fix(实例路由): 转发失败不再伪装成「任务不存活」

归属节点滚动重启时跨节点转发会连接失败，原来返回 ApiResponse.fail(40903) ——
HTTP 200 + data:null。客户端把「拿不到数据」解析成 alive=false，于是把健康的长
任务主动停掉（2026-09-18 任务 28616：跑到 66/253 被自杀，只跑到第 2/6 页）。

改为 503 + 空 body：新客户端按状态码判未知继续跑；老客户端因 body 不是 JSON、
resp.json() 抛异常也落到未知——两边都不会再把未知当成死。顺带让这类故障在
HTTP 指标里可见（原先记成 200，滚动重启期间丢了多少心跳监控完全看不到）。

admin-frontend-vue/e2e/records-tutorial.spec.ts

@@ -0,0 +1,79 @@
+import { expect, test, type Page } from '@playwright/test'
+
+// 教程管理页验收（module 记录与版本）：列表新增「版本号」列 + 上传时间列，
+// 默认按上传时间降序；版本号/上传时间表头可点击切换升降序；空版本历史行沉底。
+// 依赖 scripts/mock-admin-server.mjs 的教程包夹具（4 条，其中 1 条无版本号、1 条 09-01 历史行）。
+
+async function openTutorial(page: Page) {
+  await page.goto('/admin-vue/records/tutorial')
+  await expect(page.locator('.admin-topbar h1')).toHaveText('教程管理')
+  await expect(page.locator('.panel-box tbody tr').first()).toBeVisible()
+}
+
+const versionTexts = (page: Page) =>
+  page.locator('.panel-box tbody tr td:nth-child(3)').allTextContents()
+
+const fileTexts = (page: Page) =>
+  page.locator('.panel-box tbody tr td:nth-child(2) .file-name').allTextContents()
+
+test('test_tutorial_list_default_desc_by_upload_time', async ({ page }) => {
+  await openTutorial(page)
+
+  // 表头：版本号、上传时间（均带排序标记）
+  await expect(page.locator('.sort-version')).toHaveText(/版本号/)
+  await expect(page.locator('.sort-time')).toHaveText(/上传时间/)
+  // 默认排序状态：上传时间降序（▼），版本号未激活（▲▼）
+  await expect(page.locator('.sort-time .sort-mark')).toHaveText('▼')
+  await expect(page.locator('.sort-version .sort-mark')).toHaveText('▲▼')
+
+  // 默认按上传时间降序：09-14 → 09-12 → 09-10 → 09-01；空版本行沉底
+  expect(await versionTexts(page)).toEqual(['v3.10.0', 'v3.2.0', 'v3.1.0', '—'])
+  expect(await fileTexts(page)).toEqual([
+    '数富AI-教学客户端-v4.zip',
+    '数富AI-教学客户端-v3.zip',
+    '数富AI-教学客户端-v2.zip',
+    '数富AI-教学客户端.zip',
+  ])
+  // 当前生效 = 最新上传（夹具里 09-14 那条），不随列表排序变化
+  await expect(page.locator('.tag-active')).toHaveCount(1)
+  await expect(page.locator('.tag-active').locator('xpath=..')).toHaveText(/数富AI-教学客户端-v4\.zip/)
+})
+
+test('test_tutorial_sort_toggle_by_version_and_time', async ({ page }) => {
+  await openTutorial(page)
+
+  // 点「版本号」：首次为降序（v3.10.0 按自然序大于 v3.2.0），空版本行仍在末尾
+  await page.locator('.sort-version').click()
+  await expect(page.locator('.sort-version .sort-mark')).toHaveText('▼')
+  expect(await versionTexts(page)).toEqual(['v3.10.0', 'v3.2.0', 'v3.1.0', '—'])
+
+  // 再点一次切升序
+  await page.locator('.sort-version').click()
+  await expect(page.locator('.sort-version .sort-mark')).toHaveText('▲')
+  expect(await versionTexts(page)).toEqual(['v3.1.0', 'v3.2.0', 'v3.10.0', '—'])
+
+  // 点回「上传时间」：切换字段时回到降序默认
+  await page.locator('.sort-time').click()
+  await expect(page.locator('.sort-time .sort-mark')).toHaveText('▼')
+  await expect(page.locator('.sort-version .sort-mark')).toHaveText('▲▼')
+  expect(await fileTexts(page)).toEqual([
+    '数富AI-教学客户端-v4.zip',
+    '数富AI-教学客户端-v3.zip',
+    '数富AI-教学客户端-v2.zip',
+    '数富AI-教学客户端.zip',
+  ])
+})
+
+test('test_tutorial_upload_dialog_has_version_field', async ({ page }) => {
+  await openTutorial(page)
+
+  await page.getByRole('button', { name: '上传教程包' }).click()
+  const dialog = page.locator('.el-dialog')
+  await expect(dialog).toBeVisible()
+  await expect(dialog.locator('.el-form-item').first()).toContainText('版本号')
+  await expect(dialog.locator('input').first()).toHaveAttribute('maxlength', '64')
+
+  // 不选文件直接提交：提示仍以 zip 为必填（版本号可空）
+  await dialog.getByRole('button', { name: '上传教程包' }).click()
+  await expect(dialog.locator('.el-alert')).toContainText('请选择 zip 压缩包')
+})

admin-frontend-vue/scripts/mock-admin-server.mjs

@@ -142,6 +142,15 @@ function json(res, payload, status = 200) {
   res.end(body)
 }
 
+/* ===================== 教程包夹具（教程管理页：版本号列 + 上传时间排序验收用） ===================== */
+// 故意乱序给出，且含一条无版本号的历史行：页面默认应按上传时间降序、空版本行沉底。
+const TUTORIAL_PACKAGES = [
+  { id: 3, file_name: '数富AI-教学客户端-v3.zip', version: 'v3.2.0', object_key: 'tutorial/20260912090000-数富AI-教学客户端-v3.zip', file_size: 1048576, file_url: 'https://oss.aishufu.top/client/tutorial/t3.zip', created_at: '2026-09-12 09:00' },
+  { id: 1, file_name: '数富AI-教学客户端.zip', version: '', object_key: 'tutorial/数富AI-教学客户端.zip', file_size: 0, file_url: 'https://oss.aishufu.top/client/tutorial/legacy.zip', created_at: '2026-09-01 08:00' },
+  { id: 4, file_name: '数富AI-教学客户端-v4.zip', version: 'v3.10.0', object_key: 'tutorial/20260914103000-数富AI-教学客户端-v4.zip', file_size: 2097152, file_url: 'https://oss.aishufu.top/client/tutorial/t4.zip', created_at: '2026-09-14 10:30' },
+  { id: 2, file_name: '数富AI-教学客户端-v2.zip', version: 'v3.1.0', object_key: 'tutorial/20260910120000-数富AI-教学客户端-v2.zip', file_size: 524288, file_url: 'https://oss.aishufu.top/client/tutorial/t2.zip', created_at: '2026-09-10 12:00' },
+]
+
 /* ===================== 站内通知夹具（铃铛面板：搜索/按天分组/分页验收用） ===================== */
 const NOTIFICATIONS = [
   ...[6, 5, 4, 3, 2, 1].map((seq) => notif(seq, '2026-09-13', 22, seq, seq <= 3)),
@@ -243,6 +252,11 @@ const server = createServer((req, res) => {
   if (url === '/api/admin/notifications') {
     return json(res, notificationPage(new URL(req.url || '/', 'http://127.0.0.1').searchParams))
   }
+  if (url === '/api/admin/tutorials') {
+    // 与 Java 侧同契约：created_at 降序返回（页面"当前生效"取第一条）。
+    const items = [...TUTORIAL_PACKAGES].sort((a, b) => (a.created_at < b.created_at ? 1 : -1))
+    return json(res, { success: true, data: { items } })
+  }
   if (url.startsWith('/api/')) {
     return json(res, { success: true, data: { items: [], total: 0 } })
   }

admin-frontend-vue/src/api/device-logs.ts

@@ -0,0 +1,135 @@
+import { http } from './http'
+import { unwrap } from './envelope'
+
+/** 日志文件行（桌面客户端 / 麦象采集机上报）。 */
+export interface DeviceLogFileRow {
+  id: number
+  source: string
+  deviceId: string
+  deviceName: string | null
+  username: string | null
+  uid: number | null
+  fileName: string
+  logDate: string
+  uploadedBytes: number
+  partCount: number
+  lastUploadAt: string | null
+  createdAt: string | null
+}
+
+export interface DeviceLogPage {
+  items: DeviceLogFileRow[]
+  total: number
+  page: number
+  pageSize: number
+  /** 云端日志保留天数（页面提示用）。 */
+  retentionDays: number
+}
+
+export interface DeviceLogContent {
+  fileId: number
+  fileName: string
+  content: string
+  totalBytes: number
+  shownBytes: number
+  truncated: boolean
+}
+
+export interface DeviceLogDevice {
+  source: string
+  deviceId: string
+  deviceName: string | null
+  lastUploadAt: string | null
+}
+
+export interface DeviceLogOverride {
+  id: number
+  source: string
+  deviceId: string
+  deviceName: string | null
+  mode: string
+  updatedAt: string | null
+}
+
+export interface DeviceLogConfigData {
+  globalMode: string
+  overrides: DeviceLogOverride[]
+}
+
+export interface DeviceLogQuery {
+  source?: string
+  keyword?: string
+  startDate?: string
+  endDate?: string
+  page: number
+  pageSize: number
+}
+
+/** 分页查询日志文件列表：GET /api/admin/device-logs/files */
+export async function fetchDeviceLogFiles(params: DeviceLogQuery): Promise<DeviceLogPage> {
+  const query: Record<string, string | number> = { page: params.page, pageSize: params.pageSize }
+  if (params.source) query.source = params.source
+  if (params.keyword) query.keyword = params.keyword
+  if (params.startDate) query.startDate = params.startDate
+  if (params.endDate) query.endDate = params.endDate
+  const { data } = await http.get('/api/admin/device-logs/files', { params: query })
+  return unwrap<DeviceLogPage>(data)
+}
+
+/** 查看日志尾部内容：GET /api/admin/device-logs/content */
+export async function fetchDeviceLogContent(fileId: number, maxBytes?: number): Promise<DeviceLogContent> {
+  const query: Record<string, number> = { fileId }
+  if (maxBytes) query.maxBytes = maxBytes
+  const { data } = await http.get('/api/admin/device-logs/content', { params: query })
+  return unwrap<DeviceLogContent>(data)
+}
+
+/** 完整日志下载地址（同域 cookie 鉴权，直接给 a[href] 或 window.open 用）。 */
+export function deviceLogDownloadUrl(fileId: number): string {
+  return `/api/admin/device-logs/download?fileId=${fileId}`
+}
+
+/** 删除日志文件（片段与元数据，不可恢复）：DELETE /api/admin/device-logs/{id} */
+export async function deleteDeviceLogFile(id: number): Promise<void> {
+  const { data } = await http.delete(`/api/admin/device-logs/${id}`)
+  unwrap<unknown>(data)
+}
+
+/** 采集配置（全局默认 + 终端覆盖）：GET /api/admin/device-logs/config */
+export async function fetchDeviceLogConfig(keyword?: string): Promise<DeviceLogConfigData> {
+  const { data } = await http.get('/api/admin/device-logs/config', {
+    params: keyword ? { keyword } : undefined,
+  })
+  return unwrap<DeviceLogConfigData>(data)
+}
+
+/** 最近上报过的终端（覆盖选择用）：GET /api/admin/device-logs/devices */
+export async function fetchDeviceLogDevices(): Promise<DeviceLogDevice[]> {
+  const { data } = await http.get('/api/admin/device-logs/devices')
+  return unwrap<DeviceLogDevice[]>(data)
+}
+
+/** 设置全局采集模式：PUT /api/admin/device-logs/config/global */
+export async function updateDeviceLogGlobalMode(mode: string): Promise<void> {
+  const { data } = await http.put('/api/admin/device-logs/config/global', undefined, { params: { mode } })
+  unwrap<unknown>(data)
+}
+
+/** 设置/更新终端覆盖：PUT /api/admin/device-logs/config/device */
+export async function updateDeviceLogOverride(
+  source: string,
+  deviceId: string,
+  deviceName: string | null,
+  mode: string,
+): Promise<void> {
+  const { data } = await http.put('/api/admin/device-logs/config/device', undefined, {
+    params: { source, deviceId, deviceName: deviceName || undefined, mode },
+  })
+  unwrap<unknown>(data)
+}
+
+/** 删除终端覆盖（回落到全局默认）：DELETE /api/admin/device-logs/config/device/{id} */
+export async function deleteDeviceLogOverride(id: number): Promise<void> {
+  const { data } = await http.delete(`/api/admin/device-logs/config/device/${id}`)
+  unwrap<unknown>(data)
+}

admin-frontend-vue/src/components/NotificationBell.vue

@@ -57,19 +57,18 @@
           @input="onKeywordInput"
         />
         <div class="bell-search-days">
-          <input
+          <el-date-picker
-            v-model="startDate"
+            :model-value="dateRangeValue"
-            type="date"
+            type="daterange"
-            class="bell-date-input"
+            value-format="YYYY-MM-DD"
-            aria-label="起始日期"
+            start-placeholder="开始日期"
-            @change="reload"
+            end-placeholder="结束日期"
-          />
+            range-separator="至"
-          <span class="bell-date-sep">至</span>
+            size="small"
-          <input
+            class="bell-date-picker"
-            v-model="endDate"
+            popper-class="bell-date-popper"
-            type="date"
+            :clearable="true"
-            class="bell-date-input"
+            @update:model-value="onDateRangeChange"
-            aria-label="结束日期"
             @change="reload"
           />
           <button v-if="hasFilter" type="button" class="bell-search-reset" @click="resetFilters">
@@ -171,6 +170,23 @@ const groupedItems = computed(() => groupNotificationsByDay(items.value))
 const totalPages = computed(() => Math.max(1, Math.ceil(total.value / PAGE_SIZE)))
 const hasFilter = computed(() => Boolean(keyword.value.trim() || startDate.value || endDate.value))
 
+/**
+ * el-date-picker 的绑定桥：日期区间在内部仍用 startDate/endDate 两个 ref 表示
+ * （hasFilter/normalizedDayRange/resetFilters 都基于它们）。
+ * 此前模板写的是 v-model="dateRange"，而 dateRange 从未声明 → vue-tsc 直接报 TS2339 构建失败，
+ * 且日期筛选完全不下发参数。
+ */
+const dateRangeValue = computed<[string, string] | null>(() =>
+  startDate.value && endDate.value ? [startDate.value, endDate.value] : null
+)
+
+function onDateRangeChange(value: [string, string] | null) {
+  const start = value?.[0]
+  const end = value?.[1]
+  startDate.value = start ? String(start) : ''
+  endDate.value = end ? String(end) : ''
+}
+
 let pollTimer: number | null = null
 let searchTimer: number | null = null
 

admin-frontend-vue/src/pages/account/UserMenuAuthTree.vue

@@ -77,7 +77,7 @@ watch(
             node-key="id"
             show-checkbox
             default-expand-all
-            :props="{ label: 'name', children: 'children' }"
+            :props="{ label: 'name', children: 'children', disabled: 'disabled' }"
             @check="onCheck"
           />
         </div>
@@ -89,7 +89,7 @@ watch(
             node-key="id"
             show-checkbox
             default-expand-all
-            :props="{ label: 'name', children: 'children' }"
+            :props="{ label: 'name', children: 'children', disabled: 'disabled' }"
             @check="onCheck"
           />
         </div>

admin-frontend-vue/src/pages/account/user-menu-auth.ts

@@ -16,6 +16,11 @@ export interface MenuOptionNode {
   parentId: number | null
   /** 所属菜单类型（admin 后台 / app 前端客户端），提交时按类型分区落库。 */
   type: string
+  /**
+   * 当前操作者无权授予（后端 grantable=false）时置灰：仍展示并回显已勾选，
+   * 但不允许改勾选。非超管只能授自己已有的菜单，勾到越权项会让整笔保存回滚。
+   */
+  disabled?: boolean
   children?: MenuOptionNode[]
 }
 
@@ -39,6 +44,8 @@ export function parsePermissionMenuItem(raw: unknown, type = ''): MenuOptionNode
     sort: sortRaw === null ? 0 : sortRaw,
     parentId: parentId === null ? null : parentId,
     type,
+    // 缺省（菜单管理页等未标记的接口）按可授予处理，保持旧行为
+    disabled: record.grantable === false,
   }
 }
 

admin-frontend-vue/src/pages/asin/DedupeRegistryPage.vue

@@ -43,6 +43,10 @@ const lockedGroupId = computed<number | null>(() => {
   return groups.value.length === 1 ? groups.value[0].id : null
 })
 const jumpPage = ref('')
+/** 顺序翻页游标：上一页响应给的 nextLastId；仅在"下一页"时使用，跳页/筛选时清空走 OFFSET。 */
+const pageCursor = ref<number | null>(null)
+/** 本次请求实际使用的游标（load 时决定） */
+let pendingCursor: number | null = null
 
 const totalPages = computed(() => Math.max(1, Math.ceil(total.value / pageSize.value)))
 
@@ -85,10 +89,15 @@ function stopExportWait(): void {
 async function load(): Promise<void> {
   loading.value = true
   try {
-    const result = await fetchDedupeTotalList(toDedupeListParams(filter, page.value, pageSize.value))
+    const result = await fetchDedupeTotalList(
+      toDedupeListParams(filter, page.value, pageSize.value, pendingCursor),
+    )
     rows.value = result.items
     total.value = result.total
     if (result.page >= 1) page.value = result.page
+    // 本页响应回传的游标留作"下一页"用；空页则清空（没有更多）
+    pageCursor.value = result.nextLastId ?? null
+    pendingCursor = null
   } catch (error) {
     ElMessage.error(error instanceof Error ? error.message : '去重数据加载失败')
   } finally {
@@ -98,11 +107,15 @@ async function load(): Promise<void> {
 
 function apply(): void {
   page.value = 1
+  pageCursor.value = null
+  pendingCursor = null
   void load()
 }
 
 function changePage(next: number): void {
   if (next < 1 || next > totalPages.value) return
+  // 只有"顺序下一页"用 keyset 游标（避免深分页 offset）；跳页/回退走 OFFSET，行为不变
+  pendingCursor = next === page.value + 1 ? pageCursor.value : null
   page.value = next
   void load()
 }
@@ -119,6 +132,8 @@ function goJump(): void {
 function changeSize(size: number) {
   pageSize.value = size
   page.value = 1
+  pageCursor.value = null
+  pendingCursor = null
   void load()
 }
 

admin-frontend-vue/src/pages/asin/asin-filter.ts

@@ -24,6 +24,8 @@ export interface AsinListParams {
   groupId?: number | null
   /** 国家代码（如 DE、UK）。 */
   country?: string
+  /** 顺序翻页游标（上一页返回的 nextLastId）：传了就忽略 page 偏移，走 keyset。 */
+  lastId?: number
 }
 
 /** 序列化到 Java 分页接口的查询参数（snake_case）。 */
@@ -36,6 +38,8 @@ export interface AsinPageQuery {
   end_date?: string
   group_id?: number
   country?: string
+  /** 顺序翻页游标（keyset）：传了就忽略 page 偏移。 */
+  last_id?: number
 }
 
 function finiteInt(value: unknown): number | null {
@@ -77,5 +81,6 @@ export function toAsinPageQuery(params: AsinListParams): AsinPageQuery {
   if (params.endDate) query.end_date = params.endDate
   if (params.groupId != null) query.group_id = params.groupId
   if (params.country) query.country = params.country
+  if (typeof params.lastId === 'number' && params.lastId > 0) query.last_id = params.lastId
   return query
 }

admin-frontend-vue/src/pages/asin/dedupe-total-filter.ts

@@ -30,8 +30,10 @@ export function toDedupeListParams(
   state: DedupeTotalFilterState,
   page: number,
   pageSize: number,
+  lastId?: number | null,
 ): AsinListParams {
   const params: AsinListParams = { page, pageSize }
+  if (typeof lastId === 'number' && lastId > 0) params.lastId = lastId
   const keyword = (state.keyword || '').trim()
   const username = (state.username || '').trim()
   const country = (state.country || '').trim()

admin-frontend-vue/src/pages/asin/dedupe-total-model.ts

@@ -18,6 +18,8 @@ export interface DedupeTotalPageResult {
   total: number
   page: number
   pageSize: number
+  /** 顺序翻页游标：本页最后一行 id；下一页回传它即可走 keyset。 */
+  nextLastId?: number
 }
 
 export function emptyDedupeTotalPage(): DedupeTotalPageResult {
@@ -65,6 +67,8 @@ export function parseDedupeTotalPage(payload: unknown): DedupeTotalPageResult {
   }
   if (typeof record.total === 'number') out.total = Math.floor(record.total)
   if (typeof record.page === 'number' && record.page >= 1) out.page = Math.floor(record.page)
+  const rawNextLastId = record.nextLastId ?? record.next_last_id
+  if (typeof rawNextLastId === 'number' && rawNextLastId > 0) out.nextLastId = Math.floor(rawNextLastId)
   const rawSize = record.pageSize ?? record.page_size
   if (typeof rawSize === 'number' && rawSize >= 1) out.pageSize = Math.floor(rawSize)
   return out

admin-frontend-vue/src/pages/records/DeviceLogsPage.vue

@@ -0,0 +1,719 @@
+<script setup lang="ts">
+/** 日志管理页：桌面客户端 / 麦象采集机上报的日志浏览（仅超管）。
+ *  支持来源/日期/关键字筛选，尾部内容查看（自动滚到底、可向前加载更早内容）、
+ *  完整下载、删除；「采集配置」可调全量/精选模式（全局默认 + 终端覆盖）。 */
+import { computed, nextTick, onMounted, ref } from 'vue'
+import OldPagination from '@/components/OldPagination.vue'
+import { ElMessage } from 'element-plus'
+import { formatDateTime } from '@/utils/datetime'
+import {
+  deleteDeviceLogFile,
+  deleteDeviceLogOverride,
+  deviceLogDownloadUrl,
+  fetchDeviceLogConfig,
+  fetchDeviceLogContent,
+  fetchDeviceLogDevices,
+  fetchDeviceLogFiles,
+  updateDeviceLogGlobalMode,
+  updateDeviceLogOverride,
+  type DeviceLogContent,
+  type DeviceLogDevice,
+  type DeviceLogFileRow,
+  type DeviceLogOverride,
+} from '@/api/device-logs'
+
+const SOURCE_OPTIONS = [
+  { value: '', label: '全部来源' },
+  { value: 'client', label: '桌面客户端' },
+  { value: 'maixiang', label: '麦象采集机' },
+]
+
+function sourceLabel(source: string): string {
+  if (source === 'client') return '桌面客户端'
+  if (source === 'maixiang') return '麦象采集机'
+  return source || '—'
+}
+
+function formatBytes(bytes: number | null | undefined): string {
+  if (bytes == null || !Number.isFinite(bytes) || bytes <= 0) return '—'
+  if (bytes < 1024) return `${bytes} B`
+  const units = ['KB', 'MB', 'GB']
+  let value = bytes / 1024
+  let unitIndex = 0
+  while (value >= 1024 && unitIndex < units.length - 1) {
+    value /= 1024
+    unitIndex += 1
+  }
+  return `${value >= 100 ? value.toFixed(0) : value.toFixed(1)} ${units[unitIndex]}`
+}
+
+const loading = ref(false)
+const rows = ref<DeviceLogFileRow[]>([])
+const total = ref(0)
+const page = ref(1)
+const pageSize = ref(15)
+const sourceFilter = ref('')
+const keyword = ref('')
+const startDate = ref('')
+const endDate = ref('')
+const retentionDays = ref(7)
+
+const totalPages = computed(() => Math.max(1, Math.ceil(total.value / pageSize.value)))
+
+async function load() {
+  loading.value = true
+  try {
+    const result = await fetchDeviceLogFiles({
+      source: sourceFilter.value || undefined,
+      keyword: keyword.value.trim() || undefined,
+      startDate: startDate.value || undefined,
+      endDate: endDate.value || undefined,
+      page: page.value,
+      pageSize: pageSize.value,
+    })
+    rows.value = result?.items || []
+    total.value = Number(result?.total || 0)
+    if (result?.retentionDays) retentionDays.value = result.retentionDays
+  } catch (error) {
+    ElMessage.error(error instanceof Error ? error.message : '日志列表加载失败')
+  } finally {
+    loading.value = false
+  }
+}
+
+function search() {
+  page.value = 1
+  load()
+}
+
+function reset() {
+  sourceFilter.value = ''
+  keyword.value = ''
+  startDate.value = ''
+  endDate.value = ''
+  page.value = 1
+  load()
+}
+
+function changePage(next: number) {
+  if (next < 1 || next > totalPages.value) return
+  page.value = next
+  load()
+}
+
+function changeSize(size: number) {
+  pageSize.value = size
+  page.value = 1
+  load()
+}
+
+// ---------------------------------------------------------------- 内容查看
+
+const TAIL_STEP = 256 * 1024
+const TAIL_MAX = 8 * 1024 * 1024
+
+const viewerVisible = ref(false)
+const viewerLoading = ref(false)
+const viewerFile = ref<DeviceLogFileRow | null>(null)
+const viewerContent = ref<DeviceLogContent | null>(null)
+const viewerMaxBytes = ref(TAIL_STEP)
+const viewerPre = ref<HTMLElement | null>(null)
+
+async function openViewer(row: DeviceLogFileRow) {
+  viewerFile.value = row
+  viewerMaxBytes.value = TAIL_STEP
+  viewerContent.value = null
+  viewerVisible.value = true
+  await loadContent(true)
+}
+
+async function loadContent(scrollToBottom: boolean) {
+  if (!viewerFile.value) return
+  viewerLoading.value = true
+  try {
+    viewerContent.value = await fetchDeviceLogContent(viewerFile.value.id, viewerMaxBytes.value)
+    if (scrollToBottom) {
+      await nextTick()
+      if (viewerPre.value) viewerPre.value.scrollTop = viewerPre.value.scrollHeight
+    }
+  } catch (error) {
+    ElMessage.error(error instanceof Error ? error.message : '日志内容加载失败')
+  } finally {
+    viewerLoading.value = false
+  }
+}
+
+function loadMore() {
+  viewerMaxBytes.value = Math.min(viewerMaxBytes.value * 2, TAIL_MAX)
+  loadContent(false)
+}
+
+async function remove(row: DeviceLogFileRow) {
+  if (!window.confirm(`确认删除「${row.fileName}」（${row.deviceName || row.deviceId}）的云端日志？删除后不可恢复。`)) return
+  try {
+    await deleteDeviceLogFile(row.id)
+    ElMessage.success('已删除')
+    load()
+  } catch (error) {
+    ElMessage.error(error instanceof Error ? error.message : '删除失败')
+  }
+}
+
+// ---------------------------------------------------------------- 采集配置
+
+const configVisible = ref(false)
+const configLoading = ref(false)
+const globalMode = ref('full')
+const overrides = ref<DeviceLogOverride[]>([])
+const devices = ref<DeviceLogDevice[]>([])
+const newOverrideKey = ref('')
+const newOverrideMode = ref('selected')
+
+const MODE_OPTIONS = [
+  { value: 'full', label: '全量采集' },
+  { value: 'selected', label: '精选采集' },
+]
+
+function modeLabel(mode: string): string {
+  return mode === 'selected' ? '精选' : '全量'
+}
+
+async function openConfig() {
+  configVisible.value = true
+  await loadConfig()
+}
+
+async function loadConfig() {
+  configLoading.value = true
+  try {
+    const data = await fetchDeviceLogConfig()
+    globalMode.value = data.globalMode || 'full'
+    overrides.value = data.overrides || []
+    devices.value = (await fetchDeviceLogDevices()) || []
+  } catch (error) {
+    ElMessage.error(error instanceof Error ? error.message : '采集配置加载失败')
+  } finally {
+    configLoading.value = false
+  }
+}
+
+async function saveGlobalMode(mode: string) {
+  if (mode === globalMode.value) return
+  try {
+    await updateDeviceLogGlobalMode(mode)
+    globalMode.value = mode
+    ElMessage.success(`全局采集模式已切换为「${modeLabel(mode)}」`)
+  } catch (error) {
+    ElMessage.error(error instanceof Error ? error.message : '保存失败')
+  }
+}
+
+async function addOverride() {
+  if (!newOverrideKey.value) {
+    ElMessage.warning('请先选择终端')
+    return
+  }
+  const [source, deviceId] = newOverrideKey.value.split('|')
+  const device = devices.value.find((item) => item.source === source && item.deviceId === deviceId)
+  try {
+    await updateDeviceLogOverride(source, deviceId, device?.deviceName || null, newOverrideMode.value)
+    ElMessage.success('终端覆盖已保存')
+    newOverrideKey.value = ''
+    await loadConfig()
+  } catch (error) {
+    ElMessage.error(error instanceof Error ? error.message : '保存失败')
+  }
+}
+
+async function removeOverride(row: DeviceLogOverride) {
+  const who = row.deviceName || row.deviceId
+  if (!window.confirm(`确认删除终端「${who}」的采集覆盖（回落到全局默认）？`)) return
+  try {
+    await deleteDeviceLogOverride(row.id)
+    ElMessage.success('已删除')
+    await loadConfig()
+  } catch (error) {
+    ElMessage.error(error instanceof Error ? error.message : '删除失败')
+  }
+}
+
+onMounted(load)
+</script>
+
+<template>
+  <div class="device-logs-view">
+    <section class="panel-box">
+      <div class="logs-head">
+        <h3>日志文件列表</h3>
+        <div class="logs-head-tools">
+          <span class="retention-tip">云端仅保留 {{ retentionDays }} 天</span>
+          <button class="btn btn-ghost" type="button" @click="openConfig">采集配置</button>
+          <button class="btn" type="button" @click="load">刷新</button>
+        </div>
+      </div>
+
+      <div class="form-row logs-filter-row">
+        <div class="form-group" style="min-width: 150px">
+          <label>来源</label>
+          <select v-model="sourceFilter">
+            <option v-for="option in SOURCE_OPTIONS" :key="option.value" :value="option.value">{{ option.label }}</option>
+          </select>
+        </div>
+        <div class="form-group" style="min-width: 220px">
+          <label>设备 / 文件名</label>
+          <input v-model="keyword" type="text" placeholder="模糊搜索设备名、设备ID、文件名" @keyup.enter="search" />
+        </div>
+        <div class="form-group" style="min-width: 150px">
+          <label>日志日期（起）</label>
+          <input v-model="startDate" type="date" />
+        </div>
+        <div class="form-group" style="min-width: 150px">
+          <label>日志日期（止）</label>
+          <input v-model="endDate" type="date" />
+        </div>
+        <div class="form-group">
+          <label>&nbsp;</label>
+          <div class="filter-actions">
+            <button class="btn" type="button" @click="search">查询</button>
+            <button class="btn btn-ghost" type="button" @click="reset">重置</button>
+          </div>
+        </div>
+      </div>
+
+      <div class="logs-table-scroll">
+        <table>
+          <thead>
+            <tr>
+              <th style="width: 110px">来源</th>
+              <th style="width: 230px">设备</th>
+              <th style="width: 130px">用户</th>
+              <th style="width: 230px">文件名</th>
+              <th style="width: 110px">日志日期</th>
+              <th style="width: 100px">已收大小</th>
+              <th style="width: 80px">片段数</th>
+              <th style="width: 170px">最后更新</th>
+              <th style="width: 190px">操作</th>
+            </tr>
+          </thead>
+          <tbody>
+            <template v-if="rows.length">
+              <tr v-for="row in rows" :key="row.id">
+                <td>
+                  <span class="source-pill" :class="row.source === 'maixiang' ? 'is-maixiang' : 'is-client'">
+                    {{ sourceLabel(row.source) }}
+                  </span>
+                </td>
+                <td>
+                  <span class="device-name" :title="row.deviceId">{{ row.deviceName || '—' }}</span>
+                  <span class="device-id" :title="row.deviceId">{{ row.deviceId }}</span>
+                </td>
+                <td>
+                  <span v-if="row.username" class="user-name">{{ row.username }}</span>
+                  <span v-else-if="row.uid" class="user-name">UID {{ row.uid }}</span>
+                  <span v-else class="dim">—</span>
+                </td>
+                <td>
+                  <span class="file-name" :title="row.fileName">{{ row.fileName }}</span>
+                </td>
+                <td>{{ row.logDate }}</td>
+                <td>{{ formatBytes(row.uploadedBytes) }}</td>
+                <td>{{ row.partCount ?? 0 }}</td>
+                <td>{{ row.lastUploadAt ? formatDateTime(row.lastUploadAt) : '—' }}</td>
+                <td class="ops-cell">
+                  <button class="btn btn-sm" type="button" @click="openViewer(row)">查看</button>
+                  <a class="btn btn-sm btn-ghost dl-link" :href="deviceLogDownloadUrl(row.id)" download>下载</a>
+                  <button class="btn btn-sm btn-danger" type="button" @click="remove(row)">删除</button>
+                </td>
+              </tr>
+            </template>
+            <tr v-else-if="loading">
+              <td colspan="9" class="empty-tip">加载中...</td>
+            </tr>
+            <tr v-else>
+              <td colspan="9" class="empty-tip">
+                {{ keyword || sourceFilter || startDate || endDate ? '暂无匹配日志' : '暂无日志上报（客户端/采集机上报后自动出现在这里）' }}
+              </td>
+            </tr>
+          </tbody>
+        </table>
+      </div>
+      <OldPagination :total="total" :page="page" :page-size="pageSize" @change="changePage" @size-change="changeSize" />
+    </section>
+
+    <el-dialog v-model="viewerVisible" :title="viewerFile ? `${viewerFile.fileName}（${viewerFile.deviceName || viewerFile.deviceId}）` : '日志内容'" width="900px" top="5vh">
+      <div class="viewer-toolbar">
+        <span class="viewer-meta">
+          云端已收 {{ formatBytes(viewerContent?.totalBytes ?? viewerFile?.uploadedBytes) }}
+          <template v-if="viewerContent"> · 当前展示 {{ formatBytes(viewerContent.shownBytes) }}</template>
+          <template v-if="viewerContent?.truncated"> · 更早内容未加载</template>
+        </span>
+        <span class="viewer-actions">
+          <button class="btn btn-sm btn-ghost" type="button" :disabled="viewerLoading" @click="loadContent(true)">刷新</button>
+          <button
+            class="btn btn-sm btn-ghost"
+            type="button"
+            :disabled="viewerLoading || !viewerContent?.truncated || viewerMaxBytes >= TAIL_MAX"
+            @click="loadMore"
+          >加载更早</button>
+          <a v-if="viewerFile" class="btn btn-sm btn-ghost dl-link" :href="deviceLogDownloadUrl(viewerFile.id)" download>下载完整日志</a>
+        </span>
+      </div>
+      <pre ref="viewerPre" class="log-pre">{{ viewerLoading && !viewerContent ? '加载中...' : (viewerContent?.content || '（暂无内容）') }}</pre>
+    </el-dialog>
+
+    <el-dialog v-model="configVisible" title="采集配置" width="720px">
+      <div class="config-block">
+        <h4>全局默认模式</h4>
+        <p class="config-desc">对未单独配置的终端生效。全量=上传日志目录内全部文件；精选=排除低价值大日志（客户端排除 pywebview；麦象排除 kk-browser / 控制台 / 测试日志）。终端在下一次上报周期（约 1 分钟内）跟随新配置。</p>
+        <div class="mode-switch">
+          <button
+            v-for="option in MODE_OPTIONS"
+            :key="option.value"
+            class="mode-btn"
+            :class="{ 'is-active': globalMode === option.value }"
+            type="button"
+            @click="saveGlobalMode(option.value)"
+          >{{ option.label }}</button>
+        </div>
+      </div>
+
+      <div class="config-block">
+        <h4>终端覆盖</h4>
+        <div class="override-add">
+          <select v-model="newOverrideKey" class="override-select">
+            <option value="">选择终端（最近上报的设备）</option>
+            <option v-for="device in devices" :key="`${device.source}|${device.deviceId}`" :value="`${device.source}|${device.deviceId}`">
+              {{ sourceLabel(device.source) }} · {{ device.deviceName || device.deviceId }}（{{ device.deviceId }}）
+            </option>
+          </select>
+          <select v-model="newOverrideMode">
+            <option v-for="option in MODE_OPTIONS" :key="option.value" :value="option.value">{{ option.label }}</option>
+          </select>
+          <button class="btn btn-sm" type="button" @click="addOverride">添加/更新覆盖</button>
+        </div>
+        <table class="override-table">
+          <thead>
+            <tr>
+              <th style="width: 100px">来源</th>
+              <th>设备</th>
+              <th style="width: 80px">模式</th>
+              <th style="width: 160px">更新时间</th>
+              <th style="width: 90px">操作</th>
+            </tr>
+          </thead>
+          <tbody>
+            <tr v-for="row in overrides" :key="row.id">
+              <td>{{ sourceLabel(row.source) }}</td>
+              <td>
+                <span class="device-name">{{ row.deviceName || row.deviceId }}</span>
+                <span class="device-id">{{ row.deviceId }}</span>
+              </td>
+              <td>{{ modeLabel(row.mode) }}</td>
+              <td>{{ row.updatedAt ? formatDateTime(row.updatedAt) : '—' }}</td>
+              <td class="ops-cell">
+                <button class="btn btn-sm btn-danger" type="button" @click="removeOverride(row)">删除</button>
+              </td>
+            </tr>
+            <tr v-if="!overrides.length">
+              <td colspan="5" class="empty-tip">{{ configLoading ? '加载中...' : '暂无终端覆盖（全部跟随全局默认）' }}</td>
+            </tr>
+          </tbody>
+        </table>
+      </div>
+      <template #footer>
+        <el-button @click="configVisible = false">关闭</el-button>
+      </template>
+    </el-dialog>
+  </div>
+</template>
+
+<style scoped>
+/* 沿用「记录与版本」系列的旧后台面板视觉语言。 */
+.device-logs-view {
+  font-family: inherit;
+  color: #24384d;
+}
+.panel-box {
+  width: 100%;
+  min-width: 0;
+  padding: 20px 22px 24px;
+  border: 1px solid #d8e3ee;
+  border-radius: 14px;
+  background: linear-gradient(145deg, #ffffff, #f9fbfd);
+  box-shadow: 0 1px 2px rgba(39, 67, 94, 0.04), 0 14px 30px -24px rgba(39, 67, 94, 0.28);
+}
+h3 {
+  margin: 0;
+  font-size: 15px;
+  font-weight: 650;
+  color: #24384d;
+  letter-spacing: 0.2px;
+}
+.logs-head {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  gap: 12px;
+  margin-bottom: 16px;
+}
+.logs-head-tools {
+  display: flex;
+  align-items: center;
+  gap: 10px;
+}
+.retention-tip {
+  color: #8598ab;
+  font-size: 12.5px;
+}
+.logs-filter-row {
+  margin: 0 0 16px;
+}
+.form-row {
+  display: flex;
+  flex-wrap: wrap;
+  align-items: flex-end;
+  gap: 14px 18px;
+}
+.form-group {
+  display: flex;
+  flex-direction: column;
+  gap: 7px;
+}
+.form-group label {
+  color: #5b6f83;
+  font-size: 12.5px;
+  font-weight: 600;
+}
+.form-group input,
+.form-group select {
+  min-width: 0;
+  min-height: 42px;
+  padding: 8px 12px;
+  border: 1px solid #cbd9e6;
+  border-radius: 9px;
+  background: #f8fbfd;
+  color: #24384d;
+  font-size: 13.5px;
+  font-family: inherit;
+  color-scheme: light;
+  outline: none;
+}
+.form-group input:focus,
+.form-group select:focus {
+  background: #ffffff;
+  border-color: #5f85ad;
+  box-shadow: 0 0 0 3px rgba(95, 133, 173, 0.16);
+}
+.filter-actions {
+  display: flex;
+  gap: 10px;
+}
+.btn {
+  display: inline-flex;
+  align-items: center;
+  justify-content: center;
+  gap: 8px;
+  min-height: 42px;
+  padding: 9px 18px;
+  border: 1px solid #4f78a5;
+  border-radius: 9px;
+  background: linear-gradient(135deg, #5f85ad, #4f78a5);
+  color: #ffffff;
+  font-family: inherit;
+  font-size: 13.5px;
+  cursor: pointer;
+  box-shadow: 0 8px 18px -14px rgba(79, 120, 165, 0.72);
+  text-decoration: none;
+}
+.btn:hover:not(:disabled) {
+  background: linear-gradient(135deg, #7094ba, #5d83ac);
+}
+.btn:disabled {
+  cursor: not-allowed;
+  opacity: 0.55;
+}
+.btn-ghost {
+  background: #ffffff;
+  border-color: #c7d7e5;
+  color: #4f78a5;
+  box-shadow: none;
+}
+.btn-ghost:hover:not(:disabled) {
+  background: #edf5fb;
+  border-color: #95b1cb;
+  color: #2f5d8b;
+}
+.btn-danger {
+  background: linear-gradient(135deg, #c06d77, #b35f6a);
+  border-color: #b35f6a;
+}
+.btn-danger:hover:not(:disabled) {
+  background: linear-gradient(135deg, #cb7c84, #b96570);
+}
+.btn-sm {
+  min-height: 32px;
+  padding: 4px 12px;
+  font-size: 12.5px;
+  border-radius: 7px;
+}
+.logs-table-scroll {
+  width: 100%;
+  overflow-x: auto;
+}
+table {
+  width: 100%;
+  border-collapse: collapse;
+  font-size: 13px;
+}
+th {
+  padding: 10px 12px;
+  text-align: left;
+  color: #5b6f83;
+  font-weight: 650;
+  font-size: 12.5px;
+  border-bottom: 1px solid #dbe6f0;
+  white-space: nowrap;
+}
+td {
+  padding: 10px 12px;
+  border-bottom: 1px solid #eaf1f7;
+  vertical-align: top;
+}
+.source-pill {
+  display: inline-block;
+  padding: 2px 10px;
+  border-radius: 999px;
+  font-size: 12px;
+  font-weight: 600;
+}
+.source-pill.is-client {
+  background: #e8f1fb;
+  color: #37618f;
+}
+.source-pill.is-maixiang {
+  background: #eef7ec;
+  color: #3f7a42;
+}
+.device-name {
+  display: block;
+  font-weight: 600;
+  word-break: break-all;
+}
+.device-id {
+  display: block;
+  color: #8598ab;
+  font-size: 11.5px;
+  word-break: break-all;
+}
+.user-name {
+  font-weight: 600;
+}
+.file-name {
+  display: block;
+  word-break: break-all;
+}
+.dim {
+  color: #9db0c2;
+}
+.ops-cell {
+  display: flex;
+  gap: 8px;
+  flex-wrap: wrap;
+  border-bottom: none;
+}
+.empty-tip {
+  padding: 26px 0;
+  text-align: center;
+  color: #8598ab;
+}
+.dl-link {
+  text-decoration: none;
+}
+.viewer-toolbar {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  gap: 12px;
+  margin-bottom: 10px;
+}
+.viewer-meta {
+  color: #5b6f83;
+  font-size: 12.5px;
+}
+.viewer-actions {
+  display: flex;
+  gap: 8px;
+}
+.log-pre {
+  max-height: 62vh;
+  margin: 0;
+  padding: 12px 14px;
+  overflow: auto;
+  border: 1px solid #d8e3ee;
+  border-radius: 10px;
+  background: #0f1c29;
+  color: #d7e4f1;
+  font-family: Consolas, 'Courier New', monospace;
+  font-size: 12.5px;
+  line-height: 1.55;
+  white-space: pre;
+}
+.config-block {
+  margin-bottom: 18px;
+}
+.config-block h4 {
+  margin: 0 0 6px;
+  font-size: 13.5px;
+  color: #24384d;
+}
+.config-desc {
+  margin: 0 0 10px;
+  color: #66798d;
+  font-size: 12.5px;
+  line-height: 1.6;
+}
+.mode-switch {
+  display: flex;
+  gap: 10px;
+}
+.mode-btn {
+  min-height: 38px;
+  padding: 8px 20px;
+  border: 1px solid #c7d7e5;
+  border-radius: 9px;
+  background: #ffffff;
+  color: #4f78a5;
+  font-size: 13.5px;
+  font-family: inherit;
+  cursor: pointer;
+}
+.mode-btn.is-active {
+  border-color: #4f78a5;
+  background: linear-gradient(135deg, #5f85ad, #4f78a5);
+  color: #ffffff;
+  font-weight: 600;
+}
+.override-add {
+  display: flex;
+  gap: 10px;
+  margin-bottom: 12px;
+}
+.override-add select {
+  min-height: 38px;
+  padding: 6px 10px;
+  border: 1px solid #cbd9e6;
+  border-radius: 9px;
+  background: #f8fbfd;
+  color: #24384d;
+  font-size: 13px;
+  font-family: inherit;
+  color-scheme: light;
+}
+.override-select {
+  flex: 1;
+  min-width: 0;
+}
+.override-table td {
+  vertical-align: middle;
+}
+</style>

