fix: align spoken sentences and reuse prepared audio on seeks

Do not merge multiple sentences in one source part; keep connective prefixes with their clauses and preserve lexical tones. Seek within live current/next jobs instead of stop/rebuilding them, with epochs and in-flight guards. Retain only two played/ready descriptors for completed-cache-only reopening; misses prepare just the target, never an old batch. Preserve pause, sleep, source Unicode, inline/audio cleanup and all search behavior.

docs/TTS_PILOT.md

@@ -5,6 +5,12 @@
 
 - 固定女声01。只提供音高、即时语速、睡眠定时、上一句／下一句、暂停／继续、
   从当前阅读位置开始、回到播放位置。双击正文会选中所在句子并从那里继续播放。
+- 音频片段不会再把多句正文合成一个长块；按钮按当前句子的原文位置前后移动。
+  当前／下一批内已生成的句子直接切换，不取消重建。双击同一句可从头重播，
+  不重新合成。尚未生成的位置仍需准备，不能保证所有跳读都瞬间出声。
+- 只保留最近两批“已听且完成”的短小分段元数据（不保留额外声音／任务凭据）。
+  跳回时先用原样请求尝试后端完成缓存，`cache_only=true` 命中才取，不会因未命中
+  而排队生成整批旧文；未命中后只从所选位置重新准备。仍最多两个活跃批次任务。
 - 原文保留在搜索页内存，不修改正文或搜索高亮；后端返回原始 UTF-16 句段范围。
   高亮是句段级，不把播放时间比例冒充逐字定位。“回到播放位置”可重新打开原预览。
 - 去掉整篇两万字限制：首批目标 48 字，后续目标 480 字，每个 API 请求仍最多 800
@@ -12,6 +18,8 @@
   只保留当前／下一批任务及一个后续 MP3 预载，不整本上传或生成整本音频。
 - 词边界由浏览器中文分词器在有限窗口内提供，优先在自然标点停顿，避免把复合词、
   专有名词和数字在 60 字处硬拆。首个音频片段目标约 16 字，之后用较完整的分句。
+  “但是、同时、然而”等引导语与后续分句放在一起，避免拆成独立很短的音频；
+  保留词典正常声调，不统一改轻声、加重音或对字词任意调音量。声线听感仍需试听。
   源范围经 API／worker 验证。首个创建请求最多异步等 8 秒，首段可用立即返回，减少
   反复穿过 HF 入口轮询；这不是固定的首字速度保证。
 - 首批额外请求 `inline_first=true`：把一个已完成、最多 128 KiB 的首段 MP3 随创建

static/index.html

@@ -158,7 +158,7 @@
 </section>
 
 <script>if ("serviceWorker" in navigator) { navigator.serviceWorker.register("/sw.js", { scope: "/" }); }</script>
-<script src="/static/tts-client.js?v=inline-4"></script>
+<script src="/static/tts-client.js?v=navigation-5"></script>
 <script src="/static/app.js?v=phrases-3"></script>
 </body>
 </html>

static/tts-client.js

@@ -2,10 +2,21 @@
 // Pure, bounded rolling source planner. All offsets refer to original UTF-16.
