### System details

- **Omarchy:** 4.0.4-1 · **Hyprland:** 0.56.2 (`efb50993…`, `aq_0.15`) · **Aquamarine:** 0.15.0-2 · **uwsm:** 0.26.7-1
- **Kernel:** linux-omarchy 7.2.5-3 · **Driver:** `nvidia-open-dkms` 610.57.04-1
- **GPUs:**
  - `0000:01:00.0` NVIDIA TU117M GeForce GTX 1650 Mobile / Max-Q `[10de:1f9d]` → `nvidia`
  - `0000:06:00.0` AMD Renoir Radeon Vega iGPU `[1002:1636]` → `amdgpu`
- **Outputs:**
  - `DP-1` — Acer XV272U V3, [email protected] — on the **NVIDIA** card
  - `eDP-1` — Chimei Innolux internal panel, 1920x1080@144 — on the **AMD** iGPU

```
/sys/class/drm/card1 -> nvidia [0000:01:00.0]   card1-DP-1:      connected
/sys/class/drm/card2 -> amdgpu [0000:06:00.0]   card2-eDP-1:     connected
                                                card2-HDMI-A-1:  disconnected
```

All three detectors return true on this machine:

```
$ omarchy-hw-nvidia && echo yes        # yes
$ omarchy-hw-nvidia-gsp && echo yes    # yes
$ omarchy-hw-hybrid-gpu && echo yes    # yes
```

### What's wrong?

This is the mirror image of #11173 / #11168 / #8215, and it constrains the fix being proposed there.

`default/hypr/nvidia.lua` exports the NVIDIA VA-API/GL stack whenever `omarchy-hw-nvidia` and `omarchy-hw-nvidia-gsp` succeed:

```
LIBVA_DRIVER_NAME=nvidia
NVD_BACKEND=direct
__GLX_VENDOR_LIBRARY_NAME=nvidia
```

(confirmed present in client environments on this machine).

But **Omarchy never sets `AQ_DRM_DEVICES`** — `grep -rn 'AQ_DRM_DEVICES' /usr/share/omarchy/` returns nothing. So the render device is whatever Aquamarine picks by default, which here is the AMD iGPU, while the only external display is physically wired to the NVIDIA dGPU.

The result is a split pipeline: **compositing on AMD, video decode on NVIDIA (NVDEC), scanout back on NVIDIA.** Frames cross the PCIe bridge in both directions every frame. Chromium flickers on pages containing video; disabling hardware acceleration in the browser avoids it, which matches the cross-GPU dmabuf import failure in #11168.

The important difference from #11173: there, every display hangs off the iGPU, so "stop forcing NVIDIA VA-API" is the right answer. **Here it is the wrong answer** — NVIDIA genuinely owns the connected external output, so the correct resolution is to move the *render device* onto NVIDIA and keep NVIDIA VA-API. A fix that keys only on "is the iGPU rendering?" and drops `LIBVA_DRIVER_NAME` would regress this machine. What both cases actually need is the same invariant: **the VA-API driver and the Aquamarine render device must agree, and the render device should be the GPU that owns the connected outputs.**

### There is no Omarchy-managed place to put `AQ_DRM_DEVICES`

Two constraints make this awkward for users to fix themselves:

1. `hl.env()` in `hypr/*.lua` is too late. Aquamarine reads `AQ_DRM_DEVICES` when it opens the DRM devices at compositor startup, before the Lua config env is applied.
2. `~/.config/hypr/monitors.lua` is overwritten by `omarchy-refresh-hyprland`, so even a comment parked there does not survive.

`~/.config/uwsm/env` works (uwsm's environment preloader runs before Hyprland), but nothing in Omarchy documents or manages it for this purpose.

### Two Aquamarine findings that I could not find reported anywhere

Once NVIDIA is made the primary renderer on this hardware, two further things bite. Neither appears in any existing issue I could find.

**1. The internal panel needs `AQ_MGPU_NO_EXPLICIT=1`.** `eDP-1` hangs off the AMD card, so with NVIDIA primary it can only light up through a cross-GPU (mgpu) buffer copy. Explicit sync on imported buffers is NVIDIA's weak spot, and the atomic commit fails:

```
atomic drm request: failed to commit: Invalid argument
```

`AQ_MGPU_NO_EXPLICIT=1` disables explicit sync on mgpu buffers only and brings the panel up. (`render:explicit_sync` was removed in Hyprland 0.50, so there is no config-level escape hatch — cf. #10350.)

**2. `AQ_NO_MODIFIERS=1` is a trap — it makes the screen go black.** It is the obvious next thing to try, and it applies to *every* GPU rather than just the mgpu path. NVIDIA's KMS then refuses modifier-less scanout buffers:

```
GBM: Allocated a new buffer ... XR24 with modifier 0xffffffffffffff : INVALID
drm: Failed to submit a buffer with drmModeAddFB2
drm: Buffer failed to import to KMS
```

Every mode on `DP-1` fails to commit, the output drops to `0x0`, and the screen stays black while Hyprland itself runs fine. Worth a warning next to any mgpu guidance.

### Steps to reproduce

1. Hybrid AMD iGPU + NVIDIA (GSP) dGPU laptop, with the **external** display wired to the dGPU and the internal panel on the iGPU.
2. Log in to a stock Omarchy session. `nvidia.lua` exports `LIBVA_DRIVER_NAME=nvidia`; Aquamarine picks the AMD iGPU as render device.
3. Play a video in Chromium/Brave on the external monitor with hardware acceleration on → flicker.

### Workaround in use (verified, stable)

In `~/.config/uwsm/env`:

```bash
# Resolve by-path -> /dev/dri/cardN. by-path names must NEVER go into
# AQ_DRM_DEVICES directly: the variable is colon-separated and
# "pci-0000:01:00.0-card" contains colons (see #8776).
_aq_nvidia=$(readlink -f /dev/dri/by-path/pci-0000:01:00.0-card 2>/dev/null)
_aq_amd=$(readlink -f /dev/dri/by-path/pci-0000:06:00.0-card 2>/dev/null)
if [ -c "$_aq_nvidia" ] && [ -c "$_aq_amd" ]; then
  export AQ_DRM_DEVICES="$_aq_nvidia:$_aq_amd"
elif [ -c "$_aq_amd" ]; then
  export AQ_DRM_DEVICES="$_aq_amd"   # NVIDIA missing: don't lose the session
fi
unset _aq_nvidia _aq_amd

export AQ_MGPU_NO_EXPLICIT=1
```

Resulting session: `AQ_DRM_DEVICES=/dev/dri/card1:/dev/dri/card2`, both outputs up at native refresh, no Chromium flicker.

### Suggestions

- Make `nvidia.lua`'s VA-API/GL exports and the Aquamarine render device one decision instead of two independent ones, keyed on **which GPU drives the connected outputs** — this covers both this issue and #11173 without regressing either.
- Give `AQ_DRM_DEVICES` a supported, refresh-proof home (e.g. an Omarchy-managed `~/.config/uwsm/env` fragment), since `hypr/*.lua` structurally cannot host it.
- If NVIDIA-first ordering is ever auto-applied, pair it with `AQ_MGPU_NO_EXPLICIT=1` for iGPU-attached internal panels, resolve by-path names to `/dev/dri/cardN` (#8776), and document that `AQ_NO_MODIFIERS=1` black-screens NVIDIA scanout.

Related: #11173, #11168, #8215 (VA-API vs render device), #10350 (NVIDIA-first ordering), #8776 (colon parsing).