feat(sounds): replace call and stream effects

The originals were all 1.14s — long enough to be intrusive for events that fire whenever anyone joins, leaves, or watches a stream. Synthesised rather than sourced: no third-party file, so no licensing question in a public repository. The parameters are measured, not guessed — envelope, spectral peaks and decay taken from two reference sounds the instance owner supplied. The timbre is fundamental plus octave at near-equal strength (1.00 / 0.85 / 0.10 / 0.02) decaying to 1/e in 0.19s, with no reverb. Call join rises C4→G4, leave falls D4→G3, matching the references' intervals and their 100ms spacing. The stream pair reuses those resolution notes as single tones at ~55% the volume: they fire far more often during a broadcast, so they have to sit under the call sounds rather than beside them. Stream start is the one event that happens once per broadcast, so it can be a chord — with a low thump and a short air layer that both die inside 150ms, adding weight without length. Stream end mirrors it descending, quieter, and without the air, since brightness reads as arrival. Generators and the measured parameters are kept in tools/sfx so these can be retuned without redoing the analysis. Total size drops from 150KB to 41KB.
feat(desktop): capture system audio with process isolation
2026-09-01 13:01:06 -03:00 · 2026-09-01 00:56:34 -03:00
18 changed files with 509 additions and 3 deletions
@@ -84,7 +84,7 @@ modais de convite · configurações restantes · telas de erro
 | Problema | Causa provável | Correção |
 |---|---|---|
 | **Enviar som ao soundboard não funciona no app desktop** (funciona no navegador) | `SoundboardPopover` pede o nome do som com `window.prompt`, que o **Electron não implementa** — não abre nada e devolve vazio, então o fluxo aborta em silêncio, sem erro. `window.prompt` aparece em exatamente um lugar no projeto: esse. O resto do código já o evitava | Trocar por um campo de texto dentro do próprio popover (ou um modal reutilizável). Some o prompt e passa a funcionar igual nos dois. **Vale criar o modal de entrada genérico**, já que não existe nenhum e outras features vão precisar |
-| **Eco absurdo ao compartilhar tela com som** (app desktop) | O `loopback` do Electron captura a mistura de saída do sistema **inteiro**, que inclui o próprio Backspace tocando a voz dos outros. Essa voz volta para eles dentro da transmissão, com atraso. Não é eco acústico — é digital, então **fone não resolve**. As constraints do web (`restrictOwnAudio`) nunca chegam: o handler ignora o `_request` e monta o stream a partir do enum. É também por isso que `shareAudio` já vinha desligado por padrão no app | Sem solução limpa no Electron: nem `loopback` nem `loopbackWithMute` excluem o áudio do próprio app. Opções reais na seção abaixo |
+| **Eco absurdo ao compartilhar tela com som** (app desktop) | O `loopback` do Electron captura a mistura de saída do sistema **inteiro**, que inclui o próprio Backspace tocando a voz dos outros. Essa voz volta para eles dentro da transmissão, com atraso. Não é eco acústico — é digital, então **fone não resolve**. As constraints do web (`restrictOwnAudio`) nunca chegam: o handler ignora o `_request` e monta o stream a partir do enum. É também por isso que `shareAudio` já vinha desligado por padrão no app | **Implementado (2026-09-01), não testado.** Trocado por `electron-native-screenshare`: captura com isolamento por processo — só a janela compartilhada (include) ou tudo menos este app (exclude). O PCM vai do processo principal ao renderer por IPC e vira faixa publicada no LiveKit. **Exige gerar instalador novo e reinstalar**, e só pode ser confirmado numa máquina Windows |
 | **Bloco do Spotify dessincronizado** (mostra a faixa errada por um tempo) | Soma de duas esperas: a consulta roda a cada **20s** (`useSpotifyActivity`) e o envio pelo WebSocket ainda passa por um **debounce de 5s** no `activityStore`. Na pior hipótese os outros veem a música anterior por ~25s | Consultar de novo perto do fim da faixa (a duração é conhecida) em vez de só por intervalo fixo, e encurtar o debounce para esta fonte |
