// ==UserScript==
// @name         小鹅通 · 实时字幕（多引擎）
// @namespace    workbuddy/xiaoe-live-caption
// @version      1.1.0
// @description  为小鹅通网页视频叠加实时字幕。支持浏览器语音识别、云端 ASR（从视频音轨取声）、以及已有 SRT/VTT 字幕文件；可导出字幕。
// @author       WorkBuddy
// @match        https://*.xiaoe-tech.com/*
// @match        http://*.xiaoe-tech.com/*
// @match        https://*.xiaoeknow.com/*
// @match        http://*.xiaoeknow.com/*
// @run-at       document-idle
// @grant        GM_addStyle
// @grant        GM_setValue
// @grant        GM_getValue
// @grant        GM_registerMenuCommand
// @grant        GM_xmlhttpRequest
// @grant        unsafeWindow
// @connect      *
// @license Apache-2.0
// ==/UserScript==

/*
 * 边界声明：
 * 1. 本脚本只读取正在播放的 <video> 元素与其音轨，不请求、不解析、不解密 m3u8 / TS 分片 / 密钥。
 * 2. 默认引擎（speech）完全在本地/浏览器服务内完成，不上传任何课程音频。
 * 3. 云端 ASR 引擎只会把音频片段发送到「你自己填写的 endpoint」，密钥只存在浏览器本地 GM_setValue。
 *    启用前请确认该行为符合课程平台条款与你所在组织的数据合规要求。
 * 4. 若店铺使用自定义域名，请自行在 @match 中补一行。
 *
 * 引擎说明：
 *   speech : Web Speech API，延迟最低（<1s），零配置；但听的是麦克风，需要外放或系统音频回环，
 *           且依赖浏览器厂商的在线识别服务（Chrome 在国内常不可达，报 network 错误）。
 *   asr    : video.captureStream() 取音轨 -> 16kHz 单声道低码率 -> VAD 按句间停顿切分
 *           -> 并行请求、按序上屏 -> 你自填的 OpenAI 兼容 /v1/audio/transcriptions
 *           （或自建服务）。戴耳机可用、不拾环境音、识别质量最好，需要 API 地址与密钥。
 *           延迟 ≈ 停顿判定 320ms + 上传 + 推理，通常 1~3s；可选「预览字幕」进一步降低感知延迟。
 *           若要 <1s 的逐字实时，需要流式 ASR（WebSocket）协议，本脚本用的是文件转录接口。
 *   file   : 加载已有的 SRT / VTT 字幕（填 URL 或直接粘贴文本），零误差、零延迟。
 */

(function () {
  'use strict';

  var DEFAULTS = {
    enabled: true,
    engine: 'speech',
    lang: 'zh-CN',
    fontSize: 22,
    bottom: 12,
    hold: 4,
    maxLines: 1,
    splitMode: 'vad',
    silenceMs: 320,
    minSegMs: 700,
    maxSegMs: 4500,
    vadThreshold: 0.010,
    chunkMs: 3000,
    concurrency: 2,
    timeoutMs: 10000,
    preview: false,
    previewMs: 1400,
    audioKbps: 32,
    endpoint: '',
    apiKey: '',
    model: 'whisper-1',
    asrMode: 'multipart',
    prompt: '这是一段中文在线课程讲解，请使用简体中文并加上合适的标点符号。',
    subtitleText: '',
    subtitleUrl: ''
  };

  var cfg = loadCfg();
  var state = {
    video: null,
    overlay: null,
    panel: null,
    lines: null,
    interimEl: null,
    cues: [],
    shown: [],
    playing: false,
    running: false,
    recognition: null,
    recorder: null,
    recordTimer: null,
    audioCtx: null,
    audioStream: null,
    queue: [],
    inflight: 0,
    results: {},
    times: {},
    seq: 0,
    nextCommit: 0,
    gen: 0,
    lastLatency: 0,
    recordStream: null,
    monoStream: null,
    vadAnalyser: null,
    vadBuf: null,
    vadTimer: null,
    vadSegAt: 0,
    vadLastVoice: 0,
    vadHasVoice: false,
    vadCutting: false,
    vadSend: true,
    previewRecorder: null,
    previewTimer: null,
    previewBusy: false,
    previewSeq: 0,
    segStart: 0,
    fileCues: [],
    hideTimer: null,
    positionTimer: null,
    pollTimer: null,
    lastVideo: null,
    uiReady: false,
    lastPoll: 0,
    fileShown: ''
  };

  // ---------- 基础工具 ----------

  function loadCfg() {
    var out = {}, saved;
    try { saved = JSON.parse(GM_getValue('xlcap_cfg', '{}') || '{}'); } catch (e) { saved = {}; }
    Object.keys(DEFAULTS).forEach(function (k) {
      out[k] = Object.prototype.hasOwnProperty.call(saved, k) ? saved[k] : DEFAULTS[k];
    });
    return out;
  }

  function saveCfg() {
    GM_setValue('xlcap_cfg', JSON.stringify(cfg));
  }

  function sleep(ms) { return new Promise(function (r) { setTimeout(r, ms); }); }

  function setStatus(text, kind) {
    var el = document.getElementById('xlcap-status');
    if (el) {
      el.textContent = text;
      el.style.color = kind === 'error' ? '#ff8a80' : (kind === 'warn' ? '#ffd54f' : '#9fe1cb');
    }
    if (kind === 'error' || kind === 'warn') console.warn('[小鹅通字幕] ' + text);
  }

  function fmtSrtTime(sec) {
    var t = Math.max(0, Math.round(sec * 1000));
    var h = Math.floor(t / 3600000);
    var m = Math.floor((t % 3600000) / 60000);
    var s = Math.floor((t % 60000) / 1000);
    var ms = t % 1000;
    return String(h).padStart(2, '0') + ':' + String(m).padStart(2, '0') + ':' +
      String(s).padStart(2, '0') + ',' + String(ms).padStart(3, '0');
  }

