diff --git a/openless-all/app/scripts/audio-cue-listener-contract.test.mjs b/openless-all/app/scripts/audio-cue-listener-contract.test.mjs
new file mode 100644
index 000000000..319d18c27
--- /dev/null
+++ b/openless-all/app/scripts/audio-cue-listener-contract.test.mjs
@@ -0,0 +1,19 @@
+import assert from 'node:assert/strict';
+import { readFile } from 'node:fs/promises';
+
+const source = await readFile(new URL('../src/components/AudioCue.tsx', import.meta.url), 'utf8');
+
+// The always-alive capsule listener must never open output just by mounting, even when
+// the persisted preference is disabled. Runtime playback/recovery is covered separately.
+assert.doesNotMatch(source, /primeAudioCue/);
+assert.equal((source.match(/playRecordStartCue\(\)/g) ?? []).length, 1);
+assert.match(source, /if \(audioCueEnabledRef\.current\) playRecordStartCue\(\)/);
+assert.match(source, /state === 'recording' && prev !== 'recording'/);
+assert.match(source, /!isAndroid\(\) && nativeCueOwner === false/);
+assert.equal(
+  (source.match(/if \(!audioCueEnabledRef\.current\) disposeAudioCue\(\)/g) ?? []).length,
+  2,
+  'both initial settings and live preference changes must dispose disabled audio',
+);
+assert.match(source, /return \(\) => disposeAudioCue\(\)/);
+console.log('[audioCue.listener.test] lazy startup, preference and ownership contracts passed');
diff --git a/openless-all/app/scripts/audio-cue-runtime.test.mjs b/openless-all/app/scripts/audio-cue-runtime.test.mjs
index f854901b2..3472ba603 100644
--- a/openless-all/app/scripts/audio-cue-runtime.test.mjs
+++ b/openless-all/app/scripts/audio-cue-runtime.test.mjs
@@ -1,7 +1,7 @@
 import assert from 'node:assert/strict';
 import { mock } from 'node:test';
 import { tsImport } from 'tsx/esm/api';
-const { playRecordStartCue, stopAudioCue } = await tsImport(
+const { disposeAudioCue, playRecordStartCue, stopAudioCue } = await tsImport(
   '../src/lib/audioCue.ts',
   import.meta.url,
 );
@@ -21,6 +21,7 @@ class FakeContext {
     this.state = FakeContext.initialState;
     this.frozen = false;
     this.voices = 0;
+    this.oscillators = [];
     this.closeCalls = 0;
     this.resumeImpl = FakeContext.initialResumeImpl;
     this.destination = {};
@@ -36,13 +37,15 @@ class FakeContext {
   }
   createOscillator() {
     this.voices++;
-    return {
+    const oscillator = {
       frequency: { setValueAtTime() {} },
       connect: (gain) => gain,
       start() {},
       stop() {},
       disconnect() {},
     };
+    this.oscillators.push(oscillator);
+    return oscillator;
   }
   createGain() {
     return {
@@ -80,7 +83,6 @@ try {
   releaseOldResume();
   await Promise.resolve();
   assert.equal(stuck.voices, 0, 'late resolution cannot play through a discarded context');
-  tick(315);
   // Every request invalidates old work, including a new synchronous playback.
   healthy.state = 'suspended';
   let releaseSuperseded;
@@ -91,14 +93,12 @@ try {
   releaseSuperseded();
   await Promise.resolve();
   assert.equal(healthy.voices, 4, 'a superseded resume must not duplicate the newer cue');
-  tick(315);
   healthy.frozen = true;
   playRecordStartCue();
   tick(315);
   const recovered = FakeContext.instances[2];
   assert.equal(healthy.closeCalls, 1, 'running with a frozen clock must recover');
   assert.equal(recovered.voices, 2);
-  tick(315);
   // The original stop/time boundary survives recreation, so a late retry stays silent.
   recovered.state = 'suspended';
   playRecordStartCue();
@@ -109,7 +109,6 @@ try {
   playRecordStartCue();
   const next = FakeContext.instances[3];
   assert.equal(next.voices, 2, 'the next recording still works after a dropped cue');
-  tick(315);
   next.state = 'suspended';
   next.resumeImpl = () => {
     throw new Error('audio interrupted');
@@ -132,21 +131,116 @@ try {
   releaseRetry();
   await Promise.resolve();
   assert.equal(retry.voices, 0, 'a recreated context cannot reset the late-cue deadline');
+  assert.equal(retry.closeCalls, 1, 'a late successful resume must release its silent output');
+  FakeContext.initialState = 'running';
   playRecordStartCue();
-  assert.equal(retry.voices, 2, 'dropping the old cue must not disable the next recording');
-  tick(315);
+  const fresh = FakeContext.instances[7];
+  assert.equal(fresh.voices, 2, 'dropping the old cue must not disable the next recording');
 
