feat: pHash阈值校准+颜色直方图融合 #1658 (#1674 )

Co-authored-by: xiaoxia <dev@xiaoxiajianji.com> Co-committed-by: xiaoxia <dev@xiaoxiajianji.com>
feat(worker): 手动查重 worker task + visual_similarity/match_count 字段 #1661 (#1679 )
2026-09-04 02:06:58 +08:00 · 2026-09-04 01:28:35 +08:00 · 2026-09-04 01:03:35 +08:00 · 2026-09-04 00:32:38 +08:00 · 2026-09-04 00:11:51 +08:00 · 2026-09-03 23:16:50 +08:00
169 changed files with 12298 additions and 1147 deletions
@@ -0,0 +1 @@
+CI re-trigger after runner add-host/DNS fix. This file is harmless and not referenced.
@@ -0,0 +1,105 @@
+name: CI Base Image Build
+
+on:
+  push:
+    branches:
+      - develop
+      - main
+    paths:
+      - 'requirements-base.txt'
+      - 'requirements-dev.txt'
+      - 'infra/docker/ci.Dockerfile'
+  workflow_dispatch:
+    inputs:
+      reason:
+        description: "触发原因"
+        required: false
+        default: "手动触发 - ci-base 镜像重建"
+
+concurrency:
+  group: ci-base-image-build
+  cancel-in-progress: false
+
+jobs:
+  build-ci-base:
+    name: Build CI Base Image
+    runs-on: runtime-builder
+    timeout-minutes: 60
+    steps:
+    - name: Checkout code
+      shell: sh
+      env:
+        GITHUB_TOKEN: ${{ github.token }}
+      run: |
+        curl -sH "Authorization: token $GITHUB_TOKEN" \
+          "${GITHUB_API_URL}/repos/${GITHUB_REPOSITORY}/raw/scripts/ci/step_checkout.sh?ref=${GITHUB_SHA}" \
+          | bash
+
+    - name: Docker login to Gitea Registry
+      shell: sh
+      env:
+        GITEA_REGISTRY_USER: xiaoxia
+        GITEA_REGISTRY_TOKEN: ${{ secrets.REGISTRY_TOKEN }}
+      run: |
+        set -eu
+        for i in 1 2 3; do
+          echo "=== Docker login 尝试 $i/3 ==="
+          if docker login git.xiaoxiajianji.com -u "${GITEA_REGISTRY_USER}" -p "${GITEA_REGISTRY_TOKEN}"; then
+            echo "✅ Docker login successful"
+            break
+          fi
+          echo "❌ Docker login 失败（尝试 $i/3），5s 后重试..."
+          sleep 5
+        done
+
+    - name: Build and push CI base image
+      shell: sh
+      run: |
+        set -eu
+        IMAGE="git.xiaoxiajianji.com/xiaoxia/xiaoxia-saas/ci-base"
+        VERSION_TAG="deps-$(date +%Y%m%d-%H%M)-${GITHUB_SHA::8}"
+
+        echo "=== Building CI base image (tags: latest, ${VERSION_TAG}) ==="
+        docker build --progress=plain \
+          -f infra/docker/ci.Dockerfile \
+          -t "${IMAGE}:latest" \
+          -t "${IMAGE}:${VERSION_TAG}" \
+          .
+        echo "✅ Image built successfully"
+
+        echo "=== Pushing ${VERSION_TAG} ==="
+        docker push "${IMAGE}:${VERSION_TAG}"
+        echo "=== Pushing latest ==="
+        docker push "${IMAGE}:latest"
+        echo "✅ Pushed to Gitea Registry"
+
+    - name: Verify image
+      shell: sh
+      run: |
+        set -eu
+        IMAGE="git.xiaoxiajianji.com/xiaoxia/xiaoxia-saas/ci-base:latest"
+        echo "=== Verifying pinned deps in fresh image ==="
+        docker run --rm "${IMAGE}" /opt/xiaoxia-ci-venv/bin/python -c \
+          "import httpcore, h2, numpy, httpx; print('VERSIONS:', httpcore.__version__, h2.__version__, numpy.__version__, httpx.__version__)"
+
+    - name: Notify result
+      if: always()
+      continue-on-error: true
+      shell: sh
+      env:
+        CI_NOTIFY_WEBHOOK: ${{ secrets.CI_NOTIFY_WEBHOOK }}
+      run: |
+        set +e
+        if [ "${{ job.status }}" = "success" ]; then
+          NOTIFY_MODE=success JOB_NAME="CI Base Image Build" python3 scripts/ci_notify.py
+        else
+          NOTIFY_MODE=failure JOB_NAME="CI Base Image Build" python3 scripts/ci_notify.py
+        fi
+
+    - name: Cleanup
+      if: always()
+      shell: sh
+      run: |
+        IMAGE="git.xiaoxiajianji.com/xiaoxia/xiaoxia-saas/ci-base"
+        docker rmi "${IMAGE}:latest" 2>/dev/null || true
+        echo "Cleanup done"
@@ -0,0 +1,51 @@
+name: CI Canary Check
+on:
+  schedule:
+    - cron: '*/30 * * * *'
+  workflow_dispatch:
+
+jobs:
+  canary:
+    runs-on: ci-l2
+    timeout-minutes: 10
+    steps:
+      - name: Canary (runner -> docker -> network -> gitea)
+        run: |
+          set -e
+          echo "== runner/container basic =="
+          date; hostname; whoami
+          echo "== gitea api reachability =="
+          code=$(curl -s -o /tmp/v.json -w '%{http_code}' -m 15 "$GITHUB_API_URL/version")
+          echo "gitea api http_code=$code"
+          [ "$code" = "200" ] || { echo "::error::Gitea API unreachable, http_code=$code"; exit 1; }
+          cat /tmp/v.json; echo
+          echo "== external egress =="
+          ext=$(curl -s -o /dev/null -w '%{http_code}' -m 15 https://www.baidu.com || echo 000)
+          echo "external http_code=$ext"
+          echo "== gitea domain resolves NOT to loopback =="
+          set -o pipefail
+          ip=$(getent hosts git.xiaoxiajianji.com | awk '{print $1}' | head -1)
+          echo "git.xiaoxiajianji.com -> $ip"
+          if [ -z "$ip" ]; then
+            echo "::error::DNS resolution failed, git.xiaoxiajianji.com unresolvable"; exit 1
+          fi
+          if [ "$ip" = "127.0.0.1" ] || [ "$ip" = "::1" ]; then
+            echo "::error::Gitea domain resolves to loopback inside job container (hosts/DNS leak)"; exit 1
+          fi
+          echo "CANARY OK"
+      - name: Notify failure
+        if: failure()
+        env:
+          CI_NOTIFY_WEBHOOK: ${{ secrets.CI_NOTIFY_WEBHOOK }}
+        run: |
+          set +e
+          if [ -n "$CI_NOTIFY_WEBHOOK" ]; then
+            MSG="🚨 CI 金丝雀失败：runner->docker->网络->Gitea 链路异常，时间 $(date '+%Y-%m-%d %H:%M:%S')，请立即检查构建服务器"
+            python3 - "$CI_NOTIFY_WEBHOOK" "$MSG" <<'PY'
+          import json,sys,urllib.request
+          hook,msg=sys.argv[1],sys.argv[2]
+          data=json.dumps({"msg_type":"text","content":{"text":msg}}).encode()
+          urllib.request.urlopen(urllib.request.Request(hook,data=data,headers={"Content-Type":"application/json"}),timeout=10)
+          PY
+          fi
+          exit 0
@@ -58,10 +58,11 @@ jobs:
          echo "::warning::PR #${PR_NUMBER} 已合并，测试由合并后 push 流水线承接，PR 侧测试类 job 跳过"
          exit 0
        fi
-        # 情形2：同一 head_sha 已有在跑/排队的 push 流水线（rebase/ff 合并竞态）
+        # 情形2：同一 head_sha 已有在跑/排队/已成功的 push 流水线（rebase/ff 合并竞态；
+        #        已成功也去重——竞态窗口内两边都通过判断时，PR侧再跑全量测试纯属重复）
        DUP=$(curl -sfH "Authorization: token $GITHUB_TOKEN" \
          "${GITHUB_API_URL}/repos/${GITHUB_REPOSITORY}/actions/runs?head_sha=${HEAD_SHA}&per_page=30" \
-          | python3 -c "import json,sys; d=json.load(sys.stdin); runs=d if isinstance(d,list) else d.get('workflow_runs',d.get('runs',[])); hit=[r for r in runs if r.get('event')=='push' and r.get('status') in ('in_progress','queued','waiting','pending')]; print('true' if hit else 'false')" || echo false)
+          | python3 -c "import json,sys; d=json.load(sys.stdin); runs=d if isinstance(d,list) else d.get('workflow_runs',d.get('runs',[])); hit=[r for r in runs if r.get('event')=='push' and (r.get('status') in ('in_progress','queued','waiting','pending') or r.get('conclusion')=='success')]; print('true' if hit else 'false')" || echo false)
        if [ "$DUP" = "true" ]; then
          echo "skip_tests=true" >> $GITHUB_OUTPUT
          echo "reason=duplicate-push-run-active" >> $GITHUB_OUTPUT
@@ -80,12 +81,6 @@ jobs:
      skip_backend: ${{ steps.check.outputs.skip_backend }}
      skip_frontend: ${{ steps.check.outputs.skip_frontend }}
    steps:
-    - name: Checkout code
-      shell: sh
-      env:
-        GITHUB_TOKEN: ${{ github.token }}
-      run: |
-        curl -sfH "Authorization: token $GITHUB_TOKEN" -o /tmp/_ci_checkout.sh "${GITHUB_API_URL}/repos/${GITHUB_REPOSITORY}/raw/scripts/ci/step_checkout.sh?ref=${GITHUB_SHA}" && bash /tmp/_ci_checkout.sh
    - name: Check changed files
      id: check
      shell: bash
@@ -124,14 +119,14 @@ jobs:
        [ ${{ job.status }} = "success" ] || STATUS="error"
        START_TIME=""
        [ -f /tmp/ci_job_start_time ] && START_TIME=$(cat /tmp/ci_job_start_time)
-        python3 scripts/ci/ci_trace_report.py --service xiaoxia-saas-ci --status $STATUS --start-time "$START_TIME" || true
+        curl -sfH "Authorization: token ${GITHUB_TOKEN:-$GITEA_TOKEN}" -o /tmp/_ci_trace.py "${GITHUB_API_URL}/repos/${GITHUB_REPOSITORY}/raw/scripts/ci/ci_trace_report.py?ref=${GITHUB_SHA}" 2>/dev/null && python3 /tmp/_ci_trace.py --service xiaoxia-saas-ci --status $STATUS --start-time "$START_TIME" || true

-  validate-code-quality:
+  validate-style:
    needs: dedupe-check
    if: always() && needs.dedupe-check.outputs.skip_tests != 'true'
-    name: Validate - Code Quality
+    name: Validate - Style
    runs-on: ci-l2
-    timeout-minutes: 8
+    timeout-minutes: 6
    env:
      PIP_CACHE_DIR: /root/.cache/pip
      PIP_NO_CACHE_DIR: ''
@@ -144,16 +139,23 @@ jobs:
        GITHUB_TOKEN: ${{ github.token }}
      run: |
        curl -sfH "Authorization: token $GITHUB_TOKEN" -o /tmp/_ci_checkout.sh "${GITHUB_API_URL}/repos/${GITHUB_REPOSITORY}/raw/scripts/ci/step_checkout.sh?ref=${GITHUB_SHA}" && bash /tmp/_ci_checkout.sh
+    - name: Zombie run selfcheck
+      shell: sh
+      env:
+        GITHUB_TOKEN: ${{ github.token }}
+        GITHUB_RUN_ID: ${{ github.run_id }}
+      run: bash scripts/ci/ci_run_selfcheck.sh
    - name: Record job start time
      shell: sh
      run: bash scripts/ci/step_timer_start.sh
    - name: Cache pip dependencies
      uses: actions/cache@v4
+      continue-on-error: true
      with:
        path: /root/.cache/pip
-        key: ${{ runner.os }}-pip-codequality-${{ hashFiles('requirements*.txt') }}
+        key: ${{ runner.os }}-pip-style-${{ hashFiles('requirements*.txt') }}
        restore-keys: |
-          ${{ runner.os }}-pip-codequality-
+          ${{ runner.os }}-pip-style-
          ${{ runner.os }}-pip-
    - name: Install dependencies
      shell: sh
@@ -177,17 +179,9 @@ jobs:
          [ $i -eq 3 ] && exit 1
          sleep 5
        done
-        for i in 1 2 3; do
-          python3 -m pip install --no-binary :all: black==26.5.1 isort==8.0.1 && break
-          echo "pip install black/isort 失败，重试 $i/3..."
-          [ $i -eq 3 ] && exit 1
-          sleep 5
-        done
-    - name: Run code quality and security checks
+    - name: Run style checks
      shell: bash
-      env:
-        GITHUB_TOKEN: ${{ github.token }}
-      run: bash scripts/ci/validate_code_quality.sh
+      run: bash scripts/ci/validate_style.sh
    - name: Auto-fix formatting (black + isort)
      if: failure()
      shell: sh
@@ -203,7 +197,7 @@ jobs:
        CI_WEBHOOK_URL: ${{ secrets.CI_WEBHOOK_URL }}
      run: |
        set +e
-        FAILED_JOB="Validate - Code Quality" python3 scripts/ci_notify_failure.py
+        FAILED_JOB="Validate - Style" python3 scripts/ci_notify_failure.py
    - name: Job duration summary
      if: always()
      shell: sh
@@ -216,7 +210,7 @@ jobs:
        CI_NOTIFY_WEBHOOK: ${{ secrets.CI_NOTIFY_WEBHOOK }}
      run: |
        set +e
-        NOTIFY_MODE=failure JOB_NAME="Validate - Code Quality" python3 scripts/ci_notify.py
+        NOTIFY_MODE=failure JOB_NAME="Validate - Style" python3 scripts/ci_notify.py
    - name: Report CI trace
      if: always()
      shell: sh
@@ -229,12 +223,16 @@ jobs:
        [ -f /tmp/ci_job_start_time ] && START_TIME=$(cat /tmp/ci_job_start_time)
        python3 scripts/ci/ci_trace_report.py --service xiaoxia-saas-ci --status $STATUS --start-time "$START_TIME" || true

-  validate-type-check:
+
+  validate-security:
    needs: dedupe-check
    if: always() && needs.dedupe-check.outputs.skip_tests != 'true'
-    name: Validate - Type Check (mypy)
+    name: Validate - Security
    runs-on: ci-l2
    timeout-minutes: 8
+    env:
+      PIP_CACHE_DIR: /root/.cache/pip
+      PIP_NO_CACHE_DIR: ''
    permissions:
      contents: read
    steps:
@@ -244,9 +242,126 @@ jobs:
        GITHUB_TOKEN: ${{ github.token }}
      run: |
        curl -sfH "Authorization: token $GITHUB_TOKEN" -o /tmp/_ci_checkout.sh "${GITHUB_API_URL}/repos/${GITHUB_REPOSITORY}/raw/scripts/ci/step_checkout.sh?ref=${GITHUB_SHA}" && bash /tmp/_ci_checkout.sh
+    - name: Zombie run selfcheck
+      shell: sh
+      env:
+        GITHUB_TOKEN: ${{ github.token }}
+        GITHUB_RUN_ID: ${{ github.run_id }}
+      run: bash scripts/ci/ci_run_selfcheck.sh
    - name: Record job start time
      shell: sh
      run: bash scripts/ci/step_timer_start.sh
+    - name: Cache pip dependencies
+      uses: actions/cache@v4
+      continue-on-error: true
+      with:
+        path: /root/.cache/pip
+        key: ${{ runner.os }}-pip-security-${{ hashFiles('requirements*.txt') }}
+        restore-keys: |
+          ${{ runner.os }}-pip-security-
+          ${{ runner.os }}-pip-
+    - name: Install dependencies
+      shell: sh
+      run: |
+        set -eu
+        for i in 1 2 3; do
+          python3 -m pip install -q -r requirements-base.txt && break
+          echo "pip install requirements-base.txt 失败，重试 $i/3..."
+          [ $i -eq 3 ] && exit 1
+          sleep 5
+        done
+        for i in 1 2 3; do
+          python3 -m pip install -q -r requirements.txt && break
+          echo "pip install requirements.txt 失败，重试 $i/3..."
+          [ $i -eq 3 ] && exit 1
+          sleep 5
+        done
+        for i in 1 2 3; do
+          python3 -m pip install -q -r requirements-dev.txt && break
+          echo "pip install requirements-dev.txt 失败，重试 $i/3..."
+          [ $i -eq 3 ] && exit 1
+          sleep 5
+        done
+    - name: Run security checks
+      continue-on-error: true  # Security scan is advisory; runner failure must not block deploy
+      shell: bash
+      env:
+        GITHUB_TOKEN: ${{ github.token }}
+      run: bash scripts/ci/validate_security.sh
+    - name: CI failure notification
+      if: failure()
+      shell: sh
+      env:
+        GITEA_TOKEN: ${{ secrets.GITEA_TOKEN }}
+        CI_WEBHOOK_URL: ${{ secrets.CI_WEBHOOK_URL }}
+      run: |
+        set +e
+        FAILED_JOB="Validate - Security" python3 scripts/ci_notify_failure.py
+    - name: Job duration summary
+      if: always()
+      shell: sh
+      run: bash scripts/ci/step_timer_end.sh
+    - name: Notify on failure
+      continue-on-error: true
+      if: failure()
+      shell: sh
+      env:
+        CI_NOTIFY_WEBHOOK: ${{ secrets.CI_NOTIFY_WEBHOOK }}
+      run: |
+        set +e
+        NOTIFY_MODE=failure JOB_NAME="Validate - Security" python3 scripts/ci_notify.py
+    - name: Report CI trace
+      if: always()
+      shell: sh
+      env:
+        AGENTLOOP_LICENSE_KEY: ${{ secrets.AGENTLOOP_LICENSE_KEY }}
+      run: |
+        STATUS="ok"
+        [ ${{ job.status }} = "success" ] || STATUS="error"
+        START_TIME=""
+        [ -f /tmp/ci_job_start_time ] && START_TIME=$(cat /tmp/ci_job_start_time)
+        python3 scripts/ci/ci_trace_report.py --service xiaoxia-saas-ci --status $STATUS --start-time "$START_TIME" || true
+
+
+  validate-python:
+    needs: dedupe-check
+    if: always() && needs.dedupe-check.outputs.skip_tests != 'true'
+    name: Validate - Python (mypy + alembic)
+    runs-on: ci-l2
+    timeout-minutes: 10
+    env:
+      PIP_CACHE_DIR: /root/.cache/pip
+      PIP_NO_CACHE_DIR: ''
+      DATABASE_URL: postgresql+psycopg://postgres:postgres@host.docker.internal:5432/xiaoxia_saas
+      USE_IN_MEMORY_DB: 'false'
+      CI_USE_SHARED_PG: 'false'
+    permissions:
+      contents: read
+    steps:
+    - name: Checkout code
+      shell: sh
+      env:
+        GITHUB_TOKEN: ${{ github.token }}
+      run: |
+        curl -sfH "Authorization: token $GITHUB_TOKEN" -o /tmp/_ci_checkout.sh "${GITHUB_API_URL}/repos/${GITHUB_REPOSITORY}/raw/scripts/ci/step_checkout.sh?ref=${GITHUB_SHA}" && bash /tmp/_ci_checkout.sh
+    - name: Zombie run selfcheck
+      shell: sh
+      env:
+        GITHUB_TOKEN: ${{ github.token }}
+        GITHUB_RUN_ID: ${{ github.run_id }}
+      run: bash scripts/ci/ci_run_selfcheck.sh
+    - name: Record job start time
+      shell: sh
+      run: bash scripts/ci/step_timer_start.sh
+    - name: Cache pip dependencies
+      uses: actions/cache@v4
+      continue-on-error: true
+      with:
+        path: /root/.cache/pip
+        key: ${{ runner.os }}-pip-python-${{ hashFiles('requirements*.txt') }}
+        restore-keys: |
+          ${{ runner.os }}-pip-python-
+          ${{ runner.os }}-pip-
    - name: Install dependencies
      shell: sh
      run: |
@@ -272,6 +387,9 @@ jobs:
    - name: Run mypy type check
      shell: bash
      run: bash scripts/ci/validate_mypy.sh
+    - name: Run alembic migration validation
+      shell: bash
+      run: bash scripts/ci/validate_migration.sh
    - name: CI failure notification
      if: failure()
      shell: sh
@@ -280,7 +398,7 @@ jobs:
        CI_WEBHOOK_URL: ${{ secrets.CI_WEBHOOK_URL }}
      run: |
        set +e
-        FAILED_JOB="Validate - Type Check (mypy)" python3 scripts/ci_notify_failure.py
+        FAILED_JOB="Validate - Python (mypy + alembic)" python3 scripts/ci_notify_failure.py
    - name: Job duration summary
      if: always()
      shell: sh
@@ -293,7 +411,7 @@ jobs:
        CI_NOTIFY_WEBHOOK: ${{ secrets.CI_NOTIFY_WEBHOOK }}
      run: |
        set +e
-        NOTIFY_MODE=failure JOB_NAME="Validate - Type Check (mypy)" python3 scripts/ci_notify.py
+        NOTIFY_MODE=failure JOB_NAME="Validate - Python (mypy + alembic)" python3 scripts/ci_notify.py
    - name: Report CI trace
      if: always()
      shell: sh
@@ -306,86 +424,6 @@ jobs:
        [ -f /tmp/ci_job_start_time ] && START_TIME=$(cat /tmp/ci_job_start_time)
        python3 scripts/ci/ci_trace_report.py --service xiaoxia-saas-ci --status $STATUS --start-time "$START_TIME" || true

-  validate-migration:
-    needs: dedupe-check
-    if: always() && needs.dedupe-check.outputs.skip_tests != 'true'
-    name: Validate - Migration (alembic)
-    runs-on: ci-l2
-    timeout-minutes: 8
-    permissions:
-      contents: read
-    env:
-      DATABASE_URL: postgresql+psycopg://postgres:postgres@host.docker.internal:5432/xiaoxia_saas
-      USE_IN_MEMORY_DB: 'false'
-      CI_USE_SHARED_PG: 'true'
-    steps:
-    - name: Checkout code
-      shell: sh
-      env:
-        GITHUB_TOKEN: ${{ github.token }}
-      run: |
-        curl -sfH "Authorization: token $GITHUB_TOKEN" -o /tmp/_ci_checkout.sh "${GITHUB_API_URL}/repos/${GITHUB_REPOSITORY}/raw/scripts/ci/step_checkout.sh?ref=${GITHUB_SHA}" && bash /tmp/_ci_checkout.sh
-    - name: Record job start time
-      shell: sh
-      run: bash scripts/ci/step_timer_start.sh
-    - name: Install dependencies
-      shell: sh
-      run: |
-        set -eu
-        for i in 1 2 3; do
-          python3 -m pip install -q -r requirements-base.txt && break
-          echo "pip install requirements-base.txt 失败，重试 $i/3..."
-          [ $i -eq 3 ] && exit 1
-          sleep 5
-        done
-        for i in 1 2 3; do
-          python3 -m pip install -q -r requirements.txt && break
-          echo "pip install requirements.txt 失败，重试 $i/3..."
-          [ $i -eq 3 ] && exit 1
-          sleep 5
-        done
-        for i in 1 2 3; do
-          python3 -m pip install -q -r requirements-dev.txt && break
-          echo "pip install requirements-dev.txt 失败，重试 $i/3..."
-          [ $i -eq 3 ] && exit 1
-          sleep 5
-        done
-    - name: Run alembic migration validation
-      shell: bash
-      run: bash scripts/ci/validate_migration.sh
-    - name: CI failure notification
-      if: failure()
-      shell: sh
-      env:
-        GITEA_TOKEN: ${{ secrets.GITEA_TOKEN }}
-        CI_WEBHOOK_URL: ${{ secrets.CI_WEBHOOK_URL }}
-      run: |
-        set +e
-        FAILED_JOB="Validate - Migration (alembic)" python3 scripts/ci_notify_failure.py
-    - name: Job duration summary
-      if: always()
-      shell: sh
-      run: bash scripts/ci/step_timer_end.sh
-    - name: Notify on failure
-      continue-on-error: true
-      if: failure()
-      shell: sh
-      env:
-        CI_NOTIFY_WEBHOOK: ${{ secrets.CI_NOTIFY_WEBHOOK }}
-      run: |
-        set +e
-        NOTIFY_MODE=failure JOB_NAME="Validate - Migration (alembic)" python3 scripts/ci_notify.py
-    - name: Report CI trace
-      if: always()
-      shell: sh
-      env:
-        AGENTLOOP_LICENSE_KEY: ${{ secrets.AGENTLOOP_LICENSE_KEY }}
-      run: |
-        STATUS="ok"
-        [ ${{ job.status }} = "success" ] || STATUS="error"
-        START_TIME=""
-        [ -f /tmp/ci_job_start_time ] && START_TIME=$(cat /tmp/ci_job_start_time)
-        python3 scripts/ci/ci_trace_report.py --service xiaoxia-saas-ci --status $STATUS --start-time "$START_TIME" || true

  unit-tests:
    needs: [check-frontend-only, dedupe-check]
@@ -408,6 +446,12 @@ jobs:
        GITHUB_TOKEN: ${{ github.token }}
      run: |
        curl -sfH "Authorization: token $GITHUB_TOKEN" -o /tmp/_ci_checkout.sh "${GITHUB_API_URL}/repos/${GITHUB_REPOSITORY}/raw/scripts/ci/step_checkout.sh?ref=${GITHUB_SHA}" && bash /tmp/_ci_checkout.sh
+    - name: Zombie run selfcheck
+      shell: sh
+      env:
+        GITHUB_TOKEN: ${{ github.token }}
+        GITHUB_RUN_ID: ${{ github.run_id }}
+      run: bash scripts/ci/ci_run_selfcheck.sh
    - name: Record job start time
      shell: sh
      run: bash scripts/ci/step_timer_start.sh
@@ -416,6 +460,7 @@ jobs:
      run: bash scripts/ci/step_install_ffmpeg.sh
    - name: Cache pip dependencies
      uses: actions/cache@v4
+      continue-on-error: true
      with:
        path: /root/.cache/pip
        key: ${{ runner.os }}-pip-unittests-${{ hashFiles('requirements*.txt') }}
@@ -470,13 +515,10 @@ jobs:
    needs:
      - check-frontend-only
      - dedupe-check
-      - validate-code-quality
-      - validate-type-check
-      - validate-migration
    env:
      DATABASE_URL: postgresql+psycopg://postgres:postgres@host.docker.internal:5432/xiaoxia_saas
      USE_IN_MEMORY_DB: 'false'
-      CI_USE_SHARED_PG: 'true'
+      CI_USE_SHARED_PG: 'false'
      OSS_ACCESS_KEY_ID: placeholder
      OSS_ACCESS_KEY_SECRET: placeholder
      OSS_BUCKET_NAME: xiaoxia-autocut
@@ -488,6 +530,12 @@ jobs:
        GITHUB_TOKEN: ${{ github.token }}
      run: |
        curl -sfH "Authorization: token $GITHUB_TOKEN" -o /tmp/_ci_checkout.sh "${GITHUB_API_URL}/repos/${GITHUB_REPOSITORY}/raw/scripts/ci/step_checkout.sh?ref=${GITHUB_SHA}" && bash /tmp/_ci_checkout.sh
+    - name: Zombie run selfcheck
+      shell: sh
+      env:
+        GITHUB_TOKEN: ${{ github.token }}
+        GITHUB_RUN_ID: ${{ github.run_id }}
+      run: bash scripts/ci/ci_run_selfcheck.sh
    - name: Record job start time
      shell: sh
      run: bash scripts/ci/step_timer_start.sh
@@ -542,6 +590,12 @@ jobs:
        GITHUB_TOKEN: ${{ github.token }}
      run: |
        curl -sfH "Authorization: token $GITHUB_TOKEN" -o /tmp/_ci_checkout.sh "${GITHUB_API_URL}/repos/${GITHUB_REPOSITORY}/raw/scripts/ci/step_checkout.sh?ref=${GITHUB_SHA}" && bash /tmp/_ci_checkout.sh
+    - name: Zombie run selfcheck
+      shell: sh
+      env:
+        GITHUB_TOKEN: ${{ github.token }}
+        GITHUB_RUN_ID: ${{ github.run_id }}
+      run: bash scripts/ci/ci_run_selfcheck.sh
    - name: Record job start time
      shell: sh
      run: bash scripts/ci/step_timer_start.sh
@@ -604,11 +658,18 @@ jobs:
        GITHUB_TOKEN: ${{ github.token }}
      run: |
        curl -sfH "Authorization: token $GITHUB_TOKEN" -o /tmp/_ci_checkout.sh "${GITHUB_API_URL}/repos/${GITHUB_REPOSITORY}/raw/scripts/ci/step_checkout.sh?ref=${GITHUB_SHA}" && bash /tmp/_ci_checkout.sh
+    - name: Zombie run selfcheck
+      shell: sh
+      env:
+        GITHUB_TOKEN: ${{ github.token }}
+        GITHUB_RUN_ID: ${{ github.run_id }}
+      run: bash scripts/ci/ci_run_selfcheck.sh
    - name: Record job start time
      shell: sh
      run: bash scripts/ci/step_timer_start.sh
    - name: Cache npm dependencies
      uses: actions/cache@v4
+      continue-on-error: true
      with:
        path: /root/.npm
        key: ${{ runner.os }}-npm-vitest-${{ hashFiles('apps/web/package-lock.json') }}
@@ -695,6 +756,12 @@ jobs:
        GITHUB_TOKEN: ${{ github.token }}
      run: |
        curl -sfH "Authorization: token $GITHUB_TOKEN" -o /tmp/_ci_checkout.sh "${GITHUB_API_URL}/repos/${GITHUB_REPOSITORY}/raw/scripts/ci/step_checkout.sh?ref=${GITHUB_SHA}" && bash /tmp/_ci_checkout.sh
+    - name: Zombie run selfcheck
+      shell: sh
+      env:
+        GITHUB_TOKEN: ${{ github.token }}
+        GITHUB_RUN_ID: ${{ github.run_id }}
+      run: bash scripts/ci/ci_run_selfcheck.sh
    - name: Record job start time
      shell: sh
      run: bash scripts/ci/step_timer_start.sh
@@ -809,19 +876,13 @@ jobs:
      skip_backend: ${{ steps.check.outputs.skip_backend }}
      skip_frontend: ${{ steps.check.outputs.skip_frontend }}
    steps:
-    - name: Checkout code
-      shell: sh
-      env:
-        GITHUB_TOKEN: ${{ github.token }}
-      run: |
-        curl -sfH "Authorization: token $GITHUB_TOKEN" -o /tmp/_ci_checkout.sh "${GITHUB_API_URL}/repos/${GITHUB_REPOSITORY}/raw/scripts/ci/step_checkout.sh?ref=${GITHUB_SHA}" && bash /tmp/_ci_checkout.sh
    - name: Check changed paths
      id: check
      shell: bash
      env:
        GITHUB_TOKEN: ${{ github.token }}
      run: |
-        bash scripts/ci/ci_push_paths.sh
+        curl -sfH "Authorization: token $GITHUB_TOKEN" -o /tmp/_ci_push_paths.sh "${GITHUB_API_URL}/repos/${GITHUB_REPOSITORY}/raw/scripts/ci/ci_push_paths.sh?ref=${GITHUB_SHA}" && bash /tmp/_ci_push_paths.sh
    - name: Report CI trace
      if: always()
      shell: sh
@@ -832,7 +893,7 @@ jobs:
        [ ${{ job.status }} = "success" ] || STATUS="error"
        START_TIME=""
        [ -f /tmp/ci_job_start_time ] && START_TIME=$(cat /tmp/ci_job_start_time)
-        python3 scripts/ci/ci_trace_report.py --service xiaoxia-saas-ci --status $STATUS --start-time "$START_TIME" || true
+        curl -sfH "Authorization: token ${GITHUB_TOKEN:-$GITEA_TOKEN}" -o /tmp/_ci_trace.py "${GITHUB_API_URL}/repos/${GITHUB_REPOSITORY}/raw/scripts/ci/ci_trace_report.py?ref=${GITHUB_SHA}" 2>/dev/null && python3 /tmp/_ci_trace.py --service xiaoxia-saas-ci --status $STATUS --start-time "$START_TIME" || true

  build-staging:
    name: Build Staging ${{ matrix.service_display }} Image
@@ -873,6 +934,12 @@ jobs:
        GITHUB_TOKEN: ${{ github.token }}
      run: |
        curl -sfH "Authorization: token $GITHUB_TOKEN" -o /tmp/_ci_checkout.sh "${GITHUB_API_URL}/repos/${GITHUB_REPOSITORY}/raw/scripts/ci/step_checkout.sh?ref=${GITHUB_SHA}" && bash /tmp/_ci_checkout.sh
+    - name: Zombie run selfcheck
+      shell: sh
+      env:
+        GITHUB_TOKEN: ${{ github.token }}
+        GITHUB_RUN_ID: ${{ github.run_id }}
+      run: bash scripts/ci/ci_run_selfcheck.sh
    - name: Record job start time
      shell: sh
      run: bash scripts/ci/step_timer_start.sh
@@ -1084,9 +1151,7 @@ jobs:
    - check-push-paths
    - build-staging
    - retag-staging-skipped
-    # 显式 success() 状态检查：上游 build/retag 被路径过滤 if 跳过(skipped)时不阻塞本 job；
-    # 上游真正失败时仍然阻断（act_runner 对无状态函数的 if 隐式包 success()，纯 skipped 也会连带跳过）
-    if: success() && github.event_name == 'push' && (github.ref_name == 'main' || github.ref_name == 'develop')
+    if: (!cancelled()) && github.event_name == 'push' && (github.ref_name == 'main' || github.ref_name == 'develop')
    steps:
    - name: Checkout code
      shell: sh
@@ -1105,6 +1170,31 @@ jobs:
      run: |
        set +e
        NOTIFY_MODE=start JOB_NAME="Deploy Staging" python3 scripts/ci_notify.py
+    - name: Render .env from template
+      shell: sh
+      env:
+        STAGING_DATABASE_URL: ${{ secrets.STAGING_DATABASE_URL }}
+        STAGING_REDIS_URL: ${{ secrets.STAGING_REDIS_URL }}
+        STAGING_CELERY_BROKER_URL: ${{ secrets.STAGING_CELERY_BROKER_URL }}
+        STAGING_CELERY_RESULT_BACKEND: ${{ secrets.STAGING_CELERY_RESULT_BACKEND }}
+        STAGING_JWT_SECRET_KEY: ${{ secrets.STAGING_JWT_SECRET_KEY }}
+        STAGING_MINIO_ENDPOINT: ${{ secrets.STAGING_MINIO_ENDPOINT }}
+        STAGING_MINIO_ACCESS_KEY: ${{ secrets.STAGING_MINIO_ACCESS_KEY }}
+        STAGING_MINIO_SECRET_KEY: ${{ secrets.STAGING_MINIO_SECRET_KEY }}
+        STAGING_MINIO_BUCKET: ${{ secrets.STAGING_MINIO_BUCKET }}
+        OSS_ACCESS_KEY_ID: ${{ secrets.OSS_ACCESS_KEY_ID }}
+        OSS_ACCESS_KEY_SECRET: ${{ secrets.OSS_ACCESS_KEY_SECRET }}
+        COSYVOICE_API_KEY: ${{ secrets.COSYVOICE_API_KEY }}
+        DASHSCOPE_API_KEY: ${{ secrets.DASHSCOPE_API_KEY }}
+        MEDIAKIT_API_KEY: ${{ secrets.MEDIAKIT_API_KEY }}
+      run: |
+        set -eu
+        echo "Rendering .env from template + secrets..."
+        bash scripts/render_env.sh staging
+        echo "✅ .env rendered (file contains secrets, not printed to log)"
+        # 验证文件存在且非空
+        test -s .env.rendered
+        echo "✅ .env.rendered validated ($(wc -l < .env.rendered) lines)"
    - name: Docker login to Registry
      shell: sh
      env:
@@ -1177,9 +1267,31 @@ jobs:
        ssh -p "$staging_port" -i "$key_path" -o StrictHostKeyChecking=no "${staging_user}@${staging_host}" "echo SSH_CONNECTION_OK && hostname"
        echo "SSH connection verified"

