diff --git a/CHANGELOG.md b/CHANGELOG.md
index 597f315de..a5ec06e53 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,5 +1,17 @@
 ## Unreleased
 
+* Updated the default llama.cpp native runtime pin to
+  `leehack/llamadart-native@v0.4.1-1`, keeping the `v0.4.1` llama.cpp
+  ABI/bindings while picking up wrapper-only native fixes. Refreshed
+  the `llamadart_llama_cpp_flutter` Apple SwiftPM checksum and
+  aligned current README/website native override docs.
+
+- Stop a Qwen3-TTS audio decode at its next chunk boundary when native
+  text-to-speech is cancelled, instead of finishing the native step in
+  progress first. This needs llamadart-native v0.4.1-1 or later; older
+  runtimes keep the previous behaviour
+  ([llamadart-native#86](https://github.com/leehack/llamadart-native/issues/86),
+  [#322](https://github.com/leehack/llamadart/issues/322)).
 - Add an experimental `DecisionEngine` for Laya-style decision models (a
   ModernBERT encoder GGUF plus a safetensors head) on native llama.cpp, with
   typed `ChoiceKey`, `ScoreKey` and `NoulKey` questions
diff --git a/README.md b/README.md
index 65bca1f27..a81ee8a7f 100644
--- a/README.md
+++ b/README.md
@@ -156,7 +156,7 @@ Current default runtime pins:
 
 | Runtime | Pin |
 | --- | --- |
-| Native llama.cpp / GGUF | `leehack/llamadart-native@v0.4.1` |
+| Native llama.cpp / GGUF | `leehack/llamadart-native@v0.4.1-1` |
 | Native LiteRT-LM / `.litertlm` | `leehack/litert-lm-native@v0.17.0-6` |
 | Web llama.cpp / GGUF | `leehack/llama-web-bridge-assets@v0.1.44` |
 | Web LiteRT-LM / `.litertlm` | `@litert-lm/core@0.15.0` |
diff --git a/doc/webgpu_bridge.md b/doc/webgpu_bridge.md
index c60365631..bf89a372a 100644
--- a/doc/webgpu_bridge.md
+++ b/doc/webgpu_bridge.md
@@ -22,15 +22,16 @@ pipelines.
 Default pinned tag in the example is `v0.1.44`.
 
 That release embeds llama.cpp `v0.4.1`, matching the `hook/build.dart` native pin
-(`v0.4.1`, both built from upstream llama.cpp `v0.4.1@b29c606e28a01b1bc8c1351026a0fa6e616bf6c4`)
+(`v0.4.1-1`, both built from upstream llama.cpp `v0.4.1@b29c606e28a01b1bc8c1351026a0fa6e616bf6c4`)
 even though the bridge asset tag `v0.1.44` differs from the native runtime tag
-`v0.4.1`. Provenance for this immutable consumer artifact: release `389783936`,
+`v0.4.1-1`. Provenance for this immutable consumer artifact: release `389783936`,
 tag commit `fdafd9f8cbdb9bf99c359536595eff9a23095379`, bridge source
 `89178be67c3c84300bc1b129182bd5bc5a8e21fc`, and manifest SHA-256
-`8d61f453753ac7a7d839ac12318b70986a814748d86029993118c19454293aa9`. It retains
-the Qwen3-ASR typed speech-to-text contract introduced in `v0.1.30` and
-provisions the explicit 1 MiB Wasm stack needed for memory64 context
-construction in direct and worker modes. The chat bootstrap opts
+`8d61f453753ac7a7d839ac12318b70986a814748d86029993118c19454293aa9`. The bridge
+assets were qualified against native `v0.4.1`; `v0.4.1-1` rebuilds it from the
+same upstream commit. It retains the Qwen3-ASR typed speech-to-text contract
+introduced in `v0.1.30` and provisions the explicit 1 MiB Wasm stack needed for
+memory64 context construction in direct and worker modes. The chat bootstrap opts
 `SpeechToTextEngine` into that contract from the immutable tag; older or custom
 assets remain disabled unless the host explicitly sets
 `window.__llamadartBridgeSpeechToTextSupported = true` after equivalent
diff --git a/lib/src/backends/llama_cpp/bindings.dart b/lib/src/backends/llama_cpp/bindings.dart
index 7635b0b03..31d10f32b 100644
--- a/lib/src/backends/llama_cpp/bindings.dart
+++ b/lib/src/backends/llama_cpp/bindings.dart
@@ -6298,6 +6298,15 @@ external int llama_dart_tts_api_version();
 @ffi.Native<ffi.Void Function(ffi.Pointer<llama_dart_tts>)>()
 external void llama_dart_tts_cancel(ffi.Pointer<llama_dart_tts> tts);
 
+@ffi.Native<
+  ffi.Bool Function(ffi.Pointer<ggml_tensor>, ffi.Bool, ffi.Pointer<ffi.Void>)
+>()
+external bool llama_dart_tts_eval_callback(
+  ffi.Pointer<ggml_tensor> tensor,
+  bool ask,
+  ffi.Pointer<ffi.Void> user_data,
+);
+
 @ffi.Native<ffi.Void Function(ffi.Pointer<llama_dart_tts>)>()
 external void llama_dart_tts_free(ffi.Pointer<llama_dart_tts> tts);
 
