fix(gpu): #1970 MuseTalk worker 推理期心跳/超时900/重试收敛/短视频前置失败 (#1990)

Co-authored-by: xiaoxia <dev@xiaoxiajianji.com>
Co-committed-by: xiaoxia <dev@xiaoxiajianji.com>

.env.example

@@ -217,6 +217,7 @@ APIZERO_API_KEY=
 # GPU Worker 长期鉴权 Token，Worker 端 .env 的 GPU_WORKER_TOKEN 必须与此一致
 # 留空时 development 环境允许匿名访问（仅本地调试），staging/production 必须配置
 GPU_WORKER_TOKEN=
-# 单任务超时（秒），超过则回退 pending 或标记 failed
-GPU_TASK_TIMEOUT_SECONDS=300
+# 单任务超时（秒），processing 超过此时长无任务心跳才回退 pending 或标记 failed
+# #1970：RTX2060 6G 推理 720p 长视频需 5 分钟以上，默认 900
+GPU_TASK_TIMEOUT_SECONDS=900
 

.gitea/workflows/ci-pipeline.yml

@@ -1186,6 +1186,7 @@ jobs:
         DOUBAO_API_KEY: "${{ secrets.DOUBAO_API_KEY }}"
         DOUBAO_MODEL: "${{ secrets.DOUBAO_MODEL }}"
         DOUBAO_BASE_URL: "${{ secrets.DOUBAO_BASE_URL }}"
+        DOUBAO_VISION_MODEL: "${{ secrets.DOUBAO_VISION_MODEL }}"
         WECHAT_APP_ID: "${{ secrets.WECHAT_APP_ID }}"
         WECHAT_APP_SECRET: "${{ secrets.WECHAT_APP_SECRET }}"
         TIKHUB_API_KEY: "${{ secrets.TIKHUB_API_KEY }}"
@@ -1641,6 +1642,7 @@ jobs:
         DOUBAO_API_KEY: "${{ secrets.DOUBAO_API_KEY }}"
         DOUBAO_MODEL: "${{ secrets.DOUBAO_MODEL }}"
         DOUBAO_BASE_URL: "${{ secrets.DOUBAO_BASE_URL }}"
+        DOUBAO_VISION_MODEL: "${{ secrets.DOUBAO_VISION_MODEL }}"
         WECHAT_APP_ID: "${{ secrets.WECHAT_APP_ID }}"
         WECHAT_APP_SECRET: "${{ secrets.WECHAT_APP_SECRET }}"
         TIKHUB_API_KEY: "${{ secrets.TIKHUB_API_KEY }}"

alembic/versions/082_atom_clip_ai_tags.py

@@ -1,7 +1,7 @@
 """add ai_tags to asset_atom_clips for #1970 fragment-level AI tagging
 
-Revision ID: 081_atom_clip_ai_tags
-Revises: 080_edit_plan_clips_atom_clip_id
+Revision ID: 082_atom_clip_ai_tags
+Revises: 081_add_gpu_lipsync
 Create Date: 2026-09-18
 """
 
@@ -9,8 +9,8 @@ import sqlalchemy as sa
 
 from alembic import op
 
-revision = "081_atom_clip_ai_tags"
-down_revision = "080_edit_plan_clips_atom_clip_id"
+revision = "082_atom_clip_ai_tags"
+down_revision = "081_add_gpu_lipsync"
 branch_labels = None
 depends_on = None
 

apps/api/app/api/routes/gpu_lipsync.py

@@ -99,6 +99,7 @@ def register_worker(
         gpu_name=body.gpu_name,
         free_vram_mb=body.free_vram_mb,
         capabilities=body.capabilities,
+        task_id=body.task_id,
     )
     return GpuWorkerRegisterResponse(ok=True, server_time=datetime.now(UTC), message="ok")
 

apps/api/app/schemas/gpu_lipsync.py

@@ -22,6 +22,14 @@ class GpuWorkerRegisterRequest(BaseModel):
     gpu_name: str = Field("", max_length=200, description="GPU 型号，如 'NVIDIA GeForce RTX 2060'")
     free_vram_mb: int = Field(0, ge=0, description="当前空闲显存（MB）")
     capabilities: str = Field("musetalk", max_length=500, description="能力列表，逗号分隔，如 'musetalk'")
+    task_id: Optional[str] = Field(
+        None,
+        max_length=64,
+        description=(
+            "当前正在处理的任务 ID。Worker 推理期间定期心跳时携带，"
+            "服务端同步刷新该任务 last_heartbeat_at，防止长推理被误判超时；空闲时不传"
+        ),
+    )
 
 
 class GpuWorkerRegisterResponse(BaseModel):

apps/api/app/services/gpu_lipsync_service.py

@@ -49,7 +49,15 @@ class GpuLipsyncService:
         gpu_name: str = "",
         free_vram_mb: int = 0,
         capabilities: str = "musetalk",
+        task_id: Optional[str] = None,
     ) -> GpuWorkerModel:
+        """Worker 注册/心跳。
+
+        task_id 非空时（Worker 推理期间的任务级心跳），同步把对应 processing
+        任务的 last_heartbeat_at 续到当前时间，使长推理不会被
+        ``_recover_timed_out_tasks`` 误回退。任务已结束 / 不属于该 worker
+        （如已被超时回收重新派发）时忽略，不报错。
+        """
         now = datetime.now(UTC)
         worker = self.db.query(GpuWorkerModel).filter(GpuWorkerModel.worker_id == worker_id).one_or_none()
         if worker is None:
@@ -69,6 +77,8 @@ class GpuLipsyncService:
             worker.free_vram_mb = free_vram_mb
             worker.capabilities = capabilities or worker.capabilities
             worker.last_heartbeat_at = now
+        if task_id:
+            self._touch_task_heartbeat(task_id, worker_id, now)
         self.db.commit()
         return worker
 
@@ -78,8 +88,10 @@ class GpuLipsyncService:
         """原子地认领一条最早的 pending 任务，返回给 worker；无任务返回 None.
 
         同时会：
-        - 把 processing 状态且超时（超过 gpu_task_timeout_seconds 无心跳）的任务
-          回退为 pending（attempt++，超过 MAX_ATTEMPTS 置 failed），让其它 worker 认领。
+        - 把 processing 状态且真正超时（任务心跳停滞超过
+          gpu_task_timeout_seconds；Worker 推理期会通过 register(task_id=...)
+          续心跳，长推理不会误判）的任务回退为 pending（attempt++，超过
+          MAX_ATTEMPTS 置 failed），让其它 worker 认领。
         - 刷新 worker 心跳。
         """
         now = datetime.now(UTC)
@@ -238,6 +250,28 @@ class GpuLipsyncService:
     def _result_key(self, task_id: str) -> str:
         return f"{self.RESULT_PREFIX}{task_id}.mp4"
 
+    def _touch_task_heartbeat(self, task_id: str, worker_id: str, now: datetime) -> None:
+        """Worker 推理期间的任务级心跳：只刷新属于该 worker 且仍在 processing 的任务。
+
+        任务不存在 / 已被超时回收重新派发 / 已完成 → 静默忽略（此时旧 worker 的
+        结果上报会被结果接口按最终态处理）。
+        """
+        task = self.db.get(GpuLipsyncTaskModel, task_id)
+        if task is None:
+            return
+        if task.status != "processing" or task.worker_id != worker_id:
+            logger.info(
+                "忽略过期任务心跳 task=%s worker=%s（status=%s owner=%s）",
+                task_id,
+                worker_id,
+                task.status,
+                task.worker_id,
+            )
+            return
+        task.last_heartbeat_at = now
+        task.updated_at = now
+        self.db.flush()
+
     def _touch_worker(self, worker_id: str, now: datetime) -> None:
         if not worker_id:
             return
@@ -260,7 +294,13 @@ class GpuLipsyncService:
             self.db.flush()
 
     def _recover_timed_out_tasks(self, now: datetime) -> None:
-        """扫描 processing 状态且超时（无心跳）的任务，回退 pending 或失败."""
+        """扫描 processing 状态且真正超时的任务，回退 pending 或失败。
+
+        判定只看任务自身 last_heartbeat_at：claim 时写入，Worker 推理期间通过
+        /gpu/register(task_id=...) 每 30s 续期。因此仅在 Worker 崩溃/断网
+        （任务心跳停滞超过 gpu_task_timeout_seconds）时才回收，
+        不会因 Worker 主循环忙于推理而误回退。
+        """
         timeout = self.settings.gpu_task_timeout_seconds
         cutoff = now - timedelta(seconds=timeout)
         stuck_tasks = (

apps/web/e2e/core-douyin-extract.spec.ts

@@ -0,0 +1,117 @@
+import { expect, test, type APIRequestContext, type Page } from "@playwright/test"
+
+const PASSWORD = "SmokePass123!"
+const apiBase = process.env.E2E_API_BASE || "/api/v1"
+const apiOrigin = apiBase.endsWith("/api/v1") ? apiBase.slice(0, -"/api/v1".length) : ""
+
+async function routeBrowserApiToTestApi(page: Page) {
+  if (!apiOrigin) return
+  await page.route("**/api/v1/**", async (route) => {
+    const sourceUrl = new URL(route.request().url())
+    const response = await route.fetch({
+      url: `${apiOrigin}${sourceUrl.pathname}${sourceUrl.search}`,
+    })
+    await route.fulfill({ response })
+  })
+}
+
+async function loginWithRetry(request: APIRequestContext, email: string, password: string) {
+  for (let i = 0; i <= 2; i++) {
+    const r = await request.post(`${apiBase}/auth/login`, { data: { email, password } })
+    if (r.status() !== 429) {
+      expect(r.ok(), `login: ${await r.text()}`).toBeTruthy()
+      return (await r.json()).access_token as string
+    }
+    console.log(`[douyin] 429 retry ${i + 1}/2`)
+    await new Promise((res) => setTimeout(res, 65000))
+  }
+  throw new Error("Login retries exhausted")
+}
+
+/**
+ * #1972 抖音文案提取冒烟
+ *
+ * 路径：文案库页面 → 点「🎬 从抖音提取」→ 粘贴分享文案 → 点「开始提取」
+ * → mock /api/v1/scripts/extract-from-douyin 返回稳定文案 → 断言「新建文案」弹窗中预填了非空文案
+ */
+test.describe("Douyin Script Extraction (#1972)", () => {
+  test("extract flow: open modal, paste link, text prefilled in create modal", async ({
+    page,
+    request,
+  }) => {
+    test.setTimeout(180_000)
+    await page.setViewportSize({ width: 1440, height: 900 })
+
+    const suffix = Math.random().toString(36).slice(2, 8)
+    const email = `e2e-douyin-${suffix}@example.com`
+    await request.post(`${apiBase}/auth/register`, {
+      data: { email, password: PASSWORD, username: `e2e_dy_${suffix}` },
+    })
+    const token = await loginWithRetry(request, email, PASSWORD)
+    const authHeader = { Authorization: `Bearer ${token}` }
+
+    const proj = await request.post(`${apiBase}/projects`, {
+      headers: authHeader,
+      data: { name: `Smoke Douyin ${suffix}` },
+    })
+    const projectId = (await proj.json()).id ?? (await proj.json()).project_id
+    await request.post(`${apiBase}/asset-libraries`, {
+      headers: authHeader,
+      data: { project_id: projectId, name: "Smoke", kind: "video" },
+    })
+
+    await page.addInitScript((t: string) => {
+      window.localStorage.setItem("access_token", t)
+      window.localStorage.setItem(
+        "auth-storage",
+        JSON.stringify({ state: { token: t, user: null } }),
+      )
+    }, token)
+    await routeBrowserApiToTestApi(page)
+
+    // Mock 抖音提取接口返回稳定文案
+    const extractedText = "大家好，今天给大家推荐一款超好用的产品，性价比非常高，快来看看吧！"
+    await page.route("**/api/v1/scripts/extract-from-douyin", (route) =>
+      route.fulfill({
+        status: 200,
+        contentType: "application/json",
+        body: JSON.stringify({ text: extractedText, duration_seconds: 15 }),
+      }),
+    )
+    // 文案列表空态
+    await page.route(
+      (url) => url.pathname.endsWith("/scripts") && !url.pathname.includes("extract-from-douyin"),
+      (route) =>
+        route.fulfill({
+          status: 200,
+          contentType: "application/json",
+          body: JSON.stringify({ items: [], total: 0, page: 1, page_size: 20 }),
+        }),
+    )
+
+    await page.goto("/app/scripts")
+    // 文案库页面加载
+    await expect(page.getByText(/文案库|文案/).first()).toBeVisible({ timeout: 30000 })
+
+    // 点「🎬 从抖音提取」按钮
+    await page.getByRole("button", { name: /从抖音提取/ }).click()
+    await expect(page.getByText("从抖音视频提取文案")).toBeVisible({ timeout: 5000 })
+
+    // 在 TextArea 粘贴"抖音分享文案"
+    const textarea = page.locator(".ant-modal textarea").first()
+    await expect(textarea).toBeVisible()
+    await textarea.fill("8.88 复制打开抖音，看看【推荐视频】https://v.douyin.com/abcDEF/")
+
+    // 点「开始提取」
+    await page.getByRole("button", { name: "开始提取" }).click()
+    await expect(page.getByText(/提取中/)).toBeVisible({ timeout: 3000 })
+
+    // 等待抖音弹窗关闭，「新建文案」弹窗打开并预填提取文案
+    await expect(page.getByText("从抖音视频提取文案")).not.toBeVisible({ timeout: 15000 })
+    await expect(page.getByText("新建文案")).toBeVisible({ timeout: 5000 })
+    const createTextarea = page.locator(".ant-modal textarea").first()
+    await expect(createTextarea).toBeVisible()
+    await expect(createTextarea).toHaveValue(new RegExp(extractedText.slice(0, 10)))
+    console.log("[douyin] Extraction flow completed ✓, text length:", extractedText.length)
+  })
+})

apps/web/e2e/core-generation.spec.ts

@@ -1,4 +1,4 @@
-import { expect, test, type APIRequestContext } from "@playwright/test"
+import { expect, test, type APIRequestContext, type Page } from "@playwright/test"
 import * as fs from "node:fs"
 import * as path from "node:path"
 import { fileURLToPath } from "node:url"
@@ -8,7 +8,8 @@ const PASSWORD = "SmokePass123!"
 const apiBase = process.env.E2E_API_BASE || "/api/v1"
 const apiOrigin = apiBase.endsWith("/api/v1") ? apiBase.slice(0, -"/api/v1".length) : ""
 
