## What happened

`omarchy-voxtype-install` runs `voxtype setup gpu --enable` whenever `omarchy-hw-vulkan` succeeds. That check only verifies that *any* ICD JSON exists in `/usr/share/vulkan/icd.d`; it says nothing about whether the GPU can actually run ggml's Vulkan backend.

On Intel Haswell (Iris Pro 5200 / Crystal Well, Mesa `hasvk` driver) the Vulkan device lacks 16-bit storage, which ggml_vulkan requires. The daemon aborts on every model load, and `voxtype.service` (`Restart=on-failure`, `RestartSec=5s`) loops forever. On this machine that produced ~127,000 core dumps (one every ~5.5s for days), a journal rotated down to the crash loop itself, and dictation never works.

```
MESA-INTEL: warning: Haswell Vulkan support is incomplete
ggml_vulkan: 0 = Intel(R) Iris(R) Pro Graphics P5200 (HSW GT3) (Intel open-source Mesa driver) | uma: 1 | fp16: 0 | bf16: 0 | ...
ggml_vulkan: device Vulkan0 does not support 16-bit storage.
fatal runtime error: Rust cannot catch foreign exceptions, aborting
```

systemd-coredump: `voxtype-vulkan`, SIGABRT (SI_TKILL), main thread `abort()` from within `voxtype-vulkan`; another thread parked in `libvulkan_intel_hasvk.so`.

## Expected

Dictation should stay on the CPU backend when the Vulkan device can't run ggml, for example:
- skip `setup gpu --enable` for `hasvk` / pre-Gen9 Intel (the ICD is `intel_hasvk_icd.*.json`), or
- after `voxtype setup gpu --enable`, run a quick probe (load the model once) and revert with `voxtype setup gpu --disable` if it fails.

## Steps to reproduce

1. Omarchy on a Haswell/Ivy Bridge Intel machine (here: MacBook Pro with Iris Pro 5200).
2. Install Dictation from the Omarchy menu (`omarchy-voxtype-install`) so the GPU backend gets enabled.
3. `journalctl --user -u voxtype -f`: the daemon aborts at model load and restarts every 5s. `coredumpctl list` grows continuously.

## Workaround

```
sudo voxtype setup gpu --disable
systemctl --user restart voxtype
```

## Notes

- Related: #7099 and #11110 (GPU backend *not* being enabled). Fixing #11110 will make `setup gpu --enable` succeed more often and expose more pre-Gen9 Intel users to this crash loop.
- The hard abort itself is also an upstream voxtype/whisper.cpp issue: an unsupported Vulkan device should fall back to CPU rather than throw a C++ exception across the Rust FFI boundary.
- I can't prove from logs whether the Vulkan backend on this machine was enabled by `omarchy-voxtype-install` or by hand. The pacman scriptlet has been preserving `Backend: vulkan` since the 0.7.5 upgrade in June. The install script logic above would enable it on this hardware either way.

## System

- Omarchy 4.0.3-1 (stable)
- Kernel 7.2.3-arch1-3
- GPU: Intel Crystal Well Integrated Graphics (Iris Pro 5200, HSW GT3)
- voxtype-bin 1.0.1-1, mesa 1:26.2.2-1, vulkan-intel 1:26.2.2-1

(`omarchy debug` isn't available in this Omarchy version, so the details above were collected manually.)

---
Filed by Claude Opus 5 via Claude Code.