fix(reference): preserve raw symlink target identity

Read link targets as bytes before strict UTF-8 admission. Retain malformed targets as unreadable with source hex; preserve literal replacement characters and valid BOM-prefixed targets.

Fixes #6765

src/reference/ReferenceSourceReaderEntries.ts

@@ -377,9 +377,12 @@ const readSymlinkTarget = async (
   absolute: string,
   path: string,
 ): Promise<ReferenceSourceEntry> => {
-  let rawTarget: string;
+  let rawTarget: Buffer;
   try {
-    rawTarget = await readlink(absolute);
+    // Keep the symlink's source bytes until UTF-8 validity is known. The
+    // default string encoding can alias an invalid target with a real U+FFFD
+    // sibling during the subsequent path lookup.
+    rawTarget = await readlink(absolute, { encoding: "buffer" });
   } catch (cause: unknown) {
     // POSIX readlink reports EINVAL when the observed link became a non-link.
     // Keep this classification at the syscall, rather than hiding EINVAL from
@@ -388,7 +391,21 @@ const readSymlinkTarget = async (
       return changedSymlink(path);
     throw cause;
   }
-  const lexicalTarget = resolve(dirname(absolute), rawTarget);
+  let target: string;
+  try {
+    target = new TextDecoder("utf-8", {
+      fatal: true,
+      ignoreBOM: true,
+    }).decode(rawTarget);
+  } catch {
+    return entryFailure(
+      path,
+      "symlink",
+      "unsupported",
+      `Symlink target is not valid UTF-8 (raw target bytes: ${rawTarget.toString("hex")})`,
+    );
+  }
+  const lexicalTarget = resolve(dirname(absolute), target);
   if (!isPathWithinRoot(root, lexicalTarget))
     return {
       status: "read",