feat: embed a Flutter G4F client with Cronet, cookie jar, and origin WebView

Port the Oracle off the Python g4f container onto on-device providers: Chromium HTTP via Cronet, a shared cookie session, IterList retry, hosted OpenAI-compat aggregators, and same-origin WebView fallbacks instead of CORS fetch.

lib/g4f/providers/base_provider.dart

@@ -0,0 +1,103 @@
+import '../models.dart';
+
+/// Base abstract class for all G4F providers.
+/// Mirrors `AsyncGeneratorProvider` in xtekky/gpt4free.
+abstract class BaseProvider {
+  /// User-friendly name of the provider
+  String get name;
+
+  String get label => name;
+
+  String get url => '';
+
+  /// Whether the provider is currently believed to be working
+  bool get working;
+
+  /// List of supported models by this provider
+  List<String> get models;
+
+  /// Optional fallback default model if none specified
+  String get defaultModel;
+
+  /// Whether this provider needs a hidden WebView (slow, device-only).
+  bool get isWebView => false;
+
+  /// Per-provider deadline. Python g4f defaults to ~120s; we keep a minute.
+  Duration get requestTimeout => const Duration(seconds: 60);
+
+  /// g4f `model_aliases` (e.g. gpt-4o → provider default).
+  Map<String, String> get modelAliases => const {};
+
+  String resolveModel(String? model) {
+    if (model == null || model.isEmpty) return defaultModel;
+    if (models.contains(model)) return model;
+    final alias = modelAliases[model];
+    if (alias != null) return alias;
+    return defaultModel;
+  }
+
+  /// Whether the provider can emit incremental tokens.
+  bool get supportsStream => false;
+
+  /// Full reply. The Oracle never streams — it waits for a complete text.
+  Future<String> complete(String model, List<Message> messages) async {
+    final buf = StringBuffer();
+    await for (final chunk in createAsyncGenerator(
+      model,
+      messages,
+      stream: false,
+    ).timeout(requestTimeout)) {
+      buf.write(chunk);
+    }
+    return buf.toString();
+  }
+
+  /// Convert chat history to OpenAI-style JSON maps.
+  List<Map<String, String>> toOpenAiMessages(List<Message> messages) {
+    return [
+      for (final m in messages) {'role': m.role.name, 'content': m.content},
+    ];
+  }
+
+  /// Helper to get last user message
+  String getLastUserMessage(List<Message> messages) {
+    for (var i = messages.length - 1; i >= 0; i--) {
+      if (messages[i].role == Role.user) {
+        return messages[i].content;
+      }
+    }
+    return '';
+  }
+
+  /// Helper to extract system prompt if supported by the provider
+  String extractSystemMessage(List<Message> messages) {
+    final sys = messages.where((m) => m.role == Role.system);
+    if (sys.isNotEmpty) {
+      return sys.map((m) => m.content).join('\n');
+    }
+    return '';
+  }
+
+  /// Helper to format a full prompt string (for providers without native chat context)
+  String formatPrompt(List<Message> messages, {bool includeSystem = true}) {
+    final sb = StringBuffer();
+    for (final m in messages) {
+      if (m.role == Role.system) {
+        if (includeSystem) sb.writeln('System: ${m.content}');
+      } else if (m.role == Role.user) {
+        sb.writeln('User: ${m.content}');
+      } else {
+        sb.writeln('Assistant: ${m.content}');
+      }
+    }
+    return sb.toString();
+  }
+
+  /// Main interface to execute a chat completion synchronously/asynchronously.
+  /// Streams chunked output if `stream` is supported.
+  Stream<String> createAsyncGenerator(
+    String model,
+    List<Message> messages, {
+    bool stream = false,
+  });
+}