 globalThis.CCRDSpeechPlan = (() => {
   const ending = /[。！？!?；;\n]/u;
-  const closers = /[”’"'」』）)\t \r]/u;
   const ascii = /[A-Za-z0-9]/u;
   const words = typeof Intl.Segmenter === "function" ? new Intl.Segmenter("zh", { granularity: "word" }) : null;
-  const dangling = new Set(["的", "地", "得", "和", "与", "及", "或", "把", "被", "在", "对", "向", "为", "将", "不", "一", "很", "更", "最", "也", "还", "都", "就", "再", "又", "才", "没", "没有", "关于", "以及", "因为", "为了", "由于", "如果", "但是", "一个", "这个", "那个", "今后", "目前", "当时", "现在", "已经", "正在", "需要", "应该", "必须", "可以", "能够", "不能", "有关", "一些"]);
+  const connectors = "但是|可是|不过|然而|同时|而且|并且|因此|所以|此外|另外|于是|但是同时|与此同时|但是也|但是又|同时也|而是|或者|以及";
+  const connectorWords = connectors.split("|").sort((a, b) => b.length - a.length);
+  function connectorOnly(text) {
+    let index = 0, found = false;
+    while (index < text.length) {
+      if (/[，,、：:\s]/u.test(text[index])) { index += 1; continue; }
+      const word = connectorWords.find(value => text.startsWith(value, index));
+      if (!word) return false;
+      found = true; index += word.length;
+    }
+    return found;
+  }
+  const dangling = new Set(["的", "地", "得", "和", "与", "及", "或", "把", "被", "在", "对", "向", "为", "将", "不", "一", "很", "更", "最", "也", "还", "都", "就", "再", "又", "才", "没", "没有", "关于", "以及", "因为", "为了", "由于", "如果", ...connectors.split("|"), "一个", "这个", "那个", "今后", "目前", "当时", "现在", "已经", "正在", "需要", "应该", "必须", "可以", "能够", "不能", "有关", "一些"]);
   function numberSpans(text) {
     const label = "(?:订单编号|訂單編號|订单号|訂單號|编号|編號|手机号|手機號|电话|電話|学号|學號|邮编|郵編|ISBN(?:-13|-10)?|IP)";
     const pattern = new RegExp(label + "[ \\t]*[:：]?[ \\t]*[A-Za-z0-9]+(?:[._/\\-][A-Za-z0-9]+)*" +
@@ -31,14 +42,23 @@ globalThis.CCRDSpeechPlan = (() => {
       const position = offsets.indexOf(start), short = first && !ends.length;
       const target = offsets[Math.min(chars.length, position + (short ? 16 : 90))];
       const ceiling = offsets[Math.min(chars.length, position + (short ? 18 : 120))];
-      const natural = cuts.filter(value => value.end > start + (short ? 5 : 20) && value.end <= ceiling && /[，,。！？!?；;：:\n][”’"'」』）)\t \r]*$/u.test(text.slice(start, value.end)));
+      const hard = cuts.find(value => value.end > start && value.end <= ceiling && /[。！？!?；;\n.][”’"'」』）)\t \r]*$/u.test(text.slice(start, value.end)));
+      const natural = cuts.filter(value => value.end > start + (short ? 5 : 20) && value.end <= ceiling &&
+        /[，,：:][”’"'」』）)\t \r]*$/u.test(text.slice(start, value.end)) && !connectorOnly(text.slice(start, value.end).trim()));
       let end;
-      if (natural.length) end = short ? natural[0].end : natural[natural.length - 1].end;
+      if (hard) end = hard.end; // Never merge different spoken sentences in one part.
+      else if (natural.length) end = short ? natural[0].end : natural[natural.length - 1].end;
       else {
         const choices = cuts.filter(value => value.end > start && value.end <= target && !dangling.has(value.word));
         const any = cuts.filter(value => value.end > start && value.end <= target);
         end = choices.length ? choices[choices.length - 1].end : any.length ? any[any.length - 1].end : cuts.find(value => value.end > start)?.end || text.length;
       }
+      // Keep leading connectives with their predicate, not in a tiny independent
+      // model clip. No source replacement, phoneme gain or global tone changes.
