Rewrite durable task lifecycle, gallery indexing and subscription publication

src/main/java/com/lion/lionwebsite/Configuration/CustomBean.java

@@ -27,6 +27,9 @@ import java.util.Random;
         Chat.class, MessageEntity.class,
         AbstractMethodError.class, DeleteGalleryMessage.class, DownloadPostMessage.class, DownloadStatusMessage.class,
         IdentityMessage.class, MaintainMessage.class, ResponseMessage.class, AvailableCheckMessage.class,
+        // 协议 v2 的消息与任务操作实体：原生镜像下 Jackson/MyBatis 都要反射构造它们。
+        ProtocolHelloMessage.class, TaskCommandMessage.class, TaskReportMessage.class, TaskReportAckMessage.class,
+        TaskOperation.class,
         SubscriptionSnapshotMessage.class, SubscriptionSnapshotPayload.class, SubscriptionAccountSnapshot.class,
         SubscriptionBindingSnapshot.class, LinkPreviewOptions.class})
 public class CustomBean {

src/main/java/com/lion/lionwebsite/Configuration/GalleryIndexConfiguration.java

@@ -0,0 +1,83 @@
+package com.lion.lionwebsite.Configuration;
+
+import com.lion.lionwebsite.Dao.cache.ImageCacheMapper;
+import com.lion.lionwebsite.Domain.ImageKeyCache;
+import com.lion.lionwebsite.Service.GalleryIndexService;
+import com.lion.lionwebsite.Util.GalleryUtil;
+import org.springframework.beans.factory.annotation.Qualifier;
+import org.springframework.context.annotation.Bean;
+import org.springframework.context.annotation.Configuration;
+import org.springframework.jdbc.datasource.DataSourceTransactionManager;
+import org.springframework.transaction.PlatformTransactionManager;
+import org.springframework.transaction.support.TransactionTemplate;
+
+import javax.sql.DataSource;
+import java.util.List;
+
+/**
+ * 图片索引服务的事务装配。
+ *
+ * <p>单独成文件是为了不触碰既有的 SqlConfiguration：那里的两个 SqlSessionFactory
+ * 保持原样，这里只额外声明缓存库的事务管理器与模板。
+ *
+ * <p>事务管理器必须绑定 datasource-cache —— 索引只写缓存库，
+ * 若误用主库事务管理器，整组替换就失去原子性（删除与写入落在不同连接上）。
+ */
+@Configuration
+public class GalleryIndexConfiguration {
+
+    @Bean
+    public PlatformTransactionManager cacheTransactionManager(
+            @Qualifier("datasource-cache") DataSource cacheDataSource) {
+        return new DataSourceTransactionManager(cacheDataSource);
+    }
+
+    /** 只用于图片索引的短事务发布，超时留给上层请求超时控制，这里显式设为 30 秒。 */
+    @Bean
+    public TransactionTemplate cacheTransactionTemplate(
+            @Qualifier("cacheTransactionManager") PlatformTransactionManager cacheTransactionManager) {
+        TransactionTemplate template = new TransactionTemplate(cacheTransactionManager);
+        template.setTimeout(30);
+        return template;
+    }
+
+    /**
+     * 只读快照事务，用于一次性读取头部与页码。
+     *
+     * <p>读事务不写库，因此只读连接可以与写入并发；它的作用是让「读头部 + 读页码」
+     * 落在同一个快照上，避免恰好读到整组替换的中间状态。
+     */
+    @Bean
+    public TransactionTemplate cacheReadTransactionTemplate(
+            @Qualifier("cacheTransactionManager") PlatformTransactionManager cacheTransactionManager) {
+        TransactionTemplate template = new TransactionTemplate(cacheTransactionManager);
+        template.setTimeout(10);
+        return template;
+    }
+
+    /**
+     * 这里显式不设置 {@code readOnly}：sqlite-jdbc 不允许在已建立的连接上改只读标志
+     * （实测抛 SQLException: Cannot change read-only flag after establishing a connection），
+     * 而事务模板拿到的是连接池里的既有连接，设置只读会让每次读取都直接失败。
+     * 快照一致性由「同一个事务」保证，不需要额外的只读标志。
+     */
+
+    @Bean
+    public GalleryIndexService galleryIndexService(
+            ImageCacheMapper imageCacheMapper,
+            @Qualifier("cacheTransactionTemplate") TransactionTemplate cacheTransactionTemplate,
+            @Qualifier("cacheReadTransactionTemplate") TransactionTemplate cacheReadTransactionTemplate) {
+        return new GalleryIndexService(imageCacheMapper, cacheTransactionTemplate,
+                cacheReadTransactionTemplate,
+                GalleryIndexConfiguration::fetchPageKeys);
+    }
+
+    /**
+     * 生产环境的上游实现：沿用 GalleryUtil.parseImageKeys（HTTP + JS 解析）。
+     *
+     * <p>放在这里而不是服务内部，是为了让服务在测试中可以脱离静态工具类。
+     */
+    static List<ImageKeyCache> fetchPageKeys(String url) throws Exception {
+        return GalleryUtil.parseImageKeys(url);
+    }
+}

src/main/java/com/lion/lionwebsite/Configuration/SqlConfiguration.java

@@ -8,6 +8,11 @@ import org.springframework.boot.context.properties.ConfigurationProperties;
 import org.springframework.boot.jdbc.DataSourceBuilder;
 import org.springframework.context.annotation.Bean;
 import org.springframework.context.annotation.Configuration;
+import org.springframework.context.annotation.Primary;
+import org.springframework.jdbc.datasource.DataSourceTransactionManager;
+import org.springframework.transaction.support.TransactionTemplate;
+
+import com.zaxxer.hikari.HikariDataSource;
 
 import javax.sql.DataSource;
 
@@ -16,7 +21,9 @@ public class SqlConfiguration {
     @Bean(name = "datasource-main")
     @ConfigurationProperties(prefix = "spring.datasource-main")
     public DataSource dataSourceMain() {
-        return DataSourceBuilder.create().build();
+        DataSource dataSource = DataSourceBuilder.create().build();
+        applyBusyTimeout(dataSource);
+        return dataSource;
     }
 
     @Bean(name = "datasource-cache")
@@ -25,6 +32,18 @@ public class SqlConfiguration {
         return DataSourceBuilder.create().build();
     }
 
+    /**
+     * 给主库连接打开 busy 等待。
+     *
+     * <p>主库是单文件 SQLite，任务生命周期会并发开启短事务（创建、状态回执、删除）。
+     * 默认没有 busy 超时，第二个写事务会立刻拿到 SQLITE_BUSY 并被当作 500；
+     * 打开等待后，写冲突会按序排队重试而不是失败。
+     */
+    private void applyBusyTimeout(DataSource dataSource) {
+        if (dataSource instanceof HikariDataSource hikari)
+            hikari.setConnectionInitSql("PRAGMA busy_timeout=5000");
+    }
+
     @Bean(name = "sqlSessionFactory-main")
     public SqlSessionFactory sqlSessionFactoryMain(@Qualifier("datasource-main") DataSource dataSource) throws Exception {
         SqlSessionFactoryBean sessionFactoryBean = new SqlSessionFactoryBean();
@@ -48,4 +67,30 @@ public class SqlConfiguration {
     public SqlSessionTemplate sqlSessionTemplateCache(@Qualifier("sqlSessionFactory-cache") SqlSessionFactory sqlSessionFactory) {
         return new SqlSessionTemplate(sqlSessionFactory);
     }
+
+    /**
+     * 主库事务管理器。
+     *
+     * <p>必须显式绑定到 datasource-main：应用里有两个 DataSource，Spring Boot 的自动配置
+     * 会退避，若没有这个 bean 就没有任何事务管理器，@Transactional 会静默失效。
+     * @Primary 同时让「不指定名字就注入 TransactionManager」的旧代码继续可用。
+     */
+    @Bean(name = "mainTransactionManager")
+    @Primary
+    public DataSourceTransactionManager mainTransactionManager(@Qualifier("datasource-main") DataSource dataSource) {
+        return new DataSourceTransactionManager(dataSource);
+    }
+
+    /**
+     * 主库编程式事务模板。
+     *
+     * <p>任务生命周期需要把「短事务」放在网络调用之外，并且要求「回调里捕获异常后不再抛出」
+     * 不能错误提交，因此调用方必须用 {@code execute} 的返回状态或再次抛出决定回滚，
+     * 而不是在回调里吞掉异常（见 GalleryTaskLifecycleService）。
+     */
+    @Bean(name = "mainTransactionTemplate")
+    @Primary
+    public TransactionTemplate mainTransactionTemplate(@Qualifier("mainTransactionManager") DataSourceTransactionManager transactionManager) {
+        return new TransactionTemplate(transactionManager);
+    }
 }

src/main/java/com/lion/lionwebsite/Configuration/TaskLifecycleWiring.java

@@ -0,0 +1,57 @@
+package com.lion.lionwebsite.Configuration;
+
+import com.lion.lionwebsite.Domain.TaskOperation;
+import com.lion.lionwebsite.Service.GalleryTaskLifecycleService;
+import com.lion.lionwebsite.Service.RemoteService;
+import com.lion.lionwebsite.Service.TaskReportHandler;
+import lombok.RequiredArgsConstructor;
+import lombok.extern.slf4j.Slf4j;
+import org.springframework.boot.context.event.ApplicationReadyEvent;
+import org.springframework.context.event.EventListener;
+import org.springframework.stereotype.Component;
+
+/**
+ * 装配协议 v2 的双向依赖，并做启动恢复。
+ *
+ * <p>RemoteService 消费上报时需要生命周期服务，而生命周期服务下发命令时需要 RemoteService。
+ * 直接互相注入构造参数会形成环，所以这里在两端都构造完成之后用 setter 接上；
+ * 同时也把「启动时把中断的 SUBMITTING 判为 UNKNOWN、释放 RESERVED 占位」这一步接到应用就绪事件上。
+ *
+ * <p>顺序是不可交换的：
+ * <ol>
+ *   <li>先绑定上报处理器、下发判定解析器与 operationId 解析器；</li>
+ *   <li>再执行启动恢复（SUBMITTING → UNKNOWN、释放 RESERVED）；</li>
+ *   <li>最后才允许建立连接与自动重投。连接由本类显式触发，不再由 RemoteService 的
+ *       PostConstruct 在装配前发起，否则会把状态不明的任务盲目重投给节点。</li>
+ * </ol>
+ * 恢复失败时仍然允许连接（保持节点在线/心跳），但明确禁止自动派发。
+ */
+@Component
+@Slf4j
+@RequiredArgsConstructor
+public class TaskLifecycleWiring {
+
+    final RemoteService remoteService;
+
+    final GalleryTaskLifecycleService lifecycleService;
+
+    @EventListener(ApplicationReadyEvent.class)
+    public void wire() {
+        TaskReportHandler handler = lifecycleService;
+        remoteService.setTaskReportHandler(handler);
+        remoteService.setDownloadOperationIdResolver(lifecycleService::currentOperationId);
+        remoteService.setDispatchResolver(lifecycleService::resolveDispatch);
+        boolean recoverySafe = false;
+        try {
+            int recovered = lifecycleService.recoverInterruptedSubmissions();
+            recoverySafe = true;
+            if (recovered > 0)
+                log.warn("启动恢复：{} 个中断的创建/预占任务已处理（提交中的标为提交未知）", recovered);
+        } catch (RuntimeException e) {
+            // 恢复失败说明库中任务真实状态未知：可以保持连接，但绝不能自动重投/派发。
+            log.error("启动恢复失败，本次启动将拒绝自动重投未完成任务", e);
+        }
+        // 绑定与恢复完成之后才连接：连接成功后的自动重投受 lifecycleReady/recoverySafe 双重约束。
+        remoteService.startAfterWiring(recoverySafe);
+    }
+}

src/main/java/com/lion/lionwebsite/Controller/PublicController.java

@@ -62,16 +62,14 @@ public class PublicController {
     public String validate(String AuthCode){
         Response response = Response.generateResponse();
         User user = publicService.getUserId(AuthCode);
-        // 用 ObjectMapper 组装内层 JSON：手工 String.format 拼用户名时，
-        // 名字里带引号或反斜杠会直接产出非法 JSON，前端 JSON.parse 随即失败。
-        // 这里仍以「JSON 文本」形式放进 data（历史契约，前端按字符串再解析一次）。
+        // 身份字段直接作为结构化 data 返回；共享前端解码器兼容旧版本字符串。
         ObjectNode identity = CustomUtil.objectMapper.createObjectNode();
         identity.put("userId", user.getId());
         identity.put("username", user.getUsername());
         identity.put("isAvailable", !remoteService.isDead());
         // 管理员标记随登录一起下发，前端据此决定是否显示下载人信息与筛选。
         identity.put("isAdmin", user.getId() == UserService.ADMIN_USER_ID);
-        response.success(identity.toString());
+        response.success(identity);
         return response.toJSONString();
     }
 

src/main/java/com/lion/lionwebsite/Dao/cache/ImageCacheMapper.java

@@ -8,6 +8,8 @@ import org.apache.ibatis.annotations.Param;
 import org.apache.ibatis.annotations.Delete;
 import org.apache.ibatis.annotations.Select;
 
+import java.util.List;
+
 @Mapper
 public interface ImageCacheMapper {
     @Insert("insert into ImageKeyCache values (#{gid}, #{imgkey}, #{page})")
@@ -31,4 +33,32 @@ public interface ImageCacheMapper {
 
     @Delete("delete from gidToKey where gid=#{gid}")
     void deleteGidToKey(String gid);
+
+    /**
+     * 某个 gid 已落库的页码列表（升序）。
+     *
+     * <p>只按行数判断完整性会把「页码 1 和 3、共 2 行」当成完整索引，
+     * 因此完整性判定改为按页码集合比对——先取回页码，再校验它是否恰好是 1..pages。
+     */
+    @Select("select page from ImageKeyCache where gid=#{gid} order by page")
+    List<Integer> selectPageNumbersByGid(String gid);
+
+    /** 该 gid 下 imgkey 为空的脏行数；任何一行脏行都说明索引不可用。 */
+    @Select("select count(*) from ImageKeyCache where gid=#{gid} and (imgkey is null or trim(imgkey)='')")
+    int countInvalidImageKeysByGid(String gid);
+
+    /**
+     * 批量写入页 key（单条多值 INSERT）。
+     *
+     * <p>整本画廊的索引可能上千页，逐行 INSERT 会拉长写事务并放大锁竞争；
+     * 调用方按固定大小分片后，一个写事务里只需少量语句即可写完。
+     */
+    // MyBatis 只在注解值「以 <script> 开头」时才按 XML 动态 SQL 解析，因此这里不能有前导空白。
+    @Insert("<script>"
+            + "insert into ImageKeyCache (gid, imgkey, page) values "
+            + "<foreach collection='pages' item='p' separator=','>"
+            + "(#{p.gid}, #{p.imgkey}, #{p.page})"
+            + "</foreach>"
+            + "</script>")
+    int insertImageKeyCacheBatch(@Param("pages") List<ImageKeyCache> pages);
 }

src/main/java/com/lion/lionwebsite/Dao/normal/CollectMapper.java

@@ -20,4 +20,8 @@ public interface CollectMapper {
 
     @Select("select collector from collect where gid=#{gid}")
     ArrayList<Integer> selectCollectorByGid(int gid);
+
+    /** 任务删除完成后清理该 gid 的全部收藏记录（与删任务行同一事务）。 */
+    @Delete("delete from collect where gid=#{gid}")
+    void deleteByGid(@Param("gid") int gid);
 }

src/main/java/com/lion/lionwebsite/Dao/normal/GalleryMapper.java

@@ -36,9 +36,10 @@ public interface GalleryMapper {
 
 
     @Update("""
-            update gallery set name=#{name}, link=#{link}, language=#{language},
+            update gallery set name=#{name}, link=#{link}, language=#{language}, downloader=#{downloader},
                                     pages=#{pages}, status=#{status}, fileSize=#{fileSize}, createTime=#{createTime},
-                                    proceeding=#{proceeding}, resolution=#{resolution}, displayFileSize=#{displayFileSize}, is_download=#{is_download}
+                                    proceeding=#{proceeding}, resolution=#{resolution}, displayFileSize=#{displayFileSize},
+                                    thumb_link=#{thumb_link}, is_download=#{is_download}
                     where gid=#{gid}""")
     void updateGallery(Gallery gallery);
 
@@ -47,4 +48,8 @@ public interface GalleryMapper {
 
     @Delete("delete from gallery where gid=#{gid}")
     void deleteGalleryByGid(Integer gid);
+
+    /** 只改状态列，用于「待删除」这类不牵动其他字段的意图标记。 */
+    @Update("update gallery set status=#{status} where gid=#{gid}")
+    int updateGalleryStatus(@Param("gid") int gid, @Param("status") String status);
 }

src/main/java/com/lion/lionwebsite/Dao/normal/SubMapper.java

@@ -43,49 +43,57 @@ public interface SubMapper {
     @Delete("delete from subscription_account where id=#{id}")
     void deleteSubscriptionAccount(Integer id);
 
-    @Insert("insert into sub_bind (key, user, subscription_account_id) values (#{key}, #{user}, #{subscriptionAccountId})")
+    @Insert("insert into sub_bind (key, user, subscription_account_id, user_id) values (#{key}, #{user}, #{subscriptionAccountId}, (select id from User where username=#{user}))")
     void insertSubBind(SubBind subBind);
 
-    @Select("select sb.key, sb.user, sb.subscription_account_id as subscriptionAccountId, sa.name as subscriptionAccountName, sa.enabled as subscriptionAccountEnabled, sa.filter_high_multiplier as filterHighMultiplier from sub_bind sb left join subscription_account sa on sa.id=sb.subscription_account_id order by sb.user")
+    @Select("select sb.key, coalesce(u.username, sb.user) as user, sb.user_id as userId, sb.subscription_account_id as subscriptionAccountId, sa.name as subscriptionAccountName, sa.enabled as subscriptionAccountEnabled, sa.filter_high_multiplier as filterHighMultiplier from sub_bind sb left join User u on u.id=sb.user_id left join subscription_account sa on sa.id=sb.subscription_account_id order by sb.user")
     ArrayList<SubBind> selectAllSubBind();
 
-    @Select("select sb.key, sb.user, sb.subscription_account_id as subscriptionAccountId, sa.name as subscriptionAccountName, sa.enabled as subscriptionAccountEnabled, sa.filter_high_multiplier as filterHighMultiplier from sub_bind sb left join subscription_account sa on sa.id=sb.subscription_account_id where sb.key=#{key}")
+    @Select("select sb.key, coalesce(u.username, sb.user) as user, sb.user_id as userId, sb.subscription_account_id as subscriptionAccountId, sa.name as subscriptionAccountName, sa.enabled as subscriptionAccountEnabled, sa.filter_high_multiplier as filterHighMultiplier from sub_bind sb left join User u on u.id=sb.user_id left join subscription_account sa on sa.id=sb.subscription_account_id where sb.key=#{key}")
     SubBind selectSubBind(String key);
 
     @Select("select count(key) from sub_bind where key=#{key}")
     boolean selectSubBindExist(String key);
 
-    @Select("select count(*) from sub_bind where user=#{user}")
+    @Select("select count(*) from sub_bind where user_id=(select id from User where username=#{user}) or (user_id is null and user=#{user})")
     int countSubBindByUser(String user);
 
-    @Update("update sub_bind set subscription_account_id=#{accountId} where user=#{user}")
+    @Update("update sub_bind set subscription_account_id=#{accountId} where user_id=(select id from User where username=#{user}) or (user_id is null and user=#{user})")
     int updateSubBindAccount(@Param("user") String user, @Param("accountId") Integer accountId);
 
-    @Update("update sub_bind set key=#{key} where user=#{user}")
+    @Update("update sub_bind set key=#{key} where user_id=(select id from User where username=#{user}) or (user_id is null and user=#{user})")
     int updateSubBindKey(@Param("user") String user, @Param("key") String key);
 
-    @Delete("delete from sub_bind where user=#{user}")
+    @Delete("delete from sub_bind where user_id=(select id from User where username=#{user}) or (user_id is null and user=#{user})")
     void deleteSubBind(String user);
 
-    @Insert("insert into sub_update_record (user, ip, UA, time, location) values (#{user}, #{ip}, #{UA}, #{time}, #{location})")
+    @Insert("insert into sub_update_record (user, ip, UA, time, location, user_id) values (#{user}, #{ip}, #{UA}, #{time}, #{location}, coalesce(#{userId}, (select id from User where username=#{user})))")
+    @Options(useGeneratedKeys = true, keyProperty = "id")
     void insertSubUpdateRecord(SubUpdateRecord subUpdateRecord);
 
-    @Select("select * from sub_update_record order by time desc")
+    @Select("select id,user,ip,UA,time,location,user_id as userId from sub_update_record order by time desc")
     ArrayList<SubUpdateRecord> selectAllSubUpdateRecord();
 
-    @Delete("delete from sub_update_record where user=#{user}")
+    @Delete("delete from sub_update_record where user_id=(select id from User where username=#{user}) or (user_id is null and user=#{user})")
     void deleteSubUpdateRecord(String user);
 
-    @Select("select count(user) from sub_update_record where user=#{user}")
+    @Select("select count(user) from sub_update_record where user_id=(select id from User where username=#{user}) or (user_id is null and user=#{user})")
     Integer selectUpdateRecordCount(String user);
 
-    @Select("select min(id) from sub_update_record where user=#{user}")
+    @Select("select min(id) from sub_update_record where user_id=(select id from User where username=#{user}) or (user_id is null and user=#{user})")
     Integer selectMinUpdateRecordId(String user);
 
     /** 后台补齐归属地：只更新该用户最近一条记录，避免覆盖更早的历史记录。 */
-    @Update("update sub_update_record set location=#{location} where id=(select max(id) from sub_update_record where user=#{user})")
+    @Update("update sub_update_record set location=#{location} where id=(select max(id) from sub_update_record where user_id=(select id from User where username=#{user}) or (user_id is null and user=#{user}))")
     void updateLatestSubUpdateRecordLocation(@Param("user") String user, @Param("location") String location);
 
