Merge pull request #41 from truewhile/fix/database-playback-stability

修复数据库兼容性与 HLS 播放生命周期问题

.github/workflows/ci.yml

@@ -40,6 +40,39 @@ jobs:
       - name: go test
         run: go test ./...
 
+  database-compatibility:
+    name: SQLite / PostgreSQL ${{ matrix.postgres }}
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        postgres: [16, 17]
+    services:
+      postgres:
+        image: postgres:${{ matrix.postgres }}-alpine
+        env:
+          POSTGRES_USER: mebox
+          POSTGRES_PASSWORD: mebox-test-only
+          POSTGRES_DB: mebox_test
+        ports:
+          - 5432:5432
+        options: >-
+          --health-cmd "pg_isready -U mebox -d mebox_test"
+          --health-interval 5s
+          --health-timeout 5s
+          --health-retries 10
+    env:
+      MEBOX_TEST_POSTGRES_DSN: postgres://mebox:mebox-test-only@127.0.0.1:5432/mebox_test?sslmode=disable
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-go@v5
+        with:
+          go-version: '1.26'
+          cache: true
+
+      - name: Database regression tests
+        run: go test ./internal/database ./internal/repository -count=1
+
   frontend:
     name: Frontend (Node)
     runs-on: ubuntu-latest
@@ -78,7 +111,7 @@ jobs:
   ci-success:
     name: CI Success
     runs-on: ubuntu-latest
-    needs: [backend, frontend, smoke]
+    needs: [backend, frontend, database-compatibility, smoke]
     if: success()
     steps:
       - run: echo "CI passed, ready for release"

CONTRIBUTING.md

@@ -84,6 +84,15 @@ go test ./internal/service -run "TestOrganize|TestSubscription|TestMTeam" -count
 go test ./cmd/server -count=1
 ```
 
+数据库兼容测试默认覆盖 SQLite。连接专用测试 PostgreSQL 后可运行真实的双数据库回归：
+
+```bash
+MEBOX_TEST_POSTGRES_DSN='postgres://mebox:test-password@127.0.0.1:5432/mebox_test?sslmode=disable' \
+  go test ./internal/database ./internal/repository -count=1
+```
+
+测试账号需要创建、删除 schema 的权限；每个用例使用独立的临时 schema。CI 会分别在 PostgreSQL 16、17 上执行这些检查。
+
 ### 代码要求
 
 - Go 代码使用 `gofmt`。

Dockerfile

@@ -84,12 +84,15 @@ ENV MEBOX_APP_PORT=8080 \
 
 EXPOSE 8080
 
-HEALTHCHECK --interval=30s --timeout=5s --start-period=15s --retries=3 \
-    CMD busybox wget -q --spider http://127.0.0.1:8080/api/health || exit 1
+# 探测协议不能写死：开启 HTTPS 后 8080 只接受 TLS。脚本按数据目录里的
+# .http_scheme 选择协议并在失败时回退，无需用户改配置。
+HEALTHCHECK --interval=30s --timeout=10s --start-period=30s --retries=3 \
+    CMD ["/usr/local/bin/mebox-healthcheck"]
 
 # Tiny entrypoint that lets us run as a NAS host UID/GID via PUID/PGID without
 # rewriting /etc/passwd or /etc/group on every container start.
 COPY docker-entrypoint.sh /entrypoint.sh
-RUN chmod +x /entrypoint.sh
+COPY docker-healthcheck.sh /usr/local/bin/mebox-healthcheck
+RUN chmod +x /entrypoint.sh /usr/local/bin/mebox-healthcheck
 
 CMD ["/entrypoint.sh"]

cmd/server/application.go

@@ -88,7 +88,7 @@ func (a *application) start() error {
 		a.logger.Error("database not ready", zap.Error(err))
 		return fmt.Errorf("database not ready: %w", err)
 	}
-	if err := database.AutoMigrate(db); err != nil {
+	if err := database.AutoMigrateWithConfig(db, a.cfg, a.logger); err != nil {
 		a.logger.Error("auto-migrate failed", zap.Error(err))
 		return fmt.Errorf("auto-migrate failed: %w", err)
 	}

cmd/server/protocol_mux.go

@@ -0,0 +1,129 @@
+package main
+
+import (
+	"bufio"
+	"net"
+	"sync"
+	"time"
+)
+
+// 开启 HTTPS 后，同一个端口会同时面对两类连接：
+//
+//   - TLS 客户端（浏览器、播放器）——正常提供 HTTPS 服务；
+//   - 明确访问 http:// 的客户端——以前只会撞出一句
+//     "Client sent an HTTP request to an HTTPS server."，用户看不懂。
+//
+// 一个端口上只有 TLS 监听时无法区分两者，所以在 Accept 之后先偷看第一个字节：
+// TLS 握手的记录类型固定是 0x16（handshake），而 HTTP 请求行必然是可见字符。
+// 据此把连接分流成两路，明文那一路交给只做跳转的 http.Server，于是 http:// 会被
+// 自动重定向到 https://（见 server_manager.go 的 httpsRedirectHandler）。
+const tlsRecordHandshake = 0x16
+
+// defaultPeekTimeout 限制"偷看首字节"阶段的等待，避免只连不发的半开连接长期占住
+// goroutine。分流器按实例持有该值，便于测试用更短的超时。
+const defaultPeekTimeout = 10 * time.Second
+
+// splitListener 是监听分流器：Accept 到底层连接后按首字节投递给对应的一路。
+type splitListener struct {
+	ln      net.Listener
+	tlsCh   chan net.Conn
+	plainCh chan net.Conn
+	quit    chan struct{}
+	once    sync.Once
+	peek    time.Duration
+}
+
+func newSplitListener(ln net.Listener) *splitListener {
+	return newSplitListenerWithTimeout(ln, defaultPeekTimeout)
+}
+
+func newSplitListenerWithTimeout(ln net.Listener, peek time.Duration) *splitListener {
+	s := &splitListener{
+		ln:      ln,
+		tlsCh:   make(chan net.Conn),
+		plainCh: make(chan net.Conn),
+		quit:    make(chan struct{}),
+		peek:    peek,
+	}
+	go s.acceptLoop()
+	return s
+}
+
+// tlsListener 返回只收到 TLS 连接的 listener。
+func (s *splitListener) tlsListener() net.Listener {
+	return &channelListener{addr: s.ln.Addr(), conns: s.tlsCh, quit: s.quit}
+}
+
+// plainListener 返回只收到明文 HTTP 连接的 listener。
+func (s *splitListener) plainListener() net.Listener {
+	return &channelListener{addr: s.ln.Addr(), conns: s.plainCh, quit: s.quit}
+}
+
+func (s *splitListener) acceptLoop() {
+	for {
+		conn, err := s.ln.Accept()
+		if err != nil {
+			// 底层监听关闭（换证、关停）后两路 Accept 一起返回 net.ErrClosed，
+			// 由持有它们的 http.Server 自行收尾。
+			s.stop()
+			return
+		}
+		go s.dispatch(conn)
+	}
+}
+
+func (s *splitListener) dispatch(conn net.Conn) {
+	br := bufio.NewReader(conn)
+	_ = conn.SetReadDeadline(time.Now().Add(s.peek))
+	first, err := br.Peek(1)
+	_ = conn.SetReadDeadline(time.Time{})
+	if err != nil {
+		_ = conn.Close()
+		return
+	}
+
+	peeked := &peekedConn{Conn: conn, r: br}
+	ch := s.plainCh
+	if first[0] == tlsRecordHandshake {
+		ch = s.tlsCh
+	}
+	select {
+	case ch <- peeked: // 交给对应的 http.Server，由它负责读写与关闭
+	case <-s.quit:
+		_ = conn.Close()
+	}
+}
+
+func (s *splitListener) stop() {
+	s.once.Do(func() { close(s.quit) })
+}
+
+// channelListener 把一路 channel 包装成 net.Listener，供 http.Server.Serve 使用。
+type channelListener struct {
+	addr  net.Addr
+	conns <-chan net.Conn
+	quit  <-chan struct{}
+}
+
+func (l *channelListener) Accept() (net.Conn, error) {
+	select {
+	case conn := <-l.conns:
+		return conn, nil
+	case <-l.quit:
+		return nil, net.ErrClosed
+	}
+}
+
+// Close 是空实现：真正要释放的端口在 splitListener 的底层监听上，它一旦关闭，
+// 这里的 Accept 就会返回 net.ErrClosed，http.Server 随之退出。
+func (l *channelListener) Close() error { return nil }
+
+func (l *channelListener) Addr() net.Addr { return l.addr }
+
+// peekedConn 把已经读进 bufio 缓冲的首字节还给上层读取者。
+type peekedConn struct {
+	net.Conn
+	r *bufio.Reader
+}
+
+func (c *peekedConn) Read(p []byte) (int, error) { return c.r.Read(p) }

cmd/server/protocol_mux_test.go

@@ -0,0 +1,140 @@
+package main
+
+import (
+	"io"
+	"net"
+	"net/http"
+	"testing"
+	"time"
+)
+
+func TestSplitListenerRoutesByFirstByte(t *testing.T) {
+	raw, err := net.Listen("tcp", "127.0.0.1:0")
+	if err != nil {
+		t.Fatal(err)
+	}
+	split := newSplitListener(raw)
+	addr := raw.Addr().String()
+
+	// TLS：首字节 0x16。分流器偷看过的字节必须还能被上层读到，否则握手会失败。
+	handshake := []byte{0x16, 0x03, 0x01, 0x00, 0x2c}
+	tlsClient, err := net.Dial("tcp", addr)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer func() { _ = tlsClient.Close() }()
+	if _, err := tlsClient.Write(handshake); err != nil {
+		t.Fatal(err)
+	}
+	tlsConn := <-split.tlsCh
+	defer func() { _ = tlsConn.Close() }()
+	got := make([]byte, len(handshake))
+	if _, err := io.ReadFull(tlsConn, got); err != nil {
+		t.Fatalf("read peeked tls bytes: %v", err)
+	}
+	if string(got) != string(handshake) {
+		t.Fatalf("tls bytes = %v, want %v", got, handshake)
+	}
+
+	// 明文 HTTP：请求行走另一路，同样要读到偷看过的首字节。
+	request := "GET /api/health HTTP/1.1\r\nHost: example.com\r\n\r\n"
+	plainClient, err := net.Dial("tcp", addr)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer func() { _ = plainClient.Close() }()
+	if _, err := plainClient.Write([]byte(request)); err != nil {
+		t.Fatal(err)
+	}
+	plainConn := <-split.plainCh
+	defer func() { _ = plainConn.Close() }()
+	head := make([]byte, len("GET /"))
+	if _, err := io.ReadFull(plainConn, head); err != nil {
+		t.Fatalf("read peeked plain bytes: %v", err)
+	}
+	if string(head) != request[:len(head)] {
+		t.Fatalf("plain bytes = %q, want prefix of %q", head, request)
+	}
+
+	// 底层监听关闭后两路 Accept 都要返回 net.ErrClosed，http.Server 才能收尾。
+	_ = raw.Close()
+	for _, ln := range []net.Listener{split.tlsListener(), split.plainListener()} {
+		deadline := time.Now().Add(3 * time.Second)
+		for {
+			_, err := ln.Accept()
+			if err == net.ErrClosed {
+				break
+			}
+			if time.Now().After(deadline) {
+				t.Fatalf("Accept after close = %v, want net.ErrClosed", err)
+			}
+			time.Sleep(10 * time.Millisecond)
+		}
+	}
+}
+
+// 只连不发的半开连接不应被分给任何一路，而应在偷看超时后被丢弃。
+func TestSplitListenerDropsIdleConnection(t *testing.T) {
+	raw, err := net.Listen("tcp", "127.0.0.1:0")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer func() { _ = raw.Close() }()
+	split := newSplitListenerWithTimeout(raw, 200*time.Millisecond)
+
+	conn, err := net.Dial("tcp", raw.Addr().String())
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer func() { _ = conn.Close() }()
+
+	select {
+	case c := <-split.plainCh:
+		_ = c.Close()
+		t.Fatal("idle connection should not be dispatched")
+	case c := <-split.tlsCh:
+		_ = c.Close()
+		t.Fatal("idle connection should not be dispatched")
+	case <-time.After(time.Second):
+	}
+}
+
+func TestHTTPSRedirectTarget(t *testing.T) {
+	cases := []struct {
+		name    string
+		host    string
+		rawURL  string
+		want    string
+		wantErr bool
+	}{
+		{name: "keeps host port path and query", host: "10.0.0.2:8080", rawURL: "/library?tab=1&q=a%20b", want: "https://10.0.0.2:8080/library?tab=1&q=a%20b"},
+		{name: "root path", host: "nas.local:18080", rawURL: "/", want: "https://nas.local:18080/"},
+		{name: "keeps escaping of user input", host: "10.0.0.2", rawURL: "/a%2Fb", want: "https://10.0.0.2/a%2Fb"},
+		{name: "absolute form request line", host: "10.0.0.2:8080", rawURL: "http://10.0.0.2:8080/api/x?y=1", want: "https://10.0.0.2:8080/api/x?y=1"},
+		{name: "rejects empty host", host: "", rawURL: "/", wantErr: true},
+		{name: "rejects host with slash", host: "evil.com/path", rawURL: "/", wantErr: true},
+		{name: "rejects userinfo host", host: "user@evil.com", rawURL: "/", wantErr: true},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			req, err := http.NewRequest(http.MethodGet, tc.rawURL, nil)
+			if err != nil {
+				t.Fatal(err)
+			}
+			req.Host = tc.host
+
+			if host := redirectHostHeader(req); host == "" {
+				if !tc.wantErr {
+					t.Fatalf("redirectHostHeader(%q) rejected, want accepted", tc.host)
+				}
+				return
+			}
+			if tc.wantErr {
+				t.Fatalf("redirectHostHeader(%q) accepted, want rejected", tc.host)
+			}
+			if got := httpsRedirectURL(tc.host, req); got != tc.want {
+				t.Fatalf("redirect target = %q, want %q", got, tc.want)
+			}
+		})
+	}
+}

cmd/server/server_manager.go

@@ -7,6 +7,8 @@ import (
 	"fmt"
 	"net"
 	"net/http"
+	"os"
+	"path/filepath"
 	"strings"
 	"sync"
 	"time"
@@ -17,6 +19,14 @@ import (
 	"github.com/truewhile/MeBox/internal/service"
 )
 
+// httpSchemeFileName 是数据目录下记录当前监听协议的小文件（内容为 http 或 https）。
+//
+// 容器健康检查无法读取后台保存的 HTTPS 设置（它存在数据库里），而同一端口在开启
+// HTTPS 后只接受 TLS。若健康检查固定用 http:// 探测，每轮都会撞出
+// "client sent an HTTP request to an HTTPS server"，既刷日志又让容器永远 unhealthy。
+// 因此由服务进程在每次成功绑定监听后写下真实协议，供 docker-healthcheck.sh 选择。
+const httpSchemeFileName = ".http_scheme"
+
 // tlsPair 记录当前正在服务的证书，用于判断是否需要重新绑定监听。
 type tlsPair struct {
 	cert    tls.Certificate
@@ -35,6 +45,9 @@ type tlsPair struct {
 //   - 旧服务器随后优雅退出，正在进行的播放/请求不会被立刻掐断。
 //
 // 任何校验失败都会中止切换并保留旧监听，保证用户不会被锁在服务外面。
+//
+// 开启 HTTPS 时端口上会同时接两类连接：TLS 走主服务器，明文 HTTP 走
+// redirectSrv 并跳转到 HTTPS（见 protocol_mux.go）。
 type serverManager struct {
 	cfg     *config.Config
 	log     *zap.Logger
@@ -43,6 +56,7 @@ type serverManager struct {
 
 	mu                sync.Mutex
 	srv               *http.Server
+	redirectSrv       *http.Server
 	ln                net.Listener
 	pair              *tlsPair
 	stopCh            chan struct{}
@@ -92,11 +106,11 @@ func (m *serverManager) Reload() error {
 		return nil
 	}
 
-	oldSrv, oldLn, oldPair := m.srv, m.ln, m.pair
+	oldSrv, oldRedirect, oldLn, oldPair := m.srv, m.redirectSrv, m.ln, m.pair
 	if oldLn != nil {
 		_ = oldLn.Close() // 释放端口后再绑定新监听
 	}
-	m.srv, m.ln, m.pair = nil, nil, nil
+	m.srv, m.redirectSrv, m.ln, m.pair = nil, nil, nil, nil
 
 	firstErr := m.bind(pair)
 	if firstErr != nil {
@@ -107,6 +121,7 @@ func (m *serverManager) Reload() error {
 	}
 	// 新监听已就绪，让旧服务器在新连接切换到新监听后优雅退出。
 	m.drain(oldSrv)
+	m.drain(oldRedirect)
 	m.logServerReady()
 	m.maybeStartAutoReloadLocked()
 	return firstErr
@@ -120,11 +135,19 @@ func (m *serverManager) Shutdown(ctx context.Context) error {
 		close(m.stopCh)
 	}
 	m.mu.Lock()
-	defer m.mu.Unlock()
-	if m.srv == nil {
-		return nil
+	targets := []*http.Server{m.srv, m.redirectSrv}
+	m.mu.Unlock()
+
+	var firstErr error
+	for _, srv := range targets {
+		if srv == nil {
+			continue
+		}
+		if err := srv.Shutdown(ctx); err != nil && firstErr == nil {
+			firstErr = err
+		}
 	}
-	return m.srv.Shutdown(ctx)
+	return firstErr
 }
 
 // desiredPair 根据当前配置计算目标监听形态：nil 表示明文 HTTP，非 nil 表示 TLS。
@@ -217,25 +240,107 @@ func (m *serverManager) pairEquals(pair *tlsPair) bool {
 	return pair.version == m.pair.version
 }
 
-// bind 创建并按需启用 TLS 的监听，异步开始服务。
+// bind 创建监听，异步开始服务。
+//
+// 未开启 HTTPS 时端口只服务明文 HTTP；开启后同一个端口两种流量都要接：TLS 连接
+// 交给主服务器，明文 HTTP 连接交给只做跳转的 redirect 服务器，让 http:// 自动
+// 跳到 https://，而不是甩给用户一句 "Client sent an HTTP request to an HTTPS
+// server."。
 func (m *serverManager) bind(pair *tlsPair) error {
 	ln, err := net.Listen("tcp", m.addr)
 	if err != nil {
 		return fmt.Errorf("listen %s: %w", m.addr, err)
 	}
-	srv := &http.Server{
-		Handler:           m.handler,
+
+	if pair == nil {
+		srv := m.newHTTPServer(m.handler)
+		m.srv, m.redirectSrv, m.ln, m.pair = srv, nil, ln, nil
+		m.writeSchemeState(nil)
+		go m.serve(srv, ln)
+		return nil
+	}
+
+	split := newSplitListener(ln)
+	srv := m.newHTTPServer(m.handler)
+	tlsLn := tls.NewListener(split.tlsListener(), &tls.Config{
+		Certificates: []tls.Certificate{pair.cert},
+		MinVersion:   tls.VersionTLS12,
+	})
+	redirect := m.newHTTPServer(m.httpsRedirectHandler())
+
+	m.srv, m.redirectSrv, m.ln, m.pair = srv, redirect, ln, pair
+	m.writeSchemeState(pair)
+	go m.serve(srv, tlsLn)
+	go m.serve(redirect, split.plainListener())
+	return nil
+}
+
+func (m *serverManager) newHTTPServer(handler http.Handler) *http.Server {
+	return &http.Server{
+		Handler:           handler,
 		ReadHeaderTimeout: 15 * time.Second,
 	}
-	if pair != nil {
-		ln = tls.NewListener(ln, &tls.Config{
-			Certificates: []tls.Certificate{pair.cert},
-			MinVersion:   tls.VersionTLS12,
-		})
+}
+
+// httpsRedirectHandler 处理"开着 HTTPS 却用 http:// 访问"的请求：跳转到同端口、
+// 同路径的 https://。
+//
+// 用 302/307 而不是 301/308：HTTPS 开关可以在后台随时改，永久跳转会被浏览器长期
+// 缓存，等用户关掉 HTTPS 之后 http 访问会被带到已经不提供 TLS 的端口上打不开。
+func (m *serverManager) httpsRedirectHandler() http.Handler {
+	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		host := redirectHostHeader(r)
+		if host == "" {
+			// 没有可用的 Host（HTTP/1.0 或畸形请求头）时宁可明说，也不跳到错的地方。
+			w.Header().Set("Content-Type", "text/plain; charset=utf-8")
+			w.WriteHeader(http.StatusBadRequest)
+			fmt.Fprintf(w, "HTTPS is enabled; open https://<host>:%d instead.\n", m.cfg.App.Port)
+			return
+		}
+		status := http.StatusFound
+		if r.Method != http.MethodGet && r.Method != http.MethodHead {
+			status = http.StatusTemporaryRedirect // 保住方法与请求体
+		}
+		w.Header().Set("Location", httpsRedirectURL(host, r))
+		w.WriteHeader(status)
+	})
+}
+
+// httpsRedirectURL 拼出同端口同路径的 HTTPS 地址。
+func httpsRedirectURL(host string, r *http.Request) string {
+	target := "https://" + host + r.URL.EscapedPath()
+	if r.URL.RawQuery != "" {
+		target += "?" + r.URL.RawQuery
+	}
+	return target
+}
+
+// redirectHostHeader 取重定向目标主机。HTTP/1.1 里 Host 是唯一的主机来源（端口映射
+// 到容器 8080 时它同时也是 HTTPS 端口），这里只做字符过滤，避免把请求头里的怪字符
+// 带进 Location 造成任意跳转。
+func redirectHostHeader(r *http.Request) string {
+	host := strings.TrimSpace(r.Host)
+	if host == "" || strings.ContainsAny(host, " \t\r\n/\\@?#") {
+		return ""
+	}
+	return host
+}
+
+// writeSchemeState 把当前实际监听的协议写入数据目录，供容器健康检查按协议探测。
+// 写失败只降级为告警：健康检查脚本在状态文件缺失时会回退试另一种协议。
+func (m *serverManager) writeSchemeState(pair *tlsPair) {
+	dataDir := strings.TrimSpace(m.cfg.App.DataDir)
+	if dataDir == "" {
+		return
+	}
+	scheme := "http"
+	if pair != nil {
+		scheme = "https"
+	}
+	path := filepath.Join(dataDir, httpSchemeFileName)
+	if err := os.WriteFile(path, []byte(scheme+"\n"), 0o644); err != nil { // #nosec G304 -- path is DataDir + fixed filename.
+		m.log.Warn("write http scheme state failed", zap.String("path", path), zap.Error(err))
 	}
-	m.srv, m.ln, m.pair = srv, ln, pair
-	go m.serve(srv, ln)
-	return nil
 }
 
 func (m *serverManager) serve(s *http.Server, ln net.Listener) {
@@ -250,6 +355,9 @@ func (m *serverManager) drain(s *http.Server) {
 	if s == nil {
 		return
 	}
+	// Stop reusing old connections before Reload returns. Asynchronous Shutdown
+	// alone leaves a window where a client receives the previous HTTP/TLS mode.
+	s.SetKeepAlivesEnabled(false)
 	go func(s *http.Server) {
 		ctx, cancel := context.WithTimeout(context.Background(), 15*time.Second)
 		defer cancel()
@@ -271,7 +379,7 @@ func (m *serverManager) logServerReady() {
 		zap.String("listen", m.addr),
 	)
 	if m.pair != nil {
-		m.log.Info("HTTPS is enabled; plain HTTP is no longer served on this port",
+		m.log.Info("HTTPS is enabled; plain HTTP requests on this port are redirected to HTTPS",
 			zap.String("addr", m.addr),
 		)
 	}

cmd/server/server_manager_test.go

@@ -1,13 +1,17 @@
 package main
 
 import (
+	"context"
 	"crypto/ecdsa"
 	"crypto/elliptic"
 	"crypto/rand"
+	"crypto/tls"
 	"crypto/x509"
 	"crypto/x509/pkix"
 	"encoding/pem"
+	"io"
 	"math/big"
+	"net"
 	"net/http"
 	"os"
 	"path/filepath"
@@ -32,6 +36,8 @@ func makeTestPairPEM(t *testing.T) (certPEM, keyPEM string) {
 		NotBefore:    time.Now().Add(-time.Hour),
 		NotAfter:     time.Now().Add(24 * time.Hour),
 		DNSNames:     []string{"localhost"},
+		// 带上回环 IP，HTTPS 客户端就能在校验证书的前提下连 127.0.0.1:port。
+		IPAddresses: []net.IP{net.ParseIP("127.0.0.1")},
 	}
 	der, err := x509.CreateCertificate(rand.Reader, tpl, tpl, &priv.PublicKey, priv)
 	if err != nil {
@@ -105,3 +111,136 @@ func TestDesiredPairModes(t *testing.T) {
 		t.Fatal("version should change after files replaced")
 	}
 }
+
+// 健康检查依赖该文件判断该用 http 还是 https 探测，内容与当前监听协议必须一致，
+// 否则容器会一直显示 unhealthy 并刷 TLS handshake error 日志。
+func TestWriteSchemeState(t *testing.T) {
+	m := newTestServerManager(t)
+	m.cfg.App.DataDir = t.TempDir()
+	path := filepath.Join(m.cfg.App.DataDir, httpSchemeFileName)
+
+	m.writeSchemeState(nil)
+	if data, err := os.ReadFile(path); err != nil || strings.TrimSpace(string(data)) != "http" {
+		t.Fatalf("plain HTTP state: data=%q err=%v", data, err)
+	}
+
+	m.writeSchemeState(&tlsPair{})
+	if data, err := os.ReadFile(path); err != nil || strings.TrimSpace(string(data)) != "https" {
+		t.Fatalf("HTTPS state: data=%q err=%v", data, err)
+	}
+
+	// DataDir 为空时不应写到工作目录（测试/裸跑二进制时会污染仓库）。
+	m.cfg.App.DataDir = ""
+	_ = os.Remove(path)
+	m.writeSchemeState(nil)
+	if _, err := os.Stat(httpSchemeFileName); !os.IsNotExist(err) {
+		t.Fatalf("empty DataDir should not create %s in cwd: %v", httpSchemeFileName, err)
+	}
+}
+
+// 开启 HTTPS 后同一个端口要同时接 TLS 与明文 HTTP：明文请求跳到 HTTPS，而不是
+// 给用户看 "Client sent an HTTP request to an HTTPS server."。
+func TestServerManagerRedirectsPlainHTTPToHTTPS(t *testing.T) {
+	certPEM, keyPEM := makeTestPairPEM(t)
+	cfg := &config.Config{}
+	cfg.App.Port = 8080
+	cfg.App.DataDir = t.TempDir()
+	cfg.App.HTTPSEnabled = true
+	cfg.App.SSLCert, cfg.App.SSLKey = certPEM, keyPEM
+
+	handler := http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusOK)
+		_, _ = io.WriteString(w, "secure:"+r.URL.Path)
+	})
+	m := newServerManager(cfg, zap.NewNop(), handler)
+	m.addr = "127.0.0.1:0" // 端口 0 交给内核分配，测试之间不会撞端口
+	if err := m.Start(); err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() { _ = m.Shutdown(context.Background()) })
+
+	addr := m.ln.Addr().String()
+	// 端口 0 只在第一次分配时用；热切换要绑回同一个端口（生产里端口本来就是固定的）。
+	m.addr = addr
+	schemeFile := filepath.Join(cfg.App.DataDir, httpSchemeFileName)
+	if data, err := os.ReadFile(schemeFile); err != nil || strings.TrimSpace(string(data)) != "https" {
+		t.Fatalf("scheme state = %q err=%v, want https", data, err)
+	}
+
+	plain := &http.Client{
+		Timeout:       10 * time.Second,
+		CheckRedirect: func(*http.Request, []*http.Request) error { return http.ErrUseLastResponse },
+	}
+
+	resp, err := plain.Get("http://" + addr + "/library?tab=1")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer func() { _ = resp.Body.Close() }()
+	if resp.StatusCode != http.StatusFound {
+		t.Fatalf("plain GET status = %d, want %d", resp.StatusCode, http.StatusFound)
+	}
+	if got, want := resp.Header.Get("Location"), "https://"+addr+"/library?tab=1"; got != want {
+		t.Fatalf("Location = %q, want %q", got, want)
+	}
+
+	postResp, err := plain.Post("http://"+addr+"/api/me/login", "application/json", strings.NewReader("{}"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer func() { _ = postResp.Body.Close() }()
+	if postResp.StatusCode != http.StatusTemporaryRedirect {
+		t.Fatalf("plain POST status = %d, want %d", postResp.StatusCode, http.StatusTemporaryRedirect)
+	}
+	if got, want := postResp.Header.Get("Location"), "https://"+addr+"/api/me/login"; got != want {
+		t.Fatalf("POST Location = %q, want %q", got, want)
+	}
+
+	// 同一端口上的 TLS 连接照常拿到正常响应（校验证书，不做跳过校验）。
+	pool := x509.NewCertPool()
+	if !pool.AppendCertsFromPEM([]byte(certPEM)) {
+		t.Fatal("append test certificate to pool failed")
+	}
+	secure := &http.Client{
+		Timeout: 10 * time.Second,
+		Transport: &http.Transport{
+			TLSClientConfig: &tls.Config{RootCAs: pool, MinVersion: tls.VersionTLS12},
+		},
+	}
+	secureResp, err := secure.Get("https://" + addr + "/library")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer func() { _ = secureResp.Body.Close() }()
+	body, err := io.ReadAll(secureResp.Body)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if secureResp.StatusCode != http.StatusOK || string(body) != "secure:/library" {
+		t.Fatalf("https response = %d %q, want 200 %q", secureResp.StatusCode, body, "secure:/library")
+	}
+
+	// 关掉 HTTPS 后同一端口回到纯明文服务，不再有多余的跳转服务器。
+	cfg.App.HTTPSEnabled = false
+	if err := m.Reload(); err != nil {
+		t.Fatal(err)
+	}
+	if m.redirectSrv != nil {
+		t.Fatal("redirect server should be released when HTTPS is off")
+	}
+	if data, err := os.ReadFile(schemeFile); err != nil || strings.TrimSpace(string(data)) != "http" {
+		t.Fatalf("scheme state = %q err=%v, want http", data, err)
+	}
+	plainResp, err := plain.Get("http://" + addr + "/library")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer func() { _ = plainResp.Body.Close() }()
+	plainBody, err := io.ReadAll(plainResp.Body)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if plainResp.StatusCode != http.StatusOK || string(plainBody) != "secure:/library" {
+		t.Fatalf("plain response after disabling https = %d %q", plainResp.StatusCode, plainBody)
+	}
+}

docker-compose.metatube.yml

@@ -62,7 +62,7 @@ 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。
-      - ./media: /media
+      - ./media:/media
 
     environment:
       TZ: Asia/Shanghai
@@ -105,7 +105,9 @@ services:
       MEBOX_TRANSCODER_IDLE_TIMEOUT_SECONDS: "120"
 
     healthcheck:
-      test: ["CMD-SHELL", "busybox wget -qO- http://127.0.0.1:8080/api/health || exit 1"]
+      # 不能写死 http://：开启 HTTPS 后 8080 只接受 TLS，固定明文探测会
+      # 每轮刷一条 TLS handshake error 并让容器一直 unhealthy。
+      test: ["CMD", "/usr/local/bin/mebox-healthcheck"]
       interval: 30s
       timeout: 10s
       retries: 5

docker-compose.search.yml

@@ -115,7 +115,9 @@ services:
       MEBOX_TRANSCODER_IDLE_TIMEOUT_SECONDS: "120"
 
     healthcheck:
-      test: ["CMD-SHELL", "busybox wget -qO- http://127.0.0.1:8080/api/health || exit 1"]
+      # 不能写死 http://：开启 HTTPS 后 8080 只接受 TLS，固定明文探测会
+      # 每轮刷一条 TLS handshake error 并让容器一直 unhealthy。
+      test: ["CMD", "/usr/local/bin/mebox-healthcheck"]
       interval: 30s
       timeout: 10s
       retries: 5

docker-compose.simple.yml

@@ -73,7 +73,9 @@ services:
       - "host.docker.internal:host-gateway"
 
     healthcheck:
-      test: ["CMD-SHELL", "busybox wget -qO- http://127.0.0.1:8080/api/health || exit 1"]
+      # 不能写死 http://：开启 HTTPS 后 8080 只接受 TLS，固定明文探测会
+      # 每轮刷一条 TLS handshake error 并让容器一直 unhealthy。
+      test: ["CMD", "/usr/local/bin/mebox-healthcheck"]
       interval: 30s
       timeout: 10s
       retries: 5

docker-compose.standard.yml

@@ -44,7 +44,7 @@ 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。
-      - ./media: /media
+      - ./media:/media
 
     environment:
       TZ: Asia/Shanghai
@@ -87,7 +87,9 @@ services:
       MEBOX_TRANSCODER_IDLE_TIMEOUT_SECONDS: "120"
 
     healthcheck:
-      test: ["CMD-SHELL", "busybox wget -qO- http://127.0.0.1:8080/api/health || exit 1"]
+      # 不能写死 http://：开启 HTTPS 后 8080 只接受 TLS，固定明文探测会
+      # 每轮刷一条 TLS handshake error 并让容器一直 unhealthy。
+      test: ["CMD", "/usr/local/bin/mebox-healthcheck"]
       interval: 30s
       timeout: 10s
       retries: 5

docker-compose.yml

@@ -131,7 +131,9 @@ services:
       MEBOX_TRANSCODER_IDLE_TIMEOUT_SECONDS: "120"
 
     healthcheck:
-      test: ["CMD-SHELL", "busybox wget -qO- http://127.0.0.1:8080/api/health || exit 1"]
+      # 不能写死 http://：开启 HTTPS 后 8080 只接受 TLS，固定明文探测会
+      # 每轮刷一条 TLS handshake error 并让容器一直 unhealthy。
+      test: ["CMD", "/usr/local/bin/mebox-healthcheck"]
       interval: 30s
       timeout: 10s
       retries: 5

docker-healthcheck.sh

@@ -0,0 +1,39 @@
+#!/bin/sh
+# MeBox 容器健康检查。
+#
+# 背景：MeBox 只监听一个端口，且开启 HTTPS 后该端口只接受 TLS（明文 HTTP 不再
+# 服务）。健康检查若固定用 http:// 探测，每轮都会被 TLS 监听拒绝并在服务日志里
+# 留下一条 "client sent an HTTP request to an HTTPS server"，同时容器状态永远
+# 是 unhealthy。反过来固定用 https:// 探测，则在未开启 HTTPS 的部署上产生
+# "first record does not look like a TLS handshake"。
+#
+# 当前协议由服务进程写入 $MEBOX_APP_DATA_DIR/.http_scheme（见
+# cmd/server/server_manager.go 的 writeSchemeState）。脚本优先按该文件选择协议，
+# 文件缺失或该协议探测失败时再试另一种，避免状态文件过期导致误报不健康。
+set -u
+
+port="${MEBOX_APP_PORT:-8080}"
+data_dir="${MEBOX_APP_DATA_DIR:-/data}"
+
+scheme_file="$data_dir/.http_scheme"
+scheme=""
+if [ -r "$scheme_file" ]; then
+	scheme="$(tr -d '[:space:]' <"$scheme_file")"
+fi
+if [ "$scheme" = "https" ]; then
+	first="https"
+	second="http"
+else
+	first="http"
+	second="https"
+fi
+
+for proto in "$first" "$second"; do
+	# --no-check-certificate：用户常用自签证书或 IP 访问，校验域名必然会失败。
+	if busybox wget -q --spider -T 3 --no-check-certificate \
+		"$proto://127.0.0.1:$port/api/health" >/dev/null 2>&1; then
+		exit 0
+	fi
+done
+
+exit 1

internal/config/defaults.go

@@ -33,6 +33,7 @@ func setDefaults(v *viper.Viper) {
 	v.SetDefault("database.cache_size", -40000)
 	v.SetDefault("database.max_open_conns", defaultDatabaseMaxOpenConns)
 	v.SetDefault("database.max_idle_conns", defaultDatabaseMaxIdleConns)
+	v.SetDefault("database.postgres_trgm_index", true)
 
 	v.SetDefault("secrets.jwt_secret", "")
 

internal/config/types.go

@@ -50,7 +50,8 @@ type AppConfig struct {
 	DataDir string `mapstructure:"data_dir"`
 	WebDir  string `mapstructure:"web_dir"`
 	// HTTPSEnabled 是否仅通过 HTTPS 提供访问。启用时必须同时配置
-	// SSLCert / SSLKey（或 SSLCertPath / SSLKeyPath），保存后服务会热切换到 HTTPS。
+	// SSLCert / SSLKey（或 SSLCertPath / SSLKeyPath），保存后服务会热切换到 HTTPS；
+	// 同一端口上的明文 HTTP 请求会被跳转到 HTTPS，而不是直接拒绝。
 	HTTPSEnabled bool `mapstructure:"https_enabled"`
 	// SSLCert 是 PEM 编码的 SSL 证书内容。
 	SSLCert string `mapstructure:"ssl_cert"`
@@ -86,6 +87,10 @@ type DatabaseConfig struct {
 	CacheSize    int    `mapstructure:"cache_size"`
 	MaxOpenConns int    `mapstructure:"max_open_conns"`
 	MaxIdleConns int    `mapstructure:"max_idle_conns"`
+	// PostgresTrgmIndex 决定是否用 pg_trgm 的 GIN 三元组索引替代 SQLite 的
+	// FTS5 全文索引。开启后 ILIKE '%term%' 搜索能走索引而不是整表扫描，代价是
+	// media 表写入时要多维护 4 个 GIN 索引。仅对 PostgreSQL 生效。
+	PostgresTrgmIndex bool `mapstructure:"postgres_trgm_index"`
 }
 
 // SecretsConfig 保存 JWT / 第三方 API 密钥（不要提交值）。

internal/database/compatibility_test.go

@@ -0,0 +1,238 @@
+package database
+
+import (
+	"context"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/google/uuid"
+	"github.com/jackc/pgx/v5"
+	"github.com/jackc/pgx/v5/stdlib"
+	"gorm.io/driver/postgres"
+	"gorm.io/gorm"
+	"gorm.io/gorm/logger"
+
+	"github.com/truewhile/MeBox/internal/config"
+	"github.com/truewhile/MeBox/internal/model"
+	"github.com/truewhile/MeBox/internal/repository"
+)
+
+// PostgreSQL tests use a disposable schema, never the caller's existing tables.
+// Set MEBOX_TEST_POSTGRES_DSN to include a real PostgreSQL server in local runs.
+func compatibilityDB(t *testing.T, dialect string) *gorm.DB {
+	t.Helper()
+	var db *gorm.DB
+	if dialect == "sqlite" {
+		cfg := &config.Config{}
+		cfg.Database.Type = dialect
+		cfg.Database.DBPath = filepath.Join(t.TempDir(), "test.db")
+		cfg.Database.WALMode = true
+		cfg.Database.BusyTimeout = 5000
+		cfg.Database.MaxOpenConns = 8
+		var err error
+		db, err = Open(cfg, nil)
+		if err != nil {
+			t.Fatal(err)
+		}
+	} else {
+		dsn := os.Getenv("MEBOX_TEST_POSTGRES_DSN")
+		if dsn == "" {
+			t.Skip("set MEBOX_TEST_POSTGRES_DSN to run PostgreSQL integration tests")
+		}
+		pgcfg, err := pgx.ParseConfig(dsn)
+		if err != nil {
+			t.Fatal("invalid PostgreSQL test configuration")
+		}
+		admin := stdlib.OpenDB(*pgcfg)
+		t.Cleanup(func() { _ = admin.Close() })
+		schemaName := "mebox_test_" + strings.ReplaceAll(uuid.NewString(), "-", "")
+		if _, err := admin.Exec("CREATE SCHEMA " + quoteIdent(schemaName)); err != nil {
+			t.Fatal(err)
+		}
+		t.Cleanup(func() {
+			if _, err := admin.Exec("DROP SCHEMA " + quoteIdent(schemaName) + " CASCADE"); err != nil {
+				t.Error(err)
+			}
+		})
+		pgcfg.RuntimeParams["search_path"] = schemaName + ",public"
+		conn := stdlib.OpenDB(*pgcfg)
+		t.Cleanup(func() { _ = conn.Close() })
+		db, err = gorm.Open(postgres.New(postgres.Config{Conn: conn}), &gorm.Config{})
+		if err != nil {
+			t.Fatal(err)
+		}
+	}
+	db.Logger = logger.Default.LogMode(logger.Silent)
+	if conn, err := db.DB(); err == nil {
+		t.Cleanup(func() { _ = conn.Close() })
+	}
+	if err := AutoMigrate(db); err != nil {
+		t.Fatal(err)
+	}
+	return db
+}
+
+func TestDatabaseCompatibility(t *testing.T) {
+	for _, dialect := range []string{"sqlite", "postgres"} {
+		t.Run(dialect, func(t *testing.T) {
+			t.Run("long_paths_and_restart", func(t *testing.T) {
+				db := compatibilityDB(t, dialect)
+				// Distinct CJK characters prevent TOAST compression from hiding
+				// the PostgreSQL btree entry size limit (> 2704 bytes).
+				var path strings.Builder
+				path.WriteString("/media/")
+				for i := range 980 {
+					path.WriteRune(rune(0x4e00 + i))
+				}
+				path.WriteString(".mkv")
+				media := model.Media{Path: path.String(), Title: "Long path"}
+				if err := db.Create(&media).Error; err != nil {
+					t.Fatalf("insert long media path: %v", err)
+				}
+				if err := db.Create(&model.Media{Path: media.Path}).Error; err == nil {
+					t.Fatal("duplicate path accepted")
+				}
+				if err := AutoMigrate(db); err != nil {
+					t.Fatalf("migrate existing database: %v", err)
+				}
+				var found model.Media
+				if err := db.Where("path = ?", media.Path).First(&found).Error; err != nil {
+					t.Fatalf("lookup long path: %v", err)
+				}
+			})
+			if dialect == "postgres" {
+				t.Run("upgrade_legacy_path_indexes", func(t *testing.T) {
+					db := compatibilityDB(t, dialect)
+					for _, unique := range []string{"UNIQUE ", ""} {
+						if err := db.Exec("DROP INDEX idx_media_path").Error; err != nil {
+							t.Fatal(err)
+						}
+						if err := db.Exec("CREATE " + unique + "INDEX idx_media_path ON media(path)").Error; err != nil {
+							t.Fatal(err)
+						}
+						for range 2 {
+							if err := ensureMediaPathIndex(db); err != nil {
+								t.Fatal(err)
+							}
+						}
+						isUnique, method, found, err := postgresIndexProperties(db, "idx_media_path")
+						if err != nil || !found || isUnique || method != "hash" {
+							t.Fatalf("legacy index not upgraded: unique=%v method=%s found=%v err=%v", isUnique, method, found, err)
+						}
+					}
+				})
+			}
+			t.Run("search_ranking", func(t *testing.T) {
+				db := compatibilityDB(t, dialect)
+				// Exercise LIKE fallback in both engines.
+				if dialect == "sqlite" {
+					resetMediaSearchIndex(db)
+				}
+				for i, title := range []string{"The Alpha story", "Alpha sequel", "Alpha"} {
+					row := model.Media{Base: model.Base{ID: fmt.Sprintf("m-%d", i), CreatedAt: time.Now().Add(time.Duration(i) * time.Minute)}, Title: title, Path: fmt.Sprintf("/media/%d.mkv", i)}
+					if err := db.Create(&row).Error; err != nil {
+						t.Fatal(err)
+					}
+				}
+				items, total, err := repository.New(db).Media.SearchFilteredPage(t.Context(), "Alpha", 0, 10, repository.MediaQueryFilter{IncludeNSFW: true})
+				if err != nil {
+					t.Fatal(err)
+				}
+				if total != 3 || len(items) != 3 || items[0].Title != "Alpha" || items[1].Title != "Alpha sequel" {
+					t.Fatalf("unexpected search ranking: total=%d, items=%v", total, items)
+				}
+			})
+			t.Run("concurrent_media_upsert", func(t *testing.T) {
+				db := compatibilityDB(t, dialect)
+				repo := repository.New(db).Media
+				const workers = 8
+				// Synchronize the initial lookups to force the insert race.
