### Summary

The `kms`-dropping heuristic in `mkinitcpio.conf.d/omarchy_hooks.conf` keys off *GPU presence* rather than *whether that GPU drives a display*. On a desktop with an NVIDIA card and an idle AMD iGPU, it keeps `kms` and pulls ~130 MB of firmware for drivers that never load — the exact cost the file's own comment warns about. On a 2 GB ESP that is the difference between a working system and an unbootable one.

### Current behaviour

```bash
if [[ " ${MODULES[*]:-} " == *" nvidia_drm "* ]]; then
  _omarchy_nvidia_gpu=0
  _omarchy_other_gpu=0
  for _omarchy_pci in "${OMARCHY_PCI_DEVICES_PATH:-/sys/bus/pci/devices}"/*; do
    ...
    [[ $(<"$_omarchy_pci/class") == "0x03"* ]] || continue
    if [[ $(<"$_omarchy_pci/vendor") == "0x10de" ]]; then
      _omarchy_nvidia_gpu=1
    else
      _omarchy_other_gpu=1
    fi
  done
  if ((_omarchy_nvidia_gpu && !_omarchy_other_gpu)); then
    # drop kms
```

`kms` is dropped only when NVIDIA is the **sole** GPU. The accompanying comment explains the intent:

> Keeping the kms hook on such a system makes autodetect pull in nouveau — and ~100 MB of its GSP firmware — for a driver that never runs. [...] Hybrid systems keep kms: their iGPU still needs it for early KMS at the LUKS prompt.

That reasoning is sound for laptops, where the panel hangs off the iGPU. It misfires on a desktop where the iGPU exists but nothing is plugged into it.

### This machine

Ryzen 9 7950X exposes its Raphael iGPU, so `_omarchy_other_gpu=1` and `kms` stays. But the iGPU drives nothing:

```
/sys/class/drm/card0 -> amdgpu    card0-Writeback-1   (no real output)
/sys/class/drm/card1 -> nvidia    card1-DP-1 connected
                                  card1-DP-2 connected
```

Both monitors are on the NVIDIA card, which does its own early KMS via `nvidia.conf` (`nvidia_drm` with `modeset=1`).

### Measured cost

Built both variants against the same root, kernel `7.2.3-arch1-3`, in a tmpfs overlay:

| | initramfs | firmware |
|---|---|---|
| `kms` kept (current) | **273 MB** | 1380 files, incl. 667 amdgpu + nouveau `ga102`/`tu102` GSP |
| `kms` dropped | **134 MB** | 6 nvidia, 14 intel; no amdgpu, no nouveau |

Breakdown of the 245 MB of firmware in the current image:

```
102 MB  nvidia/ga102 + nvidia/tu102 GSP   <- nouveau; nouveau never loads
 80 MB  nvidia/610.57.04/gsp_ga10x.bin    <- needed (AD103)
 28 MB  nvidia/610.57.04/gsp_tu10x.bin    <- Turing, wrong generation
 28 MB  firmware/amdgpu/* (667 files)     <- iGPU drives nothing
```

Dropping `kms` removed 139 MB. On-disk UKI went 280 MB -> 141 MB. Boot got faster too, since Limine reads less off the ESP:

```
loader  27.387s -> 3.480s
kernel  18.101s -> 1.916s
total   1min 46s -> 1min 17s
```

Early KMS and Plymouth were unaffected — `gsp_ga10x.bin` is retained and `nvidia_drm` still comes up in the initramfs:

```
[drm] [nvidia-drm] [GPU ID 0x00000100] Loading driver
[drm] Initialized nvidia-drm 0.0.0 for 0000:01:00.0 on minor 1
fbcon: nvidia-drmdrmfb (fb0) is primary device
```

### Why this matters beyond disk usage

At 280 MB per UKI, with `limine-snapper-sync` retaining one copy per snapshot, ESP demand is `(kernels + snapshots) x 280 MB`. Against a 2 GB ESP and `NUMBER_LIMIT=5` that is ~1.7 GB, and a kernel upgrade eventually fails to write its image. The system stops booting, the user ends up stuck on a recovery subvolume, and snapshot sync silently stops. At 141 MB the same configuration needs ~700 MB and the failure mode disappears.

### Suggested fix

Refine the scan: treat a non-NVIDIA GPU as "needs kms" only when it has a **connected** connector. Something along the lines of walking `/sys/class/drm/card*/card*-*/status` and mapping each card back to its PCI device, counting `connected` only. Laptops keep `kms` as today; desktops with an idle iGPU drop it and save ~130 MB. Preserve the existing inconclusive-means-keep-kms bias for unreadable sysfs.

Separately worth considering: `gsp_tu10x.bin` (28 MB, Turing) ships in the initramfs on an Ada card. If the firmware set can be narrowed to the installed GPU generation, that is another easy saving.

### Workaround in use

A drop-in sorting after Omarchy's, so package updates re-add `kms` upstream and it gets removed again:

```bash
# /etc/mkinitcpio.conf.d/zzz-drop-kms.conf
_nokms_hooks=()
for _nokms_hook in "${HOOKS[@]}"; do
  [[ $_nokms_hook == "kms" ]] || _nokms_hooks+=("$_nokms_hook")
done
HOOKS=("${_nokms_hooks[@]}")
unset _nokms_hooks _nokms_hook
```

### Environment

```
Omarchy   4.0.3 -> 4.0.4
kernel    7.2.3-arch1-3, linux-omarchy 7.2.5-3
GPU       RTX 4080 SUPER (AD103), driver 610.57.04, both monitors on DP
iGPU      AMD Raphael (0d:00.0), amdgpu bound, no connected outputs
boot      Limine 12.8.0, UKI, ESP 2.0 GB
root      btrfs @ on LUKS2, TPM2 auto-unlock (sd-encrypt)
```