优化缩略图

docker-compose.metatube.yml

@@ -0,0 +1,216 @@
+# MeBox + MetaTube 一体部署 Docker Compose 部署文件
+#
+# 组件：
+#   MeBox + PostgreSQL + Redis + MetaTube
+#
+# 与 docker-compose.standard.yml 的区别：
+#   额外内置一个 MetaTube 后端（番号元数据刮削 / 封面人脸裁剪），
+#   并与 MeBox 共用同一个 PostgreSQL 实例，不需要再单独起一套数据库。
+#
+# 使用方式二选一：
+#   1. 保存为 docker-compose.yml 后执行：
+#        docker compose up -d
+#   2. 保留本文件名时执行：
+#        docker compose -f docker-compose.metatube.yml up -d
+#
+# 首次启动后需要做一次配置（只有这一处是手动的）：
+#   登录 MeBox → 设置 → 番号刮削：
+#     - 番号刮削引擎：MetaTube 后端服务，或 智能混合
+#     - MetaTube 服务端地址：http://metatube:8080
+#     - 访问 Token：留空（仅内部网络，未启用认证；如需认证两边填同一个值）
+#
+# 说明：
+#   - MetaTube 使用独立数据库 metatube，由 metatube-db-init 一次性容器自动创建；
+#     和 MeBox 的 mebox 库共用同一个 PostgreSQL 实例，只是分成两个库。
+#   - MetaTube 不向宿主机映射端口，只在本 Compose 内部网络提供服务。
+#   - MeBox 故意不依赖 MetaTube：MetaTube 未就绪或挂掉时，智能混合模式会自动回退到内置刮削源。
+#   - 更新全部组件：
+#       docker compose -f docker-compose.metatube.yml pull
+#       docker compose -f docker-compose.metatube.yml up -d
+#     保留本文件名时，管理面板「系统更新」需要把「Compose 安装目录」指向本文件所在目录，
+#     或把更新命令改成带 -f docker-compose.metatube.yml 的形式。
+#
+# 默认账号：
+#   admin / admin123
+
+services:
+  mebox:
+    image: ghcr.io/truewhile/mebox:latest
+
+    restart: unless-stopped
+    init: true
+    depends_on:
+      postgres:
+        condition: service_healthy
+      redis:
+        condition: service_healthy
+
+    ports:
+      - "18080:8080"
+
+    extra_hosts:
+      - "host.docker.internal:host-gateway"
+
+    volumes:
+      # 程序运行数据：JWT 密钥、运行配置、旧 SQLite 迁移源。
+      - ./data:/data
+
+      # 缓存目录：海报缓存、临时文件等。通常不用备份。
+      - ./cache:/cache
+
+      # 媒体库目录。自动整理/重命名/入库需要读写权限。
+      # NAS 示例：source: /vol1/1000/Media
+      # Windows Docker Desktop 示例：source: D:/Media
+      # create_host_path=false 可以避免路径写错时 Docker 自动创建空文件夹。
+      - ./media: /media
+
+    environment:
+      TZ: Asia/Shanghai
+
+      PUID: "0"
+      PGID: "0"
+
+      MEBOX_APP_HOST: 0.0.0.0
+      MEBOX_APP_PORT: 8080
+      MEBOX_APP_WEB_DIR: /app/web/dist
+      MEBOX_APP_DATA_DIR: /data
+      MEBOX_LOGGING_LEVEL: info
+      MEBOX_LOGGING_FORMAT: console
+      MEBOX_LOGGING_OUTPUT_PATH: /data/logs
+      MEBOX_LOGGING_MAX_SIZE_MB: "20"
+      MEBOX_LOGGING_MAX_BACKUPS: "5"
+      MEBOX_LOGGING_MAX_AGE_DAYS: "30"
+
+      MEBOX_DATABASE_TYPE: postgres
+      MEBOX_DATABASE_DSN: postgres://mebox:mebox@postgres:5432/mebox?sslmode=disable
+      MEBOX_DATABASE_DB_PATH: /data/mebox.db
+
+      # Redis 只做热缓存，源数据仍在 PostgreSQL；Redis 丢失可自动重建。
+      MEBOX_CACHE_REDIS_URL: redis://redis:6379/0
+      MEBOX_CACHE_CACHE_DIR: /cache
+
+      MEBOX_UPDATE_IMAGE: ghcr.io/truewhile/mebox:latest
+
+      # 路径换算配置。左边宿主机真实路径要和 volumes 左边保持一致。
+      MEBOX_MEDIA_DIR: /media
+      MEBOX_MEDIA_CONTAINER_DIR: /media
+      MEBOX_DOWNLOAD_DIR: /downloads
+      MEBOX_DOWNLOAD_CONTAINER_DIR: /downloads
+
+      MEBOX_TRANSCODER_ENABLED: "true"
+      MEBOX_TRANSCODER_HARDWARE_ACCEL: "false"
+      MEBOX_TRANSCODER_REALTIME: "true"
+      MEBOX_TRANSCODER_THREADS: "2"
+      MEBOX_TRANSCODER_MAX_CONCURRENT: "1"
+      MEBOX_TRANSCODER_IDLE_TIMEOUT_SECONDS: "120"
+
+    healthcheck:
+      test: ["CMD-SHELL", "busybox wget -qO- http://127.0.0.1:8080/api/health || exit 1"]
+      interval: 30s
+      timeout: 10s
+      retries: 5
+      start_period: 30s
+
+    logging:
+      driver: json-file
+      options:
+        max-size: "20m"
+        max-file: "3"
+
+  postgres:
+    image: postgres:16-alpine
+    # 首次部署允许拉取；日常更新请只 pull mebox。
+    # MetaTube 与 MeBox 共用本实例（不同数据库），无需第二套 PostgreSQL。
+    pull_policy: missing
+    restart: unless-stopped
+    environment:
+      POSTGRES_DB: mebox
+      POSTGRES_USER: mebox
+      POSTGRES_PASSWORD: mebox
+      TZ: Asia/Shanghai
+    volumes:
+      # 同时存放 mebox 与 metatube 两个数据库，备份这一个目录即可。
+      - ./postgres:/var/lib/postgresql/data
+    healthcheck:
+      test: ["CMD-SHELL", "pg_isready -h 127.0.0.1 -U mebox -d mebox"]
+      interval: 10s
+      timeout: 5s
+      retries: 10
+    logging:
+      driver: json-file
+      options:
+        max-size: "20m"
+        max-file: "3"
+
+  # 一次性容器：确保 metatube 库存在。
+  # MetaTube 自己的 -db-auto-migrate 只建表不建库，所以必须先建库。
+  # 数据库已存在时它会直接退出，重复执行无副作用。
+  metatube-db-init:
+    image: postgres:16-alpine
+    pull_policy: missing
+    restart: "no"
+    depends_on:
+      postgres:
+        condition: service_healthy
+    environment:
+      PGPASSWORD: mebox
+    entrypoint: ["/bin/sh", "-c"]
+    command:
+      - |
+        psql -h postgres -U mebox -d mebox -tAc "SELECT 1 FROM pg_database WHERE datname='metatube'" | grep -q 1 || psql -h postgres -U mebox -d mebox -c "CREATE DATABASE metatube OWNER mebox"
+    logging:
+      driver: json-file
+      options:
+        max-size: "5m"
+        max-file: "1"
+
+  metatube:
+    image: ghcr.io/metatube-community/metatube-server:latest
+    restart: unless-stopped
+    depends_on:
+      metatube-db-init:
+        condition: service_completed_successfully
+    environment:
+      TZ: Asia/Shanghai
+      # 需要走代理时在这里填写，例如 http://192.168.1.2:7890
+      HTTP_PROXY: ""
+      HTTPS_PROXY: ""
+    # 不需要 ports：MeBox 通过内部网络访问 http://metatube:8080。
+    # 只有想让局域网内其他工具直连时才映射，且建议只映射到 127.0.0.1。
+    command:
+      - -dsn
+      - postgres://mebox:mebox@postgres:5432/metatube?sslmode=disable
+      - -port
+      - "8080"
+      - -db-auto-migrate
+      - -db-prepared-stmt
+    logging:
+      driver: json-file
+      options:
+        max-size: "20m"
+        max-file: "3"
+
+  redis:
+    image: redis:7-alpine
+    pull_policy: missing
+    restart: unless-stopped
+    command:
+      - redis-server
+      - --appendonly
+      - "yes"
+      - --maxmemory
+      - 256mb
+      - --maxmemory-policy
+      - allkeys-lru
+    volumes:
+      - ./redis:/data
+    healthcheck:
+      test: ["CMD", "redis-cli", "ping"]
+      interval: 10s
+      timeout: 5s
+      retries: 10
+    logging:
+      driver: json-file
+      options:
+        max-size: "20m"
+        max-file: "3"

docker-compose.standard.yml

@@ -44,30 +44,13 @@ services:
       # NAS 示例：source: /vol1/1000/Media
       # Windows Docker Desktop 示例：source: D:/Media
       # create_host_path=false 可以避免路径写错时 Docker 自动创建空文件夹。
-      - type: bind
-        source: ./media
-        target: /media
-        bind:
-          create_host_path: false
-
-      # 下载目录。需要和下载器保存路径保持一致。
-      # NAS 示例：source: /vol1/1000/Downloads
-      # Windows Docker Desktop 示例：source: D:/Downloads
-      - type: bind
-        source: ./downloads
-        target: /downloads
-        bind:
-          create_host_path: false
-
-      # 管理面板「系统更新」需要访问 Docker 引擎。
-      # 需要一键更新 Docker 镜像时取消下一行注释。
-      # - /var/run/docker.sock:/var/run/docker.sock
+      - ./media: /media
 
     environment:
       TZ: Asia/Shanghai
 
-      PUID: "1000"
-      PGID: "1000"
+      PUID: "0"
+      PGID: "0"
 
       MEBOX_APP_HOST: 0.0.0.0
       MEBOX_APP_PORT: 8080

docs/reading-module-design.md

@@ -0,0 +1,743 @@
+# 阅读模块（Reading Module）设计与实施方案
+
+> 状态：设计稿，待评审
+> 目标版本：v0.2.0（分期落地，见第 9 节）
+> 关联现有子系统：媒体库 / 网盘存储 / 权限体系 / 任务队列
+
+---
+
+## 1. 目标与范围
+
+### 1.1 已确认的产品决策
+
+| 维度 | 决策 |
+| --- | --- |
+| 内容类型 | **电子书 + 漫画统一书架**（EPUB / TXT / PDF / MOBI 与 CBZ / CBR / 图片文件夹） |
+| 书源 | **独立书库**（不复用影视媒体库）+ **网盘直链阅读** |
+| 首版范围 | **完整版**：多用户书库权限 + 阅读统计 |
+| 阅读形态 | **滚动流式与分页翻页双模式**，用户可切换并持久化偏好 |
+
+### 1.2 明确的非目标
+
+- **不接入 Emby/Jellyfin 协议。** Emby 的 `Items` / `Views` / `PlaybackInfo` 语义围绕音视频构建，没有书籍章节与阅读进度的对应概念。强行映射会污染 `internal/service/emby_*.go` 与 `internal/handler/emby_*.go` 的既有兼容层，收益极低。阅读能力只通过 MeBox 自己的 Web UI 提供。
+- **不复用 `model.Library` / `LibraryRoot`。** `Library.Type` 的取值域是 `movie/tv/anime/music`，且被海报墙轮播（`CarouselEnabled`）、自动整理管线、Emby 视图、首页预览等链路消费。把书库塞进去会导致这些链路需要到处加 `type != "book"` 判断。
+- 首版不做：听书 TTS、在线书源（笔趣阁类）、社交分享、跨设备同步批注冲突合并。
+
+---
+
+## 2. 总体架构
+
+### 2.1 分层落位
+
+完全沿用现有分层，不引入新模式：
+
+```
+web/src/pages/Books*.tsx            ← 页面
+web/src/components/Book*.tsx         ← 阅读器与书架组件
+web/src/api/books.ts                 ← axios 封装（仿 web/src/api/library.ts）
+        ↓ /api/books/*
+internal/handler/books*.go           ← 反序列化 + 权限校验 + 响应
+internal/service/book_*.go           ← 业务策略（扫描、解析、进度、统计）
+internal/repository/book_*.go        ← 纯持久化
+internal/model/book.go               ← GORM 模型，注册进 model.AllModels()
+```
+
+新增路由注册走 `internal/handler/routes_authenticated_features.go` 的既有范式，新增一个 `registerAuthedBookRoutes(authed, svc)`，在 `registerAuthenticatedRoutes` 链上挂载。`service.Container` 与 `repository.Container` 各追加一个字段。
+
+### 2.2 与现有能力的复用点
+
+| 现有部件 | 复用方式 |
+| --- | --- |
+| `service.StreamService.ServeFile`（`internal/service/stream_file.go`） | 已用 `http.ServeContent` 处理 HEAD / Range / If-Modified-Since，**PDF 与原始文件流直接照搬这条路径** |
+| `cloud.Provider.Resolve(ctx, fileRef) (*DirectLink, error)`（`internal/service/cloud/cloud.go`） | 网盘书源的直链解析入口，`DirectLink.Proxy` 决定 302 还是反代 |
+| `model.StorageConfig`（`internal/model/storage_assistant.go`） | 直接复用为网盘书源的账号凭据载体，**不新建凭据表** |
+| `service.ImageProxy`（`internal/service/image_proxy*.go`） | 漫画页与封面的磁盘缓存 + 远程拉取 + 缩放，复用其缓存目录与命名思路 |
+| `service.PruneImageCache` / `PruneImageCachePools`（`internal/service/cache_cleanup.go`） | 现成的「按池做 LRU 淘汰 + 按保留时长淘汰」助手，书籍缓存淘汰直接复用它 |
+| `service/scheduler_local_jobs.go` | 本地定时任务的挂载点，书籍缓存清理与每日统计汇总都注册在这里 |
+| `config.CacheConfig`（`internal/config/types.go`） | 已有 `CacheDir` / `ImagesMaxSizeMB` / `ImagesOriginalsMaxSizeMB` / `ImagesOriginalsTTLHours` / `MemoryMaxSizeMB`，书籍缓存容量配置直接挂进去 |
+| `service.FileManager`（`internal/service/filemanager.go`） | 本地书源目录浏览，前端复用 `LocalDirBrowserDialog.tsx` |
+| `service.Scheduler` | 书库定时扫描（默认关闭，管理员可开） |
+| `model.UserPermission` | 新增阅读权限位，见第 7 节 |
+| `helper.Go` / `Container.stopCtx` | 后台扫描任务的生命周期管理 |
+
+---
+
+## 3. 数据模型
+
+新增文件 `internal/model/book.go`，并在 `internal/model/model.go` 的 `AllModels()` 中追加。所有表继承 `model.Base`（UUID 主键 + 时间戳 + 软删除）。
+
+**表名约定**：`internal/model` 全包**没有任何 `TableName()` 覆盖**，一律使用 GORM 默认复数化（例如 `PlaybackHistory` → `playback_histories`，可从 `internal/database/schema_migration.go` 的裸 SQL 印证）。新表沿用该约定，不引入例外。因此模型命名要保证复数化结果干净：
+
+| 模型 | 表名 |
+| --- | --- |
+| `Book` | `books` |
+| `BookLibrary` | `book_libraries` |
+| `BookSource` | `book_sources` |
+| `BookChapter` | `book_chapters` |
+| `BookProgress` | `book_progresses` |
+| `BookAnnotation` | `book_annotations` |
+| `BookFavorite` | `book_favorites` |
+| `BookReadingSession` | `book_reading_sessions` |
+| `BookDailyStat` | `book_daily_stats` |
+
+（刻意用 `BookDailyStat` 而不是 `BookStatDaily`——后者复数化会得到 `book_stat_dailies`。）
+
+### 3.1 书库与书源
+
+```go
+// BookLibrary 是独立于影视媒体库的书库。
+type BookLibrary struct {
+    Base
+    Name        string `gorm:"size:128;not null" json:"name"`
+    Kind        string `gorm:"size:16;not null;default:mixed" json:"kind"` // ebook / comic / mixed
+    CoverURL    string `gorm:"size:1024" json:"cover_url,omitempty"`
+    Enabled     bool   `gorm:"default:true" json:"enabled"`
+    SortOrder   int    `gorm:"index;default:0" json:"sort_order"`
+    LastScanAt  *time.Time `json:"last_scan_at,omitempty"`
+    ScanStatus  string `gorm:"size:16;default:idle" json:"scan_status"` // idle / scanning / error
+    ScanMessage string `gorm:"size:512" json:"scan_message,omitempty"`
+}
+
+// BookSource 是书库下的一条挂载来源：本地目录或网盘路径。
+type BookSource struct {
+    Base
+    LibraryID       string `gorm:"index;size:36;not null" json:"library_id"`
+    Name            string `gorm:"size:128" json:"name,omitempty"`
+    StorageKind     string `gorm:"size:16;not null;default:local" json:"storage_kind"` // local / cloud
+    Path            string `gorm:"size:1024;not null" json:"path"`                     // 本地绝对路径 / 网盘内路径
+    StorageConfigID string `gorm:"index;size:36" json:"storage_config_id,omitempty"`   // 复用 model.StorageConfig
+    Depth           int    `gorm:"default:3" json:"depth"`                             // 扫描递归深度上限
+    Enabled         bool   `gorm:"default:true" json:"enabled"`
+    SortOrder       int    `gorm:"default:0" json:"sort_order"`
+}
+```
+
+`StorageKind = cloud` 时，`StorageConfigID` 指向一条 `StorageConfig`（`Type` ∈ `cloud115 / clouddrive2 / openlist / emby_remote`）。凭据解密沿用 `service.CryptoService`。
+
+### 3.2 书籍与章节
+
+```go
+type Book struct {
+    Base
+    LibraryID   string `gorm:"index;size:36;not null" json:"library_id"`
+    SourceID    string `gorm:"uniqueIndex:uniq_book_source_path,priority:1;index;size:36;not null" json:"source_id"`
+    // SourcePath 在本地源是绝对路径，在网盘源是「网盘内路径」，两者都用
+    // (source_id, source_path) 做唯一键，天然隔离两个 ID 空间。
+    SourcePath  string `gorm:"uniqueIndex:uniq_book_source_path,priority:2;size:1024;not null" json:"source_path"`
+    SourceRef   string `gorm:"size:256" json:"source_ref,omitempty"` // 网盘 file id / pickcode
+    Title       string `gorm:"size:512;not null" json:"title"`
+    Author      string `gorm:"size:256;index" json:"author,omitempty"`
+    SeriesName  string `gorm:"size:256;index" json:"series_name,omitempty"`
+    Volume      int    `json:"volume"`
+    Format      string `gorm:"size:16;not null" json:"format"` // epub/txt/pdf/mobi/cbz/cbr/folder
+    MediaKind   string `gorm:"size:16;not null;default:ebook" json:"media_kind"` // ebook / comic
+    SizeBytes   int64  `json:"size_bytes"`
+    FileHash    string `gorm:"index;size:64" json:"file_hash,omitempty"` // 大小+首尾采样，去重
+    CoverURL    string `gorm:"size:1024" json:"cover_url,omitempty"`
+    Description string `gorm:"type:text" json:"description,omitempty"`
+    Language    string `gorm:"size:32" json:"language,omitempty"`
+    Tags        string `gorm:"type:text" json:"tags,omitempty"` // 逗号分隔
+    ChapterCount int   `json:"chapter_count"`
+    WordCount    int64 `json:"word_count"`
+    PageCount    int   `json:"page_count"` // 漫画总页数 / PDF 页数
+    ParseStatus  string `gorm:"size:16;default:pending" json:"parse_status"` // pending/ok/failed
+    ParseMessage string `gorm:"size:512" json:"parse_message,omitempty"`
+    NSFW         bool  `gorm:"default:false" json:"nsfw"`
+    AddedAt      time.Time `json:"added_at"`
+}
+```
+
+**唯一键说明**：`SourcePath` 上的 `uniqueIndex` 需与 `SourceID` 组成复合键（`uniq_book_source_path`，priority 1 = `source_id`）。同一本书被两个书源包含时允许重复入库，这是符合预期的（用户可能故意如此）。
+
+```go
+// BookChapter 只存索引，不存正文（见 3.4 的取舍）。
+type BookChapter struct {
+    Base
+    BookID       string `gorm:"index:idx_book_chapter,priority:1;size:36;not null" json:"book_id"`
+    Index        int    `gorm:"index:idx_book_chapter,priority:2" json:"index"`
+    Title        string `gorm:"size:512" json:"title"`
+    Level        int    `gorm:"default:1" json:"level"` // 目录嵌套层级，1 = 顶级
+    // 电子书定位：二选一
+    Href         string `gorm:"size:1024" json:"href,omitempty"`      // EPUB zip 内条目路径
+    StartOffset  int64  `json:"start_offset"`                          // TXT 字节区间
+    EndOffset    int64  `json:"end_offset"`
+    // 漫画/PDF 定位
+    PageStart    int    `json:"page_start"`
+    PageEnd      int    `json:"page_end"`
+    CharCount    int    `json:"char_count"`
+}
+```
+
+### 3.3 进度、批注、收藏、统计
+
+```go
+// BookProgress 每个用户每本书一行（复合唯一键，仿 model.PlaybackHistory 的 uniq_user_history 模式）。
+type BookProgress struct {
+    Base
+    UserID       string `gorm:"uniqueIndex:uniq_user_book,priority:1;size:36;not null" json:"user_id"`
+    BookID       string `gorm:"uniqueIndex:uniq_user_book,priority:2;size:36;not null" json:"book_id"`
+    ChapterIndex int    `gorm:"default:0" json:"chapter_index"`
+    ChapterTitle string `gorm:"size:512" json:"chapter_title,omitempty"`
+    CharOffset   int    `json:"char_offset"`  // 章内字符偏移（电子书）
+    PageIndex    int    `json:"page_index"`   // 页码（漫画 / PDF）
+    Percent      float64 `json:"percent"`      // 全书百分比，书架进度条展示用
+    ScrollRatio  float64 `json:"scroll_ratio"` // 章内滚动比例，跨端还原更精确
+    ReaderMode   string `gorm:"size:16;default:scroll" json:"reader_mode"` // scroll / paged
+    Finished     bool   `gorm:"default:false" json:"finished"`
+    TotalSeconds int64  `json:"total_seconds"`
+    LastReadAt   time.Time `gorm:"index" json:"last_read_at"`
+}
+
+type BookAnnotation struct {
+    Base
+    UserID       string `gorm:"index:idx_book_anno,priority:1;size:36;not null" json:"user_id"`
+    BookID       string `gorm:"index:idx_book_anno,priority:2;size:36;not null" json:"book_id"`
+    ChapterIndex int    `json:"chapter_index"`
+    Type         string `gorm:"size:16;not null" json:"type"` // bookmark / highlight / note
+    StartOffset  int    `json:"start_offset"`
+    EndOffset    int    `json:"end_offset"`
+    SelectedText string `gorm:"size:2048" json:"selected_text,omitempty"`
+    Note         string `gorm:"type:text" json:"note,omitempty"`
+    Color        string `gorm:"size:16" json:"color,omitempty"`
+}
+
+type BookFavorite struct {
+    Base
+    UserID string `gorm:"uniqueIndex:uniq_user_book_fav,priority:1;size:36;not null" json:"user_id"`
+    BookID string `gorm:"uniqueIndex:uniq_user_book_fav,priority:2;size:36;not null" json:"book_id"`
+}
+
+// BookReadingSession 由前端心跳驱动，服务端按小时聚合，避免行数爆炸。
+type BookReadingSession struct {
+    Base
+    UserID      string    `gorm:"index:idx_book_stat,priority:1;size:36;not null" json:"user_id"`
+    BookID      string    `gorm:"index;size:36;not null" json:"book_id"`
+    BucketStart time.Time `gorm:"index:idx_book_stat,priority:2" json:"bucket_start"` // 截断到小时
+    Seconds     int64     `json:"seconds"`
+    CharsRead   int64     `json:"chars_read"`
+    PagesRead   int       `json:"pages_read"`
+}
+
+// BookDailyStat 每日汇总，供热力图与「年度阅读报告」查询，避免实时扫 session 表。
+type BookDailyStat struct {
+    Base
+    UserID  string `gorm:"uniqueIndex:uniq_user_book_daily,priority:1;size:36;not null" json:"user_id"`
+    Day     string `gorm:"uniqueIndex:uniq_user_book_daily,priority:2;size:10;not null" json:"day"` // YYYY-MM-DD
+    Seconds int64  `json:"seconds"`
+    Chars   int64  `json:"chars"`
+    Pages   int    `json:"pages"`
+    Books   int    `json:"books"` // 当日有阅读记录的书数
+}
+```
+
+### 3.4 关键取舍：正文不入库
+
+**决策：DB 只存章节索引（偏移量 / zip 内路径 / 页码区间），正文按需从源文件读取。**
+
+理由：
+1. 网文 TXT 常见 5–50MB，漫画单册 100–800MB。入库会让 SQLite 单文件膨胀到数十 GB，直接冲击 `docker-compose.simple.yml` 的「单文件数据库好备份」定位，也会拖慢全库 VACUUM / 备份 / 数据库迁移（`internal/service/database_admin.go`）。
+2. 源文件本来就是权威副本，重复存储没有收益。
+3. EPUB 与 CBZ 本质上都是 zip，**随机读取 zip 内单个条目成本极低**（读中央目录 + 解压目标条目），不需要把整本解压落盘。
+
+代价是每次打开章节都要读源文件。缓解手段：
+- 本地源：`os.Open` + `io.SectionReader`，代价可忽略。
+- 网盘源：见 4.3 的本地缓存策略，且对已缓存的章节走本地。
+
+### 3.5 用户级字段（挂在 `model.User` 上）
+
+沿用 `AllowedLibraryIDs` 的 JSON-in-text 模式（见 `internal/model/user.go`），**不复用影视库字段**，避免两个 ID 空间交叉：
+
+```go
+// 追加到 model.User
+ReaderSettings        string `gorm:"type:text" json:"-"`               // 阅读器偏好 JSON
+AllowedBookLibraryIDs string `gorm:"type:text" json:"-"`               // 空 = 不限制
+AllowedBookLibraryList []string `gorm:"-" json:"allowed_book_library_ids,omitempty"`
+```
+
+`ReaderSettings` 结构（前端读写，服务端仅透传与长度校验）：
+
+```json
+{
+  "mode": "scroll|paged",
+  "fontSize": 18,
+  "lineHeight": 1.8,
+  "fontFamily": "serif|sans|custom",
+  "contentWidth": 720,
+  "theme": "light|sepia|dark|black",
+  "pageAnimation": "slide|fade|none",
+  "comicLayout": "single|double|auto",
+  "comicDirection": "ltr|rtl",
+  "hideScrollbar": true
+}
+```
+
+放在 `User` 行内（而非新表）的理由：与 `PlayerVolume` / `DanmakuFontSize` 等既有播放器偏好一致，读取时随用户信息一并返回，无需额外查询。
+
+---
+
+## 4. 书源与内容读取管线
+
+### 4.1 扫描流程
+
+```
+POST /api/books/libraries/:id/scan
+  → BookScannerService.ScanLibrary(ctx, libraryID)
+      1. 置 ScanStatus=scanning，通过 SSEHub 广播进度（复用 service.SSEHub）
+      2. 遍历启用的 BookSource
+         - local: filepath.WalkDir，按扩展名白名单过滤，超过 Depth 停止递归
+         - cloud: cloud.New(cfg.Type, cfg, client).List(ctx, dirID) 递归列目录
+      3. 对每个候选文件调 BookParser.ParseMeta(reader) 拿元信息 + 目录
+      4. Upsert 到 books / book_chapters（source_id + source_path 为幂等键）
+      5. 源上已消失的书标记软删除（与影视库扫描语义保持一致）
+      6. 置 ScanStatus=idle，记录 LastScanAt
+```
+
+扩展名白名单：`.epub .txt .pdf .mobi .azw3 .cbz .cbr .zip .rar`（`.zip/.rar` 仅当目录内全是图片时按漫画处理，否则跳过，防止误吞压缩包）。
+
+并发：复用 `internal/service` 现有的 worker 池写法，默认 2–4 并发解析（解析要读文件，IO 密集）。
+
+### 4.2 各格式解析策略
+
+| 格式 | 元信息 | 章节 / 页 | 正文读取 |
+| --- | --- | --- | --- |
+| **EPUB** | zip → `META-INF/container.xml` → OPF → `dc:title/dc:creator/dc:language/dc:description`；封面取 OPF `meta[name=cover]` 指向项，退化到 `guide` | 按 spine 顺序，标题取每个 XHTML 的 `<title>` 或首个 `h1..h3`；`Level` 由 nav/ncx 的嵌套深度推断 | `archive/zip` 定位 `Href` 条目，读出 XHTML → 服务端清洗后返回 |
+| **TXT** | 文件名（`书名 - 作者.txt` 模式）+ 编码探测 | 正则切分：`第[一二三四五六七八九十百千零两0-9]+[章节卷回篇]`、`Chapter\s+\d+`、`^\s*\d+\s*$`；命中不足 3 个则按固定字节窗口切片 | `io.SectionReader` 读 `[StartOffset, EndOffset)` → 按探测到的编码转 UTF-8 |
+| **PDF** | 首页/元数据（页数、标题）；封面渲染首屏，失败则留空 | 单章「正文」，`PageStart/PageEnd` = 全书页 | 原始文件流（Range），前端 pdf.js 自己解析 |
+| **CBZ / CBR** | zip/rar 条目自然排序，第一张图做封面 | 单章，页区间 = 图片条目序号 | 按页解压单条目，走图片响应路径 |
+| **图片文件夹** | 目录名 | 单章，页区间 = 排序后图片序号 | 直接读本地文件 |
+| **MOBI / AZW3** | PalmDOC / KF8 头 | 首版**只入库展示、不支持在线阅读**，详情页给出「下载原文件」入口 | — |
+
+实现细节提示：
+- 编码探测用 `golang.org/x/text`（已是 `go.mod` 间接依赖）。GBK/Big5/UTF-16LE 都要覆盖，中文网文 TXT 大量是 GBK。
+- CBR 需要 RAR 解压。建议引入纯 Go 的 `github.com/nwaples/rardecode`；若不接受新依赖，首版把 CBR 归入「只入库、不可读」。
+- EPUB XHTML 清洗**必须在服务端做**：剔除 `<script>`、`on*` 事件属性、`<iframe>`、外部 `http(s)` 资源引用，把 `src/href` 重写为 `/api/books/:id/res/*`。前端再叠一层 DOMPurify 作为纵深防御。
+
+### 4.3 网盘书籍的读取策略
+
+网盘直链的核心约束：**EPUB / CBZ 的解析必须能读到文件尾部**（zip 中央目录在末尾），但 `cloud.Provider.Resolve` 返回的是短时效 URL，且 115 直链依赖 UA/Cookie（`DirectLink.Headers`），浏览器无法直接携带。
+
+因此分两条路径：
+
+**A. 解析阶段 —— 完整拉取到缓存目录**
+
+```
+<CacheDir>/books/<sourceID>/<hash>.<ext>
+```
+
+`BookParser` 通过 `DirectLink` 拉全量文件到缓存后再解析。缓存目录复用 `config.CacheConfig.CacheDir`（默认 `<DataDir>/cache`，容器里是 `/cache`），容量上限新加一项 `CacheConfig.BooksMaxSizeMB`（默认 2GB），走 LRU 淘汰。缓存命中的书后续正文读取也直接走本地，不再回网盘。
+
+**B. 阅读阶段 —— 优先本地缓存，未命中走代理流**
+
+未缓存时由服务端反代目标 URL（`DirectLink.Proxy=true` 时同样反代），并把 `Content-Type: image/*` 或 `application/pdf` 透传给前端。反代实现直接参照 `internal/service/cloud115_hls_proxy.go` 的响应头透传白名单（`Content-Type/Content-Length/Content-Range/Accept-Ranges/ETag/Last-Modified`）。
+
+**C. 阅读进度与文件解耦** —— 代码里区分「源」「位置」：
+
+```go
+type BookLocator struct {
+    Kind        string `json:"kind"` // local / cloud
+    LocalPath   string `json:"local_path,omitempty"`
+    CloudConfig string `json:"cloud_config,omitempty"`
+    CloudRef    string `json:"cloud_ref,omitempty"`
+    Href        string `json:"href,omitempty"` // zip 内条目
+    StartOffset int64  `json:"start_offset,omitempty"`
+    EndOffset   int64  `json:"end_offset,omitempty"`
+}
+```
+
+被缓存的书 `Kind` 仍报 `cloud`（进度不绑物理位置），这样缓存被淘汰后进度依然有效。这是不把 `Book.Path` 直接存成本地缓存路径的原因。
+
+### 4.4 磁盘与容器
+
+书籍目录需要在 compose 里挂载，并在 README 的部署档位表补充说明。新缓存目录复用现有 `MEBOX_CACHE_CACHE_DIR`（`docker-compose.simple.yml` 中为 `/cache`），无需新增环境变量。
+
+---
+
+## 5. HTTP API 设计
+
+全部挂在 `/api/books/*`，注册在 `registerAuthedBookRoutes`。响应统一走 `internal/handler/response.go` 的既有助手。
+
+### 5.1 书库与扫描（管理端）
+
+| 方法 | 路径 | 权限 | 说明 |
+| --- | --- | --- | --- |
+| GET | `/api/books/libraries` | `can_read_books` | 列表，按 `AllowedBookLibraryIDs` 过滤可见性 |
+| POST | `/api/books/libraries` | `can_manage_book_library` | 新建/更新书库 |
+| DELETE | `/api/books/libraries/:id` | `can_manage_book_library` | 删除（含级联软删 books） |
+| GET | `/api/books/libraries/:id/sources` | `can_manage_book_library` | 书源列表 |
+| POST | `/api/books/libraries/:id/sources` | `can_manage_book_library` | 新增书源（本地目录 / 网盘路径） |
+| POST | `/api/books/libraries/:id/scan` | `can_manage_book_library` | 触发扫描，返回 task id |
+| GET | `/api/books/scan/status` | `can_manage_book_library` | 扫描进度轮询 |
+| GET | `/api/books/browse` | `can_manage_book_library` | 网盘路径浏览（复用 cloud Provider.List） |
+
+### 5.2 书架与详情
+
+| 方法 | 路径 | 说明 |
+| --- | --- | --- |
+| GET | `/api/books` | 书架列表。参数：`library_id`、`keyword`、`media_kind`、`format`、`tag`、`sort`（`title/added_at/last_read/progress`）、`page/page_size` |
+| GET | `/api/books/continue-reading` | 最近在读，首页「继续阅读」区块用 |
+| GET | `/api/books/:id` | 详情（元信息 + 目录树 + 当前用户进度 + 收藏态） |
+| GET | `/api/books/:id/cover` | 封面。走 `ImageProxy` 的缓存与缩放，参数 `w` |
+| GET | `/api/books/:id/chapters/:index` | 章节正文。电子书返回 `text/html`（已清洗）或 `application/json` 结构化段落 |
+| GET | `/api/books/:id/res/*path` | EPUB 内部资源（图片/字体/CSS），路径参数为 zip 内条目 |
+| GET | `/api/books/:id/pages/:index` | 漫画单页图片，`Content-Type: image/*` + 长效缓存头 |
+| GET | `/api/books/:id/file` | 原始文件流（Range），pdf.js 与「下载原文件」共用 |
+| POST | `/api/books/:id/favorite` | 收藏 / 取消收藏 |
+| DELETE | `/api/books/:id` | 删除（`can_manage_books`） |
+
+**章节响应格式（推荐 JSON 而非裸 HTML）**：
+
+```json
+{
+  "index": 12,
+  "title": "第十二章 雨夜",
+  "char_count": 3820,
+  "blocks": [
+    { "type": "p", "text": "……" },
+    { "type": "img", "src": "/api/books/xxx/res/images/1.png" }
+  ],
+  "next_index": 13,
+  "prev_index": 11
+}
+```
+
+用结构化 blocks 而非 HTML 的理由：
+1. 前端可安全渲染，不必 `dangerouslySetInnerHTML`，彻底绕开 XSS 面。
+2. 分页模式需要按节点测量高度做分栏，结构化的段落数组比操作 DOM 简单得多。
+3. 字号/行距/主题切换只需重渲染，不碰 HTML 字符串。
+
+保底方案：`?format=html` 仍返回清洗后的 HTML，供 EPUB 中复杂排版（表格、脚注、双向文字）回退。
+
+### 5.3 进度、批注、统计
+
+| 方法 | 路径 | 说明 |
+| --- | --- | --- |
+| GET | `/api/books/:id/progress` | 当前用户进度 |
+| PUT | `/api/books/:id/progress` | 保存进度。前端**节流 5s + 页面卸载时 `sendBeacon`** |
+| GET | `/api/books/:id/annotations` | 批注列表 |
+| POST | `/api/books/:id/annotations` | 新建 |
+| PATCH | `/api/books/:id/annotations/:aid` | 修改 note / color |
+| DELETE | `/api/books/:id/annotations/:aid` | 删除 |
+| POST | `/api/books/:id/heartbeat` | 阅读心跳，body 为 `{seconds, chars, pages}`，落 `book_reading_sessions` 小时桶 |
+| GET | `/api/books/reader-settings` | 读取当前用户的阅读器偏好（`model.User.ReaderSettings`） |
+| PUT | `/api/books/reader-settings` | 保存阅读器偏好（服务端只做长度与枚举校验后原样存储） |
+| GET | `/api/books/stats` | 个人统计：累计时长、已读书数、在读、近 30 天热力图 |
+| GET | `/api/books/stats/overview` | 管理员视角：全站阅读统计（`can_access_settings`） |
+
+进度写入的并发安全：`uniq_user_book` 复合唯一键 + 先 `Updates` 再 `Create` 的 upsert 模式。**参照 `internal/database/schema_migration.go` 里 `dedupePlaybackHistories` 的前车之鉴**——`PlaybackHistory` 曾因 read-then-write 产生重复行导致唯一索引进不去，新表直接写 upsert，不要复制那个 bug。
+
+---
+
+## 6. 前端设计
+
+### 6.1 路由与导航
+
+`web/src/appRoutes.tsx` 新增懒加载路由：
+
+```tsx
+const BookshelfPage = lazy(() => import('./pages/BookshelfPage').then(m => ({ default: m.BookshelfPage })))
+const BookDetailPage = lazy(() => import('./pages/BookDetailPage').then(m => ({ default: m.BookDetailPage })))
+const BookReaderPage = lazy(() => import('./pages/BookReaderPage').then(m => ({ default: m.BookReaderPage })))
+const BookStatsPage  = lazy(() => import('./pages/BookStatsPage').then(m => ({ default: m.BookStatsPage })))
+const BookLibraryPage = lazy(() => import('./pages/BookLibraryPage').then(m => ({ default: m.BookLibraryPage })))
+```
+
+```
+/books                书架
+/books/:id            书籍详情（目录、元信息、开始阅读）
+/books/:id/read       阅读器（全屏，隐藏底栏）
+/books/stats          阅读统计
+/books/library        书库管理（adminOnly）
+```
+
+`web/src/components/layoutNavigation.ts` 的改动：
+
+- `MEDIA_NAV_ITEMS` 与 `MOBILE_BOTTOM_NAV_ITEMS` 加「阅读」项（`BookOpen` 图标），普通用户可见。
+- 新增 `isBookReaderRoute(pathname)`，并在 `shouldShowMobileBottomNav` 中排除 `/books/:id/read`，与 `isPlayerRoute` 的处理一致。
+- `resolveHeaderBack` 补 `/books/...` 的返回链。
+- `LAYOUT_NAV_ITEMS` 加「书库管理」条目，`adminOnly: true`。
+
+### 6.2 页面组成
+
+```
+web/src/pages/
+  BookshelfPage.tsx            书架：筛选栏 + 网格/列表双视图 + 继续阅读横滑
+  BookDetailPage.tsx           详情：封面、元信息、目录树、进度、开始/继续阅读
+  BookReaderPage.tsx           阅读器外壳：顶栏 + 内容区 + 底部工具条 + 设置抽屉
+  BookStatsPage.tsx            统计：热力图 + 概览卡片
+  BookLibraryPage.tsx          书库管理：书库 CRUD + 书源 CRUD + 扫描触发与进度
+web/src/components/
+  book/ReaderCore.tsx          渲染内核分发（按 media_kind + format）
+  book/ScrollReader.tsx        滚动流式
+  book/PagedReader.tsx         分页翻页
+  book/ComicReader.tsx         漫画（单页/双页/右开本/预加载）
+  book/PdfReader.tsx           PDF（pdf.js）
+  book/ReaderToolbar.tsx       顶栏 + 底栏（章节、进度、目录、批注、设置）
+  book/ReaderSettingsPanel.tsx 阅读设置
+  book/ChapterTocDrawer.tsx    目录抽屉
+  book/AnnotationList.tsx      书签笔记列表
+  book/ReaderProgressBar.tsx   进度条（可拖拽跳章）
+web/src/api/books.ts           接口封装
+```
+
+对于 4 类内容，`ReaderCore` 的分发是第一层决策：
+
+| `media_kind` | `format` | 内核 |
+| --- | --- | --- |
+| ebook | epub / txt | `ScrollReader` 或 `PagedReader`（按 `settings.mode`） |
+| ebook | pdf | `PdfReader` |
+| comic | cbz / cbr / folder | `ComicReader` |
+| ebook | mobi / azw3 | 不提供阅读，仅详情页 |
+
+### 6.3 渲染内核选型（关键决策）
+
+**结论：自研内核，不引入 epub.js / foliate-js。**
+
+对比：
+
+| 方案 | 优点 | 缺点 |
+| --- | --- | --- |
+| `epub.js` | 成熟、CFI 精确定位、多列分页开箱即用 | 维护停滞；CFI 定位难以与自研进度模型（`charOffset` / `percent`）对齐；PDF/漫画仍需另做两个内核 |
+| `foliate-js` | 一套 API 覆盖 EPUB/MOBI/CBZ/PDF，排版质量高 | 生态小、文档薄、非稳定 API，需要 vendored 一份并自行承担升级风险 |
+| **自研（推荐）** | 进度模型完全可控、跨端一致；零新增重依赖；与既有 Tailwind 主题体系天然统一 | 需要自己实现分页测量、脏 HTML 清洗、资源重写 |
+
+自研方案的可行性依据：分页的本质是「CSS multi-column 布局 + `transform: translateX` 平移」，foliate-js 也是这么做的，核心约 200 行；滚动模式的虚拟化可以直接复用已有的 `react-virtuoso`（已在 `web/package.json`，用于 `VirtualMediaGrid`）。
+
+自研必须做好的三件事：
+
+1. **HTML 清洗**：服务端为主（见 4.2），前端用 `dompurify` 兜底。这是新增的唯一运行时依赖。
+2. **资源重写**：EPUB 内部图片/字体/CSS 的 `src`、`href`、`url()` 全部重写到 `/api/books/:id/res/`，否则相对路径会 404。
+3. **分页测量与重排**：容器尺寸变化（窗口 resize、字号切换、横竖屏）后必须重新分页，并把「当前段落 + 段内比例」作为锚点恢复位置，不能让用户跳回章首。
+
+### 6.4 双模式实现
+
+**滚动模式（`ScrollReader`）**
+- 章内虚拟化：单章文本通常 2k–10k 字，直接整章渲染即可；跨章用「当前章 + 前后各一章」的窗口，滚动到边界时无缝追加。
+- 进度：`IntersectionObserver` 观测可视段落，映射为 `charOffset`；`scroll_ratio` 同时上报。
+- 优势：移动端体验好，实现简单，长段落无分页误差。
+
+**分页模式（`PagedReader`）**
+- 章内：容器设为多列（`column-width: <contentWidth>`），`overflow: hidden`，通过 `translateX` 翻页；总页数由 `scrollWidth / containerWidth` 得出。
+- 跨章：翻到本章末尾自动加载下一章首页；反向同理。章首/章尾需处理「残页合并」，避免出现半屏空白页。
+- 输入：左右方向键、空格、点击左右热区、滑动手势（移动端）。`comicDirection`/`pageAnimation` 控制方向与动画。
+- 进度：`chapter_index` + `page_index` 映射回 `charOffset`。
+
+两种模式共享 `BookProgress`，切换模式时用「章 + 比率」换算，不丢位置。
+
+### 6.5 状态与持久化
+
+- 阅读器设置来自 `authStore` 的用户信息（`ReaderSettings` 反序列化），改动后 `PUT /api/books/reader-settings` 持久化 + 本地 `localStorage` 兜底（首屏渲染不等接口）。
+- 进度本地先写 `localStorage`（key `mebook:book:<id>:pos`），再节流同步服务端；页面隐藏/卸载用 `navigator.sendBeacon` 保证不丢。
+- 新增 `web/src/stores/readerSettings.ts`（zustand），与既有 `playProfile.ts` 组织方式一致。
+
+---
+
+## 7. 权限与多用户
+
+`model.UserPermission` 新增 4 位（当前 18 个字段，加后 22 位）：
+
+| 权限位 | 默认 | 含义 |
+| --- | --- | --- |
+| `can_read_books` | `true` | 书架、阅读、进度、批注 |
+| `can_manage_book_library` | `false` | 书库 / 书源 CRUD、触发扫描、网盘浏览 |
+| `can_manage_books` | `false` | 编辑书籍元信息、删除书、手动重新解析 |
+| `can_view_book_stats` | `false` | 查看全站阅读统计 |
+
+同步改动清单（**漏一处就会出现「后端有权限、前端不显示开关」的静默 bug**）：
+
+1. `internal/model/permission.go` — 字段、`NewDefaultPermission()`、`PermissionMap()`，并更新文件头注释里的数量描述（注释目前写「19项」，实际 18 个字段，顺手修正）。
+2. `web/src/types/auth.ts` — `PermissionFlags` 接口加 4 个字段。
+3. `web/src/stores/permissions.ts` — 默认值对象、中文标签映射、权限分组数组。
+4. `web/src/hooks/usePermission.ts` — 若其中有分组注释需同步。
+5. `internal/handler/permissions.go` — 权限矩阵响应（若有枚举）。
+6. `web/src/pages/AdminUsersForm.tsx` / 权限勾选 UI — 若按分组硬编码了列表。
+
+书库可见性：
+
+- 管理员在用户管理页勾选该用户可访问的书库，写入 `User.AllowedBookLibraryIDs`。
+- 空值 = 不限制（与影视库语义一致）。
+- 过滤集中在一个 `bookVisibility` 助手，与现有的 `internal/handler/visibility.go` 并列（该文件就是影视库可见性的集中判定点，并且会与 `PlayProfile.AllowedLibraryIDs` 求交集）。**阅读模块首版不接播放配置档**——`PlayProfile` 是影视播放器概念（音量、转码参数、PIN），与阅读无关；但判定入口要与它放在同一层，将来若要按配置档限制书库才不用重构。
+- **服务端强制**：`GET /api/books/:id`、章节、页面、资源（`/res/*`）、封面、原始文件流，**所有**按 ID 取内容的接口都要校验 `book.LibraryID ∈ 用户可见书库`，不能只靠书架列表过滤。这是最容易漏的越权点：`/api/books/:id/res/*path` 会直接吐出书籍内部的原始资源，漏校验等于开放全库文件读取。
+- 用户被取消书库授权后，其 `BookProgress` / `BookAnnotation` 保留不删（授权恢复即恢复），但接口一律按当前可见性判定，不因历史数据放行。
+
+---
+
+## 8. 阅读统计
+
+- **采集**：阅读器每 30s 发一次 `heartbeat`，卸载时补发一次；服务端按 `(user_id, book_id, 小时桶)` 累加，行数上限 = 用户数 × 书数 × 阅读小时数，可控。
+- **汇总**：`Scheduler` 每日 03:00 把昨天之前的 session 滚进 `BookDailyStat`（复用 `service.Scheduler` 的既有定时任务注册方式）。
+- **展示**：
+  - 个人页「阅读统计」：累计时长、读完本数、在读本数、近 30 天热力图（仿 GitHub 贡献图）、阅读类型分布（电子书 / 漫画）。
+  - 首页新增「继续阅读」横滑区块（参照 `HomePageSections.tsx` 里既有区块的写法）。
+  - 管理员视图：全站活跃度、热门书籍 Top 20（需 `can_view_book_stats`）。
+
+隐私：统计仅对本人与管理员可见；管理员视图只出聚合数据，不暴露单个用户的阅读内容。
+
+---
+
+## 9. 分期实施计划
+
+### P0 — 端到端可用（本地书库 / EPUB + TXT 电子书）
+
+目标：能扫库、能在网页上把一本书读完、关掉浏览器再打开能续读。
+
+| # | 交付物 |
+| --- | --- |
+| 1 | `internal/model/book.go` 九张表 + `AllModels()` 注册 + 迁移验证（SQLite 与 PostgreSQL 各跑一次升级） |
+| 2 | `BookLibrary` / `BookSource` / `Book` / `BookChapter` / `BookProgress` 的 repository |
+| 3 | `BookParser`：EPUB 与 TXT 解析（含 GBK 编码探测、章节正则切分、封面提取） |
+| 4 | `BookScannerService`：本地目录扫描 + upsert + 进度广播 |
+| 5 | API：书库 CRUD、书源 CRUD、扫描、书架列表、详情、章节正文、封面、进度读写、阅读器偏好读写 |
+| 6 | 前端：`BookshelfPage`、`BookDetailPage`、`BookReaderPage`（仅滚动模式）、目录抽屉、阅读设置面板 |
+| 7 | 权限：4 个权限位 + `AllowedBookLibraryIDs` 全链路（含服务端越权校验） |
+
+**验收标准**
+- 一个含 50 本 EPUB 与 20 本 GBK 编码 TXT 的目录，扫描后书架正确列出，标题/作者/封面/章节目录无误。
+- 任意一本书可连续阅读 3 章以上，刷新页面后回到原位置（误差 < 1 段）。
+- 权限为 `can_read_books=false` 的账号访问 `/api/books` 返回 403；直接请求他人书库的 `/api/books/:id/chapters/0`、`/api/books/:id/res/*`、`/api/books/:id/file` 同样被拒。
+- SQLite 单文件档与 PostgreSQL 档都能从旧版本升级启动，无迁移报错。
+
+### P1 — 漫画 + 分页模式 + 网盘直链
+
+| # | 交付物 |
+| --- | --- |
+| 1 | `ComicReader`：CBZ 解析、单页/双页、右开本、相邻页预加载 |
+| 2 | `PagedReader`：分页测量、resize 重排、跨章衔接、键鼠与手势输入 |
+| 3 | 网盘书源：`StorageKind=cloud` 的书源配置、`cloud.Provider` 接入、本地缓存目录 + LRU 淘汰 |
+| 4 | 网盘书籍的索引拉取与阅读反代（含 `Content-Range` 透传） |
+| 5 | PDF：`PdfReader`（pdf.js）+ Range 文件流接口 |
+| 6 | 图片文件夹型漫画 |
+
+**验收标准**
+- CBZ 单册 300 页可流畅翻阅，双页模式断页处理正确（避免跨章错配）。
+- 分页模式下切换字号、resize 窗口、手机横竖屏切换后，位置不跳、不出现空白页。
+- 挂在 OpenList 与 115 上的 EPUB 能正常入库并在线阅读，缓存目录达到上限后按 LRU 淘汰且不影响已有进度。
+- 20MB 以上 PDF 可跳页、可缩放。
+
+### P2 — 批注、统计与体验打磨
+
+| # | 交付物 |
+| --- | --- |
+| 1 | 划线 / 书签 / 笔记：`BookAnnotation` 接口与 UI，批注列表与跳转 |
+| 2 | 阅读统计：心跳采集、每日汇总任务、个人统计页、首页「继续阅读」区块 |
+| 3 | 管理员统计视图 + 热门书籍排行 |
+| 4 | 书库定时扫描（`Scheduler` 接入，默认关闭） |
+| 5 | 书架高级筛选与排序、合集（系列）聚合视图 |
+| 6 | MOBI/AZW3 元信息解析（仍不做在线阅读，仅提供下载） |
+| 7 | 部署文档与 compose 注释更新（书籍目录挂载说明） |
+
+### P3 — 可选增强
+听书 TTS、跨设备批注冲突合并、书源自动整理（仿 `OrganizerService`）、EPUB 阅读器内注释锚点高亮。
+
+---
+
+## 10. 风险与待拍板项
+
+### 10.1 需要你拍板的两点
+
+**① 网盘书籍的缓存策略**
+- 选项 A（本方案）：索引时完整下载到缓存目录，阅读时优先本地。省流量、体验好，但全新书首次打开有等待，且占用磁盘（默认 2GB 上限）。
+- 选项 B：完全不落盘，每次按 Range/整文件从网盘拉。省磁盘，但每次打开都要重新下载，网盘限速时体验很差。
+- 选项 C：折中——只对 EPUB/CBZ 缓存（解析必须读全文），漫画原图与 PDF 走流式。
+
+我的建议是 **C**，因为它把「必须落盘」和「可以不落盘」分开了。
+
+**② 章节正文的返回格式**
+- JSON blocks（本方案推荐）：安全、便于分页测量，但复杂 EPUB 排版（表格、脚注、竖排）会降级。
+- 清洗后 HTML：保真度高，但前端要 `dangerouslySetInnerHTML`，XSS 面更大。
+- 我的建议是 **JSON blocks 为主 + `?format=html` 回退**，两者都实现，前端在遇到 `type: "html-block"` 时回退渲染。
+
+### 10.2 技术风险
+
+| 风险 | 影响 | 缓解 |
+| --- | --- | --- |
+| 自研分页内核的边界情况多（残页、跨章、RTL、竖排） | P1 可能超期 | P0 先只做滚动模式；分页单独立项，配套 `playerPageModel.test.ts` 那样的单测 |
+| TXT 章节正则对网文变体覆盖不足 | 目录错乱 | 提供「手动重新切分」入口，规则可配（仿 `RecognitionWordsPanel` 的可配置词表模式） |
+| 网盘直链失效 / 限速 / 防盗链 | 阅读中断 | 复用现有 115 换链与 `url_cache.go` 的缓存机制；失败时前端降级为「下载原文件」 |
+| 大 TXT（>50MB）章节表行数过多 | SQLite 写入慢 | 章节超过阈值（如 5000 章）时按固定窗口粗切，或改为「按需切分 + 缓存到章节表」的惰性策略 |
+| 缓存目录膨胀 | 磁盘打满 | 容量上限 + 复用 `service.PruneImageCache` 的 LRU 清理 + 系统设置页可见 |
+| 数据库迁移对老库不兼容 | 升级失败 | 新表全部是纯新增，无列变更；不触碰 `ensurePostgresColumnCompatibility` 的既有语句 |
+
+### 10.3 不引入的新依赖清单
+
+| 依赖 | 用途 | 取舍 |
+| --- | --- | --- |
+| `dompurify` | 前端 HTML 清洗兜底 | **建议引入**（前端必需） |
+| `pdfjs-dist` | PDF 渲染 | **建议引入**（P1） |
+| `github.com/nwaples/rardecode` | CBR 解压 | 可选；不接受则 CBR 首版只入库 |
+| `epub.js` / `foliate-js` | EPUB 渲染 | **不引入**，见 6.3 |
+
+---
+
+## 11. 测试策略
+
+与项目现有测试密度对齐（`internal/service` 下大量 `_test.go`，前端有 `*.test.ts`）：
+
+**后端**
+- `book_parser_test.go`：EPUB / TXT 各准备 fixture（`testdata/` 下小体积样本），断言元信息、章节数、章节边界字节偏移、GBK 转码正确性。
+- `book_scanner_test.go`：临时目录扫描 + 重复扫描幂等 + 源文件删除后软删。
+- `book_progress_test.go`：并发 upsert 不产生重复行（直接复现 `dedupePlaybackHistories` 防的那类 bug）。
+- `book_permission_test.go`：越权矩阵，逐接口断言非可见书库返回 403/404。
+- Handler 层：仿 `internal/handler/media_test.go` 起的 `httptest` + 真实内存 SQLite。
+
+**前端**
+- `readerModel.test.ts`：模式切换时的位置换算（`scroll ↔ paged`、`charOffset ↔ pageIndex`）、百分比计算、跨章边界。
+- 分页计算的纯函数抽出单测（不含 DOM），参照 `web/src/pages/playerPageModel.test.ts` 的做法——把逻辑从组件里拔出来测，是项目已有的好传统。
+
+---
+
+## 12. 附：改动文件清单
+
+**后端新增**
+```
+internal/model/book.go
+internal/repository/book_repository.go
+internal/service/book_parser.go            EPUB / TXT / CBZ 解析
+internal/service/book_parser_epub.go
+internal/service/book_parser_txt.go
+internal/service/book_parser_comic.go
+internal/service/book_scanner.go
+internal/service/book_reader.go            章节 / 页面 / 资源的读取与清洗
+internal/service/book_progress.go
+internal/service/book_stats.go
+internal/service/book_cache.go             网盘缓存与 LRU
+internal/handler/books.go
+internal/handler/books_library.go
+internal/handler/books_reader.go
+internal/handler/routes_books.go
+```
+
+**后端修改**
+```
+internal/model/model.go                    AllModels() 追加 9 张表
+internal/model/permission.go               4 个权限位
+internal/model/user.go                     ReaderSettings / AllowedBookLibraryIDs
+internal/repository/repository.go          Container 加字段
+internal/service/service.go                Container 加字段 + Boot() 启动扫描
+internal/handler/routes_authenticated.go   挂载 registerAuthedBookRoutes
+internal/service/scheduler_local_jobs.go   书籍缓存清理 + 每日阅读统计汇总
+internal/config/types.go                   CacheConfig 加 BooksMaxSizeMB；新增 BookConfig（扫描并发等）
+docker-compose*.yml                        书籍目录挂载注释
+README.md / README_EN.md                   能力表新增「阅读」
+```
+
+**前端新增**
+```
+web/src/api/books.ts
+web/src/stores/readerSettings.ts
+web/src/pages/BookshelfPage.tsx
+web/src/pages/BookDetailPage.tsx
+web/src/pages/BookReaderPage.tsx
+web/src/pages/BookStatsPage.tsx
+web/src/pages/BookLibraryPage.tsx
+web/src/components/book/*.tsx
+```
+**前端修改**
+```
+web/src/appRoutes.tsx                       4 条路由
+web/src/components/layoutNavigation.ts      导航项、阅读器路由判定、返回链
+web/src/types/auth.ts                       权限位
+web/src/stores/permissions.ts               权限位默认值 / 标签 / 分组
+web/src/pages/HomePageSections.tsx          「继续阅读」区块
+web/src/pages/settingsGroupBooks.ts         （新增）阅读设置分组
+web/src/pages/settingsGroups.ts             把 settingsGroupBooks 加入 GROUPS 数组
+```