+        # 配置 Diff 检查：下载服务器当前 .env，对比渲染结果，检测漂移
+        echo "Running config diff check..."
+        scp -P "$staging_port" -i "$key_path" -o StrictHostKeyChecking=no \
+          "${staging_user}@${staging_host}:/var/lib/xiaoxia-saas-staging/.env" .env.current 2>/dev/null \
+          || touch .env.current  # 首次部署时文件不存在，创建空文件
+        bash scripts/config_diff_check.sh .env.rendered .env.current
+        rm -f .env.current
+        echo "Config diff check done"
+
+        # 上传渲染后的 .env 到服务器（替代服务器上旧的 .env）
+        echo "Uploading rendered .env to staging server..."
+        # 备份旧 .env
+        ssh -p "$staging_port" -i "$key_path" -o StrictHostKeyChecking=no "${staging_user}@${staging_host}" \
+          "cp -f /var/lib/xiaoxia-saas-staging/.env /var/lib/xiaoxia-saas-staging/.env.bak.\$(date +%Y%m%d%H%M%S) 2>/dev/null || true"
+        # 上传新 .env
+        scp -P "$staging_port" -i "$key_path" -o StrictHostKeyChecking=no .env.rendered \
+          "${staging_user}@${staging_host}:/var/lib/xiaoxia-saas-staging/.env"
+        echo "✅ .env uploaded to staging server"
+
        # 通过环境变量传递凭证，避免命令行引号转义问题
        cat scripts/ci_staging_deploy.sh | ssh -p "$staging_port" -i "$key_path" -o StrictHostKeyChecking=no "${staging_user}@${staging_host}" "IMAGE_TAG=${GITHUB_SHA} ACR_USERNAME=${ACR_USERNAME} ACR_PASSWORD=${ACR_PASSWORD} sh"

+        # 清理 CI runner 上的渲染文件
+        rm -f .env.rendered
+
    - name: Staging health check + auto rollback
      if: success()
      shell: sh
@@ -1246,6 +1358,12 @@ jobs:
        GITHUB_TOKEN: ${{ github.token }}
      run: |
        curl -sfH "Authorization: token $GITHUB_TOKEN" -o /tmp/_ci_checkout.sh "${GITHUB_API_URL}/repos/${GITHUB_REPOSITORY}/raw/scripts/ci/step_checkout.sh?ref=${GITHUB_SHA}" && bash /tmp/_ci_checkout.sh
+    - name: Zombie run selfcheck
+      shell: sh
+      env:
+        GITHUB_TOKEN: ${{ github.token }}
+        GITHUB_RUN_ID: ${{ github.run_id }}
+      run: bash scripts/ci/ci_run_selfcheck.sh
    - name: Record job start time
      shell: sh
      run: bash scripts/ci/step_timer_start.sh
@@ -1293,6 +1411,12 @@ jobs:
        GITHUB_TOKEN: ${{ github.token }}
      run: |
        curl -sfH "Authorization: token $GITHUB_TOKEN" -o /tmp/_ci_checkout.sh "${GITHUB_API_URL}/repos/${GITHUB_REPOSITORY}/raw/scripts/ci/step_checkout.sh?ref=${GITHUB_SHA}" && bash /tmp/_ci_checkout.sh
+    - name: Zombie run selfcheck
+      shell: sh
+      env:
+        GITHUB_TOKEN: ${{ github.token }}
+        GITHUB_RUN_ID: ${{ github.run_id }}
+      run: bash scripts/ci/ci_run_selfcheck.sh
    - name: Record job start time
      shell: sh
      run: bash scripts/ci/step_timer_start.sh
@@ -1332,12 +1456,13 @@ jobs:
    runs-on: runtime-builder
    timeout-minutes: ${{ matrix.timeout }}
    needs:
-      - validate-code-quality
-      - validate-type-check
+      - validate-style
+      - validate-security
+      - validate-python
      - unit-tests
      - frontend-lint
      - frontend-unit-test
-    if: startsWith(github.ref, 'refs/tags/v') || (github.event_name == 'push' && github.ref_name == 'main')
+    if: github.event_name == 'push' && github.ref_name == 'main' && !failure() && !cancelled()
    strategy:
      fail-fast: false
      matrix:
@@ -1367,6 +1492,12 @@ jobs:
        GITHUB_TOKEN: ${{ github.token }}
      run: |
        curl -sfH "Authorization: token $GITHUB_TOKEN" -o /tmp/_ci_checkout.sh "${GITHUB_API_URL}/repos/${GITHUB_REPOSITORY}/raw/scripts/ci/step_checkout.sh?ref=${GITHUB_SHA}" && bash /tmp/_ci_checkout.sh
+    - name: Zombie run selfcheck
+      shell: sh
+      env:
+        GITHUB_TOKEN: ${{ github.token }}
+        GITHUB_RUN_ID: ${{ github.run_id }}
+      run: bash scripts/ci/ci_run_selfcheck.sh
    - name: Record job start time
      shell: sh
      run: bash scripts/ci/step_timer_start.sh
@@ -1415,11 +1546,7 @@ jobs:
        set -eu
        REGISTRY="xiaoxia-registry.cn-hangzhou.cr.aliyuncs.com/xiaoxiakeji"
        # 根据ref类型设置镜像标签：tag用版本号，分支用分支名+sha
-        if [[ "$GITHUB_REF" == refs/tags/* ]]; then
-          TAG_NAME="${GITHUB_REF_NAME}"
-        else
-          TAG_NAME="${GITHUB_REF_NAME}-${GITHUB_SHA::8}"
-        fi
+        TAG_NAME="${GITHUB_SHA}"
        IMAGE_TAG="${REGISTRY}/${{ matrix.image_name }}:${TAG_NAME}"
        CACHE_REF="${REGISTRY}/${{ matrix.cache_name }}:main"

@@ -1481,7 +1608,7 @@ jobs:
    concurrency:
      group: deploy-production-${{ gitea.ref }}
      cancel-in-progress: false
-    if: startsWith(github.ref, 'refs/tags/v')
+    if: github.event_name == 'push' && github.ref_name == 'main'
    needs:
    - build-production
    steps:
@@ -1554,7 +1681,7 @@ jobs:
        echo "SSH connection verified"

        # 通过环境变量传递凭证，避免命令行引号转义问题
-        cat scripts/ci_production_deploy.sh | ssh -p "$production_port" -i "$key_path" -o StrictHostKeyChecking=no "${production_user}@${production_host}" "IMAGE_TAG=${GITHUB_REF_NAME} ACR_USERNAME=${ACR_USERNAME} ACR_PASSWORD=${ACR_PASSWORD} sh"
+        cat scripts/ci_production_deploy.sh | ssh -p "$production_port" -i "$key_path" -o StrictHostKeyChecking=no "${production_user}@${production_host}" "IMAGE_TAG=${GITHUB_SHA} ACR_USERNAME=${ACR_USERNAME} ACR_PASSWORD=${ACR_PASSWORD} sh"

    - name: Production health check + auto rollback
      if: success()
@@ -1613,7 +1740,7 @@ jobs:
    name: Production Browser E2E
    runs-on: runtime-builder
    timeout-minutes: 15
-    if: startsWith(github.ref, 'refs/tags/v')
+    # if: removed - runs after deploy-production succeeds
    needs: deploy-production
    steps:
    - name: Checkout code
@@ -1813,9 +1940,9 @@ jobs:
    if: always() && github.event_name == 'pull_request'
    needs:
      - check-frontend-only
-      - validate-code-quality
-      - validate-type-check
-      - validate-migration
+      - validate-style
+      - validate-security
+      - validate-python
      - unit-tests
      - integration-tests
      - frontend-lint
@@ -1823,14 +1950,6 @@ jobs:
      - build-pr
    timeout-minutes: 3
    steps:
-      - name: Checkout code
-        shell: sh
-        env:
-          GITHUB_TOKEN: ${{ github.token }}
-        run: |
-          curl -sfH "Authorization: token $GITHUB_TOKEN" -o /tmp/_ci_checkout.sh \
-            "${GITHUB_API_URL}/repos/${GITHUB_REPOSITORY}/raw/scripts/ci/step_checkout.sh?ref=${GITHUB_SHA}" && bash /tmp/_ci_checkout.sh
-
      - name: Evaluate CI Gate
        id: gate
        shell: bash
@@ -1839,9 +1958,9 @@ jobs:
          PR_NUMBER: ${{ github.event.pull_request.number }}
          PR_HEAD_SHA: ${{ github.event.pull_request.head.sha }}
          RESULT_CHECK_FRONTEND: ${{ needs.check-frontend-only.result }}
-          RESULT_CODE_QUALITY: ${{ needs.validate-code-quality.result }}
-          RESULT_TYPE_CHECK: ${{ needs.validate-type-check.result }}
-          RESULT_MIGRATION: ${{ needs.validate-migration.result }}
+          RESULT_STYLE: ${{ needs.validate-style.result }}
+          RESULT_SECURITY: ${{ needs.validate-security.result }}
+          RESULT_PYTHON: ${{ needs.validate-python.result }}
          RESULT_UNIT_TESTS: ${{ needs.unit-tests.result }}
          RESULT_INTEGRATION: ${{ needs.integration-tests.result }}
          RESULT_FRONTEND_LINT: ${{ needs.frontend-lint.result }}
@@ -1853,9 +1972,9 @@ jobs:
          echo ""
          echo "各job结果:"
          echo "  check-frontend-only: $RESULT_CHECK_FRONTEND"
-          echo "  validate-code-quality: $RESULT_CODE_QUALITY"
-          echo "  validate-type-check: $RESULT_TYPE_CHECK"
-          echo "  validate-migration: $RESULT_MIGRATION"
+          echo "  validate-style: $RESULT_STYLE"
+          echo "  validate-security: $RESULT_SECURITY"
+          echo "  validate-python: $RESULT_PYTHON"
          echo "  unit-tests: $RESULT_UNIT_TESTS"
          echo "  integration-tests: $RESULT_INTEGRATION"
          echo "  frontend-lint: $RESULT_FRONTEND_LINT"
@@ -1895,9 +2014,9 @@ jobs:
          # 必填检查项（根据PR类型决定）
          # 通用检查（所有PR都必须过）
          REQUIRED_GENERAL=(
-            "validate-code-quality:$RESULT_CODE_QUALITY"
-            "validate-type-check:$RESULT_TYPE_CHECK"
-            "validate-migration:$RESULT_MIGRATION"
+            "validate-style:$RESULT_STYLE"
+            "validate-security:$RESULT_SECURITY"
+            "validate-python:$RESULT_PYTHON"
            "frontend-lint:$RESULT_FRONTEND_LINT"
            "build-pr:$RESULT_BUILD_PR"
            "ai-code-review:$AI_REVIEW_STATUS"
@@ -1941,6 +2060,11 @@ jobs:
              echo "  ⏳ $name: pending（审查中，暂不阻塞）"
              continue
            fi
+            # Security scan cancelled/failed时不阻塞部署（runner故障不应卡住流水线）
+            if [ "$name" = "validate-security" ] && { [ "$result" = "cancelled" ] || [ "$result" = "failure" ]; }; then
+              echo "  ⚠️  $name: $result（安全扫描为非阻塞项，不卡住部署）"
+              continue
+            fi
            check_job "$name" "$result"
          done

@@ -1992,4 +2116,4 @@ jobs:
          [ "${{ steps.gate.outputs.gate_result }}" = "success" ] || STATUS="error"
          START_TIME=""
          [ -f /tmp/ci_job_start_time ] && START_TIME=$(cat /tmp/ci_job_start_time)
-          python3 scripts/ci/ci_trace_report.py --service xiaoxia-saas-ci --status $STATUS --start-time "$START_TIME" || true
+          curl -sfH "Authorization: token ${GITHUB_TOKEN:-$GITEA_TOKEN}" -o /tmp/_ci_trace.py "${GITHUB_API_URL}/repos/${GITHUB_REPOSITORY}/raw/scripts/ci/ci_trace_report.py?ref=${GITHUB_SHA}" 2>/dev/null && python3 /tmp/_ci_trace.py --service xiaoxia-saas-ci --status $STATUS --start-time "$START_TIME" || true
@@ -2,13 +2,13 @@ name: CI Trigger Monitor

 on:
  schedule:
-    - cron: '*/5 * * * *'  # 每5分钟检查一次
+    - cron: '*/10 * * * *'  # 每10分钟检查一次（与pr-auto-scan同步降频）
  workflow_dispatch:
    inputs:
      stale_threshold:
        description: 'CI未触发告警阈值（分钟）'
        required: false
-        default: '5'
+        default: '10'

 permissions:
  contents: read
@@ -0,0 +1,59 @@
+name: Playwright Base Image Build
+
+on:
+  workflow_dispatch:
+    inputs:
+      reason:
+        description: "触发原因"
+        required: false
+        default: "构建 playwright 基础镜像"
+
+jobs:
+  build-playwright:
+    name: Build Playwright Base Image
+    runs-on: runtime-builder
+    timeout-minutes: 30
+    steps:
+    - name: Docker login to Gitea Registry
+      shell: sh
+      env:
+        GITEA_REGISTRY_USER: xiaoxia
+        GITEA_REGISTRY_TOKEN: ${{ secrets.REGISTRY_TOKEN }}
+      run: |
+        set -eu
+        for i in 1 2 3; do
+          echo "=== Docker login attempt $i/3 ==="
+          if printf '%s' "${GITEA_REGISTRY_TOKEN}" | docker login git.xiaoxiajianji.com -u "${GITEA_REGISTRY_USER}" --password-stdin; then
+            echo "Docker login successful"
+            break
+          fi
+          echo "Docker login failed (attempt $i/3), retrying in 5s..."
+          sleep 5
+          [ $i -eq 3 ] && exit 1
+        done
+
+    - name: Pull, retag and push Playwright image
+      shell: sh
+      run: |
+        set -eu
+        OFFICIAL_IMAGE="mcr.microsoft.com/playwright:v1.45.0-jammy"
+        GITEA_IMAGE="git.xiaoxiajianji.com/xiaoxia/base/playwright:v1.45.0-jammy"
+
+        echo "=== Pulling official Playwright image ==="
+        docker pull "${OFFICIAL_IMAGE}"
+
+        echo "=== Tagging ==="
+        docker tag "${OFFICIAL_IMAGE}" "${GITEA_IMAGE}"
+
+        echo "=== Pushing to Gitea Registry ==="
+        docker push "${GITEA_IMAGE}"
+
+        echo "Done: ${GITEA_IMAGE}"
+
+    - name: Cleanup
+      if: always()
+      shell: sh
+      run: |
+        docker rmi "mcr.microsoft.com/playwright:v1.45.0-jammy" 2>/dev/null || true
+        docker rmi "git.xiaoxiajianji.com/xiaoxia/base/playwright:v1.45.0-jammy" 2>/dev/null || true
+        echo "Cleanup done"
@@ -3,7 +3,7 @@ name: PR Auto Scan
 # 作为短作业模式的兜底，防止事件驱动遗漏
 on:
  schedule:
-    - cron: "*/5 * * * *"  # 每5分钟扫描一次
+    # - cron: "*/15 * * * *"  # DISABLED: temporarily to stop failure spam (2026-09-02)  # 每10分钟扫描一次（脚本自带240s墙钟上限，降频减负）
  workflow_dispatch:

 permissions:
@@ -18,7 +18,7 @@ jobs:
    name: Auto Approve on CI Green
    runs-on: ci-check
    if: github.event_name == 'pull_request' && !github.event.pull_request.draft
-    timeout-minutes: 3  # 长等待模式：等CI全绿后自动合并，不遗漏任何PR
+    timeout-minutes: 10  # 等待CI全绿+审批，需要充足时间
    steps:
      - name: Checkout code
        shell: sh
@@ -61,7 +61,8 @@ jobs:
    name: Auto Merge on CI Green + Approved
    runs-on: ci-check
    if: github.event_name == 'pull_request' && !github.event.pull_request.draft && github.event.pull_request.base.ref == 'develop'
-    timeout-minutes: 3  # 短作业模式：检查一次，不满足就退出，由pr-auto-scan每5分钟定时兜底
+    needs: [auto-approve]  # 修复竞态：必须等审批完成后再尝试合并
+    timeout-minutes: 15  # 等待审批+CI就绪+合并，需要充足时间
    steps:
      - name: Checkout code
        shell: sh
@@ -24,6 +24,11 @@ ruff_cache/
 .env.production
 .env.staging
 !.env.example
+# 配置模板不受忽略规则限制
+!deploy/configs/.env.staging
+!deploy/configs/.env.production
+# 渲染后的 env 文件包含真实密钥，绝不能提交
+.env.rendered

 # OS / editor
 .DS_Store
@@ -54,3 +59,4 @@ frontend-v21-ui-prototype-final.html
 !.vscode/settings.json
 .vscode/extensions.json
 .coverage
+.env.current
@@ -0,0 +1,23 @@
+"""add duplicate_rate to generated_videos
+
+Revision ID: 059_duplicate_rate
+Revises: 058_uq_asset_lib_project_kind
+Create Date: 2026-08-31
+"""
+
+import sqlalchemy as sa
+
+from alembic import op
+
+revision = "059_duplicate_rate"
+down_revision = "058_uq_asset_lib_project_kind"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    op.add_column("generated_videos", sa.Column("duplicate_rate", sa.Float(), nullable=True))
+
+
+def downgrade() -> None:
+    op.drop_column("generated_videos", "duplicate_rate")
@@ -0,0 +1,57 @@
+"""migrate template_segments data to template_clip_configs
+
+Revision ID: 060_migrate_segments
+Revises: 059_duplicate_rate
+Create Date: 2026-08-31
+"""
+
+import sqlalchemy as sa
+
+from alembic import op
+
+revision = "060_migrate_segments"
+down_revision = "059_duplicate_rate"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    dialect = op.get_bind().dialect.name
+
+    if dialect == "postgresql":
+        config_expr = (
+            "CASE WHEN s.material_type IS NOT NULL AND s.material_type != '' "
+            "THEN json_build_object('material_type', s.material_type)::jsonb "
+            "ELSE '{}'::jsonb END"
+        )
+        empty_json = "'{}'::jsonb"
+    else:
+        config_expr = (
+            "CASE WHEN s.material_type IS NOT NULL AND s.material_type != '' "
+            "THEN JSON_OBJECT('material_type', s.material_type) "
+            "ELSE '{}' END"
+        )
+        empty_json = "'{}'"
+
+    sql_str = (
+        "INSERT INTO template_clip_configs "
+        '(id, template_id, clip_type, "order", min_duration, max_duration, '
+        "text_template, material_requirements, transition_effect, config, "
+        "created_at, updated_at) "
+        "SELECT "
+        "s.id, s.template_id, 'main', s.segment_order, "
+        "s.duration_min, s.duration_max, "
+        "'', " + empty_json + ", "
+        "'cut', " + config_expr + ", "
+        "s.created_at, s.updated_at "
+        "FROM template_segments s "
+        "WHERE NOT EXISTS ("
+        "  SELECT 1 FROM template_clip_configs c "
+        "  WHERE c.template_id = s.template_id"
+        ")"
+    )
+    op.execute(sa.text(sql_str))
+
+
+def downgrade() -> None:
+    pass
@@ -0,0 +1,26 @@
+"""add sort_order to template_categories
+
+Revision ID: 061_sort_order
+Revises: 060_migrate_segments
+Create Date: 2026-09-02
+"""
+
+import sqlalchemy as sa
+
+from alembic import op
+
+revision = "061_sort_order"
+down_revision = "060_migrate_segments"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    op.add_column(
+        "template_categories",
+        sa.Column("sort_order", sa.Integer, nullable=False, server_default="0"),
+    )
+
+
+def downgrade() -> None:
+    op.drop_column("template_categories", "sort_order")
@@ -0,0 +1,28 @@
+"""re-add edit_plan_id to generation_tasks (align staging with production)
+
+Revision ID: 062_edit_plan_id
+Revises: 061_sort_order
+Create Date: 2026-09-02
+"""
+
+import sqlalchemy as sa
+
+from alembic import op
+
+revision = "062_edit_plan_id"
+down_revision = "061_sort_order"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    op.add_column(
+        "generation_tasks",
+        sa.Column("edit_plan_id", sa.String(36), nullable=True),
+    )
+    op.create_index("ix_generation_tasks_edit_plan_id_2", "generation_tasks", ["edit_plan_id"])
+
+
+def downgrade() -> None:
+    op.drop_index("ix_generation_tasks_edit_plan_id_2", table_name="generation_tasks")
+    op.drop_column("generation_tasks", "edit_plan_id")
@@ -0,0 +1,46 @@
+"""add video_fingerprint_chunks table for per-chunk fingerprint storage
+
+Revision ID: 063_fingerprint_chunks
+Revises: 062_edit_plan_id
+Create Date: 2026-09-03
+"""
+
+import sqlalchemy as sa
+
+from alembic import op
+
+revision = "063_fingerprint_chunks"
+down_revision = "062_edit_plan_id"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    op.create_table(
+        "video_fingerprint_chunks",
+        sa.Column("id", sa.String(36), primary_key=True),
+        sa.Column("video_id", sa.String(36), nullable=False),
+        sa.Column("project_id", sa.String(36), nullable=False),
+        sa.Column("user_id", sa.String(36), nullable=False, server_default=""),
+        sa.Column("start_time_ms", sa.Integer, nullable=False),
+        sa.Column("end_time_ms", sa.Integer, nullable=False),
+        sa.Column("phash_binary", sa.String(16), nullable=False),
+        sa.Column("color_histogram", sa.JSON, nullable=False),
+        sa.Column("frame_count", sa.Integer, nullable=False, server_default="1"),
+        sa.Column(
+            "created_at",
+            sa.DateTime,
+            nullable=False,
+            server_default=sa.func.now(),
+        ),
+    )
+    op.create_index("ix_vfc_video_id", "video_fingerprint_chunks", ["video_id"])
+    op.create_index("ix_vfc_project_id", "video_fingerprint_chunks", ["project_id"])
+    op.create_index("ix_vfc_user_id", "video_fingerprint_chunks", ["user_id"])
+
+
+def downgrade() -> None:
+    op.drop_index("ix_vfc_user_id", table_name="video_fingerprint_chunks")
+    op.drop_index("ix_vfc_project_id", table_name="video_fingerprint_chunks")
+    op.drop_index("ix_vfc_video_id", table_name="video_fingerprint_chunks")
+    op.drop_table("video_fingerprint_chunks")
@@ -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")
@@ -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")
@@ -26,7 +26,7 @@ from app.schemas.asset import (
    UpdateAssetReviewRequest,
 )
 from app.schemas.tag import TagAssetsRequest
-from app.services.asset_segment_tracker import compute_asset_availability
+from app.services.asset_segment_tracker import compute_asset_availability, get_asset_recent_use_counts
 from fastapi import APIRouter, Depends, HTTPException, Query, Response

 from packages.domain.smart_match import smart_select_assets
@@ -290,7 +290,7 @@ def list_assets(
        else:
            total = asset_repository.count_by_project_ids(project_ids, status=status_list)
            # 跨项目分页：逐项目累积直到凑够一页
-            paged_items: list = []
+            paged_items = []
            offset = skip
            remaining = limit
            for pid in project_ids:
@@ -576,47 +576,93 @@ def smart_match_assets(
            request.library_id, request.kind, status=["ready"], limit=10000
        )
    else:
-        filtered_assets = asset_repository.find_by_library(
-            request.library_id, status=["ready"], limit=10000
-        )
+        filtered_assets = asset_repository.find_by_library(request.library_id, status=["ready"], limit=10000)
    total_candidates = len(filtered_assets)

-    # 调用统一智能选素材算法（kind 已在 DB 层过滤，无需重复过滤）
-    results = smart_select_assets(
-        filtered_assets,
-        limit=request.limit,
-        kind=None,
-    )
+    # ── 过滤前置：余量 + 高频使用，过滤在评分/截取 limit 之前完成 ──────────
+    # 旧实现先 smart_select_assets(limit=N) 再对这 N 条做过滤，过滤后不回补，
+    # 当排名靠前的素材恰好都被排除时返回空 items（前端回退全选，smart-match 名存实亡）。
+    # 现在先过滤全量候选，每级过滤后为空/不足则回退上一级，最后才评分截取。

-    # 结果层过滤：usable=false（零重复可切区间耗尽且历史区间均达复用上限）的素材
-    # 不返回给前端；不动 smart_select_assets 评分逻辑本身
-    filtered_results = []
-    for r in results:
+    # 1) 余量过滤：usable=False（零重复可切区间耗尽且历史区间均达复用上限）的素材排除
+    usable_assets = []
+    exhausted_assets = []
+    for a in filtered_assets:
        try:
-            avail = compute_asset_availability(r.asset)
+            avail = compute_asset_availability(a)
        except Exception:
            logger.warning(
                "smart-match 余量计算失败，按可用处理: asset_id=%s",
-                getattr(r.asset, "id", "?"),
+                getattr(a, "id", "?"),
                exc_info=True,
            )
            avail = None
        if avail is not None and not avail["usable"]:
-            logger.info(
-                "smart-match 排除已用尽素材: asset_id=%s name=%s",
-                getattr(r.asset, "id", "?"),
-                getattr(r.asset, "name", ""),
-            )
-            continue
-        filtered_results.append(r)
+            exhausted_assets.append(a)
+        else:
+            usable_assets.append(a)

+    if exhausted_assets:
+        logger.info(
+            "smart-match 余量过滤: 候选 %d，可切区间耗尽 %d",
+            len(filtered_assets), len(exhausted_assets),
+        )
+
+    # 回退策略：余量过滤后为空（全部耗尽）时，保留全部候选，不返回空结果。
+    # 宁可让用户在已耗尽素材上复用，也比 smart-match 空结果回退全选更可控
+    # （全选同样会选到这些素材，且不经过评分排序）。
+    pool = usable_assets if usable_assets else filtered_assets
+
+    # 2) 高频使用排除：同一素材在最近 5 个视频中出现超过 3 次则排除
+    MAX_RECENT_USE_COUNT = 3
+    high_freq_assets = set()
+    if pool:
+        asset_ids = [getattr(a, "id", "") for a in pool if getattr(a, "id", "")]
+        if asset_ids:
+            try:
+                use_counts = get_asset_recent_use_counts(
+                    db=asset_repository.session,
+                    asset_ids=asset_ids,
+                    recent_video_count=5,
+                )
+                for a in pool:
+                    aid = getattr(a, "id", "")
+                    count = use_counts.get(aid, 0)
+                    if count > MAX_RECENT_USE_COUNT:
+                        high_freq_assets.add(aid)
+                        logger.info(
+                            "smart-match 排除高频使用素材: asset_id=%s use_count=%d limit=%d",
+                            aid, count, MAX_RECENT_USE_COUNT,
+                        )
+                # 回退策略：排除后剩余素材不足（为空或不够 limit）时，
+                # 不再全部排除，保留全部可用素材
+                if high_freq_assets:
+                    remaining_count = len(pool) - len(high_freq_assets)
+                    enough = request.limit is None or remaining_count >= request.limit
+                    if remaining_count > 0 and enough:
+                        pool = [a for a in pool if getattr(a, "id", "") not in high_freq_assets]
+                    else:
+                        logger.info(
+                            "smart-match 高频排除后素材不足(%d<%s)，保留全部 %d 条",
+                            remaining_count,
+                            request.limit if request.limit is not None else "不限",
+                            len(pool),
+                        )
+            except Exception:
+                logger.warning("smart-match 高频使用查询失败，跳过排除", exc_info=True)
+
+    # 3) 调用统一智能选素材算法（kind 已在 DB 层过滤，无需重复过滤）
+    results = smart_select_assets(pool, limit=request.limit, kind=None)
+
+    # 扁平结构：SmartMatchItem 继承 AssetResponse，素材字段直接在条目顶层，
+    # 前端无需解析 item.asset 包装层，item.id / item.usable / 余量字段直接可读
    items = [
        SmartMatchItem(
-            asset=_to_asset_response(r.asset),
+            **_to_asset_response(r.asset).model_dump(),
            score=r.score,
            breakdown=r.breakdown,
        )
-        for r in filtered_results
+        for r in results
    ]

    return SmartMatchResponse(items=items, total_candidates=total_candidates)
@@ -456,7 +456,7 @@ async def wechat_callback(
    user = user_repository.find_by_id(response.user_id)
    binding_complete = False
    if user:
-        binding_complete = (
+        binding_complete = bool(
            user.phone_verified and user.email_verified and user.email and "@wechat.local" not in user.email
        )

@@ -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,
@@ -513,7 +513,7 @@ def generate_cover(
        if generation_task_id:
            try:
                task = gen_task_repo.get(generation_task_id)
-                if task and getattr(task, "cover_url", ""):
+                if task and getattr(task, "cover_url", ""):  # type: ignore[arg-type]
                    cover_url_from_task = task.cover_url
                    logger.info(
                        "[封面生成] 统一管道封面(步骤A-direct): plan_id=%s task_id=%s url=%s",
@@ -538,7 +538,7 @@ def generate_cover(
                    gv_task_id = getattr(gv, "generation_task_id", "") or ""
                    if gv_task_id:
                        task_a2 = gen_task_repo.get(gv_task_id)
-                        if task_a2 and getattr(task_a2, "cover_url", ""):
+                        if task_a2 and getattr(task_a2, "cover_url", ""):  # type: ignore[arg-type]
                            cover_url_from_task = task_a2.cover_url
                            logger.info(
                                "[封面生成] 封面(步骤A2-video-task): plan_id=%s video_id=%s url=%s",
@@ -747,7 +747,7 @@ def generate_cover(

        if cover_url_from_task:
            # 标题已在预览视频渲染时烧录（ASS字幕），封面帧自然包含标题
-            cover_data = {
+            cover_data: dict[str, object] = {  # type: ignore[no-redef]
                "type": "ai_frame",
                "image_url": cover_url_from_task,
                "frame_time": 0.0,
@@ -98,7 +98,7 @@ def _resolve_strategy_id_from_template(template_id: str, db: Session, user_id: s
    try:
        new_repo = SQLAlchemyEditTemplateRepository(db)
        new_template = new_repo.get(template_id)
-        if new_template and getattr(new_template, "editing_mode", ""):
+        if new_template and getattr(new_template, "editing_mode", ""):  # type: ignore[arg-type]
            mode = new_template.editing_mode.strip()
            if mode:
                logger.info(
@@ -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(
@@ -385,7 +389,7 @@ def create_generation_task(

    use_case = CreateGenerationTaskUseCase(generation_task_repository)
    count = request.count
-    created_tasks = []
+    created_tasks: list = []
    failed_tasks = []
    user_id = authenticated_user.user.id
    # 同批次任务共享 batch_id，用于视频查重时批次内比对
@@ -23,6 +23,10 @@ import re
 from app.auth import AuthenticatedUser, get_current_user
 from app.core.storage import get_storage_service
 from app.dependencies import get_asset_repository, get_db_session
+
+# 默认转场时长（与 worker 端保持一致）
+_DEFAULT_TRANSITION_DURATION = 0.5
+
 from app.services.asset_segment_tracker import (
    REUSE_RATIO_LIMIT,
    SEGMENT_EDGE_GAP,
@@ -37,10 +41,20 @@ from fastapi import APIRouter, BackgroundTasks, Depends, HTTPException, Query, s
 from sqlalchemy.orm import Session

 from packages.adapters.sqlalchemy_impl.asset_repository import SQLAlchemyAssetRepository
+from packages.adapters.sqlalchemy_impl.template_clip_config_repository import (
+    SQLAlchemyTemplateClipConfigRepository,
+)
 from packages.adapters.sqlalchemy_impl.template_repository import (
    SQLAlchemyTemplateRepository,
 )
-from packages.domain.plan_generator_utils import _calc_random_start_time
+from packages.domain.plan_generator_utils import (
+    _calc_random_start_time,
+    build_scene_segments,
+    extract_scene_points_from_metadata,
+    pick_scene_aware_start,
+    pick_start_in_scene_segment,
+)
+from packages.domain.smart_match import SCORE_RANDOM_NOISE_MAX, score_asset
 from packages.shared.mediakit_client import get_mediakit_client

 from .dependencies import get_draft_plan_id, get_editor_services
@@ -367,7 +381,6 @@ def batch_delete_editor_clips(
    return ClipBatchDeleteResponse(deleted_count=deleted, plan_id=plan_id)


-
 def _safe_segment_duration(value, default: float) -> float:
    """安全地将数据库中的时长值转换为正浮点数.