 | **Bloco do Spotify some sozinho** | Três caminhos apagam a atividade inteira: faixa **pausada** (`is_playing: false` devolve `null`), o **vão entre faixas** (o Spotify responde `204`) e qualquer falha transitória. Some e volta = a piscada que você viu | Manter a última faixa conhecida por alguns segundos antes de apagar, e enviar um estado *pausado* explícito em vez de sumir com o bloco |
 | **Barra de progresso errada / andando pausada** | Duas causas independentes: (1) o progresso é derivado de carimbos calculados com o relógio do **servidor** e desenhado contra o relógio de **quem olha** — se os relógios divergem, a barra fica deslocada; (2) a barra continua avançando localmente depois que a pessoa pausa, até a próxima consulta | Enviar o horário do servidor junto no payload para o cliente corrigir a diferença, e congelar a barra quando o estado for pausado |
@@ -21,6 +21,7 @@
    "postinstall": "electron-rebuild -f -w uiohook-napi || node -e \"console.warn('[desktop] uiohook-napi native rebuild skipped - needs build tools (make, g++, python3). Only required to RUN the desktop app; the server, web client, and Docker image are unaffected.')\""
  },
  "dependencies": {
+    "electron-native-screenshare": "^1.2.0",
    "electron-updater": "^6.3.0",
    "uiohook-napi": "^1.5.5"
  },
@@ -265,6 +265,102 @@ function applyLoginItemSettings(openAtLogin: boolean, startMinimized: boolean):
  }
 }

+// ─── Native system-audio capture ────────────────────────────────────────────
+//
+// Electron's own `audio: 'loopback'` captures the whole output mix, which
+// includes this app playing everyone else's voices — so those voices went back
+// out inside the screen share and every listener heard themselves. Not acoustic
+// echo: it is a digital copy of the output, so headphones never helped.
+//
+// This module captures with process-level isolation instead: only the shared
+// window (include mode), or everything except this app (exclude mode).
+interface NativeAudioMeta {
+  sampleRate: number;
+  channels: number;
+  bitsPerSample: number;
+  isFloat: boolean;
+}
+
+interface NativeScreenShareAudio {
+  startCapture(processId?: number, isIncludeMode?: boolean, onData?: (data: Buffer, meta: NativeAudioMeta) => void): boolean;
+  stopCapture(): boolean;
+  getPidFromWindowHandle(windowHandle: number): number;
+  isAvailable(): boolean;
+  getLoadError(): string | null;
+}
+
+let nativeAudio: NativeScreenShareAudio | null = null;
+try {
+  // Required lazily and defensively: a native module that fails to load must
+  // degrade to sharing without audio, never stop the app from starting.
+  nativeAudio = require('electron-native-screenshare') as NativeScreenShareAudio;
+  if (!nativeAudio.isAvailable()) {
+    console.warn('[Main:ScreenShare] Native audio unavailable:', nativeAudio.getLoadError());
+    nativeAudio = null;
+  }
+} catch (err) {
+  console.warn('[Main:ScreenShare] Native audio module missing:', err);
+  nativeAudio = null;
+}
+
+let nativeAudioActive = false;
+
+/**
+ * Windows and Linux hand desktopCapturer ids of the form `window:<handle>:<n>`.
+ * Recovering the handle lets us capture only that window's audio, which is
+ * better than excluding ourselves: a game's sound goes out, the rest of the
+ * desktop does not.
+ */
+function windowHandleFromSourceId(sourceId: string): number | null {
+  const match = /^window:(\d+)/.exec(sourceId);
+  if (!match) return null;
+  const handle = Number(match[1]);
+  return Number.isFinite(handle) && handle > 0 ? handle : null;
+}
+
+function startNativeAudioCapture(sourceId: string): boolean {
+  if (!nativeAudio || nativeAudioActive) return false;
+
+  let targetPid = process.pid;
+  let includeMode = false;
+
+  const handle = windowHandleFromSourceId(sourceId);
+  if (handle !== null) {
+    const pid = nativeAudio.getPidFromWindowHandle(handle);
+    // pid 0 means the handle did not resolve; fall back to excluding ourselves
+    // rather than capturing nothing.
