dsh-creation-mode 3.0.0 → 3.0.1

package/client.js

@@ -107,6 +107,37 @@ window.__ModuleLoader__.load({
 .cas-lightbox{position:fixed;inset:0;z-index:9100;background:rgba(4,5,8,.88);display:flex;align-items:center;
   justify-content:center;padding:32px;cursor:zoom-out}
 .cas-lightbox img{max-width:100%;max-height:100%;border-radius:10px;box-shadow:0 20px 60px rgba(0,0,0,.6)}
+
+/* 样式库：会话里的第三个页签（对话 / 轨迹 / 样式库） */
+.cas-lib-view{display:flex;flex-direction:column;gap:10px;padding:16px;min-height:0}
+.cas-lib-head{display:flex;align-items:center;gap:10px}
+.cas-lib-title{font-weight:600;font-size:14px}
+.cas-lib-count{font-size:11px;color:var(--dsw-alias-label-secondary);white-space:nowrap}
+.cas-lib-search{flex:1;height:30px;padding:0 10px;border-radius:8px;outline:none;min-width:0;
+  border:1px solid var(--dsw-alias-border-l1);background:var(--dsw-alias-bg-layer-2);
+  color:var(--dsw-alias-label-primary);font:13px/1 inherit}
+.cas-lib-search:focus{border-color:var(--dsw-alias-border-l2)}
+.cas-lib-chips{display:flex;flex-wrap:wrap;gap:6px}
+.cas-chip{height:24px;padding:0 9px;border-radius:999px;cursor:pointer;
+  border:1px solid var(--dsw-alias-border-l1);background:transparent;
+  color:var(--dsw-alias-label-secondary);font:11.5px/1 inherit}
+.cas-chip:hover{border-color:var(--dsw-alias-border-l2)}
+.cas-chip[data-on="1"]{background:var(--dsw-alias-label-primary);border-color:transparent;
+  color:var(--dsw-alias-bg-base)}
+.cas-lib-body{overflow:auto;min-height:240px;max-height:calc(100vh - 260px)}
+.cas-lib-grid{display:grid;grid-template-columns:repeat(auto-fill,minmax(200px,1fr));gap:10px}
+.cas-lib-card{border:1px solid var(--dsw-alias-border-l1);border-radius:9px;overflow:hidden;
+  background:var(--dsw-alias-bg-layer-1);padding:0;text-align:left;cursor:zoom-in;
+  color:inherit;font:inherit}
+.cas-lib-card:hover{border-color:var(--dsw-alias-border-l2)}
+.cas-lib-card img{width:100%;display:block;aspect-ratio:16/9;object-fit:cover;
+  background:var(--dsw-alias-bg-layer-2)}
+.cas-lib-cap{padding:7px 9px}
+.cas-lib-id{font-size:10px;font-weight:600;letter-spacing:.5px;color:var(--dsw-alias-label-secondary)}
+.cas-lib-name{margin-left:6px;font-size:12.5px}
+.cas-lib-desc{display:block;margin-top:3px;font-size:11px;color:var(--dsw-alias-label-secondary);
+  overflow:hidden;text-overflow:ellipsis;white-space:nowrap}
+.cas-lib-empty{padding:48px 0;text-align:center;font-size:13px;color:var(--dsw-alias-label-secondary)}
 `;
 
     const PALETTE = ['#ff8a3d', '#e0457b', '#7b5bff', '#3fd7c0', '#4d6bfe', '#ffc04d'];
@@ -123,1244 +154,42 @@ window.__ModuleLoader__.load({
     }
 
     /**
- * Real-time, cue-accurate score synthesis.
- *
- * Two problems have to be solved together for a cut to feel like a 卡点:
- *
- *   1. The cut must sit on a strong musical position. That is a tempo-and-phase
- *      problem: pick the beat period that puts every cut on a bar line, not merely
- *      "near a beat". (This is the same thing `web-audio-beat-detector` solves when it
- *      jointly estimates tempo and first-beat offset — here it is inverted, because the
- *      cut times are known up front and the music is composed to them.)
- *
- *   2. The bar has to be an actual groove. A drone with occasional booms does not read
- *      as an accent, because there is no pulse for the accent to break. So every bar
- *      carries kick, snare, hats and bass, and the cut lands where the ear expects the
- *      downbeat.
- *
- * On top of that, the arrangement sidechains under each cue so the impact lands in a
- * gap rather than on top of an equally loud bed.
- *
- * Everything is synthesised with Web Audio through `OfflineAudioContext`, so there are
- * no bundled assets and no network.
- */
-
-const MUSIC_PRESETS = [
-  { id: 'auto', label: '智能匹配分镜卡点' },
-  { id: 'guofeng', label: '国风水墨韵律' },
-  { id: 'epic', label: '史诗宏大震撼' },
-  { id: 'synth', label: '赛博未来科技' },
-  { id: 'pulse', label: '快节奏动感脉冲' },
-  { id: 'bright', label: '活泼轻快灵动' },
-  { id: 'chiptune', label: '8-bit 芯片音乐' },
-  { id: 'ambient', label: '沉浸深邃氛围' },
-  { id: 'none', label: '无配乐' },
-];
-
-const SCALES = {
-  pentatonic: [0, 2, 4, 7, 9],
-  minorPent: [0, 3, 5, 7, 10],
-  minor: [0, 2, 3, 5, 7, 8, 10],
-  major: [0, 2, 4, 5, 7, 9, 11],
-  dorian: [0, 2, 3, 5, 7, 9, 10],
-  phrygian: [0, 1, 3, 5, 7, 8, 10],
-};
-
-/**
- * Bar patterns, sixteen steps each: one character per sixteenth note.
- *
- * `K`/`k` kick (accent/soft), `X` kick and snare together, `S` snare,
- * `H`/`h` hat (accent/soft), `o` open hat, `.` silence.
- *
- * `four` is four-on-the-floor (a kick on every beat), `backbeat` is kick on 1 and 3
- * with the snare on 2 and 4, `sparse` leaves air for an arrival. Each family holds
- * several variants because repeating one bar verbatim is what made every score read
- * as the same loop; consecutive bars rotate through them.
- */
-const GROOVE_PATTERNS = {
-  four: [
-    'K.h.X.h.K.h.X.h.',
-    'K.hhX.hhK.hhX.hh',
-    'K.h.X.h.K.h.X.oH',
-    'K.h.X.hhK.h.X.hh',
-  ],
-  backbeat: [
-    'K.h.S.h.K.h.S.h.',
-    'K..hS..hK..hS..h',
-    'K.hhS.hhK.hhS.hh',
-    'K.h.S.h.K.h.S.oH',
-  ],
-  sparse: [
-    'K.......o.......',
-    'K.....k.o.......',
-    'K.......h...S...',
-    'K...k...o...S...',
-  ],
-};
-
-/**
- * One instrument set per mood.
- *
- * Every preset used to share a single electronic kit — the same sine kick, the same
- * noise snare, the same saw pluck — so changing mood moved only the scale, tempo and
- * reverb and every score still sounded like the same track. These sets change the
- * instrument families themselves: a taiko and a plucked zither for 国风, an orchestra
- * for 史诗, an 808/909 machine for 赛博, a marimba for 轻快, an FM bell for 氛围.
- *
- * Voice parameters are read by the primitives below; `null` drops the voice entirely.
- */
-/**
- * Instruments, each built by a different synthesis method.
- *
- * A parameter sweep over one oscillator cannot make a zither, a flute and a piano —
- * they differ in *how* they make sound, not in cutoff and decay. Each entry here names
- * the method its voice uses:
- *
- *   pluck  filtered oscillator stack, exponential decay, optional pitch bend
- *   struck additive partials with hammer noise; higher partials decay faster (piano)
- *   fm     carrier ring-modulated by a ratio (marimba, bell, e-piano)
- *   bowed  detuned saw with bow noise, vibrato and a formant peak (violin, cello, brass)
- *   wind   sine/triangle with breath noise and vibrato (笛子)
- *   chip   unfiltered pulse wave by duty cycle (NES/8-bit)
- */
-const INSTRUMENTS = {
-  // ---- 中国乐器 ----
-  zheng: {
-    synth: 'pluck', wave: 'triangle', attack: 0.0015, c0: 3600, c1: 520, bend: -0.5, decayScale: 0.9,
-    // 古筝的泛音列比钢琴更亮、更不谐和
-    partials: [{ wave: 'sine', octave: 12, gain: 0.40 }, { wave: 'sine', octave: 19, gain: 0.16 }, { wave: 'triangle', octave: 24, gain: 0.08 }],
-  },
-  dizi: {
-    synth: 'wind', wave: 'sine', attack: 0.05, decayScale: 1.7, sustain: true,
-    breath: 0.20, breathF: 2600, air: 0.05, vibrato: 5.6, vibratoDepth: 0.013, c0: 2600, c1: 2000, level: 0.85,
-  },
-  // ---- 西洋乐器 ----
-  piano: {
-    synth: 'struck', attack: 0.003, decayScale: 1.5, hammer: 0.09, hammerF: 3200,
-    // 非谐泛音 + 高次泛音衰减更快：这是钢琴与拨弦最大的区别
-    partials: [
-      { octave: 0, gain: 1 }, { octave: 12, gain: 0.34 }, { octave: 19, gain: 0.20 },
-      { octave: 24, gain: 0.10 }, { octave: 28, gain: 0.05 },
-    ],
-    c0: 2800, c1: 700,
-  },
-  violin: {
-    synth: 'bowed', wave: 'sawtooth', attack: 0.075, decayScale: 1.8, sustain: true,
-    bow: 0.09, bowF: 3000, vibrato: 5.8, vibratoDepth: 0.010, formant: 900, c0: 3200, c1: 2400, level: 0.9,
-  },
-  cello: {
-    synth: 'bowed', wave: 'sawtooth', attack: 0.11, decayScale: 2.0, sustain: true,
-    bow: 0.07, bowF: 1800, vibrato: 4.6, vibratoDepth: 0.008, formant: 520, c0: 1700, c1: 1300, level: 1.0,
-  },
-  brass: {
-    synth: 'bowed', wave: 'sawtooth', attack: 0.055, decayScale: 1.9, sustain: false,
-    bow: 0.02, bowF: 1500, vibrato: 5.2, vibratoDepth: 0.005, formant: 1200, c0: 2200, c1: 900, level: 1.4,
-  },
-  guitar: {
-    synth: 'pluck', wave: 'triangle', attack: 0.004, c0: 2600, c1: 700, decayScale: 1.0,
-    partials: [{ wave: 'sine', octave: 12, gain: 0.30 }, { wave: 'sine', octave: 19, gain: 0.12 }],
-  },
-  // ---- 打击与键盘 ----
-  marimba: { synth: 'fm', ratio: 4, index: 1.6, attack: 0.003, decayScale: 0.8, level: 1.25 },
-  bell: { synth: 'fm', ratio: 3.5, index: 1.4, attack: 0.006, decayScale: 2.6, level: 1.25 },
-  ePiano: { synth: 'fm', ratio: 2, index: 1.1, attack: 0.004, decayScale: 1.6, level: 1.2 },
-  // ---- 电子与 8-bit ----
-  acid: { synth: 'pluck', wave: 'square', attack: 0.002, c0: 4200, c1: 900, decayScale: 0.6, q: 7 },
-  stab: { synth: 'pluck', wave: 'sawtooth', attack: 0.001, c0: 5200, c1: 1400, decayScale: 0.35 },
-  padSynth: { synth: 'pluck', wave: 'sawtooth', attack: 0.01, c0: 2000, c1: 800, decayScale: 1.2 },
-  chipPulse: { synth: 'chip', duty: 0.25, attack: 0.001, decayScale: 0.55 },
-  chipSquare: { synth: 'chip', duty: 0.5, attack: 0.001, decayScale: 0.65 },
-  chipTri: { synth: 'chipTri', attack: 0.002, decayScale: 0.8 },
-};
-
-/**
- * One kit per mood: which instruments play, plus the drum design.
