fix(audio): detect transcript-wide hallucinations

frontend/src/App.tsx

@@ -40,7 +40,7 @@ type Result = {
 };
 type SearchCoverage = {available:boolean;model_version?:string;total:number;indexed:number;searchable:number;empty:number;queued:number;running:number;failed:number;low_quality?:number;pending:number;percent:number;current_model_percent:number};
 type TranscriptItem = {id:string;start_ms:number;end_ms:number;raw_text:string;language?:string;quality_score?:number};
-type TranscriptPage = {status:"ready"|"filtered"|"low_quality"|"empty"|"queued"|"running"|"failed"|"unavailable"|"pending";model_version?:string;indexed_version?:string;detected_language?:string;quality_score?:number;quality_state:"ready"|"filtered"|"low_quality"|"empty"|"legacy";rejected_segments:number;quality_flags:string[];items:TranscriptItem[];page:number;page_size:number;total:number;pages:number;job?:Pick<Job,"id"|"status"|"progress"|"message"|"error">};
+type TranscriptPage = {status:"ready"|"filtered"|"low_quality"|"empty"|"queued"|"running"|"failed"|"unavailable"|"pending";model_version?:string;indexed_version?:string;detected_language?:string;quality_score?:number;quality_state:"ready"|"filtered"|"low_quality"|"empty"|"legacy";rejected_segments:number;quality_flags:string[];aggregate_risk?:boolean;repeated_phrases?:{text:string;count:number}[];quality_score_semantics?:"rule_check";items:TranscriptItem[];page:number;page_size:number;total:number;pages:number;job?:Pick<Job,"id"|"status"|"progress"|"message"|"error">};
 type SpeechConfig = {language_policy:"zh_priority"|"auto"|"zh";quality_profile:"speed"|"balanced"|"accuracy";model_variant:"small"|"medium";index_revision:number};
 type RecognitionType = "visual"|"ocr"|"person"|"audio"|"subtitle"|"metadata";
 type UploadTask = {
@@ -628,7 +628,7 @@ function PlayerPage({ video, startMs, videos, collections, onBack, onPlay, onEdi
       </section>
       <section className={`transcript-panel ${transcriptOpen?"open":""}`}>
         <button className="transcript-toggle" aria-expanded={transcriptOpen} onClick={()=>setTranscriptOpen(value=>!value)}><span><Volume2/><span><strong>音频转写</strong><small>{transcriptLoading?"正在读取识别状态…":transcript?.status==="ready"?`${transcript.total} 个高可信时间片段`:transcript?.status==="filtered"?`${transcript.total} 个可搜索片段，部分内容已过滤`:transcript?.status==="low_quality"?"识别质量偏低，系统将进行一次增强修复":transcript?.status==="empty"?"已识别，未检测到语音":transcript?.status==="running"?`正在识别 ${Math.round((transcript.job?.progress||0)*100)}%`:transcript?.status==="queued"?"等待后台识别":transcript?.status==="failed"?"识别失败":transcript?.status==="unavailable"?"音频模型未安装":"尚未识别"}</small></span></span>{transcriptLoading?<RefreshCw className="spin"/>:transcriptOpen?<ArrowUp/>:<ArrowDown/>}</button>
-        {transcriptOpen&&<div className="transcript-body">{transcriptLoading?<AsyncNotice compact label="正在读取音频转写…"/>:<>{transcript&&<div className={`transcript-quality ${transcript.quality_state}`} role="status"><span><strong>{transcript.detected_language?`识别语言：${({zh:"中文",en:"英语",ja:"日语",ko:"韩语"} as Record<string,string>)[transcript.detected_language]||transcript.detected_language}`:"识别语言：自动"}</strong><small>{transcript.quality_state==="low_quality"?`整体质量偏低，已保留可信片段并仅自动增强一次`:transcript.quality_state==="filtered"?`已过滤 ${transcript.rejected_segments} 个不可信片段`:transcript.quality_state==="empty"?"有界语音检测未发现可识别内容":transcript.quality_state==="legacy"?"旧版转写，建议重新识别":"质量检查已通过"}{typeof transcript.quality_score==="number"?` · 可信度 ${Math.round(transcript.quality_score*100)}%`:""}</small></span>{["filtered","low_quality","legacy"].includes(transcript.quality_state)?<AlertCircle/>:<ShieldCheck/>}</div>}{transcript?.items.length?