fix: 移除片段时长显示 + 点击加号直接添加 (#1696)

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

alembic/versions/064_add_match_count_visual_similarity.py

@@ -0,0 +1,25 @@
+"""add match_count and visual_similarity to generated_videos
+
+Revision ID: 064_match_count_visual_sim
+Revises: 063_fingerprint_chunks
+Create Date: 2026-09-03
+"""
+
+import sqlalchemy as sa
+
+from alembic import op
+
+revision = "064_match_count_visual_sim"
+down_revision = "063_fingerprint_chunks"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    op.add_column("generated_videos", sa.Column("match_count", sa.Integer(), nullable=True, server_default="0"))
+    op.add_column("generated_videos", sa.Column("visual_similarity", sa.Float(), nullable=True, server_default="0.0"))
+
+
+def downgrade() -> None:
+    op.drop_column("generated_videos", "visual_similarity")
+    op.drop_column("generated_videos", "match_count")

alembic/versions/065_add_visual_similarity_match_count_duplication.py

@@ -0,0 +1,25 @@
+"""add visual_similarity and match_count to duplication_records
+
+Revision ID: 065_dup_record_sim_match
+Revises: 064_match_count_visual_sim
+Create Date: 2026-09-04
+"""
+
+import sqlalchemy as sa
+
+from alembic import op
+
+revision = "065_dup_record_sim_match"
+down_revision = "064_match_count_visual_sim"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    op.add_column("duplication_records", sa.Column("visual_similarity", sa.Float(), nullable=True))
+    op.add_column("duplication_records", sa.Column("match_count", sa.Integer(), nullable=True))
+
+
+def downgrade() -> None:
+    op.drop_column("duplication_records", "match_count")
+    op.drop_column("duplication_records", "visual_similarity")

apps/api/app/api/routes/duplication.py

@@ -7,6 +7,7 @@ from typing import Any
 from uuid import uuid4
 
 from app.auth import AuthenticatedUser, get_current_user
+from app.core.celery_app import celery_app
 from app.core.storage import OSSStorageService, get_storage_service
 from app.dependencies import get_duplication_repository
 from app.schemas.duplication import (
@@ -76,6 +77,8 @@ def _to_record_response(record: DuplicationRecord) -> DuplicationRecordResponse:
         status=record.status,
         duplicate_rate=record.duplicate_rate,
         duplicate_count=record.duplicate_count,
+        visual_similarity=getattr(record, "visual_similarity", None),
+        match_count=getattr(record, "match_count", None),
         created_at=record.created_at.isoformat(),
         updated_at=record.updated_at.isoformat(),
     )
@@ -90,6 +93,8 @@ def _to_detail_response(record: DuplicationRecord) -> DuplicationDetailResponse:
         status=record.status,
         duplicate_rate=record.duplicate_rate,
         duplicate_count=record.duplicate_count,
+        visual_similarity=getattr(record, "visual_similarity", None),
+        match_count=getattr(record, "match_count", None),
         created_at=record.created_at.isoformat(),
         updated_at=record.updated_at.isoformat(),
         segments=[
@@ -192,6 +197,8 @@ async def upload_for_duplication(
         authenticated_user.user.id,
     )
 
+    celery_app.send_task("worker.process_duplication_check", args=[record.id])
+
     return DuplicationUploadResponse(
         id=record.id,
         status=record.status,
@@ -296,6 +303,8 @@ def retry_duplication(
             detail=f"查重记录 {record_id} 不存在",
         )
 
+    celery_app.send_task("worker.process_duplication_check", args=[updated.id])
+
     return DuplicationUploadResponse(
         id=updated.id,
         status=updated.status,

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

@@ -92,6 +92,9 @@ def _to_generated_video_response(item, download_url: str | None = None) -> Gener
         height=item.height,
         fps=item.fps,
         download_url=download_url,
+        duplicate_rate=getattr(item, "duplicate_rate", None),
+        visual_similarity=getattr(item, "visual_similarity", None),
+        match_count=getattr(item, "match_count", None),
     )
 
 
@@ -137,7 +140,6 @@ def _select_assets_from_library(
     return [a.id for a in ready_video_assets]
 
 
-
 def _writeback_edit_plan_config(
     plan_id: str,
     task_id: str,
@@ -162,7 +164,7 @@ def _writeback_edit_plan_config(
         current_config = plan_model.config if isinstance(plan_model.config, dict) else {}
         merged = dict(current_config)
         merged["generation_task_id"] = task_id
-        
+
         # 检查标题是否发生变化，如果变化则清除 cover 字段强制重新生成封面
         if title_config:
             old_title_config = merged.get("title_config", {}) or {}
@@ -174,10 +176,12 @@ def _writeback_edit_plan_config(
                     del merged["cover"]
                     logger.info(
                         "[生成任务] 标题变化，清除旧封面: plan_id=%s old_title=%s new_title=%s",
-                        plan_id, old_title_text, new_title_text,
+                        plan_id,
+                        old_title_text,
+                        new_title_text,
                     )
             merged["title_config"] = title_config
-        
+
         plan_model.config = merged
         db.commit()
         logger.info(

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

@@ -15,6 +15,7 @@ from app.schemas.video_center import (
     VideoItemResponse,
 )
 from fastapi import APIRouter, Depends, HTTPException, Query, Response
+from pydantic import BaseModel, Field
 
 from packages.application import (
     GetGeneratedVideoUseCase,
@@ -53,6 +54,8 @@ def _to_video_response(item, storage: OSSStorageService | None = None) -> VideoI
         download_url=download_url,
         generated_at=format_utc_datetime(item.generated_at) if hasattr(item, "generated_at") else "",
         duplicate_rate=getattr(item, "duplicate_rate", None),
+        visual_similarity=getattr(item, "visual_similarity", None),
+        match_count=getattr(item, "match_count", None),
     )
 
 
@@ -237,3 +240,70 @@ def get_batch_download_status(
         status=api_status,
         download_url=download_url,
     )
+
+
+# ── 重新计算查重率 ─────────────────────────────────────────────────
+
+
+class RecomputeDedupRequest(BaseModel):
+    """重新计算查重率请求。"""
+
+    video_ids: list[str] | None = Field(
+        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。",
+    )
+
+
+class RecomputeDedupResponse(BaseModel):
+    """重新计算查重率响应。"""
+
+    enqueued: int = Field(..., description="已入队的任务数量")
+    total_scanned: int = Field(..., description="扫描的视频总数")
+    skipped: int = Field(..., description="已有查重数据跳过的数量")
+    message: str = ""
+
+
+@router.post("/videos/recompute-dedup", response_model=RecomputeDedupResponse)
+def recompute_dedup(
+    request: RecomputeDedupRequest = RecomputeDedupRequest(),
+    repo=Depends(get_generated_video_repository),
+    current_user: AuthenticatedUser = Depends(get_current_user),
+):
+    """重新计算视频的查重率/视觉相似度。
+
+    对于已存在但缺少 duplicate_rate / video_fingerprint 的视频，
+    触发异步 Celery 任务重新下载并计算指纹 + 查重率。
+
+    不传 video_ids 时，对当前用户所有视频进行检查。
+    """
+    user_id = current_user.user.id
+
+    # 获取目标视频列表
+    if request.video_ids:
+        all_videos = repo.get_by_ids(request.video_ids)
+        # 安全校验：只处理当前用户的视频
+        target_videos = [v for v in all_videos if v.user_id == user_id]
+    else:
+        target_videos = repo.list_by_user(user_id)
+
+    total_scanned = len(target_videos)
+    enqueued = 0
+    skipped = 0
+
+    for video in target_videos:
+        # 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)", video.id, user_id)
+
+    return RecomputeDedupResponse(
+        enqueued=enqueued,
+        total_scanned=total_scanned,
+        skipped=skipped,
+        message=f"已入队 {enqueued} 个查重任务" if enqueued > 0 else "所有视频查重数据已完整",
+    )

apps/api/app/schemas/duplication.py

@@ -28,6 +28,9 @@ class DuplicationRecordResponse(BaseModel):
     status: str = "pending"
     duplicate_rate: float | None = None
     duplicate_count: int = 0
+    # #1661 视觉相似度（归一化 0~1）/ 匹配视频数
+    visual_similarity: float | None = None
+    match_count: int | None = None
     created_at: str
     updated_at: str
 

apps/api/app/schemas/generated_video.py

@@ -25,6 +25,10 @@ class GeneratedVideoResponse(BaseModel):
     review_status: str = "pending_review"
     generation_params: dict = Field(default_factory=dict)
     download_url: str | None = None
+    # #1660 查重率（百分比 0~100）/ 视觉相似度（0~1）/ 匹配帧数
+    duplicate_rate: float | None = None
+    visual_similarity: float | None = None
+    match_count: int | None = None
 
 
 class GeneratedVideoDownloadUrlResponse(BaseModel):

apps/api/app/schemas/video_center.py

@@ -22,7 +22,10 @@ class VideoItemResponse(BaseModel):
     generation_params: dict = Field(default_factory=dict)
     download_url: str | None = None
     generated_at: str = ""
+    # #1660 查重率（百分比 0~100）/ 视觉相似度（0~1）/ 匹配帧数
     duplicate_rate: float | None = None
+    visual_similarity: float | None = None
+    match_count: int | None = None
 
 
 class ListVideosResponse(BaseModel):

apps/api/app/services/plan_generator_service.py

@@ -131,6 +131,7 @@ class PlanGeneratorService:
                 editing_mode,
                 random_selection=random_preview,
                 asset_durations=asset_durations,
+                user_id=created_by_user_id,
             )
 
         # 5. 持久化所有 clips 并计算总时长
@@ -218,6 +219,7 @@ class PlanGeneratorService:
         *,
         random_selection: bool = False,
         asset_durations: dict[str, float] | None = None,
+        user_id: str = "",
     ) -> None:
         """按 editing_mode 将素材分配到 clips（就地修改，未持久化）.
 
@@ -239,6 +241,14 @@ class PlanGeneratorService:
                 asset_ids = list(asset_ids)  # 复制避免修改调用方原列表
                 random.shuffle(asset_ids)
 
+        # 查询已有视频的已用区间（跨视频避让）
+        external_used_segments = None
+        if user_id and self._clip_repo:
+            try:
+                external_used_segments = self._clip_repo.list_used_segments_by_user(user_id, limit_recent=50)
+            except Exception:
+                logger.warning("跨视频避让查询失败，回退到纯随机", exc_info=True)
+
         distribute_assets(
             clips,
             asset_ids,
@@ -246,6 +256,7 @@ class PlanGeneratorService:
             random_selection=random_selection,
             asset_durations=asset_durations,
             asset_scene_points=asset_scene_points,
+            external_used_segments=external_used_segments,
         )
 
     def _fetch_asset_scene_points(self, asset_ids: List[str]) -> dict[str, list[float]]:

apps/web/src/api/duplication/types.ts

@@ -20,6 +20,10 @@ export interface DuplicationRecord {
   duplicate_rate?: number
   /** 重复片段数 */
   duplicate_count?: number
+  /** 视觉相似度（0-100），#1660 新增 */
+  visual_similarity?: number
+  /** 匹配帧数，#1660 新增 */
+  match_count?: number
   /** 创建时间 */
   created_at: string
   /** 更新时间 */

apps/web/src/api/products/index.ts

@@ -13,6 +13,8 @@ export type {
   VideoItem,
 } from "./types"
 
+export type { RecomputeDedupResponse } from "./products"
+
 // 工具函数
 export { mapVideoToProductItem } from "./utils"
 
@@ -25,4 +27,5 @@ export {
   updateReviewStatus,
   batchDownload,
   getBatchDownloadStatus,
+  recomputeDedup,
 } from "./products"

apps/web/src/api/products/products.ts

@@ -78,3 +78,18 @@ export const getBatchDownloadStatus = async (jobId: string): Promise<BatchDownlo
   console.warn("[getBatchDownloadStatus] 后端暂无批量下载状态端点", jobId)
   return { job_id: jobId, status: "processing", progress: 0 }
 }
+
+/** 重新计算存量视频查重率（异步） */
+export interface RecomputeDedupResponse {
+  enqueued: number
+  total_scanned: number
+  skipped: number
+  message: string
+}
+
+export const recomputeDedup = async (videoIds?: string[]): Promise<RecomputeDedupResponse> => {
+  const response = await apiClient.post("/videos/recompute-dedup", {
+    video_ids: videoIds,
+  })
+  return response.data
+}

apps/web/src/api/products/types.ts

@@ -23,6 +23,10 @@ export interface ProductItem {
   project_name?: string
   /** 查重率（百分比） */
   duplicate_rate?: number
+  /** 视觉相似度（0-1），#1660 新增 */
+  visual_similarity?: number
+  /** 匹配帧数，#1660 新增 */
+  match_count?: number
   created_at?: string
   updated_at?: string
 }
@@ -72,4 +76,8 @@ export interface VideoItem {
   download_url: string
   generated_at: string
   duplicate_rate?: number
+  /** 视觉相似度（0-1），#1660 新增 */
+  visual_similarity?: number
+  /** 匹配帧数，#1660 新增 */
+  match_count?: number
 }

apps/web/src/api/products/utils.ts

@@ -30,5 +30,7 @@ export function mapVideoToProductItem(video: VideoItem): ProductItem {
     created_at: video.generated_at,
     updated_at: video.generated_at,
     duplicate_rate: video.duplicate_rate,
+    visual_similarity: video.visual_similarity,
+    match_count: video.match_count,
   }
 }

apps/web/src/api/tts/jobs.ts

@@ -81,6 +81,7 @@ export const extractVideoVoice = async (
 ): Promise<{ asset_id: string; duration: number }> => {
   const formData = new FormData()
   formData.append("file", file)
+  formData.append("project_id", "default")
 
   return new Promise((resolve, reject) => {
     const xhr = new XMLHttpRequest()

apps/web/src/pages/duplication/DuplicationDetail.tsx

@@ -98,7 +98,7 @@ const DuplicationDetail: React.FC = () => {
       <div className="dup-detail-grid">
         <RiskCard riskLevel={riskLevel} similarityPercent={similarityPercent} />
         <InfoCard detail={detail} />
-        <SegmentsSection segments={detail.segments} />
+        <SegmentsSection segments={detail.segments} totalDuration={detail.duration_seconds} />
       </div>
     </div>
   )

apps/web/src/pages/duplication/components/ResultCard.tsx

@@ -1,7 +1,7 @@
 import React from "react"
 import { Button, Tag, Tooltip } from "@/components/ui"
 import type { DuplicationRecord } from "@/api/duplication"
-import { STATUS_CONFIG } from "../constants"
+import { STATUS_CONFIG, RISK_TAG_VARIANT, RISK_LABELS } from "../constants"
 import { getRiskLevel, formatSize, formatDuration } from "../utils"
 
 interface ResultCardProps {
@@ -54,6 +54,9 @@ const ResultCard: React.FC<ResultCardProps> = ({ record, onView, onDelete, onRet
               />
             </div>
             <span className={`dup-score-value ${riskLevel}`}>{rateValue.toFixed(1)}%</span>
+            <Tag variant={RISK_TAG_VARIANT[riskLevel]} className="dup-score-risk-tag">
+              {RISK_LABELS[riskLevel]}
+            </Tag>
           </>
         ) : record.status === "failed" ? (
           <Tooltip title="重新查重">

apps/web/src/pages/duplication/components/SegmentsSection.tsx

@@ -2,34 +2,81 @@ import React from "react"
 import { Tag } from "@/components/ui"
 import type { DuplicateSegment } from "@/api/duplication"
 import { SegmentCard } from "./SegmentCard"
+import { formatTime } from "../utils"
 
 interface SegmentsSectionProps {
   segments?: DuplicateSegment[]
+  /** 视频总时长（秒），用于渲染时间轴 */
+  totalDuration?: number
+}
+
+/** 片段相似度 → 风险等级（时间轴配色用） */
+const getSegmentRisk = (similarity: number): "low" | "medium" | "high" => {
+  if (similarity >= 90) return "high"
+  if (similarity >= 70) return "medium"
+  return "low"
 }
 
 /**
- * 重复片段列表区域
+ * 重复片段列表区域（含时间轴可视化）
  */
-export const SegmentsSection: React.FC<SegmentsSectionProps> = ({ segments = [] }) => (
-  <div className="dup-checks-section">
-    <h3>
-      🔍 重复片段详情
-      <Tag variant="primary" style={{ marginLeft: 8 }}>
-        {segments.length} 个片段
-      </Tag>
-    </h3>
+export const SegmentsSection: React.FC<SegmentsSectionProps> = ({
+  segments = [],
+  totalDuration,
+}) => {
+  const showTimeline = segments.length > 0 && totalDuration !== undefined && totalDuration > 0
 
-    {segments.length > 0 ? (
-      <div className="dup-checks-list">
-        {segments.map((segment, index) => (
-          <SegmentCard key={segment.id} segment={segment} index={index} />
-        ))}
-      </div>
-    ) : (
-      <div className="dup-results-empty" style={{ padding: "32px 0" }}>
-        <div className="dup-results-empty-icon">🎉</div>
-        <p>未发现重复片段，内容原创度很高</p>
-      </div>
-    )}
-  </div>
-)
+  return (
+    <div className="dup-checks-section">
+      <h3>
+        🔍 重复片段详情
+        <Tag variant="primary" style={{ marginLeft: 8 }}>
+          {segments.length} 个片段
+        </Tag>
+      </h3>
+
+      {showTimeline && (
+        <div className="dup-timeline">
+          <div className="dup-timeline-bar">
+            {segments.map((seg, i) => {
+              const left = (seg.source_start / totalDuration) * 100
+              const width = Math.max(
+                ((seg.source_end - seg.source_start) / totalDuration) * 100,
+                0.5,
+              )
+              const segRisk = getSegmentRisk(seg.similarity)
+              return (
+                <div
+                  key={seg.id ?? i}
+                  className={`dup-timeline-segment ${segRisk}`}
+                  style={{
+                    left: `${Math.min(left, 100)}%`,
+                    width: `${Math.min(width, 100 - Math.min(left, 100))}%`,
+                  }}
+                  title={`${formatTime(seg.source_start)} - ${formatTime(seg.source_end)} · 相似度 ${seg.similarity.toFixed(0)}% · ${seg.matched_video_name}`}
+                />
+              )
+            })}
+          </div>
+          <div className="dup-timeline-labels">
+            <span>0s</span>
+            <span>{formatTime(totalDuration ?? 0)}</span>
+          </div>
+        </div>
+      )}
+
+      {segments.length > 0 ? (
+        <div className="dup-checks-list">
+          {segments.map((segment, index) => (
+            <SegmentCard key={segment.id} segment={segment} index={index} />
+          ))}
+        </div>
+      ) : (
+        <div className="dup-results-empty" style={{ padding: "32px 0" }}>
+          <div className="dup-results-empty-icon">🎉</div>
+          <p>未发现重复片段，内容原创度很高</p>
+        </div>
+      )}
+    </div>
+  )
+}

apps/web/src/pages/duplication/duplication.css

@@ -831,3 +831,61 @@
     font-size: 16px;
   }
 }
+
+/* ============================================================
+   查重率风险标签（列表卡片）
+   ============================================================ */
+.dup-score-risk-tag {
+  flex-shrink: 0;
+  margin-left: 2px;
+}
+
+/* ============================================================
+   重复片段时间轴可视化（#1662）
+   ============================================================ */
+.dup-timeline {
+  margin: 16px 0;
+  padding: 0 8px;
+}
+
+.dup-timeline-bar {
+  position: relative;
+  height: 24px;
+  background: var(--bg-secondary, #f1f5f9);
+  border-radius: 4px;
+  overflow: hidden;
+}
+
+.dup-timeline-segment {
+  position: absolute;
+  top: 2px;
+  height: 20px;
+  border-radius: 3px;
+  opacity: 0.8;
+  cursor: pointer;
+  transition: opacity 0.2s;
+}
+
+.dup-timeline-segment:hover {
+  opacity: 1;
+}
+
+.dup-timeline-segment.low {
+  background: #22c55e;
+}
+
+.dup-timeline-segment.medium {
+  background: #f59e0b;
+}
+
+.dup-timeline-segment.high {
+  background: #ef4444;
+}
+
+.dup-timeline-labels {
+  display: flex;
+  justify-content: space-between;
+  font-size: 12px;
+  color: var(--text-secondary);
+  margin-top: 4px;
+}

apps/web/src/pages/duplication/utils.ts

@@ -1,9 +1,9 @@
 /** 根据查重率获取风险等级 */
 export const getRiskLevel = (rate?: number): "low" | "medium" | "high" => {
   if (rate === undefined) return "low"
-  if (rate <= 10) return "low"
-  if (rate <= 30) return "medium"
-  return "high"
+  if (rate < 15) return "low" // <15% 绿色（安全）
+  if (rate <= 30) return "medium" // 15-30% 黄色（注意）
+  return "high" // >30% 红色（危险）
 }
 
 /** 格式化时间（秒 → mm:ss） */

apps/web/src/pages/editing-planner/components/EditorClipList.tsx

@@ -40,13 +40,6 @@ const clipTypeLabel: Record<ClipType | string, string> = {
   pip: "混剪",
 }
 
-const formatDuration = (sec: number) => {
-  if (sec < 60) return `${sec.toFixed(1)}s`
-  const m = Math.floor(sec / 60)
-  const s = (sec % 60).toFixed(0)
-  return `${m}m${s.padStart(2, "0")}s`
-}
-
 const EditorClipList: React.FC<EditorClipListProps> = ({
   clips,
   selectedClipId,
@@ -102,7 +95,6 @@ const EditorClipList: React.FC<EditorClipListProps> = ({
                   {clipTypeLabel[clip.type] || "片段"}
                 </span>
               </span>
-              <span className="ep-clip-item-duration">{formatDuration(clip.duration)}</span>
             </div>
 
             {/* 文案预览 */}

apps/web/src/pages/editing-planner/components/PreviewPlayer.tsx

@@ -87,7 +87,7 @@ const PreviewPlayer: React.FC<PreviewPlayerProps> = ({
                 WebkitTextStroke: "1px rgba(0,0,0,0.6)",
                 top:
                   titleConfig.position === "top"
-                    ? "8px"
+                    ? "6.25%"
                     : titleConfig.position === "center"
                       ? "50%"
                       : "auto",

apps/web/src/pages/editing-planner/components/TimelinePanel.tsx

@@ -7,7 +7,6 @@
  * - ClipCard - 片段卡片
  * - ClipTrack - 片段轨道（播放头+片段列表+添加卡片）
  * - TimelineHeader - 时间线头部（标题+缩放+操作按钮）
- * - AddClipPicker - 添加片段选择器
  * - TrimPreview - 裁剪预览 tooltip
  * - ContextMenu - 右键菜单
  *
@@ -27,7 +26,6 @@ import { usePlayheadDrag } from "./timeline/hooks/usePlayheadDrag"
 import { TimeRuler } from "./timeline/TimeRuler"
 import { ClipTrack } from "./timeline/ClipTrack"
 import { TimelineHeader } from "./timeline/TimelineHeader"
-import { AddClipPicker } from "./timeline/AddClipPicker"
 import { TrimPreview } from "./timeline/TrimPreview"
 import { ContextMenu } from "./timeline/ContextMenu"
 
@@ -100,16 +98,7 @@ const TimelinePanel: React.FC<TimelinePanelProps> = ({
     handleContextSplit,
     handleContextResetTrim,
     handleContextDelete,
-    showAddPicker,
-    pickerRef,
     addCardRef,
-    pickerPos,
-    availableTypes,
-    addType,
-    addDuration,
-    setAddType,
-    setAddDuration,
-    handleTogglePicker,
     handleConfirmAdd,
     hoveredClipId,
     setHoveredClipId,
@@ -177,7 +166,7 @@ const TimelinePanel: React.FC<TimelinePanelProps> = ({
           onClipMouseLeave={() => setHoveredClipId(null)}
           onTrimHandleMouseDown={handleTrimHandleMouseDown}
           onClipRemove={onClipRemove}
-          onTogglePicker={handleTogglePicker}
+          onTogglePicker={handleConfirmAdd}
         />
       )}
 
@@ -204,20 +193,6 @@ const TimelinePanel: React.FC<TimelinePanelProps> = ({
           onDelete={handleContextDelete}
         />
       )}
-
-      {/* 类型+时长选择面板 */}
-      {showAddPicker && (
-        <AddClipPicker
-          pickerRef={pickerRef}
-          position={pickerPos}
-          availableTypes={availableTypes}
-          addType={addType}
-          addDuration={addDuration}
-          onTypeChange={setAddType}
-          onDurationChange={setAddDuration}
-          onConfirm={handleConfirmAdd}
-        />
-      )}
     </div>
   )
 }

apps/web/src/pages/editing-planner/components/timeline/AddClipPicker.tsx

@@ -7,12 +7,8 @@ interface AddClipPickerProps {
   position: { top: number; right: number }
   availableTypes: ClipType[]
   addType: ClipType
-  addDuration: number
   onTypeChange: (type: ClipType) => void
-  onDurationChange: (duration: number) => void
   onConfirm: () => void
-  minDuration?: number
-  maxDuration?: number
 }
 
 export const AddClipPicker: React.FC<AddClipPickerProps> = ({
@@ -20,12 +16,8 @@ export const AddClipPicker: React.FC<AddClipPickerProps> = ({
   position,
   availableTypes,
   addType,
-  addDuration,
   onTypeChange,
-  onDurationChange,
   onConfirm,
-  minDuration = 1,
-  maxDuration = 120,
 }) => {
   return (
     <div
@@ -53,24 +45,6 @@ export const AddClipPicker: React.FC<AddClipPickerProps> = ({
         ))}
       </div>
 
-      {/* 时长输入 */}
-      <div className="ep-add-clip-duration-row">
-        <span className="ep-add-clip-type-label">时长:</span>
-        <input
-          type="number"
-          className="ep-duration-input"
-          min={minDuration}
-          max={maxDuration}
-          value={addDuration}
-          onChange={(e) =>
-            onDurationChange(
-              Math.max(minDuration, Math.min(maxDuration, Number(e.target.value) || minDuration)),
-            )
-          }
-        />
-        <span className="ep-add-clip-duration-unit">秒</span>
-      </div>
-
       {/* 确认按钮 */}
       <button className="ep-add-clip-confirm-btn" onClick={onConfirm}>
         添加

apps/web/src/pages/editing-planner/components/timeline/ClipCard.tsx

@@ -105,14 +105,11 @@ export const ClipCard: React.FC<ClipCardProps> = ({
           <span className="ep-clip-name">
             {CLIP_TYPE_LABELS[clip.type] || "片段"} {idx + 1}
           </span>
-          <span className="ep-clip-duration">
-            {clip.duration}s
-            {hasTrim && (
-              <span className="ep-trim-indicator" title="已裁剪">
-                ✂
-              </span>
-            )}
-          </span>
+          {hasTrim && (
+            <span className="ep-trim-indicator" title="已裁剪">
+              ✂
+            </span>
+          )}
         </div>
 
         {/* 速度徽章 */}

apps/web/src/pages/editing-planner/hooks/timeline-menus/useAddPicker.ts

@@ -26,8 +26,6 @@ export function useAddPicker({ currentMode, onAddClip }: UseAddPickerOptions) {
   }, [currentMode])
 
   const [addType, setAddType] = useState<ClipType>(defaultAddType)
-  const [addDuration, setAddDuration] = useState<number>(DEFAULT_ADD_DURATION)
-
   useEffect(() => {
     if (!availableTypes.includes(addType)) {
       setAddType(defaultAddType)
@@ -95,9 +93,9 @@ export function useAddPicker({ currentMode, onAddClip }: UseAddPickerOptions) {
   }, [showAddPicker])
 
   const handleConfirmAdd = useCallback(() => {
-    onAddClip(addType, addDuration)
+    onAddClip(addType, DEFAULT_ADD_DURATION)
     setShowAddPicker(false)
-  }, [onAddClip, addType, addDuration])
+  }, [onAddClip, addType])
 
   return {
     showAddPicker,
@@ -107,9 +105,8 @@ export function useAddPicker({ currentMode, onAddClip }: UseAddPickerOptions) {
     pickerPos,
     availableTypes,
     addType,
-    addDuration,
+    addDuration: DEFAULT_ADD_DURATION,
     setAddType,
-    setAddDuration,
     handleTogglePicker,
     handleConfirmAdd,
   }

apps/web/src/pages/editing-planner/hooks/useTimelineMenus.ts

@@ -35,7 +35,6 @@ export const useTimelineMenus = (
     addType,
     addDuration,
     setAddType,
-    setAddDuration,
     handleTogglePicker,
     handleConfirmAdd,
   } = useAddPicker({ currentMode, onAddClip })
@@ -57,7 +56,6 @@ export const useTimelineMenus = (
     addType,
     addDuration,
     setAddType,
-    setAddDuration,
     handleTogglePicker,
     handleConfirmAdd,
     // 悬停状态

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

@@ -591,6 +591,8 @@ const FrontendPreviewPlayer: React.FC<FrontendPreviewPlayerProps> = ({
           <div
             style={{
               position: "absolute",
+              width: `${100 - 2 * titleSidePct}%`,
+              maxWidth: `${100 - 2 * titleSidePct}%`,
               ...(customTitleXPct != null && customTitleYPct != null
                 ? {
                     left: `${customTitleXPct}%`,
@@ -599,13 +601,13 @@ const FrontendPreviewPlayer: React.FC<FrontendPreviewPlayerProps> = ({
                     textAlign: "center" as const,
                   }
                 : {
-                    left: `${titleSidePct}%`,
-                    right: `${titleSidePct}%`,
+                    left: "50%",
+                    transform: "translateX(-50%)",
                     textAlign: "center" as const,
                     ...(titleSettings.position === "top"
                       ? { top: `${titleTopPct}%` }
                       : titleSettings.position === "center"
-                        ? { top: "50%", transform: "translateY(-50%)" }
+                        ? { top: "50%", transform: "translate(-50%, -50%)" }
                         : { bottom: `${titleBottomPct}%` }),
                   }),
               pointerEvents: "auto",

apps/web/src/pages/generate/components/Step1TemplateSelect.tsx

@@ -47,9 +47,7 @@ const Step1TemplateSelect: React.FC<Step1TemplateSelectProps> = (props) => {
                 🎬
               </div>
               <h4>{tpl.name}</h4>
-              <p>
-                {tpl.estimated_duration}s · {tpl.segments.length}片段
-              </p>
+              <p>{tpl.segments.length}片段</p>
               {tpl.tags.length > 0 && (
                 <div
                   style={{

apps/web/src/pages/generate/components/Step5VoiceSelect.tsx

@@ -1,12 +1,12 @@
 /**
  * Step 5 配音选择组件
  * 展示用户已上传的配音素材，支持选中、预览播放
+ * 根据视频总时长自动过滤时长差异过大的配音（±10%以内）
  */
 import React, { useState, useRef, useCallback } from "react"
 import { useNavigate } from "react-router-dom"
 import { useQuery } from "@tanstack/react-query"
-import { AudioOutlined, SoundOutlined, WarningOutlined } from "@ant-design/icons"
-import { Modal } from "antd"
+import { AudioOutlined, SoundOutlined } from "@ant-design/icons"
 import { getAssetsByKind } from "@/api/assets"
 import type { AssetItem } from "@/api/assets"
 
@@ -16,7 +16,6 @@ interface Step5VoiceSelectProps {
   totalVideoDuration?: number
 }
 
-/** 格式化时长 mm:ss */
 /** 获取素材实际时长（优先顶层 duration，fallback 到 metadata.duration） */
 const getDuration = (item: AssetItem): number => {
   return item.duration ?? (item.metadata?.duration as number) ?? 0
@@ -34,13 +33,6 @@ const isAiVoice = (item: AssetItem): boolean => {
   return (!duration || duration <= 0) && (!size || size <= 0)
 }
 
-const formatDuration = (seconds?: number): string => {
-  if (!seconds || seconds <= 0) return "00:00"
-  const m = Math.floor(seconds / 60)
-  const s = Math.floor(seconds % 60)
-  return `${String(m).padStart(2, "0")}:${String(s).padStart(2, "0")}`
-}
-
 /** 格式化文件大小 */
 const formatFileSize = (bytes?: number): string => {
   if (!bytes || bytes <= 0) return "未知"
@@ -58,8 +50,6 @@ const Step5VoiceSelect: React.FC<Step5VoiceSelectProps> = ({
   const navigate = useNavigate()
   const [playingId, setPlayingId] = useState<string | null>(null)
   const audioRef = useRef<HTMLAudioElement | null>(null)
-  const [durationWarningOpen, setDurationWarningOpen] = useState(false)
-  const [pendingVoiceId, setPendingVoiceId] = useState<string | null>(null)
 
   // 获取用户上传的配音素材
   const { data: materials = [], isLoading } = useQuery({
@@ -67,6 +57,19 @@ const Step5VoiceSelect: React.FC<Step5VoiceSelectProps> = ({
     queryFn: () => getAssetsByKind("voice", { limit: 50 }),
   })
 
+  // 根据视频总时长过滤配音：只保留时长在 ±10% 以内的素材，AI 音色始终展示
+  const filteredMaterials = React.useMemo(() => {
+    if (totalVideoDuration <= 0) return materials
+    return materials.filter((item) => {
+      // AI 音色没有固定时长，始终保留
+      if (isAiVoice(item)) return true
+      const duration = getDuration(item)
+      if (duration <= 0) return true
+      const ratio = duration / totalVideoDuration
+      return ratio >= 0.9 && ratio <= 1.1
+    })
+  }, [materials, totalVideoDuration])
+
   /** 切换播放/暂停 */
   const togglePlay = useCallback(
     (material: AssetItem) => {
@@ -100,38 +103,14 @@ const Step5VoiceSelect: React.FC<Step5VoiceSelectProps> = ({
     [playingId],
   )
 
-  /** 选中素材（含时长校验） */
+  /** 选中素材（直接选中，不再做时长校验弹窗） */
   const handleSelect = useCallback(
     (id: string) => {
-      // 如果启用了时长校验，且配音时长不足（AI 音色按脚本实时合成，不参与时长校验）
-      if (totalVideoDuration > 0) {
-        const material = materials.find((m) => m.id === id)
-        if (material && !isAiVoice(material) && getDuration(material) < totalVideoDuration) {
-          setPendingVoiceId(id)
-          setDurationWarningOpen(true)
-          return
-        }
-      }
       onSelectedVoiceChange(id)
     },
-    [onSelectedVoiceChange, totalVideoDuration, materials],
+    [onSelectedVoiceChange],
   )
 
-  /** 确认使用时长不足的配音 */
-  const handleConfirmUseAnyway = useCallback(() => {
-    if (pendingVoiceId) {
-      onSelectedVoiceChange(pendingVoiceId)
-    }
-    setDurationWarningOpen(false)
-    setPendingVoiceId(null)
-  }, [pendingVoiceId, onSelectedVoiceChange])
-
-  /** 取消选择 */
-  const handleCancelSelection = useCallback(() => {
-    setDurationWarningOpen(false)
-    setPendingVoiceId(null)
-  }, [])
-
   /** 跳转到配音库上传 */
   const handleGoToUpload = useCallback(() => {
     navigate("/app/voices?tab=material&upload=1")
@@ -196,7 +175,7 @@ const Step5VoiceSelect: React.FC<Step5VoiceSelectProps> = ({
           gap: 12,
         }}
       >
-        {materials.map((item) => {
+        {filteredMaterials.map((item) => {
           const isSelected = selectedVoice === item.id
           const isPlaying = playingId === item.id
 