+      if (connectorOnly(text.slice(start, end).trim()) && end < text.length) {
+        end = cuts.find(value => value.end > end && !dangling.has(value.word) && !connectorOnly(text.slice(start, value.end).trim()))?.end || text.length;
+      }
+      while (end < text.length && /[”’"'」』）)\s]/u.test(text[end])) end += 1;
       if (end <= start) throw new Error("无法确定完整词语边界");
       if (end < text.length && !/[\p{L}\p{N}]/u.test(text.slice(end))) end = text.length;
       ends.push(end); start = end;
@@ -77,9 +97,20 @@ globalThis.CCRDSpeechPlan = (() => {
       const finish = boundary(text, end++);
       if (finish) break;
     }
-    while (end < text.length && closers.test(text[end])) end += 1;
+    while (end < text.length && /[”’"'」』）)\s]/u.test(text[end])) end += 1;
     return { start: safe(text, start), end: safe(text, end) };
   }
+  function partIndex(batch, offset) {
+    const relative = offset - batch.start;
+    if (relative === 0) return 0;
+    const previous = batch.segment_ends.indexOf(relative);
+    return previous >= 0 && previous + 1 < batch.segment_ends.length ? previous + 1 : -1;
+  }
+  function previousSentence(text, offset) {
+    let cursor = sentence(text, offset).start - 1;
+    while (cursor > 0 && /[”’"'」』）)\s]/u.test(text[cursor])) cursor -= 1;
+    return sentence(text, Math.max(0, cursor)).start;
+  }
   function batch(text, offset, first = false) {
     const start = first ? wordStart(text, offset) : safe(text, offset), limit = first ? 48 : 480;
     if (start >= text.length) return null;
@@ -106,12 +137,13 @@ globalThis.CCRDSpeechPlan = (() => {
     const candidates = wordEnds(text.slice(start, offsets[Math.min(chars.length, 800)]))
       .filter(value => value.end <= stop - start && value.end > 0 && !dangling.has(value.word));
     if (candidates.length && stop < text.length) stop = start + candidates[candidates.length - 1].end;
-    const body = text.slice(start, stop);
-    return { start, end: stop, text: body, first, segment_ends: phraseEnds(body, first),
+    let body = text.slice(start, stop), ends = phraseEnds(body, first);
+    if (ends.length > 64) { stop = start + ends[63]; body = text.slice(start, stop); ends = ends.slice(0, 64); }
+    return { start, end: stop, text: body, first, segment_ends: ends,
       context_before: Array.from(text.slice(Math.max(0, start - 64), start)).slice(-32).join(""),
       context_after: Array.from(text.slice(stop, stop + 64)).slice(0, 32).join("") };
   }
-  return { safe, sentence, batch, phraseEnds, wordEnds, wordStart, numberSpans };
+  return { safe, sentence, batch, phraseEnds, wordEnds, wordStart, numberSpans, partIndex, previousSentence };
 })();
 
 if (typeof document !== "undefined") window.CCRDSpeech = (() => {
@@ -139,7 +171,12 @@ if (typeof document !== "undefined") window.CCRDSpeech = (() => {
   function clearInline(job) {
     if (job?.inlineUrl) { URL.revokeObjectURL(job.inlineUrl); job.inlineUrl = null; job.inlineConsumed = true; }
   }
+  function remember(value, job) {
+    if (job.state !== "ready" || !job.used) return;
+    value.history = [job.batch, ...value.history.filter(batch => batch.start !== job.batch.start)].slice(0, 2);
+  }
   function discard(value, job) {
+    remember(value, job);
     job.cancelled = true;
     value.jobs.delete(job.batch.start);
     clearInline(job);
@@ -179,8 +216,8 @@ if (typeof document !== "undefined") window.CCRDSpeech = (() => {
     play.textContent = owner?.wantsPlay ? "暂停" : owner?.blocked ? "重试" : "继续";
     play.setAttribute("aria-label", play.textContent + "朗读");
     play.disabled = !owner && !saved;
-    previous.disabled = !owner || owner.position <= 0;
-    next.disabled = !owner || owner.position >= owner.snapshot.text.length;
+    previous.disabled = !owner || plan.sentence(owner.snapshot.text, owner.position).start <= 0;
+    next.disabled = !owner || plan.sentence(owner.snapshot.text, owner.position).end >= owner.snapshot.text.length;
     visible.disabled = !preview;
     locate.disabled = !owner && !saved;
     $("#speech-pitch-value").textContent = (Number(pitch.value) > 0 ? "+" : "") + pitch.value;
@@ -278,7 +315,8 @@ if (typeof document !== "undefined") window.CCRDSpeech = (() => {
   function stop(message = "已停止，可继续朗读", keepalive = false) {
