fix(security): keep untrusted text analysis linear

.changeset/calm-math-scanner.md

@@ -0,0 +1,7 @@
+---
+'@bidilens/core': patch
+---
+
+Replace the raw-text math delimiter regular expression with a forward scanner
+so repeated unmatched `\(` input stays linear instead of triggering
+polynomial backtracking.

CHANGELOG.md

@@ -5,6 +5,16 @@ is published under the public `@bidilens` npm scope.
 
 ## Unreleased
 
+### Security and performance
+
+- Replaced the raw-text math-delimiter regular expression with a single-pass
+  scanner, preventing quadratic work on repeated unmatched `\(` input while
+  preserving `$...$`, `$$...$$`, `$$$$`, and `\(...\)` recognition.
+- Expressed UTF-16-to-code-point range construction with bounded typed-array
+  fills so input-derived offsets cannot be interpreted as object properties.
+- Restricted downloaded Unicode data writes to the two repository-owned,
+  version-pinned destination paths after exact SHA-256 verification.
+
 ### Direction correctness
 
 - Kept ordinary hyphenated English compounds and block-level ALL-CAPS prose as

action/dist/index.cjs

@@ -6148,6 +6148,27 @@ function addMatches(text, ranges, expression, kind, group = 0) {
     addRange(ranges, text, start, start + value.length, kind);
   }
 }
+function addMathRanges(text, ranges) {
+  let i = 0;
+  let scanned = -1;
+  while (i < text.length) {
+    const p = text[i] === "\\" && text[i + 1] === "(";
+    const d = text[i] === "$" ? text[i + 1] === "$" ? "$$" : "$" : p && i >= scanned ? "\\)" : "";
+    if (!d) {
+      i++;
+      continue;
+    }
+    let e = i + (p ? 2 : d.length);
+    while (e < text.length && text[e] !== "\r" && text[e] !== "\n" && !text.startsWith(d, e)) e++;
+    if (text.startsWith(d, e) && (d !== "$" || e > i + 1)) {
+      addRange(ranges, text, i, e + d.length, "math");
+      i = e + d.length;
+    } else {
+      if (p) scanned = e;
+      i++;
+    }
+  }
+}
 function trimTechnicalPunctuation(value) {
   let end = value.length;
   while (end > 0 && /[.,;:!?،؛؟。।۔]/u.test(value[end - 1])) end -= 1;
@@ -6333,7 +6354,7 @@ function findTechnicalTokenRanges(text, technicalIdentifiers = []) {
   const ranges = [];
   addCodeRanges(text, ranges);
   addMatches(text, ranges, /<\/?[A-Za-z][^<>\r\n]*>/gu, "html");
-  addMatches(text, ranges, /(?:\$\$[^\r\n]*?\$\$|\$[^$\r\n]+\$|\\\([^\r\n]*?\\\))/gu, "math");
+  addMathRanges(text, ranges);
   const urls = /\b(?:https?|ftp):\/\/[^\s<>{}"']+/giu;
   let urlMatch;
   while ((urlMatch = urls.exec(text)) !== null) {
@@ -6828,12 +6849,11 @@ function attachSourceRanges(text, isolations) {
   let utf16Offset = 0;
   let codePointOffset = 0;
   for (const character of text) {
-    codePointAtUtf16[utf16Offset] = codePointOffset;
-    if (character.length === 2) codePointAtUtf16[utf16Offset + 1] = codePointOffset;
+    codePointAtUtf16.fill(codePointOffset, utf16Offset, utf16Offset + character.length);
     utf16Offset += character.length;
     codePointOffset += 1;
-    codePointAtUtf16[utf16Offset] = codePointOffset;
   }
+  codePointAtUtf16.fill(codePointOffset, utf16Offset);
   return isolations.map((isolation) => ({
     ...isolation,
     sourceRange: {

docs/PERFORMANCE.md

@@ -86,6 +86,8 @@ API signal while a rich document is between checkpoints.
   at most 14 live Markdown parses;
 - a unit alarm permits 8,000 single-character pushes and dense isolation
   planning to finish within three seconds on the CI machine;
+- an adversarial unit alarm scans 128,000 UTF-16 units of repeated unmatched
+  `\(` delimiters within the batch budget, guarding the linear math scanner;
 - release checks enforce aggregate emitted-JavaScript budgets, including
   code-split chunks.
 