+    @Update("update sub_update_record set location=#{location} where id=#{id}")
+    void updateSubUpdateRecordLocation(@Param("id") int id, @Param("location") String location);
+
+    @Update("update sub_bind set user=#{newName} where user_id=#{userId} or (user_id is null and user=#{oldName})")
+    void renameBoundUser(@Param("userId") int userId, @Param("oldName") String oldName,
+                         @Param("newName") String newName);
+
     @Delete("delete from sub_update_record where id=#{id}")
     void deleteSubUpdateRecordById(int id);
 }

src/main/java/com/lion/lionwebsite/Dao/normal/TaskOperationRepository.java

@@ -0,0 +1,107 @@
+package com.lion.lionwebsite.Dao.normal;
+
+import com.lion.lionwebsite.Domain.TaskOperation;
+import org.apache.ibatis.annotations.*;
+
+/**
+ * 任务操作记录的持久化入口（主库）。
+ *
+ * <p>所有写操作都在调用方的事务里执行；这里的注解只描述 SQL，不包含业务判断。
+ * 主键为 gid，保证「同一 gid 只能有一条有效操作意图」。
+ */
+@Mapper
+public interface TaskOperationRepository {
+
+    @Select("select * from task_operation where gid=#{gid}")
+    TaskOperation selectByGid(int gid);
+
+    /** 生成该 gid 的稳定下载 operationId：只有还没有记录时才写入。 */
+    @Insert("insert or ignore into task_operation (gid, operationId, deleteOperationId, state, lastRevision," +
+            " lastReportId, updateTime) values (#{gid}, #{operationId}, null, #{state}, 0, null, #{updateTime})")
+    int insertIfAbsent(@Param("gid") int gid, @Param("operationId") String operationId,
+                       @Param("state") String state, @Param("updateTime") long updateTime);
+
+    @Insert("insert into task_operation (gid, operationId, deleteOperationId, state, lastRevision, lastReportId, updateTime)" +
+            " values (#{gid}, #{operationId}, #{deleteOperationId}, #{state}, #{lastRevision}, #{lastReportId}, #{updateTime})")
+    void insert(TaskOperation operation);
+
+    /** 推进创建意图的状态（RESERVED → SUBMITTING → SUBMITTED/UNKNOWN）。 */
+    @Update("update task_operation set state=#{state}, updateTime=#{updateTime} where gid=#{gid}")
+    int updateState(@Param("gid") int gid, @Param("state") String state, @Param("updateTime") long updateTime);
+
+    /**
+     * 带前置条件的创建状态推进：只有当前状态与 operationId 都匹配时才写入。
+     *
+     * <p>这是「提交前先确认这一代意图仍然有效」的 CAS：若期间被删除/重建，
+     * 返回值 0，调用方必须停下，绝不能再产生上游副作用。
+     */
+    @Update("update task_operation set state=#{state}, updateTime=#{updateTime}" +
+            " where gid=#{gid} and operationId=#{operationId} and state=#{expectedState}")
+    int compareAndSetCreateState(@Param("gid") int gid, @Param("operationId") String operationId,
+                                 @Param("expectedState") String expectedState, @Param("state") String state,
+                                 @Param("updateTime") long updateTime);
+
+    /**
+     * 只按当前删除意图 ID 匹配的删除完成：期间若已被删除/重建，则不改动新世代记录。
+     */
+    @Update("update task_operation set deleteOperationId=null, state=#{state}, updateTime=#{updateTime}" +
+            " where gid=#{gid} and deleteOperationId=#{deleteOperationId}")
+    int completeDeleteIfSameGeneration(@Param("gid") int gid, @Param("deleteOperationId") String deleteOperationId,
+                                       @Param("state") String state, @Param("updateTime") long updateTime);
+
+    /**
+     * 删除「本次创建意图」的预占记录：必须同时匹配 gid 与 operationId，且状态仍是
+     * RESERVED/SUBMITTING。
+     *
+     * <p>三重匹配保证释放流程只能清掉自己那一代：期间被删除/重建（新 operationId 或墓碑）
+     * 的记录一律不动，避免把新世代抹掉。
+     */
+    @Delete("delete from task_operation where gid=#{gid} and operationId=#{operationId}" +
+            " and state in ('RESERVED','SUBMITTING')")
+    int deleteReservationIfSameGeneration(@Param("gid") int gid, @Param("operationId") String operationId);
+
+    /** 记录/替换当前下载 operationId 并推进状态。 */
+    @Update("update task_operation set operationId=#{operationId}, state=#{state}, updateTime=#{updateTime} where gid=#{gid}")
+    int updateOperationId(@Param("gid") int gid, @Param("operationId") String operationId,
+                          @Param("state") String state, @Param("updateTime") long updateTime);
+
+    /**
+     * 给「有记录但缺 operationId」的历史任务补一个稳定 operationId。
+     * 只在确实缺失时写入，避免覆盖并发创建出来的新世代。
+     */
+    @Update("update task_operation set operationId=#{operationId}, updateTime=#{updateTime}" +
+            " where gid=#{gid} and operationId is null")
+    int assignOperationIdIfMissing(@Param("gid") int gid, @Param("operationId") String operationId,
+                                   @Param("updateTime") long updateTime);
+
+    /**
+     * 墓碑复活：删除完成后的 gid 允许用全新 operationId 重新创建。
+     * revision 与回执一并清零，旧操作的迟到上报会因 operationId 不匹配被丢弃。
+     */
+    @Update("update task_operation set operationId=#{operationId}, deleteOperationId=null, state=#{state}," +
+            " lastRevision=0, lastReportId=null, updateTime=#{updateTime} where gid=#{gid}")
+    int reactivate(@Param("gid") int gid, @Param("operationId") String operationId,
+                   @Param("state") String state, @Param("updateTime") long updateTime);
+
+    /** 记录删除意图与其稳定的删除操作 ID（重试必须复用同一个 ID）。 */
+    @Update("update task_operation set deleteOperationId=#{deleteOperationId}, state=#{state}, updateTime=#{updateTime} where gid=#{gid}")
+    int updateDeleteIntent(@Param("gid") int gid, @Param("deleteOperationId") String deleteOperationId,
+                           @Param("state") String state, @Param("updateTime") long updateTime);
+
+    /**
+     * 写入状态上报回执（revision + reportId）。
+     *
+     * <p>回执与任务行更新在同一个事务里提交，提交成功之后才回 ACK；
+     * 因此节点重复投递同一份上报时，只能看到已经提交过的回执，不会重复落库或重复通知。
+     */
+    @Update("update task_operation set lastRevision=#{revision}, lastReportId=#{reportId}, updateTime=#{updateTime}" +
+            " where gid=#{gid}")
+    int updateReportReceipt(@Param("gid") int gid, @Param("revision") long revision,
+                            @Param("reportId") String reportId, @Param("updateTime") long updateTime);
+
+    @Select("select * from task_operation where state=#{state}")
+    java.util.List<TaskOperation> selectByState(String state);
+
+    @Delete("delete from task_operation where gid=#{gid}")
+    int deleteByGid(int gid);
+}

src/main/java/com/lion/lionwebsite/Domain/SubBind.java

@@ -14,4 +14,10 @@ public class SubBind {
     String subscriptionAccountName;
     boolean subscriptionAccountEnabled;
     boolean filterHighMultiplier;
+    Integer userId;
+
+    public SubBind(String key, String user, Integer accountId, String accountName,
+                   boolean accountEnabled, boolean filterHighMultiplier) {
+        this(key, user, accountId, accountName, accountEnabled, filterHighMultiplier, null);
+    }
 }

src/main/java/com/lion/lionwebsite/Domain/SubUpdateRecord.java

@@ -16,4 +16,9 @@ public class SubUpdateRecord {
     String UA;
     Date time;
     String location;
+    Integer userId;
+
+    public SubUpdateRecord(int id, String user, String ip, String UA, Date time, String location) {
+        this(id, user, ip, UA, time, location, null);
+    }
 }

src/main/java/com/lion/lionwebsite/Domain/TaskOperation.java

@@ -0,0 +1,53 @@
+package com.lion.lionwebsite.Domain;
+
+import lombok.Data;
+
+/**
+ * 单个 gid 的下载生命周期操作记录（协议 v2 的持久化意图）。
+ *
+ * <p>这张表是「先持久化意图、再产生外部副作用」的落点：
+ * <ul>
+ *   <li>CREATE 意图先于上游 POST 落库，用来保证同一 gid 只有一个未完成意图；</li>
+ *   <li>提交结果不确定时只能停在 {@link #STATE_UNKNOWN}，绝不自动重复上游 POST；</li>
+ *   <li>DELETE 意图先于节点删除落库，节点确认后（或文件已不存在）才在同一事务里删任务行并留墓碑。</li>
+ * </ul>
+ *
+ * <p>旧任务行没有对应操作记录，查询时按 gid 惰性补建（见 GalleryTaskLifecycleService）。
+ */
+@Data
+public class TaskOperation {
+
+    /** 已登记创建意图，尚未向上游提交。 */
+    public static final String STATE_RESERVED = "RESERVED";
+    /** 正在向上游提交；此时不可再发第二次 POST。 */
+    public static final String STATE_SUBMITTING = "SUBMITTING";
+    /** 上游确认已接受下载请求。 */
+    public static final String STATE_SUBMITTED = "SUBMITTED";
+    /**
+     * 提交结果未知：超时、断连或在 SUBMITTING 期间重启。
+     * 上游不支持幂等，因此这里只能等对账（reconcile），不能自动重发。
+     */
+    public static final String STATE_UNKNOWN = "UNKNOWN";
+    /** 已登记删除意图，等待节点确认。 */
+    public static final String STATE_DELETE_PENDING = "DELETE_PENDING";
+    /** 节点已确认删除，任务行也随之删除；本行作为墓碑保留，用于停止迟到状态的重放。 */
+    public static final String STATE_DELETED = "DELETED";
+
+    private int gid;
+
+    /** 当前有效的下载操作 ID；同一 gid 的 DOWNLOAD 重试必须复用它。 */
+    private String operationId;
+
+    /** 独立于下载操作 ID 的删除操作 ID，删除重试复用它。 */
+    private String deleteOperationId;
+
+    private String state;
+
+    /** 最近一次应用的状态上报 revision；用于丢弃乱序/重放的旧上报。 */
+    private long lastRevision;
+
+    /** 节点已确认收到的最新上报 ID；用于重复 ACK 的幂等判断。 */
+    private String lastReportId;
+
+    private long updateTime;
+}

src/main/java/com/lion/lionwebsite/Exception/GalleryIndexException.java

@@ -0,0 +1,44 @@
+package com.lion.lionwebsite.Exception;
+
+/**
+ * 图片索引构建失败。
+ *
+ * <p>调用方需要区分三类结果，才能给出正确的用户可见响应：
+ * 链接本身有问题（{@link Reason#INVALID_LINK}，属于请求错误）、
+ * 画廊已不存在（{@link Reason#UNAVAILABLE}，重试无意义）、
+ * 上游或数据库暂时不可用（{@link Reason#TRANSIENT}，可以重试）。
+ */
+public class GalleryIndexException extends Exception {
+
+    public enum Reason {
+        /** 链接无法解析出 gid 或 key。 */
+        INVALID_LINK,
+        /** 上游明确表示画廊不存在（下架/删除/解析不到页 key）。 */
+        UNAVAILABLE,
+        /** 上游返回的页 key 无法自洽（gid 不匹配、页码重复/跳号、imgkey 为空）。 */
+        MALFORMED,
+        /** 上游波动或索引写入失败，重试可能成功。 */
+        TRANSIENT
+    }
+
+    private final Reason reason;
+
+    public GalleryIndexException(Reason reason, String message) {
+        super(message);
+        this.reason = reason;
+    }
+
+    public GalleryIndexException(Reason reason, String message, Throwable cause) {
+        super(message, cause);
+        this.reason = reason;
+    }
+
+    public Reason getReason() {
+        return reason;
+    }
+
+    /** 重试没有意义（链接错误或画廊不可用）；上游数据畸形与会话级故障仍可重试。 */
+    public boolean isPermanent() {
+        return reason == Reason.INVALID_LINK || reason == Reason.UNAVAILABLE;
+    }
+}

src/main/java/com/lion/lionwebsite/Maintenance/GalleryIndexMaintenance.java

@@ -0,0 +1,319 @@
+package com.lion.lionwebsite.Maintenance;
+
+import java.nio.file.Files;
+import java.nio.file.Path;
+import java.sql.Connection;
+import java.sql.DriverManager;
+import java.sql.PreparedStatement;
+import java.sql.ResultSet;
+import java.sql.SQLException;
+import java.sql.Statement;
+import java.time.Instant;
+import java.util.ArrayList;
+import java.util.LinkedHashMap;
+import java.util.List;
+import java.util.Map;
+
+/**
+ * 缓存库图片索引的离线体检与修复工具。
+ *
+ * <p>它是独立可执行类（只用 JDBC，不依赖 Spring / MyBatis），因此可以在服务停止后
+ * 直接对数据库文件运行，也可以先对生产库的副本演练。</p>
+ *
+ * <pre>
+ *   java -cp sqlite-jdbc.jar GalleryIndexMaintenance.java --db /path/cache.db
+ *   java -cp sqlite-jdbc.jar GalleryIndexMaintenance.java --db /path/cache.db --apply
+ * </pre>
+ *
+ * <p>它只做两类写入：删除不完整或有冲突的 gid 整组索引（gidToKey 头部加该 gid 的全部
+ * 页 key），以及删除没有头部记录的孤儿页 key。被删掉的 gid 由下一次在线阅读重新抓取，
+ * 因此不需要猜测保留哪一组 key。工具不建表、不改列、不碰其它业务表。</p>
+ */
+public final class GalleryIndexMaintenance {
+
+    /** 生产运行时数据库所在目录；默认拒绝直接写入，必须显式加 --allow-production。 */
+    private static final String PRODUCTION_DIR = "/home/lionwebsite";
+
+    private static final String PLAIN_REPAIR_NOTE =
+            "这些画廊将在下次被访问时重新抓取。";
+
+    public static void main(String[] args) {
+        Map<String, String> options = parseOptions(args);
+        if (options.containsKey("help") || !options.containsKey("db")) {
+            printUsage();
+            System.exit(options.containsKey("help") ? 0 : 2);
+            return;
+        }
+
+        Path database = Path.of(options.get("db")).toAbsolutePath().normalize();
+        boolean apply = options.containsKey("apply");
+        if (!Files.isRegularFile(database)) {
+            System.err.println("数据库不存在: " + database);
+            System.exit(2);
+            return;
+        }
+        if (apply && isProductionPath(database) && !options.containsKey("allow-production")) {
+            System.err.println("拒绝写入生产库路径 " + database
+                    + "。请先在副本上演练；确需在本机生产库执行时显式加 --allow-production，"
+                    + "并确保服务已停止。");
+            System.exit(3);
+            return;
+        }
+
+        try (Connection connection = open(database)) {
+            Report before = inspect(connection);
+            System.out.println(before.render(apply ? "修复前" : "体检结果"));
+
+            if (!apply) {
+                System.out.println();
+                System.out.println("dry-run：未做任何写入。确认无误后加 --apply 执行修复。");
+                System.exit(before.isHealthy() ? 0 : 1);
+                return;
+            }
+
+            if (before.isHealthy()) {
+                System.out.println();
+                System.out.println("索引已健康，无需修复。");
+                return;
+            }
+
+            connection.setAutoCommit(false);
+            int removed = repair(connection, before.badGids, before.orphanGids);
+            connection.commit();
+            System.out.println();
+            System.out.println("已删除 " + removed + " 个 gid 的索引整组，" + PLAIN_REPAIR_NOTE);
+
+            Report after = inspect(connection);
+            System.out.println();
+            System.out.println(after.render("修复后"));
+            if (!after.isHealthy()) {
+                System.err.println("修复后仍未达到健康状态，请人工检查。");
+                System.exit(4);
+            }
+        } catch (SQLException e) {
+            System.err.println("执行失败: " + e.getMessage());
+            e.printStackTrace(System.err);
+            System.exit(5);
+        }
+    }
+
+    private static Connection open(Path database) throws SQLException {
+        Connection connection = DriverManager.getConnection("jdbc:sqlite:" + database);
+        try (Statement statement = connection.createStatement()) {
+            statement.execute("pragma busy_timeout=5000");
+            statement.execute("pragma foreign_keys=off");
+        }
+        return connection;
+    }
+
+    /** 是否位于生产运行时目录；工具据此拒绝在未显式确认时写入生产库。 */
+    static boolean isProductionPath(Path database) {
+        return database.startsWith(Path.of(PRODUCTION_DIR).toAbsolutePath().normalize());
+    }
+
+    // ---------- 体检 ----------
+
+    static Report inspect(Connection connection) throws SQLException {
+        Report report = new Report();
+        report.hasHeaderTable = tableExists(connection, "gidToKey");
+        report.hasPageTable = tableExists(connection, "ImageKeyCache");
+        if (!report.hasHeaderTable || !report.hasPageTable)
+            return report;
+
+        report.headerRows = count(connection, "select count(*) from gidToKey");
+        report.pageRows = count(connection, "select count(*) from ImageKeyCache");
+        report.duplicateHeaderGids = count(connection,
+                "select count(*) from (select gid from gidToKey group by gid having count(*) > 1)");
+        report.duplicatePageGroups = count(connection,
+                "select count(*) from (select gid, page from ImageKeyCache group by gid, page having count(*) > 1)");
+
+        try (PreparedStatement statement = connection.prepareStatement(BAD_GID_QUERY);
+             ResultSet rows = statement.executeQuery()) {
+            while (rows.next()) {
+                String gid = rows.getString(1);
+                report.badGids.add(gid);
+                report.badReasons.put(gid, describe(rows));
+            }
+        }
+
+        try (PreparedStatement statement = connection.prepareStatement(
+                "select distinct gid from ImageKeyCache where gid not in (select gid from gidToKey)");
+             ResultSet rows = statement.executeQuery()) {
+            while (rows.next())
+                report.orphanGids.add(rows.getString(1));
+        }
+        return report;
+    }
+
+    /**
+     * 一个 gid 触发的全部不健康条件集中在一句里检查，避免多次全表扫描。
+     *
+     * <p>必须按 gid 分组后再判断：重复 gid（同一 gid 有多条头部记录）本身就是要清除的对象，
+     * 而重复行里的某一条可能恰好与页 key 数一致，逐行判断会把它当成健康记录漏掉。
+     */
+    private static final String BAD_GID_QUERY = """
+            select g.gid as gid,
+                   count(*) as header_rows,
+                   max(g.pages) as pages,
+                   min(g.key) as key,
+                   (select count(*) from ImageKeyCache i where i.gid = g.gid) as page_rows,
+                   (select count(distinct i.page) from ImageKeyCache i where i.gid = g.gid) as distinct_pages,
+                   (select coalesce(min(i.page), 0) from ImageKeyCache i where i.gid = g.gid) as min_page,
+                   (select coalesce(max(i.page), 0) from ImageKeyCache i where i.gid = g.gid) as max_page,
+                   (select count(*) from ImageKeyCache i where i.gid = g.gid
+                        and (i.imgkey is null or trim(i.imgkey) = '')) as dirty_keys
+            from gidToKey g
+            group by g.gid
+            having count(*) > 1
+                or max(g.pages) is null or max(g.pages) <= 0
+                or min(g.key) is null or trim(min(g.key)) = ''
+                or (select count(*) from ImageKeyCache i where i.gid = g.gid) <> max(g.pages)
+                or (select count(distinct i.page) from ImageKeyCache i where i.gid = g.gid) <> max(g.pages)
+                or (select coalesce(min(i.page), 0) from ImageKeyCache i where i.gid = g.gid) <> 1
+                or (select coalesce(max(i.page), 0) from ImageKeyCache i where i.gid = g.gid) <> max(g.pages)
+                or (select count(*) from ImageKeyCache i where i.gid = g.gid
+                        and (i.imgkey is null or trim(i.imgkey) = '')) > 0
+            order by g.gid
+            """;
+
+    private static String describe(ResultSet row) throws SQLException {
+        List<String> reasons = new ArrayList<>();
+        int headerRows = row.getInt("header_rows");
+        int pages = row.getInt("pages");
+        String key = row.getString("key");
+        int pageRows = row.getInt("page_rows");
+        int distinctPages = row.getInt("distinct_pages");
+        int minPage = row.getInt("min_page");
+        int maxPage = row.getInt("max_page");
+        int dirtyKeys = row.getInt("dirty_keys");
+        if (headerRows > 1) reasons.add("重复头部记录 " + headerRows + " 条");
+        if (key == null || key.isBlank()) reasons.add("key 为空");
+        if (pages <= 0) reasons.add("pages 非法(" + pages + ")");
+        if (pageRows != pages) reasons.add("页 key 行数 " + pageRows + " 与 pages " + pages + " 不符");
+        if (distinctPages != pages) reasons.add("页码去重后 " + distinctPages + " 个");
+        if (minPage != 1 || maxPage != pages) reasons.add("页码区间 " + minPage + "-" + maxPage);
+        if (dirtyKeys > 0) reasons.add("imgkey 为空 " + dirtyKeys + " 行");
+        return String.join("、", reasons);
+    }
+
+    // ---------- 修复 ----------
+
+    /** 删除不健康 gid 的整组索引与孤儿页 key，返回被移除索引的 gid 个数。 */
+    static int repair(Connection connection, List<String> badGids, List<String> orphanGids)
+            throws SQLException {
+        int touched = 0;
+        try (PreparedStatement pages = connection.prepareStatement("delete from ImageKeyCache where gid = ?");
+             PreparedStatement header = connection.prepareStatement("delete from gidToKey where gid = ?")) {
+            for (String gid : badGids) {
+                pages.setString(1, gid);
+                pages.executeUpdate();
+                header.setString(1, gid);
+                header.executeUpdate();
+                touched++;
+            }
+            for (String gid : orphanGids) {
+                pages.setString(1, gid);
+                if (pages.executeUpdate() > 0)
+                    touched++;
+            }
+        }
+        return touched;
+    }
+
+    // ---------- 基础设施 ----------
+
+    private static boolean tableExists(Connection connection, String table) throws SQLException {
+        try (PreparedStatement statement = connection.prepareStatement(
+                "select count(*) from sqlite_master where type='table' and name=?")) {
+            statement.setString(1, table);
+            try (ResultSet rows = statement.executeQuery()) {
+                return rows.next() && rows.getInt(1) > 0;
+            }
+        }
+    }
+
+    private static int count(Connection connection, String sql) throws SQLException {
+        try (Statement statement = connection.createStatement();
+             ResultSet rows = statement.executeQuery(sql)) {
+            return rows.next() ? rows.getInt(1) : 0;
+        }
+    }
+
+    private static Map<String, String> parseOptions(String[] args) {
+        Map<String, String> options = new LinkedHashMap<>();
+        for (int i = 0; i < args.length; i++) {
+            String arg = args[i];
+            if (!arg.startsWith("--"))
+                continue;
+            String name = arg.substring(2);
+            if (name.equals("apply") || name.equals("help") || name.equals("allow-production")) {
+                options.put(name, "true");
+                continue;
+            }
+            if (i + 1 >= args.length) {
+                System.err.println("缺少参数值: " + arg);
+                System.exit(2);
+            }
+            options.put(name, args[++i]);
+        }
+        return options;
+    }
+
+    private static void printUsage() {
+        System.out.println("""
+                用法: java -cp <sqlite-jdbc.jar> GalleryIndexMaintenance.java --db <cache.db> [--apply] [--allow-production]
+
+                  --db <path>            必填，缓存库文件
+                  --apply                执行修复；不加则只做 dry-run 体检
+                  --allow-production     允许写入 /home/lionwebsite 下的数据库（默认拒绝）
+
+                退出码: 0 健康或修复成功 / 1 存在待修复项(dry-run) / 2 参数错误 /
+                        3 拒绝写生产路径 / 4 修复后仍不健康 / 5 执行异常
+                """);
+    }
+
+    private GalleryIndexMaintenance() { }
+
+    /** 一次体检结果。 */
+    static final class Report {
+        boolean hasHeaderTable;
+        boolean hasPageTable;
+        int headerRows;
+        int pageRows;
+        int duplicateHeaderGids;
+        int duplicatePageGroups;
+        final List<String> badGids = new ArrayList<>();
+        final List<String> orphanGids = new ArrayList<>();
+        final Map<String, String> badReasons = new LinkedHashMap<>();
+
+        boolean isHealthy() {
+            return hasHeaderTable && hasPageTable && badGids.isEmpty() && orphanGids.isEmpty();
+        }
+
+        String render(String title) {
+            StringBuilder out = new StringBuilder();
+            out.append("=== 图片索引").append(title).append("（").append(Instant.now()).append("）===\n");
+            if (!hasHeaderTable || !hasPageTable) {
+                out.append("缺少缓存表: gidToKey=").append(hasHeaderTable)
+                   .append(" ImageKeyCache=").append(hasPageTable).append('\n');
+                return out.toString();
+            }
+            out.append("gidToKey 行数: ").append(headerRows).append('\n');
+            out.append("ImageKeyCache 行数: ").append(pageRows).append('\n');
+            out.append("重复 gid 组: ").append(duplicateHeaderGids).append('\n');
+            out.append("重复 (gid,page) 组: ").append(duplicatePageGroups).append('\n');
+            out.append("不完整或冲突 gid: ").append(badGids.size()).append('\n');
+            int limit = 20;
+            for (int i = 0; i < badGids.size() && i < limit; i++) {
+                String gid = badGids.get(i);
+                out.append("  - gid=").append(gid).append(' ')
+                   .append(badReasons.getOrDefault(gid, "")).append('\n');
+            }
+            if (badGids.size() > limit)
+                out.append("  ... 其余 ").append(badGids.size() - limit).append(" 个略\n");
+            out.append("孤儿页 key 的 gid: ").append(orphanGids.size()).append('\n');
+            out.append(isHealthy() ? "结论: 健康\n" : "结论: 需要修复\n");
+            return out.toString();
+        }
+    }
+}

src/main/java/com/lion/lionwebsite/Message/AbstractMessage.java

@@ -21,6 +21,20 @@ public class AbstractMessage {
 
     public static final byte SUBSCRIPTION_SNAPSHOT_MESSAGE = 9;
 