<><div className="transcript-list">{transcript.items.map(item=><button key={item.id} onClick={()=>{seekTo(item.start_ms/1000);void element.current?.play().catch(()=>{})}}><time>{formatTime(item.start_ms)}</time><span>{item.raw_text}</span><Play/></button>)}</div>{transcript.pages>1&&<div className="transcript-pages"><button disabled={transcript.page<=1} onClick={()=>void loadTranscript(transcript.page-1)}><ChevronLeft/>上一页</button><span>{transcript.page} / {transcript.pages}</span><button disabled={transcript.page>=transcript.pages} onClick={()=>void loadTranscript(transcript.page+1)}>下一页<ChevronRight/></button></div>}</>:<div className="transcript-empty"><VolumeX/><p>{transcript?.job?.error||transcript?.job?.message||"当前视频还没有可显示的音频转写。"}</p></div>}</>}{transcriptError&&<div className="error"><AlertCircle/>{transcriptError}</div>}<button className="secondary transcript-reindex" disabled={transcriptBusy||transcript?.status==="running"||transcript?.status==="queued"} onClick={()=>setTranscriptLanguageOpen(true)}>{transcriptBusy?<RefreshCw className="spin"/>:<RefreshCw/>}{transcript?.status==="failed"?"重试识别":"重新识别"}</button></div>}
+        {transcriptOpen&&<div className="transcript-body">{transcriptLoading?<AsyncNotice compact label="正在读取音频转写…"/>:<>{transcript&&<div className={`transcript-quality ${transcript.quality_state}`} role="status"><span><strong>{transcript.detected_language?`识别语言：${({zh:"中文",en:"英语",ja:"日语",ko:"韩语"} as Record<string,string>)[transcript.detected_language]||transcript.detected_language}`:"识别语言：自动"}</strong><small>{transcript.quality_state==="low_quality"?`检测到整片重复或低质量风险，结果仅供诊断且不进入搜索`:transcript.quality_state==="filtered"?`已过滤 ${transcript.rejected_segments} 个不可信片段`:transcript.quality_state==="empty"?"有界语音检测未发现可识别内容":transcript.quality_state==="legacy"?"旧版转写，建议重新识别":"规则检查未发现明显异常"}{typeof transcript.quality_score==="number"?` · 规则检查分 ${Math.round(transcript.quality_score*100)}%`:""}{transcript.repeated_phrases?.length?<>{" · 重复："+transcript.repeated_phrases.slice(0,3).map(item=>item.text+"×"+item.count).join("、")}</>:null}</small></span>{["filtered","low_quality","legacy"].includes(transcript.quality_state)?<AlertCircle/>:<ShieldCheck/>}</div>}{transcript?.items.length?<><div className="transcript-list">{transcript.items.map(item=><button key={item.id} onClick={()=>{seekTo(item.start_ms/1000);void element.current?.play().catch(()=>{})}}><time>{formatTime(item.start_ms)}</time><span>{item.raw_text}</span><Play/></button>)}</div>{transcript.pages>1&&<div className="transcript-pages"><button disabled={transcript.page<=1} onClick={()=>void loadTranscript(transcript.page-1)}><ChevronLeft/>上一页</button><span>{transcript.page} / {transcript.pages}</span><button disabled={transcript.page>=transcript.pages} onClick={()=>void loadTranscript(transcript.page+1)}>下一页<ChevronRight/></button></div>}</>:<div className="transcript-empty"><VolumeX/><p>{transcript?.job?.error||transcript?.job?.message||"当前视频还没有可显示的音频转写。"}</p></div>}</>}{transcriptError&&<div className="error"><AlertCircle/>{transcriptError}</div>}<button className="secondary transcript-reindex" disabled={transcriptBusy||transcript?.status==="running"||transcript?.status==="queued"} onClick={()=>setTranscriptLanguageOpen(true)}>{transcriptBusy?<RefreshCw className="spin"/>:<RefreshCw/>}{transcript?.status==="failed"?"重试识别":"重新识别"}</button></div>}
       </section>
     </div>
     <aside className={`related player-side-panel ${collectionDetail?"player-collection-directory":""}`}>
@@ -1359,7 +1359,7 @@ function SettingsPage({ preferences, sources, onPreferences }: {preferences:Pref
         {speechLoad.phase==="loading"||speechLoad.phase==="idle"?<AsyncNotice compact label="正在读取语音识别策略…"/>:speechLoad.phase==="error"&&!speechConfig?<div className="error"><AlertCircle/>{speechLoad.error}<button onClick={()=>void loadSpeech()}>重试</button></div>:speechConfig&&<>