  function parseTime(str) {
    var m = String(str).trim().match(/^(?:(\d+):)?(\d+):(\d+)[,.](\d+)$/);
    if (m) return (+(m[1] || 0)) * 3600 + (+m[2]) * 60 + (+m[3]) + (+m[4]) / 1000;
    var m2 = String(str).trim().match(/^(\d+):(\d+)[,.](\d+)$/);
    if (m2) return (+m2[1]) * 60 + (+m2[2]) + (+m2[3]) / 1000;
    return null;
  }

  function parseSubtitles(raw) {
    var text = String(raw || '').replace(/^\uFEFF/, '').replace(/\r/g, '');
    var out = [];
    var blocks = text.split(/\n\s*\n/);
    blocks.forEach(function (block) {
      var ls = block.split('\n').filter(function (l) { return l.trim() !== ''; });
      if (!ls.length) return;
      var idx = ls.findIndex(function (l) { return l.indexOf('-->') !== -1; });
      if (idx === -1) return;
      var parts = ls[idx].split('-->');
      var start = parseTime((parts[0] || '').trim().split(/\s+/)[0]);
      var end = parseTime((parts[1] || '').trim().split(/\s+/)[0]);
      if (start === null || end === null) return;
      var body = ls.slice(idx + 1).join('\n').replace(/<[^>]+>/g, '').trim();
      if (body) out.push({ start: start, end: end, text: body });
    });
    return out.sort(function (a, b) { return a.start - b.start; });
  }

  // ---------- 视频发现 ----------

  function pickVideo() {
    var list = Array.prototype.slice.call(document.querySelectorAll('video'));
    if (!list.length) return null;
    list.sort(function (a, b) {
      var ra = a.getBoundingClientRect(), rb = b.getBoundingClientRect();
      return (rb.width * rb.height) - (ra.width * ra.height);
    });
    var v = list[0];
    return (v.getBoundingClientRect().width > 200) ? v : null;
  }

  function bindVideo(v) {
    if (state.video === v) return;
    stopEngine();
    ensureUI();
    state.video = v;
    state.lastVideo = v;
    state.cues = [];
    state.shown = [];
    state.fileShown = '';
    clearInterim();
    setStatus('已找到视频，等待播放', 'ok');
    v.addEventListener('play', onPlay);
    v.addEventListener('pause', onPause);
    v.addEventListener('ended', onPause);
    v.addEventListener('seeked', clearInterim);
    if (!v.paused && !v.ended) onPlay();
  }

  function onPlay() {
    state.playing = true;
    if (cfg.enabled) startEngine();
  }

  function onPause() {
    state.playing = false;
    stopEngine();
    clearInterim();
  }

  // ---------- 字幕叠加层 ----------

  function buildOverlay() {
    GM_addStyle([
      '.xlcap-overlay{position:fixed;pointer-events:none;z-index:2147483647;overflow:hidden}',
      '.xlcap-lines{position:absolute;left:4%;right:4%;bottom:var(--xlcap-bottom,12%);text-align:center}',
      '.xlcap-line,.xlcap-interim{margin:3px 0;line-height:1.6;word-break:break-word}',
      '.xlcap-line span,.xlcap-interim span{display:inline;background:rgba(0,0,0,.74);color:#fff;',
      'padding:2px 9px;border-radius:5px;box-decoration-break:clone;-webkit-box-decoration-break:clone;',
      'box-shadow:0 1px 4px rgba(0,0,0,.5);text-shadow:0 1px 2px #000;font-size:var(--xlcap-size,22px)}',
      '.xlcap-interim span{background:rgba(0,0,0,.5);color:#d8e8ff;opacity:.85}',
      '.xlcap-panel{position:fixed;right:16px;bottom:16px;z-index:2147483647;width:288px;',
      'background:#1c2028;color:#e8edf4;border:1px solid #3a4351;border-radius:10px;',
      'max-height:80vh;overflow:auto;',
      'font:13px/1.5 system-ui,-apple-system,"Microsoft YaHei",sans-serif;box-shadow:0 6px 24px rgba(0,0,0,.45)}',
      '.xlcap-panel.xlcap-min .xlcap-body{display:none}',
      '.xlcap-head{display:flex;align-items:center;justify-content:space-between;padding:9px 11px;cursor:pointer}',
      '.xlcap-head b{font-weight:500;font-size:13px}',
      '.xlcap-body{padding:0 11px 11px;border-top:1px solid #333b47}',
      '.xlcap-status{margin:8px 0;font-size:12px;color:#9fe1cb;min-height:18px;word-break:break-all}',
      '.xlcap-row{display:flex;gap:6px;align-items:center;margin:7px 0}',
      '.xlcap-row label{flex:0 0 76px;color:#aeb9c8;font-size:12px}',
      '.xlcap-row input,.xlcap-row select,.xlcap-row textarea{flex:1;min-width:0;background:#252b35;',
      'color:#e8edf4;border:1px solid #414b59;border-radius:5px;padding:5px 6px;font-size:12px;',
      'font-family:inherit;box-sizing:border-box}',
      '.xlcap-row textarea{height:54px;resize:vertical}',
      '.xlcap-btns{display:flex;flex-wrap:wrap;gap:6px;margin-top:10px}',
      '.xlcap-btns button{background:#2f3b4c;color:#e8edf4;border:1px solid #46566c;',
      'border-radius:5px;padding:5px 9px;font-size:12px;cursor:pointer;font-family:inherit}',
      '.xlcap-btns button:hover{background:#3b4a60}',
      '.xlcap-btns button.xlcap-primary{background:#1d6f8b;border-color:#2a8bab}',
      '.xlcap-hint{color:#8895a6;font-size:11px;margin-top:9px;line-height:1.5}'
    ].join('\n'));

    var ov = document.createElement('div');
    ov.className = 'xlcap-overlay';
    ov.id = 'xlcap-overlay';
    var lines = document.createElement('div');
    lines.className = 'xlcap-lines';
    ov.appendChild(lines);
    document.body.appendChild(ov);
    state.overlay = ov;
    state.lines = lines;
    applyStyle();
    startPositionLoop();
    document.addEventListener('fullscreenchange', onFullscreenChange);
    window.addEventListener('resize', syncPosition);
    window.addEventListener('scroll', syncPosition, true);
  }

