euredact 0.3.6 → 0.3.7

package/README.md

@@ -313,7 +313,7 @@ for (const page of pages) {
   offset += page.length;
 }
 
-// page 1: "Factuur — IBAN NL91ABNA0417164300, info@jansen.nl"
+// page 1: "Factuur — IBAN NL91ABNA0417164300, info@example.nl"
 // page 7: "Telefoon 0612345678"  -> PHONE (NL), not NATIONAL_ID (DK)
 ```
 

package/dist/cjs/cache.js

@@ -1,30 +1,44 @@
 "use strict";
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.ResultCache = void 0;
-// Simple hash function (FNV-1a) - no crypto dependency needed for cache keys
-function fnv1aHash(str) {
-    let hash = 0x811c9dc5;
-    for (let i = 0; i < str.length; i++) {
-        hash ^= str.charCodeAt(i);
-        hash = (hash * 0x01000193) >>> 0;
-    }
-    return hash.toString(16).padStart(8, "0");
+const node_crypto_1 = require("node:crypto");
+// Default ceiling on retained characters, not entries. A count-only cap of
+// 1024 entries says nothing about memory: 1024 cached 1 MB documents is a
+// gigabyte of PII-bearing text held live. ~16M characters is a few tens of MB.
+const DEFAULT_MAX_CHARS = 16_000_000;
+/** Approximate retained size of a cached result, in characters. */
+function resultChars(result) {
+    let chars = result.redactedText.length;
+    for (const d of result.detections)
+        chars += d.text.length;
+    return chars;
 }
 class ResultCache {
     maxsize;
+    maxChars;
     enabled;
     store = new Map();
-    constructor(maxsize = 1024, enabled = true) {
+    sizes = new Map();
+    chars = 0;
+    constructor(maxsize = 1024, enabled = true, maxChars = DEFAULT_MAX_CHARS) {
         this.maxsize = maxsize;
+        this.maxChars = maxChars;
         this.enabled = enabled;
     }
     key(text, countries, mode) {
         const sorted = [...countries].sort();
-        const raw = `${text}|${sorted.join("|")}|${mode}`;
-        // Include raw length to reduce collision risk alongside dual FNV-1a hashes
-        const h1 = fnv1aHash(raw);
-        const h2 = fnv1aHash(raw.slice(Math.floor(raw.length / 2)) + raw.slice(0, Math.floor(raw.length / 2)));
-        return `${raw.length}:${h1}${h2}`;
+        // SHA-256, matching the Python implementation. The previous dual FNV-1a
+        // was non-cryptographic and collidable by construction: in a shared-process
+        // deployment an attacker who can guess a victim's document could submit a
+        // same-length colliding one and be served the victim's cached result,
+        // detections and raw matched PII included.
+        return (0, node_crypto_1.createHash)("sha256")
+            .update(text)
+            .update("|")
+            .update(sorted.join("|"))
+            .update("|")
+            .update(mode)
+            .digest("hex");
     }
     get(key) {
         if (!this.enabled)
@@ -41,16 +55,36 @@ class ResultCache {
     put(key, result) {
         if (!this.enabled)
             return;
-        this.store.delete(key);
+        const size = resultChars(result);
+        // A single result bigger than the whole budget would evict everything else
+        // and still not fit; skip it rather than empty the cache for one document.
+        if (size > this.maxChars) {
+            this.evict(key);
+            return;
+        }
+        this.evict(key);
         this.store.set(key, result);
-        while (this.store.size > this.maxsize) {
+        this.sizes.set(key, size);
+        this.chars += size;
+        while (this.store.size > this.maxsize || this.chars > this.maxChars) {
             const first = this.store.keys().next().value;
-            if (first !== undefined)
-                this.store.delete(first);
+            if (first === undefined)
+                break;
+            this.evict(first);
         }
     }
+    evict(key) {
+        const size = this.sizes.get(key);
+        if (size !== undefined) {
+            this.chars -= size;
+            this.sizes.delete(key);
+        }
+        this.store.delete(key);
+    }
     clear() {
         this.store.clear();
+        this.sizes.clear();
+        this.chars = 0;
     }
 }
 exports.ResultCache = ResultCache;

package/dist/cjs/rules/context.js

@@ -7,7 +7,7 @@
  * contain no IBAN, no `+CC` number and no ccTLD infers nothing at all, even
  * when the first page identified the document beyond doubt:
  *
- *     page 1   "Factuur — IBAN NL91ABNA0417164300, info@jansen.nl"
+ *     page 1   "Factuur — IBAN NL91ABNA0417164300, info@example.nl"
  *     page 7   "Telefoon 0612345678"        <- Dutch mobile, or Danish CPR?
