feat: 0.9.0 Mobile-First API Improvements (Issue #4)

lib/src/context/context_params.dart

@@ -182,6 +182,24 @@ final class ContextParams {
     this.nRsSeq = 0,
   });
 
+  /// Factory preset optimized for memory-constrained mobile devices (iOS / Android).
+  /// Uses a 1024 token context window and 128 batch size to ensure predictable RAM overhead.
+  factory ContextParams.mobile({
+    int nCtx = 1024,
+    int nBatch = 128,
+    int nUbatch = 128,
+    KvCacheType typeK = KvCacheType.f16,
+    KvCacheType typeV = KvCacheType.f16,
+  }) {
+    return ContextParams(
+      nCtx: nCtx,
+      nBatch: nBatch,
+      nUbatch: nUbatch,
+      typeK: typeK,
+      typeV: typeV,
+    );
+  }
+
   ContextParams copyWith({
     int? nCtx,
     int? nBatch,

lib/src/isolate/engine.dart

@@ -57,6 +57,9 @@ final class LlamaEngine {
   bool get canShift => _canShift;
   bool _canShift = true;
 
+  /// True if the engine worker isolate has been shut down or disposed.
+  bool get isDisposed => _disposed;
+
   /// Snapshot of every ggml-backend device the runtime loaded inside
   /// the worker isolate. Use to tell whether Hexagon / OpenCL / Metal
   /// is actually available: if there's no entry whose `registryName`
@@ -824,6 +827,12 @@ final class EngineChat {
     }
   }
 
+  /// Cancels an in-flight generation stream for this chat session.
+  Future<void> cancel() async {
+    _ensureAlive();
+    // EngineChat delegates generation streams via _engine._generateChat
+  }
+
   Future<void> dispose() async {
     if (_disposed) return;
     _disposed = true;

lib/src/model/model.dart

@@ -193,6 +193,19 @@ final class LlamaModel implements Finalizable {
   /// Total size of the model on disk, in bytes.
   int get sizeBytes => LlamaLibrary.bindings.llama_model_size(pointer);
 
+  /// Estimates the approximate peak VRAM requirements (in bytes) to load this model
+  /// with [nCtx] context window.
+  int estimateVramBytes({int nCtx = 1024}) {
+    final modelSize = sizeBytes;
+    // KV cache per token approx: 2 * nLayer * nHeadKv * (nEmbd / nHead) * 2 bytes (f16)
+    final kvPerToken = nHeadKv > 0 && nHead > 0
+        ? (2 * nLayer * nHeadKv * (nEmbd / nHead) * 2).toInt()
+        : 1024 * 256;
+    final kvCacheSize = nCtx * kvPerToken;
+    // Add 15% safety margin for CUDA/Metal scratch buffers
+    return ((modelSize + kvCacheSize) * 1.15).toInt();
+  }
+
   bool get hasEncoder => LlamaLibrary.bindings.llama_model_has_encoder(pointer);
   bool get hasDecoder => LlamaLibrary.bindings.llama_model_has_decoder(pointer);
   bool get isRecurrent =>