internal/config/defaults.go

@@ -44,9 +44,10 @@ func setDefaults(v *viper.Viper) {
 	v.SetDefault("logging.max_backups", 10)
 
 	v.SetDefault("cache.cache_dir", "./cache")
-	v.SetDefault("cache.images_max_size_mb", 500)
+	v.SetDefault("cache.images_max_size_mb", 2000)
+	v.SetDefault("cache.images_originals_max_size_mb", 256)
+	v.SetDefault("cache.images_originals_ttl_hours", 168)
 	v.SetDefault("cache.memory_max_size_mb", DefaultCacheMemoryMaxSizeMB)
-	v.SetDefault("cache.cleanup_interval_min", 60)
 	v.SetDefault("cache.redis_url", "")
 	v.SetDefault("cache.redis_prefix", "mebox")
 	v.SetDefault("cache.media_ttl_seconds", 90)

internal/config/normalize.go

@@ -47,6 +47,12 @@ func (c *Config) normalize() error {
 	if c.Cache.ImagesMaxSizeMB < 0 {
 		c.Cache.ImagesMaxSizeMB = 0
 	}
+	if c.Cache.ImagesOriginalsMaxSizeMB < 0 {
+		c.Cache.ImagesOriginalsMaxSizeMB = 0
+	}
+	if c.Cache.ImagesOriginalsTTLHours < 0 {
+		c.Cache.ImagesOriginalsTTLHours = 0
+	}
 	if c.Cache.MemoryMaxSizeMB <= 0 {
 		c.Cache.MemoryMaxSizeMB = DefaultCacheMemoryMaxSizeMB
 	}

internal/config/types.go

@@ -117,15 +117,17 @@ type LoggingConfig struct {
 type CacheConfig struct {
 	CacheDir        string `mapstructure:"cache_dir"`
 	ImagesMaxSizeMB int    `mapstructure:"images_max_size_mb"`
+	// ImagesOriginalsMaxSizeMB 限制 cache/images 下原图子配额（MB）。原图只是
+	// 生成缩略图的原料，丢了可以重新回源；0 表示不单独限制（只受总量约束）。
+	ImagesOriginalsMaxSizeMB int `mapstructure:"images_originals_max_size_mb"`
+	// ImagesOriginalsTTLHours 是原图的保留时长（小时），超过即删除；0 表示
+	// 不按时间淘汰。缩短它能让原图缓存保持在一个滚动窗口内。
+	ImagesOriginalsTTLHours int `mapstructure:"images_originals_ttl_hours"`
 	// MemoryMaxSizeMB 限制进程内 L1 缓存总字节数，JSON/对象缓存共用该预算。
-	MemoryMaxSizeMB    int    `mapstructure:"memory_max_size_mb"`
-	MaxDiskUsageMB     int    `mapstructure:"max_disk_usage_mb"`
-	TTLHours           int    `mapstructure:"ttl_hours"`
-	AutoCleanup        bool   `mapstructure:"auto_cleanup"`
-	CleanupIntervalMin int    `mapstructure:"cleanup_interval_min"`
-	RedisURL           string `mapstructure:"redis_url"`
-	RedisPrefix        string `mapstructure:"redis_prefix"`
-	MediaTTLSeconds    int    `mapstructure:"media_ttl_seconds"`
+	MemoryMaxSizeMB int    `mapstructure:"memory_max_size_mb"`
+	RedisURL        string `mapstructure:"redis_url"`
+	RedisPrefix     string `mapstructure:"redis_prefix"`
+	MediaTTLSeconds int    `mapstructure:"media_ttl_seconds"`
 	// EmbyLatestTTLSeconds 是 Emby「最新添加」(Items/Latest) 的缓存时长。
 	// 客户端刷新首页时会并发请求全部媒体库的 Latest（生产环境观察到 73 个
 	// 并发），缓存过短会让这批请求同时穿透并各自重建 payload，在低配主机

internal/handler/admin_settings.go

@@ -65,8 +65,10 @@ func updateSettingHandler(svc *service.Container) gin.HandlerFunc {
 		if req.Key == "transcode.hw_enabled" || req.Key == "transcode.hw_accel" || req.Key == "transcoder.hardware_accel" || req.Key == "transcoder.encoder" {
 			svc.Transcoder.StopAll()
 		}
-		if req.Key == "cache.images_max_size_mb" && svc.Scheduler != nil {
-			_ = svc.Scheduler.RunNowAsync(c.Request.Context(), "image_cache_cleanup")
+		if req.Key == "cache.images_max_size_mb" || req.Key == "cache.images_originals_max_size_mb" || req.Key == "cache.images_originals_ttl_hours" {
+			if svc.Scheduler != nil {
+				_ = svc.Scheduler.RunNowAsync(c.Request.Context(), "image_cache_cleanup")
+			}
 		}
 		if req.Key == "cache.memory_max_size_mb" && svc.Cache != nil {
 			svc.Cache.SetMaxSizeMB(svc.Cfg.Cache.MemoryMaxSizeMB)

internal/handler/emby_items_handlers.go

@@ -40,9 +40,28 @@ func parseEmbyItemsParams(c *gin.Context) service.ItemsParams {
 		SortOrder:        firstQueryValue(c, "SortOrder", "sortOrder", "sortorder"),
 		Limit:            limit,
 		StartIndex:       offset,
+		SeasonIndex:      parseEmbySeasonIndexQuery(c),
 	}
 }
 
+// parseEmbySeasonIndexQuery 读取客户端请求的季序号。
+//
+// Emby 客户端有两种表达方式：SeasonId（虚拟季 ID）与 Season / SeasonIndex
+// （季序号，特别篇为 0）。两者都是合法入参，SeasonId 更精确。这里只解析季序号，
+// 返回 nil 表示客户端没有按季过滤（区别于 Season=0 的特别篇）。
+func parseEmbySeasonIndexQuery(c *gin.Context) *int {
+	raw := firstQueryValue(c, "Season", "season", "SeasonIndex", "seasonIndex", "seasonindex")
+	if raw == "" {
+		return nil
+	}
+	value, err := strconv.Atoi(raw)
+	if err != nil {
+		// 客户端偶尔传入季名称之类的非数字值；按「未过滤」处理，避免整季空结果。
+		return nil
+	}
+	return &value
+}
+
 func embyFirstNonEmptyString(values ...string) string {
 	for _, value := range values {
 		if strings.TrimSpace(value) != "" {
@@ -199,8 +218,15 @@ func embyShowSeasonsHandler(svc *service.Container) gin.HandlerFunc {
 func embyShowEpisodesHandler(svc *service.Container) gin.HandlerFunc {
 	return func(c *gin.Context) {
 		parentID := firstQueryValue(c, "SeasonId", "seasonId")
+		// 客户端常用季序号而不是虚拟季 ID 请求剧集。缺了这层过滤，
+		// /Shows/{id}/Episodes?Season=2 会把整部剧的所有季都返回。
+		seasonIndex := parseEmbySeasonIndexQuery(c)
 		if parentID == "" {
 			parentID = c.Param("id")
+		} else {
+			// SeasonId 已经限定了具体季，忽略同时传来的季序号，避免两者
+			// 不一致时把结果过滤成空集。
+			seasonIndex = nil
 		}
 		params := service.ItemsParams{
 			UserID:           embyEffectiveUserID(c),
@@ -208,6 +234,7 @@ func embyShowEpisodesHandler(svc *service.Container) gin.HandlerFunc {
 			IncludeItemTypes: []string{"Episode"},
 			Recursive:        true,
 			Limit:            500,
+			SeasonIndex:      seasonIndex,
 		}
 		out, err := svc.Emby.Items(c.Request.Context(), params)
 		if err != nil {

internal/handler/emby_show_episodes_season_test.go

@@ -0,0 +1,211 @@
+package handler
+
+import (
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"github.com/gin-gonic/gin"
+	"github.com/glebarez/sqlite"
+	"go.uber.org/zap"
+	"gorm.io/gorm"
+
+	"github.com/truewhile/MeBox/internal/config"
+	"github.com/truewhile/MeBox/internal/model"
+	"github.com/truewhile/MeBox/internal/repository"
+	"github.com/truewhile/MeBox/internal/service"
+)
+
+// embySeasonEpisodesRouter builds an Emby-compatible router over one series with
+// two seasons, so /Shows/{id}/Episodes can be exercised with the query forms
+// real clients send.
+func embySeasonEpisodesRouter(t *testing.T) (*gin.Engine, string, string) {
+	t.Helper()
+	gin.SetMode(gin.TestMode)
+	db, err := gorm.Open(sqlite.Open(":memory:"), &gorm.Config{})
+	if err != nil {
+		t.Fatalf("open db: %v", err)
+	}
+	if err := db.AutoMigrate(&model.User{}, &model.Library{}, &model.Series{}, &model.Media{}, &model.Favorite{}, &model.PlaybackHistory{}, &model.Setting{}); err != nil {
+		t.Fatalf("migrate: %v", err)
+	}
+	if sqlDB, err := db.DB(); err == nil {
+		sqlDB.SetMaxOpenConns(1)
+	}
+	repos := repository.New(db)
+	if err := repos.User.Create(t.Context(), &model.User{
+		Base:         model.Base{ID: "user-1"},
+		Username:     "tester",
+		PasswordHash: "x",
+		Role:         "admin",
+		Tier:         "plus",
+		IsActive:     true,
+	}); err != nil {
+		t.Fatalf("create user: %v", err)
+	}
+	lib := model.Library{Name: "剧集", Path: "D:\\media\\tv", Type: "tv", Enabled: true}
+	if err := repos.Library.Create(t.Context(), &lib); err != nil {
+		t.Fatalf("create library: %v", err)
+	}
+	for _, m := range []model.Media{
+		{
+			Base:       model.Base{ID: "s1e1"},
+			LibraryID:  lib.ID,
+			Title:      "Test Show",
+			Path:       "D:\\media\\tv\\Test Show\\Season 01\\Test Show - S01E01.mkv",
+			SeasonNum:  1,
+			EpisodeNum: 1,
+			Container:  "mkv",
+		},
+		{
+			Base:       model.Base{ID: "s2e1"},
+			LibraryID:  lib.ID,
+			Title:      "Test Show",
+			Path:       "D:\\media\\tv\\Test Show\\Season 02\\Test Show - S02E01.mkv",
+			SeasonNum:  2,
+			EpisodeNum: 1,
+			Container:  "mkv",
+		},
+	} {
+		if err := db.Create(&m).Error; err != nil {
+			t.Fatalf("create media: %v", err)
+		}
+	}
+
+	const secret = "test-secret"
+	router := gin.New()
+	registerEmbyRoutes(router, secret, &service.Container{
+		Repo: repos,
+		Emby: service.NewEmbyService(&config.Config{}, zap.NewNop(), repos),
+	})
+
+	// Resolve the virtual series id the way a client would.
+	seriesReq := httptest.NewRequest(http.MethodGet, "/Items?ParentId="+lib.ID+"&IncludeItemTypes=Series", nil)
+	seriesReq.Header.Set("X-Emby-Token", signedTestToken(t, secret))
+	seriesRec := httptest.NewRecorder()
+	router.ServeHTTP(seriesRec, seriesReq)
+	if seriesRec.Code != http.StatusOK {
+		t.Fatalf("series lookup status=%d body=%s", seriesRec.Code, seriesRec.Body.String())
+	}
+	var seriesPayload struct {
+		Items []map[string]any `json:"Items"`
+	}
+	if err := json.Unmarshal(seriesRec.Body.Bytes(), &seriesPayload); err != nil {
+		t.Fatalf("decode series: %v", err)
+	}
+	if len(seriesPayload.Items) != 1 {
+		t.Fatalf("want one series, got %#v", seriesPayload.Items)
+	}
+	seriesID, _ := seriesPayload.Items[0]["Id"].(string)
+	if seriesID == "" {
+		t.Fatalf("series payload has no Id: %#v", seriesPayload.Items[0])
+	}
+	return router, secret, seriesID
+}
+
+func fetchEpisodeIDs(t *testing.T, router *gin.Engine, secret, path string) ([]string, int) {
+	t.Helper()
+	req := httptest.NewRequest(http.MethodGet, path, nil)
+	req.Header.Set("X-Emby-Token", signedTestToken(t, secret))
+	rec := httptest.NewRecorder()
+	router.ServeHTTP(rec, req)
+	if rec.Code != http.StatusOK {
+		t.Fatalf("%s status=%d body=%s", path, rec.Code, rec.Body.String())
+	}
+	var payload struct {
+		Items            []map[string]any `json:"Items"`
+		TotalRecordCount float64          `json:"TotalRecordCount"`
+	}
+	if err := json.Unmarshal(rec.Body.Bytes(), &payload); err != nil {
+		t.Fatalf("decode %s: %v", path, err)
+	}
+	ids := make([]string, 0, len(payload.Items))
+	for _, item := range payload.Items {
+		id, _ := item["Id"].(string)
+		ids = append(ids, id)
+	}
+	return ids, int(payload.TotalRecordCount)
+}
+
+// TestEmbyShowEpisodesHonoursSeasonQueryParam is the client-facing regression
+// test: Emby clients scope episodes with ?Season=<index> (not only SeasonId), and
+// the endpoint used to ignore it and return every season of the series.
+func TestEmbyShowEpisodesHonoursSeasonQueryParam(t *testing.T) {
+	router, secret, seriesID := embySeasonEpisodesRouter(t)
+
+	cases := []struct {
+		name    string
+		query   string
+		wantIDs []string
+	}{
+		{name: "season 1", query: "Season=1", wantIDs: []string{"s1e1"}},
+		{name: "season 2", query: "Season=2", wantIDs: []string{"s2e1"}},
+		{name: "season index alias", query: "SeasonIndex=2", wantIDs: []string{"s2e1"}},
+		{name: "no season returns all", query: "", wantIDs: []string{"s1e1", "s2e1"}},
+		{name: "unknown season is empty", query: "Season=9", wantIDs: []string{}},
+		{name: "non numeric season falls back to all", query: "Season=Specials", wantIDs: []string{"s1e1", "s2e1"}},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			path := "/Shows/" + seriesID + "/Episodes"
+			if tc.query != "" {
+				path += "?" + tc.query
+			}
+			ids, total := fetchEpisodeIDs(t, router, secret, path)
+			if len(ids) != len(tc.wantIDs) {
+				t.Fatalf("episode ids = %#v (total=%d), want %#v", ids, total, tc.wantIDs)
+			}
+			for i := range tc.wantIDs {
+				if ids[i] != tc.wantIDs[i] {
+					t.Fatalf("episode ids = %#v, want %#v", ids, tc.wantIDs)
+				}
+			}
+			if total != len(tc.wantIDs) {
+				t.Fatalf("TotalRecordCount = %d, want %d", total, len(tc.wantIDs))
+			}
+		})
+	}
+}
+
+// TestEmbyShowEpisodesSeasonIdStillWins pins that the virtual SeasonId form keeps
+// working, and that a conflicting Season query cannot empty it out.
+func TestEmbyShowEpisodesSeasonIdStillWins(t *testing.T) {
+	router, secret, seriesID := embySeasonEpisodesRouter(t)
+
+	seasonsPath := "/Shows/" + seriesID + "/Seasons"
+	req := httptest.NewRequest(http.MethodGet, seasonsPath, nil)
+	req.Header.Set("X-Emby-Token", signedTestToken(t, secret))
+	rec := httptest.NewRecorder()
+	router.ServeHTTP(rec, req)
+	if rec.Code != http.StatusOK {
+		t.Fatalf("seasons status=%d body=%s", rec.Code, rec.Body.String())
+	}
+	var seasons struct {
+		Items []map[string]any `json:"Items"`
+	}
+	if err := json.Unmarshal(rec.Body.Bytes(), &seasons); err != nil {
+		t.Fatalf("decode seasons: %v", err)
+	}
+	var season1ID string
+	for _, s := range seasons.Items {
+		if index, ok := s["IndexNumber"].(float64); ok && int(index) == 1 {
+			season1ID, _ = s["Id"].(string)
+		}
+	}
+	if season1ID == "" {
+		t.Fatalf("season 1 not found in %#v", seasons.Items)
+	}
+
+	ids, _ := fetchEpisodeIDs(t, router, secret, "/Shows/"+seriesID+"/Episodes?SeasonId="+season1ID)
+	if len(ids) != 1 || ids[0] != "s1e1" {
+		t.Fatalf("SeasonId episodes = %#v, want [s1e1]", ids)
+	}
+
+	// A stale/conflicting season number must not override the explicit season id.
+	ids, _ = fetchEpisodeIDs(t, router, secret, "/Shows/"+seriesID+"/Episodes?SeasonId="+season1ID+"&Season=2")
+	if len(ids) != 1 || ids[0] != "s1e1" {
+		t.Fatalf("SeasonId+Season episodes = %#v, want [s1e1]", ids)
+	}
+}

internal/service/cache_cleanup.go

@@ -4,7 +4,7 @@
 package service
 
 import (
-	"errors"
+	"encoding/hex"
 	"os"
 	"path/filepath"
 	"sort"
@@ -55,18 +55,287 @@ type PruneImageCacheResult struct {
 	RemainingBytes   int64
 }
 
+// imageCacheFileEntry 是池内一个可被淘汰的缓存文件。
 type imageCacheFileEntry struct {
 	path    string
+	pool    int
 	size    int64
 	modTime time.Time
 }
 
-// PruneImageCache scans imagesDir for cached image files. If the total disk usage
-// exceeds maxSizeBytes, it removes files starting from the oldest (by ModTime)
-// until disk usage falls to or below targetSizeBytes (80% of maxSizeBytes).
+// ImageCacheScope 是池内的一处文件范围。Recursive=false 时只处理该目录下的
+// 直属文件（历史版本的图片缓存是平铺在 images/ 下的，升级后仍要能清理掉）。
+type ImageCacheScope struct {
+	Root      string
+	Recursive bool
+	// DropUnknownResizeVariants 只用于派生缩略图目录：直接删除文件名不符合当前
+	// 命名（<源图键>.<宽>x<高>q<质量>.img）的历史缩放缓存。它们在新的查找路径下
+	// 永远不会被命中，留着只会占用总量配额、把有用缓存挤出去。
+	DropUnknownResizeVariants bool
+}
+
+// ImageCachePool 描述一个图片缓存池：范围 + 自身配额 + 保留时长。
 //
-// In-flight temporary files (*.tmp) are skipped to avoid corrupting concurrent writes.
-// Empty subdirectories left behind are best-effort removed.
+// 分池的意义在于两类缓存的“可再生成本”完全不同：原图只是生成缩略图的原料，
+// 丢了可以重新回源；派生缩略图是客户端热路径真正读取的成品，重新生成代价高。
+// 因此原图配小配额、短保留，总量超限时也优先淘汰原图。
+type ImageCachePool struct {
+	Name     string
+	Scopes   []ImageCacheScope
+	MaxBytes int64         // 0 = 不单独限制
+	MaxAge   time.Duration // 0 = 不按时间淘汰
+}
+
+// imageFailMarkerMaxAge 是失败标记文件的保留时长：标记只用于观测/重试，
+// 过期即视为噪音清掉（不会因为标记残留而阻止重新抓取）。
+const imageFailMarkerMaxAge = 24 * time.Hour
+
+// tempImageCacheFile 报告文件名是否是下载/缩放过程中的临时文件。它们正在被
+// 并发写入，清理时必须跳过。
+func tempImageCacheFile(name string) bool {
+	if strings.HasSuffix(name, ".tmp") {
+		return true
+	}
+	return strings.HasPrefix(name, "img-") && strings.Contains(name, ".tmp")
+}
+
+// sha256HexLength 是十六进制 sha256 摘要的长度。
+const sha256HexLength = 64
+
+// isCurrentResizeVariantName 报告文件名是否符合当前的缩放缓存命名
+// （<源图键>.<宽>x<高>q<质量>.img）。
+func isCurrentResizeVariantName(name string) bool {
+	if !strings.HasSuffix(name, ".img") {
+		return false
+	}
+	spec := strings.TrimSuffix(name, ".img")
+	key, _, found := strings.Cut(spec, ".")
+	if !found || len(key) != sha256HexLength {
+		return false
+	}
+	if _, err := hex.DecodeString(key); err != nil {
+		return false
+	}
+	_, _, ok := parseResizeVariantSuffix(name, key)
+	return ok
+}
+
+// collectImageCachePoolFiles 收集池内可淘汰的文件（跳过并发写入中的临时文件）。
+// 失败标记单独返回，它们不计入容量，只按年龄清理。
+func collectImageCachePoolFiles(pool ImageCachePool, poolIndex int, entries *[]imageCacheFileEntry, failMarkers *[]imageCacheFileEntry) {
+	visit := func(path string, info os.FileInfo, scope ImageCacheScope) {
+		if info.IsDir() {
+			return
+		}
+		name := info.Name()
+		if tempImageCacheFile(name) {
+			return
+		}
+		if scope.DropUnknownResizeVariants && strings.HasSuffix(name, ".img") && !isCurrentResizeVariantName(name) {
+			_ = os.Remove(path) // 旧版命名，永远不会命中，直接清掉
+			return
+		}
+		entry := imageCacheFileEntry{path: path, pool: poolIndex, size: info.Size(), modTime: info.ModTime()}
+		if strings.HasSuffix(name, ".fail") {
+			*failMarkers = append(*failMarkers, entry)
+			return
+		}
+		*entries = append(*entries, entry)
+	}
+	for _, scope := range pool.Scopes {
+		if strings.TrimSpace(scope.Root) == "" {
+			continue
+		}
+		if !scope.Recursive {
+			dir, err := os.Open(scope.Root) // #nosec G304 -- cache root from config.
+			if err != nil {
+				continue
+			}
+			names, err := dir.Readdir(-1)
+			_ = dir.Close()
+			if err != nil {
+				continue
+			}
+			for _, info := range names {
+				visit(filepath.Join(scope.Root, info.Name()), info, scope)
+			}
+			continue
+		}
+		_ = filepath.Walk(scope.Root, func(path string, info os.FileInfo, err error) error {
+			if err != nil || info == nil {
+				return nil // best-effort: one unreadable dir must not abort cleanup
+			}
+			visit(path, info, scope)
+			return nil
+		})
+	}
+}
+
+// removeImageCacheEntry 删除一个缓存文件，并顺带清掉它的失败标记。
+func removeImageCacheEntry(entry imageCacheFileEntry, result *PruneImageCacheResult) {
+	if err := os.Remove(entry.path); err != nil {
+		return
+	}
+	result.DeletedFiles++
+	result.FreedBytes += entry.size
+	result.RemainingBytes -= entry.size
+	_ = os.Remove(entry.path + ".fail")
+}
+
+// PruneImageCachePools 按池清理图片缓存：
+//  1. 删除超过池 MaxAge 的文件（原图按保留时长淘汰）；
+//  2. 池自身超过 MaxBytes 时按 mtime 淘汰到 80%（留水位，避免连续写入即触发）；
+//  3. 全部池合计超过 totalBytes 时，仍按“先原图、后派生”的顺序淘汰最旧文件，
+//     使总量上限始终是硬保证，同时让客户端热路径的缩略图活得更久。
+//
+// 传入的池顺序即总量超限时的淘汰优先级（排在前面的先被淘汰）。
+func PruneImageCachePools(pools []ImageCachePool, totalBytes int64) (PruneImageCacheResult, error) {
+	var result PruneImageCacheResult
+	if len(pools) == 0 {
+		return result, nil
+	}
+
+	var entries []imageCacheFileEntry
+	var failMarkers []imageCacheFileEntry
+	for i, pool := range pools {
+		collectImageCachePoolFiles(pool, i, &entries, &failMarkers)
+	}
+
+	result.TotalFilesBefore = len(entries)
+	for _, entry := range entries {
+		result.TotalBytesBefore += entry.size
+	}
+	result.RemainingBytes = result.TotalBytesBefore
+
+	// 过期的失败标记直接清掉：它们不计容量，只用于观测与重试。
+	cutoffMarkers := time.Now().Add(-imageFailMarkerMaxAge)
+	for _, marker := range failMarkers {
+		if marker.modTime.Before(cutoffMarkers) {
+			_ = os.Remove(marker.path)
+		}
+	}
+
+	now := time.Now()
+	deleted := make(map[string]bool, len(entries))
+	deleteEntry := func(entry imageCacheFileEntry) {
+		if deleted[entry.path] {
+			return
+		}
+		deleted[entry.path] = true
+		removeImageCacheEntry(entry, &result)
+	}
+
+	// 1. 按池保留时长淘汰（原图池）。
+	for i := range pools {
+		if pools[i].MaxAge <= 0 {
+			continue
+		}
+		cutoff := now.Add(-pools[i].MaxAge)
+		for _, entry := range entries {
+			if entry.pool == i && entry.modTime.Before(cutoff) {
+				deleteEntry(entry)
+			}
+		}
+	}
+
+	// 2. 池自身配额。
+	for i := range pools {
+		maxBytes := pools[i].MaxBytes
+		if maxBytes <= 0 {
+			continue
+		}
+		var poolBytes int64
+		for _, entry := range entries {
+			if entry.pool == i && !deleted[entry.path] {
+				poolBytes += entry.size
+			}
+		}
+		if poolBytes <= maxBytes {
+			continue
+		}
+		target := maxBytes * 80 / 100
+		for _, entry := range oldestFirst(entries, i, deleted) {
+			if poolBytes <= target {
+				break
+			}
+			deleteEntry(entry)
+			poolBytes -= entry.size
+		}
+	}
+
+	// 3. 总量硬上限：跨池按淘汰优先级（池顺序）再按 mtime 淘汰。
+	if totalBytes > 0 && result.RemainingBytes > totalBytes {
+		target := totalBytes * 80 / 100
+		for _, entry := range evictionOrder(entries, deleted) {
+			if result.RemainingBytes <= target {
+				break
+			}
+			deleteEntry(entry)
+		}
+	}
+
+	removeEmptyImageCacheDirs(pools)
+	return result, nil
+}
+
+// oldestFirst 返回指定池内未被删除的文件，按修改时间从旧到新。
+func oldestFirst(entries []imageCacheFileEntry, pool int, deleted map[string]bool) []imageCacheFileEntry {
+	out := make([]imageCacheFileEntry, 0, len(entries))
+	for _, entry := range entries {
+		if entry.pool == pool && !deleted[entry.path] {
+			out = append(out, entry)
+		}
+	}
+	sort.Slice(out, func(i, j int) bool { return out[i].modTime.Before(out[j].modTime) })
+	return out
+}
+
+// evictionOrder 返回总量超限时的淘汰顺序：先按池优先级（原图池在前），
+// 池内再按修改时间从旧到新。
+func evictionOrder(entries []imageCacheFileEntry, deleted map[string]bool) []imageCacheFileEntry {
+	out := make([]imageCacheFileEntry, 0, len(entries))
+	for _, entry := range entries {
+		if !deleted[entry.path] {
+			out = append(out, entry)
+		}
+	}
+	sort.SliceStable(out, func(i, j int) bool {
+		if out[i].pool != out[j].pool {
+			return out[i].pool < out[j].pool
+		}
+		return out[i].modTime.Before(out[j].modTime)
+	})
+	return out
+}
+
+// removeEmptyImageCacheDirs 清理淘汰后留下的空子目录（不含池根目录本身）。
+func removeEmptyImageCacheDirs(pools []ImageCachePool) {
+	seen := map[string]bool{}
+	for _, pool := range pools {
+		for _, scope := range pool.Scopes {
+			if !scope.Recursive || seen[scope.Root] {
+				continue
+			}
+			seen[scope.Root] = true
+			dirs := []string{}
+			_ = filepath.Walk(scope.Root, func(path string, info os.FileInfo, err error) error {
+				if err != nil || info == nil {
+					return nil
+				}
+				if info.IsDir() && path != scope.Root {
+					dirs = append(dirs, path)
+				}
+				return nil
+			})
+			for i := len(dirs) - 1; i >= 0; i-- {
+				_ = os.Remove(dirs[i])
+			}
+		}
+	}
+}
+
+// PruneImageCache 是单池版本的兼容入口：把整个图片目录当作一个池，超限时按
+// mtime 淘汰到 80%。保留它是为了「立刻清理图片缓存」这类只关心总量的调用方。
 func PruneImageCache(imagesDir string, maxSizeBytes int64) (PruneImageCacheResult, error) {
 	var result PruneImageCacheResult
 	if imagesDir == "" || maxSizeBytes <= 0 {
@@ -75,66 +344,35 @@ func PruneImageCache(imagesDir string, maxSizeBytes int64) (PruneImageCacheResul
 	if _, err := os.Stat(imagesDir); err != nil {
 		return result, nil
 	}
-
-	var (
-		dirs    []string
-		entries []imageCacheFileEntry
-	)
-
-	_ = filepath.Walk(imagesDir, func(path string, info os.FileInfo, err error) error {
-		if err != nil {
-			return nil
-		}
-		if info.IsDir() {
-			if path != imagesDir {
-				dirs = append(dirs, path)
-			}
-			return nil
-		}
-		// Skip temporary files created during image download.
-		name := info.Name()
-		if strings.HasSuffix(name, ".tmp") || strings.HasPrefix(name, "img-") && strings.Contains(name, ".tmp") {
-			return nil
-		}
-		size := info.Size()
-		result.TotalFilesBefore++
-		result.TotalBytesBefore += size
-		entries = append(entries, imageCacheFileEntry{
-			path:    path,
-			size:    size,
-			modTime: info.ModTime(),
-		})
-		return nil
-	})
-
-	result.RemainingBytes = result.TotalBytesBefore
-	if result.TotalBytesBefore <= maxSizeBytes {
-		return result, nil
+	pool := ImageCachePool{
+		Name:     "images",
+		Scopes:   []ImageCacheScope{{Root: imagesDir, Recursive: true}},
+		MaxBytes: maxSizeBytes,
 	}
-
-	// High/Low watermark: prune down to 80% of max size to leave headroom
-	// and prevent disk thrashing on consecutive writes.
-	targetSizeBytes := maxSizeBytes * 80 / 100
-
-	sort.Slice(entries, func(i, j int) bool {
-		return entries[i].modTime.Before(entries[j].modTime)
-	})
-
-	for _, entry := range entries {
-		if result.RemainingBytes <= targetSizeBytes {
-			break
-		}
-		if err := os.Remove(entry.path); err == nil || errors.Is(err, os.ErrNotExist) {
-			result.DeletedFiles++
-			result.FreedBytes += entry.size
-			result.RemainingBytes -= entry.size
-		}
-	}
-
-	// Clean up emptied subdirectories from deepest to shallowest.
-	for i := len(dirs) - 1; i >= 0; i-- {
-		_ = os.Remove(dirs[i])
-	}
-
-	return result, nil
+	return PruneImageCachePools([]ImageCachePool{pool}, maxSizeBytes)
+}
+
+// ImageCachePools 按配置组装图片缓存的两个清理池。
+//
+// 第一个池是原图：历史版本平铺在 images/ 下的旧缓存也归入此池，升级后会被
+// 逐步淘汰；第二个池是派生成品（本地缩放结果 + 挂载 Emby 按尺寸返回的成品）。
+// 池顺序即总量超限时的淘汰优先级。
+func ImageCachePools(imagesDir string, originalsMaxBytes int64, originalsMaxAge time.Duration) []ImageCachePool {
+	originals := ImageCachePool{
+		Name:     "originals",
+		MaxBytes: originalsMaxBytes,
+		MaxAge:   originalsMaxAge,
+		Scopes: []ImageCacheScope{
+			{Root: imagesDir, Recursive: false}, // 旧版平铺布局
+			{Root: filepath.Join(imagesDir, imageOriginalCacheSubdir), Recursive: true},
+		},
+	}
+	derived := ImageCachePool{
+		Name: "derived",
+		Scopes: []ImageCacheScope{
+			{Root: filepath.Join(imagesDir, imageResizeCacheSubdir), Recursive: true, DropUnknownResizeVariants: true},
+			{Root: filepath.Join(imagesDir, imageRenditionCacheSubdir), Recursive: true},
+		},
+	}
+	return []ImageCachePool{originals, derived}
 }

internal/service/cache_cleanup_pools_test.go

@@ -0,0 +1,223 @@
+package service
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+)
+
+// writePoolFile ?????????????????????????
+func writePoolFile(t *testing.T, path string, size int, modTime time.Time) {
+	t.Helper()
+	if err := os.MkdirAll(filepath.Dir(path), 0o750); err != nil {
+		t.Fatalf("mkdir %s: %v", filepath.Dir(path), err)
+	}
+	if err := os.WriteFile(path, make([]byte, size), 0o600); err != nil {
+		t.Fatalf("write %s: %v", path, err)
+	}
+	if !modTime.IsZero() {
+		if err := os.Chtimes(path, modTime, modTime); err != nil {
+			t.Fatalf("chtimes %s: %v", path, err)
+		}
+	}
+}
+
+func mustNotExist(t *testing.T, path string) {
+	t.Helper()
+	if _, err := os.Stat(path); !os.IsNotExist(err) {
+		t.Fatalf("expected %s to be deleted, got err=%v", path, err)
+	}
+}
+
+func mustExist(t *testing.T, path string) {
+	t.Helper()
+	if _, err := os.Stat(path); err != nil {
+		t.Fatalf("expected %s to survive, got err=%v", path, err)
+	}
+}
+
+// currentVariantPath ????????????????.<??x<??q<??>.img??
+// ??????????????????????????
+func currentVariantPath(dir, keyChar string, width, height, quality int) string {
+	key := strings.Repeat(keyChar, sha256HexLength)
+	return filepath.Join(dir, fmt.Sprintf("%s.%dx%dq%d.img", key, width, height, quality))
+}
+
+// TestPruneImageCachePoolsOriginalTTL ????????????????????
+// ?? TTL ??????????????????????
+func TestPruneImageCachePoolsOriginalTTL(t *testing.T) {
+	imagesDir := t.TempDir()
+	now := time.Now()
+
+	stale := filepath.Join(imagesDir, imageOriginalCacheSubdir, "stale-original")
+	fresh := filepath.Join(imagesDir, imageOriginalCacheSubdir, "fresh-original")
+	thumb := currentVariantPath(filepath.Join(imagesDir, imageResizeCacheSubdir), "b", 480, 600, 80)
+	writePoolFile(t, stale, 400, now.Add(-8*24*time.Hour))
+	writePoolFile(t, fresh, 400, now.Add(-1*time.Hour))
+	// ??????????TTL ????
+	writePoolFile(t, thumb, 400, now.Add(-8*24*time.Hour))
+
+	pools := ImageCachePools(imagesDir, 0, 7*24*time.Hour)
+	res, err := PruneImageCachePools(pools, 100*1024*1024)
+	if err != nil {
+		t.Fatalf("prune failed: %v", err)
+	}
+	if res.DeletedFiles != 1 {
+		t.Fatalf("deleted files = %d, want 1", res.DeletedFiles)
+	}
+	mustNotExist(t, stale)
+	mustExist(t, fresh)
+	mustExist(t, thumb)
+}
+
+// TestPruneImageCachePoolsEvictsOriginalsBeforeDerived ????????????
+// ????????????????????
+func TestPruneImageCachePoolsEvictsOriginalsBeforeDerived(t *testing.T) {
+	imagesDir := t.TempDir()
+	now := time.Now()
+
+	original := filepath.Join(imagesDir, imageOriginalCacheSubdir, "original")
+	rendition := filepath.Join(imagesDir, imageRenditionCacheSubdir, "rendition")
+	thumb := currentVariantPath(filepath.Join(imagesDir, imageResizeCacheSubdir), "c", 480, 600, 80)
+	// ?????????????????
+	writePoolFile(t, original, 1024*1024, now.Add(-72*time.Hour))
+	writePoolFile(t, rendition, 1024*1024, now.Add(-2*time.Hour))
+	writePoolFile(t, thumb, 1024*1024, now.Add(-1*time.Hour))
+
+	pools := ImageCachePools(imagesDir, 0, 0)
+	// ?? 3MB????2.4MB -> ????? 1.92MB??
+	res, err := PruneImageCachePools(pools, 2*1024*1024+400*1024)
+	if err != nil {
+		t.Fatalf("prune failed: %v", err)
+	}
+	if res.DeletedFiles != 2 {
+		t.Fatalf("deleted files = %d, want 2 (original + oldest derived)", res.DeletedFiles)
+	}
+	mustNotExist(t, original)
+	mustNotExist(t, rendition)
+	mustExist(t, thumb)
+	if res.RemainingBytes != 1024*1024 {
+		t.Fatalf("remaining bytes = %d, want %d", res.RemainingBytes, 1024*1024)
+	}
+}
+
+// TestPruneImageCachePoolsOriginalQuotaKeepsDerived ????????????
+// ??????????????
+func TestPruneImageCachePoolsOriginalQuotaKeepsDerived(t *testing.T) {
+	imagesDir := t.TempDir()
+	now := time.Now()
+
+	oldOriginal := filepath.Join(imagesDir, imageOriginalCacheSubdir, "old")
+	newOriginal := filepath.Join(imagesDir, imageOriginalCacheSubdir, "new")
+	thumb := currentVariantPath(filepath.Join(imagesDir, imageResizeCacheSubdir), "c", 480, 600, 80)
+	writePoolFile(t, oldOriginal, 800, now.Add(-3*time.Hour))
+	writePoolFile(t, newOriginal, 800, now.Add(-1*time.Hour))
+	writePoolFile(t, thumb, 800, now.Add(-2*time.Hour))
+
+	// ???? 1000 ??????????1600 -> ????800????? 1 ????
+	pools := ImageCachePools(imagesDir, 1000, 0)
+	res, err := PruneImageCachePools(pools, 0)
+	if err != nil {
+		t.Fatalf("prune failed: %v", err)
+	}
+	if res.DeletedFiles != 1 {
+		t.Fatalf("deleted files = %d, want 1", res.DeletedFiles)
+	}
+	mustNotExist(t, oldOriginal)
+	mustExist(t, newOriginal)
+	mustExist(t, thumb)
+}
+
+// TestPruneImageCachePoolsCleansFailMarkers ????????????????
+// ????????????????
+func TestPruneImageCachePoolsCleansFailMarkers(t *testing.T) {
+	imagesDir := t.TempDir()
+	now := time.Now()
+
+	entry := filepath.Join(imagesDir, imageOriginalCacheSubdir, "entry")
+	entryMarker := entry + ".fail"
+	orphanMarker := filepath.Join(imagesDir, imageOriginalCacheSubdir, "orphan.fail")
+	freshMarker := filepath.Join(imagesDir, imageOriginalCacheSubdir, "fresh.fail")
+
+	writePoolFile(t, entry, 900, now.Add(-10*time.Hour))
+	writePoolFile(t, entryMarker, 35, now.Add(-10*time.Hour))
+	writePoolFile(t, orphanMarker, 35, now.Add(-48*time.Hour))
+	writePoolFile(t, freshMarker, 35, now.Add(-1*time.Minute))
+
+	// ???? 500????900 ?????????????????
+	pools := ImageCachePools(imagesDir, 500, 0)
+	res, err := PruneImageCachePools(pools, 0)
+	if err != nil {
+		t.Fatalf("prune failed: %v", err)
+	}
+	if res.DeletedFiles != 1 {
+		t.Fatalf("deleted files = %d, want 1", res.DeletedFiles)
+	}
+	mustNotExist(t, entry)
+	mustNotExist(t, entryMarker)
+	mustNotExist(t, orphanMarker)
+	mustExist(t, freshMarker)
+}
+
+// TestPruneImageCachePoolsSkipsTemporaryFiles ????????????????
+// ?????????????????
+func TestPruneImageCachePoolsSkipsTemporaryFiles(t *testing.T) {
+	imagesDir := t.TempDir()
+	tmp := filepath.Join(imagesDir, imageOriginalCacheSubdir, "img-1234.tmp")
+	writePoolFile(t, tmp, 4096, time.Now().Add(-48*time.Hour))
+
+	pools := ImageCachePools(imagesDir, 1, 0)
+	if _, err := PruneImageCachePools(pools, 1); err != nil {
+		t.Fatalf("prune failed: %v", err)
+	}
+	mustExist(t, tmp)
+}
+
+// TestPruneImageCachePoolsDropsLegacyResizeVariants ????????.img????
+// ?????????????????????????????????????
+func TestPruneImageCachePoolsDropsLegacyResizeVariants(t *testing.T) {
+	imagesDir := t.TempDir()
+	resized := filepath.Join(imagesDir, imageResizeCacheSubdir)
+	legacy := filepath.Join(resized, "4f2a1b"+strings.Repeat("0", 58)+".img")
+	current := filepath.Join(resized, strings.Repeat("a", 64)+".480x600q80.img")
+	other := filepath.Join(resized, "not-a-hash.480x600q80.img")
+	writePoolFile(t, legacy, 400, time.Now())
+	writePoolFile(t, current, 400, time.Now())
+	writePoolFile(t, other, 400, time.Now())
+
+	pools := ImageCachePools(imagesDir, 0, 0)
+	res, err := PruneImageCachePools(pools, 0)
+	if err != nil {
+		t.Fatalf("prune failed: %v", err)
+	}
+	mustNotExist(t, legacy)
+	mustNotExist(t, other)
+	mustExist(t, current)
+	if res.TotalFilesBefore != 1 {
+		t.Fatalf("counted files = %d, want only the current-format variant", res.TotalFilesBefore)
+	}
+}
+
+// TestPruneImageCachePoolsCountsLegacyFlatLayout ??????images/ ????
+// ?????????????????
+func TestPruneImageCachePoolsCountsLegacyFlatLayout(t *testing.T) {
+	imagesDir := t.TempDir()
+	legacy := filepath.Join(imagesDir, "legacy-flat-cache")
+	thumb := currentVariantPath(filepath.Join(imagesDir, imageResizeCacheSubdir), "c", 480, 600, 80)
+	writePoolFile(t, legacy, 900, time.Now().Add(-30*24*time.Hour))
+	writePoolFile(t, thumb, 900, time.Now().Add(-30*24*time.Hour))
+
+	pools := ImageCachePools(imagesDir, 0, 24*time.Hour)
+	res, err := PruneImageCachePools(pools, 0)
+	if err != nil {
+		t.Fatalf("prune failed: %v", err)
+	}
+	if res.DeletedFiles != 1 {
+		t.Fatalf("deleted files = %d, want 1", res.DeletedFiles)
+	}
+	mustNotExist(t, legacy)
+	mustExist(t, thumb)
+}

internal/service/cache_cleanup_test.go

@@ -148,8 +148,8 @@ func TestSchedulerJobCleanImageCache(t *testing.T) {
 
 	scheduler := NewSchedulerService(zap.NewNop(), nil, nil, nil, nil, nil, cacheRoot)
 	// Set limit to 1MB; our file is 2MB -> should be pruned
-	scheduler.SetImagesMaxSizeMBProvider(func() int {
-		return 1
+	scheduler.SetImageCachePolicyProvider(func() ImageCachePolicy {
+		return ImageCachePolicy{TotalBytes: 1 * 1024 * 1024}
 	})
 
 	if err := scheduler.jobCleanImageCache(context.Background()); err != nil {

internal/service/cloud115_playback.go

@@ -58,12 +58,32 @@ type Cloud115PlaybackService struct {
 	repo *repository.Container
 	strm *StrmService
 