- *
- * `lead` is the melody line; `answer` is an optional second instrument that takes
- * every other melodic step, which is what turns a single line into an exchange
- * (古筝 问答 笛子) instead of the same timbre repeating.
- */
-const KITS = {
-  guofeng: {
-    trim: 1.997, // 响度校准
-    // 太鼓：鼓皮噪声为主
-    kick: { f0: 96, f1: 52, sweep: 0.12, dur: 0.55, peak: 0.34, tone: 0.55, thump: 0.36, thumpDur: 0.10, thumpF0: 1400, thumpF1: 160, thumpType: 'bandpass', thumpQ: 0.8 },
-    // 梆子：极短高 Q 木质
-    snare: { kind: 'wood', type: 'bandpass', f0: 1700, q: 4.5, dur: 0.075, peak: 0.20, attack: 0.0008, body: 430, bodyLevel: 0.10, bodyDur: 0.07 },
-    hat: { type: 'highpass', f0: 4600, dur: 0.05, peak: 0.07, attack: 0.004 },
-    lead: 'zheng',
-    answer: 'dizi',
-    bass: { wave: 'sine', attack: 0.03, sustain: true, release: 0.16, level: 0.30, cutoff: 320 },
-    pad: { voices: [{ wave: 'sine', detune: -5 }, { wave: 'triangle', detune: 5 }], attack: 0.65, sustain: true, release: 0.3, cutoff: 700, level: 0.18 },
-    riser: { type: 'bandpass', f0: 500, f1: 5200, q: 2, peak: 0.24 },
-    impact: { sub0: 92, sub1: 34, subDur: 0.9, crash: 2600, crashQ: 0.6, stab: 'triangle', stabPeak: 0.13 },
-  },
-  epic: {
-    trim: 0.611, // 响度校准
-    kick: { f0: 88, f1: 40, sweep: 0.20, dur: 0.80, peak: 0.44, tone: 0.85, thump: 0.26, thumpDur: 0.12, thumpF0: 1100, thumpF1: 120 },
-    snare: { kind: 'noise', type: 'bandpass', f0: 2100, q: 1.1, dur: 0.19, peak: 0.26, attack: 0.001, body: 185, bodyWave: 'sine', bodyLevel: 0.20, bodyDur: 0.16 },
-    hat: { type: 'highpass', f0: 6600, dur: 0.045, peak: 0.055, attack: 0.004 },
-    // 弦乐群：小提琴 + 大提琴分解，而不是铜管主奏
-    lead: 'violin',
-    answer: 'cello',
-    bass: { voices: [{ wave: 'sawtooth', detune: -7 }, { wave: 'sawtooth', detune: 0 }, { wave: 'sawtooth', detune: 7 }], attack: 0.09, sustain: true, release: 0.2, level: 0.26, cutoff: 900 },
-    pad: { voices: [{ wave: 'sawtooth', detune: -10 }, { wave: 'sawtooth', detune: -3 }, { wave: 'sawtooth', detune: 3 }, { wave: 'sawtooth', detune: 10 }], attack: 0.4, sustain: true, release: 0.3, cutoff: 2100, level: 0.14 },
-    riser: { type: 'bandpass', f0: 300, f1: 4500, q: 1.4, peak: 0.34 },
-    impact: { sub0: 76, sub1: 30, subDur: 1.1, crash: 3000, crashQ: 0.5, stab: 'sawtooth', stabPeak: 0.17 },
-  },
-  synth: {
-    trim: 4.377, // 响度校准
-    kick: { f0: 182, f1: 42, sweep: 0.055, dur: 0.55, peak: 0.60, tone: 1, thump: 0.05, thumpDur: 0.02, thumpF0: 800, thumpF1: 90 },
-    snare: { kind: 'noise', type: 'highpass', f0: 1500, q: 0.8, dur: 0.12, peak: 0.26, attack: 0.0008, body: 225, bodyLevel: 0.14, bodyDur: 0.08 },
-    hat: { type: 'highpass', f0: 9200, dur: 0.03, peak: 0.10, attack: 0.0006 },
-    lead: 'acid',
-    bass: { wave: 'sawtooth', attack: 0.004, sustain: true, release: 0.08, level: 0.30, c0: 260, c1: 420, q: 9, resonant: true },
-    pad: { voices: [{ wave: 'sawtooth', detune: -14 }, { wave: 'sawtooth', detune: 14 }], attack: 0.22, sustain: true, release: 0.25, cutoff: 1150, level: 0.13 },
-    riser: { type: 'bandpass', f0: 600, f1: 7000, q: 4, peak: 0.30 },
-    impact: { sub0: 84, sub1: 26, subDur: 0.95, crash: 4200, crashQ: 0.9, stab: 'square', stabPeak: 0.14 },
-  },
-  chiptune: {
-    trim: 14.917, // 响度校准
-    // NES 没有真正的底鼓：三角波快速下扫当鼓，是芯片音乐的标志手法
-    kick: { f0: 220, f1: 48, sweep: 0.04, dur: 0.16, peak: 0.55, tone: 1, thump: 0.12, thumpDur: 0.015, thumpF0: 1200, thumpF1: 200, thumpType: 'highpass', wave: 'triangle' },
-    // 噪声通道：短促、不做滤波塑形
-    snare: { kind: 'noise', type: 'highpass', f0: 2400, q: 0.6, dur: 0.10, peak: 0.24, attack: 0.0004 },
-    hat: { type: 'highpass', f0: 7000, dur: 0.025, peak: 0.11, attack: 0.0004 },
-    lead: 'chipPulse',
-    answer: 'chipSquare',
-    arpeggio: true,
-    // 三角波贝斯：NES 三个音源之一
-    bass: { wave: 'triangle', attack: 0.002, sustain: true, release: 0.04, level: 0.34, cutoff: 2400 },
-    pad: null,
-    riser: { type: 'highpass', f0: 2000, f1: 8000, q: 1, peak: 0.22 },
-    impact: { sub0: 120, sub1: 40, subDur: 0.4, crash: 5200, crashQ: 1.1, stab: 'square', stabPeak: 0.16 },
-  },
-  pulse: {
-    trim: 6.709, // 响度校准
-    kick: { f0: 205, f1: 52, sweep: 0.05, dur: 0.26, peak: 0.62, tone: 1, thump: 0.09, thumpDur: 0.02, thumpF0: 900, thumpF1: 100 },
-    snare: { kind: 'clap', type: 'bandpass', f0: 1800, q: 1.5, peak: 0.24 },
-    hat: { type: 'highpass', f0: 8600, dur: 0.028, peak: 0.10, attack: 0.0006 },
-    lead: 'stab',
-    answer: 'padSynth',
-    bass: { wave: 'square', attack: 0.004, sustain: true, release: 0.06, level: 0.30, cutoff: 760 },
-    pad: { voices: [{ wave: 'square', detune: -8 }, { wave: 'sawtooth', detune: 8 }], attack: 0.18, sustain: true, release: 0.2, cutoff: 2600, level: 0.09 },
-    riser: { type: 'bandpass', f0: 700, f1: 6500, q: 3, peak: 0.30 },
-    impact: { sub0: 88, sub1: 28, subDur: 0.8, crash: 3800, crashQ: 0.8, stab: 'sawtooth', stabPeak: 0.15 },
-  },
-  bright: {
-    trim: 8.116, // 响度校准
-    kick: { f0: 150, f1: 72, sweep: 0.04, dur: 0.17, peak: 0.26, tone: 0.5, thump: 0.30, thumpDur: 0.04, thumpF0: 1800, thumpF1: 300, thumpType: 'bandpass', thumpQ: 0.9 },
-    snare: { kind: 'wood', type: 'bandpass', f0: 2500, q: 4, dur: 0.055, peak: 0.20, attack: 0.0006, body: 620, bodyLevel: 0.09, bodyDur: 0.05 },
-    hat: { type: 'highpass', f0: 6200, dur: 0.06, peak: 0.085, attack: 0.006 },
-    // 马林巴主奏，钢琴应答
-    lead: 'marimba',
-    answer: 'piano',
-    bass: { wave: 'sine', attack: 0.008, sustain: true, release: 0.08, level: 0.28, cutoff: 500 },
-    pad: { voices: [{ wave: 'triangle', detune: -7 }, { wave: 'sine', detune: 7 }], attack: 0.28, sustain: true, release: 0.2, cutoff: 3200, level: 0.12 },
-    riser: { type: 'bandpass', f0: 900, f1: 6000, q: 2.5, peak: 0.22 },
-    impact: { sub0: 96, sub1: 40, subDur: 0.6, crash: 4600, crashQ: 0.9, stab: 'sine', stabPeak: 0.12 },
-  },
-  ambient: {
-    trim: 1.742, // 响度校准
-    // 完全不敲鼓
-    kick: null,
-    snare: null,
-    hat: null,
-    lead: 'bell',
-    answer: 'ePiano',
-    bass: { wave: 'sine', attack: 0.5, sustain: true, release: 0.5, level: 0.26, cutoff: 240 },
-    pad: { voices: [{ wave: 'sine', detune: -12 }, { wave: 'triangle', detune: -2 }, { wave: 'triangle', detune: 2 }, { wave: 'sine', detune: 12 }], attack: 0.9, sustain: true, release: 0.5, cutoff: 560, level: 0.18 },
-    riser: { type: 'bandpass', f0: 200, f1: 3000, q: 1.2, peak: 0.18 },
-    impact: { sub0: 64, sub1: 26, subDur: 1.6, crash: 1800, crashQ: 0.4, stab: 'sine', stabPeak: 0.10 },
-  },
-  cinematic: {
-    trim: 1.994, // 响度校准
-    kick: { f0: 122, f1: 45, sweep: 0.14, dur: 0.55, peak: 0.42, tone: 0.9, thump: 0.18, thumpDur: 0.07, thumpF0: 1000, thumpF1: 120 },
-    snare: { kind: 'noise', type: 'bandpass', f0: 2000, q: 1.0, dur: 0.15, peak: 0.24, attack: 0.001, body: 200, bodyWave: 'triangle', bodyLevel: 0.15, bodyDur: 0.1 },
-    hat: { type: 'highpass', f0: 6800, dur: 0.04, peak: 0.07, attack: 0.003 },
-    // 钢琴主奏，小提琴应答
-    lead: 'piano',
-    answer: 'violin',
