. The line and paragraph separators U+2028 and U+2029 become spaces instead of passing through inside a token. Variation selectors are kept.
• An empty whitespace set (setWhitespace('')) no longer skips control stripping. C0, C1, and bidi control characters are removed as with any other set; tab is kept in the text.

Fixed

• With many custom nickname delimiter pairs, a row near the 65,536-token limit no longer exhausts memory. The suffix scan kept one array slot per pair per token (about 75 MB peak for 32 pairs); it now keeps compact state only for pairs that open a span, so the same row peaks at about 19 MB and fits a 64 MB memory_limit.
• Confidence counts names the way the parse maps them, so a name-colliding title followed by one surname is flagged even when a single-letter suffix follows (Lord Ashcroft V), the title is wrapped ("Lord" Ashcroft, (Lord) Ashcroft), another name-colliding title follows (Lord Pastor Smith), or a stray connector trails (Lord Ashcroft and). Lord Junior is flagged for the same reason, while "Lord", Smith still reads Lord as the surname and stays unflagged. An honorific run that joins a connector or an unambiguous title (Lord and Lady Smith, Lord Rev. Smith) shows the lead is a title and is not flagged.
• An internal apostrophe no longer hides a trailing credential: John D'Angelo MD' keeps D'Angelo as the last name and MD as the suffix.
• A single-letter roman token after a given name in a comma segment stays an initial: Smith, Michelle, I gives first name Michelle and initial I.
• Replacement nickname delimiters keep literal quotes in the nickname: with setNicknameDelimiters(['[' => ']']), John ['Twas] Smith gives nickname 'Twas.
• Name::getConfidence() tokenizes the way the parser does, so NUL whitespace and nickname-shielded commas no longer split or over-count tokens, and a parsed row with many shielded commas no longer throws LengthException.
• Nicknames between surname particles no longer invent a given name: van (Bob) der Berg keeps the complete surname van der Berg.
• Combined initials no longer split digits, punctuation, or caseless characters out of a token, so Smith, A1 John keeps A1 as the first name.
• Confidence tokenization, delimiter handling, and comma credential tails scan each token a bounded number of times instead of rescanning earlier tokens.
• Nickname delimiters made of Unicode format characters are ignored like whitespace and control delimiters, since normalization strips those characters and the pair could never match.