admin-frontend-vue/src/pages/records/RecordsTutorialPage.vue

@@ -2,7 +2,8 @@
 import { formatDateTime } from '@/utils/datetime'
 /** 教程管理页：工具台首页「立即下载教程」的包体管理（上传新包 / 列表 / 下载 / 删除）。
  *  上传走浏览器直传 MinIO（presign → PUT 带进度 → confirm 落库），与软件版本管理页同一套链路；
- *  工具台始终下载"最新上传"的包（列表第一条即当前生效）。 */
+ *  上传时可填版本号（仅展示与排序用）；工具台始终下载"最新上传"的包（不随列表排序变化）。
+ *  列表默认按上传时间降序，「版本号」「上传时间」表头可点击切换升降序。 */
 import { computed, onMounted, ref, watch } from 'vue'
 import OldPagination from '@/components/OldPagination.vue'
 import { ElMessage } from 'element-plus'
@@ -22,12 +23,64 @@ const filteredItems = computed(() => {
 /** 全站分页统一：客户端分页（10/20/50/100）。 */
 const page = ref(1)
 const pageSize = ref(10)
-const pagedItems = computed(() => filteredItems.value.slice((page.value - 1) * pageSize.value, page.value * pageSize.value))
+
+/** 排序：默认按上传时间降序（最新上传在最前，与后端返回顺序一致）；点表头切换升降序。 */
+type SortKey = 'version' | 'createdAt'
+const sortKey = ref<SortKey>('createdAt')
+const sortAsc = ref(false)
+
+function toggleSort(key: SortKey) {
+  if (sortKey.value === key) {
+    sortAsc.value = !sortAsc.value
+  } else {
+    sortKey.value = key
+    sortAsc.value = false
+  }
+  page.value = 1
+}
+
+/** 排序标记：未激活 ▲▼，激活时只留方向箭头（比 ⇅/↕ 字形支持好，避免 Windows 缺字形显示方框）。 */
+function sortMark(key: SortKey): string {
+  if (sortKey.value !== key) return '▲▼'
+  return sortAsc.value ? '▲' : '▼'
+}
+
+function compareText(a: string, b: string): number {
+  return a.localeCompare(b, 'zh-Hans-CN', { numeric: true, sensitivity: 'base' })
+}
+
+const sortedItems = computed(() => {
+  const rows = [...filteredItems.value]
+  rows.sort((a, b) => {
+    // 版本号为空的历史行始终排在末尾，避免切换排序时"无版本"占满首页。
+    if (sortKey.value === 'version') {
+      const av = a.version
+      const bv = b.version
+      if (!av || !bv) {
+        if (!av && !bv) return 0
+        return av ? -1 : 1
+      }
+      const diff = compareText(av, bv)
+      return sortAsc.value ? diff : -diff
+    }
+    const at = a.createdAt
+    const bt = b.createdAt
+    if (!at || !bt) {
+      if (!at && !bt) return 0
+      return at ? -1 : 1
+    }
+    const diff = compareText(at, bt)
+    return sortAsc.value ? diff : -diff
+  })
+  return rows
+})
+
+const pagedItems = computed(() => sortedItems.value.slice((page.value - 1) * pageSize.value, page.value * pageSize.value))
 function changePage(p: number) { page.value = p }
 function changeSize(size: number) { pageSize.value = size; page.value = 1 }
-// 搜索导致数据收缩时回钳页码，避免停在空页。
+// 搜索/排序导致数据收缩时回钳页码，避免停在空页。
-watch(filteredItems, () => {
+watch(sortedItems, () => {
-  page.value = Math.min(page.value, Math.max(1, Math.ceil(filteredItems.value.length / pageSize.value)))
+  page.value = Math.min(page.value, Math.max(1, Math.ceil(sortedItems.value.length / pageSize.value)))
 })
 
 /** 当前生效包 = 列表第一条（工具台下发的就是它）。 */
@@ -80,6 +133,7 @@ async function removeOne(row: TutorialPackageItem) {
 const uploadVisible = ref(false)
 const uploading = ref(false)
 const uploadPercent = ref(0)
+const newVersion = ref('')
 const pickedFile = ref<File | null>(null)
 /** 弹窗内成功/失败文案留驻（对齐软件版本页交互）。 */
 const uploadMsg = ref('')
@@ -103,6 +157,7 @@ function onFileChange(file: File) {
 function openUpload() {
   uploadMsg.value = ''
   uploadMsgOk.value = false
+  newVersion.value = ''
   pickedFile.value = null
   uploadVisible.value = true
 }
@@ -126,12 +181,13 @@ async function submitUpload() {
   uploading.value = true
   try {
     // 浏览器直传 MinIO：presign → PUT（进度条）→ confirm 落库。
-    await uploadTutorialPackage(pickedFile.value, pickedFile.value.name, (p) => {
+    await uploadTutorialPackage(pickedFile.value, pickedFile.value.name, newVersion.value.trim(), (p) => {
       uploadPercent.value = p
       uploadMsg.value = p >= 100 ? '上传完成，正在登记教程包...' : `正在上传：${p}%`
     })
     uploadMsg.value = '上传成功，工具台「立即下载教程」已切换为该包'
     uploadMsgOk.value = true
+    newVersion.value = ''
     pickedFile.value = null
     uploadPercent.value = 0
     load()
@@ -165,8 +221,13 @@ onMounted(load)
                 <input type="checkbox" :checked="pagedItems.length > 0 && pagedItems.every((r) => selectedIds.has(r.id))" @change="toggleSelectAllPage" :disabled="!pagedItems.length" />
               </th>
               <th style="width: 260px">文件名</th>
+              <th style="width: 110px">
+                <button class="sort-th sort-version" type="button" @click="toggleSort('version')">版本号<span class="sort-mark">{{ sortMark('version') }}</span></button>
+              </th>
               <th style="width: 110px">大小</th>
-              <th style="width: 150px">上传时间</th>
+              <th style="width: 150px">
+                <button class="sort-th sort-time" type="button" @click="toggleSort('createdAt')">上传时间<span class="sort-mark">{{ sortMark('createdAt') }}</span></button>
+              </th>
               <th>下载链接</th>
               <th style="width: 170px">操作</th>
             </tr>
@@ -181,6 +242,10 @@ onMounted(load)
                   <span class="file-name" :title="row.fileName">{{ row.fileName }}</span>
                   <span v-if="row.id === activeId" class="tag-active">当前生效</span>
                 </td>
+                <td>
+                  <span v-if="row.version" class="version-cell" :title="row.version">{{ row.version }}</span>
+                  <span v-else class="dim">—</span>
+                </td>
                 <td>{{ formatFileSize(row.fileSize) }}</td>
                 <td>{{ formatDateTime(row.createdAt) }}</td>
                 <td>
@@ -195,10 +260,10 @@ onMounted(load)
               </tr>
             </template>
             <tr v-else-if="loading">
-              <td colspan="6" class="empty-tip">加载中...</td>
+              <td colspan="7" class="empty-tip">加载中...</td>
             </tr>
             <tr v-else>
-              <td colspan="6" class="empty-tip">{{ keyword ? '暂无匹配教程包' : '暂无教程包记录' }}</td>
+              <td colspan="7" class="empty-tip">{{ keyword ? '暂无匹配教程包' : '暂无教程包记录' }}</td>
             </tr>
           </tbody>
         </table>
@@ -209,6 +274,10 @@ onMounted(load)
     <el-dialog v-model="uploadVisible" title="上传教程包" width="520px">
       <p class="upload-desc">上传教程 ZIP 包后，工具台首页「立即下载教程」将以下载该包为准（以最新上传的为主）。</p>
       <el-form label-width="110px">
+        <el-form-item label="版本号">
+          <el-input v-model="newVersion" placeholder="例如：v2026.09 或留空" maxlength="64" />
+          <p class="zip-hint">仅作展示与排序用，可留空；不影响工具台按最新上传下载</p>
+        </el-form-item>
         <el-form-item label="ZIP 压缩包" required>
           <el-upload :auto-upload="false" :limit="1" accept=".zip" :on-change="(file: any) => onFileChange(file.raw)" :on-remove="() => (pickedFile = null)">
             <el-button>选择文件</el-button>
@@ -368,6 +437,46 @@ h3 {
   text-overflow: ellipsis;
   vertical-align: middle;
 }
+/* 表头排序：按钮铺满单元格，仅 hover 时加深字色，保持旧版表头观感。 */
+.sort-th {
+  display: inline-flex;
+  align-items: center;
+  gap: 4px;
+  width: 100%;
+  padding: 0;
+  border: 0;
+  background: none;
+  color: inherit;
+  font-family: inherit;
+  font-size: inherit;
+  font-weight: inherit;
+  letter-spacing: inherit;
+  text-align: left;
+  cursor: pointer;
+}
+.sort-th:hover {
+  color: #2f5d8b;
+}
+.sort-mark {
+  color: #8293a5;
+  font-size: 12px;
+}
+.sort-th:hover .sort-mark {
+  color: #5f85ad;
+}
+.sort-version {
+  width: 110px;
+}
+.sort-time {
+  width: 150px;
+}
+.version-cell {
+  display: inline-block;
+  max-width: 100%;
+  overflow: hidden;
+  text-overflow: ellipsis;
+  vertical-align: middle;
+}
 .tag-active {
   display: inline-block;
   margin-left: 6px;

admin-frontend-vue/src/pages/records/tutorial-api.ts

@@ -22,11 +22,12 @@ export interface TutorialUploadTarget {
   objectKey: string
   uploadUrl: string
   fileUrl: string
+  version: string
 }
 
 /** 第一步：向后端申请直传 PUT 预签名地址（登录会话签发，3 分钟有效）。 */
-export async function requestTutorialPresign(fileName: string): Promise<TutorialUploadTarget> {
+export async function requestTutorialPresign(fileName: string, version = ''): Promise<TutorialUploadTarget> {
-  const { data } = await http.post<unknown>(TUTORIAL_PRESIGN_ENDPOINT, null, { params: { file_name: fileName } })
+  const { data } = await http.post<unknown>(TUTORIAL_PRESIGN_ENDPOINT, null, { params: { file_name: fileName, version } })
   const core = (unwrap(data) ?? {}) as Record<string, unknown>
   const uploadUrl = typeof core.upload_url === 'string' ? core.upload_url : ''
   const objectKey = typeof core.object_key === 'string' ? core.object_key : ''
@@ -37,13 +38,14 @@ export async function requestTutorialPresign(fileName: string): Promise<Tutorial
     objectKey,
     uploadUrl,
     fileUrl: typeof core.file_url === 'string' ? core.file_url : '',
+    version: typeof core.version === 'string' ? core.version : version.trim(),
   }
 }
 
 /** 第三步：直传完成后通知后端校验对象并写入记录（data.item 为新记录行）。 */
-export async function confirmTutorialPackage(objectKey: string, fileName: string): Promise<TutorialPackageItem | null> {
+export async function confirmTutorialPackage(objectKey: string, fileName: string, version = ''): Promise<TutorialPackageItem | null> {
   const { data } = await http.post<unknown>(TUTORIAL_CONFIRM_ENDPOINT, null, {
-    params: { object_key: objectKey, file_name: fileName },
+    params: { object_key: objectKey, file_name: fileName, version },
   })
   return parseTutorialPackageUpload(data)
 }
@@ -63,9 +65,10 @@ export async function deleteTutorialPackages(ids: number[]): Promise<number> {
 export async function uploadTutorialPackage(
   file: Blob,
   fileName: string,
+  version = '',
   onProgress?: (percent: number) => void,
 ): Promise<TutorialPackageItem | null> {
-  const target = await requestTutorialPresign(fileName)
+  const target = await requestTutorialPresign(fileName, version)
   await directPut.put(target.uploadUrl, file, {
     headers: { 'Content-Type': 'application/octet-stream' },
     onUploadProgress: (event) => {
@@ -74,5 +77,5 @@ export async function uploadTutorialPackage(
       onProgress(total > 0 ? Math.min(Math.round((event.loaded / total) * 100), 100) : 0)
     },
   })
-  return confirmTutorialPackage(target.objectKey, fileName)
+  return confirmTutorialPackage(target.objectKey, fileName, target.version)
 }

admin-frontend-vue/src/pages/records/tutorial-dto.ts

@@ -4,6 +4,8 @@
 export interface TutorialPackageItem {
   id: number
   fileName: string
+  /** 版本号（上传时填写，V126 之前的历史行为空串） */
+  version: string
   objectKey: string
   fileSize: number
   fileUrl: string

admin-frontend-vue/src/pages/records/tutorial-model.ts

@@ -19,6 +19,7 @@ export function toTutorialPackageItem(raw: unknown): TutorialPackageItem | null
   return {
     id,
     fileName: text(r.file_name ?? r.fileName),
+    version: text(r.version),
     objectKey: text(r.object_key ?? r.objectKey),
     fileSize: numberOrNull(r.file_size ?? r.fileSize) ?? 0,
     fileUrl: text(r.file_url ?? r.fileUrl),

admin-frontend-vue/src/router/routes.ts

@@ -25,6 +25,7 @@ export const adminPages: AdminPageDef[] = [
   { path: 'records/history', menuKey: 'admin_history', title: '生成记录', load: () => import('@/pages/records/RecordsHistoryPage.vue') },
   { path: 'records/software-version', menuKey: 'admin_version', title: '软件版本管理', load: () => import('@/pages/records/RecordsSoftwareVersionPage.vue') },
   { path: 'records/tutorial', menuKey: 'admin_tutorial', title: '教程管理', load: () => import('@/pages/records/RecordsTutorialPage.vue') },
+  { path: 'records/device-logs', menuKey: 'admin_device_logs', title: '日志管理', load: () => import('@/pages/records/DeviceLogsPage.vue') },
   { path: 'records/digital-human-version', menuKey: 'digital_human_version', title: '数字人版本管理', load: () => import('@/pages/records/RecordsDigitalHumanVersionPage.vue') },
   { path: 'records/image-video-tasks', menuKey: 'admin_image_video_tasks', title: '视频任务记录', load: () => import('@/pages/tasks/ImageVideoTasksPage.vue') },
   { path: 'shop-center/data-tasks', menuKey: 'admin_shop_data_crawl_tasks', title: '店铺数据记录', load: () => import('@/pages/tasks/ShopDataTasksPage.vue') },

admin-frontend-vue/tests/align-tutorial.test.ts

@@ -14,16 +14,23 @@ test('align_tutorial_page_registered', () => {
 
 test('align_tutorial_page_wiring', () => {
   const page = readSource('src/pages/records/RecordsTutorialPage.vue')
-  // 上传入口：选择 zip → 直传（进度条）→ 成功提示留驻弹窗。
+  // 上传入口：版本号（可空）→ 选择 zip → 直传（进度条）→ 成功提示留驻弹窗。
   assert.match(page, /上传教程包/, '存在上传入口按钮')
   assert.match(page, /uploadTutorialPackage/, '上传走教程包直传链路')
+  assert.match(page, /label="版本号"/, '上传弹窗提供版本号输入')
+  assert.match(page, /newVersion\.value\.trim\(\)/, '版本号去空白后随上传提交')
   assert.match(page, /请选择 zip 压缩包/, '未选文件时提示')
   assert.match(page, /仅支持 \.zip 格式/, '格式校验提示')
   assert.match(page, /上传成功，工具台「立即下载教程」已切换为该包/, '成功提示说明生效位置')
   assert.match(page, /上传完成，正在登记教程包/, '进度文案')
-  // 列表：当前生效标记 + 下载 + 删除 + 空态。
+  // 列表：版本号列 + 上传时间列 + 当前生效标记 + 下载 + 删除 + 空态。
   assert.match(page, /当前生效/, '最新上传的包标记当前生效')
   assert.match(page, /formatFileSize/, '展示包体大小')
+  assert.match(page, /row\.version/, '展示版本号列')
+  assert.match(page, /toggleSort\('version'\)/, '版本号表头可点击排序')
+  assert.match(page, /toggleSort\('createdAt'\)/, '上传时间表头可点击排序')
+  assert.match(page, /const sortKey = ref<SortKey>\('createdAt'\)/, '默认排序字段为上传时间')
+  assert.match(page, /const sortAsc = ref\(false\)/, '默认降序')
   assert.match(page, /下载/, '操作列提供下载')
   assert.match(page, /<a v-if="row\.fileUrl" class="btn btn-sm dl-btn"[^>]*download/, '下载链接带 download 强制下载')
   assert.match(page, /确认删除选中的/, '批量删除二次确认')
@@ -37,7 +44,8 @@ test('align_tutorial_api_contract', () => {
   assert.match(api, /\$\{TUTORIAL_UPLOAD_ENDPOINT\}\/presign/, '直传预签名端点')
   assert.match(api, /\$\{TUTORIAL_UPLOAD_ENDPOINT\}\/confirm/, '确认端点')
   assert.match(api, /\$\{TUTORIAL_UPLOAD_ENDPOINT\}\/delete/, '删除端点')
-  assert.match(api, /params: \{ object_key: objectKey, file_name: fileName \}/, '确认回传对象 key 与文件名')
+  assert.match(api, /params: \{ file_name: fileName, version \}/, '预签名回传文件名与版本号')
+  assert.match(api, /params: \{ object_key: objectKey, file_name: fileName, version \}/, '确认回传对象 key、文件名与版本号')
   assert.match(api, /withCredentials: false/, '直传实例不挂会话拦截器')
 })
 
@@ -46,4 +54,5 @@ test('align_tutorial_model_parsers', () => {
   assert.match(model, /r\.file_name \?\? r\.fileName/, '兼容 snake/camel 文件名')
   assert.match(model, /r\.file_url \?\? r\.fileUrl/, '兼容 snake/camel 下载链接')
   assert.match(model, /r\.file_size \?\? r\.fileSize/, '兼容 snake/camel 文件大小')
+  assert.match(model, /version: text\(r\.version\)/, '解析版本号（缺失为空串，兼容历史行）')
 })

admin-frontend-vue/tests/align-user-menu-grantable.test.ts

@@ -0,0 +1,86 @@
+import test from 'node:test'
+import assert from 'node:assert/strict'
+import { readSource } from './helpers.ts'
+import {
+  buildMenuOptionTree,
+  compactDirectGrantIds,
+  parseMenuOptionList,
+  parsePermissionMenuItem,
+} from '../src/pages/account/user-menu-auth.ts'
+
+// 2026-09-16 线上事故：非超管（普通管理员）在授权树里勾到自己无权授予的菜单后，
+// 后端 ensureGrantable 抛 403 并回滚整笔事务——创建用户与保存权限双双失败，
+// 前端只显示「普通管理员只能分配自己已有的菜单权限」。
+// 修复：菜单列表按操作者标记 grantable，前端把不可授予的节点置灰不可勾。
+//
+// 注意：这里是「置灰」而非「隐藏」。授权保存是整树替换，隐藏会让超管早先授予、
+// 而操作者自己没有的菜单在提交时被当作取消勾选删掉（与 09-13「权限自己没掉」同类）。
+
+test('align_user_menu_grantable_false_maps_to_disabled_node', () => {
+  const locked = parsePermissionMenuItem(
+    { id: 5, name: '查询ASIN', parent_id: null, sort_order: 1, grantable: false },
+    'admin',
+  )
+  assert.equal(locked?.disabled, true, 'grantable=false → 节点置灰')
+
+  const allowed = parsePermissionMenuItem(
+    { id: 6, name: '店铺管理', parent_id: null, sort_order: 2, grantable: true },
+    'admin',
+  )
+  assert.equal(allowed?.disabled, false, 'grantable=true → 可勾选')
+
+  // 菜单管理页等未标记 grantable 的接口必须保持旧行为（全部可勾选）
+  const unmarked = parsePermissionMenuItem({ id: 7, name: '菜单权限配置', parent_id: null, sort_order: 3 }, 'admin')
+  assert.equal(unmarked?.disabled, false, '缺省 grantable 视为可授予')
+})
+
+test('align_user_menu_grantable_survives_tree_build', () => {
+  const nodes = parseMenuOptionList(
+    [
+      { id: 98, name: '账号与权限', parent_id: null, sort_order: 1, grantable: false },
+      { id: 7, name: '用户管理', parent_id: 98, sort_order: 1, grantable: false },
+      { id: 100, name: '店铺管理', parent_id: null, sort_order: 2, grantable: true },
+    ],
+    'admin',
+  )
+  const tree = buildMenuOptionTree(nodes)
+  const account = tree.find((node) => node.id === 98)
+  assert.equal(account?.disabled, true, '分组节点置灰')
+  assert.equal(account?.children?.[0]?.disabled, true, '子节点置灰随树保留')
+  assert.equal(tree.find((node) => node.id === 100)?.disabled, false, '可授予节点不受影响')
+})
+
+test('align_user_menu_grantable_disabled_node_still_compactable', () => {
+  // 已持有但无权授予的节点会保持勾选并原样提交，压缩逻辑不能因 disabled 漏掉它
+  const tree = buildMenuOptionTree(
+    parseMenuOptionList(
+      [
+        { id: 98, name: '账号与权限', parent_id: null, sort_order: 1, grantable: false },
+        { id: 7, name: '用户管理', parent_id: 98, sort_order: 1, grantable: false },
+      ],
+      'admin',
+    ),
+  )
+  assert.deepEqual(compactDirectGrantIds([98, 7], tree), [98], '父级已勾选时仍压缩掉后代')
+})
+
+test('align_user_menu_grantable_wired_end_to_end', () => {
+  const tree = readSource('src/pages/account/UserMenuAuthTree.vue')
+  assert.match(tree, /disabled: 'disabled'/, 'el-tree 按 disabled 键置灰节点')
+
+  const vo = readSource(
+    '../backend-java/src/main/java/com/nanri/aiimage/modules/permission/model/vo/PermissionMenuItemVo.java',
+  )
+  assert.match(vo, /private Boolean grantable;/, 'VO 暴露 grantable')
+
+  const controller = readSource(
+    '../backend-java/src/main/java/com/nanri/aiimage/modules/permission/controller/PermissionMenuController.java',
+  )
+  assert.match(controller, /permissionMenuService\.list\(requireAdmin\(request\), menuType\)/, '列表接口传入操作者')
+
+  const service = readSource(
+    '../backend-java/src/main/java/com/nanri/aiimage/modules/permission/service/PermissionMenuService.java',
+  )
+  assert.match(service, /resolveGrantableMenuIds/, '按操作者计算可授予集')
+  assert.match(service, /ensureGrantable\(operator, grantIds, userId\)/, '保存校验传入目标用户以放行既有授权')
+})

backend-java/docs/tx-duration-benchmark.md

@@ -0,0 +1,28 @@
+# 单次提交事务段耗时基线（task-138）
+
+> 本文档由 `TxDurationBenchmarkTest` 断言存在与内容，用于冻结「Python 结果提交」单次调用的
+> 耗时上界，防止提交路径劣化。**不要删除**，劣化时更新数值并说明原因。
+
+## 测量对象
+
+`SimilarAsinTaskService.submitResult(taskId, request)` 的 mock 环境单次调用，拆成两段：
+
+| 段 | 含义 |
+|---|---|
+| 计算段（prepare） | 事务外的纯计算：行裁剪、校验、payload 序列化与哈希 |
+| 事务段（persist） | `inNewTransaction` 内的落库：分片 upsert、scope state 更新 |
+
+## 基线数值
+
+- **单次事务段基线: 5**（毫秒，mock 环境，本地实测）
+- **200 分片**负载基线: 10 秒（`twoHundredChunkLoadBounded` 的上界）
+- 单次提交总耗时上界: 500 毫秒（`totalDurationReasonablePerSubmission`）
+
+`noRegressionAgainstDocumentedBaseline` 按「当前 ≤ 基线 × 3」判定劣化，因此事务段超过
+15ms 即视为回归。
+
+## 说明
+
+- 该基线取的是 mock 环境（Mapper 全部打桩）而非真实 MySQL 的耗时，用于**相对劣化**而非绝对性能。
+- 真实环境的事务段耗时会显著高于此值，本基线不适用于容量规划。
+- 测量机器与 JDK 变更后如需调整，请同步更新本文件数值。

backend-java/src/main/java/com/nanri/aiimage/common/exception/BusinessCodes.java

@@ -0,0 +1,36 @@
+package com.nanri.aiimage.common.exception;
+
+/**
+ * 业务错误码常量。
+ *
+ * <p>历史上 40901 被两种语义复用：「任务已结束」（幂等忽略）与「任务正在处理中」（分布式锁竞争，
+ * 应重试）；而 GlobalExceptionHandler 把 40901 一律转成 HTTP 200 + success=true，
+ * 导致锁竞争时 Python worker 误判回传成功并停止重试，分片数据静默丢失。
+ * 现拆分为两个码：
+ * <ul>
+ *   <li>{@link #TASK_ALREADY_FINISHED}：任务已结束，重复提交无意义 → 幂等忽略，响应 success=true；</li>
+ *   <li>{@link #TASK_BUSY}：任务正被其它请求持锁推进 → 响应 success=false，调用方应稍后重试。</li>
+ * </ul>
+ */
+public final class BusinessCodes {
+
+    private BusinessCodes() {
+    }
+
+    /** 任务已结束，拒绝重复提交。语义：幂等忽略，响应 success=true，调用方不应重试。 */
+    public static final int TASK_ALREADY_FINISHED = 40901;
+
+    /** 任务正在处理中（分布式锁竞争）。语义：资源忙，响应 success=false，调用方应稍后重试。 */
+    public static final int TASK_BUSY = 40902;
+
+    /** 任务归属其它实例，需转发。 */
+    public static final int TASK_OWNER_FORWARD = 40903;
+
+    /**
+     * 提交结果的目标任务已不存在（通常是被删除）。
+     * 语义：本次提交无意义，响应 success=false 且带该码，调用方应放弃而不是反复重试。
+     * 与 {@link #TASK_ALREADY_FINISHED} 的区别：那个还留了任务记录（可幂等忽略），
+     * 这个任务已经没了——如实报错，否则任务被误删时结果会被静默吞掉。
+     */
+    public static final int TASK_NOT_FOUND = 40401;
+}

backend-java/src/main/java/com/nanri/aiimage/common/exception/GlobalExceptionHandler.java

@@ -1,18 +1,21 @@
 package com.nanri.aiimage.common.exception;
 
 import com.nanri.aiimage.common.api.ApiResponse;
+import com.nanri.aiimage.common.security.DeviceSessionPolicy;
 import com.nanri.aiimage.common.service.TaskOwnerForwardService;
 import com.nanri.aiimage.config.TaskOperationLockConfig;
 import jakarta.servlet.http.HttpServletRequest;
 import jakarta.validation.ConstraintViolationException;
 import lombok.RequiredArgsConstructor;
 import lombok.extern.slf4j.Slf4j;
+import org.springframework.http.HttpStatus;
 import org.springframework.http.ResponseEntity;
 import org.springframework.web.bind.MethodArgumentNotValidException;
 import org.springframework.web.bind.annotation.ExceptionHandler;
 import org.springframework.web.bind.annotation.RestControllerAdvice;
 import org.springframework.web.context.request.async.AsyncRequestNotUsableException;
 import org.springframework.web.server.ResponseStatusException;
+import org.springframework.web.servlet.resource.NoResourceFoundException;
 
 import java.io.IOException;
 
@@ -52,26 +55,64 @@ public class GlobalExceptionHandler {
                     ? ApiResponse.fail(forwardEx.getMessage())
                     : ApiResponse.fail(forwardEx.getCode(), forwardEx.getMessage());
         } catch (Exception forwardEx) {
+            // 转发失败是**瞬时基础设施故障**（归属实例正在滚动重启），不是业务结论，
+            // 更不能表达成「任务不存活」。原先返回 ApiResponse.fail(40903) —— HTTP 200
+            // 加 data:null，而客户端那句 bool((resp.json().get("data") or {}).get("alive"))
+            // 会把「拿不到数据」折叠成 alive=false，于是客户端把**健康的长任务主动停掉**：
+            // 2026-09-18 任务 28616 就是这么死的（归属节点 server-110 重启窗口内，心跳经
+            // nginx 落到 server-121，转发 3 次 Connection refused 后返回空 data）。
+            // 改为 503 + 空 body：新客户端按状态码判为「未知」继续跑；老客户端因 body 不是
+            // JSON、resp.json() 抛异常，同样落到「未知」。顺带让这类故障在 HTTP 指标里可见
+            // （原先记成 200，监控完全看不到滚动重启期间丢了多少心跳）。
             log.warn("[instance-routing] forward failed taskId={} operation={} owner={} current={} msg={}",
                     ex.getTaskId(), ex.getOperation(), ex.getOwnerInstanceId(), ex.getCurrentInstanceId(),
                     forwardEx.getMessage(), forwardEx);
-            return ApiResponse.fail(40903, "任务归属实例转发失败: " + forwardEx.getMessage());
+            return ResponseEntity.status(HttpStatus.SERVICE_UNAVAILABLE).build();
         }
     }
 
     @ExceptionHandler(BusinessException.class)
-    public ApiResponse<Void> handleBusinessException(BusinessException ex) {
+    public ApiResponse<Void> handleBusinessException(BusinessException ex, HttpServletRequest request) {
-        if (Integer.valueOf(40901).equals(ex.getCode())) {
+        // 业务异常此前完全不记日志：2026-09-16 线上「保存权限/创建用户」双双失败时，
+        // 服务端只留 RequestTraceFilter 的 200 一行，根因只能靠反推响应体字节数才定位到。
+        // 401/4011（未登录、被顶下线）属于轮询类接口的常态噪声，降为 debug 以免淹没真实业务错。
+        if (isRoutineAuthNoise(ex.getCode())) {
+            log.debug("[business] {} {} code={} message={}", request.getMethod(), request.getRequestURI(),
+                    ex.getCode(), ex.getMessage());
+        } else {
+            log.warn("[business] {} {} code={} message={}", request.getMethod(), request.getRequestURI(),
+                    ex.getCode(), ex.getMessage());
+        }
+        if (Integer.valueOf(BusinessCodes.TASK_ALREADY_FINISHED).equals(ex.getCode())) {
+            // 幂等忽略：任务已结束时的重复提交无副作用，按成功返回，避免客户端反复重试
             return ApiResponse.success("任务已结束，忽略重复提交", null);
         }
+        if (Integer.valueOf(BusinessCodes.TASK_BUSY).equals(ex.getCode())) {
+            // 锁竞争：必须如实返回失败 + 可重试码，否则 worker 会把「未落库」当成功而停止重试
+            return ApiResponse.fail(BusinessCodes.TASK_BUSY, ex.getMessage());
+        }
         return ex.getCode() == null
                 ? ApiResponse.fail(ex.getMessage())
                 : ApiResponse.fail(ex.getCode(), ex.getMessage());
     }
 
+    @ExceptionHandler(NoResourceFoundException.class)
+    public ResponseEntity<ApiResponse<Void>> handleNoResourceFoundException(NoResourceFoundException ex) {
+        // 静态资源 404。绝大部分是外部扫描器在探测 /.env、/credentials、aliyun.json、oss.json
+        // 这类云凭据文件（线上单节点一天 580 条）。此前落到 handleException 里，既刷 ERROR 堆栈，
+        // 又把探测响应伪装成 HTTP 200；这里降为 debug 并如实返回 404。
+        log.debug("static resource not found: {}", ex.getMessage());
+        return ResponseEntity.status(HttpStatus.NOT_FOUND).body(ApiResponse.fail("资源不存在"));
+    }
+
+    /** 未登录 / 登录态失效 / 被其他设备顶下线：按 401 语义的常态噪声，不占 WARN。 */
+    private boolean isRoutineAuthNoise(Integer code) {
+        return Integer.valueOf(401).equals(code)
+                || Integer.valueOf(DeviceSessionPolicy.CODE_KICKED).equals(code);
+    }
+
     @ExceptionHandler(MethodArgumentNotValidException.class)
-    public ApiResponse<Void> handleMethodArgumentNotValidException(MethodArgumentNotValidException ex) {
+    public ApiResponse<Void> handleMethodArgumentNotValidException(MethodArgumentNotValidException ex) {        String message = ex.getBindingResult().getFieldError() != null
-        String message = ex.getBindingResult().getFieldError() != null
                 ? ex.getBindingResult().getFieldError().getDefaultMessage()
                 : "参数校验失败";
         return ApiResponse.fail(message);
@@ -100,7 +141,9 @@ public class GlobalExceptionHandler {
             return ApiResponse.fail("客户端已断开连接");
         }
         log.error("Unhandled exception", ex);
-        return ApiResponse.fail("服务异常: " + ex.getMessage());
+        // 不再回传原始异常信息：SQL 报错、类名与内部路径会直接暴露给调用方，便于攻击者
+        // 摸清技术栈与表结构。详情只进日志（上方 log.error 已带完整堆栈），对外统一文案。
+        return ApiResponse.fail("服务器内部错误，请稍后重试");
     }
 
     private boolean isClientAbort(Throwable ex) {

backend-java/src/main/java/com/nanri/aiimage/common/mapper/AdminUserMapper.java

@@ -1,7 +1,7 @@
-package com.nanri.aiimage.modules.permission.mapper;
+package com.nanri.aiimage.common.mapper;
 
 import com.baomidou.mybatisplus.core.mapper.BaseMapper;
-import com.nanri.aiimage.modules.permission.model.entity.AdminUserEntity;
+import com.nanri.aiimage.common.model.entity.AdminUserEntity;
 import org.apache.ibatis.annotations.Mapper;
 
 @Mapper

backend-java/src/main/java/com/nanri/aiimage/common/mapper/ShopManageGroupMapper.java

@@ -1,7 +1,7 @@
-package com.nanri.aiimage.modules.shopkey.mapper;
+package com.nanri.aiimage.common.mapper;
 
 import com.baomidou.mybatisplus.core.mapper.BaseMapper;
-import com.nanri.aiimage.modules.shopkey.model.entity.ShopManageGroupEntity;
+import com.nanri.aiimage.common.model.entity.ShopManageGroupEntity;
 import org.apache.ibatis.annotations.Mapper;
 import org.apache.ibatis.annotations.Param;
 import org.apache.ibatis.annotations.Select;

backend-java/src/main/java/com/nanri/aiimage/common/model/entity/AdminUserEntity.java

@@ -1,4 +1,4 @@
-package com.nanri.aiimage.modules.permission.model.entity;
+package com.nanri.aiimage.common.model.entity;
 
 import com.baomidou.mybatisplus.annotation.IdType;
 import com.baomidou.mybatisplus.annotation.TableField;

backend-java/src/main/java/com/nanri/aiimage/common/model/entity/ShopManageGroupEntity.java

@@ -1,4 +1,4 @@
-package com.nanri.aiimage.modules.shopkey.model.entity;
+package com.nanri.aiimage.common.model.entity;
 
 import com.baomidou.mybatisplus.annotation.IdType;
 import com.baomidou.mybatisplus.annotation.TableField;

backend-java/src/main/java/com/nanri/aiimage/common/model/vo/ProductRiskCandidateVo.java

@@ -1,4 +1,4 @@
-package com.nanri.aiimage.modules.productrisk.model.vo;
+package com.nanri.aiimage.common.model.vo;
 
 import com.fasterxml.jackson.annotation.JsonProperty;
 import io.swagger.v3.oas.annotations.media.Schema;

backend-java/src/main/java/com/nanri/aiimage/common/model/vo/ProductRiskCountryPreferenceVo.java

@@ -1,4 +1,4 @@
-package com.nanri.aiimage.modules.productrisk.model.vo;
+package com.nanri.aiimage.common.model.vo;
 
 import com.fasterxml.jackson.annotation.JsonProperty;
 import io.swagger.v3.oas.annotations.media.Schema;

backend-java/src/main/java/com/nanri/aiimage/common/model/vo/ProductRiskDashboardVo.java

@@ -1,4 +1,4 @@
-package com.nanri.aiimage.modules.productrisk.model.vo;
+package com.nanri.aiimage.common.model.vo;
 
 import com.fasterxml.jackson.annotation.JsonProperty;
 import io.swagger.v3.oas.annotations.media.Schema;

backend-java/src/main/java/com/nanri/aiimage/common/model/vo/ProductRiskHistoryVo.java

@@ -1,4 +1,4 @@
-package com.nanri.aiimage.modules.productrisk.model.vo;
+package com.nanri.aiimage.common.model.vo;
 
 import io.swagger.v3.oas.annotations.media.Schema;
 import lombok.Data;

backend-java/src/main/java/com/nanri/aiimage/common/model/vo/ProductRiskMatchShopsVo.java

@@ -1,4 +1,4 @@
-package com.nanri.aiimage.modules.productrisk.model.vo;
+package com.nanri.aiimage.common.model.vo;
 
 import io.swagger.v3.oas.annotations.media.Schema;
 import lombok.Data;

backend-java/src/main/java/com/nanri/aiimage/common/model/vo/ProductRiskResultItemVo.java

@@ -1,4 +1,4 @@
-package com.nanri.aiimage.modules.productrisk.model.vo;
+package com.nanri.aiimage.common.model.vo;
 
 import com.fasterxml.jackson.annotation.JsonAlias;
 import com.fasterxml.jackson.annotation.JsonProperty;

backend-java/src/main/java/com/nanri/aiimage/common/model/vo/ProductRiskShopQueueItemVo.java

@@ -1,4 +1,4 @@
-package com.nanri.aiimage.modules.productrisk.model.vo;
+package com.nanri.aiimage.common.model.vo;
 
 import com.fasterxml.jackson.annotation.JsonAlias;
 import com.fasterxml.jackson.annotation.JsonProperty;

backend-java/src/main/java/com/nanri/aiimage/common/model/vo/ProductRiskTaskBatchVo.java

@@ -1,4 +1,4 @@
-package com.nanri.aiimage.modules.productrisk.model.vo;
+package com.nanri.aiimage.common.model.vo;
 
 import io.swagger.v3.oas.annotations.media.Schema;
 import lombok.Data;

backend-java/src/main/java/com/nanri/aiimage/common/model/vo/ProductRiskTaskDetailVo.java

@@ -1,4 +1,4 @@
-package com.nanri.aiimage.modules.productrisk.model.vo;
+package com.nanri.aiimage.common.model.vo;
 
 import io.swagger.v3.oas.annotations.media.Schema;
 import lombok.Data;

backend-java/src/main/java/com/nanri/aiimage/common/model/vo/ProductRiskTaskItemVo.java

