fix a bug

.gitignore

@@ -9,3 +9,5 @@ sentence_api/data/v/
 sentence_api/data/work/
 sentence_api/data/attempts/
 android/sample-app/release/
+
+.tmp/

README.md

@@ -10,11 +10,17 @@ pip 安装依赖后即可直接播放视频。
 - 播放 / 暂停 / 停止、上一首 / 下一首
 - 前进 / 后退（步长可自定义）与进度条点击拖动定位
 - 快进步长可在界面底部直接填写（1-600 秒），自动保存、下次启动自动加载
-- 按句跳转：`, / .` 直接跳转上一句/下一句；勾选“按句跳转”后 `← / →` 也可
-  跳句。句边界通过线上服务生成，PC 仅上传本地提取的音频轨，服务器处理完成
-  后返回边界且不保存音频；结果按文件 SHA-256 缓存在本机。
+- 按句跳转：先点“生成句边界”。生成时会先按 SHA-256 查询线上是否已有同一
+  视频的句边界；有则直接取回，没有时才本地提取音频并上传。生成完成后才能
+  勾选“按句跳转”，之后 `, / .` 或 `← / →` 可跳句。PC 会按文件 SHA-256
+  缓存结果；生成过程有进度条。
+- 画面字幕：句边界就绪后，当前句会叠在视频画面底部。可勾选“中文字幕”
+  （快捷键 `C`）显示或关闭中文翻译；翻译由服务端生成并缓存。
+- 导出字幕：句边界就绪后可点“导出字幕”，默认保存为 UTF-8（带 BOM）的 `.ass`（中英同一条字幕避免重叠），也可另存 `.srt`。
 - AI 讲解：播放中按 `E` 或点击“AI讲解 (E)”，由服务端 AI 讲解当前句子。
 - 服务配置：点击“服务…”可填写服务器地址，或使用账号密码登录后保存令牌。
+- 调试日志：点击“调试日志”可查看/复制请求状态、耗时和错误摘要；日志保存在
+  `~/.mediaplayer_logs/mediaplayer.log`，不包含登录令牌。
 - 音量调节、静音，0.5x - 2.0x 变速播放
 - 三种循环模式：关 / 单曲循环 / 列表循环
 - 播放列表管理（打开文件、打开文件夹、删除条目、双击播放）
@@ -80,6 +86,8 @@ pip install -r requirements.txt
 | `N` / `P` | 下一首 / 上一首 |
 | `[` / `]` | 减速 / 加速（0.5x - 2.0x） |
 | `L` | 切换循环模式（关 / 单曲 / 列表） |
+| `C` | 开 / 关中文字幕 |
+| 双击画面 | 进入 / 退出全屏 |
 | `F` | 视频全屏切换 |
 | `Esc` | 退出全屏 |
 | `Ctrl+O` | 打开文件（macOS 亦支持 `Command+O`） |
@@ -101,7 +109,8 @@ pip install -r requirements.txt
 静音阈值会根据每个文件的背景噪声自动调整，带环境噪音的录音也能正常分句。
 分析在后台进行，首次播放某文件时状态栏会提示“正在分析句边界…”，完成后
 自动缓存；无法识别（如纯音乐、无音频轨）时会提示并回退为普通步进跳转。
-“按句跳转”的开关状态同样会自动保存。
+“按句跳转”和“中文字幕”的开关状态同样会自动保存。句边界就绪后，原文会叠在
+视频画面底部；打开中文字幕时会向服务器请求翻译并一并显示。
 
 ## 常见问题
 

android/oral-trainer-sdk/src/main/AndroidManifest.xml

@@ -1,3 +1,5 @@
 <manifest xmlns:android="http://schemas.android.com/apk/res/android">
     <uses-permission android:name="android.permission.INTERNET" />
+    <uses-permission android:name="android.permission.WAKE_LOCK" />
+    <uses-permission android:name="android.permission.ACCESS_WIFI_STATE" />
 </manifest>

android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/OralTrainerController.kt

@@ -10,8 +10,10 @@ import androidx.media3.common.C
 import androidx.media3.common.MediaItem
 import androidx.media3.common.PlaybackException
 import androidx.media3.common.Player
+import androidx.media3.common.TrackSelectionParameters.AudioOffloadPreferences
 import androidx.media3.exoplayer.ExoPlayer
 import androidx.media3.exoplayer.PlayerMessage
+import androidx.media3.exoplayer.SeekParameters
 import androidx.media3.exoplayer.source.DefaultMediaSourceFactory
 import java.util.concurrent.CopyOnWriteArraySet
 import kotlin.math.max
@@ -31,8 +33,15 @@ class OralTrainerController internal constructor(
     private var lastSentenceIndex: Int? = null
     private var stopAtSentenceEnd = false
     private var stopAtSentenceEndIndex: Int? = null
-    private var boundaryMessage: PlayerMessage? = null
-    private var boundaryGuardToken = 0
+    private var baseItemIndex = 0
+    private var clippedSentence: SentenceBoundary? = null
+    private var updatingPlaybackWindow = false
+    private var unclippedDurationMs = -1L
+    private var sentenceEndMessage: PlayerMessage? = null
+    private var playFromStartAfterReady = false
+    private var clipStartNeedsWarmup = false
+    private var clipStartWarmupVolume: Float? = null
+    private val finishClipStartWarmupRunnable = Runnable { finishClipStartWarmup() }
 
     var config: PlayerConfig = initialConfig
         private set
@@ -47,10 +56,10 @@ class OralTrainerController internal constructor(
             if (released) {
                 return
             }
-            enforceSentenceBoundary()
+            enforceStopAtSentenceEnd()
             notifySentenceIfChanged()
             notifySnapshot()
-            mainHandler.postDelayed(this, 250L)
+            mainHandler.postDelayed(this, IDLE_POLL_MS)
         }
     }
 
@@ -60,30 +69,63 @@ class OralTrainerController internal constructor(
         )
         player = ExoPlayer.Builder(appContext)
             .setMediaSourceFactory(mediaSourceFactory)
+            .setWakeMode(C.WAKE_MODE_NETWORK)
             .build()
             .also { exoPlayer ->
                 exoPlayer.repeatMode = Player.REPEAT_MODE_ALL
                 exoPlayer.playWhenReady = initialConfig.autoPlay
+                exoPlayer.setSeekParameters(SeekParameters.EXACT)
+                exoPlayer.setForegroundMode(true)
+                exoPlayer.trackSelectionParameters = exoPlayer.trackSelectionParameters
+                    .buildUpon()
+                    .setAudioOffloadPreferences(
+                        AudioOffloadPreferences.Builder()
+                            .setAudioOffloadMode(AudioOffloadPreferences.AUDIO_OFFLOAD_MODE_DISABLED)
+                            .build()
+                    )
+                    .build()
                 exoPlayer.addListener(object : Player.Listener {
                     override fun onPlaybackStateChanged(playbackState: Int) {
+                        if (playbackState == Player.STATE_ENDED) {
+                            if (clipStartWarmupVolume != null) {
+                                finishClipStartWarmup()
+                            } else {
+                                onClippedSentenceEnded()
+                            }
+                        }
+                        if (playbackState == Player.STATE_READY) {
+                            if (playFromStartAfterReady) {
+                                playFromStartAfterReady = false
+                                playClippedSentenceFromStart()
+                            }
+                            if (stopAtSentenceEnd) {
+                                scheduleSentenceEndGuard()
+                            }
+                            if (unclippedDurationMs <= 0L) {
+                                val duration = player.duration
+                                if (duration != C.TIME_UNSET && clippedSentence == null) {
+                                    unclippedDurationMs = duration
+                                }
+                            }
+                        }
                         notifySnapshot()
                     }
 
                     override fun onIsPlayingChanged(isPlaying: Boolean) {
-                        if (isPlaying) {
-                            refreshSentenceBoundaryGuard()
-                        } else {
-                            cancelSentenceBoundaryGuard()
-                        }
                         notifySnapshot()
                     }
 
                     override fun onMediaItemTransition(mediaItem: MediaItem?, reason: Int) {
+                        if (updatingPlaybackWindow) {
+                            notifySnapshot()
+                            return
+                        }
+                        if (clippedSentence == null) {
+                            baseItemIndex = player.currentMediaItemIndex
                             lastSentenceIndex = null
-                        cancelSentenceBoundaryGuard()
+                        }
                         listeners.forEach { it.onMediaChanged(currentTrainingItem()) }
                         notifySentenceIfChanged(force = true)
-                        refreshSentenceBoundaryGuard()
                         notifySnapshot()
                     }
 
@@ -93,7 +135,6 @@ class OralTrainerController internal constructor(
                         reason: Int,
                     ) {
                         notifySentenceIfChanged(force = true)
-                        refreshSentenceBoundaryGuard()
                         notifySnapshot()
                     }
 
@@ -140,9 +181,18 @@ class OralTrainerController internal constructor(
         lastSentenceIndex = null
         stopAtSentenceEnd = false
         stopAtSentenceEndIndex = null
-        player.setMediaItems(items.map { it.toMedia3Item() }, safeIndex, max(0L, startPositionMs))
+        cancelClipStartWarmup(restoreVolume = true)
+        cancelSentenceEndGuard()
+        baseItemIndex = safeIndex
+        clippedSentence = null
+        unclippedDurationMs = -1L
+        val startMs = max(0L, startPositionMs)
+        player.setMediaItems(items.map { it.toMedia3Item() }, safeIndex, startMs)
         player.prepare()
         player.playWhenReady = config.autoPlay
+        applyLoopMode()
+        lockSentenceAt(startMs)
+        syncPlaybackWindow(startMs)
         listeners.forEach { it.onMediaChanged(currentTrainingItem()) }
         notifySentenceIfChanged(force = true)
         notifySnapshot()
@@ -166,25 +216,32 @@ class OralTrainerController internal constructor(
             return true
         }
         mediaItems[index] = updatedItem
-        if (index == player.currentMediaItemIndex) {
+        if (index == baseItemIndex) {
             listeners.forEach { it.onMediaChanged(updatedItem) }
+            lockSentenceAt(playbackPositionMs())
+            syncPlaybackWindow()
             notifySentenceIfChanged(force = true)
-            refreshSentenceBoundaryGuard()
             notifySnapshot()
         }
         return true
     }
 
     fun play() {
+        syncPlaybackWindow()
+        if (clippedSentence != null && !player.isPlaying && shouldWarmClipStart()) {
+            playClippedSentenceFromStart()
+            return
+        }
         player.play()
-        refreshSentenceBoundaryGuard()
     }
 
     fun pause() {
         stopAtSentenceEnd = false
         stopAtSentenceEndIndex = null
-        cancelSentenceBoundaryGuard()
+        cancelClipStartWarmup(restoreVolume = true)
+        cancelSentenceEndGuard()
         player.pause()
+        applySentenceRepeatMode()
     }
 
     fun togglePlayPause() {
@@ -198,19 +255,26 @@ class OralTrainerController internal constructor(
     fun stop() {
         stopAtSentenceEnd = false
         stopAtSentenceEndIndex = null
-        cancelSentenceBoundaryGuard()
+        cancelClipStartWarmup(restoreVolume = true)
+        cancelSentenceEndGuard()
         player.stop()
     }
 
     fun seekTo(positionMs: Long) {
         stopAtSentenceEnd = false
         stopAtSentenceEndIndex = null
-        player.seekTo(max(0L, positionMs))
-        refreshSentenceBoundaryGuard()
+        cancelClipStartWarmup(restoreVolume = true)
+        cancelSentenceEndGuard()
+        val safePositionMs = max(0L, positionMs)
+        lockSentenceAt(safePositionMs)
+        syncPlaybackWindow(safePositionMs)
+        seekInCurrentWindow(safePositionMs)
+        notifySentenceIfChanged(force = true)
+        notifySnapshot()
     }
 
     fun seekRelative(deltaMs: Long) {
-        seekTo(player.currentPosition + deltaMs)
+        seekTo(playbackPositionMs() + deltaMs)
     }
 
     fun rewind(stepMs: Long = config.defaultSeekStepMs) {
@@ -222,11 +286,11 @@ class OralTrainerController internal constructor(
     }
 
     fun previousItem() {
-        player.seekToPreviousMediaItem()
+        switchBaseItem(baseItemIndex - 1)
     }
 
     fun nextItem() {
-        player.seekToNextMediaItem()
+        switchBaseItem(baseItemIndex + 1)
     }
 
     fun setPlaybackSpeed(speed: Float) {
@@ -240,20 +304,21 @@ class OralTrainerController internal constructor(
 
     fun setLoopMode(loopMode: LoopMode) {
         this.loopMode = loopMode
-        player.repeatMode = when (loopMode) {
-            LoopMode.OFF -> Player.REPEAT_MODE_OFF
-            LoopMode.ONE -> Player.REPEAT_MODE_ONE
-            LoopMode.ALL -> Player.REPEAT_MODE_ALL
-        }
+        applySentenceRepeatMode()
     }
 
     fun setSentenceMode(enabled: Boolean) {
         config = config.copy(sentenceMode = enabled)
+        syncPlaybackWindow()
     }
 
     fun setContinuousPlayback(enabled: Boolean) {
         config = config.copy(continuousPlayback = enabled)
-        refreshSentenceBoundaryGuard()
+        if (!enabled) {
+            lockSentenceAt(playbackPositionMs())
+        }
+        syncPlaybackWindow()
+        notifySentenceIfChanged(force = true)
     }
 
     fun playCurrentSentenceAndStop() {
@@ -263,24 +328,47 @@ class OralTrainerController internal constructor(
 
     /** Seek to a known sentence and stop automatically when that sentence ends. */
     fun playSentenceAndStop(startMs: Long, sentenceIndex: Int) {
+        val item = currentTrainingItem()
+        val sentence = item?.sentences?.getOrNull(sentenceIndex)
+        cancelClipStartWarmup(restoreVolume = true)
         stopAtSentenceEnd = true
         stopAtSentenceEndIndex = sentenceIndex
+        lastSentenceIndex = sentenceIndex
+        player.playWhenReady = false
+        if (item != null && sentence != null) {
+            if (!isCurrentClip(sentence)) {
+                installSentenceClip(item, sentence, 0L)
+                player.playWhenReady = false
+            }
+            playClippedSentenceFromStart()
+        } else {
+            cancelSentenceEndGuard()
             player.seekTo(max(0L, startMs))
+            player.playWhenReady = true
             player.play()
-        refreshSentenceBoundaryGuard()
+        }
+        notifySentenceIfChanged(force = true)
+        notifySnapshot()
     }
 
     fun currentTrainingItem(): TrainingMediaItem? {
-        val index = player.currentMediaItemIndex
-        return mediaItems.getOrNull(index)
+        return mediaItems.getOrNull(baseItemIndex)
     }
 
     fun currentSentence(): SentenceBoundary? {
         val item = currentTrainingItem() ?: return null
-        val position = player.currentPosition
-        return item.sentences.lastOrNull { sentence ->
-            position >= sentence.startMs && position < sentence.endMs
+        if (item.sentences.isEmpty()) {
+            return null
         }
+        clippedSentence?.let { window ->
+            return item.sentences.getOrNull(window.index)
+        }
+        if (shouldBoundCurrentSentence()) {
+            lastSentenceIndex?.let { index ->
+                item.sentences.getOrNull(index)?.let { return it }
+            }
+        }
+        return sentenceAt(item, playbackPositionMs())
     }
 
     fun seekToPreviousSentence(): Boolean {
@@ -288,11 +376,16 @@ class OralTrainerController internal constructor(
         if (item.sentences.isEmpty()) {
             return false
         }
-        val position = player.currentPosition
-        val target = item.sentences
+        val target = if (shouldBoundCurrentSentence()) {
+            val index = anchoredSentenceIndex(item) ?: 0
+            item.sentences.getOrNull((index - 1).coerceAtLeast(0)) ?: item.sentences.first()
+        } else {
+            val position = playbackPositionMs()
+            item.sentences
                 .asReversed()
                 .firstOrNull { it.startMs < position - PREVIOUS_SENTENCE_TOLERANCE_MS }
                 ?: item.sentences.first()
+        }
         seekTo(target.startMs)
         return true
     }
@@ -302,10 +395,15 @@ class OralTrainerController internal constructor(
         if (item.sentences.isEmpty()) {
             return false
         }
-        val position = player.currentPosition
-        val target = item.sentences.firstOrNull {
+        val target = if (shouldBoundCurrentSentence()) {
+            val index = anchoredSentenceIndex(item) ?: return false
+            item.sentences.getOrNull(index + 1) ?: return false
+        } else {
+            val position = playbackPositionMs()
+            item.sentences.firstOrNull {
                 it.startMs > position + NEXT_SENTENCE_TOLERANCE_MS
             } ?: return false
+        }
         seekTo(target.startMs)
         return true
     }
@@ -362,12 +460,16 @@ class OralTrainerController internal constructor(
 
     fun snapshot(): PlaybackSnapshot {
         val item = currentTrainingItem()
-        val duration = player.duration.takeUnless { it == C.TIME_UNSET } ?: -1L
+        val duration = when {
+            unclippedDurationMs > 0L -> unclippedDurationMs
+            clippedSentence != null -> clippedSentence?.endMs ?: -1L
+            else -> player.duration.takeUnless { it == C.TIME_UNSET } ?: -1L
+        }
         return PlaybackSnapshot(
             mediaId = item?.id,
-            positionMs = max(0L, player.currentPosition),
+            positionMs = playbackPositionMs(),
             durationMs = duration,
-            bufferedPositionMs = max(0L, player.bufferedPosition),
+            bufferedPositionMs = playbackBufferedPositionMs(),
             isPlaying = player.isPlaying,
             playbackState = player.playbackState.toOralTrainerState(),
             playbackSpeed = player.playbackParameters.speed,
@@ -380,7 +482,8 @@ class OralTrainerController internal constructor(
             return
         }
         released = true
-        cancelSentenceBoundaryGuard()
+        cancelClipStartWarmup(restoreVolume = false)
+        cancelSentenceEndGuard()
         mainHandler.removeCallbacks(ticker)
         listeners.clear()
         player.release()
@@ -400,127 +503,295 @@ class OralTrainerController internal constructor(
         if (!force && index == lastSentenceIndex) {
             return
         }
-        val changed = index != lastSentenceIndex
         lastSentenceIndex = index
         listeners.forEach { it.onSentenceChanged(sentence) }
-        if (changed && !config.continuousPlayback && !stopAtSentenceEnd) {
-            refreshSentenceBoundaryGuard()
-        }
     }
 
-    private fun enforceSentenceBoundary() {
-        if (!config.sentenceMode || !player.isPlaying) {
+    private fun syncPlaybackWindow(positionMs: Long = playbackPositionMs()) {
+        if (released || updatingPlaybackWindow) {
             return
         }
-        val item = currentTrainingItem() ?: return
-        if (item.sentences.isEmpty()) {
+        val item = currentTrainingItem()
+        val target = sentenceForWindow(item)
+        if (target == null || item == null) {
+            restoreUnclipped(positionMs)
             return
         }
-        val position = player.currentPosition
-        if (stopAtSentenceEnd) {
-            val index = stopAtSentenceEndIndex ?: return
-            val sentence = item.sentences.getOrNull(index) ?: return
-            if (position >= sentence.endMs) {
-                applySentenceEndAction(sentence)
-            }
+        val current = clippedSentence
+        if (current != null &&
+            current.index == target.index &&
+            current.startMs == target.startMs &&
+            current.endMs == target.endMs
+        ) {
+            applySentenceRepeatMode()
             return
         }
-        if (config.continuousPlayback) {
-            return
-        }
-        val anchorIndex = lastSentenceIndex ?: currentSentence()?.index ?: return
-        val anchor = item.sentences.getOrNull(anchorIndex) ?: return
-        if (position >= anchor.endMs) {
-            applySentenceEndAction(anchor)
-        }
+        val relativeMs = (positionMs - target.startMs).coerceAtLeast(0L)
+        installSentenceClip(item, target, relativeMs)
     }
 
-    private fun refreshSentenceBoundaryGuard() {
-        cancelSentenceBoundaryGuard()
-        if (released || !config.sentenceMode) {
-            return
+    private fun sentenceForWindow(item: TrainingMediaItem?): SentenceBoundary? {
+        if (item == null || !config.sentenceMode || item.sentences.isEmpty()) {
+            return null
         }
-        if (!player.isPlaying && !player.playWhenReady) {
-            return
-        }
-        val item = currentTrainingItem() ?: return
-        if (item.sentences.isEmpty()) {
-            return
-        }
-
-        val sentence = when {
-            stopAtSentenceEnd -> item.sentences.getOrNull(stopAtSentenceEndIndex ?: return)
+        return when {
+            stopAtSentenceEnd -> item.sentences.getOrNull(stopAtSentenceEndIndex ?: return null)
             config.continuousPlayback -> null
             else -> {
-                val index = lastSentenceIndex ?: currentSentence()?.index
-                item.sentences.getOrNull(index ?: return)
+                val index = lastSentenceIndex ?: sentenceAt(item, playbackPositionMs())?.index
+                item.sentences.getOrNull(index ?: return null)
+            }
+        }
     }
-        } ?: return
 
-        val mediaItemIndex = player.currentMediaItemIndex
-        if (player.currentPosition >= sentence.endMs) {
-            applySentenceEndAction(sentence)
+    private fun installSentenceClip(
+        item: TrainingMediaItem,
+        sentence: SentenceBoundary,
+        relativePositionMs: Long,
+    ) {
+        if (released) {
             return
         }
+        rememberUnclippedDuration()
+        cancelSentenceEndGuard()
+        val lengthMs = (sentence.endMs - sentence.startMs).coerceAtLeast(0L)
+        val startPositionMs = relativePositionMs.coerceIn(0L, lengthMs)
+        val playWhenReady = player.playWhenReady
+        val warmFromStart = startPositionMs <= CLIP_START_WARMUP_POSITION_MS
+        clipStartNeedsWarmup = warmFromStart
+        updatingPlaybackWindow = true
+        clippedSentence = sentence
+        lastSentenceIndex = sentence.index
+        player.setMediaItem(item.toMedia3Item(sentence.startMs, sentence.endMs), startPositionMs)
+        player.prepare()
+        player.playWhenReady = false
+        applySentenceRepeatMode()
+        updatingPlaybackWindow = false
+        if (playWhenReady) {
+            if (warmFromStart) {
+                playClippedSentenceFromStart()
+            } else {
+                player.playWhenReady = true
+                player.play()
+            }
+        }
+    }
 
-        val token = ++boundaryGuardToken
-        boundaryMessage = player.createMessage { _, _ ->
-            if (token != boundaryGuardToken || released) {
-                return@createMessage
+    private fun restoreUnclipped(positionMs: Long) {
+        if (released) {
+            return
         }
-            val active = currentTrainingItem() ?: return@createMessage
-            val latest = when {
-                stopAtSentenceEnd -> active.sentences.getOrNull(stopAtSentenceEndIndex ?: return@createMessage)
-                else -> active.sentences.getOrNull(sentence.index)
-            } ?: return@createMessage
-            if (player.currentMediaItemIndex != mediaItemIndex) {
-                return@createMessage
+        if (clippedSentence == null && player.mediaItemCount == mediaItems.size) {
+            applyLoopMode()
+            return
         }
-            applySentenceEndAction(latest)
+        val playWhenReady = player.playWhenReady
+        updatingPlaybackWindow = true
+        clippedSentence = null
+        clipStartNeedsWarmup = false
+        player.setMediaItems(
+            mediaItems.map { it.toMedia3Item() },
+            baseItemIndex.coerceIn(0, mediaItems.lastIndex),
+            max(0L, positionMs),
+        )
+        player.prepare()
+        player.playWhenReady = playWhenReady
+        applyLoopMode()
+        updatingPlaybackWindow = false
     }
-            .setLooper(mainHandler.looper)
-            .setPosition(mediaItemIndex, sentence.endMs)
+
+    private fun seekInCurrentWindow(positionMs: Long) {
+        val sentence = clippedSentence
+        if (sentence == null) {
+            player.seekTo(max(0L, positionMs))
+            return
+        }
+        val lengthMs = (sentence.endMs - sentence.startMs).coerceAtLeast(0L)
+        player.seekTo((positionMs - sentence.startMs).coerceIn(0L, lengthMs))
+    }
+
+    private fun onClippedSentenceEnded() {
+        val sentence = clippedSentence ?: return
+        if (stopAtSentenceEnd && stopAtSentenceEndIndex == sentence.index) {
+            handleStopAtSentenceEnd(sentence)
+            return
+        }
+        clipStartNeedsWarmup = false
+        player.seekTo(0L)
+        if (player.playbackState == Player.STATE_ENDED) {
+            player.prepare()
+        }
+        if (!config.continuousPlayback) {
+            player.play()
+        }
+        lastSentenceIndex = sentence.index
+        notifySentenceIfChanged(force = true)
+        notifySnapshot()
+    }
+
+    private fun enforceStopAtSentenceEnd() {
+        if (!stopAtSentenceEnd) {
+            return
+        }
+        val index = stopAtSentenceEndIndex ?: return
+        val sentence = currentTrainingItem()?.sentences?.getOrNull(index) ?: return
+        if (playbackPositionMs() >= sentence.endMs) {
+            handleStopAtSentenceEnd(sentence)
+        }
+    }
+
+    private fun handleStopAtSentenceEnd(sentence: SentenceBoundary) {
+        if (!stopAtSentenceEnd || stopAtSentenceEndIndex != sentence.index) {
+            return
+        }
+        cancelSentenceEndGuard()
+        stopAtSentenceEnd = false
+        stopAtSentenceEndIndex = null
+        player.pause()
+        if (clippedSentence?.index == sentence.index) {
+            player.seekTo(0L)
+            if (player.playbackState == Player.STATE_ENDED) {
+                player.prepare()
+            }
+        } else {
+            seekInCurrentWindow(sentence.startMs)
+        }
+        applySentenceRepeatMode()
+        lastSentenceIndex = sentence.index
+        notifySentenceIfChanged(force = true)
+        notifySnapshot()
+    }
+
+    private fun scheduleSentenceEndGuard() {
+        val index = stopAtSentenceEndIndex ?: return
+        val sentence = currentTrainingItem()?.sentences?.getOrNull(index) ?: return
+        cancelSentenceEndGuard()
+        val endInWindowMs = if (clippedSentence?.index == sentence.index) {
+            (sentence.endMs - sentence.startMs - 1L).coerceAtLeast(0L)
+        } else {
+            sentence.endMs
+        }
+        sentenceEndMessage = player.createMessage { _, _ ->
+            if (stopAtSentenceEnd && stopAtSentenceEndIndex == sentence.index) {
+                handleStopAtSentenceEnd(sentence)
+            }
+        }.setLooper(mainHandler.looper)
+            .setPosition(endInWindowMs)
             .setDeleteAfterDelivery(true)
             .send()
     }
 
-    private fun cancelSentenceBoundaryGuard() {
-        boundaryGuardToken++
-        boundaryMessage?.let { message ->
-            runCatching { message.cancel() }
-        }
-        boundaryMessage = null
+    private fun cancelSentenceEndGuard() {
+        sentenceEndMessage?.cancel()
+        sentenceEndMessage = null
     }
 
-    private fun applySentenceEndAction(sentence: SentenceBoundary) {
-        if (stopAtSentenceEnd && stopAtSentenceEndIndex == sentence.index) {
+    private fun switchBaseItem(targetIndex: Int) {
+        if (targetIndex !in mediaItems.indices) {
+            return
+        }
+        baseItemIndex = targetIndex
+        lastSentenceIndex = null
         stopAtSentenceEnd = false
         stopAtSentenceEndIndex = null
-            cancelSentenceBoundaryGuard()
-            player.pause()
-            player.seekTo(sentence.startMs)
-            lastSentenceIndex = sentence.index
-            notifySentenceIfChanged(force = true)
-            notifySnapshot()
-            return
+        cancelClipStartWarmup(restoreVolume = true)
+        cancelSentenceEndGuard()
+        clippedSentence = null
+        unclippedDurationMs = -1L
+        val item = mediaItems[targetIndex]
+        val firstSentence = item.sentences.firstOrNull()
+        if (shouldBoundCurrentSentence() && firstSentence != null) {
+            lastSentenceIndex = firstSentence.index
+            installSentenceClip(item, firstSentence, 0L)
+        } else {
+            player.setMediaItem(item.toMedia3Item(), 0L)
+            player.prepare()
+            applyLoopMode()
         }
-        if (config.continuousPlayback) {
-            return
-        }
-        // Loop the current sentence without leaking into the next one.
-        player.seekTo(sentence.startMs)
-        lastSentenceIndex = sentence.index
+        listeners.forEach { it.onMediaChanged(item) }
         notifySentenceIfChanged(force = true)
-        refreshSentenceBoundaryGuard()
         notifySnapshot()
     }
 
-    private fun TrainingMediaItem.toMedia3Item(): MediaItem {
+    private fun playbackPositionMs(): Long {
+        val sentence = clippedSentence ?: return max(0L, player.currentPosition)
+        val lengthMs = (sentence.endMs - sentence.startMs).coerceAtLeast(0L)
+        return sentence.startMs + player.currentPosition.coerceIn(0L, lengthMs)
+    }
+
+    private fun playbackBufferedPositionMs(): Long {
+        val sentence = clippedSentence ?: return max(0L, player.bufferedPosition)
+        val lengthMs = (sentence.endMs - sentence.startMs).coerceAtLeast(0L)
+        return sentence.startMs + player.bufferedPosition.coerceIn(0L, lengthMs)
+    }
+
+    private fun shouldBoundCurrentSentence(): Boolean {
+        return config.sentenceMode && (!config.continuousPlayback || stopAtSentenceEnd)
+    }
+
+    private fun lockSentenceAt(positionMs: Long) {
+        if (!shouldBoundCurrentSentence()) {
+            return
+        }
+        val item = currentTrainingItem() ?: return
+        lastSentenceIndex = sentenceAt(item, positionMs)?.index
+    }
+
+    private fun anchoredSentenceIndex(item: TrainingMediaItem): Int? {
+        return lastSentenceIndex?.takeIf { it in item.sentences.indices }
+            ?: clippedSentence?.index
+            ?: sentenceAt(item, playbackPositionMs())?.index
+    }
+
+    private fun sentenceAt(item: TrainingMediaItem, positionMs: Long): SentenceBoundary? {
+        return item.sentences.lastOrNull { positionMs >= it.startMs && positionMs < it.endMs }
+            ?: item.sentences.lastOrNull { it.startMs <= positionMs }
+    }
+
+    private fun rememberUnclippedDuration() {
+        if (unclippedDurationMs > 0L || clippedSentence != null) {
+            return
+        }
+        val duration = player.duration
+        if (duration != C.TIME_UNSET) {
+            unclippedDurationMs = duration
+        }
+    }
+
+    private fun applyLoopMode() {
+        player.repeatMode = when (loopMode) {
+            LoopMode.OFF -> Player.REPEAT_MODE_OFF
+            LoopMode.ONE -> Player.REPEAT_MODE_ONE
+            LoopMode.ALL -> Player.REPEAT_MODE_ALL
+        }
+    }
+
+    private fun applySentenceRepeatMode() {
+        if (clippedSentence == null) {
+            applyLoopMode()
+            return
+        }
+        player.repeatMode = Player.REPEAT_MODE_OFF
+    }
+
+    private fun TrainingMediaItem.toMedia3Item(
+        clippingStartPositionMs: Long = 0L,
+        clippingEndPositionMs: Long? = null,
+    ): MediaItem {
         val builder = MediaItem.Builder()
             .setMediaId(id)
             .setUri(uri)
         mimeType?.let(builder::setMimeType)
         customCacheKey?.let(builder::setCustomCacheKey)
+        if (clippingEndPositionMs != null) {
+            builder.setClippingConfiguration(
+                MediaItem.ClippingConfiguration.Builder()
+                    .setStartPositionMs(max(0L, clippingStartPositionMs))
+                    .setEndPositionMs(max(0L, clippingEndPositionMs))
+                    .setStartsAtKeyFrame(false)
+                    .build()
+            )
+        }
         subtitleUri?.let { uri ->
             builder.setSubtitleConfigurations(
                 listOf(
@@ -549,8 +820,93 @@ class OralTrainerController internal constructor(
         return CancellableAssessment {}
     }
 
+    private fun isCurrentClip(sentence: SentenceBoundary): Boolean {
+        val current = clippedSentence ?: return false
+        return current.index == sentence.index &&
+            current.startMs == sentence.startMs &&
+            current.endMs == sentence.endMs
+    }
+
+    private fun shouldWarmClipStart(): Boolean {
+        if (clippedSentence == null) {
+            return false
+        }
+        return clipStartNeedsWarmup || player.currentPosition <= CLIP_START_WARMUP_POSITION_MS
+    }
+
+    private fun playClippedSentenceFromStart() {
+        if (released) {
+            return
+        }
+        if (player.playbackState == Player.STATE_IDLE ||
+            player.playbackState == Player.STATE_ENDED ||
+            player.playbackState == Player.STATE_BUFFERING
+        ) {
+            playFromStartAfterReady = true
+            player.playWhenReady = false
+            if (player.playbackState == Player.STATE_IDLE || player.playbackState == Player.STATE_ENDED) {
+                player.prepare()
+            }
+            return
+        }
+        startClipStartWarmup()
+    }
+
+    private fun startClipStartWarmup() {
+        if (released) {
+            return
+        }
+        mainHandler.removeCallbacks(finishClipStartWarmupRunnable)
+        if (clipStartWarmupVolume == null) {
+            clipStartWarmupVolume = player.volume
+        }
+        player.volume = 0f
+        player.seekTo(0L)
+        if (player.playbackState == Player.STATE_ENDED) {
+            playFromStartAfterReady = true
+            player.prepare()
+            return
+        }
+        player.playWhenReady = true
+        player.play()
+        mainHandler.postDelayed(finishClipStartWarmupRunnable, CLIP_START_WARMUP_MS)
+    }
+
+    private fun finishClipStartWarmup() {
+        if (released) {
+            return
+        }
+        val restoreVolume = clipStartWarmupVolume ?: return
+        clipStartWarmupVolume = null
+        clipStartNeedsWarmup = false
+        mainHandler.removeCallbacks(finishClipStartWarmupRunnable)
+        player.seekTo(0L)
+        if (player.playbackState == Player.STATE_ENDED) {
+            player.prepare()
+        }
+        player.volume = restoreVolume
+        player.playWhenReady = true
+        player.play()
+        if (stopAtSentenceEnd) {
+            scheduleSentenceEndGuard()
+        }
+    }
+
+    private fun cancelClipStartWarmup(restoreVolume: Boolean) {
+        playFromStartAfterReady = false
+        mainHandler.removeCallbacks(finishClipStartWarmupRunnable)
+        val volume = clipStartWarmupVolume ?: return
+        clipStartWarmupVolume = null
+        if (restoreVolume) {
+            player.volume = volume
+        }
+    }
+
     private companion object {
         const val PREVIOUS_SENTENCE_TOLERANCE_MS = 250L
         const val NEXT_SENTENCE_TOLERANCE_MS = 150L
+        const val IDLE_POLL_MS = 250L
+        const val CLIP_START_WARMUP_MS = 180L
+        const val CLIP_START_WARMUP_POSITION_MS = 120L
     }
 }

android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/OralTrainerPlayerView.kt

@@ -4,6 +4,7 @@ import android.content.Context
 import android.util.AttributeSet
 import android.view.GestureDetector
 import android.view.MotionEvent
+import android.view.View
 import android.view.ViewConfiguration
 import android.widget.FrameLayout
 import androidx.media3.ui.PlayerView
@@ -16,6 +17,7 @@ class OralTrainerPlayerView @JvmOverloads constructor(
 ) : FrameLayout(context, attrs, defStyleAttr) {
     val playerView: PlayerView = PlayerView(context).apply {
         useController = false
+        setKeepContentOnPlayerReset(true)
         layoutParams = LayoutParams(LayoutParams.MATCH_PARENT, LayoutParams.MATCH_PARENT)
     }
 
@@ -122,6 +124,26 @@ class OralTrainerPlayerView @JvmOverloads constructor(
         verticalSwipeListener = listener
     }
 
+    /** Forwards touches from an overlay (for example a subtitle layer) as player gestures. */
+    fun forwardGestureFrom(source: View, event: MotionEvent): Boolean {
+        if (controller == null || !gestureControls.enabled) {
+            return false
+        }
+        val sourceLocation = IntArray(2).also(source::getLocationOnScreen)
+        val playerLocation = IntArray(2).also(::getLocationOnScreen)
+        val playerEvent = MotionEvent.obtain(event).apply {
+            offsetLocation(
+                (playerLocation[0] - sourceLocation[0]).toFloat(),
+                (playerLocation[1] - sourceLocation[1]).toFloat(),
+            )
+        }
+        return try {
+            handleLongPressSpeed(playerEvent) || gestureDetector.onTouchEvent(playerEvent)
+        } finally {
+            playerEvent.recycle()
+        }
+    }
+
     override fun onTouchEvent(event: MotionEvent): Boolean {
         return gestureDetector.onTouchEvent(event) || super.onTouchEvent(event)
     }

android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/SentenceBoundaryApi.kt

@@ -159,6 +159,11 @@ class SentenceBoundaryApi internal constructor(
                         } else {
                             sentence.getLong("reference_speech_duration_ms")
                         },
+                        zhText = if (sentence.isNull("zh_text")) {
+                            null
+                        } else {
+                            sentence.getString("zh_text").takeIf { it.isNotBlank() }
+                        },
                     )
                 )
             }

android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/TrainingModels.kt

@@ -10,6 +10,7 @@ data class SentenceBoundary @JvmOverloads constructor(
     val text: String? = null,
     val language: String? = null,
     val referenceSpeechDurationMs: Long? = null,
+    val zhText: String? = null,
 ) {
     init {
         require(index >= 0) { "Sentence index must be non-negative." }

android/sample-app/build.gradle.kts

@@ -19,7 +19,7 @@ android {
         minSdk = 26
         targetSdk = 35
         versionCode = 1
-        versionName = "0.1.4"
+        versionName = "0.1.6"
     }
 
     compileOptions {
@@ -37,6 +37,7 @@ kotlin {
 dependencies {
     implementation(project(":oral-trainer-sdk"))
     implementation("androidx.core:core-ktx:1.13.1")
+    implementation("androidx.recyclerview:recyclerview:1.3.2")
     implementation("androidx.media3:media3-common:1.11.0")
     implementation("androidx.media3:media3-exoplayer:1.11.0")
 }

android/sample-app/src/main/AndroidManifest.xml

@@ -1,6 +1,9 @@
 <manifest xmlns:android="http://schemas.android.com/apk/res/android">
     <uses-permission android:name="android.permission.INTERNET" />
     <uses-permission android:name="android.permission.RECORD_AUDIO" />
+    <uses-permission android:name="android.permission.FOREGROUND_SERVICE" />
+    <uses-permission android:name="android.permission.FOREGROUND_SERVICE_MEDIA_PLAYBACK" />
+    <uses-permission android:name="android.permission.POST_NOTIFICATIONS" />
 
     <application
         android:allowBackup="true"
@@ -19,6 +22,11 @@
             </intent-filter>
         </activity>
 
+        <service
+            android:name=".TrainingPlaybackService"
+            android:exported="false"
+            android:foregroundServiceType="mediaPlayback" />
+
         <provider
             android:name="androidx.core.content.FileProvider"
             android:authorities="${applicationId}.fileprovider"

android/sample-app/src/main/java/cn/learningpad/oraltrainer/sample/AppConfig.kt

@@ -3,6 +3,7 @@ package cn.learningpad.oraltrainer.sample
 enum class Module {
     COURSE,
     TRAIN,
+    READING,
     TEST,
     MINE,
 }
@@ -13,7 +14,6 @@ internal const val ASSESSMENT_API_KEY =
 
 internal const val SSO_LOGIN_URL = "https://d1kt.cn/login"
 
-internal const val PICK_VIDEO_REQUEST = 1001
 internal const val RECORD_AUDIO_REQUEST = 1002
 internal const val PROGRESS_MAX = 1000
 

android/sample-app/src/main/java/cn/learningpad/oraltrainer/sample/CourseCatalogController.kt

@@ -2,6 +2,7 @@ package cn.learningpad.oraltrainer.sample
 
 import cn.learningpad.oraltrainer.sdk.OralTrainerController
 import cn.learningpad.oraltrainer.sdk.OralTrainerSdk
+import cn.learningpad.oraltrainer.sdk.SentenceBoundary
 import cn.learningpad.oraltrainer.sdk.SentenceBoundaryApiCallback
 import cn.learningpad.oraltrainer.sdk.SentenceBoundaryApiResult
 import cn.learningpad.oraltrainer.sdk.TrainingVideoSummary
@@ -26,23 +27,45 @@ internal class CourseCatalogController(
         fun onSentencePlaceholder(meta: String, text: String)
         fun refreshPlayerUi()
         fun openTrainingModule()
+        fun syncSentenceExplanations(videoHash: String, sentences: List<SentenceBoundary>)
     }
 
     fun loadCatalog() {
+        loadCatalog(refreshBoundaries = false)
+    }
+
+    fun refreshCourses() {
+        if (screenState.catalogIsRefreshing) return
+        loadCatalog(refreshBoundaries = true)
+    }
+
+    private fun loadCatalog(refreshBoundaries: Boolean) {
+        val generation = screenState.beginCatalogRequest()
+        screenState.catalogIsRefreshing = refreshBoundaries
         screenState.catalogStatus = "正在同步"
         host.onCatalogStatus(screenState.catalogStatus)
         host.onCatalogLoadingPlaceholder()
         sdk.userApi.courses(tokenProvider()) { result ->
+            if (!screenState.isCurrentCatalogRequest(generation)) return@courses
             result.fold(
                 onSuccess = { courses ->
                     screenState.catalogVideos = courses.map { it.toVideoSummary() }
                     screenState.courseEnrollment.clear()
                     courses.forEach { screenState.courseEnrollment[it.videoHash] = it.enrolled }
+                    if (refreshBoundaries) {
+                        screenState.catalogStatus = "${courses.size} 个课程，正在刷新句界"
+                        host.onCatalogStatus(screenState.catalogStatus)
+                        host.onCatalogChanged()
+                        refreshAllCourseBoundaries(generation, index = 0, succeeded = 0, failed = 0)
+                    } else {
+                        screenState.catalogIsRefreshing = false
                         screenState.catalogStatus = "${screenState.catalogVideos.size} 个课程"
                         host.onCatalogStatus(screenState.catalogStatus)
                         host.onCatalogChanged()
+                    }
                 },
                 onFailure = {
+                    screenState.catalogIsRefreshing = false
                     screenState.catalogStatus = "课程加载失败：${it.message}"
                     host.onCatalogStatus(screenState.catalogStatus)
                     screenState.courseEnrollment.clear()
@@ -52,6 +75,68 @@ internal class CourseCatalogController(
         }
     }
 
+    private fun refreshAllCourseBoundaries(
+        generation: Int,
+        index: Int,
+        succeeded: Int,
+        failed: Int,
+    ) {
+        if (!screenState.isCurrentCatalogRequest(generation)) return
+        val videos = screenState.catalogVideos
+        if (index >= videos.size) {
+            screenState.catalogIsRefreshing = false
+            screenState.catalogStatus = when {
+                failed == 0 -> "${videos.size} 个课程 · 句界已刷新"
+                succeeded == 0 -> "${videos.size} 个课程 · 句界刷新失败"
+                else -> "${videos.size} 个课程 · 句界部分刷新（失败 $failed）"
+            }
+            host.onCatalogStatus(screenState.catalogStatus)
+            host.onCatalogChanged()
+            return
+        }
+
+        val video = videos[index]
+        screenState.catalogStatus =
+            "正在刷新句界 ${index + 1}/${videos.size}"
+        host.onCatalogStatus(screenState.catalogStatus)
+        sdk.sentenceBoundaryApi.fetch(video.videoHash, object : SentenceBoundaryApiCallback {
+            override fun onSuccess(result: SentenceBoundaryApiResult) {
+                if (!screenState.isCurrentCatalogRequest(generation)) return
+                screenState.courseBoundaries[result.videoHash] = result
+                host.syncSentenceExplanations(result.videoHash, result.sentences)
+                updateCatalogVideo(video.videoHash, result)
+                if (screenState.activeItemId == result.videoHash) {
+                    controller.updateSentences(result.videoHash, result.sentences)
+                    screenState.currentSentenceCount = result.sentences.size
+                    host.refreshPlayerUi()
+                }
+                host.onCatalogChanged()
+                refreshAllCourseBoundaries(generation, index + 1, succeeded + 1, failed)
+            }
+
+            override fun onError(error: Throwable) {
+                if (!screenState.isCurrentCatalogRequest(generation)) return
+                refreshAllCourseBoundaries(generation, index + 1, succeeded, failed + 1)
+            }
+        })
+    }
+
+    private fun updateCatalogVideo(videoHash: String, result: SentenceBoundaryApiResult) {
+        screenState.catalogVideos = screenState.catalogVideos.map { video ->
+            if (video.videoHash != videoHash) {
+                video
+            } else {
+                video.copy(
+                    durationMs = result.durationMs,
+                    sentenceCount = result.sentences.size,
+                )
+            }
+        }
+        if (screenState.currentCourse?.videoHash == videoHash) {
+            screenState.currentCourse = screenState.catalogVideos.firstOrNull { it.videoHash == videoHash }
+        }
+    }
+
     fun loadUserData(target: Module, onMine: () -> Unit) {
         when (target) {
             Module.COURSE -> loadCatalog()
@@ -85,23 +170,24 @@ internal class CourseCatalogController(
         }
     }
 
-    fun useCourseInPlace(video: TrainingVideoSummary) {
-        // Keep the current module (e.g. Train) while syncing course selection.
-        if (screenState.courseEnrollment[video.videoHash] != true) {
-            enrollCourse(video)
-        }
-        loadRemoteVideo(video)
-    }
-
     fun loadRemoteVideo(video: TrainingVideoSummary) {
         screenState.currentCourse = video
         screenState.activeItemId = video.videoHash
         host.onPlayerLoading(video)
         host.onCatalogChanged()
 
-        controller.loadItem(video.toTrainingMediaItem())
-        screenState.currentSentenceCount = video.sentenceCount
+        val cachedBoundary = screenState.courseBoundaries[video.videoHash]
+        if (cachedBoundary != null) {
+            host.syncSentenceExplanations(cachedBoundary.videoHash, cachedBoundary.sentences)
+        }
+        controller.loadItem(video.toTrainingMediaItem(cachedBoundary?.sentences.orEmpty()))
+        screenState.currentSentenceCount = cachedBoundary?.sentences?.size ?: video.sentenceCount
+        if (cachedBoundary == null) {
             refreshSentenceBoundaries(video, background = false)
+        } else {
+            host.refreshPlayerUi()
+            host.onStatus("课程已就绪：${cachedBoundary.sentences.size} 句")
+        }
     }
 
     fun refreshSentenceBoundaries(
@@ -117,8 +203,12 @@ internal class CourseCatalogController(
                 if (!screenState.isCurrentBoundaryRequest(requestGeneration, video.videoHash)) {
                     return
                 }
+                screenState.courseBoundaries[result.videoHash] = result
+                host.syncSentenceExplanations(result.videoHash, result.sentences)
+                updateCatalogVideo(video.videoHash, result)
                 screenState.currentSentenceCount = result.sentences.size
                 controller.updateSentences(video.videoHash, result.sentences)
+                host.onCatalogChanged()
                 host.refreshPlayerUi()
                 host.onStatus("课程已就绪：${result.sentences.size} 句")
             }

android/sample-app/src/main/java/cn/learningpad/oraltrainer/sample/MainActivity.kt

@@ -7,7 +7,6 @@ import android.content.Intent
 import android.content.IntentFilter
 import android.content.pm.PackageManager
 import android.content.res.Configuration
-import android.net.Uri
 import android.os.Bundle
 import cn.learningpad.oraltrainer.sdk.AuthSession
 import cn.learningpad.oraltrainer.sdk.GestureEvent
@@ -17,8 +16,6 @@ import cn.learningpad.oraltrainer.sdk.OralTrainerListener
 import cn.learningpad.oraltrainer.sdk.OralTrainerSdk
 import cn.learningpad.oraltrainer.sdk.PlaybackSnapshot
 import cn.learningpad.oraltrainer.sdk.SentenceBoundary
-import cn.learningpad.oraltrainer.sdk.SentenceBoundaryApiCallback
-import cn.learningpad.oraltrainer.sdk.SentenceBoundaryApiResult
 import cn.learningpad.oraltrainer.sdk.TrainingMediaItem
 import cn.learningpad.oraltrainer.sdk.TrainingVideoSummary
 
@@ -43,16 +40,24 @@ class MainActivity : Activity() {
     private lateinit var catalogController: CourseCatalogController
     private lateinit var testSession: TestSessionController
     private lateinit var explanationOverlay: SentenceExplanationOverlay
+    private lateinit var translationApi: SubtitleTranslationApi
+    private lateinit var sentenceFavoriteStore: SentenceFavoriteStore
+    private var translationRequest: CancellableTranslationRequest? = null
 
     private var progressBarDragging = false
+    private var trainingPlaybackServiceActive = false
 
     private val screenActionReceiver = object : BroadcastReceiver() {
         override fun onReceive(context: Context?, intent: Intent?) {
             if (intent?.action == Intent.ACTION_SCREEN_OFF) {
+                // Training playback intentionally keeps running with the screen off.
+                // The test module still pauses because recording/assessment state is UI-dependent.
+                if (screenState.activeModule == Module.TEST) {
                     testSession.onScreenOff()
                 }
             }
         }
+    }
 
     override fun onCreate(savedInstanceState: Bundle?) {
         super.onCreate(savedInstanceState)
@@ -73,6 +78,10 @@ class MainActivity : Activity() {
                 authSessionStore.session?.token.orEmpty()
             },
         )
+        translationApi = SubtitleTranslationApi(this, SERVER_BASE_URL) {
+            authSessionStore.session?.token.orEmpty()
+        }
+        sentenceFavoriteStore = SentenceFavoriteStore(this)
 
         screenFactory = SampleScreenFactory(
             activity = this,
@@ -147,8 +156,9 @@ class MainActivity : Activity() {
             if (bindings.hasPlayerChrome) {
                 bindings.statusText.text = "已停止配音播放"
             }
-        } else {
+        } else if (screenState.activeModule != Module.TRAIN) {
             controller.pause()
+            stopTrainingPlaybackService()
         }
         super.onPause()
     }
@@ -173,58 +183,14 @@ class MainActivity : Activity() {
         super.onBackPressed()
     }
 
-    override fun onActivityResult(requestCode: Int, resultCode: Int, data: Intent?) {
-        super.onActivityResult(requestCode, resultCode, data)
-        if (requestCode != PICK_VIDEO_REQUEST || resultCode != RESULT_OK) return
-        val uri = data?.data ?: return
-        val persistableFlags = data.flags and Intent.FLAG_GRANT_READ_URI_PERMISSION
-        if (persistableFlags != 0) {
-            runCatching { contentResolver.takePersistableUriPermission(uri, persistableFlags) }
-        }
-
-        val item = TrainingMediaItem(
-            id = uri.toString(),
-            title = uri.lastPathSegment ?: "本地视频",
-            uri = uri,
-        )
-        screenState.activeItemId = item.id
-        screenState.currentSentenceCount = 0
-        controller.loadItem(item)
-        if (bindings.hasPlayerChrome) {
-            bindings.lessonTitleText.text = item.title
-            bindings.sentenceMetaText.text = "句子边界分析中"
-            bindings.sentenceText.text = "正在匹配服务端句子边界"
-            bindings.statusText.text = "正在获取本地视频的句子边界..."
-        }
-        screenFactory.renderCatalog()
-
-        sdk.sentenceBoundaryApi.fetchForUri(uri, contentResolver, object : SentenceBoundaryApiCallback {
-            override fun onSuccess(result: SentenceBoundaryApiResult) {
-                if (screenState.activeItemId != item.id) return
-                screenState.currentSentenceCount = result.sentences.size
-                controller.loadItem(item.copy(sentences = result.sentences))
-                if (bindings.hasPlayerChrome) {
-                    bindings.statusText.text = "句子边界已加载：${result.sentences.size} 句"
-                }
-            }
-
-            override fun onError(error: Throwable) {
-                if (screenState.activeItemId != item.id) return
-                if (bindings.hasPlayerChrome) {
-                    bindings.statusText.text =
-                        "句子边界获取失败，使用默认快退/快进：${error.message.orEmpty()}"
-                    bindings.sentenceMetaText.text = "暂无句子边界"
-                    bindings.sentenceText.text = ""
-                }
-            }
-        })
-    }
-
     override fun onDestroy() {
         authCoordinator.reset()
         testSession.release()
+        stopTrainingPlaybackService()
         controller.release()
         explanationOverlay.release()
+        translationRequest?.cancel()
+        translationApi.release()
         super.onDestroy()
     }
 
@@ -264,28 +230,20 @@ class MainActivity : Activity() {
         sdk.config.userToken = null
         screenState.clearCourse()
         controller.pause()
+        stopTrainingPlaybackService()
         screenState.activeModule = Module.COURSE
         setRootView()
     }
 
-    private fun openLocalVideoPicker() {
-        @Suppress("DEPRECATION")
-        startActivityForResult(
-            Intent(Intent.ACTION_OPEN_DOCUMENT).apply {
-                addCategory(Intent.CATEGORY_OPENABLE)
-                type = "video/*"
-                addFlags(Intent.FLAG_GRANT_READ_URI_PERMISSION)
-                addFlags(Intent.FLAG_GRANT_PERSISTABLE_URI_PERMISSION)
-            },
-            PICK_VIDEO_REQUEST,
-        )
-    }
-
     private fun switchModule(module: Module) {
         if (module != Module.TEST) {
             testSession.stopDubbingIfNeeded()
         }
         screenState.activeModule = module
+        if (module != Module.TRAIN) {
+            controller.pause()
+            stopTrainingPlaybackService()
+        }
         if (module == Module.COURSE || module == Module.MINE) {
             catalogController.loadUserData(module) {
                 screenFactory.refreshMyContent(sdk.userApi, authSessionStore.session?.token.orEmpty())
@@ -305,10 +263,23 @@ class MainActivity : Activity() {
         if (screenState.activeModule == Module.MINE) {
             screenFactory.refreshMyContent(sdk.userApi, authSessionStore.session?.token.orEmpty())
         }
+        if (screenState.activeModule == Module.READING) {
+            screenFactory.scrollToReadingCurrentSentence()
+        }
     }
 
     private fun refreshCurrentUi() {
-        if (screenState.activeModule != Module.TRAIN && screenState.activeModule != Module.TEST) {
+        if (screenState.activeModule != Module.TRAIN &&
+            screenState.activeModule != Module.TEST &&
+            screenState.activeModule != Module.READING
+        ) {
+            screenFactory.setCatalogStatus(screenState.catalogStatus)
+            return
+        }
+        if (screenState.activeModule == Module.READING) {
+            screenState.currentSentenceCount = controller.currentTrainingItem()?.sentences?.size ?: 0
+            screenFactory.refreshReadingSentences()
+            ensureChineseTranslations()
             screenFactory.setCatalogStatus(screenState.catalogStatus)
             return
         }
@@ -318,6 +289,7 @@ class MainActivity : Activity() {
         screenState.currentSentenceCount = item?.sentences?.size ?: 0
         applyPlaybackSnapshot(controller.snapshot())
         applySentence(controller.currentSentence())
+        ensureChineseTranslations()
         screenFactory.setCatalogStatus(screenState.catalogStatus)
     }
 
@@ -327,9 +299,13 @@ class MainActivity : Activity() {
         }
 
         override fun onMediaChanged(item: TrainingMediaItem?) {
+            screenState.currentSentenceCount = item?.sentences?.size ?: 0
             if (bindings.hasPlayerChrome) {
                 bindings.lessonTitleText.text = item?.title ?: "未选择课程"
-                screenState.currentSentenceCount = item?.sentences?.size ?: 0
+            }
+            ensureChineseTranslations()
+            if (screenState.activeModule == Module.READING) {
+                screenFactory.refreshReadingSentences()
             }
             testSession.ensureHistoryForCurrentCourse()
             if (screenState.activeModule == Module.TEST) {
@@ -359,6 +335,11 @@ class MainActivity : Activity() {
 
     private fun applyPlaybackSnapshot(snapshot: PlaybackSnapshot) {
         if (!bindings.hasPlayerChrome) return
+        if (screenState.activeModule == Module.TRAIN && snapshot.isPlaying) {
+            startTrainingPlaybackService()
+        } else if (!snapshot.isPlaying) {
+            stopTrainingPlaybackService()
+        }
         if (!progressBarDragging) {
             bindings.timeText.text =
                 "${PlaybackFormat.time(snapshot.positionMs)} / ${PlaybackFormat.time(snapshot.durationMs)}"
@@ -368,12 +349,25 @@ class MainActivity : Activity() {
         bindings.statusText.text = PlaybackFormat.status(snapshot)
     }
 
+    private fun startTrainingPlaybackService() {
+        if (trainingPlaybackServiceActive) return
+        TrainingPlaybackService.start(this)
+        trainingPlaybackServiceActive = true
+    }
+
+    private fun stopTrainingPlaybackService() {
+        if (!trainingPlaybackServiceActive) return
+        TrainingPlaybackService.stop(this)
+        trainingPlaybackServiceActive = false
+    }
+
     private fun applySentence(sentence: SentenceBoundary?) {
         if (!bindings.hasPlayerChrome) return
         if (sentence == null) {
             bindings.sentenceMetaText.text = "暂无句子边界"
             bindings.sentenceText.text = ""
             bindings.explainSentenceButton?.isEnabled = false
+            applyChineseSubtitle(null)
             return
         }
         val countText = if (screenState.currentSentenceCount > 0) {
@@ -385,18 +379,113 @@ class MainActivity : Activity() {
             "第 ${sentence.index + 1}$countText 句  ${PlaybackFormat.time(sentence.startMs)}-${PlaybackFormat.time(sentence.endMs)}"
         bindings.sentenceText.text = sentence.text?.takeIf { it.isNotBlank() } ?: "当前句子"
         bindings.explainSentenceButton?.isEnabled = !sentence.text.isNullOrBlank()
+        applyChineseSubtitle(sentence)
         if (screenState.activeModule == Module.TEST) {
             testSession.refreshTestUi()
         }
     }
 