  *
  * Scanned alone, page 7 has nothing to go on. Sharing a context across the two

package/dist/cjs/rules/countries/index.js

@@ -160,7 +160,20 @@ const SHARED = {
         // --- Secret / API Key (assignment-based: KEY=value or KEY: value) ---
         p(types_js_1.EntityType.SECRET, String.raw `(?<=[:=] )\S{8,}|(?<=[:=])\S{8,}`, "high_entropy", "Assigned secret value", SECRET_CONTEXT, true),
         // --- Secret / API Key (entropy-based fallback for longer tokens) ---
-        p(types_js_1.EntityType.SECRET, String.raw `\b[A-Za-z0-9_\-+/]{24,}[A-Za-z0-9_\-+/=]*\b`, "high_entropy", "High-entropy token", SECRET_CONTEXT, true),
+        // Only the *trailing* \b is replaced. It could not match when the token run
+        // ended on "-", "+" or "/", so the engine backtracked across two
+        // overlapping quantified classes: an ordinary dashed rule line ("x-----...")
+        // cost O(n^2), 8.3 s for 80 KB. The lookahead simply asserts the run is
+        // over, which is what \b was there to express, and it cannot fail after a
+        // maximal run — so there is nothing to backtrack into.
+        //
+        // The leading \b stays. A class lookbehind there widened the span leftwards
+        // over any "/" in front, so an authorized_keys entry started "//...",
+        // matched the protocol-relative URL shape in the endpoint suppressor, and
+        // was dropped — leaving SSH key material in the clear. Base64 padding is
+        // now inside the span instead of trailing outside it; that is the only
+        // intended difference.
+        p(types_js_1.EntityType.SECRET, String.raw `\b[A-Za-z0-9_\-+/]{24,}[A-Za-z0-9_\-+/=]*(?![A-Za-z0-9_\-+/=])`, "high_entropy", "High-entropy token", SECRET_CONTEXT, true),
         p(types_js_1.EntityType.DOB, String.raw `\b(?:0[1-9]|[12][0-9]|3[01])[/.\-](?:0[1-9]|1[0-2])[/.\-](?:19|20)\d{2}\b`, null, "DD/MM/YYYY", DOB_CONTEXT, true),
         p(types_js_1.EntityType.DATE_OF_DEATH, String.raw `\b(?:0[1-9]|[12][0-9]|3[01])[/.\-](?:0[1-9]|1[0-2])[/.\-](?:19|20)\d{2}\b`, null, "", DATE_OF_DEATH_CONTEXT, true),
         p(types_js_1.EntityType.DOB, String.raw `\b(?:19|20)\d{2}[/.\-](?:0[1-9]|1[0-2])[/.\-](?:0[1-9]|[12][0-9]|3[01])\b`, null, "ISO YYYY-MM-DD", DOB_ISO_CONTEXT, true),

package/dist/cjs/rules/engine.js

@@ -310,6 +310,14 @@ class RuleEngine {
         const allFailed = failedSpans
             .map(f => [f.start, f.end])
             .sort((a, b) => a[0] - b[0] || a[1] - b[1]);
+        // Running maximum of the end offsets, so the fragment test below can ask
+        // "does any span starting at or before me reach past my end?" in O(1)
+        // instead of walking the prefix per candidate.
+        const failedStarts = allFailed.map(f => f[0]);
+        const failedPrefixMax = new Array(allFailed.length + 1).fill(0);
+        for (let i = 0; i < allFailed.length; i++) {
+            failedPrefixMax[i + 1] = Math.max(failedPrefixMax[i], allFailed[i][1]);
+        }
         const mergedZones = new Map();
         for (const [etype, spans] of zonesByType) {
             spans.sort((a, b) => a[0] - b[0] || a[1] - b[1]);
@@ -332,8 +340,11 @@ class RuleEngine {
             mergedZones.set(etype, { starts, ends });
         }
         const candidates = [];
+        // Whole-document facts a suppressor may need, derived at most once per call
+        // and discarded with it.