+	// transcoder 用于判断本地 HLS 档位此刻是否真的可用；未注入时按"不可用"处理，
+	// 避免向播放器推荐必然 500 的转码档位。
+	transcoder *TranscoderService
+
 	mu        sync.Mutex
 	pushState map[string]cloud115PushAttempt
 
 	proxy *Cloud115HLSProxy
 }
 
+// SetTranscoder 注入转码服务，用于标注本地 HLS 档位的可用性。
+func (s *Cloud115PlaybackService) SetTranscoder(transcoder *TranscoderService) *Cloud115PlaybackService {
+	if s != nil {
+		s.transcoder = transcoder
+	}
+	return s
+}
+
+// localTranscodeAvailable 报告本地 HLS 此刻是否可用。
+func (s *Cloud115PlaybackService) localTranscodeAvailable() bool {
+	if s == nil || s.transcoder == nil {
+		return false
+	}
+	return s.transcoder.Available()
+}
+
 func NewCloud115PlaybackService(cfg *config.Config, log *zap.Logger, repo *repository.Container, strm *StrmService) *Cloud115PlaybackService {
 	svc := &Cloud115PlaybackService{
 		cfg:       cfg,
@@ -101,7 +121,7 @@ func (s *Cloud115PlaybackService) PlaybackInfo(ctx context.Context, mediaID stri
 	}
 
 	provider := MediaPlaybackProvider(m)
-	localQualities := LocalQualityOptions(m)
+	localQualities := LocalQualityOptions(m, s.localTranscodeAvailable())
 	localDefault := DefaultLocalHLSQualityID(m)
 	info := &PlaybackInfo{
 		MediaID:        m.ID,

internal/service/emby_artwork.go

@@ -147,10 +147,18 @@ func (e *EmbyService) ImageInfos(ctx context.Context, id string) []map[string]an
 			continue
 		}
 		seen[raw] = true
+		// ImageTag 让客户端判断自己缓存的图片是否还有效：远程条目的真实
+		// ImageTags 会随远端换图变化，拿不到时才退化为条目 ID（恒定值）。
+		tag := id
+		if e.remote != nil {
+			if remoteTag := e.remote.RemoteImageTagOfEncodedID(ctx, id, imageType); remoteTag != "" {
+				tag = remoteTag
+			}
+		}
 		out = append(out, map[string]any{
 			"ImageType":  imageType,
 			"ImageIndex": 0,
-			"ImageTag":   id,
+			"ImageTag":   tag,
 		})
 	}
 	return out

internal/service/emby_compat.go

@@ -141,6 +141,10 @@ type ItemsParams struct {
 	SortOrder        string
 	Limit            int
 	StartIndex       int
+	// SeasonIndex 对应客户端的 Season / SeasonIndex 查询参数（季序号，0 为特别篇）。
+	// nil 表示不按季过滤；非 nil 时只返回该季的剧集。客户端普遍用季序号而不是
+	// 虚拟季 ID 请求剧集，缺了它 /Shows/{id}/Episodes?Season=N 会返回整部剧。
+	SeasonIndex *int
 }
 
 const (

internal/service/emby_items_cache.go

@@ -47,10 +47,20 @@ func (e *EmbyService) embyItemsCacheKey(kind string, p ItemsParams) string {
 		p.SortOrder,
 		strconv.Itoa(p.StartIndex),
 		strconv.Itoa(p.Limit),
+		formatSeasonIndexCacheKeyPart(p.SeasonIndex),
 	}, "|")))
 	return "media:emby:" + hex.EncodeToString(sum[:])
 }
 
+// formatSeasonIndexCacheKeyPart keeps "no season filter" distinct from "season 0"
+// so a client's Season=0 (specials) request never reuses an unfiltered response.
+func formatSeasonIndexCacheKeyPart(seasonIndex *int) string {
+	if seasonIndex == nil {
+		return "season:*"
+	}
+	return "season:" + strconv.Itoa(*seasonIndex)
+}
+
 func (e *EmbyService) embyLatestCacheKey(userID, parentID string, limit int) string {
 	// v2: payload tags for virtual artwork changed so clients drop cached placeholders.
 	sum := sha256.Sum256([]byte(strings.Join([]string{"latest-v2", userID, parentID, strconv.Itoa(limit)}, "|")))

internal/service/emby_items_list.go

@@ -121,16 +121,31 @@ func (e *EmbyService) mediaItems(ctx context.Context, p ItemsParams) (map[string
 
 func (e *EmbyService) episodeItems(ctx context.Context, rows []model.Media, p ItemsParams) (map[string]any, error) {
 	rows = e.filterMediaRowsForUser(ctx, rows, p.UserID)
-	if p.SearchTerm != "" {
-		filtered := rows[:0]
-		needle := strings.ToLower(p.SearchTerm)
-		for _, row := range rows {
-			if strings.Contains(strings.ToLower(row.Title), needle) || strings.Contains(strings.ToLower(row.OriginalName), needle) {
-				filtered = append(filtered, row)
+	// rows 可能来自 series 分组的内存 memo（embySeriesGroup.Episodes）。过滤必须
+	// 分配新切片：就地复用 rows[:0] 会覆写 memo 里的元素，让后续请求看到被前一次
+	// 过滤污染的剧集列表（例如按季筛选一次之后，特别篇就从缓存分组里消失了）。
+	if p.SeasonIndex != nil {
+		filtered := make([]model.Media, 0, len(rows))
+		for i := range rows {
+			if embyRowMatchesSeasonIndex(&rows[i], p.SeasonIndex) {
+				filtered = append(filtered, rows[i])
 			}
 		}
 		rows = filtered
 	}
+	if p.SearchTerm != "" {
+		needle := strings.ToLower(p.SearchTerm)
+		filtered := make([]model.Media, 0, len(rows))
+		for i := range rows {
+			if strings.Contains(strings.ToLower(rows[i].Title), needle) || strings.Contains(strings.ToLower(rows[i].OriginalName), needle) {
+				filtered = append(filtered, rows[i])
+			}
+		}
+		rows = filtered
+	}
+	// sort.SliceStable 同样会就地重排：先拷贝一份，避免把 memo 分组里的剧集顺序
+	// 按每次请求的分页/筛选结果固定下来。
+	rows = append([]model.Media(nil), rows...)
 	sort.SliceStable(rows, func(i, j int) bool {
 		if rows[i].SeasonNum != rows[j].SeasonNum {
 			return rows[i].SeasonNum < rows[j].SeasonNum

internal/service/emby_remote.go

@@ -79,6 +79,13 @@ type EmbyRemoteService struct {
 
 	personMu     sync.RWMutex
 	personImages map[string]embyRemotePersonImageRef
+
+	// imageTagMu 保护 imageTags：远程条目图片标签缓存，键为
+	// imageTagKey(accountID, remoteID, imageType)。它让图片 URL 带上远端
+	// ImageTags，从而在远端换图后让本地磁盘缓存与客户端缓存一起失效
+	// （没有它时 URL 恒定，缩略图会永久停留在旧版本）。
+	imageTagMu sync.RWMutex
+	imageTags  map[string]string
 }
 
 type embyRemotePersonImageRef struct {
@@ -86,6 +93,9 @@ type embyRemotePersonImageRef struct {
 	remoteID  string
 }
 
+// embyRemoteMaxImageTags 限制图片标签映射的条目数，避免长期运行后无界增长。
+const embyRemoteMaxImageTags = 20000
+
 // NewEmbyRemoteService 构造远程 Emby 聚合服务。
 func NewEmbyRemoteService(cfg *config.Config, log *zap.Logger, repo *repository.Container, crypto *CryptoService) *EmbyRemoteService {
 	return &EmbyRemoteService{
@@ -819,6 +829,7 @@ func (r *EmbyRemoteService) RemoteItem(ctx context.Context, mount *model.EmbyMou
 		var cached map[string]any
 		if r.cache.GetJSON(ctx, cacheKey, &cached) && len(cached) > 0 {
 			r.rememberRemotePeople(mount, cached)
+			r.rememberRemoteImageTags(embyRemoteAccountID(acct), cached)
 			return cached, nil
 		}
 	}
@@ -833,6 +844,7 @@ func (r *EmbyRemoteService) RemoteItem(ctx context.Context, mount *model.EmbyMou
 		return nil, err
 	}
 	r.rememberRemotePeople(mount, out)
+	r.rememberRemoteImageTags(embyRemoteAccountID(acct), out)
 	RewriteEmbyRemoteIDs(out, mount.ID)
 	if cacheKey != "" && len(out) > 0 {
 		r.cache.SetJSON(ctx, cacheKey, out, r.remoteMediaCacheTTL())
@@ -1153,6 +1165,150 @@ func rewriteSubtitleDeliveryURLs(src map[string]any, playURL string, cfg *EmbyRe
 	}
 }
 
+// embyRemoteImageTagType 归一化图片类型：Emby 的 Art 与 Backdrop 指同一张图，
+// 载荷里的 ImageTags 只会有 Primary / Backdrop 两个键。
+func embyRemoteImageTagType(imageType string) string {
+	switch strings.ToLower(strings.TrimSpace(imageType)) {
+	case "primary", "poster":
+		return "Primary"
+	case "backdrop", "art", "background":
+		return "Backdrop"
+	default:
+		return ""
+	}
+}
+
+// remoteItemImageTag 从远程条目载荷读取某一类图片的原始 tag。载荷可能已被
+// RewriteEmbyRemoteIDs 伪装过（tag 变成 embyremote~scope~tag），此处会还原。
+func remoteItemImageTag(item map[string]any, imageType string) string {
+	typ := embyRemoteImageTagType(imageType)
+	if item == nil || typ == "" {
+		return ""
+	}
+	var raw string
+	switch tags := item["ImageTags"].(type) {
+	case map[string]any:
+		raw = anyString(tags[typ])
+	case map[string]string:
+		raw = tags[typ]
+	}
+	if raw == "" && typ == "Backdrop" {
+		switch tags := item["BackdropImageTags"].(type) {
+		case []any:
+			if len(tags) > 0 {
+				raw = anyString(tags[0])
+			}
+		case []string:
+			if len(tags) > 0 {
+				raw = tags[0]
+			}
+		}
+	}
+	if _, original, ok := DecodeEmbyRemoteID(raw); ok {
+		return original
+	}
+	return strings.TrimSpace(raw)
+}
+
+// imageTagKey 是图片标签映射的键；不认识的图片类型返回空串（不记录）。
+func imageTagKey(accountID, remoteID, imageType string) string {
+	typ := embyRemoteImageTagType(imageType)
+	if typ == "" || strings.TrimSpace(remoteID) == "" || strings.TrimSpace(accountID) == "" {
+		return ""
+	}
+	return accountID + "|" + remoteID + "|" + typ
+}
+
+// rememberRemoteImageTagValue 记录单条图片标签（供 SeriesPrimaryImageTag 这类
+// 散落在载荷其他字段里的标签使用）。
+func (r *EmbyRemoteService) rememberRemoteImageTagValue(accountID, remoteID, imageType, tag string) {
+	if r == nil || strings.TrimSpace(tag) == "" {
+		return
+	}
+	if _, original, ok := DecodeEmbyRemoteID(tag); ok {
+		tag = original
+	}
+	key := imageTagKey(accountID, remoteID, imageType)
+	if key == "" {
+		return
+	}
+	r.imageTagMu.Lock()
+	defer r.imageTagMu.Unlock()
+	if r.imageTags == nil || len(r.imageTags) > embyRemoteMaxImageTags {
+		r.imageTags = make(map[string]string, 256)
+	}
+	r.imageTags[key] = tag
+}
+
+// rememberRemoteImageTags 记录载荷里出现的图片标签。载荷可以已被伪装。
+func (r *EmbyRemoteService) rememberRemoteImageTags(accountID string, item map[string]any) {
+	if r == nil || item == nil || strings.TrimSpace(accountID) == "" {
+		return
+	}
+	remoteID := remoteItemString(item, "Id")
+	if _, original, ok := DecodeEmbyRemoteID(remoteID); ok {
+		remoteID = original
+	}
+	if strings.TrimSpace(remoteID) == "" {
+		return
+	}
+	for _, imageType := range []string{"Primary", "Backdrop"} {
+		if tag := remoteItemImageTag(item, imageType); tag != "" {
+			r.rememberRemoteImageTagValue(accountID, remoteID, imageType, tag)
+		}
+	}
+}
+
+// remoteImageTag 查询已记录的图片标签；未知时返回空串（调用方退化为原行为）。
+func (r *EmbyRemoteService) remoteImageTag(accountID, remoteID, imageType string) string {
+	if r == nil {
+		return ""
+	}
+	key := imageTagKey(accountID, remoteID, imageType)
+	if key == "" {
+		return ""
+	}
+	r.imageTagMu.RLock()
+	defer r.imageTagMu.RUnlock()
+	return r.imageTags[key]
+}
+
+// remoteImageTagQuery 返回追加到远程图片地址后的 tag 查询片段（含 & 前缀）。
+// Emby 用 tag 作为图片 ETag/cache key：带上它之后，远端换图会改变 MeBox 的
+// 磁盘缓存键，缩略图与客户端缓存都会随之失效，而不是永久停留在旧版本。
+func (r *EmbyRemoteService) remoteImageTagQuery(accountID, remoteID, imageType string) string {
+	tag := r.remoteImageTag(accountID, remoteID, imageType)
+	if tag == "" {
+		return ""
+	}
+	return "&tag=" + url.QueryEscape(tag)
+}
+
+// RemoteImageTagOfEncodedID 按伪装 ID 解析已记录的图片标签，供兼容层
+// 回报 ImageTag（客户端据此决定是否复用自己缓存的图片）。
+func (r *EmbyRemoteService) RemoteImageTagOfEncodedID(ctx context.Context, encodedID, imageType string) string {
+	if r == nil {
+		return ""
+	}
+	mountID, remoteID, ok := DecodeEmbyRemoteID(encodedID)
+	if !ok {
+		return ""
+	}
+	_, acct, _ := r.ResolveMount(ctx, mountID)
+	if acct == nil {
+		return ""
+	}
+	return r.remoteImageTag(acct.ID, remoteID, imageType)
+}
+
+// embyRemoteAccountID 空值安全的账号 ID 读取（构建图片 URL 时可能只有账号对象）。
+func embyRemoteAccountID(acct *model.StrmAccount) string {
+	if acct == nil {
+		return ""
+	}
+	return acct.ID
+}
+
 // RemoteImageURL 构造远程图片绝对地址（由既有 ImageProxy 拉取透传）。
 func (r *EmbyRemoteService) RemoteImageURL(ctx context.Context, acct *model.StrmAccount, remoteID, imageType string) (string, error) {
 	cfg, err := r.configOf(acct)
@@ -1160,7 +1316,7 @@ func (r *EmbyRemoteService) RemoteImageURL(ctx context.Context, acct *model.Strm
 		return "", err
 	}
 	return r.embyBase(cfg) + "/Items/" + url.PathEscape(remoteID) + "/Images/" + url.PathEscape(strings.ToLower(imageType)) +
-		"?api_key=" + url.QueryEscape(cfg.Token), nil
+		"?api_key=" + url.QueryEscape(cfg.Token) + r.remoteImageTagQuery(embyRemoteAccountID(acct), remoteID, imageType), nil
 }
 
 // rememberRemotePeople 记录远程人物名称到远程人物 ID 的映射，供旧式

internal/service/emby_remote_image_tag_test.go

@@ -0,0 +1,173 @@
+package service
+
+import (
+	"strings"
+	"testing"
+
+	"go.uber.org/zap"
+
+	"github.com/truewhile/MeBox/internal/config"
+	"github.com/truewhile/MeBox/internal/model"
+	"github.com/truewhile/MeBox/internal/repository"
+)
+
+// newImageTagTestService 构造只用于图片标签映射测试的远程 Emby 服务。
+func newImageTagTestService(t *testing.T) (*EmbyRemoteService, *model.EmbyMount, *model.StrmAccount) {
+	t.Helper()
+	db := newServiceTestDB(t, &model.StrmAccount{}, &model.EmbyMount{})
+	repos := repository.New(db)
+	svc := NewEmbyRemoteService(&config.Config{}, zap.NewNop(), repos, NewCryptoService("", zap.NewNop()))
+	acct := &model.StrmAccount{
+		Base:     model.Base{ID: "acct-tag"},
+		Name:     "tag-emby",
+		Provider: model.StrmProviderEmbyRemote,
+		Enabled:  true,
+		Config:   `{"url":"http://emby.test:8096","token":"fake-token"}`,
+	}
+	mount := &model.EmbyMount{
+		Base:         model.Base{ID: "mount-tag"},
+		AccountID:    acct.ID,
+		RemoteViewID: "view-tag",
+		Enabled:      true,
+	}
+	if err := repos.StrmAccount.Create(t.Context(), acct); err != nil {
+		t.Fatalf("create account: %v", err)
+	}
+	if err := repos.EmbyMount.Create(t.Context(), mount); err != nil {
+		t.Fatalf("create mount: %v", err)
+	}
+	return svc, mount, acct
+}
+
+// testRemoteImageConfig 是图片 URL 构造所需的最小远程配置。
+func testRemoteImageConfig() *EmbyRemoteConfig {
+	return &EmbyRemoteConfig{BaseURL: "http://emby.test:8096", Token: "fake-token"}
+}
+
+// TestMapRemoteItemToMediaCarriesImageTags 下发的图片 URL 必须带上远端图片
+// 标签，否则远端换图后磁盘/客户端缓存永远不会失效。
+func TestMapRemoteItemToMediaCarriesImageTags(t *testing.T) {
+	svc, mount, acct := newImageTagTestService(t)
+
+	media := svc.MapRemoteItemToMedia(t.Context(), mount, acct, testRemoteImageConfig(), map[string]any{
+		"Id":   "movie-1",
+		"Name": "测试影片",
+		"Type": "Movie",
+		"ImageTags": map[string]any{
+			"Primary": "primary-tag-1",
+		},
+		"BackdropImageTags": []any{"backdrop-tag-1"},
+	})
+
+	if !strings.Contains(media.PosterURL, "tag=primary-tag-1") {
+		t.Fatalf("PosterURL missing remote image tag: %q", media.PosterURL)
+	}
+	if !strings.Contains(media.BackdropURL, "tag=backdrop-tag-1") {
+		t.Fatalf("BackdropURL missing remote image tag: %q", media.BackdropURL)
+	}
+}
+
+// TestMapRemoteItemToMediaCarriesImageTagsAfterRewrite 载荷先经过 ID 伪装
+// （tag 变成 embyremote~...~tag）时仍要还原出原始 tag。
+func TestMapRemoteItemToMediaCarriesImageTagsAfterRewrite(t *testing.T) {
+	svc, mount, acct := newImageTagTestService(t)
+
+	item := map[string]any{
+		"Id":   "movie-2",
+		"Name": "伪装过的影片",
+		"Type": "Movie",
+		"ImageTags": map[string]any{
+			"Primary": "primary-tag-2",
+		},
+	}
+	RewriteEmbyRemoteIDs(item, mount.ID)
+
+	media := svc.MapRemoteItemToMedia(t.Context(), mount, acct, testRemoteImageConfig(), item)
+	if !strings.Contains(media.PosterURL, "tag=primary-tag-2") {
+		t.Fatalf("PosterURL missing decoded remote image tag: %q", media.PosterURL)
+	}
+	if strings.Contains(media.PosterURL, EmbyRemoteIDPrefix) {
+		t.Fatalf("PosterURL leaked disguised tag: %q", media.PosterURL)
+	}
+}
+
+// TestRemoteImageURLCacheKeyFollowsImageTag 远端换图（tag 变化）后，图片 URL
+// 必须随之变化，缓存键才会失效。
+func TestRemoteImageURLCacheKeyFollowsImageTag(t *testing.T) {
+	svc, mount, acct := newImageTagTestService(t)
+
+	first := svc.MapRemoteItemToMedia(t.Context(), mount, acct, testRemoteImageConfig(), map[string]any{
+		"Id":        "movie-3",
+		"Name":      "换图影片",
+		"Type":      "Movie",
+		"ImageTags": map[string]any{"Primary": "old-tag"},
+	})
+	second := svc.MapRemoteItemToMedia(t.Context(), mount, acct, testRemoteImageConfig(), map[string]any{
+		"Id":        "movie-3",
+		"Name":      "换图影片",
+		"Type":      "Movie",
+		"ImageTags": map[string]any{"Primary": "new-tag"},
+	})
+
+	if first.PosterURL == second.PosterURL {
+		t.Fatalf("image URL did not change when the remote tag changed: %q", first.PosterURL)
+	}
+	if imageCacheKeyURL(first.PosterURL) == imageCacheKeyURL(second.PosterURL) {
+		t.Fatal("image cache key did not change when the remote tag changed")
+	}
+}
+
+// TestMapRemoteItemToMediaWithoutImageTagsKeepsURLTagFree 没有标签时保持原样
+// （不追加空 tag 参数）。
+func TestMapRemoteItemToMediaWithoutImageTagsKeepsURLTagFree(t *testing.T) {
+	svc, mount, acct := newImageTagTestService(t)
+
+	media := svc.MapRemoteItemToMedia(t.Context(), mount, acct, testRemoteImageConfig(), map[string]any{
+		"Id":        "movie-4",
+		"Name":      "无标签影片",
+		"Type":      "Movie",
+		"ImageTags": map[string]any{"Primary": "primary-tag-4"},
+	})
+	if !strings.Contains(media.PosterURL, "tag=primary-tag-4") {
+		t.Fatalf("PosterURL missing tag: %q", media.PosterURL)
+	}
+	if strings.Contains(media.BackdropURL, "tag=") {
+		t.Fatalf("BackdropURL should not carry a tag when absent: %q", media.BackdropURL)
+	}
+
+	unknown := svc.MapRemoteItemToMedia(t.Context(), mount, acct, testRemoteImageConfig(), map[string]any{
+		"Id":   "movie-5",
+		"Name": "未知标签影片",
+		"Type": "Movie",
+		"ImageTags": map[string]any{
+			"Primary": "primary-tag-5",
+		},
+	})
+	if strings.Contains(unknown.BackdropURL, "tag=") {
+		t.Fatalf("unknown image tag should degrade to no tag: %q", unknown.BackdropURL)
+	}
+}
+
+// TestRemoteImageURLIncludesRememberedTag 兼容层（/emby/Items/{id}/Images/...）
+// 复用同一份标签映射。
+func TestRemoteImageURLIncludesRememberedTag(t *testing.T) {
+	svc, mount, acct := newImageTagTestService(t)
+
+	svc.rememberRemoteImageTags(acct.ID, map[string]any{
+		"Id":        "movie-6",
+		"ImageTags": map[string]any{"Primary": "compat-tag"},
+	})
+
+	raw, err := svc.RemoteImageURL(t.Context(), acct, "movie-6", "Primary")
+	if err != nil {
+		t.Fatalf("RemoteImageURL failed: %v", err)
+	}
+	if !strings.Contains(raw, "tag=compat-tag") {
+		t.Fatalf("remote image URL missing tag: %q", raw)
+	}
+
+	encoded := EncodeEmbyRemoteID(mount.ID, "movie-6")
+	if got := svc.RemoteImageTagOfEncodedID(t.Context(), encoded, "Primary"); got != "compat-tag" {
+		t.Fatalf("RemoteImageTagOfEncodedID = %q, want compat-tag", got)
+	}
+}

internal/service/emby_remote_web.go

@@ -158,7 +158,7 @@ func (r *EmbyRemoteService) mapRemoteMountToLibrary(mount *model.EmbyMount, acct
 	}
 	// 远程媒体库封面只有真实存在图片标签才下发。
 	if remoteItemHasImageTag(item, "Primary") {
-		lib.CoverURL = r.remoteItemImageURL(cfg, mount.RemoteViewID, "Primary")
+		lib.CoverURL = r.remoteItemImageURL(acct, cfg, mount.RemoteViewID, "Primary")
 	}
 	return lib
 }
@@ -176,6 +176,8 @@ func (r *EmbyRemoteService) MapRemoteItemToMedia(ctx context.Context, mount *mod
 	if _, rid, ok := DecodeEmbyRemoteID(remoteID); ok {
 		remoteID = rid
 	}
+	// 记录图片标签，使下发的图片 URL 带上 tag：远端换图后缓存随之失效。
+	r.rememberRemoteImageTags(embyRemoteAccountID(acct), item)
 	seriesID := remoteItemString(item, "SeriesId")
 	if _, rid, ok := DecodeEmbyRemoteID(seriesID); ok {
 		seriesID = rid
@@ -214,10 +216,10 @@ func (r *EmbyRemoteService) MapRemoteItemToMedia(ctx context.Context, mount *mod
 	}
 	// 只有远程明确存在图片标签才下发图片 URL。
 	if remoteItemHasImageTag(item, "Primary") {
-		media.PosterURL = r.remoteItemImageURL(cfg, remoteID, "Primary")
+		media.PosterURL = r.remoteItemImageURL(acct, cfg, remoteID, "Primary")
 	}
 	if remoteItemHasImageTag(item, "Backdrop") || len(remoteBackdropTags(item)) > 0 {
-		media.BackdropURL = r.remoteItemImageURL(cfg, remoteID, "Backdrop")
+		media.BackdropURL = r.remoteItemImageURL(acct, cfg, remoteID, "Backdrop")
 	}
 	if ticks := remoteItemInt64(item, "RunTimeTicks"); ticks > 0 {
 		media.DurationSec = int(ticks / 10_000_000)
@@ -307,9 +309,11 @@ func (r *EmbyRemoteService) MapRemoteItemToMedia(ctx context.Context, mount *mod
 		}
 		// 单集通常无独立海报：若远程返回 SeriesPrimaryImageTag（需要
 		// Fields=SeriesPrimaryImage）且系列有图，则回退到系列海报。
-		if media.PosterURL == "" && seriesID != "" &&
-			strings.TrimSpace(remoteItemString(item, "SeriesPrimaryImageTag")) != "" {
-			media.PosterURL = r.remoteItemImageURL(cfg, seriesID, "Primary")
+		if media.PosterURL == "" && seriesID != "" {
+			if seriesTag := strings.TrimSpace(remoteItemString(item, "SeriesPrimaryImageTag")); seriesTag != "" {
+				r.rememberRemoteImageTagValue(embyRemoteAccountID(acct), seriesID, "Primary", seriesTag)
+				media.PosterURL = r.remoteItemImageURL(acct, cfg, seriesID, "Primary")
+			}
 		}
 	default: // Movie / Series / Season / Folder
 		media.SeasonNum = 0
@@ -914,8 +918,9 @@ func remoteItemGenres(item map[string]any) string {
 	return strings.Join(parts, ",")
 }
 
-// remoteItemImageURL 构造远程条目图片绝对地址（带 api_key；前端经 /api/img 代理）。
-func (r *EmbyRemoteService) remoteItemImageURL(cfg *EmbyRemoteConfig, remoteID, imageType string) string {
+// remoteItemImageURL 构造远程条目图片绝对地址（带 api_key 与图片 tag；前端经
+// /api/img 代理）。tag 来自载荷的 ImageTags，用于远端换图后缓存失效。
+func (r *EmbyRemoteService) remoteItemImageURL(acct *model.StrmAccount, cfg *EmbyRemoteConfig, remoteID, imageType string) string {
 	if remoteID == "" {
 		return ""
 	}
@@ -924,7 +929,7 @@ func (r *EmbyRemoteService) remoteItemImageURL(cfg *EmbyRemoteConfig, remoteID,
 		imageType = "primary"
 	}
 	return r.embyBase(cfg) + "/Items/" + url.PathEscape(remoteID) + "/Images/" + url.PathEscape(imageType) +
-		"?api_key=" + url.QueryEscape(cfg.Token)
+		"?api_key=" + url.QueryEscape(cfg.Token) + r.remoteImageTagQuery(embyRemoteAccountID(acct), remoteID, imageType)
 }
 
 // remoteItemHasImageTag 远程 item 是否带某类型图片标签（Emby 的 ImageTags map）。

internal/service/emby_season_index_test.go

@@ -0,0 +1,231 @@
+package service
+
+import (
+	"testing"
+
+	"github.com/truewhile/MeBox/internal/model"
+)
+
+// seasonIndexFixture builds one series covering every season-numbering class a
+// client can ask for: two regular seasons, a generic specials bucket (season 0)
+// and an OVA folder (negative season).
+func seasonIndexFixture(t *testing.T) (*EmbyService, string) {
+	t.Helper()
+	svc := newTestEmbyService(t)
+	lib := model.Library{Name: "剧集", Path: `F:\media\剧集`, Type: "tv", Enabled: true}
+	if err := svc.repo.Library.Create(t.Context(), &lib); err != nil {
+		t.Fatalf("create library: %v", err)
+	}
+	rows := []model.Media{
+		{
+			Base:       model.Base{ID: "s1e1"},
+			LibraryID:  lib.ID,
+			Title:      "權力的遊戲",
+			Path:       `F:\media\剧集\權力的遊戲 (2011)\Season 01\權力的遊戲 - S01E01.mkv`,
+			SeasonNum:  1,
+			EpisodeNum: 1,
+		},
+		{
+			Base:       model.Base{ID: "s1e2"},
+			LibraryID:  lib.ID,
+			Title:      "權力的遊戲",
+			Path:       `F:\media\剧集\權力的遊戲 (2011)\Season 01\權力的遊戲 - S01E02.mkv`,
+			SeasonNum:  1,
+			EpisodeNum: 2,
+		},
+		{
+			Base:       model.Base{ID: "s2e1"},
+			LibraryID:  lib.ID,
+			Title:      "權力的遊戲",
+			Path:       `F:\media\剧集\權力的遊戲 (2011)\Season 02\權力的遊戲 - S02E01.mkv`,
+			SeasonNum:  2,
+			EpisodeNum: 1,
+		},
+		{
+			Base:       model.Base{ID: "spec1"},
+			LibraryID:  lib.ID,
+			Title:      "權力的遊戲",
+			Path:       `F:\media\剧集\權力的遊戲 (2011)\Specials\權力的遊戲 - S00E01.mkv`,
+			SeasonNum:  0,
+			EpisodeNum: 1,
+		},
+		{
+			Base:       model.Base{ID: "ova1"},
+			LibraryID:  lib.ID,
+			Title:      "權力的遊戲",
+			Path:       `F:\media\剧集\權力的遊戲 (2011)\OVA\權力的遊戲 - OVA.mkv`,
+			SeasonNum:  0,
+			EpisodeNum: 1,
+		},
+	}
+	for i := range rows {
+		if err := svc.repo.DB.Create(&rows[i]).Error; err != nil {
+			t.Fatalf("create media: %v", err)
+		}
+	}
+
+	items, err := svc.Items(t.Context(), ItemsParams{
+		ParentID:         lib.ID,
+		IncludeItemTypes: []string{"Series"},
+		Limit:            10,
+	})
+	if err != nil {
+		t.Fatalf("series items: %v", err)
+	}
+	cards := items["Items"].([]map[string]any)
+	if len(cards) != 1 {
+		t.Fatalf("want one series card, got %#v", cards)
+	}
+	return svc, cards[0]["Id"].(string)
+}
+
+func episodeIDs(t *testing.T, out map[string]any) []string {
+	t.Helper()
+	items, ok := out["Items"].([]map[string]any)
+	if !ok {
+		t.Fatalf("Items has unexpected type: %#v", out["Items"])
+	}
+	ids := make([]string, 0, len(items))
+	for _, item := range items {
+		ids = append(ids, item["Id"].(string))
+	}
+	return ids
+}
+
+// itemsTotal reads TotalRecordCount, which is int on the episode path and int64
+// on the empty-envelope path.
+func itemsTotal(t *testing.T, out map[string]any) int {
+	t.Helper()
+	switch v := out["TotalRecordCount"].(type) {
+	case int:
+		return v
+	case int64:
+		return int(v)
+	default:
+		t.Fatalf("TotalRecordCount has unexpected type: %#v", out["TotalRecordCount"])
+		return 0
+	}
+}
+
+func assertEpisodeIDs(t *testing.T, got, want []string) {
+	t.Helper()
+	if len(got) != len(want) {
+		t.Fatalf("episode ids = %#v, want %#v", got, want)
+	}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Fatalf("episode ids = %#v, want %#v", got, want)
+		}
+	}
+}
+
+// TestEmbyEpisodeItemsFilterBySeasonIndex is the regression guard for clients
+// that scope episodes by season number (Season=N) instead of the virtual season
+// id: without the filter every season was returned for any requested season.
+func TestEmbyEpisodeItemsFilterBySeasonIndex(t *testing.T) {
+	svc, seriesID := seasonIndexFixture(t)
+
+	cases := []struct {
+		name      string
+		season    *int
+		wantIDs   []string
+		wantTotal int
+	}{
+		{
+			name:   "unfiltered returns every episode",
+			season: nil,
+			// episodeItems sorts by the stored season/episode numbers, so the two
+			// specials (both stored as season 0) come first in creation order.
+			wantIDs:   []string{"spec1", "ova1", "s1e1", "s1e2", "s2e1"},
+			wantTotal: 5,
+		},
+		{name: "season 1", season: intPtr(1), wantIDs: []string{"s1e1", "s1e2"}, wantTotal: 2},
+		{name: "season 2", season: intPtr(2), wantIDs: []string{"s2e1"}, wantTotal: 1},
+		{name: "season 0 is the specials bucket", season: intPtr(0), wantIDs: []string{"spec1"}, wantTotal: 1},
+		{name: "OVA season", season: intPtr(embySeasonOVA), wantIDs: []string{"ova1"}, wantTotal: 1},
+		{name: "unknown season is empty", season: intPtr(9), wantIDs: []string{}, wantTotal: 0},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			out, err := svc.Items(t.Context(), ItemsParams{
+				ParentID:         seriesID,
+				IncludeItemTypes: []string{"Episode"},
+				Recursive:        true,
+				Limit:            50,
+				SeasonIndex:      tc.season,
+			})
+			if err != nil {
+				t.Fatalf("items: %v", err)
+			}
+			assertEpisodeIDs(t, episodeIDs(t, out), tc.wantIDs)
+			if total := itemsTotal(t, out); total != tc.wantTotal {
+				t.Fatalf("TotalRecordCount = %d, want %d", total, tc.wantTotal)
+			}
+		})
+	}
+}
+
+// TestEmbySeasonIndexFilterKeepsSeasonIdsWorking pins the pre-existing
+// SeasonId behaviour: resolving through the virtual season id must still return
+// exactly that season.
+func TestEmbySeasonIndexFilterKeepsSeasonIdsWorking(t *testing.T) {
+	svc, seriesID := seasonIndexFixture(t)
+
+	items, err := svc.Items(t.Context(), ItemsParams{ParentID: seriesID, Limit: 50})
+	if err != nil {
+		t.Fatalf("seasons: %v", err)
+	}
+	seasons := items["Items"].([]map[string]any)
+	var seasonTwoID string
+	for _, season := range seasons {
+		if season["IndexNumber"] == 2 {
+			seasonTwoID = season["Id"].(string)
+		}
+	}
+	if seasonTwoID == "" {
+		t.Fatalf("season 2 missing from payload: %#v", seasons)
+	}
+
+	out, err := svc.Items(t.Context(), ItemsParams{
+		ParentID:         seasonTwoID,
+		IncludeItemTypes: []string{"Episode"},
+		Recursive:        true,
+		Limit:            50,
+	})
+	if err != nil {
+		t.Fatalf("season episodes: %v", err)
+	}
+	got := episodeIDs(t, out)
+	assertEpisodeIDs(t, got, []string{"s2e1"})
+}
+
+// TestEmbyItemsCacheKeySeparatesSeasonFilters guards against a cached unfiltered
+// response being reused for a season-scoped request (and vice versa).
+func TestEmbyItemsCacheKeySeparatesSeasonFilters(t *testing.T) {
+	svc := newTestEmbyService(t)
+	base := ItemsParams{ParentID: "series-1", IncludeItemTypes: []string{"Episode"}, Limit: 50}
+
+	unfiltered := svc.embyItemsCacheKey("items", base)
+	seasonZero := svc.embyItemsCacheKey("items", withSeasonIndex(base, 0))
+	seasonTwo := svc.embyItemsCacheKey("items", withSeasonIndex(base, 2))
+
+	if unfiltered == seasonZero {
+		t.Fatal("unfiltered and Season=0 keys must differ")
+	}
+	if unfiltered == seasonTwo {
+		t.Fatal("unfiltered and Season=2 keys must differ")
+	}
+	if seasonZero == seasonTwo {
+		t.Fatal("Season=0 and Season=2 keys must differ")
+	}
+	if again := svc.embyItemsCacheKey("items", withSeasonIndex(base, 2)); again != seasonTwo {
+		t.Fatal("cache key must be stable for the same season filter")
+	}
+}
+
+func withSeasonIndex(p ItemsParams, season int) ItemsParams {
+	p.SeasonIndex = intPtr(season)
+	return p
+}
+
+func intPtr(v int) *int { return &v }

internal/service/emby_series_ids.go

@@ -154,6 +154,16 @@ func embySeasonNumForMedia(m *model.Media) int {
 	return m.SeasonNum
 }
 
+// embyRowMatchesSeasonIndex 判断某条剧集是否属于客户端请求的季序号。
+// seasonIndex 为 nil 时不参与过滤。判定复用 embySeasonNumForMedia，保证与
+// /Shows/{id}/Seasons 暴露的 IndexNumber 完全一致（含特别篇的负数分类）。
+func embyRowMatchesSeasonIndex(m *model.Media, seasonIndex *int) bool {
+	if seasonIndex == nil {
+		return true
+	}
+	return embySeasonNumForMedia(m) == *seasonIndex
+}
+
 func embySeasonCandidates(seasonNum int) []int {
 	if seasonNum > 0 {
 		return []int{seasonNum}

internal/service/ffprobe.go

@@ -27,8 +27,19 @@ type FFprobeService struct {
 	log     *zap.Logger
 	mu      sync.RWMutex
 	limiter chan struct{}
+
+	// availMu guards the short-lived availability cache used by Available().
+	availMu        sync.Mutex
+	availCheckedAt time.Time
+	availValue     bool
 }
 
+// ffprobeAvailabilityTTL bounds how long an Available() result is reused.
+// Resolving a binary stats up to a couple dozen candidate paths, and the scanner
+// consults availability per root while deciding whether to queue backfill
+// probes. A short TTL keeps a freshly installed ffmpeg visible within seconds.
+const ffprobeAvailabilityTTL = 30 * time.Second
+
 // NewFFprobeService is the constructor.
 func NewFFprobeService(cfg *config.Config, log *zap.Logger) *FFprobeService {
 	maxConcurrent := normalizeFFprobeMaxConcurrent(cfg.App.FFprobeMaxConcurrent)
@@ -54,6 +65,29 @@ func (f *FFprobeService) SetMaxConcurrent(n int) {
 	f.limiter = make(chan struct{}, normalizeFFprobeMaxConcurrent(n))
 }
 
+// Available reports whether a probe can actually run right now: either an
+// ffprobe binary or the ffmpeg fallback must be resolvable. The scan path uses
+// this to decide whether re-queueing probes for media that still lack technical
+// metadata is worth doing — without a binary every probe would just fail.
+func (f *FFprobeService) Available() bool {
+	if f == nil || f.cfg == nil {
+		return false
+	}
+	f.availMu.Lock()
+	defer f.availMu.Unlock()
+	if !f.availCheckedAt.IsZero() && time.Since(f.availCheckedAt) < ffprobeAvailabilityTTL {
+		return f.availValue
+	}
+	available := true
+	if _, err := resolveLocalExecutable(f.cfg.App.FFprobePath, "ffprobe"); err != nil {
+		_, ffmpegErr := resolveLocalExecutable(f.cfg.App.FFmpegPath, "ffmpeg")
+		available = ffmpegErr == nil
+	}
+	f.availValue = available
+	f.availCheckedAt = time.Now()
+	return available
+}
+
 // ProbeResult is the subset of ffprobe output consumed by the scanner.
 type ProbeResult struct {
 	DurationSec int

internal/service/image_proxy.go

@@ -38,6 +38,12 @@ type ImageProxy struct {
 	mu         sync.Mutex
 	fetchGroup singleflight.Group
 
+	// directClient bypasses HTTP_PROXY / OS proxy settings. It is built once and
+	// shared: rebuilding the transport on every fetch discarded all keep-alive
+	// connections, so every burst of poster requests paid a fresh TCP (and TLS)
+	// handshake per image.
+	directClient *http.Client
+
 	// resizeSem bounds concurrent decode/resize jobs. Emby TV clients request
 	// poster grids in bursts; letting every request decode a source image at
 	// once causes CPU and memory spikes that make the whole UI feel sluggish.
@@ -65,6 +71,12 @@ const (
 	imageBrowserCacheControl     = "public, max-age=2592000, immutable"
 	imagePlaceholderCacheControl = "no-store"
 	imageMaxResizeConcurrency    = 4
+
+	// imageOriginalCacheSubdir 存放上游原图（生成各种尺寸的原料，可再生）。
+	imageOriginalCacheSubdir = "originals"
+	// imageRenditionCacheSubdir 存放挂载 Emby 直接按尺寸产出的成品图：它们
+	// 已经是客户端要的最终尺寸，和派生缩略图一样属于长期保留的热路径缓存。
+	imageRenditionCacheSubdir = "renditions"
 )
 
 // NewImageProxy is the constructor.
@@ -113,6 +125,7 @@ func NewImageProxy(cfg *config.Config, log *zap.Logger) *ImageProxy {
 	}
 
 	proxy.client = &http.Client{Timeout: 30 * time.Second, Transport: transport}
+	proxy.directClient = &http.Client{Timeout: 30 * time.Second, Transport: NewInternalTransport()}
 	return proxy
 }
 
@@ -207,10 +220,20 @@ func imageResizeConcurrency() int {
 }
 
 // Prune removes oldest cached images until disk usage is within the configured limit.
+// 原图池按独立配额与保留时长优先淘汰，客户端热路径读取的成品图最后淘汰。
 func (p *ImageProxy) Prune() (PruneImageCacheResult, error) {
 	if p.cfg == nil || p.cfg.Cache.ImagesMaxSizeMB <= 0 {
 		return PruneImageCacheResult{}, nil
 	}
-	maxBytes := int64(p.cfg.Cache.ImagesMaxSizeMB) * 1024 * 1024
-	return PruneImageCache(p.cacheDir, maxBytes)
+	var originalsMaxBytes int64
+	if p.cfg.Cache.ImagesOriginalsMaxSizeMB > 0 {
+		originalsMaxBytes = int64(p.cfg.Cache.ImagesOriginalsMaxSizeMB) * 1024 * 1024
+	}
+	var originalsMaxAge time.Duration
+	if p.cfg.Cache.ImagesOriginalsTTLHours > 0 {
+		originalsMaxAge = time.Duration(p.cfg.Cache.ImagesOriginalsTTLHours) * time.Hour
+	}
+	totalBytes := int64(p.cfg.Cache.ImagesMaxSizeMB) * 1024 * 1024
+	pools := ImageCachePools(p.cacheDir, originalsMaxBytes, originalsMaxAge)
+	return PruneImageCachePools(pools, totalBytes)
 }

internal/service/image_proxy_cache.go

@@ -4,6 +4,8 @@ import (
 	"bytes"
 	"crypto/sha256"
 	"encoding/hex"
+	"errors"
+	"fmt"
 	"io"
 	"net/http"
 	"os"
@@ -68,17 +70,7 @@ func serveCachedPlaceholder(w http.ResponseWriter) {
 	_, _ = w.Write(transparent1x1PNG)
 }
 
-func (p *ImageProxy) cloudImageCachePaths(stableKey string) (string, string, string) {
-	stableKey = strings.TrimSpace(stableKey)
-	if stableKey == "" {
-		stableKey = "unknown"
-	}
-	sum := sha256.Sum256([]byte("cloud-image:" + stableKey))
-	key := "cloud-" + hex.EncodeToString(sum[:])
-	cachePath := filepath.Join(p.cacheDir, key)
-	return key, cachePath, cachePath + ".fail"
-}
-
+// remoteImageCachePaths 返回上游原图池的路径（带 URL 校验）。
 func (p *ImageProxy) remoteImageCachePaths(raw string) (string, string, string, error) {
 	if _, err := p.validateURL(raw); err != nil {
 		return "", "", "", err
@@ -87,13 +79,120 @@ func (p *ImageProxy) remoteImageCachePaths(raw string) (string, string, string,
 	return key, cachePath, failPath, nil
 }
 
+// remoteImageCachePathsForValidated 返回上游原图池的缓存路径。
 func (p *ImageProxy) remoteImageCachePathsForValidated(raw string) (string, string, string) {
-	sum := sha256.Sum256([]byte(raw))
+	return p.remoteImageCachePathsInPool(raw, imageOriginalCacheSubdir)
+}
+
+// remoteImageCachePathsInPool 生成某个池内的缓存路径。池决定清理策略：
+// 原图池小配额 + 短保留，成品池（挂载 Emby 按尺寸产出的图）长期保留。
+func (p *ImageProxy) remoteImageCachePathsInPool(raw, pool string) (string, string, string) {
+	sum := sha256.Sum256([]byte(imageCacheKeyURL(raw)))
 	key := hex.EncodeToString(sum[:])
-	cachePath := filepath.Join(p.cacheDir, key)
+	cachePath := filepath.Join(p.cacheDir, pool, key)
 	return key, cachePath, cachePath + ".fail"
 }
 
+// remoteImageCachePathsEveryPool 返回同一个地址在所有池中的缓存路径。调用方
+// 只知道原始 URL，无法判断它是原图还是挂载 Emby 按尺寸返回的成品，因此清理
+// 类操作（refresh/retry）需要两个池都试一遍。
+func (p *ImageProxy) remoteImageCachePathsEveryPool(raw string) [][3]string {
+	pools := []string{imageOriginalCacheSubdir, imageRenditionCacheSubdir}
+	out := make([][3]string, 0, len(pools))
+	for _, pool := range pools {
+		key, cachePath, failPath := p.remoteImageCachePathsInPool(raw, pool)
+		out = append(out, [3]string{key, cachePath, failPath})
+	}
+	return out
+}
+
+// imageProxyMaxDownloadBytes 是单张图片的下载上限，防上游异常返回超大响应
+// 把磁盘和内存打满。
+const imageProxyMaxDownloadBytes = 32 << 20
+
+var (
+	// errImageCacheUnavailable 表示缓存目录不可写、连临时文件都建不出来。
+	// 它发生在读取响应体之前，因此调用方还能退回内存缓冲，保证图片不因
+	// 运维异常（磁盘只读/满）全部变成占位图。
+	errImageCacheUnavailable = errors.New("image cache is not writable")
+	// errImageCacheCommitFailed 表示响应体已被消费但无法提交到缓存（写入
+	// 中断、rename 失败等），此时无法重放响应体，只能按拉取失败处理。
+	errImageCacheCommitFailed = errors.New("image cache commit failed")
+)
+
+// cachedImageFileValid 只读取文件头判断缓存文件是否是可用图片。透明占位图
+// 与空文件都视为不可用（历史实现曾把上游失败时的占位图写进缓存）。
+func cachedImageFileValid(path string) error {
+	file, err := os.Open(path) // #nosec G304 -- cache paths are SHA-derived under cacheDir.
+	if err != nil {
+		return err
+	}
+	defer file.Close()
+	stat, err := file.Stat()
+	if err != nil || stat.IsDir() || stat.Size() <= 0 {
+		return errImageProxyNonImageContent
+	}
+	headerSize := 512
+	if stat.Size() < int64(headerSize) {
+		headerSize = int(stat.Size())
+	}
+	header := make([]byte, headerSize)
+	n, readErr := io.ReadFull(file, header)
+	if readErr != nil && !errors.Is(readErr, io.ErrUnexpectedEOF) {
+		return readErr
+	}
+	header = header[:n]
+	if !isImageContentType(detectContentType(header)) {
+		return errImageProxyNonImageContent
+	}
+	// 透明占位图恰好 67 字节；只看头部就够（detectContentType 也只读前 512 字节）。
+	if n == len(transparent1x1PNG) && bytes.Equal(header, transparent1x1PNG) {
+		return errImageProxyNonImageContent
+	}
+	return nil
+}
+
+// streamImageToCache 把上游响应体流式写入临时文件，校验确为图片后原子替换到
+// cachePath。原图常有数兆字节，旧实现每次都要先整张读进内存再写盘，在小内存
+// 主机上几个并发海报请求就能把内存顶满。
+func (p *ImageProxy) streamImageToCache(cachePath, failPath string, body io.Reader) error {
+	if err := os.MkdirAll(filepath.Dir(cachePath), 0o750); err != nil {
+		p.warn("imageproxy: cache dir failed", err)
+		return errImageCacheUnavailable
+	}
+	tmp, err := os.CreateTemp(p.cacheDir, "img-*.tmp")
+	if err != nil {
+		p.warn("imageproxy: cache temp file failed", err)
+		return errImageCacheUnavailable
+	}
+	tmpName := tmp.Name()
+	discard := func() { _ = os.Remove(tmpName) }
+
+	written, copyErr := io.Copy(tmp, io.LimitReader(body, imageProxyMaxDownloadBytes))
+	closeErr := tmp.Close()
+	if copyErr != nil || closeErr != nil {
+		discard()
+		if copyErr != nil {
+			return copyErr
+		}
+		return fmt.Errorf("%w: %v", errImageCacheCommitFailed, closeErr)
+	}
+	if written == 0 {
+		discard()
+		return errors.New("upstream image body is empty")
+	}
+	if err := cachedImageFileValid(tmpName); err != nil {
+		discard()
+		return err
+	}
+	if err := os.Rename(tmpName, cachePath); err != nil {
+		discard()
+		return fmt.Errorf("%w: %v", errImageCacheCommitFailed, err)
+	}
+	_ = os.Remove(failPath)
+	return nil
+}
+
 func serveCachedImageFile(w http.ResponseWriter, r *http.Request, key, cachePath string) bool {
 	return serveImageFile(w, r, key, cachePath, imageBrowserCacheControl)
 }
@@ -131,13 +230,6 @@ func imageFileETag(key string, stat os.FileInfo) string {
 	return `"img-` + hex.EncodeToString(sum[:8]) + "-" + strconv.FormatInt(stat.Size(), 16) + "-" + strconv.FormatInt(stat.ModTime().Unix(), 16) + `"`
 }
 