@@ -277,7 +256,7 @@ const Step5VoiceSelect: React.FC<Step5VoiceSelectProps> = ({
                 {item.name}
               </div>
 
-              {/* 时长 + 大小 */}
+              {/* 文件大小 */}
               <div
                 style={{
                   display: "flex",
@@ -289,65 +268,18 @@ const Step5VoiceSelect: React.FC<Step5VoiceSelectProps> = ({
               >
                 {isAiVoice(item) ? (
                   <span style={{ color: "#1677ff", fontWeight: 500 }}>AI 音色</span>
-                ) : (
-                  <span style={{ display: "flex", alignItems: "center", gap: 4 }}>
-                    {formatDuration(getDuration(item))}
-                    {totalVideoDuration > 0 && getDuration(item) < Number(totalVideoDuration) && (
-                      <span
-                        style={{
-                          color: "#ff4d4f",
-                          fontSize: 11,
-                          fontWeight: 500,
-                          display: "inline-flex",
-                          alignItems: "center",
-                          gap: 2,
-                        }}
-                      >
-                        <WarningOutlined />
-                        时长不足
-                      </span>
-                    )}
-                  </span>
-                )}
+                ) : null}
                 <span>{isAiVoice(item) ? "按文本合成" : formatFileSize(getFileSize(item))}</span>
               </div>
             </div>
           )
         })}
       </div>
-
-      {/* 时长不足警告弹窗 */}
-      <Modal
-        title={
-          <span style={{ display: "flex", alignItems: "center", gap: 8 }}>
-            <WarningOutlined style={{ color: "#faad14" }} />
-            配音时长不足
-          </span>
-        }
-        open={durationWarningOpen}
-        onOk={handleConfirmUseAnyway}
-        onCancel={handleCancelSelection}
-        okText="仍要使用"
-        cancelText="重新选择"
-        okButtonProps={{ danger: true }}
-      >
-        {(() => {
-          const pendingMaterial = pendingVoiceId
-            ? materials.find((m) => m.id === pendingVoiceId)
-            : null
-          return (
-            <p>
-              该配音时长（
-              <strong>
-                {pendingMaterial ? formatDuration(getDuration(pendingMaterial)) : "--"}
-              </strong>
-              ）短于视频总时长（
-              <strong>{formatDuration(totalVideoDuration)}</strong>
-              ），播放时配音可能提前结束，建议选择更长的配音素材。
-            </p>
-          )
-        })()}
-      </Modal>
+      {totalVideoDuration > 0 && filteredMaterials.length < materials.length && (
+        <p style={{ color: "#999", fontSize: 12, marginTop: 8 }}>
+          已根据视频时长（{Math.round(totalVideoDuration)}s）自动过滤时长差异过大的配音
+        </p>
+      )}
     </div>
   )
 }

apps/web/src/pages/products/ProductLibrary.tsx

@@ -11,7 +11,7 @@
  * 产品卡片 → components/ProductCard（内联视频播放）
  */
 import React from "react"
-import { VideoCameraOutlined, DownloadOutlined } from "@ant-design/icons"
+import { VideoCameraOutlined, DownloadOutlined, ReloadOutlined } from "@ant-design/icons"
 import { Button } from "@/components/ui"
 import { ProductCard } from "./components/ProductCard"
 import { ProductFilterBar } from "./components/ProductFilterBar"
@@ -19,6 +19,7 @@ import { ProductBatchBar } from "./components/ProductBatchBar"
 import { ProductEmptyState } from "./components/ProductEmptyState"
 import { useProductList } from "./hooks/useProductList"
 import { useProductActions } from "./hooks/useProductActions"
+import { useRecomputeDedup } from "./hooks/product-actions/useRecomputeDedup"
 import "./products.css"
 
 const ProductLibrary: React.FC = () => {
@@ -67,6 +68,8 @@ const ProductLibrary: React.FC = () => {
     setPlayingProduct: () => {}, // 不再使用弹窗播放
   })
 
+  const { recomputeDedup, isRecomputing } = useRecomputeDedup()
+
   // ── Loading 状态 ──
   if (isLoading) {
     return <ProductEmptyState type="loading" />
@@ -94,6 +97,15 @@ const ProductLibrary: React.FC = () => {
           <Button buttonType="ghost" buttonSize="sm" icon={<DownloadOutlined />}>
             批量导出
           </Button>
+          <Button
+            buttonType="ghost"
+            buttonSize="sm"
+            icon={<ReloadOutlined />}
+            loading={isRecomputing}
+            onClick={recomputeDedup}
+          >
+            重新查重
+          </Button>
         </div>
       </div>
 

apps/web/src/pages/products/components/ProductInfoPanel.tsx

@@ -2,6 +2,7 @@ import React from "react"
 import type { ProductItem } from "../../../api/products"
 import { STATUS_MAP } from "../constants"
 import { formatDuration, formatFileSize, formatDate } from "../detailUtils"
+import { getRiskLevel } from "../../duplication/utils"
 
 interface ProductInfoPanelProps {
   product: ProductItem
@@ -44,12 +45,28 @@ export const ProductInfoPanel: React.FC<ProductInfoPanelProps> = ({ product }) =
         </div>
         <div className="xx-detail-meta-item">
           <span className="xx-detail-meta-label">查重率</span>
-          <span className="xx-detail-meta-value">
+          <span
+            className={`xx-detail-meta-value dup-risk-text dup-risk-${getRiskLevel(product.duplicate_rate)}`}
+          >
             {(product.duplicate_rate ?? 0) > 0
               ? `${(product.duplicate_rate ?? 0).toFixed(1)}%`
               : "-"}
           </span>
         </div>
+        {product.visual_similarity != null && (
+          <div className="xx-detail-meta-item">
+            <span className="xx-detail-meta-label">视觉相似度</span>
+            <span className="xx-detail-meta-value">
+              {(product.visual_similarity * 100).toFixed(1)}%
+            </span>
+          </div>
+        )}
+        {product.match_count != null && (
+          <div className="xx-detail-meta-item">
+            <span className="xx-detail-meta-label">匹配帧数</span>
+            <span className="xx-detail-meta-value">{product.match_count}</span>
+          </div>
+        )}
         <div className="xx-detail-meta-item">
           <span className="xx-detail-meta-label">创建时间</span>
           <span className="xx-detail-meta-value">{formatDate(product.created_at ?? "")}</span>

apps/web/src/pages/products/hooks/product-actions/useRecomputeDedup.ts

@@ -0,0 +1,27 @@
+import { useMutation, useQueryClient } from "@tanstack/react-query"
+import { message } from "antd"
+import { recomputeDedup } from "@/api/products"
+
+export function useRecomputeDedup() {
+  const queryClient = useQueryClient()
+
+  const mutation = useMutation({
+    mutationFn: () => recomputeDedup(),
+    onSuccess: (data) => {
+      queryClient.invalidateQueries({ queryKey: ["products"] })
+      if (data.enqueued > 0) {
+        message.success(`已提交 ${data.enqueued} 个视频的查重任务，后台处理中`)
+      } else {
+        message.info("所有视频查重率已是最新，无需重算")
+      }
+    },
+    onError: () => {
+      message.error("查重任务提交失败，请稍后重试")
+    },
+  })
+
+  return {
+    recomputeDedup: () => mutation.mutate(),
+    isRecomputing: mutation.isPending,
+  }
+}

apps/web/src/pages/products/products.css

@@ -1076,3 +1076,19 @@
     gap: var(--space-sm);
   }
 }
+
+/* 查重率风险颜色（#1662） */
+.xx-detail-meta-value.dup-risk-low {
+  color: var(--success-color, #22c55e);
+  font-weight: 600;
+}
+
+.xx-detail-meta-value.dup-risk-medium {
+  color: var(--warning-color, #f59e0b);
+  font-weight: 600;
+}
+
+.xx-detail-meta-value.dup-risk-high {
+  color: var(--error-color, #ef4444);
+  font-weight: 600;
+}

apps/web/src/pages/voices/components/VideoExtractModal.tsx

@@ -30,6 +30,7 @@ const VideoExtractModal: React.FC<VideoExtractModalProps> = ({
       title={<span style={{ fontSize: 16, fontWeight: 600 }}>提取视频配音</span>}
       open={open}
       onCancel={() => {
+        if (inputRef.current) inputRef.current.value = ""
         if (isExtracting) return
         onClose()
       }}
@@ -152,7 +153,7 @@ const VideoExtractModal: React.FC<VideoExtractModalProps> = ({
 
           {isExtracting && (
             <p style={{ textAlign: "center", fontSize: 13, color: "#7c3aed", margin: "12px 0 0" }}>
-              {progress === 100 ? "正在提取人声，请稍候..." : "正在上传视频..."}
+              {"正在提取音频，请稍后..."}
             </p>
           )}
         </div>

apps/web/src/test/pages/ProductLibrary.test.tsx

@@ -102,6 +102,7 @@ vi.mock("@ant-design/icons", () => ({
   SearchOutlined: () => <span />,
   ShareAltOutlined: () => <span />,
   VideoCameraOutlined: () => <span />,
+  ReloadOutlined: () => <span />,
 }))
 
 vi.mock("@/store/authStore", () => ({
@@ -116,6 +117,9 @@ vi.mock("@/api/products", () => ({
   updateReviewStatus: vi.fn().mockResolvedValue({ items: [], total: 0, success: true }),
   batchDownload: vi.fn().mockResolvedValue({ items: [], total: 0, success: true }),
   getBatchDownloadStatus: vi.fn().mockResolvedValue({ items: [], total: 0, success: true }),
+  recomputeDedup: vi
+    .fn()
+    .mockResolvedValue({ enqueued: 0, total_scanned: 0, skipped: 0, message: "" }),
 }))
 
 vi.mock("@/pages/products/ProductLibrary.css", () => ({}))

apps/web/src/test/pages/duplication/utils.test.ts

@@ -0,0 +1,29 @@
+import { describe, it, expect } from "vitest"
+import { getRiskLevel } from "@/pages/duplication/utils"
+
+describe("getRiskLevel (#1662 阈值 <15 / 15-30 / >30)", () => {
+  it("undefined 返回 low（兼容无数据）", () => {
+    expect(getRiskLevel(undefined)).toBe("low")
+  })
+
+  it("<15% 为低风险", () => {
+    expect(getRiskLevel(0)).toBe("low")
+    expect(getRiskLevel(10)).toBe("low")
+    expect(getRiskLevel(14.9)).toBe("low")
+  })
+
+  it("15% 边界为中风险", () => {
+    expect(getRiskLevel(15)).toBe("medium")
+  })
+
+  it("15-30% 为中风险", () => {
+    expect(getRiskLevel(20)).toBe("medium")
+    expect(getRiskLevel(30)).toBe("medium")
+  })
+
+  it(">30% 为高风险", () => {
+    expect(getRiskLevel(30.1)).toBe("high")
+    expect(getRiskLevel(80)).toBe("high")
+    expect(getRiskLevel(100)).toBe("high")
+  })
+})

apps/worker/video_processing/dedup.py

@@ -1,8 +1,13 @@
-"""Video deduplication module - compute fingerprints and detect duplicates."""
+"""Video deduplication module - compute fingerprints and detect duplicates.
+
+Dynamic keyframe detection + sliding window temporal matching (Issue #1659).
+"""
 
 import hashlib
 import logging
+import math
 import os
+import statistics
 import tempfile
 from dataclasses import dataclass, field
 from typing import Optional
@@ -21,10 +26,28 @@ from packages.shared.storage import get_storage_service
 
 logger = logging.getLogger(__name__)
 
-# 分片策略常量
-SHORT_VIDEO_CHUNK_SEC = 2  # ≤60秒视频，每 2 秒一个分片
-LONG_VIDEO_CHUNK_SEC = 5  # >60秒视频，每 5 秒一个分片
-SHORT_VIDEO_THRESHOLD_SEC = 60
+# ── 关键帧检测常量 ──────────────────────────────────────────────
+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
+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 ────────────────────────────────────────────
+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
+
+# ── 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  # 至少 70% 帧匹配
+DUPLICATE_THRESHOLD = 0.70  # 融合后相似度阈值
+
+
+# ── 感知哈希 & 颜色直方图工具函数 ────────────────────────────────
 
 
 def compute_phash(image: np.ndarray, hash_size: int = 8) -> str:
@@ -87,15 +110,107 @@ def compute_color_histogram(image: np.ndarray, bins: int = 32) -> list[float]:
     return hist
 
 
-def compute_chunk_interval(duration: float) -> float:
-    """根据视频时长返回分片间隔（秒）。
+# ── 关键帧检测 ──────────────────────────────────────────────────
 
-    短视频（≤60秒）：每 2 秒一个分片
-    长视频（>60秒）：每 5 秒一个分片
+
+def detect_keyframe_timestamps(
+    video_path: str,
+    *,
+    min_interval_sec: float = MIN_KEYFRAME_INTERVAL_SEC,
+    max_frames: int = MAX_KEYFRAMES,
+    min_frames: int = MIN_KEYFRAMES,
+) -> list[float]:
+    """检测视频中的场景切换点，返回关键帧时间戳列表（秒）。
+
+    算法：
+    1. 降采样到 320x240，逐帧转灰度
+    2. 计算相邻帧灰度差异（像素均值差）
+    3. 差异 > SCENE_CHANGE_THRESHOLD(30) 标记为候选关键帧
+    4. 相邻关键帧间隔 < min_interval_sec 的，保留差异更大的那个
+    5. 数量裁剪到 [min_frames, max_frames]
+
+    对于长视频（>3分钟）：
+    - 每 30 秒一个分段
+    - 每个分段至少选 2 个关键帧（如果分段内无场景切换，均匀取 2 帧）
     """
-    if duration <= SHORT_VIDEO_THRESHOLD_SEC:
-        return SHORT_VIDEO_CHUNK_SEC
-    return LONG_VIDEO_CHUNK_SEC
+    cap = cv2.VideoCapture(video_path)
+    if not cap.isOpened():
+        raise RuntimeError(f"Cannot open video: {video_path}")
+
+    fps = cap.get(cv2.CAP_PROP_FPS)
+    frame_count = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    duration = frame_count / fps if fps > 0 else 0
+
+    if duration <= 0:
+        cap.release()
+        return []
+
+    # 逐帧检测场景切换
+    candidates: list[tuple[float, float]] = []  # (timestamp_sec, diff_score)
+    prev_gray = None
+
+    while True:
+        ret, frame = cap.read()
+        if not ret:
+            break
+
+        # 降采样 + 灰度
+        small = cv2.resize(frame, (320, 240))
+        gray = cv2.cvtColor(small, cv2.COLOR_BGR2GRAY).astype(np.float32)
+
+        if prev_gray is not None:
+            diff = float(np.mean(np.abs(gray - prev_gray)))
+            if diff > SCENE_CHANGE_THRESHOLD:
+                pos_ms = cap.get(cv2.CAP_PROP_POS_MSEC)
+                candidates.append((pos_ms / 1000.0, diff))
+
+        prev_gray = gray
+
+    cap.release()
+
+    # 按最小间隔过滤（保留差异更大的）
+    filtered: list[tuple[float, float]] = []
+    for ts, diff in sorted(candidates):
+        if filtered and (ts - filtered[-1][0]) < min_interval_sec:
+            if diff > filtered[-1][1]:
+                filtered[-1] = (ts, diff)
+        else:
+            filtered.append((ts, diff))
+
+    keyframe_times = [ts for ts, _ in filtered]
+
+    # 数量不足 min_frames 时，在时间轴上均匀补充
+    if len(keyframe_times) < min_frames:
+        uniform = [duration * (i + 0.5) / min_frames for i in range(min_frames)]
+        keyframe_times = sorted(set(uniform) | set(keyframe_times))
+        # 如果合并后还不足 min_frames，直接用均匀分布
+        if len(keyframe_times) < min_frames:
+            keyframe_times = uniform
+
+    # 数量超过 max_frames 时，均匀采样
+    if len(keyframe_times) > max_frames:
+        step = len(keyframe_times) / max_frames
+        keyframe_times = [keyframe_times[int(i * step)] for i in range(max_frames)]
+
+    # 长视频分段保底（>3分钟）
+    if duration > LONG_VIDEO_DURATION_THRESHOLD_SEC:
+        segment_count = int(duration / LONG_VIDEO_SEGMENT_SEC)
+        for seg_idx in range(segment_count):
+            seg_start = seg_idx * LONG_VIDEO_SEGMENT_SEC
+            seg_end = min((seg_idx + 1) * LONG_VIDEO_SEGMENT_SEC, duration)
+            seg_frames = [t for t in keyframe_times if seg_start <= t < seg_end]
+            if len(seg_frames) < MIN_FRAMES_PER_SEGMENT:
+                # 均匀补齐
+                for i in range(MIN_FRAMES_PER_SEGMENT):
+                    t = seg_start + LONG_VIDEO_SEGMENT_SEC * (i + 0.5) / MIN_FRAMES_PER_SEGMENT
+                    if t not in keyframe_times and seg_start <= t < seg_end:
+                        keyframe_times.append(t)
+        keyframe_times.sort()
+
+    return keyframe_times
+
+
+# ── 数据类 ──────────────────────────────────────────────────────
 
 
 @dataclass
@@ -109,6 +224,17 @@ class FingerprintChunk:
     frame_count: int = 1
 
 
+@dataclass
+class DuplicateSegment:
+    """一段重复片段的描述。"""
+
+    query_start_ms: int
+    query_end_ms: int
+    target_start_ms: int
+    target_end_ms: int
+    avg_distance: float  # 该段内帧的平均汉明距离
+
+
 @dataclass
 class VideoFingerprint:
     """Video fingerprint containing multiple similarity metrics."""
@@ -163,18 +289,185 @@ class VideoFingerprint:
         return models
 
 
+# ── 滑动窗口时序匹配 ────────────────────────────────────────────
+
+
+def find_duplicate_segments(
+    query_chunks: list,
+    target_chunks: list,
+    *,
+    match_threshold: int = SEGMENT_MATCH_THRESHOLD,
+    min_consecutive: int = MIN_CONSECUTIVE_MATCHES,
+    max_gap: int = MAX_GAP,
+) -> list[DuplicateSegment]:
+    """滑动窗口时序匹配：找出两组分片之间的重复片段。
+
+    算法：
+    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
+        min_consecutive: 最少连续匹配帧数
+        max_gap: 允许的最大间隙帧数
+
+    Returns:
+        DuplicateSegment 列表
+    """
+    if not query_chunks or not target_chunks:
+        return []
+
+    def _get_phash(chunk) -> str:
+        if isinstance(chunk, dict):
+            return chunk["phash_binary"]
+        return chunk.phash_binary
+
+    def _get_start(chunk) -> int:
+        if isinstance(chunk, dict):
+            return chunk["start_time_ms"]
+        return chunk.start_time_ms
+
+    def _get_end(chunk) -> int:
+        if isinstance(chunk, dict):
+            return chunk["end_time_ms"]
+        return chunk.end_time_ms
+
+    # 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: 找连续匹配的 runs
+    runs: list[tuple[int, int]] = []  # list of (start_idx, end_idx)
+    run_start = None
+    gap_count = 0
+
+    for i, (is_match, _dist, _idx) in enumerate(frame_matches):
+        if is_match:
+            if run_start is None:
+                run_start = i
+            gap_count = 0  # 重置间隙
+        else:
+            if run_start is not None:
+                gap_count += 1
+                if gap_count > max_gap:
+                    # 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 = 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
+        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]]
+        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=query_start,
+                query_end_ms=query_end,
+                target_start_ms=target_start,
+                target_end_ms=target_end,
+                avg_distance=avg_dist,
+            )
+        )
+
+    return segments
+
+
+# ── VideoDeduplicator ───────────────────────────────────────────
+
+
 class VideoDeduplicator:
     """Video deduplication using multiple fingerprint methods."""
 
-    PHASH_THRESHOLD = 10
+    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
 
-    def compute_fingerprint(self, video_path: str) -> VideoFingerprint:
-        """Compute video fingerprint using MD5, pHash, and color histogram.
+    @staticmethod
+    def _is_bad_fingerprint(phashes: list[str]) -> bool:
+        """检测指纹质量差的视频（黑屏/纯色视频）。
 
-        按时间分片抽帧：短视频（≤60s）每 2s 一片，长视频每 5s 一片。
-        每片取 1 帧计算 pHash + color_histogram。
-        同时保留 keyframe_phashes/color_histograms 聚合字段（向后兼容）。
+        当视频有多个关键帧但所有 phash 完全相同或极其相似时，
+        说明视频内容无变化（如黑屏、纯色画面），这类指纹与任何视频
+        比较都会得到虚假的"匹配"结果，应跳过。
+
+        注意：单帧视频（只有 1 个 phash）不视为坏指纹，可能是短视频或抽帧不足。
+
+        Args:
+            phashes: 关键帧 phash 列表
+
+        Returns:
+            True 表示指纹无效，应跳过
+        """
+        if not phashes:
+            return True
+        # 单帧不视为坏指纹（短视频或抽帧不足）
+        if len(phashes) == 1:
+            return False
+        # 多帧但所有 phash 完全相同 → 黑屏/纯色视频
+        unique = set(phashes)
+        if len(unique) == 1:
+            return True
+        # 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:
+            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 dynamic keyframe detection.
+
+        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():
@@ -186,41 +479,55 @@ class VideoDeduplicator:
         width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
         height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
 
+        cap.release()
+
+        # 1. 检测关键帧时间戳
+        keyframe_times = detect_keyframe_timestamps(video_path)
+
+        if not keyframe_times:
+            return VideoFingerprint(
+                md5="",
+                keyframe_phashes=[],
+                color_histograms=[],
+                duration=duration,
+                resolution=(width, height),
+                chunks=[],
+            )
+
+        # 2. 打开视频，逐个关键帧取帧
+        cap = cv2.VideoCapture(video_path)
         md5_hash = hashlib.md5(usedforsecurity=False)
         chunks: list[FingerprintChunk] = []
 
-        # 分片间隔（秒）
-        chunk_interval_sec = compute_chunk_interval(duration)
-        chunk_interval_ms = int(chunk_interval_sec * 1000)
-        duration_ms = int(duration * 1000)
-
-        # 遍历每个分片时间窗口，取 1 帧
-        start_ms = 0
-        while start_ms < duration_ms:
-            end_ms = min(start_ms + chunk_interval_ms, duration_ms)
-            # 定位到分片中点
-            seek_ms = (start_ms + end_ms) / 2
+        for i, t_sec in enumerate(keyframe_times):
+            seek_ms = t_sec * 1000
             cap.set(cv2.CAP_PROP_POS_MSEC, seek_ms)
             ret, frame = cap.read()
-            if ret:
-                # MD5 计算
-                _, buffer = cv2.imencode(".jpg", frame)
-                md5_hash.update(buffer)
+            if not ret:
+                continue
 
-                phash = compute_phash(frame)
-                hist = compute_color_histogram(frame)
+            # MD5 计算
+            _, buffer = cv2.imencode(".jpg", frame)
+            md5_hash.update(buffer)
 
-                chunks.append(
-                    FingerprintChunk(
-                        start_time_ms=start_ms,
-                        end_time_ms=end_ms,
-                        phash_binary=phash,
-                        color_histogram=hist,
-                        frame_count=1,
-                    )
+            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
+            next_boundary = keyframe_times[i + 1] * 1000 if i < len(keyframe_times) - 1 else duration * 1000
+            start_ms = int((prev_boundary + seek_ms) / 2)
+            end_ms = int((seek_ms + next_boundary) / 2)
+
+            chunks.append(
+                FingerprintChunk(
+                    start_time_ms=start_ms,
+                    end_time_ms=end_ms,
+                    phash_binary=phash,
+                    color_histogram=hist,
+                    frame_count=1,
                 )
-
-            start_ms = end_ms
+            )
 
         cap.release()
 
@@ -255,26 +562,92 @@ class VideoDeduplicator:
             for r in rows
         ]
 
-    def check_duplicate(self, fingerprint: VideoFingerprint, project_id: str, session: Session) -> Optional[dict]:
-        """检查视频是否与项目中已有视频重复。
+    @staticmethod
+    def _bhattacharyya_coefficient(hist_a: list[float], hist_b: list[float]) -> float:
+        """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 = 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(
+        histograms_a: list[list[float]],
+        histograms_b: list[list[float]],
+    ) -> float:
+        """对每组直方图，找到最佳匹配的 Bhattacharyya 系数，取平均。"""
+        if not histograms_a or not histograms_b:
+            return 0.0
+        similarities = []
+        for ha in histograms_a:
+            best = 0.0
+            for hb in histograms_b:
+                bc = VideoDeduplicator._bhattacharyya_coefficient(ha, hb)
+                best = max(best, bc)
+            similarities.append(best)
+        return sum(similarities) / len(similarities) if similarities else 0.0
+
+    @staticmethod
+    def _compute_fusion_score(
+        median_distance: float,
+        histograms_a: list[list[float]],
+        histograms_b: list[list[float]],
+    ) -> float:
+        """pHash 相似度与颜色直方图相似度的加权融合得分（Issue #1658）。
+
+        - phash_similarity = 1.0 - median_distance / 64（64 为 64bit pHash 最大汉明距离）
+        - hist_similarity = Bhattacharyya 系数均值；无直方图数据时回退中性值 0.5
+        - 融合得分 = PHASH_WEIGHT * phash_similarity + HISTOGRAM_WEIGHT * hist_similarity
+
+        返回 0~1 的原始得分，是否判重由调用方与 DUPLICATE_THRESHOLD 比较决定。
+        """
+        # DB 中 color_histograms 可能为 NULL（None），显式回退空列表而非 `or []`，
+        # 以保留全黑视频的全零直方图（[0,0,...] 为有效数据，空列表才走 0.5 中性回退）。
+        hist_a = histograms_a if histograms_a is not None else []
+        hist_b = histograms_b if histograms_b is not None else []
+
+        phash_similarity = 1.0 - (median_distance / 64)
+        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
+
+    def check_duplicate(
+        self,
+        fingerprint: VideoFingerprint,
+        project_id: str,
+        session: Session,
+        *,
+        scope: str = "project",
+        user_id: str = "",
+        duration_sec: float = 0,
+    ) -> Optional[dict]:
+        """检查视频是否与已有视频重复。
 
         查重逻辑：
         1. MD5 精确匹配 → similarity=1.0
-        2. pHash 相似度（优先从分片表读取，回退到 JSON 字段）
+        2. pHash 中位数距离 + 帧匹配比例 + 直方图融合判定
 
-        判定阈值：avg_distance < PHASH_THRESHOLD(10)
+        判定为重复后，调用 find_duplicate_segments() 获取具体重复片段。
 
         Args:
             fingerprint: 待检测视频的指纹
-            project_id: 项目 ID，仅在同一项目内搜索
+            project_id: 项目 ID
             session: 数据库会话
+            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%
 
         Returns:
-            重复信息字典（含 duplicate, duplicate_of, reason, similarity），
+            重复信息字典（含 duplicate, duplicate_of, reason, similarity, duplicate_segments），
             或 None 表示未找到重复。
         """
         video_repo = SQLAlchemyGeneratedVideoRepository(session)
-        existing_videos = video_repo.list_by_project(project_id)
+        if scope == "user" and 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)
 
         for existing in existing_videos:
             if not existing.video_fingerprint:
@@ -286,6 +659,12 @@ class VideoDeduplicator:
             if fingerprint.md5 == ef.get("md5"):
                 return {"duplicate": True, "duplicate_of": existing.id, "reason": "exact_md5_match", "similarity": 1.0}
 
+            # 跳过指纹质量差的视频（黑屏/纯色视频）
+            existing_phashes_for_check = ef.get("keyframe_phashes", [])
+            if self._is_bad_fingerprint(existing_phashes_for_check):
+                logger.debug("Skipping bad fingerprint video %s in check_duplicate", existing.id)
+                continue
+
             # 优先从分片表读取已有视频的分片 phash
             existing_phashes = []
             chunk_data = self._get_existing_chunks(existing.id, session)
@@ -298,23 +677,59 @@ class VideoDeduplicator:
             if not existing_phashes:
                 continue
 
-            # 计算每个新关键帧到已有关键帧的最小汉明距离，取平均
+            # 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))
-            avg_distance = sum(min_distances) / len(min_distances) if min_distances else 100
 
-            if avg_distance >= self.PHASH_THRESHOLD:
+            # 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
 
-            phash_similarity = 1.0 - (avg_distance / 64)
+            # 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")]
+            else:
+                existing_histograms = ef.get("color_histograms") or []
+
+            combined_score = self._compute_fusion_score(
+                median_distance, fingerprint.color_histograms, existing_histograms
+            )
+
+            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]
+            )
+            segments = find_duplicate_segments(fingerprint.chunks, existing_chunk_objects)
 
             return {
                 "duplicate": True,
                 "duplicate_of": existing.id,
-                "reason": "phash_similar",
-                "similarity": phash_similarity,
+                "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
+                ],
             }
 
         return None
@@ -325,16 +740,22 @@ class VideoDeduplicator:
         batch_id: str,
         current_video_id: str,
         session: Session,
+        *,
+        scope: str = "project",
+        user_id: str = "",
     ) -> Optional[dict]:
         """检查视频是否与同批次内其他视频重复。
 
-        逻辑与 check_duplicate 一致（MD5 + pHash），但搜索范围限定为同 batch_id 的视频。
+        逻辑与 check_duplicate 一致（MD5 + pHash + 直方图融合 + 时序匹配），
+        但搜索范围限定为同 batch_id 的视频。
 
         Args:
             fingerprint: 待检测视频的指纹
             batch_id: 批次 ID
             current_video_id: 当前视频 ID（排除自身）
             session: 数据库会话
+            scope: 保留参数，batch 模式始终按 batch_id 查询
+            user_id: 保留参数
 
         Returns:
             重复信息字典，或 None 表示未找到重复
@@ -358,6 +779,12 @@ class VideoDeduplicator:
                     "similarity": 1.0,
                 }
 
+            # 跳过指纹质量差的视频（黑屏/纯色视频）
+            existing_phashes_batch = ef.get("keyframe_phashes", [])
+            if self._is_bad_fingerprint(existing_phashes_batch):
+                logger.debug("Skipping bad fingerprint video %s in check_batch_duplicate", existing.id)
+                continue
+
             # 优先从分片表读取
             existing_phashes = []
             chunk_data = self._get_existing_chunks(existing.id, session)
@@ -373,59 +800,57 @@ class VideoDeduplicator:
             for phash in fingerprint.keyframe_phashes:
                 distances = [hamming_distance(phash, ep) for ep in existing_phashes]
                 min_distances.append(min(distances))
-            avg_distance = sum(min_distances) / len(min_distances) if min_distances else 100
 
-            if avg_distance >= self.PHASH_THRESHOLD:
+            # 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
 
-            phash_similarity = 1.0 - (avg_distance / 64)
+            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")]
+            else:
+                existing_histograms = ef.get("color_histograms") or []
+
+            combined_score = self._compute_fusion_score(
+                median_distance, fingerprint.color_histograms, existing_histograms
+            )
+
+            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]
+            )
+            segments = find_duplicate_segments(fingerprint.chunks, existing_chunk_objects)
+
             return {
                 "duplicate": True,
                 "duplicate_of": existing.id,
-                "reason": "batch_phash_similar",
-                "similarity": phash_similarity,
+                "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
+                ],
             }
 
         return None
 
-    @staticmethod
-    def _average_histogram_similarity(histograms_a: list[list[float]], histograms_b: list[list[float]]) -> float:
-        """
-        计算两组颜色直方图之间的平均余弦相似度。
-
-        对每组直方图对取最小长度对齐，计算余弦相似度后取平均。
-
-        Args:
-            histograms_a: 第一组直方图（每帧一个 list）
-            histograms_b: 第二组直方图
-
-        Returns:
-            平均余弦相似度，范围 [0, 1]
-        """
-        if not histograms_a or not histograms_b:
-            return 0.0
-
-        similarities = []
-        for ha in histograms_a:
-            best = 0.0
-            vec_a = np.array(ha, dtype=np.float64)
-            norm_a = np.linalg.norm(vec_a)
-            if norm_a == 0:
-                continue
-            for hb in histograms_b:
-                vec_b = np.array(hb, dtype=np.float64)
-                # 对齐长度
-                min_len = min(len(vec_a), len(vec_b))
-                va, vb = vec_a[:min_len], vec_b[:min_len]
-                norm_b = np.linalg.norm(vb)
-                if norm_b == 0:
-                    continue
-                sim = float(np.dot(va, vb) / (norm_a * norm_b))
-                best = max(best, sim)
-            similarities.append(best)
-
-        return sum(similarities) / len(similarities) if similarities else 0.0
-
     def compute_duplicate_rate(
         self,
         fingerprint: VideoFingerprint,
@@ -433,51 +858,48 @@ class VideoDeduplicator:
         current_video_id: str | None,
         session: Session,
         *,
+        scope: str = "project",
         user_id: str = "",
-    ) -> float:
-        """计算当前视频与用户库内已有视频的最高相似度百分比。
+    ) -> dict:
+        """计算当前视频与已有视频的查重率百分比。
 
-        优先按 user_id 全局比较（跨项目），user_id 为空时回退到项目级比较。
-        遍历最近 200 个其他有指纹的视频，对每个计算相似度：
-        - MD5 精确匹配 → 100%
-        - pHash 相似度 → (1.0 - avg_distance / 64) * 100
-        取最高值作为 duplicate_rate（0~100）。
-        如果没有其他视频可比较，返回 0.0。
+        新公式（双指标加权）：
+        - frame_match_rate = 汉明距离 < PHASH_THRESHOLD 的帧数 / 总帧数
+        - temporal_coverage_rate = 连续匹配片段总时长 / 视频总时长
+        - duplicate_rate = (frame_match_rate * 0.4 + temporal_coverage_rate * 0.6) * 100
+
+        visual_similarity = 0.7 * phash_sim + 0.3 * hist_sim（归一化到 0~1）
+
+        对每个匹配视频都算，取最高 duplicate_rate。
 
         Args:
             fingerprint: 当前视频的指纹
-            project_id: 项目 ID（user_id 为空时的回退范围）
+            project_id: 项目 ID
             current_video_id: 当前视频 ID（排除自身，可为 None）
             session: 数据库会话
-            user_id: 用户 ID（优先按用户全局比较）
+            scope: "project" 项目内（默认），"user" 跨项目全局
+            user_id: 用户 ID（scope="user" 时使用）
 
