euredact 0.3.5 → 0.3.7

package/dist/esm/rules/suppressors.js

@@ -9,15 +9,29 @@ function getContext(text, start, end) {
 }
 // Unicode-safe word boundary: negative lookahead for any letter (including accented)
 const _UWB = "(?![a-zA-Z\\u00C0-\\u024F\\u0400-\\u04FF])";
-const CURRENCY_AFTER = new RegExp(`^\\s*(?:EUR|€|\\$|USD|GBP|£|CHF|ISK|SEK|NOK|DKK|euro|euros|dollar|dollars|pond|kronor|kroner|kr)${_UWB}`, "i");
-const CURRENCY_COMMA_AFTER = new RegExp(`^[.,]\\d{1,2}\\s*(?:EUR|€|\\$|USD|GBP|£|CHF|ISK|SEK|NOK|DKK|euro|euros|kr(?:onor|oner)?|pond)${_UWB}`, "i");
-const CURRENCY_BEFORE = /(?:EUR|€|\$|USD|GBP|£|CHF|ISK|SEK|NOK|DKK)\s*$/;
-const AMOUNT_LABEL_BEFORE = /(?:Montant|Beløb|Summa|Summe|Bedrag|Amount|Total|TTC|inkl|Upphæð)\s*:?\s*$/i;
+// Symbols are held apart from the alphabetic codes because a word boundary
+// after "€" is meaningless — the symbol is not a word character. Python used a
+// bare "\b" here and so could never match "1163 €," at all; this port used the
+// lookahead below and did not have that defect. Kept split so the difference
+// cannot come back.
+const CURRENCY_SYMBOL = "€|\\$|£|¥|₺|zł|Kč|лв|kn|Ft|₽";
+const CURRENCY_WORD = "EUR|USD|GBP|CHF|ISK|SEK|NOK|DKK|PLN|CZK|HUF|RON|BGN|HRK|" +
+    "euro|euros|dollar|dollars|pond|kronor|kroner|kronur|kr|" +
+    "złoty|korun|forint|lei|leva";
+const CURRENCY_AFTER = new RegExp(`^\\s*(?:(?:${CURRENCY_SYMBOL})|(?:${CURRENCY_WORD})${_UWB})`, "i");
+const CURRENCY_COMMA_AFTER = new RegExp(`^[.,]\\d{1,2}\\s*(?:(?:${CURRENCY_SYMBOL})|(?:${CURRENCY_WORD})${_UWB})`, "i");
+const CURRENCY_BEFORE = new RegExp(`(?:(?:${CURRENCY_SYMBOL})|\\b(?:${CURRENCY_WORD}))\\s*$`, "i");
+const AMOUNT_LABEL_BEFORE = /\b(?:Montant|Beløb|Summa|Summe|Bedrag|Amount|Total|TTC|inkl|Upphæð|Importe|Importo|Valore|Wartość|Kwota|Částka|Összeg|Sum|Beloop|Prix|Preis|Price|Loyer|Miete|Huur|Rent|Betrag)\s*(?:\w+\s*)?:?\s*$/i;
 // Use Unicode-aware word boundary via \p{L} negative lookahead instead of \b
 // because JS \b is ASCII-only and fails on Unicode letters (e.g. "München" → \bm\b matches M before ü)
 const UNIT_AFTER = /^\s*(?:kg|km|cm|mm|m[²³]|ml|mg|GB|MB|KB|TB|%|jaar|maanden|weken|dagen|uur|minuten|seconden|stuks|pcs|pieces|ans|mois|semaines|jours|heures|Jahre|Monate|Wochen|Tage|Stunden)(?![a-zA-Z\u00C0-\u024F\u0400-\u04FF])/i;
 // Same missing-boundary flaw: "ref" matched the tail of "kortref".