-func freshNegativeImageCache(failPath string) bool {
-	if _, err := os.Stat(failPath); err == nil {
-		_ = os.Remove(failPath)
-	}
-	return false
-}
-
 func (p *ImageProxy) markImageFetchFailed(failPath string) {
 	if err := os.MkdirAll(filepath.Dir(failPath), 0o750); err != nil {
 		return

internal/service/image_proxy_remote.go

@@ -4,53 +4,80 @@ import (
 	"bytes"
 	"context"
 	"errors"
-	"io"
 	"net/http"
 	"os"
 	"path/filepath"
 	"strings"
 	"time"
-
-	"go.uber.org/zap"
 )
 
 var errImageProxyRequestSetup = errors.New("image proxy request setup failed")
 var errImageProxyNonImageContent = errors.New("upstream returned non-image content")
 
+// prefetchCardResizeOptions 对应前端 ARTWORK.posterCard（见 web/src/api/client.ts）：
+// 卡片是海报墙最常请求的档位，刮削阶段预生成它能让首个列表请求直接命中缓存。
+// 若前端调整该预设，这里只是白生成一份用不到的档位（约几十 KB），不影响正确性。
+var prefetchCardResizeOptions = imageResizeOptions{MaxWidth: 480, MaxHeight: 600, Quality: 80}
+
 func (p *ImageProxy) PrefetchRemote(ctx context.Context, raw string) error {
 	_, _, err := p.Fetch(ctx, raw)
 	return err
 }
 
+// PrefetchCardVariant 预取原图后再离线生成卡片档位的缩略图。刮削是后台任务，
+// 在这里做掉解码可以把海报墙首屏的 CPU 抖动移到请求路径之外。
+//
+// 预生成失败不影响预取结果：预取的成功含义是「图片可达」（刮削据此决定是否
+// 替换旧图），而缩略图只是优化，客户端首次请求时会自己生成。
+func (p *ImageProxy) PrefetchCardVariant(ctx context.Context, raw string) error {
+	if err := p.PrefetchRemote(ctx, raw); err != nil {
+		return err
+	}
+	if !isHTTPish(raw) {
+		return nil
+	}
+	_, cachePath, _ := p.remoteImageCachePathsForValidated(raw)
+	if err := p.ensureResizeCache(ctx, cachePath, prefetchCardResizeOptions); err != nil {
+		p.warn("imageproxy: prefetch card variant failed", err)
+	}
+	return nil
+}
+
 func (p *ImageProxy) RemoveCached(raw string) error {
 	if !isHTTPish(raw) {
 		return nil
 	}
-	_, cachePath, failPath, err := p.remoteImageCachePaths(raw)
-	if err != nil {
+	if _, err := p.validateURL(raw); err != nil {
 		return nil
 	}
-	if err := os.Remove(cachePath); err != nil && !errors.Is(err, os.ErrNotExist) {
-		return err
+	var firstErr error
+	for _, paths := range p.remoteImageCachePathsEveryPool(raw) {
+		cachePath, failPath := paths[1], paths[2]
+		if err := os.Remove(cachePath); err != nil && !errors.Is(err, os.ErrNotExist) {
+			firstErr = err
+		}
+		if err := os.Remove(failPath); err != nil && !errors.Is(err, os.ErrNotExist) {
+			firstErr = err
+		}
 	}
-	if err := os.Remove(failPath); err != nil && !errors.Is(err, os.ErrNotExist) {
-		return err
-	}
-	return nil
+	return firstErr
 }
 
 func (p *ImageProxy) RemoveFailed(raw string) error {
 	if !isHTTPish(raw) {
 		return nil
 	}
-	_, _, failPath, err := p.remoteImageCachePaths(raw)
-	if err != nil {
+	if _, err := p.validateURL(raw); err != nil {
 		return nil
 	}
-	if err := os.Remove(failPath); err != nil && !errors.Is(err, os.ErrNotExist) {
-		return err
+	var firstErr error
+	for _, paths := range p.remoteImageCachePathsEveryPool(raw) {
+		failPath := paths[2]
+		if err := os.Remove(failPath); err != nil && !errors.Is(err, os.ErrNotExist) {
+			firstErr = err
+		}
 	}
-	return nil
+	return firstErr
 }
 
 // Serve writes the requested image to w. Caller is expected to validate
@@ -87,7 +114,17 @@ func (p *ImageProxy) serveRemoteImage(ctx context.Context, w http.ResponseWriter
 		return err
 	}
 	host := strings.ToLower(u.Host)
-	key, cachePath, failPath := p.remoteImageCachePathsForValidated(raw)
+	// 已配置的远程 Emby 挂载：把尺寸直接转发给远端生成缩略图，缓存键也用
+	// 带尺寸的地址，这样不同尺寸各自缓存互不覆盖。
+	fetchURL := p.upstreamImageFetchURL(raw, opts)
+	// 尺寸被转发给挂载的远端时，上游返回的就是客户端要的最终尺寸成品，归入
+	// 成品池长期保留；其他上游返回的是原图，只是生成各种尺寸的原料，归入
+	// 原图池（小配额 + 短保留）。
+	pool := imageOriginalCacheSubdir
+	if fetchURL != raw {
+		pool = imageRenditionCacheSubdir
+	}
+	key, cachePath, failPath := p.remoteImageCachePathsInPool(fetchURL, pool)
 	forceRefresh := r.URL.Query().Get("refresh") != ""
 	p.removeUnusableImageCache(cachePath, failPath)
 	if !forceRefresh && p.serveCachedImage(w, r, key, cachePath, opts) {
@@ -95,7 +132,7 @@ func (p *ImageProxy) serveRemoteImage(ctx context.Context, w http.ResponseWriter
 	}
 	// No negative caching: a previously failed fetch is retried on every
 	// subsequent request, so the image recovers as soon as upstream does.
-	data, ctype, contentLength, err := p.fetchAndCacheRemoteImageShared(ctx, raw, host, cachePath, failPath)
+	result, err := p.fetchAndCacheRemoteImageShared(ctx, fetchURL, host, cachePath, failPath)
 	if err != nil {
 		if forceRefresh && p.serveCachedImage(w, r, key, cachePath, opts) {
 			return nil
@@ -111,17 +148,16 @@ func (p *ImageProxy) serveRemoteImage(ctx context.Context, w http.ResponseWriter
 	if opts.active() && p.serveResizedFromFile(w, r, cachePath, opts) {
 		return nil
 	}
-	w.Header().Set("Content-Type", ctype)
-	if contentLength != "" {
-		w.Header().Set("Content-Length", contentLength)
+	// 缓存目录不可写时的内存兜底：图片只在本次响应里直出，不落盘。
+	if len(result.data) > 0 {
+		w.Header().Set("Content-Type", result.contentType)
+		w.Header().Set("Cache-Control", imageBrowserCacheControl)
+		http.ServeContent(w, r, key, time.Now(), bytes.NewReader(result.data))
+		return nil
 	}
-	modTime := time.Now()
-	if stat, err := os.Stat(cachePath); err == nil && stat.Size() > 0 {
-		modTime = stat.ModTime()
-		w.Header().Set("ETag", imageFileETag(key, stat))
+	if !p.serveCachedImage(w, r, key, cachePath, opts) {
+		serveCachedPlaceholder(w)
 	}
-	w.Header().Set("Cache-Control", imageBrowserCacheControl)
-	http.ServeContent(w, r, key, modTime, bytes.NewReader(data))
 	return nil
 }
 
@@ -137,64 +173,32 @@ func (p *ImageProxy) serveCachedImage(w http.ResponseWriter, r *http.Request, ke
 func (p *ImageProxy) removeUnusableImageCache(cachePath, failPath string) {
 	// 只读取文件头判断缓存是否可用。旧实现每次命中远程图片缓存都会把整个
 	// 原图读进内存再丢弃，电视端批量加载海报时会产生大量无意义的磁盘 I/O。
-	file, err := os.Open(cachePath) // #nosec G304 -- cachePath is SHA-derived under cacheDir.
-	if err != nil {
+	if _, err := os.Stat(cachePath); err != nil {
 		return
 	}
-	stat, err := file.Stat()
-	if err != nil || stat.IsDir() || stat.Size() <= 0 {
-		_ = file.Close()
+	if err := cachedImageFileValid(cachePath); err != nil {
 		_ = os.Remove(cachePath)
 		_ = os.Remove(failPath)
-		return
 	}
-	headerSize := 512
-	if stat.Size() < int64(headerSize) {
-		headerSize = int(stat.Size())
-	}
-	header := make([]byte, headerSize)
-	n, readErr := io.ReadFull(file, header)
-	_ = file.Close()
-	if readErr != nil && readErr != io.ErrUnexpectedEOF {
-		_ = os.Remove(cachePath)
-		_ = os.Remove(failPath)
-		return
-	}
-	header = header[:n]
-	// A transparent placeholder is exactly 67 bytes; checking the header alone
-	// is enough for the normal image cache entries (they are much larger but
-	// detectContentType only inspects the same leading 512 bytes anyway).
-	// Close the handle before deleting: Windows refuses to delete an open file.
-	if n > 0 && isImageContentType(detectContentType(header)) &&
-		!(n == len(transparent1x1PNG) && bytes.Equal(header, transparent1x1PNG)) {
-		return
-	}
-	_ = os.Remove(cachePath)
-	_ = os.Remove(failPath)
 }
 
-func (p *ImageProxy) fetchAndCacheRemoteImage(ctx context.Context, raw, host, cachePath, failPath string) ([]byte, string, string, error) {
-	if err := os.MkdirAll(p.cacheDir, 0o750); err != nil {
-		p.log.Warn("imageproxy: mkdir failed", zap.String("dir", p.cacheDir), zap.Error(err))
-		return nil, "", "", errImageProxyRequestSetup
-	}
+func (p *ImageProxy) fetchAndCacheRemoteImage(ctx context.Context, raw, host, cachePath, failPath string) (remoteImageFetchResult, error) {
 	var lastErr error
-	for _, candidate := range p.remoteImageFetchClients() {
-		data, ctype, contentLength, err := p.fetchRemoteImageOnce(ctx, raw, host, candidate)
+	for _, candidate := range p.remoteImageFetchClients(host) {
+		result, err := p.fetchRemoteImageOnce(ctx, raw, host, candidate, cachePath, failPath)
 		if err == nil {
-			p.writeImageCache(cachePath, failPath, "img-*.tmp", data)
-			return data, ctype, contentLength, nil
+			return result, nil
 		}
 		if errors.Is(err, errImageProxyRequestSetup) {
-			return nil, "", "", err
+			return remoteImageFetchResult{}, err
 		}
 		lastErr = err
 	}
 	if p.canUseExternalImageFallback() && isDoubanImageHost(host) {
-		data, ctype, contentLength, err := fetchRemoteImageWithCurl(ctx, raw, host)
+		data, ctype, _, err := fetchRemoteImageWithCurl(ctx, raw, host)
 		if err == nil {
 			p.writeImageCache(cachePath, failPath, "img-*.tmp", data)
-			return data, ctype, contentLength, nil
+			return remoteImageFetchResult{data: data, contentType: ctype}, nil
 		}
 		logImageFetchError(p.log, "imageproxy: curl fallback failed", host, "curl", err)
 		lastErr = err
@@ -203,41 +207,27 @@ func (p *ImageProxy) fetchAndCacheRemoteImage(ctx context.Context, raw, host, ca
 	if lastErr == nil {
 		lastErr = errors.New("upstream image fetch failed")
 	}
-	return nil, "", "", redactSensitiveError(lastErr)
-}
-
-type sharedRemoteImageResult struct {
-	data          []byte
-	contentType   string
-	contentLength string
+	return remoteImageFetchResult{}, redactSensitiveError(lastErr)
 }
 
 // fetchAndCacheRemoteImageShared coalesces concurrent requests for the same
 // upstream image. A poster can appear in the hero, a shelf and the detail page
 // at the same time; without this guard every resize variant may fetch the same
 // original before the first cache write finishes.
-func (p *ImageProxy) fetchAndCacheRemoteImageShared(ctx context.Context, raw, host, cachePath, failPath string) ([]byte, string, string, error) {
+func (p *ImageProxy) fetchAndCacheRemoteImageShared(ctx context.Context, raw, host, cachePath, failPath string) (remoteImageFetchResult, error) {
 	value, err, _ := p.fetchGroup.Do(cachePath, func() (any, error) {
 		loadCtx, cancel := context.WithTimeout(context.WithoutCancel(ctx), 45*time.Second)
 		defer cancel()
-		data, contentType, contentLength, err := p.fetchAndCacheRemoteImage(loadCtx, raw, host, cachePath, failPath)
-		if err != nil {
-			return nil, err
-		}
-		return sharedRemoteImageResult{
-			data:          data,
-			contentType:   contentType,
-			contentLength: contentLength,
-		}, nil
+		return p.fetchAndCacheRemoteImage(loadCtx, raw, host, cachePath, failPath)
 	})
 	if err != nil {
-		return nil, "", "", err
+		return remoteImageFetchResult{}, err
 	}
-	result, ok := value.(sharedRemoteImageResult)
+	result, ok := value.(remoteImageFetchResult)
 	if !ok {
-		return nil, "", "", errors.New("upstream image fetch failed")
+		return remoteImageFetchResult{}, errors.New("upstream image fetch failed")
 	}
-	return result.data, result.contentType, result.contentLength, nil
+	return result, nil
 }
 
 // Fetch pulls a remote image and returns bytes plus Content-Type using cache.
@@ -259,13 +249,29 @@ func (p *ImageProxy) Fetch(ctx context.Context, raw string) ([]byte, string, err
 	}
 	// No negative caching: a previously failed fetch is retried on every
 	// subsequent request, so the image recovers as soon as upstream does.
-	data, ctype, _, err := p.fetchAndCacheRemoteImage(ctx, raw, host, cachePath, failPath)
-	return data, ctype, err
+	result, err := p.fetchAndCacheRemoteImage(ctx, raw, host, cachePath, failPath)
+	if err != nil {
+		return nil, "", err
+	}
+	// 正常路径只落盘，这里按需读回（调用方需要字节）。
+	if len(result.data) > 0 {
+		return result.data, result.contentType, nil
+	}
+	data, err := os.ReadFile(cachePath) // #nosec G304 -- cachePath is SHA-derived under cacheDir.
+	if err != nil || len(data) == 0 {
+		return nil, "", errors.New("cached image is unreadable")
+	}
+	return data, detectContentType(data), nil
 }
 
+// writeImageCache atomically writes the fetched original. The global mutex is
+// deliberately not held: os.CreateTemp already yields a unique name and
+// os.Rename is atomic, so the lock only serialized multi-megabyte disk writes
+// and made one poster's write block every other image in flight.
 func (p *ImageProxy) writeImageCache(cachePath, failPath, pattern string, data []byte) {
-	p.mu.Lock()
-	defer p.mu.Unlock()
+	if err := os.MkdirAll(filepath.Dir(cachePath), 0o750); err != nil {
+		return
+	}
 	tmp, tmpErr := os.CreateTemp(p.cacheDir, pattern)
 	if tmpErr != nil {
 		return

internal/service/image_proxy_remote_fetch.go

@@ -18,23 +18,32 @@ type remoteImageFetchClient struct {
 	client *http.Client
 }
 
-func (p *ImageProxy) remoteImageFetchClients() []remoteImageFetchClient {
+func (p *ImageProxy) remoteImageFetchClients(host string) []remoteImageFetchClient {
 	client := p.client
 	if client == nil {
 		client = NewExternalHTTPClient(30 * time.Second)
 	}
-	clients := []remoteImageFetchClient{{name: "default", client: client}}
-	if _, ok := client.Transport.(*http.Transport); ok {
-		timeout := client.Timeout
-		if timeout <= 0 {
-			timeout = 30 * time.Second
-		}
-		clients = append(clients, remoteImageFetchClient{
-			name:   "direct",
-			client: &http.Client{Timeout: timeout, Transport: NewInternalTransport()},
-		})
+	if _, ok := client.Transport.(*http.Transport); !ok {
+		return []remoteImageFetchClient{{name: "default", client: client}}
 	}
-	return clients
+	timeout := client.Timeout
+	if timeout <= 0 {
+		timeout = 30 * time.Second
+	}
+	direct := p.directClient
+	if direct == nil {
+		direct = &http.Client{Timeout: timeout, Transport: NewInternalTransport()}
+	}
+	directCandidate := remoteImageFetchClient{name: "direct", client: direct}
+	defaultCandidate := remoteImageFetchClient{name: "default", client: client}
+	// Hosts the user configured themselves — remote Emby mounts and their image
+	// endpoints — are reached over the LAN or a dedicated line. Routing those
+	// through the OS/env proxy first costs a failed attempt before every single
+	// image, so try the direct client first for them.
+	if p.isAllowedRemoteHost(host) {
+		return []remoteImageFetchClient{directCandidate, defaultCandidate}
+	}
+	return []remoteImageFetchClient{defaultCandidate, directCandidate}
 }
 
 func (p *ImageProxy) canUseExternalImageFallback() bool {
@@ -45,38 +54,63 @@ func (p *ImageProxy) canUseExternalImageFallback() bool {
 	return ok
 }
 
-func (p *ImageProxy) fetchRemoteImageOnce(ctx context.Context, raw, host string, candidate remoteImageFetchClient) ([]byte, string, string, error) {
+// remoteImageFetchResult 是一次成功拉取的结果。正常路径图片已经流式落盘
+// （data 为空，调用方从缓存文件下发/缩放）；只有缓存目录不可写、退回内存
+// 缓冲时才带 data，保证图片仍能发给客户端。
+type remoteImageFetchResult struct {
+	data        []byte
+	contentType string
+}
+
+// fetchRemoteImageOnce 拉取一次上游图片并写入 cachePath。响应体直接流式落盘，
+// 不再整张读进内存。
+func (p *ImageProxy) fetchRemoteImageOnce(ctx context.Context, raw, host string, candidate remoteImageFetchClient, cachePath, failPath string) (remoteImageFetchResult, error) {
 	req, err := http.NewRequestWithContext(ctx, http.MethodGet, raw, nil)
 	if err != nil {
 		p.log.Warn("imageproxy: build request failed", zap.String("url", redactSensitiveURL(raw)), zap.Error(redactSensitiveError(err)))
-		return nil, "", "", errImageProxyRequestSetup
+		return remoteImageFetchResult{}, errImageProxyRequestSetup
 	}
 	applyRemoteImageHeaders(req, host, raw)
 
 	resp, err := candidate.client.Do(req)
 	if err != nil {
 		logImageFetchError(p.log, "imageproxy: upstream fetch failed", host, candidate.name, err)
-		return nil, "", "", err
+		return remoteImageFetchResult{}, err
 	}
 	defer resp.Body.Close()
 	if resp.StatusCode >= 400 {
 		p.log.Warn("imageproxy: upstream returned non-OK", zap.String("host", host), zap.String("client", candidate.name), zap.String("status", resp.Status))
-		return nil, "", "", errors.New("upstream returned " + resp.Status)
+		return remoteImageFetchResult{}, errors.New("upstream returned " + resp.Status)
 	}
-	data, err := io.ReadAll(io.LimitReader(resp.Body, 32<<20))
+	if err := p.streamImageToCache(cachePath, failPath, resp.Body); err != nil {
+		if errors.Is(err, errImageCacheUnavailable) {
+			// 缓存目录不可写：响应体还没读，退回内存缓冲。
+			return p.bufferRemoteImage(resp, host, candidate.name, cachePath, failPath)
+		}
+		logImageFetchError(p.log, "imageproxy: stream image failed", host, candidate.name, err)
+		return remoteImageFetchResult{}, err
+	}
+	return remoteImageFetchResult{}, nil
+}
+
+// bufferRemoteImage 在缓存不可用时把响应体读进内存，校验为图片后尽力写入
+// 缓存（失败也不影响本次下发）。
+func (p *ImageProxy) bufferRemoteImage(resp *http.Response, host, client, cachePath, failPath string) (remoteImageFetchResult, error) {
+	data, err := io.ReadAll(io.LimitReader(resp.Body, imageProxyMaxDownloadBytes))
 	if err != nil || len(data) == 0 {
-		p.log.Warn("imageproxy: read upstream body failed", zap.String("host", host), zap.String("client", candidate.name), zap.Error(redactSensitiveError(err)))
+		p.log.Warn("imageproxy: read upstream body failed", zap.String("host", host), zap.String("client", client), zap.Error(redactSensitiveError(err)))
 		if err == nil {
 			err = errors.New("upstream image body is empty")
 		}
-		return nil, "", "", err
+		return remoteImageFetchResult{}, err
 	}
 	ctype, ok := validImageContentType(data)
 	if !ok {
-		p.log.Warn("imageproxy: upstream returned non-image content", zap.String("host", host), zap.String("client", candidate.name), zap.String("content_type", resp.Header.Get("Content-Type")))
-		return nil, "", "", errImageProxyNonImageContent
+		p.log.Warn("imageproxy: upstream returned non-image content", zap.String("host", host), zap.String("client", client), zap.String("content_type", resp.Header.Get("Content-Type")))
+		return remoteImageFetchResult{}, errImageProxyNonImageContent
 	}
-	return data, ctype, resp.Header.Get("Content-Length"), nil
+	p.writeImageCache(cachePath, failPath, "img-*.tmp", data)
+	return remoteImageFetchResult{data: data, contentType: ctype}, nil
 }
 
 func logImageFetchError(log *zap.Logger, message, host, client string, err error) {

internal/service/image_proxy_stream_test.go

@@ -0,0 +1,188 @@
+package service
+
+import (
+	"bytes"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"net/url"
+	"os"
+	"path/filepath"
+	"sync/atomic"
+	"testing"
+
+	"go.uber.org/zap"
+
+	"github.com/truewhile/MeBox/internal/config"
+)
+
+// newStreamImageProxy 构造一个允许访问指定测试上游的图片代理。
+func newStreamImageProxy(t *testing.T, upstreamURL string) *ImageProxy {
+	t.Helper()
+	parsed, err := url.Parse(upstreamURL)
+	if err != nil {
+		t.Fatalf("parse upstream url: %v", err)
+	}
+	proxy := NewImageProxy(&config.Config{Cache: config.CacheConfig{CacheDir: filepath.Join(t.TempDir(), "cache")}}, zap.NewNop())
+	proxy.SetAllowedRemoteHostsProvider(func() []string { return []string{parsed.Host} })
+	return proxy
+}
+
+// bigTestJPEG 在合法 JPEG 后面补一串数据，用来验证大图是流式落盘而不是
+// 整张读进内存（剧照/原图常有数兆字节）。
+func bigTestJPEG(t *testing.T, extraBytes int) []byte {
+	t.Helper()
+	out := make([]byte, 0, len(testJPEG)+extraBytes)
+	out = append(out, testJPEG...)
+	out = append(out, bytes.Repeat([]byte{0x5a}, extraBytes)...)
+	return out
+}
+
+// TestImageProxyStreamsRemoteImageIntoCacheFile 远程原图必须直接流式写到
+// 缓存文件，后续请求由缓存文件服务（上游只被请求一次）。
+func TestImageProxyStreamsRemoteImageIntoCacheFile(t *testing.T) {
+	payload := bigTestJPEG(t, 3<<20)
+
+	var calls int32
+	upstream := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		atomic.AddInt32(&calls, 1)
+		w.Header().Set("Content-Type", "image/jpeg")
+		_, _ = w.Write(payload)
+	}))
+	defer upstream.Close()
+
+	proxy := newStreamImageProxy(t, upstream.URL)
+	raw := upstream.URL + "/poster.jpg"
+
+	rec := httptest.NewRecorder()
+	if err := proxy.Serve(t.Context(), rec, httptest.NewRequest(http.MethodGet, "/api/img", nil), raw); err != nil {
+		t.Fatalf("Serve failed: %v", err)
+	}
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status = %d, want 200", rec.Code)
+	}
+	if !bytes.Equal(rec.Body.Bytes(), payload) {
+		t.Fatalf("body length = %d, want %d", rec.Body.Len(), len(payload))
+	}
+
+	_, cachePath, failPath := proxy.remoteImageCachePathsForValidated(raw)
+	stat, err := os.Stat(cachePath)
+	if err != nil {
+		t.Fatalf("expected the image to be streamed to %s: %v", cachePath, err)
+	}
+	if stat.Size() != int64(len(payload)) {
+		t.Fatalf("cached size = %d, want %d", stat.Size(), len(payload))
+	}
+	if _, err := os.Stat(failPath); err == nil {
+		t.Fatalf("unexpected failure marker at %s", failPath)
+	}
+	if entries, err := os.ReadDir(filepath.Dir(cachePath)); err == nil {
+		for _, entry := range entries {
+			if filepath.Ext(entry.Name()) == ".tmp" {
+				t.Fatalf("temporary file left behind: %s", entry.Name())
+			}
+		}
+	}
+
+	// 第二次请求命中缓存文件，不再回源。
+	second := httptest.NewRecorder()
+	if err := proxy.Serve(t.Context(), second, httptest.NewRequest(http.MethodGet, "/api/img", nil), raw); err != nil {
+		t.Fatalf("second Serve failed: %v", err)
+	}
+	if !bytes.Equal(second.Body.Bytes(), payload) {
+		t.Fatal("cached response body mismatch")
+	}
+	if got := atomic.LoadInt32(&calls); got != 1 {
+		t.Fatalf("upstream calls = %d, want 1 (second request must be served from cache)", got)
+	}
+}
+
+// TestImageProxyRejectsNonImageUpstreamWithoutCaching 流式写入必须在提交前
+// 校验内容，否则错误页会被永久缓存成图片。
+func TestImageProxyRejectsNonImageUpstreamWithoutCaching(t *testing.T) {
+	upstream := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "text/html")
+		_, _ = w.Write([]byte("<html><body>not an image</body></html>"))
+	}))
+	defer upstream.Close()
+
+	proxy := newStreamImageProxy(t, upstream.URL)
+	raw := upstream.URL + "/poster.jpg"
+
+	rec := httptest.NewRecorder()
+	if err := proxy.Serve(t.Context(), rec, httptest.NewRequest(http.MethodGet, "/api/img", nil), raw); err != nil {
+		t.Fatalf("Serve failed: %v", err)
+	}
+	if rec.Body.Len() != len(transparent1x1PNG) {
+		t.Fatalf("body length = %d, want placeholder %d", rec.Body.Len(), len(transparent1x1PNG))
+	}
+
+	_, cachePath, failPath := proxy.remoteImageCachePathsForValidated(raw)
+	if _, err := os.Stat(cachePath); err == nil {
+		t.Fatal("non-image response must not be cached")
+	}
+	if _, err := os.Stat(failPath); err != nil {
+		t.Fatalf("expected a failure marker next to the cache entry: %v", err)
+	}
+}
+
+// TestImageProxyServesRemoteImageWhenCacheDirUnwritable 缓存目录不可写时
+// （磁盘只读/满）必须退回内存缓冲，不能让所有图片变成占位图。
+func TestImageProxyServesRemoteImageWhenCacheDirUnwritable(t *testing.T) {
+	payload := testJPEG
+	upstream := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "image/jpeg")
+		_, _ = w.Write(payload)
+	}))
+	defer upstream.Close()
+
+	// 把一个普通文件当作目录的父级，MkdirAll 必然失败。
+	root := t.TempDir()
+	blocker := filepath.Join(root, "blocker")
+	if err := os.WriteFile(blocker, []byte("x"), 0o644); err != nil {
+		t.Fatalf("write blocker: %v", err)
+	}
+	proxy := newStreamImageProxy(t, upstream.URL)
+	proxy.cacheDir = filepath.Join(blocker, "cache", "images")
+
+	raw := upstream.URL + "/poster.jpg"
+	rec := httptest.NewRecorder()
+	if err := proxy.Serve(t.Context(), rec, httptest.NewRequest(http.MethodGet, "/api/img", nil), raw); err != nil {
+		t.Fatalf("Serve failed: %v", err)
+	}
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status = %d, want 200", rec.Code)
+	}
+	body, err := io.ReadAll(rec.Body)
+	if err != nil {
+		t.Fatalf("read body: %v", err)
+	}
+	if !bytes.Equal(body, payload) {
+		t.Fatalf("body length = %d, want %d", len(body), len(payload))
+	}
+}
+
+// TestImageProxyFetchReturnsCachedBytes Fetch 仍需返回字节（刮削写元数据用），
+// 内容来自刚写入的缓存文件。
+func TestImageProxyFetchReturnsCachedBytes(t *testing.T) {
+	payload := bigTestJPEG(t, 256<<10)
+	upstream := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "image/jpeg")
+		_, _ = w.Write(payload)
+	}))
+	defer upstream.Close()
+
+	proxy := newStreamImageProxy(t, upstream.URL)
+	raw := upstream.URL + "/poster.jpg"
+
+	data, ctype, err := proxy.Fetch(t.Context(), raw)
+	if err != nil {
+		t.Fatalf("Fetch failed: %v", err)
+	}
+	if !bytes.Equal(data, payload) {
+		t.Fatalf("fetched %d bytes, want %d", len(data), len(payload))
+	}
+	if ctype != "image/jpeg" {
+		t.Fatalf("content type = %q, want image/jpeg", ctype)
+	}
+}

internal/service/image_proxy_upstream.go

@@ -0,0 +1,63 @@
+package service
+
+import (
+	"net/url"
+	"strconv"
+	"strings"
+)
+
+// imageCacheKeyURL normalizes an upstream image URL before it is hashed into a
+// disk cache key. Credential query parameters are dropped: a remote Emby token
+// is per-account, so removing it cannot make two different images collide, but
+// it stops a token rotation from invalidating every cached poster at once.
+func imageCacheKeyURL(raw string) string {
+	u, err := url.Parse(strings.TrimSpace(raw))
+	if err != nil || u.Host == "" {
+		return raw
+	}
+	q := u.Query()
+	dropped := false
+	for key := range q {
+		switch strings.ToLower(key) {
+		case "api_key", "apikey", "x-emby-token", "x-mediabrowser-token":
+			q.Del(key)
+			dropped = true
+		}
+	}
+	if !dropped {
+		return raw
+	}
+	u.RawQuery = q.Encode()
+	return u.String()
+}
+
+// upstreamImageFetchURL forwards the client's thumbnail request to a configured
+// remote Emby mount. The remote server can produce the thumbnail itself from
+// maxWidth/maxHeight/quality, so the proxy transfers a few dozen KB instead of
+// the full-size original and skips the local decode+scale entirely (measured at
+// ~400ms for a single 2892×4096 poster).
+//
+// Other upstreams (TMDb, Douban, adult sites, ...) do not honour these
+// parameters, so their originals are still fetched and scaled locally.
+func (p *ImageProxy) upstreamImageFetchURL(raw string, opts imageResizeOptions) string {
+	if p == nil || !opts.active() {
+		return raw
+	}
+	u, err := url.Parse(raw)
+	if err != nil || u.Host == "" {
+		return raw
+	}
+	if !p.isAllowedRemoteHost(u.Hostname()) {
+		return raw
+	}
+	q := u.Query()
+	if opts.MaxWidth > 0 {
+		q.Set("maxWidth", strconv.Itoa(opts.MaxWidth))
+	}
+	if opts.MaxHeight > 0 {
+		q.Set("maxHeight", strconv.Itoa(opts.MaxHeight))
+	}
+	q.Set("quality", strconv.Itoa(opts.encodingQuality()))
+	u.RawQuery = q.Encode()
+	return u.String()
+}

internal/service/image_proxy_upstream_test.go

@@ -0,0 +1,198 @@
+package service
+
+import (
+	"bytes"
+	"net/http"
+	"net/http/httptest"
+	"net/url"
+	"path/filepath"
+	"strconv"
+	"strings"
+	"sync"
+	"testing"
+
+	"go.uber.org/zap"
+
+	"github.com/truewhile/MeBox/internal/config"
+)
+
+func newUpstreamImageProxy(t *testing.T, allowedHosts ...string) *ImageProxy {
+	t.Helper()
+	proxy := NewImageProxy(&config.Config{Cache: config.CacheConfig{CacheDir: filepath.Join(t.TempDir(), "cache")}}, zap.NewNop())
+	if len(allowedHosts) > 0 {
+		proxy.SetAllowedRemoteHostsProvider(func() []string { return allowedHosts })
+	}
+	return proxy
+}
+
+func TestUpstreamImageFetchURLForwardsResizeToConfiguredMount(t *testing.T) {
+	proxy := newUpstreamImageProxy(t, "192.168.1.50:8096")
+	raw := "http://192.168.1.50:8096/emby/Items/abc/Images/Primary?api_key=secret"
+
+	got := proxy.upstreamImageFetchURL(raw, imageResizeOptions{MaxWidth: 480, MaxHeight: 600, Quality: 82})
+	u, err := url.Parse(got)
+	if err != nil {
+		t.Fatal(err)
+	}
+	q := u.Query()
+	if q.Get("maxWidth") != "480" || q.Get("maxHeight") != "600" || q.Get("quality") != "82" {
+		t.Fatalf("forwarded query = %q, want maxWidth=480 maxHeight=600 quality=82", u.RawQuery)
+	}
+	if q.Get("api_key") != "secret" {
+		t.Fatalf("api_key = %q, want the credential preserved", q.Get("api_key"))
+	}
+	if u.Path != "/emby/Items/abc/Images/Primary" {
+		t.Fatalf("path = %q, want the item image path unchanged", u.Path)
+	}
+
+	// The default quality is applied when the client does not send one.
+	got = proxy.upstreamImageFetchURL(raw, imageResizeOptions{MaxWidth: 160})
+	if !strings.Contains(got, "quality="+strconv.Itoa(imageResizeDefaultQuality)) {
+		t.Fatalf("url = %q, want the default encoding quality forwarded", got)
+	}
+}
+
+func TestUpstreamImageFetchURLLeavesOtherUpstreamsAlone(t *testing.T) {
+	proxy := newUpstreamImageProxy(t, "192.168.1.50:8096")
+	opts := imageResizeOptions{MaxWidth: 480, MaxHeight: 600, Quality: 82}
+
+	// Not a configured mount: TMDb/Douban do not honour these parameters, so the
+	// proxy must keep fetching their originals and resize locally.
+	tmdb := "https://image.tmdb.org/t/p/original/poster.jpg"
+	if got := proxy.upstreamImageFetchURL(tmdb, opts); got != tmdb {
+		t.Fatalf("url = %q, want %q unchanged", got, tmdb)
+	}
+	// Configured mount but no size requested: keep the full original.
+	mount := "http://192.168.1.50:8096/emby/Items/abc/Images/Primary?api_key=secret"
+	if got := proxy.upstreamImageFetchURL(mount, imageResizeOptions{}); got != mount {
+		t.Fatalf("url = %q, want %q unchanged", got, mount)
+	}
+}
+
+func TestRemoteImageCacheKeyIgnoresCredentialRotation(t *testing.T) {
+	proxy := newUpstreamImageProxy(t)
+	first := "http://192.168.1.50:8096/emby/Items/abc/Images/Primary?api_key=old-token"
+	second := "http://192.168.1.50:8096/emby/Items/abc/Images/Primary?api_key=new-token"
+
+	_, firstPath, _ := proxy.remoteImageCachePathsForValidated(first)
+	_, secondPath, _ := proxy.remoteImageCachePathsForValidated(second)
+	if firstPath != secondPath {
+		t.Fatalf("token rotation changed the cache key: %q vs %q", firstPath, secondPath)
+	}
+
+	other := "http://192.168.1.50:8096/emby/Items/zzz/Images/Primary?api_key=new-token"
+	_, otherPath, _ := proxy.remoteImageCachePathsForValidated(other)
+	if otherPath == secondPath {
+		t.Fatal("different items must not share a cache key")
+	}
+
+	// A generic `token` parameter can be part of a signed URL's identity, so it
+	// must stay in the key. Only credential-ish names are dropped.
+	signed := "http://cdn.example.com/img/1.jpg?token=aaa"
+	signedOther := "http://cdn.example.com/img/1.jpg?token=bbb"
+	_, signedPath, _ := proxy.remoteImageCachePathsForValidated(signed)
+	_, signedOtherPath, _ := proxy.remoteImageCachePathsForValidated(signedOther)
+	if signedPath == signedOtherPath {
+		t.Fatal("generic token query parameters must remain part of the cache key")
+	}
+}
+
+func TestRemoteImageFetchClientsPreferDirectForConfiguredMount(t *testing.T) {
+	proxy := newUpstreamImageProxy(t, "192.168.1.50:8096")
+
+	mountClients := proxy.remoteImageFetchClients("192.168.1.50:8096")
+	if len(mountClients) != 2 || mountClients[0].name != "direct" || mountClients[1].name != "default" {
+		t.Fatalf("mount client order = %+v, want direct first", clientNames(mountClients))
+	}
+	if mountClients[0].client != proxy.directClient {
+		t.Fatal("direct fetches must reuse the shared no-proxy client")
+	}
+
+	cdnClients := proxy.remoteImageFetchClients("image.tmdb.org")
+	if len(cdnClients) != 2 || cdnClients[0].name != "default" || cdnClients[1].name != "direct" {
+		t.Fatalf("cdn client order = %+v, want default first", clientNames(cdnClients))
+	}
+}
+
+func clientNames(clients []remoteImageFetchClient) []string {
+	out := make([]string, 0, len(clients))
+	for _, c := range clients {
+		out = append(out, c.name)
+	}
+	return out
+}
+
+// TestServeRemoteImageForwardsResizeAndCachesPerSize is the regression test for
+// sluggish artwork on mounted Emby libraries: MeBox used to download the remote
+// original and scale it locally for every requested size. The size must now be
+// forwarded to the mount, and each size must get its own cache entry.
+func TestServeRemoteImageForwardsResizeAndCachesPerSize(t *testing.T) {
+	var mu sync.Mutex
+	queries := []url.Values{}
+	upstream := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		mu.Lock()
+		queries = append(queries, r.URL.Query())
+		mu.Unlock()
+		w.Header().Set("Content-Type", "image/jpeg")
+		_, _ = w.Write(testJPEG)
+	}))
+	defer upstream.Close()
+
+	u, err := url.Parse(upstream.URL)
+	if err != nil {
+		t.Fatal(err)
+	}
+	proxy := newUpstreamImageProxy(t, u.Host)
+	raw := upstream.URL + "/emby/Items/abc/Images/Primary?api_key=secret"
+
+	serve := func(t *testing.T, query string) *httptest.ResponseRecorder {
+		t.Helper()
+		rec := httptest.NewRecorder()
+		req := httptest.NewRequest(http.MethodGet, "/api/img?"+query, nil)
+		if err := proxy.Serve(t.Context(), rec, req, raw); err != nil {
+			t.Fatal(err)
+		}
+		if rec.Code != http.StatusOK {
+			t.Fatalf("status = %d, want 200", rec.Code)
+		}
+		return rec
+	}
+
+	rec := serve(t, "maxWidth=480&maxHeight=600&quality=82")
+	if got := rec.Body.Bytes(); !bytes.Equal(got, testJPEG) {
+		t.Fatalf("body = %x, want the upstream image", got)
+	}
+
+	mu.Lock()
+	first := append([]url.Values(nil), queries...)
+	mu.Unlock()
+	if len(first) != 1 {
+		t.Fatalf("upstream calls = %d, want 1", len(first))
+	}
+	if first[0].Get("maxWidth") != "480" || first[0].Get("maxHeight") != "600" || first[0].Get("quality") != "82" {
+		t.Fatalf("upstream query = %q, want the client's thumbnail request forwarded", first[0].Encode())
+	}
+
+	// Same size again: served from the disk cache, no second upstream call.
+	serve(t, "maxWidth=480&maxHeight=600&quality=82")
+	mu.Lock()
+	if len(queries) != 1 {
+		mu.Unlock()
+		t.Fatalf("upstream calls = %d, want 1 after a cache hit", len(queries))
+	}
+	mu.Unlock()
+
+	// A different size must be its own cache entry, not a re-use of the 480px file.
+	serve(t, "maxWidth=160&quality=60")
+	mu.Lock()
+	defer mu.Unlock()
+	if len(queries) != 2 {
+		t.Fatalf("upstream calls = %d, want 2 for two distinct sizes", len(queries))
+	}
+	if got := queries[1].Get("maxWidth"); got != "160" {
+		t.Fatalf("second upstream maxWidth = %q, want 160", got)
+	}
+	if got := queries[1].Get("quality"); got != "60" {
+		t.Fatalf("second upstream quality = %q, want 60", got)
+	}
+}

internal/service/image_resize.go

@@ -34,7 +34,16 @@ const (
 	// N 像素的图在内存中约需 4N 字节；没有上限时，一张异常的超大图
 	// 就能在多张并发缩略图请求下打爆小内存主机。超过该上限时直接
 	// 回退为原图直出，宁可不缩放也不冒 OOM 风险。
-	imageResizeMaxSourcePixels = 30_000_000
+	//
+	// 1200 万像素覆盖常见的高清海报（2892×4096 约 1180 万），单张解码
+	// 峰值约 48MB RGBA；原先的 3000 万在 2 核 2GB 的机器上意味着单张
+	// 就可能吃掉 120MB 以上，两个并发槽足以触发 OOM/大量换页。
+	imageResizeMaxSourcePixels = 12_000_000
+
+	// imageResizeCheapScaleRatio 是启用低成本插值的缩小比例阈值：目标尺寸
+	// 小于源图一半时，CatmullRom 的收益肉眼不可见，但耗时和临时缓冲明显
+	// 更高（电视端海报墙会同时请求几十张缩略图）。
+	imageResizeCheapScaleRatio = 0.5
 
 	// imageResizeCacheSubdir 存放缩放结果，与远程原图缓存分开放，
 	// 便于单独清理且不与原始字节流缓存互相覆盖。
@@ -156,7 +165,7 @@ func resizeImageData(data []byte, o imageResizeOptions) (out []byte, ctype strin
 		return nil, "", false, err
 	}
 	dst := image.NewRGBA(image.Rect(0, 0, dstW, dstH))
-	draw.CatmullRom.Scale(dst, dst.Bounds(), src, src.Bounds(), draw.Src, nil)
+	resizeScaled(dst, src, cfg.Width, cfg.Height)
 
 	// 只有可能带透明的源格式才需要逐像素确认，避免 JPEG 的无谓遍历。
 	// 写实海报的 PNG 通常比等价 JPEG 大一个数量级，因此在确认不含透明
@@ -176,12 +185,42 @@ func resizeImageData(data []byte, o imageResizeOptions) (out []byte, ctype strin
 	return buf.Bytes(), "image/jpeg", false, nil
 }
 
-// isOpaqueImage 逐像素确认图像不含透明像素。
+// useCheapScaleInterpolator 报告这次缩放是否该用低成本插值：目标尺寸在宽高
+// 两个方向都缩到源图一半以下时，CatmullRom 的额外采样换来的观感差异不可见。
+func useCheapScaleInterpolator(dstW, dstH, srcW, srcH int) bool {
+	if srcW <= 0 || srcH <= 0 || dstW <= 0 || dstH <= 0 {
+		return false
+	}
+	return float64(dstW)/float64(srcW) <= imageResizeCheapScaleRatio &&
+		float64(dstH)/float64(srcH) <= imageResizeCheapScaleRatio
+}
+
+// resizeScaled 把 src 缩放到 dst。大幅缩小时改用低成本插值：CatmullRom
+// 与 ApproxBiLinear 在大比例缩小下观感差异看不出来，但前者要遍历更多
+// 邻域样本，在 2 核机器上会明显拖慢海报墙的并发缩略图请求。
+func resizeScaled(dst *image.RGBA, src image.Image, srcW, srcH int) {
+	if dst == nil {
+		return
+	}
+	bounds := dst.Bounds()
+	interp := draw.Interpolator(draw.CatmullRom)
+	if useCheapScaleInterpolator(bounds.Dx(), bounds.Dy(), srcW, srcH) {
+		interp = draw.ApproxBiLinear
+	}
+	interp.Scale(dst, bounds, src, src.Bounds(), draw.Src, nil)
+}
+
+// isOpaqueImage 逐像素确认图像不含透明像素。dst 已知是 *image.RGBA，用
+// RGBAAt 直取字段可以避免 At() 的接口分派与颜色模型换算（PNG 海报每次
+// 生成缩略图都要走一遍全图扫描）。
 func isOpaqueImage(img *image.RGBA) bool {
+	if img == nil {
+		return false
+	}
 	bounds := img.Bounds()
 	for y := bounds.Min.Y; y < bounds.Max.Y; y++ {
 		for x := bounds.Min.X; x < bounds.Max.X; x++ {
-			if _, _, _, a := img.At(x, y).RGBA(); a != 0xffff {
+			if img.RGBAAt(x, y).A != 0xff {
 				return false
 			}
 		}
@@ -189,19 +228,91 @@ func isOpaqueImage(img *image.RGBA) bool {
 	return true
 }
 
-// resizeCacheKey 生成缩放结果的缓存键，覆盖源文件身份（路径 + 大小 +
-// 修改时间）与全部影响输出的参数，源文件被替换后不会命中陈旧缩略图。
-func (o imageResizeOptions) resizeCacheKey(sourceID string, stat os.FileInfo) string {
+// resizeSourceKey 生成源图身份键：只覆盖源文件身份（路径 + 大小 + mtime），不含
+// 目标尺寸。同一源图的所有尺寸档位共用这个前缀，才能互相列举与派生（小图直接
+// 从已缓存的大图缩小，而不是重新解码原图）。
+func (o imageResizeOptions) resizeSourceKey(sourceID string, stat os.FileInfo) string {
 	h := sha256.New()
-	_, _ = fmt.Fprintf(h, "v1|%s|%dx%d|q%d", sourceID, o.MaxWidth, o.MaxHeight, o.encodingQuality())
+	_, _ = fmt.Fprintf(h, "v2|%s", sourceID)
 	if stat != nil {
 		_, _ = fmt.Fprintf(h, "|%d|%d", stat.Size(), stat.ModTime().UnixNano())
 	}
 	return hex.EncodeToString(h.Sum(nil))
 }
 
-func (p *ImageProxy) resizeCachePath(key string) string {
-	return filepath.Join(p.cacheDir, imageResizeCacheSubdir, key+".img")
+// resizeCacheKey 是「源图身份 + 档位」的稳定标识，用于 ETag 等需要区分档位的
+// 场合（磁盘路径只用源图身份 + 文件名后缀）。
+func (o imageResizeOptions) resizeCacheKey(sourceID string, stat os.FileInfo) string {
+	return o.resizeSourceKey(sourceID, stat) + "." + o.resizeVariantSuffix()
+}
+
+func (p *ImageProxy) resizeCachePath(sourceKey string, o imageResizeOptions) string {
+	return filepath.Join(p.cacheDir, imageResizeCacheSubdir, sourceKey+"."+o.resizeVariantSuffix()+".img")
+}
+
+// resizeVariantSuffix 把目标尺寸与质量写进缩放结果的文件名。这样同一源图的
+// 各档尺寸可以被列举出来：请求小图时若已有更大的档位缓存，直接从它缩小即可，
+// 不必再解码多兆字节的原图（解码 480×600 约 1MB RGBA，解码原图可达数十 MB）。
+func (o imageResizeOptions) resizeVariantSuffix() string {
+	return strconv.Itoa(o.MaxWidth) + "x" + strconv.Itoa(o.MaxHeight) + "q" + strconv.Itoa(o.encodingQuality())
+}
+
+// parseResizeVariantSuffix 解析 "<sourceKey>.<w>x<h>q<q>.img" 文件名中的尺寸。
+func parseResizeVariantSuffix(name, sourceKey string) (width, height int, ok bool) {
+	prefix := sourceKey + "."
+	if !strings.HasPrefix(name, prefix) || !strings.HasSuffix(name, ".img") {
+		return 0, 0, false
+	}
+	spec := strings.TrimSuffix(strings.TrimPrefix(name, prefix), ".img")
+	dimensions, quality, found := strings.Cut(spec, "q")
+	if !found {
+		return 0, 0, false
+	}
+	w, h, found := strings.Cut(dimensions, "x")
+	if !found {
+		return 0, 0, false
+	}
+	width, err := strconv.Atoi(w)
+	if err != nil || width < 0 {
+		return 0, 0, false
+	}
+	height, err = strconv.Atoi(h)
+	if err != nil || height < 0 {
+		return 0, 0, false
+	}
+	if _, err := strconv.Atoi(quality); err != nil {
+		return 0, 0, false
+	}
+	return width, height, true
+}
+
+// largerCachedVariant 找出同一源图已缓存的、能覆盖目标尺寸的最小档位（宽高
+// 都不小于目标）。返回空串表示没有可用的档位，调用方回退到解码原图。
+func (p *ImageProxy) largerCachedVariant(key string, o imageResizeOptions) string {
+	pattern := filepath.Join(p.cacheDir, imageResizeCacheSubdir, key+".*.img")
+	matches, err := filepath.Glob(pattern)
+	if err != nil {
+		return ""
+	}
+	best := ""
+	bestPixels := 0
+	for _, path := range matches {
+		width, height, ok := parseResizeVariantSuffix(filepath.Base(path), key)
+		if !ok {
+			continue
+		}
+		if width == o.MaxWidth && height == o.MaxHeight {
+			continue // 精确档位，本该在上面就命中
+		}
+		if width < o.MaxWidth || height < o.MaxHeight {
+			continue
+		}
+		pixels := width * height
+		if best == "" || pixels < bestPixels {
+			best, bestPixels = path, pixels
+		}
+	}
+	return best
 }
 
 // acquireResizeSlot bounds CPU-heavy decode/resize work. Returning false means
@@ -265,7 +376,7 @@ func (p *ImageProxy) serveResizedFromFile(w http.ResponseWriter, r *http.Request
 	// 缓存命中必须发生在读原图和解码之前。否则电视端每次刷新海报墙都会
 	// 把已经是缩略图缓存的原图重新解码、缩放一遍，造成明显的 CPU 抖动。
 	key := o.resizeCacheKey(srcPath, stat)
-	cachePath := p.resizeCachePath(key)
+	cachePath := p.resizeCachePath(o.resizeSourceKey(srcPath, stat), o)
 	if serveCachedImageFile(w, r, key, cachePath) {
 		return true
 	}
@@ -286,7 +397,15 @@ func (p *ImageProxy) serveResizedFromFile(w http.ResponseWriter, r *http.Request
 		return true
 	}
 
-	data, err := os.ReadFile(srcPath) // #nosec G304 -- srcPath comes from an allowed local path or a SHA-derived cache path.
+	// 优先从同一源图已缓存的大尺寸档位缩小：小图（如 160px 模糊占位图）
+	// 往往能直接由已缓存的卡片图派生，代价从“解码多兆字节原图”降为
+	// “解码几百 KB 的档位文件”。
+	sourcePath := srcPath
+	if derived := p.largerCachedVariant(o.resizeSourceKey(srcPath, stat), o); derived != "" {
+		sourcePath = derived
+	}
+
+	data, err := os.ReadFile(sourcePath) // #nosec G304 -- srcPath comes from an allowed local path or a SHA-derived cache path.
 	if err != nil {
 		return false
 	}
@@ -306,6 +425,48 @@ func (p *ImageProxy) serveResizedFromFile(w http.ResponseWriter, r *http.Request
 	return true
 }
 