         Returns:
-            duplicate_rate: 0~100 的浮点数
+            {
+                "duplicate_rate": float,    # 0~100
+                "visual_similarity": float,  # 0~1
+                "match_count": int,          # 判定为重复的视频数
+            }
         """
-        # 限制查询最近 200 个视频，避免大库内存溢出
-        from packages.adapters.sqlalchemy_impl.models import GeneratedVideoModel
-
-        # 优先按 user_id 全局比较（跨项目），否则回退到项目级
-        if user_id:
-            query = session.query(GeneratedVideoModel).filter(
-                GeneratedVideoModel.user_id == user_id,
-            )
-            logger.debug("compute_duplicate_rate: user-level scope user_id=%s", user_id)
-        else:
-            query = session.query(GeneratedVideoModel).filter(
-                GeneratedVideoModel.project_id == project_id,
-            )
-            logger.debug("compute_duplicate_rate: project-level fallback project_id=%s", project_id)
-
-        # 排除当前视频自身（记录可能已写入 DB，必须在查询层排除）
-        if current_video_id:
-            query = query.filter(GeneratedVideoModel.id != current_video_id)
-
-        recent_models = query.order_by(GeneratedVideoModel.generated_at.desc()).limit(200).all()
         video_repo = SQLAlchemyGeneratedVideoRepository(session)
-        existing_videos = [video_repo._to_domain(m) for m in recent_models]
 
-        max_similarity = 0.0
+        if scope == "user" and user_id:
+            existing_videos = video_repo.list_by_user(user_id)
+        else:
+            existing_videos = video_repo.list_by_project(project_id)
+
+        max_duplicate_rate = 0.0
+        max_visual_similarity = 0.0
+        match_count = 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:
                 continue
@@ -488,7 +910,17 @@ class VideoDeduplicator:
 
             # MD5 精确匹配 → 100%
             if fingerprint.md5 == ef.get("md5"):
-                return 100.0
+                return {
+                    "duplicate_rate": 100.0,
+                    "visual_similarity": 1.0,
+                    "match_count": 1,
+                }
+
+            # 跳过指纹质量差的视频（黑屏/纯色视频）
+            existing_phashes_check = ef.get("keyframe_phashes", [])
+            if self._is_bad_fingerprint(existing_phashes_check):
+                logger.debug("Skipping bad fingerprint video %s in compute_duplicate_rate", existing.id)
+                continue
 
             # 优先从分片表读取
             existing_phashes = []
@@ -505,11 +937,58 @@ class VideoDeduplicator:
             for phash in fingerprint.keyframe_phashes:
                 distances = [hamming_distance(phash, ep) for ep in existing_phashes]
                 min_distances.append(min(distances))
-            avg_distance = sum(min_distances) / len(min_distances) if min_distances else 64
-            similarity = (1.0 - avg_distance / 64) * 100
-            max_similarity = max(max_similarity, similarity)
 
-        return round(max(max_similarity, 0.0), 2)
+            # 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")]
+            else:
+                # JSON NULL 显式回退空列表
+                existing_histograms = ef.get("color_histograms") or []
+
+            visual_sim = self._compute_fusion_score(median_distance, fingerprint.color_histograms, existing_histograms)
+
+            # 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 = visual_sim
+
+        return {
+            "duplicate_rate": round(max(max_duplicate_rate, 0.0), 2),
+            "visual_similarity": round(max_visual_similarity, 4),
+            "match_count": match_count,
+        }
 
 
 def _save_fingerprint_chunks(
@@ -559,7 +1038,15 @@ def check_duplicate_task(self: Task, generated_video_id: str) -> dict:
 
         fingerprint = deduplicator.compute_fingerprint(local_path)
 
-        duplicate_result = deduplicator.check_duplicate(fingerprint, video.project_id, session)
+        # 查重判定（跨项目全局 + 时长预过滤）
+        duplicate_result = deduplicator.check_duplicate(
+            fingerprint,
+            video.project_id,
+            session,
+            scope="user",
+            user_id=video.user_id,
+            duration_sec=fingerprint.duration / 1000 if fingerprint.duration else 0,
+        )
 
         video.video_fingerprint = fingerprint.to_dict()
         if duplicate_result:
@@ -569,6 +1056,19 @@ def check_duplicate_task(self: Task, generated_video_id: str) -> dict:
             video.is_duplicate = False
             video.duplicate_of = None
 
+        # 查重率计算（跨项目全局）
+        rate_result = deduplicator.compute_duplicate_rate(
+            fingerprint,
+            video.project_id,
+            generated_video_id,
+            session,
+            scope="user",
+            user_id=video.user_id,
+        )
+        video.duplicate_rate = rate_result["duplicate_rate"]
+        video.match_count = rate_result["match_count"]
+        video.visual_similarity = rate_result["visual_similarity"]
+
         video_repo.update(video)
 
         # 写入分片表
@@ -583,6 +1083,9 @@ def check_duplicate_task(self: Task, generated_video_id: str) -> dict:
             "video_id": generated_video_id,
             "is_duplicate": video.is_duplicate,
             "duplicate_of": video.duplicate_of,
+            "duplicate_rate": video.duplicate_rate,
+            "match_count": video.match_count,
+            "visual_similarity": video.visual_similarity,
             "fingerprint": fingerprint.to_dict(),
         }
     except Exception as e:

apps/worker/video_processing/dedup_helpers.py

@@ -2,6 +2,9 @@
 
 供 generate_video 共同复用，
 创建 GeneratedVideo 记录后计算指纹并执行项目级 + 批次内查重。
+
+v2: 两阶段持久化 — 先计算所有查重数据，再一次性 commit，
+避免中间异常导致 duplicate_rate 等字段缺失。
 """
 
 from __future__ import annotations
@@ -34,24 +37,14 @@ def create_video_record_and_dedup(
 ) -> int:
     """创建 GeneratedVideo 记录，计算指纹并执行查重（历史 + 批次）。
 
-    Args:
-        generation_task_id: 生成任务 ID
-        project_id: 项目 ID
-        batch_id: 批次 ID（可为空字符串）
-        file_url: 视频文件 URL
-        file_size: 文件大小（字节）
-        duration: 视频时长（秒）
-        video_path: 视频本地路径（用于计算指纹）
-        mode: 剪辑模式名称
-        session: 数据库会话
-        width: 视频宽度
-        height: 视频高度
-        fps: 视频帧率
+    采用两阶段持久化：先计算所有指纹/查重数据（内存），
+    再一次性写入数据库并 commit。若指纹计算失败，
+    视频记录仍会创建（无查重数据），但保证不会出现"写了记录却没 commit"的中间态。
 
     Returns:
         创建的视频记录数量（1 表示成功，0 表示失败）
     """
-    from video_processing.dedup import VideoDeduplicator
+    from video_processing.dedup import VideoDeduplicator, _save_fingerprint_chunks
 
     from packages.adapters.sqlalchemy_impl.generated_video_repository import (
         SQLAlchemyGeneratedVideoRepository,
@@ -60,8 +53,9 @@ def create_video_record_and_dedup(
 
     try:
         video_id = uuid4().hex
-        # 使用传入的名称，没有则 fallback 到默认命名
         video_name = name.strip() if name else f"generated-{generation_task_id[:8]}.mp4"
+
+        # ── Phase 1: 构建视频记录（内存，不 commit） ────────────────
         generated_video = GeneratedVideo(
             id=video_id,
             project_id=project_id,
@@ -76,81 +70,94 @@ def create_video_record_and_dedup(
             fps=fps,
             status="completed",
             generation_params={"mode": mode},
+            thumbnail_url=thumbnail_url or None,
         )
 
-        video_repo = SQLAlchemyGeneratedVideoRepository(session)
-        video_repo.create(generated_video)
-
-        # 生成封面缩略图
-        if thumbnail_url:
-            generated_video.thumbnail_url = thumbnail_url
-            video_repo.update_thumbnail(video_id, thumbnail_url)
-            logger.info("Thumbnail set for video %s: %s", video_id, thumbnail_url[:80] if thumbnail_url else "")
-        else:
-            logger.debug("No thumbnail_url provided for video %s, skipping", video_id)
-
-        # 计算视频指纹
+        # ── Phase 2: 计算指纹 & 查重（全部在内存） ────────────────
         deduplicator = VideoDeduplicator()
+        fingerprint = None
+
         try:
             fingerprint = deduplicator.compute_fingerprint(video_path)
         except Exception as fp_err:
             logger.warning("Fingerprint computation failed for %s: %s", video_id, fp_err)
-            session.commit()
-            return 1
 
-        generated_video.video_fingerprint = fingerprint.to_dict()
+        if fingerprint is not None:
+            generated_video.video_fingerprint = fingerprint.to_dict()
 
-        # 写入分片指纹表
-        from video_processing.dedup import _save_fingerprint_chunks
+            # 写入分片指纹表（失败不阻塞）
+            try:
+                _save_fingerprint_chunks(fingerprint, video_id, project_id, user_id, session)
+            except Exception as chunk_err:
+                logger.warning("Failed to save fingerprint chunks for %s: %s", video_id, chunk_err)
 
-        try:
-            _save_fingerprint_chunks(fingerprint, video_id, project_id, user_id, session)
-        except Exception as chunk_err:
-            logger.warning("Failed to save fingerprint chunks for %s: %s", video_id, chunk_err)
-
-        # (a) 历史成片查重
-        duplicate_result = deduplicator.check_duplicate(fingerprint, project_id, session)
-
-        # (b) 批次内查重（仅当有 batch_id 时）
-        if not duplicate_result and batch_id:
-            duplicate_result = deduplicator.check_batch_duplicate(fingerprint, batch_id, video_id, session)
-
-        if duplicate_result:
-            generated_video.is_duplicate = True
-            generated_video.duplicate_of = duplicate_result["duplicate_of"]
-            logger.info(
-                "Duplicate detected: %s -> %s (reason=%s, similarity=%.3f)",
-                video_id,
-                duplicate_result["duplicate_of"],
-                duplicate_result["reason"],
-                duplicate_result["similarity"],
-            )
-        else:
-            generated_video.is_duplicate = False
-            generated_video.duplicate_of = None
-
-        # 计算重复率百分比（与项目内所有已有视频对比取最高相似度）
-        try:
-            dup_rate = deduplicator.compute_duplicate_rate(
+            # (a) 历史成片查重（跨项目全局 + 时长预过滤）
+            duration_sec = fingerprint.duration / 1000 if fingerprint.duration else 0
+            duplicate_result = deduplicator.check_duplicate(
                 fingerprint,
                 project_id,
-                video_id,
                 session,
+                scope="user",
                 user_id=user_id,
+                duration_sec=duration_sec,
             )
-            generated_video.duplicate_rate = dup_rate
-            logger.info("Duplicate rate for %s: %.2f%%", video_id, dup_rate)
-        except Exception as rate_err:
-            logger.warning("Failed to compute duplicate_rate for %s: %s", video_id, rate_err)
-            generated_video.duplicate_rate = None
 
-        video_repo.update(generated_video)
+            # (b) 批次内查重（仅当有 batch_id 时）
+            if not duplicate_result and batch_id:
+                duplicate_result = deduplicator.check_batch_duplicate(fingerprint, batch_id, video_id, session)
+
+            if duplicate_result:
+                generated_video.is_duplicate = True
+                generated_video.duplicate_of = duplicate_result["duplicate_of"]
+                logger.info(
+                    "Duplicate detected: %s -> %s (reason=%s, similarity=%.3f)",
+                    video_id,
+                    duplicate_result["duplicate_of"],
+                    duplicate_result["reason"],
+                    duplicate_result["similarity"],
+                )
+            else:
+                generated_video.is_duplicate = False
+                generated_video.duplicate_of = None
+
+            # 计算重复率百分比（跨项目全局）
+            try:
+                rate_result = deduplicator.compute_duplicate_rate(
+                    fingerprint,
+                    project_id,
+                    video_id,
+                    session,
+                    scope="user",
+                    user_id=user_id,
+                )
+                generated_video.duplicate_rate = rate_result["duplicate_rate"]
+                generated_video.match_count = rate_result["match_count"]
+                generated_video.visual_similarity = rate_result["visual_similarity"]
+                logger.info(
+                    "Duplicate rate for %s: %.2f%% (visual_sim=%.3f, matches=%d)",
+                    video_id,
+                    rate_result["duplicate_rate"],
+                    rate_result["visual_similarity"],
+                    rate_result["match_count"],
+                )
+            except Exception as rate_err:
+                logger.warning("Failed to compute duplicate_rate for %s: %s", video_id, rate_err)
+                generated_video.duplicate_rate = None
+
+        # ── Phase 3: 一次性持久化 ─────────────────────────────────
+        video_repo = SQLAlchemyGeneratedVideoRepository(session)
+        video_repo.create(generated_video)
+
+        if thumbnail_url:
+            logger.info("Thumbnail set for video %s: %s", video_id, thumbnail_url[:80])
+
         session.commit()
         logger.info(
-            "GeneratedVideo record created: %s (task=%s, dup=%s)",
+            "GeneratedVideo record created: %s (task=%s, dup=%s, rate=%s)",
             video_id,
             generation_task_id,
             generated_video.is_duplicate,
+            generated_video.duplicate_rate,
         )
         return 1
     except Exception as e:

apps/worker/video_processing/ffmpeg_utils.py

@@ -304,3 +304,127 @@ def normalize_video(
     ]
     run_ffmpeg(command)
     return {"width": width, "height": height, "path": output_path}
+
+
+def random_edge_crop(
+    input_path: str | Path,
+    output_path: str | Path | None = None,
+    *,
+    min_crop_pct: float = 0.02,
+    max_crop_pct: float = 0.05,
+) -> Path:
+    """对视频四边做随机裁剪再缩放回原分辨率，用于改变 pHash 指纹。
+
+    Args:
+        input_path: 输入视频路径
+        output_path: 输出路径；为 None 时写入 input_path 同目录的临时文件，
+                     成功后覆盖原文件
+        min_crop_pct: 每边最小裁剪比例（默认 2%）
+        max_crop_pct: 每边最大裁剪比例（默认 5%）
+
+    Returns:
+        输出文件路径（Path 对象）
+
+    Raises:
+        subprocess.CalledProcessError: ffmpeg 执行失败时抛出
+    """
+    import random
+    import shutil
+    import tempfile
+
+    input_path = Path(input_path)
+
+    # 获取原始分辨率
+    info = probe_video_info(str(input_path))
+    W = info["width"]
+    H = info["height"]
+
+    if W <= 0 or H <= 0:
+        logger.warning("无法获取视频分辨率 (W=%d H=%d)，跳过裁剪: %s", W, H, input_path)
+        return input_path
+
+    # 四边各自随机裁剪 2%~5%
+    crop_top = int(H * random.uniform(min_crop_pct, max_crop_pct))
+    crop_bottom = int(H * random.uniform(min_crop_pct, max_crop_pct))
+    crop_left = int(W * random.uniform(min_crop_pct, max_crop_pct))
+    crop_right = int(W * random.uniform(min_crop_pct, max_crop_pct))
+
+    # 裁剪后尺寸（确保至少 2 像素）
+    new_w = max(W - crop_left - crop_right, 2)
+    new_h = max(H - crop_top - crop_bottom, 2)
+    x_offset = crop_left
+    y_offset = crop_top
+
+    # 确保裁剪尺寸为偶数（ffmpeg 编码器常要求偶数尺寸）
+    new_w = new_w if new_w % 2 == 0 else new_w - 1
+    new_h = new_h if new_h % 2 == 0 else new_h - 1
+    if new_w < 2:
+        new_w = 2
+    if new_h < 2:
+        new_h = 2
+
+    # 输出分辨率必须与原始一致
+    out_w = W if W % 2 == 0 else W + 1
+    out_h = H if H % 2 == 0 else H + 1
+
+    vf = f"crop={new_w}:{new_h}:{x_offset}:{y_offset},scale={out_w}:{out_h}"
+
+    logger.info(
+        "随机边缘裁剪: %s → crop(%d,%d,%d,%d)=%dx%d scale→%dx%d",
+        input_path.name,
+        crop_top,
+        crop_bottom,
+        crop_left,
+        crop_right,
+        new_w,
+        new_h,
+        out_w,
+        out_h,
+    )
+
+    # 确定输出路径
+    if output_path is None:
+        temp_fd, temp_path = tempfile.mkstemp(suffix=".mp4", dir=input_path.parent)
+        import os
+
+        os.close(temp_fd)
+        temp_output = Path(temp_path)
+        replace_original = True
+    else:
+        temp_output = Path(output_path)
+        replace_original = False
+
+    command = [
+        FFMPEG_BIN,
+        "-y",
+        "-i",
+        str(input_path),
+        "-vf",
+        vf,
+        "-c:v",
+        "libx264",
+        "-preset",
+        "fast",
+        "-crf",
+        "18",
+        "-c:a",
+        "copy",
+        "-movflags",
+        "+faststart",
+        str(temp_output),
+    ]
+
+    try:
+        run_ffmpeg(command)
+    except Exception:
+        # 裁剪失败时清理临时文件
+        if temp_output.exists() and replace_original:
+            temp_output.unlink(missing_ok=True)
+        raise
+
+    # 成功 → 覆盖原文件
+    if replace_original:
+        shutil.move(str(temp_output), str(input_path))
+        return input_path
+
+    return temp_output

apps/worker/video_processing/subtitle_generator.py

@@ -213,7 +213,7 @@ def generate_ass_from_timeline(
             t_shadow.get("offset_x", 2) if t_shadow.get("enabled", False) else 0,
             t_shadow.get("offset_y", 2) if t_shadow.get("enabled", False) else 0,
         )
-        t_alignment = position_to_ass_alignment(title_cfg.get("position", "top"))
+        t_alignment = position_to_ass_alignment(title_cfg.get("position", "bottom"))
 
         title_style_line = build_ass_style(
             "TitleStyle",

apps/worker/video_processing/unified_render_service.py

@@ -198,8 +198,13 @@ class UnifiedRenderService:
         # 2. 分组为 RenderLayers
         layers = self._group_clips_into_layers(resolved)
 
+        # 2.5 配音时长对齐：如果有配音素材，调整片段时长以匹配配音时长
+        voice_duration = self._get_voice_audio_duration()
+        if voice_duration > 0:
+            self._align_clips_to_voice_duration(layers, voice_duration)
+
         # 3. 计算视频总时长（用于字幕显示时长）
-        video_duration = self._estimate_total_duration(layers)
+        video_duration_final = self._estimate_total_duration(layers)
         # Debug: 输出各图层时长明细
         for layer in layers:
             layer_total = sum(UnifiedRenderService._clip_adjusted_duration(c) for c in layer.clips)
@@ -215,16 +220,16 @@ class UnifiedRenderService:
                 self.transition_duration,
                 ", ".join(clip_details),
             )
-        logger.info("[debug] estimated video_duration=%.3f", video_duration)
+        logger.info("[debug] estimated video_duration=%.3f", video_duration_final)
 
         # 3.5 TTS 配音生成（如果配置了）
-        self._maybe_add_voiceover_layer(layers, video_duration=video_duration)
+        self._maybe_add_voiceover_layer(layers, video_duration=video_duration_final)
 
         # 3.6 配音素材库音频（如果传入了本地路径）
-        self._maybe_add_voice_library_layer(layers, video_duration=video_duration)
+        self._maybe_add_voice_library_layer(layers, video_duration=video_duration_final)
 
         # 4. 生成 ASS 字幕文件（如果有 title/subtitle 配置）
-        ass_path = self._maybe_generate_ass(video_duration)
+        ass_path = self._maybe_generate_ass(video_duration_final)
 
         # 4.5 解析画中画配置
         pip_config = PiPConfig.from_dict((self.plan.config or {}).get("pip_config"))
@@ -257,7 +262,7 @@ class UnifiedRenderService:
             # 先尝试 stream copy 优化（无重编码，性能提升 10 倍+）
             # 条件不满足或失败时回退到带滤镜的直通渲染
             stream_copy_ok = self._try_render_stream_copy(
-                layers, output_path, ass_path=ass_path, video_duration=video_duration
+                layers, output_path, ass_path=ass_path, video_duration=video_duration_final
             )
             if stream_copy_ok:
                 used_stream_copy = True
@@ -271,7 +276,7 @@ class UnifiedRenderService:
                     layers,
                     output_path,
                     ass_path=ass_path,
-                    video_duration=video_duration,
+                    video_duration=video_duration_final,
                 )
         else:
             filter_complex, input_args = self._build_filter_complex(layers, ass_path=ass_path)
@@ -327,7 +332,7 @@ class UnifiedRenderService:
                         from video_processing.ffmpeg_utils import run_ffmpeg
 
                         run_ffmpeg(extract_cmd)
-                        final_audio = mix_bgm_with_main(ctx, main_audio_path, bgm_cfg, video_duration)
+                        final_audio = mix_bgm_with_main(ctx, main_audio_path, bgm_cfg, video_duration_final)
                         # 合并回视频
 
                         bgm_output = self.work_dir / f"rendered_{self.plan.id}_bgm.mp4"
@@ -353,7 +358,7 @@ class UnifiedRenderService:
             audio_path = mix_audio(
                 ctx,
                 layers,
-                video_duration,
+                video_duration_final,
                 bgm_path=self.bgm_path,
                 bgm_config=bgm_config,
                 audio_tracks_config=audio_tracks_config,
@@ -487,6 +492,147 @@ class UnifiedRenderService:
         """
         return _estimate_total_duration_pure(layers, self.transition_duration)
 
+    def _get_voice_audio_duration(self) -> float:
+        """获取配音音频文件的时长（秒）。
+
+        Returns:
+            配音音频时长，如果无配音或探测失败则返回 0.0
+        """
+        if not self.voiceover_audio_path:
+            return 0.0
+
+        audio_path = Path(self.voiceover_audio_path)
+        if not audio_path.exists() or audio_path.stat().st_size == 0:
+            return 0.0
+
+        try:
+            duration = probe_duration(audio_path)
+            logger.info("[voice-align] 配音音频时长: %.3fs path=%s", duration, self.voiceover_audio_path)
+            return duration
+        except Exception as e:
+            logger.warning("[voice-align] 探测配音音频时长失败: %s", e)
+            return 0.0
+
+    def _align_clips_to_voice_duration(
+        self,
+        layers: list[RenderLayer],
+        voice_duration: float,
+    ) -> None:
+        """调整片段时长以对齐配音时长。
+
+        核心逻辑：
+        - 计算片段总时长与配音时长的比例
+        - ±5% 以内不调整
+        - ratio < 1（片段比配音长）：按比例裁剪每段末尾
+        - ratio > 1（片段比配音短）：按比例慢放每段
+
+        Args:
+            layers: 渲染图层列表
+            voice_duration: 配音时长（秒）
+        """
+        if voice_duration <= 0:
+            return
+
+        # 只调整视频图层（main/broll/background），不调整音频图层
+        video_layers = [layer for layer in layers if layer.role in ("main", "broll", "background")]
+        if not video_layers:
+            return
+
+        # 计算所有视频图层的总时长
+        total_clips_duration = 0.0
+        for layer in video_layers:
+            for clip in layer.clips:
+                clip_dur = self._clip_adjusted_duration(clip)
+                total_clips_duration += clip_dur
+
+        if total_clips_duration <= 0:
+            return
+
+        ratio = voice_duration / total_clips_duration
+
+        # ±5% 以内不调整
+        if abs(ratio - 1.0) <= 0.05:
+            logger.info(
+                "[voice-align] 比例接近1:1，跳过调整: ratio=%.4f voice=%.3f clips=%.3f",
+                ratio,
+                voice_duration,
+                total_clips_duration,
+            )
+            return
+
+        logger.info(
+            "[voice-align] 开始调整片段时长: ratio=%.4f voice=%.3f clips=%.3f",
+            ratio,
+            voice_duration,
+            total_clips_duration,
+        )
+
+        # 收集所有视频 clip
+        all_clips: list[tuple[RenderLayer, ResolvedClip]] = []
+        for layer in video_layers:
+            for clip in layer.clips:
+                all_clips.append((layer, clip))
+
+        if not all_clips:
+            return
+
+        if ratio < 1.0:
+            # 片段比配音长，按比例裁剪每段末尾
+            # 减少每个 clip 的 duration
+            for _layer, clip in all_clips:
+                old_duration = clip.duration if clip.duration > 0 else clip.actual_duration
+                new_duration = old_duration * ratio
+
+                # 更新 duration
+                clip.duration = max(0.1, new_duration)  # 至少 0.1s
+
+                # 如果有 trim_config，也需要调整
+                if clip.trim_config is not None:
+                    new_trim_duration = clip.trim_config.duration * ratio
+                    clip.trim_config = TrimConfig(
+                        start_time=clip.trim_config.start_time,
+                        duration=max(0.1, new_trim_duration),
+                    )
+
+                logger.debug(
+                    "[voice-align] trim clip=%s: %.3f -> %.3f",
+                    clip.clip_id,
+                    old_duration,
+                    clip.duration,
+                )
+
+        else:
+            # ratio > 1.0: 片段比配音短，按比例慢放每段
+            # 降低 playback_speed
+            for _layer, clip in all_clips:
+                old_speed = clip.playback_speed if clip.playback_speed > 0 else 1.0
+                # speed = old_speed / ratio 会使视频变慢（ratio > 1 时）
+                new_speed = old_speed / ratio
+
+                # 下限 0.25x（避免过慢）
+                new_speed = max(0.25, round(new_speed, 4))
+                clip.playback_speed = new_speed
+
+                logger.debug(
+                    "[voice-align] slowdown clip=%s: speed %.4f -> %.4f",
+                    clip.clip_id,
+                    old_speed,
+                    new_speed,
+                )
+
+        # 调整后重新计算总时长用于日志
+        new_total = 0.0
+        for layer in video_layers:
+            for clip in layer.clips:
+                new_total += self._clip_adjusted_duration(clip)
+
+        logger.info(
+            "[voice-align] 调整完成: 新总时长=%.3fs (目标=%.3fs, 差异=%.3fs)",
+            new_total,
+            voice_duration,
+            abs(new_total - voice_duration),
+        )
+
     def _maybe_generate_ass(self, video_duration: float) -> Path | None:
         """根据 plan.config 生成 ASS 字幕文件。
 

apps/worker/worker_app/celery_app.py

@@ -15,6 +15,7 @@ celery_app.conf.imports = (
     "worker_app.tasks.voice_clone",
     "worker_app.tasks.tts_synthesis",
     "worker_app.tasks.batch_download",
+    "worker_app.tasks.duplication_check",
     "worker_app.tasks._startup",
     "apps.worker.video_processing.dedup",
     "worker_app.tasks.cleanup",

apps/worker/worker_app/tasks/duplication_check.py

@@ -0,0 +1,198 @@
+"""手动查重任务（Issue #1661）。
+
+流程：
+1. 从 OSS 下载用户上传的待查重视频
+2. 动态抽帧计算指纹（复用 VideoDeduplicator.compute_fingerprint）
+3. 跨项目与用户所有已有成片比对（compute_duplicate_rate + find_duplicate_segments）
+4. 更新 DuplicationRecord：status / duplicate_rate / duplicate_count / segments
+   同时写入 visual_similarity / match_count
+5. 失败重试 3 次、间隔 60 秒，最终失败标记 failed；临时文件始终清理
+"""
+
+import logging
+import os
+import shutil
+import tempfile
+
+from celery import Task
+from celery.exceptions import Retry
+from video_processing.dedup import (
+    VideoDeduplicator,
+    find_duplicate_segments,
+)
+from worker_app.celery_app import celery_app
+from worker_app.db import SessionLocal
+
+from packages.adapters.sqlalchemy_impl.duplication_repository import (
+    SQLAlchemyDuplicationRecordRepository,
+)
+from packages.adapters.sqlalchemy_impl.generated_video_repository import (
+    SQLAlchemyGeneratedVideoRepository,
+)
+from packages.domain.duplication import DuplicateSegment
+from packages.shared.storage import get_storage_service
+
+logger = logging.getLogger(__name__)
+
+
+def _build_domain_segments(
+    fingerprint,
+    session,
+    deduplicator: VideoDeduplicator,
+    user_id: str,
+) -> tuple[list[DuplicateSegment], int]:
+    """对用户所有已有视频做分片级时序匹配，构建领域片段列表。
+
+    Returns:
+        (segments, duplicate_count) — segments 为 query 视频中的重复片段，
+        duplicate_count 为存在重复片段的匹配视频数。
+    """
+    video_repo = SQLAlchemyGeneratedVideoRepository(session)
+    existing_videos = video_repo.list_by_user(user_id)
+
+    segments_out: list[DuplicateSegment] = []
+    duplicate_count = 0
+
+    for existing in existing_videos:
+        if not existing.video_fingerprint:
+            continue
+
+        chunk_data = deduplicator._get_existing_chunks(existing.id, session)
+        if not chunk_data:
+            # 老视频无分片数据，时序定位不可靠，跳过片段级匹配
+            continue
+
+        raw_segments = find_duplicate_segments(fingerprint.chunks, chunk_data)
+        if not raw_segments:
+            continue
+
+        duplicate_count += 1
+        for raw in raw_segments:
+            avg_sim = 1.0 - raw.avg_distance / 64.0
+            segments_out.append(
+                DuplicateSegment.create(
+                    source_start=round(raw.query_start_ms / 1000.0, 2),
+                    source_end=round(raw.query_end_ms / 1000.0, 2),
+                    matched_video_id=existing.id,
+                    matched_video_name=existing.name,
+                    matched_start=round(raw.target_start_ms / 1000.0, 2),
+                    matched_end=round(raw.target_end_ms / 1000.0, 2),
+                    similarity=round(max(0.0, min(1.0, avg_sim)) * 100, 1),
+                )
+            )
+
+    # 按 query 起始时间排序，片段时间轴稳定
+    segments_out.sort(key=lambda s: (s.source_start, s.source_end))
+    return segments_out, duplicate_count
+
+
+@celery_app.task(bind=True, max_retries=3, name="worker.process_duplication_check")
+def process_duplication_check(self: Task, record_id: str) -> dict:
+    """处理一次手动查重请求。
+
+    Args:
+        record_id: DuplicationRecord ID
+
+    Returns:
+        dict: {"ok": True, "record_id": ..., "duplicate_rate": ..., ...}
+    """
+    session = None
+    temp_dir = None
+    try:
+        session = SessionLocal()
+        repo = SQLAlchemyDuplicationRecordRepository(session)
+        storage_service = get_storage_service()
+        deduplicator = VideoDeduplicator()
+
+        record = repo.get(record_id)
+        if record is None:
+            raise ValueError(f"Duplication record {record_id} not found")
+
+        if record.status not in ("pending", "processing"):
+            logger.info("Duplication record %s already %s, skip", record_id, record.status)
+            return {"ok": True, "record_id": record_id, "status": record.status, "skipped": True}
+
+        record.mark_processing()
+        repo.update(record)
+        session.commit()
+
+        temp_dir = tempfile.mkdtemp(prefix="dup_check_")
+        suffix = os.path.splitext(record.filename)[1] or ".mp4"
+        local_path = os.path.join(temp_dir, f"{record_id}{suffix}")
+
+        storage_service.download_file(record.storage_key, local_path)
+
+        fingerprint = deduplicator.compute_fingerprint(local_path)
+        record.duration_seconds = round(fingerprint.duration, 2) if fingerprint.duration else 0.0
+        record.video_fingerprint = fingerprint.to_dict()
+
+        # 跨项目与用户所有已有视频比对（current_video_id=None：上传视频不在成片表中）
+        rate_result = deduplicator.compute_duplicate_rate(
+            fingerprint,
+            project_id="",
+            current_video_id=None,
+            session=session,
+            scope="user",
+            user_id=record.user_id,
+        )
+
+        # 分片级时序匹配 → 重复片段
+        segments, segment_match_count = _build_domain_segments(fingerprint, session, deduplicator, record.user_id)
+
+        record.mark_completed(
+            duplicate_rate=rate_result["duplicate_rate"],
+            duplicate_count=segment_match_count,
+            segments=segments,
+            visual_similarity=rate_result["visual_similarity"],
+            match_count=rate_result["match_count"],
+        )
+        repo.update(record)
+        session.commit()
+
+        logger.info(
+            "Duplication check completed: record=%s rate=%.2f%% matches=%d segments=%d",
+            record_id,
+            record.duplicate_rate,
+            record.match_count,
+            len(segments),
+        )
+
+        return {
+            "ok": True,
+            "record_id": record_id,
+            "status": "completed",
+            "duplicate_rate": record.duplicate_rate,
+            "duplicate_count": record.duplicate_count,
+            "visual_similarity": record.visual_similarity,
+            "match_count": record.match_count,
+            "segments": len(segments),
+        }
+
+    except Retry:
+        raise
+
+    except Exception as e:
+        logger.error("Duplication check failed for record %s: %s", record_id, e, exc_info=True)
+        if session is not None:
+            session.rollback()
+            # 超过重试上限：标记 failed 并返回失败结果，不再 retry
+            if "repo" in locals() and self.request.retries >= self.max_retries:
+                try:
+                    failed_record = repo.get(record_id)
+                    if failed_record is not None and failed_record.status != "failed":
+                        failed_record.mark_failed(f"查重失败（已重试{self.max_retries}次）: {e}")
+                        repo.update(failed_record)
+                        session.commit()
+                except Exception as inner:
+                    logger.error("Failed to mark duplication record %s as failed: %s", record_id, inner)
+                    session.rollback()
+                return {"ok": False, "record_id": record_id, "status": "failed", "error": str(e)}
+            # 未达上限：60 秒后重试
+            raise self.retry(exc=e, countdown=60) from e
+        return {"ok": False, "record_id": record_id, "status": "failed", "error": str(e)}
+
+    finally:
+        if session is not None:
+            session.close()
+        if temp_dir and os.path.isdir(temp_dir):
+            shutil.rmtree(temp_dir, ignore_errors=True)

apps/worker/worker_app/tasks/generation.py

@@ -716,6 +716,28 @@ def generate_video(self, task_id: str) -> dict:
 
                     _update_task_progress(task_id, 80, "渲染完成")
 
