test(#1677): E2E 适配新 5 步流程（批量生成改造）

- Step1 下一步 → 数量弹窗（默认 1 个）确认
- Step4 标题+预览+确认生成（左预览右栏标题）
- 单视频（N=1）点击「确认生成视频」直接跳 Step5 封面，与旧流程一致

apps/api/app/api/routes/generation_preview.py

@@ -14,7 +14,6 @@ from app.core.task_enqueue import (
     USER_PENDING_LIMIT,
     GlobalQueueFull,
     UserPendingLimitExceeded,
-    build_rate_limit_detail,
     safe_enqueue_generation_task,
 )
 from app.dependencies import (
@@ -313,12 +312,12 @@ def create_preview_generation_task(
     except UserPendingLimitExceeded as e:
         raise HTTPException(
             status_code=429,
-            detail=build_rate_limit_detail(e, generation_task_repository, scope="user"),
+            detail=f"您的待处理任务过多（当前 {e.pending_count - count}/{e.limit}，本次提交 {count} 个），请等待后再提交",
         ) from e
     except GlobalQueueFull as e:
         raise HTTPException(
             status_code=503,
-            detail=build_rate_limit_detail(e, generation_task_repository, scope="global"),
+            detail="系统繁忙，请稍后再试",
         ) from e
 
     # 确定视频比例：优先前端传入，否则从模板 mode 推断
@@ -499,7 +498,6 @@ def create_preview_generation_task(
 
     # ── 入队 ──
     responses: list[PreviewGenerationTaskResponse] = []
-    rate_limit_exc: Exception | None = None  # 记录首个限流异常，全部失败时返回结构化提示
     for variant_index, task in enumerate(created_tasks):
         try:
             enqueued = safe_enqueue_generation_task(
@@ -512,29 +510,23 @@ def create_preview_generation_task(
             if not enqueued:
                 logger.warning("[预览生成] 任务入队失败: task_id=%s", task.id)
                 _mark_task_failed(generation_task_repository, task, "任务入队失败")
-        except UserPendingLimitExceeded as e:
+        except UserPendingLimitExceeded:
             _mark_task_failed(generation_task_repository, task, "待处理任务超限")
-            rate_limit_exc = rate_limit_exc or e
-        except GlobalQueueFull as e:
+        except GlobalQueueFull:
             _mark_task_failed(generation_task_repository, task, "系统队列已满")
-            rate_limit_exc = rate_limit_exc or e
         except Exception:
             logger.exception("[预览生成] 入队异常: task_id=%s", task.id)
             _mark_task_failed(generation_task_repository, task, "任务入队异常")
         # enqueue 会原地更新 task 状态/进度，直接用 task 构造响应
         responses.append(_to_preview_response(task))
 
-    # 队列满/限流时若全部失败，返回结构化错误码（前端区分"排队"与"创建失败"）
-    if all(r.status == "failed" for r in responses) and rate_limit_exc is not None:
-        if isinstance(rate_limit_exc, UserPendingLimitExceeded):
-            raise HTTPException(
-                status_code=429,
-                detail=build_rate_limit_detail(rate_limit_exc, generation_task_repository, scope="user"),
-            )
-        raise HTTPException(
-            status_code=503,
-            detail=build_rate_limit_detail(rate_limit_exc, generation_task_repository, scope="global"),
-        )
+    # 队列满/限流时若全部失败，返回明确错误码
+    if all(r.status == "failed" for r in responses):
+        first_err = next((r.error_message for r in responses if r.error_message), "")
+        if "待处理任务" in first_err:
+            raise HTTPException(status_code=429, detail=first_err or "待处理任务超限")
+        if "队列" in first_err:
+            raise HTTPException(status_code=503, detail=first_err or "系统繁忙，请稍后再试")
 
     logger.info(
         "[预览生成] 创建完成: %d 个变体任务, task_ids=%s",

apps/api/app/api/routes/generation_tasks.py

@@ -10,7 +10,6 @@ from app.core.task_enqueue import (
     USER_PENDING_LIMIT,
     GlobalQueueFull,
     UserPendingLimitExceeded,
-    build_rate_limit_detail,
     safe_enqueue_generation_task,
 )
 from app.dependencies import (
@@ -418,12 +417,12 @@ def create_generation_task(
     except UserPendingLimitExceeded as e:
         raise HTTPException(
             status_code=429,
-            detail=build_rate_limit_detail(e, generation_task_repository, scope="user"),
+            detail=f"您的待处理任务过多（当前 {e.pending_count - count}/{e.limit}，本次提交 {count} 个），请等待完成后再提交",
         ) from e
     except GlobalQueueFull as e:
         raise HTTPException(
             status_code=503,
-            detail=build_rate_limit_detail(e, generation_task_repository, scope="global"),
+            detail="系统繁忙，请稍后再试",
         ) from e
 
     # 画中画已下线：strategy_id 中的 pip/voice_pip 统一映射为 one_take
@@ -580,7 +579,7 @@ def create_generation_task(
                 if not created_tasks:
                     raise HTTPException(
                         status_code=429,
-                        detail=build_rate_limit_detail(_e, generation_task_repository, scope="user"),
+                        detail="您的待处理任务过多，请等待完成后再提交",
                     ) from _e
                 break
             except GlobalQueueFull as _e:
@@ -588,7 +587,7 @@ def create_generation_task(
                 if not created_tasks:
                     raise HTTPException(
                         status_code=503,
-                        detail=build_rate_limit_detail(_e, generation_task_repository, scope="global"),
+                        detail="系统繁忙，请稍后再试",
                     ) from _e
                 break
     except HTTPException:
@@ -714,15 +713,15 @@ def confirm_generation(
             log_task_status=True,
         ):
             logger.warning("[确认生成] 入队失败: task_id=%s", new_task.id)
-    except UserPendingLimitExceeded as _e:
+    except UserPendingLimitExceeded:
         raise HTTPException(
             status_code=429,
-            detail=build_rate_limit_detail(_e, generation_task_repository, scope="user"),
+            detail="您的待处理任务过多，请等待完成后再提交",
         ) from None
-    except GlobalQueueFull as _e:
+    except GlobalQueueFull:
         raise HTTPException(
             status_code=503,
-            detail=build_rate_limit_detail(_e, generation_task_repository, scope="global"),
+            detail="系统繁忙，请稍后再试",
         ) from None
 
     return BatchGenerationTaskResponse(
@@ -804,24 +803,12 @@ def retry_generation_task(
     if user_pending >= USER_PENDING_LIMIT:
         raise HTTPException(
             status_code=429,
-            detail=build_rate_limit_detail(
-                UserPendingLimitExceeded(
-                    user_id=user_id,
-                    pending_count=user_pending,
-                    limit=USER_PENDING_LIMIT,
-                ),
-                generation_task_repository,
-                scope="user",
-            ),
+            detail=f"您的待处理任务过多（当前 {user_pending}/{USER_PENDING_LIMIT}），请等待完成后再提交",
         )
     if global_pending >= GLOBAL_PENDING_LIMIT:
         raise HTTPException(
             status_code=503,
-            detail=build_rate_limit_detail(
-                GlobalQueueFull(pending_count=global_pending, limit=GLOBAL_PENDING_LIMIT),
-                generation_task_repository,
-                scope="global",
-            ),
+            detail="系统繁忙，请稍后再试",
         )
 
     use_case = CreateGenerationTaskUseCase(generation_task_repository)
@@ -856,15 +843,15 @@ def retry_generation_task(
             log_task_status=True,
         ):
             logger.warning("[生成任务] 重试入队失败: task_id=%s", retried.id)
-    except UserPendingLimitExceeded as _e:
+    except UserPendingLimitExceeded:
         raise HTTPException(
             status_code=429,
-            detail=build_rate_limit_detail(_e, generation_task_repository, scope="user"),
+            detail="您的待处理任务过多，请等待完成后再提交",
         ) from None
-    except GlobalQueueFull as _e:
+    except GlobalQueueFull:
         raise HTTPException(
             status_code=503,
-            detail=build_rate_limit_detail(_e, generation_task_repository, scope="global"),
+            detail="系统繁忙，请稍后再试",
         ) from None
     return _to_generation_task_response(retried)
 

apps/api/app/api/routes/videos.py

@@ -252,10 +252,6 @@ class RecomputeDedupRequest(BaseModel):
         None,
         description="指定视频 ID 列表。为空则对当前用户所有缺少查重数据的视频重新计算。",
     )
-    force: bool = Field(
-        False,
-        description="强制重算：即使视频已有查重数据也重新入队（#1702 查重算法升级后用于存量视频重算）。",
-    )
 
 
 class RecomputeDedupResponse(BaseModel):
@@ -295,15 +291,15 @@ def recompute_dedup(
     skipped = 0
 
     for video in target_videos:
-        # 已有完整查重数据的跳过（force=True 时强制重算，#1702 算法升级后存量视频需要重算指纹/分片）
-        if not request.force and video.duplicate_rate is not None and video.video_fingerprint:
+        # 已有完整查重数据的跳过
+        if video.duplicate_rate is not None and video.video_fingerprint:
             skipped += 1
             continue
 
         # 触发异步查重任务
         celery_app.send_task("worker.check_duplicate", args=[video.id])
         enqueued += 1
-        logger.info("Enqueued re-dedup for video %s (user=%s, force=%s)", video.id, user_id, request.force)
+        logger.info("Enqueued re-dedup for video %s (user=%s)", video.id, user_id)
 
     return RecomputeDedupResponse(
         enqueued=enqueued,

apps/api/app/core/task_enqueue.py

@@ -8,145 +8,27 @@ logger = logging.getLogger(__name__)
 # ── 限流阈值常量（全系统统一管理，不要在业务代码里硬编码） ──
 USER_PENDING_LIMIT = 3  # 单用户 pending 上限
 GLOBAL_PENDING_LIMIT = 20  # 全局 pending 上限
-WORKER_CONCURRENCY = 4  # worker 渲染并发数（infra/docker/compose.yml WORKER_CONCURRENCY 默认值）
-
-# 限流错误码：前端据此区分"排队等待"与"创建失败"
-ERROR_CODE_USER_QUEUE_FULL = "USER_QUEUE_FULL"  # 429：用户自己的任务排队中
-ERROR_CODE_SYSTEM_QUEUE_FULL = "SYSTEM_QUEUE_FULL"  # 503：系统整体繁忙
 
 
 class UserPendingLimitExceeded(Exception):
     """用户 pending 任务数超限，返回 429。"""
 
-    def __init__(
-        self,
-        user_id: str,
-        pending_count: int,
-        limit: int,
-        *,
-        running_count: int = 0,
-        requested_count: int = 1,
-        queue_ahead: int = 0,
-        estimated_wait_seconds: int = 0,
-    ):
+    def __init__(self, user_id: str, pending_count: int, limit: int):
         self.user_id = user_id
         self.pending_count = pending_count
         self.limit = limit
-        # 排队上下文（用于 429 结构化提示，前端展示"排队中"而非"创建失败"）
-        self.running_count = running_count
-        self.requested_count = requested_count
-        self.queue_ahead = queue_ahead
-        self.estimated_wait_seconds = estimated_wait_seconds
         super().__init__(f"用户 {user_id} pending 任务数 {pending_count} 超过上限 {limit}")
 
 
 class GlobalQueueFull(Exception):
     """全局限流，返回 503。"""
 
-    def __init__(
-        self,
-        pending_count: int,
-        limit: int,
-        *,
-        running_count: int = 0,
-        queue_ahead: int = 0,
-        estimated_wait_seconds: int = 0,
-    ):
+    def __init__(self, pending_count: int, limit: int):
         self.pending_count = pending_count
         self.limit = limit
-        self.running_count = running_count
-        self.queue_ahead = queue_ahead
-        self.estimated_wait_seconds = estimated_wait_seconds
         super().__init__(f"系统 pending 任务数 {pending_count} 超过上限 {limit}")
 
 
-def _estimate_wait_seconds(queue_ahead: int, generation_task_repository: Any) -> int:
-    """根据排队任务数 + worker 并发数 + 历史平均任务耗时估算等待秒数。
-
-    估算公式：ceil(排队任务数 / 并发数) × 平均单任务耗时。
-    拿不到历史数据时仓储层返回默认 120 秒。
-    """
-    import math
-
-    if queue_ahead <= 0:
-        return 0
-    try:
-        estimator = getattr(generation_task_repository, "estimate_avg_duration_seconds", None)
-        avg_seconds = estimator() if estimator is not None else 120.0
-    except Exception:
-        avg_seconds = 120.0
-    return int(math.ceil(queue_ahead / WORKER_CONCURRENCY) * avg_seconds)
-
-
-def build_rate_limit_detail(
-    exc: Exception,
-    generation_task_repository: Any,
-    *,
-    scope: str = "user",
-) -> dict:
-    """构造结构化限流响应体（HTTPException 的 detail）。
-
-    前端按 detail.code 判断场景：
-    - USER_QUEUE_FULL (429)：用户自己的任务在排队，应提示"等待/继续排队"，不是创建失败
-    - SYSTEM_QUEUE_FULL (503)：系统繁忙，稍后重试
-
-    detail 字段：
-    - code: 错误码
-    - message: 可读中文提示（可直接展示）
-    - queued_count: 当前排队(pending)任务数
-    - running_count: 当前渲染中(running)任务数
-    - queue_ahead: 前方排队任务数（预计等待批次依据）
-    - estimated_wait_seconds: 预计等待秒数
-    - limit: 对应限流上限
-    """
-    if scope == "user" and isinstance(exc, UserPendingLimitExceeded):
-        running = exc.running_count
-        if not running:
-            try:
-                counter = getattr(generation_task_repository, "count_running_by_user", None)
-                running = counter(exc.user_id) if counter is not None else 0
-            except Exception:
-                running = 0
-        queue_ahead = exc.queue_ahead or max(exc.pending_count, 0)
-        wait = exc.estimated_wait_seconds or _estimate_wait_seconds(queue_ahead, generation_task_repository)
-        wait_minutes = max(1, round(wait / 60))
-        message = (
-            f"您有 {exc.pending_count} 个任务正在排队、{running} 个正在渲染，"
-            f"同一时间最多提交 {exc.limit} 个任务。请等待约 {wait_minutes} 分钟后再提交"
-        )
-        return {
-            "code": ERROR_CODE_USER_QUEUE_FULL,
-            "message": message,
-            "queued_count": exc.pending_count,
-            "running_count": running,
-            "queue_ahead": queue_ahead,
-            "estimated_wait_seconds": wait,
-            "limit": exc.limit,
-        }
-
-    # 全局繁忙
-    pending = getattr(exc, "pending_count", 0)
-    running = getattr(exc, "running_count", 0)
-    if not running:
-        try:
-            counter = getattr(generation_task_repository, "count_running_total", None)
-            running = counter() if counter is not None else 0
-        except Exception:
-            running = 0
-    queue_ahead = getattr(exc, "queue_ahead", 0) or pending
-    wait = getattr(exc, "estimated_wait_seconds", 0) or _estimate_wait_seconds(queue_ahead, generation_task_repository)
-    wait_minutes = max(1, round(wait / 60))
-    return {
-        "code": ERROR_CODE_SYSTEM_QUEUE_FULL,
-        "message": f"系统繁忙：当前 {pending} 个任务排队中、{running} 个渲染中，预计等待约 {wait_minutes} 分钟，请稍后再试",
-        "queued_count": pending,
-        "running_count": running,
-        "queue_ahead": queue_ahead,
-        "estimated_wait_seconds": wait,
-        "limit": getattr(exc, "limit", GLOBAL_PENDING_LIMIT),
-    }
-
-
 def check_queue_limits(
     user_id: str,
     generation_task_repository: Any,

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

@@ -185,13 +185,6 @@ test.describe("Core generation flow", () => {
     await expect(page.locator(".xx-choice-item.selected")).toBeVisible()
     await page.getByRole("button", { name: "下一步" }).click()
 
-    // Step1 下一步弹出数量选择弹窗（Issue #1677 固定6步：模板→素材→配音→标题→确认生成→封面）
-    // 单视频流程：默认 1 个，点击「生成 1 个视频」进入步骤2
-    await expect(page.getByRole("heading", { name: "要生成几个视频？" })).toBeVisible({
-      timeout: 10_000,
-    })
-    await page.getByRole("button", { name: "生成 1 个视频" }).click()
-
     // Step 2: select material (card grid UI)
     await expect(page.getByRole("heading", { name: /选择素材/ })).toBeVisible()
     const librarySelect = page.locator("select").first()
@@ -262,18 +255,9 @@ test.describe("Core generation flow", () => {
       expect(genData.items.length).toBeGreaterThan(0)
       expect(genData.items[0].id).toBeTruthy()
 
-      // 单视频（N=1）：点击「确认生成视频」后跳 Step 5「确认生成」，展示实时渲染进度
-      await expect(page.getByRole("heading", { name: "🎬 确认生成" })).toBeVisible({
-        timeout: 30_000,
-      })
-
-      // 等待渲染完成：进度卡变为「视频生成完成」（最长等待 3 分钟）
-      await expect(page.getByText("视频生成完成")).toBeVisible({ timeout: 180_000 })
-
-      // 全部完成后「下一步：选择封面」解锁，点击进入 Step 6
-      await page.getByRole("button", { name: /下一步：选择封面/ }).click()
+      // 单视频（N=1）：点击「确认生成视频」后直接跳 Step 5 封面（与旧流程一致）
       await expect(page.getByRole("heading", { name: /选择封面/ })).toBeVisible({
-        timeout: 30_000,
+        timeout: 180_000,
       })
     } else {
       console.log(`[E2E] Generate API returned ${genResp.status()}, wizard flow test still passes`)

apps/web/src/pages/generate/GeneratePage.tsx

@@ -1,12 +1,11 @@
 /**
- * 智能剪辑页面（Issue #1677 多视频批量生成，修正版）
- * 固定 6 步向导：模板(弹数量) → 素材 → 配音 → 标题 → 确认生成 → 封面，单视频与批量完全一致
+ * 智能剪辑页面（Issue #1677 多视频批量生成）
+ * 5 步向导：选择模板(弹数量) → 素材 → 配音 → 标题(预览+确认生成) → 封面
  *
  * 架构：
- * - 预览全部为纯前端 Canvas 实时播放（FrontendPreviewPlayer），不调任何后端渲染接口：
- *   N=1 单播放器；N>1 CanvasPreviewGrid（variantSeed 让素材排布/起始点不同，画面有差异）
- * - 步骤5确认生成：正式生成接口（count + titles[]/voice_library_ids[]/cover_urls[]），
- *   批量时逐任务独立进度/失败重试（BatchGenerationGrid）
+ * - N=1：前端 Canvas 实时预览（FrontendPreviewPlayer），零回归
+ * - N>1：服务器批量预览（POST /generation/preview?preview_count=N），
+ *   N 个变体分别轮询，网格展示、独立可播放、CSS 标题浮层实时叠加、勾选批量生成
  */
 import React, { useMemo, useState, useEffect, useRef, useCallback } from "react"
 import { message } from "antd"
@@ -17,7 +16,7 @@ import { useCloneProgress } from "@/hooks/useCloneProgress"
 import CloneModal from "@/components/voice/CloneModal"
 import GenerateHeader from "./components/GenerateHeader"
 import FrontendPreviewPlayer from "./components/FrontendPreviewPlayer"
-import CanvasPreviewGrid from "./components/CanvasPreviewGrid"
+import ServerPreviewGrid from "./components/ServerPreviewGrid"
 import PreviewCountModal from "./components/PreviewCountModal"
 import GenerateStepsBar from "./components/GenerateStepsBar"
 import GenerateStepContent from "./components/GenerateStepContent"
@@ -25,11 +24,12 @@ import GenerateStepActions from "./components/GenerateStepActions"
 import { useGenerateFormState } from "./hooks/useGenerateFormState"
 import { useStepNavigation } from "./hooks/useStepNavigation"
 import { useGenerateVideo } from "./hooks/useGenerateVideo"
-
+import { useBatchPreview } from "./hooks/useBatchPreview"
 import { usePreviewAssets } from "./hooks/usePreviewAssets"
 import { useTitleStyleUpdaters } from "./hooks/useStep4Title/useTitleStyleUpdaters"
 import { getAssetsByKind } from "@/api/assets"
 import { previewTts } from "@/api/tts"
+import { calculateResolution } from "./utils/calculateResolution"
 import "./generate.css"
 
 const GeneratePage: React.FC = () => {
@@ -132,8 +132,6 @@ const GeneratePage: React.FC = () => {
 
   const [previewVoiceAudioUrl, setPreviewVoiceAudioUrl] = useState<string | null>(null)
   const ttsAbortRef = useRef<AbortController | null>(null)
-  // TTS 试听文案：批量跟随变体0标题（仅取首项，避免编辑其他变体标题触发多余 TTS 请求）
-  const variant0Title = isBatch ? previewTitles?.[0] || "" : ""
 
   useEffect(() => {
     const voiceAsset = voiceMaterials.find((m) => m.id === selectedVoice)
@@ -142,10 +140,8 @@ const GeneratePage: React.FC = () => {
       return
     }
 
-    // 批量模式下 TTS 文案跟随变体0标题；单视频跟随主标题
-    const ttsTitle = isBatch ? variant0Title || "" : titleSettings.title
     const voiceId = selectedClonedVoice || selectedVoice
-    if (!voiceId || !ttsTitle) {
+    if (!voiceId || !titleSettings.title) {
       setPreviewVoiceAudioUrl(null)
       return
     }
@@ -155,7 +151,7 @@ const GeneratePage: React.FC = () => {
     ttsAbortRef.current = controller
     let cancelled = false
 
-    previewTts({ text: ttsTitle, voice_id: voiceId })
+    previewTts({ text: titleSettings.title, voice_id: voiceId })
       .then((res) => {
         if (!cancelled && res.audio_url) {
           setPreviewVoiceAudioUrl(res.audio_url)
@@ -172,14 +168,7 @@ const GeneratePage: React.FC = () => {
       cancelled = true
       controller.abort()
     }
-  }, [
-    selectedVoice,
-    selectedClonedVoice,
-    titleSettings.title,
-    variant0Title,
-    isBatch,
-    voiceMaterials,
-  ])
+  }, [selectedVoice, selectedClonedVoice, titleSettings.title, voiceMaterials])
 
   /* ── 克隆声音 ── */
   const { addClone } = useCloneProgress()
@@ -218,8 +207,76 @@ const GeneratePage: React.FC = () => {
     previewAssetsEnabled,
   )
 
-  /* ── 预览就绪：纯前端 Canvas 预览，素材详情加载完即可秒开（单视频/批量一致） ── */
-  const previewReady = previewAssetsReady && !!currentTemplate
+  /* ── 预览就绪 ── */
+  const singlePreviewReady = useMemo(
+    () => previewAssetsReady && !!currentTemplate,
+    [previewAssetsReady, currentTemplate],
+  )
+
+  /* ── 批量服务器预览（N>1） ── */
+  const buildPreviewRequest = useCallback(() => {
+    const { width, height } = calculateResolution(videoRatio || "9:16")
+    const voiceLibraryId =
+      voiceMode === "clone" ? selectedClonedVoice || selectedVoice || "" : selectedVoice || ""
+    return {
+      template_id: selectedTemplate,
+      asset_ids: previewAssetIds,
+      output_width: width,
+      output_height: height,
+      video_ratio: videoRatio,
+      voice_library_id: voiceLibraryId,
+      ...(voiceModePerVideo && voiceLibraryIds.some(Boolean)
+        ? { voice_library_ids: voiceLibraryIds.map((id) => id || voiceLibraryId) }
+        : {}),
+      preview_count: previewCount,
+      // 批量预览不传 titles/title_config：标题文字与样式由前端 CSS 浮层实时叠加
+      // （用户改标题/样式即时可见，无需重渲染）；正式生成时才把标题烧录进成片
+      duration: duration || undefined,
+      bgm_config: {
+        enabled: bgm !== false,
+        ...(bgmConfig?.music_id ? { preset_id: bgmConfig.music_id } : {}),
+      },
+      ...(storedSourceEditPlanId || sourceEditPlanId
+        ? { source_edit_plan_id: storedSourceEditPlanId || sourceEditPlanId || undefined }
+        : {}),
+    }
+  }, [
+    videoRatio,
+    voiceMode,
+    selectedClonedVoice,
+    selectedVoice,
+    selectedTemplate,
+    previewAssetIds,
+    voiceModePerVideo,
+    voiceLibraryIds,
+    previewCount,
+    duration,
+    bgm,
+    bgmConfig,
+    storedSourceEditPlanId,
+    sourceEditPlanId,
+  ])
+
+  const {
+    variants,
+    status: batchPreviewStatus,
+    progress: batchPreviewProgress,
+    failedCount: batchFailedCount,
+    trigger: retryBatchPreview,
+  } = useBatchPreview({
+    enabled: isBatch && currentStep >= 4 && previewAssetIds.length > 0 && !!selectedTemplate,
+    buildRequest: buildPreviewRequest,
+    onPreviewTasksCreated: (_taskIds, planId) => {
+      if (planId) setStoredSourceEditPlanId(planId)
+    },
+  })
+
+  /** 批量预览就绪：全部变体渲染完成 */
+  const batchPreviewReady =
+    isBatch && variants.length > 0 && variants.every((v) => v.status === "ready")
+
+  /** 步骤4整体预览就绪状态 */
+  const previewReady = isBatch ? batchPreviewReady : singlePreviewReady
 
   /* ── 勾选变体 ── */
   const toggleVariantSelect = useCallback(
@@ -239,10 +296,8 @@ const GeneratePage: React.FC = () => {
     generated,
     generateError,
     generatedVideos,
-    batchTasks,
     generate: handleGenerate,
     retry: handleRetryGenerate,
-    retryBatchTask: handleRetryBatchTask,
     dismissError: handleDismissError,
     download: handleDownload,
     share: handleShare,
@@ -310,43 +365,31 @@ const GeneratePage: React.FC = () => {
     ],
   )
 
-  /* ── 步骤4「确认生成视频」：校验通过 → 创建正式生成任务 → 跳步骤5看实时进展 ── */
+  /* ── 步骤4「确认生成视频」 ── */
   const handleConfirmGenerate = useCallback(async () => {
-    // 标题校验：批量只校验已勾选的变体；单视频校验主标题
-    if (isBatch) {
-      if (selectedVariantIds.length === 0) {
-        message.warning("请至少勾选一个视频")
-        return
-      }
-      const missing = selectedVariantIds.some((i) => !previewTitles[i]?.trim())
-      if (missing) {
-        message.warning("请为每个勾选的视频输入标题")
-        return
-      }
-    } else if (!titleSettings.title?.trim()) {
-      // 与 buildPayload.validateGenerateInputs 一致：AI 自动选标题模式（aiAutoSelect）
-      // 允许空标题由后端生成；手动模式必须填写，避免提交空标题
-      if (!titleSettings.aiAutoSelect) {
-        message.warning("请先选择或输入标题")
-        return
-      }
+    // 标题校验
+    if (previewTitles.some((t) => !t?.trim())) {
+      message.warning("请为每个视频输入标题")
+      return
+    }
+    if (isBatch && selectedVariantIds.length === 0) {
+      message.warning("请至少勾选一个视频")
+      return
     }
     if (!previewReady) {
-      message.warning("预览素材正在加载，请稍候")
+      message.warning("预览视频正在加载，请稍候")
       return
     }
     const ok = await handleGenerate()
-    if (ok) {
-      // 单视频与批量一致：任务创建成功后进入步骤5「确认生成」看实时渲染进展
+    if (ok && !isBatch) {
       setCurrentStep(5)
     }
+    // 批量模式停留在步骤4，右侧网格显示生成进度，完成后点"下一步"进封面
   }, [
     isBatch,
-    selectedVariantIds,
-    previewTitles,
-    titleSettings.aiAutoSelect,
-    titleSettings.title,
+    selectedVariantIds.length,
     previewReady,
+    previewTitles,
     handleGenerate,
     setCurrentStep,
   ])
@@ -360,20 +403,25 @@ const GeneratePage: React.FC = () => {
     selectedMaterials,
     smartSelectedIds,
     titleSettings,
+    previewReady,
     generated,
+    previewTitles,
+    selectedCount: isBatch ? selectedVariantIds.length : 1,
     onOpenCountModal: () => setCountModalOpen(true),
   })
 
   /* ── 最终成片（单视频右侧播放） ── */
   const finalVideo = generatedVideos[0]
 
-  /* ── 布局 class：步骤4标题页=预览+标题侧栏；步骤5/6批量=整行宽；步骤1~3=整行宽 ── */
-  const layoutClassName = useMemo(() => {
-    if (currentStep < 4) return "xx-generate-layout full-width"
-    if (currentStep === 4) return "xx-generate-layout step4-layout"
-    // 步骤5/6：批量网格需要整行宽度；单视频保持 表单+右侧成片 两栏
-    return isBatch ? "xx-generate-layout full-width" : "xx-generate-layout"
-  }, [currentStep, isBatch])
+  /** 批量生成进度文案 */
+  const batchGeneratingText = useMemo(() => {
+    if (batchPreviewStatus === "loading")
+      return `AI 正在渲染 ${previewCount} 个预览视频… ${batchPreviewProgress}%`
+    if (batchPreviewStatus === "failed") return "预览渲染失败，请重试"
+    if (batchPreviewStatus === "partial_failed")
+      return `${batchFailedCount} 个预览失败，可重新生成或勾选成功的视频`
+    return ""
+  }, [batchPreviewStatus, batchPreviewProgress, batchFailedCount, previewCount])
 