-  retry.state = 'suspended';
-  retry.resumeImpl = () => Promise.reject(new Error('device changed'));
+  fresh.state = 'suspended';
+  fresh.resumeImpl = () => Promise.reject(new Error('device changed'));
   FakeContext.initialState = 'running';
   playRecordStartCue();
   await Promise.resolve();
-  assert.equal(retry.closeCalls, 1, 'asynchronous resume rejection also releases the context');
-  assert.equal(FakeContext.instances[7].voices, 2, 'device-change recovery plays only one cue');
+  assert.equal(fresh.closeCalls, 1, 'asynchronous resume rejection also releases the context');
+  const finalCue = FakeContext.instances[8];
+  assert.equal(finalCue.voices, 2, 'device-change recovery plays only one cue');
   tick(315);
+  assert.equal(finalCue.closeCalls, 1, 'natural cue completion releases the output device');
+
+  // No startup priming: import/disposal does not create a context. Every new preview or
+  // recording can still play after the previous context was closed.
+  const idleCount = FakeContext.instances.length;
+  disposeAudioCue();
+  tick(1000);
+  assert.equal(FakeContext.instances.length, idleCount, 'idle cleanup never creates audio');
+  playRecordStartCue();
+  const stopped = FakeContext.instances.at(-1);
+  assert.equal(stopped.voices, 2);
+  stopAudioCue();
+  tick(30);
+  assert.equal(stopped.closeCalls, 1, 'stopping a playing cue releases output after its ramp');
+
+  playRecordStartCue();
+  const rapid = FakeContext.instances.at(-1);
+  stopAudioCue();
+  tick(10);
+  playRecordStartCue();
+  tick(20);
+  assert.equal(rapid.closeCalls, 0, 'old stop cleanup cannot close a newer cue');
+  tick(295);
+  assert.equal(rapid.closeCalls, 1, 'the newer cue also releases output when finished');
+
+  // Disabling the preference / unmounting is stronger than a recording stop: a pending
+  // resume must never revive audio or start a recovery retry after disposal.
+  let releaseDisposed;
+  FakeContext.initialState = 'suspended';
+  FakeContext.initialResumeImpl = () => new Promise((resolve) => (releaseDisposed = resolve));
+  playRecordStartCue();
+  const disposed = FakeContext.instances.at(-1);
+  disposeAudioCue();
+  const disposedCount = FakeContext.instances.length;
+  disposed.state = 'running';
+  releaseDisposed();
+  await Promise.resolve();
+  tick(1000);
+  assert.equal(disposed.closeCalls, 1);
+  assert.equal(disposed.voices, 0, 'disabled/unmounted listeners cannot play a late cue');
+  assert.equal(FakeContext.instances.length, disposedCount, 'disposal cancels recovery timers');
+
+  // A quick recording still gets feedback when resume settles within the original deadline.
+  let releaseQuick;
+  FakeContext.initialResumeImpl = () => new Promise((resolve) => (releaseQuick = resolve));
+  playRecordStartCue();
+  const quick = FakeContext.instances.at(-1);
+  stopAudioCue();
+  tick(100);
+  quick.state = 'running';
+  releaseQuick();
+  await Promise.resolve();
+  assert.equal(quick.voices, 2, 'quick recordings keep start feedback without startup priming');
+  tick(315);
+  assert.equal(quick.closeCalls, 1);
+
+  FakeContext.initialState = 'running';
+  playRecordStartCue();
+  const ended = FakeContext.instances.at(-1);
+  ended.oscillators[0].onended();
+  assert.equal(ended.closeCalls, 0, 'one ending tone cannot cut off the remaining tone');
+  ended.oscillators[1].onended();
+  assert.equal(ended.closeCalls, 1, 'the final tone releases output without waiting for a timer');
+
+  playRecordStartCue();
+  const superseded = FakeContext.instances.at(-1);
+  const oldVoices = [...superseded.oscillators];
+  playRecordStartCue();
+  for (const voice of oldVoices) voice.onended();
+  assert.equal(superseded.closeCalls, 0, 'old onended callbacks cannot close a newer cue');
+  tick(315);
+  assert.equal(superseded.closeCalls, 1);
+
+  playRecordStartCue();
+  const disabledWhilePlaying = FakeContext.instances.at(-1);
+  disposeAudioCue();
+  assert.equal(disabledWhilePlaying.closeCalls, 1, 'disabling an active cue releases immediately');
+  tick(1000);
+  assert.equal(disabledWhilePlaying.closeCalls, 1, 'stale timers cannot dispose twice');
+
+  const originalCreateOscillator = FakeContext.prototype.createOscillator;
+  FakeContext.prototype.createOscillator = () => {
+    throw new Error('audio node creation failed');
+  };
+  try {
+    assert.doesNotThrow(playRecordStartCue);
+    const failedSynthesis = FakeContext.instances.at(-1);
+    tick(315);
+    assert.equal(failedSynthesis.closeCalls, 1, 'failed synthesis cannot leak a silent output');
+  } finally {
+    FakeContext.prototype.createOscillator = originalCreateOscillator;
+  }
   console.log('[audioCue.runtime.test] playback recovery assertions passed');
 } finally {
-  stopAudioCue();
+  disposeAudioCue();
   mock.restoreAll();
   mock.timers.reset();
   if (previousWindow) Object.defineProperty(globalThis, 'window', previousWindow);
diff --git a/openless-all/app/src/components/AudioCue.tsx b/openless-all/app/src/components/AudioCue.tsx
index e8ca5a425..5de5084b9 100644
--- a/openless-all/app/src/components/AudioCue.tsx
+++ b/openless-all/app/src/components/AudioCue.tsx
@@ -5,7 +5,7 @@
 
