feat: self-host on-device model, fix FlutterGemma init, surface real errors

NOTICE

@@ -0,0 +1,17 @@
+Cachy — on-device model attribution and use restrictions
+=========================================================
+
+The optional on-device AI feature downloads and runs Gemma 3 1B (int4 .task
+build), redistributed by this project from a GitHub release for install
+convenience. Gemma is provided by Google under the Gemma Terms of Use.
+
+Your use of the Gemma model is subject to:
+  - Gemma Terms of Use:            https://ai.google.dev/gemma/terms
+  - Gemma Prohibited Use Policy:   https://ai.google.dev/gemma/prohibited_use_policy
+
+Gemma is a trademark of Google LLC. This project is not affiliated with or
+endorsed by Google. The model weights are unmodified; only the hosting location
+differs from the upstream litert-community/Gemma3-1B-IT repository.
+
+By downloading the model through Cachy you agree to the Gemma Terms of Use and
+Prohibited Use Policy above.

app/lib/data/services/local_ai/gemma_local_ai_service.dart

@@ -12,10 +12,11 @@ import '../local_store.dart';
 import 'local_ai_service.dart';
 
 /// Gemma 3 1B IT, int4 .task build — the flutter_gemma-supported model with the
-/// best JSON-following per size. HF-gated: download needs a HuggingFace token
-/// with the Gemma license accepted.
+/// best JSON-following per size. Self-hosted on our GitHub release so end users
+/// need no HuggingFace account or license token (Gemma license permits
+/// redistribution with its use-restrictions notice, shipped in NOTICE).
 const kLocalAiModelUrl =
-    'https://huggingface.co/litert-community/Gemma3-1B-IT/resolve/main/gemma3-1b-it-int4.task';
+    'https://github.com/Vatsal057/Cachy/releases/download/model-v1/gemma3-1b-it-int4.task';
 const kLocalAiModelFile = 'gemma3-1b-it-int4.task';
 const kLocalAiModelSizeLabel = '~550 MB';
 
@@ -41,8 +42,10 @@ class GemmaLocalAiService extends LocalAiService {
 
   static bool get _supported => !kIsWeb && Platform.isAndroid;
 
-  @override
-  Future<void> saveHfToken(String token) => _store.setHfToken(token);
+  /// flutter_gemma requires FlutterGemma.initialize() before any other call.
+  /// Lazy + memoized so app startup never pays for it when the feature is unused.
+  Future<void>? _gemmaInit;
+  Future<void> _ensureGemmaInit() => _gemmaInit ??= FlutterGemma.initialize();
 
   void _set(LocalAiStatus s) {
     _status = s;
@@ -53,6 +56,7 @@ class GemmaLocalAiService extends LocalAiService {
     if (!_supported) return;
     _set(const LocalAiStatus(LocalAiPhase.notInstalled));
     try {
+      await _ensureGemmaInit();
       final installed = await FlutterGemma.isModelInstalled(kLocalAiModelFile);
       if (installed) _set(const LocalAiStatus(LocalAiPhase.ready));
     } catch (e) {
@@ -65,16 +69,20 @@ class GemmaLocalAiService extends LocalAiService {
     if (!_supported || _status.phase == LocalAiPhase.downloading) return;
     _set(const LocalAiStatus(LocalAiPhase.downloading));
     try {
+      await _ensureGemmaInit();
       await FlutterGemma.installModel(modelType: ModelType.gemmaIt)
-          .fromNetwork(kLocalAiModelUrl, token: _store.hfToken)
+          .fromNetwork(kLocalAiModelUrl)
           .withProgress((int percent) {
         _set(LocalAiStatus(LocalAiPhase.downloading, progress: percent / 100));
       }).install();
       _set(const LocalAiStatus(LocalAiPhase.ready));
     } catch (e) {
       debugPrint('local-ai: download failed: $e');
+      // Surface the real cause — a generic message made this undebuggable.
+      final detail = e.toString();
       _set(LocalAiStatus(LocalAiPhase.error,
-          message: 'Download failed — check connection and HF token'));
+          message:
+              'Download failed: ${detail.length > 220 ? detail.substring(0, 220) : detail}'));
     }
   }
 
