## Summary
Closing a libvirt/QEMU VM (win11) causes the **main Chromium browser process** to die with `SIGTRAP` (internal `CHECK` failure) at the exact moment the VM's systemd scope deactivates. All Chromium windows close instantly.

## Environment
- **Distro**: Arch Linux (Omarchy)
- **Window Manager**: Hyprland (via uwsm)
- **Chromium**: `chromium 151.0.7922.173-1` (Arch package)
- **Mesa**: `mesa 1:26.2.1-1`
- **VM**: libvirt/QEMU, Windows 11 guest
- **VM Graphics**: `virtio-vga-gl` + `egl-headless` with `<gl rendernode='/dev/dri/renderD128'/>`
- **GPU**: AMD Radeon RX 6700 XT (renderD128)

## Steps to Reproduce
1. Start Chromium (multiple windows/tabs)
2. Start a libvirt VM with `virtio-vga-gl` + `egl-headless` + `rendernode=/dev/dri/renderD128`
3. Close the VM (shutdown or poweroff)
4. Observe: **All Chromium windows vanish instantly** at the exact second `machine-qemu-*-win11.scope: Deactivated`

## Evidence
- **Signal**: `SIGTRAP` (5), `si_code=SI_KERNEL` — internal `CHECK` failure
- **Process**: Main browser (`/usr/lib/chromium/chromium` — no `--type=`)
- **Stack**: `ChromeMain` → glib main loop → `gpu_data_manager_impl_private.cc:417`
- **Fatal message**: `"GPU process isn't usable. Goodbye."`
- **Timing**: Identical to `machine-qemu-*-win11.scope: Deactivated` (same second)

## What Has Been Ruled Out
| Hypothesis | Test | Result |
|------------|------|--------|
| CAC hook `pkill` | Fixed hook to child-only + dedupe | Still crashes **before** hook's `release` op runs |
| Mesa/GPU sharing | VM moved to plain `virtio-vga` (no egl-headless) | Still crashes |
| CAC/OpenSC USB | Disabled `opensc-pkcs11.so` in NSS | Still crashes (VM even picked up CAC fine) |
| GPU process | `--disable-gpu`, then SwiftShader (`--use-gl=angle --use-angle=swiftshader`) | Still crashes |
| `vm.max_map_count` | Already at 1,048,576 | N/A |
| Clock/time | No `adjtimex`/NTP step at crash moment | N/A |

## Key Observations
1. Crash is **main browser process** (no `--type=`), not a renderer/GPU child
2. `SIGTRAP` with `SI_KERNEL` = kernel-delivered fatal `CHECK` assertion
3. Occurs at **VM scope deactivation** — before libvirt hooks run their `release` phase
4. Stack trace consistently shows `ChromeMain` → glib main loop → `gpu_data_manager_impl_private.cc`
5. `07:51` SIGILL renderer crash occurred **with no VM running** — Chromium has baseline instability on this stack

## VM Configuration (Relevant Portion)
```xml
<graphics type='egl-headless'>
  <gl rendernode='/dev/dri/renderD128'/>
</graphics>
<video>
  <model type='virtio' heads='1' primary='yes' device='virtio-vga-gl'>
    <acceleration accel3d='yes'/>
  </model>
</video>
```

## Workaround Attempted
- Moving VM to `virtio-vga` (SPICE only, no egl-headless) did **not** fix it
- `--disable-gpu` and SwiftShader (`--use-gl=angle --use-angle=swiftshader --enable-unsafe-swiftshader`) did **not** fix it

## Request
This appears to be a fundamental interaction between VM teardown (cgroup/PCI/kernel event at scope deactivation) and Chromium's `GpuDataManagerImpl` fatal fallback path. The `CHECK` at `gpu_data_manager_impl_private.cc:417` fires when the GPU process is deemed "not usable" — but this happens even with software GL, suggesting the trigger is a kernel/cgroup event during VM scope teardown that Chromium's GPU data manager interprets as GPU loss.

**Possible areas to investigate:**
1. Chromium's `GpuDataManagerImpl` handling of GPU process termination during system cgroup changes
2. Hyprland/uwsm session handling of XWayland restarts during VM teardown
3. libvirt's cgroup/PCI teardown sequence affecting GPU process visibility