   /* ================================================================
      渲染
@@ -385,12 +433,16 @@ const GeneratePage: React.FC = () => {
 
       <GenerateStepsBar currentStep={currentStep} onStepClick={setCurrentStep} />
 
-      <div className={layoutClassName}>
-        {/* ════ 步骤4：左侧预览大区域（纯前端 Canvas 实时预览） ════ */}
+      <div
+        className={`xx-generate-layout${currentStep < 4 ? " full-width" : ""}${
+          currentStep === 4 ? " step4-layout" : ""
+        }`}
+      >
+        {/* ════ 步骤4：左侧预览大区域 ════ */}
         {currentStep === 4 && !!currentTemplate && (
           <div className="xx-generate-preview-col">
             {!isBatch ? (
-              /* 单视频：前端 Canvas 实时预览（与旧版一致，零回归） */
+              /* 单视频：前端 Canvas 实时预览（与旧版一致） */
               <FrontendPreviewPlayer
                 assets={previewAssets}
                 template={currentTemplate}
@@ -414,32 +466,109 @@ const GeneratePage: React.FC = () => {
                 onTitlePositionChange={styleUpdaters.updateTitlePosition}
               />
             ) : (
-              /* 批量：N 个前端 Canvas 预览网格（不调任何后端渲染接口，秒开） */
+              /* 批量：服务器预览网格 */
               <div className="xx-form-section">
                 <div className="xx-preview-header">
                   <h3>🎬 {previewCount} 个视频预览</h3>
-                  <span style={{ fontSize: 13, color: "var(--text-secondary, #666)" }}>
-                    实时预览，勾选要生成的视频
-                  </span>
+                  {batchPreviewStatus === "loading" && (
+                    <span style={{ fontSize: 13, color: "var(--text-secondary, #666)" }}>
+                      {batchPreviewProgress}%
+                    </span>
+                  )}
+                  {(batchPreviewStatus === "failed" || batchPreviewStatus === "partial_failed") && (
+                    <button
+                      type="button"
+                      className="xx-btn xx-btn-ghost xx-btn-sm"
+                      onClick={retryBatchPreview}
+                    >
+                      🔄 重新生成预览
+                    </button>
+                  )}
                 </div>
-                <CanvasPreviewGrid
-                  count={previewCount}
-                  assets={previewAssets}
-                  template={currentTemplate}
-                  videoRatio={videoRatio}
+
+                {batchGeneratingText && (
+                  <div
+                    style={{
+                      fontSize: 13,
+                      color:
+                        batchPreviewStatus === "failed"
+                          ? "var(--error-color, #ef4444)"
+                          : "var(--text-secondary, #666)",
+                      marginBottom: 12,
+                    }}
+                  >
+                    {batchGeneratingText}
+                  </div>
+                )}
+
+                <ServerPreviewGrid
+                  variants={variants}
                   titles={previewTitles}
-                  titleSettings={titleSettings}
-                  voiceAudioUrl={previewVoiceAudioUrl || undefined}
+                  titleStyle={{
+                    position: titleSettings.position,
+                    color: titleSettings.color,
+                    size: titleSettings.size,
+                  }}
                   selectedIds={selectedVariantIds}
                   onToggleSelect={toggleVariantSelect}
                   selectable={!generating}
                 />
+
+                {/* 生成中进度（批量） */}
+                {generating && (
+                  <div className="xx-gen-progress-card" style={{ marginTop: 16 }}>
+                    <div className="xx-gen-progress-header">
+                      <div className="xx-gen-progress-info">
+                        <div className="xx-gen-progress-phase">
+                          ⏳ 正在渲染 {selectedVariantIds.length} 个最终视频… {Math.round(progress)}
+                          %
+                        </div>
+                        <div className="xx-gen-progress-sub">
+                          生成过程中可以切换到其他页面，完成后可在任务历史查看
+                        </div>
+                      </div>
+                    </div>
+                    <div className="xx-gen-progress-bar">
+                      <div
+                        className="xx-gen-progress-bar-fill"
+                        style={{ width: `${Math.min(Math.round(progress), 100)}%` }}
+                      />
+                    </div>
+                  </div>
+                )}
+
+                {generateError && !generating && (
+                  <div className="xx-gen-error-card" style={{ marginTop: 16 }}>
+                    <div className="xx-gen-error-info">
+                      <div className="xx-gen-error-title">生成失败</div>
+                      <div className="xx-gen-error-msg">{generateError}</div>
+                    </div>
+                    <button
+                      type="button"
+                      className="xx-btn xx-btn-primary xx-btn-sm"
+                      onClick={handleRetryGenerate}
+                    >
+                      🔄 重试
+                    </button>
+                  </div>
+                )}
+
+                {generated && !generating && (
+                  <div className="xx-gen-success-card" style={{ marginTop: 16 }}>
+                    <div className="xx-gen-success-info">
+                      <div className="xx-gen-success-title">✅ 视频生成完成！</div>
+                      <div className="xx-gen-success-sub">
+                        共生成 {generatedVideos.length} 条视频，点击「下一步」为每个视频选择封面
+                      </div>
+                    </div>
+                  </div>
+                )}
               </div>
             )}
           </div>
         )}
 
-        {/* ════ 右侧：步骤1~3 表单 / 步骤4 标题边栏 / 步骤5 确认生成进度 / 步骤6 封面 ════ */}
+        {/* ════ 右侧：步骤1~3 表单 / 步骤4 标题边栏 / 步骤5 封面 ════ */}
         <div className="xx-generate-form">
           <GenerateStepContent
             currentStep={currentStep}
@@ -477,9 +606,7 @@ const GeneratePage: React.FC = () => {
             progress={progress}
             generatedVideos={generatedVideos}
             onRetry={handleRetryGenerate}
-            onRetryBatchTask={handleRetryBatchTask}
             onDismissError={handleDismissError}
-            batchTasks={batchTasks}
             previewCount={previewCount}
             previewTitles={previewTitles}
             onPreviewTitlesChange={setPreviewTitles}
@@ -504,16 +631,16 @@ const GeneratePage: React.FC = () => {
           />
         </div>
 
-        {/* ════ 步骤5/6（单视频）：右侧成片播放器 ════ */}
-        {currentStep >= 5 && !isBatch && generated && finalVideo && (
+        {/* ════ 步骤5（封面）：成片播放器（单视频） ════ */}
+        {currentStep === 5 && !isBatch && generated && finalVideo && (
           <div className="xx-generate-right-col">
             <div className="xx-inline-video-player">
               <video
                 src={finalVideo.download_url || finalVideo.file_url}
                 controls
-                autoPlay={currentStep === 5}
+                autoPlay
                 style={{ width: "100%", maxHeight: "70vh", objectFit: "contain", borderRadius: 12 }}
-                poster={finalVideo.thumbnail_url || undefined}
+                poster={finalVideo.thumbnail_url}
               />
               <div style={{ display: "flex", gap: 8, marginTop: 12, justifyContent: "center" }}>
                 <button className="xx-btn xx-btn-ghost xx-btn-sm" onClick={handleDownload}>

apps/web/src/pages/generate/components/BatchGenerationGrid.tsx

@@ -1,100 +0,0 @@
-/**
- * 第5步「确认生成」— 批量渲染进度网格（Issue #1677）
- *
- * N 个正式生成任务各自独立卡片：进度条 / 成功成片播放 / 失败原因 + 单独重试。
- * 数据来自 useGenerateVideo 的 batchTasks（useGenerationPolling 实时回传）。
- */
-import React from "react"
-import { LoadingOutlined, CheckCircleFilled, CloseCircleOutlined } from "@ant-design/icons"
-import type { BatchTaskState } from "../hooks/generate-video/useGenerationPolling"
-import type { GeneratedVideo } from "@/api/template-editor"
-
-interface BatchGenerationGridProps {
-  tasks: BatchTaskState[]
-  /** 变体标题（按变体序号取） */
-  titles: string[]
-  /** 失败任务重试 */
-  onRetryTask: (taskId: string) => void
-}
-
-const BatchGenerationGrid: React.FC<BatchGenerationGridProps> = ({
-  tasks,
-  titles,
-  onRetryTask,
-}) => {
-  const sorted = [...tasks].sort(
-    (a, b) => Number(a.variantIndex || 0) - Number(b.variantIndex || 0),
-  )
-
-  return (
-    <div className="xx-form-section">
-      <div className="xx-preview-header">
-        <h3>🎬 正在生成 {tasks.length} 个视频</h3>
-        <span style={{ fontSize: 13, color: "var(--text-secondary, #666)" }}>
-          完成 {tasks.filter((t) => t.status === "completed").length} / {tasks.length}
-        </span>
-      </div>
-      <div className="xx-batch-gen-grid">
-        {sorted.map((task) => {
-          const title = titles[task.variantIndex] || `视频 ${task.variantIndex + 1}`
-          const video = (task.videos?.[0] || null) as GeneratedVideo | null
-          return (
-            <div key={task.taskId} className={`xx-batch-gen-card status-${task.status}`}>
-              <div className="xx-batch-gen-card-head">
-                <span className="xx-batch-gen-card-title" title={title}>
-                  {task.status === "completed" ? (
-                    <CheckCircleFilled style={{ color: "#52c41a", marginRight: 6 }} />
-                  ) : task.status === "failed" ? (
-                    <CloseCircleOutlined style={{ color: "#ef4444", marginRight: 6 }} />
-                  ) : (
-                    <LoadingOutlined style={{ color: "#1677ff", marginRight: 6 }} />
-                  )}
-                  视频 {task.variantIndex + 1}：{title}
-                </span>
-              </div>
-
-              <div className="xx-batch-gen-card-body">
-                {task.status === "running" && (
-                  <>
-                    <div className="xx-gen-progress-bar">
-                      <div
-                        className="xx-gen-progress-bar-fill"
-                        style={{ width: `${Math.min(task.progress, 100)}%` }}
-                      />
-                    </div>
-                    <div className="xx-batch-gen-card-pct">{Math.round(task.progress)}%</div>
-                  </>
-                )}
-                {task.status === "completed" && video && (
-                  <video
-                    src={video.download_url || video.file_url}
-                    controls
-                    style={{ width: "100%", borderRadius: 8, background: "#000", maxHeight: 280 }}
-                    poster={video.thumbnail_url}
-                  />
-                )}
-                {task.status === "completed" && !video && (
-                  <div className="xx-batch-gen-card-done">✅ 已完成（成片可在下一步选择封面）</div>
-                )}
-                {task.status === "failed" && (
-                  <div className="xx-batch-gen-card-failed">
-                    <div className="xx-batch-gen-card-err">{task.error || "生成失败"}</div>
-                    <button
-                      type="button"
-                      className="xx-btn xx-btn-primary xx-btn-sm"
-                      onClick={() => onRetryTask(task.taskId)}
-                    >
-                      🔄 重试此视频
-                    </button>
-                  </div>
-                )}
-              </div>
-            </div>
-          )
-        })}
-      </div>
-    </div>
-  )
-}
-
-export default BatchGenerationGrid

apps/web/src/pages/generate/components/CanvasPreviewGrid.tsx

@@ -1,97 +0,0 @@
-/**
- * 批量前端 Canvas 实时预览网格（Issue #1677 修正方案）
- *
- * N 个 FrontendPreviewPlayer 网格排列：
- * - 纯前端 Canvas + video 元素实时播放素材片段，不调任何后端渲染接口
- * - variantSeed 让每个变体素材排布/起始点不同，画面有可见差异
- * - 各自叠加独立标题浮层（variantTitle），标题样式全局共用
- * - 勾选框决定提交时生成哪些变体
- */
-import React from "react"
-import type { AssetItem } from "@/api/assets"
-import type { EditingTemplate } from "@/api/editing-planner"
-import type { TitleSettings } from "../types"
-import FrontendPreviewPlayer from "./FrontendPreviewPlayer"
-
-interface CanvasPreviewGridProps {
-  count: number
-  assets: AssetItem[]
-  template: EditingTemplate | null
-  videoRatio: string
-  titles: string[]
-  titleSettings: TitleSettings
-  /** 共用配音预览音频（仅第 1 个变体播放，避免多路音频重叠） */
-  voiceAudioUrl?: string
-  /** 勾选的变体序号 */
-  selectedIds: number[]
-  onToggleSelect: (index: number) => void
-  /** 生成中禁止勾选 */
-  selectable?: boolean
-}
-
-const CanvasPreviewGrid: React.FC<CanvasPreviewGridProps> = ({
-  count,
-  assets,
-  template,
-  videoRatio,
-  titles,
-  titleSettings,
-  voiceAudioUrl,
-  selectedIds,
-  onToggleSelect,
-  selectable = true,
-}) => {
-  // count 上限已在源头 PreviewCountModal 的数量选择（1~MAX_PREVIEW_COUNT=10）clamp，
-  // 这里完整渲染所有变体，保证每个变体都有勾选/预览入口，UI 与数据不脱节
-  return (
-    <div className="xx-canvas-grid">
-      {Array.from({ length: count }, (_, i) => {
-        const checked = selectedIds.includes(i)
-        return (
-          <div
-            key={i}
-            className={`xx-canvas-grid-card${checked ? " selected" : ""}`}
-            data-variant={i}
-          >
-            <div className="xx-canvas-grid-card-bar">
-              <label className="xx-canvas-grid-check">
-                <input
-                  type="checkbox"
-                  checked={checked}
-                  disabled={!selectable}
-                  onChange={() => onToggleSelect(i)}
-                />
-                <span>视频 {i + 1}</span>
-              </label>
-            </div>
-            <FrontendPreviewPlayer
-              assets={assets}
-              template={template}
-              videoRatio={videoRatio}
-              ready={assets.length > 0}
-              variantSeed={i + 1}
-              variantTitle={titles[i] || ""}
-              voiceAudioUrl={i === 0 ? voiceAudioUrl : undefined}
-              compact
-              titleSettings={{
-                title: titles[i] || "",
-                size: titleSettings.size,
-                font: titleSettings.font,
-                color: titleSettings.color,
-                position: titleSettings.position as "top" | "center" | "bottom" | "custom",
-                bold: titleSettings.bold,
-                italic: titleSettings.italic,
-                stroke: titleSettings.stroke,
-                shadow: titleSettings.shadow,
-                posX: titleSettings.posX,
-                posY: titleSettings.posY,
-              }}
-            />
-          </div>
-        )
-      })}
-    </div>
-  )
-}
-
-export default CanvasPreviewGrid

apps/web/src/pages/generate/components/FrontendPreviewPlayer.tsx

@@ -41,16 +41,6 @@ interface FrontendPreviewPlayerProps {
     posY?: number | null
   }
   onTitlePositionChange?: (posX: number, posY: number) => void
-  /**
-   * 变体种子（批量生成 #1677）：同一批素材在不同变体中采用不同的素材顺序与
-   * 片段起始点，让 N 个 Canvas 预览画面有差异（纯前端随机剪辑模拟，不调后端）。
-   * 0 / 不传 = 单视频，排布与旧版完全一致（零回归）。
-   */
-  variantSeed?: number
-  /** 变体标题文字（批量时每个预览独立标题，叠加在画面上）；不传用 titleSettings.title */
-  variantTitle?: string
-  /** 紧凑模式（批量网格中使用，缩小内边距/标题尺寸） */
-  compact?: boolean
 }
 
 function formatTime(seconds: number): string {
@@ -62,23 +52,10 @@ function formatTime(seconds: number): string {
 /**
  * 将素材映射为播放片段（复用原逻辑）
  */
-/** 简单可复现随机数（mulberry32），同一种子产出稳定排布，避免每次渲染抖动 */
-function seededRandom(seed: number): () => number {
-  let a = seed >>> 0
-  return () => {
-    a |= 0
-    a = (a + 0x6d2b79f5) | 0
-    let t = Math.imul(a ^ (a >>> 15), 1 | a)
-    t = (t + Math.imul(t ^ (t >>> 7), 61 | t)) ^ t
-    return ((t ^ (t >>> 14)) >>> 0) / 4294967296
-  }
-}
-
 function buildPlaybackSegments(
   assets: AssetItem[],
   template: EditingTemplate | null,
   serverClips?: EditPlanClip[],
-  variantSeed = 0,
 ): PlaybackSegment[] {
   if (!assets.length) return []
 
@@ -102,40 +79,18 @@ function buildPlaybackSegments(
     }
   }
 
-  // Fallback: 本地构建片段
-  // variantSeed=0（单视频）：与旧行为完全一致（素材原序、起始点 0），零回归
-  // variantSeed>0（批量变体）：素材顺序按种子轮换 + 片段起始点在素材内偏移，
-  //   模拟后端"AI 随机剪辑出不同版本"，让 N 个预览画面有可见差异
+  // Fallback: 本地构建片段（与旧行为一致）
   const templateSegments = template?.segments || []
   const segments: PlaybackSegment[] = []
-  const orderedAssets = variantSeed > 0 ? [...assets] : assets
-  if (variantSeed > 0 && orderedAssets.length > 1) {
-    const rand = seededRandom(variantSeed * 7919 + 13)
-    // 素材轮换：把数组旋转 (seed % n) 位，再对后半段做一次稳定交换
-    const n = orderedAssets.length
-    const rotate = variantSeed % n
-    orderedAssets.push(...orderedAssets.splice(0, rotate))
-    const swapA = Math.floor(rand() * n)
-    const swapB = Math.floor(rand() * n)
-    if (swapA !== swapB) {
-      ;[orderedAssets[swapA], orderedAssets[swapB]] = [orderedAssets[swapB], orderedAssets[swapA]]
-    }
-  }
 
-  orderedAssets.forEach((asset, i) => {
+  assets.forEach((asset, i) => {
     const assetDuration = asset.duration || asset.metadata?.duration || 30
     const tplSeg = templateSegments[i] || templateSegments[templateSegments.length - 1]
     const segDuration = tplSeg
       ? Math.min(tplSeg.duration_max, Math.max(tplSeg.duration_min, assetDuration))
       : Math.min(assetDuration, 10)
 
-    let startTime = 0
-    if (variantSeed > 0 && assetDuration - segDuration > 1) {
-      const rand = seededRandom(variantSeed * 104729 + i * 31 + 7)
-      // 起始点在素材可用区间内随机偏移（至少留 0.5s 余量）
-      const maxStart = Math.max(0, assetDuration - segDuration - 0.5)
-      startTime = Math.round(rand() * maxStart * 10) / 10
-    }
+    const startTime = 0
     const endTime = Math.min(startTime + segDuration, assetDuration)
     const videoUrl = asset.file_url || asset.storage_key
 
@@ -154,16 +109,11 @@ const FrontendPreviewPlayer: React.FC<FrontendPreviewPlayerProps> = ({
   voiceAudioUrl,
   titleSettings,
   onTitlePositionChange,
-  variantSeed = 0,
-  variantTitle,
-  compact = false,
 }) => {
   const segments = useMemo(
-    () => buildPlaybackSegments(assets, template, serverClips, variantSeed),
-    [assets, template, serverClips, variantSeed],
+    () => buildPlaybackSegments(assets, template, serverClips),
+    [assets, template, serverClips],
   )
-  // 批量变体：标题文字取 variantTitle，样式仍由全局 titleSettings 控制
-  const effectiveTitle = variantTitle ?? titleSettings?.title
 
   // ── ASS 坐标系参数（与后端 ass_subtitle_builder.py 一致） ──
   const TITLE_MARGIN_TOP = 120
@@ -284,7 +234,7 @@ const FrontendPreviewPlayer: React.FC<FrontendPreviewPlayerProps> = ({
   // ── Canvas 播放器（WebCodecs 路径） ──
   const canvasTitle = titleSettings
     ? {
-        text: effectiveTitle || "标题预览",
+        text: titleSettings.title || "标题预览",
         fontSize: titleSettings.size,
         fontFamily: titleSettings.font || "思源黑体",
         color: titleSettings.color || "#ffffff",
@@ -570,15 +520,13 @@ const FrontendPreviewPlayer: React.FC<FrontendPreviewPlayerProps> = ({
       style={{
         position: "relative",
         width: "100%",
-        maxWidth: compact ? "100%" : 280,
-        margin: compact ? 0 : "0 auto",
+        maxWidth: 280,
         aspectRatio: "9 / 16",
         background: "#0a0a0a",
-        borderRadius: compact ? 10 : 24,
+        borderRadius: 24,
         overflow: "hidden",
-        boxShadow: compact
-          ? "inset 0 0 0 1px rgba(255,255,255,0.06)"
-          : "0 4px 6px -1px rgba(0,0,0,0.3), 0 20px 50px -12px rgba(0,0,0,0.5), inset 0 0 0 1px rgba(255,255,255,0.06)",
+        boxShadow:
+          "0 4px 6px -1px rgba(0,0,0,0.3), 0 20px 50px -12px rgba(0,0,0,0.5), inset 0 0 0 1px rgba(255,255,255,0.06)",
       }}
     >
       {/* ── Canvas 渲染层（WebCodecs 路径） ── */}
@@ -662,8 +610,8 @@ const FrontendPreviewPlayer: React.FC<FrontendPreviewPlayerProps> = ({
                         ? { top: "50%", transform: "translate(-50%, -50%)" }
                         : { bottom: `${titleBottomPct}%` }),
                   }),
-              pointerEvents: onTitlePositionChange && variantSeed === 0 ? "auto" : "none",
-              cursor: onTitlePositionChange && variantSeed === 0 ? "grab" : "default",
+              pointerEvents: "auto",
+              cursor: onTitlePositionChange ? "grab" : "default",
               touchAction: "none",
               userSelect: "none",
               WebkitUserSelect: "none",
@@ -693,7 +641,7 @@ const FrontendPreviewPlayer: React.FC<FrontendPreviewPlayerProps> = ({
                   : undefined,
               }}
             >
-              {(effectiveTitle || "").split(/[/／]/).map((part, i) => (
+              {titleSettings.title.split(/[/／]/).map((part, i) => (
                 <span key={i}>
                   {i > 0 && <br />}
                   {part}

apps/web/src/pages/generate/components/GenerateStepActions.tsx

@@ -1,10 +1,10 @@
 /**
- * GeneratePage 步骤底部操作按钮（Issue #1677 修正：固定 6 步）
+ * GeneratePage 步骤底部操作按钮（Issue #1677 改造后 5 步）
  *
  * 步骤 1~3：上一步 / 下一步
- * 步骤 4（选择标题）：「✨ 确认生成视频 / 确认生成 N 个视频」→ 创建正式生成任务，成功后跳步骤5
- * 步骤 5（确认生成）：渲染进度页，全部完成后「下一步：选择封面」；仅上一步
- * 步骤 6（选择封面）：仅上一步
+ * 步骤 4（标题+预览+确认生成）：确认生成按钮在右侧边栏底部（含勾选数量），
+ *   渲染中显示进度；生成完成后显示"下一步 → 选择封面"
+ * 步骤 5（选择封面）：仅上一步
  */
 import React from "react"
 
@@ -12,7 +12,7 @@ export interface GenerateStepActionsProps {
   currentStep: number
   onPrev: () => void
   onNext: () => void
-  /** 步骤4：确认生成视频（校验 + 创建渲染任务） */
+  /** 步骤4：确认生成视频（校验 + 创建渲染任务 + 成功后进入步骤5） */
   onConfirmGenerate: () => void | Promise<void>
   generating: boolean
   generated: boolean
@@ -41,19 +41,26 @@ const GenerateStepActions: React.FC<GenerateStepActionsProps> = ({
       )
     }
 
-    /* 步骤 4：选择标题 — 确认生成 */
+    /* 步骤 4：标题+预览+确认生成 */
     if (currentStep === 4) {
       if (generating) {
         return (
           <button className="xx-btn xx-btn-primary" disabled>
-            ⏳ 正在提交生成任务…
+            ⏳ 视频生成中…
           </button>
         )
       }
       if (generateError) {
         return (
           <button className="xx-btn xx-btn-primary" onClick={onConfirmGenerate}>
-            🔄 重新生成
+            🔄 重新生成视频
+          </button>
+        )
+      }
+      if (generated) {
+        return (
+          <button className="xx-btn xx-btn-primary" onClick={onNext}>
+            下一步：选择封面 →
           </button>
         )
       }
@@ -64,23 +71,7 @@ const GenerateStepActions: React.FC<GenerateStepActionsProps> = ({
       )
     }
 
-    /* 步骤 5：确认生成进度页 — 全部完成后下一步进封面 */
-    if (currentStep === 5) {
-      if (generated) {
-        return (
-          <button className="xx-btn xx-btn-primary" onClick={onNext}>
-            下一步：选择封面 →
-          </button>
-        )
-      }
-      return (
-        <button className="xx-btn xx-btn-primary" disabled>
-          ⏳ 视频渲染中…
-        </button>
-      )
-    }
-
-    /* 步骤 6（封面，最后一步）：无主按钮 */
+    /* 步骤 5（封面，最后一步）：无主按钮 */
     return null
   }
 

apps/web/src/pages/generate/components/GenerateStepContent.tsx

@@ -1,7 +1,6 @@
 /**
  * GeneratePage 步骤内容渲染
- * 步骤顺序（6步，Issue #1677 修正）：模板(1) → 素材(2) → 配音(3) → 标题(4) → 确认生成(5) → 封面(6)
- * 步骤4预览（Canvas 网格）与步骤5进度（批量渲染网格）由 GeneratePage 直接渲染在左侧大区域。
+ * 步骤顺序（5步，Issue #1677）：模板(1) → 素材(2) → 配音(3) → 标题+预览+确认生成(4) → 封面(5)
  */
 import React from "react"
 import type { EditingTemplate } from "@/api/editing-planner"
@@ -13,8 +12,6 @@ import Step2MaterialSelect from "../components/Step2MaterialSelect"
 import Step3VoiceWithMode from "./Step3VoiceWithMode"
 import Step4TitleSettings from "../components/Step4TitleSettings"
 import Step6CoverSettings from "../components/Step6CoverSettings"
-import BatchGenerationGrid from "./BatchGenerationGrid"
-import type { BatchTaskState } from "../hooks/generate-video/useGenerationPolling"
 import type { GeneratedVideo } from "@/api/template-editor"
 
 export interface GenerateStepContentProps {
@@ -57,10 +54,7 @@ export interface GenerateStepContentProps {
   progress: number
   generatedVideos: GeneratedVideo[]
   onRetry: () => void
-  onRetryBatchTask: (taskId: string) => void
   onDismissError: () => void
-  /** 批量：每个正式生成任务的独立状态（步骤5进度网格） */
-  batchTasks: BatchTaskState[]
   /** BGM 开关 */
   bgm: boolean
   /** BGM 配置（来自模板） */
@@ -108,14 +102,7 @@ export const GenerateStepContent: React.FC<GenerateStepContentProps> = (props) =
     selectedVoice,
     onSelectedVoiceChange,
     onServerClipsChange,
-    generating,
-    generated,
-    generateError,
-    progress,
-    onRetry,
     generatedVideos,
-    batchTasks,
-    onRetryBatchTask,
     previewCount,
     previewTitles,
     onPreviewTitlesChange,
@@ -189,62 +176,6 @@ export const GenerateStepContent: React.FC<GenerateStepContentProps> = (props) =
         />
       )
     case 5:
-      /* 确认生成页：批量=逐任务进度网格；单视频=进度状态卡（成片播放器在左侧大区域） */
-      if (previewCount > 1) {
-        return (
-          <BatchGenerationGrid
-            tasks={batchTasks}
-            titles={previewTitles}
-            onRetryTask={onRetryBatchTask}
-          />
-        )
-      }
-      /* 单视频：渲染进度 / 失败重试 / 完成提示（成片播放器在右侧栏） */
-      return (
-        <div className="xx-form-section">
-          <h3>🎬 确认生成</h3>
-          {generating && (
-            <div className="xx-gen-progress-card">
-              <div className="xx-gen-progress-header">
-                <div className="xx-gen-progress-info">
-                  <div className="xx-gen-progress-phase">
-                    ⏳ 视频渲染中… {Math.round(progress)}%
-                  </div>
-                  <div className="xx-gen-progress-sub">
-                    生成过程中可以切换到其他页面，完成后可在任务历史查看
-                  </div>
-                </div>
-              </div>
-              <div className="xx-gen-progress-bar">
-                <div
-                  className="xx-gen-progress-bar-fill"
-                  style={{ width: `${Math.min(Math.round(progress), 100)}%` }}
-                />
-              </div>
-            </div>
-          )}
-          {generateError && !generating && (
-            <div className="xx-gen-error-card">
-              <div className="xx-gen-error-info">
-                <div className="xx-gen-error-title">生成失败</div>
-                <div className="xx-gen-error-msg">{generateError}</div>
-              </div>
-              <button type="button" className="xx-btn xx-btn-primary xx-btn-sm" onClick={onRetry}>
-                🔄 重试
-              </button>
-            </div>
-          )}
-          {generated && !generating && (
-            <div className="xx-gen-success-card">
-              <div className="xx-gen-success-info">
-                <div className="xx-gen-success-title">✅ 视频生成完成！</div>
-                <div className="xx-gen-success-sub">右侧可预览成片，点击「下一步」选择封面</div>
-              </div>
-            </div>
-          )}
-        </div>
-      )
-    case 6:
       return (
         <Step6CoverSettings
           coverSettings={coverSettings}

