## What happened

After running a full system upgrade (`pacman -Syu`) shortly after a fresh Omarchy install, Hyprland began crash-looping continuously (every 30s–90min depending on GPU load), taking down Xwayland, `xdg-desktop-portal-hyprland`, and `hyprland-dialog` with it. The upgrade bumped `mesa` (1:26.1.7-1 → 1:26.2.2-1), `hyprland`/`aquamarine`/`hyprutils`/`hyprtoolkit` in lockstep, and `linux-firmware-amdgpu` (20260810-1 → 20260810-2) — all while Hyprland was already running (via the `omarchy-hyprland-reload-pause/resume` hooks).

Kernel log at every crash showed the same GPU fault, regardless of which layer was reverted:

```
amdgpu 0000:0a:00.0: [gfxhub] page fault (src_id:0 ring:24 vmid:7 pasid:N)
amdgpu 0000:0a:00.0:   in page starting at address 0x0000800... from client 0x1b (UTCL2)
amdgpu 0000:0a:00.0: GCVM_L2_PROTECTION_FAULT_STATUS:0x00701031
amdgpu 0000:0a:00.0:          Faulty UTCL2 client ID: TCP (0x8)
amdgpu 0000:0a:00.0: ring gfx_0.1.0 timeout, signaled seq=N, emitted seq=N+1
amdgpu 0000:0a:00.0: Starting gfx_0.1.0 ring reset
amdgpu 0000:0a:00.0: Ring gfx_0.1.0 reset succeeded
amdgpu 0000:0a:00.0: [drm] device wedged, but no recovery needed
```

Hyprland's own renderer then aborted right after the ring reset (`assertImpl` inside `Render::GL::CHyprOpenGLImpl::begin`, called from `Aquamarine::CDRMBackend::handlePageFlip`), or occasionally segfaulted during backend teardown.

## What was ruled out (each independently tested)

1. **Stale mesa shader cache** — cleared `~/.cache/mesa_shader_cache` and `~/.cache/radv_builtin_shaders`. Crash persisted.
2. **`aquamarine`/`hyprutils` ABI mismatch** — separately found and real (`hyprctl version` showed `hyprland` 0.56.2-2 built against `aquamarine 0.15.0`/`hyprutils 0.14.1` while the repo had already shipped `0.15.1`/`0.14.2`). Pinned back to the matching versions. Fixed the ABI warning but **did not** stop the GPU crash.
3. **`mesa` regression** — downgraded `mesa`, `vulkan-radeon`, `vulkan-mesa-implicit-layers` back to `1:26.1.7-1` (the version installed before the live upgrade). Crash persisted, identical signature.
4. **Kernel regression** — booted the alternate stock `linux` kernel (7.2.3-arch1-3) instead of `linux-omarchy` (7.2.5-3). Crash persisted, identical signature.
5. **`linux-firmware-amdgpu` regression** — downgraded from `20260810-2` to `20260810-1` (Arch Linux Archive). **This fixed it.** No further crashes/page-faults observed after this change, on both kernels and either mesa version.

## Expected behavior

`linux-firmware-amdgpu 20260810-2` should not cause GPU page-faults/ring hangs on this GPU under normal Hyprland compositing.

## System details

- Omarchy version: 4.0.4-1 (channel: stable)
- CPU: AMD Ryzen 7 5800X (desktop, single discrete GPU, not a hybrid/laptop setup — `omarchy hw hybrid gpu` returns false)
- GPU: AMD Navi 22 [1002:73df] (Radeon RX 6700/6700 XT/6750 XT family), ASUS board partner, PCI `0a:00.0`, driver `amdgpu`
- Kernel: reproduced on both `linux-omarchy 7.2.5-3` and stock `linux 7.2.3-arch1-3`
- Mesa: reproduced on both `1:26.2.2-1` and `1:26.1.7-1`
- `linux-firmware-amdgpu`: `20260810-2` reproduces the crash; `20260810-1` does not
- Hyprland: 0.56.2-2 (also reproduced crash-signature-identical after the aquamarine/hyprutils ABI mismatch was independently fixed, ruling that out as the cause)

## Steps to reproduce

1. Fresh Omarchy install, then `sudo pacman -Syu` to pull the latest `linux-firmware-amdgpu` (20260810-2).
2. Use Hyprland normally (switch workspaces, open windows) with this AMD Navi 22 GPU.
3. Within roughly 30 seconds to a few minutes of GPU activity, `dmesg`/`journalctl -k` shows a `[gfxhub] page fault` + `ring gfx_0.1.0 timeout`, and Hyprland crashes/restarts.

## Workaround

Downgrade `linux-firmware-amdgpu` to `20260810-1`:
```
sudo pacman -U https://archive.archlinux.org/packages/l/linux-firmware-amdgpu/linux-firmware-amdgpu-20260810-1-any.pkg.tar.zst
```
and pin it via `IgnorePkg` in `/etc/pacman.conf` until a fixed firmware build is available.