@@ -6403,6 +6412,23 @@ llama_dart_tts_status llama_dart_tts_reset(ffi.Pointer<llama_dart_tts> tts) {
   return llama_dart_tts_status.fromValue(_llama_dart_tts_reset(tts));
 }
 
+@ffi.Native<
+  ffi.Int Function(ffi.Pointer<llama_dart_tts>, ffi.Pointer<ffi.Int8>)
+>(symbol: 'llama_dart_tts_set_cancel_flag')
+external int _llama_dart_tts_set_cancel_flag(
+  ffi.Pointer<llama_dart_tts> tts,
+  ffi.Pointer<ffi.Int8> flag,
+);
+
+llama_dart_tts_status llama_dart_tts_set_cancel_flag(
+  ffi.Pointer<llama_dart_tts> tts,
+  ffi.Pointer<ffi.Int8> flag,
+) {
+  return llama_dart_tts_status.fromValue(
+    _llama_dart_tts_set_cancel_flag(tts, flag),
+  );
+}
+
 @ffi.Native<
   ffi.Int Function(
     ffi.Pointer<llama_dart_tts>,
@@ -8719,7 +8745,7 @@ external int mtmd_tokenize_from_parts(
   bool add_special,
 );
 
-typedef FILE = __sFILE;
+typedef FILE = _IO_FILE;
 
 const int GGML_BACKEND_META_MAX_DEVICES = 16;
 
@@ -8858,88 +8884,90 @@ final class UnnamedUnion extends ffi.Union {
   external ffi.Array<ffi.Char> val_str;
 }
 
-typedef __darwin_off_t = __int64_t;
-typedef __int64_t = ffi.LongLong;
-typedef Dart__int64_t = int;
+final class _IO_FILE extends ffi.Struct {
+  @ffi.Int()
+  external int _flags;
 
-final class __sFILE extends ffi.Struct {
-  external ffi.Pointer<ffi.UnsignedChar> _p;
+  external ffi.Pointer<ffi.Char> _IO_read_ptr;
 
-  @ffi.Int()
-  external int _r;
+  external ffi.Pointer<ffi.Char> _IO_read_end;
 
-  @ffi.Int()
-  external int _w;
+  external ffi.Pointer<ffi.Char> _IO_read_base;
 
-  @ffi.Short()
-  external int _flags;
+  external ffi.Pointer<ffi.Char> _IO_write_base;
 
-  @ffi.Short()
-  external int _file;
+  external ffi.Pointer<ffi.Char> _IO_write_ptr;
 
-  external __sbuf _bf;
+  external ffi.Pointer<ffi.Char> _IO_write_end;
 
-  @ffi.Int()
-  external int _lbfsize;
+  external ffi.Pointer<ffi.Char> _IO_buf_base;
 
-  external ffi.Pointer<ffi.Void> _cookie;
+  external ffi.Pointer<ffi.Char> _IO_buf_end;
 
-  external ffi.Pointer<
-    ffi.NativeFunction<ffi.Int Function(ffi.Pointer<ffi.Void>)>
-  >
-  _close;
+  external ffi.Pointer<ffi.Char> _IO_save_base;
 
-  external ffi.Pointer<
-    ffi.NativeFunction<
-      ffi.Int Function(ffi.Pointer<ffi.Void>, ffi.Pointer<ffi.Char>, ffi.Int)
-    >
-  >
-  _read;
+  external ffi.Pointer<ffi.Char> _IO_backup_base;
 
-  external ffi.Pointer<
-    ffi.NativeFunction<fpos_t Function(ffi.Pointer<ffi.Void>, fpos_t, ffi.Int)>
-  >
-  _seek;
+  external ffi.Pointer<ffi.Char> _IO_save_end;
 
-  external ffi.Pointer<
-    ffi.NativeFunction<
-      ffi.Int Function(ffi.Pointer<ffi.Void>, ffi.Pointer<ffi.Char>, ffi.Int)
-    >
-  >
-  _write;
+  external ffi.Pointer<_IO_marker> _markers;
 
-  external __sbuf _ub;
+  external ffi.Pointer<_IO_FILE> _chain;
 
-  external ffi.Pointer<__sFILEX> _extra;
+  @ffi.Int()
+  external int _fileno;
 
   @ffi.Int()
-  external int _ur;
+  external int _flags2;
 
-  @ffi.Array.multi([3])
-  external ffi.Array<ffi.UnsignedChar> _ubuf;
+  @__off_t()
+  external int _old_offset;
 
-  @ffi.Array.multi([1])
-  external ffi.Array<ffi.UnsignedChar> _nbuf;
+  @ffi.UnsignedShort()
+  external int _cur_column;
 
-  external __sbuf _lb;
+  @ffi.SignedChar()
+  external int _vtable_offset;
 
-  @ffi.Int()
-  external int _blksize;
+  @ffi.Array.multi([1])
+  external ffi.Array<ffi.Char> _shortbuf;
+
+  external ffi.Pointer<_IO_lock_t> _lock;
 
-  @fpos_t()
+  @__off64_t()
   external int _offset;
-}
 
-final class __sFILEX extends ffi.Opaque {}
+  external ffi.Pointer<_IO_codecvt> _codecvt;
+
+  external ffi.Pointer<_IO_wide_data> _wide_data;
+
+  external ffi.Pointer<_IO_FILE> _freeres_list;
+
+  external ffi.Pointer<ffi.Void> _freeres_buf;
 