apps/web/src/pages/generate/components/ServerPreviewGrid.tsx

@@ -0,0 +1,127 @@
+/**
+ * 批量预览网格（Issue #1677）
+ * N 个服务器渲染的预览视频，网格排列、各自独立播放、CSS 标题浮层实时叠加、勾选框批量选择
+ */
+import React from "react"
+import { LoadingOutlined, CheckCircleFilled, CloseCircleOutlined } from "@ant-design/icons"
+import type { VariantPreview } from "../hooks/useBatchPreview"
+
+interface ServerPreviewGridProps {
+  variants: VariantPreview[]
+  /** 每个变体的标题文字（实时叠加浮层） */
+  titles: string[]
+  /** 标题样式（全局共用） */
+  titleStyle: {
+    position: string
+    color: string
+    size: number
+  }
+  /** 勾选的变体索引 */
+  selectedIds: number[]
+  onToggleSelect: (index: number) => void
+  /** 是否显示勾选框（确认生成前） */
+  selectable?: boolean
+}
+
+const ServerPreviewGrid: React.FC<ServerPreviewGridProps> = ({
+  variants,
+  titles,
+  titleStyle,
+  selectedIds,
+  onToggleSelect,
+  selectable = true,
+}) => {
+  if (variants.length === 0) return null
+
+  return (
+    <div className="xx-variant-grid">
+      {variants.map((v) => {
+        const selected = selectedIds.includes(v.index)
+        const titleText = titles[v.index] || ""
+        return (
+          <div
+            key={v.index}
+            className={`xx-variant-card ${selected ? "selected" : ""} ${
+              v.status === "failed" ? "failed" : ""
+            }`}
+            onClick={() => {
+              if (selectable && v.status === "ready") onToggleSelect(v.index)
+            }}
+            role="button"
+            tabIndex={0}
+          >
+            {/* 勾选框 */}
+            {selectable && v.status === "ready" && (
+              <div className={`xx-variant-check ${selected ? "checked" : ""}`}>
+                {selected && "✓"}
+              </div>
+            )}
+
+            {/* 变体序号 */}
+            <div className="xx-variant-index">视频 {v.index + 1}</div>
+
+            {/* 视频区域 */}
+            <div className="xx-variant-video-wrap">
+              {v.status === "loading" && (
+                <div className="xx-variant-loading">
+                  <LoadingOutlined style={{ fontSize: 28, color: "#3b82f6" }} />
+                  <div className="xx-variant-progress">
+                    <div
+                      className="xx-variant-progress-bar"
+                      style={{ width: `${Math.min(v.progress, 100)}%` }}
+                    />
+                  </div>
+                  <span className="xx-variant-progress-text">{v.progress}%</span>
+                </div>
+              )}
+              {v.status === "failed" && (
+                <div className="xx-variant-failed">
+                  <CloseCircleOutlined style={{ fontSize: 28, color: "#ef4444" }} />
+                  <span>{v.error || "预览失败"}</span>
+                </div>
+              )}
+              {v.status === "ready" && v.videoUrl && (
+                <>
+                  <video
+                    src={v.videoUrl}
+                    controls
+                    style={{ width: "100%", display: "block", background: "#000", borderRadius: 8 }}
+                    onClick={(e) => e.stopPropagation()}
+                  />
+                  {/* 标题浮层（CSS 实时叠加，改标题即时可见） */}
+                  {titleText && (
+                    <div
+                      className={`xx-variant-title-overlay pos-${titleStyle.position}`}
+                      style={{
+                        color: titleStyle.color,
+                        fontSize: Math.max(13, Math.round(titleStyle.size * 0.55)),
+                        WebkitTextStroke: "0.5px rgba(0,0,0,0.6)",
+                      }}
+                      onClick={(e) => e.stopPropagation()}
+                    >
+                      {titleText}
+                    </div>
+                  )}
+                </>
+              )}
+            </div>
+
+            {/* 底部状态 */}
+            <div className="xx-variant-footer">
+              {v.status === "ready" && selected && (
+                <span className="xx-variant-ready-tag">
+                  <CheckCircleFilled style={{ color: "#52c41a" }} /> 已选择
+                </span>
+              )}
+              {v.status === "ready" && !selected && selectable && (
+                <span className="xx-variant-skip-tag">点击卡片取消/勾选</span>
+              )}
+            </div>
+          </div>
+        )
+      })}
+    </div>
+  )
+}
+
+export default ServerPreviewGrid

apps/web/src/pages/generate/components/Step4TitleSettings.tsx

@@ -1,22 +1,18 @@
 /**
- * Step 4 选择标题（Issue #1677 批量生成）
+ * Step 4 选择标题（Issue #1677 批量生成改造）
  *
- * 布局（由 GeneratePage 编排）：左侧大区域实时预览（单=大播放器，批量=Canvas 网格），
- * 右侧边栏标题设置。本组件渲染在右侧边栏：
- * - 单视频：AI 标题生成器 + AutoComplete 标题库（与旧版完全一致，零回归）
- * - 批量：N 个独立标题输入框（AutoComplete 支持标题库选择）+ 批量 AI 生成
- *   （一次生成 N 个标题，分别填入各变体，可单独换一个）
+ * 布局（由 GeneratePage 编排）：左侧大区域预览，右侧边栏标题设置。
+ * 本组件渲染在右侧边栏：
+ * - 标题文字：1 个视频 1 个输入框；N 个视频 N 个输入框各自独立
  * - 标题样式（字体/颜色/位置/大小/粗斜描边/预设）：全局统一
  */
-import React, { useMemo, useState } from "react"
-import { AutoComplete, Input, message } from "antd"
-import { LoadingOutlined } from "@ant-design/icons"
+import React from "react"
+import { AutoComplete, Input } from "antd"
 import type { TitleSettings } from "../types"
 import { POSITION_OPTIONS, FONT_OPTIONS } from "../constants"
 import { useStep4Title } from "../hooks/useStep4Title"
 import AiTitleGenerator from "./title/AiTitleGenerator"
 import TitleStylePanel from "./title/TitleStylePanel"
-import { AI_TITLE_TEMPLATES } from "../constants"
 
 interface Step4TitleSettingsProps {
   titleSettings: TitleSettings
@@ -42,36 +38,6 @@ interface Step4TitleSettingsProps {
   onPreviewTitlesChange?: (titles: string[]) => void
 }
 
-/** 从本地 AI 标题模板池按主题词生成 N 个不同标题（与单视频 AI 生成同源） */
-function buildBatchAiTitles(topic: string, count: number): string[] {
-  const styles: Array<"catchy" | "emotional" | "informative"> = [
-    "catchy",
-    "emotional",
-    "informative",
-  ]
-  const pool: string[] = []
-  styles.forEach((style) => {
-    const templates = AI_TITLE_TEMPLATES[style] || []
-    templates.forEach((tpl) => pool.push(tpl.replace(/\{topic\}/g, topic)))
-  })
-  // 洗牌后取前 count 个；不足则轮转补齐
-  const shuffled = [...pool].sort(() => Math.random() - 0.5)
-  const out: string[] = []
-  for (let i = 0; i < count; i++) {
-    out.push(shuffled[i % shuffled.length] || "")
-  }
-  return out
-}
-
-function extractTopic(text: string): string {
-  const keywords = text
-    .replace(/[，。！？、,.!?]/g, " ")
-    .split(/\s+/)
-    .filter(Boolean)
-  if (keywords.length === 0) return "这个话题"
-  return keywords.slice(0, 3).join("")
-}
-
 const Step4TitleSettings: React.FC<Step4TitleSettingsProps> = (props) => {
   const t = useStep4Title(props)
   const {
@@ -91,10 +57,8 @@ const Step4TitleSettings: React.FC<Step4TitleSettingsProps> = (props) => {
   } = props
 
   const isBatch = previewCount > 1
-  const [batchAiLoading, setBatchAiLoading] = useState(false)
-  const [batchAiTopic, setBatchAiTopic] = useState("")
 
-  /** 更新单个变体标题；变体0同步写回 titleSettings.title（全局样式面板/草稿/TTS 链路依赖） */
+  /** 更新单个变体标题；变体0同步写回 titleSettings.title（全局样式面板/草稿保存依赖） */
   const updateVariantTitle = (index: number, val: string) => {
     if (!previewTitles || !onPreviewTitlesChange) return
     const next = [...previewTitles]
@@ -105,43 +69,12 @@ const Step4TitleSettings: React.FC<Step4TitleSettingsProps> = (props) => {
     }
   }
 
-  /** 批量 AI 生成：按主题词生成标题，分别填入 N 个变体 */
-  const handleBatchAiGenerate = async (onlyEmpty = false) => {
-    if (!onPreviewTitlesChange || !previewTitles) return
-    const topic = (batchAiTopic || t.aiTitleInput || "").trim()
-    if (!topic) {
-      message.warning("请先输入主题词，例如：萌宠日常、旅行vlog")
-      return
-    }
-    setBatchAiLoading(true)
-    try {
-      // 与单视频一致：本地模板模拟 AI 生成（1200ms 体验延迟）
-      await new Promise((resolve) => setTimeout(resolve, 800))
-      const picked = buildBatchAiTitles(extractTopic(topic), previewCount)
-      const next = [...previewTitles]
-      for (let i = 0; i < previewCount; i++) {
-        if (onlyEmpty && next[i]?.trim()) continue
-        if (picked[i]) next[i] = picked[i]
-      }
-      onPreviewTitlesChange(next)
-      if (next[0]) t.updateTitle(next[0])
-      message.success(`已为 ${previewCount} 个视频生成标题，可单独修改`)
-    } finally {
-      setBatchAiLoading(false)
-    }
-  }
-
-  const titleOptions = useMemo(
-    () => t.userTitles.map((ut) => ({ label: ut.content, value: ut.content })),
-    [t.userTitles],
-  )
-
   return (
     <div className="xx-form-section xx-title-sidebar">
       <h3>📝 选择标题</h3>
 
       {!isBatch ? (
-        /* ── 单视频：原有 AI 标题 + 输入框（保持不变，零回归） ── */
+        /* ── 单视频：原有 AI 标题 + 输入框（保持不变） ── */
         <>
           {t.titleSettings.aiAutoSelect ? (
             <>
@@ -211,7 +144,7 @@ const Step4TitleSettings: React.FC<Step4TitleSettingsProps> = (props) => {
                     t.updateTitle(val || "")
                     onPreviewTitlesChange?.([val || ""])
                   }}
-                  options={titleOptions}
+                  options={t.userTitles.map((ut) => ({ label: ut.content, value: ut.content }))}
                   filterOption={(inputValue, option) => {
                     const title = (option?.label || option?.value || "") as string
                     return title.toLowerCase().includes((inputValue || "").toLowerCase())
@@ -222,7 +155,7 @@ const Step4TitleSettings: React.FC<Step4TitleSettingsProps> = (props) => {
           )}
         </>
       ) : (
-        /* ── 批量：AI 批量生成 + N 个独立标题输入框（AutoComplete 支持标题库） ── */
+        /* ── 批量：N 个独立标题输入框（CSS 浮层实时叠加到对应预览） ── */
         <div className="xx-batch-titles">
           <div
             style={{
@@ -234,52 +167,15 @@ const Step4TitleSettings: React.FC<Step4TitleSettingsProps> = (props) => {
           >
             为每个视频输入独立标题，修改会实时叠加到左侧对应视频上。标题样式（字体/颜色/位置）全局统一。
           </div>
-
-          {/* 批量 AI 标题 */}
-          <div className="xx-batch-ai-row">
-            <Input
-              placeholder="主题词，如：萌宠日常、旅行vlog"
-              value={batchAiTopic || t.aiTitleInput}
-              onChange={(e) => {
-                setBatchAiTopic(e.target.value)
-                t.setAiTitleInput(e.target.value)
-              }}
-              maxLength={30}
-              size="small"
-              style={{ flex: 1 }}
-            />
-            <button
-              type="button"
-              className="xx-btn xx-btn-primary xx-btn-sm"
-              disabled={batchAiLoading}
-              onClick={() => handleBatchAiGenerate(false)}
-            >
-              {batchAiLoading ? <LoadingOutlined /> : "✨"} 一键生成 {previewCount} 个标题
-            </button>
-            <button
-              type="button"
-              className="xx-btn xx-btn-ghost xx-btn-sm"
-              disabled={batchAiLoading}
-              onClick={() => handleBatchAiGenerate(true)}
-            >
-              补填空标题
-            </button>
-          </div>
-
           {Array.from({ length: previewCount }, (_, i) => (
             <div className="xx-form-field" key={i}>
               <label>视频 {i + 1} 标题</label>
-              <AutoComplete
+              <Input
                 placeholder={`视频 ${i + 1} 的标题…`}
                 maxLength={50}
-                style={{ width: "100%" }}
-                value={previewTitles?.[i] || undefined}
-                onChange={(val) => updateVariantTitle(i, val || "")}
-                options={titleOptions}
-                filterOption={(inputValue, option) => {
-                  const title = (option?.label || option?.value || "") as string
-                  return title.toLowerCase().includes((inputValue || "").toLowerCase())
-                }}
+                showCount
+                value={previewTitles?.[i] || ""}
+                onChange={(e) => updateVariantTitle(i, e.target.value)}
               />
             </div>
           ))}

apps/web/src/pages/generate/constants.ts

@@ -33,8 +33,7 @@ export const STEPS = [
   { key: 2, label: "选择素材" },
   { key: 3, label: "选择配音" },
   { key: 4, label: "选择标题" },
-  { key: 5, label: "确认生成" },
-  { key: 6, label: "选择封面" },
+  { key: 5, label: "选择封面" },
 ]
 
 /* ── 批量生成限制 ── */

apps/web/src/pages/generate/generate.css

@@ -3278,167 +3278,3 @@
     max-height: none;
   }
 }
-
-/* ============================================================
-   批量前端 Canvas 预览网格（Issue #1677 修正：纯前端实时预览）
-   ============================================================ */
-.xx-canvas-grid {
-  display: grid;
-  grid-template-columns: repeat(2, 1fr);
-  gap: 16px;
-}
-
-.xx-canvas-grid-card {
-  border: 2px solid var(--border-primary, #e2e8f0);
-  border-radius: 12px;
-  overflow: hidden;
-  background: #000;
-  transition: border-color 0.2s ease;
-  min-width: 0;
-}
-
-.xx-canvas-grid-card.selected {
-  border-color: var(--primary-color, #1677ff);
-  box-shadow: 0 0 0 2px rgba(22, 119, 255, 0.15);
-}
-
-.xx-canvas-grid-card-bar {
-  position: relative;
-  z-index: 2;
-  display: flex;
-  align-items: center;
-  padding: 6px 10px;
-  background: var(--bg-surface, #fff);
-  border-bottom: 1px solid var(--border-primary, #e2e8f0);
-}
-
-.xx-canvas-grid-check {
-  display: inline-flex;
-  align-items: center;
-  gap: 6px;
-  font-size: 13px;
-  font-weight: 500;
-  color: var(--text-primary, #1a1a1a);
-  cursor: pointer;
-  user-select: none;
-}
-
-.xx-canvas-grid-check input[type="checkbox"] {
-  width: 15px;
-  height: 15px;
-  cursor: pointer;
-  accent-color: var(--primary-color, #1677ff);
-}
-
-/* ============================================================
-   批量标题：AI 一键生成行（Issue #1677）
-   ============================================================ */
-.xx-batch-ai-row {
-  display: flex;
-  flex-wrap: wrap;
-  align-items: center;
-  gap: 8px;
-  padding: 10px 12px;
-  margin-bottom: 12px;
-  background: var(--bg-secondary, #f7f8fa);
-  border: 1px dashed var(--border-primary, #d9d9d9);
-  border-radius: 10px;
-}
-
-.xx-batch-ai-row .xx-form-field {
-  margin: 0;
-  flex: 1;
-  min-width: 140px;
-}
-
-.xx-batch-titles {
-  display: flex;
-  flex-direction: column;
-  gap: 10px;
-}
-
-/* ============================================================
-   第5步确认生成：批量渲染进度网格（Issue #1677）
-   ============================================================ */
-.xx-batch-gen-grid {
-  display: grid;
-  grid-template-columns: repeat(2, 1fr);
-  gap: 16px;
-}
-
-.xx-batch-gen-card {
-  border: 1px solid var(--border-primary, #e2e8f0);
-  border-radius: 12px;
-  padding: 14px;
-  background: var(--bg-surface, #fff);
-  display: flex;
-  flex-direction: column;
-  gap: 10px;
-  min-width: 0;
-}
-
-.xx-batch-gen-card.status-completed {
-  border-color: rgba(82, 196, 26, 0.4);
-  background: rgba(82, 196, 26, 0.04);
-}
-
-.xx-batch-gen-card.status-failed {
-  border-color: rgba(239, 68, 68, 0.4);
-  background: rgba(239, 68, 68, 0.04);
-}
-
-.xx-batch-gen-card-head {
-  display: flex;
-  align-items: center;
-  justify-content: space-between;
-  gap: 8px;
-}
-
-.xx-batch-gen-card-title {
-  font-size: 14px;
-  font-weight: 600;
-  color: var(--text-primary, #1a1a1a);
-  overflow: hidden;
-  text-overflow: ellipsis;
-  white-space: nowrap;
-}
-
-.xx-batch-gen-card-body {
-  display: flex;
-  flex-direction: column;
-  gap: 8px;
-}
-
-.xx-batch-gen-card-pct {
-  font-size: 13px;
-  color: var(--text-secondary, #666);
-  text-align: right;
-}
-
-.xx-batch-gen-card-done {
-  font-size: 13px;
-  color: var(--success-color, #52c41a);
-  padding: 8px 0;
-}
-
-.xx-batch-gen-card-failed {
-  display: flex;
-  flex-direction: column;
-  gap: 8px;
-  align-items: flex-start;
-}
-
-.xx-batch-gen-card-err {
-  font-size: 13px;
-  color: var(--error-color, #ef4444);
-  line-height: 1.5;
-  word-break: break-word;
-}
-
-/* ── 响应式：窄屏批量网格回退单列 ── */
-@media (max-width: 960px) {
-  .xx-canvas-grid,
-  .xx-batch-gen-grid {
-    grid-template-columns: 1fr;
-  }
-}

apps/web/src/pages/generate/hooks/generate-video/useGenerationPolling.ts

@@ -1,28 +1,14 @@
-import { useRef, useCallback, useState } from "react"
+import { useRef, useCallback } from "react"
 import { message } from "antd"
 import axios from "axios"
-import { getGenerationTask, retryTask as retryGenerationTaskApi } from "@/api/tasks/tasks"
+import { getGenerationTask } from "@/api/tasks/tasks"
 import { getGenerationTaskResults } from "@/api/template-editor"
 import { safeExtractError } from "./errorUtils"
 
-/** 批量生成时单个任务的实时状态（Issue #1677 第5步确认生成页） */
-export interface BatchTaskState {
-  taskId: string
-  /** 变体序号（0-based，与标题/封面数组对齐） */
-  variantIndex: number
-  status: "running" | "completed" | "failed"
-  progress: number
-  error: string | null
-  /** 完成后的成片视频 */
-  videos: unknown[]
-}
-
 interface UseGenerationPollingOptions {
   onProgress: (progress: number) => void
   onComplete: (videos: unknown[]) => void
   onFailed: (errorMsg: string) => void
-  /** 批量：单任务状态变化（第5步逐卡片展示） */
-  onBatchTaskUpdate?: (taskId: string, patch: Partial<BatchTaskState>) => void
 }
 
 /** 最大连续错误次数（仅对可重试错误），超过后终止轮询 */
@@ -31,25 +17,20 @@ const MAX_RETRYABLE_ERRORS = 10
 const MAX_RESULTS_RETRIES = 3
 
 /**
- * 生成状态轮询 Hook（v4 — 批量任务独立状态 + 单任务重试）
+ * 生成状态轮询 Hook（v3 — 支持批量多任务）
  *
  * startPolling(taskId) 轮询单个任务；
- * startPollingBatch(tasks) 并行轮询 N 个任务：
- * - 每个任务独立进度/状态/失败，通过 onBatchTaskUpdate 实时回传
- * - 全部成功才 onComplete（聚合视频按变体顺序）；任一失败不影响其他任务继续
- * - retryTask(taskId) 单独重试失败任务（重新轮询，后端任务仍在跑则直接接续）
+ * startPollingBatch(taskIds) 并行轮询 N 个任务，全部完成后聚合结果，
+ * 任一任务失败即整体失败（其余任务仍在后端继续，不影响）。
+ * 进度为所有任务平均值。
  */
 export function useGenerationPolling({
   onProgress,
   onComplete,
   onFailed,
-  onBatchTaskUpdate,
 }: UseGenerationPollingOptions) {
   const progressTimer = useRef<ReturnType<typeof setTimeout>[]>([])
   const cancelledRef = useRef(false)
-  /** 批量任务上下文：taskId → 变体序号 */
-  const batchContextRef = useRef<Map<string, number>>(new Map())
-  const [, forceTick] = useState(0)
 
   const clearTimer = useCallback(() => {
     cancelledRef.current = true
@@ -85,21 +66,9 @@ export function useGenerationPolling({
     return safeExtractError(msg)
   }
 
-  /**
-   * 轮询单个任务。
-   * - isBatch=true：状态变化通过 onBatchTaskUpdate 回传，不触发整体 onProgress/onComplete
-   * - resolve(videos) 成功；reject(Error) 失败
-   */
+  /** 轮询单个任务，resolve 该任务的结果视频数组；失败时 reject(new Error(msg)) */
   const pollSingleTask = useCallback(
-    (
-      taskId: string,
-      runId: number,
-      callbacks?: {
-        onTaskProgress?: (pct: number) => void
-        onTaskCompleted?: (videos: unknown[]) => void
-        onTaskFailed?: (msg: string) => void
-      },
-    ): Promise<unknown[]> => {
+    (taskId: string, runId: number, onTaskProgress?: (pct: number) => void): Promise<unknown[]> => {
       return new Promise((resolve, reject) => {
         let consecutiveErrors = 0
         let done = false
@@ -116,12 +85,9 @@ export function useGenerationPolling({
               const videos = await fetchResultsWithRetry(taskId)
               if (cancelledRef.current) return
               if (videos === null) {
-                const msg = "视频已生成，但获取结果列表失败，请稍后在任务列表查看"
-                callbacks?.onTaskFailed?.(msg)
-                reject(new Error(msg))
+                reject(new Error("视频已生成，但获取结果列表失败，请稍后在任务列表查看"))
                 return
               }
-              callbacks?.onTaskCompleted?.(videos)
               resolve(videos)
               return
             }
@@ -132,15 +98,14 @@ export function useGenerationPolling({
                 task.error_info?.error_message ||
                 task.error_message ||
                 (task.status === "cancelled" ? "任务已取消" : "视频生成失败，请联系管理员或重试")
-              const msg = safeExtractError(rawMsg)
-              callbacks?.onTaskFailed?.(msg)
-              reject(new Error(msg))
+              reject(new Error(safeExtractError(rawMsg)))
               return
             }
 
             const pct = Math.max(0, Math.min(99, Math.round(Number(task.progress) || 0)))
-            callbacks?.onTaskProgress?.(pct)
-            if (!callbacks && runId === 0) {
+            if (onTaskProgress) {
+              onTaskProgress(pct)
+            } else if (runId === 0) {
               onProgress(pct)
             }
             const timer = setTimeout(poll, 2000)
@@ -151,17 +116,13 @@ export function useGenerationPolling({
             const status = axios.isAxiosError(pollErr) ? pollErr.response?.status : undefined
             if (status && status >= 400 && status < 500) {
               done = true
-              const msg = extractErrorMessage(pollErr, status)
-              callbacks?.onTaskFailed?.(msg)
-              reject(new Error(msg))
+              reject(new Error(extractErrorMessage(pollErr, status)))
               return
             }
             consecutiveErrors += 1
             if (consecutiveErrors >= MAX_RETRYABLE_ERRORS) {
               done = true
-              const msg = "任务状态查询连续失败，请稍后在任务列表查看结果"
-              callbacks?.onTaskFailed?.(msg)
-              reject(new Error(msg))
+              reject(new Error("任务状态查询连续失败，请稍后在任务列表查看结果"))
               return
             }
             const timer = setTimeout(poll, 3000)
@@ -176,12 +137,12 @@ export function useGenerationPolling({
     [onProgress, fetchResultsWithRetry],
   )
 
-  /** 单任务轮询（单视频，兼容旧调用） */
+  /** 单任务轮询（兼容旧调用） */
   const startPolling = useCallback(
     (taskId: string) => {
       cancelledRef.current = false
-      batchContextRef.current.clear()
-      pollSingleTask(taskId, 0)
+      const runId = 0
+      pollSingleTask(taskId, runId)
         .then((videos) => {
           if (cancelledRef.current) return
           onProgress(100)
@@ -198,114 +159,48 @@ export function useGenerationPolling({
     [pollSingleTask, onProgress, onComplete, onFailed],
   )
 
-  /**
-   * 批量多任务轮询：
-   * - 每个任务独立进度/状态回传 onBatchTaskUpdate
-   * * 全部完成后按变体顺序聚合视频 onComplete
-   * - 部分失败：整体不 onFailed（第5步逐卡片展示失败+重试按钮）；全部失败才 onFailed
-   */
+  /** 批量多任务轮询：全部完成后聚合结果；任一失败即整体失败 */
   const startPollingBatch = useCallback(
-    (tasks: { taskId: string; variantIndex: number }[]) => {
+    (taskIds: string[]) => {
       cancelledRef.current = false
       const runId = Date.now()
       const progressMap = new Map<string, number>()
-      const resultMap = new Map<string, unknown[]>()
-      const failureMap = new Map<string, string>()
-      batchContextRef.current = new Map(tasks.map((t) => [t.taskId, t.variantIndex]))
 
       const reportAggregateProgress = () => {
         if (cancelledRef.current) return
-        const values = tasks.map((t) => progressMap.get(t.taskId) ?? 0)
+        const values = taskIds.map((id) => progressMap.get(id) ?? 0)
         const avg = Math.round(values.reduce((a, b) => a + b, 0) / Math.max(values.length, 1))
         onProgress(Math.min(avg, 99))
       }
 
-      const checkAllSettled = () => {
-        if (resultMap.size + failureMap.size < tasks.length) return
-        if (resultMap.size === tasks.length) {
-          onProgress(100)
-          const ordered = tasks.map((t) => resultMap.get(t.taskId) || []).flat()
-          onComplete(ordered)
-          message.success(`全部 ${tasks.length} 个视频生成完成！`)
-        } else if (resultMap.size > 0) {
-          // 部分失败：成功的视频聚合进成片列表（可进封面），失败卡片带重试按钮
-          onProgress(100)
-          const ordered = tasks
-            .filter((t) => resultMap.has(t.taskId))
-            .map((t) => resultMap.get(t.taskId) || [])
-            .flat()
-          onComplete(ordered)
-          message.warning(
-            `${failureMap.size} 个视频生成失败，可点击卡片上的「重试此视频」，成功的视频可先进入下一步`,
-          )
-        } else {
-          const firstMsg = failureMap.get(tasks[0].taskId) || "全部视频生成失败"
-          onFailed(firstMsg)
-        }
-      }
+      const tasks = taskIds.map((taskId) =>
+        pollSingleTask(taskId, runId, (pct) => {
+          progressMap.set(taskId, pct)
+          reportAggregateProgress()
+        }).then((videos) => {
+          progressMap.set(taskId, 100)
+          reportAggregateProgress()
+          return videos
+        }),
+      )
 
-      tasks.forEach(({ taskId, variantIndex }) => {
-        onBatchTaskUpdate?.(taskId, {
-          taskId,
-          variantIndex,
-          status: "running",
-          progress: 0,
-          error: null,
-          videos: [],
+      Promise.all(tasks)
+        .then((results) => {
+          if (cancelledRef.current) return
+          onProgress(100)
+          const allVideos = results.flat()
+          onComplete(allVideos)
+          message.success(`全部 ${taskIds.length} 个视频生成完成！`)
         })
-        pollSingleTask(taskId, runId, {
-          onTaskProgress: (pct) => {
-            progressMap.set(taskId, pct)
-            onBatchTaskUpdate?.(taskId, { status: "running", progress: pct })
-            reportAggregateProgress()
-          },
-          onTaskCompleted: (videos) => {
-            progressMap.set(taskId, 100)
-            resultMap.set(taskId, videos)
-            onBatchTaskUpdate?.(taskId, { status: "completed", progress: 100, videos })
-            reportAggregateProgress()
-            checkAllSettled()
-          },
-          onTaskFailed: (msg) => {
-            failureMap.set(taskId, msg)
-            onBatchTaskUpdate?.(taskId, { status: "failed", error: msg })
-            checkAllSettled()
-          },
-        }).catch(() => {
-          // 失败已在 onTaskFailed 处理，这里吞掉 Promise rejection
+        .catch((err: Error) => {
+          if (cancelledRef.current) return
+          console.error("[批量生成失败]", err.message)
+          onFailed(err.message)
+          message.error(err.message)
         })
-      })
     },
-    [pollSingleTask, onProgress, onComplete, onFailed, onBatchTaskUpdate],
+    [pollSingleTask, onProgress, onComplete, onFailed],
   )
 