-const routeBrowserApiToTestApi = async (page: import("@playwright/test").Page) => {
+/** 将浏览器侧 /api/v1 请求路由到 Playwright request 源（支持跨域） */
+async function routeBrowserApiToTestApi(page: Page) {
   if (!apiOrigin) return
   await page.route("**/api/v1/**", async (route) => {
     const sourceUrl = new URL(route.request().url())
@@ -24,276 +25,358 @@ async function loginWithRetry(
   email: string,
   password: string,
   maxRetries = 2,
-) {
+): Promise<string> {
   for (let i = 0; i <= maxRetries; i++) {
-    const response = await request.post(`${apiBase}/auth/login`, {
-      data: { email, password },
-    })
-    if (response.status() !== 429) return response
-    console.log(`[login] 触发限流，等待 65s 后重试 (${i + 1}/${maxRetries})`)
+    const resp = await request.post(`${apiBase}/auth/login`, { data: { email, password } })
+    if (resp.status() !== 429) {
+      expect(resp.ok(), `Login should succeed: ${await resp.text()}`).toBeTruthy()
+      const data = await resp.json()
+      return data.access_token
+    }
+    console.log(`[login] 429 rate limited, retry ${i + 1}/${maxRetries} after 65s`)
     await new Promise((r) => setTimeout(r, 65000))
   }
-  return request.post(`${apiBase}/auth/login`, {
-    data: { email, password },
+  throw new Error("Login failed after retries")
+}
+
+/**
+ * 注册新用户 + 建项目/视频库/上传 sample.mp4，等素材 ready。返回 { token, projectId, libraryId, assetId }。
+ */
+async function setupFreshUser(
+  request: APIRequestContext,
+  label: string,
+): Promise<{ token: string; libraryId: string; assetId: string; suffix: string }> {
+  const suffix = Math.random().toString(36).slice(2, 8)
+  const email = `e2e-${label}-${suffix}@example.com`
+  await request.post(`${apiBase}/auth/register`, {
+    data: { email, password: PASSWORD, username: `e2e_${label}_${suffix}` },
   })
-}
+  const token = await loginWithRetry(request, email, PASSWORD)
+  const auth = { Authorization: `Bearer ${token}` }
 
-type ProjectResponse = { id: string }
-type LibraryResponse = { id: string }
-type AssetListResponse = {
-  items: Array<{
-    id: string
-    name: string
-    status: string
-  }>
-}
+  const proj = await request.post(`${apiBase}/projects`, {
+    headers: auth,
+    data: { name: `Smoke ${label} ${suffix}` },
+  })
+  expect(proj.ok(), `create project: ${await proj.text()}`).toBeTruthy()
+  const projectId = (await proj.json()).id ?? (await proj.json()).project_id
 
-test.describe("Core generation flow", () => {
-  test.describe.configure({ timeout: 360_000 })
+  const lib = await request.post(`${apiBase}/asset-libraries`, {
+    headers: auth,
+    data: { project_id: projectId, name: "Smoke", kind: "video" },
+  })
+  expect(lib.ok(), `create library: ${await lib.text()}`).toBeTruthy()
+  const libraryId = (await lib.json()).id
 
-  test("walks through wizard with count modal and starts generation", async ({ page, request }) => {
-    test.setTimeout(360_000)
-
-    await routeBrowserApiToTestApi(page)
-    const suffix = Date.now().toString(36)
-    const email = `e2e-gen-${suffix}@example.com`
-    const username = `e2e_gen_${suffix}`
-    const libraryName = `E2E Gen Lib ${suffix}`
-
-    // Register
-    const register = await request.post(`${apiBase}/auth/register`, {
-      data: { email, username, password: PASSWORD, display_name: username },
-    })
-    expect(register.status()).toBe(201)
-    const registerData = (await register.json()) as { user_id: string }
-
-    // Login
-    const login = await loginWithRetry(request, email, PASSWORD)
-    expect(login.status()).toBe(200)
-    const loginData = (await login.json()) as { access_token: string }
-    const headers = { Authorization: `Bearer ${loginData.access_token}` }
-
-    // Create project
-    const project = await request.post(`${apiBase}/projects`, {
-      headers,
-      data: { name: `E2E Gen Proj ${suffix}` },
-    })
-    expect(project.status()).toBe(200)
-    const projectData = (await project.json()) as ProjectResponse
-
-    // Create asset library
-    const library = await request.post(`${apiBase}/asset-libraries`, {
-      headers,
-      data: { project_id: projectData.id, name: libraryName, kind: "video" },
-    })
-    expect(library.status()).toBe(200)
-    const libraryData = (await library.json()) as LibraryResponse
-
-    // Upload source video
-    const sourceFileName = "e2e-gen-source.mp4"
-    const sampleVideoPath = path.join(__dirname, "fixtures", "sample.mp4")
-    const sampleVideoBuffer = fs.readFileSync(sampleVideoPath)
-    const upload = await request.post(`${apiBase}/upload`, {
-      headers,
-      multipart: {
-        project_id: projectData.id,
-        library_id: libraryData.id,
-        file: {
-          name: sourceFileName,
-          mimeType: "video/mp4",
-          buffer: sampleVideoBuffer,
-        },
+  const samplePath = path.join(__dirname, "fixtures", "sample.mp4")
+  const sampleBuf = fs.readFileSync(samplePath)
+  const up = await request.post(`${apiBase}/upload`, {
+    headers: auth,
+    multipart: {
+      project_id: projectId,
+      library_id: libraryId,
+      file: {
+        name: "sample.mp4",
+        mimeType: "video/mp4",
+        buffer: sampleBuf,
       },
-    })
-    expect(upload.status()).toBe(200)
+    },
+  })
+  expect(up.ok(), `upload sample: ${await up.text()}`).toBeTruthy()
+  const assetId = (await up.json()).asset_id
+  await expect
+    .poll(
+      async () => {
+        const r = await request.get(`${apiBase}/assets/${assetId}`, { headers: auth })
+        return r.ok() ? (await r.json()).status : "pending"
+      },
+      { timeout: 90_000, intervals: [3000, 3000, 5000] },
+    )
+    .toBe("ready")
+  return { token, libraryId, assetId, suffix }
+}
 
-    // Wait for asset to be ready
-    await expect
-      .poll(
-        async () => {
-          const assets = await request.get(`${apiBase}/assets`, {
-            headers,
-            params: { library_id: libraryData.id },
-          })
-          if (!assets.ok()) return `http_${assets.status()}`
-          const data = (await assets.json()) as AssetListResponse
-          const asset = data.items.find((a) => a.name === sourceFileName)
-          if (!asset) return "missing"
-          return asset.status
-        },
-        { timeout: 30_000, intervals: [1_000, 2_000, 3_000] },
+/**
+ * #1970 智能剪辑核心冒烟（新 5 步向导）
+ *
+ * 新流程：选择模式 → 选择素材 → 选择标题 → 确认生成 → 选择封面
+ *
+ * 两条路径：
+ *   1) 随机混剪（默认）→ Step1 下一步 → 配音选择弹窗 → Step2 选素材 → 数量弹窗
+ *      → Step3 标题 → Step4 确认生成 → 断言任务创建
+ *   2) 叙事剪辑 → Step1 切模式 → 下一步 → 文案选择弹窗 → TTS 弹窗选音色（mock 合成）
+ *      → Step2 AI 提示卡可见 + 选素材 → 数量弹窗 → Step3 标题 → Step4 确认生成
+ *      → 断言任务创建
+ */
+test.describe("Core Smart-Edit Flow (#1970)", () => {
+  test("random mode: 5-step wizard creates generation task", async ({ page, request }) => {
+    test.setTimeout(600_000)
+    await page.setViewportSize({ width: 1440, height: 1000 })
+    const { token, suffix } = await setupFreshUser(request, "random")
+    const authHeader = { Authorization: `Bearer ${token}` }
+
+    // 确保默认模板存在（智能剪辑页依赖模板）
+    const tmpls = await request.get(`${apiBase}/templates`, { headers: authHeader })
+    const tmplsJson = await tmpls.json()
+    const templates = Array.isArray(tmplsJson)
+      ? tmplsJson
+      : Array.isArray(tmplsJson.items)
+        ? tmplsJson.items
+        : []
+    expect(templates.length).toBeGreaterThan(0)
+
+    // 注入登录态 + 路由 API
+    await page.addInitScript((t: string) => {
+      window.localStorage.setItem("access_token", t)
+      window.localStorage.setItem(
+        "auth-storage",
+        JSON.stringify({ state: { token: t, user: null } }),
       )
-      .toBe("ready")
+    }, token)
+    await routeBrowserApiToTestApi(page)
 
-    // GET /templates auto-creates a default template for new users
-    const templatesResp = await request.get(`${apiBase}/templates`, { headers })
-    expect(templatesResp.status(), await templatesResp.text()).toBe(200)
-    const templatesData = (await templatesResp.json()) as {
-      items: Array<{ id: string }>
-    }
-    expect(Array.isArray(templatesData.items)).toBe(true)
-    expect(templatesData.items.length).toBeGreaterThan(0)
-    const templateId = templatesData.items[0].id
-    expect(templateId).toBeTruthy()
-
-    // Set auth in localStorage
-    await page.addInitScript(
-      ({ token, user }) => {
-        localStorage.setItem("access_token", token)
-        localStorage.setItem(
-          "auth-storage",
-          JSON.stringify({
-            state: { user, isAuthenticated: true },
-            version: 0,
+    // ── 提前 mock 配音列表（VoiceSelectModal 查询 /assets?kind=voice） ──
+    await page.route(
+      (url) => url.pathname.endsWith("/assets") && url.searchParams.get("kind") === "voice",
+      (route) =>
+        route.fulfill({
+          status: 200,
+          contentType: "application/json",
+          body: JSON.stringify({
+            items: [
+              {
+                id: `asset-voice-${suffix}`,
+                name: "测试配音.mp3",
+                file_url: "data:audio/mpeg;base64,",
+                duration: 10,
+                file_size: 1024,
+                kind: "voice",
+                status: "ready",
+              },
+            ],
+            total: 1,
           }),
-        )
-      },
-      {
-        token: loginData.access_token,
-        user: {
-          id: registerData.user_id,
-          user_id: registerData.user_id,
-          email,
-          username,
-          display_name: username,
-          is_email_verified: true,
-          email_verified: true,
-        },
-      },
+        }),
     )
 
-    // Navigate to generate page
     await page.goto("/app/generate")
     await expect(page.getByRole("heading", { name: "智能剪辑" })).toBeVisible({
-      timeout: 20_000,
+      timeout: 30000,
     })
 
-    // 5步向导：素材(1)→配音(2)→标题(3)→确认生成(4)→封面(5)
+    // ── Step 1：默认随机混剪选中，点下一步 ──────────────────────────
+    await expect(page.getByText("选择模式", { exact: true })).toBeVisible()
+    await expect(page.getByText("随机混剪")).toBeVisible()
+    await page.getByRole("button", { name: /下一步/ }).click()
 
-    // ── Step 1: 素材选择 ──
-    await expect(page.getByRole("heading", { name: /选择素材/ })).toBeVisible()
-    const librarySelect = page.locator("select").first()
-    await librarySelect.selectOption({ label: libraryName })
-    const materialCard = page.getByTestId("material-card").filter({ hasText: sourceFileName })
-    await expect(materialCard).toBeVisible({ timeout: 10_000 })
-    await materialCard.click({ position: { x: 15, y: 15 } })
-    await expect(materialCard.getByTestId("material-card-check")).toBeVisible({ timeout: 5_000 })
-    await page.getByRole("button", { name: "下一步" }).click()
+    // ── 配音选择弹窗：选第一个配音 → 确认 ─────────────────────────
+    await expect(page.getByText("🎙️ 选择配音")).toBeVisible({ timeout: 5000 })
+    await page.getByText("测试配音.mp3").first().click()
+    await page.getByRole("button", { name: "确认选择" }).click()
+    await expect(page.getByText("🎙️ 选择配音")).not.toBeVisible()
 
-    // ── 数量弹窗（PreviewCountModal） ──
-    await expect(page.getByRole("heading", { name: "要生成几个视频？" })).toBeVisible({
-      timeout: 5_000,
-    })
+    // ── Step 2：选择素材 ──────────────────────────────────────────
+    await expect(page.getByText("选择素材", { exact: true })).toBeVisible({ timeout: 10000 })
+    await page.getByTestId("material-card").first().click()
+    await page.getByRole("button", { name: /下一步/ }).click()
+
+    // ── 数量弹窗：默认 1 个 → 确认 ───────────────────────────────
+    await expect(page.getByText("要生成几个视频？")).toBeVisible({ timeout: 5000 })
     await page.getByRole("button", { name: "生成 1 个视频" }).click()
 
-    // ── Step 2: 配音（新注册用户无配音素材，跳过） ──
-    await expect(page.getByRole("heading", { name: /选择配音/ })).toBeVisible({ timeout: 15000 })
-    await page.getByRole("button", { name: "下一步" }).click()
-
-    // ── Step 3: 标题设置 ──
-    await expect(page.getByRole("heading", { name: /选择标题/ })).toBeVisible({ timeout: 15000 })
-    await page.waitForTimeout(2000)
-
-    const titleInput = page.locator(".ant-select-auto-complete input")
+    // ── Step 3：填写标题 ──────────────────────────────────────────
+    await expect(page.getByText("选择标题", { exact: true })).toBeVisible({ timeout: 10000 })
+    const titleInput = page.getByPlaceholder("输入或从标题库选择")
     await expect(titleInput).toBeVisible({ timeout: 5000 })
-    await titleInput.fill(`E2E Test ${suffix}`)
+    await titleInput.fill(`测试随机剪辑 ${suffix}`)
+    await page.getByRole("button", { name: /下一步/ }).click()
 