-const REFERENCE_BEFORE = /\b(?:dossier|ref\.?|referentie|reference|référence|factuurnummer|invoice\s*(?:nr|number|no)?|bestelnummer|order\s*(?:nr|number|no)?|kenmerk|ordernummer|Aktenzeichen|numéro\s*de\s*(?:dossier|facture|commande)|bestellnummer|Rechnungsnummer|artikelnr|article\s*no|contract\s*(?:nr|number|no)?|pagina|page|Seite|blz\.?|Facture\s*n[°o]?|Faktura\s*n[°or]\.?|Lasku\s*n[°or]o?\.?|Rechnung\s*(?:Nr|n[°o])?|faktura\s*(?:nr|n[°o])?|bestilling\s*(?:nr|n[°o])?|bestelling\s*n[°or]\.?|Reikningur\s*nr|Factuur\s*n[ro]?\.?|Nota\s*n[ro]?\.?)\s*[:.]?\s*$/i;
+const REFERENCE_BEFORE = /\b(?:dossier|ref\.?|referentie|reference|référence|factuurnummer|invoice\s*(?:nr|number|no)?|bestelnummer|order\s*(?:nr|number|no)?|kenmerk|ordernummer|Aktenzeichen|numéro\s*de\s*(?:dossier|facture|commande)|bestellnummer|Rechnungsnummer|artikelnr|article\s*no|contract\s*(?:nr|number|no)?|pagina|page|Seite|blz\.?|Facture\s*n[°o]?|Faktura\s*n[°or]\.?|Lasku\s*n[°or]o?\.?|Rechnung\s*(?:Nr|n[°o])?|faktura\s*(?:nr|n[°o])?|bestilling\s*(?:nr|n[°o])?|bestelling\s*n[°or]\.?|Reikningur\s*nr|ticket|incident\s*(?:report|nr|no)?|case\s*(?:nr|no|id)?|zaaknummer|meldingsnummer|Vorgangsnummer|Vorgang|Störungsmeldung|saksnummer|ärendenummer|sagsnummer|asianumero|Factuur\s*n[ro]?\.?|Nota\s*n[ro]?\.?)\s*[:.#]?\s*$/i;
+// "#" before a number marks a reference in every language in scope — a ticket,
+// an order, a line item — and never a postcode or a national identifier.
+const HASH_BEFORE = /#\s*$/;
+// A short uppercase tag hyphenated to the number: "IR-43433", "INC-2024".
+const REF_PREFIX_BEFORE = /(?:^|[\s([])[A-Z]{2,5}-$/;
 const LEGAL_BEFORE = /(?:Art(?:ikel|icle|\.)|§|Artikel|Section|Sectie|Afdeling|paragraaf|Absatz|alinéa|punt|point|Punkt|lid)\s*$/i;
 const MATH_BEFORE = /[=+\-×÷*/]\s*$/;
 const MATH_AFTER = /^\s*[=+\-×÷*/]/;
@@ -49,7 +63,11 @@ const MONTH_NAMES = "january|february|march|april|may|june|july|august|september
     "|gennaio|febbraio|marzo|aprile|maggio|giugno|luglio|agosto|settembre|ottobre|dicembre" +
     "|enero|febrero|abril|mayo|junio|julio|septiembre|octubre|noviembre|diciembre" +
     "|janeiro|fevereiro|março|marco|maio|junho|julho|setembro|outubro|novembro|dezembro" +
-    "|marts|augusti" +
+    // "desember" (nb/nn/is) was the one December spelling missing from a list
+    // that already carried ten others — "1. desember 2025" was a Norwegian
+    // postcode because of it.
+    "|marts|augusti|desember" +
+    "|janúar|febrúar|apríl|maí|júní|júlí|ágúst|nóvember" +
     "|tammikuu|helmikuu|maaliskuu|huhtikuu|toukokuu|kesäkuu|heinäkuu|elokuu|syyskuu|lokakuu|marraskuu|joulukuu" +
     "|jaanuar|veebruar|aprill|juuni|juuli|oktoober|detsember" +
     "|janvāris|februāris|aprīlis|maijs|jūnijs|jūlijs|augusts|septembris|oktobris|novembris|decembris" +
@@ -74,7 +92,11 @@ const DATE_AFTER = /^[./-]\d{1,2}[./-]\d{1,2}\b/;
  * ISO 4217 codes made only of the consonants a Spanish plate accepts, so a money
  * amount reads as a registration. This is why "2297 DKK" was a plate.