           <fieldset><legend>识别模型</legend><div className="audio-model-variants">{(["small","medium"] as const).map(variant=>{const info=models.audio_variants?.[variant]||{};const state=models.installations?.[`audio:${variant}`] as ModelInstallation|undefined;const active=state?.status==="queued"||state?.status==="running";const progress=Math.round(Math.max(0,Math.min(1,state?.progress||0))*100);const selected=speechConfig.model_variant===variant;return <article className={`${selected?"selected":""} ${info.installed?"installed":"missing"}`} key={variant}><button type="button" className="audio-model-choice" disabled={!info.installed||active||(variant==="medium"&&!gpuId)} aria-pressed={selected} onClick={()=>setSpeechConfig({...speechConfig,model_variant:variant})}><span>{selected?<CheckCircle2/>:<Boxes/>}</span><div><strong>Whisper {variant==="medium"?"Medium INT8":"Small"}</strong><small>{variant==="medium"?"高质量 · 仅 Intel GPU · 约 0.78 GB":"标准模型 · 约 0.49 GB · 纯 CPU 设备可用"}</small></div></button>{active?<div className="audio-model-progress" role="status"><span>{state?.message||"正在下载安装"}</span><b>{state?.status==="queued"?"等待":`${progress}%`}</b><i><span style={{width:`${state?.status==="queued"?4:progress}%`}}/></i></div>:info.needs_update?<button className="secondary audio-model-install" disabled={modelBusy||Boolean(modelJob)||(variant==="medium"&&!gpuId)} onClick={()=>void installOnline("audio",variant)}><RefreshCw/>更新为 INT8 · {formatSize(info.size_bytes||0)}</button>:info.installed?<div className="audio-model-actions"><span>{selected?"当前选择":formatSize(info.size_bytes||0)}</span>{!selected&&<button className="danger-text" disabled={modelBusy} onClick={()=>void removeAudioVariant(variant)}><Trash2/>删除</button>}</div>:<button className="secondary audio-model-install" disabled={modelBusy||Boolean(modelJob)||(variant==="medium"&&!gpuId)} onClick={()=>void installOnline("audio",variant)}><Download/>{variant==="medium"&&!gpuId?"需要 Intel GPU":"下载安装"}</button>}</article>})}</div></fieldset>
           <fieldset><legend>语言检测</legend><div className="speech-policy-options">{([['zh_priority','中文优先','音轨标签只作提示；弱检测或冲突时回到中文'],['auto','智能检测','由多个有声片段多数决，音轨标签仅用于平票'],['zh','固定中文','所有任务均明确按中文解码']] as const).map(([key,label,description])=><button type="button" className={speechConfig.language_policy===key?"active":""} aria-pressed={speechConfig.language_policy===key} key={key} onClick={()=>setSpeechConfig({...speechConfig,language_policy:key})}><strong>{label}</strong><small>{description}</small>{speechConfig.language_policy===key&&<CheckCircle2/>}</button>)}</div></fieldset>
-          <fieldset><legend>质量档位</legend><div className="speech-policy-options quality">{([['speed','速度优先','单 beam 解码，适合快速建立索引'],['balanced','均衡','提高解码候选数，兼顾速度与准确率'],['accuracy','准确率优先','使用更多 beam，处理时间更长']] as const).map(([key,label,description])=><button type="button" className={speechConfig.quality_profile===key?"active":""} aria-pressed={speechConfig.quality_profile===key} key={key} onClick={()=>setSpeechConfig({...speechConfig,quality_profile:key})}><strong>{label}</strong><small>{description}</small>{speechConfig.quality_profile===key&&<CheckCircle2/>}</button>)}</div></fieldset>
+          <fieldset><legend>质量档位</legend><div className="speech-policy-options quality">{([['speed','速度优先','单 beam 解码，适合快速建立索引'],['balanced','均衡','CPU 会增加候选；Intel GPU 仍为单 beam'],['accuracy','准确率优先','CPU 使用 5 beam；Intel GPU 受限为单 beam']] as const).map(([key,label,description])=><button type="button" className={speechConfig.quality_profile===key?"active":""} aria-pressed={speechConfig.quality_profile===key} key={key} onClick={()=>setSpeechConfig({...speechConfig,quality_profile:key})}><strong>{label}</strong><small>{description}</small>{speechConfig.quality_profile===key&&<CheckCircle2/>}</button>)}</div></fieldset>
           <p className="speech-resource-note"><ShieldCheck/>Medium 在 GPU、内存或推理验证失败时会停止任务，不会回退 CPU；旧转写会保留到新结果成功写入。</p>
           {speechLoad.error&&<div className="error"><AlertCircle/>{speechLoad.error}<button onClick={()=>void loadSpeech()}>重新读取</button></div>}
           <div className="speech-policy-actions"><button className="primary" disabled={speechBusy} onClick={()=>void saveSpeechConfig()}>{speechBusy?<RefreshCw className="spin"/>:<CheckCircle2/>}保存识别策略</button><button className="secondary" disabled={speechBusy} onClick={()=>void reconcileSpeech()}><ShieldCheck/>重建旧版音频索引</button></div>{speechMessage&&<p className="status-line"><Activity/>{speechMessage}</p>}