-  /** 单独重试失败任务（第5步卡片「重试此视频」）：先调后端重试接口，再轮询 */
-  const retryTask = useCallback(
-    async (taskId: string) => {
-      if (cancelledRef.current) cancelledRef.current = false
-      const variantIndex = batchContextRef.current.get(taskId) ?? 0
-      onBatchTaskUpdate?.(taskId, { status: "running", progress: 0, error: null, videos: [] })
-      try {
-        await retryGenerationTaskApi(taskId)
-      } catch (err) {
-        // 后端不支持重试或任务不可重试：直接重新轮询（任务可能已被自动恢复）
-        console.warn("[重试任务接口调用失败，改为直接轮询]", err)
-      }
-      pollSingleTask(taskId, Date.now(), {
-        onTaskProgress: (pct) => onBatchTaskUpdate?.(taskId, { status: "running", progress: pct }),
-        onTaskCompleted: (videos) => {
-          onBatchTaskUpdate?.(taskId, { status: "completed", progress: 100, videos })
-          message.success(`视频 ${variantIndex + 1} 重试成功`)
-        },
-        onTaskFailed: (msg) => onBatchTaskUpdate?.(taskId, { status: "failed", error: msg }),
-      }).catch(() => {
-        /* 失败已在回调处理 */
-      })
-      forceTick((n) => n + 1)
-      return variantIndex
-    },
-    [pollSingleTask, onBatchTaskUpdate],
-  )
-
-  return { startPolling, startPollingBatch, retryTask, clearTimer }
+  return { startPolling, startPollingBatch, clearTimer }
 }

apps/web/src/pages/generate/hooks/useBatchPreview.ts

@@ -0,0 +1,285 @@
+/**
+ * 批量服务器预览 Hook（Issue #1677 多视频批量生成）
+ *
+ * 核心职责：
+ * 1. 调用 POST /generation/preview（preview_count=N）一次创建 N 个独立变体任务
+ * 2. 对每个变体 task_id 分别轮询 GET /generation/preview/{task_id}
+ * 3. 返回每个变体的状态/进度/视频URL，供网格播放器展示
+ *
+ * N=1 时不启用（走前端 Canvas 实时预览，零回归）；
+ * N>1 时进入标题页自动触发；素材/配音等配置变化后重新触发。
+ */
+import { useState, useCallback, useRef, useEffect } from "react"
+import { createPreview, getPreviewStatus } from "@/api/generation/preview"
+import type { CreatePreviewRequest } from "@/api/generation/types"
+
+export type VariantPreviewStatus = "loading" | "ready" | "failed"
+
+export interface VariantPreview {
+  /** 变体序号（0-based） */
+  index: number
+  taskId: string
+  status: VariantPreviewStatus
+  progress: number
+  videoUrl: string | null
+  error: string | null
+}
+
+interface UseBatchPreviewOptions {
+  /** 是否启用（仅 previewCount>1 且在标题页时启用） */
+  enabled: boolean
+  /** 构建预览请求参数（每次触发时调用，获取最新配置） */
+  buildRequest: () => CreatePreviewRequest
+  /** 批量预览任务创建成功回调（回传变体 taskId 列表与 source_edit_plan_id） */
+  onPreviewTasksCreated?: (taskIds: string[], sourceEditPlanId?: string) => void
+}
+
+interface UseBatchPreviewReturn {
+  variants: VariantPreview[]
+  /** 整体状态：loading=任一进行中，ready=全部完成，failed=有失败 */
+  status: "idle" | "loading" | "ready" | "partial_failed" | "failed"
+  /** 总进度 0-100（各变体平均值） */
+  progress: number
+  /** 失败的变体数量 */
+  failedCount: number
+  /** 手动重新触发 */
+  trigger: () => void
+}
+
+const POLL_INTERVAL = 2000
+const POLL_TIMEOUT = 180_000
+const MAX_NETWORK_RETRIES = 2
+
+/**
+ * 对配置参数做指纹，用于检测配置是否变化（标题文字/样式变化不触发重渲染，仅CSS浮层叠加）
+ */
+function buildFingerprint(req: CreatePreviewRequest): string {
+  // 不含 titles/title_config：标题文字与样式由 CSS 浮层实时叠加，变化不触发重渲染
+  return JSON.stringify({
+    t: req.template_id,
+    a: [...(req.asset_ids || [])].sort(),
+    r: req.video_ratio,
+    v: req.voice_library_id,
+    vs: req.voice_library_ids,
+    pc: req.preview_count,
+    b: req.bgm_config,
+  })
+}
+
+export function useBatchPreview({
+  enabled,
+  buildRequest,
+  onPreviewTasksCreated,
+}: UseBatchPreviewOptions): UseBatchPreviewReturn {
+  const [variants, setVariants] = useState<VariantPreview[]>([])
+  const [status, setStatus] = useState<"idle" | "loading" | "ready" | "partial_failed" | "failed">(
+    "idle",
+  )
+  const requestSeqRef = useRef(0)
+  const pollTimersRef = useRef<ReturnType<typeof setTimeout>[]>([])
+  const timeoutTimerRef = useRef<ReturnType<typeof setTimeout> | null>(null)
+  const mountedRef = useRef(true)
+
+  const buildRequestRef = useRef(buildRequest)
+  buildRequestRef.current = buildRequest
+  const onCreatedRef = useRef(onPreviewTasksCreated)
+  onCreatedRef.current = onPreviewTasksCreated
+
+  const clearTimers = useCallback(() => {
+    pollTimersRef.current.forEach((t) => clearTimeout(t))
+    pollTimersRef.current = []
+    if (timeoutTimerRef.current) {
+      clearTimeout(timeoutTimerRef.current)
+      timeoutTimerRef.current = null
+    }
+  }, [])
+
+  useEffect(() => {
+    mountedRef.current = true
+    return () => {
+      mountedRef.current = false
+      clearTimers()
+    }
+  }, [clearTimers])
+
+  /** 更新单个变体状态 */
+  const patchVariant = useCallback((taskId: string, patch: Partial<VariantPreview>) => {
+    setVariants((prev) => prev.map((v) => (v.taskId === taskId ? { ...v, ...patch } : v)))
+  }, [])
+
+  /** 轮询单个变体任务 */
+  const pollVariant = useCallback(
+    async (taskId: string, seq: number, retries = 0) => {
+      if (seq !== requestSeqRef.current || !mountedRef.current) return
+      try {
+        const st = await getPreviewStatus(taskId)
+        if (seq !== requestSeqRef.current || !mountedRef.current) return
+
+        if (st.status === "completed" && st.video_url) {
+          patchVariant(taskId, {
+            status: "ready",
+            videoUrl: st.video_url,
+            progress: 100,
+            error: null,
+          })
+          return
+        }
+        if (st.status === "failed" || st.status === "cancelled") {
+          patchVariant(taskId, {
+            status: "failed",
+            error:
+              st.status === "cancelled" ? "预览任务已取消" : st.error_message || "预览渲染失败",
+          })
+          return
+        }
+        if (typeof st.progress === "number") {
+          patchVariant(taskId, { progress: Math.round(st.progress) })
+        }
+        const timer = setTimeout(() => pollVariant(taskId, seq), POLL_INTERVAL)
+        pollTimersRef.current.push(timer)
+      } catch (err) {
+        if (seq !== requestSeqRef.current || !mountedRef.current) return
+        if (retries < MAX_NETWORK_RETRIES) {
+          console.warn(`[BatchPreview] 变体 ${taskId} 轮询网络错误，第 ${retries + 1} 次重试`, err)
+          const timer = setTimeout(() => pollVariant(taskId, seq, retries + 1), POLL_INTERVAL * 2)
+          pollTimersRef.current.push(timer)
+        } else {
+          patchVariant(taskId, { status: "failed", error: "网络错误，无法获取预览状态" })
+        }
+      }
+    },
+    [patchVariant],
+  )
+
+  /** 创建批量预览任务并开始轮询 */
+  const trigger = useCallback(() => {
+    if (!enabled) return
+    const request = buildRequestRef.current()
+    if (!request.template_id || !request.asset_ids?.length) return
+    const count = request.preview_count && request.preview_count > 1 ? request.preview_count : 0
+    if (!count) return
+
+    clearTimers()
+    const seq = ++requestSeqRef.current
+    setStatus("loading")
+    setVariants(
+      Array.from({ length: count }, (_, i) => ({
+        index: i,
+        taskId: "",
+        status: "loading" as const,
+        progress: 0,
+        videoUrl: null,
+        error: null,
+      })),
+    )
+
+    createPreview(request)
+      .then((resp) => {
+        if (seq !== requestSeqRef.current || !mountedRef.current) return
+        const items = resp.items || []
+        const taskIds = items.map((it) => it.task_id).filter(Boolean)
+        if (taskIds.length === 0) {
+          setStatus("failed")
+          setVariants((prev) =>
+            prev.map((v) => ({ ...v, status: "failed", error: "未创建预览任务" })),
+          )
+          return
+        }
+        onCreatedRef.current?.(taskIds, resp.source_edit_plan_id)
+
+        // 用返回的 task_id 填充变体（按 variant_index 对齐）
+        setVariants((prev) =>
+          prev.map((v) => {
+            const item = items.find((it) => it.variant_index === v.index) || items[v.index]
+            return item ? { ...v, taskId: item.task_id } : v
+          }),
+        )
+
+        // 超时保护
+        timeoutTimerRef.current = setTimeout(() => {
+          if (seq !== requestSeqRef.current || !mountedRef.current) return
+          setVariants((prev) =>
+            prev.map((v) =>
+              v.status === "loading"
+                ? { ...v, status: "failed", error: "预览渲染超时，请重试" }
+                : v,
+            ),
+          )
+        }, POLL_TIMEOUT)
+
+        // 分别轮询每个变体
+        items.forEach((item) => {
+          if (item.task_id) pollVariant(item.task_id, seq)
+        })
+      })
+      .catch((err: unknown) => {
+        if (seq !== requestSeqRef.current || !mountedRef.current) return
+        console.error("[BatchPreview] 创建批量预览失败:", err)
+        const errData = (err as { response?: { data?: { detail?: string; message?: string } } })
+          ?.response?.data
+        setStatus("failed")
+        setVariants((prev) =>
+          prev.map((v) => ({
+            ...v,
+            status: "failed",
+            error: errData?.detail || errData?.message || "预览任务创建失败，请重试",
+          })),
+        )
+      })
+  }, [enabled, clearTimers, pollVariant])
+
+  /* ── 自动触发 + 配置变更检测 ── */
+  const request = enabled ? buildRequest() : null
+  const currentFingerprint = request
+    ? request.template_id && request.asset_ids?.length && (request.preview_count || 1) > 1
+      ? buildFingerprint(request)
+      : ""
+    : ""
+
+  const didInitRef = useRef(false)
+  useEffect(() => {
+    if (!enabled || !currentFingerprint) {
+      didInitRef.current = false
+      requestSeqRef.current += 1
+      clearTimers()
+      setStatus("idle")
+      setVariants([])
+      return
+    }
+    if (!didInitRef.current) {
+      didInitRef.current = true
+      trigger()
+    }
+  }, [enabled, currentFingerprint, trigger, clearTimers])
+
+  // 配置变更（素材/配音/数量）→ 重新渲染；标题文字变化不触发（CSS浮层实时叠加）
+  const prevFingerprintRef = useRef(currentFingerprint)
+  useEffect(() => {
+    if (!enabled || !currentFingerprint) return
+    const prev = prevFingerprintRef.current
+    prevFingerprintRef.current = currentFingerprint
+    if (!prev || prev === currentFingerprint) return
+    trigger()
+  }, [enabled, currentFingerprint, trigger])
+
+  /* ── 派生状态 ── */
+  const progress =
+    variants.length > 0
+      ? Math.round(variants.reduce((sum, v) => sum + v.progress, 0) / variants.length)
+      : 0
+  const failedCount = variants.filter((v) => v.status === "failed").length
+  const readyCount = variants.filter((v) => v.status === "ready").length
+
+  useEffect(() => {
+    if (status !== "loading" || variants.length === 0) return
+    if (readyCount === variants.length) {
+      setStatus("ready")
+    } else if (readyCount + failedCount === variants.length && failedCount > 0) {
+      setStatus(failedCount === variants.length ? "failed" : "partial_failed")
+    }
+  }, [variants, status, readyCount, failedCount])
+
+  return { variants, status, progress, failedCount, trigger }
+}
+
+export default useBatchPreview

apps/web/src/pages/generate/hooks/useGenerateVideo.ts

@@ -2,13 +2,13 @@
  * 视频生成 Hook
  * 封装视频生成的核心逻辑、状态管理、轮询等
  */
-import { useState, useCallback, useEffect } from "react"
+import { useState, useCallback } from "react"
 import { message } from "antd"
 import { type GeneratedVideo, getEditPlanClips, createClipsFromAssets } from "@/api/template-editor"
 import { createGenerationTask } from "@/api/tasks/tasks"
 import type { UseGenerateVideoProps } from "./generate-video/types"
 import { getGenerationPhase } from "./generate-video/phase"
-import { useGenerationPolling, type BatchTaskState } from "./generate-video/useGenerationPolling"
+import { useGenerationPolling } from "./generate-video/useGenerationPolling"
 import { validateGenerateInputs } from "./generate-video/buildPayload"
 import { calculateResolution } from "../utils/calculateResolution"
 import { extractBackendError, translateError } from "./generate-video/errorUtils"
@@ -22,32 +22,6 @@ export function useGenerateVideo(props: UseGenerateVideoProps) {
   const [generated, setGenerated] = useState(false)
   const [generateError, setGenerateError] = useState<string | null>(null)
   const [generatedVideos, setGeneratedVideos] = useState<GeneratedVideo[]>([])
-  /** 批量模式：每个正式生成任务的独立状态（第5步逐卡片展示） */
-  const [batchTasks, setBatchTasks] = useState<BatchTaskState[]>([])
-
-  const handleBatchTaskUpdate = useCallback((taskId: string, patch: Partial<BatchTaskState>) => {
-    setBatchTasks((prev) => {
-      const list = prev || []
-      const idx = list.findIndex((t) => t.taskId === taskId)
-      if (idx === -1) {
-        return [
-          ...list,
-          {
-            taskId,
-            variantIndex: patch.variantIndex ?? 0,
-            status: "running",
-            progress: 0,
-            error: null,
-            videos: [],
-            ...patch,
-          },
-        ]
-      }
-      const next = [...list]
-      next[idx] = { ...next[idx], ...patch }
-      return next
-    })
-  }, [])
 
   const handleProgress = useCallback((p: number) => setProgress(p), [])
   const handleComplete = useCallback(
@@ -55,19 +29,6 @@ export function useGenerateVideo(props: UseGenerateVideoProps) {
       setGenerating(false)
       setGenerated(true)
       setGeneratedVideos(videos as GeneratedVideo[])
-      // 批量：成功任务的 videos 已通过 onBatchTaskUpdate 写入，这里同步兜底
-      setBatchTasks((prev) =>
-        (prev || []).map((t) =>
-          t.status === "completed" && t.videos.length === 0
-            ? {
-                ...t,
-                videos: (videos as GeneratedVideo[]).filter(
-                  (v) => v.generation_task_id === t.taskId,
-                ),
-              }
-            : t,
-        ),
-      )
       onGenerationSuccess?.()
     },
     [onGenerationSuccess],
@@ -77,30 +38,10 @@ export function useGenerateVideo(props: UseGenerateVideoProps) {
     setGenerateError(errorMsg)
   }, [])
 
-  /* 批量：任务状态变化时聚合已完成成片（含失败重试成功后补入），
-     按变体索引排序，供步骤6封面按勾选顺序逐个取视频 */
-  useEffect(() => {
-    if (batchTasks.length === 0) return
-    const byVariant = new Map<number, GeneratedVideo>()
-    batchTasks.forEach((t) => {
-      if (t.status === "completed" && t.videos && t.videos.length > 0) {
-        byVariant.set(t.variantIndex, t.videos[0] as GeneratedVideo)
-      }
-    })
-    const ordered = [...byVariant.entries()].sort((a, b) => a[0] - b[0]).map(([, v]) => v)
-    setGeneratedVideos((prev) => {
-      if (prev.length === ordered.length && prev.every((v, i) => v.id === ordered[i].id)) {
-        return prev
-      }
-      return ordered
-    })
-  }, [batchTasks])
-
-  const { startPolling, startPollingBatch, retryTask, clearTimer } = useGenerationPolling({
+  const { startPolling, startPollingBatch, clearTimer } = useGenerationPolling({
     onProgress: handleProgress,
     onComplete: handleComplete,
     onFailed: handleFailed,
-    onBatchTaskUpdate: handleBatchTaskUpdate,
   })
 
   /* ── 生成视频 ──
@@ -116,7 +57,6 @@ export function useGenerateVideo(props: UseGenerateVideoProps) {
     setProgress(0)
     setGenerated(false)
     setGenerateError(null)
-    setBatchTasks([])
     clearTimer()
 
     try {
@@ -230,8 +170,7 @@ export function useGenerateVideo(props: UseGenerateVideoProps) {
           throw new Error("创建任务成功但未返回任务 ID，请稍后在任务列表查看")
         }
         if (taskIds.length > 1) {
-          // 批量：任务按创建顺序与勾选变体一一对应（后端按 count 顺序创建）
-          startPollingBatch(taskIds.map((taskId, i) => ({ taskId, variantIndex: indexes[i] ?? i })))
+          startPollingBatch(taskIds)
         } else {
           startPolling(taskIds[0])
         }
@@ -257,14 +196,6 @@ export function useGenerateVideo(props: UseGenerateVideoProps) {
     generate()
   }, [generate])
 
-  /** 第5步：单独重试某个失败任务 */
-  const retryBatchTask = useCallback(
-    (taskId: string) => {
-      retryTask(taskId)
-    },
-    [retryTask],
-  )
-
   const dismissError = useCallback(() => {
     setGenerateError(null)
   }, [])
@@ -309,8 +240,6 @@ export function useGenerateVideo(props: UseGenerateVideoProps) {
     generatedVideos,
     generate,
     retry,
-    retryBatchTask,
-    batchTasks,
     dismissError,
     download,
     share,

apps/web/src/pages/generate/hooks/useStepNavigation.ts

@@ -1,11 +1,6 @@
 /**
- * GeneratePage 步骤导航（Issue #1677 修正：固定 6 步，单视频与批量一致）
- * 步骤：模板(1) → 素材(2) → 配音(3) → 标题(4) → 确认生成(5) → 封面(6)
- *
- * - 步骤4底部按钮是「确认生成视频/确认生成 N 个视频」（由 GenerateStepActions 调
- *   onConfirmGenerate），创建成功后跳转步骤5；本 hook 的 goNext 只负责 1→2→3→4
- *   和 5→6 的「下一步」。
- * - 步骤5（确认生成进度页）：渲染全部完成（generated）后「下一步」解锁进封面。
+ * GeneratePage 步骤导航（Issue #1677 改造后 5 步）
+ * 步骤：模板(1) → 素材(2) → 配音(3) → 标题+预览+确认生成(4) → 封面(5)
  */
 import { message } from "antd"
 import type { TitleSettings } from "../types"
@@ -18,8 +13,14 @@ export interface UseStepNavigationOptions {
   selectedMaterials: string[]
   smartSelectedIds: string[]
   titleSettings: TitleSettings
-  /** 是否已完成视频生成（步骤5全部渲染完成后才能进入封面） */
+  /** 预览是否已就绪（单视频=前端预览素材已加载；批量=服务器预览全部完成） */
+  previewReady: boolean
+  /** 是否已完成视频生成（步骤4确认生成后才能进入封面） */
   generated: boolean
+  /** 批量模式下每个变体的标题 */
+  previewTitles: string[]
+  /** 批量模式勾选的变体数 */
+  selectedCount: number
   /** Step1 点下一步时弹出数量选择弹窗 */
   onOpenCountModal: () => void
 }
@@ -37,7 +38,10 @@ export const useStepNavigation = (options: UseStepNavigationOptions): UseStepNav
     materialMode,
     selectedMaterials,
     smartSelectedIds,
+    previewReady,
     generated,
+    previewTitles,
+    selectedCount,
     onOpenCountModal,
   } = options
 
@@ -59,14 +63,27 @@ export const useStepNavigation = (options: UseStepNavigationOptions): UseStepNav
       message.warning("请先进行智能匹配并选择素材")
       return
     }
-    // 步骤5（确认生成）：全部渲染完成后才能下一步进封面
-    if (currentStep === 5) {
+    // Step4（标题+预览+确认生成）：标题必填 + 预览必须已加载
+    if (currentStep === 4) {
+      const allTitlesFilled = previewTitles.every((t) => t && t.trim())
+      if (!allTitlesFilled) {
+        message.warning("请为每个视频输入标题")
+        return
+      }
+      if (selectedCount === 0) {
+        message.warning("请至少勾选一个视频")
+        return
+      }
+      if (!previewReady) {
+        message.warning("预览视频正在加载，请稍候")
+        return
+      }
       if (!generated) {
-        message.warning("视频还在渲染中，请等待生成完成")
+        message.warning("请先点击「确认生成视频」完成渲染")
         return
       }
     }
-    if (currentStep < 6) {
+    if (currentStep < 5) {
       setCurrentStep((s) => s + 1)
     }
   }

apps/web/src/test/pages/generate/smoke.test.tsx

@@ -20,8 +20,7 @@ import "@/pages/generate/components/Step2MaterialSelect"
 import "@/pages/generate/components/Step4TitleSettings"
 import "@/pages/generate/components/Step5VoiceSelect"
 import "@/pages/generate/components/Step3VoiceWithMode"
-import "@/pages/generate/components/CanvasPreviewGrid"
-import "@/pages/generate/components/BatchGenerationGrid"
+import "@/pages/generate/components/ServerPreviewGrid"
 import "@/pages/generate/components/PreviewCountModal"
 import "@/pages/generate/components/PreviewVideoPanel"
 import "@/pages/generate/components/GenerateResultPanel"
@@ -49,6 +48,7 @@ describe("GeneratePage module smoke test", () => {
   })
 })
 import "@/pages/generate/hooks/useGenerateVideo"
+import "@/pages/generate/hooks/useBatchPreview"
 import "@/pages/generate/hooks/useBatchCovers"
 import "@/pages/generate/hooks/usePreviewAssets"
 import "@/pages/generate/hooks/useSegmentScheduler"

apps/worker/video_processing/dedup.py

@@ -31,65 +31,25 @@ SCENE_CHANGE_THRESHOLD = 30  # 灰度差异阈值
 MIN_KEYFRAME_INTERVAL_SEC = 1.0  # 最小关键帧间隔（秒）
 MAX_KEYFRAMES = 30  # 最大关键帧数
 MIN_KEYFRAMES = 5  # 最小关键帧数
-FINGERPRINT_SAMPLE_INTERVAL_SEC = 1.0  # 指纹采样间隔（秒）：密集均匀采样，保证两视频时序可对齐
-FINGERPRINT_MAX_SAMPLES = 30  # 长视频采样数上限（超过后采样间隔自动放宽）
 LONG_VIDEO_SEGMENT_SEC = 30  # 长视频每段秒数
 LONG_VIDEO_DURATION_THRESHOLD_SEC = 180  # 3 分钟阈值
 MIN_FRAMES_PER_SEGMENT = 2  # 长视频每段最少帧数
 
-# ── 滑动窗口匹配常量（Issue #1702 二次校准） ─────────────────────
-# 阈值经 staging 真实数据两轮回归校准（worker 容器内离线实验）：
-#   第一轮（2026-09-05）：同源对 <=12 命中 4/11，异源最小距离 24 → 定 12；
-#   第二轮（2026-09-05，证据视频 B->A 仍漏检）：扩大样本到该用户全部
-#   15 个真实成片（13 个异源候选）实测：
-#   - 同源成片对（A 20s / B、C 各 11.75s，1s 密集采样）：
-#       B->A 中位数距离 14，<=16 命中 8/11=0.73；C->A 8/11=0.73
-#   - 异源成片对（13 个真实视频）：每帧全局最近邻最小距离 18，
-#       <=16 命中帧数全部为 0（最近邻 18 仅个别帧，中位数 22~28）
-# 12 漏掉同源降重对（降重滤镜/字幕/画面扰动把距离从 ~8 推到 14~16）；
-# 16 对同源命中 0.73+ 且与异源分布（最近邻 >=18）仍有 >=2bit 安全裕度，
-# 异源 <=16 命中 0 帧，无误报空间。
-PHASH_THRESHOLD = 16
-SEGMENT_MATCH_THRESHOLD = PHASH_THRESHOLD  # 片段匹配阈值与帧匹配统一（#1702：阈值常量统一来源）
-MIN_CONSECUTIVE_MATCHES = 5  # 连续匹配默认门槛；短视频自适应 min(5, max(2, 分片数//2))
+# ── 滑动窗口匹配常量 ────────────────────────────────────────────
+SEGMENT_MATCH_THRESHOLD = 8  # 帧匹配汉明距离阈值
+MIN_CONSECUTIVE_MATCHES = 5  # 最少连续匹配帧数
 MAX_GAP = 2  # 允许的最大间隙帧数
-NEIGHBOR_WINDOW = 1  # 分片时序对齐：允许 ±1 邻接偏移（1s 密集采样下即 ±1s，缓解切点不一致）
 
 # ── 融合判定常量 ────────────────────────────────────────────────
 PHASH_WEIGHT = 0.7  # pHash 权重
 HISTOGRAM_WEIGHT = 0.3  # 直方图权重
-MATCH_RATIO_THRESHOLD = 0.7  # 全片重复（is_duplicate）至少 70% 帧匹配
-PARTIAL_COVERAGE_THRESHOLD = 0.5  # 局部复用覆盖率 >=50% 也判全片重复
+MATCH_RATIO_THRESHOLD = 0.7  # 至少 70% 帧匹配
 DUPLICATE_THRESHOLD = 0.70  # 融合后相似度阈值
 
-# ── 降重裁剪规避常量（Issue #1702） ─────────────────────────────
-# 成片强制 2-5% random_edge_crop 降重只服务外部平台；自查重指纹取中心 90%
-# 区域，使两次不同裁剪的同源画面 pHash 距离回到同分布。
-FINGERPRINT_CENTER_CROP_RATIO = 0.90
-
 
 # ── 感知哈希 & 颜色直方图工具函数 ────────────────────────────────
 
 
-def center_crop_frame(image: np.ndarray, ratio: float = FINGERPRINT_CENTER_CROP_RATIO) -> np.ndarray:
-    """取画面中心 ratio 比例区域（裁除四边边缘）。
-
-    查重指纹用：random_edge_crop 降重（2-5% 四边随机裁剪）会让同源画面 pHash
-    位翻转 12-16，污染自查重（Issue #1702）。算 pHash/颜色直方图前先居中裁除
-    边缘 10%，两次不同裁剪的同源画面中心区域基本重合，指纹不再被降重污染。
-    降重只服务外部平台，不影响内部查重。
-    """
-    if image is None or image.size == 0:
-        return image
-    h, w = image.shape[:2]
-    ch, cw = int(h * ratio), int(w * ratio)
-    if ch <= 0 or cw <= 0 or (ch >= h and cw >= w):
-        return image
-    y0 = (h - ch) // 2
-    x0 = (w - cw) // 2
-    return image[y0 : y0 + ch, x0 : x0 + cw]
-
-
 def compute_phash(image: np.ndarray, hash_size: int = 8) -> str:
     """计算图像的感知哈希（pHash），基于 DCT（离散余弦变换）。
 
@@ -141,17 +101,11 @@ def hamming_distance(hash1: str, hash2: str) -> int:
 
 
 def compute_color_histogram(image: np.ndarray, bins: int = 32) -> list[float]:
-    """Compute BGR color histogram for an image.
-
-    Issue #1702: 每个通道独立做 NORM_L1 归一化（通道内 Σ=1，是概率分布），
-    三通道拼接存储。Bhattacharyya 系数对拼接向量直接 Σ√(a*b) 会得到
-    3 通道之和（范围 [0,3]，实测 ~14.9 是旧 L2 归一化的错误结果），
-    消费方 _bhattacharyya_coefficient 按通道数平均归一到 [0,1]。
-    """
+    """Compute color histogram for an image."""
     hist = []
     for i in range(3):
         h = cv2.calcHist([image], [i], None, [bins], [0, 256])
-        h = cv2.normalize(h, h, norm_type=cv2.NORM_L1).flatten()
+        h = cv2.normalize(h, h).flatten()
         hist.extend(h)
     return hist
 
@@ -256,30 +210,6 @@ def detect_keyframe_timestamps(
     return keyframe_times
 
 
-def sample_fingerprint_timestamps(
-    duration: float,
-    *,
-    interval_sec: float = FINGERPRINT_SAMPLE_INTERVAL_SEC,
-    max_samples: int = FINGERPRINT_MAX_SAMPLES,
-) -> list[float]:
-    """指纹采样时间戳：固定间隔密集均匀采样（Issue #1702）。
-
-    动态场景检测抽帧（#1659）在两个同源视频上会各自取到不同时刻，切点/取帧
-    错位让对齐帧的 pHash 距离都很大（实测同源对最小距离 12 且配对时序错乱）。
-    改为固定 1s 间隔均匀采样后，复用片段的帧时刻天然对齐，配合 ±1 邻接窗口
-    即可检出同源/局部复用。长视频（>max_samples*interval）自动放宽间隔到
-    duration/max_samples，保证分片数有上限。
-    """
-    if duration <= 0:
-        return []
-    step = interval_sec
-    n_uniform = int(duration / step)
-    if n_uniform > max_samples:
-        step = duration / max_samples
-    count = max(1, int(duration / step))
-    return [step * (i + 0.5) for i in range(count)]
-
-
 # ── 数据类 ──────────────────────────────────────────────────────
 