+    if (pid > 0) {
+      targetPid = pid;
+      includeMode = true;
+    }
+  }
+
+  try {
+    const started = nativeAudio.startCapture(targetPid, includeMode, (data, meta) => {
+      if (!mainWindow || mainWindow.isDestroyed()) return;
+      mainWindow.webContents.send('native-audio-data', data, meta);
+    });
+    nativeAudioActive = started;
+    console.log('[Main:ScreenShare] Native audio', started ? 'started' : 'failed',
+      includeMode ? `(only pid ${targetPid})` : '(excluding self)');
+    return started;
+  } catch (err) {
+    console.error('[Main:ScreenShare] Native audio start failed:', err);
+    nativeAudioActive = false;
+    return false;
+  }
+}
+
+function stopNativeAudioCapture(): void {
+  if (!nativeAudio || !nativeAudioActive) return;
+  try {
+    nativeAudio.stopCapture();
+  } catch (err) {
+    console.warn('[Main:ScreenShare] Native audio stop failed:', err);
+  }
+  nativeAudioActive = false;
+}
+
 // ─── Tray Icon ──────────────────────────────────────────────────────────────

 function generateFallbackTrayIcon(): Electron.NativeImage {
@@ -594,6 +690,10 @@ function registerIpcHandlers(): void {
  ipcMain.handle('get-app-version', () => app.getVersion());

  // Screen share picker coordination (used by setDisplayMediaRequestHandler)
+  ipcMain.on('native-audio-stop', () => {
+    stopNativeAudioCapture();
+  });
+
  ipcMain.on('screen-share-selected', (_event, _sourceId: string | null, _shareAudio?: boolean) => {
    // Handled via ipcMain.once in the display media handler — this is just
    // a safety net to prevent unhandled-message warnings
@@ -953,7 +1053,19 @@ if (!gotTheLock) {
        //     `PulseaudioLoopbackForScreenShare` feature flag we enable above.
        //     Fails on PipeWire-only systems without pulse compat — the
        //     renderer catches that and toasts the user.
-        callback({ video: selected, ...(shareAudio ? { audio: 'loopback' } : {}) });
+        // Audio no longer rides on the Electron stream: `loopback` would put
+        // this app's own output (everyone else's voices) back into the share.
+        // The native module captures it separately, isolated by process, and
+        // the renderer turns it into the track LiveKit publishes.
+        if (shareAudio) {
+          const started = startNativeAudioCapture(sourceId);
+          if (!started) {
+            // Tell the renderer so it can say the share is going out silently,
+            // instead of the user assuming sound is included.
+            mainWindow?.webContents.send('native-audio-unavailable');
+          }
+        }
+        callback({ video: selected });
      } catch (err) {
        console.error('[Main:ScreenShare] Handler error:', err);
        // @ts-ignore — deny the request without crashing
@@ -65,6 +65,24 @@ contextBridge.exposeInMainWorld('backspace', {
  onScreenShareSources: (callback: (sources: unknown[]) => void) => {
    ipcRenderer.on('screen-share-sources', (_event, sources) => callback(sources));
  },
+  // Raw PCM from the native capture. Arrives ~50x/second; the renderer turns it
+  // into a MediaStreamTrack for LiveKit.
+  onNativeAudioData: (callback: (data: ArrayBuffer, meta: { sampleRate: number; channels: number; bitsPerSample: number; isFloat: boolean }) => void) => {
+    const listener = (_event: unknown, data: Uint8Array, meta: { sampleRate: number; channels: number; bitsPerSample: number; isFloat: boolean }) => {
+      // Copied out of the transferred buffer: reusing it across IPC messages
+      // would let a later chunk overwrite one still being read.