  */
-const CURRENCY_CODE = "(?:DKK|SEK|NOK|ISK|CZK|PLN|HUF|RON|BGN|HRK|CHF|GBP|TRY|RSD|MKD|BYN|KZT|CNY|JPY|KRW|ZAR|BRL|MXN|CLP|COP|PLZ|SKK|TRL)";
+// Crypto tickers are the same shape and the same mistake — every surviving
+// plate false positive was "4499 BTC" or a sibling. A ticker is a unit, exactly
+// as an ISO 4217 code is.
+const CURRENCY_CODE = "(?:DKK|SEK|NOK|ISK|CZK|PLN|HUF|RON|BGN|HRK|CHF|GBP|TRY|RSD|MKD|BYN|KZT|CNY|JPY|KRW|ZAR|BRL|MXN|CLP|COP|PLZ|SKK|TRL" +
+    "|BTC|ETH|LTC|XRP|BCH|XLM|XMR|DOT|SOL|ADA|TRX|DOGE|USDT|USDC)";
 const CURRENCY_CODE_TAIL = new RegExp(`\\s${CURRENCY_CODE}$`);
 const CURRENCY_CODE_AFTER = new RegExp(`^\\s*${CURRENCY_CODE}\\b`);
 /** Log-timestamp tails, which the ":"-anchored secret rule picks up. */
@@ -88,6 +110,32 @@ const NOT_A_SECRET = new RegExp("^(?:af|ap|ca|cn|eu|il|me|sa|us|gov)-(?:north|so
 const EXACT_UUID = /^[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}$/;
 const BIC_SHAPE = /^[A-Z]{6}[A-Z0-9]{2}(?:[A-Z0-9]{3})?$/;
 const BIC_CUE_BEFORE = /(?:BIC|SWIFT|BIC\/SWIFT)\s*[:=]?\s*\(?\s*$/i;
+/** An email address, exactly. Same argument as the UUID and the BIC. */
+const EXACT_EMAIL = /^[\w.!#$%&'*+/=?^`{|}~-]+@[\w-]+(?:\.[\w-]+)+$/;
+/**
+ * A URL, or the tail of one after the scheme has been split off. The secret
+ * rules anchor on ":" and "=", so "https://api.sendgrid.com/v3/mail/send" hands
+ * them "//api.sendgrid.com/v3/mail/send" and a slash-separated path scores as
+ * high-entropy. An endpoint is published documentation, not a credential.
+ */
+const URL_LIKE = /^(?:[a-z][a-z0-9+.-]*:)?\/\/[^\s/]+(?:\/\S*)?$|^[a-z0-9-]+(?:\.[a-z0-9-]+)*\.[a-z]{2,}(?::\d{1,5})?(?:\/\S*)?$/i;
+/**
+ * A URL that *carries* a credential is the opposite case, and it is the common
+ * one: "mongodb://admin:DFKDKi1eb51OOhuHPYz@rds-main.eu-west-1.rds.amazonaws.com"
+ * is a connection string with a live password in it. Suppressing those as "just
+ * a URL" lost 347 real secrets, so the endpoint test must not fire when userinfo
+ * or a credential-bearing query parameter is present.
+ */
+const URL_WITH_CREDENTIALS = /^(?:[a-z][a-z0-9+.-]*:)?\/\/[^/@\s]*:[^/@\s]*@|[?&](?:api_?key|access_?token|auth|token|secret|password|pwd|sig|signature|credential)=/i;
+/** An LDAP distinguished name: "cn=github-actions,dc=corp,dc=eu". */
+const LDAP_DN = /^(?:cn|ou|dc|uid|o|l|st|c)=[^,=]+(?:,\s*(?:cn|ou|dc|uid|o|l|st|c)=[^,=]+)+$/i;
+/**
+ * Lower-case words joined by hyphens, optionally with a short id suffix:
+ * "service-account", "data-lake-33e061". The single-word test below already
+ * rejects "sozialversicherungsnummer" on the reasoning that a real secret would
+ * be mixed case; a hyphenated compound is the same token with a separator.