 
@@ -367,33 +297,23 @@ def find_duplicate_segments(
     target_chunks: list,
     *,
     match_threshold: int = SEGMENT_MATCH_THRESHOLD,
-    min_consecutive: Optional[int] = None,
+    min_consecutive: int = MIN_CONSECUTIVE_MATCHES,
     max_gap: int = MAX_GAP,
-    neighbor_window: int = NEIGHBOR_WINDOW,
 ) -> list[DuplicateSegment]:
-    """滑动窗口时序匹配：找出两组分片之间的重复片段（Issue #1702 重构）。
+    """滑动窗口时序匹配：找出两组分片之间的重复片段。
 
     算法：
-    1. 构建 query×target 全量汉明距离矩阵；每个 query chunk 保留所有
-       距离 <= match_threshold 的候选 target 分片（与帧匹配判定同一阈值）。
-    2. 时序一致贪心对齐：沿 query 时序推进，run 内优先选择与上一匹配帧
-       目标序号连贯（|delta| <= neighbor_window+1，允许 ±1 邻接/时序偏移
-       对齐——1s 密集采样下相邻帧 pHash 接近，最近邻在目标相邻帧间
-       正/反向跳变均属正常，缓解场景切割切点、取帧错位、局部倒退）的
-       候选；同距时偏好小索引（最早对齐位置）。
-    3. 连贯匹配中允许 <= max_gap 帧间隙桥接；断裂后另起新 run——天然
-       支持局部片段复用（复用片段可出现在任意时序位置，各成独立片段）。
-    4. 连续匹配帧数 >= min_consecutive 的 run 报为重复片段。短视频自适应：
-       min_consecutive = min(5, max(2, len(query_chunks)//2))；n=1 时
-       不形成片段，由调用方匹配帧回退兜底。
+    1. 对每个 query chunk，找到 target 中汉明距离最小的 chunk
+    2. 距离 <= match_threshold 视为匹配
+    3. 找连续匹配的 run（允许 max_gap 帧间隙）
+    4. 连续匹配数 >= min_consecutive 的 run 报告为重复片段
 
     Args:
         query_chunks: 查询视频的分片列表（FingerprintChunk 或 dict）
         target_chunks: 目标视频的分片列表
-        match_threshold: 汉明距离匹配阈值（统一常量 PHASH_THRESHOLD）
-        min_consecutive: 最少连续匹配帧数；None 时按短视频自适应
+        match_threshold: 汉明距离匹配阈值
+        min_consecutive: 最少连续匹配帧数
         max_gap: 允许的最大间隙帧数
-        neighbor_window: 时序对齐允许的目标分片序号邻接窗口（正/反向均允许）
 
     Returns:
         DuplicateSegment 列表
@@ -401,94 +321,95 @@ def find_duplicate_segments(
     if not query_chunks or not target_chunks:
         return []
 
-    def _get(chunk, key):
+    def _get_phash(chunk) -> str:
         if isinstance(chunk, dict):
-            return chunk[key]
-        return getattr(chunk, key)
+            return chunk["phash_binary"]
+        return chunk.phash_binary
 
-    n, m = len(query_chunks), len(target_chunks)
-    q_ph = [_get(c, "phash_binary") for c in query_chunks]
-    t_ph = [_get(c, "phash_binary") for c in target_chunks]
+    def _get_start(chunk) -> int:
+        if isinstance(chunk, dict):
+            return chunk["start_time_ms"]
+        return chunk.start_time_ms
 
-    # Step 1: 全量距离矩阵。每个 query chunk 保留所有 <= 阈值的候选 target，
-    # 按距离升序；同距时小索引优先（取最早的对齐位置，贪心连贯推进时最保守，
-    # 不会越过复用片段末端；重复 hash 的连续帧由 Step 2 的连贯性窗口约束）。
-    candidates: list[list[tuple[int, int]]] = []  # 每 query 帧: [(target_idx, dist), ...]
-    for i in range(n):
-        dists = [hamming_distance(q_ph[i], t_ph[j]) for j in range(m)]
-        cand = [(j, d) for j, d in enumerate(dists) if d <= match_threshold]
-        cand.sort(key=lambda x: (x[1], x[0]))
-        candidates.append(cand)
+    def _get_end(chunk) -> int:
+        if isinstance(chunk, dict):
+            return chunk["end_time_ms"]
+        return chunk.end_time_ms
 
-    # 短视频自适应连续匹配门槛（Issue #1702 工单公式）：
-    # MIN_CONSECUTIVE_MATCHES = min(5, max(2, 分片数//2))。
-    # n=1 时门槛为 2 不形成片段，由 _evaluate_candidate 的匹配帧回退
-    # （temporal_coverage 按匹配帧占比估计）兜底检出，不回归。
-    if min_consecutive is None:
-        min_consecutive = min(MIN_CONSECUTIVE_MATCHES, max(2, n // 2))
+    # Step 1: 逐帧匹配
+    frame_matches: list[tuple[bool, int, int]] = []  # (is_match, min_dist, best_target_idx)
+    for qc in query_chunks:
+        qc_phash = _get_phash(qc)
+        best_dist = 64
+        best_idx = 0
+        for j, tc in enumerate(target_chunks):
+            d = hamming_distance(qc_phash, _get_phash(tc))
+            if d < best_dist:
+                best_dist = d
+                best_idx = j
+        frame_matches.append((best_dist <= match_threshold, best_dist, best_idx))
 
-    # Step 2: 时序一致贪心对齐。
-    # run 内偏好与上一匹配帧目标序号连贯（|delta| <= neighbor_window+1，
-    # 支持 ±1 邻接窗口/时序偏移对齐，正反向抖动均允许）的候选；
-    # 无连贯候选时关闭旧 run。
-    # 这天然支持局部片段复用：同一 query 视频中多个复用片段各自形成独立 run。
-    frame_matches: list[tuple[bool, int, int]] = []
-    runs: list[tuple[int, int]] = []
-    run_start: Optional[int] = None
-    run_last_t: Optional[int] = None
+    # Step 2: 找连续匹配的 runs
+    runs: list[tuple[int, int]] = []  # list of (start_idx, end_idx)
+    run_start = None
     gap_count = 0
 
-    def _matching_count(a: int, b: int) -> int:
-        return sum(1 for k in range(a, b + 1) if frame_matches[k][0])
-
-    def _close_run(a: int, b: int) -> None:
-        if b >= a and _matching_count(a, b) >= min_consecutive:
-            runs.append((a, b))
-
-    for i in range(n):
-        cand = candidates[i]
-        if run_last_t is None:
-            chosen = cand[0] if cand else None
-        else:
-            chosen = next(
-                (c for c in cand if abs(c[0] - run_last_t) <= neighbor_window + 1),
-                None,
-            )
-
-        if chosen is not None:
-            tidx, dist = chosen
-            frame_matches.append((True, dist, tidx))
+    for i, (is_match, _dist, _idx) in enumerate(frame_matches):
+        if is_match:
             if run_start is None:
                 run_start = i
-            gap_count = 0
-            run_last_t = tidx
+            gap_count = 0  # 重置间隙
         else:
-            frame_matches.append((False, match_threshold + 1, -1))
             if run_start is not None:
                 gap_count += 1
                 if gap_count > max_gap:
-                    # 非匹配帧从 i-gap_count+1 开始，run 结束于其前一帧
-                    _close_run(run_start, i - gap_count)
-                    run_start, run_last_t, gap_count = None, None, 0
+                    # 中断当前 run
+                    run_end = i - gap_count  # 最后一个匹配帧的索引
+                    # 计算 run 内的实际匹配帧数（总跨度 - 间隙数）
+                    total_gaps = sum(1 for k in range(run_start, run_end + 1) if not frame_matches[k][0])
+                    matching_count = (run_end - run_start + 1) - total_gaps
+                    if matching_count >= min_consecutive:
+                        runs.append((run_start, run_end))
+                    run_start = None
+                    gap_count = 0
 
+    # 处理末尾 run
     if run_start is not None:
-        last_idx = n - 1
+        last_idx = len(frame_matches) - 1
+        # 回退找到最后一个匹配帧的位置（跳过尾部非匹配帧）
         while last_idx >= run_start and not frame_matches[last_idx][0]:
             last_idx -= 1
-        _close_run(run_start, last_idx)
+        if last_idx >= run_start:
+            # 计算 run 内的总间隙数
+            total_gaps = sum(1 for k in range(run_start, last_idx + 1) if not frame_matches[k][0])
+            matching_count = (last_idx - run_start + 1) - total_gaps
+            if matching_count >= min_consecutive:
+                runs.append((run_start, last_idx))
 
     # Step 3: 构建 DuplicateSegment
     segments: list[DuplicateSegment] = []
     for start, end in runs:
+        query_start = _get_start(query_chunks[start])
+        query_end = _get_end(query_chunks[end])
+
+        # 取目标范围（按最佳匹配的目标 chunk 时间范围）
         target_indices = [frame_matches[k][2] for k in range(start, end + 1) if frame_matches[k][0]]
-        t_min, t_max = min(target_indices), max(target_indices)
-        avg_dist = sum(frame_matches[k][1] for k in range(start, end + 1) if frame_matches[k][0]) / len(target_indices)
+        if target_indices:
+            t_min = min(target_indices)
+            t_max = max(target_indices)
+            target_start = _get_start(target_chunks[t_min])
+            target_end = _get_end(target_chunks[t_max])
+        else:
+            target_start = _get_start(target_chunks[0])
+            target_end = _get_end(target_chunks[-1])
+
+        avg_dist = sum(frame_matches[k][1] for k in range(start, end + 1)) / (end - start + 1)
         segments.append(
             DuplicateSegment(
-                query_start_ms=_get(query_chunks[start], "start_time_ms"),
-                query_end_ms=_get(query_chunks[end], "end_time_ms"),
-                target_start_ms=_get(target_chunks[t_min], "start_time_ms"),
-                target_end_ms=_get(target_chunks[t_max], "end_time_ms"),
+                query_start_ms=query_start,
+                query_end_ms=query_end,
+                target_start_ms=target_start,
+                target_end_ms=target_end,
                 avg_distance=avg_dist,
             )
         )
@@ -502,9 +423,7 @@ def find_duplicate_segments(
 class VideoDeduplicator:
     """Video deduplication using multiple fingerprint methods."""
 
-    # Issue #1702: 阈值统一来源为模块常量 PHASH_THRESHOLD（#1658 曾收紧到 8，
-    # 后经 staging 真实同源/异源指纹分布重新校准，见 test_phash_threshold_calibration_1702）。
-    PHASH_THRESHOLD = PHASH_THRESHOLD
+    PHASH_THRESHOLD = 8  # Issue #1658: pHash 汉明距离阈值由 10 收紧到 8，降低不同视频误判率
     HISTOGRAM_THRESHOLD = 0.85
 
     @staticmethod
@@ -528,36 +447,27 @@ class VideoDeduplicator:
         # 单帧不视为坏指纹（短视频或抽帧不足）
         if len(phashes) == 1:
             return False
-        # Issue #1702: 旧逻辑"所有 phash 完全相同即判黑屏"会误杀短视频——
-        # 11s 视频只有几个不同镜头时，相邻 1s 采样帧可能 phash 完全一致（内容
-        # 连续但非黑屏）。黑屏的特征是「大量帧全部无内容」，要求至少 8 帧
-        # 且相同帧占比 >=80% 才判坏；短视频（<8 帧）只有真正单值时交给
-        # _bhattacharyya/融合分兜底，不因"帧都一样"直接跳过。
-        if len(phashes) < 8:
-            return False
+        # 多帧但所有 phash 完全相同 → 黑屏/纯色视频
         unique = set(phashes)
-        same_ratio = sum(1 for x in phashes if x == phashes[0]) / len(phashes)
-        if len(unique) == 1 and same_ratio >= 0.8:
+        if len(unique) == 1:
             return True
-        # 多帧但所有唯一 phash 之间的汉明距离都极小（<3）且占比 >=80% → 近似黑屏
+        # 多帧但所有 phash 之间的汉明距离都极小（<3）→ 近似黑屏
         phash_list = list(unique)
-        if len(phash_list) >= 2 and same_ratio >= 0.8:
-            all_distances = [
-                hamming_distance(phash_list[i], phash_list[j])
-                for i in range(len(phash_list))
-                for j in range(i + 1, len(phash_list))
-            ]
+        if len(phash_list) >= 2:
+            all_distances = []
+            for i in range(len(phash_list)):
+                for j in range(i + 1, len(phash_list)):
+                    all_distances.append(hamming_distance(phash_list[i], phash_list[j]))
             if all_distances and max(all_distances) < 3:
                 return True
         return False
 
     def compute_fingerprint(self, video_path: str) -> VideoFingerprint:
-        """Compute video fingerprint using dense uniform sampling.
+        """Compute video fingerprint using dynamic keyframe detection.
 
-        Issue #1702: 使用 sample_fingerprint_timestamps() 固定 1s 间隔密集均匀
-        采样（替代动态场景检测抽帧），保证两个同源视频复用片段的帧时刻天然
-        对齐；每帧取中心 90% 区域（center_crop_frame）计算 pHash + color_histogram，
-        绕开 random_edge_crop 降重裁剪污染；MD5 仍基于原始帧。
+        使用 detect_keyframe_timestamps() 检测内容感知关键帧，
+        在每个关键帧处取帧计算 pHash + color_histogram。
+        同时保留 MD5 计算和分片数据结构。
         """
         cap = cv2.VideoCapture(video_path)
         if not cap.isOpened():
@@ -571,8 +481,8 @@ class VideoDeduplicator:
 
         cap.release()
 
-        # 1. 固定间隔密集采样（Issue #1702：替代动态场景检测，保证跨视频时序对齐）
-        keyframe_times = sample_fingerprint_timestamps(duration)
+        # 1. 检测关键帧时间戳
+        keyframe_times = detect_keyframe_timestamps(video_path)
 
         if not keyframe_times:
             return VideoFingerprint(
@@ -596,15 +506,12 @@ class VideoDeduplicator:
             if not ret:
                 continue
 
-            # MD5 计算（基于原始帧，指纹文件级去重不受裁剪影响）
+            # MD5 计算
             _, buffer = cv2.imencode(".jpg", frame)
             md5_hash.update(buffer)
 
-            # Issue #1702: pHash / 颜色直方图基于中心 90% 区域，绕开 random_edge_crop
-            # 降重裁剪对指纹的污染（降重只服务外部平台，不污染自查重）。
-            fp_frame = center_crop_frame(frame)
-            phash = compute_phash(fp_frame)
-            hist = compute_color_histogram(fp_frame)
+            phash = compute_phash(frame)
+            hist = compute_color_histogram(frame)
 
             # 计算分片时间范围（从前一个关键帧到下一个关键帧的中点）
             prev_boundary = keyframe_times[i - 1] * 1000 if i > 0 else 0
@@ -657,22 +564,12 @@ class VideoDeduplicator:
 
     @staticmethod
     def _bhattacharyya_coefficient(hist_a: list[float], hist_b: list[float]) -> float:
-        """Bhattacharyya 系数（概率分布版，范围 [0,1]，1=完全相同）。
-
-        Issue #1702: compute_color_histogram 输出 3 通道拼接、每通道独立 NORM_L1
-        （单通道 Σ=1，三通道拼接向量 Σ=3）。旧实现直接 Σ√(a*b) 对三通道拼接向量
-        算出 ~3（旧 L2 归一化更是算出 ~14.9），不是合法的概率系数。
-        这里按两个直方图各自的总量归一：BC = Σ√(a*b) / √(Σa·Σb)。
-        - 单通道概率分布（Σa=Σb=1）：分母 1，与旧测试/教科书定义一致；
-        - 三通道拼接（Σa=Σb=3）：分母 3，结果在 [0,1]。
-        """
+        """Bhattacharyya 系数：Σ √(a[i] * b[i])，范围 [0, 1]，1=完全相同。"""
         min_len = min(len(hist_a), len(hist_b))
-        a = [max(0.0, float(x)) for x in hist_a[:min_len]]
-        b = [max(0.0, float(x)) for x in hist_b[:min_len]]
-        # max(0.0, ...) 防御上游异常负值导致 sqrt domain error
-        coeff = sum(math.sqrt(ai * bi) for ai, bi in zip(a, b, strict=False))
-        norm = math.sqrt(sum(a) * sum(b))
-        return float(coeff / norm) if norm > 0 else 0.0
+        a = hist_a[:min_len]
+        b = hist_b[:min_len]
+        # 纯标准库计算（不依赖 numpy）；max(0.0, ...) 防御上游异常负值导致 sqrt domain error
+        return float(sum(math.sqrt(max(0.0, ai * bi)) for ai, bi in zip(a, b, strict=False)))
 
     @staticmethod
     def _compute_histogram_similarity(
@@ -714,72 +611,6 @@ class VideoDeduplicator:
         hist_similarity = VideoDeduplicator._compute_histogram_similarity(hist_a, hist_b) if hist_b else 0.5
         return PHASH_WEIGHT * phash_similarity + HISTOGRAM_WEIGHT * hist_similarity
 
-    @staticmethod
-    def _evaluate_candidate(
-        fingerprint: VideoFingerprint,
-        existing_phashes: list[str],
-        existing_histograms: list,
-        existing_chunk_objects: list,
-        *,
-        query_duration_sec: float,
-    ) -> dict:
-        """评估新视频指纹与单个候选视频的相似度（Issue #1702 共享逻辑）。
-
-        指标：
-        - min_distances / frame_match_rate：每个新分片到候选视频全局最近邻的汉明距离，
-          分母取两视频分片数的较小值（支持局部片段复用：短视频复用长视频片段时不被长视频分母稀释）。
-        - temporal_coverage：时序一致连续匹配片段总时长 / 新视频时长（局部复用主指标）。
-        - fusion：pHash 中位数距离 + 颜色直方图的加权融合分。
-
-        Returns:
-            {frame_match_rate, temporal_coverage, segments, median_distance,
-             fusion, matching_frames, min_distances}
-        """
-        query_phashes = fingerprint.keyframe_phashes or []
-        if not query_phashes or not existing_phashes:
-            return {
-                "frame_match_rate": 0.0,
-                "temporal_coverage": 0.0,
-                "segments": [],
-                "median_distance": 64,
-                "fusion": 0.0,
-                "matching_frames": 0,
-                "min_distances": [],
-            }
-
-        min_distances = [min(hamming_distance(ph, ep) for ep in existing_phashes) for ph in query_phashes]
-        matching_frames = sum(1 for d in min_distances if d <= PHASH_THRESHOLD)
-        # 分母取 min(两视频分片数)：局部复用时（如 B 的 5 片复用 A 9 片中的若干片）
-        # 命中帧占比不因候选视频更长而被稀释。
-        frame_match_rate = matching_frames / min(len(query_phashes), len(existing_phashes))
-
-        segments = find_duplicate_segments(fingerprint.chunks, existing_chunk_objects)
-        duration_ms = query_duration_sec * 1000 if query_duration_sec else 0
-        if duration_ms > 0 and segments:
-            covered_ms = sum(s.query_end_ms - s.query_start_ms for s in segments)
-            temporal_coverage = min(covered_ms / duration_ms, 1.0)
-        elif matching_frames > 0:
-            # 无连续片段（时序连贯性不足）时，按匹配帧占比估计覆盖：
-            # 密集 1s 采样下每个分片≈1s 等权时间片，匹配帧数≈命中秒数。
-            temporal_coverage = min(frame_match_rate, 1.0)
-        else:
-            temporal_coverage = 0.0
-
-        median_distance = statistics.median(min_distances) if min_distances else 64
-        fusion = VideoDeduplicator._compute_fusion_score(
-            median_distance, fingerprint.color_histograms, existing_histograms
-        )
-
-        return {
-            "frame_match_rate": frame_match_rate,
-            "temporal_coverage": temporal_coverage,
-            "segments": segments,
-            "median_distance": median_distance,
-            "fusion": fusion,
-            "matching_frames": matching_frames,
-            "min_distances": min_distances,
-        }
-
     def check_duplicate(
         self,
         fingerprint: VideoFingerprint,
@@ -789,7 +620,6 @@ class VideoDeduplicator:
         scope: str = "project",
         user_id: str = "",
         duration_sec: float = 0,
-        exclude_video_id: str | None = None,
     ) -> Optional[dict]:
         """检查视频是否与已有视频重复。
 
@@ -806,9 +636,6 @@ class VideoDeduplicator:
             scope: "project" 项目内查重（默认），"user" 跨项目全局查重
             user_id: 用户 ID（scope="user" 时使用）
             duration_sec: 视频时长（秒），用于时长预过滤 ±15%
-            exclude_video_id: 排除的视频 ID（查重自身时用）。recompute-dedup
-                重算时视频记录已存在，不排除会自匹配（距离 0 分最高）导致
-                duplicate_of 指向自己（Issue #1702 连带修复）。
 
         Returns:
             重复信息字典（含 duplicate, duplicate_of, reason, similarity, duplicate_segments），
@@ -816,22 +643,13 @@ class VideoDeduplicator:
         """
         video_repo = SQLAlchemyGeneratedVideoRepository(session)
         if scope == "user" and user_id:
-            # Issue #1702: 不做 ±15% 时长预过滤。旧逻辑按 duration_sec 缩小候选窗口，
-            # 但局部片段复用的两个视频时长必然不同（证据视频 20s vs 11s，差 42%），
-            # ±15% 窗口让同源视频互相不可见 → is_duplicate 恒 False。
-            # 全量遍历同用户视频（与 compute_duplicate_rate 口径一致），异源视频由
-            # fusion/temporal_coverage 阈值天然过滤（校准：异源最小汉明距离 24）。
-            existing_videos = video_repo.list_by_user(user_id)
+            dur_min = duration_sec * 0.85 if duration_sec > 0 else 0
+            dur_max = duration_sec * 1.15 if duration_sec > 0 else 0
+            existing_videos = video_repo.list_by_user(user_id, duration_min=dur_min, duration_max=dur_max)
         else:
             existing_videos = video_repo.list_by_project(project_id)
 
-        best_score = 0.0
-        best_result: Optional[dict] = None
-
         for existing in existing_videos:
-            # 排除自身（recompute 时当前视频已在候选列表里，否则自匹配距离 0 必最高分）
-            if exclude_video_id and existing.id == exclude_video_id:
-                continue
             if not existing.video_fingerprint:
                 continue
 
@@ -859,70 +677,61 @@ class VideoDeduplicator:
             if not existing_phashes:
                 continue
 
-            # 直方图 / 分片对象（chunk 表优先，回退 JSON 字段；JSON NULL 显式回退空列表）
+            # 计算每个新关键帧到已有关键帧的最小汉明距离
+            min_distances = []
+            for phash in fingerprint.keyframe_phashes:
+                distances = [hamming_distance(phash, ep) for ep in existing_phashes]
+                min_distances.append(min(distances))
+
+            # 帧匹配比例检查
+            matching_frames = sum(1 for d in min_distances if d < self.PHASH_THRESHOLD)
+            match_ratio = matching_frames / len(min_distances) if min_distances else 0
+            if match_ratio < MATCH_RATIO_THRESHOLD:
+                continue
+
+            # 中位数距离
+            median_distance = statistics.median(min_distances) if min_distances else 64
+            if median_distance >= self.PHASH_THRESHOLD:
+                continue
+
+            # 直方图融合（chunk 表优先，回退 JSON 字段；JSON NULL 显式回退空列表）
             if chunk_data:
                 existing_histograms = [c["color_histogram"] for c in chunk_data if c.get("color_histogram")]
-                existing_chunk_objects = chunk_data
             else:
                 existing_histograms = ef.get("color_histograms") or []
-                existing_chunk_objects = [
-                    {"phash_binary": pp, "start_time_ms": 0, "end_time_ms": 0} for pp in existing_phashes
-                ]
 
-            # Issue #1702: 统一评估每个候选（含局部片段复用），不再用
-            # "frame_match_rate<0.7 整条跳过" 的硬门槛——局部复用（如 B 结尾 2s
-            # ≈ A 中间 2s）帧比例天然低，但 coverage 能检出。
-            ev = self._evaluate_candidate(
-                fingerprint,
-                existing_phashes,
-                existing_histograms,
-                existing_chunk_objects,
-                query_duration_sec=fingerprint.duration,
-            )
-            logger.debug(
-                "check_duplicate candidate=%s min_distances=%s frame_match_rate=%.3f "
-                "temporal_coverage=%.3f median=%.1f fusion=%.3f segments=%d",
-                existing.id,
-                ev["min_distances"],
-                ev["frame_match_rate"],
-                ev["temporal_coverage"],
-                ev["median_distance"],
-                ev["fusion"],
-                len(ev["segments"]),
+            combined_score = self._compute_fusion_score(
+                median_distance, fingerprint.color_histograms, existing_histograms
             )
 
-            # 全片重复判定：融合分过阈 且（帧匹配比例 >=70% 或 局部覆盖 >=50%）
-            is_full_duplicate = ev["fusion"] >= DUPLICATE_THRESHOLD and (
-                ev["frame_match_rate"] >= MATCH_RATIO_THRESHOLD or ev["temporal_coverage"] >= PARTIAL_COVERAGE_THRESHOLD
+            if combined_score < DUPLICATE_THRESHOLD:
+                continue
+
+            # 滑动窗口时序匹配：获取具体重复片段
+            existing_chunk_objects = (
+                chunk_data
+                if chunk_data
+                else [{"phash_binary": p, "start_time_ms": 0, "end_time_ms": 0} for p in existing_phashes]
             )
-            if is_full_duplicate and ev["fusion"] > best_score:
-                best_score = ev["fusion"]
-                best_result = {
-                    "duplicate": True,
-                    "duplicate_of": existing.id,
-                    "reason": "phash_histogram_fusion",
-                    "similarity": ev["fusion"],
-                    "duplicate_segments": [
-                        {
-                            "query_start_ms": s.query_start_ms,
-                            "query_end_ms": s.query_end_ms,
-                            "target_start_ms": s.target_start_ms,
-                            "target_end_ms": s.target_end_ms,
-                            "avg_distance": round(s.avg_distance, 2),
-                        }
-                        for s in ev["segments"]
-                    ],
-                }
+            segments = find_duplicate_segments(fingerprint.chunks, existing_chunk_objects)
+
+            return {
+                "duplicate": True,
+                "duplicate_of": existing.id,
+                "reason": "phash_histogram_fusion",
+                "similarity": combined_score,
+                "duplicate_segments": [
+                    {
+                        "query_start_ms": s.query_start_ms,
+                        "query_end_ms": s.query_end_ms,
+                        "target_start_ms": s.target_start_ms,
+                        "target_end_ms": s.target_end_ms,
+                        "avg_distance": round(s.avg_distance, 2),
+                    }
+                    for s in segments
+                ],
+            }
 
-        if best_result:
-            return best_result
-        logger.info(
-            "check_duplicate no match (project=%s scope=%s): %d candidates evaluated, best_fusion=%.3f",
-            project_id,
-            scope,
-            len(existing_videos),
-            best_score,
-        )
         return None
 
     def check_batch_duplicate(
@@ -954,9 +763,6 @@ class VideoDeduplicator:
         video_repo = SQLAlchemyGeneratedVideoRepository(session)
         batch_videos = video_repo.list_by_batch(batch_id)
 
-        best_score = 0.0
-        best_result: Optional[dict] = None
-
         for existing in batch_videos:
             if existing.id == current_video_id:
                 continue
@@ -990,59 +796,59 @@ class VideoDeduplicator:
             if not existing_phashes:
                 continue
 
+            min_distances = []
+            for phash in fingerprint.keyframe_phashes:
+                distances = [hamming_distance(phash, ep) for ep in existing_phashes]
+                min_distances.append(min(distances))
+
+            # 帧匹配比例检查
+            matching_frames = sum(1 for d in min_distances if d < self.PHASH_THRESHOLD)
+            match_ratio = matching_frames / len(min_distances) if min_distances else 0
+            if match_ratio < MATCH_RATIO_THRESHOLD:
+                continue
+
+            median_distance = statistics.median(min_distances) if min_distances else 64
+            if median_distance >= self.PHASH_THRESHOLD:
+                continue
+
+            # 直方图融合（chunk 表优先，回退 JSON 字段；JSON NULL 显式回退空列表）
             if chunk_data:
                 existing_histograms = [c["color_histogram"] for c in chunk_data if c.get("color_histogram")]
-                existing_chunk_objects = chunk_data
             else:
                 existing_histograms = ef.get("color_histograms") or []
-                existing_chunk_objects = [
-                    {"phash_binary": pp, "start_time_ms": 0, "end_time_ms": 0} for pp in existing_phashes
-                ]
 
-            ev = self._evaluate_candidate(
-                fingerprint,
-                existing_phashes,
-                existing_histograms,
-                existing_chunk_objects,
-                query_duration_sec=fingerprint.duration,
-            )
-            logger.debug(
-                "check_batch_duplicate candidate=%s min_distances=%s frame_match_rate=%.3f "
-                "temporal_coverage=%.3f median=%.1f fusion=%.3f segments=%d",
-                existing.id,
-                ev["min_distances"],
-                ev["frame_match_rate"],
-                ev["temporal_coverage"],
-                ev["median_distance"],
-                ev["fusion"],
-                len(ev["segments"]),
+            combined_score = self._compute_fusion_score(
+                median_distance, fingerprint.color_histograms, existing_histograms
             )
 