-    // Step 3 底部是「下一步 →」，点击进入 Step 4（确认生成）
-    await page.getByRole("button", { name: "下一步" }).click()
+    // ── Step 4：确认生成 ──────────────────────────────────────────
+    await expect(page.getByText("📋 生成配置")).toBeVisible({ timeout: 10000 })
+    await expect(page.getByText("随机混剪")).toBeVisible()
+    const confirmBtn = page.getByRole("button", { name: /确认生成视频/ })
+    await expect(confirmBtn).toBeEnabled({ timeout: 5000 })
 
-    // ── Step 4: 确认生成 ──
-    // 等待实时预览就绪（占位消失）
-    await page
-      .getByText("准备预览素材")
-      .waitFor({ state: "detached", timeout: 30_000 })
-      .catch(() => {})
+    const createTask = page.waitForResponse(
+      (r) => r.url().includes("/generation/tasks") && r.request().method() === "POST",
+      { timeout: 30000 },
+    )
+    await confirmBtn.click()
+    const taskResp = await createTask
+    expect(taskResp.ok(), `Create task: ${await taskResp.text()}`).toBeTruthy()
+    const taskId = (await taskResp.json()).id ?? (await taskResp.json()).task_id
+    console.log("[random] Generation task created:", taskId)
+    await expect(page.getByText(/正在生成|提交/)).toBeVisible({ timeout: 15000 })
+    console.log("[random] Wizard flow completed ✓")
+  })
 
-    // Step 4 底部是「✨ 确认生成视频」
-    const confirmBtn = page.locator(".xx-step-actions .xx-btn-primary").first()
-    await expect(confirmBtn).toBeVisible({ timeout: 15_000 })
+  test("narrative mode: select script + mock TTS, create generation task", async ({
+    page,
+    request,
+  }) => {
+    test.setTimeout(600_000)
+    await page.setViewportSize({ width: 1440, height: 1000 })
+    const { token, suffix } = await setupFreshUser(request, "narrative")
 
-    // 先挂 API 监听再点击
-    const generatePromise = page.waitForResponse(
-      (response) => {
-        const url = response.url()
-        const path = new URL(url).pathname
-        return response.request().method() === "POST" && path.endsWith("/generation/tasks")
-      },
-      { timeout: 30_000 },
+    await page.addInitScript((t: string) => {
+      window.localStorage.setItem("access_token", t)
+      window.localStorage.setItem(
+        "auth-storage",
+        JSON.stringify({ state: { token: t, user: null } }),
+      )
+    }, token)
+    await routeBrowserApiToTestApi(page)
+
+    // ── Mock 文案列表、音色、TTS 合成（避免真实合成） ──────────────
+    const mockScriptId = `script-mock-${suffix}`
+    const mockVoiceId = `preset-voice-${suffix}`
+    const mockJobId = `tts-job-${suffix}`
+
+    // 文案列表（ScriptSelectModal 查询 /scripts）
+    await page.route("**/api/v1/scripts**", (route) => {
+      const url = new URL(route.request().url())
+      if (url.pathname.includes("/extract-from-douyin")) {
+        route.continue()
+        return
+      }
+      route.fulfill({
+        status: 200,
+        contentType: "application/json",
+        body: JSON.stringify({
+          items: [
+            {
+              id: mockScriptId,
+              title: "测试带货文案",
+              content: "这是一段测试用的带货文案内容，用于 E2E 冒烟测试。",
+              tags: ["带货"],
+              title_category: "daihuo",
+              created_at: new Date().toISOString(),
+              updated_at: new Date().toISOString(),
+            },
+          ],
+          total: 1,
+          page: 1,
+          page_size: 200,
+        }),
+      })
+    })
+
+    // 预设音色（TtsVoiceModal 查询 GET /voices/presets）
+    await page.route("**/api/v1/voices/presets**", (route) =>
+      route.fulfill({
+        status: 200,
+        contentType: "application/json",
+        body: JSON.stringify({
+          items: [
+            {
+              voice_id: mockVoiceId,
+              name: "晓晓（女声）",
+              description: "温柔女声",
+              gender: "female",
+              language: "zh-CN",
+              preview_url: null,
+              tags: ["温柔"],
+            },
+          ],
+          total: 1,
+        }),
+      }),
     )
 
-    await confirmBtn.click()
+    // 克隆音色：空列表
+    await page.route(
+      (url) => url.pathname.endsWith("/voice-clones"),
+      (route) =>
+        route.fulfill({
+          status: 200,
+          contentType: "application/json",
+          body: JSON.stringify({ items: [] }),
+        }),
+    )
 
-    // 验证生成 API 被调用
-    const genResp = await generatePromise.catch(() => null)
-    if (!genResp) {
-      // staging 预览未就绪导致按钮校验拦截，未触发 API — 向导导航仍通过
-      console.log(
-        "[E2E] Generation API not triggered (preview not ready) — wizard navigation verified",
-      )
-    } else if (genResp.ok()) {
-      const genData = (await genResp.json()) as {
-        items: Array<{ id: string; status: string }>
-        total: number
-      }
-      expect(genData.items.length).toBeGreaterThan(0)
+    // TTS 合成：直接返回 completed 任务
+    await page.route("**/api/v1/tts/synthesize", (route) =>
+      route.fulfill({
+        status: 200,
+        contentType: "application/json",
+        body: JSON.stringify({ job_id: mockJobId, status: "queued" }),
+      }),
+    )
+    await page.route(`**/api/v1/tts/jobs/${mockJobId}/status`, (route) =>
+      route.fulfill({
+        status: 200,
+        contentType: "application/json",
+        body: JSON.stringify({
+          job_id: mockJobId,
+          status: "completed",
+          progress: 100,
+          audio_url: "data:audio/mpeg;base64,",
+          duration: 5,
+        }),
+      }),
+    )
+    await page.route(`**/api/v1/tts/jobs/${mockJobId}/save-to-library`, (route) =>
+      route.fulfill({
+        status: 200,
+        contentType: "application/json",
+        body: JSON.stringify({ id: `tts-asset-${suffix}`, name: "AI合成配音" }),
+      }),
+    )
 
-      // race：渲染完成 vs 生成失败/超时
-      const downloadReady = page
-        .getByText("视频生成完成")
-        .isVisible({ timeout: 180_000 })
-        .then((v) => (v ? "completed" : null))
-      const generationFailed = page
-        .getByText(/生成失败|重新生成/)
-        .isVisible({ timeout: 180_000 })
-        .then((v) => (v ? "failed" : null))
-
-      const outcome = await Promise.any([downloadReady, generationFailed]).catch(() => "timeout")
-
-      if (outcome === "completed") {
-        await page.getByRole("button", { name: /下一步：选择封面/ }).click()
-        await expect(page.getByRole("heading", { name: /选择封面/ })).toBeVisible({
-          timeout: 30_000,
-        })
-      } else {
-        console.log(`[E2E] Video rendering ${outcome} on staging — wizard flow verified`)
-      }
-    } else {
-      console.log(`[E2E] Generate API returned ${genResp.status()}, wizard flow test still passes`)
-    }
-
-    // 验证成品库页面加载
-    await page.goto("/app/products")
-    await expect(page).toHaveURL(/\/app\/products/)
-    await expect(page.locator(".xx-products-page")).toBeVisible({ timeout: 15_000 })
-
-    await page.unrouteAll({ behavior: "ignoreErrors" })
-  })
-
-  test("generation task API creates and lists tasks", async ({ request }) => {
-    const suffix = Date.now().toString(36)
-    const email = `e2e-gen-api-${suffix}@example.com`
-    const username = `e2e_gen_api_${suffix}`
-
-    const register = await request.post(`${apiBase}/auth/register`, {
-      data: { email, username, password: PASSWORD, display_name: username },
+    await page.goto("/app/generate")
+    await expect(page.getByRole("heading", { name: "智能剪辑" })).toBeVisible({
+      timeout: 30000,
     })
-    expect(register.status()).toBe(201)
 
-    const login = await loginWithRetry(request, email, PASSWORD)
-    expect(login.status()).toBe(200)
-    const loginData = (await login.json()) as { access_token: string }
-    const headers = { Authorization: `Bearer ${loginData.access_token}` }
+    // ── Step 1：切到叙事剪辑 → 下一步 ────────────────────────────
+    await expect(page.getByText("选择模式", { exact: true })).toBeVisible()
+    await page.getByText("叙事剪辑").click()
+    await page.getByRole("button", { name: /下一步/ }).click()
 
-    const project = await request.post(`${apiBase}/projects`, {
-      headers,
-      data: { name: `E2E API Proj ${suffix}` },
-    })
-    expect(project.status()).toBe(200)
+    // ── 文案选择弹窗：选第一条 → 确认 ─────────────────────────────
+    await expect(page.getByText("📝 选择文案")).toBeVisible({ timeout: 5000 })
+    await page.getByText("测试带货文案").first().click()
+    await page.getByRole("button", { name: "确认选择" }).click()
+    await expect(page.getByText("📝 选择文案")).not.toBeVisible()
 
-    const tasks = await request.get(`${apiBase}/tasks`, { headers })
-    expect(tasks.status()).toBe(200)
-    const tasksData = await tasks.json()
-    expect(Array.isArray(tasksData.items)).toBe(true)
+    // ── TTS 音色弹窗：选系统音色 → 合成 ─────────────────────────
+    await expect(page.getByText("🎙️ 合成配音")).toBeVisible({ timeout: 5000 })
+    await page.getByText("晓晓（女声）").first().click()
+    await page.getByRole("button", { name: "🎧 合成配音" }).click()
+    await expect(page.getByText("🎙️ 合成配音")).not.toBeVisible({ timeout: 30000 })
+
+    // ── Step 2：AI 匹配提示卡可见 + 选素材 ────────────────────────
+    await expect(page.getByText("选择素材", { exact: true })).toBeVisible({ timeout: 10000 })
+    await expect(page.getByText(/AI智能匹配/)).toBeVisible()
+    await page.getByTestId("material-card").first().click()
+    await page.getByRole("button", { name: /下一步/ }).click()
+
+    // ── 数量弹窗 ─────────────────────────────────────────────────
+    await expect(page.getByText("要生成几个视频？")).toBeVisible({ timeout: 5000 })
+    await page.getByRole("button", { name: "生成 1 个视频" }).click()
+
+    // ── Step 3：填写标题（handleScriptModalConfirm 已预填 script.title，但我们再覆盖一次） ─
+    await expect(page.getByText("选择标题", { exact: true })).toBeVisible({ timeout: 10000 })
+    const titleInput2 = page.getByPlaceholder("输入或从标题库选择")
+    await expect(titleInput2).toBeVisible({ timeout: 5000 })
+    await titleInput2.fill(`测试叙事剪辑 ${suffix}`)
+    await page.getByRole("button", { name: /下一步/ }).click()
+
+    // ── Step 4：确认生成 ──────────────────────────────────────────
+    await expect(page.getByText("📋 生成配置")).toBeVisible({ timeout: 10000 })
+    await expect(page.getByText("叙事剪辑")).toBeVisible()
+    const confirmBtn2 = page.getByRole("button", { name: /确认生成视频/ })
+    await expect(confirmBtn2).toBeEnabled({ timeout: 5000 })
+
+    const createTask2 = page.waitForResponse(
+      (r) => r.url().includes("/generation/tasks") && r.request().method() === "POST",
+      { timeout: 30000 },
+    )
+    await confirmBtn2.click()
+    const taskResp2 = await createTask2
+    expect(taskResp2.ok(), `Create task: ${await taskResp2.text()}`).toBeTruthy()
+    console.log("[narrative] Generation task created:", (await taskResp2.json()).id)
+    await expect(page.getByText(/正在生成|提交/)).toBeVisible({ timeout: 15000 })
+    console.log("[narrative] Wizard flow completed ✓")
   })
 })

apps/web/e2e/core-navigation.spec.ts

@@ -0,0 +1,105 @@
+import { expect, test, type APIRequestContext, type Page } from "@playwright/test"
+
+const PASSWORD = "SmokePass123!"
+const apiBase = process.env.E2E_API_BASE || "/api/v1"
+const apiOrigin = apiBase.endsWith("/api/v1") ? apiBase.slice(0, -"/api/v1".length) : ""
+
+async function routeBrowserApiToTestApi(page: Page) {
+  if (!apiOrigin) return
+  await page.route("**/api/v1/**", async (route) => {
+    const sourceUrl = new URL(route.request().url())
+    const response = await route.fetch({
+      url: `${apiOrigin}${sourceUrl.pathname}${sourceUrl.search}`,
+    })
+    await route.fulfill({ response })
+  })
+}
+
+async function loginWithRetry(request: APIRequestContext, email: string, password: string) {
+  for (let i = 0; i <= 2; i++) {
+    const r = await request.post(`${apiBase}/auth/login`, { data: { email, password } })
+    if (r.status() !== 429) {
+      expect(r.ok(), `login: ${await r.text()}`).toBeTruthy()
+      return (await r.json()).access_token as string
+    }
+    console.log(`[nav] 429 retry ${i + 1}/2`)
+    await new Promise((res) => setTimeout(res, 65000))
+  }
+  throw new Error("Login retries exhausted")
+}
+
+/**
+ * 核心页面导航冒烟：侧边栏主要入口能访问、文案库/配音库页面能正常加载（不出白屏/无致命 js error）
+ */
+test.describe("Core Navigation", () => {
+  let authToken: string
+
+  test.beforeAll(async ({ request }) => {
+    const suffix = Math.random().toString(36).slice(2, 8)
+    const email = `e2e-nav-${suffix}@example.com`
+    await request.post(`${apiBase}/auth/register`, {
+      data: { email, password: PASSWORD, username: `e2e_nav_${suffix}` },
+    })
+    authToken = await loginWithRetry(request, email, PASSWORD)
+    const authHeader = { Authorization: `Bearer ${authToken}` }
+    const proj = await request.post(`${apiBase}/projects`, {
+      headers: authHeader,
+      data: { name: `Smoke Nav ${suffix}` },
+    })
+    if (proj.ok()) {
+      const projectId = (await proj.json()).id ?? (await proj.json()).project_id
+      await request.post(`${apiBase}/asset-libraries`, {
+        headers: authHeader,
+        data: { project_id: projectId, name: "Nav Lib", kind: "video" },
+      })
+    }
+  })
+
+  test.beforeEach(async ({ page }) => {
+    await page.setViewportSize({ width: 1440, height: 900 })
+    await page.addInitScript((t: string) => {
+      window.localStorage.setItem("access_token", t)
+      window.localStorage.setItem(
+        "auth-storage",
+        JSON.stringify({ state: { token: t, user: null } }),
+      )
+    }, authToken)
+    await routeBrowserApiToTestApi(page)
+  })
+
+  const navCases = [
+    { path: "/app/dashboard", marker: /概览|工作台|最近/i, name: "概览" },
+    { path: "/app/generate", marker: /智能剪辑|剪辑/, name: "智能剪辑" },
+    { path: "/app/assets", marker: /视频库|素材/, name: "视频库" },
+    { path: "/app/scripts", marker: /文案/, name: "文案库" },
+    { path: "/app/voices", marker: /配音|我的音色|配音库/, name: "配音库" },
+    { path: "/app/products", marker: /成品|作品/, name: "成品库" },
+    { path: "/app/history", marker: /历史|任务/, name: "任务历史" },
+    { path: "/app/tasks", marker: /任务中心|任务列表/, name: "任务中心" },
+    { path: "/app/points", marker: /积分|我的积分/, name: "积分中心" },
+  ]
+
+  for (const c of navCases) {
+    test(`visit ${c.name} (${c.path}) loads without fatal pageerror`, async ({ page }) => {
+      const errors: Error[] = []
+      page.on("pageerror", (e) => errors.push(e))
+      await page.goto(c.path)
+      await expect(page.locator("body")).not.toBeEmpty({ timeout: 20000 })
+      // 过滤掉常见第三方/非致命错误
+      const fatal = errors.filter(
+        (e) =>
+          !/ResizeObserver|Loading chunk|network error|Failed to fetch|chunkLoadError/i.test(
+            e.message,
+          ),
+      )
+      expect(fatal, `${c.name} pageerrors: ${fatal.map((e) => e.message).join("; ")}`).toHaveLength(
+        0,
+      )
+      await expect(
+        page.getByText(c.marker).first(),
+        `${c.name} should show relevant text`,
+      ).toBeVisible({ timeout: 15000 })
+      console.log(`[nav] ${c.name} loaded ✓`)
+    })
+  }
+})