 import { useEffect, useRef, useState } from 'react';
 import { isAndroid, isTauri } from '../lib/ipc';
-import { playRecordStartCue, primeAudioCue, stopAudioCue } from '../lib/audioCue';
+import { disposeAudioCue, playRecordStartCue, stopAudioCue } from '../lib/audioCue';
 import type { CapsuleState, UserPreferences } from '../lib/types';
 
 interface CapsulePayload {
@@ -48,7 +48,10 @@ export function AudioCueListener() {
       try {
         const { getSettings } = await import('../lib/ipc');
         const prefs = await getSettings();
-        if (!cancelled) audioCueEnabledRef.current = prefs.audioCueOnRecord !== false;
+        if (!cancelled) {
+          audioCueEnabledRef.current = prefs.audioCueOnRecord !== false;
+          if (!audioCueEnabledRef.current) disposeAudioCue();
+        }
       } catch {
         // Read failure: keep the default true
       }
@@ -67,7 +70,10 @@ export function AudioCueListener() {
       const { listen } = await import('@tauri-apps/api/event');
       listen<UserPreferences>('prefs:changed', (event) => {
         const next = event.payload;
-        if (next) audioCueEnabledRef.current = next.audioCueOnRecord !== false;
+        if (next) {
+          audioCueEnabledRef.current = next.audioCueOnRecord !== false;
+          if (!audioCueEnabledRef.current) disposeAudioCue();
+        }
       })
         .then((fn) => {
           if (cancelled) fn();
@@ -81,10 +87,10 @@ export function AudioCueListener() {
     };
   }, [audioCueRuntimeEnabled]);
 
-  // Prime the AudioContext
+  // Do not prime on mount: even a silent running context holds the output device open.
+  // Dispose pending resumes as well as active voices when the owner changes or unmounts.
   useEffect(() => {
-    if (!isTauri || !audioCueRuntimeEnabled) return;
-    primeAudioCue();
+    return () => disposeAudioCue();
   }, [audioCueRuntimeEnabled]);
 
   // Listen for capsule state edges
diff --git a/openless-all/app/src/lib/audioCue.ts b/openless-all/app/src/lib/audioCue.ts
index 967fd6a34..f3787a415 100644
--- a/openless-all/app/src/lib/audioCue.ts
+++ b/openless-all/app/src/lib/audioCue.ts
@@ -69,6 +69,12 @@ let stopSeq = 0;
 const DEFERRED_CUE_LATE_THRESHOLD_MS = 400;
 const RESUME_TIMEOUT_MS = 250;
 let recoveryTimer: ReturnType<typeof setTimeout> | undefined;
+let stopTimer: ReturnType<typeof setTimeout> | undefined;
+
+function clearStopTimer(): void {
+  clearTimeout(stopTimer);
+  stopTimer = undefined;
+}
 
