nameparser/unicode.rs
1// SPDX-License-Identifier: Apache-2.0
2//! Java `org.gbif.nameparser.util.UnicodeUtils` — ported subset: `normalizeQuotes`, plus
3//! the homoglyph-replacement table (`HOMOGLYHPS`/`containsHomoglyphs`/`replaceHomoglyphs`)
4//! used by `pipeline::stripandstash`'s step 13. `decompose`/`foldToAscii`/
5//! `replaceSpecialCases`/`removeNonAscii`/`replaceNonAscii`/`unescapeUnicodeChars`/
6//! `decodeUtf8Garbage`/`containsDiacritics`/`findDiacritics` are NOT ported: no call site
7//! anywhere in this crate's (now complete) pipeline reaches them, and the full-corpus
8//! Rust-vs-Java parity cross-validation (see `BENCHMARKS.md`/`cross-validation.md`, 0
9//! mismatches) confirms their absence isn't a gap.
10
11/// Unicode apostrophe / single-quote variants that normalise to the ASCII apostrophe `'`.
12/// Verbatim transcription of Java `UnicodeUtils.SINGLE_QUOTES`
13/// (`"`´ʹʻʼʽˊˋ‘’‚‛′‵❛❜'"`).
14/// Every codepoint is a 4-hex-digit Java `\u` escape, i.e. a single UTF-16 code unit in
15/// the Basic Multilingual Plane outside the surrogate range — so iterating Rust `char`s
16/// (Unicode scalar values) reproduces Java's `char`-at-a-time (UTF-16 code unit) loop
17/// exactly for this fixed table.
18const SINGLE_QUOTES: &[char] = &[
19 '\u{0060}', '\u{00B4}', '\u{02B9}', '\u{02BB}', '\u{02BC}', '\u{02BD}', '\u{02CA}', '\u{02CB}',
20 '\u{0091}', '\u{0092}', '\u{2018}', '\u{2019}', '\u{201A}', '\u{201B}', '\u{2032}', '\u{2035}',
21 '\u{275B}', '\u{275C}', '\u{FF07}',
22];
23
24/// Unicode double-quote / double-prime variants that normalise to the ASCII double quote
25/// `"`. Verbatim transcription of Java `UnicodeUtils.DOUBLE_QUOTES`
26/// (`"“”„‟″‶❝❞〝〞〟""`).
27const DOUBLE_QUOTES: &[char] = &[
28 '\u{0093}', '\u{0094}', '\u{201C}', '\u{201D}', '\u{201E}', '\u{201F}', '\u{2033}', '\u{2036}',
29 '\u{275D}', '\u{275E}', '\u{301D}', '\u{301E}', '\u{301F}', '\u{FF02}',
30];
31
32// ---------------------------------------------------------------------------------------
33// Homoglyph table: Java `UnicodeUtils.HOMOGLYHPS`/`containsHomoglyphs`/`replaceHomoglyphs`
34// (`UnicodeUtils.java:58-139` for the static-init loader, `:194-215` and `:239-276` for the
35// public methods). Used by `pipeline::stripandstash`'s step 13 (`replaceHomoglyphs`) to
36// fold Latin look-alike characters from other scripts (e.g. Cyrillic 'а' U+0430) — and a
37// handful of archaic Latin typography, notably U+017F LATIN SMALL LETTER LONG S "ſ" — down
38// to their canonical ASCII/Latin letter.