+				var lookups sync.WaitGroup
+				lookups.Add(workers)
+				if dialect == "postgres" {
+					if err := db.Callback().Query().After("gorm:after_query").Register("test:insert_race", func(tx *gorm.DB) {
+						if tx.Statement.Table == "media" && tx.Error == gorm.ErrRecordNotFound {
+							lookups.Done()
+							lookups.Wait()
+						}
+					}); err != nil {
+						t.Fatal(err)
+					}
+				}
+				ctx, cancel := context.WithTimeout(t.Context(), 15*time.Second)
+				defer cancel()
+				errs := make(chan error, workers)
+				for range workers {
+					go func() { errs <- repo.Upsert(ctx, &model.Media{Path: "/media/shared.mkv", Title: "Shared"}) }()
+				}
+				for range workers {
+					if err := <-errs; err != nil {
+						t.Errorf("concurrent upsert: %v", err)
+					}
+				}
+				var count int64
+				if err := db.Model(&model.Media{}).Count(&count).Error; err != nil || count != 1 {
+					t.Fatalf("media count=%d, err=%v", count, err)
+				}
+			})
+			t.Run("search_filters", func(t *testing.T) {
+				db := compatibilityDB(t, dialect)
+				rows := []model.Media{
+					{Base: model.Base{ID: "keep"}, LibraryID: "lib-1", Genres: "Action", Year: 2020, Rating: 8},
+					{Base: model.Base{ID: "other-library"}, LibraryID: "lib-2", Genres: "Action", Year: 2020, Rating: 8},
+					{Base: model.Base{ID: "other-genre"}, LibraryID: "lib-1", Genres: "Comedy", Year: 2020, Rating: 8},
+					{Base: model.Base{ID: "too-old"}, LibraryID: "lib-1", Genres: "Action", Year: 1990, Rating: 8},
+					{Base: model.Base{ID: "too-new"}, LibraryID: "lib-1", Genres: "Action", Year: 2026, Rating: 8},
+					{Base: model.Base{ID: "low-rating"}, LibraryID: "lib-1", Genres: "Action", Year: 2020, Rating: 2},
+					{Base: model.Base{ID: "watched"}, LibraryID: "lib-1", Genres: "Action", Year: 2020, Rating: 8},
+				}
+				for i := range rows {
+					rows[i].Title = "Alpha " + rows[i].ID
+					rows[i].Path = "/media/" + rows[i].ID + ".mkv"
+					if err := db.Create(&rows[i]).Error; err != nil {
+						t.Fatal(err)
+					}
+				}
+				if err := db.Create(&model.PlaybackHistory{UserID: "viewer", MediaID: "watched", Completed: true}).Error; err != nil {
+					t.Fatal(err)
+				}
+				items, total, err := repository.New(db).Media.SearchFilteredPage(t.Context(), "alpha", 0, 10, repository.MediaQueryFilter{
+					IncludeNSFW: true, LibraryID: "lib-1", Genres: []string{"Action"}, YearMin: 2000, YearMax: 2025, RatingMin: 5,
+					UnwatchedOnly: true, UnwatchedUserID: "viewer",
+				})
+				if err != nil {
+					t.Fatal(err)
+				}
+				if total != 1 || len(items) != 1 || items[0].ID != "keep" {
+					t.Fatalf("search ignored filters: total=%d, items=%v", total, items)
+				}
+			})
+			t.Run("migration_preserves_zero_values", func(t *testing.T) {
+				src := compatibilityDB(t, "sqlite")
+				dst := compatibilityDB(t, dialect)
+				user := model.User{Username: "disabled", PasswordHash: "hash"}
+				if err := src.Create(&user).Error; err != nil {
+					t.Fatal(err)
+				}
+				if err := src.Model(&user).Updates(map[string]any{"is_active": false, "hide_adult": false, "player_volume": 0, "danmaku_enabled": false}).Error; err != nil {
+					t.Fatal(err)
+				}
+				if _, _, err := copyModelTables(src, dst, 500); err != nil {
+					t.Fatal(err)
+				}
+				var got model.User
+				if err := dst.First(&got, "id = ?", user.ID).Error; err != nil {
+					t.Fatal(err)
+				}
+				if got.IsActive || got.HideAdult || got.PlayerVolume != 0 || got.DanmakuEnabled {
+					t.Fatalf("migration replaced stored zero values: active=%v hide_adult=%v volume=%v danmaku=%v", got.IsActive, got.HideAdult, got.PlayerVolume, got.DanmakuEnabled)
+				}
+			})
+		})
+	}
+}

internal/database/database_admin.go

@@ -165,7 +165,7 @@ func MigrateCurrentToPostgres(src *gorm.DB, targetDSN string, batchSize int, log
 	}
 
 	// 1. 初始化目标库 Schema、类型与索引
-	if err := AutoMigrate(targetDB); err != nil {
+	if err := AutoMigrateWithConfig(targetDB, nil, log); err != nil {
 		return nil, fmt.Errorf("auto migrate target PostgreSQL: %w", err)
 	}
 

internal/database/persisted_rows.go

@@ -0,0 +1,50 @@
+package database
+
+import (
+	"fmt"
+	"reflect"
+
+	"gorm.io/gorm"
+)
+
+// InsertPreservingValues copies persisted model rows without applying creation
+// hooks or defaults. Struct CreateInBatches treats false, 0 and empty strings as
+// missing values, which would reactivate disabled accounts during migration or
+// restore. Column maps preserve the exact values already stored in the source.
+func InsertPreservingValues(db *gorm.DB, rows any, batchSize int) error {
+	stmt := &gorm.Statement{DB: db}
+	if err := stmt.Parse(rows); err != nil {
+		return err
+	}
+	slice := reflect.Indirect(reflect.ValueOf(rows))
+	if slice.Kind() != reflect.Slice {
+		return fmt.Errorf("persisted rows must be a slice, got %T", rows)
+	}
+	if len(stmt.Schema.DBNames) == 0 {
+		return fmt.Errorf("persisted rows have no database columns")
+	}
+	values := make([]map[string]any, 0, slice.Len())
+	for i := 0; i < slice.Len(); i++ {
+		row := make(map[string]any, len(stmt.Schema.DBNames))
+		for _, name := range stmt.Schema.DBNames {
+			field := stmt.Schema.FieldsByDBName[name]
+			if field.Creatable {
+				row[name], _ = field.ValueOf(db.Statement.Context, slice.Index(i))
+			}
+		}
+		values = append(values, row)
+	}
+	if len(values) == 0 {
+		return nil
+	}
+	// Bound batches by each engine's bind-parameter limit.
+	parameterLimit := 65535
+	if isSQLite(db) {
+		parameterLimit = 32766
+	}
+	maxBatch := parameterLimit / len(stmt.Schema.DBNames)
+	if batchSize <= 0 || batchSize > maxBatch {
+		batchSize = maxBatch
+	}
+	return db.Table(stmt.Schema.Table).CreateInBatches(values, batchSize).Error
+}

internal/database/schema_migration.go

@@ -1,13 +1,25 @@
 package database
 
 import (
+	"fmt"
+	"strings"
+
+	"go.uber.org/zap"
 	"gorm.io/gorm"
 
+	"github.com/truewhile/MeBox/internal/config"
 	"github.com/truewhile/MeBox/internal/model"
 )
 
 // AutoMigrate creates tables for every model registered in the model package.
 func AutoMigrate(db *gorm.DB) error {
+	return AutoMigrateWithConfig(db, nil, nil)
+}
+
+// AutoMigrateWithConfig 与 AutoMigrate 相同，另外接受配置与日志：
+// 配置用来决定可选的 PostgreSQL 搜索索引，日志用来记录耗时步骤
+// （首次在已有大库上重建索引可能需要数十秒）。两者都可以为 nil。
+func AutoMigrateWithConfig(db *gorm.DB, cfg *config.Config, log *zap.Logger) error {
 	// 必须先于 AutoMigrate：旧库中可能已有重复的 (user_id, media_id) 历史行，
 	// 不去重会导致唯一索引 uniq_user_history 创建失败。
 	if err := dedupePlaybackHistories(db); err != nil {
@@ -32,6 +44,9 @@ func AutoMigrate(db *gorm.DB) error {
 	if err := ensurePerformanceIndexes(db); err != nil {
 		return err
 	}
+	if err := ensureMediaPathIndex(db); err != nil {
+		return err
+	}
 	if err := ensureLibraryRootsCompatibility(db); err != nil {
 		return err
 	}
@@ -44,6 +59,113 @@ func AutoMigrate(db *gorm.DB) error {
 		}
 		return ensureSQLiteQueryOptimizer(db)
 	}
+	return ensurePostgresSearchIndexes(db, cfg, log)
+}
+
+// ensureMediaPathIndex 建立 media.path 的唯一约束与等值查找索引。
+//
+// 模型上不再用 gorm:"uniqueIndex"（见 model.Media.Path 的注释）：PostgreSQL 的
+// btree 单条索引项上限是 2704 字节，varchar(1024) 在中文路径下可达 3072 字节，
+// 对整列建唯一索引会让这类行插入时报
+//
+//	index row size ... exceeds btree version 4 maximum 2704
+//
+// 所以 PostgreSQL 改用 md5(path) 的表达式唯一索引（定长 32 字节，永远塞得下），
+// 同时用 hash 索引 idx_media_path 加速 path = ? 等值查找；非唯一 btree
+// 也有索引项长度限制，不能继续对完整 path 建 btree。SQLite 沿用整列唯一索引。
+func ensureMediaPathIndex(db *gorm.DB) error {
+	if db == nil || !db.Migrator().HasTable("media") {
+		return nil
+	}
+	if !isPostgres(db) {
+		return db.Exec(`CREATE UNIQUE INDEX IF NOT EXISTS idx_media_path ON media(path)`).Error
+	}
+	// 先建哈希唯一索引：此时旧索引还在，唯一性不会出现空窗。
+	if err := db.Exec(`CREATE UNIQUE INDEX IF NOT EXISTS idx_media_path_hash ON media(md5(path))`).Error; err != nil {
+		return err
+	}
+	unique, method, found, err := postgresIndexProperties(db, "idx_media_path")
+	if err != nil {
+		return err
+	}
+	if found && (unique || method != "hash") {
+		if err := db.Exec(`DROP INDEX IF EXISTS idx_media_path`).Error; err != nil {
+			return err
+		}
+	}
+	return db.Exec(`CREATE INDEX IF NOT EXISTS idx_media_path ON media USING hash(path)`).Error
+}
+
+// postgresIndexProperties 返回当前 schema 下索引的唯一性与访问方法。
+func postgresIndexProperties(db *gorm.DB, indexName string) (unique bool, method string, found bool, err error) {
+	// 用指针接 bool：没有匹配行时 Scan 不报错，nil 即代表索引不存在。
+	var row struct {
+		IsUnique *bool
+		Method   string
+	}
+	if err := db.Raw(`
+SELECT i.indisunique AS is_unique, a.amname AS method
+FROM pg_class c
+JOIN pg_index i ON i.indexrelid = c.oid
+JOIN pg_namespace n ON n.oid = c.relnamespace
+JOIN pg_am a ON a.oid = c.relam
+WHERE n.nspname = current_schema() AND c.relname = ?
+`, indexName).Scan(&row).Error; err != nil {
+		return false, "", false, err
+	}
+	if row.IsUnique == nil {
+		return false, "", false, nil
+	}
+	return *row.IsUnique, row.Method, true, nil
+}
+
+// ensurePostgresSearchIndexes 给 PostgreSQL 补上 SQLite FTS5 的替代品。
+//
+// 全文检索只在 SQLite 下存在（FTS5 虚表 + 触发器），PostgreSQL 下
+// media_search_repository 会退回 title/original_name/path/genres 的 ILIKE
+// 全表扫描。pg_trgm 的 GIN 三元组索引能让这类 '%term%' 模式（≥3 字符）走索引，
+// 把搜索从整表扫描降到索引扫描。
+//
+// 之所以可关闭：media 表写入非常频繁（每次扫描批量入库），多 4 个 GIN 索引会
+// 明显放大写入成本。默认开启以对齐 SQLite 的搜索能力，写入压力大的部署可以关掉。
+//
+// 整个过程是尽力而为：没有建扩展的权限、Postgres 版本不支持等情况下只记一条
+// 警告并继续启动，绝不让一个搜索加速项把服务挡在门外。
+func ensurePostgresSearchIndexes(db *gorm.DB, cfg *config.Config, log *zap.Logger) error {
+	if !isPostgres(db) {
+		return nil
+	}
+	if cfg != nil && !cfg.Database.PostgresTrgmIndex {
+		if log != nil {
+			log.Info("postgres trigram search index disabled; media search falls back to sequential scan")
+		}
+		return nil
+	}
+	if err := db.Exec(`CREATE EXTENSION IF NOT EXISTS pg_trgm`).Error; err != nil {
+		if log != nil {
+			log.Warn("create pg_trgm extension failed; media search falls back to sequential scan", zap.Error(err))
+		}
+		return nil
+	}
+	// 首次在已有大库上执行会重建索引（服务尚未开始接受请求），之后每次启动都因
+	// IF NOT EXISTS 直接跳过。
+	if log != nil {
+		log.Info("ensuring postgres trigram search indexes")
+	}
+	for _, stmt := range []string{
+		`CREATE INDEX IF NOT EXISTS idx_media_title_trgm ON media USING gin (title gin_trgm_ops)`,
+		`CREATE INDEX IF NOT EXISTS idx_media_original_name_trgm ON media USING gin (original_name gin_trgm_ops)`,
+		`CREATE INDEX IF NOT EXISTS idx_media_path_trgm ON media USING gin (path gin_trgm_ops)`,
+		`CREATE INDEX IF NOT EXISTS idx_media_genres_trgm ON media USING gin (genres gin_trgm_ops)`,
+	} {
+		if err := db.Exec(stmt).Error; err != nil {
+			if log != nil {
+				log.Warn("create trigram index failed; media search may fall back to sequential scan",
+					zap.String("statement", stmt), zap.Error(err))
+			}
+			return nil
+		}
+	}
 	return nil
 }
 
@@ -176,16 +298,35 @@ func ensurePostgresColumnCompatibility(db *gorm.DB) error {
 	if !isPostgres(db) {
 		return nil
 	}
-	statements := []string{
-		`ALTER TABLE media ALTER COLUMN container TYPE varchar(128)`,
-		`ALTER TABLE media ALTER COLUMN genres TYPE text`,
-		`ALTER TABLE media ALTER COLUMN series_id TYPE varchar(128)`,
-		`ALTER TABLE media ALTER COLUMN duplicate_of TYPE varchar(128)`,
-		`ALTER TABLE playback_histories ALTER COLUMN media_id TYPE varchar(128)`,
-		`ALTER TABLE favorites ALTER COLUMN media_id TYPE varchar(128)`,
-		`ALTER TABLE playlist_items ALTER COLUMN media_id TYPE varchar(128)`,
+	// 只在类型确实不匹配时才 ALTER。
+	//
+	// PostgreSQL 的 ALTER COLUMN ... TYPE 只要类型声明与目标不同就会整表重写并
+	// 重建相关索引，执行期间持有 ACCESS EXCLUSIVE 锁。此前这里每次启动都无条件
+	// 发这 7 条语句：类型其实早已一致，但几十万行的 media 表仍然被反复重写，
+	// 启动会卡住数分钟，且期间这张表连读都被阻塞。
+	changes := []struct {
+		table  string
+		column string
+		target string
+	}{
+		{"media", "container", "varchar(128)"},
+		{"media", "genres", "text"},
+		{"media", "series_id", "varchar(128)"},
+		{"media", "duplicate_of", "varchar(128)"},
+		{"playback_histories", "media_id", "varchar(128)"},
+		{"favorites", "media_id", "varchar(128)"},
+		{"playlist_items", "media_id", "varchar(128)"},
 	}
-	for _, stmt := range statements {
+	for _, change := range changes {
+		current, found, err := postgresColumnType(db, change.table, change.column)
+		if err != nil {
+			return err
+		}
+		if !found || current == change.target {
+			continue
+		}
+		stmt := "ALTER TABLE " + quoteIdent(change.table) +
+			" ALTER COLUMN " + quoteIdent(change.column) + " TYPE " + change.target
 		if err := db.Exec(stmt).Error; err != nil {
 			return err
 		}
@@ -193,6 +334,35 @@ func ensurePostgresColumnCompatibility(db *gorm.DB) error {
 	return nil
 }
 
+// postgresColumnType 返回列在当前 schema 下的规范化类型描述，例如
+// "varchar(128)" / "text"。列不存在时 found 为 false。
+func postgresColumnType(db *gorm.DB, table, column string) (value string, found bool, err error) {
+	var row struct {
+		DataType  string
+		MaxLength *int64
+	}
+	if err := db.Raw(`
+SELECT data_type AS data_type, character_maximum_length AS max_length
+FROM information_schema.columns
+WHERE table_schema = current_schema() AND table_name = ? AND column_name = ?
+`, table, column).Scan(&row).Error; err != nil {
+		return "", false, fmt.Errorf("inspect postgres column %s.%s: %w", table, column, err)
+	}
+	dataType := strings.ToLower(strings.TrimSpace(row.DataType))
+	if dataType == "" {
+		return "", false, nil
+	}
+	// information_schema 用的是 SQL 标准名，映射回 GORM 标签里的写法再比较。
+	switch dataType {
+	case "character varying", "varchar":
+		dataType = "varchar"
+	}
+	if row.MaxLength != nil && *row.MaxLength > 0 {
+		dataType = fmt.Sprintf("%s(%d)", dataType, *row.MaxLength)
+	}
+	return dataType, true, nil
+}
+
 func ensurePerformanceIndexes(db *gorm.DB) error {
 	statements := []string{
 		// 完整多级排序索引：媒体库分页与首页预览的 ORDER BY

internal/database/schema_migration_path_index_test.go

@@ -0,0 +1,54 @@
+package database
+
+import (
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"github.com/glebarez/sqlite"
+	"gorm.io/gorm"
+	"gorm.io/gorm/logger"
+
+	"github.com/truewhile/MeBox/internal/model"
+)
+
+// TestAutoMigrateCreatesMediaPathUniqueIndex 守住 media.path 的唯一约束。
+//
+// 模型上不再声明 gorm:"uniqueIndex"（PostgreSQL 的 btree 单条索引项上限是
+// 2704 字节，而 varchar(1024) 在中文路径下可达 3072 字节，对整列建唯一索引会让
+// 这类行插入直接失败），改由 AutoMigrate 里的 ensureMediaPathIndex 按方言显式
+// 建立。一旦这个调用被漏掉，扫描会静默产生重复的媒体行（同一文件被入库多次）。
+func TestAutoMigrateCreatesMediaPathUniqueIndex(t *testing.T) {
+	db, err := gorm.Open(sqlite.Open(filepath.Join(t.TempDir(), "mebox.db")), &gorm.Config{
+		Logger: logger.Default.LogMode(logger.Silent),
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	// 必须在 TempDir 清理之前关掉连接，否则 Windows 上删除库文件会失败。
+	// t.Cleanup 是后进先出：这里注册的关闭早于 t.TempDir 注册的删除。
+	if sqlDB, err := db.DB(); err == nil {
+		t.Cleanup(func() { _ = sqlDB.Close() })
+	}
+	if err := AutoMigrate(db); err != nil {
+		t.Fatalf("AutoMigrate: %v", err)
+	}
+
+	var definition string
+	if err := db.Raw(`SELECT COALESCE(sql, '') FROM sqlite_master WHERE type = 'index' AND name = 'idx_media_path'`).Scan(&definition).Error; err != nil {
+		t.Fatal(err)
+	}
+	if definition == "" {
+		t.Fatal("AutoMigrate did not create idx_media_path")
+	}
+	if !strings.Contains(strings.ToUpper(definition), "UNIQUE") {
+		t.Fatalf("idx_media_path is not unique: %q", definition)
+	}
+
+	if err := db.Create(&model.Media{Base: model.Base{ID: "m-1"}, Path: "/media/a.mkv"}).Error; err != nil {
+		t.Fatal(err)
+	}
+	if err := db.Create(&model.Media{Base: model.Base{ID: "m-2"}, Path: "/media/a.mkv"}).Error; err == nil {
+		t.Fatal("a duplicate media path was accepted; the unique constraint is not enforced")
+	}
+}

internal/database/sqlite_migration_copy.go

@@ -78,7 +78,7 @@ func copyModelTables(src, target *gorm.DB, batchSize int) (map[string]int64, int
 			if filtered.Len() > 0 {
 				filteredPtr := reflect.New(filtered.Type())
 				filteredPtr.Elem().Set(filtered)
-				if err := target.Clauses(clause.OnConflict{DoNothing: true}).CreateInBatches(filteredPtr.Interface(), batchSize).Error; err != nil {
+				if err := InsertPreservingValues(target.Clauses(clause.OnConflict{DoNothing: true}), filteredPtr.Interface(), batchSize); err != nil {
 					return tableCounts, totalCopied, fmt.Errorf("copy sqlite table %s: %w", table, err)
 				}
 				copiedForTable += int64(filtered.Len())

internal/database/sqlite_runtime.go

@@ -104,19 +104,18 @@ func (g *sqliteWriteGate) Lock(ctx context.Context) (*sqliteGateHolder, error) {
 	if ctx == nil {
 		ctx = context.Background()
 	}
-	// ctx 取消时唤醒等待者（cond 无法感知 ctx，用旁路 goroutine 广播）。
-	if done := ctx.Done(); done != nil {
-		stop := make(chan struct{})
-		defer close(stop)
-		go func() {
-			select {
-			case <-done:
-				g.cond.Broadcast()
-			case <-stop:
-			}
-		}()
+	// 广播也必须持有同一把锁，否则取消可能落在检查与 Wait 之间而丢失。
+	wake := func() {
+		g.mu.Lock()
+		defer g.mu.Unlock()
+		g.cond.Broadcast()
 	}
+	stopCancel := context.AfterFunc(ctx, wake)
+	defer stopCancel()
 	for {
+		if ctx.Err() != nil {
+			return nil, ctx.Err()
+		}
 		if g.owner == nil {
 			holder := &sqliteGateHolder{
 				id:       sqliteGateHolderSeq.Add(1),
@@ -125,16 +124,17 @@ func (g *sqliteWriteGate) Lock(ctx context.Context) (*sqliteGateHolder, error) {
 			g.owner = holder
 			return holder, nil
 		}
-		if ctx.Err() != nil {
-			return nil, ctx.Err()
-		}
-		if time.Since(g.owner.acquired) > sqliteGateHoldLimit {
+		remaining := sqliteGateHoldLimit - time.Since(g.owner.acquired)
+		if remaining <= 0 {
 			// 持有者疑似 panic 泄漏（After 回调未执行）：强制回收。
 			g.owner = nil
 			g.cond.Broadcast()
 			continue
 		}
+		// 无后续写入或取消时，也要在租约到期后主动唤醒等待者。
+		timer := time.AfterFunc(remaining, wake)
 		g.cond.Wait()
+		timer.Stop()
 	}
 }
 

internal/database/sqlite_runtime_test.go

@@ -0,0 +1,79 @@
+package database
+
+import (
+	"context"
+	"errors"
+	"testing"
+	"time"
+)
+
+func TestSQLiteWriteGateRejectsCancelledContext(t *testing.T) {
+	ctx, cancel := context.WithCancel(t.Context())
+	cancel()
+	gate := newSQLiteWriteGate()
+	holder, err := gate.Lock(ctx)
+	defer gate.Unlock(holder)
+	if !errors.Is(err, context.Canceled) {
+		t.Fatalf("Lock error=%v, want context.Canceled", err)
+	}
+}
+
+func TestSQLiteWriteGateWakesForExpiredHolder(t *testing.T) {
+	gate := newSQLiteWriteGate()
+	holder, err := gate.Lock(t.Context())
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer gate.Unlock(holder)
+	// Make the lease expire shortly, without waiting a full minute.
+	gate.mu.Lock()
+	holder.acquired = time.Now().Add(-sqliteGateHoldLimit + 20*time.Millisecond)
+	gate.mu.Unlock()
+	done := make(chan *sqliteGateHolder, 1)
+	go func() {
+		next, _ := gate.Lock(t.Context())
+		done <- next
+	}()
+	select {
+	case next := <-done:
+		gate.Unlock(holder) // A stale holder must not unlock its successor.
+		gate.mu.Lock()
+		owner := gate.owner
+		gate.mu.Unlock()
+		gate.Unlock(next)
+		if next == nil || owner != next {
+			t.Fatal("expired holder released the new owner's lock")
+		}
+	case <-time.After(time.Second):
+		gate.Unlock(holder)
+		gate.Unlock(<-done)
+		t.Fatal("waiter was not woken when the holder expired")
+	}
+}
+
+func TestSQLiteWriteGateWakesForCancellation(t *testing.T) {
+	gate := newSQLiteWriteGate()
+	holder, err := gate.Lock(t.Context())
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer gate.Unlock(holder)
+	ctx, cancel := context.WithCancel(t.Context())
+	done := make(chan error, 1)
+	go func() {
+		next, err := gate.Lock(ctx)
+		gate.Unlock(next)
+		done <- err
+	}()
+	cancel()
+	select {
+	case err := <-done:
+		if !errors.Is(err, context.Canceled) {
+			t.Fatalf("Lock error=%v, want context.Canceled", err)
+		}
+	case <-time.After(time.Second):
+		gate.Unlock(holder)
+		<-done
+		t.Fatal("waiter was not woken for cancellation")
+	}
+}

internal/handler/media.go

@@ -810,6 +810,20 @@ func streamHandler(svc *service.Container) gin.HandlerFunc {
 				c.JSON(http.StatusNotFound, gin.H{"error": "not found"})
 				return
 			}
+			// 远程条目可能是「整部剧/季/目录」：这类条目自身没有媒体源，远程 Emby 对它
+			// 们的 /Videos/{id}/stream 会直接回 500（Unable to cast object of type
+			// 'Series' to type 'IHasMediaSources'）。网页端、外部播放器、第三方客户端都
+			// 可能把这种 ID 送进来，这里统一改投第一集：302 让客户端自己跟过去，播放器
+			// 原有的 token / proxy 查询参数一并保留。
+			if playable := svc.EmbyRemote.PlayableRemoteID(ctx, acct, remoteID); playable != "" && playable != remoteID {
+				target := "/api/stream/" + service.EncodeEmbyRemoteID(mountID, playable)
+				if rawQuery := c.Request.URL.RawQuery; rawQuery != "" {
+					target += "?" + rawQuery
+				}
+				setRedirectNoStoreHeaders(c)
+				c.Redirect(http.StatusFound, target)
+				return
+			}
 			// 远程 Emby 条目的同源转发：VR 全景需要浏览器读帧（见 stream_proxy.go），
 			// 与挂载账号的 proxy_play 开关等价。
 			if mount.ProxyPlay || wantSameOriginProxy(c) {

internal/handler/media_search_test.go

@@ -0,0 +1,111 @@
+package handler
+
+import (
+	"encoding/json"
+	"fmt"
+	"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/middleware"
+	"github.com/truewhile/MeBox/internal/model"
+	"github.com/truewhile/MeBox/internal/repository"
+	"github.com/truewhile/MeBox/internal/service"
+)
+
+func TestSearchMediaSeriesKeepsRemoteLibrariesSeparate(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	upstream := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		_ = json.NewEncoder(w).Encode(map[string]any{
+			"TotalRecordCount": 1,
+			"Items": []map[string]any{{
+				"Id": "same-series", "Name": "龙珠改", "Type": "Series",
+				"ProviderIds": map[string]string{"Tmdb": "61709"},
+			}},
+		})
+	}))
+	defer upstream.Close()
+
+	db, err := gorm.Open(sqlite.Open(":memory:"), &gorm.Config{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := db.AutoMigrate(&model.Library{}, &model.Media{}, &model.StrmAccount{}, &model.EmbyMount{}, &model.Setting{}, &model.User{}, &model.PlayProfile{}); err != nil {
+		t.Fatal(err)
+	}
+	repos := repository.New(db)
+	lib := model.Library{Name: "本地番剧", Path: "/media/anime", Type: "anime", Enabled: true}
+	if err := repos.Library.Create(t.Context(), &lib); err != nil {
+		t.Fatal(err)
+	}
+	rows := []model.Media{
+		{Base: model.Base{ID: "local-1"}, LibraryID: lib.ID, Title: "龙珠改", Path: "/media/anime/龙珠改/Season 1/ep1.mkv", SeasonNum: 1, EpisodeNum: 1, TMDbID: 61709},
+		{Base: model.Base{ID: "local-2"}, LibraryID: lib.ID, Title: "龙珠改", Path: "/media/anime/龙珠改/Season 1/ep2.mkv", SeasonNum: 1, EpisodeNum: 2, TMDbID: 61709},
+	}
+	if err := db.Create(&rows).Error; err != nil {
+		t.Fatal(err)
+	}
+	rawConfig, _ := json.Marshal(map[string]string{"url": upstream.URL, "token": "fake-token"})
+	acct := model.StrmAccount{Base: model.Base{ID: "account"}, Name: "Emby", Provider: model.StrmProviderEmbyRemote, Config: string(rawConfig), Enabled: true}
+	if err := repos.StrmAccount.Create(t.Context(), &acct); err != nil {
+		t.Fatal(err)
+	}
+	expectedIDs := map[string]string{lib.ID: "local-1"}
+	for _, suffix := range []string{"a", "b"} {
+		mount := model.EmbyMount{Base: model.Base{ID: "mount-" + suffix}, AccountID: acct.ID, RemoteViewID: "view-" + suffix, RemoteViewName: "番剧" + suffix, CollectionType: "tvshows", Enabled: true}
+		if err := repos.EmbyMount.Create(t.Context(), &mount); err != nil {
+			t.Fatal(err)
+		}
+		expectedIDs[service.EncodeEmbyRemoteID(mount.ID, mount.RemoteViewID)] = service.EncodeEmbyRemoteID(mount.ID, "same-series")
+	}
+	svc := &service.Container{
+		Repo:       repos,
+		Media:      service.NewMediaService(&config.Config{}, zap.NewNop(), repos),
+		EmbyRemote: service.NewEmbyRemoteService(&config.Config{}, zap.NewNop(), repos, service.NewCryptoService("", zap.NewNop())),
+	}
+
+	request := func(params string) ([]model.Media, int) {
+		t.Helper()
+		w := httptest.NewRecorder()
+		c, _ := gin.CreateTestContext(w)
+		c.Set(middleware.CtxUserID, "admin")
+		c.Set(middleware.CtxUserRole, "admin")
+		c.Request = httptest.NewRequest(http.MethodGet, "/api/media?q=龙珠改&group_series=1&"+params, nil)
+		searchMediaHandler(svc)(c)
+		if w.Code != http.StatusOK {
+			t.Fatalf("search status=%d, body=%s", w.Code, w.Body.String())
+		}
+		var response struct {
+			Items []model.Media `json:"items"`
+			Total int           `json:"total"`
+		}
+		if err := json.Unmarshal(w.Body.Bytes(), &response); err != nil {
+			t.Fatal(err)
+		}
+		return response.Items, response.Total
+	}
+
+	items, _ := request("limit=8")
+	if len(items) != len(expectedIDs) {
+		t.Fatalf("got %d results, want one local series and two remote series: %#v", len(items), items)
+	}
+	seen := make(map[string]bool)
+	for page := 1; page <= len(expectedIDs); page++ {
+		items, total := request(fmt.Sprintf("page=%d&page_size=1", page))
+		if total != len(expectedIDs) || len(items) != 1 {
+			t.Fatalf("page %d: total=%d items=%#v, want total=3 and one item", page, total, items)
+		}
+		item := items[0]
+		if expectedIDs[item.DisplayLibraryID] != item.ID || seen[item.DisplayLibraryID] || item.DisplayLibraryName == "" {
+			t.Fatalf("page %d lost or repeated a library's series identity: %#v", page, item)
+		}
+		seen[item.DisplayLibraryID] = true
+	}
+}

internal/handler/media_test.go

@@ -26,7 +26,9 @@ import (
 )
 
 func TestQueueLibraryRootScanImportsNewLibraryMedia(t *testing.T) {
-	db, err := gorm.Open(sqlite.Open(":memory:"), &gorm.Config{})
+	// The scanner and polling loop use separate connections; :memory: gives
+	// each connection its own empty database.
+	db, err := gorm.Open(sqlite.Open(filepath.Join(t.TempDir(), "scan.db")+"?_pragma=journal_mode(WAL)&_pragma=busy_timeout(5000)"), &gorm.Config{})
 	if err != nil {
 		t.Fatal(err)
 	}

internal/handler/stream_emby_remote_redirect_test.go

@@ -0,0 +1,200 @@
+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"
+)
+
+// 首页轮播/剧集卡片交给播放器的是「整部剧」的伪装 ID，它自身没有媒体源：远程 Emby
+// 对它的取流请求会回 500（Unable to cast 'Series' to 'IHasMediaSources'）。这里锁定
+// 播放接口的行为——改投第一集，且客户端原有的查询参数（token / profile / proxy）
+// 必须原样带过去，否则播放器跟过去之后会掉鉴权。
+func TestStreamRedirectsRemoteContainerToPlayableEpisode(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	upstream := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		switch r.URL.Path {
+		case "/emby/Users/uid-1/Items/series-100":
+			_ = json.NewEncoder(w).Encode(map[string]any{
+				"Id": "series-100", "Name": "某部新番", "Type": "Series", "IsFolder": true,
+			})
+		case "/emby/Users/uid-1/Items":
+			_ = json.NewEncoder(w).Encode(map[string]any{
+				"Items": []map[string]any{
+					map[string]any{
+						"Id": "ep-1", "Name": "第1集", "Type": "Episode", "SeriesId": "series-100",
+						"ParentIndexNumber": 1, "IndexNumber": 1,
+						"MediaSources": []any{map[string]any{"Id": "ep-1-src"}},
+					},
+				},
+			})
+		default:
+			w.WriteHeader(http.StatusNotFound)
+		}
+	}))
+	defer upstream.Close()
+
+	db, err := gorm.Open(sqlite.Open(":memory:"), &gorm.Config{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := db.AutoMigrate(
+		&model.User{}, &model.UserPermission{}, &model.Setting{}, &model.Library{}, &model.Media{},
+		&model.StrmAccount{}, &model.EmbyMount{},
+	); err != nil {
+		t.Fatal(err)
+	}
+	repos := repository.New(db)
+	cfg := &config.Config{}
+	cfg.Secrets.JWTSecret = "stream-redirect-secret"
+	log := zap.NewNop()
+
+	remote := service.NewEmbyRemoteService(cfg, log, repos, service.NewCryptoService("", log))
+	svc := &service.Container{Cfg: cfg, Repo: repos, EmbyRemote: remote}
+
+	rawConfig, _ := json.Marshal(map[string]string{
+		"url":            upstream.URL,
+		"api_key":        "test-api-key",
+		"remote_user_id": "uid-1",
+	})
+	if err := repos.StrmAccount.Create(t.Context(), &model.StrmAccount{
+		Base:     model.Base{ID: "acct-1"},
+		Name:     "远程 Emby",
+		Provider: model.StrmProviderEmbyRemote,
+		Config:   string(rawConfig),
+		Enabled:  true,
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if err := repos.EmbyMount.Create(t.Context(), &model.EmbyMount{
+		Base:           model.Base{ID: "mount-1"},
+		AccountID:      "acct-1",
+		RemoteViewID:   "view-1",
+		RemoteViewName: "新番连载",
+		CollectionType: "tvshows",
+		Enabled:        true,
+	}); err != nil {
+		t.Fatal(err)
+	}
+
+	router := gin.New()
+	router.GET("/api/stream/:id", streamHandler(svc))
+
+	req := httptest.NewRequest(
+		http.MethodGet,
+		"/api/stream/"+service.EncodeEmbyRemoteID("mount-1", "series-100")+"?token=tok-1&profile=p-1",
+		nil,
+	)
+	w := httptest.NewRecorder()
+	router.ServeHTTP(w, req)
+
+	if w.Code != http.StatusFound {
+		t.Fatalf("status = %d body=%s, want 302", w.Code, w.Body.String())
+	}
+	want := "/api/stream/" + service.EncodeEmbyRemoteID("mount-1", "ep-1") + "?token=tok-1&profile=p-1"
+	if got := w.Header().Get("Location"); got != want {
+		t.Fatalf("Location = %q, want %q", got, want)
+	}
+}
+
+// 本身可播的远程条目（单集/电影）不能被改写，必须照原样进入原来的代理/直连分支。
+func TestStreamKeepsPlayableRemoteItem(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	var streamCalls int
+	upstream := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		switch r.URL.Path {
+		case "/emby/Users/uid-1/Items/ep-1":
+			_ = json.NewEncoder(w).Encode(map[string]any{
+				"Id": "ep-1", "Name": "第1集", "Type": "Episode", "SeriesId": "series-100",
+				"ParentIndexNumber": 1, "IndexNumber": 1,
+				"MediaSources": []any{map[string]any{"Id": "ep-1-src"}},
+			})
+		case "/emby/Videos/ep-1/stream":
+			streamCalls++
+			w.Header().Set("Content-Type", "video/mp4")
+			w.WriteHeader(http.StatusOK)
+			_, _ = w.Write([]byte("bytes"))
+		default:
+			w.WriteHeader(http.StatusNotFound)
+		}
+	}))
+	defer upstream.Close()
+
+	db, err := gorm.Open(sqlite.Open(":memory:"), &gorm.Config{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := db.AutoMigrate(
+		&model.User{}, &model.UserPermission{}, &model.Setting{}, &model.Library{}, &model.Media{},
+		&model.StrmAccount{}, &model.EmbyMount{},
+	); err != nil {
+		t.Fatal(err)
+	}
+	repos := repository.New(db)
+	cfg := &config.Config{}
+	cfg.Secrets.JWTSecret = "stream-keep-secret"
+	log := zap.NewNop()
+
+	remote := service.NewEmbyRemoteService(cfg, log, repos, service.NewCryptoService("", log))
+	emby := service.NewEmbyService(cfg, log, repos).SetEmbyRemote(remote)
+	svc := &service.Container{Cfg: cfg, Repo: repos, EmbyRemote: remote, Emby: emby}
+
+	rawConfig, _ := json.Marshal(map[string]string{
+		"url":            upstream.URL,
+		"api_key":        "test-api-key",
+		"remote_user_id": "uid-1",
+	})
+	if err := repos.StrmAccount.Create(t.Context(), &model.StrmAccount{
+		Base:     model.Base{ID: "acct-1"},
+		Name:     "远程 Emby",
+		Provider: model.StrmProviderEmbyRemote,
+		Config:   string(rawConfig),
+		Enabled:  true,
+	}); err != nil {
+		t.Fatal(err)
+	}
+	// proxy_play=true：直接走反向代理分支，便于断言确实取到了流。
+	if err := repos.EmbyMount.Create(t.Context(), &model.EmbyMount{
+		Base:           model.Base{ID: "mount-1"},
+		AccountID:      "acct-1",
+		RemoteViewID:   "view-1",
+		RemoteViewName: "新番连载",
+		CollectionType: "tvshows",
+		ProxyPlay:      true,
+		Enabled:        true,
+	}); err != nil {
+		t.Fatal(err)
+	}
+
+	router := gin.New()
+	router.GET("/api/stream/:id", streamHandler(svc))
+
+	req := httptest.NewRequest(http.MethodGet, "/api/stream/"+service.EncodeEmbyRemoteID("mount-1", "ep-1"), nil)
+	w := httptest.NewRecorder()
+	router.ServeHTTP(w, req)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("status = %d body=%s, want 200", w.Code, w.Body.String())
+	}
+	if got := w.Body.String(); got != "bytes" {
+		t.Fatalf("body = %q, want the proxied stream", got)
+	}
+	if streamCalls != 1 {
+		t.Fatalf("upstream stream calls = %d, want 1", streamCalls)
+	}
+}

internal/model/library_media.go

@@ -26,13 +26,17 @@ type LibraryRoot struct {
 // Media 是单个可播放项。剧集链接到 SeriesID；电影 SeriesID == ""。
 type Media struct {
 	Base
-	LibraryID       string  `gorm:"index;size:36;index:idx_media_library_release,priority:1" json:"library_id"`
-	LibraryRootID   string  `gorm:"index;size:36" json:"library_root_id,omitempty"`
-	SeriesID        string  `gorm:"index;size:128" json:"series_id,omitempty"`
-	Title           string  `gorm:"size:255;not null" json:"title"`
-	OriginalName    string  `gorm:"size:255" json:"original_name,omitempty"`
-	EpisodeTitle    string  `gorm:"size:255" json:"episode_title,omitempty"`
-	Path            string  `gorm:"uniqueIndex;size:1024;not null" json:"path"`
+	LibraryID     string `gorm:"index;size:36;index:idx_media_library_release,priority:1" json:"library_id"`
+	LibraryRootID string `gorm:"index;size:36" json:"library_root_id,omitempty"`
+	SeriesID      string `gorm:"index;size:128" json:"series_id,omitempty"`
+	Title         string `gorm:"size:255;not null" json:"title"`
+	OriginalName  string `gorm:"size:255" json:"original_name,omitempty"`
+	EpisodeTitle  string `gorm:"size:255" json:"episode_title,omitempty"`
+	// Path 的唯一性由 database 包按方言显式建索引，这里刻意不用 uniqueIndex：
+	// PostgreSQL 的 btree 单条索引项上限 2704 字节，而 varchar(1024) 在中文路径
+	// 下可达 3072 字节，直接对整列建唯一索引会让长路径的扫描插入直接失败
+	// （SQLite 没有这个限制）。见 database.ensureMediaPathIndex。
+	Path            string  `gorm:"size:1024;not null" json:"path"`
 	RelativePath    string  `gorm:"size:1024" json:"relative_path,omitempty"`
 	SizeBytes       int64   `json:"size_bytes"`
 	DurationSec     int     `json:"duration_sec"`

internal/repository/dialect.go

@@ -0,0 +1,21 @@
+package repository
+
+import "gorm.io/gorm"
+
+// LikeOperator 返回当前方言下「与 SQLite LIKE 行为一致」的模式匹配运算符。
+//
+// SQLite 的 LIKE 对 ASCII 不区分大小写（默认编译选项），PostgreSQL 的 LIKE 区分
+// 大小写、只有 ILIKE 才不区分。同一个搜索词在两种方言下必须给出同样的结果集，
+// 否则换库之后「inception」就再也搜不到「Inception」。
+//
+// 供构造原始 SQL 片段的调用方（含 service 层）使用。
+func LikeOperator(db *gorm.DB) string {
+	return likeOperator(db)
+}
+
+func likeOperator(db *gorm.DB) string {
+	if db != nil && db.Dialector != nil && db.Dialector.Name() == "postgres" {
+		return "ILIKE"
+	}
+	return "LIKE"
+}

internal/repository/dialect_test.go

@@ -0,0 +1,54 @@
+package repository
+
+import (
+	"strings"
+	"testing"
+
+	"gorm.io/driver/postgres"
+	"gorm.io/gorm"
+)
+
+// TestSegmentPrewarmQueryBindsBooleanFound 守住 PostgreSQL 下片段预取查询的合法性。
+//
+// media_segment_fetches.found 是布尔列，而 PostgreSQL 没有 boolean = integer
+// 运算符：写成 f.found = 0 / = 1 会让计划任务里的片段预取固定报
+//
+//	operator does not exist: boolean = integer (SQLSTATE 42883)
+//
+// SQLite 一直能跑，所以这个错只在 PostgreSQL 上暴露（表现为日志里周期性出现
+// 该错误、片头片尾数据永远补不上）。这里断言语句里只能出现占位符绑定。
+func TestSegmentPrewarmQueryBindsBooleanFound(t *testing.T) {
+	query := segmentPrewarmCandidatesQuery
+	lower := strings.ToLower(query)
+	for _, forbidden := range []string{"found = 0", "found = 1", "found=0", "found=1", "found = true", "found = false"} {
+		if strings.Contains(lower, forbidden) {
+			t.Fatalf("预取查询把布尔列与字面量比较（%q），PostgreSQL 会直接报错：\n%s", forbidden, query)
+		}
+	}
+	if got := strings.Count(query, "f.found = ?"); got != 2 {
+		t.Fatalf("expected 2 placeholder-bound found comparisons, got %d:\n%s", got, query)
+	}
+}
+
+// TestLikeOperatorMatchesSQLiteCaseInsensitivity 守住搜索大小写行为的方言差异。
+//
+// SQLite 的 LIKE 对 ASCII 不区分大小写，PostgreSQL 的 LIKE 区分大小写：同一个
+// 搜索词在两种方言下必须给出同样的结果集，PostgreSQL 必须走 ILIKE。
+func TestLikeOperatorMatchesSQLiteCaseInsensitivity(t *testing.T) {
+	pg, err := gorm.Open(postgres.New(postgres.Config{
+		DSN: "postgres://mebox:mebox@127.0.0.1:5432/mebox?sslmode=disable",
+	}), &gorm.Config{
+		DryRun:                 true,
+		DisableAutomaticPing:   true,
+		SkipDefaultTransaction: true,
+	})
+	if err != nil {
+		t.Fatalf("open dry-run postgres: %v", err)
+	}
+	if got := LikeOperator(pg); got != "ILIKE" {
+		t.Fatalf("postgres LikeOperator = %q, want ILIKE", got)
+	}
+	if got := LikeOperator(nil); got != "LIKE" {
+		t.Fatalf("nil LikeOperator = %q, want LIKE", got)
+	}
+}

internal/repository/history_repository.go

@@ -33,6 +33,20 @@ func (r *HistoryRepository) UpsertProgress(ctx context.Context, h *model.Playbac
 	return r.upsert(ctx, h, true)
 }
 
+// historyColumn 生成「当前表」限定的列引用。
+//
+// PostgreSQL 的 ON CONFLICT DO UPDATE ... WHERE 里，未限定的列名会同时可解析到
+// 目标表和本次插入的行，于是报
+//
+//	column reference "session_started_at_ms" is ambiguous (SQLSTATE 42702)
+//
+// 结果是 POST /api/history 一律 500，观看进度与续播完全存不下来。SQLite 不做这项
+// 检查，所以这个错只在 PostgreSQL 上暴露。clause.CurrentTable 由 GORM 在构建语句
+// 时替换成真实表名，两种方言都能正确渲染。
+func historyColumn(name string) clause.Column {
+	return clause.Column{Table: clause.CurrentTable, Name: name}
+}
+
 func (r *HistoryRepository) upsert(ctx context.Context, h *model.PlaybackHistory, versioned bool) error {
 	updates := playbackHistoryAssignments(h)
 	if versioned {
@@ -49,16 +63,16 @@ func (r *HistoryRepository) upsert(ctx context.Context, h *model.PlaybackHistory
 		onConflict.Where = clause.Where{Exprs: []clause.Expression{
 			clause.Or(
 				clause.Lt{
-					Column: clause.Column{Name: "session_started_at_ms"},
+					Column: historyColumn("session_started_at_ms"),
 					Value:  h.SessionStartedAtMs,
 				},
 				clause.And(
 					clause.Eq{
-						Column: clause.Column{Name: "session_started_at_ms"},
+						Column: historyColumn("session_started_at_ms"),
 						Value:  h.SessionStartedAtMs,
 					},
 					clause.Lte{
-						Column: clause.Column{Name: "sequence"},
+						Column: historyColumn("sequence"),
 						Value:  h.Sequence,
 					},
 				),

internal/repository/history_repository_dialect_test.go

@@ -0,0 +1,65 @@
+package repository
+
+import (
+	"context"
+	"strings"
+	"testing"
+
+	"github.com/truewhile/MeBox/internal/model"
+)
+
+// TestHistoryProgressUpsertQualifiesConflictWhereColumns 守住 PostgreSQL 下进度上报
+// 的 upsert 语句合法性。
+//
+// ON CONFLICT DO UPDATE 的 WHERE 里如果不给列名加表限定，PostgreSQL 会因为这些列
+// 同时出现在本次插入的行里而拒绝该语句：
+//
+//	column reference "session_started_at_ms" is ambiguous (SQLSTATE 42702)
+//
+// 后果是 POST /api/history 全部 500，观看进度与续播完全存不下来。SQLite 不做这项
+// 检查（旧写法在 SQLite 下一直正常），所以只能在 PostgreSQL 方言下断言。
+func TestHistoryProgressUpsertQualifiesConflictWhereColumns(t *testing.T) {
+	repos, rec := newPostgresDryRunRepos(t)
+	ctx := context.Background()
+
+	if err := repos.History.UpsertProgress(ctx, &model.PlaybackHistory{
+		UserID:             "u-1",
+		MediaID:            "m-1",
+		PositionMs:         1_000,
+		DurationMs:         2_000,
+		SessionID:          "session-1",
+		SessionStartedAtMs: 111,
+		Sequence:           2,
+	}); err != nil {
+		t.Fatalf("build upsert: %v", err)
+	}
+
+	sql := rec.all()
+	if strings.TrimSpace(sql) == "" {
+		t.Fatal("no SQL captured from UpsertProgress")
+	}
+	if !strings.Contains(sql, "ON CONFLICT") {
+		t.Fatalf("expected an ON CONFLICT upsert, got:\n%s", sql)
+	}
+
+	// 只看 ON CONFLICT 之后的部分：SET 的赋值目标和 CASE 里的 duration_ms 引用
+	// 属于另一套语法规则，不在这里断言。
+	conflict := sql[strings.Index(sql, "ON CONFLICT"):]
+	whereIdx := strings.Index(conflict, " WHERE ")
+	if whereIdx < 0 {
+		t.Fatalf("ON CONFLICT clause has no DO UPDATE ... WHERE guard:\n%s", sql)
+	}
+	where := conflict[whereIdx:]
+
+	for _, want := range []string{
+		`"playback_histories"."session_started_at_ms"`,
+		`"playback_histories"."sequence"`,
+	} {
+		if !strings.Contains(where, want) {
+			t.Fatalf("DO UPDATE WHERE is missing the table-qualified %s; it is ambiguous on PostgreSQL:\n%s", want, sql)
+		}
+	}
+	if strings.Contains(where, `("session_started_at_ms"`) || strings.Contains(where, ` AND "sequence"`) {
+		t.Fatalf("DO UPDATE WHERE still references unqualified columns:\n%s", sql)
+	}
+}

internal/repository/media_repository.go

@@ -63,38 +63,38 @@ type MediaQueryFilter struct {
 
 func applyMediaQueryFilter(q *gorm.DB, filter MediaQueryFilter) *gorm.DB {
 	if !filter.IncludeNSFW {
-		q = q.Where("nsfw = ?", false)
+		q = q.Where("media.nsfw = ?", false)
 	}
 	if len(filter.HiddenLibraryIDs) > 0 {
-		q = q.Where("library_id NOT IN ?", filter.HiddenLibraryIDs)
+		q = q.Where("media.library_id NOT IN ?", filter.HiddenLibraryIDs)
 	}
 	if len(filter.AllowedLibraryIDs) > 0 {
-		q = q.Where("library_id IN ?", filter.AllowedLibraryIDs)
+		q = q.Where("media.library_id IN ?", filter.AllowedLibraryIDs)
 	}
 	if libraryID := strings.TrimSpace(filter.LibraryID); libraryID != "" {
-		q = q.Where("library_id = ?", libraryID)
+		q = q.Where("media.library_id = ?", libraryID)
 	}
 	if seriesID := strings.TrimSpace(filter.SeriesID); seriesID != "" {
-		q = q.Where("series_id = ?", seriesID)
+		q = q.Where("media.series_id = ?", seriesID)
 	}
 	if len(filter.Genres) > 0 {
 		q = q.Where(genreMatchClause(filter.Genres), genreMatchArgs(filter.Genres)...)