@@ -1,4 +1,4 @@
-package com.nanri.aiimage.modules.productrisk.model.vo;
+package com.nanri.aiimage.common.model.vo;
 
 import com.fasterxml.jackson.annotation.JsonProperty;
 import com.nanri.aiimage.modules.shopmatch.model.vo.ShopMatchTaskStageVo;

backend-java/src/main/java/com/nanri/aiimage/common/module/TaskModuleRegistry.java

@@ -0,0 +1,111 @@
+package com.nanri.aiimage.common.module;
+
+import java.util.LinkedHashMap;
+import java.util.List;
+import java.util.Map;
+import java.util.Set;
+import java.util.stream.Collectors;
+
+/**
+ * 任务模块清单的**单一来源**（2026-09 全维度审查 G6）。
+ *
+ * 背景：新增一个工具模块此前要同时改 5 处枚举（结果文件 Job 白名单、按天清理清单、
+ * 站内通知模块名、任务心跳分支、陈旧判死巡检线），其中 4 处没有任何自检，
+ * 漏改表现为"功能静默不生效"。现在把**数据驱动的三处**（结果文件 Job、按天清理、通知中文名）
+ * 统一从这里派生，并由 {@code TaskModuleCoverageTest} 守住覆盖面。
+ *
+ * 心跳与陈旧判死两处是代码分支驱动（switch / 逐模块委派），无法纯数据派生，
+ * 新增模块时仍需实现对应 Handler，测试会提示缺失。
+ */
+public final class TaskModuleRegistry {
+
+    /**
+     * 单个任务模块的声明。
+     *
+     * @param type            moduleType（与 biz_file_task.module_type 一致）
+     * @param label           站内通知/界面用中文名
+     * @param resultFileJob   是否产出结果文件（需要注册 ResultFileJobHandler）
+     * @param ageCleanup      是否参与按天清理（ModuleCleanupProperties）
+     * @param delegatedStaleCheck 是否以"委派"方式并入 stale-check 巡检线
+     *                        （DeleteBrandStaleTaskService.delegatedStaleChecks，必须逐模块登记动作）
+     * @param selfScheduledStaleCheck 该模块的陈旧判死自带更快的调度（publish 60s / collect-data 30s），
+     *                        刻意不并入 2 分钟一轮的巡检线——并入会显著拉长判死时延。
+     *                        这类模块的"不在委托名单"是设计差异，不是漏接；由覆盖面测试固定。
+     */
+    public record Module(String type, String label, boolean resultFileJob, boolean ageCleanup,
+                         boolean delegatedStaleCheck, boolean selfScheduledStaleCheck) {
+    }
+
+    private static final List<Module> MODULES = List.of(
+            new Module("PUBLISH", "上架", true, false, false, true),
+            new Module("DEDUPE", "数据去重", false, true, false, false),
+            new Module("SPLIT", "数据拆分", false, true, false, false),
+            new Module("CONVERT", "格式转换", false, true, false, false),
+            new Module("DELETE_BRAND", "删除ASIN", true, true, false, false),
+            new Module("PRODUCT_RISK_RESOLVE", "商品风险解决", true, true, false, false),
+            new Module("PRICE_TRACK", "跟价", true, true, false, false),
+            new Module("SHOP_MATCH", "定时匹配", true, true, false, false),
+            new Module("PATROL_DELETE", "巡店删除", true, true, false, false),
+            new Module("QUERY_ASIN", "查询ASIN", true, true, false, false),
+            new Module("WITHDRAW", "取款", true, true, false, false),
+            new Module("APPEARANCE_PATENT", "外观专利检测", true, true, true, false),
+            new Module("SIMILAR_ASIN", "货源查询", true, true, true, false),
+            new Module("COLLECT_DATA", "采集数据", true, true, false, true),
+            new Module("SHOP_DATA_CRAWL", "店铺数据抓取", true, false, true, false),
+            new Module("BRAND", "品牌检测", true, false, true, false)
+    );
+
+    private TaskModuleRegistry() {
+    }
+
+    public static List<Module> modules() {
+        return MODULES;
+    }
+
+    /** 全部 moduleType。 */
+    public static Set<String> moduleTypes() {
+        return MODULES.stream().map(Module::type).collect(Collectors.toUnmodifiableSet());
+    }
+
+    /** moduleType → 中文名（站内通知等展示用）。 */
+    public static Map<String, String> labels() {
+        Map<String, String> labels = new LinkedHashMap<>();
+        for (Module module : MODULES) {
+            labels.put(module.type(), module.label());
+        }
+        return Map.copyOf(labels);
+    }
+
+    /** 产出结果文件的模块（需要注册 Handler 的模块）。 */
+    public static Set<String> resultFileJobModuleTypes() {
+        return MODULES.stream().filter(Module::resultFileJob).map(Module::type)
+                .collect(Collectors.toUnmodifiableSet());
+    }
+
+    /** 参与按天清理的模块。 */
+    public static Set<String> ageCleanupModuleTypes() {
+        return MODULES.stream().filter(Module::ageCleanup).map(Module::type)
+                .collect(Collectors.toUnmodifiableSet());
+    }
+
+    /** 自带更快巡检节奏、刻意不并入集中巡检线的模块。 */
+    public static Set<String> selfScheduledStaleCheckModuleTypes() {
+        return MODULES.stream().filter(Module::selfScheduledStaleCheck).map(Module::type)
+                .collect(Collectors.toUnmodifiableSet());
+    }
+
+    /** 以"委派"方式并入 stale-check 巡检线的模块（DeleteBrandStaleTaskService 必须逐模块登记动作）。 */
+    public static Set<String> delegatedStaleCheckModuleTypes() {
+        return MODULES.stream().filter(Module::delegatedStaleCheck).map(Module::type)
+                .collect(Collectors.toUnmodifiableSet());
+    }
+
+    public static String labelOf(String moduleType) {
+        for (Module module : MODULES) {
+            if (module.type().equals(moduleType)) {
+                return module.label();
+            }
+        }
+        return moduleType;
+    }
+}

backend-java/src/main/java/com/nanri/aiimage/common/security/AdminAuthSupport.java

@@ -1,24 +1,26 @@
-package com.nanri.aiimage.modules.admin.support;
+package com.nanri.aiimage.common.security;
 
 import com.nanri.aiimage.common.exception.BusinessException;
-import com.nanri.aiimage.modules.auth.service.JwtService;
+import com.nanri.aiimage.common.security.JwtService;
-import com.nanri.aiimage.modules.auth.support.DeviceSessionPolicy;
+import com.nanri.aiimage.common.security.DeviceSessionPolicy;
-import com.nanri.aiimage.modules.permission.mapper.AdminUserMapper;
+import com.nanri.aiimage.common.mapper.AdminUserMapper;
-import com.nanri.aiimage.modules.permission.model.entity.AdminUserEntity;
+import com.nanri.aiimage.common.model.entity.AdminUserEntity;
 import io.jsonwebtoken.Claims;
 import jakarta.servlet.http.Cookie;
 import jakarta.servlet.http.HttpServletRequest;
 import lombok.RequiredArgsConstructor;
+import lombok.extern.slf4j.Slf4j;
 import org.springframework.http.HttpHeaders;
 import org.springframework.stereotype.Component;
 
-import com.nanri.aiimage.modules.auth.config.AuthProperties;
+import com.nanri.aiimage.common.security.AuthProperties;
 import org.springframework.beans.factory.annotation.Value;
 
 import java.nio.charset.StandardCharsets;
 import java.nio.file.Files;
 import java.nio.file.Path;
 
+@Slf4j
 @Component
 @RequiredArgsConstructor
 public class AdminAuthSupport {
@@ -62,8 +64,30 @@ public class AdminAuthSupport {
         return user;
     }
 
-    /** 当前用户必须是管理员或超级管理员，否则抛 403。 */
+    /**
-    public AdminUserEntity requireAdmin(HttpServletRequest request) {
+     * 解析当前请求 JWT 中**签名的**设备标识（deviceId claim）；识别不出时返回空串。
+     *
+     * <p>无 token、token 过期/非法、内部令牌通道调用一律返回空串——调用方必须把空串
+     * 当作"来源不明"做保守判定，绝不据此放宽任何限制。只认签名 claim，不接受
+     * X-Device-Id 请求头（头由客户端可控，见 {@link DeviceSessionPolicy} 类注释）。</p>
+     *
+     * <p>本方法只做识别、不做鉴权，因此解析失败不抛异常，仅记日志后返回空串，
+     * 避免把匿名/内部调用直接升级成 401。</p>
+     */
+    public String currentDeviceId(HttpServletRequest request) {
+        String token = resolveToken(request);
+        if (token == null || token.isBlank()) {
+            return "";
+        }
+        try {
+            return DeviceSessionPolicy.claimDeviceId(jwtService.parse(token));
+        } catch (Exception ex) {
+            log.warn("[auth] 解析 token 取设备标识失败，按来源不明处理: {}", ex.getMessage());
+            return "";
+        }
+    }
+
+    /** 当前用户必须是管理员或超级管理员，否则抛 403。 */    public AdminUserEntity requireAdmin(HttpServletRequest request) {
         AdminUserEntity user = requireUser(request);
         String role = currentRole(user);
         if (role == null) {
@@ -143,7 +167,10 @@ public class AdminAuthSupport {
         if (expectedToken.isBlank() || suppliedToken == null || suppliedToken.isBlank()) {
             return false;
         }
-        return expectedToken.equals(suppliedToken.trim());
+        // 常量时间比较：逐字节 equals 可被计时侧信道逐位试探出内部令牌
+        return java.security.MessageDigest.isEqual(
+                expectedToken.getBytes(java.nio.charset.StandardCharsets.UTF_8),
+                suppliedToken.trim().getBytes(java.nio.charset.StandardCharsets.UTF_8));
     }
 
     /**

backend-java/src/main/java/com/nanri/aiimage/common/security/AuthProperties.java

@@ -1,4 +1,4 @@
-package com.nanri.aiimage.modules.auth.config;
+package com.nanri.aiimage.common.security;
 
 import lombok.Data;
 import org.springframework.boot.context.properties.ConfigurationProperties;

backend-java/src/main/java/com/nanri/aiimage/common/security/DeviceSessionPolicy.java

@@ -1,4 +1,4 @@
-package com.nanri.aiimage.modules.auth.support;
+package com.nanri.aiimage.common.security;
 
 import com.nanri.aiimage.common.exception.BusinessException;
 import io.jsonwebtoken.Claims;

backend-java/src/main/java/com/nanri/aiimage/common/security/JwtService.java

@@ -1,6 +1,6 @@
-package com.nanri.aiimage.modules.auth.service;
+package com.nanri.aiimage.common.security;
 
-import com.nanri.aiimage.modules.auth.config.AuthProperties;
+import com.nanri.aiimage.common.security.AuthProperties;
 import io.jsonwebtoken.Claims;
 import io.jsonwebtoken.JwtException;
 import io.jsonwebtoken.Jwts;
@@ -26,6 +26,26 @@ public class JwtService {
 
     private volatile SecretKey cachedKey;
 
+    @org.springframework.beans.factory.annotation.Value("${spring.profiles.active:}")
+    private String activeProfiles;
+
+    /**
+     * 生产 profile 下密钥缺失即拒绝启动（2026-09 全维度审查）。
+     *
+     * <p>此前未配置时静默回退到内置默认密钥——那是公开值，任何人可据此伪造 token；
+     * 一次 env 丢失就会让全站鉴权形同虚设，只留一行 warn 日志不足以拦住发布。
+     * 本地/测试 profile 保持宽松（否则开发无法启动）。
+     */
+    @jakarta.annotation.PostConstruct
+    void requireSecretInServerProfile() {
+        boolean serverProfile = activeProfiles != null && activeProfiles.contains("server");
+        if (serverProfile && (props.getJwtSecret() == null || props.getJwtSecret().isBlank())) {
+            throw new IllegalStateException(
+                    "server profile 下必须配置 aiimage.auth.jwt-secret（环境变量 AIIMAGE_JWT_SECRET）——"
+                            + "缺失时会回退到公开的默认密钥，token 可被任意伪造");
+        }
+    }
+
     private SecretKey signingKey() {
         SecretKey key = cachedKey;
         if (key == null) {

backend-java/src/main/java/com/nanri/aiimage/common/service/NodeSharedStore.java

@@ -0,0 +1,179 @@
+package com.nanri.aiimage.common.service;
+
+import com.fasterxml.jackson.databind.ObjectMapper;
+import lombok.extern.slf4j.Slf4j;
+import org.springframework.data.redis.core.StringRedisTemplate;
+
+import java.time.Duration;
+import java.util.concurrent.ConcurrentHashMap;
+import java.util.concurrent.atomic.AtomicLong;
+
+/**
+ * 跨节点的小容量状态存储（2026-09 全维度审查 D2）。
+ *
+ * <p>用途：导入/导出任务的进度、归属用户、分组等状态。原先只存节点本地内存，
+ * 客户端轮询落到另一节点就报"任务不存在"（nginx 的 user_id 亲和只覆盖常态）。
+ * 现在本地 Map 作快路径、Redis 作跨节点真源，任何节点都能读到。
+ *
+ * <p>写节流：调用方可能**逐行**刷新进度（数十万行），逐次写 Redis 不可接受；
+ * 默认 500ms 内只写一次（终态用 {@link #putNow} 立即落库）。
+ *
+ * <p>语义取舍：读-改-写不保证原子（进度类状态可接受）；{@code redis}/{@code objectMapper}
+ * 为空时退化为纯本地（单测与未注入场景）。
+ */
+@Slf4j
+public final class NodeSharedStore<K, V> {
+
+    private static final long DEFAULT_WRITE_THROTTLE_MILLIS = 500L;
+
+    private final String keyPrefix;
+    private final Duration ttl;
+    private final Class<V> valueType;
+    private final StringRedisTemplate redis;
+    private final ObjectMapper objectMapper;
+    private final long writeThrottleMillis;
+
+    private final ConcurrentHashMap<K, V> local = new ConcurrentHashMap<>();
+    private final ConcurrentHashMap<K, AtomicLong> lastWriteAt = new ConcurrentHashMap<>();
+
+    public NodeSharedStore(String keyPrefix,
+                           Duration ttl,
+                           Class<V> valueType,
+                           StringRedisTemplate redis,
+                           ObjectMapper objectMapper) {
+        this(keyPrefix, ttl, valueType, redis, objectMapper, DEFAULT_WRITE_THROTTLE_MILLIS);
+    }
+
+    public NodeSharedStore(String keyPrefix,
+                           Duration ttl,
+                           Class<V> valueType,
+                           StringRedisTemplate redis,
+                           ObjectMapper objectMapper,
+                           long writeThrottleMillis) {
+        this.keyPrefix = keyPrefix;
+        this.ttl = ttl;
+        this.valueType = valueType;
+        this.redis = redis;
+        this.objectMapper = objectMapper;
+        this.writeThrottleMillis = Math.max(0L, writeThrottleMillis);
+    }
+
+    /** 本地优先；本地没有则读 Redis 并回填本地（跨节点可见）。 */
+    public V get(K key) {
+        if (key == null) {
+            return null;
+        }
+        V cached = local.get(key);
+        if (cached != null) {
+            return cached;
+        }
+        V remote = readRemote(key);
+        if (remote != null) {
+            local.put(key, remote);
+        }
+        return remote;
+    }
+
+    public boolean containsKey(K key) {
+        return get(key) != null;
+    }
+
+    /** 写入并（按节流）同步到 Redis。 */
+    public void put(K key, V value) {
+        if (key == null || value == null) {
+            return;
+        }
+        local.put(key, value);
+        if (!throttleAllowsWrite(key)) {
+            return;
+        }
+        writeRemote(key, value);
+    }
+
+    /** 立即写入 Redis（终态、归属等一次性状态用）。 */
+    public void putNow(K key, V value) {
+        if (key == null || value == null) {
+            return;
+        }
+        local.put(key, value);
+        writeRemote(key, value);
+    }
+
+    /** 删除（本地 + Redis），返回删除前的值（可能为 null）。 */
+    public V remove(K key) {
+        if (key == null) {
+            return null;
+        }
+        V previous = local.remove(key);
+        lastWriteAt.remove(key);
+        if (redis != null) {
+            try {
+                redis.delete(fullKey(key));
+            } catch (Exception ex) {
+                log.warn("[node-shared-store] 删除远端状态失败 key={} msg={}", fullKey(key), ex.getMessage());
+            }
+        }
+        return previous;
+    }
+
+    /** 本地快照（仅用于日志/统计；不含其它节点的写入）。 */
+    public int localSize() {
+        return local.size();
+    }
+
+    /**
+     * 本节点已知条目的快照（副本，可安全遍历）。
+     *
+     * <p>用途：保留期清理等维护动作需要遍历键；跨节点的过期回收由 Redis TTL 兜底，
+     * 因此这里只返回本节点写入过的条目即可。
+     */
+    public java.util.Map<K, V> localEntriesSnapshot() {
+        return new java.util.LinkedHashMap<>(local);
+    }
+
+    private boolean throttleAllowsWrite(K key) {
+        if (writeThrottleMillis <= 0L) {
+            return true;
+        }
+        long now = System.currentTimeMillis();
+        AtomicLong last = lastWriteAt.computeIfAbsent(key, ignored -> new AtomicLong(0L));
+        long previous = last.get();
+        if (now - previous < writeThrottleMillis) {
+            return false;
+        }
+        return last.compareAndSet(previous, now);
+    }
+
+    private V readRemote(K key) {
+        if (redis == null || objectMapper == null) {
+            return null;
+        }
+        try {
+            String json = redis.opsForValue().get(fullKey(key));
+            if (json == null || json.isBlank()) {
+                return null;
+            }
+            return objectMapper.readValue(json, valueType);
+        } catch (Exception ex) {
+            log.warn("[node-shared-store] 读取远端状态失败 key={} msg={}", fullKey(key), ex.getMessage());
+            return null;
+        }
+    }
+
+    private void writeRemote(K key, V value) {
+        if (redis == null || objectMapper == null) {
+            return;
+        }
+        try {
+            String json = objectMapper.writeValueAsString(value);
+            redis.opsForValue().set(fullKey(key), json, ttl);
+        } catch (Exception ex) {
+            // 写失败不影响本地进度（下次 put 会重试），只留线索
+            log.warn("[node-shared-store] 写入远端状态失败 key={} msg={}", fullKey(key), ex.getMessage());
+        }
+    }
+
+    private String fullKey(K key) {
+        return keyPrefix + ":" + key;
+    }
+}

backend-java/src/main/java/com/nanri/aiimage/common/service/ResultDownloadResolver.java

@@ -0,0 +1,46 @@
+package com.nanri.aiimage.common.service;
+
+import com.nanri.aiimage.common.exception.BusinessException;
+import com.nanri.aiimage.modules.file.service.oss.OssStorageService;
+import com.nanri.aiimage.modules.task.mapper.FileResultMapper;
+import com.nanri.aiimage.modules.task.model.entity.FileResultEntity;
+import lombok.RequiredArgsConstructor;
+import org.springframework.stereotype.Component;
+
+import java.util.Objects;
+
+/**
+ * 结果文件下载直链解析（2026-09 全维度审查去重）。
+ *
+ * <p>此前 11 个业务模块各写一份逐字相同的 {@code resolveResultDownloadUrl}，下载鉴权口径
+ * 或 OSS 直链规则一调整就要改 11 处；similarasin 已自行分叉成返回 record 的第 12 种写法，
+ * 说明"改的时候漏一个"已经开始发生。
+ *
+ * <p>文件名解析（{@code resolveResultDownloadFilename}）**未**收口：各模块兜底文件名策略
+ * 确实不同（模块名+id / 源文件名派生 stem），属业务差异而非重复。
+ */
+@Component
+@RequiredArgsConstructor
+public class ResultDownloadResolver {
+
+    private final FileResultMapper fileResultMapper;
+    private final OssStorageService ossStorageService;
+
+    /**
+     * 解析结果文件下载直链。
+     *
+     * @param resultId   结果行 id（biz_file_result）
+     * @param userId     当前用户 id，必须与结果行归属一致
+     * @param moduleType 期望的模块类型
+     */
+    public String resolveUrl(Long resultId, Long userId, String moduleType) {
+        FileResultEntity row = fileResultMapper.selectById(resultId);
+        if (row == null || !moduleType.equals(row.getModuleType()) || !Objects.equals(userId, row.getUserId())) {
+            throw new BusinessException("记录不存在");
+        }
+        if (row.getResultFileUrl() == null || row.getResultFileUrl().isBlank()) {
+            throw new BusinessException("暂无可下载文件");
+        }
+        return ossStorageService.generateFreshDownloadUrl(row.getResultFileUrl());
+    }
+}

backend-java/src/main/java/com/nanri/aiimage/common/service/TaskOwnerForwardService.java

@@ -13,10 +13,12 @@ import org.springframework.http.ResponseEntity;
 import org.springframework.http.client.SimpleClientHttpRequestFactory;
 import org.springframework.stereotype.Service;
 import org.springframework.util.StreamUtils;
+import org.springframework.web.client.ResourceAccessException;
 import org.springframework.web.client.RestClient;
 import org.springframework.web.util.ContentCachingRequestWrapper;
 
 import java.io.IOException;
+import java.net.ConnectException;
 import java.util.Enumeration;
 import java.util.List;
 import java.util.Set;
@@ -28,6 +30,11 @@ public class TaskOwnerForwardService {
 
     public static final String FORWARDED_HEADER = "X-AIIMAGE-Owner-Forwarded";
 
+    /** 连接类失败的重试次数（含首次）。对端滚动重启时通常几秒内即可恢复。 */
+    private static final int CONNECT_RETRY_TIMES = 3;
+    /** 第 n 次重试前的退避：1s、2s（总等待不超过 3s，不长时间占用请求线程）。 */
+    private static final long CONNECT_RETRY_BACKOFF_MILLIS = 1000L;
+
     private static final Set<String> HOP_BY_HOP_HEADERS = Set.of(
             "connection",
             "keep-alive",
@@ -58,12 +65,60 @@ public class TaskOwnerForwardService {
         HttpHeaders headers = copyForwardHeaders(request, ex.getCurrentInstanceId());
         log.info("[instance-routing] forward {} taskId={} operation={} owner={} current={} url={}",
                 method, ex.getTaskId(), ex.getOperation(), ex.getOwnerInstanceId(), ex.getCurrentInstanceId(), url);
+        return forwardWithConnectRetry(method, url, headers, body, ex);
+    }
+
+    /**
+     * 转发带连接级重试。
+     *
+     * <p>对端实例在部署窗口（原地换 JAR + 两节点滚动重启）内会有几秒的 Connection refused。
+     * 连接都没建立起来说明请求没到达对端，此时重放是安全的；而读超时不重试——对端可能
+     * 已经在处理，盲目重放会造成重复提交。线上由此丢过用户提交的结果。
+     */
+    private ResponseEntity<byte[]> forwardWithConnectRetry(HttpMethod method, String url,
+                                                           HttpHeaders headers, byte[] body,
+                                                           TaskOwnerMismatchException ex) {
+        RuntimeException lastError = null;
+        for (int attempt = 1; attempt <= CONNECT_RETRY_TIMES; attempt++) {
+            try {
                 return restClient().method(method)
                         .uri(url)
                         .headers(target -> target.addAll(headers))
                         .body(body)
                         .retrieve()
                         .toEntity(byte[].class);
+            } catch (ResourceAccessException accessError) {
+                if (!isConnectFailure(accessError)) {
+                    throw accessError;
+                }
+                lastError = accessError;
+                log.warn("[instance-routing] 转发连接失败，第 {}/{} 次 url={} taskId={} 原因={}",
+                        attempt, CONNECT_RETRY_TIMES, url, ex.getTaskId(), accessError.getMessage());
+                if (attempt < CONNECT_RETRY_TIMES) {
+                    sleepQuietly(CONNECT_RETRY_BACKOFF_MILLIS * attempt);
+                }
+            }
+        }
+        throw lastError;
+    }
+
+    private static boolean isConnectFailure(Throwable error) {
+        Throwable cursor = error;
+        while (cursor != null) {
+            if (cursor instanceof ConnectException) {
+                return true;
+            }
+            cursor = cursor.getCause();
+        }
+        return false;
+    }
+
+    private static void sleepQuietly(long millis) {
+        try {
+            Thread.sleep(millis);
+        } catch (InterruptedException interrupted) {
+            Thread.currentThread().interrupt();
+        }
     }
 
     private String resolveUrl(TaskOwnerMismatchException ex, String path) {

backend-java/src/main/java/com/nanri/aiimage/common/util/BoundedLruCache.java

@@ -0,0 +1,70 @@
+package com.nanri.aiimage.common.util;
+
+import java.util.Collections;
+import java.util.LinkedHashMap;
+import java.util.Map;
+import java.util.function.Function;
+
+/**
+ * 有界 LRU 缓存：容量超限时自动淘汰最久未使用的条目。
+ *
+ * <p>用于按「外部端点」缓存长生命周期资源（HttpClient / RestClient）。这类资源各自持有
+ * 连接池与 selector 线程，无界累积会持续泄漏线程与内存：代理端点每次提取往往是新的
+ * IP:port（jikip 提取），无上限的缓存只增不减。
+ *
+ * <p>淘汰时只从缓存移除引用，不做显式关闭：JDK 的 HttpClientImpl 注册了 Cleaner，
+ * 对象不可达后由 GC 回收并关闭其 selector 线程；显式关闭反而可能打断仍在途的请求。
+ */
+public final class BoundedLruCache<K, V> {
+
+    /** 默认容量：代理端点数量级远小于此，足够覆盖热点端点又不至于累积。 */
+    public static final int DEFAULT_MAX_SIZE = 64;
+
+    private final int maxSize;
+    private final Map<K, V> store;
+
+    public BoundedLruCache() {
+        this(DEFAULT_MAX_SIZE);
+    }
+
+    public BoundedLruCache(int maxSize) {
+        this.maxSize = Math.max(1, maxSize);
+        // accessOrder=true 使 get 也刷新顺序（真正的 LRU）；synchronizedMap 保证其线程安全
+        this.store = Collections.synchronizedMap(new LinkedHashMap<>(16, 0.75f, true) {
+            @Override
+            protected boolean removeEldestEntry(Map.Entry<K, V> eldest) {
+                return size() > BoundedLruCache.this.maxSize;
+            }
+        });
+    }
+
+    /**
+     * 取缓存值，缺失时用 loader 计算并放入。
+     *
+     * <p>与 {@code ConcurrentHashMap.computeIfAbsent} 不同，此处不保证 loader 的原子性：
+     * 并发首次访问同一 key 时可能各自构造一次，随后其中一个覆盖另一个。对
+     * HttpClient/RestClient 这类构造廉价且幂等的资源可接受，换来的是锁粒度更小。
+     */
+    public V computeIfAbsent(K key, Function<K, V> loader) {
+        V existing = store.get(key);
+        if (existing != null) {
+            return existing;
+        }
+        V created = loader.apply(key);
+        store.put(key, created);
+        return created;
+    }
+
+    public int size() {
+        return store.size();
+    }
+
+    /** 当前容量上限，供日志与测试断言使用。 */
+    public int maxSize() {
+        return maxSize;
+    }
+
+    public void clear() {
+        store.clear();
+    }
+}

backend-java/src/main/java/com/nanri/aiimage/common/util/ExcelStreamReader.java

@@ -56,6 +56,13 @@ public final class ExcelStreamReader {
         default void onHeader(String sheetName, Integer sheetNo, Map<Integer, String> headerMap) throws Exception {
         }
 
+        /**
+         * 表头之后回调本次 sheet 的近似总行数（EasyExcel 基于 sheet 尺寸，可能为 null）。
+         * 流式解析无法在读完前得到精确行数，需要展示进度/做前置上限校验的调用方可用它近似。
+         */
+        default void onSheetTotal(String sheetName, Integer sheetNo, Integer approximateTotalRows) throws Exception {
+        }
+
         void onRow(String sheetName, Integer sheetNo, int rowIndex, Map<Integer, String> headerMap, Map<Integer, String> rowMap) throws Exception;
     }
 
@@ -77,6 +84,7 @@ public final class ExcelStreamReader {
             currentHeaderMap = normalizedHeadMap;
             try {
                 handler.onHeader(sheetName(context), sheetNo(context), currentHeaderMap);
+                handler.onSheetTotal(sheetName(context), sheetNo(context), approximateTotalRows(context));
             } catch (RuntimeException ex) {
                 throw ex;
             } catch (Exception ex) {
@@ -109,6 +117,15 @@ public final class ExcelStreamReader {
         public void doAfterAllAnalysed(AnalysisContext context) {
         }
 
+        private Integer approximateTotalRows(AnalysisContext context) {
+            try {
+                return context.readSheetHolder() == null ? null
+                        : context.readSheetHolder().getApproximateTotalRowNumber();
+            } catch (Exception ex) {
+                return null;
+            }
+        }
+
         private String sheetName(AnalysisContext context) {
             return context.readSheetHolder() == null ? "" : context.readSheetHolder().getSheetName();
         }

backend-java/src/main/java/com/nanri/aiimage/common/util/PendingDeleteJournal.java

@@ -0,0 +1,138 @@
+package com.nanri.aiimage.common.util;
+
+import lombok.extern.slf4j.Slf4j;
+
+import java.io.IOException;
+import java.nio.charset.StandardCharsets;
+import java.nio.file.Files;
+import java.nio.file.Path;
+import java.nio.file.StandardCopyOption;
+import java.nio.file.StandardOpenOption;
+import java.util.ArrayList;
+import java.util.Collection;
+import java.util.LinkedHashSet;
+import java.util.List;
+import java.util.Set;
+
+/**
+ * 待删对象队列的本地落盘日志：进程重启后仍能恢复「待删除对象」清单。
+ *
+ * <p>使用场景：对象存储删除失败后的补偿队列原本只在节点内存里，重启即丢，
+ * 对应对象会一直残留在桶里直到生命周期规则过期。这里用一行一个 objectKey 的
+ * 追加日志做最小持久化——入队追加、队列收敛后整体重写、启动时回放。
+ *
+ * <p>并发：所有方法内部同步，调用方无需额外加锁；文件损坏/读写异常只记日志不抛出，
+ * 保证补偿链路本身不会因为落盘失败而中断业务。
+ *
+ * <p>上限：{@code maxEntries} 用于防止日志在异常堆积时无界增长（超出后丢弃最早的记录，
+ * 与内存队列的容量准入语义一致——丢的是「待删对象」，最坏结果是对象残留）。
+ */
+@Slf4j
+public final class PendingDeleteJournal {
+
+    private final Path path;
+    private final int maxEntries;
+    private final Object lock = new Object();
+
+    public PendingDeleteJournal(Path path, int maxEntries) {
+        this.path = path;
+        this.maxEntries = Math.max(1, maxEntries);
+    }
+
+    public Path getPath() {
+        return path;
+    }
+
+    /** 追加一条待删对象（重复追加由回放时的 Set 语义去重）。 */
+    public void record(String objectKey) {
+        if (objectKey == null || objectKey.isBlank()) {
+            return;
+        }
+        synchronized (lock) {
+            try {
+                Path parent = path.getParent();
+                if (parent != null) {
+                    Files.createDirectories(parent);
+                }
+                Files.writeString(path, sanitize(objectKey) + System.lineSeparator(),
+                        StandardCharsets.UTF_8, StandardOpenOption.CREATE, StandardOpenOption.APPEND);
+            } catch (Exception ex) {
+                log.warn("[delete-journal] 追加待删对象失败 path={} msg={}", path, ex.getMessage());
+            }
+        }
+    }
+
+    /** 启动回放：返回日志中的待删对象（按首次出现顺序去重）。 */
+    public List<String> readAll() {
+        synchronized (lock) {
+            if (!Files.isRegularFile(path)) {
+                return List.of();
+            }
+            List<String> lines;
+            try {
+                lines = Files.readAllLines(path, StandardCharsets.UTF_8);
+            } catch (Exception ex) {
+                log.warn("[delete-journal] 读取待删对象日志失败 path={} msg={}", path, ex.getMessage());
+                return List.of();
+            }
+            Set<String> unique = new LinkedHashSet<>();
+            for (String line : lines) {
+                if (line == null) {
+                    continue;
+                }
+                String key = line.trim();
+                if (!key.isEmpty()) {
+                    unique.add(key);
+                }
+            }
+            return new ArrayList<>(unique);
+        }
+    }
+
+    /** 队列收敛（成功删除/引用仍在）后整体重写为剩余的待删对象；剩余为空则删除日志。 */
+    public void rewrite(Collection<String> remaining) {
+        synchronized (lock) {
+            Set<String> keep = new LinkedHashSet<>();
+            if (remaining != null) {
+                for (String key : remaining) {
+                    if (key != null && !key.isBlank()) {
+                        keep.add(sanitize(key));
+                        if (keep.size() >= maxEntries) {
+                            break;
+                        }
+                    }
+                }
+            }
+            try {
+                if (keep.isEmpty()) {
+                    Files.deleteIfExists(path);
+                    return;
+                }
+                Path parent = path.getParent();
+                if (parent != null) {
+                    Files.createDirectories(parent);
+                }
+                Path tmp = path.resolveSibling(path.getFileName() + ".tmp");
+                StringBuilder content = new StringBuilder();
+                for (String key : keep) {
+                    content.append(key).append(System.lineSeparator());
+                }
+                Files.writeString(tmp, content.toString(), StandardCharsets.UTF_8,
+                        StandardOpenOption.CREATE, StandardOpenOption.TRUNCATE_EXISTING);
+                try {
+                    Files.move(tmp, path, StandardCopyOption.REPLACE_EXISTING, StandardCopyOption.ATOMIC_MOVE);
+                } catch (IOException atomicUnsupported) {
+                    // 少数文件系统不支持 ATOMIC_MOVE，退化为普通替换
+                    Files.move(tmp, path, StandardCopyOption.REPLACE_EXISTING);
+                }
+            } catch (Exception ex) {
+                log.warn("[delete-journal] 重写待删对象日志失败 path={} msg={}", path, ex.getMessage());
+            }
+        }
+    }
+
+    /** 单行一条记录：去掉换行避免破坏行结构。 */
+    private static String sanitize(String objectKey) {
+        return objectKey.replace('\n', ' ').replace('\r', ' ').trim();
+    }
+}

backend-java/src/main/java/com/nanri/aiimage/common/util/SecretMasking.java

@@ -0,0 +1,52 @@
+package com.nanri.aiimage.common.util;
+
+import java.net.URI;
+
+/**
+ * 敏感串脱敏工具（2026-09 全维度审查后收口）。
+ *
+ * <p>背景：代理提取链接形态为 {@code http://user:pass@host:port}，代码中曾有多处
+ * 直接把整串打进日志，导致用户代理账号密码长期留在应用日志与 docker logs 里。
+ */
+public final class SecretMasking {
+
+    private SecretMasking() {
+    }
+
+    /** 通用掩码：保留前 4 与后 4 字符；过短则整体掩掉。 */
+    public static String mask(String value) {
+        if (value == null || value.isBlank()) {
+            return "";
+        }
+        String text = value.trim();
+        if (text.length() <= 8) {
+            return "****";
+        }
+        return text.substring(0, 4) + "****" + text.substring(text.length() - 4);
+    }
+
+    /** 代理掩码：隐去账号密码，保留 scheme://host:port 便于运维核对。 */
+    public static String maskProxy(String value) {
+        if (value == null || value.isBlank()) {
+            return "";
+        }
+        try {
+            URI uri = URI.create(value.trim());
+            if (uri.getHost() == null || uri.getHost().isBlank()) {
+                return mask(value);
+            }
+            StringBuilder masked = new StringBuilder();
+            masked.append(uri.getScheme() == null ? "http" : uri.getScheme()).append("://");
+            if (uri.getUserInfo() != null && !uri.getUserInfo().isBlank()) {
+                masked.append("***@");
+            }
+            masked.append(uri.getHost());
+            if (uri.getPort() > 0) {
+                masked.append(':').append(uri.getPort());
+            }
+            return masked.toString();
+        } catch (Exception ex) {
+            return mask(value);
+        }
+    }
+}

backend-java/src/main/java/com/nanri/aiimage/common/util/ThreadPools.java

@@ -0,0 +1,47 @@
+package com.nanri.aiimage.common.util;
+
+import java.util.concurrent.ExecutorService;
+import java.util.concurrent.LinkedBlockingQueue;
+import java.util.concurrent.ThreadPoolExecutor;
+import java.util.concurrent.TimeUnit;
+
+/**
+ * 有界线程池工厂（2026-09 全维度审查补）。
+ *
+ * <p>业务里多处使用 {@code Executors.newFixedThreadPool}：它内部是**无界**
+ * {@code LinkedBlockingQueue}，任务堆积时永远不会触发拒绝策略，会把内存吃满
+ * （表现为 OOM，或整机因 GC 变慢导致所有任务一起劣化）。
+ * 统一改为有界队列 + {@code CallerRunsPolicy}：队列满时在提交线程执行，形成天然背压。
+ */
+public final class ThreadPools {
+
+    private ThreadPools() {
+    }
+
+    /** 默认队列容量：足以吸收突发，又不至于无界堆积。 */
+    public static final int DEFAULT_QUEUE_CAPACITY = 512;
+
+    /** 有界固定线程池（daemon 线程，空闲可回收）。 */
+    public static ExecutorService boundedFixed(String threadNamePrefix, int threads) {
+        return boundedFixed(threadNamePrefix, threads, DEFAULT_QUEUE_CAPACITY);
+    }
+
+    /** 有界固定线程池（显式队列容量）。 */
+    public static ExecutorService boundedFixed(String threadNamePrefix, int threads, int queueCapacity) {
+        ThreadPoolExecutor executor = new ThreadPoolExecutor(
+                Math.max(1, threads),
+                Math.max(1, threads),
+                // keepAliveTime 必须 > 0：下面开了 allowCoreThreadTimeOut，
+                // 传 0 会让构造器直接抛 "Core threads must have nonzero keep alive times"
+                60L, TimeUnit.SECONDS,
+                new LinkedBlockingQueue<>(Math.max(1, queueCapacity)),
+                runnable -> {
+                    Thread thread = new Thread(runnable, threadNamePrefix);
+                    thread.setDaemon(true);
+                    return thread;
+                },
+                new ThreadPoolExecutor.CallerRunsPolicy());
+        executor.allowCoreThreadTimeOut(true);
+        return executor;
+    }
+}

backend-java/src/main/java/com/nanri/aiimage/config/AdminApiGuardFilter.java

@@ -3,7 +3,8 @@ package com.nanri.aiimage.config;
 import com.fasterxml.jackson.databind.ObjectMapper;
 import com.nanri.aiimage.common.api.ApiResponse;
 import com.nanri.aiimage.common.exception.BusinessException;
-import com.nanri.aiimage.modules.admin.support.AdminAuthSupport;
+import com.nanri.aiimage.common.security.AdminAuthSupport;
+import com.nanri.aiimage.common.security.DeviceSessionPolicy;
 import jakarta.servlet.FilterChain;
 import jakarta.servlet.ServletException;
 import jakarta.servlet.http.HttpServletRequest;
@@ -63,6 +64,24 @@ public class AdminApiGuardFilter extends OncePerRequestFilter {
     private static final String[] USER_TOOL_PREFIXES = {
             "/api/collect-data",
             "/api/price-track",
+            // 2026-09 全维度审查补：以下前缀此前完全在守卫范围之外。
+            // /api/files 匿名可上传（2GB/次，可耗尽临时盘）；/api/digital-human 匿名可发布/删除版本；
+            // /api/brand 的 fileUrl 曾可直接请求任意地址（/api/image-video 与 /api/task-file-jobs
+            // 客户端零调用，已移入下方无条件名单）。
+            "/api/files",
+            "/api/digital-human",
+            "/api/brand",
+            "/api/appearance-patent",
+            "/api/similar-asin",
+            "/api/query-asin",
+            "/api/patrol-delete",
+            "/api/product-risk-resolve",
+            "/api/shop-match",
+            "/api/shop-data-crawl",
+            "/api/withdraw",
+            // /api/tasks/{taskId}/interrupted 仅凭 taskId 即可把 RUNNING 任务置为 FAILED，
+            // 匿名遍历 taskId 就能批量打断线上任务
+            "/api/tasks",
     };
 
     /**
@@ -72,6 +91,16 @@ public class AdminApiGuardFilter extends OncePerRequestFilter {
     private static final String[] SELF_SERVICE_PREFIXES = {
             "/api/user-secrets",
             "/api/notifications",
+            // 2026-09-14：这三组前缀**桌面端 Python 侧零调用**（实测），只有带 JWT 的网页端在用，
+            // 因此不必等客户端铺开即可无条件收紧（其余用户态前缀仍在 user-tool-guard-enabled 开关后面）
+            "/api/image-video",
+            "/api/task-file-jobs",
+            // 其全部接口已在 controller 内 requireAdmin（含可换取员工店铺登录令牌的 /shops/open），
+            // 纳入守卫是"鉴权失败返回统一 401 体"的第二层
+            "/api/ziniao",
+            // 2026-09 全维度审查补：内部端点此前仅靠 controller 自校验令牌，纳入守卫后
+            // 不带令牌的请求直接 401（带可信令牌的仍由 doFilterInternal 放行）
+            "/api/internal",
     };
 
     private final AdminAuthSupport adminAuthSupport;
@@ -125,7 +154,14 @@ public class AdminApiGuardFilter extends OncePerRequestFilter {
             ApiResponse<Void> body = ex.getCode() == null
                     ? ApiResponse.fail(ex.getMessage())
                     : ApiResponse.fail(ex.getCode(), ex.getMessage());
+            // 401（登录已过期）与被顶下线是前端定时轮询（/api/notifications/summary、/api/user-secrets 等）
+            // 的常态：线上单节点一天近 3000 条，会把真实业务错误淹没。与 GlobalExceptionHandler
+            // 的 isRoutineAuthNoise 同一口径降为 debug。
+            if (isRoutineAuthNoise(ex.getCode())) {
+                log.debug("[admin-guard] {} {} rejected: {}", request.getMethod(), request.getRequestURI(), ex.getMessage());
+            } else {
                 log.warn("[admin-guard] {} {} rejected: {}", request.getMethod(), request.getRequestURI(), ex.getMessage());
+            }
             response.setStatus(HttpServletResponse.SC_OK);
             response.setContentType("application/json;charset=UTF-8");
             response.getWriter().write(objectMapper.writeValueAsString(body));
@@ -140,6 +176,12 @@ public class AdminApiGuardFilter extends OncePerRequestFilter {
         chain.doFilter(request, response);
     }
 