+      callback(data.buffer.slice(data.byteOffset, data.byteOffset + data.byteLength) as ArrayBuffer, meta);
+    };
+    ipcRenderer.on('native-audio-data', listener);
+    return () => ipcRenderer.off('native-audio-data', listener);
+  },
+  onNativeAudioUnavailable: (callback: () => void) => {
+    const listener = () => callback();
+    ipcRenderer.on('native-audio-unavailable', listener);
+    return () => ipcRenderer.off('native-audio-unavailable', listener);
+  },
+  stopNativeAudio: () => ipcRenderer.send('native-audio-stop'),
+
  selectScreenSource: (sourceId: string | null, shareAudio?: boolean) => {
    ipcRenderer.send('screen-share-selected', sourceId, shareAudio ?? true);
  },
@@ -61,6 +61,11 @@ interface BackspaceElectronAPI {
  // Screen share picker coordination
  onScreenShareSources: (callback: (sources: ElectronScreenSource[]) => void) => void;
  selectScreenSource: (sourceId: string | null, shareAudio?: boolean) => void;
+  onNativeAudioData?: (
+    callback: (data: ArrayBuffer, meta: { sampleRate: number; channels: number; bitsPerSample: number; isFloat: boolean }) => void,
+  ) => () => void;
+  onNativeAudioUnavailable?: (callback: () => void) => () => void;
+  stopNativeAudio?: () => void;

  // Instance URL management
  getInstanceUrl: () => Promise<string | null>;
@@ -0,0 +1,126 @@
+import { Room, Track, LocalAudioTrack } from 'livekit-client';
+import { isElectron } from '../platform/platform';
+
+/**
+ * Turns the desktop app's native system-audio capture into a track LiveKit can
+ * publish.
+ *
+ * The native module hands the main process raw PCM, which arrives here in
+ * chunks over IPC. Web Audio has no "push samples" input, so each chunk is
+ * scheduled as a buffer source on a running cursor — the standard way to play a
+ * live stream without gaps.
+ *
+ * Exists because Electron's `audio: 'loopback'` captures the whole output mix,
+ * this app's own playback included, so every listener heard themselves echo.
+ */
+class NativeScreenAudio {
+  private ctx: AudioContext | null = null;
+  private destination: MediaStreamAudioDestinationNode | null = null;
+  private track: LocalAudioTrack | null = null;
+  private unsubscribe: (() => void) | null = null;
+  private nextStartTime = 0;
+
+  /**
+   * Held ahead of the cursor so a late chunk does not land in the past and get
+   * dropped. Small enough not to be noticeable against the video.
+   */
+  private static readonly BUFFER_AHEAD_S = 0.08;
+  /** Beyond this the stream has stalled; resync rather than drift forever. */
+  private static readonly MAX_DRIFT_S = 0.5;
+
+  isActive(): boolean {
+    return this.track !== null;
+  }
+
+  async start(room: Room): Promise<boolean> {
+    if (!isElectron() || !window.backspace?.onNativeAudioData) return false;
+    if (this.track) return true;
+
+    this.ctx = new AudioContext();
+    this.destination = this.ctx.createMediaStreamDestination();
+    this.nextStartTime = 0;
+
+    this.unsubscribe = window.backspace.onNativeAudioData((data, meta) => {
+      this.enqueue(data, meta);
+    });
+
+    const mediaTrack = this.destination.stream.getAudioTracks()[0];
+    if (!mediaTrack) {
+      await this.stop();
+      return false;
+    }
+
+    this.track = new LocalAudioTrack(mediaTrack);
+    try {
+      await room.localParticipant.publishTrack(this.track, {
+        source: Track.Source.ScreenShareAudio,
+      });
+      return true;
+    } catch (err) {
+      console.error('[NativeScreenAudio] publish failed', err);
+      await this.stop();
+      return false;
+    }
+  }
+
+  private enqueue(data: ArrayBuffer, meta: { sampleRate: number; channels: number; isFloat: boolean; bitsPerSample: number }): void {
+    const ctx = this.ctx;
+    const destination = this.destination;
+    if (!ctx || !destination) return;
+
+    // Only float32 is handled: it is what every supported platform reports.
+    // Anything else is dropped rather than played as noise.