+        const scratch = new suppressors_js_1.SuppressionScratch(text);
         for (const { match, hasValidValidator } of validated) {
-            if ((0, suppressors_js_1.shouldSuppress)(text, match))
+            if ((0, suppressors_js_1.shouldSuppress)(text, match, scratch))
                 continue;
             // A validator-less match inside a failed-validation span of its own type
             // is demoted below every other candidate rather than deleted. Demotion
@@ -361,13 +372,31 @@ class RuleEngine {
             // default — the generic phone pattern kept claiming the tail of the
             // rejected national number.
             if (match.patternDef.validator === null && !demoted) {
+                // Upper bound of the spans starting at or before match.start.
+                let lo = 0, hi = failedStarts.length;
+                while (lo < hi) {
+                    const mid = (lo + hi) >>> 1;
+                    if (failedStarts[mid] <= match.start)
+                        lo = mid + 1;
+                    else
+                        hi = mid;
+                }
                 let fragment = false;
-                for (const [zs, ze] of allFailed) {
-                    if (zs > match.start)
-                        break;
-                    if (match.end <= ze && (match.start > zs || match.end < ze)) {
-                        fragment = true;
-                        break;
+                const reach = failedPrefixMax[lo];
+                if (reach > match.end) {
+                    // Something starting no later than us ends strictly later: we are a
+                    // proper subset of it.
+                    fragment = true;
+                }
+                else if (reach === match.end) {
+                    // Same end offset — a proper subset only if some span with that end
+                    // starts strictly earlier. Scanning backwards finds the enclosing
+                    // span first, and the equal case is rare.
+                    for (let i = lo - 1; i >= 0; i--) {
+                        if (allFailed[i][1] === match.end && allFailed[i][0] < match.start) {
+                            fragment = true;
+                            break;
+                        }
                     }
                 }
                 if (fragment)
@@ -418,6 +447,7 @@ class RuleEngine {
                 length: match.end - match.start,
                 countryScore: countryScores.get(match.countryCode) ?? 0,
                 inScope: declared === null || declared.has(match.countryCode) ? 1 : 0,
+                tier: 0,
                 det: {
                     entityType: match.patternDef.entityType,
                     start: match.start,
@@ -432,7 +462,7 @@ class RuleEngine {
         }
         for (const d of (0, structural_js_1.detectStructuralDob)(text)) {
             candidates.push({ blindPriority: 1, cue: 0, priority: 1, length: d.end - d.start,
-                countryScore: 0, inScope: 1, det: d });
+                countryScore: 0, inScope: 1, tier: 0, det: d });
         }
         const detections = this.deduplicate(candidates);
         detections.sort((a, b) => a.start - b.start || b.end - a.end);
@@ -466,7 +496,10 @@ class RuleEngine {
             if (seen === undefined || c.blindPriority > seen)
                 spanTier.set(key, c.blindPriority);
         }
-        const tierOf = (c) => spanTier.get(`${c.det.start}:${c.det.end}`);
+        // Resolved once per candidate rather than inside the comparator, which ran
+        // it O(n log n) times and rebuilt the key string on every comparison.
+        for (const c of candidates)
+            c.tier = spanTier.get(`${c.det.start}:${c.det.end}`);
         // Span length outranks *everything*, including validation.
         //
         // A shorter candidate winning re-cuts a longer one and leaves the
@@ -492,7 +525,7 @@ class RuleEngine {
         // country-blind up to and including `start`, and (length, start) already
         // identifies a span uniquely.