+ */
+const HYPHENATED_WORDS = /^[a-z]{2,}(?:-[a-z]{2,})+(?:-[0-9a-f]{4,8})?$/;
 /** A dotted quad with every octet in range. */
 const DOTTED_QUAD = /^(?:(?:25[0-5]|2[0-4]\d|1\d\d|[1-9]?\d)\.){3}(?:25[0-5]|2[0-4]\d|1\d\d|[1-9]?\d)$/;
 /** Two plausible years joined by a hyphen — a school year, a contract term. */
@@ -117,7 +165,7 @@ function suppressSecretOverStructured(text, match) {
     if (match.patternDef.entityType !== EntityType.SECRET)
         return false;
     const token = coreToken(match.text);
-    if (EXACT_UUID.test(token))
+    if (EXACT_UUID.test(token) || EXACT_EMAIL.test(token))
         return true;
     if (BIC_SHAPE.test(token)) {
         return BIC_CUE_BEFORE.test(text.slice(Math.max(0, match.start - 24), match.start));
@@ -142,6 +190,12 @@ function suppressSecretNotASecret(_text, match) {
         return true;
     if (NOT_A_SECRET.test(bare))
         return true;
+    if (URL_LIKE.test(bare) && !URL_WITH_CREDENTIALS.test(bare))
+        return true;
+    if (LDAP_DN.test(bare))
+        return true;
+    if (HYPHENATED_WORDS.test(bare))
+        return true;
     const alpha = /^[A-Za-zÀ-ÿ]+$/.test(bare);
     const oneWord = bare === bare.toLowerCase() ||
         (bare.slice(0, 1) === bare.slice(0, 1).toUpperCase() &&
@@ -202,15 +256,34 @@ function suppressUnits(text, match) {
     const [, after] = getContext(text, match.start, match.end);
     return UNIT_AFTER.test(after);
 }
+/**
+ * Suppress numbers preceded by reference/invoice/dossier keywords.
+ *
+ * POSTAL_CODE belongs here for the same reason every other numeric type does,
+ * and its absence was an oversight: a five-digit ticket number sits in exactly
+ * the shape a German or French postcode occupies, so "Ticket #94730" and
+ * "Incident report IR-43433" were masked as addresses.
+ *
+ * "#" and "XX-" are checked adjacently rather than through the 150-character
+ * keyword window. That window is what made the postal rule claim years in
+ * dates; widening its reach to fix a different symptom would repeat the error.
+ */
+// Hoisted out of the suppressor bodies: rebuilding these per candidate cost
+// ~260k Set allocations per 1 MB document, and GC was 10.7% of the profile.
+const REFERENCE_TYPES = new Set([EntityType.PHONE, EntityType.NATIONAL_ID, EntityType.SSN, EntityType.TAX_ID, EntityType.IBAN, EntityType.CHAMBER_OF_COMMERCE, EntityType.POSTAL_CODE]);
+const LEGAL_TYPES = new Set([EntityType.PHONE, EntityType.NATIONAL_ID, EntityType.POSTAL_CODE]);
 function suppressReference(text, match) {
-    const applicable = new Set([EntityType.PHONE, EntityType.NATIONAL_ID, EntityType.SSN, EntityType.TAX_ID, EntityType.IBAN, EntityType.CHAMBER_OF_COMMERCE]);
+    const applicable = REFERENCE_TYPES;
     if (!applicable.has(match.patternDef.entityType))
         return false;
     const [before] = getContext(text, match.start, match.end);
-    return REFERENCE_BEFORE.test(before);
+    if (REFERENCE_BEFORE.test(before))
+        return true;
+    const adjacent = text.slice(Math.max(0, match.start - 8), match.start);
+    return HASH_BEFORE.test(adjacent) || REF_PREFIX_BEFORE.test(adjacent);
 }
 function suppressLegal(text, match) {
-    const applicable = new Set([EntityType.PHONE, EntityType.NATIONAL_ID, EntityType.POSTAL_CODE]);
+    const applicable = LEGAL_TYPES;
     if (!applicable.has(match.patternDef.entityType))
         return false;
     const [before] = getContext(text, match.start, match.end);
@@ -433,8 +506,15 @@ function enclosingLine(text, start, end) {
  */
 function structuralUnit(text, start, end) {
     const [lineStart, lineEnd] = enclosingLine(text, start, end);
+    // Both walks stop as soon as the paragraph is too wide to be used, because
+    // everything past that point is discarded by the cap below anyway. Applying
+    // the cap only to the *result* meant that on a document with no blank lines
+    // each walk ran to the top and bottom of the whole text — O(document) per
+    // candidate, O(n^2) overall.
