llama-cpp: fix model hash

Replace three custom Prometheus textfile collector scripts with
dedicated community-maintained exporters:

- jellyfin-collector.nix (25 LoC shell) -> rebelcore/jellyfin_exporter
  Metric: jellyfin_active_streams -> count(jellyfin_now_playing_state)
  Bonus: per-session labels (user, title, device, codec info)

- qbittorrent-collector.nix (40 LoC shell) -> anriha/qbittorrent-metrics-exporter
  Metric: qbittorrent_{download,upload}_bytes_per_second -> qbit_{dl,up}speed
  Bonus: per-torrent metrics with category/tag aggregation

- intel-gpu-collector.nix + .py (130 LoC Python) -> mike1808/igpu-exporter
  Metric: intel_gpu_engine_busy_percent -> igpu_engines_busy_percent
  Bonus: persistent daemon vs oneshot timer, no streaming JSON parser

All three run as persistent daemons scraped by Prometheus, replacing
the textfile-collector pattern of systemd timers writing .prom files.
Dashboard PromQL queries updated to match new metric names.

services/grafana/exporters.nix

@@ -15,7 +15,9 @@ in
   # -- Jellyfin Prometheus Exporter --
   # Replaces custom jellyfin-collector.nix textfile timer.
   # Exposes per-session metrics (jellyfin_now_playing_state) and library stats.
-  systemd.services.jellyfin-exporter = lib.mkIf (config.services.grafana.enable && config.services.jellyfin.enable) {
+  systemd.services.jellyfin-exporter =
+    lib.mkIf (config.services.grafana.enable && config.services.jellyfin.enable)
+      {
         description = "Prometheus exporter for Jellyfin";
         after = [
           "network.target"
@@ -23,7 +25,8 @@ in
         ];
         wantedBy = [ "multi-user.target" ];
         serviceConfig = {
-      ExecStart = lib.getExe (pkgs.writeShellApplication {
+          ExecStart = lib.getExe (
+            pkgs.writeShellApplication {
               name = "jellyfin-exporter-wrapper";
               runtimeInputs = [ pkgs.jellyfin-exporter ];
               text = ''
@@ -32,7 +35,8 @@ in
                   --jellyfin.token="$(cat "$CREDENTIALS_DIRECTORY/jellyfin-api-key")" \
                   --web.listen-address=127.0.0.1:${toString jellyfinExporterPort}
               '';
-      });
+            }
+          );
           Restart = "on-failure";
           RestartSec = "10s";
           DynamicUser = true;
@@ -49,7 +53,9 @@ in
   # Replaces custom qbittorrent-collector.nix textfile timer.
   # Exposes per-torrent metrics (qbit_dlspeed, qbit_upspeed) and aggregate stats.
   # qBittorrent runs in a VPN namespace; the exporter reaches it via namespace address.
-  systemd.services.qbittorrent-exporter = lib.mkIf (config.services.grafana.enable && config.services.qbittorrent.enable) {
+  systemd.services.qbittorrent-exporter =
+    lib.mkIf (config.services.grafana.enable && config.services.qbittorrent.enable)
+      {
         description = "Prometheus exporter for qBittorrent";
         after = [
           "network.target"
@@ -57,7 +63,9 @@ in
         ];
         wantedBy = [ "multi-user.target" ];
         serviceConfig = {
-      ExecStart = lib.getExe' inputs.qbittorrent-metrics-exporter.packages.${pkgs.system}.default "qbittorrent-metrics-exporter";
+          ExecStart =
+            lib.getExe' inputs.qbittorrent-metrics-exporter.packages.${pkgs.system}.default
+              "qbittorrent-metrics-exporter";
           Restart = "on-failure";
           RestartSec = "10s";
           DynamicUser = true;

services/llama-cpp/llama-cpp.nix

@@ -16,7 +16,7 @@ in
     model = toString (
       pkgs.fetchurl {
         url = "https://huggingface.co/unsloth/gemma-4-E4B-it-GGUF/resolve/main/gemma-4-E4B-it-Q4_K_M.gguf";
-        sha256 = "ced37f54b80068fe65e95c6dd79ac88cddc227e179fd1040b8f751b1e5bdf849";
+        sha256 = "sha256-4bxEJwn+eAqksuybIsFqf83/VC8X8B7Q4yAxFNKPnzQ=";
       }
     );
     port = service_configs.ports.private.llama_cpp.port;