         const sorted = [...candidates].sort((a, b) => b.length - a.length || // longest first     } country-blind
-            tierOf(b) - tierOf(a) || // best tier here    } — decides
+            b.tier - a.tier || // best tier here    } — decides
             a.det.start - b.det.start || // leftmost          } WHICH span
             b.cue - a.cue || // local cue         } country-aware
             b.priority - a.priority || // country priority  } — decides the

package/dist/cjs/rules/suppressors.js

@@ -1,5 +1,6 @@
 "use strict";
 Object.defineProperty(exports, "__esModule", { value: true });
+exports.SuppressionScratch = void 0;
 exports.shouldSuppress = shouldSuppress;
 const types_js_1 = require("../types.js");
 const bicRegistry_js_1 = require("./bicRegistry.js");
@@ -271,8 +272,12 @@ function suppressUnits(text, match) {
  * 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([types_js_1.EntityType.PHONE, types_js_1.EntityType.NATIONAL_ID, types_js_1.EntityType.SSN, types_js_1.EntityType.TAX_ID, types_js_1.EntityType.IBAN, types_js_1.EntityType.CHAMBER_OF_COMMERCE, types_js_1.EntityType.POSTAL_CODE]);
+const LEGAL_TYPES = new Set([types_js_1.EntityType.PHONE, types_js_1.EntityType.NATIONAL_ID, types_js_1.EntityType.POSTAL_CODE]);
 function suppressReference(text, match) {
-    const applicable = new Set([types_js_1.EntityType.PHONE, types_js_1.EntityType.NATIONAL_ID, types_js_1.EntityType.SSN, types_js_1.EntityType.TAX_ID, types_js_1.EntityType.IBAN, types_js_1.EntityType.CHAMBER_OF_COMMERCE, types_js_1.EntityType.POSTAL_CODE]);
+    const applicable = REFERENCE_TYPES;
     if (!applicable.has(match.patternDef.entityType))
         return false;
     const [before] = getContext(text, match.start, match.end);
@@ -282,7 +287,7 @@ function suppressReference(text, match) {
     return HASH_BEFORE.test(adjacent) || REF_PREFIX_BEFORE.test(adjacent);
 }
 function suppressLegal(text, match) {
-    const applicable = new Set([types_js_1.EntityType.PHONE, types_js_1.EntityType.NATIONAL_ID, types_js_1.EntityType.POSTAL_CODE]);
+    const applicable = LEGAL_TYPES;
     if (!applicable.has(match.patternDef.entityType))
         return false;
     const [before] = getContext(text, match.start, match.end);
@@ -505,8 +510,15 @@ 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() === "")
@@ -515,6 +527,8 @@ 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)
@@ -551,28 +565,74 @@ 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]+$/.test(word))
+            continue;
+        const i = m.index;
+        const before = i > 0 ? text[i - 1] : "";
+        if (before === "@")
+            continue;
+        if (before === "." && i >= 2 && /[A-Za-z0-9]/.test(text[i - 2]))
+            continue;
+        const after = text.slice(i + word.length, i + word.length + 8);
+        if (after.startsWith("@"))
+            continue;
+        if (/^\.[a-zA-Z]{2,6}\b/.test(after))
+            continue;
+        words.add(word);
+    }
+    return words;
+}
+/**
+ * Per-document scratch space for suppressors that would otherwise re-derive the
+ * same whole-document facts for every candidate.
+ *
+ * Built lazily and held only for the duration of one `detect()` call, so no
+ * document text outlives the call that supplied it.
+ */
+class SuppressionScratch {
+    text;
+    words = null;
+    constructor(text) {
+        this.text = text;
+    }
+    documentWords() {
+        if (this.words === null)
+            this.words = buildDocumentWords(this.text);
+        return this.words;
+    }
+}
+exports.SuppressionScratch = SuppressionScratch;
+/**
+ * Does `token` also occur as an ordinary lowercase word in this document?
+ *
+ * `hospital`/`HOSPITAL`, `gegevens`/`GEGEVENS` — a token that appears in
+ * ordinary case elsewhere in the same document is a word, not a bank code.
+ * Email and domain contexts are excluded, so `ing.nl` does not vouch for a
+ * heading. Measured on the corpus this alone identifies ~78% of the BIC false
+ * positives with no dictionaries and no labelled data.
+ */
+function occursAsLowercaseWord(text, token, scratch) {
     if (!/^[A-Za-z]+$/.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?
@@ -610,25 +670,29 @@ 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 !== types_js_1.EntityType.BIC)
         return false;
     const token = match.text.trim();
     if ((0, bicRegistry_js_1.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*$/i;
 // An international dialling prefix earlier on the same line, with nothing but
 // number punctuation in between: these digits belong to the phone detector.
 const DIALLING_PREFIX_BEFORE = /\+\d{1,3}[\d\s\-().]*$/;
@@ -740,13 +804,13 @@ const TYPE_SUPPRESSORS = {
     [types_js_1.EntityType.SECRET]: [suppressSecretOverStructured, suppressSecretNotASecret],
     [types_js_1.EntityType.CHAMBER_OF_COMMERCE]: [suppressReference],
 };
-function shouldSuppress(text, match) {
+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/cjs/rules/validators.js

@@ -225,6 +225,18 @@ 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.