     const value = owner;
     if (value) {
-      saved = { snapshot: value.snapshot, position: value.position };
+      for (const job of value.jobs.values()) remember(value, job);
+      saved = { snapshot: value.snapshot, position: value.position, history: value.history, pitch: value.pitch };
       value.cancelled = true; clearTimeout(value.timer);
       for (const controller of value.controllers) if (!controller.createsJob) controller.abort();
       for (const job of value.jobs.values()) { clearInline(job); release(job.id, keepalive); }
@@ -390,18 +428,27 @@ if (typeof document !== "undefined") window.CCRDSpeech = (() => {
     acceptInline(job, data);
   }
   async function update(value, job) {
-    if (!live(value) || job.cancelled || !value.wantsPlay || job.state === "ready" || job.blocked || Date.now() < job.retryAt || checkSleep()) return;
+    if (!live(value) || job.cancelled || job.pending || !value.wantsPlay || job.state === "ready" || job.blocked || Date.now() < job.retryAt || checkSleep()) return;
+    job.pending = true;
     try {
       if (Date.now() > job.deadline) throw new Error("语音排队超时，点击继续重试");
       const data = job.id ? await request(`/tts/jobs/${job.id}`, {}, value) : await request("/tts/jobs", { method: "POST", headers: { "Content-Type": "application/json" },
         body: JSON.stringify({ text: job.batch.text, voice: "zf_001", speed: 1.2, pitch_semitones: value.pitch,
-          unsupported_policy: "pause", progressive: true, segment_ends: job.batch.segment_ends, wait_first: job.batch.first, inline_first: job.batch.first,
+          unsupported_policy: "pause", progressive: true, segment_ends: job.batch.segment_ends, wait_first: job.batch.first,
+          inline_first: job.batch.first && value.part === 0,
+          cache_only: job.restore === true,
           context_before: job.batch.context_before,
           context_after: job.batch.context_after, request_id: job.nonce }) }, value, job.batch.first ? 22000 : 15000);
       if (!live(value) || job.cancelled || value.jobs.get(job.batch.start) !== job) { release(data.id); return; }
       accept(job, data, value); job.failures = 0; job.retryAt = Date.now() + (job.parts.length ? 650 : 250);
     } catch (error) {
       if (!live(value) || job.cancelled) return;
+      if (job.restore && error.status === 410) {
+        // Metadata history is not an audio cache promise. A miss must NOT queue
+        // a large old batch; create just the requested new foreground sentence.
+        start(value.snapshot, value.position, value.wantsPlay, true);
+        return;
+      }
       job.failures += 1;
       const transient = [404, 410, 429, 503].includes(error.status) || error.name === "AbortError" || error instanceof TypeError;
       job.retryAt = Date.now() + 1000 * (error.retryAfter || Math.min(20, 2 ** job.failures));
@@ -416,7 +463,7 @@ if (typeof document !== "undefined") window.CCRDSpeech = (() => {
         notice(error.message + (job.blocked ? " · 点击继续重试" : " · 稍后重试，进度保留"));
       }
       controls();
-    }
+    } finally { job.pending = false; }
   }
   function advance(value) {
     if (!live(value) || !value.wantsPlay || value.blocked || checkSleep()) return;
@@ -426,6 +473,7 @@ if (typeof document !== "undefined") window.CCRDSpeech = (() => {
     if (!value.current) {
       const part = job.parts[value.part];
       if (part) {
+        job.used = true;
         value.current = { job, part }; value.position = job.batch.start + part.start;
         switching = true; audio.src = part.index === 0 && job.inlineUrl ? job.inlineUrl : ORIGIN + part.audio_url; audio.load(); applyRate();
         notice("朗读中"); mark(); controls();
@@ -442,6 +490,9 @@ if (typeof document !== "undefined") window.CCRDSpeech = (() => {
     // One following API batch and one following immutable MP3 only. No work on
     // pause, no document-wide plan/DOM/audio cache and no whole-book POST.
     if (!finishClip && value.current && job.state === "ready" && job.batch.end < value.snapshot.text.length && !value.jobs.has(job.batch.end)) {
+      // A completed preceding batch can remain usable briefly after a seek, but
+      // never keep a third subscription or add extra speculative inference.
