terra: ollama (ROCm) + LibreChat with local persistent memory

Local LLM server on the 6800 XT (ollama-rocm, gfx1030 needs no
HSA_OVERRIDE_GFX_VERSION) fronted by a LibreChat web UI, talking to it
over the OpenAI-compatible /v1 route. Also wires up LibreChat's
persistent-memory feature, which needed its own agent+model plus a
custom extraction prompt: the default 3b model couldn't reliably tell
the user's stated facts apart from its own boilerplate, and even a
tuned prompt didn't fix that — so memory extraction now reuses
gemma4:12b, the same model as the daily-driver chat endpoint.

flake.lock bump pulls in the ollama and librechat NixOS modules.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

hosts/terra/configuration.nix

@@ -17,6 +17,7 @@ in
     ../../services/vpn/tailscale.nix
     ../../services/desktop/desktop-hyprland.nix
     ../../services/desktop/desktop-apps.nix
+    ../../services/desktop/librechat.nix
   ];
 
   networking.hostName = "terra";
@@ -69,6 +70,34 @@ in
   # whisper.cpp/llama.cpp containers in ~/Data/Dev/repos/content-trigger-scanner.
   users.users.darman.extraGroups = [ "render" "video" ];
 
+  # ---- ollama (local LLM server, ROCm on the 6800 XT) ----
+  # Navi 21 is gfx1030 — officially supported by ROCm, so no
+  # rocmOverrideGfx/HSA_OVERRIDE_GFX_VERSION needed (that's for gpus ROCm
+  # doesn't recognize, e.g. RDNA1/gfx101x). The upstream module runs the
+  # service under DynamicUser with SupplementaryGroups=["render"] and
+  # DeviceAllow for char-kfd/char-drm/char-fb already, so unlike jellyfin's
+  # static user it needs no extraGroups wiring here.
+  services.ollama = {
+    enable = true;
+    package = pkgs.ollama-rocm;
+    # keep in sync with services/desktop/librechat.nix's endpoints.custom
+    # default model — LibreChat's config schema needs a non-empty default
+    # even though fetch=true replaces it with whatever's actually pulled.
+    # gemma4:12b: general chat/coding daily driver, fits fully in 16G VRAM —
+    # also doubles as the memory-extraction agent (see librechat.nix): a
+    # 3b model (llama3.2:3b, dropped) couldn't reliably tell the user's
+    # stated facts apart from its own boilerplate, e.g. saving "I am an AI
+    # assistant with tool calling capabilities" as the user's personal_info
+    # after "Hi I'm Erik Simon". Reusing gemma4:12b for both roles also means
+    # no second model needs to swap into VRAM while it's already the active
+    # chat model.
+    # qwen3.6:35b-a3b: MoE (3B active/36B total), ~24GB Q4_K_M — doesn't fit
+    # in VRAM alone, so ollama offloads the inactive experts to CPU RAM.
+    # Sparse activation makes that far less painful than it'd be for a dense
+    # model this size, but still expect it to run slower than the two above.
+    loadModels = [ "gemma4:12b" "qwen3.6:35b-a3b" ];
+  };
+
   # ---- Dev-data disks — NOT in disko, mounted read-write, never wiped ----
   fileSystems."/mnt/hdd_01" = {
     device = "/dev/disk/by-uuid/b8445126-ec6d-4f88-818a-d9e13031d9a4";