-    bass: { voices: [{ wave: 'triangle', detune: -4 }, { wave: 'sine', detune: 4 }], attack: 0.02, sustain: true, release: 0.12, level: 0.30, cutoff: 600 },
-    pad: { voices: [{ wave: 'sawtooth', detune: -8 }, { wave: 'sawtooth', detune: 0 }, { wave: 'sawtooth', detune: 8 }], attack: 0.5, sustain: true, release: 0.3, cutoff: 1350, level: 0.14 },
-    riser: { type: 'bandpass', f0: 400, f1: 5000, q: 2, peak: 0.28 },
-    impact: { sub0: 80, sub1: 28, subDur: 0.95, crash: 3400, crashQ: 0.7, stab: 'triangle', stabPeak: 0.14 },
-  },
-};
-
-/** Unknown ids (including `auto` and the retired `lofi`) keep a distinct default. */
-function kitFor(mood) {
-  return KITS[mood] || KITS.cinematic;
-}
-
-function mulberry32(seed) {
-  let a = seed >>> 0;
-  return () => {
-    a = (a + 0x6D2B79F5) | 0;
-    let t = Math.imul(a ^ (a >>> 15), 1 | a);
-    t = (t + Math.imul(t ^ (t >>> 7), 61 | t)) ^ t;
-    return ((t ^ (t >>> 14)) >>> 0) / 4294967296;
-  };
-}
-
-function midiToHz(midi) {
-  return 440 * 2 ** ((midi - 69) / 12);
-}
-
-function createNoiseBuffer(ctx, seconds, rng) {
-  const len = Math.max(1, Math.floor(ctx.sampleRate * seconds));
-  const buf = ctx.createBuffer(1, len, ctx.sampleRate);
-  const data = buf.getChannelData(0);
-  for (let i = 0; i < len; i++) data[i] = rng() * 2 - 1;
-  return buf;
-}
-
-function createReverbBuffer(ctx, seconds, decay, rng) {
-  const len = Math.max(1, Math.floor(ctx.sampleRate * seconds));
-  const buf = ctx.createBuffer(2, len, ctx.sampleRate);
-  for (let ch = 0; ch < 2; ch++) {
-    const data = buf.getChannelData(ch);
-    for (let i = 0; i < len; i++) data[i] = (rng() * 2 - 1) * (1 - i / len) ** decay;
-  }
-  return buf;
-}
-
-/**
- * Turn the free-form music description into synthesis parameters.
- *
- * Both Chinese and English words are matched, because the description is whatever the
- * Agent wrote for this clip rather than a fixed preset id.
- */
-function analyzePrompt(promptOrId) {
-  const text = String(promptOrId || '').toLowerCase();
-  const has = (...words) => words.some((w) => text.includes(w));
-
-  if (has('国风', '水墨', '中国风', '古', '山水', '禅', 'guofeng', 'chinese', 'ink')) {
-    return {
-      mood: 'guofeng', scale: SCALES.pentatonic, root: 52, bpm: 88,
-      chords: [0, 3, 4, 1, 0, 4], wave: 'triangle', pluckDecay: 0.9,
-      groove: 'sparse', reverbWet: 0.42, cutoff: 2400, impact: 0.85,
-    };
-  }
-  if (has('史诗', '宏大', '震撼', '大片', '战争', '英雄', 'epic', 'cinematic', 'trailer')) {
-    return {
-      mood: 'epic', scale: SCALES.minor, root: 41, bpm: 80,
-      chords: [0, 5, 3, 4], wave: 'sawtooth', pluckDecay: 0.45,
-      groove: 'backbeat', reverbWet: 0.38, cutoff: 1600, impact: 1.0,
-    };
-  }
-  if (has('赛博', '科技', '未来', 'crt', '故障', 'synth', 'cyber', 'electronic', 'techno')) {
-    return {
-      mood: 'synth', scale: SCALES.minor, root: 43, bpm: 124,
-      chords: [0, 3, 5, 2], wave: 'sawtooth', pluckDecay: 0.3,
-      groove: 'four', reverbWet: 0.24, cutoff: 2800, impact: 0.92,
-    };
-  }
-  if (has('轻快', '活泼', '卡通', '趣味', '明快', 'bright', 'playful', 'happy')) {
-    return {
-      mood: 'bright', scale: SCALES.major, root: 48, bpm: 116,
-      chords: [0, 4, 5, 3], wave: 'sine', pluckDecay: 0.35,
-      groove: 'backbeat', reverbWet: 0.2, cutoff: 3200, impact: 0.7,
-    };
-  }
-  if (has('快', '脉冲', '燃', '节奏', '紧张', 'pulse', 'drive', 'energy', 'fast')) {
-    return {
-      mood: 'pulse', scale: SCALES.dorian, root: 45, bpm: 128,
-      chords: [0, 2, 3, 0], wave: 'sawtooth', pluckDecay: 0.25,
-      groove: 'four', reverbWet: 0.18, cutoff: 2200, impact: 0.9,
-    };
-  }
-  if (has('8bit', '8-bit', '芯片', '红白机', 'fc', 'nes', 'chiptune', '像素游戏', 'gameboy', 'gb')) {
-    return {
-      mood: 'chiptune', scale: SCALES.major, root: 60, bpm: 150,
-      chords: [0, 4, 5, 3], wave: 'square', pluckDecay: 0.16,
-      groove: 'backbeat', reverbWet: 0.12, cutoff: 6000, impact: 0.7,
-    };
-  }
-  if (has('氛围', '静谧', '空灵', '孤独', 'ambient', 'calm', 'drone', 'space')) {
-    return {
-      mood: 'ambient', scale: SCALES.minorPent, root: 45, bpm: 72,
-      chords: [0, 5, 3, 4], wave: 'sine', pluckDecay: 0.8,
-      groove: 'sparse', reverbWet: 0.5, cutoff: 1200, impact: 0.6,
-    };
-  }
-  return {
-    mood: 'cinematic', scale: SCALES.minorPent, root: 45, bpm: 96,
-    chords: [0, 3, 4, 2], wave: 'triangle', pluckDecay: 0.45,
-    groove: 'backbeat', reverbWet: 0.34, cutoff: 2000, impact: 0.88,
-  };
-}
-
-/**
- * Choose the beat period whose bar lines land on the clip's cuts.
- *
- * Every cut is a musical arrival, so it should fall on a downbeat. Candidate tempos are
- * scored by the mean distance between each cut and the nearest bar line, with a small
- * preference for the mood's intended tempo when two candidates fit equally well.
- *
- * @param cuts - cut times in seconds.
- * @param duration - clip length.
- * @param hintedBpm - the mood's tempo, used to break ties and when there are no cuts.
- * @returns `{ beat, bar, bpm, fit }` where `fit` is 1 for a perfect alignment.
- */
-function fitGroove(cuts, duration, hintedBpm) {
-  const target = Number(hintedBpm) || 96;
-  const usable = (Array.isArray(cuts) ? cuts : [])
-    .map(Number)
-    .filter((t) => Number.isFinite(t) && t > 0.15 && t < duration - 0.1);
-
-  if (!usable.length) {
-    const beat = 60 / target;
-    return { beat, bar: beat * 4, bpm: target, fit: 0 };
-  }
-
-  let best = null;
-  for (let bpm = 60; bpm <= 190; bpm += 1) {
-    const beat = 60 / bpm;
-    const bar = beat * 4;
-    // A bar shorter than ~0.35s has no room for a groove; longer than 3s drags.
-    if (bar < 0.35 || bar > 3) continue;
-
-    let err = 0;
-    for (const t of usable) {
-      const k = t / bar;
-      err += Math.abs(k - Math.round(k));
-    }
-    err /= usable.length;
-
-    // Ties go to the tempo the mood asked for, so a well-fitting clip keeps its feel.
-    const comfort = Math.abs(bpm - target) / (target * 8);
-    const score = err * 6 + comfort;
-    if (!best || score < best.score) best = { beat, bar, bpm, err, score };
-  }
-
-  if (!best) {
-    const beat = 60 / target;
-    return { beat, bar: beat * 4, bpm: target, fit: 0 };
-  }
-  return { beat: best.beat, bar: best.bar, bpm: best.bpm, fit: Math.max(0, 1 - best.err * 4) };
-}
-
-/**
- * Synthesise a score tailored to one clip.
- *
- * @param musicPromptOrPreset - free-form description (or a legacy preset id).
- * @param options - `{ duration, cuts, cues, seed, sampleRate }`. `cuts` are storyboard
- *   edits; `cues` are in-shot visual peaks. Both get a hit, sample-accurately.
- */
-async function renderMusic(musicPromptOrPreset, options = {}) {
-  const duration = Math.max(0.5, Number(options.duration) || 8);
-  const sampleRate = Number(options.sampleRate) || 44100;
-  const seed = Number(options.seed) || 20261103;
-  const rng = mulberry32(seed);
-  const spec = analyzePrompt(musicPromptOrPreset);
-  // The instrument set for this mood; every voice below reads from it.
-  const K = kitFor(spec.mood);
-
-  // Collect every moment the picture needs an accent on.