+    if (!meta.isFloat || meta.bitsPerSample !== 32) return;
+
+    const samples = new Float32Array(data);
+    const channels = Math.max(1, meta.channels);
+    const frames = Math.floor(samples.length / channels);
+    if (frames === 0) return;
+
+    const buffer = ctx.createBuffer(channels, frames, meta.sampleRate);
+    for (let channel = 0; channel < channels; channel++) {
+      const channelData = buffer.getChannelData(channel);
+      // Interleaved in, planar out. The bounds are computed from the buffer's
+      // own length, but the index signature is still optional under
+      // noUncheckedIndexedAccess — a truncated final chunk reads as silence
+      // rather than NaN, which would click.
+      for (let frame = 0; frame < frames; frame++) {
+        channelData[frame] = samples[frame * channels + channel] ?? 0;
+      }
+    }
+
+    const source = ctx.createBufferSource();
+    source.buffer = buffer;
+    source.connect(destination);
+
+    const now = ctx.currentTime;
+    if (this.nextStartTime < now || this.nextStartTime > now + NativeScreenAudio.MAX_DRIFT_S) {
+      this.nextStartTime = now + NativeScreenAudio.BUFFER_AHEAD_S;
+    }
+    source.start(this.nextStartTime);
+    this.nextStartTime += buffer.duration;
+  }
+
+  async stop(): Promise<void> {
+    this.unsubscribe?.();
+    this.unsubscribe = null;
+    window.backspace?.stopNativeAudio?.();
+
+    if (this.track) {
+      try {
+        this.track.stop();
+      } catch { /* already gone */ }
+      this.track = null;
+    }
+    this.destination = null;
+    if (this.ctx) {
+      try {
+        await this.ctx.close();
+      } catch { /* already closed */ }
+      this.ctx = null;
+    }
+    this.nextStartTime = 0;
+  }
+}
+
+export const nativeScreenAudio = new NativeScreenAudio();
@@ -1,5 +1,7 @@
 import { Room, Track, BackupCodecPolicy } from 'livekit-client';
 import { useVoiceStore } from '../stores/voiceStore';
+import { nativeScreenAudio } from './nativeScreenAudio';
+import { isElectron } from '../platform/platform';
 import type { ScreenShareConfig } from '../stores/voiceStore';
 import { getStreamingLimits } from '../stores/settingsStore';
 import { getPublisherPC, getMediaStreamTrack } from './livekitInternals';
@@ -253,7 +255,10 @@ export async function startScreenShare(room: Room): Promise<boolean> {
  try {
    // For native mode: omit resolution constraint to capture at display's full native resolution
    const captureOptions: any = {
-      audio: config.shareAudio ? {
+      // In the desktop app audio never comes through this stream: Electron
+      // ignores these constraints and its loopback would carry our own output
+      // back into the share. The native capture below supplies it instead.
+      audio: config.shareAudio && !isElectron() ? {
        // Chrome 141+: exclude this tab's own audio from system audio capture
        // @ts-ignore — restrictOwnAudio is not yet in all TS type definitions
        restrictOwnAudio: true,
@@ -287,6 +292,14 @@ export async function startScreenShare(room: Room): Promise<boolean> {
      return false;
    }

+    // Desktop only, and only after the video track exists: publishing the audio
+    // first would briefly show a share with sound and no picture. A failure
+    // here leaves the share running silently rather than tearing it down.
+    if (isElectron() && config.shareAudio) {
+      const ok = await nativeScreenAudio.start(room);
+      if (!ok) console.warn('[SS] native system audio unavailable — sharing without sound');
+    }
+
    // Set content hint from builder (motion for gaming, detail for text)
    const screenPub = room.localParticipant.getTrackPublications()
      .find(p => p.source === Track.Source.ScreenShare);
@@ -416,6 +429,9 @@ function scheduleEncoderDetection(room: Room): void {
 // ---------------------------------------------------------------------------

 export async function stopScreenShare(room: Room): Promise<void> {
+  // Stopped first: leaving the native capture running would keep reading system
+  // audio after the share is gone.
