Changelog¶

All notable changes to this project will be documented in this file.

The format follows Keep a Changelog. Version numbers use the MAJOR.MINOR.PATCH shape but follow disarm's own release policy — patch = fixes/cleanups/docs, minor = features or major refactors, and the major component denotes support status, not API compatibility (see RELEASING.md).

Project renamed translit → disarm (#264). Historical entries below predate the rename and refer to the old identity (translit-rs on PyPI, the translit import package, the _translit native module); they are left unchanged because they were accurate for their release. Entries from this point on use the disarm identity.

[Unreleased]¶

Added¶

• is_case_fold_stable — ask whether a value is a stable identity key before you key a table on it (#619). Answers fold_case(x) == x.lower(). A False says some other string folds to the same value, which is the precondition node-tar's PathReservations guard missed in CVE-2026-23950: groß.txt and gross.txt are one path on a case-insensitive filesystem. Available in Rust (api::is_case_fold_stable, and on DisarmStr), Python (is_case_fold_stable, Text.is_case_fold_stable), Node (isCaseFoldStable), Ruby (Disarm.case_fold_stable?), the C ABI and Java.

It states a fact about the string, not suspicion. groß is an ordinary German word and file an ordinary ligature, so the predicate reads True for ordinary text and is deliberately kept out of has_anomalies and out of the CVE detector panel — folding it in would flag ordinary German and every Greek word ending in sigma. What to do about a False is the caller's decision: reserve both forms, reject the name, or key the table on fold_case rather than str.lower().

str.lower() is the comparison basis and str.casefold() is not. Casefolding performs the very transform under test, so a predicate written against it answers True for every string in Unicode — the substitution #617 already made once, now pinned by a test.

Not a per-character table, because a per-character table is wrong for Greek. ΟΔΟΣ ("street") lowercases to οδος and folds to οδοσ, yet Σ agrees with itself in isolation; U+03A3 is the only code point in Unicode whose lowercase mapping depends on its neighbours, which a Tier-3 test asserts by enumeration rather than by assertion. The implementation is allocation-free for anything that contains no capital sigma — ASCII short-circuits, everything else scans the folding table in place — and falls back to the exact string comparison for the rest, so it cannot drift from what fold_case actually does.

CVE-2026-23950's Detected by column stops reading —. The collision itself is still a property of a pair of names and no single-string predicate can report it (#620 tracks that); the precondition is what moved. Measured limit: the issue paired this with CVE-2019-19844, and that half does not hold — that row's probe turns on U+0131 DOTLESS I, which folds and lowercases to itself and collides through .upper() instead, so the predicate is silent on it.

• Ten CVEs on encoding rather than code points, and the survey method behind them (36 → 46 rows). Found by sweeping NVD across the operations disarm performs, then verified one ID at a time.

Class	Added
Overlong / invalid byte sequences	CVE-2024-46954, CVE-2026-44288, CVE-2009-4142
Lone surrogates	CVE-2022-31116, CVE-2025-64439, CVE-2008-4066
Full-width evasion of a detector	CVE-2007-2688, CVE-2001-0669
Encoding layers disarm does not own	CVE-2022-3782, CVE-2006-2753

The byte-level rows are the first on the page whose input is not a str. decode_to_utf8 replaces overlong sequences rather than decoding them, so the Ghostscript traversal never materializes and strict=True refuses outright. CVE-2026-44288 names the correct behaviour exactly — protobufjs decoded overlong sequences "to canonical characters instead of replacing them" — which is what makes not-affected measurable rather than asserted.

CVE-2025-64439 is a Unicode edge case reaching RCE through an error path: illegal surrogates made msgpack serialization fail in LangGraph, and the fallback was JSON deserialization of untrusted data. disarm substitutes rather than drops, which is what keeps key<U+DC00>value from colliding with keyvalue — the CVE-2022-31116 shape.

CVE-2007-2688 is the Threat Model's ordering rule eighteen years early. Cisco IPS, Check Point and IBM ISS Proventia all shipped the same missing normalization step in the same month. It is also the same fold TestFullwidthUnmaskingHazard pins as a hazard — both readings are correct, and pipeline position decides which applies.

docs/security/cve-validation.md now records how rows are found: the NVD sweeps by mechanism rather than by product, the per-ID verification that caught CVE-2017-20190 having no CVSS at all, and the non-CVE research that informed rows — Paul Butler on variation-selector smuggling, and the CoreText Telugu crash.

• The comparator table no longer contradicts the matrix. CVE-2026-23950 rendered as Neutralized in the matrix and as no under both disarm columns in the comparison a hundred lines below, because the comparison scores every row against two fixed disarm entry points and that row is neutralized by fold_case.

The columns stay fixed on purpose — decancer and unidecode each expose exactly one entry point, so letting disarm pick a different function per row would flatter it. Instead the three rows whose matrix neutralizer is neither fixed column now carry a † naming the function that does own them, so a no reads as not this function rather than not disarm. test_named_elsewhere_matches_the_registry derives that set from the registry rather than trusting the benchmark's copy, and also checks the named function is one the row actually lists.

• Comparator corpus caught up with the matrix, and gated against falling behind again (15 → 22 of 36 rows). The corpus stopped growing when the matrix went from 20 to 36 rows, and the drift gate at the time only compared it against NEUTRALIZABLE — which had stopped growing too. Both sides moved together, so nothing failed and 21 rows fell out of the comparison silently.

Seven comparable rows are now compared: the four terminal-control CVEs (CVE-2025-55754, CVE-2024-52005, CVE-2023-43620, CVE-2023-37275), the zalgo row (CVE-2017-20190), the Latin kra (CVE-2019-11721) and the sharp-s path collision (CVE-2026-23950).

The other 14 are named rather than left unexplained: 11 out of scope (nothing neutralizes them, so nothing to compare), 2 not-affected (a cost property, not a transformation), 1 detected without being neutralized. test_every_registry_row_is_compared_or_has_a_reason_not_to now checks each row against the registry rather than against another list that can drift with it.

A flaw in the comparison predicate came out of adding the sharp-s row. The harness neutralized case with str.casefold(), which performs Unicode full case folding — and therefore maps ß to ss itself. Every tool would have passed CVE-2026-23950 by measuring Python rather than the tool. Switched to str.lower(), which leaves ß alone; no other row changed.

Scores on the widened corpus: canonicalize and strip_obfuscation 20/22, decancer 16/22, unidecode 14/22. The single row where disarm reads no and unidecode reads yes is CVE-2026-23950, and it is a deliberate refusal — folding ß in the confusable table would rewrite ordinary German, so the key builders own that collision. The page says so next to the table.

• Normalization cost as a CVE class, and a fourth disposition to describe it (33 → 36). CVE-2026-3276 (CPython unicodedata.normalize() on alternating-CCC runs, CWE-407), CVE-2023-46695 (Django NFKC on Windows, re-reported three times since) and CVE-2017-20190 ("Zalgo text", disputed, deferred, and carrying no CVSS score at all — only SSVC).

The input here is not a disguise, it is a bill, and the existing vocabulary could not say what disarm's relationship to it is. not-affected now means the CVE is a defect in another implementation of something disarm also does, and disarm's implementation was measured and does not have it — distinct from out-of-scope, which means disarm does not stop the attack. It is a stronger claim, so it is gated: identical output to CPython across all four forms, and a linear-cost bound.

disarm is not uniformly faster, and the page says so. Over nine input shapes at 20,000 characters it runs between 6× faster (CJK compatibility ideographs) and 10× slower (already-normalized text, where CPython's quick-check short-circuits and disarm does more work). An earlier informal measurement suggested a ~700× margin; that was a cold-start artefact and does not survive best-of-N timing.

The bound is the actual defense. canonicalize collapses a 2,000-mark pile to at most four characters, and is_zalgo flags the CVE-2026-3276 payload too — rejecting the input costs less than normalizing it quickly. What disarm does not do is bound input length, which is what the Frigate/Yeti/spbu_se_site CVEs in this family are actually about; test_disarm_does_not_bound_input_length pins that distinction so the cap is not misread as a resource limit.

• Eleven more CVEs, in the classes the matrix was thinnest on (22 → 33).

Class	Added
Ordering: normalize-then-validate	CVE-2026-28289, CVE-2024-43093, CVE-2023-41889, CVE-2023-52081
Terminal control sequences	CVE-2025-55754, CVE-2024-52005, CVE-2023-43620, CVE-2023-37275
Address bar and deny lists	CVE-2019-11721, CVE-2023-4399
Case-folding path collision	CVE-2026-23950

The ordering rows are out of scope on purpose. disarm can produce the canonical form; it cannot make a caller look at it first. CVE-2026-28289 states the bug in the Threat Model's own terms — NVD describes a TOCTOU weakness where "the dot-prefix check occurs before sanitization removes invisible characters". CVE-2024-43093 is in CISA's Known Exploited Vulnerabilities catalog.

The terminal class is neutralized and entirely undetected. All six escape-sequence rows are cleaned by strip_log_injection and reported by nothing, which is the sharpest argument yet for cleaning unconditionally rather than screening first. test_no_detector_reports_any_terminal_control_row pins that as a class-level claim.

CVE-2026-23950 closes a loop. node-tar's symlink poisoning turns on the ß/ss path collision — the single code point the CVE-2019-19844 exhaustive scan identified as the one the confusable table deliberately leaves alone, because ß is a real German letter. The key builders collide it; the canonicalizers correctly do not.

Two schema changes fell out of the additions. cvss and cvss_version are now optional, because CVE-2017-20190 has no CVSS record at all — only SSVC — and inventing a number would be worse than an empty cell. v4.0 joins the accepted versions.

One probe needed adjusting rather than one assertion: ASCII | is itself a TR39 confusable source, so a curl evil|sh payload tripped is_confusable for reasons unrelated to its CVE. The derived-detector gate caught it as a false positive.

• Corrected: there is no single call that neutralizes every vector (#609 follow-up). The CVE page said canonicalize was the one call to make when the attack is unknown. That was wrong, and its gate could not catch it, because every vector in the matrix at the time happened to be one canonicalize handles.

CVE-2017-7833 (Firefox, 5.3) is the vector that breaks it: a single Arabic vowel mark riding a Latin letter. One mark sits below the zalgo threshold, so is_zalgo correctly returns False, and canonicalize caps combining marks rather than removing them (#429) — so the spoof never collapses onto the genuine host.

CVE-2017-5383 (Firefox, 5.3) is the mirror image, and rules out the obvious replacement. strip_obfuscation removes the mark, but renders punctuation confusables as their names — U+2010 HYPHEN becomes the word "hyphen" — so it never folds to ASCII -. Neither preset dominates the other and they fail on different inputs, so "5/6 each, pick either" is the wrong read.

Measured across the matrix plus both, no entry point clears everything. catalog_key comes closest — the only one carrying both a confusable step and strip_accents — and it has no format-stripping step, so the Tags block of CVE-2025-32711 goes straight through. The answer is a composition: canonicalize(strip_zalgo(text, max_marks=0)).

TestOneCall now gates both halves, including a test that fails if any single entry point ever does become sufficient, so the guidance is revisited rather than left stale. has_bidi_conflict also stops reading zero: CVE-2017-7833's Arabic mark beside Latin letters is exactly the strong-direction mix it asks about.

• CVE matrix: comparator columns, split entry-point roles, and a measured "one call" answer (#607 follow-up). Three gaps in the published matrix, all of them the kind that make a table look more useful than it is.

canonicalize is the single call, and that is now measured rather than recommended. The matrix said which entry point handles which CVE; a defender's actual question is which call to make when the attack is unknown. canonicalize handles all thirteen neutralizable vectors, as do canonicalize_strict, strip_obfuscation and the llm_guardrail / rag_ingest profiles. strip_format handles eight — it has no confusable step, so every row needing a fold survives it. Every score is pinned in TestOneCallSuperset.

Detection has no such answer, and the asymmetry is the point. No detector covers the matrix, and neither does all of them together: CVE-2023-24329, CVE-2008-2383, CVE-2019-9535, CVE-2025-32711 and CVE-2019-9636 are silent to every one. All five are still neutralized by canonicalize, which settles the pipeline question — clean unconditionally, and use the detectors to decide whether to alert, never whether to clean. A pipeline that screens first and cleans only what it flagged forwards those five untouched.

Neutralizers and detectors are separate fields. They were one entry_points list, which read as though any name on it would defend the row. CVE-2019-19844 is the clearest case: its neutralizers detect nothing and its only detector rewrites nothing — and it was mislabelled Neutralized when it is also Detected. Each row's detector list is now derived, not written: the suite runs the row's vector through every detector and asserts the list matches what fired.

Comparator columns. benchmarks/cve_comparators.py runs disarm, decancer and unidecode over the same vectors under one predicate, and regenerates the published table. disarm 13/13, decancer 8/13, unidecode 10/13 — but the score is the least interesting part. unidecode maps by sound, so its homoglyph passes are decided by which homoglyph the attacker picked (рroduсt → rrodust, not product); decancer reorders bidi text rather than stripping it, leaving U+202E in the output. decancer-py==0.4.1 joins the bench extra.

• CVE validation suite (tests/test_cve_vectors.py, docs/security/cve-validation.md). The docs encourage security use; nothing checked that against a named attack. Twenty published CVEs are now reconstructed from the vector each one describes and asserted against disarm's real behaviour, in the CI gate, across six classes:

Class	CVEs
Source code and identifiers	CVE-2021-42574, CVE-2021-42694
Identity and account takeover	CVE-2019-19844, CVE-2013-7236, CVE-2020-12063
Filesystem and paths	CVE-2014-9390, CVE-2009-3376, CVE-2023-33955
Hostnames and URLs	CVE-2017-7832, CVE-2023-24329, CVE-2019-9636
Terminal output and logs	CVE-2008-2383, CVE-2019-9535
ML / LLM input	CVE-2025-32711, CVE-2024-5184, CVE-2024-5565, CVE-2023-29374, CVE-2023-36258, CVE-2024-3098, CVE-2023-32786

Seven rows are out-of-scope negatives, and that is the point. A suite that recorded only wins would quietly convert THREAT_MODEL.md's "no guarantee that any class of attack is fully neutralized" into a coverage claim. The negatives are asserted so the limits cannot drift into untested marketing — including two rows where canonicalizing in the wrong pipeline position makes an attack worse: ＿＿ｉｍｐｏｒｔ＿＿ becomes executable ASCII under NFKC (CVE-2024-3098's safe_eval blocklist), and ＃ becomes a real # that moves where a host ends (CVE-2019-9636).

Four findings came out of measuring rather than assuming. has_bidi_conflict() is correctly False for every Trojan Source payload — they are ASCII plus controls, with no strong RTL run — so it is not the detector for that family. collapse_whitespace() leaves a leading NUL, so it does not close CVE-2023-24329 on its own. ᴀ (U+1D00) has no uppercase mapping at all, so it is not a CVE-2019-19844 vector despite looking like the obvious one; the real collision class (non-ASCII code points whose .upper() is pure ASCII) is exactly ten members wide, walked exhaustively over all of Unicode in ~0.2s, and nine fold at canonicalize_strict while ß closes only under fold_case. And there is no "old CVEs are CVSS v2" rule: NVD backfilled a v3.1 score for CVE-2014-9390 while leaving CVE-2013-7236 and CVE-2009-3376 v2.0-only, so each row records the revision it quotes.

ml_normalize() passes all twelve bidi controls, PUA, and homoglyphs through unchanged — it is a tokenizer-hygiene preset, not a screen. The llm_guardrail and rag_ingest profiles are the entry points for untrusted text, and that is now asserted rather than implied.

The suite is mutation-checked: neutering any of nine entry points turns it red, so no assertion passes vacuously. TestDocsMatrixDrift derives the published table's scores and disposition wording from the registry, so a row cannot be softened in Markdown alone.

• 31 real-attacker confusable mappings TR39 does not carry (#597). Miss-mining the BitCore subset of the BitAbuse corpus (Lee et al., NAACL Findings 2025) with benchmarks/adversarial_eval surfaced codepoints that attackers substitute for basic-Latin letters and that TR39 does not list as sources at all, so normalize_confusables left them unfolded. ɴ→n alone accounts for 4,576 real occurrences; the top three (ɴ, ʍ, ɾ) are 74% of the Tier-1 mass.

They arrive in a new file, data/confusables_attested.tsv, not in confusables_supplement.tsv. That file declares itself cross-script and pins its provenance to one measured dataset above a stated danger threshold (#336); 18 of these sources are Latin folding to Latin, and none comes from that dataset. Two admission criteria, two provenance stories, two files — the generator merges them, but the audit trails stay separate.

The contract widens, and that is deliberate. Tier 1 (23 rows) are optical twins. Tier 2a (7) are positional — an attacker reached for an Armenian, Georgian or runic glyph by its place in the word, not because it resembles the letter — and Tier 2b (1) is convention. Admitting 2a means this table now encodes observed attacker substitution, which is wider than visual confusability; Unicode would not accept those rows upstream. The tier is recorded per row, and the widened rule is stated in docs/user-guide/confusables.md and THREAT_MODEL.md rather than left implicit.

Two details worth knowing. µ U+00B5 is folded alongside μ U+03BC, because NFKC maps one to the other and folding only one would make the result depend on the input's normalization form. And six rows fold an uppercase source to a lowercase target (Ƿ Ʌ Ա Ⴝ Ⴍ Ⴓ), which no generated row does — the generated pipeline reconciles case to the source and these bypass it. The attested form is kept, because the evidence is the letter the attacker meant, not its case.

Per the #39/#40 guardrail the corpora are measuring instruments, never optimization targets: the synthetic BitViper tail — 254 further codepoints, 733,029 occurrences, 98.5% of all novel misses — is excluded by construction, and no row is justified by a benchmark score. Latin table 2,189 → 2,220 mappings; the Cyrillic table is untouched, since seven of these sources already fold there and the new rows set the Latin column only.

• The C header is committed and drift-gated (#580). bindings/cabi/disarm.h was gitignored and regenerated inside the CI step that then compiled smoke.c against it. Both sides therefore moved together: a signature change plus a matching call-site change was self-consistent and passed. That is how a widened disarm_normalize_confusables (2 args → 3) reached review on #574 with every check green — a human reading the diff caught it, no test did.

The header is now a committed baseline, the same way src/metadata.rs and generated/parity.yaml are committed and drift-checked rather than regenerated blind. CI diffs the regenerated header against it in the cabi job, so every ABI change is a visible diff someone has to approve. A second test pins the arity of the entry points that shipped before 0.14, so a widening fails by name rather than as one line in a 346-line diff.

The gate does not judge whether a change is breaking — it makes changes visible. Deciding that a diff is additive (a new _opts entry point) rather than breaking (a widened signature) stays the reviewer's job.

• Selectable digit-mapping policy (#561). disarm folds a non-Latin digit to the ASCII digit; upstream TR39 folds several of them to a Latin letter (० → o, ೦ → O, ١ → l). Neither is wrong — numeric is right for prose, where a Devanagari zero really is a zero, and the letter is right for an identifier skeleton, whose only job is to make two confusable identifiers collide. The divergence was fixed in the table with no way to select the other side, so it read as a defect to anyone scoring disarm against a TR39-derived benchmark and cost points silently.

digit_policy now selects it: "numeric" (default, unchanged behaviour) or "tr39". Reaches Python (normalize_confusables(..., digit_policy=…) and Text), Rust (api::normalize_confusables_with + the DigitPolicy enum), Node (digitPolicy option), Ruby (digit_policy: keyword), Java/Kotlin (DigitPolicy enum), and the C ABI.

The Rust surface adds a second function rather than a third parameter on normalize_confusables: that is the crate's most-used security primitive and the policy is rarely set, so widening it would tax every call site for something almost none of them need. normalize_confusables(text, target) is unchanged.

The divergent rows are generated, not hand-maintained: scripts/gen_confusables.py already computes both sides — it makes this exact choice at generation time via enforce_digit_target (#439) — so the discarded alternative is now emitted as src/tables/data/confusables_digit_tr39.tsv (45 rows) and build.rs turns it into an override PHF. An override set rather than a second full table, so the two policies cannot drift on the rows they agree on, which is all but 45 of them.

Scope: the policy is a property of the normalize_confusables entry point only. The presets (canonicalize, catalog_key, search_key, …) serve prose and keys, where numeric is unambiguously right, so they have no switch. Hostname analysis also stays numeric — selecting TR39 there would silently change what is_suspicious_hostname flags, which is a security-behaviour change and belongs in its own issue.

Scope, second axis: "tr39" applies to the Latin target only; with any other target script it is a no-op and the fold behaves exactly as "numeric". The override set is generated from the Latin table and its values are TR39's Latin-script targets, so consulting it under target_script = "cyrillic" emitted Latin letters into a Cyrillic skeleton (० folded to o, not 0) and invented folds for sources the Cyrillic table deliberately has no row for. Three of those leaked outputs (Ʌ, o, rn) are themselves confusable with Cyrillic, so the fold did not even reach a fixed point.