+      for (const other of Array.from(value.jobs.values())) if (other !== job) discard(value, other);
       const batch = plan.batch(value.snapshot.text, job.batch.end);
       value.jobs.set(batch.start, newJob(batch));
     }
@@ -449,30 +500,42 @@ if (typeof document !== "undefined") window.CCRDSpeech = (() => {
     if (following && preloader.getAttribute("src") !== ORIGIN + following.audio_url) preloader.src = ORIGIN + following.audio_url;
   }
   async function pump(value) {
-    if (!live(value) || value.ticking || checkSleep() || !value.wantsPlay) return;
-    clearTimeout(value.timer); value.ticking = true;
+    if (!live(value) || value.pumpEpoch === value.seekEpoch || checkSleep() || !value.wantsPlay) return;
+    const epoch = value.seekEpoch;
+    clearTimeout(value.timer); value.pumpEpoch = epoch;
     try {
       const current = value.jobs.get(value.batch.start);
       if (current) await update(value, current);
+      if (!live(value) || epoch !== value.seekEpoch) return;
       advance(value); // Never wait for speculative work before playing a part.
       if (!live(value) || !value.wantsPlay) return;
       for (const job of Array.from(value.jobs.values())) if (job.batch.start !== current?.batch.start) {
-        await update(value, job); advance(value);
+        await update(value, job);
+        if (!live(value) || epoch !== value.seekEpoch) return;
+        advance(value);
       }
     } finally {
-      value.ticking = false;
-      if (live(value) && value.wantsPlay) value.timer = setTimeout(() => pump(value), value.current ? 600 : 250);
+      if (epoch === value.seekEpoch) {
+        value.pumpEpoch = null;
+        if (live(value) && value.wantsPlay) value.timer = setTimeout(() => pump(value), value.current ? 600 : 250);
+      }
     }
   }
-  function start(snapshot, offset, playing = true) {
+  function start(snapshot, offset, playing = true, skipHistory = false) {
     createAudio();
     const mode = timer.value, deadline = sleepDeadline;
+    if (owner) for (const job of owner.jobs.values()) remember(owner, job);
+    const recent = owner || saved;
+    const history = recent?.snapshot.id === snapshot.id && recent.snapshot.text === snapshot.text && recent.pitch === Number(pitch.value) ? recent.history.slice() : [];
+    const restored = !skipHistory && history.find(batch => plan.partIndex(batch, offset) >= 0);
     stop("正在准备首句…");
-    const batch = plan.batch(snapshot.text, offset, true);
+    const batch = restored || plan.batch(snapshot.text, offset, true);
     if (!batch) { notice("已到正文末尾"); return; }
-    const value = { snapshot, batch, position: batch.start, part: 0, current: null, pitch: Number(pitch.value), wantsPlay: playing,
-      jobs: new Map(), controllers: new Set(), cancelled: false, blocked: false };
-    value.jobs.set(batch.start, newJob(batch)); owner = value; saved = null;
+    const value = { snapshot, batch, position: restored ? offset : batch.start, part: restored ? plan.partIndex(batch, offset) : 0,
+      current: null, pitch: Number(pitch.value), wantsPlay: playing, history,
+      jobs: new Map(), controllers: new Set(), cancelled: false, blocked: false, seekEpoch: 0, pumpEpoch: null };
+    const job = newJob(batch); job.restore = !!restored;
+    value.jobs.set(batch.start, job); owner = value; saved = null;
     timer.value = mode; armSleep();
     if (deadline) { sleepDeadline = deadline; if (checkSleep()) return; }
     $("#speech-title").textContent = snapshot.title || "朗读";
@@ -483,6 +546,7 @@ if (typeof document !== "undefined") window.CCRDSpeech = (() => {
   function fromPreview(offset = visibleOffset()) {
     if (!preview || !preview.text.trim()) { if (typeof showToast === "function") showToast("请等待预览正文加载完成"); return; }
     if (owner && owner.snapshot.id !== preview.id && !confirm("切换到当前预览会停止原朗读，继续吗？")) return;
+    if (owner?.snapshot.id === preview.id && owner.snapshot.text === preview.text && seek(owner, offset, true)) return;
     start({ id: preview.id, text: preview.text, title: preview.title }, offset);
   }
   function resume() {
@@ -503,8 +567,34 @@ if (typeof document !== "undefined") window.CCRDSpeech = (() => {
   function navigate(direction) {
     if (!owner) return;
     const sentence = plan.sentence(owner.snapshot.text, owner.position);
-    const target = direction > 0 ? sentence.end : plan.sentence(owner.snapshot.text, Math.max(0, sentence.start - 1)).start;
-    if (target < owner.snapshot.text.length) start(owner.snapshot, target, owner.wantsPlay);
+    const target = direction > 0 ? sentence.end : plan.previousSentence(owner.snapshot.text, owner.position);
+    if (target < owner.snapshot.text.length && !seek(owner, target, owner.wantsPlay)) start(owner.snapshot, target, owner.wantsPlay);
+  }
+  function seek(value, offset, playing) {
+    if (!live(value) || value.settingsChanged || value.pitch !== Number(pitch.value)) return false;
+    const job = Array.from(value.jobs.values()).find(item => !item.cancelled && !item.blocked && plan.partIndex(item.batch, offset) >= 0);
+    if (!job) return false;
+    const index = plan.partIndex(job.batch, offset), current = value.current;
+    value.seekEpoch += 1; clearTimeout(value.timer); clearPreload();
+    value.position = offset; value.batch = job.batch; value.part = index; value.wantsPlay = playing;
+    value.blocked = false;
+    // Do not let synthesis of a now irrelevant batch hold up the target job.