-final class __sbuf extends ffi.Struct {
-  external ffi.Pointer<ffi.UnsignedChar> _base;
+  @ffi.Size()
+  external int __pad5;
 
   @ffi.Int()
-  external int _size;
+  external int _mode;
+
+  @ffi.Array.multi([20])
+  external ffi.Array<ffi.Char> _unused2;
 }
 
-typedef fpos_t = __darwin_off_t;
+final class _IO_codecvt extends ffi.Opaque {}
+
+typedef _IO_lock_t = ffi.Void;
+typedef Dart_IO_lock_t = void;
+
+final class _IO_marker extends ffi.Opaque {}
+
+final class _IO_wide_data extends ffi.Opaque {}
+
+typedef __off64_t = ffi.Long;
+typedef Dart__off64_t = int;
+typedef __off_t = ffi.Long;
+typedef Dart__off_t = int;
 typedef ggml_abort_callback =
     ffi.Pointer<ffi.NativeFunction<ggml_abort_callbackFunction>>;
 typedef ggml_abort_callbackFunction =
diff --git a/lib/src/backends/llama_cpp/llama_cpp_backend.dart b/lib/src/backends/llama_cpp/llama_cpp_backend.dart
index 7757e5229..3e259e247 100644
--- a/lib/src/backends/llama_cpp/llama_cpp_backend.dart
+++ b/lib/src/backends/llama_cpp/llama_cpp_backend.dart
@@ -44,23 +44,30 @@ class NativeLlamaBackend
   int _lifecycleEpoch = 0;
   final LlamaWorkerEntrypoint _workerEntrypoint;
   final Duration _workerStartupTimeout;
+  final Allocator _textToSpeechCancelFlagAllocator;
   Pointer<Int8>? _activeCancelToken;
   void Function()? _activeGenerationCleanup;
   void Function()? _activeFreeToken;
   bool _textToSpeechActive = false;
   bool _textToSpeechCancelRequested = false;
   bool _textToSpeechRequestSent = false;
+  _NativeCancelFlag? _textToSpeechCancelFlag;
 
   bool _isReady = false;
   LlamaLogLevel _currentLogLevel = LlamaLogLevel.warn;
 
   /// Creates a new [NativeLlamaBackend] and initializes its ports.
+  ///
+  /// [textToSpeechCancelFlagAllocator] allocates and frees the one-byte cancel
+  /// flag of each text-to-speech synthesis.
   NativeLlamaBackend({
     SendPort? initialSendPort,
     LlamaWorkerEntrypoint workerEntrypoint = llamaWorkerEntry,
     Duration workerStartupTimeout = const Duration(seconds: 30),
+    Allocator textToSpeechCancelFlagAllocator = malloc,
   }) : _workerEntrypoint = workerEntrypoint,
-       _workerStartupTimeout = workerStartupTimeout {
+       _workerStartupTimeout = workerStartupTimeout,
+       _textToSpeechCancelFlagAllocator = textToSpeechCancelFlagAllocator {
     if (initialSendPort != null) {
       _sendPort = initialSendPort;
       _isReady = true;
@@ -758,6 +765,9 @@ class NativeLlamaBackend
   }
 