+ */
 function validateVin(candidate) {
     const c = candidate.replace(/[\s\-]/g, "").toUpperCase();
     if (c.length !== 17)

package/dist/cjs/sdk.js

@@ -7,20 +7,44 @@ const engine_js_1 = require("./rules/engine.js");
 const evidence_js_1 = require("./rules/evidence.js");
 const types_js_1 = require("./types.js");
 const DATE_TYPES = new Set([types_js_1.EntityType.DOB, types_js_1.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 ${this.mapping.size} ` +
+                    `raw PII values and is never evicted. Call clear() between workloads ` +
+                    `to release them.`);
+            }
         }
         return this.mapping.get(text);
     }
     clear() {
         this.counters.clear();
         this.mapping.clear();
+        this.warned = false;
     }
 }
 const DEFAULT_MAX_INPUT_LENGTH = 10_485_760; // ~10 MB of text
@@ -110,14 +134,32 @@ class EuRedact {
             detections = detections.filter(d => !DATE_TYPES.has(d.entityType));
         }
         detections.sort((a, b) => a.start - b.start || b.end - a.end);
-        let redacted = text;
+        // Labels are resolved right-to-left because the referential mapper numbers
+        // each entity type in call order, and that order is part of the output
+        // contract. The string itself is then assembled in a single forward pass:
+        // rebuilding it per detection copied the whole document each time, which is
+        // O(document x detections) — 1.68 s of pure copying on a 1 MB document with
+        // 15,000 detections, versus 2 ms here.
+        const replacements = new Array(detections.length);
         for (let i = detections.length - 1; i >= 0; i--) {
             const det = detections[i];
-            const replacement = referentialIntegrity
+            replacements[i] = referentialIntegrity
                 ? this.referentialMapper.getLabel(det.text, det.entityType)
                 : `[${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 = [...(0, evidence_js_1.weightsToRanking)(scores).entries()]

package/dist/esm/cache.d.ts

@@ -1,11 +1,15 @@
 import type { RedactResult } from "./types.js";
 export declare class ResultCache {
     private maxsize;
+    private maxChars;
     private enabled;
     private store;
-    constructor(maxsize?: number, enabled?: boolean);
+    private sizes;
+    private chars;
+    constructor(maxsize?: number, enabled?: boolean, maxChars?: number);
     key(text: string, countries: string[], mode: string): string;
     get(key: string): RedactResult | null;
     put(key: string, result: RedactResult): void;
+    private evict;
     clear(): void;
 }

package/dist/esm/cache.js

@@ -1,27 +1,41 @@
-// Simple hash function (FNV-1a) - no crypto dependency needed for cache keys
-function fnv1aHash(str) {
-    let hash = 0x811c9dc5;
-    for (let i = 0; i < str.length; i++) {
-        hash ^= str.charCodeAt(i);
-        hash = (hash * 0x01000193) >>> 0;
-    }
-    return hash.toString(16).padStart(8, "0");
+import { createHash } from "node:crypto";
+// Default ceiling on retained characters, not entries. A count-only cap of
+// 1024 entries says nothing about memory: 1024 cached 1 MB documents is a
+// gigabyte of PII-bearing text held live. ~16M characters is a few tens of MB.
+const DEFAULT_MAX_CHARS = 16_000_000;
+/** Approximate retained size of a cached result, in characters. */
+function resultChars(result) {
+    let chars = result.redactedText.length;
+    for (const d of result.detections)
+        chars += d.text.length;
+    return chars;
 }
 export class ResultCache {
     maxsize;
+    maxChars;
     enabled;
     store = new Map();
-    constructor(maxsize = 1024, enabled = true) {
+    sizes = new Map();
+    chars = 0;
+    constructor(maxsize = 1024, enabled = true, maxChars = DEFAULT_MAX_CHARS) {
         this.maxsize = maxsize;
+        this.maxChars = maxChars;
         this.enabled = enabled;
     }
     key(text, countries, mode) {
         const sorted = [...countries].sort();
-        const raw = `${text}|${sorted.join("|")}|${mode}`;
-        // Include raw length to reduce collision risk alongside dual FNV-1a hashes
-        const h1 = fnv1aHash(raw);
-        const h2 = fnv1aHash(raw.slice(Math.floor(raw.length / 2)) + raw.slice(0, Math.floor(raw.length / 2)));
-        return `${raw.length}:${h1}${h2}`;
+        // SHA-256, matching the Python implementation. The previous dual FNV-1a
+        // was non-cryptographic and collidable by construction: in a shared-process
+        // deployment an attacker who can guess a victim's document could submit a
+        // same-length colliding one and be served the victim's cached result,
+        // detections and raw matched PII included.