+    // ---- 协议 v2（可靠任务）。旧类型 0..9 的字段保持不变，不能新增字段。 ----
+
+    /** 能力协商：{@code {protocolVersion: 2, features: ["reliable-task-v2"]}}。 */
+    public static final byte PROTOCOL_HELLO_MESSAGE = 10;
+
+    /** 带稳定 operationId 的任务命令（DOWNLOAD/DELETE）。 */
+    public static final byte TASK_COMMAND_MESSAGE = 11;
+
+    /** 带 reportId + revision 的状态上报，替代旧的无确认上报。 */
+    public static final byte TASK_REPORT_MESSAGE = 12;
+
+    /** 主站提交事务成功后的上报回执，节点据此清理快照。 */
+    public static final byte TASK_REPORT_ACK_MESSAGE = 13;
+
     public byte messageType;
 
     public int messageId;

src/main/java/com/lion/lionwebsite/Message/MessageCodec.java

@@ -45,6 +45,10 @@ public class MessageCodec extends ByteToMessageCodec<AbstractMessage> {
             case AbstractMessage.MAINTAIN_MESSAGE -> objectMapper.readValue(metadata, MaintainMessage.class);
             case AbstractMessage.AVAILABLE_CHECK_MESSAGE -> objectMapper.readValue(metadata, AvailableCheckMessage.class);
             case AbstractMessage.SUBSCRIPTION_SNAPSHOT_MESSAGE -> objectMapper.readValue(metadata, SubscriptionSnapshotMessage.class);
+            case AbstractMessage.PROTOCOL_HELLO_MESSAGE -> objectMapper.readValue(metadata, ProtocolHelloMessage.class);
+            case AbstractMessage.TASK_COMMAND_MESSAGE -> objectMapper.readValue(metadata, TaskCommandMessage.class);
+            case AbstractMessage.TASK_REPORT_MESSAGE -> objectMapper.readValue(metadata, TaskReportMessage.class);
+            case AbstractMessage.TASK_REPORT_ACK_MESSAGE -> objectMapper.readValue(metadata, TaskReportAckMessage.class);
             default -> null;
         };
 

src/main/java/com/lion/lionwebsite/Message/ProtocolHelloMessage.java

@@ -0,0 +1,30 @@
+package com.lion.lionwebsite.Message;
+
+import lombok.Data;
+import lombok.NoArgsConstructor;
+
+import java.util.ArrayList;
+import java.util.List;
+
+/** 协议 v2 能力协商（类型 10）。双方都以对方声明了 version>=2 为前提才启用新消息。 */
+@Data
+@NoArgsConstructor
+public class ProtocolHelloMessage extends AbstractMessage {
+    {
+        messageType = PROTOCOL_HELLO_MESSAGE;
+    }
+
+    public static final int PROTOCOL_V2 = 2;
+
+    public static final String FEATURE_RELIABLE_TASK_V2 = "reliable-task-v2";
+
+    private int protocolVersion = PROTOCOL_V2;
+
+    private List<String> features = new ArrayList<>(List.of(FEATURE_RELIABLE_TASK_V2));
+
+    /** 对端是否声明了 v2 可靠任务能力。 */
+    public boolean supportsReliableTasks() {
+        return protocolVersion >= PROTOCOL_V2
+                && features != null && features.contains(FEATURE_RELIABLE_TASK_V2);
+    }
+}

src/main/java/com/lion/lionwebsite/Message/TaskCommandMessage.java

@@ -0,0 +1,29 @@
+package com.lion.lionwebsite.Message;
+
+import lombok.Data;
+
+/**
+ * 任务命令（类型 11）。
+ *
+ * <p>只有确认对端支持 v2 之后才发送。同一逻辑操作在重试/重连后必须复用同一个
+ * {@link #operationId}，节点据此返回原结果而不重复执行副作用。
+ */
+@Data
+public class TaskCommandMessage extends AbstractMessage {
+    {
+        messageType = TASK_COMMAND_MESSAGE;
+    }
+
+    public static final String ACTION_DOWNLOAD = "DOWNLOAD";
+
+    public static final String ACTION_DELETE = "DELETE";
+
+    /** 稳定操作 ID：DOWNLOAD 与 DELETE 各自独立。 */
+    private String operationId;
+
+    private String action;
+
+    private int gid;
+
+    private String name;
+}

src/main/java/com/lion/lionwebsite/Message/TaskReportAckMessage.java

@@ -0,0 +1,29 @@
+package com.lion.lionwebsite.Message;
+
+import lombok.Data;
+import lombok.NoArgsConstructor;
+
+/**
+ * 上报回执（类型 13）。
+ *
+ * <p>只在任务行与回执已经提交之后才发送；重复上报（同一 reportId）也会得到回执，
+ * 让节点停止重放。{@link #result} 0 表示已落库。
+ */
+@Data
+@NoArgsConstructor
+public class TaskReportAckMessage extends AbstractMessage {
+    {
+        messageType = TASK_REPORT_ACK_MESSAGE;
+    }
+
+    public static final byte RESULT_SUCCESS = 0;
+
+    private String reportId;
+
+    private byte result;
+
+    public TaskReportAckMessage(String reportId, byte result) {
+        this.reportId = reportId;
+        this.result = result;
+    }
+}

src/main/java/com/lion/lionwebsite/Message/TaskReportMessage.java

@@ -0,0 +1,32 @@
+package com.lion.lionwebsite.Message;
+
+import com.lion.lionwebsite.Domain.GalleryTask;
+import lombok.Data;
+
+import java.util.LinkedHashMap;
+import java.util.Map;
+
+/**
+ * 节点状态上报（类型 12）。
+ *
+ * <p>相比旧的无确认上报，这里带上单调递增的 {@link #revision} 与唯一 {@link #reportId}：
+ * 主站据此丢弃乱序/重放的旧快照，并在事务提交后回 {@link TaskReportAckMessage}，
+ * 节点只有在收到匹配回执后才清理自己的快照。
+ */
+@Data
+public class TaskReportMessage extends AbstractMessage {
+    {
+        messageType = TASK_REPORT_MESSAGE;
+    }
+
+    /** 本次上报的唯一 ID，重复投递同一份快照时保持不变。 */
+    private String reportId;
+
+    /** 节点侧单调递增版本号；主站只接受比已记录更大的 revision。 */
+    private long revision;
+
+    private GalleryTask[] galleryTasks;
+
+    /** gid → 产生该状态的下载 operationId，用于丢弃属于旧操作的迟到上报。 */
+    private Map<Integer, String> operationIds = new LinkedHashMap<>();
+}

src/main/java/com/lion/lionwebsite/Service/GalleryIndexService.java

@@ -0,0 +1,262 @@
+package com.lion.lionwebsite.Service;
+
+import com.lion.lionwebsite.Dao.cache.ImageCacheMapper;
+import com.lion.lionwebsite.Domain.GidToKey;
+import com.lion.lionwebsite.Domain.ImageKeyCache;
+import com.lion.lionwebsite.Exception.GalleryIndexException;
+import com.lion.lionwebsite.Util.GalleryUtil;
+import com.lion.lionwebsite.Util.SingleFlight;
+import org.slf4j.Logger;
+import org.slf4j.LoggerFactory;
+import org.springframework.transaction.support.TransactionOperations;
+
+import java.util.ArrayList;
+import java.util.HashSet;
+import java.util.List;
+import java.util.Set;
+
+/**
+ * 图片索引（gid 到每页 imgkey）的唯一读写入口。
+ *
+ * <p>这个类替换掉原先「读缓存 / 写缓存 / 回滚」分散在业务服务里、并且逐页 INSERT
+ * 的写法。三件事在这里被固定下来：
+ *
+ * <ol>
+ *   <li><b>同 gid 合并</b>：并发请求同一本画廊时只有一个线程真的去抓上游，
+ *       其余线程等待同一个结果（含失败），因此不会出现两个构建者互相删除对方写入的索引。</li>
+ *   <li><b>抓取在事务外</b>：上游 HTTP 可能耗时十几秒，绝不能占着 SQLite 写事务，
+ *       否则整个缓存库的写入都会排队。</li>
+ *   <li><b>整组短事务发布</b>：一个事务里先删该 gid 的旧索引，再批量写入新索引。
+ *       失败整体回滚，磁盘上仍保留旧的完整索引；只有写入成功这一个状态是对外可见的。</li>
+ * </ol>
+ *
+ * <p>依赖全部由构造函数注入：mapper 负责落库，operations 负责事务边界，
+ * upstream 负责把 mpv 页解析成页 key 列表。测试可以直接给假的实现，
+ * 不需要起 Spring 上下文，也不需要网络。
+ */
+public class GalleryIndexService {
+
+    private static final Logger log = LoggerFactory.getLogger(GalleryIndexService.class);
+
+    /** 单条 INSERT 携带的最大行数，避免 SQL 过长。 */
+    static final int BATCH_SIZE = 200;
+
+    private final ImageCacheMapper mapper;
+    private final TransactionOperations operations;
+    private final TransactionOperations readOperations;
+    private final GalleryKeyUpstream upstream;
+    private final SingleFlight<String, GidToKey> builds = new SingleFlight<>();
+
+    public GalleryIndexService(ImageCacheMapper mapper, TransactionOperations operations,
+                              GalleryKeyUpstream upstream) {
+        this(mapper, operations, operations, upstream);
+    }
+
+    /**
+     * @param operations     发布索引的写事务（删除加批量写入必须整体生效）
+     * @param readOperations 读取索引的一致性快照事务；传 null 表示不额外开启读事务
+     */
+    public GalleryIndexService(ImageCacheMapper mapper, TransactionOperations operations,
+                               TransactionOperations readOperations, GalleryKeyUpstream upstream) {
+        if (mapper == null || operations == null || upstream == null)
+            throw new IllegalArgumentException("mapper/operations/upstream 不能为空");
+        this.mapper = mapper;
+        this.operations = operations;
+        this.readOperations = readOperations;
+        this.upstream = upstream;
+    }
+
+    /**
+     * 保证该画廊的索引可用，返回 gid 到 key 的映射。
+     *
+     * @param url 画廊链接，形如 https://exhentai.org/g/1596929/f08534d87d/
+     * @throws GalleryIndexException 链接错误 / 画廊不可用 / 暂时失败，见 GalleryIndexException.Reason
+     */
+    public GidToKey ensureIndex(String url) throws GalleryIndexException {
+        return ensureIndex(url, this::buildAndPublish);
+    }
+
+    /**
+     * 同 ensureIndex，但允许调用方提供构建动作。
+     *
+     * <p>离线重建需要「先检查再决定是否重建」的不同策略，因此把构建步骤作为参数暴露出来；
+     * 生产调用使用默认实现即可。
+     */
+    GidToKey ensureIndex(String url, IndexBuilder builder) throws GalleryIndexException {
+        GalleryLink link = GalleryLink.parse(url);
+        GidToKey cached = lookup(link);
+        if (cached != null)
+            return cached;
+        try {
+            return builds.run(link.gid(), () -> {
+                // 进入临界区后必须再看一眼：等待期间别的线程可能已经把索引建好了。
+                GidToKey existing = lookup(link);
+                return existing != null ? existing : builder.build(url);
+            });
+        } catch (GalleryIndexException e) {
+            throw e;
+        } catch (Exception e) {
+            throw new GalleryIndexException(GalleryIndexException.Reason.TRANSIENT,
+                    "构建图片索引失败 gid=" + link.gid(), e);
+        }
+    }
+
+    /**
+     * 已缓存且完整则返回索引，否则返回 null。
+     *
+     * <p>完整性由页码集合判定，而不是行数：历史重复行与半截写入都会让行数与页数不符。
+     */
+    GidToKey lookup(GalleryLink link) {
+        if (readOperations != null)
+            return readOperations.execute(status -> lookupInTransaction(link));
+        return lookupInTransaction(link);
+    }
+
+    /**
+     * 在同一个读事务里读头部与页码，得到一个一致的快照。
+     *
+     * <p>两个独立的 SELECT 之间可能夹着另一个线程的整组替换，于是头部是旧的、页码是新的
+     * （或相反），完整性判断会得出错误结论。SQLite 的读事务能看到固定快照，避免这种交错。
+     */
+    private GidToKey lookupInTransaction(GalleryLink link) {
+        GidToKey header = mapper.selectKeyByGid(link.gid());
+        if (header == null)
+            return null;
+        if (header.getPages() <= 0 || header.getKey() == null || header.getKey().isBlank())
+            return null;
+        if (!isComplete(header))
+            return null;
+        return header;
+    }
+
+    private boolean isComplete(GidToKey header) {
+        List<Integer> pages = mapper.selectPageNumbersByGid(header.getGid());
+        if (pages == null || pages.size() != header.getPages())
+            return false;
+        if (mapper.countInvalidImageKeysByGid(header.getGid()) != 0)
+            return false;
+        // 页码必须是严格递增的 1..pages，重复或跳号都视为不完整。
+        for (int i = 0; i < pages.size(); i++) {
+            Integer page = pages.get(i);
+            if (page == null || page != i + 1)
+                return false;
+        }
+        return true;
+    }
+
+    /** 抓上游（事务外）后短事务整组替换。 */
+    private GidToKey buildAndPublish(String url) throws GalleryIndexException {
+        GalleryLink link = GalleryLink.parse(url);
+        List<ImageKeyCache> pageKeys;
+        try {
+            // 抓取必须发生在事务之外：上游可能耗时十几秒，占着 SQLite 写事务会拖垮整个缓存库。
+            pageKeys = upstream.fetchPageKeys(url);
+        } catch (Exception e) {
+            throw new GalleryIndexException(GalleryIndexException.Reason.TRANSIENT,
+                    "抓取图片索引失败 gid=" + link.gid(), e);
+        }
+        if (pageKeys == null || pageKeys.isEmpty())
+            throw new GalleryIndexException(GalleryIndexException.Reason.UNAVAILABLE,
+                    "该图片已下架或已被删除 gid=" + link.gid());
+
+        // 落库前先校验上游数据自洽：整组替换一旦开始就没有回头路，
+        // 与其把畸形页码写进唯一约束再回滚，不如在此之前就判为失败。
+        validatePageKeys(link, pageKeys);
+
+        GidToKey header = new GidToKey();
+        header.setGid(link.gid());
+        header.setKey(link.key());
+        header.setPages(pageKeys.size());
+
+        publish(header, pageKeys);
+        return header;
+    }
+
+    /**
+     * 校验上游返回的页 key 是否自洽。
+     *
+     * <p>四条要求缺一不可，否则索引会以「看起来完整」的样子写进库：
+     * gid 属于请求的那本画廊、页码恰好是 1..N、页码不重复、imgkey 非空。
+     * 任一不满足都抛出 {@link GalleryIndexException.Reason#MALFORMED}，此时尚未开启写事务，
+     * 库中旧索引保持原样。
+     */
+    void validatePageKeys(GalleryLink link, List<ImageKeyCache> pageKeys) throws GalleryIndexException {
+        int expected = pageKeys.size();
+        Set<Integer> seen = new HashSet<>(expected * 2);
+        for (ImageKeyCache row : pageKeys) {
+            if (row == null || row.getGid() == null || !row.getGid().equals(link.gid()))
+                throw malformed(link, "上游页 key 的 gid 与请求不一致");
+            if (row.getImgkey() == null || row.getImgkey().isBlank())
+                throw malformed(link, "上游页 key 的 imgkey 为空");
+            int page = row.getPage();
+            if (page < 1 || page > expected)
+                throw malformed(link, "上游页码越界: " + page + "（共 " + expected + " 页）");
+            if (!seen.add(page))
+                throw malformed(link, "上游页码重复: " + page);
+        }
+        if (seen.size() != expected)
+            throw malformed(link, "上游页码不连续，去重后 " + seen.size() + " 个，期望 " + expected + " 个");
+    }
+
+    private GalleryIndexException malformed(GalleryLink link, String detail) {
+        log.warn("上游图片索引非法 gid={} detail={}", link.gid(), detail);
+        return new GalleryIndexException(GalleryIndexException.Reason.MALFORMED,
+                detail + " gid=" + link.gid());
+    }
+
+    /**
+     * 在单个事务里整组替换索引。
+     *
+     * <p>失败时事务整体回滚：要么看到旧的完整索引，要么看到新的完整索引，
+     * 不存在「头部已更新、页 key 缺失」的中间态。
+     */
+    void publish(GidToKey header, List<ImageKeyCache> pageKeys) {
+        operations.executeWithoutResult(status -> {
+            mapper.deleteImageKeyCacheByGid(header.getGid());
+            mapper.deleteGidToKey(header.getGid());
+            mapper.insertGidToKey(header);
+            writeBatches(pageKeys);
+        });
+        log.info("图片索引已整体替换 gid={} pages={}", header.getGid(), header.getPages());
+    }
+
+    private void writeBatches(List<ImageKeyCache> pageKeys) {
+        for (int start = 0; start < pageKeys.size(); start += BATCH_SIZE) {
+            int end = Math.min(start + BATCH_SIZE, pageKeys.size());
+            mapper.insertImageKeyCacheBatch(new ArrayList<>(pageKeys.subList(start, end)));
+        }
+    }
+
+    /** 便于日志与错误信息使用的解析结果。 */
+    record GalleryLink(String gid, String key) {
+
+        static GalleryLink parse(String url) throws GalleryIndexException {
+            String[] segments = url == null ? null : url.split("/");
+            Integer gid = GalleryUtil.parseGid(url);
+            if (gid == null || segments == null || segments.length <= 5 || segments[5].isBlank())
+                throw new GalleryIndexException(GalleryIndexException.Reason.INVALID_LINK,
+                        "链接错误");
+            return new GalleryLink(String.valueOf(gid), segments[5].strip());
+        }
+    }
+
+    /**
+     * 上游取页 key 的端口。
+     *
+     * <p>生产实现只是把 URL 转交给 GalleryUtil.parseImageKeys；
+     * 单独抽出来是为了让本服务在测试里完全脱离网络与静态工具类。
+     */
+    @FunctionalInterface
+    public interface GalleryKeyUpstream {
+        List<ImageKeyCache> fetchPageKeys(String url) throws Exception;
+    }
+
+    /**
+     * 构建索引的动作。允许抛出业务异常，因此不用 java.util.function.Function
+     * （后者不能抛受检异常，会让调用方被迫包装成运行时异常）。
+     */
+    @FunctionalInterface
+    interface IndexBuilder {
+        GidToKey build(String url) throws GalleryIndexException;
+    }
+}

src/main/java/com/lion/lionwebsite/Service/GalleryManageService.java

@@ -46,6 +46,16 @@ public class GalleryManageService {
 
     final PushService pushService;
 
+    /**
+     * 任务生命周期（协议 v2）。创建/删除都先走它做持久化意图，再产生外部副作用。
+     * 用字段注入是为了兼容现有构造式 mock 测试与旧调用点。
+     */
+    @org.springframework.beans.factory.annotation.Autowired(required = false)
+    GalleryTaskLifecycleService lifecycleService;
+
+    @org.springframework.beans.factory.annotation.Autowired
+    GalleryIndexService galleryIndexService;
+
     /**
      * 创建任务
      * @param link             任务链接
@@ -79,6 +89,20 @@ public class GalleryManageService {
             pushService.taskCreateReport(user.getUsername(), taskName, response);
             return response.toJSONString();
         }
+
+        // 协议 v2 生命周期：gid 唯一意图先落库（占位行「提交中」），提交前持久化 SUBMITTING，
+        // 上游结果不确定时报 UNKNOWN 而不是自动重发；差额额度与任务行在同一短事务里写入。
+        if (lifecycleService != null) {
+            response = lifecycleService.createTask(gid, taskName, link, targetResolution, user.getId());
+            Gallery current = galleryMapper.selectGalleryByGid(gid);
+            taskName = current == null || current.getName() == null ? taskName : current.getName();
+            // 业务结果由生命周期服务给出，data 用结构化对象下发（前端解码器兼容旧字符串）。
+            if (response.isSuccess() && current != null)
+                response.success(objectMapper.valueToTree(current));
+            pushService.taskCreateReport(user.getUsername(), taskName, response);
+            return response.toJSONString();
+        }
+
         Gallery gallery;
 
         //判断数据库中是否有这个任务，有则返回状态
@@ -124,7 +148,7 @@ public class GalleryManageService {
 
         // Do not overwrite an immediate node status with the original submitted state.
         Gallery current = galleryMapper.selectGalleryByGid(gallery.getGid());
-        response.success((current == null ? gallery : current).toString());
+        response.success(objectMapper.valueToTree(current == null ? gallery : current));
         pushService.taskCreateReport(user.getUsername(), taskName, response);
 
         return response.toJSONString();
@@ -184,7 +208,7 @@ public class GalleryManageService {
             try {
                 gallery = GalleryUtil.parse(link, false, null);
                 if (gallery != null)
-                    response.success(objectMapper.valueToTree(gallery).toString());
+                    response.success(objectMapper.valueToTree(gallery));
                 else
                     response.failure("查询失败");
             } catch (Exception e) {
@@ -193,7 +217,7 @@ public class GalleryManageService {
             }
 
         else
-            response.success(objectMapper.valueToTree(gallery).toString());
+            response.success(objectMapper.valueToTree(gallery));
 
         return response.toJSONString();
     }
@@ -212,7 +236,7 @@ public class GalleryManageService {
             response.failure("未找到该图片，请使用链接");
 
         else
-            response.success(gallery.toString());
+            response.success(objectMapper.valueToTree(gallery));
 
         return response.toJSONString();
     }
@@ -245,7 +269,7 @@ public class GalleryManageService {
         if (userId == UserService.ADMIN_USER_ID)
             fillDownloaderNames(galleries);
 
-        response.success(objectMapper.valueToTree(galleries).toString());
+        response.success(objectMapper.valueToTree(galleries));
         return response.toJSONString();
     }
 
@@ -302,7 +326,7 @@ public class GalleryManageService {
         Gallery gallery = galleryMapper.selectGalleryByName("%" + name + "%");
 
         if (gallery != null)
-            response.success(objectMapper.valueToTree(gallery).toString());
+            response.success(objectMapper.valueToTree(gallery));
         else
             response.failure("没有找到该名字的图片");
 
@@ -320,7 +344,7 @@ public class GalleryManageService {
 
         Gallery[] galleries = galleryMapper.selectGalleryByDownloader(userMapper.selectUserByAuthCode(AuthCode).getId());
         if (galleries.length > 0)
-            response.success(objectMapper.valueToTree(galleries).toString());
+            response.success(objectMapper.valueToTree(galleries));
         else
             response.failure("您未下载图片");
 
@@ -358,8 +382,46 @@ public class GalleryManageService {
             return response.toJSONString();
         }
 
-        // 通过授权后才落库并通知节点，避免被拒请求仍删除节点文件。
+        // 通过授权后才产生外部副作用。协议 v2 的删除顺序：
+        // 先落「待删除」意图 → 节点确认（或文件已不存在）→ 才在同一事务删任务行并留墓碑。
+        // 失败时任务行与意图都保留，可沿原记录重试并复用同一个删除 operationId。
         log.info("删除图片{}", gallery.getName());
+        if (lifecycleService != null) {
+            // 被拒绝（例如任务仍在提交中）时必须回业务失败：
+            // 绝不能退避到裸删除，否则会与并发的创建 POST 同时执行并写回已删/重建的 gid。
+            GalleryTaskLifecycleService.DeleteIntent intent = lifecycleService.beginDelete(gallery.getGid());
+            if (!intent.accepted()) {
+                response.failure(intent.message());
+                log.info("拒绝删除 gid={}：{}", gid, intent.message());
+                return response.toJSONString();
+            }
+            byte result = remoteService.sendDeleteCommand(intent.deleteOperationId(),
+                    gallery.getGid(), gallery.getName());
+            switch (result) {
+                case 0 -> {
+                    lifecycleService.completeDelete(gallery.getGid(), intent.deleteOperationId());
+                    response.success();
+                }
+                // 文件已不存在等于目标达成：同样完成本地清理，避免留下永远删不掉的任务行。
+                case ErrorCode.FILE_NOT_FOUND -> {
+                    lifecycleService.completeDelete(gallery.getGid(), intent.deleteOperationId());
+                    response.success();
+                }
+                case ErrorCode.IO_ERROR -> {
+                    lifecycleService.failDelete(gallery.getGid());
+                    response.failure("图片:" + gallery.getName() + "删除失败，IO错误");
+                }
+                // 节点无响应/超时：保留记录与待删除意图，可重试。
+                default -> {
+                    lifecycleService.failDelete(gallery.getGid());
+                    response.failure("图片:" + gallery.getName() + "删除失败，节点无响应");
+                }
+            }
+            if (!response.isSuccess())
+                log.info(response.getData());
+            return response.toJSONString();
+        }
+
         galleryMapper.deleteGalleryByGid(gallery.getGid());
         switch (remoteService.deleteGallery(gallery)) {
             case ErrorCode.IO_ERROR -> response.failure("图片:" + gallery.getName() + "删除失败，IO错误");
@@ -399,76 +461,25 @@ public class GalleryManageService {
                 lastResetAmountTime == null || lastResetAmountTime.getValue() == null
                         ? "" : lastResetAmountTime.getValue());
 
-        response.success(objectMapper.valueToTree(data).toString());
+        response.success(objectMapper.valueToTree(data));
         return response.toJSONString();
     }
 
     public String cacheImagesKey(String url) {
-        Response response = Response.generateResponse();
-
-        // 畸形链接在这里就拦掉：段数不足会让 url.split("/")[5] 抛 ArrayIndexOutOfBoundsException，
-        // 而 parseGid 也会对无 /g/ 的链接返回 null，两者都不该变成 500。
-        String[] segments = url == null ? null : url.split("/");
-        Integer parsedGid = parseGidFromLink(url);
-        if (parsedGid == null || segments == null || segments.length <= 5) {
-            response.failure("链接错误");
-            return response.toJSONString();
-        }
-
-        String gid = String.valueOf(parsedGid);
-        GidToKey gidToKey = imageCacheMapper.selectKeyByGid(gid);
-        //已缓存过且页 key 完整，直接返回
-        if(gidToKey != null && isIndexComplete(gidToKey)) {
-            return response.success(objectMapper.valueToTree(gidToKey)).toJSONString();
-        }
-        // 半截缓存（历史崩溃或旧版本写入中断）先清干净再整体重建，否则残留的页 key 会重复累积。
-        if (gidToKey != null) {
-            log.warn("图片索引缓存不完整，重建 gid={} 记录页数={}", gid, gidToKey.getPages());
-            rollbackImageIndex(gid);
-        }
-
-        boolean inserted = false;
         try {
-            gidToKey = new GidToKey();
-            gidToKey.setGid(gid);
-            gidToKey.setKey(segments[5].strip());
-            ArrayList<ImageKeyCache> imageKeyCaches = GalleryUtil.parseImageKeys(url);
-            if(imageKeyCaches == null)
-                return response.failure("该图片已下架或已被删除").toJSONString();
-            gidToKey.setPages(imageKeyCaches.size());
-            imageCacheMapper.insertGidToKey(gidToKey);
-            inserted = true;
-            for (ImageKeyCache imageKeyCache : imageKeyCaches)
-                imageCacheMapper.insertImageKeyCache(imageKeyCache);
-            response.success(objectMapper.valueToTree(gidToKey));
-        }catch (Exception e){
-            // 索引必须整体生效：先写 gidToKey 再逐页写 key，中途失败会留下「gidToKey 命中、
-            // 页 key 缺失」的半截缓存，后续请求直接返回已缓存而永远取不到图。
-            // 因此任何异常都回滚已写入的部分再回业务失败。Jackson 3 的解析异常继承
-            // RuntimeException 而非 IOException，只用 IOException 会漏掉，这里统一兜住。
-            if (inserted)
-                rollbackImageIndex(gid);
-            log.warn("缓存图片索引失败 gid={} errorType={}", gid, e.getClass().getSimpleName());
-            response.failure("网络波动或其他异常");
-        }
-        return  response.toJSONString();
-    }
-
-    /** 页 key 行数与 gidToKey 记录不一致时视为半截缓存，需要重建。 */
-    private boolean isIndexComplete(GidToKey gidToKey) {
-        Integer pages = gidToKey.getPages();
-        if (pages == null || pages <= 0)
-            return false;
-        return imageCacheMapper.countImageKeyCacheByGid(gidToKey.getGid()) == pages;
-    }
-
-    /** 删除某个 gid 已写入的索引（页 key 与 gidToKey），用于回滚或重建。 */
-    private void rollbackImageIndex(String gid) {
-        try {
-            imageCacheMapper.deleteImageKeyCacheByGid(gid);
-            imageCacheMapper.deleteGidToKey(gid);
-        } catch (Exception e) {
-            log.warn("回滚图片索引缓存失败 gid={}", gid, e);
+            GidToKey index = galleryIndexService.ensureIndex(url);
+            return Response.generateResponse().success(objectMapper.valueToTree(index)).toJSONString();
+        } catch (com.lion.lionwebsite.Exception.GalleryIndexException error) {
+            String message = switch (error.getReason()) {
+                case INVALID_LINK -> "链接错误";
+                case UNAVAILABLE -> "该图片已下架或已被删除";
+                case MALFORMED, TRANSIENT -> "网络波动或其他异常";
+            };
+            log.warn("缓存图片索引失败 reason={}", error.getReason());
+            return Response._failure(message);
+        } catch (RuntimeException error) {
+            log.warn("缓存图片索引失败 errorType={}", error.getClass().getSimpleName());
+            return Response._failure("网络波动或其他异常");
         }
     }
 