• ml_normalize reaches every binding. It was Rust + Python only, recorded in scripts/parity.py as a deliberate scope decision. That decision does not survive contact with what the preset is for: it is the ML/NLP entry point, so keeping it Python-only meant a Node or JVM model pipeline could not use disarm for the thing disarm built it to do. Now exposed as mlNormalize (Node, Java/Kotlin), Disarm.ml_normalize (Ruby), String.mlNormalize (Kotlin extension), and disarm_ml_normalize (C ABI), each with the fold_case switch from #559.

• Multi-codepoint confusable sources — contraction (#562). The confusable tables map one codepoint to one-or-more (0271 → rn), so expansion always worked. Contraction — recognising that rn may stand in for m — could not be expressed at all: the source column of both TSVs is a single hex codepoint in every data row. This was a schema change before it was a data change.

is_suspicious_hostname(host, contractions=True) / api::analyze_hostname_with(host, true) now folds ASCII digraphs that can impersonate a single letter into the canonical form, so arnazon.com canonicalizes to amazon.com. Reaches Python, Rust, Node, Ruby, Java/Kotlin, and the C ABI.

Off by default and confined to the hostname path. Unconditional contraction is worse than none: rn → m is right for arnazon and wrong for earnings, turnip, born. A hostname is the one place where the threat model justifies those false positives and there is no running prose to corrupt. It is not reachable from normalize_confusables at any setting; a general-text mode would need its own disambiguation story.

Three rules, each with recorded provenance: rn → m is the one TR39 itself sanctions (it reduces m to the sequence rn, and 17 sources fold to rn, the dominant multi-character target in the file); vv → w and cl → d are disarm additions from the IDN homograph literature. Every rule is a false-positive source, so the bar is "documented real-world technique", not "plausible".

Matching is leftmost-longest over an Aho-Corasick automaton (reusing the dependency #242 already brought in), and applied per label, so a digraph can never form across a dot. One pass is a fixed point by construction: build.rs asserts no rule's output occurs inside any rule's input, so a pass cannot expose a fresh match, and a data edit that introduced such a chain fails the build.

Argument style follows each ecosystem: an options object in Node/TypeScript, keyword arguments in Ruby, a MlNormalizeOptions builder in Java (mirroring the existing TransliterateOptions), default parameters on the Kotlin extension, and positional arguments with a nullable lang in the C ABI.

ml_normalize is removed from SCOPE_REVIEW in scripts/parity.py; the op now reads ✓ across rust/python/ruby/node in the parity matrix.

• Nightly Hypothesis run (.github/workflows/nightly-hypothesis.yml). Tier 2 is excluded from PR CI on purpose (~440 tests, non-deterministic, slow) and is not in the Tier-3 release gate either, so it ran only when a developer happened to run the full suite locally. #570 sat undetected in exactly that gap.

Runs at 03:17 UTC and on demand, with --hypothesis-seed=random so consecutive nights explore different input space, and ORACLE_MAXEX=20000 — 10× the local default — for the adversarial-oracle suite, the one env-tunable budget in the tier and the suite that found #570. A failure opens (or comments on) a single rolling issue rather than disappearing into a run log.

Deliberately not a required check and not in the publish path: a probabilistic suite must never be able to block a security release. It reports; a human triages.

• The bundled confusables.txt version is readable at runtime (#560). Nothing in the library reported which upstream release the confusable tables were folded from. The number existed — in the TSV header and in docs/provenance.md — but both are build-time artifacts, so a deployment could not answer "is my fold stale?" without inferring it from behaviour. It is now exposed everywhere: disarm::api::CONFUSABLES_VERSION (and api::confusables_version()) in Rust, disarm.CONFUSABLES_VERSION in Python, confusablesVersion() in Node and Java/Kotlin, Disarm.confusables_version in Ruby, and disarm_confusables_version() in the C ABI.

build.rs parses the value out of the TSV header it already reads, so the constant cannot drift from the data it describes, and the build fails if that header stops naming a version. Both confusable tables are folded from one upstream release, which build.rs asserts, so a single constant covers them.

There is deliberately no library-wide UNICODE_VERSION: disarm's bundled tables track different releases (confusables 17.0.0, case folding 16.0, East Asian width 15.1.0), so one number would be wrong for three of the four. See docs/provenance.md for the full table and the per-language accessors.

• ml_normalize takes fold_case=False (#559). The preset folds case deliberately, and that is defensible — most tokenizers are uncased. What was missing was a way to turn it off for one call. A caller who wants everything else the preset does (NFKC, demojize, transliterate, strip-accents, control and zero-width removal, whitespace folding) in front of a cased model now has a route to it. The fold is destructive and cannot be undone downstream, and an uncased evaluation harness cannot measure what it costs.

Default true/True, so existing behaviour is unchanged. The flag drops Step::FoldCase and nothing else; the no-fold step list is derived from the folded one by a const fn, so the two cannot drift, and a build-time assertion fires if the pipeline ever stops containing exactly one fold step.

ml_normalize is Rust + Python only (Node/Ruby/Java/C-ABI do not expose it — a standing scope decision recorded in scripts/parity.py), so the flag reaches disarm::api::ml_normalize, disarm.ml_normalize, and Text.ml_normalize.

Note fold_case=False restores case, not diacritics: strip_accents is a separate step and still runs, so José becomes Jose. See below.

• Confusables coverage introspection (#563). find_untranslatable has existed for transliteration since #184; there was no confusables analogue, so answering "which sources does disarm not fold?" meant building a harness outside the library against a cached copy of confusables.txt. Two read-only accessors now answer it from inside:

• unmapped_confusables(target) — every source in the bundled upstream file that the chosen table does not fold, sorted.

• find_unmapped_confusables(text, target) — the same question for one input, in the shape of find_untranslatable: (char, byte_offset) in order of appearance.

Both reach Rust, Python, Node, Ruby, Java/Kotlin, and the C ABI.

The denominator is generated: scripts/gen_confusables.py now emits src/tables/data/confusables_upstream_sources.tsv (the source set it already read and discarded), and build.rs turns it into a PHF set. The exposure set is derived at runtime — upstream sources minus the resolved table's keys — so it cannot go stale against the table it describes, and one denominator covers both targets. The existing confusable TSVs regenerate byte-identically; this change is purely additive.

The per-input scan composes exactly as the fold does (#475/#477/#483), so a decomposed homoglyph whose precomposed form is mapped counts as covered, and offsets anchor to the caller's string rather than to the composed intermediate — matching find_untranslatable's guarantee.

Read the result as exposure, not as a score: a tool at 95% per-source coverage is one query away from the other 5%, and this set is where an adaptive attacker goes. Nothing is filtered out, which means the Latin set contains five ASCII characters (%, 0, 1, I, m) — TR39 is a skeleton transform (m→rn, I/1→l, 0→O) and disarm deliberately does not apply those rows, so a scan over ordinary English reports the letter m. Documented on every surface; a coverage report that quietly drops rows reads as coverage it does not have.

Changed (breaking)¶

• disarm::api::ml_normalize gains a fourth parameter, fold_case: bool. Rust callers must add true to keep the current behaviour: ml_normalize(text, lang, emoji_style) → ml_normalize(text, lang, emoji_style, true). Python, which takes it as a keyword with a default, is unaffected. (#559)

Fixed¶

• The two CVE rows behind "no single call" are closed, and they had to close together (#614, #615). Each was one of the exactly two vectors that made docs/security/cve-validation.md say no entry point cleared everything, so fixing one alone would have left the other failing and forced the guidance to be rewritten twice.

#614 — strip_obfuscation named confusables instead of folding them. 49 code points appear in both emoji_single.tsv and confusables_to_latin.tsv, and most are not emoji: typographic punctuation, currency, math operators, CJK brackets. They reach the emoji table from CLDR annotationsDerived, which names non-emoji characters. strip_obfuscation("€xample.com") produced "euro xample.com", so the spoof and the genuine host stopped being equal rather than becoming equal — CVE-2017-5383 surviving a preset documented as maximum-strength deobfuscation.

Not fixed the way the issue proposed. Reordering the confusable fold before demojize would break idempotency: punctuation inside emoji names (the ’ in "woman’s hat") has to be folded by the confusable pass. That ordering is documented three times and pinned by tests/test_presets.py. Instead the overlap is derived at build time as an intersection of the two tables — so it cannot drift the way a curated list would — and demojize skips those rows inside comparison presets only. Standalone demojize("I ❤ €5") still returns "I red heart euro 5", which is what that function is for. build.rs asserts the count is 49, so a table refresh that claims another confusable source fails the build instead of widening the gap silently.

#615 — canonicalize cannot cap its way out of an eclipsing mark. The anti-zalgo step is a count, and by count one Arabic shadda is indistinguishable from one acute accent, so no threshold removes CVE-2017-7833's spoof and keeps café. The discriminator that works was already in disarm's script data: strip a combining mark whose own Script is a specific script differing from its base's, and keep Inherited marks, which attach to anything. That is UTS #39's mixed-script reasoning applied per grapheme rather than per string.

It runs in canonicalize_strict only. The rule is destructive for scholarly transliteration, IPA and linguistic transcription, where marks from one script legitimately sit on bases of another — the corpus least able to notice. canonicalize is deliberately still one short, and there is a test asserting that rather than leaving it implied. Verified against nine legitimate samples in five scripts, including Arabic with its own vowel marks: all pass through completely unchanged.

The step sits after the confusable fold, and that ordering is load-bearing. Placed before it, а (Cyrillic) + U+0489 (Cyrillic mark) agrees on the first pass, then the fold rewrites the base to Latin a and the next pass strips the mark — f(f(x)) != f(x). The property test canonicalize_strict_idempotent caught it; deciding against the final base script is the only stable point.

canonicalize_strict and strip_obfuscation now each clear the whole matrix, so TestOneCall's guard is inverted rather than deleted: it asserted that closing a gap should fail loudly, it did, and it now asserts the two sufficient entry points stay sufficient while every other one stays short. The published advice is unchanged and its reason has moved — from "nothing suffices" to "the two that suffice are the two most destructive ones", which is the same conclusion for a caller who has to forward the text they cleaned.

• is_suspicious_hostname() now catches tags, variation selectors, noncharacters and PUA, and stops reporting a noncharacter as Arabic (#610). Third in the sequence after #603 (bidi controls) and #605 (zero-width). 17 of 18 sampled code points passed the screen clean and all 18 survived into canonical.

The one that did flag was flagging for the wrong reason, and it is the #605 bug in a class #605 did not cover: U+FDD0 sits in the Arabic Presentation Forms range, so the script detector read it as a letter and paypal<U+FDD0>.evil.com reported scripts=['Latin', 'Arabic'] with mixed_script=True. Widening the existing per-label strip fixes that by construction rather than by special case, because the strip already runs before detect_scripts.

Reported on the existing has_invisible field rather than four new ones, so no binding payload changes: no new field on the Ruby positional tuple and no change to the Java jni_sig! string. is_invisible_in_hostname composes the four class predicates that src/invisibles.rs already carried.

Private use and the variation selectors are included because RFC 5892 puts all four of the classes added here in DISALLOWED outright. (The pre-existing zero-width set is not uniformly disallowed — U+200C/U+200D are CONTEXTJ, conditionally permitted — and flagging those remains the deliberate fail-closed policy #605 chose, not a reading of the RFC.) Both have legitimate uses in ordinary text, so a general-text detector needs its own argument for them; that is tracked separately. Measured against the full suite including the adversarial-oracle clean corpus: no false positives.

The tag block is the reason this is a security fix rather than tidying. U+E0061–U+E007A spell arbitrary Latin invisibly, and the screen previously called such a hostname clean and returned the payload intact in canonical — the combination that turns a detector into a laundering step.

• search_key, catalog_key and sort_key no longer keep 134 characters that transliterate() deletes (#602). A character the table maps to the empty string is not unknown — it is a decision the table already made, "this has no ASCII form, drop it". ErrorMode::Preserve was excepting itself from those mappings on the reading that an empty mapping is a kind of failure the caller asked to keep, so the three presets that pass Preserve kept the characters verbatim while TextPipeline(transliterate=True), which passes Ignore, dropped them correctly.

catalog_key made it worse by running the confusable fold after transliteration, so a leaked Cyrillic soft sign was folded onto Latin b: Пьеса became pbesa, a key containing a letter that appears in neither the input nor its romanisation. It is now pesa.

ErrorMode still governs what happens to characters the table has nothing to say about, so a genuinely unmapped code point is preserved exactly as before. Verified by a full-range scan reproducing the issue's own predicate: zero leaks remain.

One property test asserted that Preserve never returns empty output. That held only because of the exception removed here, and could not have been true in general — a string of nothing but empty-mapped characters legitimately transliterates to nothing, which is why its generator already had to exclude combining marks. It is restated as the invariant that actually defines the mode: Preserve output is never shorter than Ignore output, which needs no generator exclusions at all.

• A new control anomaly kind — has_anomalies goes from 11 CVE rows to 18 (#612). A non-whitespace control (NUL, ESC, BEL, DEL, the C1 block) is never legitimate in text, and nothing reported one. strip_control_chars has removed them since #433, so the transform existed and the detector did not.

The reason they were invisible is worth recording: the introducers are plain ASCII, so the ASCII fast path in the token classifier — which exists because the invisible, bidi, zalgo and mixed-script branches can only fire above U+007F — skipped them entirely. The new branch runs before that gate.

Presence, not position. #612 framed this as an "edge" question because it started from whitespace trimming, but a control hides things wherever it sits: the last character of "malicious\u001b\\" is a backslash, so an edge-only rule would call that token clean while the escape introducer sits one place in.

The whitespace-class controls are excluded, reusing is_fold_whitespace rather than restating the set. TAB, LF, VT, FF, CR, U+001C–U+001F and NEL are real separators that collapse_whitespace folds to a space, and flagging them would fire on every multi-line string.

This closes seven rows that docs/security/cve-validation.md listed as reported by nothing: CVE-2023-24329 (leading NUL) and the whole terminal-control class (CVE-2008-2383, CVE-2019-9535, CVE-2025-55754, CVE-2024-52005, CVE-2023-43620, CVE-2023-37275). The three that remain undetected are a different shape — a fold collision, a length budget, a table lookup — so no further character class will close them, and the page now says so.

Deliberately not added: leading/trailing whitespace detection, which #612 also asked for. inspect_anomalies documents itself as flagging characters "disguising a real word", and padding disguises nothing; a kind for it would fire on ordinary text.

• The Node AnomalyKind union shipped without bidi_mixed. It was added to the Rust enum in #412 and never mirrored, so a TypeScript caller matching on it got a type error for a kind the library really returns. Nothing caught it, because the value crosses napi as a bare String and index.ts casts. Node is the only binding that restates the set — every other surface passes it through as a string — so a drift gate now reads the as_str arms out of src/anomalies.rs and compares them to the union, plus a second test asserting every kind is reachable from some input.

• PRESETS["ml_normalize"] was missing two of the nine steps it claims to describe (#600). PRESETS is a hand-maintained Python mirror of the const STEPS arrays in src/presets.rs; nothing executes it, and it had drifted. The mirror listed seven steps, omitting the transliterate step and the second demojize that #498 added after strip_accents. test_preset_steps_exact did not catch it because it compares the mirror against a literal in the test file, and both were written from the same wrong reading — so the test pinned the drift instead of detecting it. Mirror and test are now correct against Rust.

list_profiles() also said presets are "step-lists defined in Python". They are defined in Rust. That sentence is how the drift went unnoticed, so it is corrected too, and a comment on PRESETS now states what the dict is and what it is not.

A structural gate that parses the Rust arrays is not part of this change: the step lists use composite variants (FixedPoint, ConfusablesNfcFixedPoint) that the mirror flattens, so a real gate needs per-preset expansion rules and deserves its own issue rather than a fragile parser bolted on here.

• Documentation: three functions whose names promise more than they check.
• has_bidi_conflict now says plainly that it is not the RLO check (#599). It reads letters, so "invoice\u202Egpj.exe" returns False — the two conditions are disjoint and a string can satisfy either, both or neither. The docstrings route to inspect_anomalies (kind bidi) for detection and strip_bidi for removal, and note that strip_bidi does not close the real-letter case, because there is no format character to remove. docs/concepts/which-function.md gains the two bidi rows its threat-model table lacked; its only previous mention of bidi was in a cost column, so the page could not answer "how do I detect a bidi attack".
• get_pipeline() now states that profile names and PRESETS keys are disjoint namespaces, so get_pipeline("canonicalize") raising is expected rather than a bug (#600).

• ml_normalize's documented limits stopped at homoglyphs. They now cover the other two: all twelve bidi controls and every PUA code point pass through unchanged (#608). strip_control handles Cc; bidi controls are Cf. No behaviour change — the preset is tokenizer hygiene, and llm_guardrail / rag_ingest already exist for untrusted input.

• Python can now call strip_control_chars and strip_zero_width_chars directly (#616). They already existed in the Rust core (disarm::api) and in the C ABI, Java/Kotlin, Node and Ruby bindings. Python was the only surface without them, so control-stripping there meant constructing a TextPipeline rather than calling a function — unlike the ten sibling strip_* operations, four of which (strip_tags, strip_pua, strip_noncharacters, strip_variation_selectors) are narrower and are plain functions. Both are now exported, and Text gains the matching fluent methods.

The parity matrix recorded the gap as deliberate and named the substitute as collapse_whitespace(strip_control=True) — a signature that has never existed; collapse_whitespace takes only text. That record lived in PROVIDED_VIA in scripts/parity.py, so anyone consulting the matrix for the Python equivalent was sent to a TypeError. Both entries are removed and the matrix regenerated.

• collapse_whitespace gains a property test covering control characters. The existing no_leading_trailing_whitespace property draws from \PC*, which excludes controls, so the trim invariant was never tested against them. It holds: measured exhaustively over the cross product of whitespace, controls and letters for lengths 1–4, and over 200,000 random strings, with zero cases where the output starts or ends with whitespace. Reported as a trim bug in #612; that report was wrong and is retracted there. What looked like a defeated trim is the space between a leading control and the word, which is interior by the same rule that makes "a\u{0}b" keep both of its spaces. No behaviour change — the test closes the coverage gap that made the question open.

• is_suspicious_hostname() now catches zero-width and invisible characters, and no longer reports a phantom script for them (#605). Sibling of #603, for the characters that carry no direction at all — U+200B–U+200D, U+2060–U+2064, U+FEFF and U+180E. Eight of the ten passed the screen clean, and all ten survived into canonical.

The two that did flag were flagging for the wrong reason. U+FEFF sits in the Arabic Presentation Forms block and U+180E in the Mongolian block, so the script detector read each as a letter: paypal<BOM>.evil.com reported scripts=['Latin', 'Arabic'] and mixed_script=True. Right verdict, wrong evidence — and any caller keying policy on scripts was told an ASCII-looking hostname contained Arabic.

A new HostnameAnalysis.has_invisible reports the finding and is folded into suspicious. It is additive and disjoint from bidi_control (#603). The characters are removed per label, before script analysis, not on the joined hostname afterwards — so scripts, mixed_script, has_confusables and canonical are all computed on what a reader actually sees, and the phantom-script bug is fixed by construction rather than special-cased.

U+200C ZWNJ and U+200D ZWJ are flagged unconditionally. IDNA2008 CONTEXTJ permits them only in narrow joining contexts that a spoof screen has no reason to honour.

Exposed across all six surfaces (Rust core, Python, Node, Ruby, Java/Kotlin, C ABI).

• is_suspicious_hostname() now catches bidi control characters, and canonical no longer carries them (#603). Every UAX #9 bidi control — the overrides U+202D/U+202E, the embeddings U+202A–U+202C, the isolates U+2066–U+2069 and the marks U+200E/U+200F/U+061C — passed the hostname screen clean, paypal<RLO>moc.evil.com among them. The verdict was derived entirely from bidi_conflict (#412), which reads strong-direction letters and is structurally blind to a format character.

The ACE path was never affected: idna::domain_to_unicode rejects these codepoints and the decode failure already failed closed. What slipped through was the literal-Unicode label, which reached the pass-through arm uninspected — exactly the form a hostname takes in a log line, a mail header or a UI label, where the name never resolves and the display spoof is the whole attack.

Two changes. A new HostnameAnalysis.bidi_control field reports the finding and is folded into suspicious; it is additive and disjoint from bidi_conflict, whose #412 meaning is unchanged (a string can set either, both or neither). And the controls are stripped before canonical is built, so a caller who screens a hostname and then renders that field can no longer render the spoof they were told was absent.

Exposed across all six surfaces (Rust core, Python, Node, Ruby, Java/Kotlin, C ABI). The character set is now defined once, in scripts::is_bidi_control; presets::is_bidi_or_format was refactored to build on it rather than keep a second copy, so the hostname screen and strip_bidi cannot drift apart.