39// ---------------------------------------------------------------------------------------
40
41use std::borrow::Cow;
42use std::collections::HashMap;
43use std::sync::LazyLock;
44use unicode_normalization::UnicodeNormalization;
45
46/// The homoglyph table, embedded at compile time (copied verbatim from
47/// `name-parser-api/src/main/resources/unicode/homoglyphs.txt`, itself sourced from
48/// <https://raw.githubusercontent.com/codebox/homoglyph/master/raw_data/chars.txt> per the
49/// resource file's own header comment — `diff`-verified byte-for-byte identical to the
50/// Java classpath resource). Each line lists a "canonical" character (the line's first
51/// char) followed by every codepoint the codebox project considers a visual look-alike of
52/// it. 1861 lines total (`\n`-delimited; the file's own last line has no trailing
53/// terminator, the same off-by-one `wc -l` undercounts documented on
54/// `pipeline::blacklisted_epithets`'s `BLACKLIST_TXT`) — but see [`HOMOGLYPHS`]'s own doc
55/// comment: Java's loader, and this port, only ever read a small prefix of it.
56const HOMOGLYPHS_TXT: &str = include_str!("../resources/homoglyphs.txt");
57
58/// U+00D7 MULTIPLICATION SIGN, the botanical hybrid marker ("Abies × Picea"). Java
59/// `NameFormatter.HYBRID_MARKER` (`NameFormatter.java:14`: `public static final char
60/// HYBRID_MARKER = '×';`). Excluded from the table [`HOMOGLYPHS`] builds below — even
61/// though it appears, unremoved, as one of the look-alikes on `homoglyphs.txt`'s `x` row
62/// (line 88) — so a genuine hybrid marker is never folded into a plain "x", which would
63/// silently destroy hybrid-name semantics. Java's own `UnicodeUtilsTest.containsHomoglyphs`
64/// asserts exactly this: `"Abies × Picea"` is NOT flagged as containing a homoglyph.
65const HYBRID_MARKER: char = '\u{00D7}';
66
67/// Java `ignoredCanonicals` (`UnicodeUtils.java:67`): `new CharArraySet(new char[]{' ',
68/// '\'', '-', '﹘'})`. A `homoglyphs.txt` row whose canonical (first) character is one of
69/// these four is skipped ENTIRELY — no homoglyph entries at all are taken from that row.
70/// U+0020 SPACE, U+0027 APOSTROPHE, U+002D HYPHEN-MINUS, U+FE58 SMALL EM DASH.
71const IGNORED_CANONICALS: [char; 4] = [' ', '\'', '-', '\u{FE58}'];
72
73/// Java's local `ignore` set (`UnicodeUtils.java:93-103`, rebuilt from the literal
74/// `"‘’“”"` on every loop iteration but with identical fixed content
75/// each time — a code-style quirk, not per-row-varying data). Curly single/double-quote
76/// variants are never captured as a homoglyph VALUE on ANY row — quote normalisation is
77/// [`normalize_quotes`]'s job, not this table's.
78const QUOTE_IGNORE: [char; 4] = ['\u{2018}', '\u{2019}', '\u{201C}', '\u{201D}'];
79
80/// Java `DIACRITICS` (`UnicodeUtils.java:33-56`, static-init block): standalone
81/// combining/spacing diacritic marks, excluded from ever being captured as a homoglyph
82/// VALUE (a bare combining accent isn't meaningfully "the same as" some base letter, even
83/// where the auto-generated codebox table lists it as a look-alike on some canonical's
84/// row). Verbatim transcription of the Java string literal
85/// `"´˝\` ̏ˆˇ˘ ̑¸¨· ̡ ̢ ̉ ̛ˉ˛ ˚˳῾᾿"` with its 7 embedded literal spaces filtered out
86/// (`.filter(cp -> cp != 32)`, exactly like the Java) — 21 codepoints, extracted
87/// programmatically from the Java source (not hand-transcribed from the hard-to-read
88/// rendered glyphs, several of which are zero-width combining marks that visually vanish
89/// or merge with a neighbour when displayed).