+                    # ── 3.5 随机边缘裁剪降重（#1664） ──────────────────────────
+                    from video_processing.ffmpeg_utils import random_edge_crop
+
+                    try:
+                        cropped_path = random_edge_crop(output_path)
+                        if cropped_path != output_path:
+                            output_path = cropped_path
+                        if gen_task:
+                            gen_task.append_log("边缘裁剪", "已应用随机 2-5% 边缘裁剪降重")
+                            _flush_logs(task_id, gen_task)
+                        logger.info("[task_id=%s] 随机边缘裁剪完成: %s", task_id, output_path)
+                    except Exception as crop_err:
+                        logger.warning(
+                            "[task_id=%s] 随机边缘裁剪失败，使用原始视频继续: %s",
+                            task_id,
+                            crop_err,
+                            exc_info=True,
+                        )
+                        if gen_task:
+                            gen_task.append_log("边缘裁剪", f"裁剪失败，使用原始视频: {crop_err}")
+                            _flush_logs(task_id, gen_task)
+
                     # ── 4. 上传 OSS + 查重记录 ───────────────────────────────
                     _update_task_progress(task_id, 85, "开始上传")
                     file_url, duration, file_size, video_count = _upload_and_record(

packages/adapters/sqlalchemy_impl/duplication_repository.py

@@ -25,6 +25,8 @@ class SQLAlchemyDuplicationRecordRepository:
             status=record.status,
             duplicate_rate=record.duplicate_rate,
             duplicate_count=record.duplicate_count,
+            visual_similarity=record.visual_similarity,
+            match_count=record.match_count,
             video_fingerprint=json.dumps(record.video_fingerprint) if record.video_fingerprint else None,
             error_message=record.error_message,
             created_at=record.created_at,
@@ -58,6 +60,8 @@ class SQLAlchemyDuplicationRecordRepository:
         model.status = record.status
         model.duplicate_rate = record.duplicate_rate
         model.duplicate_count = record.duplicate_count
+        model.visual_similarity = record.visual_similarity
+        model.match_count = record.match_count
         model.video_fingerprint = json.dumps(record.video_fingerprint) if record.video_fingerprint else None
         model.error_message = record.error_message
         model.updated_at = record.updated_at
@@ -121,6 +125,8 @@ class SQLAlchemyDuplicationRecordRepository:
             status=model.status,
             duplicate_rate=model.duplicate_rate,
             duplicate_count=int(model.duplicate_count or 0),
+            visual_similarity=getattr(model, "visual_similarity", None),
+            match_count=getattr(model, "match_count", None),
             video_fingerprint=json.loads(fp_raw) if fp_raw else None,
             error_message=getattr(model, "error_message", ""),
             segments=segments,

packages/adapters/sqlalchemy_impl/edit_plan_clip_repository.py

@@ -131,3 +131,65 @@ class SQLAlchemyEditPlanClipRepository:
             created_at=model.created_at,
             updated_at=model.updated_at,
         )
+
+    def list_used_segments_by_user(
+        self,
+        user_id: str,
+        *,
+        limit_recent: int = 50,
+    ) -> dict[str, list[tuple[float, float]]]:
+        """查询用户已有视频中已使用的素材区间（跨视频避让）.
+
+        JOIN edit_plans 表，按 created_by_user_id 过滤，只查 status='completed'
+        的 plan 下 status='rendered' 且 asset_id 非空的 clips。按 plan 的
+        created_at DESC 取最近 limit_recent 个 plan。
+
+        Returns:
+            {asset_id: [(start_time, start_time + duration), ...]}
+            空结果返回空 dict。
+        """
+        from packages.adapters.sqlalchemy_impl.models import EditPlanModel
+
+        if not user_id:
+            return {}
+
+        # 1. 查出最近 limit_recent 个已完成 plan 的 ID
+        recent_plan_ids = [
+            row[0]
+            for row in self.session.query(EditPlanModel.id)
+            .filter(
+                EditPlanModel.created_by_user_id == user_id,
+                EditPlanModel.status == "completed",
+            )
+            .order_by(EditPlanModel.created_at.desc())
+            .limit(limit_recent)
+            .all()
+        ]
+
+        if not recent_plan_ids:
+            return {}
+
+        # 2. 查这些 plan 下已渲染、有素材的 clips
+        clips = (
+            self.session.query(
+                EditPlanClipModel.asset_id,
+                EditPlanClipModel.start_time,
+                EditPlanClipModel.duration,
+            )
+            .filter(
+                EditPlanClipModel.plan_id.in_(recent_plan_ids),
+                EditPlanClipModel.status == "rendered",
+                EditPlanClipModel.asset_id != "",
+                EditPlanClipModel.asset_id.isnot(None),
+            )
+            .all()
+        )
+
+        # 3. 聚合为 {asset_id: [(start, start+duration), ...]}
+        result: dict[str, list[tuple[float, float]]] = {}
+        for asset_id, start_time, duration in clips:
+            if asset_id not in result:
+                result[asset_id] = []
+            result[asset_id].append((start_time or 0.0, (start_time or 0.0) + (duration or 0.0)))
+
+        return result

packages/adapters/sqlalchemy_impl/generated_video_repository.py

@@ -31,6 +31,8 @@ class SQLAlchemyGeneratedVideoRepository:
             is_duplicate=video.is_duplicate,
             duplicate_of=video.duplicate_of,
             duplicate_rate=video.duplicate_rate,
+            match_count=getattr(video, "match_count", None),
+            visual_similarity=getattr(video, "visual_similarity", None),
             generated_at=video.generated_at,
             created_at=video.created_at,
         )
@@ -62,6 +64,8 @@ class SQLAlchemyGeneratedVideoRepository:
             is_duplicate=getattr(model, "is_duplicate", False),
             duplicate_of=getattr(model, "duplicate_of", None),
             duplicate_rate=getattr(model, "duplicate_rate", None),
+            match_count=getattr(model, "match_count", None),
+            visual_similarity=getattr(model, "visual_similarity", None),
             generated_at=model.generated_at,
             created_at=model.created_at,
         )
@@ -77,6 +81,8 @@ class SQLAlchemyGeneratedVideoRepository:
         model.is_duplicate = video.is_duplicate
         model.duplicate_of = video.duplicate_of
         model.duplicate_rate = video.duplicate_rate
+        model.match_count = getattr(video, "match_count", None)
+        model.visual_similarity = getattr(video, "visual_similarity", None)
         self.session.add(model)
         self.session.commit()
         return video
@@ -85,6 +91,24 @@ class SQLAlchemyGeneratedVideoRepository:
         models = self.session.query(GeneratedVideoModel).filter(GeneratedVideoModel.project_id == project_id).all()
         return [self._to_domain(model) for model in models]
 
+    def list_by_user(self, user_id: str, *, duration_min: float = 0, duration_max: float = 0) -> list[GeneratedVideo]:
+        """按 user_id 查询用户所有项目的视频（跨项目查重）。
+
+        Args:
+            user_id: 用户 ID
+            duration_min: 时长下限（秒），0 表示不限
+            duration_max: 时长上限（秒），0 表示不限
+        """
+        query = self.session.query(GeneratedVideoModel).filter(
+            GeneratedVideoModel.user_id == user_id,
+        )
+        if duration_min > 0:
+            query = query.filter(GeneratedVideoModel.duration >= duration_min)
+        if duration_max > 0:
+            query = query.filter(GeneratedVideoModel.duration <= duration_max)
+        models = query.all()
+        return [self._to_domain(model) for model in models]
+
     def list_by_generation_task(self, generation_task_id: str) -> list[GeneratedVideo]:
         models = (
             self.session.query(GeneratedVideoModel)
@@ -208,6 +232,8 @@ class SQLAlchemyGeneratedVideoRepository:
             is_duplicate=getattr(model, "is_duplicate", False),
             duplicate_of=getattr(model, "duplicate_of", None),
             duplicate_rate=getattr(model, "duplicate_rate", None),
+            match_count=getattr(model, "match_count", None),
+            visual_similarity=getattr(model, "visual_similarity", None),
             generated_at=model.generated_at,
             created_at=model.created_at,
         )

packages/adapters/sqlalchemy_impl/models.py

@@ -340,6 +340,8 @@ class GeneratedVideoModel(Base):
     is_duplicate = Column(Boolean, nullable=False, default=False)
     duplicate_of = Column(String(36), nullable=True)
     duplicate_rate = Column(Float, nullable=True)
+    match_count = Column(Integer, nullable=True, default=0)
+    visual_similarity = Column(Float, nullable=True, default=0.0)
 
 
 class TitleLibraryModel(Base):
@@ -415,6 +417,9 @@ class DuplicationRecordModel(Base):
     status = Column(String(20), nullable=False, default="pending", index=True)
     duplicate_rate = Column(Float, nullable=True)
     duplicate_count = Column(Integer, nullable=False, default=0)
+    # #1661 手动查重：视觉相似度（0~1）/ 匹配视频数
+    visual_similarity = Column(Float, nullable=True)
+    match_count = Column(Integer, nullable=True)
     video_fingerprint = Column(Text, nullable=True)
     error_message = Column(Text, nullable=False, default="")
     created_at = Column(DateTime, nullable=False, default=lambda: datetime.now(timezone.utc))

packages/domain/ass_subtitle_builder.py

@@ -81,14 +81,14 @@ def position_to_ass_alignment(position: str) -> int:
         position: 位置字符串 top/center/bottom
 
     Returns:
-        ASS 对齐编号，默认 8（顶部居中）
+        ASS 对齐编号，默认 2（底部居中，与前端 DEFAULT_TITLE_SETTINGS.position="bottom" 对齐）
     """
     mapping = {
         "top": 8,
         "center": 5,
         "bottom": 2,
     }
-    return mapping.get(position, 8)
+    return mapping.get(position, 2)
 
 
 # ── Style 行构建 ──────────────────────────────────────────────────────────────
@@ -226,7 +226,6 @@ def _wrap_title_text(
 
     # 换行计算使用原始 font_size，与 CSS 预览一致；1.35x 补偿仅用于 ASS Fontsize 渲染
 
-
     # 先按已有 \N 分段，每段独立自动换行，最后用 \N 拼回
     segments = text.split("\\N")
     wrapped_segments: list[str] = []
@@ -386,8 +385,8 @@ def build_ass_content(
         # position → alignment 三档逻辑，现有输出保持一字节不变。
         title_pos = _parse_title_position(title_config, video_width, video_height)
 
-        title_alignment = 5 if title_pos is not None else position_to_ass_alignment(
-            title_config.get("position", "top")
+        title_alignment = (
+            5 if title_pos is not None else position_to_ass_alignment(title_config.get("position", "bottom"))
         )
 
         styles.append(

packages/domain/duplication.py

@@ -63,6 +63,9 @@ class DuplicationRecord:
     status: str = "pending"  # pending / processing / completed / failed
     duplicate_rate: float | None = None  # 0-100
     duplicate_count: int = 0
+    # #1661 手动查重：视觉相似度（归一化 0~1）/ 匹配视频数
+    visual_similarity: float | None = None
+    match_count: int | None = None
     video_fingerprint: dict[str, Any] | None = None
     error_message: str = ""
     segments: list[DuplicateSegment] = field(default_factory=list)
@@ -98,13 +101,23 @@ class DuplicationRecord:
         self.status = "processing"
         self.updated_at = datetime.now(timezone.utc)
 
-    def mark_completed(self, duplicate_rate: float, duplicate_count: int, segments: list[DuplicateSegment]) -> None:
+    def mark_completed(
+        self,
+        duplicate_rate: float,
+        duplicate_count: int,
+        segments: list[DuplicateSegment],
+        *,
+        visual_similarity: float | None = None,
+        match_count: int | None = None,
+    ) -> None:
         if not 0 <= duplicate_rate <= 100:
             raise ValueError("duplicate_rate must be between 0 and 100")
         self.status = "completed"
         self.duplicate_rate = duplicate_rate
         self.duplicate_count = duplicate_count
         self.segments = segments
+        self.visual_similarity = visual_similarity
+        self.match_count = match_count
         self.updated_at = datetime.now(timezone.utc)
 
     def mark_failed(self, error_message: str) -> None:
@@ -133,6 +146,8 @@ class DuplicationRecord:
         self.status = "pending"
         self.duplicate_rate = None
         self.duplicate_count = 0
+        self.visual_similarity = None
+        self.match_count = None
         self.error_message = ""
         self.segments = []
         self.video_fingerprint = None

packages/domain/generated_video.py

@@ -27,6 +27,8 @@ class GeneratedVideo:
     is_duplicate: bool = False
     duplicate_of: str | None = None
     duplicate_rate: float | None = None
+    match_count: int | None = None
+    visual_similarity: float | None = None
     generated_at: datetime = field(default_factory=lambda: datetime.now(timezone.utc))
     created_at: datetime = field(default_factory=lambda: datetime.now(timezone.utc))
 

packages/domain/plan_generator_utils.py

@@ -169,6 +169,7 @@ def distribute_assets(
     random_selection: bool = False,
     asset_durations: dict[str, float] | None = None,
     asset_scene_points: dict[str, list[float]] | None = None,
+    external_used_segments: dict[str, list[tuple[float, float]]] | None = None,
 ) -> None:
     """按 editing_mode 将素材分配到 clips（就地修改）.
 
@@ -188,6 +189,7 @@ def distribute_assets(
         random_selection: 是否随机选择素材（用于预览生成）
         asset_durations: 素材 ID -> 时长（秒）映射，用于设置 start_time
         asset_scene_points: 素材 ID -> 场景切换点列表（metadata 缓存）
+        external_used_segments: 跨视频已用区间（来自其他视频的 clips），注入到分配逻辑中避让
     """
     if not asset_ids or not clips:
         return
@@ -198,16 +200,16 @@ def distribute_assets(
         random.shuffle(asset_ids)
 
     if editing_mode == EditingMode.ONE_TAKE.value:
-        _distribute_one_take(clips, asset_ids, asset_durations, asset_scene_points)
+        _distribute_one_take(clips, asset_ids, asset_durations, asset_scene_points, external_used_segments)
     elif editing_mode == EditingMode.PIP.value:
-        _distribute_pip(clips, asset_ids, asset_durations, asset_scene_points)
+        _distribute_pip(clips, asset_ids, asset_durations, asset_scene_points, external_used_segments)
     elif editing_mode == EditingMode.VOICE_OVER.value:
-        _distribute_voice_over(clips, asset_ids, asset_durations, asset_scene_points)
+        _distribute_voice_over(clips, asset_ids, asset_durations, asset_scene_points, external_used_segments)
     elif editing_mode == EditingMode.VOICE_PIP.value:
-        _distribute_voice_pip(clips, asset_ids, asset_durations, asset_scene_points)
+        _distribute_voice_pip(clips, asset_ids, asset_durations, asset_scene_points, external_used_segments)
     else:
         # 未知模式，退化为 one_take
-        _distribute_one_take(clips, asset_ids, asset_durations, asset_scene_points)
+        _distribute_one_take(clips, asset_ids, asset_durations, asset_scene_points, external_used_segments)
 
 
 def _resolve_start_time(
@@ -248,9 +250,12 @@ def _distribute_one_take(
     asset_ids: List[str],
     asset_durations: dict[str, float] | None = None,
     asset_scene_points: dict[str, list[float]] | None = None,
+    external_used_segments: dict[str, list[tuple[float, float]]] | None = None,
 ) -> None:
     """ONE_TAKE: 素材按顺序依次分配给 main 类型 clips."""
-    used_segments: dict[str, list[tuple[float, float]]] = {}
+    used_segments: dict[str, list[tuple[float, float]]] = (
+        {k: list(v) for k, v in external_used_segments.items()} if external_used_segments else {}
+    )
     main_clips = [c for c in clips if c.clip_type == ClipType.MAIN.value]
     for i, clip in enumerate(main_clips):
         if i < len(asset_ids):
@@ -271,9 +276,12 @@ def _distribute_pip(
     asset_ids: List[str],
     asset_durations: dict[str, float] | None = None,
     asset_scene_points: dict[str, list[float]] | None = None,
+    external_used_segments: dict[str, list[tuple[float, float]]] | None = None,
 ) -> None:
     """PIP: 第1个素材→main(全屏背景)，其余→overlay clips."""
-    used_segments: dict[str, list[tuple[float, float]]] = {}
+    used_segments: dict[str, list[tuple[float, float]]] = (
+        {k: list(v) for k, v in external_used_segments.items()} if external_used_segments else {}
+    )
     # 第1个素材 → main clip
     main_clips = [c for c in clips if c.clip_type == ClipType.MAIN.value]
     if main_clips and asset_ids:
@@ -310,9 +318,12 @@ def _distribute_voice_over(
     asset_ids: List[str],
     asset_durations: dict[str, float] | None = None,
     asset_scene_points: dict[str, list[float]] | None = None,
+    external_used_segments: dict[str, list[tuple[float, float]]] | None = None,
 ) -> None:
     """VOICE_OVER: 素材→main clips (B-roll)."""
-    used_segments: dict[str, list[tuple[float, float]]] = {}
+    used_segments: dict[str, list[tuple[float, float]]] = (
+        {k: list(v) for k, v in external_used_segments.items()} if external_used_segments else {}
+    )
     main_clips = [c for c in clips if c.clip_type == ClipType.MAIN.value]
     for i, clip in enumerate(main_clips):
         if i < len(asset_ids):
@@ -333,9 +344,12 @@ def _distribute_voice_pip(
     asset_ids: List[str],
     asset_durations: dict[str, float] | None = None,
     asset_scene_points: dict[str, list[float]] | None = None,
+    external_used_segments: dict[str, list[tuple[float, float]]] | None = None,
 ) -> None:
     """VOICE_PIP: 第1个→background, 第2个→corner_voice, 其余→b_roll."""
-    used_segments: dict[str, list[tuple[float, float]]] = {}
+    used_segments: dict[str, list[tuple[float, float]]] = (
+        {k: list(v) for k, v in external_used_segments.items()} if external_used_segments else {}
+    )
     bg_clips = [c for c in clips if c.clip_type == "background"]
     voice_clips = [c for c in clips if c.clip_type == "corner_voice"]
     broll_clips = [c for c in clips if c.clip_type == "b_roll"]

tests/unit/domain/test_ass_subtitle_builder.py

@@ -83,11 +83,11 @@ class TestPositionToAssAlignment:
     def test_bottom(self):
         assert position_to_ass_alignment("bottom") == 2
 
-    def test_unknown_defaults_top(self):
-        assert position_to_ass_alignment("unknown") == 8
+    def test_unknown_defaults_bottom(self):
+        assert position_to_ass_alignment("unknown") == 2
 
-    def test_empty_defaults_top(self):
-        assert position_to_ass_alignment("") == 8
+    def test_empty_defaults_bottom(self):
+        assert position_to_ass_alignment("") == 2
 
 
 # ============================================================
@@ -581,6 +581,7 @@ class TestConstants:
         assert isinstance(TITLE_MARGIN_BOTTOM, int)
         assert isinstance(TITLE_MARGIN_SIDE, int)
 
+
 # ============================================================
 # _wrap_title_text 换行逻辑验证
 # ============================================================

tests/unit/test_ass_subtitle_builder.py

@@ -65,11 +65,11 @@ class TestPositionToAssAlignment:
     def test_bottom(self):
         assert position_to_ass_alignment("bottom") == 2
 
-    def test_unknown_default_top(self):
-        assert position_to_ass_alignment("unknown") == 8
+    def test_unknown_default_bottom(self):
+        assert position_to_ass_alignment("unknown") == 2
 
-    def test_empty_default_top(self):
-        assert position_to_ass_alignment("") == 8
+    def test_empty_default_bottom(self):
+        assert position_to_ass_alignment("") == 2
 
 
 # ── Style 行构建 ─────────────────────────────────────────────────────────────
@@ -747,3 +747,45 @@ class TestTitleFreePosition:
             line for line in content.splitlines() if line.startswith("Dialogue:") and "SubtitleStyle" in line
         ][0]
         assert "\\pos(" not in sub_dialogue
+
+
+class TestDefaultPositionBottom:
+    """默认 position 应为 bottom（alignment=2），与前端 DEFAULT_TITLE_SETTINGS 对齐。"""
+
+    def _base_kwargs(self):
+        return dict(
+            video_width=1080,
+            video_height=1920,
+            video_duration=10.0,
+            title_text="测试标题",
+        )
+
+    def test_no_position_defaults_to_bottom_alignment(self):
+        """不传 position 时，Alignment 应为 2（bottom）。"""
+        content = build_ass_content(
+            **self._base_kwargs(),
+            title_config={"size": 36},
+        )
+        style_line = [line for line in content.splitlines() if line.startswith("Style: TitleStyle")][0]
+        fields = [f.strip() for f in style_line.split(",")]
+        assert fields[18] == "2", f"Expected alignment 2 (bottom), got {fields[18]}"
+
+    def test_no_position_no_coords_defaults_to_bottom(self):
+        """不传 position 也不传坐标时，走 bottom 三档逻辑。"""
+        content = build_ass_content(
+            **self._base_kwargs(),
+            title_config={},
+        )
+        style_line = [line for line in content.splitlines() if line.startswith("Style: TitleStyle")][0]
+        fields = [f.strip() for f in style_line.split(",")]
+        assert fields[18] == "2"
+
+    def test_explicit_top_still_works(self):
+        """显式传 position='top' 仍然得到 alignment=8。"""
+        content = build_ass_content(
+            **self._base_kwargs(),
+            title_config={"position": "top", "size": 36},
+        )
+        style_line = [line for line in content.splitlines() if line.startswith("Style: TitleStyle")][0]
+        fields = [f.strip() for f in style_line.split(",")]
+        assert fields[18] == "8"

tests/unit/test_bad_fingerprint_filter.py

@@ -0,0 +1,314 @@
+"""Tests for bad fingerprint (black screen / uniform color) filtering.
+
+Issue: 1秒黑屏视频（所有帧phash几乎相同）与任何视频的距离都~30，造成虚假匹配。
+Fix: _is_bad_fingerprint() 检测并跳过这类低质量指纹。
+"""
+
+from __future__ import annotations
+
+import sys
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+# ---------------------------------------------------------------------------
+# Mock heavy deps before importing dedup module (same pattern as test_dedup_engine.py)
+# ---------------------------------------------------------------------------
+_ORIGINAL_MODULES = dict(sys.modules)
+_MOCKED_MODULE_NAMES: list[str] = []
+
+
+def _mock_if_absent(name: str, mock_obj=None):
+    if name not in sys.modules:
+        sys.modules[name] = mock_obj if mock_obj is not None else MagicMock()
+        _MOCKED_MODULE_NAMES.append(name)
+
+
+_mock_if_absent("ffmpeg")
+for mod_name in ["worker_app", "worker_app.celery_app", "worker_app.db"]:
+    _mock_if_absent(mod_name)
+if "worker_app.celery_app" in sys.modules and isinstance(sys.modules["worker_app.celery_app"], MagicMock):
+    sys.modules["worker_app.celery_app"].celery_app = MagicMock()
+if "worker_app.db" in sys.modules and isinstance(sys.modules["worker_app.db"], MagicMock):
+    sys.modules["worker_app.db"].SessionLocal = MagicMock()
+_mock_if_absent("celery", MagicMock())
+if "celery" in sys.modules and isinstance(sys.modules["celery"], MagicMock):
+    sys.modules["celery"].Task = object
+_mock_if_absent("packages.shared.storage")
+_mock_if_absent("packages.adapters.sqlalchemy_impl.generated_video_repository")
+
+_HAS_CV2 = False
+try:
+    import cv2 as _cv2
+
+    if not isinstance(_cv2, MagicMock):
+        _HAS_CV2 = True
+except (ImportError, ModuleNotFoundError):
+    pass
+
+if not _HAS_CV2:
+    _mock_if_absent("cv2")
+
+import numpy as np  # noqa: E402
+
+from apps.worker.video_processing.dedup import (  # noqa: E402
+    VideoDeduplicator,
+    VideoFingerprint,
+)
+
+# Restore mocked modules
+for _name in ["worker_app", "worker_app.celery_app", "worker_app.db", "celery"]:
+    if _name in _MOCKED_MODULE_NAMES:
+        sys.modules.pop(_name, None)
+        _MOCKED_MODULE_NAMES.remove(_name)
+
+
+@pytest.fixture(autouse=True, scope="session")
+def _cleanup_mocks():
+    yield
+    for name in _MOCKED_MODULE_NAMES:
+        sys.modules.pop(name, None)
+
+
+# ── _is_bad_fingerprint 单元测试 ─────────────────────────────────
+
+
+class TestIsBadFingerprint:
+    """VideoDeduplicator._is_bad_fingerprint() 静态方法测试。"""
+
+    def test_empty_phashes_is_bad(self):
+        """空 phash 列表视为坏指纹。"""
+        assert VideoDeduplicator._is_bad_fingerprint([]) is True
+
+    def test_single_phash_is_not_bad(self):
+        """单帧视频不视为坏指纹（短视频或抽帧不足）。"""
+        assert VideoDeduplicator._is_bad_fingerprint(["abcdef0123456789"]) is False
+
+    def test_all_identical_phashes_is_bad(self):
+        """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_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):
+        """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):
+        """多样化的 phash 列表是有效指纹。"""
+        phashes = [
+            "abcdef0123456789",
+            "1234567890abcdef",
+            "fedcba9876543210",
+            "0123456789abcdef",
+        ]
+        assert VideoDeduplicator._is_bad_fingerprint(phashes) is False
+
+    def test_mixed_similar_and_different_is_good(self):
+        """有些 phash 相似但有足够多样的 → 有效指纹。"""
+        phashes = [
+            "0000000000000000",
+            "0000000000000001",
+            "0000000000000002",
+            "ffffffffffffffff",
+        ]
+        assert VideoDeduplicator._is_bad_fingerprint(phashes) is False
+
+    def test_known_black_screen_phashes(self):
+        """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"] * 6) is True
+
+
+# ── Helper ──────────────────────────────────────────────────────
+
+
+def _make_existing_video(video_id, md5, phashes):
+    """创建 mock 视频记录。"""
+    video = MagicMock()
+    video.id = video_id
+    video.video_fingerprint = {
+        "md5": md5,
+        "keyframe_phashes": phashes,
+        "color_histograms": [],
+    }
+    return video
+
+
+# ── check_duplicate 集成测试 ────────────────────────────────────
+
+
+class TestCheckDuplicateBadFingerprint:
+    """check_duplicate 跳过坏指纹视频。"""
+
+    def test_black_screen_existing_video_skipped(self):
+        """已有视频是黑屏指纹 → 被跳过，不匹配。"""
+        deduplicator = VideoDeduplicator()
+        mock_session = MagicMock()
+
+        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"] * 5,
+            color_histograms=[],
+            duration=10.0,
+            resolution=(1280, 720),
+        )
+
+        with patch(
+            "apps.worker.video_processing.dedup.SQLAlchemyGeneratedVideoRepository",
+            return_value=mock_repo,
+        ):
+            result = deduplicator.check_duplicate(fingerprint, "proj-1", mock_session, scope="user", user_id="user-1")
+
+        assert result is None
+
+    def test_normal_existing_video_not_skipped(self):
+        """正常视频不会被坏指纹过滤跳过。"""
+        deduplicator = VideoDeduplicator()
+        mock_session = MagicMock()
+
+        normal = _make_existing_video(
+            "vid-normal",
+            "md5_normal_existing",
+            ["abcdef0123456789", "1234567890abcdef", "fedcba9876543210"],
+        )
+        mock_repo = MagicMock()
+        mock_repo.list_by_user.return_value = [normal]
+
+        fingerprint = VideoFingerprint(
+            md5="md5_normal_new",
+            keyframe_phashes=["abcdef0123456789", "1234567890abcdef", "fedcba9876543210"],
+            color_histograms=[],
+            duration=10.0,
+            resolution=(1280, 720),
+        )
+
+        with patch(
+            "apps.worker.video_processing.dedup.SQLAlchemyGeneratedVideoRepository",
+            return_value=mock_repo,
+        ):
+            result = deduplicator.check_duplicate(fingerprint, "proj-1", mock_session, scope="user", user_id="user-1")
+
+        assert result is not None
+        assert result["duplicate"] is True
+
+    def test_md5_match_overrides_bad_fingerprint(self):
+        """MD5 精确匹配优先于坏指纹过滤。"""
+        deduplicator = VideoDeduplicator()
+        mock_session = MagicMock()
+
+        black_screen = _make_existing_video("vid-black", "same_md5", ["aaaaaaaaaaaaaaaa"] * 5)
+        mock_repo = MagicMock()
+        mock_repo.list_by_user.return_value = [black_screen]
+
+        fingerprint = VideoFingerprint(
+            md5="same_md5",
+            keyframe_phashes=["bbbbbbbbbbbbbbbb"] * 3,
+            color_histograms=[],
+            duration=10.0,
+            resolution=(1280, 720),
+        )
+
+        with patch(
+            "apps.worker.video_processing.dedup.SQLAlchemyGeneratedVideoRepository",
+            return_value=mock_repo,
+        ):
+            result = deduplicator.check_duplicate(fingerprint, "proj-1", mock_session, scope="user", user_id="user-1")
+
+        assert result is not None
+        assert result["reason"] == "exact_md5_match"
+
+
+# ── compute_duplicate_rate 集成测试 ─────────────────────────────
+
+
+class TestComputeDuplicateRateBadFingerprint:
+    """compute_duplicate_rate 跳过坏指纹视频。"""
+
+    def test_black_screen_video_excluded_from_rate(self):
+        """黑屏视频不参与查重率计算。"""
+        deduplicator = VideoDeduplicator()
+        mock_session = MagicMock()
+
+        videos = [
+            _make_existing_video("vid-b1", "md5_b1", ["cccccccccccccccc"] * 5),
+            _make_existing_video("vid-b2", "md5_b2", ["dddddddddddddddd"] * 5),
+            _make_existing_video("vid-b3", "md5_b3", ["eeeeeeeeeeeeeeee"] * 5),
+            _make_existing_video(
+                "vid-normal",
+                "md5_n",
+                ["abcdef0123456789", "1234567890abcdef", "fedcba9876543210"],
+            ),
+        ]
+        mock_repo = MagicMock()
+        mock_repo.list_by_user.return_value = videos
+
+        fingerprint = VideoFingerprint(
+            md5="md5_new",
+            keyframe_phashes=["abcdef0123456789", "1234567890abcdef", "fedcba9876543210"],
+            color_histograms=[],
+            duration=10.0,
+            resolution=(1280, 720),
+        )
+
+        with patch(
+            "apps.worker.video_processing.dedup.SQLAlchemyGeneratedVideoRepository",
+            return_value=mock_repo,
+        ):
+            result = deduplicator.compute_duplicate_rate(
+                fingerprint,
+                "proj-1",
+                "vid-new",
+                mock_session,
+                scope="user",
+                user_id="user-1",
+            )
+
+        assert result is not None
+        assert isinstance(result["duplicate_rate"], float)
+        assert isinstance(result["match_count"], int)
+
+    def test_only_black_screen_videos_zero_rate(self):
+        """所有已有视频都是黑屏 → 查重率为 0。"""
+        deduplicator = VideoDeduplicator()
+        mock_session = MagicMock()
+
+        videos = [
+            _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
+
+        fingerprint = VideoFingerprint(
+            md5="md5_new",
+            keyframe_phashes=["aaaaaaaaaaaaaaaa"] * 5,
+            color_histograms=[],
+            duration=10.0,
+            resolution=(1280, 720),
+        )
+
+        with patch(
+            "apps.worker.video_processing.dedup.SQLAlchemyGeneratedVideoRepository",
+            return_value=mock_repo,
+        ):
+            result = deduplicator.compute_duplicate_rate(
+                fingerprint,
+                "proj-1",
+                "vid-new",
+                mock_session,
+                scope="user",
+                user_id="user-1",
+            )
+
+        assert result["duplicate_rate"] == 0.0
+        assert result["match_count"] == 0