+  await nativeScreenAudio.stop();
  try {
    await room.localParticipant.setScreenShareEnabled(false);
  } catch (err) {
@@ -0,0 +1,32 @@
+# Geradores dos efeitos sonoros
+
+Os `.ogg` em `packages/web/public/sounds/` para call e transmissão são
+sintetizados por estes scripts — nenhum arquivo de terceiros, nenhuma questão
+de licença num repositório público.
+
+Os parâmetros não foram escolhidos a gosto: saíram da **medição** de dois sons
+de referência fornecidos pelo dono da instância (análise de envelope, picos
+espectrais e tempo de decaimento). O timbre resultante é fundamental + oitava
+quase na mesma intensidade, com pouco acima disso:
+
+| Harmônico | Proporção medida |
+|---|---|
+| 1x (fundamental) | 1.00 |
+| 2x (oitava)      | 0.85 |
+| 3x               | 0.10 |
+| 4x               | 0.02 |
+
+Decaimento a 1/e: ~0,19 s. Sem reverberação.
+
+| Script | Gera | Notas |
+|---|---|---|
+| `build.py`    | `user_join`, `user_leave` | C4→G4 subindo / D4→G3 descendo |
+| `stream.py`   | `stream_user_joined`, `stream_user_left` | nota única (G4 / G3), ~55% do volume |
+| `started4.py` | `stream_started` | acorde de dó + golpe grave + sopro de ar |
+| `ended.py`    | `stream_ended` | arpejo descendo, resolve na raiz grave |
+
+Para regerar: `python3 <script>.py` produz `.wav`, depois converter com
+`ffmpeg -i X.wav -c:a libvorbis -q:a 5 -ar 48000 X.ogg`.
+
+Ajustes comuns: `HARM` muda o timbre, `tau` a duração, `step` a velocidade do
+arpejo, e o alvo em `write()` o volume (pico RMS).
@@ -0,0 +1,47 @@
+import math, wave, struct
+SR=48000
+# Proporcoes medidas nas referencias: fundamental, oitava, 3o, 4o
+HARM=(1.00, 0.85, 0.10, 0.02)
+TAU=0.19          # decaimento a 1/e medido
+ATTACK=0.004
+
+def note(f, dur):
+    n=int(SR*dur); o=[]
+    for i in range(n):
+        t=i/SR
+        s=sum(a*math.sin(2*math.pi*f*(k+1)*t) for k,a in enumerate(HARM))
+        o.append(s*min(1.0,t/ATTACK)*math.exp(-t/TAU))
+    return o
+
+def mix(layers,total):
+    b=[0.0]*total
+    for off,s in layers:
+        for i,v in enumerate(s):
+            if off+i<total: b[off+i]+=v
+    return b
+
+def peak_rms(x, win=int(SR*0.01)):
+    best=0.0
+    for i in range(0,len(x)-win,win):
+        seg=x[i:i+win]
+        best=max(best, math.sqrt(sum(v*v for v in seg)/len(seg)))
+    return best
+
+def write(name, buf, target_rms):
+    cur=peak_rms(buf)
+    g=target_rms/max(1e-9,cur)
+    # nao deixa estourar mesmo casando o volume da referencia
+    pk=max(abs(s*g) for s in buf)
+    if pk>0.97: g*=0.97/pk
+    with wave.open(name,'w') as w:
+        w.setnchannels(1); w.setsampwidth(2); w.setframerate(SR)
+        w.writeframes(b''.join(struct.pack('<h',int(max(-1,min(1,s*g))*32767)) for s in buf))
+    print(f"{name}: pico RMS {peak_rms([s*g for s in buf]):.3f}, {len(buf)/SR:.2f}s")
+
+C4,D4,G3,G4 = 261.63, 293.66, 196.00, 392.00
+GAP=int(SR*0.10)   # 100ms entre as notas, medido
+