+    private fun applyChineseSubtitle(sentence: SentenceBoundary?) {
+        if (!bindings.hasPlayerChrome) return
+        val showChinese = screenState.activeModule == Module.TRAIN &&
+            screenState.chineseSubtitlesEnabled &&
+            sentence != null &&
+            !sentence.text.isNullOrBlank()
+        if (!showChinese || sentence == null) {
+            bindings.sentenceZhText.visibility = android.view.View.GONE
+            return
+        }
+        bindings.sentenceZhText.visibility = android.view.View.VISIBLE
+        bindings.sentenceZhText.text = when {
+            !sentence.zhText.isNullOrBlank() -> sentence.zhText
+            screenState.chineseTranslationLoading -> "正在翻译中文字幕…"
+            screenState.chineseTranslationError.isNotBlank() -> screenState.chineseTranslationError
+            else -> "暂无中文字幕"
+        }
+    }
+
+    private fun applyTranslatedSentences(itemId: String, translations: Map<Int, String>) {
+        if (translations.isEmpty()) return
+        val current = controller.currentTrainingItem() ?: return
+        if (current.id != itemId) return
+        val updated = current.sentences.map { sentence ->
+            sentence.copy(zhText = translations[sentence.index] ?: sentence.zhText)
+        }
+        controller.updateSentences(current.id, updated)
+        screenState.courseBoundaries[current.id]?.let { cached ->
+            screenState.courseBoundaries[current.id] = cached.copy(sentences = updated)
+        }
+        applyChineseSubtitle(controller.currentSentence())
+        screenFactory.refreshReadingTranslations()
+    }
+
+    private fun favoriteUserKey(): String =
+        SentenceFavoriteStore.userKey(authSessionStore.session?.token.orEmpty())
+
+    private fun ensureChineseTranslations() {
+        val translationEnabled = when (screenState.activeModule) {
+            Module.TRAIN -> screenState.chineseSubtitlesEnabled
+            Module.READING -> screenState.readingTranslationVisible
+            else -> false
+        }
+        if (!translationEnabled) {
+            return
+        }
+        val item = controller.currentTrainingItem() ?: return
+        val sentences = item.sentences
+        val needsTranslation = sentences.any { !it.text.isNullOrBlank() && it.zhText.isNullOrBlank() }
+        if (!needsTranslation) {
+            screenState.chineseTranslationLoading = false
+            screenState.chineseTranslationError = ""
+            applyChineseSubtitle(controller.currentSentence())
+            screenFactory.refreshReadingTranslations()
+            return
+        }
+        val generation = screenState.beginTranslationRequest()
+        screenState.chineseTranslationLoading = true
+        screenState.chineseTranslationError = ""
+        applyChineseSubtitle(controller.currentSentence())
+        screenFactory.refreshReadingTranslations()
+        translationRequest?.cancel()
+        val expectedIndexes = sentences.mapNotNull { sentence ->
+            sentence.index.takeIf { !sentence.text.isNullOrBlank() }
+        }
+        translationRequest = translationApi.translate(
+            item.id,
+            expectedIndexes,
+            object : SubtitleTranslationCallback {
+                override fun onProgress(result: SubtitleTranslationResult) {
+                    if (!screenState.isCurrentTranslationRequest(generation, item.id)) return
+                    applyTranslatedSentences(item.id, result.translations)
+                }
+
+                override fun onSuccess(result: SubtitleTranslationResult) {
+                    if (!screenState.isCurrentTranslationRequest(generation, item.id)) return
+                    applyTranslatedSentences(item.id, result.translations)
+                    screenState.chineseTranslationLoading = false
+                    screenState.chineseTranslationError = ""
+                    applyChineseSubtitle(controller.currentSentence())
+                    if (bindings.hasPlayerChrome) {
+                        bindings.statusText.text = "已加载中文字幕"
+                    }
+                }
+
+                override fun onError(error: Throwable) {
+                    if (!screenState.isCurrentTranslationRequest(generation, item.id)) return
+                    screenState.chineseTranslationLoading = false
+                    screenState.chineseTranslationError = "中文字幕翻译失败"
+                    applyChineseSubtitle(controller.currentSentence())
+                    if (bindings.hasPlayerChrome) {
+                        bindings.statusText.text = "中文字幕翻译失败：${error.message.orEmpty()}"
+                    }
+                }
+            },
+        )
+    }
+
     private fun screenActions() = object : SampleScreenActions {
         override fun onModuleSelected(module: Module) = switchModule(module)
 
-        override fun onImportLocalVideo() = openLocalVideoPicker()
-
-        override fun onRefreshCatalog() = catalogController.loadCatalog()
-
         override fun onSeekProgressPreview(progress: Int) {
             if (!bindings.hasPlayerChrome) return
             val duration = controller.snapshot().durationMs
@@ -427,13 +516,66 @@ class MainActivity : Activity() {
             }
         }
 
+        override fun onChineseSubtitlesChanged(enabled: Boolean) {
+            screenState.chineseSubtitlesEnabled = enabled
+            if (bindings.hasPlayerChrome) {
+                bindings.statusText.text = if (enabled) {
+                    "已开启中文字幕"
+                } else {
+                    "已关闭中文字幕"
+                }
+            }
+            if (enabled) {
+                ensureChineseTranslations()
+            } else {
+                translationRequest?.cancel()
+                screenState.chineseTranslationLoading = false
+                applyChineseSubtitle(controller.currentSentence())
+            }
+        }
+
+        override fun onReadingTranslationVisible(visible: Boolean) {
+            screenState.readingTranslationVisible = visible
+            if (!visible) {
+                translationRequest?.cancel()
+                screenState.chineseTranslationLoading = false
+            } else {
+                ensureChineseTranslations()
+            }
+            screenFactory.refreshReadingTranslations()
+        }
+
+        override fun isSentenceFavorite(sentence: SentenceBoundary): Boolean {
+            val item = controller.currentTrainingItem() ?: return false
+            return sentence.index in sentenceFavoriteStore.indexes(favoriteUserKey(), item.id)
+        }
+
+        override fun toggleSentenceFavorite(sentence: SentenceBoundary): Boolean {
+            val item = controller.currentTrainingItem() ?: return false
+            return sentenceFavoriteStore.toggle(favoriteUserKey(), item.id, sentence.index)
+        }
+
+        override fun onSentenceDoubleTapped(sentence: SentenceBoundary) {
+            screenState.continuousPlaybackEnabled = false
+            switchModule(Module.TRAIN)
+            controller.setContinuousPlayback(false)
+            controller.seekTo(sentence.startMs)
+            controller.play()
+            if (bindings.hasPlayerChrome) {
+                bindings.statusText.text = "已定位到第 ${sentence.index + 1} 句，单句循环播放中"
+            }
+        }
+
         override fun onExplainCurrentSentence() {
             val sentence = controller.currentSentence() ?: return
             if (sentence.text.isNullOrBlank()) return
             controller.pause()
+            val item = controller.currentTrainingItem()
             explanationOverlay.show(
                 sentence,
-                controller.currentTrainingItem()?.title.orEmpty(),
+                item?.title.orEmpty(),
+                item?.id.orEmpty(),
+                item?.sentences.orEmpty(),
             )
         }
 
@@ -449,8 +591,8 @@ class MainActivity : Activity() {
 
         override fun onReplaySentence() {
             testSession.stopDubbingIfNeeded()
-            testSession.refreshTestUi()
             controller.playCurrentSentenceAndStop()
+            testSession.refreshTestUi()
         }
 
         override fun onNextSentence() {
@@ -471,14 +613,18 @@ class MainActivity : Activity() {
             catalogController.selectCourse(video)
         }
 
-        override fun onTrainCatalogVideoClick(video: TrainingVideoSummary) {
-            catalogController.useCourseInPlace(video)
-        }
+        override fun onRefreshCourses() = catalogController.refreshCourses()
 
         override fun onLogout() = logout()
 
         override fun nickname(): String =
             authSessionStore.session?.user?.nickname.orEmpty()
+
+        override fun favoriteSentences(): List<SentenceBoundary> {
+            val item = controller.currentTrainingItem() ?: return emptyList()
+            val indexes = sentenceFavoriteStore.indexes(favoriteUserKey(), item.id)
+            return item.sentences.filter { it.index in indexes }
+        }
     }
 
     private fun catalogHost() = object : CourseCatalogController.Host {
@@ -513,6 +659,13 @@ class MainActivity : Activity() {
             screenFactory.renderCatalog()
             refreshCurrentUi()
         }
+
+        override fun syncSentenceExplanations(
+            videoHash: String,
+            sentences: List<SentenceBoundary>,
+        ) {
+            SentenceExplanationStore(this@MainActivity).sync(videoHash, sentences)
+        }
     }
 
     private fun testHost() = object : TestSessionController.Host {

android/sample-app/src/main/java/cn/learningpad/oraltrainer/sample/MarkdownText.kt

@@ -0,0 +1,130 @@
+package cn.learningpad.oraltrainer.sample
+
+import android.widget.TextView
+import androidx.core.text.HtmlCompat
+
+/** Renders common LLM Markdown into a TextView without extra libraries. */
+internal object MarkdownText {
+    private val heading = Regex("^(#{1,6})\\s*(.*)$")
+    private val unordered = Regex("^[-*+]\\s+(.*)$")
+    private val ordered = Regex("^\\d+[.)]\\s+(.*)$")
+    private val boldItalic = Regex("\\*\\*\\*(.+?)\\*\\*\\*")
+    private val bold = Regex("(\\*\\*|__)(.+?)\\1")
+    private val italic = Regex("(?<!\\*)\\*(?!\\*)(.+?)(?<!\\*)\\*(?!\\*)")
+    private val italicUnderscore = Regex("(?<!\\w)_(?!_)(.+?)(?<!_)_(?!\\w)")
+    private val inlineCode = Regex("`([^`]+)`")
+    private val image = Regex("!\\[([^\\]]*)]\\(([^)]+)\\)")
+    private val link = Regex("\\[([^\\]]+)]\\(([^)]+)\\)")
+
+    fun set(view: TextView, markdown: String) {
+        view.text = render(markdown)
+    }
+
+    fun render(markdown: String): CharSequence {
+        return HtmlCompat.fromHtml(toHtml(markdown), HtmlCompat.FROM_HTML_MODE_COMPACT)
+    }
+
+    internal fun toHtml(markdown: String): String {
+        val normalized = markdown.replace("\r\n", "\n").replace('\r', '\n')
+        val html = StringBuilder()
+        var inFence = false
+        val code = StringBuilder()
+        var listTag: String? = null
+
+        fun closeList() {
+            listTag?.let { html.append("</$it>") }
+            listTag = null
+        }
+
+        fun openList(tag: String) {
+            if (listTag != tag) {
+                closeList()
+                listTag = tag
+                html.append('<').append(tag).append('>')
+            }
+        }
+
+        for (line in normalized.split('\n')) {
+            val trimmed = line.trim()
+            if (trimmed.startsWith("```")) {
+                if (inFence) {
+                    html.append("<pre><code>").append(code).append("</code></pre>")
+                    code.setLength(0)
+                    inFence = false
+                } else {
+                    closeList()
+                    inFence = true
+                }
+                continue
+            }
+            if (inFence) {
+                if (code.isNotEmpty()) code.append('\n')
+                code.append(escape(line))
+                continue
+            }
+            if (trimmed.isEmpty()) {
+                closeList()
+                continue
+            }
+
+            val headingMatch = heading.matchEntire(trimmed)
+            if (headingMatch != null) {
+                closeList()
+                val level = headingMatch.groupValues[1].length
+                html.append("<h$level>").append(inline(headingMatch.groupValues[2])).append("</h$level>")
+                continue
+            }
+            val unorderedMatch = unordered.matchEntire(trimmed)
+            if (unorderedMatch != null) {
+                openList("ul")
+                html.append("<li>").append(inline(unorderedMatch.groupValues[1])).append("</li>")
+                continue
+            }
+            val orderedMatch = ordered.matchEntire(trimmed)
+            if (orderedMatch != null) {
+                openList("ol")
+                html.append("<li>").append(inline(orderedMatch.groupValues[1])).append("</li>")
+                continue
+            }
+            closeList()
+            if (trimmed.startsWith(">")) {
+                html.append("<blockquote>")
+                    .append(inline(trimmed.removePrefix(">").trimStart()))
+                    .append("</blockquote>")
+            } else {
+                html.append("<p>").append(inline(trimmed)).append("</p>")
+            }
+        }
+        if (inFence) {
+            html.append("<pre><code>").append(code).append("</code></pre>")
+        }
+        closeList()
+        return html.toString()
+    }
+
+    private fun inline(text: String): String {
+        val codes = mutableListOf<String>()
+        var current = escape(text)
+        current = inlineCode.replace(current) { match ->
+            codes += "<code>${match.groupValues[1]}</code>"
+            "%%MDCODE${codes.lastIndex}%%"
+        }
+        current = image.replace(current) { match -> escape(match.groupValues[1]) }
+        current = link.replace(current) { match ->
+            """<a href="${match.groupValues[2]}">${match.groupValues[1]}</a>"""
+        }
+        current = boldItalic.replace(current) { match -> "<b><i>${match.groupValues[1]}</i></b>" }
+        current = bold.replace(current) { match -> "<b>${match.groupValues[2]}</b>" }
+        current = italic.replace(current) { match -> "<i>${match.groupValues[1]}</i>" }
+        current = italicUnderscore.replace(current) { match -> "<i>${match.groupValues[1]}</i>" }
+        codes.forEachIndexed { index, value ->
+            current = current.replace("%%MDCODE$index%%", value)
+        }
+        return current
+    }
+
+    private fun escape(text: String): String =
+        text.replace("&", "&amp;")
+            .replace("<", "&lt;")
+            .replace(">", "&gt;")
+}

android/sample-app/src/main/java/cn/learningpad/oraltrainer/sample/SampleBindings.kt

@@ -13,12 +13,14 @@ internal class SampleBindings {
     lateinit var lessonTitleText: TextView
     lateinit var statusText: TextView
     lateinit var sentenceText: TextView
+    lateinit var sentenceZhText: TextView
     lateinit var sentenceMetaText: TextView
     lateinit var timeText: TextView
     lateinit var speedText: TextView
     lateinit var progressBar: SeekBar
     lateinit var catalogList: LinearLayout
     lateinit var catalogStatusText: TextView
+    var catalogRefreshButton: Button? = null
     lateinit var testStatusText: TextView
     lateinit var scoreSummaryText: TextView
     lateinit var scoreDetailText: TextView
@@ -29,11 +31,13 @@ internal class SampleBindings {
     var dubbingStatusText: TextView? = null
     var explainSentenceButton: ImageButton? = null
     var myContentList: LinearLayout? = null
+    var favoriteSentenceList: LinearLayout? = null
 
     val hasPlayerChrome: Boolean
         get() = ::lessonTitleText.isInitialized &&
             ::statusText.isInitialized &&
             ::sentenceText.isInitialized &&
+            ::sentenceZhText.isInitialized &&
             ::sentenceMetaText.isInitialized &&
             ::timeText.isInitialized &&
             ::speedText.isInitialized &&

android/sample-app/src/main/java/cn/learningpad/oraltrainer/sample/SampleScreenActions.kt

@@ -1,5 +1,6 @@
 package cn.learningpad.oraltrainer.sample
 
+import cn.learningpad.oraltrainer.sdk.SentenceBoundary
 import cn.learningpad.oraltrainer.sdk.TrainingVideoSummary
 
 internal interface ModuleTabActions {
@@ -8,6 +9,7 @@ internal interface ModuleTabActions {
 
 internal interface CourseScreenActions {
     fun onSelectCourse(video: TrainingVideoSummary)
+    fun onRefreshCourses()
 }
 
 internal interface SentenceExplainActions {
@@ -16,13 +18,18 @@ internal interface SentenceExplainActions {
 
 internal interface TrainScreenActions :
     SentenceExplainActions {
-    fun onImportLocalVideo()
-    fun onRefreshCatalog()
     fun onSeekProgressPreview(progress: Int)
     fun onSeekStart()
     fun onSeekStop(progress: Int)
     fun onContinuousPlaybackChanged(enabled: Boolean)
-    fun onTrainCatalogVideoClick(video: TrainingVideoSummary)
+    fun onChineseSubtitlesChanged(enabled: Boolean)
+}
+
+internal interface ReadingScreenActions {
+    fun onReadingTranslationVisible(visible: Boolean)
+    fun isSentenceFavorite(sentence: SentenceBoundary): Boolean
+    fun toggleSentenceFavorite(sentence: SentenceBoundary): Boolean
+    fun onSentenceDoubleTapped(sentence: SentenceBoundary)
 }
 
 internal interface TestScreenActions :
@@ -43,11 +50,14 @@ internal interface TestScreenActions :
 internal interface MineScreenActions {
     fun onLogout()
     fun nickname(): String
+    fun favoriteSentences(): List<SentenceBoundary>
+    fun toggleSentenceFavorite(sentence: SentenceBoundary): Boolean
 }
 
 internal interface SampleScreenActions :
     ModuleTabActions,
     CourseScreenActions,
     TrainScreenActions,
+    ReadingScreenActions,
     TestScreenActions,
     MineScreenActions

android/sample-app/src/main/java/cn/learningpad/oraltrainer/sample/SampleScreenFactory.kt

@@ -7,6 +7,7 @@ import android.widget.LinearLayout
 import cn.learningpad.oraltrainer.sample.ui.CourseScreen
 import cn.learningpad.oraltrainer.sample.ui.MineScreen
 import cn.learningpad.oraltrainer.sample.ui.PlayerChrome
+import cn.learningpad.oraltrainer.sample.ui.ReadingScreen
 import cn.learningpad.oraltrainer.sample.ui.SampleWidgets
 import cn.learningpad.oraltrainer.sample.ui.TestScreen
 import cn.learningpad.oraltrainer.sample.ui.TrainScreen
@@ -33,13 +34,16 @@ internal class SampleScreenFactory(
     private val courseScreen = CourseScreen(activity, bindings, screenState, widgets, actions)
     private val trainScreen = TrainScreen(
         activity = activity,
-        bindings = bindings,
-        screenState = screenState,
-        widgets = widgets,
         playerChrome = playerChrome,
-        actions = actions,
     )
     private val testScreen = TestScreen(activity, bindings, screenState, playerChrome, actions)
+    private val readingScreen = ReadingScreen(
+        activity = activity,
+        screenState = screenState,
+        controller = controller,
+        widgets = widgets,
+        actions = actions,
+    )
     private val mineScreen = MineScreen(activity, bindings, screenState, widgets, actions)
 
     fun createContentView(): View {
@@ -53,36 +57,54 @@ internal class SampleScreenFactory(
 
         return LinearLayout(activity).apply {
             orientation = LinearLayout.VERTICAL
+            layoutParams = LinearLayout.LayoutParams(
+                LinearLayout.LayoutParams.MATCH_PARENT,
+                LinearLayout.LayoutParams.MATCH_PARENT,
+            )
             setBackgroundColor(SampleTheme.background)
             addView(widgets.moduleTabBar(actions))
             when (screenState.activeModule) {
                 Module.COURSE -> addView(courseScreen.createSection())
                 Module.MINE -> addView(mineScreen.createSection())
                 Module.TRAIN -> trainScreen.attachPortraitSections(this)
+                Module.READING -> addView(readingScreen.createSection())
                 Module.TEST -> testScreen.attachSections(this)
             }
         }
     }
 
+    fun refreshReadingSentences() {
+        if (screenState.activeModule == Module.READING) {
+            readingScreen.renderSentences()
+        }
+    }
+
+    fun refreshReadingTranslations() {
+        if (screenState.activeModule == Module.READING) {
+            readingScreen.refreshTranslations()
+        }
+    }
+
+    fun scrollToReadingCurrentSentence() {
+        if (screenState.activeModule == Module.READING) {
+            readingScreen.scrollToCurrentSentence()
+        }
+    }
+
     fun renderCatalog() {
-        when (screenState.activeModule) {
-            Module.COURSE -> courseScreen.renderCatalog()
-            Module.TRAIN -> trainScreen.renderCatalogStrip()
-            else -> Unit
+        if (screenState.activeModule == Module.COURSE) {
+            courseScreen.renderCatalog()
         }
     }
 
     fun showCatalogLoading() {
-        when (screenState.activeModule) {
-            Module.COURSE -> courseScreen.showLoading()
-            Module.TRAIN -> trainScreen.showCatalogLoading()
-            else -> Unit
+        if (screenState.activeModule == Module.COURSE) {
+            courseScreen.showLoading()
         }
     }
 
     fun setCatalogStatus(text: String) {
         courseScreen.setStatus(text)
-        trainScreen.setCatalogStatus(text)
     }
 
     fun refreshMyContent(userApi: UserApi, token: String) {

android/sample-app/src/main/java/cn/learningpad/oraltrainer/sample/SampleTheme.kt

@@ -19,6 +19,7 @@ internal object SampleTheme {
     val textDark = Color.rgb(18, 24, 31)
     val textMuted = Color.rgb(151, 162, 174)
     val textSubtle = Color.rgb(88, 98, 108)
+    val readingTranslation = Color.rgb(147, 197, 253)
     val danger = Color.rgb(220, 38, 38)
     val textOverlay = Color.rgb(163, 172, 182)
 }

android/sample-app/src/main/java/cn/learningpad/oraltrainer/sample/SentenceExplanationApi.kt

@@ -91,7 +91,7 @@ internal class SentenceExplanationApi(
             connection.useCaches = false
             connection.setRequestProperty("Accept", "application/json")
             connection.setRequestProperty("Content-Type", "application/json")
-            connection.setRequestProperty("User-Agent", "OralTrainerSample/0.1.4 (Android)")
+            connection.setRequestProperty("User-Agent", "OralTrainerSample/0.1.5 (Android)")
             tokenProvider().takeIf { it.isNotBlank() }?.let {
                 connection.setRequestProperty("Authorization", "Bearer $it")
             }

android/sample-app/src/main/java/cn/learningpad/oraltrainer/sample/SentenceExplanationOverlay.kt

@@ -18,11 +18,20 @@ internal class SentenceExplanationOverlay(
     private val activity: Activity,
     private val api: SentenceExplanationApi,
 ) {
+    private val store = SentenceExplanationStore(activity)
     private var currentOverlay: View? = null
     private var currentRequest: CancellableExplanationRequest? = null
     private var requestGeneration = 0
 
-    fun show(sentence: SentenceBoundary, courseTitle: String) {
+    fun show(
+        sentence: SentenceBoundary,
+        courseTitle: String,
+        mediaId: String,
+        sentences: List<SentenceBoundary> = emptyList(),
+    ) {
+        if (sentences.isNotEmpty()) {
+            store.sync(mediaId, sentences)
+        }
         close()
         val generation = ++requestGeneration
         val root = FrameLayout(activity).apply {
@@ -51,13 +60,38 @@ internal class SentenceExplanationOverlay(
             setTextColor(SampleTheme.textDark)
             textSize = 19f
             typeface = Typeface.DEFAULT_BOLD
+            layoutParams = LinearLayout.LayoutParams(0, ViewGroup.LayoutParams.WRAP_CONTENT, 1f)
+        }
+        val saveHint = TextView(activity).apply {
+            text = HINT_UNSAVED
+            setTextColor(SampleTheme.textSubtle)
+            textSize = 12f
+            setPadding(6.dp, 0, 0, 0)
+        }
+        val saveButton = TextView(activity).apply {
+            text = "保存"
+            setTextColor(Color.WHITE)
+            textSize = 13f
+            typeface = Typeface.DEFAULT_BOLD
+            gravity = Gravity.CENTER
+            background = rounded(SampleTheme.button, 8f)
+            setPadding(12.dp, 6.dp, 12.dp, 6.dp)
+            isEnabled = false
+            alpha = 0.45f
+        }
+        val titleRow = LinearLayout(activity).apply {
+            orientation = LinearLayout.HORIZONTAL
+            gravity = Gravity.CENTER_VERTICAL
+            addView(title)
+            addView(saveButton)
+            addView(saveHint)
         }
         val subtitle = TextView(activity).apply {
             text = sentence.text?.takeIf { it.isNotBlank() } ?: courseTitle
             setTextColor(SampleTheme.accentDeep)
             textSize = 14f
             setPadding(0, 8.dp, 0, 12.dp)
-            maxLines = 3
+            maxLines = 5
         }
         val message = TextView(activity).apply {
             text = "AI 正在讲解这句话…"
@@ -86,7 +120,7 @@ internal class SentenceExplanationOverlay(
             ).apply { topMargin = 16.dp }
             setOnClickListener { close() }
         }
-        card.addView(title)
+        card.addView(titleRow)
         card.addView(subtitle)
         card.addView(scroll)
         card.addView(close)
@@ -94,6 +128,21 @@ internal class SentenceExplanationOverlay(
 
         addToWindow(root)
         currentOverlay = root
+
+        val cached = store.load(mediaId, sentence.index)
+        if (!cached.isNullOrBlank()) {
+            showMarkdown(message, cached)
+            markSaved(saveButton, saveHint)
+            return
+        }
+
+        var loadedExplanation: String? = null
+        saveButton.setOnClickListener {
+            val explanation = loadedExplanation ?: return@setOnClickListener
+            if (store.save(mediaId, sentence.index, explanation)) {
+                markSaved(saveButton, saveHint)
+            }
+        }
         currentRequest = api.explain(
             sentence.text.orEmpty(),
             courseTitle,
@@ -101,14 +150,19 @@ internal class SentenceExplanationOverlay(
             sentence.language,
             object : SentenceExplanationCallback {
                 override fun onSuccess(result: SentenceExplanationResult) {
-                    if (generation == requestGeneration) message.text = result.explanation
+                    if (generation != requestGeneration) return
+                    loadedExplanation = result.explanation
+                    showMarkdown(message, result.explanation)
+                    saveButton.isEnabled = true
+                    saveButton.alpha = 1f
                 }
 
                 override fun onError(error: Throwable) {
-                    if (generation == requestGeneration) {
+                    if (generation != requestGeneration) return
                     message.text = "讲解失败：${error.message ?: "请稍后重试"}"
                     message.setTextColor(SampleTheme.danger)
-                    }
+                    saveButton.isEnabled = false
+                    saveButton.alpha = 0.45f
                 }
             },
         )
@@ -128,6 +182,18 @@ internal class SentenceExplanationOverlay(
         api.release()
     }
 
+    private fun showMarkdown(message: TextView, explanation: String) {
+        message.setTextColor(SampleTheme.textDark)
+        MarkdownText.set(message, explanation)
+    }
+
+    private fun markSaved(saveButton: TextView, saveHint: TextView) {
+        saveButton.text = "已保存"
+        saveButton.isEnabled = false
+        saveButton.alpha = 0.7f
+        saveHint.text = HINT_SAVED
+    }
+
     private fun addToWindow(view: View) {
         (activity.window?.decorView as? ViewGroup)?.addView(view)
     }
@@ -147,4 +213,9 @@ internal class SentenceExplanationOverlay(
         cornerRadius = radiusDp.dp.toFloat()
         strokeColor?.let { setStroke(1.dp, it) }
     }
+
+    private companion object {
+        const val HINT_UNSAVED = "（下次直接看本地）"
+        const val HINT_SAVED = "（已保存本地）"
+    }
 }

android/sample-app/src/main/java/cn/learningpad/oraltrainer/sample/SentenceExplanationStore.kt

@@ -0,0 +1,62 @@
+package cn.learningpad.oraltrainer.sample
+
+import android.content.Context
+import cn.learningpad.oraltrainer.sdk.SentenceBoundary
+import java.io.File
+import java.util.Locale
+
+/** Persists sentence explanations under app filesDir for reuse without another API call. */
+internal class SentenceExplanationStore(context: Context) {
+    private val rootDir = File(context.applicationContext.filesDir, "sentence-explanations")
+
+    fun load(mediaId: String, sentenceIndex: Int): String? {
+        val file = fileFor(mediaId, sentenceIndex) ?: return null
+        if (!file.isFile || file.length() <= 0L) return null
+        return file.readText()
+    }
+
+    fun save(mediaId: String, sentenceIndex: Int, explanation: String): Boolean {
+        val text = explanation.trim()
+        if (text.isEmpty()) return false
+        val file = fileFor(mediaId, sentenceIndex) ?: return false
+        file.parentFile?.mkdirs()
+        file.writeText(text)
+        return file.isFile && file.length() > 0L
+    }
+
+    fun sync(mediaId: String, sentences: List<SentenceBoundary>) {
+        if (sentences.isEmpty()) return
+        val directoryName = sanitize(mediaId) ?: return
+        val dir = File(rootDir, directoryName)
+        val revision = revisionFor(sentences)
+        val marker = File(dir, REVISION_FILE)
+        val previous = marker.takeIf { it.isFile }?.readText()?.trim().orEmpty()
+        if (previous == revision) return
+        if (dir.exists()) {
+            dir.deleteRecursively()
+        }
+        dir.mkdirs()
+        marker.writeText(revision)
+    }
+
+    private fun fileFor(mediaId: String, sentenceIndex: Int): File? {
+        val directoryName = sanitize(mediaId) ?: return null
+        return File(File(rootDir, directoryName), "$sentenceIndex.txt")
+    }
+
+    companion object {
+        private const val REVISION_FILE = "revision.txt"
+
+        fun revisionFor(sentences: List<SentenceBoundary>): String =
+            sentences.joinToString("\n") { sentence ->
+                "${sentence.index}\t${sentence.text.orEmpty().trim()}"
+            }
+
+        fun sanitize(mediaId: String): String? {
+            val filtered = mediaId.trim().lowercase(Locale.US)
+                .replace(Regex("[^a-z0-9._-]"), "_")
+                .trim('_')
+            return filtered.takeIf { it.isNotBlank() && it != "." && it != ".." }
+        }
+    }
+}

android/sample-app/src/main/java/cn/learningpad/oraltrainer/sample/SentenceFavoriteStore.kt

@@ -0,0 +1,55 @@
+package cn.learningpad.oraltrainer.sample
+
+import android.content.Context
+import java.security.MessageDigest
+
+/** Persists reading-mode favorites locally, scoped by signed-in user and video. */
+internal class SentenceFavoriteStore(context: Context) {
+    private val preferences =
+        context.getSharedPreferences(PREFERENCES_NAME, Context.MODE_PRIVATE)
+
+    fun toggle(userKey: String, mediaId: String, sentenceIndex: Int): Boolean {
+        val key = storageKey(userKey, mediaId) ?: return false
+        val favorites = readIndexes(key)
+        val isFavorite = sentenceIndex in favorites
+        val updated = if (isFavorite) {
+            favorites - sentenceIndex
+        } else {
+            favorites + sentenceIndex
+        }
+        preferences.edit()
+            .putString(key, updated.sorted().joinToString(SEPARATOR))
+            .apply()
+        return !isFavorite
+    }
+
+    fun indexes(userKey: String, mediaId: String): Set<Int> {
+        return storageKey(userKey, mediaId)?.let(::readIndexes).orEmpty()
+    }
+
+    private fun readIndexes(key: String): Set<Int> {
+        return preferences.getString(key, null).orEmpty()
+            .split(SEPARATOR)
+            .mapNotNull { it.trim().toIntOrNull() }
+            .toSet()
+    }
+
+    private fun storageKey(userKey: String, mediaId: String): String? {
+        val normalizedUser = userKey.trim()
+        val normalizedMedia = mediaId.trim()
+        if (normalizedUser.isEmpty() || normalizedMedia.isEmpty()) return null
+        return "$KEY_PREFIX$normalizedUser:$normalizedMedia"
+    }
+
+    companion object {
+        private const val PREFERENCES_NAME = "sentence_reading_favorites"
+        private const val KEY_PREFIX = "indexes"
+        private const val SEPARATOR = ","
+
+        fun userKey(token: String): String {
+            val digest = MessageDigest.getInstance("SHA-256")
+                .digest(token.toByteArray(Charsets.UTF_8))
+            return digest.joinToString("") { "%02x".format(it) }.take(24)
+        }
+    }
+}

android/sample-app/src/main/java/cn/learningpad/oraltrainer/sample/SubtitleTranslationApi.kt

@@ -0,0 +1,197 @@
+package cn.learningpad.oraltrainer.sample
+
+import android.content.Context
+import android.os.Handler
+import android.os.Looper
+import android.os.SystemClock
+import org.json.JSONObject
+import java.io.IOException
+import java.net.HttpURLConnection
+import java.net.SocketTimeoutException
+import java.net.URL
+import java.util.concurrent.ExecutorService
+import java.util.concurrent.Executors
+import java.util.concurrent.Future
+
+data class SubtitleTranslationResult(
+    val videoHash: String,
+    val translations: Map<Int, String>,
+    val complete: Boolean = false,
+    val status: String = "",
+    val error: String = "",
+)
+
+interface SubtitleTranslationCallback {
+    fun onProgress(result: SubtitleTranslationResult) {}
+    fun onSuccess(result: SubtitleTranslationResult)
+    fun onError(error: Throwable)
+}
+
+class CancellableTranslationRequest internal constructor(
+    private val future: Future<*>,
+) {
+    fun cancel() {
+        future.cancel(true)
+    }
+}
+
+/** Calls the app server; AI credentials never live in the mobile app. */
+internal class SubtitleTranslationApi(
+    context: Context,
+    private val baseUrl: String,
+    private val tokenProvider: () -> String,
+) {
+    private val appContext = context.applicationContext
+    private val mainHandler = Handler(Looper.getMainLooper())
+    private val executor: ExecutorService = Executors.newCachedThreadPool()
+
+    fun translate(
+        videoHash: String,
+        expectedIndexes: Collection<Int> = emptyList(),
+        callback: SubtitleTranslationCallback,
+    ): CancellableTranslationRequest {
+        val future = executor.submit {
+            try {
+                val result = loadBlocking(videoHash, expectedIndexes) { partial ->
+                    if (!Thread.currentThread().isInterrupted) {
+                        mainHandler.post { callback.onProgress(partial) }
+                    }
+                }
+                if (!Thread.currentThread().isInterrupted) {
+                    mainHandler.post { callback.onSuccess(result) }
+                }
+            } catch (error: Throwable) {
+                if (!Thread.currentThread().isInterrupted) {
+                    mainHandler.post { callback.onError(error) }
+                }
+            }
+        }
+        return CancellableTranslationRequest(future)
+    }
+
+    fun release() {
+        executor.shutdownNow()
+    }
+
+    private fun loadBlocking(
+        videoHash: String,
+        expectedIndexes: Collection<Int>,
+        onProgress: (SubtitleTranslationResult) -> Unit,
+    ): SubtitleTranslationResult {
+        val expected = expectedIndexes.toSet()
+        val deadline = SystemClock.elapsedRealtime() + TRANSLATION_DEADLINE_MS
+        var kickJob = true
+        var lastError: Throwable? = null
+        var lastCount = -1
+        var stalledRounds = 0
+        while (SystemClock.elapsedRealtime() < deadline) {
+            if (Thread.currentThread().isInterrupted) {
+                throw InterruptedException("Subtitle translation cancelled")
+            }
+            val stored = runCatching {
+                requestBlocking(videoHash, method = "GET", readTimeoutMs = 15_000)
+            }.onFailure { lastError = it }.getOrNull()
+            if (stored == null) {
+                kickJob = true
+            } else {
+                onProgress(stored)
+                if (isComplete(stored, expected)) {
+                    return stored
+                }
+                if (stored.translations.size > lastCount) {
+                    lastCount = stored.translations.size
+                    stalledRounds = 0
+                } else {
+                    stalledRounds += 1
+                }
+                if (stored.status != "running" && stalledRounds >= 2) {
+                    kickJob = true
+                }
+            }
+            if (kickJob) {
+                val posted = runCatching {
+                    requestBlocking(videoHash, method = "POST", readTimeoutMs = 30_000)
+                }.onFailure { error ->
+                    lastError = error
+                    if (error !is SocketTimeoutException) {
+                        stalledRounds += 1
+                    }
+                }.getOrNull()
+                kickJob = false
+                if (posted != null) {
+                    onProgress(posted)
+                    if (isComplete(posted, expected)) {
+                        return posted
+                    }
+                    if (posted.status == "error") {
+                        lastError = IOException(posted.error.ifBlank { "Subtitle translation failed" })
+                    }
+                }
+            }
+            Thread.sleep(POLL_INTERVAL_MS)
+        }
+        throw lastError ?: IOException("Subtitle translation timed out")
+    }
+
+    private fun isComplete(
+        result: SubtitleTranslationResult,
+        expected: Set<Int>,
+    ): Boolean {
+        if (expected.isNotEmpty()) {
+            return expected.all { result.translations.containsKey(it) }
+        }
+        return result.complete
+    }
+
+    private fun requestBlocking(
+        videoHash: String,
+        method: String,
+        readTimeoutMs: Int,
+    ): SubtitleTranslationResult {
+        val endpoint = baseUrl.trimEnd('/') + "/api/v1/videos/$videoHash/subtitle-translation"
+        val connection = URL(endpoint).openConnection() as HttpURLConnection
+        try {
+            connection.requestMethod = method
+            connection.connectTimeout = 15_000
+            connection.readTimeout = readTimeoutMs
+            connection.doOutput = false
+            connection.useCaches = false
+            connection.setRequestProperty("Accept", "application/json")
+            connection.setRequestProperty("User-Agent", "OralTrainerSample/0.1.5 (Android)")
+            tokenProvider().takeIf { it.isNotBlank() }?.let {
+                connection.setRequestProperty("Authorization", "Bearer $it")
+            }
+            val statusCode = connection.responseCode
+            val stream = if (statusCode in 200..299) connection.inputStream else connection.errorStream
+            val response = stream?.bufferedReader(Charsets.UTF_8)?.use { it.readText() }.orEmpty()
+            if (statusCode !in 200..299) {
+                throw IOException("Subtitle translation API returned HTTP $statusCode: $response")
+            }
+            val root = JSONObject(response)
+            val items = root.getJSONArray("translations")
+            val translations = buildMap {
+                for (index in 0 until items.length()) {
+                    val item = items.getJSONObject(index)
+                    val zh = item.optString("zh_text")
+                    if (zh.isNotBlank()) {
+                        put(item.getInt("index"), zh)
+                    }
+                }
+            }
+            return SubtitleTranslationResult(
+                videoHash = root.getString("video_hash"),
+                translations = translations,
+                complete = root.optBoolean("complete", false),
+                status = root.optString("status"),
+                error = root.optString("error"),
+            )
+        } finally {
+            connection.disconnect()
+        }
+    }
+
+    private companion object {
+        const val POLL_INTERVAL_MS = 1_500L
+        const val TRANSLATION_DEADLINE_MS = 15 * 60 * 1000L
+    }
+}

android/sample-app/src/main/java/cn/learningpad/oraltrainer/sample/TrainingPlaybackService.kt

@@ -0,0 +1,79 @@
+package cn.learningpad.oraltrainer.sample
+
+import android.app.Notification
+import android.app.NotificationChannel
+import android.app.NotificationManager
+import android.app.PendingIntent
+import android.app.Service
+import android.content.Context
+import android.content.Intent
+import android.content.pm.ServiceInfo
+import android.os.Build
+import android.os.IBinder
+
+/** Keeps the training player alive after the screen turns off or the app backgrounds. */
+class TrainingPlaybackService : Service() {
+
+    override fun onBind(intent: Intent?): IBinder? = null
+
+    override fun onCreate() {
+        super.onCreate()
+        val manager = getSystemService(NOTIFICATION_SERVICE) as NotificationManager
+        manager.createNotificationChannel(
+            NotificationChannel(
+                CHANNEL_ID,
+                getString(R.string.training_playback_channel_name),
+                NotificationManager.IMPORTANCE_LOW,
+            ),
+        )
+    }
+
+    override fun onStartCommand(intent: Intent?, flags: Int, startId: Int): Int {
+        val notification = buildNotification()
+        if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.Q) {
+            startForeground(
+                NOTIFICATION_ID,
+                notification,
+                ServiceInfo.FOREGROUND_SERVICE_TYPE_MEDIA_PLAYBACK,
+            )
+        } else {
+            startForeground(NOTIFICATION_ID, notification)
+        }
+        return START_NOT_STICKY
+    }
+
+    private fun buildNotification(): Notification {
+        val contentIntent = PendingIntent.getActivity(
+            this,
+            0,
+            Intent(this, MainActivity::class.java).addFlags(Intent.FLAG_ACTIVITY_SINGLE_TOP),
+            PendingIntent.FLAG_IMMUTABLE,
+        )
+        return Notification.Builder(this, CHANNEL_ID)
+            .setContentTitle(getString(R.string.training_playback_notification_title))
+            .setContentText(getString(R.string.training_playback_notification_text))
+            .setSmallIcon(R.drawable.ic_notification_play)
+            .setContentIntent(contentIntent)
+            .setOngoing(true)
+            .setCategory(Notification.CATEGORY_TRANSPORT)
+            .build()
+    }
+
+    companion object {
+        private const val CHANNEL_ID = "training_playback"
+        private const val NOTIFICATION_ID = 1001
+
+        fun start(context: Context) {
+            val intent = Intent(context, TrainingPlaybackService::class.java)
+            if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.O) {
+                context.startForegroundService(intent)
+            } else {
+                context.startService(intent)
+            }
+        }
+
+        fun stop(context: Context) {
+            context.stopService(Intent(context, TrainingPlaybackService::class.java))
+        }
+    }
+}

android/sample-app/src/main/java/cn/learningpad/oraltrainer/sample/TrainingScreenState.kt

@@ -1,5 +1,6 @@
 package cn.learningpad.oraltrainer.sample
 
+import cn.learningpad.oraltrainer.sdk.SentenceBoundaryApiResult
 import cn.learningpad.oraltrainer.sdk.TrainingVideoSummary
 
 internal class TrainingScreenState {
@@ -7,23 +8,46 @@ internal class TrainingScreenState {
     var currentSentenceCount = 0
     var catalogVideos: List<TrainingVideoSummary> = emptyList()
     val courseEnrollment = mutableMapOf<String, Boolean>()
+    val courseBoundaries = mutableMapOf<String, SentenceBoundaryApiResult>()
     var catalogStatus = "正在同步"
+    var catalogIsRefreshing = false
     var activeModule = Module.COURSE
     var currentCourse: TrainingVideoSummary? = null
     var continuousPlaybackEnabled = false
+    var chineseSubtitlesEnabled = false
+    var readingTranslationVisible = true
+    var chineseTranslationLoading = false
+    var chineseTranslationError = ""
     var testSubtitleVisible = false
     var landscapeSubtitleVisible = true
 
     private var boundaryRequestGeneration = 0
+    private var catalogRequestGeneration = 0
+    private var translationRequestGeneration = 0
 
     fun beginBoundaryRequest(): Int = ++boundaryRequestGeneration
 
     fun isCurrentBoundaryRequest(generation: Int, itemId: String): Boolean =
         boundaryRequestGeneration == generation && activeItemId == itemId
 
+    fun beginCatalogRequest(): Int = ++catalogRequestGeneration
+
+    fun isCurrentCatalogRequest(generation: Int): Boolean =
+        catalogRequestGeneration == generation
+
+    fun beginTranslationRequest(): Int = ++translationRequestGeneration
+
+    fun isCurrentTranslationRequest(generation: Int, itemId: String): Boolean =
+        translationRequestGeneration == generation && activeItemId == itemId
+
     fun clearCourse() {
         currentCourse = null
         activeItemId = ""
         currentSentenceCount = 0
+        catalogVideos = emptyList()
+        courseEnrollment.clear()
+        courseBoundaries.clear()
+        catalogStatus = "正在同步"
+        catalogIsRefreshing = false
     }
 }

android/sample-app/src/main/java/cn/learningpad/oraltrainer/sample/ui/CourseScreen.kt

@@ -3,6 +3,7 @@ package cn.learningpad.oraltrainer.sample.ui
 import android.app.Activity
 import android.graphics.Color
 import android.graphics.Typeface
+import android.view.Gravity
 import android.view.View
 import android.view.ViewGroup
 import android.widget.LinearLayout
@@ -34,18 +35,31 @@ internal class CourseScreen(
             orientation = LinearLayout.VERTICAL
             setPadding(16.dp, 8.dp, 16.dp, 16.dp)
         }
-        return ScrollView(activity).apply {
-            layoutParams = LinearLayout.LayoutParams(ViewGroup.LayoutParams.MATCH_PARENT, 0, 1f)
-            addView(LinearLayout(activity).apply {
-                orientation = LinearLayout.VERTICAL
-                setPadding(0.dp, 12.dp, 0.dp, 16.dp)
+        val header = LinearLayout(activity).apply {
+            orientation = LinearLayout.HORIZONTAL
+            gravity = Gravity.CENTER_VERTICAL
+            setPadding(16.dp, 0.dp, 16.dp, 0.dp)
             addView(TextView(activity).apply {
                 text = "课程首页"
                 textSize = 26f
                 typeface = Typeface.DEFAULT_BOLD
                 setTextColor(Color.WHITE)
-                    setPadding(16.dp, 0.dp, 16.dp, 0.dp)
+                layoutParams = LinearLayout.LayoutParams(0, ViewGroup.LayoutParams.WRAP_CONTENT, 1f)
             })
+            bindings.catalogRefreshButton = widgets.headerButton("刷新课程") {
+                actions.onRefreshCourses()
+            }.also { button ->
+                button.isEnabled = !screenState.catalogIsRefreshing
+                button.alpha = if (screenState.catalogIsRefreshing) 0.55f else 1f
+                addView(button)
+            }
+        }
+        return ScrollView(activity).apply {
+            layoutParams = LinearLayout.LayoutParams(ViewGroup.LayoutParams.MATCH_PARENT, 0, 1f)
+            addView(LinearLayout(activity).apply {
+                orientation = LinearLayout.VERTICAL
+                setPadding(0.dp, 12.dp, 0.dp, 16.dp)
+                addView(header)
                 addView(TextView(activity).apply {
                     text = "选择一门课程开始训练"
                     textSize = 13f
@@ -83,5 +97,10 @@ internal class CourseScreen(
         if (bindings.hasCatalogChrome) {
             bindings.catalogStatusText.text = text
         }
+        bindings.catalogRefreshButton?.let { button ->
+            button.isEnabled = !screenState.catalogIsRefreshing
+            button.alpha = if (screenState.catalogIsRefreshing) 0.55f else 1f
+            button.text = if (screenState.catalogIsRefreshing) "刷新中" else "刷新课程"
+        }
     }
 }

android/sample-app/src/main/java/cn/learningpad/oraltrainer/sample/ui/MineScreen.kt

@@ -16,6 +16,7 @@ import cn.learningpad.oraltrainer.sample.SampleBindings
 import cn.learningpad.oraltrainer.sample.SampleTheme
 import cn.learningpad.oraltrainer.sample.TrainingScreenState
 import cn.learningpad.oraltrainer.sample.UiStyled
+import cn.learningpad.oraltrainer.sdk.SentenceBoundary
 import cn.learningpad.oraltrainer.sdk.UserApi
 
 internal class MineScreen(
@@ -33,6 +34,9 @@ internal class MineScreen(
         bindings.myContentList = LinearLayout(activity).apply {
             orientation = LinearLayout.VERTICAL
         }
+        bindings.favoriteSentenceList = LinearLayout(activity).apply {
+            orientation = LinearLayout.VERTICAL
+        }
         return ScrollView(activity).apply {
             layoutParams = LinearLayout.LayoutParams(ViewGroup.LayoutParams.MATCH_PARENT, 0, 1f)
             addView(LinearLayout(activity).apply {
@@ -52,6 +56,13 @@ internal class MineScreen(
                         44.dp,
                     ).withMargins(0, 12.dp, 0, 20.dp)
                 })
+                addView(TextView(activity).apply {
+                    text = "本视频收藏句子"
+                    textSize = 18f
+                    typeface = Typeface.DEFAULT_BOLD
+                    setTextColor(Color.WHITE)
+                })
+                addView(bindings.favoriteSentenceList)
                 addView(TextView(activity).apply {
                     text = "我的测试结果"
                     textSize = 18f
@@ -64,6 +75,7 @@ internal class MineScreen(
     }
 
     fun refreshContent(userApi: UserApi, token: String) {
+        renderFavoriteSentences()
         val list = bindings.myContentList ?: return
         list.removeAllViews()
         val course = screenState.currentCourse
@@ -125,6 +137,74 @@ internal class MineScreen(
         }
     }
 
+    private fun renderFavoriteSentences() {
+        val list = bindings.favoriteSentenceList ?: return
+        list.removeAllViews()
+        if (screenState.currentCourse == null) {
+            list.addView(widgets.infoCard("请先在「课程」中选择一门课程"))
+            return
+        }
+        val favorites = actions.favoriteSentences()
+        if (favorites.isEmpty()) {
+            list.addView(widgets.infoCard("当前课程还没有收藏句子"))
+            return
+        }
+        favorites.forEach { sentence ->
+            list.addView(favoriteCard(sentence) {
+                if (!actions.toggleSentenceFavorite(sentence)) {
+                    renderFavoriteSentences()
+                }
+            })
+        }
+    }
+
+    private fun favoriteCard(
+        sentence: SentenceBoundary,
+        onRemove: () -> Unit,
+    ): View {
+        val foreign = sentence.text?.takeIf { it.isNotBlank() } ?: "未提供文本"
+        val translation = sentence.zhText?.takeIf { it.isNotBlank() }
+        val cardText = buildString {
+            append("第 ${sentence.index + 1} 句")
+            append("\n$foreign")
+            translation?.let { append("\n$it") }
+        }
+        return LinearLayout(activity).apply {
+            orientation = LinearLayout.HORIZONTAL
+            gravity = Gravity.CENTER_VERTICAL
+            background = rounded(SampleTheme.selected, 8f)
+            setPadding(12.dp, 10.dp, 10.dp, 10.dp)
+            layoutParams = LinearLayout.LayoutParams(
+                ViewGroup.LayoutParams.MATCH_PARENT,
+                ViewGroup.LayoutParams.WRAP_CONTENT,
+            ).withMargins(0, 0, 0, 8.dp)
+            addView(TextView(activity).apply {
+                text = cardText
+                setTextColor(SampleTheme.textDark)
+                textSize = 13f
+                layoutParams = LinearLayout.LayoutParams(0, ViewGroup.LayoutParams.WRAP_CONTENT, 1f)
+            })
+            addView(Button(activity).apply {
+                text = "移除"
+                isAllCaps = false
+                setTextColor(Color.WHITE)
+                textSize = 12f
+                typeface = Typeface.DEFAULT_BOLD
+                minWidth = 0
+                minHeight = 0
+                minimumWidth = 0
+                minimumHeight = 0
+                setPadding(10.dp, 0, 10.dp, 0)
+                background = rounded(SampleTheme.danger, 8f)
+                layoutParams = LinearLayout.LayoutParams(
+                    ViewGroup.LayoutParams.WRAP_CONTENT,
+                    32.dp,
+                ).withMargins(8.dp, 0, 0, 0)
+                setOnClickListener { onRemove() }
+            })
+        }
+    }
+
     private fun deleteResult(userApi: UserApi, token: String, attemptId: String) {
         userApi.deleteResult(token, attemptId) { result ->
             result.fold(

android/sample-app/src/main/java/cn/learningpad/oraltrainer/sample/ui/PlayerChrome.kt

@@ -4,7 +4,11 @@ import android.app.Activity
 import android.content.pm.ActivityInfo
 import android.content.res.ColorStateList
 import android.graphics.Color
+import android.graphics.drawable.GradientDrawable
+import android.graphics.drawable.InsetDrawable
 import android.graphics.Typeface
+import android.text.TextUtils
+import android.view.MotionEvent
 import android.view.Gravity
 import android.view.View
 import android.view.ViewGroup
@@ -14,9 +18,12 @@ import android.widget.FrameLayout
 import android.widget.ImageButton
 import android.widget.ImageView
 import android.widget.LinearLayout
+import android.widget.ScrollView
 import android.widget.SeekBar
 import android.widget.Switch
 import android.widget.TextView
+import androidx.core.view.ViewCompat
+import androidx.core.view.WindowInsetsCompat
 import cn.learningpad.oraltrainer.sample.Module
 import cn.learningpad.oraltrainer.sample.PROGRESS_MAX
 import cn.learningpad.oraltrainer.sample.R
@@ -40,16 +47,38 @@ internal class PlayerChrome(
     override val uiContext get() = activity
 
     fun createPlayerSection(): View =
-        FrameLayout(activity).apply {
+        LinearLayout(activity).apply {
+            orientation = LinearLayout.VERTICAL
             background = rounded(SampleTheme.surface, 8f, SampleTheme.border)
             clipToOutline = true
             layoutParams = LinearLayout.LayoutParams(
                 ViewGroup.LayoutParams.MATCH_PARENT,
-                0,
-                1f,
+                ViewGroup.LayoutParams.WRAP_CONTENT,
             ).withMargins(16.dp, 0, 16.dp, 10.dp)
+            addView(
+                createPlayerOverlay().apply {
+                    layoutParams = LinearLayout.LayoutParams(
+                        ViewGroup.LayoutParams.MATCH_PARENT,
+                        ViewGroup.LayoutParams.WRAP_CONTENT,
+                    )
+                },
+            )
+            addView(
+                object : FrameLayout(activity) {
+                    override fun onMeasure(widthMeasureSpec: Int, heightMeasureSpec: Int) {
+                        val width = MeasureSpec.getSize(widthMeasureSpec)
+                        val height = (width * 9) / 16
+                        super.onMeasure(
+                            MeasureSpec.makeMeasureSpec(width, MeasureSpec.EXACTLY),
+                            MeasureSpec.makeMeasureSpec(height, MeasureSpec.EXACTLY),
+                        )
+                    }
+                }.apply {
+                    layoutParams = LinearLayout.LayoutParams(
+                        ViewGroup.LayoutParams.MATCH_PARENT,
+                        ViewGroup.LayoutParams.WRAP_CONTENT,
+                    )
                     addView(createPlayerView())
-            addView(createPlayerOverlay())
                     addView(
                         createRotationButton().apply {
                             layoutParams = FrameLayout.LayoutParams(
@@ -62,6 +91,8 @@ internal class PlayerChrome(
                             }
                         },
                     )
+                },
+            )
         }
 
     fun createSentenceSection(): View =
@@ -70,16 +101,94 @@ internal class PlayerChrome(
             setPadding(18.dp, 16.dp, 18.dp, 16.dp)
             layoutParams = LinearLayout.LayoutParams(
                 ViewGroup.LayoutParams.MATCH_PARENT,
-                ViewGroup.LayoutParams.WRAP_CONTENT,
+                0,
+                1f,
             ).withMargins(16.dp, 0, 16.dp, 12.dp)
         }
 
     fun createLandscapeContentView(): View =
-        FrameLayout(activity).apply {
+        LinearLayout(activity).apply {
+            orientation = LinearLayout.VERTICAL
             setBackgroundColor(Color.BLACK)
+            ViewCompat.setOnApplyWindowInsetsListener(this) { view, insets ->
+                val bars = insets.getInsets(
+                    WindowInsetsCompat.Type.systemBars() or WindowInsetsCompat.Type.displayCutout(),
+                )
+                val gestures = insets.getInsets(WindowInsetsCompat.Type.mandatorySystemGestures())
+                view.setPadding(
+                    bars.left,
+                    bars.top,
+                    bars.right,
+                    maxOf(bars.bottom, gestures.bottom, 24.dp),
+                )
+                WindowInsetsCompat.CONSUMED
+            }
+            addView(
+                createPlayerOverlay(compact = true).apply {
+                    layoutParams = LinearLayout.LayoutParams(
+                        ViewGroup.LayoutParams.MATCH_PARENT,
+                        ViewGroup.LayoutParams.WRAP_CONTENT,
+                    )
+                },
+            )
+            addView(
+                FrameLayout(activity).apply {
+                    layoutParams = LinearLayout.LayoutParams(
+                        ViewGroup.LayoutParams.MATCH_PARENT,
+                        0,
+                        1f,
+                    )
+                    addView(
+                        object : FrameLayout(activity) {
+                            override fun onMeasure(widthMeasureSpec: Int, heightMeasureSpec: Int) {
+                                val maxWidth = MeasureSpec.getSize(widthMeasureSpec)
+                                val maxHeight = MeasureSpec.getSize(heightMeasureSpec)
+                                val heightFromWidth = if (maxWidth > 0) (maxWidth * 9) / 16 else 0
+                                val width: Int
+                                val height: Int
+                                if (heightFromWidth <= maxHeight) {
+                                    width = maxWidth
+                                    height = heightFromWidth
+                                } else {
+                                    height = maxHeight
+                                    width = (maxHeight * 16) / 9
+                                }
+                                super.onMeasure(
+                                    MeasureSpec.makeMeasureSpec(width, MeasureSpec.EXACTLY),
+                                    MeasureSpec.makeMeasureSpec(height, MeasureSpec.EXACTLY),
+                                )
+                            }
+                        }.apply {
+                            layoutParams = FrameLayout.LayoutParams(
+                                ViewGroup.LayoutParams.WRAP_CONTENT,
+                                ViewGroup.LayoutParams.WRAP_CONTENT,
+                                Gravity.TOP or Gravity.CENTER_HORIZONTAL,
+                            )
                             addView(createPlayerView())
-            addView(createPlayerOverlay())
-            val subtitleContainer = LinearLayout(activity).apply {
+                        },
+                    )
+                    val subtitleContainer = object : LinearLayout(activity) {
+                        private var forwardTouchToPlayer = false
+
+                        override fun dispatchTouchEvent(event: MotionEvent): Boolean {
+                            if (event.actionMasked == MotionEvent.ACTION_DOWN) {
+                                forwardTouchToPlayer = !isTouchOverInteractiveControl(this, event)
+                            }
+                            val forwarded = if (forwardTouchToPlayer) {
+                                bindings.playerView.forwardGestureFrom(this, event)
+                            } else {
+                                false
+                            }
+                            val handled = forwarded || super.dispatchTouchEvent(event)
+                            if (event.actionMasked == MotionEvent.ACTION_UP ||
+                                event.actionMasked == MotionEvent.ACTION_CANCEL
+                            ) {
+                                forwardTouchToPlayer = false
+                            }
+                            return handled
+                        }
+
+                    }.apply {
                         orientation = LinearLayout.VERTICAL
                         layoutParams = FrameLayout.LayoutParams(
                             ViewGroup.LayoutParams.MATCH_PARENT,
@@ -98,6 +207,8 @@ internal class PlayerChrome(
                     addView(subtitleContainer)
                     applyLandscapeSubtitleState(subtitleContainer, animate = false)
                     installLandscapeSwipeGesture(subtitleContainer)
+                },
+            )
         }
 
     fun createSentenceContent(translucent: Boolean): LinearLayout {
@@ -106,6 +217,7 @@ internal class PlayerChrome(
             textSize = 17f
             typeface = Typeface.DEFAULT_BOLD
             maxLines = 1
+            ellipsize = TextUtils.TruncateAt.END
         }
         bindings.sentenceMetaText = TextView(activity).apply {
             setTextColor(if (translucent) SampleTheme.accentLight else SampleTheme.accentDeep)
@@ -117,7 +229,14 @@ internal class PlayerChrome(
             textSize = 22f
             typeface = Typeface.DEFAULT_BOLD
             setLineSpacing(2.dp.toFloat(), 1.05f)
-            setPadding(0, 10.dp, 0, 10.dp)
+            setPadding(0, 10.dp, 0, 4.dp)
+        }
+        bindings.sentenceZhText = TextView(activity).apply {
+            setTextColor(if (translucent) SampleTheme.textOverlay else SampleTheme.textSubtle)
+            textSize = 16f
+            setLineSpacing(2.dp.toFloat(), 1.05f)
+            setPadding(0, 0, 0, 8.dp)
+            visibility = View.GONE
         }
         bindings.progressBar = SeekBar(activity).apply {
             max = PROGRESS_MAX
@@ -157,16 +276,37 @@ internal class PlayerChrome(
                 background = rounded(SampleTheme.overlayFaint, 0f, null)
                 setPadding(18.dp, 14.dp, 18.dp, 16.dp)
             }
+            val topContent = LinearLayout(activity).apply {
+                orientation = LinearLayout.VERTICAL
+                if (screenState.activeModule != Module.TRAIN) {
                     addView(bindings.lessonTitleText)
+                }
                 if (screenState.activeModule == Module.TRAIN) {
                     addView(createContinuousPlaybackRow(translucent))
+                    addView(createChineseSubtitleRow(translucent))
                 }
                 addView(bindings.sentenceMetaText)
                 addView(LinearLayout(activity).apply {
                     orientation = LinearLayout.HORIZONTAL
-                gravity = Gravity.CENTER_VERTICAL
+                    gravity = Gravity.TOP
+                    addView(
+                        LinearLayout(activity).apply {
+                            orientation = LinearLayout.VERTICAL
                             addView(
                                 bindings.sentenceText,
+                                LinearLayout.LayoutParams(
+                                    ViewGroup.LayoutParams.MATCH_PARENT,
+                                    ViewGroup.LayoutParams.WRAP_CONTENT,
+                                ),
+                            )
+                            addView(
+                                bindings.sentenceZhText,
+                                LinearLayout.LayoutParams(
+                                    ViewGroup.LayoutParams.MATCH_PARENT,
+                                    ViewGroup.LayoutParams.WRAP_CONTENT,
+                                ),
+                            )
+                        },
                         LinearLayout.LayoutParams(
                             0,
                             ViewGroup.LayoutParams.WRAP_CONTENT,
@@ -179,6 +319,26 @@ internal class PlayerChrome(
                         LinearLayout.LayoutParams(32.dp, 32.dp).apply { marginStart = 10.dp },
                     )
                 })
+            }
+            addView(
+                ScrollView(activity).apply {
+                    isFillViewport = true
+                    isVerticalScrollBarEnabled = false
+                    overScrollMode = View.OVER_SCROLL_NEVER
+                    addView(
+                        topContent,
+                        LinearLayout.LayoutParams(
+                            ViewGroup.LayoutParams.MATCH_PARENT,
+                            ViewGroup.LayoutParams.WRAP_CONTENT,
+                        ),
+                    )
+                },
+                LinearLayout.LayoutParams(
+                    ViewGroup.LayoutParams.MATCH_PARENT,
+                    if (translucent) ViewGroup.LayoutParams.WRAP_CONTENT else 0,
+                    if (translucent) 0f else 1f,
+                ),
+            )
             addView(bindings.progressBar)
             addView(bindings.statusText)
         }
@@ -193,16 +353,15 @@ internal class PlayerChrome(
             )
         }.also { bindings.playerView = it }
 