 function clearRecoveryTimer(): void {
   clearTimeout(recoveryTimer);
@@ -217,13 +223,34 @@ function stopVoices(): void {
 /** Close/stop the cue: stop playing nodes and mark "a stop happened meanwhile" so pending resume callbacks can judge lateness. */
 export function stopAudioCue(): void {
   stopSeq++;
+  const ctx = sharedCtx;
+  const wasPlaying = activeVoices.length > 0;
   stopVoices();
+  // Allow the short release ramp to finish. A pending resume keeps its existing bounded
+  // recovery window so quick recordings still receive their start feedback.
+  if (ctx && wasPlaying) {
+    clearRecoveryTimer();
+    clearStopTimer();
+    const myPlay = playSeq;
+    stopTimer = setTimeout(() => {
+      if (myPlay === playSeq && sharedCtx === ctx) discardContext(ctx);
+    }, 30);
+  }
+}
+
+/** Disable/unmount: cancel even pending resumes, without creating an audio context. */
+export function disposeAudioCue(): void {
+  playSeq++;
+  stopSeq++;
+  clearRecoveryTimer();
+  clearStopTimer();
+  if (sharedCtx) discardContext(sharedCtx);
 }
 
 // Actually schedule the synthesis nodes. Must be called while the AudioContext is running
 // (not suspended): when suspended, currentTime is frozen at the pause moment, so nodes would
 // be scheduled at stale timestamps -> silent and piling up.
-function scheduleCueVoices(ctx: AudioContext): void {
+function scheduleCueVoices(ctx: AudioContext, onComplete: () => void): void {
   // On rapid hotkey presses, stop the previous round first to keep overlap from building up.
   // Use stopVoices, not stopAudioCue: this must not invalidate its own generation.
   stopVoices();
@@ -255,6 +282,7 @@ function scheduleCueVoices(ctx: AudioContext): void {
         } catch {
           // noop
         }
+        if (activeVoices.length === 0) onComplete();
       };
     } catch {
       // One tone failing to create/schedule must not affect the rest.
@@ -262,32 +290,10 @@ function scheduleCueVoices(ctx: AudioContext): void {
   }
 }
 
-/**
- * Prime the AudioContext: create and resume early so the ctx is already running when recording
- * actually starts and playRecordStartCue can schedule synchronously, bypassing the
- * suspended->resume async race — the root cause of lost cues on quick recordings.
- * Note: under a strict autoplay policy a resume without a user gesture may be rejected
- * (silently degraded), in which case priming does nothing and the playSeq/stopSeq + late
- * threshold fallback in playRecordStartCue applies. Tauri WebViews are usually lenient about
- * the first resume, so priming typically lets the common path take the synchronous branch.
- * Call once when the capsule window mounts.
- */
-export function primeAudioCue(): void {
-  // getContext already rebuilds a closed ctx; here we only need to wake suspended/interrupted.
-  const ctx = getContext();
-  if (!ctx) return;
-  if (audioContextActionForState(ctx.state) === 'resume') {
-    try {
-      void ctx.resume().catch(() => undefined);
-    } catch {
-      discardContext(ctx);
-    }
-  }
-}
-
 /** Play the "recording started" cue. Silently degrades when Web Audio is unavailable or the context is suspended and cannot be resumed. */
 export function playRecordStartCue(): void {
   clearRecoveryTimer();
+  clearStopTimer();
   const myPlay = ++playSeq;
   const stopAtRequest = stopSeq;
   const requestedAt = nowMs();
@@ -318,12 +324,21 @@ function playRecordStartCueOnce(
   };
   const schedule = () => {
     const startedAt = ctx.currentTime;
-    scheduleCueVoices(ctx);
+    scheduleCueVoices(ctx, () => {
+      if (myPlay !== playSeq || sharedCtx !== ctx) return;
+      clearRecoveryTimer();
+      discardContext(ctx);
+    });
     // Some audio interruptions still report running while the audio clock is frozen;
     // proactively retire the ctx before the next keypress.
     recoveryTimer = setTimeout(
       () => {
+        if (myPlay !== playSeq || sharedCtx !== ctx) return;
         if (ctx.currentTime <= startedAt) recover();
+        // A connected, silent AudioContext still holds the system output device (and on
+        // macOS prevents idle sleep). Close after the cue tail and clock-health check,
+        // including when onended is delayed or synthesis failed. Next play is lazy.
+        else discardContext(ctx);
       },
       cueTotalDurationMs(recordStartCueTones()) + 50,
     );
@@ -351,6 +366,9 @@ function playRecordStartCueOnce(
     } else if (!runningAfterResume) {
       // Even if this round is already late, clear the bad context so the next recording can recover.
       recover();
+    } else {
+      // A resumed but already-late cue must not leave a silent output stream alive.
+      discardContext(ctx);
     }
   };