   Future<void> _disposeWorker() async {
+    // A synthesis claimed after this point sends its flag to the next worker,
+    // so this dispose must not free it.
+    final textToSpeechCancelFlag = _textToSpeechCancelFlag;
     _lifecycleEpoch += 1;
     _isReady = false;
     final startup = _isolateStart;
@@ -792,6 +802,7 @@ class NativeLlamaBackend
     _workerLogPort = null;
     // Worker is gone; free the token if a terminal response did not already.
     _activeFreeToken?.call();
+    textToSpeechCancelFlag?.free();
     _activeCancelToken = null;
     _activeGenerationCleanup = null;
     _activeFreeToken = null;
@@ -886,10 +897,13 @@ class NativeLlamaBackend
     _textToSpeechActive = true;
     _textToSpeechCancelRequested = false;
     _textToSpeechRequestSent = false;
+    final cancelFlag = _NativeCancelFlag(_textToSpeechCancelFlagAllocator);
+    _textToSpeechCancelFlag = cancelFlag;
     try {
       await _ensureIsolate();
     } catch (_) {
       _textToSpeechActive = false;
+      cancelFlag.free();
       rethrow;
     }
     final rp = ReceivePort();
@@ -899,6 +913,7 @@ class NativeLlamaBackend
         contextHandle,
         mmContextHandle,
         request,
+        cancelFlag.address,
         rp.sendPort,
       ),
     );
@@ -942,6 +957,7 @@ class NativeLlamaBackend
       return await completer.future;
     } finally {
       _textToSpeechActive = false;
+      cancelFlag.free();
       await subscription.cancel();
       rp.close();
     }
@@ -953,6 +969,7 @@ class NativeLlamaBackend
       return;
     }
     _textToSpeechCancelRequested = true;
+    _textToSpeechCancelFlag?.raise();
     if (_textToSpeechRequestSent) {
       _sendPort?.send(TextToSpeechCancelRequest());
     }
@@ -1096,3 +1113,34 @@ class NativeLlamaBackend
     throw Exception("Unknown response during chat template application");
   }
 }
+
+/// A one-byte cancel flag that the worker isolate reads, including from native
+/// code while a text-to-speech step runs.
+///
+/// The backend frees it after the worker's terminal response for the request
+/// that carries it, after the worker acknowledges a dispose that began after
+/// the flag was allocated, or when that request was never sent; never on
+/// cancel or on a timer.
+final class _NativeCancelFlag {
+  final Allocator _allocator;
+  Pointer<Int8>? _pointer;
+
+  /// Allocates the flag and lowers it.
+  _NativeCancelFlag(this._allocator) : _pointer = _allocator<Int8>()..value = 0;
+
+  /// The flag's native address.
+  int get address => _pointer!.address;
+
+  /// Raises the flag unless it was freed.
+  void raise() => _pointer?.value = 1;
+
+  /// Frees the flag; later calls do nothing.
+  void free() {
+    final pointer = _pointer;
+    if (pointer == null) {
+      return;
+    }
+    _pointer = null;
+    _allocator.free(pointer);
+  }
+}
diff --git a/lib/src/backends/llama_cpp/llama_cpp_service.dart b/lib/src/backends/llama_cpp/llama_cpp_service.dart
index b9f09d4d5..0bb36e476 100644
--- a/lib/src/backends/llama_cpp/llama_cpp_service.dart
+++ b/lib/src/backends/llama_cpp/llama_cpp_service.dart
@@ -68,6 +68,10 @@ typedef _TtsStepDart =
     int Function(Pointer<llama_dart_tts>, Pointer<llama_dart_tts_progress>);
 typedef _TtsCancelNative = Void Function(Pointer<llama_dart_tts>);
 typedef _TtsCancelDart = void Function(Pointer<llama_dart_tts>);
+typedef _TtsSetCancelFlagNative =
+    Int32 Function(Pointer<llama_dart_tts>, Pointer<Int8>);
+typedef _TtsSetCancelFlagDart =
+    int Function(Pointer<llama_dart_tts>, Pointer<Int8>);
 typedef _TtsResetNative = Int32 Function(Pointer<llama_dart_tts>);
 typedef _TtsResetDart = int Function(Pointer<llama_dart_tts>);
 typedef _TtsGetOutputInfoNative =
@@ -2968,6 +2972,14 @@ class LlamaCppService {
     throw LlamaUnsupportedException(_ttsUnavailableMessage());
   }
 
+  ggml_backend_sched_eval_callback _ttsEvalCallback() {
+    try {
+      return _resolveTtsApi().evalCallback;
+    } on LlamaUnsupportedException {
+      return nullptr;
+    }
+  }
+
   String _ttsUnavailableMessage() {
     return 'Native text-to-speech is unavailable in this runtime bundle '
         '(missing llama_dart_tts_* ABI v$LLAMA_DART_TTS_API_VERSION symbols). '
@@ -6874,9 +6886,11 @@ class LlamaCppService {
     final mmProjPathPtr = mmProjPath.toNativeUtf8();
     Pointer<mtmd_context> mmCtx = nullptr;
     try {
-      final ctxParams = _mtmdContextParamsDefault();
-      ctxParams.use_gpu = _modelToMtmdUseGpu[modelHandle] ?? true;
-      mmCtx = _mtmdInitFromFile(mmProjPathPtr.cast(), model.pointer, ctxParams);
+      mmCtx = _mtmdInitFromFile(
+        mmProjPathPtr.cast(),
+        model.pointer,
+        _mtmdContextParamsFor(modelHandle),
+      );
     } finally {
       malloc.free(mmProjPathPtr);
     }
@@ -6891,6 +6905,36 @@ class LlamaCppService {
     return handle;
   }
 
+  mtmd_context_params _mtmdContextParamsFor(int modelHandle) {
+    final ctxParams = _mtmdContextParamsDefault();
+    ctxParams.use_gpu = _modelToMtmdUseGpu[modelHandle] ?? true;
+    ctxParams.cb_eval = _ttsEvalCallback();
+    return ctxParams;
+  }
+
+  /// The `cb_eval` address [createMultimodalContext] passes to mtmd: that of
+  /// `llama_dart_tts_eval_callback` in the library providing the text-to-
+  /// speech API, or 0 when that API is unavailable or its library does not
+  /// export the callback.
+  int debugMtmdEvalCallbackAddressForTesting() =>
+      _mtmdContextParamsFor(-1).cb_eval.address;
+
+  /// Whether the text-to-speech API this service resolves has
+  /// `llama_dart_tts_eval_callback` and `llama_dart_tts_set_cancel_flag`;
+  /// null when that API is unavailable.
+  ({bool evalCallback, bool setCancelFlag})? debugTtsCancelExportsForTesting() {
+    final _TtsApi api;
+    try {
+      api = _resolveTtsApi();
+    } on LlamaUnsupportedException {
+      return null;
+    }
+    return (
+      evalCallback: api.evalCallback != nullptr,
+      setCancelFlag: api.setCancelFlag != null,
+    );
+  }
+
   /// Frees the multimodal context (projector).
   void freeMultimodalContext(int mmContextHandle) {
     final mmCtx = _mtmdContexts.remove(mmContextHandle);
@@ -8198,12 +8242,21 @@ class LlamaCppService {
   }
 