+    for (const other of Array.from(value.jobs.values())) {
+      if (other !== job && other.batch.start !== job.batch.end && other.state !== "ready") discard(value, other);
+      else if (!playing && other !== job && other.state !== "ready") discard(value, other);
+    }
+    if (current?.part.index === 0 && (current.job !== job || current.part.index !== index)) clearInline(current.job);
+    if (current?.job === job && current.part.index === index && audio.readyState >= 1 && !audio.error) {
+      audio.currentTime = 0;
+      if (playing) playAudio(value); else audio.pause();
+    } else {
+      value.current = null; switching = true; audio.pause(); audio.removeAttribute("src"); audio.load();
+      advance(value);
+    }
+    show(); mark(); controls();
+    notice(playing ? (value.current ? "朗读中" : "正在准备所选句…") : "已选择句子，点击继续播放");
+    pump(value);
+    return true;
   }
   function previewLoaded(value) {
     clearMark();

tests/test_tts_browser.py

@@ -34,10 +34,11 @@ class SpeechReaderBrowserTests(unittest.TestCase):
                         context = browser.new_context(viewport={"width": 390, "height": 844} if mobile else {"width": 1440, "height": 900}, reduced_motion="reduce")
                         context.add_init_script("delete Navigator.prototype.serviceWorker;new MutationObserver(()=>{const a=document.querySelector('#speech-audio');if(a)a.muted=true;}).observe(document,{childList:true,subtree:true});")
                         context.add_init_script("window.__sourceRange=(body,offset)=>{const w=document.createTreeWalker(body,NodeFilter.SHOW_TEXT);let n;while(n=w.nextNode()){if(offset<n.length){const r=document.createRange();r.setStart(n,offset);r.setEnd(n,offset+1);return r;}offset-=n.length;}};")
+                        context.add_init_script("window.__spokenText=()=>document.querySelector('.speech-current')?.textContent||Array.from(CSS.highlights?.get('ccrd-speech')||[])[0]?.toString()||'';")
                         context.add_init_script("window.__audioBlobs=new Set();window.__blobCount=0;const make=URL.createObjectURL.bind(URL),drop=URL.revokeObjectURL.bind(URL);URL.createObjectURL=blob=>{const url=make(blob);__audioBlobs.add(url);__blobCount++;return url;};URL.revokeObjectURL=url=>{__audioBlobs.delete(url);drop(url);};")
                         if mobile:
                             context.add_init_script("Object.defineProperty(CSS,'highlights',{value:null,configurable:true});")
-                        posts, deleted, paths, jobs, errors, held = [], [], [], {}, [], []
+                        posts, deleted, paths, jobs, errors, held, cached = [], [], [], {}, [], [], set()
                         hold_next, offline, missing, bad = [False], [False], [False], [False]
                         inline_mode = ["valid"]
                         headers = {"Access-Control-Allow-Origin": SEARCH, "Access-Control-Allow-Headers": "Content-Type,Range", "Access-Control-Allow-Methods": "GET,POST,DELETE,OPTIONS"}
@@ -69,13 +70,19 @@ class SpeechReaderBrowserTests(unittest.TestCase):
                                 if path == "/tts/jobs" and request.method == "POST":
                                     body = request.post_data_json
                                     posts.append(body)
+                                    key = json.dumps({name: body.get(name) for name in ("text", "voice", "speed", "pitch_semitones", "segment_ends", "context_before", "context_after", "unsupported_policy")}, sort_keys=True)
+                                    hit = key in cached
+                                    if body.get("cache_only") and not hit:
+                                        reply(route, {"error": {"code": "audio_expired", "message": "已完成的句段缓存已过期"}}, 410)
+                                        return
+                                    cached.add(key)
                                     id = f"{len(posts):032d}"
                                     length = len(body["text"].encode("utf-16-le")) // 2
                                     middle = len(body["text"][:20].encode("utf-16-le")) // 2
                                     spans = list(zip([0] + body["segment_ends"][:-1], body["segment_ends"]))