@@ -403,9 +416,7 @@ def _get_template_segments(
        if clip_configs:
            result = []
            for cc in clip_configs:
-                dur_min = _safe_segment_duration(
-                    cc.min_duration, _DEFAULT_EDITOR_CLIP_DURATION
-                )
+                dur_min = _safe_segment_duration(cc.min_duration, _DEFAULT_EDITOR_CLIP_DURATION)
                dur_max = _safe_segment_duration(
                    cc.max_duration or cc.min_duration,
                    _DEFAULT_EDITOR_CLIP_DURATION,
@@ -414,7 +425,25 @@ def _get_template_segments(
                result.append((cc.order, dur_min, dur_max))
            return sorted(result, key=lambda x: x[0])
    except Exception:
-        logger.warning("新模板系统查询clip_configs失败，回退到旧系统", exc_info=True)
+        logger.warning("新模板系统查询clip_configs失败（主表可能不存在），直接查clip_configs表", exc_info=True)
+
+    # 兜底：直接查 template_clip_configs 表（片段表有 template_id 外键，不依赖模板主表）
+    try:
+        direct_repo = SQLAlchemyTemplateClipConfigRepository(db)
+        direct_configs = direct_repo.list_by_template(template_id)
+        if direct_configs:
+            result = []
+            for cc in direct_configs:
+                dur_min = _safe_segment_duration(cc.min_duration, _DEFAULT_EDITOR_CLIP_DURATION)
+                dur_max = _safe_segment_duration(
+                    cc.max_duration or cc.min_duration,
+                    _DEFAULT_EDITOR_CLIP_DURATION,
+                )
+                dur_min, dur_max = min(dur_min, dur_max), max(dur_min, dur_max)
+                result.append((cc.order, dur_min, dur_max))
+            return sorted(result, key=lambda x: x[0])
+    except Exception:
+        logger.warning("直接查clip_configs表也失败，继续回退旧系统", exc_info=True)

    # 回退到旧模板系统（template_segments表）
    try:
@@ -452,6 +481,12 @@ def _recommended_time_conflicts(
    return False


+# 向后兼容别名：镜头段构建/段内取点逻辑已下沉到 packages.domain.plan_generator_utils，
+# 旧测试与历史代码仍按 clips._build_scene_segments / _pick_start_in_scene_segment 导入
+_build_scene_segments = build_scene_segments
+_pick_start_in_scene_segment = pick_start_in_scene_segment
+
+
 def _get_mediakit_recommendations(
    asset_ids: list[str],
    asset_repo,
@@ -492,7 +527,7 @@ def _get_mediakit_recommendations(
        prompt = (
            "请分析每段视频，找出最精彩的5秒片段应该从哪个时间点开始。"
            "考虑因素：画面清晰度、主体是否明确、是否有明显的动作或场景变化。"
-            '请严格以JSON数组格式返回，不要包含其他文字：'
+            "请严格以JSON数组格式返回，不要包含其他文字："
            '[{"asset_id": "素材ID", "recommended_start_time": 12.5, "reason": "原因"}]'
        )

@@ -545,9 +580,7 @@ def _get_mediakit_recommendations(

            # 尝试正则提取
            if not parsed:
-                time_match = re.search(
-                    r'recommended_start_time["\s:]+([\d.]+)', content_text
-                )
+                time_match = re.search(r'recommended_start_time["\s:]+([\d.]+)', content_text)
                if time_match:
                    try:
                        recommendations[asset_id] = float(time_match.group(1))
@@ -566,6 +599,47 @@ def _get_mediakit_recommendations(
        return {}


+def _calc_plan_internal_duplicate_rate(clips_data: list[dict]) -> float:
+    """估算单条成片内部重复率（%）.
+
+    检查本条成片中同一素材是否有重叠的时间区间。
+    重叠时长 / 成片总时长 * 100 = 内部重复率。
+    这是一个轻量估算，不依赖视频指纹；完整查重由 worker 异步完成。
+    """
+    if not clips_data:
+        return 0.0
+
+    # 按素材分组
+    by_asset: dict[str, list[tuple[float, float]]] = {}
+    total_duration = 0.0
+    for c in clips_data:
+        aid = c.get("asset_id", "")
+        if not aid:
+            continue
+        start = c.get("start_time", 0.0)
+        end = start + c.get("duration", 0.0)
+        by_asset.setdefault(aid, []).append((start, end))
+        total_duration += c.get("duration", 0.0)
+
+    if total_duration <= 0:
+        return 0.0
+
+    # 检查同素材内的区间重叠
+    overlap_duration = 0.0
+    for segments in by_asset.values():
+        if len(segments) < 2:
+            continue
+        segments_sorted = sorted(segments, key=lambda s: s[0])
+        last_end = segments_sorted[0][1]
+        for start, end in segments_sorted[1:]:
+            overlap = max(0.0, min(end, last_end) - start)
+            if overlap > 0:
+                overlap_duration += overlap
+            last_end = max(last_end, end)
+
+    return round(overlap_duration / total_duration * 100, 1)
+
+
@router.post("/clips/from-assets", response_model=ClipsFromAssetsResponse)
 def create_clips_from_assets_editor(
    template_id: str,
@@ -598,26 +672,43 @@ def create_clips_from_assets_editor(
            detail="模板没有片段配置，无法创建片段",
        )

-    if not body.asset_ids:
+    # 防御：schema validator 已过滤 null/空串，这里再归一化一次，
+    # 避免异常入参（undefined → null）导致后续 /assets/{id} 404 / 422
+    asset_ids = [str(aid).strip() for aid in (body.asset_ids or []) if isinstance(aid, str) and aid.strip()]
+    if not asset_ids:
        raise HTTPException(
            status_code=status.HTTP_400_BAD_REQUEST,
            detail="素材列表为空，无法创建片段",
        )

    # 2. 获取素材实际时长（去重查询）
-    unique_asset_ids = list(dict.fromkeys(body.asset_ids))
+    unique_asset_ids = list(dict.fromkeys(asset_ids))
    asset_durations: dict[str, float] = {}
+    asset_smart_scores: dict[str, float] = {}
+    # 素材 metadata 中缓存的场景切换点（由后台 MediaKit SceneChange 检测写入）：
+    # 有缓存时片段起点从随机镜头段中选取（不同片段来自不同镜头），无缓存回退随机起点
+    asset_scene_points: dict[str, list[float]] = {}
    for asset_id in unique_asset_ids:
        asset = asset_repo.get(asset_id)
        if asset and hasattr(asset, "duration"):
            asset_durations[asset_id] = float(asset.duration or 0.0)
+            # 计算 smart_match 综合评分，用于候选排序
+            smart_score, _ = score_asset(asset)
+            asset_smart_scores[asset_id] = smart_score
+            # 读取场景切换点缓存（新素材未检测过时为 None，走随机起点兜底）
+            cached_points = extract_scene_points_from_metadata(getattr(asset, "metadata", None))
+            if cached_points:
+                asset_scene_points[asset_id] = cached_points
+    logger.info(
+        "from-assets 场景缓存命中: %d/%d 个素材有场景切换点",
+        len(asset_scene_points),
+        len(unique_asset_ids),
+    )

    # 3. 在内存中计算所有片段数据（使用随机起始时间，不调用MediaKit）
    # 读取素材 metadata 中持久化的历史已用区间（跨任务/跨调用去重），
    # 格式与 _calc_random_start_time 的 used_segments 参数一致
-    used_segments: dict[str, list[tuple[float, float]]] = get_used_segments(
-        db, unique_asset_ids
-    )
+    used_segments: dict[str, list[tuple[float, float]]] = get_used_segments(db, unique_asset_ids)
    # 受控复用回调：可用区间耗尽时复用最久未用且未达复用上限（3次）的历史区间，
    # 复用片段时长累加到 reused_durations 供 15% 占比控制
    reused_durations: dict[str, float] = {}
@@ -645,19 +736,52 @@ def create_clips_from_assets_editor(
            return False
        return reused_durations.get(aid, 0.0) / assigned > REUSE_RATIO_LIMIT

-    for i, (_seg_order, dur_min, dur_max) in enumerate(segments):
+    # 素材耗尽标志：某轮循环中所有素材均被跳过时为 True
+    all_assets_exhausted = False
+
+    # 计算转场重叠补偿：每个 clip 需要额外增加的时长
+    # 目标：渲染后视频总时长 = 模板设定的各片段时长之和
+    # 公式：每 clip 增加 (n_segments - 1) * td / n_segments
+    n_segments = len(segments)
+    if n_segments > 1:
+        transition_compensation = (n_segments - 1) * _DEFAULT_TRANSITION_DURATION / n_segments
+    else:
+        transition_compensation = 0.0
+
+    # 打乱 segments 的处理顺序（分配素材的顺序随机化），但最终 clips_data 按原始 order 排序
+    shuffled_indices = list(range(len(segments)))
+    random.shuffle(shuffled_indices)
+
+    for idx in shuffled_indices:
+        _seg_order, dur_min, dur_max = segments[idx]
        # 在 segment 的 duration_min ~ duration_max 之间随机取值（保留一位小数）
        raw_duration = random.uniform(dur_min, dur_max)
+        # 加上转场补偿，确保最终输出时长 = 模板设定总时长
+        raw_duration += transition_compensation

-        # 轮询分配素材：跳过时长缺失、复用占比已超 15% 阈值的素材；
+        # 贪心分配素材：按"已使用次数"升序排列候选素材（使用最少的优先），
+        # 同次数随机打散，避免"A-B-C-D"的固定组合反复出现。
+        # 跳过时长缺失、复用占比已超 10% 阈值的素材；
        # 选中后计算起点，若该素材可用区间耗尽且复用被闸门拒绝（calc 返回 None），
-        # 继续轮询下一个素材
+        # 继续尝试下一个素材
        asset_id = ""
        clip_duration = 0.0
        start_time: float | None = None
-        n_assets = len(body.asset_ids)
-        for offset in range(n_assets):
-            candidate = body.asset_ids[(i + offset) % n_assets]
+        # 动态按使用次数排序：优先选使用最少的素材，同次数随机打散
+        asset_use_counts = {aid: len(used_segments.get(aid, [])) for aid in asset_ids}
+        # 排序键：smart_match 评分（注入随机噪声）→ 使用次数 → 纯随机。
+        # 噪声让得分接近的素材排名每次浮动，避免同一批素材反复选出相同组合，
+        # 从素材组合层面降低成片查重率；分差 > SCORE_RANDOM_NOISE_MAX 时排名稳定，
+        # 质量差距显著的素材仍保持优先级。
+        sorted_candidates = sorted(
+            asset_ids,
+            key=lambda aid: (
+                -(asset_smart_scores.get(aid, 0.0) + random.uniform(0.0, SCORE_RANDOM_NOISE_MAX)),
+                asset_use_counts.get(aid, 0),
+                random.random(),
+            ),
+        )
+        for candidate in sorted_candidates:
            candidate_total = asset_durations.get(candidate, 0.0)
            if candidate_total <= 0:
                continue
@@ -671,16 +795,30 @@ def create_clips_from_assets_editor(
                    candidate,
                )
                continue
-            # 随机起始时间（不调用 MediaKit，保证接口快速返回）；100 次避不开
-            # 历史区间时走受控复用回调（复用片段累加 reused_durations，回调内部
-            # 预判复用后占比超 15% 则拒绝并返回 None）
-            candidate_start = _calc_random_start_time(
-                candidate,
-                candidate_duration,
-                asset_durations,
-                used_segments,
-                on_exhausted=reuse_cb,
-            )
+            # 起始时间选取（不调用 MediaKit，保证接口快速返回）：
+            # 1) 素材有场景切换点缓存时，优先从随机镜头段中选起点（不同片段来自不同镜头，
+            #    画面内容本质不同），与 used_segments 做冲突避让（含 1.5s 边缘间隙）
+            # 2) 无缓存 / 镜头段全冲突 → _calc_random_start_time 随机起点兜底；
+            #    100 次避不开历史区间时走受控复用回调（复用片段累加 reused_durations，
+            #    回调内部预判复用后占比超 10% 则拒绝并返回 None）
+            candidate_start = None
+            if candidate in asset_scene_points:
+                candidate_start = pick_scene_aware_start(
+                    candidate,
+                    candidate_duration,
+                    asset_durations,
+                    asset_scene_points,
+                    used_segments,
+                    edge_gap=SEGMENT_EDGE_GAP,
+                )
+            if candidate_start is None:
+                candidate_start = _calc_random_start_time(
+                    candidate,
+                    candidate_duration,
+                    asset_durations,
+                    used_segments,
+                    on_exhausted=reuse_cb,
+                )
            if candidate_start is None:
                # 该素材可用区间耗尽且复用被闸门/use_count 上限拒绝 → 尝试下一素材
                logger.info(
@@ -695,28 +833,23 @@ def create_clips_from_assets_editor(

        if not asset_id or start_time is None:
            # 所有素材时长缺失、复用占比超阈值，或区间耗尽且复用被拒 → 素材可切区间不足
+            all_assets_exhausted = True
            raise HTTPException(
                status_code=status.HTTP_400_BAD_REQUEST,
                detail="素材可切区间不足，请补充新素材",
            )

        # 记录已使用时间段（内存，供本次后续片段避开）
-        used_segments.setdefault(asset_id, []).append(
-            (start_time, start_time + clip_duration)
-        )
-        asset_assigned_durations[asset_id] = (
-            asset_assigned_durations.get(asset_id, 0.0) + clip_duration
-        )
+        used_segments.setdefault(asset_id, []).append((start_time, start_time + clip_duration))
+        asset_assigned_durations[asset_id] = asset_assigned_durations.get(asset_id, 0.0) + clip_duration
        # 同步写入素材 metadata（不 commit，与下方 replace_all_clips_transactional
        # 处于同一事务，任一步失败整体回滚，不留脏数据）；
        # 复用区间与历史记录高度重叠时 record 内部自动累加 use_count
-        record_used_segments(
-            db, asset_id, start_time, start_time + clip_duration, plan_id
-        )
+        record_used_segments(db, asset_id, start_time, start_time + clip_duration, plan_id)

        clips_data.append(
            {
-                "order": i,
+                "order": _seg_order,
                "asset_id": asset_id,
                "start_time": start_time,
                "duration": clip_duration,
@@ -724,6 +857,9 @@ def create_clips_from_assets_editor(
            }
        )

+    # 按原始 segment order 排序，确保 clips_data 的 order 字段有序（0,1,2,3...）
+    clips_data.sort(key=lambda c: c["order"])
+
    # 4. 事务性替换：清空旧片段 → 创建新片段 → 标记ready（单事务，失败自动回滚）
    created_count = plan_svc.replace_all_clips_transactional(plan_id, clips_data)

@@ -743,11 +879,31 @@ def create_clips_from_assets_editor(
        unique_asset_ids,
    )

-    # 6. 立即返回响应
+    # 6. 估算成片内部重复率（本条成片中同一素材的重叠片段时长占比）
+    dup_rate = _calc_plan_internal_duplicate_rate(clips_data)
+    duplicate_warning = None
+    if dup_rate > 50:
+        duplicate_warning = f"查重率 {dup_rate:.1f}% 超过50%，建议更换素材或模板"
+        logger.warning(
+            "from-assets 成片查重率超标: plan_id=%s dup_rate=%.1f%%",
+            plan_id,
+            dup_rate,
+        )
+
+    # 7. 素材耗尽提示
+    exhaustion_warning = None
+    if all_assets_exhausted and created_count < len(segments):
+        exhaustion_warning = (
+            "素材可切区间不足，部分片段使用了复用素材。" "建议：1) 补充更多素材到素材库 2) 使用不同的素材组合生成"
+        )
+
+    # 8. 立即返回响应
    return ClipsFromAssetsResponse(
        created_count=created_count,
        plan_id=plan_id,
        clip_ids=[],
+        duplicate_warning=duplicate_warning,
+        exhaustion_warning=exhaustion_warning,
    )


@@ -755,7 +911,15 @@ def _update_mediakit_recommendations_async(  # pragma: no cover
    plan_id: str,
    asset_ids: list[str],
 ) -> None:
-    """后台任务：调用 MediaKit 智能选片并更新片段的起始时间.
+    """后台任务：使用 SceneChange 智能选帧并更新片段的起始时间.
+
+    优先使用 SceneChange 策略检测视频镜头切换点，将每个素材按镜头段拆分，
+    各片段优先从不同镜头段中选取起始时间，实现「不同片段展示不同场景」的效果。
+
+    降级策略：
+    1. SceneChange 优先 → detect_scene_changes 内部已含 TimeInterval 降级
+    2. 若 detect_scene_changes 仍返回 None → 回退到旧的 analyze_videos 方式
+    3. 所有方式都失败 → 保持现有随机 start_time，不影响视频生成

    此函数在后台异步执行，不影响接口响应时间。
    失败时静默处理，不影响已创建的片段。
@@ -777,12 +941,6 @@ def _update_mediakit_recommendations_async(  # pragma: no cover
        asset_repo = SQLAlchemyAssetRepository(db)
        plan_svc = EditPlanService(db)

-        # 调用 MediaKit 获取推荐时间
-        recommendations = _get_mediakit_recommendations(asset_ids, asset_repo)
-        if not recommendations:
-            logger.info("后台任务: MediaKit 无推荐结果，跳过更新")
-            return
-
        # 查询该 plan 的所有片段（分批获取，避免硬编码 limit 截断）
        batch_size = 500
        all_clips = []
@@ -803,21 +961,18 @@ def _update_mediakit_recommendations_async(  # pragma: no cover

        # 批量预加载所有涉及的素材（消除 N+1 查询）
        unique_asset_ids = list({getattr(c, "asset_id", "") or "" for c in clips} - {""})
-        assets_map: dict[str, object] = {
-            a.id: a for a in asset_repo.find_by_ids(unique_asset_ids)
-        }
+        assets_map: dict[str, object] = {a.id: a for a in asset_repo.find_by_ids(unique_asset_ids)}

-        # 按 asset_id 预分组片段时间段（消除 O(N^2) 嵌套循环）
-        clips_by_asset: dict[str, list[tuple[str, float, float]]] = defaultdict(list)
+        # 按 asset_id 预分组片段对象（按 order 排序，保证按模板顺序分配镜头段）
+        clips_by_asset: dict[str, list] = defaultdict(list)
        for clip in clips:
            aid = getattr(clip, "asset_id", "") or ""
-            if aid and clip.start_time is not None:
-                clips_by_asset[aid].append(
-                    (clip.id, clip.start_time, clip.start_time + clip.duration)
-                )
+            if aid:
+                clips_by_asset[aid].append(clip)
+        for aid in clips_by_asset:
+            clips_by_asset[aid].sort(key=lambda c: c.order)

-        # 读取素材全部历史已用区间（跨任务/跨 plan 持久化记录）：
-        # MediaKit 挪点必须与随机选片一样避让历史区间，否则会把片段挪回已用过的画面
+        # 读取素材全部历史已用区间（跨任务/跨 plan 持久化记录）
        historical_segments = get_used_segments(db, unique_asset_ids)

        # 已更新的片段ID（用于排除已移动的旧时间段）
@@ -826,16 +981,22 @@ def _update_mediakit_recommendations_async(  # pragma: no cover
        updated_segments: dict[str, list[tuple[float, float]]] = {}
        updated_count = 0

-        # 遍历片段，按 asset_id 匹配推荐时间
-        for clip in clips:
-            asset_id = getattr(clip, "asset_id", "") or ""
-            if not asset_id or asset_id not in recommendations:
+        # 尝试获取存储服务（用于生成视频 URL）
+        try:
+            storage = get_storage_service()
+        except Exception:
+            logger.warning("后台任务: 获取存储服务失败，跳过 SceneChange 更新")
+            return
+
+        # 获取 MediaKit 客户端
+        client = get_mediakit_client()
+
+        # 对每个素材，检测场景切换点并分配镜头段
+        for asset_id in unique_asset_ids:
+            asset_clips = clips_by_asset.get(asset_id, [])
+            if not asset_clips:
                continue

-            recommended_start = recommendations[asset_id]
-            clip_duration = clip.duration
-
-            # 从预加载字典获取素材（O(1) 查找）
            asset = assets_map.get(asset_id)
            if not asset:
                continue
@@ -843,96 +1004,177 @@ def _update_mediakit_recommendations_async(  # pragma: no cover
            if asset_total <= 0:
                continue

-            # 推荐时间 + 片段时长不能超过素材总时长
-            if recommended_start + clip_duration > asset_total:
-                logger.info(
-                    "后台任务: 推荐时间越界，跳过: asset_id=%s recommended=%.2f duration=%.1f total=%.1f",
-                    asset_id,
-                    recommended_start,
-                    clip_duration,
-                    asset_total,
-                )
-                continue
-
-            # 构建排除当前片段及已更新片段后的占用列表（O(M)，M=同素材片段数）
-            other_segments: list[tuple[float, float]] = [
-                (cs, ce)
-                for cid, cs, ce in clips_by_asset.get(asset_id, [])
-                if cid != clip.id and cid not in updated_clip_ids
-            ]
-            other_segments.extend(updated_segments.get(asset_id, []))
-            # 并入该素材全部历史已用区间（含其他 plan/其他任务），set 去重：
-            # 本 plan 片段创建时已写入历史记录
-            # 并入该素材全部历史已用区间（含其他 plan/其他任务）。
-            # set 去重前先归一化精度（round 3 位），避免浮点尾差导致逻辑相同的
-            # 区间（如 1.0 与 1.0000000001）被误判为不同区间
-            def _norm(segs):
-                return {(round(float(a), 3), round(float(b), 3)) for a, b in segs}
-
-            other_segments = list(
-                _norm(other_segments) | _norm(historical_segments.get(asset_id, []))
-            )
-
-            # 检查推荐时间是否与同 plan 片段或历史已用区间冲突（含 0.3s 边缘间隙）：
-            # 冲突时放弃该推荐、保留原随机起点（不硬挪到已用过的画面）
-            if _recommended_time_conflicts(recommended_start, clip_duration, other_segments):
-                logger.info(
-                    "后台任务: 推荐时间与同片/历史区间冲突，保留原起点: asset_id=%s recommended=%.2f",
-                    asset_id,
-                    recommended_start,
-                )
-                continue
-
-            # 逐个更新并捕获异常（单点失败不影响其他片段）
-            try:
-                old_start = clip.start_time
-                old_end = old_start + clip_duration
-                # MediaKit 移动片段起点 + 同步素材 metadata 区间记录放在同一事务：
-                # 删旧区间记录（按 plan_id + 旧 start 匹配，兼容无 plan_id 的旧数据）、
-                # 写新区间，最后统一 commit；任一步失败整体 rollback，
-                # 保证 clip.start_time 与 metadata.used_time_ranges 不出现不一致。
-                plan_svc.update_clip(clip.id, start_time=recommended_start)
+            # 获取素材视频 URL
+            video_url: str | None = None
+            storage_key = getattr(asset, "storage_key", None) or ""
+            mime = getattr(asset, "mime_type", "") or ""
+            if storage_key and mime.startswith("video/"):
                try:
-                    if remove_used_segment(
-                        db, asset_id, old_start, old_end, plan_id=plan_id
-                    ):
-                        record_used_segments(
-                            db,
-                            asset_id,
-                            recommended_start,
-                            recommended_start + clip_duration,
-                            plan_id,
-                        )
-                except Exception as me:
-                    logger.warning(
-                        "后台任务: 同步素材区间记录失败，回滚本次片段更新: clip_id=%s error=%s",
-                        clip.id,
-                        me,
+                    video_url = storage.get_download_url(storage_key)
+                except Exception as e:
+                    logger.warning("后台任务: 获取素材URL失败: asset_id=%s error=%s", asset_id, e)
+
+            # 构建该素材的占用区间列表（排除已更新片段）
+            def _get_other_segments(asset_id_inner, clip_id_inner):
+                segs: list[tuple[float, float]] = []
+                for c in clips_by_asset.get(asset_id_inner, []):
+                    cid = c.id
+                    if cid != clip_id_inner and cid not in updated_clip_ids:
+                        segs.append((c.start_time, c.start_time + c.duration))
+                segs.extend(updated_segments.get(asset_id_inner, []))
+
+                # 并入历史已用区间
+                def _norm(segs_in):
+                    return {(round(float(a), 3), round(float(b), 3)) for a, b in segs_in}
+
+                return list(_norm(segs) | _norm(historical_segments.get(asset_id_inner, [])))
+
+            # 优先使用 SceneChange 策略
+            scene_segments: list[tuple[float, float]] = []
+            # 先查素材 metadata 中的场景点缓存：命中则直接复用，跳过 MediaKit 检测
+            # （缓存由本任务首次检测后写入，跨任务/跨 plan 复用）
+            cached_points = extract_scene_points_from_metadata(getattr(asset, "metadata", None))
+            if cached_points:
+                scene_segments = build_scene_segments(cached_points, asset_total)
+                logger.info(
+                    "后台任务: 命中场景点缓存: asset_id=%s scenes=%d",
+                    asset_id,
+                    len(scene_segments),
+                )
+
+            if not scene_segments and client.is_available and video_url:
+                scene_changes = client.detect_scene_changes(video_url)
+                if scene_changes is not None:
+                    scene_segments = build_scene_segments(scene_changes, asset_total)
+                    logger.info(
+                        "后台任务: 素材场景检测完成: asset_id=%s scenes=%d",
+                        asset_id,
+                        len(scene_segments),
                    )
-                    db.rollback()
-                    continue
-                db.commit()
-                updated_count += 1
-                updated_clip_ids.add(clip.id)
-            except Exception as ue:
-                logger.warning(
-                    "后台任务: 单个片段更新失败: clip_id=%s error=%s", clip.id, ue
+                    # 检测结果写入素材 metadata 缓存：首次生成用随机起点，
+                    # 检测完成后后续生成的渲染前同步路径即可读缓存选镜头段
+                    try:
+                        existing_meta = dict(getattr(asset, "metadata", None) or {})
+                        existing_meta["scene_change_points"] = scene_changes
+                        asset.metadata = existing_meta  # type: ignore[attr-defined]
+                        asset_repo.update(asset)  # type: ignore[arg-type]
+                        logger.info(
+                            "后台任务: 场景点已写入素材缓存: asset_id=%s points=%d",
+                            asset_id,
+                            len(scene_changes),
+                        )
+                    except Exception as cache_err:
+                        # 缓存写入失败不影响本次片段更新
+                        logger.warning(
+                            "后台任务: 场景点缓存写入失败: asset_id=%s error=%s",
+                            asset_id,
+                            cache_err,
+                        )
+
+            # SceneChange 未获得有效结果 → 尝试 analyze_videos 作为 fallback
+            if not scene_segments and video_url:
+                fallback_recs = _get_mediakit_recommendations([asset_id], asset_repo)
+                if fallback_recs and asset_id in fallback_recs:
+                    # analyze_videos 只返回单个推荐点，转为单镜头段
+                    rec_start = fallback_recs[asset_id]
+                    scene_segments = [(rec_start, asset_total)]
+                    logger.info(
+                        "后台任务: 使用 analyze_videos fallback: asset_id=%s start=%.2f",
+                        asset_id,
+                        rec_start,
+                    )
+
+            if not scene_segments:
+                # 所有方式都失败 → 保持现有随机 start_time
+                logger.info(
+                    "后台任务: SceneChange 与 analyze_videos 均无结果，保持随机起点: asset_id=%s",
+                    asset_id,
                )
-                try:
-                    db.rollback()
-                except Exception:
-                    pass
                continue