-    private fun createPlayerOverlay(): View {
+    private fun createPlayerOverlay(compact: Boolean = false): View {
         val overlay = LinearLayout(activity).apply {
             orientation = LinearLayout.HORIZONTAL
             gravity = Gravity.CENTER_VERTICAL
-            setPadding(12.dp, 12.dp, 12.dp, 12.dp)
-            layoutParams = FrameLayout.LayoutParams(
-                ViewGroup.LayoutParams.MATCH_PARENT,
-                ViewGroup.LayoutParams.WRAP_CONTENT,
-                Gravity.TOP,
-            )
+            if (compact) {
+                setPadding(12.dp, 4.dp, 12.dp, 0)
+            } else {
+                setPadding(12.dp, 10.dp, 12.dp, 8.dp)
+            }
         }
         bindings.timeText = widgets.overlayPill("--:-- / --:--").apply {
             layoutParams = LinearLayout.LayoutParams(0, ViewGroup.LayoutParams.WRAP_CONTENT, 1f)
@@ -217,11 +376,27 @@ internal class PlayerChrome(
         ImageButton(activity).apply {
             setImageResource(R.drawable.ic_screen_rotation)
             imageTintList = ColorStateList.valueOf(Color.WHITE)
-            background = rounded(Color.argb(178, 9, 12, 16), 22.dp.toFloat())
+            val backgroundSize = 22.dp
+            val buttonSize = 44.dp
+            val inset = (buttonSize - backgroundSize) / 2
+            background = InsetDrawable(
+                GradientDrawable().apply {
+                    shape = GradientDrawable.OVAL
+                    setColor(Color.argb(102, 9, 12, 16))
+                    setSize(backgroundSize, backgroundSize)
+                },
+                inset,
+            )
             scaleType = ImageView.ScaleType.CENTER_INSIDE
             setPadding(10.dp, 10.dp, 10.dp, 10.dp)
             contentDescription = "横屏"
-            layoutParams = LinearLayout.LayoutParams(44.dp, 44.dp)
+            minimumWidth = 0
+            minimumHeight = 0
+            clipToOutline = false
+            if (android.os.Build.VERSION.SDK_INT >= android.os.Build.VERSION_CODES.M) {
+                foreground = null
+            }
+            layoutParams = LinearLayout.LayoutParams(buttonSize, buttonSize)
             setOnClickListener {
                 activity.requestedOrientation = ActivityInfo.SCREEN_ORIENTATION_SENSOR_LANDSCAPE
             }
@@ -276,6 +451,29 @@ internal class PlayerChrome(
             })
         }
 
+    private fun createChineseSubtitleRow(translucent: Boolean): View =
+        LinearLayout(activity).apply {
+            orientation = LinearLayout.HORIZONTAL
+            gravity = Gravity.CENTER_VERTICAL
+            setPadding(0, 6.dp, 0, 0)
+            addView(TextView(activity).apply {
+                text = "中文字幕"
+                setTextColor(if (translucent) SampleTheme.textOverlay else SampleTheme.textSubtle)
+                textSize = 13f
+                typeface = Typeface.DEFAULT_BOLD
+                layoutParams = LinearLayout.LayoutParams(0, ViewGroup.LayoutParams.WRAP_CONTENT, 1f)
+            })
+            addView(Switch(activity).apply {
+                isChecked = screenState.chineseSubtitlesEnabled
+                buttonTintList = ColorStateList.valueOf(
+                    if (translucent) SampleTheme.accentLight else SampleTheme.accent,
+                )
+                setOnCheckedChangeListener { _, checked ->
+                    seekActions.onChineseSubtitlesChanged(checked)
+                }
+            })
+        }
+
     private fun installLandscapeSwipeGesture(subtitleContainer: View) {
         bindings.playerView.setVerticalSwipeListener { isSwipeUp ->
             if (isSwipeUp == screenState.landscapeSubtitleVisible) return@setVerticalSwipeListener
@@ -327,4 +525,33 @@ internal class PlayerChrome(
         }
         view.startAnimation(animation)
     }
+
+    /** Lets controls own their touches while the subtitle background drives player gestures. */
+    private fun isTouchOverInteractiveControl(root: View, event: MotionEvent): Boolean =
+        hasInteractiveTouchTarget(root, event.rawX, event.rawY)
+
+    private fun hasInteractiveTouchTarget(view: View, screenX: Float, screenY: Float): Boolean {
+        if (view.visibility != View.VISIBLE || !view.isEnabled) {
+            return false
+        }
+        if (view.isClickable || view.isLongClickable || view.isContextClickable) {
+            val location = IntArray(2).also(view::getLocationOnScreen)
+            if (screenX >= location[0] &&
+                screenX < location[0] + view.width &&
+                screenY >= location[1] &&
+                screenY < location[1] + view.height
+            ) {
+                return true
+            }
+        }
+        if (view !is ViewGroup) {
+            return false
+        }
+        for (index in view.childCount - 1 downTo 0) {
+            if (hasInteractiveTouchTarget(view.getChildAt(index), screenX, screenY)) {
+                return true
+            }
+        }
+        return false
+    }
 }

android/sample-app/src/main/java/cn/learningpad/oraltrainer/sample/ui/ReadingScreen.kt

@@ -0,0 +1,306 @@
+package cn.learningpad.oraltrainer.sample.ui
+
+import android.app.Activity
+import android.graphics.Color
+import android.graphics.Typeface
+import android.os.SystemClock
+import android.view.Gravity
+import android.view.View
+import android.view.ViewGroup
+import android.widget.Button
+import android.widget.LinearLayout
+import android.widget.Switch
+import android.widget.TextView
+import androidx.recyclerview.widget.LinearLayoutManager
+import androidx.recyclerview.widget.RecyclerView
+import cn.learningpad.oraltrainer.sample.ReadingScreenActions
+import cn.learningpad.oraltrainer.sample.SampleTheme
+import cn.learningpad.oraltrainer.sample.TrainingScreenState
+import cn.learningpad.oraltrainer.sample.UiStyled
+import cn.learningpad.oraltrainer.sdk.OralTrainerController
+import cn.learningpad.oraltrainer.sdk.SentenceBoundary
+
+/** Lazy-loading full-text reading view for the selected course. */
+internal class ReadingScreen(
+    private val activity: Activity,
+    private val screenState: TrainingScreenState,
+    private val controller: OralTrainerController,
+    private val widgets: SampleWidgets,
+    private val actions: ReadingScreenActions,
+) : UiStyled {
+    override val uiContext get() = activity
+
+    private var readingAdapter: ReadingSentenceAdapter? = null
+    private var titleText: TextView? = null
+    private var sentenceListView: RecyclerView? = null
+    private var lastClickedSentenceIndex = Int.MIN_VALUE
+    private var lastClickedAtMs = 0L
+
+    fun createSection(): View {
+        val root = LinearLayout(activity).apply {
+            orientation = LinearLayout.VERTICAL
+            setPadding(16.dp, 12.dp, 16.dp, 16.dp)
+        }
+        val adapter = ReadingSentenceAdapter()
+        readingAdapter = adapter
+        val list = RecyclerView(activity).apply {
+            layoutManager = LinearLayoutManager(activity)
+            this.adapter = adapter
+            itemAnimator = null
+            setItemViewCacheSize(12)
+            clipToPadding = false
+            setPadding(0, 0, 0, 16.dp)
+            layoutParams = LinearLayout.LayoutParams(
+                ViewGroup.LayoutParams.MATCH_PARENT,
+                0,
+                1f,
+            )
+        }
+        sentenceListView = list
+
+        val title = TextView(activity).apply {
+            setTextColor(Color.WHITE)
+            textSize = 15f
+            typeface = Typeface.DEFAULT_BOLD
+            setPadding(0, 0, 0, 6.dp)
+        }
+        titleText = title
+        root.addView(title)
+        root.addView(createTranslationToggle())
+        root.addView(list)
+        renderSentences()
+
+        return root
+    }
+
+    fun renderSentences() {
+        val item = controller.currentTrainingItem()
+        val baseTitle = screenState.currentCourse?.title.orEmpty()
+            .ifBlank { item?.title ?: "阅读模式" }
+        val sentences = item?.sentences.orEmpty()
+        titleText?.text = if (sentences.isEmpty()) {
+            baseTitle
+        } else {
+            "$baseTitle · ${sentences.size} 句"
+        }
+        readingAdapter?.submitSentences(sentences)
+    }
+
+    fun refreshTranslations() {
+        readingAdapter?.notifyDataSetChanged()
+    }
+
+    fun refreshFavorites() {
+        readingAdapter?.notifyDataSetChanged()
+    }
+
+    fun scrollToCurrentSentence() {
+        val list = sentenceListView ?: return
+        val currentSentence = controller.currentSentence() ?: return
+        val item = controller.currentTrainingItem() ?: return
+        val position = item.sentences.indexOfFirst { it.index == currentSentence.index }
+        if (position < 0) return
+        list.post {
+            if (sentenceListView !== list) return@post
+            val layoutManager = list.layoutManager as? LinearLayoutManager ?: return@post
+            val offset = (list.height * 0.25f).toInt().coerceAtLeast(0)
+            layoutManager.scrollToPositionWithOffset(position, offset)
+        }
+    }
+
+    private fun createTranslationToggle(): View =
+        LinearLayout(activity).apply {
+            orientation = LinearLayout.HORIZONTAL
+            gravity = Gravity.CENTER_VERTICAL
+            setPadding(0, 4.dp, 0, 10.dp)
+            addView(TextView(activity).apply {
+                text = "中文翻译"
+                setTextColor(Color.WHITE)
+                textSize = 15f
+                typeface = Typeface.DEFAULT_BOLD
+                layoutParams = LinearLayout.LayoutParams(0, ViewGroup.LayoutParams.WRAP_CONTENT, 1f)
+            })
+            addView(Switch(activity).apply {
+                isChecked = screenState.readingTranslationVisible
+                buttonTintList = android.content.res.ColorStateList.valueOf(SampleTheme.accent)
+                setOnCheckedChangeListener { _, checked ->
+                    actions.onReadingTranslationVisible(checked)
+                }
+            })
+        }
+
+    private fun translationText(sentence: SentenceBoundary): String = when {
+        !sentence.zhText.isNullOrBlank() -> sentence.zhText.orEmpty()
+        screenState.chineseTranslationLoading -> "正在翻译中文…"
+        screenState.chineseTranslationError.isNotBlank() -> screenState.chineseTranslationError
+        else -> "暂无中文翻译"
+    }
+
+    private inner class ReadingSentenceAdapter :
+        RecyclerView.Adapter<RecyclerView.ViewHolder>() {
+
+        private var sentences: List<SentenceBoundary> = emptyList()
+
+        init {
+            setHasStableIds(true)
+        }
+
+        override fun getItemId(position: Int): Long {
+            return if (sentences.isEmpty()) EMPTY_ID else sentences[position].index.toLong()
+        }
+
+        override fun getItemViewType(position: Int): Int {
+            return if (sentences.isEmpty()) TYPE_EMPTY else TYPE_SENTENCE
+        }
+
+        override fun onCreateViewHolder(parent: ViewGroup, viewType: Int): RecyclerView.ViewHolder {
+            return if (viewType == TYPE_EMPTY) {
+                val empty = TextView(activity).apply {
+                    text = "当前课程还没有可阅读的句子文本"
+                    setTextColor(SampleTheme.textDark)
+                    textSize = 13f
+                    background = rounded(SampleTheme.selected, 8f)
+                    setPadding(12.dp, 10.dp, 12.dp, 10.dp)
+                    layoutParams = RecyclerView.LayoutParams(
+                        ViewGroup.LayoutParams.MATCH_PARENT,
+                        ViewGroup.LayoutParams.WRAP_CONTENT,
+                    )
+                }
+                EmptyViewHolder(empty)
+            } else {
+                SentenceViewHolder(createSentenceView())
+            }
+        }
+
+        override fun onBindViewHolder(holder: RecyclerView.ViewHolder, position: Int) {
+            if (holder !is SentenceViewHolder || sentences.isEmpty()) return
+            val sentence = sentences[position]
+            attachDoubleTap(holder.itemView, sentence)
+            holder.text.text =
+                "${sentence.index + 1}. ${sentence.text?.takeIf { it.isNotBlank() } ?: "未提供文本"}"
+            holder.translation.visibility = if (screenState.readingTranslationVisible) {
+                View.VISIBLE
+            } else {
+                View.GONE
+            }
+            holder.translation.text = translationText(sentence)
+            bindFavoriteButton(holder.favoriteButton, sentence)
+        }
+
+        override fun getItemCount(): Int {
+            return if (sentences.isEmpty()) 1 else sentences.size
+        }
+
+        fun submitSentences(updated: List<SentenceBoundary>) {
+            sentences = updated
+            notifyDataSetChanged()
+        }
+
+        private fun createSentenceView(): View {
+            val text = TextView(activity).apply {
+                setTextColor(Color.WHITE)
+                textSize = 16f
+            }
+            val translation = TextView(activity).apply {
+                setTextColor(SampleTheme.readingTranslation)
+                textSize = 14f
+                setPadding(0, 6.dp, 0, 0)
+            }
+            val favoriteButton = Button(activity).apply {
+                isAllCaps = false
+                setTextColor(Color.WHITE)
+                textSize = 12f
+                typeface = Typeface.DEFAULT_BOLD
+                minWidth = 0
+                minHeight = 0
+                minimumWidth = 0
+                minimumHeight = 0
+                setPadding(12.dp, 0, 12.dp, 0)
+                layoutParams = LinearLayout.LayoutParams(
+                    ViewGroup.LayoutParams.WRAP_CONTENT,
+                    32.dp,
+                ).withMargins(10.dp, 0, 0, 0)
+            }
+            val textColumn = LinearLayout(activity).apply {
+                orientation = LinearLayout.VERTICAL
+                layoutParams = LinearLayout.LayoutParams(0, ViewGroup.LayoutParams.WRAP_CONTENT, 1f)
+                addView(text)
+                addView(translation)
+            }
+            return LinearLayout(activity).apply {
+                orientation = LinearLayout.HORIZONTAL
+                gravity = Gravity.CENTER_VERTICAL
+                background = rounded(SampleTheme.surface, 10f, SampleTheme.border)
+                setPadding(14.dp, 12.dp, 10.dp, 12.dp)
+                layoutParams = RecyclerView.LayoutParams(
+                    ViewGroup.LayoutParams.MATCH_PARENT,
+                    ViewGroup.LayoutParams.WRAP_CONTENT,
+                ).apply {
+                    bottomMargin = 10.dp
+                }
+                addView(textColumn)
+                addView(favoriteButton)
+            }
+        }
+
+        private fun bindFavoriteButton(button: Button, sentence: SentenceBoundary) {
+            val isFavorite = actions.isSentenceFavorite(sentence)
+            button.setText(if (isFavorite) "已收藏" else "收藏")
+            button.background = rounded(
+                if (isFavorite) SampleTheme.accent else SampleTheme.button,
+                8f,
+            )
+            button.setOnClickListener {
+                val favorite = actions.toggleSentenceFavorite(sentence)
+                button.setText(if (favorite) "已收藏" else "收藏")
+                button.background = rounded(
+                    if (favorite) SampleTheme.accent else SampleTheme.button,
+                    8f,
+                )
+            }
+        }
+
+        private fun attachDoubleTap(view: View, sentence: SentenceBoundary) {
+            view.isClickable = true
+            view.setOnClickListener {
+                val now = SystemClock.elapsedRealtime()
+                val isDoubleTap = lastClickedSentenceIndex == sentence.index &&
+                    now - lastClickedAtMs <= DOUBLE_TAP_INTERVAL_MS
+                if (isDoubleTap) {
+                    lastClickedSentenceIndex = Int.MIN_VALUE
+                    lastClickedAtMs = 0L
+                    actions.onSentenceDoubleTapped(sentence)
+                } else {
+                    lastClickedSentenceIndex = sentence.index
+                    lastClickedAtMs = now
+                }
+            }
+        }
+
+        private inner class EmptyViewHolder(root: View) : RecyclerView.ViewHolder(root)
+
+        private inner class SentenceViewHolder(
+            root: View,
+        ) : RecyclerView.ViewHolder(root) {
+            val text: TextView
+            val translation: TextView
+            val favoriteButton: Button
+
+            init {
+                val textColumn = root as ViewGroup
+                val content = textColumn.getChildAt(0) as ViewGroup
+                text = content.getChildAt(0) as TextView
+                translation = content.getChildAt(1) as TextView
+                favoriteButton = textColumn.getChildAt(1) as Button
+            }
+        }
+
+    }
+
+    private companion object {
+        private const val TYPE_EMPTY = 0
+        private const val TYPE_SENTENCE = 1
+        private const val EMPTY_ID = Long.MIN_VALUE
+        private const val DOUBLE_TAP_INTERVAL_MS = 300L
+    }
+}

android/sample-app/src/main/java/cn/learningpad/oraltrainer/sample/ui/SampleWidgets.kt

@@ -30,6 +30,7 @@ internal class SampleWidgets(
             setPadding(16.dp, 40.dp, 16.dp, 0)
             addView(moduleTab("课程", Module.COURSE, actions))
             addView(moduleTab("训练", Module.TRAIN, actions))
+            addView(moduleTab("阅读", Module.READING, actions))
             addView(moduleTab("测试", Module.TEST, actions))
             addView(moduleTab("我的", Module.MINE, actions))
         }
@@ -153,40 +154,4 @@ internal class SampleWidgets(
         }
     }
 
-    fun videoCard(
-        id: String,
-        title: String,
-        meta: String,
-        onClick: () -> Unit,
-    ): View {
-        val selected = screenState.activeItemId == id
-        val titleColor = if (selected) SampleTheme.textDark else Color.WHITE
-        val metaColor = if (selected) SampleTheme.accentDeep else SampleTheme.textMuted
-        return LinearLayout(activity).apply {
-            orientation = LinearLayout.VERTICAL
-            gravity = Gravity.CENTER_VERTICAL
-            isClickable = true
-            background = if (selected) {
-                rounded(SampleTheme.selected, 8f, SampleTheme.accent)
-            } else {
-                rounded(SampleTheme.surface, 8f, SampleTheme.border)
-            }
-            setPadding(14.dp, 12.dp, 14.dp, 12.dp)
-            layoutParams = LinearLayout.LayoutParams(218.dp, 84.dp).withMargins(0, 0, 10.dp, 0)
-            setOnClickListener { onClick() }
-            addView(TextView(activity).apply {
-                text = title
-                setTextColor(titleColor)
-                textSize = 15f
-                typeface = Typeface.DEFAULT_BOLD
-                maxLines = 2
-            })
-            addView(TextView(activity).apply {
-                text = meta
-                setTextColor(metaColor)
-                textSize = 12f
-                setPadding(0, 6.dp, 0, 0)
-            })
-        }
-    }
 }

android/sample-app/src/main/java/cn/learningpad/oraltrainer/sample/ui/TrainScreen.kt

@@ -5,23 +5,14 @@ import android.graphics.Color
 import android.graphics.Typeface
 import android.view.Gravity
 import android.view.View
-import android.view.ViewGroup
-import android.widget.HorizontalScrollView
 import android.widget.LinearLayout
 import android.widget.TextView
-import cn.learningpad.oraltrainer.sample.SampleBindings
 import cn.learningpad.oraltrainer.sample.SampleTheme
-import cn.learningpad.oraltrainer.sample.TrainScreenActions
-import cn.learningpad.oraltrainer.sample.TrainingScreenState
 import cn.learningpad.oraltrainer.sample.UiStyled
 
 internal class TrainScreen(
     private val activity: Activity,
-    private val bindings: SampleBindings,
-    private val screenState: TrainingScreenState,
-    private val widgets: SampleWidgets,
     private val playerChrome: PlayerChrome,
-    private val actions: TrainScreenActions,
 ) : UiStyled {
     override val uiContext get() = activity
 
@@ -29,47 +20,13 @@ internal class TrainScreen(
         root.addView(createHeader())
         root.addView(playerChrome.createPlayerSection())
         root.addView(playerChrome.createSentenceSection())
-        root.addView(createCatalogSection())
-    }
-
-    fun renderCatalogStrip() {
-        if (!bindings.hasCatalogChrome) return
-        bindings.catalogList.removeAllViews()
-        bindings.catalogStatusText.text = screenState.catalogStatus
-        screenState.catalogVideos.forEach { video ->
-            val enrolled = screenState.courseEnrollment[video.videoHash] == true
-            bindings.catalogList.addView(
-                widgets.videoCard(
-                    id = video.videoHash,
-                    title = video.title,
-                    meta = if (enrolled) {
-                        "已选修 · ${video.sentenceCount} 句"
-                    } else {
-                        "未选修 · 点击选修并进入"
-                    },
-                    onClick = { actions.onTrainCatalogVideoClick(video) },
-                ),
-            )
-        }
-    }
-
-    fun showCatalogLoading() {
-        if (!bindings.hasCatalogChrome) return
-        bindings.catalogList.removeAllViews()
-        bindings.catalogList.addView(widgets.infoCard("课程加载中..."))
-    }
-
-    fun setCatalogStatus(text: String) {
-        if (bindings.hasCatalogChrome) {
-            bindings.catalogStatusText.text = text
-        }
     }
 
     private fun createHeader(): View {
-        val titleBlock = LinearLayout(activity).apply {
-            orientation = LinearLayout.VERTICAL
+        return LinearLayout(activity).apply {
+            orientation = LinearLayout.HORIZONTAL
             gravity = Gravity.CENTER_VERTICAL
-            layoutParams = LinearLayout.LayoutParams(0, ViewGroup.LayoutParams.WRAP_CONTENT, 1f)
+            setPadding(18.dp, 16.dp, 18.dp, 12.dp)
             addView(TextView(activity).apply {
                 text = "八哥口语"
                 setTextColor(Color.WHITE)
@@ -80,52 +37,10 @@ internal class TrainScreen(
             addView(TextView(activity).apply {
                 text = "外语跟读训练神器"
                 setTextColor(SampleTheme.textMuted)
-                textSize = 20f
-                setPadding(1.dp, 4.dp, 0, 0)
+                textSize = 18f
+                includeFontPadding = false
+                setPadding(12.dp, 6.dp, 0, 0)
             })
         }
-        return LinearLayout(activity).apply {
-            orientation = LinearLayout.HORIZONTAL
-            gravity = Gravity.CENTER_VERTICAL
-            setPadding(18.dp, 16.dp, 18.dp, 12.dp)
-            addView(titleBlock)
-            addView(widgets.headerButton("导入") { actions.onImportLocalVideo() })
-            addView(widgets.headerButton("刷新") { actions.onRefreshCatalog() })
-        }
-    }
-
-    private fun createCatalogSection(): View {
-        bindings.catalogStatusText = TextView(activity).apply {
-            setTextColor(SampleTheme.textMuted)
-            textSize = 13f
-            gravity = Gravity.END
-            text = "正在同步"
-            layoutParams = LinearLayout.LayoutParams(0, ViewGroup.LayoutParams.WRAP_CONTENT, 1f)
-        }
-        bindings.catalogList = LinearLayout(activity).apply {
-            orientation = LinearLayout.HORIZONTAL
-            setPadding(0, 10.dp, 0, 4.dp)
-        }
-        val header = LinearLayout(activity).apply {
-            orientation = LinearLayout.HORIZONTAL
-            gravity = Gravity.CENTER_VERTICAL
-            addView(TextView(activity).apply {
-                text = "课程"
-                setTextColor(Color.WHITE)
-                textSize = 16f
-                typeface = Typeface.DEFAULT_BOLD
-            })
-            addView(bindings.catalogStatusText)
-        }
-        val scroller = HorizontalScrollView(activity).apply {
-            isHorizontalScrollBarEnabled = false
-            addView(bindings.catalogList)
-        }
-        return LinearLayout(activity).apply {
-            orientation = LinearLayout.VERTICAL
-            setPadding(16.dp, 0, 16.dp, 14.dp)
-            addView(header)
-            addView(scroller)
-        }
     }
 }

android/sample-app/src/main/res/drawable/ic_notification_play.xml

@@ -0,0 +1,9 @@
+<vector xmlns:android="http://schemas.android.com/apk/res/android"
+    android:width="24dp"
+    android:height="24dp"
+    android:viewportWidth="24"
+    android:viewportHeight="24">
+    <path
+        android:fillColor="#FFFFFFFF"
+        android:pathData="M8,5v14l11,-7z" />
+</vector>

android/sample-app/src/main/res/values-zh/strings.xml

@@ -1,4 +1,7 @@
 <?xml version="1.0" encoding="utf-8"?>
 <resources>
     <string name="app_name">八哥口语</string>
+    <string name="training_playback_channel_name">训练播放</string>
+    <string name="training_playback_notification_title">口语训练</string>
+    <string name="training_playback_notification_text">正在后台播放</string>
 </resources>

android/sample-app/src/main/res/values/strings.xml

@@ -1,3 +1,6 @@
 <resources>
     <string name="app_name">MynaTalk</string>
+    <string name="training_playback_channel_name">Training playback</string>
+    <string name="training_playback_notification_title">Oral training</string>
+    <string name="training_playback_notification_text">Playing in the background</string>
 </resources>

ios/OralTrainer.xcodeproj/project.pbxproj

@@ -10,6 +10,7 @@
 		AA0000000000000000000090 /* AuthAPI.swift in Sources */ = {isa = PBXBuildFile; fileRef = AA0000000000000000000091 /* AuthAPI.swift */; };
 		AA00000000000000000000A1 /* UserHistory.swift in Sources */ = {isa = PBXBuildFile; fileRef = AA00000000000000000000A2 /* UserHistory.swift */; };
 		AA00000000000000000000A3 /* TrainingHistoryStore.swift in Sources */ = {isa = PBXBuildFile; fileRef = AA00000000000000000000A4 /* TrainingHistoryStore.swift */; };
+		AA00000000000000000000B0 /* SentenceExplanationStore.swift in Sources */ = {isa = PBXBuildFile; fileRef = AA00000000000000000000B1 /* SentenceExplanationStore.swift */; };
 		AA0000000000000000000092 /* AuthView.swift in Sources */ = {isa = PBXBuildFile; fileRef = AA0000000000000000000093 /* AuthView.swift */; };
 		AA0000000000000000000040 /* Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = AA0000000000000000000020 /* Assets.xcassets */; };
 		AA0000000000000000000042 /* OralTrainerApp.swift in Sources */ = {isa = PBXBuildFile; fileRef = AA0000000000000000000022 /* OralTrainerApp.swift */; };
@@ -41,6 +42,7 @@
 		AA0000000000000000000091 /* AuthAPI.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = AuthAPI.swift; sourceTree = "<group>"; };
 		AA00000000000000000000A2 /* UserHistory.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = UserHistory.swift; sourceTree = "<group>"; };
 		AA00000000000000000000A4 /* TrainingHistoryStore.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = TrainingHistoryStore.swift; sourceTree = "<group>"; };
+		AA00000000000000000000B1 /* SentenceExplanationStore.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = SentenceExplanationStore.swift; sourceTree = "<group>"; };
 		AA0000000000000000000093 /* AuthView.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = AuthView.swift; sourceTree = "<group>"; };
 		AA000000000000000000000A /* OralTrainer.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = OralTrainer.app; sourceTree = BUILT_PRODUCTS_DIR; };
 		AA000000000000000000001F /* Info.plist */ = {isa = PBXFileReference; lastKnownFileType = text.plist.xml; path = Info.plist; sourceTree = "<group>"; };
@@ -131,6 +133,7 @@
 				AA000000000000000000002A /* SentenceBoundaryAPI.swift */,
 				AA000000000000000000002B /* ImitationAssessor.swift */,
 				AA00000000000000000000A4 /* TrainingHistoryStore.swift */,
+				AA00000000000000000000B1 /* SentenceExplanationStore.swift */,
 				AA00000000000000000000A2 /* UserHistory.swift */,
 				AA000000000000000000007E /* DubbingCompositor.swift */,
 				AA0000000000000000000081 /* AudioRecorderSession.swift */,
@@ -267,6 +270,7 @@
 				AA0000000000000000000090 /* AuthAPI.swift in Sources */,
 				AA00000000000000000000A1 /* UserHistory.swift in Sources */,
 				AA00000000000000000000A3 /* TrainingHistoryStore.swift in Sources */,
+				AA00000000000000000000B0 /* SentenceExplanationStore.swift in Sources */,
 				AA0000000000000000000092 /* AuthView.swift in Sources */,
 				AA0000000000000000000053 /* TestView.swift in Sources */,
 				AA0000000000000000000054 /* FullscreenPlayerView.swift in Sources */,

ios/OralTrainer.xcodeproj/project.xcworkspace/xcuserdata/liushuming.xcuserdatad/IDEFindNavigatorScopes.plist

@@ -0,0 +1,5 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<array/>
+</plist>

ios/OralTrainer/AppDelegate.swift

@@ -9,4 +9,38 @@ final class AppDelegate: NSObject, UIApplicationDelegate {
     ) -> UIInterfaceOrientationMask {
         AppDelegate.orientationLock
     }
+
+    static func lockOrientation(_ mask: UIInterfaceOrientationMask) {
+        orientationLock = mask
+        guard let scene = UIApplication.shared.connectedScenes
+            .compactMap({ $0 as? UIWindowScene })
+            .first(where: { $0.activationState == .foregroundActive })
+            ?? UIApplication.shared.connectedScenes.compactMap({ $0 as? UIWindowScene }).first
+        else {
+            return
+        }
+
+        func update(_ controller: UIViewController?) {
+            guard let controller else { return }
+            controller.setNeedsUpdateOfSupportedInterfaceOrientations()
+            update(controller.presentedViewController)
+            controller.children.forEach { update($0) }
+        }
+
+        let requested: UIInterfaceOrientationMask
+        if mask.contains(.landscapeRight) {
+            requested = .landscapeRight
+        } else if mask.contains(.landscapeLeft) {
+            requested = .landscapeLeft
+        } else {
+            requested = mask
+        }
+
+        scene.windows.forEach { update($0.rootViewController) }
+        scene.requestGeometryUpdate(.iOS(interfaceOrientations: requested)) { _ in }
+        DispatchQueue.main.async {
+            scene.windows.forEach { update($0.rootViewController) }
+            scene.requestGeometryUpdate(.iOS(interfaceOrientations: requested)) { _ in }
+        }
+    }
 }

ios/OralTrainer/Config.swift

@@ -15,6 +15,9 @@ enum AppConfig {
     static let connectTimeout: TimeInterval = 15
     static let readTimeout: TimeInterval = 30
     static let assessmentTimeout: TimeInterval = 180
+    static let translationTimeout: TimeInterval = 30
+    static let translationDeadline: TimeInterval = 900
+    static let translationPollIntervalNs: UInt64 = 1_500_000_000
 
     static let sha256Pattern = #"^[0-9a-fA-F]{64}$"#
 

ios/OralTrainer/Models/PlaybackModels.swift

@@ -63,3 +63,22 @@ enum BoundaryState: Equatable {
     case loaded(count: Int)
     case failed(message: String)
 }
+
+/// 云端视频按需流式播放；若本地已有完整缓存文件则优先使用。
+enum VideoCacheState: Equatable {
+    case notRequired
+    case downloading
+    case ready
+    case failed(message: String)
+
+    var statusMessage: String? {
+        switch self {
+        case .notRequired, .ready:
+            return nil
+        case .downloading:
+            return "正在缓冲视频…"
+        case .failed(let message):
+            return "视频加载失败：\(message)"
+        }
+    }
+}

ios/OralTrainer/Models/TrainingModels.swift

@@ -7,6 +7,7 @@ struct SentenceBoundary: Identifiable, Equatable, Decodable {
     let text: String?
     let language: String?
     let referenceSpeechDurationMs: Int64?
+    let zhText: String?
 
     var id: Int { index }
 
@@ -15,6 +16,48 @@ struct SentenceBoundary: Identifiable, Equatable, Decodable {
         case startMs = "start_ms"
         case endMs = "end_ms"
         case referenceSpeechDurationMs = "reference_speech_duration_ms"
+        case zhText = "zh_text"
+    }
+
+    init(from decoder: Decoder) throws {
+        let container = try decoder.container(keyedBy: CodingKeys.self)
+        index = try container.decode(Int.self, forKey: .index)
+        startMs = try container.decode(Int64.self, forKey: .startMs)
+        endMs = try container.decode(Int64.self, forKey: .endMs)
+        text = try container.decodeIfPresent(String.self, forKey: .text)
+        language = try container.decodeIfPresent(String.self, forKey: .language)
+        referenceSpeechDurationMs = try container.decodeIfPresent(Int64.self, forKey: .referenceSpeechDurationMs)
+        zhText = try container.decodeIfPresent(String.self, forKey: .zhText)
+    }
+
+    init(
+        index: Int,
+        startMs: Int64,
+        endMs: Int64,
+        text: String?,
+        language: String?,
+        referenceSpeechDurationMs: Int64?,
+        zhText: String? = nil
+    ) {
+        self.index = index
+        self.startMs = startMs
+        self.endMs = endMs
+        self.text = text
+        self.language = language
+        self.referenceSpeechDurationMs = referenceSpeechDurationMs
+        self.zhText = zhText
+    }
+
+    func withZhText(_ zhText: String?) -> SentenceBoundary {
+        SentenceBoundary(
+            index: index,
+            startMs: startMs,
+            endMs: endMs,
+            text: text,
+            language: language,
+            referenceSpeechDurationMs: referenceSpeechDurationMs,
+            zhText: zhText
+        )
     }
 }
 
@@ -66,6 +109,19 @@ struct TrainingVideoSummary: Identifiable, Decodable {
         }
         return URL(string: absolute)
     }
+
+    func copy(durationMs: Int64? = self.durationMs, sentenceCount: Int = self.sentenceCount) -> Self {
+        TrainingVideoSummary(
+            videoHash: videoHash,
+            title: title,
+            streamUrl: streamUrl,
+            durationMs: durationMs,
+            sizeBytes: sizeBytes,
+            language: language,
+            sentenceCount: sentenceCount,
+            status: status
+        )
+    }
 }
 
 struct TrainingMediaItem: Identifiable, Equatable {

ios/OralTrainer/Player/PlayerController.swift

@@ -10,6 +10,7 @@ final class PlayerController: ObservableObject {
     @Published private(set) var currentSentence: SentenceBoundary?
     @Published var boundaryState: BoundaryState = .idle
     @Published var transientMessage: String?
+    @Published private(set) var cacheState: VideoCacheState = .notRequired
 
     let player = AVPlayer()
 
@@ -30,8 +31,9 @@ final class PlayerController: ObservableObject {
     private var swipeBackAnchorTime: TimeInterval = 0
     private var speedBeforeLongPress: Float?
     private var transientTask: Task<Void, Never>?
-    private var cacheSwapTask: Task<Void, Never>?
+    private var cacheFallbackTask: Task<Void, Never>?
     private var activePlaybackURL: URL?
+    private var didAttemptFullFileFallback = false
     private var sentenceEndTimeObserver: Any?
     private var sentenceEndGuardToken = 0
 
@@ -39,6 +41,8 @@ final class PlayerController: ObservableObject {
     /// AVPlayer 在个别句子边界会把 seek 结果上报得比目标早 1ms 左右。
     /// 如果恰好落在句边界前，tick 会把句子重新判定回上一句；稍微进入目标句可消除该问题。
     private static let sentenceSeekInsetMs: Int64 = 20
+    private static let previousSentenceToleranceMs: Int64 = 250
+    private static let nextSentenceToleranceMs: Int64 = 150
 
     private var timeObserver: Any?
     private var endObserver: NSObjectProtocol?
@@ -88,33 +92,29 @@ final class PlayerController: ObservableObject {
     @MainActor
     func load(_ item: TrainingMediaItem) {
         cancelSentenceBoundaryGuard()
+        cacheFallbackTask?.cancel()
+        cacheFallbackTask = nil
+        didAttemptFullFileFallback = false
         lastSentenceIndex = nil
         swipeBackAnchorIndex = nil
         pendingSeekPositionMs = nil
         stopAtSentenceEnd = false
         stopAtSentenceEndIndex = nil
+        playbackIntent = false
         mediaItem = item
         currentItem = item
         boundaryState = .idle
-        let playbackURL: URL
-        if item.isLocalFile {
-            playbackURL = item.url
-        } else if let cachedURL = StreamingVideoCache.shared.cachedFileIfAvailable(for: item.url) {
-            playbackURL = cachedURL
-        } else {
-            playbackURL = item.url
-        }
 
-        activePlaybackURL = playbackURL
-        let playerItem = AVPlayerItem(url: playbackURL)
-        attachItemObservers(to: playerItem)
-        player.replaceCurrentItem(with: playerItem)
-        if config.autoPlay {
-            player.play()
+        if item.isLocalFile {
+            cacheState = .notRequired
+            preparePlaybackItem(with: item.url)
+        } else if let cachedURL = StreamingVideoCache.shared.cachedFileIfAvailable(for: item.url) {
+            cacheState = .ready
+            preparePlaybackItem(with: cachedURL)
+        } else {
+            cacheState = .notRequired
+            preparePlaybackItem(with: item.url)
         }
-        playbackIntent = config.autoPlay
-        scheduleCachePrefetchIfNeeded(for: item, activeURL: playbackURL)
-        tick()
     }
 
     /// 句边界请求完成后只更新训练数据，不重新创建 AVPlayerItem。
@@ -201,8 +201,13 @@ final class PlayerController: ObservableObject {
         guard let item = mediaItem, !item.sentences.isEmpty else {
             return false
         }
-        guard let currentIndex = sentenceAnchorIndex() else { return false }
-        let target = item.sentences[max(0, currentIndex - 1)]
+        // 用播放位置查找目标，而不是用可能已经过期的 sentence index 推导。
+        // 播放点落在两句之间的空隙时也能正确定位前一句。
+        let position = referencePositionMs()
+        let target = item.sentences
+            .reversed()
+            .first { $0.startMs < position - Self.previousSentenceToleranceMs }
+            ?? item.sentences[0]
         seekToSentenceBeginning(target)
         return true
     }
@@ -212,12 +217,14 @@ final class PlayerController: ObservableObject {
         guard let item = mediaItem, !item.sentences.isEmpty else {
             return false
         }
-        guard let currentIndex = sentenceAnchorIndex() else { return false }
-        let targetIndex = currentIndex + 1
-        guard item.sentences.indices.contains(targetIndex) else {
+        // Android 端同样按当前位置向后查找下一个 startMs。即使边界表存在句间
+        // 空隙，或 AVPlayer 的时间回调仍停留在上一句，也能跳到实际的下一句。
+        let position = referencePositionMs()
+        guard let target = item.sentences.first(where: {
+            $0.startMs > position + Self.nextSentenceToleranceMs
+        }) else {
             return false
         }
-        let target = item.sentences[targetIndex]
         seekToSentenceBeginning(target)
         return true
     }
@@ -309,21 +316,27 @@ final class PlayerController: ObservableObject {
     }
 
     /// 只播放当前句子，播完自动暂停并回到句首（评测“播放原句”）。
+    /// 必须先打上句尾标记再 seek，这样 completion 里的 forwardPlaybackEndTime
+    /// 会卡在当前句 endMs，而不是等 250ms 观察器才停。
     func playCurrentSentenceAndStop() {
         guard let sentence = currentSentence else {
             return
         }
-        performSeek(to: sentenceSeekPosition(sentence))
         stopAtSentenceEnd = true
         stopAtSentenceEndIndex = sentence.index
-        play()
+        lastSentenceIndex = sentence.index
+        playbackIntent = true
+        performSeek(to: sentenceSeekPosition(sentence), resumePlayback: true)
     }
 
     // MARK: - 设置
 
     func setContinuousPlayback(_ enabled: Bool) {
         config.continuousPlayback = enabled
-        refreshSentenceBoundaryGuard()
+        if !enabled {
+            lastSentenceIndex = currentSentence?.index ?? lastSentenceIndex
+        }
+        refreshSentenceBoundaryGuard(force: true)
     }
 
     func setSentenceMode(_ enabled: Bool) {
@@ -465,41 +478,30 @@ final class PlayerController: ObservableObject {
         pendingSeekPositionMs ?? currentPositionMs()
     }
 
-    @MainActor
-    private func scheduleCachePrefetchIfNeeded(for item: TrainingMediaItem, activeURL: URL) {
-        cacheSwapTask?.cancel()
-        cacheSwapTask = nil
-        guard !item.isLocalFile,
-              activeURL != item.url else {
-            return
-        }
-
-        let remoteURL = item.url
-        cacheSwapTask = Task { [weak self] in
-            guard let cachedURL = await StreamingVideoCache.shared.prefetch(remoteURL) else {
-                return
-            }
-            self?.swapToCachedFileIfNeeded(remoteURL: remoteURL, cachedURL: cachedURL)
-        }
-    }
-
-    @MainActor
-    private func swapToCachedFileIfNeeded(remoteURL: URL, cachedURL: URL) {
-        guard !Task.isCancelled,
-              mediaItem?.url == remoteURL,
-	              activePlaybackURL != cachedURL else {
-            return
-        }
-
-        let targetPositionMs = pendingSeekPositionMs ?? currentPositionMs()
-        let newItem = AVPlayerItem(url: cachedURL)
-        activePlaybackURL = cachedURL
-
-        // 递增 seek generation，让旧远程 Item 上未完成的 seek 回调失效。
-        seekGeneration += 1
+    private func preparePlaybackItem(with playbackURL: URL) {
+        let newItem = makePlayerItem(url: playbackURL)
+        activePlaybackURL = playbackURL
         attachItemObservers(to: newItem)
         player.replaceCurrentItem(with: newItem)
-        performSeek(to: targetPositionMs)
+        let shouldPlay = config.autoPlay || playbackIntent
+        if shouldPlay {
+            player.play()
+        }
+        playbackIntent = shouldPlay
+        tick()
+    }
+
+    private func makePlayerItem(url: URL) -> AVPlayerItem {
+        guard !url.isFileURL else {
+            return AVPlayerItem(url: url)
+        }
+        var options: [String: Any] = [:]
+        let ext = url.pathExtension.lowercased()
+        if ext.isEmpty || ["mp4", "m4v", "mov"].contains(ext) {
+            options["AVURLAssetOverrideMIMETypeKey"] = "video/mp4"
+        }
+        let asset = AVURLAsset(url: url, options: options.isEmpty ? nil : options)
+        return AVPlayerItem(asset: asset, automaticallyLoadedAssetKeys: ["playable", "tracks", "duration"])
     }
 
     /// AVPlayer 的 seek 是异步的；所有 seek 必须统一走这里，避免播放循环、
@@ -511,6 +513,11 @@ final class PlayerController: ObservableObject {
     ) {
         let targetPositionMs = max(0, positionMs)
         let shouldResumePlayback = resumePlayback ?? playbackIntent
+        // 必须在发起 seek 前清掉上一句的硬边界。否则从第 1 句右划到第 2 句时，
+        // AVPlayerItem 上的 forwardPlaybackEndTime 仍是第 1 句 endMs；seek 可能
+        // 被这个旧边界截断，UI 先显示第 2 句，然后立刻又回到第 1 句。
+        cancelSentenceBoundaryGuard()
+        clearForwardPlaybackEnd()
         seekGeneration += 1
         let generation = seekGeneration
         pendingSeekPositionMs = targetPositionMs
@@ -529,9 +536,11 @@ final class PlayerController: ObservableObject {
                     self.player.play()
                 }
                 completion?()
+                // seek 已真正落在目标句后，再按目标句注册新的句尾硬边界。
+                self.refreshSentenceBoundaryGuard(force: shouldResumePlayback)
+                self.tick()
             }
         }
-        refreshSentenceBoundaryGuard()
         tick()
     }
 
@@ -760,6 +769,13 @@ final class PlayerController: ObservableObject {
             tick()
             return
         }
+        // 评测“播放原句”即使处于连续播放，也要用当前句硬边界，不能滑进下一句。
+        if stopAtSentenceEnd,
+           let index = stopAtSentenceEndIndex,
+           let sentence = mediaItem?.sentences.first(where: { $0.index == index }) {
+            applySentenceEndAction(sentence)
+            return
+        }
         // 若最后一个句子的 endMs 就是视频末尾，forwardPlaybackEndTime 和自然结束
         // 是同一时间点。此时必须执行“当前句”的句尾动作，而不能整条视频回零。
         if config.sentenceMode,
@@ -808,12 +824,57 @@ final class PlayerController: ObservableObject {
                 if item.status == .failed {
                     let message = item.error?.localizedDescription ?? "未知错误"
                     self.showTransientMessage("播放失败：\(message)")
+                } else if item.status == .readyToPlay {
+                    self.considerFullFileFallback(for: item)
                 }
                 self.tick()
             }
         }
     }
 
+    private func considerFullFileFallback(for item: AVPlayerItem) {
+        guard !didAttemptFullFileFallback,
+              let media = mediaItem,
+              !media.isLocalFile,
+              !(activePlaybackURL?.isFileURL ?? false) else {
+            return
+        }
+        let hasVisibleVideo = item.presentationSize.height >= 2 && item.presentationSize.width >= 2
+        if hasVisibleVideo {
+            return
+        }
+        didAttemptFullFileFallback = true
+        cacheFallbackTask?.cancel()
+        cacheFallbackTask = Task { @MainActor [weak self] in
+            try? await Task.sleep(nanoseconds: 800_000_000)
+            guard let self, !Task.isCancelled else {
+                return
+            }
+            guard self.mediaItem?.id == media.id,
+                  self.player.currentItem === item,
+                  item.presentationSize.height < 2 else {
+                return
+            }
+            do {
+                let cachedURL = try await StreamingVideoCache.shared.prefetch(media.url)
+                guard self.mediaItem?.id == media.id else {
+                    return
+                }
+                let shouldPlay = self.playbackIntent
+                let positionMs = self.pendingSeekPositionMs ?? self.currentPositionMs()
+                self.preparePlaybackItem(with: cachedURL)
+                self.cacheState = .ready
+                if positionMs > 0 {
+                    self.performSeek(to: positionMs, resumePlayback: shouldPlay)
+                }
+            } catch is CancellationError {
+                return
+            } catch {
+                self.showTransientMessage("视频画面加载失败：\(error.localizedDescription)")
+            }
+        }
+    }
+
     private func configureAudioSession() {
         let session = AVAudioSession.sharedInstance()
         try? session.setCategory(.playback, mode: .moviePlayback)

ios/OralTrainer/Player/PlayerSurfaceView.swift

@@ -13,15 +13,20 @@ struct PlayerSurfaceView: UIViewRepresentable {
 
     func makeUIView(context: Context) -> VideoSurfaceView {
         let view = VideoSurfaceView()
-        view.playerLayer.player = controller.player
         view.controller = controller
         view.onVerticalSwipe = onVerticalSwipe
+        view.attach(player: controller.player)
+        view.setContentHuggingPriority(.defaultLow, for: .vertical)
+        view.setContentHuggingPriority(.defaultLow, for: .horizontal)
+        view.setContentCompressionResistancePriority(.defaultLow, for: .vertical)
+        view.setContentCompressionResistancePriority(.defaultLow, for: .horizontal)
         return view
     }
 
     func updateUIView(_ uiView: VideoSurfaceView, context: Context) {
-        uiView.playerLayer.player = controller.player
+        uiView.controller = controller
         uiView.onVerticalSwipe = onVerticalSwipe
+        uiView.attach(player: controller.player)
     }
 }
 
@@ -29,13 +34,9 @@ final class VideoSurfaceView: UIView {
     weak var controller: PlayerController?
     var onVerticalSwipe: ((Bool) -> Void)?
 