-            is_full_duplicate = ev["fusion"] >= DUPLICATE_THRESHOLD and (
-                ev["frame_match_rate"] >= MATCH_RATIO_THRESHOLD or ev["temporal_coverage"] >= PARTIAL_COVERAGE_THRESHOLD
+            if combined_score < DUPLICATE_THRESHOLD:
+                continue
+
+            # 滑动窗口时序匹配
+            existing_chunk_objects = (
+                chunk_data
+                if chunk_data
+                else [{"phash_binary": p, "start_time_ms": 0, "end_time_ms": 0} for p in existing_phashes]
             )
-            if is_full_duplicate and ev["fusion"] > best_score:
-                best_score = ev["fusion"]
-                best_result = {
-                    "duplicate": True,
-                    "duplicate_of": existing.id,
-                    "reason": "batch_phash_histogram_fusion",
-                    "similarity": ev["fusion"],
-                    "duplicate_segments": [
-                        {
-                            "query_start_ms": s.query_start_ms,
-                            "query_end_ms": s.query_end_ms,
-                            "target_start_ms": s.target_start_ms,
-                            "target_end_ms": s.target_end_ms,
-                            "avg_distance": round(s.avg_distance, 2),
-                        }
-                        for s in ev["segments"]
-                    ],
-                }
+            segments = find_duplicate_segments(fingerprint.chunks, existing_chunk_objects)
+
+            return {
+                "duplicate": True,
+                "duplicate_of": existing.id,
+                "reason": "batch_phash_histogram_fusion",
+                "similarity": combined_score,
+                "duplicate_segments": [
+                    {
+                        "query_start_ms": s.query_start_ms,
+                        "query_end_ms": s.query_end_ms,
+                        "target_start_ms": s.target_start_ms,
+                        "target_end_ms": s.target_end_ms,
+                        "avg_distance": round(s.avg_distance, 2),
+                    }
+                    for s in segments
+                ],
+            }
 
-        if best_result:
-            return best_result
-        logger.info("check_batch_duplicate no match (batch=%s): best_fusion=%.3f", batch_id, best_score)
         return None
 
     def compute_duplicate_rate(
@@ -1091,7 +897,8 @@ class VideoDeduplicator:
         max_duplicate_rate = 0.0
         max_visual_similarity = 0.0
         match_count = 0
-        evaluated = 0
+
+        total_duration_ms = fingerprint.duration if fingerprint.duration else 0
 
         for existing in existing_videos:
             if current_video_id and existing.id == current_video_id:
@@ -1126,63 +933,57 @@ class VideoDeduplicator:
             if not existing_phashes or not fingerprint.keyframe_phashes:
                 continue
 
-            # 直方图 / 分片对象（chunk 表优先，回退 JSON 字段；JSON NULL 显式回退空列表）
+            min_distances = []
+            for phash in fingerprint.keyframe_phashes:
+                distances = [hamming_distance(phash, ep) for ep in existing_phashes]
+                min_distances.append(min(distances))
+
+            # frame_match_rate
+            total_frames = len(min_distances)
+            if total_frames == 0:
+                continue
+            matching_frames = sum(1 for d in min_distances if d < self.PHASH_THRESHOLD)
+            frame_match_rate = matching_frames / total_frames
+
+            # 帧匹配比例太低则跳过
+            if frame_match_rate < 0.3:
+                continue
+
+            # temporal_coverage_rate via find_duplicate_segments
+            existing_chunk_objects = (
+                chunk_data
+                if chunk_data
+                else [{"phash_binary": p, "start_time_ms": 0, "end_time_ms": 0} for p in existing_phashes]
+            )
+            segments = find_duplicate_segments(fingerprint.chunks, existing_chunk_objects)
+
+            if total_duration_ms > 0 and segments:
+                covered_ms = sum(s.query_end_ms - s.query_start_ms for s in segments)
+                temporal_coverage_rate = min(covered_ms / total_duration_ms, 1.0)
+            else:
+                temporal_coverage_rate = 0.0
+
+            # duplicate_rate = 0.4 * frame_match_rate + 0.6 * temporal_coverage_rate
+            dup_rate = (frame_match_rate * 0.4 + temporal_coverage_rate * 0.6) * 100
+
+            # visual_similarity (融合相似度，归一化 0~1)
+            median_distance = statistics.median(min_distances) if min_distances else 64
             if chunk_data:
                 existing_histograms = [c["color_histogram"] for c in chunk_data if c.get("color_histogram")]
-                existing_chunk_objects = chunk_data
             else:
+                # JSON NULL 显式回退空列表
                 existing_histograms = ef.get("color_histograms") or []
-                existing_chunk_objects = [
-                    {"phash_binary": pp, "start_time_ms": 0, "end_time_ms": 0} for pp in existing_phashes
-                ]
 
-            # Issue #1702: 统一评估；frame_match_rate 分母为 min(两视频分片数)，
-            # temporal_coverage 时长量纲在 _evaluate_candidate 内统一为毫秒。
-            ev = self._evaluate_candidate(
-                fingerprint,
-                existing_phashes,
-                existing_histograms,
-                existing_chunk_objects,
-                query_duration_sec=fingerprint.duration,
-            )
-            evaluated += 1
-            logger.debug(
-                "compute_duplicate_rate candidate=%s min_distances=%s frame_match_rate=%.3f "
-                "temporal_coverage=%.3f median=%.1f fusion=%.3f segments=%d",
-                existing.id,
-                ev["min_distances"],
-                ev["frame_match_rate"],
-                ev["temporal_coverage"],
-                ev["median_distance"],
-                ev["fusion"],
-                len(ev["segments"]),
-            )
+            visual_sim = self._compute_fusion_score(median_distance, fingerprint.color_histograms, existing_histograms)
 
-            # Issue #1702: 去掉 "frame_match_rate<0.3 整条跳过" 硬门槛——
-            # 局部片段复用帧比例天然低；coverage 为主指标，0 匹配自然得 0 分。
-            # duplicate_rate = 0.4 * frame_match_rate + 0.6 * temporal_coverage
-            dup_rate = (min(ev["frame_match_rate"], 1.0) * 0.4 + ev["temporal_coverage"] * 0.6) * 100
-
-            # 全片重复计数与 check_duplicate 判定口径一致
-            if ev["fusion"] >= DUPLICATE_THRESHOLD and (
-                ev["frame_match_rate"] >= MATCH_RATIO_THRESHOLD or ev["temporal_coverage"] >= PARTIAL_COVERAGE_THRESHOLD
-            ):
+            # 判定是否为重复（融合分数超过阈值）
+            if visual_sim >= DUPLICATE_THRESHOLD:
                 match_count += 1
 
             if dup_rate > max_duplicate_rate:
                 max_duplicate_rate = dup_rate
-                max_visual_similarity = ev["fusion"]
+                max_visual_similarity = visual_sim
 
-        logger.info(
-            "compute_duplicate_rate done (project=%s scope=%s): evaluated=%d max_rate=%.2f%% "
-            "max_visual_sim=%.3f matches=%d",
-            project_id,
-            scope,
-            evaluated,
-            max_duplicate_rate,
-            max_visual_similarity,
-            match_count,
-        )
         return {
             "duplicate_rate": round(max(max_duplicate_rate, 0.0), 2),
             "visual_similarity": round(max_visual_similarity, 4),
@@ -1198,20 +999,18 @@ def _save_fingerprint_chunks(
     session: Session,
 ) -> None:
     """将指纹分片数据批量写入 video_fingerprint_chunks 表。幂等：已有数据时跳过。"""
+    # 幂等检查：已有分片数据则跳过
+    existing_count = (
+        session.query(VideoFingerprintChunkModel).filter(VideoFingerprintChunkModel.video_id == video_id).count()
+    )
+    if existing_count > 0:
+        logger.debug("Fingerprint chunks already exist for video %s (%d chunks), skipping", video_id, existing_count)
+        return
+
     if not fingerprint.chunks:
         logger.warning("No chunks in fingerprint for video %s, skipping chunk save", video_id)
         return
 
-    # Issue #1702: recompute-dedup 重算时指纹算法已变（中心裁剪 + 新阈值），
-    # 旧分片必须替换而非跳过（旧实现"有数据就跳过"导致重算不刷新分片表）。
-    deleted = (
-        session.query(VideoFingerprintChunkModel)
-        .filter(VideoFingerprintChunkModel.video_id == video_id)
-        .delete(synchronize_session=False)
-    )
-    if deleted:
-        logger.info("Replaced %d stale fingerprint chunks for video %s", deleted, video_id)
-
     chunk_models = fingerprint.to_chunk_models(video_id, project_id, user_id)
     session.bulk_save_objects(chunk_models)
     logger.info("Saved %d fingerprint chunks for video %s", len(chunk_models), video_id)
@@ -1233,17 +1032,9 @@ def check_duplicate_task(self: Task, generated_video_id: str) -> dict:
             raise ValueError(f"Generated video {generated_video_id} not found")
 
         local_path = os.path.join(temp_dir, f"{generated_video_id}.mp4")
-        # Issue #1702: recompute 走的是 OSS 重新下载路径（正常生成流程用本地渲染文件，
-        # 不经此任务）。成片真实 OSS key 是生成时的
-        # generated/projects/{pid}/tasks/{task_id}/rendered_*.mp4（见 generation.py
-        # _upload_and_record），旧代码硬编码 projects/{pid}/generated/{vid}/{vid}.mp4
-        # 这个从不存在的 key，导致所有 recompute 任务下载 404、查重数据永远无法重算。
-        # 优先从 file_url 解析真实 key，旧 key 模式仅作回退。
-        download_key = getattr(video, "file_url", "") or ""
-        if not download_key:
-            download_key = f"projects/{video.project_id}/generated/{generated_video_id}/{generated_video_id}.mp4"
-            logger.warning("video %s has no file_url, falling back to legacy key %s", generated_video_id, download_key)
-        storage_service.download_file(download_key, local_path)
+        storage_service.download_file(
+            f"projects/{video.project_id}/generated/{generated_video_id}/{generated_video_id}.mp4", local_path
+        )
 
         fingerprint = deduplicator.compute_fingerprint(local_path)
 
@@ -1254,10 +1045,7 @@ def check_duplicate_task(self: Task, generated_video_id: str) -> dict:
             session,
             scope="user",
             user_id=video.user_id,
-            # Issue #1702: fingerprint.duration 单位已经是秒，旧代码 /1000 导致
-            # ±15% 时长预过滤窗口缩到 ~0.013s，scope=user 的跨项目查重永远返回 None。
-            duration_sec=fingerprint.duration if fingerprint.duration else 0,
-            exclude_video_id=generated_video_id,
+            duration_sec=fingerprint.duration / 1000 if fingerprint.duration else 0,
         )
 
         video.video_fingerprint = fingerprint.to_dict()

apps/worker/video_processing/dedup_helpers.py

@@ -92,8 +92,7 @@ def create_video_record_and_dedup(
                 logger.warning("Failed to save fingerprint chunks for %s: %s", video_id, chunk_err)
 
             # (a) 历史成片查重（跨项目全局 + 时长预过滤）
-            # Issue #1702: fingerprint.duration 单位是秒，旧代码 /1000 让时长预过滤失效
-            duration_sec = fingerprint.duration if fingerprint.duration else 0
+            duration_sec = fingerprint.duration / 1000 if fingerprint.duration else 0
             duplicate_result = deduplicator.check_duplicate(
                 fingerprint,
                 project_id,
@@ -101,7 +100,6 @@ def create_video_record_and_dedup(
                 scope="user",
                 user_id=user_id,
                 duration_sec=duration_sec,
-                exclude_video_id=video_id,
             )
 
             # (b) 批次内查重（仅当有 batch_id 时）

apps/worker/worker_app/celery_app.py

@@ -22,18 +22,10 @@ celery_app.conf.imports = (
 )
 
 # Celery Beat 定时任务调度
-# 注：worker 单实例内嵌 beat（entrypoint-worker.sh -B），定时任务不会重复执行
 celery_app.conf.beat_schedule = {
-    # pending 任务超时清理：worker 停止消费后，卡 pending 的任务 15 分钟内释放限流名额
     "cleanup-stale-pending-tasks": {
         "task": "worker.cleanup_stale_pending_tasks",
-        "schedule": 300.0,  # 每 5 分钟（秒）
-        "options": {"expires": 240},  # 4 分钟过期，避免堆积
-    },
-    # running 孤儿任务巡检：容器重启/进程被杀后卡 running 的任务，20 分钟无更新则判失败
-    "cleanup-stale-running-tasks": {
-        "task": "worker.cleanup_stale_running_tasks",
-        "schedule": 300.0,  # 每 5 分钟（秒）
-        "options": {"expires": 240},
+        "schedule": 600.0,  # 每 10 分钟（秒）
+        "options": {"expires": 300},  # 5 分钟过期，避免堆积
     },
 }

apps/worker/worker_app/tasks/_startup.py

@@ -7,34 +7,11 @@ from worker_app.db import SessionLocal
 
 logger = logging.getLogger(__name__)
 
+# 孤儿任务超时阈值：渲染任务超过此时间未更新则视为卡死
+ORPHAN_TASK_TIMEOUT_MINUTES = 10
 
-def cleanup_stale_running_with_session(repo, timeout_minutes: int) -> int:
-    """清理超时未更新的 running GenerationTask（可注入 repo 的纯核心，便于单测）。
-
-    Returns:
-        清理的任务数量
-    """
-    return repo.cleanup_stale_running(timeout_minutes)
-
-
-def cleanup_stale_pending_with_session(repo, timeout_minutes: int) -> int:
-    """清理超时 pending GenerationTask（可注入 repo 的纯核心，便于单测）。
-
-    Returns:
-        清理的任务数量
-    """
-    return repo.cleanup_stale_pending(timeout_minutes)
-
-
-# 孤儿任务超时阈值：running 任务超过此时间无进度更新则视为卡死。
-# 依据：worker.generate_video 硬超时 time_limit=11 分钟，正常任务不可能超过；
-# 20 分钟阈值覆盖硬超时 + 重试 + 余量，绝不误杀正常任务。
-ORPHAN_TASK_TIMEOUT_MINUTES = 20
-
-# Pending 任务超时阈值：任务创建后超过此时间仍未被 worker 拉取，
-# 说明 worker 已停止消费（容器异常/卡死），清掉释放限流名额。
-# 依据：满队列（20 pending）× 平均 2 分钟 / 并发 4 ≈ 10 分钟，15 分钟留余量。
-PENDING_TASK_TIMEOUT_MINUTES = 15
+# Pending 任务超时阈值：pending 任务在队列中等待超过此时间则自动清理
+PENDING_TASK_TIMEOUT_MINUTES = 30
 
 
 def cleanup_orphan_tasks(timeout_minutes: int = ORPHAN_TASK_TIMEOUT_MINUTES) -> int:  # pragma: no cover
@@ -55,11 +32,9 @@ def cleanup_orphan_tasks(timeout_minutes: int = ORPHAN_TASK_TIMEOUT_MINUTES) ->
 
     try:
         session = SessionLocal()
-        try:
-            repo = SQLAlchemyGenerationTaskRepository(session)
-            count = cleanup_stale_running_with_session(repo, timeout_minutes)
-        finally:
-            session.close()
+        repo = SQLAlchemyGenerationTaskRepository(session)
+        count = repo.cleanup_stale_running(timeout_minutes)
+        session.close()
         if count > 0:
             logger.warning("清理了 %d 个超时的孤儿 GenerationTask（超过 %d 分钟未更新）", count, timeout_minutes)
         else:
@@ -130,7 +105,7 @@ def cleanup_stale_pending_tasks(timeout_minutes: int = PENDING_TASK_TIMEOUT_MINU
     session = SessionLocal()
     try:
         repo = SQLAlchemyGenerationTaskRepository(session)
-        count = cleanup_stale_pending_with_session(repo, timeout_minutes)
+        count = repo.cleanup_stale_pending(timeout_minutes)
         if count > 0:
             logger.warning("清理了 %d 个超时的 pending GenerationTask（超过 %d 分钟未处理）", count, timeout_minutes)
         else:

apps/worker/worker_app/tasks/cleanup.py

@@ -1,18 +1,14 @@
 """定期清理任务 — Celery Beat 调度。
 
 包含：
-- cleanup_stale_pending_tasks: 定期清理卡在 pending 超时的 generation_tasks（worker 停止消费时占位）
-- cleanup_stale_running_tasks: 定期清理卡在 running 超时的 generation_tasks（容器重启/进程被杀后的孤儿）
+- cleanup_stale_pending_tasks: 定期清理卡在 pending 超时的 generation_tasks
 """
 
 import logging
 
 from celery import shared_task
 from worker_app.tasks._startup import (
-    ORPHAN_TASK_TIMEOUT_MINUTES,
     PENDING_TASK_TIMEOUT_MINUTES,
-    cleanup_orphan_tasks,
-    cleanup_stale_jobs,
     cleanup_stale_pending_tasks,
 )
 
@@ -23,12 +19,12 @@ logger = logging.getLogger(__name__)
 def scheduled_cleanup_stale_pending(timeout_minutes: int = PENDING_TASK_TIMEOUT_MINUTES) -> dict:
     """Celery Beat 调度的定期任务：清理超时的 pending 任务。
 
-    每 5 分钟执行一次（由 celery_app.py 的 beat_schedule 配置），
+    每 10 分钟执行一次（由 celery_app.py 的 beat_schedule 配置），
     查找所有 status='pending' 且 created_at < NOW() - timeout_minutes
-    的 generation_tasks，批量更新为 failed，释放限流名额。
+    的 generation_tasks，批量更新为 failed。
 
     Args:
-        timeout_minutes: 超时时间（分钟），默认 15 分钟
+        timeout_minutes: 超时时间（分钟），默认 30 分钟
 
     Returns:
         {"cleaned": int}
@@ -37,33 +33,3 @@ def scheduled_cleanup_stale_pending(timeout_minutes: int = PENDING_TASK_TIMEOUT_
     if count > 0:
         logger.info("[Beat] 清理了 %d 个超时 pending 任务（超时阈值 %d 分钟）", count, timeout_minutes)
     return {"cleaned": count}
-
-
-@shared_task(name="worker.cleanup_stale_running_tasks")
-def scheduled_cleanup_stale_running(timeout_minutes: int = ORPHAN_TASK_TIMEOUT_MINUTES) -> dict:
-    """Celery Beat 调度的定期任务：清理超时的 running 孤儿任务。
-
-    每 5 分钟执行一次。worker_ready 信号只在 worker 启动时清一次，
-    若 worker 没重启但任务卡死（上传挂起、进程 OOM 被内核杀掉等），
-    任务会永久卡在 running 占位。此任务做持续兜底：
-    查找 status='running' 且 updated_at < NOW() - timeout_minutes 的任务，
-    标记为 failed（原因：容器重启/超时中断），同时清理 Job 表孤儿。
-
-    Args:
-        timeout_minutes: 超时时间（分钟），默认 20 分钟
-        （worker.generate_video 硬超时 11 分钟，正常任务不可能超过 20 分钟）
-
-    Returns:
-        {"generation_tasks": int, "jobs": int}
-    """
-    gen_count = cleanup_orphan_tasks(timeout_minutes)
-    job_count = cleanup_stale_jobs(timeout_minutes)
-    total = gen_count + job_count
-    if total > 0:
-        logger.warning(
-            "[Beat] 清理孤儿任务: running GenerationTask=%d, Job=%d（超时阈值 %d 分钟）",
-            gen_count,
-            job_count,
-            timeout_minutes,
-        )
-    return {"generation_tasks": gen_count, "jobs": job_count}

packages/adapters/sqlalchemy_impl/generation_task_repository.py

@@ -138,52 +138,6 @@ class SQLAlchemyGenerationTaskRepository:
             .count()
         )
 
-    def count_running_by_user(self, user_id: str) -> int:
-        """统计指定用户处于 running 状态的任务数（用于限流提示展示）。"""
-        return (
-            self.session.query(GenerationTaskModel)
-            .filter(
-                GenerationTaskModel.created_by_user_id == user_id,
-                GenerationTaskModel.status == GenerationTaskStatus.RUNNING.value,
-            )
-            .count()
-        )
-
-    def count_running_total(self) -> int:
-        """统计全局处于 running 状态的任务数（worker 实际在执行的任务数）。"""
-        return (
-            self.session.query(GenerationTaskModel)
-            .filter(GenerationTaskModel.status == GenerationTaskStatus.RUNNING.value)
-            .count()
-        )
-
-    def estimate_avg_duration_seconds(self, limit: int = 20, default_seconds: float = 120.0) -> float:
-        """估算最近完成任务的平均耗时（秒），用于 429 限流提示的等待预估。
-
-        取最近 N 条 completed 任务的 (completed_at - started_at) 平均值；
-        无足够历史数据时返回 default_seconds。
-        用 Python 侧计算差值，避免 SQLite/PostgreSQL 方言差异。
-        """
-        rows = (
-            self.session.query(GenerationTaskModel.started_at, GenerationTaskModel.completed_at)
-            .filter(
-                GenerationTaskModel.status == GenerationTaskStatus.COMPLETED.value,
-                GenerationTaskModel.started_at.isnot(None),
-                GenerationTaskModel.completed_at.isnot(None),
-            )
-            .order_by(GenerationTaskModel.completed_at.desc())
-            .limit(limit)
-            .all()
-        )
-        durations = [
-            (completed - started).total_seconds()
-            for started, completed in rows
-            if completed and started and (completed - started).total_seconds() > 0
-        ]
-        if not durations:
-            return default_seconds
-        return sum(durations) / len(durations)
-
     def list_recent_by_user(self, user_id: str, limit: int = 5) -> list[GenerationTask]:
         models = (
             self.session.query(GenerationTaskModel)

packages/ports/generation_task_repository.py

@@ -20,12 +20,6 @@ class GenerationTaskRepository(Protocol):
 
     def count_pending_total(self) -> int: ...
 
-    def count_running_by_user(self, user_id: str) -> int: ...
-
-    def count_running_total(self) -> int: ...
-
-    def estimate_avg_duration_seconds(self, limit: int = 20, default_seconds: float = 120.0) -> float: ...
-
     def list_recent_by_user(self, user_id: str, limit: int = 5) -> list[GenerationTask]: ...
 
     def list_by_source_edit_plan(self, plan_id: str) -> list[GenerationTask]: ...

tests/unit/test_bad_fingerprint_filter.py

@@ -85,18 +85,17 @@ class TestIsBadFingerprint:
         assert VideoDeduplicator._is_bad_fingerprint(["abcdef0123456789"]) is False
 
     def test_all_identical_phashes_is_bad(self):
-        """>=8 帧且所有 phash 完全相同 → 黑屏/纯色视频（#1702：短帧不误杀）。"""
-        phashes = ["aaaaaaaaaaaaaaaa"] * 10
+        """多帧但所有 phash 完全相同 → 黑屏/纯色视频。"""
+        phashes = ["aaaaaaaaaaaaaaaa"] * 5
         assert VideoDeduplicator._is_bad_fingerprint(phashes) is True
 
-    def test_short_identical_phashes_not_bad(self):
-        """<8 帧完全相同不判坏——短视频内容连续时相邻采样帧 phash 天然相同（#1702）。"""
-        assert VideoDeduplicator._is_bad_fingerprint(["bbbbbbbbbbbbbbbb"] * 5) is False
-        assert VideoDeduplicator._is_bad_fingerprint(["bbbbbbbbbbbbbbbb", "bbbbbbbbbbbbbbbb"]) is False
+    def test_two_identical_phashes_is_bad(self):
+        """两帧完全相同也视为坏指纹。"""
+        assert VideoDeduplicator._is_bad_fingerprint(["bbbbbbbbbbbbbbbb", "bbbbbbbbbbbbbbbb"]) is True
 
     def test_all_very_similar_phashes_is_bad(self):
-        """>=8 帧 phash 之间的汉明距离都 < 3 且高占比 → 近似黑屏。"""
-        phashes = ["0000000000000000"] * 8 + ["0000000000000001", "0000000000000002"]
+        """多帧 phash 之间的汉明距离都 < 3 → 近似黑屏。"""
+        phashes = ["0000000000000000", "0000000000000001", "0000000000000002"]
         assert VideoDeduplicator._is_bad_fingerprint(phashes) is True
 
     def test_diverse_phashes_is_good(self):
@@ -123,9 +122,7 @@ class TestIsBadFingerprint:
         """已知黑屏视频的 phash 特征（全零或均匀分布）。"""
         assert VideoDeduplicator._is_bad_fingerprint(["0000000000000000"] * 10) is True
         assert VideoDeduplicator._is_bad_fingerprint(["ffffffffffffffff"] * 8) is True
-        assert VideoDeduplicator._is_bad_fingerprint(["9999999999999966"] * 8) is True
-        # <8 帧不判坏（#1702 短视频保护）
-        assert VideoDeduplicator._is_bad_fingerprint(["9999999999999966"] * 5) is False
+        assert VideoDeduplicator._is_bad_fingerprint(["9999999999999966"] * 6) is True
 
 
 # ── Helper ──────────────────────────────────────────────────────
@@ -154,13 +151,13 @@ class TestCheckDuplicateBadFingerprint:
         deduplicator = VideoDeduplicator()
         mock_session = MagicMock()
 
-        black_screen = _make_existing_video("vid-black", "md5_black", ["aaaaaaaaaaaaaaaa"] * 10)
+        black_screen = _make_existing_video("vid-black", "md5_black", ["aaaaaaaaaaaaaaaa"] * 5)
         mock_repo = MagicMock()
         mock_repo.list_by_user.return_value = [black_screen]
 
         fingerprint = VideoFingerprint(
             md5="md5_normal",
-            keyframe_phashes=["aaaaaaaaaaaaaaaa"] * 10,
+            keyframe_phashes=["aaaaaaaaaaaaaaaa"] * 5,
             color_histograms=[],
             duration=10.0,
             resolution=(1280, 720),
@@ -209,7 +206,7 @@ class TestCheckDuplicateBadFingerprint:
         deduplicator = VideoDeduplicator()
         mock_session = MagicMock()
 
-        black_screen = _make_existing_video("vid-black", "same_md5", ["aaaaaaaaaaaaaaaa"] * 10)
+        black_screen = _make_existing_video("vid-black", "same_md5", ["aaaaaaaaaaaaaaaa"] * 5)
         mock_repo = MagicMock()
         mock_repo.list_by_user.return_value = [black_screen]
 