+        return createHash("sha256")
+            .update(text)
+            .update("|")
+            .update(sorted.join("|"))
+            .update("|")
+            .update(mode)
+            .digest("hex");
     }
     get(key) {
         if (!this.enabled)
@@ -38,15 +52,35 @@ export class ResultCache {
     put(key, result) {
         if (!this.enabled)
             return;
-        this.store.delete(key);
+        const size = resultChars(result);
+        // A single result bigger than the whole budget would evict everything else
+        // and still not fit; skip it rather than empty the cache for one document.
+        if (size > this.maxChars) {
+            this.evict(key);
+            return;
+        }
+        this.evict(key);
         this.store.set(key, result);
-        while (this.store.size > this.maxsize) {
+        this.sizes.set(key, size);
+        this.chars += size;
+        while (this.store.size > this.maxsize || this.chars > this.maxChars) {
             const first = this.store.keys().next().value;
-            if (first !== undefined)
-                this.store.delete(first);
+            if (first === undefined)
+                break;
+            this.evict(first);
         }
     }
+    evict(key) {
+        const size = this.sizes.get(key);
+        if (size !== undefined) {
+            this.chars -= size;
+            this.sizes.delete(key);
+        }
+        this.store.delete(key);
+    }
     clear() {
         this.store.clear();
+        this.sizes.clear();
+        this.chars = 0;
     }
 }

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

@@ -6,7 +6,7 @@
  * contain no IBAN, no `+CC` number and no ccTLD infers nothing at all, even
  * when the first page identified the document beyond doubt:
  *
- *     page 1   "Factuur — IBAN NL91ABNA0417164300, info@jansen.nl"
+ *     page 1   "Factuur — IBAN NL91ABNA0417164300, info@example.nl"
  *     page 7   "Telefoon 0612345678"        <- Dutch mobile, or Danish CPR?
  *
  * Scanned alone, page 7 has nothing to go on. Sharing a context across the two

package/dist/esm/rules/context.js

@@ -6,7 +6,7 @@
  * contain no IBAN, no `+CC` number and no ccTLD infers nothing at all, even
  * when the first page identified the document beyond doubt:
  *
- *     page 1   "Factuur — IBAN NL91ABNA0417164300, info@jansen.nl"
+ *     page 1   "Factuur — IBAN NL91ABNA0417164300, info@example.nl"
  *     page 7   "Telefoon 0612345678"        <- Dutch mobile, or Danish CPR?
  *
  * Scanned alone, page 7 has nothing to go on. Sharing a context across the two

package/dist/esm/rules/countries/index.js

@@ -157,7 +157,20 @@ const SHARED = {
         // --- Secret / API Key (assignment-based: KEY=value or KEY: value) ---
         p(EntityType.SECRET, String.raw `(?<=[:=] )\S{8,}|(?<=[:=])\S{8,}`, "high_entropy", "Assigned secret value", SECRET_CONTEXT, true),
         // --- Secret / API Key (entropy-based fallback for longer tokens) ---
-        p(EntityType.SECRET, String.raw `\b[A-Za-z0-9_\-+/]{24,}[A-Za-z0-9_\-+/=]*\b`, "high_entropy", "High-entropy token", SECRET_CONTEXT, true),
+        // Only the *trailing* \b is replaced. It could not match when the token run
+        // ended on "-", "+" or "/", so the engine backtracked across two
+        // overlapping quantified classes: an ordinary dashed rule line ("x-----...")
+        // cost O(n^2), 8.3 s for 80 KB. The lookahead simply asserts the run is
+        // over, which is what \b was there to express, and it cannot fail after a
+        // maximal run — so there is nothing to backtrack into.