 	}
 	if filter.YearMin > 0 {
-		q = q.Where("year >= ?", filter.YearMin)
+		q = q.Where("media.year >= ?", filter.YearMin)
 	}
 	if filter.YearMax > 0 {
-		q = q.Where("year <= ?", filter.YearMax)
+		q = q.Where("media.year <= ?", filter.YearMax)
 	}
 	if filter.RatingMin > 0 {
-		q = q.Where("rating >= ?", filter.RatingMin)
+		q = q.Where("media.rating >= ?", filter.RatingMin)
 	}
 	if filter.UnwatchedOnly {
 		userID := strings.TrimSpace(filter.UnwatchedUserID)
 		// 没有用户上下文时忽略该条件：否则会把整个库筛成空，看起来像「坏了」。
 		if userID != "" {
 			q = q.Where(
-				"id NOT IN (SELECT media_id FROM playback_histories WHERE user_id = ? AND completed = ?)",
+				"media.id NOT IN (SELECT media_id FROM playback_histories WHERE user_id = ? AND completed = ?)",
 				userID, true,
 			)
 		}
@@ -113,7 +113,7 @@ func applyMediaQueryFilter(q *gorm.DB, filter MediaQueryFilter) *gorm.DB {
 func genreMatchClause(genres []string) string {
 	clauses := make([]string, 0, len(genres))
 	for range genres {
-		clauses = append(clauses, "',' || REPLACE(REPLACE(TRIM(genres), ' ', ''), '，', ',') || ',' LIKE ?")
+		clauses = append(clauses, "',' || REPLACE(REPLACE(TRIM(media.genres), ' ', ''), '，', ',') || ',' LIKE ?")
 	}
 	return "(" + strings.Join(clauses, " OR ") + ")"
 }
@@ -122,7 +122,7 @@ func genreMatchClause(genres []string) string {
 //
 // 必须与 genreMatchClause 的列端处理完全对称：
 //   - TRIM                   → TrimSpace
-//   - REPLACE(…, ' ', '')    → ReplaceAll(…, " ", "")   ← 多词类型（如 "Science Fiction"）
+//   - REPLACE(…, ' ', ”)    → ReplaceAll(…, " ", "")   ← 多词类型（如 "Science Fiction"）
 //   - REPLACE(…, '，', ',')  → ReplaceAll(…, "，", ",")
 func genreMatchArgs(genres []string) []any {
 	args := make([]any, 0, len(genres))

internal/repository/media_repository_upsert.go

@@ -7,6 +7,7 @@ import (
 	"strings"
 
 	"gorm.io/gorm"
+	"gorm.io/gorm/clause"
 
 	"github.com/truewhile/MeBox/internal/model"
 )
@@ -95,6 +96,9 @@ func (r *MediaRepository) UpsertBatchWithAliases(ctx context.Context, items []Me
 
 // indexByIDBestEffort 在事务提交后按 ID 回读最新行并同步搜索索引。
 func (r *MediaRepository) indexByIDBestEffort(ctx context.Context, ids []string) {
+	if _, ok := r.searchBackend.(MediaSearchSyncBackend); !ok {
+		return
+	}
 	for _, id := range ids {
 		if id == "" {
 			continue
@@ -193,14 +197,19 @@ func (r *MediaRepository) findOrCreateMediaByPath(ctx context.Context, db *gorm.
 		if m.ScrapeStatus == "" {
 			m.ScrapeStatus = "pending"
 		}
-		if createErr := db.WithContext(ctx).Create(m).Error; createErr == nil {
-			return *m, true, false, nil
-		} else if retryErr := db.WithContext(ctx).Unscoped().Where("path = ?", m.Path).First(&existing).Error; retryErr != nil {
-			return model.Media{}, false, false, createErr
-		} else {
-			// 并发插入竞态：重查已命中既有行，直接走更新分支。
-			return existing, false, false, nil
+		// PostgreSQL 唯一键异常会中止整个事务，之后重查也会失败。
+		// DO NOTHING 保持事务可用，并兼容 path 的表达式唯一索引。
+		result := db.WithContext(ctx).Clauses(clause.OnConflict{DoNothing: true}).Create(m)
+		if result.Error != nil {
+			return model.Media{}, false, false, result.Error
 		}
+		if result.RowsAffected > 0 {
+			return *m, true, false, nil
+		}
+		if err := db.WithContext(ctx).Unscoped().Where("path = ?", m.Path).First(&existing).Error; err != nil {
+			return model.Media{}, false, false, err
+		}
+		return existing, false, false, nil
 	}
 	if err != nil {
 		return model.Media{}, false, false, err

internal/repository/media_search_repository.go

@@ -6,6 +6,7 @@ import (
 	"unicode"
 
 	"gorm.io/gorm"
+	"gorm.io/gorm/clause"
 
 	"github.com/truewhile/MeBox/internal/model"
 )
@@ -86,7 +87,7 @@ func (r *MediaRepository) searchFilteredFTS(ctx context.Context, query string, o
 		Joins("JOIN media_search_fts ON media_search_fts.rowid = media.rowid").
 		Where("media.deleted_at IS NULL").
 		Where("media_search_fts MATCH ?", ftsQuery)
-	q = applyQualifiedMediaQueryFilter(q, filter)
+	q = applyMediaQueryFilter(q, filter)
 	if err := q.Count(&total).Error; err != nil {
 		return nil, 0, false
 	}
@@ -106,11 +107,8 @@ func (r *MediaRepository) searchFilteredLIKE(ctx context.Context, query string,
 	q := r.db.WithContext(ctx).Model(&model.Media{})
 	q = applyMediaQueryFilter(q, filter)
 	// SQLite 的 LIKE 对 ASCII 不区分大小写；Postgres 的 LIKE 区分大小写，
-	// 需用 ILIKE 保持两端搜索行为一致。
-	likeOp := "LIKE"
-	if r.db.Dialector != nil && r.db.Dialector.Name() == "postgres" {
-		likeOp = "ILIKE"
-	}
+	// 需用 ILIKE 保持两端搜索行为一致（见 dialect.go 的 likeOperator）。
+	likeOp := likeOperator(r.db)
 	terms := mediaSearchTerms(query)
 	for _, term := range terms {
 		like := "%" + escapeLike(term) + "%"
@@ -125,10 +123,10 @@ func (r *MediaRepository) searchFilteredLIKE(ctx context.Context, query string,
 	if query != "" {
 		prefix := escapeLike(query) + "%"
 		exact := query
-		q = q.Order(gorm.Expr(
+		q = q.Order(clause.OrderBy{Expression: gorm.Expr(
 			"CASE WHEN title = ? THEN 0 WHEN original_name = ? THEN 1 WHEN title "+likeOp+" ? ESCAPE '\\' THEN 2 WHEN original_name "+likeOp+" ? ESCAPE '\\' THEN 3 ELSE 4 END, created_at desc",
 			exact, exact, prefix, prefix,
-		))
+		)})
 	} else {
 		q = q.Order("created_at desc")
 	}
@@ -136,22 +134,6 @@ func (r *MediaRepository) searchFilteredLIKE(ctx context.Context, query string,
 	return items, total, err
 }
 
-func applyQualifiedMediaQueryFilter(q *gorm.DB, filter MediaQueryFilter) *gorm.DB {
-	if !filter.IncludeNSFW {
-		q = q.Where("media.nsfw = ?", false)
-	}
-	if len(filter.HiddenLibraryIDs) > 0 {
-		q = q.Where("media.library_id NOT IN ?", filter.HiddenLibraryIDs)
-	}
-	if len(filter.AllowedLibraryIDs) > 0 {
-		q = q.Where("media.library_id IN ?", filter.AllowedLibraryIDs)
-	}
-	if seriesID := strings.TrimSpace(filter.SeriesID); seriesID != "" {
-		q = q.Where("media.series_id = ?", seriesID)
-	}
-	return q
-}
-
 func mediaFTSQuery(query string) string {
 	terms := mediaSearchTerms(query)
 	if len(terms) == 0 {

internal/repository/media_segment_repository.go

@@ -86,21 +86,15 @@ func (r *MediaSegmentRepository) UpsertFetch(ctx context.Context, row *model.Med
 	return r.db.WithContext(ctx).Clauses(onConflict).Create(row).Error
 }
 
-// ListPrewarmCandidates returns media that should be refreshed from source:
-// no ledger, expired miss (fetched_at < missBefore), or expired hit (fetched_at < hitBefore).
-// Recently played items come first so hot titles recover coverage sooner.
-func (r *MediaSegmentRepository) ListPrewarmCandidates(
-	ctx context.Context,
-	source string,
-	missBefore, hitBefore time.Time,
-	limit int,
-) ([]model.Media, error) {
-	if r == nil || limit <= 0 {
-		return nil, nil
-	}
-	rows := make([]model.Media, 0, limit)
-	// 可查询：有 TMDb，且是剧集（有季集）或电影（无季集）。
-	err := r.db.WithContext(ctx).Raw(`
+// segmentPrewarmCandidatesQuery 是片段预取候选查询。
+//
+// 抽成包级常量是为了让方言相关的回归测试能直接断言语句文本：DryRun 模式下
+// GORM 不会把 Raw 语句交给 logger，无法从执行路径上捕获生成后的 SQL。
+//
+// 注意 found 必须用占位符绑定布尔值，不能写 f.found = 0 / = 1：
+// PostgreSQL 没有 boolean = integer 运算符，那样写会固定报 SQLSTATE 42883
+// （SQLite 宽容，所以只在 PostgreSQL 上暴露）。
+const segmentPrewarmCandidatesQuery = `
 SELECT m.*
 FROM media m
 LEFT JOIN media_segment_fetches f
@@ -119,11 +113,29 @@ WHERE m.deleted_at IS NULL
   )
   AND (
     f.id IS NULL
-    OR (f.found = 0 AND f.fetched_at < ?)
-    OR (f.found = 1 AND f.fetched_at < ?)
+    OR (f.found = ? AND f.fetched_at < ?)
+    OR (f.found = ? AND f.fetched_at < ?)
   )
 ORDER BY ph.last_played DESC
 LIMIT ?
-`, source, missBefore, hitBefore, limit).Scan(&rows).Error
+`
+
+// ListPrewarmCandidates returns media that should be refreshed from source:
+// no ledger, expired miss (fetched_at < missBefore), or expired hit (fetched_at < hitBefore).
+// Recently played items come first so hot titles recover coverage sooner.
+func (r *MediaSegmentRepository) ListPrewarmCandidates(
+	ctx context.Context,
+	source string,
+	missBefore, hitBefore time.Time,
+	limit int,
+) ([]model.Media, error) {
+	if r == nil || limit <= 0 {
+		return nil, nil
+	}
+	rows := make([]model.Media, 0, limit)
+	// 可查询：有 TMDb，且是剧集（有季集）或电影（无季集）。
+	err := r.db.WithContext(ctx).
+		Raw(segmentPrewarmCandidatesQuery, source, false, missBefore, true, hitBefore, limit).
+		Scan(&rows).Error
 	return rows, err
 }

internal/repository/reader_repository.go

@@ -6,6 +6,7 @@ import (
 	"errors"
 
 	"gorm.io/gorm"
+	"gorm.io/gorm/clause"
 
 	"github.com/truewhile/MeBox/internal/model"
 )
@@ -15,6 +16,17 @@ type ReaderRepository struct {
 	db *gorm.DB
 }
 
+// readerOrderBy 组装方言安全的 ORDER BY。
+//
+// 用 clause.Column 让 GORM 按当前方言给列名加引号（SQLite 反引号 / PostgreSQL 双引号），
+// 而不是把引号硬编码进排序字符串。reader_books.order、reader_chapters.index、
+// reader_replace_rules.order 都撞上了 SQL 关键字（order 在 PostgreSQL 里是保留字）：
+// 之前写成 .Order("`order` ASC")，SQLite 下正常，PostgreSQL 下反引号被词法分析成
+// 运算符，直接报 `syntax error at or near "order"`，书架 / 目录 / 替换规则整页 500。
+func readerOrderBy(columns ...clause.OrderByColumn) clause.OrderBy {
+	return clause.OrderBy{Columns: columns}
+}
+
 // ListSources 用户的书源列表（书源按用户独立，按 customOrder 排序）。
 func (r *ReaderRepository) ListSources(ctx context.Context, userID string) ([]model.ReaderBookSource, error) {
 	var out []model.ReaderBookSource
@@ -153,7 +165,11 @@ func (r *ReaderRepository) SaveSourceState(ctx context.Context, st *model.Reader
 // ListBooks 用户书架（按 order 排序）。
 func (r *ReaderRepository) ListBooks(ctx context.Context, userID string) ([]model.ReaderBook, error) {
 	var out []model.ReaderBook
-	err := r.db.WithContext(ctx).Where("user_id = ?", userID).Order("`order` ASC, updated_at DESC").Find(&out).Error
+	err := r.db.WithContext(ctx).Where("user_id = ?", userID).
+		Order(readerOrderBy(
+			clause.OrderByColumn{Column: clause.Column{Name: "order"}},
+			clause.OrderByColumn{Column: clause.Column{Name: "updated_at"}, Desc: true},
+		)).Find(&out).Error
 	return out, err
 }
 
@@ -231,7 +247,10 @@ func (r *ReaderRepository) SwitchBookOrigin(ctx context.Context, book *model.Rea
 // ListChapters 按序取章节。
 func (r *ReaderRepository) ListChapters(ctx context.Context, bookID string) ([]model.ReaderChapter, error) {
 	var out []model.ReaderChapter
-	err := r.db.WithContext(ctx).Where("book_id = ?", bookID).Order("`index` ASC").Find(&out).Error
+	err := r.db.WithContext(ctx).Where("book_id = ?", bookID).
+		Order(readerOrderBy(
+			clause.OrderByColumn{Column: clause.Column{Name: "index"}},
+		)).Find(&out).Error
 	return out, err
 }
 
@@ -365,9 +384,17 @@ func (r *ReaderRepository) CountChaptersByBook(ctx context.Context, bookIDs []st
 }
 
 // GetChapter 取指定章节。
+//
+// index 在 PostgreSQL 里是保留字，必须让 GORM 按方言加引号（SQLite 反引号 /
+// PostgreSQL 双引号）。此前写成硬编码的 "book_id = ? AND `index` = ?"，
+// PostgreSQL 会把反引号词法分析成运算符并报 syntax error at or near "`"。
 func (r *ReaderRepository) GetChapter(ctx context.Context, bookID string, index int) (*model.ReaderChapter, error) {
 	var out model.ReaderChapter
-	if err := r.db.WithContext(ctx).First(&out, "book_id = ? AND `index` = ?", bookID, index).Error; err != nil {
+	err := r.db.WithContext(ctx).
+		Where(clause.Eq{Column: clause.Column{Name: "book_id"}, Value: bookID}).
+		Where(clause.Eq{Column: clause.Column{Name: "index"}, Value: index}).
+		First(&out).Error
+	if err != nil {
 		return nil, err
 	}
 	return &out, nil
@@ -433,7 +460,11 @@ func (r *ReaderRepository) SaveBookGroups(ctx context.Context, row *model.Reader
 // ListReplaceRules 用户替换规则（按 order 排序）。
 func (r *ReaderRepository) ListReplaceRules(ctx context.Context, userID string) ([]model.ReaderReplaceRule, error) {
 	var out []model.ReaderReplaceRule
-	err := r.db.WithContext(ctx).Where("user_id = ?", userID).Order("`order` DESC, created_at ASC").Find(&out).Error
+	err := r.db.WithContext(ctx).Where("user_id = ?", userID).
+		Order(readerOrderBy(
+			clause.OrderByColumn{Column: clause.Column{Name: "order"}, Desc: true},
+			clause.OrderByColumn{Column: clause.Column{Name: "created_at"}},
+		)).Find(&out).Error
 	return out, err
 }
 

internal/repository/reader_repository_dialect_test.go

@@ -0,0 +1,116 @@
+package repository
+
+import (
+	"context"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"gorm.io/driver/postgres"
+	"gorm.io/gorm"
+	gormlogger "gorm.io/gorm/logger"
+)
+
+// sqlCaptureLogger 只记录 GORM 生成的 SQL，不打印任何日志。
+type sqlCaptureLogger struct {
+	mu    sync.Mutex
+	lines []string
+}
+
+func (l *sqlCaptureLogger) LogMode(gormlogger.LogLevel) gormlogger.Interface { return l }
+func (l *sqlCaptureLogger) Info(context.Context, string, ...interface{})     {}
+func (l *sqlCaptureLogger) Warn(context.Context, string, ...interface{})     {}
+func (l *sqlCaptureLogger) Error(context.Context, string, ...interface{})    {}
+
+func (l *sqlCaptureLogger) Trace(_ context.Context, _ time.Time, fc func() (string, int64), _ error) {
+	sql, _ := fc()
+	l.mu.Lock()
+	l.lines = append(l.lines, sql)
+	l.mu.Unlock()
+}
+
+func (l *sqlCaptureLogger) all() string {
+	l.mu.Lock()
+	defer l.mu.Unlock()
+	return strings.Join(l.lines, "\n")
+}
+
+// newPostgresDryRunRepos 返回使用 PostgreSQL 方言的 repository 容器：
+// DryRun 只构建 SQL 不执行，DisableAutomaticPing 让它无需真实数据库也能离线运行，
+// SkipDefaultTransaction 则挡掉 GORM 默认事务包装——那一步会真的去开事务连接，
+// 在 DryRun 下同样会拨号（ToSQL 内部也是这么绕开的）。于是可以直接断言生成语句
+// 在 PostgreSQL 下是否合法。
+func newPostgresDryRunRepos(t *testing.T) (*Container, *sqlCaptureLogger) {
+	t.Helper()
+	rec := &sqlCaptureLogger{}
+	db, err := gorm.Open(postgres.New(postgres.Config{
+		DSN: "postgres://mebox:mebox@127.0.0.1:5432/mebox?sslmode=disable",
+	}), &gorm.Config{
+		DryRun:                 true,
+		DisableAutomaticPing:   true,
+		SkipDefaultTransaction: true,
+		Logger:                 rec,
+	})
+	if err != nil {
+		t.Fatalf("open dry-run postgres: %v", err)
+	}
+	return New(db), rec
+}
+
+// TestReaderListsUseDialectQuotedReservedColumns 守住阅读子系统列表查询在
+// PostgreSQL 下的合法性。reader_books.order / reader_chapters.index /
+// reader_replace_rules.order 都是关键字级列名（order 在 PostgreSQL 里是保留字）。
+//
+// 排序条件曾经写成硬编码反引号的原始字符串，例如 .Order("`order` ASC")：
+// SQLite 正常，PostgreSQL 会把反引号词法分析成运算符，直接报
+// `syntax error at or near "order"` (SQLSTATE 42601)，书架 / 目录 / 替换规则整页 500。
+// 现在改用 clause.Column，由 GORM 按方言加引号（SQLite 反引号 / PostgreSQL 双引号）。
+func TestReaderListsUseDialectQuotedReservedColumns(t *testing.T) {
+	ctx := context.Background()
+	cases := []struct {
+		name  string
+		run   func(repos *Container) error
+		wants []string
+	}{
+		{
+			name:  "books",
+			run:   func(repos *Container) error { _, err := repos.Reader.ListBooks(ctx, "user-1"); return err },
+			wants: []string{`"order"`, `"updated_at" DESC`},
+		},
+		{
+			name:  "chapters",
+			run:   func(repos *Container) error { _, err := repos.Reader.ListChapters(ctx, "book-1"); return err },
+			wants: []string{`"index"`},
+		},
+		{
+			name:  "replace rules",
+			run:   func(repos *Container) error { _, err := repos.Reader.ListReplaceRules(ctx, "user-1"); return err },
+			wants: []string{`"order" DESC`, `"created_at"`},
+		},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			repos, rec := newPostgresDryRunRepos(t)
+			if err := tc.run(repos); err != nil {
+				t.Fatalf("build query: %v", err)
+			}
+			sql := rec.all()
+			if strings.TrimSpace(sql) == "" {
+				t.Fatal("no SQL captured from the repository call")
+			}
+			if strings.Contains(sql, "`") {
+				t.Fatalf("SQL uses backtick quoting, which PostgreSQL rejects:\n%s", sql)
+			}
+			if !strings.Contains(sql, "ORDER BY") {
+				t.Fatalf("SQL has no ORDER BY clause:\n%s", sql)
+			}
+			for _, want := range tc.wants {
+				if !strings.Contains(sql, want) {
+					t.Fatalf("SQL missing %q:\n%s", want, sql)
+				}
+			}
+		})
+	}
+}

internal/service/backup.go

@@ -1,9 +1,13 @@
 // Package service — database hot backup and restore.
 //
-// BackupService uses SQLite's "VACUUM INTO" (available since 3.27.0) to
-// produce a consistent copy of the database without blocking writers.
-// Backup files are stored under {data_dir}/backups/ with a timestamp
-// suffix. The admin UI can list / download / delete / restore them.
+// Backup files are stored under {data_dir}/backups/ with a timestamp suffix.
+// The admin UI can list / download / delete / restore them.
+//
+// 两种方言的备份形态不同：
+//   - SQLite：用 "VACUUM INTO"（3.27.0 起可用）产出一个一致的库文件，
+//     扩展名 .db；恢复即「替换库文件 + 重启」。
+//   - PostgreSQL：没有单文件语义，用逻辑备份（见 backup_postgres.go），
+//     扩展名 .json.gz；恢复是在单个事务内重建全部表。
 package service
 
 import (
@@ -65,18 +69,34 @@ func (b *BackupService) backupFilePath(filename string) (string, error) {
 	return path, nil
 }
 
-// Create produces a new hot backup via "VACUUM INTO".
+// isPostgres reports whether the service is backed by PostgreSQL.
+//
+// PostgreSQL 不支持 SQLite 的 "VACUUM INTO"，按文件替换恢复也会写错目标，
+// 因此备份/恢复必须按方言分支，不能共用一条路径。
+func (b *BackupService) isPostgres() bool {
+	return b.db != nil && b.db.Dialector != nil && b.db.Dialector.Name() == "postgres"
+}
+
+// Create produces a new hot backup. SQLite uses "VACUUM INTO"; PostgreSQL uses
+// a logical dump.
 func (b *BackupService) Create(ctx context.Context) (*BackupInfo, error) {
 	dir := b.backupDir()
 	if err := os.MkdirAll(dir, 0o750); err != nil {
 		return nil, err
 	}
 	ts := time.Now().Format("20060102_150405")
-	dst := filepath.Join(dir, fmt.Sprintf("mebox_%s.db", ts))
 
-	// VACUUM INTO creates a clean, non-WAL copy atomically.
-	if err := b.db.WithContext(ctx).Exec("VACUUM INTO ?", dst).Error; err != nil {
-		return nil, fmt.Errorf("VACUUM INTO failed: %w", err)
+	var (
+		dst string
+		err error
+	)
+	if b.isPostgres() {
+		dst, err = b.createPostgresBackup(ctx, ts)
+	} else {
+		dst, err = b.createSQLiteBackup(ctx, ts)
+	}
+	if err != nil {
+		return nil, err
 	}
 	stat, err := os.Stat(dst)
 	if err != nil {
@@ -91,6 +111,15 @@ func (b *BackupService) Create(ctx context.Context) (*BackupInfo, error) {
 	}, nil
 }
 
+func (b *BackupService) createSQLiteBackup(ctx context.Context, stamp string) (string, error) {
+	dst := filepath.Join(b.backupDir(), fmt.Sprintf("mebox_%s.db", stamp))
+	// VACUUM INTO creates a clean, non-WAL copy atomically.
+	if err := b.db.WithContext(ctx).Exec("VACUUM INTO ?", dst).Error; err != nil {
+		return "", fmt.Errorf("VACUUM INTO failed: %w", err)
+	}
+	return dst, nil
+}
+
 // List returns every backup file in the backup directory, newest first.
 func (b *BackupService) List() ([]BackupInfo, error) {
 	dir := b.backupDir()
@@ -103,7 +132,7 @@ func (b *BackupService) List() ([]BackupInfo, error) {
 	}
 	out := make([]BackupInfo, 0, len(entries))
 	for _, e := range entries {
-		if e.IsDir() || !strings.HasSuffix(e.Name(), ".db") {
+		if e.IsDir() || !isBackupFilename(e.Name()) {
 			continue
 		}
 		info, err := e.Info()
@@ -132,9 +161,10 @@ func (b *BackupService) Delete(filename string) error {
 	return os.Remove(path)
 }
 
-// Restore copies a backup over the live database using VACUUM INTO in
-// reverse. WARNING: this is destructive — the live DB will be replaced.
-// Callers should shut down the server after this call.
+// Restore replaces the live database with the given backup.
+// WARNING: this is destructive. SQLite restores by swapping the database file
+// (the caller should shut down the server afterwards); PostgreSQL restores by
+// rebuilding every table inside a single transaction.
 func (b *BackupService) Restore(ctx context.Context, filename string) error {
 	src, err := b.backupFilePath(filename)
 	if err != nil {
@@ -143,6 +173,19 @@ func (b *BackupService) Restore(ctx context.Context, filename string) error {
 	if _, err := os.Stat(src); err != nil {
 		return fmt.Errorf("backup not found: %s", filename)
 	}
+	if b.isPostgres() {
+		if !strings.HasSuffix(filename, postgresBackupSuffix) {
+			return errors.New("该备份是 SQLite 文件备份，当前运行在 PostgreSQL 上，无法恢复")
+		}
+		return b.restorePostgresBackup(ctx, src)
+	}
+	if !strings.HasSuffix(filename, ".db") {
+		return errors.New("该备份是 PostgreSQL 逻辑备份，当前运行在 SQLite 上，无法恢复")
+	}
+	return b.restoreSQLiteBackup(filename, src)
+}
+
+func (b *BackupService) restoreSQLiteBackup(filename, src string) error {
 	dbPath, err := filepath.Abs(b.cfg.Database.DBPath)
 	if err != nil || dbPath == "" {
 		return errors.New("invalid database path")
@@ -171,7 +214,13 @@ func (b *BackupService) Restore(ctx context.Context, filename string) error {
 	return nil
 }
 
-// isValidBackupFilename rejects path traversal attempts and non-.db files.
+// isBackupFilename reports whether name looks like a backup produced by either
+// dialect (SQLite ".db" or PostgreSQL ".json.gz").
+func isBackupFilename(name string) bool {
+	return strings.HasSuffix(name, ".db") || strings.HasSuffix(name, postgresBackupSuffix)
+}
+
+// isValidBackupFilename rejects path traversal attempts and non-backup files.
 func isValidBackupFilename(name string) bool {
 	if name == "" {
 		return false
@@ -179,8 +228,5 @@ func isValidBackupFilename(name string) bool {
 	if strings.ContainsAny(name, "/\\") || strings.Contains(name, "..") {
 		return false
 	}
-	if !strings.HasSuffix(name, ".db") {
-		return false
-	}
-	return true
+	return isBackupFilename(name)
 }

internal/service/backup_postgres.go

@@ -0,0 +1,280 @@
+package service
+
+import (
+	"compress/gzip"
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"os"
+	"path/filepath"
+	"reflect"
+	"time"
+
+	"gorm.io/gorm"
+	"gorm.io/gorm/schema"
+
+	"github.com/truewhile/MeBox/internal/database"
+	"github.com/truewhile/MeBox/internal/model"
+)
+
+// PostgreSQL 逻辑备份 / 恢复。
+//
+// SQLite 的备份是「VACUUM INTO 出一个一致的库文件，恢复时替换文件」，而
+// PostgreSQL 既没有 VACUUM INTO，数据也由服务端进程管理，没有可替换的单文件，
+// 所以这里实现逻辑备份：把每张表按主键顺序分块序列化成 JSON Lines（gzip 压缩）
+// 写入 {data_dir}/backups/。
+//
+// 分块序列化而不是整表 marshal：media 表在真实媒体库上可达几十万行、每行还带
+// overview / path 等长文本，整表进内存会把小内存的 NAS 打爆。
+//
+// 恢复在单个事务内「先清空全部模型表、再逐块插入」：任何一步失败（包括备份文件
+// 读到一半损坏）都整体回滚，不会留下半清空的库。因此恢复不需要重启服务。
+const (
+	postgresBackupSuffix   = ".json.gz"
+	postgresBackupFormat   = "mebox-postgres-logical"
+	postgresBackupVersion  = 1
+	backupDumpChunkSize    = 1000
+	backupRestoreBatchSize = 500
+)
+
+type postgresBackupHeader struct {
+	Format    string    `json:"format"`
+	Version   int       `json:"version"`
+	CreatedAt time.Time `json:"created_at"`
+}
+
+// postgresBackupRow 是「列名 → JSON 值」的一行。
+//
+// 刻意逐列反射序列化，而不是直接 json.Marshal 模型结构体：模型里有大量带
+// json:"-" 的列（Base.DeletedAt、User.AllowedLibraryIDs、APIConfig.APIKey、
+// StrmAccount.Config、PlayProfile.PINHash、ReaderBookSource.RawJSON…），走结构体
+// 自带的序列化会把这些列整列丢掉——软删除行会在恢复后「复活」，加密后的第三方
+// 密钥与 PIN 会全部丢失。
+type postgresBackupRow map[string]json.RawMessage
+
+type postgresBackupChunk struct {
+	Table string              `json:"table"`
+	Rows  []postgresBackupRow `json:"rows"`
+}
+
+// parseModelSchema 返回模型对应的表名与 GORM schema（逐列反射所需）。
+func parseModelSchema(db *gorm.DB, m any) (table string, sch *schema.Schema, err error) {
+	stmt := &gorm.Statement{DB: db}
+	if err := stmt.Parse(m); err != nil {
+		return "", nil, err
+	}
+	return stmt.Schema.Table, stmt.Schema, nil
+}
+
+// postgresBackupModelSchemas 返回「表名 → 模型 schema / 结构体类型」，用于把备份
+// 里的行还原成 GORM 能插入的具体类型。
+func postgresBackupModelSchemas(db *gorm.DB) (map[string]*schema.Schema, map[string]reflect.Type, error) {
+	schemas := make(map[string]*schema.Schema, len(model.AllModels()))
+	types := make(map[string]reflect.Type, len(model.AllModels()))
+	for _, m := range model.AllModels() {
+		table, sch, err := parseModelSchema(db, m)
+		if err != nil {
+			return nil, nil, err
+		}
+		schemas[table] = sch
+		types[table] = reflect.TypeOf(m).Elem()
+	}
+	return schemas, types, nil
+}
+
+// createPostgresBackup 把全库导出成一个 gzip 压缩的 JSON Lines 文件。
+func (b *BackupService) createPostgresBackup(ctx context.Context, stamp string) (string, error) {
+	dst := filepath.Join(b.backupDir(), "mebox_"+stamp+postgresBackupSuffix)
+	file, err := os.OpenFile(dst, os.O_CREATE|os.O_TRUNC|os.O_WRONLY, 0o600)
+	if err != nil {
+		return "", err
+	}
+	complete := false
+	defer func() {
+		_ = file.Close()
+		if !complete {
+			// 半截备份文件比没有备份更危险：留着会让人以为有可用的还原点。
+			_ = os.Remove(dst)
+		}
+	}()
+
+	zw := gzip.NewWriter(file)
+	enc := json.NewEncoder(zw)
+	if err := enc.Encode(postgresBackupHeader{
+		Format:    postgresBackupFormat,
+		Version:   postgresBackupVersion,
+		CreatedAt: time.Now(),
+	}); err != nil {
+		return "", fmt.Errorf("写入备份头失败: %w", err)
+	}
+	for _, m := range model.AllModels() {
+		table, sch, err := parseModelSchema(b.db, m)
+		if err != nil {
+			return "", err
+		}
+		if !b.db.Migrator().HasTable(table) {
+			continue
+		}
+		if err := b.dumpPostgresTable(ctx, m, sch, table, enc); err != nil {
+			return "", err
+		}
+	}
+	if err := zw.Close(); err != nil {
+		return "", fmt.Errorf("压缩备份失败: %w", err)
+	}
+	if err := file.Sync(); err != nil {
+		return "", fmt.Errorf("写入备份失败: %w", err)
+	}
+	complete = true
+	return dst, nil
+}
+
+// dumpPostgresTable 按主键分块导出一张表。FindInBatches 会自己按主键升序游标
+// 翻页（WHERE pk > last），不会像 OFFSET 分页那样在大表上做 O(n²) 扫描。
+func (b *BackupService) dumpPostgresTable(ctx context.Context, m any, sch *schema.Schema, table string, enc *json.Encoder) error {
+	// 软删除行也要备份：恢复后应还原成删除前的样子，而不是让它们「复活」。
+	slicePtr := reflect.New(reflect.SliceOf(reflect.TypeOf(m).Elem()))
+	err := b.db.WithContext(ctx).Unscoped().