     let paraStart = lineStart;
     while (paraStart > 0) {
+        if (lineEnd - paraStart > MAX_UNIT_CHARS)
+            return text.slice(lineStart, lineEnd);
         const prevEnd = paraStart - 1;
         const prevStart = text.lastIndexOf("\n", prevEnd - 1) + 1;
         if (text.slice(prevStart, prevEnd).trim() === "")
@@ -443,6 +523,8 @@ function structuralUnit(text, start, end) {
     }
     let paraEnd = lineEnd;
     while (paraEnd < text.length) {
+        if (paraEnd - paraStart > MAX_UNIT_CHARS)
+            return text.slice(lineStart, lineEnd);
         const nextStart = paraEnd + 1;
         let nextEnd = text.indexOf("\n", nextStart);
         if (nextEnd === -1)
@@ -479,28 +561,73 @@ function escapeRe(s) {
  * heading. Measured on the corpus this alone identifies ~78% of the BIC false
  * positives with no dictionaries and no labelled data.
  */
-function occursAsLowercaseWord(text, token) {
+/**
+ * Every all-letter word in the document that is not part of an email address
+ * or a domain name.
+ *
+ * `\b<word>\b` can only match a maximal run of word characters, so collecting
+ * the runs once answers the question for every candidate. Re-scanning the whole
+ * document per candidate made this quadratic: 1.58 s on a 400 KB document
+ * dense in bank codes, against ~1 ms to build this set.
+ */
+function buildDocumentWords(text) {
+    const words = new Set();
+    const re = /[A-Za-z0-9_]+/g;
+    let m;
+    while ((m = re.exec(text)) !== null) {
+        const word = m[0];
+        if (!/^[A-Za-z]+$/.test(word))
+            continue;
+        const i = m.index;
+        const before = i > 0 ? text[i - 1] : "";
+        if (before === "@")
+            continue;
+        if (before === "." && i >= 2 && /[A-Za-z0-9]/.test(text[i - 2]))
+            continue;
+        const after = text.slice(i + word.length, i + word.length + 8);
+        if (after.startsWith("@"))
+            continue;
+        if (/^\.[a-zA-Z]{2,6}\b/.test(after))
+            continue;
+        words.add(word);
+    }
+    return words;
+}
+/**
+ * Per-document scratch space for suppressors that would otherwise re-derive the
+ * same whole-document facts for every candidate.
+ *
+ * Built lazily and held only for the duration of one `detect()` call, so no
+ * document text outlives the call that supplied it.
+ */
+export class SuppressionScratch {
+    text;
+    words = null;
+    constructor(text) {
+        this.text = text;
+    }
+    documentWords() {
+        if (this.words === null)
+            this.words = buildDocumentWords(this.text);
+        return this.words;
+    }
+}
+/**
+ * Does `token` also occur as an ordinary lowercase word in this document?
+ *
+ * `hospital`/`HOSPITAL`, `gegevens`/`GEGEVENS` — a token that appears in
+ * ordinary case elsewhere in the same document is a word, not a bank code.
+ * Email and domain contexts are excluded, so `ing.nl` does not vouch for a
+ * heading. Measured on the corpus this alone identifies ~78% of the BIC false
+ * positives with no dictionaries and no labelled data.
+ */
+function occursAsLowercaseWord(text, token, scratch) {
     if (!/^[A-Za-z]+$/.test(token))
         return false;
     const lower = token.toLowerCase();
     const capitalised = lower[0].toUpperCase() + lower.slice(1);
-    for (const form of [lower, capitalised]) {
-        // Cheap pre-check: a plain substring test almost always answers "no".