+# join: quinta subindo — mesmas notas da referencia
+write('user_join.wav',  mix([(0,note(C4,0.55)),(GAP,note(G4,0.55))], int(SR*0.62)), 0.365)
+# leave: quinta descendo
+write('user_leave.wav', mix([(0,note(D4,0.55)),(GAP,note(G3,0.55))], int(SR*0.62)), 0.440)
@@ -0,0 +1,53 @@
+import math, wave, struct
+SR=48000
+HARM=(1.00, 0.85, 0.10, 0.02)
+
+def note(f, dur, tau, amp=1.0):
+    n=int(SR*dur); o=[]
+    for i in range(n):
+        t=i/SR
+        s=sum(a*math.sin(2*math.pi*f*(k+1)*t) for k,a in enumerate(HARM))
+        o.append(s*amp*min(1.0,t/0.004)*math.exp(-t/tau))
+    return o
+
+def thump(f, dur, amp):
+    n=int(SR*dur); o=[]; ph=0.0
+    for i in range(n):
+        t=i/SR
+        ph += 2*math.pi*(f*(1+1.2*math.exp(-t/0.02)))/SR
+        o.append(math.sin(ph)*amp*math.exp(-t/(dur*0.30)))
+    return o
+
+def mix(layers,total):
+    b=[0.0]*total
+    for off,s in layers:
+        for i,v in enumerate(s):
+            if off+i<total: b[off+i]+=v
+    return b
+
+def peak_rms(x, win=int(SR*0.01)):
+    return max(math.sqrt(sum(v*v for v in x[i:i+win])/win) for i in range(0,len(x)-win,win))
+
+def write(name, buf, target):
+    g=target/max(1e-9,peak_rms(buf))
+    pk=max(abs(s*g) for s in buf)
+    if pk>0.97: g*=0.97/pk
+    with wave.open(name,'w') as w:
+        w.setnchannels(1); w.setsampwidth(2); w.setframerate(SR)
+        w.writeframes(b''.join(struct.pack('<h',int(max(-1,min(1,s*g))*32767)) for s in buf))
+    print(f"{name}: pico RMS {peak_rms([s*g for s in buf]):.3f}, {len(buf)/SR:.2f}s")
+
+C3,C4,E4,G4,C5 = 130.81,261.63,329.63,392.00,523.25
+step=int(SR*0.033)
+
+# Espelho do inicio: o arpejo DESCE e resolve na raiz grave, que entra por
+# ultimo. Sem o sopro de ar — brilho sugere chegada, nao encerramento.
+# Mais curto e mais baixo: encerrar merece menos destaque que comecar.
+write('stream_ended.wav', mix([
+    (0,      note(C5, 0.42, 0.15, 0.70)),
+    (step,   note(G4, 0.44, 0.16)),
+    (2*step, note(E4, 0.46, 0.17)),
+    (3*step, note(C4, 0.48, 0.19)),
+    (3*step, note(C3, 0.50, 0.21, 0.80)),
+    (3*step, thump(85, 0.13, 0.40)),
+], int(SR*0.55)), 0.34)
@@ -0,0 +1,65 @@
+import math, random, wave, struct
+SR=48000; random.seed(11)
+HARM=(1.00, 0.85, 0.10, 0.02)
+
+def note(f, dur, tau, amp=1.0):
+    n=int(SR*dur); o=[]
+    for i in range(n):
+        t=i/SR
+        s=sum(a*math.sin(2*math.pi*f*(k+1)*t) for k,a in enumerate(HARM))
+        o.append(s*amp*min(1.0,t/0.004)*math.exp(-t/tau))
+    return o
+
+def thump(f, dur, amp):
+    """Instrumento 2: golpe grave com queda de tom. Da impacto no ataque e
+    morre em 90ms — engorda sem alongar."""
+    n=int(SR*dur); o=[]; ph=0.0
+    for i in range(n):
+        t=i/SR
+        ph += 2*math.pi*(f*(1+1.2*math.exp(-t/0.02)))/SR
+        o.append(math.sin(ph)*amp*math.exp(-t/(dur*0.30)))
+    return o
+
+def air(dur, amp, cutoff=5000):
+    """Instrumento 3: sopro curto de ruido filtrado. Brilho no inicio, sem cauda."""