+		FindInBatches(slicePtr.Interface(), backupDumpChunkSize, func(_ *gorm.DB, _ int) error {
+			rows, err := encodePostgresBackupRows(ctx, sch, slicePtr.Elem())
+			if err != nil {
+				return fmt.Errorf("编码表 %s 失败: %w", table, err)
+			}
+			return enc.Encode(postgresBackupChunk{Table: table, Rows: rows})
+		}).Error
+	if err != nil {
+		return fmt.Errorf("导出表 %s 失败: %w", table, err)
+	}
+	return nil
+}
+
+// backupPersistableFields 返回真正对应数据库列的字段。
+//
+// GORM 的 Schema.Fields 里还带着 gorm:"-" 的字段（例如 User.AllowedLibraryList、
+// LibraryTagList），它们的 DBName 是空字符串。不过滤的话多个「列」会共用空键：
+// 编码时互相覆盖，解码时把某个字段的值写到另一个字段上，报出
+// 「cannot unmarshal number into Go value of type []string」这类与真实数据无关的错。
+func backupPersistableFields(sch *schema.Schema) []*schema.Field {
+	out := make([]*schema.Field, 0, len(sch.Fields))
+	for _, field := range sch.Fields {
+		if field.DBName == "" {
+			continue
+		}
+		out = append(out, field)
+	}
+	return out
+}
+
+// encodePostgresBackupRows 逐列反射序列化一个批次。
+func encodePostgresBackupRows(ctx context.Context, sch *schema.Schema, slice reflect.Value) ([]postgresBackupRow, error) {
+	fields := backupPersistableFields(sch)
+	rows := make([]postgresBackupRow, 0, slice.Len())
+	for i := 0; i < slice.Len(); i++ {
+		element := slice.Index(i)
+		row := make(postgresBackupRow, len(fields))
+		for _, field := range fields {
+			value := field.ReflectValueOf(ctx, element)
+			if !value.IsValid() {
+				continue
+			}
+			raw, err := json.Marshal(value.Interface())
+			if err != nil {
+				return nil, fmt.Errorf("列 %s: %w", field.DBName, err)
+			}
+			row[field.DBName] = raw
+		}
+		rows = append(rows, row)
+	}
+	return rows, nil
+}
+
+// decodePostgresBackupRows 把「列名 → JSON 值」还原成可直接插入的具体切片。
+func decodePostgresBackupRows(ctx context.Context, sch *schema.Schema, modelType reflect.Type, rows []postgresBackupRow) (any, error) {
+	fields := backupPersistableFields(sch)
+	slice := reflect.MakeSlice(reflect.SliceOf(modelType), len(rows), len(rows))
+	for i, row := range rows {
+		element := slice.Index(i)
+		for _, field := range fields {
+			raw, ok := row[field.DBName]
+			if !ok {
+				continue
+			}
+			holder := reflect.New(field.FieldType)
+			if err := json.Unmarshal(raw, holder.Interface()); err != nil {
+				return nil, fmt.Errorf("列 %s: %w", field.DBName, err)
+			}
+			if err := field.Set(ctx, element, holder.Elem().Interface()); err != nil {
+				return nil, fmt.Errorf("列 %s: %w", field.DBName, err)
+			}
+		}
+	}
+	return slice.Interface(), nil
+}
+
+// restorePostgresBackup 用逻辑备份重建全库。
+func (b *BackupService) restorePostgresBackup(ctx context.Context, src string) error {
+	file, err := os.Open(src)
+	if err != nil {
+		return err
+	}
+	defer file.Close()
+	zr, err := gzip.NewReader(file)
+	if err != nil {
+		return fmt.Errorf("备份文件不是有效的 gzip: %w", err)
+	}
+	defer zr.Close()
+
+	dec := json.NewDecoder(zr)
+	var header postgresBackupHeader
+	if err := dec.Decode(&header); err != nil {
+		return fmt.Errorf("备份头损坏: %w", err)
+	}
+	if header.Format != postgresBackupFormat || header.Version != postgresBackupVersion {
+		return fmt.Errorf("不支持的备份格式 %q (v%d)", header.Format, header.Version)
+	}
+	schemas, modelTypes, err := postgresBackupModelSchemas(b.db)
+	if err != nil {
+		return err
+	}
+
+	// 恢复必须在单个事务里完成：清空成功后中途失败若不回滚，库就空了。
+	return b.db.WithContext(ctx).Transaction(func(tx *gorm.DB) error {
+		// 逆序清空，保证 library_roots → libraries 这类外键先删子表。
+		models := model.AllModels()
+		for i := len(models) - 1; i >= 0; i-- {
+			m := models[i]
+			if !tx.Migrator().HasTable(m) {
+				continue
+			}
+			if err := tx.Session(&gorm.Session{AllowGlobalUpdate: true}).Unscoped().Delete(m).Error; err != nil {
+				return fmt.Errorf("清空表失败: %w", err)
+			}
+		}
+		for {
+			var chunk postgresBackupChunk
+			decodeErr := dec.Decode(&chunk)
+			if errors.Is(decodeErr, io.EOF) {
+				return nil
+			}
+			if decodeErr != nil {
+				return fmt.Errorf("备份文件损坏: %w", decodeErr)
+			}
+			sch, ok := schemas[chunk.Table]
+			if !ok {
+				return fmt.Errorf("备份包含未知表 %s，可能来自不兼容的版本", chunk.Table)
+			}
+			if len(chunk.Rows) == 0 {
+				continue
+			}
+			rows, err := decodePostgresBackupRows(ctx, sch, modelTypes[chunk.Table], chunk.Rows)
+			if err != nil {
+				return fmt.Errorf("表 %s 解码失败: %w", chunk.Table, err)
+			}
+			if err := database.InsertPreservingValues(tx.Unscoped(), rows, backupRestoreBatchSize); err != nil {
+				return fmt.Errorf("恢复表 %s 失败: %w", chunk.Table, err)
+			}
+		}
+	})
+}

internal/service/backup_postgres_test.go

@@ -0,0 +1,226 @@
+package service
+
+import (
+	"compress/gzip"
+	"context"
+	"os"
+	"path/filepath"
+	"reflect"
+	"testing"
+
+	"go.uber.org/zap"
+
+	"github.com/truewhile/MeBox/internal/config"
+	"github.com/truewhile/MeBox/internal/model"
+)
+
+func newBackupTestService(t *testing.T) (*BackupService, *config.Config) {
+	t.Helper()
+	cfg := &config.Config{}
+	cfg.App.DataDir = t.TempDir()
+	db := newServiceTestDB(t)
+	svc := &BackupService{cfg: cfg, log: zap.NewNop(), db: db}
+	// Create 会先建目录；直接调各方言的 create* 助手时需要它已存在。
+	if err := os.MkdirAll(svc.backupDir(), 0o750); err != nil {
+		t.Fatal(err)
+	}
+	return svc, cfg
+}
+
+// TestPostgresLogicalBackupRoundTrip 验证「备份 → 改库 → 恢复」后整库回到备份时的
+// 样子，包括软删除行与 json:"-" 的列。
+//
+// 逻辑备份/恢复本身是方言无关的 GORM 代码（PostgreSQL 专有的只有文件后缀与
+// isPostgres() 分派），所以可以直接在 SQLite 上跑。
+func TestPostgresLogicalBackupRoundTrip(t *testing.T) {
+	svc, _ := newBackupTestService(t)
+	ctx := context.Background()
+	db := svc.db
+
+	lib := model.Library{Base: model.Base{ID: "lib-1"}, Name: "Movies", Path: "/media/movies", Type: "movie", Enabled: true}
+	if err := db.Create(&lib).Error; err != nil {
+		t.Fatal(err)
+	}
+	if err := db.Create(&model.LibraryRoot{Base: model.Base{ID: "root-1"}, LibraryID: lib.ID, Path: "/media/movies"}).Error; err != nil {
+		t.Fatal(err)
+	}
+	if err := db.Create(&model.Media{Base: model.Base{ID: "m-keep"}, LibraryID: lib.ID, Title: "Keep", Path: "/media/movies/keep.mkv"}).Error; err != nil {
+		t.Fatal(err)
+	}
+	if err := db.Create(&model.Media{Base: model.Base{ID: "m-gone"}, LibraryID: lib.ID, Title: "Gone", Path: "/media/movies/gone.mkv"}).Error; err != nil {
+		t.Fatal(err)
+	}
+	if err := db.Delete(&model.Media{}, "id = ?", "m-gone").Error; err != nil {
+		t.Fatal(err)
+	}
+	if err := db.Create(&model.Setting{Key: "k", Value: "v"}).Error; err != nil {
+		t.Fatal(err)
+	}
+	// 下面这些列都带 json:"-"：走结构体自带的 JSON 序列化会被整列丢掉，
+	// 恢复之后用户权限、置顶库、加密后的密钥会一起消失。
+	user := model.User{
+		Base:                 model.Base{ID: "u-1"},
+		Username:             "alice",
+		PasswordHash:         "hash",
+		Role:                 "user",
+		AllowedLibraryIDs:    `["lib-1"]`,
+		PinnedLibraryIDs:     `["lib-1"]`,
+		HideAdult:            true,
+		IsActive:             true,
+		DanmakuEnabled:       true,
+		PlayerVr360GuideSeen: true,
+	}
+	if err := db.Create(&user).Error; err != nil {
+		t.Fatal(err)
+	}
+	if err := db.Create(&model.APIConfig{Base: model.Base{ID: "cfg-1"}, Provider: "tmdb", APIKey: "ciphertext", Enabled: true}).Error; err != nil {
+		t.Fatal(err)
+	}
+
+	dump, err := svc.createPostgresBackup(ctx, "20260101_000000")
+	if err != nil {
+		t.Fatalf("createPostgresBackup: %v", err)
+	}
+	if dump == "" {
+		t.Fatal("createPostgresBackup returned an empty path")
+	}
+
+	// 破坏当前数据：加一条备份里没有的、硬删一条备份里有的。
+	if err := db.Create(&model.Media{Base: model.Base{ID: "m-extra"}, LibraryID: lib.ID, Title: "Extra", Path: "/media/movies/extra.mkv"}).Error; err != nil {
+		t.Fatal(err)
+	}
+	if err := db.Unscoped().Delete(&model.Media{}, "id = ?", "m-keep").Error; err != nil {
+		t.Fatal(err)
+	}
+
+	if err := svc.restorePostgresBackup(ctx, dump); err != nil {
+		t.Fatalf("restorePostgresBackup: %v", err)
+	}
+
+	var ids []string
+	if err := db.Unscoped().Model(&model.Media{}).Order("id asc").Pluck("id", &ids).Error; err != nil {
+		t.Fatal(err)
+	}
+	if want := []string{"m-gone", "m-keep"}; !reflect.DeepEqual(ids, want) {
+		t.Fatalf("media ids after restore = %v, want %v", ids, want)
+	}
+
+	// 软删除状态必须原样还原，而不是被「复活」。
+	var gone model.Media
+	if err := db.Unscoped().First(&gone, "id = ?", "m-gone").Error; err != nil {
+		t.Fatal(err)
+	}
+	if !gone.DeletedAt.Valid {
+		t.Fatal("soft-deleted media was revived by restore")
+	}
+	var live int64
+	if err := db.Model(&model.Media{}).Where("id = ?", "m-keep").Count(&live).Error; err != nil {
+		t.Fatal(err)
+	}
+	if live != 1 {
+		t.Fatalf("live media count = %d, want 1", live)
+	}
+
+	var value string
+	if err := db.Model(&model.Setting{}).Where("key = ?", "k").Select("value").Scan(&value).Error; err != nil {
+		t.Fatal(err)
+	}
+	if value != "v" {
+		t.Fatalf("setting value = %q, want v", value)
+	}
+	var libCount int64
+	if err := db.Model(&model.Library{}).Count(&libCount).Error; err != nil {
+		t.Fatal(err)
+	}
+	if libCount != 1 {
+		t.Fatalf("library count = %d, want 1", libCount)
+	}
+	var rootCount int64
+	if err := db.Model(&model.LibraryRoot{}).Count(&rootCount).Error; err != nil {
+		t.Fatal(err)
+	}
+	if rootCount != 1 {
+		t.Fatalf("library root count = %d, want 1", rootCount)
+	}
+
+	// json:"-" 的列必须同样被备份：它们是权限与凭据，丢了比丢一条媒体严重得多。
+	var restoredUser model.User
+	if err := db.First(&restoredUser, "id = ?", "u-1").Error; err != nil {
+		t.Fatal(err)
+	}
+	if restoredUser.PasswordHash != "hash" {
+		t.Fatalf("password hash = %q, want hash", restoredUser.PasswordHash)
+	}
+	if restoredUser.AllowedLibraryIDs != `["lib-1"]` {
+		t.Fatalf(`allowed library ids = %q, want ["lib-1"]`, restoredUser.AllowedLibraryIDs)
+	}
+	if restoredUser.PinnedLibraryIDs != `["lib-1"]` {
+		t.Fatalf(`pinned library ids = %q, want ["lib-1"]`, restoredUser.PinnedLibraryIDs)
+	}
+	if !restoredUser.HideAdult || !restoredUser.IsActive || !restoredUser.DanmakuEnabled || !restoredUser.PlayerVr360GuideSeen {
+		t.Fatalf("boolean columns lost: %+v", restoredUser)
+	}
+
+	var restoredConfig model.APIConfig
+	if err := db.First(&restoredConfig, "id = ?", "cfg-1").Error; err != nil {
+		t.Fatal(err)
+	}
+	if restoredConfig.APIKey != "ciphertext" {
+		t.Fatalf("api key = %q, want ciphertext", restoredConfig.APIKey)
+	}
+}
+
+// TestRestorePostgresBackupKeepsDataOnCorruptFile 验证损坏的备份不会在回滚后留下
+// 半清空的库：恢复是先清空再插入，必须整体在一个事务里，否则中途失败就把库清掉了。
+func TestRestorePostgresBackupKeepsDataOnCorruptFile(t *testing.T) {
+	svc, _ := newBackupTestService(t)
+	ctx := context.Background()
+	if err := svc.db.Create(&model.Setting{Key: "keep", Value: "me"}).Error; err != nil {
+		t.Fatal(err)
+	}
+
+	corrupt := filepath.Join(t.TempDir(), "broken.json.gz")
+	file, err := os.Create(corrupt)
+	if err != nil {
+		t.Fatal(err)
+	}
+	zw := gzip.NewWriter(file)
+	// 头部完整、正文半截：解码在清空表之后才会失败。
+	if _, err := zw.Write([]byte(
+		`{"format":"mebox-postgres-logical","version":1,"created_at":"2026-01-01T00:00:00Z"}` + "\n" +
+			`{"table":"media","rows":[`)); err != nil {
+		t.Fatal(err)
+	}
+	if err := zw.Close(); err != nil {
+		t.Fatal(err)
+	}
+	if err := file.Close(); err != nil {
+		t.Fatal(err)
+	}
+
+	if err := svc.restorePostgresBackup(ctx, corrupt); err == nil {
+		t.Fatal("corrupt backup should fail to restore")
+	}
+
+	var value string
+	if err := svc.db.Model(&model.Setting{}).Where("key = ?", "keep").Select("value").Scan(&value).Error; err != nil {
+		t.Fatal(err)
+	}
+	if value != "me" {
+		t.Fatalf("a failed restore wiped the database: settings keep = %q", value)
+	}
+}
+
+// TestRestoreRejectsBackupFromOtherDialect 验证备份文件与当前方言不匹配时给出明确
+// 错误，而不是把 SQLite 文件写到 PostgreSQL 部署的库路径上（那是原实现的隐患）。
+func TestRestoreRejectsBackupFromOtherDialect(t *testing.T) {
+	svc, _ := newBackupTestService(t)
+	pgStyle := filepath.Join(svc.backupDir(), "mebox_20260101_000000.json.gz")
+	if err := os.WriteFile(pgStyle, []byte("not-gzip"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	// SQLite 路径下不允许恢复 PostgreSQL 逻辑备份。
+	if err := svc.Restore(context.Background(), filepath.Base(pgStyle)); err == nil {
+		t.Fatal("SQLite restore accepted a PostgreSQL logical backup")
+	}
+}

internal/service/backup_zero_values_test.go

@@ -0,0 +1,34 @@
+package service
+
+import (
+	"testing"
+
+	"github.com/truewhile/MeBox/internal/model"
+)
+
+func TestPostgresLogicalBackupPreservesZeroValues(t *testing.T) {
+	svc, _ := newBackupTestService(t)
+	user := model.User{Username: "disabled", PasswordHash: "hash"}
+	if err := svc.db.Create(&user).Error; err != nil {
+		t.Fatal(err)
+	}
+	if err := svc.db.Model(&user).Updates(map[string]any{
+		"is_active": false, "hide_adult": false, "player_volume": 0, "danmaku_enabled": false,
+	}).Error; err != nil {
+		t.Fatal(err)
+	}
+	dump, err := svc.createPostgresBackup(t.Context(), "zero-values")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := svc.restorePostgresBackup(t.Context(), dump); err != nil {
+		t.Fatal(err)
+	}
+	var got model.User
+	if err := svc.db.First(&got, "id = ?", user.ID).Error; err != nil {
+		t.Fatal(err)
+	}
+	if got.IsActive || got.HideAdult || got.PlayerVolume != 0 || got.DanmakuEnabled {
+		t.Fatalf("restore replaced zero values: active=%v hide_adult=%v volume=%v danmaku=%v", got.IsActive, got.HideAdult, got.PlayerVolume, got.DanmakuEnabled)
+	}
+}

internal/service/danmaku_remote_resolver_test.go

@@ -0,0 +1,157 @@
+package service
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"go.uber.org/zap"
+
+	"github.com/truewhile/MeBox/internal/config"
+	"github.com/truewhile/MeBox/internal/model"
+	"github.com/truewhile/MeBox/internal/repository"
+)
+
+// newDanmakuRemoteResolverTest 起一个假的远程 Emby（账号 + 挂载齐备），返回接好
+// 弹幕解析器的 builder。
+func newDanmakuRemoteResolverTest(t *testing.T, handler http.Handler) *serviceContainerBuilder {
+	t.Helper()
+	server := httptest.NewServer(handler)
+	t.Cleanup(server.Close)
+
+	db := newServiceTestDB(t, &model.StrmAccount{}, &model.EmbyMount{})
+	repos := repository.New(db)
+	cfg := &config.Config{}
+	remote := NewEmbyRemoteService(cfg, zap.NewNop(), repos, NewCryptoService("", zap.NewNop()))
+	remote.SetRuntimeCache(NewRuntimeCacheService(cfg, zap.NewNop()))
+
+	rawConfig, _ := json.Marshal(map[string]string{
+		"url":            server.URL,
+		"api_key":        "test-api-key",
+		"remote_user_id": "uid-1",
+	})
+	if err := repos.StrmAccount.Create(t.Context(), &model.StrmAccount{
+		Base:     model.Base{ID: "acct-1"},
+		Name:     "远程 Emby",
+		Provider: model.StrmProviderEmbyRemote,
+		Config:   string(rawConfig),
+		Enabled:  true,
+	}); err != nil {
+		t.Fatalf("create account: %v", err)
+	}
+	if err := repos.EmbyMount.Create(t.Context(), &model.EmbyMount{
+		Base:           model.Base{ID: "mount-1"},
+		AccountID:      "acct-1",
+		RemoteViewID:   "view-1",
+		RemoteViewName: "新番连载",
+		CollectionType: "tvshows",
+		Enabled:        true,
+	}); err != nil {
+		t.Fatalf("create mount: %v", err)
+	}
+	return &serviceContainerBuilder{
+		cfg:   cfg,
+		log:   zap.NewNop(),
+		repos: repos,
+		c:     &Container{Cfg: cfg, Log: zap.NewNop(), Repo: repos, EmbyRemote: remote},
+	}
+}
+
+func danmakuResolverTestEpisode(id string) map[string]any {
+	return map[string]any{
+		"Id":   id,
+		"Name": "第 1 集",
+		"Type": "Episode",
+		"Path": "/media/anime/PSYREN/" + id + ".mkv",
+		"MediaSources": []any{
+			map[string]any{"Id": id + "-src", "Container": "mkv", "Size": int64(1 << 20)},
+		},
+	}
+}
+
+// 目录类条目（整部剧/季）自身没有媒体源，拿它的 ID 去请求远程 /Videos/{id}/stream 只会
+// 回 500，哈希层于是每次都白白跳过。这条用例保证弹幕解析器跟随 /api/stream 的规则先
+// 落到第一集：哈希算的是真正要播的那条流。
+func TestResolveDanmakuRemoteMediaUsesPlayableChildForContainer(t *testing.T) {
+	var requested []string
+	b := newDanmakuRemoteResolverTest(t, http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		requested = append(requested, r.URL.Path)
+		w.Header().Set("Content-Type", "application/json")
+		switch r.URL.Path {
+		case "/emby/Users/uid-1/Items/series-100":
+			_ = json.NewEncoder(w).Encode(map[string]any{
+				"Id": "series-100", "Name": "PSYREN", "Type": "Series", "IsFolder": true,
+			})
+		case "/emby/Users/uid-1/Items":
+			if got := r.URL.Query().Get("ParentId"); got != "series-100" {
+				t.Errorf("ParentId = %q, want series-100", got)
+			}
+			_ = json.NewEncoder(w).Encode(map[string]any{
+				"Items": []map[string]any{
+					{"Id": "season-1", "Type": "Season", "IndexNumber": 1},
+					danmakuResolverTestEpisode("ep-1"),
+				},
+			})
+		case "/emby/Users/uid-1/Items/ep-1":
+			_ = json.NewEncoder(w).Encode(danmakuResolverTestEpisode("ep-1"))
+		default:
+			w.WriteHeader(http.StatusNotFound)
+		}
+	}))
+
+	media, streamURL, err := b.resolveDanmakuRemoteMedia(
+		context.Background(), EncodeEmbyRemoteID("mount-1", "series-100"))
+	if err != nil {
+		t.Fatalf("resolve failed: %v", err)
+	}
+	if media == nil {
+		t.Fatal("resolve returned no media metadata")
+	}
+	if !strings.Contains(streamURL, "/Videos/ep-1/stream") {
+		t.Fatalf("stream url = %q, want the playable child ep-1", streamURL)
+	}
+	if strings.Contains(streamURL, "series-100") {
+		t.Fatalf("stream url = %q, must not point at the container id (remote Emby answers 500)", streamURL)
+	}
+	if !strings.Contains(strings.Join(requested, " "), "/Items/ep-1") {
+		t.Fatalf("metadata must come from the playable child, requests = %v", requested)
+	}
+}
+
+// 本身可播的条目（单集/电影）不能被改写：直接取它自己的流地址。
+func TestResolveDanmakuRemoteMediaKeepsPlayableItem(t *testing.T) {
+	b := newDanmakuRemoteResolverTest(t, http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		if r.URL.Path != "/emby/Users/uid-1/Items/ep-2" {
+			t.Errorf("unexpected upstream path %q", r.URL.Path)
+		}
+		_ = json.NewEncoder(w).Encode(danmakuResolverTestEpisode("ep-2"))
+	}))
+
+	media, streamURL, err := b.resolveDanmakuRemoteMedia(
+		context.Background(), EncodeEmbyRemoteID("mount-1", "ep-2"))
+	if err != nil {
+		t.Fatalf("resolve failed: %v", err)
+	}
+	if media == nil {
+		t.Fatal("resolve returned no media metadata")
+	}
+	if !strings.Contains(streamURL, "/Videos/ep-2/stream") {
+		t.Fatalf("stream url = %q, want /Videos/ep-2/stream", streamURL)
+	}
+}
+
+// 不是远程伪装 ID 时直接报错，不去猜一个可以拉取的地址。
+func TestResolveDanmakuRemoteMediaRejectsPlainID(t *testing.T) {
+	b := newDanmakuRemoteResolverTest(t, http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		t.Error("plain local ids must not reach the remote Emby")
+		w.WriteHeader(http.StatusNotFound)
+	}))
+
+	if _, _, err := b.resolveDanmakuRemoteMedia(context.Background(), "local-media-1"); err == nil {
+		t.Fatal("expected an error for a non remote-emby id")
+	}
+}

internal/service/emby_episodic_path_test.go

@@ -0,0 +1,82 @@
+package service
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/truewhile/MeBox/internal/model"
+)
+
+// TestLikelyEpisodicPathSQLNormalisesSeparator 守住剧集路径判定的转义差异。
+//
+// 此前 embyLikelyEpisodicPathSQL 会额外生成一份反斜杠模式
+// （strings.ReplaceAll(pattern, "/", `\`)）。PostgreSQL 的 LIKE 默认就把反斜杠
+// 当转义符，模式结尾的 `\%` 被解析成「字面百分号」而不是「反斜杠 + 通配符」，
+// 那些模式在 PostgreSQL 下永不命中；SQLite 没有默认转义符所以一直正常。
+// 现在统一在 SQL 里把分隔符归一化成 '/'，模式本身不再含反斜杠。
+func TestLikelyEpisodicPathSQLNormalisesSeparator(t *testing.T) {
+	clause, args := embyLikelyEpisodicPathSQL()
+	if !strings.Contains(clause, "REPLACE(media.path") {
+		t.Fatalf("clause no longer normalises the path separator:\n%s", clause)
+	}
+	if got := strings.Count(clause, "LIKE ?"); got != len(args) {
+		t.Fatalf("clause has %d placeholders but %d args:\n%s", got, len(args), clause)
+	}
+	for _, arg := range args {
+		pattern, ok := arg.(string)
+		if !ok {
+			t.Fatalf("unexpected pattern type %T", arg)
+		}
+		if strings.Contains(pattern, `\`) {
+			t.Fatalf("pattern %q still contains a backslash; PostgreSQL's LIKE would treat it as an escape character", pattern)
+		}
+		if !strings.Contains(pattern, "/") {
+			t.Fatalf("pattern %q lost its separator", pattern)
+		}
+	}
+}
+
+// TestFilterLikelyEpisodicPathsHandlesWindowsStylePaths 验证反斜杠路径（Windows
+// 宿主机入库的媒体）仍然被判成剧集并从电影库查询里排除。
+func TestFilterLikelyEpisodicPathsHandlesWindowsStylePaths(t *testing.T) {
+	db := newServiceTestDB(t)
+	episodic := model.Media{
+		Base:       model.Base{ID: "ep-1"},
+		Path:       `D:\Media\国产剧\我的剧\Season 1\我的剧 - S01E01.mkv`,
+		SeasonNum:  1,
+		EpisodeNum: 1,
+	}
+	movie := model.Media{
+		Base: model.Base{ID: "mv-1"},
+		Path: `D:\Media\电影\某电影\某电影.mkv`,
+	}
+	if err := db.Create(&episodic).Error; err != nil {
+		t.Fatal(err)
+	}
+	if err := db.Create(&movie).Error; err != nil {
+		t.Fatal(err)
+	}
+
+	q := filterLikelyEpisodicPathsFromMovieQuery(
+		db.Model(&model.Media{}).Where("season_num > 0 OR episode_num > 0"),
+	)
+	var rows []model.Media
+	if err := q.Find(&rows).Error; err != nil {
+		t.Fatal(err)
+	}
+	if len(rows) != 0 {
+		t.Fatalf("windows-style episodic path was not excluded, got %d rows", len(rows))
+	}
+
+	// 反证：电影类型的行必须保留。
+	var kept []model.Media
+	keepQ := filterLikelyEpisodicPathsFromMovieQuery(
+		db.Model(&model.Media{}).Where("path LIKE ?", `D:\Media\电影\%`),
+	)
+	if err := keepQ.Find(&kept).Error; err != nil {
+		t.Fatal(err)
+	}
+	if len(kept) != 1 || kept[0].ID != "mv-1" {
+		t.Fatalf("plain movie row was dropped from the movie query: %+v", kept)
+	}
+}

internal/service/emby_items_list.go

@@ -7,6 +7,7 @@ import (
 	"time"
 
 	"github.com/truewhile/MeBox/internal/model"
+	"github.com/truewhile/MeBox/internal/repository"
 )
 
 func (e *EmbyService) mediaItems(ctx context.Context, p ItemsParams) (map[string]any, error) {
@@ -21,7 +22,11 @@ func (e *EmbyService) mediaItems(ctx context.Context, p ItemsParams) (map[string
 		q = q.Where("library_id IN ? OR series_id = ?", e.mergedLibraryIDs(ctx, p.ParentID), p.ParentID)
 	}
 	if p.SearchTerm != "" {
-		q = q.Where("title LIKE ? OR original_name LIKE ?", "%"+p.SearchTerm+"%", "%"+p.SearchTerm+"%")
+		// ILIKE（PostgreSQL）/ LIKE（SQLite）由方言决定，避免换库后搜索词
+		// 大小写不匹配就搜不到结果。
+		likeOp := repository.LikeOperator(e.repo.DB)
+		term := "%" + p.SearchTerm + "%"
+		q = q.Where("title "+likeOp+" ? OR original_name "+likeOp+" ?", term, term)
 	}
 	if containsEmbyFilter(p.Filters, "IsFavorite") {
 		if strings.TrimSpace(p.UserID) == "" {
@@ -264,7 +269,11 @@ func (e *EmbyService) seriesItemsForLibrary(ctx context.Context, libraryID strin
 		q = q.Where("library_id IN ?", e.mergedLibraryIDs(ctx, libraryID))
 	}
 	if p.SearchTerm != "" {
-		q = q.Where("title LIKE ? OR original_name LIKE ?", "%"+p.SearchTerm+"%", "%"+p.SearchTerm+"%")
+		// ILIKE（PostgreSQL）/ LIKE（SQLite）由方言决定，避免换库后搜索词
+		// 大小写不匹配就搜不到结果。
+		likeOp := repository.LikeOperator(e.repo.DB)
+		term := "%" + p.SearchTerm + "%"
+		q = q.Where("title "+likeOp+" ? OR original_name "+likeOp+" ?", term, term)
 	}
 	if containsEmbyFilter(p.Filters, "IsFavorite") {
 		if strings.TrimSpace(p.UserID) == "" {

internal/service/emby_movie_items.go

@@ -9,6 +9,7 @@ import (
 	"gorm.io/gorm"
 
 	"github.com/truewhile/MeBox/internal/model"
+	"github.com/truewhile/MeBox/internal/repository"
 )
 
 // movieLibraryHasEpisodicContent 报告电影类型库里是否混入了「剧集结构」内容
@@ -51,7 +52,11 @@ func (e *EmbyService) movieLibraryItems(ctx context.Context, p ItemsParams) (map
 		q = e.applyUserMediaVisibility(ctx, q, p.UserID)
 		q = q.Where("library_id IN ?", libIDs)
 		if p.SearchTerm != "" {
-			q = q.Where("title LIKE ? OR original_name LIKE ?", "%"+p.SearchTerm+"%", "%"+p.SearchTerm+"%")
+			// 方言相关的 ILIKE/LIKE：PostgreSQL 的 LIKE 区分大小写，
+			// 直接用 LIKE 会让客户端搜索的大小写与入库标题不一致时搜不到。
+			likeOp := repository.LikeOperator(e.repo.DB)
+			term := "%" + p.SearchTerm + "%"
+			q = q.Where("title "+likeOp+" ? OR original_name "+likeOp+" ?", term, term)
 		}
 		if containsEmbyFilter(p.Filters, "IsFavorite") {
 			if strings.TrimSpace(p.UserID) == "" {
@@ -266,6 +271,14 @@ func filterLikelyEpisodicPathsFromMovieQuery(q *gorm.DB) *gorm.DB {
 	return q.Where("NOT ((media.season_num > 0 OR media.episode_num > 0) AND ("+clause+"))", args...)
 }
 
+// embyEpisodicPathExpr 把路径分隔符归一化成 '/' 后再小写化，供剧集路径判定使用。
+//
+// 反斜杠归一化必须放在 SQL 里做，而不是再生成一份反斜杠模式：PostgreSQL 的
+// LIKE 默认就把反斜杠当转义符，模式结尾的 `\%` 会被解析成「字面百分号」而不是
+// 「反斜杠 + 通配符」，那些模式在 PostgreSQL 下永不命中（SQLite 没有默认转义符，
+// 所以只有换库之后才会暴露）。归一化列值后两种方言结果一致。
+const embyEpisodicPathExpr = `LOWER(REPLACE(media.path, '\', '/'))`
+
 func embyLikelyEpisodicPathSQL() (string, []any) {
 	patterns := []string{
 		"%/season %/%", "%/season.%/%", "%/season-%/%", "%/season_%/%",
@@ -274,15 +287,11 @@ func embyLikelyEpisodicPathSQL() (string, []any) {
 		"%/电视剧/%", "%/剧集/%", "%/连续剧/%", "%/短剧/%", "%/国产剧/%", "%/国剧/%", "%/大陆剧/%", "%/华语剧/%", "%/国产电视剧/%", "%/大陆电视剧/%", "%/华语电视剧/%", "%/欧美剧/%", "%/欧美电视剧/%", "%/美剧/%", "%/英剧/%", "%/日韩剧/%", "%/日韩电视剧/%", "%/日剧/%", "%/韩剧/%", "%/港剧/%", "%/台剧/%", "%/港台剧/%", "%/泰剧/%",
 		"%/日番/%", "%/国漫/%", "%/番剧/%", "%/动漫/%", "%/特别篇/%", "%/特別篇/%", "%/番外/%", "%/特典/%",
 	}
-	clauses := make([]string, 0, len(patterns)*2)
-	args := make([]any, 0, len(patterns)*2)
+	clauses := make([]string, 0, len(patterns))
+	args := make([]any, 0, len(patterns))
 	for _, pattern := range patterns {
-		clauses = append(clauses, "LOWER(media.path) LIKE ?")
+		clauses = append(clauses, embyEpisodicPathExpr+" LIKE ?")
 		args = append(args, pattern)
-		if strings.Contains(pattern, "/") {
-			clauses = append(clauses, "LOWER(media.path) LIKE ?")
-			args = append(args, strings.ReplaceAll(pattern, "/", `\`))
-		}
 	}
 	return strings.Join(clauses, " OR "), args
 }

internal/service/emby_remote.go

@@ -1587,7 +1587,22 @@ func (r *EmbyRemoteService) proxyVideoStreamOnLine(ctx context.Context, w http.R
 	defer resp.Body.Close()
 	if resp.StatusCode >= 400 {
 		data, _ := io.ReadAll(io.LimitReader(resp.Body, 256))
-		return redactSensitiveError(fmt.Errorf("远程 Emby 视频流失败(%d): %s", resp.StatusCode, strings.TrimSpace(string(data))))
+		detail := strings.TrimSpace(string(data))
+		// 远程 Emby 对「整部剧/目录」这类自身没有媒体源的条目会直接 500，正文是
+		// 一句 .NET 异常（如 Unable to cast object of type 'Series' to type
+		// 'IHasMediaSources'）。原样丢给用户没人看得懂，也没有排查线索——细节进日志，
+		// 给用户一句能照着做的提示。
+		if r.log != nil {
+			r.log.Warn("remote emby video stream rejected",
+				zap.String("remote_id", remoteID),
+				zap.Int("status", resp.StatusCode),
+				zap.String("detail", detail),
+			)
+		}
+		if resp.StatusCode >= 500 {
+			return redactSensitiveError(fmt.Errorf("远程 Emby 无法播放该条目（HTTP %d）：请在详情页选择具体集数后重试", resp.StatusCode))
+		}
+		return redactSensitiveError(fmt.Errorf("远程 Emby 视频流失败(%d): %s", resp.StatusCode, detail))
 	}
 	for _, header := range []string{"Content-Type", "Content-Length", "Content-Range", "Accept-Ranges", "ETag", "Cache-Control"} {
 		if value := resp.Header.Get(header); value != "" {

internal/service/emby_remote_playable_id_test.go

@@ -0,0 +1,160 @@
+package service
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"strconv"
+	"sync/atomic"
+	"testing"
+
+	"go.uber.org/zap"
+
+	"github.com/truewhile/MeBox/internal/config"
+	"github.com/truewhile/MeBox/internal/model"
+	"github.com/truewhile/MeBox/internal/repository"
+)
+
+// newPlayableIDTestService 起一个假的远程 Emby，并建好账号。
+func newPlayableIDTestService(t *testing.T, handler http.Handler) (*EmbyRemoteService, *model.StrmAccount) {
+	t.Helper()
+	server := httptest.NewServer(handler)
+	t.Cleanup(server.Close)
+
+	db := newServiceTestDB(t, &model.StrmAccount{}, &model.EmbyMount{})
+	repos := repository.New(db)
+	cfg := &config.Config{}
+	remote := NewEmbyRemoteService(cfg, zap.NewNop(), repos, NewCryptoService("", zap.NewNop()))
+	remote.SetRuntimeCache(NewRuntimeCacheService(cfg, zap.NewNop()))
+
+	rawConfig, _ := json.Marshal(map[string]string{
+		"url":            server.URL,
+		"api_key":        "test-api-key",
+		"remote_user_id": "uid-1",
+	})
+	acct := &model.StrmAccount{
+		Base:     model.Base{ID: "acct-1"},
+		Name:     "远程 Emby",
+		Provider: model.StrmProviderEmbyRemote,
+		Config:   string(rawConfig),
+		Enabled:  true,
+	}
+	if err := repos.StrmAccount.Create(t.Context(), acct); err != nil {
+		t.Fatalf("create account: %v", err)
+	}
+	return remote, acct
+}
+
+func playableIDTestEpisode(id string, index, sources int) map[string]any {
+	item := map[string]any{
+		"Id":                id,
+		"Name":              "第" + strconv.Itoa(index) + "集",
+		"Type":              "Episode",
+		"SeriesId":          "series-100",
+		"ParentIndexNumber": 1,
+		"IndexNumber":       index,
+	}
+	if sources > 0 {
+		item["MediaSources"] = []any{map[string]any{"Id": id + "-source"}}
+	}
+	return item
+}
+
+// 整部剧的 ID 自身没有媒体源，远程 Emby 对它的取流请求会回 500
+// （Unable to cast 'Series' to 'IHasMediaSources'）。这条用例保证我们会换成第一集，
+// 而不是把这种 ID 直接交给远程。
+func TestPlayableRemoteIDResolvesContainerToFirstEpisode(t *testing.T) {
+	var itemsCalls int64
+	remote, acct := newPlayableIDTestService(t, http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		switch r.URL.Path {
+		case "/emby/Users/uid-1/Items/series-100":
+			_ = json.NewEncoder(w).Encode(map[string]any{
+				"Id": "series-100", "Name": "剧一", "Type": "Series", "IsFolder": true,
+			})
+		case "/emby/Users/uid-1/Items/ep-2":
+			_ = json.NewEncoder(w).Encode(playableIDTestEpisode("ep-2", 2, 1))
+		case "/emby/Users/uid-1/Items":
+			atomic.AddInt64(&itemsCalls, 1)
+			if got := r.URL.Query().Get("ParentId"); got != "series-100" {
+				t.Errorf("ParentId = %q, want series-100", got)
+			}
+			// 故意乱序，并让第一条没有媒体源：只有真按季/集排序、且检查过媒体源，
+			// 才会选中 ep-1。
+			_ = json.NewEncoder(w).Encode(map[string]any{
+				"Items": []map[string]any{
+					{"Id": "season-1", "Type": "Season", "IndexNumber": 1},
+					playableIDTestEpisode("ep-3", 3, 1),
+					playableIDTestEpisode("ep-1", 1, 1),
+					playableIDTestEpisode("ep-2", 2, 1),
+				},
+			})
+		default:
+			w.WriteHeader(http.StatusNotFound)
+		}
+	}))
+
+	ctx := context.Background()
+	if got := remote.PlayableRemoteID(ctx, acct, "series-100"); got != "ep-1" {
+		t.Fatalf("series id resolved to %q, want ep-1", got)
+	}
+	// 播放器对同一个地址会反复发 Range 请求，解析结果必须命中缓存，不再打远程。
+	if got := remote.PlayableRemoteID(ctx, acct, "series-100"); got != "ep-1" {
+		t.Fatalf("cached resolve = %q, want ep-1", got)
+	}
+	if n := atomic.LoadInt64(&itemsCalls); n != 1 {
+		t.Fatalf("episode query issued %d times, want 1 (cached)", n)
+	}
+}
+
+// 本身带媒体源的条目（电影/单集）必须原样返回，不能被改写。
+func TestPlayableRemoteIDKeepsPlayableItem(t *testing.T) {
+	var queries int64
+	remote, acct := newPlayableIDTestService(t, http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		atomic.AddInt64(&queries, 1)
+		w.Header().Set("Content-Type", "application/json")
+		if r.URL.Path != "/emby/Users/uid-1/Items/ep-2" {
+			t.Errorf("unexpected upstream path %s", r.URL.Path)
+		}
+		_ = json.NewEncoder(w).Encode(playableIDTestEpisode("ep-2", 2, 1))
+	}))
+
+	if got := remote.PlayableRemoteID(context.Background(), acct, "ep-2"); got != "ep-2" {
+		t.Fatalf("playable item = %q, want ep-2", got)
+	}
+	if n := atomic.LoadInt64(&queries); n != 1 {
+		t.Fatalf("upstream queries = %d, want 1", n)
+	}
+}
+
+// 远程不可用/查不到条目时保持原 ID：让原有流程去报错，别把播放地址改成空。
+func TestPlayableRemoteIDKeepsOriginalOnFailure(t *testing.T) {
+	remote, acct := newPlayableIDTestService(t, http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.WriteHeader(http.StatusNotFound)
+	}))
+
+	if got := remote.PlayableRemoteID(context.Background(), acct, "series-100"); got != "series-100" {
+		t.Fatalf("failure path = %q, want the original id", got)
+	}
+}
+
+// 容器下没有任何带媒体源的子条目时同样返回原 ID（下一层会给友好报错）。
+func TestPlayableRemoteIDKeepsContainerWithoutPlayableChild(t *testing.T) {
+	remote, acct := newPlayableIDTestService(t, http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		if r.URL.Path == "/emby/Users/uid-1/Items" {
+			_ = json.NewEncoder(w).Encode(map[string]any{
+				"Items": []map[string]any{
+					{"Id": "season-1", "Type": "Season", "IndexNumber": 1},
+				},
+			})
+			return
+		}
+		_ = json.NewEncoder(w).Encode(map[string]any{"Id": "series-100", "Type": "Series", "IsFolder": true})
+	}))
+
+	if got := remote.PlayableRemoteID(context.Background(), acct, "series-100"); got != "series-100" {
+		t.Fatalf("container without playable child = %q, want the original id", got)
+	}
+}

internal/service/emby_remote_stream_proxy_test.go

@@ -0,0 +1,84 @@
+package service
+
+import (
+	"context"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"go.uber.org/zap"
+)
+
+// 远程条目自身没有媒体源时（例如把「整部剧」的 ID 当播放目标），远程 Emby 会回
+// 500 + 一句 .NET 异常。异常原文不能直接丢给用户，但要能从日志里查得到。
+func TestProxyVideoStreamOnLineHidesUpstreamException(t *testing.T) {
+	const upstreamBody = "Unable to cast object of type 'MediaBrowser.Controller.Entities.TV.Series' to type 'MediaBrowser.Controller.Entities.IHasMediaSources'."
+
+	var gotPath, gotStatic, gotRange, gotToken string
+	upstream := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, req *http.Request) {
+		gotPath = req.URL.Path
+		gotStatic = req.URL.Query().Get("Static")
+		gotRange = req.Header.Get("Range")
+		gotToken = req.Header.Get("X-Emby-Token")
+		w.WriteHeader(http.StatusInternalServerError)
+		_, _ = w.Write([]byte(upstreamBody))
+	}))
+	defer upstream.Close()
+
+	svc := &EmbyRemoteService{stream: upstream.Client(), log: zap.NewNop()}
+	cfg := &EmbyRemoteConfig{BaseURL: upstream.URL, Token: "tok-1"}
+
+	req := httptest.NewRequest(http.MethodGet, "/api/stream/embyremote~acct~222679", nil)
+	req.Header.Set("Range", "bytes=0-1023")
+	rec := httptest.NewRecorder()
+
+	err := svc.proxyVideoStreamOnLine(context.Background(), rec, req, cfg, "222679")
+	if err == nil {
+		t.Fatal("expected an error for upstream 500")
+	}
+	if gotPath != "/emby/Videos/222679/stream" {
+		t.Fatalf("upstream path = %q", gotPath)
+	}
+	if gotStatic != "true" {
+		t.Fatalf("Static = %q, want true", gotStatic)
+	}
+	if gotRange != "bytes=0-1023" {
+		t.Fatalf("Range = %q, want forwarded range", gotRange)
+	}
+	if gotToken != "tok-1" {
+		t.Fatalf("X-Emby-Token = %q", gotToken)
+	}
+	if rec.Body.Len() != 0 {
+		t.Fatalf("nothing should be written to the client on upstream failure, got %q", rec.Body.String())
+	}
+
+	msg := err.Error()
+	if strings.Contains(msg, "Unable to cast") || strings.Contains(msg, "MediaBrowser") {
+		t.Fatalf("raw upstream exception must not leak to clients: %q", msg)
+	}
+	if !strings.Contains(msg, "无法播放该条目") {
+		t.Fatalf("error should tell the user what to do, got %q", msg)
+	}
+}
+
+// 4xx（例如远程条目被删除/无权限）保持原有措辞，带着上游说明便于排查。
+func TestProxyVideoStreamOnLineKeepsClientErrorDetail(t *testing.T) {
+	upstream := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.WriteHeader(http.StatusNotFound)
+		_, _ = w.Write([]byte("Item not found"))
+	}))
+	defer upstream.Close()
+
+	svc := &EmbyRemoteService{stream: upstream.Client(), log: zap.NewNop()}
+	cfg := &EmbyRemoteConfig{BaseURL: upstream.URL, Token: "tok-1"}
+
+	req := httptest.NewRequest(http.MethodGet, "/api/stream/embyremote~acct~404", nil)
+	err := svc.proxyVideoStreamOnLine(context.Background(), httptest.NewRecorder(), req, cfg, "404")
+	if err == nil {
+		t.Fatal("expected an error for upstream 404")
+	}
+	if !strings.Contains(err.Error(), "404") || !strings.Contains(err.Error(), "Item not found") {
+		t.Fatalf("client error detail should be kept, got %q", err.Error())
+	}
+}

internal/service/emby_remote_web.go

@@ -820,6 +820,123 @@ func (r *EmbyRemoteService) remoteItemSeriesID(ctx context.Context, acct *model.
 	return remoteItemString(out, "SeriesId")
 }
 
+// PlayableRemoteID 把「自身没有媒体源」的远程条目换成可直接播放的条目 ID。
+//
+// 远程 Emby 只为带媒体源的条目提供 /Videos/{id}/stream：拿整部剧/季/目录的 ID 去
+// 请求，它会回 500 加一句 .NET 异常（Unable to cast object of type 'Series' to
+// type 'IHasMediaSources'）。首页轮播、剧集库卡片、外部播放器与第三方客户端都可能
+// 把这种 ID 交过来，统一在这里回落到第一集，保证拿到的 ID 真能取到流。
+//
+// 条目本身可播、或解析失败时返回原 ID（调用方拿到后用原逻辑处理即可，无需区分错误）。
+func (r *EmbyRemoteService) PlayableRemoteID(ctx context.Context, acct *model.StrmAccount, remoteID string) string {
+	remoteID = strings.TrimSpace(remoteID)
+	if r == nil || acct == nil || remoteID == "" {
+		return remoteID
+	}
+	cfg, err := r.remoteConfigWithToken(ctx, acct)
+	if err != nil {
+		return remoteID
+	}
+	cacheKey := r.remoteCacheKey("playable-remote-id", acct.ID, remoteID)
+	if r.cache != nil {
+		var cached string
+		// 播放器会对同一个地址发很多次 Range 请求，判定结果必须走缓存，否则每次拖动
+		// 进度条都要多打两次远程 Emby。缓存比常规媒体缓存长：条目可播性是静态属性。
+		if r.cache.GetJSON(ctx, cacheKey, &cached) {
+			if cached == "" {
+				return remoteID
+			}
+			return cached
+		}
+	}
+	resolved := r.resolvePlayableRemoteID(ctx, acct, cfg, remoteID)
+	// 解析失败不写缓存：远程抖动时下次照样重试，而不是把「播不了」记住十分钟。
+	if r.cache != nil && resolved != "" {
+		r.cache.SetJSON(ctx, cacheKey, resolved, playableRemoteIDCacheTTL)
+	}
+	if resolved == "" {
+		return remoteID
+	}
+	return resolved
+}
+
+// playableRemoteIDCacheTTL 是「条目可播性」判定的缓存时长。
+const playableRemoteIDCacheTTL = 10 * time.Minute
+
+// resolvePlayableRemoteID 返回 remoteID 对应的可播条目：本身带媒体源时原样返回，
+// 容器条目（整部剧/季/目录）返回排在最前的、确实带媒体源的子条目。
+func (r *EmbyRemoteService) resolvePlayableRemoteID(ctx context.Context, acct *model.StrmAccount, cfg *EmbyRemoteConfig, remoteID string) string {
+	var item map[string]any
+	if err := r.doGet(ctx, acct, cfg, "/Users/"+url.PathEscape(r.remoteUserID(cfg))+"/Items/"+url.PathEscape(remoteID), nil, &item); err != nil {
+		return ""
+	}
+	if remoteItemHasMediaSources(item) || !remoteItemIsContainer(item) {
+		return remoteID
+	}
+
+	q := url.Values{}
+	q.Set("ParentId", remoteID)
+	q.Set("Recursive", "true")
+	q.Set("IncludeItemTypes", "Episode,Movie,Video")
+	q.Set("Fields", "MediaSources")
+	q.Set("SortBy", "ParentIndexNumber,IndexNumber,SortName")
+	q.Set("Limit", "50")
+	var body struct {
+		Items []map[string]any `json:"Items"`
+	}
+	if err := r.doGet(ctx, acct, cfg, "/Users/"+url.PathEscape(r.remoteUserID(cfg))+"/Items", q, &body); err != nil {
+		return ""
+	}
+
+	// 远程 SortBy 只覆盖它认识的字段（不同 Emby/Jellyfin 版本支持度不一），
+	// 这里再按季/集排一次，确保拿到的是第一集而不是服务端返回顺序里的第一条。
+	children := body.Items
+	sort.SliceStable(children, func(i, j int) bool {
+		if a, b := remoteItemInt(children[i], "ParentIndexNumber"), remoteItemInt(children[j], "ParentIndexNumber"); a != b {
+			return a < b
+		}
+		if a, b := remoteItemInt(children[i], "IndexNumber"), remoteItemInt(children[j], "IndexNumber"); a != b {
+			return a < b
+		}
+		return remoteItemString(children[i], "SortName") < remoteItemString(children[j], "SortName")
+	})
+	for _, child := range children {
+		if id := remoteItemString(child, "Id"); id != "" && remoteItemHasMediaSources(child) {
+			return id
+		}
+	}
+	return ""
+}
+
+// remoteItemHasMediaSources 该条目是否自带媒体源（即能直接取流）。
+func remoteItemHasMediaSources(item map[string]any) bool {
+	if item == nil {
+		return false
+	}
+	switch sources := item["MediaSources"].(type) {
+	case []any:
+		return len(sources) > 0
+	case []map[string]any:
+		return len(sources) > 0
+	}
+	return false
+}
+
+// remoteItemIsContainer 该条目是否是「装别人」的容器：整部剧/季/目录/合集。
+func remoteItemIsContainer(item map[string]any) bool {
+	switch remoteItemString(item, "Type") {
+	case "Series", "Season", "Folder", "BoxSet", "CollectionFolder", "AggregateFolder", "Playlist":
+		return true
+	}
+	// 类型缺失时退回 IsFolder（Emby 的目录类条目都会带）。
+	if _, ok := item["Type"]; !ok {
+		if isFolder, okBool := item["IsFolder"].(bool); okBool {
+			return isFolder
+		}
+	}
+	return false
+}
+
 // ─── 远程 item JSON 取值辅助 ────────────────────────────────────────────────
 
 func anyString(v any) string {

internal/service/ffmpeg_tools_test.go

@@ -49,6 +49,8 @@ func TestSafeZipTargetRejectsTraversal(t *testing.T) {
 }
 
 func TestFFmpegToolsStatusNoPanic(t *testing.T) {
+	t.Setenv("PATH", t.TempDir())
+	t.Chdir(t.TempDir())
 	svc := NewFFmpegToolsService(&config.Config{}, zap.NewNop(), nil)
 	st := svc.Status(context.Background())
 	for _, key := range []string{"installing", "message", "error", "install_dir", "target", "ffmpeg", "ffprobe"} {

internal/service/image_proxy.go

@@ -14,14 +14,12 @@ package service
 
 import (
 	"context"
-	"errors"
 	"net"
 	"net/http"
 	"path/filepath"
 	"runtime"
 	"strings"
 	"sync"
-	"syscall"
 	"time"
 
 	"go.uber.org/zap"
@@ -103,53 +101,14 @@ func NewImageProxy(cfg *config.Config, log *zap.Logger) *ImageProxy {
 	// from image.tmdb.org via their HTTP proxy without extra config. On
 	// Windows we also honor the current user's system proxy settings.