• ẞ uppercases to SS, not B — German is no longer corrupted (#597). #595 recovered ẞ (U+1E9E, the capital sharp S — official German orthography since 2017) by folding its TR39 prototype ß through ASCII_FOLD to b, then reconciling case. That produced B, so STRAẞE became STRABE, GROẞ became GROB and FUẞBALL became FUBBALL.

ß.to_uppercase() is the two-character SS, and that is the right answer: STRAẞE and STRASSE are the same word, so folding to SS makes them collide — which is what a skeleton is for. A genuine multi-character case mapping now wins over ASCII_FOLD. Two rows move, ẞ and Ꟗ (Middle Scots S, same prototype). Ᏸ U+13F0 is the counter-case that keeps the rule honest: a Cherokee letter shaped like B with no case expansion of its own, so it still folds to B.

tests/test_accented_latin_fidelity.py missed this because its German fixture is lowercase Straße, which was preserved throughout. It now carries the uppercase forms, the lowercase asymmetry, and the Cherokee counter-case.

• Six Latin homoglyphs now fold, and the Latin lambdas collide with the Greek one (#593). filter_latin_homoglyphs recovers a Latin-script source whose TR39 prototype is a single basic-ASCII graphic. ASCII_FOLD runs later, in generate_mappings, so a row whose prototype that table already knows was discarded before it could be consulted — the same shape as #587 and #590: a filter dropping a row before the pass that would have made it valid.

source	prototype	now
ẞ U+1E9E	ß	B
Ꟗ U+A7D6	ß	B
ꞵ U+A7B5	ß	b
ꝫ U+A76B	ȝ	z
Ꟛ U+A7DA	Ʌ	A
Ƛ U+A7DC	Ʌ	A

TR39 puts ٨ ۸ Λ Ꟛ in one confusable class, and the Latin lambdas did not collide with the Greek one they are defined to be confusable with — the class had three distinct skeletons. It reads A throughout now.

Two candidates are deliberately excluded. ţ (U+0163) and ț (U+021B) reach the same prototype, but folding them strips a cedilla and a comma-below; ț is ordinary Romanian orthography, and normalize_confusables promises accented Latin comes through intact. The guard tests the source's own canonical decomposition rather than a codepoint list, so a future confusables.txt cannot smuggle a new accented source past it. It applies only to the rows this pass newly recovers: Ç, ç and Ǿ reach a bare ASCII prototype only because strip_combining removed the mark from TR39's target, and they have folded since long before this — #586's fixed-point loop is built on Ç → C.

• Cherokee YE folds to B, not SS (#593). fix_case_mismatch uppercased the prototype before ASCII_FOLD could see it, and ß.upper() is the two-character SS, which then escaped the fold because that only fires on a single character. In a table about visual confusability, Ᏸ (U+13F0) is a B-shape. Both call sites now fold to ASCII before reconciling case; the blast radius was measured at this one row.

Latin table 2,183 → 2,189 mappings. The count gate added in #591 caught all five documented figures immediately, which is what it was for.

• The Kotlin extensions keep their published JVM signatures (#588). A Kotlin default argument compiles to one JVM method plus a synthetic $default bridge, not to an overload per arity. Adding a defaulted parameter therefore deletes the signature that shipped, and anything compiled against the previous artifact gets NoSuchMethodError — Java callers, and Kotlin callers that have not been recompiled.

It had happened twice, unnoticed both times. dev.disarm:disarm-kotlin:0.13.0 published normalizeConfusables(String, TargetScript) and analyzeHostname(String); #574 and #562 each added a default and removed one. #562 made the identical break in the C ABI, where it was caught and reverted by #580 — the C surface has a committed, drift-gated header and this one had nothing.

All seventeen public extensions with default arguments now carry @JvmOverloads, which restores both lost signatures and preserves every arity from here on. JvmSignatureTest is the gate: it reads the compiled facade by reflection rather than the source text, so adding a parameter without the annotation turns it red. The policy is recorded in BINDINGS.md.

The cost is generated methods rather than maintained ones — slugify has twelve defaults and emits thirteen arities, taking the facade to 105 public static methods.

• The tr39 digit-policy overrides now honour the ASCII contract (#587). #341 made ASCII the contract for the Latin confusable tables. The override set was written later, for #561, and never joined it: write_digit_tr39_overrides took upstream's raw target with only strip_combining applied, bypassing the ASCII_FOLD pass every value in the main table goes through. So digit_policy="tr39" put back exactly the residue #341 had removed.

Four of the 46 rows carried a non-ASCII value. Three had a clear ASCII representative and now use it:

source	TR39 target	now
٨ U+0668	Ʌ U+0245	a
۸ U+06F8	Ʌ U+0245	a
⁰ U+2070	º U+00BA	o

This was not only cosmetic. TR39 puts ٨ ۸ Λ Ꟛ in one confusable class, and the un-folded value made the class stop colliding: ٨ gave Ʌ, Λ gave A, Ꟛ gave itself. Three skeletons for one class defeats the only thing a skeleton is for. After the fix the class reads a, a, A, Ꟛ — the digits now collide with the lambda case-insensitively. Ꟛ (U+A7DA) is still unmapped, for an unrelated reason: filter_latin_homoglyphs only recovers a Latin-script source whose prototype is basic ASCII, and Ʌ is not, so that row is dropped before ASCII_FOLD is ever consulted. That is its own gap, not this one.

The fourth, ⁹ → ꝰ (U+A770 MODIFIER LETTER US), has no clear ASCII representative. Rather than ship the residue, the row is dropped and tr39 falls back to the numeric reading for that codepoint, so ⁹ folds to 9 under both policies. The override set is 45 rows, every value ASCII, and build.rs now asserts it — the assertion the sibling table blocks already carried and this one did not.

The documentation claim is corrected too. Every surface said tr39 "folds several digits to a Latin letter"; three of the 45 rows do not land on a letter — ٠ and ۰ fold to ., and 𑣣 folds to the two characters rn. That is now stated wherever the policy is documented, because a caller building a label- or path-shaped key needs to know a delimiter can appear.

• normalize_confusables now reaches a fixed point in every binding, not just Python (#586). 0.11.1 shipped #523 as "normalize_confusables is now idempotent and complete on confusable + combining-mark input". That was true of one of the two call paths. #361 had already wired the public Rust API to the single-pass normalize_confusables_cow a month earlier; #523 added the fixed-point loop to the owned Layer-1 form and its tests, and never touched src/api/safety.rs. Python reaches the core through the fixed path. Rust, Node, Ruby, Java, Kotlin and the C ABI reach it through the other one.

So the same call answered differently depending on the language, and the non-Python answer could still be confusable:

Input	Before, outside Python	Python, and now everywhere
U+04AA U+0327	U+0043 U+0327 — is_confusable says true	U+0043
U+00A5 U+0300	U+0059 U+0300	U+1EF2

For a primitive whose whole job is producing a skeleton two identifiers can be compared on, returning output that the library's own detector still flags is the failure that matters. An exhaustive sweep of the BMP crossed with composing marks finds 28 base characters affected, not just the two above.

Layer 1 gains normalize_confusables_fixed_cow, the borrowing form of the fixed-point fold, and the public API calls it. The owned form now delegates to it rather than carrying a second copy of the loop, so there is one implementation to keep correct. The borrow-on-no-op guarantee (#352) is unchanged: input with nothing to fold is already a fixed point, so the common case still never allocates.

Guarded at every level: spot cases on the Layer-2 API, parity tests in the Node, Ruby, Java, Kotlin and C-ABI suites, and a Tier-3 sweep over the BMP × composing marks for both idempotence and residual confusability. The Tier-3 entry is deliberately separate from #523's lib-level sweep, because that sweep tests Layer 1 — testing the layer below the one the bindings call is how this survived a year.

• ⁰ and ᵒ now fold; a block-range gap had been dropping them (#590). is_latin_or_common in gen_confusables.py enumerates Latin block ranges and jumps from 0x007F to 0x00C0, leaving U+0080–U+00BF uncovered. º (U+00BA) is category Lo with Script=Latin and lives in that hole, so filter_direct read it as a non-Latin target and discarded both upstream rows pointing at it.

The symptom was an asymmetry between the only two superscript digits upstream TR39 carries — it lists visual lookalikes, and no letter resembles a superscript four:

	before	now
⁰ U+2070	unmapped, passed through	0 (numeric), º (tr39)
⁹ U+2079	9 (numeric), ꝰ (tr39)	unchanged

⁹ targets ꝰ in Latin Extended-D, so its row survived and the #89 digit rule rewrote it to the ASCII digit. ⁰ was discarded before that rule could run. With the row restored, ⁰ gains a tr39 override too, so the pair is now symmetric.

The fix admits only the three Latin letters in the gap — ª µ º — and gives ASCII_FOLD the two ordinal indicators, keeping #341's ASCII contract. Opening the whole range instead would pull in 58 rows targeting punctuation and symbols (·, °, ¶, ©), which is what that contract exists to prevent. is_latin, the source-side predicate, has the identical gap; closing it there was measured to change nothing, since no upstream row uses those three as a source, so it is documented rather than changed blind.

Table sizes move accordingly: Latin 2,181 → 2,183 mappings, tr39 overrides 45 → 46. Every documented count was re-measured against the regenerated tables rather than adjusted by hand, which turned up two that were already stale before this change — the Latin table was described as ~2,063 mappings (actual 2,183) and the Cyrillic as ~1,369 (actual 1,349).

tests/test_doc_table_counts.py existed to prevent exactly that drift, and the two files it gates were accurate. The three surfaces carrying the same figure were not gated, and had drifted by 118 rows unnoticed: src/tables/confusables_data.rs, python/disarm/_api.py, and the target-script table in the confusables user guide. All three are now gated against the same source of truth, taking the check from 5 figures to 11.

• Restored the 1-argument disarm_analyze_hostname C ABI (#580). #562 widened it to take contractions, but that symbol shipped in 0.13.0 and callers are linked against the 1-argument form — widening it breaks them at link time. The contraction pass moved to a new disarm_analyze_hostname_opts(host, contractions), with the original delegating to it, matching disarm_transliterate / _opts and disarm_normalize_confusables / _opts.

Found by the header drift gate added in this same change, on its first real run against a moved main — which is the argument for the gate. Nothing else in CI could see it: the smoke test regenerates the header and compiles against it in one step, so a widened signature plus a matching call-site change is self-consistent and passes.

• sanitize_filename is now a fixed point on the first pass (#570). The trailing-dot trim ran in finalize_name, after the extension split it invalidates. Input ending in a . — literal, or produced by transliteration (· U+00B7, … U+2026) — had that dot taken as the "extension", leaving an earlier dot inside the stem; trimming it then moved the boundary, so the next call split elsewhere and stripped a separator that had become stem-trailing. sanitize_filename("a*.b.") gave "a_.b", then "a.b".

The trim now runs before the split, so the boundary the split sees is the one the output has. finalize_name is unchanged and still required — the extension branch re-prepends '.', and it owns the empty / "." / ".." fallback.

The guarantee asserted is stronger than idempotence: a caller sanitizes once, so if the first pass returned something a second pass would change, the single-pass answer was already wrong. Two systems sanitizing a different number of times derived different filenames from one input, which defeats dedup on sanitized names.

Found by the Hypothesis tier, which ran nowhere automatic — see the nightly workflow below.

• Restored a Kotlin test lost in a rebase. DisarmKtTest.coverageIntrospection, added with the coverage-introspection API (#563), was silently dropped when that branch was rebased through a 14-file conflict. The Kotlin source survived, so nothing failed to compile and the loss was invisible until the JVM surface was audited for this change.

• 16 confusable rows the generator was silently dropping (#558). scripts/gen_confusables.py's filter_latin_homoglyphs pass required the TR39 prototype to be a single basic ASCII letter. That quietly excluded every Latin-script letter whose prototype is an ASCII digit or punctuation mark — Ʒ→3, Ȣ→8, Ꝯ→9, ǃ→!, Ɂ→?, ꝸ→&, ꞉→:, ꞌ→' and eight more. Nothing distinguished them from the þ→p / ſ→f rows already in the table except the category of the target, so this was a table gap rather than a policy decision. The predicate is now is_basic_ascii_graphic and the 16 rows are folded.

This is the letter-impersonates-a-digit direction only. The reverse — a digit source folding to a look-alike letter — is still guarded by enforce_digit_target (#439); normalize_confusables("०") remains "0".

Whitespace is deliberately excluded from the widening: TR39 folds the whole Zs/Zl/Zp family to a space, but collapse_whitespace already owns that from an explicit core-defined set (#433), and a second copy in the confusables table would be a divergent duplicate of the whitespace policy.

The remaining residue is now triaged and written down in docs/provenance.md rather than inferred: 5 deliberate ASCII skeleton divergences, 16 whitespace rows owned elsewhere, and ~4,300 sources whose upstream target is non-Latin, for which a to-Latin table is the wrong home. unmapped_confusables() (#563) makes the split recomputable at any time, so the closed gap cannot silently reopen.

Documentation¶

• The required status checks are named correctly again. CONTRIBUTING.md, AGENTS.md and SECURITY.md told contributors to wait for "Rust checks passed" and "Python checks passed". #583 collapsed those contexts into a single roll-up, so neither exists: branch protection requires "All checks passed", "DCO sign-off" and "iai estimated-cycles gate". A contributor following the old text waits for a check that will never report.

• The four drift gates are described in one place. CONTRIBUTING.md → Drift gates now names what each guards and when it fails: the committed disarm.h (#580), the 11 documented row counts in test_doc_table_counts.py (#591), the build.rs ASCII assertions (extended to the tr39 overrides earlier in this release by #587), and JvmSignatureTest over the published Kotlin JVM signatures (#588). Each entry names the CI check that reports it, so a red run leads straight to the gate. Documentation only — no gate changes behaviour here. Two of them read a build product rather than source text, which is why they catch what source-level assertions miss.

It also records the habit those gates exist to enforce: when you regenerate a table, read the data diff, not just the test output. A generator change can silently remove rows and leave the suite green — which is how an over-broad filter deleted Ç → C during #593.

• The Tier 3 listing matches tier3.yml again. CONTRIBUTING.md documented two of the five steps the workflow runs; exhaustive_grapheme (#174), exhaustive_confusables (#586) and the lib-level ignored sweep were missing. AGENTS.md gained the confusables entry and the note on why it is deliberately separate from the Layer-1 sweep.

• Accented-Latin fidelity is a strip_obfuscation property, not a confusables property (#564). normalize_confusables preserves accented Latin where strip_obfuscation destroys it, at identical homoglyph recovery — because strip_accents sits in the strip_obfuscation bundle, not in the confusable primitive. Nothing in the docs said so, so a reader could reasonably conclude that disarm destroys accented Latin as a matter of course, and a benchmark cell measuring the bundle could be read as measuring the fold.

docs/security/adversarial-defense.md gains a "What each entry point costs you" section: the worked comparison, the structural explanation, and a threat-model → entry-point → cost table covering all six entry points. The confusables user guide gains the short version with a cross-link. Both are doctested, and tests/test_accented_latin_fidelity.py pins the claims — including that the loss is attributable to strip_accents specifically, and that ml_normalize folds no confusables at any fold_case setting, so it is not a homoglyph defence.

Filed and fixed together with #559 because both have the same shape: a destructive step baked into a bundle with no documented route to the non-destructive path.

[0.13.0] — 2026-08-17¶

Added¶

• Full HostnameAnalysis across the Node, Ruby, and Java/Kotlin bindings (#549). A new analyzeHostname / analyze_hostname returns the complete analysis (verdict + all granular signals, including whole_script_confusable) — previously those bindings exposed only the .suspicious boolean. Node returns a HostnameAnalysis object, Ruby a Hash, and Java/Kotlin a dev.disarm.HostnameAnalysis record (the List<List<String>> labelScripts and List<Boolean> labelWholeScriptConfusable are marshalled directly across the JNI boundary). The boolean isSuspiciousHostname predicate is unchanged. Mirrors how the anomaly report is already exposed. The C-ABI gains the matching structured entry points in the same window (see below).
• Structured reports across the C-ABI (#553). Five new #[ffi_export] entry points return their report as a JSON string (freed with the existing disarm_string_free): disarm_analyze_hostname (the full HostnameAnalysis, including whole_script_confusable / label_whole_script_confusable), disarm_inspect_anomalies (per-finding kind/token/start/end/detail/reason, taking a JSON word-array lexicon so the leet/segmentation branches match the other bindings, not just the structural ones), disarm_inspect_auto_lang (script + chosen language + discriminators), and the fallible disarm_lang_info / disarm_script_info metadata lookups. JSON is the one transport for every nested shape (List<List<String>>, List<Boolean>, optionals) — no repr(C) mirror structs, trivially parsed by any C/Go/Swift/ ctypes consumer. serde_json is a C-ABI-only dependency; the pure core still carries no serde. The scalar predicates (disarm_is_suspicious_hostname, …) are unchanged.
• Whole-script-confusable signal on HostnameAnalysis (#545). Two additive fields — whole_script_confusable (any label qualifies) and the per-label label_whole_script_confusable — name the fact that discriminates a whole-script spoof (аррӏе.com → skeleton apple.com, every letter a confusable) from a genuine non-Latin domain (москва.рф, whose м/к/в survive the skeleton). A label qualifies when it is single-script, non-Latin, and its confusable skeleton is entirely Latin. It is a graded signal, not a verdict, and is deliberately not folded into suspicious: on its own it fires on short non-Latin ccTLDs (ру→py) and on real words whose every letter is a confusable (оса→oca). The precise, low-false-positive policy — whole_script_confusable(non-TLD label) ∧ Latin TLD — is caller-side (disarm does not model registrable boundaries). Exposed on the Rust and Python surfaces; the other bindings expose only .suspicious today and are tracked separately (#549).

Documentation¶

• Clarified that is_suspicious_hostname's suspicious flag is a maximally conservative screen (an any-character confusable test flags essentially every non-Latin hostname), not a precise verdict, and moved whole-script confusables in THREAT_MODEL.md from out of scope to a defined mechanism with its stated irreducible false-positive class. Completed the HostnameAnalysis field table in the predicates docs.
• Upgrading guide + stability-contract clarifications (#546, #547, #548). Added docs/upgrading.md (a new top-level nav section, distinct from Migration) with the cumulative table of public renames since 0.9 and a !!! danger note on the is_safe_hostname → is_suspicious_hostname boolean-polarity inversion. Restated SECURITY.md's supported-version window as a self-maintaining rule (was the stale 0.6.x). Extended the semver data-change clause in docs/RUST_API.md to name the security surfaces (is_suspicious_hostname, normalize_confusables, …), recorded the bundled Unicode/UTS#39 data versions in docs/provenance.md (+ provenance headers on the two confusables tables), and clarified that "removed in 1.0" refers to the RELEASING.md commercial-support milestone, not the next release.
• Surfaced disarm's measured BitAbuse recovery in the coverage docs (#543). Re-ran the adversarial-eval harness against the full corpus on v0.12.0 (325,580 rows) — strip_obfuscation now recovers 65.3% word-level (up from 64.1% on 0.6.3) with 81.7% of non-ASCII perturbation occurrences folded — and added disarm's own row to the coverage spectrum in docs/security/adversarial-defense.md and THREAT_MODEL.md (previously only the ~35% class baseline and ~96% ceiling appeared, inviting readers to transfer ~35% onto disarm). The word-level metric is defined inline with line-exact (5.8%) stated alongside, and the BitAbuse figure is explicitly separated from the near-identical TR39-space XMR = 0.634. Retired the divergent pre-harness baseline in the benchmark README in favour of the committed report, and documented a manual pre-release refresh cadence.

[0.12.0] — 2026-08-15¶

Added¶

• JVM bindings — Java + Kotlin, on Maven Central (#540, closes #43). disarm now ships to the JVM ecosystem: dev.disarm:disarm (idiomatic Java API, a fat JAR with per-platform JNI native libraries for macOS/Linux/Windows on x86-64 and aarch64) and dev.disarm:disarm-kotlin (extension functions + default arguments). The full core surface is exposed — transliteration, confusable folding, slugify / sanitizeFilename, reusable Pipeline and Lexicon handles, grapheme/script queries, and the structured anomaly/inspection reports. As a new binding, this is a lockstep minor across every registry (crates.io / PyPI / npm / RubyGems / Maven Central).
• First-class C-ABI substrate (dev.disarm disarm-cabi / disarm_ffi). A reusable, unsafe-free C ABI over the core (via safer-ffi), consumed by the JVM binding and the basis for the forthcoming Go bindings (#47). The JNI layer and C ABI carry no unsafe (safe handle registry, jni_mangle-generated exports).

Documentation¶

• Documented the full set of version-bump sites a release touches (#525).

Internal¶

• Routine dependency and CI-action updates (dependabot).