@@ -517,10 +528,24 @@ public class GalleryManageService {
 
         Gallery[] galleries = galleryMapper.selectUnDoneGalleries();
         if(galleries != null && galleries.length != 0) {
-            log.info("发送未下载完成图片至节点,{}本", galleries.length);
-            for (Gallery gallery : galleries)
+            // 只重投「确实已提交上游」的任务：提交未知/预占中/待删除的任务重投会变成
+            // 未计量副作用（节点侧显示下载完成但没有任何额度记录）。
+            int sent = 0;
+            int skipped = 0;
+            for (Gallery gallery : galleries) {
+                if (lifecycleService != null
+                        && !lifecycleService.resolveDispatch(gallery.getGid()).dispatchable()) {
+                    skipped++;
+                    continue;
+                }
                 remoteService.addGalleryToQueue(gallery);
-            response.success(String.format("发送未下载完成图片至节点,%s本", galleries.length));
+                sent++;
+            }
+            log.info("发送未下载完成图片至节点,{}本（跳过需要对账的{}本）", sent, skipped);
+            if (sent == 0)
+                response.failure("没有可安全重投的任务，需先人工对账提交未知的任务");
+            else
+                response.success(String.format("发送未下载完成图片至节点,%s本", sent));
         }else{
             response.failure("当前没有未下载完成的图片");
         }
@@ -536,6 +561,13 @@ public class GalleryManageService {
             return Response._success("下载完成");
         if(remoteService.isDead())
             return Response._failure("节点不在线，无法重试");
+        // 提交结果未知（UNKNOWN）或正在删除的任务不能重试：上游不支持幂等，重发会让
+        // 「可能已受理但未计量」的任务在节点侧变成下载完成却 0 额度。这类任务需人工对账。
+        if (lifecycleService != null) {
+            var decision = lifecycleService.resolveDispatch(gid);
+            if (decision.kind() == GalleryTaskLifecycleService.DispatchDecision.Kind.REFUSE)
+                return Response._failure("该任务提交结果未知或正在删除，请先人工对账后再重试");
+        }
 
         RemoteService.RetryResult retryResult = remoteService.retryGallery(gallery);
         if(retryResult.success())

src/main/java/com/lion/lionwebsite/Service/GalleryTaskLifecycleService.java

@@ -0,0 +1,783 @@
+package com.lion.lionwebsite.Service;
+
+import com.lion.lionwebsite.Dao.normal.*;
+import com.lion.lionwebsite.Domain.*;
+import com.lion.lionwebsite.Exception.ResolutionNotMatchException;
+import com.lion.lionwebsite.Message.TaskCommandMessage;
+import com.lion.lionwebsite.Message.TaskReportMessage;
+import com.lion.lionwebsite.Util.PreparedGallery;
+import com.lion.lionwebsite.Util.Response;
+import lombok.extern.slf4j.Slf4j;
+import org.springframework.beans.factory.annotation.Qualifier;
+import org.springframework.stereotype.Service;
+import org.springframework.transaction.support.TransactionTemplate;
+
+import java.io.IOException;
+import java.util.ArrayList;
+import java.util.List;
+
+import static com.lion.lionwebsite.Util.CustomUtil.objectMapper;
+
+/**
+ * 下载任务的持久化生命周期（协议 v2）。
+ *
+ * <p>核心约束：
+ * <ol>
+ *   <li><b>先保存意图，再产生外部副作用</b>：同一 gid 的唯一 CREATE 意图先落库（占位任务行 + RESERVED 操作记录），
+ *       通过 gid 的主键唯一约束把并发创建收敛成「一个赢家」；</li>
+ *   <li><b>SUBMITTING 之后绝不自动重发</b>：上游不支持幂等，提交结果不确定时只能落 UNKNOWN，等显式对账/解决；</li>
+ *   <li><b>提交结果与额度用一个短事务写入</b>：Gallery 与 weekUsedAmount 要么一起成功，要么一起回滚，
+ *       事务提交之后才把 DOWNLOAD 命令下发给节点；</li>
+ *   <li><b>删除失败保留记录</b>：先落待删除意图，节点确认（或文件已不存在）后才在同一事务里删任务行并留墓碑；
+ *       失败则保留 Gallery 与 DELETE_PENDING 意图，可沿原记录重试且复用同一删除 operationId。</li>
+ * </ol>
+ *
+ * <p>所有网络调用都发生在事务之外；事务回调内部不做 I/O，也不吞异常（吞掉会错误提交）。
+ */
+@Service
+@Slf4j
+public class GalleryTaskLifecycleService implements TaskReportHandler {
+
+    final GalleryMapper galleryMapper;
+
+    final CollectMapper collectMapper;
+
+    final CustomConfigurationMapper configurationMapper;
+
+    final UserMapper userMapper;
+
+    final TaskOperationRepository taskOperationRepository;
+
+    final UpstreamGateway upstreamGateway;
+
+    final RemoteService remoteService;
+
+    /**
+     * 短全局创建预占锁。
+     *
+     * <p>只有「登记唯一 CREATE 意图」这一小段数据库写入在锁内，网络调用一律在锁外。
+     * 单进程内它把并发同 gid 创建直接收敛成一个赢家；跨进程的正确性仍由
+     * {@code task_operation}/{@code gallery} 的主键唯一约束兜底。
+     */
+    private final Object reservationLock = new Object();
+
+    /**
+     * 测试钩子：进入上游只读解析之前执行。
+     * 用来让并发删除发生在预占之后、SUBMITTING 之前。生产环境是空实现。
+     */
+    volatile Runnable beforePrepare = () -> { };
+
+    /**
+     * 测试钩子：SUBMITTING 已持久化之后、真正 POST 上游之前执行。
+     * 用来制造「提交瞬间任务被删除」的 late generation 场景。生产环境是空实现。
+     */
+    volatile Runnable afterMarkSubmitting = () -> { };
+
+    /** 主库事务模板；测试可注入内存 sqlite 的事务模板。 */
+    TransactionTemplate transactions;
+
+    public GalleryTaskLifecycleService(GalleryMapper galleryMapper,
+                                       CollectMapper collectMapper,
+                                       CustomConfigurationMapper configurationMapper,
+                                       UserMapper userMapper,
+                                       TaskOperationRepository taskOperationRepository,
+                                       UpstreamGateway upstreamGateway,
+                                       RemoteService remoteService,
+                                       @Qualifier("mainTransactionTemplate") TransactionTemplate transactions) {
+        this.galleryMapper = galleryMapper;
+        this.collectMapper = collectMapper;
+        this.configurationMapper = configurationMapper;
+        this.userMapper = userMapper;
+        this.taskOperationRepository = taskOperationRepository;
+        this.upstreamGateway = upstreamGateway;
+        this.remoteService = remoteService;
+        this.transactions = transactions;
+    }
+
+    /** 测试用：替换事务模板或在没有 Spring 事务时退化为直连执行。 */
+    void setTransactions(TransactionTemplate transactions) {
+        this.transactions = transactions;
+    }
+
+    /**
+     * 一次「登记创建意图」的结果。
+     *
+     * <p>只有赢家能继续做外部副作用；输家（已存在有效意图 / 数据库竞争失败）
+     * 一律转成业务失败，绝不再 POST 上游。
+     */
+    private enum ReserveOutcome {
+        /** 已成功登记唯一 CREATE 意图，可以做只读解析与提交。 */
+        RESERVED,
+        /** 同一 gid 已存在有效意图（或任务行）。 */
+        DUPLICATE,
+        /** 登记时的数据库竞争/失败：没有产生任何副作用，必须转业务失败。 */
+        CONTENDED
+    }
+
+    /** 一次「登记创建意图」的返回：结果 + 赢家持有的操作记录。 */
+    private record Reservation(ReserveOutcome outcome, TaskOperation operation) {
+    }
+
+    /**
+     * 单次下发决策。
+     *
+     * <p>把「能不能把这个任务交给节点」从「下载是否已提交」里拆出来：
+     * 未确认上游提交结果（UNKNOWN / 预占 / 待删除 / 删除墓碑）绝不能下发，
+     * 否则未计量的人工对账任务可能在节点侧变成「下载完成却 0 额度」。
+     */
+    public static final class DispatchDecision {
+
+        public enum Kind {
+            /** 当前 operation 已是 SUBMITTED：可以按稳定 operationId 下发。 */
+            READY,
+            /** 迁移前的旧任务行：没有操作记录但历史上确实提交过下载。 */
+            LEGACY,
+            /** 不可下发：状态不明或被删除，下发会产生未计量副作用。 */
+            REFUSE
+        }
+
+        private final Kind kind;
+
+        private final String operationId;
+
+        private DispatchDecision(Kind kind, String operationId) {
+            this.kind = kind;
+            this.operationId = operationId;
+        }
+
+        public Kind kind() {
+            return kind;
+        }
+
+        public String operationId() {
+            return operationId;
+        }
+
+        /** 只有 READY/LEGACY 允许真正下发到节点。 */
+        public boolean dispatchable() {
+            return kind != Kind.REFUSE && operationId != null;
+        }
+
+        /** 拒绝下发（供调用方与测试构造）。 */
+        public static DispatchDecision refuse() {
+            return REFUSE;
+        }
+
+        /** 可直接下发（供测试构造）。 */
+        public static DispatchDecision ready(String operationId) {
+            return new DispatchDecision(Kind.READY, operationId);
+        }
+    }
+
+    private static final DispatchDecision REFUSE = new DispatchDecision(DispatchDecision.Kind.REFUSE, null);
+
+    /**
+     * 一次删除登记的返回。
+     *
+     * <p>用显式结果替代「null 即退避到裸删除」的旧语义：
+     * 被拒绝时调用方必须回业务失败，不能绕过待删除意图直接删节点文件。
+     */
+    public record DeleteIntent(boolean accepted, String deleteOperationId, String message) {
+
+        static DeleteIntent refused(String message) {
+            return new DeleteIntent(false, null, message);
+        }
+
+        static DeleteIntent accepted(String deleteOperationId) {
+            return new DeleteIntent(true, deleteOperationId, null);
+        }
+    }
+
+    /** 在事务里执行；没有配置事务模板时候直接执行（仅便于构造 mock 测试）。 */
+    private <T> T inTransaction(java.util.function.Supplier<T> body) {
+        TransactionTemplate template = transactions;
+        if (template == null)
+            return body.get();
+        return template.execute(status -> {
+            try {
+                return body.get();
+            } catch (RuntimeException e) {
+                // 回调里必须把异常继续抛出，否则 Spring 会把只读了一半的事务当成正常提交。
+                status.setRollbackOnly();
+                throw e;
+            }
+        });
+    }
+
+    /**
+     * 创建任务的持久化前置：在同一个事务里登记唯一 CREATE 意图。
+     *
+     * <p>同时写入可见的占位任务行（状态「提交中」），这样提交期间任务列表就能解析该 gid，
+     * 不会出现「上游已受理但本地查不到」的空窗。
+     *
+     * @return 已登记意图（赢家）返回它；同一 gid 已有有效意图返回 {@code null}（输家）。
+     */
+    private Reservation reserveCreateIntent(int gid, String placeholderName, int downloaderId) {
+        // 锁内只做数据库写入：把并发同 gid 创建串行化，网络调用留在锁外。
+        synchronized (reservationLock) {
+            try {
+                TaskOperation operation = inTransaction(() -> {
+                    TaskOperation existingOperation = taskOperationRepository.selectByGid(gid);
+                    Gallery existingGallery = galleryMapper.selectGalleryByGid(gid);
+                    if (existingOperation != null) {
+                        // 只有「删除已完成」的墓碑才允许用全新 operationId 重新创建（显式重建）。
+                        if (!TaskOperation.STATE_DELETED.equals(existingOperation.getState()))
+                            return null;
+                        if (existingGallery != null)
+                            return null;
+                        String operationId = TaskProtocol.newDownloadOperationId();
+                        taskOperationRepository.reactivate(gid, operationId, TaskOperation.STATE_RESERVED, now());
+                        insertPlaceholder(gid, placeholderName, downloaderId);
+                        return taskOperationRepository.selectByGid(gid);
+                    }
+                    if (existingGallery != null)
+                        return null;
+                    TaskOperation created = new TaskOperation();
+                    created.setGid(gid);
+                    created.setOperationId(TaskProtocol.newDownloadOperationId());
+                    created.setState(TaskOperation.STATE_RESERVED);
+                    created.setUpdateTime(now());
+                    taskOperationRepository.insert(created);
+                    insertPlaceholder(gid, placeholderName, downloaderId);
+                    return created;
+                });
+                return operation == null
+                        ? new Reservation(ReserveOutcome.DUPLICATE, null)
+                        : new Reservation(ReserveOutcome.RESERVED, operation);
+            } catch (RuntimeException e) {
+                // 数据库竞争/失败发生在任何外部副作用之前，绝不能穿透成 500，也不能继续提交。
+                log.warn("登记创建意图失败，按业务失败处理 gid={} error={}", gid, e.getClass().getSimpleName());
+                return new Reservation(ReserveOutcome.CONTENDED, null);
+            }
+        }
+    }
+
+    /** 写入占位任务行：让提交期间的任务在列表里可见。 */
+    private void insertPlaceholder(int gid, String name, int downloaderId) {
+        Gallery placeholder = new Gallery();
+        placeholder.setGid(gid);
+        placeholder.setName(name);
+        placeholder.setStatus(TaskStatus.PENDING_SUBMIT);
+        placeholder.setDownloader(downloaderId);
+        placeholder.setCreateTime(System.currentTimeMillis() / 1000);
+        placeholder.setProceeding(0);
+        galleryMapper.insertGallery(placeholder);
+    }
+
+    /**
+     * 标记为正在提交，并用 CAS 确认这一代意图仍然有效。
+     *
+     * <p>只有当前状态与 operationId 都匹配（RESERVED → SUBMITTING）才返回 true；
+     * 期间被删除或重建返回 false，调用方必须停下，绝不能再 POST 上游。
+     *
+     * @return true 表示已持久化 SUBMITTING，可以安全产生上游副作用
+     */
+    boolean markSubmitting(String operationId, int gid) {
+        return inTransaction(() -> taskOperationRepository.compareAndSetCreateState(gid, operationId,
+                TaskOperation.STATE_RESERVED, TaskOperation.STATE_SUBMITTING, now()) == 1);
+    }
+
+    /**
+     * 提交结果未知（超时/断连/中断）。保留意图，等待人工对账，不自动重发。
+     *
+     * <p>TaskOperation 与 Gallery 状态在同一个短事务里更新：
+     * 操作落 UNKNOWN 而任务行仍显示「提交中」会让运营侧误以为还能自动重试，
+     * 因此两边必须一致地表达「不确定，需要人工对账」。
+     */
+    void markUnknown(String operationId, int gid) {
+        inTransaction(() -> {
+            int advanced = taskOperationRepository.compareAndSetCreateState(gid, operationId,
+                    TaskOperation.STATE_SUBMITTING, TaskOperation.STATE_UNKNOWN, now());
+            // 只有本次这一代确实从 SUBMITTING 推进到 UNKNOWN 时才同步任务行：
+            // 若期间已被删除/重建（CAS 返回 0），绝不能再改那条记录。
+            if (advanced != 1)
+                return null;
+            Gallery gallery = galleryMapper.selectGalleryByGid(gid);
+            if (gallery != null && !TaskStatus.UNKNOWN.equals(gallery.getStatus()))
+                galleryMapper.updateGalleryStatus(gid, TaskStatus.UNKNOWN);
+            return null;
+        });
+    }
+
+    /**
+     * 提交成功后的「一次性短事务」：写入 Gallery 行、累计周用量、把操作推进为 SUBMITTED。
+     * 三者同一个事务，任何一个失败都一起回滚，避免出现「任务存在但用量未计」这类不一致。
+     */
+    /**
+     * 提交成功后的落库（带世代校验）。
+     *
+     * @param expectedOperationId 本次提交所归属的 operationId；只有它仍是当前有效意图时才能落库，
+     *                            否则（被删除/重建）必须抛错，避免复活已删 gid 或污染新世代
+     */
+    int persistSubmitted(Gallery gallery, int downloaderId, String expectedOperationId) {
+        return inTransaction(() -> {
+            TaskOperation operation = taskOperationRepository.selectByGid(gallery.getGid());
+            // 严格匹配本次世代：任务被删除、被重建或换了 operationId 时都不能落库。
+            if (operation == null || TaskOperation.STATE_DELETED.equals(operation.getState())
+                    || TaskOperation.STATE_DELETE_PENDING.equals(operation.getState()))
+                throw new IllegalStateException("任务已删除或没有有效创建意图，禁止落库复活 gid=" + gallery.getGid());
+            if (expectedOperationId == null || !expectedOperationId.equals(operation.getOperationId()))
+                throw new IllegalStateException("创建意图世代不匹配，禁止落库 gid=" + gallery.getGid());
+            // 幂等：同一意图重复落库（例如上游已提交后重跑）不得把用量累计两次。
+            boolean alreadyCounted = TaskOperation.STATE_SUBMITTED.equals(operation.getState());
+            gallery.setDownloader(downloaderId);
+            gallery.set_download(true);
+            if (!TaskStatus.SUBMITTED.equals(gallery.getStatus()))
+                gallery.setStatus(TaskStatus.SUBMITTED);
+            Gallery existing = galleryMapper.selectGalleryByGid(gallery.getGid());
+            if (existing == null)
+                galleryMapper.insertGallery(gallery);
+            else
+                galleryMapper.updateGallery(gallery);
+            if (!alreadyCounted)
+                configurationMapper.incrementConfiguration(CustomConfiguration.WEEK_USED_AMOUNT, gallery.getFileSize());
+            taskOperationRepository.updateState(gallery.getGid(), TaskOperation.STATE_SUBMITTED, now());
+            return 1;
+        });
+    }
+
+    /**
+     * 预占式创建：登记 gid 唯一意图（提交任务行占位）→ 只读解析上游 → 标 SUBMITTING → 真正提交。
+     *
+     * <p>上游提交发生在事务之外，且只在 SUBMITTING 已持久化之后执行一次。网络异常一律落 UNKNOWN。
+     */
+    public Response createTask(int gid, String placeholderName, String link, String targetResolution, int downloaderId) {
+        Response response = Response.generateResponse();
+        Reservation reservation = reserveCreateIntent(gid, placeholderName, downloaderId);
+        if (reservation.outcome() == ReserveOutcome.DUPLICATE) {
+            response.failure("任务队列已有此任务，请点击查找任务");
+            return response;
+        }
+        if (reservation.outcome() == ReserveOutcome.CONTENDED) {
+            // 预占阶段的数据库竞争不产生任何副作用；转业务失败，绝不下发上游、绝不穿透成 500。
+            response.failure("任务提交失败，请稍后重试");
+            return response;
+        }
+
+        // 解析上游是只读步骤；真正的 POST 只在 SUBMITTING 持久化之后发生。
+        // 生命周期约束：RESERVED 阶段被中断是安全的（可释放重来），SUBMITTING 之后一律按 UNKNOWN 处理。
+        return submitPreparedIntent(gid, reservation.operation(), link, targetResolution, downloaderId);
+    }
+
+    private Response submitPreparedIntent(int gid, TaskOperation operation, String link, String targetResolution,
+                                          int downloaderId) {
+        Response response = Response.generateResponse();
+        beforePrepare.run();
+        PreparedGallery prepared;
+        try {
+            prepared = upstreamGateway.prepare(link);
+        } catch (IOException e) {
+            log.warn("解析任务链接失败 gid={}", gid, e);
+            releaseReservation(gid, operation.getOperationId());
+            response.failure("IO错误，可能是网络波动");
+            return response;
+        } catch (RuntimeException e) {
+            log.warn("解析任务链接异常 gid={}", gid, e);
+            releaseReservation(gid, operation.getOperationId());
+            response.failure("提交任务失败，未知原因，请检查链接是否正常");
+            return response;
+        }
+        if (prepared == null || prepared.gallery() == null) {
+            releaseReservation(gid, operation.getOperationId());
+            response.failure("提交任务失败，未知原因，请检查链接是否正常");
+            return response;
+        }
+
+        Gallery gallery = prepared.gallery();
+        gallery.setGid(gid);
+
+        // CAS 确认这一代意图仍然有效（未被并发删除/重建），并把状态持久化为 SUBMITTING。
+        // 只有成功之后才允许真正 POST；失败说明任务已不在当前生命周期里，直接放弃。
+        if (!markSubmitting(operation.getOperationId(), gid)) {
+            log.warn("创建意图已失效（可能被并发删除），放弃提交 gid={} operationId={}",
+                    gid, operation.getOperationId());
+            response.failure("任务已被删除或状态已变化，请刷新任务列表");
+            return response;
+        }
+        afterMarkSubmitting.run();
+
+        Gallery submitted;
+        try {
+            submitted = upstreamGateway.submit(prepared, targetResolution);
+        } catch (ResolutionNotMatchException e) {
+            releaseReservation(gid, operation.getOperationId());
+            response.failure("提交失败，分辨率不存在");
+            return response;
+        } catch (IOException | RuntimeException e) {
+            // 超时或断连时上游可能已经受理，无法确认：只能记 UNKNOWN，绝不自动重发。
+            log.warn("上游提交结果未知 gid={} operationId={}", gid, operation.getOperationId(), e);
+            markUnknown(operation.getOperationId(), gid);
+            response.failure("上游提交结果未知，请稍后在任务列表中确认，不要重复提交");
+            return response;
+        }
+        if (submitted == null || !TaskStatus.SUBMITTED.equals(submitted.getStatus())) {
+            markUnknown(operation.getOperationId(), gid);
+            response.failure("上游提交结果未知，请稍后在任务列表中确认，不要重复提交");
+            return response;
+        }
+        submitted.setGid(gid);
+        try {
+            persistSubmitted(submitted, downloaderId, operation.getOperationId());
+        } catch (RuntimeException e) {
+            // 上游已受理但本地落库失败：保留 SUBMITTING 意图并标 UNKNOWN，任务行仍在列表里可对账。
+            log.error("提交成功后落库失败 gid={} operationId={}", gid, operation.getOperationId(), e);
+            markUnknown(operation.getOperationId(), gid);
+            response.failure("任务已提交，但本地保存失败，请在任务列表中确认");
+            return response;
+        }
+        // 只有事务提交之后才向节点下发命令；重复下发由节点按 operationId 去重。
+        byte dispatchResult = dispatchDownload(gid, submitted.getName());
+        Gallery current = galleryMapper.selectGalleryByGid(gid);
+        Gallery payload = current == null ? submitted : current;
+        if (dispatchResult == 0)
+            response.success(objectMapper.valueToTree(payload));
+        else
+            response.failure("任务已保存，但节点未确认接收；请刷新任务列表后重试");
+        return response;
+    }
+
+    /**
+     * 提交成功后的落库 + 下发。落库（含用量）在一个短事务里完成，事务提交后才向节点发命令。
+     *
+     * <p>先解析当前世代再落库：世代不存在（未提交过 / 已删除）时不落库也不下发。
+     *
+     * @return 已持久化的 Gallery；数据库失败或世代无效时抛出异常，调用方回业务失败且不下发节点。
+     */
+    public Gallery persistAndDispatch(Gallery gallery, int downloaderId) {
+        DispatchDecision decision = resolveDispatch(gallery.getGid());
+        if (!decision.dispatchable())
+            throw new IllegalStateException("任务没有可下发的有效世代 gid=" + gallery.getGid());
+        persistSubmitted(gallery, downloaderId, decision.operationId());
+        dispatchDownload(gallery.getGid(), gallery.getName());
+        return gallery;
+    }
+
+    /** 向节点下发 DOWNLOAD 命令（复用当前有效 operationId），返回节点结果码。 */
+    public byte dispatchDownload(int gid, String name) {
+        DispatchDecision decision = resolveDispatch(gid);
+        if (!decision.dispatchable())
+            return -1;
+        return remoteService.sendTaskCommand(TaskCommandMessage.ACTION_DOWNLOAD, decision.operationId(), gid, name);
+    }
+
+    /**
+     * 判定一个 gid 能否下发给节点，并为迁移前的旧任务补齐稳定 operationId。
+     *
+     * <p>规则：
+     * <ul>
+     *   <li>v2 任务只有 {@code SUBMITTED} 可下发；{@code UNKNOWN} 必须人工对账，
+     *       否则未计量的未知任务会在节点侧变成「下载完成却 0 额度」；</li>
+     *   <li>预占中的 {@code RESERVED}/{@code SUBMITTING}、待删除与删除墓碑都不下发；</li>
+     *   <li>没有任何操作记录、但任务行确实存在且状态已是「已提交」及之后的历史任务，
+     *       补建一个稳定 operationId 后再下发（这是唯一的 legacy 兼容路径）。</li>
+     * </ul>
+     */
+    public DispatchDecision resolveDispatch(int gid) {
+        TaskOperation operation = taskOperationRepository.selectByGid(gid);
+        if (operation != null && TaskOperation.STATE_SUBMITTED.equals(operation.getState())
+                && operation.getOperationId() != null)
+            return new DispatchDecision(DispatchDecision.Kind.READY, operation.getOperationId());
+        // 没有操作记录，或只有「旧世代」记录（缺 operationId）：必须先确认任务行确实已提交过，
+        // 再补一个稳定 operationId；UNKNOWN/预占/待删除一律拒绝。
+        if (operation != null && !TaskOperation.STATE_SUBMITTED.equals(operation.getState()))
+            return REFUSE;
+        if (!isLegacySubmitted(gid))
+            return REFUSE;
+        String freshId = ensureStableOperationId(gid);
+        return freshId == null ? REFUSE : new DispatchDecision(DispatchDecision.Kind.LEGACY, freshId);
+    }
+
+    /** 迁移前的旧任务行：没有操作记录，但状态表明上游确实已受理过下载，可安全重投。 */
+    private boolean isLegacySubmitted(int gid) {
+        Gallery gallery = galleryMapper.selectGalleryByGid(gid);
+        if (gallery == null || gallery.getStatus() == null)
+            return false;
+        return switch (gallery.getStatus()) {
+            case TaskStatus.SUBMITTED, TaskStatus.DOWNLOADING, TaskStatus.WAITING_COMPRESS,
+                 TaskStatus.COMPRESSING -> true;
+            // 提交未知 / 提交中 / 待删除 / 删除失败 一律不算已提交，不能下发。
+            default -> false;
+        };
+    }
+
+    /**
+     * 为一个已经真正提交过的旧任务持久化稳定 operationId。
+     *
+     * <p>必须是短事务写入、先落库后返回，之后重投/重连才复用同一个 ID；
+     * 用 {@code insert or ignore} + 回读，避免与并发创建互相覆盖。
+     *
+     * @return 稳定 operationId；无法安全建立时返回 null（调用方不下发）
+     */
+    String ensureStableOperationId(int gid) {
+        // 与创建预占共用同一把短锁（锁内只有数据库写），并发为同一 gid 补 ID 时只能有一个赢家。
+        synchronized (reservationLock) {
+            String candidate = TaskProtocol.newDownloadOperationId();
+            return inTransaction(() -> {
+                if (taskOperationRepository.selectByGid(gid) == null)
+                    taskOperationRepository.insertIfAbsent(gid, candidate, TaskOperation.STATE_SUBMITTED, now());
+                else
+                    taskOperationRepository.assignOperationIdIfMissing(gid, candidate, now());
+                TaskOperation current = taskOperationRepository.selectByGid(gid);
+                if (current == null || current.getOperationId() == null)
+                    return null;
+                // 只有确实处于 SUBMITTED 的世代才可下发；并发写入的其他状态一律拒绝。
+                return TaskOperation.STATE_SUBMITTED.equals(current.getState()) ? current.getOperationId() : null;
+            });
+        }
+    }
+
+    /**
+     * 取消预占（解析失败、分辨率不存在等：还没有产生任何上游副作用）。
+     *
+     * <p>只删除与本次创建 intent 匹配的 RESERVED 记录：若期间该 gid 已被删除并重建
+     * （新的 operationId / 墓碑），这里绝不能把新世代记录抹掉。占位行同理只删「提交中」。
+     */
+    private void releaseReservation(int gid, String operationId) {
+        inTransaction(() -> {
+            // 只删与本代 operationId 匹配且仍停在 RESERVED/SUBMITTING 的记录。
+            int removed = taskOperationRepository.deleteReservationIfSameGeneration(gid, operationId);
+            // 只有确实删掉了本代预占记录，才能删对应的占位行：
+            // 若期间已出现新的 RESERVED 代，删除返回 0，此时绝不能误删新占位。
+            if (removed == 1) {
+                Gallery existing = galleryMapper.selectGalleryByGid(gid);
+                if (existing != null && TaskStatus.PENDING_SUBMIT.equals(existing.getStatus()))
+                    galleryMapper.deleteGalleryByGid(gid);
+            }
+            return null;
+        });
+    }
+
+    /**
+     * 删除的持久化部分。
+     *
+     * <p>正在创建意图（RESERVED/SUBMITTING）的任务不能删除：此刻 DELETE 与创建 POST 会同时发生，
+     * 随后创建落库还会把已删/重建的 gid 写回。这里直接拒绝，由用户在提交结果明确后再操作。
+     *
+     * @return 显式结果：被拒时调用方必须回业务失败，不能退避到裸删除
+     */
+    public DeleteIntent beginDelete(int gid) {
+        return inTransaction(() -> {
+            Gallery gallery = galleryMapper.selectGalleryByGid(gid);
+            if (gallery == null)
+                return DeleteIntent.refused("删除失败，该图片不存在");
+            TaskOperation operation = taskOperationRepository.selectByGid(gid);
+            if (operation != null && (TaskOperation.STATE_RESERVED.equals(operation.getState())
+                    || TaskOperation.STATE_SUBMITTING.equals(operation.getState())))
+                return DeleteIntent.refused("任务正在提交中，请等待提交结果明确后再删除");
+            if (operation == null) {
+                // 历史任务行没有操作记录：直接落一条 DELETE_PENDING 意图，避免中间态缺字段。
+                TaskOperation fresh = new TaskOperation();
+                fresh.setGid(gid);
+                fresh.setOperationId(TaskProtocol.newDownloadOperationId());
+                fresh.setDeleteOperationId(TaskProtocol.newDeleteOperationId());
+                fresh.setState(TaskOperation.STATE_DELETE_PENDING);
+                fresh.setUpdateTime(now());
+                taskOperationRepository.insert(fresh);
+                galleryMapper.updateGalleryStatus(gid, TaskStatus.DELETE_PENDING);
+                return DeleteIntent.accepted(fresh.getDeleteOperationId());
+            }
+            if (operation.getDeleteOperationId() == null) {
+                operation.setDeleteOperationId(TaskProtocol.newDeleteOperationId());
+                operation.setUpdateTime(now());
+            }
+            taskOperationRepository.updateDeleteIntent(gid, operation.getDeleteOperationId(),
+                    TaskOperation.STATE_DELETE_PENDING, now());
+            galleryMapper.updateGalleryStatus(gid, TaskStatus.DELETE_PENDING);
+            return DeleteIntent.accepted(operation.getDeleteOperationId());
+        });
+    }
+
+    /**
+     * 删除的完成部分：节点已确认（或文件已不存在）后，在一个短事务里删任务行、收藏并留墓碑。
+     *
+     * <p>事务提交前若失败，任务行与 DELETE_PENDING 意图都保留，可以继续重试。
+     * 删除意图 ID 必须匹配（防止并发删除/重建时抹掉别的一代），随后只把 DELETE_PENDING
+     * 世代推进为 DELETED；墓碑（DELETED 且无删除意图）不允许在这里被重新激活。
+     *
+     * @return true 表示本次删除意图已完成并留下墓碑
+     */
+    public boolean completeDelete(int gid, String deleteOperationId) {
+        return inTransaction(() -> {
+            TaskOperation operation = taskOperationRepository.selectByGid(gid);
+            if (operation == null || operation.getDeleteOperationId() == null
+                    || !operation.getDeleteOperationId().equals(deleteOperationId))
+                return false;
+            // 先清收藏再删任务行：收藏一侧在部分历史库上带有指向 gallery 的外键。
+            collectMapper.deleteByGid(gid);
+            galleryMapper.deleteGalleryByGid(gid);
+            int updated = taskOperationRepository.completeDeleteIfSameGeneration(gid, deleteOperationId,
+                    TaskOperation.STATE_DELETED, now());
+            return updated == 1;
+        });
+    }
+
+    /** 删除失败：保留任务行并记一条可重试的失败状态。 */
+    public void failDelete(int gid) {
+        inTransaction(() -> {
+            Gallery gallery = galleryMapper.selectGalleryByGid(gid);
+            if (gallery == null)
+                return null;
+            gallery.setStatus(TaskStatus.DELETE_FAILED);
+            galleryMapper.updateGallery(gallery);
+            return null;
+        });
+    }
+
+    /**
+     * 启动恢复：把上次中断在 SUBMITTING 的意图判为 UNKNOWN，并安全释放还停在 RESERVED 的预占。
+     *
+     * <p>这段恢复必须发生在任何自动重投之前：
+     * <ul>
+     *   <li>{@code SUBMITTING}：上游可能已经受理但本地无法确认，只能落 UNKNOWN（操作与任务行同时
+     *       更新为「提交未知」），交人工对账，绝不自动再 POST；</li>
+     *   <li>{@code RESERVED}：还没有产生任何上游副作用，重启后可以直接释放预占
+     *       （删掉匹配的 RESERVED 记录与「提交中」占位行），让用户重新提交。</li>
+     * </ul>
+     *
+     * @return 本次恢复（判未知 + 释放预占）处理的任务数
+     */
+    public int recoverInterruptedSubmissions() {
+        return inTransaction(() -> {
+            int recovered = 0;
+            for (TaskOperation operation : taskOperationRepository.selectByState(TaskOperation.STATE_SUBMITTING)) {
+                int advanced = taskOperationRepository.compareAndSetCreateState(operation.getGid(), operation.getOperationId(),
+                        TaskOperation.STATE_SUBMITTING, TaskOperation.STATE_UNKNOWN, now());
+                if (advanced != 1)
+                    continue;
+                Gallery gallery = galleryMapper.selectGalleryByGid(operation.getGid());
+                if (gallery != null && !TaskStatus.UNKNOWN.equals(gallery.getStatus()))
+                    galleryMapper.updateGalleryStatus(operation.getGid(), TaskStatus.UNKNOWN);
+                recovered++;
+            }
+            for (TaskOperation operation : taskOperationRepository.selectByState(TaskOperation.STATE_RESERVED)) {
+                // 还没 POST：释放匹配的预占记录与占位行，允许重启后重新提交。
+                // 只有确实删掉了这条本代 RESERVED 记录，才删其占位行；否则不动（可能已换代）。
+                int removed = taskOperationRepository.deleteReservationIfSameGeneration(operation.getGid(),
+                        operation.getOperationId());
+                if (removed == 1) {
+                    Gallery gallery = galleryMapper.selectGalleryByGid(operation.getGid());
+                    if (gallery != null && TaskStatus.PENDING_SUBMIT.equals(gallery.getStatus()))
+                        galleryMapper.deleteGalleryByGid(operation.getGid());
+                }
+                recovered++;
+            }
+            return recovered;
+        });
+    }
+
+    /** 已完成状态上报的处理：丢弃旧 revision、旧 operation、待删除任务与重复上报。 */
+    @Override
+    public CommittedReport apply(TaskReportMessage report) {
+        if (report == null || report.getReportId() == null || report.getGalleryTasks() == null)
+            return null;
+        Applied applied;
+        try {
+            applied = inTransaction(() -> applyInTransaction(report));
+        } catch (RuntimeException e) {
+            // 数据库失败：不返回可回执结果，节点会保留快照并稍后重投；失败写入绝不清任务。
+            log.warn("应用节点上报失败，未回执 reportId={}", report.getReportId(), e);
+            return null;
+        }
+        if (applied == null)
+            return null;
+        return new CommittedReport(report.getReportId(), applied.newlyCompleted(), applied.acceptedTasks());
+    }
+
+    /** 事务内应用上报的结果：新完成的任务与真正被接受的进度更新。 */
+    private record Applied(List<Gallery> newlyCompleted, List<GalleryTask> acceptedTasks) {
+    }
+
+    /**
+     * 在一个事务里应用上报。
+     *
+     * @return 非 null 表示已经提交（可以回 ACK，重复上报也会返回结果以停止重放）。
+     */
+    private Applied applyInTransaction(TaskReportMessage report) {
+        long revision = report.getRevision();
+        List<Gallery> newlyCompleted = new ArrayList<>();
+        List<GalleryTask> acceptedTasks = new ArrayList<>();
+        for (GalleryTask task : report.getGalleryTasks()) {
+            if (task == null)
+                continue;
+            TaskOperation operation = taskOperationRepository.selectByGid(task.getGid());
+            Gallery gallery = galleryMapper.selectGalleryByGid(task.getGid());
+            if (gallery == null) {
+                // 未知 gid（任务已删除）：忽略但 ACK，让节点停止重放。
+                log.debug("忽略节点上报的未知任务 gid={}", task.getGid());
+                continue;
+            }
+            String expectedOperationId = report.getOperationIds() == null
+                    ? null : report.getOperationIds().get(task.getGid());
+            if (operation == null) {
+                // 旧任务行（迁移前创建）没有操作记录：按显式生成策略补建并采纳上报里的 operationId。
+                operation = adoptGeneration(task.getGid(), expectedOperationId);
+            } else if (TaskOperation.STATE_DELETED.equals(operation.getState())) {
+                // 墓碑必须永久忽略旧上报：删除已完成的任务只能由显式 CREATE 重新建立新世代。
+                // （gallery 已被删除时上面的 null 分支已经提前忽略，这里防御性兜底。）
+                log.debug("忽略已删除 gid 的迟到上报 gid={}", task.getGid());
+                continue;
+            } else if (TaskOperation.STATE_DELETE_PENDING.equals(operation.getState())) {
+                log.info("忽略待删除任务的迟到上报 gid={}", task.getGid());
+                continue;
+            } else if (operation.getLastRevision() > 0 && revision <= operation.getLastRevision()) {
+                log.debug("丢弃过期上报 gid={} revision={} lastRevision={}",
+                        task.getGid(), revision, operation.getLastRevision());
+                continue;
+            } else if (operation.getOperationId() != null) {
+                // v2 任务必须有归属：缺 operationId 或与当前世代不符都不能改当前任务，
+                // 否则「删除后重建」的新世代会被旧上报污染。
+                if (expectedOperationId == null || !operation.getOperationId().equals(expectedOperationId)) {
+                    log.info("丢弃无法归属/旧世代的迟到上报 gid={} 上报={} 当前={}",
+                            task.getGid(), expectedOperationId, operation.getOperationId());
+                    continue;
+                }
+            }
+
+            String status = TaskStatus.fromReport(task.getStatus());
+            if (status == null) {
+                log.warn("未知上报状态码 gid={} status={}", task.getGid(), task.getStatus());
+                continue;
+            }
+            gallery.setProceeding(task.getProceeding());
+            if (task.getName() != null)
+                gallery.setName(task.getName());
+            boolean justCompleted = TaskStatus.COMPLETE.equals(status)
+                    && !TaskStatus.COMPLETE.equals(gallery.getStatus());
+            gallery.setStatus(status);
+            galleryMapper.updateGallery(gallery);
+            taskOperationRepository.updateReportReceipt(task.getGid(), revision, report.getReportId(), now());
+            acceptedTasks.add(task);
+            if (justCompleted)
+                newlyCompleted.add(gallery);
+        }
+        return new Applied(newlyCompleted, acceptedTasks);
+    }
+
+    /**
+     * 为迁移前创建、尚无操作记录的历史任务建立一次「世代」：优先采纳上报里的 operationId；
+     * 旧节点不带 operationId 时保留 null，让后续同世代的进度上报仍能被接受。
+     *
+     * <p>不在这里凭空生成随机 ID：那会让第一份进度上报之后的所有旧上报都因「无法归属」被丢弃。
+     * 真正的稳定 operationId 会在首次自动下发时用 {@code ensureStableOperationId} 落库，
+     * 之后该世代即可被严格匹配。不会覆盖任何已有任务状态。
+     */
+    private TaskOperation adoptGeneration(int gid, String reportedOperationId) {
+        TaskOperation operation = new TaskOperation();
+        operation.setGid(gid);
+        operation.setOperationId(reportedOperationId);
+        operation.setState(TaskOperation.STATE_SUBMITTED);
+        operation.setUpdateTime(now());
+        taskOperationRepository.insert(operation);
+        return operation;
+    }
+
+    private static long now() {
+        return System.currentTimeMillis() / 1000;
+    }
+
+    /** 供装配层解析当前有效 operationId（避免装配代码访问包私有字段）。 */
+    public String currentOperationId(int gid) {
+        TaskOperation operation = taskOperationRepository.selectByGid(gid);
+        return operation == null ? null : operation.getOperationId();
+    }
+}