-            updated_segments.setdefault(asset_id, []).append(
-                (recommended_start, recommended_start + clip_duration)
-            )
-            logger.info(
-                "后台任务: 更新片段起始时间: clip_id=%s asset_id=%s start_time=%.2f",
-                clip.id,
-                asset_id,
-                recommended_start,
-            )
+            # 为每个片段分配不同的镜头段
+            scene_segments_pool = list(scene_segments)  # 可消费的镜头段池
+            for clip in asset_clips:
+                clip_duration = clip.duration
+                recommended_start: float | None = None
+
+                # 从镜头段池中依次尝试，选一个不冲突的
+                for seg_idx, (seg_start, seg_end) in enumerate(scene_segments_pool):
+                    candidate_start = pick_start_in_scene_segment(seg_start, seg_end, clip_duration)
+                    if candidate_start is None:
+                        continue  # 镜头段太短，跳过
+
+                    # 检查越界
+                    if candidate_start + clip_duration > asset_total:
+                        continue
+
+                    # 检查与已用区间冲突
+                    other_segs = _get_other_segments(asset_id, clip.id)
+                    if _recommended_time_conflicts(candidate_start, clip_duration, other_segs):
+                        continue
+
+                    recommended_start = candidate_start
+                    # 消费该镜头段（从池中移除，下一个片段用不同镜头段）
+                    scene_segments_pool.pop(seg_idx)
+                    break
+
+                if recommended_start is None:
+                    # 镜头段用完或都冲突 → 尝试 _calc_random_start_time 兜底
+                    used_segs_for_calc: dict[str, list[tuple[float, float]]] = {
+                        asset_id: _get_other_segments(asset_id, clip.id)
+                    }
+                    fallback_start = _calc_random_start_time(
+                        asset_id,
+                        clip_duration,
+                        {asset_id: asset_total},
+                        used_segs_for_calc,
+                    )
+                    if fallback_start is None:
+                        continue  # 完全无法分配，保持原起点
+                    recommended_start = fallback_start
+
+                # 更新片段起始时间
+                try:
+                    old_start = clip.start_time
+                    old_end = old_start + clip_duration
+
+                    plan_svc.update_clip(clip.id, start_time=recommended_start)
+                    try:
+                        if remove_used_segment(db, asset_id, old_start, old_end, plan_id=plan_id):
+                            record_used_segments(
+                                db,
+                                asset_id,
+                                recommended_start,
+                                recommended_start + clip_duration,
+                                plan_id,
+                            )
+                    except Exception as me:
+                        logger.warning(
+                            "后台任务: 同步素材区间记录失败，回滚本次片段更新: clip_id=%s error=%s",
+                            clip.id,
+                            me,
+                        )
+                        db.rollback()
+                        continue
+                    db.commit()
+                    updated_count += 1
+                    updated_clip_ids.add(clip.id)
+                    updated_segments.setdefault(asset_id, []).append(
+                        (recommended_start, recommended_start + clip_duration)
+                    )
+                    logger.info(
+                        "后台任务: 更新片段起始时间(场景选帧): clip_id=%s asset_id=%s start_time=%.2f",
+                        clip.id,
+                        asset_id,
+                        recommended_start,
+                    )
+                except Exception as ue:
+                    logger.warning("后台任务: 单个片段更新失败: clip_id=%s error=%s", clip.id, ue)
+                    try:
+                        db.rollback()
+                    except Exception:
+                        pass
+                    continue

        logger.info("后台任务完成: plan_id=%s 成功更新 %d 个片段", plan_id, updated_count)

@@ -950,4 +1192,3 @@ def _update_mediakit_recommendations_async(  # pragma: no cover
                db.close()
            except Exception:
                pass
-
@@ -41,17 +41,17 @@ def list_editor_transition_presets(
    _: AuthenticatedUser = Depends(get_current_user),
 ) -> TransitionPresetListResponse:
    """获取转场预设列表"""
-    from packages.domain.transition_presets import TRANSITION_PRESETS
+    from packages.domain.transition_presets import TRANSITION_PRESET_LIBRARY

    items = [
        {
-            "id": p["id"],
-            "name": p["name"],
-            "category": p.get("category", "通用"),
-            "duration": p.get("default_duration", 0.5),
-            "description": p.get("description", ""),
+            "id": p.id,
+            "name": p.name,
+            "category": p.category,
+            "duration": p.default_duration,
+            "description": p.description,
        }
-        for p in TRANSITION_PRESETS
+        for p in TRANSITION_PRESET_LIBRARY
    ]
    return TransitionPresetListResponse(items=items, total=len(items))

@@ -123,17 +123,17 @@ def list_editor_filter_presets(
    _: AuthenticatedUser = Depends(get_current_user),
 ) -> FilterPresetListResponse:
    """获取滤镜预设列表"""
-    from packages.domain.filter_presets import FILTER_PRESETS
+    from packages.domain.filter_presets import FILTER_PRESET_LIBRARY

    items = [
        {
-            "id": p["id"],
-            "name": p["name"],
-            "category": p.get("category", "通用"),
-            "thumbnail": p.get("thumbnail", ""),
-            "description": p.get("description", ""),
+            "id": p.id,
+            "name": p.name,
+            "category": p.category,
+            "thumbnail": p.lut_url,
+            "description": p.description,
        }
-        for p in FILTER_PRESETS
+        for p in FILTER_PRESET_LIBRARY
    ]
    return FilterPresetListResponse(items=items, total=len(items))

@@ -167,7 +167,18 @@ class ClipsFromAssetsRequest(BaseModel):

    asset_ids: List[str] = Field(..., min_length=1, max_length=200, description="素材 ID 列表，按顺序追加到时间线末尾")
    clip_type: str = Field(default="main", description="片段类型，默认 main")
-    required_clips_count: Optional[int] = Field(default=None, ge=1, le=200, description="要求创建的片段数量；不传则等于素材数量")
+    required_clips_count: Optional[int] = Field(
+        default=None, ge=1, le=200, description="要求创建的片段数量；不传则等于素材数量"
+    )
+
+    @validator("asset_ids", pre=True)
+    def _drop_invalid_asset_ids(cls, v):  # noqa: N805
+        """容错过滤：前端异常情况下可能把 undefined 序列化成 null 或空串混入
+        asset_ids（会直接 422 或导致后续 /assets/{id} 404），这里统一剔除。
+        过滤后为空时由 Field(min_length=1) / 路由层 400 兜底。"""
+        if not isinstance(v, list):
+            return v
+        return [x for x in v if isinstance(x, str) and x.strip()]


 class ClipsFromAssetsResponse(BaseModel):
@@ -178,6 +189,8 @@ class ClipsFromAssetsResponse(BaseModel):
    plan_id: str = ""
    message: str = ""
    clip_ids: List[str] = Field(default_factory=list, description="创建的片段ID列表")
+    duplicate_warning: Optional[str] = Field(default=None, description="查重率超标警告")
+    exhaustion_warning: Optional[str] = Field(default=None, description="素材耗尽警告")


 # ── 封面配置 ────────────────────────────────────────────────────────────────
@@ -2,7 +2,9 @@

 from __future__ import annotations

+import json
 import logging
+import subprocess
 import tempfile
 from pathlib import Path
 from typing import Any, Optional
@@ -430,6 +432,8 @@ def save_tts_job_to_library(
    storage_key = f"uploads/voice/tts/{job.id}.{audio_format}"

    tmp_path: Path | None = None
+    audio_duration: float | None = None
+    file_size = 0
    try:
        with tempfile.NamedTemporaryFile(suffix=f".{audio_format}", delete=False) as tmp:
            tmp_path = Path(tmp.name)
@@ -445,6 +449,23 @@ def save_tts_job_to_library(
            )
        file_size = tmp_path.stat().st_size
        storage_service.upload_file(tmp_path, storage_key, content_type=content_type)
+
+        # 从音频文件提取时长（ffprobe），作为 job.duration 的兜底
+        try:
+            proc = subprocess.run(
+                [
+                    "ffprobe", "-v", "quiet", "-print_format", "json",
+                    "-show_format", str(tmp_path),
+                ],
+                capture_output=True, text=True, timeout=10,
+            )
+            if proc.returncode == 0:
+                fmt = json.loads(proc.stdout).get("format", {})
+                dur = float(fmt.get("duration", 0))
+                if dur > 0:
+                    audio_duration = dur
+        except Exception:
+            logger.warning("ffprobe 提取时长失败: job_id=%s", job.id, exc_info=True)
    except HTTPException:
        raise
    except Exception as e:
@@ -482,7 +503,7 @@ def save_tts_job_to_library(
        mime_type=content_type,
        metadata=metadata_,
        file_size=file_size,
-        duration=job.duration or None,
+        duration=job.duration or audio_duration or None,
        status=AssetStatus.READY,
        classification_status=ClassificationStatus.PENDING,  # 音频不参与内容分类，保持 pending 与 ingest 链路一致
        uploaded_by_user_id=user_id,
@@ -23,6 +23,7 @@ from app.schemas.upload import (
 from fastapi import APIRouter, Depends, File, Form, HTTPException, UploadFile, status

 from packages.application import SubmitIngestJobCommand, SubmitIngestJobUseCase
+from packages.domain import Asset, AssetStatus

 logger = logging.getLogger(__name__)

@@ -80,6 +81,40 @@ def _validate_mime_type(content_type: str | None) -> str:
    return base_type


+def _infer_mime_type_from_storage_key(storage_key: str) -> str:
+    """从 storage_key 推断 MIME 类型（与 worker 端保持一致）。"""
+    lower_filename = storage_key.rsplit("/", 1)[-1].lower()
+    _MIME_MAP = {
+        ".mov": "video/quicktime", ".mp4": "video/mp4", ".avi": "video/x-msvideo",
+        ".mkv": "video/x-matroska", ".webm": "video/webm",
+        ".png": "image/png", ".gif": "image/gif", ".bmp": "image/bmp",
+        ".svg": "image/svg+xml", ".jpg": "image/jpeg", ".jpeg": "image/jpeg",
+        ".mp3": "audio/mpeg", ".wav": "audio/wav", ".ogg": "audio/ogg",
+        ".flac": "audio/flac", ".m4a": "audio/x-m4a",
+    }
+    for ext, mime in _MIME_MAP.items():
+        if lower_filename.endswith(ext):
+            return mime
+    return "video/mp4"  # default
+
+
+def _create_pending_asset(
+    asset_repository, project_id, library_id, storage_key, filename, mime_type, user_id, file_hash=""
+):
+    """立即创建一条 PROCESSING 状态的 Asset 记录，使前端能马上看到新素材。"""
+    asset = Asset.create(
+        project_id=project_id,
+        library_id=library_id,
+        name=filename,
+        storage_key=storage_key,
+        mime_type=mime_type,
+        status=AssetStatus.PROCESSING,
+        uploaded_by_user_id=user_id,
+        file_hash=file_hash,
+    )
+    return asset_repository.create(asset)
+
+
 def _submit_ingest_job(
    project_id: str,
    library_id: str,
@@ -209,6 +244,20 @@ async def complete_direct_upload(
                url=storage_service.get_url(normalized_key),
            )

+    # 立即创建 Asset 记录（PROCESSING 状态），使前端刷新后即可看到新素材
+    filename = normalized_key.rsplit("/", 1)[-1]
+    mime_type = _infer_mime_type_from_storage_key(normalized_key)
+    pending_asset = _create_pending_asset(
+        asset_repository=asset_repository,
+        project_id=request.project_id,
+        library_id=request.library_id,
+        storage_key=normalized_key,
+        filename=filename,
+        mime_type=mime_type,
+        user_id=authenticated_user.user.id,
+        file_hash=request.file_hash,
+    )
+
    job = _submit_ingest_job(
        project_id=request.project_id,
        library_id=request.library_id,
@@ -216,7 +265,12 @@ async def complete_direct_upload(
        ingest_job_repository=ingest_job_repository,
        file_hash=request.file_hash,
    )
-    return DirectUploadCompleteResponse(storage_key=normalized_key, ingest_job_id=job.id, url=storage_service.get_url(normalized_key))
+    return DirectUploadCompleteResponse(
+        storage_key=normalized_key,
+        ingest_job_id=job.id,
+        asset_id=pending_asset.id,
+        url=storage_service.get_url(normalized_key),
+    )


@router.post(
@@ -284,6 +338,18 @@ async def upload_asset(
            detail=f"Failed to upload file: {type(error).__name__}",
        ) from error

+    # 立即创建 Asset 记录（PROCESSING 状态），使前端刷新后即可看到新素材
+    pending_asset = _create_pending_asset(
+        asset_repository=asset_repository,
+        project_id=project_id,
+        library_id=library_id,
+        storage_key=storage_key,
+        filename=safe_filename,
+        mime_type=validated_content_type,
+        user_id=authenticated_user.user.id,
+        file_hash=file_hash,
+    )
+
    job = _submit_ingest_job(
        project_id=project_id,
        library_id=library_id,
@@ -295,5 +361,6 @@ async def upload_asset(
    return UploadAssetResponse(
        storage_key=storage_key,
        ingest_job_id=job.id,
+        asset_id=pending_asset.id,
        url=file_url,
    )
@@ -52,6 +52,9 @@ def _to_video_response(item, storage: OSSStorageService | None = None) -> VideoI
        generation_params=item.generation_params,
        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),
    )


@@ -6,12 +6,26 @@
 from __future__ import annotations

 import logging
+import shutil
+import subprocess
+import tempfile
 import time
+from pathlib import Path
 from typing import Literal, Optional
+from uuid import uuid4

 from app.api.routes._helpers import get_user_plan
 from app.auth import AuthenticatedUser, get_current_user
-from app.dependencies import get_audio_url_signer, get_cosyvoice_service, get_db_session, get_user_repository
+from app.core.storage import get_storage_service
+from app.dependencies import (
+    get_asset_library_repository,
+    get_asset_repository,
+    get_audio_url_signer,
+    get_cosyvoice_service,
+    get_db_session,
+    get_project_repository,
+    get_user_repository,
+)
 from app.schemas.voice import (
    PresetVoiceItemResponse,
    PresetVoiceListResponse,
@@ -24,7 +38,7 @@ from app.schemas.voice_library import (
    UpdateVoiceLibraryRequest,
    VoiceLibraryItemResponse,
 )
-from fastapi import APIRouter, Depends, HTTPException, Query, Response, status
+from fastapi import APIRouter, Depends, File, Form, HTTPException, Query, Response, UploadFile, status
 from sqlalchemy.orm import Session

 from packages.adapters.sqlalchemy_impl.voice_clone_profile_repository import SQLAlchemyVoiceCloneProfileRepository
@@ -40,8 +54,12 @@ from packages.application.voice_library.use_cases import (
    QuotaExceededError,
    UpdateVoiceLibraryUseCase,
 )
+from packages.domain import Asset, AssetStatus
+from packages.domain.classification import AssetLibraryKind, ClassificationStatus
+from packages.domain.entities import AssetLibrary
 from packages.domain.preset_voices import PRESET_VOICES, get_preset_voice_by_id
 from packages.ports.user_repository import UserRepository
+from packages.shared.storage import SharedStorageService

 router = APIRouter()
 logger = logging.getLogger(__name__)
@@ -507,3 +525,243 @@ def delete_voice(
    if not deleted:
        raise HTTPException(status_code=status.HTTP_404_NOT_FOUND, detail="Voice not found")
    return
+
+
+# ── 提取视频配音 ─────────────────────────────────────────────────────
+
+# 支持的视频格式
+EXTRACT_VIDEO_MIMES = frozenset({"video/mp4", "video/quicktime", "video/webm", "video/x-msvideo"})
+MAX_EXTRACT_SIZE = 500 * 1024 * 1024  # 500MB
+
+
+@router.post(
+    "/extract-voice",
+    status_code=status.HTTP_201_CREATED,
+)
+def extract_voice_from_video(
+    file: UploadFile = File(...),
+    project_id: str = Form(...),
+    authenticated_user: AuthenticatedUser = Depends(get_current_user),
+    project_repository=Depends(get_project_repository),
+    asset_library_repository=Depends(get_asset_library_repository),
+    asset_repository=Depends(get_asset_repository),
+    storage_service: SharedStorageService = Depends(get_storage_service),
+    sign_url=Depends(get_audio_url_signer),
+):
+    """从上传的视频中提取人声配音。
+
+    流程：
+    1. 接收视频文件（mp4/mov/webm）
+    2. ffmpeg 提取音频 + 降噪 + 编码为 mp3
+    3. 上传到 OSS，创建 Asset 记录到配音素材库
+    4. 返回素材信息（时长、文件大小、URL）
+    """
+    user_id = authenticated_user.user.id
+
+    # 校验文件类型
+    content_type = file.content_type or ""
+    if content_type and content_type not in EXTRACT_VIDEO_MIMES:
+        # 兜底：按扩展名判断
+        ext = (file.filename or "").rsplit(".", 1)[-1].lower()
+        ext_to_mime = {"mp4": "video/mp4", "mov": "video/quicktime", "webm": "video/webm", "avi": "video/x-msvideo"}
+        if ext not in ext_to_mime:
+            raise HTTPException(
+                status_code=status.HTTP_400_BAD_REQUEST,
+                detail="仅支持 mp4/mov/webm/avi 格式的视频文件",
+            )
+        content_type = ext_to_mime[ext]
+
+    # 找到（或自动创建）用户 voice 素材库（复用 TTS 的逻辑）
+    library = _find_or_create_voice_library_for_extract(
+        user_id=user_id,
+        project_repository=project_repository,
+        asset_library_repository=asset_library_repository,
+    )
+
+    tmp_dir = None
+    try:
+        tmp_dir = Path(tempfile.mkdtemp(prefix="voice_extract_"))
+        video_path = tmp_dir / f"input_{uuid4().hex[:8]}_{file.filename or 'video.mp4'}"
+        audio_path = tmp_dir / f"output_{uuid4().hex[:8]}.mp3"
+
+        # 保存上传的视频到临时文件
+        with open(video_path, "wb") as f:
+            total = 0
+            while chunk := file.file.read(1024 * 1024):  # 1MB chunks
+                total += len(chunk)
+                if total > MAX_EXTRACT_SIZE:
+                    raise HTTPException(
+                        status_code=status.HTTP_413_REQUEST_ENTITY_TOO_LARGE,
+                        detail="视频文件过大，最大支持 500MB",
+                    )
+                f.write(chunk)
+
+        if video_path.stat().st_size == 0:
+            raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="视频文件为空")
+
+        # ffmpeg: 提取音频 + 降噪 + 编码 mp3
+        # 滤镜链：highpass(去低频噪声) → afftdn(FFT降噪) → lowpass(去高频噪声)
+        ffmpeg_cmd = [
+            "ffmpeg",
+            "-y",
+            "-i",
+            str(video_path),
+            "-vn",  # 不要视频
+            "-af",
+            "highpass=f=80,afftdn=nf=-25:tn=1,lowpass=f=8000",
+            "-acodec",
+            "libmp3lame",
+            "-ab",
+            "192k",
+            "-ar",
+            "44100",
+            "-ac",
+            "1",  # 单声道（人声足够）
+            str(audio_path),
+        ]
+
+        result = subprocess.run(
+            ffmpeg_cmd,
+            capture_output=True,
+            timeout=300,  # 5 分钟超时
+        )
+
+        if result.returncode != 0:
+            stderr_text = result.stderr.decode("utf-8", errors="replace")[-500:]
+            logger.error("ffmpeg 提取配音失败: %s", stderr_text)
+            raise HTTPException(
+                status_code=status.HTTP_422_UNPROCESSABLE_ENTITY,
+                detail="视频音频提取失败，可能该视频没有音轨或格式不支持",
+            )
+
+        if not audio_path.exists() or audio_path.stat().st_size == 0:
+            raise HTTPException(
+                status_code=status.HTTP_422_UNPROCESSABLE_ENTITY,
+                detail="音频提取结果为空",
+            )
+
+        # 获取音频时长
+        duration = _get_audio_duration(audio_path)
+        file_size = audio_path.stat().st_size
+
+        # 上传到 OSS
+        audio_ext = "mp3"
+        storage_key = f"uploads/voice/extracted/{uuid4().hex}.{audio_ext}"
+        storage_service.upload_file(audio_path, storage_key, content_type="audio/mpeg")
+
+        # 创建 Asset 记录
+        original_name = (file.filename or "video").rsplit(".", 1)[0]
+        asset_name = f"{original_name}-配音"
+
+        asset = Asset.create(
+            project_id=library.project_id,
+            library_id=library.id,
+            name=asset_name,
+            storage_key=storage_key,
+            mime_type="audio/mpeg",
+            metadata={
+                "source": "video_extract",
+                "original_video": file.filename or "unknown",
+            },
+            file_size=file_size,
+            duration=duration,
+            status=AssetStatus.READY,
+            classification_status=ClassificationStatus.PENDING,
+            uploaded_by_user_id=user_id,
+        )
+        asset = asset_repository.create(asset)
+
+        return {
+            "id": asset.id,
+            "name": asset.name,
+            "audio_url": sign_url(storage_key),
+            "duration": duration,
+            "file_size": file_size,
+            "status": "completed",
+            "source": "video_extract",
+        }
+
+    except HTTPException:
+        raise
+    except subprocess.TimeoutExpired:
+        raise HTTPException(
+            status_code=status.HTTP_504_GATEWAY_TIMEOUT,
+            detail="视频处理超时，请尝试较短的视频",
+        ) from None
+    except Exception as e:
+        logger.exception("提取视频配音失败: %s", e)
+        raise HTTPException(
+            status_code=status.HTTP_500_INTERNAL_SERVER_ERROR,
+            detail="提取配音失败，请稍后重试",
+        ) from e
+    finally:
+        # 清理临时文件
+        if tmp_dir and Path(tmp_dir).exists():
+            shutil.rmtree(tmp_dir, ignore_errors=True)
+
+
+def _find_or_create_voice_library_for_extract(*, user_id, project_repository, asset_library_repository):
+    """为用户找到或创建 voice 素材库（与 TTS 保存逻辑一致）。"""
+    projects = project_repository.find_accessible_projects(user_id)
+    if not projects:
+        raise HTTPException(
+            status_code=status.HTTP_400_BAD_REQUEST,
+            detail="没有可用的项目，请先创建项目",
+        )
+
+    for project in projects:
+        for lib in asset_library_repository.find_by_project(project.id):
+            kind = lib.kind.value if hasattr(lib.kind, "value") else lib.kind
+            if kind == AssetLibraryKind.VOICE.value:
+                return lib
+
+    # 自动创建
+    from sqlalchemy.exc import IntegrityError
+
+    project = projects[0]
+    library = AssetLibrary.create(
+        project_id=project.id,
+        name="配音素材库",
+        kind=AssetLibraryKind.VOICE,
+    )
+    try:
+        return asset_library_repository.create(library)
+    except IntegrityError:
+        session = getattr(asset_library_repository, "session", None)
+        if session is not None:
+            try:
+                session.rollback()
+            except Exception:
+                pass
+        for lib in asset_library_repository.find_by_project(project.id):
+            kind = lib.kind.value if hasattr(lib.kind, "value") else lib.kind
+            if kind == AssetLibraryKind.VOICE.value:
+                return lib
+        raise HTTPException(
+            status_code=status.HTTP_500_INTERNAL_SERVER_ERROR,
+            detail="配音素材库创建失败",
+        ) from None
+
+
+def _get_audio_duration(audio_path: Path) -> float:
+    """用 ffprobe 获取音频时长（秒）。"""
+    try:
+        result = subprocess.run(
+            [
+                "ffprobe",
+                "-v",
+                "quiet",
+                "-show_entries",
+                "format=duration",
+                "-of",
+                "csv=p=0",
+                str(audio_path),
+            ],
+            capture_output=True,
+            timeout=10,
+        )
+        if result.returncode == 0 and result.stdout.strip():
+            return float(result.stdout.strip())
+    except (ValueError, subprocess.TimeoutExpired):
+        pass
+    return 0.0
@@ -129,10 +129,13 @@ class SmartMatchRequest(BaseModel):
    )


-class SmartMatchItem(BaseModel):
-    """智能选素材结果条目。"""
+class SmartMatchItem(AssetResponse):
+    """智能选素材结果条目（扁平结构）。
+
+    素材字段（id/usable/余量等）直接挂在条目顶层，前端拿到 item 即可读 item.id，
+    与 AssetResponse 字段完全一致；score/breakdown 为智能匹配附加的评分字段。
+    """

-    asset: AssetResponse
    score: float = Field(..., ge=0, le=100, description="综合得分 0-100")
    breakdown: dict[str, float] = Field(default_factory=dict, description="各维度得分明细")

@@ -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

@@ -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):
@@ -22,6 +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):
@@ -3,7 +3,7 @@
 在素材 metadata（assets.classification_result JSON）中持久化已使用的片段时间区间，
 供 from-assets 创建片段时避开历史区间，实现跨任务/跨调用的片段去重；
 素材可用区间耗尽后进入受控复用：允许有限次数（MAX_RANGE_USE_COUNT）复用最久未用
-的历史区间，配合调用方的成片复用占比控制（MAX_REUSE_RATIO = 15%），把任意两条
+的历史区间，配合调用方的成片复用占比控制（MAX_REUSE_RATIO = 10%），把任意两条
 成片的画面重复率控制在阈值内。

 metadata 中的记录字段 ``used_time_ranges``::
@@ -40,14 +40,14 @@ logger = logging.getLogger(__name__)
 USED_RANGES_KEY = "used_time_ranges"

 # ── 受控复用配置常量 ─────────────────────────────────────────────────────────
-MAX_RANGE_USE_COUNT = 3
+MAX_RANGE_USE_COUNT = 2
 """单条历史区间最多被使用次数（含首次），达到后不再参与复用。"""

-REUSE_RATIO_LIMIT = 0.15
-"""单条成片中，单个素材的复用片段累计时长 / 该素材在成片中的总时长上限（15%）。
+REUSE_RATIO_LIMIT = 0.10
+"""单条成片中，单个素材的复用片段累计时长 / 该素材在成片中的总时长上限（10%）。
 超过则该素材不再分配新片段（调用方在轮询分配时跳过）。"""

-SEGMENT_EDGE_GAP = 0.3
+SEGMENT_EDGE_GAP = 1.5
 """冲突判定边缘间隙（秒）：历史区间按 [start-gap, end+gap] 扩边后参与冲突检测，
 避免两条片段首尾紧贴导致画面观感重复；记录仍存实际值。"""

@@ -397,12 +397,12 @@ def make_reuse_callback(
        db: SQLAlchemy session
        asset_durations: 素材 ID -> 总时长（回调需要素材总时长做边界约束）
        reused_tracker: 可选的 ``{asset_id: 累计复用时长}``，回调成功返回复用区间时
-            会把本次片段时长累加进去，供调用方统计成片复用占比（15% 阈值）。
+            会把本次片段时长累加进去，供调用方统计成片复用占比（10% 阈值）。
        assigned_tracker: 可选的 ``{asset_id: 已分配片段总时长}``，配合 ratio_limit
            在复用前预判：若复用本片段后占比 (reused + clip_duration) /
            (assigned + clip_duration) 超过 ratio_limit，则拒绝复用、返回 None
            （保证成片复用占比不超阈值）。
-        ratio_limit: 单条成片复用时长占比上限，默认 15%。
+        ratio_limit: 单条成片复用时长占比上限，默认 10%。

    Returns:
        回调函数 ``(asset_id, clip_duration) -> (start, end) | None``。
@@ -438,3 +438,57 @@ def make_reuse_callback(
        return result

    return _reuse
+
+
+def get_asset_recent_use_counts(
+    db: Session,
+    asset_ids: list[str],
+    recent_video_count: int = 5,
+) -> dict[str, int]:
+    """统计每个素材在最近 N 个不同 plan_id 中的使用次数。
+
+    遍历素材 metadata 中的 used_time_ranges，统计有多少个不同的 plan_id（去重），
+    返回 {asset_id: count}。只统计最近 recent_video_count 个不同 plan_id 的使用次数。
+
+    Args:
+        db: 数据库会话
+        asset_ids: 素材 ID 列表
+        recent_video_count: 统计最近多少个不同 plan_id
+
+    Returns:
+        {asset_id: 在最近 recent_video_count 个 plan 中的使用次数}
+    """
+    if not asset_ids:
+        return {}
+
+    result: dict[str, int] = {}
+    models = db.query(AssetModel).filter(AssetModel.id.in_(asset_ids)).all()
+    for model in models:
+        meta = _read_meta(model)
+        ranges = meta.get(USED_RANGES_KEY) or []
+        if not ranges:
+            result[model.id] = 0
+            continue
+
+        # 按 created_at 倒序收集不同 plan_id
+        sorted_ranges = sorted(
+            ranges,
+            key=lambda r: r.get("created_at") or "",
+            reverse=True,
+        )
+        recent_plan_ids: set[str] = set()
+        for r in sorted_ranges:
+            plan_id = r.get("plan_id")
+            if plan_id:
+                recent_plan_ids.add(plan_id)
+            if len(recent_plan_ids) >= recent_video_count:
+                break
+
+        result[model.id] = len(recent_plan_ids)
+
+    # 未找到的素材计为 0
+    for aid in asset_ids:
+        if aid not in result:
+            result[aid] = 0
+
+    return result
@@ -13,6 +13,7 @@
 from __future__ import annotations

 import logging
+import random
 from typing import Any, List

 from sqlalchemy.orm import Session
@@ -29,9 +30,11 @@ from packages.domain.editing_mode import EditingMode
 from packages.domain.plan_generator_utils import (
    create_clips_from_configs,
    distribute_assets,
+    extract_scene_points_from_metadata,
    generate_default_clips,
    map_clip_types_for_mode,
 )
+from packages.domain.smart_match import SCORE_RANDOM_NOISE_MAX, score_asset
 from packages.domain.template_clip_config import TemplateClipConfig

 logger = logging.getLogger(__name__)
@@ -128,6 +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
@@ -215,19 +219,82 @@ 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）.

-        委托给 plan_generator_utils.distribute_assets 纯函数。
+        先用 smart_match 评分对素材排序（高分优先），再委托给
+        plan_generator_utils.distribute_assets 纯函数完成分配。
        """
+        # 预览随机模式：素材顺序已 shuffle，纯随机起点即可，不读 DB 评分/缓存
+        asset_scene_points: dict[str, list[float]] = {}
+        if not random_selection:
+            # 正式生成：smart_match 评分排序（高分优先）+ 场景切换点缓存
+            if self._asset_repo:
+                asset_ids = self._sort_assets_by_smart_score(asset_ids)
+                # 读取素材 metadata 中的场景切换点缓存（后台 SceneChange 检测写入）：
+                # 有缓存的素材片段起点从随机镜头段选取，无缓存走随机起点兜底
+                asset_scene_points = self._fetch_asset_scene_points(asset_ids)
+
+                # 正式生成也随机重排片段顺序（降重，默认开启无开关）
+                # smart_match 决定选哪些素材，shuffle 只改变分配到 clips 的顺序
+                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,
            editing_mode,
            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]]:
+        """从素材 metadata 读取场景切换点缓存（无缓存的素材不包含在结果中）。"""
+        points_map: dict[str, list[float]] = {}
+        if not self._asset_repo:
+            return points_map
+        for asset_id in asset_ids:
+            asset = self._asset_repo.get(asset_id)
+            if asset:
+                points = extract_scene_points_from_metadata(getattr(asset, "metadata", None))
+                if points:
+                    points_map[asset_id] = points
+        return points_map
+
+    def _sort_assets_by_smart_score(self, asset_ids: List[str]) -> List[str]:
+        """按 smart_match 综合评分降序排列素材 ID（注入随机噪声）。
+
+        评分高的素材（质量好、时长合适、新鲜、使用次数少）倾向排在前面；
+        排序时给每个素材的得分注入 0~SCORE_RANDOM_NOISE_MAX 的随机噪声，
+        使得分接近的素材排名每次浮动，避免一键生成反复选出相同素材组合，
+        从素材组合层面降低成片查重率。分差大于噪声上限时排名保持稳定。
+        """
+        scored: list[tuple[str, float]] = []
+        for asset_id in asset_ids:
+            asset = self._asset_repo.get(asset_id)
+            if asset:
+                score, _ = score_asset(asset)
+                scored.append((asset_id, score))
+            else:
+                scored.append((asset_id, 0.0))
+        # 评分 + 随机噪声后按降序排列
+        scored.sort(
+            key=lambda x: x[1] + random.uniform(0.0, SCORE_RANDOM_NOISE_MAX),
+            reverse=True,
+        )
+        return [aid for aid, _ in scored]
+
    def _fetch_asset_durations(self, asset_ids: List[str]) -> dict[str, float]:
        """从数据库获取素材时长信息.

