diff --git a/docs/desktop/index.mdx b/docs/desktop/index.mdx
index fc00878..fbcf9e4 100644
--- a/docs/desktop/index.mdx
+++ b/docs/desktop/index.mdx
@@ -51,4 +51,6 @@ The RTX 5060 Ti in a USB4 enclosure (Minisforum `TBGAA`, bolt policy `iommu`) is
 
 Session side: `conf.d/10-env.fish` sets `__EGL_VENDOR_LIBRARY_FILENAMES` (Mesa) and `VK_DRIVER_FILES` (RADV), see [fish](/shell/fish#10-envfish). niri and Umbriel both start through a fish login shell, so the compositor, Noctalia, and every app inherit them. Without them, glvnd and the Vulkan loader load the NVIDIA vendor libraries into every process, and those processes hold `/dev/nvidiactl`. Hyprland gets the same pins from `~/.config/uwsm/env-hyprland`, see [Hyprland](/desktop/hyprland#monitors).
 
+Never expose `nvidia_icd.json` to LM Studio. When its engine survey reaches the eGPU through NVIDIA's Vulkan driver, the card falls off the bus (`Xid 79`) and the machine hard-resets (`reset reason … data fabric sync flood`). The survey runs when the app starts and when the `llmster` daemon starts, and any `lms` command starts that daemon. It happened twice on 2026-10-17. CUDA on the same card is stable, so select the `llama.cpp … nvidia-cuda` engine in LM Studio.
+
 Check: `nvidia-smi` lists the card, and `fuser -v /dev/nvidia*` shows only CUDA processes.
diff --git a/dot_config/fish b/dot_config/fish
index 3b91330..102e1c1 160000
--- a/dot_config/fish
+++ b/dot_config/fish
@@ -1 +1 @@
-Subproject commit 3b913300af8315f9e1f509156255842cca13ca3f
+Subproject commit 102e1c1e7ebba7ebe39e2cf9b7c920ce5eedc8da
diff --git a/dot_local/private_share/applications/ai.elementlabs.lmstudio.desktop b/dot_local/private_share/applications/ai.elementlabs.lmstudio.desktop
deleted file mode 100644
index 8e32934..0000000
--- a/dot_local/private_share/applications/ai.elementlabs.lmstudio.desktop
+++ /dev/null
@@ -1,18 +0,0 @@
-[Desktop Entry]
-# Replaces lmstudio-bin's lmstudio.desktop (hidden by lmstudio.desktop here).
-# The ID matches the window app_id and the scratchpad spawn
-# (`gtk-launch ai.elementlabs.lmstudio` in umbriel/rules.toml and
-# omarchy-toggle-scratch). The session pins Vulkan to RADV; LM Studio's Vulkan
-# backend also gets the NVIDIA ICD so it can run models on the Thunderbolt eGPU.
-# The iGPU stays Vulkan device 0. Keep the env in sync with fish conf.d/10-env.fish.
-Type=Application
-Name=LM Studio
-GenericName=Large Language Model Studio
-Comment=A desktop app for exploring and running large language models locally
-Exec=env VK_DRIVER_FILES=/usr/share/vulkan/icd.d/radeon_icd.json:/usr/share/vulkan/icd.d/nvidia_icd.json MESA_VK_DEVICE_SELECT=1002:1586 lm-studio %U
-Icon=lmstudio-bin
-Categories=Development;ArtificialIntelligence;
-Terminal=false
-StartupNotify=true
-StartupWMClass=ai.elementlabs.lmstudio
-MimeType=text/plain;
diff --git a/dot_local/private_share/applications/lmstudio.desktop b/dot_local/private_share/applications/lmstudio.desktop
deleted file mode 100644
index 5d49201..0000000
--- a/dot_local/private_share/applications/lmstudio.desktop
+++ /dev/null
@@ -1,5 +0,0 @@
-[Desktop Entry]
-# Hides lmstudio-bin's entry; ai.elementlabs.lmstudio.desktop replaces it.
-Type=Application
-Name=LM Studio
-Hidden=true