### What happens

The Omarchy shell (`quickshell -n -p /usr/share/omarchy/shell`) segfaults while the audio panel rebuilds its device lists. Quickshell restarts on its own, so the bar blinks and comes back. I've had six crashes with an identical stack between 2026-09-23 and 2026-10-02.

### Stack (same in all six cores)

```
#4  libQt6QmlModels.so.6 + 0xb3530          <- SIGSEGV
#5  libQt6QmlModels.so.6 + 0xb3f01
#6  libQt6Qml.so.6 + 0x3880a7
#7  QV4::QObjectWrapper::getQmlProperty
#8  QV4::QQmlContextWrapper::getPropertyAndBase
#9  QV4::ExecutionContext::getProperty
#10 QV4::Runtime::TypeofName::call           <- top-level `typeof module` in Model.js
#11 (JIT)
#13 QV4::Module::evaluate
#15 QQmlEnginePrivate::createInternalContext  <- Model.js evaluated per delegate context
#17 QQmlObjectCreator::createInstance
#19 QQmlIncubatorPrivate::incubate
#21 libQt6QmlModels.so.6 + 0x8c87c
#22 QQuickRepeater::regenerate
#23 QQuickRepeater::setModel
...
#36 QV4::Runtime::StoreNameSloppy::call       <- displayAudioSinks = listSnapshot(...)
#46 QQmlBoundSignalExpression::evaluate       <- Timer onTriggered, or onOpenedChanged
```

Path: `refreshDisplayAudioModels()` in `plugins/panels/audio/Panel.qml` assigns `displayAudioSinks`, the `Repeater` at line 853 rebuilds, and a `SinkRow` delegate is incubated. Because `Model.js` isn't `.pragma library`, it's evaluated again in the new delegate context. Its top-level `typeof module` check (Model.js:236) looks names up through the delegate's context. That context's `modelData` QObject is already null, and the lookup crashes with a vtable call on null inside QmlModels. Frames #4-#6 have no debug symbols. The rest are from `coredumpctl info` and match address for address across all six cores.

### Trigger

The crashes happen whenever PipeWire nodes are removed while the panel is open or opening:

- 09-26: AirPods Max A2DP/HFP transports failing every ~20 s
- 09-29: a PipeWire error on an object, a metadata rebind, and the default sink going 73 -> null -> 82 -> 73 (no Bluetooth involved)
- 10-02: a browser audio stream ending (PwLink/PwNode objects destroyed just before the crash in the crashed instance's log)

The snapshot and 75 ms timer added to Panel.qml (the comment at lines 103-107) don't seem to be enough. The snapshot is a JS array of PwNode QObject pointers, and a node can still be destroyed after it's taken and before or during delegate incubation. A related suspect: `audioSinks`/`audioSources` (lines 93-94) fall back to `cachedAudioSinks`/`cachedAudioSources` whenever the live list is momentarily empty. Those cached arrays can hold nodes that PipeWire has already destroyed. The bluetooth panel doesn't hit this because its rows are built from primitives rather than live node objects.

### Possible fixes

- Snapshot primitives (id, name, description, volume, muted) into the Repeater model and look the live node up by id when the user acts on a row, as the bluetooth panel does.
- Make `Model.js` `.pragma library`, or drop the top-level `typeof module` check, so it isn't re-evaluated in every delegate context.
- Filter null entries out of the snapshot and cached arrays before handing them to the Repeater.

### Expected

Opening the audio panel, or audio devices and streams appearing and disappearing, shouldn't crash the shell.

### Steps to reproduce

Not deterministic. Keep the audio panel open, or open it, while sinks or streams are being added and removed (for example a flapping Bluetooth headset, switching the default sink, or starting and stopping browser audio).

### System

- Omarchy 4.0.4-1
- quickshell 0.3.1-1, qt6-declarative 6.11.2-1
- pipewire 1.6.8-1, wireplumber 0.5.17-1
- Hyprland 0.56.2, kernel 7.2.5-3-omarchy
- AMD Ryzen 7 9800X3D, ASUS TUF GAMING B850-E WIFI

Possibly related crashes with the same Qt frames but different QML (bar/tray Repeaters): #13011, #12822, #12938.

Filed by Claude Opus 5.5 via Claude Code.