-        if (!text.includes(form))
-            continue;
-        const re = new RegExp(`\\b${escapeRe(form)}\\b`, "g");
-        let m;
-        while ((m = re.exec(text)) !== null) {
-            const before = m.index > 0 ? text[m.index - 1] : "";
-            const after = text.slice(m.index + form.length, m.index + form.length + 8);
-            if (before === "@" || (before === "." && m.index >= 2 && /[A-Za-z0-9]/.test(text[m.index - 2])))
-                continue;
-            if (after.startsWith("@") || /^\.[a-zA-Z]{2,6}\b/.test(after))
-                continue;
-            return true;
-        }
-    }
-    return false;
+    const words = scratch ? scratch.documentWords() : buildDocumentWords(text);
+    return words.has(lower) || words.has(capitalised);
 }
 /**
  * Is the candidate positioned as a section heading rather than a value?
@@ -538,25 +665,29 @@ function isHeadingShape(text, start, end, token) {
  * ordinary lowercase word — so it is applied on both emitting paths. It scans
  * the whole document, so it runs only on the paths that would otherwise emit.
  */
-function suppressBicWithoutEvidence(text, match) {
+function suppressBicWithoutEvidence(text, match, scratch) {
     if (match.patternDef.entityType !== EntityType.BIC)
         return false;
     const token = match.text.trim();
     if (isRegisteredBic(token))
-        return occursAsLowercaseWord(text, token);
+        return occursAsLowercaseWord(text, token, scratch);
     if (isHeadingShape(text, match.start, match.end, token))
         return true;
     // The token itself is blanked out so it cannot vouch for itself.
     const unit = structuralUnit(text, match.start, match.end).split(token).join(" ".repeat(token.length));
     if (BIC_KEYWORD.test(unit) || IBAN_SHAPE.test(unit) || BANK_BLOCK.test(unit)) {
-        return occursAsLowercaseWord(text, token);
+        return occursAsLowercaseWord(text, token, scratch);
     }
     return true;
 }
 // ── Postal code: digits belonging to a longer identifier ────────────────
 // An identifier label immediately before the digits. A postal code is never
 // introduced this way; an SVNr, policy number or service number always is.
-const ID_CUE_BEFORE = /(?:[\w\-]*Nr|[\w\-]*N[°ºo]|[\w\-]*Nummer|[\w\-]*Numero|[\w\-]*Numéro|No|number|num|Kennzahl|Aktenzeichen|Az|e-?card|Polizze|Police|Policen)\.?\s*:?\s*$/i;
+// The `[\w\-]*` prefix is factored out of the five alternatives that share it.
+// Repeated per alternative, the engine retried the same unbounded prefix five
+// times per starting position — the most expensive single regex in the engine
+// at 197 ms per 1 MB document. Same language, 8x faster on the worst case.
+const ID_CUE_BEFORE = /(?:[\w\-]*(?:Nr|N[°ºo]|Nummer|Numero|Numéro)|No|number|num|Kennzahl|Aktenzeichen|Az|e-?card|Polizze|Police|Policen)\.?\s*:?\s*$/i;
 // An international dialling prefix earlier on the same line, with nothing but
 // number punctuation in between: these digits belong to the phone detector.