90const DIACRITICS: [char; 21] = [
91 '\u{00B4}', '\u{02DD}', '\u{0060}', '\u{030F}', '\u{02C6}', '\u{02C7}', '\u{02D8}', '\u{0311}',
92 '\u{00B8}', '\u{00A8}', '\u{00B7}', '\u{0321}', '\u{0322}', '\u{0309}', '\u{031B}', '\u{02C9}',
93 '\u{02DB}', '\u{02DA}', '\u{02F3}', '\u{1FFE}', '\u{1FBF}',
94];
95
96/// Java `UnicodeUtils.HOMOGLYHPS` (`UnicodeUtils.java:59-139`, static-init block): a
97/// codepoint -> canonical-char map, built once from [`HOMOGLYPHS_TXT`]. Faithful port of
98/// the loader's exact algorithm, verified against a compiled run of the real Java
99/// `LineReader` over the real resource file (see the loop body comments below for the
100/// specific findings that verification surfaced):
101///
102/// 1. `'ſ'` (U+017F LATIN SMALL LETTER LONG S) -> `'s'` is inserted FIRST, before any table
103/// row loads, so it wins over row `f`'s own listing of `'ſ'` as one of ITS look-alikes
104/// (`homoglyphs.txt` line 72, canonical `f`) — the loop below never overwrites an
105/// existing key. Java `UnicodeUtils.java:71`: `homoglyphs.put(toCodePoint('ſ'), 's');`.
106/// 2. Each remaining `homoglyphs.txt` line is a candidate row, in file order. Java's
107/// `LineReader` (`skipBlank=true, skipComments=true`, the defaults its 1-arg
108/// `InputStream` constructor chains to) skips blank and `#`-comment lines before they
109/// ever reach this loop. A plain `line.is_empty()` check stands in for `LineReader`'s
110/// fuller `StringUtils.isBlank` (which additionally treats any all-whitespace line as
111/// blank): the ONE line in this file that's all-whitespace-but-not-empty (line 9, a run
112/// of Unicode space variants: U+0020 U+0020 U+1680 U+2000-U+200A U+2028 U+2029 U+202F
113/// U+205F) is NOT actually blank under Java's exact `Character.isWhitespace` (which
114/// excludes U+2007 FIGURE SPACE and U+202F NARROW NO-BREAK SPACE, both present on that
115/// line) — so Java's `LineReader` does NOT skip it either, and it reaches
116/// `UnicodeUtils`'s own loop body same as any other row. It's discarded there instead,
117/// by step 3's `IGNORED_CANONICALS` check (its canonical, the first char, is U+0020
118/// SPACE) — so the outcome is identical either way; `line.is_empty()` here is not a
119/// faithfulness gap.
120/// 3. The row's canonical is its first char (`.chars().next()` on a non-empty `str` is
121/// exactly Java's `line.charAt(0)` here — the canonical is never a surrogate pair, and a
122/// Rust `char` is already a full Unicode scalar value). If the canonical is one of
123/// [`IGNORED_CANONICALS`], the entire row is skipped.
124/// 4. Otherwise, every remaining char on the row becomes a candidate homoglyph VALUE,
125/// filtered exactly as `UnicodeUtils.java:104-111` filters it: codepoint > 128 (skip
126/// ASCII/Latin-1-control), not [`HYBRID_MARKER`], not in [`DIACRITICS`], not already a
127/// map key (first row wins — relevant beyond just the `'ſ'` pre-seed whenever the same
128/// look-alike appears on more than one row), not in [`QUOTE_IGNORE`].