-    override class var layerClass: AnyClass {
-        AVPlayerLayer.self
-    }
+    private let videoLayer = AVPlayerLayer()
 
-    var playerLayer: AVPlayerLayer {
-        layer as! AVPlayerLayer
-    }
+    var playerLayer: AVPlayerLayer { videoLayer }
 
     private let minSwipeDistance: CGFloat = 48
     private let minSwipeVelocity: CGFloat = 160
@@ -55,7 +56,11 @@ final class VideoSurfaceView: UIView {
     override init(frame: CGRect) {
         super.init(frame: frame)
         backgroundColor = .black
-        playerLayer.videoGravity = .resizeAspect
+        isOpaque = true
+        clipsToBounds = true
+        videoLayer.videoGravity = .resizeAspect
+        videoLayer.backgroundColor = UIColor.black.cgColor
+        layer.addSublayer(videoLayer)
 
         tapRecognizer = UITapGestureRecognizer(target: self, action: #selector(handleTap))
         panRecognizer = UIPanGestureRecognizer(target: self, action: #selector(handlePan))
@@ -82,6 +87,25 @@ final class VideoSurfaceView: UIView {
         fatalError("init(coder:) has not been implemented")
     }
 
+    override func layoutSubviews() {
+        super.layoutSubviews()
+        CATransaction.begin()
+        CATransaction.setDisableActions(true)
+        videoLayer.frame = bounds
+        CATransaction.commit()
+    }
+
+    func attach(player: AVPlayer) {
+        if videoLayer.player !== player {
+            videoLayer.player = player
+        }
+        if videoLayer.superlayer !== layer {
+            layer.addSublayer(videoLayer)
+        }
+        videoLayer.frame = bounds
+        videoLayer.isHidden = false
+    }
+
     @objc private func handleTap() {
         controller?.handleGesture(.singleTap)
     }
@@ -96,7 +120,6 @@ final class VideoSurfaceView: UIView {
         case .began:
             resetSwipeTracking()
         case .changed:
-            // 在滑动达到阈值时立即派发，并且同一个手势只允许派发一次。
             evaluateSwipe(dx: dx, dy: dy, velocity: velocity)
         case .ended, .cancelled, .failed:
             if !didDispatchSwipe {
@@ -112,7 +135,6 @@ final class VideoSurfaceView: UIView {
         guard !didDispatchSwipe else { return }
 
         if swipeAxis == nil {
-            // 只有方向足够明确时锁定轴向，避免轻微斜滑被误判为上下滑。
             if abs(dx) > abs(dy) * 1.2 {
                 swipeAxis = .horizontal
             } else if abs(dy) > abs(dx) * 1.2 {

ios/OralTrainer/Resources/Info.plist

@@ -33,6 +33,8 @@
 	<string>需要麦克风权限才能进行录音评测</string>
 	<key>UILaunchScreen</key>
 	<dict/>
+	<key>UIRequiresFullScreen</key>
+	<true/>
 	<key>UISupportedInterfaceOrientations</key>
 	<array>
 		<string>UIInterfaceOrientationPortrait</string>

ios/OralTrainer/Services/ImitationAssessor.swift

@@ -88,16 +88,20 @@ enum ImitationAssessor {
         appendFile(&body, boundary: boundary, name: "audio", filename: filename, mimeType: mimeType, data: recordingData)
         body.append(Data("--\(boundary)--\r\n".utf8))
 
+        var headers = [
+            "Accept": "application/json",
+            "Content-Type": "multipart/form-data; boundary=\(boundary)",
+            "X-Client-Key": AppConfig.assessmentAPIKey,
+        ]
+        if !AppConfig.authToken.isEmpty {
+            headers["Authorization"] = "Bearer \(AppConfig.authToken)"
+        }
         let result = try await HTTPClient.execute(
             baseURL: AppConfig.serverBaseURL,
             path: path,
             method: "POST",
             timeout: AppConfig.assessmentTimeout,
-            headers: [
-                "Accept": "application/json",
-                "Content-Type": "multipart/form-data; boundary=\(boundary)",
-                "X-Client-Key": AppConfig.assessmentAPIKey,
-            ],
+            headers: headers,
             body: body
         )
         guard (200...299).contains(result.statusCode) else {

ios/OralTrainer/Services/SentenceBoundaryAPI.swift

@@ -46,6 +46,40 @@ struct SentenceExplanationResponse: Decodable {
     }
 }
 
+struct SubtitleTranslationItem: Decodable {
+    let index: Int
+    let zhText: String
+
+    private enum CodingKeys: String, CodingKey {
+        case index
+        case zhText = "zh_text"
+    }
+}
+
+struct SubtitleTranslationResponse: Decodable {
+    let videoHash: String
+    let translations: [SubtitleTranslationItem]
+    let complete: Bool?
+    let status: String?
+    let error: String?
+
+    private enum CodingKeys: String, CodingKey {
+        case translations
+        case videoHash = "video_hash"
+        case complete
+        case status
+        case error
+    }
+
+    var mapped: [Int: String] {
+        var values: [Int: String] = [:]
+        for item in translations where !item.zhText.isEmpty {
+            values[item.index] = item.zhText
+        }
+        return values
+    }
+}
+
 extension SentenceBoundaryAPI {
     static func explain(
         _ sentence: SentenceBoundary,
@@ -79,4 +113,110 @@ extension SentenceBoundaryAPI {
             .decode(SentenceExplanationResponse.self, from: result.body)
             .explanation
     }
+
+    static func fetchStoredSubtitles(videoHash: String) async throws -> SubtitleTranslationResponse {
+        try await requestSubtitles(videoHash: videoHash, method: "GET", timeout: AppConfig.readTimeout)
+    }
+
+    static func translateSubtitles(videoHash: String) async throws -> SubtitleTranslationResponse {
+        try await requestSubtitles(videoHash: videoHash, method: "POST", timeout: AppConfig.translationTimeout)
+    }
+
+    static func loadSubtitles(
+        videoHash: String,
+        expectedIndexes: [Int],
+        onProgress: (([Int: String]) -> Void)? = nil
+    ) async throws -> [Int: String] {
+        let expected = Set(expectedIndexes)
+        let deadline = Date().addingTimeInterval(AppConfig.translationDeadline)
+        var kickJob = true
+        var lastError: Error?
+        var lastCount = -1
+        var stalledRounds = 0
+        while Date() < deadline {
+            try Task.checkCancellation()
+            if let stored = try? await fetchStoredSubtitles(videoHash: videoHash) {
+                let mapped = stored.mapped
+                onProgress?(mapped)
+                if isComplete(mapped, expected: expected, payload: stored) {
+                    return mapped
+                }
+                if mapped.count > lastCount {
+                    lastCount = mapped.count
+                    stalledRounds = 0
+                } else {
+                    stalledRounds += 1
+                }
+                if stored.status != "running" && stalledRounds >= 2 {
+                    kickJob = true
+                }
+            } else {
+                kickJob = true
+            }
+            if kickJob {
+                do {
+                    let posted = try await translateSubtitles(videoHash: videoHash)
+                    let mapped = posted.mapped
+                    onProgress?(mapped)
+                    if isComplete(mapped, expected: expected, payload: posted) {
+                        return mapped
+                    }
+                    if posted.status == "error" {
+                        lastError = HTTPClientError.badStatus(503, posted.error ?? "Subtitle translation failed")
+                    }
+                } catch {
+                    lastError = error
+                    if !isTimeout(error) {
+                        stalledRounds += 1
+                    }
+                }
+                kickJob = false
+            }
+            try await Task.sleep(nanoseconds: AppConfig.translationPollIntervalNs)
+        }
+        throw lastError ?? HTTPClientError.badStatus(408, "Subtitle translation timed out")
+    }
+
+    private static func isComplete(
+        _ mapped: [Int: String],
+        expected: Set<Int>,
+        payload: SubtitleTranslationResponse
+    ) -> Bool {
+        if !expected.isEmpty {
+            return expected.isSubset(of: Set(mapped.keys))
+        }
+        return payload.complete == true
+    }
+
+    private static func isTimeout(_ error: Error) -> Bool {
+        let nsError = error as NSError
+        return nsError.domain == NSURLErrorDomain && nsError.code == NSURLErrorTimedOut
+    }
+
+    private static func requestSubtitles(
+        videoHash: String,
+        method: String,
+        timeout: TimeInterval
+    ) async throws -> SubtitleTranslationResponse {
+        let hash = videoHash.lowercased()
+        guard hash.range(of: AppConfig.sha256Pattern, options: .regularExpression) != nil else {
+            throw SentenceBoundaryAPIError.invalidHash
+        }
+        var headers: [String: String] = ["Accept": "application/json"]
+        if !AppConfig.authToken.isEmpty {
+            headers["Authorization"] = "Bearer \(AppConfig.authToken)"
+        }
+        let result = try await HTTPClient.execute(
+            baseURL: AppConfig.serverBaseURL,
+            path: "api/v1/videos/\(hash)/subtitle-translation",
+            method: method,
+            timeout: timeout,
+            headers: headers
+        )
+        guard (200...299).contains(result.statusCode) else {
+            let detail = String(data: result.body, encoding: .utf8) ?? ""
+            throw HTTPClientError.badStatus(result.statusCode, detail)
+        }
+        return try JSONDecoder().decode(SubtitleTranslationResponse.self, from: result.body)
+    }
 }

ios/OralTrainer/Services/SentenceExplanationStore.swift

@@ -0,0 +1,67 @@
+import Foundation
+
+final class SentenceExplanationStore {
+    static let shared = SentenceExplanationStore()
+
+    private let root: URL
+
+    init() {
+        let documents = FileManager.default.urls(for: .documentDirectory, in: .userDomainMask)[0]
+        root = documents.appendingPathComponent("SentenceExplanations", isDirectory: true)
+        try? FileManager.default.createDirectory(at: root, withIntermediateDirectories: true)
+    }
+
+    func load(mediaId: String, sentenceIndex: Int) -> String? {
+        let url = fileURL(mediaId: mediaId, sentenceIndex: sentenceIndex)
+        guard FileManager.default.fileExists(atPath: url.path) else { return nil }
+        let text = try? String(contentsOf: url, encoding: .utf8)
+        let trimmed = text?.trimmingCharacters(in: .whitespacesAndNewlines) ?? ""
+        return trimmed.isEmpty ? nil : trimmed
+    }
+
+    @discardableResult
+    func save(mediaId: String, sentenceIndex: Int, explanation: String) -> Bool {
+        let trimmed = explanation.trimmingCharacters(in: .whitespacesAndNewlines)
+        guard !trimmed.isEmpty else { return false }
+        let url = fileURL(mediaId: mediaId, sentenceIndex: sentenceIndex)
+        try? FileManager.default.createDirectory(at: url.deletingLastPathComponent(), withIntermediateDirectories: true)
+        do {
+            try trimmed.write(to: url, atomically: true, encoding: .utf8)
+            return true
+        } catch {
+            return false
+        }
+    }
+
+    func sync(mediaId: String, sentences: [SentenceBoundary]) {
+        guard !sentences.isEmpty else { return }
+        let directory = root.appendingPathComponent(Self.sanitize(mediaId), isDirectory: true)
+        let revision = Self.revision(for: sentences)
+        let marker = directory.appendingPathComponent("revision.txt")
+        let previous = (try? String(contentsOf: marker, encoding: .utf8))?
+            .trimmingCharacters(in: .whitespacesAndNewlines) ?? ""
+        guard previous != revision else { return }
+        try? FileManager.default.removeItem(at: directory)
+        try? FileManager.default.createDirectory(at: directory, withIntermediateDirectories: true)
+        try? revision.write(to: marker, atomically: true, encoding: .utf8)
+    }
+
+    private func fileURL(mediaId: String, sentenceIndex: Int) -> URL {
+        let directory = root.appendingPathComponent(Self.sanitize(mediaId), isDirectory: true)
+        return directory.appendingPathComponent("\(sentenceIndex).txt")
+    }
+
+    static func revision(for sentences: [SentenceBoundary]) -> String {
+        sentences.map { sentence in
+            let text = (sentence.text ?? "").trimmingCharacters(in: .whitespacesAndNewlines)
+            return "\(sentence.index)\t\(text)"
+        }.joined(separator: "\n")
+    }
+
+    static func sanitize(_ mediaId: String) -> String {
+        let allowed = CharacterSet.alphanumerics.union(CharacterSet(charactersIn: "._-"))
+        let filtered = mediaId.lowercased().unicodeScalars.map { allowed.contains($0) ? Character($0) : "_" }
+        let value = String(filtered).trimmingCharacters(in: CharacterSet(charactersIn: "_"))
+        return value.isEmpty || value == "." || value == ".." ? "unknown" : value
+    }
+}

ios/OralTrainer/Services/StreamingVideoCache.swift

@@ -1,12 +1,10 @@
 import CryptoKit
 import Foundation
 
-/// iOS 端的持久化整文件视频缓存。
+/// iOS 端的可选整文件视频缓存。
 ///
-/// Android 端使用 ExoPlayer 的 CacheDataSource 边播边缓存；AVFoundation 没有公开
-/// 等价的持久字节缓存 API。这里在后台把远程 MP4 下载成完整本地文件，下次装载时
-/// 直接交给 AVPlayer，并在本次播放完成下载后热切换到本地文件，避免回跳句子时
-/// 再次依赖网络。
+/// 播放默认走远程 URL 按需加载（AVPlayer HTTP Range）。若本地已有完整缓存文件，
+/// 播放器会优先使用该文件，避免重复拉流。
 @MainActor
 final class StreamingVideoCache {
     static let shared = StreamingVideoCache()
@@ -18,7 +16,7 @@ final class StreamingVideoCache {
 
     private let directory: URL
     private let fileManager = FileManager.default
-    private var downloadsByURL: [URL: Task<URL?, Never>] = [:]
+    private var downloadsByURL: [URL: Task<URL, Error>] = [:]
 
     private init() {
         let caches = fileManager.urls(for: .cachesDirectory, in: .userDomainMask)[0]
@@ -44,31 +42,40 @@ final class StreamingVideoCache {
         return file
     }
 
-    /// 重复调用同一 URL 时复用同一个后台下载；完成后返回本地文件。
-    func prefetch(_ remoteURL: URL) async -> URL? {
+    /// 重复调用同一 URL 时复用同一个后台下载；只有下载完成后才返回本地文件。
+    func prefetch(_ remoteURL: URL) async throws -> URL {
         if let cached = cachedFileIfAvailable(for: remoteURL) {
             return cached
         }
-        guard isCacheable(remoteURL), let destination = destinationFile(for: remoteURL) else {
-            return nil
+        guard let destination = destinationFile(for: remoteURL) else {
+            throw StreamingVideoCacheError.unsupportedURL
+        }
+        guard isCacheable(remoteURL) else {
+            throw StreamingVideoCacheError.unsupportedURL
         }
 
         if let existing = downloadsByURL[remoteURL] {
-            return await existing.value
+            return try await existing.value
         }
 
         let task = Task {
-            await Self.download(from: remoteURL, to: destination)
+            try await Self.download(from: remoteURL, to: destination)
         }
         downloadsByURL[remoteURL] = task
 
-        let result = await task.value
+        let result: URL
+        do {
+            result = try await task.value
+        } catch {
+            if !task.isCancelled {
+                downloadsByURL[remoteURL] = nil
+            }
+            throw error
+        }
         if downloadsByURL[remoteURL]?.isCancelled != true {
             downloadsByURL[remoteURL] = nil
         }
-        if result != nil {
         evictIfNeeded(keeping: remoteURL)
-        }
         return result
     }
 
@@ -137,19 +144,17 @@ final class StreamingVideoCache {
         }
     }
 
-    private static func download(from remoteURL: URL, to destination: URL) async -> URL? {
+    private static func download(from remoteURL: URL, to destination: URL) async throws -> URL {
         var request = URLRequest(url: remoteURL)
         request.timeoutInterval = AppConfig.connectTimeout
         request.setValue(AppConfig.userAgent, forHTTPHeaderField: "User-Agent")
         request.cachePolicy = .reloadIgnoringLocalCacheData
 
-        do {
         let (tempFile, response) = try await URLSession.shared.download(for: request)
+        do {
             guard let http = response as? HTTPURLResponse, (200...299).contains(http.statusCode) else {
                 let status = (response as? HTTPURLResponse)?.statusCode ?? -1
-                FileManager.default.removeItemIfPossible(at: tempFile)
-                logFailure(remoteURL, message: "HTTP \(status)")
-                return nil
+                throw StreamingVideoCacheError.downloadFailed(message: "HTTP \(status)")
             }
 
             let fileManager = FileManager.default
@@ -160,11 +165,16 @@ final class StreamingVideoCache {
             )
             try fileManager.moveItem(at: tempFile, to: destination)
             return destination
-        } catch is CancellationError {
-            return nil
         } catch {
+            FileManager.default.removeItemIfPossible(at: tempFile)
+            if let cacheError = error as? StreamingVideoCacheError {
+                logFailure(remoteURL, message: cacheError.localizedDescription)
+                throw cacheError
+            }
             logFailure(remoteURL, message: error.localizedDescription)
-            return nil
+            throw StreamingVideoCacheError.downloadFailed(
+                message: error.localizedDescription
+            )
         }
     }
 
@@ -175,6 +185,20 @@ final class StreamingVideoCache {
     }
 }
 
+enum StreamingVideoCacheError: LocalizedError {
+    case unsupportedURL
+    case downloadFailed(message: String)
+
+    var errorDescription: String? {
+        switch self {
+        case .unsupportedURL:
+            return "仅支持 MP4/M4V/MOV 整文件缓存"
+        case .downloadFailed(let message):
+            return message
+        }
+    }
+}
+
 private extension FileManager {
     @discardableResult
     func removeItemIfPossible(at url: URL) -> Bool {

ios/OralTrainer/Views/CatalogView.swift

@@ -10,6 +10,7 @@ final class CatalogViewModel: ObservableObject {
 
     private var baseURL = AppConfig.serverBaseURL
     private var hasLoaded = false
+    private var boundaryDocuments: [String: SentenceBoundaryDocument] = [:]
 
     func loadIfNeeded() async {
         guard !hasLoaded else {
@@ -31,6 +32,42 @@ final class CatalogViewModel: ObservableObject {
         }
     }
 
+    func preloadBoundaries(for videos: [TrainingVideoSummary]) async -> (
+        videos: [TrainingVideoSummary],
+        succeeded: Int,
+        failed: Int
+    ) {
+        var updatedVideos = videos
+        var succeeded = 0
+        var failed = 0
+
+        for (index, video) in videos.enumerated() {
+            do {
+                let document = try await SentenceBoundaryAPI.fetch(videoHash: video.videoHash)
+                boundaryDocuments[document.videoHash] = document
+                SentenceExplanationStore.shared.sync(
+                    mediaId: document.videoHash,
+                    sentences: document.sentences
+                )
+                updatedVideos[index] = video.copy(
+                    durationMs: document.durationMs,
+                    sentenceCount: document.sentences.count
+                )
+                if activeVideoHash == video.videoHash {
+                    controller?.updateBoundaries(document.sentences)
+                    controller?.boundaryState = .loaded(count: document.sentences.count)
+                    controller?.showTransientMessage("课程已就绪：\(document.sentences.count) 句")
+                }
+                succeeded += 1
+            } catch {
+                failed += 1
+            }
+        }
+
+        videos = updatedVideos
+        return (updatedVideos, succeeded, failed)
+    }
+
     /// 与安卓端 loadRemoteVideo() 一致：先以无句边界装载，再拉取句边界后重新装载。
     func select(_ video: TrainingVideoSummary) {
         guard let controller = controller else {
@@ -41,6 +78,23 @@ final class CatalogViewModel: ObservableObject {
             controller.boundaryState = .failed(message: "无法解析视频流地址")
             return
         }
+        if let document = boundaryDocuments[video.videoHash] {
+            SentenceExplanationStore.shared.sync(
+                mediaId: document.videoHash,
+                sentences: document.sentences
+            )
+            let item = TrainingMediaItem(
+                id: video.videoHash,
+                title: video.title,
+                url: url,
+                sentences: document.sentences
+            )
+            controller.load(item)
+            controller.boundaryState = .loaded(count: document.sentences.count)
+            controller.showTransientMessage("课程已就绪：\(document.sentences.count) 句")
+            return
+        }
+
         let loadingItem = TrainingMediaItem(id: video.videoHash, title: video.title, url: url)
         controller.load(loadingItem)
         controller.boundaryState = .loading(isLocal: false)
@@ -51,6 +105,11 @@ final class CatalogViewModel: ObservableObject {
                 guard activeVideoHash == video.videoHash else {
                     return
                 }
+                boundaryDocuments[document.videoHash] = document
+                SentenceExplanationStore.shared.sync(
+                    mediaId: document.videoHash,
+                    sentences: document.sentences
+                )
                 controller.updateBoundaries(document.sentences)
                 controller.boundaryState = .loaded(count: document.sentences.count)
                 controller.showTransientMessage("课程已就绪：\(document.sentences.count) 句")
@@ -95,6 +154,10 @@ final class CatalogViewModel: ObservableObject {
                 guard activeVideoHash == hash else {
                     return
                 }
+                SentenceExplanationStore.shared.sync(
+                    mediaId: document.videoHash,
+                    sentences: document.sentences
+                )
                 controller.updateBoundaries(document.sentences)
                 controller.boundaryState = .loaded(count: document.sentences.count)
                 controller.showTransientMessage("句子边界已加载：\(document.sentences.count) 句")

ios/OralTrainer/Views/FullscreenPlayerView.swift

@@ -3,12 +3,39 @@ import SwiftUI
 /// 与安卓端横屏布局一致：全屏播放 + 半透明句面板 + 退出全屏按钮。
 struct FullscreenPlayerView: View {
     @ObservedObject var controller: PlayerController
+    var showsLessonTitle: Bool = true
+    var showChineseSubtitles: Bool = false
+    var chineseSubtitlePlaceholder: String = "正在翻译中文字幕…"
     let onExit: () -> Void
 
     @State private var subtitleVisible = true
     @State private var explanationSentence: SentenceBoundary?
 
     var body: some View {
+        GeometryReader { proxy in
+            let fakeLandscape = proxy.size.height > proxy.size.width
+            playerChrome
+                .frame(
+                    width: fakeLandscape ? proxy.size.height : proxy.size.width,
+                    height: fakeLandscape ? proxy.size.width : proxy.size.height
+                )
+                .rotationEffect(.degrees(fakeLandscape ? -90 : 0))
+                .frame(width: proxy.size.width, height: proxy.size.height)
+        }
+        .background(Color.black)
+        .ignoresSafeArea()
+        .statusBarHidden(true)
+        .sheet(item: $explanationSentence) { sentence in
+            SentenceExplanationSheet(
+                mediaId: controller.currentItem?.id ?? "",
+                courseTitle: controller.currentItem?.title ?? "口语训练课程",
+                sentence: sentence,
+                sentences: controller.currentItem?.sentences ?? []
+            )
+        }
+    }
+
+    private var playerChrome: some View {
         ZStack {
             Color.black.ignoresSafeArea()
             PlayerSurfaceView(controller: controller) { isSwipeUp in
@@ -32,18 +59,28 @@ struct FullscreenPlayerView: View {
                 Spacer()
                 if subtitleVisible {
                     VStack(alignment: .leading, spacing: 6) {
+                        if showsLessonTitle {
                             Text(controller.currentItem?.title ?? "未选择课程")
                                 .font(.system(size: 17, weight: .bold))
                                 .foregroundColor(.white)
                                 .lineLimit(1)
+                        }
                         Text(sentenceMetaText)
                             .font(.system(size: 13))
                             .foregroundColor(Theme.textOverlay)
                         HStack(alignment: .center, spacing: 8) {
+                            VStack(alignment: .leading, spacing: 4) {
                                 Text(sentenceTextText)
                                     .font(.system(size: 15))
                                     .foregroundColor(.white)
                                     .lineLimit(2)
+                                if let chineseSubtitleText = chineseSubtitleText {
+                                    Text(chineseSubtitleText)
+                                        .font(.system(size: 14))
+                                        .foregroundColor(Theme.textOverlay)
+                                        .lineLimit(2)
+                                }
+                            }
                             Spacer()
                             Button {
                                 guard let sentence = controller.currentSentence,
@@ -95,12 +132,6 @@ struct FullscreenPlayerView: View {
             }
             .padding(16)
         }
-        .sheet(item: $explanationSentence) { sentence in
-            SentenceExplanationSheet(
-                courseTitle: controller.currentItem?.title ?? "口语训练课程",
-                sentence: sentence
-            )
-        }
     }
 
     @State private var progressDragging = false
@@ -183,7 +214,24 @@ struct FullscreenPlayerView: View {
         }
     }
 
+    private var chineseSubtitleText: String? {
+        guard showChineseSubtitles else {
+            return nil
+        }
+        if let zh = controller.currentSentence?.zhText?.trimmingCharacters(in: .whitespacesAndNewlines),
+           !zh.isEmpty {
+            return zh
+        }
+        if (controller.currentSentence?.text ?? "").trimmingCharacters(in: .whitespacesAndNewlines).isEmpty {
+            return nil
+        }
+        return chineseSubtitlePlaceholder
+    }
+
     private var statusText: String {
+        if let message = controller.cacheState.statusMessage {
+            return message
+        }
         if let message = controller.transientMessage {
             return message
         }

ios/OralTrainer/Views/PlayerSectionView.swift

@@ -6,6 +6,9 @@ struct PlayerSectionView: View {
     var showsContinuousToggle: Bool
     @Binding var continuousPlayback: Bool
     var showsSubtitle: Binding<Bool>? = .constant(true)
+    var showsLessonTitle: Bool = true
+    var showChineseSubtitles: Binding<Bool>? = nil
+    var chineseSubtitlePlaceholder: String = "正在翻译中文字幕…"
     var onFullscreen: () -> Void
 
     @State private var progressDragging = false
@@ -22,8 +25,10 @@ struct PlayerSectionView: View {
         .padding(.horizontal, 16)
         .sheet(item: $explanationSentence) { sentence in
             SentenceExplanationSheet(
+                mediaId: controller.currentItem?.id ?? "",
                 courseTitle: controller.currentItem?.title ?? "口语训练课程",
-                sentence: sentence
+                sentence: sentence,
+                sentences: controller.currentItem?.sentences ?? []
             )
         }
     }
@@ -73,10 +78,12 @@ struct PlayerSectionView: View {
 
     private var sentencePanel: some View {
         VStack(alignment: .leading, spacing: 6) {
+            if showsLessonTitle {
                 Text(controller.currentItem?.title ?? "未选择课程")
                     .font(.system(size: 17, weight: .bold))
                     .foregroundColor(.white)
                     .lineLimit(1)
+            }
             if showsContinuousToggle {
                 HStack(spacing: 0) {
                     Text("连续播放")
@@ -89,14 +96,34 @@ struct PlayerSectionView: View {
                 }
                 .padding(.vertical, 2)
             }
+            if let showChineseSubtitles = showChineseSubtitles {
+                HStack(spacing: 0) {
+                    Text("中文字幕")
+                        .font(.system(size: 13, weight: .bold))
+                        .foregroundColor(Theme.textSubtle)
+                    Spacer()
+                    Toggle("", isOn: showChineseSubtitles)
+                        .labelsHidden()
+                        .tint(Theme.accent)
+                }
+                .padding(.vertical, 2)
+            }
             Text(sentenceMetaText)
                 .font(.system(size: 13))
                 .foregroundColor(Theme.textSubtle)
             HStack(alignment: .center, spacing: 8) {
+                VStack(alignment: .leading, spacing: 4) {
                     Text(sentenceTextText)
                         .font(.system(size: 15))
                         .foregroundColor(.white)
-                    .lineLimit(2)
+                        .lineLimit(4)
+                    if let chineseSubtitleText = chineseSubtitleText {
+                        Text(chineseSubtitleText)
+                            .font(.system(size: 14))
+                            .foregroundColor(Theme.textMuted)
+                            .lineLimit(4)
+                    }
+                }
                 Spacer()
                 Button {
                     guard let sentence = controller.currentSentence, !(sentence.text ?? "").isEmpty else {
@@ -201,7 +228,24 @@ struct PlayerSectionView: View {
         }
     }
 
+    private var chineseSubtitleText: String? {
+        guard showChineseSubtitles?.wrappedValue == true else {
+            return nil
+        }
+        if let zh = controller.currentSentence?.zhText?.trimmingCharacters(in: .whitespacesAndNewlines),
+           !zh.isEmpty {
+            return zh
+        }
+        if (controller.currentSentence?.text ?? "").trimmingCharacters(in: .whitespacesAndNewlines).isEmpty {
+            return nil
+        }
+        return chineseSubtitlePlaceholder
+    }
+
     private var statusText: String {
+        if let message = controller.cacheState.statusMessage {
+            return message
+        }
         if let message = controller.transientMessage {
             return message
         }
@@ -227,11 +271,15 @@ struct PlayerSectionView: View {
 }
 
 struct SentenceExplanationSheet: View {
+    let mediaId: String
     let courseTitle: String
     let sentence: SentenceBoundary
+    var sentences: [SentenceBoundary] = []
 
+    @Environment(\.dismiss) private var dismiss
     @State private var explanation = ""
     @State private var errorMessage: String?
+    @State private var saved = false
 
     var body: some View {
         NavigationView {
@@ -249,10 +297,7 @@ struct SentenceExplanationSheet: View {
                             .font(.system(size: 15))
                             .foregroundColor(Theme.textSubtle)
                     } else {
-                        Text(explanation)
-                            .font(.system(size: 16))
-                            .foregroundColor(Theme.textDark)
-                            .lineSpacing(5)
+                        MarkdownExplanationText(text: explanation)
                     }
                 }
                 .frame(maxWidth: .infinity, alignment: .leading)
@@ -261,8 +306,38 @@ struct SentenceExplanationSheet: View {
             .background(Color.white.ignoresSafeArea())
             .navigationTitle("句子讲解")
             .navigationBarTitleDisplayMode(.inline)
+            .toolbar {
+                ToolbarItem(placement: .navigationBarTrailing) {
+                    HStack(spacing: 6) {
+                        Button(saved ? "已保存" : "保存") {
+                            saveExplanation()
+                        }
+                        .disabled(saved || explanation.isEmpty || errorMessage != nil)
+                        Text(saved ? "（已保存本地）" : "（下次直接看本地）")
+                            .font(.system(size: 12))
+                            .foregroundColor(Theme.textMuted)
+                        Button {
+                            dismiss()
+                        } label: {
+                            Image(systemName: "xmark.circle.fill")
+                                .font(.system(size: 22, weight: .semibold))
+                                .foregroundColor(Theme.textMuted)
+                        }
+                        .accessibilityLabel("关闭句子讲解")
+                    }
+                }
+            }
         }
         .task {
+            SentenceExplanationStore.shared.sync(mediaId: mediaId, sentences: sentences)
+            if let cached = SentenceExplanationStore.shared.load(
+                mediaId: mediaId,
+                sentenceIndex: sentence.index
+            ) {
+                explanation = cached
+                saved = true
+                return
+            }
             guard explanation.isEmpty, errorMessage == nil else { return }
             do {
                 explanation = try await SentenceBoundaryAPI.explain(
@@ -274,4 +349,64 @@ struct SentenceExplanationSheet: View {
             }
         }
     }
+
+    private func saveExplanation() {
+        guard !saved, !explanation.isEmpty, errorMessage == nil else { return }
+        saved = SentenceExplanationStore.shared.save(
+            mediaId: mediaId,
+            sentenceIndex: sentence.index,
+            explanation: explanation
+        )
+    }
+}
+
+private struct MarkdownExplanationText: View {
+    let text: String
+
+    var body: some View {
+        Text(Self.attributed(from: text))
+            .font(.system(size: 16))
+            .foregroundColor(Theme.textDark)
+            .tint(Theme.buttonBlue)
+            .lineSpacing(5)
+            .frame(maxWidth: .infinity, alignment: .leading)
+    }
+
+    private static func attributed(from source: String) -> AttributedString {
+        let options = AttributedString.MarkdownParsingOptions(
+            interpretedSyntax: .full,
+            failurePolicy: .returnPartiallyParsedIfPossible
+        )
+        if let parsed = try? AttributedString(
+            markdown: preservingLineBreaks(source),
+            options: options
+        ) {
+            return parsed
+        }
+        return AttributedString(source)
+    }
+
+    private static func preservingLineBreaks(_ source: String) -> String {
+        let normalized = source
+            .replacingOccurrences(of: "\r\n", with: "\n")
+            .replacingOccurrences(of: "\r", with: "\n")
+        var result = ""
+        var inFence = false
+        let lines = normalized.split(separator: "\n", omittingEmptySubsequences: false)
+        for (index, rawLine) in lines.enumerated() {
+            let line = String(rawLine)
+            let trimmed = line.trimmingCharacters(in: .whitespaces)
+            if trimmed.hasPrefix("```") {
+                inFence.toggle()
+            }
+            result += line
+            guard index < lines.count - 1 else { continue }
+            if inFence || trimmed.isEmpty {
+                result += "\n"
+            } else {
+                result += "\n\n"
+            }
+        }
+        return result
+    }
 }

ios/OralTrainer/Views/RootView.swift

@@ -9,11 +9,16 @@ struct RootView: View {
     @StateObject private var historyManager = TrainingHistoryManager()
     @State private var activeModule: Module = .course
     @State private var continuousPlayback = false
+    @State private var showChineseSubtitles = false
+    @State private var chineseTranslationLoading = false
+    @State private var chineseTranslationError = ""
+    @State private var chineseTranslationTask: Task<Void, Never>?
     @State private var showImporter = false
     @State private var showFullscreenPlayer = false
     @State private var courses: [Course] = []
     @State private var currentCourse: Course?
     @State private var courseStatusText = "正在同步"
+    @State private var isRefreshingCourses = false
     @State private var showCourseGuardToast = false
     @State private var managedResults: [UserResultSummary] = []
     @State private var managedDubShares: [UserDubShareSummary] = []
@@ -52,6 +57,8 @@ struct RootView: View {
                         controller: controller,
                         catalogVM: catalogVM,
                         continuousPlayback: $continuousPlayback,
+                        showChineseSubtitles: $showChineseSubtitles,
+                        chineseSubtitlePlaceholder: chineseSubtitlePlaceholder,
                         onFullscreen: enterFullscreen
                     )
                 case .test:
@@ -87,9 +94,27 @@ struct RootView: View {
                 Task { await loadCourses() }
             }
         }
+        .onChange(of: showChineseSubtitles) { enabled in
+            if enabled {
+                ensureChineseTranslations()
+            } else {
+                chineseTranslationTask?.cancel()
+                chineseTranslationLoading = false
+            }
+        }
+        .onChange(of: controller.currentItem?.id) { _ in
+            ensureChineseTranslations()
+        }
+        .onChange(of: controller.currentItem?.sentences.count) { _ in
+            ensureChineseTranslations()
+        }
         .onChange(of: activeModule) { module in
             if module == .train {
                 controller.setContinuousPlayback(continuousPlayback)
+                ensureChineseTranslations()
+            } else if module == .test {
+                // 与安卓测试页一致：按当前句窗口播放，避免滑进下一句。
+                controller.setContinuousPlayback(false)
             } else {
                 controller.setContinuousPlayback(true)
             }
@@ -109,10 +134,17 @@ struct RootView: View {
             syncCurrentCourse(with: videoHash)
         }
         .fullScreenCover(isPresented: $showFullscreenPlayer) {
-            FullscreenPlayerView(controller: controller) {
+            FullscreenPlayerView(
+                controller: controller,
+                showsLessonTitle: activeModule != .train,
+                showChineseSubtitles: showChineseSubtitles,
+                chineseSubtitlePlaceholder: chineseSubtitlePlaceholder
+            ) {
                 exitFullscreen()
             }
-            .onAppear(perform: enterFullscreen)
+            .onAppear {
+                AppDelegate.lockOrientation(.landscape)
+            }
         }
     }
 
@@ -203,9 +235,27 @@ struct RootView: View {
                     .font(.system(size: 26, weight: .bold))
                     .foregroundColor(.white)
                     .padding(.horizontal, 16)
+                HStack(spacing: 10) {
                     Text("选择一门课程开始训练")
                         .font(.system(size: 13))
                         .foregroundColor(Theme.textMuted)
+                    Spacer()
+                    Button {
+                        Task { await loadCourses() }
+                    } label: {
+                        Text(isRefreshingCourses ? "刷新中" : "刷新课程")
+                            .font(.system(size: 13, weight: .bold))
+                            .foregroundColor(.white)
+                            .padding(.horizontal, 12)
+                            .frame(height: 32)
+                            .background(Theme.buttonBlue)
+                            .clipShape(RoundedRectangle(cornerRadius: 8))
+                    }
+                    .buttonStyle(.plain)
+                    .disabled(isRefreshingCourses)
+                    .opacity(isRefreshingCourses ? 0.6 : 1)
+                    .padding(.trailing, 16)
+                }
                 .padding(.horizontal, 16)
                 .padding(.top, 4)
                 Text(courseStatusText)
@@ -504,10 +554,50 @@ struct RootView: View {
     // MARK: - Load Courses
 
     private func loadCourses() async {
+        guard !isRefreshingCourses else { return }
+        isRefreshingCourses = true
+        defer { isRefreshingCourses = false }
         courseStatusText = "正在同步"
         do {
-            courses = try await AuthAPI.courses(token: auth.token)
-            courseStatusText = "\(courses.count) 个课程"
+            var fetched = try await AuthAPI.courses(token: auth.token)
+            var boundarySucceeded = 0
+            var boundaryFailed = 0
+            courses = fetched
+            if !fetched.isEmpty {
+                courseStatusText = "\(fetched.count) 个课程，正在刷新句界"
+                let refreshed = await catalogVM.preloadBoundaries(
+                    for: fetched.map { $0.toVideoSummary() }
+                )
+                boundarySucceeded = refreshed.succeeded
+                boundaryFailed = refreshed.failed
+                let refreshedByHash = Dictionary(uniqueKeysWithValues: refreshed.videos.map { ($0.videoHash, $0) })
+                fetched = fetched.map { course in
+                    guard let video = refreshedByHash[course.videoHash] else { return course }
+                    return Course(
+                        videoHash: course.videoHash,
+                        title: course.title,
+                        durationMs: video.durationMs,
+                        language: course.language,
+                        sentenceCount: video.sentenceCount,
+                        streamUrl: course.streamUrl,
+                        enrolled: course.enrolled
+                    )
+                }
+                if currentCourse != nil,
+                   let video = refreshed.videos.first(where: { $0.videoHash == currentCourse?.videoHash }) {
+                    currentCourse = fetched.first(where: { $0.videoHash == video.videoHash })
+                }
+            }
+            courses = fetched
+            if fetched.isEmpty {
+                courseStatusText = "0 个课程"
+            } else if boundaryFailed == 0 {
+                courseStatusText = "\(fetched.count) 个课程 · 句界已刷新"
+            } else if boundarySucceeded == 0 {
+                courseStatusText = "\(fetched.count) 个课程 · 句界刷新失败"
+            } else {
+                courseStatusText = "\(fetched.count) 个课程 · 句界部分刷新（失败 \(boundaryFailed)）"
+            }
         } catch {
             courseStatusText = "课程加载失败：\(error.localizedDescription)"
         }
@@ -516,25 +606,84 @@ struct RootView: View {
     // MARK: - Fullscreen
 
     private func enterFullscreen() {
-        AppDelegate.orientationLock = UIDevice.current.userInterfaceIdiom == .pad ? .all : .landscape
-        updateSupportedInterfaceOrientations()
+        AppDelegate.lockOrientation(.landscape)
         showFullscreenPlayer = true
     }
 
     private func exitFullscreen() {
-        AppDelegate.orientationLock = UIDevice.current.userInterfaceIdiom == .pad ? .all : .portrait
-        updateSupportedInterfaceOrientations()
+        let restored: UIInterfaceOrientationMask =
+            UIDevice.current.userInterfaceIdiom == .pad ? .all : .portrait
+        AppDelegate.lockOrientation(restored)
         showFullscreenPlayer = false
     }
 
-    private func updateSupportedInterfaceOrientations() {
-        let scenes = UIApplication.shared.connectedScenes.compactMap { $0 as? UIWindowScene }
-        let controllers = scenes.flatMap { $0.windows }.compactMap(\.rootViewController)
-        for controller in controllers {
-            if #available(iOS 16.0, *) {
-                controller.setNeedsUpdateOfSupportedInterfaceOrientations()
-            } else {
-                UIViewController.attemptRotationToDeviceOrientation()
+    private var chineseSubtitlePlaceholder: String {
+        if chineseTranslationLoading {
+            return "正在翻译中文字幕…"
+        }
+        if !chineseTranslationError.isEmpty {
+            return chineseTranslationError
+        }
+        return "暂无中文字幕"
+    }
+
+    @MainActor
+    private func applyChineseTranslations(_ map: [Int: String], mediaId: String) {
+        guard controller.currentItem?.id == mediaId else { return }
+        let updated = (controller.currentItem?.sentences ?? []).map { sentence in
+            sentence.withZhText(map[sentence.index] ?? sentence.zhText)
+        }
+        controller.updateBoundaries(updated)
+    }
+
+    private func ensureChineseTranslations() {
+        guard showChineseSubtitles, activeModule == .train else {
+            return
+        }
+        guard let item = controller.currentItem else {
+            return
+        }
+        let needsTranslation = item.sentences.contains {
+            !($0.text ?? "").trimmingCharacters(in: .whitespacesAndNewlines).isEmpty &&
+                ($0.zhText ?? "").trimmingCharacters(in: .whitespacesAndNewlines).isEmpty
+        }
+        guard needsTranslation else {
+            chineseTranslationLoading = false
+            chineseTranslationError = ""
+            return
+        }
+        chineseTranslationTask?.cancel()
+        chineseTranslationLoading = true
+        chineseTranslationError = ""
+        let mediaId = item.id
+        chineseTranslationTask = Task {
+            do {
+                let expectedIndexes = item.sentences.compactMap { sentence -> Int? in
+                    let text = (sentence.text ?? "").trimmingCharacters(in: .whitespacesAndNewlines)
+                    return text.isEmpty ? nil : sentence.index
+                }
+                let map = try await SentenceBoundaryAPI.loadSubtitles(
+                    videoHash: mediaId,
+                    expectedIndexes: expectedIndexes
+                ) { partial in
+                    Task { @MainActor in
+                        applyChineseTranslations(partial, mediaId: mediaId)
+                    }
+                }
+                guard !Task.isCancelled else { return }
+                await MainActor.run {
+                    applyChineseTranslations(map, mediaId: mediaId)
+                    chineseTranslationLoading = false
+                    chineseTranslationError = ""
+                    controller.showTransientMessage("已加载中文字幕")
+                }
+            } catch {
+                guard !Task.isCancelled else { return }
+                await MainActor.run {
+                    chineseTranslationLoading = false
+                    chineseTranslationError = "中文字幕翻译失败"
+                    controller.showTransientMessage("中文字幕翻译失败：\(error.localizedDescription)")
+                }
             }
         }
     }

ios/OralTrainer/Views/TestView.swift

@@ -35,7 +35,7 @@ struct TestView: View {
                 PlayerSectionView(
                     controller: controller,
                     showsContinuousToggle: false,
-                    continuousPlayback: .constant(true),
+                    continuousPlayback: .constant(false),
                     showsSubtitle: $testSubtitleVisible,
                     onFullscreen: onFullscreen
                 )
@@ -321,7 +321,7 @@ struct TestView: View {
         dubSegments[sentence.index] = audioURL
         dubRawSegments[sentence.index] = rawPersistedURL
         mergedDubURL = nil
-        submitAssessment(recordingURL: url)
+        submitAssessment(recordingURL: audioURL)
     }
 
     private func restorePlaybackAudioSession() {

ios/OralTrainer/Views/TrainingView.swift

@@ -4,6 +4,8 @@ struct TrainingView: View {
     @ObservedObject var controller: PlayerController
     @ObservedObject var catalogVM: CatalogViewModel
     @Binding var continuousPlayback: Bool
+    @Binding var showChineseSubtitles: Bool
+    var chineseSubtitlePlaceholder: String = "正在翻译中文字幕…"
     let onFullscreen: () -> Void
 
     var body: some View {
@@ -13,6 +15,9 @@ struct TrainingView: View {
                     controller: controller,
                     showsContinuousToggle: true,
                     continuousPlayback: $continuousPlayback,
+                    showsLessonTitle: false,
+                    showChineseSubtitles: $showChineseSubtitles,
+                    chineseSubtitlePlaceholder: chineseSubtitlePlaceholder,
                     onFullscreen: onFullscreen
                 )
                 CatalogView(vm: catalogVM)
@@ -27,5 +32,8 @@ struct TrainingView: View {
                     : "已关闭连续播放，将循环播放当前句子"
             )
         }
+        .onChange(of: showChineseSubtitles) { enabled in
+            controller.showTransientMessage(enabled ? "已开启中文字幕" : "已关闭中文字幕")
+        }
     }
 }

ios/README.md

@@ -3,7 +3,8 @@
 iOS 端与 [android/](../android/README.md) 示例 App 功能对齐的移动端：
 
 - 在线视频目录（`GET /api/v1/videos`，仅展示 `ready` 课程，自动解析相对流地址）
-- AVPlayer 在线播放 + 本地缓存导入；单击播放/暂停
+- 云端视频先完整下载为本地整文件缓存，下载完成后才允许 AVPlayer 播放；
+  支持本地缓存导入；单击播放/暂停
 - 左划 / 右划：上一句 / 下一句（无句边界时回退为 10 秒快退 / 快进）
 - 长按播放区：按触摸点水平位置变速（左慢右快，0.5x - 2.0x），松手恢复原速
 - 按句跳转：句边界来自 `GET /api/v1/videos/{sha256}/sentence-boundaries`

moss/README.md

@@ -0,0 +1,468 @@
+# MOSS-Transcribe-Diarize 转写服务部署
+
+基于 [OpenMOSS/MOSS-Transcribe-Diarize](https://github.com/OpenMOSS/MOSS-Transcribe-Diarize)
+0.9B 的端到端语音转写服务。单模型联合完成转写与说话人分离，输出带
+`[S01]`、`[S02]` 等说话人标签和精确时间戳的结构化转写结果，适合多人对话、
+会议、播客、访谈等多说话人场景。
+
+与 Whisper 的区别：Whisper 只做语音转文字，不区分说话人；MOSS 在转写的
+同时输出说话人标签，`sentence_api` 无需额外集成即可自动按说话人分句。
+
+## 目录结构
+
+```text
+moss/
+  README.md                        本文档
+  systemd/moss-transcribe.service  systemd 服务模板
+  verify.sh                        部署验证脚本
+```
+
+## 硬件与系统要求
+
+| 项目 | 要求 |
+|------|------|
+| GPU  | NVIDIA，显存 ≥ 8 GB（推荐 16 GB 以上） |
+| CUDA | 12.9+（vLLM 方案）或 13.0+（SGLang Omni 方案） |
+| 驱动 | ≥ 535（支持 CUDA 12.x）或 ≥ 570（支持 CUDA 13） |
+| 系统 | Ubuntu 22.04 / 24.04 |
+| Python | 3.12 |
+| 其他 | FFmpeg、Git、curl |
+
+模型权重约 2 GB（bf16），推理时加上音频编码器和 KV 缓存实际占用约 3–6 GB。
+如果与其他 GPU 服务共存，需确保剩余显存充足。
+
+## 推理后端选择
+
+MOSS 官方提供两种推理后端，按 CUDA 版本选择：
+
+| 后端 | CUDA 要求 | 适用场景 |
+|------|-----------|----------|
+| **SGLang Omni**（推荐） | CUDA 13.0+ | 官方推荐后端，长音频性能更好 |
+| **vLLM** | CUDA 12.x / 13.x | CUDA 12 环境或 SGLang 安装遇到问题时使用 |
+
+两者提供相同的 OpenAI 兼容接口（`POST /v1/audio/transcriptions`），
+切换后端不需要改 `sentence_api` 的任何配置。
+
+---
+
+## 方案 A：SGLang Omni 部署（CUDA 13，推荐）
+
+### 前置条件
+
+确认驱动支持 CUDA 13：
+
+```bash
+nvidia-smi
+# 查看 CUDA Version 列，需 ≥ 13.0
+```
+
+### 安装系统依赖
+
+```bash
+sudo apt update
+sudo apt install -y ffmpeg git curl
+sudo apt install -y ninja-build
+```
+
+### 安装 uv
+
+```bash
+curl -LsSf https://astral.sh/uv/install.sh | sh
+# 重新加载 shell 或执行 source ~/.bashrc 使 uv 命令生效
+```
+
+### 创建目录和虚拟环境
+
+```bash
+sudo mkdir -p /opt/moss-transcribe
+sudo chown "$USER":"$USER" /opt/moss-transcribe
+cd /opt/moss-transcribe
+
+git clone https://github.com/OpenMOSS/MOSS-Transcribe-Diarize.git
+cd MOSS-Transcribe-Diarize
+
+uv venv --python 3.12 .venv
+source .venv/bin/activate
+```
+
+### 安装推理框架
+
+```bash
+uv pip install -e ".[torch-runtime]" --torch-backend=auto
+uv pip install sglang-omni
+uv pip install ninja
+```
+
+> `ninja` 是 FlashInfer JIT 编译 CUDA kernel 所需的构建工具，缺少它会在启动时
+> 报 `FileNotFoundError: 'ninja'`。
+>
+> 必须用 `apt install ninja-build` 安装系统级 ninja（`/usr/bin/ninja`）。
+> `uv pip install ninja` 只安装到虚拟环境 `.venv/bin/`，systemd 服务的 PATH
+> 不包含它。如果只能用 venv 里的 ninja，在 systemd 服务的 `[Service]` 段加：
+>
+> ```ini
+> Environment=PATH=/opt/moss-transcribe/MOSS-Transcribe-Diarize/.venv/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
+> ```
+
+### 下载模型
+
+```bash
+# 国内服务器使用 HF 镜像加速
+export HF_ENDPOINT=https://hf-mirror.com
+
+hf download OpenMOSS-Team/MOSS-Transcribe-Diarize
+```
+
+模型下载到 `~/.cache/huggingface/hub/`，约 2 GB。
+
+### 启动服务
+
+```bash
+source /opt/moss-transcribe/MOSS-Transcribe-Diarize/.venv/bin/activate
+
+sgl-omni serve \
+  --model-path OpenMOSS-Team/MOSS-Transcribe-Diarize \
+  --host 127.0.0.1 \
+  --port 8001 \
+  --mem-fraction-static 0.35
+```
+
+**参数说明：**
+
+| 参数 | 默认值 | 说明 |
+|------|--------|------|
+| `--host` | `127.0.0.1` | 只监听本机，不要暴露公网 |
+| `--port` | `8001` | 服务端口（Whisper 用 9000，避免冲突） |
+| `--mem-fraction-static` | 自动 | 静态显存分配比例（作用于所有 SGLang engine stage）；与其他 GPU 服务共存时调低 |
+
+> 与 llama-server、Whisper 等其他 GPU 服务共存时，将
+> `--mem-fraction-static` 设为 `0.30`–`0.40`（约 7–10 GB），避免 OOM。
+> 独占 GPU 时可设为 `0.80` 获得最大并发能力。
+
+首次启动会加载模型到 GPU，等待日志出现 `Uvicorn running on ...` 即启动完成。
+
+---
+
+## 方案 B：vLLM 部署（CUDA 12 / 13 兼容）
+
+如果 CUDA 版本低于 13.0 或 SGLang Omni 安装遇到兼容性问题，改用 vLLM。
+
+### 前置条件
+
+同方案 A，但不要求 CUDA 13。
+
+### 安装系统依赖和 uv
+
+与方案 A 相同（见上文）。
+
+### 创建目录和虚拟环境
+
+```bash
+sudo mkdir -p /opt/moss-transcribe
+sudo chown "$USER":"$USER" /opt/moss-transcribe
+cd /opt/moss-transcribe
+
+uv venv --python 3.12 .venv
+source .venv/bin/activate
+```
+
+### 安装 vLLM
+
+vLLM 从 0.27.1 起包含 MOSS-Transcribe-Diarize 的模型注册。
+
+```bash
+# CUDA 13 环境
+uv pip install -U "vllm[audio]" --torch-backend=auto
+```
+
+如果安装的 vLLM 版本不支持 MOSS 模型（启动时报 `model not found`），
+需要使用官方指定的 pinned nightly 构建：
+
+```bash
+# CUDA 12 环境（cu129）
+uv pip install -U vllm \
+  --torch-backend=auto \
+  --extra-index-url https://wheels.vllm.ai/68b4a1d582818e67adc903bf1b8fc5a5447da2fa/cu129
+```
+
+> 上游构建地址可能随版本更新，正式部署前对照
+> [MOSS 官方仓库](https://github.com/OpenMOSS/MOSS-Transcribe-Diarize) Quickstart。
+
+### 启动服务
+
+```bash
+source /opt/moss-transcribe/.venv/bin/activate
+
+vllm serve OpenMOSS-Team/MOSS-Transcribe-Diarize \
+  --host 127.0.0.1 \
+  --port 8001 \
+  --trust-remote-code
+```
+
+> `--trust-remote-code` 是必需的：模型使用自定义 HuggingFace processor 代码。
+
+首次启动会自动从 HuggingFace 下载模型权重。国内服务器建议在启动前设置：
+
+```bash
+export HF_ENDPOINT=https://hf-mirror.com
+```
+
+---
+
+## 注册为 systemd 服务
+
+### SGLang Omni 版本
+
+复制服务文件到系统目录并修改实际路径：
+
+```bash
+sudo cp /path/to/repo/moss/systemd/moss-transcribe.service /etc/systemd/system/
+```
+
+确认 `ExecStart` 路径与实际安装位置一致：
+
+```ini
+ExecStart=/opt/moss-transcribe/MOSS-Transcribe-Diarize/.venv/bin/sgl-omni serve ...
+```
+
+### vLLM 版本
+
+将 `ExecStart` 改为：
+
+```ini
+ExecStart=/opt/moss-transcribe/.venv/bin/vllm serve OpenMOSS-Team/MOSS-Transcribe-Diarize --host 127.0.0.1 --port 8001 --trust-remote-code
+```
+
+并删除 `Environment=HF_ENDPOINT=...` 以外的 SGLang 特有参数。
+
+### 启用并启动
+
+```bash
+sudo systemctl daemon-reload
+sudo systemctl enable --now moss-transcribe
+sudo journalctl -u moss-transcribe -f   # 观察启动日志
+```
+
+首次启动会下载模型权重，看到 `running on ...` 表示服务就绪。
+
+---
+
+## 验证部署
+
+使用仓库自带的验证脚本：
+
+```bash
+cd /opt/moss-transcribe
+bash verify.sh
+
+# 用真实语音文件验证（推荐）
+bash verify.sh --file /path/to/multi-speaker.wav
+
+# 同时检查 sentence_api 集成状态
+bash verify.sh --api-url http://127.0.0.1:8000
+```
+
+或手动发起转写请求：
+
+```bash
+curl -X POST http://127.0.0.1:8001/v1/audio/transcriptions \
+  -F model=OpenMOSS-Team/MOSS-Transcribe-Diarize \
+  -F file=@/path/to/test.wav \
+  -F response_format=verbose_json \
+  -F temperature=0 \
+  -F max_new_tokens=65536
+```
+
+**正常响应**应包含：
+
+- `text`：完整转写文本
+- `segments`：分段数组，每段包含 `start`、`end`、`text`、`speaker`
+
+```json
+{
+  "text": "大家好，今天我们讨论项目进度。 好的，我先汇报我这边的进展。",
+  "segments": [
+    {"start": 0.48, "end": 2.12, "text": "大家好，今天我们讨论项目进度。", "speaker": "S01"},
+    {"start": 2.55, "end": 5.80, "text": "好的，我先汇报我这边的进展。", "speaker": "S02"}
+  ]
+}
+```
+
+**说话人标签说明：** `S01`、`S02` 等由模型自动分配，是匿名编号。
+同一说话人在同一段音频中标签一致；不同音频之间标签无关联。
+
+---
+
+## 接入 sentence_api
+
+在 `sentence_api/.env` 中确认以下配置：
+
+```dotenv
+MOSS_TRANSCRIBE_URL=http://127.0.0.1:8001/v1/audio/transcriptions
+MOSS_MODEL=OpenMOSS-Team/MOSS-Transcribe-Diarize
+```
+
+然后重建 API 容器：
+
+```bash
+cd /opt/oral-trainer/sentence_api
+docker compose up -d --build
+curl http://127.0.0.1:8000/healthz   # 确认 moss_configured 为 true
+```
+
+已上传的视频可在管理后台点击"重新处理"，重新转写后会带上说话人标签。
+
+---
+
+## 从 Whisper 切换到 MOSS
+
+如果服务器上同时部署了 Whisper（Speaches）和 MOSS，切换步骤：
+
+1. **停止 Whisper 服务**（释放约 4 GB 显存）：
+
+   ```bash
+   cd /opt/whisper
+   docker compose down
+   sudo systemctl disable --now whisper-transcribe   # 如果注册了 systemd
+   ```
+
+2. **修改 `sentence_api/.env`**：
+
+   ```dotenv
+   MOSS_TRANSCRIBE_URL=http://127.0.0.1:8001/v1/audio/transcriptions
+   MOSS_MODEL=OpenMOSS-Team/MOSS-Transcribe-Diarize
+   ```
+
+3. **重建 API**：
+
+   ```bash
+   cd /opt/oral-trainer/sentence_api
+   docker compose up -d --build
+   ```
+
+4. **验证**：
+
+   ```bash
+   curl http://127.0.0.1:8000/healthz
+   ```
+
+不需要删除 Whisper 的部署文件，后续切回只需恢复 `.env` 并重新启动。
+
+---
+
+## 显存分配参考
+
+以 NVIDIA L4（24 GB）为例，MOSS 与其他服务共存的显存分配：
+
+| 服务 | 显存占用 | 说明 |
+|------|----------|------|
+| llama-server | 8.4 GB | 保留 |
+| MOSS (SGLang, mem-fraction=0.35) | ~8 GB | 新增 |
+| Speaches (Whisper) | 4 GB | 停掉后释放 |
+| **MOSS + llama-server** | ~16 GB / 24 GB | 安全 |
+
+独占 GPU 时可将 `--mem-fraction-static` 调到 `0.80`（约 19 GB），支持更多并发。
+
+---
+
+## API 参数
+
+| 参数 | 类型 | 默认值 | 说明 |
+|------|------|--------|------|
+| `file` | file | 必填 | 上传的音频/视频文件 |
+| `model` | string | 服务端默认 | 模型标识符 |
+| `language` | string | 未设置 | 可选的语言提示（如 `zh`、`en`） |
+| `response_format` | string | `json` | `json`（原始文本）、`verbose_json`（分段）或 `text` |
+| `temperature` | float | `0.0` | 采样温度，转写建议保持 `0` |
+| `max_new_tokens` | int | `5120` | 最大生成 token 数；长音频/多人对话调大到 `65536` |
+| `prompt` | string | 未设置 | 可选的自定义指令或热词提示 |
+
+### 热词提示
+
+通过 `prompt` 参数为专有名词、人名、产品名等提供提示，可提高识别准确率：
+
+```bash
+curl -X POST http://127.0.0.1:8001/v1/audio/transcriptions \
+  -F model=OpenMOSS-Team/MOSS-Transcribe-Diarize \
+  -F file=@meeting.wav \
+  -F response_format=verbose_json \
+  -F prompt="热词提示：MOSS, Transcribe, Diarize, vLLM" \
+  -F temperature=0
+```
+
+---
+
+## 常见问题
+
+### 模型下载慢或失败
+
+```bash
+# 方法一：使用 HF 镜像
+export HF_ENDPOINT=https://hf-mirror.com
+hf download OpenMOSS-Team/MOSS-Transcribe-Diarize
+
+# 方法二：配置 HTTP 代理
+export HTTP_PROXY=http://your-proxy:port
+export HTTPS_PROXY=http://your-proxy:port
+```
+
+### 启动报 CUDA out of memory
+
+与其他 GPU 服务共存时，降低 `--mem-fraction-static`（如 `0.30`）；
+或停掉暂时不需要的服务（如 Whisper）释放显存。
+
+### 启动报 model not found
+
+vLLM 版本过低，不包含 MOSS 模型注册。使用官方指定的 pinned nightly
+（见方案 B 安装部分），或升级到 vLLM ≥ 0.27.1。
+
+### 长音频转写截断
+
+请求时添加 `-F max_new_tokens=65536`，确保解码器能完整生成整段转写。
+
+### 说话人标签偶尔混淆
+
+多人重叠说话（同时说话）时 0.9B 模型可能偶尔混淆，这是模型本身的局限。
+MOSS 采用端到端 joint diarization，比 Whisper + pyannote 拼接方案更稳定，
+但如果场景对说话人准确率要求极高，可以考虑使用在线的 MOSS Pro 版本。
+
+### 服务只监听 127.0.0.1，外部怎么访问
+
+这是有意设计——转写服务不应暴露公网。`sentence_api` 通过 Docker 内部
+网络或宿主机 `127.0.0.1` 访问即可。如果 `sentence_api` 运行在 Docker
+容器里且无法访问宿主机的 `127.0.0.1:8001`，将 `MOSS_TRANSCRIBE_URL` 改为
+`http://host.docker.internal:8001/v1/audio/transcriptions`（Linux 需在
+`docker-compose.yml` 中添加 `extra_hosts: ["host.docker.internal:host-gateway"]`）。
+
+---
+
+## 更新与回滚
+
+### 更新模型
+
+```bash
+export HF_ENDPOINT=https://hf-mirror.com
+hf download OpenMOSS-Team/MOSS-Transcribe-Diarize
+sudo systemctl restart moss-transcribe
+```
+
+### 更新推理框架
+
+```bash
+cd /opt/moss-transcribe/MOSS-Transcribe-Diarize
+source .venv/bin/activate
+uv pip install -U sglang-omni    # 或 vllm
+sudo systemctl restart moss-transcribe
+```
+
+### 回滚
+
+```bash
+# 回滚到上一个 systemd 配置
+sudo systemctl revert moss-transcribe
+
+# 或手动恢复备份的服务文件
+sudo cp /path/to/backup/moss-transcribe.service.bak /etc/systemd/system/moss-transcribe.service
+sudo systemctl daemon-reload
+sudo systemctl restart moss-transcribe
+```

moss/systemd/moss-transcribe-vllm.service

@@ -0,0 +1,16 @@
+[Unit]
+Description=MOSS Transcribe Diarize Service (vLLM)
+After=network-online.target
+
+[Service]
+Type=simple
+User=lsm
+WorkingDirectory=/opt/moss-transcribe
+Environment=HF_ENDPOINT=https://hf-mirror.com
+ExecStart=/opt/moss-transcribe/.venv/bin/vllm serve OpenMOSS-Team/MOSS-Transcribe-Diarize --host 127.0.0.1 --port 8001 --trust-remote-code
+Restart=on-failure
+RestartSec=10
+TimeoutStartSec=1800
+
+[Install]
+WantedBy=multi-user.target

moss/systemd/moss-transcribe.service

@@ -0,0 +1,20 @@
+[Unit]
+Description=MOSS Transcribe Diarize Service (SGLang Omni)
+After=network-online.target
+
+[Service]
+Type=simple
+User=lsm
+WorkingDirectory=/opt/moss-transcribe/MOSS-Transcribe-Diarize
+Environment=HF_ENDPOINT=https://hf-mirror.com
+ExecStart=/opt/moss-transcribe/MOSS-Transcribe-Diarize/.venv/bin/sgl-omni serve \
+  --model-path OpenMOSS-Team/MOSS-Transcribe-Diarize \
+  --host 127.0.0.1 \
+  --port 8001 \
+  --mem-fraction-static 0.35
+Restart=on-failure
+RestartSec=10
+TimeoutStartSec=1800
+
+[Install]
+WantedBy=multi-user.target

moss/verify.sh

@@ -0,0 +1,130 @@
+#!/usr/bin/env bash
+#
+# 验证 MOSS-Transcribe-Diarize 转写服务：
+#   1. 检查 GPU 与服务运行状态
+#   2. 用测试音频发起 verbose_json 转写请求，校验响应契约和说话人标签
+#   3. （可选）检查 oral-trainer-api 的集成状态
+#
+# 用法：
+#   ./verify.sh
+#   ./verify.sh --file /path/to/speech.wav
+#   ./verify.sh --url http://127.0.0.1:8001 --model OpenMOSS-Team/MOSS-Transcribe-Diarize
+#   ./verify.sh --api-url http://127.0.0.1:8000
+
+set -euo pipefail
+
+MOSS_URL="${MOSS_URL:-http://127.0.0.1:8001}"
+MOSS_MODEL="${MOSS_MODEL:-OpenMOSS-Team/MOSS-Transcribe-Diarize}"
+TEST_FILE=""
+API_URL=""
+
+usage() {
+    sed -n '2,9p' "$0" | sed 's/^# \{0,1\}//'
+}
+
+while [[ $# -gt 0 ]]; do
+    case "$1" in
+        --url) MOSS_URL="$2"; shift 2 ;;
+        --model) MOSS_MODEL="$2"; shift 2 ;;
+        --file) TEST_FILE="$2"; shift 2 ;;
+        --api-url) API_URL="$2"; shift 2 ;;
+        -h|--help) usage; exit 0 ;;
+        *) echo "未知参数: $1"; usage; exit 1 ;;
+    esac
+done
+
+echo "==> MOSS 地址: $MOSS_URL"
+echo "==> 模型: $MOSS_MODEL"
+
+echo "==> 检查 GPU"
+if command -v nvidia-smi >/dev/null 2>&1; then
+    nvidia-smi --query-gpu=name,memory.total,memory.used,driver_version --format=csv,noheader | head -5
+else
+    echo "警告: 未找到 nvidia-smi，无法确认 GPU 状态。"
+fi
+
+echo "==> 检查服务健康"
+if curl -fsS --max-time 10 "$MOSS_URL/health" > /dev/null 2>&1; then
+    echo "OK: /health 返回 200"
+elif curl -fsS --max-time 10 "$MOSS_URL/healthz" > /dev/null 2>&1; then
+    echo "OK: /healthz 返回 200"
+else
+    echo "警告: /health 和 /healthz 均不可达，服务可能尚未就绪。"
+fi
+
+TMP_DIR="$(mktemp -d)"
+trap 'rm -rf "$TMP_DIR"' EXIT
+
+if [[ -n "$TEST_FILE" ]]; then
+    AUDIO_FILE="$TEST_FILE"
+else
+    echo "==> 生成测试音频（3 秒 440Hz 正弦波）"
+    if ! command -v ffmpeg >/dev/null 2>&1; then
+        echo "错误: 未找到 ffmpeg，请安装或用 --file 指定真实语音文件。" >&2
+        exit 1
+    fi
+    AUDIO_FILE="$TMP_DIR/tone.wav"
+    ffmpeg -hide_banner -loglevel error -f lavfi \
+        -i "sine=frequency=440:duration=3" -ar 16000 -ac 1 -y "$AUDIO_FILE"
+fi
+
+if [[ ! -f "$AUDIO_FILE" ]]; then
+    echo "错误: 音频文件不存在: $AUDIO_FILE" >&2
+    exit 1
+fi
+
+echo "==> 发起转写请求（verbose_json）"
+START_TS="$(date +%s)"
+RESPONSE="$(curl -sS --max-time 600 \
+    -X POST "$MOSS_URL/v1/audio/transcriptions" \
+    -F "model=$MOSS_MODEL" \
+    -F "file=@$AUDIO_FILE" \
+    -F "response_format=verbose_json" \
+    -F "temperature=0" \
+    -F "max_new_tokens=65536")"
+ELAPSED="$(( $(date +%s) - START_TS ))"
+
+echo "$RESPONSE" | python3 -c '
+import json, sys
+
+payload = json.load(sys.stdin)
+text = payload.get("text")
+segments = payload.get("segments")
+
+assert isinstance(text, str), "响应缺少 text 字段"
+assert isinstance(segments, list), "响应缺少 segments 字段（需要 verbose_json）"
+
+speakers = set()
+for index, segment in enumerate(segments):
+    start = segment.get("start")
+    end = segment.get("end")
+    speaker = segment.get("speaker")
+    if not (isinstance(start, (int, float)) and isinstance(end, (int, float)) and end > start >= 0):
+        raise AssertionError(f"segments[{index}] 缺少合法 start/end")
+    if speaker:
+        speakers.add(speaker)
+
+print(f"OK: text={text!r}")
+print(f"OK: segments={len(segments)} 条")
+if speakers:
+    print(f"OK: 说话人={sorted(speakers)}")
+else:
+    print("注意: 未见 speaker 字段；单说话人或模型未输出标签（正常弦音无多人对话）。")
+'
+echo "==> 转写耗时: ${ELAPSED}s"
+
+if [[ -n "$API_URL" ]]; then
+    echo "==> 检查 API 集成 ($API_URL/healthz)"
+    curl -sS --max-time 15 "$API_URL/healthz" | python3 -c '
+import json, sys
+
+payload = json.load(sys.stdin)
+configured = payload.get("moss_configured")
+if configured is not True:
+    raise SystemExit(f"错误: moss_configured={configured!r}，请检查 .env 中的 MOSS_TRANSCRIBE_URL")
+print("OK: moss_configured=true")
+'
+fi
+
+echo "全部检查通过。建议再用真实多人对话音频复核说话人分离效果："
+echo "  $0 --file /path/to/multi-speaker.wav"

sentence_api/.env.example

@@ -9,6 +9,7 @@ MAX_VIDEO_UPLOAD_BYTES=12884901888
 MAX_ATTEMPT_UPLOAD_BYTES=52428800
 KEEP_ATTEMPT_AUDIO=false
 ASSESSMENT_PASS_SCORE=70
+LOG_LEVEL=INFO
 