apps/worker/video_processing/render_adapter.py

@@ -493,6 +493,41 @@ class RenderAdapter:
             logger.warning("ASR 服务初始化失败，自动字幕将不可用: %s", e)
             return None
 
+    def _resolve_clip_has_text(self, clips: list[Any]) -> list[bool] | None:
+        """#1970：按源视频片段顺序解析 atom_clip.ai_tags.has_text。
+
+        顺序与 UnifiedRenderService 的「非 audio 源片段」口径一致。
+        仅当 atom_clip 存在 ai_tags 字典且 has_text 显式为 False 时标记为
+        无文字（允许 hflip）；atom_clip_id 缺失、ai_tags 未生成、has_text 为
+        true/null/非布尔值时一律按有文字处理（保守不翻转）。
+        查询失败时返回 None，渲染层回退到全保守路径。
+        """
+        video_clips = [c for c in clips if getattr(c, "clip_type", "main") != "audio"]
+        atom_ids: list[str] = []
+        seen: set[str] = set()
+        for c in video_clips:
+            atom_id = getattr(c, "atom_clip_id", "") or ""
+            if atom_id and atom_id not in seen:
+                seen.add(atom_id)
+                atom_ids.append(atom_id)
+        if not atom_ids:
+            return None
+        try:
+            from packages.adapters.sqlalchemy_impl.asset_atom_clip_repository import (
+                SQLAlchemyAssetAtomClipRepository,
+            )
+
+            atom_clips = SQLAlchemyAssetAtomClipRepository(self._db).find_by_ids(atom_ids)
+        except Exception as exc:
+            logger.warning("[render-adapter] atom_clip ai_tags 查询失败，hflip 全量保守处理: %s", exc)
+            return None
+        has_text_map: dict[str, bool] = {}
+        for ac in atom_clips:
+            ai_tags = getattr(ac, "ai_tags", None)
+            no_text = isinstance(ai_tags, dict) and ai_tags.get("has_text") is False
+            has_text_map[ac.id] = not no_text
+        return [has_text_map.get((getattr(c, "atom_clip_id", "") or ""), True) for c in video_clips]
+
     def _do_render(
         self,
         plan: Any,
@@ -542,6 +577,7 @@ class RenderAdapter:
         )
 
         # 4. 执行统一渲染
+        clip_has_text = self._resolve_clip_has_text(clips)
         render_svc = UnifiedRenderService(
             plan=plan,
             clips=clips,
@@ -552,6 +588,7 @@ class RenderAdapter:
             bgm_path=bgm_path,
             asr_service=asr_service,
             voiceover_audio_path=voiceover_audio_path,
+            clip_has_text=clip_has_text,
         )
         result = render_svc.render()
 

apps/worker/video_processing/unified_render_service.py

@@ -155,6 +155,7 @@ class UnifiedRenderService:
         asr_service: Any = None,  # ASRService 实例，用于自动生成字幕
         bgm_path: str | None = None,  # BGM 本地文件路径
         voiceover_audio_path: str | None = None,  # 配音素材库音频本地路径
+        clip_has_text: list[bool] | None = None,  # 源视频片段是否有文字（来自 atom_clip.ai_tags.has_text）
     ):
         self.plan = plan
         self.clips = clips
@@ -167,6 +168,8 @@ class UnifiedRenderService:
         self.asr_service = asr_service
         self.bgm_path = bgm_path
         self.voiceover_audio_path = voiceover_audio_path
+        # #1970：片段级文字检测（顺序与非 audio 的源视频片段一致）；None 表示无可靠检测，保守不翻转
+        self._clip_has_text = clip_has_text
         self._transition_engine = TransitionEngine(default_duration=transition_duration)
         self._speed_engine = SpeedEngine()
         self._asr_timeline_cache: Any = None  # ASR 字幕结果缓存，避免重复调用
@@ -186,7 +189,9 @@ class UnifiedRenderService:
 
         种子 hash(generation_task_id + video_index)%10000，同一任务重渲结果一致。
         dedup_enabled=False 时返回 None，调用方不注入任何微变换。
-        P1 字幕检测：无可靠的片段文字轨道信息，hflip 一律关闭（宁可不翻转）。
+        hflip 放开（#1970）：clip_has_text 来自 atom_clip.ai_tags.has_text，
+        仅 AI 明确判定无文字的片段可参与 50% 翻转；未打标签 / has_text 为
+        true/null 或缺位时一律视为有文字，保持保守不翻转。
         """
         if self._micro_plan_loaded:
             return self._micro_plan_cache
@@ -200,11 +205,14 @@ class UnifiedRenderService:
             cfg = self.plan.config or {}
             task_id = str(cfg.get("generation_task_id", "") or "")
             video_index = int(cfg.get("video_index", 0) or 0)
+            # self._clip_has_text 顺序与非 audio 源片段一致；
+            # None（未提供检测，如内存直渲/旧任务）→ 纯函数层按全有文字保守处理；
+            # 列表短于片段数时缺位片段同样按有文字处理
             self._micro_plan_cache = build_micro_transform_plan(
                 task_id,
                 video_index,
                 clip_count,
-                clip_has_text=None,  # P1 保守策略：全部按有文字处理，不翻转
+                clip_has_text=self._clip_has_text,
                 enable_bgm_offset=bool(cfg.get("bgm")),
             )
         except Exception as e:

apps/worker/worker_app/celery_app.py

@@ -28,6 +28,10 @@ celery_app.conf.imports = (
     "worker_app.tasks.health",
     "worker_app.tasks.ingest",
     "worker_app.tasks.atom_clips",
+    # #1970 片段级 AI 标签：必须显式 import 注册，否则 worker 报
+    # "Received unregistered task of type 'worker.tag_atom_clip'"
+    "worker_app.tasks.atom_clip_tagging",
+    "worker_app.tasks.backfill_atom_clip_tags",
     "worker_app.tasks.classification",
     "worker_app.tasks.generation",
     "worker_app.tasks.voice_extraction",

apps/worker/worker_app/tasks/atom_clip_tagging.py

@@ -25,11 +25,14 @@ logger = get_task_logger(__name__)
 
 
 @celery_app.task(name="worker.tag_atom_clip", bind=True, max_retries=2, default_retry_delay=10)
-def tag_atom_clip_task(self, atom_clip_id: str) -> dict:
+def tag_atom_clip_task(self, atom_clip_id: str, force: bool = False) -> dict:
     """为单个原子片段生成 AI 标签.
 
     Args:
         atom_clip_id: 原子片段 ID。
+        force: True 时允许覆盖只有 inherited_tags 的降级记录
+            （视觉 API 曾失败写入的占位标签，#1970）。
+            已有完整标签（含 has_text）始终跳过，保证幂等。
 
     Returns:
         任务结果 dict：status / clip_id / ai_tags（部分字段）。
@@ -43,9 +46,11 @@ def tag_atom_clip_task(self, atom_clip_id: str) -> dict:
         if clip is None:
             return {"status": "skipped", "reason": "clip not found", "clip_id": atom_clip_id}
 
-        # 已有标签则跳过（幂等）
+        # 已有完整标签则跳过（幂等）；force 仅放行缺失 has_text 的降级记录
         if clip.ai_tags is not None:
-            return {"status": "skipped", "reason": "already tagged", "clip_id": atom_clip_id}
+            has_real_tags = isinstance(clip.ai_tags, dict) and "has_text" in clip.ai_tags
+            if has_real_tags or not force:
+                return {"status": "skipped", "reason": "already tagged", "clip_id": atom_clip_id}
 
         # 获取素材信息
         asset = asset_repo.find_by_id(clip.asset_id)

apps/worker/worker_app/tasks/backfill_atom_clip_tags.py

@@ -30,6 +30,7 @@ def backfill_atom_clip_tags(
     batch_size: int = DEFAULT_BATCH_SIZE,
     batch_interval: int = DEFAULT_BATCH_INTERVAL,
     max_clips: int = 0,
+    force: bool = False,
 ) -> dict:
     """批量回填未打标的 atom_clips.
 
@@ -37,6 +38,8 @@ def backfill_atom_clip_tags(
         batch_size: 每批处理数量，默认 10。
         batch_interval: 每批间隔秒数，默认 5。
         max_clips: 最大处理总数，0 表示不限。
+        force: True 时连同只有 inherited_tags 的降级记录一起强制重打
+            （视觉 API 曾失败、DOUBAO_VISION_MODEL 修复后重跑用，#1970）。
 
     Returns:
         任务结果 dict：total_submitted / batches。
@@ -52,7 +55,7 @@ def backfill_atom_clip_tags(
             remaining = max_clips - total_submitted if max_clips > 0 else batch_size
             fetch_limit = min(batch_size, remaining) if max_clips > 0 else batch_size
 
-            untagged = atom_repo.find_untagged(limit=fetch_limit)
+            untagged = atom_repo.find_untagged(limit=fetch_limit, include_downgraded=force)
             if not untagged:
                 break
 
@@ -62,6 +65,7 @@ def backfill_atom_clip_tags(
                     celery_app.send_task(
                         "worker.tag_atom_clip",
                         args=[clip.id],
+                        kwargs={"force": force},
                     )
                     total_submitted += 1
                 except Exception as e:

deploy/configs/.env.production

@@ -234,6 +234,9 @@ DOUBAO_TIMEOUT=60
 # 最大重试次数
 DOUBAO_MAX_RETRIES=2
 
+# 视觉模型 Endpoint ID（支持图片/视频理解的模型）
+DOUBAO_VISION_MODEL=${DOUBAO_VISION_MODEL}
+
 
 # ==================== 微信开放平台 OAuth（网页扫码登录）====================
 # 回调域名：xiaoxiajianji.com（微信开放平台已配置）
@@ -255,4 +258,4 @@ APIZERO_API_KEY=${APIZERO_API_KEY}
 
 # ==================== GPU MuseTalk Worker（反向轮询） ====================
 GPU_WORKER_TOKEN=${GPU_WORKER_TOKEN}
-GPU_TASK_TIMEOUT_SECONDS=300
+GPU_TASK_TIMEOUT_SECONDS=900

deploy/configs/.env.staging

@@ -251,6 +251,9 @@ DOUBAO_TIMEOUT=60
 # 最大重试次数
 DOUBAO_MAX_RETRIES=2
 
+# 视觉模型 Endpoint ID（支持图片/视频理解的模型）
+DOUBAO_VISION_MODEL=${DOUBAO_VISION_MODEL}
+
 
 # ==================== 微信开放平台 OAuth（网页扫码登录）====================
 # 回调域名：xiaoxiajianji.com（微信开放平台已配置）
@@ -272,4 +275,4 @@ APIZERO_API_KEY=${APIZERO_API_KEY}
 
 # ==================== GPU MuseTalk Worker（反向轮询） ====================
 GPU_WORKER_TOKEN=${GPU_WORKER_TOKEN}
-GPU_TASK_TIMEOUT_SECONDS=300
+GPU_TASK_TIMEOUT_SECONDS=900

deploy/gpu_worker/.env.example

@@ -19,7 +19,12 @@ MUSE_TALK_URL=http://127.0.0.1:7861
 # 轮询/心跳/超时（秒）
 POLL_INTERVAL=5
 HEARTBEAT_INTERVAL=15
-REQUEST_TIMEOUT=300
+# 下载/推理/上传 HTTP 超时，需与服务端 GPU_TASK_TIMEOUT_SECONDS 对齐（默认 900）
+REQUEST_TIMEOUT=900
 