@@ -286,8 +283,8 @@ class TestComputeDuplicateRateBadFingerprint:
         mock_session = MagicMock()
 
         videos = [
-            _make_existing_video("vid-b1", "md5_b1", ["aaaaaaaaaaaaaaaa"] * 10),
-            _make_existing_video("vid-b2", "md5_b2", ["cccccccccccccccc"] * 5),  # hamming(a,c)=32 > PHASH_THRESHOLD
+            _make_existing_video("vid-b1", "md5_b1", ["aaaaaaaaaaaaaaaa"] * 5),
+            _make_existing_video("vid-b2", "md5_b2", ["bbbbbbbbbbbbbbbb"] * 5),
         ]
         mock_repo = MagicMock()
         mock_repo.list_by_user.return_value = videos

tests/unit/test_dedup_1702_zero_rate_fix.py

@@ -1,548 +0,0 @@
-"""Issue #1702 — 查重率恒为 0% 修复：单测.
-
-覆盖验收要求：
-1. 同源不同裁剪的两个视频能检出非 0 相似度（指纹中心裁剪绕开降重 + 阈值校准）
-2. 局部片段复用（B 结尾 2s ≈ A 中间 2s）能检出
-3. 异源视频不误报（相似度接近 0）
-4. N=1 现有流程不回归
-5. P1 确定性 bug：时长预过滤单位 /1000、直方图归一化、temporal_coverage 量纲、阈值比较统一
-6. P0：±1 邻接对齐、短视频自适应连续门槛
-7. P2：0 匹配也要落日志
-"""
-
-from __future__ import annotations
-
-import logging
-import sys
-from pathlib import Path
-from unittest.mock import MagicMock, patch
-
-import pytest
-
-sys.modules.setdefault("cv2", MagicMock())
-
-ROOT = Path(__file__).resolve().parents[2]
-sys.path.insert(0, str(ROOT / "apps" / "worker"))
-sys.path.insert(0, str(ROOT / "packages"))
-
-
-from video_processing.dedup import (  # noqa: E402
-    PHASH_THRESHOLD,
-    SEGMENT_MATCH_THRESHOLD,
-    FingerprintChunk,
-    VideoDeduplicator,
-    VideoFingerprint,
-    find_duplicate_segments,
-)
-
-# ── helpers ────────────────────────────────────────────────────
-
-
-def _h(d: int) -> str:
-    """64-bit phash with exactly d bits set vs zero hash."""
-    bits = ["0"] * 64
-    for i in range(d):
-        bits[i] = "1"
-    return f"{int(''.join(bits), 2):016x}"
-
-
-def _chunk(phash: str, t0: float, t1: float):
-
-    return FingerprintChunk(
-        start_time_ms=int(t0 * 1000),
-        end_time_ms=int(t1 * 1000),
-        phash_binary=phash,
-        color_histogram=[],
-        frame_count=1,
-    )
-
-
-def _fingerprint(phashes, duration, chunks=None, md5="fp-md5-x"):
-
-    return VideoFingerprint(
-        md5=md5,
-        keyframe_phashes=list(phashes),
-        color_histograms=[],
-        duration=duration,
-        resolution=(1280, 720),
-        chunks=chunks or [],
-    )
-
-
-def _video(vid, phashes, duration=10.0, project_id="proj1"):
-    from packages.domain import GeneratedVideo
-
-    return GeneratedVideo(
-        id=vid,
-        project_id=project_id,
-        generation_task_id=f"task-{vid}",
-        name=f"video-{vid}.mp4",
-        file_url=f"https://example.com/{vid}.mp4",
-        file_size=1000,
-        duration=duration,
-        width=1280,
-        height=720,
-        fps=25.0,
-        video_fingerprint={"md5": f"md5-{vid}", "keyframe_phashes": list(phashes)},
-    )
-
-
-def _rate(deduplicator, fp, videos, session=None):
-    session_magic = MagicMock()
-    # 分片表无数据 -> 回退 JSON keyframe_phashes
-    session_magic.query.return_value.filter.return_value.order_by.return_value.all.return_value = []
-    with patch("video_processing.dedup.SQLAlchemyGeneratedVideoRepository") as MockRepo:
-        repo = MockRepo.return_value
-        repo.list_by_project.return_value = videos
-        repo.list_by_user.return_value = videos
-        return deduplicator.compute_duplicate_rate(fp, "proj1", "new-vid", session_magic, scope="project")
-
-
-def _check(deduplicator, fp, videos, scope="project", **kw):
-    session_magic = MagicMock()
-    session_magic.query.return_value.filter.return_value.order_by.return_value.all.return_value = []
-    with patch("video_processing.dedup.SQLAlchemyGeneratedVideoRepository") as MockRepo:
-        repo = MockRepo.return_value
-        repo.list_by_project.return_value = videos
-        repo.list_by_user.return_value = videos
-        return deduplicator.check_duplicate(fp, "proj1", session_magic, scope=scope, **kw)
-
-
-# ── P0-1/P0-2: 同源不同裁剪（距离 6~10）检出非 0 ──────────────
-
-
-class TestSameSourceDifferentCrop:
-    """同源成片：random_edge_crop 后 pHash 距离 6~10，应检出非 0 相似度。"""
-
-    def test_same_source_high_similarity_detected(self):
-
-        ddp = VideoDeduplicator()
-        # 新视频 5 个分片，每个 phash 与已有视频对应分片距离 6（< 阈值）
-        base = [_h(0) for _ in range(5)]
-        new = [_h(6) for _ in range(5)]
-        existing = _video("v-old", base, duration=11.0)
-        chunks = [_chunk(h, i * 2.2, (i + 1) * 2.2) for i, h in enumerate(new)]
-        fp = _fingerprint(new, 11.0, chunks=chunks)
-
-        result = _rate(ddp, fp, [existing], MagicMock())
-        assert result["duplicate_rate"] > 0
-        assert result["visual_similarity"] > 0
-
-    def test_same_source_distance_at_threshold_still_detected(self):
-        """距离正好等于阈值（<=）也要算匹配——阈值比较统一为 <=。"""
-
-        assert PHASH_THRESHOLD <= 16, "阈值应经真实数据校准保持在能检出同源裁剪/降重对的范围（#1702 二次校准为 16）"
-        ddp = VideoDeduplicator()
-        base = [_h(0) for _ in range(6)]
-        new = [_h(PHASH_THRESHOLD) for _ in range(6)]
-        existing = _video("v-old", base, duration=12.0)
-        chunks = [_chunk(h, i * 2, (i + 1) * 2) for i, h in enumerate(new)]
-        fp = _fingerprint(new, 12.0, chunks=chunks)
-
-        result = _rate(ddp, fp, [existing], MagicMock())
-        assert result["duplicate_rate"] > 0
-
-
-# ── P0-2: 局部片段复用（B 结尾 2s ≈ A 中间 2s） ────────────────
-
-
-class TestPartialReuse:
-    def test_partial_reuse_tail_overlap_detected(self):
-        """新视频 6 片，最后 2 片命中已有视频中间 2 片（距离 4），其余不匹配。
-
-        旧逻辑 frame_match_rate=2/6≈0.33（<0.3 硬跳过边界）+ MIN_CONSECUTIVE=5
-        导致完全检不出；新逻辑 coverage 为主指标 + 自适应门槛应检出。
-        """
-
-        ddp = VideoDeduplicator()
-        # 已有 8 片：索引 3、4 是被复用的镜头
-        old = [_h(20 + i) for i in range(8)]
-        # 新视频 6 片：最后 2 片对应 old[3], old[4]，距离 4；其余距离 30
-        new = [_h(50 + i) for i in range(4)] + [_h(4)] * 2
-        # 让 new[4] 与 old[3] 距离 4、new[5] 与 old[4] 距离 4（构造近似）
-        new[4] = f"{int('1' * 4 + '0' * 60, 2):016x}"
-        new[5] = f"{int('1' * 4 + '0' * 60, 2):016x}"
-        old[3] = _h(0)
-        old[4] = _h(0)
-
-        existing = _video("v-old", old, duration=16.0)
-        chunks = [_chunk(h, i * 2, (i + 1) * 2) for i, h in enumerate(new)]
-        fp = _fingerprint(new, 12.0, chunks=chunks)
-
-        result = _rate(ddp, fp, [existing], MagicMock())
-        # 局部复用：duplicate_rate 必须非 0
-        assert result["duplicate_rate"] > 0
-
-    def test_short_video_adaptive_consecutive_threshold(self):
-        """11s/5 片短视频：MIN_CONSECUTIVE 自适应 min(5, max(2, 5//2))=2，
-        2 片连续命中即报片段（旧值 5 让短视频永远无法报片段）。"""
-
-        q = [
-            FingerprintChunk(0, 2000, "f" * 16, []),
-            FingerprintChunk(2000, 4000, "0" * 16, []),
-            FingerprintChunk(4000, 6000, f"{int('11110000', 2):016x}", []),
-        ]
-        t = [
-            FingerprintChunk(0, 2000, "f" * 16, []),
-            FingerprintChunk(2000, 4000, "0" * 16, []),
-            FingerprintChunk(4000, 6000, "e" * 16, []),
-        ]
-        # 3 片视频自适应门槛 = min(5, max(2, 3//2)) = 2
-        segs = find_duplicate_segments(q, t)
-        assert len(segs) >= 1
-
-
-# ── P0-3: ±1 邻接窗口对齐 ─────────────────────────────────────
-
-
-class TestNeighborAlignment:
-    def test_neighbor_window_absorbs_boundary_jitter(self):
-        """切点错位导致目标索引偏移 ±1 时，连续匹配不应被中断。"""
-
-        q = [FingerprintChunk(i * 1000, (i + 1) * 1000, f"{i:016x}", []) for i in range(4)]
-        # 目标：前 3 片与 q 相同，但第 3 片最佳匹配偏移 +1（t[4]），t[3] 是无关内容
-        t_hashes = [f"{i:016x}" for i in range(3)] + ["f" * 16, f"{3:016x}"]
-        t = [FingerprintChunk(i * 1000, (i + 1) * 1000, h, []) for i, h in enumerate(t_hashes)]
-        segs = find_duplicate_segments(q, t)
-        # q[0],q[1] 精确匹配 t[0],t[1]；q[2]->t[2]；q[3]->t[4]（步进 2，窗口 ±1 内）
-        assert len(segs) >= 1
-        assert segs[0].query_end_ms >= 3000
-
-
-# ── P0-5 / 验收：异源不误报 ───────────────────────────────────
-
-
-class TestDifferentSourceNoFalsePositive:
-    def test_unrelated_videos_near_zero(self):
-
-        ddp = VideoDeduplicator()
-        # 异源：所有分片距离 >= 20
-        old = [_h(40 + i * 3 % 20) for i in range(6)]
-        new = [_h(0 + i) for i in range(6)]
-        existing = _video("v-old", old, duration=12.0)
-        chunks = [_chunk(h, i * 2, (i + 1) * 2) for i, h in enumerate(new)]
-        fp = _fingerprint(new, 12.0, chunks=chunks)
-
-        result = _rate(ddp, fp, [existing], MagicMock())
-        assert result["duplicate_rate"] == 0
-        assert result["visual_similarity"] < 0.7
-        assert result["match_count"] == 0
-
-    def test_check_duplicate_returns_none_for_unrelated(self):
-
-        ddp = VideoDeduplicator()
-        old = [_h(40 + i) for i in range(6)]
-        new = [_h(i) for i in range(6)]
-        existing = _video("v-old", old, duration=12.0)
-        fp = _fingerprint(new, 12.0)
-
-        result = _check(ddp, fp, [existing])
-        assert result is None
-
-
-# ── N=1 不回归 ────────────────────────────────────────────────
-
-
-class TestSingleChunkNoRegression:
-    def test_single_chunk_identical_detected(self):
-
-        ddp = VideoDeduplicator()
-        h = _h(2)
-        existing = _video("v-old", [h], duration=3.0)
-        chunks = [_chunk(h, 0, 3000)]
-        fp = _fingerprint([h], 3.0, chunks=chunks)
-        result = _rate(ddp, fp, [existing], MagicMock())
-        assert result["duplicate_rate"] > 0
-
-    def test_single_chunk_md5_exact_match(self):
-
-        ddp = VideoDeduplicator()
-        existing = _video("v-old", [_h(0)], duration=3.0)
-        existing.video_fingerprint["md5"] = "same"
-        fp = _fingerprint([_h(0)], 3.0, md5="same")
-        result = _check(ddp, fp, [existing])
-        assert result is not None
-        assert result["reason"] == "exact_md5_match"
-
-
-# ── P1-6: 时长预过滤单位 bug ──────────────────────────────────
-
-
-class TestDurationPrefilterUnit:
-    def test_user_scope_skips_duration_prefilter(self):
-        """Issue #1702: scope=user 跨项目查重不做 ±15% 时长预过滤。
-
-        旧逻辑 duration/1000 单位 bug 先修成秒，但 ±15% 窗口与局部片段复用
-        根本矛盾——复用片段的两个视频时长必然不同（证据视频 20s vs 11s 差 42%），
-        窗口内找不到对方导致 is_duplicate 恒 False。最终口径：scope=user 全量
-        遍历同用户视频（与 compute_duplicate_rate 一致），不传 duration_min/max。
-        """
-
-        ddp = VideoDeduplicator()
-        fp = _fingerprint([_h(0)], 13.5)
-        session_magic = MagicMock()
-        session_magic.query.return_value.filter.return_value.order_by.return_value.all.return_value = []
-        with patch("video_processing.dedup.SQLAlchemyGeneratedVideoRepository") as MockRepo:
-            repo = MockRepo.return_value
-            repo.list_by_user.return_value = []
-            ddp.check_duplicate(fp, "proj1", session_magic, scope="user", user_id="u1", duration_sec=fp.duration)
-            args, kwargs = repo.list_by_user.call_args
-            # 全量查询：不带任何时长过滤参数（局部复用必须跨时长比较）
-            assert "duration_min" not in kwargs
-            assert "duration_max" not in kwargs
-            assert args == ("u1",) or args == ()
-
-
-# ── P1-7: 颜色直方图归一化 ────────────────────────────────────
-
-
-class TestHistogramNormalization:
-    def test_bhattacharyya_coefficient_in_unit_range(self):
-        """Bhattacharyya 系数必须在 [0,1]（旧 L2 + 3 通道拼接算出 ~14.9）。"""
-
-        # 3 通道拼接、每通道概率分布（Σ=1）
-        hist_a = [0.5, 0.5] + [0.0] * 94 + [0.5, 0.5] + [0.0] * 94 + [0.5, 0.5] + [0.0] * 94
-        # 长度裁剪到 96（3 通道 × 32 bins）
-        hist_a = ([0.5, 0.5] + [0.0] * 30) * 3
-        hist_b = ([0.5, 0.5] + [0.0] * 30) * 3
-
-        coeff = VideoDeduplicator._bhattacharyya_coefficient(hist_a, hist_b)
-        assert 0.0 <= coeff <= 1.0
-        assert coeff > 0.99  # 完全相同 -> 1.0
-
-    def test_bhattacharyya_disjoint_hist_low(self):
-
-        hist_a = ([1.0] + [0.0] * 31) * 3
-        hist_b = ([0.0] * 31 + [1.0]) * 3
-        coeff = VideoDeduplicator._bhattacharyya_coefficient(hist_a, hist_b)
-        assert coeff < 0.05
-
-
-# ── P1-8: temporal_coverage 量纲 ──────────────────────────────
-
-
-class TestTemporalCoverageUnits:
-    def test_coverage_uses_milliseconds(self):
-        """命中片段 6s / 视频 12s -> coverage=0.5；旧 bug 把 duration（秒）当毫秒，
-        covered_ms(6000)/duration(12) = 500 -> min(1.0)=1.0 误判 100% 覆盖。"""
-
-        ddp = VideoDeduplicator()
-        old = [_h(0) for _ in range(6)]
-        new = [_h(0) for _ in range(3)] + [_h(30) for _ in range(3)]
-        existing = _video("v-old", old, duration=12.0)
-        # 新视频 12s，前 6s（3 片）与 old 相同
-        chunks = [_chunk(h, i * 2, (i + 1) * 2) for i, h in enumerate(new)]
-        fp = _fingerprint(new, 12.0, chunks=chunks)
-        result = _rate(ddp, fp, [existing], MagicMock())
-        # coverage 应约 0.5（3 片 × 2s = 6s / 12s），duplicate_rate ≈ (0.5*0.4 + 0.5*0.6)*100 = 50
-        assert 30 < result["duplicate_rate"] < 70
-
-
-# ── P1-9: 阈值比较统一 ────────────────────────────────────────
-
-
-class TestThresholdConsistency:
-    def test_frame_and_segment_thresholds_same_source(self):
-
-        assert SEGMENT_MATCH_THRESHOLD == PHASH_THRESHOLD
-        assert VideoDeduplicator.PHASH_THRESHOLD == PHASH_THRESHOLD
-
-
-# ── P2: 0 匹配也要有日志痕迹 ──────────────────────────────────
-
-
-class TestZeroMatchLogging:
-    def test_no_match_emits_info_log(self, caplog):
-
-        ddp = VideoDeduplicator()
-        old = [_h(40 + i) for i in range(5)]
-        existing = _video("v-old", old, duration=10.0)
-        fp = _fingerprint([_h(i) for i in range(5)], 10.0)
-
-        with caplog.at_level(logging.INFO, logger="video_processing.dedup"):
-            result = _check(ddp, fp, [existing])
-        assert result is None
-        assert any("no match" in r.message for r in caplog.records)
-
-
-# ── recompute 任务下载路径（#1702 连带修复：旧硬编码 key 404） ─────
-
-
-class TestRecomputeDownloadPath:
-    """recompute-dedup 走 check_duplicate_task，需要从 OSS 重新下载成片。
-
-    旧代码硬编码 projects/{pid}/generated/{vid}/{vid}.mp4（从不存在），
-    真实 key 在 file_url：generated/projects/{pid}/tasks/{tid}/rendered_*.mp4。
-    """
-
-    def test_task_downloads_from_file_url(self):
-        import inspect
-
-        import video_processing.dedup as dedup_mod
-
-        source = inspect.getsource(dedup_mod.check_duplicate_task)
-        # 下载 key 必须来自 video.file_url
-        assert 'getattr(video, "file_url"' in source or "video.file_url" in source
-        # 旧的硬编码 key 只能作为回退存在，不能是主路径
-        assert "falling back to legacy key" in source
-        # download_file 接收的是派生 key 而非硬编码 f-string
-        assert "storage_service.download_file(download_key" in source
-        assert '/generated/{generated_video_id}/{generated_video_id}.mp4"' not in source.replace(
-            'download_key = f"projects/{video.project_id}/generated/{generated_video_id}/{generated_video_id}.mp4"',
-            "",
-        )
-
-
-# ── check_duplicate 排除自身（#1702 连带修复：recompute 自匹配） ─────
-
-
-class TestCheckDuplicateExcludesSelf:
-    def test_exclude_video_id_skips_self_match(self):
-        """recompute 时当前视频已在候选列表：自匹配距离 0 分会让 duplicate_of
-        指向自己。exclude_video_id 必须跳过自身，返回真实的其他匹配或 None。
-        """
-
-        ddp = VideoDeduplicator()
-        h = _h(0)
-        # 候选列表里同时放「自己」（完全相同）和一个异源视频
-        self_video = _video("v-self", [h], duration=10.0)
-        other_video = _video("v-other", [_h(40 + i) for i in range(3)], duration=10.0)
-        fp = _fingerprint([h], 10.0)
-
-        session = MagicMock()
-        session.query.return_value.filter.return_value.order_by.return_value.all.return_value = []
-
-        # 不传 exclude → 自匹配命中（错误行为复现）
-        with patch("video_processing.dedup.SQLAlchemyGeneratedVideoRepository") as MockRepo:
-            MockRepo.return_value.list_by_project.return_value = [self_video, other_video]
-            result = ddp.check_duplicate(fp, "proj1", session)
-        assert result is not None and result["duplicate_of"] == "v-self"
-
-        # 传 exclude_video_id → 跳过自己，异源不匹配 → None
-        with patch("video_processing.dedup.SQLAlchemyGeneratedVideoRepository") as MockRepo:
-            MockRepo.return_value.list_by_project.return_value = [self_video, other_video]
-            result = ddp.check_duplicate(fp, "proj1", session, exclude_video_id="v-self")
-        assert result is None
-
-        # 排除自己后，真实同源其他视频仍能检出
-        real_dup = _video("v-real", [h], duration=10.0)
-        with patch("video_processing.dedup.SQLAlchemyGeneratedVideoRepository") as MockRepo:
-            MockRepo.return_value.list_by_project.return_value = [self_video, real_dup]
-            result = ddp.check_duplicate(fp, "proj1", session, exclude_video_id="v-self")
-        assert result is not None and result["duplicate_of"] == "v-real"
-
-
-# ── 阈值 16 二次校准 + 时序抖动对齐（#1702 第二轮真实数据校准） ──────
-
-
-class TestThreshold16Calibration:
-    """二次校准：staging 15 个真实成片实测——同源降重对中位数距离 14、
-    <=16 命中 8/11=0.73；异源 13 个候选每帧全局最近邻最小距离 18、<=16
-    命中全 0。阈值 16 检出同源且异源零误报（>=2bit 安全裕度）。"""
-
-    def test_threshold_calibrated_to_16(self):
-        assert PHASH_THRESHOLD == 16
-
-    @staticmethod
-    def _variant(phash: str, d: int) -> str:
-        """在 phash 基础上翻转恰好 d 个低位 bit → 与原哈希汉明距离恰为 d。"""
-        v = int(phash, 16)
-        for b in range(d):
-            v ^= 1 << b
-        return f"{v:016x}"
-
-    def test_distance_18_unrelated_not_matched(self):
-        """距离 18（异源实测最小最近邻距离）不判匹配，距离 16 判匹配。"""
-        ddp = VideoDeduplicator()
-        # 多样化 base（相邻帧各不相同，避免黑屏过滤器）
-        base = [_h(i + 4) for i in range(8)]
-        near = [self._variant(h, 16) for h in base]  # 同源降重：每帧距离恰 16
-        far = [self._variant(h, 18) for h in base]  # 异源边界：每帧距离恰 18
-
-        fp_near = _fingerprint(near, 8.0, chunks=[_chunk(h, i, i + 1) for i, h in enumerate(near)])
-        fp_far = _fingerprint(far, 8.0, chunks=[_chunk(h, i, i + 1) for i, h in enumerate(far)])
-
-        r_near = _rate(ddp, fp_near, [_video("v-base", base, duration=8.0)])
-        r_far = _rate(ddp, fp_far, [_video("v-base", base, duration=8.0)])
-
-        assert r_near["duplicate_rate"] > 0, "距离16的同源降重对必须检出"
-        assert r_far["duplicate_rate"] == 0.0, "距离18的异源对不得误报"
-        assert r_far["match_count"] == 0
-
-    def test_deduped_pair_frame_match_rate_over_threshold(self):
-        """真实场景比例：11 帧中 8 帧距离 <=16（0.73 >= 0.7），
-        其余 3 帧异源距离（>=18）——frame_match_rate 必须过 0.7 门槛。"""
-        ddp = VideoDeduplicator()
-        base = [_h(i + 4) for i in range(11)]
-        near = [self._variant(h, 14) for h in base[:8]]  # 中位数 14 的同源降重帧
-        # 异源帧用完全不同前缀（与 base 距离 >=30）
-        far = [_h(52 + i) for i in range(3)]
-        query = near + far
-
-        fp = _fingerprint(query, 11.0, chunks=[_chunk(h, i, i + 1) for i, h in enumerate(query)])
-        r = _rate(ddp, fp, [_video("v-base", base, duration=11.0)])
-        # frame_match_rate=8/11=0.73、时序片段覆盖 ~0.73
-        # → duplicate_rate = 0.4*0.73+0.6*0.73 ≈ 73%（空直方图回退下 fusion=0.6965
-        # 略低于 is_duplicate 的 0.70 判定阈值，故此处断言查重率而非 match_count；
-        # 真实视频带颜色直方图时 fusion≈0.80，staging A-C 实测 is_duplicate=True）
-        assert r["duplicate_rate"] >= 70.0
-
-
-class TestTemporalJitterAlignment:
-    """时序对齐允许目标索引正/反向 ±(neighbor_window+1) 抖动。
-
-    密集 1s 采样下相邻帧 pHash 接近，全局最近邻会在目标相邻帧间
-    正负 1 跳变（场景切割/取帧错位/局部倒退）；旧逻辑只允许正向
-    delta，把同源连续匹配拆碎，min_consecutive 门槛够不上而漏检。
-    """
-
-    def test_backward_jitter_keeps_run_continuous(self):
-        """匹配目标索引序列 0,1,2,1,2,3（含一次 -1 倒退）应保持同一 run。"""
-        from video_processing.dedup import find_duplicate_segments
-
-        # 构造 target 相邻帧 pHash 相同（距离0），query 帧的最近邻在
-        # target[1]/target[2] 之间抖动；全部 <= 阈值
-        t_hash = _h(0)
-        other = _h(40)
-        # target: 帧0-3 相同场景，帧4+ 异源
-        t_chunks = [_chunk(t_hash, i, i + 1) for i in range(4)] + [_chunk(other, i, i + 1) for i in range(4, 8)]
-        # query 6 帧同场景（最近邻会落到 target 0~3，索引可正可负）
-        q_chunks = [_chunk(t_hash, i, i + 1) for i in range(6)]
-
-        segments = find_duplicate_segments(q_chunks, t_chunks)
-        assert segments, "含 ±1 时序抖动的连续匹配必须形成片段"
-        # 6 帧匹配 >= min_consecutive(min(5,max(2,6//2))=5)，报为一个片段
-        assert len(segments) == 1
-        seg = segments[0]
-        assert seg.query_end_ms - seg.query_start_ms >= 5000
-
-    def test_large_backward_jump_breaks_run(self):
-        """目标索引倒退 > neighbor_window+1（如从 5 跳回 0）不属于抖动，
-        不桥接为同一片段；孤立短匹配 < min_consecutive 不报片段。"""
-        from video_processing.dedup import find_duplicate_segments
-
-        # 异源段：9-bit 不重叠段（相邻段隔 3 bit），跨段距离 18~24 > 阈值 16
-        def _bit_seg(start):
-            bits = ["0"] * 64
-            for b in range(9):
-                bits[start + b] = "1"
-            return f"{int(''.join(bits), 2):016x}"
-
-        t_hash = _bit_seg(0)  # 复用场景：bit 0-8
-        t_other = [_bit_seg(22 + 4 * i) for i in range(4)]  # target 异源段
-        q_other = [_bit_seg(40 + 4 * i) for i in range(3)]  # query 异源段
-        # target: 帧0 同场景；帧1-4 异源；帧5-6 同场景
-        t_chunks = (
-            [_chunk(t_hash, 0, 1)]
-            + [_chunk(t_other[i - 1], i, i + 1) for i in range(1, 5)]
-            + [_chunk(t_hash, i, i + 1) for i in range(5, 7)]
-        )
-        # query: 帧0 匹配 target[0]；帧1-3 异源（与 target 任何帧距离 >16）；帧4-5 匹配 target[5,6]
-        q_chunks = (
-            [_chunk(t_hash, 0, 1)]
-            + [_chunk(q_other[i - 1], i, i + 1) for i in range(1, 4)]
-            + [_chunk(t_hash, i, i + 1) for i in range(4, 6)]
-        )
-        segments = find_duplicate_segments(q_chunks, t_chunks)
-        # 两段各 1、2 帧 < min_consecutive=5 → 不报片段（大跳跃不桥接）
-        assert segments == []

tests/unit/test_dedup_engine.py

@@ -358,11 +358,11 @@ class TestVideoDeduplicatorCheckDuplicate:
         finally:
             self._restore_repo(mod, orig)
 
-    def test_highest_score_match_returned(self, deduplicator, mock_session):
-        """Issue #1702: 遍历所有候选取融合分最高者（旧逻辑首个过阈即返回）。"""
-        # vid-1: 距离=1 bit（0x03 XOR 0x01 = 0x02 → 1 bit），通过阈值
+    def test_first_match_returned(self, deduplicator, mock_session):
+        """返回第一个通过阈值的匹配（非最优匹配）。"""
+        # vid-1: 距离=2 bits（0x03 XOR 0x01 = 0x02 → 1 bit），通过阈值
         vid1 = self._make_existing_video("vid-1", "md5_1", phashes=["0000000000000003"])
-        # vid-2: 距离=0 bits（完全匹配），融合分更高
+        # vid-2: 距离=0 bits（完全匹配）
         vid2 = self._make_existing_video("vid-2", "md5_2", phashes=["0000000000000001"])
 
         mock_repo = MagicMock()
@@ -380,8 +380,8 @@ class TestVideoDeduplicatorCheckDuplicate:
         try:
             result = deduplicator.check_duplicate(fingerprint, "proj-1", mock_session)
             assert result is not None
-            # 两个候选都过阈，返回融合分最高的 vid-2（距离 0 < 1）
-            assert result["duplicate_of"] == "vid-2"
+            # 返回第一个通过阈值的匹配（vid-1 距离=1 < 10）
+            assert result["duplicate_of"] == "vid-1"
         finally:
             self._restore_repo(mod, orig)
 

tests/unit/test_dedup_pure.py

@@ -185,14 +185,11 @@ class TestBhattacharyyaCoefficient:
     """_bhattacharyya_coefficient Bhattacharyya 系数测试."""
 
     def test_identical_histograms(self):
-        """完全相同的直方图系数为1.0（#1702：按 Σ 归一，概率分布语义）。"""
-        hist = [0.5, 0.5, 0.0, 0.0]  # Σ=1 的概率分布
+        """完全相同的直方图系数为1.0."""
+        hist = [0.5, 0.5, 0.0, 0.3]
         bc = VideoDeduplicator._bhattacharyya_coefficient(hist, hist)
-        assert bc == pytest.approx(1.0)
-        # 非归一化输入也归一到 1.0（三通道拼接 Σ=3 的等价情形）
-        hist3 = [0.5, 0.5, 0.0, 0.3]
-        bc3 = VideoDeduplicator._bhattacharyya_coefficient(hist3, hist3)
-        assert bc3 == pytest.approx(1.0)
+        # Σ √(a[i]*a[i]) = Σ a[i] = 1.0 (normalized)
+        assert bc == pytest.approx(sum(h for h in hist))
 
     def test_zero_histograms(self):
         """全零直方图系数为0."""
@@ -205,10 +202,10 @@ class TestBhattacharyyaCoefficient:
         assert bc == pytest.approx(0.0)
 
     def test_different_lengths(self):
-        """不同长度直方图取最小长度对齐，并按各自总量归一（#1702 概率分布语义）。"""
-        # 对齐到前 2 维：coeff = 2，norm = √(Σa·Σb) = √(2·2) = 2 → 1.0
+        """不同长度直方图取最小长度对齐."""
         bc = VideoDeduplicator._bhattacharyya_coefficient([1.0, 1.0, 0.0, 0.0], [1.0, 1.0])
-        assert bc == pytest.approx(1.0)
+        # 对齐到前2维: √(1*1) + √(1*1) = 2.0
+        assert bc == pytest.approx(2.0)
 
     def test_known_value(self):
         """已知值验证."""

