fix(ai): reject non-ASCII base64 that Node decodes by its low byte

Node reads code units above 0xFF by their low byte, so U+012B decodes
as '+' and keeps the decoded length the input predicts. Require ASCII
(~1-2 ms for 43M chars) before the length check.

src/backend/util/mediaSniff.test.ts

@@ -35,6 +35,7 @@ describe('decodeStrictBase64', () => {
         ['the URL-safe alphabet', 'a-b_'],
         ['whitespace', 'YW Jj'],
         ['a character outside the alphabet', 'YW*j'],
+        ['a non-ASCII character whose low byte is in the alphabet', 'YWJī'],
         ['padding mid-string', 'YQ==YWJj'],
         ['three pad characters', 'Y==='],
     ])('rejects %s', (_label, value) => {

src/backend/util/mediaSniff.ts

@@ -59,6 +59,9 @@ const MP4_BRANDS = new Set([
  */
 export function decodeStrictBase64(value: string): Buffer | null {
     if (value.length === 0 || value.length % 4 !== 0) return null;
+    // Node decodes a code unit above 0xFF by its low byte (`ī` as `+`), so
+    // only ASCII is decoded.
+    if (Buffer.byteLength(value, 'utf8') !== value.length) return null;
     // Node also decodes the URL-safe alphabet.
     if (value.includes('-') || value.includes('_')) return null;
     const padding = value.endsWith('==') ? 2 : value.endsWith('=') ? 1 : 0;