 	transport := NewExternalTransport()
-	if proxyConfiguredForImageFetch() {
-		// 走本地代理（如 127.0.0.1:7890）时，拨号目标是代理本身，
-		// 连接层 SSRF 校验会误杀本地回环代理；此时沿用 URL 级校验。
-		log.Info("image proxy: outbound proxy detected, connection-level SSRF guard disabled")
-	} else {
-		// 仅 URL 解析层的 isPrivateHost 可被十进制/十六进制 IP、解析到
-		// 私网的域名与 DNS rebinding 绕过；在拨号层对最终连接 IP 做二次
-		// 校验（含重定向后的每条连接）堵住该旁路。
-		// 用户明确配置的远程挂载源（如内网 Emby）豁免该私网限制。
-		dialer := &net.Dialer{
-			Timeout: 15 * time.Second,
-			Control: func(_, address string, _ syscall.RawConn) error {
-				host, _, err := net.SplitHostPort(address)
-				if err != nil {
-					return err
-				}
-				if proxy.isAllowedRemoteHost(host) {
-					return nil
-				}
-				ip := net.ParseIP(host)
-				if ip == nil {
-					return errors.New("image proxy: refusing non-IP dial target")
-				}
-				if isPrivateIP(ip) {
-					return errors.New("image proxy: requests to private/internal hosts are not allowed")
-				}
-				return nil
-			},
-		}
-		transport.DialContext = dialer.DialContext
-	}
-
+	proxy.guardTransport(transport)
+	directTransport := NewInternalTransport()
+	proxy.guardTransport(directTransport)
 	proxy.client = &http.Client{Timeout: 30 * time.Second, Transport: transport}
-	proxy.directClient = &http.Client{Timeout: 30 * time.Second, Transport: NewInternalTransport()}
+	proxy.directClient = &http.Client{Timeout: 30 * time.Second, Transport: directTransport}
 	return proxy
 }
 
-// proxyConfiguredForImageFetch 探测环境变量或系统代理是否会影响图片抓取。
-func proxyConfiguredForImageFetch() bool {
-	req, err := http.NewRequest(http.MethodGet, "https://image.tmdb.org/", nil)
-	if err != nil {
-		return false
-	}
-	proxy, err := ProxyFromEnvironmentOrSystem(req)
-	return err == nil && proxy != nil
-}
-
 // SetLibraryRootsProvider injects a callback that returns the current set of
 // media library root directories. Sidecar posters live under these roots
 // (which are arbitrary, user-defined, and not necessarily under the

internal/service/image_proxy_allowed_hosts.go

@@ -0,0 +1,193 @@
+package service
+
+import (
+	"context"
+	"net"
+	"net/url"
+	"sort"
+	"strings"
+
+	"github.com/truewhile/MeBox/internal/repository"
+)
+
+// 显式配置的服务地址也要进入图片代理的私网放行名单。
+//
+// 背景（现场 2026-10-16）：MetaTube 刮削后端被配置成 docker 服务名
+// `http://mebox-metatube-1:8081`，刮削结果里的 poster_url 就是这个地址。
+// 图片代理的连接层 SSRF 校验只放行远程 Emby 挂载主机，于是每一次封面回源都
+// 被拒（日志 warn: requests to private/internal hosts are not allowed），
+// 前端长期只拿到占位图。
+//
+// 这里的取舍与 isPrivateHost 的注释一致：私网校验防的是「用户可控 URL 打到
+// 内网」，而管理员显式配置的服务地址属于受信输入，和远程 Emby 挂载同级。
+func configuredServiceImageHosts(values ...string) []string {
+	seen := make(map[string]struct{}, len(values))
+	out := make([]string, 0, len(values))
+	for _, raw := range values {
+		host := imageHostFromBaseURL(raw)
+		if host == "" {
+			continue
+		}
+		if _, ok := seen[host]; ok {
+			continue
+		}
+		seen[host] = struct{}{}
+		out = append(out, host)
+	}
+	sort.Strings(out)
+	return out
+}
+
+// internalServiceImageHosts 只保留指向内网/局域网的服务地址。
+//
+// 公网域名（image.tmdb.org、api.themoviedb.org 等）必须留在名单之外：放行后
+// 图片代理会优先走直连客户端，而 GFW 环境恰恰需要先用 HTTP(S)_PROXY。把它们
+// 加进来会让每张海报都白付一次直连失败。
+func internalServiceImageHosts(values ...string) []string {
+	hosts := configuredServiceImageHosts(values...)
+	out := make([]string, 0, len(hosts))
+	for _, host := range hosts {
+		if isInternalServiceHost(host) {
+			out = append(out, host)
+		}
+	}
+	return out
+}
+
+// isInternalServiceHost 判断主机名是否只在内网可达：私网/回环/链路本地 IP、
+// 单标签服务名（docker compose 服务名如 mebox-metatube-1），或常见的内部域名后缀。
+func isInternalServiceHost(host string) bool {
+	host = strings.ToLower(strings.TrimSpace(host))
+	if host == "" {
+		return false
+	}
+	if ip := net.ParseIP(host); ip != nil {
+		return isPrivateIP(ip)
+	}
+	// 单标签名（不含点）不是公网可解析的 FQDN，只可能是容器名/主机名。
+	if !strings.Contains(host, ".") {
+		return true
+	}
+	for _, suffix := range []string{".local", ".internal", ".lan", ".home.arpa", ".localhost"} {
+		if strings.HasSuffix(host, suffix) {
+			return true
+		}
+	}
+	return false
+}
+
+// imageHostFromBaseURL 从配置里的 base URL 取出小写主机名（不含端口）。
+// 容忍缺少 scheme 的写法（`mebox-metatube-1:8081`），并拒绝非 http/https。
+func imageHostFromBaseURL(raw string) string {
+	raw = strings.TrimSpace(raw)
+	if raw == "" {
+		return ""
+	}
+	if !strings.Contains(raw, "://") {
+		raw = "http://" + raw
+	}
+	u, err := url.Parse(raw)
+	if err != nil {
+		return ""
+	}
+	switch strings.ToLower(u.Scheme) {
+	case "http", "https":
+	default:
+		return ""
+	}
+	return strings.ToLower(strings.TrimSpace(u.Hostname()))
+}
+
+// mergeImageProxyAllowedHosts 合并多组放行主机（远程 Emby 挂载 + 显式配置的
+// 服务地址），去重时忽略大小写与端口。
+func mergeImageProxyAllowedHosts(groups ...[]string) []string {
+	seen := make(map[string]struct{})
+	out := make([]string, 0)
+	for _, group := range groups {
+		for _, raw := range group {
+			host := strings.ToLower(strings.TrimSpace(raw))
+			if host == "" {
+				continue
+			}
+			if h, _, err := net.SplitHostPort(host); err == nil {
+				host = strings.ToLower(strings.TrimSpace(h))
+			}
+			if host == "" {
+				continue
+			}
+			if _, ok := seen[host]; ok {
+				continue
+			}
+			seen[host] = struct{}{}
+			out = append(out, host)
+		}
+	}
+	sort.Strings(out)
+	return out
+}
+
+// imageMintingSettingKeys 是「其取值会被拼进图片 URL」的设置项。
+//
+// 只收录这类设置：放行名单每多一条就是多一处 SSRF 例外，因此 app.server_url、
+// strm.base_url 这些只出现在播放地址里的设置不在此列（它们由 streamURL 使用，
+// 与 /api/img 无关）。
+var imageMintingSettingKeys = []string{
+	"adult.scraper.metatube_server",
+	"adult.scraper.builtin_javdb_url",
+	"adult.scraper.builtin_javbus_url",
+}
+
+// imageMintingAPIProviders 是「其 base_url 会被拼进图片 URL」的 api_configs 行。
+//
+// 同样刻意收窄：api_configs 里每个 provider 都有一行出厂默认值，其中 metatube
+// 默认指向 http://127.0.0.1:7700。把整张表都放行会让每次安装都无条件放开回环
+// 地址，而只有真正用 MetaTube 刮削的部署才需要这个例外。
+var imageMintingAPIProviders = []string{"metatube"}
+
+// configuredServiceImageHostsFromRepo 从持久化配置里收集「会被写进图片 URL」的
+// 内网服务地址。只在构造阶段调用一次，所以直接读库、忽略错误——读不到只是少
+// 一条放行，不影响启动。
+func configuredServiceImageHostsFromRepo(ctx context.Context, repos *repository.Container) []string {
+	if repos == nil {
+		return nil
+	}
+	values := make([]string, 0, 4)
+
+	if repos.Setting != nil {
+		for _, key := range imageMintingSettingKeys {
+			if val, err := repos.Setting.Get(ctx, key); err == nil && strings.TrimSpace(val) != "" {
+				values = append(values, val)
+			}
+		}
+	}
+
+	if repos.ApiConfig != nil {
+		// 用 List 而不是 FindByProvider：后者走 First()，缺行时 gorm 会打一条
+		// "record not found" 日志，而「没配置 MetaTube」是完全正常的状态。
+		if rows, err := repos.ApiConfig.List(ctx); err == nil {
+			for _, row := range rows {
+				if !imageMintingAPIProvider(row.Provider) {
+					continue
+				}
+				if strings.TrimSpace(row.BaseURL) != "" {
+					values = append(values, row.BaseURL)
+				}
+				// Extra 是逗号/分号分隔的备用源列表（见 api_config.go 的 adult 默认值）。
+				values = append(values, adultConfiguredBases(row.Extra)...)
+			}
+		}
+	}
+
+	return internalServiceImageHosts(values...)
+}
+
+func imageMintingAPIProvider(provider string) bool {
+	provider = strings.ToLower(strings.TrimSpace(provider))
+	for _, allowed := range imageMintingAPIProviders {
+		if provider == allowed {
+			return true
+		}
+	}
+	return false
+}
+

internal/service/image_proxy_allowed_hosts_test.go

@@ -0,0 +1,249 @@
+package service
+
+import (
+	"context"
+	"sort"
+	"testing"
+
+	"github.com/glebarez/sqlite"
+	"gorm.io/gorm"
+
+	"github.com/truewhile/MeBox/internal/model"
+	"github.com/truewhile/MeBox/internal/repository"
+)
+
+// 现场（2026-10-16）：13 条媒体记录的 poster_url 指向
+// http://mebox-metatube-1:8081/v1/images/primary/... ，而图片代理只放行
+// 远程 Emby 挂载主机。MetaTube 是用户显式配置的刮削后端（settings 里的
+// adult.scraper.metatube_server），它跑在 docker 私网里，于是每一次封面
+// 回源都被连接层 SSRF 校验拒掉，日志刷 warn，前端长期只拿到占位图。
+//
+// 结论：显式配置的服务地址必须和远程 Emby 一样被加入放行名单。
+func TestConfiguredServiceImageHostsExtractsHostnames(t *testing.T) {
+	got := configuredServiceImageHosts(
+		"http://mebox-metatube-1:8081",
+		"https://api.themoviedb.org/3",
+		"http://192.168.1.50:8096/emby",
+	)
+	want := []string{"192.168.1.50", "api.themoviedb.org", "mebox-metatube-1"}
+	if len(got) != len(want) {
+		t.Fatalf("hosts = %v, want %v", got, want)
+	}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Fatalf("hosts = %v, want %v", got, want)
+		}
+	}
+}
+
+func TestConfiguredServiceImageHostsNormalizesAndDedupes(t *testing.T) {
+	got := configuredServiceImageHosts(
+		"  HTTP://MeBox-MetaTube-1:8081  ",
+		"http://mebox-metatube-1:8081/other/path",
+		"mebox-metatube-1:8081",
+	)
+	if len(got) != 1 || got[0] != "mebox-metatube-1" {
+		t.Fatalf("hosts = %v, want exactly [mebox-metatube-1]", got)
+	}
+}
+
+func TestConfiguredServiceImageHostsSkipsUnusableValues(t *testing.T) {
+	got := configuredServiceImageHosts(
+		"",
+		"   ",
+		"not a url at all",
+		"http://",
+		"://missing-scheme",
+		"/relative/path",
+		"ftp://files.example.com",
+	)
+	if len(got) != 0 {
+		t.Fatalf("hosts = %v, want none", got)
+	}
+}
+
+func TestConfiguredServiceImageHostsPreservesIPv6(t *testing.T) {
+	got := configuredServiceImageHosts("http://[fd00::5]:8081")
+	if len(got) != 1 || got[0] != "fd00::5" {
+		t.Fatalf("hosts = %v, want [fd00::5]", got)
+	}
+}
+
+func TestImageProxyAllowlistCombinesMountsAndConfiguredServices(t *testing.T) {
+	// 回归：合并后的名单必须同时覆盖远程 Emby 挂载与显式配置的服务，
+	// 否则两组图片里必有一组被 SSRF 校验挡住。
+	proxy := newUpstreamImageProxy(t)
+	proxy.SetAllowedRemoteHostsProvider(func() []string {
+		return mergeImageProxyAllowedHosts(
+			[]string{"192.168.1.50"},
+			configuredServiceImageHosts("http://mebox-metatube-1:8081"),
+		)
+	})
+
+	for _, host := range []string{"192.168.1.50", "mebox-metatube-1"} {
+		if !proxy.isAllowedRemoteHost(host) {
+			t.Fatalf("host %q must bypass the private-IP guard", host)
+		}
+	}
+	if proxy.isAllowedRemoteHost("evil.example.com") {
+		t.Fatal("unrelated hosts must stay blocked")
+	}
+}
+
+func TestMergeImageProxyAllowedHostsDedupesCaseInsensitively(t *testing.T) {
+	got := mergeImageProxyAllowedHosts(
+		[]string{"192.168.1.50:8096", "MeBox-MetaTube-1"},
+		[]string{"mebox-metatube-1", "", "  ", "192.168.1.50"},
+	)
+	sort.Strings(got)
+	want := []string{"192.168.1.50", "mebox-metatube-1"}
+	if len(got) != len(want) {
+		t.Fatalf("merged = %v, want %v", got, want)
+	}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Fatalf("merged = %v, want %v", got, want)
+		}
+	}
+}
+
+// 公网上游必须留在名单之外：放行后图片代理会优先走直连，而 GFW 环境恰恰要靠
+// HTTP(S)_PROXY 才能取到 image.tmdb.org 这类海报。只有内网地址才需要豁免。
+func TestInternalServiceImageHostsKeepsOnlyInternalAddresses(t *testing.T) {
+	got := internalServiceImageHosts(
+		"http://mebox-metatube-1:8081",
+		"http://127.0.0.1:7700",
+		"http://192.168.1.50:8096",
+		"https://api.themoviedb.org/3",
+		"https://image.tmdb.org",
+		"https://api.bgm.tv",
+		"https://javdb.com",
+		"http://nas.local:7700",
+		"http://media.internal",
+	)
+	want := []string{
+		"127.0.0.1",
+		"192.168.1.50",
+		"mebox-metatube-1",
+		"media.internal",
+		"nas.local",
+	}
+	sort.Strings(want)
+	if len(got) != len(want) {
+		t.Fatalf("hosts = %v, want %v", got, want)
+	}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Fatalf("hosts = %v, want %v", got, want)
+		}
+	}
+}
+
+func TestIsInternalServiceHost(t *testing.T) {
+	internal := []string{
+		"mebox-metatube-1", "127.0.0.1", "10.1.2.3", "172.18.0.5",
+		"192.168.1.50", "fd00::5", "nas.local", "media.internal",
+		"box.lan", "db.home.arpa",
+	}
+	for _, host := range internal {
+		if !isInternalServiceHost(host) {
+			t.Errorf("isInternalServiceHost(%q) = false, want true", host)
+		}
+	}
+	external := []string{
+		"", "api.themoviedb.org", "image.tmdb.org", "www.javbus.com",
+		"pics.dmm.co.jp", "8.8.8.8", "example.com",
+	}
+	for _, host := range external {
+		if isInternalServiceHost(host) {
+			t.Errorf("isInternalServiceHost(%q) = true, want false", host)
+		}
+	}
+}
+
+// 端到端装配：从真实的 settings / api_configs 里读出内网服务地址。
+// 现场数据是 settings.adult.scraper.metatube_server = http://mebox-metatube-1:8081，
+// 而 api_configs.metatube.base_url 还是出厂默认的 http://127.0.0.1:7700。
+func TestConfiguredServiceImageHostsFromRepoReadsSettingsAndAPIConfigs(t *testing.T) {
+	db, err := gorm.Open(sqlite.Open(":memory:"), &gorm.Config{})
+	if err != nil {
+		t.Fatalf("failed to open sqlite: %v", err)
+	}
+	if err := db.AutoMigrate(&model.Setting{}, &model.APIConfig{}); err != nil {
+		t.Fatalf("migrate: %v", err)
+	}
+	repos := repository.New(db)
+	ctx := context.Background()
+
+	if err := repos.Setting.Set(ctx, "adult.scraper.metatube_server", "http://mebox-metatube-1:8081"); err != nil {
+		t.Fatal(err)
+	}
+	if err := repos.Setting.Set(ctx, "adult.scraper.builtin_javdb_url", "https://javdb.com"); err != nil {
+		t.Fatal(err)
+	}
+	for _, row := range []model.APIConfig{
+		{Provider: "metatube", BaseURL: "http://127.0.0.1:7700", Enabled: true},
+		{Provider: "tmdb", BaseURL: "https://api.themoviedb.org/3", Enabled: true},
+		{Provider: "openai", BaseURL: "http://192.168.1.77:11434/v1", Enabled: true},
+	} {
+		if err := repos.ApiConfig.Create(ctx, &row); err != nil {
+			t.Fatal(err)
+		}
+	}
+
+	got := configuredServiceImageHostsFromRepo(ctx, repos)
+	// javdb.com / api.themoviedb.org 是公网地址，必须留在名单之外（否则图片
+	// 代理会改走直连，GFW 环境下反而取不到图）；openai 的 base_url 虽然指向
+	// 内网，但它从不产出图片 URL，同样不应成为 SSRF 例外。
+	want := []string{"127.0.0.1", "mebox-metatube-1"}
+	sort.Strings(want)
+	if len(got) != len(want) {
+		t.Fatalf("hosts = %v, want %v", got, want)
+	}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Fatalf("hosts = %v, want %v", got, want)
+		}
+	}
+}
+
+// 放行名单是 SSRF 例外，必须收窄：只有「会产出图片 URL」的配置才入选。
+func TestConfiguredServiceImageHostsFromRepoIgnoresUnrelatedInternalServices(t *testing.T) {
+	db, err := gorm.Open(sqlite.Open(":memory:"), &gorm.Config{})
+	if err != nil {
+		t.Fatalf("failed to open sqlite: %v", err)
+	}
+	if err := db.AutoMigrate(&model.Setting{}, &model.APIConfig{}); err != nil {
+		t.Fatalf("migrate: %v", err)
+	}
+	repos := repository.New(db)
+	ctx := context.Background()
+
+	// 只配置了内网的 OpenAI 兼容网关与内网下载器，没有任何图片来源。
+	if err := repos.Setting.Set(ctx, "app.server_url", "http://10.0.0.2:1314"); err != nil {
+		t.Fatal(err)
+	}
+	if err := repos.Setting.Set(ctx, "strm.base_url", "http://10.0.0.2:1314"); err != nil {
+		t.Fatal(err)
+	}
+	for _, row := range []model.APIConfig{
+		{Provider: "openai", BaseURL: "http://192.168.1.77:11434/v1", Enabled: true},
+		{Provider: "tmdb", BaseURL: "https://api.themoviedb.org/3", Enabled: true},
+	} {
+		if err := repos.ApiConfig.Create(ctx, &row); err != nil {
+			t.Fatal(err)
+		}
+	}
+
+	if got := configuredServiceImageHostsFromRepo(ctx, repos); len(got) != 0 {
+		t.Fatalf("hosts = %v, want none: unrelated internal services must not become SSRF exceptions", got)
+	}
+}
+
+func TestConfiguredServiceImageHostsFromRepoToleratesMissingRepos(t *testing.T) {
+	if got := configuredServiceImageHostsFromRepo(context.Background(), nil); len(got) != 0 {
+		t.Fatalf("hosts = %v, want none", got)
+	}
+}
+
+

internal/service/image_proxy_dial_allowlist_test.go

@@ -0,0 +1,112 @@
+package service
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"net/url"
+	"path/filepath"
+	"testing"
+
+	"go.uber.org/zap"
+
+	"github.com/truewhile/MeBox/internal/config"
+)
+
+// hostAndPortOf 从测试服务器的 URL 里取出主机名与端口。
+func hostAndPortOf(t *testing.T, raw string) (string, string) {
+	t.Helper()
+	u, err := url.Parse(raw)
+	if err != nil {
+		t.Fatalf("parse upstream url: %v", err)
+	}
+	return u.Hostname(), u.Port()
+}
+
+// 现场（2026-10-16，生产容器内实测）：放行名单里明明有 mebox-metatube-1，
+// 回源仍然失败于
+//
+//	image proxy: requests to private/internal hosts are not allowed
+//
+// 原因：连接层校验（dialer 的 Control 回调）拿到的 address 是**解析后的 IP**，
+// 而名单里存的是主机名（ConfiguredRemoteHosts 与
+// configuredServiceImageHostsFromRepo 都返回 u.Hostname()），两者永远不相等：
+// 那段 isAllowedRemoteHost 分支对主机名条目是死代码。
+//
+// 远程 Emby 挂载用 IP 配置时看不出问题，用主机名（NAS 名、docker 服务名）配置
+// 时同样中招。这个用例锁住「按主机名放行」这条路径。
+func TestImageProxyDialsAllowlistedHostname(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 := NewImageProxy(
+		&config.Config{Cache: config.CacheConfig{CacheDir: filepath.Join(t.TempDir(), "cache")}},
+		zap.NewNop(),
+	)
+	// 名单里放「主机名」而不是 IP —— 这正是生产里 mebox-metatube-1 的形态。
+	proxy.SetAllowedRemoteHostsProvider(func() []string { return []string{"localhost"} })
+
+	// 直接用该代理自己的 transport 拨号，才能命中连接层校验。
+	_, port := hostAndPortOf(t, upstream.URL)
+	resp, err := proxy.client.Get("http://localhost:" + port + "/poster.jpg")
+	if err != nil {
+		t.Fatalf("allowlisted hostname must be dialable, got: %v", err)
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("status = %d, want 200", resp.StatusCode)
+	}
+}
+
+// 反向约束：未放行的主机名解析到私网地址时仍必须被拒。
+// 这是连接层校验存在的理由（DNS rebinding），不能为了修上面那条而拆掉。
+func TestImageProxyStillBlocksUnlistedHostnameResolvingToPrivateIP(t *testing.T) {
+	upstream := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_, _ = w.Write(testJPEG)
+	}))
+	defer upstream.Close()
+
+	proxy := NewImageProxy(
+		&config.Config{Cache: config.CacheConfig{CacheDir: filepath.Join(t.TempDir(), "cache")}},
+		zap.NewNop(),
+	)
+	// 名单为空：localhost 解析到回环地址，必须被拦。
+	proxy.SetAllowedRemoteHostsProvider(func() []string { return nil })
+
+	_, port := hostAndPortOf(t, upstream.URL)
+	for _, client := range []*http.Client{proxy.client, proxy.directClient} {
+		resp, err := client.Get("http://localhost:" + port + "/poster.jpg")
+		if err == nil {
+			resp.Body.Close()
+			t.Fatal("a non-allowlisted hostname resolving to a private IP must be refused")
+		}
+	}
+}
+
+func TestImageProxyAllowsSelectedLocalProxy(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()
+	proxyURL, err := url.Parse(upstream.URL)
+	if err != nil {
+		t.Fatal(err)
+	}
+	p := NewImageProxy(&config.Config{}, zap.NewNop())
+	transport := NewExternalTransport()
+	transport.Proxy = http.ProxyURL(proxyURL)
+	p.guardTransport(transport)
+	t.Cleanup(transport.CloseIdleConnections)
+	client := &http.Client{Transport: transport}
+	resp, err := client.Get("http://images.example.com/poster.jpg")
+	if err != nil {
+		t.Fatalf("configured local proxy must remain usable: %v", err)
+	}
+	resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("status=%d, want 200", resp.StatusCode)
+	}
+}

internal/service/image_proxy_live_verify_test.go

@@ -0,0 +1,152 @@
+// 在真实 docker 网络里验证图片代理的私网放行名单修复。
+//
+// 现场（修复前）：MetaTube 刮削结果里的 poster_url 指向
+// http://mebox-metatube-1:8081/... ，而放行名单只含远程 Emby 挂载主机，
+// 于是每一次封面回源都被连接层 SSRF 校验拒掉，前端长期只拿到占位图。
+//
+// 本文件用与 service_builder.initImageProxy 相同的装配方式复现那条路径。
+// 真实回源那一步只在能解析到 MetaTube 服务名的环境里跑（生产容器网络）；
+// 开发机上会 Skip，因为那属于环境缺失，不是回归。
+//
+// 在生产机上运行的姿势（无需 Go 工具链）：
+//
+//	GOOS=linux GOARCH=amd64 CGO_ENABLED=0 go test -c -o /tmp/verify.test ./internal/service
+//	scp /tmp/verify.test root@HOST:/tmp/ && ssh root@HOST \
+//	  'docker cp /tmp/verify.test mebox:/tmp/ && docker exec mebox /tmp/verify.test \
+//	   -test.run TestImageProxyFetchesFromMetaTubeService -test.v'
+package service
+
+import (
+	"context"
+	"errors"
+	"net"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"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"
+)
+
+// 与生产环境一致的两个取值（取自 10.0.0.2 上真实的 settings / api_configs）。
+const (
+	metaTubeServer = "http://mebox-metatube-1:8081"
+	metaTubeLegacy = "http://127.0.0.1:7700"
+	sampleImageURL = "http://mebox-metatube-1:8081/v1/images/primary/JAV321/dsam-15"
+)
+
+// newMetaTubeVerificationRepos 建一个带现场取值的临时库（不碰生产数据）。
+func newMetaTubeVerificationRepos(t *testing.T) (*repository.Container, *ImageProxy) {
+	t.Helper()
+	db, err := gorm.Open(sqlite.Open(":memory:"), &gorm.Config{})
+	if err != nil {
+		t.Fatalf("open sqlite: %v", err)
+	}
+	if err := db.AutoMigrate(&model.Setting{}, &model.APIConfig{}); err != nil {
+		t.Fatalf("migrate: %v", err)
+	}
+	repos := repository.New(db)
+	ctx := context.Background()
+	if err := repos.Setting.Set(ctx, "adult.scraper.metatube_server", metaTubeServer); err != nil {
+		t.Fatalf("seed setting: %v", err)
+	}
+	if err := repos.ApiConfig.Create(ctx, &model.APIConfig{
+		Provider: "metatube", BaseURL: metaTubeLegacy, Enabled: true,
+	}); err != nil {
+		t.Fatalf("seed metatube api config: %v", err)
+	}
+	// 一个公网上游，用来确认它没有混进放行名单。
+	if err := repos.ApiConfig.Create(ctx, &model.APIConfig{
+		Provider: "tmdb", BaseURL: "https://api.themoviedb.org/3", Enabled: true,
+	}); err != nil {
+		t.Fatalf("seed tmdb api config: %v", err)
+	}
+
+	cfg := &config.Config{Cache: config.CacheConfig{CacheDir: filepath.Join(t.TempDir(), "cache")}}
+	cfg.Secrets.JWTSecret = "verify-secret"
+	proxy := NewImageProxy(cfg, zap.NewNop())
+	// 与 service_builder.initImageProxy 相同的装配方式（无远程 Emby 挂载：
+	// 本次只验证「显式配置的服务地址」这条路径）。
+	proxy.SetAllowedRemoteHostsProvider(func() []string {
+		return mergeImageProxyAllowedHosts(
+			nil,
+			configuredServiceImageHostsFromRepo(ctx, repos),
+		)
+	})
+	return repos, proxy
+}
+
+// TestImageProxyAllowlistIncludesConfiguredMetaTube 确认内网服务地址被放行、
+// 公网上游没有被放行。
+func TestImageProxyAllowlistIncludesConfiguredMetaTube(t *testing.T) {
+	repos, proxy := newMetaTubeVerificationRepos(t)
+	t.Logf("allowlist from config: %v",
+		configuredServiceImageHostsFromRepo(context.Background(), repos))
+
+	for _, host := range []string{"mebox-metatube-1", "127.0.0.1"} {
+		if !proxy.isAllowedRemoteHost(host) {
+			t.Fatalf("host %q must be allowlisted (it is an explicitly configured service)", host)
+		}
+	}
+	if proxy.isAllowedRemoteHost("api.themoviedb.org") {
+		t.Fatal("public upstreams must stay off the allowlist so image fetches keep using the proxy")
+	}
+}
+
+// TestImageProxyReachesMetaTubeService 证明放行名单真的解除了私网拦截。
+//
+// 现场（修复前）回源必然失败于
+//
+//	dial tcp 172.18.0.5:8081: image proxy: requests to private/internal hosts are not allowed
+//
+// 也就是连接根本没发生。修复后连接成功，失败只可能来自上游自己的 HTTP 状态
+// （MetaTube 冷缓存时对未刮削过的番号返回 500）。
+//
+// 因此这个用例断言的不是「拿到 200」，而是**「不再是被私网校验拒绝」**：
+// 后者才是本次修复的对象，前者取决于 MetaTube 缓存里有没有那条记录。
+func TestImageProxyReachesMetaTubeService(t *testing.T) {
+	_, proxy := newMetaTubeVerificationRepos(t)
+
+	data, ctype, err := proxy.Fetch(context.Background(), sampleImageURL)
+	if err != nil {
+		if isMetaTubeUnreachable(err) {
+			t.Skipf("mebox-metatube-1 is not reachable from this environment: %v", err)
+		}
+		if isPrivateHostRejection(err) {
+			t.Fatalf("the allowlist did not take effect: %v", err)
+		}
+		// 连接已经建立，只是上游自己返回了错误状态（MetaTube 冷缓存）。
+		t.Logf("reached metaTube; upstream returned its own error: %v", err)
+		return
+	}
+	t.Logf("fetched %d bytes (content-type=%s)", len(data), ctype)
+	if len(data) == 0 {
+		t.Fatal("fetched an empty image body")
+	}
+}
+
+// isPrivateHostRejection 报告错误是否来自连接层的私网校验。
+// 这正是本次修复必须消除的那一类失败。
+func isPrivateHostRejection(err error) bool {
+	return err != nil && strings.Contains(err.Error(), "requests to private/internal hosts are not allowed")
+}
+
+// isMetaTubeUnreachable 区分「环境里没有这个服务」与「放行名单没生效」。
+func isMetaTubeUnreachable(err error) bool {
+	var dnsErr *net.DNSError
+	if errors.As(err, &dnsErr) {
+		return true
+	}
+	msg := strings.ToLower(err.Error())
+	for _, needle := range []string{"no such host", "server misbehaving", "i/o timeout", "connection refused"} {
+		if strings.Contains(msg, needle) {
+			return true
+		}
+	}
+	return false
+}

internal/service/image_proxy_remote_fetch.go

@@ -32,7 +32,9 @@ func (p *ImageProxy) remoteImageFetchClients(host string) []remoteImageFetchClie
 	}
 	direct := p.directClient
 	if direct == nil {
-		direct = &http.Client{Timeout: timeout, Transport: NewInternalTransport()}
+		transport := NewInternalTransport()
+		p.guardTransport(transport)
+		direct = &http.Client{Timeout: timeout, Transport: transport}
 	}
 	directCandidate := remoteImageFetchClient{name: "direct", client: direct}
 	defaultCandidate := remoteImageFetchClient{name: "default", client: client}

internal/service/image_proxy_transport.go

@@ -0,0 +1,67 @@
+package service
+
+import (
+	"context"
+	"errors"
+	"net"
+	"net/http"
+	"net/url"
+	"strings"
+	"sync"
+	"syscall"
+	"time"
+)
+
+// guardTransport checks resolved destination addresses even when a deployment
+// has an outbound proxy. Only actual proxy endpoints and explicitly configured
+// image services may use private addresses; NO_PROXY and direct fallback
+// requests still need the connection-level check against DNS rebinding.
+func (p *ImageProxy) guardTransport(transport *http.Transport) {
+	var proxyEndpoints sync.Map
+	if selectProxy := transport.Proxy; selectProxy != nil {
+		transport.Proxy = func(req *http.Request) (*url.URL, error) {
+			proxyURL, err := selectProxy(req)
+			if err == nil && proxyURL != nil {
+				port := proxyURL.Port()
+				if port == "" {
+					switch strings.ToLower(proxyURL.Scheme) {
+					case "https":
+						port = "443"
+					case "socks5", "socks5h":
+						port = "1080"
+					default:
+						port = "80"
+					}
+				}
+				proxyEndpoints.Store(net.JoinHostPort(strings.ToLower(proxyURL.Hostname()), port), true)
+			}
+			return proxyURL, err
+		}
+	}
+	guardDialer := &net.Dialer{
+		Timeout: 15 * time.Second,
+		Control: func(_, address string, _ syscall.RawConn) error {
+			host, _, err := net.SplitHostPort(address)
+			if err != nil {
+				return err
+			}
+			ip := net.ParseIP(host)
+			if ip == nil || isPrivateIP(ip) {
+				return errors.New("image proxy: requests to private/internal hosts are not allowed")
+			}
+			return nil
+		},
+	}
+	trustedDialer := &net.Dialer{Timeout: 15 * time.Second}
+	transport.DialContext = func(ctx context.Context, network, address string) (net.Conn, error) {
+		host, _, err := net.SplitHostPort(address)
+		if err != nil {
+			return nil, err
+		}
+		_, isProxyEndpoint := proxyEndpoints.Load(strings.ToLower(address))
+		if isProxyEndpoint || p.isAllowedRemoteHost(host) {
+			return trustedDialer.DialContext(ctx, network, address)
+		}
+		return guardDialer.DialContext(ctx, network, address)
+	}
+}

internal/service/local_availability.go

@@ -50,9 +50,12 @@ func LookupLocalAvailability(ctx context.Context, repo *repository.Container, ti
 		return out
 	}
 	like := "%" + query + "%"
+	// 方言相关的 ILIKE/LIKE：PostgreSQL 下用 LIKE 会因标题大小写与查询不一致而
+	// 漏判「本地已有该剧」，进而反复提示订阅缺失。
+	likeOp := repository.LikeOperator(repo.DB)
 	var rows []model.Media
 	if err := repo.DB.WithContext(ctx).
-		Where("title LIKE ? OR original_name LIKE ? OR path LIKE ?", like, like, like).
+		Where("title "+likeOp+" ? OR original_name "+likeOp+" ? OR path "+likeOp+" ?", like, like, like).
 		Order("season_num asc, episode_num asc, episode_fraction asc, created_at desc").
 		Limit(2000).
 		Find(&rows).Error; err != nil {

internal/service/media_search_test.go

@@ -0,0 +1,69 @@
+package service
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/truewhile/MeBox/internal/model"
+)
+
+func TestGroupMediaSeriesItemsKeepsLibrariesSeparate(t *testing.T) {
+	for _, tc := range []struct {
+		name  string
+		media model.Media
+	}{
+		{"remote series with TMDb ID", model.Media{TMDbID: 61709}},
+		{"remote series with Bangumi ID", model.Media{BangumiID: 61709}},
+		{"flat episode with TMDb ID", model.Media{SeasonNum: 1, EpisodeNum: 1, TMDbID: 61709}},
+		{"flat episode with Bangumi ID", model.Media{SeasonNum: 1, EpisodeNum: 1, BangumiID: 61709}},
+		{"flat episode with Douban ID", model.Media{SeasonNum: 1, EpisodeNum: 1, DoubanID: "61709"}},
+		{"flat episode with TVDB ID", model.Media{SeasonNum: 1, EpisodeNum: 1, TheTVDBID: "61709"}},
+		{"shared series ID", model.Media{SeriesID: "same-series"}},
+		{"same title and directory", model.Media{Path: "/media/anime/龙珠改/Season 1/ep1.mkv", SeasonNum: 1, EpisodeNum: 1}},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			first, second := tc.media, tc.media
+			first.ID, first.LibraryID, first.LibraryName, first.Title = "first", "anime-a", "番剧库甲", "龙珠改"
+			second.ID, second.LibraryID, second.LibraryName, second.Title = "second", "anime-b", "番剧库乙", "龙珠改"
+			if strings.HasPrefix(tc.name, "remote") {
+				first.ID = EncodeEmbyRemoteID("mount-a", "same-id")
+				second.ID = EncodeEmbyRemoteID("mount-b", "same-id")
+			}
+
+			items := GroupMediaSeriesItems([]model.Media{first, second})
+			if len(items) != 2 {
+				t.Fatalf("got %d search results, want one per library: %#v", len(items), items)
+			}
+			seen := make(map[string]bool)
+			for _, item := range items {
+				seen[item.LibraryID] = true
+				if item.Title != "龙珠改" || item.LibraryName == "" {
+					t.Fatalf("search result lost title or library metadata: %#v", item)
+				}
+			}
+			if !seen[first.LibraryID] || !seen[second.LibraryID] {
+				t.Fatalf("search results lost a library: %#v", items)
+			}
+		})
+	}
+}
+
+func TestGroupMediaSeriesItemsFoldsEpisodesWithinDisplayLibrary(t *testing.T) {
+	items := []model.Media{
+		{Base: model.Base{ID: "a-2"}, LibraryID: "source-a", DisplayLibraryID: "anime-a", Title: "龙珠改", Path: "/media/anime/龙珠改/Season 1/ep2.mkv", SeasonNum: 1, EpisodeNum: 2, TMDbID: 61709},
+		{Base: model.Base{ID: "a-1"}, LibraryID: "source-b", DisplayLibraryID: "anime-a", Title: "龙珠改", Path: "/media/anime/龙珠改/Season 1/ep1.mkv", SeasonNum: 1, EpisodeNum: 1, TMDbID: 61709},
+		{Base: model.Base{ID: "b-2"}, LibraryID: "source-a", DisplayLibraryID: "anime-b", Title: "龙珠改", Path: "/media/anime/龙珠改/Season 1/ep2.mkv", SeasonNum: 1, EpisodeNum: 2, TMDbID: 61709},
+		{Base: model.Base{ID: "b-1"}, LibraryID: "source-a", DisplayLibraryID: "anime-b", Title: "龙珠改", Path: "/media/anime/龙珠改/Season 1/ep1.mkv", SeasonNum: 1, EpisodeNum: 1, TMDbID: 61709},
+	}
+	grouped := GroupMediaSeriesItems(items)
+	if len(grouped) != 2 {
+		t.Fatalf("got %d results, want one series per display library: %#v", len(grouped), grouped)
+	}
+	seen := make(map[string]bool)
+	for _, item := range grouped {
+		if item.EpisodeNum != 1 || seen[item.DisplayLibraryID] {
+			t.Fatalf("expected the first episode from each display library: %#v", grouped)
+		}
+		seen[item.DisplayLibraryID] = true
+	}
+}

internal/service/media_series.go

@@ -524,9 +524,16 @@ func groupMediaSeriesCardsByKeys(items []model.Media, keys []string) []SeriesCar
 // GroupMediaSeriesItems folds episode-level rows into one representative media
 // row per series. It is intended for search surfaces where applying a small
 // limit before series grouping would otherwise return several episodes from
-// the same show.
+// the same show. Each display library keeps its own search result.
 func GroupMediaSeriesItems(items []model.Media) []model.Media {
-	cards := groupMediaSeriesCards(items)
+	_, keys := resolveMediaSeriesKeys(items)
+	for i, item := range items {
+		// External IDs and SeriesID can be shared by different libraries. Scope
+		// search grouping to the same library used by the result's label/link,
+		// while preserving episode folding within merged display libraries.