129/// 5. After a (non-ignored-canonical) row is fully processed, Java breaks the loop
130/// (`UnicodeUtils.java:128`: `if (lr.getRow() > 175 || 'ɸ' == canonical) break;`,
131/// comment "skip all rare chars"). `lr.getRow()` is the row's own 1-indexed PHYSICAL
132/// line number in the source file (confirmed by compiling and running the real
133/// `LineReader` against the real resource file) — i.e. exactly this loop's `row`
134/// variable below, a straight `enumerate()` index, not a count of rows actually
135/// accepted. The `> 175` numeric fallback is dead code against the actual resource file
136/// — canonical `'ɸ'` (U+0278, physical line 151) always fires first — but ported
137/// verbatim anyway, matching this port's established precedent of preserving
138/// apparently-dead Java guards faithfully (see e.g. `stash_trailing_strain_code`'s
139/// `DIGITS_ONLY` guard in `pipeline::stripandstash`). Net effect: only `homoglyphs.txt`
140/// lines 1-151 are ever consulted — lines 1-8 are the file's own comment header (all
141/// `#`-prefixed), line 9 is the ignored space-canonical row, 2 more rows in range (14,
142/// canonical `'`; 20, canonical `-`) are also [`IGNORED_CANONICALS`], and the remaining
143/// 140 rows (lines 10-151, minus those 2) each contribute at least one map entry — lines
144/// 152-1861 (the file's long CJK "duplicate codepoint" tail, e.g. `"𦰶𦰶"`) are never
145/// read into the map at all — by Java's own design, not a deferral this port introduces.
146static HOMOGLYPHS: LazyLock<HashMap<char, char>> = LazyLock::new(|| {
147 let mut map = HashMap::new();
148 map.insert('\u{017F}', 's'); // 'ſ' -> 's', pre-seeded (see step 1 above)
149
150 for (idx, line) in HOMOGLYPHS_TXT.lines().enumerate() {
151 let row = idx + 1; // 1-indexed physical line number == Java `lr.getRow()`
152 if line.is_empty() || line.starts_with('#') {
153 continue;
154 }
155 let canonical = line.chars().next().expect("checked non-empty above");
156 if IGNORED_CANONICALS.contains(&canonical) {
157 continue;
158 }
159 // Rust-only: a digit row listing ASCII letters too ("0Oo…", "1Il…") folds a look-alike
160 // letter to the ASCII letter of its case, not to the digit — a Cyrillic "о" in
161 // "оleiformis" is an "o", and Java made it "0leiformis".
162 let letter = |upper: bool| {
163 line.chars()
164 .skip(1)
165 .find(|c| c.is_ascii_alphabetic() && c.is_ascii_uppercase() == upper)
166 };
167 let (upper_letter, lower_letter) = if canonical.is_ascii_digit() {
168 (letter(true), letter(false))
169 } else {
170 (None, None)
171 };
172 for cp in line.chars().skip(1) {
173 if (cp as u32) > 128
174 && cp != HYBRID_MARKER
175 && !DIACRITICS.contains(&cp)
176 && !map.contains_key(&cp)
177 && !QUOTE_IGNORE.contains(&cp)
178 {
179 let folded = match (cp.is_uppercase(), cp.is_lowercase()) {
180 (true, _) => upper_letter,
181 (_, true) => lower_letter,
182 _ => None,
183 };
184 map.insert(cp, folded.unwrap_or(canonical));
185 }
186 }
187 if row > 175 || canonical == 'ɸ' {
188 break;
189 }
190 }
191 map
192});
193
194/// Java `UnicodeUtils.containsHomoglyphs(CharSequence)` (`UnicodeUtils.java:194-196`,
195/// delegating to `findHomoglyph` at `:203-215`). Returns true if `s` contains at least one
196/// character that is a known homoglyph of a Latin letter. Java's version short-circuits via
197/// a codepoint lower/upper-bound check before the map lookup (an `Int2CharMap`-specific
198/// micro-optimisation); a `HashMap<char, _>` lookup here is already O(1) average without
199/// needing that bound check reproduced.