  function applyStyle() {
    if (!state.overlay) return;
    state.overlay.style.setProperty('--xlcap-size', cfg.fontSize + 'px');
    state.overlay.style.setProperty('--xlcap-bottom', cfg.bottom + '%');
  }

  function onFullscreenChange() {
    syncPosition();
  }

  function startPositionLoop() {
    if (state.positionTimer) return;
    state.positionTimer = setInterval(function () {
      syncPosition();
      tickFileEngine();
    }, 250);
  }

  function syncPosition() {
    var ov = state.overlay, v = state.video;
    if (!ov) return;
    var fs = document.fullscreenElement || document.webkitFullscreenElement;
    if (fs) {
      var host = (fs.tagName === 'VIDEO') ? (fs.parentElement || fs) : fs;
      if (ov.parentNode !== host) host.appendChild(ov);
      ov.style.position = 'absolute';
      ov.style.left = '0'; ov.style.top = '0';
      ov.style.width = '100%'; ov.style.height = '100%';
      return;
    }
    if (ov.parentNode !== document.body) document.body.appendChild(ov);
    ov.style.position = 'fixed';
    if (!v) { ov.style.display = 'none'; return; }
    var r = v.getBoundingClientRect();
    if (r.width < 100 || r.height < 100 || r.bottom < 0 || r.top > window.innerHeight) {
      ov.style.display = 'none';
      return;
    }
    ov.style.display = 'block';
    ov.style.left = r.left + 'px';
    ov.style.top = r.top + 'px';
    ov.style.width = r.width + 'px';
    ov.style.height = r.height + 'px';
  }

  function renderLines() {
    if (!state.lines) return;
    while (state.lines.firstChild) state.lines.removeChild(state.lines.firstChild);
    state.shown.slice(-Math.max(1, cfg.maxLines)).forEach(function (s) {
      var d = document.createElement('div');
      d.className = 'xlcap-line';
      var sp = document.createElement('span');
      sp.textContent = s;
      d.appendChild(sp);
      state.lines.appendChild(d);
    });
    state.interimEl = null;
  }

  function showLine(text) {
    if (!text) return;
    state.shown.push(text);
    if (state.shown.length > Math.max(1, cfg.maxLines)) state.shown.shift();
    renderLines();
    clearTimeout(state.hideTimer);
    state.hideTimer = setTimeout(function () {
      state.shown = [];
      renderLines();
      clearInterim();
    }, Math.max(1000, cfg.hold * 1000));
  }

  function showInterim(text) {
    if (!state.lines) return;
    clearInterim();
    if (!text) return;
    var d = document.createElement('div');
    d.className = 'xlcap-interim';
    d.id = 'xlcap-interim-el';
    var sp = document.createElement('span');
    sp.textContent = text;
    d.appendChild(sp);
    state.lines.appendChild(d);
    state.interimEl = d;
  }

  function clearInterim() {
    var el = document.getElementById('xlcap-interim-el');
    if (el && el.parentNode) el.parentNode.removeChild(el);
    state.interimEl = null;
  }

  // ---------- 字幕记录与导出 ----------

  function addCue(text, start, end) {
    text = String(text || '').trim();
    if (!text) return;
    var prev = state.cues[state.cues.length - 1];
    if (prev && prev.text === text && start - prev.end < 2) { prev.end = end; return; }
    state.cues.push({ text: text, start: start, end: end });
    showLine(text);
  }

  function exportSrt() {
    var cues = state.cues.slice();
    if (!cues.length) { setStatus('还没有可导出的字幕', 'warn'); return; }
    var srt = cues.map(function (c, i) {
      return (i + 1) + '\n' + fmtSrtTime(c.start) + ' --> ' + fmtSrtTime(Math.max(c.end, c.start + 0.8)) +
        '\n' + c.text + '\n';
    }).join('\n');
    var blob = new Blob(['\uFEFF' + srt], { type: 'text/plain;charset=utf-8' });
    var a = document.createElement('a');
    a.href = URL.createObjectURL(blob);
    a.download = '小鹅通字幕-' + new Date().toISOString().slice(0, 19).replace(/[:T]/g, '-') + '.srt';
    document.body.appendChild(a);
    a.click();
    setTimeout(function () { URL.revokeObjectURL(a.href); a.remove(); }, 1000);
    setStatus('已导出 ' + cues.length + ' 条字幕', 'ok');
  }

  // ---------- 引擎 1：浏览器语音识别 ----------

  function startSpeech() {
    var uw = (typeof unsafeWindow !== 'undefined') ? unsafeWindow : window;
    var SR = window.SpeechRecognition || window.webkitSpeechRecognition ||
      uw.SpeechRecognition || uw.webkitSpeechRecognition;
    if (!SR) {
      setStatus('此浏览器不支持 Web Speech API，请在设置里切换为「云端 ASR」', 'error');
      return false;
    }
    var rec = new SR();
    rec.lang = cfg.lang;
    rec.continuous = true;
    rec.interimResults = true;
    rec.maxAlternatives = 1;
    rec.onresult = function (e) {
      for (var i = e.resultIndex; i < e.results.length; i++) {
        var r = e.results[i];
        var txt = (r[0] && r[0].transcript || '').trim();
        if (!txt) continue;
        if (r.isFinal) {
          clearInterim();
          var now = state.video ? state.video.currentTime : 0;
          addCue(txt, Math.max(0, now - 2.5), now);
        } else {
          showInterim(txt);
        }
      }
    };
    rec.onerror = function (e) {
      if (e.error === 'network') {
        setStatus('浏览器识别服务不可达（国内常见）。请在设置里切换为「云端 ASR」或「字幕文件」', 'error');
      } else if (e.error === 'not-allowed' || e.error === 'service-not-allowed') {
        setStatus('麦克风权限被拒绝，请在地址栏允许麦克风后重试', 'error');
      } else if (e.error !== 'no-speech' && e.error !== 'aborted') {
        setStatus('语音识别错误：' + e.error, 'error');
      }
    };
    rec.onend = function () {
      if (state.running && state.playing && cfg.engine === 'speech' && cfg.enabled) {
        setTimeout(function () { try { rec.start(); } catch (e) { /* 已在停止 */ } }, 350);
      }
    };
    try { rec.start(); } catch (e) { setStatus('无法启动语音识别：' + e.message, 'error'); return false; }
    state.recognition = rec;
    setStatus('语音识别中｜请外放声音，或把系统音频回环到麦克风', 'ok');
    return true;
  }

  function stopSpeech() {
    if (!state.recognition) return;
    var rec = state.recognition;
    state.recognition = null;
    rec.onend = null;
    rec.onresult = null;
    rec.onerror = null;
    try { rec.abort(); } catch (e) { /* noop */ }
  }