+		keys[i] = seriesFingerprint(mediaTargetLibraryID(item), keys[i])
+	}
+	cards := groupMediaSeriesCardsByKeys(items, keys)
 	if len(cards) == 0 {
 		return []model.Media{}
 	}

internal/service/service_builder.go

@@ -159,28 +159,41 @@ func (b *serviceContainerBuilder) initContentServices() {
 	b.c.Stream.SetStrmPlayTargetResolver(b.c.Strm.ResolvePlayTargetWithUA)
 	// 弹幕识别需要把远程 Emby 条目解析为 Media 元数据及可拉取前 16MB 的直链 URL。
 	if b.c.EmbyRemote != nil {
-		b.c.Danmaku.SetRemoteMediaResolver(func(ctx context.Context, encodedID string) (*model.Media, string, error) {
-			mountID, remoteID, ok := DecodeEmbyRemoteID(encodedID)
-			if !ok {
-				return nil, "", errors.New("invalid emby remote id")
-			}
-			mount, acct, err := b.c.EmbyRemote.ResolveMount(ctx, mountID)
-			if err != nil || mount == nil || acct == nil {
-				if err != nil {
-					return nil, "", err
-				}
-				return nil, "", errors.New("emby mount or account not found")
-			}
-			m, err := b.c.EmbyRemote.RemoteMediaDetail(ctx, mount, acct, remoteID)
-			if err != nil || m == nil {
-				return nil, "", err
-			}
-			streamURL, _ := b.c.EmbyRemote.WebStreamURL(ctx, acct, remoteID)
-			return m, streamURL, nil
-		})
+		b.c.Danmaku.SetRemoteMediaResolver(b.resolveDanmakuRemoteMedia)
 	}
 }
 
+// resolveDanmakuRemoteMedia 把远程 Emby 伪 ID 解析成弹幕识别要用的 Media 与直链 URL。
+//
+// 目录类条目（整部剧/季/目录）自身没有媒体源，拿它的 ID 去请求远程
+// /Videos/{id}/stream 只会回 500（Unable to cast object of type 'Series' to type
+// 'IHasMediaSources'），哈希层于是每次都在这里白白失败、退回按文件名搜索。所以先按
+// 与 /api/stream 相同的规则回落到第一集：算出来的哈希对得上真正播放的那条流，元数据
+// （时长、文件名）也才是播放中那一集的。
+func (b *serviceContainerBuilder) resolveDanmakuRemoteMedia(ctx context.Context, encodedID string) (*model.Media, string, error) {
+	mountID, remoteID, ok := DecodeEmbyRemoteID(encodedID)
+	if !ok {
+		return nil, "", errors.New("invalid emby remote id")
+	}
+	mount, acct, err := b.c.EmbyRemote.ResolveMount(ctx, mountID)
+	if err != nil || mount == nil || acct == nil {
+		if err != nil {
+			return nil, "", err
+		}
+		return nil, "", errors.New("emby mount or account not found")
+	}
+	playable := b.c.EmbyRemote.PlayableRemoteID(ctx, acct, remoteID)
+	if strings.TrimSpace(playable) == "" {
+		playable = remoteID
+	}
+	m, err := b.c.EmbyRemote.RemoteMediaDetail(ctx, mount, acct, playable)
+	if err != nil || m == nil {
+		return nil, "", err
+	}
+	streamURL, _ := b.c.EmbyRemote.WebStreamURL(ctx, acct, playable)
+	return m, streamURL, nil
+}
+
 func (b *serviceContainerBuilder) initAccessAndStorageServices() {
 	b.c.PlayProfiles = NewPlayProfileService(b.log, b.repos)
 	b.c.Permissions = NewPermissionService(b.log, b.repos)
@@ -240,10 +253,21 @@ func (b *serviceContainerBuilder) initIdentityServices() {
 func (b *serviceContainerBuilder) initImageProxy() {
 	b.c.ImageProxy = NewImageProxy(b.cfg, b.log)
 	b.c.ImageProxy.SetLibraryRootsProvider(b.libraryRoots)
+	// 放行名单 = 远程 Emby 挂载主机 + 用户显式配置的内网服务地址。
+	//
+	// 只放行挂载主机是不够的：MetaTube 这类刮削后端会把「容器服务名」写进
+	// poster_url（现场：http://mebox-metatube-1:8081），连接层 SSRF 校验会把
+	// 每一次封面回源都拒掉，前端长期只拿到占位图。名单本身有 30s 缓存，
+	// 所以管理员改完设置无需重启即可生效。
+	b.c.ImageProxy.SetAllowedRemoteHostsProvider(func() []string {
+		var mounts []string
+		if b.c.EmbyRemote != nil {
+			mounts = b.c.EmbyRemote.ConfiguredRemoteHosts(context.Background())
+		}
+		services := configuredServiceImageHostsFromRepo(context.Background(), b.repos)
+		return mergeImageProxyAllowedHosts(mounts, services)
+	})
 	if b.c.EmbyRemote != nil {
-		b.c.ImageProxy.SetAllowedRemoteHostsProvider(func() []string {
-			return b.c.EmbyRemote.ConfiguredRemoteHosts(context.Background())
-		})
 		// 远程 Emby 的图片 URL 会带着签发时的 api_key 长期缓存/下发；轮换后
 		// URL 里的旧值必然过期。回源时改用账号当前 token，并在上游 401/403 时
 		// 重新登录一次再重试，封面图不再依赖人工改账号。

internal/service/transcoder.go

@@ -59,6 +59,9 @@ type TranscoderService struct {
 	availMu        sync.Mutex
 	availCheckedAt time.Time
 	availValue     bool
+	availKey       string
+	ffmpegMu       sync.Mutex
+	ffmpegCache    ffmpegValidationCache
 }
 
 // hlsJob holds the live state of one ffmpeg run.
@@ -83,6 +86,11 @@ type hlsJob struct {
 	quality string
 	// done is closed when the ffmpeg goroutine fully exits (after process death).
 	done chan struct{}
+	// Completed outputs stay reusable until idle eviction without occupying a
+	// running-process slot. The binary and config are fixed for this generation.
+	completed  bool
+	ffmpegPath string
+	config     *config.Config
 }
 
 var (
@@ -117,24 +125,37 @@ func (t *TranscoderService) Available() bool {
 	if t == nil || t.cfg == nil {
 		return false
 	}
-	if !t.cfg.Transcoder.Enabled {
+	cfg := t.transcodeConfig()
+	if !cfg.Transcoder.Enabled {
 		return false
 	}
+	encoder := ffmpegArgSettingsFromConfig(cfg, nil).encoder
+	key := cfg.App.FFmpegPath + "\x00" + encoder
 	t.availMu.Lock()
 	defer t.availMu.Unlock()
-	if !t.availCheckedAt.IsZero() && time.Since(t.availCheckedAt) < transcodeAvailabilityTTL {
-		return t.availValue
+	if !t.availValue && t.availKey == key && !t.availCheckedAt.IsZero() && time.Since(t.availCheckedAt) < transcodeAvailabilityTTL {
+		return false
 	}
-	_, err := t.resolveFFmpegPath()
+	_, err := t.resolveFFmpegPathFor(context.Background(), cfg.App.FFmpegPath, encoder)
 	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()
+	t.availKey = key
 	return available
 }
 
+func (t *TranscoderService) transcodeConfig() *config.Config {
+	config.RuntimeMu.RLock()
+	defer config.RuntimeMu.RUnlock()
+	return &config.Config{
+		App:        config.AppConfig{FFmpegPath: t.cfg.App.FFmpegPath, VAAPIDevice: t.cfg.App.VAAPIDevice},
+		Transcoder: t.cfg.Transcoder,
+	}
+}
+
 // 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)
@@ -174,7 +195,8 @@ func (t *TranscoderService) EnsureJobFromSubtitle(ctx context.Context, mediaID s
 // EnsureJobFromSubtitleQuality is EnsureJobFromSubtitle with an explicit local
 // HLS quality profile. Empty qualityID keeps the global transcoder config.
 func (t *TranscoderService) EnsureJobFromSubtitleQuality(ctx context.Context, mediaID string, startSec float64, seekGen int64, subtitleStream int, qualityID string) (string, error) {
-	if !t.cfg.Transcoder.Enabled {
+	cfg := t.transcodeConfig()
+	if !cfg.Transcoder.Enabled {
 		return "", ErrTranscodeDisabled
 	}
 	if startSec < 0 {
@@ -203,12 +225,17 @@ func (t *TranscoderService) EnsureJobFromSubtitleQuality(ctx context.Context, me
 
 	t.mu.Lock()
 	if existing, ok := t.jobs[mediaID]; ok {
-		if sameHLSConfiguration(existing, startSec, subtitleStream, qualityID) {
+		reusable := true
+		if existing.completed {
+			info, err := os.Stat(t.PlaylistPath(mediaID))
+			reusable = err == nil && info.Size() > 0
+		}
+		if reusable && sameHLSConfiguration(existing, startSec, subtitleStream, qualityID) {
 			t.touchJobLocked(mediaID)
 			t.mu.Unlock()
 			return t.PlaylistPath(mediaID), nil
 		}
-		if !shouldReplaceHLSJobConfigurationQuality(existing, startSec, seekGen, subtitleStream, qualityID) {
+		if reusable && !shouldReplaceHLSJobConfigurationQuality(existing, startSec, seekGen, subtitleStream, qualityID) {
 			t.touchJobLocked(mediaID)
 			t.mu.Unlock()
 			return t.PlaylistPath(mediaID), nil
@@ -230,7 +257,13 @@ func (t *TranscoderService) EnsureJobFromSubtitleQuality(ctx context.Context, me
 	}
 	input.Quality = quality
 	t.maybeFillDuration(ctx, m, input)
-	if _, err := t.resolveFFmpegPath(); err != nil {
+	settings := ffmpegArgSettingsFromConfig(cfg, quality)
+	jobEncoder := settings.encoder
+	if settings.noScale || subtitleStream >= 0 {
+		jobEncoder = ""
+	}
+	bin, err := t.resolveFFmpegPathFor(ctx, cfg.App.FFmpegPath, jobEncoder)
+	if err != nil {
 		return "", err
 	}
 
@@ -242,29 +275,15 @@ func (t *TranscoderService) EnsureJobFromSubtitleQuality(ctx context.Context, me
 	}
 
 	t.mu.Lock()
-	if existing, ok := t.jobs[mediaID]; ok {
-		if sameHLSConfiguration(existing, startSec, subtitleStream, qualityID) {
-			t.touchJobLocked(mediaID)
-			t.mu.Unlock()
-			return t.PlaylistPath(mediaID), nil
-		}
-		if !shouldReplaceHLSJobConfigurationQuality(existing, startSec, seekGen, subtitleStream, qualityID) {
-			t.touchJobLocked(mediaID)
-			t.mu.Unlock()
-			return t.PlaylistPath(mediaID), nil
-		}
-		prev := t.detachJobLocked(mediaID)
-		t.mu.Unlock()
-		waitJobExit(prev, 12*time.Second)
-		t.mu.Lock()
-	}
-	if max := t.maxConcurrent(); max > 0 && len(t.jobs) >= max {
+	if max := t.maxConcurrent(); max > 0 && t.runningJobCountLocked() >= max {
 		t.mu.Unlock()
 		return "", ErrTranscodeBusy
 	}
 
 	jobCtx, cancel := context.WithCancel(context.Background())
-	jobEncoder := t.effectiveEncoder()
+	if jobEncoder == "" {
+		jobEncoder = "software"
+	}
 	if subtitleStream >= 0 {
 		jobEncoder = "software (subtitle burn-in)"
 	}
@@ -280,6 +299,8 @@ func (t *TranscoderService) EnsureJobFromSubtitleQuality(ctx context.Context, me
 		subtitleStream: subtitleStream,
 		quality:        qualityID,
 		done:           make(chan struct{}),
+		ffmpegPath:     bin,
+		config:         cfg,
 	}
 	t.jobs[mediaID] = job
 	t.mu.Unlock()

internal/service/transcoder_availability_test.go

@@ -13,6 +13,8 @@ import (
 // be false both when transcoding is disabled by configuration and when no usable
 // ffmpeg can be resolved.
 func TestTranscoderAvailableGatesOnConfigAndBinary(t *testing.T) {
+	t.Setenv("PATH", t.TempDir())
+	t.Chdir(t.TempDir())
 	cases := []struct {
 		name       string
 		enabled    bool

internal/service/transcoder_ffmpeg_probe.go

@@ -11,29 +11,17 @@ import (
 	"time"
 
 	"go.uber.org/zap"
+
+	"github.com/truewhile/MeBox/internal/config"
 )
 
 func (t *TranscoderService) runFFmpeg(ctx context.Context, job *hlsJob, input transcodeInput) {
-	bin, err := t.resolveFFmpegPath()
-	if err != nil {
-		t.log.Warn("ffmpeg unavailable", zap.String("media_id", job.mediaID), zap.Error(err))
-		t.mu.Lock()
-		delete(t.jobs, job.mediaID)
-		t.mu.Unlock()
-		t.hub.Publish("transcode", map[string]any{
-			"media_id": job.mediaID,
-			"status":   "error",
-			"error":    err.Error(),
-		})
-		return
-	}
-
 	playlist := filepath.Join(job.outputDir, "index.m3u8")
 	segments := filepath.Join(job.outputDir, "seg_%05d.ts")
 
-	args := buildFFmpegArgsForInput(t.cfg, input, playlist, segments)
+	args := buildFFmpegArgsForInput(job.config, input, playlist, segments)
 
-	cmd := exec.CommandContext(ctx, bin, args...) // #nosec G204 -- bin is resolved by resolveFFmpegPath and args are passed without a shell.
+	cmd := exec.CommandContext(ctx, job.ffmpegPath, args...) // #nosec G204 -- the binary was validated before creating this job.
 	setFFmpegSysProcAttr(cmd)
 	cmd.Stderr = os.Stderr
 
@@ -50,7 +38,8 @@ func (t *TranscoderService) runFFmpeg(ctx context.Context, job *hlsJob, input tr
 		"start_sec": input.StartSec,
 	})
 
-	if err := cmd.Run(); err != nil && !errors.Is(ctx.Err(), context.Canceled) {
+	err := cmd.Run()
+	if err != nil && !errors.Is(ctx.Err(), context.Canceled) {
 		t.log.Warn("ffmpeg exited",
 			zap.String("media_id", job.mediaID),
 			zap.Float64("start_sec", input.StartSec),
@@ -59,11 +48,18 @@ func (t *TranscoderService) runFFmpeg(ctx context.Context, job *hlsJob, input tr
 	}
 
 	t.mu.Lock()
-	// Only drop the map entry if we are still the registered generation.
+	// Only update the map entry if we are still the registered generation.
 	if cur, ok := t.jobs[job.mediaID]; ok && cur == job {
-		delete(t.jobs, job.mediaID)
+		if err == nil && ctx.Err() == nil {
+			job.completed = true
+		} else {
+			delete(t.jobs, job.mediaID)
+		}
 	}
 	t.mu.Unlock()
+	if err != nil || ctx.Err() != nil {
+		job.cancel()
+	}
 
 	t.hub.Publish("transcode", map[string]any{
 		"media_id": job.mediaID,
@@ -73,19 +69,53 @@ func (t *TranscoderService) runFFmpeg(ctx context.Context, job *hlsJob, input tr
 }
 
 func (t *TranscoderService) resolveFFmpegPath() (string, error) {
-	t.mu.Lock()
-	configuredPath := strings.TrimSpace(t.cfg.App.FFmpegPath)
-	t.mu.Unlock()
+	cfg := t.transcodeConfig()
+	return t.resolveFFmpegPathFor(context.Background(), cfg.App.FFmpegPath, ffmpegArgSettingsFromConfig(cfg, nil).encoder)
+}
+
+type ffmpegValidationCache struct {
+	configuredPath string
+	path           string
+	encoder        string
+	size           int64
+	modTime        time.Time
+	checkedAt      time.Time
+}
+
+func (t *TranscoderService) resolveFFmpegPathFor(ctx context.Context, configuredPath, encoder string) (string, error) {
+	t.ffmpegMu.Lock()
+	defer t.ffmpegMu.Unlock()
+	if err := ctx.Err(); err != nil {
+		return "", err
+	}
+	configuredPath = strings.TrimSpace(configuredPath)
+	cached := t.ffmpegCache
+	if cached.path != "" && cached.encoder == encoder &&
+		(configuredPath == cached.configuredPath || configuredPath == cached.path) &&
+		time.Since(cached.checkedAt) < transcodeAvailabilityTTL {
+		if info, err := os.Stat(cached.path); err == nil && info.Size() == cached.size && info.ModTime().Equal(cached.modTime) {
+			return cached.path, nil
+		}
+	}
 
 	var lastErr error
 	for _, bin := range executableCandidates(configuredPath, "ffmpeg") {
-		if err := validateFFmpegForTranscode(context.Background(), bin, t.effectiveEncoder()); err != nil {
+		if err := validateFFmpegForTranscode(ctx, bin, encoder); err != nil {
+			if ctx.Err() != nil {
+				return "", ctx.Err()
+			}
 			lastErr = err
 			continue
 		}
-		t.mu.Lock()
-		t.cfg.App.FFmpegPath = bin
-		t.mu.Unlock()
+		if info, err := os.Stat(bin); err == nil {
+			t.ffmpegCache = ffmpegValidationCache{configuredPath: configuredPath, path: bin, encoder: encoder,
+				size: info.Size(), modTime: info.ModTime(), checkedAt: time.Now()}
+		}
+		config.RuntimeMu.Lock()
+		if strings.TrimSpace(t.cfg.App.FFmpegPath) == configuredPath {
+			t.cfg.App.FFmpegPath = bin
+		}
+		config.RuntimeMu.Unlock()
 		return bin, nil
 	}
 	if lastErr != nil {
@@ -95,10 +125,11 @@ func (t *TranscoderService) resolveFFmpegPath() (string, error) {
 }
 
 func (t *TranscoderService) effectiveEncoder() string {
-	if !t.cfg.Transcoder.HardwareAccel {
+	cfg := t.transcodeConfig()
+	if !cfg.Transcoder.HardwareAccel {
 		return ""
 	}
-	return normalizedHardwareEncoder(t.cfg.Transcoder.Encoder)
+	return normalizedHardwareEncoder(cfg.Transcoder.Encoder)
 }
 
 func normalizedHardwareEncoder(encoder string) string {
@@ -160,6 +191,7 @@ func hasFFmpegListEntry(output, name string) bool {
 
 // HumanFFmpegProfile is exposed for the admin UI / settings view.
 func (t *TranscoderService) HumanFFmpegProfile() string {
+	cfg := t.transcodeConfig()
 	return fmt.Sprintf("ffmpeg=%s, encoder=%s, output=%s",
-		t.cfg.App.FFmpegPath, t.cfg.Transcoder.Encoder, filepath.Join(t.cfg.Cache.CacheDir, "hls"))
+		cfg.App.FFmpegPath, cfg.Transcoder.Encoder, filepath.Join(t.cfg.Cache.CacheDir, "hls"))
 }

internal/service/transcoder_jobs.go

@@ -6,41 +6,58 @@ import (
 	"time"
 
 	"go.uber.org/zap"
+
+	"github.com/truewhile/MeBox/internal/config"
 )
 
 // WaitReady blocks (with a deadline) until the playlist file shows up on
 // disk for the *current* job generation. Stale playlists left behind by a
 // failed RemoveAll / still-exiting ffmpeg must not unblock a mid-file restart.
 func (t *TranscoderService) WaitReady(ctx context.Context, mediaID string, timeout time.Duration) bool {
-	deadline := time.Now().Add(timeout)
+	timer := time.NewTimer(timeout)
+	defer timer.Stop()
+	ticker := time.NewTicker(250 * time.Millisecond)
+	defer ticker.Stop()
 	for {
+		if ctx.Err() != nil {
+			return false
+		}
 		t.mu.Lock()
 		job, ok := t.jobs[mediaID]
 		var started time.Time
+		var completed bool
 		if ok {
 			started = job.startedAt
+			completed = job.completed
 		}
 		t.mu.Unlock()
+		if !ok {
+			return false // A failed/exited job cannot become ready later.
+		}
 		if ok {
 			if info, err := os.Stat(t.PlaylistPath(mediaID)); err == nil {
 				// Allow a small clock skew; reject anything older than this job.
 				if !info.ModTime().Before(started.Add(-2 * time.Second)) {
 					t.mu.Lock()
-					if j, exists := t.jobs[mediaID]; exists {
+					if j, exists := t.jobs[mediaID]; exists && j == job {
 						j.playlistOK = true
+						t.mu.Unlock()
+						return true
 					}
 					t.mu.Unlock()
-					return true
 				}
 			}
 		}
-		if time.Now().After(deadline) || ctx.Err() != nil {
-			return false
+		if completed {
+			return false // A completed job cannot regenerate a missing playlist.
 		}
 		select {
 		case <-ctx.Done():
 			return false
-		case <-time.After(250 * time.Millisecond):
+		case <-timer.C:
+			return false
+		case <-ticker.C:
+		case <-job.done:
 		}
 	}
 }
@@ -101,6 +118,9 @@ func (t *TranscoderService) Active() []ActiveJob {
 	defer t.mu.Unlock()
 	out := make([]ActiveJob, 0, len(t.jobs))
 	for _, j := range t.jobs {
+		if j.completed {
+			continue
+		}
 		out = append(out, ActiveJob{
 			MediaID:    j.mediaID,
 			Encoder:    j.encoder,
@@ -111,7 +131,19 @@ func (t *TranscoderService) Active() []ActiveJob {
 	return out
 }
 
+func (t *TranscoderService) runningJobCountLocked() int {
+	n := 0
+	for _, job := range t.jobs {
+		if !job.completed {
+			n++
+		}
+	}
+	return n
+}
+
 func (t *TranscoderService) maxConcurrent() int {
+	config.RuntimeMu.RLock()
+	defer config.RuntimeMu.RUnlock()
 	if t.cfg.Transcoder.MaxConcurrent <= 0 {
 		return 1
 	}
@@ -119,6 +151,8 @@ func (t *TranscoderService) maxConcurrent() int {
 }
 
 func (t *TranscoderService) idleTimeout() time.Duration {
+	config.RuntimeMu.RLock()
+	defer config.RuntimeMu.RUnlock()
 	if t.cfg.Transcoder.IdleTimeoutSeconds <= 0 {
 		return 120 * time.Second
 	}
@@ -137,21 +171,38 @@ func (t *TranscoderService) monitorIdle(ctx context.Context, job *hlsJob) {
 		case <-ticker.C:
 			t.mu.Lock()
 			current, ok := t.jobs[job.mediaID]
-			if !ok {
+			if !ok || current != job {
 				t.mu.Unlock()
 				return
 			}
 			idleFor := time.Since(current.lastAccess)
 			t.mu.Unlock()
-			if idleFor >= timeout {
-				t.log.Info("transcode idle timeout",
-					zap.String("media_id", job.mediaID),
-					zap.Duration("idle_for", idleFor),
-					zap.Duration("timeout", timeout),
-				)
-				t.StopJob(job.mediaID)
+			if idleFor >= timeout && t.stopIdleJob(job, timeout) {
 				return
 			}
 		}
 	}
 }
+
+func (t *TranscoderService) stopIdleJob(job *hlsJob, timeout time.Duration) bool {
+	gate := t.mediaStartGate(job.mediaID)
+	gate.Lock()
+	defer gate.Unlock()
+	t.mu.Lock()
+	current := t.jobs[job.mediaID]
+	if current != job {
+		t.mu.Unlock()
+		return true
+	}
+	idleFor := time.Since(job.lastAccess)
+	if idleFor < timeout {
+		t.mu.Unlock()
+		return false
+	}
+	t.detachJobLocked(job.mediaID)
+	t.mu.Unlock()
+	waitJobExit(job, 12*time.Second)
+	t.log.Info("transcode idle timeout", zap.String("media_id", job.mediaID),
+		zap.Duration("idle_for", idleFor), zap.Duration("timeout", timeout))
+	return true
+}

internal/service/transcoder_lifecycle_test.go

@@ -0,0 +1,228 @@
+package service
+
+import (
+	"context"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"runtime"
+	"strings"
+	"testing"
+	"time"
+
+	"go.uber.org/zap"
+
+	"github.com/truewhile/MeBox/internal/config"
+	"github.com/truewhile/MeBox/internal/model"
+	"github.com/truewhile/MeBox/internal/repository"
+)
+
+func TestHLSCompletedTranscodeStaysReadyAndIsReused(t *testing.T) {
+	bin, err := exec.LookPath("ffmpeg")
+	if err != nil {
+		t.Skip("ffmpeg is required for the real HLS lifecycle test")
+	}
+	ctx, cancel := context.WithTimeout(t.Context(), 10*time.Second)
+	defer cancel()
+	source := filepath.Join(t.TempDir(), "short.mp4")
+	cmd := exec.CommandContext(ctx, bin, "-hide_banner", "-loglevel", "error", "-f", "lavfi", "-i", "testsrc=size=64x64:rate=10", "-t", "1", "-c:v", "libx264", "-pix_fmt", "yuv420p", "-threads", "1", source)
+	if output, err := cmd.CombinedOutput(); err != nil {
+		t.Fatalf("generate source: %v: %s", err, output)
+	}
+	cfg := &config.Config{}
+	cfg.App.FFmpegPath = bin
+	cfg.Cache.CacheDir = t.TempDir()
+	cfg.Transcoder.Enabled = true
+	cfg.Transcoder.MaxConcurrent = 1
+	cfg.Transcoder.MaxHeight = 64
+	cfg.Transcoder.SegmentSeconds = 1
+	cfg.Transcoder.Threads = 1
+	db := newServiceTestDB(t)
+	media := model.Media{Path: source, DurationSec: 1}
+	if err := db.Create(&media).Error; err != nil {
+		t.Fatal(err)
+	}
+	svc := NewTranscoderService(cfg, zap.NewNop(), repository.New(db), NewHub(zap.NewNop()))
+	t.Cleanup(svc.StopAll)
+	playlist, err := svc.EnsureJob(ctx, media.ID)
+	if err != nil {
+		t.Fatal(err)
+	}
+	svc.mu.Lock()
+	job := svc.jobs[media.ID]
+	svc.mu.Unlock()
+	if job == nil {
+		t.Fatal("transcode lost its job before the playlist could be read")
+	}
+	select {
+	case <-job.done:
+	case <-ctx.Done():
+		t.Fatal("short transcode did not finish")
+	}
+	data, err := os.ReadFile(playlist)
+	if err != nil || !strings.Contains(string(data), "#EXT-X-ENDLIST") {
+		t.Fatalf("transcode did not produce a complete playlist: err=%v, data=%s", err, data)
+	}
+	if !svc.WaitReady(ctx, media.ID, 100*time.Millisecond) {
+		t.Fatal("completed HLS playlist is incorrectly reported as not ready")
+	}
+	if len(svc.Active()) != 0 {
+		t.Fatal("completed transcode still occupies an active job slot")
+	}
+	if _, err := svc.EnsureJob(ctx, media.ID); err != nil {
+		t.Fatal(err)
+	}
+	svc.mu.Lock()
+	reused := svc.jobs[media.ID] == job
+	svc.mu.Unlock()
+	if !reused {
+		t.Fatal("refreshing a completed playlist started a second transcode")
+	}
+	other := model.Media{Path: source + ".second", DurationSec: 1}
+	if err := os.WriteFile(other.Path, mustReadTestFile(t, source), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if err := db.Create(&other).Error; err != nil {
+		t.Fatal(err)
+	}
+	if _, err := svc.EnsureJob(ctx, other.ID); err != nil {
+		t.Fatalf("completed cache must not consume MaxConcurrent: %v", err)
+	}
+}
+
+func mustReadTestFile(t *testing.T, path string) []byte {
+	t.Helper()
+	data, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return data
+}
+
+func TestHLSWaitReadyReturnsImmediatelyForMissingJob(t *testing.T) {
+	svc := NewTranscoderService(&config.Config{}, zap.NewNop(), nil, NewHub(zap.NewNop()))
+	started := time.Now()
+	if svc.WaitReady(t.Context(), "failed", time.Second) {
+		t.Fatal("missing job must not be ready")
+	}
+	if elapsed := time.Since(started); elapsed >= 500*time.Millisecond {
+		t.Fatalf("missing/failed job waited for the readiness deadline: %v", elapsed)
+	}
+}
+
+func countingFFmpeg(t *testing.T) string {
+	t.Helper()
+	if runtime.GOOS == "windows" {
+		t.Skip("test ffmpeg uses a POSIX shell")
+	}
+	bin := filepath.Join(t.TempDir(), "ffmpeg")
+	script := `#!/bin/sh
+printf '%s\n' "$*" >> "$0.calls"
+case " $* " in
+  *" -encoders "*) printf ' V..... libx264\n V..... h264_nvenc\n' ;;
+  *" -muxers "*) printf ' E hls\n' ;;
+esac
+`
+	if err := os.WriteFile(bin, []byte(script), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	return bin
+}
+
+func TestFFmpegValidationIsCachedAndInvalidatedByEncoder(t *testing.T) {
+	bin := countingFFmpeg(t)
+	cfg := &config.Config{}
+	cfg.App.FFmpegPath = bin
+	svc := NewTranscoderService(cfg, zap.NewNop(), nil, NewHub(zap.NewNop()))
+	for range 3 {
+		if _, err := svc.resolveFFmpegPath(); err != nil {
+			t.Fatal(err)
+		}
+	}
+	calls := strings.Count(string(mustReadTestFile(t, bin+".calls")), "\n")
+	if calls != 2 {
+		t.Fatalf("repeated resolution spawned ffmpeg %d times, want one encoder/muxer probe pair", calls)
+	}
+	ApplyRuntimeSetting(cfg, "transcoder.hardware_accel", "true")
+	ApplyRuntimeSetting(cfg, "transcoder.encoder", "nvenc")
+	if _, err := svc.resolveFFmpegPath(); err != nil {
+		t.Fatal(err)
+	}
+	calls = strings.Count(string(mustReadTestFile(t, bin+".calls")), "\n")
+	if calls != 4 {
+		t.Fatalf("encoder change must revalidate capabilities, calls=%d", calls)
+	}
+	changed := time.Now().Add(time.Second)
+	if err := os.Chtimes(bin, changed, changed); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := svc.resolveFFmpegPath(); err != nil {
+		t.Fatal(err)
+	}
+	calls = strings.Count(string(mustReadTestFile(t, bin+".calls")), "\n")
+	if calls != 6 {
+		t.Fatalf("binary replacement must invalidate validation, calls=%d", calls)
+	}
+}
+
+func TestHLSIdleCleanupPreservesReplacementAndRecentActivity(t *testing.T) {
+	svc := NewTranscoderService(&config.Config{}, zap.NewNop(), nil, NewHub(zap.NewNop()))
+	done := make(chan struct{})
+	close(done)
+	cancelled := false
+	old := &hlsJob{mediaID: "m", lastAccess: time.Now().Add(-time.Hour)}
+	current := &hlsJob{mediaID: "m", lastAccess: time.Now(), done: done, cancel: func() { cancelled = true }}
+	svc.jobs["m"] = current
+	if !svc.stopIdleJob(old, time.Minute) || cancelled || svc.jobs["m"] != current {
+		t.Fatal("an old watchdog must leave the replacement generation intact")
+	}
+	if svc.stopIdleJob(current, time.Minute) || cancelled {
+		t.Fatal("recent segment access must prevent idle cancellation")
+	}
+	current.lastAccess = time.Now().Add(-time.Hour)
+	if !svc.stopIdleJob(current, time.Minute) || !cancelled || svc.jobs["m"] != nil {
+		t.Fatal("an idle generation must release its resources")
+	}
+}
+
+func TestHLSOriginalQualityValidatesSoftwareEncoder(t *testing.T) {
+	bin := countingFFmpeg(t)
+	softwareOnly := strings.ReplaceAll(string(mustReadTestFile(t, bin)), ` V..... h264_nvenc\n`, "")
+	if err := os.WriteFile(bin, []byte(softwareOnly), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	t.Setenv("PATH", t.TempDir())
+	t.Chdir(t.TempDir())
+	cfg := &config.Config{}
+	cfg.App.FFmpegPath = bin
+	cfg.Transcoder.Enabled = true
+	cfg.Transcoder.HardwareAccel = true
+	cfg.Transcoder.Encoder = "nvenc"
+	cfg.Cache.CacheDir = t.TempDir()
+	db := newServiceTestDB(t)
+	source := filepath.Join(t.TempDir(), "movie.mkv")
+	if err := os.WriteFile(source, []byte("source"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	media := model.Media{Path: source, DurationSec: 60}
+	if err := db.Create(&media).Error; err != nil {
+		t.Fatal(err)
+	}
+	svc := NewTranscoderService(cfg, zap.NewNop(), repository.New(db), NewHub(zap.NewNop()))
+	t.Cleanup(svc.StopAll)
+	if _, err := svc.EnsureJobFromSubtitleQuality(t.Context(), media.ID, 0, 1, -1, "source"); err != nil {
+		t.Fatalf("source quality uses software and must not require NVENC: %v", err)
+	}
+	svc.mu.Lock()
+	job := svc.jobs[media.ID]
+	svc.mu.Unlock()
+	select {
+	case <-job.done:
+	case <-time.After(time.Second):
+		t.Fatal("fake ffmpeg did not finish")
+	}
+	calls := string(mustReadTestFile(t, bin+".calls"))
+	if strings.Count(calls, "\n") != 3 || !strings.Contains(calls, "-c:v libx264") {
+		t.Fatalf("startup must probe once and encode with libx264, calls=%s", calls)
+	}
+}

web/src/api/client.ts

@@ -2,7 +2,7 @@ import axios, { AxiosError, type InternalAxiosRequestConfig } from 'axios'
 
 import { useAuthStore } from '../stores/auth'
 import { getActivePlayProfileId, getActivePlayProfilePinToken } from '../stores/playProfile'
-import type { RefreshOutcome } from '../utils/authRefresh'
+import { createTokenPreflight, type RefreshOutcome } from '../utils/authRefresh'
 
 // Single axios instance used by every API helper. Adds the JWT to outgoing
 // requests and refreshes it on 401. Only an explicit server-side rejection of
@@ -48,8 +48,27 @@ function isRefreshRequest(config?: InternalAxiosRequestConfig | null): boolean {
   return Boolean(config?.url?.includes('/auth/refresh'))
 }
 
+// 请求发出前的令牌预检。本地持久化的 token 往往已经过期，而首屏有多个各自
+// 独立的 useEffect 会并发发请求；不预检的话它们会一起拿到 401，控制台刷出
+// 一整屏红字（现场：2026-10-16 12:49，7 个请求同时 401），用户还要等拦截器
+// 刷新完才看到内容。预检把刷新提到请求之前，且刷新协调器本身是单飞的。
+const preflightAccessToken = createTokenPreflight({
+  readAccessToken: () => useAuthStore.getState().token,
+  refreshSession: () => useAuthStore.getState().refreshSession(),
+})
+
 // Add auth token to requests
-api.interceptors.request.use((config) => {
+api.interceptors.request.use(async (config) => {
+  // 刷新请求自己绝不能走预检，否则会递归刷新。
+  if (!isRefreshRequest(config)) {
+    try {
+      await preflightAccessToken()
+    } catch {
+      // 会话已被服务端判定失效：直接失败，不发无意义的 401 请求。
+      // 清空会话与跳转登录页由刷新协调器和 401 拦截器负责，保持一致。
+      throw new axios.Cancel('session expired')
+    }
+  }
   const token = useAuthStore.getState().token
   if (token) {
     config.headers = config.headers ?? {}

web/src/api/danmaku.ts

@@ -102,7 +102,10 @@ export const danmakuAPI = {
     api
       .get<DanmakuFetchResult>(`/danmaku/${encodeURIComponent(mediaId)}`, {
         params: { kw: options.kw || undefined, episodeId: options.episodeId || undefined },
-        timeout: 20_000,
+        // 服务端依次做「16MB 前缀哈希 → 上游匹配 → 评论库」三段串行请求，每段的上游
+        // 超时都是 20 秒。客户端要是也卡在 20 秒，就会在服务端正要降级返回（哈希超时
+        // 后改按文件名搜索）时把请求撤掉：用户看不到弹幕，服务端还记一条 404。
+        timeout: 45_000,
       })
       .then((r) => r.data),
 

web/src/api/playback.ts

@@ -42,9 +42,9 @@ function publicOriginHeader() {
 }
 
 export const playbackAPI = {
-  getResume: (mediaId: string) =>
+  getResume: (mediaId: string, signal?: AbortSignal) =>
     api
-      .get<{ position_ms: number; duration_ms: number; completed: boolean }>(`/playback/${mediaId}/resume`)
+      .get<{ position_ms: number; duration_ms: number; completed: boolean }>(`/playback/${encodeURIComponent(mediaId)}/resume`, { signal })
       .then((r) => r.data),
 
   // 片头/片尾片段：播放开始后再调用，服务端可能需要几秒去外部数据库取数，

web/src/appRoutes.tsx

@@ -6,6 +6,7 @@ const HomePageLoader = () => import('./pages/HomePage').then((m) => ({ default:
 const LibraryPageLoader = () => import('./pages/LibraryPage').then((m) => ({ default: m.LibraryPage }))
 const LibrariesPageLoader = () => import('./pages/LibrariesPage').then((m) => ({ default: m.LibrariesPage }))
 const FavouritesPageLoader = () => import('./pages/FavouritesPage').then((m) => ({ default: m.FavouritesPage }))
+const MediaSearchPageLoader = () => import('./pages/MediaSearchPage').then((m) => ({ default: m.MediaSearchPage }))
 const PlaylistsPageLoader = () => import('./pages/PlaylistsPage').then((m) => ({ default: m.PlaylistsPage }))
 const PlaylistDetailPageLoader = () =>
   import('./pages/PlaylistDetailPage').then((m) => ({ default: m.PlaylistDetailPage }))
@@ -21,6 +22,7 @@ const HomePage = lazy(HomePageLoader)
 const LibraryPage = lazy(LibraryPageLoader)
 const LibrariesPage = lazy(LibrariesPageLoader)
 const FavouritesPage = lazy(FavouritesPageLoader)
+const MediaSearchPage = lazy(MediaSearchPageLoader)
 const PlaylistsPage = lazy(PlaylistsPageLoader)
 const PlaylistDetailPage = lazy(PlaylistDetailPageLoader)
 const MediaDetailPage = lazy(MediaDetailPageLoader)
@@ -73,6 +75,7 @@ export function prefetchCommonRouteChunks() {
     LibrariesPageLoader,
     LibraryPageLoader,
     MediaDetailPageLoader,
+    MediaSearchPageLoader,
     FavouritesPageLoader,
     WatchHistoryPageLoader,
     PlaylistsPageLoader,
@@ -98,6 +101,7 @@ export type AppRoute = {
 
 export const appRoutes: AppRoute[] = [
   { index: true, element: <HomePage /> },
+  { path: 'search', element: <MediaSearchPage /> },
   { path: 'libraries', element: <LibrariesPage /> },
   { path: 'library/:id', element: <LibraryPage /> },
   { path: 'favourites', element: <FavouritesPage /> },

web/src/components/Layout.tsx

@@ -25,6 +25,7 @@ function isMediaView(pathname: string, search: string): boolean {
 
   return (
     pathname === '/' ||
+    pathname === '/search' ||
     pathname === '/libraries' ||
     pathname.startsWith('/library') ||
     pathname.startsWith('/media') ||

web/src/components/LayoutHeaderSections.tsx

@@ -7,7 +7,9 @@ import { ARTWORK, imageURL } from '../api/client'
 import { mediaAPI } from '../api/library'
 import { readerAPI, type ReaderBook } from '../api/reader'
 import type { Media, PlayProfile, User } from '../types'
-import { favouriteMediaLink } from '../utils/mediaNavigation'
+import { favouriteMediaLink, mediaSearchLink, mediaSearchReturnPath } from '../utils/mediaNavigation'
+import { useAuthStore } from '../stores/auth'
+import { usePlayProfileStore } from '../stores/playProfile'
 import { resolveHeaderBack } from './layoutNavigation'
 import { LayoutThemeToggle } from './LayoutThemeToggle'
 import { LayoutUserMenu } from './LayoutUserMenu'
@@ -56,7 +58,11 @@ export function LayoutHeader({
   pathname = '',
 }: LayoutHeaderProps) {
   const navigate = useNavigate()
-  const headerBack = resolveHeaderBack(pathname)
+  const location = useLocation()
+  const searchFrom = mediaSearchReturnPath(location.state)
+  const headerBack = searchFrom && (pathname.startsWith('/library/') || pathname.startsWith('/media/'))
+    ? { to: searchFrom, label: '搜索结果' }
+    : resolveHeaderBack(pathname)
 
   return (
     <header className="relative z-30 flex h-16 shrink-0 items-center justify-between gap-2 border-b border-[var(--app-border)] bg-[var(--app-header-bg)] px-3 backdrop-blur-md sm:h-20 sm:gap-4 sm:px-4 md:px-8">
@@ -295,7 +301,12 @@ function LayoutHeaderBookSearch() {
 const SEARCH_PAGE_SIZE = 8
 
 function LayoutHeaderSearch() {
-  const [query, setQuery] = useState('')
+  const location = useLocation()
+  const userId = useAuthStore((s) => s.user?.id)
+  const profileId = usePlayProfileStore((s) => s.activeProfileId)
+  const [query, setQuery] = useState(() =>
+    location.pathname === '/search' ? new URLSearchParams(location.search).get('q') ?? '' : '',
+  )
   const [isOpen, setIsOpen] = useState(false)
   const [loading, setLoading] = useState(false)
   const [loadingMore, setLoadingMore] = useState(false)
@@ -312,7 +323,6 @@ function LayoutHeaderSearch() {
   const resultsRef = useRef<Media[]>([])
   const hasMoreRef = useRef(false)
   const navigate = useNavigate()
-  const location = useLocation()
   const locationKey = `${location.pathname}${location.search}`
   const prevLocationKeyRef = useRef(locationKey)
 
@@ -392,14 +402,21 @@ function LayoutHeaderSearch() {
       }
     }, 250)
 
-    return () => clearTimeout(timer)
-  }, [query])
+    return () => {
+      clearTimeout(timer)
+      // 卸载或关键词变化时，已发出的旧请求也不能覆盖当前结果。
+      searchSeqRef.current += 1
+    }
+  }, [query, userId, profileId])
 
   useLayoutEffect(() => {
     if (prevLocationKeyRef.current === locationKey) return
     prevLocationKeyRef.current = locationKey
-    dismissSearchRef.current(true)
-  }, [locationKey])
+    dismissSearchRef.current(false)
+    if (location.pathname === '/search') {
+      setQuery(new URLSearchParams(location.search).get('q') ?? '')
+    }
+  }, [locationKey, location.pathname, location.search])
 
   useEffect(() => {
     function handlePointerDown(e: PointerEvent) {
@@ -446,28 +463,34 @@ function LayoutHeaderSearch() {
     }
   }
 
-  const handleSelect = (item: Media) => {
+  const openSearchPage = () => {
     if (suppressInputRef.current) return
-    const to = favouriteMediaLink(item)
-    dismissSearch(true)
-    navigate(to)
+    const trimmed = query.trim()
+    if (!trimmed) return
+    dismissSearch(false)
+    navigate(mediaSearchLink(trimmed))
   }
 
   const handleKeyDown = (e: React.KeyboardEvent) => {
     if (e.key === 'Escape') {
       dismissSearch(false)
-    } else if (e.key === 'Enter' && results.length > 0) {
-      handleSelect(results[0])
+    } else if (e.key === 'Enter' && !e.nativeEvent.isComposing && e.keyCode !== 229) {
+      e.preventDefault()
+      openSearchPage()
     }
   }
 
   return (
     <div ref={containerRef} className="relative w-full">
       <div className="relative flex items-center">
-        <Search
-          size={15}
-          className="absolute left-3 text-[var(--app-muted)] pointer-events-none transition-colors group-focus-within:text-brand-500 sm:left-3.5 sm:text-[16px]"
-        />
+        <button
+          type="button"
+          onClick={openSearchPage}
+          aria-label="搜索媒体"
+          className="absolute left-3 rounded text-[var(--app-muted)] transition-colors hover:text-brand-500 sm:left-3.5"
+        >
+          <Search size={15} />
+        </button>
         <input
           ref={inputRef}
           type="text"
@@ -488,6 +511,7 @@ function LayoutHeaderSearch() {
           }}
           onKeyDown={handleKeyDown}
           placeholder="搜索媒体…"
+          aria-label="搜索媒体关键词"
           className="w-full h-9 sm:h-10 pl-8 sm:pl-10 pr-8 sm:pr-9 rounded-xl sm:rounded-2xl border border-[var(--app-border)] bg-[var(--app-panel)] text-xs sm:text-sm text-[var(--app-text)] placeholder:text-[var(--app-muted)] shadow-sm outline-none transition-all duration-200 focus:border-brand-500 focus:ring-2 focus:ring-brand-500/20 focus:bg-[var(--app-panel-elevated)]"
         />
         {loading ? (
@@ -498,6 +522,7 @@ function LayoutHeaderSearch() {
             onClick={() => {
               dismissSearch(true)
             }}
+            aria-label="清空媒体搜索"
             className="absolute right-3 text-[var(--app-muted)] hover:text-[var(--app-text)] p-0.5 rounded-lg"
           >
             <X size={15} />
@@ -538,15 +563,11 @@ function LayoutHeaderSearch() {
             ) : (
               <div className="space-y-1">
                 {results.map((item) => (
-                  <button
+                  <Link
                     key={item.id}
-                    type="button"
-                    onPointerDown={(e) => {
-                      if (e.button !== 0) return
-                      e.preventDefault()
-                      handleSelect(item)
-                    }}
-                    onClick={() => handleSelect(item)}
+                    to={favouriteMediaLink(item)}
+                    state={{ from: locationKey }}
+                    onClick={() => dismissSearch(false)}
                     className="flex w-full items-center gap-3 rounded-xl p-2 text-left transition-colors hover:bg-[var(--app-hover)] group"
                   >
                     <div className="relative h-12 w-9 shrink-0 overflow-hidden rounded-lg bg-[var(--app-panel-soft)]">
@@ -590,7 +611,7 @@ function LayoutHeaderSearch() {
                         )}
                       </div>
                     </div>
-                  </button>
+                  </Link>
                 ))}
                 {loadingMore && (
                   <div className="flex items-center justify-center gap-2 py-3 text-[10px] text-[var(--app-muted)]">
@@ -605,6 +626,15 @@ function LayoutHeaderSearch() {
                 )}
               </div>
             )}
+            <Link
+              to={mediaSearchLink(query)}
+              onClick={() => dismissSearch(false)}
+              className="sticky -bottom-2 flex items-center justify-center gap-2 border-t border-[var(--app-border)] bg-[var(--app-panel)] px-3 py-3 text-xs font-bold text-brand-500 hover:bg-[var(--app-hover)]"
+            >
+              <Search size={14} />
+              查看全部搜索结果
+              <span className="hidden text-[10px] font-normal text-[var(--app-muted)] sm:inline">按回车搜索</span>
+            </Link>
           </div>
         ) : null}
     </div>

web/src/components/MediaCard.tsx

@@ -10,7 +10,7 @@ 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, linkState, onClick, actions, renderActions, compact,
+  media, progress, count, rating, linkTo, linkState, onClick, actions, renderActions, compact, showLibrary,
 }: {
   media: Media
   progress?: number
@@ -22,6 +22,7 @@ export const MediaCard = memo(function MediaCard({
   actions?: ReactNode
   renderActions?: (media: Media) => ReactNode
   compact?: boolean
+  showLibrary?: boolean
 }) {
   const ref = useRef<HTMLDivElement>(null)
   const href = linkTo ?? `/media/${media.id}`
@@ -31,6 +32,7 @@ export const MediaCard = memo(function MediaCard({
   const blurredPosterSrc = imageURL(media.poster_url, media.updated_at, ARTWORK.posterTiny)
   const displayRating = rating ?? media.rating
   const versionCount = media.versions?.length ?? 0
+  const libraryName = media.display_library_name || media.library_name
   // renderActions 延迟到卡片自身渲染时才调用，保证 memo 生效
   const actionContent = actions ?? renderActions?.(media)
 
@@ -131,6 +133,11 @@ export const MediaCard = memo(function MediaCard({
           }`}>
             {media.title}
           </p>
+          {showLibrary && libraryName && (
+            <p title={libraryName} className="truncate text-[10px] text-[var(--app-muted)] sm:text-xs">
+              {libraryName}
+            </p>
+          )}
           {/* 单行、不折行：年份可截断，徽标永不压缩。折行会让同一行里各卡片的
               徽标落位高低不一，看起来就是「错位」，所以宁可极窄屏上截断年份。
               min-h 统一行高，避免有/无徽标的卡片年份基线差几像素。 */}

web/src/components/layoutNavigation.ts

@@ -74,7 +74,7 @@ export const MOBILE_BOTTOM_NAV_ITEMS: MobileBottomNavItem[] = [
   },
 ]
 
-const ROOT_MEDIA_PATHS = new Set(['/', '/libraries', '/favourites', '/playlists', '/history'])
+const ROOT_MEDIA_PATHS = new Set(['/', '/search', '/libraries', '/favourites', '/playlists', '/history'])
 
 /** True only for the fullscreen player route (`/play` or `/play/:id`), not `/playlists`. */
 export function isPlayerRoute(pathname: string): boolean {
@@ -131,7 +131,7 @@ export function resolveHeaderBack(pathname: string): HeaderBackTarget | null {
   if (pathname === '/history/stats') {
     return { to: '/history', label: '观看历史' }
   }
-  if (pathname === '/playlists' || pathname === '/favourites' || pathname === '/history') {
+  if (pathname === '/search' || pathname === '/playlists' || pathname === '/favourites' || pathname === '/history') {
     return { to: '/', label: '首页' }
   }
   if (pathname.startsWith('/playlist/')) {
@@ -171,6 +171,7 @@ export function shouldShowMobileBottomNav(pathname: string): boolean {
   if (isPlayerRoute(pathname)) return false
   return (
     pathname === '/' ||
+    pathname === '/search' ||
     pathname === '/libraries' ||
     pathname.startsWith('/library') ||
     pathname.startsWith('/media') ||

web/src/hooks/useScrollMemory.test.ts

@@ -11,6 +11,7 @@ function check(name: string, condition: boolean) {
 }
 
 check('home remembers scroll', shouldRememberScroll('/'))
+check('media search remembers scroll', shouldRememberScroll('/search'))
 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'))

web/src/hooks/useScrollMemory.ts

@@ -63,6 +63,7 @@ function writeScrollPosition(key: string, value: number): void {
 export function shouldRememberScroll(pathname: string): boolean {
   return (
     pathname === '/' ||
+    pathname === '/search' ||
     pathname === '/libraries' ||
     pathname.startsWith('/library/') ||
     pathname.startsWith('/media/')

web/src/hooks/useWebSocket.ts

@@ -1,38 +1,13 @@
 import { useEffect, useRef } from 'react'
 
 import { useAuthStore } from '../stores/auth'
+import { isAccessTokenStale } from '../utils/authRefresh'
 
 // 前 5 次沿用原快速退避间隔；之后进入 60s 慢速重试并不再放弃，
 // 服务重启或网络恢复后仍能自动重连（清理函数可随时取消定时器）。
 const FAST_RECONNECT_ATTEMPTS = 5
 const SLOW_RECONNECT_INTERVAL = 60_000
 
-// token 过期前 30 秒即视为失效，给刷新留出余量。
-const TOKEN_EXPIRY_SKEW_MS = 30_000
-
-// decodeJwtExpiryMs 解析 JWT 的 exp 声明并换算成毫秒时间戳。
-// 解析失败（非 JWT / 结构异常）返回 null，调用方据此放行重连。
-function decodeJwtExpiryMs(token: string): number | null {
-  const segment = token.split('.')[1]
-  if (!segment) return null
-  try {
-    // JWT 使用 base64url 编码，需先还原字符集并补齐 padding。
-    const normalized = segment.replace(/-/g, '+').replace(/_/g, '/')
-    const padded = normalized.padEnd(Math.ceil(normalized.length / 4) * 4, '=')
-    const claims = JSON.parse(atob(padded)) as { exp?: unknown }
-    return typeof claims.exp === 'number' ? claims.exp * 1000 : null
-  } catch {
-    return null
-  }
-}
-
-// isTokenExpired 判断 token 是否已过期或即将过期。
-function isTokenExpired(token: string): boolean {
-  const expiry = decodeJwtExpiryMs(token)
-  if (expiry === null) return false
-  return Date.now() >= expiry - TOKEN_EXPIRY_SKEW_MS
-}
-
 // useWebSocket opens a single connection to /api/ws and dispatches every
 // message to the supplied handler. Auto-reconnects with back-off while the
 // auth token is present; after the fast retries are exhausted it keeps a
@@ -68,7 +43,7 @@ export function useWebSocket(onEvent: (topic: string, payload: unknown) => void)
 
     const open = () => {
       if (closed) return
-      if (isTokenExpired(token)) {
+      if (isAccessTokenStale(token)) {
         refreshExpiredToken()
         return
       }
@@ -92,7 +67,7 @@ export function useWebSocket(onEvent: (topic: string, payload: unknown) => void)
       ws.onclose = () => {
         if (closed) return
         // token 过期时不要继续退避重试，交给刷新流程处理。
-        if (isTokenExpired(token)) {
+        if (isAccessTokenStale(token)) {
           refreshExpiredToken()
           return
         }

web/src/pages/LibraryPage.tsx

@@ -32,12 +32,14 @@ import { useLibrarySeriesSelection } from './useLibrarySeriesSelection'
 import { useLibraryAdminActions } from './useLibraryAdminActions'
 import { usePermission } from '../hooks/usePermission'
 import { useFavourites } from '../hooks/useFavourites'
+import { mediaSearchReturnPath } from '../utils/mediaNavigation'
 
 export function LibraryPage() {
   const { id = '' } = useParams()
   const [searchParams, setSearchParams] = useSearchParams()
   const location = useLocation()
   const navigate = useNavigate()
+  const searchFrom = mediaSearchReturnPath(location.state)
   const role = useAuthStore((s) => s.user?.role)
   const canFavorite = usePermission('can_favorite')
   const { isFavourite, toggleFavourite } = useFavourites()
@@ -358,7 +360,7 @@ export function LibraryPage() {
           void handleToggleFavourite(selectedSeries.rep.id)
         }}
         seriesToolBusy={seriesToolBusy}
-        onBack={clearSelectedSeries}
+        onBack={() => searchFrom ? navigate(searchFrom) : clearSelectedSeries()}
         onSmartScrape={handleSeriesSmartScrape}
         onManualScrape={() => setManualSeriesScrapeOpen(true)}
         onMetadataEdit={() => setSeriesMetadataEditOpen(true)}

web/src/pages/MediaSearchPage.tsx

@@ -0,0 +1,182 @@
+import { useEffect, useState } from 'react'
+import { useSearchParams } from 'react-router-dom'
+import { ChevronLeft, ChevronRight, LoaderCircle, RefreshCw, Search } from 'lucide-react'
+
+import { mediaAPI, type MediaSearchPage as SearchPageData } from '../api/library'
+import { MediaCard } from '../components/MediaCard'
+import { PageBackButton } from '../components/PageBackButton'
+import { normalizeListPosition } from '../hooks/useListPositionMemory'
+import { useAuthStore } from '../stores/auth'
+import { usePlayProfileStore } from '../stores/playProfile'
+import { favouriteMediaLink, mediaSearchLink } from '../utils/mediaNavigation'
+
+const PAGE_SIZE = 48
+const CACHE_TTL_MS = 5 * 60_000
+// 返回详情前的结果可立即显示；按账号、播放档案、关键词和页码隔离。
+const pageCache = new Map<string, { at: number; data: SearchPageData }>()
+
+function readCachedPage(key: string): SearchPageData | null {
+  const cached = pageCache.get(key)
+  if (!cached) return null
+  if (Date.now() - cached.at >= CACHE_TTL_MS) {
+    pageCache.delete(key)
+    return null
+  }
+  return cached.data
+}
+
+export function MediaSearchPage() {
+  const [params, setParams] = useSearchParams()
+  const query = (params.get('q') ?? '').trim()
+  const page = normalizeListPosition(params.get('page'))
+  const userId = useAuthStore((s) => s.user?.id)
+  const profileId = usePlayProfileStore((s) => s.activeProfileId)
+  const cacheKey = JSON.stringify([userId, profileId, query, page])
+
+  const changePage = (next: number) => {
+    const nextParams = new URLSearchParams({ q: query })
+    if (next > 1) nextParams.set('page', String(next))
+    setParams(nextParams)
+  }
+
+  return (
+    <MediaSearchResults
+      key={cacheKey}
+      cacheKey={cacheKey}
+      query={query}
+      page={page}
+      onPageChange={changePage}
+    />
+  )
+}
+
+function MediaSearchResults({ cacheKey, query, page, onPageChange }: {
+  cacheKey: string
+  query: string
+  page: number
+  onPageChange: (page: number) => void
+}) {
+  const [data, setData] = useState(() => readCachedPage(cacheKey))
+  const [loading, setLoading] = useState(() => !!query && !readCachedPage(cacheKey))
+  const [error, setError] = useState('')
+  const [revision, setRevision] = useState(0)
+
+  useEffect(() => {
+    if (!query) return
+    const cached = readCachedPage(cacheKey)
+    if (cached) {
+      setData(cached)
+      setLoading(false)
+      return
+    }
+
+    let cancelled = false
+    setLoading(true)
+    setError('')
+    mediaAPI.searchPage(query, page, PAGE_SIZE, { groupSeries: true })
+      .then((result) => {
+        if (cancelled) return
+        // 限制缓存大小，避免长时间搜索大量关键词后占用过多内存。
+        if (pageCache.size >= 20) pageCache.delete(pageCache.keys().next().value!)