[0.11.1] — 2026-07-13¶

Fixed¶

• ml_normalize is now idempotent on NFKD-exposed symbol bases (#498). The "cldr" preset was non-idempotent on negated-relation symbols whose NFKD decomposition strips a combining overlay to expose a nameable base (e.g. ≇ U+2247 → ≅ U+2245 + U+0338 overlay). Because demojize ran before accent-stripping, the freshly-exposed base was only named on a second call, so ml_normalize(x) != ml_normalize(ml_normalize(x)) across the enumerated 17-member negated-symbol class. A second CLDR demojize pass now runs right after accent-stripping, so bases exposed within a call reach a true fixed point in a single call.
• normalize_confusables is now idempotent and complete on confusable + combining-mark input (#523). Confusable folding and canonical composition interact both ways — a fold can expose a composition (¥+◌̀ → Y+◌̀ → Ỳ) and a composition can expose a new fold (Ҫ+◌̧ → Ç → C, since Ç is itself a confusable) — so a single pass was not always a fixed point, and could even leave a residual confusable in the output. The fold/compose pass now iterates to a fixed point, restoring both idempotency and completeness. Guarded by a new exhaustive Tier-3 test over every confusable × combining-mark pair (~9M).

Documentation¶

• Documented unidecode() Cyrillic soft/hard-sign collisions (#511). The compatibility unidecode() path maps the Cyrillic soft sign (ь) and hard sign (ъ) to the empty string, so otherwise-distinct inputs can collide; the limitation is now called out with a pointer to the script-aware transliterate(…, lang=…) path that preserves them.

Internal¶

• Binding publishers are gated on the published core (#500). The npm and RubyGems publish jobs build their native addon/gem against disarm as a crates.io dependency, so on a release they now wait for the core crate to land on crates.io before building instead of racing it — removing the manual re-run every prior release required. No library-behavior change.
• Held phf/phf_codegen at 0.13 to preserve MSRV 1.81 (#510). phf 0.14 moves to edition 2024 / Rust 1.85; a dependabot group now keeps the pair in lockstep and pins it below 0.14 until the MSRV is deliberately raised.

[0.11.0] — 2026-06-21¶

Performance¶

• Transliterate recovers most of the form-invariance compose-at-lookup regression. The #475/#477/#481 boundary added a needs_composition pre-scan that walked every non-ASCII input a second time (UTF-8 decode + is_combining_mark trie lookup per character) before transliterating, so the hot path paid two full passes where it used to pay one — the latin/unidecode comparator ratio fell ~18 → ~11 across #474 → #480. The mark/jamo detection is now fused into the engine's existing decode loop: the fast pass bails to the compose path only when it actually meets a combining mark, so mark-free input (the common case) makes a single pass. needs_composition (still used by the confusables fold) also drops the trie lookup for a range fast-path over U+0000–058F. Behaviour is identical (verified by the exhaustive, formal, and form-invariance suites); Rust-level micro-bench gains, pre → post: latin −33%, cyrillic −20%, mixed −21%, greek −12% ns/char.

Added¶

• Strip invisible & non-interchange code points in the security presets (#413). The presets a service puts in front of an LLM, a logger, or a denylist now neutralize the dominant 2024–25 "ASCII smuggling" channels and the adjacent non-interchange classes that survive NFKC and the existing zero-width passes: the Unicode Tags block (U+E0000–U+E007F, including the previously-missed U+E0001), variation selectors, the Combining Grapheme Joiner (U+034F, a denylist-evasion blocker), noncharacters, and the Private Use Area; the Braille Pattern Blank (U+2800) now folds to a space rather than surviving as invisible padding. None of this is a blanket delete — a well-formed emoji subdivision flag (U+1F3F4 … U+E007F) is preserved, and display_clean keeps the VS15/VS16 presentation selectors after a base and preserves the PUA (icon fonts), while the comparison presets (security_clean, normalize_user_input, strip_obfuscation) strip it. Four standalone helpers — strip_tags, strip_variation_selectors, strip_noncharacters, strip_pua — are exposed across the Rust core and the Python, Node, and Ruby bindings for composing policy directly. Output change: the comparison presets now remove these classes; idempotency is preserved (a terminal NFC recomposes any base+mark adjacency a strip creates).

• Bidi-direction conflict detection (has_bidi_conflict, #412). A new primitive that flags text mixing strong left-to-right and strong right-to-left characters — the precondition for Unicode Bidi display-reordering and the structural signal behind "BiDi Swap"-style spoofs (an LTR brand label stacked on an RTL domain, varonis.com.ו.קום). Unlike a U+202x override check, it fires on the real letters. Derived from disarm's own script ranges (no new table); exposed across the Rust core (disarm::api::has_bidi_conflict) and the Python (has_bidi_conflict, Text.has_bidi_conflict), Node (hasBidiConflict) and Ruby (Disarm.bidi_conflict?) bindings.

• HostnameAnalysis direction fields (#412). The Python HostnameAnalysis gains bidi_conflict (folded into suspicious), cross_label_script (the broader, non-folded cross-label fact), and label_scripts (per-label resolved scripts, left to right) for position-aware caller policy.

• Anomaly detection: has_anomalies / inspect_anomalies (#389). An out-of-place-character detector: it flags text disguising a real word via a cross-script homoglyph, leet, single-letter segmentation, a zero-width / bidi control, or zalgo, and reports a technical fact, not intent (like the hostname analysis). Built on the core's own primitives plus a caller-supplied common-word lexicon (used only by the leet/segmentation branches; the others are script-agnostic). Exposed across the Rust core (disarm::api) and the Python, Ruby, and Node bindings, with a per-language usage page. A dated defensive publication — published as prior art so the method stays freely usable.

• Reusable anomaly lexicon handle (Lexicon). The binding has_anomalies / inspect_anomalies functions rebuilt a hash set from the caller's word list on every call; a new opaque Lexicon class lets callers build the set once and reuse it across many calls (disarm.Lexicon(words) in Python, new Lexicon(words) in Node, Disarm::Lexicon.new(words) in Ruby). Both functions accept either the raw word collection (unchanged, back-compatible) or a Lexicon. The Rust core already amortizes this (it takes &HashSet<String>), so this closes the gap only the FFI bindings had.

• Node.js docs + doc-example gate (#44). A docs/node/ getting-started page and API reference plug into the language-neutral structure (#50), with Node.js added to the Getting started and API Reference nav. Every Node // => example is executed against the built addon by scripts/check_doc_node_examples.mjs — the Node analogue of the Sybil/Rust/Ruby doc gates — wired into the node CI job (which now also triggers on docs/**), so the examples can't rot.

• Node.js binding (#44). A new bindings/node/ napi-rs addon exposes the pure-Rust core to Node with a fully-typed, idiomatic TypeScript surface — camelCase functions, options objects with sensible defaults, string-union token types, and a DisarmError / DisarmInvalidArgument class hierarchy. It covers the full plain-function surface (transliterate, confusables, slugify, normalization, text cleaning, graphemes, filenames, reverse/untranslatable, script analysis) and ships .d.ts types. Two layers, like the gem: a raw napi shim (src/lib.rs) under a hand-written index.ts. Built + vitest-tested in CI against the in-repo core (the #374 drift gate, now node/"Node checks passed"), with a publish-node.yml release workflow (per-platform prebuilds + npm provenance) so npm i disarm needs no Rust toolchain.

• Ruby: filename, reverse-transliteration, and script-analysis ops (#375). Completes the plain-function parity backfill: sanitize_filename (platform:/max_length:/preserve_extension:), reverse_transliterate(lang:) (:el/:ru/:uk), find_untranslatable (→ { char:, offset: } hashes), detect_scripts, mixed_script?, and inspect_auto_lang (→ a :script/:chosen_lang/:reason/:discriminators_hit hash) — thin wrappers over the core disarm::api.

• Ruby: grapheme-cluster operations (#375). The binding gains grapheme_len, grapheme_split, grapheme_truncate, grapheme_width, and terminal_width — user-perceived-character counting/splitting/truncation and East Asian Width display measurement (ambiguous_wide: false by default), thin wrappers over the core disarm::api. Continues the Ruby↔core parity backfill (#375) and unblocks the graphemes Ruby docs.

• Ruby: normalization + text-cleaning primitives (#375). The binding gains normalize / normalized? (NFC/NFD/NFKC/NFKD), collapse_whitespace, strip_control_chars, strip_zero_width_chars, strip_bidi, and strip_zalgo / zalgo? — the first batch of the Ruby↔core parity backfill (#375), which unblocks honest normalization/text-cleaning Ruby docs. Each is a thin keyword-argument wrapper over the core disarm::api, carrying the core's defaults (normalize(form: :nfc), strip_zalgo(max_marks: 2), zalgo?(threshold: 3)).

• CI: the Ruby binding is built and RSpec'd against the local core on every PR (#374). A new ruby job in ci.yml compiles the gem (Ruby 3.1–3.3) and runs rake spec against the in-repo core — not the published one — on any PR that touches the binding or the core it wraps. It injects a CI-only [patch.crates-io] redirect so an unreleased core API change is actually exercised; the registry-core build in publish-ruby.yml is unchanged. A core change that breaks the gem (like the 0.10 tuple→struct return that shipped a broken gem, #364–#367) now fails the new "Ruby checks passed" gate on the PR that introduces it, not silently at release.

• CI: the docs' Rust and Ruby usage examples are now executed gates (#50). The per-language usage tabs are no longer illustrative — each is run in CI, the way the Python tabs already are (Sybil). scripts/check_doc_rust_examples.py extracts every ``rust doc block, compiles and runs it against the pure core with#![deny(unused_must_use)](so an example that discards its result fails);scripts/check_doc_ruby_examples.rbevals every Ruby# =>line against the freshly-built gem. The Rust gate runs in theDoc testsjob; the Ruby gate runs in the Ruby workflow, now also triggered ondocs/**`. Catches the signature/output drift that the tabs introduced (which had shipped as non-compiling Rust until this gate).

• Ruby: transliterate now accepts a lang: language profile. Previously the Ruby binding's transliterate exposed only scheme:, so it could not reach the core's per-language profiles (a parity gap vs Python/Rust). lang: accepts a String or Symbol and composes with scheme: — e.g. Disarm.transliterate("Київ", lang: :uk) # => "Kyiv". Implemented over the core's Transliterate builder via a generalized _transliterate_opts shim.

Changed¶

• Re-point Greek small letter iota U+03B9 to the i-class, reverting #343 (#436). #343 had re-pointed the bare iota from i/і to the l/vertical-bar class (l/ӏ) to unify {ι, ӏ, ا}. That split the iota family — the accented iotas U+03AF (ί) and U+03CA (ϊ) still folded to i in the same table — and was shadowed in the security presets: security_clean and strip_obfuscation run NFKC first, which decomposes the accented iotas to bare iota, so under #343 the whole family folded to l there. It also contradicted the upstream Unicode TR39 mapping (03B9 → 0069, i.e. i) and missed the dominant spoof — normalize_confusables("bιtcoin") returned "bltcoin" instead of colliding with "bitcoin". The bare iota now folds to i (latin)/і (cyrillic), consistent with its accented forms, so the entire iota family folds to the i-class under normalize_confusables and the NFKC-first presets, and the ι-for-i spoof is caught (bιtcoin → bitcoin). The genuine full-height bars — ӏ (palochka U+04CF), ا (alef U+0627), and the U+2502/U+FFE8 bars (#245) — stay in the l-class. The only confusable-table change is the single iota row in each target.

• security_clean and normalize_user_input no longer neutralize path separators (#431, reverses #248). The presets previously rewrote / and \ to _ and collapsed .. runs so the output was safe to drop into a filesystem path. That is sink-specific output sanitization — out of scope for the canonicalization presets per THREAT_MODEL.md — and it corrupted legitimate input: URLs, file paths, and any /- or \-bearing string came back mangled ("https://example.com/path" → "https:__example.com_path"). The presets now pass separators through verbatim. Upgrading: if you fed preset output straight into a filesystem path, defend traversal at the sink instead — call sanitize_filename on the final path component, or validate against your own allowlist. A confusable fraction/division slash that NFKC folds to a real / is still normalized to / (that is canonicalization working as intended); it is just no longer rewritten away. The internal neutralize_path_separators helper is removed.

• collapse_whitespace folds the full whitespace set and the blank-rendering code points; control/zero-width stripping is now a separate step (#433). collapse_whitespace was category-driven and also deleted controls and zero-width characters inline. It now folds whitespace only, to a single space, over an explicit core-defined set: the line controls (TAB/LF/VT/FF/CR), the information separators (U+001C–U+001F), NEL, the Zs/Zl/Zp spaces, and a blank-rendering set that category detection cannot reach — U+2800 Braille blank and the Hangul fillers U+115F/U+1160/U+3164/U+FFA0 (e.g. aㅤb → a b). Breaking: collapse_whitespace drops its strip_control / strip_zero_width parameters (Rust, Python, Node, Ruby) — it no longer deletes anything. Compose strip_control_chars / strip_zero_width_chars before it for the old behaviour; the presets do this internally, so their output is unchanged except for the line-control fix below. strip_control_chars now preserves the whitespace controls (CR/VT/FF/NEL/U+001C–U+001F) so the fold can turn them into a space; it still removes NUL, DEL, and the rest of the C0/C1 block. The PRESETS metadata now lists the explicit strip_control / strip_zero_width steps.

• security_clean now caps combining marks (anti-zalgo, #429). The preset left zalgo-stacked tokens intact, so a mark-stacked admin did not match its base form in a denylist/dedup comparison security_clean is meant to canonicalize. It now caps combining marks at 2 per base (the same threshold normalize_user_input already used), removing abusive stacking while preserving legitimate diacritics — security_clean stays accent-preserving (café → café, Việt → Việt; full accent folding remains in search_key/sort_key). The cap runs after the invisible/control strip so a stripped character between marks cannot split a run and hide the count (#121), and idempotency is verified by the raw-equality property test. Output change: inputs with more than two stacked marks per base are now capped.

• is_suspicious_hostname and has_anomalies now flag bidi-direction conflicts (#412). These detectors strengthen as disarm grows. A hostname that mixes strong-LTR and strong-RTL characters (the "BiDi Swap" shape, e.g. varonis.com.ו.קום) is now flagged suspicious via the new bidi_conflict signal — previously it slipped past mixed_script (which is per-label) and was only caught incidentally, if at all. The anomaly detector gains a bidi_mixed finding kind for a token mixing strong-LTR and strong-RTL letters: it is the precise, reorder-capable subset of mixed_script and additionally catches non-Latin RTL mixes (e.g. Cyrillic+Hebrew) the Latin-anchored mixed_script rule could not see. Behaviour change: some inputs that previously reported mixed_script (Latin+Hebrew/Arabic) now report bidi_mixed, and some that reported clean now flag. bidi_conflict=False / no bidi_mixed is not a safety guarantee.

• sort_key now preserves base accented characters (#99.1). sort_key is documented as a collation key — accented forms should stay distinct so the accent survives for ordering — but it shared search_key's full transliteration pass, so it ASCII-folded every accent ("Über" → "uber") and produced output identical to search_key. It now transliterates only non-Latin scripts, preserving Latin accents (sort_key("Über") → "über", sort_key("Café") → "café") while still folding Cyrillic/Greek/etc. to a consistent Latin form ("Война и мир" → "voyna i mir"). search_key and catalog_key are unchanged — they still fold accents for exact-match lookup and dedup. A language profile no longer expands an accented Latin letter in a sort key (sort_key("Über", lang="de") is "über", not "ueber"). Output change: persisted sort keys for accented-Latin input will differ from 0.10 and should be regenerated. Applies across the Rust core and the Python, Ruby, and Node bindings.

• Docs: synced the public XMR benchmark claims to the v2 note (#399). The README, the docs landing page, the adversarial-defense page, and the unidecode-migration guide led with the v1 curated-set headline (XMR = 1.000 on the hand-curated pairs). They now lead with the v2 broad-sample measurement over the 1,314 single-codepoint TR39 sources whose skeleton is a single Latin letter: instance XMR 0.634 / 0.682 (95% CI) with ~95% per-source coverage (stated as a distinct quantity), plus the NFKC (0.103) and TR39-skeleton-oracle (1.000, by construction) baselines, citing the v2 DOI 10.5281/zenodo.20618323. The curated 1.000 is retained only as a labeled sanity check, and the curated set is described correctly (18 hand-curated Cyrillic pairs; the 19 Greek pairs were a separate experiment). CITATION.cff is bumped to 0.11.0 with the note DOI.

• Docs: Node.js usage tabs across the guide pages (#44). The twelve guide pages that carry Python/Rust/Ruby tabs now also show a runnable Node tab — 38 tabs in all, matching the Ruby coverage. Every Node example is executed against the built addon by the doc gate (scripts/check_doc_node_examples.mjs).

• Docs: completed the language-neutral restructure (#50). The Adversarial-Text Defense concept page now shows Python/Rust/Ruby usage tabs (no bare Python), and the stale untabbed user-guide/getting-started.md was removed in favour of the per-language getting-started guides (now linked from the index nav). With every published binding carrying install + quickstart + API and mkdocs build --strict clean, all four #50 acceptance criteria are met.

• Docs: Ruby usage tabs across the guide pages unblocked by the parity backfill (#375/#50). The normalization, text-cleaning, graphemes, filenames, and language-detection guides now show a runnable Ruby tab beside Python and Rust — 17 tabs in all. Every Ruby example is executed against the built gem by the doc gate, so the tabs cannot rot.

• Docs: language-neutral scaffold — first phase of the docs restructure (#50). Reshaped the documentation IA toward "language-neutral concept core + per-language specifics": a neutral landing headline (no longer "for Python") that routes by ecosystem; per-language Getting started pages under docs/python/, docs/rust/, and docs/ruby/; a shared docs/concepts/which-function.md concept page (lifting the #328 decision table into the neutral layer); and an mkdocs.yml nav reorganized into Getting started / Concepts / Guide / API Reference (Python · Rust) / Architecture / Migration / Reference / Project. Folded six previously orphaned pages into the nav. No library behaviour change; the per-topic concept/usage split and per-language example tabs land in following phases.

• Docs/metadata: scope transliterate() vs the TR39 confusable functions (#328). The headline identity led with "TR39 confusable analysis", while the most discoverable function, transliterate(), performs the opposite mapping — phonetic BGN/PCGN romanization (Cyrillic р → r), not TR39 visual confusable folding (р → p). Clarified across every entry point with no behaviour change: the identity one-liner (README, docs/index.md, Cargo.toml, pyproject.toml, mkdocs.yml, CITATION.cff) now says visual confusable analysis and phonetic transliteration; a new "Which function do I want?" decision table sits near the top of the README and docs landing page; and transliterate()'s docstring (hence docs/api/transforms.md) and the README Quick Start block now state it is romanization, not homoglyph defense, pointing to normalize_confusables() / strip_obfuscation() for the latter.

Deprecated¶

• Presets renamed to mechanism names; old names deprecated (#430). The three presets whose *_clean / normalize_user_input names overpromised safety — flagged as documentation defects in THREAT_MODEL.md — are renamed to names that describe their mechanism. The rename is byte-stable (old(x) == new(x) for all inputs):

Old name (deprecated)	New name
security_clean	canonicalize
display_clean	strip_format
normalize_user_input	canonicalize_strict

The old names remain as deprecated aliases across every binding — Rust (free functions + DisarmStr methods, #[deprecated(since = "0.11.0")]), Python (each emits a DeprecationWarning; the Text builder's .security_clean() / .display_clean() methods and the PRESETS keys are aliased too), Node (securityClean, @deprecated), and Ruby (Disarm.security_clean, warns with category: :deprecated). They are removed in 1.0. catalog_key, search_key, sort_key, ml_normalize, and strip_obfuscation are unchanged.