  // ---------- 引擎 2：视频音轨 -> 云端 ASR ----------

  function pickMime() {
    if (typeof MediaRecorder === 'undefined') return null;
    var list = ['audio/webm;codecs=opus', 'audio/webm', 'audio/mp4', 'audio/ogg;codecs=opus'];
    for (var i = 0; i < list.length; i++) {
      if (MediaRecorder.isTypeSupported && MediaRecorder.isTypeSupported(list[i])) return list[i];
    }
    return '';
  }

  // captureStream -> AudioContext(16kHz) -> Analyser(VAD) -> 单声道 MediaStream
  // 单声道 + 16kHz 既是 Whisper 的原生输入，也能把上传体积压到最小。
  function buildAudioGraph(stream) {
    var AC = window.AudioContext || window.webkitAudioContext;
    var ctx = null;
    try {
      try { ctx = new AC({ sampleRate: 16000 }); } catch (e) { ctx = new AC(); }
      if (ctx.state === 'suspended' && ctx.resume) ctx.resume().catch(function () {});
      var src = ctx.createMediaStreamSource(stream);
      var an = ctx.createAnalyser();
      an.fftSize = 1024;
      an.smoothingTimeConstant = 0;
      var dst = ctx.createMediaStreamDestination();
      try { dst.channelCount = 1; dst.channelCountMode = 'explicit'; } catch (e) { /* noop */ }
      src.connect(an);
      an.connect(dst);
      state.audioCtx = ctx;
      state.vadAnalyser = an;
      state.vadBuf = new Float32Array(an.fftSize);
      state.monoStream = dst.stream;
      return dst.stream;
    } catch (e) {
      setStatus('音频重采样不可用，改用原始音轨：' + e.message, 'warn');
      try { if (ctx) ctx.close(); } catch (e2) { /* noop */ }
      state.audioCtx = null;
      state.vadAnalyser = null;
      return stream;
    }
  }

  function rms() {
    var an = state.vadAnalyser;
    if (!an || !state.vadBuf) return 0;
    an.getFloatTimeDomainData(state.vadBuf);
    var s = 0, b = state.vadBuf;
    for (var i = 0; i < b.length; i++) s += b[i] * b[i];
    return Math.sqrt(s / b.length);
  }

  function resetVadSegment() {
    var now = Date.now();
    state.vadSegAt = now;
    state.vadLastVoice = now;
    state.vadHasVoice = false;
  }

  function probeAndStart(mime) {
    var n = 0, peak = 0;
    var tick = setInterval(function () {
      peak = Math.max(peak, rms());
      if (++n < 34) return;
      clearInterval(tick);
      if (!state.running || cfg.engine !== 'asr') return;
      if (peak <= 0.0015) {
        setStatus('音轨是静音的（DRM 或跨域限制）。建议改用「语音识别」引擎', 'warn');
      } else {
        setStatus('音轨正常｜按停顿切分中…', 'ok');
      }
      beginRecording(mime);
    }, 30);
    state.vadTimer = tick;
  }

  // 短片段相邻处常有重叠识别，去掉上一句尾部与本句头部重复的字。
  function trimOverlap(prev, cur) {
    if (!prev) return cur;
    var max = Math.min(prev.length, cur.length, 14);
    for (var k = max; k >= 2; k--) {
      if (prev.slice(-k) === cur.slice(0, k)) return cur.slice(k);
    }
    return cur;
  }

  function callASR(blob, mime) {
    return new Promise(function (resolve, reject) {
      var headers = {};
      if (cfg.apiKey) headers.Authorization = 'Bearer ' + cfg.apiKey;
      var data;
      if (cfg.asrMode === 'raw') {
        headers['Content-Type'] = mime || 'application/octet-stream';
        data = blob;
      } else {
        var fd = new FormData();
        fd.append('file', blob, 'chunk.' + (mime.indexOf('mp4') !== -1 ? 'mp4' : 'webm'));
        fd.append('model', cfg.model || 'whisper-1');
        if (cfg.lang) fd.append('language', cfg.lang.split('-')[0]);
        fd.append('response_format', 'json');
        if (cfg.prompt) fd.append('prompt', cfg.prompt);
        data = fd;
      }
      GM_xmlhttpRequest({
        method: 'POST',
        url: cfg.endpoint,
        data: data,
        headers: headers,
        timeout: Math.max(3000, cfg.timeoutMs),
        onload: function (r) {
          if (r.status < 200 || r.status >= 300) {
            reject(new Error('HTTP ' + r.status + ' ' + String(r.responseText || '').slice(0, 160)));
            return;
          }
          var j;
          try { j = JSON.parse(r.responseText || '{}'); } catch (e) {
            reject(new Error('返回不是 JSON：' + String(r.responseText || '').slice(0, 160)));
            return;
          }
          var text = j.text || (j.data && j.data.text) || j.result || '';
          if (!text) { reject(new Error('识别结果为空，请检查返回结果格式')); return; }
          resolve(String(text).trim());
        },
        onerror: function () { reject(new Error('请求失败：地址不可达、被 CORS/扩展授权拦截，或网络错误')); },
        ontimeout: function () { reject(new Error('识别超时，可增大超时时间或换更快的端点')); }
      });
    });
  }