+// ensureResizeCache 在请求路径之外预先生成某档缩略图（刮削预取用）。已存在
+// 时直接返回，无需调用方再次解码。
+func (p *ImageProxy) ensureResizeCache(ctx context.Context, srcPath string, o imageResizeOptions) error {
+	if p == nil || !o.active() {
+		return nil
+	}
+	stat, err := os.Stat(srcPath)
+	if err != nil || stat.IsDir() || stat.Size() <= 0 {
+		return errors.New("image source unavailable")
+	}
+	sourceKey := o.resizeSourceKey(srcPath, stat)
+	cachePath := p.resizeCachePath(sourceKey, o)
+	if _, err := os.Stat(cachePath); err == nil {
+		return nil
+	}
+	release, ok := p.acquireResizeSlotFor(ctx, stat.Size())
+	if !ok {
+		return errors.New("image resize slot unavailable")
+	}
+	defer release()
+	if _, err := os.Stat(cachePath); err == nil {
+		return nil
+	}
+	sourcePath := srcPath
+	if derived := p.largerCachedVariant(sourceKey, o); derived != "" {
+		sourcePath = derived
+	}
+	data, err := os.ReadFile(sourcePath) // #nosec G304 -- srcPath is an allowed local path or a SHA-derived cache path.
+	if err != nil {
+		return err
+	}
+	out, _, unchanged, err := resizeImageData(data, o)
+	if err != nil {
+		return err
+	}
+	if unchanged {
+		return nil // 源图已在目标尺寸内，客户端会直接使用原文件。
+	}
+	p.writeResizeCache(cachePath, out)
+	return nil
+}
+
 // compactImageSkipBytes 是“直接出原图”的体积上限。超过它的原图（多兆字节
 // 的剧照、未压缩 sidecar）仍然走缩放，避免把大文件直接塞给电视端。
 // ponytail: 200KB 覆盖这台机器上的典型海报（平均约 100KB）；更大的图仍尝试缩放。
@@ -344,14 +505,14 @@ func isCompactWebImage(header []byte) bool {
 }
 
 // writeResizeCache 原子写入缩放结果；失败只记日志，不影响本次响应。
+// 与 writeImageCache 一样不再持有全局锁：临时文件名唯一、rename 原子，
+// 持锁只会把缩略图的写盘和原图的写盘串成一条队。
 func (p *ImageProxy) writeResizeCache(cachePath string, data []byte) {
 	dir := filepath.Dir(cachePath)
 	if err := os.MkdirAll(dir, 0o750); err != nil {
 		p.warn("imageproxy: resize cache mkdir failed", err)
 		return
 	}
-	p.mu.Lock()
-	defer p.mu.Unlock()
 	tmp, err := os.CreateTemp(dir, "resized-*.tmp")
 	if err != nil {
 		p.warn("imageproxy: resize cache temp failed", err)

internal/service/image_resize_budget_test.go

@@ -0,0 +1,113 @@
+package service
+
+import (
+	"bytes"
+	"errors"
+	"image"
+	"image/color"
+	"image/jpeg"
+	"testing"
+)
+
+// encodeTestJPEG 生成一张便于压缩的测试用 JPEG（比 PNG 更适合构造大尺寸样本）。
+func encodeTestJPEG(t *testing.T, w, h int) []byte {
+	t.Helper()
+	img := image.NewRGBA(image.Rect(0, 0, w, h))
+	for y := 0; y < h; y++ {
+		for x := 0; x < w; x++ {
+			img.Set(x, y, color.RGBA{R: uint8(x % 32 * 8), G: uint8(y % 32 * 8), B: 150, A: 255})
+		}
+	}
+	var buf bytes.Buffer
+	if err := jpeg.Encode(&buf, img, &jpeg.Options{Quality: 70}); err != nil {
+		t.Fatalf("encode test jpeg: %v", err)
+	}
+	return buf.Bytes()
+}
+
+// TestResizeImageDataRejectsOverPixelBudget 锁定解码像素预算：超过上限的
+// 原图必须放弃缩放（回退原图直出），而不是在小内存主机上尝试解码。
+func TestResizeImageDataRejectsOverPixelBudget(t *testing.T) {
+	// 12_000_000 像素预算之上：3000×5000 = 1500 万。
+	data := encodeTestJPEG(t, 3000, 5000)
+
+	if _, _, _, err := resizeImageData(data, imageResizeOptions{MaxWidth: 400}); !errors.Is(err, errImageResizeTooLarge) {
+		t.Fatalf("expected errImageResizeTooLarge, got %v", err)
+	}
+}
+
+// TestResizeImageDataAcceptsRealisticPoster 常见高清海报（2892×4096，约
+// 1180 万像素）必须仍在预算之内，否则真实海报会退化成直出多兆字节原图。
+func TestResizeImageDataAcceptsRealisticPoster(t *testing.T) {
+	data := encodeTestJPEG(t, 2892, 4096)
+
+	out, ctype, unchanged, err := resizeImageData(data, imageResizeOptions{MaxWidth: 480, MaxHeight: 600})
+	if err != nil {
+		t.Fatalf("resizeImageData failed: %v", err)
+	}
+	if unchanged {
+		t.Fatal("expected a resized result, got the original bytes")
+	}
+	if ctype != "image/jpeg" {
+		t.Fatalf("content type = %q, want image/jpeg", ctype)
+	}
+	cfg, _, err := image.DecodeConfig(bytes.NewReader(out))
+	if err != nil {
+		t.Fatalf("decode resized output: %v", err)
+	}
+	if cfg.Width != 424 || cfg.Height != 600 {
+		t.Fatalf("resized to %dx%d, want 424x600", cfg.Width, cfg.Height)
+	}
+}
+
+// TestUseCheapScaleInterpolator 大比例缩小走低成本插值，轻微缩小仍保持高质量。
+func TestUseCheapScaleInterpolator(t *testing.T) {
+	cases := []struct {
+		name                   string
+		dstW, dstH, srcW, srcH int
+		want                   bool
+	}{
+		{"poster to card", 480, 600, 2892, 4096, true},
+		{"hero to strip", 480, 320, 1920, 1080, true},
+		{"exactly half is cheap", 960, 540, 1920, 1080, true},
+		{"slight shrink", 900, 540, 1000, 600, false},
+		{"upscale", 1200, 800, 600, 400, false},
+		{"zero source", 100, 100, 0, 0, false},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			if got := useCheapScaleInterpolator(tc.dstW, tc.dstH, tc.srcW, tc.srcH); got != tc.want {
+				t.Fatalf("useCheapScaleInterpolator(%d,%d,%d,%d) = %v, want %v",
+					tc.dstW, tc.dstH, tc.srcW, tc.srcH, got, tc.want)
+			}
+		})
+	}
+}
+
+// TestIsOpaqueImageDetectsAlpha 透明度检测：全不透明才允许转 JPEG。
+func TestIsOpaqueImageDetectsAlpha(t *testing.T) {
+	opaque := image.NewRGBA(image.Rect(0, 0, 4, 4))
+	for y := 0; y < 4; y++ {
+		for x := 0; x < 4; x++ {
+			opaque.Set(x, y, color.RGBA{R: 10, G: 20, B: 30, A: 255})
+		}
+	}
+	if !isOpaqueImage(opaque) {
+		t.Fatal("expected fully opaque image to be reported opaque")
+	}
+
+	withAlpha := image.NewRGBA(image.Rect(0, 0, 4, 4))
+	for y := 0; y < 4; y++ {
+		for x := 0; x < 4; x++ {
+			withAlpha.Set(x, y, color.RGBA{R: 10, G: 20, B: 30, A: 255})
+		}
+	}
+	withAlpha.Set(2, 2, color.RGBA{R: 10, G: 20, B: 30, A: 128})
+	if isOpaqueImage(withAlpha) {
+		t.Fatal("expected image with a translucent pixel to be reported non-opaque")
+	}
+
+	if isOpaqueImage(nil) {
+		t.Fatal("expected nil image to be reported non-opaque")
+	}
+}

internal/service/image_resize_derive_test.go

@@ -0,0 +1,184 @@
+package service
+
+import (
+	"bytes"
+	"image"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"go.uber.org/zap"
+)
+
+func TestParseResizeVariantSuffixRoundTrip(t *testing.T) {
+	opts := imageResizeOptions{MaxWidth: 480, MaxHeight: 600, Quality: 80}
+	name := "abc123." + opts.resizeVariantSuffix() + ".img"
+
+	width, height, ok := parseResizeVariantSuffix(name, "abc123")
+	if !ok {
+		t.Fatalf("expected %q to parse", name)
+	}
+	if width != 480 || height != 600 {
+		t.Fatalf("parsed %dx%d, want 480x600", width, height)
+	}
+
+	for _, bad := range []string{"abc123.img", "abc124.480x600q80.img", "abc123.480x600.img", "abc123.axbq80.img"} {
+		if _, _, ok := parseResizeVariantSuffix(bad, "abc123"); ok {
+			t.Fatalf("expected %q to be rejected", bad)
+		}
+	}
+}
+
+func TestLargerCachedVariantPicksSmallestCoveringVariant(t *testing.T) {
+	proxy := &ImageProxy{cacheDir: filepath.Join(t.TempDir(), "cache"), log: zap.NewNop()}
+	dir := filepath.Join(proxy.cacheDir, imageResizeCacheSubdir)
+	if err := os.MkdirAll(dir, 0o750); err != nil {
+		t.Fatal(err)
+	}
+
+	key := "deadbeef"
+	writeVariant := func(o imageResizeOptions) string {
+		path := proxy.resizeCachePath(key, o)
+		if err := os.WriteFile(path, []byte("x"), 0o600); err != nil {
+			t.Fatal(err)
+		}
+		return path
+	}
+	// 更小的档位不能用于派生（会放大），更大的档位里要挑最小的那个。
+	writeVariant(imageResizeOptions{MaxWidth: 96, Quality: 60})
+	large := writeVariant(imageResizeOptions{MaxWidth: 1920, MaxHeight: 1080, Quality: 80})
+	card := writeVariant(imageResizeOptions{MaxWidth: 480, MaxHeight: 600, Quality: 80})
+
+	got := proxy.largerCachedVariant(key, imageResizeOptions{MaxWidth: 160, Quality: 60})
+	if got != card {
+		t.Fatalf("largerCachedVariant = %q, want the card variant %q", got, card)
+	}
+
+	// 目标尺寸已与大档位相同（甚至更大）时没有可用来源。
+	if got := proxy.largerCachedVariant(key, imageResizeOptions{MaxWidth: 1920, MaxHeight: 1080, Quality: 80}); got != "" {
+		t.Fatalf("expected no derivable source when the target equals the only covering variant, got %q", got)
+	}
+	if got := proxy.largerCachedVariant(key, imageResizeOptions{MaxWidth: 2400, Quality: 80}); got != "" {
+		t.Fatalf("expected no derivable source for oversize targets, got %q", got)
+	}
+	if large == "" {
+		t.Fatal("large variant should have been written")
+	}
+}
+
+// TestServeResizedDerivesSmallVariantFromCachedCard 小尺寸档位必须能从已缓存
+// 的大尺寸档位派生，而不是每次都重新解码原图。
+//
+// 验证方式：生成卡片档位后，把原图内容替换成同长度、同 mtime 的非图片数据
+// （缓存键不变，但原图已无法解码）。若小图请求仍能返回正确的缩略图，就说明它
+// 是从缓存档位派生的；否则只能退回原图直出（此处会失败，因为原图已不是图片）。
+func TestServeResizedDerivesSmallVariantFromCachedCard(t *testing.T) {
+	dir := t.TempDir()
+	source := filepath.Join(dir, "poster.png")
+	original := encodeTestPNG(t, 1200, 1800, 255)
+	if err := os.WriteFile(source, original, 0o644); err != nil {
+		t.Fatal(err)
+	}
+	// 固定 mtime 并读回平台量化后的值，保证替换内容后缓存键不变。
+	fixed := time.Now().Add(-time.Hour).Truncate(time.Second)
+	if err := os.Chtimes(source, fixed, fixed); err != nil {
+		t.Fatal(err)
+	}
+	srcStat, err := os.Stat(source)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	proxy := &ImageProxy{cacheDir: filepath.Join(dir, "cache"), log: zap.NewNop()}
+
+	card := imageResizeOptions{MaxWidth: 480, MaxHeight: 600, Quality: 80}
+	cardRec := httptest.NewRecorder()
+	if !proxy.serveResizedFromFile(cardRec, httptest.NewRequest("GET", "/x?maxWidth=480", nil), source, card) {
+		t.Fatal("expected the card variant request to be handled")
+	}
+	cardCfg, _, err := image.DecodeConfig(bytes.NewReader(cardRec.Body.Bytes()))
+	if err != nil {
+		t.Fatalf("decode card variant: %v", err)
+	}
+	if cardCfg.Width != 400 || cardCfg.Height != 600 {
+		t.Fatalf("card variant = %dx%d, want 400x600", cardCfg.Width, cardCfg.Height)
+	}
+
+	// 把原图换成同样长度的非图片数据，并恢复完全相同的 mtime，保持缓存键不变。
+	garbage := bytes.Repeat([]byte{0x11}, len(original))
+	if err := os.WriteFile(source, garbage, 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.Chtimes(source, srcStat.ModTime(), srcStat.ModTime()); err != nil {
+		t.Fatal(err)
+	}
+
+	tiny := imageResizeOptions{MaxWidth: 160, Quality: 60}
+	tinyRec := httptest.NewRecorder()
+	if !proxy.serveResizedFromFile(tinyRec, httptest.NewRequest("GET", "/x?maxWidth=160", nil), source, tiny) {
+		t.Fatal("expected the tiny variant request to be handled")
+	}
+	tinyCfg, _, err := image.DecodeConfig(bytes.NewReader(tinyRec.Body.Bytes()))
+	if err != nil {
+		t.Fatalf("decode tiny variant: %v (the request fell back to the placeholder instead of deriving from the cached card)", err)
+	}
+	if tinyCfg.Width != 160 || tinyCfg.Height != 240 {
+		t.Fatalf("tiny variant = %dx%d, want 160x240", tinyCfg.Width, tinyCfg.Height)
+	}
+}
+
+// TestPrefetchCardVariantWarmsResizeCache 刮削预取要顺带生成卡片档位，
+// 首个海报墙请求即可命中缩放缓存。
+func TestPrefetchCardVariantWarmsResizeCache(t *testing.T) {
+	payload := encodeTestPNG(t, 1200, 1800, 255)
+	upstream := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "image/png")
+		_, _ = w.Write(payload)
+	}))
+	defer upstream.Close()
+
+	proxy := newStreamImageProxy(t, upstream.URL)
+	raw := upstream.URL + "/poster.png"
+
+	if err := proxy.PrefetchCardVariant(t.Context(), raw); err != nil {
+		t.Fatalf("PrefetchCardVariant failed: %v", err)
+	}
+
+	_, cachePath, _ := proxy.remoteImageCachePathsForValidated(raw)
+	stat, err := os.Stat(cachePath)
+	if err != nil {
+		t.Fatalf("expected the original to be cached: %v", err)
+	}
+	variantPath := proxy.resizeCachePath(prefetchCardResizeOptions.resizeSourceKey(cachePath, stat), prefetchCardResizeOptions)
+	if _, err := os.Stat(variantPath); err != nil {
+		t.Fatalf("expected the card variant to be pre-generated at %s: %v", variantPath, err)
+	}
+
+	// 预取失败时仍要返回错误（刮削据此保留旧图）。
+	bad := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusBadGateway)
+	}))
+	defer bad.Close()
+	failing := newStreamImageProxy(t, bad.URL)
+	if err := failing.PrefetchCardVariant(t.Context(), bad.URL+"/poster.png"); err == nil {
+		t.Fatal("expected an error for an unreachable candidate image")
+	}
+}
+
+// TestPrefetchCardVariantSurvivesUnresizableImage 上游返回的图无法缩放时，
+// 预取仍视为成功（图片可达即可），不能因此让刮削回退到旧图。
+func TestPrefetchCardVariantSurvivesUnresizableImage(t *testing.T) {
+	upstream := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "image/jpeg")
+		_, _ = w.Write(testJPEG)
+	}))
+	defer upstream.Close()
+
+	proxy := newStreamImageProxy(t, upstream.URL)
+	if err := proxy.PrefetchCardVariant(t.Context(), upstream.URL+"/cover.jpg"); err != nil {
+		t.Fatalf("PrefetchCardVariant failed: %v", err)
+	}
+}

internal/service/manual_scrape_test.go

@@ -199,7 +199,7 @@ func TestManualSearchFallsBackToMovieFolderForGenericQuery(t *testing.T) {
 			http.NotFound(w, r)
 			return
 		}
-		if r.URL.Query().Get("query") != "inception" {
+		if !strings.EqualFold(r.URL.Query().Get("query"), "inception") {
 			_ = json.NewEncoder(w).Encode(map[string]any{"results": []any{}})
 			return
 		}
@@ -248,7 +248,7 @@ func TestManualSearchFallsBackToMovieFolderForGenericQuery(t *testing.T) {
 	if len(results) != 1 || results[0].TMDbID != 27205 {
 		t.Fatalf("manual search results=%#v, want folder fallback candidate; queries=%v", results, queries)
 	}
-	if len(queries) < 2 || queries[0] != "00000" || queries[len(queries)-1] != "inception" {
+	if len(queries) < 2 || queries[0] != "00000" || !strings.EqualFold(queries[len(queries)-1], "inception") {
 		t.Fatalf("manual search queries=%v, want explicit query then folder fallback", queries)
 	}
 }

internal/service/nfo.go

@@ -76,7 +76,10 @@ type episodeNFO struct {
 }
 
 // ExportOne writes a movie.nfo file next to the media file. Existing files
-// are overwritten so a re-scrape always reflects the latest metadata.
+// are overwritten so a re-scrape always reflects the latest metadata. When the
+// folder already holds a sidecar the reader would pick (movie.nfo /
+// <dirname>.nfo), that file is updated in place instead of adding a second,
+// divergent NFO.
 func (s *NFOService) ExportOne(ctx context.Context, mediaID string) (string, error) {
 	m, err := s.repo.Media.FindByID(ctx, mediaID)
 	if err != nil {
@@ -191,12 +194,49 @@ func WriteMediaNFO(m *model.Media) (string, error) {
 		return "", err
 	}
 	dst := nfoPath(resolveMappedDestinationPath(m.Path))
+	if m.SeasonNum <= 0 && m.EpisodeNum <= 0 {
+		dst = nfoExportTarget(dst)
+	}
 	if err := os.WriteFile(dst, []byte(xml.Header+string(out)+"\n"), 0o644); err != nil { // #nosec G306 -- NFO sidecars must remain readable by media players.
 		return "", err
 	}
 	return dst, nil
 }
 
+// nfoExportTarget picks the file an export should write for a movie.
+//
+// Besides the canonical "<base>.nfo", the reader (findMovieNFO) also accepts
+// movie.nfo and "<dirname>.nfo". Writing the canonical name into a folder that
+// already carries one of those left two divergent NFOs for the same movie, and
+// because the reader prefers the canonical one the pre-existing file quietly went
+// stale. Reuse whichever file the reader would already have picked, in the same
+// precedence order, so an export updates the existing sidecar in place.
+//
+// Episodes are deliberately excluded by the caller: for an episode the
+// alternatives are series-level files, and writing episode metadata there would
+// corrupt the series identity.
+func nfoExportTarget(canonical string) string {
+	dir := filepath.Dir(canonical)
+	candidates := []string{
+		filepath.Base(canonical),
+		"movie.nfo",
+		filepath.Base(dir) + ".nfo",
+	}
+	for _, name := range candidates {
+		if name == "" || name == "." {
+			continue
+		}
+		candidate := filepath.Join(dir, name)
+		if candidate == canonical {
+			continue
+		}
+		if info, err := os.Stat(candidate); err == nil && !info.IsDir() {
+			return candidate
+		}
+	}
+	return canonical
+}
+
 func splitNFOList(value string) []string {
 	parts := strings.Split(value, ",")
 	out := make([]string, 0, len(parts))

internal/service/nfo_export_duplicate_test.go

@@ -0,0 +1,145 @@
+package service
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"github.com/truewhile/MeBox/internal/model"
+)
+
+func writeNFOTestFile(t *testing.T, path, content string) {
+	t.Helper()
+	if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(path, []byte(content), 0o644); err != nil {
+		t.Fatal(err)
+	}
+}
+
+// A folder that already carries the Kodi/Emby convention "movie.nfo" must be
+// updated in place. Exporting used to always write "<filename>.nfo", leaving two
+// divergent NFOs for one movie — and since the reader prefers "<filename>.nfo",
+// the pre-existing movie.nfo silently went stale.
+func TestWriteMediaNFOUpdatesExistingMovieNFO(t *testing.T) {
+	dir := t.TempDir()
+	mediaPath := filepath.Join(dir, "The Shawshank Redemption (1994).mp4")
+	writeNFOTestFile(t, mediaPath, "video")
+	existing := filepath.Join(dir, "movie.nfo")
+	writeNFOTestFile(t, existing, `<?xml version="1.0"?><movie><title>旧标题</title></movie>`)
+
+	media := &model.Media{
+		Title:        "肖申克的救赎",
+		Year:         1994,
+		Path:         mediaPath,
+		Overview:     "希望让人自由。",
+		TMDbID:       278,
+		ScrapeStatus: "matched",
+	}
+	dst, err := WriteMediaNFO(media)
+	if err != nil {
+		t.Fatalf("WriteMediaNFO: %v", err)
+	}
+	if filepath.Clean(dst) != filepath.Clean(existing) {
+		t.Fatalf("export wrote %q, want the existing movie.nfo at %q", dst, existing)
+	}
+	if _, err := os.Stat(filepath.Join(dir, "The Shawshank Redemption (1994).nfo")); err == nil {
+		t.Fatal("export created a duplicate <filename>.nfo next to movie.nfo")
+	}
+	body, err := os.ReadFile(existing)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if strings.Contains(string(body), "旧标题") || !strings.Contains(string(body), "肖申克的救赎") {
+		t.Fatalf("movie.nfo was not refreshed: %s", string(body))
+	}
+}
+
+// With no pre-existing sidecar the canonical <filename>.nfo is still the target.
+func TestWriteMediaNFOWritesCanonicalNameByDefault(t *testing.T) {
+	dir := t.TempDir()
+	mediaPath := filepath.Join(dir, "Inception (2010).mkv")
+	writeNFOTestFile(t, mediaPath, "video")
+
+	dst, err := WriteMediaNFO(&model.Media{Title: "盗梦空间", Year: 2010, Path: mediaPath})
+	if err != nil {
+		t.Fatalf("WriteMediaNFO: %v", err)
+	}
+	want := filepath.Join(dir, "Inception (2010).nfo")
+	if filepath.Clean(dst) != filepath.Clean(want) {
+		t.Fatalf("export wrote %q, want %q", dst, want)
+	}
+}
+
+// Re-exporting over an existing canonical sidecar must keep updating that file.
+func TestWriteMediaNFOOverwritesCanonicalSidecar(t *testing.T) {
+	dir := t.TempDir()
+	mediaPath := filepath.Join(dir, "Inception (2010).mkv")
+	writeNFOTestFile(t, mediaPath, "video")
+	canonical := filepath.Join(dir, "Inception (2010).nfo")
+	writeNFOTestFile(t, canonical, `<?xml version="1.0"?><movie><title>old</title></movie>`)
+
+	dst, err := WriteMediaNFO(&model.Media{Title: "盗梦空间", Year: 2010, Path: mediaPath})
+	if err != nil {
+		t.Fatalf("WriteMediaNFO: %v", err)
+	}
+	if filepath.Clean(dst) != filepath.Clean(canonical) {
+		t.Fatalf("export wrote %q, want %q", dst, canonical)
+	}
+	body, _ := os.ReadFile(canonical)
+	if strings.Contains(string(body), ">old<") {
+		t.Fatal("existing canonical sidecar was not refreshed")
+	}
+}
+
+// Episode exports must never be redirected to the series-level movie.nfo /
+// <dirname>.nfo files: those describe the show, not the episode.
+func TestWriteMediaNFOEpisodeIgnoresSeriesLevelSidecars(t *testing.T) {
+	dir := t.TempDir()
+	mediaPath := filepath.Join(dir, "Show - S01E01.mkv")
+	writeNFOTestFile(t, mediaPath, "video")
+	seriesNFO := filepath.Join(dir, "movie.nfo")
+	writeNFOTestFile(t, seriesNFO, `<?xml version="1.0"?><movie><title>series level</title></movie>`)
+
+	dst, err := WriteMediaNFO(&model.Media{
+		Title:      "Show",
+		Path:       mediaPath,
+		SeasonNum:  1,
+		EpisodeNum: 1,
+	})
+	if err != nil {
+		t.Fatalf("WriteMediaNFO: %v", err)
+	}
+	if filepath.Clean(dst) == filepath.Clean(seriesNFO) {
+		t.Fatal("episode export must not overwrite the series-level movie.nfo")
+	}
+	if filepath.Clean(dst) != filepath.Clean(filepath.Join(dir, "Show - S01E01.nfo")) {
+		t.Fatalf("episode export wrote %q, want the episode sidecar", dst)
+	}
+	body, _ := os.ReadFile(seriesNFO)
+	if !strings.Contains(string(body), "series level") {
+		t.Fatal("series-level movie.nfo was modified by an episode export")
+	}
+}
+
+// The <dirname>.nfo convention must be reused too.
+func TestWriteMediaNFOUpdatesDirectoryNamedNFO(t *testing.T) {
+	parent := t.TempDir()
+	dir := filepath.Join(parent, "Interstellar (2014)")
+	mediaPath := filepath.Join(dir, "Interstellar (2014).mkv")
+	writeNFOTestFile(t, mediaPath, "video")
+	dirNFO := filepath.Join(dir, "Interstellar (2014).nfo")
+	// The directory-named sidecar is the only pre-existing candidate besides the
+	// canonical name, so the canonical name is skipped and this one is reused.
+	writeNFOTestFile(t, dirNFO, `<?xml version="1.0"?><movie><title>old</title></movie>`)
+
+	dst, err := WriteMediaNFO(&model.Media{Title: "星际穿越", Year: 2014, Path: mediaPath})
+	if err != nil {
+		t.Fatalf("WriteMediaNFO: %v", err)
+	}
+	if filepath.Clean(dst) != filepath.Clean(dirNFO) {
+		t.Fatalf("export wrote %q, want %q", dst, dirNFO)
+	}
+}

internal/service/organizer_reclassify_metadata_test.go

@@ -6,6 +6,7 @@ import (
 	"net/http/httptest"
 	"os"
 	"path/filepath"
+	"strings"
 	"testing"
 
 	"go.uber.org/zap"
@@ -24,7 +25,10 @@ func TestOrganizeDirectoryCleansReleaseNoiseBeforeMetadataClassify(t *testing.T)
 		}
 		query := r.URL.Query().Get("query")
 		queries = append(queries, query)
-		if query != "motherhood of taihang" {
+		// Real providers match the query case-insensitively; keep this stub
+		// faithful so the test guards release-noise cleaning rather than the
+		// historical lowercased query spelling.
+		if !strings.EqualFold(query, "motherhood of taihang") {
 			_ = json.NewEncoder(w).Encode(map[string]any{"results": []any{}})
 			return
 		}
@@ -81,7 +85,7 @@ func TestOrganizeDirectoryCleansReleaseNoiseBeforeMetadataClassify(t *testing.T)
 	if _, err := os.Stat(want); err != nil {
 		t.Fatalf("organized media missing at %q: %v; items=%#v queries=%v", want, err, res.Items, queries)
 	}
-	if len(queries) == 0 || queries[0] != "motherhood of taihang" {
+	if len(queries) == 0 || !strings.EqualFold(queries[0], "motherhood of taihang") {
 		t.Fatalf("first metadata query = %q, want cleaned title; all queries=%v", firstQuery(queries), queries)
 	}
 }

internal/service/organizer_scrape_providers_test.go

@@ -74,7 +74,9 @@ func TestOrganizeDirectoryUsesAdultMetadataBeforeRename(t *testing.T) {
 
 func TestOrganizeDirectoryUsesBangumiForAnimeRename(t *testing.T) {
 	upstream := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-		if r.URL.Path != "/search/subject/frieren" {
+		// Bangumi takes the keyword in the path; compare case-insensitively so the
+		// stub does not pin the query spelling (titles now keep their original case).
+		if !strings.EqualFold(r.URL.Path, "/search/subject/frieren") {
 			http.NotFound(w, r)
 			return
 		}

internal/service/playback_quality.go

@@ -211,13 +211,25 @@ func DefaultCloud115Quality(options []PlaybackQuality) string {
 }
 
 // LocalQualityOptions 返回 MeBox 本地 HLS 可用的画质档位。
-func LocalQualityOptions(m *model.Media) []PlaybackQuality {
+//
+// transcodeAvailable 为 false（未装 ffmpeg、或转码被全局关闭）时档位仍然返回，
+// 但 Available 为 false：播放器据此直接走直连播放，而不是先请求 /api/hls 拿到
+// 一个必然失败的 500 再回退。档位列表保留是为了让前端仍能展示"需要转码"的说明。
+func LocalQualityOptions(m *model.Media, transcodeAvailable bool) []PlaybackQuality {
 	sourceHeight := mediaSourceHeight(m)
 	presets := []PlaybackQuality{
-		{ID: "source", Label: "原画", Height: sourceHeight, Source: "local", Available: true, Note: "本地 HLS，保持源分辨率"},
-		{ID: "1080", Label: "1080P", Height: 1080, Source: "local", Available: true},
-		{ID: "720", Label: "720P", Height: 720, Source: "local", Available: true},
-		{ID: "480", Label: "480P", Height: 480, Source: "local", Available: true},
+		{ID: "source", Label: "原画", Height: sourceHeight, Source: "local", Available: transcodeAvailable, Note: "本地 HLS，保持源分辨率"},
+		{ID: "1080", Label: "1080P", Height: 1080, Source: "local", Available: transcodeAvailable},
+		{ID: "720", Label: "720P", Height: 720, Source: "local", Available: transcodeAvailable},
+		{ID: "480", Label: "480P", Height: 480, Source: "local", Available: transcodeAvailable},
+	}
+	if !transcodeAvailable {
+		for i := range presets {
+			presets[i].RequiresTranscode = true
+			if presets[i].Note == "" {
+				presets[i].Note = "需要 ffmpeg，当前不可用"
+			}
+		}
 	}
 	out := make([]PlaybackQuality, 0, len(presets))
 	for _, preset := range presets {

internal/service/playback_quality_test.go

@@ -61,20 +61,45 @@ func TestCloud115QualityOptionsListsAvailableOnlyOnce(t *testing.T) {
 
 func TestLocalQualityOptionsFollowSourceHeight(t *testing.T) {
 	media := &model.Media{Height: 720}
-	options := LocalQualityOptions(media)
+	options := LocalQualityOptions(media, true)
 	if _, ok := findPlaybackQuality(options, "1080"); ok {
 		t.Fatal("720p source should not list 1080P")
 	}
 	for _, id := range []string{"source", "720", "480"} {
-		if _, ok := findPlaybackQuality(options, id); !ok {
+		quality, ok := findPlaybackQuality(options, id)
+		if !ok {
 			t.Fatalf("local quality %s missing", id)
 		}
+		if !quality.Available {
+			t.Fatalf("local quality %s should be available while transcoding works", id)
+		}
 	}
 	if got := DefaultLocalHLSQualityID(media); got != "720" {
 		t.Fatalf("default local quality = %q, want 720", got)
 	}
 }
 
+// Without a usable ffmpeg the local HLS renditions must not advertise themselves
+// as available: the player would request /api/hls/... and take a guaranteed 500
+// before falling back to direct play.
+func TestLocalQualityOptionsMarkUnavailableWithoutTranscoder(t *testing.T) {
+	media := &model.Media{Height: 1080}
+	for _, id := range []string{"source", "1080", "720", "480"} {
+		for _, available := range []bool{true, false} {
+			quality, ok := findPlaybackQuality(LocalQualityOptions(media, available), id)
+			if !ok {
+				t.Fatalf("local quality %s missing", id)
+			}
+			if quality.Available != available {
+				t.Fatalf("quality %s availability = %v, want %v", id, quality.Available, available)
+			}
+			if !available && !quality.RequiresTranscode {
+				t.Fatalf("unavailable quality %s should be flagged as requiring transcoding", id)
+			}
+		}
+	}
+}
+
 func TestLocalHLSQualityFFmpegArgs(t *testing.T) {
 	cfg := &config.Config{}
 	cfg.Transcoder.MaxHeight = 720

internal/service/recognition_words_test.go

@@ -20,8 +20,8 @@ One\.Piece\.S01E(89[2-9]|9\d{2}|10\d{2})\.1999 => 海贼王.S21E\1.1999 && S21E
 
 func TestCleanQueryWithRecognitionDisabledByDefaultRepoNil(t *testing.T) {
 	title, year := CleanQueryWithRecognition(t.Context(), nil, "Dune.2021.2160p.WEB-DL.mkv")
-	if title != "dune" || year != 2021 {
-		t.Fatalf("CleanQueryWithRecognition = %q/%d, want dune/2021", title, year)
+	if title != "Dune" || year != 2021 {
+		t.Fatalf("CleanQueryWithRecognition = %q/%d, want Dune/2021", title, year)
 	}
 }
 

internal/service/runtime_cache.go

@@ -36,6 +36,16 @@ type RuntimeCacheService struct {
 	limit     int
 	maxBytes  int64
 	bytesUsed int64
+	// seq stamps every insert/access with a strictly increasing number. Two
+	// entries written or touched inside the same clock tick carry an identical
+	// lastUsed, and Go map iteration order is random, so eviction used to pick an
+	// arbitrary "oldest" entry. seq turns recency into a total order.
+	seq uint64
+}
+
+func (c *RuntimeCacheService) nextSeqLocked() uint64 {
+	c.seq++
+	return c.seq
 }
 
 type runtimeCacheItem struct {
@@ -43,6 +53,7 @@ type runtimeCacheItem struct {
 	expiresAt  time.Time
 	staleUntil time.Time
 	lastUsed   time.Time
+	seq        uint64
 	size       int64
 }
 
@@ -55,6 +66,7 @@ type runtimeObjectItem struct {
 	value     any
 	expiresAt time.Time
 	lastUsed  time.Time
+	seq       uint64
 	size      int64
 }
 
@@ -225,6 +237,7 @@ func (c *RuntimeCacheService) GetObject(key string) (any, bool) {
 		return nil, false
 	}
 	item.lastUsed = now
+	item.seq = c.nextSeqLocked()
 	c.obj[fullKey] = item
 	c.mu.Unlock()
 	return item.value, true
@@ -252,6 +265,7 @@ func (c *RuntimeCacheService) SetObject(key string, value any, ttl time.Duration
 		value:     value,
 		expiresAt: now.Add(ttl),
 		lastUsed:  now,
+		seq:       c.nextSeqLocked(),
 		size:      size,
 	}
 	c.bytesUsed += size
@@ -334,6 +348,7 @@ func (c *RuntimeCacheService) getMemoryWithStale(key string) ([]byte, bool, bool
 		return nil, false, false
 	}
 	item.lastUsed = now
+	item.seq = c.nextSeqLocked()
 	c.memory[key] = item
 	return item.raw, true, !now.Before(item.expiresAt)
 }
@@ -376,6 +391,7 @@ func (c *RuntimeCacheService) setMemoryBytesWithStale(key string, raw []byte, fr
 		expiresAt:  now.Add(freshTTL),
 		staleUntil: now.Add(staleTTL),
 		lastUsed:   now,
+		seq:        c.nextSeqLocked(),
 		size:       size,
 	}
 	c.bytesUsed += size
@@ -424,20 +440,36 @@ func (c *RuntimeCacheService) evictExpiredLocked(now time.Time) {
 	}
 }
 
+// evictOldestLocked removes the least recently used entry across both caches.
+//
+// Recency is ordered by (lastUsed, seq): entries written within the same clock
+// tick share lastUsed, and map iteration order is randomized, so comparing
+// timestamps alone let eviction drop a just-written entry instead of the older
+// one. seq breaks those ties deterministically.
 func (c *RuntimeCacheService) evictOldestLocked() bool {
 	var (
 		oldestKey  string
 		oldestKind byte
 		oldestAt   time.Time
+		oldestSeq  uint64
 	)
+	older := func(at time.Time, seq uint64) bool {
+		if oldestKind == 0 {
+			return true
+		}
+		if at.Equal(oldestAt) {
+			return seq < oldestSeq
+		}
+		return at.Before(oldestAt)
+	}
 	for key, item := range c.memory {
-		if oldestKind == 0 || item.lastUsed.Before(oldestAt) {
-			oldestKey, oldestKind, oldestAt = key, 'm', item.lastUsed
+		if older(item.lastUsed, item.seq) {
+			oldestKey, oldestKind, oldestAt, oldestSeq = key, 'm', item.lastUsed, item.seq
 		}
 	}
 	for key, item := range c.obj {
-		if oldestKind == 0 || item.lastUsed.Before(oldestAt) {
-			oldestKey, oldestKind, oldestAt = key, 'o', item.lastUsed
+		if older(item.lastUsed, item.seq) {
+			oldestKey, oldestKind, oldestAt, oldestSeq = key, 'o', item.lastUsed, item.seq
 		}
 	}
 	switch oldestKind {

internal/service/runtime_cache_eviction_test.go

@@ -0,0 +1,108 @@
+package service
+
+import (
+	"context"
+	"strings"
+	"testing"
+	"time"
+)
+
+// Eviction must be deterministic when entries share a lastUsed timestamp.
+//
+// Regression: recency was compared by lastUsed alone. Entries written within the
+// same clock tick carry identical timestamps and Go map iteration order is
+// randomized, so eviction could drop a just-written entry instead of the oldest
+// one — which made TestRuntimeCacheSetMaxSizeEvictsImmediately fail intermittently
+// in a full-package run.
+func TestRuntimeCacheEvictionBreaksTimestampTiesDeterministically(t *testing.T) {
+	// Repeat because the bug only showed up for some map iteration orders.
+	for attempt := 0; attempt < 50; attempt++ {
+		cache := newRuntimeCacheForTest(t, 128)
+		const perEntry = 700 << 10
+
+		// Two entries large enough that a 1MiB budget can hold only one. Writing
+		// them back to back puts them in the same clock tick often enough to hit
+		// the tie almost immediately.
+		cache.SetObject("old", strings.Repeat("a", perEntry), time.Minute)
+		cache.SetObject("new", strings.Repeat("b", perEntry), time.Minute)
+
+		cache.SetMaxSizeMB(1)
+
+		if _, ok := cache.GetObject("old"); ok {
+			t.Fatalf("attempt %d: oldest entry survived the budget drop", attempt)
+		}
+		if _, ok := cache.GetObject("new"); !ok {
+			t.Fatalf("attempt %d: newest entry was evicted instead of the oldest", attempt)
+		}
+	}
+}
+
+// Accessing an entry refreshes its recency even when it lands in the same tick as
+// the write of a competing entry.
+func TestRuntimeCacheAccessRefreshesEvictionOrder(t *testing.T) {
+	cache := newRuntimeCacheForTest(t, 128)
+	const perEntry = 700 << 10
+
+	cache.SetObject("first", strings.Repeat("a", perEntry), time.Minute)
+	cache.SetObject("second", strings.Repeat("b", perEntry), time.Minute)
+	// Touch the older entry so it becomes the more recent one.
+	if _, ok := cache.GetObject("first"); !ok {
+		t.Fatal("first entry should be present before the budget drop")
+	}
+
+	cache.SetMaxSizeMB(1)
+
+	if _, ok := cache.GetObject("second"); ok {
+		t.Fatal("the untouched entry should have been evicted first")
+	}
+	if _, ok := cache.GetObject("first"); !ok {
+		t.Fatal("the just-accessed entry should have been kept")
+	}
+}
+
+// Eviction picks the oldest across the byte cache and the object cache together.
+func TestRuntimeCacheEvictionComparesBothCaches(t *testing.T) {
+	cache := newRuntimeCacheForTest(t, 128)
+	const perEntry = 700 << 10
+
+	// Byte-cache entry written first, object-cache entry second.
+	payload := strings.Repeat("c", perEntry)
+	cache.SetJSON(context.Background(), "bytes", payload, time.Minute)
+	cache.SetObject("object", strings.Repeat("d", perEntry), time.Minute)
+
+	cache.SetMaxSizeMB(1)
+
+	if _, ok := cache.GetObject("object"); !ok {
+		t.Fatal("the newer object entry must be kept")
+	}
+	var decoded string
+	if cache.GetJSON(context.Background(), "bytes", &decoded) {
+		t.Fatal("the older byte-cache entry should have been evicted")
+	}
+}
+
+// The sequence counter must hand out strictly increasing values, otherwise ties
+// would resolve randomly again.
+func TestRuntimeCacheSequenceIsMonotonic(t *testing.T) {
+	cache := newRuntimeCacheForTest(t, 16)
+	seen := map[uint64]bool{}
+	cache.mu.Lock()
+	var previous uint64
+	for i := 0; i < 100; i++ {
+		seq := cache.nextSeqLocked()
+		if seen[seq] {
+			cache.mu.Unlock()
+			t.Fatalf("sequence %d handed out twice", seq)
+		}
+		if seq <= previous && previous != 0 {
+			cache.mu.Unlock()
+			t.Fatalf("sequence went backwards: %d after %d", seq, previous)
+		}
+		seen[seq] = true
+		previous = seq
+	}
+	cache.mu.Unlock()
+	if len(seen) != 100 {
+		t.Fatalf("expected 100 distinct sequence numbers, got %d", len(seen))
+	}
+}

internal/service/runtime_settings.go

@@ -119,6 +119,20 @@ func ApplyRuntimeSetting(cfg *config.Config, key, value string) {
 			}
 			cfg.Cache.ImagesMaxSizeMB = n
 		}
+	case "cache.images_originals_max_size_mb":
+		if n, err := strconv.Atoi(value); err == nil {
+			if n < 0 {
+				n = 0
+			}
+			cfg.Cache.ImagesOriginalsMaxSizeMB = n
+		}
+	case "cache.images_originals_ttl_hours":
+		if n, err := strconv.Atoi(value); err == nil {
+			if n < 0 {
+				n = 0
+			}
+			cfg.Cache.ImagesOriginalsTTLHours = n
+		}
 	case "cache.memory_max_size_mb":
 		n := config.DefaultCacheMemoryMaxSizeMB
 		if value != "" {

internal/service/sanitize_dotdir_test.go

@@ -0,0 +1,79 @@
+package service
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+
+	"go.uber.org/zap"
+)
+
+// Dot-prefixed directory names are legal and common (hidden media folders,
+// ".staging" drop dirs). cleanEntryName used to trim dots from BOTH ends, which
+// silently rewrote `.media` into `media`.
+func TestCleanEntryNameKeepsLeadingDots(t *testing.T) {
+	cases := []struct {
+		name  string
+		isDir bool
+		want  string
+	}{
+		{name: ".media", isDir: true, want: ".media"},
+		{name: ".tmp-live-media", isDir: true, want: ".tmp-live-media"},
+		{name: ".staging", isDir: true, want: ".staging"},
+		{name: "Season 01.", isDir: true, want: "Season 01"},
+		{name: "  trailing space  ", isDir: true, want: "trailing space"},
+		{name: ".hidden.mkv", isDir: false, want: ".hidden.mkv"},
+		{name: "poster.jpg", isDir: false, want: "poster.jpg"},
+		{name: "file:name?test*<foo>|bar\".mkv", isDir: false, want: "file nametestfoobar.mkv"},
+		// Names that are nothing but dots must still collapse to a safe placeholder.
+		{name: ".", isDir: true, want: "unnamed"},
+		{name: "..", isDir: true, want: "unnamed"},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			if got := cleanEntryName(tc.name, tc.isDir); got != tc.want {
+				t.Fatalf("cleanEntryName(%q, %v) = %q, want %q", tc.name, tc.isDir, got, tc.want)
+			}
+		})
+	}
+}
+
+// sanitizeLocalPath must preserve a dot-prefixed directory in the middle of an
+// absolute path. Regression: artwork sidecars for a library at
+// `<root>/.tmp-live-media/电影/...` were written to `<root>/tmp-live-media/...`
+// instead — a sibling tree outside the media folder.
+func TestSanitizeLocalPathKeepsDotPrefixedSegments(t *testing.T) {
+	dir := filepath.Join(string(filepath.Separator)+"media", ".hidden-lib", "电影", "Inception (2010)")
+	got := sanitizeLocalPath(dir)
+	if got != dir {
+		t.Fatalf("sanitizeLocalPath(%q) = %q, want the path unchanged", dir, got)
+	}
+}
+
+// End-to-end guard for the observed bug: writing scraped artwork into a media
+// folder located under a dot-prefixed library root must land inside that folder,
+// never in a dot-stripped sibling directory.
+func TestArtworkWriteKeepsDotPrefixedLibraryPath(t *testing.T) {
+	root := t.TempDir()
+	libraryRoot := filepath.Join(root, ".media")
+	mediaDir := filepath.Join(libraryRoot, "电影", "Inception (2010)")
+	if err := os.MkdirAll(mediaDir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+
+	svc := &ScraperService{log: zap.NewNop()}
+	dst := svc.writeArtworkDataToPath(mediaDir, "Inception (2010)-poster", "image/jpeg", []byte{0xff, 0xd8, 0xff, 0xdb, 0x00})
+	if dst == "" {
+		t.Fatal("artwork write reported no destination")
+	}
+	if filepath.Dir(filepath.Clean(dst)) != filepath.Clean(mediaDir) {
+		t.Fatalf("artwork written to %q, want a file inside %q", dst, mediaDir)
+	}
+	if _, err := os.Stat(filepath.Join(mediaDir, "Inception (2010)-poster.jpg")); err != nil {
+		t.Fatalf("sidecar missing from the media folder: %v", err)
+	}
+	// A dot-stripped sibling must not be created.
+	if _, err := os.Stat(filepath.Join(root, "media")); err == nil {
+		t.Fatal("dot-stripped sibling directory was created")
+	}
+}

internal/service/scanner_incremental_test.go

@@ -114,11 +114,11 @@ func TestScanLibraryUsesISOParentFolderForScrapeIdentity(t *testing.T) {
 	if err := repos.DB.First(&media).Error; err != nil {
 		t.Fatal(err)
 	}
-	if media.Title != "dune part two" || media.Year != 2024 || media.ScrapeStatus != "pending" {
+	if media.Title != "Dune Part Two" || media.Year != 2024 || media.ScrapeStatus != "pending" {
 		t.Fatalf("ISO scrape identity = title=%q year=%d status=%q", media.Title, media.Year, media.ScrapeStatus)
 	}
 	candidates := scrapeQueryCandidates(&media, &lib)
-	if len(candidates) == 0 || candidates[0] != "dune part two" {
+	if len(candidates) == 0 || candidates[0] != "Dune Part Two" {
 		t.Fatalf("ISO scrape candidates = %#v", candidates)
 	}
 }
@@ -161,7 +161,7 @@ func TestScanLibraryRepairsPreviouslyUnmatchedGenericISO(t *testing.T) {
 	if err := repos.DB.First(&media, "id = ?", legacy.ID).Error; err != nil {
 		t.Fatal(err)
 	}
-	if media.Title != "dune part two" || media.Year != 2024 || media.ScrapeStatus != "pending" {
+	if media.Title != "Dune Part Two" || media.Year != 2024 || media.ScrapeStatus != "pending" {
 		t.Fatalf("repaired ISO = title=%q year=%d status=%q", media.Title, media.Year, media.ScrapeStatus)
 	}
 }

internal/service/scanner_local_ingest.go

@@ -122,11 +122,52 @@ func (s *ScannerService) localMediaScanState(in localMediaScanStateInput) (bool,
 	existing, exists := in.existingMedia[in.cleanPath]
 	isNewMedia := !exists
 	if exists && in.ext != ".strm" && existing.SizeBytes == in.size && !localMetadataNeedsRefresh(existing, in.localMeta) && !localDerivedMetadataNeedsRefresh(existing, in.incoming) {
+		// Unchanged rows are skipped, but a row that never received ffprobe data
+		// still needs a backfill probe so installing ffmpeg later repairs the library.
+		s.queueProbeBackfillIfNeeded(in.path, in.ext, existing)
 		return isNewMedia, true
 	}
 	return isNewMedia, false
 }
 