   /// Synthesizes complete float32 PCM with the stable native TTS wrapper.
+  ///
+  /// [cancelFlagAddress] is the address of a one-byte flag that another
+  /// isolate sets to nonzero to cancel. It is read before the native task is
+  /// created and, when the runtime exports `llama_dart_tts_set_cancel_flag`,
+  /// attached to the task after startup so native code reads it while a step
+  /// runs. It must stay allocated until this future completes; it is last read
+  /// before then.
   Future<BackendTextToSpeechResult> synthesizeTextToSpeech(
     int contextHandle,
     int mmContextHandle,
-    BackendTextToSpeechRequest request, {
+    BackendTextToSpeechRequest request,
+    int cancelFlagAddress, {
     void Function(BackendTextToSpeechProgress progress)? onProgress,
   }) async {
+    final cancelFlag = Pointer<Int8>.fromAddress(cancelFlagAddress);
     final context = _contexts[contextHandle];
     final mtmd = _mtmdContexts[mmContextHandle];
     if (context == null ||
@@ -8276,7 +8329,7 @@ class LlamaCppService {
       // Yields so a cancel already queued for this synthesis can be processed
       // before the native task setup, which cannot be interrupted.
       await Future<void>.delayed(Duration.zero);
-      if (_ttsCancelPending) {
+      if (_ttsCancelPending || cancelFlag.value != 0) {
         _ttsCancelPending = false;
         throw LlamaTextToSpeechException(
           'Text-to-speech synthesis was cancelled.',
@@ -8319,6 +8372,7 @@ class LlamaCppService {
         api: api,
         task: task,
       );
+      api.setCancelFlag?.call(task, cancelFlag);
 
       while (true) {
         progress.ref.struct_size = sizeOf<llama_dart_tts_progress>();
@@ -8544,6 +8598,8 @@ class _TtsApi {
   final _TtsGetOutputInfoDart getOutputInfo;
   final _TtsReadPcmDart readPcm;
   final _TtsLastErrorDart lastError;
+  final _TtsSetCancelFlagDart? setCancelFlag;
+  final ggml_backend_sched_eval_callback evalCallback;
 
   const _TtsApi({
     required this.apiVersion,
@@ -8558,10 +8614,14 @@ class _TtsApi {
     required this.getOutputInfo,
     required this.readPcm,
     required this.lastError,
+    required this.setCancelFlag,
+    required this.evalCallback,
   });
 
   static _TtsApi? tryLoad(DynamicLibrary library) {
     try {
+      const setCancelFlagSymbol = 'llama_dart_tts_set_cancel_flag';
+      const evalCallbackSymbol = 'llama_dart_tts_eval_callback';
       return _TtsApi(
         apiVersion: library
             .lookupFunction<_TtsApiVersionNative, _TtsApiVersionDart>(
@@ -8603,6 +8663,17 @@ class _TtsApi {
             .lookupFunction<_TtsLastErrorNative, _TtsLastErrorDart>(
               'llama_dart_tts_last_error',
             ),
+        setCancelFlag: library.providesSymbol(setCancelFlagSymbol)
+            ? library.lookupFunction<
+                _TtsSetCancelFlagNative,
+                _TtsSetCancelFlagDart
+              >(setCancelFlagSymbol)
+            : null,
+        evalCallback: library.providesSymbol(evalCallbackSymbol)
+            ? library.lookup<
+                NativeFunction<ggml_backend_sched_eval_callbackFunction>
+              >(evalCallbackSymbol)
+            : nullptr,
       );
     } catch (_) {
       return null;
diff --git a/lib/src/backends/llama_cpp/worker.dart b/lib/src/backends/llama_cpp/worker.dart
index b2dd81d4a..2b3904a7c 100644
--- a/lib/src/backends/llama_cpp/worker.dart
+++ b/lib/src/backends/llama_cpp/worker.dart
@@ -289,6 +289,7 @@ void runLlamaWorkerForTesting(
                   message.contextHandle,
                   message.mmContextHandle,
                   message.request,
+                  message.cancelFlagAddress,
                   onProgress: (progress) {
                     message.sendPort.send(
                       TextToSpeechProgressResponse(progress),
diff --git a/lib/src/backends/llama_cpp/worker_messages.dart b/lib/src/backends/llama_cpp/worker_messages.dart
index d3a68360c..c6fe50ff6 100644
--- a/lib/src/backends/llama_cpp/worker_messages.dart
+++ b/lib/src/backends/llama_cpp/worker_messages.dart
@@ -352,11 +352,19 @@ class TextToSpeechSynthesizeRequest extends WorkerRequest {
   /// Synthesis inputs and sampling parameters.
   final BackendTextToSpeechRequest request;
 
+  /// Address of the one-byte cancel flag the sender owns.
+  ///
+  /// The sender sets the byte to nonzero to cancel and keeps it allocated
+  /// until this request's terminal response arrives or the worker acknowledges
+  /// a [DisposeRequest].
+  final int cancelFlagAddress;
+
   /// Creates a TTS synthesis request.
   TextToSpeechSynthesizeRequest(
     this.contextHandle,
     this.mmContextHandle,
     this.request,
+    this.cancelFlagAddress,
     super.sendPort,
   );
 }
diff --git a/lib/src/hook/native_release_pins.dart b/lib/src/hook/native_release_pins.dart
index 92ef44d98..38ee2afcf 100644
--- a/lib/src/hook/native_release_pins.dart
+++ b/lib/src/hook/native_release_pins.dart
@@ -5,7 +5,7 @@
 library;
 