+    n=int(SR*dur)
+    y=0.0; a=1-math.exp(-2*math.pi*cutoff/SR); o=[]
+    for i in range(n):
+        t=i/SR
+        y += a*(random.uniform(-1,1)-y)
+        o.append(y*amp*min(1.0,t/0.008)*math.exp(-t/(dur*0.22)))
+    return o
+
+def mix(layers,total):
+    b=[0.0]*total
+    for off,s in layers:
+        for i,v in enumerate(s):
+            if off+i<total: b[off+i]+=v
+    return b
+
+def peak_rms(x, win=int(SR*0.01)):
+    return max(math.sqrt(sum(v*v for v in x[i:i+win])/win) for i in range(0,len(x)-win,win))
+
+def write(name, buf, target):
+    g=target/max(1e-9,peak_rms(buf))
+    pk=max(abs(s*g) for s in buf)
+    if pk>0.97: g*=0.97/pk
+    with wave.open(name,'w') as w:
+        w.setnchannels(1); w.setsampwidth(2); w.setframerate(SR)
+        w.writeframes(b''.join(struct.pack('<h',int(max(-1,min(1,s*g))*32767)) for s in buf))
+    print(f"{name}: pico RMS {peak_rms([s*g for s in buf]):.3f}, {len(buf)/SR:.2f}s")
+
+C3,G3,C4,E4,G4,C5 = 130.81,196.00,261.63,329.63,392.00,523.25
+step=int(SR*0.033)   # arpejo bem rapido: acorde montado em 100ms
+
+# Decaimentos ~metade dos anteriores (0.20 vs 0.38). Total 0.62s vs 1.08s.
+write('stream_started.wav', mix([
+    (0,      thump(85, 0.14, 0.55)),        # impacto
+    (0,      air(0.13, 0.16)),              # brilho
+    (0,      note(C3, 0.60, 0.22, 0.80)),
+    (0,      note(G3, 0.58, 0.20, 0.45)),
+    (0,      note(C4, 0.56, 0.20)),
+    (step,   note(E4, 0.54, 0.19)),
+    (2*step, note(G4, 0.52, 0.18)),
+    (3*step, note(C5, 0.50, 0.16, 0.70)),
+], int(SR*0.62)), 0.42)
@@ -0,0 +1,31 @@
+import math, wave, struct
+SR=48000
+HARM=(1.00, 0.85, 0.10, 0.02)   # mesmo timbre medido nas referencias
+ATTACK=0.004
+
+def note(f, dur, tau):
+    n=int(SR*dur); o=[]
+    for i in range(n):
+        t=i/SR
+        s=sum(a*math.sin(2*math.pi*f*(k+1)*t) for k,a in enumerate(HARM))
+        o.append(s*min(1.0,t/ATTACK)*math.exp(-t/tau))
+    return o
+
+def peak_rms(x, win=int(SR*0.01)):
+    return max(math.sqrt(sum(v*v for v in x[i:i+win])/win)
+               for i in range(0,len(x)-win,win))
+
+def write(name, buf, target):
+    g=target/max(1e-9,peak_rms(buf))
+    pk=max(abs(s*g) for s in buf)
+    if pk>0.97: g*=0.97/pk
+    with wave.open(name,'w') as w:
+        w.setnchannels(1); w.setsampwidth(2); w.setframerate(SR)
+        w.writeframes(b''.join(struct.pack('<h',int(max(-1,min(1,s*g))*32767)) for s in buf))
+    print(f"{name}: pico RMS {peak_rms([s*g for s in buf]):.3f}, {len(buf)/SR:.2f}s")
+
+G3, G4 = 196.00, 392.00
+# Decaimento mais curto (0.13 vs 0.19) e volume ~55% do som de call:
+# entrar/sair da tela dispara muitas vezes durante uma transmissao.
+write('stream_user_joined.wav', note(G4, 0.42, 0.13), 0.20)
+write('stream_user_left.wav',   note(G3, 0.42, 0.13), 0.20)