+// queueProbeBackfillIfNeeded re-queues ffprobe for an unchanged media row that
+// still carries no technical metadata, reporting whether a probe was queued.
+//
+// The incremental skip above only compares file size and metadata, so a library
+// scanned while ffprobe was unavailable kept duration/resolution/codec empty
+// forever: installing ffmpeg and rescanning did nothing, and the only way out was
+// a per-item manual "探测媒体轨". Only the probe is re-queued here — the row
+// itself is not rewritten, so scraped metadata and titles stay untouched.
+func (s *ScannerService) queueProbeBackfillIfNeeded(path, ext string, existing existingLocalMedia) bool {
+	if !mediaExtensionSupportsProbe(ext) || !localMediaProbeDataMissing(existing) {
+		return false
+	}
+	// Without a resolvable ffprobe/ffmpeg every queued probe would just fail, so
+	// only queue when probing is actually possible right now.
+	if s == nil || s.probe == nil || !s.probe.Available() {
+		return false
+	}
+	if !s.queueLocalMediaProbe(path) {
+		return false
+	}
+	if s.log != nil {
+		s.log.Debug("rescan queued ffprobe backfill for media without track metadata", zap.String("path", path))
+	}
+	return true
+}
+
+// localMediaProbeDataMissing reports whether a media row has none of the fields
+// ffprobe fills in. Requiring every field to be empty keeps genuinely probed
+// media (which always yields at least a codec or a duration) out of the backfill
+// path.
+func localMediaProbeDataMissing(existing existingLocalMedia) bool {
+	return existing.DurationSec <= 0 &&
+		existing.Width <= 0 &&
+		existing.Height <= 0 &&
+		strings.TrimSpace(existing.VideoCodec) == "" &&
+		strings.TrimSpace(existing.AudioCodec) == ""
+}
+
 type localScanMediaInput struct {
 	lib           *model.Library
 	root          *model.LibraryRoot

internal/service/scanner_local_metadata_test.go

@@ -263,7 +263,7 @@ func TestScanLibraryRefreshesStaleNoMatchDerivedMetadata(t *testing.T) {
 	if err := db.First(&media, "path = ?", mediaPath).Error; err != nil {
 		t.Fatal(err)
 	}
-	if media.Title != "hntv spring festival gala" || media.SeasonNum != 1 || media.EpisodeNum != 202 || media.ScrapeStatus != "pending" {
+	if media.Title != "Hntv Spring Festival Gala" || media.SeasonNum != 1 || media.EpisodeNum != 202 || media.ScrapeStatus != "pending" {
 		t.Fatalf("stale no_match row was not refreshed: title=%q s=%d e=%d status=%q", media.Title, media.SeasonNum, media.EpisodeNum, media.ScrapeStatus)
 	}
 }

internal/service/scanner_probe_backfill_test.go

@@ -0,0 +1,167 @@
+package service
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+
+	"go.uber.org/zap"
+
+	"github.com/truewhile/MeBox/internal/config"
+	"github.com/truewhile/MeBox/internal/model"
+	"github.com/truewhile/MeBox/internal/repository"
+)
+
+// newProbeBackfillScanner builds a scanner over the given repository whose
+// ffprobe "binary" is a temp file, so probe availability is deterministic
+// regardless of what happens to be installed on the test machine.
+func newProbeBackfillScanner(t *testing.T, repos *repository.Container, probeAvailable bool) *ScannerService {
+	t.Helper()
+	cfg := &config.Config{}
+	probePath := filepath.Join(t.TempDir(), "not-a-real-ffprobe")
+	if probeAvailable {
+		if err := os.WriteFile(probePath, []byte("stub"), 0o755); err != nil {
+			t.Fatalf("write stub probe: %v", err)
+		}
+	}
+	// Point the probe and its ffmpeg fallback at paths that only exist when this
+	// test wants them to, so Available() cannot pick up a system binary.
+	cfg.App.FFprobePath = probePath
+	cfg.App.FFmpegPath = probePath
+	return NewScannerService(cfg, zap.NewNop(), repos, NewHub(zap.NewNop()), NewFFprobeService(cfg, zap.NewNop()), nil)
+}
+
+func newProbeBackfillRepo(t *testing.T) *repository.Container {
+	t.Helper()
+	return repository.New(newServiceTestDB(t, &model.Library{}, &model.Media{}, &model.Setting{}))
+}
+
+func probeBackfillScanner(t *testing.T, probeAvailable bool) *ScannerService {
+	t.Helper()
+	return newProbeBackfillScanner(t, newProbeBackfillRepo(t), probeAvailable)
+}
+
+// localMediaProbeDataMissing decides which unchanged rows are worth re-probing.
+func TestLocalMediaProbeDataMissing(t *testing.T) {
+	cases := []struct {
+		name     string
+		existing existingLocalMedia
+		want     bool
+	}{
+		{name: "never probed", existing: existingLocalMedia{}, want: true},
+		{name: "duration only", existing: existingLocalMedia{DurationSec: 148}, want: false},
+		{name: "resolution only", existing: existingLocalMedia{Width: 1920, Height: 1080}, want: false},
+		{name: "video codec only", existing: existingLocalMedia{VideoCodec: "h264"}, want: false},
+		{name: "audio codec only", existing: existingLocalMedia{AudioCodec: "aac"}, want: false},
+		{name: "fully probed", existing: existingLocalMedia{DurationSec: 148, Width: 1920, Height: 1080, VideoCodec: "h264", AudioCodec: "aac"}, want: false},
+		{name: "blank codecs are not data", existing: existingLocalMedia{VideoCodec: "  ", AudioCodec: " "}, want: true},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			if got := localMediaProbeDataMissing(tc.existing); got != tc.want {
+				t.Fatalf("localMediaProbeDataMissing(%+v) = %v, want %v", tc.existing, got, tc.want)
+			}
+		})
+	}
+}
+
+// A rescan must re-queue ffprobe for media that still has no technical metadata,
+// otherwise a library scanned before ffmpeg was installed stays empty forever.
+func TestRescanQueuesProbeBackfillForUnprobedMedia(t *testing.T) {
+	sc := probeBackfillScanner(t, true)
+
+	if !sc.queueProbeBackfillIfNeeded("/media/movie.mkv", ".mkv", existingLocalMedia{}) {
+		t.Fatal("unprobed media should get a backfill probe queued")
+	}
+	if sc.queueProbeBackfillIfNeeded("/media/movie.mkv", ".mkv", existingLocalMedia{DurationSec: 148, VideoCodec: "h264"}) {
+		t.Fatal("already probed media must not be re-queued")
+	}
+}
+
+// Backfill is pointless without a usable binary: every probe would fail, so the
+// scanner must not flood the probe queue.
+func TestRescanSkipsProbeBackfillWhenProbeUnavailable(t *testing.T) {
+	sc := probeBackfillScanner(t, false)
+
+	if sc.probe.Available() {
+		t.Skip("a system ffprobe/ffmpeg is resolvable in this environment; availability gating cannot be exercised")
+	}
+	if sc.queueProbeBackfillIfNeeded("/media/movie.mkv", ".mkv", existingLocalMedia{}) {
+		t.Fatal("no probe should be queued while ffprobe/ffmpeg is unavailable")
+	}
+}
+
+// Formats that mediaExtensionSupportsProbe excludes must never enter the backfill
+// path, even when the row looks unprobed.
+func TestRescanSkipsProbeBackfillForUnprobeableExtensions(t *testing.T) {
+	sc := probeBackfillScanner(t, true)
+
+	for _, ext := range []string{".strm", ".iso"} {
+		if sc.queueProbeBackfillIfNeeded("/media/item"+ext, ext, existingLocalMedia{}) {
+			t.Fatalf("ext %s must not be queued for a probe backfill", ext)
+		}
+	}
+}
+
+// End-to-end guard: after a scan leaves a row without technical metadata, the
+// incremental-skip path for an unchanged file must still schedule a backfill
+// probe rather than skipping the file forever.
+func TestScanLibraryQueuesBackfillOnUnchangedRescan(t *testing.T) {
+	repos := newProbeBackfillRepo(t)
+	root := t.TempDir()
+	lib := model.Library{Name: "Movies", Path: root, Type: "movie", Enabled: true}
+	if err := repos.Library.Create(t.Context(), &lib); err != nil {
+		t.Fatal(err)
+	}
+	file := filepath.Join(root, "Some Movie (2021).mkv")
+	if err := os.WriteFile(file, []byte("data"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	first := newProbeBackfillScanner(t, repos, true)
+	res, err := first.ScanLibrary(t.Context(), lib.ID)
+	if err != nil {
+		t.Fatalf("first scan: %v", err)
+	}
+	if res.Added != 1 {
+		t.Fatalf("first scan added=%d, want 1", res.Added)
+	}
+
+	var existing existingLocalMedia
+	if err := repos.DB.Model(&model.Media{}).
+		Select("duration_sec", "width", "height", "video_codec", "audio_codec", "size_bytes").
+		Where("path = ?", file).Scan(&existing).Error; err != nil {
+		t.Fatal(err)
+	}
+	if !localMediaProbeDataMissing(existing) {
+		t.Fatalf("fixture precondition failed, media already has probe data: %+v", existing)
+	}
+
+	// A fresh scanner instance has no probe in flight for this path, which is what
+	// a later rescan (e.g. after the operator installs ffmpeg) looks like.
+	second := newProbeBackfillScanner(t, repos, true)
+	_, skipUnchanged := second.localMediaScanState(localMediaScanStateInput{
+		ctx:           t.Context(),
+		path:          file,
+		cleanPath:     filepath.Clean(file),
+		ext:           ".mkv",
+		size:          existing.SizeBytes,
+		existingMedia: map[string]existingLocalMedia{filepath.Clean(file): existing},
+	})
+	if !skipUnchanged {
+		t.Fatal("unchanged file should take the incremental-skip path")
+	}
+
+	// Control: a different unprobed path on the same instance still queues, so the
+	// instance is demonstrably able to schedule probes.
+	other := filepath.Join(root, "Another Movie (2022).mkv")
+	if !second.queueProbeBackfillIfNeeded(other, ".mkv", existing) {
+		t.Fatal("control path should queue a backfill probe")
+	}
+	// Therefore the incremental-skip path above must already have queued this path:
+	// queueing reserves the path, so a second attempt is refused. If the skip path
+	// had not queued it, this call would have succeeded.
+	if second.queueProbeBackfillIfNeeded(file, ".mkv", existing) {
+		t.Fatal("unchanged unprobed media was not queued by the incremental-skip path")
+	}
+}

internal/service/scheduler.go

@@ -39,7 +39,7 @@ type SchedulerService struct {
 	cacheDir         string
 	now              func() time.Time
 
-	imagesMaxSizeMBProvider func() int
+	imagesPolicyProvider func() ImageCachePolicy
 
 	mu     sync.Mutex
 	stopCh chan struct{}
@@ -59,15 +59,22 @@ func (s *SchedulerService) SetOrganizePipeline(pipeline *OrganizePipelineService
 	s.organizePipeline = pipeline
 }
 
-func (s *SchedulerService) SetImagesMaxSizeMBProvider(fn func() int) {
-	s.imagesMaxSizeMBProvider = fn
+// ImageCachePolicy 是一次图片缓存清理要用的策略（全部为 0 表示不做任何清理）。
+type ImageCachePolicy struct {
+	TotalBytes     int64
+	OriginalsBytes int64
+	OriginalsAge   time.Duration
 }
 
-func (s *SchedulerService) imagesMaxSizeMB() int {
-	if s.imagesMaxSizeMBProvider != nil {
-		return s.imagesMaxSizeMBProvider()
+func (s *SchedulerService) SetImageCachePolicyProvider(fn func() ImageCachePolicy) {
+	s.imagesPolicyProvider = fn
+}
+
+func (s *SchedulerService) imageCachePolicy() ImageCachePolicy {
+	if s.imagesPolicyProvider != nil {
+		return s.imagesPolicyProvider()
 	}
-	return 0
+	return ImageCachePolicy{}
 }
 
 // scheduledJob is one recurring task.

internal/service/scheduler_local_jobs.go

@@ -178,18 +178,19 @@ func (s *SchedulerService) jobCleanTranscodeCache(ctx context.Context) error {
 	return walkAndPrune(s.cacheDir+"/hls", cutoff)
 }
 
-// jobCleanImageCache prunes image proxy cache files when disk usage exceeds the configured limit.
+// jobCleanImageCache prunes the image proxy cache: 原图按保留时长与独立配额
+// 优先淘汰，总量超限时再淘汰派生成品。
 func (s *SchedulerService) jobCleanImageCache(ctx context.Context) error {
 	if s.cacheDir == "" {
 		return nil
 	}
-	maxMB := s.imagesMaxSizeMB()
-	if maxMB <= 0 {
+	policy := s.imageCachePolicy()
+	if policy.TotalBytes <= 0 && policy.OriginalsBytes <= 0 && policy.OriginalsAge <= 0 {
 		return nil
 	}
 	imagesDir := filepath.Join(s.cacheDir, "images")
-	maxSizeBytes := int64(maxMB) * 1024 * 1024
-	res, err := PruneImageCache(imagesDir, maxSizeBytes)
+	pools := ImageCachePools(imagesDir, policy.OriginalsBytes, policy.OriginalsAge)
+	res, err := PruneImageCachePools(pools, policy.TotalBytes)
 	if err != nil {
 		if s.log != nil {
 			s.log.Warn("scheduled image cache cleanup failed", zap.Error(err))

internal/service/scraper_artwork.go

@@ -29,7 +29,7 @@ func (s *ScraperService) prepareScrapedArtworkURL(ctx context.Context, mediaID,
 		timeout = 8 * time.Second
 	}
 	fetchCtx, cancel := context.WithTimeout(context.WithoutCancel(ctx), timeout)
-	err := s.images.PrefetchRemote(fetchCtx, candidate)
+	err := s.images.PrefetchCardVariant(fetchCtx, candidate)
 	cancel()
 	if err != nil {
 		if originalSource != "" {

internal/service/scraper_query_case_test.go

@@ -0,0 +1,68 @@
+package service
+
+import (
+	"strings"
+	"testing"
+)
+
+// TestCleanQueryPreservesOriginalTitleCase is the regression guard for English
+// titles being lowercased on ingest: "The.Matrix.1999.1080p.BluRay.x264-AMIABLE"
+// used to be stored as "the matrix", which also made metadata providers miss it.
+func TestCleanQueryPreservesOriginalTitleCase(t *testing.T) {
+	cases := []struct {
+		in        string
+		wantTitle string
+		wantYear  int
+	}{
+		{"The.Matrix.1999.1080p.BluRay.x264-AMIABLE.mkv", "The Matrix", 1999},
+		{"The.Shawshank.Redemption.1994.1080p.BluRay.mkv", "The Shawshank Redemption", 1994},
+		{"Interstellar.2014.2160p.WEB-DL.mkv", "Interstellar", 2014},
+		{"Breaking.Bad.S01E01.1080p.WEB-DL.mkv", "Breaking Bad", 0},
+		{"Fast.and.Furious.2001.mkv", "Fast and Furious", 2001},
+		{"spider-man.2002.1080p.mkv", "spider man", 2002},
+	}
+	for _, tc := range cases {
+		t.Run(tc.in, func(t *testing.T) {
+			gotTitle, gotYear := CleanQuery(tc.in)
+			if gotTitle != tc.wantTitle || gotYear != tc.wantYear {
+				t.Errorf("CleanQuery(%q) = (%q, %d), want (%q, %d)",
+					tc.in, gotTitle, gotYear, tc.wantTitle, tc.wantYear)
+			}
+		})
+	}
+}
+
+// The release-tag rules run on lowercased tokens, so restoring the display case
+// must not change which tokens survive. Anything that starts matching or
+// dropping tokens only because of casing (e.g. patEP's boundary class seeing an
+// uppercase letter) would show up here.
+func TestCleanQueryCaseRestoreDoesNotChangeTokenSelection(t *testing.T) {
+	// Both spellings must select exactly the same title tokens.
+	pairs := [][2]string{
+		{"The.Matrix.1999.1080p.BluRay.mkv", "the.matrix.1999.1080p.bluray.mkv"},
+		{"SE7EN.1995.1080p.BluRay.mkv", "se7en.1995.1080p.bluray.mkv"},
+		{"WEB.Therapy.S01E01.1080p.mkv", "web.therapy.s01e01.1080p.mkv"},
+		{"For.All.Mankind.S05E06.4K.mkv", "for.all.mankind.s05e06.4k.mkv"},
+	}
+	for _, pair := range pairs {
+		upper, upperYear := CleanQuery(pair[0])
+		lower, lowerYear := CleanQuery(pair[1])
+		if !equalFoldASCII(upper, lower) || upperYear != lowerYear {
+			t.Errorf("case changed token selection: %q -> (%q, %d) vs %q -> (%q, %d)",
+				pair[0], upper, upperYear, pair[1], lower, lowerYear)
+		}
+	}
+}
+
+func equalFoldASCII(a, b string) bool {
+	return strings.EqualFold(a, b)
+}
+
+// A filename with no case information (all lowercase, as many release names are)
+// must stay as-is rather than being title-cased: MeBox cannot invent casing.
+func TestCleanQueryKeepsLowercaseInputLowercase(t *testing.T) {
+	title, _ := CleanQuery("the.matrix.1999.1080p.mkv")
+	if title != "the matrix" {
+		t.Fatalf("CleanQuery = %q, want %q", title, "the matrix")
+	}
+}

internal/service/scraper_query_clean.go

@@ -172,10 +172,53 @@ func CleanQuery(raw string) (title string, year int) {
 		}
 		out = append(out, w)
 	}
-	title = strings.TrimSpace(strings.Join(out, " "))
+	title = strings.TrimSpace(strings.Join(restoreQueryTokenCase(name, out), " "))
 	return title, year
 }
 
+// queryTokenSeparators mirrors the separator split used while tokenising a
+// cleaned filename.
+var queryTokenSeparators = []string{".", "_", "-", "[", "]", "(", ")", "×"}
+
+// restoreQueryTokenCase maps the surviving lowercase tokens of a cleaned
+// filename back onto the spelling used in the original name.
+//
+// Tokenisation itself stays case-insensitive: the release-tag rules in this file
+// rely on lowercased tokens (and patEP's [^a-z] boundary would misfire on
+// "SE7EN"-style names if it ran on mixed case). Rebuilding the display title from
+// the original tokens afterwards keeps "The.Matrix.1999.1080p..." as "The Matrix"
+// instead of overwriting the user's media with a lowercased title — the lowercased
+// form also used to poison metadata lookups. Tokens with no original counterpart
+// (fragments produced by a removal) fall back to their lowercase form.
+func restoreQueryTokenCase(name string, tokens []string) []string {
+	if len(tokens) == 0 {
+		return tokens
+	}
+	originalByLower := make(map[string]string, len(tokens))
+	normalized := bracketedTag.ReplaceAllString(name, " ")
+	for _, sep := range queryTokenSeparators {
+		normalized = strings.ReplaceAll(normalized, sep, " ")
+	}
+	for _, token := range strings.Fields(normalized) {
+		key := strings.ToLower(token)
+		if key == "" || key == releaseSignalToken {
+			continue
+		}
+		if _, exists := originalByLower[key]; !exists {
+			originalByLower[key] = token
+		}
+	}
+	cased := make([]string, 0, len(tokens))
+	for _, token := range tokens {
+		if original, ok := originalByLower[token]; ok {
+			cased = append(cased, original)
+			continue
+		}
+		cased = append(cased, token)
+	}
+	return cased
+}
+
 func isASCIIWord(s string) bool {
 	if s == "" {
 		return false

internal/service/scraper_query_clean_regression_test.go

@@ -12,10 +12,10 @@ func TestCleanQueryKeepsCommonEnglishTitleWords(t *testing.T) {
 		wantTitle string
 		wantYear  int
 	}{
-		{"Mad.Max.1979.1080p.BluRay.mkv", "mad max", 1979},
-		{"Max.Payne.2008.1080p.WEB-DL.mkv", "max payne", 2008},
-		{"Web.Therapy.S01E01.1080p.WEB-DL.mkv", "web therapy", 0},
-		{"The.Web.2019.1080p.mkv", "the web", 2019},
+		{"Mad.Max.1979.1080p.BluRay.mkv", "Mad Max", 1979},
+		{"Max.Payne.2008.1080p.WEB-DL.mkv", "Max Payne", 2008},
+		{"Web.Therapy.S01E01.1080p.WEB-DL.mkv", "Web Therapy", 0},
+		{"The.Web.2019.1080p.mkv", "The Web", 2019},
 	}
 	for _, tc := range cases {
 		t.Run(tc.in, func(t *testing.T) {
@@ -43,8 +43,8 @@ func TestCleanQueryDoesNotTruncateAmbiguousTagToNothing(t *testing.T) {
 // truncation but must not discard the real title words that follow it.
 func TestCleanQueryKeepsTitleAfterYearPrefix(t *testing.T) {
 	cases := map[string]string{
-		"2019.Avatar.1080p.BluRay.mkv":     "avatar",
-		"2024.Dune.Part.Two.2160p.WEB.mkv": "dune part two",
+		"2019.Avatar.1080p.BluRay.mkv":     "Avatar",
+		"2024.Dune.Part.Two.2160p.WEB.mkv": "Dune Part Two",
 	}
 	for in, want := range cases {
 		t.Run(in, func(t *testing.T) {
@@ -62,9 +62,9 @@ func TestCleanQueryKeepsTitleAfterYearPrefix(t *testing.T) {
 func TestCleanQueryStillDropsTagsAfterReleaseSignal(t *testing.T) {
 	cases := map[string]string{
 		"复仇者联盟4.2019.BD.1080p.mkv":               "复仇者联盟4",
-		"The.Matrix.1999.1080p.WEB-DL.H265.mp4":  "the matrix",
-		"Oppenheimer.2023.2160p.UHD.BluRay.mkv":  "oppenheimer",
-		"Interstellar.2014.4k.hdr.dts.atmos.mkv": "interstellar",
+		"The.Matrix.1999.1080p.WEB-DL.H265.mp4":  "The Matrix",
+		"Oppenheimer.2023.2160p.UHD.BluRay.mkv":  "Oppenheimer",
+		"Interstellar.2014.4k.hdr.dts.atmos.mkv": "Interstellar",
 	}
 	for in, want := range cases {
 		t.Run(in, func(t *testing.T) {

internal/service/scraper_query_movie_test.go

@@ -30,7 +30,7 @@ func TestScrapeQueryCandidatesUseMovieFolderForGenericFilename(t *testing.T) {
 	if len(got) == 0 {
 		t.Fatal("scrapeQueryCandidates returned no candidates")
 	}
-	if got[0] != "inception" {
+	if got[0] != "Inception" {
 		t.Fatalf("first query candidate = %q, want movie folder title; all candidates=%#v", got[0], got)
 	}
 	for _, candidate := range got {
@@ -55,7 +55,7 @@ func TestScrapeQueryCandidatesUseMovieLibraryRootWhenMountedAtMovieFolder(t *tes
 	if len(got) == 0 {
 		t.Fatal("scrapeQueryCandidates returned no candidates")
 	}
-	if got[0] != "inception" {
+	if got[0] != "Inception" {
 		t.Fatalf("first query candidate = %q, want movie library root title; all candidates=%#v", got[0], got)
 	}
 }
@@ -73,7 +73,7 @@ func TestScrapeQueryCandidatesDoNotUseMovieCollectionFolderAsTitle(t *testing.T)
 	if len(got) == 0 {
 		t.Fatal("scrapeQueryCandidates returned no candidates")
 	}
-	if got[0] != "the hunger games catching fire" {
+	if got[0] != "The Hunger Games Catching Fire" {
 		t.Fatalf("first query candidate = %q, want individual movie title; all candidates=%#v", got[0], got)
 	}
 	for _, candidate := range got {
@@ -92,7 +92,10 @@ func TestEnrichOneUsesMovieFolderWhenFilenameIsGeneric(t *testing.T) {
 			http.NotFound(w, r)
 			return
 		}
-		if r.URL.Query().Get("query") != "inception" {
+		// Real metadata providers treat the query case-insensitively; this stub
+		// must too, otherwise it would only be pinned to the historical lowercased
+		// query spelling instead of the folder-fallback behaviour it guards.
+		if !strings.EqualFold(r.URL.Query().Get("query"), "inception") {
 			_ = json.NewEncoder(w).Encode(map[string]any{"results": []any{}})
 			return
 		}
@@ -149,7 +152,7 @@ func TestEnrichOneUsesMovieFolderWhenFilenameIsGeneric(t *testing.T) {
 	if got.ScrapeStatus != "matched" || got.TMDbID != 27205 || got.Title != "Inception" {
 		t.Fatalf("generic filename scrape did not use folder title: status=%q tmdb=%d title=%q queries=%v", got.ScrapeStatus, got.TMDbID, got.Title, queries)
 	}
-	if len(queries) == 0 || queries[0] != "inception" {
+	if len(queries) == 0 || !strings.EqualFold(queries[0], "inception") {
 		t.Fatalf("first tmdb query = %q, want folder title; all queries=%v", firstQuery(queries), queries)
 	}
 }

internal/service/scraper_query_test.go

@@ -23,19 +23,19 @@ func TestCleanQuery(t *testing.T) {
 		wantTitle string
 		wantYear  int
 	}{
-		{"Inception.2010.1080p.BluRay.x264.mkv", "inception", 2010},
-		{"The_Matrix_(1999).1080p.WEB-DL.H265.mp4", "the matrix", 1999},
+		{"Inception.2010.1080p.BluRay.x264.mkv", "Inception", 2010},
+		{"The_Matrix_(1999).1080p.WEB-DL.H265.mp4", "The Matrix", 1999},
 		{"interstellar.2014.4k.hdr.dts.atmos.mkv", "interstellar", 2014},
-		{"My Movie 2022 [HDR] (1080p) [TGx].mp4", "my movie", 2022},
-		{"NoYearOrTags.mkv", "noyearortags", 0},
-		{"亏成首富从游戏开始 The Richest in Game - S01E11 - 4K.mp4", "亏成首富从游戏开始 the richest in game", 0},
+		{"My Movie 2022 [HDR] (1080p) [TGx].mp4", "My Movie", 2022},
+		{"NoYearOrTags.mkv", "NoYearOrTags", 0},
+		{"亏成首富从游戏开始 The Richest in Game - S01E11 - 4K.mp4", "亏成首富从游戏开始 The Richest in Game", 0},
 		{"紫川.2024.S02E24.第24集.2160p.WEB-DL.H.265-ColorTV.mkv", "紫川", 2024},
 		{"紫川 (2024) {tmdb-247590}", "紫川", 2024},
-		{"HNTV.Spring.Festival.Gala.FPS.HLG-QHStudio.S01E202-DD5.QHstudIo.6.4K.ts", "hntv spring festival gala", 0},
-		{"Hntv Spring Festival Gala S01e (2026)", "hntv spring festival gala", 2026},
-		{"Motherhood Of Taihang Aac2 Mweb - S01E01-Aac2.Mweb.mkv", "motherhood of taihang", 0},
-		{"For All Mankind Atvp Hhweb - S05E06-DDP5.HHWEB.4K.mkv", "for all mankind", 0},
-		{"Hntv Spring Festival Gala Fps Hlg Qhstudio S01e (2026)", "hntv spring festival gala", 2026},
+		{"HNTV.Spring.Festival.Gala.FPS.HLG-QHStudio.S01E202-DD5.QHstudIo.6.4K.ts", "HNTV Spring Festival Gala", 0},
+		{"Hntv Spring Festival Gala S01e (2026)", "Hntv Spring Festival Gala", 2026},
+		{"Motherhood Of Taihang Aac2 Mweb - S01E01-Aac2.Mweb.mkv", "Motherhood Of Taihang", 0},
+		{"For All Mankind Atvp Hhweb - S05E06-DDP5.HHWEB.4K.mkv", "For All Mankind", 0},
+		{"Hntv Spring Festival Gala Fps Hlg Qhstudio S01e (2026)", "Hntv Spring Festival Gala", 2026},
 	}
 	for _, tc := range cases {
 		t.Run(tc.in, func(t *testing.T) {
@@ -75,7 +75,7 @@ func TestScrapeQueryCandidatesCleanDirtySeriesFolder(t *testing.T) {
 	if len(got) == 0 {
 		t.Fatal("scrapeQueryCandidates returned no candidates")
 	}
-	if got[0] != "motherhood of taihang" {
+	if got[0] != "Motherhood Of Taihang" {
 		t.Fatalf("first query candidate = %q, want cleaned series title; all candidates=%#v", got[0], got)
 	}
 	for _, candidate := range got {

internal/service/scraper_write_metadata_files.go

@@ -1,6 +1,7 @@
 package service
 
 import (
+	"bytes"
 	"context"
 	"os"
 	"path/filepath"
@@ -152,6 +153,13 @@ func (s *ScraperService) writeArtworkDataToPath(dir, name, ctype string, data []
 		return ""
 	}
 	dst := filepath.Join(dir, name+imageExtForContentType(ctype))
+	// 内容与现有文件完全一致时不重写：重写会刷新源文件 mtime，而缩略图缓存键
+	// 包含源文件大小与 mtime，一次「内容没变」的重新刮削会让该条目所有尺寸的
+	// 缩略图一并作废，下次浏览全部重新解码。
+	if artworkFileUnchanged(dst, data) {
+		s.log.Debug("scrape artwork unchanged", zap.String("dst", dst))
+		return dst
+	}
 	tmp, err := os.CreateTemp(dir, "img-*.tmp")
 	if err != nil {
 		s.log.Warn("scrape artwork temp create failed", zap.String("dir", dir), zap.Error(err))
@@ -186,6 +194,26 @@ func (s *ScraperService) writeArtworkDataToPath(dir, name, ctype string, data []
 	return dst
 }
 
+// artworkFileUnchanged 报告 path 是否已经就是 data 这些字节。
+//
+// 刮削在刷新海报地址时会无条件重写 sidecar 图片，即使下载回来的字节一模一样。
+// 缩略图缓存键包含源文件的大小与 mtime，因此这种无谓的重写会让该条目所有尺寸
+// 的缩略图一起失效；保持一致时直接复用原文件可以保住 mtime。
+func artworkFileUnchanged(path string, data []byte) bool {
+	if len(data) == 0 {
+		return false
+	}
+	info, err := os.Stat(path)
+	if err != nil || info.IsDir() || info.Size() != int64(len(data)) {
+		return false
+	}
+	existing, err := os.ReadFile(path) // #nosec G304 -- path is built from a sanitized media directory.
+	if err != nil {
+		return false
+	}
+	return bytes.Equal(existing, data)
+}
+
 // imageExtForContentType maps a detected image MIME type to a file extension.
 // Unknown image types fall back to a generic ".img" so we never write an empty
 // extension that could confuse media players.

internal/service/scraper_write_metadata_files_test.go

@@ -9,6 +9,7 @@ import (
 	"os"
 	"path/filepath"
 	"testing"
+	"time"
 
 	"github.com/glebarez/sqlite"
 	"go.uber.org/zap"
@@ -62,6 +63,47 @@ func TestWriteArtworkDataToPathReplacesExistingSidecar(t *testing.T) {
 	}
 }
 
+// TestWriteArtworkDataToPathSkipsUnchangedSidecar 验证"内容一致就不重写"：
+// 缩略图缓存键包含源文件的大小与 mtime，无谓的重写会让该条目所有尺寸的缩略图
+// 一起失效，下次浏览全部重新解码。
+func TestWriteArtworkDataToPathSkipsUnchangedSidecar(t *testing.T) {
+	scraper := &ScraperService{log: zap.NewNop()}
+	mediaDir := t.TempDir()
+	dst := scraper.writeArtworkDataToPath(mediaDir, "poster", "image/jpeg", testJPEG)
+	if dst == "" {
+		t.Fatal("expected a written destination path")
+	}
+	// 把 mtime 拨回过去，再写入完全相同的字节：文件必须原样保留。
+	past := time.Now().Add(-2 * time.Hour)
+	if err := os.Chtimes(dst, past, past); err != nil {
+		t.Fatal(err)
+	}
+	if got := scraper.writeArtworkDataToPath(mediaDir, "poster", "image/jpeg", testJPEG); got != dst {
+		t.Fatalf("destination = %q, want %q", got, dst)
+	}
+	info, err := os.Stat(dst)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if time.Since(info.ModTime()) < time.Minute {
+		t.Fatalf("mtime = %v, want the original past timestamp (identical artwork was rewritten)", info.ModTime())
+	}
+
+	// 内容变化时仍然必须覆盖。
+	changed := append([]byte(nil), testJPEG...)
+	changed[len(changed)-1] = 0x00
+	if got := scraper.writeArtworkDataToPath(mediaDir, "poster", "image/jpeg", changed); got != dst {
+		t.Fatalf("destination = %q, want %q", got, dst)
+	}
+	data, err := os.ReadFile(dst)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !bytes.Equal(data, changed) {
+		t.Fatal("changed artwork was not written")
+	}
+}
+
 func TestShouldCropAdultPosterWhenTMDbArtworkMatchedCodePath(t *testing.T) {
 	media := &model.Media{
 		Path:      filepath.Join(t.TempDir(), "IPX-235.mp4"),

internal/service/service_builder.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"errors"
 	"strings"
+	"time"
 
 	"go.uber.org/zap"
 
@@ -120,6 +121,8 @@ func (b *serviceContainerBuilder) initContentServices() {
 	b.c.Audit = NewAuditService(b.log, b.repos)
 	b.c.Strm = NewStrmService(b.cfg, b.log, b.repos, b.c.Crypto)
 	b.c.Cloud115 = NewCloud115PlaybackService(b.cfg, b.log, b.repos, b.c.Strm)
+	// 本地 HLS 档位的可用性取决于 ffmpeg 是否可用；未注入时按不可用处理。
+	b.c.Cloud115.SetTranscoder(b.c.Transcoder)
 	// ffmpeg/ffprobe 一键下载安装（data/tools/ffmpeg/）。
 	b.c.FFTools = NewFFmpegToolsService(b.cfg, b.log, b.repos)
 	// 弹幕 hash 识别需要把 strm 指向解析成可拉取的直链/本地路径。
@@ -164,11 +167,20 @@ func (b *serviceContainerBuilder) initAccessAndStorageServices() {
 	)
 	b.c.Scheduler.SetTaskTracker(b.c.Tasks)
 	b.c.Scheduler.SetOrganizePipeline(b.c.OrganizePipeline)
-	b.c.Scheduler.SetImagesMaxSizeMBProvider(func() int {
+	b.c.Scheduler.SetImageCachePolicyProvider(func() ImageCachePolicy {
 		if b.cfg == nil {
-			return 0
+			return ImageCachePolicy{}
 		}
-		return b.cfg.Cache.ImagesMaxSizeMB
+		policy := ImageCachePolicy{
+			TotalBytes: int64(b.cfg.Cache.ImagesMaxSizeMB) * 1024 * 1024,
+		}
+		if b.cfg.Cache.ImagesOriginalsMaxSizeMB > 0 {
+			policy.OriginalsBytes = int64(b.cfg.Cache.ImagesOriginalsMaxSizeMB) * 1024 * 1024
+		}
+		if b.cfg.Cache.ImagesOriginalsTTLHours > 0 {
+			policy.OriginalsAge = time.Duration(b.cfg.Cache.ImagesOriginalsTTLHours) * time.Hour
+		}
+		return policy
 	})
 }
 

internal/service/strm_service.go

@@ -892,6 +892,10 @@ func truncateStringRuneSafe(s string, maxBytes int) string {
 
 // cleanEntryName 清理单个目录名或文件名中的非法字符、控制字符、尾部点空格及 Windows 保留字，
 // 确保在 Windows (NTFS/FAT)、Linux (ext4/btrfs/xfs) 及 NAS/SMB 挂载环境下均安全可用。
+//
+// 只裁剪**尾部**的点与空格：那是 Windows 明确禁止的部分。前导点是合法且常见的
+// （隐藏目录/文件、.staging 之类），早先按 ". " 双向裁剪会把 `.media` 静默改写成
+// `media`，导致刮削海报边车被写到去掉点后的平行目录、媒体目录反而拿不到图。
 func cleanEntryName(name string, isDir bool) string {
 	name = strings.TrimSpace(name)
 	if name == "" {
@@ -900,7 +904,7 @@ func cleanEntryName(name string, isDir bool) string {
 
 	if isDir {
 		clean := sanitizeFilename(name)
-		clean = strings.Trim(clean, ". ")
+		clean = strings.TrimRight(clean, ". ")
 		if clean == "" {
 			return "unnamed"
 		}
@@ -921,7 +925,7 @@ func cleanEntryName(name string, isDir bool) string {
 	}
 
 	cleanBase := sanitizeFilename(base)
-	cleanBase = strings.Trim(cleanBase, ". ")
+	cleanBase = strings.TrimRight(cleanBase, ". ")
 	if cleanBase == "" {
 		cleanBase = "unnamed"
 	}

internal/service/subtitle.go

@@ -410,14 +410,10 @@ func (s *SubtitleService) Serve(ctx context.Context, mediaID, sub string, w io.W
 		}
 		return s.serveEmbedded(ctx, m, index, w)
 	}
-	abs, err := filepath.Abs(sub)
+	abs, err := readExternalSubtitlePath(m, sub)
 	if err != nil {
 		return err
 	}
-	mediaDir, _ := filepath.Abs(filepath.Dir(m.Path))
-	if !pathWithin(abs, mediaDir) {
-		return fmt.Errorf("path escape")
-	}
 
 	f, err := os.Open(abs) // #nosec G304 -- abs is constrained to the media file directory with pathWithin.
 	if err != nil {
@@ -428,14 +424,17 @@ func (s *SubtitleService) Serve(ctx context.Context, mediaID, sub string, w io.W
 	if err != nil {
 		return err
 	}
+	// 非 UTF-8 的外挂字幕（UTF-16、GBK/Big5 等）必须先归一化：浏览器只能按
+	// UTF-8 解析 <track> 内容，否则整篇都会变成替换字符。
+	text := decodeSubtitleText(body)
 
 	switch strings.ToLower(filepath.Ext(abs)) {
 	case ".vtt":
-		_, err = w.Write(body)
+		_, err = io.WriteString(w, text)
 	case ".srt":
-		_, err = w.Write([]byte(srtToVTT(string(body))))
+		_, err = io.WriteString(w, srtToVTT(text))
 	case ".ass", ".ssa":
-		_, err = w.Write([]byte(assToVTT(string(body))))
+		_, err = io.WriteString(w, assToVTT(text))
 	default:
 		return errors.New("unsupported subtitle format")
 	}
@@ -517,12 +516,42 @@ func (s *SubtitleService) ServeASS(ctx context.Context, mediaID, sub string, w i
 	}
 	switch strings.ToLower(filepath.Ext(sub)) {
 	case ".ass", ".ssa":
-		return s.ServeRaw(ctx, mediaID, sub, w)
+		// 外挂 ASS 同样要归一化成 UTF-8：libass 只认 UTF-8，UTF-16/GBK 的
+		// 字幕交给它会解析不到任何事件，表现为「字幕选中了却不显示」。
+		abs, err := readExternalSubtitlePath(m, sub)
+		if err != nil {
+			return err
+		}
+		f, err := os.Open(abs) // #nosec G304 -- abs is constrained to the media file directory with pathWithin.
+		if err != nil {
+			return err
+		}
+		defer f.Close()
+		body, err := io.ReadAll(f)
+		if err != nil {
+			return err
+		}
+		_, err = io.WriteString(w, decodeSubtitleText(body))
+		return err
 	default:
 		return errors.New("subtitle is not ASS/SSA")
 	}
 }
 
+// readExternalSubtitlePath 校验外挂字幕路径（必须落在媒体文件所在目录内）并返回
+// 绝对路径。Serve 与 ServeASS 共用，避免两处各自实现出现安全口径不一致。
+func readExternalSubtitlePath(m *model.Media, sub string) (string, error) {
+	abs, err := filepath.Abs(sub)
+	if err != nil {
+		return "", err
+	}
+	mediaDir, _ := filepath.Abs(filepath.Dir(m.Path))
+	if !pathWithin(abs, mediaDir) {
+		return "", fmt.Errorf("path escape")
+	}
+	return abs, nil
+}
+
 func (s *SubtitleService) serveEmbeddedASS(ctx context.Context, media *model.Media, streamIndex int, w io.Writer) error {
 	input, err := s.resolveInput(ctx, media)
 	if err != nil {

internal/service/subtitle_encoding.go

@@ -0,0 +1,167 @@
+package service
+
+import (
+	"bytes"
+	"encoding/binary"
+	"math"
+	"regexp"
+	"strings"
+	"unicode"
+	"unicode/utf16"
+	"unicode/utf8"
+
+	"golang.org/x/text/encoding"
+	"golang.org/x/text/encoding/japanese"
+	"golang.org/x/text/encoding/korean"
+	"golang.org/x/text/encoding/simplifiedchinese"
+	"golang.org/x/text/encoding/traditionalchinese"
+)
+
+// 外挂字幕文件的编码归一化。
+//
+// 字幕文件由字幕组/工具生成，编码并不统一：除 UTF-8 外，中文圈常见 GBK/GB18030，
+// 港台常见 Big5，部分 Windows 工具（Aegisub 早期版本、Subtitle Edit 的某些导出）
+// 会写成带 BOM 的 UTF-16LE。浏览器 <track> 与 libass（JASSUB）都只按 UTF-8 解析，
+// 后端如果原样下发字节，非 UTF-8 的字幕会整篇变成替换字符——表现就是「字幕轨已经
+// 选中，但画面上一个字都不显示」。所以外挂字幕在服务端就要统一解码成 UTF-8。
+//
+// 注意：只用于下发给浏览器/播放器的文本路径。Emby 兼容层的 /subtitles 原始字节
+// 接口（ServeRaw）必须保持字节一致，不能经过这里。
+
+// 各编码的 BOM 前缀。
+var (
+	utf8BOM    = []byte{0xEF, 0xBB, 0xBF}
+	utf16BEBOM = []byte{0xFE, 0xFF}
+	utf16LEBOM = []byte{0xFF, 0xFE}
+)
+
+// subtitleLegacyEncodings 是无 BOM 时的候选编码，按优先级排列。
+// GB18030 是 GBK/GB2312 的超集，中文外挂字幕绝大多数是它，因此排在最前——
+// 多个候选都能解出合法字幕时，靠前的优先。
+var subtitleLegacyEncodings = []struct {
+	name string
+	enc  encoding.Encoding
+}{
+	{"gb18030", simplifiedchinese.GB18030},
+	{"big5", traditionalchinese.Big5},
+	{"shift_jis", japanese.ShiftJIS},
+	{"euc-kr", korean.EUCKR},
+}
+
+// subtitleTimecodePattern 匹配 SRT（00:00:01,000 --> ...）与 WebVTT
+// （00:00:01.000 --> ...）的时间行。
+var subtitleTimecodePattern = regexp.MustCompile(`\d{1,2}:\d{2}:\d{2}[,.]\d{1,3}\s*-->`)
+
+// decodeSubtitleText 把字幕文件字节归一化成 UTF-8 文本。识别不出编码时原样返回，
+// 保证不会比「直接透传」更糟。
+func decodeSubtitleText(raw []byte) string {
+	switch {
+	case bytes.HasPrefix(raw, utf8BOM):
+		return string(raw[len(utf8BOM):])
+	case bytes.HasPrefix(raw, utf16LEBOM):
+		return decodeUTF16(raw[len(utf16LEBOM):], false)
+	case bytes.HasPrefix(raw, utf16BEBOM):
+		return decodeUTF16(raw[len(utf16BEBOM):], true)
+	}
+
+	// 无 BOM 且是合法 UTF-8：这是绝大多数情况，直接返回。
+	if utf8.Valid(raw) {
+		return string(raw)
+	}
+
+	best := ""
+	bestScore := math.MinInt
+	consider := func(text string) {
+		if !utf8.ValidString(text) || !looksLikeSubtitle(text) {
+			return
+		}
+		score := subtitlePlausibility(text)
+		if score > bestScore {
+			best, bestScore = text, score
+		}
+	}
+
+	// 无 BOM 的 UTF-16 少见（多来自手工编辑），但一旦是它，所有单字节编码都会
+	// 解出带大量 NUL 的垃圾，因此先用「格式特征」筛掉这些候选。
+	consider(decodeUTF16(raw, false))
+	consider(decodeUTF16(raw, true))
+
+	for _, candidate := range subtitleLegacyEncodings {
+		text, err := candidate.enc.NewDecoder().Bytes(raw)
+		if err != nil {
+			continue
+		}
+		consider(string(text))
+	}
+	if best != "" {
+		return best
+	}
+	return string(raw)
+}
+
+// decodeUTF16 按给定字节序把 UTF-16 字节解码为字符串（含代理对）。
+func decodeUTF16(raw []byte, bigEndian bool) string {
+	units := make([]uint16, 0, len(raw)/2)
+	for i := 0; i+1 < len(raw); i += 2 {
+		if bigEndian {
+			units = append(units, binary.BigEndian.Uint16(raw[i:i+2]))
+		} else {
+			units = append(units, binary.LittleEndian.Uint16(raw[i:i+2]))
+		}
+	}
+	return string(utf16.Decode(units))
+}
+
+// looksLikeSubtitle 判断解码结果是否真的像一份字幕。解码器「没报错」并不代表
+// 解对了——很多编码能把任意字节解成合法文本，只有格式特征才能确认。
+func looksLikeSubtitle(text string) bool {
+	head := text
+	if len(head) > 4096 {
+		head = head[:4096]
+	}
+	if strings.Contains(head, "[Script Info]") || strings.Contains(head, "Dialogue:") {
+		return true // ASS/SSA
+	}
+	if strings.HasPrefix(strings.TrimSpace(head), "WEBVTT") {
+		return true
+	}
+	return subtitleTimecodePattern.MatchString(head) // SRT / WebVTT 时间行
+}
+
+// subtitlePlausibility 给解码结果打分，返回 -100..100 的可信度：正常文本字符加分，
+// 替换字符（解码失败）、控制字符、未分配码位扣分。
+//
+// 刻意用比例而不是字符数量：单字节回退会把一个双字节字符拆成两个字符，字符多的
+// 候选看起来「更正常」，实际是解错了——GBK 中文字幕被当成 Shift_JIS 解，就会得到
+// 一串半角片假名，数量比正确结果还多。半角片假名单独扣分正是为了压住这种情况。
+func subtitlePlausibility(text string) int {
+	total := 0
+	weight := 0
+	for _, r := range text {
+		total++
+		switch {
+		case r == utf8.RuneError:
+			weight -= 8
+		case r == '\n' || r == '\r' || r == '\t':
+			// 正常换行不计分也不扣分。
+		case r < 0x20 || r == 0x7F:
+			weight -= 8
+		case unicode.Is(unicode.Co, r) || unicode.Is(unicode.Cn, r):
+			weight -= 4
+		case r >= 0xFF61 && r <= 0xFF9F:
+			// 半角片假名：GBK 汉字被按 Shift_JIS 解出来的典型产物。
+			weight -= 2
+		case unicode.In(r, unicode.Han, unicode.Hiragana, unicode.Katakana, unicode.Hangul),
+			unicode.Is(unicode.Latin, r), unicode.IsDigit(r), unicode.IsPunct(r):
+			weight++
+		case unicode.IsSpace(r):
+			// 空格（含全角空格）不参与打分。
+		default:
+			weight--
+		}
+	}
+	if total == 0 {
+		return 0
+	}
+	return weight * 100 / total
+}

internal/service/subtitle_encoding_test.go

@@ -0,0 +1,169 @@
+package service
+
+import (
+	"bytes"
+	"encoding/binary"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"unicode/utf16"
+
+	"github.com/glebarez/sqlite"
+	"go.uber.org/zap"
+	"golang.org/x/text/encoding/simplifiedchinese"
+	"gorm.io/gorm"
+
+	"github.com/truewhile/MeBox/internal/config"
+	"github.com/truewhile/MeBox/internal/model"
+	"github.com/truewhile/MeBox/internal/repository"
+)
+
+const assFixture = "[Script Info]\n" +
+	"Title:Railgun 01 BD\n" +
+	"ScriptType:v4.00+\n" +
+	"\n" +
+	"[Events]\n" +
+	"Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text\n" +
+	"Dialogue: 0,0:00:01.00,0:00:02.00,Default,,0,0,0,,放て！心に刻んだ夢を\n"
+
+const srtFixture = "1\n" +
+	"00:00:01,000 --> 00:00:02,000\n" +
+	"只有我的超电磁炮\n" +
+	"\n"
+
+func encodeUTF16(s string, bigEndian bool, withBOM bool) []byte {
+	units := utf16.Encode([]rune(s))
+	out := make([]byte, 0, len(units)*2+2)
+	if withBOM {
+		if bigEndian {
+			out = append(out, 0xFE, 0xFF)
+		} else {
+			out = append(out, 0xFF, 0xFE)
+		}
+	}
+	buf := make([]byte, 2)
+	for _, unit := range units {
+		if bigEndian {
+			binary.BigEndian.PutUint16(buf, unit)
+		} else {
+			binary.LittleEndian.PutUint16(buf, unit)
+		}
+		out = append(out, buf...)
+	}
+	return out
+}
+
+func encodeGBK(t *testing.T, s string) []byte {
+	t.Helper()
+	encoded, err := simplifiedchinese.GBK.NewEncoder().Bytes([]byte(s))
+	if err != nil {
+		t.Fatalf("encode GBK: %v", err)
+	}
+	return encoded
+}
+
+// 这是本次线上问题的核心场景：字幕组给的 ASS 是带 BOM 的 UTF-16LE，
+// 之前原样下发，浏览器按 UTF-8 解开后整篇是替换字符，一个字都渲染不出来。
+func TestDecodeSubtitleTextNormalisesLegacyEncodings(t *testing.T) {
+	cases := []struct {
+		name string
+		raw  []byte
+	}{
+		{"utf8", []byte(assFixture)},
+		{"utf8-bom", append([]byte{0xEF, 0xBB, 0xBF}, []byte(assFixture)...)},
+		{"utf16le-bom", encodeUTF16(assFixture, false, true)},
+		{"utf16be-bom", encodeUTF16(assFixture, true, true)},
+		{"utf16le-no-bom", encodeUTF16(assFixture, false, false)},
+		{"gb18030", encodeGBK(t, assFixture)},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			got := decodeSubtitleText(tc.raw)
+			if !strings.Contains(got, "[Script Info]") {
+				t.Fatalf("decoded text lost ASS header: %q", got)
+			}
+			if !strings.Contains(got, "Dialogue:") {
+				t.Fatalf("decoded text lost Dialogue line: %q", got)
+			}
+			if !strings.Contains(got, "放て！心に刻んだ夢を") {
+				t.Fatalf("decoded text lost non-ASCII content: %q", got)
+			}
+			if strings.ContainsRune(got, 0) || strings.ContainsRune(got, '\uFFFD') {
+				t.Fatalf("decoded text still contains NUL or replacement runes: %q", got)
+			}
+		})
+	}
+}
+
+func TestDecodeSubtitleTextKeepsUnknownBytes(t *testing.T) {
+	raw := []byte("not a subtitle at all: \x81\x82\x83\x84")
+	if got := decodeSubtitleText(raw); got != string(raw) {
+		t.Fatalf("decodeSubtitleText mangled non-subtitle bytes: %q", got)
+	}
+}
+
+// 接口层回归：外挂 UTF-16 字幕经 /subtitles/:id/ass（libass 路径）与
+// /subtitles/:id（WebVTT 降级路径）下发时都必须是可解析的 UTF-8。
+func TestSubtitleHandlersNormaliseUTF16Sidecar(t *testing.T) {
+	db, err := gorm.Open(sqlite.Open("file:subtitle-encoding?mode=memory&cache=shared"), &gorm.Config{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := db.AutoMigrate(&model.Library{}, &model.Media{}); err != nil {
+		t.Fatal(err)
+	}
+
+	dir := t.TempDir()
+	videoPath := filepath.Join(dir, "Railgun S01E01.mkv")
+	assPath := filepath.Join(dir, "Railgun S01E01.ass")
+	srtPath := filepath.Join(dir, "Railgun S01E01.zh.srt")
+	if err := os.WriteFile(videoPath, []byte("video"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(assPath, encodeUTF16(assFixture, false, true), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(srtPath, encodeGBK(t, srtFixture), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	media := model.Media{Title: "Railgun", Path: videoPath}
+	if err := db.Create(&media).Error; err != nil {
+		t.Fatal(err)
+	}
+	svc := NewSubtitleService(&config.Config{}, zap.NewNop(), repository.New(db))
+
+	var buf bytes.Buffer
+	if err := svc.ServeASS(t.Context(), media.ID, assPath, &buf); err != nil {
+		t.Fatalf("ServeASS: %v", err)
+	}
+	ass := buf.String()
+	if !bytes.HasPrefix(buf.Bytes(), []byte("[Script Info]")) {
+		t.Fatalf("ServeASS output is not UTF-8 ASS: %q", ass)
+	}
+	if !strings.Contains(ass, "放て！心に刻んだ夢を") {
+		t.Fatalf("ServeASS output lost dialogue text: %q", ass)
+	}
+
+	buf.Reset()
+	if err := svc.Serve(t.Context(), media.ID, assPath, &buf); err != nil {
+		t.Fatalf("Serve(ass): %v", err)
+	}
+	vtt := buf.String()
+	if !strings.HasPrefix(vtt, "WEBVTT") || !strings.Contains(vtt, "-->") {
+		t.Fatalf("ASS->WebVTT fallback produced no cues: %q", vtt)
+	}
+	if !strings.Contains(vtt, "放て！心に刻んだ夢を") {
+		t.Fatalf("ASS->WebVTT fallback lost dialogue text: %q", vtt)
+	}
+
+	buf.Reset()
+	if err := svc.Serve(t.Context(), media.ID, srtPath, &buf); err != nil {
+		t.Fatalf("Serve(srt): %v", err)
+	}
+	vtt = buf.String()
+	if !strings.Contains(vtt, "只有我的超电磁炮") {
+		t.Fatalf("GBK SRT was not transcoded: %q", vtt)
+	}
+}

internal/service/test_db_test.go

@@ -1,13 +1,24 @@
 package service
 
 import (
+	"reflect"
 	"testing"
 