tests/unit/test_dedup_v2.py

@@ -103,7 +103,6 @@ from video_processing.dedup import (  # noqa: E402
     MIN_CONSECUTIVE_MATCHES,
     MIN_KEYFRAME_INTERVAL_SEC,
     MIN_KEYFRAMES,
-    PHASH_THRESHOLD,
     PHASH_WEIGHT,
     SCENE_CHANGE_THRESHOLD,
     SEGMENT_MATCH_THRESHOLD,
@@ -270,17 +269,22 @@ class TestFindDuplicateSegments:
         注意：使用不同的 hash 对，确保后半部分帧距离 > 阈值。
         """
         same_hash = "aaaaaaaaaaaaaaaa"
-        # 4 帧匹配，后面 6 帧用与匹配哈希距离 32 的不匹配哈希（> PHASH_THRESHOLD=16）
-        nomatch_hash = "cccccccccccccccc"  # hamming(aaaa, cccc)=32
+        # 4 帧匹配，后面 6 帧各自不同（在 query 和 target 中使用不同 hash）
         chunks_a = [_make_chunk(i * 1000, (i + 1) * 1000, same_hash) for i in range(4)] + [
-            _make_chunk(i * 1000, (i + 1) * 1000, nomatch_hash) for i in range(4, 10)
+            _make_chunk(i * 1000, (i + 1) * 1000, "bbbbbbbbbbbbbbbb") for i in range(4, 10)
         ]
         chunks_b = [_make_chunk(i * 1000, (i + 1) * 1000, same_hash) for i in range(4)] + [
-            _make_chunk(i * 1000, (i + 1) * 1000, nomatch_hash) for i in range(4, 10)
+            _make_chunk(i * 1000, (i + 1) * 1000, "cccccccccccccccc") for i in range(4, 10)
         ]
-        # hamming(aaaa..., cccc...) = 32 > PHASH_THRESHOLD(16)，后半段不匹配；
-        # 前 4 帧匹配 < min_consecutive=5，不形成片段
+
+        # hamming("bbbb...", "cccc...") should be > 8 (SEGMENT_MATCH_THRESHOLD)
+        # b=1011, c=1100 → 4 bits differ per hex digit × 16 digits = 64 bits total? No...
+        # Actually: hamming_distance("bbbbbbbbbbbbbbbb", "cccccccccccccccc")
+        # b=0xb=1011, c=0xc=1100 → XOR=0111=0x7 → 3 bits per digit × 16 = 48
+        # That's > 8 so won't match
+
         segments = find_duplicate_segments(chunks_a, chunks_b)
+        # 只有 4 帧匹配（< min_consecutive=5），所以不报告
         assert segments == []
 
     def test_max_gap_behavior(self):
@@ -289,12 +293,10 @@ class TestFindDuplicateSegments:
         关键：间隙帧必须在 query 和 target 中使用不同 hash，使其真正不匹配。
         """
         match_hash = "aaaaaaaaaaaaaaaa"
-        # 间隙/尾部哈希与 match_hash 及彼此之间汉明距离均 >64 (> PHASH_THRESHOLD=16)，
-        # 确保在 ±(neighbor_window+1) 时序抖动对齐窗口内也不会误匹配
-        gap_hash_a = "ffffffffffffffff"  # hamming(a,f)=128
-        gap_hash_b = "9999999999999999"  # hamming(a,9)=128, hamming(f,9)=128
-        tail_hash_a = "7777777777777777"  # hamming(a,7)=192
-        tail_hash_b = "1111111111111111"  # hamming(a,1)=192, hamming(7,1)=128
+        gap_hash_a = "bbbbbbbbbbbbbbbb"  # query 端
+        gap_hash_b = "cccccccccccccccc"  # target 端（与 query 端距离 > 8）
+        tail_hash_a = "dddddddddddddddd"
+        tail_hash_b = "eeeeeeeeeeeeeeee"
 
         # 5 帧匹配, 1 帧间隙, 3 帧匹配, 5 帧不匹配
         hashes_a = [match_hash] * 5 + [gap_hash_a] + [match_hash] * 3 + [tail_hash_a] * 5
@@ -316,10 +318,10 @@ class TestFindDuplicateSegments:
     def test_max_gap_exceeded(self):
         """间隙超过 max_gap → 分成两段."""
         match_hash = "aaaaaaaaaaaaaaaa"
-        gap_hash_a = "ffffffffffffffff"  # hamming(a,f)=128
-        gap_hash_b = "9999999999999999"  # hamming(a,9)=128
-        tail_hash_a = "7777777777777777"  # hamming(a,7)=192
-        tail_hash_b = "1111111111111111"  # hamming(a,1)=192
+        gap_hash_a = "bbbbbbbbbbbbbbbb"
+        gap_hash_b = "cccccccccccccccc"
+        tail_hash_a = "dddddddddddddddd"
+        tail_hash_b = "eeeeeeeeeeeeeeee"
 
         # 5 帧匹配, 3 帧间隙 (> max_gap=2), 5 帧匹配, 5 帧不匹配
         hashes_a = [match_hash] * 5 + [gap_hash_a] * 3 + [match_hash] * 5 + [tail_hash_a] * 5
@@ -482,10 +484,8 @@ class TestBackwardCompatibility:
         chunks_b = [{"phash_binary": "aaaaaaaaaaaaaaaa", "start_time_ms": 0, "end_time_ms": 5000}]
 
         segments = find_duplicate_segments(chunks_a, chunks_b)
-        # Issue #1702: 自适应门槛 min(5, max(2, 1//2))=2，1 帧不成段；
-        # N=1 的检出由 _evaluate_candidate 匹配帧回退兜底（见 test_dedup_1702）。
-        # 这里只要求不崩溃。
-        assert isinstance(segments, list)
+        # 1 帧 < min_consecutive=5，不会报重复
+        assert segments == []
 
 
 # ── TestConstants ───────────────────────────────────────────────
@@ -495,11 +495,8 @@ class TestConstants:
     """常量值验证 — 使用已在模块顶部导入的常量，避免重新 import."""
 
     def test_segment_match_threshold(self):
-        # Issue #1702 二次校准：阈值经 staging 真实数据两轮回归——
-        # 第一轮同源 4/11、异源 min=24 定 12；第二轮扩样本（15 个真实成片）
-        # 同源降重对中位数距离 14、<=16 命中 8/11=0.73，异源 13 个候选
-        # <=16 命中全 0、最近邻最小距离 18 → 校准为 16。
-        assert SEGMENT_MATCH_THRESHOLD == PHASH_THRESHOLD == 16
+        # 从已导入的 find_duplicate_segments 默认参数间接验证
+        assert SEGMENT_MATCH_THRESHOLD == 8
 
     def test_min_consecutive_matches(self):
         assert MIN_CONSECUTIVE_MATCHES == 5

tests/unit/test_duplicate_rate_scope.py

@@ -132,14 +132,9 @@ class TestCheckDuplicateScopeUser:
 
 
 class TestDurationPrefilter:
-    """Issue #1702: scope=user 跨项目查重不做时长预过滤。
+    """test_duration_prefilter：时长 ±15% 过滤."""
 
-    局部片段复用的两个视频时长必然不同（证据视频 20s vs 11s，差 42%），
-    旧的 ±15% 窗口会让同源视频互相不可见 → is_duplicate 恒 False。
-    全量遍历同用户视频，异源视频由 fusion/temporal_coverage 阈值天然过滤。
-    """
-
-    def test_user_scope_no_duration_filter(self):
+    def test_duration_prefilter_passes_correct_range(self):
         from video_processing.dedup import VideoDeduplicator
 
         deduplicator = VideoDeduplicator()
@@ -158,13 +153,12 @@ class TestDurationPrefilter:
                 duration_sec=30.0,
             )
 
-        # scope=user 全量遍历：位置参数只传 user_id，kwargs 不含时长过滤
+        # Should pass duration_min=25.5, duration_max=34.5 (30 ± 15%)
         call_args = mock_repo.list_by_user.call_args
-        assert call_args[0] == ("user1",)
-        assert "duration_min" not in call_args[1]
-        assert "duration_max" not in call_args[1]
+        assert call_args[1]["duration_min"] == pytest.approx(25.5, abs=0.1)
+        assert call_args[1]["duration_max"] == pytest.approx(34.5, abs=0.1)
 
-    def test_user_scope_no_duration_filter_when_zero(self):
+    def test_no_duration_prefilter_when_zero(self):
         from video_processing.dedup import VideoDeduplicator
 
         deduplicator = VideoDeduplicator()
@@ -184,34 +178,8 @@ class TestDurationPrefilter:
             )
 
         call_args = mock_repo.list_by_user.call_args
-        assert "duration_min" not in call_args[1]
-        assert "duration_max" not in call_args[1]
-
-    def test_project_scope_also_no_duration_filter(self):
-        """scope=project 走 list_by_project，本来就不做时长过滤。"""
-        from video_processing.dedup import VideoDeduplicator
-
-        deduplicator = VideoDeduplicator()
-        fingerprint = _make_fingerprint(duration_ms=30000)
-        session = MagicMock()
-
-        with patch("video_processing.dedup.SQLAlchemyGeneratedVideoRepository") as MockRepo:
-            mock_repo = MockRepo.return_value
-            mock_repo.list_by_project.return_value = []
-            deduplicator.check_duplicate(
-                fingerprint,
-                "proj1",
-                session,
-                scope="project",
-                user_id="user1",
-                duration_sec=30.0,
-            )
-
-        mock_repo.list_by_project.assert_called_once()
-        call_args = mock_repo.list_by_project.call_args
-        assert call_args[0] == ("proj1",)
-        assert "duration_min" not in call_args[1]
-        assert "duration_max" not in call_args[1]
+        assert call_args[1]["duration_min"] == 0
+        assert call_args[1]["duration_max"] == 0
 
 
 class TestComputeDuplicateRateFormula:

tests/unit/test_fingerprint_chunks.py

@@ -3,7 +3,7 @@
 覆盖：
 - 分片策略：60秒视频 → 30片，120秒视频 → 24片
 - VideoFingerprint.to_chunk_models() 输出正确
-- _save_fingerprint_chunks 替换语义（Issue #1702：重算时先删旧分片再写入）
+- _save_fingerprint_chunks 幂等性（已有数据跳过）
 - to_dict() 向后兼容
 """
 
@@ -169,15 +169,11 @@ class TestVideoFingerprintToChunkModels:
         assert models == []
 
 
-class TestSaveFingerprintChunksReplace:
-    """测试 _save_fingerprint_chunks 替换语义（Issue #1702）。
+class TestSaveFingerprintChunksIdempotent:
+    """测试 _save_fingerprint_chunks 幂等性。"""
 
-    重算查重时指纹算法已升级（中心裁剪 + 新采样/阈值），旧分片必须先删除
-    再写入新分片，否则 recompute-dedup 永远读到旧指纹、修复对存量视频不生效。
-    """
-
-    def test_save_replaces_existing(self):
-        """已有分片数据时：先删除旧分片，再写入新分片。"""
+    def test_save_skips_existing(self):
+        """已有分片数据时跳过写入。"""
         fp = VideoFingerprint(
             md5="abc",
             keyframe_phashes=["a1b2"],
@@ -190,22 +186,16 @@ class TestSaveFingerprintChunksReplace:
         )
 
         session = MagicMock()
-        # Mock: 删除旧分片返回 3（旧算法留下的 3 条分片）
-        session.query.return_value.filter.return_value.delete.return_value = 3
+        # Mock: 已有 1 条分片数据
+        session.query.return_value.filter.return_value.count.return_value = 1
 
         _save_fingerprint_chunks(fp, video_id="v1", project_id="p1", user_id="u1", session=session)
 
-        # 必须先执行删除
-        session.query.return_value.filter.return_value.delete.assert_called_once()
-        # 新分片必须写入
-        session.bulk_save_objects.assert_called_once()
-        saved_models = session.bulk_save_objects.call_args[0][0]
-        assert len(saved_models) == 1
-        assert saved_models[0].video_id == "v1"
-        assert saved_models[0].phash_binary == "a1b2"
+        # bulk_save_objects 不应被调用
+        session.bulk_save_objects.assert_not_called()
 
     def test_save_writes_new(self):
-        """无旧分片时直接写入。"""
+        """无分片数据时写入。"""
         fp = VideoFingerprint(
             md5="abc",
             keyframe_phashes=["a1b2"],
@@ -218,12 +208,12 @@ class TestSaveFingerprintChunksReplace:
         )
 
         session = MagicMock()
-        # Mock: 无旧分片
-        session.query.return_value.filter.return_value.delete.return_value = 0
+        # Mock: 无分片数据
+        session.query.return_value.filter.return_value.count.return_value = 0
 
         _save_fingerprint_chunks(fp, video_id="v1", project_id="p1", user_id="u1", session=session)
 
-        session.query.return_value.filter.return_value.delete.assert_called_once()
+        # bulk_save_objects 应被调用一次
         session.bulk_save_objects.assert_called_once()
         saved_models = session.bulk_save_objects.call_args[0][0]
         assert len(saved_models) == 1
@@ -231,7 +221,7 @@ class TestSaveFingerprintChunksReplace:
         assert saved_models[0].phash_binary == "a1b2"
 
     def test_save_skips_no_chunks(self):
-        """指纹无 chunks 时跳过（不删不写）。"""
+        """指纹无 chunks 时跳过。"""
         fp = VideoFingerprint(
             md5="abc",
             keyframe_phashes=[],
@@ -242,11 +232,11 @@ class TestSaveFingerprintChunksReplace:
         )
 
         session = MagicMock()
+        session.query.return_value.filter.return_value.count.return_value = 0
 
         _save_fingerprint_chunks(fp, video_id="v1", project_id="p1", user_id="u1", session=session)
 
-        # 无 chunks：不查询、不删除、不写入
-        session.query.assert_not_called()
+        # bulk_save_objects 不应被调用
         session.bulk_save_objects.assert_not_called()
 
 

tests/unit/test_phash_threshold_calibration_1658.py

@@ -102,7 +102,6 @@ from video_processing.dedup import (  # noqa: E402
     DUPLICATE_THRESHOLD,
     HISTOGRAM_WEIGHT,
     MATCH_RATIO_THRESHOLD,
-    PHASH_THRESHOLD,
     PHASH_WEIGHT,
     VideoDeduplicator,
 )
@@ -129,17 +128,11 @@ _ZERO_HIST = [0.0] * 96  # 全黑视频的全零直方图（有效数据）
 
 
 class TestThresholdCalibration:
-    """pHash 阈值校准（#1658 收紧到 8，#1702 两轮真实数据重校准 12→16）。
+    """pHash 阈值由 10 收紧到 8（Issue #1658）。"""
 
-    #1702 第一轮 staging 离线实验：同帧两次 2-5% 随机裁剪距离 4~10；同源成片
-    （密集 1s 采样）<=12 命中 4/11、异源成片最小距离 24 → 初定 12。
-    #1702 第二轮（证据视频 B->A 仍漏检）扩样本到该用户 15 个真实成片实测：
-    同源降重对中位数距离 14、<=16 命中 8/11=0.73；异源 13 个候选 <=16 命中
-    全 0、每帧全局最近邻最小距离 18 → 校准为 16（与异源仍有 >=2bit 裕度）。
-    """
-
-    def test_phash_threshold_is_calibrated(self):
-        assert VideoDeduplicator.PHASH_THRESHOLD == PHASH_THRESHOLD == 16
+    def test_phash_threshold_is_8(self):
+        """PHASH_THRESHOLD 必须为 8（旧值 10 会放过 8~9 汉明距离的不同视频）。"""
+        assert VideoDeduplicator.PHASH_THRESHOLD == 8
 
     def test_match_ratio_threshold_constant(self):
         assert MATCH_RATIO_THRESHOLD == 0.7
@@ -151,22 +144,22 @@ class TestThresholdCalibration:
         assert PHASH_WEIGHT == 0.7
         assert HISTOGRAM_WEIGHT == 0.3
 
-    def test_threshold_matching_semantics(self):
-        """阈值比较统一为 <=（帧匹配与片段匹配同一口径）。
+    def test_threshold_tightening_excludes_distance_8_and_9(self):
+        """距离 8、9 的帧：旧阈值 10 下算匹配，新阈值 8 下不算匹配。
 
-        场景：5 个关键帧距离为 [10, 14, 16, 18, 26]。
-        - <=16（#1702 二次校准阈值）：3 帧匹配 → 0.6 < 0.7，被帧比例门槛
-          拦截（异源安全边界：真实数据异源最近邻最小距离 18，<=16 命中 0）
-        - 距离正好 16 的同源降重帧应算匹配（< 与 <= 口径统一）
+        场景：5 个关键帧距离为 [7, 7, 7, 9, 9]。
+        - 旧阈值 10：5 帧全部 < 10 → match_ratio = 1.0（误放过）
+        - 新阈值 8：仅 3 帧 < 8 → match_ratio = 0.6 < 0.7（正确跳过）
         """
-        distances = [10, 14, 16, 18, 26]
-        matched = sum(1 for d in distances if d <= VideoDeduplicator.PHASH_THRESHOLD)
-        assert matched == 3
-        assert matched / len(distances) == 0.6
-        assert matched / len(distances) < MATCH_RATIO_THRESHOLD
+        distances = [7, 7, 7, 9, 9]
 
-        # 异源安全边界（实测最小距离 18）及以上绝不匹配
-        assert not any(d <= VideoDeduplicator.PHASH_THRESHOLD for d in (18, 24, 26, 30))
+        matched_old = sum(1 for d in distances if d < 10)
+        assert matched_old == 5  # 旧行为：全匹配 → 误判风险
+
+        matched_new = sum(1 for d in distances if d < VideoDeduplicator.PHASH_THRESHOLD)
+        assert matched_new == 3
+        assert matched_new / len(distances) == 0.6
+        assert matched_new / len(distances) < MATCH_RATIO_THRESHOLD  # 被帧比例门槛拦截
 
 
 # ── TestComputeFusionScore：统一融合得分方法 ────────────────────

tests/unit/test_task_fault_tolerance_1709.py

@@ -1,315 +0,0 @@
-"""Issue #1709 任务容错：孤儿任务恢复 + 429 限流结构化提示。
-
-覆盖：
-1. 仓储层：count_running_by_user/count_running_total 计数正确（预览/正式任务都计入）
-2. 仓储层：estimate_avg_duration_seconds 耗时估算（有历史/无历史）
-3. 限流核心：build_rate_limit_detail 返回结构化 code/message/排队数/预计等待
-4. worker 侧：cleanup_stale_running/pending 核心函数——中断任务被重置为 failed
-   且原因写明（容器重启/超时中断），正常任务不受影响
-"""
-
-import sys
-from datetime import datetime, timedelta, timezone
-from pathlib import Path
-from unittest.mock import MagicMock
-
-sys.path.insert(0, str(Path(__file__).resolve().parents[2] / "apps" / "api"))
-sys.path.insert(0, str(Path(__file__).resolve().parents[2] / "apps" / "worker"))
-
-# 预注入 mock worker_app.db，防止真实数据库连接初始化（与其他 worker 测试同模式）
-_mock_db = MagicMock()
-_mock_db.SessionLocal = MagicMock()
-sys.modules.setdefault("worker_app.db", _mock_db)
-
-from app.core import task_enqueue  # noqa: E402
-from sqlalchemy import create_engine, text  # noqa: E402
-from sqlalchemy.orm import sessionmaker  # noqa: E402
-from worker_app.tasks import _startup  # noqa: E402
-
-from packages.adapters.sqlalchemy_impl.generation_task_repository import (  # noqa: E402
-    SQLAlchemyGenerationTaskRepository,
-)
-from packages.adapters.sqlalchemy_impl.models import Base  # noqa: E402
-from packages.domain import GenerationTask, GenerationTaskStatus  # noqa: E402
-
-
-def _repository():
-    engine = create_engine("sqlite:///:memory:", connect_args={"check_same_thread": False})
-    Base.metadata.create_all(engine)
-    session = sessionmaker(bind=engine)()
-    return SQLAlchemyGenerationTaskRepository(session), session, engine
-
-
-def _make_task(**kwargs) -> GenerationTask:
-    defaults = dict(
-        project_id="proj-1",
-        asset_library_id="lib-1",
-        created_by_user_id="user-1",
-    )
-    defaults.update(kwargs)
-    return GenerationTask.create(**defaults)
-
-
-def _age_task(engine, task_id, *, updated_minutes=None, created_minutes=None):
-    """用 SQL 直接把 updated_at/created_at 改到过去（模拟孤儿任务）。"""
-    sets, params = [], {"id": task_id}
-    if updated_minutes is not None:
-        sets.append("updated_at = :uts")
-        params["uts"] = datetime.now(timezone.utc) - timedelta(minutes=updated_minutes)
-    if created_minutes is not None:
-        sets.append("created_at = :cts")
-        params["cts"] = datetime.now(timezone.utc) - timedelta(minutes=created_minutes)
-    with engine.connect() as conn:
-        conn.execute(text(f"UPDATE generation_tasks SET {', '.join(sets)} WHERE id = :id"), params)
-        conn.commit()
-
-
-# ---------------------------------------------------------------------------
-# 1. running 计数（限流"渲染中"数量）
-# ---------------------------------------------------------------------------
-
-
-def test_count_running_by_user_mix_statuses():
-    """count_running_by_user 只统计该用户 running，不含 pending/completed/failed。"""
-    repo, _, _ = _repository()
-    t1 = _make_task(project_id="p1")
-    repo.create(t1)  # pending
-    t2 = _make_task(project_id="p2")
-    repo.create(t2)
-    t2.mark_processing()
-    repo.update(t2)
-    t3 = _make_task(project_id="p3")
-    repo.create(t3)
-    t3.mark_processing()
-    repo.update(t3)
-    t4 = _make_task(project_id="p4")
-    repo.create(t4)
-    t4.mark_processing()
-    repo.update(t4)
-    t4.mark_completed()
-    repo.update(t4)
-    t5 = _make_task(project_id="p5", created_by_user_id="user-2")
-    repo.create(t5)
-    t5.mark_processing()
-    repo.update(t5)
-
-    assert repo.count_running_by_user("user-1") == 2
-    assert repo.count_running_by_user("user-2") == 1
-    assert repo.count_running_total() == 3
-
-
-def test_count_running_total_empty():
-    repo, _, _ = _repository()
-    assert repo.count_running_total() == 0
-    assert repo.count_running_by_user("nobody") == 0
-
-
-def test_preview_tasks_counted_in_running():
-    """预览任务（is_preview=True，工单实测卡 80% 的那种）同样计入 running。"""
-    repo, _, _ = _repository()
-    t = _make_task(is_preview=True)
-    repo.create(t)
-    t.mark_processing()
-    repo.update(t)
-    assert repo.count_running_by_user("user-1") == 1
-    assert repo.count_running_total() == 1
-
-
-# ---------------------------------------------------------------------------
-# 2. 平均耗时估算（429 等待预估依据）
-# ---------------------------------------------------------------------------
-
-
-def _complete_task(repo, engine, task, duration_seconds: float):
-    repo.create(task)
-    task.mark_processing()
-    repo.update(task)
-    task.mark_completed()
-    repo.update(task)
-    now = datetime.now(timezone.utc)
-    with engine.connect() as conn:
-        conn.execute(
-            text("UPDATE generation_tasks SET started_at = :s, completed_at = :c WHERE id = :id"),
-            {"s": now - timedelta(seconds=duration_seconds), "c": now, "id": task.id},
-        )
-        conn.commit()
-
-
-def test_estimate_avg_duration_with_history():
-    """有历史完成任务时返回平均耗时（秒）。"""
-    repo, _, engine = _repository()
-    _complete_task(repo, engine, _make_task(project_id="p1"), 60.0)
-    _complete_task(repo, engine, _make_task(project_id="p2"), 180.0)
-
-    avg = repo.estimate_avg_duration_seconds(default_seconds=120.0)
-    assert 119.0 < avg < 121.0  # (60+180)/2 = 120
-
-
-def test_estimate_avg_duration_no_history_returns_default():
-    """无历史数据时返回默认值。"""
-    repo, _, _ = _repository()
-    assert repo.estimate_avg_duration_seconds(default_seconds=90.0) == 90.0
-
-
-# ---------------------------------------------------------------------------
-# 3. build_rate_limit_detail 结构化提示（前端区分"排队"与"创建失败"）
-# ---------------------------------------------------------------------------
-
-
-def test_user_rate_limit_detail_structure():
-    """429 用户限流：返回 USER_QUEUE_FULL + 排队/渲染数 + 预计等待。"""
-    repo, _, _ = _repository()
-    for i in range(2):  # 2 个渲染中
-        t = _make_task(project_id=f"rp{i}")
-        repo.create(t)
-        t.mark_processing()
-        repo.update(t)
-
-    exc = task_enqueue.UserPendingLimitExceeded(user_id="user-1", pending_count=3, limit=3)
-    detail = task_enqueue.build_rate_limit_detail(exc, repo, scope="user")
-
-    assert detail["code"] == task_enqueue.ERROR_CODE_USER_QUEUE_FULL
-    assert detail["queued_count"] == 3
-    assert detail["running_count"] == 2
-    assert detail["limit"] == 3
-    assert detail["estimated_wait_seconds"] > 0
-    assert "排队" in detail["message"]
-    assert "user-1" not in detail["message"]  # 不泄露内部 ID
-
-
-def test_global_rate_limit_detail_structure():
-    """503 全局繁忙：返回 SYSTEM_QUEUE_FULL。"""
-    repo, _, _ = _repository()
-    exc = task_enqueue.GlobalQueueFull(pending_count=20, limit=20)
-    detail = task_enqueue.build_rate_limit_detail(exc, repo, scope="global")
-
-    assert detail["code"] == task_enqueue.ERROR_CODE_SYSTEM_QUEUE_FULL
-    assert detail["queued_count"] == 20
-    assert detail["limit"] == 20
-    assert detail["estimated_wait_seconds"] > 0
-    assert "系统繁忙" in detail["message"]
-
-
-def test_wait_estimate_uses_concurrency():
-    """等待预估：排队 8 个 / 并发 4 = 2 批 × 平均耗时。"""
-
-    class FakeRepo:
-        def estimate_avg_duration_seconds(self, limit=20, default_seconds=120.0):
-            return 100.0
-
-    wait = task_enqueue._estimate_wait_seconds(8, FakeRepo())
-    assert wait == 200  # ceil(8/4)=2 批 × 100 秒
-
-
-def test_wait_estimate_repo_without_methods_uses_default():
-    """仓储没有新方法（旧 mock/鸭子类型）时用默认 120 秒兜底，不抛错。"""
-
-    class LegacyRepo:
-        """只实现旧接口的仓储（模拟未升级的调用方）。"""
-
-        def count_pending_total(self):
-            return 0
-
-    wait = task_enqueue._estimate_wait_seconds(4, LegacyRepo())
-    assert wait == 120  # ceil(4/4)=1 批 × 120 默认
-
-
-def test_rate_limit_detail_running_count_falls_back_to_zero():
-    """仓储不支持 running 计数时，running_count 优雅降级为 0。"""
-
-    class LegacyRepo:
-        def count_pending_total(self):
-            return 0
-
-    exc = task_enqueue.GlobalQueueFull(pending_count=20, limit=20)
-    detail = task_enqueue.build_rate_limit_detail(exc, LegacyRepo(), scope="global")
-    assert detail["running_count"] == 0
-    assert detail["code"] == task_enqueue.ERROR_CODE_SYSTEM_QUEUE_FULL
-
-
-# ---------------------------------------------------------------------------
-# 4. worker 清理核心：中断任务被重置（worker 重启/超时恢复）
-# ---------------------------------------------------------------------------
-
-
-def test_worker_cleanup_resets_interrupted_running_task():
-    """模拟 worker 重启：running 超 20 分钟无更新的任务被重置为 failed，原因写明。"""
-    repo, _, engine = _repository()
-
-    t = _make_task(is_preview=True)  # 预览任务
-    repo.create(t)
-    t.mark_processing()  # running
-    repo.update(t)
-    _age_task(engine, t.id, updated_minutes=25)  # 25 分钟无进度更新
-
-    cleaned = _startup.cleanup_stale_running_with_session(repo, 20)
-    assert cleaned == 1
-
-    saved = repo.get(t.id)
-    assert saved.status == GenerationTaskStatus.FAILED
-    assert "中断" in saved.error_message
-    assert saved.error_info.get("error_type") == "WorkerInterrupted"
-    assert saved.completed_at is not None
-
-
-def test_worker_cleanup_keeps_healthy_running_task():
-    """正常运行中（5 分钟前有更新）的任务不被误杀。"""
-    repo, _, engine = _repository()
-
-    t = _make_task()
-    repo.create(t)
-    t.mark_processing()
-    repo.update(t)
-    _age_task(engine, t.id, updated_minutes=5)
-
-    assert _startup.cleanup_stale_running_with_session(repo, 20) == 0
-    assert repo.get(t.id).status == GenerationTaskStatus.RUNNING
-
-
-def test_worker_cleanup_resets_stale_pending_task():
-    """卡 pending 超 15 分钟（worker 停止消费）的任务被重置，释放限流名额。"""
-    repo, _, engine = _repository()
-
-    t = _make_task(is_preview=True)
-    repo.create(t)  # 一直 pending
-    _age_task(engine, t.id, created_minutes=20)
-
-    cleaned = _startup.cleanup_stale_pending_with_session(repo, 15)
-    assert cleaned == 1
-
-    saved = repo.get(t.id)
-    assert saved.status == GenerationTaskStatus.FAILED
-    assert saved.error_info.get("error_type") == "PendingTimeout"
-    # 释放名额后 pending 计数归零，新请求不再被 429 误伤
-    assert repo.count_pending_total() == 0
-
-
-def test_worker_cleanup_pending_keeps_recent():
-    """刚创建 3 分钟的 pending 任务不清理。"""
-    repo, _, engine = _repository()
-
-    t = _make_task()
-    repo.create(t)
-    _age_task(engine, t.id, created_minutes=3)
-
-    assert _startup.cleanup_stale_pending_with_session(repo, 15) == 0
-    assert repo.get(t.id).status == GenerationTaskStatus.PENDING
-
-
-def test_worker_cleanup_multiple_orphans_all_reset():
-    """3 个卡死 running 任务（工单实测：3 个预览卡 80% 超 10 小时）全部恢复。"""
-    repo, _, engine = _repository()
-
-    ids = []
-    for i in range(3):
-        t = _make_task(project_id=f"p{i}", is_preview=True)
-        repo.create(t)
-        t.mark_processing()
-        repo.update(t)
-        _age_task(engine, t.id, updated_minutes=600)  # 10 小时
-        ids.append(t.id)
-
-    cleaned = _startup.cleanup_stale_running_with_session(repo, 20)
-    assert cleaned == 3
-    for tid in ids:
-        assert repo.get(tid).status == GenerationTaskStatus.FAILED