 MOSS_TRANSCRIBE_URL=http://127.0.0.1:8001/v1/audio/transcriptions
 MOSS_MODEL=OpenMOSS-Team/MOSS-Transcribe-Diarize
@@ -24,6 +25,8 @@ AI_API_URL=https://api.example.com/v1/chat/completions
 AI_API_KEY=replace-with-your-ai-api-key
 AI_MODEL=gpt-4o-mini
 AI_TIMEOUT_SECONDS=45
+# 后台批量翻译可使用更长时间；实际翻译超时 = max(AI_TIMEOUT_SECONDS, AI_TRANSLATION_TIMEOUT_SECONDS)
+AI_TRANSLATION_TIMEOUT_SECONDS=120
 
 # 火山引擎对象存储（TOS / 火山 OSS）。桶请设置为公共读，或把下方地址配置为可公开访问的 CDN 域名。
 VOLCANO_OSS_ACCESS_KEY=replace-with-volcano-access-key

sentence_api/DEPLOYMENT.md

@@ -5,7 +5,8 @@
 - `oral-trainer-api`：视频上传、管理后台、边界查询、Range 视频播放和朗读评分。
 - GPU 转写服务（二选一，只在服务器内网监听）：
   - `whisper/`：faster-whisper（Speaches），推荐，部署最简单，见第 2A 节。
-  - `MOSS-Transcribe-Diarize`：带说话人分离，见第 2 节。
+  - `MOSS-Transcribe-Diarize`：带说话人分离，见第 2 节和
+    [`moss/README.md`](../moss/README.md)。
 
 推荐 Ubuntu 22.04/24.04、Python 3.12、FFmpeg、NVIDIA GPU 和已安装的 NVIDIA 驱动。MOSS 0.9B 的实际显存占用受推理框架、并发和音频长度影响，生产环境建议从 16 GB 以上显存开始验证。
 
@@ -50,6 +51,8 @@ cp -a whisper/. /opt/whisper/   # 用 whisper/. 而不是 whisper/*，否则不
 cd /opt/whisper
 
 cp .env.example .env          # 按需修改 WHISPER_MODEL
+# 默认从 DaoCloud 拉 ghcr 镜像；能直连 GitHub 时在 .env 里设
+# WHISPER_IMAGE=ghcr.io/speaches-ai/speaches:latest-cuda
 docker compose up -d
 docker compose logs -f whisper   # 首次启动下载模型（约 3 GB）
 ```
@@ -155,8 +158,9 @@ curl -X POST http://127.0.0.1:8001/v1/audio/transcriptions \
 `sentence_api` 不以 Whisper 返回的 segment 直接作为句子：文本会跨 segment 累积，
 只有碰到句号（`.`）、问号（`?`）或逗号前的单词数达到 50 时才算一句结束；
 逗号前内容不足 50 个单词时不在此处断句，会继续与后续内容合并。Whisper 在没有
-标点处断开的 segment 会自动合并成同一句。管理后台编辑句子文本时，换行并保存会把
-该句拆成多句；也支持将相邻两句合并成一句；拆分点优先使用对该句音频再次转写得到的词级时间戳，转写不可用时按
+标点处断开的 segment 会自动合并成同一句。管理后台的文本保存、换行拆分占位、相邻句
+合并和结束时间微调会立即入库，但不会触发翻译或转写；关闭句子对话框时才统一执行
+一轮拆分时间戳处理和翻译。拆分点优先使用对原句音频再次转写得到的词级时间戳，转写不可用时按
 文本长度估算。管理后台的结束时间旁支持微调按钮；调整后下一句的起始时间联动更新，
 并保证相邻句子不重叠。若转写服务支持词级时间戳
 （`timestamp_granularities[]=word`，Whisper/Speaches 支持），句子的结束时间用
@@ -194,6 +198,33 @@ AI_MODEL=gpt-4o-mini
 `http://127.0.0.1:9000/v1/audio/transcriptions`，并把 `MOSS_MODEL` 设为
 `Systran/faster-whisper-large-v3`。
 
+Dockerfile 默认使用国内镜像源，避免构建时从 Docker Hub / Debian / PyPI 直连过慢：
+
+- 基础镜像：`docker.m.daocloud.io/library/python:3.12-slim`
+- apt：`mirrors.aliyun.com`
+- pip：清华 `https://pypi.tuna.tsinghua.edu.cn/simple`
+
+如需覆盖，可在 `sentence_api/.env` 或构建命令前设置：
+
+```bash
+export PYTHON_IMAGE=python:3.12-slim   # 能直连 Docker Hub 时
+export PIP_INDEX_URL=https://mirrors.aliyun.com/pypi/simple
+export PIP_TRUSTED_HOST=mirrors.aliyun.com
+export APT_MIRROR=mirrors.tuna.tsinghua.edu.cn
+```
+
+拉 Docker Hub 官方镜像仍然慢时，把仓库里的 `sentence_api/docker-daemon.json.example`
+拷到 Docker 配置并重启：
+
+```bash
+sudo mkdir -p /etc/docker
+sudo cp sentence_api/docker-daemon.json.example /etc/docker/daemon.json
+sudo systemctl restart docker
+```
+
+`registry-mirrors` 只加速 `docker.io`。Whisper 的 `ghcr.io` 镜像请用 compose 里默认的
+DaoCloud 前缀（见第 2A 节）。
+
 启动：
 
 ```bash

sentence_api/Dockerfile

@@ -1,10 +1,27 @@
-FROM python:3.12-slim
+ARG PYTHON_IMAGE=docker.m.daocloud.io/library/python:3.12-slim
+FROM ${PYTHON_IMAGE}
+
+ARG APT_MIRROR=mirrors.aliyun.com
+ARG PIP_INDEX_URL=https://pypi.tuna.tsinghua.edu.cn/simple
+ARG PIP_TRUSTED_HOST=pypi.tuna.tsinghua.edu.cn
 
 ENV PYTHONDONTWRITEBYTECODE=1 \
     PYTHONUNBUFFERED=1 \
-    PIP_NO_CACHE_DIR=1
+    PIP_NO_CACHE_DIR=1 \
+    PIP_DISABLE_PIP_VERSION_CHECK=1 \
+    PIP_INDEX_URL=${PIP_INDEX_URL} \
+    PIP_TRUSTED_HOST=${PIP_TRUSTED_HOST}
 
-RUN apt-get update \
+RUN set -eux; \
+    if [ -f /etc/apt/sources.list.d/debian.sources ]; then \
+      sed -i "s|deb.debian.org|${APT_MIRROR}|g; s|security.debian.org|${APT_MIRROR}|g" \
+        /etc/apt/sources.list.d/debian.sources; \
+    fi; \
+    if [ -f /etc/apt/sources.list ]; then \
+      sed -i "s|deb.debian.org|${APT_MIRROR}|g; s|security.debian.org|${APT_MIRROR}|g" \
+        /etc/apt/sources.list; \
+    fi; \
+    apt-get update \
     && apt-get install -y --no-install-recommends ffmpeg ca-certificates \
     && rm -rf /var/lib/apt/lists/*
 

sentence_api/README.md

@@ -59,15 +59,30 @@ POST   /api/v1/videos/{sha256}/sentences/{index}/assessments
 POST   /api/v1/admin/videos
 PUT    /api/v1/admin/videos/raw
 POST   /api/v1/admin/videos/{sha256}/process
+POST   /api/v1/admin/videos/{sha256}/repair-gaps
+GET    /api/v1/admin/videos/{sha256}/subtitle-translation
+POST   /api/v1/admin/videos/{sha256}/subtitle-translation
+POST   /api/v1/admin/videos/{sha256}/subtitle-translation/repair
 PUT    /api/v1/admin/videos/{sha256}/sentences/{index}
+PUT    /api/v1/admin/videos/{sha256}/sentences/batch
 PUT    /api/v1/admin/videos/{sha256}/sentences/{index}/split
 PUT    /api/v1/admin/videos/{sha256}/sentences/{index}/boundary
+PUT    /api/v1/admin/users/password
 DELETE /api/v1/admin/videos/{sha256}
 ```
 
 Admin endpoints use `X-Admin-Key` when `ADMIN_API_KEY` is configured. Assessment
 requests use `X-Client-Key` when `CLIENT_API_KEY` is configured.
 
+Administrators can set a local password for an SSO-linked account with:
+
+```bash
+curl -X PUT https://your-server/api/v1/admin/users/password \
+  -H "X-Admin-Key: $ADMIN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"username":"bobcoc","password":"new-password"}'
+```
+
 `POST /api/v1/audio/sentence-boundaries` accepts an audio file from an
 authenticated client, generates boundaries in memory/temp storage only, and
 returns a sentence document. It does not register or archive the uploaded
@@ -90,6 +105,21 @@ The endpoint requires a signed-in user and accepts only the sentence and course
 context. `AI_API_URL` and `AI_API_KEY` remain on the server and are never sent
 to the mobile clients.
 
+Chinese captions are stored on the course in the server database and shared by
+every user. `GET /api/v1/videos/{sha256}/subtitle-translation` returns the saved
+captions without calling the AI, including `complete` and `status` so clients can
+poll long videos. `POST` starts missing-caption generation in the background and
+returns quickly; it translates in small windows with neighboring context, runs
+several windows in parallel, retries and splits a window if the model response is
+invalid, and writes each finished window immediately so later viewers can reuse
+it. English sentence edits clear the saved Chinese captions for that course.
+
+Desktop playback of a local file can `POST` the same endpoint with the locally
+generated English cues (`duration_ms`, `algorithm_version`, `sentences`). If that
+SHA-256 is not already a course, the server stores the English subtitles and
+translates them; later `GET`/`POST` calls poll and reuse the saved Chinese
+captions. Official course sentences are never overwritten by a local upload.
+
 ## Assessment
 
 ```bash

sentence_api/UPDATE.md

@@ -180,8 +180,8 @@ sudo systemctl restart whisper-transcribe
 | 现象 | 原因 | 处理 |
 | --- | --- | --- |
 | 更新后接口行为没变 | 忘了 `--build`，容器还在跑旧镜像 | 执行 `docker compose up -d --build` |
-| 构建时 pip 安装超时 | 服务器访问 PyPI 受限 | 给 Dockerfile 的 pip install 加 `-i https://pypi.tuna.tsinghua.edu.cn/simple` 后重新构建 |
-| 构建时拉基础镜像超时 | Docker Hub 网络问题 | 检查 `/etc/docker/daemon.json` 的 `registry-mirrors`（已配置 DaoCloud） |
+| 构建时 pip 安装超时 | 服务器访问 PyPI 受限 | Dockerfile 默认已用清华 PyPI；仍慢可改 `PIP_INDEX_URL` 后 `--build` |
+| 构建时拉基础镜像超时 | Docker Hub 网络问题 | Dockerfile 默认从 DaoCloud 拉 `python:3.12-slim`；也可把 `docker-daemon.json.example` 拷到 `/etc/docker/daemon.json` 后 `systemctl restart docker` |
 | `healthz` 里 `moss_configured: false` | `.env` 里 `MOSS_TRANSCRIBE_URL` 为空或改了未重建 | 改 `.env` 后执行 `docker compose up -d` |
 | Whisper 9000 端口连不上 | 服务未运行或容器被删 | `systemctl status whisper-transcribe`；容器被删时用 `systemctl start` 重建 |
 | API 重启后视频卡在 `processing` | 重启中断了处理流程 | 后台点“重新处理”即可 |

sentence_api/ai_tutor.py

@@ -1,7 +1,11 @@
-"""AI-backed sentence explanations for the mobile clients."""
+"""AI-backed sentence explanations and subtitle translation."""
 
+from concurrent.futures import ThreadPoolExecutor, as_completed
 from dataclasses import dataclass
+import json
 import logging
+import re
+import threading
 from urllib.parse import urlsplit
 
 import requests
@@ -9,6 +13,17 @@ import requests
 
 logger = logging.getLogger(__name__)
 
+_JSON_FENCE_RE = re.compile(r"^```(?:json)?\s*|\s*```$", re.IGNORECASE)
+_CJK_RE = re.compile("[\u3400-\u4dbf\u4e00-\u9fff\uf900-\ufaff]")
+_NUMBERED_LINE_RE = re.compile(r"^\s*[\[(]?(\d+)[\])]?[\.、:：\)]\s*(.+?)\s*$")
+_TRANSLATE_BATCH_SENTENCES = 12
+_TRANSLATE_BATCH_CHARS = 2_200
+_TRANSLATE_CONTEXT_BEFORE = 8
+_TRANSLATE_CONTEXT_AFTER = 4
+_TRANSLATE_WORKERS = 4
+_TRANSLATE_TIMEOUT_SECONDS = 120.0
+_TRANSLATE_SPLIT_THRESHOLD = 4
+
 
 class AITutorError(RuntimeError):
     """Raised when the configured AI service cannot answer a request."""
@@ -20,6 +35,7 @@ class AITutor:
     api_key: str
     model: str
     timeout_seconds: float
+    translation_timeout_seconds: float = _TRANSLATE_TIMEOUT_SECONDS
 
     @property
     def available(self) -> bool:
@@ -54,15 +70,302 @@ class AITutor:
         language_note = f"句子语言：{language.strip()}。" if language.strip() else ""
         system_prompt = (
             "你是一位耐心的外语口语老师。请用简体中文讲解用户提供的影视/课程句子。"
-            "先给出自然的中文翻译，再解释关键词、常用搭配或语法要点，并简要说明这句话适合怎样跟读。"
-            "回答使用简明清晰的分段短文，总长度不超过 300 字，不要编造视频里没有的信息。"
+            "先给出自然的中文翻译，再解释关键词、常用搭配、语法要点，并简要说明这句话适合怎样跟读。"
+            "回答使用简明清晰的分段短文，总长度不超过 600 字，不要编造视频里没有的信息。"
         )
         user_prompt = (
             f"课程：{context}\n位置：{position}\n{language_note}句子：{sentence_text}"
         )
+        return self._complete(system_prompt, user_prompt, temperature=0.3)
+
+    def translate_sentences(
+        self,
+        sentences: list[tuple[int, str]],
+        *,
+        course_title: str = "",
+        existing: dict[int, str] | None = None,
+        persist=None,
+        terms: list[tuple[str, str]] | None = None,
+    ) -> dict[int, str]:
+        if not self.available:
+            raise AITutorError("AI translation is not configured.")
+        if not sentences:
+            raise AITutorError("No sentences were provided for translation.")
+
+        ordered = [(index, text.strip()) for index, text in sentences if text.strip()]
+        if not ordered:
+            raise AITutorError("No sentence text was provided for translation.")
+
+        glossary = _normalize_translation_terms(terms)
+        source_by_index = {index: text for index, text in ordered}
+        translations = {
+            index: text.strip()
+            for index, text in (existing or {}).items()
+            if isinstance(text, str)
+            and text.strip()
+            and not translation_needs_repair(source_by_index.get(index, ""), text)
+        }
+        pending = [(index, text) for index, text in ordered if index not in translations]
+        if not pending:
+            return translations
+
+        persist_lock = threading.Lock()
+
+        def save_partial(partial: dict[int, str]) -> None:
+            if not partial:
+                return
+            translations.update(partial)
+            if persist is None:
+                return
+            with persist_lock:
+                persist(partial)
+
+        batches = _translation_batches(pending, ordered)
+        workers = min(_TRANSLATE_WORKERS, len(batches)) or 1
+        failures: list[tuple[list[tuple[int, str]], Exception]] = []
+
+        def run_batch(batch: list[tuple[int, str]]) -> dict[int, str]:
+            return self._translate_batch_resilient(
+                batch,
+                all_sentences=ordered,
+                existing=translations,
+                course_title=course_title,
+                terms=glossary,
+            )
+
+        if workers == 1:
+            for batch in batches:
+                try:
+                    save_partial(run_batch(batch))
+                except Exception as exc:  # noqa: BLE001 - collect and retry below
+                    failures.append((batch, exc))
+        else:
+            with ThreadPoolExecutor(max_workers=workers) as pool:
+                futures = {pool.submit(run_batch, batch): batch for batch in batches}
+                for future in as_completed(futures):
+                    batch = futures[future]
+                    try:
+                        save_partial(future.result())
+                    except Exception as exc:  # noqa: BLE001
+                        failures.append((batch, exc))
+
+        for batch, error in failures:
+            logger.warning(
+                "Translation batch failed, retrying with smaller windows: size=%s error=%s",
+                len(batch),
+                error,
+            )
+            save_partial(
+                self._translate_batch_resilient(
+                    batch,
+                    all_sentences=ordered,
+                    existing=translations,
+                    course_title=course_title,
+                    terms=glossary,
+                    force_split=len(batch) > 1,
+                )
+            )
+
+        missing = [index for index, _ in pending if not translations.get(index, "").strip()]
+        if missing:
+            raise AITutorError("AI translation did not cover every sentence.")
+        if glossary:
+            source_by_index = {index: text for index, text in ordered}
+            translations = {
+                index: _apply_translation_terms(source_by_index.get(index, ""), zh_text, glossary)
+                for index, zh_text in translations.items()
+            }
+        return translations
+
+    def _translate_batch_resilient(
+        self,
+        batch: list[tuple[int, str]],
+        *,
+        all_sentences: list[tuple[int, str]],
+        existing: dict[int, str],
+        course_title: str,
+        terms: list[tuple[str, str]] | None = None,
+        force_split: bool = False,
+    ) -> dict[int, str]:
+        if not batch:
+            return {}
+        if force_split and len(batch) > 1:
+            return self._translate_split_batch(
+                batch,
+                all_sentences=all_sentences,
+                existing=existing,
+                course_title=course_title,
+                terms=terms,
+            )
+
+        last_error: Exception | None = None
+        attempts = 2 if len(batch) <= _TRANSLATE_SPLIT_THRESHOLD else 1
+        for _ in range(attempts):
+            try:
+                return self._translate_batch(
+                    batch,
+                    all_sentences=all_sentences,
+                    existing=existing,
+                    course_title=course_title,
+                    terms=terms,
+                )
+            except AITutorError as exc:
+                last_error = exc
+                logger.warning(
+                    "Translation batch attempt failed: size=%s indexes=%s error=%s",
+                    len(batch),
+                    [index for index, _ in batch],
+                    exc,
+                )
+
+        if len(batch) > 1:
+            return self._translate_split_batch(
+                batch,
+                all_sentences=all_sentences,
+                existing=existing,
+                course_title=course_title,
+                terms=terms,
+            )
+        if last_error is not None:
+            raise last_error
+        raise AITutorError("AI translation did not cover every sentence.")
+
+    def _translate_split_batch(
+        self,
+        batch: list[tuple[int, str]],
+        *,
+        all_sentences: list[tuple[int, str]],
+        existing: dict[int, str],
+        course_title: str,
+        terms: list[tuple[str, str]] | None = None,
+    ) -> dict[int, str]:
+        mid = max(1, len(batch) // 2)
+        merged: dict[int, str] = {}
+        for part in (batch[:mid], batch[mid:]):
+            merged.update(
+                self._translate_batch_resilient(
+                    part,
+                    all_sentences=all_sentences,
+                    existing=existing,
+                    course_title=course_title,
+                    terms=terms,
+                )
+            )
+        return merged
+
+    def _translate_batch(
+        self,
+        batch: list[tuple[int, str]],
+        *,
+        all_sentences: list[tuple[int, str]],
+        existing: dict[int, str],
+        course_title: str,
+        terms: list[tuple[str, str]] | None = None,
+    ) -> dict[int, str]:
+        context = course_title.strip() or "口语训练课程"
+        existing = dict(existing)
+        before, after = _neighbor_context(batch, all_sentences)
+        expected_indexes = [index for index, _ in batch]
+        glossary = _normalize_translation_terms(terms)
+        system_prompt = (
+            "你是专业的影视字幕翻译。请结合前后文语境，把指定编号的外语句子翻译成自然、口语化的简体中文。"
+            "必须一句对一句，不能合并、拆分、跳过或改写编号。"
+            "只输出 JSON 对象，键是句子编号，值是对应中文，不要输出解释或 markdown。"
+        )
+        if glossary:
+            system_prompt += "必须严格使用用户指定的专有名词译法，禁止自行音译成其他写法。"
+        user_parts = [
+            f"课程：{context}",
+            f"本次需要翻译 {len(batch)} 句，编号必须与输入完全一致。",
+        ]
+        if glossary:
+            user_parts.append("专有名词译法（必须严格遵守，原文出现时译文必须使用指定中文）：")
+            user_parts.extend(f"{source} => {zh}" for source, zh in glossary)
+        previous = [
+            (index, source, existing[index])
+            for index, source in before
+            if existing.get(index, "").strip()
+        ]
+        if previous:
+            user_parts.append("已有译文（供用词一致，不要重复输出）：")
+            user_parts.extend(
+                f"[{index}] {source} => {zh}" for index, source, zh in previous[-6:]
+            )
+        elif before:
+            user_parts.append("前文（只供理解语境，不要翻译）：")
+            user_parts.extend(f"[{index}] {text}" for index, text in before)
+        if after:
+            user_parts.append("后文（只供理解语境，不要翻译）：")
+            user_parts.extend(f"[{index}] {text}" for index, text in after)
+        user_parts.append("待翻译句子：")
+        user_parts.extend(f"[{index}] {text}" for index, text in batch)
+        user_parts.append(
+            "输出示例：{" + ",".join(f'"{index}":"中文"' for index in expected_indexes[:2]) + "}"
+        )
+        user_prompt = "\n".join(user_parts)
+        content = self._complete(
+            system_prompt,
+            user_prompt,
+            temperature=0.2,
+            timeout_seconds=max(self.timeout_seconds, self.translation_timeout_seconds),
+        )
+        try:
+            parsed = _parse_translation_payload(content, expected_indexes)
+        except AITutorError:
+            logger.warning(
+                "Translation response parse failed: endpoint=%s model=%s "
+                "expected_indexes=%s prompt=%r raw_content=%r",
+                self.request_url,
+                self.model,
+                expected_indexes,
+                user_prompt[:2000],
+                content[:2000],
+            )
+            raise
+        source_by_index = {index: text for index, text in batch}
+        missing = [
+            index
+            for index in expected_indexes
+            if translation_needs_repair(source_by_index.get(index, ""), parsed.get(index, ""))
+        ]
+        if missing:
+            invalid = [
+                {
+                    "index": index,
+                    "source": source_by_index.get(index, ""),
+                    "translation": parsed.get(index, ""),
+                }
+                for index in missing
+            ]
+            logger.warning(
+                "Translation response failed validation: endpoint=%s model=%s "
+                "expected_indexes=%s parsed=%r invalid=%r prompt=%r raw_content=%r",
+                self.request_url,
+                self.model,
+                expected_indexes,
+                parsed,
+                invalid,
+                user_prompt[:2000],
+                content[:2000],
+            )
+            raise AITutorError("AI translation returned an unexpected response.")
+        return {
+            index: _apply_translation_terms(source_by_index.get(index, ""), parsed[index], glossary)
+            for index in expected_indexes
+        }
+
+    def _complete(
+        self,
+        system_prompt: str,
+        user_prompt: str,
+        *,
+        temperature: float,
+        timeout_seconds: float | None = None,
+    ) -> str:
         headers = {"Content-Type": "application/json"}
         if self.api_key:
             headers["Authorization"] = f"Bearer {self.api_key}"
+        timeout = self.timeout_seconds if timeout_seconds is None else timeout_seconds
         try:
             response = requests.post(
                 self.request_url,
@@ -73,14 +376,24 @@ class AITutor:
                         {"role": "system", "content": system_prompt},
                         {"role": "user", "content": user_prompt},
                     ],
-                    "temperature": 0.3,
+                    "temperature": temperature,
                     "stream": False,
                 },
-                timeout=self.timeout_seconds,
+                timeout=timeout,
             )
+        except requests.Timeout as exc:
+            logger.warning(
+                "AI request timed out: endpoint=%s model=%s timeout_seconds=%s",
+                self.request_url,
+                self.model,
+                timeout,
+            )
+            raise AITutorError(
+                f"AI service request timed out after {timeout:g} seconds."
+            ) from exc
         except requests.RequestException as exc:
             logger.exception(
-                "AI explanation request failed: endpoint=%s model=%s",
+                "AI request failed: endpoint=%s model=%s",
                 self.request_url,
                 self.model,
             )
@@ -88,7 +401,7 @@ class AITutor:
 
         if response.status_code < 200 or response.status_code >= 300:
             logger.warning(
-                "AI explanation returned HTTP %s: endpoint=%s model=%s body=%s",
+                "AI service returned HTTP %s: endpoint=%s model=%s body=%s",
                 response.status_code,
                 self.request_url,
                 self.model,
@@ -97,19 +410,246 @@ class AITutor:
             raise AITutorError(f"AI service returned HTTP {response.status_code}.")
         try:
             body = response.json()
-            explanation = body["choices"][0]["message"]["content"]
+            content = body["choices"][0]["message"]["content"]
         except (KeyError, IndexError, TypeError, ValueError) as exc:
             logger.warning(
-                "AI explanation returned an unexpected body: endpoint=%s body=%s",
+                "AI service returned an unexpected body: endpoint=%s body=%s",
                 self.request_url,
                 response.text[:1000],
             )
             raise AITutorError("AI service returned an unexpected response.") from exc
-        if not isinstance(explanation, str) or not explanation.strip():
+        if not isinstance(content, str) or not content.strip():
             logger.warning(
-                "AI explanation returned empty content: endpoint=%s body=%s",
+                "AI service returned empty content: endpoint=%s body=%s",
                 self.request_url,
                 response.text[:1000],
             )
             raise AITutorError("AI service returned an empty explanation.")
-        return explanation.strip()
+        return content.strip()
+
+
+def translation_needs_repair(source: str | None, zh_text: str | None) -> bool:
+    """Return True when a sentence still needs a Chinese translation."""
+    source_text = (source or "").strip()
+    if not source_text:
+        return False
+    zh = (zh_text or "").strip()
+    if not zh:
+        return True
+    if _CJK_RE.search(zh):
+        return False
+    return _CJK_RE.search(source_text) is None
+
+
+def _translation_batches(
+    pending: list[tuple[int, str]],
+    all_sentences: list[tuple[int, str]],
+) -> list[list[tuple[int, str]]]:
+    position = {index: offset for offset, (index, _) in enumerate(all_sentences)}
+    ordered_pending = sorted(pending, key=lambda item: position[item[0]])
+    runs: list[list[tuple[int, str]]] = []
+    current: list[tuple[int, str]] = []
+    last_pos: int | None = None
+    for item in ordered_pending:
+        pos = position[item[0]]
+        if current and last_pos is not None and pos != last_pos + 1:
+            runs.append(current)
+            current = []
+        current.append(item)
+        last_pos = pos
+    if current:
+        runs.append(current)
+
+    batches: list[list[tuple[int, str]]] = []
+    for run in runs:
+        batch: list[tuple[int, str]] = []
+        batch_chars = 0
+        for item in run:
+            extra = len(item[1]) + 8
+            too_many = len(batch) >= _TRANSLATE_BATCH_SENTENCES
+            too_long = batch and batch_chars + extra > _TRANSLATE_BATCH_CHARS
+            if too_many or too_long:
+                batches.append(batch)
+                batch = []
+                batch_chars = 0
+            batch.append(item)
+            batch_chars += extra
+        if batch:
+            batches.append(batch)
+    return batches
+
+
+def _neighbor_context(
+    batch: list[tuple[int, str]],
+    all_sentences: list[tuple[int, str]],
+) -> tuple[list[tuple[int, str]], list[tuple[int, str]]]:
+    position = {index: offset for offset, (index, _) in enumerate(all_sentences)}
+    first = position[batch[0][0]]
+    last = position[batch[-1][0]]
+    before = all_sentences[max(0, first - _TRANSLATE_CONTEXT_BEFORE):first]
+    after = all_sentences[last + 1:last + 1 + _TRANSLATE_CONTEXT_AFTER]
+    return before, after
+
+
+def _parse_translation_payload(content: str, expected_indexes: list[int] | int) -> dict[int, str]:
+    if isinstance(expected_indexes, int):
+        expected = list(range(expected_indexes))
+    else:
+        expected = list(expected_indexes)
+    text = _JSON_FENCE_RE.sub("", content.strip()).strip()
+    parse_errors: list[str] = []
+    for candidate in (text, *_json_snippets(text)):
+        try:
+            parsed = _translations_from_json(json.loads(candidate), expected)
+        except (TypeError, ValueError) as exc:
+            parsed = None
+            parse_errors.append(str(exc))
+        if parsed:
+            return parsed
+
+    numbered: dict[int, str] = {}
+    for line in text.splitlines():
+        match = _NUMBERED_LINE_RE.match(line)
+        if match is None:
+            continue
+        numbered[int(match.group(1))] = match.group(2).strip()
+    by_original = {
+        index: numbered[index]
+        for index in expected
+        if numbered.get(index, "").strip()
+    }
+    if len(by_original) == len(expected):
+        return by_original
+    by_order = {
+        expected[offset]: numbered[offset + 1]
+        for offset in range(len(expected))
+        if numbered.get(offset + 1, "").strip()
+    }
+    if len(by_order) == len(expected):
+        return by_order
+    logger.warning(
+        "Translation payload parser exhausted all formats: expected_indexes=%s "
+        "json_parse_errors=%r numbered_line_keys=%s content=%r",
+        expected,
+        parse_errors,
+        sorted(numbered),
+        text[:2000],
+    )
+    raise AITutorError("AI translation returned an unexpected response.")
+
+
+def _json_snippets(text: str) -> list[str]:
+    snippets: list[str] = []
+    for opener, closer in (("{", "}"), ("[", "]")):
+        start = text.find(opener)
+        while start != -1:
+            depth = 0
+            in_string = False
+            escape = False
+            for offset, char in enumerate(text[start:], start=start):
+                if in_string:
+                    if escape:
+                        escape = False
+                    elif char == "\\":
+                        escape = True
+                    elif char == '"':
+                        in_string = False
+                    continue
+                if char == '"':
+                    in_string = True
+                elif char == opener:
+                    depth += 1
+                elif char == closer:
+                    depth -= 1
+                    if depth == 0:
+                        snippets.append(text[start:offset + 1])
+                        break
+            start = text.find(opener, start + 1)
+    return snippets
+
+
+def _translations_from_json(data: object, expected: list[int]) -> dict[int, str] | None:
+    expected_set = set(expected)
+    if isinstance(data, dict):
+        nested = None
+        for key in ("translations", "sentences", "zh", "items"):
+            if key in data:
+                nested = data[key]
+                break
+        if nested is not None:
+            parsed = _translations_from_json(nested, expected)
+            if parsed:
+                return parsed
+        values: dict[int, str] = {}
+        for raw_key, raw_value in data.items():
+            try:
+                index = int(raw_key)
+            except (TypeError, ValueError):
+                continue
+            if index not in expected_set or not isinstance(raw_value, str) or not raw_value.strip():
+                continue
+            values[index] = raw_value.strip()
+        if len(values) == len(expected):
+            return values
+        return None
+    if isinstance(data, list) and len(data) == len(expected):
+        if all(isinstance(item, str) and item.strip() for item in data):
+            return {
+                index: item.strip()
+                for index, item in zip(expected, data)
+            }
+        if all(isinstance(item, dict) for item in data):
+            values = {}
+            for item in data:
+                raw_index = item.get("index")
+                zh = item.get("zh_text") or item.get("zh") or item.get("text")
+                try:
+                    index = int(raw_index)
+                except (TypeError, ValueError):
+                    continue
+                if index not in expected_set or not isinstance(zh, str) or not zh.strip():
+                    continue
+                values[index] = zh.strip()
+            if len(values) == len(expected):
+                return values
+    return None
+
+
+def _normalize_translation_terms(terms: list[tuple[str, str]] | None) -> list[tuple[str, str]]:
+    cleaned: list[tuple[str, str]] = []
+    seen: dict[str, int] = {}
+    for item in terms or []:
+        if not isinstance(item, (tuple, list)) or len(item) < 2:
+            continue
+        source = str(item[0] or "").strip()[:120]
+        zh = str(item[1] or "").strip()[:120]
+        if not source or not zh:
+            continue
+        key = source.casefold()
+        if key in seen:
+            cleaned[seen[key]] = (source, zh)
+            continue
+        seen[key] = len(cleaned)
+        cleaned.append((source, zh))
+        if len(cleaned) >= 50:
+            break
+    return cleaned
+
+
+def _term_pattern(source: str) -> re.Pattern[str]:
+    escaped = re.escape(source)
+    if re.fullmatch(r"[A-Za-z][A-Za-z0-9'’.\- ]*", source):
+        return re.compile(rf"(?<![A-Za-z0-9]){escaped}(?![A-Za-z0-9])", re.IGNORECASE)
+    return re.compile(escaped)
+
+
+def _apply_translation_terms(
+    source_text: str,
+    zh_text: str,
+    terms: list[tuple[str, str]] | None,
+) -> str:
+    del source_text
+    result = zh_text
+    for source, zh in _normalize_translation_terms(terms):
+        result = _term_pattern(source).sub(zh, result)
+    return result

sentence_api/config.py

@@ -26,10 +26,12 @@ class Settings:
     moss_timeout_seconds: float
     moss_max_new_tokens: int
     moss_end_padding_ms: int
+    log_level: str
     ai_api_url: str
     ai_api_key: str
     ai_model: str
     ai_timeout_seconds: float
+    ai_translation_timeout_seconds: float
     pass_score: float
     volcano_oss_access_key: str
     volcano_oss_secret_key: str
@@ -65,10 +67,14 @@ class Settings:
             moss_timeout_seconds=float(os.getenv("MOSS_TIMEOUT_SECONDS", "1800")),
             moss_max_new_tokens=int(os.getenv("MOSS_MAX_NEW_TOKENS", "65536")),
             moss_end_padding_ms=int(os.getenv("MOSS_END_PADDING_MS", "300")),
+            log_level=os.getenv("LOG_LEVEL", "INFO").strip().upper() or "INFO",
             ai_api_url=os.getenv("AI_API_URL", "").strip().rstrip("/"),
             ai_api_key=os.getenv("AI_API_KEY", "").strip(),
             ai_model=os.getenv("AI_MODEL", "gpt-4o-mini"),
             ai_timeout_seconds=float(os.getenv("AI_TIMEOUT_SECONDS", "45")),
+            ai_translation_timeout_seconds=float(
+                os.getenv("AI_TRANSLATION_TIMEOUT_SECONDS", "120")
+            ),
             pass_score=float(os.getenv("ASSESSMENT_PASS_SCORE", "70")),
             volcano_oss_access_key=os.getenv("VOLCANO_OSS_ACCESS_KEY", "").strip(),
             volcano_oss_secret_key=os.getenv("VOLCANO_OSS_SECRET_KEY", "").strip(),

sentence_api/docker-compose.yml

@@ -3,6 +3,11 @@ services:
     build:
       context: ..
       dockerfile: sentence_api/Dockerfile
+      args:
+        PYTHON_IMAGE: ${PYTHON_IMAGE:-docker.m.daocloud.io/library/python:3.12-slim}
+        APT_MIRROR: ${APT_MIRROR:-mirrors.aliyun.com}
+        PIP_INDEX_URL: ${PIP_INDEX_URL:-https://pypi.tuna.tsinghua.edu.cn/simple}
+        PIP_TRUSTED_HOST: ${PIP_TRUSTED_HOST:-pypi.tuna.tsinghua.edu.cn}
     restart: unless-stopped
     network_mode: host
     env_file:

sentence_api/docker-daemon.json.example

@@ -0,0 +1,5 @@
+{
+  "registry-mirrors": [
+    "https://docker.m.daocloud.io"
+  ]
+}

sentence_api/main.py

@@ -24,11 +24,13 @@ from fastapi import (
     Request,
     UploadFile,
 )
+from pydantic import ValidationError
 from fastapi.responses import FileResponse, RedirectResponse, Response
 from fastapi.staticfiles import StaticFiles
 from starlette.concurrency import run_in_threadpool
 from starlette.requests import ClientDisconnect
 from .ai_tutor import AITutor, AITutorError
+from .translation_jobs import SubtitleTranslationService
 
 from .assessment import AssessmentService
 from .auth import AuthenticatedUser
@@ -39,11 +41,21 @@ from .models import (
     AuthRequest,
     AuthResponse,
     CourseSummary,
+    CourseTitleUpdate,
     SentenceBoundaryAdjust,
+    SentenceGapRepairRequest,
+    SentenceGapRepairResponse,
     SentenceBoundary,
     SentenceBoundaryDocument,
     SentenceExplanationRequest,
     SentenceExplanationResponse,
+    AdminSubtitleTranslationResponse,
+    AdminSentenceBatchItem,
+    AdminSentenceBatchUpdate,
+    LocalSubtitleTranslationRequest,
+    SubtitleRetranslateRequest,
+    SubtitleTranslationResponse,
+    PasswordResetRequest,
     SentenceTextUpdate,
     UserDubShareSummary,
     SsoExchangeRequest,
@@ -89,6 +101,10 @@ def create_app(
     oss_uploader: Optional[VolcanoOSSUploader] = None,
 ) -> FastAPI:
     service_settings = settings or Settings.from_env()
+    logging.basicConfig(
+        level=getattr(logging, service_settings.log_level, logging.INFO),
+        format="%(asctime)s %(levelname)s %(name)s %(message)s",
+    )
     service_settings.ensure_directories()
     video_repository = repository or VideoRepository(service_settings.database_path)
     legacy_store = store or BoundaryStore(service_settings.legacy_boundaries_file)
@@ -103,6 +119,7 @@ def create_app(
         api_key=service_settings.ai_api_key,
         model=service_settings.ai_model,
         timeout_seconds=service_settings.ai_timeout_seconds,
+        translation_timeout_seconds=service_settings.ai_translation_timeout_seconds,
     )
     oss = oss_uploader or VolcanoOSSUploader(service_settings)
     processor = VideoProcessor(service_settings, video_repository, moss, oss)
@@ -125,6 +142,10 @@ def create_app(
     application.state.oss_uploader = oss
     application.state.video_processor = processor
     application.state.assessment_service = assessment_service
+    application.state.subtitle_translations = SubtitleTranslationService(
+        video_repository,
+        lambda: application.state.ai_tutor,
+    )
 
     static_dir = Path(__file__).resolve().parent / "static"
     if static_dir.is_dir():
@@ -331,6 +352,59 @@ def create_app(
             raise HTTPException(status_code=503, detail=str(exc)) from exc
         return SentenceExplanationResponse(explanation=explanation)
 
+    def stored_subtitle_translations(
+        document: SentenceBoundaryDocument,
+    ) -> SubtitleTranslationResponse:
+        return application.state.subtitle_translations.snapshot(document)
+
+    @application.get(
+        "/api/v1/videos/{video_hash}/subtitle-translation",
+        response_model=SubtitleTranslationResponse,
+    )
+    def get_subtitle_translation(
+        video_hash: str = SHA256_PATH,
+        user: AuthenticatedUser = Depends(current_user),
+    ) -> SubtitleTranslationResponse:
+        del user
+        snapshot = application.state.subtitle_translations.current(video_hash)
+        if snapshot is None:
+            raise HTTPException(status_code=404, detail="Course is not available.")
+        return snapshot
+
+    @application.post(
+        "/api/v1/videos/{video_hash}/subtitle-translation",
+        response_model=SubtitleTranslationResponse,
+    )
+    async def translate_subtitles(
+        request: Request,
+        video_hash: str = SHA256_PATH,
+        user: AuthenticatedUser = Depends(current_user),
+    ) -> SubtitleTranslationResponse:
+        del user
+        raw = await request.body()
+        local_document = None
+        if raw.strip():
+            try:
+                payload = LocalSubtitleTranslationRequest.model_validate_json(raw)
+                local_document = SentenceBoundaryDocument(
+                    video_hash=video_hash,
+                    duration_ms=payload.duration_ms,
+                    algorithm_version=payload.algorithm_version,
+                    sentences=payload.sentences,
+                )
+            except ValidationError as exc:
+                raise HTTPException(status_code=422, detail=exc.errors()) from exc
+        try:
+            if local_document is not None:
+                return application.state.subtitle_translations.ensure_from_local(local_document)
+            return application.state.subtitle_translations.ensure(video_hash)
+        except KeyError:
+            raise HTTPException(status_code=404, detail="Course is not available.") from None
+        except ValueError as exc:
+            raise HTTPException(status_code=422, detail=str(exc)) from exc
+        except AITutorError as exc:
+            raise HTTPException(status_code=503, detail=str(exc)) from exc
+
     @application.post(
         "/api/v1/audio/sentence-boundaries",
         response_model=SentenceBoundaryDocument,
@@ -441,6 +515,19 @@ def create_app(
             raise HTTPException(status_code=404, detail="User was not found.")
         return Response(status_code=204)
 
+    @application.put(
+        "/api/v1/admin/users/password",
+        status_code=204,
+        dependencies=[Depends(require_admin)],
+    )
+    def reset_user_password(
+        payload: PasswordResetRequest,
+    ) -> Response:
+        updated = video_repository.set_user_password(payload.username, payload.password)
+        if not updated:
+            raise HTTPException(status_code=404, detail="User was not found.")
+        return Response(status_code=204)
+
     @application.get("/api/v1/videos", response_model=VideoListResponse)
     def list_videos() -> VideoListResponse:
         return VideoListResponse(
@@ -467,9 +554,13 @@ def create_app(
         media_path = service_settings.videos_dir / row["stored_filename"]
         if not media_path.is_file():
             raise HTTPException(status_code=404, detail="Stored video file is missing.")
+        media_type = row.get("content_type") or "video/mp4"
+        if not str(media_type).lower().startswith("video/"):
+            guessed = mimetypes.guess_type(str(row.get("stored_filename") or ""))[0]
+            media_type = guessed if guessed and guessed.lower().startswith("video/") else "video/mp4"
         return FileResponse(
             media_path,
-            media_type=row["content_type"],
+            media_type=media_type,
             headers={"Accept-Ranges": "bytes", "Cache-Control": "public, max-age=3600"},
         )
 
@@ -606,6 +697,127 @@ def create_app(
         background_tasks.add_task(_process_safely, processor, video_hash)
         return {"video_hash": video_hash.lower(), "status": "processing_queued"}
 
+    @application.post(
+        "/api/v1/admin/videos/{video_hash}/repair-gaps",
+        response_model=SentenceGapRepairResponse,
+        dependencies=[Depends(require_admin)],
+    )
+    async def repair_sentence_gaps(
+        payload: Optional[SentenceGapRepairRequest] = None,
+        video_hash: str = SHA256_PATH,
+    ) -> SentenceGapRepairResponse:
+        request = payload or SentenceGapRepairRequest()
+        try:
+            result = await run_in_threadpool(
+                processor.repair_gaps,
+                video_hash,
+                request.min_gap_ms,
+            )
+        except KeyError:
+            raise HTTPException(status_code=404, detail="Video was not found.") from None
+        except FileNotFoundError as exc:
+            raise HTTPException(status_code=404, detail=str(exc)) from exc
+        except ValueError as exc:
+            raise HTTPException(status_code=400, detail=str(exc)) from exc
+        except RuntimeError as exc:
+            raise HTTPException(status_code=503, detail=str(exc)) from exc
+        if result["inserted_count"]:
+            application.state.subtitle_translations.start_missing(video_hash)
+        return SentenceGapRepairResponse(video_hash=video_hash.lower(), **result)
+
+    @application.get(
+        "/api/v1/admin/videos/{video_hash}/subtitle-translation",
+        response_model=AdminSubtitleTranslationResponse,
+        dependencies=[Depends(require_admin)],
+    )
+    def get_admin_subtitle_translation(
+        video_hash: str = SHA256_PATH,
+    ) -> AdminSubtitleTranslationResponse:
+        snapshot = application.state.subtitle_translations.admin_snapshot(video_hash)
+        if snapshot is None:
+            raise HTTPException(status_code=404, detail="Course is not available.")
+        return snapshot
+
+    @application.post(
+        "/api/v1/admin/videos/{video_hash}/subtitle-translation",
+        response_model=AdminSubtitleTranslationResponse,
+        dependencies=[Depends(require_admin)],
+    )
+    def retranslate_subtitles(
+        payload: Optional[SubtitleRetranslateRequest] = None,
+        video_hash: str = SHA256_PATH,
+    ) -> AdminSubtitleTranslationResponse:
+        request = payload or SubtitleRetranslateRequest()
+        try:
+            return application.state.subtitle_translations.retranslate(
+                video_hash,
+                terms=[(item.source, item.zh) for item in request.terms],
+            )
+        except KeyError:
+            raise HTTPException(status_code=404, detail="Course is not available.") from None
+        except ValueError as exc:
+            raise HTTPException(status_code=422, detail=str(exc)) from exc
+        except AITutorError as exc:
+            raise HTTPException(status_code=503, detail=str(exc)) from exc
+
+    @application.post(
+        "/api/v1/admin/videos/{video_hash}/subtitle-translation/repair",
+        response_model=AdminSubtitleTranslationResponse,
+        dependencies=[Depends(require_admin)],
+    )
+    def repair_subtitle_translation(
+        video_hash: str = SHA256_PATH,
+    ) -> AdminSubtitleTranslationResponse:
+        try:
+            return application.state.subtitle_translations.repair_missing(video_hash)
+        except KeyError:
+            raise HTTPException(status_code=404, detail="Course is not available.") from None
+        except ValueError as exc:
+            raise HTTPException(status_code=422, detail=str(exc)) from exc
+        except AITutorError as exc:
+            raise HTTPException(status_code=503, detail=str(exc)) from exc
+
+    @application.put(
+        "/api/v1/admin/videos/{video_hash}/title",
+        response_model=VideoSummary,
+        dependencies=[Depends(require_admin)],
+    )
+    def update_course_title(
+        payload: CourseTitleUpdate,
+        video_hash: str = SHA256_PATH,
+    ) -> VideoSummary:
+        video = video_repository.update_title(video_hash, payload.title)
+        if video is None:
+            raise HTTPException(status_code=404, detail="Video was not found.")
+        return _video_summary(video, service_settings)
+
+    @application.put(
+        "/api/v1/admin/videos/{video_hash}/sentences/batch",
+        response_model=List[SentenceBoundary],
+        dependencies=[Depends(require_admin)],
+    )
+    def batch_admin_sentences(
+        payload: AdminSentenceBatchUpdate,
+        video_hash: str = SHA256_PATH,
+    ) -> List[SentenceBoundary]:
+        try:
+            result = _replace_admin_sentences(
+                repository=video_repository,
+                transcriber=moss,
+                settings=service_settings,
+                video_hash=video_hash,
+                payload=payload,
+            )
+        except KeyError:
+            raise HTTPException(status_code=404, detail="Course is not available.") from None
+        except ValueError as exc:
+            raise HTTPException(status_code=400, detail=str(exc)) from exc
+        if payload.process_timings:
+            application.state.subtitle_translations.refresh(video_hash)
+        else:
+            application.state.subtitle_translations.invalidate(video_hash)
+        return result
+
     @application.put(
         "/api/v1/admin/videos/{video_hash}/sentences/{sentence_index}",
         response_model=SentenceBoundary,
@@ -624,6 +836,7 @@ def create_app(
         )
         if sentence is None:
             raise HTTPException(status_code=404, detail="Sentence was not found.")
+        application.state.subtitle_translations.invalidate(video_hash)
         return sentence
 
     @application.put(
@@ -659,6 +872,7 @@ def create_app(
             raise HTTPException(status_code=400, detail=str(exc)) from exc
         if result is None:
             raise HTTPException(status_code=404, detail="Sentence was not found.")
+        application.state.subtitle_translations.refresh(video_hash)
         return result
 
 
@@ -680,6 +894,7 @@ def create_app(
             raise HTTPException(status_code=400, detail=str(exc)) from exc
         if sentence is None:
             raise HTTPException(status_code=404, detail="Sentence was not found.")
+        application.state.subtitle_translations.invalidate(video_hash)
         return sentence
 
     @application.put(
@@ -1016,6 +1231,161 @@ def _process_safely(processor: VideoProcessor, video_hash: str) -> None:
         logger.exception("Video processing failed for %s", video_hash)
 
 
+def _replace_admin_sentences(
+    *,
+    repository: VideoRepository,
+    transcriber: Transcriber,
+    settings: Settings,
+    video_hash: str,
+    payload: AdminSentenceBatchUpdate,
+) -> List[SentenceBoundary]:
+    """Apply queued admin edits once, instead of after every split/merge."""
+    video = repository.get_video(video_hash)
+    old_document = repository.get_document(video_hash)
+    if video is None or old_document is None:
+        raise KeyError(video_hash)
+
+    old_sentences = {
+        (sentence.start_ms, sentence.end_ms): sentence
+        for sentence in old_document.sentences
+    }
+    rows: List[SentenceBoundary] = []
+    index = 0
+    position = 0
+    while position < len(payload.sentences):
+        item = payload.sentences[position]
+        if item.source_start_ms is None or item.source_end_ms is None:
+            if "\n" in item.text or "\r" in item.text:
+                raise ValueError("Unsplit batch sentences cannot contain line breaks.")
+            rows.append(
+                SentenceBoundary(
+                    index=index,
+                    start_ms=item.start_ms,
+                    end_ms=item.end_ms,
+                    text=item.text.strip(),
+                    language=item.language,
+                    zh_text=item.zh_text,
+                    reference_speech_duration_ms=item.reference_speech_duration_ms,
+                )
+            )
+            index += 1
+            position += 1
+            continue
+
+        source_key = (item.source_start_ms, item.source_end_ms)
+        group_end = position + 1
+        while group_end < len(payload.sentences):
+            following = payload.sentences[group_end]
+            if following.source_start_ms != source_key[0] or following.source_end_ms != source_key[1]:
+                break
+            group_end += 1
+
+        group = payload.sentences[position:group_end]
+        source = old_sentences.get(source_key)
+        if source is None:
+            if payload.process_timings:
+                # A split may already have been persisted with provisional
+                # boundaries. Reconstruct its audio range from the saved rows so
+                # timing refinement can be retried after a dialog is reopened.
+                missing = [
+                    item
+                    for item in group
+                    if (item.start_ms, item.end_ms) not in old_sentences
+                ]
+                if missing:
+                    raise ValueError("The original sentence for a queued split no longer exists.")
+                source = SentenceBoundary(
+                    index=0,
+                    start_ms=source_key[0],
+                    end_ms=source_key[1],
+                    text="\n".join(item.text.strip() for item in group),
+                    language=next(
+                        (item.language for item in group if item.language is not None),
+                        None,
+                    ),
+                    reference_speech_duration_ms=sum(
+                        item.reference_speech_duration_ms for item in group
+                    ),
+                )
+            else:
+                source = None
+        if len(group) == 1:
+            if "\n" in group[0].text or "\r" in group[0].text:
+                raise ValueError("Unsplit batch sentences cannot contain line breaks.")
+            rows.append(
+                SentenceBoundary(
+                    index=index,
+                    start_ms=group[0].start_ms,
+                    end_ms=group[0].end_ms,
+                    text=group[0].text.strip(),
+                    language=group[0].language,
+                    zh_text=group[0].zh_text,
+                    reference_speech_duration_ms=group[0].reference_speech_duration_ms,
+                )
+            )
+        elif not payload.process_timings:
+            for item in group:
+                rows.append(
+                    SentenceBoundary(
+                        index=index,
+                        start_ms=item.start_ms,
+                        end_ms=item.end_ms,
+                        text=item.text.strip(),
+                        language=item.language,
+                        zh_text=item.zh_text,
+                        reference_speech_duration_ms=item.reference_speech_duration_ms,
+                    )
+                )
+                index += 1
+        else:
+            split_text = "\n".join(item.text.strip() for item in group if item.text.strip())
+            replacements = split_sentence_text(
+                transcriber=transcriber,
+                media_path=settings.videos_dir / video["stored_filename"],
+                sentence=source,
+                text=split_text,
+                language=next(
+                    (item.language for item in group if item.language is not None),
+                    source.language,
+                ),
+            )
+            if len(replacements) != len(group):
+                raise ValueError("Split result count does not match the queued sentences.")
+            for replacement, item in zip(replacements, group):
+                rows.append(
+                    SentenceBoundary(
+                        index=index,
+                        start_ms=replacement.start_ms,
+                        end_ms=replacement.end_ms,
+                        text=item.text.strip(),
+                        language=(
+                            item.language
+                            if item.language is not None
+                            else source.language
+                        ),
+                        zh_text=None,
+                        reference_speech_duration_ms=replacement.reference_speech_duration_ms,
+                    )
+                )
+                index += 1
+        index = len(rows)
+        position = group_end
+
+    duration_ms = video["duration_ms"]
+    if duration_ms is None:
+        raise ValueError("Course duration is not available.")
+    previous_end = 0
+    for row in rows:
+        if row.start_ms < previous_end:
+            raise ValueError("Batch sentences must be sorted and non-overlapping.")
+        if row.end_ms > duration_ms:
+            raise ValueError("Batch sentences cannot exceed the course duration.")
+        previous_end = row.end_ms
+
+    document = old_document.model_copy(update={"sentences": rows})
+    return repository.replace_sentences(document).sentences
+
+
 def _video_summary(row: Dict[str, Any], settings: Settings) -> VideoSummary:
     if row.get("remote_url"):
         stream_url = row["remote_url"]

sentence_api/models.py

@@ -12,6 +12,7 @@ class SentenceBoundary(BaseModel):
     end_ms: int = Field(gt=0)
     text: Optional[str] = None
     language: Optional[str] = None
+    zh_text: Optional[str] = None
     reference_speech_duration_ms: Optional[int] = Field(default=None, gt=0)
 
     @model_validator(mode="after")
@@ -51,7 +52,7 @@ class SentenceBoundaryDocument(BaseModel):
         return self
 
 
-VideoStatus = Literal["uploaded", "processing", "ready", "failed"]
+VideoStatus = Literal["uploaded", "processing", "ready", "failed", "local_subtitles"]
 
 
 class VideoSummary(BaseModel):
@@ -92,6 +93,57 @@ class SentenceTextUpdate(BaseModel):
     language: Optional[str] = Field(default=None, max_length=32)
 
 
+class AdminSentenceBatchItem(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    start_ms: int = Field(ge=0)
+    end_ms: int = Field(gt=0)
+    text: str = Field(min_length=1, max_length=4000)
+    language: Optional[str] = Field(default=None, max_length=32)
+    zh_text: Optional[str] = Field(default=None, max_length=4000)
+    reference_speech_duration_ms: int = Field(gt=0)
+    source_start_ms: Optional[int] = Field(default=None, ge=0)
+    source_end_ms: Optional[int] = Field(default=None, gt=0)
+
+    @model_validator(mode="after")
+    def validate_range(self):
+        if self.end_ms <= self.start_ms:
+            raise ValueError("end_ms must be greater than start_ms")
+        if (self.source_start_ms is None) != (self.source_end_ms is None):
+            raise ValueError("source_start_ms and source_end_ms must be set together")
+        elif self.source_start_ms is not None and self.source_end_ms <= self.source_start_ms:
+            raise ValueError("source end_ms must be greater than source start_ms")
+        return self
+
+
+class AdminSentenceBatchUpdate(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    sentences: List[AdminSentenceBatchItem] = Field(min_length=1, max_length=10000)
+    process_timings: bool = Field(default=True)
+
+
+class CourseTitleUpdate(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    title: str = Field(max_length=200)
+
+    @field_validator("title")
+    @classmethod
+    def validate_title(cls, value: str) -> str:
+        cleaned = value.strip()
+        if not cleaned:
+            raise ValueError("title cannot be empty")
+        return cleaned
+
+
+class PasswordResetRequest(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    username: str = Field(min_length=3, max_length=50)
+    password: str = Field(min_length=8, max_length=128)
+
+
 class SentenceExplanationRequest(BaseModel):
     model_config = ConfigDict(extra="forbid")
 
@@ -105,12 +157,83 @@ class SentenceExplanationResponse(BaseModel):
     explanation: str = Field(min_length=1, max_length=6000)
 
 
+class SubtitleTranslationItem(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    index: int = Field(ge=0)
+    zh_text: str = Field(min_length=1, max_length=4000)
+
+
+class SubtitleTranslationResponse(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    video_hash: str
+    translations: List[SubtitleTranslationItem]
+    complete: bool = True
+    status: str = "ready"
+    error: Optional[str] = None
+
+
+class LocalSubtitleTranslationRequest(BaseModel):
+    """Upload locally generated English subtitles so the server can translate them."""
+
+    model_config = ConfigDict(extra="forbid")
+
+    duration_ms: int = Field(gt=0)
+    algorithm_version: str = Field(default="local-pc", min_length=1, max_length=64)
+    sentences: List[SentenceBoundary] = Field(min_length=1)
+
+
+class TranslationTerm(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    source: str = Field(min_length=1, max_length=120)
+    zh: str = Field(min_length=1, max_length=120)
+
+    @field_validator("source", "zh")
+    @classmethod
+    def strip_term(cls, value: str) -> str:
+        cleaned = value.strip()
+        if not cleaned:
+            raise ValueError("translation term cannot be empty")
+        return cleaned
+
+
+class SubtitleRetranslateRequest(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    terms: List[TranslationTerm] = Field(default_factory=list, max_length=50)
+
+
+class AdminSubtitleTranslationResponse(SubtitleTranslationResponse):
+    terms: List[TranslationTerm] = Field(default_factory=list)
+    missing_count: int = Field(default=0, ge=0)
+    attempted_count: int = Field(default=0, ge=0)
+
+
 class SentenceBoundaryAdjust(BaseModel):
     model_config = ConfigDict(extra="forbid")
 
     delta_ms: int = Field(ge=-10_000, le=10_000)
 