-# 单个任务本地最大重试次数（首次失败后再重试 N 次，默认 2）
-TASK_MAX_RETRY=2
+# 单个任务本地最大重试次数（仅网络/MuseTalk 瞬时错误才重试，默认 1）
+TASK_MAX_RETRY=1
+# 推理期间任务心跳间隔（秒，独立线程，无需改动）
+TASK_HEARTBEAT_INTERVAL=30
+# 输入视频最短时长（秒），小于则直接上报失败，不调用 MuseTalk
+MIN_VIDEO_DURATION_SECONDS=3

deploy/gpu_worker/README.md

@@ -89,8 +89,9 @@ SaaS 后端部署完成后需配置：
 | 日志 `MuseTalk 健康检查未通过` | 本地 MuseTalk 没启动，或端口不是 7861；`curl http://127.0.0.1:7861/health` 验证 |
 | 任务长时间不被拉取 | Worker 和服务端连不上；检查 API_BASE_URL 是否可达、Token 是否正确 |
 | 推理后上传 OSS 失败 | 本地出口网络被防火墙拦截 OSS 域名（oss-cn-hangzhou.aliyuncs.com） |
-| 服务端看到任务回退到 pending 重试 | Worker 心跳超时（默认 5 分钟）；Worker 进程崩溃或推理卡死超过 5 分钟 |
-| 日志 `MuseTalk 推理超时` | 视频太长或显存不足；可临时调大 REQUEST_TIMEOUT，或限制输入视频时长 |
+| 服务端看到任务回退到 pending 重试 | 任务心跳真正超时（默认 900s）：Worker 进程崩溃/断网，或推理彻底卡死；正常长推理期间心跳线程每 30s 续期，不会回退 |
+| 日志 `MuseTalk 推理超时或连接失败` | 视频太长或显存不足；可临时调大 REQUEST_TIMEOUT（服务端 GPU_TASK_TIMEOUT_SECONDS 需同步调大），或限制输入视频时长 |
+| 日志 `视频过短（x.xxs < 3s）` | 输入视频不足 3s，MuseTalk 对短视频会 division by zero，已在本地直接上报失败；可用 MIN_VIDEO_DURATION_SECONDS 调整阈值 |
 
 ## 七、安全注意事项
 

deploy/gpu_worker/gpu_worker.py

@@ -9,9 +9,12 @@
   WORKER_ID           本机唯一 ID（默认 hostname+网卡MAC 后4位）
   MUSE_TALK_URL       本地 MuseTalk 地址，默认 http://127.0.0.1:7861
   POLL_INTERVAL       轮询间隔秒，默认 5
-  HEARTBEAT_INTERVAL  心跳间隔秒，默认 15
-  REQUEST_TIMEOUT     HTTP 请求超时秒，默认 60
-  TASK_MAX_RETRY      单个任务最大重试次数（在 Worker 本地的重试），默认 2
+  HEARTBEAT_INTERVAL  空闲心跳间隔秒，默认 15
+  REQUEST_TIMEOUT     HTTP 请求超时秒（下载/推理/上传统一使用），默认 900
+                      需与服务端 GPU_TASK_TIMEOUT_SECONDS（默认 900）对齐
+  TASK_MAX_RETRY      单任务本地最大重试次数（仅对瞬时错误重试），默认 1
+  TASK_HEARTBEAT_INTERVAL  推理期间任务心跳间隔秒，默认 30
+  MIN_VIDEO_DURATION_SECONDS  最短输入视频时长秒，小于则直接上报失败，默认 3
 
 用法：
   python gpu_worker.py
@@ -19,13 +22,13 @@
 
 from __future__ import annotations
 
-import json
 import logging
 import os
 import platform
 import socket
 import sys
 import tempfile
+import threading
 import time
 import uuid
 from pathlib import Path
@@ -54,8 +57,17 @@ class Config:
     muse_talk_url: str = _env("MUSE_TALK_URL", "http://127.0.0.1:7861").rstrip("/")
     poll_interval: float = float(_env("POLL_INTERVAL", "5"))
     heartbeat_interval: float = float(_env("HEARTBEAT_INTERVAL", "15"))
-    request_timeout: float = float(_env("REQUEST_TIMEOUT", "300"))
-    task_max_retry: int = int(_env("TASK_MAX_RETRY", "2"))
+    # #1970：RTX2060 6G 处理 720p 长视频可能 >5min；与服务端
+    # GPU_TASK_TIMEOUT_SECONDS 默认值对齐为 900，避免推理被本地/服务端先掐断。
+    request_timeout: float = float(_env("REQUEST_TIMEOUT", "900"))
+    # 本地只在网络/MuseTalk 瞬时错误时重试 1 次；服务端 MAX_ATTEMPTS=3
+    # 负责跨 worker/真正超时后的重派发，总尝试次数不再相乘放大。
+    task_max_retry: int = int(_env("TASK_MAX_RETRY", "1"))
+    # 推理期间任务心跳间隔（独立线程 POST /gpu/register 带 task_id）
+    task_heartbeat_interval: float = float(_env("TASK_HEARTBEAT_INTERVAL", "30"))
+    # 输入视频最短时长（秒）：过短（如 1s）MuseTalk 会 division by zero，
+    # 本地前置拦截，直接上报 failed，不浪费 GPU 时间
+    min_video_duration_seconds: float = float(_env("MIN_VIDEO_DURATION_SECONDS", "3"))
     worker_id: str = _env("WORKER_ID", "")
 
     @classmethod
@@ -96,8 +108,12 @@ def _check_musetalk_health() -> tuple[bool, dict]:
         return False, {"error": str(exc)}
 
 
-def _register() -> bool:
-    """向服务端注册 / 心跳，附带 GPU 信息."""
+def _register(task_id: Optional[str] = None) -> bool:
+    """向服务端注册 / 心跳，附带 GPU 信息。
+
+    推理期间的心跳线程传 task_id：服务端会同步刷新该 processing 任务的
+    last_heartbeat_at，防止长推理被误判超时回收。
+    """
     ok, info = _check_musetalk_health()
     free_vram = int(info.get("free_vram_mb", 0) or 0) if isinstance(info, dict) else 0
     gpu_name = info.get("gpu_name", "") if isinstance(info, dict) else ""
@@ -111,6 +127,8 @@ def _register() -> bool:
         "free_vram_mb": free_vram,
         "capabilities": "musetalk",
     }
+    if task_id:
+        payload["task_id"] = task_id
     try:
         r = requests.post(
             f"{Config.api_base_url}/api/v1/gpu/register",
@@ -181,11 +199,13 @@ def _download(url: str, path: Path) -> bool:
         return False
 
 
-def _call_musetalk(video_path: Path, audio_path: Path, out_path: Path) -> tuple[bool, float, str]:
+def _call_musetalk(video_path: Path, audio_path: Path, out_path: Path) -> tuple[bool, float, str, bool]:
     """调用本地 MuseTalk /inference.
 
-    返回 (success, duration_seconds, error_msg).
+    返回 (success, duration_seconds, error_msg, retryable)。
     duration 用 ffprobe 读结果视频，失败填 0。
+    retryable 仅对瞬时错误（连接失败/超时/5xx）为 True；HTTP 4xx、结果过小
+    等确定性失败不重试，直接上报服务端（服务端 MAX_ATTEMPTS 再决定是否重派发）。
     """
     try:
         with open(video_path, "rb") as vf, open(audio_path, "rb") as af:
@@ -199,17 +219,20 @@ def _call_musetalk(video_path: Path, audio_path: Path, out_path: Path) -> tuple[
                 timeout=Config.request_timeout,
             )
         if r.status_code != 200:
-            return False, 0.0, f"MuseTalk HTTP {r.status_code}: {r.text[:500]}"
+            retryable = r.status_code >= 500
+            return False, 0.0, f"MuseTalk HTTP {r.status_code}: {r.text[:500]}", retryable
         out_path.parent.mkdir(parents=True, exist_ok=True)
         out_path.write_bytes(r.content)
         if out_path.stat().st_size < 1024:
-            return False, 0.0, f"MuseTalk 返回结果过小 ({out_path.stat().st_size} bytes)"
+            # 确定性失败（推理产物异常），本地重试大概率还是坏的，不重试
+            return False, 0.0, f"MuseTalk 返回结果过小 ({out_path.stat().st_size} bytes)", False
         duration = _probe_duration(out_path)
-        return True, duration, ""
-    except requests.exceptions.Timeout:
-        return False, 0.0, f"MuseTalk 推理超时（>{Config.request_timeout}s）"
+        return True, duration, "", False
+    except (requests.exceptions.Timeout, requests.exceptions.ConnectionError):
+        # 瞬时网络/超时错误，允许本地重试 1 次
+        return False, 0.0, f"MuseTalk 推理超时或连接失败（>{Config.request_timeout}s）", True
     except Exception as exc:
-        return False, 0.0, f"MuseTalk 调用异常: {exc}"
+        return False, 0.0, f"MuseTalk 调用异常: {exc}", False
 
 
 def _probe_duration(path: Path) -> float:
@@ -219,9 +242,13 @@ def _probe_duration(path: Path) -> float:
 
         out = subprocess.check_output(
             [
-                "ffprobe", "-v", "error",
-                "-show_entries", "format=duration",
-                "-of", "default=noprint_wrappers=1:nokey=1",
+                "ffprobe",
+                "-v",
+                "error",
+                "-show_entries",
+                "format=duration",
+                "-of",
+                "default=noprint_wrappers=1:nokey=1",
                 str(path),
             ],
             stderr=subprocess.DEVNULL,
@@ -276,42 +303,91 @@ def _report_result(task_id: str, success: bool, duration: float = 0.0, error_msg
         return False
 
 
+class TaskHeartbeat(threading.Thread):
+    """推理期间的任务心跳线程。
+
+    主循环的空闲心跳在 ``_handle_task`` 同步阻塞（下载/推理/上传最长 900s）
+    期间无法发送，服务端会因任务 last_heartbeat_at 停滞而误判超时回退 pending。
+    本线程每 task_heartbeat_interval 秒（默认 30s）POST /gpu/register 并
+    携带当前 task_id，让服务端持续续期任务心跳；任务处理结束 stop()。
+    """
+
+    def __init__(self, task_id: str, interval: float):
+        super().__init__(daemon=True, name=f"hb-{task_id[:8]}")
+        self.task_id = task_id
+        self.interval = max(5.0, interval)
+        self._stop_event = threading.Event()
+
+    def run(self) -> None:
+        # 先立即发一次，再按间隔循环（首次心跳失败不影响主流程）
+        while not self._stop_event.is_set():
+            try:
+                if _register(self.task_id):
+                    logger.debug("任务 %s 心跳已发送", self.task_id)
+            except Exception as exc:  # noqa: BLE001
+                logger.warning("任务 %s 心跳异常（忽略）: %s", self.task_id, exc)
+            self._stop_event.wait(self.interval)
+
+    def stop(self) -> None:
+        self._stop_event.set()
+
+
 def _handle_task(task: dict) -> None:
-    """处理一条任务（整个串行流程：下载→推理→上传→上报）."""
+    """处理一条任务（整个串行流程：下载→时长校验→推理→上传→上报）。"""
     task_id = task["task_id"]
     logger.info("开始处理任务 %s", task_id)
-    with tempfile.TemporaryDirectory(prefix="musetalk_") as tmpdir:
-        tmp = Path(tmpdir)
-        video_path = tmp / "input.mp4"
-        audio_path = tmp / "input_audio.bin"
-        out_path = tmp / "output.mp4"
+    # 领取任务后立即启动任务级心跳线程，覆盖下载/推理/上报全过程
+    hb = TaskHeartbeat(task_id, Config.task_heartbeat_interval)
+    hb.start()
+    try:
+        with tempfile.TemporaryDirectory(prefix="musetalk_") as tmpdir:
+            tmp = Path(tmpdir)
+            video_path = tmp / "input.mp4"
+            audio_path = tmp / "input_audio.bin"
+            out_path = tmp / "output.mp4"
 
-        # 1. 下载
-        if not _download(task["video_url"], video_path):
-            _report_result(task_id, False, 0.0, "下载人物视频失败")
-            return
-        if not _download(task["audio_url"], audio_path):
-            _report_result(task_id, False, 0.0, "下载驱动音频失败")
-            return
+            # 1. 下载
+            if not _download(task["video_url"], video_path):
+                _report_result(task_id, False, 0.0, "下载人物视频失败")
+                return
+            if not _download(task["audio_url"], audio_path):
+                _report_result(task_id, False, 0.0, "下载驱动音频失败")
+                return
 
-        # 2. 推理（本地重试）
-        success = False
-        duration = 0.0
-        err = ""
-        for attempt in range(Config.task_max_retry + 1):
-            if attempt > 0:
-                logger.info("任务 %s 第 %d 次重试...", task_id, attempt + 1)
-                time.sleep(2)
-            success, duration, err = _call_musetalk(video_path, audio_path, out_path)
-            if success:
-                break
-        if not success:
-            logger.error("任务 %s 推理失败: %s", task_id, err)
-            _report_result(task_id, False, 0.0, err)
-            return
+            # 2. 输入时长前置校验：短视频 MuseTalk 会 division by zero，
+            #    直接上报 failed，不浪费 GPU 时间。ffprobe 不可用/读失败（0.0）
+            #    时不拦截，交给 MuseTalk 处理，避免误杀。
+            video_duration = _probe_duration(video_path)
+            if video_duration and video_duration < Config.min_video_duration_seconds:
+                msg = (
+                    f"视频过短（{video_duration:.2f}s < {Config.min_video_duration_seconds:.0f}s），"
+                    "MuseTalk 无法处理"
+                )
+                logger.error("任务 %s %s", task_id, msg)
+                _report_result(task_id, False, 0.0, msg)
+                return
 
-        # 3. 上报结果（multipart 同时上传文件 → API 代为 PUT 到 OSS，逻辑最稳）
-        _report_success_with_file(task_id, duration, out_path)
+            # 3. 推理（本地仅对瞬时错误重试）
+            success = False
+            duration = 0.0
+            err = ""
+            retryable = False
+            for attempt in range(Config.task_max_retry + 1):
+                if attempt > 0:
+                    logger.info("任务 %s 第 %d 次重试（瞬时错误）...", task_id, attempt + 1)
+                    time.sleep(2)
+                success, duration, err, retryable = _call_musetalk(video_path, audio_path, out_path)
+                if success or not retryable:
+                    break
+            if not success:
+                logger.error("任务 %s 推理失败: %s", task_id, err)
+                _report_result(task_id, False, 0.0, err)
+                return
+
+            # 4. 上报结果（multipart 同时上传文件 → API 代为 PUT 到 OSS，逻辑最稳）
+            _report_success_with_file(task_id, duration, out_path)
+    finally:
+        hb.stop()
 
 
 def _report_success_with_file(task_id: str, duration: float, file_path: Path) -> None:

packages/adapters/sqlalchemy_impl/asset_atom_clip_repository.py

@@ -91,15 +91,21 @@ class SQLAlchemyAssetAtomClipRepository:
         self.session.commit()
         return count > 0
 
