diff --git a/LOTUS_TODO.md b/LOTUS_TODO.md index 92001885c..8f413bbf8 100644 --- a/LOTUS_TODO.md +++ b/LOTUS_TODO.md @@ -106,7 +106,7 @@ Per-slice bug hunt (5 agents: theming · calls · messaging · threads/presence/ - [x] **[Med]** `DenoiseTester.play()` (Settings → Calls A/B model test) leaks the denoise model node — calls `ctx.close()` but never `denoise.dispose()` (inconsistent with `stopLive`, which disposes) → leaks the DeepFilterNet/DTLN worker/WASM per press. **FIXED** (`c9d9d914`): `stopPlayback` now mirrors `stopLive` (dispose model + gate), and a generation token also closes the rapid-click / stop-during-load / unmount-during-load leak windows (3 review passes, all 6 interleavings traced). - [ ] **[Med] [live]** PiP auto-spotlight never released on return to the call room — the release branch sits inside the `if (!pipMode) return` guard, so screenshare→PiP→back leaves spotlight forced on and `pipAutoSpotlightRef` stuck `true`. `CallEmbedProvider.tsx:733-744`. - [x] **[Low]** DenoiseTester async paths (`getUserMedia`) have no mounted-guard → ctx/stream leak + setState-after-unmount if Settings closes during the mic prompt. **FIXED** (`c9d9d914`): a `mountedRef` guards `startLive`/`startRecord` after the `getUserMedia` await (and `play()` after its model load); the ref is set on mount, not only cleared on unmount, so it survives a StrictMode/Activity remount. -- [ ] **[Low]** Soundboard 30s safety timeout never cleared on natural clip end (`CallSoundboard.tsx:115`); `PrescreenControls` `PermissionStatus.onchange` not removed on unmount (`PrescreenControls.tsx:22-28`). +- [x] **[Low]** Soundboard 30s safety timeout never cleared on natural clip end (`CallSoundboard.tsx:115`); `PrescreenControls` `PermissionStatus.onchange` not removed on unmount (`PrescreenControls.tsx:22-28`). **FIXED** (`56561627`): per-play timer token cleared on end/unmount (identity-guarded so a stale clip can't disarm a newer one); permission `onchange` detached + `cancelled`-guarded setState. - [ ] **[Low] [live]** Call-to-call switch disposes the embed without an explicit `HangupCall` → possible transient ghost RTC membership until EC's unload-leave fires. **Theming / visuals**