@@ -0,0 +1,174 @@
+#!/usr/bin/env python3
+"""存量指纹重建脚本 — 为已有视频生成 video_fingerprint_chunks 分片数据。
+
+功能：
+- 查询 generated_videos 中 video_fingerprint IS NOT NULL 但尚无分片数据的视频
+- 从 OSS 下载视频 → 用新的分片算法重新计算指纹 → 写入分片表
+- 支持 --dry-run（只打印不写入）和 --batch-size（默认 50）
+- 幂等：已存在分片数据的视频跳过
+
+用法：
+    # 预览（不写入）
+    python rebuild_fingerprint_chunks.py --dry-run
+
+    # 执行重建
+    python rebuild_fingerprint_chunks.py --batch-size 50
+"""
+
+from __future__ import annotations
+
+import argparse
+import logging
+import os
+import sys
+import tempfile
+
+# 确保可以 import worker_app 和 packages
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "..", "..", "worker"))
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "..", ".."))
+
+logging.basicConfig(
+    level=logging.INFO,
+    format="%(asctime)s [%(levelname)s] %(name)s: %(message)s",
+)
+logger = logging.getLogger("rebuild_fingerprint_chunks")
+
+
+def find_videos_needing_rebuild(session, batch_size: int) -> list[dict]:
+    """查询需要重建分片指纹的视频。"""
+    from sqlalchemy import and_
+
+    from packages.adapters.sqlalchemy_impl.models import GeneratedVideoModel, VideoFingerprintChunkModel
+
+    # 有 video_fingerprint 的视频
+    has_fingerprint = GeneratedVideoModel.video_fingerprint.isnot(None)
+    has_fingerprint = and_(has_fingerprint, GeneratedVideoModel.video_fingerprint != "")
+
+    # 排除已有分片数据的视频
+    subq = session.query(VideoFingerprintChunkModel.video_id).distinct().subquery()
+    no_chunks = ~GeneratedVideoModel.id.in_(subq)
+
+    videos = (
+        session.query(GeneratedVideoModel)
+        .filter(and_(has_fingerprint, no_chunks))
+        .order_by(GeneratedVideoModel.generated_at.desc())
+        .limit(batch_size)
+        .all()
+    )
+
+    return [
+        {
+            "id": v.id,
+            "project_id": v.project_id,
+            "user_id": v.user_id or "",
+            "duration": v.duration,
+        }
+        for v in videos
+    ]
+
+
+def rebuild_one(video_info: dict, dry_run: bool = False) -> int:
+    """重建单个视频的分片数据。返回写入的 chunk 数量。"""
+    from video_processing.dedup import VideoDeduplicator, _save_fingerprint_chunks
+    from worker_app.db import SessionLocal
+
+    from packages.adapters.sqlalchemy_impl.models import VideoFingerprintChunkModel
+    from packages.shared.storage import get_storage_service
+
+    video_id = video_info["id"]
+    project_id = video_info["project_id"]
+    user_id = video_info["user_id"]
+
+    if dry_run:
+        logger.info("[DRY-RUN] Would rebuild video %s (project=%s)", video_id, project_id)
+        return 0
+
+    session = SessionLocal()
+    temp_dir = tempfile.mkdtemp()
+
+    try:
+        # 再次检查幂等性
+        existing_count = (
+            session.query(VideoFingerprintChunkModel).filter(VideoFingerprintChunkModel.video_id == video_id).count()
+        )
+        if existing_count > 0:
+            logger.info("Video %s already has %d chunks, skipping", video_id, existing_count)
+            return 0
+
+        # 下载视频
+        storage_service = get_storage_service()
+        local_path = os.path.join(temp_dir, f"{video_id}.mp4")
+        storage_key = f"projects/{project_id}/generated/{video_id}/{video_id}.mp4"
+        storage_service.download_file(storage_key, local_path)
+
+        # 重新计算指纹
+        deduplicator = VideoDeduplicator()
+        fingerprint = deduplicator.compute_fingerprint(local_path)
+
+        # 写入分片表
+        _save_fingerprint_chunks(fingerprint, video_id, project_id, user_id, session)
+        session.commit()
+
+        chunk_count = len(fingerprint.chunks)
+        logger.info("Rebuilt %d chunks for video %s", chunk_count, video_id)
+        return chunk_count
+
+    except Exception as e:
+        logger.error("Failed to rebuild video %s: %s", video_id, e)
+        session.rollback()
+        return -1
+    finally:
+        session.close()
+        import shutil
+
+        shutil.rmtree(temp_dir, ignore_errors=True)
+
+
+def main():
+    parser = argparse.ArgumentParser(description="存量指纹重建脚本")
+    parser.add_argument("--dry-run", action="store_true", help="只打印不写入")
+    parser.add_argument("--batch-size", type=int, default=50, help="每批处理数量（默认 50）")
+    parser.add_argument("--total-limit", type=int, default=0, help="总处理数量限制（0=不限制）")
+    args = parser.parse_args()
+
+    from worker_app.db import SessionLocal
+
+    session = SessionLocal()
+
+    try:
+        videos = find_videos_needing_rebuild(session, args.batch_size)
+        logger.info("Found %d videos needing rebuild", len(videos))
+
+        if args.dry_run:
+            for v in videos:
+                logger.info("[DRY-RUN] Video %s | project=%s | duration=%.1fs", v["id"], v["project_id"], v["duration"])
+            return
+
+        total_chunks = 0
+        processed = 0
+        failed = 0
+
+        for v in videos:
+            if args.total_limit > 0 and processed >= args.total_limit:
+                break
+
+            result = rebuild_one(v, dry_run=False)
+            if result < 0:
+                failed += 1
+            else:
+                total_chunks += result
+                processed += 1
+
+        logger.info(
+            "Rebuild complete: processed=%d, chunks=%d, failed=%d",
+            processed,
+            total_chunks,
+            failed,
+        )
+
+    finally:
+        session.close()
+
+
+if __name__ == "__main__":
+    main()
@@ -1848,10 +1848,9 @@
    },
    "node_modules/@testing-library/dom": {
      "version": "10.4.1",
-      "resolved": "https://registry.npmmirror.com/@testing-library/dom/-/dom-10.4.1.tgz",
+      "resolved": "https://registry.npmjs.org/@testing-library/dom/-/dom-10.4.1.tgz",
      "integrity": "sha512-o4PXJQidqJl82ckFaXUeoAW+XysPLauYI43Abki5hABd853iMhitooc6znOnczgbTYmEP6U6/y1ZyKAIsvMKGg==",
      "dev": true,
-      "license": "MIT",
      "peer": true,
      "dependencies": {
        "@babel/code-frame": "^7.10.4",
@@ -1938,10 +1937,9 @@
    },
    "node_modules/@types/aria-query": {
      "version": "5.0.4",
-      "resolved": "https://registry.npmmirror.com/@types/aria-query/-/aria-query-5.0.4.tgz",
+      "resolved": "https://registry.npmjs.org/@types/aria-query/-/aria-query-5.0.4.tgz",
      "integrity": "sha512-rfT93uj5s0PRL7EzccGMs3brplhcrghnDoV26NqKhCAS1hVo+WdNsPvE/yb6ilfr5hi2MEk6d5EWJTKdxg8jVw==",
      "dev": true,
-      "license": "MIT",
      "peer": true
    },
    "node_modules/@types/babel__core": {
@@ -3113,10 +3111,9 @@
    },
    "node_modules/dom-accessibility-api": {
      "version": "0.5.16",
-      "resolved": "https://registry.npmmirror.com/dom-accessibility-api/-/dom-accessibility-api-0.5.16.tgz",
+      "resolved": "https://registry.npmjs.org/dom-accessibility-api/-/dom-accessibility-api-0.5.16.tgz",
      "integrity": "sha512-X7BJ2yElsnOJ30pZF4uIIDfBEVgF4XEBxL9Bxhy6dnrm5hkzqmsWHGTiHqRiITNhMyFLyAiWndIJP7Z1NTteDg==",
      "dev": true,
-      "license": "MIT",
      "peer": true
    },
    "node_modules/dunder-proto": {
@@ -4457,10 +4454,9 @@
    },
    "node_modules/lz-string": {
      "version": "1.5.0",
-      "resolved": "https://registry.npmmirror.com/lz-string/-/lz-string-1.5.0.tgz",
+      "resolved": "https://registry.npmjs.org/lz-string/-/lz-string-1.5.0.tgz",
      "integrity": "sha512-h5bgJWpxJNswbU7qCrV0tIKQCaS3blPDrqKWx+QxzuzL1zGUzij9XCWLrSLsJPu5t+eWA/ycetzYAO5IOMcWAQ==",
      "dev": true,
-      "license": "MIT",
      "peer": true,
      "bin": {
        "lz-string": "bin/bin.js"
@@ -5008,10 +5004,9 @@
    },
    "node_modules/pretty-format": {
      "version": "27.5.1",
-      "resolved": "https://registry.npmmirror.com/pretty-format/-/pretty-format-27.5.1.tgz",
+      "resolved": "https://registry.npmjs.org/pretty-format/-/pretty-format-27.5.1.tgz",
      "integrity": "sha512-Qb1gy5OrP5+zDf2Bvnzdl3jsTf1qXVMazbvCoKhtKqVs4/YK4ozX4gKQJJVyNe+cajNPn0KoC0MC3FUmaHWEmQ==",
      "dev": true,
-      "license": "MIT",
      "peer": true,
      "dependencies": {
        "ansi-regex": "^5.0.1",
@@ -5024,10 +5019,9 @@
    },
    "node_modules/pretty-format/node_modules/ansi-styles": {
      "version": "5.2.0",
-      "resolved": "https://registry.npmmirror.com/ansi-styles/-/ansi-styles-5.2.0.tgz",
+      "resolved": "https://registry.npmjs.org/ansi-styles/-/ansi-styles-5.2.0.tgz",
      "integrity": "sha512-Cxwpt2SfTzTtXcfOlzGEee8O+c+MmUgGrNiBcXnuWxuFJHe6a5Hz7qwhwe5OgaSYI0IJvkLqWX1ASG+cJOkEiA==",
      "dev": true,
-      "license": "MIT",
      "peer": true,
      "engines": {
        "node": ">=10"
@@ -5735,10 +5729,9 @@
    },
    "node_modules/react-is": {
      "version": "17.0.2",
-      "resolved": "https://registry.npmmirror.com/react-is/-/react-is-17.0.2.tgz",
+      "resolved": "https://registry.npmjs.org/react-is/-/react-is-17.0.2.tgz",
      "integrity": "sha512-w2GsyukL62IJnlaff/nRegPQR94C/XXamvMWmSHRJ4y7Ts/4ocGRmTHvOs8PSE6pB3dWOrD/nueuU5sduBsQ4w==",
      "dev": true,
-      "license": "MIT",
      "peer": true
    },
    "node_modules/react-refresh": {
@@ -69,10 +69,13 @@ interface SmartMatchWrappedItem {
  breakdown?: unknown
 }

-export const smartMatchAssets = async (libraryId: string): Promise<SmartMatchResult> => {
-  const response = await apiClient.post("/assets/smart-match", {
-    library_id: libraryId,
-  })
+export const smartMatchAssets = async (
+  libraryId: string,
+  limit?: number,
+): Promise<SmartMatchResult> => {
+  const payload: Record<string, unknown> = { library_id: libraryId }
+  if (limit && limit > 0) payload.limit = limit
+  const response = await apiClient.post("/assets/smart-match", payload)
  const rawItems: SmartMatchWrappedItem[] = response.data?.items ?? []
  const items = rawItems
    .map((it) =>
@@ -2,12 +2,17 @@
 * 素材诊断 API
 */
 import apiClient from "../client"
+import { getOrCreateDefaultProject } from "../projects"
 import type { AssetDiagnosis } from "./types"

 /** 获取素材诊断信息（可选 asset_id 查单素材，否则全局诊断） */
-export const getAssetDiagnosis = async (assetId?: string): Promise<AssetDiagnosis> => {
+export const getAssetDiagnosis = async (
+  assetId?: string,
+  projectId?: string,
+): Promise<AssetDiagnosis> => {
+  const pid = projectId ?? (await getOrCreateDefaultProject()).id
  const params: Record<string, string> = {}
  if (assetId) params.asset_id = assetId
-  const response = await apiClient.get("/asset-diagnosis", { params })
+  const response = await apiClient.get(`/projects/${pid}/asset-diagnosis`, { params })
  return response.data
 }
@@ -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 */
@@ -23,6 +23,10 @@ export interface ProductItem {
  project_name?: string
  /** 查重率（百分比） */
  duplicate_rate?: number
+  /** 视觉相似度（0-100），#1660 新增 */
+  visual_similarity?: number
+  /** 匹配帧数，#1660 新增 */
+  match_count?: number
  created_at?: string
  updated_at?: string
 }
@@ -71,4 +75,9 @@ export interface VideoItem {
  generation_params: Record<string, unknown>
  download_url: string
  generated_at: string
+  duplicate_rate?: number
+  /** 视觉相似度（0-100），#1660 新增 */
+  visual_similarity?: number
+  /** 匹配帧数，#1660 新增 */
+  match_count?: number
 }
@@ -29,7 +29,8 @@ export function mapVideoToProductItem(video: VideoItem): ProductItem {
    // 后端字段名为 generated_at，映射为 created_at 供前端统一使用
    created_at: video.generated_at,
    updated_at: video.generated_at,
-    // 后端 /videos 接口暂无 duplicate_rate 字段
-    duplicate_rate: undefined,
+    duplicate_rate: video.duplicate_rate,
+    visual_similarity: video.visual_similarity,
+    match_count: video.match_count,
  }
 }
@@ -28,4 +28,5 @@ export {
  deleteTTSJob,
  getTtsVoices,
  previewTts,
+  extractVideoVoice,
 } from "./jobs"
@@ -70,3 +70,56 @@ export const previewTts = async (data: TTSPreviewRequest): Promise<TTSPreviewRes
  const response = await apiClient.post<TTSPreviewResponse>("/tts/preview", data)
  return response.data
 }
+
+/**
+ * 从视频中提取配音（上传视频 → 后端提取人声 → 保存到配音素材库）
+ * 支持 mp4/mov/webm 格式
+ */
+export const extractVideoVoice = async (
+  file: File,
+  onProgress?: (percent: number) => void,
+): Promise<{ asset_id: string; duration: number }> => {
+  const formData = new FormData()
+  formData.append("file", file)
+
+  return new Promise((resolve, reject) => {
+    const xhr = new XMLHttpRequest()
+    xhr.open("POST", "/api/v1/voices/extract-voice")
+
+    // 携带认证 token（从 localStorage 获取，与 apiClient 拦截器一致）
+    const token = localStorage.getItem("access_token")
+    if (token) {
+      xhr.setRequestHeader("Authorization", `Bearer ${token}`)
+    }
+
+    xhr.timeout = 10 * 60 * 1000 // 10 分钟超时
+
+    xhr.upload.onprogress = (e) => {
+      if (e.lengthComputable && onProgress) {
+        onProgress(Math.round((e.loaded / e.total) * 100))
+      }
+    }
+
+    xhr.onload = () => {
+      if (xhr.status >= 200 && xhr.status < 300) {
+        try {
+          resolve(JSON.parse(xhr.responseText))
+        } catch {
+          reject(new Error("服务器返回数据解析失败"))
+        }
+      } else {
+        try {
+          const err = JSON.parse(xhr.responseText)
+          reject(new Error(err.detail || err.message || `提取失败: HTTP ${xhr.status}`))
+        } catch {
+          reject(new Error(`提取失败: HTTP ${xhr.status}`))
+        }
+      }
+    }
+
+    xhr.onerror = () => reject(new Error("网络错误，请检查网络连接"))
+    xhr.ontimeout = () => reject(new Error("上传超时（10分钟），请检查网络或尝试更小的文件"))
+
+    xhr.send(formData)
+  })
+}
@@ -297,7 +297,7 @@ const CloneModal: React.FC<CloneModalProps> = ({ open, onClose, onSuccess }) =>
                    buttonSize="sm"
                    onClick={() => {
                      handleClose()
-                      navigate("/app/voice-materials")
+                      navigate("/app/voices?tab=material&upload=1")
                    }}
                  >
                    去配音库上传
@@ -3,9 +3,11 @@ import { useQuery } from "@tanstack/react-query"
 import {
  getAssetLibraries,
  getAssets,
+  ensureDefaultLibrary,
  type AssetLibraryItem,
  type AssetItem as ApiAssetItem,
 } from "@/api/assets"
+import { getOrCreateDefaultProject } from "@/api/projects"
 import { mapLibrary, mapAsset, type AssetItem, type LibraryItem } from "../types"

 /**
@@ -16,7 +18,18 @@ export function useAssetsData() {
  /* ── 视频库列表查询 ── */
  const { data: apiLibraries = [], isLoading: libLoading } = useQuery<AssetLibraryItem[], Error>({
    queryKey: ["asset-libraries"],
-    queryFn: getAssetLibraries,
+    queryFn: async () => {
+      const libs = await getAssetLibraries()
+      // 如果没有 video 类型的库，自动创建默认视频素材库（与 useVoiceMaterials 保持一致）
+      const hasVideoLib = libs.some((lib) => lib.kind === "video")
+      if (!hasVideoLib) {
+        const project = await getOrCreateDefaultProject()
+        await ensureDefaultLibrary({ project_id: project.id, kind: "video" })
+        // 创建后重新拉取最新列表
+        return getAssetLibraries()
+      }
+      return libs
+    },
    staleTime: 60_000,
  })

@@ -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>
  )
@@ -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">
@@ -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>
+  )
+}
@@ -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;
+}
@@ -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） */
@@ -98,8 +98,8 @@ export function useTemplateSave(options: UseTemplateSaveOptions) {
        estimated_duration: totalDuration,
        segments: clips.map((c, i) => ({
          segment_order: i,
-          duration_min: Math.max(1, c.duration - 2),
-          duration_max: c.duration + 2,
+          duration_min: c.duration,
+          duration_max: c.duration,
          material_type: c.type === "voice" ? "voiceover" : "video",
          transition: c.transition
            ? { type: c.transition.type, duration: c.transition.duration }
@@ -338,12 +338,15 @@ const GeneratePage: React.FC = () => {
                size: titleSettings.size,
                font: titleSettings.font,
                color: titleSettings.color,
-                position: titleSettings.position as "top" | "center" | "bottom",
+                position: titleSettings.position as "top" | "center" | "bottom" | "custom",
                bold: titleSettings.bold,
                italic: titleSettings.italic,
                stroke: titleSettings.stroke,
                shadow: titleSettings.shadow,
+                posX: titleSettings.posX,
+                posY: titleSettings.posY,
              }}
+              onTitlePositionChange={styleUpdaters.updateTitlePosition}
            />
          )}
          {currentStep >= 5 && generated && finalVideo && (
@@ -32,12 +32,15 @@ interface FrontendPreviewPlayerProps {
    size: number
    font: string
    color: string
-    position: "top" | "center" | "bottom"
+    position: "top" | "center" | "bottom" | "custom"
    bold?: boolean
    italic?: boolean
    stroke?: boolean
    shadow?: boolean
+    posX?: number | null
+    posY?: number | null
  }
+  onTitlePositionChange?: (posX: number, posY: number) => void
 }

 function formatTime(seconds: number): string {
@@ -105,6 +108,7 @@ const FrontendPreviewPlayer: React.FC<FrontendPreviewPlayerProps> = ({
  serverClips,
  voiceAudioUrl,
  titleSettings,
+  onTitlePositionChange,
 }) => {
  const segments = useMemo(
    () => buildPlaybackSegments(assets, template, serverClips),
@@ -126,6 +130,69 @@ const FrontendPreviewPlayer: React.FC<FrontendPreviewPlayerProps> = ({
        return { width: 1080, height: 1920 }
    }
  })()
+  const customTitleXPct =
+    titleSettings?.posX != null && playRes.width > 0
+      ? (titleSettings.posX / playRes.width) * 100
+      : null
+  const customTitleYPct =
+    titleSettings?.posY != null && playRes.height > 0
+      ? (titleSettings.posY / playRes.height) * 100
+      : null
+
+  // ── 拖拽状态（用 ref 避免在每帧渲染中触发重渲染）──
+  const draggingTitleRef = useRef(false)
+  const titleDragRef = useRef<HTMLDivElement>(null)
+  const handleTitlePointerDown = useCallback(
+    (e: React.PointerEvent<HTMLDivElement>) => {
+      if (!onTitlePositionChange || !playerContainerRef.current) return
+      e.preventDefault()
+      e.stopPropagation()
+      ;(e.target as Element).setPointerCapture(e.pointerId)
+      draggingTitleRef.current = true
+      ;(e.currentTarget as HTMLDivElement).style.cursor = "grabbing"
+    },
+    [onTitlePositionChange],
+  )
+  const handleTitlePointerMove = useCallback((e: React.PointerEvent<HTMLDivElement>) => {
+    if (!draggingTitleRef.current || !playerContainerRef.current) return
+    e.preventDefault()
+    e.stopPropagation()
+    // 拖拽过程中直接修改 DOM，不触发 React 渲染（避免频繁重渲染导致换行）
+    if (titleDragRef.current) {
+      const rect = playerContainerRef.current.getBoundingClientRect()
+      const relX = Math.max(0, Math.min(rect.width, e.clientX - rect.left))
+      const relY = Math.max(0, Math.min(rect.height, e.clientY - rect.top))
+      const xpct = (relX / rect.width) * 100
+      const ypct = (relY / rect.height) * 100
+      titleDragRef.current.style.left = `${xpct}%`
+      titleDragRef.current.style.top = `${ypct}%`
+    }
+  }, [])
+  const handleTitlePointerUp = useCallback(
+    (e: React.PointerEvent<HTMLDivElement>) => {
+      if (!draggingTitleRef.current) return
+      draggingTitleRef.current = false
+      // 拖拽结束时才调用 onTitlePositionChange 保存最终位置
+      if (onTitlePositionChange && playerContainerRef.current) {
+        const rect = playerContainerRef.current.getBoundingClientRect()
+        const relX = Math.max(0, Math.min(rect.width, e.clientX - rect.left))
+        const relY = Math.max(0, Math.min(rect.height, e.clientY - rect.top))
+        const posX = Math.round((relX / rect.width) * playRes.width)
+        const posY = Math.round((relY / rect.height) * playRes.height)
+        onTitlePositionChange(posX, posY)
+      }
+      ;(e.currentTarget as HTMLDivElement).style.cursor = "grab"
+      try {
+        if ((e.currentTarget as Element).hasPointerCapture(e.pointerId)) {
+          ;(e.currentTarget as Element).releasePointerCapture(e.pointerId)
+        }
+      } catch {
+        /* ignore */
+      }
+    },
+    [onTitlePositionChange, playRes.width, playRes.height],
+  )
+
  const playerContainerRef = useRef<HTMLDivElement>(null)
  const [containerHeight, setContainerHeight] = useState(0)
  useEffect(() => {
@@ -524,15 +591,36 @@ const FrontendPreviewPlayer: React.FC<FrontendPreviewPlayerProps> = ({
          <div
            style={{
              position: "absolute",
-              left: `${titleSidePct}%`,
-              right: `${titleSidePct}%`,
-              textAlign: "center",
-              ...(titleSettings.position === "top"
-                ? { top: `${titleTopPct}%` }
-                : titleSettings.position === "center"
-                  ? { top: "50%", transform: "translateY(-50%)" }
-                  : { bottom: `${titleBottomPct}%` }),
+              ...(customTitleXPct != null && customTitleYPct != null
+                ? {
+                    left: `${customTitleXPct}%`,
+                    top: `${customTitleYPct}%`,
+                    transform: "translate(-50%, -50%)",
+                    textAlign: "center" as const,
+                  }
+                : {
+                    left: `${titleSidePct}%`,
+                    right: `${titleSidePct}%`,
+                    textAlign: "center" as const,
+                    ...(titleSettings.position === "top"
+                      ? { top: `${titleTopPct}%` }
+                      : titleSettings.position === "center"
+                        ? { top: "50%", transform: "translateY(-50%)" }
+                        : { bottom: `${titleBottomPct}%` }),
+                  }),
+              pointerEvents: "auto",
+              cursor: onTitlePositionChange ? "grab" : "default",
+              touchAction: "none",
+              userSelect: "none",
+              WebkitUserSelect: "none",
+              padding: "8px 12px",
+              boxShadow: "inset 0 0 0 16px transparent",
            }}
+            ref={titleDragRef}
+            onPointerDown={handleTitlePointerDown}
+            onPointerMove={handleTitlePointerMove}
+            onPointerUp={handleTitlePointerUp}
+            onPointerCancel={handleTitlePointerUp}
          >
            <span
              style={{
@@ -134,7 +134,7 @@ const Step5VoiceSelect: React.FC<Step5VoiceSelectProps> = ({

  /** 跳转到配音库上传 */
  const handleGoToUpload = useCallback(() => {
-    navigate("/app/voices")
+    navigate("/app/voices?tab=material&upload=1")
  }, [navigate])

  // 加载中状态
@@ -23,7 +23,7 @@ const MaterialModeTabs: React.FC<MaterialModeTabsProps> = ({ mode, onModeChange
        onClick={() => onModeChange("auto")}
        type="button"
      >
-        选择视频库自动匹配
+        AI智能匹配
      </button>
    </div>
  )
@@ -1,5 +1,7 @@
 /**
 * 标题预设样式网格
+ * 双图层渲染：底层=描边轮廓(text-shadow模拟)，上层=填充色
+ * 避免 -webkit-text-stroke 在 Chromium 中吞掉填充色的问题
 */
 import React from "react"
 import { getFontFamily } from "../../constants"
@@ -7,7 +9,10 @@ import { getFontFamily } from "../../constants"
 interface TitlePresetItem {
  key: string
  label: string
-  previewStyle: React.CSSProperties
+  previewStyle: React.CSSProperties & {
+    _strokeColor?: string
+    _strokeWidth?: number
+  }
 }

 interface TitlePresetsGridProps {
@@ -17,6 +22,32 @@ interface TitlePresetsGridProps {
  fontFamily?: string
 }

+/**
+ * 用 text-shadow 模拟描边轮廓（8方向 + 4对角 = 12层阴影）
+ */
+function buildStrokeShadow(color: string, width: number): string {
+  const w = width
+  const parts: string[] = []
+  // 4 cardinal directions
+  parts.push(`${w}px 0 ${color}`)
+  parts.push(`${-w}px 0 ${color}`)
+  parts.push(`0 ${w}px ${color}`)
+  parts.push(`0 ${-w}px ${color}`)
+  // 4 diagonal directions
+  const d = Math.round(w * 0.71 * 10) / 10 // 0.71 ≈ sqrt(2)/2
+  parts.push(`${d}px ${d}px ${color}`)
+  parts.push(`${-d}px ${d}px ${color}`)
+  parts.push(`${d}px ${-d}px ${color}`)
+  parts.push(`${-d}px ${-d}px ${color}`)
+  // 4 extra mid-points for smoother stroke
+  const h = Math.round(w * 0.5 * 10) / 10
+  parts.push(`${w}px ${h}px ${color}`)
+  parts.push(`${w}px ${-h}px ${color}`)
+  parts.push(`${-w}px ${h}px ${color}`)
+  parts.push(`${-w}px ${-h}px ${color}`)
+  return parts.join(", ")
+}
+
 const TitlePresetsGrid: React.FC<TitlePresetsGridProps> = ({
  presets,
  activePreset,
@@ -27,20 +58,53 @@ const TitlePresetsGrid: React.FC<TitlePresetsGridProps> = ({
    <div className="xx-title-presets-grid">
      {presets.map((p) => {
        const isActive = activePreset === p.key
+        const { _strokeColor, _strokeWidth, ...fillStyle } = p.previewStyle
+        const ff = getFontFamily(fontFamily || "思源黑体")
+
+        // 底层：描边轮廓（用 text-shadow 模拟粗描边）
+        const strokeStyle: React.CSSProperties = {
+          color: _strokeColor || "transparent",
+          textShadow:
+            _strokeColor && _strokeWidth
+              ? buildStrokeShadow(_strokeColor, _strokeWidth)
+              : undefined,
+          fontWeight: fillStyle.fontWeight,
+          fontSize: fillStyle.fontSize,
+          lineHeight: 1,
+        }
+
+        // 上层：仅填充色 + 可选 textShadow（发光/投影效果）
+        const topStyle: React.CSSProperties = {
+          color: fillStyle.color,
+          textShadow: fillStyle.textShadow,
+          fontWeight: fillStyle.fontWeight,
+          fontSize: fillStyle.fontSize,
+          lineHeight: 1,
+        }
+
        return (
          <button
            key={p.key}
            className={`xx-title-preset-card${isActive ? " active" : ""}`}
            onClick={() => onApply(p.key)}
-            title={p.label}
          >
-            <span
-              className="xx-title-preset-preview-text"
-              style={{ ...p.previewStyle, fontFamily: getFontFamily(fontFamily || "思源黑体") }}
-            >
-              标题
+            <span className="xx-title-preset-preview-text" style={{ position: "relative" }}>
+              {/* 底层：描边轮廓 */}
+              <span
+                aria-hidden
+                style={{
+                  ...strokeStyle,
+                  fontFamily: ff,
+                  position: "absolute",
+                  top: 0,
+                  left: 0,
+                }}
+              >
+                T
+              </span>
+              {/* 上层：填充色 */}
+              <span style={{ ...topStyle, fontFamily: ff, position: "relative" }}>T</span>
            </span>
-            <span className="xx-title-preset-card-label">{p.label}</span>
          </button>
        )
      })}
@@ -42,6 +42,7 @@ export const POSITION_OPTIONS = [
  { value: "top", label: "顶部" },
  { value: "center", label: "居中" },
  { value: "bottom", label: "底部" },
+  { value: "custom", label: "自定义" },
 ]