 	"github.com/glebarez/sqlite"
 	"gorm.io/gorm"
 	"gorm.io/gorm/logger"
+
+	"github.com/truewhile/MeBox/internal/model"
 )
 
+// newServiceTestDB opens an in-memory DB for service tests.
+//
+// The full model set is always migrated, on top of any explicitly requested
+// models. Service code probes tables that a given test may not care about
+// (settings for adult visibility, play_profiles for playback, media for library
+// counts); with a partial schema those probes fail with "no such table" and the
+// default GORM logger floods the test output, burying real failures. Migrating
+// everything keeps fixtures faithful to production and the output quiet.
 func newServiceTestDB(t *testing.T, models ...any) *gorm.DB {
 	t.Helper()
 	db, err := gorm.Open(sqlite.Open(":memory:"), &gorm.Config{Logger: logger.Default.LogMode(logger.Silent)})
@@ -17,10 +28,21 @@ func newServiceTestDB(t *testing.T, models ...any) *gorm.DB {
 	if sqlDB, err := db.DB(); err == nil {
 		t.Cleanup(func() { _ = sqlDB.Close() })
 	}
-	if len(models) > 0 {
-		if err := db.AutoMigrate(models...); err != nil {
-			t.Fatal(err)
+	toMigrate := model.AllModels()
+	seen := make(map[string]struct{}, len(toMigrate)+len(models))
+	for _, m := range toMigrate {
+		seen[reflect.TypeOf(m).String()] = struct{}{}
+	}
+	for _, m := range models {
+		key := reflect.TypeOf(m).String()
+		if _, ok := seen[key]; ok {
+			continue
 		}
+		seen[key] = struct{}{}
+		toMigrate = append(toMigrate, m)
+	}
+	if err := db.AutoMigrate(toMigrate...); err != nil {
+		t.Fatal(err)
 	}
 	return db
 }

internal/service/transcoder.go

@@ -54,6 +54,11 @@ type TranscoderService struct {
 	startGates  sync.Map // mediaID -> *sync.Mutex
 	strmResolve func(ctx context.Context, raw string) (*StrmPlayResult, error)
 	probe       *FFprobeService
+
+	// availMu guards the cached ffmpeg availability used by Available().
+	availMu        sync.Mutex
+	availCheckedAt time.Time
+	availValue     bool
 }
 
 // hlsJob holds the live state of one ffmpeg run.
@@ -99,6 +104,37 @@ func NewTranscoderService(cfg *config.Config, log *zap.Logger, repo *repository.
 	}
 }
 
+// transcodeAvailabilityTTL bounds how long an Available() result is reused.
+// Validating ffmpeg actually spawns it, and /playback is polled on every player
+// load, so the answer is cached briefly. It stays short so a freshly installed
+// ffmpeg becomes usable without restarting the server.
+const transcodeAvailabilityTTL = 20 * time.Second
+
+// Available reports whether HLS transcoding can run right now: transcoding must
+// be enabled by configuration and a usable ffmpeg must resolve. Callers use it to
+// stop advertising local HLS renditions that would only fail with a 500.
+func (t *TranscoderService) Available() bool {
+	if t == nil || t.cfg == nil {
+		return false
+	}
+	if !t.cfg.Transcoder.Enabled {
+		return false
+	}
+	t.availMu.Lock()
+	defer t.availMu.Unlock()
+	if !t.availCheckedAt.IsZero() && time.Since(t.availCheckedAt) < transcodeAvailabilityTTL {
+		return t.availValue
+	}
+	_, err := t.resolveFFmpegPath()
+	available := err == nil
+	if !available && t.log != nil {
+		t.log.Debug("local HLS renditions unavailable", zap.Error(err))
+	}
+	t.availValue = available
+	t.availCheckedAt = time.Now()
+	return available
+}
+
 // HLSDir is the per-media directory that holds index.m3u8 + segment files.
 func (t *TranscoderService) HLSDir(mediaID string) string {
 	return filepath.Join(t.cfg.Cache.CacheDir, "hls", mediaID)

internal/service/transcoder_availability_test.go

@@ -0,0 +1,67 @@
+package service
+
+import (
+	"path/filepath"
+	"testing"
+
+	"go.uber.org/zap"
+
+	"github.com/truewhile/MeBox/internal/config"
+)
+
+// Available() gates the local HLS renditions advertised to the player, so it must
+// be false both when transcoding is disabled by configuration and when no usable
+// ffmpeg can be resolved.
+func TestTranscoderAvailableGatesOnConfigAndBinary(t *testing.T) {
+	cases := []struct {
+		name       string
+		enabled    bool
+		ffmpegPath string
+	}{
+		{name: "disabled by config", enabled: false, ffmpegPath: filepath.Join(t.TempDir(), "missing-ffmpeg")},
+		{name: "no ffmpeg binary", enabled: true, ffmpegPath: filepath.Join(t.TempDir(), "missing-ffmpeg")},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			cfg := &config.Config{}
+			cfg.Transcoder.Enabled = tc.enabled
+			cfg.App.FFmpegPath = tc.ffmpegPath
+			svc := NewTranscoderService(cfg, zap.NewNop(), nil, NewHub(zap.NewNop()))
+
+			if svc.Available() {
+				t.Fatalf("Available() = true, want false (enabled=%v path=%q)", tc.enabled, tc.ffmpegPath)
+			}
+		})
+	}
+}
+
+// A nil service or nil config must not panic: playback info is built on paths
+// where wiring may be partial.
+func TestTranscoderAvailableNilSafe(t *testing.T) {
+	var nilSvc *TranscoderService
+	if nilSvc.Available() {
+		t.Fatal("nil transcoder should report unavailable")
+	}
+	svc := &TranscoderService{}
+	if svc.Available() {
+		t.Fatal("transcoder without config should report unavailable")
+	}
+}
+
+// Playback info must not advertise local HLS renditions when the playback service
+// has no transcoder wired in.
+func TestCloud115PlaybackServiceWithoutTranscoderMarksLocalQualitiesUnavailable(t *testing.T) {
+	svc := NewCloud115PlaybackService(&config.Config{}, zap.NewNop(), nil, nil)
+	if svc.localTranscodeAvailable() {
+		t.Fatal("playback service without a transcoder must report local HLS unavailable")
+	}
+	options := LocalQualityOptions(nil, svc.localTranscodeAvailable())
+	if len(options) == 0 {
+		t.Fatal("local quality presets should still be listed for the UI")
+	}
+	for _, option := range options {
+		if option.Available {
+			t.Fatalf("quality %s advertised as available without a transcoder", option.ID)
+		}
+	}
+}

internal/service/walk.go

@@ -14,14 +14,22 @@ type walkInfo struct {
 }
 
 // walk traverses root depth-first calling fn for every entry. Hidden
-// directories (starting with ".") are skipped.
+// directories (starting with ".") below the root are skipped.
+//
+// The root itself is deliberately exempt from the hidden-directory rule.
+// WalkDir reports the root with its own base name, so a root such as
+// "/media/.staging" used to match the hidden check and yield filepath.SkipDir
+// immediately — the entire tree was skipped silently, with no error and no
+// warning. A library or organize source may legitimately live in a
+// dot-prefixed directory, so only children are filtered.
 func walk(root string, fn func(string, walkInfo) error) error {
-	return filepath.WalkDir(root, func(path string, d fs.DirEntry, err error) error {
+	rootPath := filepath.Clean(root)
+	return filepath.WalkDir(rootPath, func(path string, d fs.DirEntry, err error) error {
 		if err != nil {
 			return nil // best effort — keep walking
 		}
 		name := d.Name()
-		if d.IsDir() && name != "." && len(name) > 1 && name[0] == '.' {
+		if d.IsDir() && path != rootPath && name != "." && len(name) > 1 && name[0] == '.' {
 			return filepath.SkipDir
 		}
 		info := walkInfo{isDir: d.IsDir()}

internal/service/walk_dotdir_test.go

@@ -0,0 +1,115 @@
+package service
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+
+	"go.uber.org/zap"
+
+	"github.com/truewhile/MeBox/internal/config"
+	"github.com/truewhile/MeBox/internal/model"
+)
+
+func writeWalkFile(t *testing.T, path string) {
+	t.Helper()
+	if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
+		t.Fatalf("mkdir %s: %v", filepath.Dir(path), err)
+	}
+	if err := os.WriteFile(path, []byte("data"), 0o644); err != nil {
+		t.Fatalf("write %s: %v", path, err)
+	}
+}
+
+func collectWalkPaths(t *testing.T, root string) []string {
+	t.Helper()
+	var seen []string
+	if err := walk(root, func(path string, info walkInfo) error {
+		if !info.isDir {
+			seen = append(seen, path)
+		}
+		return nil
+	}); err != nil {
+		t.Fatalf("walk %s: %v", root, err)
+	}
+	return seen
+}
+
+// A dot-prefixed walk root must not be treated as a hidden directory: doing so
+// skipped the whole tree silently, so a library rooted at e.g. "/media/.staging"
+// indexed nothing and the organizer reported "0 organized" with no error.
+func TestWalkIncludesDotPrefixedRoot(t *testing.T) {
+	parent := t.TempDir()
+	dotRoot := filepath.Join(parent, ".dotroot")
+	want := filepath.Join(dotRoot, "Movie (2021)", "Movie (2021).mkv")
+	writeWalkFile(t, want)
+
+	got := collectWalkPaths(t, dotRoot)
+	if len(got) != 1 || filepath.Clean(got[0]) != filepath.Clean(want) {
+		t.Fatalf("walk over dot-prefixed root = %#v, want the one file inside it", got)
+	}
+}
+
+// Hidden directories *below* the root are still skipped.
+func TestWalkStillSkipsHiddenChildDirectories(t *testing.T) {
+	root := t.TempDir()
+	visible := filepath.Join(root, "Movie (2021)", "Movie (2021).mkv")
+	hidden := filepath.Join(root, ".trash", "Old Movie.mkv")
+	deepHidden := filepath.Join(root, "Season 01", ".thumbnails", "thumb.mkv")
+	writeWalkFile(t, visible)
+	writeWalkFile(t, hidden)
+	writeWalkFile(t, deepHidden)
+
+	got := collectWalkPaths(t, root)
+	if len(got) != 1 || filepath.Clean(got[0]) != filepath.Clean(visible) {
+		t.Fatalf("walk = %#v, want only the visible file", got)
+	}
+}
+
+// End-to-end: a library whose root directory starts with "." must be scanned.
+func TestScanLibraryIndexesDotPrefixedRoot(t *testing.T) {
+	sc, repos := newScannerTestEnv(t)
+	root := filepath.Join(t.TempDir(), ".media")
+	lib := model.Library{Name: "Movies", Path: root, Type: "movie", Enabled: true}
+	if err := repos.Library.Create(t.Context(), &lib); err != nil {
+		t.Fatal(err)
+	}
+	writeWalkFile(t, filepath.Join(root, "Movie (2021)", "Movie (2021).mkv"))
+
+	res, err := sc.ScanLibrary(t.Context(), lib.ID)
+	if err != nil {
+		t.Fatalf("scan: %v", err)
+	}
+	if res.Added != 1 {
+		t.Fatalf("scan added=%d (errors=%v), want 1 for a dot-prefixed library root", res.Added, res.Errors)
+	}
+	if got := countMedia(t, repos); got != 1 {
+		t.Fatalf("media count = %d, want 1", got)
+	}
+}
+
+// End-to-end: organizing from a dot-prefixed source directory must find files.
+func TestOrganizeDirectoryReadsDotPrefixedSource(t *testing.T) {
+	repos := newOrganizerTestRepo(t)
+	cfg := &config.Config{}
+	cfg.Organizer.SmartClassify = true
+	organizer := NewOrganizerService(cfg, zap.NewNop(), repos)
+
+	root := t.TempDir()
+	src := filepath.Join(root, ".downloads")
+	dest := filepath.Join(root, "media")
+	writeWalkFile(t, filepath.Join(src, "Oppenheimer.2023.2160p.WEB-DL.H265", "oppenheimer-2160p.mkv"))
+
+	res, err := organizer.OrganizeDirectory(t.Context(), OrganizeOptions{
+		SourcePath:   src,
+		DestPath:     dest,
+		TransferMode: TransferCopy,
+	})
+	if err != nil {
+		t.Fatalf("organize directory: %v", err)
+	}
+	if res.Organized != 1 {
+		t.Fatalf("organized=%d skipped=%d errors=%v, want the file under a dot-prefixed source to be organized",
+			res.Organized, res.Skipped, res.Errors)
+	}
+}

web/public/artwork-cache-sw.js

@@ -1,8 +1,22 @@
 const ARTWORK_CACHE_PREFIX = 'mebox-artwork-'
-const ARTWORK_CACHE = `${ARTWORK_CACHE_PREFIX}v2`
+const ARTWORK_CACHE = `${ARTWORK_CACHE_PREFIX}v3`
 const MIN_CACHEABLE_ARTWORK_BYTES = 128
 const STRIP_QUERY_KEYS = ['token', 'profile_id', 'profile_pin_token']
 
+// 浏览器侧作品缓存的容量上限。
+//
+// 没有上限时 Cache Storage 会一直增长，浏览器在存储压力下会整体清空该
+// origin 的缓存，失效时机完全不可控。条数上限每次写入都检查（只是读一遍
+// 键，开销小）；字节上限最多每两分钟统计一次，因为需要逐个读取已存响应的
+// Content-Length。两者都淘汰到上限的 80%，避免刚清完又立刻触发。
+const MAX_CACHE_ENTRIES = 500
+const MAX_CACHE_BYTES = 64 * 1024 * 1024
+const TRIM_TARGET_RATIO = 0.8
+const BYTE_TRIM_INTERVAL_MS = 2 * 60 * 1000
+
+let lastByteTrimAt = 0
+let trimInFlight = null
+
 function isArtworkRequest(url) {
   if (url.origin !== self.location.origin) return false
   if (url.pathname === '/api/img') return true
@@ -54,10 +68,69 @@ async function cacheArtwork(request) {
   if (cacheResponse) {
     await cache.put(cacheKey, cacheResponse)
     await deleteOldArtworkVariants(cache, cacheKey)
+    await trimArtworkCache(cache)
   }
   return response
 }
 
+// trimArtworkCache 把缓存裁剪回容量上限之内，按写入顺序淘汰（最旧的先删）。
+// 同一时刻只允许一次裁剪在跑。
+async function trimArtworkCache(cache) {
+  if (trimInFlight) return trimInFlight
+  trimInFlight = runArtworkTrim(cache).catch(() => undefined)
+  try {
+    await trimInFlight
+  } finally {
+    trimInFlight = null
+  }
+}
+
+async function runArtworkTrim(cache) {
+  const keys = await cache.keys()
+  if (keys.length > MAX_CACHE_ENTRIES) {
+    const keep = Math.floor(MAX_CACHE_ENTRIES * TRIM_TARGET_RATIO)
+    await deleteOldestArtwork(cache, keys, keys.length - keep)
+    return
+  }
+
+  const now = Date.now()
+  if (now - lastByteTrimAt < BYTE_TRIM_INTERVAL_MS) return
+  lastByteTrimAt = now
+
+  const sizes = []
+  let total = 0
+  for (const key of keys) {
+    const response = await cache.match(key)
+    const size = cacheableResponseSize(response)
+    sizes.push(size)
+    total += size
+  }
+  if (total <= MAX_CACHE_BYTES) return
+
+  const target = Math.floor(MAX_CACHE_BYTES * TRIM_TARGET_RATIO)
+  let remaining = total
+  const evicted = []
+  for (let i = 0; i < keys.length && remaining > target; i += 1) {
+    remaining -= sizes[i]
+    evicted.push(keys[i])
+  }
+  await Promise.all(evicted.map((key) => cache.delete(key)))
+}
+
+// cacheableResponseSize 用 Content-Length 估算体积。MeBox 的图片响应由
+// ServeContent 生成，始终带该头；缺失时按 0 计（只是少算，不会误删）。
+function cacheableResponseSize(response) {
+  if (!response) return 0
+  const size = Number(response.headers.get('Content-Length') || '0')
+  return Number.isFinite(size) && size > 0 ? size : 0
+}
+
+// deleteOldestArtwork 按 cache.keys() 的顺序（写入顺序）删除最旧的若干条。
+async function deleteOldestArtwork(cache, keys, count) {
+  const victims = keys.slice(0, Math.max(0, count))
+  await Promise.all(victims.map((key) => cache.delete(key)))
+}
+
 async function cloneCacheableArtworkResponse(response) {
   if (!response.ok) return null
   const contentType = response.headers.get('Content-Type') || ''

web/src/api/client.ts

@@ -230,6 +230,27 @@ export type ImageURLOptions =
       quality?: number
     }
 
+// ARTWORK is the shared set of thumbnail sizes the whole app requests.
+//
+// The server caches thumbnails per (source file, maxWidth, maxHeight, quality):
+// every extra combination is another full decode of the original image (measured
+// at 60-120ms for a 1920x1080 backdrop, up to ~800ms for a large poster) plus
+// another cache file, so do not invent sizes at the call site — pick a preset.
+// Keeping this list short is what stops a single page from triggering a dozen
+// different decodes for the same artwork.
+export const ARTWORK = {
+  /** 最小一档（160px）：卡片的模糊占位图，以及 30-40px 级别的列表缩略图。 */
+  posterTiny: { maxWidth: 160, quality: 60 },
+  /** 海报列表与卡片（2:3 容器，紧凑布局共用同一份缓存）。 */
+  posterCard: { maxWidth: 480, maxHeight: 600, quality: 80 },
+  /** 海报详情页、剧集详情页头部与首页大图。 */
+  posterDetail: { maxWidth: 560, maxHeight: 840, quality: 80 },
+  /** 剧照大图：首页 hero、影片详情页背景。 */
+  backdropHero: { maxWidth: 1920, maxHeight: 1080, quality: 80 },
+  /** 剧照小图：媒体库封面、剧集条目、搜索结果等。 */
+  backdropStrip: { maxWidth: 480, maxHeight: 320, quality: 80 },
+}
+
 export function imageURL(remote?: string, version?: string, options: ImageURLOptions = false): string {
   if (!remote) return ''
   const versionQuery = version ? `v=${encodeURIComponent(version)}` : ''

web/src/components/Layout.tsx

@@ -89,6 +89,7 @@ export function Layout() {
         )}
         <LayoutWorkspace
           routeKey={location.pathname}
+          scrollKey={`${location.pathname}${location.search}`}
           userKey={user?.id}
           showMobileBottomNav={showMobileBottomNav}
         />

web/src/components/LayoutHeaderSections.tsx

@@ -3,7 +3,7 @@ import { createPortal } from 'react-dom'
 import { Link, useLocation, useNavigate } from 'react-router-dom'
 import { ArrowLeft, Film, LoaderCircle, Menu, Search, Star, X } from 'lucide-react'
 
-import { imageURL } from '../api/client'
+import { ARTWORK, imageURL } from '../api/client'
 import { mediaAPI } from '../api/library'
 import type { Media, PlayProfile, User } from '../types'
 import { favouriteMediaLink } from '../utils/mediaNavigation'
@@ -368,7 +368,7 @@ function LayoutHeaderSearch() {
                     <div className="relative h-12 w-9 shrink-0 overflow-hidden rounded-lg bg-[var(--app-panel-soft)]">
                       {item.poster_url ? (
                         <img
-                          src={imageURL(item.poster_url, item.updated_at)}
+                          src={imageURL(item.poster_url, item.updated_at, ARTWORK.posterCard)}
                           alt=""
                           className="h-full w-full object-cover"
                           loading="lazy"

web/src/components/LayoutSections.tsx

@@ -31,6 +31,8 @@ type LayoutSidebarsProps = Omit<
 
 type LayoutWorkspaceProps = {
   routeKey: string
+  /** 滚动记忆的 key：含 query，让同一路径下的不同视图（如剧集面板）各记一份位置。 */
+  scrollKey?: string
   userKey?: string
   showMobileBottomNav?: boolean
 }
@@ -113,10 +115,11 @@ export function LayoutSidebars({
 
 export function LayoutWorkspace({
   routeKey,
+  scrollKey,
   userKey = 'anonymous',
   showMobileBottomNav = false,
 }: LayoutWorkspaceProps) {
-  useScrollMemory(routeKey, userKey)
+  useScrollMemory(scrollKey ?? routeKey, userKey)
 
   const bottomPad = showMobileBottomNav
     ? 'pb-[calc(3.75rem+env(safe-area-inset-bottom,0px))] lg:pb-10'

web/src/components/ManualScrapeDialogSections.tsx

@@ -1,6 +1,6 @@
 import { Check, LoaderCircle, Search, Sparkles, X } from 'lucide-react'
 
-import { imageURL } from '../api/client'
+import { ARTWORK, imageURL } from '../api/client'
 import type { ManualScrapeCandidate } from '../api/library'
 import {
   candidateIDText,
@@ -202,7 +202,7 @@ function ManualScrapeCandidateRow({
     <div className="flex min-w-0 max-w-full flex-col gap-4 overflow-hidden rounded-xl border border-sand-200 bg-white p-3 shadow-sm sm:flex-row">
       <div className="h-28 w-20 shrink-0 overflow-hidden rounded-lg bg-sand-100">
         {item.poster_url ? (
-          <img src={imageURL(item.poster_url)} alt={item.title} loading="lazy" decoding="async" className="h-full w-full object-cover" referrerPolicy="no-referrer" />
+          <img src={imageURL(item.poster_url, undefined, ARTWORK.posterCard)} alt={item.title} loading="lazy" decoding="async" className="h-full w-full object-cover" referrerPolicy="no-referrer" />
         ) : (
           <div className="flex h-full items-center justify-center text-xs text-sand-500">无海报</div>
         )}

web/src/components/MediaCard.tsx

@@ -1,7 +1,7 @@
 import { memo, useEffect, useRef, useState, type ReactNode } from 'react'
 import { Link } from 'react-router-dom'
 import { Film, Play, Layers, Star } from 'lucide-react'
-import { imageURL } from '../api/client'
+import { ARTWORK, imageURL } from '../api/client'
 import type { Media } from '../types'
 
 const ACTION_OVERLAY_CLASS =
@@ -10,13 +10,14 @@ const ACTION_OVERLAY_CLASS =
 // memo：父级状态变化（如轮播切图、其它卡片操作）不再级联重渲染所有卡片。
 // 注意 actions/renderActions 必须引用稳定（用 renderActions 传函数）memo 才生效。
 export const MediaCard = memo(function MediaCard({
-  media, progress, count, rating, linkTo, onClick, actions, renderActions, compact,
+  media, progress, count, rating, linkTo, linkState, onClick, actions, renderActions, compact,
 }: {
   media: Media
   progress?: number
   count?: number
   rating?: number
   linkTo?: string
+  linkState?: { from?: string }
   onClick?: () => void
   actions?: ReactNode
   renderActions?: (media: Media) => ReactNode
@@ -25,12 +26,9 @@ export const MediaCard = memo(function MediaCard({
   const ref = useRef<HTMLDivElement>(null)
   const href = linkTo ?? `/media/${media.id}`
   const [posterFit, setPosterFit] = useState<'cover' | 'contain'>('cover')
-  const posterSrc = imageURL(media.poster_url, media.updated_at, {
-    maxWidth: compact ? 320 : 480,
-    maxHeight: compact ? 480 : 600,
-    quality: 82,
-  })
-  const blurredPosterSrc = imageURL(media.poster_url, media.updated_at, { maxWidth: 160, quality: 60 })
+  // 固定档位而不是在调用点拼尺寸：紧凑布局与普通卡片共用同一份缩略图缓存。
+  const posterSrc = imageURL(media.poster_url, media.updated_at, ARTWORK.posterCard)
+  const blurredPosterSrc = imageURL(media.poster_url, media.updated_at, ARTWORK.posterTiny)
   const displayRating = rating ?? media.rating
   const versionCount = media.versions?.length ?? 0
   // renderActions 延迟到卡片自身渲染时才调用，保证 memo 生效
@@ -82,14 +80,6 @@ export const MediaCard = memo(function MediaCard({
             </div>
           )}
 
-          {/* Episode count badge：统一放右下角（桌面端与移动端一致），与左上角评分彻底错开 */}
-          {(count !== undefined && count > 1) || (count === undefined && versionCount > 1) ? (
-            <span className="absolute bottom-2 right-2 inline-flex items-center gap-1 rounded-lg border border-white/15 bg-[#111827]/90 px-1.5 py-0.5 text-[10px] font-bold text-white shadow-sm">
-              <Layers size={10} className="shrink-0 text-[#c9954a]" />
-              <span>{count !== undefined && count > 1 ? `${count} 集` : `${versionCount} 版本`}</span>
-            </span>
-          ) : null}
-
           {/* Rating Badge */}
           {displayRating > 0 && (
             <span className="absolute left-2 top-2 inline-flex items-center gap-0.5 rounded-lg border border-white/15 bg-[#111827]/90 px-1.5 py-0.5 text-[10px] font-bold text-[#c9954a] shadow-sm">
@@ -137,13 +127,23 @@ export const MediaCard = memo(function MediaCard({
           }`}>
             {media.title}
           </p>
-          <div className="flex items-center justify-between text-[11px] font-bold uppercase tracking-wider text-[var(--app-muted)]">
+          <div className="flex items-center justify-between gap-2 text-[11px] font-bold uppercase tracking-wider text-[var(--app-muted)]">
             <span>{media.year > 0 ? media.year : "未知年份"}</span>
-            {media.video_codec && (
-              <span className="rounded-xl border border-[var(--app-border)] bg-[var(--app-panel-soft)] px-1.5 py-0.5 text-[var(--app-subtle)]">
-                {media.video_codec}
-              </span>
-            )}
+            <span className="flex items-center gap-1.5">
+              {/* 集数/版本数徽标：放在海报下方的信息行而不是压在画面上。海报底部通常
+                  自带片名，角标会盖住它（Game of Thrones / Breaking Bad 都很明显）。 */}
+              {(count !== undefined && count > 1) || (count === undefined && versionCount > 1) ? (
+                <span className="inline-flex items-center gap-1 rounded-xl border border-[var(--app-border)] bg-[var(--app-panel-soft)] px-1.5 py-0.5 text-[var(--app-subtle)]">
+                  <Layers size={10} className="shrink-0 text-[#c9954a]" />
+                  <span>{count !== undefined && count > 1 ? `${count} 集` : `${versionCount} 版本`}</span>
+                </span>
+              ) : null}
+              {media.video_codec && (
+                <span className="rounded-xl border border-[var(--app-border)] bg-[var(--app-panel-soft)] px-1.5 py-0.5 text-[var(--app-subtle)]">
+                  {media.video_codec}
+                </span>
+              )}
+            </span>
           </div>
         </div>
       </div>
@@ -167,7 +167,7 @@ export const MediaCard = memo(function MediaCard({
   if (actionContent) {
     return (
       <div className="group relative block">
-        <Link to={href} className="block">
+        <Link to={href} state={linkState} className="block">
           {card}
         </Link>
         <div className={ACTION_OVERLAY_CLASS}>
@@ -178,7 +178,7 @@ export const MediaCard = memo(function MediaCard({
   }
 
   return (
-    <Link to={href} className="group block">
+    <Link to={href} state={linkState} className="group block">
         {card}
     </Link>
   )

web/src/components/PlayerControls.tsx

@@ -4,6 +4,7 @@ import {
   Captions,
   CaptionsOff,
   Check,
+  ChevronDown,
   FastForward,
   Layers,
   ListVideo,
@@ -46,6 +47,9 @@ import {
   PLAYER_MENU_LABEL,
   PLAYER_POPOVER,
   PLAYER_RANGE_OVERLAY,
+  PLAYER_SHEET,
+  PLAYER_SHEET_BODY,
+  PLAYER_SHEET_HEADER,
   PLAYER_TEXT_BUTTON,
   PLAYER_TRACK,
   PLAYER_TRACK_FILL,
@@ -193,10 +197,17 @@ type PlayerControlsProps = {
   streamOffset?: number
   /** Absolute seek on the full timeline; return true when handled (e.g. HLS restart). */
   onSeekAbsolute?: (seconds: number) => boolean
+  /**
+   * 竖屏剧场模式：控制栏按钮放大到触控尺寸（44px 起），清晰度/倍速/
+   * 选集/字幕弹层改为视频区底部的动作面板，避免右侧抽屉与悬浮小弹层
+   * 在窄屏下被裁剪。桌面端保持原样。
+   */
+  theater?: boolean
 }
 
 export function PlayerControls({
   videoRef,
+  theater = false,
   volume: volumeProp = 1,
   onVolumeChange,
   onVolumeCommit,
@@ -271,6 +282,8 @@ export function PlayerControls({
   const speedMenuRef = useRef<HTMLDivElement | null>(null)
   const [settingsMenuOpen, setSettingsMenuOpen] = useState(false)
   const settingsMenuRef = useRef<HTMLDivElement | null>(null)
+  /** 竖屏剧场模式的设置面板（相对整个屏幕固定）——点击面板内部不应把它关掉。 */
+  const settingsSheetRef = useRef<HTMLDivElement | null>(null)
   const hideTimerRef = useRef<ReturnType<typeof setTimeout> | null>(null)
   const isScrubbingRef = useRef(false)
   /** 控制栏当前是否处于「必须保持显示」的状态（悬停/拖进度/菜单或面板打开）。 */
@@ -382,6 +395,10 @@ export function PlayerControls({
   useEffect(() => {
     if (!subtitleMenuOpen && !qualityMenuOpen && !speedMenuOpen && !settingsMenuOpen) return
     const onDocClick = (e: MouseEvent) => {
+      // 剧场模式的设置面板挂在操作栏之外，先判断：点面板内部不关任何弹层。
+      if (settingsSheetRef.current && settingsSheetRef.current.contains(e.target as Node)) {
+        return
+      }
       if (subtitleMenuRef.current && !subtitleMenuRef.current.contains(e.target as Node)) {
         setSubtitleMenuOpen(false)
       }
@@ -910,7 +927,7 @@ export function PlayerControls({
   const compactOnlySections = (
     <>
       {showQuality && hasQualityGroups && (
-        <div className="sm:hidden">
+        <div className={theater ? undefined : 'sm:hidden'}>
           <p className={PLAYER_MENU_LABEL}>清晰度</p>
           <QualityItems
             groups={qualityGroups}
@@ -919,7 +936,7 @@ export function PlayerControls({
           />
         </div>
       )}
-      <div className="sm:hidden">
+      <div className={theater ? undefined : 'sm:hidden'}>
         <p className={PLAYER_MENU_LABEL}>播放速度</p>
         <SpeedItems
           playbackRate={playbackRateProp}
@@ -927,7 +944,7 @@ export function PlayerControls({
         />
       </div>
       {onTogglePlaylist && (
-        <div className="sm:hidden">
+        <div className={theater ? undefined : 'sm:hidden'}>
           <p className={PLAYER_MENU_LABEL}>选集</p>
           <button
             type="button"
@@ -947,7 +964,7 @@ export function PlayerControls({
         </div>
       )}
       {subs.length > 0 && (
-        <div className="sm:hidden">
+        <div className={theater ? undefined : 'sm:hidden'}>
           <p className={PLAYER_MENU_LABEL}>字幕</p>
           <SubtitleItems
             subs={subs}
@@ -965,7 +982,7 @@ export function PlayerControls({
           />
         </div>
       )}
-      <div className={`sm:hidden ${PLAYER_MENU_DIVIDER}`} />
+      <div className={theater ? PLAYER_MENU_DIVIDER : `sm:hidden ${PLAYER_MENU_DIVIDER}`} />
     </>
   )
 
@@ -1028,6 +1045,28 @@ export function PlayerControls({
     <>
     {seekOverlay}
     {playbackRateOverlay}
+    {theater && settingsMenuOpen ? (
+      <div
+        ref={settingsSheetRef}
+        className={PLAYER_SHEET}
+        onClick={(e) => e.stopPropagation()}
+        onPointerDown={(e) => e.stopPropagation()}
+        onPointerUp={(e) => e.stopPropagation()}
+      >
+        <div className={PLAYER_SHEET_HEADER}>
+          <span className="text-[13px] font-semibold text-white/90">播放器设置</span>
+          <button
+            type="button"
+            onClick={() => setSettingsMenuOpen(false)}
+            className={`${PLAYER_ICON_BUTTON} h-10 w-10`}
+            title="关闭设置"
+          >
+            <ChevronDown size={20} />
+          </button>
+        </div>
+        <div className={PLAYER_SHEET_BODY}>{settingsPanel}</div>
+      </div>
+    ) : null}
     <div
       className={`pointer-events-auto absolute inset-x-0 bottom-0 z-20 bg-gradient-to-t from-black/85 via-black/45 to-transparent px-2.5 pb-1.5 pt-12 transition-opacity duration-300 sm:px-3.5 sm:pb-2.5 ${
         uiVisible ? 'opacity-100' : 'pointer-events-none opacity-0'
@@ -1042,11 +1081,11 @@ export function PlayerControls({
     >
       {/* 进度条：单独占一行贴着画面底部，轨道只有 3px，悬停才变粗并露出圆点手柄 */}
       <div
-        className="group/seek relative flex h-5 w-full items-center"
+        className={`group/seek relative flex w-full items-center ${theater ? 'h-9' : 'h-5'}`}
         onMouseMove={handleSeekHover}
         onMouseLeave={() => setSeekHover(null)}
       >
-        <div className={`${PLAYER_TRACK} group-hover/seek:h-[5px]`}>
+        <div className={`${PLAYER_TRACK} ${theater ? 'h-[5px]' : ''} group-hover/seek:h-[5px]`}>
           <div
             className={`${PLAYER_TRACK_FILL} bg-white/30`}
             style={{ width: `${bufferedRatio * 100}%` }}
@@ -1057,7 +1096,7 @@ export function PlayerControls({
           />
         </div>
         <div
-          className="pointer-events-none absolute h-3 w-3 -translate-x-1/2 rounded-full bg-white opacity-0 shadow-md transition-opacity duration-150 group-hover/seek:opacity-100"
+          className={`pointer-events-none absolute h-3 w-3 -translate-x-1/2 rounded-full bg-white shadow-md transition-opacity duration-150 group-hover/seek:opacity-100 ${theater ? 'opacity-100' : 'opacity-0'}`}
           style={{ left: `${playedRatio * 100}%` }}
         />
         {tooltipVisible && durationSafe > 0 ? (
@@ -1085,14 +1124,14 @@ export function PlayerControls({
         />
       </div>
 
-      <div className="mt-0.5 flex items-center gap-0.5 sm:mt-1 sm:gap-1">
+      <div className={`flex items-center ${theater ? 'mt-1 gap-1' : 'mt-0.5 gap-0.5 sm:mt-1 sm:gap-1'}`}>
         {/* 播放 / 暂停 */}
         <button
           onClick={togglePlay}
-          className={`${PLAYER_ICON_BUTTON} h-8 w-8 sm:h-9 sm:w-9`}
+          className={`${PLAYER_ICON_BUTTON} ${theater ? 'h-11 w-11' : 'h-8 w-8 sm:h-9 sm:w-9'}`}
           title={playing ? '暂停 (Space)' : '播放 (Space)'}
         >
-          {playing ? <Pause size={19} /> : <Play size={19} />}
+          {playing ? <Pause size={theater ? 24 : 19} /> : <Play size={theater ? 24 : 19} />}
         </button>
 
         {/* 上一集 / 下一集：单条媒体（电影）没有上下集时整组隐藏，少两个死按钮 */}
@@ -1101,7 +1140,7 @@ export function PlayerControls({
             <button
               onClick={onPrevEpisode}
               disabled={!hasPrevEpisode}
-              className={PLAYER_ICON_BUTTON}
+              className={`${PLAYER_ICON_BUTTON} ${theater ? 'h-11 w-11' : ''}`}
               title={
                 hasPrevEpisode
                   ? prevEpisodeTitle
@@ -1115,7 +1154,7 @@ export function PlayerControls({
             <button
               onClick={onNextEpisode}
               disabled={!hasNextEpisode}
-              className={PLAYER_ICON_BUTTON}
+              className={`${PLAYER_ICON_BUTTON} ${theater ? 'h-11 w-11' : ''}`}
               title={
                 hasNextEpisode
                   ? nextEpisodeTitle
@@ -1223,7 +1262,7 @@ export function PlayerControls({
           type="button"
           onClick={onOpenDanmaku}
           disabled={!onOpenDanmaku}
-          className={`${PLAYER_ICON_BUTTON} ${danmakuOpen ? 'bg-white/15' : ''}`}
+          className={`${PLAYER_ICON_BUTTON} ${theater ? 'h-11 w-11' : ''} ${danmakuOpen ? 'bg-white/15' : ''}`}
           title={danmakuOpen ? '弹幕设置（已打开）' : '弹幕设置'}
         >
           <span
@@ -1330,16 +1369,16 @@ export function PlayerControls({
               setSubtitleMenuOpen(false)
               setSettingsMenuOpen((v) => !v)
             }}
-            className={`${PLAYER_ICON_BUTTON} ${settingsMenuOpen ? 'bg-white/15 text-white' : ''}`}
+            className={`${PLAYER_ICON_BUTTON} ${theater ? 'h-11 w-11' : ''} ${settingsMenuOpen ? 'bg-white/15 text-white' : ''}`}
             title="播放器设置"
           >
-            <Settings2 size={17} />
+            <Settings2 size={theater ? 21 : 17} />
           </button>
           {/* 识别到 VR 素材但还没进 VR 时，用一个小圆点提示设置里有东西可开 */}
           {vr360Detected && !vr360 && (
             <span className="pointer-events-none absolute right-0.5 top-0.5 h-1.5 w-1.5 rounded-full bg-amber-300" />
           )}
-          {settingsMenuOpen && (
+          {settingsMenuOpen && !theater && (
             <div className={`${PLAYER_POPOVER} absolute bottom-11 right-0 max-h-[70vh] w-[15rem] overflow-y-auto`}>
               {settingsPanel}
             </div>
@@ -1358,10 +1397,10 @@ export function PlayerControls({
 
         <button
           onClick={toggleFullscreen}
-          className={PLAYER_ICON_BUTTON}
+          className={`${PLAYER_ICON_BUTTON} ${theater ? 'h-11 w-11' : ''}`}
           title={fullscreen ? '退出全屏 (F)' : '全屏 (F)'}
         >
-          {fullscreen ? <Minimize size={17} /> : <Maximize size={17} />}
+          {fullscreen ? <Minimize size={theater ? 21 : 17} /> : <Maximize size={theater ? 21 : 17} />}
         </button>
       </div>
     </div>

web/src/components/PlayerDanmakuPanel.tsx

@@ -2,7 +2,7 @@ import { useEffect, useState } from 'react'
 import { Check, ChevronDown, ChevronRight, Eye, EyeOff, Film, Hash, KeyRound, Loader2, MessageSquareText, RefreshCw, Search, Server, Settings2, Sparkles, Tag, X } from 'lucide-react'
 
 import type { DanmakuAnime, DanmakuEpisode, DanmakuLoadedInfo } from '../api/danmaku'
-import { PLAYER_DRAWER, PLAYER_ICON_BUTTON, PLAYER_PANEL_HEADER } from './playerTheme'
+import { PLAYER_DRAWER, PLAYER_ICON_BUTTON, PLAYER_PANEL_HEADER, PLAYER_SHEET, PLAYER_SHEET_BODY, PLAYER_SHEET_HEADER } from './playerTheme'
 
 // PlayerDanmakuPanel — the on-player danmaku control panel. It displays
 // the matched danmaku details (anime title, episode title, comment count,
@@ -56,6 +56,11 @@ type PlayerDanmakuPanelProps = {
   danmakuInfo?: DanmakuLoadedInfo | null
   onSelectEpisode: (episodeId: number, animeTitle: string, episodeTitle: string) => void
   onResetAuto: () => void
+  /**
+   * 竖屏剧场模式：弹幕面板改为视频区底部的动作面板，而不是右侧抽屉，
+   * 窄屏下搜索框与滑杆不再被压扁。
+   */
+  theater?: boolean
 }
 
 export function PlayerDanmakuPanel({
@@ -90,6 +95,7 @@ export function PlayerDanmakuPanel({
   danmakuInfo,
   onSelectEpisode,
   onResetAuto,
+  theater = false,
 }: PlayerDanmakuPanelProps) {
   const [draft, setDraft] = useState(search)
   const [advancedOpen, setAdvancedOpen] = useState(false)
@@ -201,20 +207,26 @@ export function PlayerDanmakuPanel({
       onClick={(e) => e.stopPropagation()}
       onPointerDown={(e) => e.stopPropagation()}
       onPointerUp={(e) => e.stopPropagation()}
-      className={`absolute inset-y-0 right-0 z-30 ${PLAYER_DRAWER}`}
+      className={
+        theater ? PLAYER_SHEET : `absolute inset-y-0 right-0 z-30 ${PLAYER_DRAWER}`
+      }
     >
-      <div className={PLAYER_PANEL_HEADER}>
+      <div className={theater ? PLAYER_SHEET_HEADER : PLAYER_PANEL_HEADER}>
         <div className="flex min-w-0 items-center gap-2 text-sm font-semibold">
           <MessageSquareText size={16} className="shrink-0 text-rose-400" />
           <span className="truncate">弹幕设置</span>
           {settingsSaving && <Loader2 size={12} className="shrink-0 animate-spin text-rose-300" />}
         </div>
-        <button onClick={onClose} className={PLAYER_ICON_BUTTON} title="关闭 (Esc)">
-          <X size={16} />
+        <button
+          onClick={onClose}
+          className={`${PLAYER_ICON_BUTTON} ${theater ? 'h-10 w-10' : ''}`}
+          title="关闭 (Esc)"
+        >
+          <X size={theater ? 20 : 16} />
         </button>
       </div>
 
-      <div className="min-h-0 flex-1 overflow-y-auto overscroll-contain px-4 py-3.5">
+      <div className={theater ? PLAYER_SHEET_BODY : 'min-h-0 flex-1 overflow-y-auto overscroll-contain px-4 py-3.5'}>
       {/* 弹幕总开关。从操作栏的「弹」按钮直接进来就到了这里，所以这一项就是
           用户找的「开关弹幕」；文案跟旧的按钮提示保持一致，避免换个说法让人找不到。 */}
       <label className="mb-3 flex cursor-pointer items-center justify-between rounded-lg bg-white/5 px-2.5 py-2 text-sm transition hover:bg-white/10">

web/src/components/PlayerMobileTheaterInfo.tsx

@@ -0,0 +1,199 @@
+import { useEffect, useRef } from 'react'
+import type { ReactNode } from 'react'
+import { ChevronDown, Layers, ListVideo } from 'lucide-react'
+
+import type { Media } from '../types'
+import { mediaVersionLabel, mediaVersionMatches } from '../utils/mediaVersion'
+
+// PlayerMobileTheaterInfo — 竖屏剧场模式下视频下方的可滚动内容区。
+//
+// 竖屏手机不再用全屏居中黑边布局：视频贴顶按 16:9 自适应高度，
+// 下面这块区域接管标题、选集、版本切换和简介，信息密度与
+// 主流手机视频应用保持一致。桌面端与横屏手机不渲染它。
+//
+// 选集刻意放在这里而不是弹层里：从操作栏点「选集」只做「展开 + 滚到这里」，
+// 播放画面始终可见，不会像浮层那样盖住正在看的内容。
+
+type PlayerMobileTheaterInfoProps = {
+  media: Media | null
+  title: string
+  subtitle?: string
+  playbackModeLabel?: string
+  qualityLabel?: string
+  episodes: Media[]
+  currentMediaId: string
+  currentEpisodeIndex: number
+  currentVersions?: Media[]
+  onSelectEpisode: (media: Media) => void
+  onSelectVersion?: (media: Media) => void
+  /** 选集列表是否展开（展开后由 playlistPanel 接管，提供季/搜索/版本）。 */
+  playlistOpen: boolean
+  onTogglePlaylist: () => void
+  /** 展开状态下的选集内容（复用播放器的选集面板，内嵌渲染）。 */
+  playlistPanel?: ReactNode
+  /**
+   * 每次从操作栏点「选集」都会 +1：即使列表已经展开，也要把这块滚回视野内。
+   */
+  revealToken?: number
+}
+
+function episodeShortLabel(ep: Media): string {
+  if (ep.episode_num > 0) return `${ep.episode_num}`
+  const title = ep.episode_title?.trim() || ep.title?.trim() || ''
+  return title ? title.slice(0, 4) : '·'
+}
+
+function episodeFullLabel(ep: Media): string {
+  if (ep.episode_num > 0) {
+    const title = ep.episode_title?.trim()
+    return title && title !== ep.title?.trim() ? `第 ${ep.episode_num} 集 · ${title}` : `第 ${ep.episode_num} 集`
+  }
+  return ep.episode_title?.trim() || ep.title?.trim() || '未命名'
+}
+
+export function PlayerMobileTheaterInfo({
+  media,
+  title,
+  subtitle,
+  playbackModeLabel,
+  qualityLabel,
+  episodes,
+  currentMediaId,
+  currentEpisodeIndex,
+  currentVersions = [],
+  onSelectEpisode,
+  onSelectVersion,
+  playlistOpen,
+  onTogglePlaylist,
+  playlistPanel,
+  revealToken = 0,
+}: PlayerMobileTheaterInfoProps) {
+  const episodeSectionRef = useRef<HTMLElement | null>(null)
+  const hasEpisodeList = episodes.length > 1
+  const versionsToSwitch = currentVersions.length > 1 ? currentVersions : []
+  const overview = media?.overview?.trim() || ''
+
+  // 从操作栏点「选集」：把选集区滚回视野内，用户不用自己往上翻。
+  useEffect(() => {
+    if (!revealToken) return
+    episodeSectionRef.current?.scrollIntoView({ block: 'start', behavior: 'smooth' })
+  }, [revealToken])
+
+  return (
+    <div className="min-h-0 flex-1 overflow-y-auto overscroll-contain bg-black px-4 pb-8 pt-3 text-white">
+      <h1 className="text-[15px] font-semibold leading-snug text-white/95">{title || '加载中…'}</h1>
+      {subtitle ? <p className="mt-1 text-xs text-white/55">{subtitle}</p> : null}
+      {(playbackModeLabel || qualityLabel) && (
+        <div className="mt-2 flex flex-wrap items-center gap-1.5">
+          {qualityLabel ? (
+            <span className="rounded-md bg-white/10 px-2 py-0.5 text-[11px] font-medium text-white/80">
+              {qualityLabel}
+            </span>
+          ) : null}
+          {playbackModeLabel ? (
+            <span className="rounded-md bg-white/10 px-2 py-0.5 text-[11px] text-white/60">
+              {playbackModeLabel}
+            </span>
+          ) : null}
+        </div>
+      )}
+
+      {hasEpisodeList ? (
+        <section ref={episodeSectionRef} className="mt-4 scroll-mt-2">
+          <div className="mb-2 flex items-center justify-between">
+            <h2 className="flex items-center gap-1.5 text-[13px] font-semibold text-white/90">
+              <ListVideo size={14} className="text-rose-400" />
+              选集
+              <span className="font-mono text-[11px] font-normal text-white/45">
+                {currentEpisodeIndex >= 0
+                  ? `${currentEpisodeIndex + 1}/${episodes.length}`
+                  : `${episodes.length} 集`}
+              </span>
+            </h2>
+            <button
+              type="button"
+              onClick={onTogglePlaylist}
+              className="flex min-h-8 items-center gap-1 rounded-md px-2 text-[11px] text-white/60 transition hover:bg-white/10 hover:text-white"
+            >
+              {playlistOpen ? '收起' : '全部'}
+              <ChevronDown
+                size={13}
+                className={`transition-transform ${playlistOpen ? 'rotate-180' : ''}`}
+              />
+            </button>
+          </div>
+
+          {playlistOpen && playlistPanel ? (
+            playlistPanel
+          ) : (
+            <div className="scrollbar-hide -mx-4 flex gap-1.5 overflow-x-auto px-4 pb-1">
+              {episodes.map((ep) => {
+                const active = mediaVersionMatches(ep, currentMediaId)
+                return (
+                  <button
+                    key={ep.id}
+                    type="button"
+                    onClick={() => onSelectEpisode(ep)}
+                    title={episodeFullLabel(ep)}
+                    className={`flex h-10 min-w-12 shrink-0 items-center justify-center rounded-lg px-2 text-xs font-medium tabular-nums transition ${
+                      active
+                        ? 'bg-rose-500 text-white'
+                        : 'bg-white/[0.07] text-white/80 active:bg-white/15'
+                    }`}
+                  >
+                    {episodeShortLabel(ep)}
+                  </button>
+                )
+              })}
+            </div>
+          )}
+        </section>
+      ) : null}
+
+      {/* 多版本切换：展开选集后由选集面板顶部的版本区接管，这里不重复一份 */}
+      {versionsToSwitch.length > 0 && !playlistOpen ? (
+        <section className="mt-4">
+          <h2 className="mb-2 flex items-center gap-1.5 text-[13px] font-semibold text-white/90">
+            <Layers size={13} className="text-rose-400" />
+            版本
+            <span className="font-mono text-[11px] font-normal text-white/45">
+              {versionsToSwitch.length} 个
+            </span>
+          </h2>
+          <div className="space-y-1.5">
+            {versionsToSwitch.map((version) => {
+              const active = mediaVersionMatches(version, currentMediaId)
+              return (
+                <button
+                  key={version.id}
+                  type="button"
+                  disabled={active}
+                  onClick={() => onSelectVersion?.(version)}
+                  className={`flex w-full items-center gap-2 rounded-lg border px-2.5 py-2 text-left transition disabled:cursor-default ${
+                    active
+                      ? 'border-rose-500/50 bg-rose-500/15 text-white'
+                      : 'border-white/5 bg-white/5 text-white/85 active:bg-white/10'
+                  }`}
+                >
+                  <span className="min-w-0 flex-1 truncate text-xs font-medium">
+                    {mediaVersionLabel(version)}
+                  </span>
+                  {active ? <span className="shrink-0 text-[10px] text-rose-300">播放中</span> : null}
+                </button>
+              )
+            })}
+          </div>
+        </section>
+      ) : null}
+
+      {overview ? (
+        <section className="mt-4">
+          <h2 className="mb-1.5 text-[13px] font-semibold text-white/90">简介</h2>
+          <p className="text-xs leading-relaxed text-white/60">{overview}</p>
+        </section>
+      ) : null}
+
+      <p className="mt-5 text-center text-[10px] text-white/30">将手机横放可获得全屏沉浸式播放体验</p>
+    </div>
+  )
+}

web/src/components/PlayerPlaylistPanel.tsx

@@ -27,6 +27,11 @@ type PlayerPlaylistPanelProps = {
   onSelectEpisode: (media: Media) => void
   /** 切换到同一集/同一条目的另一个版本。 */
   onSelectVersion?: (media: Media) => void
+  /**
+   * 竖屏剧场模式：不再以浮层形式盖住播放画面，而是内嵌在视频下方的
+   * 内容区里（外层负责标题与「收起」按钮），选集不再遮挡播放器。
+   */
+  inline?: boolean
 }
 