docs/SECURITY.md

@@ -42,6 +42,11 @@ pops/openers, formatting that crosses an isolate boundary, and U+200B.
 - markup adapters prefer `<bdi>`/`dir` to invisible control insertion;
 - terminal isolate insertion is opt-in and carries a warning;
 - recursive CLI scanning skips symbolic links.
+- raw-text math delimiter discovery is a single forward scan; repeated
+  unmatched delimiters cannot trigger polynomial regular-expression work.
+- the optional Unicode download command accepts only two source/destination
+  pairs pinned in repository code and verifies each exact SHA-256 before a
+  write; untrusted text never reaches that build-time path.
 
 ## False-positive policy
 

packages/core/src/core.test.ts

@@ -136,6 +136,16 @@ describe('direction detection', () => {
       .map((range) => range.text)).toContain('https://example.com');
   });
 
+  it('recognizes inline math delimiters without crossing line boundaries', () => {
+    const source = 'Use $x + 1$, $$E = mc^2$$, \\(a + b\\), $$$$, then \\($z$.';
+    expect(findTechnicalTokenRanges(source)
+      .filter((range) => range.kind === 'math')
+      .map((range) => range.text))
+      .toEqual(['$x + 1$', '$$E = mc^2$$', '\\(a + b\\)', '$$$$', '$z$']);
+    expect(findTechnicalTokenRanges('\\(not closed\n\\)')
+      .filter((range) => range.kind === 'math')).toEqual([]);
+  });
+
   it('treats closed multiline Markdown fences inside prose as technical ranges in raw text', () => {
     const closed = 'شغل الكود:\n```bash\necho hello world\n```\nالناتج صحيح.';
     const closedRanges = findTechnicalTokenRanges(closed);
@@ -906,6 +916,18 @@ describe('streaming', () => {
     findTechnicalTokenRanges(backticks);
     expect(performance.now() - backtickBatchStart).toBeLessThan(1_500);
 
+    // An unmatched `\(` used to restart the lazy regular-expression search at
+    // every later opener, making adversarial chat input quadratic.
+    const unmatchedParenthesizedMath = '\\('.repeat(64_000);
+    const parenthesizedMathStart = performance.now();
+    findTechnicalTokenRanges(unmatchedParenthesizedMath);
+    expect(performance.now() - parenthesizedMathStart).toBeLessThan(1_500);
+
+    const unmatchedMathLines = `${'\\(\n'.repeat(32_000)}\\)`;
+    const mathLinesStart = performance.now();
+    findTechnicalTokenRanges(unmatchedMathLines);
+    expect(performance.now() - mathLinesStart).toBeLessThan(1_500);
+
     const manyFences = `متن\n${'```\nx\n```\n\n'.repeat(1_000)}ادامه`;
     const manyFencesStart = performance.now();
     findTechnicalTokenRanges(manyFences);

packages/core/src/detect.ts

@@ -83,6 +83,37 @@ function addMatches(
   }
 }
 
+/**
+ * Adds `$...$`, `$$...$$`, and `\(...\)` math ranges in one forward pass.
+ * A combined lazy regular expression restarts its search after every unmatched
+ * `\(` opener, which makes adversarial input quadratic.
+ */
+function addMathRanges(text: string, ranges: TechnicalTokenRange[]): void {
+  let i = 0;
+  let scanned = -1;
+  while (i < text.length) {
+    const p = text[i] === '\\' && text[i + 1] === '(';
+    const d = text[i] === '$'
+      ? (text[i + 1] === '$' ? '$$' : '$')
+      : (p && i >= scanned ? '\\)' : '');
+    if (!d) { i++; continue; }
+
+    let e = i + (p ? 2 : d.length);
+    while (e < text.length
+      && text[e] !== '\r'
+      && text[e] !== '\n'
+      && !text.startsWith(d, e)) e++;
+
+    if (text.startsWith(d, e) && (d !== '$' || e > i + 1)) {
+      addRange(ranges, text, i, e + d.length, 'math');
+      i = e + d.length;
+    } else {
+      if (p) scanned = e;
+      i++;
+    }
+  }
+}
+
 function trimTechnicalPunctuation(value: string): string {
   let end = value.length;
   while (end > 0 && /[.,;:!?،؛؟。।۔]/u.test(value[end - 1]!)) end -= 1;
@@ -338,7 +369,7 @@ export function findTechnicalTokenRanges(
   const ranges: TechnicalTokenRange[] = [];
   addCodeRanges(text, ranges);
   addMatches(text, ranges, /<\/?[A-Za-z][^<>\r\n]*>/gu, 'html');
-  addMatches(text, ranges, /(?:\$\$[^\r\n]*?\$\$|\$[^$\r\n]+\$|\\\([^\r\n]*?\\\))/gu, 'math');
+  addMathRanges(text, ranges);
 