+    /** 登录态过期 / 被其他设备顶下线：前端轮询的常态噪声，不占 WARN。 */
+    private static boolean isRoutineAuthNoise(Integer code) {
+        return Integer.valueOf(401).equals(code)
+                || Integer.valueOf(DeviceSessionPolicy.CODE_KICKED).equals(code);
+    }
+
     /** 命中受保护前缀（/api/admin、/debug、自助接口、用户态工具前缀及其子路径）才进入鉴权，其余请求直接放行。 */
     private boolean isGuarded(String uri) {
         if (matchesPrefix(uri, ADMIN_API_PREFIX) || matchesPrefix(uri, DEBUG_PREFIX)) {

backend-java/src/main/java/com/nanri/aiimage/config/AppearancePatentProperties.java

@@ -30,15 +30,24 @@ public class AppearancePatentProperties {
     private int llmFirstAttemptReadTimeoutMillis = 60000;
     private int llmBatchSize = 10;
     /**
-     * 批内行级并发数，默认等于批量大小
+     * 批内行级并发数。批次串行提交，每行串行发 2 个 LLM 请求，
+     * 故该值≈单任务对 LLM 网关的瞬时并发；默认与品牌检测同为 5，避免多任务并行时成倍放大。
      */
-    private int llmRowConcurrency = 10;
+    private int llmRowConcurrency = 5;
     /**
      * 每行每个 LLM 请求的重试次数（含首次）
      */
     private int llmRetryTimes = 3;
     private int staleTimeoutMinutes = 30;
     private String staleFinalizeCron = "0 */2 * * * *";
+
+    /**
+     * 二次判死线（心跳正常但连续 N 分钟无结果分片上报）阈值，分钟；<=0 关闭本线。
+     * 既有判定以 Redis heartbeat stale 为主信号，但心跳随 Python HTTP 心跳每分钟刷新——
+     * 主线程卡死时心跳线程照发，任务永远不 stale（与生产 28131 同型缺口）。
+     * 本线改看 biz_task_scope_state.last_chunk_at（仅分片上传时刷新）。
+     */
+    private int noResultUploadTimeoutMinutes = 180;
     /**
      * 末尾不足一批的数据等待该时长后强制提交检测。
      * Python 回传超时后会直接封口；这个配置处理 Python 慢回传但仍未超时的零头批次。

backend-java/src/main/java/com/nanri/aiimage/config/BrandCheckProperties.java

@@ -10,10 +10,27 @@ public class BrandCheckProperties {
     private String path = "/brand_check";
     private String token = "";
     private String defaultStrategy = "Terms";
-    /** 单品牌商标查询的重试次数（含首次），连续失败超过该次数才判定为查询失败。 */
+    /**
-    private int retryTimes = 3;
+     * 单品牌商标查询的重试次数（含首次），连续失败超过该次数才判定为查询失败。
-    /** 每次查询失败后到下一次重试前的等待毫秒数。 */
+     * 原为 3：16890 偶发限流几秒内即恢复，3 次（前两次间隔各 1s）恢复不了就把结论
+     * 写成「查询失败」，对客户是硬伤；2026-09-14 与客户端品牌一致提到 10 次。
+     */
+    private int retryTimes = 10;
+    /** 每次查询失败后到下一次重试前的等待毫秒数（基准值，按重试轮次递增）。 */
     private int retryIntervalMillis = 1000;
+    /**
+     * 单次重试等待的上限毫秒数。等待按 retryIntervalMillis × 第几次重试 递增后封顶，
+     * 10 次尝试合计约 45s，既能覆盖限流窗口又不会让单个品牌长时间占住查询线程。
+     */
+    private int retryMaxIntervalMillis = 10000;
+    /**
+     * 一次批量检查的总耗时上限（毫秒）：整批用尽后不再重试，剩余品牌按「查询失败」收尾。
+     * 上限的意义不是省时间，而是给「分片回传」这类同步调用方一个时延上界——上游卡死时
+     * 单品牌 10 次重试曾把一次回传拖到 103.5 秒（taskId 28599），客户端重试预算耗尽后
+     * 中止了整个采集。首轮请求始终执行，故上游只是略慢时不会误降级。
+     * 设为 0 或负数表示不限制。
+     */
+    private int totalTimeoutMillis = 90000;
     private int connectTimeoutMillis = 10000;
     private int readTimeoutMillis = 60000;
 }

backend-java/src/main/java/com/nanri/aiimage/config/BrandProgressProperties.java

@@ -10,4 +10,11 @@ public class BrandProgressProperties {
     private long failedTtlHours = 2;
     private long heartbeatTimeoutMinutes = 15;
     private String staleCheckCron = "0 */2 * * * *";
+
+    /**
+     * 二次判死线（心跳正常但连续 N 分钟无结果上报）阈值，分钟；<=0 关闭本线。
+     * 既有心跳线以 updated_at/last_heartbeat_at 陈旧为判据，而前端心跳会持续刷新它们——
+     * 主线程卡死时心跳线程照发，任务永远命不中。本线改看结果上报时写入的 last_result_at。
+     */
+    private long noResultUploadTimeoutMinutes = 180;
 }

backend-java/src/main/java/com/nanri/aiimage/config/DeleteBrandProgressProperties.java

@@ -45,6 +45,19 @@ public class DeleteBrandProgressProperties {
      */
     private long withdrawStaleTimeoutMinutes = 30;
 
+    /**
+     * 「心跳正常但连续 N 分钟无结果分片上报」的二次判死阈值（分钟），默认 3 小时。
+     *
+     * <p>既有各模块心跳线的候选条件是 updated_at 陈旧，而客户端心跳会持续刷新 updated_at——
+     * 主线程卡死时心跳线程照发，任务永远命不中（生产 28131 卡死 12h+ 仍 RUNNING）。
+     * 本线改用 biz_task_scope_state.last_chunk_at（只随结果分片上报刷新）作判据，
+     * 从未上报（无 scope 行或全 NULL）的任务跳过，避免误杀首批较慢的正常任务。
+     */
+    private long noResultUploadTimeoutMinutes = 180;
+
+    /** 二次判死线开关：false = 整段不扫描（观察期与回滚用，改环境变量即生效）。 */
+    private boolean noResultUploadCheckEnabled = true;
+
     /**
      * 组装结果文件时允许缺失的最大分片数：缺失不超过该值时降级组装（缺失行状态写"未回传"），
      * 保证任务不因少量分片丢失整体白跑；超出时任务失败，但仍会在重试耗尽后尝试

backend-java/src/main/java/com/nanri/aiimage/config/DeviceLogOssProperties.java

@@ -0,0 +1,38 @@
+package com.nanri.aiimage.config;
+
+import lombok.Data;
+import org.springframework.boot.context.properties.ConfigurationProperties;
+
+/**
+ * 设备日志对象存储配置：指向主机B 独立部署的 MinIO 实例（非业务 MinIO）。
+ *
+ * <p>日志体积大、只保留 7 天，独立实例便于单独设生命周期规则与容量管理，
+ * 不挤占业务桶（nanri-ai-images 等）。endpoint 为空时上报接口直接失败，
+ * 不做静默回退（避免日志悄悄落到其他存储上而无人知情）。
+ */
+@Data
+@ConfigurationProperties(prefix = "aiimage.device-log-oss")
+public class DeviceLogOssProperties {
+
+    private String endpoint;
+    private String accessKeyId;
+    private String accessKeySecret;
+    private String bucket;
+
+    /**
+     * 日志保留天数：查询侧按此过滤（早于今天的 N-1 天不展示），
+     * 对象过期由 MinIO 桶生命周期规则在部署时同步设置（两侧口径保持一致）。
+     */
+    private Integer retentionDays;
+
+    public boolean configured() {
+        return endpoint != null && !endpoint.isBlank()
+                && accessKeyId != null && !accessKeyId.isBlank()
+                && accessKeySecret != null && !accessKeySecret.isBlank()
+                && bucket != null && !bucket.isBlank();
+    }
+
+    public int retentionDaysOrDefault() {
+        return retentionDays == null || retentionDays < 1 ? 7 : retentionDays;
+    }
+}

backend-java/src/main/java/com/nanri/aiimage/config/HttpClientPool.java

@@ -1,17 +1,21 @@
 package com.nanri.aiimage.config;
 
+import com.nanri.aiimage.common.util.BoundedLruCache;
+import lombok.extern.slf4j.Slf4j;
 import org.springframework.http.client.ClientHttpRequestFactory;
 import org.springframework.http.client.JdkClientHttpRequestFactory;
 
+import java.io.IOException;
+import java.io.InputStream;
 import java.net.Authenticator;
 import java.net.InetSocketAddress;
 import java.net.PasswordAuthentication;
 import java.net.ProxySelector;
 import java.net.URI;
 import java.net.http.HttpClient;
+import java.net.http.HttpRequest;
+import java.net.http.HttpResponse;
 import java.time.Duration;
-import java.util.Map;
-import java.util.concurrent.ConcurrentHashMap;
 
 /**
  * Task 77：外部 HTTP 客户端统一连接复用池。
@@ -20,6 +24,7 @@ import java.util.concurrent.ConcurrentHashMap;
  * 每次请求都重新建连。各客户端按自身超时创建独立的
  * JdkClientHttpRequestFactory（共享底层连接池），RestClient 单例懒加载。
  */
+@Slf4j
 public class HttpClientPool {
 
     private static volatile HttpClient sharedHttpClient;
@@ -55,6 +60,38 @@ public class HttpClientPool {
         }
     }
 
+    /**
+     * 打开远程文件流（带超时），调用方负责关闭返回的流。
+     *
+     * <p>替代裸 {@code URI.create(url).toURL().openStream()}：后者走 JVM 默认超时（0 = 无限），
+     * 上游半开连接或挂起时会把 Tomcat 工作线程无限占用（管理端批量打包可同时挂多个）。
+     * 返回的流是流式的，适用于「服务端代理下载 OSS 文件转发给浏览器」这类不落盘场景。
+     *
+     * @param url     远程地址
+     * @param timeout 等待响应超时（连接建立 + 响应头）；非法值钳制到 1 秒
+     * @throws IOException          非 2xx 响应或网络异常
+     * @throws InterruptedException 线程被中断
+     */
+    public static InputStream openStreamWithTimeout(String url, Duration timeout) throws IOException, InterruptedException {
+        Duration effective = (timeout == null || timeout.isZero() || timeout.isNegative())
+                ? Duration.ofSeconds(1)
+                : timeout;
+        HttpRequest request = HttpRequest.newBuilder(URI.create(url))
+                .timeout(effective)
+                .GET()
+                .build();
+        HttpResponse<InputStream> response = sharedHttpClient().send(request, HttpResponse.BodyHandlers.ofInputStream());
+        if (response.statusCode() / 100 != 2) {
+            try {
+                response.body().close();
+            } catch (Exception ignored) {
+                // 关闭失败不影响错误上报
+            }
+            throw new IOException("远程文件返回非 2xx: HTTP " + response.statusCode());
+        }
+        return response.body();
+    }
+
     /** 按 readTimeout（毫秒）创建共享连接池工厂；非法值钳制到最小正数。 */
     public static ClientHttpRequestFactory requestFactory(int readTimeoutMillis) {
         return requestFactory(readTimeoutMillis, null);
@@ -67,8 +104,12 @@ public class HttpClientPool {
     public static ClientHttpRequestFactory requestFactory(int readTimeoutMillis, String proxyUrl) {
         long safeReadTimeout = Math.max(1L, readTimeoutMillis);
         long callTimeout = configuredCallTimeoutMillis;
-        if (callTimeout > 0L) {
+        if (callTimeout > 0L && safeReadTimeout > callTimeout) {
-            safeReadTimeout = Math.min(safeReadTimeout, callTimeout);
+            // 读超时以调用方显式值为准，不再被全局 call-timeout 截断：
+            // LLM 长思考配的是 180s（llm-read-timeout-millis），曾被静默压到 90s，
+            // 导致请求在 90s 被掐断 → 上层重试 → 付费网关二次计费（2026-09-15 修复）。
+            // 各调用方的超时已由各自的 HttpConfigResolver 钳制，此处不再二次收敛。
+            log.debug("读超时 {}ms 超过全局 call-timeout {}ms，按调用方显式值生效", safeReadTimeout, callTimeout);
         }
         JdkClientHttpRequestFactory factory =
                 new JdkClientHttpRequestFactory(httpClientFor(proxyUrl));
@@ -130,5 +171,6 @@ public class HttpClientPool {
     private record ProxyEndpoint(String host, int port, String userInfo) {
     }
 
-    private static final Map<ProxyEndpoint, HttpClient> PROXY_CLIENTS = new ConcurrentHashMap<>();
+    private static final BoundedLruCache<ProxyEndpoint, HttpClient> PROXY_CLIENTS =
+            new BoundedLruCache<>(BoundedLruCache.DEFAULT_MAX_SIZE);
 }

backend-java/src/main/java/com/nanri/aiimage/config/ModuleCleanupProperties.java

@@ -1,5 +1,6 @@
 package com.nanri.aiimage.config;
 
+import com.nanri.aiimage.common.module.TaskModuleRegistry;
 import lombok.Data;
 import org.springframework.boot.context.properties.ConfigurationProperties;
 
@@ -19,5 +20,6 @@ public class ModuleCleanupProperties {
     private int batchSize = 500;
     // SHOP_DATA_CRAWL keeps one per-shop daily workbook in its task service
     // and must not be removed by the age-based sweep.
-    private List<String> moduleTypes = new ArrayList<>(List.of("DEDUPE", "SPLIT", "CONVERT", "DELETE_BRAND", "PRODUCT_RISK_RESOLVE", "PRICE_TRACK", "SHOP_MATCH", "PATROL_DELETE", "QUERY_ASIN", "WITHDRAW", "APPEARANCE_PATENT", "SIMILAR_ASIN", "COLLECT_DATA"));
+    /** 参与按天清理的模块：取自模块注册表（G6）。 */
+    private List<String> moduleTypes = new ArrayList<>(TaskModuleRegistry.ageCleanupModuleTypes());
 }

backend-java/src/main/java/com/nanri/aiimage/config/NotificationProperties.java

@@ -41,6 +41,38 @@ public class NotificationProperties {
     /** jikip 代理接口探测开关（复用余额查询接口判活，使用 user-secret 的 jikip 配置）。 */
     private boolean jikipProbeEnabled = true;
 
+    /** 麦象（18960 任务调度）异常扫描开关：任务停滞/失败/队列积压 → 管理员通知。 */
+    private boolean maixiangScanEnabled = true;
+
+    /**
+     * 麦象后台接口令牌（18960 console token）。留空=跳过麦象异常扫描——
+     * 该令牌与「跟价任务 API 地址」(priceTrackApiUrl) 一起构成后台只读接口的访问凭据。
+     */
+    private String maixiangConsoleToken = "";
+
+    /** 麦象批量任务停滞阈值（分钟）：status=0/1 且超过该时长无更新视为卡住。 */
+    private int maixiangStuckMinutes = 60;
+
+    /** 麦象单任务滞留阈值（分钟）：创建超时仍未完成（status=0/1）视为滞留/无人消费。 */
+    private int maixiangSingleStuckMinutes = 30;
+
+    /** 麦象任务失败告警阈值（条）：近 30 分钟窗口内失败数达到该值才告警。 */
+    private int maixiangFailMinCount = 1;
+
+    /** 麦象队列积压阈值（条）：task:queue 待处理数达到该值告警。 */
+    private int maixiangQueuePendingThreshold = 300;
+
+    /** 麦象队列积压阈值（条）：task:processing 处理中数达到该值告警。 */
+    private int maixiangQueueProcessingThreshold = 100;
+
     /** 已读通知保留天数（超期自动清理），默认 90 天。 */
     private int readRetentionDays = 90;
+
+    /**
+     * 未读通知保留天数，默认 180 天（比已读长一倍）。
+     *
+     * <p>未读通知此前永不清理：不看铃铛的用户会无限累积。保留期给得更宽，
+     * 是因为未读意味着"用户可能还没看到"，但也不能永远留着。
+     */
+    private int unreadRetentionDays = 180;
 }

backend-java/src/main/java/com/nanri/aiimage/config/PropertiesConfig.java

@@ -4,6 +4,6 @@ import org.springframework.boot.context.properties.EnableConfigurationProperties
 import org.springframework.context.annotation.Configuration;
 
 @Configuration
-@EnableConfigurationProperties({OssProperties.class, TransientStorageProperties.class, StorageProperties.class, BrandProgressProperties.class, DeleteBrandProgressProperties.class, BrandCheckProperties.class, ZiniaoProperties.class, ModuleCleanupProperties.class, TaskPressureProperties.class, TaskImageCacheCleanupProperties.class, AppearancePatentProperties.class, SimilarAsinProperties.class, ImageVideoProperties.class, InstanceRoutingProperties.class, CapacityPlanProperties.class, UserSecretProperties.class, NotificationProperties.class})
+@EnableConfigurationProperties({OssProperties.class, TransientStorageProperties.class, StorageProperties.class, BrandProgressProperties.class, DeleteBrandProgressProperties.class, BrandCheckProperties.class, ZiniaoProperties.class, ModuleCleanupProperties.class, TaskPressureProperties.class, TaskImageCacheCleanupProperties.class, AppearancePatentProperties.class, SimilarAsinProperties.class, ImageVideoProperties.class, InstanceRoutingProperties.class, CapacityPlanProperties.class, UserSecretProperties.class, NotificationProperties.class, DeviceLogOssProperties.class})
 public class PropertiesConfig {
 }

backend-java/src/main/java/com/nanri/aiimage/config/SchedulingConfig.java

@@ -20,8 +20,15 @@ public class SchedulingConfig {
 
     private static final ZoneId BUSINESS_ZONE = ZoneId.of("Asia/Shanghai");
 
+    /**
+     * 调度线程池：承载全站 30+ 个 @Scheduled（含 imagevideo 1s 派发、5s 轮询、结果文件 worker 15s 等高频任务）。
+     *
+     * <p>此前默认 4 线程：任一慢任务（如结果文件组装被内联执行时）都会把兜底类任务
+     * （StaleTaskRepair 心跳判死、陈旧扫描、历史清理）顺延，而兜底任务被顺延会直接放大线上故障面。
+     * 提到 16 并保持可配（aiimage.scheduling.pool-size）。
+     */
     @Bean
-    public TaskScheduler taskScheduler(@Value("${aiimage.scheduling.pool-size:4}") int poolSize) {
+    public TaskScheduler taskScheduler(@Value("${aiimage.scheduling.pool-size:16}") int poolSize) {
         ThreadPoolTaskScheduler scheduler = new ThreadPoolTaskScheduler();
         scheduler.setPoolSize(Math.max(1, poolSize));
         scheduler.setThreadNamePrefix("aiimage-scheduling-");

backend-java/src/main/java/com/nanri/aiimage/config/SimilarAsinProperties.java

@@ -23,6 +23,14 @@ public class SimilarAsinProperties {
     private int staleTimeoutMinutes = 30;
     private String staleFinalizeCron = "0 */2 * * * *";
 
+    /**
+     * 二次判死线（心跳正常但连续 N 分钟无结果分片上报）阈值，分钟；<=0 关闭本线。
+     * 既有判定以 Redis heartbeat stale 为主信号，但心跳随 Python HTTP 心跳每分钟刷新——
+     * 主线程卡死时心跳线程照发，任务永远不 stale（与生产 28131 同型缺口）。
+     * 本线改看 biz_task_scope_state.last_chunk_at（仅分片上传时刷新）。
+     */
+    private int noResultUploadTimeoutMinutes = 180;
+
     /**
      * 末尾零头 batch 的强制 flush 阈值（分钟）：当不足 llmBatchSize 的零头 row
      * 长时间挂着（Python 慢回传）时触发提交。

backend-java/src/main/java/com/nanri/aiimage/config/TaskFileJobConfig.java

@@ -33,6 +33,7 @@ import io.micrometer.core.instrument.MeterRegistry;
 import org.springframework.beans.factory.ObjectProvider;
 import org.springframework.beans.factory.annotation.Qualifier;
 import org.springframework.beans.factory.annotation.Value;
+import com.nanri.aiimage.common.module.TaskModuleRegistry;
 import org.springframework.context.annotation.Bean;
 import org.springframework.context.annotation.Configuration;
 import org.springframework.core.task.TaskExecutor;
@@ -52,11 +53,11 @@ import java.util.concurrent.Semaphore;
 public class TaskFileJobConfig {
 
     /** 结果文件 Job 支持的全部 moduleType（启动校验枚举源，见 ResultFileJobHandlerRegistry.validateCoverage） */
-    public static final Set<String> RESULT_FILE_JOB_MODULE_TYPES = Set.of(
+    /**
-            "SHOP_MATCH", "PRICE_TRACK", "PRODUCT_RISK_RESOLVE",
+     * 结果文件 Job 支持的 moduleType：取自模块注册表（G6），
-            "PUBLISH", "QUERY_ASIN", "SHOP_DATA_CRAWL", "WITHDRAW",
+     * 与 {@code ResultFileJobHandlerRegistry.validateCoverage} 的启动自检配合使用。
-            "PATROL_DELETE", "APPEARANCE_PATENT", "SIMILAR_ASIN",
+     */
-            "DELETE_BRAND", "BRAND", "COLLECT_DATA");
+    public static final Set<String> RESULT_FILE_JOB_MODULE_TYPES = TaskModuleRegistry.resultFileJobModuleTypes();
 
     @Bean("taskFileJobDispatchExecutor")
     public TaskExecutor taskFileJobDispatchExecutor(

backend-java/src/main/java/com/nanri/aiimage/config/TransientStorageProperties.java

@@ -29,7 +29,14 @@ public class TransientStorageProperties {
      */
     private long maxTotalConcurrentOperations = 0;
     private long acquirePermitTimeoutMillis = 2000;
-    private long baseRetryDelayMillis = 500;
+    /**
+     * 首次重试前的基础退避。
+     *
+     * <p>线上高频的重试诱因是 `unexpected end of stream`——那是**立即失败**（连接被 RustFS
+     * 重置后 OkHttp 读响应即报错），不是等超时，所以 500ms 基本是白等：每天上千次累计十几分钟。
+     * 降到 200ms 保留退避语义（真遇到服务端过载仍会退让），又不至于让用户等太久。
+     */
+    private long baseRetryDelayMillis = 200;
     private long maxRetryDelayMillis = 5000;
     private long retryJitterMillis = 250;
     private long failureWindowSeconds = 60;
@@ -37,7 +44,19 @@ public class TransientStorageProperties {
     private long failureCooldownMillis = 10000;
     private int dispatcherMaxRequests = 56;
     private int dispatcherMaxRequestsPerHost = 56;
+    /**
+     * 空闲连接保留数。
+     *
+     * <p>2026-09-17 曾试过设 0（彻底不复用）来验证"unexpected end of stream 是复用死连接导致的"
+     * 这一假设——**实测照旧失败**（新容器起来后第一次请求就中招）。至此已排除公网链路、
+     * keepAlive 过长、连接复用三项；用 mc 并发压 200 个小对象也全部成功，说明服务端没问题。
+     * 剩余方向指向 MinIO Java SDK / OkHttp 与 RustFS 的协议细节，故恢复默认的连接复用。
+     */
     private int connectionPoolMaxIdle = 5;
+    /**
+     * 空闲连接在池里的保留时长。曾由 300000 调到 30000 试图减少 unexpected end of stream，
+     * 实测无改善（该现象与连接复用无关，见 {@link #connectionPoolMaxIdle} 的排查记录），故恢复原值。
+     */
     private long connectionPoolKeepAliveMillis = 300000;
     private long warnPayloadBytes = 5L * 1024 * 1024;
     private long maxPayloadBytes = 50L * 1024 * 1024;
@@ -48,6 +67,12 @@ public class TransientStorageProperties {
      */
     private long maxDecompressedPayloadBytes = 100L * 1024 * 1024;
     private boolean fallbackToLocalOnOversize = true;
+    /**
+     * 上传失败（已重试+熔断）时是否回落到本机磁盘。默认 false：
+     * 多实例/容器化部署下 local 指针只有写入它的那个实例能读，宁可让本次写入失败，
+     * 也不要把跨节点不可读的脏指针交给调用方；仅单机部署才应打开。
+     */
+    private boolean fallbackToLocalOnError = false;
     private boolean deleteRetryEnabled = true;
     private String deleteRetryCron = "0 */5 * * * *";
     private int deleteRetryQueueCapacity = 10000;

backend-java/src/main/java/com/nanri/aiimage/config/UserSecretProperties.java

@@ -22,6 +22,13 @@ public class UserSecretProperties {
     /** 单轮巡检时间预算（分钟），超时中断本轮。 */
     private int checkBudgetMinutes = 20;
 
+    /**
+     * 检测请求使用的 LLM 模型：独立于业务任务模型（业务用 gemini-3.8-flash 等），
+     * 选便宜的可用模型，只验证密钥有效性与链路连通，降低每次检测与巡检的成本。
+     * 用 lite 而非 mini：mini 在中继分组下无可用渠道（503 model_not_found），实测 lite 可路由。
+     */
+    private String checkModel = "doubao-seed-2-0-lite-260215";
+
     /**
      * 检测出口代理提取链接（选配）：配置后检测请求优先经该代理 IP 发出，
      * 代理不可用时自动回退直连；留空则全部直连。

backend-java/src/main/java/com/nanri/aiimage/modules/admin/controller/AdminConsoleController.java

@@ -1,8 +1,8 @@
 package com.nanri.aiimage.modules.admin.controller;
 
 import com.nanri.aiimage.common.api.ApiResponse;
-import com.nanri.aiimage.modules.admin.support.AdminAuthSupport;
+import com.nanri.aiimage.common.security.AdminAuthSupport;
-import com.nanri.aiimage.modules.permission.model.entity.AdminUserEntity;
+import com.nanri.aiimage.common.model.entity.AdminUserEntity;
 import com.nanri.aiimage.modules.permission.model.vo.PermissionMenuItemVo;
 import com.nanri.aiimage.modules.permission.service.PermissionMenuService;
 import com.nanri.aiimage.modules.permission.service.support.AdminMenuTreeBuilder;
@@ -76,8 +76,10 @@ public class AdminConsoleController {
     @Operation(summary = "当前登录管理员的可见后台菜单树")
     public ApiResponse<Map<String, Object>> currentUserMenus(HttpServletRequest request) {
         AdminUserEntity operator = adminAuthSupport.requireAdminOrInternal(request);
+        // 补全祖先分组：部分授权用户（只授权了子页面）也要看到「一级分组 + 子页面」层级，
+        // 与超管的菜单组织顺序一致；分组节点无页面路由，不构成权限扩展。
         List<PermissionMenuItemVo> menus = permissionMenuService.getUserColumnPermissions(
-                operator, operator.getId(), PermissionMenuService.MENU_TYPE_ADMIN);
+                operator, operator.getId(), PermissionMenuService.MENU_TYPE_ADMIN, true);
         menus = AdminMenuTreeBuilder.filterByMenuType(menus, PermissionMenuService.MENU_TYPE_ADMIN);
         List<Map<String, Object>> items = AdminMenuTreeBuilder.toMapList(AdminMenuTreeBuilder.build(menus));
         return ApiResponse.success(Map.of("items", items));

backend-java/src/main/java/com/nanri/aiimage/modules/admin/controller/AdminUserController.java

@@ -5,8 +5,8 @@ import com.nanri.aiimage.modules.admin.model.dto.AdminUserCreateRequest;
 import com.nanri.aiimage.modules.admin.model.dto.AdminUserUpdateRequest;
 import com.nanri.aiimage.modules.admin.model.vo.AdminUserListVo;
 import com.nanri.aiimage.modules.admin.service.AdminUserService;
-import com.nanri.aiimage.modules.admin.support.AdminAuthSupport;
+import com.nanri.aiimage.common.security.AdminAuthSupport;
-import com.nanri.aiimage.modules.permission.model.entity.AdminUserEntity;
+import com.nanri.aiimage.common.model.entity.AdminUserEntity;
 import io.swagger.v3.oas.annotations.Operation;
 import io.swagger.v3.oas.annotations.tags.Tag;
 import jakarta.servlet.http.HttpServletRequest;

backend-java/src/main/java/com/nanri/aiimage/modules/admin/service/AdminUserService.java

@@ -8,14 +8,14 @@ import com.nanri.aiimage.modules.admin.model.dto.AdminUserUpdateRequest;
 import com.nanri.aiimage.modules.admin.model.vo.AdminBriefVo;
 import com.nanri.aiimage.modules.admin.model.vo.AdminUserItemVo;
 import com.nanri.aiimage.modules.admin.model.vo.AdminUserListVo;
-import com.nanri.aiimage.modules.admin.support.AdminAuthSupport;
+import com.nanri.aiimage.common.security.AdminAuthSupport;
 import com.nanri.aiimage.modules.admin.util.PinyinAbbrUtil;
 import com.nanri.aiimage.modules.auth.util.WerkzeugPasswordEncoder;
-import com.nanri.aiimage.modules.permission.mapper.AdminUserMapper;
+import com.nanri.aiimage.common.mapper.AdminUserMapper;
 import com.nanri.aiimage.modules.permission.model.dto.UserColumnPermissionUpdateRequest;
-import com.nanri.aiimage.modules.permission.model.entity.AdminUserEntity;
+import com.nanri.aiimage.common.model.entity.AdminUserEntity;
 import com.nanri.aiimage.modules.permission.service.PermissionMenuService;
-import com.nanri.aiimage.modules.usersecret.service.UserApiSecretService;
+import com.nanri.aiimage.modules.admin.spi.UserSecretCleanupPort;
 import lombok.RequiredArgsConstructor;
 import lombok.extern.slf4j.Slf4j;
 import org.springframework.dao.DuplicateKeyException;
@@ -43,7 +43,7 @@ public class AdminUserService {
     private final WerkzeugPasswordEncoder passwordEncoder;
     private final AdminAuthSupport adminAuthSupport;
     private final PermissionMenuService permissionMenuService;
-    private final UserApiSecretService userApiSecretService;
+    private final UserSecretCleanupPort userSecretCleanupPort;
 
     public AdminUserListVo listUsers(AdminUserEntity currentUser, Integer page, Integer pageSize,
                                      String username, Long createdById, String roleFilter) {
@@ -303,7 +303,7 @@ public class AdminUserService {
         if (affected == 0) {
             throw new BusinessException("用户不存在");
         }
-        int secretRows = userApiSecretService.adminClearByUser(uid);
+        int secretRows = userSecretCleanupPort.adminClearByUser(uid);
         log.info("[admin-user] 用户已删除 uid={} 级联清理密钥行={}", uid, secretRows);
     }
 

backend-java/src/main/java/com/nanri/aiimage/modules/admin/spi/UserSecretCleanupPort.java

@@ -0,0 +1,12 @@
+package com.nanri.aiimage.modules.admin.spi;
+
+/**
+ * 删除用户时级联清理其密钥数据（2026-09 边界收敛：admin → usersecret 的类依赖改为端口）。
+ *
+ * <p>实现方在 usersecret 模块；返回清理的行数用于日志。
+ */
+public interface UserSecretCleanupPort {
+
+    /** 清理该用户的全部密钥相关行，返回受影响行数。 */
+    int adminClearByUser(Long userId);
+}

backend-java/src/main/java/com/nanri/aiimage/modules/appconfig/controller/KdFlowController.java

@@ -0,0 +1,45 @@
+package com.nanri.aiimage.modules.appconfig.controller;
+
+import com.nanri.aiimage.common.api.ApiResponse;
+import com.nanri.aiimage.modules.appconfig.service.KdFlowService;
+import io.swagger.v3.oas.annotations.Operation;
+import io.swagger.v3.oas.annotations.tags.Tag;
+import jakarta.servlet.http.HttpServletRequest;
+import lombok.RequiredArgsConstructor;
+import lombok.extern.slf4j.Slf4j;
+import org.springframework.web.bind.annotation.PostMapping;
+import org.springframework.web.bind.annotation.RequestBody;
+import org.springframework.web.bind.annotation.RestController;
+
+import java.util.Map;
+
+/**
+ * 工作台「开店流程」模块访问密码校验（公开接口，密码本身就是凭据，不额外要求登录态）。
+ *
+ * <p>客户端只在用户点开「开店流程」分组时调用一次；返回体只给 ok 与中文提示，
+ * 不回显服务端配置的密码。
+ */
+@Slf4j
+@RestController
+@RequiredArgsConstructor
+@Tag(name = "开店流程访问校验", description = "工作台「开店流程」模块访问密码的服务端校验")
+public class KdFlowController {
+
+    private final KdFlowService kdFlowService;
+
+    @PostMapping("/api/kd-flow/verify")
+    @Operation(summary = "校验开店流程访问密码",
+            description = "密码存 app_config.kd_flow_password；改密码只需 UPDATE 该行，客户端无需重新发布")
+    public ApiResponse<Map<String, Object>> verify(@RequestBody(required = false) Map<String, String> body,
+                                                   HttpServletRequest request) {
+        String input = body == null ? null : body.get("password");
+        boolean ok = kdFlowService.matches(input);
+        // 只记输入长度与结果，绝不回显密码本身
+        log.info("[开店流程] 校验请求 remoteAddr={} 输入为空={} 结果={}",
+                request.getRemoteAddr(), input == null || input.isBlank(), ok ? "通过" : "拒绝");
+        if (!ok) {
+            return ApiResponse.fail("密码错误");
+        }
+        return ApiResponse.success("验证通过", Map.of("ok", true));
+    }
+}

backend-java/src/main/java/com/nanri/aiimage/modules/appconfig/mapper/AppConfigMapper.java

@@ -0,0 +1,9 @@
+package com.nanri.aiimage.modules.appconfig.mapper;
+
+import com.baomidou.mybatisplus.core.mapper.BaseMapper;
+import com.nanri.aiimage.modules.appconfig.model.entity.AppConfigEntity;
+import org.apache.ibatis.annotations.Mapper;
+
+@Mapper
+public interface AppConfigMapper extends BaseMapper<AppConfigEntity> {
+}

backend-java/src/main/java/com/nanri/aiimage/modules/appconfig/model/entity/AppConfigEntity.java

@@ -0,0 +1,32 @@
+package com.nanri.aiimage.modules.appconfig.model.entity;
+
+import com.baomidou.mybatisplus.annotation.IdType;
+import com.baomidou.mybatisplus.annotation.TableId;
+import com.baomidou.mybatisplus.annotation.TableName;
+import lombok.Data;
+
+import java.time.LocalDateTime;
+
+/**
+ * 通用应用配置（键值）。首个用途：工作台「开店流程」模块访问密码（key = kd_flow_password）。
+ * <p>只放这类低价值、需要"改一行即生效"的口令，不放密钥类敏感配置。
+ */
+@Data
+@TableName("app_config")
+public class AppConfigEntity {
+
+    @TableId(type = IdType.AUTO)
+    private Long id;
+
+    /** 配置键（唯一） */
+    private String configKey;
+
+    /** 配置值 */
+    private String configValue;
+
+    /** 说明 */
+    private String remark;
+
+    /** 更新时间，由数据库 CURRENT_TIMESTAMP 维护 */
+    private LocalDateTime updatedAt;
+}

backend-java/src/main/java/com/nanri/aiimage/modules/appconfig/service/KdFlowService.java

@@ -0,0 +1,48 @@
+package com.nanri.aiimage.modules.appconfig.service;
+
+import com.baomidou.mybatisplus.core.conditions.query.LambdaQueryWrapper;
+import com.nanri.aiimage.modules.appconfig.mapper.AppConfigMapper;
+import com.nanri.aiimage.modules.appconfig.model.entity.AppConfigEntity;
+import lombok.RequiredArgsConstructor;
+import lombok.extern.slf4j.Slf4j;
+import org.springframework.stereotype.Service;
+
+/**
+ * 工作台「开店流程」模块访问密码的服务端校验。
+ *
+ * <p>此前密码写死在客户端源码（KD_FLOW_PASSWORD），改密码必须重新打包装包发给全部用户；
+ * 改由服务端比对后，改密码只需 UPDATE app_config 一行（key = kd_flow_password）。
+ *
+ * <p>不缓存：调用频次极低（用户点一次分组头一次），且改密码后应立即生效。
+ */
+@Slf4j
+@Service
+@RequiredArgsConstructor
+public class KdFlowService {
+
+    /** app_config 中存放开店流程访问密码的键名 */
+    public static final String PASSWORD_KEY = "kd_flow_password";
+
+    private final AppConfigMapper appConfigMapper;
+
+    /** 读取服务端配置的密码；未配置返回 null。 */
+    public String configuredPassword() {
+        AppConfigEntity row = appConfigMapper.selectOne(new LambdaQueryWrapper<AppConfigEntity>()
+                .eq(AppConfigEntity::getConfigKey, PASSWORD_KEY)
+                .last("LIMIT 1"));
+        return row == null ? null : row.getConfigValue();
+    }
+
+    /** 校验用户输入的密码。未配置密码时一律判失败（宁可锁死也不放行）。 */
+    public boolean matches(String input) {
+        String expect = configuredPassword();
+        if (expect == null || expect.isBlank()) {
+            log.warn("[开店流程] app_config 未配置 {}，本次校验一律判失败", PASSWORD_KEY);
+            return false;
+        }
+        String actual = input == null ? "" : input.trim();
+        boolean ok = expect.equals(actual);
+        log.info("[开店流程] 服务端校验 输入长度={} 结果={}", actual.length(), ok ? "通过" : "不通过");
+        return ok;
+    }
+}

backend-java/src/main/java/com/nanri/aiimage/modules/appearancepatent/controller/AppearancePatentController.java

@@ -34,6 +34,7 @@ import java.net.URI;
 import java.nio.charset.StandardCharsets;
 import com.nanri.aiimage.modules.task.model.dto.TaskProgressLightRequest;
 import com.nanri.aiimage.modules.task.model.vo.TaskProgressLightBatchVo;
+import com.nanri.aiimage.modules.task.service.TaskProgressOwnershipSupport;
 
 @RestController
 @RequiredArgsConstructor
@@ -43,6 +44,7 @@ import com.nanri.aiimage.modules.task.model.vo.TaskProgressLightBatchVo;
 public class AppearancePatentController {
 
     private final AppearancePatentTaskService service;
+    private final TaskProgressOwnershipSupport progressOwnershipSupport;
 
     @PostMapping("/parse")
     @Operation(summary = "解析 Excel 并创建任务", description = "解析上传后的 Excel 文件，提取 id、ASIN、国家、URL、标题等字段。返回给前端的数据只包含整数 id 和 n_1 行；n_2、n_3 等子行会保存在 OSS 解析载荷中，用于最终结果补齐。创建后的任务状态为 PENDING，不会自动推送 Python。")
@@ -93,21 +95,26 @@ public class AppearancePatentController {
             @RequestParam("user_id") Long userId,
             @Parameter(description = "历史记录条数，默认 50，最大 100", example = "50")
             @RequestParam(value = "limit", required = false, defaultValue = "50") Integer limit) {
-        return ApiResponse.success(service.history(userId, limit));
+        // 上限钳制（2026-09）：文档写"最大 100"但此前无实际校验
+        int safeLimit = limit == null ? 100 : Math.min(Math.max(1, limit), 100);
+        return ApiResponse.success(service.history(userId, safeLimit));
     }
 
     @PostMapping("/tasks/progress/batch")
     @Operation(summary = "批量查询任务进度", description = "前端只对活跃任务调用该接口，建议 6 秒一次。接口只返回轻量任务状态，不返回明细结果。")
     public ApiResponse<AppearancePatentTaskBatchVo> progress(@Valid @RequestBody AppearancePatentTaskBatchRequest request) {
-        return ApiResponse.success(service.progressBatch(request.getTaskIds()));
+        // 归属过滤：只查询属于该用户的任务（userId 未传时保持原行为，兼容未升级调用方）
+        return ApiResponse.success(service.progressBatch(progressOwnershipSupport
+                .filterOwnedTaskIds(request.getTaskIds(), request.getUserId())));
     }
 
     @PostMapping("/tasks/progress/light")
     @Operation(summary = "批量查询任务轻量进度（前端轮询专用）",
-            description = "与 progress/batch 同输入，只返回轻量字段（taskId/status/statusCode/fileStatus/fileError/fileReady/updatedAt），"
+            description = "与 progress/batch 同输入，只返回轻量字段（taskId/status/statusCode/fileStatus/fileError/fileReady/updatedAt/rowsVersion），"
-                    + "不查明细/result 行，响应体更小；旧 progress/batch 端点保留不动。")
+                    + "不返回行明细内容（仅多一次结果行版本聚合查询），响应体更小；旧 progress/batch 端点保留不动。")
     public ApiResponse<TaskProgressLightBatchVo> progressLight(@Valid @RequestBody TaskProgressLightRequest request) {
-        return ApiResponse.success(service.progressLight(request.getTaskIds()));
+        // 透传 userId：此前传 null = 不过滤归属，匿名遍历 taskId 即可读他人任务状态（2026-09 审查）
+        return ApiResponse.success(service.progressLight(request.getTaskIds(), request.getUserId()));
     }
 