  // 并行发送，但按片段序号顺序上屏，避免快片段被慢片段拖住或乱序。
  function pumpQueue() {
    var limit = Math.max(1, Math.min(4, cfg.concurrency));
    while (state.inflight < limit && state.queue.length) submitSeg(state.queue.shift());
    while (state.queue.length > 8) {
      var old = state.queue.shift();
      state.results[old.seq] = '';
      setStatus('识别跟不上播放，已跳过过期片段；可调大「最长片段」或增加并行数', 'warn');
    }
    commitReady();
  }

  function submitSeg(item) {
    var gen = state.gen;
    state.inflight++;
    callASR(item.blob, item.mime).then(function (text) {
      if (gen === state.gen) state.results[item.seq] = text || '';
    }).catch(function (e) {
      if (gen === state.gen) {
        state.results[item.seq] = '';
        setStatus('识别失败：' + e.message, 'error');
      }
    }).then(function () {
      if (gen !== state.gen) return;
      state.inflight--;
      state.lastLatency = (Date.now() - item.createdAt) / 1000;
      commitReady();
      pumpQueue();
    });
  }

  function commitReady() {
    while (Object.prototype.hasOwnProperty.call(state.results, state.nextCommit)) {
      var seq = state.nextCommit;
      var text = state.results[seq];
      var tm = state.times[seq] || { start: 0, end: 0 };
      delete state.results[seq];
      delete state.times[seq];
      state.nextCommit++;
      if (!text) continue;
      var prev = state.cues.length ? state.cues[state.cues.length - 1].text : '';
      var merged = trimOverlap(prev, text);
      if (!merged) continue;
      clearInterim();
      addCue(merged, tm.start, tm.end);
      setStatus('云端 ASR｜末条延迟 ' + state.lastLatency.toFixed(1) + 's｜已出 ' + state.cues.length + ' 条', 'ok');
    }
  }

  function startASR() {
    if (!cfg.endpoint) {
      setStatus('请先在设置里填写 ASR 接口地址（endpoint）', 'warn');
      openSettings();
      return false;
    }
    var v = state.video;
    if (!v || typeof v.captureStream !== 'function') {
      setStatus('浏览器不支持 captureStream，无法读取视频音轨', 'error');
      return false;
    }
    var mime = pickMime();
    if (mime === null) {
      setStatus('浏览器不支持 MediaRecorder，请换 Chrome / Edge', 'error');
      return false;
    }
    var stream;
    try { stream = v.captureStream(); } catch (e) {
      setStatus('读取音轨失败：' + e.message + '（加密或跨域受限时可改用「语音识别」引擎）', 'error');
      return false;
    }
    var tracks = stream.getAudioTracks();
    if (!tracks.length) {
      setStatus('未捕获到音轨（DRM/加密流会这样）。请改用「语音识别」引擎', 'error');
      return false;
    }
    state.audioStream = new MediaStream(tracks);
    state.gen++;
    state.seq = 0;
    state.nextCommit = 0;
    state.results = {};
    state.times = {};
    state.queue = [];
    state.inflight = 0;
    state.lastLatency = 0;
    state.vadCutting = false;
    state.vadSend = true;
    state.previewSeq = 0;
    state.recordStream = buildAudioGraph(state.audioStream);
    setStatus('正在检测音轨…', 'ok');
    if (!state.vadAnalyser) beginRecording(mime);
    else probeAndStart(mime);
    return true;
  }

  function beginRecording(mime) {
    var opts = { audioBitsPerSecond: Math.max(16000, cfg.audioKbps * 1000) };
    if (mime) opts.mimeType = mime;
    var mimeBase = (mime || 'audio/webm').split(';')[0];
    var rec;
    try { rec = new MediaRecorder(state.recordStream || state.audioStream, opts); }
    catch (e) {
      try { rec = new MediaRecorder(state.audioStream, opts); }
      catch (e2) { setStatus('无法创建录音器：' + e2.message, 'error'); return; }
    }
    var chunks = [];
    rec.ondataavailable = function (e) { if (e.data && e.data.size) chunks.push(e.data); };
    rec.onstop = function () {
      var blob = new Blob(chunks, { type: mimeBase });
      chunks = [];
      var send = state.vadSend;
      state.vadSend = true;
      state.vadCutting = false;
      if (!state.running || cfg.engine !== 'asr') return;
      if (send && blob.size > 800) {
        var seq = state.seq++;
        state.times[seq] = {
          start: state.segStart,
          end: state.video ? state.video.currentTime : state.segStart
        };
        state.queue.push({ seq: seq, blob: blob, mime: mimeBase, createdAt: Date.now() });
        pumpQueue();
      }
      state.segStart = state.video ? state.video.currentTime : 0;
      resetVadSegment();
      if (state.playing && rec.state === 'inactive') {
        try { rec.start(); } catch (e) { setStatus('录音器重启失败：' + e.message, 'error'); }
      }
    };
    state.recorder = rec;
    state.segStart = state.video ? state.video.currentTime : 0;
    resetVadSegment();
    rec.start();
    startSplitter(rec);
    if (cfg.preview) startPreview(mime, mimeBase);
    setStatus(cfg.splitMode === 'vad' && state.vadAnalyser
      ? '云端 ASR｜按停顿切分 · 并行 ' + cfg.concurrency + ' 路'
      : '云端 ASR｜固定 ' + (cfg.chunkMs / 1000).toFixed(1) + 's 切片', 'ok');
  }

  // VAD：检测到句间停顿就立刻发片段，而不是攒够固定时长。
  function startSplitter(rec) {
    if (cfg.splitMode !== 'vad' || !state.vadAnalyser) {
      state.recordTimer = setInterval(function () {
        if (rec.state === 'recording') {
          state.vadSend = true;
          try { rec.stop(); } catch (e) { /* noop */ }
        }
      }, Math.max(1000, cfg.chunkMs));
      return;
    }
    var threshold = Math.max(0.002, cfg.vadThreshold);
    var silenceMs = Math.max(120, cfg.silenceMs);
    var minMs = Math.max(200, cfg.minSegMs);
    var maxMs = Math.max(minMs + 500, cfg.maxSegMs);
    var silentTicks = 0;
    state.vadTimer = setInterval(function () {
      if (!state.running || cfg.engine !== 'asr' || rec.state !== 'recording' || state.vadCutting) return;
      var now = Date.now();
      if (rms() > threshold) {
        state.vadLastVoice = now;
        state.vadHasVoice = true;
        silentTicks = 0;
      } else if (++silentTicks === 200) {
        setStatus('长时间没检测到人声：可把「音量阈值」调低，或确认音轨非静音', 'warn');
      }
      var dur = now - state.vadSegAt;
      var cut = false;
      if (state.vadHasVoice && dur >= minMs && (now - state.vadLastVoice) >= silenceMs) cut = true;
      else if (state.vadHasVoice && dur >= maxMs) cut = true;
      else if (!state.vadHasVoice && dur >= maxMs * 2) { state.vadSend = false; cut = true; }
      if (!cut) return;
      state.vadCutting = true;
      try { rec.stop(); } catch (e) { state.vadCutting = false; }
    }, 30);
  }