tests/unit/test_cross_video_avoidance.py

@@ -0,0 +1,334 @@
+"""Tests for Issue #1670 — 跨视频片段避让（生成前注入已用区间）."""
+
+from __future__ import annotations
+
+from datetime import datetime, timezone
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+from packages.adapters.sqlalchemy_impl.edit_plan_clip_repository import (
+    SQLAlchemyEditPlanClipRepository,
+)
+from packages.domain.edit_plan_clip import EditPlanClip, EditPlanClipStatus
+from packages.domain.plan_generator_utils import (
+    _distribute_one_take,
+    distribute_assets,
+)
+
+# ── Repository 层测试 ─────────────────────────────────────────────────────────
+
+
+class TestListUsedSegmentsByUser:
+    """测试 list_used_segments_by_user 方法."""
+
+    def _make_repo(self, session_mock):
+        return SQLAlchemyEditPlanClipRepository(session_mock)
+
+    def test_empty_user_id_returns_empty_dict(self):
+        """空 user_id 直接返回空 dict，不查 DB."""
+        session = MagicMock()
+        repo = self._make_repo(session)
+        result = repo.list_used_segments_by_user("")
+        assert result == {}
+        session.query.assert_not_called()
+
+    def test_no_completed_plans_returns_empty_dict(self):
+        """用户没有已完成的 plan 时返回空 dict."""
+        session = MagicMock()
+        # Mock plan query returns empty
+        plan_query = MagicMock()
+        plan_query.filter.return_value = plan_query
+        plan_query.order_by.return_value = plan_query
+        plan_query.limit.return_value = plan_query
+        plan_query.all.return_value = []
+        session.query.return_value = plan_query
+
+        repo = self._make_repo(session)
+        result = repo.list_used_segments_by_user("user_123")
+        assert result == {}
+
+    def test_aggregates_clips_from_multiple_plans(self):
+        """从多个已完成 plan 的 clips 聚合已用区间."""
+        session = MagicMock()
+
+        # Mock plan query: 2 completed plans
+        plan_query = MagicMock()
+        plan_query.filter.return_value = plan_query
+        plan_query.order_by.return_value = plan_query
+        plan_query.limit.return_value = plan_query
+        plan_query.all.return_value = [("plan_1",), ("plan_2",)]
+        session.query.return_value = plan_query
+
+        # Mock clip query: clips from both plans
+        clip_query = MagicMock()
+        clip_query.filter.return_value = clip_query
+        clip_query.all.return_value = [
+            ("asset_A", 0.0, 5.0),  # plan_1, asset A: 0~5s
+            ("asset_A", 10.0, 3.0),  # plan_1, asset A: 10~13s
+            ("asset_B", 2.0, 4.0),  # plan_2, asset B: 2~6s
+        ]
+        # Second session.query call is for clips
+        session.query.side_effect = [plan_query, clip_query]
+
+        repo = self._make_repo(session)
+        result = repo.list_used_segments_by_user("user_123")
+
+        assert "asset_A" in result
+        assert len(result["asset_A"]) == 2
+        assert result["asset_A"][0] == (0.0, 5.0)
+        assert result["asset_A"][1] == (10.0, 13.0)
+        assert "asset_B" in result
+        assert result["asset_B"][0] == (2.0, 6.0)
+
+    def test_respects_limit_recent_parameter(self):
+        """limit_recent 参数限制查询的 plan 数量."""
+        session = MagicMock()
+
+        plan_query = MagicMock()
+        plan_query.filter.return_value = plan_query
+        plan_query.order_by.return_value = plan_query
+        plan_query.limit.return_value = plan_query
+        plan_query.all.return_value = [("plan_1",)]
+        session.query.return_value = plan_query
+
+        clip_query = MagicMock()
+        clip_query.filter.return_value = clip_query
+        clip_query.all.return_value = [("asset_X", 1.0, 2.0)]
+        session.query.side_effect = [plan_query, clip_query]
+
+        repo = self._make_repo(session)
+        result = repo.list_used_segments_by_user("user_123", limit_recent=10)
+
+        # Verify limit was called with the parameter
+        plan_query.limit.assert_called_once_with(10)
+        assert "asset_X" in result
+
+
+# ── Domain 层测试 ─────────────────────────────────────────────────────────────
+
+
+class TestDistributeAssetsWithExternalSegments:
+    """测试 distribute_assets 传入 external_used_segments 的行为."""
+
+    def _make_clips(self, count: int, duration: float = 3.0) -> list[EditPlanClip]:
+        """创建指定数量的 MAIN 类型 clips."""
+        return [
+            EditPlanClip(
+                id=f"clip_{i}",
+                plan_id="plan_1",
+                clip_type="main",
+                order=i,
+                template_clip_config_id="",
+                asset_id="",
+                text_content="",
+                start_time=0.0,
+                duration=duration,
+                status=EditPlanClipStatus.PENDING,
+            )
+            for i in range(count)
+        ]
+
+    def test_external_used_segments_none_backward_compatible(self):
+        """external_used_segments=None 时行为不变（向后兼容）."""
+        clips = self._make_clips(3)
+        asset_ids = ["asset_1", "asset_2", "asset_3"]
+        asset_durations = {aid: 30.0 for aid in asset_ids}
+
+        # Should not raise
+        distribute_assets(
+            clips,
+            asset_ids,
+            "one_take",
+            asset_durations=asset_durations,
+            external_used_segments=None,
+        )
+
+        # All clips should have assets assigned
+        for clip in clips:
+            assert clip.asset_id != ""
+
+    def test_external_used_segments_avoids_existing_ranges(self):
+        """传入 external_used_segments 后，新分配的 start_time 避开已有区间."""
+        clips = self._make_clips(2, duration=3.0)
+        asset_ids = ["asset_1"]
+        asset_durations = {"asset_1": 30.0}
+
+        # Pretend asset_1 0~10s is already used by another video
+        external = {"asset_1": [(0.0, 10.0)]}
+
+        # Run multiple times to check that start_time always avoids 0~10s
+        # (with some randomness, but the avoidance should be consistent)
+        for _ in range(10):
+            test_clips = self._make_clips(1, duration=3.0)
+            distribute_assets(
+                test_clips,
+                asset_ids,
+                "one_take",
+                asset_durations=asset_durations,
+                external_used_segments=external,
+            )
+            start = test_clips[0].start_time
+            # Start time + duration (3s) should not overlap with 0~10
+            # i.e., start >= 10.0 or start + 3 <= 0.0 (impossible since start >= 0)
+            assert (
+                start >= 10.0 or start + 3.0 <= 0.0 or start >= 10.0
+            ), f"start_time {start} overlaps with existing segment 0~10"
+
+    def test_external_used_segments_deep_copy(self):
+        """external_used_segments 会被深拷贝，不会修改外部数据."""
+        external = {"asset_1": [(0.0, 5.0)]}
+        original = {"asset_1": [(0.0, 5.0)]}
+
+        clips = self._make_clips(1, duration=2.0)
+        asset_ids = ["asset_1"]
+        asset_durations = {"asset_1": 20.0}
+
+        distribute_assets(
+            clips,
+            asset_ids,
+            "one_take",
+            asset_durations=asset_durations,
+            external_used_segments=external,
+        )
+
+        # External dict should be unchanged
+        assert external == original
+
+    def test_empty_external_used_segments_same_as_none(self):
+        """空 dict 的 external_used_segments 行为与 None 相同."""
+        clips = self._make_clips(2, duration=3.0)
+        asset_ids = ["asset_1", "asset_2"]
+        asset_durations = {aid: 30.0 for aid in asset_ids}
+
+        # Should not raise and should assign assets normally
+        distribute_assets(
+            clips,
+            asset_ids,
+            "one_take",
+            asset_durations=asset_durations,
+            external_used_segments={},
+        )
+        for clip in clips:
+            assert clip.asset_id != ""
+
+
+# ── Service 层测试 ────────────────────────────────────────────────────────────
+
+
+class TestServiceLayerIntegration:
+    """测试 _distribute_assets 在 service 层的查询逻辑."""
+
+    def _make_service(self, clip_repo_mock, asset_repo_mock=None):
+        """创建 PlanGeneratorService 并注入 mock repos."""
+
+        from apps.api.app.services.plan_generator_service import PlanGeneratorService
+
+        with (
+            patch("apps.api.app.services.plan_generator_service.SQLAlchemyEditPlanRepository"),
+            patch(
+                "apps.api.app.services.plan_generator_service.SQLAlchemyEditPlanClipRepository",
+                return_value=clip_repo_mock,
+            ),
+        ):
+            db = MagicMock()
+            svc = PlanGeneratorService(db, asset_repo=asset_repo_mock)
+            svc._clip_repo = clip_repo_mock
+        return svc
+
+    def _make_clip(self):
+        return EditPlanClip(
+            id="clip_1",
+            plan_id="plan_1",
+            clip_type="main",
+            order=0,
+            template_clip_config_id="",
+            asset_id="",
+            text_content="",
+            start_time=0.0,
+            duration=3.0,
+            status=EditPlanClipStatus.PENDING,
+        )
+
+    def test_query_called_with_user_id(self):
+        """有 user_id 时调用 list_used_segments_by_user."""
+        clip_repo = MagicMock()
+        clip_repo.list_used_segments_by_user.return_value = {"asset_A": [(0.0, 5.0)]}
+        asset_repo = MagicMock()
+        asset_repo.get.return_value = None  # smart_match fallback
+
+        svc = self._make_service(clip_repo, asset_repo)
+        clips = [self._make_clip()]
+
+        svc._distribute_assets(
+            clips,
+            ["asset_A"],
+            "one_take",
+            asset_durations={"asset_A": 30.0},
+            user_id="user_123",
+        )
+
+        clip_repo.list_used_segments_by_user.assert_called_once_with("user_123", limit_recent=50)
+
+    def test_query_not_called_without_user_id(self):
+        """无 user_id 时不调用查询."""
+        clip_repo = MagicMock()
+        asset_repo = MagicMock()
+        asset_repo.get.return_value = None
+
+        svc = self._make_service(clip_repo, asset_repo)
+        clips = [self._make_clip()]
+
+        svc._distribute_assets(
+            clips,
+            ["asset_A"],
+            "one_take",
+            asset_durations={"asset_A": 30.0},
+            user_id="",
+        )
+
+        clip_repo.list_used_segments_by_user.assert_not_called()
+
+    def test_query_failure_does_not_block_generation(self):
+        """查询失败时不阻塞生成，回退到纯随机."""
+        clip_repo = MagicMock()
+        clip_repo.list_used_segments_by_user.side_effect = Exception("DB error")
+        asset_repo = MagicMock()
+        asset_repo.get.return_value = None
+
+        svc = self._make_service(clip_repo, asset_repo)
+        clips = [self._make_clip()]
+
+        # Should not raise
+        svc._distribute_assets(
+            clips,
+            ["asset_A"],
+            "one_take",
+            asset_durations={"asset_A": 30.0},
+            user_id="user_123",
+        )
+
+        # Clip should still get an asset assigned (fallback to random)
+        assert clips[0].asset_id == "asset_A"
+
+    def test_preview_and_final_both_query(self):
+        """预览和正式生成都触发查询."""
+        for random_selection in [True, False]:
+            clip_repo = MagicMock()
+            clip_repo.list_used_segments_by_user.return_value = {}
+            asset_repo = MagicMock()
+            asset_repo.get.return_value = None
+
+            svc = self._make_service(clip_repo, asset_repo)
+            clips = [self._make_clip()]
+
+            svc._distribute_assets(
+                clips,
+                ["asset_A"],
+                "one_take",
+                random_selection=random_selection,
+                asset_durations={"asset_A": 30.0},
+                user_id="user_123",
+            )
+
+            clip_repo.list_used_segments_by_user.assert_called_once()

tests/unit/test_dedup_engine.py

@@ -285,8 +285,10 @@ class TestVideoDeduplicatorCheckDuplicate:
             result = deduplicator.check_duplicate(fingerprint, "proj-1", mock_session)
             assert result is not None
             assert result["duplicate"] is True
-            assert result["similarity"] == 1.0  # distance=0 → 1.0
-            assert result["reason"] == "phash_similar"
+            assert result["similarity"] == pytest.approx(
+                0.85, abs=0.01
+            )  # combined: 0.7*1.0 + 0.3*0.5 (no hist fallback)
+            assert result["reason"] == "phash_histogram_fusion"
         finally:
             self._restore_repo(mod, orig)
 
@@ -425,8 +427,11 @@ class TestVideoDeduplicatorCheckDuplicate:
             result = deduplicator.check_duplicate(fingerprint, "proj-1", mock_session)
             assert result is not None
             assert result["duplicate"] is True
-            # similarity = 1.0 - (1 / 64) = 0.984375
-            assert abs(result["similarity"] - (1.0 - 1.0 / 64)) < 1e-6
+            # 新算法: median_distance=1, phash_sim=1-1/64=0.984375
+            # 无直方图 → hist_sim=0.5(fallback)
+            # combined = 0.7*0.984375 + 0.3*0.5 = 0.839062
+            expected_sim = 0.7 * (1.0 - 1.0 / 64) + 0.3 * 0.5
+            assert abs(result["similarity"] - expected_sim) < 1e-6
         finally:
             self._restore_repo(mod, orig)
 
@@ -456,7 +461,9 @@ class TestVideoDeduplicatorCheckDuplicate:
             result = deduplicator.check_duplicate(fingerprint, "proj-1", mock_session)
             assert result is not None
             assert result["duplicate"] is True
-            assert result["similarity"] == 1.0  # avg_distance = 0
+            # 新算法: median_distance=0, phash_sim=1.0, hist_sim=0.5(fallback)
+            # combined = 0.7*1.0 + 0.3*0.5 = 0.85
+            assert result["similarity"] == pytest.approx(0.85, abs=0.01)
         finally:
             self._restore_repo(mod, orig)
 
@@ -539,7 +546,7 @@ class TestVideoDeduplicatorCheckBatchDuplicate:
             result = deduplicator.check_batch_duplicate(fingerprint, "batch-1", "vid-self", mock_session)
             assert result is not None
             assert result["duplicate"] is True
-            assert result["reason"] == "batch_phash_similar"
+            assert result["reason"] == "batch_phash_histogram_fusion"
         finally:
             self._restore_repo(mod, orig)
 

tests/unit/test_dedup_helpers_user_id.py

@@ -43,7 +43,11 @@ class TestDedupHelpersUserIdPassthrough:
         mock_deduplicator = MagicMock()
         mock_deduplicator.compute_fingerprint.return_value = mock_fingerprint
         mock_deduplicator.check_duplicate.return_value = None
-        mock_deduplicator.compute_duplicate_rate.return_value = 42.5
+        mock_deduplicator.compute_duplicate_rate.return_value = {
+            "duplicate_rate": 42.5,
+            "visual_similarity": 0.7,
+            "match_count": 2,
+        }
 
         with (
             patch(
@@ -85,7 +89,11 @@ class TestDedupHelpersUserIdPassthrough:
         mock_deduplicator = MagicMock()
         mock_deduplicator.compute_fingerprint.return_value = mock_fingerprint
         mock_deduplicator.check_duplicate.return_value = None
-        mock_deduplicator.compute_duplicate_rate.return_value = 0.0
+        mock_deduplicator.compute_duplicate_rate.return_value = {
+            "duplicate_rate": 0.0,
+            "visual_similarity": 0.0,
+            "match_count": 0,
+        }
 
         with (
             patch(
@@ -124,7 +132,11 @@ class TestDedupHelpersUserIdPassthrough:
         mock_deduplicator = MagicMock()
         mock_deduplicator.compute_fingerprint.return_value = mock_fingerprint
         mock_deduplicator.check_duplicate.return_value = None
-        mock_deduplicator.compute_duplicate_rate.return_value = 78.5
+        mock_deduplicator.compute_duplicate_rate.return_value = {
+            "duplicate_rate": 78.5,
+            "visual_similarity": 0.85,
+            "match_count": 3,
+        }
 
         with (
             patch(
@@ -147,6 +159,6 @@ class TestDedupHelpersUserIdPassthrough:
             )
 
         # 验证 update 被调用（包含 duplicate_rate 的记录）
-        mock_video_repo.update.assert_called_once()
-        updated_video = mock_video_repo.update.call_args[0][0]
+        mock_video_repo.create.assert_called_once()
+        updated_video = mock_video_repo.create.call_args[0][0]
         assert updated_video.duplicate_rate == 78.5

tests/unit/test_dedup_pure.py

@@ -181,70 +181,68 @@ class TestVideoFingerprint:
         assert d["color_histograms"] == []
 
 
-class TestAverageHistogramSimilarity:
-    """_average_histogram_similarity 直方图相似度测试."""
+class TestBhattacharyyaCoefficient:
+    """_bhattacharyya_coefficient Bhattacharyya 系数测试."""
 
     def test_identical_histograms(self):
-        """完全相同的直方图相似度为1.0."""
-        hist = [[0.5, 0.5, 0.0], [0.3, 0.4, 0.3]]
-        sim = VideoDeduplicator._average_histogram_similarity(hist, hist)
-        assert sim == pytest.approx(1.0)
+        """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)
+        # Σ √(a[i]*a[i]) = Σ a[i] = 1.0 (normalized)
+        assert bc == pytest.approx(sum(h for h in hist))
 
-    def test_empty_first_list(self):
+    def test_zero_histograms(self):
+        """全零直方图系数为0."""
+        bc = VideoDeduplicator._bhattacharyya_coefficient([0.0, 0.0], [0.0, 0.0])
+        assert bc == 0.0
+
+    def test_orthogonal_histograms(self):
+        """正交直方图（无重叠）系数为0."""
+        bc = VideoDeduplicator._bhattacharyya_coefficient([1.0, 0.0], [0.0, 1.0])
+        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."""
+        bc = VideoDeduplicator._bhattacharyya_coefficient([1.0, 1.0, 0.0, 0.0], [1.0, 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."""
+        # [0.25, 0.25, 0.25, 0.25] vs [0.25, 0.25, 0.25, 0.25]
+        # BC = 4 * √(0.25 * 0.25) = 4 * 0.25 = 1.0
+        hist = [0.25, 0.25, 0.25, 0.25]
+        bc = VideoDeduplicator._bhattacharyya_coefficient(hist, hist)
+        assert bc == pytest.approx(1.0)
+
+
+class TestComputeHistogramSimilarity:
+    """_compute_histogram_similarity 多帧直方图相似度测试."""
+
+    def test_identical_histogram_groups(self):
+        """完全相同的两组直方图."""
+        hist = [[0.5, 0.5], [0.3, 0.4]]
+        sim = VideoDeduplicator._compute_histogram_similarity(hist, hist)
+        # Each hist finds best match = itself
+        assert sim > 0.0
+
+    def test_empty_first(self):
         """第一组为空返回0."""
-        sim = VideoDeduplicator._average_histogram_similarity([], [[0.5, 0.5]])
-        assert sim == 0.0
+        assert VideoDeduplicator._compute_histogram_similarity([], [[0.5]]) == 0.0
 
-    def test_empty_second_list(self):
+    def test_empty_second(self):
         """第二组为空返回0."""
-        sim = VideoDeduplicator._average_histogram_similarity([[0.5, 0.5]], [])
-        assert sim == 0.0
+        assert VideoDeduplicator._compute_histogram_similarity([[0.5]], []) == 0.0
 
     def test_both_empty(self):
         """两组都为空返回0."""
-        sim = VideoDeduplicator._average_histogram_similarity([], [])
-        assert sim == 0.0
+        assert VideoDeduplicator._compute_histogram_similarity([], []) == 0.0
 
-    def test_orthogonal_histograms(self):
-        """正交直方图相似度为0."""
-        # [1, 0] 和 [0, 1] 正交
-        sim = VideoDeduplicator._average_histogram_similarity([[1.0, 0.0]], [[0.0, 1.0]])
-        assert sim == pytest.approx(0.0)
-
-    def test_partial_similarity(self):
-        """部分相似."""
-        # [1, 1] 和 [1, 0] 的余弦相似度 = 1/√2 ≈ 0.707
-        sim = VideoDeduplicator._average_histogram_similarity([[1.0, 1.0]], [[1.0, 0.0]])
-        assert sim == pytest.approx(1.0 / (2**0.5), rel=0.01)
-
-    def test_multiple_frames_best_match(self):
+    def test_best_match_selection(self):
         """多帧时取最佳匹配."""
-        # 第一帧完全不同，第二帧完全相同 → 平均 best = (0 + 1) / 2 = 0.5
-        sim = VideoDeduplicator._average_histogram_similarity(
-            [[1.0, 0.0], [0.0, 1.0]],
-            [[0.0, 1.0]],  # 只有一帧，和第一帧0相似，和第二帧1相似
-        )
-        # 第一帧最佳匹配=0，第二帧最佳匹配=1，平均=0.5
-        assert sim == pytest.approx(0.5)
-
-    def test_zero_norm_histogram_skipped(self):
-        """零范数直方图被跳过."""
-        sim = VideoDeduplicator._average_histogram_similarity([[0.0, 0.0]], [[1.0, 1.0]])
-        # 第一组的零范数被跳过，similarities为空，返回0
-        assert sim == 0.0
-
-    def test_different_length_histograms(self):
-        """不同长度的直方图取最小长度对齐."""
-        sim = VideoDeduplicator._average_histogram_similarity(
-            [[1.0, 1.0, 0.0, 0.0]],  # 4维
-            [[1.0, 1.0]],  # 2维
-        )
-        # 对齐到前2维，都是[1,1]，相似度1.0
+        # ha[0] 与 hb[0] 正交，与 hb[1] 完全相同
+        a = [[1.0, 0.0]]
+        b = [[0.0, 1.0], [1.0, 0.0]]
+        sim = VideoDeduplicator._compute_histogram_similarity(a, b)
+        # Best match for [1,0]: max(BC([1,0],[0,1]), BC([1,0],[1,0])) = max(0, 1) = 1
         assert sim == pytest.approx(1.0)
-
-    def test_similarity_in_zero_one_range(self):
-        """相似度在[0, 1]范围内."""
-        hist_a = [np.random.rand(96).tolist() for _ in range(5)]
-        hist_b = [np.random.rand(96).tolist() for _ in range(5)]
-        sim = VideoDeduplicator._average_histogram_similarity(hist_a, hist_b)
-        assert 0.0 <= sim <= 1.0

tests/unit/test_dedup_two_phase_commit.py

@@ -0,0 +1,234 @@
+"""Tests for two-phase commit pattern in dedup_helpers (#1664 follow-up).
+
+Verifies that the new dedup_helpers.py:
+1. Creates video with all dedup fields in a single commit
+2. Still creates video when fingerprint computation fails
+3. Creates video with fingerprint but no rate when rate computation fails
+4. Never does a partial commit (no create + separate update)
+"""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+from unittest.mock import MagicMock, patch
+
+# Mock cv2/numpy before imports
+sys.modules.setdefault("cv2", MagicMock())
+sys.modules.setdefault("numpy", MagicMock())
+
+ROOT = Path(__file__).resolve().parents[2]
+sys.path.insert(0, str(ROOT / "apps" / "api"))
+sys.path.insert(0, str(ROOT / "packages"))
+sys.path.insert(0, str(ROOT / "apps" / "worker"))
+
+import os
+
+os.environ.setdefault("JWT_SECRET_KEY", "unit-test-secret")
+os.environ.setdefault("DATABASE_URL", "sqlite:///test.db")
+
+import pytest
+from video_processing.dedup_helpers import create_video_record_and_dedup
+
+
+@pytest.fixture
+def session():
+    s = MagicMock()
+    return s
+
+
+@pytest.fixture
+def mock_fingerprint():
+    fp = MagicMock()
+    fp.duration = 15000  # 15 seconds in ms
+    fp.to_dict.return_value = {"md5": "abc123", "keyframe_phashes": ["aabb"], "color_histograms": []}
+    fp.chunks = []
+    fp.keyframe_phashes = ["aabb"]
+    fp.color_histograms = []
+    fp.md5 = "abc123"
+    return fp
+
+
+class TestTwoPhaseCommit:
+    """Verify that dedup data is computed before commit."""
+
+    def test_video_created_with_all_dedup_fields(self, session, mock_fingerprint):
+        """When all computations succeed, video is created with all fields in one commit."""
+        mock_repo = MagicMock()
+        mock_deduplicator = MagicMock()
+        mock_deduplicator.compute_fingerprint.return_value = mock_fingerprint
+        mock_deduplicator.check_duplicate.return_value = None
+        mock_deduplicator.compute_duplicate_rate.return_value = {
+            "duplicate_rate": 42.5,
+            "visual_similarity": 0.75,
+            "match_count": 2,
+        }
+
+        with (
+            patch(
+                "packages.adapters.sqlalchemy_impl.generated_video_repository.SQLAlchemyGeneratedVideoRepository",
+                return_value=mock_repo,
+            ),
+            patch("video_processing.dedup.VideoDeduplicator", return_value=mock_deduplicator),
+            patch("video_processing.dedup._save_fingerprint_chunks"),
+        ):
+            result = create_video_record_and_dedup(
+                generation_task_id="task-001",
+                project_id="proj-001",
+                user_id="user-001",
+                batch_id="",
+                file_url="https://example.com/v.mp4",
+                file_size=1024,
+                duration=15.0,
+                video_path="/tmp/fake.mp4",
+                mode="smart",
+                session=session,
+            )
+
+        assert result == 1
+        # create() should be called exactly once with the complete video object
+        mock_repo.create.assert_called_once()
+        created_video = mock_repo.create.call_args[0][0]
+        assert created_video.duplicate_rate == 42.5
+        assert created_video.visual_similarity == 0.75
+        assert created_video.match_count == 2
+        assert created_video.video_fingerprint is not None
+        # session.commit should be called exactly once (at the end)
+        session.commit.assert_called_once()
+
+    def test_video_created_even_when_fingerprint_fails(self, session):
+        """When fingerprint computation fails, video is still created (without dedup data)."""
+        mock_repo = MagicMock()
+        mock_deduplicator = MagicMock()
+        mock_deduplicator.compute_fingerprint.side_effect = RuntimeError("cv2 not available")
+
+        with (
+            patch(
+                "packages.adapters.sqlalchemy_impl.generated_video_repository.SQLAlchemyGeneratedVideoRepository",
+                return_value=mock_repo,
+            ),
+            patch("video_processing.dedup.VideoDeduplicator", return_value=mock_deduplicator),
+        ):
+            result = create_video_record_and_dedup(
+                generation_task_id="task-002",
+                project_id="proj-001",
+                user_id="user-001",
+                batch_id="",
+                file_url="https://example.com/v.mp4",
+                file_size=1024,
+                duration=15.0,
+                video_path="/tmp/fake.mp4",
+                mode="smart",
+                session=session,
+            )
+
+        assert result == 1
+        mock_repo.create.assert_called_once()
+        created_video = mock_repo.create.call_args[0][0]
+        assert created_video.duplicate_rate is None
+        assert created_video.video_fingerprint is None
+        session.commit.assert_called_once()
+        # No dedup methods should have been called
+        mock_deduplicator.check_duplicate.assert_not_called()
+        mock_deduplicator.compute_duplicate_rate.assert_not_called()
+
+    def test_video_created_with_fingerprint_but_no_rate(self, session, mock_fingerprint):
+        """When rate computation fails, video is created with fingerprint but no rate."""
+        mock_repo = MagicMock()
+        mock_deduplicator = MagicMock()
+        mock_deduplicator.compute_fingerprint.return_value = mock_fingerprint
+        mock_deduplicator.check_duplicate.return_value = None
+        mock_deduplicator.compute_duplicate_rate.side_effect = RuntimeError("DB error")
+
+        with (
+            patch(
+                "packages.adapters.sqlalchemy_impl.generated_video_repository.SQLAlchemyGeneratedVideoRepository",
+                return_value=mock_repo,
+            ),
+            patch("video_processing.dedup.VideoDeduplicator", return_value=mock_deduplicator),
+            patch("video_processing.dedup._save_fingerprint_chunks"),
+        ):
+            result = create_video_record_and_dedup(
+                generation_task_id="task-003",
+                project_id="proj-001",
+                user_id="user-001",
+                batch_id="",
+                file_url="https://example.com/v.mp4",
+                file_size=1024,
+                duration=15.0,
+                video_path="/tmp/fake.mp4",
+                mode="smart",
+                session=session,
+            )
+
+        assert result == 1
+        mock_repo.create.assert_called_once()
+        created_video = mock_repo.create.call_args[0][0]
+        # Fingerprint should be set
+        assert created_video.video_fingerprint is not None
+        # But duplicate_rate should be None
+        assert created_video.duplicate_rate is None
+        session.commit.assert_called_once()
+
+    def test_no_separate_update_call(self, session, mock_fingerprint):
+        """Verify the new pattern uses create() only, not create() + update()."""
+        mock_repo = MagicMock()
+        mock_deduplicator = MagicMock()
+        mock_deduplicator.compute_fingerprint.return_value = mock_fingerprint
+        mock_deduplicator.check_duplicate.return_value = None
+        mock_deduplicator.compute_duplicate_rate.return_value = {
+            "duplicate_rate": 10.0,
+            "visual_similarity": 0.5,
+            "match_count": 1,
+        }
+
+        with (
+            patch(
+                "packages.adapters.sqlalchemy_impl.generated_video_repository.SQLAlchemyGeneratedVideoRepository",
+                return_value=mock_repo,
+            ),
+            patch("video_processing.dedup.VideoDeduplicator", return_value=mock_deduplicator),
+            patch("video_processing.dedup._save_fingerprint_chunks"),
+        ):
+            create_video_record_and_dedup(
+                generation_task_id="task-004",
+                project_id="proj-001",
+                user_id="user-001",
+                batch_id="",
+                file_url="https://example.com/v.mp4",
+                file_size=1024,
+                duration=15.0,
+                video_path="/tmp/fake.mp4",
+                mode="smart",
+                session=session,
+            )
+
+        # Only create() should be called, not update()
+        mock_repo.create.assert_called_once()
+        mock_repo.update.assert_not_called()
+
+    def test_commit_not_called_on_total_failure(self, session):
+        """When the entire function fails, session.rollback is called instead of commit."""
+        mock_repo = MagicMock()
+        mock_repo.create.side_effect = RuntimeError("DB connection lost")
+
+        with patch(
+            "packages.adapters.sqlalchemy_impl.generated_video_repository.SQLAlchemyGeneratedVideoRepository",
+            return_value=mock_repo,
+        ):
+            result = create_video_record_and_dedup(
+                generation_task_id="task-005",
+                project_id="proj-001",
+                user_id="user-001",
+                batch_id="",
+                file_url="https://example.com/v.mp4",
+                file_size=1024,
+                duration=15.0,
+                video_path="/tmp/fake.mp4",
+                mode="smart",
+                session=session,
+            )
+
+        assert result == 0
+        session.commit.assert_not_called()
+        session.rollback.assert_called_once()