src/main/java/com/lion/lionwebsite/Service/QueryService.java

@@ -54,35 +54,16 @@ public class QueryService {
         }
 
         Document parse = Jsoup.parse(result);
-        Elements elements = parse.select("body > div.ido > div:nth-child(2) > table > tbody > tr");
-
-        ArrayList<GalleryForQuery> galleries = new ArrayList<>();
-        ListIterator<Element> elementListIterator = elements.listIterator();
+        java.util.List<GalleryForQuery> galleries;
         try {
-            elementListIterator.next();                                             //skip firstlink
-        }catch (NullPointerException | NoSuchElementException e){
-            response.failure("没有搜索到结果");
-            return response.toJSONString();
+            galleries = com.lion.lionwebsite.Util.GallerySearchParser.parse(parse);
+        } catch (IOException | RuntimeException error) {
+            log.warn("上游搜索页面格式不支持: {}", error.getClass().getSimpleName());
+            return response.failure("上游搜索页面格式不支持").toJSONString();
         }
+        if (galleries.isEmpty()) return response.failure("没有搜索到结果").toJSONString();
 
-        while (elementListIterator.hasNext()){
-            Element element = elementListIterator.next();
-            GalleryForQuery gallery = new GalleryForQuery();
-            gallery.setType(element.child(0).text());                                                      //type
-            String src = element.child(1).select("img").attr("src");
-            if (src.startsWith("data"))
-                src = element.child(1).select("img").attr("data-src");
-            src = src.replace("https://s.exhentai.org", "");
-            gallery.setThumbnailUrl(src);                                 //thumbnailSrc
-            gallery.setUploadTime(element.child(1).select("div > div [onclick]").get(1).text());          //uploadTime
-            gallery.setLink(element.child(2).child(0).attr("href"));                         //link
-            gallery.setGid(GalleryUtil.parseGid(gallery.getLink()) + "");                                           //gid
-            gallery.setName(element.child(2).child(0).child(0).text());                           //name
-            gallery.setPage(Integer.parseInt(element.child(3).child(1).text().split(" ")[0]));   //page
-            galleries.add(gallery);
-        }
-
-        response.success(objectMapper.valueToTree(galleries).toString());
+        response.success(objectMapper.valueToTree(galleries));
         Elements nextLink = parse.select("#unext");
         if(nextLink.hasAttr("href"))
             response.set("next", nextLink.attr("href"));

src/main/java/com/lion/lionwebsite/Service/RemoteService.java

@@ -37,6 +37,7 @@ import java.util.concurrent.TimeUnit;
 import java.util.concurrent.TimeoutException;
 import java.util.concurrent.atomic.AtomicInteger;
 import java.util.concurrent.atomic.AtomicBoolean;
+import java.util.List;
 
 @Service
 @Data
@@ -76,12 +77,131 @@ public class RemoteService {
 
     final SubscriptionStandbySnapshotService subscriptionStandbySnapshotService;
 
+    /**
+     * 协议 v2 上报处理器。用 setter 注入而不是构造参数，避免 RemoteService 与
+     * GalleryTaskLifecycleService 之间的依赖环（后者需要 RemoteService 下发命令）。
+     */
+    volatile TaskReportHandler taskReportHandler;
+
+    /**
+     * gid → 当前有效的下载 operationId。由生命周期服务在装配期注册（保持依赖方向单向）。
+     * 未注册或返回 null 时，重投/重试退化为旧的无 operationId 命令。
+     */
+    volatile java.util.function.IntFunction<String> downloadOperationIdResolver;
+
+    /**
+     * gid → 下发判定（是否允许下发 + 用哪个稳定 operationId）。
+     *
+     * <p>替代「能解析出 operationId 就下发」的旧语义：UNKNOWN / 预占中 / 待删除的任务
+     * 必须被明确拒绝，否则未计量的人工对账任务可能在节点侧变成「下载完成却 0 额度」。
+     */
+    volatile java.util.function.IntFunction<GalleryTaskLifecycleService.DispatchDecision> dispatchResolver;
+
+    /**
+     * 生命周期装配是否已经完成（handler/resolver 已绑定 + 启动恢复已执行）。
+     *
+     * <p>在它为 true 之前，绝不自动重投未完成任务：启动早期还不知道哪些任务是
+     * SUBMITTED、哪些是 UNKNOWN，盲目下发会制造未计量副作用。
+     */
+    volatile boolean lifecycleReady;
+
+    /**
+     * 启动恢复是否安全完成。恢复失败时拒绝自动重投，避免在状态不明的库上继续派发。
+     */
+    volatile boolean recoverySafe;
+
+    /**
+     * 节点是否声明了 v2 可靠任务能力。只有协商成功之后才发送类型 11..13，
+     * 否则继续走旧的无确认上报，保持向后兼容。
+     */
+    volatile boolean v2Negotiated;
+
+    /** 本次连接的 Hello 协商结果；收到节点 Hello（或超时）后完成。 */
+    volatile CompletableFuture<Boolean> helloNegotiation;
+
+    /** 明确「未下发」的结果码：与 -1（发送失败/超时）区分开。 */
+    public static final byte REFUSED = -3;
+
+    /** 事务提交之后的通知/推送执行器：外部延迟（Telegram、慢浏览器）不得阻塞状态提交。 */
+    final ExecutorService notificationExecutor = Executors.newSingleThreadExecutor(r -> {
+        Thread thread = new Thread(r, "gallery-notify");
+        thread.setDaemon(true);
+        return thread;
+    });
+
+    /**
+     * 协议握手的专用线程。hello 等待响应是阻塞调用，绝不能放在 Netty 事件循环上，
+     * 否则响应本身要由同一个事件循环投递，会直接死锁。
+     */
+    final ExecutorService handshakeExecutor = Executors.newSingleThreadExecutor(r -> {
+        Thread thread = new Thread(r, "node-protocol-handshake");
+        thread.setDaemon(true);
+        return thread;
+    });
+
     final ExecutorService subscriptionSyncExecutor = Executors.newSingleThreadExecutor(r -> {
         Thread thread = new Thread(r, "subscription-standby-sync");
         thread.setDaemon(true);
         return thread;
     });
 
+    /** 测试/装配用：绑定上报处理器。 */
+    public void setTaskReportHandler(TaskReportHandler taskReportHandler) {
+        this.taskReportHandler = taskReportHandler;
+    }
+
+    public void setDownloadOperationIdResolver(java.util.function.IntFunction<String> resolver) {
+        this.downloadOperationIdResolver = resolver;
+    }
+
+    public void setDispatchResolver(java.util.function.IntFunction<GalleryTaskLifecycleService.DispatchDecision> resolver) {
+        this.dispatchResolver = resolver;
+    }
+
+    /**
+     * 由装配层在「绑定 handler/resolver + 完成启动恢复」之后调用：标记可以安全派发并建立连接。
+     *
+     * <p>初始连接不再放在 {@code @PostConstruct}：那时 handler/resolver 都还没绑定，
+     * 连接成功会把尚未确认的 UNKNOWN/预占任务当作普通任务重投。
+     *
+     * @param recoverySafe 启动恢复是否成功；false 时只允许连接，不允许自动重投
+     */
+    public void startAfterWiring(boolean recoverySafe) {
+        this.recoverySafe = recoverySafe;
+        this.lifecycleReady = true;
+        initChannel();
+    }
+
+    /** 把 gid 解析成下发判定；未注册解析器或解析失败时按「拒绝下发」处理。 */
+    private GalleryTaskLifecycleService.DispatchDecision resolveDispatchDecision(int gid) {
+        java.util.function.IntFunction<GalleryTaskLifecycleService.DispatchDecision> resolver = dispatchResolver;
+        if (resolver == null)
+            return null;
+        try {
+            return resolver.apply(gid);
+        } catch (Exception e) {
+            log.warn("解析任务下发判定失败 gid={}", gid, e);
+            return null;
+        }
+    }
+
+    private String resolveDownloadOperationId(int gid) {
+        java.util.function.IntFunction<String> resolver = downloadOperationIdResolver;
+        if (resolver == null)
+            return null;
+        try {
+            return resolver.apply(gid);
+        } catch (Exception e) {
+            log.warn("解析任务 operationId 失败 gid={}", gid, e);
+            return null;
+        }
+    }
+
+    /** 节点是否已完成 v2 能力协商（供健康检查/测试观察）。 */
+    public boolean isReliableProtocolReady() {
+        return v2Negotiated;
+    }
+
     /**
      * 处理节点上报的任务状态。原先这些 JDBC 读写直接跑在 Netty 的 IO 线程上，
      * N 条上报会阻塞该 IO 线程，连带拖慢心跳与请求响应；改为单线程顺序执行，
@@ -120,10 +240,14 @@ public class RemoteService {
     @Value("${subscription.standby.sync-enabled:false}")
     boolean subscriptionSyncEnabled;
 