  // 可选预览：独立录音器每 previewMs 发一小段，结果以半透明临时字幕先显示。
  function startPreview(mime, mimeBase) {
    var opts = { audioBitsPerSecond: Math.max(16000, cfg.audioKbps * 1000) };
    if (mime) opts.mimeType = mime;
    var rec;
    try { rec = new MediaRecorder(state.recordStream || state.audioStream, opts); }
    catch (e) { return; }
    var chunks = [];
    rec.ondataavailable = function (e) { if (e.data && e.data.size) chunks.push(e.data); };
    rec.onstop = function () {
      var blob = new Blob(chunks, { type: mimeBase });
      chunks = [];
      if (!state.running || cfg.engine !== 'asr' || !cfg.preview) return;
      if (blob.size > 800 && !state.previewBusy) {
        state.previewBusy = true;
        var my = ++state.previewSeq;
        var gen = state.gen;
        callASR(blob, mimeBase).then(function (t) {
          if (gen === state.gen && my === state.previewSeq && t && cfg.engine === 'asr') showInterim(t);
        }).catch(function () {}).then(function () { state.previewBusy = false; });
      }
      if (state.playing && rec.state === 'inactive') {
        try { rec.start(); } catch (e) { /* noop */ }
      }
    };
    rec.start();
    state.previewRecorder = rec;
    state.previewTimer = setInterval(function () {
      if (rec.state === 'recording') { try { rec.stop(); } catch (e) { /* noop */ } }
    }, Math.max(800, cfg.previewMs));
  }

  function stopASR() {
    state.gen++;
    [state.recordTimer, state.vadTimer, state.previewTimer].forEach(function (t) { clearInterval(t); });
    state.recordTimer = state.vadTimer = state.previewTimer = null;
    [state.recorder, state.previewRecorder].forEach(function (rec) {
      if (!rec) return;
      rec.onstop = null;
      rec.ondataavailable = null;
      try { if (rec.state !== 'inactive') rec.stop(); } catch (e) { /* noop */ }
    });
    state.recorder = null;
    state.previewRecorder = null;
    [state.audioStream, state.monoStream].forEach(function (s) {
      if (!s) return;
      try { s.getTracks().forEach(function (t) { t.stop(); }); } catch (e) { /* noop */ }
    });
    state.audioStream = null;
    state.monoStream = null;
    state.recordStream = null;
    if (state.audioCtx) { try { state.audioCtx.close(); } catch (e) { /* noop */ } }
    state.audioCtx = null;
    state.vadAnalyser = null;
    state.vadBuf = null;
    state.queue = [];
    state.results = {};
    state.times = {};
    state.inflight = 0;
    state.previewBusy = false;
    state.vadCutting = false;
    state.vadSend = true;
    clearInterim();
  }

  // ---------- 引擎 3：已有字幕文件 ----------

  function loadSubtitleText(raw) {
    var cues = parseSubtitles(raw);
    state.fileCues = cues;
    if (!cues.length) {
      setStatus('没解析出字幕，请检查 SRT / VTT 格式', 'error');
      return false;
    }
    setStatus('已加载 ' + cues.length + ' 条字幕，随时间轴显示', 'ok');
    return true;
  }

  function startFileEngine() {
    if (state.fileCues.length) {
      setStatus('字幕文件引擎运行中｜共 ' + state.fileCues.length + ' 条', 'ok');
      return true;
    }
    if (cfg.subtitleText) return loadSubtitleText(cfg.subtitleText);
    if (cfg.subtitleUrl) {
      setStatus('正在下载字幕文件…', 'ok');
      GM_xmlhttpRequest({
        method: 'GET',
        url: cfg.subtitleUrl,
        timeout: 15000,
        onload: function (r) {
          if (r.status >= 200 && r.status < 300) {
            cfg.subtitleText = r.responseText;
            saveCfg();
            loadSubtitleText(r.responseText);
          } else setStatus('字幕下载失败：HTTP ' + r.status, 'error');
        },
        onerror: function () { setStatus('字幕下载失败：地址不可达或跨域被拦截', 'error'); },
        ontimeout: function () { setStatus('字幕下载超时', 'error'); }
      });
      return true;
    }
    setStatus('请在设置里粘贴 SRT/VTT 文本或填字幕文件地址', 'warn');
    openSettings();
    return false;
  }

  function tickFileEngine() {
    if (cfg.engine !== 'file' || !state.playing || !state.video) return;
    var t = state.video.currentTime;
    var want = '';
    for (var i = 0; i < state.fileCues.length; i++) {
      var c = state.fileCues[i];
      if (t >= c.start && t <= c.end) { want = c.text; break; }
      if (c.start > t) break;
    }
    if (want === state.fileShown) return;
    state.fileShown = want;
    state.shown = want ? [want] : [];
    renderLines();
  }

  // ---------- 引擎调度 ----------

  function startEngine() {
    if (state.running || !state.video) return;
    if (!cfg.enabled) return;
    state.running = true;
    var ok = false;
    if (cfg.engine === 'asr') ok = startASR();
    else if (cfg.engine === 'file') ok = startFileEngine();
    else ok = startSpeech();
    if (!ok) state.running = false;
  }

  function stopEngine() {
    state.running = false;
    stopSpeech();
    stopASR();
    state.fileCues = [];
    state.fileShown = '';
  }

  function switchEngine(name) {
    var wasPlaying = state.playing;
    clearTimeout(state.hideTimer);
    stopEngine();
    cfg.engine = name;
    saveCfg();
    syncPanel();
    if (wasPlaying && cfg.enabled) startEngine();
    else setStatus('已切换到「' + engineLabel(name) + '」，等待播放', 'ok');
  }

  function engineLabel(n) {
    return n === 'asr' ? '云端 ASR' : (n === 'file' ? '字幕文件' : '语音识别');
  }