200pub fn contains_homoglyphs(s: &str) -> bool {
201 s.chars().any(|c| HOMOGLYPHS.contains_key(&c))
202}
203
204/// Java `UnicodeUtils.replaceHomoglyphs(CharSequence, boolean)` / its 3-arg sibling
205/// (`UnicodeUtils.java:239-276`), specialised to the only shape any ported call site ever
206/// needs: `inclHyphens=false` and no `keep` set. `pipeline::stripandstash`'s step 13 (the
207/// sole caller) always passes `inclHyphens=false` — Unicode hyphen variants are already
208/// normalised to ASCII `-` by the immediately-preceding step 12 (`normaliseHyphens`), per
209/// that Java method's own comment ("Hyphen homoglyphs are intentionally excluded — those
210/// are normalised above already") — so `HYPHEN_HOMOGLYHPS` (`UnicodeUtils.java:141-167`)
211/// and the `keep`-list parameter are both unreachable dead weight for this port's actual
212/// call graph and are not ported. Replaces every character present as a key in
213/// [`HOMOGLYPHS`] with its mapped canonical; every other character passes through
214/// unchanged.
215pub fn replace_homoglyphs(s: &str) -> String {
216 s.chars()
217 .map(|c| *HOMOGLYPHS.get(&c).unwrap_or(&c))
218 .collect()
219}
220
221/// Faithful port of Java `String.trim()`: strips only leading/trailing chars whose
222/// codepoint is <= U+0020 (NOT the full Unicode White_Space set that Rust's str::trim uses).
223/// Use this everywhere the Java source calls `.trim()`.
224///
225/// Java's `String.trim()` only removes characters with codepoint values 0–32 (space through
226/// control chars), while Rust's `str::trim()` removes the entire Unicode `White_Space` category.
227/// This means non-breaking space (U+00A0) and other Unicode spaces are trimmed by Rust but NOT
228/// by Java, causing empty/length guard logic to diverge.
229pub fn java_trim(s: &str) -> &str {
230 s.trim_matches(|c: char| (c as u32) <= 0x20)
231}
232
233/// Folds every Unicode space separator (category `Zs`: NBSP U+00A0, U+1680, U+2000–U+200A, the
234/// narrow NBSP U+202F, U+205F, U+3000) to an ASCII space. The parser's own whitespace handling is
235/// ASCII-only (Java parity: `String.trim()`, `Character.isWhitespace`, and the `(?-u:\s)` regexes
236/// all skip NBSP), so an NBSP left in place glues words together and hides keywords from every
237/// stage — `Assimineidae\u{a0}incertae\u{a0}sedis` read as a monomial with the author
238/// `incertae sedis`. Runs once on the raw input, before [`java_trim`], so the guards and every
239/// later stage see plain spaces. Borrows when there is nothing to fold (the common case).
240pub fn normalize_spaces(x: &str) -> Cow<'_, str> {
241 if x.chars().any(is_unicode_space) {
242 Cow::Owned(
243 x.chars()
244 .map(|c| if is_unicode_space(c) { ' ' } else { c })
245 .collect(),
246 )
247 } else {
248 Cow::Borrowed(x)
249 }
250}
251
252/// The input clean-up every parse starts with: [`normalize_spaces`], invisible format characters
253/// removed — zero-width space and joiners U+200B–U+200D, the word joiner U+2060 and the byte order
254/// mark U+FEFF (the first cell of a UTF-8 file with a BOM) — and the fullwidth ASCII forms
255/// U+FF01–U+FF5E folded to ASCII, as East Asian keyboards type them (`(Tubangui,1928)`). The
256/// homoglyph table cannot do the latter: it maps fullwidth `l` and `I` to the digit `1`. Not the
257/// soft hyphen U+00AD: in the data it is mostly half of the Windows-1252 mojibake `Ã\u{ad}` for
258/// `í`, repaired later, or a stand-in for a real hyphen (`Miranda\u{ad}Ribeiro`). Decomposed
259/// letters are composed (NFC): a combining accent after its letter (`Za\u{301}gors\u{30c}ek`)
260/// otherwise splits the word, and Java read `Za gors ek`. Borrows when there is nothing to change.