                                     parts = [{"index": index, "start": begin, "end": end, "audio_url": f"/tts/audio/{id}/parts/{index}.mp3", "pitch_semitones": body["pitch_semitones"], "bytes": len(mp3)} for index, (begin, end) in enumerate(spans)]
                                     if bad[0]: parts[0]["audio_url"] = "https://untrusted.test/leak"
-                                    jobs[id] = {"id": id, "state": "ready", "source_segments": [{"start": begin, "end": end} for begin, end in spans],
+                                    jobs[id] = {"id": id, "state": "ready", "cache_hit": hit, "source_segments": [{"start": begin, "end": end} for begin, end in spans],
                                                 "progress": {"total_parts": len(parts), "parts": parts}, "result": {"pitch_semitones": body["pitch_semitones"]}}
                                     if body.get("inline_first") and inline_mode[0] != "omit":
                                         jobs[id]["first_audio"] = {"index": 0, "encoding": "base64", "content_type": "audio/mpeg", "bytes": len(mp3),
@@ -194,14 +201,48 @@ class SpeechReaderBrowserTests(unittest.TestCase):
                             self.assertGreater(len(posts), before)
                             self.assertTrue(posts[before]["text"].startswith("第0000段"))
                             page.wait_for_function("!document.querySelector('#speech-audio').paused", polling=100)
+                            page.wait_for_timeout(300)
                             before = len(posts)
                             page.locator("#speech-next").click()
-                            page.wait_for_timeout(300)
-                            self.assertTrue(posts[before]["text"].startswith("𠮷※😀"), posts[before]["text"])
+                            page.wait_for_function("__spokenText().startsWith('𠮷※��')&&!document.querySelector('#speech-audio').paused")
+                            self.assertEqual(len(posts), before, "Next must reuse the already-generated job rather than resubmit")
                             before = len(posts)
                             page.locator("#speech-previous").click()
+                            page.wait_for_function("__spokenText().startsWith('第0000段')&&!document.querySelector('#speech-audio').paused")
+                            self.assertEqual(len(posts), before, "Previous must reuse the already-generated job rather than resubmit")
+                            page.mouse.dblclick(box["x"], box["y"])
+                            page.wait_for_function("__spokenText().startsWith('第0000段')&&!document.querySelector('#speech-audio').paused")
+                            self.assertEqual(len(posts), before, "Double-click within buffered source must not regenerate")
+                            # Crossing the two-job window must restore the exact
+                            # completed batch as cache-only, not new inference.
+                            page.locator("#speech-next").click()
+                            page.wait_for_function("__spokenText().startsWith('𠮷※😀')")
+                            page.locator("#speech-next").click()
+                            page.wait_for_function("__spokenText().startsWith('第0001段')&&!document.querySelector('#speech-audio').paused")
                             page.wait_for_timeout(300)
-                            self.assertTrue(posts[before]["text"].startswith("第0000段"), posts[before]["text"])
+                            before = len(posts)
+                            page.locator("#speech-previous").click()
+                            page.wait_for_function("__spokenText().startsWith('𠮷※😀')&&!document.querySelector('#speech-audio').paused")
+                            restored = [row for row in posts[before:] if row.get("cache_only")]
+                            self.assertEqual(len(restored), 1)
+                            self.assertTrue(restored[0]["text"].startswith("第0000段"))
+                            page.locator("#speech-previous").click()
+                            page.wait_for_function("__spokenText().startsWith('第0000段')&&!document.querySelector('#speech-audio').paused")
+                            # Eviction must cause only one cache lookup miss,
+                            # then new synthesis starts at the target source.
