diff --git a/src/engine/pipeline.rs b/src/engine/pipeline.rs
index f39f537..0bcfa3b 100644
--- a/src/engine/pipeline.rs
+++ b/src/engine/pipeline.rs
@@ -285,17 +285,23 @@ pub async fn phase_clean_persons(db: &Database) -> (DirtySet, usize) {
 /// 这样可以防止停用词（如 "photo"、"model" 等通用词）被错误识别为人物。
 pub async fn phase_clean_stop_word_persons(db: &Database) -> (DirtySet, usize) {
     let persons = db.get_all_persons_with_counts().await;
-    let dict = tag_dict::get_tag_dict();
+
+    // 预先复制停用词集合到本地 HashSet，避免持 RwLockReadGuard 穿越 await
+    let stop_words: HashSet<String> = {
+        let dict = tag_dict::get_tag_dict();
+        dict.get_stop_words().clone()
+    };
+
     let mut dirty = DirtySet::new();
     let mut deleted = 0usize;
 
     for (pid, name, _albums, _images) in &persons {
         // 检查人名本身是否在停用词中
-        let in_stop_words = dict.is_stop_word(name)
-            || dict.is_stop_word(&name.to_lowercase());
-        // 检查中文核心是否在停用词中（例如 "杨晨晨" 不在，但 "photo" 可能在）
+        let in_stop_words = stop_words.contains(name)
+            || stop_words.contains(&name.to_lowercase());
+        // 检查中文核心是否在停用词中
         let cn_in_stop_words = title_parser::extract_cn(name)
-            .map(|cn| dict.is_stop_word(&cn))
+            .map(|cn| stop_words.contains(&cn))
             .unwrap_or(false);
 
         if in_stop_words || cn_in_stop_words {