261pub fn normalize_input(x: &str) -> Cow<'_, str> {
262 let x: Cow<'_, str> = if x
263 .chars()
264 .any(unicode_normalization::char::is_combining_mark)
265 {
266 Cow::Owned(x.nfc().collect())
267 } else {
268 Cow::Borrowed(x)
269 };
270 if !x
271 .chars()
272 .any(|c| is_unicode_space(c) || is_invisible(c) || is_fullwidth(c))
273 {
274 return x;
275 }
276 Cow::Owned(
277 x.chars()
278 .filter(|&c| !is_invisible(c))
279 .map(|c| {
280 if is_unicode_space(c) {
281 ' '
282 } else if is_fullwidth(c) {
283 char::from_u32(c as u32 - 0xFEE0).unwrap_or(c)
284 } else {
285 c
286 }
287 })
288 .collect(),
289 )
290}
291
292/// A fullwidth ASCII form, U+FF01–U+FF5E: `!` … `~`.
293pub fn is_fullwidth(c: char) -> bool {
294 matches!(c, '\u{FF01}'..='\u{FF5E}')
295}
296
297fn is_invisible(c: char) -> bool {
298 matches!(c, '\u{200B}'..='\u{200D}' | '\u{2060}' | '\u{FEFF}')
299}
300
301fn is_unicode_space(c: char) -> bool {
302 matches!(
303 c,
304 '\u{00A0}' | '\u{1680}' | '\u{2000}'..='\u{200A}' | '\u{202F}' | '\u{205F}' | '\u{3000}'
305 )
306}
307
308/// Normalises the many unicode apostrophe / single-quote variants to the ASCII apostrophe
309/// (') and the unicode double-quote variants to the ASCII double quote ("). Author names
310/// and quoted/provisional names routinely arrive with curly, prime, modifier-letter,
311/// low-9, fullwidth or Windows-1252 quotes; collapsing them to ASCII keeps tokenisation
312/// and the parsed output consistent regardless of the input's quote style.
313///
314/// Faithful port of Java `UnicodeUtils.normalizeQuotes(String)`. Java's `null`-in/`null`-out
315/// case is not ported: this takes and returns non-nullable Rust types, so callers threading
316/// an optional authorship string (e.g. `Option<&str>`) map over it instead
317/// (`authorship.map(normalize_quotes)`), which reproduces the same null-forwarding at the
318/// `Option` level.
319pub fn normalize_quotes(x: &str) -> String {
320 x.chars()
321 .map(|c| {
322 if SINGLE_QUOTES.contains(&c) {
323 '\''
324 } else if DOUBLE_QUOTES.contains(&c) {
325 '"'
326 } else {
327 c
328 }
329 })
330 .collect()
331}
332
333#[cfg(test)]
334mod tests {
335 use super::*;
336
337 // ---- Homoglyph table ----
338
339 /// `homoglyphs.txt` lines 10-151 (142 physical lines) are the candidate data rows once
340 /// the 8-line `#`-comment header and line 9's ignored space-canonical row are behind us
341 /// (see [`HOMOGLYPHS`]'s own doc comment); 2 more of those 142 (line 14, canonical `'`;
342 /// line 20, canonical `-`) are also [`IGNORED_CANONICALS`], leaving 140 rows that each
343 /// contribute at least one homoglyph codepoint (independently verified against the
344 /// source file — no row contributes zero) — plus the `'ſ'` pre-seed, for an exact total
345 /// of 1751 map entries. Asserted exactly (not just "a lot"), as a direct regression
346 /// guard against a bad resource copy or a loader loop that silently stops early/never
347 /// reaches `'ɸ'`/double-counts/under-filters.