-    @PostConstruct
-    void init() {
-        initChannel();
-    }
+    /**
+     * 这里刻意不再自动连接。
+     *
+     * <p>原实现在 PostConstruct 里立即 initChannel()，而那一刻 TaskLifecycleWiring
+     * 还没绑定上报处理器、operationId 解析器，也没执行启动恢复：连接成功后会马上把
+     * 未完成任务（含 UNKNOWN/预占）当作普通任务重投给节点。
+     * 现在改为由装配层在完成绑定与恢复后调用 startAfterWiring(boolean)。
+     */
 
     public boolean initChannel(){
         if (stopping || !connecting.compareAndSet(false, true))
@@ -175,12 +299,32 @@ public class RemoteService {
             channel = channelFuture.channel();
             closeMonitorSocket();
             channel.writeAndFlush(new IdentityMessage("lionwebsite"));
-
-            //子节点上线时，发送未完成的任务
-            resetUndone();
-            // 节点刚上线时可能没有任何快照（例如刚重启），必须推一次；
-            // 这同时会刷新节点「最近收到快照」的时刻，避免内容未变时被判过期。
-            requestSubscriptionSync();
+            // 协议 v2：先发旧的 Identity，再发 hello 做能力协商。
+            // 协商完成之前不允许下发任何任务，也不允许自动重投未完成任务。
+            CompletableFuture<Boolean> negotiation = queueHelloHandshake();
+            // 在协商线程上等待结果，再决定重投；绝不阻塞 Netty 事件循环。
+            handshakeExecutor.execute(() -> {
+                boolean reliable = awaitNegotiation(negotiation);
+                // 订阅快照与任务重投都要求生命周期装配已完成；装配前只保持连接。
+                requestSubscriptionSync();
+                if (!reliable) {
+                    log.info("节点未确认 v2 能力，跳过自动重投未完成任务");
+                    return;
+                }
+                if (!lifecycleReady) {
+                    log.info("生命周期装配未就绪，跳过自动重投未完成任务");
+                    return;
+                }
+                if (!recoverySafe) {
+                    log.error("启动恢复未安全完成，拒绝自动重投未完成任务");
+                    return;
+                }
+                try {
+                    resetUndone();
+                } catch (RuntimeException e) {
+                    log.warn("重投未完成任务失败", e);
+                }
+            });
             return true;
         }catch (Exception e){
             log.error("connect node failed, wait for node back online", e);
@@ -297,6 +441,8 @@ public class RemoteService {
         failPendingRequests();
         if (channel != null) channel.close();
         statusApplyExecutor.shutdownNow();
+        notificationExecutor.shutdownNow();
+        handshakeExecutor.shutdownNow();
         subscriptionSyncExecutor.shutdownNow();
         networkGroup.shutdownGracefully();
         eventLoopGroup.shutdownGracefully();
@@ -315,7 +461,11 @@ public class RemoteService {
     }
 
     public void resetUndone(){
-        if (channelFuture.channel() == null || !channelFuture.channel().isActive())
+        // 只有「连接在线 + v2 已协商 + 生命周期装配完成 + 启动恢复安全完成」才允许批量重投；
+        // 任何一项不满足都说明还不知道任务真实状态，盲目下发会制造未计量副作用。
+        if (channelFuture == null || channelFuture.channel() == null || !channelFuture.channel().isActive())
+            return;
+        if (!v2Negotiated || !lifecycleReady || !recoverySafe)
             return;
 
         Gallery[] galleries = galleryMapper.selectUnDoneGalleries();
@@ -328,15 +478,135 @@ public class RemoteService {
     }
 
     public byte addGalleryToQueue(Gallery gallery){
+        GalleryTaskLifecycleService.DispatchDecision decision = resolveDispatchDecision(gallery.getGid());
+        if (decision != null) {
+            // 明确拒绝（UNKNOWN/预占/待删除等）：绝不能退化为无 operationId 的旧命令。
+            if (!decision.dispatchable())
+                return REFUSED;
+            return sendTaskCommand(TaskCommandMessage.ACTION_DOWNLOAD, decision.operationId(),
+                    gallery.getGid(), gallery.getName());
+        }
+        // 兼容：装配层未注册判定解析器时（如单元测试），退回只按 operationId 解析的旧行为。
+        String operationId = resolveDownloadOperationId(gallery.getGid());
+        if (operationId != null)
+            return sendTaskCommand(TaskCommandMessage.ACTION_DOWNLOAD, operationId, gallery.getGid(), gallery.getName());
+        return legacyDownloadCommand(gallery.getGid(), gallery.getName());
+    }
+
+    /** 旧的无 operationId 下载命令，供未启用 v2 或无法解析 operationId 时退避使用。 */
+    private byte legacyDownloadCommand(int gid, String name) {
         GalleryTask galleryTask = new GalleryTask();
-        galleryTask.setGid(gallery.getGid());
-        galleryTask.setName(gallery.getName());
+        galleryTask.setGid(gid);
+        galleryTask.setName(name);
         DownloadPostMessage message = new DownloadPostMessage();
         message.setGalleryTask(galleryTask);
         return sendRequest(message, 10, TimeUnit.SECONDS);
     }
 
+    /**
+     * 在专用线程上执行 hello 协商，并返回一个在「收到节点 Hello」或「等待超时」后完成的 future。
+     *
+     * <p>节点的应答是 Hello 而不是 ResponseMessage，因此绝不能走 sendRequest 的响应等待器
+     * （那会白等 10 秒超时，排在后面的自动重投也只能 10 秒后才开始）。
+     * 这里直接 writeAndFlush，并让 channelRead 的 Hello 分支完成 future；
+     * 超时用 completeOnTimeout 处理，写失败立刻完成。
+     */
+    private CompletableFuture<Boolean> queueHelloHandshake() {
+        // 10 秒内没等到节点 Hello 就按「未协商」完成，避免后续重投被无限期挂起。
+        CompletableFuture<Boolean> result = new CompletableFuture<Boolean>()
+                .completeOnTimeout(false, 10, TimeUnit.SECONDS);
+        helloNegotiation = result;
+        if (stopping) {
+            result.complete(false);
+            return result;
+        }
+        try {
+            handshakeExecutor.execute(() -> {
+                v2Negotiated = false;
+                try {
+                    Channel target = channel;
+                    if (target == null || !target.isActive()) {
+                        result.complete(false);
+                        return;
+                    }
+                    // Hello 不写入 promiseHashMap：节点回的是 Hello，没有 ResponseMessage 等待者。
+                    target.writeAndFlush(new ProtocolHelloMessage()).addListener(future -> {
+                        if (!future.isSuccess()) {
+                            if (!stopping)
+                                log.warn("协议协商发送失败: {}", future.cause() == null
+                                        ? "unknown" : future.cause().getClass().getSimpleName());
+                            result.complete(false);
+                        }
+                    });
+                } catch (RuntimeException e) {
+                    log.warn("协议协商失败", e);
+                    result.complete(false);
+                }
+            });
+        } catch (java.util.concurrent.RejectedExecutionException e) {
+            if (!stopping) log.warn("安排协议协商失败", e);
+            result.complete(false);
+        }
+        return result;
+    }
+
+    /** 等待协商结果；超时或中断都按「未协商」处理（宁可不下发，也不盲目重投）。 */
+    boolean awaitNegotiation(CompletableFuture<Boolean> negotiation) {
+        if (negotiation == null)
+            return v2Negotiated;
+        try {
+            // future 自身已在 10 秒时按「未协商」完成，这里留少量余量即可。
+            return negotiation.get(12, TimeUnit.SECONDS);
+        } catch (InterruptedException e) {
+            Thread.currentThread().interrupt();
+            return false;
+        } catch (Exception e) {
+            return false;
+        }
+    }
+
+    /**
+     * 下发一条带稳定 operationId 的任务命令；未协商 v2 时退化为旧的 DownloadPostMessage，
+     * 保证新旧节点组合仍可用（旧节点上无法保证可靠 ACK，这是已知的兼容边界）。
+     */
+    public byte sendTaskCommand(String action, String operationId, int gid, String name) {
+        if (!v2Negotiated)
+            return TaskCommandMessage.ACTION_DELETE.equals(action)
+                    ? legacyDeleteCommand(gid, name)
+                    : legacyDownloadCommand(gid, name);
+        TaskCommandMessage command = new TaskCommandMessage();
+        command.setOperationId(operationId);
+        command.setAction(action);
+        command.setGid(gid);
+        command.setName(name);
+        return sendRequest(command, 10, TimeUnit.SECONDS);
+    }
+
+    /** 删除命令：带独立稳定 operationId，重试复用同一个 ID。 */
+    public byte sendDeleteCommand(String operationId, int gid, String name) {
+        if (!v2Negotiated)
+            return legacyDeleteCommand(gid, name);
+        TaskCommandMessage command = new TaskCommandMessage();
+        command.setOperationId(operationId);
+        command.setAction(TaskCommandMessage.ACTION_DELETE);
+        command.setGid(gid);
+        command.setName(name);
+        return sendRequest(command, 10, TimeUnit.SECONDS);
+    }
+
+    /** 旧的无 operationId 删除命令，供未协商 v2 时退避使用。 */
+    private byte legacyDeleteCommand(int gid, String name) {
+        Gallery gallery = new Gallery();
+        gallery.setGid(gid);
+        gallery.setName(name);
+        return deleteGallery(gallery);
+    }
+
     public RetryResult retryGallery(Gallery gallery){
+        // 提交结果未知或被删除的任务不能重试：节点可能已经把未计量任务当成新下载。
+        GalleryTaskLifecycleService.DispatchDecision decision = resolveDispatchDecision(gallery.getGid());
+        if (decision != null && !decision.dispatchable())
+            return new RetryResult(false, "该任务提交结果未知或正在删除，需要人工对账后再重试");
         CompletableFuture<String> statusFuture = new CompletableFuture<>();
         retryStatusWaiters.computeIfAbsent(gallery.getGid(), ignored -> new CopyOnWriteArrayList<>())
                 .add(statusFuture);
@@ -447,6 +717,18 @@ public class RemoteService {
                     }
                 });
             }
+            else if(msg instanceof ProtocolHelloMessage hello) {
+                // 节点声明了 v2 可靠任务能力；此后才允许使用类型 11..13。
+                v2Negotiated = hello.supportsReliableTasks();
+                log.info("节点协议协商完成 v2={} version={}", v2Negotiated, hello.getProtocolVersion());
+                CompletableFuture<Boolean> negotiation = helloNegotiation;
+                if (negotiation != null)
+                    negotiation.complete(v2Negotiated);
+            }
+            else if(msg instanceof TaskReportMessage report) {
+                // 事务提交与 ACK 都放在单线程消费队列里顺序执行，避免并发写库。
+                statusApplyExecutor.execute(() -> consumeV2Report(report));
+            }
             else if(msg instanceof ResponseMessage rsm) {
                 Promise<AbstractMessage> promise = promiseHashMap.remove(rsm.messageId);
                 if(promise != null)
@@ -456,6 +738,55 @@ public class RemoteService {
             }
         }
 
+        /**
+         * 消费协议 v2 上报：短事务提交 → 事务成功后回 ACK → 再用独立执行器做通知/推送。
+         *
+         * <p>顺序很关键：数据库失败时既不能回 ACK，也不能清任务；慢通知不能阻塞这里，
+         * 否则会拖住后续上报的提交。
+         */
+        private void consumeV2Report(TaskReportMessage report) {
+            try {
+                TaskReportHandler handler = taskReportHandler;
+                if (handler == null) {
+                    log.warn("收到 v2 上报但未装配处理器，暂不回执 reportId={}", report.getReportId());
+                    return;
+                }
+                TaskReportHandler.CommittedReport committed = handler.apply(report);
+                if (committed == null) {
+                    // 提交失败：不回 ACK，节点保留快照后重投。
+                    log.warn("v2 上报未提交，等待节点重投 reportId={}", report.getReportId());
+                    return;
+                }
+                // 提交成功之后才回执，让节点清理快照。
+                Channel target = channel;
+                if (target != null && target.isActive())
+                    target.writeAndFlush(new TaskReportAckMessage(committed.reportId(), TaskReportAckMessage.RESULT_SUCCESS));
+                dispatchNotifications(report, committed);
+            } catch (Exception e) {
+                log.warn("处理 v2 上报失败 reportId={}", report.getReportId(), e);
+            }
+        }
+
+        /** 通知/推送在独立执行器里执行，外部延迟不会阻塞状态提交。 */
+        private void dispatchNotifications(TaskReportMessage report, TaskReportHandler.CommittedReport committed) {
+            // 只推送事务里真正被接受的更新（含下载中/压缩中等中间状态）；
+            // 过期、旧世代、待删除任务已被过滤掉，不会把它们带到前端。
+            GalleryTask[] accepted = committed.acceptedTasks() == null
+                    ? new GalleryTask[0]
+                    : committed.acceptedTasks().toArray(new GalleryTask[0]);
+            List<Gallery> newlyCompleted = committed.newlyCompleted();
+            notificationExecutor.execute(() -> {
+                try {
+                    for (Gallery gallery : newlyCompleted)
+                        pushService.downloadComplete(gallery);
+                    if (accepted.length > 0)
+                        webSocketService.updateTaskProcessing(accepted);
+                } catch (Exception e) {
+                    log.warn("任务完成通知/推送失败", e);
+                }
+            });
+        }
+
         /** 把节点上报的任务状态写回库里并推送给前端，按上报顺序逐个处理。 */
         private void applyReportedStatus(GalleryTask[] galleryTasks) {
             for (GalleryTask galleryTask : galleryTasks) {

src/main/java/com/lion/lionwebsite/Service/SubService.java

@@ -17,6 +17,10 @@ import lombok.extern.slf4j.Slf4j;
 import org.jsoup.Jsoup;
 import org.jsoup.select.Elements;
 import org.springframework.stereotype.Service;
+import org.springframework.beans.factory.annotation.Autowired;
+import org.springframework.beans.factory.annotation.Qualifier;
+import org.springframework.transaction.support.TransactionTemplate;
+import jakarta.annotation.PreDestroy;
 
 import java.io.IOException;
 import java.nio.charset.StandardCharsets;
@@ -39,6 +43,10 @@ import java.util.function.Supplier;
         final RemoteService remoteService;
         final SubscriptionStateCoordinator stateCoordinator;
 
+        @Autowired
+        @Qualifier("mainTransactionTemplate")
+        TransactionTemplate transactions;
+
         /**
          * 归属地查询是同步外呼（最长十几秒），且失败只影响记录里的一列展示。
          * 若放在订阅分发路径上，用户取一次订阅就得先等它完成，因此改为后台补齐。
@@ -48,11 +56,12 @@ import java.util.function.Supplier;
         Executor locationExecutor = newFixedLocationExecutor();
 
         private static ExecutorService newFixedLocationExecutor() {
-            return Executors.newFixedThreadPool(2, r -> {
+            return new java.util.concurrent.ThreadPoolExecutor(2, 2, 0, java.util.concurrent.TimeUnit.MILLISECONDS,
+                    new java.util.concurrent.ArrayBlockingQueue<>(128), r -> {
                 Thread thread = new Thread(r, "sub-location-resolver");
                 thread.setDaemon(true);
                 return thread;
-            });
+            }, new java.util.concurrent.ThreadPoolExecutor.AbortPolicy());
         }
 
     public String insertSubscriptionAccount(String name, String upstreamKey, boolean filterHighMultiplier, boolean enabled) {
@@ -211,6 +220,16 @@ import java.util.function.Supplier;
     }
 
     public void updateSub(HttpServletResponse response, HttpServletRequest request, String client, String key) {
+        Lock lock = stateCoordinator.readLock();
+        lock.lock();
+        try {
+            distributeSubscription(response, request, client, key);
+        } finally {
+            lock.unlock();
+        }
+    }
+
+    private void distributeSubscription(HttpServletResponse response, HttpServletRequest request, String client, String key) {
         if (key == null || client == null)
             return;
         SubBind subBind = subMapper.selectSubBind(key);
@@ -221,7 +240,7 @@ import java.util.function.Supplier;
         String ip = resolveClientIp(request);
         String ua = request.getHeader("User-Agent");
         if (ua == null) return;
-        recordUpdate(subBind.getUser(), ip, ua);
+        recordUpdate(subBind, ip, ua);
         if (!"v2".equals(client) && !"cat".equals(client)) {
             sendStatus(response, HttpServletResponse.SC_BAD_REQUEST, "client error");
             return;
@@ -259,24 +278,34 @@ import java.util.function.Supplier;
                 || "0:0:0:0:0:0:0:1".equals(address);
     }
 
-    private void recordUpdate(String user, String ip, String ua) {
+    private void recordUpdate(SubBind binding, String ip, String ua) {
+        String user = binding.getUser();
         // 先落一条 unknown 记录并立即返回，归属地随后在后台补齐；
         // 这样取订阅的延迟不再取决于第三方定位站点的响应时间。
-        subMapper.insertSubUpdateRecord(new SubUpdateRecord(0, user, ip, ua, new Date(), "unknown"));
+        SubUpdateRecord record = new SubUpdateRecord(0, user, ip, ua, new Date(), "unknown");
+        record.setUserId(binding.getUserId());
+        subMapper.insertSubUpdateRecord(record);
         if (subMapper.selectUpdateRecordCount(user) > 10)
             subMapper.deleteSubUpdateRecordById(subMapper.selectMinUpdateRecordId(user));
-        locationExecutor.execute(() -> enrichLocation(user, ip));
+        // Capture the generated row ID: another request may arrive before the lookup finishes.
+        if (record.getId() > 0) {
+            try {
+                locationExecutor.execute(() -> enrichLocation(record.getId(), ip));
+            } catch (java.util.concurrent.RejectedExecutionException busy) {
+                log.debug("位置查询队列已满，保留 unknown recordId={}", record.getId());
+            }
+        }
     }
 
-    /** 后台补齐最近一条记录的归属地；任何失败都只记日志，不影响已落库的记录。 */
-    private void enrichLocation(String user, String ip) {
+    /** 后台按记录 ID 补齐归属地；任何失败都只记日志，不影响已落库的记录。 */
+    private void enrichLocation(int recordId, String ip) {
         try {
             String location = resolveLocation(ip);
             if (location == null || location.isBlank() || "unknown".equals(location))
                 return;
-            subMapper.updateLatestSubUpdateRecordLocation(user, location);
+            subMapper.updateSubUpdateRecordLocation(recordId, location);
         } catch (Exception e) {
-            log.debug("补齐订阅更新记录归属地失败 user={}", user, e);
+            log.debug("补齐订阅更新记录归属地失败 recordId={}", recordId, e);
         }
     }
 
@@ -310,20 +339,26 @@ import java.util.function.Supplier;
         Lock stateLock = stateCoordinator.writeLock();
         stateLock.lock();
         try {
-            return action.get();
+            return transactions == null ? action.get() : transactions.execute(ignored -> action.get());
         } finally {
             stateLock.unlock();
         }
     }
 
+    @PreDestroy
+    void shutdownLocationResolver() {
+        if (locationExecutor instanceof ExecutorService service)
+            service.shutdownNow();
+    }
+
     private SubscriptionAccount firstEnabledAccount() {
         return subMapper.selectAllSubscriptionAccounts().stream().filter(SubscriptionAccount::isEnabled).findFirst().orElse(null);
     }
 
-    private String accountJson(SubscriptionAccount account) {
+    private tools.jackson.databind.JsonNode accountJson(SubscriptionAccount account) {
         account.setV2Url(refreshService.v2Url(account));
         account.setClashUrl(refreshService.clashUrl(account));
-        return CustomUtil.objectMapper.valueToTree(account).toString();
+        return CustomUtil.objectMapper.valueToTree(account);
     }
 
     private static void sendStatus(HttpServletResponse response, int status, String message) {

src/main/java/com/lion/lionwebsite/Service/SubscriptionRefreshService.java

@@ -2,7 +2,9 @@ package com.lion.lionwebsite.Service;
 
 import com.lion.lionwebsite.Dao.normal.SubMapper;
 import com.lion.lionwebsite.Domain.SubscriptionAccount;
+import com.lion.lionwebsite.Util.ClashSubscriptionParser;
 import com.lion.lionwebsite.Util.SubscriptionClientProfile;
+import com.lion.lionwebsite.Util.SubscriptionCacheStore;
 import lombok.RequiredArgsConstructor;
 import lombok.extern.slf4j.Slf4j;
 import org.apache.hc.client5.http.classic.methods.HttpGet;
@@ -57,6 +59,12 @@ public class SubscriptionRefreshService {
     /** 时间源，测试可替换以固定成功时刻。 */
     Clock clock = Clock.systemDefaultZone();
 
+    /**
+     * 分代缓存存储。默认按 {@link #cacheRoot} 懒构造；测试可覆写为受控子类来模拟写入失败。
+     * 保持为普通字段而不是构造参数，是为了不改动既有两参构造器（Spring 与老测试都在用）。
+     */
+    SubscriptionCacheStore cacheStore;
+
     private long refreshSequence;
     // Accessed only while holding the coordinator write lock.
     private final Map<Integer, Long> latestRefresh = new HashMap<>();
@@ -78,18 +86,19 @@ public class SubscriptionRefreshService {
 
         // Network access and parsing never hold the shared subscription lock.
         try {
+            boolean filter = account.isFilterHighMultiplier();
             String v2 = processV2(firstLine(download(v2Url(account), SubscriptionClientProfile.forAccount(accountId, false))),
-                    account.isFilterHighMultiplier(), highMultiplierThreshold);
-            List<String> clash = processClash(download(clashUrl(account), SubscriptionClientProfile.forAccount(accountId, true)),
-                    account.isFilterHighMultiplier(), highMultiplierThreshold);
+                    filter, highMultiplierThreshold);
+            // 按 YAML 语义过滤高倍节点并同步清理 proxy-groups 引用，不再靠行缩进猜测。
+            String clash = ClashSubscriptionParser.transform(
+                    download(clashUrl(account), SubscriptionClientProfile.forAccount(accountId, true)),
+                    filter, highMultiplierThreshold);
             stateLock.lock();
             try {
                 if (!isCurrent(account, version))
                     return false;
-                Path dir = Paths.get(cacheRoot, String.valueOf(accountId));
-                Files.createDirectories(dir);
-                atomicWrite(dir.resolve("v2ray.txt"), v2.getBytes(StandardCharsets.UTF_8));
-                atomicWrite(dir.resolve("clash.yaml"), String.join("\n", clash).concat("\n").getBytes(StandardCharsets.UTF_8));
+                // 两份内容写进同一分代，校验通过后指针最后发布：任一步失败都保留上次完整版本。
+                store().publish(accountId, v2.getBytes(StandardCharsets.UTF_8), clash.getBytes(StandardCharsets.UTF_8));
                 subMapper.markSubscriptionRefreshSuccess(accountId, clock.millis());
                 return true;
             } finally {
@@ -162,56 +171,6 @@ public class SubscriptionRefreshService {
         return URLDecoder.decode(node.substring(hash + 1), StandardCharsets.UTF_8);
     }
 
-    private List<String> processClash(List<String> source, boolean filter, double threshold) {
-        Set<String> removed = new HashSet<>();
-        List<String> result = new ArrayList<>();
-        boolean inProxies = false;
-        boolean skipNode = false;
-        for (String line : source) {
-            if (line.equals("proxies:")) {
-                inProxies = true;
-                skipNode = false;
-                result.add(line);
-                continue;
-            }
-            if (line.equals("proxy-groups:")) {
-                inProxies = false;
-                skipNode = false;
-                result.add(line);
-                continue;
-            }
-            if (inProxies && line.matches("^\\s{2}-\\s+name:.*")) {
-                String name = clashName(line);
-                skipNode = filter && isHigh(name, threshold);
-                if (skipNode)
-                    removed.add(name);
-                else
-                    result.add(line);
-                continue;
-            }
-            if (skipNode)
-                continue;
-            if (!inProxies && !removed.isEmpty() && line.trim().startsWith("- ")) {
-                String ref = line.trim().substring(2).trim();
-                if (removed.contains(unquote(ref)))
-                    continue;
-            }
-            result.add(line);
-        }
-        return result;
-    }
-
-    private static String clashName(String line) {
-        int index = line.indexOf("name:");
-        return unquote(line.substring(index + 5).trim());
-    }
-
-    private static String unquote(String value) {
-        if (value.length() >= 2 && ((value.startsWith("\"") && value.endsWith("\"")) || (value.startsWith("'") && value.endsWith("'"))))
-            return value.substring(1, value.length() - 1);
-        return value;
-    }
-
     private static boolean isHigh(String name, double threshold) {
         Matcher matcher = MULTIPLIER.matcher(name);
         return matcher.find() && Double.parseDouble(matcher.group(1)) > threshold;
@@ -251,22 +210,17 @@ public class SubscriptionRefreshService {
         }
     }
 
-    private static void atomicWrite(Path target, byte[] data) throws IOException {
-        Path temp = target.resolveSibling(target.getFileName() + ".tmp");
-        Files.write(temp, data);
-        try {
-            Files.move(temp, target, StandardCopyOption.ATOMIC_MOVE, StandardCopyOption.REPLACE_EXISTING);
-        } catch (AtomicMoveNotSupportedException e) {
-            Files.move(temp, target, StandardCopyOption.REPLACE_EXISTING);
-        }
-    }
-
+    /**
+     * 当前应分发的缓存文件。只在一整个版本有效时返回真实路径；否则返回一个保证不存在的
+     * 哨兵路径，使分发端按「订阅暂不可用」处理，而不是发出半份旧缓存。
+     */
     public Path cachedPath(Integer accountId, String client) {
-        return Paths.get(cacheRoot, String.valueOf(accountId), client.equals("v2") ? "v2ray.txt" : "clash.yaml");
+        SubscriptionCacheStore store = store();
+        return store.resolve(accountId, client).orElseGet(() -> store.unavailablePath(accountId, client));
     }
 
     public boolean hasCompleteCache(Integer accountId) {
-        return Files.isRegularFile(cachedPath(accountId, "v2")) && Files.isRegularFile(cachedPath(accountId, "cat"));
+        return store().hasComplete(accountId);
     }
 
     public void invalidateCache(Integer accountId) {
@@ -274,12 +228,15 @@ public class SubscriptionRefreshService {
         stateLock.lock();
         try {
             latestRefresh.remove(accountId); // In-flight responses must not restore invalidated content.
-            Files.deleteIfExists(cachedPath(accountId, "v2"));
-            Files.deleteIfExists(cachedPath(accountId, "cat"));
-        } catch (IOException e) {
-            log.warn("清理失效订阅缓存失败 accountId={}", accountId, e);
+            store().invalidate(accountId);
         } finally {
             stateLock.unlock();
         }
     }
+
+    private synchronized SubscriptionCacheStore store() {
+        if (cacheStore == null)
+            cacheStore = new SubscriptionCacheStore(cacheRoot);
+        return cacheStore;
+    }
 }

src/main/java/com/lion/lionwebsite/Service/SubscriptionStandbySnapshotService.java