 /* ── 标题字体选项 ── */
@@ -77,10 +78,11 @@ export const TITLE_PRESETS = [
    label: "经典白字",
    style: { size: 28, color: "#ffffff", bold: true, italic: false, stroke: true, shadow: false },
    previewStyle: {
-      fontWeight: 700,
      color: "#ffffff",
-      WebkitTextStroke: "1px #000000",
-      fontSize: "20px",
+      _strokeColor: "#000000",
+      _strokeWidth: 2,
+      fontWeight: 700,
+      fontSize: "32px",
    },
  },
  {
@@ -88,45 +90,54 @@ export const TITLE_PRESETS = [
    label: "黑金质感",
    style: { size: 32, color: "#d4a843", bold: true, italic: false, stroke: false, shadow: true },
    previewStyle: {
-      fontWeight: 700,
      color: "#d4a843",
+      fontWeight: 700,
+      fontSize: "32px",
      textShadow: "1px 1px 3px rgba(0,0,0,0.8)",
-      fontSize: "20px",
    },
  },
  {
    key: "fresh_minimal",
    label: "清新简约",
    style: { size: 24, color: "#333333", bold: false, italic: false, stroke: false, shadow: false },
-    previewStyle: { fontWeight: 400, color: "#333333", fontSize: "18px" },
+    previewStyle: {
+      color: "#e8e8e8",
+      fontWeight: 400,
+      fontSize: "32px",
+    },
  },
  {
    key: "variety_show",
    label: "综艺花字",
    style: { size: 36, color: "#ff4081", bold: true, italic: false, stroke: true, shadow: true },
    previewStyle: {
-      fontWeight: 900,
      color: "#ff4081",
-      WebkitTextStroke: "1.5px #ffffff",
+      _strokeColor: "#ffffff",
+      _strokeWidth: 2,
+      fontWeight: 900,
+      fontSize: "32px",
      textShadow: "2px 2px 4px rgba(0,0,0,0.5)",
-      fontSize: "22px",
    },
  },
  {
    key: "business",
    label: "商务极简",
    style: { size: 24, color: "#1a1a1a", bold: false, italic: false, stroke: false, shadow: false },
-    previewStyle: { fontWeight: 400, color: "#1a1a1a", fontSize: "17px" },
+    previewStyle: {
+      color: "#e0e0e0",
+      fontWeight: 400,
+      fontSize: "32px",
+    },
  },
  {
    key: "retro_film",
    label: "复古胶片",
    style: { size: 28, color: "#e8d5b7", bold: false, italic: false, stroke: false, shadow: true },
    previewStyle: {
-      fontWeight: 400,
      color: "#e8d5b7",
+      fontWeight: 400,
+      fontSize: "32px",
      textShadow: "2px 2px 6px rgba(0,0,0,0.7)",
-      fontSize: "18px",
    },
  },
  {
@@ -134,10 +145,10 @@ export const TITLE_PRESETS = [
    label: "霓虹发光",
    style: { size: 32, color: "#00e5ff", bold: true, italic: false, stroke: false, shadow: true },
    previewStyle: {
-      fontWeight: 700,
      color: "#00e5ff",
+      fontWeight: 700,
+      fontSize: "32px",
      textShadow: "0 0 4px #00e5ff, 0 0 8px #00e5ff, 0 0 16px rgba(0,229,255,0.5)",
-      fontSize: "20px",
    },
  },
  {
@@ -145,10 +156,202 @@ export const TITLE_PRESETS = [
    label: "手写字",
    style: { size: 28, color: "#333333", bold: false, italic: false, stroke: false, shadow: true },
    previewStyle: {
+      color: "#e0e0e0",
      fontWeight: 400,
-      color: "#333333",
+      fontSize: "32px",
      textShadow: "1px 1px 2px rgba(0,0,0,0.3)",
-      fontSize: "20px",
+    },
+  },
+  {
+    key: "outline_yellow",
+    label: "黄色描边",
+    style: { size: 28, color: "#ffd54f", bold: true, italic: false, stroke: true, shadow: false },
+    previewStyle: {
+      color: "#ffd54f",
+      _strokeColor: "#000000",
+      _strokeWidth: 2,
+      fontWeight: 700,
+      fontSize: "32px",
+    },
+  },
+  {
+    key: "outline_pink",
+    label: "粉色描边",
+    style: { size: 28, color: "#ff80ab", bold: true, italic: false, stroke: true, shadow: false },
+    previewStyle: {
+      color: "#ff80ab",
+      _strokeColor: "#000000",
+      _strokeWidth: 2,
+      fontWeight: 700,
+      fontSize: "32px",
+    },
+  },
+  {
+    key: "outline_blue",
+    label: "蓝色描边",
+    style: { size: 28, color: "#82b1ff", bold: true, italic: false, stroke: true, shadow: false },
+    previewStyle: {
+      color: "#82b1ff",
+      _strokeColor: "#000000",
+      _strokeWidth: 2,
+      fontWeight: 700,
+      fontSize: "32px",
+    },
+  },
+  {
+    key: "outline_green",
+    label: "绿色描边",
+    style: { size: 28, color: "#69f0ae", bold: true, italic: false, stroke: true, shadow: false },
+    previewStyle: {
+      color: "#69f0ae",
+      _strokeColor: "#000000",
+      _strokeWidth: 2,
+      fontWeight: 700,
+      fontSize: "32px",
+    },
+  },
+  {
+    key: "outline_gray",
+    label: "灰色描边",
+    style: { size: 28, color: "#bdbdbd", bold: true, italic: false, stroke: true, shadow: false },
+    previewStyle: {
+      color: "#bdbdbd",
+      _strokeColor: "#000000",
+      _strokeWidth: 2,
+      fontWeight: 700,
+      fontSize: "32px",
+    },
+  },
+  {
+    key: "bg_white",
+    label: "白底黑字",
+    style: { size: 28, color: "#1a1a1a", bold: true, italic: false, stroke: false, shadow: false },
+    previewStyle: {
+      color: "#1a1a1a",
+      fontWeight: 700,
+      fontSize: "32px",
+      background: "#ffffff",
+      borderRadius: "4px",
+      padding: "2px 6px",
+    },
+  },
+  {
+    key: "bg_yellow",
+    label: "黄底黑字",
+    style: { size: 28, color: "#1a1a1a", bold: true, italic: false, stroke: false, shadow: false },
+    previewStyle: {
+      color: "#1a1a1a",
+      fontWeight: 700,
+      fontSize: "32px",
+      background: "#ffd54f",
+      borderRadius: "4px",
+      padding: "2px 6px",
+    },
+  },
+  {
+    key: "bg_pink",
+    label: "粉底黑字",
+    style: { size: 28, color: "#1a1a1a", bold: true, italic: false, stroke: false, shadow: false },
+    previewStyle: {
+      color: "#1a1a1a",
+      fontWeight: 700,
+      fontSize: "32px",
+      background: "#ff80ab",
+      borderRadius: "4px",
+      padding: "2px 6px",
+    },
+  },
+  {
+    key: "bg_red",
+    label: "红底白字",
+    style: { size: 28, color: "#ffffff", bold: true, italic: false, stroke: false, shadow: false },
+    previewStyle: {
+      color: "#ffffff",
+      fontWeight: 700,
+      fontSize: "32px",
+      background: "#ef5350",
+      borderRadius: "4px",
+      padding: "2px 6px",
+    },
+  },
+  {
+    key: "neon_orange",
+    label: "橙色发光",
+    style: { size: 32, color: "#ff9100", bold: true, italic: false, stroke: false, shadow: true },
+    previewStyle: {
+      color: "#ff9100",
+      fontWeight: 700,
+      fontSize: "32px",
+      textShadow: "0 0 4px #ff9100, 0 0 8px #ff9100, 0 0 16px rgba(255,145,0,0.5)",
+    },
+  },
+  {
+    key: "neon_purple",
+    label: "紫色发光",
+    style: { size: 32, color: "#d500f9", bold: true, italic: false, stroke: false, shadow: true },
+    previewStyle: {
+      color: "#d500f9",
+      fontWeight: 700,
+      fontSize: "32px",
+      textShadow: "0 0 4px #d500f9, 0 0 8px #d500f9, 0 0 16px rgba(213,0,249,0.5)",
+    },
+  },
+  {
+    key: "bordered_white",
+    label: "白字绿框",
+    style: { size: 28, color: "#ffffff", bold: true, italic: false, stroke: false, shadow: false },
+    previewStyle: {
+      color: "#ffffff",
+      fontWeight: 700,
+      fontSize: "32px",
+      background: "#1a1a1a",
+      border: "2px solid #69f0ae",
+      borderRadius: "4px",
+      padding: "2px 6px",
+    },
+  },
+  {
+    key: "gradient_warm",
+    label: "暖色渐变",
+    style: { size: 32, color: "#ff6d00", bold: true, italic: false, stroke: false, shadow: true },
+    previewStyle: {
+      color: "#ff6d00",
+      fontWeight: 700,
+      fontSize: "32px",
+      textShadow: "0 0 6px rgba(255,109,0,0.6), 1px 1px 2px rgba(0,0,0,0.5)",
+    },
+  },
+  {
+    key: "gradient_cool",
+    label: "冷色渐变",
+    style: { size: 32, color: "#00b0ff", bold: true, italic: false, stroke: false, shadow: true },
+    previewStyle: {
+      color: "#00b0ff",
+      fontWeight: 700,
+      fontSize: "32px",
+      textShadow: "0 0 6px rgba(0,176,255,0.6), 1px 1px 2px rgba(0,0,0,0.5)",
+    },
+  },
+  {
+    key: "shadow_deep",
+    label: "深影白字",
+    style: { size: 28, color: "#ffffff", bold: true, italic: false, stroke: false, shadow: true },
+    previewStyle: {
+      color: "#ffffff",
+      fontWeight: 700,
+      fontSize: "32px",
+      textShadow: "2px 2px 4px rgba(0,0,0,0.8), 0 0 8px rgba(0,0,0,0.4)",
+    },
+  },
+  {
+    key: "soft_gold",
+    label: "柔光金",
+    style: { size: 28, color: "#ffd54f", bold: true, italic: false, stroke: false, shadow: true },
+    previewStyle: {
+      color: "#ffd54f",
+      fontWeight: 700,
+      fontSize: "32px",
+      textShadow: "0 0 6px rgba(255,213,79,0.5), 1px 1px 2px rgba(0,0,0,0.4)",
    },
  },
 ]
@@ -1733,50 +1733,40 @@
 /* 标题预设卡片网格 */
 .xx-title-presets-grid {
  display: grid;
-  grid-template-columns: repeat(4, 1fr);
-  gap: 10px;
+  grid-template-columns: repeat(6, 52px);
+  gap: 1px;
 }

 .xx-title-preset-card {
  display: flex;
-  flex-direction: column;
  align-items: center;
  justify-content: center;
-  padding: 14px 8px;
-  background: var(--bg-secondary);
+  width: 52px;
+  height: 52px;
+  padding: 0;
+  background: #404040;
  border: 2px solid transparent;
-  border-radius: var(--radius-sm);
+  border-radius: 8px;
  cursor: pointer;
  transition: all 0.15s;
-  text-align: center;
 }

 .xx-title-preset-card:hover {
-  border-color: var(--primary-200);
-  background: var(--bg-primary);
+  border-color: #666;
+  background: #4d4d4d;
 }

 .xx-title-preset-card.active {
-  border-color: var(--primary-color);
-  background: var(--primary-50);
+  border-color: #409eff;
+  background: #4d4d4d;
 }

 .xx-title-preset-preview-text {
-  line-height: 1.4;
-  margin-bottom: 6px;
+  font-size: 32px;
+  line-height: 1;
  user-select: none;
 }

-.xx-title-preset-card-label {
-  font-size: 11px;
-  color: var(--text-secondary);
-}
-
-.xx-title-preset-card.active .xx-title-preset-card-label {
-  color: var(--primary-color);
-  font-weight: 500;
-}
-
 /* 样式按钮组 */
 .xx-style-btns {
  display: flex;
@@ -1,12 +1,34 @@
 import { useState, useCallback } from "react"
 import { message } from "antd"
 import type { AssetItem } from "@/api/assets"
+import type { TemplateSegment } from "@/api/templates/types"
 import { smartMatchAssets, isAssetUsable } from "@/api/assets"

 interface UseSmartMatchOptions {
  libraryId: string
  materials: { items: AssetItem[]; total: number }
  onSmartSelectedIdsChange: (ids: string[]) => void
+  /** 当前模板的 segments，用于根据总时长计算 limit */
+  templateSegments?: TemplateSegment[]
+}
+
+/** 默认 limit（拿不到目标时长时的兜底上限） */
+const DEFAULT_LIMIT = 10
+/** 每个素材切片按 15 秒估算所需素材数 */
+const SECONDS_PER_ASSET = 15
+
+/**
+ * 根据模板 segments 计算所需素材数量上限。
+ * 取每个 segment 的 duration_min 之和作为目标视频总时长，
+ * 再按 15 秒/素材估算需要多少个素材，且保证不少于片段数（每个片段至少 1 个素材）；
+ * 结果钳制到 [1, 200] 区间（后端 limit 上限 200）。
+ */
+function computeLimitFromSegments(segments?: TemplateSegment[]): number {
+  if (!segments || segments.length === 0) return DEFAULT_LIMIT
+  const totalSeconds = segments.reduce((sum, seg) => sum + (seg.duration_min || 0), 0)
+  if (totalSeconds <= 0) return DEFAULT_LIMIT
+  const limit = Math.max(segments.length, Math.ceil(totalSeconds / SECONDS_PER_ASSET))
+  return Math.max(1, Math.min(limit, 200))
 }

 /**
@@ -18,6 +40,7 @@ export function useSmartMatch({
  libraryId,
  materials,
  onSmartSelectedIdsChange,
+  templateSegments,
 }: UseSmartMatchOptions) {
  const [smartMatching, setSmartMatching] = useState(false)
  const [hasMatched, setHasMatched] = useState(false)
@@ -42,12 +65,14 @@ export function useSmartMatch({
    setSmartMatching(true)

    try {
+      // 根据目标视频时长计算合理的素材数量上限，避免"有几个选几个"
+      const limit = computeLimitFromSegments(templateSegments)
+
      // 调用后端智能匹配 API（后端也会排除已用尽素材，这里前端兜底过滤）
-      // items 已在 API 层归一化（兼容后端 {asset, score} 包装结构）；
-      // 这里再过滤一遍无 id/已用尽项，杜绝 undefined id 流入预览链路
-      const result = await smartMatchAssets(libraryId)
+      const result = await smartMatchAssets(libraryId, limit)
+      // 兜底过滤：id 为空或不可用的素材不参与匹配（smartMatchAssets 已做归一化，这里双保险）
      const matched = (result.items ?? []).filter((a) => !!a?.id && isAssetUsable(a))
-      const matchedIds = matched.map((a) => a.id)
+      const matchedIds = matched.map((a: AssetItem) => a.id)

      if (matchedIds.length > 0) {
        onSmartSelectedIdsChange(matchedIds)
@@ -67,7 +92,7 @@ export function useSmartMatch({
    } finally {
      setSmartMatching(false)
    }
-  }, [libraryId, materials.items, onSmartSelectedIdsChange])
+  }, [libraryId, materials.items, onSmartSelectedIdsChange, templateSegments])

  /* ── 换一批 = 重新触发智能匹配 ── */
  const handleRefreshMatch = useCallback(async () => {
@@ -233,7 +233,7 @@ export function useCanvasPlayer(
    fontSize: number
    fontFamily: string
    color: string
-    position: "top" | "center" | "bottom"
+    position: "top" | "center" | "bottom" | "custom"
    bold?: boolean
    stroke?: boolean
    shadow?: boolean
@@ -27,6 +27,8 @@ const DEFAULT_TITLE_SETTINGS: TitleSettings = {
  stroke: true,
  shadow: false,
  color: "#ffffff",
+  posX: null,
+  posY: null,
 }

 export interface GenerateFormState {
@@ -117,6 +117,14 @@ export function useGenerateVideo(props: UseGenerateVideoProps) {
                  font_size: props.titleSettings.size,
                  font_color: props.titleSettings.color,
                  position: props.titleSettings.position,
+                  ...(props.titleSettings.position === "custom" &&
+                  props.titleSettings.posX != null &&
+                  props.titleSettings.posY != null
+                    ? {
+                        pos_x: Math.round(props.titleSettings.posX),
+                        pos_y: Math.round(props.titleSettings.posY),
+                      }
+                    : {}),
                  bold: props.titleSettings.bold,
                  stroke: props.titleSettings.stroke,
                  shadow: props.titleSettings.shadow,
@@ -51,6 +51,7 @@ export function useStep2Materials({
    libraryId: selectedLibraryId,
    materials: selectableMaterials,
    onSmartSelectedIdsChange,
+    templateSegments,
  })

  /* ── 自动触发智能匹配：选择视频库后自动调用 ── */
@@ -46,7 +46,16 @@ export function useTitleStyleUpdaters({

  const updatePosition = useCallback(
    (position: string) => {
-      onTitleSettingsChange({ ...titleSettings, position })
+      // 切回三档预设时清掉自定义坐标
+      onTitleSettingsChange({ ...titleSettings, position, posX: null, posY: null })
+    },
+    [titleSettings, onTitleSettingsChange],
+  )
+
+  /** 拖拽更新自定义位置（由预览播放器调用） */
+  const updateTitlePosition = useCallback(
+    (posX: number, posY: number) => {
+      onTitleSettingsChange({ ...titleSettings, position: "custom", posX, posY })
    },
    [titleSettings, onTitleSettingsChange],
  )
@@ -111,6 +120,7 @@ export function useTitleStyleUpdaters({
    updateTitle,
    toggleAiAutoSelect,
    updatePosition,
+    updateTitlePosition,
    updateFont,
    updateSize,
    updateColor,
@@ -16,6 +16,9 @@ export interface TitleSettings {
  stroke: boolean
  shadow: boolean
  color: string
+  /** 自由位置坐标（PlayRes 像素），仅当 position="custom" 时有效 */
+  posX: number | null
+  posY: number | null
 }

 /* ── 智能匹配结果 ── */
@@ -1,6 +1,6 @@
 /**
 * 成片库页面 — V21 设计系统
- * 卡片网格布局，支持视频播放/下载/分享、批量操作、筛选
+ * 卡片网格布局，支持视频内联播放/下载/分享、批量操作、筛选
 *
 * 主组件仅保留 Hook 组装与整体布局
 * 列表查询 → hooks/useProductList
@@ -8,15 +8,12 @@
 * 筛选栏 → components/ProductFilterBar
 * 批量操作栏 → components/ProductBatchBar
 * 空状态 → components/ProductEmptyState
- * 产品卡片 → components/ProductCard
- * 视频播放 → components/VideoPlayer
+ * 产品卡片 → components/ProductCard（内联视频播放）
 */
-import React, { useState } from "react"
+import React from "react"
 import { VideoCameraOutlined, DownloadOutlined } from "@ant-design/icons"
 import { Button } from "@/components/ui"
-import type { ProductItem } from "./types"
 import { ProductCard } from "./components/ProductCard"
-import { VideoPlayer } from "./components/VideoPlayer"
 import { ProductFilterBar } from "./components/ProductFilterBar"
 import { ProductBatchBar } from "./components/ProductBatchBar"
 import { ProductEmptyState } from "./components/ProductEmptyState"
@@ -53,13 +50,9 @@ const ProductLibrary: React.FC = () => {
    clearSelection,
  } = useProductList()

-  /* 播放器 */
-  const [playingProduct, setPlayingProduct] = useState<ProductItem | null>(null)
-
  const {
    handleDownload,
    handleShare,
-    handleViewDetail,
    handleDelete,
    handlePublish,
    handleReviewStatusChange,
@@ -71,7 +64,7 @@ const ProductLibrary: React.FC = () => {
    selectedIds,
    clearSelection,
    products,
-    setPlayingProduct,
+    setPlayingProduct: () => {}, // 不再使用弹窗播放
  })

  // ── Loading 状态 ──
@@ -146,7 +139,6 @@ const ProductLibrary: React.FC = () => {
              isSelected={selectedIds.has(product.id)}
              batchMode={batchMode}
              onToggleSelect={handleToggleSelect}
-              onPlay={setPlayingProduct}
              onDownload={handleDownload}
              onShare={handleShare}
              onDelete={handleDelete}
@@ -158,17 +150,6 @@ const ProductLibrary: React.FC = () => {
      ) : (
        <ProductEmptyState type="empty" />
      )}
-
-      {/* 视频播放弹窗 */}
-      {playingProduct && (
-        <VideoPlayer
-          product={playingProduct}
-          onClose={() => setPlayingProduct(null)}
-          onDownload={handleDownload}
-          onShare={handleShare}
-          onViewDetail={handleViewDetail}
-        />
-      )}
    </div>
  )
 }
@@ -1,4 +1,4 @@
-import React from "react"
+import React, { useState, useRef, useCallback, useEffect } from "react"
 import { Popconfirm } from "antd"
 import {
  CheckOutlined,
@@ -7,17 +7,17 @@ import {
  ShareAltOutlined,
  DeleteOutlined,
  CloudUploadOutlined,
+  VideoCameraOutlined,
 } from "@ant-design/icons"
 import type { ProductItem } from "../types"
 import { statusConfig, reviewStatusConfig } from "../constants"
-import { formatTime } from "../utils"
+import { formatTime, formatSize } from "../utils"

 interface ProductCardProps {
  product: ProductItem
  isSelected: boolean
  batchMode: boolean
  onToggleSelect: (id: string) => void
-  onPlay: (product: ProductItem) => void
  onDownload: (product: ProductItem) => void
  onShare: (product: ProductItem) => void
  onDelete: (id: string) => void
@@ -30,7 +30,6 @@ export const ProductCard: React.FC<ProductCardProps> = ({
  isSelected,
  batchMode,
  onToggleSelect,
-  onPlay,
  onDownload,
  onShare,
  onDelete,
@@ -38,17 +37,39 @@ export const ProductCard: React.FC<ProductCardProps> = ({
  onReviewStatusChange,
 }) => {
  const st = statusConfig[product.status]
+  const videoRef = useRef<HTMLVideoElement>(null)
+  const [isPlaying, setIsPlaying] = useState(false)
+  const [aspectRatio, setAspectRatio] = useState<string | null>(null)
+  const [hasVideo, setHasVideo] = useState(!!product.videoUrl)

  /** 查重率样式 */
  const dupClass =
    product.duplicateRate <= 5 ? "good" : product.duplicateRate <= 15 ? "warn" : "bad"

-  /** 点击卡片 */
-  const handleCardClick = () => {
-    if (batchMode) {
-      onToggleSelect(product.id)
+  /** 视频元数据加载后获取实际比例 */
+  const handleVideoLoaded = useCallback(() => {
+    const v = videoRef.current
+    if (v && v.videoWidth > 0 && v.videoHeight > 0) {
+      setAspectRatio(`${v.videoWidth} / ${v.videoHeight}`)
+    }
+  }, [])
+
+  /** 视频播放/暂停结束事件 */
+  const handlePlay = useCallback(() => setIsPlaying(true), [])
+  const handlePause = useCallback(() => setIsPlaying(false), [])
+
+  /** 视频出错时降级为缩略图 */
+  const handleVideoError = useCallback(() => setHasVideo(false), [])
+
+  /** 点击缩略图区域：有视频则内联播放，否则不响应 */
+  const handleThumbClick = () => {
+    if (batchMode) return
+    if (!hasVideo || !videoRef.current) return
+    const v = videoRef.current
+    if (v.paused) {
+      v.play()
    } else {
-      onPlay(product)
+      v.pause()
    }
  }

@@ -58,6 +79,24 @@ export const ProductCard: React.FC<ProductCardProps> = ({
    onToggleSelect(product.id)
  }

+  /** 卡片容器点击（非批量模式下不再触发弹窗播放） */
+  const handleCardClick = () => {
+    if (batchMode) {
+      onToggleSelect(product.id)
+    }
+  }
+
+  /** 组件卸载时暂停视频 */
+  useEffect(() => {
+    const v = videoRef.current
+    return () => {
+      v?.pause()
+    }
+  }, [])
+
+  /** 缩略图区域 style：有实际比例则用实际比例，否则默认 16:9 */
+  const thumbStyle: React.CSSProperties = aspectRatio ? { aspectRatio } : { aspectRatio: "16 / 9" }
+
  return (
    <div
      className={`xx-product-card${isSelected ? " selected" : ""}${product.isPublished ? " published" : ""}`}
@@ -102,21 +141,61 @@ export const ProductCard: React.FC<ProductCardProps> = ({
        </div>
      )}

-      {/* 缩略图 */}
-      <div className="xx-product-thumb">
-        {product.thumbnailUrl ? (
-          <img
-            className="xx-product-thumb-bg"
-            src={product.thumbnailUrl}
-            alt={product.name}
-            style={{ width: "100%", height: "100%", objectFit: "cover" }}
-          />
+      {/* 视频/缩略图区域 */}
+      <div className="xx-product-thumb" style={thumbStyle}>
+        {hasVideo ? (
+          <>
+            <video
+              ref={videoRef}
+              className="xx-product-thumb-video"
+              src={product.videoUrl}
+              preload="metadata"
+              onLoadedMetadata={handleVideoLoaded}
+              onPlay={handlePlay}
+              onPause={handlePause}
+              onEnded={handlePause}
+              onError={handleVideoError}
+              controls={isPlaying}
+              onClick={(e) => {
+                e.stopPropagation()
+                handleThumbClick()
+              }}
+            />
+            {/* 未播放时显示播放按钮覆盖层 */}
+            {!isPlaying && (
+              <div className="xx-product-play" onClick={handleThumbClick}>
+                <PlayCircleOutlined />
+              </div>
+            )}
+          </>
        ) : (
-          <div className="xx-product-thumb-bg" style={{ background: product.gradient }} />
+          <>
+            {product.thumbnailUrl ? (
+              <img
+                className="xx-product-thumb-bg"
+                src={product.thumbnailUrl}
+                alt={product.name}
+                style={{ width: "100%", height: "100%", objectFit: "cover" }}
+              />
+            ) : (
+              <div
+                className="xx-product-thumb-bg"
+                style={{
+                  background: product.gradient,
+                  display: "grid",
+                  placeItems: "center",
+                  color: "rgba(255,255,255,0.3)",
+                  fontSize: "48px",
+                }}
+              >
+                <VideoCameraOutlined />
+              </div>
+            )}
+            <div className="xx-product-play">
+              <PlayCircleOutlined />
+            </div>
+          </>
        )}
-        <div className="xx-product-play">
-          <PlayCircleOutlined />
-        </div>
        {product.duration > 0 && (
          <span className="xx-product-duration">{formatTime(product.duration)}</span>
        )}
@@ -131,11 +210,20 @@ export const ProductCard: React.FC<ProductCardProps> = ({
          <span className={`xx-product-status ${st.className}`}>{st.text}</span>
          <span className="xx-product-date">{product.date}</span>
        </div>
-        {product.duplicateRate > 0 && (
-          <span className={`xx-product-dup-rate ${dupClass}`}>
-            查重率：{product.duplicateRate.toFixed(1)}%
+        <div className="xx-product-meta-grid">
+          <span className="xx-product-meta-item">分辨率：{product.resolution || "-"}</span>
+          <span className="xx-product-meta-item">
+            时长：{product.duration > 0 ? formatTime(product.duration) : "-"}
          </span>
-        )}
+          <span className="xx-product-meta-item">大小：{formatSize(product.fileSize)}</span>
+          <span
+            className={`xx-product-meta-item xx-product-dup-rate${
+              product.duplicateRate > 0 ? ` ${dupClass}` : ""
+            }`}
+          >
+            查重率：{product.duplicateRate != null ? `${product.duplicateRate.toFixed(1)}%` : "-"}
+          </span>
+        </div>
      </div>

      {/* 操作按钮 */}
@@ -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 +45,26 @@ 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.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>
@@ -91,8 +91,6 @@
  height: 18px;
  border: 2px solid var(--border-color);
  border-radius: var(--radius-xs);
-  display: grid;
-  place-items: center;
  transition: var(--transition-all);
  background: var(--bg-primary);
  flex-shrink: 0;
@@ -137,7 +135,7 @@
   ============================================================ */
 .xx-products-grid {
  display: grid;
-  grid-template-columns: repeat(4, 1fr);
+  grid-template-columns: repeat(auto-fill, minmax(220px, 1fr));
  gap: var(--space-md);
 }

@@ -187,8 +185,6 @@
  height: 22px;
  border: 2px solid rgba(255, 255, 255, 0.8);
  border-radius: var(--radius-xs);
-  display: grid;
-  place-items: center;
  background: rgba(0, 0, 0, 0.3);
  backdrop-filter: blur(4px);
  cursor: pointer;
@@ -260,13 +256,10 @@

 /* 缩略图区域 */
 .xx-product-thumb {
-  aspect-ratio: 9 / 16;
-  max-height: 220px;
  position: relative;
  overflow: hidden;
-  display: grid;
-  place-items: center;
  color: var(--text-inverse);
+  background: var(--color-gray-950);
 }

 .xx-product-thumb-bg {
@@ -276,23 +269,41 @@
  background-position: center;
 }

+/* 内联视频播放器（cover 填满容器，竖屏视频不留左右空白） */
+.xx-product-thumb-video {
+  width: 100%;
+  height: 100%;
+  object-fit: cover;
+  display: block;
+}
+
+.xx-product-thumb-video[controls] {
+  cursor: pointer;
+}
+
 .xx-product-play {
-  position: relative;
-  z-index: 1;
-  width: 44px;
-  height: 44px;
-  border-radius: var(--radius-full);
+  position: absolute;
+  top: 50%;
+  left: 50%;
+  transform: translate(-50%, -50%);
+  z-index: 2;
+  width: 52px;
+  height: 52px;
+  border-radius: 50%;
  background: rgba(0, 0, 0, 0.45);
  backdrop-filter: blur(4px);
-  display: grid;
-  place-items: center;
-  font-size: var(--font-size-md);
-  transition: var(--transition-all);
+  display: flex;
+  align-items: center;
+  justify-content: center;
+  font-size: 28px;
+  color: #fff;
+  transition: all 0.2s;
+  cursor: pointer;
 }

 .xx-product-card:hover .xx-product-play {
  background: var(--primary-color);
-  transform: scale(1.1);
+  transform: translate(-50%, -50%) scale(1.1);
 }

 /* 时长标签 */
@@ -312,7 +323,7 @@

 /* 卡片信息区 */
 .xx-product-info {
-  padding: 14px;
+  padding: 8px 10px;
  display: flex;
  flex-direction: column;
  gap: var(--space-sm);
@@ -335,6 +346,21 @@
  gap: var(--space-xs);
 }

+/* 卡片信息网格：分辨率/时长 一行，大小/查重率 一行 */
+.xx-product-meta-grid {
+  display: grid;
+  grid-template-columns: 1fr 1fr;
+  gap: 4px var(--space-sm);
+}
+
+.xx-product-meta-item {
+  font-size: var(--font-size-xs);
+  color: var(--text-secondary);
+  white-space: nowrap;
+  overflow: hidden;
+  text-overflow: ellipsis;
+}
+
 /* 状态标签 */
 .xx-product-status {
  padding: 2px 10px;
@@ -479,8 +505,6 @@
  z-index: 1000;
  background: rgba(0, 0, 0, 0.75);
  backdrop-filter: blur(8px);
-  display: grid;
-  place-items: center;
  animation: player-fade-in 0.25s ease-out;
 }

@@ -519,8 +543,6 @@
  background: var(--color-gray-950);
  aspect-ratio: 9 / 16;
  max-height: 60vh;
-  display: grid;
-  place-items: center;
 }

 .xx-player-video-wrap video {
@@ -542,8 +564,6 @@
  backdrop-filter: blur(4px);
  color: var(--text-inverse);
  font-size: 28px;
-  display: grid;
-  place-items: center;
  cursor: pointer;
  border: none;
  transition: var(--transition-all);
@@ -640,8 +660,6 @@
  backdrop-filter: blur(4px);
  color: var(--text-inverse);
  font-size: var(--font-size-md);
-  display: grid;
-  place-items: center;
  cursor: pointer;
  border: none;
  transition: var(--transition-all);
@@ -1058,3 +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;
+}
@@ -1,6 +1,11 @@
 import { useMemo, useEffect } from "react"
 import { useQuery, useMutation, useQueryClient } from "@tanstack/react-query"
-import { getAssetsByKind, getAssetLibraries, createAssetLibrary } from "@/api/assets"
+import {
+  getAssetsByKind,
+  getAssetLibraries,
+  createAssetLibrary,
+  type AssetItem,
+} from "@/api/assets"
 import { type VoiceMaterial, mapAssetToMaterial } from "../../types"

 interface UseVoiceMaterialDataOptions {
@@ -44,6 +49,15 @@ export function useVoiceMaterialData({ keyword, gender, tagIds }: UseVoiceMateri
    queryKey: ["assets", "voice", { keyword, gender, tag_ids: tagIds }],
    queryFn: () => getAssetsByKind("voice", { keyword, gender, tag_ids: tagIds }),
    staleTime: 30_000,
+    // 列表中存在上传中/处理中素材时每 3s 轮询；全部就绪后自动停止
+    refetchInterval: (query) => {
+      const items = (query.state.data as AssetItem[] | undefined) ?? []
+      const processing = items.some((a) => {
+        const st = a.status ?? ""
+        return st === "uploading" || st === "ingesting" || st === "processing" || st === "pending"
+      })
+      return processing ? 3000 : false
+    },
  })

  const materials: VoiceMaterial[] = useMemo(() => assets.map(mapAssetToMaterial), [assets])
@@ -41,7 +41,7 @@ export const mapAssetToMaterial = (asset: AssetItem): VoiceMaterial => {
    tagIds: Array.isArray(asset.tag_ids) ? asset.tag_ids : [],
    fileName: asset.storage_key?.split("/").pop() || asset.name,
    fileSize: asset.file_size || 0,
-    duration: (meta.duration as number) || 0,
+    duration: asset.duration || (meta.duration as number) || 0,
    mimeType: asset.mime_type || "audio/mpeg",
    createdAt: asset.created_at || new Date().toISOString(),
    fileUrl: asset.file_url,
@@ -14,8 +14,14 @@
 * 弹窗集合 → components/VoiceModals
 * Toast 提示 → components/VoiceToasts
 */
-import React, { useCallback, useState } from "react"
-import { UploadOutlined, AudioOutlined, RobotOutlined } from "@ant-design/icons"
+import React, { useCallback, useEffect, useState } from "react"
+import { useSearchParams } from "react-router-dom"
+import {
+  UploadOutlined,
+  AudioOutlined,
+  RobotOutlined,
+  VideoCameraOutlined,
+} from "@ant-design/icons"
 import { Button } from "@/components/ui"
 import PageHead from "@/components/layout/PageHead"
 import { type AssetItem } from "@/api/assets"
@@ -34,6 +40,8 @@ import { useTtsSynthesize } from "./hooks/useTtsSynthesize"
 import { useVoiceUpload } from "./hooks/useVoiceUpload"
 import { useMaterialDelete } from "./hooks/useMaterialDelete"
 import { useMaterialBatchDelete } from "./hooks/useMaterialBatchDelete"
+import { useVideoExtract } from "./hooks/useVideoExtract"
+import VideoExtractModal from "./components/VideoExtractModal"
 import "./voices.css"

 let toastIdSeq = 0
@@ -158,6 +166,32 @@ const VoiceLibrary: React.FC = () => {
    handleUploadClose,
  } = useVoiceUpload({ showToast })