348 #[test]
349 fn resource_loads_exactly_1751_entries() {
350 assert_eq!(
351 HOMOGLYPHS.len(),
352 1751,
353 "loaded homoglyph count drifted — see this test's own doc comment"
354 );
355 }
356
357 /// Java `UnicodeUtilsTest.replaceHomoglyphs` (`UnicodeUtilsTest.java:117`): long-s
358 /// folds to plain "s" (NOT "f", even though `ſ` sits on `homoglyphs.txt`'s `f`-canonical
359 /// row too — the manual pre-seed wins), while `æ` — a genuine Latin ligature, not a
360 /// look-alike — is untouched; this table is not `foldToAscii`. Same shape as the
361 /// corpus's `specificEpithet`-affecting rows ("Musca domeſtica", "Amphisbæna
362 /// fuliginoſa") — see the sibling test below and `pipeline::stripandstash`'s golden
363 /// harness `ALLOWLIST` (now empty, see that file's own history).
364 #[test]
365 fn long_s_folds_to_s_but_ae_ligature_is_untouched() {
366 assert_eq!(
367 replace_homoglyphs("Coccinella 2-puſtulata Linnæus, 1758"),
368 "Coccinella 2-pustulata Linnæus, 1758"
369 );
370 }
371
372 /// The two golden-corpus names this port's `replace_homoglyphs` stub used to leave
373 /// untouched (`crates/nameparser/tests/parse_golden.rs`, `specificEpithet`/`warnings`
374 /// mismatches, now 0). `genus="Amphisbæna"` keeps its `æ` — only the specific epithet's
375 /// `ſ` is a homoglyph here.
376 #[test]
377 fn corpus_long_s_names_fold_to_s() {
378 assert_eq!(replace_homoglyphs("Musca domeſtica"), "Musca domestica");
379 assert_eq!(
380 replace_homoglyphs("Amphisbæna fuliginoſa"),
381 "Amphisbæna fuliginosa"
382 );
383 }
384
385 /// Java `UnicodeUtilsTest.containsHomoglyphs` (`UnicodeUtilsTest.java:77`, comment
386 /// "hybrid marker is fine in out domain!"): U+00D7 '×' is excluded from the table even
387 /// though it sits on the `x`-canonical row, so a genuine hybrid-formula name survives
388 /// `replace_homoglyphs` untouched.
389 #[test]
390 fn hybrid_marker_is_not_a_homoglyph() {
391 assert!(!contains_homoglyphs("Abies × Picea"));
392 assert_eq!(replace_homoglyphs("Abies × Picea"), "Abies × Picea");
393 }
394
395 /// Java `UnicodeUtilsTest.replaceHomoglyphs` (`UnicodeUtilsTest.java:119-120`, the
396 /// `composed` string): digraphs/ligatures round-trip unchanged — this table only
397 /// touches genuine cross-script look-alikes, never legitimate distinct Latin letters.
398 #[test]
399 fn ligatures_and_composed_letters_are_untouched() {
400 let composed = "æÆœŒIJijLjljȸȹßstſtfffiflffiffl";
401 assert!(!contains_homoglyphs(composed));
402 assert_eq!(replace_homoglyphs(composed), composed);
403 }
404
405 /// A Cyrillic 'а' (U+0430, `homoglyphs.txt` line 67's `a`-canonical row) look-alike
406 /// mid-word folds to plain "a", flagging as a homoglyph.
407 #[test]
408 fn cyrillic_a_lookalike_folds_to_latin_a() {
409 assert!(contains_homoglyphs("Aus \u{0430}bus"));
410 assert_eq!(replace_homoglyphs("Aus \u{0430}bus"), "Aus abus");
411 }
412
413 /// A plain-ASCII scientific name contains no homoglyphs and round-trips byte-identical.
414 #[test]
415 fn plain_ascii_name_has_no_homoglyphs() {
416 assert!(!contains_homoglyphs("Abies alba Mill."));
417 assert_eq!(replace_homoglyphs("Abies alba Mill."), "Abies alba Mill.");
418 }
419
420 /// `¡` (U+00A1) is a Win-1252 transcription artefact handled by the SEPARATE
421 /// `repairWin1252Artefacts` step (`pipeline::stripandstash`'s `repair_win1252_artefacts`
422 /// step 14) — it is not in the homoglyph table itself. Java
423 /// `UnicodeUtilsTest.replaceHomoglyphs` (`UnicodeUtilsTest.java:115`, comment "should
424 /// this be a homoglyph?") asserts the same non-membership.