-  const rawCuts = Array.isArray(options.cuts) ? options.cuts : [];
-  const rawCues = Array.isArray(options.cues) ? options.cues : [];
-  const allCues = [];
-
-  for (const c of rawCuts) {
-    const t = Number(c);
-    if (Number.isFinite(t) && t > 0.05 && t < duration - 0.05) {
-      allCues.push({ time: t, type: 'cut', power: 1 });
-    }
-  }
-  for (const cue of rawCues) {
-    const t = typeof cue === 'number' ? cue : Number(cue.time ?? cue.t);
-    if (Number.isFinite(t) && t > 0.05 && t < duration - 0.05) {
-      allCues.push({
-        time: t,
-        type: (cue && cue.type) || 'hit',
-        power: (cue && Number(cue.power)) || 0.75,
-      });
-    }
-  }
-
-  allCues.sort((a, b) => a.time - b.time);
-  const cues = [];
-  for (const c of allCues) {
-    if (!cues.some((prev) => Math.abs(prev.time - c.time) < 0.07)) cues.push(c);
-  }
-
-  const totalFrames = Math.ceil(duration * sampleRate);
-  const ctx = new OfflineAudioContext(2, totalFrames, sampleRate);
-
-  // ---- master chain: headroom, glue, then a hard ceiling so impacts never clip ----
-  const master = ctx.createGain();
-  master.gain.value = 0.6;
-
-  const glue = ctx.createDynamicsCompressor();
-  glue.threshold.value = -16;
-  glue.knee.value = 8;
-  glue.ratio.value = 4;
-  glue.attack.value = 0.004;
-  glue.release.value = 0.18;
-
-  const ceiling = ctx.createDynamicsCompressor();
-  ceiling.threshold.value = -3;
-  ceiling.knee.value = 0;
-  ceiling.ratio.value = 20;
-  ceiling.attack.value = 0.001;
-  ceiling.release.value = 0.08;
-
-  // A soft ceiling of last resort. The two compressors above shape the mix, but they
-  // are dynamics processors, not limiters: a bright 8-bit arpeggio at full level can
-  // still overshoot. tanh keeps that from clipping on the way out.
-  const softClip = ctx.createWaveShaper();
-  const curve = new Float32Array(1024);
-  for (let i = 0; i < curve.length; i += 1) {
-    const x = (i / (curve.length - 1)) * 2 - 1;
-    curve[i] = Math.tanh(x * 1.35);
-  }
-  softClip.curve = curve;
-
-  master.connect(glue);
-  glue.connect(ceiling);
-  ceiling.connect(softClip);
-  softClip.connect(ctx.destination);
-
-  const verb = ctx.createConvolver();
-  verb.buffer = createReverbBuffer(ctx, Math.min(2.4, duration * 0.8), 2.2, rng);
-  const wet = ctx.createGain();
-  wet.gain.value = spec.reverbWet;
-  verb.connect(wet);
-  // `wet` joins the kit bus below, not `master`: routing the reverb past the trim let
-  // a wet preset (氛围 uses 0.5) ignore its own loudness calibration.
-
-  const fade = ctx.createGain();
-  fade.gain.setValueAtTime(0.0001, 0);
-  fade.gain.linearRampToValueAtTime(1, Math.min(0.15, duration * 0.04));
-  // This only removes the tail click. The ending itself is a cadence written by the
-  // arrangement below, so the last chord rings instead of being faded out mid-loop.
-  fade.gain.setValueAtTime(1, Math.max(0.15, duration - 0.06));
-  fade.gain.linearRampToValueAtTime(0.0001, duration);
-  fade.connect(master);
-
-  const { beat, bar, bpm, fit } = fitGroove(rawCuts, duration, spec.bpm);
-
-  /**
-   * The bed bus carries the groove. Cue impacts connect past it to `kitBus`, so when
-   * the bed ducks the hit is exposed rather than masked.
-   *
-   * `kitBus` is the per-kit trim: instruments differ in loudness by more than 20 dB
-   * (a bowed string ensemble against a bare pulse wave), so each kit carries one
-   * calibration gain instead of every voice guessing its own level.
-   */
-  const kitBus = ctx.createGain();
-  kitBus.gain.value = K.trim ?? 1;
-  kitBus.connect(fade);
-  wet.connect(kitBus);
-
-  const bed = ctx.createGain();
-  bed.gain.value = 1;
-  bed.connect(kitBus);
-
-  const scale = SCALES[spec.scale] || SCALES.minorPent;
-
-  const noiseBuf = createNoiseBuffer(ctx, 1.5, rng);
-  const hatsBuf = createNoiseBuffer(ctx, 0.2, rng);
-
-  // ---- voices ----------------------------------------------------------------
-  //
-  // Each voice is a thin wrapper over a primitive, and the primitive reads its
-  // parameters from the mood's kit. The call signatures stay as they were, so the
-  // arrangement and ending code above is unchanged.
-
-  /** Filtered noise burst: snare, clap, hat, shaker, woodblock, cymbal. */
-  const noiseBurst = (t, power, dest, k, overrides) => {
-    const cfg = overrides ? { ...k, ...overrides } : k;
-    const src = ctx.createBufferSource();
-    src.buffer = hatsBuf;
-    const f = ctx.createBiquadFilter();
-    f.type = cfg.type || 'bandpass';
-    f.frequency.setValueAtTime(cfg.f0, t);
-    if (cfg.f1) f.frequency.exponentialRampToValueAtTime(cfg.f1, t + cfg.dur);
-    f.Q.value = cfg.q ?? 1;
-    const g = ctx.createGain();
-    const attack = cfg.attack ?? 0.001;
-    g.gain.setValueAtTime(0.0001, t);
-    g.gain.linearRampToValueAtTime(cfg.peak * power, t + attack);
-    g.gain.exponentialRampToValueAtTime(0.0001, t + cfg.dur);
-    src.connect(f);
-    f.connect(g);
-    g.connect(dest);
-    if (cfg.verb) g.connect(verb);
-    src.start(t);
-    src.stop(t + cfg.dur + 0.03);
-  };
-
-  /**
-   * Drum: 808/909 kick, taiko, timpani, tom, or a soft brush thump.
-   *
-   * `tone` and `thump` set the balance between the pitched body and the skin noise.
-   * Every kit used to be a bare sine sweep, which is the single most "electronic"
-   * thing in the score; a frame drum is mostly noise over a short low tone, while an
-   * 808 is almost pure tone.
-   */
-  const playKick = (t, power, dest) => {
-    const k = K.kick;
-    if (!k) return;
-    if (k.tone !== 0) {
-      const osc = ctx.createOscillator();
-      osc.type = 'sine';
-      osc.frequency.setValueAtTime(k.f0, t);
-      osc.frequency.exponentialRampToValueAtTime(k.f1, t + k.sweep);
-      const g = ctx.createGain();
-      g.gain.setValueAtTime(k.peak * (k.tone ?? 1) * power, t);
-      g.gain.exponentialRampToValueAtTime(0.0001, t + k.dur);
-      osc.connect(g);
-      g.connect(dest);
-      osc.start(t);
-      osc.stop(t + k.dur + 0.02);
-    }
-    // The skin: noise the pitched body alone cannot give.
-    if (k.thump > 0) {
-      noiseBurst(t, k.thump * power, dest, {
-        type: k.thumpType || 'lowpass',
-        f0: k.thumpF0 ?? 900, f1: k.thumpF1 ?? 110,
-        dur: k.thumpDur || 0.05, q: k.thumpQ ?? 0.7, peak: 1, attack: 0.0005,
-      });
-    }
-  };
-
-  const playSnare = (t, power, dest) => {
-    const k = K.snare;
-    if (!k) return;
-    if (k.kind === 'clap') {
-      // A clap is three very short bursts, not one long noise hit.
-      for (let i = 0; i < 3; i++) {
-        noiseBurst(t + i * 0.011, power * (1 - i * 0.16), dest, k, { dur: 0.05 + i * 0.02, attack: 0.0005 });
-      }
-      return;
-    }
-    noiseBurst(t, power, dest, k);
-    if (k.body) {
-      // A body tone under the noise so the hit reads on small speakers too.
-      const body = ctx.createOscillator();
-      body.type = k.bodyWave || 'triangle';
-      body.frequency.setValueAtTime(k.body, t);
-      body.frequency.exponentialRampToValueAtTime(k.body * 0.72, t + 0.07);
-      const bg = ctx.createGain();
-      bg.gain.setValueAtTime(k.bodyLevel * power, t);
-      bg.gain.exponentialRampToValueAtTime(0.0001, t + (k.bodyDur ?? 0.1));
-      body.connect(bg);
-      bg.connect(dest);
-      body.start(t);
-      body.stop(t + (k.bodyDur ?? 0.1) + 0.03);
-    }
-  };
-
-  const playHat = (t, power, dest, open = false) => {
-    const k = K.hat;
-    if (!k) return;
-    noiseBurst(t, power, dest, k, open ? { dur: Math.max(k.dur, 0.16) } : undefined);
-  };
-
-  /**
-   * General tonal voice: bass, pad, strings, brass.
-   *
-   * One oscillator per entry in `voices`, so a preset can stack detuned layers for a
-   * string ensemble instead of the single sine every preset used to get.