     @PostMapping("/tasks/{taskId}/activate")

backend-java/src/main/java/com/nanri/aiimage/modules/appearancepatent/model/dto/AppearancePatentTaskBatchRequest.java

@@ -12,4 +12,8 @@ public class AppearancePatentTaskBatchRequest {
     @NotEmpty
     @Schema(description = "需要查询进度的任务 ID 列表。前端只传正在轮询的活跃任务；后端会批量查询，避免每个任务单独请求。", example = "[3938,3939]", requiredMode = Schema.RequiredMode.REQUIRED)
     private List<Long> taskIds;
+
+    /** 当前用户 ID（可省略；传入时仅返回该用户的任务，用于进度接口归属过滤）。 */
+    @Schema(description = "当前用户 ID（可省略；传入时仅返回该用户的任务）", example = "1")
+    private Long userId;
 }

backend-java/src/main/java/com/nanri/aiimage/modules/appearancepatent/service/AppearancePatentTaskHeartbeatSpi.java

@@ -0,0 +1,30 @@
+package com.nanri.aiimage.modules.appearancepatent.service;
+
+import com.nanri.aiimage.modules.task.model.dto.TaskHeartbeatRequest;
+import com.nanri.aiimage.modules.task.model.entity.FileTaskEntity;
+import com.nanri.aiimage.modules.task.spi.TaskModuleHeartbeatSpi;
+import lombok.RequiredArgsConstructor;
+import org.springframework.stereotype.Service;
+
+/**
+ * 外观专利检测 的任务心跳实现（2026-09 全维度审查 G5）。
+ *
+ * <p>心跳逻辑（缓存刷新 / 任务缓存回写）从 task 模块收回本模块，task 侧只依赖 SPI 接口，
+ * 消除 task → 业务模块的编译期依赖。
+ */
+@Service
+@RequiredArgsConstructor
+public class AppearancePatentTaskHeartbeatSpi implements TaskModuleHeartbeatSpi {
+
+    private final AppearancePatentTaskCacheService cacheService;
+
+    @Override
+    public String moduleType() {
+        return "APPEARANCE_PATENT";
+    }
+
+    @Override
+    public void touchHeartbeat(Long taskId, TaskHeartbeatRequest request) {
+        cacheService.touchTaskHeartbeat(taskId);
+    }
+}

backend-java/src/main/java/com/nanri/aiimage/modules/appearancepatent/service/AppearancePatentTaskPullSpiImpl.java

@@ -0,0 +1,60 @@
+package com.nanri.aiimage.modules.appearancepatent.service;
+
+import com.nanri.aiimage.modules.appearancepatent.model.dto.AppearancePatentParsedPayloadDto;
+import com.nanri.aiimage.modules.appearancepatent.model.vo.AppearancePatentParsedGroupVo;
+import com.nanri.aiimage.modules.task.model.entity.FileTaskEntity;
+import com.nanri.aiimage.modules.task.spi.ClientTaskPullSpi;
+import lombok.RequiredArgsConstructor;
+import lombok.extern.slf4j.Slf4j;
+import org.springframework.stereotype.Service;
+
+import java.util.LinkedHashMap;
+import java.util.List;
+import java.util.Map;
+
+/**
+ * 外观专利的客户端兜底拉取实现。
+ *
+ * <p>Python 消费端需要 groups（解析分组，页面上是现拉 /queue-payload 再入队），
+ * 这里直接复用同一个 service 方法；payload 与页面保持一致（含 prompt / api_key）。
+ */
+@Slf4j
+@Service
+@RequiredArgsConstructor
+public class AppearancePatentTaskPullSpiImpl implements ClientTaskPullSpi {
+
+    private static final String QUEUE_TYPE = "appearance-patent-run";
+
+    private final AppearancePatentTaskService taskService;
+    private final AppearancePatentTaskCacheService taskCacheService;
+
+    @Override
+    public String moduleType() {
+        return AppearancePatentTaskService.MODULE_TYPE;
+    }
+
+    @Override
+    public Map<String, Object> buildQueuePayload(FileTaskEntity task) {
+        AppearancePatentParsedPayloadDto payload = taskService.queuePayload(task.getId(), task.getUserId());
+        List<AppearancePatentParsedGroupVo> groups = payload.getGroups() == null ? List.of() : payload.getGroups();
+        if (groups.isEmpty()) {
+            // 空 groups 在 Python 侧会被静默跳过（"groups/rows is empty, skip"），宁可在服务端直接判失败
+            log.warn("[appearance-patent] 兜底拉取失败：解析分组为空 taskId={}", task.getId());
+            return null;
+        }
+        Map<String, Object> data = new LinkedHashMap<>();
+        data.put("taskId", task.getId());
+        data.put("user_id", task.getUserId());
+        data.put("prompt", payload.getAiPrompt());
+        data.put("api_key", payload.getApiKey());
+        data.put("groups", groups);
+        log.info("[appearance-patent] 兜底载荷已组装 taskId={} groups={}", task.getId(), groups.size());
+        return Map.of("type", QUEUE_TYPE, "data", data);
+    }
+
+    @Override
+    public void onClaimed(FileTaskEntity task) {
+        // 对齐 activate：刷新模块缓存心跳，让页面立刻看到 RUNNING
+        taskCacheService.touchTaskHeartbeat(task.getId());
+    }
+}

backend-java/src/main/java/com/nanri/aiimage/modules/appearancepatent/service/AppearancePatentTaskService.java

@@ -7,6 +7,7 @@ import com.baomidou.mybatisplus.core.conditions.update.LambdaUpdateWrapper;
 import com.fasterxml.jackson.databind.JsonNode;
 import com.fasterxml.jackson.databind.ObjectMapper;
 import com.nanri.aiimage.common.exception.BusinessException;
+import com.nanri.aiimage.modules.appearancepatent.service.support.AppearancePatentExcelParser;
 import com.nanri.aiimage.common.exception.TaskOwnerMismatchException;
 import com.nanri.aiimage.common.service.DistributedJobLockService;
 import com.nanri.aiimage.common.util.GroupResultPropagator;
@@ -39,6 +40,7 @@ import com.nanri.aiimage.modules.task.mapper.FileResultMapper;
 import com.nanri.aiimage.modules.task.mapper.FileTaskMapper;
 import com.nanri.aiimage.modules.task.mapper.TaskChunkMapper;
 import com.nanri.aiimage.modules.task.mapper.TaskScopeStateMapper;
+import com.nanri.aiimage.modules.task.model.dto.TaskScopeLastChunkDto;
 import com.nanri.aiimage.modules.task.model.entity.FileResultEntity;
 import com.nanri.aiimage.modules.task.model.entity.FileTaskEntity;
 import com.nanri.aiimage.modules.task.model.entity.TaskFileJobEntity;
@@ -67,6 +69,8 @@ import org.springframework.stereotype.Service;
 import org.springframework.transaction.PlatformTransactionManager;
 import org.springframework.transaction.annotation.Transactional;
 import org.springframework.transaction.TransactionDefinition;
+import org.springframework.transaction.support.TransactionSynchronization;
+import org.springframework.transaction.support.TransactionSynchronizationManager;
 import org.springframework.transaction.support.TransactionTemplate;
 
 import java.io.File;
@@ -81,6 +85,7 @@ import java.util.ArrayList;
 import java.util.Comparator;
 import java.util.LinkedHashMap;
 import java.util.LinkedHashSet;
+import java.util.HashMap;
 import java.util.List;
 import java.util.Locale;
 import java.util.Map;
@@ -103,8 +108,8 @@ public class AppearancePatentTaskService {
 
     public static final String MODULE_TYPE = "APPEARANCE_PATENT";
 
-    public TaskProgressLightBatchVo progressLight(List<Long> taskIds) {
+    public TaskProgressLightBatchVo progressLight(List<Long> taskIds, Long userId) {
-        return taskProgressLightAssembler.assemble(MODULE_TYPE, null, taskIds);
+        return taskProgressLightAssembler.assemble(MODULE_TYPE, userId, taskIds);
     }
     private static final String STATUS_PENDING = "PENDING";
     private static final String STATUS_RUNNING = "RUNNING";
@@ -121,6 +126,8 @@ public class AppearancePatentTaskService {
 
     private final LocalFileStorageService localFileStorageService;
     private final OssStorageService ossStorageService;
+    /** 结果下载直链解析（2026-09 从本类抽到 common，消除 11 处逐字重复） */
+    private final com.nanri.aiimage.common.service.ResultDownloadResolver resultDownloadResolver;
     private final StorageProperties storageProperties;
     private final FileTaskMapper fileTaskMapper;
     private final FileResultMapper fileResultMapper;
@@ -277,12 +284,20 @@ public class AppearancePatentTaskService {
             throw new BusinessException("任务不存在");
         }
         ensureTaskOwnedByCurrentInstance(task, "activate");
-        if (STATUS_SUCCESS.equals(task.getStatus()) || STATUS_FAILED.equals(task.getStatus())) {
+        // 只允许 PENDING→RUNNING（条件更新）：与客户端「兜底拉取」的原子认领互斥，
+        // 谁先翻转谁执行，避免页面与客户端重复执行同一任务
+        int updated = fileTaskMapper.update(null, new LambdaUpdateWrapper<FileTaskEntity>()
+                .eq(FileTaskEntity::getId, taskId)
+                .eq(FileTaskEntity::getStatus, STATUS_PENDING)
+                .set(FileTaskEntity::getStatus, STATUS_RUNNING)
+                .set(FileTaskEntity::getUpdatedAt, LocalDateTime.now()));
+        if (updated == 0) {
+            FileTaskEntity latest = fileTaskMapper.selectById(taskId);
+            if (latest != null && STATUS_RUNNING.equals(latest.getStatus())) {
+                throw new BusinessException("任务已在执行中（可能已由客户端自动接管），无需重复启动");
+            }
             throw new BusinessException("任务已结束");
         }
-        task.setStatus(STATUS_RUNNING);
-        task.setUpdatedAt(LocalDateTime.now());
-        fileTaskMapper.updateById(task);
         taskCacheService.touchTaskHeartbeat(taskId);
     }
 
@@ -409,6 +424,53 @@ public class AppearancePatentTaskService {
 
     public void submitResult(Long taskId, AppearancePatentSubmitResultRequest request) {
         submitResultLocked(taskId, request);
+        // 分片补传后自动恢复终态失败的组装 job —— 走与删除品牌同一套口径，
+        // 此前外观专利没有接该入口，分片补齐后只能人工重置 job（线上任务 28459 即如此）。
+        maybeRecoverTerminalFailedAssemble(taskId);
+    }
+
+    /**
+     * 补传恢复：此前因分片缺失导致组装 job 重试耗尽（终态失败），客户端补传缺口后
+     * 把失败的组装 job 重置为 PENDING 重新派发（{@code resetTerminalFailedForRecovery} 自行补发 dispatch 事件）。
+     *
+     * <p>best-effort：恢复失败不得影响补传本身——分片已经落库，恢复只是让后续组装继续推进。
+     * 常态（无终态失败 job）下只查两次即返回，不触发分片扫描。
+     */
+    private void maybeRecoverTerminalFailedAssemble(Long taskId) {
+        if (taskId == null || taskId <= 0) {
+            return;
+        }
+        try {
+            FileResultEntity result = findResultRecord(taskId);
+            if (result == null) {
+                return;
+            }
+            if (taskFileJobService.hasSuccessfulAssembleJob(taskId, MODULE_TYPE, result.getId())) {
+                return;
+            }
+            if (!taskFileJobService.isTerminalFailedAssembleJob(taskId, MODULE_TYPE, result.getId())) {
+                return;
+            }
+            if (!isResultSubmissionComplete(taskId)) {
+                log.info("[appearance-patent] 组装 job 终态失败但分片仍未补传完整，暂不恢复 taskId={} resultId={}",
+                        taskId, result.getId());
+                return;
+            }
+            log.info("[appearance-patent] 分片已补传完整，恢复终态失败的组装 job taskId={} resultId={}",
+                    taskId, result.getId());
+            taskFileJobService.resetTerminalFailedForRecovery(taskId, MODULE_TYPE, result.getId());
+        } catch (Exception ex) {
+            log.warn("[appearance-patent] 补传恢复检查失败（不影响本次补传）taskId={} err={}", taskId, ex.getMessage(), ex);
+        }
+    }
+
+    /** 按 task 取结果行（不创建）；不存在返回 null。 */
+    private FileResultEntity findResultRecord(Long taskId) {
+        List<FileResultEntity> rows = fileResultMapper.selectList(new LambdaQueryWrapper<FileResultEntity>()
+                .eq(FileResultEntity::getTaskId, taskId)
+                .eq(FileResultEntity::getModuleType, MODULE_TYPE)
+                .last("limit 1"));
+        return rows == null || rows.isEmpty() ? null : rows.getFirst();
     }
 
     private void submitResultLocked(Long taskId, AppearancePatentSubmitResultRequest request) {
@@ -509,20 +571,28 @@ public class AppearancePatentTaskService {
         scheduleLlmPipelineForSubmittedChunk(context);
     }
 
+    /**
+     * 删除任务。
+     *
+     * <p>事务边界：远端载荷删除与缓存清理移到事务提交后执行 —— 单次删除超时 120s、重试 3 次，
+     * 放在 DB 事务里会让连接被长时间占用（项目在 TaskScopePayloadStorageService 已写明该约束）。
+     */
     @Transactional
     public void deleteTask(Long taskId, Long userId) {
         FileTaskEntity task = fileTaskMapper.selectById(taskId);
         if (task == null || !MODULE_TYPE.equals(task.getModuleType()) || !Objects.equals(userId, task.getUserId())) {
             throw new BusinessException("任务不存在");
         }
+        List<String> payloads = collectTransientTaskPayloads(taskId);
         fileResultMapper.delete(new LambdaQueryWrapper<FileResultEntity>().eq(FileResultEntity::getTaskId, taskId).eq(FileResultEntity::getModuleType, MODULE_TYPE));
-        deleteTransientTaskPayloads(taskId);
         taskScopeStateMapper.delete(new LambdaQueryWrapper<TaskScopeStateEntity>().eq(TaskScopeStateEntity::getTaskId, taskId).eq(TaskScopeStateEntity::getModuleType, MODULE_TYPE));
         taskChunkMapper.delete(new LambdaQueryWrapper<TaskChunkEntity>().eq(TaskChunkEntity::getTaskId, taskId).eq(TaskChunkEntity::getModuleType, MODULE_TYPE));
-        taskCacheService.deleteTaskCache(taskId);
         // 同步清理 task_file_job，避免被删除任务遗留的 PENDING/FAILED 行被 TaskResultFileJobWorker 反复扫描出 task not found。
         taskFileJobService.deleteTaskJobs(taskId, MODULE_TYPE);
         fileTaskMapper.deleteById(taskId);
+        // 事务提交后再做远端删除与缓存清理
+        deletePayloadsAfterCommit(payloads, taskId);
+        runAfterCommit(() -> taskCacheService.deleteTaskCache(taskId));
     }
 
     public void deleteHistory(Long resultId, Long userId) {
@@ -536,14 +606,8 @@ public class AppearancePatentTaskService {
     }
 
     public String resolveResultDownloadUrl(Long resultId, Long userId) {
-        FileResultEntity row = fileResultMapper.selectById(resultId);
+        // 2026-09 去重：与原实现等价（并修掉 userId.equals 的潜在 NPE），实现收口到 common
-        if (row == null || !MODULE_TYPE.equals(row.getModuleType()) || !Objects.equals(userId, row.getUserId())) {
+        return resultDownloadResolver.resolveUrl(resultId, userId, MODULE_TYPE);
-            throw new BusinessException("记录不存在");
-        }
-        if (row.getResultFileUrl() == null || row.getResultFileUrl().isBlank()) {
-            throw new BusinessException("暂无可下载文件");
-        }
-        return ossStorageService.generateFreshDownloadUrl(row.getResultFileUrl());
     }
 
     public String resolveResultDownloadFilename(Long resultId, Long userId) {
@@ -596,6 +660,7 @@ public class AppearancePatentTaskService {
                 }
             }
         }
+        finalizeNoUploadStaleTasks();
     }
 
     public void debugFinalizeStaleTask(Long taskId) {
@@ -656,6 +721,64 @@ public class AppearancePatentTaskService {
         return updatedMillis <= thresholdMillis;
     }
 
+    /**
+     * 二次判死：Python 心跳正常（Redis heartbeat 新鲜）但连续 N 分钟无结果分片上报。
+     *
+     * <p>既有判定以 Redis heartbeat 为 stale 主信号（P1-7），但该心跳随 Python 的 HTTP 心跳
+     * 每分钟刷新——主线程卡死时心跳线程照发，任务永远不 stale（与生产 28131 同型缺口）。
+     * 本线改看 biz_task_scope_state.last_chunk_at（仅在 persistSubmittedChunk 上传分片时刷新），
+     * 命中后走既有 finalizeStaleTask（封口上传 + LLM 收尾，不粗暴杀任务）。
+     * 从未上报（无 scope 行或全 NULL）的任务跳过，避免误杀首批较慢的正常任务。
+     */
+    private void finalizeNoUploadStaleTasks() {
+        long minutes = properties.getNoResultUploadTimeoutMinutes();
+        if (minutes <= 0) {
+            return;
+        }
+        List<FileTaskEntity> tasks = listStaleFinalizeCandidates();
+        if (tasks.isEmpty()) {
+            return;
+        }
+        LocalDateTime cutoff = LocalDateTime.now().minusMinutes(minutes);
+        long heartbeatThresholdMillis = LocalDateTime.now()
+                .minusMinutes(Math.max(1, properties.getStaleTimeoutMinutes()))
+                .atZone(ZoneId.systemDefault()).toInstant().toEpochMilli();
+        Map<Long, LocalDateTime> lastResultAtByTaskId = new HashMap<>();
+        for (TaskScopeLastChunkDto dto : taskScopeStateMapper.selectLastChunkAtByTaskIds(
+                tasks.stream().map(FileTaskEntity::getId).toList())) {
+            if (dto.taskId() != null && dto.lastChunkAt() != null) {
+                lastResultAtByTaskId.put(dto.taskId(), dto.lastChunkAt());
+            }
+        }
+        for (FileTaskEntity task : tasks) {
+            if (isHeartbeatStale(task, heartbeatThresholdMillis)) {
+                // 心跳已 stale：归既有心跳线处理
+                continue;
+            }
+            LocalDateTime lastResultAt = lastResultAtByTaskId.get(task.getId());
+            if (lastResultAt == null || lastResultAt.isAfter(cutoff)) {
+                continue;
+            }
+            TaskDistributedLockService.LockHandle lockHandle = acquireTaskLock(task.getId(), 0L);
+            if (lockHandle == null) {
+                continue;
+            }
+            try (lockHandle) {
+                log.warn("[appearance-patent] 心跳正常但 {} 分钟无结果分片上报，按卡死收尾 taskId={} lastResultAt={}",
+                        minutes, task.getId(), lastResultAt);
+                String error = "Python heartbeat alive but no result chunk uploaded for " + minutes + " minutes";
+                if (transactionManager != null) {
+                    inNewTransaction(() -> {
+                        finalizeStaleTask(task.getId(), error);
+                        return null;
+                    });
+                } else {
+                    finalizeStaleTask(task.getId(), error);
+                }
+            }
+        }
+    }
+
     /**
      * /result 提交的纯计算阶段（无事务注解）：读任务 + 校验归属 + 分组展开 +
      * 序列化 + payload 存储 + 哈希预计算，全部在事务开始前完成。
@@ -989,6 +1112,14 @@ public class AppearancePatentTaskService {
         long activeAssembleJobs = taskFileJobService.countActiveAssembleJobs(taskId, MODULE_TYPE);
         log.info("[appearance-patent] stale recovery probe taskId={} uploadComplete={} activeAssembleJobs={} persistedRows={}",
                 taskId, uploadComplete, activeAssembleJobs, hasPersistedResultRows(taskId));
+        // 已有「重试耗尽且已终态收尾」的 assemble job：说明恢复已经试过、缺失是永久的。
+        // 再重建只会每 30 秒空转一轮，而且恢复过程刷新任务心跳会让任务永远 RUNNING
+        // （线上任务 28459 实测：48 分钟里每隔 30 秒重建一次 job）。返回 false 交给
+        // finalizeStaleTask 按失败收尾，用户看到明确失败而不是无限等待。
+        if (taskFileJobService.hasExhaustedAssembleJob(taskId, MODULE_TYPE)) {
+            log.warn("[appearance-patent] stale recovery 放弃：已有重试耗尽的 assemble job，按失败收尾 taskId={}", taskId);
+            return false;
+        }
         if (!hasPersistedResultRows(taskId)) {
             log.info("[appearance-patent] stale recovery aborted because no persisted rows taskId={}", taskId);
             return false;
@@ -1139,6 +1270,7 @@ public class AppearancePatentTaskService {
         if (rows == null || rows.isEmpty()) {
             return;
         }
+        String conflictDetail = "未发生冲突";
         for (int attempt = 1; attempt <= CHUNK_PAYLOAD_MERGE_RETRY_LIMIT; attempt++) {
             TaskChunkEntity chunk = taskChunkMapper.selectOne(new LambdaQueryWrapper<TaskChunkEntity>()
                     .eq(TaskChunkEntity::getTaskId, taskId)
@@ -1180,13 +1312,35 @@ public class AppearancePatentTaskService {
                 transientPayloadStorageService.deleteReplacedPayloadIfNeeded(oldPayload, storedPayload);
                 return;
             }
-            transientPayloadStorageService.deletePayloadIfPresent(storedPayload);
+            // CAS 冲突：读回行上的当前哈希，否则只有一句 conflict 无从判断是谁改的（线上 28459 至今未能定位）
+            String currentHash = currentPayloadHash(chunk.getId());
+            conflictDetail = "期望hash=" + oldPayloadHash + " 当前hash=" + currentHash;
             if (attempt < CHUNK_PAYLOAD_MERGE_RETRY_LIMIT) {
-                log.warn("[appearance-patent] chunk payload CAS conflict, retry merge taskId={} scopeHash={} chunk={} attempt={}/{}",
+                log.warn("[appearance-patent] chunk payload CAS conflict, retry merge taskId={} scopeHash={} chunk={} attempt={}/{} 期望hash={} 当前hash={}",
-                        taskId, scopeHash, chunkIndex, attempt, CHUNK_PAYLOAD_MERGE_RETRY_LIMIT);
+                        taskId, scopeHash, chunkIndex, attempt, CHUNK_PAYLOAD_MERGE_RETRY_LIMIT, oldPayloadHash, currentHash);
+                // 还要重试：这次写的对象会被下次重写，先删掉避免堆积
+                transientPayloadStorageService.deletePayloadIfPresent(storedPayload);
+            } else {
+                // 终局失败：保留这次写入的版本化对象，作为读路径「同槽位兄弟对象」兜底的恢复源。
+                // 行没指过去不该让该分片永久判死——删掉它才是线上 28459 丢数据的形态。
+                log.error("[appearance-patent] chunk payload 合并冲突重试耗尽，保留兜底对象 taskId={} scopeHash={} chunk={} 期望hash={} 当前hash={} 保留对象={}",
+                        taskId, scopeHash, chunkIndex, oldPayloadHash, currentHash, storedPayload);
             }
         }
-        throw new IllegalStateException("appearance patent chunk payload update conflict");
+        throw new IllegalStateException("appearance patent chunk payload update conflict " + conflictDetail);
+    }
+
+    /** 读回 chunk 行当前的 payload_hash，用于 CAS 冲突定位（行已不存在/读取失败时返回可读标记）。 */
+    private String currentPayloadHash(Long chunkId) {
+        if (chunkId == null) {
+            return "chunkId 为空";
+        }
+        try {
+            TaskChunkEntity latest = taskChunkMapper.selectById(chunkId);
+            return latest == null ? "行已不存在" : latest.getPayloadHash();
+        } catch (Exception ex) {
+            return "读取失败:" + ex.getMessage();
+        }
     }
 
     private List<AppearancePatentResultRowDto> expandRows(List<AppearancePatentResultRowDto> representatives,
@@ -2296,45 +2450,35 @@ public class AppearancePatentTaskService {
     }
 
 
+    /**
+     * 解析单个源文件（2026-09 审查 C5 改造）。
+     *
+     * <p>改用流式解析器 {@link AppearancePatentExcelParser}（EasyExcel SAX ✓，语义与旧 DOM 实现一致、
+     * 已有单测）取出行级原始值，再在本方法内做原有的业务处理（分组键、状态过滤、字段补齐），
+     * 不再把用户源文件整表读进堆（几十万行曾会占 1~2GB）。
+     */
     private ParsedWorkbook parseWorkbook(File input, AppearancePatentSourceFileDto source) {
-        DataFormatter formatter = new DataFormatter();
+        int maxParseRows = Math.max(1, properties.getMaxParseRows());
-        try (FileInputStream fis = new FileInputStream(input); Workbook workbook = WorkbookFactory.create(fis)) {
+        AppearancePatentExcelParser.ParsedSheet sheet = new AppearancePatentExcelParser().parse(input, maxParseRows);
-            Sheet sheet = workbook.getSheetAt(0);
+        List<String> headers = sheet.headers();
-            Row header = sheet.getRow(0);
-            if (header == null) {
-                throw new BusinessException("Excel 表头为空");
-            }
-            Map<String, Integer> headerMap = buildHeaderMap(header, formatter);
-            List<String> headers = readHeaders(header, formatter);
-            int idCol = findRequiredHeader(headerMap, "id");
-            int asinCol = findRequiredHeader(headerMap, "asin");
-            int countryCol = findRequiredHeader(headerMap, "国家", "country");
-            int priceCol = findOptionalHeaderExact(headerMap, "价格", "price");
-            int skuCol = findOptionalHeaderExact(headerMap,
-                    "sku", "seller sku", "seller_sku", "merchant sku", "merchant_sku", "商品sku", "商品 sku", "库存sku");
-            int urlCol = findOptionalHeaderExact(headerMap,
-                    "url", "rul", "link", "image", "img", "pic", "picture",
-                    "链接", "商品链接", "图片", "商品图片", "主图", "商品主图", "图片链接", "主图链接");
-            int titleCol = findOptionalHeaderExact(headerMap,
-                    "标题", "title", "listing title", "product title", "商品标题", "商品名称", "产品名称");
-
         int statusCol = FailedStatusRowFilter.findStatusColumnIndex(headers);
+        boolean includeBlankStatusRows = statusCol >= 0 && isAppearanceResultWorkbook(headers);
+
+        Map<String, Integer> headerIndex = new LinkedHashMap<>();
+        for (int i = 0; i < headers.size(); i++) {
+            headerIndex.putIfAbsent(headers.get(i), i);
+        }
 
         List<ParsedAppearanceRow> parsedRows = new ArrayList<>();
         int total = 0;
         int dropped = 0;
         int validRows = 0;
-            int maxParseRows = Math.max(1, properties.getMaxParseRows());
         String currentBlockBaseId = "";
         String currentGroupKey = "";
-            for (int i = 1; i <= sheet.getLastRowNum(); i++) {
+        for (AppearancePatentExcelParser.AppearanceExcelRow parsed : sheet.rows()) {
-                Row row = sheet.getRow(i);
+            String id = parsed.id() == null ? "" : parsed.id();
-                if (row == null) {
+            String asin = parsed.asin() == null ? "" : parsed.asin().toUpperCase(Locale.ROOT);
-                    continue;
+            String country = parsed.country() == null ? "" : parsed.country();
-                }
-                String id = cell(row, idCol, formatter);
-                String asin = cell(row, asinCol, formatter).toUpperCase(Locale.ROOT);
-                String country = cell(row, countryCol, formatter);
             if (id.isBlank() && asin.isBlank() && country.isBlank()) {
                 continue;
             }
@@ -2350,7 +2494,7 @@ public class AppearancePatentTaskService {
             AppearancePatentParsedRowVo vo = new AppearancePatentParsedRowVo();
             vo.setSourceFileKey(source.getFileKey());
             vo.setSourceFilename(firstNonBlank(source.getOriginalFilename(), input.getName()));
-                vo.setRowIndex(i + 1);
+            vo.setRowIndex(parsed.rowIndex());
             vo.setSourceId(id);
             vo.setDisplayId(normalizeDisplayId(id));
             String rowBaseId = baseId(vo.getDisplayId());
@@ -2362,18 +2506,20 @@ public class AppearancePatentTaskService {
             vo.setRowToken(buildRowToken(source.getFileKey(), vo.getRowIndex()));
             vo.setAsin(asin);
             vo.setCountry(country);
-                vo.setPrice(priceCol >= 0 ? cell(row, priceCol, formatter) : "");
+            vo.setPrice(parsed.price() == null ? "" : parsed.price());
-                vo.setSku(skuCol >= 0 ? cell(row, skuCol, formatter) : "");
+            vo.setSku(parsed.sku() == null ? "" : parsed.sku());
-                vo.setUrl(urlCol >= 0 ? cell(row, urlCol, formatter) : "");
+            vo.setUrl(parsed.url() == null ? "" : parsed.url());
-                vo.setTitle(titleCol >= 0 ? cell(row, titleCol, formatter) : "");
+            vo.setTitle(parsed.title() == null ? "" : parsed.title());
-                vo.setValues(readRowValues(row, headers, formatter));
+            vo.setValues(parsed.values() == null ? new LinkedHashMap<>() : new LinkedHashMap<>(parsed.values()));
-                parsedRows.add(new ParsedAppearanceRow(vo, statusCol >= 0 ? cell(row, statusCol, formatter) : ""));
+            String statusValue = statusCol >= 0 && parsed.values() != null
+                    ? parsed.values().getOrDefault(statusHeaderName(headers, statusCol), "")
+                    : "";
+            parsedRows.add(new ParsedAppearanceRow(vo, statusValue == null ? "" : statusValue));
         }
         List<AppearancePatentParsedRowVo> allRows = parsedRows.stream()
                 .map(ParsedAppearanceRow::row)
                 .toList();
         hydratePromptFields(allRows);
-            boolean includeBlankStatusRows = statusCol >= 0 && isAppearanceResultWorkbook(headers);
         FailedStatusRowFilter.FilterResult<ParsedAppearanceRow> filteredRows =
                 FailedStatusRowFilter.retainRows(
                         parsedRows,
@@ -2392,13 +2538,12 @@ public class AppearancePatentTaskService {
         if (allRows.isEmpty()) {
             throw new BusinessException("no valid appearance patent rows");
         }
-            return new ParsedWorkbook(total, dropped, headers, allRows);
+        return new ParsedWorkbook(total, dropped, new ArrayList<>(headers), allRows);
-        } catch (BusinessException ex) {
-            throw ex;
-        } catch (Exception ex) {
-            log.warn("[appearance-patent] parse failed file={} err={}", input, ex.getMessage());
-            throw new BusinessException("解析 Excel 失败");
     }
+
+    /** 状态列的原始表头名（流式行只带"表头→值"映射，状态值需按表头名取回）。 */
+    private static String statusHeaderName(List<String> headers, int statusCol) {
+        return statusCol >= 0 && statusCol < headers.size() ? headers.get(statusCol) : "";
     }
 
     private void hydratePromptFields(List<AppearancePatentParsedRowVo> rows) {
@@ -2848,12 +2993,35 @@ public class AppearancePatentTaskService {
         } catch (Exception ex) {
             log.warn("[appearance-patent] read chunk payload failed taskId={} chunk={} err={}",
                     chunk.getTaskId(), chunk.getChunkIndex(), ex.getMessage());
+            if (isPayloadMissing(ex)) {
+                // payload 对象已不在（被清理或从未写入）：重试多少次都读不回来。继续抛会让
+                // ASSEMBLE_RESULT job 的终态回调每轮重跑兜底组装 → 再读同一个缺失对象 → 无限循环
+                // （线上任务 28459 每 10~30 秒重试一次）。跳过该分片，让任务按已有分片出部分结果，
+                // 与品牌/相似ASIN「失败也产出可下载的部分结果」同一口径。
+                log.warn("[appearance-patent] chunk payload 已不存在，跳过该分片（任务按已有分片出结果）"
+                        + " taskId={} chunk={}", chunk.getTaskId(), chunk.getChunkIndex());
+                return rows;
+            }
             throw new BusinessException("appearance patent chunk payload read failed chunk="
                     + chunk.getChunkIndex() + ": " + ex.getMessage(), ex);
         }
         return rows;
     }
 
+    /** payload 对象已不存在（RustFS 返回 NoSuchKey：message 为 "The specified key does not exist."）。
+     *  只有这种"重试也没用"的缺失才允许跳过；网络类失败仍照旧抛出以便重试。 */
+    private static boolean isPayloadMissing(Throwable error) {
+        Throwable cursor = error;
+        while (cursor != null) {
+            String message = cursor.getMessage();
+            if (message != null && message.contains("does not exist")) {
+                return true;
+            }
+            cursor = cursor.getCause();
+        }
+        return false;
+    }
+
     private String rowKey(AppearancePatentParsedRowVo row) {
         if (row == null) {
             return "";
@@ -2935,18 +3103,27 @@ public class AppearancePatentTaskService {
         }
     }
 
-    private void deleteTransientTaskPayloads(Long taskId) {
+    /** 只读收集任务范围/分片载荷指针，供事务提交后做远端删除。 */
+    private List<String> collectTransientTaskPayloads(Long taskId) {
         if (taskId == null || taskId <= 0) {
-            return;
+            return List.of();
         }
+        List<String> payloads = new ArrayList<>();
         List<TaskScopeStateEntity> scopes = taskScopeStateMapper.selectList(new LambdaQueryWrapper<TaskScopeStateEntity>()
                 .select(TaskScopeStateEntity::getParsedPayloadJson, TaskScopeStateEntity::getStateJson)
                 .eq(TaskScopeStateEntity::getTaskId, taskId)
                 .eq(TaskScopeStateEntity::getModuleType, MODULE_TYPE));
         if (scopes != null) {
             for (TaskScopeStateEntity scope : scopes) {
-                transientPayloadStorageService.deletePayloadIfPresent(scope.getParsedPayloadJson());
+                if (scope == null) {
-                transientPayloadStorageService.deletePayloadIfPresent(scope.getStateJson());
+                    continue;
+                }
+                if (scope.getParsedPayloadJson() != null && !scope.getParsedPayloadJson().isBlank()) {
+                    payloads.add(scope.getParsedPayloadJson());
+                }
+                if (scope.getStateJson() != null && !scope.getStateJson().isBlank()) {
+                    payloads.add(scope.getStateJson());
+                }
             }
         }
         List<TaskChunkEntity> chunks = taskChunkMapper.selectList(new LambdaQueryWrapper<TaskChunkEntity>()
@@ -2955,10 +3132,50 @@ public class AppearancePatentTaskService {
                 .eq(TaskChunkEntity::getModuleType, MODULE_TYPE));
         if (chunks != null) {
             for (TaskChunkEntity chunk : chunks) {
-                transientPayloadStorageService.deletePayloadIfPresent(chunk.getPayloadJson());
+                if (chunk != null && chunk.getPayloadJson() != null && !chunk.getPayloadJson().isBlank()) {
+                    payloads.add(chunk.getPayloadJson());
                 }
             }
         }
+        return payloads;
+    }
+
+    /** 事务提交后逐条删远端载荷；无事务时立即执行（与 TaskResultItemService 同款兜底）。 */
+    private void deletePayloadsAfterCommit(List<String> payloads, Long taskId) {
+        if (payloads == null || payloads.isEmpty()) {
+            return;
+        }
+        runAfterCommit(() -> {
+            for (String payload : payloads) {
+                try {
+                    transientPayloadStorageService.deletePayloadIfPresent(payload);
+                } catch (Exception ex) {
+                    log.warn("[appearance-patent] 载荷删除失败 taskId={} msg={}", taskId, ex.getMessage());
+                }
+            }
+        });
+    }
+
+    /**
+     * 删除范围/分片的远端载荷（保留给无事务的清理链路调用）。
+     */
+    private void deleteTransientTaskPayloads(Long taskId) {
+        deletePayloadsAfterCommit(collectTransientTaskPayloads(taskId), taskId);
+    }
+
+    /** 有活动事务则注册 afterCommit，否则立即执行。 */
+    private void runAfterCommit(Runnable action) {
+        if (TransactionSynchronizationManager.isSynchronizationActive()) {
+            TransactionSynchronizationManager.registerSynchronization(new TransactionSynchronization() {
+                @Override
+                public void afterCommit() {
+                    action.run();
+                }
+            });
+            return;
+        }
+        action.run();
+    }
 
     private record SubmitContext(FileTaskEntity task,
                                  String scopeKey,

backend-java/src/main/java/com/nanri/aiimage/modules/appearancepatent/service/support/AppearancePatentExcelParser.java

@@ -1,17 +1,17 @@
 package com.nanri.aiimage.modules.appearancepatent.service.support;
 
 import com.nanri.aiimage.common.exception.BusinessException;
+import com.nanri.aiimage.common.util.ExcelStreamReader;
 import lombok.extern.slf4j.Slf4j;
-import org.apache.poi.ss.usermodel.DataFormatter;
-import org.apache.poi.ss.usermodel.Row;
-import org.apache.poi.ss.usermodel.Sheet;
-import org.apache.poi.ss.usermodel.Workbook;
-import org.apache.poi.ss.usermodel.WorkbookFactory;
 
+import java.io.BufferedInputStream;
 import java.io.File;
 import java.io.FileInputStream;
+import java.io.IOException;
 import java.io.InputStream;
+import java.io.PushbackInputStream;
 import java.util.ArrayList;
+import java.util.Arrays;
 import java.util.LinkedHashMap;
 import java.util.List;
 import java.util.Locale;
@@ -37,9 +37,8 @@ public class AppearancePatentExcelParser {
         if (input == null) {
             throw new IllegalArgumentException("input must not be null");
         }
-        try (FileInputStream fis = new FileInputStream(input);
+        try (InputStream inputStream = new FileInputStream(input)) {
-             Workbook workbook = WorkbookFactory.create(fis)) {
+            return readStreaming(inputStream, maxRows);
-            return parseWorkbook(workbook, maxRows);
         } catch (BusinessException ex) {
             throw ex;
         } catch (Exception ex) {
@@ -52,8 +51,8 @@ public class AppearancePatentExcelParser {
         if (input == null) {
             throw new IllegalArgumentException("input must not be null");
         }
-        try (Workbook workbook = WorkbookFactory.create(input)) {
+        try {
-            return parseWorkbook(workbook, DEFAULT_MAX_ROWS);
+            return readStreaming(input, DEFAULT_MAX_ROWS);
         } catch (BusinessException ex) {
             throw ex;
         } catch (Exception ex) {
@@ -62,83 +61,154 @@ public class AppearancePatentExcelParser {
         }
     }
 
-    private ParsedSheet parseWorkbook(Workbook workbook, int maxRows) {
+    /**
-        int safeMaxRows = Math.max(1, maxRows);
+     * 流式解析（EasyExcel SAX），替代 POI WorkbookFactory 全量 DOM 加载：
-        DataFormatter formatter = new DataFormatter();
+     * 大表不再整表驻留堆内存，行数上限在迭代过程中即时生效（超限抛错）。
-        Sheet sheet = workbook.getSheetAt(0);
+     * 语义与原 POI 路径一致：cell 归一化、表头别名匹配、空行跳过、必填表头缺失抛错、无字段截断。
-        Row header = sheet.getRow(0);
+     * 注意不关闭传入的 InputStream（由调用方负责）。
-        if (header == null) {
+     */
+    private ParsedSheet readStreaming(InputStream inputStream, int maxRows) throws Exception {
+        // EasyExcel 会把非 zip 文本当 CSV 解析成功；原 WorkbookFactory 只认 xlsx/xls，
+        // 这里先做文件魔数校验，保持「垃圾文件→解析 Excel 失败」的语义并拒绝 CSV 误解析。
+        PushbackInputStream pb = new PushbackInputStream(new BufferedInputStream(inputStream, 8192), 8);
+        requireExcelMagic(pb);
+        SheetContext ctx = new SheetContext(Math.max(1, maxRows));
+        ExcelStreamReader.readFirstSheet(pb, new ExcelStreamReader.SheetRowHandler() {
+            @Override
+            public void onHeader(String sheetName, Integer sheetNo, Map<Integer, String> headerMap) {
+                ctx.initHeader(headerMap);
+            }
+
+            @Override
+            public void onRow(String sheetName, Integer sheetNo, int rowIndex,
+                              Map<Integer, String> headerMap, Map<Integer, String> rowMap) {
+                ctx.consumeRow(rowIndex, rowMap);
+            }
+        });
+        if (ctx.headers == null) {
+            // 空表 / 表头行整行为空（EasyExcel 不上报表头回调）：与旧实现 header == null 一致
             throw new BusinessException("Excel 表头为空");
         }
-        Map<String, Integer> headerMap = buildHeaderMap(header, formatter);
+        return new ParsedSheet(ctx.headers, ctx.rows);
-        List<String> headers = readHeaders(header, formatter);
-        int idCol = findRequiredHeader(headerMap, "id");
-        int asinCol = findRequiredHeader(headerMap, "asin");
-        int countryCol = findRequiredHeader(headerMap, "国家", "country");
-        int priceCol = findOptionalHeaderExact(headerMap, "价格", "price");
-        int skuCol = findOptionalHeaderExact(headerMap,
-                "sku", "seller sku", "seller_sku", "merchant sku", "merchant_sku", "商品sku", "商品 sku", "库存sku");
-        int urlCol = findOptionalHeaderExact(headerMap,
-                "url", "rul", "link", "image", "img", "pic", "picture",
-                "链接", "商品链接", "图片", "商品图片", "主图", "商品主图", "图片链接", "主图链接");
-        int titleCol = findOptionalHeaderExact(headerMap,
-                "标题", "title", "listing title", "product title", "商品标题", "商品名称", "产品名称");
-
-        List<AppearanceExcelRow> rows = new ArrayList<>();
-        for (int i = 1; i <= sheet.getLastRowNum(); i++) {
-            Row row = sheet.getRow(i);
-            if (row == null) {
-                continue;
-            }
-            String id = cell(row, idCol, formatter);
-            String asin = cell(row, asinCol, formatter).toUpperCase(Locale.ROOT);
-            String country = cell(row, countryCol, formatter);
-            if (id.isBlank() && asin.isBlank() && country.isBlank()) {
-                continue;
-            }
-            if (rows.size() >= safeMaxRows) {
-                throw new BusinessException("解析行数超过上限: " + safeMaxRows);
-            }
-            rows.add(new AppearanceExcelRow(
-                    i + 1,
-                    id,
-                    asin,
-                    country,
-                    priceCol >= 0 ? cell(row, priceCol, formatter) : "",
-                    skuCol >= 0 ? cell(row, skuCol, formatter) : "",
-                    urlCol >= 0 ? cell(row, urlCol, formatter) : "",
-                    titleCol >= 0 ? cell(row, titleCol, formatter) : "",
-                    readRowValues(row, headers, formatter)));
-        }
-        return new ParsedSheet(headers, rows);
     }
 