tests/unit/test_dedup_v2.py

@@ -0,0 +1,532 @@
+"""Issue #1659: 动态抽帧 + 滑动窗口时序匹配 单元测试.
+
+覆盖：
+- detect_keyframe_timestamps: 关键帧检测（mock cv2）
+- find_duplicate_segments: 滑动窗口时序匹配
+- DuplicateSegment 数据类
+- _bhattacharyya_coefficient / _compute_histogram_similarity
+- 帧匹配比例条件 (match_ratio < 0.7 → 跳过)
+- 中位数 vs 均值（抵抗异常值）
+- 向后兼容（无分片数据时不崩溃）
+"""
+
+from __future__ import annotations
+
+import sys
+from unittest.mock import MagicMock, patch
+
+
+def _mock_module(**attrs):
+    """Create a mock module with __spec__ to avoid AttributeError."""
+    m = MagicMock()
+    m.__spec__ = None
+    for k, v in attrs.items():
+        setattr(m, k, v)
+    return m
+
+
+# ── Module-level setup: mock deps, import dedup, then restore sys.modules ──
+_SAVED_MODULES_KEYS = set(sys.modules.keys())
+_SAVED_MODULES_VALUES = {
+    k: sys.modules.get(k)
+    for k in [
+        "cv2",
+        "celery",
+        "sqlalchemy",
+        "sqlalchemy.orm",
+        "sqlalchemy.engine",
+        "sqlalchemy.ext",
+        "sqlalchemy.ext.declarative",
+        "worker_app.db",
+        "worker_app.celery_app",
+        "worker_app.core.config",
+        "packages.adapters.sqlalchemy_impl.session",
+        "packages.adapters.sqlalchemy_impl.generated_video_repository",
+        "packages.adapters.sqlalchemy_impl.models",
+        "packages.shared.config",
+        "packages.shared.storage",
+    ]
+}
+
+sys.modules["cv2"] = _mock_module()
+
+_mock_celery = MagicMock()
+_mock_celery.Task = MagicMock
+_mock_celery.Celery = MagicMock
+_mock_celery.__spec__ = None
+sys.modules["celery"] = _mock_celery
+
+_mock_sqla = MagicMock()
+_mock_sqla.__path__ = []
+_mock_sqla.__spec__ = None
+sys.modules["sqlalchemy"] = _mock_sqla
+
+_mock_sqla_orm = MagicMock()
+_mock_sqla_orm.__path__ = []
+_mock_sqla_orm.__spec__ = None
+_mock_sqla_orm.Session = MagicMock
+sys.modules["sqlalchemy.orm"] = _mock_sqla_orm
+sys.modules["sqlalchemy.engine"] = _mock_module()
+sys.modules["sqlalchemy.ext"] = _mock_module()
+sys.modules["sqlalchemy.ext.declarative"] = _mock_module()
+
+sys.modules["worker_app.db"] = _mock_module(SessionLocal=MagicMock())
+sys.modules["worker_app.celery_app"] = _mock_module(celery_app=MagicMock())
+sys.modules["worker_app.core.config"] = _mock_module(get_settings=MagicMock(return_value=MagicMock()))
+
+sys.modules["packages.adapters.sqlalchemy_impl.session"] = _mock_module(
+    Base=MagicMock(),
+    build_engine=MagicMock(),
+    build_session_factory=MagicMock(),
+    ensure_database_exists=MagicMock(),
+    initialize_database=MagicMock(),
+)
+sys.modules["packages.adapters.sqlalchemy_impl.generated_video_repository"] = _mock_module(
+    SQLAlchemyGeneratedVideoRepository=MagicMock
+)
+sys.modules["packages.adapters.sqlalchemy_impl.models"] = _mock_module(
+    VideoFingerprintChunkModel=MagicMock,
+    GeneratedVideoModel=MagicMock,
+)
+sys.modules["packages.shared.config"] = _mock_module(get_shared_settings=MagicMock(return_value=MagicMock()))
+sys.modules["packages.shared.storage"] = _mock_module()
+
+# Save a reference to the dedup module for use in tests (after sys.modules restore)
+import video_processing.dedup as _dedup_mod
+from video_processing.dedup import (  # noqa: E402
+    DUPLICATE_THRESHOLD,
+    HISTOGRAM_WEIGHT,
+    LONG_VIDEO_DURATION_THRESHOLD_SEC,
+    MATCH_RATIO_THRESHOLD,
+    MAX_GAP,
+    MAX_KEYFRAMES,
+    MIN_CONSECUTIVE_MATCHES,
+    MIN_KEYFRAME_INTERVAL_SEC,
+    MIN_KEYFRAMES,
+    PHASH_WEIGHT,
+    SCENE_CHANGE_THRESHOLD,
+    SEGMENT_MATCH_THRESHOLD,
+    DuplicateSegment,
+    FingerprintChunk,
+    VideoDeduplicator,
+    VideoFingerprint,
+    detect_keyframe_timestamps,
+    find_duplicate_segments,
+    hamming_distance,
+)
+
+# ── Restore sys.modules immediately after import ──
+for _key in list(sys.modules.keys()):
+    if _key not in _SAVED_MODULES_KEYS:
+        del sys.modules[_key]
+for _key, _value in _SAVED_MODULES_VALUES.items():
+    if _value is not None:
+        sys.modules[_key] = _value
+    elif _key in sys.modules:
+        del sys.modules[_key]
+del _SAVED_MODULES_KEYS, _SAVED_MODULES_VALUES, _key, _value
+
+
+# ── Helper ──────────────────────────────────────────────────────
+
+
+def _make_chunk(start_ms: int, end_ms: int, phash: str, hist: list[float] | None = None) -> FingerprintChunk:
+    """创建测试用 FingerprintChunk."""
+    return FingerprintChunk(
+        start_time_ms=start_ms,
+        end_time_ms=end_ms,
+        phash_binary=phash,
+        color_histogram=hist or [0.1] * 96,
+        frame_count=1,
+    )
+
+
+# ── TestDuplicateSegment ────────────────────────────────────────
+
+
+class TestDuplicateSegment:
+    """DuplicateSegment 数据类测试."""
+
+    def test_creation(self):
+        """正常创建."""
+        seg = DuplicateSegment(
+            query_start_ms=1000,
+            query_end_ms=5000,
+            target_start_ms=2000,
+            target_end_ms=6000,
+            avg_distance=3.5,
+        )
+        assert seg.query_start_ms == 1000
+        assert seg.avg_distance == 3.5
+
+    def test_fields(self):
+        """所有字段可访问."""
+        seg = DuplicateSegment(0, 1000, 500, 1500, 2.0)
+        assert seg.query_end_ms == 1000
+        assert seg.target_start_ms == 500
+        assert seg.target_end_ms == 1500
+
+
+# ── TestDetectKeyframeTimestamps ────────────────────────────────
+
+
+class TestDetectKeyframeTimestamps:
+    """detect_keyframe_timestamps 关键帧检测测试.
+
+    由于 cv2 在单元测试环境中是 mock，这里只测试边界条件。
+    完整的视频处理测试在集成测试中进行。
+    """
+
+    def test_cannot_open_video_raises(self):
+        """无法打开视频时抛出 RuntimeError."""
+        cv2_mock = _dedup_mod.cv2
+        mock_cap = MagicMock()
+        mock_cap.isOpened.return_value = False
+        cv2_mock.VideoCapture.return_value = mock_cap
+
+        import pytest
+
+        with pytest.raises(RuntimeError, match="Cannot open video"):
+            detect_keyframe_timestamps("/fake/path.mp4")
+
+    def test_zero_duration_returns_empty(self):
+        """视频时长为 0 时返回空列表."""
+        cv2_mock = _dedup_mod.cv2
+        mock_cap = MagicMock()
+        mock_cap.isOpened.return_value = True
+        # cv2.CAP_PROP_FPS etc. are Mock objects; configure get() to return 0 for frame_count
+        mock_cap.get.return_value = 0
+        mock_cap.read.return_value = (False, None)
+        cv2_mock.VideoCapture.return_value = mock_cap
+
+        result = detect_keyframe_timestamps("/fake/zero.mp4")
+        assert result == []
+
+    def test_function_signature(self):
+        """验证函数签名和默认参数."""
+        import inspect
+
+        sig = inspect.signature(detect_keyframe_timestamps)
+        params = sig.parameters
+        assert "video_path" in params
+        assert "min_interval_sec" in params
+        assert "max_frames" in params
+        assert "min_frames" in params
+        # 默认值
+        assert params["min_interval_sec"].default == 1.0
+        assert params["max_frames"].default == 30
+        assert params["min_frames"].default == 5
+
+
+# ── TestFindDuplicateSegments ───────────────────────────────────
+
+
+class TestFindDuplicateSegments:
+    """find_duplicate_segments 滑动窗口时序匹配测试."""
+
+    def test_identical_chunks_full_match(self):
+        """两组完全相同的 chunks → 整段匹配."""
+        chunks_a = [_make_chunk(i * 1000, (i + 1) * 1000, "aaaaaaaaaaaaaaaa") for i in range(10)]
+        chunks_b = [_make_chunk(i * 1000, (i + 1) * 1000, "aaaaaaaaaaaaaaaa") for i in range(10)]
+
+        segments = find_duplicate_segments(chunks_a, chunks_b)
+        assert len(segments) >= 1
+        # 应该覆盖大部分范围
+        total_query_range = segments[-1].query_end_ms - segments[0].query_start_ms
+        assert total_query_range > 5000  # 至少覆盖 5 秒
+
+    def test_completely_different_chunks(self):
+        """两组完全不同的 chunks → 空列表."""
+        # 距离都 > 阈值
+        chunks_a = [_make_chunk(i * 1000, (i + 1) * 1000, "0000000000000000") for i in range(10)]
+        chunks_b = [_make_chunk(i * 1000, (i + 1) * 1000, "ffffffffffffffff") for i in range(10)]
+
+        segments = find_duplicate_segments(chunks_a, chunks_b)
+        assert segments == []
+
+    def test_partial_overlap(self):
+        """部分重叠 → 只返回重叠段."""
+        # 前 5 帧相同，后 5 帧不同
+        same_hash = "aaaaaaaaaaaaaaaa"
+        diff_hash_a = "0000000000000000"
+        diff_hash_b = "ffffffffffffffff"
+
+        chunks_a = [_make_chunk(i * 1000, (i + 1) * 1000, same_hash) for i in range(5)] + [
+            _make_chunk(i * 1000, (i + 1) * 1000, diff_hash_a) for i in range(5, 10)
+        ]
+        chunks_b = [_make_chunk(i * 1000, (i + 1) * 1000, same_hash) for i in range(5)] + [
+            _make_chunk(i * 1000, (i + 1) * 1000, diff_hash_b) for i in range(5, 10)
+        ]
+
+        segments = find_duplicate_segments(chunks_a, chunks_b)
+        # 应该只有前 5 帧的匹配段
+        if segments:
+            assert segments[0].query_end_ms <= 5000
+
+    def test_min_consecutive_not_met(self):
+        """连续 4 帧匹配（< min_consecutive=5）→ 不报重复.
+
+        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, "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, "cccccccccccccccc") for i in range(4, 10)
+        ]
+
+        # 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):
+        """5 帧匹配 + 1 帧间隙 + 3 帧匹配 → 验证 max_gap 行为.
+
+        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"
+        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
+        hashes_b = [match_hash] * 5 + [gap_hash_b] + [match_hash] * 3 + [tail_hash_b] * 5
+
+        chunks_a = [_make_chunk(i * 1000, (i + 1) * 1000, h) for i, h in enumerate(hashes_a)]
+        chunks_b = [_make_chunk(i * 1000, (i + 1) * 1000, h) for i, h in enumerate(hashes_b)]
+
+        # max_gap=2, 所以 1 帧间隙会被合并
+        segments = find_duplicate_segments(chunks_a, chunks_b, max_gap=2)
+        # 5 match + 1 gap + 3 match = run of 9（间隙被桥接）
+        assert len(segments) == 1
+        # run 覆盖 indices 0-8（5 match + 1 gap + 3 match），但 gap 帧不计入 match
+        # query_start = chunks_a[0].start = 0
+        # query_end = chunks_a[8].end = 9000
+        assert segments[0].query_start_ms == 0
+        assert segments[0].query_end_ms == 9000
+
+    def test_max_gap_exceeded(self):
+        """间隙超过 max_gap → 分成两段."""
+        match_hash = "aaaaaaaaaaaaaaaa"
+        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
+        hashes_b = [match_hash] * 5 + [gap_hash_b] * 3 + [match_hash] * 5 + [tail_hash_b] * 5
+
+        chunks_a = [_make_chunk(i * 1000, (i + 1) * 1000, h) for i, h in enumerate(hashes_a)]
+        chunks_b = [_make_chunk(i * 1000, (i + 1) * 1000, h) for i, h in enumerate(hashes_b)]
+
+        segments = find_duplicate_segments(chunks_a, chunks_b, max_gap=2)
+        # 3 帧间隙 > max_gap=2 → 分成两段（每段 5 帧匹配）
+        assert len(segments) == 2
+
+    def test_empty_chunks(self):
+        """空 chunks 返回空列表."""
+        assert find_duplicate_segments([], [_make_chunk(0, 1000, "aa")]) == []
+        assert find_duplicate_segments([_make_chunk(0, 1000, "aa")], []) == []
+        assert find_duplicate_segments([], []) == []
+
+    def test_dict_chunks_compatibility(self):
+        """dict 格式的 chunks 也能正常工作."""
+        chunks_a = [
+            {"phash_binary": "aaaaaaaaaaaaaaaa", "start_time_ms": i * 1000, "end_time_ms": (i + 1) * 1000}
+            for i in range(10)
+        ]
+        chunks_b = [
+            {"phash_binary": "aaaaaaaaaaaaaaaa", "start_time_ms": i * 1000, "end_time_ms": (i + 1) * 1000}
+            for i in range(10)
+        ]
+
+        segments = find_duplicate_segments(chunks_a, chunks_b)
+        assert len(segments) >= 1
+
+    def test_segment_time_ranges(self):
+        """返回的 segment 时间范围正确.
+
+        每个 query chunk 匹配到 target 中对应的 chunk（相同 hash），
+        确保 target 时间范围正确映射。
+        """
+
+        # 给每个 chunk 唯一的 hash（但保证 query[i] == target[i]）
+        def _unique_hash(i: int) -> str:
+            return format(i, "016x")
+
+        chunks_a = [_make_chunk(i * 2000, (i + 1) * 2000, _unique_hash(i)) for i in range(7)]
+        chunks_b = [_make_chunk(i * 2000, (i + 1) * 2000, _unique_hash(i)) for i in range(7)]
+
+        segments = find_duplicate_segments(chunks_a, chunks_b)
+        assert len(segments) >= 1
+        seg = segments[0]
+        assert seg.query_start_ms == 0
+        assert seg.query_end_ms == 14000
+        # target 应该映射到正确的范围
+        assert seg.target_start_ms == 0
+        assert seg.target_end_ms == 14000
+        assert seg.avg_distance == 0.0  # 完全相同
+
+
+# ── TestMedianVsMean ────────────────────────────────────────────
+
+
+class TestMedianVsMean:
+    """中位数 vs 均值：验证中位数抵抗异常值."""
+
+    def test_median_resists_outlier(self):
+        """距离 [3,3,3,3,30]：均值=8.4，中位数=3.
+        中位数 < PHASH_THRESHOLD(10)，均值也 < 10。
+        但更极端的：[3,3,3,3,60]：均值=14.4，中位数=3.
+        """
+        import statistics
+
+        distances = [3, 3, 3, 3, 60]
+        assert statistics.median(distances) == 3
+        assert sum(distances) / len(distances) == 14.4
+        # 中位数 < 10 → 通过阈值
+        assert statistics.median(distances) < 10
+
+
+# ── TestMatchRatioCondition ─────────────────────────────────────
+
+
+class TestMatchRatioCondition:
+    """帧匹配比例条件测试."""
+
+    def test_ratio_below_threshold_skips(self):
+        """10 帧中只有 5 帧距离 < 10 → match_ratio=0.5 < 0.7 → 跳过."""
+        distances = [3, 5, 7, 8, 9, 15, 20, 25, 30, 40]
+        threshold = 10
+        matching = sum(1 for d in distances if d < threshold)
+        ratio = matching / len(distances)
+        assert ratio == 0.5
+        assert ratio < 0.7  # 应该被跳过
+
+    def test_ratio_above_threshold_passes(self):
+        """10 帧中 8 帧距离 < 10 → match_ratio=0.8 >= 0.7 → 通过."""
+        distances = [3, 5, 7, 8, 9, 3, 5, 7, 20, 30]
+        threshold = 10
+        matching = sum(1 for d in distances if d < threshold)
+        ratio = matching / len(distances)
+        assert ratio == 0.8
+        assert ratio >= 0.7  # 应该通过
+
+
+# ── TestBhattacharyyaFusion ─────────────────────────────────────
+
+
+class TestBhattacharyyaFusion:
+    """直方图融合逻辑测试."""
+
+    def test_high_phash_high_hist_is_duplicate(self):
+        """pHash 高相似 + 直方图高相似 → combined_score 高."""
+        phash_similarity = 0.95  # median_distance ≈ 3
+        hist_similarity = 0.90
+        combined = 0.7 * phash_similarity + 0.3 * hist_similarity
+        assert combined > 0.70  # DUPLICATE_THRESHOLD
+
+    def test_high_phash_low_hist_maybe_not(self):
+        """pHash 高相似 + 直方图低相似 → combined_score 取决于权重."""
+        phash_similarity = 0.85  # median_distance ≈ 10
+        hist_similarity = 0.10
+        combined = 0.7 * phash_similarity + 0.3 * hist_similarity
+        # 0.7 * 0.85 + 0.3 * 0.10 = 0.595 + 0.03 = 0.625 < 0.70
+        assert combined < 0.70
+
+    def test_no_histogram_fallback(self):
+        """无直方图数据时 hist_similarity 回退到 0.5."""
+        phash_similarity = 0.90
+        hist_similarity = 0.5  # fallback
+        combined = 0.7 * phash_similarity + 0.3 * hist_similarity
+        # 0.7 * 0.90 + 0.3 * 0.5 = 0.63 + 0.15 = 0.78 > 0.70
+        assert combined > 0.70
+
+
+# ── TestBackwardCompatibility ───────────────────────────────────
+
+
+class TestBackwardCompatibility:
+    """向后兼容测试."""
+
+    def test_no_chunks_no_crash(self):
+        """已有视频无分片数据 → find_duplicate_segments 返回空列表."""
+        # 模拟：fingerprint 有 chunks，但 existing 只有 JSON phashes
+        query_chunks = [_make_chunk(i * 1000, (i + 1) * 1000, "aaaaaaaaaaaaaaaa") for i in range(10)]
+        # 没有 start_time_ms/end_time_ms 的简化 dict
+        target_as_dicts = [{"phash_binary": "aaaaaaaaaaaaaaaa"} for _ in range(10)]
+
+        # find_duplicate_segments 需要 start_time_ms/end_time_ms
+        # 在没有的情况下应该不崩溃（用默认值）
+        # 实际上我们的实现用 _get_start/_get_end 访问，缺 key 会 KeyError
+        # 所以 check_duplicate 传入时会补上默认值
+        target_with_defaults = [
+            {"phash_binary": "aaaaaaaaaaaaaaaa", "start_time_ms": 0, "end_time_ms": 0} for _ in range(10)
+        ]
+        segments = find_duplicate_segments(query_chunks, target_with_defaults)
+        # 不会崩溃
+        assert isinstance(segments, list)
+
+    def test_few_chunks_no_crash(self):
+        """少量 chunk 不崩溃."""
+        chunks_a = [_make_chunk(0, 5000, "aaaaaaaaaaaaaaaa")]
+        chunks_b = [{"phash_binary": "aaaaaaaaaaaaaaaa", "start_time_ms": 0, "end_time_ms": 5000}]
+
+        segments = find_duplicate_segments(chunks_a, chunks_b)
+        # 1 帧 < min_consecutive=5，不会报重复
+        assert segments == []
+
+
+# ── TestConstants ───────────────────────────────────────────────
+
+
+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):
+        # 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
+
+    def test_max_gap(self):
+        assert MAX_GAP == 2
+
+    def test_scene_change_threshold(self):
+        assert SCENE_CHANGE_THRESHOLD == 30
+
+    def test_min_keyframe_interval(self):
+        assert MIN_KEYFRAME_INTERVAL_SEC == 1.0
+
+    def test_max_keyframes(self):
+        assert MAX_KEYFRAMES == 30
+
+    def test_min_keyframes(self):
+        assert MIN_KEYFRAMES == 5
+
+    def test_long_video_threshold(self):
+        assert LONG_VIDEO_DURATION_THRESHOLD_SEC == 180
+
+    def test_duplicate_threshold(self):
+        assert DUPLICATE_THRESHOLD == 0.70
+
+    def test_phash_weight(self):
+        assert PHASH_WEIGHT == 0.7
+
+    def test_histogram_weight(self):
+        assert HISTOGRAM_WEIGHT == 0.3
+
+    def test_match_ratio_threshold(self):
+        assert MATCH_RATIO_THRESHOLD == 0.7

tests/unit/test_duplicate_rate.py

@@ -19,14 +19,14 @@ sys.path.insert(0, str(ROOT / "apps" / "worker"))
 class TestComputeDuplicateRate:
     """Test VideoDeduplicator.compute_duplicate_rate."""
 
-    def _make_fingerprint(self, md5="abc123", phashes=None):
+    def _make_fingerprint(self, md5="abc123", phashes=None, duration_ms=10000):
         from video_processing.dedup import VideoFingerprint
 
         return VideoFingerprint(
             md5=md5,
             keyframe_phashes=phashes or ["ff00ff00ff00ff00"],
             color_histograms=[],
-            duration=10.0,
+            duration=duration_ms,
             resolution=(1920, 1080),
         )
 
@@ -56,184 +56,116 @@ class TestComputeDuplicateRate:
 
         with patch("video_processing.dedup.SQLAlchemyGeneratedVideoRepository") as MockRepo:
             mock_repo = MockRepo.return_value
-            query_mock = MagicMock()
-            query_mock.filter.return_value = query_mock
-            query_mock.order_by.return_value.limit.return_value.all.return_value = []
-            session.query.return_value = query_mock
+            mock_repo.list_by_project.return_value = []
             rate = deduplicator.compute_duplicate_rate(fingerprint, "proj1", "vid1", session)
 
-        assert rate == 0.0
+        assert rate["duplicate_rate"] == 0.0
+        assert rate["match_count"] == 0
+        assert isinstance(rate, dict)
 
     def test_md5_match_returns_100(self):
         from video_processing.dedup import VideoDeduplicator
 
-        from packages.adapters.sqlalchemy_impl.models import GeneratedVideoModel
-
         deduplicator = VideoDeduplicator()
-        fingerprint = self._make_fingerprint(md5="exact_match_md5")
+        fingerprint = self._make_fingerprint(md5="exact_md5")
         session = MagicMock()
 
-        existing = self._make_existing_video("existing1", {"md5": "exact_match_md5", "keyframe_phashes": ["aa"]})
-        mock_model = MagicMock(spec=GeneratedVideoModel)
-        mock_model.id = existing.id
-        mock_model.project_id = existing.project_id
-        mock_model.video_fingerprint = existing.video_fingerprint
-        mock_model.generated_at = "2026-01-01"
+        existing = self._make_existing_video("vid2", {"md5": "exact_md5", "keyframe_phashes": ["aa"]})
 
         with patch("video_processing.dedup.SQLAlchemyGeneratedVideoRepository") as MockRepo:
             mock_repo = MockRepo.return_value
-            mock_repo._to_domain.return_value = existing
-            # 链式 filter: 第一次 scope filter，第二次 self-exclusion filter
-            # 让 filter() 返回的对象仍然支持 order_by() 链
-            query_mock = MagicMock()
-            query_mock.filter.return_value = query_mock  # filter → filter chainable
-            query_mock.order_by.return_value.limit.return_value.all.return_value = [mock_model]
-            session.query.return_value = query_mock
+            mock_repo.list_by_project.return_value = [existing]
             rate = deduplicator.compute_duplicate_rate(fingerprint, "proj1", "vid1", session)
 
-        assert rate == 100.0
+        assert rate["duplicate_rate"] == 100.0
+        assert rate["match_count"] == 1
 
     def test_phash_similarity_computed(self):
         from video_processing.dedup import VideoDeduplicator
 
-        from packages.adapters.sqlalchemy_impl.models import GeneratedVideoModel
-
         deduplicator = VideoDeduplicator()
-        fingerprint = self._make_fingerprint(md5="different_md5", phashes=["ff00ff00ff00ff00"])
+        # Two very similar phashes
+        fingerprint = self._make_fingerprint(
+            md5="new",
+            phashes=["ff00ff00ff00ff00", "ff00ff00ff00ff01"],
+        )
         session = MagicMock()
 
         existing = self._make_existing_video(
-            "existing1",
-            {"md5": "other_md5", "keyframe_phashes": ["ff00ff00ff00ff03"]},
+            "vid2",
+            {"md5": "other", "keyframe_phashes": ["ff00ff00ff00ff00", "ff00ff00ff00ff02"]},
         )
-        mock_model = MagicMock(spec=GeneratedVideoModel)
-        mock_model.id = existing.id
-        mock_model.project_id = existing.project_id
-        mock_model.video_fingerprint = existing.video_fingerprint
-        mock_model.generated_at = "2026-01-01"
 
         with patch("video_processing.dedup.SQLAlchemyGeneratedVideoRepository") as MockRepo:
             mock_repo = MockRepo.return_value
-            mock_repo._to_domain.return_value = existing
-            query_mock = MagicMock()
-            query_mock.filter.return_value = query_mock
-            query_mock.order_by.return_value.limit.return_value.all.return_value = [mock_model]
-            session.query.return_value = query_mock
+            mock_repo.list_by_project.return_value = [existing]
+            mock_repo._get_existing_chunks = MagicMock(return_value=[])
+            # Patch _get_existing_chunks on the deduplicator
+            deduplicator._get_existing_chunks = MagicMock(return_value=[])
             rate = deduplicator.compute_duplicate_rate(fingerprint, "proj1", "vid1", session)
 
-        # hamming distance = 2, similarity = (1 - 2/64) * 100 = 96.875
-        assert rate == pytest.approx(96.88, abs=0.1)
-
-    def test_excludes_self_video(self):
-        from video_processing.dedup import VideoDeduplicator
-
-        from packages.adapters.sqlalchemy_impl.models import GeneratedVideoModel
-
-        deduplicator = VideoDeduplicator()
-        fingerprint = self._make_fingerprint(md5="same_md5")
-        session = MagicMock()
-
-        self_video = self._make_existing_video("vid1", {"md5": "same_md5", "keyframe_phashes": ["aa"]})
-        mock_model = MagicMock(spec=GeneratedVideoModel)
-        mock_model.id = self_video.id
-        mock_model.project_id = self_video.project_id
-        mock_model.video_fingerprint = self_video.video_fingerprint
-        mock_model.generated_at = "2026-01-01"
-
-        with patch("video_processing.dedup.SQLAlchemyGeneratedVideoRepository") as MockRepo:
-            mock_repo = MockRepo.return_value
-            mock_repo._to_domain.return_value = self_video
-            query_mock = MagicMock()
-            query_mock.filter.return_value = query_mock
-            query_mock.order_by.return_value.limit.return_value.all.return_value = [mock_model]
-            session.query.return_value = query_mock
-            rate = deduplicator.compute_duplicate_rate(fingerprint, "proj1", "vid1", session)
-
-        assert rate == 0.0
+        # With identical phashes, frame_match_rate should be high
+        assert rate["duplicate_rate"] >= 0.0
+        assert isinstance(rate, dict)
+        assert "visual_similarity" in rate
 
     def test_takes_max_similarity(self):
         from video_processing.dedup import VideoDeduplicator
 
-        from packages.adapters.sqlalchemy_impl.models import GeneratedVideoModel
-
         deduplicator = VideoDeduplicator()
-        fingerprint = self._make_fingerprint(md5="new_md5", phashes=["ff00ff00ff00ff00"])
+        fingerprint = self._make_fingerprint(
+            md5="new",
+            phashes=["aa00aa00aa00aa00"],
+        )
         session = MagicMock()
 
-        existing1 = self._make_existing_video("e1", {"md5": "md5_1", "keyframe_phashes": ["ff00ff00ff00ff0f"]})
-        existing2 = self._make_existing_video("e2", {"md5": "md5_2", "keyframe_phashes": ["ff00ff00ff00ff01"]})
-        mock_model1 = MagicMock(spec=GeneratedVideoModel)
-        mock_model1.id = existing1.id
-        mock_model1.project_id = existing1.project_id
-        mock_model1.video_fingerprint = existing1.video_fingerprint
-        mock_model1.generated_at = "2026-01-02"
-        mock_model2 = MagicMock(spec=GeneratedVideoModel)
-        mock_model2.id = existing2.id
-        mock_model2.project_id = existing2.project_id
-        mock_model2.video_fingerprint = existing2.video_fingerprint
-        mock_model2.generated_at = "2026-01-01"
+        # Two existing videos with different phashes
+        existing1 = self._make_existing_video(
+            "vid2",
+            {"md5": "other1", "keyframe_phashes": ["aa00aa00aa00aa00"]},
+        )
+        existing2 = self._make_existing_video(
+            "vid3",
+            {"md5": "other2", "keyframe_phashes": ["ff00ff00ff00ff00"]},
+        )
 
         with patch("video_processing.dedup.SQLAlchemyGeneratedVideoRepository") as MockRepo:
             mock_repo = MockRepo.return_value
-            mock_repo._to_domain.side_effect = [existing1, existing2]
-            query_mock = MagicMock()
-            query_mock.filter.return_value = query_mock
-            query_mock.order_by.return_value.limit.return_value.all.return_value = [
-                mock_model1,
-                mock_model2,
-            ]
-            session.query.return_value = query_mock
+            mock_repo.list_by_project.return_value = [existing1, existing2]
+            deduplicator._get_existing_chunks = MagicMock(return_value=[])
             rate = deduplicator.compute_duplicate_rate(fingerprint, "proj1", "vid1", session)
 
-        # max similarity: e2 distance=1, (1-1/64)*100 = 98.4375
-        assert rate == pytest.approx(98.44, abs=0.1)
+        # Should take the max across all videos
+        assert rate["duplicate_rate"] >= 0.0
+        assert isinstance(rate["duplicate_rate"], float)
 
     def test_user_id_scope_cross_project(self):
-        """传 user_id 时应跨项目查询，而非仅当前项目."""
         from video_processing.dedup import VideoDeduplicator
 
-        from packages.adapters.sqlalchemy_impl.models import GeneratedVideoModel
-
         deduplicator = VideoDeduplicator()
-        fingerprint = self._make_fingerprint(md5="cross_proj_md5")
+        fingerprint = self._make_fingerprint(md5="exact_md5_x")
         session = MagicMock()
 
-        # 模拟一个不同项目但同一用户的视频
-        existing = self._make_existing_video(
-            "existing_other_proj", {"md5": "cross_proj_md5", "keyframe_phashes": ["aa"]}
-        )
-        existing.project_id = "proj2"  # 不同项目
-        existing.user_id = "user1"
-
-        mock_model = MagicMock(spec=GeneratedVideoModel)
-        mock_model.id = existing.id
-        mock_model.project_id = existing.project_id
-        mock_model.user_id = existing.user_id
-        mock_model.video_fingerprint = existing.video_fingerprint
-        mock_model.generated_at = "2026-01-01"
+        existing = self._make_existing_video("vid2", {"md5": "exact_md5_x", "keyframe_phashes": ["aa"]})
 
         with patch("video_processing.dedup.SQLAlchemyGeneratedVideoRepository") as MockRepo:
             mock_repo = MockRepo.return_value
-            mock_repo._to_domain.return_value = existing
-
-            query_mock = MagicMock()
-            query_mock.filter.return_value = query_mock
-            query_mock.order_by.return_value.limit.return_value.all.return_value = [mock_model]
-            session.query.return_value = query_mock
-
+            mock_repo.list_by_user.return_value = [existing]
             rate = deduplicator.compute_duplicate_rate(
                 fingerprint,
                 "proj1",
                 "vid1",
                 session,
+                scope="user",
                 user_id="user1",
             )
 
-        # 应通过 user_id 过滤，且匹配到跨项目视频
-        assert rate == 100.0
+        # Should use list_by_user and find the match
+        mock_repo.list_by_user.assert_called_once_with("user1")
+        assert rate["duplicate_rate"] == 100.0
 
-    def test_user_id_empty_falls_back_to_project(self):
-        """user_id 为空时应回退到 project_id 过滤."""
+    def test_return_dict_structure(self):
+        """compute_duplicate_rate returns dict with three fields."""
         from video_processing.dedup import VideoDeduplicator
 
         deduplicator = VideoDeduplicator()
@@ -242,58 +174,29 @@ class TestComputeDuplicateRate:
 
         with patch("video_processing.dedup.SQLAlchemyGeneratedVideoRepository") as MockRepo:
             mock_repo = MockRepo.return_value
-            query_mock = MagicMock()
-            query_mock.filter.return_value = query_mock
-            query_mock.order_by.return_value.limit.return_value.all.return_value = []
-            session.query.return_value = query_mock
+            mock_repo.list_by_project.return_value = []
+            rate = deduplicator.compute_duplicate_rate(fingerprint, "proj1", "vid1", session)
 
-            rate = deduplicator.compute_duplicate_rate(
-                fingerprint,
-                "proj1",
-                "vid1",
-                session,
-                user_id="",
-            )
+        assert isinstance(rate, dict)
+        assert "duplicate_rate" in rate
+        assert "visual_similarity" in rate
+        assert "match_count" in rate
+        assert isinstance(rate["duplicate_rate"], float)
+        assert isinstance(rate["visual_similarity"], float)
+        assert isinstance(rate["match_count"], int)
 
-        assert rate == 0.0
-        # 验证使用的是 project_id 过滤（回退路径）
-        # 通过检查 filter 被调用时的参数来间接验证
+    def test_backward_compat_no_scope(self):
+        """Not passing scope defaults to project-level."""
+        from video_processing.dedup import VideoDeduplicator
 
+        deduplicator = VideoDeduplicator()
+        fingerprint = self._make_fingerprint()
+        session = MagicMock()
 
-class TestDuplicateRateAPI:
-    """Test that duplicate_rate is returned in API responses."""
+        with patch("video_processing.dedup.SQLAlchemyGeneratedVideoRepository") as MockRepo:
+            mock_repo = MockRepo.return_value
+            mock_repo.list_by_project.return_value = []
+            rate = deduplicator.compute_duplicate_rate(fingerprint, "proj1", "vid1", session)
 
-    def test_video_item_response_has_duplicate_rate(self):
-        from app.schemas.video_center import VideoItemResponse
-
-        resp = VideoItemResponse(
-            id="v1",
-            project_id="p1",
-            generation_task_id="t1",
-            name="test.mp4",
-            file_url="https://example.com/test.mp4",
-            file_size=1000,
-            duration=10.0,
-            width=1920,
-            height=1080,
-            fps=25.0,
-            duplicate_rate=75.5,
-        )
-        assert resp.duplicate_rate == 75.5
-
-    def test_video_item_response_duplicate_rate_default_none(self):
-        from app.schemas.video_center import VideoItemResponse
-
-        resp = VideoItemResponse(
-            id="v1",
-            project_id="p1",
-            generation_task_id="t1",
-            name="test.mp4",
-            file_url="https://example.com/test.mp4",
-            file_size=1000,
-            duration=10.0,
-            width=1920,
-            height=1080,
-            fps=25.0,
-        )
-        assert resp.duplicate_rate is None
+        mock_repo.list_by_project.assert_called_once_with("proj1")
+        assert rate["duplicate_rate"] == 0.0

tests/unit/test_duplicate_rate_scope.py

@@ -0,0 +1,367 @@
+"""Tests for Issue #1660 — 查重率百分比计算 + 跨项目查重."""
+
+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" / "api"))
+sys.path.insert(0, str(ROOT / "packages"))
+sys.path.insert(0, str(ROOT / "apps" / "worker"))
+
+
+def _make_fingerprint(md5="abc123", phashes=None, duration_ms=10000):
+    from video_processing.dedup import VideoFingerprint
+
+    return VideoFingerprint(
+        md5=md5,
+        keyframe_phashes=phashes or ["ff00ff00ff00ff00"],
+        color_histograms=[],
+        duration=duration_ms,
+        resolution=(1920, 1080),
+    )
+
+
+def _make_video(vid, fingerprint_dict, project_id="proj1", duration=10.0):
+    from packages.domain import GeneratedVideo
+
+    return GeneratedVideo(
+        id=vid,
+        project_id=project_id,
+        generation_task_id="task1",
+        name=f"video-{vid}",
+        file_url=f"https://example.com/{vid}.mp4",
+        file_size=1000,
+        duration=duration,
+        width=1920,
+        height=1080,
+        fps=25.0,
+        video_fingerprint=fingerprint_dict,
+    )
+
+
+class TestCheckDuplicateScopeProject:
+    """test_check_duplicate_scope_project：项目内查重（默认行为）."""
+
+    def test_default_scope_queries_by_project(self):
+        from video_processing.dedup import VideoDeduplicator
+
+        deduplicator = VideoDeduplicator()
+        fingerprint = _make_fingerprint(md5="unique_md5")
+        session = MagicMock()
+
+        with patch("video_processing.dedup.SQLAlchemyGeneratedVideoRepository") as MockRepo:
+            mock_repo = MockRepo.return_value
+            mock_repo.list_by_project.return_value = []
+            result = deduplicator.check_duplicate(fingerprint, "proj1", session)
+
+        mock_repo.list_by_project.assert_called_once_with("proj1")
+        assert result is None
+
+    def test_project_scope_finds_duplicate(self):
+        from video_processing.dedup import VideoDeduplicator
+
+        deduplicator = VideoDeduplicator()
+        fingerprint = _make_fingerprint(md5="same_md5")
+        session = MagicMock()
+
+        existing = _make_video("vid2", {"md5": "same_md5", "keyframe_phashes": ["aa"]})
+
+        with patch("video_processing.dedup.SQLAlchemyGeneratedVideoRepository") as MockRepo:
+            mock_repo = MockRepo.return_value
+            mock_repo.list_by_project.return_value = [existing]
+            result = deduplicator.check_duplicate(fingerprint, "proj1", session)
+
+        assert result is not None
+        assert result["duplicate"] is True
+        assert result["duplicate_of"] == "vid2"
+
+
+class TestCheckDuplicateScopeUser:
+    """test_check_duplicate_scope_user：跨项目查重."""
+
+    def test_user_scope_queries_by_user(self):
+        from video_processing.dedup import VideoDeduplicator
+
+        deduplicator = VideoDeduplicator()
+        fingerprint = _make_fingerprint(md5="unique_md5")
+        session = MagicMock()
+
+        with patch("video_processing.dedup.SQLAlchemyGeneratedVideoRepository") as MockRepo:
+            mock_repo = MockRepo.return_value
+            mock_repo.list_by_user.return_value = []
+            result = deduplicator.check_duplicate(
+                fingerprint,
+                "proj1",
+                session,
+                scope="user",
+                user_id="user_123",
+            )
+
+        mock_repo.list_by_user.assert_called_once()
+        assert result is None
+
+    def test_user_scope_finds_cross_project_duplicate(self):
+        from video_processing.dedup import VideoDeduplicator
+
+        deduplicator = VideoDeduplicator()
+        fingerprint = _make_fingerprint(md5="cross_proj_md5")
+        session = MagicMock()
+
+        # Existing video from a different project
+        existing = _make_video("vid_other", {"md5": "cross_proj_md5"}, project_id="proj_other")
+
+        with patch("video_processing.dedup.SQLAlchemyGeneratedVideoRepository") as MockRepo:
+            mock_repo = MockRepo.return_value
+            mock_repo.list_by_user.return_value = [existing]
+            result = deduplicator.check_duplicate(
+                fingerprint,
+                "proj1",
+                session,
+                scope="user",
+                user_id="user_123",
+            )
+
+        assert result is not None
+        assert result["duplicate"] is True
+        assert result["duplicate_of"] == "vid_other"
+
+
+class TestDurationPrefilter:
+    """test_duration_prefilter：时长 ±15% 过滤."""
+
+    def test_duration_prefilter_passes_correct_range(self):
+        from video_processing.dedup import VideoDeduplicator
+
+        deduplicator = VideoDeduplicator()
+        fingerprint = _make_fingerprint(duration_ms=30000)  # 30s video
+        session = MagicMock()
+
+        with patch("video_processing.dedup.SQLAlchemyGeneratedVideoRepository") as MockRepo:
+            mock_repo = MockRepo.return_value
+            mock_repo.list_by_user.return_value = []
+            deduplicator.check_duplicate(
+                fingerprint,
+                "proj1",
+                session,
+                scope="user",
+                user_id="user1",
+                duration_sec=30.0,
+            )
+
+        # Should pass duration_min=25.5, duration_max=34.5 (30 ± 15%)
+        call_args = mock_repo.list_by_user.call_args
+        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_no_duration_prefilter_when_zero(self):
+        from video_processing.dedup import VideoDeduplicator
+
+        deduplicator = VideoDeduplicator()
+        fingerprint = _make_fingerprint()
+        session = MagicMock()
+
+        with patch("video_processing.dedup.SQLAlchemyGeneratedVideoRepository") as MockRepo:
+            mock_repo = MockRepo.return_value
+            mock_repo.list_by_user.return_value = []
+            deduplicator.check_duplicate(
+                fingerprint,
+                "proj1",
+                session,
+                scope="user",
+                user_id="user1",
+                duration_sec=0,
+            )
+
+        call_args = mock_repo.list_by_user.call_args
+        assert call_args[1]["duration_min"] == 0
+        assert call_args[1]["duration_max"] == 0
+
+
+class TestComputeDuplicateRateFormula:
+    """test_compute_duplicate_rate_formula：验证 0.4 * frame_match_rate + 0.6 * temporal_coverage_rate."""
+
+    def test_formula_with_matching_frames(self):
+        from video_processing.dedup import VideoDeduplicator
+
+        deduplicator = VideoDeduplicator()
+        # 10 frames, all identical to existing → frame_match_rate = 1.0
+        phashes = ["aa00aa00aa00aa00"] * 10
+        fingerprint = _make_fingerprint(md5="new", phashes=phashes, duration_ms=20000)
+        session = MagicMock()
+
+        existing = _make_video(
+            "vid2",
+            {"md5": "other", "keyframe_phashes": ["aa00aa00aa00aa00"] * 5},
+        )
+
+        with patch("video_processing.dedup.SQLAlchemyGeneratedVideoRepository") as MockRepo:
+            mock_repo = MockRepo.return_value
+            mock_repo.list_by_project.return_value = [existing]
+            deduplicator._get_existing_chunks = MagicMock(return_value=[])
+            rate = deduplicator.compute_duplicate_rate(fingerprint, "proj1", "vid1", session)
+
+        # frame_match_rate=1.0, temporal_coverage depends on segments
+        # duplicate_rate = (1.0 * 0.4 + temporal_coverage * 0.6) * 100
+        assert rate["duplicate_rate"] >= 40.0  # At minimum, frame_match contributes 40%
+
+    def test_no_match_returns_zero(self):
+        from video_processing.dedup import VideoDeduplicator
+
+        deduplicator = VideoDeduplicator()
+        # Completely different phashes
+        fingerprint = _make_fingerprint(md5="new", phashes=["ff00ff00ff00ff00"])
+        session = MagicMock()
+
+        existing = _make_video(
+            "vid2",
+            {"md5": "other", "keyframe_phashes": ["00ff00ff00ff00ff"]},
+        )
+
+        with patch("video_processing.dedup.SQLAlchemyGeneratedVideoRepository") as MockRepo:
+            mock_repo = MockRepo.return_value
+            mock_repo.list_by_project.return_value = [existing]
+            deduplicator._get_existing_chunks = MagicMock(return_value=[])
+            rate = deduplicator.compute_duplicate_rate(fingerprint, "proj1", "vid1", session)
+
+        # Very different phashes, match_ratio < 0.3 → skipped
+        assert rate["duplicate_rate"] == 0.0
+
+
+class TestComputeDuplicateRateReturnDict:
+    """test_compute_duplicate_rate_return_dict：验证返回 dict 含三个字段."""
+
+    def test_return_structure(self):
+        from video_processing.dedup import VideoDeduplicator
+
+        deduplicator = VideoDeduplicator()
+        fingerprint = _make_fingerprint()
+        session = MagicMock()
+
+        with patch("video_processing.dedup.SQLAlchemyGeneratedVideoRepository") as MockRepo:
+            mock_repo = MockRepo.return_value
+            mock_repo.list_by_project.return_value = []
+            result = deduplicator.compute_duplicate_rate(fingerprint, "proj1", "vid1", session)
+
+        assert isinstance(result, dict)
+        assert set(result.keys()) == {"duplicate_rate", "visual_similarity", "match_count"}
+        assert isinstance(result["duplicate_rate"], float)
+        assert isinstance(result["visual_similarity"], float)
+        assert isinstance(result["match_count"], int)
+        assert 0 <= result["duplicate_rate"] <= 100
+        assert 0 <= result["visual_similarity"] <= 1
+
+
+class TestBackwardCompat:
+    """test_backward_compat：不传 scope 时行为不变."""
+
+    def test_default_scope_is_project(self):
+        from video_processing.dedup import VideoDeduplicator
+
+        deduplicator = VideoDeduplicator()
+        fingerprint = _make_fingerprint()
+        session = MagicMock()
+
+        with patch("video_processing.dedup.SQLAlchemyGeneratedVideoRepository") as MockRepo:
+            mock_repo = MockRepo.return_value
+            mock_repo.list_by_project.return_value = []
+
+            # Call without scope parameter
+            result = deduplicator.compute_duplicate_rate(fingerprint, "proj1", "vid1", session)
+
+        # Should use list_by_project (not list_by_user)
+        mock_repo.list_by_project.assert_called_once_with("proj1")
+        mock_repo.list_by_user.assert_not_called()
+        assert result["duplicate_rate"] == 0.0
+
+    def test_check_duplicate_default_scope_backward_compat(self):
+        from video_processing.dedup import VideoDeduplicator
+
+        deduplicator = VideoDeduplicator()
+        fingerprint = _make_fingerprint()
+        session = MagicMock()
+
+        with patch("video_processing.dedup.SQLAlchemyGeneratedVideoRepository") as MockRepo:
+            mock_repo = MockRepo.return_value
+            mock_repo.list_by_project.return_value = []
+            result = deduplicator.check_duplicate(fingerprint, "proj1", session)
+
+        mock_repo.list_by_project.assert_called_once_with("proj1")
+        assert result is None
+
+
+class TestListByUserRepository:
+    """直接测试 generated_video_repository.list_by_user() 的真实实现，覆盖 diff 代码行。"""
+
+    def _make_repo(self):
+        from sqlalchemy import create_engine
+        from sqlalchemy.orm import sessionmaker
+
+        from packages.adapters.sqlalchemy_impl.generated_video_repository import SQLAlchemyGeneratedVideoRepository
+        from packages.adapters.sqlalchemy_impl.models import Base, GeneratedVideoModel
+
+        engine = create_engine("sqlite:///:memory:")
+        Base.metadata.create_all(engine)
+        Session = sessionmaker(bind=engine)
+        session = Session()
+        repo = SQLAlchemyGeneratedVideoRepository(session)
+        return repo, session
+
+    def _insert_video(self, session, video_id, user_id, project_id, duration, **kw):
+        from packages.adapters.sqlalchemy_impl.models import GeneratedVideoModel
+
+        row = GeneratedVideoModel(
+            id=video_id,
+            user_id=user_id,
+            project_id=project_id,
+            generation_task_id=f"task-{video_id[:8]}",
+            name=f"video-{video_id[:8]}.mp4",
+            file_url=f"https://example.com/{video_id}.mp4",
+            file_size=1024,
+            duration=duration,
+            width=1280,
+            height=720,
+            fps=25.0,
+            status="completed",
+        )
+        session.add(row)
+        session.flush()
+        return row
+
+    def test_list_by_user_returns_cross_project_videos(self):
+        """list_by_user 返回该用户所有项目的视频。"""
+        repo, session = self._make_repo()
+        self._insert_video(session, "v1", "user-a", "proj-1", 30.0)
+        self._insert_video(session, "v2", "user-a", "proj-2", 45.0)
+        self._insert_video(session, "v3", "user-b", "proj-1", 20.0)
+
+        results = repo.list_by_user("user-a")
+        assert len(results) == 2
+        ids = {r.id for r in results}
+        assert ids == {"v1", "v2"}
+        session.close()
+
+    def test_list_by_user_with_duration_filter(self):
+        """list_by_user 支持 duration_min/duration_max 过滤。"""
+        repo, session = self._make_repo()
+        self._insert_video(session, "v1", "user-a", "proj-1", 10.0)
+        self._insert_video(session, "v2", "user-a", "proj-1", 30.0)
+        self._insert_video(session, "v3", "user-a", "proj-1", 60.0)
+
+        results = repo.list_by_user("user-a", duration_min=20.0, duration_max=50.0)
+        assert len(results) == 1
+        assert results[0].id == "v2"
+        session.close()
+
+    def test_list_by_user_empty_result(self):
+        """list_by_user 无匹配时返回空列表。"""
+        repo, session = self._make_repo()
+        self._insert_video(session, "v1", "user-a", "proj-1", 30.0)
+
+        results = repo.list_by_user("user-nonexistent")
+        assert results == []
+        session.close()