 /// `leehack/llamadart-native` release tag downloaded for the llama.cpp runtime.
-const llamaCppTag = 'v0.4.1';
+const llamaCppTag = 'v0.4.1-1';
 
 /// `leehack/litert-lm-native` release tag downloaded for the LiteRT-LM runtime.
 const liteRtLmReleaseTag = 'v0.17.0-6';
diff --git a/packages/llamadart_llama_cpp_flutter/CHANGELOG.md b/packages/llamadart_llama_cpp_flutter/CHANGELOG.md
index 7bbc8f80a..6350b70be 100644
--- a/packages/llamadart_llama_cpp_flutter/CHANGELOG.md
+++ b/packages/llamadart_llama_cpp_flutter/CHANGELOG.md
@@ -1,3 +1,7 @@
+## Unreleased
+
+* Updated Apple SwiftPM native pin to `leehack/llamadart-native@v0.4.1-1`.
+
 ## 0.0.19
 
 * Updated Apple SwiftPM native pin to `leehack/llamadart-native@v0.4.1`.
diff --git a/packages/llamadart_llama_cpp_flutter/README.md b/packages/llamadart_llama_cpp_flutter/README.md
index 13566e3f1..993820443 100644
--- a/packages/llamadart_llama_cpp_flutter/README.md
+++ b/packages/llamadart_llama_cpp_flutter/README.md
@@ -20,7 +20,7 @@ dependencies:
 This package has no runtime Dart API of its own. Import `package:llamadart`
 normally from the core package and use `LlamaBackend()` / `LlamaEngine` there.
 
-The Apple SwiftPM manifest pins `leehack/llamadart-native@v0.4.1`.
+The Apple SwiftPM manifest pins `leehack/llamadart-native@v0.4.1-1`.
 
 Source for this package lives in
 `packages/llamadart_llama_cpp_flutter` in the
diff --git a/packages/llamadart_llama_cpp_flutter/darwin/llamadart_llama_cpp_flutter/Package.swift b/packages/llamadart_llama_cpp_flutter/darwin/llamadart_llama_cpp_flutter/Package.swift
index bd96187d9..8f7d808d4 100644
--- a/packages/llamadart_llama_cpp_flutter/darwin/llamadart_llama_cpp_flutter/Package.swift
+++ b/packages/llamadart_llama_cpp_flutter/darwin/llamadart_llama_cpp_flutter/Package.swift
@@ -4,7 +4,7 @@ import PackageDescription
 
 let packageRoot = URL(fileURLWithPath: #filePath).deletingLastPathComponent()
 let artifactsRoot = packageRoot.appendingPathComponent("Artifacts")
-let llamaCppTag = "v0.4.1"
+let llamaCppTag = "v0.4.1-1"
 
 func localArtifactPath(_ name: String) -> String? {
     let path = artifactsRoot.appendingPathComponent(name).path
@@ -47,7 +47,7 @@ let package = Package(
             repository: "leehack/llamadart-native",
             artifactName: "llamadart-native-apple-xcframework-\(llamaCppTag).zip",
             tag: llamaCppTag,
-            checksum: "16175a4084a14af63cc66e78b9b2a8173d2e14db86775af280c859438bd8616a"
+            checksum: "bab353ce17dfbac9afe65ade73751d67f110b1807a1fb365b4d1df2db25a5418"
         ),
         .target(
             name: "llamadart_llama_cpp_flutter",
diff --git a/test/e2e/backends/text_to_speech_e2e_test.dart b/test/e2e/backends/text_to_speech_e2e_test.dart
index db6ee24b5..adfe280ea 100644
--- a/test/e2e/backends/text_to_speech_e2e_test.dart
+++ b/test/e2e/backends/text_to_speech_e2e_test.dart
@@ -3,14 +3,30 @@
 @Timeout(Duration(minutes: 15))
 library;
 
+import 'dart:ffi';
 import 'dart:io';
 
+import 'package:ffi/ffi.dart';
 import 'package:llamadart/llamadart.dart';
+import 'package:llamadart/src/backends/llama_cpp/llama_cpp_service.dart';
 import 'package:test/test.dart';
 
 const _modelPathKey = 'LLAMADART_TTS_MODEL_PATH';
 const _mmprojPathKey = 'LLAMADART_TTS_MMPROJ_PATH';
 const _outputPathKey = 'LLAMADART_TTS_OUTPUT_PATH';
+const _serviceRequest = BackendTextToSpeechRequest(
+  text: 'Hello from llamadart.',
+  language: 'english',
+  maxFrames: 64,
+  seed: 1,
+);
+final _cancelled = throwsA(
+  isA<LlamaTextToSpeechException>().having(
+    (error) => error.message,
+    'message',
+    'Text-to-speech synthesis was cancelled.',
+  ),
+);
 