@@ -6,6 +6,7 @@ import com.lion.lionwebsite.Domain.SubBind;
 import com.lion.lionwebsite.Domain.SubscriptionAccount;
 import com.lion.lionwebsite.Message.*;
 import com.lion.lionwebsite.Util.CustomUtil;
+import com.lion.lionwebsite.Util.SubscriptionCacheStore;
 import lombok.RequiredArgsConstructor;
 import org.springframework.beans.factory.annotation.Value;
 import org.springframework.stereotype.Service;
@@ -16,8 +17,6 @@ import javax.crypto.spec.SecretKeySpec;
 import java.io.ByteArrayOutputStream;
 import java.io.IOException;
 import java.nio.charset.StandardCharsets;
-import java.nio.file.Files;
-import java.nio.file.Path;
 import java.security.MessageDigest;
 import java.util.*;
 import java.util.zip.GZIPOutputStream;
@@ -39,6 +38,12 @@ public class SubscriptionStandbySnapshotService {
     @Value("${subscription.cache-root:sub/accounts}")
     String cacheRoot;
 
+    /**
+     * 与 {@link SubscriptionRefreshService} 共用的分代缓存解析器。
+     * 测试可直接覆写；未设置时按 {@link #cacheRoot} 懒构造。
+     */
+    SubscriptionCacheStore cacheStore;
+
     @PostConstruct
     void validateConfiguration() {
         if (syncEnabled && (syncSecret == null || syncSecret.isBlank()))
@@ -51,12 +56,14 @@ public class SubscriptionStandbySnapshotService {
         try {
         Map<Integer, SubscriptionAccountSnapshot> accountMap = new HashMap<>();
         for (SubscriptionAccount account : subMapper.selectAllSubscriptionAccounts()) {
-            Path v2Path = cachedPath(account.getId(), "v2");
-            Path clashPath = cachedPath(account.getId(), "cat");
-            if (!account.isEnabled() || !Files.isRegularFile(v2Path) || !Files.isRegularFile(clashPath))
+            if (!account.isEnabled())
                 continue;
-            byte[] v2 = Files.readAllBytes(v2Path);
-            byte[] clash = Files.readAllBytes(clashPath);
+            // 从同一分代读出成对内容，保证快照里两份文件同版本；无完整版本则跳过该账号。
+            Optional<SubscriptionCacheStore.Pair> pair = store().read(account.getId());
+            if (pair.isEmpty())
+                continue;
+            byte[] v2 = pair.get().v2();
+            byte[] clash = pair.get().clash();
             SubscriptionAccountSnapshot snapshot = new SubscriptionAccountSnapshot();
             snapshot.setAccountId(account.getId());
             snapshot.setEnabled(true);
@@ -132,7 +139,9 @@ public class SubscriptionStandbySnapshotService {
 
     private static String hex(byte[] input) { return HexFormat.of().formatHex(input); }
 
-    private Path cachedPath(Integer accountId, String client) {
-        return Path.of(cacheRoot, String.valueOf(accountId), "v2".equals(client) ? "v2ray.txt" : "clash.yaml");
+    private synchronized SubscriptionCacheStore store() {
+        if (cacheStore == null)
+            cacheStore = new SubscriptionCacheStore(cacheRoot);
+        return cacheStore;
     }
 }

src/main/java/com/lion/lionwebsite/Service/TaskProtocol.java

@@ -0,0 +1,23 @@
+package com.lion.lionwebsite.Service;
+
+import java.util.UUID;
+
+/** 协议 v2 的操作 ID 生成与解析。稳定 ID 是「节点重发/重试不重复执行副作用」的依据。 */
+public final class TaskProtocol {
+
+    private TaskProtocol() {
+    }
+
+    /** 生成一个新的稳定操作 ID（下载或删除），形如 {@code op-<uuid>}。 */
+    public static String newOperationId(String action) {
+        return action + "-" + UUID.randomUUID();
+    }
+
+    public static String newDownloadOperationId() {
+        return newOperationId("dl");
+    }
+
+    public static String newDeleteOperationId() {
+        return newOperationId("del");
+    }
+}

src/main/java/com/lion/lionwebsite/Service/TaskReportHandler.java

@@ -0,0 +1,37 @@
+package com.lion.lionwebsite.Service;
+
+import com.lion.lionwebsite.Domain.Gallery;
+import com.lion.lionwebsite.Domain.GalleryTask;
+import com.lion.lionwebsite.Message.TaskReportMessage;
+
+import java.util.List;
+
+/**
+ * 协议 v2 状态上报的消费入口。
+ *
+ * <p>定义成接口是为了断开依赖环：RemoteService 需要把上报交给生命周期服务，
+ * 而生命周期服务需要有 RemoteService 才能下发命令。RemoteService 通过可选的 setter
+ * 注入本接口（见 {@code RemoteService#setTaskReportHandler}），因此构造期没有环。
+ */
+public interface TaskReportHandler {
+
+    /**
+     * 提交一份上报。
+     *
+     * @return 已成功提交（可回 ACK）时返回提交结果；数据库失败等「不可回 ACK」的情况返回 null。
+     *         返回 null 时节点会保留快照并稍后重投，这正是「失败写入不得清任务」的要求。
+     */
+    CommittedReport apply(TaskReportMessage report);
+
+    /**
+     * 已提交的结果。
+     *
+     * @param reportId       本次上报 ID（回执用）
+     * @param newlyCompleted 本次「首次」完成的任务，通知与 WebSocket 由独立执行器在事务之后处理，
+     *                       因此慢通知不会阻塞数据库消费线程；重复上报不会重复出现在这里
+     * @param acceptedTasks  事务内实际被接受的进度更新（含下载中/压缩中等中间状态，已排除过期、
+     *                       旧世代与待删除任务），WebSocket 只推这一批，避免前端看到被丢弃的迟到状态
+     */
+    record CommittedReport(String reportId, List<Gallery> newlyCompleted, List<GalleryTask> acceptedTasks) {
+    }
+}

src/main/java/com/lion/lionwebsite/Service/TaskStatus.java

@@ -0,0 +1,45 @@
+package com.lion.lionwebsite.Service;
+
+import com.lion.lionwebsite.Domain.GalleryTask;
+
+/**
+ * 任务业务状态码到展示文案的唯一映射点。
+ *
+ * <p>状态以稳定代码表达，中文文案只出现在展示层（WebSocket/列表），
+ * 避免把「失败/待重试/结果未知」塞进成功失败提示字符串。
+ */
+public final class TaskStatus {
+
+    public static final String SUBMITTED = "已提交";
+    /** 已登记创建意图、尚未确认上游提交；用于让任务列表在提交期间也能看到占位任务。 */
+    public static final String PENDING_SUBMIT = "提交中";
+    public static final String DOWNLOADING = "下载中";
+    public static final String WAITING_COMPRESS = "等待压缩";
+    public static final String COMPRESSING = "压缩中";
+    public static final String COMPLETE = "下载完成";
+    public static final String UNKNOWN = "提交未知";
+    public static final String DELETE_PENDING = "待删除";
+    public static final String DELETE_FAILED = "删除失败";
+
+    private TaskStatus() {
+    }
+
+    /**
+     * 节点上报的进度状态映射为业务文案；未知状态码返回 null，由调用方忽略。
+     *
+     * <p>GalleryTask 里的状态码是可变 static byte，不是编译期常量，因此不能用 switch case；
+     * 这里用整数比较，避免为改常量引入风险。
+     */
+    public static String fromReport(byte reportStatus) {
+        int status = reportStatus;
+        if (status == GalleryTask.DOWNLOADING)
+            return DOWNLOADING;
+        if (status == GalleryTask.DOWNLOAD_COMPLETE)
+            return WAITING_COMPRESS;
+        if (status == GalleryTask.COMPRESSING)
+            return COMPRESSING;
+        if (status == GalleryTask.COMPRESS_COMPLETE)
+            return COMPLETE;
+        return null;
+    }
+}

src/main/java/com/lion/lionwebsite/Service/UpstreamGateway.java

@@ -0,0 +1,47 @@
+package com.lion.lionwebsite.Service;
+
+import com.lion.lionwebsite.Domain.Gallery;
+import com.lion.lionwebsite.Exception.ResolutionNotMatchException;
+import com.lion.lionwebsite.Util.GalleryUtil;
+import com.lion.lionwebsite.Util.PreparedGallery;
+import org.springframework.stereotype.Component;
+
+import java.io.IOException;
+
+/**
+ * 上游图站的读写边界。
+ *
+ * <p>分成两步是协议 v2 的核心约束：
+ * <ol>
+ *   <li>{@link #prepare} 只做读取（页面、分辨率），绝不产生下载副作用；</li>
+ *   <li>调用方必须在持久化 SUBMITTING 之后才调用 {@link #submit}，因为那一步才真正 POST。</li>
+ * </ol>
+ *
+ * <p>拆成接口是为了让生命周期服务可以在测试里注入假实现，也方便后续把抓取/解析/提交
+ * 进一步分开；默认实现 {@link GalleryUtilAdapter} 直接转发到现有 GalleryUtil。
+ */
+public interface UpstreamGateway {
+
+    /** 只读解析：失败返回 null 表示链接无效或页面不可用，不产生任何副作用。 */
+    PreparedGallery prepare(String link) throws IOException;
+
+    /** 真正提交下载请求。抛异常表示结果未知，由调用方落 UNKNOWN，不做自动重发。 */
+    Gallery submit(PreparedGallery prepared, String targetResolution)
+            throws IOException, ResolutionNotMatchException;
+
+    /** 默认实现：转发现有 GalleryUtil 的拆分方法。 */
+    @Component
+    class GalleryUtilAdapter implements UpstreamGateway {
+
+        @Override
+        public PreparedGallery prepare(String link) throws IOException {
+            return GalleryUtil.prepare(link);
+        }
+
+        @Override
+        public Gallery submit(PreparedGallery prepared, String targetResolution)
+                throws IOException, ResolutionNotMatchException {
+            return GalleryUtil.submit(prepared, targetResolution);
+        }
+    }
+}

src/main/java/com/lion/lionwebsite/Service/UserService.java

@@ -3,6 +3,7 @@ package com.lion.lionwebsite.Service;
 import com.lion.lionwebsite.Dao.normal.CollectMapper;
 import com.lion.lionwebsite.Dao.normal.GalleryMapper;
 import com.lion.lionwebsite.Dao.normal.UserMapper;
+import com.lion.lionwebsite.Dao.normal.SubMapper;
 import com.lion.lionwebsite.Domain.User;
 import com.lion.lionwebsite.Interceptor.TaskHandlerInterceptor;
 import com.lion.lionwebsite.Util.CustomUtil;
@@ -10,8 +11,13 @@ import com.lion.lionwebsite.Util.Response;
 import lombok.RequiredArgsConstructor;
 import lombok.extern.slf4j.Slf4j;
 import org.springframework.stereotype.Service;
+import org.springframework.beans.factory.annotation.Autowired;
+import org.springframework.beans.factory.annotation.Qualifier;
+import org.springframework.transaction.support.TransactionTemplate;
 
 import java.util.ArrayList;
+import java.util.concurrent.locks.Lock;
+import java.util.function.Supplier;
 
 import static com.lion.lionwebsite.Util.CustomUtil.objectMapper;
 
@@ -28,6 +34,11 @@ public class UserService{
 
     final TaskHandlerInterceptor taskHandlerInterceptor;
 
+    @Autowired SubMapper subMapper;
+    @Autowired SubscriptionStateCoordinator subscriptionState;
+    @Autowired RemoteService remoteService;
+    @Autowired @Qualifier("mainTransactionTemplate") TransactionTemplate transactions;
+
     public String addAuthCode(String targetAuthCode, String people) {
         Response response = Response.generateResponse();
 
@@ -66,12 +77,22 @@ public class UserService{
         Response response = Response.generateResponse();
 
         try{
-            if(userMapper.isExist(targetAuthCode) != 0){
-                userMapper.updateUsername(targetAuthCode, newUsername);
-                response.success("修改成功");
-            }else {
-                response.failure("授权码不存在");
-            }
+            return coordinated(() -> {
+                if (userMapper.isExist(targetAuthCode) == 0)
+                    return response.failure("授权码不存在").toJSONString();
+                if (newUsername == null || newUsername.isBlank())
+                    return response.failure("用户名不能为空").toJSONString();
+                String name = newUsername.trim();
+                User current = userMapper.selectUserByAuthCode(targetAuthCode);
+                User existing = userMapper.selectUserByUsername(name);
+                if (existing != null && (current == null || existing.getId() != current.getId()))
+                    return response.failure("用户名已存在").toJSONString();
+                userMapper.updateUsername(targetAuthCode, name);
+                if (subMapper != null && current != null)
+                    subMapper.renameBoundUser(current.getId(), current.getUsername(), name);
+                if (remoteService != null) remoteService.requestSubscriptionSync();
+                return response.success("修改成功").toJSONString();
+            });
         }catch (Exception e){
             log.error("修改用户名失败", e);
             response.failure("修改失败");
@@ -85,17 +106,22 @@ public class UserService{
 
         if(userMapper.isExist(targetAuthCode) != 0)
             try{
-                User user = userMapper.selectUserByAuthCode(targetAuthCode);
-
-                //修改下载人以及取消收藏
-                ArrayList<Integer> galleries = collectMapper.selectGidByCollector(user.getId());
-                for(Integer gallery: galleries)
-                    collectMapper.disCollect(gallery, user.getId());
-                galleryMapper.updateGalleryDownloader(3, user.getId());
-
-                userMapper.deleteUserByAuthCode(targetAuthCode);
+                coordinated(() -> {
+                    User user = userMapper.selectUserByAuthCode(targetAuthCode);
+                    ArrayList<Integer> galleries = collectMapper.selectGidByCollector(user.getId());
+                    for(Integer gallery: galleries)
+                        collectMapper.disCollect(gallery, user.getId());
+                    galleryMapper.updateGalleryDownloader(ADMIN_USER_ID, user.getId());
+                    if (subMapper != null) {
+                        subMapper.deleteSubUpdateRecord(user.getUsername());
+                        subMapper.deleteSubBind(user.getUsername());
+                    }
+                    userMapper.deleteUserByAuthCode(targetAuthCode);
+                    return "deleted";
+                });
                 response.success("删除成功");
                 taskHandlerInterceptor.updateAuthCodes();
+                if (remoteService != null) remoteService.requestSubscriptionSync();
             }catch (Exception e){
                 log.error("删除授权码失败", e);
                 response.failure("删除失败");
@@ -108,27 +134,29 @@ public class UserService{
     }
 
     public String alterStatus(String AuthCode, boolean isEnable){
-        Response response = Response.generateResponse();
-
-        if(userMapper.isExist(AuthCode) == 1){
-            int id = userMapper.selectUserByAuthCode(AuthCode).getId();
-            userMapper.updateIsEnableById(id, isEnable);
-            galleryMapper.updateGalleryDownloader(3, id);
-            response.success("修改用户状态成功");
-            taskHandlerInterceptor.updateAuthCodes();
+        try {
+            String result = coordinated(() -> {
+                if (userMapper.isExist(AuthCode) != 1)
+                    return Response._failure("该用户不存在");
+                int id = userMapper.selectUserByAuthCode(AuthCode).getId();
+                userMapper.updateIsEnableById(id, isEnable);
+                if (!isEnable) galleryMapper.updateGalleryDownloader(ADMIN_USER_ID, id);
+                return Response._success("修改用户状态成功");
+            });
+            // Refresh the in-memory identity view only after the database commits.
+            if (result.contains("\"result\":\"success\"")) taskHandlerInterceptor.updateAuthCodes();
+            return result;
+        } catch (RuntimeException error) {
+            log.warn("修改用户状态失败: {}", error.getClass().getSimpleName());
+            return Response._failure("修改用户状态失败");
         }
-        else{
-            response.failure("该用户不存在");
-        }
-
-        return response.toJSONString();
     }
 
     public String getAllUser() {
         Response response = Response.generateResponse();
         User[] users = userMapper.selectAllUser();
 
-        response.success(objectMapper.valueToTree(users).toString());
+        response.success(objectMapper.valueToTree(users));
 
         return response.toJSONString();
     }
@@ -150,4 +178,14 @@ public class UserService{
         User user = userMapper.selectUserByAuthCode(AuthCode);
         return user != null && user.getId() == ADMIN_USER_ID;
     }
+
+    private String coordinated(Supplier<String> action) {
+        Lock lock = subscriptionState == null ? null : subscriptionState.writeLock();
+        if (lock != null) lock.lock();
+        try {
+            return transactions == null ? action.get() : transactions.execute(ignored -> action.get());
+        } finally {
+            if (lock != null) lock.unlock();
+        }
+    }
 }

src/main/java/com/lion/lionwebsite/Util/ClashSubscriptionParser.java

@@ -0,0 +1,45 @@
+package com.lion.lionwebsite.Util;
+
+import org.yaml.snakeyaml.Yaml;
+import org.yaml.snakeyaml.DumperOptions;
+import java.util.*;
+import java.util.regex.Pattern;
+
+/** YAML semantics rather than indentation determine nodes and references. */
+public final class ClashSubscriptionParser {
+    private static final Pattern MULTIPLIER = Pattern.compile("(\\d+(?:\\.\\d+)?)x\\s*$", Pattern.CASE_INSENSITIVE);
+    private ClashSubscriptionParser() { }
+
+    public static String transform(List<String> lines, boolean filter, double threshold) {
+        Object loaded = new Yaml().load(String.join("\n", lines));
+        if (!(loaded instanceof Map<?, ?> source)) throw new IllegalArgumentException("Clash 上游必须为 YAML 对象");
+        if (!(source.get("proxies") instanceof List<?> proxies)) throw new IllegalArgumentException("Clash 上游缺少 proxies 列表");
+        Map<Object, Object> output = new LinkedHashMap<>(source);
+        Set<String> removed = new HashSet<>();
+        List<Object> kept = new ArrayList<>();
+        for (Object node : proxies) {
+            if (!(node instanceof Map<?, ?> fields) || !(fields.get("name") instanceof String name))
+                throw new IllegalArgumentException("Clash 节点缺少名称");
+            var multiplier = MULTIPLIER.matcher(name);
+            if (filter && multiplier.find() && Double.parseDouble(multiplier.group(1)) > threshold)
+                removed.add(name);
+            else kept.add(node);
+        }
+        output.put("proxies", kept);
+        if (source.get("proxy-groups") instanceof List<?> groups) {
+            List<Object> result = new ArrayList<>();
+            for (Object group : groups) {
+                if (!(group instanceof Map<?, ?> fields)) throw new IllegalArgumentException("Clash 分组必须为对象");
+                Map<Object, Object> copy = new LinkedHashMap<>(fields);
+                if (fields.get("proxies") instanceof List<?> references)
+                    copy.put("proxies", references.stream().filter(ref -> !removed.contains(ref)).toList());
+                result.add(copy);
+            }
+            output.put("proxy-groups", result);
+        }
+        DumperOptions options = new DumperOptions();
+        options.setDefaultFlowStyle(DumperOptions.FlowStyle.BLOCK);
+        options.setPrettyFlow(true);
+        return new Yaml(options).dump(output);
+    }
+}

src/main/java/com/lion/lionwebsite/Util/GalleryPageParser.java

@@ -0,0 +1,72 @@
+package com.lion.lionwebsite.Util;
+
+import com.lion.lionwebsite.Domain.Gallery;
+import org.jsoup.nodes.Document;
+import org.jsoup.nodes.Element;
+import java.io.IOException;
+import java.util.*;
+import java.util.regex.Pattern;
+
+/** Pure parsing: upstream requests and download submission live outside this class. */
+public final class GalleryPageParser {
+    private GalleryPageParser() { }
+
+    public static Gallery metadata(Document document, String link, int gid) throws IOException {
+        String title = document.select("#gn").text();
+        if (title.isBlank()) title = document.select("#gj").text();
+        if (title.isBlank()) throw new IOException("上游页面缺少画廊名称，可能是错误页");
+        Map<String, String> fields = new HashMap<>();
+        for (Element row : document.select("#gdd tr")) {
+            var cells = row.select("td");
+            if (cells.size() < 2) continue;
+            String name = cells.get(0).text().replace(":", "").strip().toLowerCase(Locale.ROOT);
+            fields.put(name, cells.get(1).text().strip());
+        }
+        var pages = Pattern.compile("\\d+").matcher(fields.getOrDefault("length", ""));
+        if (!pages.find()) throw new IOException("上游页面缺少画廊页数");
+        Gallery result = new Gallery();
+        result.setGid(gid); result.setLink(link); result.setName(title + " [" + gid + "]");
+        result.setCreateTime(System.currentTimeMillis() / 1000); result.setProceeding(0);
+        result.setLanguage(fields.getOrDefault("language", "").replace("\u00a0", " "));
+        try {
+            result.setPages(Integer.parseInt(pages.group()));
+            result.setFileSize(CustomUtil.stringToFileSize(fields.getOrDefault("file size", "").replace(" ", "")));
+        } catch (RuntimeException failure) { throw new IOException("上游画廊页数或文件大小格式不支持", failure); }
+        if (result.getPages() <= 0) throw new IOException("上游画廊页数必须为正数");
+        result.setDisplayFileSize(CustomUtil.fileSizeToString(result.getFileSize()));
+        Element thumbnail = document.selectFirst("#gd1 [style*=url]");
+        var image = Pattern.compile("url\\(\\s*['\"]?(.*?)['\"]?\\s*\\)").matcher(thumbnail == null ? "" : thumbnail.attr("style"));
+        String imageUrl = image.find() ? image.group(1) : document.select("#gd1 img").attr("src");
+        result.setThumb_link(imageUrl.replace("https://s.exhentai.org", ""));
+        return result;
+    }
+
+    public static String archiveLink(Document document) throws IOException {
+        for (Element anchor : document.select("#gd5 a")) {
+            var matcher = Pattern.compile("https?://[^'\"\\s]+").matcher(anchor.attr("onclick"));
+            if (matcher.find()) return matcher.group();
+            String href = anchor.absUrl("href");
+            if (!href.isBlank() && (anchor.text().toLowerCase(Locale.ROOT).contains("download") || href.contains("archive")))
+                return href;
+        }
+        throw new IOException("上游页面缺少归档入口");
+    }
+
+    public static Map<String, String> resolutions(Document document) throws IOException {
+        var expression = Pattern.compile("^(\\d+x|Original)\\s+([\\d.]+)\\s*([A-Za-z]+)$", Pattern.CASE_INSENSITIVE);
+        TreeMap<Integer, String> sizes = new TreeMap<>();
+        String original = null;
+        for (Element paragraph : document.select("#db td p")) {
+            var matcher = expression.matcher(paragraph.text().strip());
+            if (!matcher.matches()) continue;
+            String size = matcher.group(2) + matcher.group(3);
+            if (matcher.group(1).equalsIgnoreCase("Original")) original = size;
+            else sizes.put(Integer.parseInt(matcher.group(1).replace("x", "")), size);
+        }
+        Map<String, String> result = new LinkedHashMap<>();
+        sizes.forEach((width, size) -> result.put(width + "x", size));
+        if (original != null) result.put("Original", original);
+        if (result.isEmpty()) throw new IOException("上游页面没有可用分辨率");
+        return result;
+    }
+}

src/main/java/com/lion/lionwebsite/Util/GallerySearchParser.java

@@ -0,0 +1,45 @@
+package com.lion.lionwebsite.Util;
+
+import com.lion.lionwebsite.Domain.GalleryForQuery;
+import org.jsoup.nodes.Document;
+import org.jsoup.nodes.Element;
+import java.io.IOException;
+import java.util.ArrayList;
+import java.util.List;
+import java.util.regex.Pattern;
+
+/** Search rows are located by gallery links, independent of table nesting and headings. */
+public final class GallerySearchParser {
+    private GallerySearchParser() { }
+    public static List<GalleryForQuery> parse(Document document) throws IOException {
+        List<GalleryForQuery> result = new ArrayList<>();
+        for (Element row : document.select("tr")) {
+            Element link = row.selectFirst("a[href*=/g/]");
+            if (link == null) continue;
+            String href = link.attr("href");
+            Integer gid = GalleryUtil.parseGid(href);
+            if (gid == null) continue;
+            GalleryForQuery gallery = new GalleryForQuery();
+            gallery.setGid(gid.toString()); gallery.setLink(href);
+            gallery.setName(link.select(".glink").isEmpty() ? link.text() : link.select(".glink").text());
+            Element type = row.selectFirst(".cn, .cs");
+            gallery.setType(type == null ? row.select("td").first().text() : type.text());
+            Element image = row.selectFirst("img");
+            String thumbnail = image == null ? "" : image.attr("src");
+            if (thumbnail.isBlank() || thumbnail.startsWith("data:")) thumbnail = image == null ? "" : image.attr("data-src");
+            gallery.setThumbnailUrl(thumbnail.replace("https://s.exhentai.org", ""));
+            var pages = Pattern.compile("(\\d+)\\s+pages?", Pattern.CASE_INSENSITIVE).matcher(row.text());
+            if (!pages.find()) throw new IOException("上游搜索结果缺少页数 gid=" + gid);
+            gallery.setPage(Integer.parseInt(pages.group(1)));
+            Element date = row.selectFirst("[id^=posted_]");
+            if (date != null) gallery.setUploadTime(date.text());
+            else {
+                var dates = Pattern.compile("\\d{4}-\\d{2}-\\d{2}(?:\\s+\\d{2}:\\d{2})?").matcher(row.text());
+                gallery.setUploadTime(dates.find() ? dates.group() : "");
+            }
+            if (gallery.getName().isBlank()) throw new IOException("上游搜索结果缺少名称 gid=" + gid);
+            result.add(gallery);
+        }
+        return result;
+    }
+}