+  // ── 提取视频配音 ──────────────────────────────────────
+  const {
+    extractOpen,
+    extractFile,
+    extractProgress,
+    isExtracting,
+    setExtractOpen,
+    handleFileSelect: handleExtractFileSelect,
+    handleExtract,
+    handleExtractClose,
+  } = useVideoExtract({ showToast })
+
+  // ── URL 参数自动打开上传弹窗 ────────────────────────────
+  const [searchParams, setSearchParams] = useSearchParams()
+
+  useEffect(() => {
+    if (searchParams.get("upload") === "1") {
+      setActiveTab("material")
+      setUploadOpen(true)
+      // 一次性触发器：清理 upload 参数，避免切换 Tab 时重复触发
+      const next = new URLSearchParams(searchParams)
+      next.delete("upload")
+      setSearchParams(next, { replace: true })
+    }
+  }, [searchParams, setActiveTab, setUploadOpen, setSearchParams])
+
  // ── 切换 Tab 时停止播放 ───────────────────────────────
  const handleTabChange = useCallback(
    (tab: VoiceTabKey) => {
@@ -185,6 +219,14 @@ const VoiceLibrary: React.FC = () => {
      >
        上传音频
      </Button>
+      <Button
+        buttonType="primary"
+        buttonSize="sm"
+        icon={<VideoCameraOutlined />}
+        onClick={() => setExtractOpen(true)}
+      >
+        提取视频配音
+      </Button>
      <Button
        buttonType="ghost"
        buttonSize="sm"
@@ -285,7 +327,18 @@ const VoiceLibrary: React.FC = () => {
        />
      )}

-      {/* ── 弹窗集合 ──────────────────────────────────── */}
+      {/* ── 视频提取配音弹窗 ─────────────────────────────── */}
+      <VideoExtractModal
+        open={extractOpen}
+        file={extractFile}
+        progress={extractProgress}
+        isExtracting={isExtracting}
+        onClose={handleExtractClose}
+        onFileSelect={handleExtractFileSelect}
+        onExtract={handleExtract}
+      />
+
+      {/* ── 弹窗集合 ─────────────────────────────────── */}
      <VoiceModals
        cloneModalOpen={cloneModalOpen}
        onCloneClose={() => setCloneModalOpen(false)}
@@ -136,6 +136,9 @@ export const MaterialVoiceTab: React.FC<MaterialVoiceTabProps> = ({
            const material = mapAssetToMaterial(asset)
            // duration 优先取顶层（后端从 metadata 提取），兜底 metadata
            const cardDuration = asset.duration || material.duration || 0
+            // AI 生成素材标识：兼容旧素材（无 source 字段但有 tts_job_id）
+            const meta = asset.metadata as Record<string, unknown>
+            const isAiMaterial = meta?.source === "tts_job" || !!meta?.tts_job_id
            const isPlaying = playingId === asset.id
            const isSelected = selectedIds.has(asset.id)
            // 播放中以 audio 真实时长为准，未播放显示卡片时长
@@ -182,8 +185,11 @@ export const MaterialVoiceTab: React.FC<MaterialVoiceTabProps> = ({
                </div>

                <div className="xx-voice-info vmat-info">
-                  <div className="xx-voice-name" title={asset.name}>
-                    {asset.name}
+                  <div className="xx-voice-name-row">
+                    <div className="xx-voice-name" title={asset.name}>
+                      {asset.name}
+                    </div>
+                    {isAiMaterial && <span className="vmat-ai-badge">AI</span>}
                  </div>
                  <div className="xx-voice-subtitle">
                    {asset.file_size ? `${formatFileSize(asset.file_size)}` : "--"}
@@ -0,0 +1,207 @@
+import React, { useRef } from "react"
+import { Modal } from "antd"
+import { InboxOutlined, CloseOutlined } from "@ant-design/icons"
+
+interface VideoExtractModalProps {
+  open: boolean
+  file: File | null
+  progress: number | null
+  isExtracting: boolean
+  onClose: () => void
+  onFileSelect: (file: File | null) => void
+  onExtract: () => void
+}
+
+const ACCEPT_TYPES = ".mp4,.mov,.webm"
+
+const VideoExtractModal: React.FC<VideoExtractModalProps> = ({
+  open,
+  file,
+  progress,
+  isExtracting,
+  onClose,
+  onFileSelect,
+  onExtract,
+}) => {
+  const inputRef = useRef<HTMLInputElement>(null)
+
+  return (
+    <Modal
+      title={<span style={{ fontSize: 16, fontWeight: 600 }}>提取视频配音</span>}
+      open={open}
+      onCancel={() => {
+        if (isExtracting) return
+        onClose()
+      }}
+      footer={null}
+      width={480}
+      maskClosable={!isExtracting}
+    >
+      {!file ? (
+        <div
+          className="vmat-upload-dropzone"
+          onClick={() => inputRef.current?.click()}
+          style={{
+            border: "2px dashed #d9d9d9",
+            borderRadius: 8,
+            padding: "40px 20px",
+            textAlign: "center",
+            cursor: "pointer",
+            transition: "border-color 0.3s",
+          }}
+          onMouseEnter={(e) => (e.currentTarget.style.borderColor = "#7c3aed")}
+          onMouseLeave={(e) => (e.currentTarget.style.borderColor = "#d9d9d9")}
+        >
+          <InboxOutlined style={{ fontSize: 32, color: "#7c3aed", marginBottom: 12 }} />
+          <p style={{ margin: "0 0 8px", fontSize: 14, color: "#333" }}>点击选择视频文件</p>
+          <span style={{ fontSize: 12, color: "#999" }}>支持 MP4、MOV、WebM 格式</span>
+          <input
+            ref={inputRef}
+            type="file"
+            accept={ACCEPT_TYPES}
+            style={{ display: "none" }}
+            onChange={(e) => {
+              const f = e.target.files?.[0]
+              if (f) onFileSelect(f)
+            }}
+          />
+        </div>
+      ) : (
+        <div>
+          <div
+            style={{
+              display: "flex",
+              alignItems: "center",
+              justifyContent: "space-between",
+              padding: "12px 16px",
+              background: "#fafafa",
+              borderRadius: 8,
+              marginBottom: 16,
+            }}
+          >
+            <span
+              style={{
+                flex: 1,
+                overflow: "hidden",
+                textOverflow: "ellipsis",
+                whiteSpace: "nowrap",
+                fontSize: 14,
+                fontWeight: 500,
+              }}
+              title={file.name}
+            >
+              {file.name}
+            </span>
+            <span style={{ fontSize: 12, color: "#999", marginLeft: 8, flexShrink: 0 }}>
+              {(file.size / (1024 * 1024)).toFixed(1)} MB
+            </span>
+            {!isExtracting && (
+              <button
+                type="button"
+                onClick={() => {
+                  if (inputRef.current) inputRef.current.value = ""
+                  onFileSelect(null)
+                }}
+                style={{
+                  border: "none",
+                  background: "none",
+                  cursor: "pointer",
+                  color: "#999",
+                  marginLeft: 8,
+                  fontSize: 14,
+                }}
+                aria-label="移除文件"
+              >
+                <CloseOutlined />
+              </button>
+            )}
+          </div>
+
+          {progress !== null && (
+            <div style={{ marginBottom: 12 }}>
+              <div
+                style={{
+                  height: 6,
+                  background: "#f0f0f0",
+                  borderRadius: 3,
+                  overflow: "hidden",
+                }}
+              >
+                <div
+                  style={{
+                    height: "100%",
+                    width: `${progress}%`,
+                    background: "linear-gradient(90deg, #7c3aed, #a78bfa)",
+                    borderRadius: 3,
+                    transition: "width 0.3s",
+                  }}
+                />
+              </div>
+              <div
+                style={{
+                  textAlign: "right",
+                  fontSize: 12,
+                  color: "#999",
+                  marginTop: 4,
+                }}
+              >
+                {progress}%
+              </div>
+            </div>
+          )}
+
+          {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..."}
+            </p>
+          )}
+        </div>
+      )}
+
+      <div
+        style={{
+          display: "flex",
+          justifyContent: "flex-end",
+          gap: 8,
+          marginTop: 24,
+        }}
+      >
+        <button
+          type="button"
+          onClick={onClose}
+          disabled={isExtracting}
+          style={{
+            padding: "6px 16px",
+            borderRadius: 6,
+            border: "1px solid #d9d9d9",
+            background: "#fff",
+            cursor: isExtracting ? "not-allowed" : "pointer",
+            fontSize: 14,
+            opacity: isExtracting ? 0.5 : 1,
+          }}
+        >
+          取消
+        </button>
+        <button
+          type="button"
+          onClick={onExtract}
+          disabled={!file || isExtracting}
+          style={{
+            padding: "6px 16px",
+            borderRadius: 6,
+            border: "none",
+            background: !file || isExtracting ? "#d9d9d9" : "#7c3aed",
+            color: "#fff",
+            cursor: !file || isExtracting ? "not-allowed" : "pointer",
+            fontSize: 14,
+            fontWeight: 500,
+          }}
+        >
+          {isExtracting ? "提取中..." : "开始提取"}
+        </button>
+      </div>
+    </Modal>
+  )
+}
+
+export default VideoExtractModal
@@ -0,0 +1,74 @@
+import { useState, useCallback } from "react"
+import { useQueryClient } from "@tanstack/react-query"
+import { extractVideoVoice } from "@/api/tts"
+
+/**
+ * 视频提取配音 Hook
+ * 封装视频上传弹窗状态、提取进度、提取 mutation 逻辑
+ */
+interface UseVideoExtractProps {
+  showToast: (message: string, type: "success" | "error") => void
+}
+
+export function useVideoExtract({ showToast }: UseVideoExtractProps) {
+  const queryClient = useQueryClient()
+
+  const [extractOpen, setExtractOpen] = useState(false)
+  const [extractFile, setExtractFile] = useState<File | null>(null)
+  const [extractProgress, setExtractProgress] = useState<number | null>(null)
+  const [isExtracting, setIsExtracting] = useState(false)
+
+  const handleExtractClose = useCallback(() => {
+    setExtractOpen(false)
+    setExtractFile(null)
+    setExtractProgress(null)
+    setIsExtracting(false)
+  }, [])
+
+  const handleExtract = useCallback(async () => {
+    if (!extractFile) return
+    setIsExtracting(true)
+    setExtractProgress(0)
+    try {
+      await extractVideoVoice(extractFile, (p) => setExtractProgress(p))
+      // 刷新素材列表
+      queryClient.invalidateQueries({ queryKey: ["assets", "voice"] })
+      queryClient.invalidateQueries({ queryKey: ["voice-materials"] })
+      showToast("视频配音提取成功", "success")
+      handleExtractClose()
+    } catch (err: unknown) {
+      const msg = err instanceof Error ? err.message : "提取失败，请重试"
+      showToast(msg, "error")
+    } finally {
+      setIsExtracting(false)
+      setExtractProgress(null)
+    }
+  }, [extractFile, queryClient, showToast, handleExtractClose])
+
+  const handleFileSelect = useCallback(
+    (file: File | null) => {
+      if (!file) {
+        setExtractFile(null)
+        return
+      }
+      const validTypes = ["video/mp4", "video/quicktime", "video/webm"]
+      if (!validTypes.includes(file.type)) {
+        showToast("仅支持 MP4、MOV、WebM 格式的视频文件", "error")
+        return
+      }
+      setExtractFile(file)
+    },
+    [showToast],
+  )
+
+  return {
+    extractOpen,
+    setExtractOpen,
+    extractFile,
+    extractProgress,
+    isExtracting,
+    handleFileSelect,
+    handleExtract,
+    handleExtractClose,
+  }
+}
@@ -193,6 +193,24 @@
  overflow: hidden;
  text-overflow: ellipsis;
  flex: 1;
+  display: flex;
+  align-items: center;
+}
+
+/* AI 配音标识 */
+.vmat-ai-badge {
+  display: inline-block;
+  margin-left: 6px;
+  padding: 1px 6px;
+  font-size: 11px;
+  font-weight: 600;
+  color: #7c3aed;
+  background: #f3f0ff;
+  border: 1px solid #ddd6fe;
+  border-radius: 4px;
+  line-height: 16px;
+  vertical-align: middle;
+  flex-shrink: 0;
 }

 .xx-voice-star {
@@ -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")
+  })
+})
@@ -1,11 +1,17 @@
-"""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
+from dataclasses import dataclass, field
 from typing import Optional
+from uuid import uuid4

 import cv2
 import numpy as np
@@ -15,10 +21,34 @@ from worker_app.celery_app import celery_app
 from worker_app.db import SessionLocal

 from packages.adapters.sqlalchemy_impl.generated_video_repository import SQLAlchemyGeneratedVideoRepository
+from packages.adapters.sqlalchemy_impl.models import VideoFingerprintChunkModel
 from packages.shared.storage import get_storage_service

 logger = logging.getLogger(__name__)

+# ── 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:
    """计算图像的感知哈希（pHash），基于 DCT（离散余弦变换）。
@@ -80,6 +110,131 @@ def compute_color_histogram(image: np.ndarray, bins: int = 32) -> list[float]:
    return hist


+# ── 关键帧检测 ──────────────────────────────────────────────────
+
+
+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）
+    """
+    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
+class FingerprintChunk:
+    """单个分片指纹数据。"""
+
+    start_time_ms: int
+    end_time_ms: int
+    phash_binary: str
+    color_histogram: list[float]
+    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."""
@@ -89,6 +244,7 @@ class VideoFingerprint:
    color_histograms: list[list[float]]
    duration: float
    resolution: tuple[int, int]
+    chunks: list[FingerprintChunk] = field(default_factory=list)

    def to_dict(self) -> dict:
        # 注意：color_histograms 里的值可能是 np.float32（来自 cv2.normalize），
@@ -101,17 +257,182 @@ class VideoFingerprint:
            "color_histograms": native_histograms,
            "duration": float(self.duration),
            "resolution": [int(self.resolution[0]), int(self.resolution[1])],
+            "chunks": [
+                {
+                    "start_time_ms": c.start_time_ms,
+                    "end_time_ms": c.end_time_ms,
+                    "phash_binary": c.phash_binary,
+                    "color_histogram": [float(v) for v in c.color_histogram],
+                    "frame_count": c.frame_count,
+                }
+                for c in self.chunks
+            ],
        }

+    def to_chunk_models(self, video_id: str, project_id: str, user_id: str = "") -> list[VideoFingerprintChunkModel]:
+        """将分片数据转为 SQLAlchemy Model 列表，用于批量写入 video_fingerprint_chunks 表。"""
+        models = []
+        for chunk in self.chunks:
+            models.append(
+                VideoFingerprintChunkModel(
+                    id=uuid4().hex,
+                    video_id=video_id,
+                    project_id=project_id,
+                    user_id=user_id,
+                    start_time_ms=chunk.start_time_ms,
+                    end_time_ms=chunk.end_time_ms,
+                    phash_binary=chunk.phash_binary,
+                    color_histogram=[float(v) for v in chunk.color_histogram],
+                    frame_count=chunk.frame_count,
+                )
+            )
+        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."""
+        """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():
            raise RuntimeError(f"Cannot open video: {video_path}")
@@ -122,55 +443,175 @@ class VideoDeduplicator:
        width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
        height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))

-        md5_hash = hashlib.md5(usedforsecurity=False)
-        keyframe_phashes = []
-        color_histograms = []
+        cap.release()

-        frame_interval = max(1, frame_count // 10)
-        for i in range(0, frame_count, frame_interval):
-            cap.set(cv2.CAP_PROP_POS_FRAMES, i)
+        # 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] = []
+
+        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 not ret:
                continue

+            # MD5 计算
            _, buffer = cv2.imencode(".jpg", frame)
            md5_hash.update(buffer)

-            keyframe_phashes.append(compute_phash(frame))
-            color_histograms.append(compute_color_histogram(frame))
+            phash = compute_phash(frame)
+            hist = compute_color_histogram(frame)
+
+            # 计算分片时间范围（从前一个关键帧到下一个关键帧的中点）
+            prev_boundary = keyframe_times[i - 1] * 1000 if i > 0 else 0
+            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,
+                )
+            )

        cap.release()

+        # 向后兼容：聚合 keyframe_phashes / color_histograms
+        keyframe_phashes = [c.phash_binary for c in chunks]
+        color_histograms = [c.color_histogram for c in chunks]
+
        return VideoFingerprint(
            md5=md5_hash.hexdigest(),
            keyframe_phashes=keyframe_phashes,
            color_histograms=color_histograms,
            duration=duration,
            resolution=(width, height),
+            chunks=chunks,
        )

-    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。
+    def _get_existing_chunks(self, video_id: str, session: Session) -> list[dict]:
+        """从 video_fingerprint_chunks 表读取分片数据。返回空列表表示无分片数据。"""
+        rows = (
+            session.query(VideoFingerprintChunkModel)
+            .filter(VideoFingerprintChunkModel.video_id == video_id)
+            .order_by(VideoFingerprintChunkModel.start_time_ms)
+            .all()
+        )
+        return [
+            {
+                "phash_binary": r.phash_binary,
+                "color_histogram": r.color_histogram,
+                "start_time_ms": r.start_time_ms,
+                "end_time_ms": r.end_time_ms,
+            }
+            for r in rows
+        ]

-        判定逻辑（按优先级）：
-        1. MD5 精确匹配：完全一致则 similarity=1.0，立即返回
-        2. pHash 相似度：计算新视频每帧 phash 与已有视频每帧 phash 的最小汉明距离，
-           取所有帧的平均值 avg_distance。若 avg_distance < PHASH_THRESHOLD(10)，
-           则判定为重复，similarity = 1.0 - (avg_distance / 64)
+    @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)))

-        注意：返回第一个通过阈值的匹配（非最优匹配）。
+    @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 中位数距离 + 帧匹配比例 + 直方图融合判定
+
+        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:
@@ -182,28 +623,71 @@ class VideoDeduplicator:
            if fingerprint.md5 == ef.get("md5"):
                return {"duplicate": True, "duplicate_of": existing.id, "reason": "exact_md5_match", "similarity": 1.0}

-            # 感知哈希相似度
-            existing_phashes = ef.get("keyframe_phashes", [])
+            # 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)
+            if chunk_data:
+                existing_phashes = [c["phash_binary"] for c in chunk_data]
+            else:
+                # 回退：从 JSON 字段读取（存量旧视频）
+                existing_phashes = ef.get("keyframe_phashes", [])
+
            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
@@ -214,16 +698,22 @@ 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
@@ -247,7 +737,14 @@ class VideoDeduplicator:
                    "similarity": 1.0,
                }

-            existing_phashes = ef.get("keyframe_phashes", [])
+            # 优先从分片表读取
+            existing_phashes = []
+            chunk_data = self._get_existing_chunks(existing.id, session)
+            if chunk_data:
+                existing_phashes = [c["phash_binary"] for c in chunk_data]
+            else:
+                existing_phashes = ef.get("keyframe_phashes", [])
+
            if not existing_phashes:
                continue

@@ -255,58 +752,214 @@ 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。
+    def compute_duplicate_rate(
+        self,
+        fingerprint: VideoFingerprint,
+        project_id: str,
+        current_video_id: str | None,
+        session: Session,
+        *,
+        scope: str = "project",
+        user_id: str = "",
+    ) -> dict:
+        """计算当前视频与已有视频的查重率百分比。

-        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。
+        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:
-            histograms_a: 第一组直方图（每帧一个 list）
-            histograms_b: 第二组直方图
+            fingerprint: 当前视频的指纹
+            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
+            scope: "project" 项目内（默认），"user" 跨项目全局
+            user_id: 用户 ID（scope="user" 时使用）

        Returns:
-            平均余弦相似度，范围 [0, 1]
+            {
+                "duplicate_rate": float,    # 0~100
+                "visual_similarity": float,  # 0~1
+                "match_count": int,          # 判定为重复的视频数
+            }
        """
-        if not histograms_a or not histograms_b:
-            return 0.0
+        video_repo = SQLAlchemyGeneratedVideoRepository(session)

-        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:
+        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
+            if not existing.video_fingerprint:
                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
+            ef = existing.video_fingerprint
+
+            # MD5 精确匹配 → 100%
+            if fingerprint.md5 == ef.get("md5"):
+                return {
+                    "duplicate_rate": 100.0,
+                    "visual_similarity": 1.0,
+                    "match_count": 1,
+                }
+
+            # 优先从分片表读取
+            existing_phashes = []
+            chunk_data = self._get_existing_chunks(existing.id, session)
+            if chunk_data:
+                existing_phashes = [c["phash_binary"] for c in chunk_data]
+            else:
+                existing_phashes = ef.get("keyframe_phashes", [])
+
+            if not existing_phashes or not fingerprint.keyframe_phashes:
+                continue
+
+            min_distances = []
+            for phash in fingerprint.keyframe_phashes:
+                distances = [hamming_distance(phash, ep) for ep in existing_phashes]
+                min_distances.append(min(distances))
+
+            # 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(
+    fingerprint: VideoFingerprint,
+    video_id: str,
+    project_id: str,
+    user_id: str,
+    session: Session,
+) -> None:
+    """将指纹分片数据批量写入 video_fingerprint_chunks 表。幂等：已有数据时跳过。"""
+    # 幂等检查：已有分片数据则跳过
+    existing_count = (
+        session.query(VideoFingerprintChunkModel).filter(VideoFingerprintChunkModel.video_id == video_id).count()
+    )
+    if existing_count > 0:
+        logger.debug("Fingerprint chunks already exist for video %s (%d chunks), skipping", video_id, existing_count)
+        return
+
+    if not fingerprint.chunks:
+        logger.warning("No chunks in fingerprint for video %s, skipping chunk save", video_id)
+        return
+
+    chunk_models = fingerprint.to_chunk_models(video_id, project_id, user_id)
+    session.bulk_save_objects(chunk_models)
+    logger.info("Saved %d fingerprint chunks for video %s", len(chunk_models), video_id)


@celery_app.task(bind=True, max_retries=3, name="worker.check_duplicate")
@@ -331,7 +984,15 @@ 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:
@@ -341,7 +1002,24 @@ 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
+        _save_fingerprint_chunks(fingerprint, generated_video_id, video.project_id, video.user_id, session)
+
        session.commit()

        logger.info(f"Duplicate check completed for video {generated_video_id}: is_duplicate={video.is_duplicate}")
@@ -351,6 +1029,9 @@ 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:
@@ -100,8 +100,24 @@ def create_video_record_and_dedup(

        generated_video.video_fingerprint = fingerprint.to_dict()

-        # (a) 历史成片查重
-        duplicate_result = deduplicator.check_duplicate(fingerprint, project_id, session)
+        # 写入分片指纹表
+        from video_processing.dedup import _save_fingerprint_chunks
+
+        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（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,
+            session,
+            scope="user",
+            user_id=user_id,
+            duration_sec=duration_sec,
+        )

        # (b) 批次内查重（仅当有 batch_id 时）
        if not duplicate_result and batch_id:
@@ -121,6 +137,30 @@ def create_video_record_and_dedup(
            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
+
        video_repo.update(generated_video)
        session.commit()
        logger.info(
@@ -200,6 +200,22 @@ class UnifiedRenderService:

        # 3. 计算视频总时长（用于字幕显示时长）
        video_duration = 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)
+            clip_details = [
+                f"{c.clip_id}(dur={c.duration:.3f},actual={c.actual_duration:.3f},speed={getattr(c, 'playback_speed', 1.0):.4f})"
+                for c in layer.clips
+            ]
+            logger.info(
+                "[debug] layer=%s clips=%d total=%.3f transition_duration=%.3f details=%s",
+                layer.role,
+                len(layer.clips),
+                layer_total,
+                self.transition_duration,
+                ", ".join(clip_details),
+            )
+        logger.info("[debug] estimated video_duration=%.3f", video_duration)

        # 3.5 TTS 配音生成（如果配置了）
        self._maybe_add_voiceover_layer(layers, video_duration=video_duration)
@@ -1422,6 +1438,7 @@ class UnifiedRenderService:
            if trim_segments and len(trim_segments) > 1:
                # 多段裁剪：展开为多个 clip
                resolved_segments = TrimEngine.resolve_segments(trim_segments, actual_duration)
+                configured_speed = getattr(clip, "playback_speed", 1.0) or 1.0
                for i, seg in enumerate(resolved_segments):
                    # 每个段生成一个独立的 ResolvedClip
                    seg_clip_id = f"{clip.id}_seg_{seg.segment_id}"
@@ -1429,6 +1446,19 @@ class UnifiedRenderService:
                    seg_start = seg.trim.start_time
                    seg_duration = seg.trim.duration

+                    # 多段裁剪：如果段的时长超过素材实际时长，减速补偿
+                    seg_speed = configured_speed
+                    if actual_duration > 0 and seg_duration > actual_duration + 0.05:
+                        seg_speed = max(0.25, round(configured_speed * actual_duration / seg_duration, 4))
+                        logger.info(
+                            "[debug] multi-seg clip=%s seg=%s duration=%.3f actual=%.3f → speed=%.4f",
+                            clip.id,
+                            seg.segment_id,
+                            seg_duration,
+                            actual_duration,
+                            seg_speed,
+                        )
+
                    rc = ResolvedClip(
                        clip_id=seg_clip_id,
                        asset_id=asset_id,
@@ -1439,7 +1469,7 @@ class UnifiedRenderService:
                        duration=seg_duration,
                        transition_effect=clip.transition_effect or "cut",
                        transition_duration=getattr(clip, "transition_duration", 0.0) or 0.0,
-                        playback_speed=getattr(clip, "playback_speed", 1.0) or 1.0,
+                        playback_speed=seg_speed,
                        config={**clip_config, "_segment_id": seg.segment_id},
                        actual_duration=actual_duration,
                        trim_config=seg.trim,
@@ -1461,6 +1491,7 @@ class UnifiedRenderService:
            effective_trim: TrimConfig | None = None
            final_start = clip.start_time
            final_duration = clip.duration
+            configured_speed = getattr(clip, "playback_speed", 1.0) or 1.0

            if trim_config is not None and actual_duration > 0:
                effective_trim = trim_config.validate_and_resolve(actual_duration)
@@ -1474,6 +1505,25 @@ class UnifiedRenderService:
                    final_start = 0.0
                    final_duration = actual_duration

+            # 素材实际时长不足以覆盖配置的时长时，降低播放速度来补偿
+            # 例如：配置4s但素材只有3s → speed=0.75x，用满3s素材达到4s输出
+            if actual_duration > 0 and final_duration > actual_duration + 0.05:
+                compensated_speed = actual_duration / final_duration
+                # 保留用户设置的速度（如果已减速则叠加）
+                final_speed = configured_speed * compensated_speed
+                # 下限 0.25x
+                final_speed = max(0.25, round(final_speed, 4))
+                logger.info(
+                    "[debug] clip=%s duration=%.3f actual=%.3f → 减速补偿 speed=%.4f (configured=%.3f)",
+                    clip.id,
+                    final_duration,
+                    actual_duration,
+                    final_speed,
+                    configured_speed,
+                )
+            else:
+                final_speed = configured_speed
+
            rc = ResolvedClip(
                clip_id=clip.id,
                asset_id=asset_id,
@@ -1484,13 +1534,25 @@ class UnifiedRenderService:
                duration=final_duration,
                transition_effect=clip.transition_effect or "cut",
                transition_duration=getattr(clip, "transition_duration", 0.0) or 0.0,
-                playback_speed=getattr(clip, "playback_speed", 1.0) or 1.0,
+                playback_speed=final_speed,
                config=clip_config,
                actual_duration=actual_duration,
                trim_config=effective_trim,
            )
            resolved.append(rc)

+            # Debug日志：记录每个clip的时长信息
+            eff_dur = _clip_effective_duration_pure(final_duration, actual_duration)
+            logger.info(
+                "[debug] resolved clip=%s duration=%.3f actual=%.3f effective=%.3f speed=%.4f start=%.3f",
+                clip.id,
+                final_duration,
+                actual_duration,
+                eff_dur,
+                final_speed,
+                final_start,
+            )
+
        # 按 order 排序
        resolved.sort(key=lambda c: c.order)
        return resolved
@@ -1683,7 +1745,7 @@ class UnifiedRenderService:
                        if d > 0:
                            layer_dur = d
                            break
-                    xfade_filter, _ = self._transition_engine.build_xfade_chain(
+                    xfade_filter, xfade_estimated_dur = self._transition_engine.build_xfade_chain(
                        clip_durations=layer_durations,
                        clip_video_labels=layer_labels,
                        transitions=layer_transitions,
@@ -1692,6 +1754,13 @@ class UnifiedRenderService:
                    )
                    if xfade_filter:
                        filter_parts.append(xfade_filter)
+                    logger.info(
+                        "[unified-render] layer=%s xfade: clips=%d durations=%s estimated_dur=%.3f",
+                        layer.role,
+                        len(layer_labels),
+                        [round(d, 3) for d in layer_durations],
+                        xfade_estimated_dur,
+                    )
                layer_output_labels[layer.role] = out_label

        # Step 3: 合成各层
@@ -1915,8 +1984,14 @@ class UnifiedRenderService:
    def _clip_effective_duration(clip: ResolvedClip) -> float:
        """计算 clip 的有效时长（原速 trim 后时长）。

+        如果 playback_speed < 1（为补偿素材不足而减速），返回配置的 duration，
+        而非 min(duration, actual_duration)。
        实际实现移至 packages.domain.render_layer_utils.clip_effective_duration。
        """
+        speed = getattr(clip, "playback_speed", 1.0) or 1.0
+        # 减速场景：duration 已通过降低 playback_speed 补偿，返回配置的 duration
+        if speed < 1.0 - 1e-6 and clip.duration > 0:
+            return clip.duration
        return _clip_effective_duration_pure(clip.duration, clip.actual_duration)

    # ── 画中画（PiP）相关方法 ──────────────────────────────────────────────────
@@ -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",
@@ -0,0 +1,196 @@
+"""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()
+            # 本次是最后一次执行机会（retries 从 0 计数，达到 max_retries 说明重试已耗尽），
+            # 标记 failed；否则保持 pending 由 Celery 60 秒后重试
+            try:
+                if "repo" in locals() and self.request.retries >= self.max_retries:
+                    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()
+        raise self.retry(exc=e, countdown=60) from 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)
@@ -630,7 +630,7 @@ def ingest_asset(job_id: str) -> dict:
                name=filename,
                storage_key=job.storage_key,
                mime_type=mime_type,
-                metadata={"ingest_error": error_reason},
+                metadata={"source": "upload", "ingest_error": error_reason},
                file_size=int(metadata.get("size_bytes", 0)),
                duration=float(metadata.get("duration", 0)),
                width=int(metadata.get("width", 0)),
@@ -656,24 +656,57 @@ def ingest_asset(job_id: str) -> dict:
                "error": error_reason,
            }