 
+class SentenceGapRepairRequest(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    min_gap_ms: int = Field(default=8000, ge=2000, le=180_000)
+
+
+class SentenceGapRepairResponse(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    video_hash: str
+    min_gap_ms: int = Field(ge=1)
+    gap_count: int = Field(ge=0)
+    transcribed_gaps: int = Field(ge=0)
+    inserted_count: int = Field(ge=0)
+    sentence_count: int = Field(ge=0)
+
+
 class TextSubstitution(BaseModel):
     expected: str
     actual: str

sentence_api/processing.py

@@ -4,14 +4,16 @@ import subprocess
 import tempfile
 import uuid
 from pathlib import Path
-from typing import List, Optional
+from typing import Any, Dict, List, Optional, Sequence, Tuple
 
 from .audio_metrics import (
     AudioAnalysisError,
     analyze_samples,
     decode_audio_mono,
+    FRAME_SAMPLES,
     refine_sentence_end_ms,
     refine_sentence_start_ms,
+    speech_frame_mask,
 )
 from .config import Settings
 from .generate_boundaries import ALGORITHM_VERSION, make_entry, media_duration_ms
@@ -27,8 +29,72 @@ from .transcription import (
 
 
 logger = logging.getLogger(__name__)
-MOSS_ALGORITHM_VERSION = "moss-punctuation-v3"
+MOSS_ALGORITHM_VERSION = "moss-punctuation-v4"
 SILENCE_ALGORITHM_VERSION = "silence-rms-v1"
+GAP_REPAIR_MIN_MS = 8_000
+GAP_REPAIR_INSET_MS = 150
+GAP_REPAIR_MIN_SPEECH_MS = 400
+GAP_REPAIR_MIN_SENTENCE_MS = 100
+GAP_REPAIR_TARGET_CHUNK_MS = 15_000
+GAP_REPAIR_MAX_CHUNK_MS = 20_000
+GAP_REPAIR_MIN_CHUNK_MS = 5_000
+GAP_REPAIR_MIN_SPLIT_SILENCE_MS = 120
+GAP_REPAIR_FALLBACK_LANGUAGE = "en"
+_CUES_ID = b"\x1e\x53\xb6\x80"
+_CLUSTER_ID = b"\x1f\x43\xb6\x75"
+_HEADER_SCAN_SIZE = 64 * 1024
+_WEBM_SUFFIXES = {".webm", ".mkv"}
+
+
+def _webm_cues_at_front(filepath: Path) -> bool:
+    """Return True if WebM/MKV CUES appear before the first Cluster."""
+    try:
+        with open(filepath, "rb") as f:
+            header = f.read(_HEADER_SCAN_SIZE)
+    except OSError:
+        return True  # can't read → assume OK, don't touch
+    cues_pos = header.find(_CUES_ID)
+    cluster_pos = header.find(_CLUSTER_ID)
+    if cues_pos != -1 and (cluster_pos == -1 or cues_pos < cluster_pos):
+        return True
+    return False
+
+
+def _ensure_webm_cues_front(filepath: Path) -> None:
+    """Re-mux a WebM/MKV file (stream copy) if CUES are not at the front.
+
+    YouTube-dl / yt-dlp writes CUES at the end by default, which slows HTTP
+    Range seeking. This runs once during upload processing and takes seconds.
+    """
+    if filepath.suffix.lower() not in _WEBM_SUFFIXES:
+        return
+    if _webm_cues_at_front(filepath):
+        return
+    ffmpeg = shutil.which("ffmpeg")
+    if not ffmpeg:
+        logger.warning("ffmpeg not found; skipping WebM CUES optimization.")
+        return
+    tmp = filepath.with_suffix(".cues_tmp" + filepath.suffix)
+    try:
+        result = subprocess.run(
+            [ffmpeg, "-y", "-loglevel", "error",
+             "-i", str(filepath), "-c", "copy",
+             "-cues_to_front", "1", str(tmp)],
+            capture_output=True, timeout=300,
+        )
+        if result.returncode == 0 and tmp.is_file() and tmp.stat().st_size > 0:
+            tmp.replace(filepath)
+            logger.info("WebM CUES moved to front: %s", filepath.name)
+        else:
+            logger.warning(
+                "WebM CUES re-mux failed for %s: rc=%d stderr=%s",
+                filepath.name, result.returncode, result.stderr.decode()[-200:],
+            )
+    except (subprocess.TimeoutExpired, OSError) as exc:
+        logger.warning("WebM CUES re-mux error for %s: %s", filepath.name, exc)
+    finally:
+        tmp.unlink(missing_ok=True)
+
 
 
 class VideoProcessor:
@@ -55,6 +121,7 @@ class VideoProcessor:
         self.repository.mark_processing(video_hash)
         work_path: Optional[Path] = None
         try:
+            _ensure_webm_cues_front(media_path)
             if self.transcriber.available:
                 work_path = self.settings.work_dir / f"{video_hash}-{uuid.uuid4().hex}.wav"
                 extract_audio(media_path, work_path)
@@ -69,7 +136,11 @@ class VideoProcessor:
                 )
                 if not document.sentences:
                     raise RuntimeError("MOSS returned no timestamped speech segments.")
-                self._finish_processing(document, transcript.text)
+                self._finish_processing(
+                    document,
+                    transcript.text,
+                    moss_segments=_serialize_moss_segments(transcript),
+                )
             else:
                 entry, _ = make_entry(media_path, video_hash=video_hash)
                 document = SentenceBoundaryDocument(
@@ -90,13 +161,18 @@ class VideoProcessor:
         self,
         document: SentenceBoundaryDocument,
         transcription: Optional[str],
+        moss_segments: Optional[List[Dict[str, Any]]] = None,
     ) -> None:
         if self.oss_uploader is None or not self.oss_uploader.enabled:
-            self.repository.save_processing_result(document, transcription, ready=True)
+            self.repository.save_processing_result(
+                document, transcription, ready=True, moss_segments=moss_segments
+            )
             return
 
         # Keep the course invisible while the mandatory OSS upload is running.
-        self.repository.save_processing_result(document, transcription, ready=False)
+        self.repository.save_processing_result(
+            document, transcription, ready=False, moss_segments=moss_segments
+        )
         video = self.repository.get_video(document.video_hash)
         if video is None:
             raise ValueError(f"Unknown video: {document.video_hash}")
@@ -108,6 +184,338 @@ class VideoProcessor:
         )
         self.repository.mark_oss_uploaded(document.video_hash, remote_url)
 
+    def repair_gaps(
+        self,
+        video_hash: str,
+        min_gap_ms: int = GAP_REPAIR_MIN_MS,
+    ) -> Dict[str, int]:
+        """Re-transcribe long unsentenced gaps without replacing existing cues."""
+        if min_gap_ms < 1:
+            raise ValueError("min_gap_ms must be positive")
+        if not self.transcriber.available:
+            raise RuntimeError("MOSS transcription is not configured on this server.")
+        video = self.repository.get_video(video_hash)
+        if video is None:
+            raise KeyError(video_hash)
+        document = self.repository.get_document(video_hash)
+        if document is None or not document.sentences:
+            raise ValueError("Sentence boundaries are not available for this video.")
+        media_path = self.settings.videos_dir / video["stored_filename"]
+        if not media_path.is_file():
+            raise FileNotFoundError(f"Stored video is missing: {media_path.name}")
+
+        gaps = find_sentence_gaps(document, min_gap_ms)
+        inserted: List[SentenceBoundary] = []
+        transcribed_gaps = 0
+        logger.info(
+            "Gap repair started: video=%s min_gap_ms=%d language=%s gaps=%d "
+            "sentences=%d duration_ms=%d transcriber=%s",
+            video_hash,
+            min_gap_ms,
+            video.get("language") or "auto",
+            len(gaps),
+            len(document.sentences),
+            document.duration_ms,
+            type(self.transcriber).__name__,
+        )
+        if gaps:
+            language = video.get("language")
+            with tempfile.TemporaryDirectory(prefix="gap-repair-") as temp_dir:
+                for gap_index, (raw_start, raw_end) in enumerate(gaps):
+                    logger.info(
+                        "Gap %d/%d: raw=%d-%dms duration=%dms",
+                        gap_index + 1,
+                        len(gaps),
+                        raw_start,
+                        raw_end,
+                        raw_end - raw_start,
+                    )
+                    clipped = _inset_gap(raw_start, raw_end, document.duration_ms)
+                    if clipped is None:
+                        logger.warning(
+                            "Gap %d/%d skipped after inset: raw=%d-%dms",
+                            gap_index + 1,
+                            len(gaps),
+                            raw_start,
+                            raw_end,
+                        )
+                        continue
+                    start_ms, end_ms = clipped
+                    chunk_path = Path(temp_dir) / f"gap-{gap_index}-full.wav"
+                    logger.info(
+                        "Gap %d/%d extracted range=%d-%dms duration=%dms file=%s",
+                        gap_index + 1,
+                        len(gaps),
+                        start_ms,
+                        end_ms,
+                        end_ms - start_ms,
+                        chunk_path,
+                    )
+                    try:
+                        extract_audio_segment(
+                            media_path,
+                            chunk_path,
+                            start_ms=start_ms,
+                            duration_ms=end_ms - start_ms,
+                        )
+                        samples, sample_rate = decode_audio_mono(chunk_path)
+                        metrics = analyze_samples(samples, sample_rate)
+                    except (AudioAnalysisError, RuntimeError) as exc:
+                        logger.warning(
+                            "Skipping sentence gap %s-%sms: %s",
+                            start_ms,
+                            end_ms,
+                            exc,
+                            exc_info=True,
+                        )
+                        continue
+                    logger.info(
+                        "Gap %d/%d audio metrics: duration_ms=%d speech_duration_ms=%d "
+                        "internal_silence_ms=%d pause_ratio=%.3f threshold_ms=%d",
+                        gap_index + 1,
+                        len(gaps),
+                        metrics.recording_duration_ms,
+                        metrics.speech_duration_ms,
+                        metrics.internal_silence_ms,
+                        metrics.internal_pause_ratio,
+                        GAP_REPAIR_MIN_SPEECH_MS,
+                    )
+                    if metrics.speech_duration_ms < GAP_REPAIR_MIN_SPEECH_MS:
+                        logger.warning(
+                            "Gap %d/%d skipped as silence: speech_duration_ms=%d < %d",
+                            gap_index + 1,
+                            len(gaps),
+                            metrics.speech_duration_ms,
+                            GAP_REPAIR_MIN_SPEECH_MS,
+                        )
+                        continue
+                    relative_chunks = _plan_gap_chunks(samples, sample_rate, end_ms - start_ms)
+                    logger.info(
+                        "Gap %d/%d split into %d transcription chunks: %s",
+                        gap_index + 1,
+                        len(gaps),
+                        len(relative_chunks),
+                        relative_chunks,
+                    )
+                    gap_inserted_count = 0
+                    for chunk_index, (chunk_offset_start, chunk_offset_end) in enumerate(
+                        relative_chunks
+                    ):
+                        chunk_start = start_ms + chunk_offset_start
+                        chunk_end = start_ms + chunk_offset_end
+                        chunk_path = (
+                            Path(temp_dir) / f"gap-{gap_index}-chunk-{chunk_index}.wav"
+                        )
+                        try:
+                            extract_audio_segment(
+                                media_path,
+                                chunk_path,
+                                start_ms=chunk_start,
+                                duration_ms=chunk_end - chunk_start,
+                            )
+                            chunk_samples, chunk_sample_rate = decode_audio_mono(chunk_path)
+                            chunk_metrics = analyze_samples(
+                                chunk_samples, chunk_sample_rate
+                            )
+                        except (AudioAnalysisError, RuntimeError) as exc:
+                            logger.warning(
+                                "Could not prepare gap chunk %d/%d %s-%sms: %s",
+                                chunk_index + 1,
+                                len(relative_chunks),
+                                chunk_start,
+                                chunk_end,
+                                exc,
+                                exc_info=True,
+                            )
+                            continue
+                        if chunk_metrics.speech_duration_ms < GAP_REPAIR_MIN_SPEECH_MS:
+                            logger.info(
+                                "Skipping quiet gap chunk %d/%d: range=%d-%dms "
+                                "speech_duration_ms=%d",
+                                chunk_index + 1,
+                                len(relative_chunks),
+                                chunk_start,
+                                chunk_end,
+                                chunk_metrics.speech_duration_ms,
+                            )
+                            continue
+
+                        languages_to_try = [language]
+                        if not language or language.strip().lower() == "auto":
+                            fallback = GAP_REPAIR_FALLBACK_LANGUAGE
+                            if fallback not in languages_to_try:
+                                languages_to_try.append(fallback)
+
+                        transcript = None
+                        for attempt_index, attempt_language in enumerate(languages_to_try):
+                            try:
+                                logger.info(
+                                    "Transcribing gap chunk %d/%d: gap=%d/%d "
+                                    "range=%d-%dms language=%s",
+                                    chunk_index + 1,
+                                    len(relative_chunks),
+                                    gap_index + 1,
+                                    len(gaps),
+                                    chunk_start,
+                                    chunk_end,
+                                    attempt_language or "auto",
+                                )
+                                transcript = self.transcriber.transcribe(
+                                    chunk_path, attempt_language
+                                )
+                            except Exception as exc:
+                                logger.warning(
+                                    "Gap chunk transcription failed for %s %s-%sms: %s",
+                                    video_hash,
+                                    chunk_start,
+                                    chunk_end,
+                                    exc,
+                                    exc_info=True,
+                                )
+                                continue
+                            logger.info(
+                                "Gap chunk transcript returned: gap=%d/%d chunk=%d/%d "
+                                "segments=%d text=%r",
+                                gap_index + 1,
+                                len(gaps),
+                                chunk_index + 1,
+                                len(relative_chunks),
+                                len(transcript.segments),
+                                transcript.text,
+                            )
+                            if transcript.segments:
+                                break
+                            if attempt_index + 1 < len(languages_to_try):
+                                logger.warning(
+                                    "Gap chunk returned no segments; retrying with %s: "
+                                    "range=%d-%dms",
+                                    languages_to_try[attempt_index + 1],
+                                    chunk_start,
+                                    chunk_end,
+                                )
+
+                        if transcript is None or not transcript.segments:
+                            logger.warning(
+                                "Gap chunk produced no transcript segments: gap=%d/%d "
+                                "chunk=%d/%d range=%d-%dms attempts=%s",
+                                gap_index + 1,
+                                len(gaps),
+                                chunk_index + 1,
+                                len(relative_chunks),
+                                chunk_start,
+                                chunk_end,
+                                [item or "auto" for item in languages_to_try],
+                            )
+                            continue
+
+                        gap_document = document_from_transcript(
+                            video_hash=video_hash,
+                            duration_ms=chunk_end - chunk_start,
+                            transcript=transcript,
+                            language=language,
+                            audio_path=chunk_path,
+                            end_padding_ms=self.settings.moss_end_padding_ms,
+                        )
+                        shifted = _shift_gap_sentences(
+                            gap_document.sentences,
+                            offset_ms=chunk_start,
+                            limit_end_ms=chunk_end,
+                        )
+                        if not shifted:
+                            logger.warning(
+                                "Gap chunk produced no insertable sentences: gap=%d/%d "
+                                "chunk=%d/%d transcript_segments=%d "
+                                "boundary_sentences=%d range=%d-%dms",
+                                gap_index + 1,
+                                len(gaps),
+                                chunk_index + 1,
+                                len(relative_chunks),
+                                len(transcript.segments),
+                                len(gap_document.sentences),
+                                chunk_start,
+                                chunk_end,
+                            )
+                            for boundary in gap_document.sentences:
+                                logger.warning(
+                                    "Candidate boundary before shift: start=%d end=%d text=%r",
+                                    boundary.start_ms,
+                                    boundary.end_ms,
+                                    boundary.text,
+                                )
+                            continue
+                        inserted.extend(shifted)
+                        gap_inserted_count += len(shifted)
+                        logger.info(
+                            "Gap chunk produced %d insertable sentences: gap=%d/%d "
+                            "chunk=%d/%d range=%d-%dms",
+                            len(shifted),
+                            gap_index + 1,
+                            len(gaps),
+                            chunk_index + 1,
+                            len(relative_chunks),
+                            chunk_start,
+                            chunk_end,
+                        )
+
+                    if gap_inserted_count:
+                        transcribed_gaps += 1
+                        logger.info(
+                            "Gap %d/%d produced %d insertable sentences total: range=%d-%dms",
+                            gap_index + 1,
+                            len(gaps),
+                            gap_inserted_count,
+                            start_ms,
+                            end_ms,
+                        )
+                    else:
+                        logger.warning(
+                            "Gap %d/%d produced no insertable sentences after chunking: "
+                            "range=%d-%dms chunks=%d",
+                            gap_index + 1,
+                            len(gaps),
+                            start_ms,
+                            end_ms,
+                            len(relative_chunks),
+                        )
+
+
+        if not inserted:
+            logger.warning(
+                "Gap repair finished without inserts: video=%s gaps=%d "
+                "transcribed_gaps=%d inserted=0",
+                video_hash,
+                len(gaps),
+                transcribed_gaps,
+            )
+            return {
+                "min_gap_ms": min_gap_ms,
+                "gap_count": len(gaps),
+                "transcribed_gaps": transcribed_gaps,
+                "inserted_count": 0,
+                "sentence_count": len(document.sentences),
+            }
+
+        merged = _merge_sentence_lists(document.sentences, inserted)
+        updated = document.model_copy(update={"sentences": merged})
+        saved = self.repository.replace_sentences(updated)
+        logger.info(
+            "Gap repair finished: video=%s gaps=%d transcribed_gaps=%d inserted=%d "
+            "sentences=%d->%d",
+            video_hash,
+            len(gaps),
+            transcribed_gaps,
+            len(saved.sentences) - len(document.sentences),
+            len(document.sentences),
+            len(saved.sentences),
+        )
+        return {
+            "min_gap_ms": min_gap_ms,
+            "gap_count": len(gaps),
+            "transcribed_gaps": transcribed_gaps,
+            "inserted_count": len(saved.sentences) - len(document.sentences),
+            "sentence_count": len(saved.sentences),
+        }
+
 
 def extract_audio(media_path: Path, output_path: Path) -> None:
     ffmpeg = shutil.which("ffmpeg")
@@ -176,6 +584,169 @@ def extract_audio_segment(
         raise RuntimeError(f"Could not extract sentence audio: {message[-2000:]}")
 
 
+def find_sentence_gaps(
+    document: SentenceBoundaryDocument,
+    min_gap_ms: int = GAP_REPAIR_MIN_MS,
+) -> List[Tuple[int, int]]:
+    """Return [start_ms, end_ms) ranges that are missing sentence coverage."""
+    if min_gap_ms < 1:
+        raise ValueError("min_gap_ms must be positive")
+    gaps: List[Tuple[int, int]] = []
+    cursor = 0
+    for sentence in document.sentences:
+        if sentence.start_ms - cursor >= min_gap_ms:
+            gaps.append((cursor, sentence.start_ms))
+        cursor = max(cursor, sentence.end_ms)
+    if document.duration_ms - cursor >= min_gap_ms:
+        gaps.append((cursor, document.duration_ms))
+    return gaps
+
+
+def _inset_gap(
+    start_ms: int,
+    end_ms: int,
+    duration_ms: int,
+    inset_ms: int = GAP_REPAIR_INSET_MS,
+) -> Optional[Tuple[int, int]]:
+    start = start_ms + (inset_ms if start_ms > 0 else 0)
+    end = end_ms - (inset_ms if end_ms < duration_ms else 0)
+    if end - start < GAP_REPAIR_MIN_SPEECH_MS:
+        return None
+    return start, end
+
+
+def _plan_gap_chunks(
+    samples,
+    sample_rate: int,
+    duration_ms: int,
+) -> List[Tuple[int, int]]:
+    """Split a long gap near silence so no ASR request exceeds Whisper's window."""
+    if duration_ms <= GAP_REPAIR_MAX_CHUNK_MS:
+        return [(0, duration_ms)]
+
+    chunks: List[Tuple[int, int]] = []
+    cursor = 0
+    while cursor < duration_ms:
+        remaining = duration_ms - cursor
+        if remaining <= GAP_REPAIR_MAX_CHUNK_MS:
+            chunks.append((cursor, duration_ms))
+            break
+
+        preferred = min(cursor + GAP_REPAIR_TARGET_CHUNK_MS, duration_ms)
+        lower = min(cursor + GAP_REPAIR_MIN_CHUNK_MS, preferred)
+        upper = min(cursor + GAP_REPAIR_MAX_CHUNK_MS, duration_ms)
+        cut = _find_silence_cut(
+            samples,
+            sample_rate,
+            lower,
+            upper,
+            preferred,
+            min_silence_ms=GAP_REPAIR_MIN_SPLIT_SILENCE_MS,
+        )
+        cut = max(cursor + 1, min(duration_ms, cut))
+        chunks.append((cursor, cut))
+        cursor = cut
+    return chunks
+
+
+def _find_silence_cut(
+    samples,
+    sample_rate: int,
+    lower_ms: int,
+    upper_ms: int,
+    preferred_ms: int,
+    *,
+    min_silence_ms: int,
+) -> int:
+    """Choose the silence midpoint nearest ``preferred_ms``."""
+    frame_ms = 1000.0 * FRAME_SAMPLES / sample_rate
+    min_silence_frames = max(1, int(round(min_silence_ms / frame_ms)))
+    start_sample = max(0, int(lower_ms / 1000 * sample_rate))
+    end_sample = min(samples.size, int(upper_ms / 1000 * sample_rate) + 1)
+    if end_sample <= start_sample:
+        return preferred_ms
+
+    try:
+        speech = speech_frame_mask(samples[start_sample:end_sample], sample_rate)
+    except AudioAnalysisError:
+        return preferred_ms
+
+    run_start: Optional[int] = None
+    candidates: List[Tuple[int, int]] = []
+    for index, is_speech in enumerate(speech):
+        if not is_speech:
+            if run_start is None:
+                run_start = index
+            continue
+        if run_start is not None:
+            candidates.append((run_start, index))
+            run_start = None
+    if run_start is not None:
+        candidates.append((run_start, len(speech)))
+
+    valid = [
+        (run_start, run_end)
+        for run_start, run_end in candidates
+        if run_end - run_start >= min_silence_frames
+    ]
+    if not valid:
+        return preferred_ms
+
+    def distance(run: Tuple[int, int]) -> float:
+        midpoint_frame = (run[0] + run[1]) / 2.0
+        midpoint_ms = lower_ms + midpoint_frame * frame_ms
+        return abs(midpoint_ms - preferred_ms)
+
+    run_start, run_end = min(valid, key=distance)
+    midpoint_ms = lower_ms + ((run_start + run_end) / 2.0) * frame_ms
+    return int(round(midpoint_ms))
+
+
+def _shift_gap_sentences(
+    sentences: Sequence[SentenceBoundary],
+    *,
+    offset_ms: int,
+    limit_end_ms: int,
+) -> List[SentenceBoundary]:
+    shifted: List[SentenceBoundary] = []
+    previous_end = offset_ms
+    for sentence in sentences:
+        start_ms = max(previous_end, sentence.start_ms + offset_ms)
+        end_ms = min(limit_end_ms, sentence.end_ms + offset_ms)
+        text = (sentence.text or "").strip()
+        if not text or end_ms - start_ms < GAP_REPAIR_MIN_SENTENCE_MS:
+            continue
+        shifted.append(
+            sentence.model_copy(
+                update={
+                    "index": len(shifted),
+                    "start_ms": start_ms,
+                    "end_ms": end_ms,
+                    "text": text,
+                    "zh_text": None,
+                }
+            )
+        )
+        previous_end = end_ms
+    return shifted
+
+
+def _merge_sentence_lists(
+    existing: Sequence[SentenceBoundary],
+    inserted: Sequence[SentenceBoundary],
+) -> List[SentenceBoundary]:
+    combined = [*existing, *inserted]
+    combined.sort(key=lambda sentence: (sentence.start_ms, sentence.end_ms))
+    merged: List[SentenceBoundary] = []
+    previous_end = 0
+    for sentence in combined:
+        if sentence.start_ms < previous_end:
+            continue
+        merged.append(sentence.model_copy(update={"index": len(merged)}))
+        previous_end = sentence.end_ms
+    return merged
+
+
 def split_sentence_text(
     *,
     transcriber: Transcriber,
@@ -484,3 +1055,32 @@ def _media_duration_ms(path: Path) -> int:
     from .generate_boundaries import media_duration_ms
 
     return media_duration_ms(path)
+
+
+def _serialize_moss_segments(transcript: Transcript) -> List[Dict[str, Any]]:
+    rows: List[Dict[str, Any]] = []
+    for segment in transcript.segments:
+        text = segment.text.strip()
+        if not text:
+            continue
+        words = None
+        if segment.words:
+            words = [
+                {
+                    "start_seconds": word.start_seconds,
+                    "end_seconds": word.end_seconds,
+                    "text": word.text,
+                }
+                for word in segment.words
+            ]
+        rows.append(
+            {
+                "start_seconds": segment.start_seconds,
+                "end_seconds": segment.end_seconds,
+                "text": text,
+                "speaker": segment.speaker,
+                "words": words,
+            }
+        )
+    return rows
+

sentence_api/repository.py

@@ -2,7 +2,7 @@ import json
 import sqlite3
 from datetime import datetime, timedelta, timezone
 from pathlib import Path
-from typing import Any, Dict, List, Optional
+from typing import Any, Dict, List, Optional, Sequence, Tuple
 import uuid
 
 from .auth import hash_password, hash_token, new_token, verify_password
@@ -56,6 +56,7 @@ class VideoRepository:
                     end_ms INTEGER NOT NULL,
                     text TEXT,
                     language TEXT,
+                    zh_text TEXT,
                     reference_speech_duration_ms INTEGER,
                     PRIMARY KEY (video_hash, sentence_index),
                     FOREIGN KEY (video_hash) REFERENCES videos(video_hash) ON DELETE CASCADE
@@ -120,6 +121,20 @@ class VideoRepository:
                     FOREIGN KEY (share_id) REFERENCES dub_shares(share_id) ON DELETE CASCADE,
                     UNIQUE (share_id, sentence_index)
                 );
+
+                CREATE TABLE IF NOT EXISTS moss_segments (
+                    video_hash TEXT NOT NULL,
+                    segment_index INTEGER NOT NULL,
+                    start_seconds REAL NOT NULL,
+                    end_seconds REAL NOT NULL,
+                    start_ms INTEGER NOT NULL,
+                    end_ms INTEGER NOT NULL,
+                    text TEXT NOT NULL,
+                    speaker TEXT,
+                    words_json TEXT,
+                    PRIMARY KEY (video_hash, segment_index),
+                    FOREIGN KEY (video_hash) REFERENCES videos(video_hash) ON DELETE CASCADE
+                );
                 """
             )
 
@@ -179,6 +194,22 @@ class VideoRepository:
             if "remote_url" not in video_columns:
                 connection.execute("ALTER TABLE videos ADD COLUMN remote_url TEXT")
 
+        with self._connect() as connection:
+            sentence_columns = {
+                row["name"]
+                for row in connection.execute("PRAGMA table_info(sentences)")
+            }
+            if "zh_text" not in sentence_columns:
+                connection.execute("ALTER TABLE sentences ADD COLUMN zh_text TEXT")
+
+        with self._connect() as connection:
+            video_columns = {
+                row["name"]
+                for row in connection.execute("PRAGMA table_info(videos)")
+            }
+            if "translation_glossary" not in video_columns:
+                connection.execute("ALTER TABLE videos ADD COLUMN translation_glossary TEXT")
+
     def upsert_upload(
         self,
         *,
@@ -230,7 +261,24 @@ class VideoRepository:
             )
 
     def mark_processing(self, video_hash: str) -> None:
-        self._update_status(video_hash, "processing", None)
+        normalized_hash = normalize_video_hash(video_hash)
+        with self._connect() as connection:
+            connection.execute(
+                "DELETE FROM sentences WHERE video_hash = ?",
+                (normalized_hash,),
+            )
+            connection.execute(
+                "DELETE FROM moss_segments WHERE video_hash = ?",
+                (normalized_hash,),
+            )
+            connection.execute(
+                """
+                UPDATE videos
+                SET status = ?, error_message = NULL, transcription = NULL, updated_at = ?
+                WHERE video_hash = ?
+                """,
+                ("processing", utc_now(), normalized_hash),
+            )
 
     def mark_failed(self, video_hash: str, message: str) -> None:
         self._update_status(video_hash, "failed", message[:4000])
@@ -247,12 +295,17 @@ class VideoRepository:
         document: SentenceBoundaryDocument,
         transcription: Optional[str],
         ready: bool = True,
+        moss_segments: Optional[List[Dict[str, Any]]] = None,
     ) -> None:
         with self._connect() as connection:
             connection.execute(
                 "DELETE FROM sentences WHERE video_hash = ?",
                 (document.video_hash,),
             )
+            connection.execute(
+                "DELETE FROM moss_segments WHERE video_hash = ?",
+                (document.video_hash,),
+            )
             connection.executemany(
                 """
                 INSERT INTO sentences (
@@ -288,6 +341,111 @@ class VideoRepository:
                     document.video_hash,
                 ),
             )
+            if moss_segments:
+                connection.executemany(
+                    """
+                    INSERT INTO moss_segments (
+                        video_hash, segment_index, start_seconds, end_seconds,
+                        start_ms, end_ms, text, speaker, words_json
+                    ) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)
+                    """,
+                    [
+                        _moss_segment_row(document.video_hash, index, segment)
+                        for index, segment in enumerate(moss_segments)
+                    ],
+                )
+
+    def import_local_subtitle_document(
+        self,
+        document: SentenceBoundaryDocument,
+    ) -> SentenceBoundaryDocument:
+        """Save locally generated English subtitles for translation.
+
+        Official course videos are left unchanged. A previous local-only
+        import is replaced when the English cues changed.
+        """
+        normalized_hash = normalize_video_hash(document.video_hash)
+        document = document.model_copy(update={"video_hash": normalized_hash})
+        existing_video = self.get_video(normalized_hash)
+        existing_document = self.get_document(normalized_hash)
+        if existing_video is not None and existing_video.get("status") != "local_subtitles":
+            if existing_document is None:
+                raise KeyError(normalized_hash)
+            return existing_document
+        if existing_document is not None and _english_fingerprint(existing_document) == _english_fingerprint(document):
+            return existing_document
+
+        now = utc_now()
+        language = next((sentence.language for sentence in document.sentences if sentence.language), None)
+        with self._connect() as connection:
+            if existing_video is None:
+                connection.execute(
+                    """
+                    INSERT INTO videos (
+                        video_hash, title, filename, stored_filename, content_type,
+                        size_bytes, duration_ms, language, status, error_message,
+                        algorithm_version, transcription, remote_url, created_at, updated_at
+                    ) VALUES (?, ?, ?, ?, ?, ?, ?, ?, 'local_subtitles', NULL, ?, NULL, NULL, ?, ?)
+                    """,
+                    (
+                        normalized_hash,
+                        "Local subtitles",
+                        "local-subtitles.json",
+                        "",
+                        "application/json",
+                        0,
+                        document.duration_ms,
+                        language,
+                        document.algorithm_version,
+                        now,
+                        now,
+                    ),
+                )
+            else:
+                connection.execute(
+                    """
+                    UPDATE videos
+                    SET duration_ms = ?, language = COALESCE(?, language),
+                        algorithm_version = ?, status = 'local_subtitles',
+                        error_message = NULL, updated_at = ?
+                    WHERE video_hash = ?
+                    """,
+                    (
+                        document.duration_ms,
+                        language,
+                        document.algorithm_version,
+                        now,
+                        normalized_hash,
+                    ),
+                )
+            connection.execute(
+                "DELETE FROM sentences WHERE video_hash = ?",
+                (normalized_hash,),
+            )
+            connection.executemany(
+                """
+                INSERT INTO sentences (
+                    video_hash, sentence_index, start_ms, end_ms, text,
+                    language, zh_text, reference_speech_duration_ms
+                ) VALUES (?, ?, ?, ?, ?, ?, NULL, ?)
+                """,
+                [
+                    (
+                        normalized_hash,
+                        sentence.index,
+                        sentence.start_ms,
+                        sentence.end_ms,
+                        sentence.text,
+                        sentence.language,
+                        sentence.reference_speech_duration_ms,
+                    )
+                    for sentence in document.sentences
+                ],
+            )
+        imported = self.get_document(normalized_hash)
+        if imported is None:
+            raise KeyError(normalized_hash)
+        return imported
 
     def mark_oss_uploaded(self, video_hash: str, remote_url: str) -> None:
         with self._connect() as connection:
@@ -300,6 +458,22 @@ class VideoRepository:
                 (remote_url, utc_now(), normalize_video_hash(video_hash)),
             )
 
+    def update_title(self, video_hash: str, title: str) -> Optional[Dict[str, Any]]:
+        normalized_hash = normalize_video_hash(video_hash)
+        cleaned_title = title.strip()[:200]
+        if not cleaned_title:
+            raise ValueError("Title cannot be empty.")
+
+        with self._connect() as connection:
+            cursor = connection.execute(
+                "UPDATE videos SET title = ?, updated_at = ? WHERE video_hash = ?",
+                (cleaned_title, utc_now(), normalized_hash),
+            )
+            if cursor.rowcount == 0:
+                return None
+
+        return self.get_video(normalized_hash)
+
     def list_videos(self) -> List[Dict[str, Any]]:
         with self._connect() as connection:
             rows = connection.execute(
@@ -307,6 +481,7 @@ class VideoRepository:
                 SELECT v.*, COUNT(s.sentence_index) AS sentence_count
                 FROM videos v
                 LEFT JOIN sentences s ON s.video_hash = v.video_hash
+                WHERE v.status != 'local_subtitles'
                 GROUP BY v.video_hash
                 ORDER BY v.created_at DESC
                 """
@@ -327,6 +502,32 @@ class VideoRepository:
             ).fetchone()
         return dict(row) if row else None
 
+    def get_moss_segments(self, video_hash: str) -> List[Dict[str, Any]]:
+        with self._connect() as connection:
+            rows = connection.execute(
+                """
+                SELECT * FROM moss_segments
+                WHERE video_hash = ?
+                ORDER BY segment_index
+                """,
+                (normalize_video_hash(video_hash),),
+            ).fetchall()
+        segments = []
+        for row in rows:
+            item = dict(row)
+            raw_words = item.pop("words_json", None)
+            words = None
+            if raw_words:
+                try:
+                    parsed = json.loads(raw_words)
+                except json.JSONDecodeError:
+                    parsed = None
+                if isinstance(parsed, list):
+                    words = parsed
+            item["words"] = words
+            segments.append(item)
+        return segments
+
     def get_document(self, video_hash: str) -> Optional[SentenceBoundaryDocument]:
         video = self.get_video(video_hash)
         if video is None or video["duration_ms"] is None:
@@ -340,18 +541,55 @@ class VideoRepository:
             video_hash=video["video_hash"],
             duration_ms=video["duration_ms"],
             algorithm_version=video["algorithm_version"] or "unknown",
-            sentences=[
-                SentenceBoundary(
-                    index=row["sentence_index"],
-                    start_ms=row["start_ms"],
-                    end_ms=row["end_ms"],
-                    text=row["text"],
-                    language=row["language"],
-                    reference_speech_duration_ms=row["reference_speech_duration_ms"],
+            sentences=[_sentence_from_row(row) for row in rows],
         )
-                for row in rows
+
+    def replace_sentences(
+        self,
+        document: SentenceBoundaryDocument,
+    ) -> SentenceBoundaryDocument:
+        """Replace stored sentences while preserving video status and translations."""
+        normalized_hash = normalize_video_hash(document.video_hash)
+        with self._connect() as connection:
+            existing = connection.execute(
+                "SELECT video_hash FROM videos WHERE video_hash = ?",
+                (normalized_hash,),
+            ).fetchone()
+            if existing is None:
+                raise KeyError(normalized_hash)
+            connection.execute(
+                "DELETE FROM sentences WHERE video_hash = ?",
+                (normalized_hash,),
+            )
+            connection.executemany(
+                """
+                INSERT INTO sentences (
+                    video_hash, sentence_index, start_ms, end_ms, text,
+                    language, zh_text, reference_speech_duration_ms
+                ) VALUES (?, ?, ?, ?, ?, ?, ?, ?)
+                """,
+                [
+                    (
+                        normalized_hash,
+                        sentence.index,
+                        sentence.start_ms,
+                        sentence.end_ms,
+                        sentence.text,
+                        sentence.language,
+                        sentence.zh_text,
+                        sentence.reference_speech_duration_ms,
+                    )
+                    for sentence in document.sentences
                 ],
             )
+            connection.execute(
+                "UPDATE videos SET updated_at = ? WHERE video_hash = ?",
+                (utc_now(), normalized_hash),
+            )
+        saved = self.get_document(normalized_hash)
+        if saved is None:
+            raise KeyError(normalized_hash)
+        return saved
 
     def update_sentence_text(
         self,
@@ -364,7 +602,8 @@ class VideoRepository:
         with self._connect() as connection:
             cursor = connection.execute(
                 """
-                UPDATE sentences SET text = ?, language = COALESCE(?, language)
+                UPDATE sentences SET text = ?, language = COALESCE(?, language),
+                    zh_text = NULL
                 WHERE video_hash = ? AND sentence_index = ?
                 """,
                 (text.strip(), language, normalized_hash, sentence_index),
@@ -379,14 +618,7 @@ class VideoRepository:
                 "SELECT * FROM sentences WHERE video_hash = ? AND sentence_index = ?",
                 (normalized_hash, sentence_index),
             ).fetchone()
-        return SentenceBoundary(
-            index=row["sentence_index"],
-            start_ms=row["start_ms"],
-            end_ms=row["end_ms"],
-            text=row["text"],
-            language=row["language"],
-            reference_speech_duration_ms=row["reference_speech_duration_ms"],
-        )
+        return _sentence_from_row(row)
 
     def get_sentence(
         self,
@@ -404,14 +636,7 @@ class VideoRepository:
             ).fetchone()
         if row is None:
             return None
-        return SentenceBoundary(
-            index=row["sentence_index"],
-            start_ms=row["start_ms"],
-            end_ms=row["end_ms"],
-            text=row["text"],
-            language=row["language"],
-            reference_speech_duration_ms=row["reference_speech_duration_ms"],
-        )
+        return _sentence_from_row(row)
 
     def replace_sentence(
         self,
@@ -523,6 +748,10 @@ class VideoRepository:
                         replacement.reference_speech_duration_ms,
                     ),
                 )
+            connection.execute(
+                "UPDATE sentences SET zh_text = NULL WHERE video_hash = ?",
+                (normalized_hash,),
+            )
             connection.execute(
                 "UPDATE videos SET updated_at = ? WHERE video_hash = ?",
                 (utc_now(), normalized_hash),
@@ -536,17 +765,7 @@ class VideoRepository:
                 (normalized_hash, sentence_index, sentence_index + len(rows)),
             ).fetchall()
 
-        return [
-            SentenceBoundary(
-                index=row["sentence_index"],
-                start_ms=row["start_ms"],
-                end_ms=row["end_ms"],
-                text=row["text"],
-                language=row["language"],
-                reference_speech_duration_ms=row["reference_speech_duration_ms"],
-            )
-            for row in result_rows
-        ]
+        return [_sentence_from_row(row) for row in result_rows]
 
 
     def merge_with_next_sentence(
@@ -595,7 +814,7 @@ class VideoRepository:
                 """
                 UPDATE sentences
                 SET end_ms = ?, text = ?, language = ?,
-                    reference_speech_duration_ms = ?
+                    reference_speech_duration_ms = ?, zh_text = NULL
                 WHERE video_hash = ? AND sentence_index = ?
                 """,
                 (
@@ -648,14 +867,7 @@ class VideoRepository:
                 (normalized_hash, sentence_index),
             ).fetchone()
 
-        return SentenceBoundary(
-            index=row["sentence_index"],
-            start_ms=row["start_ms"],
-            end_ms=row["end_ms"],
-            text=row["text"],
-            language=row["language"],
-            reference_speech_duration_ms=row["reference_speech_duration_ms"],
-        )
+        return _sentence_from_row(row)
 
     def adjust_sentence_boundary(
         self,
@@ -760,17 +972,83 @@ class VideoRepository:
                 (normalized_hash, *updated_indexes),
             ).fetchall()
 
-        return [
-            SentenceBoundary(
-                index=row["sentence_index"],
-                start_ms=row["start_ms"],
-                end_ms=row["end_ms"],
-                text=row["text"],
-                language=row["language"],
-                reference_speech_duration_ms=row["reference_speech_duration_ms"],
+        return [_sentence_from_row(row) for row in result_rows]
+
+    def get_translation_glossary(self, video_hash: str) -> List[Tuple[str, str]]:
+        video = self.get_video(video_hash)
+        if video is None:
+            return []
+        return _parse_translation_glossary(video.get("translation_glossary"))
+
+    def save_translation_glossary(
+        self,
+        video_hash: str,
+        terms: Sequence[Any],
+    ) -> bool:
+        normalized_hash = normalize_video_hash(video_hash)
+        payload = json.dumps(
+            [
+                {"source": source, "zh": zh}
+                for source, zh in _normalize_translation_glossary(terms)
+            ],
+            ensure_ascii=False,
         )
-            for row in result_rows
-        ]
+        with self._connect() as connection:
+            cursor = connection.execute(
+                """
+                UPDATE videos
+                SET translation_glossary = ?, updated_at = ?
+                WHERE video_hash = ?
+                """,
+                (payload, utc_now(), normalized_hash),
+            )
+            return cursor.rowcount > 0
+
+    def clear_sentence_translations(
+        self,
+        video_hash: str,
+        indexes: Optional[Sequence[int]] = None,
+    ) -> None:
+        normalized_hash = normalize_video_hash(video_hash)
+        with self._connect() as connection:
+            if indexes is None:
+                connection.execute(
+                    "UPDATE sentences SET zh_text = NULL WHERE video_hash = ?",
+                    (normalized_hash,),
+                )
+                return
+            unique_indexes = list(dict.fromkeys(indexes))
+            if not unique_indexes:
+                return
+            connection.executemany(
+                """
+                UPDATE sentences SET zh_text = NULL
+                WHERE video_hash = ? AND sentence_index = ?
+                """,
+                [(normalized_hash, index) for index in unique_indexes],
+            )
+
+    def save_sentence_translations(
+        self,
+        video_hash: str,
+        translations: Dict[int, str],
+    ) -> Optional[SentenceBoundaryDocument]:
+        normalized_hash = normalize_video_hash(video_hash)
+        with self._connect() as connection:
+            for index, zh_text in translations.items():
+                connection.execute(
+                    """
+                    UPDATE sentences
+                    SET zh_text = ?
+                    WHERE video_hash = ? AND sentence_index = ?
+                    """,
+                    (zh_text.strip(), normalized_hash, index),
+                )
+            connection.execute(
+                "UPDATE videos SET updated_at = ? WHERE video_hash = ?",
+                (utc_now(), normalized_hash),
+            )
+        return self.get_document(normalized_hash)
 
     def create_dub_share(
         self,
@@ -984,6 +1262,22 @@ class VideoRepository:
             return None
         return self.get_user_by_id(row["id"])
 
+    def set_user_password(self, username: str, password: str) -> bool:
+        """Set a local password for either a normal or SSO-linked account."""
+        username = username.strip()
+        if not username:
+            return False
+        with self._connect() as connection:
+            cursor = connection.execute(
+                """
+                UPDATE users
+                SET password_hash = ?
+                WHERE username = ? COLLATE NOCASE
+                """,
+                (hash_password(password), username),
+            )
+        return cursor.rowcount > 0
+
     def create_session(self, user_id: str) -> str:
         token = new_token()
         now = datetime.now(timezone.utc)
@@ -1171,6 +1465,28 @@ class VideoRepository:
         return video
 
 
+
+def _english_fingerprint(document: SentenceBoundaryDocument) -> tuple:
+    return tuple(
+        (sentence.index, sentence.start_ms, sentence.end_ms, (sentence.text or "").strip())
+        for sentence in document.sentences
+    )
+
+
+def _sentence_from_row(row: sqlite3.Row) -> SentenceBoundary:
+    keys = set(row.keys())
+    zh_text = row["zh_text"] if "zh_text" in keys else None
+    return SentenceBoundary(
+        index=row["sentence_index"],
+        start_ms=row["start_ms"],
+        end_ms=row["end_ms"],
+        text=row["text"],
+        language=row["language"],
+        zh_text=zh_text,
+        reference_speech_duration_ms=row["reference_speech_duration_ms"],
+    )
+
+
 def _join_adjacent_sentence_text(
     left: str,
     right: str,
@@ -1206,3 +1522,61 @@ def _is_cjk_char(char: str) -> bool:
         or 0x20000 <= code <= 0x2CEAF
     )
 
+
+def _moss_segment_row(video_hash: str, index: int, segment: Dict[str, Any]) -> tuple:
+    start_seconds = float(segment["start_seconds"])
+    end_seconds = float(segment["end_seconds"])
+    words = segment.get("words")
+    words_json = json.dumps(words, ensure_ascii=False) if words else None
+    speaker = segment.get("speaker")
+    return (
+        video_hash,
+        index,
+        start_seconds,
+        end_seconds,
+        int(round(start_seconds * 1000)),
+        int(round(end_seconds * 1000)),
+        str(segment.get("text") or "").strip(),
+        str(speaker) if speaker is not None else None,
+        words_json,
+    )
+
+
+def _normalize_translation_glossary(terms: Sequence[Any]) -> List[Tuple[str, str]]:
+    cleaned: List[Tuple[str, str]] = []
+    seen: Dict[str, int] = {}
+    for item in terms or []:
+        source = ""
+        zh = ""
+        if isinstance(item, dict):
+            source = str(item.get("source") or "")
+            zh = str(item.get("zh") or "")
+        elif isinstance(item, (tuple, list)) and len(item) >= 2:
+            source = str(item[0] or "")
+            zh = str(item[1] or "")
+        source = source.strip()[:120]
+        zh = zh.strip()[:120]
+        if not source or not zh:
+            continue
+        key = source.casefold()
+        if key in seen:
+            cleaned[seen[key]] = (source, zh)
+            continue
+        seen[key] = len(cleaned)
+        cleaned.append((source, zh))
+        if len(cleaned) >= 50:
+            break
+    return cleaned
+
+
+def _parse_translation_glossary(raw: Any) -> List[Tuple[str, str]]:
+    if not raw:
+        return []
+    if isinstance(raw, str):
+        try:
+            raw = json.loads(raw)
+        except json.JSONDecodeError:
+            return []
+    if not isinstance(raw, list):
+        return []
+    return _normalize_translation_glossary(raw)

sentence_api/static/admin.css

@@ -81,7 +81,7 @@ table { width: 100%; border-collapse: collapse; font-size: 13px; }
 th, td { padding: 10px 12px; text-align: left; border-bottom: 1px solid #e3e7ea; vertical-align: middle; }
 th { position: sticky; top: 0; z-index: 1; color: #44515b; background: #edf1f3; font-weight: 600; }
 tr:last-child td { border-bottom: 0; }
-.actions-column { min-width: 280px; }
+.actions-column { min-width: 340px; }
 .actions { display: flex; gap: 6px; white-space: nowrap; }
 .actions button, .actions a { min-height: 30px; padding: 5px 9px; font-size: 12px; }
 .actions a { display: inline-flex; align-items: center; border: 1px solid #aab4bc; border-radius: 4px; color: #26323c; text-decoration: none; }
@@ -103,6 +103,9 @@ dialog {
 dialog::backdrop { background: rgb(20 25 28 / 55%); }
 .dialog-heading { padding: 16px 18px; display: flex; justify-content: space-between; align-items: center; border-bottom: 1px solid #dce1e5; }
 .dialog-heading span { font-size: 12px; color: #68747d; }
+.dialog-heading-actions { display: flex; align-items: center; gap: 8px; }
+.gap-row td { background: #fff8ee; color: #9a5b12; font-style: italic; }
+.missing-zh-row .zh-text { color: #a45d08; }
 .sentence-scroll { max-height: calc(100vh - 120px); overflow: auto; }
 .sentence-scroll textarea { width: min(600px, 45vw); min-width: 260px; min-height: 58px; resize: vertical; padding: 7px; border: 1px solid #b8c0c7; border-radius: 3px; }
 .sentence-scroll select { min-width: 90px; }
@@ -119,3 +122,11 @@ dialog::backdrop { background: rgb(20 25 28 / 55%); }
   #upload-form { grid-template-columns: 1fr; }
   .actions-column { min-width: 250px; }
 }
+
+.zh-text { margin-top: 6px; color: #66727d; font-size: 12px; white-space: pre-wrap; line-height: 1.45; }
+.retranslate-form { padding: 16px 18px 20px; display: grid; gap: 12px; }
+.retranslate-form p { margin: 0; font-size: 13px; color: #4f5b64; line-height: 1.55; }
+.retranslate-form label { display: grid; gap: 6px; font-size: 13px; color: #4f5b64; }
+.retranslate-form textarea { width: 100%; min-height: 160px; resize: vertical; }
+.retranslate-form code { font-size: 12px; }
+.retranslate-actions { display: flex; justify-content: flex-end; }

sentence_api/static/admin.html

@@ -96,13 +96,38 @@
     </section>
   </main>
 
+  <dialog id="retranslate-dialog">
+    <div class="dialog-heading">
+      <div>
+        <h2 id="retranslate-title">重新翻译中文字幕</h2>
+        <span id="retranslate-meta"></span>
+      </div>
+      <button id="close-retranslate-dialog" class="icon-button" type="button" aria-label="关闭">X</button>
+    </div>
+    <form id="retranslate-form" class="retranslate-form">
+      <p>会清空当前中文字幕并重新翻译。可指定专有名词译法，每行一条，格式为 <code>原文=中文</code>。例如 <code>Harimoto=哈瑞么托</code>，也可以把错误译名写在左边：<code>哈达马托=哈瑞么托</code>。</p>
+      <label>
+        <span>专有名词译法</span>
+        <textarea id="retranslate-terms" placeholder="Harimoto=哈瑞么托"></textarea>
+      </label>
+      <div class="retranslate-actions">
+        <button id="repair-translations" type="button">只补译缺失</button>
+        <button class="primary" type="submit">开始重新翻译</button>
+      </div>
+    </form>
+  </dialog>
+
   <dialog id="sentence-dialog">
     <div class="dialog-heading">
       <div>
         <h2 id="sentence-title">句子文本</h2>
         <span id="sentence-meta"></span>
       </div>
-      <button id="close-dialog" class="icon-button" type="button" aria-label="关闭">X</button>
+      <div class="dialog-heading-actions">
+        <button id="repair-gaps" type="button">修复空白</button>
+        <button id="repair-missing-zh" type="button">补译缺失</button>
+        <button id="close-dialog" class="icon-button" type="button" aria-label="关闭" title="关闭时批量处理暂存修改">X</button>
+      </div>
     </div>
     <div class="sentence-scroll">
       <table>
@@ -114,6 +139,6 @@
     </div>
   </dialog>
 
-  <script src="/static/admin.js?v=admin-merge-sentence-v1" defer></script>
+  <script src="/static/admin.js?v=admin-sentence-batch-v5" defer></script>
 </body>
 </html>

sentence_api/static/admin.js

@@ -1,7 +1,17 @@
+const GAP_REPAIR_MIN_MS = 8000;
+const CJK_RE = /[\u3400-\u4dbf\u4e00-\u9fff\uf900-\ufaff]/;
+
 const state = {
   adminKey: localStorage.getItem("oralTrainerAdminKey") || "",
   videos: [],
   users: [],
+  retranslateVideo: null,
+  sentenceVideo: null,
+  sentenceDraft: null,
+  sentenceBaseMeta: "",
+  sentencePending: false,
+  sentenceSaving: false,
+  sentenceNeedsProcessing: false,
 };
 
 const elements = {
@@ -22,6 +32,15 @@ const elements = {
   sentenceMeta: document.querySelector("#sentence-meta"),
   sentenceRows: document.querySelector("#sentence-rows"),
   closeDialog: document.querySelector("#close-dialog"),
+  repairGaps: document.querySelector("#repair-gaps"),
+  repairMissingZh: document.querySelector("#repair-missing-zh"),
+  retranslateDialog: document.querySelector("#retranslate-dialog"),
+  retranslateTitle: document.querySelector("#retranslate-title"),
+  retranslateMeta: document.querySelector("#retranslate-meta"),
+  retranslateForm: document.querySelector("#retranslate-form"),
+  retranslateTerms: document.querySelector("#retranslate-terms"),
+  repairTranslations: document.querySelector("#repair-translations"),
+  closeRetranslateDialog: document.querySelector("#close-retranslate-dialog"),
   userRows: document.querySelector("#user-rows"),
   userCount: document.querySelector("#user-count"),
   userEmptyState: document.querySelector("#user-empty-state"),
@@ -38,6 +57,26 @@ elements.refresh.addEventListener("click", () => {
   loadUsers();
 });
 elements.closeDialog.addEventListener("click", () => elements.dialog.close());
+elements.dialog.addEventListener("close", handleSentenceDialogClose);
+elements.repairGaps.addEventListener("click", () => {
+  if (state.sentencePending) {
+    elements.serviceState.textContent = "句子修改已暂存，请先关闭对话框完成处理";
+    return;
+  }
+  if (state.sentenceVideo) repairGaps(state.sentenceVideo, { reloadSentences: true });
+});
+elements.repairMissingZh.addEventListener("click", () => {
+  if (state.sentencePending) {
+    elements.serviceState.textContent = "句子修改已暂存，请先关闭对话框完成处理";
+    return;
+  }
+  if (state.sentenceVideo) repairMissingTranslations(state.sentenceVideo, { reloadSentences: true });
+});
+elements.closeRetranslateDialog.addEventListener("click", () => elements.retranslateDialog.close());
+elements.repairTranslations.addEventListener("click", () => {
+  if (state.retranslateVideo) repairMissingTranslations(state.retranslateVideo, { closeRetranslate: true });
+});
+elements.retranslateForm.addEventListener("submit", submitRetranslate);
 elements.uploadForm.addEventListener("submit", uploadVideo);
 elements.videoFile.addEventListener("change", () => {
   if (!elements.videoFile.files[0]) return;
@@ -181,7 +220,10 @@ function actionCell(video) {
   play.rel = "noopener";
   play.textContent = "播放";
   actions.append(play);
+  actions.append(button("改标题", () => renameVideo(video)));
   actions.append(button("句子", () => openSentences(video)));
+  actions.append(button("修复空白", () => repairGaps(video)));
+  actions.append(button("重新翻译", () => openRetranslate(video)));
   actions.append(button("重新处理", () => reprocess(video.video_hash)));
   const remove = button("删除", () => deleteVideo(video));
   remove.className = "danger";
@@ -198,6 +240,28 @@ function button(label, handler) {
   return result;
 }
 
+async function renameVideo(video) {
+  const title = window.prompt("请输入新的课程标题", video.title || "");
+  if (title === null) return;
+  const cleanedTitle = title.trim();
+  if (!cleanedTitle) {
+    elements.serviceState.textContent = "课程标题不能为空";
+    return;
+  }
+  if (cleanedTitle === video.title) return;
+  try {
+    await api(`/api/v1/admin/videos/${video.video_hash}/title`, {
+      method: "PUT",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify({ title: cleanedTitle }),
+    });
+    elements.serviceState.textContent = "课程标题已更新";
+    await loadVideos();
+  } catch (error) {
+    elements.serviceState.textContent = `标题修改失败: ${error.message}`;
+  }
+}
+
 function uploadVideo(event) {
   event.preventDefault();
   const file = document.querySelector("#video-file").files[0];
@@ -247,25 +311,306 @@ function readXhrError(request) {
 }
 
 async function openSentences(video) {
+  if (state.sentenceSaving) {
+    elements.serviceState.textContent = "句子批量处理中，请稍后再打开";
+    return;
+  }
+  if (
+    state.sentenceVideo?.video_hash === video.video_hash &&
+    Array.isArray(state.sentenceDraft)
+  ) {
+    renderSentenceDraft(video);
+    if (!elements.dialog.open) elements.dialog.showModal();
+    return;
+  }
   try {
     const detail = await api(`/api/v1/videos/${video.video_hash}`);
+    state.sentenceVideo = video;
     elements.sentenceTitle.textContent = detail.video.title;
-    const sentences = detail.boundaries?.sentences || [];
-    elements.sentenceMeta.textContent = `${sentences.length} 句 / ${detail.video.video_hash}`;
-    elements.sentenceRows.replaceChildren();
-    for (const sentence of sentences) {
-      elements.sentenceRows.append(sentenceRow(video, sentence, sentences.length));
-    }
+    state.sentenceDraft = (detail.boundaries?.sentences || []).map((sentence) => ({
+      ...sentence,
+      originStartMs: sentence.start_ms,
+      originEndMs: sentence.end_ms,
+      sourceStartMs: undefined,
+      sourceEndMs: undefined,
+    }));
+    state.sentencePending = false;
+    state.sentenceSaving = false;
+    state.sentenceNeedsProcessing = false;
+    const durationMs = detail.video.duration_ms || 0;
+    const sentences = state.sentenceDraft;
+    const gaps = collectSentenceGaps(sentences, durationMs);
+    const missingZh = sentences.filter(translationNeedsRepair).length;
+    const gapLabel = gaps.length ? ` / ${gaps.length} 段空白待修复` : "";
+    const zhLabel = missingZh ? ` / ${missingZh} 句未译` : "";
+    state.sentenceBaseMeta = `${sentences.length} 句${gapLabel}${zhLabel} / ${detail.video.video_hash}`;
+    renderSentenceDraft(video);
     if (!elements.dialog.open) elements.dialog.showModal();
   } catch (error) {
     elements.serviceState.textContent = `句子加载失败: ${error.message}`;
   }
 }
 