src/main/java/com/lion/lionwebsite/Util/GalleryUtil.java

@@ -82,6 +82,12 @@ public class GalleryUtil {
      * @throws ResolutionNotMatchException  没有目标分辨率
      */
     public static Gallery parse(String url, boolean isDownload, String targetResolution) throws IOException, ResolutionNotMatchException {
+        PreparedGallery prepared = prepare(url);
+        if (prepared == null) return null;
+        return isDownload ? submit(prepared, targetResolution) : prepared.gallery();
+    }
+
+    public static PreparedGallery prepare(String url) throws IOException {
         String origin = verifyLink(url);
 
         if(origin == null){
@@ -94,66 +100,26 @@ public class GalleryUtil {
             log.warn("链接无法解析出 gid，按无效链接处理");
             return null;
         }
-        Gallery gallery = new Gallery();
-        gallery.setLink(url);
-        gallery.setCreateTime(System.currentTimeMillis()/1000);
-        gallery.setGid(gid);
-        gallery.setProceeding(0);
-
-        //访问图片页面
-        String galleryPage = requests(url, GET, null, null);
-        Document galleryDoc = Jsoup.parse(galleryPage);
-
-        //收集图片基本信息
-        gallery.setName(galleryDoc.select("#gn").text() + " [" + gallery.getGid() + ']');
-        gallery.setLanguage(galleryDoc.select("#gdd > table > tbody > tr:nth-child(4) > td.gdt2").text().replace("%nbps", ""));
-        gallery.setPages(Integer.parseInt(galleryDoc.select("#gdd > table > tbody > tr:nth-child(6) > td.gdt2").text().split(" ")[0]));
-        gallery.setFileSize(CustomUtil.stringToFileSize(galleryDoc.select("#gdd > table > tbody > tr:nth-child(5) > td.gdt2").text().replace(" ", "")));
-        gallery.setDisplayFileSize(CustomUtil.fileSizeToString(gallery.getFileSize()));
-        String thumb_link = galleryDoc.select("#gd1 > div").getFirst().attr("style");
-        gallery.setThumb_link(thumb_link.substring(thumb_link.indexOf("url(") + 4, thumb_link.lastIndexOf(")")).replace("https://s.exhentai.org", ""));
-
-        //查找下载页面链接并初始化参数
-        String download_link = galleryDoc.select("#gd5 > p:nth-child(2) > a").attr("onclick").split("'")[1];
+        Document galleryDoc = Jsoup.parse(requests(url, GET, null, null), url);
+        Gallery gallery = GalleryPageParser.metadata(galleryDoc, url, gid);
+        String archiveLink = GalleryPageParser.archiveLink(galleryDoc);
         HashMap<String, String> headers = new HashMap<>();
         headers.put("origin", origin);
+        headers.put("referer", archiveLink);
+        Document downloadDoc = Jsoup.parse(requests(archiveLink, GET, headers, null), archiveLink);
+        String download_link = downloadDoc.select("#hathdl_form").attr("abs:action");
+        if (download_link.isBlank()) throw new IOException("上游页面缺少下载提交表单");
+        gallery.setAvailableResolution(GalleryPageParser.resolutions(downloadDoc));
+        gallery.setStatus("等待确认下载");
+        return new PreparedGallery(gallery, download_link, origin);
+    }
+
+    public static Gallery submit(PreparedGallery prepared, String targetResolution) throws IOException, ResolutionNotMatchException {
+        Gallery gallery = prepared.gallery();
+        String download_link = prepared.formUrl();
+        HashMap<String, String> headers = new HashMap<>();
+        headers.put("origin", prepared.origin());
         headers.put("referer", download_link);
-
-        //访问下载页面，获取分辨率
-        String downloadPage = requests(download_link, GET, headers, null);
-        Document downloadDoc = Jsoup.parse(downloadPage);
-        Elements resolutions = downloadDoc.select("#db > div > table > tbody > tr > td");
-        download_link = downloadDoc.select("#hathdl_form").attr("action");
-
-        Map<String, String> availableResolution = new LinkedHashMap<>();
-        Map<String, String> tempResolutions = new LinkedHashMap<>();
-        ArrayList<Integer> tempResolutionKeys = new ArrayList<>();
-
-        for(Element element: resolutions){
-            String[] parameter = element.select("td > p").text().split(" ");
-            if(parameter[1].contains("N/A"))
-                continue;
-            tempResolutions.put(parameter[0], parameter[1] + parameter[2]);
-            if(parameter[0].contains("x")){
-                tempResolutionKeys.add(Integer.valueOf(parameter[0].replace("x", "")));
-            }
-        }
-
-        tempResolutionKeys.sort(Comparator.comparingInt(o -> o));
-
-        for(Integer resolution: tempResolutionKeys){
-            availableResolution.put(resolution.toString()+"x", tempResolutions.get(resolution+"x"));
-        }
-
-        availableResolution.put("Original", tempResolutions.get("Original"));
-        gallery.setAvailableResolution(availableResolution);
-
-        //如果不是下载，则直接返回
-        if(!isDownload){
-            gallery.setStatus("等待确认下载");
-            return gallery;
-        }
-
         //如果目标分辨率不存在，抛出错误
         if(!gallery.getAvailableResolution().containsKey(targetResolution)){
             log.warn("目标分辨率不存在，可用分辨率: {}", gallery.getAvailableResolution());
@@ -161,15 +127,15 @@ public class GalleryUtil {
         }
 
         gallery.setResolution(targetResolution);
-        gallery.setFileSize(CustomUtil.stringToFileSize(availableResolution.get(targetResolution).replace(" ", "")));
+        gallery.setFileSize(CustomUtil.stringToFileSize(gallery.getAvailableResolution().get(targetResolution).replace(" ", "")));
         gallery.setDisplayFileSize(CustomUtil.fileSizeToString(gallery.getFileSize()));
 
         //提交下载请求
         HashMap<String, String> body = new HashMap<>();
         body.put("payload", FORM_DATA);
         body.put("hathdl_xres", targetResolution.replace("x", "").replace("Original", "org"));
-        downloadPage = requests(download_link, POST, headers, body);
-        downloadDoc = Jsoup.parse(downloadPage);
+        String downloadPage = requests(download_link, POST, headers, body);
+        Document downloadDoc = Jsoup.parse(downloadPage);
 
         //判断下载请求是否提交成功
         if(downloadDoc.select("#db > p:nth-child(2) > strong").text().startsWith("#"))
@@ -192,14 +158,12 @@ public class GalleryUtil {
         if(content.trim().isEmpty())
             return null;
         Document document = Jsoup.parse(content);
-        if(document.select("body > script").size() < 2)
-            return null;
-        Element script = document.select("body > script").get(1);
-        String[] scripts = script.html().split("\n");
+        String script = document.select("script").stream().map(Element::data).collect(java.util.stream.Collectors.joining("\n"));
+        if (!script.contains("imagelist") || !script.contains("mpvkey")) return null;
         ArrayList<ImageKeyCache> imageKeyCaches = new ArrayList<>();
         AtomicInteger page = new AtomicInteger(1);
-        gid2MpvKey.put(gid, scripts[1].split("=")[1].replace(";", "").replace("\"", "").replace(" ", ""));
-        JsonNode nodes = parseImagelist(scripts[2]);
+        gid2MpvKey.put(gid, JavaScriptAssignments.string(script, "mpvkey"));
+        JsonNode nodes = objectMapper.readTree(JavaScriptAssignments.array(script, "imagelist"));
         nodes.forEach((n) -> {
             ImageKeyCache imageKeyCache = new ImageKeyCache();
             imageKeyCache.setGid(gid);
@@ -247,12 +211,14 @@ public class GalleryUtil {
             gid2MpvKey.remove(parseGid(url) + "");
             return;
         }
-        Document document = Jsoup.parse(content);
-        Element script = document.select("body > script").get(1);
-        String[] scripts = script.html().split("\n");
-        String mpvKey = scripts[1].split("=")[1].replace(";", "").replace("\"", "").replace(" ", "");
-        gid2MpvKey.put(parseGid(url) + "", mpvKey);
-        log.info("刷新key:{}", mpvKey);
+        try {
+            String scripts = Jsoup.parse(content).select("script").stream().map(Element::data)
+                    .collect(java.util.stream.Collectors.joining("\n"));
+            gid2MpvKey.put(parseGid(url) + "", JavaScriptAssignments.string(scripts, "mpvkey"));
+        } catch (IOException e) {
+            gid2MpvKey.remove(parseGid(url) + "");
+            log.warn("上游mpv脚本无法解析，缓存已失效");
+        }
     }
 
     public static String getImageUrl(String mpvKey, ImageKeyCache imageKeyCache)  {

src/main/java/com/lion/lionwebsite/Util/JavaScriptAssignments.java

@@ -0,0 +1,34 @@
+package com.lion.lionwebsite.Util;
+
+import java.io.IOException;
+import java.util.regex.Pattern;
+
+/** Extracts data assignments without depending on script position or line layout. */
+public final class JavaScriptAssignments {
+    private JavaScriptAssignments() { }
+
+    public static String string(String source, String name) throws IOException {
+        var matcher = Pattern.compile("\\b" + Pattern.quote(name) + "\\s*=\\s*(['\"])(.*?)\\1", Pattern.DOTALL).matcher(source);
+        if (!matcher.find()) throw new IOException("上游脚本缺少 " + name);
+        return matcher.group(2);
+    }
+
+    public static String array(String source, String name) throws IOException {
+        var matcher = Pattern.compile("\\b" + Pattern.quote(name) + "\\s*=\\s*\\[").matcher(source);
+        if (!matcher.find()) throw new IOException("上游脚本缺少 " + name);
+        int start = matcher.end() - 1, depth = 0; char quote = 0; boolean escaped = false;
+        for (int i = start; i < source.length(); i++) {
+            char c = source.charAt(i);
+            if (quote != 0) {
+                if (escaped) escaped = false;
+                else if (c == '\\') escaped = true;
+                else if (c == quote) quote = 0;
+                continue;
+            }
+            if (c == '"' || c == '\'') quote = c;
+            else if (c == '[') depth++;
+            else if (c == ']' && --depth == 0) return source.substring(start, i + 1);
+        }
+        throw new IOException("上游脚本的 " + name + " 不完整");
+    }
+}

src/main/java/com/lion/lionwebsite/Util/PreparedGallery.java

@@ -0,0 +1,6 @@
+package com.lion.lionwebsite.Util;
+
+import com.lion.lionwebsite.Domain.Gallery;
+
+/** A read-only upstream preparation; creating this value never submits a download. */
+public record PreparedGallery(Gallery gallery, String formUrl, String origin) { }

src/main/java/com/lion/lionwebsite/Util/SubscriptionCacheStore.java

@@ -0,0 +1,328 @@
+package com.lion.lionwebsite.Util;
+
+import lombok.extern.slf4j.Slf4j;
+
+import java.io.IOException;
+import java.io.StringReader;
+import java.nio.charset.StandardCharsets;
+import java.nio.file.AtomicMoveNotSupportedException;
+import java.nio.file.Files;
+import java.nio.file.Path;
+import java.nio.file.Paths;
+import java.nio.file.StandardCopyOption;
+import java.security.MessageDigest;
+import java.util.Comparator;
+import java.util.HexFormat;
+import java.util.List;
+import java.util.Optional;
+import java.util.Properties;
+import java.util.concurrent.atomic.AtomicLong;
+import java.util.stream.Stream;
+
+/**
+ * 订阅缓存的分代存储：v2ray 与 clash 两份内容必须同时生效。
+ *
+ * <p>旧实现把两份文件各自原子替换，第二份失败时用户会拿到「新 v2ray + 旧 clash」的
+ * 混合版本。这里把两份文件写进同一个不可变分代目录，校验通过后用一次原子替换发布
+ * 指针（manifest），因此抓取、解析、写入或发布任何一步失败，都只会保留上一个完整分代。
+ *
+ * <p>目录结构：
+ * <pre>
+ * &lt;cacheRoot&gt;/&lt;accountId&gt;/current                      指针（原子替换，最后发布）
+ * &lt;cacheRoot&gt;/&lt;accountId&gt;/generations/&lt;generation&gt;/v2ray.txt
+ * &lt;cacheRoot&gt;/&lt;accountId&gt;/generations/&lt;generation&gt;/clash.yaml
+ * &lt;cacheRoot&gt;/&lt;accountId&gt;/v2ray.txt + clash.yaml          旧结构，两份都在才按 legacy 读
+ * </pre>
+ *
+ * <p>读取始终从指针指向的同一个分代里取两份文件，所以拿到的一定是同版本内容。
+ */
+@Slf4j
+public class SubscriptionCacheStore {
+    private static final String MANIFEST_FILE = "current";
+    private static final String GENERATION_DIR = "generations";
+    private static final String V2_FILE = "v2ray.txt";
+    private static final String CLASH_FILE = "clash.yaml";
+    /** 保留目录名，结构上永远不存在，用于表达「当前没有可分发版本」。 */
+    private static final String UNAVAILABLE_DIR = ".unavailable";
+    private static final String MANIFEST_GENERATION = "generation";
+    private static final String MANIFEST_V2_SHA = "v2-sha256";
+    private static final String MANIFEST_CLASH_SHA = "clash-sha256";
+    /** 旧结构读出来的版本的 generation 标记；真实分代名是定宽时间戳，不会与之碰撞。 */
+    private static final String LEGACY_GENERATION = "legacy";
+    /** 保留分代数：当前与上一个，足够回退且不会无限增长。 */
+    private static final int RETAINED_GENERATIONS = 2;
+    private static final AtomicLong SEQUENCE = new AtomicLong();
+
+    private final Path root;
+
+    public SubscriptionCacheStore(String cacheRoot) {
+        this(Paths.get(cacheRoot));
+    }
+
+    public SubscriptionCacheStore(Path root) {
+        this.root = root;
+    }
+
+    /** 同一个分代目录里读出的一对内容；两份文件必然同版本。 */
+    public record Pair(String generation, byte[] v2, byte[] clash) { }
+
+    public Path accountDir(int accountId) {
+        return root.resolve(String.valueOf(accountId));
+    }
+
+    /** 旧结构的账号根文件路径。仅供兼容测试定位 legacy 文件，不作为回退目标。 */
+    public Path legacyPath(int accountId, String client) {
+        return accountDir(accountId).resolve(fileName(client));
+    }
+
+    /**
+     * 没有完整可分发版本时返回的哨兵路径。
+     *
+     * <p>它保证不存在，因此调用方用 {@code Files.isRegularFile} 判定就会走「订阅暂不可用」
+     * 分支，而不会被半份 legacy 文件骗过。
+     */
+    public Path unavailablePath(int accountId, String client) {
+        return accountDir(accountId).resolve(UNAVAILABLE_DIR).resolve(fileName(client));
+    }
+
+    private static String fileName(String client) {
+        return "v2".equals(client) ? V2_FILE : CLASH_FILE;
+    }
+
+    /**
+     * 写入一份新分代并在最后发布指针。
+     *
+     * <p>失败时抛出异常，且不改变指针与上一个分代，调用方据此保留上次的完整版本。
+     */
+    public Pair publish(int accountId, byte[] v2, byte[] clash) throws IOException {
+        Path account = accountDir(accountId);
+        Files.createDirectories(account);
+        String generation = nextGeneration();
+        Path dir = account.resolve(GENERATION_DIR).resolve(generation);
+        boolean published = false;
+        try {
+            Files.createDirectories(dir);
+            writeFile(dir.resolve(V2_FILE), v2);
+            writeFile(dir.resolve(CLASH_FILE), clash);
+            // 发布前校验：两份文件都必须能按写入内容读回，避免把半截内容挂上指针。
+            verify(dir.resolve(V2_FILE), v2);
+            verify(dir.resolve(CLASH_FILE), clash);
+            publishManifest(account, generation, sha256(v2), sha256(clash));
+            published = true;
+        } finally {
+            if (!published)
+                deleteRecursively(dir);
+        }
+        prune(account, generation);
+        removeLegacy(account);
+        return new Pair(generation, v2, clash);
+    }
+
+    /** 读取当前发布内容；只有构成一整个有效版本时才返回。 */
+    public Optional<Pair> read(int accountId) {
+        return Optional.ofNullable(load(accountDir(accountId)));
+    }
+
+    /**
+     * 解析当前应分发的单个文件。
+     *
+     * <p>与 {@link #read}、{@link #hasComplete} 使用同一份配对契约：只有两份文件
+     * 齐全且哈希校验通过时才返回路径，因此分发路径不会发出半份或已损坏的缓存。
+     */
+    public Optional<Path> resolve(int accountId, String client) {
+        Path account = accountDir(accountId);
+        Pair pair = load(account);
+        if (pair == null)
+            return Optional.empty();
+        Path path = LEGACY_GENERATION.equals(pair.generation())
+                ? account.resolve(fileName(client))
+                : account.resolve(GENERATION_DIR).resolve(pair.generation()).resolve(fileName(client));
+        return Files.isRegularFile(path) ? Optional.of(path) : Optional.empty();
+    }
+
+    /** 是否已有可分发的一对内容（分代两份都在且哈希匹配，或旧结构两份都在）。 */
+    public boolean hasComplete(int accountId) {
+        return load(accountDir(accountId)) != null;
+    }
+
+    /** 删除该账号的全部缓存（指针、分代、旧结构文件），使后续请求判定为无缓存。 */
+    public void invalidate(int accountId) {
+        Path account = accountDir(accountId);
+        deleteIfExists(account.resolve(MANIFEST_FILE));
+        deleteRecursively(account.resolve(GENERATION_DIR));
+        deleteIfExists(account.resolve(V2_FILE));
+        deleteIfExists(account.resolve(CLASH_FILE));
+    }
+
+    /** 写入接缝：测试覆写以模拟第二份文件写入失败。 */
+    protected void writeFile(Path target, byte[] data) throws IOException {
+        atomicWrite(target, data);
+    }
+
+    /** 指针发布接缝：测试覆写以模拟 manifest 发布失败。 */
+    protected void publishManifest(Path accountDir, String generation, String v2Sha256, String clashSha256)
+            throws IOException {
+        atomicWrite(accountDir.resolve(MANIFEST_FILE),
+                manifestText(generation, v2Sha256, clashSha256).getBytes(StandardCharsets.UTF_8));
+    }
+
+    /**
+     * 读取「一整个版本」：一次读入指针快照，再按该快照定位分代并校验双文件哈希。
+     *
+     * <p>指针只读一次，因此并发 {@link #publish} 切换指针时不会出现「用新分代名配旧哈希」
+     * 或哈希字段缺失导致的空指针；要么拿到完整的新版本，要么拿到完整的旧版本。
+     */
+    private Pair load(Path account) {
+        Properties manifest = readManifest(account);
+        if (manifest == null) {
+            // 指针文件不存在才是「旧结构」。指针存在但读不出来说明它已损坏，同样按无缓存处理。
+            return Files.isRegularFile(account.resolve(MANIFEST_FILE)) ? null : loadLegacy(account);
+        }
+        Pair generation = loadGeneration(account, manifest);
+        if (generation != null)
+            return generation;
+        // 指针存在却读不出完整版本（半写、被篡改、内容缺失）时不得回退到 legacy：
+        // 那会发出一个比指针更旧的版本，掩盖损坏。此时按无可用缓存处理。
+        return null;
+    }
+
+    private Pair loadGeneration(Path account, Properties manifest) {
+        String generation = manifest.getProperty(MANIFEST_GENERATION);
+        if (generation == null || generation.isBlank() || generation.contains("/") || generation.contains("..")) {
+            log.warn("订阅缓存指针内容非法");
+            return null;
+        }
+        String v2Sha256 = manifest.getProperty(MANIFEST_V2_SHA);
+        String clashSha256 = manifest.getProperty(MANIFEST_CLASH_SHA);
+        if (v2Sha256 == null || clashSha256 == null) {
+            log.warn("订阅缓存指针缺少哈希字段 generation={}", generation);
+            return null;
+        }
+        Path dir = account.resolve(GENERATION_DIR).resolve(generation);
+        byte[] v2 = readFile(dir.resolve(V2_FILE));
+        byte[] clash = readFile(dir.resolve(CLASH_FILE));
+        if (v2 == null || clash == null) {
+            log.warn("订阅缓存分代内容缺失 generation={}", generation);
+            return null;
+        }
+        if (!matches(v2, v2Sha256) || !matches(clash, clashSha256)) {
+            log.warn("订阅缓存分代内容校验失败 generation={}", generation);
+            return null;
+        }
+        return new Pair(generation, v2, clash);
+    }
+
+    /** 旧结构：只有账号根下两份文件都在才构成一个有效版本。 */
+    private Pair loadLegacy(Path account) {
+        byte[] v2 = readFile(account.resolve(V2_FILE));
+        byte[] clash = readFile(account.resolve(CLASH_FILE));
+        if (v2 == null || clash == null)
+            return null;
+        return new Pair(LEGACY_GENERATION, v2, clash);
+    }
+
+    /** 指针读取接缝：包内可见，便于测试统计「每次查询只读一次指针」。 */
+    Properties readManifest(Path account) {
+        Path manifest = account.resolve(MANIFEST_FILE);
+        if (!Files.isRegularFile(manifest))
+            return null;
+        try {
+            Properties properties = new Properties();
+            properties.load(new StringReader(Files.readString(manifest, StandardCharsets.UTF_8)));
+            return properties;
+        } catch (IOException e) {
+            log.warn("读取订阅缓存指针失败: {}", e.getMessage());
+            return null;
+        }
+    }
+
+    private static byte[] readFile(Path path) {
+        try {
+            return Files.isRegularFile(path) ? Files.readAllBytes(path) : null;
+        } catch (IOException e) {
+            log.warn("读取订阅缓存文件失败 path={} error={}", path, e.getMessage());
+            return null;
+        }
+    }
+
+    private static void verify(Path path, byte[] expected) throws IOException {
+        byte[] written = Files.readAllBytes(path);
+        if (written.length != expected.length || !MessageDigest.isEqual(written, expected))
+            throw new IOException("订阅缓存写入校验失败 " + path.getFileName());
+    }
+
+    private static boolean matches(byte[] content, String expectedSha256) {
+        return expectedSha256 != null && expectedSha256.equals(sha256(content));
+    }
+
+    private static String manifestText(String generation, String v2Sha256, String clashSha256) {
+        return MANIFEST_GENERATION + "=" + generation + "\n"
+                + MANIFEST_V2_SHA + "=" + v2Sha256 + "\n"
+                + MANIFEST_CLASH_SHA + "=" + clashSha256 + "\n";
+    }
+
+    /** 分代名定宽可排序：时间戳在前保证单调，序号消除同毫秒碰撞。 */
+    private static String nextGeneration() {
+        return String.format("%013d-%06d", System.currentTimeMillis(), SEQUENCE.incrementAndGet() % 1_000_000);
+    }
+
+    private void prune(Path account, String current) {
+        Path dir = account.resolve(GENERATION_DIR);
+        List<Path> generations;
+        try (Stream<Path> entries = Files.list(dir)) {
+            generations = entries.filter(Files::isDirectory)
+                    .sorted(Comparator.comparing(path -> path.getFileName().toString(), Comparator.reverseOrder()))
+                    .toList();
+        } catch (IOException e) {
+            log.warn("清理订阅缓存分代失败 path={} error={}", dir, e.getMessage());
+            return;
+        }
+        for (int index = RETAINED_GENERATIONS; index < generations.size(); index++) {
+            Path stale = generations.get(index);
+            if (!stale.getFileName().toString().equals(current))
+                deleteRecursively(stale);
+        }
+    }
+
+    /** 分代发布成功后旧结构文件被取代，避免两份来源同时存在。 */
+    private void removeLegacy(Path account) {
+        deleteIfExists(account.resolve(V2_FILE));
+        deleteIfExists(account.resolve(CLASH_FILE));
+    }
+
+    private static void atomicWrite(Path target, byte[] data) throws IOException {
+        Path temp = target.resolveSibling(target.getFileName() + ".tmp");
+        Files.write(temp, data);
+        try {
+            Files.move(temp, target, StandardCopyOption.ATOMIC_MOVE, StandardCopyOption.REPLACE_EXISTING);
+        } catch (AtomicMoveNotSupportedException e) {
+            Files.move(temp, target, StandardCopyOption.REPLACE_EXISTING);
+        }
+    }
+
+    private static void deleteIfExists(Path path) {
+        try {
+            Files.deleteIfExists(path);
+        } catch (IOException e) {
+            log.warn("删除订阅缓存文件失败 path={} error={}", path, e.getMessage());
+        }
+    }
+
+    private static void deleteRecursively(Path path) {
+        if (!Files.exists(path))
+            return;
+        try (Stream<Path> entries = Files.walk(path)) {
+            entries.sorted(Comparator.reverseOrder()).forEach(SubscriptionCacheStore::deleteIfExists);
+        } catch (IOException e) {
+            log.warn("删除订阅缓存目录失败 path={} error={}", path, e.getMessage());
+        }
+    }
+
+    private static String sha256(byte[] input) {
+        try {
+            return HexFormat.of().formatHex(MessageDigest.getInstance("SHA-256").digest(input));
+        } catch (Exception e) {
+            throw new IllegalStateException(e);
+        }
+    }
+}