+        //
+        // The leading \b stays. A class lookbehind there widened the span leftwards
+        // over any "/" in front, so an authorized_keys entry started "//...",
+        // matched the protocol-relative URL shape in the endpoint suppressor, and
+        // was dropped — leaving SSH key material in the clear. Base64 padding is
+        // now inside the span instead of trailing outside it; that is the only
+        // intended difference.
+        p(EntityType.SECRET, String.raw `\b[A-Za-z0-9_\-+/]{24,}[A-Za-z0-9_\-+/=]*(?![A-Za-z0-9_\-+/=])`, "high_entropy", "High-entropy token", SECRET_CONTEXT, true),
         p(EntityType.DOB, String.raw `\b(?:0[1-9]|[12][0-9]|3[01])[/.\-](?:0[1-9]|1[0-2])[/.\-](?:19|20)\d{2}\b`, null, "DD/MM/YYYY", DOB_CONTEXT, true),
         p(EntityType.DATE_OF_DEATH, String.raw `\b(?:0[1-9]|[12][0-9]|3[01])[/.\-](?:0[1-9]|1[0-2])[/.\-](?:19|20)\d{2}\b`, null, "", DATE_OF_DEATH_CONTEXT, true),
         p(EntityType.DOB, String.raw `\b(?:19|20)\d{2}[/.\-](?:0[1-9]|1[0-2])[/.\-](?:0[1-9]|[12][0-9]|3[01])\b`, null, "ISO YYYY-MM-DD", DOB_ISO_CONTEXT, true),

package/dist/esm/rules/engine.js

@@ -1,6 +1,6 @@
 import { DetectionSource, EntityType } from "../types.js";
 import { MultiPatternMatcher } from "./matchers.js";
-import { shouldSuppress } from "./suppressors.js";
+import { shouldSuppress, SuppressionScratch } from "./suppressors.js";
 import { detectStructuralDob } from "./structural.js";
 import { COUNTRY_CONFIGS } from "./countries/index.js";
 import * as evidenceMod from "./evidence.js";
@@ -273,6 +273,14 @@ export class RuleEngine {
         const allFailed = failedSpans
             .map(f => [f.start, f.end])
             .sort((a, b) => a[0] - b[0] || a[1] - b[1]);
+        // Running maximum of the end offsets, so the fragment test below can ask
+        // "does any span starting at or before me reach past my end?" in O(1)
+        // instead of walking the prefix per candidate.
+        const failedStarts = allFailed.map(f => f[0]);
+        const failedPrefixMax = new Array(allFailed.length + 1).fill(0);
+        for (let i = 0; i < allFailed.length; i++) {
+            failedPrefixMax[i + 1] = Math.max(failedPrefixMax[i], allFailed[i][1]);
+        }
         const mergedZones = new Map();
         for (const [etype, spans] of zonesByType) {
             spans.sort((a, b) => a[0] - b[0] || a[1] - b[1]);
@@ -295,8 +303,11 @@ export class RuleEngine {
             mergedZones.set(etype, { starts, ends });
         }
         const candidates = [];
+        // Whole-document facts a suppressor may need, derived at most once per call
+        // and discarded with it.
+        const scratch = new SuppressionScratch(text);
         for (const { match, hasValidValidator } of validated) {
-            if (shouldSuppress(text, match))
+            if (shouldSuppress(text, match, scratch))
                 continue;
             // A validator-less match inside a failed-validation span of its own type
             // is demoted below every other candidate rather than deleted. Demotion
@@ -324,13 +335,31 @@ export class RuleEngine {
             // default — the generic phone pattern kept claiming the tail of the
             // rejected national number.
             if (match.patternDef.validator === null && !demoted) {
+                // Upper bound of the spans starting at or before match.start.
+                let lo = 0, hi = failedStarts.length;
+                while (lo < hi) {
+                    const mid = (lo + hi) >>> 1;
+                    if (failedStarts[mid] <= match.start)
+                        lo = mid + 1;
+                    else
+                        hi = mid;
+                }
                 let fragment = false;
-                for (const [zs, ze] of allFailed) {
-                    if (zs > match.start)
-                        break;
-                    if (match.end <= ze && (match.start > zs || match.end < ze)) {
-                        fragment = true;
-                        break;
+                const reach = failedPrefixMax[lo];
+                if (reach > match.end) {
+                    // Something starting no later than us ends strictly later: we are a
+                    // proper subset of it.