 void main() {
   test('synthesizes a playable WAV through the public API', () async {
@@ -86,6 +102,85 @@ void main() {
       await engine.dispose();
     }
   });
+
+  test('a cancel flag raised before synthesis starts cancels it', () async {
+    await _withTextToSpeechService((service, context, projector) async {
+      final flag = calloc<Int8>()..value = 1;
+      try {
+        final progress = <BackendTextToSpeechProgress>[];
+        await expectLater(
+          service.synthesizeTextToSpeech(
+            context,
+            projector,
+            _serviceRequest,
+            flag.address,
+            onProgress: progress.add,
+          ),
+          _cancelled,
+        );
+        expect(progress, isEmpty);
+      } finally {
+        calloc.free(flag);
+      }
+    });
+  });
+
+  test('a cancel flag raised between steps stops the next step only when the '
+      'runtime can attach it', () async {
+    await _withTextToSpeechService((service, context, projector) async {
+      final attachable =
+          service.debugTtsCancelExportsForTesting()?.setCancelFlag ?? false;
+      final flag = calloc<Int8>();
+      try {
+        var steps = 0;
+        final synthesis = service.synthesizeTextToSpeech(
+          context,
+          projector,
+          _serviceRequest,
+          flag.address,
+          onProgress: (_) {
+            steps += 1;
+            flag.value = 1;
+          },
+        );
+        if (attachable) {
+          await expectLater(synthesis, _cancelled);
+          expect(steps, 1);
+        } else {
+          expect((await synthesis).samples, isNotEmpty);
+          expect(steps, greaterThan(1));
+        }
+      } finally {
+        calloc.free(flag);
+      }
+    });
+  });
+}
+
+Future<void> _withTextToSpeechService(
+  Future<void> Function(LlamaCppService service, int context, int projector)
+  body,
+) async {
+  final modelPath = _requiredFile(_modelPathKey);
+  final mmprojPath = _requiredFile(_mmprojPathKey);
+  if (modelPath == null || mmprojPath == null) {
+    return;
+  }
+  final service = LlamaCppService()..setLogLevel(LlamaLogLevel.none);
+  try {
+    service.initializeBackend();
+    const modelParams = ModelParams(
+      contextSize: 4096,
+      preferredBackend: GpuBackend.cpu,
+      gpuLayers: 0,
+    );
+    final model = service.loadModel(modelPath, modelParams);
+    final context = service.createContext(model, modelParams);
+    final projector = service.createMultimodalContext(model, mmprojPath);
+    await body(service, context, projector);
+  } finally {
+    service.dispose();
+  }
 }
 
 String? _requiredFile(String environmentKey) {
diff --git a/test/integration/backends/llama_cpp/native_symbol_integration_test.dart b/test/integration/backends/llama_cpp/native_symbol_integration_test.dart
index 17a67434f..162ab1c29 100644
--- a/test/integration/backends/llama_cpp/native_symbol_integration_test.dart
+++ b/test/integration/backends/llama_cpp/native_symbol_integration_test.dart
@@ -68,6 +68,11 @@ const _mtmdChunkEvalSymbols = [
   'mtmd_helper_decode_image_chunk',
 ];
 
+const _ttsCancelSymbols = [
+  'llama_dart_tts_eval_callback',
+  'llama_dart_tts_set_cancel_flag',
+];
+
 const _aloraMetadataSymbols = [
   'llama_adapter_get_alora_n_invocation_tokens',
   'llama_adapter_get_alora_invocation_tokens',
@@ -978,6 +983,39 @@ void main() {
       );
     });
 
+    test('the TTS API has both cancel exports and mtmd contexts get its eval '
+        'callback', () {
+      final wrapper = _llamadartWrapperLibraryFileOrNull();
+      expect(
+        wrapper,
+        isNotNull,
+        reason: 'Expected the llama.cpp wrapper library.',
+      );
+      final service = LlamaCppService();
+      final exports = service.debugTtsCancelExportsForTesting();
+      final address = service.debugMtmdEvalCallbackAddressForTesting();
+      expect(service.getStartupDiagnostics(), isEmpty);
+
+      if (!_fileContainsAscii(wrapper!, 'llama_dart_tts_step')) {
+        expect(exports, isNull);
+        expect(address, 0);
+        return;
+      }
+      expect(
+        _ttsCancelSymbols
+            .where((symbol) => _fileContainsAscii(wrapper, symbol))
+            .toList(),
+        _ttsCancelSymbols,
+      );
+      expect(exports, (evalCallback: true, setCancelFlag: true));
+      expect(
+        address,
+        ffi.DynamicLibrary.open(
+          wrapper.path,
+        ).lookup<ffi.Void>('llama_dart_tts_eval_callback').address,
+      );
+    });
+
     test('Verify core llama symbols are resolvable', () {
       expect(() => llama_backend_init(), returnsNormally);
       expect(llama_version().cast<Utf8>().toDartString(), isNotEmpty);
diff --git a/test/unit/backends/llama_cpp/llama_cpp_backend_test.dart b/test/unit/backends/llama_cpp/llama_cpp_backend_test.dart
index d5ae89f39..c3c0fd0b4 100644
--- a/test/unit/backends/llama_cpp/llama_cpp_backend_test.dart
+++ b/test/unit/backends/llama_cpp/llama_cpp_backend_test.dart
@@ -2,9 +2,12 @@
 library;
 
 import 'dart:async';
+import 'dart:ffi';
+import 'dart:io';
 import 'dart:isolate';
 import 'dart:typed_data';
 
+import 'package:ffi/ffi.dart';
 import 'package:llamadart/src/backends/backend.dart';
 import 'package:llamadart/src/backends/llama_cpp/llama_cpp_backend.dart';
 import 'package:llamadart/src/backends/llama_cpp/llama_cpp_service.dart';
@@ -339,6 +342,164 @@ void main() {
       },
     );
 