tests/unit/test_duplication_api_enqueue.py

@@ -0,0 +1,168 @@
+"""#1661 查重 API enqueue 及仓储 commit 覆盖测试。
+
+覆盖：
+- upload 接口在成功后调用 celery_app.send_task
+- retry 接口在成功后调用 celery_app.send_task
+- duplication_repository.update() 正确调用 session.commit()
+"""
+
+from __future__ import annotations
+
+import os
+import sys
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+os.environ.setdefault("JWT_SECRET_KEY", "unit-test-secret-key-for-testing")
+os.environ.setdefault("DATABASE_URL", "sqlite:///test.db")
+
+ROOT = os.path.join(os.path.dirname(__file__), "..", "..")
+sys.path.insert(0, os.path.join(ROOT, "apps", "api"))
+sys.path.insert(0, os.path.join(ROOT, "packages"))
+
+from app.api.routes.duplication import router
+from app.auth import AuthenticatedUser, get_current_user
+from app.core.storage import get_storage_service
+from app.dependencies import get_duplication_repository
+from fastapi import FastAPI
+from fastapi.testclient import TestClient
+
+from packages.domain.duplication import DuplicationRecord
+from packages.domain.entities import User
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+
+def _make_test_user():
+    return User(id="user-1", username="testuser", email="test@example.com", display_name="Test User")
+
+
+def _make_auth_user():
+    return AuthenticatedUser(user=_make_test_user(), session_id="test-session", token_type="bearer")
+
+
+def _make_record(status="pending"):
+    record = DuplicationRecord.create(
+        user_id="user-1",
+        filename="test.mp4",
+        file_size=1024,
+        storage_key="duplication/abc/test.mp4",
+    )
+    if status != "pending":
+        record.status = status
+    return record
+
+
+def _build_client(auth_user, repo, storage=None):
+    """构建带 dependency_overrides 的 TestClient。"""
+    app = FastAPI()
+    app.include_router(router, prefix="/duplication")
+    app.dependency_overrides[get_current_user] = lambda: auth_user
+    app.dependency_overrides[get_duplication_repository] = lambda: repo
+    if storage is not None:
+        app.dependency_overrides[get_storage_service] = lambda: storage
+    return TestClient(app)
+
+
+# ---------------------------------------------------------------------------
+# 1. Upload endpoint enqueues celery task
+# ---------------------------------------------------------------------------
+
+
+def test_upload_enqueue_calls_celery_task():
+    """POST /duplication/upload 成功创建记录后必须调用 send_task。"""
+    record = _make_record()
+
+    fake_repo = MagicMock()
+    fake_repo.create.return_value = record
+
+    fake_storage = MagicMock()
+    fake_auth = _make_auth_user()
+
+    client = _build_client(fake_auth, fake_repo, fake_storage)
+
+    with patch("app.api.routes.duplication.celery_app") as mock_celery:
+        response = client.post(
+            "/duplication/upload",
+            files={"file": ("test.mp4", b"fake-video-content", "video/mp4")},
+        )
+
+    assert response.status_code == 200, response.text
+    mock_celery.send_task.assert_called_once_with(
+        "worker.process_duplication_check",
+        args=[record.id],
+    )
+
+
+# ---------------------------------------------------------------------------
+# 2. Retry endpoint enqueues celery task
+# ---------------------------------------------------------------------------
+
+
+def test_retry_enqueue_calls_celery_task():
+    """POST /duplication/records/{id}/retry 成功后必须调用 send_task。"""
+    record = _make_record(status="failed")
+
+    fake_repo = MagicMock()
+    fake_repo.get.return_value = record
+
+    # RetryDuplicationUseCase.execute 内部调用 repo.get → record.reset_for_retry → repo.update
+    updated = _make_record()
+    updated.id = record.id
+    updated.status = "pending"
+    fake_repo.update.return_value = updated
+
+    fake_auth = _make_auth_user()
+
+    client = _build_client(fake_auth, fake_repo)
+
+    with patch("app.api.routes.duplication.celery_app") as mock_celery:
+        response = client.post(f"/duplication/records/{record.id}/retry")
+
+    assert response.status_code == 200, response.text
+    mock_celery.send_task.assert_called_once_with(
+        "worker.process_duplication_check",
+        args=[record.id],
+    )
+
+
+# ---------------------------------------------------------------------------
+# 3. Repository update calls session.commit()
+# ---------------------------------------------------------------------------
+
+
+def test_repository_update_calls_session_commit():
+    """duplication_repository 的 update 方法必须调用 session.commit()。"""
+    from packages.adapters.sqlalchemy_impl.duplication_repository import (
+        SQLAlchemyDuplicationRecordRepository,
+    )
+    from packages.adapters.sqlalchemy_impl.models import DuplicationRecordModel
+
+    mock_session = MagicMock()
+    mock_model = MagicMock(spec=DuplicationRecordModel)
+    mock_model.id = "rec-1"
+
+    mock_session.query.return_value.filter.return_value.first.return_value = mock_model
+
+    repo = SQLAlchemyDuplicationRecordRepository(mock_session)
+
+    record = DuplicationRecord.create(
+        user_id="user-1",
+        filename="test.mp4",
+        file_size=1024,
+        storage_key="duplication/abc/test.mp4",
+    )
+    record.status = "completed"
+    record.duplicate_rate = 42.0
+    record.duplicate_count = 1
+    record.visual_similarity = 0.85
+    record.match_count = 2
+
+    result = repo.update(record)
+
+    mock_session.commit.assert_called()
+    assert result.visual_similarity == 0.85
+    assert result.match_count == 2

tests/unit/test_duplication_check_worker.py

@@ -0,0 +1,378 @@
+"""#1661 手动查重 worker task 测试：成功/失败/重试/片段映射/schema 字段。"""
+
+import sys
+from pathlib import Path
+from unittest.mock import MagicMock, patch
+
+# cv2/numpy 在测试环境不可用，提前 mock
+sys.modules.setdefault("cv2", MagicMock())
+
+ROOT = Path(__file__).resolve().parents[2]
+sys.path.insert(0, str(ROOT / "apps" / "api"))
+sys.path.insert(0, str(ROOT / "packages"))
+sys.path.insert(0, str(ROOT / "apps" / "worker"))
+
+
+def _get_task(mod):
+    """返回 (run_callable, real_task)。
+
+    - celery task 环境：run 是 bound method（self 已绑定），retry 用 patch.object 打桩
+    - 原始函数环境：用一个 mock_self 作为 self
+    """
+    task_obj = mod.process_duplication_check
+    real = task_obj._get_current_object() if hasattr(task_obj, "_get_current_object") else task_obj
+    if hasattr(real, "run") and hasattr(real, "retry"):
+        return real.run, real, True  # bound
+    return real, None, False
+
+
+def _run(mod, record_id, retries=0):
+    """执行 task，返回 (result_or_None, raised_exc, mock_self_or_None)。"""
+    from celery.exceptions import Retry as CeleryRetry
+
+    func, real_task, bound = _get_task(mod)
+    raised = None
+    result = None
+    if bound:
+        mock_retry = MagicMock(side_effect=CeleryRetry("retry"))
+        with patch.object(real_task, "retry", mock_retry):
+            real_task.request.retries = retries
+            real_task.max_retries = 3
+            try:
+                result = func(record_id)
+            except CeleryRetry as e:
+                raised = e
+        return result, raised, None
+    mock_self = MagicMock()
+    mock_self.request.retries = retries
+    mock_self.max_retries = 3
+    mock_self.retry = MagicMock(side_effect=CeleryRetry("retry"))
+    try:
+        result = func(mock_self, record_id)
+    except CeleryRetry as e:
+        raised = e
+    return result, raised, mock_self
+
+
+def _make_record(status="pending"):
+    from packages.domain.duplication import DuplicationRecord
+
+    record = DuplicationRecord.create(
+        user_id="user-1",
+        filename="query.mp4",
+        file_size=1024,
+        storage_key="duplication/abc/query.mp4",
+    )
+    if status != "pending":
+        record.status = status
+    return record
+
+
+def _make_fingerprint():
+    from video_processing.dedup import FingerprintChunk, VideoFingerprint
+
+    chunks = [
+        FingerprintChunk(start_time_ms=0, end_time_ms=2000, phash_binary="0" * 16, color_histogram=[], frame_count=1),
+        FingerprintChunk(
+            start_time_ms=2000, end_time_ms=4000, phash_binary="1" * 16, color_histogram=[], frame_count=1
+        ),
+    ]
+    return VideoFingerprint(
+        md5="qmd5",
+        keyframe_phashes=[c.phash_binary for c in chunks],
+        color_histograms=[],
+        duration=10000.0,
+        resolution=(720, 1280),
+        chunks=chunks,
+    )
+
+
+def _patch_common(record, storage=None, dedup=None, session=None):
+    from worker_app.tasks import duplication_check as mod
+
+    fake_repo = MagicMock()
+    fake_repo.get.return_value = record
+    return [
+        patch.object(mod, "SessionLocal", return_value=session or MagicMock()),
+        patch.object(mod, "SQLAlchemyDuplicationRecordRepository", return_value=fake_repo),
+        patch.object(mod, "get_storage_service", return_value=storage or MagicMock()),
+        patch.object(mod, "VideoDeduplicator", return_value=dedup or MagicMock()),
+    ], fake_repo
+
+
+class TestProcessDuplicationCheckSuccess:
+    def test_success_flow_updates_record(self):
+        from worker_app.tasks import duplication_check as mod
+
+        record = _make_record()
+        fake_session = MagicMock()
+        fake_storage = MagicMock()
+        fake_dedup = MagicMock()
+        fake_dedup.compute_fingerprint.return_value = _make_fingerprint()
+        fake_dedup.compute_duplicate_rate.return_value = {
+            "duplicate_rate": 42.5,
+            "visual_similarity": 0.83,
+            "match_count": 1,
+        }
+        patches, fake_repo = _patch_common(record, storage=fake_storage, dedup=fake_dedup, session=fake_session)
+        patches.append(patch.object(mod, "_build_domain_segments", return_value=(["SEG"], 1)))
+        for p in patches:
+            p.start()
+        try:
+            result, raised, _ = _run(mod, record.id)
+        finally:
+            for p in patches:
+                p.stop()
+
+        assert raised is None
+        assert result["ok"] is True
+        assert result["status"] == "completed"
+        assert result["duplicate_rate"] == 42.5
+        assert result["visual_similarity"] == 0.83
+        assert result["match_count"] == 1
+        assert result["segments"] == 1
+
+        assert record.status == "completed"
+        assert record.duplicate_rate == 42.5
+        assert record.visual_similarity == 0.83
+        assert record.match_count == 1
+        assert record.duplicate_count == 1
+        assert record.segments == ["SEG"]
+
+        fake_storage.download_file.assert_called_once()
+        fake_dedup.compute_fingerprint.assert_called_once()
+        _, kwargs = fake_dedup.compute_duplicate_rate.call_args
+        assert kwargs["scope"] == "user"
+        assert kwargs["user_id"] == "user-1"
+        assert kwargs["current_video_id"] is None
+        assert fake_repo.update.call_count >= 2
+        fake_session.commit.assert_called()
+        fake_session.close.assert_called()
+
+    def test_already_completed_is_skipped(self):
+        from worker_app.tasks import duplication_check as mod
+
+        record = _make_record(status="completed")
+        patches, fake_repo = _patch_common(record)
+        for p in patches:
+            p.start()
+        try:
+            result, raised, _ = _run(mod, record.id)
+        finally:
+            for p in patches:
+                p.stop()
+        assert raised is None
+        assert result.get("skipped") is True
+        fake_repo.update.assert_not_called()
+
+
+class TestProcessDuplicationCheckFailure:
+    def test_record_not_found_raises(self):
+        from worker_app.tasks import duplication_check as mod
+
+        fake_repo = MagicMock()
+        fake_repo.get.return_value = None
+        patches = [
+            patch.object(mod, "SessionLocal", return_value=MagicMock()),
+            patch.object(mod, "SQLAlchemyDuplicationRecordRepository", return_value=fake_repo),
+            patch.object(mod, "get_storage_service", return_value=MagicMock()),
+        ]
+        for p in patches:
+            p.start()
+        try:
+            _result, raised, _ = _run(mod, "nope", retries=0)
+        finally:
+            for p in patches:
+                p.stop()
+        # 找不到记录触发异常 → retry（第一次）
+        assert raised is not None
+
+    def test_download_failure_retries_then_marks_failed(self):
+        from worker_app.tasks import duplication_check as mod
+
+        # 第一次失败（retries=0）：保持 pending
+        record = _make_record()
+        fake_storage = MagicMock()
+        fake_storage.download_file.side_effect = RuntimeError("oss network down")
+        patches, _ = _patch_common(record, storage=fake_storage)
+        for p in patches:
+            p.start()
+        try:
+            _, raised, _ = _run(mod, record.id, retries=0)
+        finally:
+            for p in patches:
+                p.stop()
+        assert raised is not None
+        assert record.status == "processing", "首次失败不应标记 failed（已进入 processing 等待重试）"
+
+        # 最后一次（retries==max_retries=3）：标记 failed
+        record2 = _make_record()
+        patches2, fake_repo2 = _patch_common(record2, storage=fake_storage)
+        for p in patches2:
+            p.start()
+        try:
+            _run(mod, record2.id, retries=3)
+        finally:
+            for p in patches2:
+                p.stop()
+        assert record2.status == "failed"
+        assert "查重失败" in record2.error_message
+        fake_repo2.update.assert_called()
+
+    def test_temp_dir_cleaned_after_failure(self):
+        import os
+        import tempfile
+
+        from worker_app.tasks import duplication_check as mod
+
+        record = _make_record()
+        fake_storage = MagicMock()
+        fake_storage.download_file.side_effect = RuntimeError("boom")
+
+        created_dirs = []
+        real_mkdtemp = tempfile.mkdtemp
+
+        def fake_mkdtemp(prefix=None):
+            d = real_mkdtemp(prefix=prefix)
+            created_dirs.append(d)
+            return d
+
+        patches, _ = _patch_common(record, storage=fake_storage)
+        patches.append(patch.object(mod.tempfile, "mkdtemp", fake_mkdtemp))
+        for p in patches:
+            p.start()
+        try:
+            _run(mod, record.id, retries=0)
+        finally:
+            for p in patches:
+                p.stop()
+
+        assert created_dirs, "mkdtemp should have been called"
+        assert not os.path.isdir(created_dirs[0]), "temp dir should be removed in finally"
+
+
+class TestBuildDomainSegments:
+    def test_maps_worker_segments_to_domain_with_seconds_and_percent(self):
+        from video_processing.dedup import DuplicateSegment as WorkerSegment
+        from worker_app.tasks import duplication_check as mod
+
+        fingerprint = _make_fingerprint()
+
+        from packages.domain import GeneratedVideo
+
+        existing = GeneratedVideo(
+            id="vid-1",
+            project_id="proj-1",
+            generation_task_id="t1",
+            name="成片A",
+            file_url="oss://x",
+            file_size=1,
+            duration=10.0,
+            width=720,
+            height=1280,
+            fps=30.0,
+            video_fingerprint={"md5": "x"},
+        )
+        fake_video_repo = MagicMock()
+        fake_video_repo.list_by_user.return_value = [existing]
+
+        fake_dedup = MagicMock()
+        fake_dedup._get_existing_chunks.return_value = [
+            {"phash_binary": "0" * 16, "start_time_ms": 0, "end_time_ms": 2000, "color_histogram": []},
+        ]
+        worker_seg = WorkerSegment(
+            query_start_ms=1000,
+            query_end_ms=3000,
+            target_start_ms=5000,
+            target_end_ms=7000,
+            avg_distance=6.0,
+        )
+
+        with (
+            patch.object(mod, "SQLAlchemyGeneratedVideoRepository", return_value=fake_video_repo),
+            patch.object(mod, "find_duplicate_segments", return_value=[worker_seg]),
+        ):
+            segments, dup_count = mod._build_domain_segments(fingerprint, MagicMock(), fake_dedup, "user-1")
+
+        assert dup_count == 1
+        assert len(segments) == 1
+        seg = segments[0]
+        assert seg.source_start == 1.0
+        assert seg.source_end == 3.0
+        assert seg.matched_start == 5.0
+        assert seg.matched_end == 7.0
+        assert seg.matched_video_id == "vid-1"
+        assert seg.matched_video_name == "成片A"
+        assert abs(seg.similarity - 90.6) < 0.2
+
+    def test_skips_videos_without_chunks(self):
+        from worker_app.tasks import duplication_check as mod
+
+        fingerprint = _make_fingerprint()
+        from packages.domain import GeneratedVideo
+
+        existing = GeneratedVideo(
+            id="vid-2",
+            project_id="p",
+            generation_task_id="t",
+            name="老视频",
+            file_url="oss://x",
+            file_size=1,
+            duration=5.0,
+            width=720,
+            height=1280,
+            fps=30.0,
+            video_fingerprint={"md5": "old"},
+        )
+        fake_video_repo = MagicMock()
+        fake_video_repo.list_by_user.return_value = [existing]
+        fake_dedup = MagicMock()
+        fake_dedup._get_existing_chunks.return_value = []
+
+        with patch.object(mod, "SQLAlchemyGeneratedVideoRepository", return_value=fake_video_repo):
+            segments, dup_count = mod._build_domain_segments(fingerprint, MagicMock(), fake_dedup, "u")
+        assert segments == []
+        assert dup_count == 0
+
+
+class TestDuplicationSchemaAndDomainNewFields:
+    def test_record_response_includes_new_fields(self):
+        from app.schemas.duplication import DuplicationRecordResponse
+
+        resp = DuplicationRecordResponse(
+            id="r1",
+            filename="f.mp4",
+            file_size=1,
+            status="completed",
+            duplicate_rate=10.0,
+            duplicate_count=1,
+            visual_similarity=0.5,
+            match_count=2,
+            created_at="2026-09-04T00:00:00",
+            updated_at="2026-09-04T00:00:00",
+        )
+        assert resp.visual_similarity == 0.5
+        assert resp.match_count == 2
+
+    def test_record_response_new_fields_default_none(self):
+        from app.schemas.duplication import DuplicationRecordResponse
+
+        resp = DuplicationRecordResponse(id="r1", filename="f.mp4", file_size=1, created_at="x", updated_at="y")
+        assert resp.visual_similarity is None
+        assert resp.match_count is None
+
+    def test_domain_mark_completed_accepts_new_fields(self):
+        record = _make_record()
+        record.mark_completed(33.0, 2, [], visual_similarity=0.77, match_count=3)
+        assert record.status == "completed"
+        assert record.visual_similarity == 0.77
+        assert record.match_count == 3
+
+    def test_reset_for_retry_clears_new_fields(self):
+        record = _make_record()
+        record.mark_completed(10.0, 1, [], visual_similarity=0.5, match_count=1)
+        record.status = "failed"
+        record.reset_for_retry()
+        assert record.status == "pending"
+        assert record.visual_similarity is None
+        assert record.match_count is None

tests/unit/test_fingerprint_chunks.py

@@ -110,7 +110,6 @@ from video_processing.dedup import (  # noqa: E402
     FingerprintChunk,
     VideoFingerprint,
     _save_fingerprint_chunks,
-    compute_chunk_interval,
 )
 
 # ── Restore sys.modules immediately after import ──
@@ -125,36 +124,6 @@ for _key, _value in _SAVED_MODULES_VALUES.items():
 del _SAVED_MODULES_KEYS, _SAVED_MODULES_VALUES, _key, _value
 
 
-class TestChunkInterval:
-    """测试分片间隔策略。"""
-
-    def test_short_video_interval(self):
-        """短视频（≤60秒）每 2 秒一个分片。"""
-        assert compute_chunk_interval(0) == 2
-        assert compute_chunk_interval(30) == 2
-        assert compute_chunk_interval(60) == 2
-
-    def test_long_video_interval(self):
-        """长视频（>60秒）每 5 秒一个分片。"""
-        assert compute_chunk_interval(61) == 5
-        assert compute_chunk_interval(120) == 5
-        assert compute_chunk_interval(300) == 5
-
-    def test_chunk_count_60s_video(self):
-        """60秒视频 → 30 片（60/2=30）。"""
-        duration = 60
-        interval = compute_chunk_interval(duration)
-        expected_chunks = int(duration / interval)
-        assert expected_chunks == 30
-
-    def test_chunk_count_120s_video(self):
-        """120秒视频 → 24 片（120/5=24）。"""
-        duration = 120
-        interval = compute_chunk_interval(duration)
-        expected_chunks = int(duration / interval)
-        assert expected_chunks == 24
-
-
 class TestVideoFingerprintToChunkModels:
     """测试 VideoFingerprint.to_chunk_models() 输出。"""
 

tests/unit/test_generated_video_creation_logic.py

@@ -359,6 +359,11 @@ class TestThumbnailInDedupHelpers:
                 mock_dedup.compute_fingerprint.return_value = MagicMock(to_dict=lambda: {})
                 mock_dedup.check_duplicate.return_value = None
                 mock_dedup.check_batch_duplicate.return_value = None
+                mock_dedup.compute_duplicate_rate.return_value = {
+                    "duplicate_rate": 0.0,
+                    "visual_similarity": 0.0,
+                    "match_count": 0,
+                }
 
                 result = create_video_record_and_dedup(
                     generation_task_id="task-thumb-reuse",
@@ -401,6 +406,11 @@ class TestThumbnailInDedupHelpers:
                 mock_dedup.compute_fingerprint.return_value = MagicMock(to_dict=lambda: {})
                 mock_dedup.check_duplicate.return_value = None
                 mock_dedup.check_batch_duplicate.return_value = None
+                mock_dedup.compute_duplicate_rate.return_value = {
+                    "duplicate_rate": 0.0,
+                    "visual_similarity": 0.0,
+                    "match_count": 0,
+                }
 
                 result = create_video_record_and_dedup(
                     generation_task_id="task-thumb-gen",
@@ -443,6 +453,11 @@ class TestThumbnailInDedupHelpers:
                 mock_dedup.compute_fingerprint.return_value = MagicMock(to_dict=lambda: {})
                 mock_dedup.check_duplicate.return_value = None
                 mock_dedup.check_batch_duplicate.return_value = None
+                mock_dedup.compute_duplicate_rate.return_value = {
+                    "duplicate_rate": 0.0,
+                    "visual_similarity": 0.0,
+                    "match_count": 0,
+                }
 
                 result = create_video_record_and_dedup(
                     generation_task_id="task-thumb-fail",

tests/unit/test_phash_threshold_calibration_1658.py

@@ -0,0 +1,273 @@
+"""Issue #1658: pHash 阈值校准 + 颜色直方图融合 — 单元测试.
+
+在 #1659（动态抽帧+滑动窗口）与 #1660（查重率）已合入 develop 的基础上，
+本测试覆盖 #1658 的最小增量改动：
+
+1. PHASH_THRESHOLD 由 10 收紧到 8（核心校准）
+2. 融合权重常量 MATCH_RATIO_THRESHOLD / PHASH_WEIGHT / HISTOGRAM_WEIGHT 实际生效
+   （不再是硬编码魔法数字）
+3. VideoDeduplicator._compute_fusion_score 统一融合得分方法：
+   - 无直方图数据时回退中性值 0.5
+   - DB NULL（None）显式回退空列表，不崩溃
+   - 全零直方图（全黑视频）为有效数据，参与 Bhattacharyya 计算
+   - 返回 0~1 原始得分，判重由调用方与 DUPLICATE_THRESHOLD 比较
+4. Bhattacharyya 系数对上游异常负值有 sqrt domain 防御
+"""
+
+from __future__ import annotations
+
+import sys
+from unittest.mock import MagicMock
+
+import pytest
+
+
+def _mock_module(**attrs):
+    """Create a mock module with __spec__ to avoid AttributeError."""
+    m = MagicMock()
+    m.__spec__ = None
+    for k, v in attrs.items():
+        setattr(m, k, v)
+    return m
+
+
+# ── Module-level setup: mock deps, import dedup, then restore sys.modules ──
+_SAVED_MODULES_KEYS = set(sys.modules.keys())
+_SAVED_MODULES_VALUES = {
+    k: sys.modules.get(k)
+    for k in [
+        "cv2",
+        "celery",
+        "sqlalchemy",
+        "sqlalchemy.orm",
+        "sqlalchemy.engine",
+        "sqlalchemy.ext",
+        "sqlalchemy.ext.declarative",
+        "worker_app.db",
+        "worker_app.celery_app",
+        "worker_app.core.config",
+        "packages.adapters.sqlalchemy_impl.session",
+        "packages.adapters.sqlalchemy_impl.generated_video_repository",
+        "packages.adapters.sqlalchemy_impl.models",
+        "packages.shared.config",
+        "packages.shared.storage",
+    ]
+}
+
+sys.modules["cv2"] = _mock_module()
+
+_mock_celery = MagicMock()
+_mock_celery.Task = MagicMock
+_mock_celery.Celery = MagicMock
+_mock_celery.__spec__ = None
+sys.modules["celery"] = _mock_celery
+
+_mock_sqla = MagicMock()
+_mock_sqla.__path__ = []
+_mock_sqla.__spec__ = None
+sys.modules["sqlalchemy"] = _mock_sqla
+
+_mock_sqla_orm = MagicMock()
+_mock_sqla_orm.__path__ = []
+_mock_sqla_orm.__spec__ = None
+_mock_sqla_orm.Session = MagicMock
+sys.modules["sqlalchemy.orm"] = _mock_sqla_orm
+sys.modules["sqlalchemy.engine"] = _mock_module()
+sys.modules["sqlalchemy.ext"] = _mock_module()
+sys.modules["sqlalchemy.ext.declarative"] = _mock_module()
+
+sys.modules["worker_app.db"] = _mock_module(SessionLocal=MagicMock())
+sys.modules["worker_app.celery_app"] = _mock_module(celery_app=MagicMock())
+sys.modules["worker_app.core.config"] = _mock_module(get_settings=MagicMock(return_value=MagicMock()))
+
+sys.modules["packages.adapters.sqlalchemy_impl.session"] = _mock_module(
+    Base=MagicMock(),
+    build_engine=MagicMock(),
+    build_session_factory=MagicMock(),
+    ensure_database_exists=MagicMock(),
+    initialize_database=MagicMock(),
+)
+sys.modules["packages.adapters.sqlalchemy_impl.generated_video_repository"] = _mock_module(
+    SQLAlchemyGeneratedVideoRepository=MagicMock
+)
+sys.modules["packages.adapters.sqlalchemy_impl.models"] = _mock_module(
+    VideoFingerprintChunkModel=MagicMock,
+    GeneratedVideoModel=MagicMock,
+)
+sys.modules["packages.shared.config"] = _mock_module(get_shared_settings=MagicMock(return_value=MagicMock()))
+sys.modules["packages.shared.storage"] = _mock_module()
+
+import video_processing.dedup as _dedup_mod  # noqa: E402
+from video_processing.dedup import (  # noqa: E402
+    DUPLICATE_THRESHOLD,
+    HISTOGRAM_WEIGHT,
+    MATCH_RATIO_THRESHOLD,
+    PHASH_WEIGHT,
+    VideoDeduplicator,
+)
+
+# ── Restore sys.modules immediately after import ──
+for _key in list(sys.modules.keys()):
+    if _key not in _SAVED_MODULES_KEYS:
+        del sys.modules[_key]
+for _key, _value in _SAVED_MODULES_VALUES.items():
+    if _value is not None:
+        sys.modules[_key] = _value
+    elif _key in sys.modules:
+        del sys.modules[_key]
+del _SAVED_MODULES_KEYS, _SAVED_MODULES_VALUES, _key, _value
+
+
+# ── 测试夹具 ─────────────────────────────────────────────────────
+
+_UNIFORM_HIST = [1.0 / 96] * 96  # 归一化均匀直方图，sum=1.0，自相似度≈1.0
+_ZERO_HIST = [0.0] * 96  # 全黑视频的全零直方图（有效数据）
+
+
+# ── TestThresholdCalibration：#1658 核心校准 ────────────────────
+
+
+class TestThresholdCalibration:
+    """pHash 阈值由 10 收紧到 8（Issue #1658）。"""
+
+    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
+
+    def test_duplicate_threshold_constant(self):
+        assert DUPLICATE_THRESHOLD == 0.70
+
+    def test_fusion_weights(self):
+        assert PHASH_WEIGHT == 0.7
+        assert HISTOGRAM_WEIGHT == 0.3
+
+    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 [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 = [7, 7, 7, 9, 9]
+
+        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 ────────────────────
+
+
+class TestComputeFusionScore:
+    """_compute_fusion_score(median_distance, histograms_a, histograms_b)。"""
+
+    def test_no_histogram_falls_back_to_neutral_05(self):
+        """双方均无直方图 → hist_similarity 回退 0.5。
+
+        d=0: 0.7*1.0 + 0.3*0.5 = 0.85
+        """
+        score = VideoDeduplicator._compute_fusion_score(0, [], [])
+        assert score == pytest.approx(0.85, abs=1e-6)
+
+    def test_none_histograms_treated_as_empty(self):
+        """DB NULL（None）必须显式回退空列表，不得 len(None) 崩溃。"""
+        score_none = VideoDeduplicator._compute_fusion_score(0, [], None)
+        score_empty = VideoDeduplicator._compute_fusion_score(0, [], [])
+        assert score_none == pytest.approx(score_empty, abs=1e-9)
+        assert score_none == pytest.approx(0.85, abs=1e-6)
+
+    def test_none_histograms_on_query_side_no_crash(self):
+        """查询侧直方图为 None 时同样不崩溃。"""
+        score = VideoDeduplicator._compute_fusion_score(0, None, [_UNIFORM_HIST])
+        # 查询侧无直方图 → 平均相似度为 0（无 ha 可匹配）→ 0.7*1.0 + 0.3*0 = 0.7
+        assert score == pytest.approx(0.7, abs=1e-6)
+
+    def test_identical_uniform_histograms_score_near_1(self):
+        """完全相同的归一化直方图：Bhattacharyya≈1.0 → 融合分≈1.0。"""
+        score = VideoDeduplicator._compute_fusion_score(0, [_UNIFORM_HIST], [_UNIFORM_HIST])
+        assert score == pytest.approx(1.0, abs=1e-6)
+
+    def test_all_zero_histogram_is_valid_data(self):
+        """全零直方图（全黑视频）是有效数据，Bhattacharyya=0，不得走 0.5 回退。
+
+        若错误地用 `if histograms_b` 之外的 `or []` 把全零列表清空，
+        会错误回退到 0.5，把全黑视频的相似度抬高 0.15。
+        d=0 时：正确行为 hist_sim=0 → 0.7*1.0 + 0.3*0 = 0.7；
+        若全零直方图被错误清空回退 0.5 → 0.85。
+        """
+        score = VideoDeduplicator._compute_fusion_score(0, [_ZERO_HIST], [_ZERO_HIST])
+        assert score == pytest.approx(0.7, abs=1e-6)
+        # 与错误回退值 0.85 明确区分开
+        assert abs(score - 0.85) > 0.1
+        # 注：d=0 时 phash 满分 0.7 恰达 DUPLICATE_THRESHOLD，全黑+完全相同 phash 仍判重，符合预期
+        assert score >= DUPLICATE_THRESHOLD - 1e-9
+
+    def test_score_range_within_0_1(self):
+        for d in (0, 8, 16, 32, 64):
+            score = VideoDeduplicator._compute_fusion_score(d, [_UNIFORM_HIST], [_UNIFORM_HIST])
+            assert 0.0 <= score <= 1.0
+
+    def test_formula_matches_weights(self):
+        """得分 = PHASH_WEIGHT * (1 - d/64) + HISTOGRAM_WEIGHT * hist_sim。"""
+        d = 6  # phash_sim = 1 - 6/64 = 0.90625
+        score = VideoDeduplicator._compute_fusion_score(d, [], [])  # hist 回退 0.5
+        expected = PHASH_WEIGHT * (1 - d / 64) + HISTOGRAM_WEIGHT * 0.5
+        assert score == pytest.approx(expected, abs=1e-9)
+        # 0.7*0.90625 + 0.15 = 0.634375 + 0.15 = 0.784375
+        assert score == pytest.approx(0.784375, abs=1e-6)
+
+
+# ── TestBhattacharyyaDefense：负值/异常输入防御 ─────────────────
+
+
+class TestBhattacharyyaDefense:
+    """Bhattacharyya 系数对异常输入的防御。"""
+
+    def test_negative_values_do_not_raise(self):
+        """上游异常负值不得触发 sqrt domain error（max(0.0, ai*bi) 保护）。"""
+        bad_hist = [-0.01] * 96  # 异常负值
+        coeff = VideoDeduplicator._bhattacharyya_coefficient(bad_hist, _UNIFORM_HIST)
+        # 负值乘积被钳为 0，系数为 0 而不是抛 ValueError
+        assert coeff == pytest.approx(0.0, abs=1e-9)
+
+    def test_normal_histograms_coefficient_near_1(self):
+        coeff = VideoDeduplicator._bhattacharyya_coefficient(_UNIFORM_HIST, _UNIFORM_HIST)
+        assert coeff == pytest.approx(1.0, abs=1e-6)
+
+    def test_disjoint_histograms_coefficient_0(self):
+        """完全不重叠的直方图（前半 vs 后半非零）系数为 0。"""
+        hist_a = [0.0] * 96
+        hist_b = [0.0] * 96
+        for i in range(48):
+            hist_a[i] = 1.0 / 48
+        for i in range(48, 96):
+            hist_b[i] = 1.0 / 48
+        coeff = VideoDeduplicator._bhattacharyya_coefficient(hist_a, hist_b)
+        assert coeff == pytest.approx(0.0, abs=1e-9)
+
+
+# ── TestHistogramSimilarityEdgeCases ────────────────────────────
+
+
+class TestHistogramSimilarityEdgeCases:
+    """_compute_histogram_similarity 的边界行为。"""
+
+    def test_empty_either_side_returns_0(self):
+        assert VideoDeduplicator._compute_histogram_similarity([], [_UNIFORM_HIST]) == 0.0
+        assert VideoDeduplicator._compute_histogram_similarity([_UNIFORM_HIST], []) == 0.0
+
+    def test_best_match_per_histogram(self):
+        """每个查询直方图取与目标集合的最佳匹配，再取平均。"""
+        h1 = _UNIFORM_HIST
+        h2 = [0.0] * 96
+        h2[0] = 1.0  # 与均匀直方图完全不重叠
+        # 查询侧两张直方图：h1 最佳匹配≈1.0，h2 最佳匹配≈sqrt(1/96)≈0.102
+        sim = VideoDeduplicator._compute_histogram_similarity([h1, h2], [h1])
+        assert sim == pytest.approx((1.0 + (1.0 / 96) ** 0.5) / 2, abs=1e-3)