+                    fragment = true;
+                }
+                else if (reach === match.end) {
+                    // Same end offset — a proper subset only if some span with that end
+                    // starts strictly earlier. Scanning backwards finds the enclosing
+                    // span first, and the equal case is rare.
+                    for (let i = lo - 1; i >= 0; i--) {
+                        if (allFailed[i][1] === match.end && allFailed[i][0] < match.start) {
+                            fragment = true;
+                            break;
+                        }
                     }
                 }
                 if (fragment)
@@ -381,6 +410,7 @@ export class RuleEngine {
                 length: match.end - match.start,
                 countryScore: countryScores.get(match.countryCode) ?? 0,
                 inScope: declared === null || declared.has(match.countryCode) ? 1 : 0,
+                tier: 0,
                 det: {
                     entityType: match.patternDef.entityType,
                     start: match.start,
@@ -395,7 +425,7 @@ export class RuleEngine {
         }
         for (const d of detectStructuralDob(text)) {
             candidates.push({ blindPriority: 1, cue: 0, priority: 1, length: d.end - d.start,
-                countryScore: 0, inScope: 1, det: d });
+                countryScore: 0, inScope: 1, tier: 0, det: d });
         }
         const detections = this.deduplicate(candidates);
         detections.sort((a, b) => a.start - b.start || b.end - a.end);
@@ -429,7 +459,10 @@ export class RuleEngine {
             if (seen === undefined || c.blindPriority > seen)
                 spanTier.set(key, c.blindPriority);
         }
-        const tierOf = (c) => spanTier.get(`${c.det.start}:${c.det.end}`);
+        // Resolved once per candidate rather than inside the comparator, which ran
+        // it O(n log n) times and rebuilt the key string on every comparison.
+        for (const c of candidates)
+            c.tier = spanTier.get(`${c.det.start}:${c.det.end}`);
         // Span length outranks *everything*, including validation.
         //
         // A shorter candidate winning re-cuts a longer one and leaves the
@@ -455,7 +488,7 @@ export class RuleEngine {
         // country-blind up to and including `start`, and (length, start) already
         // identifies a span uniquely.
         const sorted = [...candidates].sort((a, b) => b.length - a.length || // longest first     } country-blind
-            tierOf(b) - tierOf(a) || // best tier here    } — decides
+            b.tier - a.tier || // best tier here    } — decides
             a.det.start - b.det.start || // leftmost          } WHICH span
             b.cue - a.cue || // local cue         } country-aware
             b.priority - a.priority || // country priority  } — decides the

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

@@ -6,4 +6,17 @@ export interface RawMatch {
     patternDef: PatternDef;
     countryCode: string;
 }
-export declare function shouldSuppress(text: string, match: RawMatch): boolean;
+/**
+ * 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 declare class SuppressionScratch {
+    private readonly text;
+    private words;
+    constructor(text: string);
+    documentWords(): Set<string>;
+}
+export declare function shouldSuppress(text: string, match: RawMatch, scratch?: SuppressionScratch): boolean;

package/dist/esm/rules/suppressors.js

@@ -268,8 +268,12 @@ function suppressUnits(text, match) {
  * 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, EntityType.POSTAL_CODE]);
+    const applicable = REFERENCE_TYPES;
     if (!applicable.has(match.patternDef.entityType))
         return false;
     const [before] = getContext(text, match.start, match.end);
@@ -279,7 +283,7 @@ function suppressReference(text, match) {
     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);
@@ -502,8 +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() === "")
@@ -512,6 +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)
@@ -548,28 +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?
@@ -607,25 +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*$/i;
 // An international dialling prefix earlier on the same line, with nothing but
 // number punctuation in between: these digits belong to the phone detector.
 const DIALLING_PREFIX_BEFORE = /\+\d{1,3}[\d\s\-().]*$/;
@@ -737,13 +799,13 @@ const TYPE_SUPPRESSORS = {
     [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

@@ -1,6 +1,6 @@
 {
   "name": "euredact",
-  "version": "0.3.6",
+  "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",
@@ -65,8 +65,9 @@
     "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"