diff --git a/modules/story-summary/vector/utils/tokenizer.js b/modules/story-summary/vector/utils/tokenizer.js
index 4122b12..6c150d0 100644
--- a/modules/story-summary/vector/utils/tokenizer.js
+++ b/modules/story-summary/vector/utils/tokenizer.js
@@ -199,8 +199,8 @@ function segmentByScript(text) {
 // 实体保护（最长匹配占位符替换）
 // ═══════════════════════════════════════════════════════════════════════════
 
-// 使用 Unicode Private Use Area (PUA) 字符作为边界，避免控制字符在分词器中产生不可控行为
-const PLACEHOLDER_PREFIX = '\uE000ENT_';
+// 使用纯 PUA 字符序列作为占位符，避免拉丁字母泄漏到分词结果
+const PLACEHOLDER_PREFIX = '\uE000\uE010';
 const PLACEHOLDER_SUFFIX = '\uE001';
 
 /**
@@ -264,21 +264,25 @@ function maskEntities(text) {
 function unmaskTokens(tokens, entities) {
     if (!entities.size) return tokens;
 
-    return tokens.map(token => {
-        // token 本身就是一个占位符
+    return tokens.flatMap(token => {
+        // token 本身就是一个完整占位符
         if (entities.has(token)) {
-            return entities.get(token);
+            return [entities.get(token)];
         }
 
-        // token 中包含占位符（结巴可能把占位符和其他字符连在一起）
-        let result = token;
-        for (const [placeholder, original] of entities) {
-            if (result.includes(placeholder)) {
-                result = result.replace(placeholder, original);
+        // token 中包含 PUA 字符 → 检查是否包含完整占位符
+        if (/[\uE000-\uE0FF]/.test(token)) {
+            for (const [placeholder, original] of entities) {
+                if (token.includes(placeholder)) {
+                    return [original];
+                }
             }
+            // 纯 PUA 碎片，丢弃
+            return [];
         }
 
-        return result;
+        // 普通 token，原样保留
+        return [token];
     });
 }