-   */
-  const tonal = (midi, t, length, level, dest, k) => {
-    const lp = ctx.createBiquadFilter();
-    lp.type = 'lowpass';
-    if (k.c1) {
-      lp.frequency.setValueAtTime(k.c0 ?? k.cutoff ?? 2000, t);
-      lp.frequency.exponentialRampToValueAtTime(k.c1, t + Math.min(length, 0.5));
-    } else {
-      lp.frequency.value = k.cutoff ?? 2000;
-    }
-    if (k.q) lp.Q.value = k.q;
-
-    const g = ctx.createGain();
-    const attack = k.attack ?? 0.01;
-    g.gain.setValueAtTime(0.0001, t);
-    g.gain.linearRampToValueAtTime(level, t + attack);
-    if (k.sustain) {
-      const release = k.release ?? 0.12;
-      g.gain.setValueAtTime(level, Math.max(t + attack, t + length - release));
-      g.gain.linearRampToValueAtTime(0.0001, t + length);
-    } else {
-      g.gain.exponentialRampToValueAtTime(0.0001, t + length);
-    }
-
-    for (const v of (k.voices || [{ wave: k.wave || 'sine', detune: 0 }])) {
-      const osc = ctx.createOscillator();
-      osc.type = v.wave;
-      osc.detune.value = v.detune || 0;
-      osc.frequency.setValueAtTime(midiToHz(midi), t);
-      if (k.vibrato) {
-        const lfo = ctx.createOscillator();
-        lfo.frequency.value = k.vibrato;
-        const lg = ctx.createGain();
-        lg.gain.value = midiToHz(midi) * 0.008;
-        lfo.connect(lg);
-        lg.connect(osc.frequency);
-        lfo.start(t);
-        lfo.stop(t + length + 0.05);
-      }
-      osc.connect(lp);
-      osc.start(t);
-      osc.stop(t + length + 0.04);
-    }
-    lp.connect(g);
-    g.connect(dest);
-    if (k.verb !== false) g.connect(verb);
-  };
-
-  /** FM voice: marimba, bell. Two oscillators give a struck, inharmonic attack. */
-  const fmVoice = (midi, t, dur, level, dest, k) => {
-    const hz = midiToHz(midi);
-    const car = ctx.createOscillator();
-    car.type = 'sine';
-    car.frequency.value = hz;
-    const mod = ctx.createOscillator();
-    mod.type = 'sine';
-    mod.frequency.value = hz * (k.ratio ?? 2);
-    const mg = ctx.createGain();
-    mg.gain.setValueAtTime(hz * (k.index ?? 1.5), t);
-    mg.gain.exponentialRampToValueAtTime(hz * 0.02, t + dur * 0.6);
-    mod.connect(mg);
-    mg.connect(car.frequency);
-
-    const g = ctx.createGain();
-    g.gain.setValueAtTime(0.0001, t);
-    g.gain.linearRampToValueAtTime(level, t + (k.attack ?? 0.004));
-    g.gain.exponentialRampToValueAtTime(0.0001, t + dur);
-    car.connect(g);
-    g.connect(dest);
-    if (k.verb !== false) g.connect(verb);
-    car.start(t);
-    car.stop(t + dur + 0.03);
-    mod.start(t);
-    mod.stop(t + dur + 0.03);
-  };
-
-  const playBass = (midi, t, length, power) => {
-    tonal(midi, t, length, (K.bass.level ?? 0.32) * power, bed, K.bass);
-  };
-
-  /** NES-style pulse wave; `duty` sets the harmonic content (0.125/0.25/0.5). */
-  const pulseWave = (duty) => {
-    const size = 32;
-    const real = new Float32Array(size);
-    const imag = new Float32Array(size);
-    for (let k = 1; k < size; k += 1) imag[k] = (2 / (k * Math.PI)) * Math.sin(Math.PI * k * duty);
-    return ctx.createPeriodicWave(real, imag);
-  };
-  const pulseCache = new Map();
-
-  /** Filtered noise used as a breath, bow or hammer transient. */
-  const transient = (t, dur, level, kind, freq, q = 1) => {
-    const src = ctx.createBufferSource();
-    src.buffer = hatsBuf;
-    const f = ctx.createBiquadFilter();
-    f.type = kind;
-    f.frequency.value = freq;
-    f.Q.value = q;
-    const g = ctx.createGain();
-    g.gain.setValueAtTime(0.0001, t);
-    g.gain.linearRampToValueAtTime(level, t + Math.min(0.006, dur * 0.25));
-    g.gain.exponentialRampToValueAtTime(0.0001, t + dur);
-    src.connect(f);
-    f.connect(g);
-    g.connect(bed);
-    src.start(t);
-    src.stop(t + dur + 0.02);
-  };
-
-  /** Vibrato on one oscillator, in cents relative to its own pitch. */
-  const applyVibrato = (osc, t, dur, rate, depth) => {
-    if (!rate) return;
-    const lfo = ctx.createOscillator();
-    lfo.frequency.value = rate;
-    const g = ctx.createGain();
-    g.gain.value = midiToHz(60) * depth * 12;
-    lfo.connect(g);
-    g.connect(osc.frequency);
-    lfo.start(t);
-    lfo.stop(t + dur + 0.05);
-  };
-
-  /** Additive stack with faster decay on higher partials — a struck string. */
-  const struckVoice = (midi, t, dur, level, k) => {
-    const lp = ctx.createBiquadFilter();
-    lp.type = 'lowpass';
-    lp.frequency.setValueAtTime(k.c0 ?? 2800, t);
-    lp.frequency.exponentialRampToValueAtTime(k.c1 ?? 700, t + dur);
-    const out = ctx.createGain();
-    out.gain.setValueAtTime(0.0001, t);
-    out.gain.linearRampToValueAtTime(level, t + (k.attack ?? 0.003));
-    out.gain.exponentialRampToValueAtTime(0.0001, t + dur);
-    for (const part of (k.partials || [{ octave: 0, gain: 1 }])) {
-      const osc = ctx.createOscillator();
-      osc.type = 'sine';
-      osc.frequency.setValueAtTime(midiToHz(midi + part.octave), t);
-      // Upper partials die first; that two-stage decay is what reads as "piano".
-      const pg = ctx.createGain();
-      const partialDur = part.octave === 0 ? dur : dur * Math.max(0.25, 1 - part.octave / 40);
-      pg.gain.setValueAtTime(0.0001, t);
-      pg.gain.linearRampToValueAtTime(part.gain, t + (k.attack ?? 0.003));
-      pg.gain.exponentialRampToValueAtTime(0.0001, t + partialDur);
-      osc.connect(pg);
-      pg.connect(lp);
-      osc.start(t);
-      osc.stop(t + partialDur + 0.03);
-    }
-    if (k.hammer) transient(t, 0.03, k.hammer * level, 'highpass', k.hammerF ?? 3200);
-    lp.connect(out);
-    out.connect(bed);
-    out.connect(verb);
-  };
-
-  /** Bowed string: saw with bow noise, vibrato and a formant peak. */
-  const bowedVoice = (midi, t, dur, level, k) => {
-    const lp = ctx.createBiquadFilter();
-    lp.type = 'lowpass';
-    lp.frequency.setValueAtTime(k.c0 ?? 3000, t);
-    if (k.sustain) lp.frequency.linearRampToValueAtTime(k.c1 ?? 2000, t + Math.min(0.3, dur));
-    else lp.frequency.exponentialRampToValueAtTime(k.c1 ?? 900, t + dur);
-    const peak = ctx.createBiquadFilter();
-    peak.type = 'peaking';
-    peak.frequency.value = k.formant ?? 900;
-    peak.Q.value = 1.1;
-    peak.gain.value = 7;
-    const g = ctx.createGain();
-    const attack = k.attack ?? 0.07;
-    g.gain.setValueAtTime(0.0001, t);
-    g.gain.linearRampToValueAtTime(level, t + attack);
-    if (k.sustain) {
-      g.gain.setValueAtTime(level, Math.max(t + attack, t + dur - 0.12));
-      g.gain.linearRampToValueAtTime(0.0001, t + dur);
-    } else {
-      g.gain.exponentialRampToValueAtTime(0.0001, t + dur);
-    }
-    for (const detune of [-7, 7]) {
-      const osc = ctx.createOscillator();
-      osc.type = k.wave || 'sawtooth';
-      osc.detune.value = detune;
-      osc.frequency.setValueAtTime(midiToHz(midi), t);
-      applyVibrato(osc, t, dur, k.vibrato, k.vibratoDepth ?? 0.008);
-      osc.connect(lp);
-      osc.start(t);
-      osc.stop(t + dur + 0.04);
-    }
-    if (k.bow) transient(t, 0.05, k.bow * level, 'bandpass', k.bowF ?? 2600, 0.8);
-    lp.connect(peak);
-    peak.connect(g);
-    g.connect(bed);
-    g.connect(verb);
-  };
-
-  /** 笛子: a hollow tone plus breath noise, with vibrato on the held part. */
-  const windVoice = (midi, t, dur, level, k) => {
-    const lp = ctx.createBiquadFilter();
-    lp.type = 'lowpass';
-    lp.frequency.setValueAtTime(k.c0 ?? 2400, t);
-    if (k.sustain) lp.frequency.linearRampToValueAtTime(k.c1 ?? 1800, t + Math.min(0.4, dur));
-    const g = ctx.createGain();
-    const attack = k.attack ?? 0.05;
-    g.gain.setValueAtTime(0.0001, t);
-    g.gain.linearRampToValueAtTime(level, t + attack);
-    g.gain.setValueAtTime(level, Math.max(t + attack, t + dur - 0.1));
-    g.gain.linearRampToValueAtTime(0.0001, t + dur);
-    // A flute is nearly a sine; its character lives in the breath and the vibrato.
-    const osc = ctx.createOscillator();
-    osc.type = k.wave || 'sine';
-    osc.frequency.setValueAtTime(midiToHz(midi), t);
-    applyVibrato(osc, t, dur, k.vibrato, k.vibratoDepth ?? 0.012);
-    const wobble = ctx.createOscillator();
-    wobble.type = 'triangle';
-    wobble.frequency.setValueAtTime(midiToHz(midi + 12), t);
-    const wg = ctx.createGain();
-    wg.gain.value = 0.18;
-    wobble.connect(wg);
-    wg.connect(lp);
-    osc.connect(lp);
-    osc.start(t);
-    osc.stop(t + dur + 0.04);
-    wobble.start(t);
-    wobble.stop(t + dur + 0.04);
-    if (k.breath) transient(t, 0.09, k.breath * level, 'bandpass', k.breathF ?? 2600, 0.7);
-    if (k.air) transient(t + 0.06, Math.max(0.1, dur - 0.06), k.air * level, 'highpass', 3600);
-    lp.connect(g);
-    g.connect(bed);
-    g.connect(verb);
-  };
-
-  /** Unfiltered pulse/triangle: the 8-bit voice. Nothing shapes it but its envelope. */
-  const chipVoice = (midi, t, dur, level, k) => {
-    const osc = ctx.createOscillator();
-    if (k.synth === 'chipTri') {
-      osc.type = 'triangle';
-    } else {
-      const duty = k.duty ?? 0.5;
-      if (!pulseCache.has(duty)) pulseCache.set(duty, pulseWave(duty));
-      osc.setPeriodicWave(pulseCache.get(duty));
-    }
-    osc.frequency.setValueAtTime(midiToHz(midi), t);
-    const g = ctx.createGain();
-    const attack = k.attack ?? 0.001;
-    g.gain.setValueAtTime(0.0001, t);
-    g.gain.linearRampToValueAtTime(level, t + attack);
-    g.gain.setValueAtTime(level, Math.max(t + attack, t + dur - 0.02));
-    g.gain.linearRampToValueAtTime(0.0001, t + dur);
-    osc.connect(g);
-    // A touch of the room keeps it from sounding like a bare test tone.