@@ -82,6 +90,7 @@ class GemmaLocalAiService extends LocalAiService {
   Future<void> delete() async {
     if (!_supported) return;
     try {
+      await _ensureGemmaInit();
       await FlutterGemma.uninstallModel(kLocalAiModelFile);
     } catch (e) {
       debugPrint('local-ai: delete failed: $e');
@@ -98,6 +107,7 @@ class GemmaLocalAiService extends LocalAiService {
     if (!canStructure) return null;
     InferenceModel? model;
     try {
+      await _ensureGemmaInit();
       model = await FlutterGemma.getActiveModel(maxTokens: 2048);
       final session = await model.createSession(temperature: 0.3, topK: 40);
       try {

app/lib/data/services/local_ai/local_ai_service.dart

@@ -44,9 +44,6 @@ abstract class LocalAiService extends ChangeNotifier {
   bool get canStructure =>
       enabled && status.phase == LocalAiPhase.ready;
 
-  /// Persist the HuggingFace token used for the license-gated model download.
-  Future<void> saveHfToken(String token);
-
   /// Download + install the model (~550 MB — caller shows the Wi-Fi warning).
   Future<void> download();
 

app/lib/data/services/local_store.dart

@@ -21,7 +21,6 @@ class LocalStore {
   static const _apiBaseUrlKey = 'api_base_url';
   static const _splitPaneFractionKey = 'split_pane_fraction';
   static const _localAiEnabledKey = 'local_ai_enabled';
-  static const _hfTokenKey = 'hf_token';
 
   static Future<LocalStore> open() async =>
       LocalStore(await SharedPreferences.getInstance());
@@ -48,14 +47,6 @@ class LocalStore {
   bool get localAiEnabled => _prefs.getBool(_localAiEnabledKey) ?? false;
   Future<void> setLocalAiEnabled(bool v) => _prefs.setBool(_localAiEnabledKey, v);
 
-  /// HuggingFace token for the license-gated local model download.
-  String? get hfToken {
-    final t = _prefs.getString(_hfTokenKey);
-    return (t == null || t.isEmpty) ? null : t;
-  }
-
-  Future<void> setHfToken(String token) => _prefs.setString(_hfTokenKey, token);
-
   /// Clears user identity and onboarding state, effectively signing the user
   /// out and forcing them back through the name-entry gate on next launch.
   Future<void> clearUser() async {

app/lib/ui/features/profile/views/profile_screen.dart

@@ -478,30 +478,14 @@ class _OfflineAiSection extends StatelessWidget {
   }
 
   Future<void> _confirmDownload(BuildContext context, LocalAiService ai) async {
-    final controller = TextEditingController();
     final ok = await showDialog<bool>(
       context: context,
       builder: (ctx) => AlertDialog(
         title: const Text('Download model?'),
-        content: Column(
-          mainAxisSize: MainAxisSize.min,
-          crossAxisAlignment: CrossAxisAlignment.start,
-          children: [
-            const Text(
-                '$kLocalAiModelSizeLabel download — Wi-Fi recommended. The model '
-                'is license-gated on HuggingFace: paste a free HF token that has '
-                'accepted the Gemma license.'),
-            const SizedBox(height: 12),
-            TextField(
-              controller: controller,
-              decoration: const InputDecoration(
-                labelText: 'HuggingFace token (hf_…)',
-                border: OutlineInputBorder(),
-              ),
-              obscureText: true,
-            ),
-          ],
-        ),
+        content: const Text(
+            '$kLocalAiModelSizeLabel download — Wi-Fi recommended. Runs entirely '
+            'on your phone once installed; no account or sign-in needed. Keep the '
+            'app open until it finishes.'),
         actions: [
           TextButton(
               onPressed: () => Navigator.pop(ctx, false),
@@ -513,8 +497,6 @@ class _OfflineAiSection extends StatelessWidget {
       ),
     );
     if (ok != true) return;
-    final token = controller.text.trim();
-    if (token.isNotEmpty) await ai.saveHfToken(token);
     await ai.download();
   }
 

app/test/local_ai_test.dart

@@ -46,9 +46,6 @@ class FakeLocalAiService extends LocalAiService {
     notifyListeners();
   }
 
-  @override
-  Future<void> saveHfToken(String token) async {}
-
   @override
   Future<void> download() async {
     _phase = LocalAiPhase.downloading;