-    def find_untagged(self, limit: int = 100) -> list[AssetAtomClip]:
-        """查找 ai_tags IS NULL 的片段，用于回填."""
-        models = (
-            self.session.query(AssetAtomClipModel)
-            .filter(AssetAtomClipModel.ai_tags.is_(None))
-            .order_by(AssetAtomClipModel.created_at.asc())
-            .limit(limit)
-            .all()
-        )
+    def find_untagged(self, limit: int = 100, include_downgraded: bool = False) -> list[AssetAtomClip]:
+        """查找未完成 AI 打标的片段，用于回填.
+
+        默认仅匹配 ai_tags IS NULL；include_downgraded=True 时额外包含
+        只有 inherited_tags 的降级记录（视觉 API 失败时写入，无 has_text 字段），
+        供强制回填（#1970 force backfill）使用。
+        """
+        query = self.session.query(AssetAtomClipModel)
+        if include_downgraded:
+            # as_string() → JSON/JSONB ->> 取值；NULL 记录或缺 has_text 键
+            # （降级记录）均为 NULL，has_text 为 true/false 的完整记录被排除
+            query = query.filter(AssetAtomClipModel.ai_tags["has_text"].as_string().is_(None))
+        else:
+            query = query.filter(AssetAtomClipModel.ai_tags.is_(None))
+        models = query.order_by(AssetAtomClipModel.created_at.asc()).limit(limit).all()
         return [self._to_domain(m) for m in models]
 
     def _to_model(self, clip: AssetAtomClip) -> AssetAtomClipModel:

packages/config/base.py

@@ -84,8 +84,11 @@ class SharedSettings(BaseSettings):
     # Worker 用这个长期 Token 鉴权（不是用户 JWT）。多 Worker 共用同一个 Token；
     # worker_id 用于区分具体机器。生产必须配置；development 留空会跳过校验。
     gpu_worker_token: str = ""
-    # GPU 任务超时（秒）：超过此时长仍未完成则标记为 failed，可重新 poll
-    gpu_task_timeout_seconds: int = 300
+    # GPU 任务超时（秒）：processing 状态超过此时长（以任务心跳为准）才回退
+    # pending / failed。#1970：RTX2060 6G 推理 720p 长视频需 5 分钟以上，300→900。
+    # Worker 推理期间每 30s 通过 /gpu/register(task_id=...) 续心跳，
+    # 只有真正超时或 Worker 明确上报 failed 才会回退。
+    gpu_task_timeout_seconds: int = 900
     # 结果预签名 URL 有效期（秒）
     gpu_result_url_expires: int = 3600
     # 输入预签名 URL 有效期（秒，需留出 Worker 下载时间）

scripts/render_env.sh

@@ -57,7 +57,7 @@ if [ "$TARGET_ENV" = "staging" ]; then
 fi
 
 # 共用 secrets 直接导出（如果存在）
-SHARED_SECRETS="OSS_ACCESS_KEY_ID OSS_ACCESS_KEY_SECRET COSYVOICE_API_KEY DASHSCOPE_API_KEY MEDIAKIT_API_KEY DOUBAO_API_KEY DOUBAO_MODEL DOUBAO_BASE_URL WECHAT_APP_ID WECHAT_APP_SECRET TIKHUB_API_KEY APIZERO_API_KEY GPU_WORKER_TOKEN"
+SHARED_SECRETS="OSS_ACCESS_KEY_ID OSS_ACCESS_KEY_SECRET COSYVOICE_API_KEY DASHSCOPE_API_KEY MEDIAKIT_API_KEY DOUBAO_API_KEY DOUBAO_MODEL DOUBAO_BASE_URL DOUBAO_VISION_MODEL WECHAT_APP_ID WECHAT_APP_SECRET TIKHUB_API_KEY APIZERO_API_KEY GPU_WORKER_TOKEN"
 for var in $SHARED_SECRETS; do
   value="${!var:-}"
   # 已经在环境中了，无需额外操作

tests/unit/test_1970_celery_ai_tag_registration.py

@@ -0,0 +1,82 @@
+"""#1970 AI 标签 Celery 任务注册回归测试。
+
+背景：staging 上 worker.generate_atom_clips 正常派发 tag_atom_clip，
+但消费端报 "Received unregistered task of type 'worker.tag_atom_clip'"，
+根因是 celery_app.conf.imports 漏列任务模块，worker 进程从未 import 之。
+
+注意：tests/unit 下大量旧测试在 import 期向 sys.modules 注入
+worker_app.celery_app 的 MagicMock 且不还原，全量收集时会污染本测试，
+因此这里用 AST 静态解析 + 隔离子进程验证，不依赖 sys.modules 状态。
+"""
+
+from __future__ import annotations
+
+import ast
+import os
+import subprocess
+import sys
+from pathlib import Path
+
+REPO_ROOT = Path(__file__).resolve().parents[2]
+CELERY_APP_PY = REPO_ROOT / "apps" / "worker" / "worker_app" / "celery_app.py"
+
+REQUIRED_MODULES = (
+    "worker_app.tasks.atom_clip_tagging",
+    "worker_app.tasks.backfill_atom_clip_tags",
+)
+
+
+def _conf_imports_values() -> set[str]:
+    """从 celery_app.py AST 中提取 celery_app.conf.imports 元组的字符串项。"""
+    tree = ast.parse(CELERY_APP_PY.read_text(encoding="utf-8"))
+    values: set[str] = set()
+    for node in ast.walk(tree):
+        if not (isinstance(node, ast.Assign) and len(node.targets) == 1):
+            continue
+        target = node.targets[0]
+        # celery_app.conf.imports = (...) 或 conf.imports = (...)
+        if not (isinstance(target, ast.Attribute) and target.attr == "imports"):
+            continue
+        if isinstance(node.value, (ast.Tuple, ast.List)):
+            for elt in node.value.elts:
+                if isinstance(elt, ast.Constant) and isinstance(elt.value, str):
+                    values.add(elt.value)
+    return values
+
+
+def test_ai_tag_modules_in_celery_imports():
+    imports = _conf_imports_values()
+    for module in REQUIRED_MODULES:
+        assert module in imports, f"{module} 未加入 celery_app.conf.imports"
+
+
+def test_ai_tag_tasks_registered_in_isolated_process():
+    """隔离子进程（无 conftest / 无 sys.modules mock）真实加载 Celery app。"""
+    # 模拟 worker 启动时按 conf.imports import 任务模块的行为；
+    # 只导入 AI 标签两个模块（其他模块依赖 cv2 等本地未安装的重依赖）。
+    code = (
+        "import importlib, sys; "
+        "from worker_app.celery_app import celery_app; "
+        "mods = [m for m in celery_app.conf.imports or () "
+        "if 'atom_clip_tagging' in m or 'backfill_atom_clip_tags' in m]; "
+        "[importlib.import_module(m) for m in mods]; "
+        "missing = [n for n in "
+        "['worker.tag_atom_clip', 'worker.backfill_atom_clip_tags'] "
+        "if n not in celery_app.tasks]; "
+        "sys.exit(1 if missing or len(mods) < 2 else 0)"
+    )
+    env = os.environ.copy()
+    paths = [
+        str(REPO_ROOT),
+        str(REPO_ROOT / "apps" / "worker"),
+        str(REPO_ROOT / "packages"),
+    ]
+    env["PYTHONPATH"] = os.pathsep.join(paths) + os.pathsep + env.get("PYTHONPATH", "")
+    result = subprocess.run(
+        [sys.executable, "-c", code],
+        capture_output=True,
+        text=True,
+        env=env,
+        timeout=60,
+    )
+    assert result.returncode == 0, "隔离子进程中任务未注册成功:\n" f"stdout={result.stdout}\nstderr={result.stderr}"

tests/unit/test_1970_force_backfill_tags.py

@@ -0,0 +1,347 @@
+"""#1970 force 回填降级 AI 标签记录的回归测试。
+
+背景：DOUBAO_VISION_MODEL 未配置时，tagger 降级写入
+{"inherited_tags": [...]}（非 NULL），默认 backfill 只捞 ai_tags IS NULL，
+这批记录永远不会重打。force=True 时应纳入降级记录，并在打标成功后覆盖。
+
+覆盖：
+- find_untagged(include_downgraded) 的 SQL 过滤（SQLite 验证跨库 JSON 取值）
+- tag_atom_clip_task 的 force 跳过/放行/覆盖逻辑
+- backfill_atom_clip_tags(force=True) 给 tag 任务传 kwargs={"force": True}
+"""
+
+from __future__ import annotations
+
+from datetime import UTC, datetime
+from types import SimpleNamespace
+
+import pytest
+from sqlalchemy import create_engine
+from sqlalchemy.orm import sessionmaker
+
+from packages.adapters.sqlalchemy_impl.asset_atom_clip_repository import (
+    SQLAlchemyAssetAtomClipRepository,
+)
+from packages.adapters.sqlalchemy_impl.models import AssetAtomClipModel
+
+# ── 仓储层：find_untagged 过滤 ─────────────────────────────────────────────
+
+
+@pytest.fixture
+def repo_session():
+    engine = create_engine("sqlite:///:memory:")
+    AssetAtomClipModel.__table__.create(engine)
+    SessionTest = sessionmaker(bind=engine)
+    session = SessionTest()
+    now = datetime.now(UTC)
+    session.add_all(
+        [
+            AssetAtomClipModel(
+                id="c-null",
+                asset_id="a1",
+                start_time=0,
+                end_time=1,
+                duration=1,
+                clip_index=0,
+                tags=[],
+                ai_tags=None,
+                created_at=now,
+            ),
+            AssetAtomClipModel(
+                id="c-downgraded-empty",
+                asset_id="a1",
+                start_time=1,
+                end_time=2,
+                duration=1,
+                clip_index=1,
+                tags=[],
+                ai_tags={"inherited_tags": []},
+                created_at=now,
+            ),
+            AssetAtomClipModel(
+                id="c-downgraded-tags",
+                asset_id="a1",
+                start_time=2,
+                end_time=3,
+                duration=1,
+                clip_index=2,
+                tags=[],
+                ai_tags={"inherited_tags": ["口播"]},
+                created_at=now,
+            ),
+            AssetAtomClipModel(
+                id="c-tagged-true",
+                asset_id="a1",
+                start_time=3,
+                end_time=4,
+                duration=1,
+                clip_index=3,
+                tags=[],
+                ai_tags={"has_text": True, "scene": ["室内"], "inherited_tags": []},
+                created_at=now,
+            ),
+            AssetAtomClipModel(
+                id="c-tagged-false",
+                asset_id="a1",
+                start_time=4,
+                end_time=5,
+                duration=1,
+                clip_index=4,
+                tags=[],
+                ai_tags={"has_text": False, "inherited_tags": ["风景"]},
+                created_at=now,
+            ),
+        ]
+    )
+    session.commit()
+    # SQLAlchemy JSON 在 SQLite 下把 None 序列化为 'null' 字符串，
+    # 而生产 PostgreSQL 存的是真 SQL NULL；用原生 SQL 对齐生产语义。
+    from sqlalchemy import text
+
+    session.execute(text("UPDATE asset_atom_clips SET ai_tags = NULL WHERE id = 'c-null'"))
+    session.commit()
+    yield session
+    session.close()
+
+
+def test_find_untagged_default_only_null(repo_session):
+    repo = SQLAlchemyAssetAtomClipRepository(repo_session)
+    ids = {c.id for c in repo.find_untagged(limit=100)}
+    assert ids == {"c-null"}
+
+
+def test_find_untagged_include_downgraded(repo_session):
+    repo = SQLAlchemyAssetAtomClipRepository(repo_session)
+    ids = {c.id for c in repo.find_untagged(limit=100, include_downgraded=True)}
+    # NULL + 两条降级记录；含 has_text=true/false 的完整记录都排除
+    assert ids == {"c-null", "c-downgraded-empty", "c-downgraded-tags"}
+
+
+# ── 任务层：tag_atom_clip_task 的 force 语义 ───────────────────────────────
+
+
+def _import_tag_task_module():
+    from worker_app.tasks import atom_clip_tagging as mod
+
+    return mod
+
+
+def _call_tag_task(mod, clip_id, force):
+    """直接调用任务，兼容两种环境。
+
+    全量收集时旧测试向 sys.modules 注入 celery_app MagicMock（其 task
+    装饰器原样返回裸函数），此时是普通函数需显式传 self=None；
+    正常 Celery 环境下属性是 Task 代理对象（非普通 function），
+    已绑定 self，按业务签名直接调用即可。
+    """
+    import inspect
+
+    obj = mod.tag_atom_clip_task
+    if inspect.isfunction(obj):
+        return obj(None, clip_id, force=force)
+    return obj(clip_id, force=force)
+
+
+def test_tag_task_skips_downgraded_without_force(monkeypatch):
+    mod = _import_tag_task_module()
+    monkeypatch.setattr(
+        mod,
+        "SessionLocal",
+        lambda: SimpleNamespace(
+            rollback=lambda: None,
+            close=lambda: None,
+        ),
+    )
+
+    class _Repo:
+        def __init__(self, db):
+            pass
+
+        def find_by_id(self, clip_id):
+            return SimpleNamespace(
+                id=clip_id,
+                ai_tags={"inherited_tags": []},
+            )
+
+    monkeypatch.setattr(mod, "SQLAlchemyAssetAtomClipRepository", _Repo)
+
+    result = _call_tag_task(mod, "clip-downgraded", force=False)
+    assert result["status"] == "skipped"
+    assert result["reason"] == "already tagged"
+
+
+def test_tag_task_force_retags_downgraded_and_overwrites(monkeypatch):
+    mod = _import_tag_task_module()
+    updated: dict[str, dict] = {}
+
+    class _FakeSession:
+        def rollback(self):
+            pass
+
+        def close(self):
+            pass
+
+    monkeypatch.setattr(mod, "SessionLocal", _FakeSession)
+
+    class _AtomRepo:
+        def __init__(self, db):
+            pass
+
+        def find_by_id(self, clip_id):
+            return SimpleNamespace(
+                id=clip_id,
+                asset_id="asset-1",
+                start_time=0.0,
+                end_time=2.0,
+                tags=["旧标签"],
+                ai_tags={"inherited_tags": []},
+            )
+
+        def update_ai_tags(self, clip_id, ai_tags):
+            updated[clip_id] = ai_tags
+
+    class _AssetRepo:
+        def __init__(self, db):
+            pass
+
+        def find_by_id(self, asset_id):
+            return SimpleNamespace(id=asset_id, storage_key="k/video.mp4")
+
+    monkeypatch.setattr(mod, "SQLAlchemyAssetAtomClipRepository", _AtomRepo)
+    monkeypatch.setattr(mod, "SQLAlchemyAssetRepository", _AssetRepo)
+
+    class _Storage:
+        def get_download_url(self, key, expires_seconds=3600):
+            return "https://example.com/signed.mp4"
+
+    monkeypatch.setattr(mod, "get_shared_storage_service", lambda: _Storage())
+    monkeypatch.setattr(mod, "get_doubao_client", lambda: object())
+    monkeypatch.setattr(mod, "get_mediakit_client", lambda: None)
+
+    new_tags = {
+        "scene": ["室内"],
+        "objects": ["人物"],
+        "action": ["说话"],
+        "shot": "中景",
+        "has_text": True,
+        "inherited_tags": ["旧标签"],
+    }
+    monkeypatch.setattr(mod, "tag_atom_clip", lambda **kw: new_tags)
+
+    result = _call_tag_task(mod, "clip-downgraded", force=True)
+    assert result["status"] == "completed"
+    assert result["has_ai_tags"] is True
+    assert updated["clip-downgraded"] == new_tags
+
+
+def test_tag_task_force_still_skips_complete_tags(monkeypatch):
+    mod = _import_tag_task_module()
+    monkeypatch.setattr(
+        mod,
+        "SessionLocal",
+        lambda: SimpleNamespace(rollback=lambda: None, close=lambda: None),
+    )
+
+    class _Repo:
+        def __init__(self, db):
+            pass
+
+        def find_by_id(self, clip_id):
+            return SimpleNamespace(
+                id=clip_id,
+                ai_tags={"has_text": False, "inherited_tags": []},
+            )
+
+    monkeypatch.setattr(mod, "SQLAlchemyAssetAtomClipRepository", _Repo)
+
+    result = _call_tag_task(mod, "clip-complete", force=True)
+    assert result["status"] == "skipped"
+    assert result["reason"] == "already tagged"
+
+
+# ── backfill 任务：force 透传到 send_task ──────────────────────────────────
+
+
+def test_backfill_force_passes_kwarg(monkeypatch):
+    from worker_app.tasks import backfill_atom_clip_tags as bmod
+
+    sent: list[tuple] = []
+
+    class _FakeSession:
+        def close(self):
+            pass
+
+    monkeypatch.setattr(bmod, "SessionLocal", _FakeSession)
+
+    class _AtomRepo:
+        def __init__(self, db):
+            self.calls: list[bool] = []
+
+        def find_untagged(self, limit, include_downgraded=False):
+            self.calls.append(include_downgraded)
+            # 第一批返回一条降级记录，第二批返回空结束循环
+            if len(self.calls) == 1:
+                return [SimpleNamespace(id="clip-1")]
+            return []
+
+    repo_holder = {}
+
+    def _repo_factory(db):
+        repo = _AtomRepo(db)
+        repo_holder["repo"] = repo
+        return repo
+
+    monkeypatch.setattr(bmod, "SQLAlchemyAssetAtomClipRepository", _repo_factory)
+
+    def _send_task(name, args=None, kwargs=None):
+        sent.append((name, args, kwargs))
+
+    monkeypatch.setattr(bmod.celery_app, "send_task", _send_task)
+
+    result = bmod.backfill_atom_clip_tags(batch_size=10, batch_interval=0, force=True)
+
+    assert result["status"] == "completed"
+    assert result["total_submitted"] == 1
+    assert repo_holder["repo"].calls == [True, True]
+    assert sent == [
+        ("worker.tag_atom_clip", ["clip-1"], {"force": True}),
+    ]
+
+
+def test_backfill_default_does_not_force(monkeypatch):
+    from worker_app.tasks import backfill_atom_clip_tags as bmod
+
+    sent_kwargs: list[dict | None] = []
+
+    class _FakeSession:
+        def close(self):
+            pass
+
+    monkeypatch.setattr(bmod, "SessionLocal", _FakeSession)
+
+    class _AtomRepo:
+        def __init__(self, db):
+            self.calls: list[bool] = []
+
+        def find_untagged(self, limit, include_downgraded=False):
+            self.calls.append(include_downgraded)
+            return [SimpleNamespace(id="clip-null")] if self.calls == [False] else []
+
+    holder = {}
+
+    def _repo_factory(db):
+        holder["repo"] = _AtomRepo(db)
+        return holder["repo"]
+
+    monkeypatch.setattr(bmod, "SQLAlchemyAssetAtomClipRepository", _repo_factory)
+    monkeypatch.setattr(
+        bmod.celery_app,
+        "send_task",
+        lambda name, args=None, kwargs=None: sent_kwargs.append(kwargs),
+    )
+
+    result = bmod.backfill_atom_clip_tags(batch_size=10, batch_interval=0)
+
+    assert result["total_submitted"] == 1
+    assert holder["repo"].calls == [False, False]
+    assert sent_kwargs == [{"force": False}]