-    g.connect(bed);
-    g.connect(verb);
-    osc.start(t);
-    osc.stop(t + dur + 0.02);
-  };
-
-  /** Dispatch one melodic note to whichever instrument the kit names. */
-  const playInstrument = (name, midi, t, vol, decay) => {
-    const k = INSTRUMENTS[name] || INSTRUMENTS.piano;
-    const dur = Math.max(0.08, decay * (k.decayScale ?? 1));
-    const level = vol * (k.level ?? 1);
-    switch (k.synth) {
-      case 'fm': fmVoice(midi, t, dur, level, bed, k); break;
-      case 'struck': struckVoice(midi, t, dur, level, k); break;
-      case 'bowed': bowedVoice(midi, t, dur, level, k); break;
-      case 'wind': windVoice(midi, t, dur, level, k); break;
-      case 'chip':
-      case 'chipTri': chipVoice(midi, t, dur, level, k); break;
-      default: {
-        // Struck/plucked string: fast attack, exponential decay, filter sweep, and a
-        // stack of partials; `q` gives a synth stab its resonant onset.
-        const lp = ctx.createBiquadFilter();
-        lp.type = 'lowpass';
-        lp.frequency.setValueAtTime(k.c0 ?? spec.cutoff, t);
-        lp.frequency.exponentialRampToValueAtTime(k.c1 ?? 420, t + dur);
-        if (k.q) lp.Q.value = k.q;
-        const g = ctx.createGain();
-        g.gain.setValueAtTime(0.0001, t);
-        g.gain.linearRampToValueAtTime(level, t + (k.attack ?? 0.005));
-        g.gain.exponentialRampToValueAtTime(0.0001, t + dur);
-        const layers = [{ wave: k.wave || spec.wave, gain: 1 }];
-        for (const part of (k.partials || [])) layers.push(part);
-        for (const layer of layers) {
-          const osc = ctx.createOscillator();
-          osc.type = layer.wave;
-          osc.frequency.setValueAtTime(midiToHz(midi + (layer.octave || 0)), t);
-          if (k.bend) osc.frequency.exponentialRampToValueAtTime(midiToHz(midi + (layer.octave || 0) + k.bend), t + Math.min(0.25, dur));
-          const lg = ctx.createGain();
-          lg.gain.value = layer.gain;
-          osc.connect(lg);
-          lg.connect(lp);
-          osc.start(t);
-          osc.stop(t + dur + 0.04);
+     * 把用户自己选的音频文件解码成 AudioBuffer，并铺满整段片长。
+     *
+     * 这是唯一一处「不是模型写的配乐」还能进片子的入口，而且必须由用户主动选文件；
+     * 插件自己不会生成任何兜底配乐。
+     */
+    async function audioFileToBuffer(file, duration, sampleRate = 44100) {
+      const bytes = await file.arrayBuffer();
+      const ctx = new OfflineAudioContext(2, Math.ceil(duration * sampleRate), sampleRate);
+      const decoded = await ctx.decodeAudioData(bytes);
+      const targetLength = Math.ceil(duration * decoded.sampleRate);
+      const out = ctx.createBuffer(2, targetLength, decoded.sampleRate);
+
+      for (let ch = 0; ch < 2; ch += 1) {
+        const from = decoded.getChannelData(Math.min(ch, decoded.numberOfChannels - 1));
+        const to = out.getChannelData(ch);
+        const take = Math.min(targetLength, from.length);
+        to.set(from.subarray(0, take));
+        // 素材比片子短就首尾相接地铺满，不做拉伸，免得变速变调。
+        if (take < targetLength && from.length > 0) {
+          let offset = take;
+          while (offset < targetLength) {
+            const chunk = Math.min(from.length, targetLength - offset);
+            to.set(from.subarray(0, chunk), offset);
+            offset += chunk;
+          }
         }
-        lp.connect(g);
-        g.connect(bed);
-        g.connect(verb);
       }
-    }
-  };
-
-  /** The kit's melody line; `answer` alternates with it for a call-and-response. */
-  const playPluck = (midi, t, vol, decay, useAnswer) => {
-    const name = useAnswer && K.answer ? K.answer : K.lead;
-    playInstrument(name, midi, t, vol, decay);
-  };
-
-  const playPad = (midi, t, length, vol) => {
-    const k = K.pad;
-    if (!k) return;
-    const attack = Math.min(k.attack ?? 0.35, length * 0.4);
-    tonal(midi, t, length, vol * (k.level ?? 1), bed, { ...k, attack });
-  };
-
-  const playRiser = (targetT, power) => {
-    const k = K.riser || { type: 'bandpass', f0: 400, f1: 6000, q: 3, peak: 0.3 };
-    const lead = Math.min(0.4, targetT);
-    if (lead < 0.12) return;
-    const startT = targetT - lead;
-    const src = ctx.createBufferSource();
-    src.buffer = noiseBuf;
-    src.loop = true;
-    const bp = ctx.createBiquadFilter();
-    bp.type = k.type;
-    bp.Q.value = k.q;
-    bp.frequency.setValueAtTime(k.f0, startT);
-    bp.frequency.exponentialRampToValueAtTime(k.f1, targetT);
-    const g = ctx.createGain();
-    g.gain.setValueAtTime(0.0001, startT);
-    g.gain.linearRampToValueAtTime(k.peak * power, targetT - 0.02);
-    g.gain.linearRampToValueAtTime(0.0001, targetT);
-    src.connect(bp);
-    bp.connect(g);
-    g.connect(bed);
-    src.start(startT);
-    src.stop(targetT + 0.03);
-  };
-
-  /** The cut accent: sub boom + noise crash + an in-key chord stab. */
-  const playImpact = (t, power, rootMidi) => {
-    const sub = ctx.createOscillator();
-    sub.type = 'sine';
-    sub.frequency.setValueAtTime(72, t);
-    sub.frequency.exponentialRampToValueAtTime(28, t + 0.6);
-    const sg = ctx.createGain();
-    sg.gain.setValueAtTime(0.42 * power, t);
-    sg.gain.exponentialRampToValueAtTime(0.0001, t + 0.75);
-    sub.connect(sg);
-    sg.connect(kitBus);
-    sub.start(t);
-    sub.stop(t + 0.8);
-
-    playKick(t, 1.1 * power, kitBus);
-
-    // The crash bright and length come from the kit too: a gong for 国风, a bright
-    // metal crash for 赛博, a soft wash for 氛围.
-    const crash = ctx.createBufferSource();
-    crash.buffer = noiseBuf;
-    const bp = ctx.createBiquadFilter();
-    bp.type = 'bandpass';
-    bp.frequency.value = K.impact.crash;
-    bp.Q.value = K.impact.crashQ;
-    const cg = ctx.createGain();
-    cg.gain.setValueAtTime(0.3 * power, t);
-    cg.gain.exponentialRampToValueAtTime(0.0001, t + 0.9);
-    crash.connect(bp);
-    bp.connect(cg);
-    cg.connect(kitBus);
-    cg.connect(verb);
-    crash.start(t);
-    crash.stop(t + 1);
-
-    // A stab makes the hit harmonic instead of a bare boom.
-    const stabNotes = [0, 7, 12];
-    for (const off of stabNotes) {
-      const osc = ctx.createOscillator();
-      osc.type = K.impact.stab;
-      osc.frequency.value = midiToHz(rootMidi + off);
-      const g = ctx.createGain();
-      g.gain.setValueAtTime(0.0001, t);
-      g.gain.linearRampToValueAtTime(K.impact.stabPeak * power, t + 0.008);
-      g.gain.exponentialRampToValueAtTime(0.0001, t + 0.5);
-      osc.connect(g);
-      g.connect(kitBus);
-      g.connect(verb);
-      osc.start(t);
-      osc.stop(t + 0.55);
-    }
-  };
-
-  // ---- sidechain: the bed ducks under every cue ----
-  bed.gain.setValueAtTime(1, 0);
-  for (const cue of cues) {
-    const from = Math.max(0, cue.time - 0.1);
-    const depth = 0.2 + (1 - Math.min(1, cue.power)) * 0.28;
-    bed.gain.setValueAtTime(1, from);
-    bed.gain.linearRampToValueAtTime(depth, cue.time);
-    bed.gain.setValueAtTime(depth, cue.time + 0.045);
-    bed.gain.linearRampToValueAtTime(1, Math.min(duration, cue.time + 0.2));
-  }
-
-  // ---- arrangement: bars of real groove, with a lead-in and a written ending ----
-  //
-  // The clip is not "N copies of one bar". A three-part shape is imposed instead:
-  // the first bar thins out as a lead-in, the middle bars rotate through the pattern
-  // variants so no two consecutive bars are identical, and the last bar stops the
-  // groove and resolves on the tonic. Previously every bar repeated one hardcoded
-  // pattern and the piece simply faded, which read as the same loop cut off.
-  const totalBars = Math.max(1, Math.ceil(duration / bar));
-  const cueTimes = new Set(cues.map((c) => c.time.toFixed(2)));
-  const isCueAt = (t) => cueTimes.has(t.toFixed(2));
-
-  const variants = GROOVE_PATTERNS[spec.groove] || GROOVE_PATTERNS.backbeat;
-  const endingBar = totalBars - 1;
-  const endingT = endingBar * bar;
-  // Only spend a bar on a lead-in when the clip is long enough to afford one.
-  const hasIntro = totalBars >= 3;
-  const tonic = spec.root + (scale[spec.chords[0] % scale.length] || 0);
-
-  for (let b = 0; b < totalBars; b++) {
-    const barT = b * bar;
-    if (barT >= duration) break;
-    const isEnding = b === endingBar;
-    const isIntro = hasIntro && b === 0;
-    const barLen = Math.min(bar, duration - barT);
-    const stepDur = bar / 16;
-    const chordIdx = spec.chords[b % spec.chords.length];
-    const rootNote = spec.root + (scale[chordIdx % scale.length] || 0);
-
-    if (isEnding) {
-      // Hold the tonic under the cadence for as long as the clip still has room.
-      const ring = Math.max(0.4, duration - barT);
-      playPad(tonic, barT, ring, 0.2);
-      playBass(tonic - 12, barT, ring, 0.95);
-    } else {
-      // A final bar usually gets cut short; only start a pad the bar can hold.