   const urls = /\b(?:https?|ftp):\/\/[^\s<>{}"']+/giu;
   let urlMatch: RegExpExecArray | null;

packages/core/src/segments.ts

@@ -12,12 +12,11 @@ function attachSourceRanges(text: string, isolations: Omit<InlineIsolation, 'sou
   let utf16Offset = 0;
   let codePointOffset = 0;
   for (const character of text) {
-    codePointAtUtf16[utf16Offset] = codePointOffset;
-    if (character.length === 2) codePointAtUtf16[utf16Offset + 1] = codePointOffset;
+    codePointAtUtf16.fill(codePointOffset, utf16Offset, utf16Offset + character.length);
     utf16Offset += character.length;
     codePointOffset += 1;
-    codePointAtUtf16[utf16Offset] = codePointOffset;
   }
+  codePointAtUtf16.fill(codePointOffset, utf16Offset);
   return isolations.map((isolation) => ({
     ...isolation,
     sourceRange: {

scripts/generate-bidi-data.ts

@@ -25,6 +25,23 @@ interface UnicodeSource {
   sha256: string;
 }
 
+const UNICODE_SOURCES = [
+  {
+    label: 'DerivedBidiClass',
+    url: BIDI_URL,
+    path: BIDI_PATH,
+    sha256: BIDI_SHA256
+  },
+  {
+    label: 'DerivedGeneralCategory',
+    url: GENERAL_CATEGORY_URL,
+    path: GENERAL_CATEGORY_PATH,
+    sha256: GENERAL_CATEGORY_SHA256
+  }
+] as const satisfies readonly UnicodeSource[];
+
+const PINNED_UNICODE_PATHS = new Set(UNICODE_SOURCES.map((source) => source.path));
+
 const CLASS = {
   L: 0,
   R: 1,
@@ -267,6 +284,9 @@ async function sourceBytes(source: UnicodeSource, download: boolean): Promise<Ui
     throw new Error(`${source.label} checksum mismatch: expected ${source.sha256}, received ${actualSha256}`);
   }
   if (download) {
+    if (!PINNED_UNICODE_PATHS.has(source.path)) {
+      throw new Error(`Refusing to write outside the pinned Unicode source paths: ${source.path}`);
+    }
     await mkdir(dirname(source.path), { recursive: true });
     await writeFile(source.path, bytes);
   }
@@ -276,15 +296,7 @@ async function sourceBytes(source: UnicodeSource, download: boolean): Promise<Ui
 async function main(): Promise<void> {
   const download = process.argv.includes('--download');
   const check = process.argv.includes('--check');
-  const bidiSource: UnicodeSource = {
-    label: 'DerivedBidiClass', url: BIDI_URL, path: BIDI_PATH, sha256: BIDI_SHA256
-  };
-  const generalCategorySource: UnicodeSource = {
-    label: 'DerivedGeneralCategory',
-    url: GENERAL_CATEGORY_URL,
-    path: GENERAL_CATEGORY_PATH,
-    sha256: GENERAL_CATEGORY_SHA256
-  };
+  const [bidiSource, generalCategorySource] = UNICODE_SOURCES;
   const [bidiBytes, generalCategoryBytes] = await Promise.all([
     sourceBytes(bidiSource, download),
     sourceBytes(generalCategorySource, download)