tests/unit/test_1970_gpu_worker_heartbeat.py

@@ -0,0 +1,254 @@
+"""#1970 GPU Worker 修复单测.
+
+覆盖 deploy/gpu_worker/gpu_worker.py（独立部署脚本，不在 apps/packages 包内，
+按文件路径动态加载）：
+1. 默认配置：REQUEST_TIMEOUT=900 / TASK_MAX_RETRY=1 / 心跳 30s / 最短 3s；
+2. 推理期心跳线程 POST /gpu/register 带 task_id，任务结束能停；
+3. <3s 短视频直接上报失败，不调用 MuseTalk；
+4. _call_musetalk 仅对 5xx/网络瞬时错误标记 retryable，4xx 不重试；
+5. _handle_task 只对 retryable 错误本地重试 1 次。
+"""
+
+from __future__ import annotations
+
+import importlib.util
+import os
+import sys
+import time
+from pathlib import Path
+from unittest import mock
+
+import pytest
+
+ROOT = Path(__file__).resolve().parents[2]
+WORKER_PATH = ROOT / "deploy" / "gpu_worker" / "gpu_worker.py"
+
+
+def _load_worker_module():
+    spec = importlib.util.spec_from_file_location("gpu_worker_standalone_1970", WORKER_PATH)
+    mod = importlib.util.module_from_spec(spec)
+    sys.modules[spec.name] = mod
+    spec.loader.exec_module(mod)
+    return mod
+
+
+@pytest.fixture
+def worker():
+    return _load_worker_module()
+
+
+# ── 默认配置 ───────────────────────────────────────────────────────
+
+
+def test_config_defaults_900_and_retry_one(monkeypatch):
+    """CI/本机若显式导出过这些 env，说明是运维覆盖，不应拿默认值断言；
+    因此只在四个 env 全部缺失时校验脚本内置默认值（#1970：900/1/30/3）。"""
+    keys = (
+        "REQUEST_TIMEOUT",
+        "TASK_MAX_RETRY",
+        "TASK_HEARTBEAT_INTERVAL",
+        "MIN_VIDEO_DURATION_SECONDS",
+    )
+    if any(k in os.environ for k in keys):
+        pytest.skip("环境显式设置了 worker 超时/重试变量，跳过默认值断言")
+    for key in keys:
+        monkeypatch.delenv(key, raising=False)
+    mod = _load_worker_module()
+    assert mod.Config.request_timeout == 900.0
+    assert mod.Config.task_max_retry == 1
+    assert mod.Config.task_heartbeat_interval == 30.0
+    assert mod.Config.min_video_duration_seconds == 3.0
+
+
+# ── register 携带 task_id ──────────────────────────────────────────
+
+
+def test_register_payload_includes_task_id_only_when_provided(worker, monkeypatch):
+    captured = []
+
+    class _Resp:
+        status_code = 200
+        text = ""
+
+    def _fake_post(url, json=None, headers=None, timeout=None):
+        captured.append(json)
+        return _Resp()
+
+    monkeypatch.setattr(worker.requests, "post", _fake_post)
+    monkeypatch.setattr(worker, "_check_musetalk_health", lambda: (True, {}))
+
+    assert worker._register("task-abc") is True
+    assert captured[-1]["task_id"] == "task-abc"
+    assert captured[-1]["worker_id"]
+
+    worker._register()  # 空闲心跳不带 task_id
+    assert "task_id" not in captured[-1]
+
+
+# ── 推理期心跳线程 ─────────────────────────────────────────────────
+
+
+def test_task_heartbeat_thread_sends_and_stops(worker, monkeypatch):
+    calls = []
+
+    def _fake_register(task_id=None):
+        calls.append(task_id)
+        return True
+
+    monkeypatch.setattr(worker, "_register", _fake_register)
+    hb = worker.TaskHeartbeat("task-hb1", interval=5)
+    hb.start()
+    time.sleep(0.3)  # 启动后立即发一次
+    hb.stop()
+    hb.join(timeout=2)
+    assert not hb.is_alive()
+    assert calls and all(c == "task-hb1" for c in calls)
+
+
+# ── 短视频前置拦截 ─────────────────────────────────────────────────
+
+
+def test_handle_task_short_video_reports_failed_without_inference(worker, monkeypatch, tmp_path):
+    video = tmp_path / "input.mp4"
+    video.write_bytes(b"fake-mp4-bytes")
+    audio = tmp_path / "input_audio.bin"
+    audio.write_bytes(b"fake-audio")
+    reports = []
+
+    monkeypatch.setattr(worker, "_register", lambda *a, **k: True)
+    monkeypatch.setattr(worker, "_download", lambda url, path: True)
+    # ffprobe 读出 1.2s → 低于 3s 阈值
+    monkeypatch.setattr(worker, "_probe_duration", lambda path: 1.2)
+
+    def _boom(*a, **k):
+        raise AssertionError("短视频不应调用 MuseTalk 推理")
+
+    monkeypatch.setattr(worker, "_call_musetalk", _boom)
+    monkeypatch.setattr(
+        worker,
+        "_report_result",
+        lambda task_id, success, duration=0.0, error_msg="": reports.append((task_id, success, error_msg)) or True,
+    )
+
+    task = {
+        "task_id": "task-short",
+        "video_url": "https://example.com/v.mp4",
+        "audio_url": "https://example.com/a.bin",
+    }
+    worker._handle_task(task)
+
+    assert len(reports) == 1
+    tid, ok, err = reports[0]
+    assert tid == "task-short"
+    assert ok is False
+    assert "视频过短" in err
+    assert "3" in err
+
+
+def test_handle_task_probe_failure_does_not_block(worker, monkeypatch):
+    """ffprobe 不可用（duration=0.0）时不能误杀，应继续推理."""
+    reports = []
+    monkeypatch.setattr(worker, "_register", lambda *a, **k: True)
+    monkeypatch.setattr(worker, "_download", lambda url, path: True)
+    monkeypatch.setattr(worker, "_probe_duration", lambda path: 0.0)
+    monkeypatch.setattr(
+        worker,
+        "_call_musetalk",
+        lambda v, a, o: (True, 8.0, "", False),
+    )
+    uploaded = []
+    monkeypatch.setattr(
+        worker,
+        "_report_success_with_file",
+        lambda task_id, duration, path: uploaded.append((task_id, duration)),
+    )
+    monkeypatch.setattr(worker, "_report_result", lambda *a, **k: True)
+
+    worker._handle_task({"task_id": "task-probe0", "video_url": "u", "audio_url": "u"})
+    assert uploaded == [("task-probe0", 8.0)]
+    assert reports == []
+
+
+# ── 重试语义：仅瞬时错误重试 ───────────────────────────────────────
+
+
+def test_call_musetalk_4xx_not_retryable_5xx_retryable(worker, monkeypatch, tmp_path):
+    video = tmp_path / "v.mp4"
+    audio = tmp_path / "a.bin"
+    video.write_bytes(b"v")
+    audio.write_bytes(b"a")
+    out = tmp_path / "o.mp4"
+
+    class _Resp:
+        def __init__(self, code, body=b"x" * 2048):
+            self.status_code = code
+            self.content = body
+            self.text = "err"
+
+    # 4xx：确定性失败，不重试
+    monkeypatch.setattr(worker.requests, "post", lambda *a, **k: _Resp(400))
+    ok, _, _, retryable = worker._call_musetalk(video, audio, out)
+    assert ok is False and retryable is False
+
+    monkeypatch.setattr(worker.requests, "post", lambda *a, **k: _Resp(503))
+    ok, _, _, retryable = worker._call_musetalk(video, audio, out)
+    assert ok is False and retryable is True
+
+    # 连接异常：瞬时错误，可重试
+    import requests as _requests
+
+    def _conn_err(*a, **k):
+        raise _requests.exceptions.ConnectionError("reset")
+
+    monkeypatch.setattr(worker.requests, "post", _conn_err)
+    ok, _, _, retryable = worker._call_musetalk(video, audio, out)
+    assert ok is False and retryable is True
+
+
+def test_handle_task_retries_once_for_transient_then_succeeds(worker, monkeypatch):
+    calls = []
+
+    def _fake_call(v, a, o):
+        calls.append(1)
+        if len(calls) == 1:
+            return False, 0.0, "MuseTalk HTTP 503: busy", True
+        return True, 6.5, "", False
+
+    monkeypatch.setattr(worker, "_register", lambda *a, **k: True)
+    monkeypatch.setattr(worker, "_download", lambda url, path: True)
+    monkeypatch.setattr(worker, "_probe_duration", lambda path: 12.0)
+    monkeypatch.setattr(worker, "_call_musetalk", _fake_call)
+    monkeypatch.setattr(worker, "time", mock.MagicMock())  # 重试 sleep 立即返回
+    uploaded = []
+    monkeypatch.setattr(
+        worker,
+        "_report_success_with_file",
+        lambda task_id, duration, path: uploaded.append((task_id, duration)),
+    )
+
+    worker._handle_task({"task_id": "t-retry", "video_url": "u", "audio_url": "u"})
+    assert len(calls) == 2
+    assert uploaded == [("t-retry", 6.5)]
+
+
+def test_handle_task_no_retry_for_deterministic_failure(worker, monkeypatch):
+    calls = []
+
+    def _fake_call(v, a, o):
+        calls.append(1)
+        return False, 0.0, "MuseTalk HTTP 400: bad input", False
+
+    reports = []
+    monkeypatch.setattr(worker, "_register", lambda *a, **k: True)
+    monkeypatch.setattr(worker, "_download", lambda url, path: True)
+    monkeypatch.setattr(worker, "_probe_duration", lambda path: 12.0)
+    monkeypatch.setattr(worker, "_call_musetalk", _fake_call)
+    monkeypatch.setattr(
+        worker,
+        "_report_result",
+        lambda task_id, success, duration=0.0, error_msg="": reports.append(error_msg) or True,
+    )
+
+    worker._handle_task({"task_id": "t-4xx", "video_url": "u", "audio_url": "u"})
+    assert len(calls) == 1  # 4xx 本地不重试，直接交服务端决定
+    assert reports and "400" in reports[0]