-      if (barLen > beat * 1.2) playPad(rootNote, barT, barLen, isIntro ? 0.12 : 0.16);
-      playBass(rootNote - 12, barT, Math.min(beat * 1.8, barLen), isIntro ? 0.7 : 0.9);
-    }
-
-    const pattern = variants[b % variants.length];
-    for (let s = 0; s < 16; s++) {
-      const t = barT + s * stepDur;
-      if (t >= duration - 0.02) break;
-      if (isCueAt(t)) continue; // the impact covers this step
 
-      if (isEnding) {
-        // One downbeat, then silence so the resolution is heard rather than filled.
-        if (s === 0) playKick(t, 1, bed);
-        continue;
+      const fadeLen = Math.min(Math.floor(out.sampleRate * 0.4), targetLength);
+      for (let ch = 0; ch < 2; ch += 1) {
+        const data = out.getChannelData(ch);
+        for (let i = 0; i < fadeLen; i += 1) data[targetLength - 1 - i] *= i / fadeLen;
       }
-      if (isIntro) {
-        // Thin lead-in: a soft pulse and one open hat, no full groove yet.
-        if (s === 0) playKick(t, 0.55, bed);
-        if (s === 8) playHat(t, 0.3, bed, true);
-        continue;
-      }
-
-      const mark = pattern[s] || '.';
-      if (mark === 'K') playKick(t, 1, bed);
-      else if (mark === 'k') playKick(t, 0.7, bed);
-      else if (mark === 'X') { playKick(t, 1, bed); playSnare(t, 0.8, bed); }
-      else if (mark === 'S') playSnare(t, 0.88, bed);
-      else if (mark === 'H') playHat(t, 0.7, bed, false);
-      else if (mark === 'h') playHat(t, 0.45, bed, false);
-      else if (mark === 'o') playHat(t, 0.6, bed, true);
-    }
-
-    // Melodic movement, rotated per bar so the phrase does not repeat verbatim.
-    if (!isEnding) {
-      const steps = isIntro ? [0] : [0, 6 + (b % 2) * 2, 10];
-      steps.forEach((s, stepIndex) => {
-        const t = barT + s * stepDur;
-        if (t >= duration - 0.05) return;
-        if (isCueAt(t)) return;
-        const degree = scale[Math.floor(rng() * scale.length)];
-        const octave = rng() > 0.45 ? 12 : 0;
-        const midi = rootNote + degree + octave;
-        const vol = 0.13 + rng() * 0.05;
-        // The second instrument answers on alternating steps, so a phrase is a交换
-        // between two timbres rather than one colour repeating.
-        const answer = stepIndex % 2 === 1 && K.answer;
-        if (K.arpeggio) {
-          // 8-bit lead writing: cycle the chord quickly instead of holding one note.
-          const arp = [midi, midi + 4, midi + 7, midi + 12];
-          const gap = Math.min(0.055, stepDur / 4);
-          arp.forEach((note, i) => playPluck(note, t + i * gap, vol * 0.85, gap * 3.2, i % 2 === 1));
-        } else {
-          playPluck(midi, t, vol, spec.pluckDecay, answer);
-        }
-      });
+      return out;
     }
 
-    // Fill into the next cut when one lands on this bar's end.
-    const nextCut = isEnding ? undefined
-      : cues.find((c) => c.time > barT + bar * 0.6 && c.time < barT + bar + 0.02);
-    if (nextCut && beat > 0.12) {
-      const fillStart = nextCut.time - beat * 0.5;
-      for (let k = 0; k < 3; k++) {
-        const t = fillStart + k * (beat * 0.5) / 3;
-        if (t > barT && t < duration - 0.02 && !isCueAt(t)) playSnare(t, 0.35 + k * 0.18, bed);
-      }
-    }
-  }
-
-  // ---- the cuts themselves: riser then impact, sample-accurate ----
-  for (const cue of cues) {
-    playRiser(cue.time, cue.power);
-    const chordIdx = Math.max(0, Math.floor(cue.time / bar)) % spec.chords.length;
-    const rootNote = spec.root + (scale[spec.chords[chordIdx] % scale.length] || 0);
-    playImpact(cue.time, cue.power * spec.impact, rootNote);
-  }
-
-  // ---- the ending: a real cadence on the tonic, not a cut ----
-  //
-  // A tonic triad that rings to the last sample, plus one impact as the final button.
-  // This is what makes a five-second clip feel finished instead of looped-and-stopped.
-  if (endingT < duration - 0.05) {
-    const ring = duration - endingT;
-    // Keep the cadence out of any sidechain dip that lands on the same downbeat.
-    bed.gain.setValueAtTime(1, endingT);
-    for (const degree of [0, 2, 4]) {
-      const tone = tonic + (scale[degree % scale.length] || 0);
-      playPluck(tone + 12, endingT, 0.1, Math.min(ring, 1.5));
-      playPluck(tone, endingT + 0.012, 0.08, Math.min(ring, 1.5));
-    }
-    playPluck(tonic + 24, endingT, 0.07, Math.min(ring, 1.2));
-    if (!isCueAt(endingT)) playImpact(endingT, 0.7 * spec.impact, tonic);
-  }
-
-  const buffer = await ctx.startRendering();
-  // Report the tempo actually used, so a caller can log why the cuts fit.
-  buffer.__groove = {
-    bpm, bar, fit: Number(fit.toFixed(3)), cues: cues.length,
-    bars: totalBars, groove: spec.groove, variants: variants.length,
-  };
-  return buffer;
-}
-
-/** Audition a score through the speakers. */
-function previewMusic(promptOrPreset, duration, options = {}) {
-  const ctx = new AudioContext();
-  let stopped = false;
-  let source = null;
-
-  const pending = renderMusic(promptOrPreset, { ...options, duration, sampleRate: ctx.sampleRate })
-    .then((buffer) => {
-      if (stopped) { void ctx.close(); return; }
-      source = ctx.createBufferSource();
-      source.buffer = buffer;
-      source.connect(ctx.destination);
-      source.onended = () => { void ctx.close(); };
-      source.start(0);
-    })
-    .catch(() => { void ctx.close(); });
-
-  return {
-    stop() {
-      stopped = true;
-      try { if (source) source.stop(); } catch { /* already ended */ }
-      void pending;
-      void ctx.close();
-    },
-  };
-}
-
-/** Decode a user-supplied file and fit it to the clip length. */
-async function audioFileToBuffer(file, duration, sampleRate) {
-  const bytes = await file.arrayBuffer();
-  const ctx = new OfflineAudioContext(2, Math.ceil(duration * (sampleRate || 44100)), sampleRate || 44100);
-  const decoded = await ctx.decodeAudioData(bytes);
-  const targetLength = Math.ceil(duration * decoded.sampleRate);
-  const out = ctx.createBuffer(2, targetLength, decoded.sampleRate);
-
-  for (let ch = 0; ch < 2; ch++) {
-    const from = decoded.getChannelData(Math.min(ch, decoded.numberOfChannels - 1));
-    const to = out.getChannelData(ch);
-    const take = Math.min(targetLength, from.length);
-    to.set(from.subarray(0, take));
-    if (take < targetLength && from.length > 0) {
-      let offset = take;
-      while (offset < targetLength) {
-        const chunk = Math.min(from.length, targetLength - offset);
-        to.set(from.subarray(0, chunk), offset);
-        offset += chunk;
-      }
-    }
-  }
-
-  const fadeLen = Math.min(Math.floor(out.sampleRate * 0.4), targetLength);
-  for (let ch = 0; ch < 2; ch++) {
-    const data = out.getChannelData(ch);
-    for (let i = 0; i < fadeLen; i++) data[targetLength - 1 - i] *= i / fadeLen;
-  }
-  return out;
-}
-
-
-
-
     /**
      * Container preference, best first.
      *
@@ -1557,8 +386,9 @@ async function audioFileToBuffer(file, duration, sampleRate) {
 
     /**
      * Drive a detached document frame frame by frame and record it through MediaRecorder.
-     * @param opts.music - a preset id, or an AudioBuffer for a user-supplied track.
-     *   Any audio is mixed at render time, so video and score start together.
+     * @param opts.scoreBuffer - 模型写的配乐，已经离线渲染好的 AudioBuffer。
+     * @param opts.musicBuffer - 用户自己选的音频文件，优先级高于 scoreBuffer。
+     *   两者都不给就是无声导出；这里不会凭空合成配乐。
      * @param opts.cuts - cut times in seconds; the score is fitted and accented to them.
      */
     /**
@@ -1602,24 +432,6 @@ async function audioFileToBuffer(file, duration, sampleRate) {
       ].filter(Boolean).join(''));
     }
 
-    /**
-     * Stable per-clip seed for the score engine.
-     *
-     * `renderMusic` falls back to one hardcoded seed, so before this every clip in
-     * every style drew the identical random sequence — the same melodic contours,
-     * the same noise and reverb buffers. Hashing the clip id keeps one clip's score
-     * reproducible while letting different clips differ.
-     */
-    function seedOf(text) {
-      let h = 2166136261;
-      const s = String(text || 'clip');
-      for (let i = 0; i < s.length; i += 1) {
-        h ^= s.charCodeAt(i);
-        h = Math.imul(h, 16777619);
-      }
-      return (h >>> 0) || 1;
-    }
-
     async function recordAnimation(iframe, opts) {
       const win = iframe && iframe.contentWindow;
       if (!win) throw new Error('预览还没准备好');
@@ -1633,28 +445,16 @@ async function audioFileToBuffer(file, duration, sampleRate) {
       const videoTrack = stream.getVideoTracks()[0];
       const tracks = [videoTrack];
 
-      // Audio, when requested, is mixed into the same recording session so the two
-      // tracks share one clock.
+      // Audio is mixed into the same recording session so the two tracks share one clock.
+      // There is no engine behind this: the buffer must be the model's score (or a file
+      // the user picked). Missing it is an error, not a reason to synthesize something.
       let audioCtx;
       let source;
       let audioBuffer;
       let analyser;
-      const wantsMusic = opts.music && opts.music !== 'none';
-      if (wantsMusic || opts.musicBuffer) {
-        if (opts.musicBuffer) {
-          audioBuffer = opts.musicBuffer;
-        } else if (opts.scoreBuffer) {
-          // Reuse the buffer already rendered for auditioning: same audio as the preview.