-        # Create Asset
-        asset = Asset.create(
-            project_id=job.project_id,
-            library_id=job.library_id,
-            name=filename,
-            storage_key=job.storage_key,
-            mime_type=mime_type,
-            metadata=metadata,
-            file_size=int(metadata.get("size_bytes", 0)),
-            duration=float(metadata.get("duration", 0)),
-            width=int(metadata.get("width", 0)),
-            height=int(metadata.get("height", 0)),
-            codec=metadata.get("codec") or None,
-            status=AssetStatus.READY,
-            file_hash=job.file_hash,
-            thumbnail_url=thumbnail_url,
-        )
-        asset_repo.create(asset)
+        # 查找已存在的 Asset 记录（由 API 端在上传完成时立即创建为 PROCESSING 状态）
+        existing_asset = None
+        try:
+            existing_asset = asset_repo.find_by_storage_key(job.storage_key)
+        except Exception:
+            logger.warning("find_by_storage_key not available, trying fallback lookup")
+
+        if existing_asset is None:
+            # 兜底：如果 API 端没有预先创建 Asset（旧版本兼容），则创建新记录
+            logger.info("No pre-created asset found for storage_key=%s, creating new", job.storage_key)
+            metadata["source"] = "upload"
+            asset = Asset.create(
+                project_id=job.project_id,
+                library_id=job.library_id,
+                name=filename,
+                storage_key=job.storage_key,
+                mime_type=mime_type,
+                metadata=metadata,
+                file_size=int(metadata.get("size_bytes", 0)),
+                duration=float(metadata.get("duration", 0)),
+                width=int(metadata.get("width", 0)),
+                height=int(metadata.get("height", 0)),
+                codec=metadata.get("codec") or None,
+                status=AssetStatus.READY,
+                file_hash=job.file_hash,
+                thumbnail_url=thumbnail_url,
+            )
+            asset_repo.create(asset)
+        else:
+            # 更新已有的 Asset 记录，补充元数据并将状态改为 READY
+            asset = existing_asset
+            asset.mime_type = mime_type
+            metadata["source"] = "upload"
+            asset.metadata = metadata
+            asset.file_size = int(metadata.get("size_bytes", 0))
+            asset.duration = float(metadata.get("duration", 0))
+            asset.width = int(metadata.get("width", 0))
+            asset.height = int(metadata.get("height", 0))
+            codec_val = metadata.get("codec")
+            if codec_val:
+                asset.codec = str(codec_val)
+            fps_val = metadata.get("fps")
+            if fps_val:
+                try:
+                    asset.fps = float(fps_val)
+                except (ValueError, TypeError):
+                    pass
+            asset.status = AssetStatus.READY
+            asset.thumbnail_url = thumbnail_url
+            asset.updated_at = datetime.now(timezone.utc)
+            asset_repo.update(asset)

        # Update job status to COMPLETED
        job.status = IngestJobStatus.COMPLETED
@@ -692,15 +725,37 @@ def ingest_asset(job_id: str) -> dict:
        db.rollback()
        logger.error(f"Failed to ingest asset {job_id}: {e}")

-        # Update job status to FAILED
+        # Update job status to FAILED and mark pre-created Asset as ERROR
        try:
            job_repo = SQLAlchemyIngestJobRepository(db)
+            asset_repo = SQLAlchemyAssetRepository(db)
            job = job_repo.get(job_id)
            if job:
                job.status = IngestJobStatus.FAILED
                job.error_message = str(e)
                job.updated_at = datetime.now(timezone.utc)
                job_repo.update(job)
+
+                # 将上传时创建的占位 Asset（PROCESSING/UPLOADING）标记为 ERROR，
+                # 避免素材永远卡在中间状态
+                try:
+                    existing = asset_repo.find_by_storage_key(job.storage_key)
+                    if existing and existing.status in (
+                        AssetStatus.PROCESSING,
+                        AssetStatus.UPLOADING,
+                    ):
+                        existing.status = AssetStatus.ERROR
+                        existing.metadata = {**(existing.metadata or {}), "ingest_error": str(e)}
+                        existing.updated_at = datetime.now(timezone.utc)
+                        asset_repo.update(existing)
+                        logger.info(
+                            "Marked asset as ERROR due to ingest failure: asset_id=%s job_id=%s",
+                            existing.id,
+                            job_id,
+                        )
+                except Exception as asset_err:
+                    logger.warning("Failed to mark asset as ERROR: %s", asset_err)
+
                db.commit()
        except Exception:
            db.rollback()
@@ -0,0 +1,216 @@
+# ============================================================
+# 小虾 SaaS — Production 环境配置模板
+# ============================================================
+# 使用方式：复制为 /var/lib/xiaoxia-saas-production/.env 并填入实际密钥
+# 敏感值标记为 ${PLACEHOLDER}，部署前必须替换为真实值
+# ============================================================
+
+
+# ==================== 应用基本配置 ====================
+
+# 应用名称
+APP_NAME=xiaoxia-saas
+
+# 环境标识
+APP_ENV=production
+
+# 关闭 Debug 模式
+DEBUG=false
+
+# 应用基础 URL（前端页面地址）
+APP_BASE_URL=https://xiaoxiajianji.com
+
+# 对外公开的 API 基础 URL（用于生成回调链接等）
+PUBLIC_API_BASE_URL=https://api.xiaoxiajianji.com
+
+# API 服务监听地址
+API_HOST=0.0.0.0
+
+# API 服务监听端口
+API_PORT=8001
+
+# 生产环境关闭自动建表，使用 alembic migration
+AUTO_CREATE_SCHEMA=false
+
+
+# ==================== 数据库配置 ====================
+
+# 数据库连接串（格式：postgresql+psycopg://user:password@host:port/dbname）
+# ${DATABASE_URL} — 替换为实际的 Production PostgreSQL 连接串
+DATABASE_URL=${DATABASE_URL}
+
+# 连接池大小（常驻连接数）
+DATABASE_POOL_SIZE=20
+
+# 连接池最大溢出连接数（pool_size + max_overflow = 最大并发连接数）
+DATABASE_MAX_OVERFLOW=10
+
+# 获取连接超时时间（秒）
+DATABASE_POOL_TIMEOUT=30
+
+# 连接回收时间（秒），防止数据库端主动断开导致的死连接
+DATABASE_POOL_RECYCLE=3600
+
+# 不使用内存数据库
+USE_IN_MEMORY_DB=false
+
+
+# ==================== Redis 配置 ====================
+
+# Redis 连接 URL（格式：redis://[:password@]host:port/db）
+# ${REDIS_URL} — 替换为实际的 Production Redis 连接串
+REDIS_URL=${REDIS_URL}
+
+# 启用 Redis Session 存储（多实例部署必须开启）
+ENABLE_REDIS_SESSIONS=true
+
+
+# ==================== Celery 任务队列 ====================
+
+# Celery Broker（任务分发），使用 Redis db0
+CELERY_BROKER_URL=${CELERY_BROKER_URL}
+
+# Celery Result Backend（任务结果存储），使用 Redis db1
+CELERY_RESULT_BACKEND=${CELERY_RESULT_BACKEND}
+
+
+# ==================== Worker 配置 ====================
+
+# Worker 进程名称
+WORKER_NAME=xiaoxia-saas-worker
+
+# Worker 并发数（同时执行的任务数）
+WORKER_CONCURRENCY=4
+
+# 每个子进程最多处理多少任务后重启（防止内存泄漏）
+WORKER_MAX_TASKS_PER_CHILD=1000
+
+
+# ==================== JWT 认证配置 ====================
+
+# JWT 签名密钥 — 必须设置为强随机字符串（至少32字符）
+# ${JWT_SECRET_KEY} — 替换为实际的随机密钥
+JWT_SECRET_KEY=${JWT_SECRET_KEY}
+
+# JWT 签名算法
+JWT_ALGORITHM=HS256
+
+# Access Token 过期时间（分钟）
+JWT_ACCESS_TOKEN_EXPIRE_MINUTES=30
+
+# Refresh Token 过期时间（天）
+JWT_REFRESH_TOKEN_EXPIRE_DAYS=30
+
+
+# ==================== 邮件配置 ====================
+
+# 邮件功能尚未上线，暂时关闭
+ENABLE_EMAIL_DELIVERY=false
+
+# SMTP 服务器地址
+SMTP_HOST=
+
+# SMTP 端口
+SMTP_PORT=587
+
+# SMTP 用户名（邮件功能上线后配置）
+SMTP_USER=
+
+# SMTP 密码（邮件功能上线后配置）
+SMTP_PASSWORD=
+
+# 发件人邮箱（邮件功能上线后配置）
+SMTP_FROM_EMAIL=
+
+# 发件人显示名称
+SMTP_FROM_NAME=小虾 SaaS
+
+# 启用 TLS
+SMTP_USE_TLS=true
+
+
+# ==================== 阿里云 OSS 配置 ====================
+
+# OSS 区域 endpoint
+OSS_ENDPOINT=oss-cn-hangzhou.aliyuncs.com
+
+# OSS Access Key ID
+# ${OSS_ACCESS_KEY_ID} — 替换为实际的 OSS Access Key ID
+OSS_ACCESS_KEY_ID=${OSS_ACCESS_KEY_ID}
+
+# OSS Access Key Secret
+# ${OSS_ACCESS_KEY_SECRET} — 替换为实际的 OSS Access Key Secret
+OSS_ACCESS_KEY_SECRET=${OSS_ACCESS_KEY_SECRET}
+
+# OSS Bucket 名称
+OSS_BUCKET_NAME=xiaoxia-autocut
+
+# 直传最大文件大小（MB）
+OSS_DIRECT_UPLOAD_MAX_MB=2000
+
+# 直传签名有效期（秒）
+OSS_DIRECT_UPLOAD_EXPIRE_SECONDS=900
+
+
+# ==================== CORS 配置 ====================
+
+# 允许跨域的前端域名列表，逗号分隔
+CORS_ORIGINS_RAW=https://xiaoxiajianji.com,https://api.xiaoxiajianji.com
+
+
+# ==================== 生成文件路径 ====================
+
+# 容器内生成文件目录（固定值，勿改）
+GENERATED_FILES_DIR=/app/generated
+
+# 生成文件 URL 前缀
+GENERATED_FILES_URL_PREFIX=/generated-files
+
+# 主机上生成文件目录（供 Docker volume bind mount 使用）
+GENERATED_FILES_HOST_DIR=/var/lib/xiaoxia-saas-production/generated
+
+
+# ==================== 渲染引擎配置 ====================
+
+# 渲染引擎选择：legacy（旧引擎，稳定）/ unified（新架构）
+RENDER_ENGINE=legacy
+
+
+# ==================== CosyVoice 语音合成 ====================
+
+# 阿里云百灵语音合成服务 API Key
+# ${COSYVOICE_API_KEY} — 替换为实际的 CosyVoice API Key
+COSYVOICE_API_KEY=${COSYVOICE_API_KEY}
+
+# API 基础 URL
+COSYVOICE_BASE_URL=https://dashscope.aliyuncs.com/api/v1
+
+# 模型选择：cosyvoice-v3-flash（推荐）/ cosyvoice-v3-plus
+COSYVOICE_MODEL=cosyvoice-v3-flash
+
+# 音色：v3 系列系统音色带 _v3 后缀
+COSYVOICE_VOICE=longxiaoxia_v3
+
+# 采样率
+COSYVOICE_SAMPLE_RATE=22050
+
+# 输出格式
+COSYVOICE_FORMAT=wav
+
+# 音色克隆模型名（固定值）
+COSYVOICE_CLONE_MODEL=voice-enrollment
+
+# DashScope 通用 API Key（与 CosyVoice 共用）
+DASHSCOPE_API_KEY=${DASHSCOPE_API_KEY}
+
+
+# ==================== MediaKit 视频理解（火山引擎）====================
+
+MEDIAKIT_API_KEY=${MEDIAKIT_API_KEY}
+MEDIAKIT_BASE_URL=https://mediakit.cn-beijing.volces.com/api/v1
+MEDIAKIT_TIMEOUT=60
+
+
+# ==================== 监控（可选）====================
+# Sentry DSN（取消注释并填入实际值以启用错误追踪）
+# SENTRY_DSN=${SENTRY_DSN}
@@ -0,0 +1,233 @@
+# ============================================================
+# 小虾 SaaS — Staging 环境配置模板
+# ============================================================
+# 使用方式：复制为 /var/lib/xiaoxia-saas-staging/.env 并填入实际密钥
+# 敏感值标记为 ${PLACEHOLDER}，部署前必须替换为真实值
+# ============================================================
+
+
+# ==================== 应用基本配置 ====================
+
+# 应用名称
+APP_NAME=xiaoxia-saas
+
+# 环境标识
+APP_ENV=staging
+
+# Staging 开启 Debug 模式便于排查问题
+DEBUG=true
+
+# 应用基础 URL（前端页面地址）
+APP_BASE_URL=https://staging.xiaoxiajianji.com
+
+# 对外公开的 API 基础 URL（用于生成回调链接等）
+PUBLIC_API_BASE_URL=https://staging-api.xiaoxiajianji.com
+
+# API 服务监听地址
+API_HOST=0.0.0.0
+
+# API 服务监听端口
+API_PORT=8000
+
+# 生产/预发布环境关闭自动建表，使用 alembic migration
+AUTO_CREATE_SCHEMA=false
+
+
+# ==================== 数据库配置 ====================
+
+# 数据库连接串（格式：postgresql+psycopg://user:password@host:port/dbname）
+# ${DATABASE_URL} — 替换为实际的 Staging PostgreSQL 连接串
+DATABASE_URL=${DATABASE_URL}
+
+# 连接池大小（常驻连接数）
+DATABASE_POOL_SIZE=20
+
+# 连接池最大溢出连接数（pool_size + max_overflow = 最大并发连接数）
+DATABASE_MAX_OVERFLOW=10
+
+# 获取连接超时时间（秒）
+DATABASE_POOL_TIMEOUT=30
+
+# 连接回收时间（秒），防止数据库端主动断开导致的死连接
+DATABASE_POOL_RECYCLE=3600
+
+# 不使用内存数据库
+USE_IN_MEMORY_DB=false
+
+
+# ==================== Redis 配置 ====================
+
+# Redis 连接 URL（格式：redis://[:password@]host:port/db）
+# ${REDIS_URL} — 替换为实际的 Staging Redis 连接串
+REDIS_URL=${REDIS_URL}
+
+# 启用 Redis Session 存储（多实例部署必须开启）
+ENABLE_REDIS_SESSIONS=true
+
+
+# ==================== Celery 任务队列 ====================
+
+# Celery Broker（任务分发），使用 Redis db0
+CELERY_BROKER_URL=${CELERY_BROKER_URL}
+
+# Celery Result Backend（任务结果存储），使用 Redis db1
+CELERY_RESULT_BACKEND=${CELERY_RESULT_BACKEND}
+
+
+# ==================== Worker 配置 ====================
+
+# Worker 进程名称
+WORKER_NAME=xiaoxia-saas-worker
+
+# Worker 并发数（同时执行的任务数）
+WORKER_CONCURRENCY=1
+
+# 每个子进程最多处理多少任务后重启（防止内存泄漏）
+WORKER_MAX_TASKS_PER_CHILD=1000
+
+
+# ==================== JWT 认证配置 ====================
+
+# JWT 签名密钥 — 必须设置为强随机字符串（至少32字符）
+# ${JWT_SECRET_KEY} — 替换为实际的随机密钥
+JWT_SECRET_KEY=${JWT_SECRET_KEY}
+
+# JWT 签名算法
+JWT_ALGORITHM=HS256
+
+# Access Token 过期时间（分钟）
+JWT_ACCESS_TOKEN_EXPIRE_MINUTES=1440
+
+# Refresh Token 过期时间（天）
+JWT_REFRESH_TOKEN_EXPIRE_DAYS=30
+
+
+# ==================== 邮件配置 ====================
+
+# 邮件功能尚未上线，暂时关闭
+ENABLE_EMAIL_DELIVERY=false
+
+# SMTP 服务器地址
+SMTP_HOST=smtp.gmail.com
+
+# SMTP 端口
+SMTP_PORT=587
+
+# SMTP 用户名（邮件功能上线后配置）
+SMTP_USER=
+
+# SMTP 密码（邮件功能上线后配置）
+SMTP_PASSWORD=
+
+# 发件人邮箱（邮件功能上线后配置）
+SMTP_FROM_EMAIL=
+
+# 发件人显示名称
+SMTP_FROM_NAME=小虾 SaaS
+
+# 启用 TLS
+SMTP_USE_TLS=true
+
+
+# ==================== 阿里云 OSS 配置 ====================
+
+# OSS 区域 endpoint
+OSS_ENDPOINT=oss-cn-hangzhou.aliyuncs.com
+
+# OSS Access Key ID
+# ${OSS_ACCESS_KEY_ID} — 替换为实际的 OSS Access Key ID
+OSS_ACCESS_KEY_ID=${OSS_ACCESS_KEY_ID}
+
+# OSS Access Key Secret
+# ${OSS_ACCESS_KEY_SECRET} — 替换为实际的 OSS Access Key Secret
+OSS_ACCESS_KEY_SECRET=${OSS_ACCESS_KEY_SECRET}
+
+# OSS Bucket 名称
+OSS_BUCKET_NAME=xiaoxia-autocut
+
+# 直传最大文件大小（MB）
+OSS_DIRECT_UPLOAD_MAX_MB=2000
+
+# 直传签名有效期（秒）
+OSS_DIRECT_UPLOAD_EXPIRE_SECONDS=900
+
+
+# ==================== MinIO 配置（Staging 独有）====================
+# Staging 环境使用 MinIO 替代 OSS 进行文件存储测试
+
+# MinIO 服务 Endpoint
+# ${MINIO_ENDPOINT} — 替换为实际的 MinIO 地址
+MINIO_ENDPOINT=${MINIO_ENDPOINT}
+
+# MinIO Access Key
+# ${MINIO_ACCESS_KEY} — 替换为实际的 MinIO Access Key
+MINIO_ACCESS_KEY=${MINIO_ACCESS_KEY}
+
+# MinIO Secret Key
+# ${MINIO_SECRET_KEY} — 替换为实际的 MinIO Secret Key
+MINIO_SECRET_KEY=${MINIO_SECRET_KEY}
+
+# MinIO Bucket 名称
+MINIO_BUCKET_NAME=${MINIO_BUCKET_NAME}
+
+# 是否使用 SSL 连接 MinIO
+MINIO_USE_SSL=false
+
+
+# ==================== CORS 配置 ====================
+
+# 允许跨域的前端域名列表，逗号分隔
+CORS_ORIGINS_RAW=https://staging.xiaoxiajianji.com,https://staging-api.xiaoxiajianji.com
+
+
+# ==================== 生成文件路径 ====================
+
+# 容器内生成文件目录（固定值，勿改）
+GENERATED_FILES_DIR=/app/generated
+
+# 生成文件 URL 前缀
+GENERATED_FILES_URL_PREFIX=/generated-files
+
+# 主机上生成文件目录（供 Docker volume bind mount 使用）
+GENERATED_FILES_HOST_DIR=/var/lib/xiaoxia-saas-staging/generated
+
+
+# ==================== 渲染引擎配置 ====================
+
+# 渲染引擎选择：legacy（旧引擎，稳定）/ unified（新架构）
+RENDER_ENGINE=legacy
+
+
+# ==================== CosyVoice 语音合成 ====================
+
+# 阿里云百灵语音合成服务 API Key
+# ${COSYVOICE_API_KEY} — 替换为实际的 CosyVoice API Key
+COSYVOICE_API_KEY=${COSYVOICE_API_KEY}
+
+# API 基础 URL
+COSYVOICE_BASE_URL=https://dashscope.aliyuncs.com/api/v1
+
+# 模型选择：cosyvoice-v3-flash（推荐）/ cosyvoice-v3-plus
+COSYVOICE_MODEL=cosyvoice-v3-flash
+
+# 音色：v3 系列系统音色带 _v3 后缀
+COSYVOICE_VOICE=longxiaoxia_v3
+
+# 采样率
+COSYVOICE_SAMPLE_RATE=22050
+
+# 输出格式
+COSYVOICE_FORMAT=wav
+
+# 音色克隆模型名（固定值）
+COSYVOICE_CLONE_MODEL=voice-enrollment
+
+# DashScope 通用 API Key（与 CosyVoice 共用）
+DASHSCOPE_API_KEY=${DASHSCOPE_API_KEY}
+
+
+# ==================== MediaKit 视频理解（火山引擎）====================
+
+MEDIAKIT_API_KEY=${MEDIAKIT_API_KEY}
+MEDIAKIT_BASE_URL=https://mediakit.cn-beijing.volces.com/api/v1
+MEDIAKIT_TIMEOUT=60
@@ -0,0 +1,51 @@
+server {
+    listen 80;
+    server_name _;
+    root /usr/share/nginx/html;
+    index index.html;
+
+    gzip on;
+    gzip_vary on;
+    gzip_min_length 1024;
+    gzip_types text/plain text/css text/xml text/javascript application/javascript application/json application/xml+rss;
+
+    client_max_body_size 800m;
+
+    # SPA routing - index.html 禁止缓存，确保每次获取最新版本
+    location / {
+        try_files $uri /index.html;
+    }
+
+    # API proxy — Production 环境代理到 production API 容器
+    resolver 127.0.0.11 valid=10s;
+    resolver_timeout 5s;
+    location /api/ {
+        proxy_pass http://xiaoxia-api-production:8000/api/;
+        proxy_set_header Host $host;
+        proxy_set_header X-Real-IP $remote_addr;
+        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
+        proxy_set_header X-Forwarded-Proto $scheme;
+        proxy_read_timeout 300s;
+        proxy_send_timeout 300s;
+        proxy_request_buffering off;
+    }
+
+    # Generated files — 通过 alias 映射容器内 /app/generated/ 目录
+    location /generated-files/ {
+        alias /app/generated/;
+    }
+
+    # Assets with legacy fallback — 部署期间兼容旧版缓存的 hash 文件名
+    # 先在当前镜像中找，找不到去 legacy-assets 目录找（从旧版本容器中备份的）
+    location ^~ /assets/ {
+        expires 1y;
+        add_header Cache-Control "public, immutable";
+        try_files $uri /assets-legacy$uri =404;
+    }
+
+    # 静态资源长缓存
+    location ~* \.(js|css|png|jpg|jpeg|gif|ico|svg|woff|woff2|ttf|eot)$ {
+        expires 1y;
+        add_header Cache-Control "public, immutable";
+    }
+}
@@ -0,0 +1,50 @@
+server {
+    listen 80;
+    server_name _;
+    root /usr/share/nginx/html;
+    index index.html;
+
+    gzip on;
+    gzip_vary on;
+    gzip_min_length 1024;
+    gzip_types text/plain text/css text/xml text/javascript application/javascript application/json application/xml+rss;
+
+    client_max_body_size 800m;
+
+    # SPA routing - index.html 禁止缓存，确保每次获取最新版本
+    location = /index.html {
+        add_header Cache-Control "no-cache, no-store, must-revalidate";
+        add_header Pragma "no-cache";
+        expires 0;
+    }
+
+    # SPA fallback
+    location / {
+        try_files $uri /index.html;
+    }
+
+    # API proxy — Staging 环境代理到 staging API 容器
+    resolver 127.0.0.11 valid=10s;
+    resolver_timeout 5s;
+    location /api/ {
+        proxy_pass http://xiaoxia-api-staging:8000/api/;
+        proxy_set_header Host $host;
+        proxy_set_header X-Real-IP $remote_addr;
+        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
+        proxy_set_header X-Forwarded-Proto $scheme;
+        proxy_read_timeout 300s;
+        proxy_send_timeout 300s;
+        proxy_request_buffering off;
+    }
+
+    # Generated files — 通过 alias 映射容器内 /app/generated/ 目录
+    location /generated-files/ {
+        alias /app/generated/;
+    }
+
+    # 静态资源长缓存
+    location ~* \.(js|css|png|jpg|jpeg|gif|ico|svg|woff|woff2|ttf|eot)$ {
+        expires 1y;
+        add_header Cache-Control "public, immutable";
+    }
+}
@@ -30,6 +30,14 @@ RUN apt-get update \
    && apt-get install -y --no-install-recommends docker-ce-cli docker-buildx-plugin \
    && rm -rf /var/lib/apt/lists/*

+# Pre-bake ffmpeg: unit-tests run in fresh containers each time; installing ffmpeg
+# on every job cost ~24 min (apt update + hundreds of codec deps). Bake it into the
+# image so step_install_ffmpeg.sh detects it and exits instantly.
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends ffmpeg \
+    && ffmpeg -version | head -1 \
+    && rm -rf /var/lib/apt/lists/*
+
 # Pre-install base deps (layer cache)
 COPY requirements-base.txt ./
 RUN python -m venv "$VIRTUAL_ENV" \
@@ -175,9 +175,14 @@ services:
      - xiaoxia-net
    
    # =========================================
-    # 重要: 生产环境不要添加任何 volume 挂载到 /usr/share/nginx/html
-    # 这会导致静态文件被覆盖，返回 403 错误
+    # Nginx 配置运行时覆盖
+    # 确保容器使用正确环境的 nginx 配置，即使镜像构建时使用了默认配置
+    # 注意: 只覆盖 /etc/nginx/conf.d/default.conf，不挂载 /usr/share/nginx/html
    # =========================================
+    environment:
+      - NGINX_ENV=${ENV:-staging}
+    volumes:
+      - ./nginx-${ENV:-staging}.conf:/etc/nginx/conf.d/default.conf:ro
    
    healthcheck:
      test: ["CMD", "wget", "--spider", "-q", "http://127.0.0.1:80"]
@@ -208,7 +213,7 @@ volumes:
      # 重要: 确保主机目录存在且有正确权限
      # Staging: /var/lib/xiaoxia-saas-staging/generated
      # Production: /var/lib/xiaoxia-saas-production/generated
-      device: ${GENERATED_FILES_HOST_DIR:-/var/lib/xiaoxia-saas-staging/generated}
+      device: ${GENERATED_FILES_HOST_DIR:?GENERATED_FILES_HOST_DIR must be set in .env}

 # ===========================================
 # 网络配置
@@ -1,5 +1,6 @@
 # Build stage
 FROM git.xiaoxiajianji.com/xiaoxia/base/node:20 AS builder
+ARG SOURCE_HASH=""
 WORKDIR /app
 ARG VITE_API_URL=https://saas-api.xiaoxiajianji.com
 ENV VITE_API_URL=$VITE_API_URL
@@ -18,8 +19,10 @@ COPY apps/web/ ./

 # 构建：TS增量编译 + Vite构建，tsbuildinfo用cache mount持久化
 # node_modules直接使用镜像中已安装的（layer缓存保证完整性）
+# SOURCE_HASH 变化时强制重新执行（防止 buildkit 幽灵缓存命中）
 RUN --mount=type=cache,target=/app/apps/web/.tscache,sharing=locked \
    mkdir -p .tscache \
+    && echo "SOURCE_HASH=${SOURCE_HASH}" > .cache_bust \
    && ./node_modules/.bin/tsc --incremental --tsBuildInfoFile .tscache/tsconfig.tsbuildinfo \
    && ./node_modules/.bin/vite build

@@ -127,6 +127,13 @@ class InMemoryAssetRepository:
        items = [a for a in self._assets.values() if tag_set.issubset(set(a.tag_ids))]
        return items[skip : skip + limit]

+    def find_by_storage_key(self, storage_key: str) -> Asset | None:
+        """按 storage_key 查找素材。"""
+        for asset in self._assets.values():
+            if asset.storage_key == storage_key:
+                return asset
+        return None
+
    def find_by_library_and_file_hash(
        self,
        library_id: str,
@@ -99,8 +99,8 @@ class SessionStore(SessionStorePort):
        session_id: str,
        user_id: str,
        refresh_token: str,
-        device_info: str,
-        ip_address: str,
+        device_info: str = "",
+        ip_address: str = "",
        expires_in_seconds: int = 30 * 24 * 60 * 60,  # 30 天
    ) -> bool:
        """
@@ -96,7 +96,7 @@ class EmailService(EmailServicePort):
        except Exception as e:
            return False, str(e)

-    def send_verification_email(
+    def send_verification_email(  # type: ignore[override]
        self,
        to_email: str,
        username: str,
@@ -165,7 +165,7 @@ class EmailService(EmailServicePort):

        return self.send_email(to_email, subject, html_body, text_body)

-    def send_password_reset_email(
+    def send_password_reset_email(  # type: ignore[override]
        self,
        to_email: str,
        username: str,
@@ -426,6 +426,13 @@ class SQLAlchemyAssetRepository:
        models = self.session.query(AssetModel).filter(AssetModel.id.in_(ids)).offset(skip).limit(limit).all()
        return [self._to_domain(m) for m in models]

+    def find_by_storage_key(self, storage_key: str) -> Asset | None:
+        """按 storage_key（对应 DB 中的 file_url）查找素材。"""
+        model = self.session.query(AssetModel).filter(AssetModel.file_url == storage_key).first()
+        if model is None:
+            return None
+        return self._to_domain(model)
+
    def find_by_library_and_file_hash(
        self,
        library_id: str,
@@ -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,
@@ -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
@@ -30,6 +30,9 @@ class SQLAlchemyGeneratedVideoRepository:
            video_fingerprint=json.dumps(video.video_fingerprint) if video.video_fingerprint else None,
            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,
        )
@@ -60,6 +63,9 @@ class SQLAlchemyGeneratedVideoRepository:
            video_fingerprint=json.loads(getattr(model, "video_fingerprint", "null") or "null"),
            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,
        )
@@ -74,6 +80,9 @@ class SQLAlchemyGeneratedVideoRepository:
        model.video_fingerprint = json.dumps(video.video_fingerprint) if video.video_fingerprint else None
        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
@@ -82,6 +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)
@@ -204,6 +231,9 @@ class SQLAlchemyGeneratedVideoRepository:
            video_fingerprint=json.loads(getattr(model, "video_fingerprint", "null") or "null"),
            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,
        )
@@ -289,6 +289,7 @@ class GenerationTaskModel(Base):
    completed_at = Column(DateTime, nullable=True)
    created_by_user_id = Column(String(36), nullable=False, default="", index=True)
    source_edit_plan_id = Column(String(36), nullable=True, index=True)
+    edit_plan_id = Column(String(36), nullable=True, index=True)
    asset_select_mode = Column(String(20), nullable=False, default="")
    batch_id = Column(String(36), nullable=False, default="", index=True)
    video_title = Column(String(255), nullable=False, default="")
@@ -338,6 +339,9 @@ class GeneratedVideoModel(Base):
    video_fingerprint = Column(Text, nullable=True)
    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):
@@ -413,6 +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))
@@ -496,6 +503,7 @@ class TemplateCategoryModel(Base):
    id = Column(String(36), primary_key=True)
    user_id = Column(String(36), nullable=False, index=True)
    name = Column(String(100), nullable=False)
+    sort_order = Column(Integer, nullable=False, default=0)
    created_at = Column(DateTime, nullable=False, default=lambda: datetime.now(timezone.utc))


@@ -617,3 +625,20 @@ class CoverTemplateModel(Base):
    config = Column(JSON, nullable=False, default=dict)
    created_at = Column(DateTime(timezone=True), nullable=False, default=lambda: datetime.now(timezone.utc))
    updated_at = Column(DateTime(timezone=True), nullable=False, default=lambda: datetime.now(timezone.utc))
+
+
+class VideoFingerprintChunkModel(Base):
+    """分片视频指纹 — 每个视频按时间分片存储 pHash + color_histogram."""
+
+    __tablename__ = "video_fingerprint_chunks"
+
+    id = Column(String(36), primary_key=True)
+    video_id = Column(String(36), nullable=False, index=True)
+    project_id = Column(String(36), nullable=False, index=True)
+    user_id = Column(String(36), nullable=False, index=True, default="")
+    start_time_ms = Column(Integer, nullable=False)
+    end_time_ms = Column(Integer, nullable=False)
+    phash_binary = Column(String(16), nullable=False)
+    color_histogram = Column(JSON, nullable=False)
+    frame_count = Column(Integer, nullable=False, default=1)
+    created_at = Column(DateTime, nullable=False, default=lambda: datetime.now(timezone.utc))
--- a/Show More
+++ b/Show More
				`@@ -0,0 +1 @@`
				`CI re-trigger after runner add-host/DNS fix. This file is harmless and not referenced.`