  // ---------- 控制面板 ----------

  function buildPanel() {
    var p = document.createElement('div');
    p.className = 'xlcap-panel';
    p.id = 'xlcap-panel';
    p.innerHTML = [
      '<div class="xlcap-head" id="xlcap-head"><b>小鹅通 · 实时字幕</b><span id="xlcap-toggle">收起 ▾</span></div>',
      '<div class="xlcap-body">',
      '  <div class="xlcap-status" id="xlcap-status">初始化中…</div>',
      '  <div class="xlcap-row"><label>识别引擎</label><select id="xlcap-engine">',
      '    <option value="speech">语音识别（零配置）</option>',
      '    <option value="asr">云端 ASR（推荐）</option>',
      '    <option value="file">已有 SRT / VTT 字幕</option>',
      '  </select></div>',
      '  <div class="xlcap-btns">',
      '    <button id="xlcap-power" class="xlcap-primary">开启 / 暂停字幕</button>',
      '    <button id="xlcap-export">导出 SRT</button>',
      '    <button id="xlcap-settings">设置 ▾</button>',
      '  </div>',
      '  <div id="xlcap-settings-box" style="display:none">',
      '    <div class="xlcap-row"><label>识别语言</label><input id="xlcap-lang" placeholder="zh-CN"></div>',
      '    <div class="xlcap-row"><label>ASR 地址</label><input id="xlcap-endpoint" placeholder="https://…/v1/audio/transcriptions"></div>',
      '    <div class="xlcap-row"><label>API Key</label><input id="xlcap-key" type="password" placeholder="仅存本地"></div>',
      '    <div class="xlcap-row"><label>模型名</label><input id="xlcap-model" placeholder="whisper-1"></div>',
      '    <div class="xlcap-row"><label>请求格式</label><select id="xlcap-mode">',
      '      <option value="multipart">multipart（OpenAI 兼容）</option>',
      '      <option value="raw">raw（裸音频 POST）</option></select></div>',
      '    <div class="xlcap-row"><label>切分方式</label><select id="xlcap-split">',
      '      <option value="vad">按停顿（低延迟）</option>',
      '      <option value="fixed">固定切片</option></select></div>',
      '    <div class="xlcap-row"><label>停顿判定ms</label><input id="xlcap-silence" type="number" min="120" step="20"></div>',
      '    <div class="xlcap-row"><label>最短片段ms</label><input id="xlcap-minseg" type="number" min="200" step="100"></div>',
      '    <div class="xlcap-row"><label>最长片段ms</label><input id="xlcap-maxseg" type="number" min="1000" step="500"></div>',
      '    <div class="xlcap-row"><label>音量阈值</label><input id="xlcap-vad" type="number" step="0.002" min="0.002" max="0.1"></div>',
      '    <div class="xlcap-row"><label>并行请求</label><input id="xlcap-conc" type="number" min="1" max="4"></div>',
      '    <div class="xlcap-row"><label>音频码率k</label><input id="xlcap-kbps" type="number" min="16" max="96"></div>',
      '    <div class="xlcap-row"><label>超时(ms)</label><input id="xlcap-timeout" type="number" min="3000" step="1000"></div>',
      '    <div class="xlcap-row"><label>预览字幕</label><select id="xlcap-preview">',
      '      <option value="off">关闭（省请求）</option>',
      '      <option value="on">开启（请求翻倍）</option></select></div>',
      '    <div class="xlcap-row"><label>预览间隔ms</label><input id="xlcap-prevms" type="number" min="800" step="200"></div>',
      '    <div class="xlcap-row"><label>固定切片ms</label><input id="xlcap-chunk" type="number" min="1000" step="500"></div>',
      '    <div class="xlcap-row"><label>字幕字号</label><input id="xlcap-font" type="number" min="12" max="48"></div>',
      '    <div class="xlcap-row"><label>底部距离%</label><input id="xlcap-bottom" type="number" min="2" max="45"></div>',
      '    <div class="xlcap-row"><label>字幕 URL</label><input id="xlcap-url" placeholder="可选，.srt / .vtt"></div>',
      '    <div class="xlcap-row"><label>字幕文本</label><textarea id="xlcap-text" placeholder="也可直接粘贴 SRT / VTT 内容"></textarea></div>',
      '    <div class="xlcap-btns"><button id="xlcap-save" class="xlcap-primary">保存设置</button></div>',
      '    <div class="xlcap-hint">按停顿切分：句间静音达到「停顿判定」就立刻发包，不必攒够固定时长；音频已压成 16kHz 单声道低码率，请求并行发送、结果按序上屏，状态栏会显示末条延迟。开启「预览字幕」会额外周期性发短片段，感知更即时但请求量约翻倍。语音识别引擎听的是麦克风，戴耳机需外放；云端 ASR 会把音频发到你填的地址，启用前请确认合规。</div>',
      '  </div>',
      '</div>'
    ].join('\n');
    document.body.appendChild(p);
    state.panel = p;

    document.getElementById('xlcap-head').addEventListener('click', function () {
      p.classList.toggle('xlcap-min');
      document.getElementById('xlcap-toggle').textContent = p.classList.contains('xlcap-min') ? '展开 ▸' : '收起 ▾';
    });
    document.getElementById('xlcap-engine').addEventListener('change', function (e) { switchEngine(e.target.value); });
    document.getElementById('xlcap-power').addEventListener('click', function () {
      cfg.enabled = !cfg.enabled;
      saveCfg();
      if (cfg.enabled) { setStatus('已开启', 'ok'); if (state.playing) startEngine(); }
      else { stopEngine(); setStatus('已暂停', 'warn'); state.shown = []; renderLines(); clearInterim(); }
      syncPanel();
    });
    document.getElementById('xlcap-export').addEventListener('click', exportSrt);
    document.getElementById('xlcap-settings').addEventListener('click', function () {
      var box = document.getElementById('xlcap-settings-box');
      var open = box.style.display === 'none';
      box.style.display = open ? 'block' : 'none';
      this.textContent = open ? '设置 ▴' : '设置 ▾';
    });
    document.getElementById('xlcap-save').addEventListener('click', saveFromPanel);
    syncPanel();
  }