-          audioBuffer = opts.scoreBuffer;
-        } else {
-          if (opts.onStage) opts.onStage('正在合成配乐…');
-          audioBuffer = await renderMusic(opts.music, {
-            duration: opts.duration,
-            seed: opts.musicSeed ?? seedOf(opts.clipId),
-            cuts: opts.cuts,
-            cues: opts.cues,
-          });
-        }
+      const wantsAudio = opts.scoreBuffer || opts.musicBuffer;
+      if (wantsAudio) {
+        audioBuffer = opts.musicBuffer || opts.scoreBuffer;
         audioCtx = new AudioContext();
         const destination = audioCtx.createMediaStreamDestination();
         source = audioCtx.createBufferSource();
@@ -2094,25 +894,30 @@ async function audioFileToBuffer(file, duration, sampleRate) {
       }, [stopAudio]);
 
       // Render the score once so auditioning is instant and export reuses the buffer.
+      // 只有模型写的 score_js 会被渲染：没有兜底引擎，拿不到就是没声音，并说明原因。
       useEffect(() => {
         if (custom) { scoreRef.current = custom; return undefined; }
-        const hasModelScore = !!anim.scoreUrl;
-        if (!hasModelScore && (!music || music === 'none')) { scoreRef.current = null; return undefined; }
+        if (music === 'none') { scoreRef.current = null; setScoreBuffer(null); return undefined; }
+        if (!anim.scoreUrl) {
+          scoreRef.current = null;
+          setScoreBuffer(null);
+          setNote('这一版没有配乐代码，无法出声；请让 Agent 重新提交 score_js');
+          return undefined;
+        }
         let cancelled = false;
         scoreRef.current = null;
 
-        const build = hasModelScore
-          // Prefer the score the model wrote for this clip; the engine is the fallback.
-          ? () => renderModelScore(anim.scoreUrl, { duration, cuts, cues, beats })
-              .then((buf) => buf || renderMusic(music || 'cinematic', { duration, cuts, cues, seed: seedOf(anim.id) }))
-          : () => renderMusic(music, { duration, cuts, cues, seed: seedOf(anim.id) });
-
-        build()
+        renderModelScore(anim.scoreUrl, { duration, cuts, cues, beats })
           .then((buf) => {
             if (cancelled) return;
+            if (!buf) {
+              setNote('配乐代码没有渲染出声音，请让 Agent 修 score_js');
+              setScoreBuffer(null);
+              return;
+            }
             scoreRef.current = buf;
             // Hand the buffer to the timeline so it can draw the true waveform.
-            setScoreBuffer(buf || null);
+            setScoreBuffer(buf);
           })
           .catch(() => {});
         return () => { cancelled = true; };
@@ -2166,18 +971,20 @@ async function audioFileToBuffer(file, duration, sampleRate) {
         setStage('');
         try {
           // The score may still be rendering when export is pressed; wait for it so the
-          // exported audio is the same one the preview played.
+          // exported audio is the same one the preview played. 没有兜底引擎：等不到就报错。
           let score = scoreRef.current;
           if (!score && !custom && anim.scoreUrl) {
             setStage('正在合成配乐…');
             score = await renderModelScore(anim.scoreUrl, { duration, cuts, cues, beats }).catch(() => null);
           }
+          if (!custom && music !== 'none' && !score) {
+            throw new Error('配乐还没合成出来，无法导出。请让 Agent 重新提交 score_js。');
+          }
           const { blob, container, hasAudio, audioPeak } = await recordAnimation(frameRef.current, {
             duration,
             fps,
             cuts,
             container: activeFormat,
-            music: custom ? 'none' : music,
             musicBuffer: custom || undefined,
             scoreBuffer: score || undefined,
             clipId: anim.id,
@@ -2509,6 +1316,10 @@ async function audioFileToBuffer(file, duration, sampleRate) {
           // to the real edits rather than to a fixed tempo.
           const board = state.storyboards.find((b) => b.id === args.storyboard_id)
             || state.storyboards[state.storyboards.length - 1];
+          // The card can only play the model's score if it knows where it lives.
+          // 先读回执里的「配乐地址」，读不到就直接按 id 拼：配乐现在是必填项，每次保存
+          // 都一定写了 <id>.score.js，所以这个地址永远有效——不再依赖回执文案的措辞。
+          const scoreUrl = (new RegExp('配乐地址：([^\\s]+)').exec(text) || [])[1] || `${API}/score/${id}`;
           return {
             ...state,
             animations: [...state.animations, {
@@ -2519,7 +1330,10 @@ async function audioFileToBuffer(file, duration, sampleRate) {
               duration: Number(args.duration) || 8,
               fps: Number(args.fps) || 30,
               music: String(args.music || (board && board.music) || ''),
+              scoreUrl,
               cuts: board ? board.cuts || [] : [],
+              cues: Array.isArray(args.cues) ? args.cues : [],
+              beats: Array.isArray(args.beats) ? args.beats : [],
               warnings: warn ? warn.split('；').map((s) => s.trim()).filter(Boolean) : [],
             }],
           };
@@ -2621,6 +1435,113 @@ async function audioFileToBuffer(file, duration, sampleRate) {
       );
     }
 
+    /**
+     * 样式库视图：把 193 种已上架风格连同参考图一次列全。
+     *
+     * 目录来自 Host 的 /code-animation-api/styles，参考图走 /still/<id>，两者都只
+     * 认已上架的风格，所以这里看到的就是 Agent 能选到的那一批。
+     *
+     * 它注册成 conversation.view，也就是「对话 / 轨迹」旁边多出来的那个页签。
+     */
+    function StyleLibrary({ sessionId, useSessions }) {
+      const selector = useMemo(() => (state) => {
+        const value = state.byId[sessionId]?.projectionValues?.agentPreset;
+        return typeof value === 'string' ? value : undefined;
+      }, [sessionId]);
+      const preset = useSessions(selector);
+      const [data, setData] = useState(null);
+      const [error, setError] = useState('');
+      const [query, setQuery] = useState('');
+      const [series, setSeries] = useState('*');
+      const [zoom, setZoom] = useState('');
+
+      useEffect(() => {
+        let alive = true;
+        fetch(`${API}/styles`)
+          .then((response) => (response.ok ? response.json() : Promise.reject(new Error(`HTTP ${response.status}`))))
+          .then((payload) => { if (alive) setData(payload); })
+          .catch((cause) => { if (alive) setError(cause.message || '目录加载失败'); });
+        return () => { alive = false; };
+      }, []);
+
+      useEffect(() => {
+        const onKey = (event) => { if (event.key === 'Escape') setZoom(''); };
+        window.addEventListener('keydown', onKey);
+        return () => window.removeEventListener('keydown', onKey);
+      }, []);
+
+      const all = (data && data.styles) || [];
+      const list = useMemo(() => {
+        const keyword = query.trim().toLowerCase();
+        return all.filter((style) => {
+          if (series !== '*' && style.series !== series) return false;
+          if (!keyword) return true;
+          return [style.id, style.name, style.en, style.description, style.fit, (style.tags || []).join(' ')]
+            .join(' ').toLowerCase().includes(keyword);
+        });
+      }, [all, query, series]);
+
+      const chips = [{ id: '*', name: '全部' }, ...((data && data.series) || [])];
+
+      // 这个页签对所有会话都注册，但只有创作模式的会话有内容可看。
+      if (preset !== 'code-animation') {
+        return h('div', { className: 'cas-lib-view' },
+          h('style', null, CSS),
+          h('div', { className: 'cas-lib-empty' }, '样式库只在创作模式的会话里可用'));
+      }
+
+      return h('div', { className: 'cas-lib-view' },
+        h('style', null, CSS),
+        h('div', { className: 'cas-lib-head' },
+          h('span', { className: 'cas-lib-title' }, '样式库'),
+          h('span', { className: 'cas-lib-count' }, error || `${list.length} / ${all.length}`),
+          h('input', {
+            className: 'cas-lib-search',
+            placeholder: '搜编号、名字、气质或题材，例如 水墨 / 赛博 / B3',
+            value: query,
+            onChange: (event) => setQuery(event.target.value),
+          }),
+        ),
+        h('div', { className: 'cas-lib-chips' },
+          chips.map((chip) => h('button', {
+            key: chip.id,
+            type: 'button',
+            className: 'cas-chip',
+            'data-on': series === chip.id ? '1' : '0',
+            onClick: () => setSeries(chip.id),
+          }, chip.id === '*' ? chip.name : `${chip.id} ${chip.name}`)),
+        ),
+        h('div', { className: 'cas-lib-body' },
+          error
+            ? h('div', { className: 'cas-lib-empty' }, `目录加载失败：${error}`)
+            : !data
+              ? h('div', { className: 'cas-lib-empty' }, '正在加载…')
+              : !list.length
+                ? h('div', { className: 'cas-lib-empty' }, '没有匹配的风格')
+                : h('div', { className: 'cas-lib-grid' },
+                  list.map((style) => h('button', {
+                    key: style.id,
+                    type: 'button',
+                    className: 'cas-lib-card',
+                    title: [style.description, style.fit].filter(Boolean).join(' · '),
+                    onClick: () => setZoom(style.stillUrl),
+                  },
+                  h('img', { src: style.stillUrl, alt: `${style.id} ${style.name}`, loading: 'lazy' }),
+                  h('div', { className: 'cas-lib-cap' },
+                    h('span', { className: 'cas-lib-id' }, style.id),
+                    h('span', { className: 'cas-lib-name' }, style.name),
+                    h('span', { className: 'cas-lib-desc' }, style.description || style.fit || ''),
+                  ),
+                  ))),
+        ),
+        zoom
+          ? h('div', {
+            className: 'cas-lightbox',
+            onClick: () => setZoom(''),
+          }, h('img', { src: zoom, alt: '' }))
+          : null);
+    }
+
     return {
       inject: ['slots', 'uiConversation'],
       apply(ctx) {
@@ -2644,6 +1565,16 @@ async function audioFileToBuffer(file, duration, sampleRate) {
             key: 'save_storyboard',
           }, StoryboardRow);
         });
+
+        // 样式库做成「对话 / 轨迹」旁边的第三个页签：view 槽的每个注册就是页签行
+        // 里的一项，内容由同一个组件渲染。非创作模式的会话也能点开，里面只说明
+        // 它属于创作模式，不报错也不空白。
+        ctx.slots.inject('conversation.view', () => ctx.slots.register({
+          name: 'conversation.view',
+          id: 'styles',
+          order: 20,
+          label: '样式库',
+        }, StyleLibrary));
       },
     };
   },