-    private Map<String, Integer> buildHeaderMap(Row header, DataFormatter formatter) {
+    /** 校验文件头魔数：xlsx=PK(zip)、xls=OLE2(CFD)。判非抛「解析 Excel 失败」；用 unread 回退已读字节。 */
+    private void requireExcelMagic(PushbackInputStream in) throws IOException {
+        byte[] head = new byte[8];
+        int n = 0;
+        while (n < head.length) {
+            int r = in.read(head, n, head.length - n);
+            if (r < 0) {
+                break;
+            }
+            n += r;
+        }
+        if (n > 0) {
+            in.unread(head, 0, n);
+        }
+        boolean isZip = n >= 4 && head[0] == 'P' && head[1] == 'K' && (head[2] == 3 || head[2] == 5 || head[2] == 7);
+        boolean isOle2 = n >= 8
+                && (head[0] & 0xFF) == 0xD0 && (head[1] & 0xFF) == 0xCF
+                && (head[2] & 0xFF) == 0x11 && (head[3] & 0xFF) == 0xE0;
+        if (!isZip && !isOle2) {
+            log.warn("[appearance-patent] parse rejected non-excel magic head={}", Arrays.copyOf(head, Math.max(n, 0)));
+            throw new BusinessException("解析 Excel 失败");
+        }
+    }
+
+    /** 单表解析上下文：表头就绪后逐行累积结果行。 */
+    private final class SheetContext {
+
+        private final int maxRows;
+        private final List<AppearanceExcelRow> rows = new ArrayList<>();
+        private List<String> headers;
+        private int idCol;
+        private int asinCol;
+        private int countryCol;
+        private int priceCol;
+        private int skuCol;
+        private int urlCol;
+        private int titleCol;
+
+        SheetContext(int maxRows) {
+            this.maxRows = maxRows;
+        }
+
+        void initHeader(Map<Integer, String> rawHeaderMap) {
+            // EasyExcel 回调给出 列号 → 表头文本，与旧 Row 遍历等价（缺列一般为 null/空串）
+            int lastColumnCount = lastColumnCount(rawHeaderMap);
             Map<String, Integer> map = new LinkedHashMap<>();
-        for (int i = 0; i < header.getLastCellNum(); i++) {
+            List<String> headerNames = new ArrayList<>();
-            String val = normalize(formatter.formatCellValue(header.getCell(i)));
+            for (int i = 0; i < lastColumnCount; i++) {
+                String val = normalize(rawHeaderMap.getOrDefault(i, ""));
+                headerNames.add(val.isBlank() ? "列" + (i + 1) : val);
                 if (!val.isBlank()) {
                     map.putIfAbsent(val.toLowerCase(Locale.ROOT), i);
                 }
             }
-        return map;
+            this.headers = headerNames;
+            this.idCol = findRequiredHeader(map, "id");
+            this.asinCol = findRequiredHeader(map, "asin");
+            this.countryCol = findRequiredHeader(map, "国家", "country");
+            this.priceCol = findOptionalHeaderExact(map, "价格", "price");
+            this.skuCol = findOptionalHeaderExact(map,
+                    "sku", "seller sku", "seller_sku", "merchant sku", "merchant_sku", "商品sku", "商品 sku", "库存sku");
+            this.urlCol = findOptionalHeaderExact(map,
+                    "url", "rul", "link", "image", "img", "pic", "picture",
+                    "链接", "商品链接", "图片", "商品图片", "主图", "商品主图", "图片链接", "主图链接");
+            this.titleCol = findOptionalHeaderExact(map,
+                    "标题", "title", "listing title", "product title", "商品标题", "商品名称", "产品名称");
         }
 
-    private List<String> readHeaders(Row header, DataFormatter formatter) {
+        void consumeRow(int rowIndex, Map<Integer, String> rowMap) {
-        List<String> headers = new ArrayList<>();
+            if (headers == null) {
-        for (int i = 0; i < header.getLastCellNum(); i++) {
+                return;
-            String val = normalize(formatter.formatCellValue(header.getCell(i)));
-            headers.add(val.isBlank() ? "列" + (i + 1) : val);
             }
-        return headers;
+            // EasyExcel rowIndex 从 0 起（0 为表头），POI 原实现行号同样 0 起并 +1 展示
+            String id = streamCell(rowMap, idCol);
+            String asin = streamCell(rowMap, asinCol).toUpperCase(Locale.ROOT);
+            String country = streamCell(rowMap, countryCol);
+            if (id.isBlank() && asin.isBlank() && country.isBlank()) {
+                return;
+            }
+            if (rows.size() >= maxRows) {
+                throw new BusinessException("解析行数超过上限: " + maxRows);
             }
-
-    private Map<String, String> readRowValues(Row row, List<String> headers, DataFormatter formatter) {
             Map<String, String> values = new LinkedHashMap<>();
             for (int i = 0; i < headers.size(); i++) {
-            values.put(headers.get(i), cell(row, i, formatter));
+                values.put(headers.get(i), streamCell(rowMap, i));
+            }
+            rows.add(new AppearanceExcelRow(
+                    rowIndex + 1,
+                    id,
+                    asin,
+                    country,
+                    priceCol >= 0 ? streamCell(rowMap, priceCol) : "",
+                    skuCol >= 0 ? streamCell(rowMap, skuCol) : "",
+                    urlCol >= 0 ? streamCell(rowMap, urlCol) : "",
+                    titleCol >= 0 ? streamCell(rowMap, titleCol) : "",
+                    values));
+        }
+
+        private int lastColumnCount(Map<Integer, String> rowMap) {
+            if (rowMap == null || rowMap.isEmpty()) {
+                return 0;
+            }
+            int lastColumnCount = 0;
+            for (Integer columnIndex : rowMap.keySet()) {
+                if (columnIndex != null && columnIndex >= lastColumnCount) {
+                    lastColumnCount = columnIndex + 1;
+                }
+            }
+            return lastColumnCount;
+        }
+
+        private String streamCell(Map<Integer, String> rowMap, int col) {
+            if (col < 0 || rowMap == null) {
+                return "";
+            }
+            return normalize(rowMap.get(col));
         }
-        return values;
     }
 
     private int findRequiredHeader(Map<String, Integer> map, String... names) {
@@ -177,10 +247,6 @@ public class AppearancePatentExcelParser {
         return normalized.replaceAll("[\\s_\\-()（）\\[\\]{}:：/\\\\]+", "");
     }
 
-    private String cell(Row row, int col, DataFormatter formatter) {
-        return col < 0 ? "" : normalize(formatter.formatCellValue(row.getCell(col)));
-    }
-
     private String normalize(String val) {
         return val == null ? "" : val.replace(String.valueOf((char) 0xFEFF), "").replace((char) 0x3000, ' ').trim().replaceAll("\\s+", " ");
     }

backend-java/src/main/java/com/nanri/aiimage/modules/auth/service/AuthService.java

@@ -3,12 +3,12 @@ package com.nanri.aiimage.modules.auth.service;
 import com.baomidou.mybatisplus.core.conditions.query.LambdaQueryWrapper;
 import com.baomidou.mybatisplus.core.conditions.update.LambdaUpdateWrapper;
 import com.nanri.aiimage.common.exception.BusinessException;
-import com.nanri.aiimage.modules.auth.config.AuthProperties;
+import com.nanri.aiimage.common.security.AuthProperties;
 import com.nanri.aiimage.modules.auth.mapper.LoginUserMapper;
 import com.nanri.aiimage.modules.auth.model.dto.LoginRequest;
 import com.nanri.aiimage.modules.auth.model.entity.LoginUserEntity;
 import com.nanri.aiimage.modules.auth.model.vo.LoginResultVo;
-import com.nanri.aiimage.modules.auth.support.DeviceSessionPolicy;
+import com.nanri.aiimage.common.security.DeviceSessionPolicy;
 import com.nanri.aiimage.modules.auth.util.WerkzeugPasswordEncoder;
 import com.nanri.aiimage.modules.permission.service.PermissionMenuService;
 import io.jsonwebtoken.Claims;
@@ -18,6 +18,10 @@ import org.springframework.http.ResponseCookie;
 import org.springframework.stereotype.Service;
 
 import java.time.Duration;
+import java.time.Instant;
+import java.util.Map;
+import java.util.concurrent.ConcurrentHashMap;
+import com.nanri.aiimage.common.security.JwtService;
 
 @Service
 @RequiredArgsConstructor
@@ -30,6 +34,50 @@ public class AuthService {
     private final PermissionMenuService permissionMenuService;
     private final AuthProperties authProperties;
 
+    /**
+     * 登录失败计数与锁定（2026-09 全维度审查补）：生产是公网域名，此前无任何失败限制
+     * 即可无限撞库。内存实现（双节点各自计数，防护效果减半但不引入新依赖），
+     * 达到阈值后锁定 15 分钟。
+     */
+    private static final int LOGIN_FAIL_LIMIT = 10;
+    private static final Duration LOGIN_LOCK_DURATION = Duration.ofMinutes(15);
+    private final Map<String, FailRecord> loginFailures = new ConcurrentHashMap<>();
+
+    /** 失败计数；lockedUntil 非空表示已锁定。 */
+    private record FailRecord(int count, Instant lockedUntil) {
+    }
+
+    private void assertLoginNotLocked(String username) {
+        FailRecord record = loginFailures.get(username);
+        if (record != null && record.lockedUntil() != null && record.lockedUntil().isAfter(Instant.now())) {
+            long minutes = Duration.between(Instant.now(), record.lockedUntil()).toMinutes() + 1;
+            log.warn("[auth] 登录已锁定 username={} 剩余约 {} 分钟", username, minutes);
+            throw new BusinessException("登录失败次数过多，请 " + minutes + " 分钟后再试");
+        }
+    }
+
+    private void recordLoginFailure(String username) {
+        // 防无界增长：积累较多时清掉未锁定的过期记录（登录接口调用频次低）
+        if (loginFailures.size() > 1000) {
+            Instant now = Instant.now();
+            loginFailures.entrySet().removeIf(e -> e.getValue().lockedUntil() == null
+                    || e.getValue().lockedUntil().isBefore(now));
+        }
+        loginFailures.compute(username, (key, old) -> {
+            int count = (old == null ? 0 : old.count()) + 1;
+            Instant lockedUntil = count >= LOGIN_FAIL_LIMIT ? Instant.now().plus(LOGIN_LOCK_DURATION) : null;
+            if (lockedUntil != null) {
+                log.warn("[auth] 登录失败达阈值，锁定 username={} count={} minutes={}",
+                        username, count, LOGIN_LOCK_DURATION.toMinutes());
+            }
+            return new FailRecord(count, lockedUntil);
+        });
+    }
+
+    private void clearLoginFailures(String username) {
+        loginFailures.remove(username);
+    }
+
     public LoginResultVo login(LoginRequest request) {
         String username = trim(request.getUsername());
         String password = request.getPassword() == null ? "" : request.getPassword();
@@ -41,12 +89,16 @@ public class AuthService {
             throw new BusinessException("缺少设备ID，请在桌面端打开");
         }
 
+        assertLoginNotLocked(username);
+
         LoginUserEntity user = loginUserMapper.selectOne(new LambdaQueryWrapper<LoginUserEntity>()
                 .eq(LoginUserEntity::getUsername, username)
                 .last("LIMIT 1"));
         if (user == null || !passwordEncoder.matches(password, user.getPasswordHash())) {
+            recordLoginFailure(username);
             throw new BusinessException("用户名或密码错误");
         }
+        clearLoginFailures(username);
 
         boolean isAdmin = user.getIsAdmin() != null && user.getIsAdmin() == 1;
         // 单设备登录：登录成功即把账号绑定到当前设备（last-login-wins），原设备下一次请求被顶下线

backend-java/src/main/java/com/nanri/aiimage/modules/brand/client/BrandCheckClient.java

@@ -95,10 +95,14 @@ public class BrandCheckClient {
 
     public BrandCheckBatchResult checkAll(List<String> brands, String strategy) {
         List<String> distinctBrands = distinctNonBlank(brands);
+        // 整批共用一个耗时预算：上游 16890 卡死时，单品牌 10 次重试曾把一次分片回传拖到
+        // 103.5 秒（taskId 28599），客户端重试预算耗尽后中止了整个采集。预算用尽即停止重试。
+        long budgetMillis = properties.getTotalTimeoutMillis();
+        long deadlineNanos = budgetMillis > 0L ? System.nanoTime() + budgetMillis * 1_000_000L : Long.MAX_VALUE;
         List<CompletableFuture<BrandCheckOutcome>> futures = new ArrayList<>(distinctBrands.size());
         for (String brand : distinctBrands) {
             futures.add(CompletableFuture.supplyAsync(
-                    () -> checkOneBrand(brand, strategy), checkExecutor));
+                    () -> checkOneBrand(brand, strategy, deadlineNanos), checkExecutor));
         }
         List<Object> failedData = new ArrayList<>();
         List<Object> queryFailedData = new ArrayList<>();
@@ -110,20 +114,29 @@ public class BrandCheckClient {
         return new BrandCheckBatchResult(distinctBrands, failedData, queryFailedData);
     }
 
-    private BrandCheckOutcome checkOneBrand(String brand, String strategy) {
+    private BrandCheckOutcome checkOneBrand(String brand, String strategy, long deadlineNanos) {
         int attempts = Math.max(1, properties.getRetryTimes());
         BrandCheckResponse response = null;
         Exception lastFailure = null;
         for (int attempt = 1; attempt <= attempts; attempt++) {
+            // 预算用尽就不再重试，按查询失败收尾。只掐「重试」不打断已发出的请求，
+            // 故最坏耗时 ≈ 预算 + 一次请求的读超时；首轮始终执行，避免上游只是慢一点时被误降级。
+            if (attempt > 1 && System.nanoTime() >= deadlineNanos) {
+                log.warn("[brand-check] 整批耗时预算用尽，停止重试 brand={} attempt={}/{} lastErr={}",
+                        brand, attempt, attempts,
+                        lastFailure == null ? "query_faild_data 持续非空" : lastFailure.getMessage());
+                break;
+            }
             try {
                 response = check(brand, strategy);
             } catch (Exception ex) {
                 lastFailure = ex;
                 response = null;
                 if (attempt < attempts) {
-                    log.warn("[brand-check] 请求异常将重试 brand={} attempt={}/{} err={}",
+                    long retryDelayMillis = retryDelayMillis(attempt);
-                            brand, attempt, attempts, ex.getMessage());
+                    log.warn("[brand-check] 请求异常将重试 brand={} attempt={}/{} 等待={}ms err={}",
-                    sleepBeforeRetry();
+                            brand, attempt, attempts, retryDelayMillis, ex.getMessage());
+                    sleepBeforeRetry(retryDelayMillis);
                 }
                 continue;
             }
@@ -132,9 +145,10 @@ public class BrandCheckClient {
                 return new BrandCheckOutcome(nullToEmpty(response.getFaildData()), List.of());
             }
             if (attempt < attempts) {
-                log.warn("[brand-check] 服务端返回查询失败将重试 brand={} attempt={}/{}",
+                long retryDelayMillis = retryDelayMillis(attempt);
-                        brand, attempt, attempts);
+                log.warn("[brand-check] 服务端返回查询失败将重试 brand={} attempt={}/{} 等待={}ms",
-                sleepBeforeRetry();
+                        brand, attempt, attempts, retryDelayMillis);
+                sleepBeforeRetry(retryDelayMillis);
             }
         }
         log.warn("[brand-check] 重试耗尽判定查询失败 brand={} attempts={} lastErr={}",
@@ -144,8 +158,18 @@ public class BrandCheckClient {
                 response == null ? List.of(brand) : nullToEmpty(response.getQueryFaildData()));
     }
 
-    private void sleepBeforeRetry() {
+    /**
-        long delayMillis = Math.max(0L, properties.getRetryIntervalMillis());
+     * 第 attempt 次重试前的等待毫秒数：按基准间隔随轮次递增后封顶。
+     * 限流窗口通常只有几秒，固定 1s 间隔反复打过去救不回来；递增等待能覆盖窗口，
+     * 封顶则保证单个品牌不会长时间占住查询线程（并发度只有 3）。
+     */
+    private long retryDelayMillis(int attempt) {
+        long base = Math.max(0L, properties.getRetryIntervalMillis());
+        long cap = Math.max(base, properties.getRetryMaxIntervalMillis());
+        return Math.min(base * Math.max(1, attempt), cap);
+    }
+
+    private void sleepBeforeRetry(long delayMillis) {
         if (delayMillis <= 0L) {
             return;
         }

backend-java/src/main/java/com/nanri/aiimage/modules/brand/controller/InternalBrandTaskController.java

@@ -0,0 +1,56 @@
+package com.nanri.aiimage.modules.brand.controller;
+
+import com.nanri.aiimage.common.api.ApiResponse;
+import com.nanri.aiimage.common.exception.BusinessException;
+import com.nanri.aiimage.common.security.AdminAuthSupport;
+import com.nanri.aiimage.modules.brand.model.dto.BrandTaskAbortRequest;
+import com.nanri.aiimage.modules.brand.service.BrandTaskService;
+import io.swagger.v3.oas.annotations.Operation;
+import io.swagger.v3.oas.annotations.tags.Tag;
+import jakarta.servlet.http.HttpServletRequest;
+import lombok.RequiredArgsConstructor;
+import lombok.extern.slf4j.Slf4j;
+import org.springframework.web.bind.annotation.PathVariable;
+import org.springframework.web.bind.annotation.PostMapping;
+import org.springframework.web.bind.annotation.RequestBody;
+import org.springframework.web.bind.annotation.RequestMapping;
+import org.springframework.web.bind.annotation.RestController;
+
+import java.util.Map;
+
+/**
+ * 内部接口：品牌任务中止上报，供主机 A 品牌检测服务（15126）在爬取不可继续时调用
+ * （如 WIPO 连续限流熔断）。Java 侧用已收到的结果分片部分组装结果文件并落 failed +
+ * 真实原因，避免任务悬挂到心跳超时被判「前端长时间无响应」、已跑出的数据无法下载。
+ *
+ * <p>鉴权：仅凭 X-Internal-Token（与容器 AIIMAGE_INTERNAL_TOKEN / 宿主机
+ * ~/.aiimage/internal-token 同值）。/api/internal 前缀虽在 AdminApiGuardFilter 兜底
+ * 名单内、可信令牌会放行，controller 仍须自校验——防止配置漂移时匿名可达。
+ */
+@RestController
+@RequiredArgsConstructor
+@Slf4j
+@RequestMapping("/api/internal/brand")
+@Tag(name = "内部接口", description = "服务间调用（X-Internal-Token 鉴权）。")
+public class InternalBrandTaskController {
+
+    private final BrandTaskService brandTaskService;
+    private final AdminAuthSupport adminAuthSupport;
+
+    @PostMapping("/tasks/{taskId}/abort")
+    @Operation(summary = "上报品牌任务中止（爬取方调用）",
+            description = "用已收到的结果分片部分组装结果文件（未检测品牌单独成 sheet）并落 failed + 真实原因；幂等，终态任务直接返回。")
+    public ApiResponse<Map<String, Object>> abortTask(HttpServletRequest request,
+                                                      @PathVariable Long taskId,
+                                                      @RequestBody(required = false) BrandTaskAbortRequest body) {
+        if (!adminAuthSupport.isTrustedInternalToken(request)) {
+            log.warn("[internal-brand-abort] 拒绝未携带可信内部令牌的请求 taskId={} remoteAddr={}",
+                    taskId, request.getRemoteAddr());
+            throw new BusinessException(401, "未授权");
+        }
+        String errorMessage = body == null ? null : body.getErrorMessage();
+        log.info("[internal-brand-abort] 收到中止上报 taskId={} remoteAddr={} msg={}",
+                taskId, request.getRemoteAddr(), errorMessage);
+        return ApiResponse.success(brandTaskService.abortTask(taskId, errorMessage));
+    }
+}

backend-java/src/main/java/com/nanri/aiimage/modules/brand/mapper/BrandCrawlTaskMapper.java

@@ -3,7 +3,31 @@ package com.nanri.aiimage.modules.brand.mapper;
 import com.baomidou.mybatisplus.core.mapper.BaseMapper;
 import com.nanri.aiimage.modules.brand.model.entity.BrandCrawlTaskEntity;
 import org.apache.ibatis.annotations.Mapper;
+import org.apache.ibatis.annotations.Param;
+import org.apache.ibatis.annotations.Select;
+
+import java.time.LocalDateTime;
+import java.util.List;
 
 @Mapper
 public interface BrandCrawlTaskMapper extends BaseMapper<BrandCrawlTaskEntity> {
+
+    /**
+     * 查询超过保留期的终态品牌检测任务 id（保留期清理用，只取 id 不拉整行——历史行的
+     * file_paths/result_paths JSON 字段可能很大）。
+     *
+     * <p>终态集合与 {@code BrandTaskService} 的状态机一致（success/failed/cancelled），
+     * pending/running 绝不返回（删了正在跑的任务，结果回传会找不到任务行）。
+     * 时间线用 updated_at，与 BrandTaskStaleRepairSpiImpl 的陈旧判定同款口径，
+     * 可命中 V120 的 idx_brand_crawl_task_status_updated(status, updated_at) 索引。
+     */
+    @Select("""
+            SELECT id FROM brand_crawl_tasks
+            WHERE status IN ('success', 'failed', 'cancelled')
+              AND updated_at < #{cutoff}
+            ORDER BY id ASC
+            LIMIT #{batchSize}
+            """)
+    List<Long> selectExpiredTerminalTaskIds(@Param("cutoff") LocalDateTime cutoff,
+                                            @Param("batchSize") int batchSize);
 }

backend-java/src/main/java/com/nanri/aiimage/modules/brand/model/dto/BrandFileAggregateCacheDto.java

@@ -18,4 +18,7 @@ public class BrandFileAggregateCacheDto {
     private Boolean completed = false;
     private List<BrandInvalidBrandDto> invalidBrands = new ArrayList<>();
     private List<String> queryFailedBrands = new ArrayList<>();
+    /** 已判定保留的品牌：与 invalidBrands / queryFailedBrands 一起构成「已检测品牌」，
+     *  失败任务的未检测品牌 = 源文件品牌 - 三者并集（部分组装时写「未检测品牌」sheet）。 */
+    private List<String> keptBrands = new ArrayList<>();
 }

backend-java/src/main/java/com/nanri/aiimage/modules/brand/model/dto/BrandTaskAbortRequest.java

@@ -0,0 +1,13 @@
+package com.nanri.aiimage.modules.brand.model.dto;
+
+import io.swagger.v3.oas.annotations.media.Schema;
+import lombok.Data;
+
+@Data
+@Schema(description = "品牌任务中止上报请求（内部接口，由爬取方 15126 调用）。")
+public class BrandTaskAbortRequest {
+
+    @Schema(description = "中止原因，会原样写入任务 error_message，前端任务列表展示该文案。",
+            example = "连续 8 次请求被 WIPO 限流（返回 Forbidden），已中止任务；请检查代理配置或错峰重跑")
+    private String errorMessage;
+}

backend-java/src/main/java/com/nanri/aiimage/modules/brand/service/BrandCrawlTaskHeartbeatSpi.java

@@ -0,0 +1,97 @@
+package com.nanri.aiimage.modules.brand.service;
+
+import com.baomidou.mybatisplus.core.conditions.query.LambdaQueryWrapper;
+import com.baomidou.mybatisplus.core.conditions.update.LambdaUpdateWrapper;
+import com.nanri.aiimage.modules.brand.mapper.BrandCrawlTaskMapper;
+import com.nanri.aiimage.modules.brand.model.entity.BrandCrawlTaskEntity;
+import com.nanri.aiimage.modules.task.model.dto.TaskHeartbeatRequest;
+import com.nanri.aiimage.modules.task.model.vo.TaskHeartbeatVo;
+import com.nanri.aiimage.modules.task.spi.BrandTaskHeartbeatSpi;
+import lombok.RequiredArgsConstructor;
+import lombok.extern.slf4j.Slf4j;
+import org.springframework.stereotype.Service;
+
+import java.time.LocalDateTime;
+
+/**
+ * 品牌检测任务（brand_crawl_tasks）的心跳/中断实现（2026-09 全维度审查 G5）。
+ *
+ * <p>品牌任务有独立的表与状态字面量（running/pending/cancelled），原先这段逻辑散在
+ * {@code TaskHeartbeatService} 里并直接依赖本模块的 Mapper 与进度缓存；现整体收回本模块。
+ */
+@Service
+@RequiredArgsConstructor
+@Slf4j
+public class BrandCrawlTaskHeartbeatSpi implements BrandTaskHeartbeatSpi {
+
+    private static final String MODULE_BRAND = "BRAND";
+    private static final String BRAND_STATUS_RUNNING = "running";
+
+    private final BrandCrawlTaskMapper brandCrawlTaskMapper;
+    private final BrandTaskProgressCacheService brandTaskProgressCacheService;
+
+    @Override
+    public TaskHeartbeatVo heartbeat(Long taskId, TaskHeartbeatRequest request) {
+        BrandCrawlTaskEntity task = selectTask(taskId);
+        if (task == null) {
+            return null;
+        }
+        String status = task.getStatus();
+        if (!BRAND_STATUS_RUNNING.equals(status)) {
+            log.warn("[task-heartbeat] brand task is not running taskId={} actualUserId={} status={}",
+                    task.getId(), task.getUserId(), status);
+            return TaskHeartbeatVo.notAlive(MODULE_BRAND, status, "task is not running");
+        }
+        int updated = brandCrawlTaskMapper.update(null, new LambdaUpdateWrapper<BrandCrawlTaskEntity>()
+                .eq(BrandCrawlTaskEntity::getId, task.getId())
+                .eq(BrandCrawlTaskEntity::getStatus, BRAND_STATUS_RUNNING)
+                .set(BrandCrawlTaskEntity::getUpdatedAt, LocalDateTime.now()));
+        if (updated <= 0) {
+            BrandCrawlTaskEntity latest = brandCrawlTaskMapper.selectById(task.getId());
+            log.warn("[task-heartbeat] brand task heartbeat update missed taskId={} actualUserId={} status={} latestStatus={}",
+                    task.getId(), task.getUserId(), status,
+                    latest == null ? null : latest.getStatus());
+            return TaskHeartbeatVo.notAlive(MODULE_BRAND, latest == null ? status : latest.getStatus(),
+                    "task is not running");
+        }
+        brandTaskProgressCacheService.touchHeartbeat(
+                task.getId(),
+                request == null ? null : request.getPhase(),
+                request == null ? null : request.getCurrent(),
+                request == null ? null : request.getTotal());
+        return TaskHeartbeatVo.alive(MODULE_BRAND, BRAND_STATUS_RUNNING);
+    }
+
+    @Override
+    public TaskHeartbeatVo markInterrupted(Long taskId, String reason) {
+        BrandCrawlTaskEntity task = selectTask(taskId);
+        if (task == null) {
+            return null;
+        }
+        String status = task.getStatus();
+        if ("running".equalsIgnoreCase(status) || "pending".equalsIgnoreCase(status)) {
+            // 与 file 分支统一改为条件更新：整行 updateById 会拿读取快照覆盖并发写入的字段
+            // （客户端重启上报与品牌任务自身状态流转同时发生时）
+            int updated = brandCrawlTaskMapper.update(null, new LambdaUpdateWrapper<BrandCrawlTaskEntity>()
+                    .eq(BrandCrawlTaskEntity::getId, task.getId())
+                    .in(BrandCrawlTaskEntity::getStatus, "running", "pending", "RUNNING", "PENDING")
+                    .set(BrandCrawlTaskEntity::getStatus, "cancelled")
+                    .set(BrandCrawlTaskEntity::getErrorMessage, reason)
+                    .set(BrandCrawlTaskEntity::getUpdatedAt, LocalDateTime.now()));
+            if (updated > 0) {
+                log.warn("[task-interrupted] brand task marked cancelled by client restart taskId={} reason={}",
+                        taskId, reason);
+                return TaskHeartbeatVo.notAlive(MODULE_BRAND, "cancelled", "marked cancelled");
+            }
+        }
+        log.info("[task-interrupted] brand task not in running/pending, skipped taskId={} status={}", taskId, status);
+        return TaskHeartbeatVo.notAlive(MODULE_BRAND, status, "task is not running");
+    }
+
+    private BrandCrawlTaskEntity selectTask(Long taskId) {
+        LambdaQueryWrapper<BrandCrawlTaskEntity> brandQuery = new LambdaQueryWrapper<BrandCrawlTaskEntity>()
+                .eq(BrandCrawlTaskEntity::getId, taskId)
+                .last("limit 1");
+        return brandCrawlTaskMapper.selectOne(brandQuery);
+    }
+}

backend-java/src/main/java/com/nanri/aiimage/modules/brand/service/BrandResultFileJobHandler.java

@@ -7,8 +7,8 @@ import com.nanri.aiimage.modules.task.model.entity.TaskFileJobEntity;
 
 /**
  * BRAND 结果文件 Job Handler（04 注册表）。
- * 注意：resultFileUrl 解析特例（resolveResultObjectKey，无 resultId 也走）
+ * resultFileUrl 解析特例（resolveResultObjectKey，无 resultId 也走）由本 Handler 的
- * 保留在 Worker 公共路径 resolveResultFileUrl，Handler 不接管 URL 解析；
+ * resolveResultFileUrl 钩子承担（2026-09：原实现在 Worker 内直接 import 本模块，构成 task → 业务依赖）；
  * cleanup 为空（原 cleanupAfterSuccess 无 BRAND 分支）。
  */
 public class BrandResultFileJobHandler implements ResultFileJobHandler {
@@ -32,4 +32,12 @@ public class BrandResultFileJobHandler implements ResultFileJobHandler {
         brandTaskService.processResultFileJob(job);
         return true;
     }
+
+    @Override
+    public String resolveResultFileUrl(TaskFileJobEntity job) {
+        if (job == null || job.getTaskId() == null) {
+            return null;
+        }
+        return brandTaskService.resolveResultObjectKey(job.getTaskId());
+    }
 }

backend-java/src/main/java/com/nanri/aiimage/modules/brand/service/BrandTaskProgressCacheService.java

@@ -4,11 +4,13 @@ import com.fasterxml.jackson.databind.ObjectMapper;
 import com.nanri.aiimage.config.BrandProgressProperties;
 import lombok.extern.slf4j.Slf4j;
 import org.springframework.data.redis.core.StringRedisTemplate;
+import org.springframework.data.redis.core.script.DefaultRedisScript;
 import org.springframework.stereotype.Service;
 
 import java.time.Duration;
 import java.time.Instant;
 import java.util.LinkedHashMap;
+import java.util.List;
 import java.util.Map;
 
 @Service
@@ -21,6 +23,9 @@ public class BrandTaskProgressCacheService {
     public static final String PHASE_FAILED = "failed";
     private static final Duration FINALIZE_LOCK_TTL = Duration.ofMinutes(10);
 
+    /** 本实例（进程）的锁持有者标识：释放锁时用它校验"锁还是我的"。 */
+    private final String lockOwnerToken = java.util.UUID.randomUUID().toString();
+
     private final StringRedisTemplate stringRedisTemplate;
     private final BrandProgressProperties brandProgressProperties;
     @SuppressWarnings("unused")
@@ -55,6 +60,8 @@ public class BrandTaskProgressCacheService {
         values.put("finished_files", String.valueOf(Math.max(finishedFiles, 0)));
         values.put("updated_at", now);
         values.put("last_heartbeat_at", now);
+        // 结果上报专属信号（二次判死线用）：心跳/touchHeartbeat 不写它，只有结果回传才刷新
+        values.put("last_result_at", now);
         try {
             stringRedisTemplate.opsForHash().putAll(key, values);
             stringRedisTemplate.expire(key, ttl());
@@ -129,8 +136,10 @@ public class BrandTaskProgressCacheService {
 
     public boolean acquireFinalizeLock(Long taskId) {
         try {
+            // value 用持有者 token（而非时间戳）：释放时需要它来校验"锁还是我的"，
+            // 否则锁因 TTL 到期被他方持有后，本线程的裸 delete 会误删他人的锁
             Boolean ok = stringRedisTemplate.opsForValue()
-                    .setIfAbsent(buildFinalizeLockKey(taskId), String.valueOf(Instant.now().toEpochMilli()), FINALIZE_LOCK_TTL);
+                    .setIfAbsent(buildFinalizeLockKey(taskId), lockOwnerToken, FINALIZE_LOCK_TTL);
             return Boolean.TRUE.equals(ok);
         } catch (Exception ex) {
             log.warn("[brand-progress-cache] acquire finalize lock degraded taskId={} msg={}", taskId, ex.getMessage());
@@ -140,7 +149,12 @@ public class BrandTaskProgressCacheService {
 
     public void releaseFinalizeLock(Long taskId) {
         try {
-            stringRedisTemplate.delete(buildFinalizeLockKey(taskId));
+            // Lua 原子校验后删除（2026-09 全维度审查）：此前是裸 delete，锁已过期（TTL 到期、
+            // 他人已持有）时会把别人的锁删掉，导致同一任务被两个线程同时收尾。
+            stringRedisTemplate.execute(new DefaultRedisScript<>(
+                            "if redis.call('get', KEYS[1]) == ARGV[1] then return redis.call('del', KEYS[1]) else return 0 end",
+                            Long.class),
+                    List.of(buildFinalizeLockKey(taskId)), lockOwnerToken);
         } catch (Exception ex) {
             log.warn("[brand-progress-cache] release finalize lock degraded taskId={} msg={}", taskId, ex.getMessage());
         }

backend-java/src/main/java/com/nanri/aiimage/modules/brand/service/BrandTaskRetentionService.java

@@ -0,0 +1,93 @@
+package com.nanri.aiimage.modules.brand.service;
+
+import com.nanri.aiimage.common.service.DistributedJobLockService;
+import com.nanri.aiimage.modules.brand.mapper.BrandCrawlTaskMapper;
+import lombok.RequiredArgsConstructor;
+import lombok.extern.slf4j.Slf4j;
+import org.springframework.beans.factory.annotation.Value;
+import org.springframework.scheduling.annotation.Scheduled;
+import org.springframework.stereotype.Service;
+
+import java.time.Duration;
+import java.time.LocalDateTime;
+import java.util.List;
+
+/**
+ * 品牌检测任务（brand_crawl_tasks）的保留期清理（2026-09 审核：该表只增不删，永久累积）。
+ *
+ * <p>该表自建、不写 biz_file_task，故不在 ModuleHistoryCleanupService 的清理名单里，
+ * 此前没有任何删除路径。这里只查过期终态任务的 id，逐个走
+ * {@link BrandTaskService#deleteTask(Long)} 既有删除入口 —— 它已处理任务行删除 +
+ * 存储数据清理（brandTaskStorageService.deleteTaskData）+ 进度缓存清理，本类不重新实现删除逻辑。
+ *
+ * <p>双节点用 job 锁保证单实例执行；每批小批量（默认 50）逐个删，避免单次跑太久占住锁。
+ */
+@Service
+@RequiredArgsConstructor
+@Slf4j
+public class BrandTaskRetentionService {
+
+    /** 单轮最多处理的批次数（每批 batchSize 个任务），剩余留给下一轮。 */
+    static final int MAX_BATCHES_PER_RUN = 20;
+
+    private final BrandCrawlTaskMapper brandCrawlTaskMapper;
+    private final BrandTaskService brandTaskService;
+    private final DistributedJobLockService distributedJobLockService;
+
+    @Value("${aiimage.brand.task-retention-days:90}")
+    private int retentionDays = 90;
+
+    @Value("${aiimage.brand.task-retention-batch-size:50}")
+    private int retentionBatchSize = 50;
+
+    @Scheduled(cron = "${aiimage.brand.task-retention-cron:0 45 4 * * *}")
+    public void purgeExpiredTasks() {
+        int days = Math.max(1, retentionDays);
+        int batchSize = Math.max(1, retentionBatchSize);
+        LocalDateTime cutoff = LocalDateTime.now().minusDays(days);
+
+        // 单轮 20 批 × 50 个任务的删除（含存储数据清理）可能跑较久，锁 TTL 给足 30 分钟
+        DistributedJobLockService.LockHandle lockHandle =
+                distributedJobLockService.tryLock("brand:task-retention", Duration.ofMinutes(30));
+        if (lockHandle == null) {
+            log.info("[brand-retention] 任务保留清理跳过：另一实例持锁");
+            return;
+        }
+        try (lockHandle) {
+            int totalDeleted = 0;
+            int totalFailed = 0;
+            int batches = 0;
+            while (batches < MAX_BATCHES_PER_RUN) {
+                // 只取 id：历史行的 file_paths/result_paths JSON 字段可能很大
+                List<Long> taskIds = brandCrawlTaskMapper.selectExpiredTerminalTaskIds(cutoff, batchSize);
+                if (taskIds.isEmpty()) {
+                    break;
+                }
+                batches++;
+                int deletedInBatch = 0;
+                for (Long taskId : taskIds) {
+                    try {
+                        brandTaskService.deleteTask(taskId);
+                        deletedInBatch++;
+                    } catch (Exception ex) {
+                        // 单个任务删除失败只记日志继续：一个坏任务不能卡住整轮
+                        log.warn("[brand-retention] 任务删除失败 taskId={} msg={}", taskId, ex.getMessage(), ex);
+                    }
+                }
+                totalDeleted += deletedInBatch;
+                totalFailed += taskIds.size() - deletedInBatch;
+                if (deletedInBatch == 0) {
+                    log.warn("[brand-retention] 整批任务均删除失败，提前结束本轮以免反复重试同一批");
+                    break;
+                }
+                if (taskIds.size() < batchSize) {
+                    break;
+                }
+            }
+            log.info("[brand-retention] 任务保留清理完成 cutoff={} retentionDays={} deleted={} failed={} batches={}",
+                    cutoff, days, totalDeleted, totalFailed, batches);
+        } catch (Exception ex) {
+            log.warn("[brand-retention] 任务保留清理失败 cutoff={} msg={}", cutoff, ex.getMessage(), ex);
+        }
+    }
+}

backend-java/src/main/java/com/nanri/aiimage/modules/brand/service/BrandTaskService.java

@@ -8,7 +8,9 @@ import com.fasterxml.jackson.core.type.TypeReference;
 import com.fasterxml.jackson.databind.ObjectMapper;
 import com.nanri.aiimage.common.exception.BusinessException;
 import com.nanri.aiimage.common.service.DistributedJobLockService;
+import com.nanri.aiimage.common.util.ExcelStreamReader;
 import com.nanri.aiimage.config.BrandProgressProperties;
+import com.nanri.aiimage.config.HttpClientPool;
 import com.nanri.aiimage.config.StorageProperties;
 import com.nanri.aiimage.modules.brand.mapper.BrandCrawlTaskMapper;
 import com.nanri.aiimage.modules.brand.model.dto.BrandCrawlResultFileDto;
@@ -38,12 +40,7 @@ import com.nanri.aiimage.modules.task.service.TaskProgressSnapshotService;
 import jakarta.annotation.PreDestroy;
 import lombok.RequiredArgsConstructor;
 import lombok.extern.slf4j.Slf4j;
-import org.apache.poi.ss.usermodel.Cell;
-import org.apache.poi.ss.usermodel.DataFormatter;
-import org.apache.poi.ss.usermodel.Row;
 import org.apache.poi.ss.usermodel.Sheet;
-import org.apache.poi.ss.usermodel.Workbook;
-import org.apache.poi.ss.usermodel.WorkbookFactory;
 import org.apache.poi.xssf.streaming.SXSSFWorkbook;
 import org.springframework.stereotype.Service;
 
