feat(local-inference): replace ollama with llama-swap + llama.cpp on zix790prors

- Add local-inference NixOS role using llama-swap (from nixpkgs-unstable)
  with llama.cpp (CUDA-enabled, from nixpkgs-unstable)
- Serves Qwen3.6-35B-A3B via HuggingFace auto-download with --cpu-moe
- Add nixosSpecialArgs for nixpkgs-unstable module access
- Configure opencode with llama-local provider pointing to zix790prors:8080
- Update gptel from Ollama backend to OpenAI-compatible llama-swap backend
- Remove ollama service from zix790prors

machines/zix790prors/configuration.nix

@@ -26,6 +26,18 @@ with lib;
       x11 = true;
     };
     kodi.enable = true;
+    local-inference = {
+      enable = true;
+      host = "zix790prors.oglehome";
+      openFirewall = true;
+      models = {
+        "Qwen3.6-35B-A3B" = {
+          hf-model = "unsloth/Qwen3.6-35B-A3B-GGUF:UD-Q8_K_XL";
+          aliases = [ "Qwen3.6-35B-A3B" ];
+          cpu-moe = true;
+        };
+      };
+    };
     nfs-mounts.enable = true;
     nvidia = {
       enable = true;
@@ -56,12 +68,6 @@ with lib;
     ${pkgs.xorg.xrandr}/bin/xrandr --output DP-0 --mode 3440x1440 --rate 164.90 --primary
   '';
 
-  services.ollama = {
-    enable = true;
-    acceleration = "cuda";
-    loadModels = [ "gpt-oss" "deepseek-r1" "qwen3:30b" ];
-  };
-
   # This option defines the first version of NixOS you have installed on this particular machine,
   # and is used to maintain compatibility with application data (e.g. databases) created on older NixOS versions.
   #