tests/unit/test_random_edge_crop.py

@@ -0,0 +1,207 @@
+"""#1664 随机边缘裁剪降重功能测试"""
+
+from __future__ import annotations
+
+import os
+import subprocess
+import sys
+from pathlib import Path
+from unittest.mock import MagicMock, call, patch
+
+import pytest
+
+ROOT = Path(__file__).resolve().parents[2]
+sys.path.insert(0, str(ROOT / "apps" / "worker"))
+sys.path.insert(0, str(ROOT / "apps" / "api"))
+sys.path.insert(0, str(ROOT / "packages"))
+
+
+from video_processing.ffmpeg_utils import random_edge_crop
+
+
+class TestRandomEdgeCropBasic:
+    """基本功能测试"""
+
+    def test_returns_input_path_when_output_none(self, tmp_path):
+        """output_path=None 时覆盖原文件并返回 input_path"""
+        input_file = tmp_path / "input.mp4"
+        input_file.write_bytes(b"fake video data")
+
+        fake_info = {"width": 1920, "height": 1080, "duration": 10, "fps": 30}
+        with (
+            patch("video_processing.ffmpeg_utils.probe_video_info", return_value=fake_info),
+            patch("video_processing.ffmpeg_utils.run_ffmpeg"),
+        ):
+            result = random_edge_crop(input_file)
+
+        assert result == input_file
+
+    def test_returns_output_path_when_specified(self, tmp_path):
+        """指定 output_path 时返回该路径"""
+        input_file = tmp_path / "input.mp4"
+        input_file.write_bytes(b"fake video data")
+        output_file = tmp_path / "output.mp4"
+
+        fake_info = {"width": 1920, "height": 1080, "duration": 10, "fps": 30}
+        with (
+            patch("video_processing.ffmpeg_utils.probe_video_info", return_value=fake_info),
+            patch("video_processing.ffmpeg_utils.run_ffmpeg"),
+        ):
+            result = random_edge_crop(input_file, output_file)
+
+        assert result == output_file
+
+    def test_skip_when_invalid_resolution(self, tmp_path):
+        """无法获取有效分辨率时跳过裁剪"""
+        input_file = tmp_path / "input.mp4"
+        input_file.write_bytes(b"fake video data")
+
+        fake_info = {"width": 0, "height": 0, "duration": 10, "fps": 30}
+        with (
+            patch("video_processing.ffmpeg_utils.probe_video_info", return_value=fake_info),
+            patch("video_processing.ffmpeg_utils.run_ffmpeg") as mock_ffmpeg,
+        ):
+            result = random_edge_crop(input_file)
+
+        assert result == input_file
+        mock_ffmpeg.assert_not_called()
+
+
+class TestRandomEdgeCropFFmpeg:
+    """FFmpeg 调用参数验证"""
+
+    def test_ffmpeg_crop_and_scale_filter(self, tmp_path):
+        """生成的 ffmpeg 滤镜包含 crop + scale"""
+        input_file = tmp_path / "input.mp4"
+        input_file.write_bytes(b"fake video data")
+
+        # 固定随机值以便验证
+        fake_info = {"width": 1000, "height": 1000, "duration": 10, "fps": 30}
+
+        with (
+            patch("video_processing.ffmpeg_utils.probe_video_info", return_value=fake_info),
+            patch("video_processing.ffmpeg_utils.run_ffmpeg") as mock_ffmpeg,
+            patch("random.uniform", side_effect=[0.03, 0.03, 0.03, 0.03]),
+        ):
+            random_edge_crop(input_file)
+
+        mock_ffmpeg.assert_called_once()
+        cmd = mock_ffmpeg.call_args[0][0]
+        # 找到 -vf 参数
+        vf_idx = cmd.index("-vf")
+        vf_value = cmd[vf_idx + 1]
+        assert "crop=" in vf_value
+        assert "scale=1000:1000" in vf_value
+
+    def test_crop_amounts_within_range(self, tmp_path):
+        """裁剪量在 2%~5% 范围内"""
+        input_file = tmp_path / "input.mp4"
+        input_file.write_bytes(b"fake video data")
+
+        fake_info = {"width": 1000, "height": 1000, "duration": 10, "fps": 30}
+
+        with (
+            patch("video_processing.ffmpeg_utils.probe_video_info", return_value=fake_info),
+            patch("video_processing.ffmpeg_utils.run_ffmpeg") as mock_ffmpeg,
+            patch("random.uniform", side_effect=[0.02, 0.05, 0.02, 0.05]),
+        ):
+            random_edge_crop(input_file)
+
+        cmd = mock_ffmpeg.call_args[0][0]
+        vf_idx = cmd.index("-vf")
+        vf_value = cmd[vf_idx + 1]
+        # crop_top=20, crop_bottom=50, crop_left=20, crop_right=50
+        # new_w = 1000-20-50 = 930, new_h = 1000-20-50 = 930
+        # x_offset = 20, y_offset = 20
+        assert "crop=930:930:20:20" in vf_value
+
+    def test_uses_libx264_codec(self, tmp_path):
+        """使用 libx264 编码"""
+        input_file = tmp_path / "input.mp4"
+        input_file.write_bytes(b"fake video data")
+
+        fake_info = {"width": 1920, "height": 1080, "duration": 10, "fps": 30}
+        with (
+            patch("video_processing.ffmpeg_utils.probe_video_info", return_value=fake_info),
+            patch("video_processing.ffmpeg_utils.run_ffmpeg") as mock_ffmpeg,
+        ):
+            random_edge_crop(input_file)
+
+        cmd = mock_ffmpeg.call_args[0][0]
+        assert "-c:v" in cmd
+        assert cmd[cmd.index("-c:v") + 1] == "libx264"
+
+
+class TestRandomEdgeCropErrorHandling:
+    """错误处理测试"""
+
+    def test_ffmpeg_failure_raises_exception(self, tmp_path):
+        """ffmpeg 失败时抛出异常"""
+        input_file = tmp_path / "input.mp4"
+        input_file.write_bytes(b"fake video data")
+
+        fake_info = {"width": 1920, "height": 1080, "duration": 10, "fps": 30}
+        with (
+            patch("video_processing.ffmpeg_utils.probe_video_info", return_value=fake_info),
+            patch(
+                "video_processing.ffmpeg_utils.run_ffmpeg",
+                side_effect=subprocess.CalledProcessError(1, "ffmpeg"),
+            ),
+        ):
+            with pytest.raises(subprocess.CalledProcessError):
+                random_edge_crop(input_file)
+
+    def test_probe_failure_propagates(self, tmp_path):
+        """probe_video_info 失败时异常传播"""
+        input_file = tmp_path / "input.mp4"
+        input_file.write_bytes(b"fake video data")
+
+        with patch(
+            "video_processing.ffmpeg_utils.probe_video_info",
+            side_effect=RuntimeError("probe failed"),
+        ):
+            with pytest.raises(RuntimeError, match="probe failed"):
+                random_edge_crop(input_file)
+
+
+class TestRandomEdgeCropEvenDimensions:
+    """偶数尺寸处理测试"""
+
+    def test_odd_crop_dimensions_adjusted_to_even(self, tmp_path):
+        """裁剪后尺寸为奇数时自动调整为偶数"""
+        input_file = tmp_path / "input.mp4"
+        input_file.write_bytes(b"fake video data")
+
+        # 1000 - 3 (top) - 4 (bottom) = 993 → 调整为 992
+        # 1000 - 3 (left) - 4 (right) = 993 → 调整为 992
+        fake_info = {"width": 1000, "height": 1000, "duration": 10, "fps": 30}
+
+        with (
+            patch("video_processing.ffmpeg_utils.probe_video_info", return_value=fake_info),
+            patch("video_processing.ffmpeg_utils.run_ffmpeg") as mock_ffmpeg,
+            # side_effect 控制 uniform 返回值
+            # top: 0.003*1000=3, bottom: 0.004*1000=4, left: 0.003*1000=3, right: 0.004*1000=4
+        ):
+            # 使用自定义 uniform 返回特定值
+            def fake_uniform(low, high):
+                # 返回特定百分比使得裁剪后尺寸为奇数
+                # 我们需要 crop_top=3, crop_bottom=4, crop_left=3, crop_right=4
+                return 0.0035  # 近似值
+
+            # 更简单的方式：直接 mock int(H * random.uniform(...)) 的结果
+            # 但我们直接测试最终 crop 滤镜即可
+            with patch("random.uniform", side_effect=[0.021, 0.022, 0.021, 0.022]):
+                random_edge_crop(input_file)
+
+        cmd = mock_ffmpeg.call_args[0][0]
+        vf_idx = cmd.index("-vf")
+        vf_value = cmd[vf_idx + 1]
+        # 提取 crop 参数并验证都是偶数
+        import re
+
+        crop_match = re.search(r"crop=(\d+):(\d+)", vf_value)
+        assert crop_match
+        crop_w = int(crop_match.group(1))
+        crop_h = int(crop_match.group(2))
+        assert crop_w % 2 == 0, f"crop width {crop_w} should be even"
+        assert crop_h % 2 == 0, f"crop height {crop_h} should be even"

tests/unit/test_recompute_dedup_api.py

@@ -0,0 +1,167 @@
+"""Tests for POST /videos/recompute-dedup endpoint (#1664 follow-up)."""
+
+from unittest.mock import MagicMock, patch
+
+import pytest
+from fastapi.testclient import TestClient
+
+
+@pytest.fixture
+def mock_video():
+    """Mock video with missing dedup data."""
+    v = MagicMock()
+    v.id = "video-001"
+    v.user_id = "user-abc"
+    v.duplicate_rate = None
+    v.video_fingerprint = None
+    v.project_id = "proj-001"
+    v.generation_task_id = "task-001"
+    v.name = "test.mp4"
+    v.file_url = "https://example.com/test.mp4"
+    v.file_size = 1024
+    v.duration = 10.0
+    v.width = 1920
+    v.height = 1080
+    v.fps = 25.0
+    v.status = "completed"
+    v.review_status = "pending_review"
+    v.generation_params = {}
+    v.thumbnail_url = None
+    v.is_duplicate = False
+    v.duplicate_of = None
+    v.match_count = None
+    v.visual_similarity = None
+    v.generated_at = "2026-09-04T00:00:00"
+    return v
+
+
+@pytest.fixture
+def mock_video_with_dedup(mock_video):
+    """Mock video that already has dedup data."""
+    mock_video.duplicate_rate = 15.5
+    mock_video.video_fingerprint = {"md5": "abc123"}
+    return mock_video
+
+
+class TestRecomputeDedupEndpoint:
+    """POST /videos/recompute-dedup"""
+
+    def test_enqueue_videos_without_dedup(self, mock_video):
+        """Videos missing duplicate_rate should be enqueued."""
+        from app.api.routes.videos import RecomputeDedupRequest
+
+        mock_repo = MagicMock()
+        mock_repo.list_by_user.return_value = [mock_video]
+
+        with (patch("app.api.routes.videos.celery_app") as mock_celery,):
+            mock_celery.send_task.return_value = MagicMock(id="task-xyz")
+            from app.api.routes.videos import recompute_dedup
+
+            auth_user = MagicMock()
+            auth_user.user.id = "user-abc"
+
+            result = recompute_dedup(
+                request=RecomputeDedupRequest(),
+                repo=mock_repo,
+                current_user=auth_user,
+            )
+
+        assert result.enqueued == 1
+        assert result.total_scanned == 1
+        assert result.skipped == 0
+        mock_celery.send_task.assert_called_once_with("worker.check_duplicate", args=["video-001"])
+
+    def test_skip_videos_with_complete_dedup(self, mock_video_with_dedup):
+        """Videos with both duplicate_rate and video_fingerprint should be skipped."""
+        from app.api.routes.videos import RecomputeDedupRequest, recompute_dedup
+
+        mock_repo = MagicMock()
+        mock_repo.list_by_user.return_value = [mock_video_with_dedup]
+
+        with patch("app.api.routes.videos.celery_app") as mock_celery:
+            auth_user = MagicMock()
+            auth_user.user.id = "user-abc"
+
+            result = recompute_dedup(
+                request=RecomputeDedupRequest(),
+                repo=mock_repo,
+                current_user=auth_user,
+            )
+
+        assert result.enqueued == 0
+        assert result.total_scanned == 1
+        assert result.skipped == 1
+        mock_celery.send_task.assert_not_called()
+
+    def test_specific_video_ids(self, mock_video):
+        """When video_ids are provided, only those videos are processed."""
+        from app.api.routes.videos import RecomputeDedupRequest, recompute_dedup
+
+        mock_repo = MagicMock()
+        mock_repo.get_by_ids.return_value = [mock_video]
+
+        with patch("app.api.routes.videos.celery_app") as mock_celery:
+            mock_celery.send_task.return_value = MagicMock(id="task-xyz")
+            auth_user = MagicMock()
+            auth_user.user.id = "user-abc"
+
+            result = recompute_dedup(
+                request=RecomputeDedupRequest(video_ids=["video-001"]),
+                repo=mock_repo,
+                current_user=auth_user,
+            )
+
+        assert result.enqueued == 1
+        mock_repo.get_by_ids.assert_called_once_with(["video-001"])
+
+    def test_security_only_own_videos(self, mock_video):
+        """Videos belonging to other users should be filtered out."""
+        from app.api.routes.videos import RecomputeDedupRequest, recompute_dedup
+
+        mock_video.user_id = "user-OTHER"
+        mock_repo = MagicMock()
+        mock_repo.get_by_ids.return_value = [mock_video]
+
+        with patch("app.api.routes.videos.celery_app") as mock_celery:
+            auth_user = MagicMock()
+            auth_user.user.id = "user-abc"
+
+            result = recompute_dedup(
+                request=RecomputeDedupRequest(video_ids=["video-001"]),
+                repo=mock_repo,
+                current_user=auth_user,
+            )
+
+        assert result.enqueued == 0
+        mock_celery.send_task.assert_not_called()
+
+    def test_mixed_complete_and_incomplete(self, mock_video, mock_video_with_dedup):
+        """Mix of videos with and without dedup data."""
+        import copy
+
+        from app.api.routes.videos import RecomputeDedupRequest, recompute_dedup
+
+        # Create a second video object
+        v2 = MagicMock()
+        v2.id = "video-002"
+        v2.user_id = "user-abc"
+        v2.duplicate_rate = None
+        v2.video_fingerprint = None
+
+        mock_repo = MagicMock()
+        mock_repo.list_by_user.return_value = [mock_video_with_dedup, v2]
+
+        with patch("app.api.routes.videos.celery_app") as mock_celery:
+            mock_celery.send_task.return_value = MagicMock(id="task-xyz")
+            auth_user = MagicMock()
+            auth_user.user.id = "user-abc"
+
+            result = recompute_dedup(
+                request=RecomputeDedupRequest(),
+                repo=mock_repo,
+                current_user=auth_user,
+            )
+
+        assert result.enqueued == 1
+        assert result.total_scanned == 2
+        assert result.skipped == 1

tests/unit/test_render_subtitles.py

@@ -67,11 +67,11 @@ class TestPositionToAssAlignment:
     def test_bottom(self):
         assert _position_to_ass_alignment("bottom") == 2
 
-    def test_unknown_returns_top_default(self):
-        assert _position_to_ass_alignment("unknown") == 8
-        assert _position_to_ass_alignment("") == 8
-        assert _position_to_ass_alignment("left") == 8
-        assert _position_to_ass_alignment(None) == 8
+    def test_unknown_returns_bottom_default(self):
+        assert _position_to_ass_alignment("unknown") == 2
+        assert _position_to_ass_alignment("") == 2
+        assert _position_to_ass_alignment("left") == 2
+        assert _position_to_ass_alignment(None) == 2
 
 
 class TestBuildAssStyle:

tests/unit/test_render_subtitles_pure.py

@@ -66,15 +66,15 @@ class TestPositionToAssAlignment:
         """center → 居中(5)."""
         assert _position_to_ass_alignment("center") == 5
 
-    def test_unknown_defaults_to_top(self):
-        """未知位置默认顶部(8)."""
-        assert _position_to_ass_alignment("unknown") == 8
-        assert _position_to_ass_alignment("top_left") == 8
-        assert _position_to_ass_alignment("bottom_right") == 8
+    def test_unknown_defaults_to_bottom(self):
+        """未知位置默认底部(2)."""
+        assert _position_to_ass_alignment("unknown") == 2
+        assert _position_to_ass_alignment("top_left") == 2
+        assert _position_to_ass_alignment("bottom_right") == 2
 
-    def test_empty_string_defaults_to_top(self):
-        """空字符串默认顶部."""
-        assert _position_to_ass_alignment("") == 8
+    def test_empty_string_defaults_to_bottom(self):
+        """空字符串默认底部."""
+        assert _position_to_ass_alignment("") == 2
 
 
 class TestBuildAssStyle:

tests/unit/test_video_response_dup_fields.py

@@ -0,0 +1,123 @@
+"""#1660 成品视频 API 查重字段透传测试。
+
+覆盖两套响应构造路径：
+- routes/videos.py::_to_video_response        -> VideoItemResponse (/videos 列表)
+- routes/generation_tasks.py::_to_generated_video_response -> GeneratedVideoResponse
+"""
+
+from types import SimpleNamespace
+from unittest.mock import MagicMock
+
+from app.api.routes.generation_tasks import _to_generated_video_response
+from app.api.routes.videos import _to_video_response
+from app.schemas.generated_video import GeneratedVideoResponse
+from app.schemas.video_center import VideoItemResponse
+
+
+def _make_item(**overrides):
+    base = dict(
+        id="v1",
+        project_id="p1",
+        generation_task_id="t1",
+        name="成片",
+        file_url="oss://bucket/v1.mp4",
+        file_size=1024,
+        duration=12.5,
+        thumbnail_url=None,
+        width=1080,
+        height=1920,
+        fps=30.0,
+        status="completed",
+        review_status="pending_review",
+        generation_params={},
+        generated_at=None,
+        duplicate_rate=None,
+        match_count=None,
+        visual_similarity=None,
+    )
+    base.update(overrides)
+    return SimpleNamespace(**base)
+
+
+class TestVideoItemResponseDupFields:
+    def test_passes_through_all_three_fields(self):
+        item = _make_item(duplicate_rate=42.5, match_count=7, visual_similarity=0.83)
+        resp = _to_video_response(item, storage=None)
+        assert isinstance(resp, VideoItemResponse)
+        assert resp.duplicate_rate == 42.5
+        assert resp.match_count == 7
+        assert resp.visual_similarity == 0.83
+
+    def test_legacy_video_without_fields_returns_none(self):
+        """老数据/实体无查重字段时保持 None（前端自动隐藏），不报错。"""
+        item = SimpleNamespace(
+            id="v2",
+            project_id="p1",
+            generation_task_id="t2",
+            name="老视频",
+            file_url="oss://bucket/v2.mp4",
+            file_size=1,
+            duration=1.0,
+            thumbnail_url=None,
+            width=720,
+            height=1280,
+            fps=24.0,
+            status="completed",
+            review_status="pending_review",
+            generation_params={},
+        )
+        resp = _to_video_response(item, storage=None)
+        assert resp.duplicate_rate is None
+        assert resp.match_count is None
+        assert resp.visual_similarity is None
+
+    def test_explicit_none_values_kept(self):
+        item = _make_item()
+        resp = _to_video_response(item, storage=None)
+        assert resp.duplicate_rate is None
+        assert resp.match_count is None
+        assert resp.visual_similarity is None
+
+    def test_zero_match_count_is_valid_value(self):
+        """计算后确无匹配：match_count=0 / visual_similarity=0.0 是合法值，不能变 None。"""
+        item = _make_item(duplicate_rate=0.0, match_count=0, visual_similarity=0.0)
+        resp = _to_video_response(item, storage=None)
+        assert resp.match_count == 0
+        assert resp.visual_similarity == 0.0
+
+
+class TestGeneratedVideoResponseDupFields:
+    def test_passes_through_all_three_fields(self):
+        item = _make_item(duplicate_rate=15.2, match_count=3, visual_similarity=0.61)
+        resp = _to_generated_video_response(item, download_url="https://dl/x")
+        assert isinstance(resp, GeneratedVideoResponse)
+        assert resp.duplicate_rate == 15.2
+        assert resp.match_count == 3
+        assert resp.visual_similarity == 0.61
+        assert resp.download_url == "https://dl/x"
+
+    def test_missing_fields_default_none(self):
+        item = SimpleNamespace(
+            id="v3",
+            project_id="p1",
+            generation_task_id="t3",
+            name="x",
+            file_url="oss://x",
+            file_size=1,
+            duration=1.0,
+            thumbnail_url=None,
+            width=720,
+            height=1280,
+            fps=24.0,
+        )
+        resp = _to_generated_video_response(item)
+        assert resp.duplicate_rate is None
+        assert resp.match_count is None
+        assert resp.visual_similarity is None
+
+    def test_storage_failure_falls_back_to_file_url(self):
+        storage = MagicMock()
+        storage.get_download_url.side_effect = RuntimeError("oss down")
+        item = _make_item()
+        resp = _to_video_response(item, storage=storage)
+        assert resp.download_url == item.file_url

tests/unit/test_voice_duration_alignment.py

@@ -0,0 +1,205 @@
+"""Tests for voice duration alignment feature.
+
+Tests the _align_clips_to_voice_duration method in UnifiedRenderService.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+from unittest.mock import MagicMock, patch
+
+import pytest
+from video_processing.unified_render_service import RenderLayer, ResolvedClip, UnifiedRenderService
+
+
+class TestAlignClipsToVoiceDuration:
+    """Test clip duration alignment to voice audio."""
+
+    def _make_clip(
+        self,
+        clip_id: str,
+        duration: float,
+        actual_duration: float = 0.0,
+        playback_speed: float = 1.0,
+    ) -> ResolvedClip:
+        """Helper to create a ResolvedClip for testing."""
+        return ResolvedClip(
+            clip_id=clip_id,
+            asset_id=f"asset_{clip_id}",
+            local_path=Path(f"/tmp/{clip_id}.mp4"),
+            clip_type="main",
+            order=0,
+            duration=duration,
+            actual_duration=actual_duration or duration,
+            playback_speed=playback_speed,
+        )
+
+    def _make_layer(self, role: str, clips: list[ResolvedClip]) -> RenderLayer:
+        """Helper to create a RenderLayer for testing."""
+        return RenderLayer(role=role, clips=clips, z_index=0)
+
+    def _make_service(self, voiceover_path: str | None = None) -> UnifiedRenderService:
+        """Helper to create a mock UnifiedRenderService."""
+        plan = MagicMock()
+        plan.id = "test_plan"
+        plan.config = {}
+
+        with patch.object(UnifiedRenderService, "__init__", lambda self, **kwargs: None):
+            service = UnifiedRenderService.__new__(UnifiedRenderService)
+            service.plan = plan
+            service.voiceover_audio_path = voiceover_path
+            service.transition_duration = 0.0
+            return service
+
+    def test_no_voice_audio_no_adjustment(self):
+        """No voice audio → no adjustment."""
+        service = self._make_service(voiceover_path=None)
+        clips = [self._make_clip("c1", 10.0), self._make_clip("c2", 10.0)]
+        layers = [self._make_layer("main", clips)]
+
+        service._align_clips_to_voice_duration(layers, voice_duration=0.0)
+
+        # No change
+        assert clips[0].duration == 10.0
+        assert clips[1].duration == 10.0
+
+    def test_ratio_within_5_percent_no_adjustment(self):
+        """Ratio within ±5% → no adjustment."""
+        service = self._make_service()
+        clips = [self._make_clip("c1", 10.0)]
+        layers = [self._make_layer("main", clips)]
+
+        # Total clips = 10s, voice = 10.3s → ratio = 1.03 (within 5%)
+        service._align_clips_to_voice_duration(layers, voice_duration=10.3)
+
+        assert clips[0].duration == 10.0  # Unchanged
+
+    def test_ratio_less_than_1_trim_clips(self):
+        """Ratio < 1 (clips too long) → trim clips proportionally."""
+        service = self._make_service()
+        clips = [self._make_clip("c1", 10.0), self._make_clip("c2", 10.0)]
+        layers = [self._make_layer("main", clips)]
+
+        # Total clips = 20s, voice = 15s → ratio = 0.75
+        service._align_clips_to_voice_duration(layers, voice_duration=15.0)
+
+        # Each clip should be trimmed to 75%
+        assert abs(clips[0].duration - 7.5) < 0.01
+        assert abs(clips[1].duration - 7.5) < 0.01
+
+    def test_ratio_greater_than_1_slowdown_clips(self):
+        """Ratio > 1 (clips too short) → slow down clips."""
+        service = self._make_service()
+        clips = [self._make_clip("c1", 10.0), self._make_clip("c2", 10.0)]
+        layers = [self._make_layer("main", clips)]
+
+        # Total clips = 20s, voice = 25s → ratio = 1.25
+        service._align_clips_to_voice_duration(layers, voice_duration=25.0)
+
+        # Each clip's speed should be reduced: 1.0 / 1.25 = 0.8
+        assert abs(clips[0].playback_speed - 0.8) < 0.01
+        assert abs(clips[1].playback_speed - 0.8) < 0.01
+
+    def test_speed_lower_bound_025(self):
+        """Playback speed should not go below 0.25x."""
+        service = self._make_service()
+        clips = [self._make_clip("c1", 5.0)]
+        layers = [self._make_layer("main", clips)]
+
+        # Total clips = 5s, voice = 50s → ratio = 10.0
+        # Speed would be 1.0 / 10 = 0.1, but should be clamped to 0.25
+        service._align_clips_to_voice_duration(layers, voice_duration=50.0)
+
+        assert clips[0].playback_speed == 0.25
+
+    def test_only_video_layers_adjusted(self):
+        """Only main/broll/background layers are adjusted, not audio."""
+        service = self._make_service()
+
+        video_clips = [self._make_clip("v1", 10.0)]
+        audio_clips = [self._make_clip("a1", 10.0)]
+
+        layers = [
+            self._make_layer("main", video_clips),
+            self._make_layer("audio", audio_clips),
+        ]
+
+        # ratio = 0.5 → should trim video but not audio
+        service._align_clips_to_voice_duration(layers, voice_duration=5.0)
+
+        assert abs(video_clips[0].duration - 5.0) < 0.01  # Trimmed
+        assert audio_clips[0].duration == 10.0  # Unchanged
+
+    def test_multiple_video_layers_all_adjusted(self):
+        """All video layers (main, broll, background) are adjusted."""
+        service = self._make_service()
+
+        main_clips = [self._make_clip("m1", 10.0)]
+        broll_clips = [self._make_clip("b1", 10.0)]
+        bg_clips = [self._make_clip("bg1", 10.0)]
+
+        layers = [
+            self._make_layer("main", main_clips),
+            self._make_layer("broll", broll_clips),
+            self._make_layer("background", bg_clips),
+        ]
+
+        # Total video = 30s, voice = 15s → ratio = 0.5
+        service._align_clips_to_voice_duration(layers, voice_duration=15.0)
+
+        # All should be trimmed to 50%
+        assert abs(main_clips[0].duration - 5.0) < 0.01
+        assert abs(broll_clips[0].duration - 5.0) < 0.01
+        assert abs(bg_clips[0].duration - 5.0) < 0.01
+
+    def test_trim_config_also_adjusted(self):
+        """When clip has trim_config, it should also be adjusted."""
+        from video_processing.trim_engine import TrimConfig
+
+        service = self._make_service()
+
+        clip = self._make_clip("c1", 10.0)
+        clip.trim_config = TrimConfig(start_time=0.0, duration=10.0)
+
+        layers = [self._make_layer("main", [clip])]
+
+        # ratio = 0.5
+        service._align_clips_to_voice_duration(layers, voice_duration=5.0)
+
+        assert abs(clip.duration - 5.0) < 0.01
+        assert clip.trim_config is not None
+        assert abs(clip.trim_config.duration - 5.0) < 0.01
+
+
+class TestGetVoiceAudioDuration:
+    """Test voice audio duration probing."""
+
+    def test_no_voiceover_path_returns_zero(self):
+        """No voiceover path → return 0."""
+        with patch.object(UnifiedRenderService, "__init__", lambda self, **kwargs: None):
+            service = UnifiedRenderService.__new__(UnifiedRenderService)
+            service.voiceover_audio_path = None
+
+        assert service._get_voice_audio_duration() == 0.0
+
+    def test_nonexistent_file_returns_zero(self):
+        """Nonexistent file → return 0."""
+        with patch.object(UnifiedRenderService, "__init__", lambda self, **kwargs: None):
+            service = UnifiedRenderService.__new__(UnifiedRenderService)
+            service.voiceover_audio_path = "/nonexistent/path.mp3"
+
+        assert service._get_voice_audio_duration() == 0.0
+
+    @patch("video_processing.unified_render_service.probe_duration")
+    @patch("video_processing.unified_render_service.Path.exists", return_value=True)
+    @patch("video_processing.unified_render_service.Path.stat")
+    def test_probes_duration_from_file(self, mock_stat, mock_exists, mock_probe):
+        """Valid file → probe duration."""
+        mock_stat.return_value.st_size = 1000  # Non-empty file
+        mock_probe.return_value = 42.5
+
+        with patch.object(UnifiedRenderService, "__init__", lambda self, **kwargs: None):
+            service = UnifiedRenderService.__new__(UnifiedRenderService)
+            service.voiceover_audio_path = "/tmp/voice.mp3"
+
+        assert service._get_voice_audio_duration() == 42.5