Fixed¶

• Hardening-review follow-ups (M-2, M-3, L-1, L-2). A pass over the 2026-06-20 deep review closed four small correctness/perf/security gaps: (M-2) the eager normalize_confusables no longer unconditionally allocates and rebuilds the string on a pure-ASCII / already-folded no-op — it now delegates to the borrowing form, sharing its borrow-on-no-op fast path; (M-3) that borrowing form skips the needs_composition char-decode scan on pure-ASCII input via a cheap is_ascii() short-circuit; (L-1) the Slugifier / UniqueSlugifier default= constructor kwarg — which crosses the str→Rust boundary in __init__, outside the @_surrogate_safe-guarded __call__ — is now WTF-8→UTF-8 scrubbed, so a lone-surrogate default no longer raises UnicodeEncodeError (closing the last gap in the #476 contract); (L-2) the fast-path guard's Confusables step now sets its marks bit itself rather than relying on a preceding Nfkc step, so a hypothetical Confusables-only preset can't let the guard skip a decomposed homoglyph the fold would recover. Per-cluster allocation in compose.rs (L-6) is also removed via a reused NFC scratch buffer, and a stale generator comment claiming U+0344 is "unmapped" is corrected (it maps to the empty string, so its output-neutral row is emitted, not skipped).
• normalize_confusables is idempotent on an excluded singleton followed by an unrelated mark. The compose-at-lookup pass (#481) recovered a composition-excluded precomposed singleton (ড় U+09DC = ড + nukta) only by a whole-cluster widening-map lookup. When such a singleton was followed by an unrelated combining mark (a visarga), the cluster's .nfc() decomposed the singleton (ড় ◌ঃ → ড nukta visarga) and the trailing mark made the lookup miss the 2-char ড nukta key — so it stayed decomposed. The fold then oscillated (a bare ড় composes, ড় + mark decomposes), i.e. nc(nc(x)) != nc(x), surfaced by a normalize_confusables_idempotent proptest seed. The lookup now matches the widening map by greedy longest prefix at each position in the cluster (bounded by a build-time-emitted EXCLUDED_COMPOSITIONS_MAX_KEY_CHARS), so the excluded head recomposes and any trailing marks are kept — idempotent and form-invariant. The mark-free hot path and the common single-mark cluster are unchanged.
• sanitize_filename never returns an empty name or a . / .. directory reference; leading/trailing dot hygiene (#485, #487). Three correctness gaps with one root: the extension branch re-prepended '.' and was exempt from the stem's dot trim. (1) The empty string bypassed the never-empty fallback and returned "" — os.path.join(dir, "") targets the directory, a write-target footgun. (2) Trailing dots/spaces survived ("report..." -> "report.", "CON." -> "_CON."), which Windows then silently strips at the filesystem layer. (3) A separator-then-dot-like input reduced to a bare "." ("_" + U+00B7 -> "."), a current-directory reference. A shared finalize_name now runs on the fully assembled name: it trims leading and trailing dots and spaces, and falls back to "_" for an empty, ".", or ".." result — so the output is always non-empty, never a directory reference, and never a leading/trailing-dot dotfile, across both return paths. A 50-case attacker battery (path traversal, Unicode separator homoglyphs, control/NUL, RTLO/bidi, the ADS colon, dot hygiene, the separator-plus-dot class) and non-emptiness/idempotency property tests lock the defenses in. (A separate idempotency gap from the extension-split boundary moving between passes is tracked in #488.)

• Class-based entrypoints honor the malformed-Unicode (surrogate) contract (#476). The #469 boundary adapter wrapped every module-level _core callable, but the class entrypoints that cross the str → Rust boundary on construction or in a method were not covered: Lexicon(["a\ud83d…"]) raised UnicodeEncodeError, and so did calling a Slugifier / UniqueSlugifier / TextPipeline on surrogate-laced text. They now apply the same WTF-8 → UTF-8 scrub-and-retry: the three callable classes guard their __call__, and Lexicon (a frozen, non-subclassable PyO3 class) is guarded by a metaclass proxy whose construction scrubs while isinstance still recognizes every real handle, so the has_anomalies / inspect_anomalies prebuilt-handle dispatch is unaffected. The dynamic surrogate audit is extended to enumerate the exported classes (covered or reviewed-exempt), so a future class with a text surface fails the audit rather than silently skipping the contract.

• Hangul romanization is invariant to the input's normal form (#483). A precomposed syllable run was romanized with inter-syllable spaces (처리 → "cheo ri"), but the same text decomposed to conjoining jamo (NFD) romanized contiguously ("cheori"), so the output depended on the normal form. Conjoining jamo are General_Category=Lo, so #479's General_Category=Mark compose-at-lookup gate never fired on them. The fix composes an L + V [+ T] jamo run into its syllable by the standard Unicode index arithmetic (no table, no normalization pass), gated on a cheap jamo range check, as a sibling to the existing mark-composition path — so the decomposed form takes the same per-code-point path as the precomposed one. transliterate, slugify, unidecode, and slugify_unicode now agree across NFC/NFD/NFKD on Hangul; the precomposed output ("cheo ri") is unchanged. Partial jamo (a lone L, or L + T with no vowel) are left alone. Cosmetic spacing only — both forms always recovered the same Korean reading; this was the last NFC/NFD gap on the transliterate path.

• Close the raw-vs-normalized residual the #477 oracle could not see (#481). The form-invariance audit compared the normal forms against each other but never against the raw precomposed input, so a composition-excluded code point passed green while still degrading: Devanagari क़ U+0958 transliterated "qa" raw, but its canonical decomposition KA + nukta is composition-excluded, so every normal form degraded to "ka" (the mark dropped). Closed entirely with build-time data, no runtime canonicalization pass (so the #478 decompose-then-recompose regression class cannot recur): an exclusion-inclusive compose map widens #479's compose-at-lookup so a base+mark exclusion reaches its precomposed scalar (KA+nukta → QA, shin+sin-dot → שׂ U+FB2B, Tibetan vowel stacks, the Hebrew presentation forms), and the two real Greek-oxia confusable singletons (U+1F77/U+1F79 → i/o) become char-table rows. The map is gated on the precomposed target being mapped, which keeps an unmapped operator (FORKING U+2ADC = NONFORKING + U+0338) from cycling with the NFKC recovery. The audit now asserts f(raw) == f(NFC) == f(NFD) == f(NFKD) for the transliterate family and confusable detection, with a small characterized tail (two Greek accent-punctuation code points, and the benign spoof-resolutions where normalizing a look-alike to its genuine character — Kelvin U+212A → K — flips detection). The ~1,027 non-target singletons (ά U+1F71 vs U+03AC, the same Greek letter, neither a Latin confusable) are deliberately left as benign re-encoding: normalize_confusables is a targeted fold, not a normalizer.

• Confusable folding, detection, and transliteration are invariant to the input's normal form (#475, #477). The confusables maps and the transliteration tables are keyed per code point on the precomposed form (ї U+0457 → i / yi), so a decomposed input (і U+0456 + combining diaeresis U+0308) reached only the base entry and the mark survived — an attacker could evade the recovery, or flip is_confusable, just by sending NFD. The confusables fold and detect (normalize_confusables, is_confusable) and every public str → str recovery entrypoint — transliterate, unidecode, and the whole slugify* family (including the Unicode-preserving slugify_unicode) — now compose each base + combining-mark cluster at lookup time, so the result is invariant to the input's normal form (f(NFC(x)) == f(NFD(x)) == f(NFKD(x)), and likewise for the is_confusable predicate). The composition is compose-only (it never decomposes): a composition-excluded presentation form such as Hebrew שׂ U+FB2B keeps its own table entry (→ s), where a naïve "NFC the input first" would have decomposed it and changed the output. It is gated on a combining-mark check, so mark-free input (ASCII, CJK, precomposed letters) keeps its borrow/zero-allocation fast path; it composes the full cluster (Vietnamese ệ, polytonic Greek ᾷ, and Brahmic two-part vowels like Bengali ো). A self-guarding audit enumerates the public entrypoints and asserts the invariant, so a future entrypoint that forgets to normalize fails the test, not in production.

• Digit confusables fold to their digit, not a look-alike letter (#439). The confusable maps mapped many non-ASCII digit sources to letters or punctuation — Arabic-Indic ٠→., ١→l, ٥→o, Devanagari/Bengali/NKO zeros→o/O, and the Unicode 16 outlined digits 𜳰→O / 𜳱→l. The root cause: gen_confusables.py classifies digits via unicodedata, so running it under a Python whose Unicode table is older than the bundled confusables.txt silently mis-folds any digit that table doesn't yet know. The generator now (a) folds every Nd digit source to its canonical ASCII digit and (b) refuses to run under a Unicode table older than the data (warning on any mismatch). The maps are regenerated: every digit spoof now canonicalizes to the plain digit (٠/०/𜳰 → 0), keeping numbers numeric (the llm_guardrail "digits are never remapped to letters" guarantee).

• sort_key / search_key / catalog_key are now idempotent across scripts and cases (#419). The transliterating key presets ran transliterate before fold_case, so a cased letter whose folded form is in the table but whose original is not — e.g. a Georgian Mtavruli capital Ჱ (U+1CB1), absent from the table, folds to Mkhedruli ჱ which transliterates to he — only transliterated on the second pass, violating f(f(x)) == f(x). fold_case now runs before transliterate so both passes see the same form. search_key/catalog_key additionally fold again after transliterate, since full transliteration can emit uppercase ASCII (£→GBP, №→No) that the pre-fold can't reach — those keys are now lowercase and stable. Output change: a few currency/symbol inputs that previously produced uppercase keys now fold to lowercase. Idempotency is pinned by per-preset property tests.

• security_clean / normalize_user_input idempotency on duplicate combining marks (#434, #416 residual). A duplicate combining mark broke the single NFC → confusables → NFC sandwich: NFC composed only one mark onto the base, the TR39 fold dropped it, and the recomposing NFC reattached the spare mark — re-creating a foldable composed character the next call would consume, so f(f(x)) != f(x) ("c"+◌̧+◌̧ → "ç" then "c"). The confusable fold is now iterated to a fixed point (each pass removes ≥1 mark, so it converges in a couple of iterations), making both presets true fixed points. The #416 Hypothesis idempotency property is re-broadened and the normalize_user_input Rust proptest strengthened from nfc-modulo to raw equality.

• Line controls no longer join tokens in collapse_whitespace (#433). TAB and LF folded to a space, but VT, FF, CR, NEL, and the information separators (U+001C–U+001F) were deleted — so a + CR + b became ab while a + LF + b became a b. All of them are Unicode whitespace; deleting them was an invisible-join (coalescence) vector. They now all fold to a single space, so a\rb → a b. The blank-rendering Braille and Hangul fillers, which category detection passed straight through, are folded too.

• security_clean / sort_key idempotency on invisible-separated combining marks (#416). When an invisible code point separated a base character from a combining mark (e.g. "a" + U+200B + combining acute + "b"), the leading NFKC passed over the still-separated mark and the later zero-width strip then left the base and mark adjacent but decomposed — so the composed form appeared only on the second call, violating the documented f(f(x)) == f(x) invariant (which THREAT_MODEL.md classifies as a vulnerability). An NFC pass after the strips now recomposes the adjacency on the first call, in the Rust core, so every binding inherits it. For security_clean a second, deeper cause was also fixed: TR39 confusable skeletoning is not normalization-stable (it drops the diacritic on some composed accented letters — ç→c, ø→o — but not the decomposed form, and can emit a decomposed skeleton like Ý→Y+◌́), so the confusable fold is now sandwiched between two NFC passes and the pipeline is a verified fixed point under a strengthened raw-equality proptest. Output change: for these previously non-idempotent inputs the first call now returns the composed NFC form. sort_key was affected only because it began preserving accents in #411 (search_key/catalog_key, which fold accents away, were never affected). A separate, pre-existing sort_key non-idempotency (transliterate-before-fold-case on a case pair) is tracked in #419.

Internal¶

• Dependency-freshness audit across every manifest + full dependabot coverage. Dependabot only watched the root cargo/uv/actions manifests, so the binding crates rotted a full major unseen (napi 2→3, magnus 0.7→0.8). .github/dependabot.yml now watches every manifest — the core crate and both binding workspaces (cargo), the Node package (npm), and the Ruby bundle (bundler) — and a new dev-time scripts/audit_dependencies.py audits all of them against their registries in one command (--strict to fail on a major lag), run weekly by the dependency-audit workflow. The guard makes any future config gap visible instead of silent. See DEPENDENCY_UPGRADES.md. The DCO check now exempts trusted GitHub App bots (*[bot] authors, e.g. dependabot[bot]) — matching the official DCO app's default — so dependabot's PRs can finally satisfy branch protection and auto-merge instead of every bump being silently blocked.

• The Tier 3 exhaustive+formal gate now guards every publish, not just PyPI/crates.io (#159, #395). The pre-publish regimen — the exhaustive Rust domain tests (#[ignore]) and the Python formal invariants (@pytest.mark.formal) — moved out of an inline job in publish.yml into a reusable workflow_call workflow (.github/workflows/tier3.yml) that all four publish paths depend on: the PyPI wheel, the crates.io core, the RubyGems gem, and the npm addon. Previously only the wheel and the core were gated, so a release whose core failed the exhaustive net could still ship the bindings. Also wired the exhaustive grapheme-integrity suite (exhaustive_grapheme, #174) into the gate alongside exhaustive_transliterate — it was documented "run before release" but had never actually been in the release workflow.

• Binding publish workflows build against the in-repo core on non-publish events (#374, #396). publish-ruby.yml's test job and publish-node.yml's build job compiled the binding against the published core, so a pre-release binding that calls a core API not yet on crates.io (e.g. has_anomalies before this release) failed to build on every PR/push — red on main until the matching core shipped. They now apply the same CI-only [patch.crates-io] redirect to the in-repo core that ci.yml's drift gate uses, but only on push / pull_request; on release / workflow_dispatch the shipped gem and prebuilt addon still build against the published core, unchanged.

• Node binding: bumped vitest 3 → 4, dropping a vulnerable dev-only esbuild (#392, #394). The Node binding's test runner pulled in esbuild 0.27.7 — a dev-only transitive dependency, never part of the published npm package — which carried two HIGH advisories (GHSA-gv7w-rqvm-qjhr, GHSA-g7r4-m6w7-qqqr). vitest 4 pulls vite 8, which demotes esbuild to an optional peer dependency, so the vulnerable package drops out of the resolved tree entirely (npm audit reports zero vulnerabilities). The Node test matrix is unchanged (20/22).

[0.10.0] — 2026-06-15¶

The multi-language milestone (epic #326): disarm becomes a publishable, pyo3-free Rust crate with a first-class idiomatic Rust API, gains a Ruby binding, and adds opt-in diagnostic logging — all over a single shared pure-Rust core. The Python package is unchanged for callers (same import disarm surface); the work is the core extraction and the new non-Python surfaces.

Added¶

• Pure-Rust core, published to crates.io (#38, #42). The default build is now the pyo3-free core (default = []); the Python extension is the opt-in extension-module feature, so cargo add disarm pulls a clean Rust library with no libpython in its dependency tree (enforced by a CI gate: the default cargo tree -e no-dev tree must contain no pyo3, matched case-insensitively). The codebase is organized in three layers: Layer-1 pub(crate) algorithm cores, Layer-2 the public disarm::api, and Layer-3b the feature-gated pyo3 shims — all consuming one implementation.
• Idiomatic Rust API (disarm::api) (#352, #361, #362). The semver-governed crates.io surface: typed enums (TargetScript, Scheme, NormalizationForm, UrlComponent, Platform, ReverseLang) that each round-trip via as_str/Display/FromStr; the Transliterate builder with Scheme / OnUnknown (which carries its replacement in the Replace(String) variant); an opaque Error with a stable ErrorKind/code(); Cow<'_, str> borrow-on-no-op returns; a graphemes() iterator; the SlugConfig builder; the DisarmStr extension trait for method-call syntax; named #[non_exhaustive] struct returns (EncodingDetection, DecodedText, HostnameAnalysis, Untranslatable — no anonymous tuples); and a guarded process-global registration API (register_lang / register_replacements / remove_replacement / clear_replacements / seal_registrations) that enforces the registration cap and the one-way seal latch. Two contract tests fail CI if a pub fn ever returns a tuple or a token enum loses its round-trip.
• Ruby bindings — the disarm RubyGem (#45, #357). A magnus-based native extension wrapping the pure-Rust core (no Python), with an idiomatic Ruby surface: keyword arguments with defaults, symbol tokens (:latin, :strict_iso9, …), a single transliterate(text, scheme:), and a Disarm::Error < StandardError hierarchy. Precompiled platform gems (Linux x86_64/aarch64, macOS x86_64/arm64, Windows) install with no local Rust toolchain.
• Opt-in, binding-neutral diagnostic logging (#208, #358). Behind the log / log-content features (off by default), the core emits structured records at API boundaries via the log facade — zero cost when off (the macros compile to nothing) and never inside a per-codepoint hot loop (enforced by a source-scan test). Default-level records carry metadata only (lengths, counts, flags, durations, error codes — never input or output content, enforced by a redaction sentinel test); the log-content TRACE escape hatch routes its truncated samples through disarm's own strip_log_injection (dogfooding) so a log line can never forge a record.

Changed¶

• Native module renamed disarm._disarm → disarm._core (#42). The public Python API is unchanged — callers import disarm. The native module name is an implementation detail the public surface doesn't require; the package's own internals (and the type-stub drift checks) reference disarm._core directly, so any consumer reaching into it should update the path.

Fixed¶

• Confusables: cross-script ASCII folds and additive Greek/Cyrillic pairs (#341, #342, #343), plus the halfwidth vertical form U+FFE8 residue (#245).
• Terminal width: corrected the additivity-across-space precondition (#279).

Security¶

• HAI-SDLC hardening pass over the Rust core (#360): a deep multi-pass review (0 critical / 0 high) actioned into 21 fixes — tightened a hostname IPv6-literal zone-id check, added limit-rejection logging, a unique-slug truncation-error fix, and an allocation-free is_normalized, among others.

Internal¶

• Wired Tier 3 (exhaustive + formal) into the release/publish gate (#159, epic #326). publish.yml now runs a tier3 job on the release/publish trigger that executes the exhaustive Rust domain tests (cargo test --no-default-features --test exhaustive_transliterate -- --ignored) and the Python formal invariants (pytest -m formal, against a freshly built wheel). Every wheel/sdist build job and the publish job needs: it, so a Tier-3 failure blocks the upload to PyPI — closing the gap where these tiers were a manual pre-release step. They remain excluded from fast PR CI; the #[ignore] / @pytest.mark.formal markers are untouched.
• Split the 1,200-line src/api.rs into cohesive submodules (api/{safety,text,transliterate,presets}.rs) re-exported from api/mod.rs, with the DisarmStr trait in the hub (#361). No public-path change.
• translit-rs 0.8.2 redirect shim published so the old PyPI name points users at disarm (#264 follow-up).

[0.9.1] — 2026-06-13¶

Added¶

• strip_log_injection(text, *, replacement='\ufffd', keep_tab=False) (#307). A stateless, character-level encoder that makes untrusted text safe to write as a log line: it replaces CR/LF/NEL/LS/PS (record forging), NUL/C0/C1 controls (parser corruption), and ESC/DEL (terminal hijack via ANSI escapes) with replacement (default U+FFFD). \t is neutralized by default (keep_tab=False) to block TSV/logfmt column injection. Idempotent; ASCII-clean fast-path returns the original object; never emits a raw CR/LF/ESC. It owns the log-record and operator-terminal sinks but makes no HTML-log-viewer-safety claim (that is stored XSS — encode at the viewer with escape_html) and is not a log4shell defense (see Threat Model).
• escape_html(text) and percent_encode(text, *, component) output encoders (#311). Standalone terminal encoders applied at the output sink — deliberately not TextPipeline/PROFILES steps (a pipeline is context-free; baking encoding in invites double-encoding and wrong-context escaping). escape_html escapes the five HTML metacharacters for element/quoted-attribute context (ASCII fast-path returns the original object; not idempotent by design). percent_encode does RFC 3986 percent-encoding for a required Component (PATH/SEGMENT/QUERY/FORM; UTF-8 byte-based, ASCII output, FORM uses space→+). Both are mechanism-named and carry the #306 scope-boundary discipline: they are the narrow, context-pinned exception to "disarm is not an output sanitizer," not a general XSS/injection defense (see Threat Model).

Changed (breaking)¶

• Renamed is_safe_hostname() → is_suspicious_hostname() and inverted its boolean. The old name asserted a safety it cannot guarantee — safe=True only meant "no mixed-script label and no bundled-table confusable found," yet whole-script spoofs and out-of-table confusables still returned safe=True (the false-assurance pattern #306/#308/#309 removed elsewhere, but as a literal safe boolean a caller branches on). The function now returns (suspicious, analysis) where suspicious=True means a problem was detected; the result struct SafeHostnameDetails → HostnameAnalysis, field safe → suspicious (inverted). The granular scripts / mixed_script / has_confusables / canonical fields are unchanged. No alias — invert call sites: safe, d = is_safe_hostname(h) → suspicious, a = is_suspicious_hostname(h). (#313)
• Renamed policy profile web_input_sanitize → normalize_web_input. Follows the sanitize_user_input → normalize_user_input rename: "sanitize" wrongly implied output/injection safety, and was especially misleading here because this profile is lighter than normalize_user_input() (NFKC + confusables only; no bidi/zero-width/control/zalgo stripping). Use get_pipeline("normalize_web_input"). No alias is kept.
• Renamed sanitize_user_input() → normalize_user_input(). The old name implied output sanitization (injection safety); this preset performs input Unicode normalization only and is not an XSS/SQL defense (see Threat Model). The PRESETS registry key changes to match ("normalize_user_input"). No alias is kept — update call sites directly.

Documentation¶

• Stated the XSS/injection scope boundary explicitly (#306): README, the docs site, and THREAT_MODEL now say plainly that disarm normalizes input and is not an output sanitizer — it performs no HTML/JS/SQL/shell escaping and never replaces context-aware output encoding at the sink (NFKC can even surface ASCII metacharacters from fullwidth lookalikes). This boundary is the conceptual basis for the renames and the new output encoders in this release.