+        pageCache.set(cacheKey, { at: Date.now(), data: result })
+        setData(result)
+      })
+      .catch(() => {
+        if (!cancelled) setError('搜索失败，请稍后重试')
+      })
+      .finally(() => {
+        if (!cancelled) setLoading(false)
+      })
+
+    return () => { cancelled = true }
+  }, [cacheKey, query, page, revision])
+
+  const refresh = () => {
+    pageCache.delete(cacheKey)
+    setRevision((value) => value + 1)
+  }
+
+  const items = data?.items ?? []
+  const total = data?.total ?? items.length
+  const totalPages = Math.max(1, Math.ceil(total / PAGE_SIZE))
+  const from = mediaSearchLink(query, page)
+
+  return (
+    <div className="space-y-6">
+      <PageBackButton to="/" label="返回首页" compact />
+      <div className="flex items-start justify-between gap-4">
+        <div className="min-w-0 space-y-2">
+          <h1 className="font-display text-2xl font-bold text-[var(--app-text)] sm:text-3xl">搜索结果</h1>
+          <p className="break-words text-sm text-[var(--app-muted)]" aria-live="polite">
+            {query
+              ? <>“{query}”{data && !loading ? ` · 共 ${total} 条结果` : ''}</>
+              : '在顶部搜索框输入媒体名称，按回车开始搜索'}
+          </p>
+        </div>
+        {query && (
+          <button
+            type="button"
+            onClick={refresh}
+            disabled={loading}
+            className="inline-flex shrink-0 items-center gap-2 rounded-xl border border-[var(--app-border)] bg-[var(--app-panel)] px-3 py-2 text-xs text-[var(--app-subtle)] hover:bg-[var(--app-hover)] disabled:opacity-50"
+          >
+            <RefreshCw size={14} />
+            刷新结果
+          </button>
+        )}
+      </div>
+
+      {loading && (
+        <div role="status" className="flex items-center justify-center gap-2 py-20 text-sm text-[var(--app-muted)]">
+          <LoaderCircle size={18} className="animate-spin text-brand-500" />
+          正在搜索…
+        </div>
+      )}
+
+      {error && (
+        <div role="alert" className="rounded-2xl border border-red-400/30 bg-[var(--app-panel)] p-5 text-sm text-red-400">
+          {error}
+          <button type="button" onClick={refresh} className="ml-3 underline">重试</button>
+        </div>
+      )}
+
+      {!loading && !error && query && items.length === 0 && (
+        <div className="flex flex-col items-center gap-3 rounded-2xl border border-[var(--app-border)] bg-[var(--app-panel)] px-6 py-16 text-center text-[var(--app-muted)]">
+          <Search size={28} />
+          <p className="break-all text-sm">未搜索到与 “{query}” 相关的媒体内容</p>
+          <p className="text-xs">试试其他名称或更简短的关键词</p>
+        </div>
+      )}
+
+      {!loading && 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((item) => (
+            <MediaCard key={item.id} media={item} showLibrary linkTo={favouriteMediaLink(item)} linkState={{ from }} />
+          ))}
+        </div>
+      )}
+
+      {!loading && !error && query && (totalPages > 1 || page > 1) && (
+        <nav aria-label="搜索结果分页" className="flex items-center justify-center gap-4 py-4 text-sm">
+          <button
+            type="button"
+            disabled={page <= 1}
+            onClick={() => onPageChange(page - 1)}
+            className="inline-flex items-center gap-1 rounded-xl border border-[var(--app-border)] bg-[var(--app-panel)] px-3 py-2 hover:bg-[var(--app-hover)] disabled:cursor-not-allowed disabled:opacity-40"
+          >
+            <ChevronLeft size={16} />上一页
+          </button>
+          <span className="text-[var(--app-muted)]">第 {page} / {totalPages} 页</span>
+          <button
+            type="button"
+            disabled={page >= totalPages}
+            onClick={() => onPageChange(page + 1)}
+            className="inline-flex items-center gap-1 rounded-xl border border-[var(--app-border)] bg-[var(--app-panel)] px-3 py-2 hover:bg-[var(--app-hover)] disabled:cursor-not-allowed disabled:opacity-40"
+          >
+            下一页<ChevronRight size={16} />
+          </button>
+        </nav>
+      )}
+    </div>
+  )
+}

web/src/pages/PlayerPage.tsx

@@ -20,7 +20,7 @@ import { useAuthStore } from '../stores/auth'
 import type { Media, PlaybackInfo, PlaybackQuality, PlaybackSegment, PlaybackSegmentKind } from '../types'
 import { getSeriesKey, seriesTitleFromPath } from '../utils/groupSeries'
 import { formatEpisodeLabel } from '../utils/episodeNumber'
-import { mediaVersionMatches, mediaVersionsOf } from '../utils/mediaVersion'
+import { autoPlayEpisodeFor, mediaVersionMatches, mediaVersionsOf } from '../utils/mediaVersion'
 import { DEFAULT_PLAYBACK_RATE, normalizePlaybackRate } from '../utils/playbackRate'
 import { resolveActiveSkip, skippedNoticeText, toSkipSegments, type SkipPrompt } from '../utils/skipSegments'
 import { isRemoteEmbyID } from '../utils/remoteEmby'
@@ -37,8 +37,10 @@ import {
   type SubtitleStylePreset,
 } from '../utils/subtitleDisplay'
 import { pickPlayerMode, needsTranscodeForBrowser, isDirectStreamMedia, isStrmMedia, type PlayerMode } from './playerPageModel'
-import { classifyDirectPlayError, DIRECT_SEEK_GRACE_MS } from './directPlayError'
+import { classifyDirectPlayError, directPlayFailureCopy, directStallTimeoutMs, DIRECT_SEEK_GRACE_MS } from './directPlayError'
+import type { DirectPlayFailureReason } from './directPlayError'
 import { canSeekDirectNow, DIRECT_SEEK_WAIT_MS } from './directPlaySeek'
+import { applyPendingHlsSeek, attachNativeHls, bindPendingHlsSeek, hlsManifestTimeoutMs, playbackPositionSec } from './hlsPlayback'
 import { apiErrorMessage } from './StrmManagePage'
 import { PlayerTopBar } from './PlayerTopBar'
 import { PlayerVideoStage } from './PlayerVideoStage'
@@ -200,6 +202,13 @@ export function PlayerPage() {
   // 必须在切换前先记下来，等直连 loadedmetadata 后再跳。
   const pendingDirectStartRef = useRef(0)
   const hlsSourceRef = useRef<'cloud' | 'local'>('local')
+  // 起播看门狗要等到 20 秒后才判定，判定时用 ref 读最新的「直连解码」开关，避免
+  // 把回调塞进播放 effect 的依赖里反复重建。
+  const directOnlyRef = useRef(false)
+  // 当前挂在画面上的「播放失败」横幅文案。这类提示描述的是「刚才那一下没起来」，
+  // 画面真正播起来之后必须撤掉：上游冷启动慢 10 秒就会先误报一次，而播放随后
+  // 是正常的，留在画面上就成了「编码不兼容」的假结论。
+  const transientPlayErrorRef = useRef('')
 
   // 弹幕控制：状态来自 /api/danmaku/config 初始值，用户在面板里实时调整。
   const [danmakuOpen, setDanmakuOpen] = useState(false)
@@ -519,6 +528,7 @@ export function PlayerPage() {
     directRetryRef.current = false
     retryingDirectRef.current = false
     directSeekGraceUntilRef.current = 0
+    transientPlayErrorRef.current = ''
     pendingDirectSeekRef.current = null
     directSeekWaitCleanupRef.current?.()
     directSeekWaitCleanupRef.current = null
@@ -553,6 +563,24 @@ export function PlayerPage() {
     }
   }, [])
 
+  /**
+   * 挂一条「播放起不来」的横幅。走这个入口的提示都会记下来，一旦画面真的开始
+   * 播放（playing）就被撤掉；切媒体、切线路时也会清空。信息类提示（例如「可能
+   * 只有画面没有声音」）不走这里，它们播起来之后依然成立。
+   */
+  const showPlayFailure = useCallback((message: string) => {
+    transientPlayErrorRef.current = message
+    setPlayerError(message)
+  }, [])
+
+  const clearPlayFailure = useCallback(() => {
+    const stale = transientPlayErrorRef.current
+    if (!stale) return
+    transientPlayErrorRef.current = ''
+    // 只撤掉自己挂的那条：期间可能已经换了别的提示。
+    setPlayerError((prev) => (prev === stale ? '' : prev))
+  }, [])
+
   const clearDirectSeekWait = useCallback(() => {
     directSeekWaitCleanupRef.current?.()
     directSeekWaitCleanupRef.current = null
@@ -742,6 +770,7 @@ export function PlayerPage() {
         if (cancelled) return
         setMedia(m)
         setPlayerError('')
+        transientPlayErrorRef.current = ''
         setLoadError('')
         const isDirect = isDirectStreamMedia(m)
         const auto = pickPlayerMode(m)
@@ -875,6 +904,13 @@ export function PlayerPage() {
   // 加载统一播放能力：115 云端清晰度 + 本地 HLS 清晰度。
   useEffect(() => {
     if (!mediaId) return
+    // 远程 Emby 条目只存在于远程库里，本地 media 表没有它的伪 ID，问这个接口只会
+    // 拿到 404（控制台里就是一条红字）；而它的播放方式恒为「远程直连、不转码」，
+    // 也没有任何可切换的档位。播放地址由 /api/stream/:id 负责解析。
+    if (isRemoteEmbyID(mediaId)) {
+      setPlaybackInfo(null)
+      return
+    }
     let cancelled = false
     mediaAPI
       .playbackInfo(mediaId)
@@ -957,14 +993,15 @@ export function PlayerPage() {
   }, [])
 
   const switchToLocalHLS = useCallback(
-    (position = 0) => {
+    (position?: number) => {
+      pendingSeekRef.current = null
       setHlsSource('local')
       setCloudWaiting(false)
       setSelectedQuality((prev) => {
         if (playbackInfo?.local_qualities?.some((quality) => quality.id === prev)) return prev
         return defaultLocalQualityId(playbackInfo)
       })
-      if (position > 2) setHlsStartSec(position)
+      if (position !== undefined) setHlsStartSec(Math.max(0, position))
       setPlaybackMode('hls')
     },
     [playbackInfo, setPlaybackMode],
@@ -1041,10 +1078,10 @@ export function PlayerPage() {
           : undefined
       if (preferred?.source === 'cloud') {
         setHlsSource('cloud')
+        // 云 HLS 是整段时间轴，等待云转码期间也要保留目标位置。
+        pendingSeekRef.current = position > 2 ? position : null
         if (preferred.available) {
           setCloudWaiting(false)
-          // 云 HLS 是整段时间轴：切过去后按当前位置跳一下，别让用户从头看。
-          pendingSeekRef.current = position > 2 ? position : null
         } else {
           setCloudWaiting(true)
           setCloudWaitMessage(preferred.note || `正在等待 115 转码 ${preferred.label}…`)
@@ -1109,8 +1146,9 @@ export function PlayerPage() {
     if (!playbackInfo || !initialHlsRequestRef.current) return
     if (directOnly || !mediaId) return
     initialHlsRequestRef.current = false
-    enterHlsPlayback(0, playbackInfo)
-  }, [directOnly, enterHlsPlayback, mediaId, playbackInfo])
+    const position = Math.max(resumePosition, playbackPositionSec(mode, hlsSource, hlsStartSec, ref.current?.currentTime || 0))
+    enterHlsPlayback(position, playbackInfo)
+  }, [directOnly, enterHlsPlayback, hlsSource, hlsStartSec, mediaId, mode, playbackInfo, resumePosition])
 
   // 直连播放只发现外挂字幕；只有 HLS 模式需要探测可烧录的内嵌字幕。
   useEffect(() => {
@@ -1145,13 +1183,122 @@ export function PlayerPage() {
   const mediaRef = useRef(media)
   useEffect(() => {
     mediaRef.current = media
-  }, [media])
+    directOnlyRef.current = directOnly
+  }, [directOnly, media])
+
+  /**
+   * 直连失败后的统一回退：按当前源与设置决定切 115 云转码 / 本地转码 / 明确报错。
+   * error 分支与起播看门狗（armDirectStallWatchdog）共用同一套判定。
+   *
+   * reason：'stall' = 看门狗等到期还没数据；'error' = 媒体元素报了 error（重试过
+   * 一次仍然失败）。两者文案必须分开——超时的真实原因多半是上游慢，不是编码不兼容。
+   */
+  const fallbackDirectPlay = useCallback((reason: DirectPlayFailureReason = 'error') => {
+    if (modeRef.current !== 'direct') return
+    // 宽限期内又触发的迟到 fallback 定时器：直接丢掉。
+    if (Date.now() < directSeekGraceUntilRef.current) return
+    const current = ref.current
+    if (current && current.readyState >= HTMLMediaElement.HAVE_CURRENT_DATA) return
+    if (isRemoteEmbyID(mediaRef.current?.id) || isDirectStreamMedia(mediaRef.current)) {
+      const copy = directPlayFailureCopy('remote-emby', reason)
+      showPlayFailure(copy.banner)
+      toast.error(copy.toast)
+    } else if (directOnly) {
+      const copy = directPlayFailureCopy('direct-only', reason)
+      showPlayFailure(copy.banner)
+      toast.error(copy.toast)
+    } else if (playbackInfo?.provider === 'cloud115') {
+      const preferred =
+        findPlaybackQualityById(playbackInfo, selectedQuality) ??
+        findPlaybackQualityById(playbackInfo, playbackInfo.default_quality)
+      if (preferred?.source === 'cloud' && preferred.available) {
+        setCloudWaiting(false)
+        setHlsSource('cloud')
+        setPlaybackMode('hls')
+      } else if (preferred?.source === 'cloud') {
+        setHlsSource('cloud')
+        setCloudWaiting(true)
+        setCloudWaitMessage(preferred.note || `正在等待 115 转码 ${preferred.label}…`)
+        void startCloudTranscode(Number(preferred.id) || 4)
+      } else {
+        switchToLocalHLS(current?.currentTime || 0)
+      }
+    } else if (hlsUnavailable) {
+      showPlayFailure('直接播放失败，且 HLS 转码不可用。请检查文件是否存在，或配置本机 ffmpeg 后使用 HLS 转码播放。')
+      toast.error('直接播放失败，HLS 转码不可用')
+    } else {
+      toast.error('直接播放失败，切换到 HLS 转码')
+      switchToLocalHLS(current?.currentTime || 0)
+    }
+  }, [
+    directOnly,
+    hlsUnavailable,
+    playbackInfo,
+    selectedQuality,
+    setPlaybackMode,
+    showPlayFailure,
+    startCloudTranscode,
+    switchToLocalHLS,
+  ])
+
+  /**
+   * 安排一次直连回退判定。STRM/115 直链冷启动比本地文件慢，给足缓冲时间再判失败。
+   * 只由 handleVideoError（媒体元素真的报了 error）触发，所以原因按 'error' 记。
+   */
+  const scheduleDirectFallback = useCallback(() => {
+    clearFallbackTimer()
+    const fallbackDelayMs = isStrmMedia(mediaRef.current) ? 4000 : 1500
+    fallbackTimerRef.current = setTimeout(() => fallbackDirectPlay('error'), fallbackDelayMs)
+  }, [clearFallbackTimer, fallbackDirectPlay])
+
+  // 看门狗定时器可能比一次渲染活得更久：始终调用最新一版回退逻辑，避免用
+  // 「起播那一刻」捕获的 playbackInfo（那时可能还是 null）去判定，把本该走
+  // 115 云转码的条目误判成本地转码。
+  const fallbackDirectPlayRef = useRef<(reason?: DirectPlayFailureReason) => void>(() => {})
+  useEffect(() => {
+    fallbackDirectPlayRef.current = fallbackDirectPlay
+  }, [fallbackDirectPlay])
+
+  /**
+   * 直连起播看门狗。
+   *
+   * 直连没有 hls.js 那种 manifestLoadingTimeOut 兜底：请求挂住、或上游回了不带
+   * error 事件的响应时，<video> 会一直停在 readyState=0——不出画面、不报错、
+   * 也不触发 handleVideoError，播放器就静默卡死（表现为「进去不自动播放」，
+   * 只能退出重进或手动切播放方式）。这里给每次直连起播挂一个兜底定时器：
+   * 到点仍是 readyState=0（连元数据都没解析到）就按失败处理；只要已经拿到
+   * metadata 就说明链路是通的，只是慢，交给 playing 事件收尾。
+   *
+   * 期限本身要按「失败之后还有没有补救动作」来定：能切 HLS 的源早点换线路对
+   * 用户有利；远程 Emby / 直连解码没有可退的线路，等不到数据时唯一的产出是
+   * 一条提示，所以必须容忍跨网上游的冷启动（见 directStallTimeoutMs）。
+   */
+  const armDirectStallWatchdog = useCallback(() => {
+    clearFallbackTimer()
+    const currentMedia = mediaRef.current
+    const canFallback = !isDirectStreamMedia(currentMedia) && !directOnlyRef.current
+    const timeoutMs = directStallTimeoutMs({
+      isStrm: isStrmMedia(currentMedia),
+      canFallback,
+    })
+    fallbackTimerRef.current = setTimeout(() => {
+      fallbackTimerRef.current = null
+      if (modeRef.current !== 'direct') return
+      if (Date.now() < directSeekGraceUntilRef.current) return
+      const current = ref.current
+      if (current && current.readyState >= HTMLMediaElement.HAVE_METADATA) return
+      // 这里只可能是「一直没数据」，按超时文案报，不要说成编码不兼容。
+      fallbackDirectPlayRef.current('stall')
+    }, timeoutMs)
+  }, [clearFallbackTimer])
 
   useEffect(() => {
     if (!mediaId || !ref.current) return
     const currentMedia = mediaRef.current
     if (!currentMedia) return
     let cancelled = false
+    let cleanupHlsResume: (() => void) | undefined
+    let onDirectMetadata: (() => void) | undefined
     teardownHls()
 
     const video = ref.current
@@ -1168,13 +1315,14 @@ export function PlayerPage() {
       void import('hls.js').then(({ default: HlsCtor }) => {
         if (cancelled || !ref.current) return
         if (HlsCtor.isSupported()) {
+          cleanupHlsResume = bindPendingHlsSeek(video, pendingSeekRef)
           const hls = new HlsCtor({
             enableWorker: true,
             lowLatencyMode: false,
             // The server waits up to 45s for the first segment. Slow two-core
             // hosts and remote STRM sources regularly need more than hls.js's
             // 10s default, which otherwise aborts a healthy transcode.
-            manifestLoadingTimeOut: 60_000,
+            manifestLoadingTimeOut: hlsManifestTimeoutMs(hlsSource, hlsStartSec),
             // 115 云 HLS 的 master.m3u8 每次都要现调 115 接口（实测 0.5–1.3s），
             // 还可能撞上 409「正在转码」。只重试 1 次的话，一次抖动就判死刑，
             // 直接掉到本地转码。
@@ -1188,15 +1336,7 @@ export function PlayerPage() {
           hls.loadSource(url)
           hls.attachMedia(video)
           hls.on(HlsCtor.Events.MANIFEST_PARSED, () => {
-            if (pendingSeekRef.current !== null) {
-              const target = pendingSeekRef.current
-              pendingSeekRef.current = null
-              try {
-                video.currentTime = target
-              } catch {
-                // ignore
-              }
-            }
+            applyPendingHlsSeek(video, pendingSeekRef)
             void video.play().catch(() => undefined)
             // .strm 入库时常缺 duration；转码启动时会补探测，这里刷新一次给进度条总时长。
             if (durationSec > 0) return
@@ -1224,10 +1364,12 @@ export function PlayerPage() {
               }
               setHlsUnavailable(true)
               if (playbackProvider === 'cloud115') {
-                setPlayerError('115 云端和本地转码均不可用，请检查账号授权或稍后重试。')
+                showPlayFailure('115 云端和本地转码均不可用，请检查账号授权或稍后重试。')
                 toast.error('115 云端和本地转码均不可用')
                 return
               }
+              // 这条同时是「为什么可能没声音」的说明，播起来之后依然成立，
+              // 所以不走 showPlayFailure（不随 playing 撤掉）。
               setPlayerError('HLS 转码不可用，正在尝试直接播放原始文件。若出现有画面无声音，通常是 MKV/AC3/EAC3 音轨需要配置本机 ffmpeg 转码为 AAC。')
               toast.error('HLS 转码失败，尝试切换到直接播放')
               setPlaybackMode('direct')
@@ -1244,18 +1386,17 @@ export function PlayerPage() {
           } catch {
             // ignore
           }
-          video.src = url
-          void video.play().catch(() => undefined)
+          cleanupHlsResume = attachNativeHls(video, url, pendingSeekRef)
         } else {
           setHlsUnavailable(true)
-          setPlayerError('当前浏览器不支持 HLS，正在尝试直接播放。')
+          showPlayFailure('当前浏览器不支持 HLS，正在尝试直接播放。')
           toast.error('当前浏览器不支持 HLS，降级到直接播放')
           setPlaybackMode('direct')
         }
       }).catch(() => {
         if (cancelled) return
         setHlsUnavailable(true)
-        setPlayerError('HLS 播放组件加载失败，正在尝试直接播放。')
+        showPlayFailure('HLS 播放组件加载失败，正在尝试直接播放。')
         setPlaybackMode('direct')
       })
     } else {
@@ -1277,41 +1418,48 @@ export function PlayerPage() {
         )
         pendingDirectStartRef.current = 0
         if (resumeAt > 0) {
-          video.addEventListener(
-            'loadedmetadata',
-            () => {
-              // STRM/115 直链要等 seekable 覆盖目标位置再写 currentTime（否则会被
-              // 钳回 0），并顺带打开误报宽限期，避免这次跳转被当成播放失败又弹回 HLS。
-              if (isStrmMedia(currentMedia)) {
-                seekStrmDirectTo(resumeAt)
-                return
-              }
-              try {
-                video.currentTime = resumeAt
-              } catch {
-                // ignore
-              }
-            },
-            { once: true },
-          )
+          onDirectMetadata = () => {
+            // STRM/115 直链要等 seekable 覆盖目标位置再写 currentTime（否则会被
+            // 钳回 0），并顺带打开误报宽限期，避免这次跳转被当成播放失败又弹回 HLS。
+            if (isStrmMedia(currentMedia)) {
+              seekStrmDirectTo(resumeAt)
+              return
+            }
+            try {
+              video.currentTime = resumeAt
+            } catch {
+              // ignore
+            }
+          }
+          video.addEventListener('loadedmetadata', onDirectMetadata, { once: true })
         }
         video.src = url
         void video.play().catch(() => undefined)
+        // 起播兜底：上游不回 error 时会一直停在 readyState=0，见 armDirectStallWatchdog。
+        armDirectStallWatchdog()
       }
       if (hlsUnavailable && needsTranscodeForBrowser(currentMedia)) {
         setPlayerError('当前正在直连播放原始文件；此封装或音轨浏览器兼容性有限，可能只有画面没有声音。请配置本机 ffmpeg 后切回 HLS 转码播放。')
       }
     }
-    const onPlaying = () => clearFallbackTimer()
+    const onPlaying = () => {
+      clearFallbackTimer()
+      // 画面真的动了：之前挂的「起不来」提示一律作废。
+      clearPlayFailure()
+    }
     video.addEventListener('playing', onPlaying)
     return () => {
       cancelled = true
       video.removeEventListener('playing', onPlaying)
+      if (onDirectMetadata) video.removeEventListener('loadedmetadata', onDirectMetadata)
+      cleanupHlsResume?.()
       teardownHls()
     }
   }, [
     activeBurnedSubtitleStream,
+    armDirectStallWatchdog,
     clearFallbackTimer,
+    clearPlayFailure,
     cloudReloadNonce,
     cloudWaiting,
     hlsSource,
@@ -1324,6 +1472,7 @@ export function PlayerPage() {
     seekStrmDirectTo,
     selectedQuality,
     setPlaybackMode,
+    showPlayFailure,
     switchToLocalHLS,
     teardownHls,
     vr360DirectProxy,
@@ -1352,19 +1501,22 @@ export function PlayerPage() {
     if (!id) return
     setResumePosition(0)
     setInitialSeekDone(false)
+    const controller = new AbortController()
     playbackAPI
-      .getResume(id)
+      .getResume(id, controller.signal)
       .then((progress) => {
+        if (controller.signal.aborted) return
         if (progress.position_ms > 2000 && !progress.completed) {
           setResumePosition(progress.position_ms / 1000)
         }
       })
       .catch(() => undefined)
+    return () => controller.abort()
   }, [id])
 
   useEffect(() => {
     if (!resumePosition || initialSeekDone) return
-    if (mode === 'hls') {
+    if (mode === 'hls' && hlsSource === 'local') {
       // Restart transcode near the resume point instead of seeking a short partial playlist.
       if (Math.abs(hlsStartSec - resumePosition) > 2) {
         setHlsStartSec(resumePosition)
@@ -1383,6 +1535,7 @@ export function PlayerPage() {
     let seekStarted = false
     let onPlaying: (() => void) | undefined
     let onCanPlay: (() => void) | undefined
+    let onSeeked: (() => void) | undefined
 
     const notifyResume = () => {
       const m = Math.floor(resumePosition / 60)
@@ -1399,7 +1552,7 @@ export function PlayerPage() {
         return
       }
       seekStarted = true
-      if (isStrmMedia(mediaRef.current)) {
+      if (mode === 'direct' && isStrmMedia(mediaRef.current)) {
         seekStrmDirectTo(resumePosition, {
           timeoutMs: 6000,
           onSettled: () => {
@@ -1426,7 +1579,8 @@ export function PlayerPage() {
       directSeekGraceUntilRef.current = Date.now() + DIRECT_SEEK_GRACE_MS
       directRetryRef.current = false
       clearFallbackTimer()
-      const onSeeked = () => {
+      onSeeked = () => {
+        if (cancelled) return
         setInitialSeekDone(true)
         notifyResume()
       }
@@ -1439,7 +1593,7 @@ export function PlayerPage() {
     }
 
     // STRM/云盘：等真正起播后再续播跳转，避免 302 未完成就 seek。
-    const deferForStrm = isStrmMedia(mediaRef.current)
+    const deferForStrm = mode === 'direct' && isStrmMedia(mediaRef.current)
     if (deferForStrm) {
       onPlaying = () => {
         performResumeSeek()
@@ -1462,8 +1616,9 @@ export function PlayerPage() {
       cancelled = true
       if (onPlaying) video.removeEventListener('playing', onPlaying)
       if (onCanPlay) video.removeEventListener('canplay', onCanPlay)
+      if (onSeeked) video.removeEventListener('seeked', onSeeked)
     }
-  }, [resumePosition, initialSeekDone, mode, hlsStartSec, clearFallbackTimer, seekStrmDirectTo])
+  }, [resumePosition, initialSeekDone, mode, hlsSource, hlsStartSec, clearFallbackTimer, seekStrmDirectTo])
 
   // 使用 ref 实时同步进度计算所需的状态，避免每次 hlsStartSec 改变都触发 cleanup 并误上报旧进度
   const hlsStartSecRef = useRef(hlsStartSec)
@@ -1589,6 +1744,17 @@ export function PlayerPage() {
     return playlistEpisodes.findIndex((e) => mediaVersionMatches(e, media.id))
   }, [media, playlistEpisodes])
 
+  // 用剧集 ID 打开播放器时（首页轮播、剧集库卡片、剧集详情页的播放按钮），先落到第一集：
+  // 这类条目没有可播文件，直接请求 /api/stream/<剧集ID> 会被远程 Emby 以 500 打回。
+  useEffect(() => {
+    const target = autoPlayEpisodeFor(media, id, playlistEpisodes)
+    if (!target) return
+    navigate(
+      { pathname: `/play/${target.id}`, search: location.search },
+      { replace: true, state: location.state },
+    )
+  }, [media, id, playlistEpisodes, navigate, location.search, location.state])
+
   const prevEpisode = useMemo(() => {
     if (currentEpisodeIndex > 0) {
       return playlistEpisodes[currentEpisodeIndex - 1]
@@ -2059,28 +2225,22 @@ export function PlayerPage() {
       toast.error('图片字幕需要开启 HLS 转码后才能显示')
       return
     }
+    const position = playbackPositionSec(mode, hlsSource, hlsStartSec, ref.current?.currentTime || 0)
     if (nextTrack?.delivery === 'burn') {
       // PGS 等图片字幕必须走本地 FFmpeg 烧录，不能交给云端 HLS。
-      switchToLocalHLS(ref.current?.currentTime || 0)
-    }
-    const burnChanged =
-      oldTrack?.delivery === 'burn' || nextTrack?.delivery === 'burn'
-    if (burnChanged && mode === 'hls' && ref.current) {
-      setHlsStartSec(hlsStartSec + (ref.current.currentTime || 0))
+      switchToLocalHLS(position)
+    } else if (oldTrack?.delivery === 'burn' && mode === 'hls') {
+      setHlsStartSec(position)
     }
     setSubtitleIndex(index)
-    if (nextTrack?.delivery === 'burn' && mode !== 'hls' && !directOnly && !isDirectStream) {
-      setHlsStartSec(ref.current?.currentTime || 0)
-      setPlaybackMode('hls')
-    }
-  }, [directOnly, hlsStartSec, isDirectStream, mode, setPlaybackMode, subs, subtitleIndex, switchToLocalHLS])
+  }, [directOnly, hlsSource, hlsStartSec, isDirectStream, mode, subs, subtitleIndex, switchToLocalHLS])
 
   const selectPlaybackQuality = useCallback(
     (quality: PlaybackQuality) => {
       const video = ref.current
       const localTime = video?.currentTime || 0
       // 云 HLS 的 currentTime 本来就在绝对时间轴上；本地 HLS 的 t=0 对应 hlsStartSec。
-      const position = mode === 'hls' && hlsSource === 'local' ? hlsStartSec + localTime : localTime
+      const position = playbackPositionSec(mode, hlsSource, hlsStartSec, localTime)
       if (quality.source === 'original') {
         setCloudWaiting(false)
         setHlsSource('local')
@@ -2092,12 +2252,13 @@ export function PlayerPage() {
       if (quality.source === 'cloud') {
         setSelectedQuality(quality.id)
         setHlsSource('cloud')
+        setHlsStartSec(position)
+        pendingSeekRef.current = position > 2 ? position : null
         // 用户手动选了云端档位：这是一次明确的重新尝试，重置自动重连的次数窗口。
         cloudReloadsRef.current = []
         if (quality.available) {
           setCloudWaiting(false)
           setCloudWaitMessage('')
-          if (position > 2) pendingSeekRef.current = position
           setPlaybackMode('hls')
         } else {
           setCloudWaiting(true)
@@ -2117,7 +2278,8 @@ export function PlayerPage() {
         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)
+      pendingSeekRef.current = null
+      setHlsStartSec(position)
       setPlaybackMode('hls')
     },
     [hlsSource, hlsStartSec, mode, setPlaybackMode, startCloudTranscode],
@@ -2150,7 +2312,7 @@ export function PlayerPage() {
   const handleVideoError = useCallback(() => {
     const video = ref.current
     if (mode !== 'direct') {
-      setPlayerError('视频播放失败，请检查文件是否存在，或确认 ffmpeg 已正确配置。')
+      showPlayFailure('视频播放失败，请检查文件是否存在，或确认 ffmpeg 已正确配置。')
       toast.error('视频播放失败，请检查文件是否存在')
       return
     }
@@ -2174,16 +2336,21 @@ export function PlayerPage() {
       directRetryRef.current = true
       retryingDirectRef.current = true
       try {
-        // STRM/115 直连：load() 会清空进度从头播。软恢复即可。
-        if (isStrmMedia(mediaRef.current) || pendingDirectSeekRef.current !== null) {
-          const pending = pendingDirectSeekRef.current
-          void video.play().catch(() => undefined)
-          if (pending !== null && pending > 1) {
-            seekStrmDirectTo(pending)
-          }
-        } else {
-          video.load()
-          void video.play().catch(() => undefined)
+        // 必须真正重新发起请求。媒体元素加载失败后 networkState 已经是
+        // NETWORK_NO_SOURCE（不是 NETWORK_EMPTY），此时只调 play() 不会重跑资源
+        // 选取算法：既不会重新请求，也不会再触发一次 error，于是没人去挂 fallback
+        // 定时器——上游 403/502 一次就把播放器钉死在「direct + 空 src」上，没有
+        // 错误提示也不会自动播放，只能退出重进或手动切播放方式。
+        //
+        // classifyDirectPlayError 只在 readyState < HAVE_CURRENT_DATA 时返回
+        // 'retry'，本来就没有可保住的播放进度，所以 load() 是安全的；需要续播
+        // 的位置由 seekStrmDirectTo 在 metadata 就绪后补回。
+        const resumeAt = pendingDirectSeekRef.current
+        pendingDirectSeekRef.current = null
+        video.load()
+        void video.play().catch(() => undefined)
+        if (resumeAt !== null && resumeAt > 1) {
+          seekStrmDirectTo(resumeAt)
         }
       } finally {
         retryingDirectRef.current = false
@@ -2191,60 +2358,8 @@ export function PlayerPage() {
       return
     }
 
-    const fallbackDirectPlay = () => {
-      if (modeRef.current !== 'direct') return
-      // 宽限期内又触发的迟到 fallback 定时器：直接丢掉。
-      if (Date.now() < directSeekGraceUntilRef.current) return
-      const current = ref.current
-      if (current && current.readyState >= HTMLMediaElement.HAVE_CURRENT_DATA) return
-      if (isRemoteEmbyID(mediaRef.current?.id) || isDirectStreamMedia(mediaRef.current)) {
-        setPlayerError('直接播放失败。该媒体为远程 Emby 挂载直连播放（不进行转码）；当前浏览器可能不支持该视频编码或音频格式，建议使用外部播放器（如 PotPlayer / VLC / IINA）播放。')
-        toast.error('直接播放失败，建议使用外部播放器')
-      } else if (directOnly) {
-        setPlayerError('直接播放失败。当前为「客户端直连解码」模式，宿主机不转码；请使用支持该编码/封装的播放器（如 Infuse / VLC / Emby 客户端）播放，或关闭直连解码模式。')
-        toast.error('直接播放失败（客户端直连解码模式）')
-      } else if (playbackInfo?.provider === 'cloud115') {
-        const preferred =
-          findPlaybackQualityById(playbackInfo, selectedQuality) ??