@@ -53,6 +50,8 @@ import java.io.FileOutputStream;
 import java.io.IOException;
 import java.io.InputStream;
 import java.net.URI;
+import java.net.http.HttpRequest;
+import java.net.http.HttpResponse;
 import java.nio.file.Files;
 import java.time.Duration;
 import java.time.Instant;
@@ -82,6 +81,21 @@ public class BrandTaskService {
 
     private static final DateTimeFormatter DATETIME_FORMATTER = DateTimeFormatter.ofPattern("yyyy-MM-dd HH:mm");
     private static final Duration STALE_CHECK_LOCK_TTL = Duration.ofMinutes(10);
+    /** 源文件下载请求超时：比普通 API 调用宽松（源文件可能几十 MB），但必须有上限。 */
+    private static final Duration SOURCE_DOWNLOAD_TIMEOUT = Duration.ofMinutes(5);
+
+    /** SSRF 防护：禁止请求云元数据与环回地址（正常源文件都在自家 OSS/MinIO 域名上）。 */
+    private static final java.util.Set<String> BLOCKED_SOURCE_HOSTS = java.util.Set.of(
+            "169.254.169.254", "metadata.google.internal", "metadata", "localhost",
+            "127.0.0.1", "0.0.0.0", "::1", "[::1]");
+
+    private static boolean isBlockedSourceHost(String host) {
+        if (host == null || host.isBlank()) {
+            return true;
+        }
+        String normalized = host.trim().toLowerCase(java.util.Locale.ROOT);
+        return BLOCKED_SOURCE_HOSTS.contains(normalized) || normalized.endsWith(".localhost");
+    }
     private static final long RESULT_SUBMIT_WAIT_MILLIS = 5 * 60 * 1000L;
     private static final String STATUS_PENDING = "pending";
     private static final String STATUS_RUNNING = "running";
@@ -92,8 +106,9 @@ public class BrandTaskService {
     /** 结果文件并发上传数：OSS/MinIO 上传互不依赖，3 并发平衡收益与内存占用。 */
     private static final int RESULT_UPLOAD_CONCURRENCY = 3;
 
-    private final ExecutorService resultUploadExecutor = Executors.newFixedThreadPool(
+    // 有界队列线程池：newFixedThreadPool 用的是无界队列，任务堆积时不会拒绝、会把内存吃满
-            RESULT_UPLOAD_CONCURRENCY, namedThreadFactory("brand-result-upload"));
+    private final ExecutorService resultUploadExecutor = com.nanri.aiimage.common.util.ThreadPools
+            .boundedFixed("brand-result-upload", RESULT_UPLOAD_CONCURRENCY);
 
     @PreDestroy
     void shutdownResultUploadExecutor() {
@@ -671,23 +686,8 @@ public class BrandTaskService {
         }
 
         brandTaskProgressCacheService.updatePhase(taskId, BrandTaskProgressCacheService.PHASE_ASSEMBLING, totalCount, totalCount);
-        File outputDir = FileUtil.mkdir(FileUtil.file(storageProperties.getLocalTempDir(), "brand-result", String.valueOf(taskId)));
-        List<OutputEntry> outputEntries = new ArrayList<>();
         try {
-            for (BrandSourceFileDto sourceFile : sourceFiles) {
+            Map<String, Object> resultPaths = assembleAndUploadResult(taskId, strategy, sourceFiles, cachedByUrl, aggregates, false);
-                BrandFileAggregateCacheDto aggregate = aggregates.get(sourceFile.getFileUrl());
-                BrandParsedFileCacheDto cachedFile = cachedByUrl.get(sourceFile.getFileUrl());
-                if (cachedFile == null) {
-                    throw new BusinessException("缺少原始缓存数据: " + sourceFile.getFileUrl());
-                }
-                File sourceLocalFile = resolveSourceFile(sourceFile);
-                String originalFilename = resolveOriginalFilename(sourceFile, sourceLocalFile);
-                File outputFile = buildNamedOutputFile(outputDir, buildResultFilename(originalFilename));
-                writeBrandWorkbook(outputFile, strategy, cachedFile, aggregate);
-                outputEntries.add(new OutputEntry(sourceLocalFile, originalFilename, outputFile, originalFilename));
-            }
-            brandTaskProgressCacheService.updatePhase(taskId, BrandTaskProgressCacheService.PHASE_UPLOADING, totalCount, totalCount);
-            Map<String, Object> resultPaths = buildAndUploadResult(taskId, outputEntries);
             int updated = brandCrawlTaskMapper.update(null, new LambdaUpdateWrapper<BrandCrawlTaskEntity>()
                     .eq(BrandCrawlTaskEntity::getId, taskId)
                     .ne(BrandCrawlTaskEntity::getStatus, STATUS_CANCELLED)
@@ -721,11 +721,150 @@ public class BrandTaskService {
                 throw businessException;
             }
             throw new BusinessException(ex.getMessage());
+        }
+    }
+
+    /**
+     * 组装并上传结果文件。partial=false 为成功收尾（要求全部文件分片收齐，由调用方校验）；
+     * partial=true 为失败中止的部分组装：用已收到的分片出结果，未检测品牌单独成 sheet。
+     */
+    private Map<String, Object> assembleAndUploadResult(Long taskId,
+                                                        String strategy,
+                                                        List<BrandSourceFileDto> sourceFiles,
+                                                        Map<String, BrandParsedFileCacheDto> cachedByUrl,
+                                                        Map<String, BrandFileAggregateCacheDto> aggregates,
+                                                        boolean partial) throws IOException {
+        File outputDir = FileUtil.mkdir(FileUtil.file(storageProperties.getLocalTempDir(), "brand-result", String.valueOf(taskId)));
+        List<OutputEntry> outputEntries = new ArrayList<>();
+        try {
+            for (BrandSourceFileDto sourceFile : sourceFiles) {
+                BrandParsedFileCacheDto cachedFile = cachedByUrl.get(sourceFile.getFileUrl());
+                if (cachedFile == null) {
+                    if (!partial) {
+                        throw new BusinessException("缺少原始缓存数据: " + sourceFile.getFileUrl());
+                    }
+                    log.warn("[brand-assemble] taskId={} 原始缓存数据缺失，跳过该文件 fileUrl={}", taskId, sourceFile.getFileUrl());
+                    continue;
+                }
+                BrandFileAggregateCacheDto aggregate = aggregates.get(sourceFile.getFileUrl());
+                if (aggregate == null) {
+                    if (!partial) {
+                        throw new BusinessException("缺少结果聚合数据: " + sourceFile.getFileUrl());
+                    }
+                    aggregate = new BrandFileAggregateCacheDto();
+                    aggregate.setFileUrl(sourceFile.getFileUrl());
+                }
+                List<String> undetectedBrands = partial ? resolveUndetectedBrands(cachedFile, aggregate) : List.of();
+                File sourceLocalFile = resolveSourceFile(sourceFile);
+                String originalFilename = resolveOriginalFilename(sourceFile, sourceLocalFile);
+                File outputFile = buildNamedOutputFile(outputDir, buildResultFilename(originalFilename));
+                writeBrandWorkbook(outputFile, strategy, cachedFile, aggregate, undetectedBrands);
+                outputEntries.add(new OutputEntry(sourceLocalFile, originalFilename, outputFile, originalFilename));
+            }
+            if (outputEntries.isEmpty()) {
+                throw new BusinessException("没有可组装的结果文件");
+            }
+            brandTaskProgressCacheService.updatePhase(taskId, BrandTaskProgressCacheService.PHASE_UPLOADING,
+                    sourceFiles.size(), sourceFiles.size());
+            return buildAndUploadResult(taskId, outputEntries);
         } finally {
             cleanupBrandResultTempFiles(outputEntries, outputDir);
         }
     }
 
+    /** 未检测品牌 = 源文件品牌 -（保留 ∪ 不符合品牌 ∪ 查询失败品牌）；按源文件出现顺序去重。 */
+    private List<String> resolveUndetectedBrands(BrandParsedFileCacheDto cachedFile, BrandFileAggregateCacheDto aggregate) {
+        List<Map<String, Object>> sourceRows = cachedFile.getRows() == null ? List.of() : cachedFile.getRows();
+        Set<String> handled = new LinkedHashSet<>();
+        handled.addAll(normalizeBrandSet(aggregate.getKeptBrands()));
+        handled.addAll(normalizeBrandSetFromInvalids(aggregate.getInvalidBrands()));
+        handled.addAll(normalizeBrandSet(aggregate.getQueryFailedBrands()));
+        LinkedHashSet<String> undetected = new LinkedHashSet<>();
+        for (Map<String, Object> rowData : sourceRows) {
+            String brand = normalizeCellText(Objects.toString(rowData.getOrDefault("品牌", ""), ""));
+            if (!brand.isBlank() && !handled.contains(brand)) {
+                undetected.add(brand);
+            }
+        }
+        return new ArrayList<>(undetected);
+    }
+
+    /**
+     * 爬取方（15126）中止上报：任务不可能再收到剩余分片时调用（如 WIPO 连续限流熔断）。
+     * 用已收到的分片部分组装结果文件并落 failed + 真实原因——否则 Java 侧任务会悬挂到
+     * 心跳超时被判「前端长时间无响应」，且已跑出的数据因没有 result_paths 无法下载。
+     * 幂等：任务已是终态时直接返回，不覆盖既有结果。
+     */
+    public Map<String, Object> abortTask(Long taskId, String errorMessage) {
+        if (taskId == null || taskId <= 0) {
+            throw new BusinessException("taskId invalid");
+        }
+        try (TaskDistributedLockService.LockHandle ignored = acquireTaskLockOrThrow(taskId, RESULT_SUBMIT_WAIT_MILLIS)) {
+            BrandCrawlTaskEntity task = requireTask(taskId);
+            String status = blankToDefault(task.getStatus(), STATUS_PENDING);
+            Map<String, Object> data = new LinkedHashMap<>();
+            data.put("taskId", taskId);
+            boolean hasResult = task.getResultPaths() != null && !task.getResultPaths().isBlank();
+            // success/cancelled 不动；failed 已有结果也不重复组装。failed 且无结果
+            // （如被心跳超时兜底判失败的历史任务）允许补组装——存量补救路径。
+            if (STATUS_SUCCESS.equalsIgnoreCase(status) || STATUS_CANCELLED.equalsIgnoreCase(status)
+                    || (STATUS_FAILED.equalsIgnoreCase(status) && hasResult)) {
+                log.info("[brand-abort] taskId={} 已是终态且无需补组装 status={} hasResult={}，跳过",
+                        taskId, status, hasResult);
+                data.put("status", status);
+                data.put("resultGenerated", false);
+                return data;
+            }
+            String message = blankToDefault(errorMessage,
+                    blankToDefault(task.getErrorMessage(), "品牌检测任务已中止"));
+            List<BrandSourceFileDto> sourceFiles = parseSourceFiles(task.getFilePaths());
+            int totalCount = sourceFiles.size();
+            Map<String, BrandParsedFileCacheDto> cachedByUrl =
+                    indexCachedFiles(brandTaskStorageService.getParsedPayload(taskId));
+            // 先从分片重建聚合再读取：缓存的 state_json 可能是旧版本（缺后加字段，
+            // 如 keptBrands）或与已落库分片不一致，直接读会把已检测品牌误判为未检测
+            for (BrandSourceFileDto sourceFile : sourceFiles) {
+                brandTaskStorageService.refreshFileAggregate(taskId, sourceFile.getFileUrl());
+            }
+            Map<String, BrandFileAggregateCacheDto> aggregates = brandTaskStorageService.getAllFileAggregates(taskId);
+            boolean hasChunk = aggregates.values().stream()
+                    .anyMatch(item -> item != null && defaultInteger(item.getReceivedChunkCount()) > 0);
+            Map<String, Object> resultPaths = null;
+            if (hasChunk && !cachedByUrl.isEmpty() && !sourceFiles.isEmpty()) {
+                try {
+                    resultPaths = assembleAndUploadResult(taskId, normalizeStrategy(task.getStrategy()),
+                            sourceFiles, cachedByUrl, aggregates, true);
+                    log.info("[brand-abort] taskId={} 部分结果组装完成 files={}", taskId, sourceFiles.size());
+                } catch (Exception ex) {
+                    log.warn("[brand-abort] taskId={} 部分结果组装失败（仅标记失败） msg={}", taskId, ex.getMessage(), ex);
+                }
+            } else {
+                log.info("[brand-abort] taskId={} 无已收到分片，跳过结果组装 receivedAggregates={}", taskId, aggregates.size());
+            }
+            int finishedCount = brandTaskStorageService.countCompletedFiles(taskId);
+            LambdaUpdateWrapper<BrandCrawlTaskEntity> wrapper = new LambdaUpdateWrapper<BrandCrawlTaskEntity>()
+                    .eq(BrandCrawlTaskEntity::getId, taskId)
+                    .in(BrandCrawlTaskEntity::getStatus, STATUS_PENDING, STATUS_RUNNING, STATUS_FAILED)
+                    .set(BrandCrawlTaskEntity::getStatus, STATUS_FAILED)
+                    .set(BrandCrawlTaskEntity::getErrorMessage, message)
+                    .set(BrandCrawlTaskEntity::getProgressCurrent, finishedCount)
+                    .set(BrandCrawlTaskEntity::getProgressTotal, totalCount);
+            if (resultPaths != null) {
+                wrapper.set(BrandCrawlTaskEntity::getResultPaths, JSONUtil.toJsonStr(resultPaths));
+            }
+            int updated = brandCrawlTaskMapper.update(null, wrapper);
+            if (updated > 0) {
+                brandTaskProgressCacheService.markFailed(taskId, message);
+                saveBrandProgressSnapshot(taskId, STATUS_FAILED, totalCount, finishedCount, 1, message);
+            }
+            log.info("[brand-abort] taskId={} aborted updated={} resultGenerated={} finishedFiles={}/{} msg={}",
+                    taskId, updated, resultPaths != null, finishedCount, totalCount, message);
+            data.put("status", STATUS_FAILED);
+            data.put("resultGenerated", resultPaths != null);
+            return data;
+        }
+    }
+
     private void ensureNoDuplicateResultFiles(List<BrandCrawlResultFileDto> resultFiles) {
         Set<String> seen = new LinkedHashSet<>();
         for (BrandCrawlResultFileDto resultFile : resultFiles) {
@@ -812,21 +951,49 @@ public class BrandTaskService {
     }
 
     private File downloadSourceFile(String fileUrl) {
+        URI uri;
         try {
-            URI uri = URI.create(fileUrl);
+            uri = URI.create(fileUrl);
+        } catch (Exception ex) {
+            log.warn("[brand] 源文件地址非法 fileUrl={} err={}", fileUrl, ex.getMessage());
+            throw new BusinessException("下载源文件失败: 地址非法");
+        }
+        // SSRF 防护（2026-09 全维度审查）：fileUrl 来自请求体，若不校验则匿名调用方可让服务端
+        // 请求内网/云元数据地址（169.254.169.254 等）。正常源文件都落在自家 OSS/MinIO 上。
+        if (isBlockedSourceHost(uri.getHost())) {
+            log.warn("[brand] 拒绝下载疑似 SSRF 的源文件地址 host={} fileUrl={}", uri.getHost(), fileUrl);
+            throw new BusinessException("源文件地址不合法");
+        }
         String filename = FileUtil.getName(uri.getPath());
         if (filename == null || filename.isBlank()) {
             filename = "brand-source.xlsx";
         }
         String suffix = FileUtil.extName(filename);
         File downloadDir = FileUtil.mkdir(FileUtil.file(storageProperties.getLocalTempDir(), "brand-source-download"));
+        try {
             File tempFile = Files.createTempFile(downloadDir.toPath(), "brand_", suffix.isBlank() ? "" : "." + suffix).toFile();
-            try (InputStream inputStream = uri.toURL().openStream()) {
+            // 走统一连接池 + 显式超时。此前 uri.toURL().openStream() 无任何超时（JVM 默认 0 = 无限），
+            // 上游半开连接会把 Tomcat 工作线程无限占用；超时值比普通 API 调用宽松（源文件可能几十 MB）
+            HttpRequest downloadRequest = HttpRequest.newBuilder(uri)
+                    .timeout(SOURCE_DOWNLOAD_TIMEOUT)
+                    .GET()
+                    .build();
+            HttpResponse<InputStream> response = HttpClientPool.sharedHttpClient()
+                    .send(downloadRequest, HttpResponse.BodyHandlers.ofInputStream());
+            if (response.statusCode() / 100 != 2) {
+                log.warn("[brand] 下载源文件返回非 2xx fileUrl={} status={}", fileUrl, response.statusCode());
+                throw new BusinessException("下载源文件失败: HTTP " + response.statusCode());
+            }
+            try (InputStream inputStream = response.body()) {
                 FileUtil.writeFromStream(inputStream, tempFile);
             }
             return tempFile;
+        } catch (BusinessException ex) {
+            throw ex;
         } catch (Exception ex) {
-            throw new BusinessException("下载源文件失败");
+            // 此前该 catch 既不记日志也不带 cause，线上无法定位
+            log.warn("[brand] 下载源文件失败 fileUrl={} err={}", fileUrl, ex.getMessage(), ex);
+            throw new BusinessException("下载源文件失败", ex);
         }
     }
 
@@ -835,54 +1002,60 @@ public class BrandTaskService {
     }
 
     private ParsedBrandFile parseBrandFile(File inputFile) {
-        DataFormatter formatter = new DataFormatter();
+        // 2026-09 全维度审查 C5：改逐行流式解析（ExcelStreamReader → EasyExcel SAX）。
-        try (FileInputStream fis = new FileInputStream(inputFile); Workbook workbook = WorkbookFactory.create(fis)) {
+        // 此前 WorkbookFactory.create 把整表读成 DOM，品牌源文件（几十万行）会整表驻留堆内存。
-            Sheet sheet = workbook.getSheetAt(0);
+        // 解析结果（columns/rows/uniqueBrands/行号）与错误文案不变，行号沿用 sheet 绝对 0 基行号 +1。
-            Row headerRow = sheet.getRow(0);
+        if (!hasExcelMagic(inputFile)) {
-            if (headerRow == null) {
+            log.warn("[brand] 源文件不是 Excel（疑似 CSV/文本），拒绝解析 file={}", inputFile.getName());
-                throw new BusinessException("Excel 表头为空");
+            throw new BusinessException("读取 Excel 失败");
         }
-            List<String> columns = extractHeaders(headerRow, formatter);
+        BrandFileSheetContext context = new BrandFileSheetContext();
-            if (columns.isEmpty()) {
+        try {
-                throw new BusinessException("未读取到有效表头");
+            ExcelStreamReader.readFirstSheet(inputFile, context);
-            }
+        } catch (BusinessException ex) {
-            Map<String, Integer> headerIndexes = buildHeaderIndexes(headerRow, formatter, columns);
+            throw ex;
-            List<Map<String, Object>> rows = new ArrayList<>();
-            Set<String> uniqueBrands = new LinkedHashSet<>();
-            for (int rowNum = 1; rowNum <= sheet.getLastRowNum(); rowNum++) {
-                Row row = sheet.getRow(rowNum);
-                if (row == null) {
-                    continue;
-                }
-                Map<String, Object> rowData = new LinkedHashMap<>();
-                for (String column : columns) {
-                    Integer index = headerIndexes.get(column);
-                    String value = index == null ? "" : normalizeCellText(formatter.formatCellValue(row.getCell(index)));
-                    rowData.put(column, value);
-                }
-                String brand = normalizeCellText(Objects.toString(rowData.getOrDefault("品牌", ""), ""));
-                if (brand.isBlank()) {
-                    rowData.put("__rowIndex", rowNum + 1);
-                    rows.add(rowData);
-                    continue;
-                }
-                if (uniqueBrands.add(brand)) {
-                    rowData.put("__rowIndex", rowNum + 1);
-                    rows.add(rowData);
-                }
-            }
-            return new ParsedBrandFile(sheet.getSheetName(), columns, rows, new ArrayList<>(uniqueBrands));
         } catch (IOException ex) {
             throw new BusinessException("读取 Excel 失败");
         }
+        return context.finish();
     }
 
-    private List<String> extractHeaders(Row headerRow, DataFormatter formatter) {
+    /**
+     * 文件魔数校验：xlsx=PK(zip)、xls=OLE2(CFD)。
+     * EasyExcel 会把非 zip 文本当 CSV 静默解析成功，而旧 POI WorkbookFactory 只认 xlsx/xls（垃圾文件直接失败），
+     * 这里保持「非 Excel 源文件 → 读取失败」的语义（与 SimilarAsinExcelParser 同口径）。
+     */
+    private boolean hasExcelMagic(File file) {
+        byte[] head = new byte[8];
+        try (InputStream inputStream = new FileInputStream(file)) {
+            int n = inputStream.readNBytes(head, 0, head.length);
+            boolean isZip = n >= 4 && head[0] == 'P' && head[1] == 'K' && (head[2] == 3 || head[2] == 5 || head[2] == 7);
+            boolean isOle2 = n >= 8
+                    && (head[0] & 0xFF) == 0xD0 && (head[1] & 0xFF) == 0xCF
+                    && (head[2] & 0xFF) == 0x11 && (head[3] & 0xFF) == 0xE0;
+            return isZip || isOle2;
+        } catch (IOException ex) {
+            return false;
+        }
+    }
+
+    /** 品牌源文件流式解析上下文：表头解析 + 按品牌去重累积结果行（内存不随整表 DOM 放大）。 */
+    private final class BrandFileSheetContext implements ExcelStreamReader.SheetRowHandler {
+
+        private final List<Map<String, Object>> rows = new ArrayList<>();
+        private final Set<String> uniqueBrands = new LinkedHashSet<>();
+        private List<String> columns;
+        private Map<String, Integer> headerIndexes;
+        private String sheetName = "";
+
+        @Override
+        public void onHeader(String sheetName, Integer sheetNo, Map<Integer, String> headerMap) {
+            this.sheetName = sheetName == null ? "" : sheetName;
+            int lastColumnCount = lastColumnCount(headerMap);
             List<String> headers = new ArrayList<>();
             Set<String> seen = new LinkedHashSet<>();
-        for (int i = 0; i < headerRow.getLastCellNum(); i++) {
+            for (int i = 0; i < lastColumnCount; i++) {
-            Cell cell = headerRow.getCell(i);
+                String value = normalizeHeaderValue(headerMap.get(i));
-            String value = normalizeHeaderValue(cell == null ? null : formatter.formatCellValue(cell));
                 if (value.isBlank() || seen.contains(value)) {
                     continue;
                 }
@@ -892,24 +1065,70 @@ public class BrandTaskService {
                     break;
                 }
             }
-        return headers;
+            if (headers.isEmpty()) {
+                throw new BusinessException("未读取到有效表头");
             }
-
+            Map<String, Integer> indexes = new LinkedHashMap<>();
-    private Map<String, Integer> buildHeaderIndexes(Row headerRow, DataFormatter formatter, List<String> columns) {
+            Set<String> allowed = new LinkedHashSet<>(headers);
-        Map<String, Integer> headerIndexes = new LinkedHashMap<>();
+            for (int i = 0; i < lastColumnCount; i++) {
-        Set<String> allowed = new LinkedHashSet<>(columns);
+                String value = normalizeHeaderValue(headerMap.get(i));
-        for (int i = 0; i < headerRow.getLastCellNum(); i++) {
+                if (value.isBlank() || indexes.containsKey(value) || !allowed.contains(value)) {
-            Cell cell = headerRow.getCell(i);
-            String value = normalizeHeaderValue(cell == null ? null : formatter.formatCellValue(cell));
-            if (value.isBlank() || headerIndexes.containsKey(value) || !allowed.contains(value)) {
                     continue;
                 }
-            headerIndexes.put(value, i);
+                indexes.put(value, i);
                 if ("缩略图地址8".equals(value)) {
                     break;
                 }
             }
-        return headerIndexes;
+            this.columns = headers;
+            this.headerIndexes = indexes;
+        }
+
+        @Override
+        public void onRow(String sheetName, Integer sheetNo, int rowIndex,
+                          Map<Integer, String> headerMap, Map<Integer, String> rowMap) {
+            if (columns == null) {
+                // 表头行整行为空时 EasyExcel 不上报表头回调：与旧实现 headerRow == null 一致
+                throw new BusinessException("Excel 表头为空");
+            }
+            Map<String, Object> rowData = new LinkedHashMap<>();
+            for (String column : columns) {
+                Integer index = headerIndexes.get(column);
+                String value = index == null ? "" : normalizeCellText(rowMap.get(index));
+                rowData.put(column, value);
+            }
+            String brand = normalizeCellText(Objects.toString(rowData.getOrDefault("品牌", ""), ""));
+            if (brand.isBlank()) {
+                rowData.put("__rowIndex", rowIndex + 1);
+                rows.add(rowData);
+                return;
+            }
+            if (uniqueBrands.add(brand)) {
+                rowData.put("__rowIndex", rowIndex + 1);
+                rows.add(rowData);
+            }
+        }
+
+        private ParsedBrandFile finish() {
+            if (columns == null) {
+                // 空表（无任何行）或表头行缺失：与旧实现 headerRow == null 一致
+                throw new BusinessException("Excel 表头为空");
+            }
+            return new ParsedBrandFile(sheetName, columns, rows, new ArrayList<>(uniqueBrands));
+        }
+
+        private int lastColumnCount(Map<Integer, String> headerMap) {
+            if (headerMap == null || headerMap.isEmpty()) {
+                return 0;
+            }
+            int lastColumnCount = 0;
+            for (Integer columnIndex : headerMap.keySet()) {
+                if (columnIndex != null && columnIndex >= lastColumnCount) {
+                    lastColumnCount = columnIndex + 1;
+                }
+            }
+            return lastColumnCount;
+        }
     }
 
     private String normalizeHeaderValue(String value) {
@@ -939,7 +1158,8 @@ public class BrandTaskService {
     private void writeBrandWorkbook(File outputFile,
                                     String strategy,
                                     BrandParsedFileCacheDto cachedFile,
-                                    BrandFileAggregateCacheDto resultFile) throws IOException {
+                                    BrandFileAggregateCacheDto resultFile,
+                                    List<String> undetectedBrands) throws IOException {
         String actualStrategy = normalizeStrategy(strategy);
         SXSSFWorkbook workbook = new SXSSFWorkbook(200);
         workbook.setCompressTempFiles(true);
@@ -953,6 +1173,10 @@ public class BrandTaskService {
             }
 
             Set<String> invalidBrandSet = normalizeBrandSetFromInvalids(resultFile.getInvalidBrands());
+            // 未检测品牌（任务中止、分片没到齐）：主 sheet 剔除这些行、整体挪到独立 sheet，
+            // 避免用户把「没查过」的行误当成「已通过检测」上架
+            Set<String> undetectedBrandSet = normalizeBrandSet(undetectedBrands);
+            List<Map<String, Object>> undetectedRows = new ArrayList<>();
             List<Map<String, Object>> sourceRows = cachedFile.getRows() == null ? List.of() : cachedFile.getRows();
             int writeRowIndex = 1;
             for (Map<String, Object> rowData : sourceRows) {
@@ -960,6 +1184,10 @@ public class BrandTaskService {
                 if (!brand.isBlank() && invalidBrandSet.contains(brand)) {
                     continue;
                 }
+                if (!brand.isBlank() && undetectedBrandSet.contains(brand)) {
+                    undetectedRows.add(rowData);
+                    continue;
+                }
                 var row = mainSheet.createRow(writeRowIndex++);
                 for (int colIndex = 0; colIndex < columns.size(); colIndex++) {
                     String column = columns.get(colIndex);
@@ -967,6 +1195,23 @@ public class BrandTaskService {
                 }
             }
 
+            if (!undetectedRows.isEmpty()) {
+                var undetectedSheet = workbook.createSheet("未检测品牌");
+                var undetectedHeader = undetectedSheet.createRow(0);
+                for (int i = 0; i < columns.size(); i++) {
+                    undetectedHeader.createCell(i).setCellValue(columns.get(i));
+                }
+                for (int i = 0; i < undetectedRows.size(); i++) {
+                    Map<String, Object> rowData = undetectedRows.get(i);
+                    var row = undetectedSheet.createRow(i + 1);
+                    for (int colIndex = 0; colIndex < columns.size(); colIndex++) {
+                        String column = columns.get(colIndex);
+                        row.createCell(colIndex).setCellValue(Objects.toString(rowData.getOrDefault(column, ""), ""));
+                    }
+                }
+                applyBrandSheetWidths(undetectedSheet, columns.size());
+            }
+
             var invalidSheet = workbook.createSheet("不符合品牌");
             var invalidHeader = invalidSheet.createRow(0);
             invalidHeader.createCell(0).setCellValue("品牌");
@@ -1292,6 +1537,55 @@ public class BrandTaskService {
             failStaleRunningTask(task.getId(), "前端长时间无响应，任务已自动失败");
             }
         }
+        failNoUploadStaleRunningTasks();
+    }
+    }
+
+    /**
+     * 二次判死：前端心跳正常但连续 N 分钟无结果上报（治「心跳续命」的假活任务）。
+     *
+     * <p>既有心跳线候选条件是 updated_at/last_heartbeat_at 陈旧，而前端心跳会持续刷新它们——
+     * 主线程卡死时心跳线程照发，任务永远命不中（与生产 28131 同型缺口）。
+     * 本线候选取「心跳新鲜 + 创建超过 N 分钟」，判据用结果上报时写入 progress hash 的
+     * last_result_at（仅 saveProgressFromResult 写）；从未上报（无该字段）跳过，避免误杀首批较慢的正常任务。
+     */
+    private void failNoUploadStaleRunningTasks() {
+        long minutes = brandProgressProperties.getNoResultUploadTimeoutMinutes();
+        if (minutes <= 0) {
+            return;
+        }
+        LocalDateTime heartbeatThreshold = LocalDateTime.now()
+                .minusMinutes(brandProgressProperties.getHeartbeatTimeoutMinutes());
+        LocalDateTime cutoff = LocalDateTime.now().minusMinutes(minutes);
+        List<BrandCrawlTaskEntity> runningTasks = brandCrawlTaskMapper.selectList(new LambdaQueryWrapper<BrandCrawlTaskEntity>()
+                .eq(BrandCrawlTaskEntity::getStatus, STATUS_RUNNING)
+                .ge(BrandCrawlTaskEntity::getUpdatedAt, heartbeatThreshold)
+                .lt(BrandCrawlTaskEntity::getCreatedAt, cutoff));
+        for (BrandCrawlTaskEntity task : runningTasks) {
+            TaskDistributedLockService.LockHandle taskLockHandle = acquireTaskLock(task.getId(), 0L);
+            if (taskLockHandle == null) {
+                continue;
+            }
+            try (taskLockHandle) {
+                Map<Object, Object> progress = brandTaskProgressCacheService.getProgress(task.getId());
+                long lastResultAt = 0L;
+                try {
+                    lastResultAt = Long.parseLong(String.valueOf(progress.getOrDefault("last_result_at", "0")));
+                } catch (Exception ignored) {
+                }
+                if (lastResultAt <= 0L) {
+                    // 从未上报结果：保守跳过（首批可能较慢）
+                    continue;
+                }
+                LocalDateTime lastResult = LocalDateTime.ofInstant(Instant.ofEpochMilli(lastResultAt), ZoneId.systemDefault());
+                if (lastResult.isAfter(cutoff)) {
+                    continue;
+                }
+                log.warn("[brand-stale-check] no-upload failing taskId={} lastResultAt={} timeoutMinutes={}",
+                        task.getId(), lastResult, minutes);
+                failStaleRunningTask(task.getId(),
+                        "连续 " + minutes + " 分钟无结果回传，疑似客户端任务卡死，任务已自动失败（最后结果上报 " + lastResult + "）");
+            }
         }
     }
 

backend-java/src/main/java/com/nanri/aiimage/modules/brand/service/BrandTaskStaleRepairSpiImpl.java

@@ -0,0 +1,47 @@
+package com.nanri.aiimage.modules.brand.service;
+
+import com.baomidou.mybatisplus.core.conditions.query.LambdaQueryWrapper;
+import com.baomidou.mybatisplus.core.conditions.update.LambdaUpdateWrapper;
+import com.nanri.aiimage.modules.brand.mapper.BrandCrawlTaskMapper;
+import com.nanri.aiimage.modules.brand.model.entity.BrandCrawlTaskEntity;
+import com.nanri.aiimage.modules.task.spi.BrandTaskStaleRepairSpi;
+import lombok.RequiredArgsConstructor;
+import org.springframework.stereotype.Service;
+
+import java.time.LocalDateTime;
+import java.util.List;
+
+/**
+ * brand_crawl_tasks 的陈旧修复实现（2026-09 全维度审查 G5：逻辑从 task 模块收回本模块）。
+ */
+@Service
+@RequiredArgsConstructor
+public class BrandTaskStaleRepairSpiImpl implements BrandTaskStaleRepairSpi {
+
+    private static final String STATUS_PENDING = "pending";
+    private static final String STATUS_RUNNING = "running";
+    private static final String STATUS_FAILED = "failed";
+
+    private final BrandCrawlTaskMapper brandCrawlTaskMapper;
+
+    @Override
+    public List<Long> failStaleBrandTasks(LocalDateTime cutoff, int limit) {
+        List<BrandCrawlTaskEntity> stale = brandCrawlTaskMapper.selectList(
+                new LambdaQueryWrapper<BrandCrawlTaskEntity>()
+                        .select(BrandCrawlTaskEntity::getId)
+                        .in(BrandCrawlTaskEntity::getStatus, STATUS_PENDING, STATUS_RUNNING)
+                        .lt(BrandCrawlTaskEntity::getUpdatedAt, cutoff)
+                        .last("limit " + Math.max(1, limit)));
+        if (stale.isEmpty()) {
+            return List.of();
+        }
+        List<Long> ids = stale.stream().map(BrandCrawlTaskEntity::getId).toList();
+        brandCrawlTaskMapper.update(null, new LambdaUpdateWrapper<BrandCrawlTaskEntity>()
+                .in(BrandCrawlTaskEntity::getId, ids)
+                .in(BrandCrawlTaskEntity::getStatus, STATUS_PENDING, STATUS_RUNNING)
+                .lt(BrandCrawlTaskEntity::getUpdatedAt, cutoff)
+                .set(BrandCrawlTaskEntity::getStatus, STATUS_FAILED)
+                .set(BrandCrawlTaskEntity::getErrorMessage, "任务长期无心跳，已自动失败"));
+        return ids;
+    }
+}

backend-java/src/main/java/com/nanri/aiimage/modules/brand/service/BrandTaskStorageService.java

@@ -13,9 +13,12 @@ import com.nanri.aiimage.modules.task.model.entity.TaskChunkEntity;
 import com.nanri.aiimage.modules.task.model.entity.TaskScopeStateEntity;
 import com.nanri.aiimage.modules.task.service.TransientPayloadStorageService;
 import lombok.RequiredArgsConstructor;
+import lombok.extern.slf4j.Slf4j;
 import org.springframework.dao.DuplicateKeyException;
 import org.springframework.stereotype.Service;
 import org.springframework.transaction.annotation.Transactional;
+import org.springframework.transaction.support.TransactionSynchronization;
+import org.springframework.transaction.support.TransactionSynchronizationManager;
 
 import java.nio.charset.StandardCharsets;
 import java.security.MessageDigest;
@@ -27,6 +30,7 @@ import java.util.Map;
 
 @Service
 @RequiredArgsConstructor
+@Slf4j
 public class BrandTaskStorageService {
 
     private static final String MODULE_TYPE = "BRAND";
@@ -236,6 +240,13 @@ public class BrandTaskStorageService {
         return new ChunkStoreResult(false, completed, countCompletedFiles(taskId), aggregate);
     }
 
+    /**
+     * 删除任务的全部范围/分片数据。
+     *
+     * <p>事务边界：远端载荷删除移到事务提交后执行 —— 单次删除超时 120s、重试 3 次，
+     * 放在 DB 事务里会让连接被长时间占用（项目在 TaskScopePayloadStorageService 已写明该约束）。
+     * 提交后再删还有个好处：引用计数查询看到的是删除完成后的 DB 状态，判断更准确。
+     */
     @Transactional
     public void deleteTaskData(Long taskId) {
         if (taskId == null || taskId <= 0) {
@@ -245,27 +256,68 @@ public class BrandTaskStorageService {
                 .select(TaskScopeStateEntity::getParsedPayloadJson, TaskScopeStateEntity::getStateJson)
                 .eq(TaskScopeStateEntity::getTaskId, taskId)
                 .eq(TaskScopeStateEntity::getModuleType, MODULE_TYPE));
-        if (states != null) {
-            for (TaskScopeStateEntity state : states) {
-                transientPayloadStorageService.deletePayloadIfPresent(state.getParsedPayloadJson());
-                transientPayloadStorageService.deletePayloadIfPresent(state.getStateJson());
-            }
-        }
         List<TaskChunkEntity> chunks = taskChunkMapper.selectList(new LambdaQueryWrapper<TaskChunkEntity>()
                 .select(TaskChunkEntity::getPayloadJson)
                 .eq(TaskChunkEntity::getTaskId, taskId)
                 .eq(TaskChunkEntity::getModuleType, MODULE_TYPE));
-        if (chunks != null) {
-            for (TaskChunkEntity chunk : chunks) {
-                transientPayloadStorageService.deletePayloadIfPresent(chunk.getPayloadJson());
-            }
-        }
         taskChunkMapper.delete(new LambdaQueryWrapper<TaskChunkEntity>()
                 .eq(TaskChunkEntity::getTaskId, taskId)
                 .eq(TaskChunkEntity::getModuleType, MODULE_TYPE));
         taskScopeStateMapper.delete(new LambdaQueryWrapper<TaskScopeStateEntity>()
                 .eq(TaskScopeStateEntity::getTaskId, taskId)
                 .eq(TaskScopeStateEntity::getModuleType, MODULE_TYPE));
+        deletePayloadsAfterCommit(states, chunks, taskId);
+    }
+
+    /** 事务提交后逐条删远端载荷；无事务时立即执行（与 TaskResultItemService 同款兜底）。 */
+    private void deletePayloadsAfterCommit(List<TaskScopeStateEntity> states,
+                                           List<TaskChunkEntity> chunks,
+                                           Long taskId) {
+        List<String> payloads = new ArrayList<>();
+        if (states != null) {
+            for (TaskScopeStateEntity state : states) {
+                if (state == null) {
+                    continue;
+                }
+                if (state.getParsedPayloadJson() != null && !state.getParsedPayloadJson().isBlank()) {
+                    payloads.add(state.getParsedPayloadJson());
+                }
+                if (state.getStateJson() != null && !state.getStateJson().isBlank()) {
+                    payloads.add(state.getStateJson());
+                }
+            }
+        }
+        if (chunks != null) {
+            for (TaskChunkEntity chunk : chunks) {
+                if (chunk != null && chunk.getPayloadJson() != null && !chunk.getPayloadJson().isBlank()) {
+                    payloads.add(chunk.getPayloadJson());
+                }
+            }
+        }
+        if (payloads.isEmpty()) {
+            return;
+        }
+        if (TransactionSynchronizationManager.isSynchronizationActive()) {
+            TransactionSynchronizationManager.registerSynchronization(new TransactionSynchronization() {
+                @Override
+                public void afterCommit() {
+                    deletePayloadsNow(payloads, taskId);
+                }
+            });
+            return;
+        }
+        deletePayloadsNow(payloads, taskId);
+    }
+
+    private void deletePayloadsNow(List<String> payloads, Long taskId) {
+        for (String payload : payloads) {
+            try {
+                transientPayloadStorageService.deletePayloadIfPresent(payload);
+            } catch (Exception ex) {
+                // 事务已提交，远端删除失败只记日志，不影响任务数据清理结果
+                log.warn("[brand-storage] 载荷删除失败 taskId={} msg={}", taskId, ex.getMessage());
+            }
+        }
     }
 
     private void saveAggregate(Long taskId,
@@ -325,6 +377,7 @@ public class BrandTaskStorageService {
         aggregate.setCompleted(false);
         aggregate.setInvalidBrands(new ArrayList<>());
         aggregate.setQueryFailedBrands(new ArrayList<>());
+        aggregate.setKeptBrands(new ArrayList<>());
         return aggregate;
     }
 
@@ -362,6 +415,12 @@ public class BrandTaskStorageService {
         if (file.getQueryFailedBrands() != null && !file.getQueryFailedBrands().isEmpty()) {
             aggregate.getQueryFailedBrands().addAll(file.getQueryFailedBrands());
         }
+        if (file.getKeptRows() != null && !file.getKeptRows().isEmpty()) {
+            if (aggregate.getKeptBrands() == null) {
+                aggregate.setKeptBrands(new ArrayList<>());
+            }
+            aggregate.getKeptBrands().addAll(file.getKeptRows());
+        }
     }
 
     private BrandFileAggregateCacheDto rebuildAggregate(Long taskId, String scopeKey, String scopeHash) {
@@ -379,6 +438,7 @@ public class BrandTaskStorageService {
                 aggregate.setCompleted(false);
                 aggregate.setInvalidBrands(new ArrayList<>());
                 aggregate.setQueryFailedBrands(new ArrayList<>());
+                aggregate.setKeptBrands(new ArrayList<>());
             }
             return aggregate;
         }
@@ -466,11 +526,8 @@ public class BrandTaskStorageService {
         try {
             MessageDigest digest = MessageDigest.getInstance("SHA-256");
             byte[] bytes = digest.digest(normalize(value).getBytes(StandardCharsets.UTF_8));
-            StringBuilder sb = new StringBuilder(bytes.length * 2);
+            // HexFormat 替代逐字节 String.format("%02x")：后者每次走 32 次 Formatter（热路径按行调用）
-            for (byte b : bytes) {
+            return java.util.HexFormat.of().formatHex(bytes);
-                sb.append(String.format("%02x", b));
-            }
-            return sb.toString();
         } catch (Exception ex) {
             throw new IllegalStateException("failed to hash brand scope", ex);
         }

backend-java/src/main/java/com/nanri/aiimage/modules/collectdata/controller/CollectDataController.java

@@ -29,6 +29,7 @@ import org.springframework.web.bind.annotation.RequestParam;
 import org.springframework.web.bind.annotation.RestController;
 import com.nanri.aiimage.modules.task.model.dto.TaskProgressLightRequest;
 import com.nanri.aiimage.modules.task.model.vo.TaskProgressLightBatchVo;
+import com.nanri.aiimage.modules.task.service.TaskProgressOwnershipSupport;
 
 @RestController
 @RequiredArgsConstructor
@@ -37,6 +38,7 @@ import com.nanri.aiimage.modules.task.model.vo.TaskProgressLightBatchVo;
 public class CollectDataController {
 
     private final CollectDataService service;
+    private final TaskProgressOwnershipSupport progressOwnershipSupport;
 