+                            cached.clear()
+                            page.locator("#speech-next").click()
+                            page.wait_for_function("__spokenText().startsWith('𠮷※😀')")
+                            page.locator("#speech-next").click()
+                            page.wait_for_function("__spokenText().startsWith('第0001段')&&!document.querySelector('#speech-audio').paused")
+                            page.wait_for_timeout(200)
+                            before = len(posts)
+                            page.locator("#speech-previous").click()
+                            page.wait_for_function("__spokenText().startsWith('𠮷※😀')&&!document.querySelector('#speech-audio').paused")
+                            self.assertEqual(len([row for row in posts[before:] if row.get("cache_only")]), 1)
+                            self.assertTrue(any(not row.get("cache_only") and row["text"].startswith("𠮷※😀") for row in posts[before:]))
+                            page.locator("#speech-previous").click()
+                            page.wait_for_function("__spokenText().startsWith('第0000段')&&!document.querySelector('#speech-audio').paused")
                             # Pitch changes regenerate only from the current
                             # source clip, while speed remains instantaneous.
                             before = len(posts)
@@ -260,8 +301,9 @@ class SpeechReaderBrowserTests(unittest.TestCase):
                             self.assertEqual(page.evaluate("__audioBlobs.size"), 0)
                             # Optional delivery can be absent or malformed: use
                             # only the canonical GET, without skipping source.
-                            for mode in ("omit", "invalid", "decode-error", "paused-error"):
+                            for index, mode in enumerate(("omit", "invalid", "decode-error", "paused-error")):
                                 inline_mode[0] = mode
+                                page.evaluate("id=>openPreview(id,'',{scroll:false})", "CCRD:1" if index % 2 == 0 else "CCRD:0")
                                 first_path = f"/tts/audio/{len(posts) + 1:032d}/parts/0.mp3"
                                 page.locator("[data-speech-toggle]").click()
                                 page.wait_for_function("!document.querySelector('#speech-audio').paused", polling=100)

tests/test_tts_plan.cjs

@@ -40,4 +40,28 @@ for (const source of ["中共中央批转文化部党组关于今后文化艺术
     offset=part.end;
   }
 }
+for (const source of ["第一句。第二句继续说。第三句接着说。".repeat(50), "第一句。\n\n“第二句继续说。”  第三句接着说。".repeat(20)]) {
+  let offset = 0;
+  while (offset < source.length) {
+    const batch = plan.batch(source, offset, offset === 0);
+    let start = batch.start;
+    for (const end of batch.segment_ends) {
+      assert(plan.sentence(source, start).end >= batch.start + end, 'a clip must not merge separate sentences');
+      assert.equal(plan.partIndex(batch, start), batch.segment_ends.indexOf(end));
+      start = batch.start + end;
+    }
+    offset = batch.end;
+  }
+}
+const spaced = "第一句。\n\n第二句。  第三句。";
+assert.equal(plan.previousSentence(spaced, spaced.indexOf('第三句')), spaced.indexOf('第二句'));
+assert.equal(plan.previousSentence(spaced, spaced.indexOf('第二句')), 0);
+for (const source of ['但是，我们需要继续检查。', '同时，也要把后文读完整。', '但是，同时，我们仍然应该继续检查。']) {
+  const batch=plan.batch(source,0,true);
+  let start=0;
+  for (const end of batch.segment_ends) {
+    assert(!/^(?:但是|同时|[，,、\s])+$/u.test(batch.text.slice(start,end)), 'connectives must not become isolated synthesis clips');
+    start=end;
+  }
+}
 console.log("Rolling source plan: lossless >20k, non-BMP, numeric atoms, bounded batches/context PASS");