  function syncPanel() {
    var q = function (id) { return document.getElementById(id); };
    if (!q('xlcap-panel')) return;
    q('xlcap-engine').value = cfg.engine;
    q('xlcap-lang').value = cfg.lang;
    q('xlcap-endpoint').value = cfg.endpoint;
    q('xlcap-key').value = cfg.apiKey;
    q('xlcap-model').value = cfg.model;
    q('xlcap-mode').value = cfg.asrMode;
    q('xlcap-split').value = cfg.splitMode;
    q('xlcap-silence').value = cfg.silenceMs;
    q('xlcap-minseg').value = cfg.minSegMs;
    q('xlcap-maxseg').value = cfg.maxSegMs;
    q('xlcap-vad').value = cfg.vadThreshold;
    q('xlcap-conc').value = cfg.concurrency;
    q('xlcap-kbps').value = cfg.audioKbps;
    q('xlcap-timeout').value = cfg.timeoutMs;
    q('xlcap-preview').value = cfg.preview ? 'on' : 'off';
    q('xlcap-prevms').value = cfg.previewMs;
    q('xlcap-chunk').value = cfg.chunkMs;
    q('xlcap-font').value = cfg.fontSize;
    q('xlcap-bottom').value = cfg.bottom;
    q('xlcap-url').value = cfg.subtitleUrl;
    q('xlcap-text').value = cfg.subtitleText;
    q('xlcap-power').textContent = cfg.enabled ? '暂停字幕' : '开启字幕';
  }

  function saveFromPanel() {
    var q = function (id) { return document.getElementById(id).value; };
    cfg.lang = q('xlcap-lang') || DEFAULTS.lang;
    cfg.endpoint = q('xlcap-endpoint').trim();
    cfg.apiKey = q('xlcap-key').trim();
    cfg.model = q('xlcap-model') || DEFAULTS.model;
    cfg.asrMode = q('xlcap-mode');
    cfg.splitMode = q('xlcap-split') === 'fixed' ? 'fixed' : 'vad';
    cfg.silenceMs = Math.max(120, parseInt(q('xlcap-silence'), 10) || DEFAULTS.silenceMs);
    cfg.minSegMs = Math.max(200, parseInt(q('xlcap-minseg'), 10) || DEFAULTS.minSegMs);
    cfg.maxSegMs = Math.max(cfg.minSegMs + 500, parseInt(q('xlcap-maxseg'), 10) || DEFAULTS.maxSegMs);
    cfg.vadThreshold = Math.min(0.1, Math.max(0.002, parseFloat(q('xlcap-vad')) || DEFAULTS.vadThreshold));
    cfg.concurrency = Math.min(4, Math.max(1, parseInt(q('xlcap-conc'), 10) || DEFAULTS.concurrency));
    cfg.audioKbps = Math.min(96, Math.max(16, parseInt(q('xlcap-kbps'), 10) || DEFAULTS.audioKbps));
    cfg.timeoutMs = Math.max(3000, parseInt(q('xlcap-timeout'), 10) || DEFAULTS.timeoutMs);
    cfg.preview = q('xlcap-preview') === 'on';
    cfg.previewMs = Math.max(800, parseInt(q('xlcap-prevms'), 10) || DEFAULTS.previewMs);
    cfg.chunkMs = Math.max(1000, parseInt(q('xlcap-chunk'), 10) || DEFAULTS.chunkMs);
    cfg.fontSize = Math.min(48, Math.max(12, parseInt(q('xlcap-font'), 10) || DEFAULTS.fontSize));
    cfg.bottom = Math.min(45, Math.max(2, parseInt(q('xlcap-bottom'), 10) || DEFAULTS.bottom));
    cfg.subtitleUrl = q('xlcap-url').trim();
    cfg.subtitleText = q('xlcap-text');
    saveCfg();
    applyStyle();
    state.fileCues = [];
    if (cfg.subtitleText) loadSubtitleText(cfg.subtitleText);
    setStatus('设置已保存' + (cfg.engine === 'asr' && cfg.endpoint ? '，播放时生效' : ''), 'ok');
  }

  function openSettings() {
    var box = document.getElementById('xlcap-settings-box');
    if (box && box.style.display === 'none') {
      box.style.display = 'block';
      state.panel.classList.remove('xlcap-min');
      var b = document.getElementById('xlcap-settings');
      if (b) b.textContent = '设置 ▴';
    }
  }

  // ---------- 启动 ----------

  function ensureUI() {
    if (state.uiReady) return;
    state.uiReady = true;
    buildPanel();
    GM_registerMenuCommand('字幕：开启 / 暂停', function () {
      var b = document.getElementById('xlcap-power');
      if (b) b.click();
    });
    GM_registerMenuCommand('字幕：切换引擎', function () {
      var order = ['speech', 'asr', 'file'];
      switchEngine(order[(order.indexOf(cfg.engine) + 1) % order.length]);
    });
    GM_registerMenuCommand('字幕：导出 SRT', exportSrt);
    GM_registerMenuCommand('字幕：设置', openSettings);
    setStatus(cfg.engine === 'asr' && !cfg.endpoint
      ? '请先在设置里填 ASR 接口地址' : '准备就绪｜' + engineLabel(cfg.engine),
      cfg.engine === 'asr' && !cfg.endpoint ? 'warn' : 'ok');
  }

  function boot() {
    buildOverlay();
    var poll = function () {
      var now = Date.now();
      if (now - state.lastPoll < 700) return;
      state.lastPoll = now;
      var v = pickVideo();
      if (v) bindVideo(v);
    };
    poll();
    state.pollTimer = setInterval(poll, 1200);
    if (document.documentElement) {
      new MutationObserver(poll).observe(document.documentElement, {
        childList: true, subtree: true
      });
    }
  }

  if (document.readyState === 'loading') {
    document.addEventListener('DOMContentLoaded', boot);
  } else {
    boot();
  }
})();