     @PostMapping("/parse")
     @Operation(summary = "解析 Excel 并创建任务", description = "解析上传后的 Excel 文件，按行入库到 biz_collect_data_item，并保存任务筛选条件。任务初始状态为 PENDING。")
@@ -112,15 +114,18 @@ public class CollectDataController {
     @PostMapping("/tasks/progress/batch")
     @Operation(summary = "批量查询任务进度")
     public ApiResponse<CollectDataTaskBatchVo> progressBatch(@Valid @RequestBody CollectDataTaskBatchRequest request) {
-        return ApiResponse.success(service.progressBatch(request.getTaskIds()));
+        // 归属过滤：只查询属于该用户的任务（userId 未传时保持原行为，兼容未升级调用方）
+        return ApiResponse.success(service.progressBatch(progressOwnershipSupport
+                .filterOwnedTaskIds(request.getTaskIds(), request.getUserId())));
     }
 
     @PostMapping("/tasks/progress/light")
     @Operation(summary = "批量查询任务轻量进度（前端轮询专用）",
-            description = "与 progress/batch 同输入，只返回轻量字段（taskId/status/statusCode/fileStatus/fileError/fileReady/updatedAt），"
+            description = "与 progress/batch 同输入，只返回轻量字段（taskId/status/statusCode/fileStatus/fileError/fileReady/updatedAt/rowsVersion），"
-                    + "不查明细/result 行，响应体更小；旧 progress/batch 端点保留不动。")
+                    + "不返回行明细内容（仅多一次结果行版本聚合查询），响应体更小；旧 progress/batch 端点保留不动。")
     public ApiResponse<TaskProgressLightBatchVo> progressLight(@Valid @RequestBody TaskProgressLightRequest request) {
-        return ApiResponse.success(service.progressLight(request.getTaskIds()));
+        // 透传 userId：此前传 null = 不过滤归属，匿名遍历 taskId 即可读他人任务状态（2026-09 审查）
+        return ApiResponse.success(service.progressLight(request.getTaskIds(), request.getUserId()));
     }
 
     @DeleteMapping("/tasks/{taskId}")

backend-java/src/main/java/com/nanri/aiimage/modules/collectdata/model/dto/CollectDataTaskBatchRequest.java

@@ -12,4 +12,8 @@ public class CollectDataTaskBatchRequest {
     @NotEmpty
     @Schema(description = "任务 ID 列表", requiredMode = Schema.RequiredMode.REQUIRED)
     private List<Long> taskIds;
+
+    /** 当前用户 ID（可省略；传入时仅返回该用户的任务，用于进度接口归属过滤）。 */
+    @Schema(description = "当前用户 ID（可省略；传入时仅返回该用户的任务）", example = "1")
+    private Long userId;
 }

backend-java/src/main/java/com/nanri/aiimage/modules/collectdata/service/CollectDataItemCleanupSpiImpl.java

@@ -0,0 +1,29 @@
+package com.nanri.aiimage.modules.collectdata.service;
+
+import com.baomidou.mybatisplus.core.conditions.query.LambdaQueryWrapper;
+import com.nanri.aiimage.modules.collectdata.mapper.CollectDataItemMapper;
+import com.nanri.aiimage.modules.collectdata.model.entity.CollectDataItemEntity;
+import com.nanri.aiimage.modules.task.spi.CollectDataItemCleanupSpi;
+import lombok.RequiredArgsConstructor;
+import org.springframework.stereotype.Service;
+
+import java.util.List;
+
+/**
+ * 采集明细行的历史清理实现（2026-09 全维度审查 G5：逻辑从 task 模块收回本模块）。
+ */
+@Service
+@RequiredArgsConstructor
+public class CollectDataItemCleanupSpiImpl implements CollectDataItemCleanupSpi {
+
+    private final CollectDataItemMapper collectDataItemMapper;
+
+    @Override
+    public int deleteItemsByTaskIds(List<Long> taskIds) {
+        if (taskIds == null || taskIds.isEmpty()) {
+            return 0;
+        }
+        return collectDataItemMapper.delete(new LambdaQueryWrapper<CollectDataItemEntity>()
+                .in(CollectDataItemEntity::getTaskId, taskIds));
+    }
+}

backend-java/src/main/java/com/nanri/aiimage/modules/collectdata/service/CollectDataService.java

@@ -4,9 +4,11 @@ import cn.hutool.core.util.IdUtil;
 import cn.hutool.core.io.FileUtil;
 import cn.hutool.crypto.digest.DigestUtil;
 import com.baomidou.mybatisplus.core.conditions.query.LambdaQueryWrapper;
+import com.baomidou.mybatisplus.core.conditions.update.LambdaUpdateWrapper;
 import com.fasterxml.jackson.core.type.TypeReference;
 import com.fasterxml.jackson.databind.JsonNode;
 import com.fasterxml.jackson.databind.ObjectMapper;
+import com.nanri.aiimage.common.exception.BusinessCodes;
 import com.nanri.aiimage.common.exception.BusinessException;
 import com.nanri.aiimage.common.util.ExcelStreamReader;
 import com.nanri.aiimage.modules.collectdata.mapper.CollectDataCountryPrefMapper;
@@ -49,6 +51,7 @@ import com.nanri.aiimage.modules.task.mapper.TaskChunkMapper;
 import com.nanri.aiimage.modules.task.mapper.TaskResultItemMapper;
 import com.nanri.aiimage.modules.task.mapper.TaskScopeStateMapper;
 import com.nanri.aiimage.modules.task.model.dto.TaskHeartbeatRequest;
+import com.nanri.aiimage.modules.task.model.dto.TaskScopeLastChunkDto;
 import com.nanri.aiimage.modules.task.model.entity.FileResultEntity;
 import com.nanri.aiimage.modules.task.model.entity.FileTaskEntity;
 import com.nanri.aiimage.modules.task.model.entity.TaskChunkEntity;
@@ -104,8 +107,8 @@ public class CollectDataService {
      */
     private static final String LEGACY_MODULE_TYPE = "collectdata";
 
-    public TaskProgressLightBatchVo progressLight(List<Long> taskIds) {
+    public TaskProgressLightBatchVo progressLight(List<Long> taskIds, Long userId) {
-        return taskProgressLightAssembler.assemble(MODULE_TYPE, null, taskIds);
+        return taskProgressLightAssembler.assemble(MODULE_TYPE, userId, taskIds);
     }
     public static final int DEFAULT_PAGE_SIZE = 50;
 
@@ -177,6 +180,10 @@ public class CollectDataService {
     @Value("${aiimage.collect-data.stale-timeout-minutes:30}")
     private long staleTimeoutMinutes;
 
+    /** 二次判死线（心跳正常但连续 N 分钟无结果分片上报）阈值，分钟；<=0 关闭本线。 */
+    @Value("${aiimage.collect-data.no-result-upload-timeout-minutes:180}")
+    private long noResultUploadTimeoutMinutes;
+
     @Value("${aiimage.collect-data.max-source-file-bytes:0}")
     private Long maxSourceFileBytes;
 
@@ -379,42 +386,73 @@ public class CollectDataService {
     @Transactional
     public void activateTask(Long taskId, Long userId) {
         FileTaskEntity task = requireTask(taskId, userId);
-        if (STATUS_SUCCESS.equals(task.getStatus()) || STATUS_FAILED.equals(task.getStatus())) {
+        // 只允许 PENDING→RUNNING（条件更新）：既堵住 TOCTOU（/fail 抢先标 FAILED 后被整行
+        // updateById 复活成 RUNNING），又与客户端「兜底拉取」的原子认领互斥，谁先翻转谁执行
+        int updated = fileTaskMapper.update(null, new LambdaUpdateWrapper<FileTaskEntity>()
+                .eq(FileTaskEntity::getId, task.getId())
+                .eq(FileTaskEntity::getStatus, STATUS_PENDING)
+                .set(FileTaskEntity::getStatus, STATUS_RUNNING)
+                .set(FileTaskEntity::getUpdatedAt, LocalDateTime.now()));
+        if (updated == 0) {
+            FileTaskEntity latest = fileTaskMapper.selectById(task.getId());
+            if (latest != null && STATUS_RUNNING.equals(latest.getStatus())) {
+                throw new BusinessException("任务已在执行中（可能已由客户端自动接管），无需重复启动");
+            }
             throw new BusinessException("任务已结束");
         }
-        task.setStatus(STATUS_RUNNING);
-        task.setUpdatedAt(LocalDateTime.now());
-        fileTaskMapper.updateById(task);
     }
 
     @Transactional
     public void failTask(Long taskId, Long userId, String error) {
         FileTaskEntity task = requireTask(taskId, userId);
-        if (STATUS_SUCCESS.equals(task.getStatus()) || STATUS_FAILED.equals(task.getStatus())) {
-            return;
-        }
         String message = firstNonBlank(error, "collect-data task dispatch failed");
         FileResultEntity result = ensureTaskResult(task);
         CollectDataStats stats = loadStats(task);
+        boolean alreadyTerminal = STATUS_SUCCESS.equals(task.getStatus()) || STATUS_FAILED.equals(task.getStatus());
+        if (!alreadyTerminal) {
             result.setSuccess(0);
             result.setErrorMessage(message);
             result.setRowCount(stats.finalRowCount);
             fileResultMapper.updateById(result);
-        task.setStatus(STATUS_FAILED);
-        task.setErrorMessage(message);
-        task.setFailedFileCount(1);
-        task.setUpdatedAt(LocalDateTime.now());
-        task.setFinishedAt(LocalDateTime.now());
             persistStats(task, stats);
-        fileTaskMapper.updateById(task);
+        }
+        // 条件更新：客户端报错（/fail）与结果文件组装完成（processResultFileJob 写 SUCCESS）
+        // 可能并发 —— 无条件 updateById 会把已生成的 SUCCESS 覆盖成 FAILED（用户拿不到下载）或反之
+        int updated = fileTaskMapper.update(null, new LambdaUpdateWrapper<FileTaskEntity>()
+                .eq(FileTaskEntity::getId, task.getId())
+                .notIn(FileTaskEntity::getStatus, STATUS_SUCCESS, STATUS_FAILED)
+                .set(FileTaskEntity::getStatus, STATUS_FAILED)
+                .set(FileTaskEntity::getErrorMessage, message)
+                .set(FileTaskEntity::getFailedFileCount, 1)
+                .set(FileTaskEntity::getUpdatedAt, LocalDateTime.now())
+                .set(FileTaskEntity::getFinishedAt, LocalDateTime.now()));
+        if (updated == 0) {
+            log.info("[collect-data] failTask 跳过写入：任务已是终态 taskId={} status={}", taskId, task.getStatus());
+        }
     }
 
-    @Transactional
+    /**
+     * 进度心跳。
+     *
+     * <p>事务边界：Redis 任务锁在事务外获取（自旋等待最长 TASK_LOCK_WAIT_MILLIS，
+     * 放在 @Transactional 里会白占一个 Hikari 连接），DB 段（统计持久化 + 任务行更新）
+     * 仍在一个事务内。
+     */
     public void updateProgress(Long taskId, TaskHeartbeatRequest request) {
         if (taskId == null || taskId <= 0 || request == null) {
             return;
         }
         try (TaskDistributedLockService.LockHandle ignored = acquireTaskLockOrThrow(taskId)) {
+            if (transactionTemplate == null) {
+                // 单测场景（@InjectMocks 未注入事务模板）：退化为直接执行 DB 段
+                updateProgressLocked(taskId, request);
+                return;
+            }
+            transactionTemplate.executeWithoutResult(status -> updateProgressLocked(taskId, request));
+        }
+    }
+
+    private void updateProgressLocked(Long taskId, TaskHeartbeatRequest request) {
         FileTaskEntity task = fileTaskMapper.selectById(taskId);
         if (task == null || !MODULE_TYPE.equals(task.getModuleType()) || !STATUS_RUNNING.equals(task.getStatus())) {
             return;
@@ -471,7 +509,6 @@ public class CollectDataService {
         fileTaskMapper.updateById(task);
         lastProgressFlushMillis = System.currentTimeMillis();
     }
-    }
 
     /**
      * 节流判定：progressThrottleMillis>0 且距上次实际落库未超过窗口 → 合并写（跳过 UPDATE）。
@@ -509,6 +546,7 @@ public class CollectDataService {
             for (FileTaskEntity task : tasks) {
                 finalizeStaleTask(task.getId(), threshold, timeoutMinutes);
             }
+            finalizeNoUploadStaleTasks(threshold);
         }
     }
 
@@ -548,6 +586,86 @@ public class CollectDataService {
         }
     }
 
+    /**
+     * 二次判死：心跳正常但连续 N 分钟无结果分片上报（治「心跳续命」的假活任务）。
+     *
+     * <p>既有心跳线候选条件是 updated_at 陈旧，而客户端心跳会持续刷新 updated_at——
+     * 主线程卡死（浏览器自动化等待/异常）时心跳线程照发，任务永远命不中。
+     * 本线候选取「心跳新鲜（既有线放过）+ 创建超过 N 分钟」，判据用
+     * biz_task_scope_state.last_chunk_at（只随结果分片上报刷新）；
+     * 从未上报（无 scope 行或全 NULL）的任务跳过，避免误杀首批较慢的正常任务。
+     */
+    private void finalizeNoUploadStaleTasks(LocalDateTime heartbeatThreshold) {
+        long minutes = noResultUploadTimeoutMinutes;
+        if (minutes <= 0) {
+            return;
+        }
+        LocalDateTime cutoff = LocalDateTime.now().minusMinutes(minutes);
+        List<FileTaskEntity> candidates = fileTaskMapper.selectList(new LambdaQueryWrapper<FileTaskEntity>()
+                .eq(FileTaskEntity::getModuleType, MODULE_TYPE)
+                .eq(FileTaskEntity::getStatus, STATUS_RUNNING)
+                .ge(FileTaskEntity::getUpdatedAt, heartbeatThreshold)
+                .lt(FileTaskEntity::getCreatedAt, cutoff)
+                .orderByAsc(FileTaskEntity::getCreatedAt)
+                .last("limit 200"));
+        if (candidates.isEmpty()) {
+            return;
+        }
+        Map<Long, LocalDateTime> lastResultAtByTaskId = new HashMap<>();
+        for (TaskScopeLastChunkDto dto : taskScopeStateMapper.selectLastChunkAtByTaskIds(
+                candidates.stream().map(FileTaskEntity::getId).toList())) {
+            if (dto.taskId() != null && dto.lastChunkAt() != null) {
+                lastResultAtByTaskId.put(dto.taskId(), dto.lastChunkAt());
+            }
+        }
+        for (FileTaskEntity task : candidates) {
+            LocalDateTime lastResultAt = lastResultAtByTaskId.get(task.getId());
+            if (lastResultAt == null || lastResultAt.isAfter(cutoff)) {
+                continue;
+            }
+            finalizeNoUploadStaleTask(task.getId(), lastResultAt, minutes);
+        }
+    }
+
+    /** 单个任务的二次判死收尾：锁内复查心跳后，有分片→组装部分工作簿；无分片→标失败。 */
+    private void finalizeNoUploadStaleTask(Long taskId, LocalDateTime lastResultAt, long minutes) {
+        try (TaskDistributedLockService.LockHandle lock =
+                     taskDistributedLockService.acquire(MODULE_TYPE, taskId, 0L)) {
+            if (lock == null) {
+                return;
+            }
+            FileTaskEntity task = fileTaskMapper.selectById(taskId);
+            LocalDateTime heartbeatFreshAfter = LocalDateTime.now().minusMinutes(Math.max(1L, staleTimeoutMinutes));
+            if (task == null
+                    || !MODULE_TYPE.equals(task.getModuleType())
+                    || !STATUS_RUNNING.equals(task.getStatus())
+                    || task.getUpdatedAt() == null
+                    || task.getUpdatedAt().isBefore(heartbeatFreshAfter)) {
+                // 已终态、或心跳在排队期间回落到陈旧（交回既有心跳线处理）
+                return;
+            }
+            if (taskFileJobService.countUnfinishedAssembleJobs(taskId, MODULE_TYPE) > 0L) {
+                return;
+            }
+            FileResultEntity result = ensureTaskResult(task);
+            CollectDataStats stats = loadStats(task);
+            if (hasReceivedChunks(taskId)) {
+                enqueueFinalWorkbook(task, result, stats);
+                log.warn("[collect-data] no-upload stale task enqueued partial workbook taskId={} lastResultAt={} timeoutMinutes={} finalRows={}",
+                        taskId, lastResultAt, minutes, stats.finalRowCount);
+                return;
+            }
+            markTaskFailed(task, result,
+                    "连续 " + minutes + " 分钟无结果回传，疑似客户端任务卡死，任务已自动失败（最后结果上报 " + lastResultAt + "）",
+                    stats);
+            log.warn("[collect-data] no-upload stale task failed without chunks taskId={} lastResultAt={} timeoutMinutes={}",
+                    taskId, lastResultAt, minutes);
+        } catch (Exception ex) {
+            log.warn("[collect-data] no-upload stale task finalization failed taskId={} msg={}",
+                    taskId, ex.getMessage(), ex);
+        }
+    }
+
     public CollectDataDashboardVo dashboard(Long userId) {
         CollectDataDashboardVo vo = new CollectDataDashboardVo();
         vo.setPendingTaskCount(countActiveTasks(userId));
@@ -664,6 +782,39 @@ public class CollectDataService {
             throw new BusinessException("request is empty");
         }
         ensureRustfsPayloadStorageEnabled();
+
+        // 锁外预检：任务不存在/已结束时立即失败，不为终态任务白跑去重查询与品牌检测。
+        // 只做快速失败，并发正确性仍由锁内的重读复核保证。
+        FileTaskEntity probe = fileTaskMapper.selectById(taskId);
+        if (probe == null || !MODULE_TYPE.equals(probe.getModuleType())) {
+            throw new BusinessException("任务不存在");
+        }
+        if (STATUS_SUCCESS.equals(probe.getStatus()) || STATUS_FAILED.equals(probe.getStatus())) {
+            log.warn("[collect-data] 任务已结束，拒绝重复提交 taskId={} status={}", taskId, probe.getStatus());
+            throw new BusinessException(40901, "任务已结束，拒绝重复提交");
+        }
+
+        // 归一化 / 去重过滤 / 品牌检测放在锁外：品牌检测是同步远程调用，上游 16890 抖动时
+        // 单品牌 10 次重试合计上百秒（taskId 28599 实测：chunk 回传在锁内等品牌检测 103.5 秒，
+        // 期间心跳与客户端重试全部撞 40902 拿不到锁，客户端 5 次重试预算耗尽后中止整个采集）。
+        // 这几步只依赖本批入参、不写任务状态，放锁外不改变 chunk 落库的串行语义。
+        List<CollectDataResultRowVo> rows = normalizeSubmitRows(request.getItems());
+        buildParseLimits().validateChunkRowCount(rows.size());
+        // 「结果文件」sheet 需要 Python 回传的全量数据（不经任何后端过滤），
+        // 因此 chunk payload 写入未筛除空 ASIN 的全量 rows；
+        // 仅当 ASIN 非空时才进入去重 / 无效品牌 / 品牌检测链路并落 biz_task_result_item。
+        List<CollectDataResultRowVo> rowsForFiltering = new ArrayList<>(rows.size());
+        for (CollectDataResultRowVo row : rows) {
+            if (row.getAsin() != null && !row.getAsin().isBlank()) {
+                rowsForFiltering.add(row);
+            }
+        }
+        long prepareStartAt = System.currentTimeMillis();
+        CollectDataBatchQuery.FilterResult filtered = collectDataBatchQuery.filter(rowsForFiltering);
+        CollectDataBrandBatchFilter.BrandBatchOutcome brandOutcome = brandBatchFilter.filter(filtered.kept());
+        log.info("[collect-data] 锁外预处理完成 taskId={} rows={} 去重后={} 品牌检测耗时={}ms",
+                taskId, rows.size(), filtered.kept().size(), System.currentTimeMillis() - prepareStartAt);
+
         try (TaskDistributedLockService.LockHandle ignored = acquireTaskLockOrThrow(taskId)) {
             FileTaskEntity task = fileTaskMapper.selectById(taskId);
             if (task == null || !MODULE_TYPE.equals(task.getModuleType())) {
@@ -701,25 +852,23 @@ public class CollectDataService {
                 return buildSubmitVo(task, result, chunkIndex, chunkTotal, request, 0);
             }
 
-            List<CollectDataResultRowVo> rows = normalizeSubmitRows(request.getItems());
-            buildParseLimits().validateChunkRowCount(rows.size());
             CollectDataStats stats = loadStats(task);
             stats.receivedRows += rows.size();
             stats.currentChunkRows = rows.size();
-
-            // 「结果文件」sheet 需要 Python 回传的全量数据（不经任何后端过滤），
-            // 因此 chunk payload 写入未筛除空 ASIN 的全量 rows；
-            // 仅当 ASIN 非空时才进入去重 / 无效品牌 / 品牌检测链路并落 biz_task_result_item。
-            List<CollectDataResultRowVo> rowsForFiltering = new ArrayList<>(rows.size());
-            for (CollectDataResultRowVo row : rows) {
-                if (row.getAsin() != null && !row.getAsin().isBlank()) {
-                    rowsForFiltering.add(row);
-                }
-            }
-            CollectDataBatchQuery.FilterResult filtered = collectDataBatchQuery.filter(rowsForFiltering);
             stats.dedupeFilteredCount += filtered.dedupeFilteredCount();
             stats.invalidFilteredCount += filtered.invalidFilteredCount();
-            List<CollectDataResultRowVo> accepted = filterByBrandCheck(filtered.kept(), stats);
+            stats.brandRejectedCount += brandOutcome.rejected().size();
+            stats.brandQueryFailedCount += brandOutcome.queryFailed().size();
+            // 只有被服务端判定为无效品牌（rejected）的行才写入无效品牌表；
+            // queryFailed 是品牌检测服务端到端失败（如缺 X-Token 422/超时/限流），
+            // 一并写入会把故障期间被误伤的品牌永久拉黑，后续同品牌商品全部被
+            // invalidFiltered 过滤（task-27265 实测：通过 Python 过滤的 21 行中
+            // 8 行查询失败被写表后误杀，最终结果 Excel 0 行）。rejected 为空时
+            // 不触发任何写入，避免空批次无意义调用。
+            if (!brandOutcome.rejected().isEmpty()) {
+                invalidAsinBatchWriter.writeBatch(brandOutcome.rejected());
+            }
+            List<CollectDataResultRowVo> accepted = brandOutcome.accepted();
             // 结果明细改为 chunk 级存储：整个 chunk 的 accepted 行共享一个
             // RustFS 对象（deterministic key，同 chunk 重提覆盖同一对象），
             // biz_task_result_item.payload_json 只存 {chunk, offset, payload} 引用。
@@ -754,6 +903,14 @@ public class CollectDataService {
 
             if (request.getError() != null && !request.getError().isBlank()) {
                 markTaskFailed(task, result, request.getError(), stats);
+                // 失败但已收到分片：照常组装结果文件，让用户能下载已采集的数据。
+                // 此前失败分支只标失败不组装，已落库的数据也没有任何结果文件可下载
+                // （taskId 28599：55 个分片全部收到、187 行明细已落库，用户却拿不到文件）。
+                if (hasReceivedChunks(taskId)) {
+                    enqueueFinalWorkbook(task, result, stats);
+                    log.warn("[collect-data] 任务失败仍组装部分结果 taskId={} error={} finalRows={}",
+                            taskId, request.getError(), stats.finalRowCount);
+                }
             } else if (Boolean.TRUE.equals(request.getDone())) {
                 enqueueFinalWorkbook(task, result, stats);
             } else {
@@ -815,22 +972,6 @@ public class CollectDataService {
         return rows;
     }
 
-    private List<CollectDataResultRowVo> filterByBrandCheck(List<CollectDataResultRowVo> rows, CollectDataStats stats) {
-        CollectDataBrandBatchFilter.BrandBatchOutcome outcome = brandBatchFilter.filter(rows);
-        stats.brandRejectedCount += outcome.rejected().size();
-        stats.brandQueryFailedCount += outcome.queryFailed().size();
-        // 只有被服务端判定为无效品牌（rejected）的行才写入无效品牌表；
-        // queryFailed 是品牌检测服务端到端失败（如缺 X-Token 422/超时/限流），
-        // 一并写入会把故障期间被误伤的品牌永久拉黑，后续同品牌商品全部被
-        // invalidFiltered 过滤（task-27265 实测：通过 Python 过滤的 21 行中
-        // 8 行查询失败被写表后误杀，最终结果 Excel 0 行）。rejected 为空时
-        // 不触发任何写入，避免空批次无意义调用。
-        if (!outcome.rejected().isEmpty()) {
-            invalidAsinBatchWriter.writeBatch(outcome.rejected());
-        }
-        return outcome.accepted();
-    }
-
     private void persistChunk(Long taskId,
                               String scopeKey,
                               String scopeHash,
@@ -893,7 +1034,8 @@ public class CollectDataService {
         result.setResultFileSize(0L);
         result.setResultContentType(CONTENT_TYPE_XLSX);
         result.setRowCount(stats.finalRowCount);
-        result.setErrorMessage(null);
+        // 不清 errorMessage：失败任务的部分结果组装也走这里，清掉会让用户看不到真实失败原因
+        // （成功路径的 errorMessage 本来就为 null，无需清理）。
         fileResultMapper.updateById(result);
         task.setUpdatedAt(LocalDateTime.now());
         fileTaskMapper.updateById(task);
@@ -942,6 +1084,22 @@ public class CollectDataService {
                     stats.summaries,
                     batch -> streamRawRows(task.getId(), batch));
             String objectKey = ossStorageService.uploadResultFile(xlsx, MODULE_TYPE);
+
+            // 复用同一份 stats 更新 finalRowCount 后再持久化，避免重复 loadStats 丢失 summaries。
+            stats.finalRowCount = (int) finalRowCount;
+            persistStats(task, stats);
+
+            // 失败原因先留存：下面的乐观写入会清空 result.errorMessage，任务已被判失败时要用它恢复。
+            String failureReason = result.getErrorMessage();
+            if (failureReason == null || failureReason.isBlank()) {
+                failureReason = task.getErrorMessage();
+            }
+            if (failureReason == null || failureReason.isBlank()) {
+                failureReason = "任务失败，结果文件为已采集的部分数据";
+            }
+
+            // 结果行先按成功乐观写入：保持「结果文件先于任务成功落库」的时序，
+            // 万一进程在这两步之间退出，任务仍是 RUNNING，会被陈旧巡检重新组装（可自愈）。
             result.setResultFilename(filename);
             result.setResultFileUrl(objectKey);
             result.setResultFileSize(xlsx.length());
@@ -951,16 +1109,27 @@ public class CollectDataService {
             result.setErrorMessage(null);
             fileResultMapper.updateById(result);
 
-            // 复用同一份 stats 更新 finalRowCount 后再持久化，避免重复 loadStats 丢失 summaries。
+            // 条件更新：任务可能已被判失败（客户端上报失败 / 陈旧判死与结果文件组装并发）——
-            stats.finalRowCount = (int) finalRowCount;
+            // 无条件 updateById 会把 FAILED 覆盖成 SUCCESS（错误信息被清、用户看到「假成功」）。
-            persistStats(task, stats);
+            // 结果文件已生成，故失败态下仍保留文件，只是不覆盖状态。
-            task.setStatus(STATUS_SUCCESS);
+            int updated = fileTaskMapper.update(null, new LambdaUpdateWrapper<FileTaskEntity>()
-            task.setSuccessFileCount(1);
+                    .eq(FileTaskEntity::getId, task.getId())
-            task.setFailedFileCount(0);
+                    .ne(FileTaskEntity::getStatus, STATUS_FAILED)
-            task.setErrorMessage(null);
+                    .set(FileTaskEntity::getStatus, STATUS_SUCCESS)
-            task.setUpdatedAt(LocalDateTime.now());
+                    .set(FileTaskEntity::getSuccessFileCount, 1)
-            task.setFinishedAt(LocalDateTime.now());
+                    .set(FileTaskEntity::getFailedFileCount, 0)
-            fileTaskMapper.updateById(task);
+                    .set(FileTaskEntity::getErrorMessage, null)
+                    .set(FileTaskEntity::getUpdatedAt, LocalDateTime.now())
+                    .set(FileTaskEntity::getFinishedAt, LocalDateTime.now()));
+            if (updated == 0) {
+                // 任务已是 FAILED：结果记录改回失败语义并保留真实原因，但文件 URL 照常保留，
+                // 用户看到「失败 + 原因」的同时仍能下载已采集的部分结果。
+                result.setSuccess(0);
+                result.setErrorMessage(failureReason);
+                fileResultMapper.updateById(result);
+                log.warn("[collect-data] 结果文件已生成但任务已是 FAILED，保留失败态与原因 taskId={} rows={} reason={}",
+                        task.getId(), finalRowCount, failureReason);
+            }
         } finally {
             FileUtil.del(xlsx);
         }
@@ -1085,7 +1254,8 @@ public class CollectDataService {
     private TaskDistributedLockService.LockHandle acquireTaskLockOrThrow(Long taskId) {
         TaskDistributedLockService.LockHandle lockHandle = taskDistributedLockService.acquire(MODULE_TYPE, taskId, TASK_LOCK_WAIT_MILLIS);
         if (lockHandle == null) {
-            throw new BusinessException(40901, "任务正在处理，请稍后再试");
+            log.warn("[collect-data] 任务锁竞争，拒绝本次提交 taskId={} waitMillis={}", taskId, TASK_LOCK_WAIT_MILLIS);
+            throw new BusinessException(BusinessCodes.TASK_BUSY, "任务正在处理，请稍后再试");
         }
         return lockHandle;
     }

backend-java/src/main/java/com/nanri/aiimage/modules/collectdata/service/CollectDataTaskHeartbeatSpi.java

@@ -0,0 +1,28 @@
+package com.nanri.aiimage.modules.collectdata.service;
+
+import com.nanri.aiimage.modules.task.model.dto.TaskHeartbeatRequest;
+import com.nanri.aiimage.modules.task.spi.TaskModuleHeartbeatSpi;
+import lombok.RequiredArgsConstructor;
+import org.springframework.stereotype.Service;
+
+/**
+ * 采集数据的任务心跳实现（2026-09 全维度审查 G5）。
+ *
+ * <p>进度由采集模块自己的导入进度表维护，心跳直接把客户端上报的进度写进去。
+ */
+@Service
+@RequiredArgsConstructor
+public class CollectDataTaskHeartbeatSpi implements TaskModuleHeartbeatSpi {
+
+    private final CollectDataService collectDataService;
+
+    @Override
+    public String moduleType() {
+        return "COLLECT_DATA";
+    }
+
+    @Override
+    public void touchHeartbeat(Long taskId, TaskHeartbeatRequest request) {
+        collectDataService.updateProgress(taskId, request);
+    }
+}

backend-java/src/main/java/com/nanri/aiimage/modules/collectdata/service/CollectDataTaskPullSpiImpl.java

@@ -0,0 +1,71 @@
+package com.nanri.aiimage.modules.collectdata.service;
+
+import com.fasterxml.jackson.databind.JsonNode;
+import com.fasterxml.jackson.databind.ObjectMapper;
+import com.nanri.aiimage.modules.task.model.entity.FileTaskEntity;
+import com.nanri.aiimage.modules.task.spi.ClientTaskPullSpi;
+import lombok.RequiredArgsConstructor;
+import lombok.extern.slf4j.Slf4j;
+import org.springframework.stereotype.Service;
+
+import java.util.LinkedHashMap;
+import java.util.Map;
+
+/**
+ * 集采（collect-data）的客户端兜底拉取实现。
+ *
+ * <p>Python 消费端需要 taskId / totalRows / pageSize / filters（明细行自行按 /items 分页拉取）。
+ * filters 取自任务行 request_json 里 parse 时落库的那份（CollectDataService#persistParsedTask），
+ * 序列化后就是 Python 读取的 camelCase 键（countryCode/minAmount/...）。
+ */
+@Slf4j
+@Service
+@RequiredArgsConstructor
+public class CollectDataTaskPullSpiImpl implements ClientTaskPullSpi {
+
+    private static final String QUEUE_TYPE = "collect-data-run";
+    private static final String TASK_TYPE = "collect-data";
+
+    private final ObjectMapper objectMapper;
+
+    @Override
+    public String moduleType() {
+        return CollectDataService.MODULE_TYPE;
+    }
+
+    @Override
+    public Map<String, Object> buildQueuePayload(FileTaskEntity task) {
+        JsonNode request = parseJson(task.getRequestJson());
+        if (request == null) {
+            log.warn("[collect-data] 兜底拉取失败：任务请求参数缺失或不可解析 taskId={}", task.getId());
+            return null;
+        }
+        JsonNode filtersNode = request.get("filters");
+        Map<String, Object> filters = filtersNode == null || filtersNode.isNull()
+                ? Map.of()
+                : objectMapper.convertValue(filtersNode, Map.class);
+        JsonNode stats = parseJson(task.getResultJson());
+        int totalRows = stats == null ? 0 : stats.path("totalRows").asInt(0);
+        Map<String, Object> data = new LinkedHashMap<>();
+        data.put("taskId", task.getId());
+        data.put("taskNo", task.getTaskNo());
+        data.put("taskType", TASK_TYPE);
+        data.put("totalRows", totalRows);
+        data.put("pageSize", CollectDataService.DEFAULT_PAGE_SIZE);
+        data.put("filters", filters);
+        log.info("[collect-data] 兜底载荷已组装 taskId={} totalRows={} filters={}", task.getId(), totalRows, filters);
+        return Map.of("type", QUEUE_TYPE, "data", data);
+    }
+
+    private JsonNode parseJson(String json) {
+        if (json == null || json.isBlank()) {
+            return null;
+        }
+        try {
+            return objectMapper.readTree(json);
+        } catch (Exception ex) {
+            log.warn("[collect-data] 兜底拉取解析任务 JSON 失败 err={}", ex.getMessage());
+            return null;
+        }
+    }
+}

backend-java/src/main/java/com/nanri/aiimage/modules/collectdata/util/CollectDataResultDetailCodec.java

@@ -148,11 +148,8 @@ public class CollectDataResultDetailCodec {
         try {
             MessageDigest digest = MessageDigest.getInstance("SHA-256");
             byte[] bytes = digest.digest((value == null ? "" : value).getBytes(StandardCharsets.UTF_8));
-            StringBuilder sb = new StringBuilder(bytes.length * 2);
+            // HexFormat 替代逐字节 String.format("%02x")：后者每次走 32 次 Formatter（热路径按行调用）
-            for (byte b : bytes) {
+            return java.util.HexFormat.of().formatHex(bytes);
-                sb.append(String.format("%02x", b));
-            }
-            return sb.toString();
         } catch (Exception ex) {
             throw new IllegalStateException("chunk detail ref hash failed", ex);
         }

backend-java/src/main/java/com/nanri/aiimage/modules/collectdata/util/CollectDataResultItemBatchWriter.java

@@ -1,6 +1,7 @@
 package com.nanri.aiimage.modules.collectdata.util;
 
 import com.baomidou.mybatisplus.core.conditions.query.LambdaQueryWrapper;
+import com.nanri.aiimage.common.exception.BusinessException;
 import com.nanri.aiimage.modules.collectdata.model.vo.CollectDataResultRowVo;
 import com.nanri.aiimage.modules.task.mapper.TaskResultItemMapper;
 import com.nanri.aiimage.modules.task.model.entity.TaskResultItemEntity;
@@ -133,12 +134,14 @@ public class CollectDataResultItemBatchWriter {
         }
 
         int written = 0;
+        int failedBatches = 0;
         for (int from = 0; from < toUpsert.size(); from += batchSize) {
             int to = Math.min(from + batchSize, toUpsert.size());
             List<TaskResultItemEntity> batch = toUpsert.subList(from, to);
             try {
                 written += taskResultItemMapper.upsertBatch(batch);
             } catch (RuntimeException ex) {
+                failedBatches++;
                 log.warn("[collect-data] upsert result item batch failed, skip batch {}..{} taskId={}",
                         from, to, taskId, ex);
                 // 失败批的新行未落库，从增量计数中扣除，避免任务内累计虚高；
@@ -150,6 +153,16 @@ public class CollectDataResultItemBatchWriter {
                 }
             }
         }
+        if (failedBatches > 0) {
+            // biz_task_result_item 是结果 Excel「明细」sheet 的唯一数据源：静默跳批会让任务
+            // 以 SUCCESS 收尾但明细缺行，且与「结果文件」sheet 的汇总数量对不上（假成功）。
+            // 抛出让本次 chunk 提交明确失败：worker（search_spider）识别 success=false 后会重试，
+            // 重提按 payload_hash 幂等（已落库行跳过、未落库行补插），最终收敛为完整数据。
+            log.error("[collect-data] 明细写入存在失败批次，拒绝本次提交 taskId={} failedBatches={} totalBatches={}",
+                    taskId, failedBatches, (toUpsert.size() + batchSize - 1) / batchSize);
+            throw new BusinessException("采集结果明细写入失败，请稍后重试（失败批次 "
+                    + failedBatches + "/" + ((toUpsert.size() + batchSize - 1) / batchSize) + "）");
+        }
         return new UpsertCounts(written, skipped, newlyInserted);
     }
 
@@ -157,11 +170,8 @@ public class CollectDataResultItemBatchWriter {
         try {
             java.security.MessageDigest digest = java.security.MessageDigest.getInstance("SHA-256");
             byte[] bytes = digest.digest((value == null ? "" : value).getBytes(java.nio.charset.StandardCharsets.UTF_8));
-            StringBuilder sb = new StringBuilder(bytes.length * 2);
+            // HexFormat 替代逐字节 String.format("%02x")：后者每次走 32 次 Formatter（热路径按行调用）
-            for (byte b : bytes) {
+            return java.util.HexFormat.of().formatHex(bytes);
-                sb.append(String.format("%02x", b));
-            }
-            return sb.toString();
         } catch (Exception ex) {
             throw new IllegalStateException("结果明细 hash 计算失败", ex);
         }

backend-java/src/main/java/com/nanri/aiimage/modules/convert/service/ConvertRunService.java

@@ -278,12 +278,14 @@ public class ConvertRunService {
         File outputDir = FileUtil.mkdir(FileUtil.file(storageProperties.getLocalTempDir(), "convert-result"));
         List<GeneratedConvertFile> generatedFiles = new ArrayList<>();
         Map<String, BufferedWriter> writers = new LinkedHashMap<>();
+        try {
+            // 创建循环纳入 try：第 N 个 writer 创建失败时，前 N-1 个已打开的句柄会因 finally
+            // 尚未生效而泄漏（同族 SplitRunService.SplitChunkWriter 已用 try/finally 处理）
             for (String outputFilename : outputFilenames) {
                 File outputFile = buildNamedOutputFile(outputDir, outputFilename);
                 generatedFiles.add(new GeneratedConvertFile(outputFilename, outputFile));
                 writers.put(outputFilename, Files.newBufferedWriter(outputFile.toPath(), StandardCharsets.UTF_8));
             }
-        try {
             streamTxtRowsToOutputs(inputFile, templateEntity, writers);
         } finally {
             IOException closeException = null;

backend-java/src/main/java/com/nanri/aiimage/modules/dedupe/controller/DedupeTotalDataController.java

@@ -2,7 +2,7 @@ package com.nanri.aiimage.modules.dedupe.controller;
 
 import com.nanri.aiimage.common.api.ApiResponse;
 import com.nanri.aiimage.common.exception.BusinessException;
-import com.nanri.aiimage.modules.admin.support.AdminAuthSupport;
+import com.nanri.aiimage.common.security.AdminAuthSupport;
 import com.nanri.aiimage.common.util.DownloadHeaderUtil;
 import com.nanri.aiimage.modules.dedupe.model.dto.DedupeTotalDataCreateRequest;
 import com.nanri.aiimage.modules.dedupe.model.dto.DedupeTotalDataUpdateRequest;
@@ -12,7 +12,7 @@ import com.nanri.aiimage.modules.dedupe.model.vo.DedupeTotalDataImportVo;
 import com.nanri.aiimage.modules.dedupe.model.vo.DedupeTotalDataItemVo;
 import com.nanri.aiimage.modules.dedupe.model.vo.DedupeTotalDataPageVo;
 import com.nanri.aiimage.modules.dedupe.service.DedupeTotalDataService;
-import com.nanri.aiimage.modules.permission.model.entity.AdminUserEntity;
+import com.nanri.aiimage.common.model.entity.AdminUserEntity;
 import com.nanri.aiimage.modules.permission.service.PermissionMenuService;
 import io.swagger.v3.oas.annotations.Operation;
 import io.swagger.v3.oas.annotations.Parameter;
@@ -88,10 +88,12 @@ public class DedupeTotalDataController {
             @RequestParam(name = "end_date", required = false) @DateTimeFormat(iso = DateTimeFormat.ISO.DATE) LocalDate endDate,
             @Parameter(description = "分组 ID") @RequestParam(name = "group_id", required = false) Long groupId,
             @Parameter(description = "国家代码（如 DE、UK）") @RequestParam(name = "country", required = false) String country,
+            @Parameter(description = "顺序翻页游标（上一页返回的 nextLastId；传了就忽略 page 偏移）")
+            @RequestParam(name = "last_id", required = false) Long lastId,
             HttpServletRequest request) {
         RequestOperator operator = requireDedupeTotalDataAccess(request);
         return ApiResponse.success(dedupeTotalDataService.page(
-                page, pageSize, keyword, username, startDate, endDate, groupId, country, operator.id()));
+                page, pageSize, keyword, username, startDate, endDate, groupId, country, lastId, operator.id()));
     }
 
     @GetMapping("/export")