 /**
@@ -50,6 +55,7 @@ export function PlayerPlaylistPanel({
   episodes,
   onSelectEpisode,
   onSelectVersion,
+  inline = false,
 }: PlayerPlaylistPanelProps) {
   const [filterText, setFilterText] = useState('')
   /** 正在挑选版本的剧集 id；null = 宫格直接播放。 */
@@ -133,19 +139,29 @@ export function PlayerPlaylistPanel({
       // 舞台用 pointerdown/pointerup 判断「移动端轻触画面」，抽屉上的触摸不该参与。
       onPointerDown={(e) => e.stopPropagation()}
       onPointerUp={(e) => e.stopPropagation()}
-      className={`absolute inset-y-0 right-0 z-30 ${PLAYER_DRAWER}`}
+      className={
+        inline
+          ? 'flex flex-col'
+          : `absolute inset-y-0 right-0 z-30 ${PLAYER_DRAWER}`
+      }
     >
-      <div className={PLAYER_PANEL_HEADER}>
-        <div className="flex min-w-0 items-center gap-2 text-sm font-semibold">
-          <span className="truncate">{hasEpisodeList ? '选集' : '版本'}</span>
-          <span className="shrink-0 font-mono text-[11px] font-normal text-white/45">
-            {hasEpisodeList ? `${episodes.length} 集` : `${versionsToSwitch.length} 个版本`}
-          </span>
+      {/* 内嵌在竖屏剧场布局的视频下方时不自带标题栏：外层已经提供了
+          「选集 / 收起」标题，重复一层只会挤掉内容。 */}
+      {inline ? null : (
+        <div className={PLAYER_PANEL_HEADER}>
+          <div className="flex min-w-0 items-center gap-2 text-sm font-semibold">
+            <span className="truncate">{hasEpisodeList ? '选集' : '版本'}</span>
+            <span className="shrink-0 font-mono text-[11px] font-normal text-white/45">
+              {hasEpisodeList ? `${episodes.length} 集` : `${versionsToSwitch.length} 个版本`}
+            </span>
+          </div>
+          <button onClick={onClose} className={PLAYER_ICON_BUTTON} title="关闭 (Esc)">
+            <X size={16} />
+          </button>
         </div>
-        <button onClick={onClose} className={PLAYER_ICON_BUTTON} title="关闭 (Esc)">
-          <X size={16} />
-        </button>
-      </div>
+      )}
+
+      <div className={inline ? undefined : 'flex min-h-0 flex-1 flex-col overflow-hidden'}>
 
       {/* 当前条目的版本切换：多版本时置顶，随时可换 */}
       {versionsToSwitch.length > 0 && (
@@ -206,7 +222,7 @@ export function PlayerPlaylistPanel({
 
       {/* 季选择 Tabs（若有多季） */}
       {hasEpisodeList && seasonGroups.length > 1 && (
-        <div className="no-scrollbar flex shrink-0 items-center gap-1.5 overflow-x-auto border-b border-white/10 px-3 py-2">
+        <div className="scrollbar-hide flex shrink-0 items-center gap-1.5 overflow-x-auto border-b border-white/10 px-3 py-2">
           {seasonGroups.map(({ season, episodes: sesEps }) => {
             const isSelected = selectedSeason === season
             const isPlayingThisSeason = sesEps.some((e) => mediaVersionMatches(e, currentMediaId))
@@ -255,15 +271,21 @@ export function PlayerPlaylistPanel({
         </div>
       )}
 
-      {/* 集数宫格 */}
+      {/* 集数宫格：内嵌模式整块跟着外层一起滚动，不再自己滚 */}
       {hasEpisodeList ? (
-        <div className="min-h-0 flex-1 overflow-y-auto p-3 select-none">
+        <div className={inline ? 'p-3 select-none' : 'min-h-0 flex-1 overflow-y-auto p-3 select-none'}>
           {filteredEpisodes.length === 0 ? (
             <div className="py-10 text-center text-xs text-white/40">
               {filterText ? '未找到匹配的剧集' : '暂无剧集列表'}
             </div>
           ) : (
-            <div className="grid grid-cols-5 gap-1.5 sm:grid-cols-6">
+            <div
+              className={
+                inline
+                  ? 'grid grid-cols-6 gap-1.5'
+                  : 'grid grid-cols-5 gap-1.5 sm:grid-cols-6'
+              }
+            >
               {filteredEpisodes.map((ep) => {
                 const isPlaying = mediaVersionMatches(ep, currentMediaId)
                 const versions = ep.versions && ep.versions.length > 1 ? ep.versions : []
@@ -285,7 +307,7 @@ export function PlayerPlaylistPanel({
                       ref={isPlaying ? activeItemRef : null}
                       onClick={() => onSelectEpisode(ep)}
                       title={cellTitle}
-                      className={`flex h-9 w-full items-center justify-center rounded-md text-xs font-medium tabular-nums transition ${
+                      className={`flex ${inline ? 'h-10' : 'h-9'} w-full items-center justify-center rounded-md text-xs font-medium tabular-nums transition ${
                         isPlaying
                           ? 'bg-rose-500 text-white shadow-[0_0_0_1px_rgba(255,255,255,0.2)_inset]'
                           : 'bg-white/[0.06] text-white/80 hover:bg-white/15 hover:text-white'
@@ -308,7 +330,7 @@ export function PlayerPlaylistPanel({
                             current === ep.id ? null : ep.id,
                           )
                         }}
-                        className={`absolute right-0 top-0 flex h-3.5 w-3.5 items-center justify-center rounded-bl-md rounded-tr-md text-[9px] font-semibold leading-none transition ${
+                        className={`absolute right-0 top-0 flex ${inline ? 'h-5 w-5 text-[10px]' : 'h-3.5 w-3.5 text-[9px]'} items-center justify-center rounded-bl-md rounded-tr-md font-semibold leading-none transition ${
                           versionsForEpisodeId === ep.id
                             ? 'bg-amber-300 text-black'
                             : 'bg-black/45 text-amber-200 hover:bg-black/70'
@@ -326,11 +348,12 @@ export function PlayerPlaylistPanel({
         </div>
       ) : (
         versionsToSwitch.length === 0 && (
-          <div className="flex flex-1 items-center justify-center px-4 text-center text-xs text-white/40">
+          <div className="flex items-center justify-center px-4 py-8 text-center text-xs text-white/40">
             暂无可切换的内容
           </div>
         )
       )}
+      </div>
     </div>
   )
 }

web/src/components/playerTheme.ts

@@ -28,6 +28,27 @@ export const PLAYER_POPOVER =
 export const PLAYER_DRAWER =
   `flex h-full w-[88%] flex-col border-white/10 ${PANEL_BG} text-white ${PANEL_SHADOW} backdrop-blur-xl sm:w-[380px] sm:border-l`
 
+/**
+ * 底部动作面板（竖屏剧场模式）：贴着整个屏幕底部向上弹出，高度按视口
+ * 计算（最高 86dvh），因此手机上能容纳的选项远多于贴视频区弹出的版本。
+ *
+ * 用 fixed 而不是 absolute：播放区在竖屏剧场模式里只有 16:9 一小块，
+ * 若以它为定位基准，面板最高只能到视频高度的 72%，内容会被压得很窄。
+ */
+export const PLAYER_SHEET =
+  `player-sheet-height fixed inset-x-0 bottom-0 z-40 flex min-h-[10rem] flex-col overflow-hidden rounded-t-2xl border border-b-0 border-white/10 ${PANEL_BG} text-white ${PANEL_SHADOW} backdrop-blur-xl`
+
+/** 底部动作面板的标题栏（带关闭按钮）。 */
+export const PLAYER_SHEET_HEADER =
+  'flex shrink-0 items-center justify-between gap-2 border-b border-white/10 px-4 py-2.5'
+
+/**
+ * 底部动作面板的可滚动内容区：底部垫上安全区高度避免 iPhone 小横条遮挡，
+ * 并把面板内所有按钮抬到 40px，手机上手指点得准。
+ */
+export const PLAYER_SHEET_BODY =
+  'min-h-0 flex-1 overflow-y-auto overscroll-contain px-3 pb-[calc(0.75rem+env(safe-area-inset-bottom,0px))] pt-2 [&_button]:min-h-10'
+
 /** 操作栏里的图标按钮。所有图标按钮共用同一尺寸与悬停反馈，排在一起才整齐。 */
 export const PLAYER_ICON_BUTTON =
   'flex h-7 w-7 shrink-0 items-center justify-center rounded-md text-white/85 transition hover:bg-white/15 hover:text-white disabled:cursor-not-allowed disabled:text-white/30 disabled:hover:bg-transparent sm:h-8 sm:w-8'

web/src/hooks/useIsMobileTheater.ts

@@ -0,0 +1,32 @@
+import { useEffect, useState } from 'react'
+
+const THEATER_MAX_WIDTH = 768
+
+function queryTheater(): boolean {
+  if (typeof window === 'undefined') return false
+  // 窄屏 + 竖屏才进剧场模式：横屏手机自动回到全屏沉浸式布局。
+  return window.innerWidth < THEATER_MAX_WIDTH && window.innerHeight >= window.innerWidth
+}
+
+/**
+ * 是否使用移动端竖屏剧场布局。
+ *
+ * 桌面端与横屏手机保持原来的全屏居中播放器；只有竖屏手机进
+ * 「视频贴顶 + 下方内容区」的剧场模式，避免 16:9 视频在竖屏里
+ * 上下大黑边、控制栏远离画面的问题。
+ */
+export function useIsMobileTheater(): boolean {
+  const [theater, setTheater] = useState(queryTheater)
+
+  useEffect(() => {
+    const onChange = () => setTheater(queryTheater())
+    window.addEventListener('resize', onChange)
+    window.addEventListener('orientationchange', onChange)
+    return () => {
+      window.removeEventListener('resize', onChange)
+      window.removeEventListener('orientationchange', onChange)
+    }
+  }, [])
+
+  return theater
+}

web/src/hooks/useScrollMemory.test.ts

@@ -1,4 +1,10 @@
-import { shouldPersistScrollSample, shouldRememberScroll } from './useScrollMemory.ts'
+import {
+  isJumpIntentKey,
+  isScrollIntentKey,
+  shouldPersistClampedSample,
+  shouldPersistScrollSample,
+  shouldRememberScroll,
+} from './useScrollMemory.ts'
 
 function check(name: string, condition: boolean) {
   if (!condition) throw new Error(`useScrollMemory: ${name}`)
@@ -7,6 +13,7 @@ function check(name: string, condition: boolean) {
 check('home remembers scroll', shouldRememberScroll('/'))
 check('library list remembers scroll', shouldRememberScroll('/libraries'))
 check('library detail remembers scroll', shouldRememberScroll('/library/lib-1'))
+check('library series panel remembers scroll', shouldRememberScroll('/library/lib-1'))
 check('media detail remembers scroll', shouldRememberScroll('/media/media-1'))
 check('settings does not remember scroll', !shouldRememberScroll('/settings'))
 check('player does not remember scroll', !shouldRememberScroll('/play/media-1'))
@@ -48,4 +55,48 @@ check(
   }),
 )
 
+// 恢复途中内容还没长出来（滚不动）或被钳在矮内容底部时，scrollTop 是浏览器钳制
+// 出来的值，不能写回存储，否则一次误触就把记忆位置清成 0。
+check(
+  'clamped sample while list is still loading is ignored',
+  !shouldPersistClampedSample({ current: 0, maxScroll: 0, saved: 1800 }),
+)
+check(
+  'clamped sample at the bottom of a shrunken list is ignored',
+  !shouldPersistClampedSample({ current: 1200, maxScroll: 1200, saved: 1800 }),
+)
+check(
+  'sample at the bottom of a completely scrolled list stays persistable',
+  shouldPersistClampedSample({ current: 1800, maxScroll: 1800, saved: 1800 }),
+)
+check(
+  'reachable scroll position is persisted',
+  shouldPersistClampedSample({ current: 900, maxScroll: 1200, saved: 1800 }),
+)
+check(
+  'reaching the remembered target is persisted',
+  shouldPersistClampedSample({ current: 1800, maxScroll: 2600, saved: 1800 }),
+)
+check(
+  'deeper than the remembered target is persisted',
+  shouldPersistClampedSample({ current: 3000, maxScroll: 4000, saved: 1800 }),
+)
+check(
+  'intentional scroll to top of a tall list is persisted',
+  shouldPersistClampedSample({ current: 0, maxScroll: 4000, saved: 1800 }),
+)
+check(
+  'unscrollable page without memory cannot persist',
+  !shouldPersistClampedSample({ current: 0, maxScroll: 0, saved: 0 }),
+)
+
+check('page down counts as scroll intent', isScrollIntentKey('PageDown'))
+check('space counts as scroll intent', isScrollIntentKey(' '))
+check('typing does not count as scroll intent', !isScrollIntentKey('a'))
+check('shortcut does not count as scroll intent', !isScrollIntentKey('Meta'))
+check('page down allows a long jump', isJumpIntentKey('PageDown'))
+check('home allows a long jump', isJumpIntentKey('Home'))
+check('arrow key keeps smooth-only intent', !isJumpIntentKey('ArrowDown'))
+check('space keeps smooth-only intent', !isJumpIntentKey(' '))
+
 console.log('useScrollMemory.test.ts ok')

web/src/hooks/useScrollMemory.ts

@@ -3,6 +3,36 @@ import { useLayoutEffect } from 'react'
 const SCROLL_STORAGE_PREFIX = 'mebox.scroll.'
 const scrollPositions = new Map<string, number>()
 
+/**
+ * 用户滚动意图的有效期：滚轮 / 触摸 / 翻页键之后这段时间内的 scroll 事件才算
+ * “用户自己滚出来的位置”。连续滚动（含惯性）期间每个 scroll 事件都会续期，
+ * 所以正常滑动不会中途过期，而恢复回填 / 虚拟列表量高这类程序化位移会被排除。
+ */
+const USER_SCROLL_INTENT_TTL_MS = 800
+
+/** 平滑滚动（滚轮/触摸/方向键）单次采样允许的最大“向上跳”距离。 */
+function maxSmoothJumpUp(viewportHeight: number): number {
+  return Math.max(viewportHeight * 2, 1200)
+}
+
+/** 会滚动容器的按键。 */
+export function isScrollIntentKey(key: string): boolean {
+  return (
+    key === 'PageDown' ||
+    key === 'PageUp' ||
+    key === 'Home' ||
+    key === 'End' ||
+    key === 'ArrowDown' ||
+    key === 'ArrowUp' ||
+    key === ' '
+  )
+}
+
+/** 一次按键可能直接跳很远（首页/末页/翻页），这些不算异常向上跳。 */
+export function isJumpIntentKey(key: string): boolean {
+  return key === 'PageDown' || key === 'PageUp' || key === 'Home' || key === 'End'
+}
+
 function readScrollPosition(key: string): number {
   const cached = scrollPositions.get(key)
   if (cached !== undefined) return cached
@@ -56,21 +86,54 @@ export function shouldPersistScrollSample(input: {
   return true
 }
 
+/**
+ * 判断一次滚动采样是不是“用户真正选中的位置”。
+ *
+ * 内容还在加载（容器根本滚不动）或被钳在“变矮内容”的底部时，scrollTop 是浏览器
+ * 钳制出来的值：恢复途中任何一次滚轮 / 触摸 / 按键都会读到这个值，写回存储就会把
+ * 记忆位置抹成 0（表现为返回媒体库时永远停在顶部，且之后再也不会记录）。
+ */
+export function shouldPersistClampedSample(input: {
+  current: number
+  maxScroll: number
+  saved: number
+}): boolean {
+  const { current, maxScroll, saved } = input
+  if (maxScroll <= 0) {
+    return false
+  }
+  if (current >= maxScroll && current < saved) {
+    return false
+  }
+  return true
+}
+
 /**
  * 记住列表页的滚动位置。页面内容会异步长高，因此恢复期间会监听内容高度，
  * 直到目标位置可达；期间用户主动滚动会立即接管，避免和恢复逻辑抢滚动条。
+ *
+ * 恢复还没完成时，容器里的 scrollTop 往往是浏览器钳制出来的值（内容还在加载
+ * 占位，或被钳在变矮内容的底部）。这种采样既不写回存储，也不当成用户“接管”，
+ * 否则一次误触的滚轮 / 触摸就会把记忆位置清成 0。
+ *
+ * 只有带“用户滚动意图”（滚轮 / 触摸 / 翻页键 / 拖滚动条）的 scroll 事件才会写
+ * 存储：恢复回填、虚拟列表挂载量高、浏览器钳制造成的位移都不算用户选择的位置。
+ *
+ * routeKey 允许包含 query：媒体库详情页的剧集面板（`/library/:id?series=...`）
+ * 与网格视图是两块不同内容，各自记一份位置，互不覆盖。
  */
-export function useScrollMemory(pathname: string, userKey = 'anonymous'): void {
+export function useScrollMemory(routeKey: string, userKey = 'anonymous'): void {
   useLayoutEffect(() => {
     const el = document.getElementById('app-main-scroll')
     if (!el) return
 
+    const pathname = routeKey.split('?')[0]
     if (!shouldRememberScroll(pathname)) {
       el.scrollTop = 0
       return
     }
 
-    const key = `${userKey}:${pathname}`
+    const key = `${userKey}:${routeKey}`
     const saved = readScrollPosition(key)
     let restoring = saved > 0
     let observer: ResizeObserver | null = null
@@ -79,6 +142,36 @@ export function useScrollMemory(pathname: string, userKey = 'anonymous'): void {
     let restorePumpUntil = 0
     let lastSaved = saved
     let lastHeight = el.scrollHeight
+    // 最近一次“用户主动滚动”的时间戳与类型（滚轮 / 触摸 / 翻页键 / 拖滚动条）。
+    // 只有用户自己滚出来的位置才写回存储：恢复逻辑、虚拟列表挂载量高、浏览器
+    // 钳制造成的 scrollTop 变化都不代表用户想要的位置，写进去就会把记忆抹掉。
+    let userScrollUntil = 0
+    let userScrollAllowsJump = false
+    const markUserScroll = (allowsJump = false) => {
+      const now = performance.now()
+      // 上一段意图已过期：新的一段从“只允许小幅向上跳”开始。
+      if (userScrollUntil <= now) {
+        userScrollAllowsJump = false
+      }
+      userScrollUntil = now + USER_SCROLL_INTENT_TTL_MS
+      if (allowsJump) {
+        userScrollAllowsJump = true
+      }
+    }
+    const userScrolling = () => userScrollUntil > performance.now()
+
+    const currentMaxScroll = () => Math.max(0, el.scrollHeight - el.clientHeight)
+
+    // 只有“用户真正能滚到”的位置才算有效采样：内容还在加载占位（滚不动）或
+    // 被钳在变矮内容的底部时，scrollTop 都是浏览器钳制出来的值。
+    const sampleIsTrustworthy = (current: number) =>
+      shouldPersistClampedSample({ current, maxScroll: currentMaxScroll(), saved })
+
+    const adoptScrollPosition = (current: number) => {
+      lastSaved = current
+      lastHeight = el.scrollHeight
+      writeScrollPosition(key, current)
+    }
 
     const stopRestore = () => {
       if (restoreFrame) {
@@ -97,14 +190,18 @@ export function useScrollMemory(pathname: string, userKey = 'anonymous'): void {
       if (!restoring) return
       restoring = false
       stopRestore()
-      lastSaved = Math.round(el.scrollTop)
-      lastHeight = el.scrollHeight
-      writeScrollPosition(key, lastSaved)
+      const current = Math.round(el.scrollTop)
+      if (!sampleIsTrustworthy(current)) {
+        // 目标一直不可达（内容仍比记忆位置矮 / 还在加载）：保留原记忆值，
+        // 不要把钳制出来的位置写回去。
+        return
+      }
+      adoptScrollPosition(current)
     }
 
     const tryRestore = () => {
       if (!restoring) return
-      const max = Math.max(0, el.scrollHeight - el.clientHeight)
+      const max = currentMaxScroll()
       const target = Math.min(saved, max)
       if (target > 0 && Math.abs(el.scrollTop - target) > 1) {
         el.scrollTop = target
@@ -125,8 +222,30 @@ export function useScrollMemory(pathname: string, userKey = 'anonymous'): void {
     }
 
     const saveNow = () => {
-      if (restoring) return
       const current = Math.round(el.scrollTop)
+      const userDriven = userScrolling()
+      // 惯性/连续滚动期间持续续期，避免长距离滑动中途被判定为“非用户滚动”。
+      if (userDriven) markUserScroll(userScrollAllowsJump)
+      // 内容还没长回来时读到的 scrollTop 不是用户选的位置，既不能写存储，
+      // 也不能当成“用户接管”。
+      if (!sampleIsTrustworthy(current)) return
+      // 一瞬间大幅向上跳（恢复回填、虚拟列表挂载量高、浏览器钳制）只有按键翻页 /
+      // 拖滚动条这类意图才可能是用户行为，否则直接丢弃。
+      const upJump = lastSaved > current ? lastSaved - current : 0
+      if (upJump > maxSmoothJumpUp(el.clientHeight) && !(userDriven && userScrollAllowsJump)) {
+        return
+      }
+      if (restoring) {
+        if (!userDriven || Math.abs(current - saved) <= 1) return
+        // 恢复途中用户真的滚到了别处：交还控制权并记录这个位置，避免恢复
+        // 逻辑继续和用户抢滚动条。
+        restoring = false
+        stopRestore()
+        adoptScrollPosition(current)
+        return
+      }
+      // 非用户滚动（恢复逻辑回填、虚拟列表挂载/量高、浏览器钳制）不写存储。
+      if (!userDriven) return
       const height = el.scrollHeight
       if (
         !shouldPersistScrollSample({
@@ -136,7 +255,6 @@ export function useScrollMemory(pathname: string, userKey = 'anonymous'): void {
           lastHeight,
         })
       ) {
-        lastHeight = height
         return
       }
       lastHeight = height
@@ -145,13 +263,13 @@ export function useScrollMemory(pathname: string, userKey = 'anonymous'): void {
       writeScrollPosition(key, current)
     }
 
-    const cancelRestore = () => {
-      if (!restoring) return
-      restoring = false
-      stopRestore()
-      lastSaved = Math.round(el.scrollTop)
-      lastHeight = el.scrollHeight
-      writeScrollPosition(key, lastSaved)
+    // 用户滚动意图：滚轮 / 触摸 = 平滑滚动；翻页键 / 拖滚动条 = 允许直接跳很远。
+    const onScrollIntent = () => {
+      markUserScroll(false)
+    }
+
+    const onJumpIntent = () => {
+      markUserScroll(true)
     }
 
     if (saved <= 0) {
@@ -171,42 +289,37 @@ export function useScrollMemory(pathname: string, userKey = 'anonymous'): void {
     }
 
     const onPointerDown = (event: PointerEvent) => {
+      // 点在容器本身 = 拖滚动条或在空白处按下：允许大幅跳转。
       if (event.target === el) {
-        cancelRestore()
+        onJumpIntent()
       }
       saveNow()
     }
 
     const onKeyDown = (event: KeyboardEvent) => {
-      if (
-        event.key === 'PageDown' ||
-        event.key === 'PageUp' ||
-        event.key === 'Home' ||
-        event.key === 'End' ||
-        event.key === 'ArrowDown' ||
-        event.key === 'ArrowUp' ||
-        event.key === ' '
-      ) {
-        cancelRestore()
+      if (isJumpIntentKey(event.key)) {
+        onJumpIntent()
+      } else if (isScrollIntentKey(event.key)) {
+        onScrollIntent()
       }
       saveNow()
     }
 
     el.addEventListener('scroll', saveNow, { passive: true })
     el.addEventListener('pointerdown', onPointerDown, true)
-    window.addEventListener('wheel', cancelRestore, { passive: true, capture: true })
-    window.addEventListener('touchstart', cancelRestore, { passive: true, capture: true })
+    window.addEventListener('wheel', onScrollIntent, { passive: true, capture: true })
+    window.addEventListener('touchstart', onScrollIntent, { passive: true, capture: true })
     window.addEventListener('keydown', onKeyDown, true)
 
     return () => {
       el.removeEventListener('scroll', saveNow)
       el.removeEventListener('pointerdown', onPointerDown, true)
-      window.removeEventListener('wheel', cancelRestore, true)
-      window.removeEventListener('touchstart', cancelRestore, true)
+      window.removeEventListener('wheel', onScrollIntent, true)
+      window.removeEventListener('touchstart', onScrollIntent, true)
       window.removeEventListener('keydown', onKeyDown, true)
       stopRestore()
       // 清理时写入最后一次有效位置，避免依赖已被钳制的 el.scrollTop。
       writeScrollPosition(key, lastSaved)
     }
-  }, [pathname, userKey])
+  }, [routeKey, userKey])
 }

web/src/index.css

@@ -272,6 +272,14 @@ body {
 }
 .scrollbar-hide::-webkit-scrollbar { display: none; }
 
+/* ── 播放器底部动作面板的高度上限 ──
+   手机上按视口高度算（86%），老浏览器不认识 dvh 时退回 vh，避免面板
+   在 flex 容器里被内容撑出屏幕。 */
+.player-sheet-height {
+  max-height: 86vh;
+  max-height: 86dvh;
+}
+
 /* ─── Reduced Motion ─── */
 @media (prefers-reduced-motion: reduce) {
   *, *::before, *::after {

web/src/pages/AdminLibraryTable.tsx

@@ -2,7 +2,7 @@ import { useEffect, useRef, useState, type DragEvent, type MouseEvent, type Reac
 import { createPortal } from 'react-dom'
 import { Folder, GripVertical, Image, MoreVertical, Plus, Power, PowerOff, RefreshCw, Save, Trash2 } from 'lucide-react'
 
-import { imageURL } from '../api/client'
+import { ARTWORK, imageURL } from '../api/client'
 import { LocalDirBrowserDialog } from '../components/LocalDirBrowserDialog'
 import type { Library, LibraryRoot } from '../types'
 import type { RootDraft } from './adminLibraryPanelModel'
@@ -172,7 +172,7 @@ function LibraryTableRow({ library, dragging, dragOver, onDragStart, onDragOver,
         <div className="flex items-center gap-2">
           {library.cover_url && (
             <img
-              src={imageURL(library.cover_url, library.updated_at)}
+              src={imageURL(library.cover_url, library.updated_at, ARTWORK.backdropStrip)}
               alt=""
               loading="lazy"
               decoding="async"

web/src/pages/AdminUsersTable.tsx

@@ -46,13 +46,13 @@ export function AdminUsersTable({
       <table className="min-w-[900px] w-full text-left text-sm">
         <thead className="text-xs uppercase tracking-wider text-sand-500">
           <tr>
-            <th className="py-2">用户名</th>
-            <th>角色</th>
-            <th>媒体库权限</th>
-            <th>状态</th>
+            <th className="py-2 whitespace-nowrap">用户名</th>
+            <th className="whitespace-nowrap">角色</th>
+            <th className="whitespace-nowrap">媒体库权限</th>
+            <th className="whitespace-nowrap">状态</th>
             <th>权限说明</th>
-            <th>最近登录</th>
-            <th className="text-right">操作</th>
+            <th className="whitespace-nowrap">最近登录</th>
+            <th className="whitespace-nowrap text-right">操作</th>
           </tr>
         </thead>
         <tbody>
@@ -72,8 +72,8 @@ export function AdminUsersTable({
                   </span>
                 )}
               </td>
-              <td className="text-ink-100">{u.role === 'admin' ? '管理员' : '观看用户'}</td>
-              <td>
+              <td className="text-ink-100 whitespace-nowrap">{u.role === 'admin' ? '管理员' : '观看用户'}</td>
+              <td className="whitespace-nowrap">
                 {u.role === 'admin' ? (
                   <span className="inline-flex items-center rounded-full bg-sand-100 px-2.5 py-0.5 text-xs text-sand-600 font-medium">
                     全库 (管理员)
@@ -94,13 +94,13 @@ export function AdminUsersTable({
                   </button>
                 )}
               </td>
-              <td className={u.is_active ? 'text-green-500' : 'text-red-400'}>
+              <td className={`whitespace-nowrap ${u.is_active ? 'text-green-500' : 'text-red-400'}`}>
                 {u.is_active ? '正常' : '已禁用'}
               </td>
-              <td className="text-ink-50">
+              <td className="text-ink-50 max-w-[22rem]">
                 {u.role === 'admin' ? '全部管理权限' : '仅浏览/播放/外部播放器，无下载与文件操作'}
               </td>
-              <td className="text-ink-50">
+              <td className="text-ink-50 whitespace-nowrap">
                 <span className="inline-flex flex-wrap items-center gap-2">
                   <span>{u.last_login_at ? new Date(u.last_login_at).toLocaleString() : '从未登录'}</span>
                   {u.realtime_online && <span className="rounded border border-green-400/40 px-1.5 py-0.5 text-[11px] text-green-500">在线</span>}

web/src/pages/DlnaPage.tsx

@@ -20,7 +20,7 @@ export function DlnaPage() {
     dlnaAPI
       .list(force)
       .then(setDevices)
-      .catch(() => toast.error('设备发现失败,容器网络可能不支持组播'))
+      .catch(() => toast.error('设备发现失败，容器网络可能不支持组播'))
       .finally(() => setScanning(false))
   }
 
@@ -94,7 +94,7 @@ export function DlnaPage() {
       {devices.length === 0 && !scanning && (
         <div className="glass-panel">
           <p className="text-ink-100">
-            未发现任何 DLNA 设备。请确保:服务器与设备在同一局域网,容器使用 host 网络模式,
+            未发现任何 DLNA 设备。请确保：服务器与设备在同一局域网，容器使用 host 网络模式，
             目标设备已开启 DLNA / 屏幕镜像。
           </p>
         </div>

web/src/pages/FavouritesPage.tsx

@@ -74,7 +74,7 @@ export function FavouritesPage() {
       {items.length > 0 && (
         <div className="grid grid-cols-2 gap-4 sm:grid-cols-3 md:grid-cols-4 lg:grid-cols-5 xl:grid-cols-6">
           {items.map((m) => (
-            <MediaCard key={m.id} media={m} />
+            <MediaCard key={m.id} media={m} linkState={{ from: '/favourites' }} />
           ))}
         </div>
       )}

web/src/pages/HomePageSections.tsx

@@ -17,7 +17,7 @@ import {
   Tv,
 } from 'lucide-react'
 
-import { imageURL } from '../api/client'
+import { ARTWORK, imageURL } from '../api/client'
 import { useInViewOnce } from '../hooks/useInViewOnce'
 import { useLazyPreviewBatch } from '../hooks/useLazyPreviewBatch'
 import { useRememberedListPosition } from '../hooks/useListPositionMemory'
@@ -171,7 +171,7 @@ export function HomeCarouselSection({
         {visual && (
           <img
             key={visual + currentItem.id}
-            src={imageURL(visual, currentItem.updated_at, { maxWidth: 1920, maxHeight: 1080, quality: 78 })}
+            src={imageURL(visual, currentItem.updated_at, ARTWORK.backdropHero)}
             alt=""
             fetchPriority="high"
             decoding="async"
@@ -283,7 +283,7 @@ export function HomeCarouselSection({
             </div>
             {poster && (
               <img
-                src={imageURL(poster, currentItem.updated_at, { maxWidth: 560, quality: 84 })}
+                src={imageURL(poster, currentItem.updated_at, ARTWORK.posterDetail)}
                 alt={currentItem.title}
                 fetchPriority="high"
                 decoding="async"
@@ -516,7 +516,7 @@ export function HomeLibraryCard({
                 src={imageURL(
                   src,
                   version,
-                  library.cover_url ? undefined : { maxWidth: 480, maxHeight: 320, quality: 78 },
+                  library.cover_url ? undefined : ARTWORK.backdropStrip,
                 )}
                 alt=""
                 loading="lazy"
@@ -750,7 +750,7 @@ function ContinueCard({ media, progress }: { media: Media; progress: number }) {
       <div className="relative h-20 w-14 shrink-0 overflow-hidden rounded-xl bg-[var(--app-panel-soft)]">
         {media.poster_url ? (
           <img
-            src={imageURL(media.poster_url, media.updated_at, { maxWidth: 180, maxHeight: 240, quality: 78 })}
+            src={imageURL(media.poster_url, media.updated_at, ARTWORK.posterCard)}
             alt=""
             className="h-full w-full object-cover transition-transform duration-500 group-hover:scale-105"
             loading="lazy"

web/src/pages/LibrarySeriesDetailHeader.tsx

@@ -2,7 +2,7 @@ import { isRemoteEmbyID } from '../utils/remoteEmby'
 import { Link } from 'react-router-dom'
 import { ArrowLeft, Database, FileText, Film, FolderInput, Pencil, Play, Search, Sparkles, Trash2 } from 'lucide-react'
 
-import { imageURL } from '../api/client'
+import { ARTWORK, imageURL } from '../api/client'
 import { ExternalPlayerButton } from '../components/ExternalPlayerButton'
 import { MediaFavouriteButton } from '../components/MediaFavouriteButton'
 import type { Media } from '../types'
@@ -70,7 +70,7 @@ export function LibrarySeriesDetailHeader({
         <div className="w-40 shrink-0 overflow-hidden rounded-xl bg-sand-200 shadow-card">
           {series.rep.poster_url ? (
             <img
-              src={imageURL(series.rep.poster_url, series.rep.updated_at)}
+              src={imageURL(series.rep.poster_url, series.rep.updated_at, ARTWORK.posterDetail)}
               alt={series.rep.title}
               loading="lazy"
               decoding="async"

web/src/pages/LibrarySeriesEpisodes.tsx

@@ -1,7 +1,7 @@
 import { Link } from 'react-router-dom'
 import { Play, Search } from 'lucide-react'
 
-import { imageURL } from '../api/client'
+import { ARTWORK, imageURL } from '../api/client'
 import { ExternalPlayerButton } from '../components/ExternalPlayerButton'
 import type { Media } from '../types'
 import {
@@ -87,11 +87,7 @@ export function LibrarySeriesEpisodes({
                   >
                     {ep.backdrop_url || ep.poster_url ? (
                       <img
-                        src={imageURL(ep.backdrop_url || ep.poster_url || '', ep.updated_at, {
-                          maxWidth: 320,
-                          maxHeight: 180,
-                          quality: 76,
-                        })}
+                        src={imageURL(ep.backdrop_url || ep.poster_url || '', ep.updated_at, ARTWORK.backdropStrip)}
                         alt=""
                         loading="lazy"
                         decoding="async"

web/src/pages/MediaDetailArtwork.tsx

@@ -2,7 +2,7 @@ import { motion } from 'framer-motion'
 import { FileText, Play } from 'lucide-react'
 import { Link } from 'react-router-dom'
 
-import { imageURL } from '../api/client'
+import { ARTWORK, imageURL } from '../api/client'
 import type { Media } from '../types'
 
 type MediaDetailArtworkProps = {
@@ -14,11 +14,7 @@ export function MediaDetailBackdrop({ media }: MediaDetailArtworkProps) {
     <div className="absolute inset-0 h-[480px] z-0 overflow-hidden">
       {media.backdrop_url || media.poster_url ? (
         <img
-          src={imageURL(media.backdrop_url || media.poster_url || '', media.updated_at, {
-            maxWidth: 1920,
-            maxHeight: 1080,
-            quality: 76,
-          })}
+          src={imageURL(media.backdrop_url || media.poster_url || '', media.updated_at, ARTWORK.backdropHero)}
           alt=""
           decoding="async"
           fetchPriority="low"
@@ -42,11 +38,7 @@ export function MediaDetailPoster({ media }: MediaDetailArtworkProps) {
       >
         {media.poster_url ? (
           <img
-            src={imageURL(media.poster_url, media.updated_at, {
-              maxWidth: 560,
-              maxHeight: 840,
-              quality: 84,
-            })}
+            src={imageURL(media.poster_url, media.updated_at, ARTWORK.posterDetail)}
             alt={media.title}
             decoding="async"
             fetchPriority="high"

web/src/pages/MediaDetailPageSections.tsx

@@ -85,7 +85,7 @@ export function MediaDetailBackButton({ onBack }: { onBack: () => void }) {
         className="btn-ghost gap-2 bg-white/80 shadow-sm backdrop-blur hover:bg-white"
       >
         <ArrowLeft size={16} />
-        <span>返回媒体库</span>
+        <span>返回</span>
       </button>
     </div>
   )

web/src/pages/PlayerPage.tsx

@@ -39,6 +39,8 @@ import { classifyDirectPlayError } from './directPlayError'
 import { apiErrorMessage } from './StrmManagePage'
 import { PlayerTopBar } from './PlayerTopBar'
 import { PlayerVideoStage } from './PlayerVideoStage'
+import { PlayerMobileTheaterInfo } from '../components/PlayerMobileTheaterInfo'
+import { useIsMobileTheater } from '../hooks/useIsMobileTheater'
 import { PlayerDanmakuPanel } from '../components/PlayerDanmakuPanel'
 import { PlayerPlaylistPanel } from '../components/PlayerPlaylistPanel'
 import {
@@ -197,6 +199,8 @@ export function PlayerPage() {
   // 选集 / 播放列表状态
   const [playlistEpisodes, setPlaylistEpisodes] = useState<Media[]>([])
   const [playlistOpen, setPlaylistOpen] = useState(false)
+  // 竖屏剧场模式：每次从操作栏点「选集」都 +1，把视频下方的选集区滚回视野。
+  const [episodeRevealToken, setEpisodeRevealToken] = useState(0)
 
   const teardownHls = useCallback(() => {
     if (hlsRef.current) {
@@ -1257,6 +1261,16 @@ export function PlayerPage() {
       if (next) setDanmakuOpen(false)
       return next
     })
+    setEpisodeRevealToken((token) => token + 1)
+  }, [])
+
+  // 操作栏 / 设置面板里的「选集」入口。竖屏剧场模式下选集区就在视频下方，
+  // 这里只负责展开并滚过去，不把已经展开的列表收起来（收起由选集区自己的
+  // 「收起」按钮负责），这样播放画面永远不会被浮层挡住。
+  const openPlaylistEntry = useCallback(() => {
+    setPlaylistOpen(true)
+    setDanmakuOpen(false)
+    setEpisodeRevealToken((token) => token + 1)
   }, [])
 
   const toggleDanmakuOpen = useCallback(() => {
@@ -1525,6 +1539,13 @@ export function PlayerPage() {
       setHlsSource('local')
       setCloudWaiting(false)
       setCloudWaitMessage('')
+      // 本地档位不可用（宿主机没有可用的 ffmpeg）时不要进入必然失败的 HLS：
+      // 保持直连播放并说明怎么修，避免播放器先打一个 500 再回退。
+      if (!quality.available) {
+        setPlaybackMode('direct')
+        toast.error(quality.note || '当前无法进行 HLS 转码，请在「系统设置 → 常规」安装 ffmpeg 后重试')
+        return
+      }
       if (position > 2) setHlsStartSec(position)
       setPlaybackMode('hls')
     },
@@ -1655,6 +1676,10 @@ export function PlayerPage() {
     media?.title?.trim() ||
     ''
 
+  // 竖屏手机走剧场布局：视频贴顶按 16:9 自适应，下方是标题/选集/简介。
+  // 桌面端与横屏手机保持原来的全屏居中沉浸式布局。
+  const isMobileTheater = useIsMobileTheater()
+
   return (
     <div className="relative flex h-full w-full flex-1 flex-col overflow-hidden bg-black">
       <PlayerTopBar
@@ -1667,6 +1692,7 @@ export function PlayerPage() {
         onBack={goBack}
       />
       <PlayerVideoStage
+        theater={isMobileTheater}
         media={media}
         loadError={loadError}
         playerError={playerError}
@@ -1708,7 +1734,7 @@ export function PlayerPage() {
         playlistOpen={playlistOpen}
         hasPlaylist={playlistEpisodes.length > 1 || currentVersions.length > 1}
         hasVersions={currentVersions.length > 1}
-        onTogglePlaylist={togglePlaylistOpen}
+        onTogglePlaylist={isMobileTheater ? openPlaylistEntry : togglePlaylistOpen}
         knownDuration={media?.duration_sec || 0}
         streamOffset={mode === 'hls' && hlsSource === 'local' ? hlsStartSec : 0}
         onSeekAbsolute={mode === 'hls' && hlsSource === 'local' ? handleSeekAbsolute : undefined}
@@ -1729,18 +1755,21 @@ export function PlayerPage() {
         showVr360Guide={Boolean(vr360) && vr360GuideSeen === false}
         onDismissVr360Guide={dismissVr360Guide}
         playlistPanel={
-          <PlayerPlaylistPanel
-            open={playlistOpen}
-            onClose={() => setPlaylistOpen(false)}
-            currentMediaId={media?.id ?? ''}
-            currentVersions={currentVersions}
-            episodes={playlistEpisodes}
-            onSelectEpisode={playEpisode}
-            onSelectVersion={switchVersion}
-          />
+          isMobileTheater
+            ? undefined
+            : <PlayerPlaylistPanel
+                open={playlistOpen}
+                onClose={() => setPlaylistOpen(false)}
+                currentMediaId={media?.id ?? ''}
+                currentVersions={currentVersions}
+                episodes={playlistEpisodes}
+                onSelectEpisode={playEpisode}
+                onSelectVersion={switchVersion}
+              />
         }
         danmakuPanel={
           <PlayerDanmakuPanel
+            theater={isMobileTheater}
             open={danmakuOpen}
             onClose={() => setDanmakuOpen(false)}
             enabled={danmakuEnabled}
@@ -1775,6 +1804,36 @@ export function PlayerPage() {
           />
         }
       />
+      {isMobileTheater ? (
+        <PlayerMobileTheaterInfo
+          media={media?.id === id ? media : null}
+          title={topBarTitle}
+          subtitle={topBarSubtitle}
+          playbackModeLabel={playbackModeLabel}
+          qualityLabel={showQuality ? qualityLabel : undefined}
+          episodes={playlistEpisodes}
+          currentMediaId={media?.id ?? ''}
+          currentEpisodeIndex={currentEpisodeIndex}
+          currentVersions={currentVersions}
+          onSelectEpisode={playEpisode}
+          onSelectVersion={switchVersion}
+          playlistOpen={playlistOpen}
+          onTogglePlaylist={togglePlaylistOpen}
+          revealToken={episodeRevealToken}
+          playlistPanel={
+            <PlayerPlaylistPanel
+              inline
+              open={playlistOpen}
+              onClose={() => setPlaylistOpen(false)}
+              currentMediaId={media?.id ?? ''}
+              currentVersions={currentVersions}
+              episodes={playlistEpisodes}
+              onSelectEpisode={playEpisode}
+              onSelectVersion={switchVersion}
+            />
+          }
+        />
+      ) : null}
     </div>
   )
 }

web/src/pages/PlayerVideoStage.tsx

@@ -156,6 +156,11 @@ function buildSubtitleRenderGroups(
 
 type PlayerVideoStageProps = {
   media: Media | null
+  /**
+   * 竖屏剧场模式：舞台不再全屏居中，而是贴顶按 16:9 自适应高度；
+   * 弹层改为底部动作面板，避免右侧抽屉与悬浮小弹层在窄屏下被裁剪。
+   */
+  theater?: boolean
   /** 媒体元数据加载失败提示（非空时替代「加载中」展示）。 */
   loadError?: string
   playerError: string
@@ -234,6 +239,7 @@ type PlayerVideoStageProps = {
 
 export function PlayerVideoStage({
   media,
+  theater = false,
   loadError,
   playerError,
   subs,
@@ -725,12 +731,22 @@ export function PlayerVideoStage({
   const assTrack = subtitleIndex >= 0 ? subs[subtitleIndex] : undefined
   // VR 模式下画面由 WebGL 画布输出，舞台窗口本身就是「镜头」，
   // 不再按视频宽高比留黑边。
+  //
+  // 竖屏剧场模式（theater）：舞台不再全屏居中，视频贴顶按容器宽度
+  // 16:9 自适应高度，下方留给标题/选集/简介内容区。stage 本体的
+  // 高度只包住视频区，控制栏与弹层都相对这块区域定位。
   const wrapperStyle = vr360
     ? {
         width: '100%',
         height: '100%',
       }
-    : videoRatio
+    : theater
+      ? {
+          aspectRatio: '16 / 9',
+          width: '100%',
+          maxHeight: '100%',
+        }
+      : videoRatio
       ? {
           aspectRatio: `${videoRatio}`,
           width: isWiderThanStage ? '100%' : 'auto',
@@ -747,7 +763,11 @@ export function PlayerVideoStage({
     <div
       ref={stageRef}
       data-player-stage
-      className="relative flex h-full w-full flex-1 items-center justify-center overflow-hidden bg-black"
+      className={
+        theater
+          ? 'relative flex w-full shrink-0 items-start justify-center overflow-hidden bg-black'
+          : 'relative flex h-full w-full flex-1 items-center justify-center overflow-hidden bg-black'
+      }
       onPointerDown={handleStagePointerDown}
       onPointerMove={handleStagePointerMove}
       onMouseLeave={() => setLockHovered(false)}
@@ -851,6 +871,7 @@ export function PlayerVideoStage({
             ) : null}
           </div>
           <PlayerControls
+            theater={theater}
             videoRef={videoRef}
             volume={playerVolume}
             onVolumeChange={onPlayerVolumeChange}

web/src/pages/ProfileManagementPage.tsx

@@ -154,7 +154,7 @@ export function ProfileManagementPage() {
         <div className="glass-panel py-12 text-center">
           <div className="mb-2 text-4xl">👤</div>
           <p className="font-medium text-ink-600">暂无 Profile</p>
-          <p className="text-sm text-ink-50">点击右上角"创建 Profile"开始</p>
+          <p className="text-sm text-ink-50">点击右上角「创建 Profile」开始</p>
         </div>
       )}
 

web/src/pages/ScraperQueuePage.tsx

@@ -19,7 +19,7 @@ import {
   X,
 } from 'lucide-react'
 
-import { imageURL } from '../api/client'
+import { ARTWORK, imageURL } from '../api/client'
 import { scraperAPI } from '../api/scraper'
 import type { ScrapeQueueSnapshot, ScrapeTask, ScrapeTaskStatus } from '../types/scraper'
 import { apiErrorMessage, formatTime, taskStatusMeta } from './StrmManagePage'
@@ -756,7 +756,7 @@ export function ScraperQueuePage({ embedded = false }: { embedded?: boolean }) {
                           <div className="flex items-center gap-2">
                             {task.poster_url ? (
                               <img
-                                src={imageURL(task.poster_url)}
+                                src={imageURL(task.poster_url, undefined, ARTWORK.posterTiny)}
                                 alt=""
                                 className="h-10 w-7 rounded object-cover border border-gray-200 shrink-0"
                                 onError={(e) => {

web/src/pages/WatchHistoryPage.tsx

@@ -4,7 +4,7 @@ import { Clock, Play, Trash2 } from 'lucide-react'
 import toast from 'react-hot-toast'
 
 import { historyAPI } from '../api/history'
-import { imageURL } from '../api/client'
+import { ARTWORK, imageURL } from '../api/client'
 import { confirmAction } from '../components/confirmAction'
 import { isRemoteEmbyID } from '../utils/remoteEmby'
 import type { HistoryItem, Media } from '../types'
@@ -117,7 +117,7 @@ export function WatchHistoryPage() {
               <div className="h-16 w-12 shrink-0 overflow-hidden rounded-lg bg-surface-900">
                 {m.poster_url ? (
                   <img
-                    src={imageURL(m.poster_url, m.updated_at)}
+                    src={imageURL(m.poster_url, m.updated_at, ARTWORK.posterCard)}
                     alt={m.title}
                     className="h-full w-full object-cover"
                     referrerPolicy="no-referrer"
@@ -127,6 +127,7 @@ export function WatchHistoryPage() {
               <div className="min-w-0 flex-1 space-y-1">
                 <Link
                   to={`/media/${m.id}`}
+                  state={{ from: '/history' }}
                   className="block truncate font-medium text-ink-600 transition hover:text-brand-500"
                 >
                   {m.title}
@@ -150,6 +151,7 @@ export function WatchHistoryPage() {
               <div className="flex w-full items-center gap-2 sm:w-auto sm:shrink-0">
                 <Link
                   to={`/play/${m.id}`}
+                  state={{ from: '/history' }}
                   className="neon-button !px-3 !py-1 !text-xs"
                 >
                   <Play size={12} /> 继续

web/src/pages/scraper-queue/ScrapeDetailModal.tsx

@@ -1,7 +1,7 @@
 import { Link } from 'react-router-dom'
 import { AlertCircle, Ban, Copy, ExternalLink, RefreshCw, Sparkles, Trash2, X } from 'lucide-react'
 
-import { imageURL } from '../../api/client'
+import { ARTWORK, imageURL } from '../../api/client'
 import type { ScrapeTask } from '../../types/scraper'
 import { formatTime, taskStatusMeta } from '../StrmManagePage'
 import { PROVIDER_LABELS, TYPE_LABELS } from './scrapeLabels'
@@ -52,7 +52,7 @@ export function ScrapeDetailModal({
             <div className="flex gap-4 rounded-2xl border border-brand-500/20 bg-primary-400/5 p-4">
               {task.poster_url && (
                 <img
-                  src={imageURL(task.poster_url)}
+                  src={imageURL(task.poster_url, undefined, ARTWORK.posterTiny)}
                   alt=""
                   className="h-28 w-20 rounded-xl object-cover border border-brand-500/30 shadow-md shrink-0"
                 />