Security¶

• Supply-chain hardening (#260): added cargo deny (license allow-list, banned/wildcard crates, crates.io-only sources via deny.toml) to the required Rust checks passed gate, alongside the existing cargo audit. Releases now attach a CycloneDX SBOM (*.cdx.json) of the Rust dependency graph, and PyPI distributions carry PEP 740 build-provenance attestations via OIDC Trusted Publishing. Verification is documented in SECURITY.md.
• Bumped pyo3 0.24 → 0.29, resolving two upstream advisories: GHSA-36hh-v3qg-5jq4 (HIGH — out-of-bounds read in nth/nth_back for PyList/PyTuple iterators) and GHSA-chgr-c6px-7xpp (MEDIUM — missing Sync bound on PyCFunction::new_closure closures). Includes the binding-layer API migration the bump requires (GIL with_gil/allow_threads → attach/detach, PyObject → Py<PyAny>, downcast_exact → cast_exact); no functional change to any transform. (#315)

Internal¶

• Docs: build the MkDocs site in CI and deploy to Cloudflare Pages (served at the unchanged docs.disarm.dev), replacing the Read the Docs trigger. mkdocs build --strict runs in GitHub Actions (Python-only — mkdocstrings parses source statically); push to main deploys production, PRs get preview deploys. Legacy /en/latest/* URLs 301 to root via docs/_redirects. Removed .readthedocs.yaml and RTD_TOKEN. (#314)
• CI: replaced the custom conversations-resolved.yml workflow with GitHub's native Require conversation resolution before merging branch-protection setting. The bespoke "Conversations resolved" status check (#55) was flaky — stale check runs lingered after threads were resolved and blocked otherwise-green PRs. Behavior is unchanged (unresolved review threads still block merge), now enforced by the built-in gate instead of a workflow + required status check.

[0.9.0] — 2026-06-11¶

The first release under the disarm name — the continuation of translit-rs (last released as 0.8.1). See #264 for the rename rationale. The 0.0.0 entries on PyPI / crates.io / npm are name-reservation placeholders, not releases; 0.9.0 is the first functional disarm release.

Changed¶

• Renamed the project from translit to disarm (#264). This unifies the distribution and import names under a single disarm:
• PyPI distribution translit-rs → disarm; import translit → import disarm.
• Native module translit._translit → disarm._disarm; crate translit → disarm.
• Console script translit → disarm.
• Breaking: the public base exception TranslitError → DisarmError (the subclasses InvalidArgumentError / ResourceLimitError / UnsupportedError keep their names). DisarmError remains a ValueError subclass, so except ValueError keeps working.
• Breaking: the context-dictionary environment variable TRANSLIT_DICT_DIR → DISARM_DICT_DIR.
• Canonical URLs moved to https://disarm.dev / https://docs.disarm.dev; the repository moved to https://github.com/raeq/disarm.

Fixed¶

• uv.lock now declares requires-python = ">=3.10", matching pyproject.toml (it had drifted to >=3.9 after the 3.10 floor landed in #277).

[0.8.1] — 2026-06-11¶

The final translit-rs release and the close of the 0.8 performance-hardening arc. The project continues as disarm from 0.9.0 (#264); 0.8.1 exists to publish honest, production-true benchmark numbers before the rename.

Changed¶

• Benchmarks now run in the fresh-string regime (#277, #302): every timed call receives a newly constructed str, the way production traffic always does. The prior cached-object measurement let CPython's per-object AsUTF8 cache hide ~105–137 ns/call of UTF-8 encode cost that only translit pays (pure-Python comparators never call AsUTF8), flattering it. JSON records now carry regime: fresh-string/v2; pre-flip history is the cached v1 regime and must not be compared across regimes.
• README short-string figures updated to the measured fresh-regime values: ~17× vs Unidecode (Latin), ~14× (mixed scripts), ~13× (Cyrillic/Greek); ~65 ns ASCII passthrough; the four-cell Unidecode-own sweep still holds (~1.3× on Unidecode's strongest case to ~25×), with a methodology note explaining the regime.

[0.8.0] — 2026-06-11¶

A performance and hardening release. The headline is a benchmark-gated optimisation programme (#233) that makes short-string transliterate roughly 15–21× faster than Unidecode (up from ~7–9×) and beats Unidecode on its own benchmark, while shrinking the library's static and resident memory. Alongside it, a Unicode-security hardening sweep tightens is_safe_hostname, the security presets, and the stateful slugifiers. Most changes are behaviour-preserving; the exceptions are called out under Upgrade notes.

Upgrade notes¶

• Minimum Python is now 3.10 (was 3.9). The extension targets the stable-ABI floor abi3-py310, so a single wheel runs on 3.10+ and the per-call Python→Rust path crosses the boundary only once (#277). Python 3.9 wheels are no longer produced.
• is_safe_hostname now flags every mixed-script label as unsafe (#254), not only the four Latin-paired high-risk combinations. A label combining two scripts with no Latin confusable (e.g. Greek + Cyrillic) previously reported safe=True; it now returns safe=False. This also flags benign combinations (e.g. Latin + CJK) — read the mixed_script / scripts fields if you need a more permissive policy. The check fails closed by design.
• Security presets no longer synthesise path separators (#248): confusable characters that normalise to /, \, or .. can no longer pass through the security/filename presets to forge path structure.
• rag_ingest now runs the confusables step (#258): Unicode homoglyph spoofs are canonicalised during RAG ingestion instead of surviving it. Output of the rag_ingest preset may change for homoglyph-bearing input.
• Stateful slugifiers validate lang at construction (Slugify, UniqueSlugify), closing the gap the 0.7.0 validation pushdown missed (#257); an invalid lang= now raises instead of being silently ignored. UniqueSlugify also honours property mutations made after construction (#249).
• Auto-language discriminator behaviour was reconciled with its documented contract (#253) — auto-detection results may differ for a few ambiguous inputs.
• Correctness edge cases fixed (#255), which may change output: reverse transliteration of all-caps digraphs and a grapheme_truncate overflow case.

Performance¶

• Short-string transliterate: ~15–21× faster than Unidecode (#277). A call now crosses the Python→Rust boundary exactly once with Rust-side keyword defaults, extracts UTF-8 zero-copy, and returns already-ASCII input as the original str object via a borrowed Cow — roughly 70 ns with no allocation.
• Beats Unidecode on its own benchmark (#281): translit wins all four cells of Unidecode's expect_ascii/expect_nonascii × ASCII/non-ASCII matrix, including Unidecode's strongest (ASCII-passthrough) case.
• Smaller static tables (#237): the default BMP transliteration table became a two-level page-table + interned-blob trie (~1 MB → ~58 KB), hanzi→pinyin a dense interned array (~600 KB → ~50 KB), and the 11,172 Hangul romanisations a single packed blob. No runtime data loading; no unsafe.
• Zero-copy context dictionaries (#238): the Arabic/Persian/Hebrew dictionaries are read once and indexed by (offset, len) spans instead of parsed into nested HashMaps of owned strings — roughly halving their resident memory. Lookup is binary search; the two-step bigram path allocates no per-token key.
• Linear-time scanning via Aho-Corasick (#242): global and slug replacements use longest/first-match automata instead of repeated per-position probing; the UniqueSlugify collision counter is amortised; and multi-codepoint emoji are matched through a code-point trie.
• Per-character hot-loop improvements — resolve-once language tables, block-table dispatch, ASCII-run skipping (#235); fewer copies on the ASCII/identity path (#236); chunked batch extraction that caps peak memory (#239); single-pass strict mode, O(u)→O(1) in time and space (#240); further ASCII fast-paths and removal of O(n·k) scans (#252).
• A benchmark harness with a deterministic iai-callgrind estimated-cycle gate guards every PR against regressions in CI (#234).

Note: the batch (list[str]) API's advantage over a Python loop has narrowed for short strings now that a scalar call is ~70 ns — for tiny inputs it is at rough parity. Its durable value is the single GIL-released crossing (thread parallelism), not a raw per-call speedup. See docs/performance.md.

Added¶

• TextPipeline(preset=…) constructor and related new-surface ergonomics (#259).
• CLI: slugify honours --lang; the strip_bidi / strip_zalgo steps are exposed; error output is cleaned up (#250).
• The errors parameter annotation now includes "strict" in the callable-module and Text wrappers (#247).

Changed¶

• docs/performance.md rewritten so every claim is CI-executed (Sybil) or linked to a recorded measurement, with a stated margin policy, varied scenarios, a prominent "where we are slower" section, and a credit paragraph for Unidecode and its lineage (#291).

Internal¶

• Resource-limit constants centralised in a single src/limits.rs module so the library's resource posture has one audit surface (#256).
• Cross-cutting Rust-core helpers (apply_replacements, emit_warning) de-duplicated (#251).
• Incorrect docstring examples in the Python wrapper modules corrected (#246).

[0.7.0] — 2026-06-10¶

A feature and architecture release. Headlines: a unified, catchable exception hierarchy; terminal column-width measurement (terminal_width / grapheme_width); native errors="strict" transliteration; LLM/RAG guardrail pipeline presets; and a substantial push of validation and configuration logic down into the Rust core, so the upcoming multi-language bindings inherit one behaviour instead of reimplementing it. Most changes are behaviour-preserving; the exceptions are called out under Upgrade notes.

Upgrade notes¶

• Exceptions now form a hierarchy. Every library error subclasses TranslitError, with InvalidArgumentError, ResourceLimitError, and UnsupportedError beneath it. TranslitError remains a ValueError subclass, so existing except ValueError keeps working. Several error message strings were enriched/standardised (#186, #187) — code matching exact message text may need updating; code matching exception types is unaffected.
• lang= is validated even for ASCII input (#197). A binding-side ASCII fast path previously skipped language validation, so transliterate("abc", lang="zz") silently returned the input; it now raises InvalidArgumentError, matching how non-ASCII input always behaved.
• slugify_filename / Slugify(safe_chars=…) output corrected (see Fixed): slugify_filename("My Report.pdf") now returns "My_Report.pdf", not "My.Report_pdf". Output for inputs that use safe_chars may change.
• New modes: errors="strict" for transliterate (#184) and decode_to_utf8(strict=True) (#189).

Added¶

• terminal_width / grapheme_width (#224): terminal column width per grapheme cluster (UAX #11 East Asian Width). Wide/fullwidth and emoji-presented clusters are 2 columns; combining marks, controls, and zero-width characters are 0. Ambiguous characters are 1 by default, or 2 with ambiguous_wide=True. Width data is generated at build time from the pinned UCD (no runtime data, no unsafe). Measures cells, not pixels; tabs are not expanded.
• errors="strict" + find_untranslatable (#184): strict transliteration raises on the first untranslatable character (reporting it and its byte offset); find_untranslatable returns all of them without raising.
• Guardrail pipeline presets (#139): TextPipeline gains strip_bidi and strip_zalgo steps and the llm_guardrail / rag_ingest named profiles for LLM/RAG input sanitisation.
• get_pipeline / list_profiles (#229): the named policy-profile registry now lives in the Rust core; the Python helpers are thin wrappers over it.
• decode_to_utf8(strict=True) (#189): raise on lossy/replacement decoding instead of silently substituting U+FFFD.

Changed¶

• Unified exception hierarchy (#183): the Python error surface is a TranslitError base with categorised subclasses; sites that previously raised bare ValueError are unified (foundation laid in 0.6.3 via #181).
• Validation moved into the Rust core (#185, #217, #229, #230, #231): enum validation, the transliterate() argument-conflict matrix, non-negative max_length / max_graphemes checks, safe_chars, and min_confidence range-checking now live in the core, so other bindings enforce the identical contract without reimplementing it. The Python layer keeps only type guards.
• Actionable error messages (#186, #187): weak messages now name the offending value, list valid options, and suggest a "did you mean…?" where applicable; message style is standardised across the surface.
• Error cause chains (#188): wrapped errors surface the underlying cause via __cause__ rather than flattening it into the message.
• TextPipeline step ordering (#174) is derived from a single source of truth, removing drift between configuration and execution order.
• All-ASCII preset fast path (#198): presets skip the NFKC pass for pure-ASCII input (behaviour-preserving).

Fixed¶

• slugify_filename / Slugify(safe_chars=…) preserved safe characters at the wrong positions — slugify_filename("My Report.pdf") returned "My.Report_pdf" instead of the awesome-slugify-correct "My_Report.pdf". safe_chars are now handled natively in the Rust core: kept verbatim and treated as word characters so they hold their position (#156, #230). The prior test only covered a dot-free input, so the bug was uncaught; regression tests now cover filenames with extensions, multiple dots, and UniqueSlugify + max_length.
• slugify(default=…) is now sanitised through the same slug pipeline (so a caller-supplied fallback cannot smuggle path-traversal or URL metacharacters into output documented as URL-safe), threads through the stateful Slugifier / UniqueSlugifier forms, and a negative max_length now raises a catchable InvalidArgumentError on both the scalar and batch paths instead of an uncatchable OverflowError (#193, #169).
• Low-severity hardening bundle (#200): eight small robustness fixes (bounds, overflow, and edge-case handling) gathered into one pass.

Security¶

• The RustSec advisory audit (cargo-audit) now blocks merge via the required "Rust checks passed" gate on every PR — an advisory can land on a dependency without any code change here (#195).

Removed¶

• Docker image build/publish and its Trivy CVE scan (#138). translit is a pip install-first library; previously published images remain as historical artifacts, but no new ones are produced. Install the CLI via pip install translit-rs.

Documentation¶

• Executable cookbook (#154, #91, #140, #156, #172): a Sybil doc-test harness with a CI gate, unidecode→translit migration recipes, an "LLM pipelines" page, a tokenizer-preprocessing page, and an anti-rot lint that turned 307 decorative # => claims into checked assertions.
• normalize-first canonicalisation recipe (#174) and a formal-verification assurance taxonomy (#223 — proof-by-exhaustion / structural / property-tested, tagging each I1–I7 invariant), plus grapheme-integrity property tests (#174).
• The project adopted the Developer Certificate of Origin (#165); all commits are signed off. The custom-emoji-provider 9-codepoint window cap is now documented (#199).

[0.6.3] — 2026-06-08¶

A correctness, maintenance, and architecture-foundation release. No output-affecting changes — every fix is behaviour-preserving and the one new public behaviour (slugify(default=...)) is opt-in. Headline: a pure-Rust error model is now in place, laying the foundation for the multi-language bindings on the roadmap.

Upgrade notes¶

• No output-affecting changes. Existing output and every exception type/message are unchanged.
• New opt-in: slugify(text, default="…") returns the fallback when the input has no sluggable characters (emoji / punctuation / zero-width) instead of "". default=None (the default) preserves the prior empty-string behaviour.

Added¶

• slugify(default=...) — opt-in fallback for inputs that would otherwise slug to the empty string, closing an empty-slug routing hazard (#97).

Fixed¶

• PRESETS["strip_obfuscation"] metadata now reflects the real pipeline order (confusables runs after demojize), matching src/presets.rs (#141).
• Lock-poison recovery now emits a Python UserWarning naming the recovered table, instead of a silent stderr line (#117).
• docs/api/exceptions.md corrected — TranslitError inherits from ValueError (not Exception), and every example message string now matches the real output (#182).

Changed (internal — behaviour-preserving)¶

• Error model (#181, part of #180): a pure-Rust Error enum (thiserror) with a stable code() per variant and a single From<Error> for PyErr boundary; ~35 error sites migrated off in-core PyErr construction. Removes the core↔PyO3 coupling and lays the foundation for non-Python bindings. Python exception types and messages are unchanged.
• Dependencies: phf / phf_codegen 0.11 → 0.13, criterion 0.5 → 0.8, chardetng 0.1 → 1.0 — each migrated and verified behaviour-preserving (#146, #153, #164).
• build.rs now auto-discovers language override tables — adding a language is just dropping in a translit_lang_*.tsv (#74).
• Generated .pyi stubs are now guarded by a stub/binary signature drift-check, which caught and fixed 18 stale stub signatures (#76).

Maintenance¶

• Split python/translit/__init__.py (2,683 lines) into _api.py + _presets.py (#73).
• Split tests/integration_transliterate.rs by script family (#75).
• Process: a required "Conversations resolved" merge gate (#55); a documented dependency-upgrade methodology with Dependabot cooldown + auto-merge (DEPENDENCY_UPGRADES.md, RELEASING.md).

[0.6.2] — 2026-06-07¶

A correctness, security, performance and maintenance release triaged from a post-0.6.1 issue sweep (#101–#132). No public API removed; one small new public behaviour (slugify(save_order=True) now functions). Two output-affecting fixes — see Upgrade notes.

Upgrade notes (output-affecting)¶

• slugify(save_order=True) was an accepted no-op; it now strips only leading/trailing stopwords (preserving interior word order), matching python-slugify (#118). If you passed save_order=True, slug output changes.
• decode_to_utf8 default min_confidence 0.5 → 0.95 (#103). The old default was inert (the detector only reports 0.50/0.95, and 0.50 < 0.50 is false), so it never rejected. It now requires high confidence by default; pass min_confidence=0.0 to accept any guess. (No practical change today — the detector currently always reports 0.95.)

Fixed¶

• #102 — UniqueSlugify no longer panics across the FFI boundary on a multibyte separator + small max_length (byte slice landed mid-codepoint; now uses floor_char_boundary).
• #101 — context bigram disambiguation tier was unreachable (it reset on every inter-word space); it now resets only on hard boundaries, so the tier fires in normal prose.
• #104 — set_emoji_provider now obeys seal_registrations() (the provider swap previously defeated the seal).
• #103 — decode_to_utf8 default confidence now actually gates (see notes).
• #107 — a corrupt context dictionary now reports a distinct "corrupt" error instead of the misleading "not found" remedy (DictState enum).
• #121 — PRESETS["sanitize_user_input"] now reflects the real pipeline order (strip invisibles before zalgo); Python registry and Rust doc aligned.
• #129 — Text.transliterate() stub now declares the tones/context parameters the implementation accepts.
• #131 — Slugify(uids=...) emits a correct wrong-class warning rather than a spurious deprecation warning.
• #122 — disambiguated the _compat should_warn nested ternary.

Security¶

• #105 — added a cargo audit (RustSec advisory) CI job and a cargo Dependabot ecosystem.
• #132 — added a Trivy CVE scan of the published image to the release workflow (SARIF → Security tab, fails on fixable HIGH/CRITICAL) + .trivyignore.
• #106 — Rust diagnostics now route through Python warnings instead of bare eprintln!, so applications can capture/suppress them.

Performance (output-preserving)¶

• #108 codepoint-range diacritic checks in tokenize(); #109 mem::take per token boundary; #110 single ch.nfkc() pass on the NFKC fallback; #111 lowered MAX_CAPACITY_HINT 256 MiB → 8 MiB; #112/#113 emoji matching uses stack buffers + a fixed sliding window (no per-char Vec/String); #114 slugify uses Cow (no eager to_owned); #115 context tokenize() returns borrowed (Cow) slices of the input — zero per-token allocation (Rust API: the crate-internal context::Token.text changed from String to Cow<'_, str>; no effect on the Python API); #116 clamped the ContextDict capacity hint.

Maintenance¶

• #118 implemented slugify(save_order=True); #119 SlugConfig::from_pyargs dedupes the four slugify PyO3 entrypoints; #120 _build_slug_kwargs helper; #123 seal-enforcement docs on each tables:: mutator; #124 infallibility comments; #125 typed _CallableModule.__call__ kwargs; #126 corrected recover_lock doc; #127 documented the lazy-import workaround; #128 renamed _mutation_generation → _registration_generation; #130 annotated the defence-in-depth conflict check.

[0.6.1] — 2026-06-07¶

A bug-fix and test-hardening release. No public API was removed and no new public names were added. One fix changes key output for inputs containing invisible characters — see Upgrade notes.

Upgrade notes (output-affecting fix)¶

• search_key / catalog_key / sort_key now strip bidi overrides and soft-hyphen / format characters (#93). Previously a value stored with an invisible character (e.g. "pass­word", "user‮txt") produced a different key from its clean equivalent, so dedup and lookup silently missed. The new key is the correct one; if you persist these keys, regenerate any that were computed over text that could contain invisible characters.

Fixed¶

• #93 — key functions (search_key/catalog_key/sort_key) leaked bidi and soft-hyphen characters, so visually-identical inputs produced non-colliding keys. They now strip_bidi after NFKC, matching the other canonicalization presets.
• #82 — Greek reverse transliteration (transliterate(text, target="el")) left literal Latin letters in the output ("psychi" → "ψyχη"). The forward direction romanizes Υ/υ as Y/y (including the ου/αυ/ευ diphthongs), so the el reverse table now maps Y/y back to Greek; round-trips no longer leak Latin letters.
• #69 — transliterate() resolved conflicting kwargs differently for str vs list input (one path silently dropped target, the other context). Conflicts are now checked once, before the dispatch, so both raise identically: context+target and context+tones raise ValueError.
• #72 — translit.unidecode() now mirrors the Unidecode 1.3 signature unidecode(string, errors="ignore", replace_str="?"), mapping Unidecode's errors modes (ignore/replace/preserve/strict) onto the native error handling, instead of raising TypeError on those kwargs.
• #95 — Greek Extended polytonic capitals for omicron/upsilon/omega/rho were corrupted, emitting unrelated Latin letters (Ὅμηρος → Xmiros, Ὑγίεια → Pgieia). Corrected all 50 affected entries to the proper base romanization, consistent with the monotonic forms (Ὅμηρος → Omiros).
• #99.3 — a typo'd form=/errors= value now raises even for pure-ASCII input. Previously the ASCII fast-path returned before reaching Rust, so the bad enum silently no-opped on ASCII and only raised on the first non-ASCII string. Validation now runs before the fast-path in normalize() and transliterate().