425 #[test]
426 fn win1252_artefacts_are_not_homoglyphs() {
427 assert!(!contains_homoglyphs("¡i"));
428 assert_eq!(replace_homoglyphs("¡i"), "¡i");
429 }
430
431 #[test]
432 fn folds_curly_single_quotes_to_ascii_apostrophe() {
433 assert_eq!(normalize_quotes("d\u{2019}Urville"), "d'Urville");
434 assert_eq!(normalize_quotes("\u{2018}Aus\u{2019}"), "'Aus'");
435 }
436
437 #[test]
438 fn normalize_input_drops_invisibles_and_folds_spaces_and_fullwidth_forms() {
439 assert_eq!(
440 normalize_input("\u{feff}Abies\u{a0}alba Mi\u{200d}ll.\u{200b}"),
441 "Abies alba Mill."
442 );
443 assert_eq!(normalize_input("(Tubangui,1928)"), "(Tubangui,1928)");
444 assert_eq!(normalize_input("Mill."), "Mill.");
445 assert_eq!(normalize_input("Za\u{301}gors\u{30c}ek"), "Zágoršek");
446 assert!(matches!(normalize_input("Abies alba L."), Cow::Borrowed(_)));
447 // the soft hyphen stays: mojibake for `í`, or a stand-in for a hyphen
448 assert_eq!(normalize_input("Ort\u{c3}\u{ad}z"), "Ort\u{c3}\u{ad}z");
449 }
450
451 #[test]
452 fn normalize_spaces_folds_unicode_space_separators_only() {
453 assert_eq!(
454 normalize_spaces("Abies\u{a0}alba\u{202f}L.\u{3000}x"),
455 "Abies alba L. x"
456 );
457 assert!(matches!(
458 normalize_spaces("Abies alba\tL."),
459 Cow::Borrowed(_)
460 ));
461 }
462
463 #[test]
464 fn folds_curly_double_quotes_to_ascii_double_quote() {
465 assert_eq!(normalize_quotes("\u{201C}Aus bus\u{201D}"), "\"Aus bus\"");
466 }
467
468 #[test]
469 fn folds_backtick_acute_and_fullwidth_apostrophe_variants() {
470 assert_eq!(normalize_quotes("\u{0060}"), "'");
471 assert_eq!(normalize_quotes("\u{00B4}"), "'");
472 assert_eq!(normalize_quotes("\u{FF07}"), "'");
473 }
474
475 #[test]
476 fn folds_low9_and_prime_double_quote_variants() {
477 assert_eq!(normalize_quotes("\u{201E}Aus\u{201F}"), "\"Aus\"");
478 assert_eq!(normalize_quotes("\u{2033}"), "\"");
479 }
480
481 #[test]
482 fn leaves_plain_ascii_and_already_normalised_quotes_untouched() {
483 assert_eq!(normalize_quotes("O'Brien"), "O'Brien");
484 assert_eq!(normalize_quotes("\"Aus bus\" Smith"), "\"Aus bus\" Smith");
485 assert_eq!(normalize_quotes("Abies alba Mill."), "Abies alba Mill.");
486 }
487
488 #[test]
489 fn java_trim_strips_only_up_to_u0020() {
490 assert_eq!(java_trim(" x\t"), "x"); // ASCII space + tab (<=0x20) stripped
491 assert_eq!(java_trim("\n x \r"), "x"); // newline/CR stripped
492 assert_eq!(java_trim("\u{00A0}x\u{00A0}"), "\u{00A0}x\u{00A0}"); // NBSP (>0x20) NOT stripped
493 assert_eq!(java_trim("\u{00A0}"), "\u{00A0}"); // NBSP-only stays non-empty, matching Java
494 }
495}