+function renderSentenceDraft(video) {
+  const sentences = state.sentenceDraft || [];
+  const durationMs = state.sentenceVideo?.duration_ms || 0;
+  const gaps = collectSentenceGaps(sentences, durationMs);
+  const missingZh = sentences.filter(translationNeedsRepair).length;
+  elements.sentenceMeta.textContent = [
+    state.sentenceBaseMeta,
+    state.sentencePending ? "有暂存修改" : "",
+    state.sentenceNeedsProcessing ? "已入库待处理" : "",
+    state.sentenceSaving ? "正在处理" : "",
+  ].filter(Boolean).join(" / ");
+  elements.repairGaps.disabled = gaps.length === 0 || state.sentencePending;
+  elements.repairMissingZh.disabled = missingZh === 0 || state.sentencePending;
+  elements.closeDialog.disabled = state.sentenceSaving;
+  elements.closeDialog.textContent = state.sentenceSaving ? "..." : "X";
+  elements.sentenceRows.replaceChildren();
+  let gapIndex = 0;
+  for (const sentence of sentences) {
+    if (gapIndex < gaps.length && gaps[gapIndex].end_ms === sentence.start_ms) {
+      elements.sentenceRows.append(gapRow(gaps[gapIndex]));
+      gapIndex += 1;
+    }
+    elements.sentenceRows.append(sentenceRow(video, sentence, sentences.length));
+  }
+  if (gapIndex < gaps.length) {
+    elements.sentenceRows.append(gapRow(gaps[gapIndex]));
+  }
+}
+
+async function handleSentenceDialogClose() {
+  if (!state.sentenceDraft || state.sentenceSaving) return;
+  if (!state.sentencePending && !state.sentenceNeedsProcessing) {
+    resetSentenceDraft();
+    return;
+  }
+  if (state.sentencePending) {
+    const persisted = await saveSentenceDraft({ processTimings: false });
+    if (!persisted) {
+      return;
+    }
+  }
+  const processed = await saveSentenceDraft({ processTimings: true });
+  if (!processed) {
+    return;
+  }
+  resetSentenceDraft();
+  loadVideos();
+}
+
+async function saveSentenceDraft({ processTimings = false } = {}) {
+  const video = state.sentenceVideo;
+  if (!video) return false;
+  if (!processTimings && !state.sentencePending) return true;
+  if (processTimings && !state.sentencePending && !state.sentenceNeedsProcessing) return true;
+  materializeSentenceDraftSplits(video);
+  state.sentenceSaving = true;
+  renderSentenceDraft(video);
+  try {
+    const body = {
+      process_timings: processTimings,
+      sentences: state.sentenceDraft.map(({
+        index,
+        originStartMs,
+        originEndMs,
+        sourceStartMs,
+        sourceEndMs,
+        ...sentence
+      }) => ({
+        ...sentence,
+        ...(sourceStartMs !== undefined && sourceEndMs !== undefined ? {
+          source_start_ms: sourceStartMs,
+          source_end_ms: sourceEndMs,
+        } : {}),
+      }),
+      ),
+    };
+    const saved = await api(
+      `/api/v1/admin/videos/${video.video_hash}/sentences/batch`,
+      {
+        method: "PUT",
+        headers: { "Content-Type": "application/json" },
+        body: JSON.stringify(body),
+      },
+    );
+    const previousByRange = new Map(
+      state.sentenceDraft.map((sentence) => [`${sentence.start_ms}:${sentence.end_ms}`, sentence]),
+    );
+    state.sentenceDraft = saved.map((sentence) => {
+      const previous = previousByRange.get(`${sentence.start_ms}:${sentence.end_ms}`);
+      return {
+        ...sentence,
+        originStartMs: previous?.originStartMs ?? sentence.start_ms,
+        originEndMs: previous?.originEndMs ?? sentence.end_ms,
+        sourceStartMs: previous?.sourceStartMs,
+        sourceEndMs: previous?.sourceEndMs,
+      };
+    });
+    state.sentencePending = false;
+    state.sentenceNeedsProcessing = !processTimings;
+    return true;
+  } catch (error) {
+    elements.serviceState.textContent = processTimings
+      ? `句子批量处理失败: ${error.message}`
+      : `句子入库失败: ${error.message}`;
+    return false;
+  } finally {
+    state.sentenceSaving = false;
+    if (state.sentenceDraft) renderSentenceDraft(video);
+  }
+}
+
+function materializeSentenceDraftSplits(video) {
+  for (const sentence of [...(state.sentenceDraft || [])]) {
+    if (/\r?\n/.test(sentence.text || "")) {
+      splitDraftSentence(video, sentence, sentence.text, sentence.language);
+    }
+  }
+}
+
+function resetSentenceDraft() {
+  state.sentenceDraft = null;
+  state.sentencePending = false;
+  state.sentenceSaving = false;
+  state.sentenceNeedsProcessing = false;
+  state.sentenceVideo = null;
+  state.sentenceBaseMeta = "";
+}
+
+function markSentenceDraftChanged(video) {
+  state.sentencePending = true;
+  renderSentenceDraft(video);
+}
+
+function markSentenceDraftPending() {
+  state.sentencePending = true;
+  elements.sentenceMeta.textContent = [
+    state.sentenceBaseMeta,
+    "有暂存修改",
+  ].filter(Boolean).join(" / ");
+  elements.repairGaps.disabled = true;
+  elements.repairMissingZh.disabled = true;
+}
+
+function joinDraftSentenceText(left, right, language) {
+  const leftText = (left || "").trim();
+  const rightText = (right || "").trim();
+  if (!leftText) return rightText;
+  if (!rightText) return leftText;
+  const lang = (language || "").toLowerCase();
+  if (lang.startsWith("zh") || lang.startsWith("ja") || lang.startsWith("ko")) {
+    return leftText + rightText;
+  }
+  const leftEnd = leftText.slice(-1);
+  const rightStart = rightText.slice(0, 1);
+  const cjk = /[\u3400-\u4dbf\u4e00-\u9fff\uf900-\ufaff]/;
+  if (cjk.test(leftEnd) || cjk.test(rightStart)) return leftText + rightText;
+  if (rightStart && ",.;:!?)]}、，。！？；：”’》〉】".includes(rightStart)) {
+    return leftText + rightText;
+  }
+  return `${leftText} ${rightText}`;
+}
+
+function splitDraftSentence(video, sentence, text, language) {
+  const lines = text.split(/\r?\n/).map((line) => line.trim()).filter(Boolean);
+  const duration = Math.max(1, sentence.end_ms - sentence.start_ms);
+  const weights = lines.map((line) => Math.max(1, line.replace(/\s+/g, "").length));
+  const totalWeight = Math.max(1, weights.reduce((total, weight) => total + weight, 0));
+  const sourceStart = sentence.sourceStartMs ?? sentence.originStartMs ?? sentence.start_ms;
+  const sourceEnd = sentence.sourceEndMs ?? sentence.originEndMs ?? sentence.end_ms;
+  const replacements = [];
+  let cursor = sentence.start_ms;
+  let remaining = lines.length;
+  let weightSum = 0;
+  for (const [lineIndex, line] of lines.entries()) {
+    const isLast = lineIndex === lines.length - 1;
+    weightSum += weights[lineIndex];
+    let end = sentence.end_ms;
+    if (!isLast) {
+      end = sentence.start_ms + Math.round(duration * weightSum / totalWeight);
+      end = Math.min(Math.max(end, cursor + 1), sentence.end_ms - (remaining - 1));
+    }
+    const sourceDuration = Math.max(
+      1,
+      (sentence.reference_speech_duration_ms || duration) * weights[lineIndex] / totalWeight,
+    );
+    replacements.push({
+      index: sentence.index + lineIndex,
+      start_ms: cursor,
+      end_ms: end,
+      text: line,
+      language: language || sentence.language,
+      zh_text: null,
+      reference_speech_duration_ms: Math.max(1, Math.round(sourceDuration)),
+      sourceStartMs: sourceStart,
+      sourceEndMs: sourceEnd,
+      originStartMs: sourceStart,
+      originEndMs: sourceEnd,
+    });
+    cursor = end;
+    remaining -= 1;
+  }
+  state.sentenceDraft.splice(sentence.index, 1, ...replacements);
+  renumberSentenceDraft();
+  markSentenceDraftChanged(video);
+}
+
+function renumberSentenceDraft() {
+  (state.sentenceDraft || []).forEach((sentence, index) => {
+    sentence.index = index;
+  });
+}
+
+function collectSentenceGaps(sentences, durationMs, minGapMs = GAP_REPAIR_MIN_MS) {
+  const gaps = [];
+  let cursor = 0;
+  for (const sentence of sentences) {
+    if (sentence.start_ms - cursor >= minGapMs) {
+      gaps.push({ start_ms: cursor, end_ms: sentence.start_ms });
+    }
+    cursor = Math.max(cursor, sentence.end_ms);
+  }
+  const endMs = durationMs || cursor;
+  if (endMs - cursor >= minGapMs) {
+    gaps.push({ start_ms: cursor, end_ms: endMs });
+  }
+  return gaps;
+}
+
+function gapRow(gap) {
+  const row = document.createElement("tr");
+  row.className = "gap-row";
+  row.append(
+    textCell(""),
+    textCell(`${formatDuration(gap.start_ms)} — ${formatDuration(gap.end_ms)}`),
+    textCell(`未识别空白 ${formatDuration(gap.end_ms - gap.start_ms)}`),
+    textCell(""),
+    textCell(""),
+  );
+  return row;
+}
+
+function translationNeedsRepair(sentence) {
+  const source = (sentence.text || "").trim();
+  const zh = (sentence.zh_text || "").trim();
+  if (!source) return false;
+  if (!zh) return true;
+  if (CJK_RE.test(zh)) return false;
+  return !CJK_RE.test(source);
+}
+
 function sentenceRow(video, sentence, totalCount) {
   const row = document.createElement("tr");
+  if (translationNeedsRepair(sentence)) row.className = "missing-zh-row";
   const textArea = document.createElement("textarea");
   textArea.value = sentence.text || "";
+  textArea.addEventListener("input", () => {
+    sentence.text = textArea.value;
+    sentence.zh_text = null;
+    markSentenceDraftPending();
+  });
   const language = document.createElement("select");
   for (const [value, label] of [["", "自动"], ["en", "英语"], ["zh", "中文"]]) {
     const option = document.createElement("option");
@@ -274,55 +619,77 @@ function sentenceRow(video, sentence, totalCount) {
     option.selected = value === (sentence.language || "");
     language.append(option);
   }
+  language.addEventListener("change", () => {
+    sentence.language = language.value || null;
+    markSentenceDraftPending();
+  });
   const textContainer = document.createElement("td");
-  textContainer.append(textArea);
+  const zhNote = document.createElement("div");
+  zhNote.className = "zh-text";
+  zhNote.textContent = translationNeedsRepair(sentence)
+    ? (sentence.zh_text ? `仍是英文：${sentence.zh_text}` : "暂无中文翻译")
+    : sentence.zh_text;
+  textContainer.append(textArea, zhNote);
   const languageContainer = document.createElement("td");
   languageContainer.append(language);
   const saveContainer = document.createElement("td");
   const save = button("保存", async () => {
     save.disabled = true;
+    save.textContent = "保存中";
     try {
-      const body = JSON.stringify({ text: textArea.value, language: language.value || null });
-      if (/\r?\n/.test(textArea.value)) {
-        await api(`/api/v1/admin/videos/${video.video_hash}/sentences/${sentence.index}/split`, {
-          method: "PUT",
-          headers: { "Content-Type": "application/json" },
-          body,
-        });
-        save.textContent = "已拆分";
-        await openSentences(video);
+      if (/\r?\n/.test(sentence.text || "")) {
+        splitDraftSentence(video, sentence, sentence.text, language.value || null);
       } else {
-        await api(`/api/v1/admin/videos/${video.video_hash}/sentences/${sentence.index}`, {
-          method: "PUT",
-          headers: { "Content-Type": "application/json" },
-          body,
-        });
-        save.textContent = "已保存";
+        sentence.text = textArea.value;
+        sentence.language = language.value || null;
+        sentence.zh_text = null;
+        markSentenceDraftChanged(video);
       }
-    } catch (error) {
-      elements.serviceState.textContent = `句子保存失败: ${error.message}`;
-      save.textContent = error.message;
+      await saveSentenceDraft({ processTimings: false });
     } finally {
-      window.setTimeout(() => { save.disabled = false; save.textContent = "保存"; }, 1600);
+      save.disabled = false;
+      save.textContent = "保存";
     }
   });
   saveContainer.append(save);
   if (sentence.index < totalCount - 1) {
     const merge = button("合并下句", async () => {
-      const preview = `${(textArea.value || sentence.text || "").trim()} + 下一句`;
+      const preview = `${(sentence.text || "").trim()} + 下一句`;
       const confirmed = window.confirm(`确定将第 ${sentence.index + 1} 句与下一句合并成一句吗？\n${preview}`);
       if (!confirmed) return;
+      const sourceStart = sentence.start_ms;
+      const sourceEnd = sentence.end_ms;
       merge.disabled = true;
+      merge.textContent = "入库中";
       try {
-        await api(`/api/v1/admin/videos/${video.video_hash}/sentences/${sentence.index}/merge`, {
-          method: "PUT",
+        if (state.sentencePending) {
+          const persisted = await saveSentenceDraft({ processTimings: false });
+          if (!persisted) return;
+        }
+        const targetIndex = state.sentenceDraft.findIndex(
+          (item) => item.start_ms === sourceStart && item.end_ms === sourceEnd,
+        );
+        if (targetIndex < 0) throw new Error("找不到要合并的句子");
+        const merged = await api(
+          `/api/v1/admin/videos/${video.video_hash}/sentences/${targetIndex}/merge`,
+          { method: "PUT" },
+        );
+        state.sentenceDraft.splice(targetIndex, 2, {
+          ...merged,
+          originStartMs: merged.start_ms,
+          originEndMs: merged.end_ms,
+          sourceStartMs: undefined,
+          sourceEndMs: undefined,
         });
-        merge.textContent = "已合并";
-        await openSentences(video);
+        renumberSentenceDraft();
+        state.sentencePending = false;
+        state.sentenceNeedsProcessing = true;
       } catch (error) {
         elements.serviceState.textContent = `句子合并失败: ${error.message}`;
-        merge.textContent = error.message;
-        window.setTimeout(() => { merge.disabled = false; merge.textContent = "合并下句"; }, 1600);
+      } finally {
+        merge.disabled = false;
+        merge.textContent = "合并下句";
+        renderSentenceDraft(video);
       }
     });
     saveContainer.append(merge);
@@ -350,41 +717,141 @@ function boundaryControl(video, sentence) {
   const status = document.createElement("span");
   status.className = "boundary-status";
 
-  async function adjust(deltaMs, changedButton) {
-    const buttons = [minus, plus];
-    buttons.forEach((item) => { item.disabled = true; });
-    status.textContent = "";
-    try {
-      await api(
-        `/api/v1/admin/videos/${video.video_hash}/sentences/${sentence.index}/boundary`,
-        {
-          method: "PUT",
-          headers: { "Content-Type": "application/json" },
-          body: JSON.stringify({ delta_ms: deltaMs }),
-        },
-      );
-      await openSentences(video);
-    } catch (error) {
-      elements.serviceState.textContent = `分割点调整失败: ${error.message}`;
-      status.textContent = "失败";
-      buttons.forEach((item) => { item.disabled = false; });
-      window.setTimeout(() => { status.textContent = ""; }, 1800);
-    }
-    if (changedButton) {
-      window.setTimeout(() => {
-        changedButton.textContent = deltaMs < 0 ? "−0.1" : "+0.1";
-      }, 900);
+  async function adjust(deltaMs) {
+    const next = state.sentenceDraft[sentence.index + 1];
+    if (!next) return;
+    const nextBoundary = sentence.end_ms + deltaMs;
+    if (
+      nextBoundary <= sentence.start_ms ||
+      nextBoundary >= next.end_ms
+    ) {
+      elements.serviceState.textContent = "分割点必须保留在两句时间范围内";
+      return;
     }
+    sentence.end_ms = nextBoundary;
+    next.start_ms = nextBoundary;
+    sentence.sourceStartMs = undefined;
+    sentence.sourceEndMs = undefined;
+    next.sourceStartMs = undefined;
+    next.sourceEndMs = undefined;
+    markSentenceDraftChanged(video);
+    await saveSentenceDraft({ processTimings: false });
   }
 
-  const minus = button("−0.1", (event) => adjust(event.shiftKey ? -10 : -100, minus));
-  const plus = button("+0.1", (event) => adjust(event.shiftKey ? 10 : 100, plus));
+  const minus = button("−0.1", (event) => adjust(event.shiftKey ? -10 : -100));
+  const plus = button("+0.1", (event) => adjust(event.shiftKey ? 10 : 100));
   minus.title = "结束点提前 0.1 秒；按住 Shift 微调 0.01 秒";
   plus.title = "结束点延后 0.1 秒；按住 Shift 微调 0.01 秒";
   container.append(minus, label, plus, status);
   return container;
 }
 
+async function openRetranslate(video) {
+  state.retranslateVideo = video;
+  elements.retranslateTitle.textContent = `重新翻译：${video.title}`;
+  elements.retranslateMeta.textContent = video.video_hash;
+  elements.retranslateTerms.value = "";
+  try {
+    const payload = await api(`/api/v1/admin/videos/${video.video_hash}/subtitle-translation`);
+    elements.retranslateTerms.value = formatTranslationTerms(payload.terms || []);
+    const missing = payload.missing_count || 0;
+    const status = missing
+      ? `${missing} 句未译`
+      : (payload.complete ? "中文字幕已就绪" : (payload.status === "running" ? "正在翻译" : "尚未翻译完成"));
+    elements.retranslateMeta.textContent = `${status} / ${video.video_hash}`;
+    elements.repairTranslations.disabled = missing === 0;
+  } catch (error) {
+    elements.serviceState.textContent = `加载翻译信息失败: ${error.message}`;
+  }
+  if (!elements.retranslateDialog.open) elements.retranslateDialog.showModal();
+}
+
+function formatTranslationTerms(terms) {
+  return (terms || []).map((item) => `${item.source}=${item.zh}`).join("\n");
+}
+
+function parseTranslationTerms(text) {
+  const terms = [];
+  for (const rawLine of (text || "").split(/\r?\n/)) {
+    const line = rawLine.trim();
+    if (!line || line.startsWith("#")) continue;
+    const match = line.match(/^(.+?)\s*(?:=|=>|：|:)\s*(.+)$/);
+    if (!match) {
+      throw new Error(`无法解析专有名词：${line}`);
+    }
+    const source = match[1].trim();
+    const zh = match[2].trim();
+    if (!source || !zh) {
+      throw new Error(`无法解析专有名词：${line}`);
+    }
+    terms.push({ source, zh });
+  }
+  return terms;
+}
+
+async function submitRetranslate(event) {
+  event.preventDefault();
+  const video = state.retranslateVideo;
+  if (!video) return;
+  let terms;
+  try {
+    terms = parseTranslationTerms(elements.retranslateTerms.value);
+  } catch (error) {
+    elements.serviceState.textContent = error.message;
+    return;
+  }
+  const submit = elements.retranslateForm.querySelector("button[type='submit']");
+  submit.disabled = true;
+  try {
+    const payload = await api(`/api/v1/admin/videos/${video.video_hash}/subtitle-translation`, {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify({ terms }),
+    });
+    const status = payload.complete ? "中文字幕已重新翻译" : "已开始重新翻译";
+    elements.serviceState.textContent = terms.length ? `${status}，已应用 ${terms.length} 条专有名词` : status;
+    elements.retranslateDialog.close();
+    await loadVideos();
+  } catch (error) {
+    elements.serviceState.textContent = `重新翻译失败: ${error.message}`;
+  } finally {
+    submit.disabled = false;
+  }
+}
+
+async function repairMissingTranslations(video, options = {}) {
+  const confirmed = window.confirm(
+    `将只补译“${video.title}”里缺失或仍是英文的中文，已有中文会保留。继续？`,
+  );
+  if (!confirmed) return;
+  const buttons = [elements.repairMissingZh, elements.repairTranslations].filter(Boolean);
+  buttons.forEach((button) => { button.disabled = true; });
+  elements.serviceState.textContent = "正在补译缺失的中文…";
+  try {
+    const payload = await api(`/api/v1/admin/videos/${video.video_hash}/subtitle-translation/repair`, {
+      method: "POST",
+    });
+    const attempted = payload.attempted_count || 0;
+    if (!attempted) {
+      elements.serviceState.textContent = "没有需要补译的句子";
+    } else if (payload.complete) {
+      elements.serviceState.textContent = `补译完成：处理了 ${attempted} 句`;
+    } else {
+      const remaining = payload.missing_count ? `，还剩 ${payload.missing_count} 句` : "";
+      elements.serviceState.textContent = `已开始补译 ${attempted} 句${remaining}，请稍后刷新`;
+    }
+    if (options.closeRetranslate) elements.retranslateDialog.close();
+    await loadVideos();
+    if (options.reloadSentences) await openSentences(video);
+  } catch (error) {
+    elements.serviceState.textContent = `补译失败: ${error.message}`;
+  } finally {
+    if (!options.reloadSentences) {
+      buttons.forEach((button) => { button.disabled = false; });
+    }
+  }
+}
+
 async function reprocess(videoHash) {
   try {
     await api(`/api/v1/admin/videos/${videoHash}/process`, { method: "POST" });
@@ -394,6 +861,36 @@ async function reprocess(videoHash) {
   }
 }
 
+async function repairGaps(video, options = {}) {
+  const confirmed = window.confirm(
+    `将对“${video.title}”中间隔超过 8 秒的空白重新识别。现有句子会保留，可能需要几分钟。继续？`,
+  );
+  if (!confirmed) return;
+  const button = elements.repairGaps;
+  if (button) button.disabled = true;
+  elements.serviceState.textContent = "正在重新识别空白段…";
+  try {
+    const payload = await api(`/api/v1/admin/videos/${video.video_hash}/repair-gaps`, {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify({ min_gap_ms: GAP_REPAIR_MIN_MS }),
+    });
+    if (payload.inserted_count) {
+      elements.serviceState.textContent = `空白修复完成：补了 ${payload.inserted_count} 句（${payload.transcribed_gaps}/${payload.gap_count} 段）`;
+    } else if (payload.gap_count) {
+      elements.serviceState.textContent = `发现 ${payload.gap_count} 段空白，但没有识别到新句子`;
+    } else {
+      elements.serviceState.textContent = "没有发现需要修复的空白段";
+    }
+    await loadVideos();
+    if (options.reloadSentences) await openSentences(video);
+  } catch (error) {
+    elements.serviceState.textContent = `空白修复失败: ${error.message}`;
+  } finally {
+    if (button && !options.reloadSentences) button.disabled = false;
+  }
+}
+
 async function deleteVideo(video) {
   if (!window.confirm(`删除“${video.title}”？`)) return;
   try {

sentence_api/tests/test_ai_tutor.py

@@ -1,6 +1,11 @@
 import pytest
 
-from sentence_api.ai_tutor import AITutor, AITutorError
+import json
+import logging
+import re
+import threading
+
+from sentence_api.ai_tutor import AITutor, AITutorError, _apply_translation_terms, _parse_translation_payload
 
 
 class FakeResponse:
@@ -71,3 +76,238 @@ def test_ai_tutor_rejects_unexpected_response(monkeypatch):
 
     with pytest.raises(AITutorError):
         tutor.explain("Hello.")
+
+
+
+def test_ai_tutor_translates_full_document(monkeypatch):
+    requests_payloads = []
+
+    def fake_post(endpoint, headers, json, timeout):
+        requests_payloads.append((endpoint, headers, json, timeout))
+        return FakeResponse(body={"choices": [{"message": {"content": '{"0":"你好。","1":"见到你很高兴。"}'}}]})
+
+    monkeypatch.setattr("sentence_api.ai_tutor.requests.post", fake_post)
+    tutor = AITutor("https://ai.test/v1", "secret", "test-model", 1.5)
+
+    translated = tutor.translate_sentences(
+        [(0, "Hello."), (1, "Nice to meet you.")],
+        course_title="Greetings",
+    )
+
+    assert translated == {0: "你好。", 1: "见到你很高兴。"}
+    _endpoint, _headers, payload, timeout = requests_payloads[0]
+    user_prompt = payload["messages"][1]["content"]
+    assert "[0] Hello." in user_prompt
+    assert "[1] Nice to meet you." in user_prompt
+    assert timeout == 120.0
+
+
+def test_ai_tutor_translation_timeout_can_be_configured(monkeypatch):
+    request_payloads = []
+
+    def fake_post(endpoint, headers, json, timeout):
+        request_payloads.append((json, timeout))
+        return _translation_response_from_prompt(json)
+
+    monkeypatch.setattr("sentence_api.ai_tutor.requests.post", fake_post)
+    tutor = AITutor(
+        "https://ai.test/v1",
+        "secret",
+        "test-model",
+        1,
+        translation_timeout_seconds=180,
+    )
+
+    translated = tutor.translate_sentences([(0, "Hello.")])
+
+    assert translated == {0: "中文0"}
+    assert request_payloads[0][1] == 180
+
+
+def _translation_response_from_prompt(payload):
+    user_prompt = payload["messages"][1]["content"]
+    section = user_prompt.split("待翻译句子：", 1)[1]
+    indexes = [int(match) for match in re.findall(r"^\[(\d+)\]", section, re.M)]
+    body = {str(index): f"中文{index}" for index in indexes}
+    return FakeResponse(
+        body={"choices": [{"message": {"content": json.dumps(body, ensure_ascii=False)}}]}
+    )
+
+
+def test_ai_tutor_translates_long_input_in_batches(monkeypatch):
+    requests_payloads = []
+    lock = threading.Lock()
+
+    def fake_post(endpoint, headers, json, timeout):
+        with lock:
+            requests_payloads.append(json)
+        return _translation_response_from_prompt(json)
+
+    monkeypatch.setattr("sentence_api.ai_tutor.requests.post", fake_post)
+    tutor = AITutor("https://ai.test/v1", "secret", "test-model", 1)
+    sentences = [(index, f"This is sentence {index} with enough text.") for index in range(30)]
+
+    translated = tutor.translate_sentences(sentences, course_title="Long talk")
+
+    assert translated == {index: f"中文{index}" for index in range(30)}
+    assert len(requests_payloads) >= 3
+    first_prompt = requests_payloads[0]["messages"][1]["content"]
+    assert "前文" in first_prompt or "后文" in first_prompt or "待翻译句子：" in first_prompt
+
+
+def test_ai_tutor_skips_cached_sentences_and_persists_windows(monkeypatch):
+    persisted = []
+
+    def fake_post(endpoint, headers, json, timeout):
+        return _translation_response_from_prompt(json)
+
+    monkeypatch.setattr("sentence_api.ai_tutor.requests.post", fake_post)
+    tutor = AITutor("https://ai.test/v1", "secret", "test-model", 1)
+    translated = tutor.translate_sentences(
+        [(0, "Hello."), (1, "Thanks."), (2, "Bye.")],
+        existing={0: "你好。"},
+        persist=persisted.append,
+        course_title="Greetings",
+    )
+
+    assert translated[0] == "你好。"
+    assert translated[1] == "中文1"
+    assert translated[2] == "中文2"
+    assert persisted
+    assert 0 not in persisted[0]
+
+
+def test_ai_tutor_splits_invalid_windows_and_retries(monkeypatch):
+    calls = {"count": 0}
+
+    def fake_post(endpoint, headers, json, timeout):
+        calls["count"] += 1
+        if calls["count"] == 1:
+            return FakeResponse(body={"choices": [{"message": {"content": "not-json"}}]})
+        return _translation_response_from_prompt(json)
+
+    monkeypatch.setattr("sentence_api.ai_tutor.requests.post", fake_post)
+    tutor = AITutor("https://ai.test/v1", "secret", "test-model", 1)
+    sentences = [(index, f"This is sentence {index} with enough text.") for index in range(12)]
+
+    translated = tutor.translate_sentences(sentences)
+
+    assert translated == {index: f"中文{index}" for index in range(12)}
+    assert calls["count"] >= 3
+
+
+def test_ai_tutor_includes_name_terms_and_rewrites_output(monkeypatch):
+    requests_payloads = []
+
+    def fake_post(endpoint, headers, json, timeout):
+        requests_payloads.append(json)
+        return FakeResponse(
+            body={"choices": [{"message": {"content": '{"0":"哈达马托击败了 Harimoto。"}'}}]}
+        )
+
+    monkeypatch.setattr("sentence_api.ai_tutor.requests.post", fake_post)
+    tutor = AITutor("https://ai.test/v1", "secret", "test-model", 1)
+
+    translated = tutor.translate_sentences(
+        [(0, "Harimoto won.")],
+        course_title="Table tennis",
+        terms=[("Harimoto", "哈瑞么托"), ("哈达马托", "哈瑞么托")],
+    )
+
+    assert translated == {0: "哈瑞么托击败了 哈瑞么托。"}
+    user_prompt = requests_payloads[0]["messages"][1]["content"]
+    assert "Harimoto => 哈瑞么托" in user_prompt
+    assert "哈达马托 => 哈瑞么托" in user_prompt
+
+
+def test_apply_translation_terms_replaces_english_and_wrong_chinese():
+    assert _apply_translation_terms(
+        "Harimoto won.",
+        "哈达马托击败了 Harimoto。",
+        [("Harimoto", "哈瑞么托"), ("哈达马托", "哈瑞么托")],
+    ) == "哈瑞么托击败了 哈瑞么托。"
+
+
+def test_parse_translation_payload_accepts_fenced_json_and_numbered_lines():
+    assert _parse_translation_payload('```json\n["一", "二"]\n```', 2) == {0: "一", 1: "二"}
+    assert _parse_translation_payload("1. 一\n2. 二", 2) == {0: "一", 1: "二"}
+    assert _parse_translation_payload('prefix {"12":"甲","13":"乙"} suffix', [12, 13]) == {
+        12: "甲",
+        13: "乙",
+    }
+
+
+def test_ai_tutor_retranslates_english_existing_cache(monkeypatch):
+    def fake_post(endpoint, headers, json, timeout):
+        return _translation_response_from_prompt(json)
+
+    monkeypatch.setattr("sentence_api.ai_tutor.requests.post", fake_post)
+    tutor = AITutor("https://ai.test/v1", "secret", "test-model", 1)
+    translated = tutor.translate_sentences(
+        [(0, "Hello."), (1, "Thanks.")],
+        existing={0: "Hello.", 1: "谢谢。"},
+    )
+
+    assert translated[0] == "中文0"
+    assert translated[1] == "谢谢。"
+
+
+def test_ai_tutor_retries_when_translation_stays_english(monkeypatch):
+    calls = {"count": 0}
+
+    def fake_post(endpoint, headers, json, timeout):
+        calls["count"] += 1
+        if calls["count"] == 1:
+            return FakeResponse(
+                body={"choices": [{"message": {"content": '{"0":"Hello."}'}}]}
+            )
+        return FakeResponse(
+            body={"choices": [{"message": {"content": '{"0":"你好。"}'}}]}
+        )
+
+    monkeypatch.setattr("sentence_api.ai_tutor.requests.post", fake_post)
+    tutor = AITutor("https://ai.test/v1", "secret", "test-model", 1)
+    translated = tutor.translate_sentences([(0, "Hello.")])
+
+    assert translated == {0: "你好。"}
+    assert calls["count"] >= 2
+
+
+def test_ai_tutor_logs_raw_response_when_translation_json_is_invalid(monkeypatch, caplog):
+    def fake_post(endpoint, headers, json, timeout):
+        return FakeResponse(
+            body={"choices": [{"message": {"content": "definitely not translation json"}}]}
+        )
+
+    monkeypatch.setattr("sentence_api.ai_tutor.requests.post", fake_post)
+    tutor = AITutor("https://ai.test/v1", "secret", "test-model", 1)
+
+    with caplog.at_level(logging.WARNING, logger="sentence_api.ai_tutor"):
+        with pytest.raises(AITutorError):
+            tutor.translate_sentences([(0, "Hello.")])
+
+    assert any(
+        "Translation response parse failed" in record.getMessage()
+        and "definitely not translation json" in record.getMessage()
+        for record in caplog.records
+    )
+
+
+def test_ai_tutor_logs_invalid_translation_before_repair(monkeypatch, caplog):
+    def fake_post(endpoint, headers, json, timeout):
+        return FakeResponse(
+            body={"choices": [{"message": {"content": '{"0":"Still English."}'}}]}
+        )
+
+    monkeypatch.setattr("sentence_api.ai_tutor.requests.post", fake_post)
+    tutor = AITutor("https://ai.test/v1", "secret", "test-model", 1)
+
+    with caplog.at_level(logging.WARNING, logger="sentence_api.ai_tutor"):
+        with pytest.raises(AITutorError):
+            tutor.translate_sentences([(0, "Hello.")])
+
+    assert any(
+        "Translation response failed validation" in record.getMessage()
+        and "Still English." in record.getMessage()
+        for record in caplog.records
+    )

sentence_api/tests/test_api.py

@@ -1,4 +1,6 @@
 import json
+import threading
+import time
 from dataclasses import replace
 
 from fastapi.testclient import TestClient
@@ -14,8 +16,20 @@ from sentence_api.store import BoundaryStore
 
 VIDEO_HASH = "a" * 64
 
+def wait_for_subtitle_translation(client, video_hash, timeout=5.0):
+    deadline = time.time() + timeout
+    last = None
+    while time.time() < deadline:
+        last = client.get(f"/api/v1/videos/{video_hash}/subtitle-translation")
+        if last.status_code == 200 and last.json().get("complete"):
+            return last
+        time.sleep(0.05)
+    raise AssertionError(f"subtitle translation did not complete: {getattr(last, 'text', last)}")
 
-def make_client(tmp_path):
+
+
+
+def make_client(tmp_path, admin_api_key=""):
     index_path = tmp_path / "boundaries.json"
     index_path.write_text(
         json.dumps({
@@ -36,6 +50,7 @@ def make_client(tmp_path):
         Settings.from_env(),
         data_dir=tmp_path / "data",
         legacy_boundaries_file=index_path,
+        admin_api_key=admin_api_key,
     )
     settings.ensure_directories()
     client = TestClient(
@@ -116,6 +131,27 @@ def test_adhoc_audio_boundaries_require_login(tmp_path):
     assert response.status_code == 401
 
 
+def test_admin_can_reset_sso_user_password(tmp_path):
+    client = make_client(tmp_path, admin_api_key="test-admin-key")
+    repository = client.app.state.video_repository
+    created = repository.create_sso_user(
+        username="bobcoc",
+        nickname="Bob",
+        sso_provider="d1kt",
+    )
+    assert created is not None
+    assert repository.authenticate_user("bobcoc", "anything-123") is None
+
+    response = client.put(
+        "/api/v1/admin/users/password",
+        json={"username": "bobcoc", "password": "new-password-123"},
+        headers={"X-Admin-Key": "test-admin-key"},
+    )
+
+    assert response.status_code == 204
+    assert repository.authenticate_user("bobcoc", "new-password-123") is not None
+
+
 def test_sentence_explanation_requires_login(tmp_path):
     index_path = tmp_path / "boundaries.json"
     index_path.write_text("{}", encoding="utf-8")
@@ -150,7 +186,7 @@ def test_sentence_explanation_uses_server_side_ai_client(tmp_path, monkeypatch):
         return "这句话表示打招呼。"
 
     class StubAITutor:
-        pass
+        available = True
 
     stub = StubAITutor()
     stub.explain = explain
@@ -180,6 +216,319 @@ def test_sentence_explanation_uses_server_side_ai_client(tmp_path, monkeypatch):
     ]
 
 
+
+def _ready_video(repository):
+    repository.upsert_upload(
+        video_hash=VIDEO_HASH,
+        title="Greetings",
+        filename="lesson.mp4",
+        stored_filename=f"{VIDEO_HASH}.mp4",
+        content_type="video/mp4",
+        size_bytes=5,
+        language="en",
+    )
+    repository.save_processing_result(
+        SentenceBoundaryDocument(
+            video_hash=VIDEO_HASH,
+            duration_ms=9000,
+            algorithm_version="test",
+            sentences=[
+                SentenceBoundary(index=0, start_ms=0, end_ms=4000, text="Hello.", language="en", reference_speech_duration_ms=3000),
+                SentenceBoundary(index=1, start_ms=4000, end_ms=9000, text="Nice to meet you.", language="en", reference_speech_duration_ms=4000),
+            ],
+        ),
+        None,
+    )
+
+
+def test_subtitle_translation_requires_login(tmp_path):
+    client = make_client(tmp_path)
+    client.headers.clear()
+    response = client.post(f"/api/v1/videos/{VIDEO_HASH}/subtitle-translation")
+    assert response.status_code == 401
+
+
+LOCAL_SUBTITLE_HASH = "b" * 64
+
+
+def _stub_translator(monkeypatch, client, requested=None):
+    requested = requested if requested is not None else []
+
+    def translate_sentences(sentences, **kwargs):
+        requested.append((sentences, kwargs))
+        return {index: f"中文{index}" for index, _text in sentences}
+
+    class StubAITutor:
+        available = True
+
+    stub = StubAITutor()
+    stub.translate_sentences = translate_sentences
+    monkeypatch.setattr(client.app.state, "ai_tutor", stub)
+    return requested
+
+
+def test_local_english_subtitles_can_be_uploaded_for_translation(tmp_path, monkeypatch):
+    client = make_client(tmp_path)
+    requested = _stub_translator(monkeypatch, client)
+
+    missing = client.post(f"/api/v1/videos/{LOCAL_SUBTITLE_HASH}/subtitle-translation")
+    assert missing.status_code == 404
+
+    response = client.post(
+        f"/api/v1/videos/{LOCAL_SUBTITLE_HASH}/subtitle-translation",
+        json={
+            "duration_ms": 9000,
+            "algorithm_version": "local-pc",
+            "sentences": [
+                {"index": 0, "start_ms": 0, "end_ms": 4000, "text": "Hello.", "language": "en"},
+                {"index": 1, "start_ms": 4000, "end_ms": 9000, "text": "Nice to meet you.", "language": "en"},
+            ],
+        },
+    )
+    assert response.status_code == 200, response.text
+    if not response.json().get("complete"):
+        response = wait_for_subtitle_translation(client, LOCAL_SUBTITLE_HASH)
+    payload = response.json()
+    assert payload["complete"] is True
+    assert payload["translations"] == [
+        {"index": 0, "zh_text": "中文0"},
+        {"index": 1, "zh_text": "中文1"},
+    ]
+    assert requested[0][0] == [(0, "Hello."), (1, "Nice to meet you.")]
+
+    stored = client.get(f"/api/v1/videos/{LOCAL_SUBTITLE_HASH}/subtitle-translation")
+    assert stored.status_code == 200
+    assert stored.json()["translations"] == payload["translations"]
+
+    courses = client.get("/api/v1/courses")
+    assert courses.status_code == 200
+    assert all(item["video_hash"] != LOCAL_SUBTITLE_HASH for item in courses.json())
+
+    videos = client.get("/api/v1/videos")
+    assert videos.status_code == 200
+    assert all(item["video_hash"] != LOCAL_SUBTITLE_HASH for item in videos.json()["videos"])
+
+
+def test_local_subtitle_upload_does_not_overwrite_course_sentences(tmp_path, monkeypatch):
+    client = make_client(tmp_path)
+    _ready_video(client.app.state.video_repository)
+    requested = _stub_translator(monkeypatch, client)
+
+    response = client.post(
+        f"/api/v1/videos/{VIDEO_HASH}/subtitle-translation",
+        json={
+            "duration_ms": 9000,
+            "algorithm_version": "local-pc",
+            "sentences": [
+                {"index": 0, "start_ms": 0, "end_ms": 4000, "text": "Changed.", "language": "en"},
+                {"index": 1, "start_ms": 4000, "end_ms": 9000, "text": "Also changed.", "language": "en"},
+            ],
+        },
+    )
+    assert response.status_code == 200, response.text
+    if not response.json().get("complete"):
+        response = wait_for_subtitle_translation(client, VIDEO_HASH)
+    assert requested[0][0] == [(0, "Hello."), (1, "Nice to meet you.")]
+    document = client.app.state.video_repository.get_document(VIDEO_HASH)
+    assert [sentence.text for sentence in document.sentences] == ["Hello.", "Nice to meet you."]
+
+
+def test_subtitle_translation_uses_full_document_and_is_cached(tmp_path, monkeypatch):
+    client = make_client(tmp_path)
+    _ready_video(client.app.state.video_repository)
+    requested = []
+
+    def translate_sentences(sentences, **kwargs):
+        requested.append((sentences, kwargs))
+        return {index: f"中文{index}" for index, _text in sentences}
+
+    class StubAITutor:
+        available = True
+
+    stub = StubAITutor()
+    stub.translate_sentences = translate_sentences
+    monkeypatch.setattr(client.app.state, "ai_tutor", stub)
+
+    response = client.post(f"/api/v1/videos/{VIDEO_HASH}/subtitle-translation")
+    assert response.status_code == 200, response.text
+    if not response.json().get("complete"):
+        response = wait_for_subtitle_translation(client, VIDEO_HASH)
+    payload = response.json()
+    assert payload["video_hash"] == VIDEO_HASH
+    assert payload["complete"] is True
+    assert payload["status"] == "ready"
+    assert payload["translations"] == [
+        {"index": 0, "zh_text": "中文0"},
+        {"index": 1, "zh_text": "中文1"},
+    ]
+    document = client.app.state.video_repository.get_document(VIDEO_HASH)
+    assert [sentence.zh_text for sentence in document.sentences] == ["中文0", "中文1"]
+
+    second = client.post(f"/api/v1/videos/{VIDEO_HASH}/subtitle-translation")
+    assert second.status_code == 200
+    assert second.json()["translations"] == payload["translations"]
+    assert second.json()["complete"] is True
+    assert len(requested) == 1
+    assert requested[0][0] == [(0, "Hello."), (1, "Nice to meet you.")]
+
+
+def test_subtitle_translation_is_shared_with_other_users(tmp_path, monkeypatch):
+    client = make_client(tmp_path)
+    _ready_video(client.app.state.video_repository)
+    requested = []
+
+    def translate_sentences(sentences, **kwargs):
+        requested.append((sentences, kwargs))
+        return {index: f"中文{index}" for index, _text in sentences}
+
+    class StubAITutor:
+        available = True
+
+    stub = StubAITutor()
+    stub.translate_sentences = translate_sentences
+    monkeypatch.setattr(client.app.state, "ai_tutor", stub)
+
+    first = client.post(f"/api/v1/videos/{VIDEO_HASH}/subtitle-translation")
+    assert first.status_code == 200, first.text
+    if not first.json().get("complete"):
+        first = wait_for_subtitle_translation(client, VIDEO_HASH)
+
+    stored = client.get(f"/api/v1/videos/{VIDEO_HASH}/subtitle-translation")
+    assert stored.status_code == 200
+    assert stored.json()["translations"] == first.json()["translations"]
+    assert stored.json()["complete"] is True
+
+    other = TestClient(client.app)
+    signup = other.post(
+        "/api/v1/auth/register",
+        json={"username": "viewer2", "password": "password-123"},
+    )
+    assert signup.status_code == 201, signup.text
+    other.headers.update({"Authorization": f"Bearer {signup.json()['token']}"})
+
+    shared = other.get(f"/api/v1/videos/{VIDEO_HASH}/subtitle-translation")
+    assert shared.status_code == 200
+    assert shared.json()["translations"] == [
+        {"index": 0, "zh_text": "中文0"},
+        {"index": 1, "zh_text": "中文1"},
+    ]
+    boundaries = other.get(f"/api/v1/videos/{VIDEO_HASH}/sentence-boundaries")
+    assert boundaries.status_code == 200
+    assert [item["zh_text"] for item in boundaries.json()["sentences"]] == ["中文0", "中文1"]
+
+    reused = other.post(f"/api/v1/videos/{VIDEO_HASH}/subtitle-translation")
+    assert reused.status_code == 200
+    assert reused.json()["translations"] == first.json()["translations"]
+    assert reused.json()["complete"] is True
+    assert len(requested) == 1
+
+
+
+def test_subtitle_translation_does_not_block_http_on_long_jobs(tmp_path, monkeypatch):
+    client = make_client(tmp_path)
+    _ready_video(client.app.state.video_repository)
+    started = threading.Event()
+    release = threading.Event()
+    requested = []
+
+    def translate_sentences(sentences, **kwargs):
+        requested.append(sentences)
+        persist = kwargs.get("persist")
+        if persist is not None:
+            persist({0: "中文0"})
+        started.set()
+        assert release.wait(timeout=5)
+        if persist is not None:
+            persist({1: "中文1"})
+        return {0: "中文0", 1: "中文1"}
+
+    class StubAITutor:
+        available = True
+
+    stub = StubAITutor()
+    stub.translate_sentences = translate_sentences
+    monkeypatch.setattr(client.app.state, "ai_tutor", stub)
+    client.app.state.subtitle_translations.ai_tutor = stub
+    client.app.state.subtitle_translations.post_wait_seconds = 0.2
+
+    response = client.post(f"/api/v1/videos/{VIDEO_HASH}/subtitle-translation")
+    assert started.wait(timeout=2)
+    assert response.status_code == 200, response.text
+    payload = response.json()
+    assert payload["complete"] is False
+    assert payload["status"] == "running"
+    if {"index": 0, "zh_text": "中文0"} not in payload["translations"]:
+        payload = client.get(f"/api/v1/videos/{VIDEO_HASH}/subtitle-translation").json()
+    assert {"index": 0, "zh_text": "中文0"} in payload["translations"]
+
+    stored = client.get(f"/api/v1/videos/{VIDEO_HASH}/subtitle-translation")
+    assert stored.status_code == 200
+    assert stored.json()["status"] == "running"
+    assert stored.json()["complete"] is False
+
+    release.set()
+    finished = wait_for_subtitle_translation(client, VIDEO_HASH)
+    assert finished.json()["complete"] is True
+    assert finished.json()["translations"] == [
+        {"index": 0, "zh_text": "中文0"},
+        {"index": 1, "zh_text": "中文1"},
+    ]
+    assert len(requested) == 1
+
+
+def test_admin_can_retranslate_subtitles_with_name_terms(tmp_path, monkeypatch):
+    client = make_client(tmp_path)
+    _ready_video(client.app.state.video_repository)
+    requested = []
+
+    def translate_sentences(sentences, **kwargs):
+        requested.append((sentences, kwargs.get("terms")))
+        terms = dict(kwargs.get("terms") or [])
+        return {
+            index: terms.get("Harimoto", "哈达马托") + f":{text}"
+            for index, text in sentences
+        }
+
+    class StubAITutor:
+        available = True
+
+    stub = StubAITutor()
+    stub.translate_sentences = translate_sentences
+    monkeypatch.setattr(client.app.state, "ai_tutor", stub)
+    client.app.state.subtitle_translations.ai_tutor = stub
+    client.app.state.subtitle_translations.post_wait_seconds = 0.2
+
+    first = client.post(f"/api/v1/videos/{VIDEO_HASH}/subtitle-translation")
+    if not first.json().get("complete"):
+        first = wait_for_subtitle_translation(client, VIDEO_HASH)
+    assert first.json()["translations"] == [
+        {"index": 0, "zh_text": "哈达马托:Hello."},
+        {"index": 1, "zh_text": "哈达马托:Nice to meet you."},
+    ]
+
+    response = client.post(
+        f"/api/v1/admin/videos/{VIDEO_HASH}/subtitle-translation",
+        json={"terms": [{"source": "Harimoto", "zh": "哈瑞么托"}]},
+    )
+    assert response.status_code == 200, response.text
+    if not response.json().get("complete"):
+        response = wait_for_subtitle_translation(client, VIDEO_HASH)
+        stored = client.get(f"/api/v1/admin/videos/{VIDEO_HASH}/subtitle-translation")
+        assert stored.status_code == 200
+        payload = stored.json()
+    else:
+        payload = response.json()
+    assert payload["translations"] == [
+        {"index": 0, "zh_text": "哈瑞么托:Hello."},
+        {"index": 1, "zh_text": "哈瑞么托:Nice to meet you."},
+    ]
+    assert payload["terms"] == [{"source": "Harimoto", "zh": "哈瑞么托"}]
+    assert requested[-1][1] == [("Harimoto", "哈瑞么托")]
+    assert client.app.state.video_repository.get_translation_glossary(VIDEO_HASH) == [
+        ("Harimoto", "哈瑞么托")
+    ]
+
+
 def test_oss_video_url_is_used_for_catalog_and_content(tmp_path):
     index_path = tmp_path / "boundaries.json"
     index_path.write_text("{}", encoding="utf-8")
@@ -344,6 +693,10 @@ def test_admin_can_split_sentence_and_reindexes_following_sentences(tmp_path):
         ),
         None,
     )
+    repository.save_sentence_translations(
+        VIDEO_HASH,
+        {0: "之前。", 1: "第一段第二段", 2: "下一句。"},
+    )
 
     class UnavailableTranscriber:
         available = False
@@ -375,6 +728,7 @@ def test_admin_can_split_sentence_and_reindexes_following_sentences(tmp_path):
         "Next sentence.",
     ]
     assert [item.index for item in document.sentences] == [0, 1, 2, 3]
+    assert all(not (item.zh_text or "").strip() for item in document.sentences)
 
 
 def test_admin_can_merge_adjacent_sentences_and_reindexes_following_sentences(tmp_path):
@@ -432,6 +786,11 @@ def test_admin_can_merge_adjacent_sentences_and_reindexes_following_sentences(tm
         ),
         None,
     )
+    repository.save_sentence_translations(
+        VIDEO_HASH,
+        {0: "之前。", 1: "第一段", 2: "第二段", 3: "下一句。"},
+    )
+    (settings.videos_dir / f"{VIDEO_HASH}.mp4").write_bytes(b"video")
 
     class UnavailableTranscriber:
         available = False
@@ -465,6 +824,7 @@ def test_admin_can_merge_adjacent_sentences_and_reindexes_following_sentences(tm
     assert [item.index for item in document.sentences] == [0, 1, 2]
     assert [item.start_ms for item in document.sentences] == [0, 2000, 9000]
     assert [item.end_ms for item in document.sentences] == [2000, 9000, 12000]
+    assert [item.zh_text for item in document.sentences] == ["之前。", None, "下一句。"]
 
     last_response = client.put(
         f"/api/v1/admin/videos/{VIDEO_HASH}/sentences/2/merge",
@@ -473,6 +833,257 @@ def test_admin_can_merge_adjacent_sentences_and_reindexes_following_sentences(tm
     assert last_response.json()["detail"] == "No next sentence to merge with."
 