Performance¶

• #70 — the batch entry points (transliterate, slugify, normalize, strip_accents on list[str]) now release the GIL around their pure-Rust compute loop via py.allow_threads. Multi-threaded callers processing large batches now get real parallelism (~1.8× wall-clock with two threads) instead of serialising on the interpreter lock. Output is unchanged. Documented in the new "Concurrency (GIL)" section of docs/performance.md.

Documentation¶

• #94 — strict_iso9 is no longer described as "ISO 9:1995". It emits ASCII digraphs (ж→zh, ч→ch, ш→sh), not the standard's diacritics (ž/č/š) — translit tables are ASCII-only by design. Docstrings, the data-file header, and the docs now describe it as a scholarly ASCII (ISO 9-style) transliteration and warn it is not ISO 9-conformant. No behavior change.
• #98 — docs/user-guide/transliteration.md no longer instructs users to pip install translit-rs[arabic|hebrew|context] (those empty extras were removed in 0.6.0); it now documents the bootstrap_dicts.sh / TRANSLIT_DICT_DIR path, matching the README and the runtime error message.

• #99.1 / #99.2 — fixed two false docstrings: sort_key no longer claims to preserve accents (it folds them via transliteration, coinciding with search_key), and slugify no longer documents a pretranslate kwarg it never had.

• #84 — corrected the README throughput table (Cyrillic ~106M chars/sec, slugify ~712K slugs/sec on commodity 4-vCPU hardware) and added a hardware/methodology footnote; added a matching variance note to docs/performance.md.

• #77 — fixed the Text fluent-builder docstring example (normalize is keyword-only: .normalize(form="NFC")), reconciled the language-profile count (README now agrees with the docs at 83), and documented the context kwarg in the transliterate() docstring.

Internal / tests¶

• #78 — added adversarial coverage for the raw-bytes decode path (detect_encoding / decode_to_utf8): deterministic hostile-byte cases in CI plus a Hypothesis st.binary() fuzz suite proving no-panic and invariant-preservation. Documented in THREAT_MODEL.md that the decode path has no input-size cap (caller's responsibility, per the 0.6.0 cap removal).
• #79 — added a single-vs-batch kwarg parity regression test across the full kwarg matrix and a multi-script corpus (the tones batch drop fixed in 0.6.0 can no longer recur silently).

[0.6.0] — 2026-06-07¶

A hardening and bug-fix release. Two new opt-in helpers (dedup_batch, make_cached_transliterator) make this a minor bump; no public API was removed. Several fixes change output for specific inputs — read Upgrade notes before upgrading if you cache or persist transliterator/normalizer output.

Upgrade notes (output-affecting fixes)¶

Each of these was a bug; the new output is the correct one. If you store or cache results that were keyed on the old (buggy) behaviour, regenerate them:

• register_replacements() now actually applies. It was a silent no-op — the registered table was never consulted. Registered replacements now take effect across transliterate() (scalar, list, and context=True). If you registered replacements and (knowingly or not) relied on them being ignored, output changes.
• transliterate(list, tones=True) now returns toned pinyin (was silently toneless on the list path); transliterate(list, target=…, tones=True) now raises ValueError for the forward-only parameter (was silently ignored).
• normalize_confusables(text, target="cyrillic") no longer maps characters onto invisible combining marks (28 such mappings removed).
• strip_obfuscation now folds intra-Latin ASCII homoglyphs (þ→p, ſ→f, ı→i, …) and is idempotent; sanitize_user_input is idempotent for control/invisible characters between combining marks; demojize no longer inserts a stray space after a tab/newline that precedes an emoji.
• Context-aware transliteration (context=True, ar/fa/he) distribution changed. The empty arabic/hebrew/context pip extras have been removed (they never installed anything). The ~37 MB dictionaries are no longer tracked in git, and are not shipped in the wheel. Context mode now loads dictionaries from $TRANSLIT_DICT_DIR (build them with scripts/bootstrap_dicts.sh), or use the embed-dicts Cargo feature for a self-contained build. A packaged pip-installable distribution is tracked in #56/#60.
• decode_to_utf8 default min_confidence changed 0.0 → 0.5. Low-confidence encoding guesses are now rejected by default instead of silently accepted; pass min_confidence=0.0 to restore the old behaviour. (#66)
• Unknown lang codes now raise instead of silently falling back (#68). A typo'd code (lang="RU", lang="russian") used to behave exactly like lang=None — quietly-wrong output — while errors=/form= rejected bad values. transliterate, slugify, sanitize_filename, catalog_key, search_key, sort_key, and ml_normalize now raise TranslitError listing the valid codes. "auto", the nb/nn/da aliases, and register_lang() codes are accepted. (target= already validated.)

Changed¶

• No library-imposed input-size limit (#80, #65). The 10 MiB input cap on transliterate, normalize, fold_case, and the preset pipelines has been removed — it was paternalistic, inconsistently applied (the ASCII fast path bypassed it; slugify/normalize_confusables/strip_zalgo never had it), and the threat model already disclaims DoS. All operations are linear time and memory; bounding untrusted input is the caller's responsibility, documented in the threat model and docstrings. The single retained size guard is the register_replacements output amplification bound (a tiny input can expand to an enormous string via a caller-registered value — an amplification a caller's own input check cannot foresee). Backward-compatible: only previously-rejected large inputs now succeed.
• External wording: capability, not promise. Security-relevant features are now described as mechanisms (TR39 confusable mapping, bidi/zalgo stripping, hostname analysis) rather than outcome guarantees. Package descriptions, README, and docs no longer claim to "prevent"/"neutralize" attacks or achieve "perfect" recovery; the XMR benchmark figure is always stated with its tested-pairs scope. Engineering rigor is held to a high internal bar (see below); the external surface promises nothing it cannot measure.

Added¶

• dedup_batch(texts, …) — transliterate a list, processing each distinct value once and mapping back (large win for repeated/categorical data; ~146× on a high-locality column). Stateless — no cache to invalidate; unique values are chunked at the 100k batch cap. (#31)
• make_cached_transliterator(maxsize=…, …) — opt-in LRU-cached single-string transliterator with options fixed at construction. Self-invalidating: the next call after any register_lang/register_replacements/remove_replacement/ clear_replacements clears the cache (via an internal table-generation counter), so it never serves stale results. Never enabled by default. (#31)
• THREAT_MODEL.md — defines in-scope mechanisms, explicit out-of-scope items (confusables outside the bundled TR39 table, whole-script and multi-character confusables, Unicode-version skew, semantic attacks, DoS), and a vulnerability-vs- known-limitation policy, grounded in the literature (Holgers 2006, Deng 2020, BitAbuse 2025).
• SECURITY.md rewritten on real footing: supported-version policy stated, triage scope defined, and linked to the threat model.
• Security-invariant property tests + fuzzing. proptest invariants in Rust (src/presets.rs) assert no-panic, idempotence, and "no bidi/format control survives" for strip_obfuscation / security_clean / sanitize_user_input / strip_bidi across the Unicode input space; a deterministic, CI-gating adversarial attack-corpus regression (tests/test_attack_corpus.py: homoglyph / zalgo / invisible / bidi / combined, XMR-style); and a cargo-fuzz harness (fuzz/) for continuous coverage-guided fuzzing of the defense pipelines.
• Confusable coverage for intra-Latin homoglyphs of basic ASCII letters (e.g. þ→p, ſ→f, ı→i, ƒ→f, Ɩ→l, ꜱ→s). The TR39 generator previously skipped all Latin-script sources for the Latin target, dropping ~83 genuine homoglyphs of A–Z/a–z; normalize_confusables/strip_obfuscation now fold them. Single-letter Latin confusable coverage of UTS#39 is now complete.
• Pinned data/confusables.txt (UTS#39 17.0.0) as the reproducible, version- controlled input for scripts/gen_confusables.py (--download refreshes it), and a tests/test_confusable_coverage.py gate against Unicode-version drift.

Fixed¶

• register_replacements() was a silent no-op — the global table was stored but never consulted by transliterate(). It now applies as a longest-match pre-pass (no cascade) across the scalar, list, and context=True forward paths, including ASCII-keyed replacements that previously bypassed Rust via the Python fast path. (#51)
• tones= on the list/batch path was dropped: transliterate(["北京"], tones=True) returned toneless pinyin while the scalar path returned toned, and transliterate([...], target=…, tones=True) silently ignored the forward-only parameter instead of raising. Both now match the scalar path. (#14, #15)
• normalize_confusables(target="cyrillic") emitted invisible combining marks — 28 mappings folded a visible character onto a combining Cyrillic-Extended mark (an obfuscation vector). The generator now excludes combining-mark targets. (#24)
• script_info("CanadianAboriginal")["context_aware"] raised KeyError — the entry omitted a required ScriptMeta field; a completeness guard now prevents recurrence. (#18)
• Context path skipped strict_iso9/gost7034 mutual-exclusion validation — transliterate(text, context=True, strict_iso9=True, gost7034=True) now raises ValueError like the non-context path; the missing-dictionary error hint is now language-specific (he→hebrew). (#18)
• demojize inserted a stray space after a tab/newline preceding an emoji ("a\t😀" → "a\t grinning face"); it now checks for any whitespace. (#12)
• Compatibility digit variants fold to digits, not letters (#89). The confusables table mapped Mathematical Alphanumeric digits 𝟎/𝟏 (and the other four families, plus superscripts) to the look-alike letters O/l, so normalize_confusables("𝟏𝟎") gave "lO" and strip_obfuscation corrupted digit runs. The generator now folds any character whose NFKC form is an ASCII digit to that digit. They remain detected as confusable (is_confusable), but canonicalize to the correct number. (ASCII 0/1 were already unaffected.)
• NFKC-compatible Latin is recovered instead of dropped to [?] (#81). Mathematical Alphanumeric Symbols (𝕳𝖊𝖑𝖑𝖔 𝟙𝟚𝟛 → Hello 123), presentation ligatures (fi/fl → fi/fl), and superscripts (x² → x2) now transliterate: an unmapped non-ASCII char is NFKC-decomposed and re-tried before the error fallback. This matches unidecode/anyascii and closes a filter-evasion ("fancy text") gap. Purely additive — only chars that were previously [?] are affected; emoji (no ASCII decomposition) still map to [?].
• Defense pipelines are now idempotent (bugs found by the property tests):
• strip_obfuscation: emoji whose CLDR name contains typographic punctuation (e.g. 👒 → woman’s hat, U+2019 ’) weren't folded because confusables ran before demojize; a second pass folded ’→'. Confusables now runs after demojize.
• sanitize_user_input: an invisible or control character between combining marks (e.g. soft-hyphen, NUL) split a mark-run, so removing it after zalgo-capping merged runs that a second pass then capped differently. Bidi, zero-width, and control characters are now stripped before zalgo-capping.
• Build-time and doc corrections: build.rs now rejects malformed \u{…} escapes in TSV data; embedded-dictionary parse errors are logged (not silently dropped); and numerous stale docstrings/comments were corrected (script_to_lang returns ISO 639-1 or 639-3; normalize() ASCII fast-path; list single-Rust-call caveats).

Security¶

• seal_registrations() / registrations_sealed() (#64, high). The register_lang/register_replacements APIs mutate process-global tables consulted by every transliterate/slugify/catalog_key/… call, so in a multi-tenant or web process one import or request handler could silently alter everyone's canonicalization. seal_registrations() is a one-way latch: after it is called, register/remove/clear raise TranslitError. The registration APIs are now documented as startup-only/single-writer. Separately, a poisoned lock no longer resets registrations to defaults (a panic in one thread could previously wipe another caller's registered languages) — it now recovers the data as-is.
• is_safe_hostname now decodes IDN/xn-- labels (#63, high). Previously an xn-- ACE label was pure ASCII → single-script → reported safe, so the on-the-wire form of the IDN homograph attack (a Cyrillic xn--80ak6aa92e.com "apple" spoof) sailed through — the exact blind spot for a library marketing idn/anti-spoofing. ACE labels are now UTS#46-decoded (via the idna crate) before script/confusable analysis; a malformed ACE label is treated as unsafe. Non-xn-- labels are untouched (no false positives on, e.g., my_host.local).
• is_safe_hostname fails closed (#67.1). A confusable-check error no longer silently degrades to "not confusable" (unwrap_or(false)) → "safe"; it now marks the hostname unsafe.
• strip_bidi/display_clean now also strip deprecated format controls (U+206A–U+206F) and interlinear annotation marks (U+FFF9–U+FFFB) (#67.2), which were previously only handled as transliteration-table entries.
• NFKC×confusables composition pinned (#67.3). Added a regression test fixing the exact set of NFKC-ASCII results that normalize_confusables re-maps (`→', "→'', |→l) so a data/ordering change — e.g. reintroducing digit→letter — fails loudly; and that presets resolve NFKC/TR39 conflicts (ſ→s) via NFKC.
• Context dictionaries are no longer loaded from a CWD-relative path (#61). load_dict_from_fs previously probed ./data/{name}_dict.bin first, so a process whose working directory an attacker influences (or where they can drop ./data/) could inject a substitute dictionary and silently change ar/fa/he output. Dictionaries now load only from $TRANSLIT_DICT_DIR (explicit opt-in) or the crate's own absolute data/ path in source builds.
• Supply-chain: corpus inputs are verified/pinned (#62). The Tashkeela corpus archive is now checksum-verified before it feeds the builders (fail-closed — an unpinned checksum aborts unless ALLOW_UNVERIFIED_CORPUS=1), and the Project Ben Yehuda corpus is fetched at a pinned commit instead of an unpinned live HEAD.
• ContextDict::from_bytes is fully bounds-checked. A malformed or truncated context dictionary previously caused an out-of-bounds panic (the crate is unsafe_code = forbid, so a panic aborts the process). Every read is now bounds-checked and section offsets are validated; capacity hints are clamped. Added truncation/bogus-offset/u32::MAX-count unit tests. (#18)
• register_replacements expansion is bounded. Replacement values are caller-controlled and unbounded; a small input with a large value could expand past the transliterate input cap. Output is now bounded during construction and rejected once it would exceed MAX_TRANSLITERATE_INPUT_BYTES. (#51)

Internal / tests¶

• 170 deterministic tests were excluded from CI. A module-level pytestmark = pytest.mark.hypothesis in test_filename_regressions.py and test_case_folding.py (filename-security and case-folding regressions) deselected the entire files under CI's -m "not hypothesis" filter; only ~10 were actual property tests. The mark is now scoped to the property-test class in each file, so the deterministic tests run in CI. (#12)
• New tests: register_replacements (unit + Hypothesis property), context-dict parser robustness, resolve_auto_lang for all 18 scripts added in v0.3.0+, and a SCRIPT_META field-completeness guard.
• CI/workflow hygiene: concurrency group on secret-scan, uv.lock in the benchmark path filter, and CodeQL no longer triggered by Rust-only changes.

[0.5.0] — 2026-06-06¶

Added¶

• Context-aware transliteration for abjad scripts (Arabic, Persian, Hebrew). transliterate(text, context=True) uses dictionary-based vowel restoration with bigram context disambiguation to produce readable romanized text instead of consonant skeletons.
• Arabic: Tashkeela corpus (65.7M words), 182K unigrams + 200K bigrams. Covers 99%+ of newspaper vocabulary.
• Hebrew: Project Ben Yehuda corpus (11.4M words), 227K unigrams + 200K bigrams. Covers literary Hebrew.
• Persian: 266 curated common words + optional Wiktionary expansion (14.9K entries available via harvester script).
• list_context_langs(): returns language codes that support context=True (currently ["ar", "fa", "he"]).
• LangMeta.context field: "full", "partial", or "none" — enables web/WASM clients to show/hide a context toggle per language.
• ScriptMeta.context_aware field: bool — enables toggle per detected script.
• Dictionary build tooling:
• scripts/build_arabic_dict.py — corpus-based Arabic dictionary builder
• scripts/build_hebrew_dict.py — corpus-based Hebrew dictionary builder
• scripts/build_persian_dict.py — curated vocabulary Persian builder
• scripts/harvest_wiktionary_persian.py — Wiktionary Persian harvester
• scripts/bootstrap_dicts.sh — reproducible bootstrap from zero with pinned checksums. All parameters auditable, no manual steps.
• Abjad transliteration documentation (docs/user-guide/abjad-transliteration.md) covering all three languages, standards used, comparison with other systems.
• pip extras: pip install translit-rs[arabic], [hebrew], [context] for optional context dictionary installation.
• Rust context engine (src/context.rs): binary dictionary reader, Arabic/Hebrew tokenizer, three-tier resolve (bigram → unigram → context-free fallback), lazy-loaded global singletons via OnceLock.
• 28 context-aware tests (8 Arabic, 14 Persian, 6 Hebrew).

Changed¶

• Repositioning (docs + metadata only — no API or coverage changes). The project now leads with its differentiated, proven core: Unicode adversarial-text defense and canonicalization (TR39 visual confusable mapping), with standards-based Latin/Cyrillic/Greek transliteration as the supporting pillar and CJK/Indic/other scripts framed as best-effort, unidecode-compatible coverage.
• Rewrote the package description, keywords, and classifiers (added Topic :: Security) across pyproject.toml, Cargo.toml, and mkdocs.yml to surface the security use case for discovery.
• Restructured README.md / docs/index.md to lead with defense; introduced an explicit three-tier coverage model (core / compatibility / best-effort).
• Added an Adversarial-Text Defense guide (docs/security/adversarial-defense.md) documenting the phonetic-vs-visual distinction, the XMR metric, and benchmark evidence; elevated security to a top-level docs navigation section.
• Reframed the Unidecode migration guide: the unidecode alias is for romanization compatibility, not security (it cannot reverse homoglyph attacks).

Fixed¶

• Linux x86_64 wheels are now built as cp39-abi3 instead of a version-specific cp38-cp38 wheel. Previously the only published x86_64 Linux wheel targeted CPython 3.8, so pip fell back to a source build (requiring a Rust toolchain) on Linux x86_64 for Python 3.9+. The publish workflow now pins the build interpreter and guards against the regression. (#26)
• Documentation: corrected the built-in language-profile count (inconsistently reported as 64 in one place; now consistently 83), and fixed several homoglyph code examples whose expected output was wrong (e.g. leading-character ordering in strip_obfuscation examples). All README/doc examples are now verified against the built library.

Security¶

• Pinned all third-party GitHub Actions to commit SHAs across the CI and release workflows (resolves the CodeQL actions/unpinned-tag findings) and added .github/dependabot.yml to keep them current. This hardens the release pipeline, which uses PyPI trusted publishing (id-token: write).
• Bumped dev/docs dependencies flagged by Dependabot: Pygments → 2.20.0 and pytest → 9.0.3 (the pytest bump applies on Python ≥ 3.10; Python 3.9 stays on pytest 8.4.2, since pytest 9 requires ≥ 3.10). Both are development-only — the package has no runtime dependencies.

Notes¶

• No public API, language registry, or script coverage was removed. All existing imports, language codes, and the pinned API surface are unchanged.

[0.4.0] — 2026-03-29¶

Added¶

• strip_obfuscation() preset pipeline: maximum-strength text deobfuscation using TR39 confusable mapping (visual similarity). Neutralizes homoglyph spoofing, zalgo abuse, invisible character injection, and bidi attacks. Does NOT transliterate — chain with transliterate() explicitly if romanization is also needed. Pipeline: NFKC → strip_zalgo(max_marks=0) → confusables → strip_bidi → strip_zero_width → demojize → strip_accents → fold_case → collapse_whitespace.
• lang_info() and script_info() APIs: return structured metadata (display name, script, region) for any language code or script. Backed by LANG_META (83 entries) and SCRIPT_META (55 entries) with import-time drift assertions.
• 18 new language codes: ban (Balinese), bax (Bamum), bug (Buginese), chr (Cherokee), cjm (Cham), cop (Coptic), khb (Tai Lue), lis (Lisu), mni (Meitei), nod (Northern Thai), nqo (N'Ko), sat (Santali), su (Sundanese), syr (Syriac), tdd (Tai Le), tl (Tagalog), tzm (Tamazight), vai (Vai). Total: 83 languages.
• 10 new Script enum members: Bamum, Buginese, Cham, Lisu, MeeteiMayek, OlChiki, Sundanese, Tagalog, TaiTham, Tifinagh. Total: 57 scripts.
• Transliteration provenance documentation (docs/provenance.md): per-block audit of which formal romanization standard each Unicode block follows.
• API surface stability tests (tests/test_api_stability.py): 133 tests locking down function signatures, class methods, enum members, TypedDicts, protocol interfaces, and __all__ exports.
• Mutation testing survivor killers (tests/test_mutant_killers.py): 92 tests targeting forward-only parameter validation, default parameter sensitivity, pipeline step tuples, and boundary checks.
• Language consistency audit (scripts/audit_language_consistency.py): checks 11 registration points for Rust/Python/docs/test alignment. Wired into pre-push gate.
• 283 empty-string mappings for combining marks and zero-width characters in translit_default.tsv — these are now silently stripped instead of producing [?].
• docs/index.md is now generated from README.md via scripts/generate_docs_index.sh — single source of truth, no more drift.

Fixed¶

• strip_obfuscation() homoglyph resolution: used phonetic transliteration (Cyrillic р→r, с→s) instead of TR39 visual confusable mapping (р→p, с→c). Removed transliterate from the pipeline; confusables now handles homoglyphs.
• Combining marks produce [?]: transliterate("n\u0303") returned "n[?]" instead of "n". Added empty-string TSV mappings for all Combining Diacritical Marks (U+0300–U+036F), Extended (U+1AB0–U+1AFF), Supplement (U+1DC0–U+1DFF), Symbols (U+20D0–U+20F0), and Half Marks (U+FE20–U+FE2F).
• Zero-width characters produce [?]: transliterate("a\u200Bb") returned "a[?]b". Added empty-string mappings for ZWS, ZWNJ, ZWJ, word joiner, BOM, soft hyphen, bidi marks, and line/paragraph separators.
• TextPipeline confusable ordering: confusables ran before transliterate, creating mixed-script gibberish on Cyrillic/Greek input. Swapped execution order so transliterate runs first (matching catalog_key preset).
• demojize() adjacent emoji concatenation: demojize("🔥🔥") returned "firefire" instead of "fire fire". Added space padding between adjacent emoji-to-text replacements.
• SCRIPT_RANGES sort order: MeeteiMayek Extensions was misplaced, breaking binary search for Ethiopic Extended-A. Added test_script_ranges_sorted invariant.
• Tibetan incorrectly documented as Wylie: actual mappings use Indic-phonetic romanization (ཅ→cha, not Wylie's ca).

Changed¶

• BREAKING: transliterate_batch(), slugify_batch(), normalize_batch(), and strip_accents_batch() removed. The base functions now accept both str and list[str] via @typing.overload. Pass a list to get batch processing: transliterate(["café", "naïve"]) → ["cafe", "naive"].
• BREAKING: strip_obfuscation() no longer transliterates. Uses TR39 confusables (visual mapping) instead. lang= parameter removed. Chain with transliterate() explicitly if romanization is also needed.
• CI restructured: lint/test on PRs only (not push-to-main), hypothesis tests excluded (~4s vs ~46s), CodeQL moved to workflow file with path filtering, benchmarks split to own workflow.
• Pinned ruff==0.15.4 in CI and pyproject.toml to prevent format drift.
• Python 3.9 remains a supported runtime (requires-python = ">=3.9", abi3-py39) but was removed from the release CI matrix; CI runs on Python 3.10+ because tests use PEP 604 (X | Y) syntax without from __future__ import annotations.

[0.3.0] — 2026-03-28¶

Added¶

• Unicode coverage expansion: 2,553 new codepoints across 33 Unicode blocks, bringing total translit_default.tsv entries from 6,633 to 9,186.

Tier 1 — Forms and extensions (~1,741 codepoints): - Fullwidth ASCII (FF01–FF5E): 94 characters, mechanical offset mapping - Halfwidth Hangul (FFA0–FFDC): 66 characters via compatibility jamo - Enclosed/Circled Alphanumerics (2460–24FF): 160 characters (①→1, Ⓐ→A) - Superscript/Subscript (2070–209F): 29 characters mapped to base forms - Roman Numerals (2160–2188): 41 characters (Ⅰ→I, Ⅱ→II, ... Ⅻ→XII) - Modifier Letters (02B0–02FF): 80 characters (ʰ→h, ʷ→w) - IPA/Phonetic Extensions (0250–02AF): 96 characters (ɑ→a, ʃ→sh, ŋ→ng) - Greek Extended (1F00–1FFF): 233 characters (polytonic → base Greek → Latin) - Hangul Jamo (1100–11FF): 256 individual jamo components - Kangxi Radicals (2F00–2FD5): 214 radical forms → pinyin via CJK decomposition - CJK Compatibility Ideographs (F900–FAFF): 472 characters → pinyin via canonical decomposition targets

Tier 2 — Living scripts (~812 codepoints): - Gap-filling for 7 partially-covered scripts: Balinese, Canadian Syllabics, Cherokee, Coptic, N'Ko, Syriac, Vai - 10 new abugida scripts with virama/inherent-vowel handling: Sundanese, Tai Tham, Cham, Batak, Buginese, Tagalog, Hanunoo, Buhid, Tagbanwa, Meetei Mayek - 4 new alphabetic/syllabic scripts: Tifinagh, Lisu, Ol Chiki, Bamum

• Unicode range constants for 12 new scripts in src/unicode_ranges.rs: SUNDANESE, TAI_THAM, CHAM, BATAK, BUGINESE, TAGALOG, HANUNOO, BUHID, TAGBANWA, MEETEI_MAYEK, MEETEI_MAYEK_EXT.
• 10 new *_char_role() functions in src/transliterate.rs for abugida virama handling (Sundanese, Tai Tham, Cham, Batak, Buginese, Tagalog, Hanunoo, Buhid, Tagbanwa, Meetei Mayek).
• scripts/generate_unicode_expansion.py: reproducible generator script for all Tier 1 and Tier 2 TSV entries (1,310 lines).
• cargo-clippy pre-commit hook mirroring CI -D warnings to catch lints before push.
• Callable module: import translit; translit("Москва", lang="auto") now works as a shorthand for translit.transliterate(...). Uses in-place __class__ mutation to preserve unittest.mock.patch compatibility.

Fixed¶

• Finnish transliteration: removed incorrect alias fi→sv. Finnish ä/ö are independent phonemes (→a/o via default table), not ae/oe variants as in Swedish/German. Hämäläinen now correctly produces Hamalainen.
• Icelandic transliteration: removed incorrect ð→dh and Ð→Dh overrides. Default table already maps ð→d (ICAO/passport standard). Retained Æ→Ae override (differs from default AE). Icelandic override count reduced from 6 to 2.
• clippy manual_range_patterns lint in buginese_char_role: collapsed 0x1A17 | 0x1A18 | 0x1A19..=0x1A1B to 0x1A17..=0x1A1B.
• errors="preserve" dropping visible characters: characters with explicit empty-string TSV mappings (e.g. U+060E Arabic Poetic Verse Sign, U+30FC Katakana Prolonged Sound Mark) are now preserved instead of silently dropped when errors="preserve" is set.

Changed¶

• is_indic() and indic_char_role() expanded to cover all 11 new Brahmic/abugida script ranges.
• lookup_lang(): Finnish no longer dispatches to Swedish override table; falls through to default.
• Icelandic language TSV (translit_lang_is.tsv) reduced from 6 to 2 entries.
• ml_normalize preset: switched transliteration from Preserve to Ignore error mode — ML pipelines need clean ASCII output, not preserved non-ASCII.

[0.2.0] — 2026-03-27¶

Added¶

• Exhaustive testing framework — three layers of machine-verifiable assurance:
• Compile-time assertions (build.rs): all transliteration table values asserted ASCII-only, entry count sanity checks (Hanzi ≥20k, BMP ≥5k, confusables ≥1k). Build fails if any assertion is violated.
• Exhaustive domain tests (Rust): 16 tests covering all 11,172 Hangul syllables, full BMP (63,488 codepoints) for ASCII output and idempotence, all 20,992 CJK ideographs, all 51 compatibility jamo, and structural verification of 15 Indic script blocks. Zero sampling gaps.
• Stated invariant specifications (Python): 7 stated invariants (I1–I7) verified via exhaustive enumeration and Hypothesis — ASCII passthrough, ASCII output, idempotence, no exceptions, determinism, input size bound, output length bound.
• Two-tier test architecture: formal tests gated behind #[ignore] (Rust) and @pytest.mark.formal (Python) so they don't slow everyday development. Run before release with cargo test -- --ignored and pytest -m formal.
• CLAUDE.md: project-level development guide for automated agents — documents build commands, test tiers, and code conventions.
• list_scripts() function for programmatic script discovery.
• docs/formal-verification.md: specification document for exhaustive testing methodology.
• Comprehensive overhaul of docs/architecture/testing-guarantees.md with exhaustive testing differentiator analysis and alternative library comparison.

Changed¶

• IndicRole enum and indic_char_role() / script-specific char_role functions changed from private to pub for integration test access (parent modules remain #[doc(hidden)]).
• tables::hangul module changed from mod to pub mod for integration test access.
• Hangul const assertions added: JUNGSEONG_COUNT, JONGSEONG_COUNT, total syllable count, and compatibility jamo range verified at compile time.
• Total test count: 2,900+ (up from 1,678 in 0.1.5).

[0.1.5] — 2026-03-27¶

Added¶

• Reverse transliteration: transliterate(text, target="ru") converts Latin → native script for Russian, Ukrainian, and Greek. PHF tables generated at build time from inverted language TSV data.
• Toned pinyin: transliterate("北京", tones=True) returns "běi jīng" with tone marks. Toned readings sourced from Unihan kMandarin field for all 20,924 CJK Unified Ideographs.
• ISO 9:1995 scholarly Cyrillic: transliterate(text, strict_iso9=True) for scholarly romanization. GOST R 7.0.34 variant via gost7034=True.
• Japanese Kunrei-shiki (lang="ja-kunrei"): alternative romanization profile, bringing total language count to 65.
• Ancient scripts: Coptic, Gothic, Old Italic, Runic, Ogham transliteration tables.
• CLI short aliases: t (transliterate), s (slugify), n (normalize), p (pipeline), d (demojize) — e.g. translit t "café".
• CLI --target flag: translit t --target ru "Moskva" for reverse transliteration.
• CLI --tones, --strict-iso9, --gost7034 flags for transliterate subcommand.
• CLI --lang flag for slugify subcommand.
• console_scripts entry point: translit command available after pip install translit-rs.
• docs/cli.md: comprehensive CLI documentation with piping, exit codes, examples.
• Links section in README.md and docs/index.md for RTD ↔ GitHub cross-references.

Changed¶

• transliterate() API unified: reverse_transliterate() merged into transliterate() via target parameter. Old function removed.
• transliterate_impl Rust signature now takes 7 arguments (added tones: bool).
• Updated benchmark numbers after tones parameter addition (15–46% regression in transliteration hot path due to additional branch; throughput now 450M chars/sec Latin, 130M chars/sec Cyrillic).
• Performance documentation updated across 4 files to reflect current benchmark results.

Fixed¶

• clippy format_push_string lint in build.rs — replaced push_str(&format!()) with write!().
• clippy unreadable_literal in PHF-generated reverse_translit_phf.rs — suppressed via inner attribute in src/reverse.rs.
• All 219 integration test call sites updated for 7-argument transliterate_impl.

[0.1.4] — 2026-03-25¶

Added¶

• lang="auto" script-based language detection: When lang="auto" is passed to transliterate(), slugify(), TextPipeline, Slugifier, or any other call site, the library detects the dominant non-Latin script in the input and maps it to a default language code automatically. Maps 28 scripts to language codes (e.g. Cyrillic→ru, Han→zh, Hiragana/Katakana→ja, Thai→th). Zero overhead for lang=None or explicit lang codes.
• LANG_AUTO constant ("auto") in translit._enums.
• Georgian transliteration (lang="ka"): 114 TSV entries covering Mkhedruli, Mtavruli, and supplement ranges. BGN/PCGN national romanization.
• Armenian transliteration (lang="hy"): 86 TSV entries covering uppercase, lowercase, and 5 ligatures (U+FB13–FB17). BGN/PCGN romanization.
• Sinhala transliteration (lang="si"): 90 TSV entries. Extended Indic Brahmic engine range from 0x0900..=0x0D7F to 0x0900..=0x0DFF with dedicated sinhala_char_role() function for Sinhala-specific offsets.
• Thai transliteration (lang="th"): 87 TSV entries using RTGS romanization. New ScriptClass::Tai with tone-mark stripping and cancellation handling.
• Lao transliteration (lang="lo"): 67 TSV entries using BGN/PCGN romanization. Shares Tai engine with Thai via offset masking.
• Ethiopic transliteration (lang="am"): 307 TSV entries for Ge'ez alphasyllabary (34 consonant bases × 7 vowel orders + labialized forms + digits). Pure data addition — no engine changes needed.
• Myanmar transliteration (lang="my"): 89 TSV entries. New myanmar_char_role() for Brahmic engine with virama (U+1039) and asat (U+103A) support. Medials (U+103B–103E) classified as dependent vowels.
• Khmer transliteration (lang="km"): 110 TSV entries. New khmer_char_role() for Brahmic engine with coeng (U+17D2) as virama. All consonants normalized to inherent 'a' regardless of series.
• Tibetan transliteration (lang="bo"): 147 TSV entries. New tibetan_char_role() for Brahmic engine with halanta (U+0F84) and subjoined consonants (U+0F90–0FBC).
• Unicode range constants: TIBETAN (0x0F00–0x0FFF), MYANMAR (0x1000–0x109F), KHMER (0x1780–0x17FF) in src/unicode_ranges.rs.
• Comprehensive test coverage: example-based tests for all 9 new scripts, property-based tests (hypothesis + proptest), multi-script mixture tests.
• Built-in language count: 51 → 60.

Changed¶

• is_indic() extended to include Tibetan, Myanmar, and Khmer ranges for Brahmic abugida processing.
• indic_char_role() dispatches to script-specific functions for Sinhala, Tibetan, Myanmar, and Khmer codepoint ranges.

[0.1.3] — 2026-03-25¶

Added¶

• strip_control and strip_zero_width now work as independent pipeline steps without requiring collapse_whitespace=True. Previously they were silently ignored when collapse_whitespace was disabled.
• strip_control_chars() and strip_zero_width_chars() standalone Rust functions for filtering without whitespace collapsing.
• decimal and hexadecimal flags in SlugConfig are now functional. Setting decimal=False preserves &#NNN; entities; hexadecimal=False preserves &#xHHH; entities. Previously these flags were accepted but silently ignored.
• Rust integration tests: tests/integration_emoji.rs (10 tests), tests/integration_slugify.rs (20 tests), tests/integration_transliterate.rs (21 tests), tests/integration_whitespace.rs (12 tests).

Changed¶

• TextPipeline parameters strip_control and strip_zero_width changed from bool (default True) to bool | None (default None). When None, they inherit from collapse_whitespace — True if collapse_whitespace=True, False otherwise. Set explicitly to True for standalone use without collapse_whitespace. This is backward compatible: existing code that passes collapse_whitespace=True gets the same behavior as before.
• steps() now reports strip_control and strip_zero_width as separate entries when active, giving full visibility into pipeline behavior.
• Pipeline step order updated: normalize → confusables → demojize → strip_accents → transliterate → fold_case → strip_control → strip_zero_width → collapse_whitespace.
• Migrated from once_cell to std::sync::LazyLock / OnceLock; MSRV bumped to 1.80. Removed once_cell dependency.
• needs_cjk_space() match arm tightened from wildcard _ to explicit Ideograph | Hangul | Kana to match the call-site is_cjk guard.

Fixed¶

• decode_entities() corrupting multi-byte UTF-8 characters (BUG-1). The function used bytes[i] as char which treated each continuation byte as a separate Latin-1 codepoint (e.g. café → café). Now advances by full UTF-8 characters.
• decode_numeric_entity_skip() panicking on malformed &# followed by multi-byte UTF-8 (BUG-2). The skip function walked through continuation bytes looking for ;, landing inside a multi-byte character. Now stops at the first non-ASCII byte.

Performance¶

• ASCII fast-path in demojize_impl and demojize_rust: pure-ASCII text returns immediately without Vec<char> allocation or emoji scanning.
• filter_stopwords replaced intermediate Vec<_> + .join() with a pre-allocated String fold, removing one allocation per slugify call.

[0.1.2] — 2026-03-25¶

Added¶

• Python 3.14 support (classifier and CI test matrix).
• ruff check --fix pre-commit hook for automatic lint fixing.
• CI publish workflow using pypa/gh-action-pypi-publish with OIDC trusted publishers.
• Multi-platform wheel builds: Linux (x86_64, aarch64), macOS (Intel, ARM64), Windows.
• steps() method on _TextPipeline type stub.

Changed¶

• Resolved all clippy pedantic warnings instead of suppressing them — reduced lint suppressions from 48 to 22 (remaining are genuine PyO3 constraints). Fixes include: combined identical match arms, replaced manual counters with .enumerate(), moved item declarations before statements, used clone_into(), merged identical branches, fixed doc comment formatting.
• Widened stopwords and replacements type stubs from strict tuple/list to Sequence for better mypy compatibility.
• Applied ruff format to all Python source and test files.
• Switched docs publish from deprecated maturin upload to pypa/gh-action-pypi-publish.
• macOS Intel wheels now cross-compiled on ARM64 runner (macos-14) instead of deprecated macos-13.
• CI doctests now run against installed package (not source tree) with explicit shell: bash for Windows compatibility.

Fixed¶

• TextPipeline.explain() doctest: output format is normalize (NFC) not normalize (form=NFC).
• from __future__ import annotations placement in test files (must follow module docstring, not precede it).
• Malformed HTML entity test expectation: decode_entities("&#xyz;") correctly returns "", not "yz;".
• Rust benchmark CI: target bench_core binary explicitly to avoid passing Criterion flags to the test harness.
• Ruff lint fixes: unsorted imports in test_encoding.py, unused import is_mixed_script in test_security_invariants.py.
• Read the Docs trigger workflow: simplified curl status handling, graceful warning when RTD_TOKEN is missing.
• Removed incorrect PyPy classifier (abi3 is CPython-only).

[0.1.1] — 2026-03-25¶

Added¶

• src/unicode_ranges.rs — named constants for all Unicode codepoint ranges used by the library, eliminating magic numbers scattered across modules.
• tests/test_concurrency.py — concurrent access tests for LANG_TABLES and HANGUL_CACHE, plus malformed Unicode input tests.
• Code coverage reporting in CI (pytest-cov, XML report uploaded as artifact).
• CLOCK$, KEYBD$, SCREEN$, COM0, LPT0 added to Windows reserved filename list.
• casefold() alias for fold_case() — matches str.casefold() naming.
• remove_accents() alias for strip_accents() — matches sklearn/ML ecosystem naming.
• Compatibility parameter aliases: replacement_text/max_len on sanitize_filename() (pathvalidate), greedy/preferred_aliases on is_confusable() (confusable_homoglyphs), delimiters on demojize() (emoji library).
• Complete API documentation for 19 previously undocumented exported functions: precompiled pipelines, grapheme clusters, encoding detection, Text builder, is_safe_hostname, demojize, strip_bidi, EmojiProvider protocol.
• Three new API reference pages: Precompiled Pipelines, Grapheme Clusters, Encoding.
• "Guides by role" section in docs/index.md and README.md.
• Performance section in README.md with benchmark numbers.
• Script enum documentation expanded from 28 to all 41 members.

Changed¶

• transliterate_impl refactored: capacity estimation extracted to estimate_capacity(), character classification to classify_char(), and CJK spacing logic to needs_cjk_space().
• All RwLock accesses now recover from lock poisoning using .unwrap_or_else(|e| e.into_inner()) instead of silently falling through.
• Lambda closures in _compat.py replaced with named inner functions for clarity.
• emoji.rs write!() call no longer uses .unwrap() (infallible, documented with a // SAFETY comment).
• MkDocs theme switched from material to readthedocs.
• All documentation references updated from "unirust" to "translit".
• Development status promoted from Alpha to Beta.
• Package renamed from translit to translit-rs on PyPI (interim until PEP 541 grants the translit name). Python import remains import translit.

Fixed¶

• Type stub _text.pyi imported from wrong module name (unirust → translit).
• Type stub _translit.pyi missing min_confidence parameter on _decode_to_utf8.
• Type stub _text.pyi missing grapheme_split, grapheme_truncate, catalog_key methods.
• security_clean() pipeline step order corrected in 5+ locations: strip_bidi runs before collapse_whitespace (matching Rust implementation).
• catalog_key() step order corrected: transliterate before strip_accents.
• Stale PyO3 boundary overhead corrected from ~4µs to ~240ns in docs and code comments.

Deprecated¶

• translit._compat awesome-slugify compatibility layer (Slugify, UniqueSlugify, slugify_* instances) — planned removal in v1.0.

[0.1.0] — 2026-01-01¶

Added¶

• Initial release.
• Unicode transliteration for 60 language profiles.
• Slugification, normalization, confusable detection, filename sanitization.
• Emoji demojization with ZWJ sequence support.
• Backward-compatible layers for Unidecode and awesome-slugify.