tests/unit/test_1970_hflip_ai_tags.py

@@ -0,0 +1,185 @@
+"""#1970 hflip 放开（has_text 来自 atom_clip.ai_tags）端到端参数链路测试。
+
+覆盖：
+1. UnifiedRenderService 传入 clip_has_text 后微变换计划的翻转门控；
+2. RenderAdapter._resolve_clip_has_text 按 atom_clip.ai_tags.has_text
+   解析布尔列表（显式 False 才可翻转，其余保守），失败回退 None；
+3. 纯函数层在「混合有/无文字」列表下的行为（顺序对齐）。
+"""
+
+from __future__ import annotations
+
+from types import SimpleNamespace
+from unittest.mock import MagicMock, patch
+
+import pytest
+from video_processing.micro_transform_pure import build_micro_transform_plan
+
+
+def _make_service(plan_config: dict | None = None, clip_has_text=None):
+    from video_processing.unified_render_service import UnifiedRenderService
+
+    svc = object.__new__(UnifiedRenderService)
+    svc.plan = MagicMock()
+    svc.plan.config = plan_config or {}
+    svc.plan.id = "plan-1"
+    svc.plan.clips = []
+    svc._micro_plan_cache = None
+    svc._micro_plan_loaded = False
+    svc._clip_has_text = clip_has_text
+    return svc
+
+
+def _clip(clip_id: str, atom_clip_id: str = "", clip_type: str = "main"):
+    return SimpleNamespace(id=clip_id, atom_clip_id=atom_clip_id, clip_type=clip_type)
+
+
+def _atom(clip_id: str, ai_tags):
+    return SimpleNamespace(id=clip_id, ai_tags=ai_tags)
+
+
+class TestServiceClipHasText:
+    def test_none_stays_conservative(self):
+        # 未注入检测列表：所有片段一律不翻转
+        svc = _make_service({"generation_task_id": "t1"}, clip_has_text=None)
+        plan = svc._get_micro_transform_plan(30)
+        assert plan is not None
+        assert all(c.has_text for c in plan.clips)
+        assert all(not c.hflip for c in plan.clips)
+
+    def test_explicit_no_text_allows_hflip(self):
+        # AI 明确判定无文字：允许参与 50% 翻转（40 段应至少出现一些翻转）
+        svc = _make_service({"generation_task_id": "t-allow"}, clip_has_text=[False] * 40)
+        plan = svc._get_micro_transform_plan(40)
+        assert plan is not None
+        assert all(not c.has_text for c in plan.clips)
+        assert any(c.hflip for c in plan.clips)
+        assert all(not c.hflip or not c.has_text for c in plan.clips)
+
+    def test_all_text_never_flips(self):
+        svc = _make_service({"generation_task_id": "t-text"}, clip_has_text=[True] * 40)
+        plan = svc._get_micro_transform_plan(40)
+        assert all(c.has_text for c in plan.clips)
+        assert all(not c.hflip for c in plan.clips)
+
+    def test_mixed_order_alignment(self):
+        # 仅第 0、2 个片段无文字；has_text 标记必须与片段序号严格对齐
+        svc = _make_service({"generation_task_id": "t-mix"}, clip_has_text=[False, True, False, True])
+        plan = svc._get_micro_transform_plan(4)
+        assert [c.has_text for c in plan.clips] == [False, True, False, True]
+        assert all(not plan.clips[i].hflip for i in (1, 3))
+        for i in (0, 2):
+            # 无文字片段的翻转由 50% 种子决定，但允许翻转（不强制一定翻）
+            assert plan.clips[i].has_text is False
+
+    def test_list_shorter_than_clips_missing_are_conservative(self):
+        # 列表短于片段数：缺位片段按有文字处理
+        svc = _make_service({"generation_task_id": "t-short"}, clip_has_text=[False])
+        plan = svc._get_micro_transform_plan(3)
+        assert [c.has_text for c in plan.clips] == [False, True, True]
+        assert not plan.clips[1].hflip and not plan.clips[2].hflip
+
+    def test_plan_reproducible_with_real_list(self):
+        cfg = {"generation_task_id": "task-x", "video_index": 1}
+        flags = [False, True, False, False, True]
+        p1 = _make_service(cfg, clip_has_text=flags)._get_micro_transform_plan(5)
+        p2 = _make_service(dict(cfg), clip_has_text=list(flags))._get_micro_transform_plan(5)
+        assert [c.hflip for c in p1.clips] == [c.hflip for c in p2.clips]
+
+
+class TestPureMixedFlags:
+    def test_pure_function_mixed_flags(self):
+        plan = build_micro_transform_plan("seed-1", 0, 4, clip_has_text=[False, True, False, True])
+        assert [c.has_text for c in plan.clips] == [False, True, False, True]
+        # 有文字片段绝不翻转
+        assert not plan.clips[1].hflip and not plan.clips[3].hflip
+
+
+class TestResolveClipHasText:
+    def _adapter(self):
+        from video_processing.render_adapter import RenderAdapter
+
+        return RenderAdapter(MagicMock())
+
+    def test_no_atom_ids_returns_none(self):
+        adapter = self._adapter()
+        clips = [_clip("c1", ""), _clip("c2", "")]
+        assert adapter._resolve_clip_has_text(clips) is None
+
+    def test_explicit_false_only_maps_to_false(self):
+        adapter = self._adapter()
+        clips = [
+            _clip("c1", "a1"),
+            _clip("c2", "a2"),
+            _clip("c3", "a3"),
+            _clip("c4", "a4"),
+            _clip("c5", "a5"),
+        ]
+        atoms = [
+            _atom("a1", {"has_text": False}),  # 明确无文字 → False
+            _atom("a2", {"has_text": True}),  # 有文字
+            _atom("a3", None),  # 标签未生成
+            _atom("a4", {"scene": ["工厂"]}),  # has_text 缺失（null）
+            _atom("a5", {"has_text": "false"}),  # 非布尔 → 保守
+        ]
+        with patch(
+            "packages.adapters.sqlalchemy_impl.asset_atom_clip_repository."
+            "SQLAlchemyAssetAtomClipRepository.find_by_ids",
+            return_value=atoms,
+        ):
+            result = adapter._resolve_clip_has_text(clips)
+        assert result == [False, True, True, True, True]
+
+    def test_audio_clips_excluded_and_order_kept(self):
+        adapter = self._adapter()
+        clips = [
+            _clip("c1", "a1", clip_type="main"),
+            _clip("bgm", "", clip_type="audio"),
+            _clip("c2", "a2", clip_type="pip"),
+        ]
+        atoms = [
+            _atom("a1", {"has_text": False}),
+            _atom("a2", {"has_text": False}),
+        ]
+        with patch(
+            "packages.adapters.sqlalchemy_impl.asset_atom_clip_repository."
+            "SQLAlchemyAssetAtomClipRepository.find_by_ids",
+            return_value=atoms,
+        ) as mock_find:
+            result = adapter._resolve_clip_has_text(clips)
+        # 只查非 audio 片段的 atom id，且顺序为 main → pip
+        assert mock_find.call_args.args[0] == ["a1", "a2"]
+        assert result == [False, False]
+
+    def test_missing_atom_record_defaults_true(self):
+        adapter = self._adapter()
+        clips = [_clip("c1", "a1"), _clip("c2", "a2")]
+        with patch(
+            "packages.adapters.sqlalchemy_impl.asset_atom_clip_repository."
+            "SQLAlchemyAssetAtomClipRepository.find_by_ids",
+            return_value=[_atom("a1", {"has_text": False})],  # a2 查不到
+        ):
+            result = adapter._resolve_clip_has_text(clips)
+        assert result == [False, True]
+
+    def test_query_failure_returns_none(self):
+        adapter = self._adapter()
+        clips = [_clip("c1", "a1")]
+        with patch(
+            "packages.adapters.sqlalchemy_impl.asset_atom_clip_repository."
+            "SQLAlchemyAssetAtomClipRepository.find_by_ids",
+            side_effect=RuntimeError("db down"),
+        ):
+            assert adapter._resolve_clip_has_text(clips) is None
+
+    def test_duplicate_atom_ids_queried_once(self):
+        adapter = self._adapter()
+        clips = [_clip("c1", "a1"), _clip("c2", "a1")]
+        with patch(
+            "packages.adapters.sqlalchemy_impl.asset_atom_clip_repository."
+            "SQLAlchemyAssetAtomClipRepository.find_by_ids",
+            return_value=[_atom("a1", {"has_text": False})],
+        ) as mock_find:
+            result = adapter._resolve_clip_has_text(clips)
+        assert mock_find.call_args.args[0] == ["a1"]
+        assert result == [False, False]

tests/unit/test_1970_micro_transform_render.py

@@ -22,6 +22,7 @@ def _make_service(plan_config: dict | None = None, clips=None):
     svc.plan.clips = clips or []
     svc._micro_plan_cache = None
     svc._micro_plan_loaded = False
+    svc._clip_has_text = None
     return svc
 
 
@@ -159,6 +160,7 @@ class TestStreamCopyGate:
         svc.clips = [source]
         svc._micro_plan_cache = None
         svc._micro_plan_loaded = False
+        svc._clip_has_text = None
         resolved = ResolvedClip(
             clip_id="c1",
             asset_id="a1",

tests/unit/test_gpu_lipsync_service.py

@@ -181,6 +181,73 @@ def test_register_worker_creates_then_updates(svc):
     assert w2.created_at == w.created_at  # 没新建
 
 
+def test_register_with_task_id_refreshes_task_heartbeat(svc):
+    """#1970 推理期心跳：register(task_id=...) 只刷新本 worker 的 processing 任务."""
+    from packages.adapters.sqlalchemy_impl.models import GpuWorkerModel
+
+    t = svc.create_task(video_url="v", audio_url="a")
+    svc.poll_task("w-1")
+    svc.db.refresh(t)
+    old_hb = t.last_heartbeat_at
+    assert t.status == "processing"
+    # 模拟时间流逝后心跳到达
+    svc.db.query(GpuWorkerModel).filter_by(worker_id="w-1").update(
+        {"last_heartbeat_at": old_hb - timedelta(seconds=300)}
+    )
+    svc.db.commit()
+    svc.register_worker("w-1", task_id=t.id)
+    svc.db.refresh(t)
+    assert t.last_heartbeat_at > old_hb
+    assert t.status == "processing"  # 心跳不改变状态
+    # worker 表心跳也被刷新
+    w = svc.db.query(GpuWorkerModel).filter_by(worker_id="w-1").one()
+    assert w.last_heartbeat_at > old_hb
+
+
+def test_register_task_heartbeat_ignores_finished_or_foreign_task(svc):
+    """任务已 done，或已被超时回收重新派发给别的 worker 时，旧心跳必须忽略."""
+    from packages.adapters.sqlalchemy_impl.models import GpuLipsyncTaskModel, GpuWorkerModel
+
+    # 场景 1：任务已完成 → register 带 task_id 不得改写任务心跳
+    t = svc.create_task(video_url="v", audio_url="a")
+    svc.poll_task("w-1")
+    done = svc.report_result(t.id, "w-1", success=True, duration_seconds=10.0)
+    hb_when_done = done.last_heartbeat_at
+    svc.register_worker("w-1", task_id=t.id)
+    svc.db.refresh(t)
+    assert t.status == "done"
+    assert t.last_heartbeat_at == hb_when_done  # 没被改写
+
+    # 场景 2：任务超时回收后被 w-2 重新认领，旧 worker w-1 的迟到心跳无效
+    t2 = svc.create_task(video_url="v2", audio_url="a2")
+    svc.poll_task("w-1")
+    svc.db.refresh(t2)
+    t2.last_heartbeat_at = datetime.now(UTC) - timedelta(days=1)
+    svc.db.commit()
+    claimed = svc.poll_task("w-2")  # 触发回收并由 w-2 重新认领
+    assert claimed is not None and claimed.id == t2.id
+    owner_hb = claimed.last_heartbeat_at
+    # 把 w-2 的 worker 心跳拨早，确认旧心跳不会影响任务归属
+    svc.db.query(GpuWorkerModel).filter_by(worker_id="w-2").update(
+        {"last_heartbeat_at": owner_hb - timedelta(seconds=600)}
+    )
+    svc.db.commit()
+    svc.register_worker("w-1", task_id=t2.id)  # 旧 worker 迟到心跳
+    svc.db.refresh(t2)
+    assert t2.worker_id == "w-2"
+    assert t2.status == "processing"
+    assert t2.last_heartbeat_at == owner_hb
+
+    # 场景 3：不存在的 task_id 不报错
+    svc.register_worker("w-1", task_id="nonexistent-id")
+    assert svc.db.get(GpuLipsyncTaskModel, "nonexistent-id") is None
+
+
+def test_default_gpu_task_timeout_is_900(svc):
+    """#1970 默认超时 300→900，覆盖 RTX2060 长视频推理."""
+    assert svc.settings.gpu_task_timeout_seconds == 900
+
+
 # ── get_by_lipsync_job ─────────────────────────────────────────────