 const DIALLING_PREFIX_BEFORE = /\+\d{1,3}[\d\s\-().]*$/;
@@ -661,20 +792,20 @@ const TYPE_SUPPRESSORS = {
     [EntityType.NATIONAL_ID]: [suppressCurrency, suppressUnits, suppressReference, suppressLegal, suppressMath, suppressNatidAsPassport, suppressSeNatidAsOrg],
     [EntityType.SSN]: [suppressCurrency, suppressUnits, suppressReference, suppressMath],
     [EntityType.TAX_ID]: [suppressCurrency, suppressUnits, suppressReference, suppressMath, suppressTaxidAsIpAddress],
-    [EntityType.POSTAL_CODE]: [suppressCurrency, suppressUnits, suppressMath, suppressLegal, suppressYearAsPostal, suppressPostalInsideIban, suppressPostalAsHouseNumber, suppressPostalInLongerIdentifier],
+    [EntityType.POSTAL_CODE]: [suppressCurrency, suppressUnits, suppressMath, suppressLegal, suppressYearAsPostal, suppressPostalInsideIban, suppressPostalAsHouseNumber, suppressPostalInLongerIdentifier, suppressReference],
     [EntityType.BIC]: [suppressBicWithoutEvidence],
     [EntityType.BANK_ACCOUNT]: [suppressReference],
     [EntityType.LICENSE_PLATE]: [suppressPlateInCompound, suppressDePlateUnknownDistrict, suppressPlateAsCurrencyAmount],
     [EntityType.SECRET]: [suppressSecretOverStructured, suppressSecretNotASecret],
     [EntityType.CHAMBER_OF_COMMERCE]: [suppressReference],
 };
-export function shouldSuppress(text, match) {
+export function shouldSuppress(text, match, scratch) {
     if (suppressSequential(text, match))
         return true;
     const typeSups = TYPE_SUPPRESSORS[match.patternDef.entityType];
     if (typeSups) {
         for (const s of typeSups) {
-            if (s(text, match))
+            if (s(text, match, scratch))
                 return true;
         }
     }

package/dist/esm/rules/validators.d.ts

@@ -17,6 +17,18 @@ export declare function validateVatFr(candidate: string): boolean;
 export declare function validateVatLu(candidate: string): boolean;
 export declare function validateGermanTaxId(candidate: string): boolean;
 export declare function validateFrenchNir(candidate: string): boolean;
+/**
+ * ISO 3779 VIN validation: 17 characters, no I/O/Q. **Shape only.**
+ *
+ * The position-9 check digit is mandatory for North American VINs (WMI 1-5)
+ * and unused elsewhere, and it is deliberately **not** enforced. Measured over
+ * the 152,300-document corpus, enforcing it turned 1,502 labelled VINs into
+ * misses — every North American VIN whose check digit does not survive being
+ * written down. A VIN that fails its checksum is still a VIN sitting in the
+ * text; dropping it leaves it unredacted, which is a leak, while keeping it
+ * costs at worst an over-masked 17-character token. Matches the Python SDK's
+ * `validate_vin`, which carries the full reasoning.
+ */
 export declare function validateVin(candidate: string): boolean;
 /**
  * BIC/SWIFT structural validation (ISO 9362): BBBB CC LL (PPP).

package/dist/esm/rules/validators.js

@@ -179,6 +179,18 @@ export function validateFrenchNir(candidate) {
     const check = parseInt(numeric.slice(13, 15));
     return (97 - (first13 % 97)) === check;
 }
+/**
+ * ISO 3779 VIN validation: 17 characters, no I/O/Q. **Shape only.**
+ *
+ * The position-9 check digit is mandatory for North American VINs (WMI 1-5)
+ * and unused elsewhere, and it is deliberately **not** enforced. Measured over
+ * the 152,300-document corpus, enforcing it turned 1,502 labelled VINs into
+ * misses — every North American VIN whose check digit does not survive being
+ * written down. A VIN that fails its checksum is still a VIN sitting in the
+ * text; dropping it leaves it unredacted, which is a leak, while keeping it
+ * costs at worst an over-masked 17-character token. Matches the Python SDK's
+ * `validate_vin`, which carries the full reasoning.
+ */
 export function validateVin(candidate) {
     const c = candidate.replace(/[\s\-]/g, "").toUpperCase();
     if (c.length !== 17)

package/dist/esm/sdk.js

@@ -4,20 +4,44 @@ import { RuleEngine } from "./rules/engine.js";
 import { weightsToRanking } from "./rules/evidence.js";
 import { EntityType } from "./types.js";
 const DATE_TYPES = new Set([EntityType.DOB, EntityType.DATE_OF_DEATH]);
+/** Entry count at which a one-time warning is emitted. */
+const MAPPING_WARN_THRESHOLD = 100_000;
+/**
+ * Maps real PII values to consistent labels within a session.