 
+
+def test_admin_batch_applies_merge_and_split_once(tmp_path, monkeypatch):
+    settings = replace(Settings.from_env(), data_dir=tmp_path / "data")
+    settings.ensure_directories()
+    repository = VideoRepository(settings.database_path)
+    repository.upsert_upload(
+        video_hash=VIDEO_HASH,
+        title="Batch edit lesson",
+        filename="lesson.mp4",
+        stored_filename=f"{VIDEO_HASH}.mp4",
+        content_type="video/mp4",
+        size_bytes=5,
+        language="en",
+    )
+    repository.save_processing_result(
+        SentenceBoundaryDocument(
+            video_hash=VIDEO_HASH,
+            duration_ms=12000,
+            algorithm_version="test",
+            sentences=[
+                SentenceBoundary(index=0, start_ms=0, end_ms=2000, text="Before.", language="en", reference_speech_duration_ms=1800),
+                SentenceBoundary(index=1, start_ms=2000, end_ms=5000, text="First part", language="en", reference_speech_duration_ms=2800),
+                SentenceBoundary(index=2, start_ms=5000, end_ms=9000, text="second part", language="en", reference_speech_duration_ms=3600),
+                SentenceBoundary(index=3, start_ms=9000, end_ms=12000, text="Next sentence.", language="en", reference_speech_duration_ms=2800),
+            ],
+        ),
+        None,
+    )
+    repository.save_sentence_translations(
+        VIDEO_HASH,
+        {0: "之前。", 1: "第一段", 2: "第二段", 3: "下一句。"},
+    )
+    (settings.videos_dir / f"{VIDEO_HASH}.mp4").write_bytes(b"video")
+
+    calls = []
+    from sentence_api.transcription import Transcript, TranscriptionSegment, WordTimestamp
+
+    class OnceTranscriber:
+        available = True
+
+        def transcribe(self, audio_path, language=None):
+            calls.append(audio_path)
+            return Transcript(
+                text="Next sentence.",
+                segments=[
+                    TranscriptionSegment(
+                        start_seconds=0.0,
+                        end_seconds=1.0,
+                        text="Next sentence.",
+                        words=[
+                            WordTimestamp(0.0, 0.2, "Next"),
+                            WordTimestamp(0.2, 1.0, "sentence."),
+                        ],
+                    )
+                ],
+            )
+
+    monkeypatch.setattr(
+        "sentence_api.processing.extract_audio_segment",
+        lambda media_path, output_path, *, start_ms, duration_ms: output_path.write_bytes(b"audio"),
+    )
+    client = TestClient(
+        create_app(
+            BoundaryStore(tmp_path / "boundaries.json"),
+            settings=settings,
+            repository=repository,
+            transcriber=OnceTranscriber(),
+        )
+    )
+    response = client.put(
+        f"/api/v1/admin/videos/{VIDEO_HASH}/sentences/batch",
+        json={
+            "sentences": [
+                {"start_ms": 0, "end_ms": 2000, "text": "Before.", "language": "en", "reference_speech_duration_ms": 1800},
+                {"start_ms": 2000, "end_ms": 9000, "text": "First part second part", "language": "en", "reference_speech_duration_ms": 6400},
+                {"start_ms": 9000, "end_ms": 9200, "text": "Next", "language": "en", "reference_speech_duration_ms": 1000, "source_start_ms": 9000, "source_end_ms": 12000},
+                {"start_ms": 9200, "end_ms": 12000, "text": "sentence.", "language": "en", "reference_speech_duration_ms": 1000, "source_start_ms": 9000, "source_end_ms": 12000},
+            ]
+        },
+    )
+
+    assert response.status_code == 200, response.text
+    payload = response.json()
+    assert [item["text"] for item in payload] == [
+        "Before.",
+        "First part second part",
+        "Next",
+        "sentence.",
+    ]
+    assert [item["start_ms"] for item in payload] == [0, 2000, 9000, 9600]
+    assert [item["end_ms"] for item in payload] == [2000, 9000, 9600, 12000]
+    assert [item["reference_speech_duration_ms"] for item in payload] == [1800, 6400, 862, 1938]
+    assert len(calls) == 1
+    document = repository.get_document(VIDEO_HASH)
+    assert [item.text for item in document.sentences] == [
+        "Before.",
+        "First part second part",
+        "Next",
+        "sentence.",
+    ]
+    assert all(not (item.zh_text or "").strip() for item in document.sentences)
+
+
+def test_admin_batch_can_persist_provisional_split_without_processing(tmp_path, monkeypatch):
+    settings = replace(Settings.from_env(), data_dir=tmp_path / "data")
+    settings.ensure_directories()
+    repository = VideoRepository(settings.database_path)
+    repository.upsert_upload(
+        video_hash=VIDEO_HASH,
+        title="Provisional split lesson",
+        filename="lesson.mp4",
+        stored_filename=f"{VIDEO_HASH}.mp4",
+        content_type="video/mp4",
+        size_bytes=5,
+        language="en",
+    )
+    repository.save_processing_result(
+        SentenceBoundaryDocument(
+            video_hash=VIDEO_HASH,
+            duration_ms=12000,
+            algorithm_version="test",
+            sentences=[
+                SentenceBoundary(index=0, start_ms=0, end_ms=5000, text="First part second part", language="en", reference_speech_duration_ms=4600),
+                SentenceBoundary(index=1, start_ms=5000, end_ms=12000, text="Unchanged.", language="en", reference_speech_duration_ms=6400),
+            ],
+        ),
+        None,
+    )
+    repository.save_sentence_translations(VIDEO_HASH, {1: "未变化。"})
+    (settings.videos_dir / f"{VIDEO_HASH}.mp4").write_bytes(b"video")
+    from sentence_api.transcription import Transcript, TranscriptionSegment, WordTimestamp
+
+    calls = []
+
+    class DeferredTranscriber:
+        available = True
+        allowed = False
+
+        def transcribe(self, audio_path, language=None):
+            assert self.allowed, "provisional save must not transcribe"
+            calls.append(audio_path)
+            return Transcript(
+                text="First part second part",
+                segments=[
+                    TranscriptionSegment(
+                        start_seconds=0.0,
+                        end_seconds=1.0,
+                        text="First part second part",
+                        words=[
+                            WordTimestamp(0.0, 0.2, "First"),
+                            WordTimestamp(0.2, 1.0, "part second part"),
+                        ],
+                    )
+                ],
+            )
+
+    monkeypatch.setattr(
+        "sentence_api.processing.extract_audio_segment",
+        lambda media_path, output_path, *, start_ms, duration_ms: output_path.write_bytes(b"audio"),
+    )
+    client = TestClient(
+        create_app(
+            BoundaryStore(tmp_path / "boundaries.json"),
+            settings=settings,
+            repository=repository,
+            transcriber=DeferredTranscriber(),
+        )
+    )
+    response = client.put(
+        f"/api/v1/admin/videos/{VIDEO_HASH}/sentences/batch",
+        json={
+            "process_timings": False,
+            "sentences": [
+                {"start_ms": 0, "end_ms": 2500, "text": "First part", "language": "en", "zh_text": None, "reference_speech_duration_ms": 2300, "source_start_ms": 0, "source_end_ms": 5000},
+                {"start_ms": 2500, "end_ms": 5000, "text": "second part", "language": "en", "zh_text": None, "reference_speech_duration_ms": 2300, "source_start_ms": 0, "source_end_ms": 5000},
+                {"start_ms": 5000, "end_ms": 12000, "text": "Unchanged.", "language": "en", "zh_text": "未变化。", "reference_speech_duration_ms": 6400},
+            ],
+        },
+    )
+
+    assert response.status_code == 200, response.text
+    payload = response.json()
+    assert [item["text"] for item in payload] == ["First part", "second part", "Unchanged."]
+    assert [item["start_ms"] for item in payload] == [0, 2500, 5000]
+    assert [item["zh_text"] for item in payload] == [None, None, "未变化。"]
+    assert calls == []
+    document = repository.get_document(VIDEO_HASH)
+    assert [item.text for item in document.sentences] == ["First part", "second part", "Unchanged."]
+    assert [item.zh_text for item in document.sentences] == [None, None, "未变化。"]
+
+    client.app.state.transcriber.allowed = True
+    processed = client.put(
+        f"/api/v1/admin/videos/{VIDEO_HASH}/sentences/batch",
+        json={
+            "process_timings": True,
+            "sentences": [
+                {"start_ms": 0, "end_ms": 2500, "text": "First part", "language": "en", "zh_text": None, "reference_speech_duration_ms": 2300, "source_start_ms": 0, "source_end_ms": 5000},
+                {"start_ms": 2500, "end_ms": 5000, "text": "second part", "language": "en", "zh_text": None, "reference_speech_duration_ms": 2300, "source_start_ms": 0, "source_end_ms": 5000},
+                {"start_ms": 5000, "end_ms": 12000, "text": "Unchanged.", "language": "en", "zh_text": "未变化。", "reference_speech_duration_ms": 6400},
+            ],
+        },
+    )
+
+    assert processed.status_code == 200, processed.text
+    processed_payload = processed.json()
+    assert [item["start_ms"] for item in processed_payload] == [0, 1000, 5000]
+    assert [item["end_ms"] for item in processed_payload] == [1000, 5000, 12000]
+    assert len(calls) == 1
+
+
+def test_admin_sentence_merge_does_not_start_subtitle_translation(tmp_path, monkeypatch):
+    client = make_client(tmp_path)
+    _ready_video(client.app.state.video_repository)
+    requested = []
+
+    def translate_sentences(sentences, **kwargs):
+        requested.append([(index, text) for index, text in sentences])
+        return {index: f"中文:{text}" for index, text in sentences}
+
+    class StubAITutor:
+        available = True
+
+    stub = StubAITutor()
+    stub.translate_sentences = translate_sentences
+    monkeypatch.setattr(client.app.state, "ai_tutor", stub)
+    client.app.state.subtitle_translations.ai_tutor = stub
+    client.app.state.subtitle_translations.post_wait_seconds = 0.2
+
+    first = client.post(f"/api/v1/videos/{VIDEO_HASH}/subtitle-translation")
+    if not first.json().get("complete"):
+        first = wait_for_subtitle_translation(client, VIDEO_HASH)
+    assert first.json()["translations"] == [
+        {"index": 0, "zh_text": "中文:Hello."},
+        {"index": 1, "zh_text": "中文:Nice to meet you."},
+    ]
+
+    merge = client.put(f"/api/v1/admin/videos/{VIDEO_HASH}/sentences/0/merge")
+    assert merge.status_code == 200, merge.text
+    assert merge.json()["text"] == "Hello. Nice to meet you."
+
+    snapshot = client.get(f"/api/v1/videos/{VIDEO_HASH}/subtitle-translation")
+    assert snapshot.status_code == 200
+    assert snapshot.json()["status"] == "idle"
+    assert snapshot.json()["complete"] is False
+    assert snapshot.json()["translations"] == []
+    assert len(requested) == 1
+    document = client.app.state.video_repository.get_document(VIDEO_HASH)
+    assert [sentence.text for sentence in document.sentences] == ["Hello. Nice to meet you."]
+    assert [sentence.zh_text for sentence in document.sentences] == [None]
+
+
 def test_admin_boundary_adjustment_updates_linked_next_start(tmp_path):
     settings = replace(Settings.from_env(), data_dir=tmp_path / "data")
     settings.ensure_directories()
@@ -540,3 +1151,264 @@ def test_admin_boundary_adjustment_updates_linked_next_start(tmp_path):
     document = repository.get_document(VIDEO_HASH)
     assert [item.end_ms for item in document.sentences] == [4250, 9000]
     assert [item.start_ms for item in document.sentences] == [0, 4250]
+
+
+def test_admin_can_update_course_title(tmp_path):
+    settings = replace(Settings.from_env(), data_dir=tmp_path / "data")
+    settings.ensure_directories()
+    repository = VideoRepository(settings.database_path)
+    repository.upsert_upload(
+        video_hash=VIDEO_HASH,
+        title="Old course title",
+        filename="lesson.mp4",
+        stored_filename=f"{VIDEO_HASH}.mp4",
+        content_type="video/mp4",
+        size_bytes=5,
+        language="en",
+    )
+    client = TestClient(
+        create_app(
+            BoundaryStore(tmp_path / "boundaries.json"),
+            settings=settings,
+            repository=repository,
+        )
+    )
+
+    response = client.put(
+        f"/api/v1/admin/videos/{VIDEO_HASH}/title",
+        json={"title": "  New course title  "},
+    )
+
+    assert response.status_code == 200
+    assert response.json()["title"] == "New course title"
+    assert repository.get_video(VIDEO_HASH)["title"] == "New course title"
+
+    missing_response = client.put(
+        f"/api/v1/admin/videos/{'b' * 64}/title",
+        json={"title": "Missing course"},
+    )
+    assert missing_response.status_code == 404
+
+
+
+def test_admin_can_repair_sentence_gaps(tmp_path, monkeypatch):
+    settings = replace(Settings.from_env(), data_dir=tmp_path / "data")
+    settings.ensure_directories()
+    repository = VideoRepository(settings.database_path)
+    repository.upsert_upload(
+        video_hash=VIDEO_HASH,
+        title="Repair gaps lesson",
+        filename="lesson.mp4",
+        stored_filename=f"{VIDEO_HASH}.mp4",
+        content_type="video/mp4",
+        size_bytes=5,
+        language="en",
+    )
+    (settings.videos_dir / f"{VIDEO_HASH}.mp4").write_bytes(b"fake-video")
+    repository.save_processing_result(
+        SentenceBoundaryDocument(
+            video_hash=VIDEO_HASH,
+            duration_ms=18000,
+            algorithm_version="test",
+            sentences=[
+                SentenceBoundary(
+                    index=0,
+                    start_ms=0,
+                    end_ms=2000,
+                    text="Hello.",
+                    language="en",
+                    reference_speech_duration_ms=1800,
+                ),
+                SentenceBoundary(
+                    index=1,
+                    start_ms=15000,
+                    end_ms=18000,
+                    text="Goodbye.",
+                    language="en",
+                    reference_speech_duration_ms=2500,
+                ),
+            ],
+        ),
+        "Hello. Goodbye.",
+    )
+    repository.save_sentence_translations(VIDEO_HASH, {0: "你好。", 1: "再见。"})
+
+    from sentence_api.transcription import Transcript, TranscriptionSegment, _wav_bytes
+    import numpy as np
+
+    class GapTranscriber:
+        available = True
+
+        def transcribe(self, audio_path, language=None):
+            return Transcript(
+                text="The missing sentence.",
+                segments=[
+                    TranscriptionSegment(
+                        start_seconds=1.0,
+                        end_seconds=5.0,
+                        text="The missing sentence.",
+                    )
+                ],
+            )
+
+    def fake_extract(media_path, output_path, *, start_ms, duration_ms):
+        sample_rate = 16000
+        t = np.arange(int(max(1.0, duration_ms / 1000) * sample_rate)) / sample_rate
+        samples = (0.25 * np.sin(2 * np.pi * 220 * t)).astype(np.float32)
+        output_path.write_bytes(_wav_bytes(samples, sample_rate))
+
+    monkeypatch.setattr("sentence_api.processing.extract_audio_segment", fake_extract)
+    client = TestClient(
+        create_app(
+            BoundaryStore(tmp_path / "boundaries.json"),
+            settings=settings,
+            repository=repository,
+            transcriber=GapTranscriber(),
+        )
+    )
+
+    response = client.post(f"/api/v1/admin/videos/{VIDEO_HASH}/repair-gaps")
+
+    assert response.status_code == 200, response.text
+    payload = response.json()
+    assert payload["gap_count"] == 1
+    assert payload["inserted_count"] == 1
+    assert payload["sentence_count"] == 3
+    document = repository.get_document(VIDEO_HASH)
+    assert [item.text for item in document.sentences] == [
+        "Hello.",
+        "The missing sentence.",
+        "Goodbye.",
+    ]
+    assert [item.zh_text for item in document.sentences] == ["你好。", None, "再见。"]
+
+
+def test_admin_gap_repair_requires_transcriber(tmp_path):
+    settings = replace(Settings.from_env(), data_dir=tmp_path / "data")
+    settings.ensure_directories()
+    repository = VideoRepository(settings.database_path)
+    repository.upsert_upload(
+        video_hash=VIDEO_HASH,
+        title="Repair gaps lesson",
+        filename="lesson.mp4",
+        stored_filename=f"{VIDEO_HASH}.mp4",
+        content_type="video/mp4",
+        size_bytes=5,
+        language="en",
+    )
+    (settings.videos_dir / f"{VIDEO_HASH}.mp4").write_bytes(b"fake-video")
+    repository.save_processing_result(
+        SentenceBoundaryDocument(
+            video_hash=VIDEO_HASH,
+            duration_ms=18000,
+            algorithm_version="test",
+            sentences=[
+                SentenceBoundary(index=0, start_ms=0, end_ms=2000, text="Hello."),
+                SentenceBoundary(index=1, start_ms=15000, end_ms=18000, text="Goodbye."),
+            ],
+        ),
+        None,
+    )
+
+    class UnavailableTranscriber:
+        available = False
+
+    client = TestClient(
+        create_app(
+            BoundaryStore(tmp_path / "boundaries.json"),
+            settings=settings,
+            repository=repository,
+            transcriber=UnavailableTranscriber(),
+        )
+    )
+
+    response = client.post(f"/api/v1/admin/videos/{VIDEO_HASH}/repair-gaps")
+    assert response.status_code == 503
+
+
+def test_admin_can_repair_missing_and_english_translations(tmp_path, monkeypatch):
+    client = make_client(tmp_path)
+    repository = client.app.state.video_repository
+    _ready_video(repository)
+    repository.save_sentence_translations(
+        VIDEO_HASH,
+        {0: "你好。", 1: "Nice to meet you."},
+    )
+
+    requested = []
+
+    def translate_sentences(sentences, **kwargs):
+        requested.append([index for index, _text in sentences])
+        existing = dict(kwargs.get("existing") or {})
+        result = {}
+        for index, _text in sentences:
+            current = existing.get(index, "")
+            if current and any("\u4e00" <= char <= "\u9fff" for char in current):
+                result[index] = current
+            else:
+                result[index] = f"补译{index}"
+        return result
+
+    class StubAITutor:
+        available = True
+
+    stub = StubAITutor()
+    stub.translate_sentences = translate_sentences
+    monkeypatch.setattr(client.app.state, "ai_tutor", stub)
+    client.app.state.subtitle_translations.ai_tutor = stub
+    client.app.state.subtitle_translations.post_wait_seconds = 0.2
+
+    snapshot = client.get(f"/api/v1/admin/videos/{VIDEO_HASH}/subtitle-translation")
+    assert snapshot.status_code == 200
+    assert snapshot.json()["missing_count"] == 1
+
+    response = client.post(
+        f"/api/v1/admin/videos/{VIDEO_HASH}/subtitle-translation/repair"
+    )
+    assert response.status_code == 200, response.text
+    payload = response.json()
+    if not payload.get("complete"):
+        finished = wait_for_subtitle_translation(client, VIDEO_HASH)
+        payload = client.get(f"/api/v1/admin/videos/{VIDEO_HASH}/subtitle-translation").json()
+        payload["attempted_count"] = response.json()["attempted_count"]
+        payload["complete"] = finished.json()["complete"]
+
+    assert payload["attempted_count"] == 1
+    assert payload["complete"] is True
+    assert payload["missing_count"] == 0
+    document = repository.get_document(VIDEO_HASH)
+    assert [sentence.zh_text for sentence in document.sentences] == ["你好。", "补译1"]
+    assert requested
+
+
+def test_admin_translation_repair_keeps_complete_chinese(tmp_path, monkeypatch):
+    client = make_client(tmp_path)
+    repository = client.app.state.video_repository
+    _ready_video(repository)
+    repository.save_sentence_translations(
+        VIDEO_HASH,
+        {0: "你好。", 1: "见到你很高兴。"},
+    )
+    called = {"count": 0}
+
+    def translate_sentences(sentences, **kwargs):
+        called["count"] += 1
+        return {index: f"新{index}" for index, _text in sentences}
+
+    class StubAITutor:
+        available = True
+
+    stub = StubAITutor()
+    stub.translate_sentences = translate_sentences
+    monkeypatch.setattr(client.app.state, "ai_tutor", stub)
+    client.app.state.subtitle_translations.ai_tutor = stub
+
+    response = client.post(
+        f"/api/v1/admin/videos/{VIDEO_HASH}/subtitle-translation/repair"
+    )
+    assert response.status_code == 200, response.text
+    assert response.json()["attempted_count"] == 0
+    assert response.json()["missing_count"] == 0
+    assert called["count"] == 0
+    document = repository.get_document(VIDEO_HASH)
+    assert [sentence.zh_text for sentence in document.sentences] == ["你好。", "见到你很高兴。"]

sentence_api/tests/test_processing.py

@@ -2,10 +2,14 @@ import numpy as np
 
 from sentence_api.processing import (
     MOSS_ALGORITHM_VERSION,
+    VideoProcessor,
     document_from_transcript,
+    find_sentence_gaps,
+    _plan_gap_chunks,
     split_sentence_text,
 )
-from sentence_api.models import SentenceBoundary
+from sentence_api.models import SentenceBoundary, SentenceBoundaryDocument
+from sentence_api.repository import VideoRepository
 from sentence_api.transcription import (
     Transcript,
     TranscriptionSegment,
@@ -158,3 +162,415 @@ def test_split_sentence_text_uses_whisper_word_timestamps(tmp_path):
         ("Alpha beta", 1000, 3000),
         ("Gamma delta", 3000, 5000),
     ]
+
+
+def test_processor_persists_original_moss_segments(tmp_path, monkeypatch):
+    from sentence_api.config import Settings
+    from dataclasses import replace
+
+    settings = replace(Settings.from_env(), data_dir=tmp_path / "data")
+    settings.ensure_directories()
+    media_path = settings.videos_dir / f"{VIDEO_HASH}.mp4"
+    media_path.write_bytes(b"video")
+    repository = VideoRepository(settings.database_path)
+    repository.upsert_upload(
+        video_hash=VIDEO_HASH,
+        title="MOSS lesson",
+        filename="lesson.mp4",
+        stored_filename=f"{VIDEO_HASH}.mp4",
+        content_type="video/mp4",
+        size_bytes=5,
+        language="en",
+    )
+    transcript = Transcript(
+        text="Hello world. Next.",
+        segments=[
+            TranscriptionSegment(
+                start_seconds=0.0,
+                end_seconds=1.2,
+                text="Hello world.",
+                speaker="S01",
+                words=[WordTimestamp(0.0, 0.5, "Hello"), WordTimestamp(0.5, 1.2, "world.")],
+            ),
+            TranscriptionSegment(
+                start_seconds=1.2,
+                end_seconds=2.4,
+                text="Next.",
+                speaker="S01",
+            ),
+        ],
+    )
+
+    class AvailableTranscriber:
+        available = True
+
+        def transcribe(self, audio_path, language=None):
+            return transcript
+
+    monkeypatch.setattr("sentence_api.processing._ensure_webm_cues_front", lambda path: None)
+    monkeypatch.setattr(
+        "sentence_api.processing.extract_audio",
+        lambda src, dst: dst.write_bytes(b"RIFF"),
+    )
+    monkeypatch.setattr("sentence_api.processing._media_duration_ms", lambda path: 2400)
+    monkeypatch.setattr(
+        "sentence_api.processing.document_from_transcript",
+        lambda **kwargs: SentenceBoundaryDocument(
+            video_hash=kwargs["video_hash"],
+            duration_ms=2400,
+            algorithm_version=MOSS_ALGORITHM_VERSION,
+            sentences=[
+                SentenceBoundary(
+                    index=0,
+                    start_ms=0,
+                    end_ms=1200,
+                    text="Hello world.",
+                    language="en",
+                    reference_speech_duration_ms=1100,
+                ),
+                SentenceBoundary(
+                    index=1,
+                    start_ms=1200,
+                    end_ms=2400,
+                    text="Next.",
+                    language="en",
+                    reference_speech_duration_ms=1100,
+                ),
+            ],
+        ),
+    )
+
+    VideoProcessor(settings, repository, AvailableTranscriber()).process(VIDEO_HASH)
+
+    segments = repository.get_moss_segments(VIDEO_HASH)
+    assert [(item["start_ms"], item["end_ms"], item["text"], item["speaker"]) for item in segments] == [
+        (0, 1200, "Hello world.", "S01"),
+        (1200, 2400, "Next.", "S01"),
+    ]
+    assert segments[0]["words"] == [
+        {"start_seconds": 0.0, "end_seconds": 0.5, "text": "Hello"},
+        {"start_seconds": 0.5, "end_seconds": 1.2, "text": "world."},
+    ]
+    assert segments[1]["words"] is None
+
+
+def test_reprocess_without_moss_clears_original_segments(tmp_path, monkeypatch):
+    from sentence_api.config import Settings
+    from dataclasses import replace
+
+    settings = replace(Settings.from_env(), data_dir=tmp_path / "data")
+    settings.ensure_directories()
+    media_path = settings.videos_dir / f"{VIDEO_HASH}.mp4"
+    media_path.write_bytes(b"video")
+    repository = VideoRepository(settings.database_path)
+    repository.upsert_upload(
+        video_hash=VIDEO_HASH,
+        title="MOSS lesson",
+        filename="lesson.mp4",
+        stored_filename=f"{VIDEO_HASH}.mp4",
+        content_type="video/mp4",
+        size_bytes=5,
+        language="en",
+    )
+    repository.save_processing_result(
+        SentenceBoundaryDocument(
+            video_hash=VIDEO_HASH,
+            duration_ms=2400,
+            algorithm_version=MOSS_ALGORITHM_VERSION,
+            sentences=[
+                SentenceBoundary(
+                    index=0,
+                    start_ms=0,
+                    end_ms=2400,
+                    text="Hello.",
+                    language="en",
+                    reference_speech_duration_ms=2000,
+                )
+            ],
+        ),
+        "Hello.",
+        moss_segments=[
+            {
+                "start_seconds": 0.0,
+                "end_seconds": 2.4,
+                "text": "Hello.",
+                "speaker": "S01",
+            }
+        ],
+    )
+    assert repository.get_moss_segments(VIDEO_HASH)
+
+    class UnavailableTranscriber:
+        available = False
+
+    monkeypatch.setattr("sentence_api.processing._ensure_webm_cues_front", lambda path: None)
+    monkeypatch.setattr(
+        "sentence_api.processing.make_entry",
+        lambda *args, **kwargs: (
+            {
+                "duration_ms": 2400,
+                "sentences": [
+                    SentenceBoundary(
+                        index=0,
+                        start_ms=0,
+                        end_ms=2400,
+                        text=None,
+                        language="en",
+                    )
+                ],
+            },
+            {},
+        ),
+    )
+    VideoProcessor(settings, repository, UnavailableTranscriber()).process(VIDEO_HASH)
+    assert repository.get_moss_segments(VIDEO_HASH) == []
+
+
+def test_mark_processing_clears_previous_sentence_data(tmp_path):
+    repository = VideoRepository(tmp_path / "oral.sqlite3")
+    repository.upsert_upload(
+        video_hash=VIDEO_HASH,
+        title="MOSS lesson",
+        filename="lesson.mp4",
+        stored_filename=f"{VIDEO_HASH}.mp4",
+        content_type="video/mp4",
+        size_bytes=5,
+        language="en",
+    )
+    repository.save_processing_result(
+        SentenceBoundaryDocument(
+            video_hash=VIDEO_HASH,
+            duration_ms=2400,
+            algorithm_version=MOSS_ALGORITHM_VERSION,
+            sentences=[
+                SentenceBoundary(
+                    index=0,
+                    start_ms=0,
+                    end_ms=1200,
+                    text="Old first.",
+                    language="en",
+                    reference_speech_duration_ms=1000,
+                ),
+                SentenceBoundary(
+                    index=1,
+                    start_ms=1200,
+                    end_ms=2400,
+                    text="Old second.",
+                    language="en",
+                    reference_speech_duration_ms=1000,
+                ),
+            ],
+        ),
+        "Old first. Old second.",
+        moss_segments=[
+            {
+                "start_seconds": 0.0,
+                "end_seconds": 1.2,
+                "text": "Old first.",
+                "speaker": "S01",
+            }
+        ],
+    )
+
+    repository.mark_processing(VIDEO_HASH)
+
+    video = repository.get_video(VIDEO_HASH)
+    document = repository.get_document(VIDEO_HASH)
+    assert video["status"] == "processing"
+    assert video["transcription"] is None
+    assert video["sentence_count"] == 0
+    assert document is not None
+    assert document.sentences == []
+    assert repository.get_moss_segments(VIDEO_HASH) == []
+
+
+
+
+def test_find_sentence_gaps_detects_leading_middle_and_trailing():
+    document = SentenceBoundaryDocument(
+        video_hash=VIDEO_HASH,
+        duration_ms=60000,
+        algorithm_version="test",
+        sentences=[
+            SentenceBoundary(index=0, start_ms=10000, end_ms=12000, text="A"),
+            SentenceBoundary(index=1, start_ms=25000, end_ms=27000, text="B"),
+        ],
+    )
+
+    assert find_sentence_gaps(document, min_gap_ms=8000) == [
+        (0, 10000),
+        (12000, 25000),
+        (27000, 60000),
+    ]
+
+
+def test_find_sentence_gaps_ignores_short_pauses():
+    document = SentenceBoundaryDocument(
+        video_hash=VIDEO_HASH,
+        duration_ms=9000,
+        algorithm_version="test",
+        sentences=[
+            SentenceBoundary(index=0, start_ms=0, end_ms=4000, text="A"),
+            SentenceBoundary(index=1, start_ms=4500, end_ms=9000, text="B"),
+        ],
+    )
+
+    assert find_sentence_gaps(document) == []
+
+
+def test_plan_gap_chunks_splits_long_gap_at_silence():
+    samples = np.concatenate(
+        [_speech(14.0), _silence(0.6), _speech(16.0)]
+    ).astype(np.float32)
+    chunks = _plan_gap_chunks(samples, SAMPLE_RATE, int(samples.size / SAMPLE_RATE * 1000))
+
+    assert len(chunks) == 2
+    first_start, first_end = chunks[0]
+    second_start, second_end = chunks[1]
+    assert first_start == 0
+    assert 13_000 <= first_end <= 17_000
+    assert second_start == first_end
+    assert second_end == int(samples.size / SAMPLE_RATE * 1000)
+    assert all(end - start <= 20_000 for start, end in chunks)
+
+
+def test_repair_gaps_inserts_transcribed_sentences_and_keeps_existing(tmp_path, monkeypatch):
+    from dataclasses import replace
+
+    from sentence_api.config import Settings
+
+    settings = replace(Settings.from_env(), data_dir=tmp_path / "data")
+    settings.ensure_directories()
+    repository = VideoRepository(settings.database_path)
+    repository.upsert_upload(
+        video_hash=VIDEO_HASH,
+        title="Gap repair lesson",
+        filename="lesson.mp4",
+        stored_filename=f"{VIDEO_HASH}.mp4",
+        content_type="video/mp4",
+        size_bytes=5,
+        language="en",
+    )
+    (settings.videos_dir / f"{VIDEO_HASH}.mp4").write_bytes(b"fake-video")
+    repository.save_processing_result(
+        SentenceBoundaryDocument(
+            video_hash=VIDEO_HASH,
+            duration_ms=18000,
+            algorithm_version="test",
+            sentences=[
+                SentenceBoundary(
+                    index=0,
+                    start_ms=0,
+                    end_ms=2000,
+                    text="Hello.",
+                    language="en",
+                    reference_speech_duration_ms=1800,
+                ),
+                SentenceBoundary(
+                    index=1,
+                    start_ms=15000,
+                    end_ms=18000,
+                    text="Goodbye.",
+                    language="en",
+                    reference_speech_duration_ms=2500,
+                ),
+            ],
+        ),
+        "Hello. Goodbye.",
+    )
+    repository.save_sentence_translations(VIDEO_HASH, {0: "你好。", 1: "再见。"})
+
+    class GapTranscriber:
+        available = True
+        calls = 0
+
+        def transcribe(self, audio_path, language=None):
+            self.calls += 1
+            return Transcript(
+                text="The missing sentence.",
+                segments=[
+                    TranscriptionSegment(
+                        start_seconds=1.0,
+                        end_seconds=5.0,
+                        text="The missing sentence.",
+                    )
+                ],
+            )
+
+    def fake_extract(media_path, output_path, *, start_ms, duration_ms):
+        output_path.write_bytes(_wav_bytes(_speech(max(1.0, duration_ms / 1000)), SAMPLE_RATE))
+
+    monkeypatch.setattr("sentence_api.processing.extract_audio_segment", fake_extract)
+    transcriber = GapTranscriber()
+    result = VideoProcessor(settings, repository, transcriber).repair_gaps(VIDEO_HASH)
+
+    assert result["gap_count"] == 1
+    assert result["transcribed_gaps"] == 1
+    assert result["inserted_count"] == 1
+    assert transcriber.calls == 1
+    document = repository.get_document(VIDEO_HASH)
+    assert [sentence.text for sentence in document.sentences] == [
+        "Hello.",
+        "The missing sentence.",
+        "Goodbye.",
+    ]
+    assert document.sentences[0].zh_text == "你好。"
+    assert document.sentences[1].zh_text is None
+    assert document.sentences[2].zh_text == "再见。"
+    assert document.sentences[0].end_ms <= document.sentences[1].start_ms
+    assert document.sentences[1].end_ms <= document.sentences[2].start_ms
+    assert document.sentences[1].start_ms >= 2000
+    assert document.sentences[1].end_ms <= 15000
+
+
+def test_repair_gaps_skips_silent_gaps(tmp_path, monkeypatch):
+    from dataclasses import replace
+
+    from sentence_api.config import Settings
+
+    settings = replace(Settings.from_env(), data_dir=tmp_path / "data")
+    settings.ensure_directories()
+    repository = VideoRepository(settings.database_path)
+    repository.upsert_upload(
+        video_hash=VIDEO_HASH,
+        title="Silent gap lesson",
+        filename="lesson.mp4",
+        stored_filename=f"{VIDEO_HASH}.mp4",
+        content_type="video/mp4",
+        size_bytes=5,
+        language="en",
+    )
+    (settings.videos_dir / f"{VIDEO_HASH}.mp4").write_bytes(b"fake-video")
+    repository.save_processing_result(
+        SentenceBoundaryDocument(
+            video_hash=VIDEO_HASH,
+            duration_ms=18000,
+            algorithm_version="test",
+            sentences=[
+                SentenceBoundary(index=0, start_ms=0, end_ms=2000, text="Hello."),
+                SentenceBoundary(index=1, start_ms=15000, end_ms=18000, text="Goodbye."),
+            ],
+        ),
+        None,
+    )
+
+    class TrackingTranscriber:
+        available = True
+        calls = 0
+
+        def transcribe(self, audio_path, language=None):
+            self.calls += 1
+            return Transcript(text="hallucinated", segments=[])
+
+    def fake_extract(media_path, output_path, *, start_ms, duration_ms):
+        output_path.write_bytes(_wav_bytes(_silence(max(1.0, duration_ms / 1000)), SAMPLE_RATE))
+
+    monkeypatch.setattr("sentence_api.processing.extract_audio_segment", fake_extract)
+    transcriber = TrackingTranscriber()
+    result = VideoProcessor(settings, repository, transcriber).repair_gaps(VIDEO_HASH)
+
+    assert result["gap_count"] == 1
+    assert result["inserted_count"] == 0
+    assert transcriber.calls == 0
+    document = repository.get_document(VIDEO_HASH)
+    assert [sentence.text for sentence in document.sentences] == ["Hello.", "Goodbye."]

sentence_api/tests/test_transcription.py

@@ -423,6 +423,36 @@ def test_split_sentences_ignores_am_pm_abbreviations():
     ]
 
 
+def test_split_sentences_ignores_clock_times_with_pm_period():
+    compact = TranscriptionSegment(
+        start_seconds=0.0,
+        end_seconds=6.0,
+        text="The briefing is at 3:00 PM. Eastern officials will attend later.",
+    )
+    assert [s.text for s in split_sentences_at_punctuation([compact])] == [
+        "The briefing is at 3:00 PM. Eastern officials will attend later."
+    ]
+
+    spoken = TranscriptionSegment(
+        start_seconds=0.0,
+        end_seconds=6.0,
+        text="We start at 5 pm. Eastern time and then break.",
+    )
+    assert [s.text for s in split_sentences_at_punctuation([spoken])] == [
+        "We start at 5 pm. Eastern time and then break."
+    ]
+
+    morning = TranscriptionSegment(
+        start_seconds=0.0,
+        end_seconds=6.0,
+        text="It begins at 9:30 AM. Please arrive early.",
+    )
+    assert [s.text for s in split_sentences_at_punctuation([morning])] == [
+        "It begins at 9:30 AM.",
+        "Please arrive early.",
+    ]
+
+
 def test_split_sentences_merges_segment_ending_with_abbreviation():
     segments = [
         TranscriptionSegment(0.0, 5.0, "The meeting ends at 5 p.m."),
@@ -460,6 +490,77 @@ def test_split_sentences_ignores_common_abbreviations():
     ]
 
 
+def test_split_sentences_merges_same_speaker_line_wraps():
+    segments = [
+        TranscriptionSegment(
+            11.18,
+            15.45,
+            "[S01]Let's begin with some scheduling announcements for the remainder of President Trump's very",
+        ),
+        TranscriptionSegment(15.45, 16.29, "[S01]busy week."),
+        TranscriptionSegment(
+            16.29,
+            21.8,
+            "[S01]Today at 3:00 PM, President Trump will host a Black History Month reception in the East",
+        ),
+        TranscriptionSegment(21.8, 23.29, "[S01]Room of the White House."),
+    ]
+    sentences = split_sentences_at_punctuation(segments)
+    assert [s.text for s in sentences] == [
+        (
+            "Let's begin with some scheduling announcements for the remainder of "
+            "President Trump's very busy week."
+        ),
+        (
+            "Today at 3:00 PM, President Trump will host a Black History Month "
+            "reception in the East Room of the White House."
+        ),
+    ]
+    assert sentences[0].start_seconds == 11.18
+    assert sentences[0].end_seconds == 16.29
+    assert sentences[1].start_seconds == 16.29
+    assert sentences[1].end_seconds == 23.29
+
+
+def test_split_sentences_splits_when_speaker_changes():
+    segments = [
+        TranscriptionSegment(4.63, 5.93, "[S01]Hello, everybody."),
+        TranscriptionSegment(6.0, 6.53, "[S02]Hello."),
+        TranscriptionSegment(6.56, 7.83, "[S01]Good afternoon."),
+    ]
+    sentences = split_sentences_at_punctuation(segments)
+    assert [s.text for s in sentences] == [
+        "Hello, everybody.",
+        "Hello.",
+        "Good afternoon.",
+    ]
+
+
+def test_split_sentences_keeps_same_speaker_sentences_apart():
+    segments = [
+        TranscriptionSegment(6.56, 7.83, "[S01]Good afternoon."),
+        TranscriptionSegment(7.93, 9.86, "[S01]Thank you for your patience, as always."),
+    ]
+    sentences = split_sentences_at_punctuation(segments)
+    assert [s.text for s in sentences] == [
+        "Good afternoon.",
+        "Thank you for your patience, as always.",
+    ]
+
+
+def test_split_sentences_ignores_initials_and_version_numbers():
+    segment = TranscriptionSegment(
+        start_seconds=0.0,
+        end_seconds=8.0,
+        text="Donald J. Trump discussed version 1.9 today. Then the meeting ended.",
+    )
+    sentences = split_sentences_at_punctuation([segment])
+    assert [s.text for s in sentences] == [
+        "Donald J. Trump discussed version 1.9 today.",
+        "Then the meeting ended.",
+    ]
+
+
 def test_post_audio_requests_word_timestamps_and_falls_back(monkeypatch):
     calls = []
 

sentence_api/transcription.py

@@ -1,4 +1,5 @@
 import io
+import logging
 import mimetypes
 import re
 import wave
@@ -12,6 +13,9 @@ import numpy as np
 from .audio_metrics import AudioAnalysisError, decode_audio_mono, speech_frame_mask
 
 
+logger = logging.getLogger(__name__)
+
+
 # Whisper was trained on 30-second clips. Sending longer audio in a single request
 # makes the model enter a repetition loop (every later segment becomes the same
 # sentence while timestamps keep advancing), so long audio is split into short
@@ -93,13 +97,46 @@ class MossTranscriber:
         self, audio_path: Path, language: Optional[str]
     ) -> Transcript:
         content_type = mimetypes.guess_type(audio_path.name)[0] or "application/octet-stream"
+        logger.info(
+            "Transcription request: endpoint=%s model=%s audio=%s language=%s",
+            self.endpoint,
+            self.model,
+            audio_path.name,
+            language or "auto",
+        )
         with audio_path.open("rb") as audio_file:
             payload = self._post_audio(audio_file, audio_path.name, content_type, language)
         raw_text = str(payload.get("text") or "").strip()
+        raw_segments = payload.get("segments")
         segments = _parse_json_segments(payload.get("segments"))
+        logger.info(
+            "Transcription response: audio=%s raw_text_chars=%d raw_segments=%s "
+            "parsed_segments=%d text=%r",
+            audio_path.name,
+            len(raw_text),
+            len(raw_segments) if isinstance(raw_segments, list) else "missing",
+            len(segments),
+            _log_text_preview(raw_text),
+        )
         if not segments:
             segments = _parse_compact_segments(raw_text)
-        plain_text = " ".join(segment.text for segment in segments).strip() or raw_text
+            if segments:
+                logger.info(
+                    "Recovered %d compact-format segments from response text: audio=%s",
+                    len(segments),
+                    audio_path.name,
+                )
+        if not segments:
+            logger.warning(
+                "Transcription returned no usable timestamped segments: audio=%s "
+                "payload_keys=%s raw_segments=%s text=%r",
+                audio_path.name,
+                sorted(payload.keys()),
+                _log_segments_preview(raw_segments),
+                _log_text_preview(raw_text, 500),
+            )
+        plain_text = " ".join(segment.text for segment in segments).strip()
+        plain_text = _strip_speaker_tags(plain_text) or _strip_speaker_tags(raw_text)
         return Transcript(text=plain_text, segments=segments)
 
     def _transcribe_chunked(
@@ -116,11 +153,38 @@ class MossTranscriber:
                 "audio/wav",
                 language,
             )
+            raw_chunk_segments = payload.get("segments")
             chunk_segments = _parse_json_segments(
                 payload.get("segments"), compression_limit=COMPRESSION_RATIO_LIMIT
             )
+            logger.info(
+                "Chunk transcription response: range=%.3f-%.3fs raw_segments=%s "
+                "parsed_segments=%d text=%r",
+                start_seconds,
+                end_seconds,
+                len(raw_chunk_segments) if isinstance(raw_chunk_segments, list) else "missing",
+                len(chunk_segments),
+                _log_text_preview(str(payload.get("text") or "")),
+            )
             if not chunk_segments:
                 chunk_segments = _parse_compact_segments(str(payload.get("text") or "").strip())
+                if chunk_segments:
+                    logger.info(
+                        "Recovered %d compact-format segments from chunk text: range=%.3f-%.3fs",
+                        len(chunk_segments),
+                        start_seconds,
+                        end_seconds,
+                    )
+            if not chunk_segments:
+                logger.warning(
+                    "Chunk returned no usable timestamped segments: range=%.3f-%.3fs "
+                    "payload_keys=%s raw_segments=%s text=%r",
+                    start_seconds,
+                    end_seconds,
+                    sorted(payload.keys()),
+                    _log_segments_preview(raw_chunk_segments),
+                    _log_text_preview(str(payload.get("text") or ""), 500),
+                )
             for segment in chunk_segments:
                 words = None
                 if segment.words:
@@ -142,7 +206,9 @@ class MossTranscriber:
                     )
                 )
         segments.sort(key=lambda item: (item.start_seconds, item.end_seconds))
-        plain_text = " ".join(segment.text for segment in segments).strip()
+        plain_text = _strip_speaker_tags(
+            " ".join(segment.text for segment in segments).strip()
+        )
         return Transcript(text=plain_text, segments=segments)
 
     def _post_audio(
@@ -169,8 +235,20 @@ class MossTranscriber:
                 word_payload, audio_file, filename, content_type
             )
             if response.status_code in (400, 422):
+                logger.warning(
+                    "Transcription endpoint rejected word timestamps; retrying without them: "
+                    "status=%d body=%s",
+                    response.status_code,
+                    _log_text_preview(getattr(response, "text", ""), 1000),
+                )
                 audio_file.seek(0)
                 response = self._post_once(data, audio_file, filename, content_type)
+            if response.status_code >= 400:
+                logger.warning(
+                    "Transcription endpoint returned HTTP %d: body=%s",
+                    response.status_code,
+                    _log_text_preview(getattr(response, "text", ""), 2000),
+                )
             response.raise_for_status()
             return response.json()
         except (httpx.HTTPError, ValueError) as exc:
@@ -243,9 +321,76 @@ _PUNCTUATION_BOUNDARY = re.compile(r"[.?]+|,+")
 MIN_WORDS_FOR_COMMA_SPLIT = 50
 _ABBREVIATION_RE = re.compile(
     r"(?i)\b(?:a\.m\.|p\.m\.|mr\.|mrs\.|ms\.|dr\.|prof\.|st\.|vs\.|etc\.|"
-    r"e\.g\.|i\.e\.|no\.|approx\.|fig\.|inc\.|ltd\.|jr\.|sr\.|u\.s\.|u\.k\.)"
+    r"e\.g\.|i\.e\.|no\.|approx\.|fig\.|inc\.|ltd\.|jr\.|sr\.|u\.s\.|u\.k\.|"
+    r"ph\.d\.|m\.d\.|b\.a\.|m\.a\.|d\.c\.|p\.s\.)"
     r"(?![A-Za-z0-9])"
 )
+_TIME_MERIDIEM_RE = re.compile(
+    r"(?i)\b(?:[01]?\d|2[0-3])(?::[0-5]\d)?\s*(?:a\.m\.|p\.m\.|am\.|pm\.)"
+    r"(?![A-Za-z0-9])"
+)
+_TIME_CONTINUATION_RE = re.compile(
+    r"(?i)^(eastern|western|central|pacific|mountain|time|"
+    r"et|est|edt|pt|pst|pdt|ct|cst|cdt|mt|mst|mdt|gmt|utc|local)\b"
+)
+_SPEAKER_TAG_RE = re.compile(r"\[(S\d+)\]")
+
+
+def _strip_speaker_tags(text: str) -> str:
+    """Remove inline ``[S01]`` speaker tags from a text string."""
+    return _SPEAKER_TAG_RE.sub("", text).strip()
+
+
+def _expand_speaker_segments(
+    segment: TranscriptionSegment,
+) -> List[TranscriptionSegment]:
+    """Split a segment at inline ``[S01]`` speaker tags and strip the tags.
+
+    MOSS sometimes embeds speaker tags directly in the text (rather than in
+    the ``speaker`` JSON field). Each tag marks the start of a new speaker
+    turn, so the text is split there and the tags are removed from the
+    visible transcription.
+    """
+    text = segment.text
+    matches = list(_SPEAKER_TAG_RE.finditer(text))
+    if not matches:
+        return [segment]
+
+    sub_segments: List[TranscriptionSegment] = []
+    total = len(matches)
+    duration = segment.end_seconds - segment.start_seconds
+    for index, match in enumerate(matches):
+        start_pos = match.end()
+        if index + 1 < total:
+            end_pos = matches[index + 1].start()
+        else:
+            end_pos = len(text)
+        piece = text[start_pos:end_pos].strip()
+        if not piece:
+            continue
+        # Allocate time proportionally across sub-segments.
+        frac_start = index / total
+        frac_end = (index + 1) / total
+        sub_start = segment.start_seconds + duration * frac_start
+        sub_end = segment.start_seconds + duration * frac_end
+        words = None
+        if segment.words:
+            sub_words = [
+                w
+                for w in segment.words
+                if w.start_seconds >= sub_start and w.end_seconds <= sub_end
+            ]
+            words = sub_words or None
+        sub_segments.append(
+            TranscriptionSegment(
+                start_seconds=sub_start,
+                end_seconds=sub_end,
+                text=piece,
+                speaker=match.group(1),
+                words=words,
+            )
+        )
+    return sub_segments or [segment]
 
 
 def split_sentences_at_punctuation(
@@ -257,14 +402,20 @@ def split_sentences_at_punctuation(
     across segments and a sentence is closed at periods/question marks. Commas
     are also eligible boundaries, but only when the text before the comma has
     enough words; otherwise the comma stays with the following text.
+
+    Inline ``[S01]`` speaker tags are stripped. A new sentence starts when the
+    speaker actually changes; same-speaker line wraps stay merged until
+    sentence punctuation. Decimal points, initials, and common abbreviations
+    are not treated as sentence ends.
     """
     sentences: List[TranscriptionSegment] = []
     pieces: List[str] = []
     start_seconds: Optional[float] = None
     last_segment_end: Optional[float] = None
+    current_speaker: Optional[str] = None
 
     def flush(end_seconds: float) -> None:
-        nonlocal pieces, start_seconds
+        nonlocal pieces, start_seconds, current_speaker
         text = " ".join(piece for piece in pieces if piece).strip()
         if text and start_seconds is not None and end_seconds > start_seconds:
             sentences.append(
@@ -272,19 +423,35 @@ def split_sentences_at_punctuation(
                     start_seconds=start_seconds,
                     end_seconds=end_seconds,
                     text=text,
+                    speaker=current_speaker,
                 )
             )
         pieces = []
         start_seconds = None
+        current_speaker = None
+
+    expanded: List[TranscriptionSegment] = []
+    for seg in segments:
+        expanded.extend(_expand_speaker_segments(seg))
 
     for segment in sorted(
-        segments, key=lambda item: (item.start_seconds, item.end_seconds)
+        expanded, key=lambda item: (item.start_seconds, item.end_seconds)
     ):
         text = segment.text.strip()
         if not text:
             continue
+        # MOSS often tags every phrase/line with [S01] even when the speaker
+        # did not change. Only a real speaker change starts a new sentence.
+        if (
+            pieces
+            and segment.speaker is not None
+            and current_speaker is not None
+            and segment.speaker != current_speaker
+        ):
+            flush(segment.start_seconds)
         if not pieces:
             start_seconds = segment.start_seconds
+        current_speaker = segment.speaker
         seg_pieces: List[Tuple[str, Optional[str]]] = []
         cursor = 0
         for match in _PUNCTUATION_BOUNDARY.finditer(text):
@@ -322,6 +489,7 @@ def split_sentences_at_punctuation(
                 end_seconds = ends[index] if index < len(ends) else segment.end_seconds
                 flush(end_seconds)
                 start_seconds = end_seconds
+                current_speaker = segment.speaker
     if pieces:
         flush(last_segment_end if last_segment_end is not None else start_seconds)
     return sentences
@@ -340,31 +508,70 @@ def _is_sentence_end_run(text: str, run_start: int, run_end: int) -> bool:
     punctuation = text[run_start:run_end]
     if "?" in punctuation:
         return True
-    if (
-        run_start > 0
-        and run_end < len(text)
-        and text[run_start - 1].isdigit()
-        and text[run_end].isdigit()
-    ):
-        return False  # decimal point, e.g. "0.4"
+    if _is_decimal_point(text, run_start, run_end):
+        return False
     if len(punctuation) > 1:
         return True  # ellipsis or "!?" etc.
     if _is_abbreviation_run(text, run_start, run_end):
         return False  # e.g. "a.m.", "p.m.", "Mr."
+    if _is_time_meridiem_run(text, run_start, run_end):
+        return False  # e.g. "3:00 PM.", "5 pm."
+    if _is_single_letter_initial(text, run_start, run_end):
+        return False  # e.g. "Donald J. Trump"
     rest = text[run_end:].lstrip()
     if rest and rest[0].islower():
         return False  # period followed by a lowercase word is an abbreviation
     return True
 
 
+def _is_decimal_point(text: str, run_start: int, run_end: int) -> bool:
+    if run_end - run_start != 1 or text[run_start] != ".":
+        return False
+    if run_start == 0 or run_end >= len(text):
+        return False
+    return text[run_start - 1].isdigit() and text[run_end].isdigit()
+
+
+def _is_single_letter_initial(text: str, run_start: int, run_end: int) -> bool:
+    if run_end - run_start != 1 or text[run_start] != ".":
+        return False
+    if run_start == 0 or not text[run_start - 1].isalpha():
+        return False
+    return run_start == 1 or not text[run_start - 2].isalpha()
+
+
+def _meridiem_continues(text: str, after: int) -> bool:
+    rest = text[after:].lstrip()
+    if not rest:
+        return True
+    if rest[0].islower():
+        return True
+    return _TIME_CONTINUATION_RE.match(rest) is not None
+
+
+def _is_time_meridiem_run(text: str, run_start: int, run_end: int) -> bool:
+    for match in _TIME_MERIDIEM_RE.finditer(text):
+        if not (match.start() < run_end <= match.end()):
+            continue
+        if run_end < match.end():
+            return True
+        return _meridiem_continues(text, match.end())
+    return False
+
+
 def _is_abbreviation_run(text: str, run_start: int, run_end: int) -> bool:
     for match in _ABBREVIATION_RE.finditer(text):
         if not (match.start() < run_end <= match.end()):
             continue
-        if match.group(0).lower() == "no.":
+        token = match.group(0).lower()
+        if token == "no.":
             rest = text[match.end():].lstrip()
             if not rest or not rest[0].isdigit():
                 return False  # "No." as a reply ends the sentence
+        if token in {"a.m.", "p.m."}:
+            if run_end < match.end():
+                return True
+            return _meridiem_continues(text, match.end())
         return True
     return False
 
@@ -424,6 +631,30 @@ def _wav_duration_seconds(path: Path) -> Optional[float]:
     return None
 
 
+def _log_text_preview(text: str, limit: int = 240) -> str:
+    text = " ".join(text.split())
+    return text if len(text) <= limit else f"{text[:limit]}... [{len(text)} chars]"
+
+
+def _log_segments_preview(raw_segments, limit: int = 5):
+    if not isinstance(raw_segments, list):
+        return raw_segments
+    preview = []
+    for item in raw_segments[:limit]:
+        if not isinstance(item, dict):
+            preview.append(item)
+            continue
+        preview.append(
+            {
+                "start": item.get("start"),
+                "end": item.get("end"),
+                "compression_ratio": item.get("compression_ratio"),
+                "text": _log_text_preview(str(item.get("text") or "")),
+            }
+        )
+    return {"count": len(raw_segments), "items": preview}
+
+
 def _wav_bytes(samples: np.ndarray, sample_rate: int) -> bytes:
     pcm = np.clip(np.round(samples * 32767), -32768, 32767).astype("<i2")
     buffer = io.BytesIO()

sentence_api/translation_jobs.py

@@ -0,0 +1,306 @@
+"""Background subtitle translation so long videos do not block HTTP."""
+
+from __future__ import annotations
+
+import logging
+import threading
+from dataclasses import dataclass
+
+from .ai_tutor import AITutor, AITutorError, translation_needs_repair
+from .models import (
+    AdminSubtitleTranslationResponse,
+    SentenceBoundaryDocument,
+    SubtitleTranslationItem,
+    SubtitleTranslationResponse,
+    TranslationTerm,
+)
+from .repository import VideoRepository
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class _Job:
+    video_hash: str
+    done: threading.Event
+    lock: threading.Lock
+    running: bool = True
+    cancelled: bool = False
+    error: str | None = None
+    thread: threading.Thread | None = None
+
+
+class SubtitleTranslationService:
+    def __init__(
+        self,
+        repository: VideoRepository,
+        ai_tutor: AITutor,
+        *,
+        post_wait_seconds: float = 8.0,
+    ) -> None:
+        self.repository = repository
+        self._ai_tutor = ai_tutor
+        self.post_wait_seconds = post_wait_seconds
+        self._lock = threading.Lock()
+        self._jobs: dict[str, _Job] = {}
+
+    @property
+    def ai_tutor(self) -> AITutor:
+        value = self._ai_tutor
+        return value() if callable(value) else value
+
+    @ai_tutor.setter
+    def ai_tutor(self, value) -> None:
+        self._ai_tutor = value
+
+    def snapshot(self, document: SentenceBoundaryDocument) -> SubtitleTranslationResponse:
+        translations = [
+            SubtitleTranslationItem(index=sentence.index, zh_text=sentence.zh_text)
+            for sentence in document.sentences
+            if sentence.zh_text and sentence.zh_text.strip()
+        ]
+        needed = [
+            sentence
+            for sentence in document.sentences
+            if (sentence.text or "").strip()
+        ]
+        complete = bool(needed) and all((sentence.zh_text or "").strip() for sentence in needed)
+        job = self._jobs.get(document.video_hash)
+        running = False
+        error: str | None = None
+        if job is not None:
+            with job.lock:
+                running = job.running
+                error = job.error
+        if complete:
+            status = "ready"
+            error = None
+        elif running:
+            status = "running"
+        elif error:
+            status = "error"
+        else:
+            status = "idle"
+        return SubtitleTranslationResponse(
+            video_hash=document.video_hash,
+            translations=translations,
+            complete=complete,
+            status=status,
+            error=error,
+        )
+
+    def current(self, video_hash: str) -> SubtitleTranslationResponse | None:
+        document = self.repository.get_document(video_hash)
+        if document is None:
+            return None
+        return self.snapshot(document)
+
+    def invalidate(self, video_hash: str) -> None:
+        """Stop an in-flight job so stale translations cannot be written back."""
+        with self._lock:
+            job = self._jobs.pop(video_hash, None)
+        if job is None:
+            return
+        with job.lock:
+            job.cancelled = True
+
+    def refresh(self, video_hash: str) -> None:
+        """Drop saved Chinese subtitles and translate the current sentences again."""
+        self.invalidate(video_hash)
+        self.repository.clear_sentence_translations(video_hash)
+        if not getattr(self.ai_tutor, "available", False):
+            return
+        try:
+            self._start_if_needed(video_hash)
+        except (KeyError, ValueError, AITutorError):
+            logger.exception("Could not restart subtitle translation: video=%s", video_hash)
+
+    def start_missing(self, video_hash: str) -> None:
+        """Translate sentences that still lack Chinese text, keeping existing ones."""
+        if not getattr(self.ai_tutor, "available", False):
+            return
+        document = self.repository.get_document(video_hash)
+        if document is None:
+            return
+        if self.snapshot(document).complete:
+            return
+        try:
+            self._start_if_needed(video_hash)
+        except (KeyError, ValueError, AITutorError):
+            logger.exception(
+                "Could not start subtitle translation after gap repair: video=%s",
+                video_hash,
+            )
+
+    def repair_missing(self, video_hash: str) -> AdminSubtitleTranslationResponse:
+        """Re-translate missing or still-English sentences, keeping good Chinese."""
+        if self.repository.get_video(video_hash) is None:
+            raise KeyError(video_hash)
+        document = self.repository.get_document(video_hash)
+        if document is None:
+            raise KeyError(video_hash)
+        missing = missing_translation_indexes(document)
+        if missing:
+            self.invalidate(video_hash)
+            self.repository.clear_sentence_translations(video_hash, missing)
+            snapshot = self.ensure(video_hash)
+        else:
+            snapshot = self.snapshot(document)
+        result = self._with_terms(snapshot)
+        return result.model_copy(update={"attempted_count": len(missing)})
+
+    def admin_snapshot(self, video_hash: str) -> AdminSubtitleTranslationResponse | None:
+        current = self.current(video_hash)
+        if current is None:
+            return None
+        return self._with_terms(current)
+
+    def retranslate(
+        self,
+        video_hash: str,
+        terms: list[tuple[str, str]] | None = None,
+    ) -> AdminSubtitleTranslationResponse:
+        if self.repository.get_video(video_hash) is None:
+            raise KeyError(video_hash)
+        if terms is not None:
+            self.repository.save_translation_glossary(video_hash, terms)
+        self.invalidate(video_hash)
+        self.repository.clear_sentence_translations(video_hash)
+        return self._with_terms(self.ensure(video_hash))
+
+    def _with_terms(self, snapshot: SubtitleTranslationResponse) -> AdminSubtitleTranslationResponse:
+        terms = [
+            TranslationTerm(source=source, zh=zh)
+            for source, zh in self.repository.get_translation_glossary(snapshot.video_hash)
+        ]
+        document = self.repository.get_document(snapshot.video_hash)
+        missing_count = 0 if document is None else len(missing_translation_indexes(document))
+        payload = snapshot.model_dump()
+        payload["terms"] = terms
+        payload["missing_count"] = missing_count
+        payload["attempted_count"] = 0
+        return AdminSubtitleTranslationResponse(**payload)
+
+    def ensure_from_local(self, document: SentenceBoundaryDocument) -> SubtitleTranslationResponse:
+        """Accept locally generated English subtitles, then translate like a course."""
+        self.repository.import_local_subtitle_document(document)
+        return self.ensure(document.video_hash)
+
+    def ensure(self, video_hash: str) -> SubtitleTranslationResponse:
+        if not getattr(self.ai_tutor, "available", False):
+            raise AITutorError("AI translation is not configured.")
+        document = self.repository.get_document(video_hash)
+        if document is None:
+            raise KeyError(video_hash)
+        current = self.snapshot(document)
+        if current.complete:
+            return current
+        payload = [
+            (sentence.index, sentence.text or "")
+            for sentence in document.sentences
+            if (sentence.text or "").strip()
+        ]
+        if not payload:
+            raise ValueError("No sentence text is available to translate.")
+        job = self._start_if_needed(video_hash)
+        wait_seconds = max(0.0, float(self.post_wait_seconds))
+        if wait_seconds > 0:
+            job.done.wait(timeout=wait_seconds)
+        latest = self.current(video_hash)
+        if latest is None:
+            raise KeyError(video_hash)
+        return latest
+
+    def _start_if_needed(self, video_hash: str) -> _Job:
+        with self._lock:
+            existing = self._jobs.get(video_hash)
+            if existing is not None:
+                with existing.lock:
+                    still_running = (
+                        not existing.cancelled
+                        and (
+                            existing.running
+                            or (
+                                existing.thread is not None
+                                and existing.thread.is_alive()
+                            )
+                        )
+                    )
+                if still_running:
+                    return existing
+            job = _Job(
+                video_hash=video_hash,
+                done=threading.Event(),
+                lock=threading.Lock(),
+            )
+            thread = threading.Thread(
+                target=self._run_job,
+                args=(job,),
+                name=f"subtitle-translate-{video_hash[:8]}",
+                daemon=True,
+            )
+            job.thread = thread
+            self._jobs[video_hash] = job
+            thread.start()
+            return job
+
+    def _persist(self, job: _Job, translations: dict[int, str]) -> None:
+        with job.lock:
+            if job.cancelled:
+                return
+            self.repository.save_sentence_translations(job.video_hash, translations)
+
+    def _run_job(self, job: _Job) -> None:
+        try:
+            with job.lock:
+                if job.cancelled:
+                    return
+            document = self.repository.get_document(job.video_hash)
+            if document is None:
+                raise AITutorError("Course is not available.")
+            payload = [
+                (sentence.index, sentence.text or "")
+                for sentence in document.sentences
+                if (sentence.text or "").strip()
+            ]
+            if not payload:
+                return
+            existing = {
+                sentence.index: sentence.zh_text
+                for sentence in document.sentences
+                if sentence.zh_text and sentence.zh_text.strip()
+            }
+            video = self.repository.get_video(job.video_hash)
+            translated = self.ai_tutor.translate_sentences(
+                payload,
+                course_title=(video["title"] if video is not None else "") or "",
+                existing=existing,
+                persist=lambda part: self._persist(job, part),
+                terms=self.repository.get_translation_glossary(job.video_hash),
+            )
+            if translated:
+                self._persist(job, translated)
+            with job.lock:
+                if not job.cancelled:
+                    job.error = None
+        except Exception as exc:  # noqa: BLE001 - surface job failure to clients
+            with job.lock:
+                cancelled = job.cancelled
+            if cancelled:
+                return
+            logger.exception("Subtitle translation job failed: video=%s", job.video_hash)
+            with job.lock:
+                job.error = str(exc)
+        finally:
+            with job.lock:
+                job.running = False
+            job.done.set()
+
+
+def missing_translation_indexes(document: SentenceBoundaryDocument) -> list[int]:
+    return [
+        sentence.index
+        for sentence in document.sentences
+        if translation_needs_repair(sentence.text, sentence.zh_text)
+    ]

whisper/.env.example

@@ -15,3 +15,7 @@ WHISPER_TTL=-1
 
 # 模型下载地址：国内默认走 hf-mirror.com 镜像；能直连 Hugging Face 时改为空值（HF_ENDPOINT=）
 HF_ENDPOINT=https://hf-mirror.com
+
+# 国内拉取 ghcr 较慢时走 DaoCloud；能直连 GitHub 时改为：
+# WHISPER_IMAGE=ghcr.io/speaches-ai/speaches:latest-cuda
+WHISPER_IMAGE=docker.m.daocloud.io/ghcr.io/speaches-ai/speaches:latest-cuda

whisper/README.md

@@ -9,7 +9,7 @@ OpenAI 兼容语音转写服务，使用 faster-whisper（CTranslate2）推理
 
 ```text
 whisper/
-  docker-compose.yml           GPU 服务编排（ghcr.io/speaches-ai/speaches:latest-cuda）
+  docker-compose.yml           GPU 服务编排（默认经 DaoCloud 拉 ghcr 镜像）
   .env.example                 配置模板（模型、端口、计算类型）
   systemd/whisper-transcribe.service  systemd 模板
   verify.sh                    部署验证脚本
@@ -27,7 +27,7 @@ sudo chown "$USER":"$USER" /opt/whisper
 cp -a whisper/. /opt/whisper/   # 注意用 whisper/. 而不是 whisper/*，否则不会复制 .env.example 等隐藏文件
 cd /opt/whisper
 
-cp .env.example .env          # 按需修改 WHISPER_MODEL
+cp .env.example .env          # 按需修改 WHISPER_MODEL / WHISPER_IMAGE
 docker compose up -d
 docker compose logs -f whisper   # 首次启动下载模型（约 3 GB），看到启动完成即可
 ```
@@ -79,9 +79,11 @@ L4（24 GB 显存）可同时常驻两个模型；Speaches 支持在请求的 `m
 
 ## 常见问题
 
+- 拉取 Speaches 镜像慢：默认镜像为 `docker.m.daocloud.io/ghcr.io/speaches-ai/speaches:latest-cuda`；
+  能直连 GitHub 时在 `.env` 设 `WHISPER_IMAGE=ghcr.io/speaches-ai/speaches:latest-cuda`。
 - 首次启动下载模型慢或失败：默认已使用国内镜像 `https://hf-mirror.com`（可在 `.env`
   的 `HF_ENDPOINT` 覆盖）；仍失败可改用代理，在 `docker-compose.yml` 的
   `environment` 里加 `HTTP_PROXY`/`HTTPS_PROXY` 后 `docker compose up -d`。
 - 容器报 CUDA 错误：确认宿主机驱动 ≥ 535，并已安装 NVIDIA Container Toolkit。
 - 转写结果没有说话人：Whisper 不做说话人分离，需要该功能请改用
-  `MOSS-Transcribe-Diarize`（见 `sentence_api/DEPLOYMENT.md`）。
+  `MOSS-Transcribe-Diarize`（见 [`moss/README.md`](../moss/README.md)）。

whisper/docker-compose.yml

@@ -1,6 +1,6 @@
 services:
   whisper:
-    image: ghcr.io/speaches-ai/speaches:latest-cuda
+    image: ${WHISPER_IMAGE:-docker.m.daocloud.io/ghcr.io/speaches-ai/speaches:latest-cuda}
     container_name: whisper-transcribe
     restart: unless-stopped
     ports: