perf: cache Kokoro TTS model in service worker

Add CacheFirst runtimeCaching rules for Hugging Face model requests
(huggingface.co and cdn-lfs) with 90-day expiration. Add
getKokoroProgress() export to tts.ts for tracking download progress.
Add TTS model download progress bar overlay in FightViewer.vue that
shows during model loading and auto-hides when complete.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

frontend/src/components/FightViewer.vue

@@ -12,6 +12,7 @@ import {
   prefetchQuestion, prefetchAnswer, prefetchNarration,
   sfxRandomComedy, sfxRandomFail, sfxVineBoom, sfxEmotionalDamage,
 } from '../game/sounds'
+import { isKokoroLoading, getKokoroProgress } from '../game/tts'
 
 interface Round {
   roundNumber: number
@@ -53,6 +54,8 @@ let cleanupTimerHandle: ReturnType<typeof setTimeout> | null = null
 let destroyed = false
 
 const isReplaying = ref(false)
+const ttsProgress = ref(-1)
+let ttsProgressTimer: ReturnType<typeof setInterval> | null = null
 const displayHpA = ref(100)
 const displayHpB = ref(100)
 const currentRound = ref(0)
@@ -181,6 +184,7 @@ onUnmounted(() => {
   sceneReady.value = false
   if (scene) { scene.destroy(); scene = null }
   if (cleanupTimerHandle) { clearTimeout(cleanupTimerHandle); cleanupTimerHandle = null }
+  if (ttsProgressTimer) { clearInterval(ttsProgressTimer); ttsProgressTimer = null }
   if (typeof speechSynthesis !== 'undefined') speechSynthesis.cancel()
 })
 
@@ -321,10 +325,18 @@ async function replay() {
   if (isReplaying.value || !props.fight.botA || !props.fight.botB) return
   isReplaying.value = true
   showingFinal.value = false
+  // Poll TTS model download progress while loading
+  ttsProgressTimer = setInterval(() => {
+    ttsProgress.value = isKokoroLoading() ? getKokoroProgress() : -1
+  }, 200)
   try { await _doReplay() } catch (e) {
     if (e instanceof Error && e.message === 'unmounted') return
     console.error('[FightViewer] replay error:', e)
-  } finally { isReplaying.value = false }
+  } finally {
+    isReplaying.value = false
+    if (ttsProgressTimer) { clearInterval(ttsProgressTimer); ttsProgressTimer = null }
+    ttsProgress.value = -1
+  }
 }
 
 async function _doReplay() {
@@ -361,10 +373,18 @@ async function _doReplay() {
   for (const round of props.fight.rounds) {
     currentRound.value = round.roundNumber
 
+    // Parse challenge data EARLY so we can prefetch TTS during overlays
+    const challenge = JSON.parse(round.challengeData)
+    const doTTS = soundOn.value && !insanityMode.value
+    const questionText = challenge.displayPrompt || challenge.prompt
+
     if (insanityMode.value) {
       // Insanity: minimal overlays, no fanfares
       await showOverlay(`R${round.roundNumber}`, '#00f0ff', 150)
     } else {
+      // Prefetch question audio NOW — worker generates while overlays play (~1.8s)
+      if (doTTS && questionText) prefetchQuestion(questionText)
+
       fanfareRound(round.roundNumber)
       await showOverlay(`ROUND ${round.roundNumber}`, '#00f0ff', 700)
       await sleep(80)
@@ -375,27 +395,25 @@ async function _doReplay() {
       await sleep(80)
     }
 
-    // === TTS-synced question + answer flow ===
-    const challenge = JSON.parse(round.challengeData)
-    const doTTS = soundOn.value && !insanityMode.value
-
     // Both fighters showboat while the question is being asked
     if (scene && !insanityMode.value) {
       scene.startShowboating('a')
       scene.startShowboating('b')
     }
 
-    // 1. Show question in log AND speak it
+    // 1. Show question in log AND speak it (audio should be pre-generated by now)
     logItems.value.push(
       { type: 'header', round: round.roundNumber, text: `ROUND ${round.roundNumber}: ${challengeLabel(round.challengeType)}`, color: 'neon-purple' },
     )
     scrollLog(); await sleep(insanityMode.value ? 30 : 100)
     logItems.value.push(
-      { type: 'prompt', round: round.roundNumber, text: challenge.displayPrompt || challenge.prompt, color: 'text-muted' },
+      { type: 'prompt', round: round.roundNumber, text: questionText, color: 'text-muted' },
     )
     scrollLog()
-    if (doTTS && (challenge.displayPrompt || challenge.prompt)) {
+    if (doTTS && questionText) {
-      await speakQuestion(challenge.displayPrompt || challenge.prompt)
+      // Prefetch bot A answer while question plays
+      if (round.botAResponse) prefetchAnswer(props.fight.botA!.name, round.botAResponse)
+      await speakQuestion(questionText)
       await sleep(150)
     } else {
       await sleep(insanityMode.value ? 50 : 600)
@@ -412,6 +430,8 @@ async function _doReplay() {
         scene.showSpeechBubble('a', round.botAResponse.slice(0, 60), 5)
         scene.startTalking('a')
       }
+      // Prefetch bot B answer while bot A talks
+      if (doTTS && round.botBResponse) prefetchAnswer(props.fight.botB!.name, round.botBResponse)
       if (doTTS) await speakAnswer(props.fight.botA!.name, round.botAResponse)
       else await sleep(insanityMode.value ? 30 : 800)
       scene?.stopTalking('a')
@@ -429,6 +449,8 @@ async function _doReplay() {
         scene.showSpeechBubble('b', round.botBResponse.slice(0, 60), 5)
         scene.startTalking('b')
       }
+      // Prefetch narration while bot B talks
+      if (doTTS && round.narration) prefetchNarration(round.narration)
       if (doTTS) await speakAnswer(props.fight.botB!.name, round.botBResponse)
       else await sleep(insanityMode.value ? 30 : 800)
       scene?.stopTalking('b')
@@ -705,6 +727,15 @@ async function _doReplay() {
       <!-- Canvas + floating overlays -->
       <div ref="canvasContainer" class="flex-1 relative min-h-0" :class="{ 'glitch-container': glitching }">
         <canvas ref="canvasRef" class="w-full h-full block" />
+        <!-- TTS model download progress -->
+        <div v-if="ttsProgress >= 0 && ttsProgress < 100" class="absolute bottom-1 left-2 right-2 z-30">
+          <div class="bg-black/60 rounded px-2 py-1 flex items-center gap-2">
+            <span class="text-[10px] text-text-muted font-mono whitespace-nowrap">TTS {{ Math.round(ttsProgress) }}%</span>
+            <div class="flex-1 h-1 bg-white/10 rounded overflow-hidden">
+              <div class="h-full bg-neon-green transition-all duration-200" :style="{ width: ttsProgress + '%' }" />
+            </div>
+          </div>
+        </div>
 
         <!-- Floating announcement -->
         <Transition name="announce">

frontend/src/game/tts.ts

@@ -121,9 +121,12 @@ const _pending = new Map<number, {
 // --- Audio state (main thread only) ---
 let _audioCtx: AudioContext | null = null
 const audioCache = new Map<string, AudioBuffer>()
-const MAX_CACHE = 30
+const MAX_CACHE = 50
 const activeSources: Set<AudioBufferSourceNode> = new Set()
 
+// In-flight generation dedup: cache key → promise (prevents duplicate worker calls)
+const _inflight = new Map<string, Promise<AudioBuffer | null>>()
+
 function getAudioCtx(): AudioContext {
   if (!_audioCtx) _audioCtx = new AudioContext()
   if (_audioCtx.state === 'suspended') _audioCtx.resume().catch(() => {})
@@ -145,6 +148,12 @@ export function isKokoroLoading(): boolean {
   return _workerLoading
 }
 
+/** Current model download progress (0-100), or -1 if not loading */
+let _loadProgress = -1
+export function getKokoroProgress(): number {
+  return _loadProgress
+}
+
 /** Handle messages from the TTS worker */
 function _handleWorkerMessage(e: MessageEvent) {
   const msg = e.data
@@ -177,13 +186,15 @@ export async function initKokoro(onProgress?: (pct: number) => void): Promise<vo
         if (msg.type === 'init-done') {
           _workerReady = true
           _workerLoading = false
+          _loadProgress = 100
           // Switch to persistent handler for generate responses
           _worker!.onmessage = _handleWorkerMessage
           resolve()
         } else if (msg.type === 'init-failed') {
           reject(new Error(msg.error))
-        } else if (msg.type === 'progress' && onProgress) {
+        } else if (msg.type === 'progress') {
-          onProgress(msg.progress)
+          _loadProgress = msg.progress
+          if (onProgress) onProgress(msg.progress)
         }
       }
       _worker!.onerror = (e) => {
@@ -202,11 +213,18 @@ export async function initKokoro(onProgress?: (pct: number) => void): Promise<vo
   }
 }
 
-// Common phrases to pre-generate so they play instantly
+// Common phrases to pre-generate so they play instantly.
+// Must match EXACT text sent by fanfareRound/fanfareFight/announce calls.
 const PRECACHE_PHRASES: Array<{ text: string; profile: string }> = [
-  { text: 'Round one!', profile: 'announcer' },
+  { text: 'Round 1', profile: 'announcer' },
+  { text: 'Round 2', profile: 'announcer' },
+  { text: 'Round 3', profile: 'announcer' },
+  { text: 'Round 4', profile: 'announcer' },
+  { text: 'Round 5', profile: 'announcer' },
   { text: 'Fight!', profile: 'announcer' },
   { text: 'K. O.!', profile: 'announcer' },
+  { text: 'Finish it!', profile: 'announcer' },
+  { text: 'Flawless victory!', profile: 'deep' },
 ]
 
 async function _precacheCommon() {
@@ -249,10 +267,23 @@ async function _generateAndCache(text: string, profile: string): Promise<AudioBu
   const cached = audioCache.get(key)
   if (cached) return cached
 
+  // Dedup: if already generating this exact audio, reuse the in-flight promise
+  const existing = _inflight.get(key)
+  if (existing) return existing
+
+  const promise = _doGenerate(text, profile, key)
+  _inflight.set(key, promise)
+  try {
+    return await promise
+  } finally {
+    _inflight.delete(key)
+  }
+}
+
+async function _doGenerate(text: string, profile: string, key: string): Promise<AudioBuffer | null> {
   const mapping = VOICE_MAP[profile] || DEFAULT_VOICE
   const raw = await _workerGenerate(text, mapping.voice, mapping.speed)
 
-  // Convert Float32Array → AudioBuffer (lightweight, main thread)
   const ctx = getAudioCtx()
   const buf = ctx.createBuffer(1, raw.audio.length, raw.sampleRate)
   buf.getChannelData(0).set(raw.audio)
@@ -334,6 +365,12 @@ export function kokoroSpeak(
   return true
 }
 
+/** Pre-generate audio in the worker so it's cached when needed. Fire-and-forget. */
+export function kokoroPrefetch(text: string, profileName: string): void {
+  if (!_workerReady) return
+  _generateAndCache(text, profileName).catch(() => {})
+}
+
 /** Stop all currently playing kokoro audio */
 export function kokoroStop() {
   for (const src of activeSources) {

frontend/vite.config.ts

@@ -58,6 +58,16 @@ export default defineConfig({
             method: 'GET',
             options: { cacheName: 'api-cache' },
           },
+          {
+            urlPattern: /^https:\/\/huggingface\.co\/.*Kokoro.*\.onnx/i,
+            handler: 'CacheFirst',
+            options: { cacheName: 'kokoro-model', expiration: { maxEntries: 10, maxAgeSeconds: 60 * 60 * 24 * 90 } },
+          },
+          {
+            urlPattern: /^https:\/\/cdn-lfs.*\.huggingface\.co\/.*/i,
+            handler: 'CacheFirst',
+            options: { cacheName: 'kokoro-model-lfs', expiration: { maxEntries: 20, maxAgeSeconds: 60 * 60 * 24 * 90 } },
+          },
         ],
       },
     }),