-          findPlaybackQualityById(playbackInfo, playbackInfo.default_quality)
-        if (preferred?.source === 'cloud' && preferred.available) {
-          setCloudWaiting(false)
-          setHlsSource('cloud')
-          setPlaybackMode('hls')
-        } else if (preferred?.source === 'cloud') {
-          setHlsSource('cloud')
-          setCloudWaiting(true)
-          setCloudWaitMessage(preferred.note || `正在等待 115 转码 ${preferred.label}…`)
-          void startCloudTranscode(Number(preferred.id) || 4)
-        } else {
-          switchToLocalHLS(current?.currentTime || 0)
-        }
-      } else if (hlsUnavailable) {
-        setPlayerError('直接播放失败，且 HLS 转码不可用。请检查文件是否存在，或配置本机 ffmpeg 后使用 HLS 转码播放。')
-        toast.error('直接播放失败，HLS 转码不可用')
-      } else {
-        toast.error('直接播放失败，切换到 HLS 转码')
-        switchToLocalHLS(current?.currentTime || 0)
-      }
-    }
-
-    clearFallbackTimer()
-    // STRM/115 直链冷启动比本地文件慢，给足缓冲时间再判定失败。
-    const fallbackDelayMs = isStrmMedia(mediaRef.current) ? 4000 : 1500
-    fallbackTimerRef.current = setTimeout(fallbackDirectPlay, fallbackDelayMs)
-  }, [
-    clearFallbackTimer,
-    directOnly,
-    hlsUnavailable,
-    mediaId,
-    mode,
-    playbackInfo,
-    seekStrmDirectTo,
-    selectedQuality,
-    setPlaybackMode,
-    startCloudTranscode,
-    switchToLocalHLS,
-  ])
+    scheduleDirectFallback()
+  }, [mediaId, mode, scheduleDirectFallback, seekStrmDirectTo, showPlayFailure])
 
   const changeSubtitlePosition = useCallback((nextPosition: SubtitlePosition) => {
     setSubtitlePosition(nextPosition)

web/src/pages/directPlayError.ts

@@ -6,6 +6,65 @@ const HAVE_CURRENT_DATA = 2
 /** 程序化续播/跳转后忽略直连误报的宽限期（毫秒）。 */
 export const DIRECT_SEEK_GRACE_MS = 12_000
 
+/** 直连起播看门狗的等待上限（毫秒）。 */
+export const DIRECT_STALL_TIMEOUT_MS = 6_000
+/** STRM / 云盘直链要等 302 换链，冷启动比本地文件慢，给长一点的期限。 */
+export const DIRECT_STALL_TIMEOUT_STRM_MS = 12_000
+/**
+ * 没有可退线路（远程 Emby 挂载、客户端直连解码）时的等待上限。
+ *
+ * 这两种源都不在宿主机转码，等不到数据时没有任何补救动作——唯一的输出是给用户
+ * 一条提示，所以宁可多等：跨网上游的冷启动实测要 8–11 秒才吐第一段数据，6 秒
+ * 就判失败只会把「上游慢」写成「浏览器不支持这个编码」，而画面其实几秒后就播上了。
+ */
+export const DIRECT_STALL_TIMEOUT_NO_FALLBACK_MS = 20_000
+
+/**
+ * 起播看门狗该等多久才把「还没起来」当成「起不来」。
+ *
+ * canFallback=false：直连失败没有可切的线路（远程 Emby / 直连解码），失败提示
+ * 是唯一产出，所以要容忍慢启动。canFallback=true：还能切 HLS，早点换线路对用户
+ * 更有利，保持较短期限。
+ */
+export function directStallTimeoutMs(input: { isStrm?: boolean; canFallback: boolean }): number {
+  if (!input.canFallback) return DIRECT_STALL_TIMEOUT_NO_FALLBACK_MS
+  return input.isStrm ? DIRECT_STALL_TIMEOUT_STRM_MS : DIRECT_STALL_TIMEOUT_MS
+}
+
+/** 直连起播失败的两类起因：看门狗等超时（上游一直没数据）/ 媒体元素真的报了 error。 */
+export type DirectPlayFailureReason = 'stall' | 'error'
+
+/** 直连失败时用户看到的来源：远程 Emby 挂载 / 客户端直连解码。 */
+export type DirectPlayFailureKind = 'remote-emby' | 'direct-only'
+
+/**
+ * 直连失败的横幅与 toast 文案。
+ *
+ * 超时和真报错的文案必须分开：超时的真实原因是上游慢，不该断言「浏览器不支持这个
+ * 编码」（实测最常见的情形就是慢，画面随后自己就播上了）。
+ */
+export function directPlayFailureCopy(
+  kind: DirectPlayFailureKind,
+  reason: DirectPlayFailureReason = 'error',
+): { banner: string; toast: string } {
+  if (reason === 'stall') {
+    return {
+      banner:
+        kind === 'remote-emby'
+          ? '直接播放失败：远程 Emby 挂载的流长时间没有返回数据（该线路不转码，只能直连）。可稍后重试，或用外部播放器（如 PotPlayer / VLC / IINA）播放。'
+          : '直接播放失败：客户端直连解码模式下宿主机不转码，流长时间没有返回数据。可稍后重试，或关闭直连解码模式改用转码播放。',
+      toast: '直连流长时间没有响应',
+    }
+  }
+  return {
+    banner:
+      kind === 'remote-emby'
+        ? '直接播放失败。该媒体为远程 Emby 挂载直连播放（不进行转码）；当前浏览器可能不支持该视频编码或音频格式，建议使用外部播放器（如 PotPlayer / VLC / IINA）播放。'
+        : '直接播放失败。当前为「客户端直连解码」模式，宿主机不转码；请使用支持该编码/封装的播放器（如 Infuse / VLC / Emby 客户端）播放，或关闭直连解码模式。',
+    toast: '直接播放失败，建议使用外部播放器',
+  }
+}
+
 /**
  * Chromium 常把被中断的直连（换 src、插入 track、302 未完成就 seek）
  * 报成 error。真正不兼容应 fallback；瞬时中断应忽略或静默重试一次。

web/src/pages/hlsPlayback.test.ts

@@ -0,0 +1,58 @@
+import { strict as assert } from 'node:assert'
+import { applyPendingHlsSeek, attachNativeHls, bindPendingHlsSeek, hlsManifestTimeoutMs, playbackPositionSec } from './hlsPlayback.ts'
+
+class FakeVideo extends EventTarget {
+  src = ''
+  rejectSeek = false
+  position = 0
+  playCalls = 0
+  get currentTime() { return this.position }
+  set currentTime(value: number) {
+    if (this.rejectSeek) throw new Error('metadata not ready')
+    this.position = value
+  }
+  play() { this.playCalls += 1; return Promise.resolve() }
+}
+
+const video = new FakeVideo()
+const pending = { current: 420 as number | null }
+const detach = attachNativeHls(video, '/cloud/master.m3u8', pending)
+assert.equal(video.src, '/cloud/master.m3u8')
+assert.equal(video.playCalls, 1)
+assert.equal(video.currentTime, 0, 'seek must wait for the new source metadata')
+video.rejectSeek = true
+video.dispatchEvent(new Event('loadedmetadata'))
+assert.equal(pending.current, 420, 'a rejected seek must survive until canplay')
+video.rejectSeek = false
+video.dispatchEvent(new Event('canplay'))
+assert.equal(video.currentTime, 420, 'native HLS must restore the absolute cloud position')
+assert.equal(pending.current, null)
+pending.current = 600
+detach()
+video.dispatchEvent(new Event('loadedmetadata'))
+assert.equal(video.currentTime, 420, 'a detached source must not apply late metadata events')
+assert.equal(pending.current, 600)
+
+const msePending = { current: 120 as number | null }
+video.rejectSeek = true
+assert.equal(applyPendingHlsSeek(video, msePending), false)
+assert.equal(msePending.current, 120)
+video.rejectSeek = false
+assert.equal(applyPendingHlsSeek(video, msePending), true)
+assert.equal(msePending.current, null)
+const lateMsePending = { current: 900 as number | null }
+video.rejectSeek = true
+const detachMse = bindPendingHlsSeek(video, lateMsePending)
+assert.equal(applyPendingHlsSeek(video, lateMsePending), false)
+video.rejectSeek = false
+video.dispatchEvent(new Event('loadedmetadata'))
+assert.equal(video.currentTime, 900, 'hls.js seeks rejected during manifest parsing must retry after metadata')
+assert.equal(lateMsePending.current, null)
+detachMse()
+assert.ok(hlsManifestTimeoutMs('local', 1800) > 120_000, 'deep seeks must outlast the server readiness deadline')
+assert.equal(hlsManifestTimeoutMs('local', 0), 60_000)
+assert.equal(hlsManifestTimeoutMs('cloud', 1800), 60_000)
+assert.equal(playbackPositionSec('hls', 'cloud', 420, 600), 600, 'cloud subtitle/quality switches must not add an old local offset')
+assert.equal(playbackPositionSec('hls', 'local', 420, 60), 480, 'local subtitle switches preserve the source position')
+assert.equal(playbackPositionSec('direct', 'local', 420, 60), 60)
+console.log('hlsPlayback.test.ts ok')

web/src/pages/hlsPlayback.ts

@@ -0,0 +1,43 @@
+type HlsVideo = Pick<HTMLVideoElement, 'src' | 'currentTime' | 'addEventListener' | 'removeEventListener' | 'play'>
+type PendingSeek = { current: number | null }
+
+/** Local mid-file restarts may wait 120 seconds on the server before a playlist exists. */
+export function hlsManifestTimeoutMs(source: 'cloud' | 'local', startSec: number): number {
+  return source === 'local' && startSec > 0.05 ? 135_000 : 60_000
+}
+
+/** Keep a pending seek until the media element accepts it; metadata can arrive later. */
+export function applyPendingHlsSeek(video: Pick<HlsVideo, 'currentTime'>, pending: PendingSeek): boolean {
+  if (pending.current === null) return true
+  if (!Number.isFinite(pending.current) || pending.current < 0) return false
+  try {
+    video.currentTime = pending.current
+    pending.current = null
+    return true
+  } catch {
+    return false
+  }
+}
+
+export function bindPendingHlsSeek(video: HlsVideo, pending: PendingSeek): () => void {
+  const restore = () => { applyPendingHlsSeek(video, pending) }
+  video.addEventListener('loadedmetadata', restore)
+  video.addEventListener('canplay', restore)
+  return () => {
+    video.removeEventListener('loadedmetadata', restore)
+    video.removeEventListener('canplay', restore)
+  }
+}
+
+/** Native HLS needs the same seek restoration as hls.js, with removable listeners. */
+export function attachNativeHls(video: HlsVideo, url: string, pending: PendingSeek): () => void {
+  const detach = bindPendingHlsSeek(video, pending)
+  video.src = url
+  void video.play().catch(() => undefined)
+  return detach
+}
+
+/** Only locally transcoded HLS uses a relative timeline starting at startSec. */
+export function playbackPositionSec(mode: 'direct' | 'hls', source: 'cloud' | 'local', startSec: number, currentTime: number): number {
+  return mode === 'hls' && source === 'local' ? startSec + currentTime : currentTime
+}

web/src/pages/playerPageModel.test.ts

@@ -1,4 +1,11 @@
-import { classifyDirectPlayError } from './directPlayError.ts'
+import {
+  classifyDirectPlayError,
+  directPlayFailureCopy,
+  directStallTimeoutMs,
+  DIRECT_STALL_TIMEOUT_MS,
+  DIRECT_STALL_TIMEOUT_NO_FALLBACK_MS,
+  DIRECT_STALL_TIMEOUT_STRM_MS,
+} from './directPlayError.ts'
 
 function check(name: string, cond: boolean) {
   if (!cond) throw new Error(`classifyDirectPlayError: ${name}`)
@@ -96,4 +103,36 @@ check(
   }) === 'ignore',
 )
 
+// 起播看门狗的期限：没有可退线路（远程 Emby / 直连解码）时必须容忍慢启动，
+// 否则上游冷启动的 8–11 秒会被误报成「编码不兼容」。
+check('local direct stall waits 6s', directStallTimeoutMs({ canFallback: true }) === DIRECT_STALL_TIMEOUT_MS)
+check(
+  'strm direct stall waits 12s',
+  directStallTimeoutMs({ isStrm: true, canFallback: true }) === DIRECT_STALL_TIMEOUT_STRM_MS,
+)
+check(
+  'no fallback line waits 20s',
+  directStallTimeoutMs({ canFallback: false }) === DIRECT_STALL_TIMEOUT_NO_FALLBACK_MS,
+)
+check(
+  'no fallback line ignores strm tuning',
+  directStallTimeoutMs({ isStrm: true, canFallback: false }) === DIRECT_STALL_TIMEOUT_NO_FALLBACK_MS,
+)
+
+const stallCopy = directPlayFailureCopy('remote-emby', 'stall')
+check('stall copy blames the upstream, not the codec', !stallCopy.banner.includes('不支持'))
+check('stall copy keeps the failure prefix', stallCopy.banner.startsWith('直接播放失败'))
+const decodeCopy = directPlayFailureCopy('remote-emby', 'error')
+check('decode copy still mentions unsupported codec', decodeCopy.banner.includes('不支持'))
+check(
+  'copy defaults to the decode wording',
+  directPlayFailureCopy('remote-emby').banner === decodeCopy.banner,
+)
+check(
+  'direct-only copy names the client-decode mode',
+  directPlayFailureCopy('direct-only', 'stall').banner.includes('客户端直连解码') &&
+    directPlayFailureCopy('direct-only', 'error').banner.includes('客户端直连解码'),
+)
+check('stall toast is not the external-player advice', stallCopy.toast !== decodeCopy.toast)
+
 console.log('playerPageModel.test.ts ok')

web/src/pages/reader/ReaderReplacePage.tsx

@@ -35,7 +35,6 @@ export default function ReaderReplacePage() {
       .then(setRules)
       .catch(() => toast.error('加载替换规则失败'))
   }
-  // eslint-disable-next-line react-hooks/exhaustive-deps
   useEffect(load, [])
 
   const save = async () => {

web/src/pages/reader/ReaderSourcesPage.tsx

@@ -49,7 +49,6 @@ export default function ReaderSourcesPage() {
       .then(setSources)
       .catch(() => toast.error('加载书源失败'))
   }
-  // eslint-disable-next-line react-hooks/exhaustive-deps
   useEffect(load, [])
 
   // 读取书源文件（选择或拖入），走与粘贴相同的导入链路

web/src/pages/reader/ReaderViewPage.tsx

@@ -1,4 +1,4 @@
-import { memo, useCallback, useEffect, useLayoutEffect, useMemo, useRef, useState, type CSSProperties } from 'react'
+import { memo, useCallback, useEffect, useLayoutEffect, useMemo, useRef, useState, type CSSProperties, type ReactNode } from 'react'
 import { useParams } from 'react-router-dom'
 import toast from 'react-hot-toast'
 import {
@@ -266,6 +266,19 @@ function settleTextDrag(dragEl: HTMLDivElement, turn: 1 | -1, dx: number, stride
   })
 }
 
+function ReaderMenuButton({ icon, label, onClick }: { icon: ReactNode; label: string; onClick: () => void }) {
+  return (
+    <button
+      type="button"
+      onClick={onClick}
+      className="flex flex-1 flex-col items-center gap-1 py-1 opacity-80 hover:opacity-100"
+    >
+      {icon}
+      <span className="text-2xs">{label}</span>
+    </button>
+  )
+}
+
 export default function ReaderViewPage() {
   const { bookId = '' } = useParams()
   const settings = useReaderSettingsStore()
@@ -316,6 +329,7 @@ export default function ReaderViewPage() {
   const [page, setPage] = useState(0)
   const [pageCount, setPageCount] = useState(1)
   const [vw, setVw] = useState(0)
+  const [scrollProgress, setScrollProgress] = useState(0)
   // 文本翻页：位移与页脚先走 DOM，React 的 page state 等到过渡结束后再同步。
   // 若在同一次点击/触摸里同步 setPage，整页重渲染会堵住下一次绘制，
   // 命令式 transform 也要等卡完才开始——点按和滑动都会顿一下。
@@ -338,8 +352,13 @@ export default function ReaderViewPage() {
   const paragraphWarmRef = useRef<Uint8Array | null>(null)
   /** 当前窗口中心页；-1 表示窗口失效（映射刚重算过），下次应用时重建。 */
   const warmCenterRef = useRef(-1)
-  vwRef.current = vw
-  pageCountRef.current = pageCount
+  const currentChapter = chapterIndex !== null ? chapters[chapterIndex] : null
+  const chapterTitle = currentChapter?.title ?? ''
+  useLayoutEffect(() => {
+    vwRef.current = vw
+    pageCountRef.current = pageCount
+    chapterTitleRef.current = chapterTitle
+  }, [vw, pageCount, chapterTitle])
   // 音频/漫画媒体状态
   const [media, setMedia] = useState<ReaderChapterContent | null>(null)
   const [restorePos, setRestorePos] = useState(0) // 音频秒数 / 漫画图片序号
@@ -448,6 +467,7 @@ export default function ReaderViewPage() {
   // ── 加载书籍与章节 ──
   useEffect(() => {
     let cancelled = false
+    const cache = contentCache.current
     ;(async () => {
       try {
         setLoadingStage('book')
@@ -510,7 +530,7 @@ export default function ReaderViewPage() {
     })()
     return () => {
       cancelled = true
-      contentCache.current.clear()
+      cache.clear()
     }
   }, [bookId, reloadKey])
 
@@ -910,6 +930,16 @@ export default function ReaderViewPage() {
     if (menuOpen) flushTextPageReactSync()
   }, [menuOpen, flushTextPageReactSync])
 
+  // Only refresh the scroll slider while its menu is visible, keeping scroll
+  // events on the reading path free of whole-page React renders.
+  useLayoutEffect(() => {
+    if (!menuOpen || settings.pageMode !== 'scroll') return
+    const el = scrollRef.current
+    if (el) {
+      setScrollProgress(Math.round((el.scrollTop / Math.max(1, el.scrollHeight - el.clientHeight)) * 1000))
+    }
+  }, [menuOpen, settings.pageMode, content])
+
   // 滚动模式恢复进度。
   //
   // 同时清掉 dataset.last：它是「上次保存的滚动百分比」，换章后必须重算。
@@ -1355,9 +1385,6 @@ export default function ReaderViewPage() {
     return () => el.removeEventListener('wheel', onWheel)
   }, [settings.pageMode, menuOpen, contentType, goPrev, goNext])
 
-  const currentChapter = chapterIndex !== null ? chapters[chapterIndex] : null
-  chapterTitleRef.current = currentChapter?.title ?? ''
-
   // 顶栏高度量一次：菜单打开时正文下移，开头几行不被顶栏压住。
   // 用 transform 而不是 padding，避免改变视口高度触发重新分页。
   const menuShiftStyle = useMemo(
@@ -1397,11 +1424,7 @@ export default function ReaderViewPage() {
       value:
         settings.pageMode === 'page'
           ? page + 1
-          : Math.round(
-              ((scrollRef.current?.scrollTop ?? 0) /
-                Math.max(1, (scrollRef.current?.scrollHeight ?? 1) - (scrollRef.current?.clientHeight ?? 1))) *
-                1000,
-            ),
+          : scrollProgress,
       onChange: (v: number) => {
         if (settings.pageMode === 'page') {
           commitTextPage(v - 1)
@@ -1569,6 +1592,7 @@ export default function ReaderViewPage() {
                 const max = el.scrollHeight - el.clientHeight
                 if (max > 0) {
                   const pct = el.scrollTop / max
+                  if (menuOpen) setScrollProgress(Math.round(pct * 1000))
                   if (shouldSaveScrollProgress(pct, el.dataset.last)) {
                     el.dataset.last = formatScrollPct(pct)
                     savePos(Math.round(el.scrollTop))
@@ -1753,62 +1777,35 @@ export default function ReaderViewPage() {
                 用 flex 等分而不是 grid-cols-N：动作数量随内容类型和书源能力变化，
                 grid 列数得写死成字面量类名，多一个按钮就要再加一档。 */}
             <div className="flex pt-1" style={{ color: theme.text }}>
-              {([
-                { icon: <LayoutList size={18} />, label: '目录', action: () => openPanel('toc') },
-                { icon: <BookOpen size={18} />, label: '界面', action: () => openPanel('style') },
-                {
-                  icon: settings.night ? <Sun size={18} /> : <Moon size={18} />,
-                  label: settings.night ? '日间' : '夜间',
-                  action: () => settings.toggleNight(),
-                },
-                {
-                  icon: settings.pageMode === 'page' ? <ScrollText size={18} /> : <ListEnd size={18} />,
-                  label: settings.pageMode === 'page' ? '滚动' : '翻页',
-                  action: () => settings.setPageMode(settings.pageMode === 'page' ? 'scroll' : 'page'),
-                },
-                // 双页铺开只对漫画的翻页模式有意义；窗口太窄时并排两页没法看，
-                // 干脆不显示入口（不是灰掉，避免用户以为坏了）。
-                ...(contentType === 'image' && settings.pageMode === 'page' && windowWidth >= COMIC_SPREAD_MIN_WIDTH
-                  ? [
-                      {
-                        icon: <Columns2 size={18} />,
-                        label: settings.comicDoublePage ? '单页' : '双页',
-                        action: () => settings.setComicDoublePage(!settings.comicDoublePage),
-                      },
-                    ]
-                  : []),
-                // 全屏：只在浏览器支持元素全屏时出现（iOS Safari 不支持，见上方注释）。
-                ...(fullscreenSupported
-                  ? [
-                      {
-                        icon: fullscreen ? <Minimize size={18} /> : <Maximize size={18} />,
-                        label: fullscreen ? '还原' : '全屏',
-                        action: toggleFullscreen,
-                      },
-                    ]
-                  : []),
-                ...(canSwitchSource
-                  ? [
-                      {
-                        icon: <ArrowLeftRight size={18} />,
-                        label: '换源',
-                        action: () => {
-                          void openSourcePicker()
-                        },
-                      },
-                    ]
-                  : []),
-              ]).map((item) => (
-                <button
-                  key={item.label}
-                  type="button"
-                  onClick={item.action}
-                  className="flex flex-1 flex-col items-center gap-1 py-1 opacity-80 hover:opacity-100"
-                >
-                  {item.icon}
-                  <span className="text-2xs">{item.label}</span>
-                </button>
-              ))}
+              <ReaderMenuButton icon={<LayoutList size={18} />} label="目录" onClick={() => openPanel('toc')} />
+              <ReaderMenuButton icon={<BookOpen size={18} />} label="界面" onClick={() => openPanel('style')} />
+              <ReaderMenuButton
+                icon={settings.night ? <Sun size={18} /> : <Moon size={18} />}
+                label={settings.night ? '日间' : '夜间'}
+                onClick={() => settings.toggleNight()}
+              />
+              <ReaderMenuButton
+                icon={settings.pageMode === 'page' ? <ScrollText size={18} /> : <ListEnd size={18} />}
+                label={settings.pageMode === 'page' ? '滚动' : '翻页'}
+                onClick={() => settings.setPageMode(settings.pageMode === 'page' ? 'scroll' : 'page')}
+              />
+              {contentType === 'image' && settings.pageMode === 'page' && windowWidth >= COMIC_SPREAD_MIN_WIDTH && (
+                <ReaderMenuButton
+                  icon={<Columns2 size={18} />}
+                  label={settings.comicDoublePage ? '单页' : '双页'}
+                  onClick={() => settings.setComicDoublePage(!settings.comicDoublePage)}
+                />
+              )}
+              {fullscreenSupported && (
+                <ReaderMenuButton
+                  icon={fullscreen ? <Minimize size={18} /> : <Maximize size={18} />}
+                  label={fullscreen ? '还原' : '全屏'}
+                  onClick={toggleFullscreen}
+                />
+              )}
+              {canSwitchSource && (
+                <ReaderMenuButton icon={<ArrowLeftRight size={18} />} label="换源" onClick={() => { void openSourcePicker() }} />
+              )}
             </div>
 
             {/* 界面设置面板（主题 / 字号 / 行距 / 段距 / 漫画图片尺寸） */}

web/src/pages/reader/bookshelfModel.test.ts

@@ -1,5 +1,5 @@
-import type { ReaderBook } from '../../api/reader'
-import { formatRelativeTime, sortShelfBooks, unreadChapters } from './bookshelfModel'
+import type { ReaderBook } from '../../api/reader.ts'
+import { formatRelativeTime, sortShelfBooks, unreadChapters } from './bookshelfModel.ts'
 
 function check(name: string, condition: boolean) {
   if (!condition) throw new Error(`bookshelfModel: ${name}`)

web/src/pages/settingsGroupGeneral.ts

@@ -142,7 +142,7 @@ export const generalSettingsGroup: SettingGroup = {
       key: 'https.enabled',
       label: '启用 HTTPS',
       type: 'toggle',
-      hint: '开启后服务将仅通过 HTTPS 访问，保存后会自动切换到 https 地址，原 http 地址不再提供服务。必须先在下文配置好 SSL 证书与私钥（内容或路径任选其一，两者需匹配）才能启用；首次切换因浏览器以 origin（协议+域名+端口）隔离登录态，需要重新登录一次。',
+      hint: '开启后服务仅通过 HTTPS 提供访问，用 http:// 打开同一地址会被自动跳转到 https://（临时跳转，随时关掉 HTTPS 都不会被浏览器长期记住）；保存后当前页面会自动切换到 https 地址。必须先在下文配置好 SSL 证书与私钥（内容或路径任选其一，两者需匹配）才能启用；首次切换因浏览器以 origin（协议+域名+端口）隔离登录态，需要重新登录一次。',
       defaultValue: 'false',
     },
     {

web/src/utils/authRefresh.test.ts

@@ -5,6 +5,9 @@ import {
   RefreshRequestError,
   createRefreshCoordinator,
   createSingleFlightRunner,
+  createTokenPreflight,
+  decodeJwtExpiryMs,
+  isAccessTokenStale,
   refreshFailureOutcome,
   type RefreshOutcome,
   type RefreshTokens,
@@ -30,6 +33,149 @@ check('timeout is transient', refreshFailureOutcome(408) === 'transient')
 check('429 is transient', refreshFailureOutcome(429) === 'transient')
 check('unexpected status is transient', refreshFailureOutcome(418) === 'transient')
 
+// ── JWT 过期判定：首屏放行前必须能识别出「已经过期的 access token」 ────────
+// 现场（2026-10-16 12:49）：刷新页面时本地 token 已过期，React 首屏并发发出
+// 7 个受保护请求，全部 401，控制台一次性刷出 7 行红字；约 2 秒后才由拦截器
+// 刷新成功并重放。修复要求首屏就认得过期 token，先刷新再放行。
+function jwtWithExp(exp: number | string | null): string {
+  const encode = (value: unknown) =>
+    Buffer.from(JSON.stringify(value)).toString('base64url')
+  const claims: Record<string, unknown> = { uid: 'u1' }
+  if (exp !== null) claims.exp = exp
+  return `${encode({ alg: 'HS256', typ: 'JWT' })}.${encode(claims)}.signature`
+}
+
+const NOW = 1_700_000_000_000
+
+check('jwt expiry is decoded to milliseconds',
+  decodeJwtExpiryMs(jwtWithExp(NOW / 1000 + 3600)) === NOW + 3600_000)
+check('jwt without exp decodes to null', decodeJwtExpiryMs(jwtWithExp(null)) === null)
+check('jwt with non-numeric exp decodes to null', decodeJwtExpiryMs(jwtWithExp('soon')) === null)
+check('opaque token decodes to null', decodeJwtExpiryMs('not-a-jwt') === null)
+check('empty token decodes to null', decodeJwtExpiryMs('') === null)
+check('malformed payload decodes to null', decodeJwtExpiryMs('aaa.!!!.ccc') === null)
+
+// 已过期 / 即将过期的 token 必须在首屏就被拦下。
+check('expired token is stale', isAccessTokenStale(jwtWithExp(NOW / 1000 - 1), NOW))
+check('token expiring inside the skew is stale',
+  isAccessTokenStale(jwtWithExp(NOW / 1000 + 20), NOW))
+check('token outside the skew is fresh',
+  isAccessTokenStale(jwtWithExp(NOW / 1000 + 3600), NOW) === false)
+// 边界：正好落在 skew 上算过期（留出余量，避免刷新途中再次过期）。
+check('token exactly at the skew boundary is stale',
+  isAccessTokenStale(jwtWithExp(NOW / 1000 + 30), NOW))
+
+// 没有 token 或 token 无法解析时不能判定为「过期」：
+// 前者由路由守卫处理，后者强行刷新会变成无休止的刷新循环。
+check('missing token is not stale', isAccessTokenStale(null, NOW) === false)
+check('opaque token is not stale', isAccessTokenStale('opaque', NOW) === false)
+
+// ── 发送前预检：本地 token 已过期就先刷新，再让请求带新令牌出去 ────────────
+// 现场（2026-10-16 12:49）：刷新页面时本地 token 已过期，React 首屏并发发出
+// 7 个受保护请求，全部 401，控制台一次刷出 7 行红字；约 2 秒后才由 401 拦截器
+// 刷新成功并重放。修复要求请求在发出前就认得过期 token，先刷新再发送。
+async function expectPreflight(
+  name: string,
+  run: Promise<void>,
+  want: 'resolved' | 'rejected',
+) {
+  let outcome: 'resolved' | 'rejected' = 'resolved'
+  try {
+    await run
+  } catch {
+    outcome = 'rejected'
+  }
+  if (outcome !== want) throw new Error(`authRefresh: ${name} = ${outcome}, want ${want}`)
+}
+
+function preflightHarness(overrides: {
+  accessToken?: string | null
+  refresh?: () => Promise<RefreshOutcome>
+}) {
+  const state = { token: overrides.accessToken === undefined ? 'at-1' : overrides.accessToken }
+  const calls = { refresh: 0 }
+  const ensure = createTokenPreflight({
+    readAccessToken: () => state.token,
+    refreshSession: async () => {
+      calls.refresh += 1
+      if (overrides.refresh) return overrides.refresh()
+      state.token = jwtWithExp(Date.now() / 1000 + 3600)
+      return 'refreshed'
+    },
+    now: () => NOW,
+  })
+  return { state, calls, ensure }
+}
+
+{
+  // 令牌新鲜：不得产生任何刷新请求。
+  const { calls, ensure } = preflightHarness({ accessToken: jwtWithExp(NOW / 1000 + 3600) })
+  await expectPreflight('fresh token passes straight through', ensure(), 'resolved')
+  check('fresh token skips the refresh', calls.refresh === 0)
+}
+
+{
+  // 核心回归：过期令牌必须先刷新，且刷新完成后才返回（调用方据此发请求）。
+  const { state, calls, ensure } = preflightHarness({ accessToken: jwtWithExp(NOW / 1000 - 60) })
+  await expectPreflight('stale token refreshes before sending', ensure(), 'resolved')
+  check('stale token refreshed exactly once', calls.refresh === 1)
+  check('preflight handed out a fresh token',
+    !isAccessTokenStale(state.token, NOW))
+}
+
+{
+  // 并发首屏：7 个请求同时预检，只允许刷新一次（协调器本身是单飞，
+  // 这里确认预检不会绕过它）。
+  const { calls, ensure } = preflightHarness({ accessToken: jwtWithExp(NOW / 1000 - 60) })
+  await Promise.all([ensure(), ensure(), ensure(), ensure(), ensure(), ensure(), ensure()])
+  check('concurrent preflight refreshes once', calls.refresh === 1)
+}
+
+{
+  // 瞬时故障（部署/重启窗口）：保留会话，照常发出请求——
+  // 页面自己会重试，不该把人挡在门外。
+  const { ensure } = preflightHarness({
+    accessToken: jwtWithExp(NOW / 1000 - 60),
+    refresh: async () => 'transient',
+  })
+  await expectPreflight('transient failure still sends the request', ensure(), 'resolved')
+}
+
+{
+  // 服务端明确判定凭证失效：不发请求，直接失败，避免刷出无意义的 401。
+  const { ensure } = preflightHarness({
+    accessToken: jwtWithExp(NOW / 1000 - 60),
+    refresh: async () => 'invalid',
+  })
+  await expectPreflight('revoked session refuses to send', ensure(), 'rejected')
+}
+
+{
+  // 未登录：交给路由守卫和 401 路径处理，预检不替它们做决定。
+  const { calls, ensure } = preflightHarness({ accessToken: null })
+  await expectPreflight('anonymous request is not preflighted', ensure(), 'resolved')
+  check('anonymous request does not refresh', calls.refresh === 0)
+}
+
+{
+  // 非 JWT 令牌（Emby 等外部客户端签发）：读不出 exp，按新鲜处理，
+  // 绝不能因此陷入刷新循环。
+  const { calls, ensure } = preflightHarness({ accessToken: 'opaque-token' })
+  await expectPreflight('opaque token passes through', ensure(), 'resolved')
+  check('opaque token does not refresh', calls.refresh === 0)
+}
+
+{
+  // 刷新抛出异常（网络不可达）：按瞬时故障处理，请求照发。
+  const { ensure } = preflightHarness({
+    accessToken: jwtWithExp(NOW / 1000 - 60),
+    refresh: async () => {
+      throw new Error('network down')
+    },
+  })
+  await expectPreflight('throwing refresh still sends the request', ensure(), 'resolved')
+}
+
 // ── 单飞：并发调用共享同一次执行，结束后可重新执行 ────────────────────────
 {
   let calls = 0

web/src/utils/authRefresh.ts

@@ -15,6 +15,46 @@
 /** 一次刷新尝试的结果。 */
 export type RefreshOutcome = 'refreshed' | 'invalid' | 'transient'
 
+/**
+ * access token 过期前多久就视为失效。刷新要留出余量：刚好卡在 exp 上
+ * 刷新，重放的原请求可能在服务端又已经过期。
+ */
+export const TOKEN_EXPIRY_SKEW_MS = 30_000
+
+/**
+ * 解析 JWT 的 exp 声明并换算成毫秒时间戳。
+ *
+ * 解析失败（非 JWT、base64url 损坏、缺少 exp、exp 不是数字）一律返回 null：
+ * 调用方据此放行，绝不能把「读不懂」当成「已过期」，否则不透明的令牌会导致
+ * 无休止的刷新循环。
+ */
+export function decodeJwtExpiryMs(token: string | null | undefined): number | null {
+  if (!token) return null
+  const segment = token.split('.')[1]
+  if (!segment) return null
+  try {
+    // JWT 使用 base64url 编码，需先还原字符集并补齐 padding。
+    const normalized = segment.replace(/-/g, '+').replace(/_/g, '/')
+    const padded = normalized.padEnd(Math.ceil(normalized.length / 4) * 4, '=')
+    const claims = JSON.parse(atob(padded)) as { exp?: unknown }
+    return typeof claims.exp === 'number' ? claims.exp * 1000 : null
+  } catch {
+    return null
+  }
+}
+
+/**
+ * token 是否已过期或即将过期。
+ *
+ * 首屏放行前用它做一次判断：本地持久化的 token 往往已经过期，直接发请求只
+ * 会一次性刷出一屏 401。读不出 exp 的令牌按「新鲜」处理（见 decodeJwtExpiryMs）。
+ */
+export function isAccessTokenStale(token: string | null | undefined, nowMs = Date.now()): boolean {
+  const expiry = decodeJwtExpiryMs(token)
+  if (expiry === null) return false
+  return nowMs >= expiry - TOKEN_EXPIRY_SKEW_MS
+}
+
 /** 刷新接口返回的令牌对。 */
 export interface RefreshTokens {
   token: string
@@ -101,6 +141,44 @@ export function createBrowserExclusiveRunner(
   }
 }
 
+export interface SessionGateDeps {
+  /** 读取当前 access token（可能来自持久化存储，往往已经过期）。 */
+  readAccessToken: () => string | null
+  /** 刷新会话；语义与 useAuthStore.refreshSession 一致。 */
+  refreshSession: () => Promise<RefreshOutcome>
+  /** 可注入的时间源，便于测试。 */
+  now?: () => number
+}
+
+/**
+ * 构造请求前预检：每个请求发出前调用，必要时先把过期的令牌换新。
+ *
+ * 本地持久化的 access token 常常已经过期。此时若直接发请求，首屏那些各自
+ * 独立的 useEffect 会在同一帧里并发打出受保护请求，全部拿到 401，控制台
+ * 一次刷出一整屏红字，用户还要等 401 拦截器刷新完才能看到内容
+ * （现场：2026-10-16 12:49，7 个请求同时 401）。
+ *
+ * 预检把「先刷新、再发送」放在请求发出之前：
+ *   - token 新鲜（或无法解析 exp）→ 立即返回；
+ *   - token 已过期 → 等待刷新（刷新协调器自带单飞，并发请求只发一次）；
+ *   - 刷新遇到瞬时故障 → 照常发送（页面自己会重试）；
+ *   - 服务端明确判定凭证失效 → 抛出，请求根本不发出去。
+ *
+ * 未登录（没有 token）直接返回：跳登录页是路由守卫的职责。
+ */
+export function createTokenPreflight(deps: SessionGateDeps): () => Promise<void> {
+  return async () => {
+    const token = deps.readAccessToken()
+    if (!token) return
+    const now = deps.now ?? Date.now
+    if (!isAccessTokenStale(token, now())) return
+    const outcome = await deps.refreshSession().catch((): RefreshOutcome => 'transient')
+    if (outcome === 'invalid') {
+      throw new Error('session expired')
+    }
+  }
+}
+
 export interface RefreshCoordinatorDeps {
   readAccessToken: () => string | null
   readRefreshToken: () => string | null

web/src/utils/chapterGroups.test.ts

@@ -1,4 +1,4 @@
-import { CHAPTER_GROUP_SIZE, buildChapterGroups, chapterGroupIndexOf } from './chapterGroups'
+import { CHAPTER_GROUP_SIZE, buildChapterGroups, chapterGroupIndexOf } from './chapterGroups.ts'
 
 // 目录分段分组的纯逻辑回归测试（对齐 utils/readerBookGroups.test.ts 的写法）。
 

web/src/utils/mediaNavigation.ts

@@ -2,6 +2,17 @@ import type { Media } from '../types'
 import { isRemoteEmbyID } from './remoteEmby'
 import { getSeriesKey, isEpisodeLike, pathLooksEpisodic } from './groupSeries'
 
+export function mediaSearchLink(query: string, page = 1): string {
+  const params = new URLSearchParams({ q: query.trim() })
+  if (page > 1) params.set('page', String(page))
+  return `/search?${params}`
+}
+
+export function mediaSearchReturnPath(state: unknown): string | undefined {
+  if (!state || typeof state !== 'object' || !('from' in state)) return undefined
+  return typeof state.from === 'string' && state.from.startsWith('/search?') ? state.from : undefined
+}
+
 function pathLooksLikeMovie(media: Media): boolean {
   const path = (media.path || media.display_library_path || media.library_path || '').toLowerCase()
   return /[\\/](?:movie|movies|film|films|电影)[\\/]/i.test(path)

web/src/utils/mediaVersion.test.ts

@@ -1,4 +1,5 @@
 import {
+  autoPlayEpisodeFor,
   isStrmMedia,
   mediaVersionFileName,
   mediaVersionLabel,
@@ -64,4 +65,25 @@ check('strm filename drops outer extension', mediaVersionFileName(cloudRow) ===
 check('missing path has no filename', mediaVersionFileName(media({ path: '' })) === '')
 check('isStrmMedia detects strm url', isStrmMedia(cloudRow))
 
+// 首页轮播/剧集卡片交给播放器的是「整部剧」ID：选集列表里没有它自己，
+// 播放器要自动落到第一集，否则 /api/stream/<剧集ID> 会被远程 Emby 打回 500。
+const series = media({ id: 'embyremote~mount-1~222679', title: '某部新番', path: '/anime/某部新番', size_bytes: 0 })
+const ep1 = media({ id: 'embyremote~mount-1~222681', path: '/anime/某部新番/Season 1/S1E01.mp4', size_bytes: 299580311 })
+const ep2 = media({ id: 'embyremote~mount-1~222683', path: '/anime/某部新番/Season 1/S1E02.mp4', size_bytes: 299580312 })
+
+check('series falls back to its first episode', autoPlayEpisodeFor(series, series.id, [ep1, ep2])?.id === ep1.id)
+check('playable episode stays untouched', autoPlayEpisodeFor(ep1, ep1.id, [ep1, ep2]) === null)
+check('single-item movie list stays untouched', autoPlayEpisodeFor(ep1, ep1.id, [{ ...ep1 }]) === null)
+check('waits until the requested item is loaded', autoPlayEpisodeFor(null, series.id, [ep1]) === null)
+check('ignores a response for another id', autoPlayEpisodeFor(ep2, series.id, [ep1]) === null)
+check('empty episode list has no fallback', autoPlayEpisodeFor(series, series.id, []) === null)
+check(
+  'entry present as another version stays untouched',
+  autoPlayEpisodeFor(
+    media({ id: 'v-4k' }),
+    'v-4k',
+    [media({ id: 'v-4k', versions: [local4k, local1080] })],
+  ) === null,
+)
+
 console.log('mediaVersion.test.ts: ok')

web/src/utils/mediaVersion.ts

@@ -195,6 +195,32 @@ export function mediaVersionSourceLabel(media: Media): string {
   return isStrmMedia(media) ? '云端直链' : '本地文件'
 }
 
+/**
+ * 挑出「播放器该自动跳转到的那一集」。
+ *
+ * 传给 /play/:id 的可能是「整部剧」这类自身没有可播文件的条目——首页轮播、远程 Emby
+ * 剧集库的卡片、剧集详情页的播放按钮都会把剧集 ID 直接交给播放器。这种条目的选集接口
+ * 返回的是它下面的剧集，列表里没有它自己；若直接拿剧集 ID 去请求
+ * /api/stream/<剧集ID>，远程 Emby 会返回 HTTP 500（Unable to cast object of type
+ * 'Series' to type 'IHasMediaSources'），用户看到的就是「播放失败」。此时换成第一集，
+ * 播放、进度、上下集都落在真正可播的条目上。
+ *
+ * 返回 null 表示无需跳转：条目还没加载完、条目本身可播（含以其它版本形式出现在列表里）、
+ * 或该条目下没有任何可选剧集。
+ */
+export function autoPlayEpisodeFor(
+  media: Media | null | undefined,
+  requestedId: string,
+  episodes: Media[],
+): Media | null {
+  if (!media || !requestedId || media.id !== requestedId) return null
+  if (episodes.length === 0) return null
+  if (episodes.some((episode) => mediaVersionMatches(episode, media.id))) return null
+  const first = episodes[0]
+  if (!first?.id || first.id === media.id) return null
+  return first
+}
+
 /** 版本文件名（去掉目录，并剥掉 .strm 外层扩展名），用作版本选项的次要说明。 */
 export function mediaVersionFileName(media: Media): string {
   const path = (media.path || '').replace(/\\/g, '/').trim()

web/src/utils/readerBookGroups.test.ts

@@ -10,7 +10,7 @@ import {
   normalizeBookGroups,
   reorderBookGroups,
   resolveSelectedGroupId,
-} from './readerBookGroups'
+} from './readerBookGroups.ts'
 
 // 书架分组的纯逻辑回归测试（对齐 utils/libraryTags.test.ts 的写法）。