+ *
+ * The mapping is keyed on the **raw PII value** and is never evicted — evicting
+ * would hand a previously seen value a second label and quietly break
+ * referential integrity. Two consequences worth designing around:
+ *
+ * - It grows for as long as the process runs. Call `clear()` between workloads;
+ *   a warning is emitted once the mapping passes `MAPPING_WARN_THRESHOLD`.
+ * - Labels are shared by every caller of the same instance, including every
+ *   caller of the module-level `redact()`. A label repeated across two
+ *   documents reveals that they contain the same underlying value, so give each
+ *   tenant its own `EuRedact` instance rather than sharing the module-level one.
+ */
 class ReferentialMapper {
     counters = new Map();
     mapping = new Map();
+    warned = false;
     getLabel(text, entityType) {
         if (!this.mapping.has(text)) {
             const count = (this.counters.get(entityType) ?? 0) + 1;
             this.counters.set(entityType, count);
             this.mapping.set(text, `${entityType}_${count}`);
+            if (!this.warned && this.mapping.size > MAPPING_WARN_THRESHOLD) {
+                this.warned = true;
+                console.warn(`[euredact] Referential integrity mapping holds ${this.mapping.size} ` +
+                    `raw PII values and is never evicted. Call clear() between workloads ` +
+                    `to release them.`);
+            }
         }
         return this.mapping.get(text);
     }
     clear() {
         this.counters.clear();
         this.mapping.clear();
+        this.warned = false;
     }
 }
 const DEFAULT_MAX_INPUT_LENGTH = 10_485_760; // ~10 MB of text
@@ -107,14 +131,32 @@ export class EuRedact {
             detections = detections.filter(d => !DATE_TYPES.has(d.entityType));
         }
         detections.sort((a, b) => a.start - b.start || b.end - a.end);
-        let redacted = text;
+        // Labels are resolved right-to-left because the referential mapper numbers
+        // each entity type in call order, and that order is part of the output
+        // contract. The string itself is then assembled in a single forward pass:
+        // rebuilding it per detection copied the whole document each time, which is
+        // O(document x detections) — 1.68 s of pure copying on a 1 MB document with
+        // 15,000 detections, versus 2 ms here.
+        const replacements = new Array(detections.length);
         for (let i = detections.length - 1; i >= 0; i--) {
             const det = detections[i];
-            const replacement = referentialIntegrity
+            replacements[i] = referentialIntegrity
                 ? this.referentialMapper.getLabel(det.text, det.entityType)
                 : `[${det.entityType}]`;
-            redacted = redacted.slice(0, det.start) + replacement + redacted.slice(det.end);
         }
+        const parts = [];
+        let pos = 0;
+        for (let i = 0; i < detections.length; i++) {
+            const det = detections[i];
+            // Spans reach here deduplicated and non-overlapping; one starting behind
+            // the cursor would splice into the middle of a label, so drop it.
+            if (det.start < pos)
+                continue;
+            parts.push(text.slice(pos, det.start), replacements[i]);
+            pos = det.end;
+        }
+        parts.push(text.slice(pos));
+        const redacted = parts.join("");
         // Report the inference so it can be audited. Spans in `evidence` are
         // offsets into the normalised text, matching `detections`.
         const inferredCountries = [...weightsToRanking(scores).entries()]

package/package.json

@@ -1,7 +1,8 @@
 {
   "name": "euredact",
-  "version": "0.3.5",
+  "version": "0.3.7",
   "description": "European PII redaction SDK -- fast, zero-dependency, rule-based detection and redaction of personal data across 31 European countries",
+  "homepage": "https://euredact.dev",
   "license": "Apache-2.0",
   "type": "module",
   "main": "./dist/cjs/index.js",
@@ -64,7 +65,11 @@
     "dump": "node --import tsx src/__tests__/dump.ts"
   },
   "devDependencies": {
-    "typescript": "^5.4.0",
-    "tsx": "^4.7.0"
+    "@types/node": "22.10.2",
+    "tsx": "^4.7.0",
+    "typescript": "^5.4.0"
+  },
+  "bugs": {
+    "url": "https://github.com/euredact/euredact/issues"
   }
 }