+    test(
+      'a cancel before the request is sent is already in its flag',
+      () async {
+        harness.holdTextToSpeech = true;
+        final pending = backend.synthesizeTextToSpeech(
+          22,
+          33,
+          const BackendTextToSpeechRequest(text: 'Cancel early.'),
+        );
+        backend.cancelTextToSpeech();
+        await harness.textToSpeechStarted.future;
+        await Future<void>.delayed(Duration.zero);
+        expect(harness.cancelFlagAtRequest, 1);
+        expect(
+          harness.received.whereType<TextToSpeechCancelRequest>(),
+          hasLength(1),
+        );
+        harness.finishHeldTextToSpeech();
+        await pending;
+      },
+    );
+
+    test(
+      'frees the cancel flag on the terminal response, not on cancel',
+      () async {
+        final allocator = _RecordingAllocator();
+        addTearDown(allocator.release);
+        final flagBackend = NativeLlamaBackend(
+          initialSendPort: harness.sendPort,
+          textToSpeechCancelFlagAllocator: allocator,
+        );
+        harness.holdTextToSpeech = true;
+        final pending = flagBackend.synthesizeTextToSpeech(
+          22,
+          33,
+          const BackendTextToSpeechRequest(text: 'Hold.'),
+        );
+        await harness.textToSpeechStarted.future;
+        final address = harness.received
+            .whereType<TextToSpeechSynthesizeRequest>()
+            .last
+            .cancelFlagAddress;
+        expect(allocator.allocated, [address]);
+        expect(harness.cancelFlagAtRequest, 0);
+
+        flagBackend.cancelTextToSpeech();
+        await Future<void>.delayed(Duration.zero);
+        expect(Pointer<Int8>.fromAddress(address).value, 1);
+        expect(allocator.freed, isEmpty);
+
+        harness.finishHeldTextToSpeech();
+        await pending;
+        expect(allocator.freed, [address]);
+        await flagBackend.dispose();
+        expect(allocator.freed, [address]);
+      },
+    );
+
+    test(
+      'dispose frees a running synthesis cancel flag after the worker acks',
+      () async {
+        final allocator = _RecordingAllocator();
+        addTearDown(allocator.release);
+        final flagBackend = NativeLlamaBackend(
+          initialSendPort: harness.sendPort,
+          textToSpeechCancelFlagAllocator: allocator,
+        );
+        harness.holdTextToSpeech = true;
+        final pending = flagBackend.synthesizeTextToSpeech(
+          22,
+          33,
+          const BackendTextToSpeechRequest(text: 'Hold.'),
+        );
+        await harness.textToSpeechStarted.future;
+        final flag = Pointer<Int8>.fromAddress(
+          harness.received
+              .whereType<TextToSpeechSynthesizeRequest>()
+              .last
+              .cancelFlagAddress,
+        );
+
+        harness.holdNextDispose = true;
+        final disposed = flagBackend.dispose();
+        await harness.disposeReceived.future;
+        expect(flag.value, 1);
+        expect(allocator.freed, isEmpty);
+
+        harness.acknowledgeHeldDispose();
+        await disposed;
+        expect(allocator.freed, [flag.address]);
+
+        flag.value = 0;
+        flagBackend.cancelTextToSpeech();
+        expect(flag.value, 0);
+
+        harness.finishHeldTextToSpeech();
+        await pending;
+        expect(allocator.freed, [flag.address]);
+      },
+    );
+
+    test('a failed worker startup frees the cancel flag', () async {
+      final allocator = _RecordingAllocator();
+      addTearDown(allocator.release);
+      final flagBackend = NativeLlamaBackend(
+        workerEntrypoint: _failingInitializationWorkerEntry,
+        textToSpeechCancelFlagAllocator: allocator,
+      );
+      addTearDown(flagBackend.dispose);
+
+      await expectLater(
+        flagBackend.synthesizeTextToSpeech(
+          22,
+          33,
+          const BackendTextToSpeechRequest(text: 'Never sent.'),
+        ),
+        throwsA(isA<LlamaBackendInitializationException>()),
+      );
+      expect(allocator.allocated, hasLength(1));
+      expect(allocator.freed, allocator.allocated);
+    });
+
+    test(
+      'dispose keeps the cancel flag of a synthesis claimed during it',
+      () async {
+        final allocator = _RecordingAllocator();
+        addTearDown(allocator.release);
+        final flagBackend = NativeLlamaBackend(
+          workerEntrypoint: _delayedTextToSpeechRejectingWorkerEntry,
+          textToSpeechCancelFlagAllocator: allocator,
+        );
+        addTearDown(flagBackend.dispose);
+        final startup = flagBackend.modelLoad(
+          'never.gguf',
+          const ModelParams(),
+        );
+        final disposed = flagBackend.dispose();
+        final synthesis = flagBackend.synthesizeTextToSpeech(
+          22,
+          33,
+          const BackendTextToSpeechRequest(text: 'Next worker.'),
+        );
+        await expectLater(
+          startup,
+          throwsA(isA<LlamaBackendInitializationException>()),
+        );
+        await disposed;
+        expect(allocator.allocated, hasLength(1));
+        expect(allocator.freed, isEmpty);
+
+        await expectLater(
+          synthesis,
+          throwsA(isA<LlamaTextToSpeechException>()),
+        );
+        expect(allocator.freed, allocator.allocated);
+      },
+    );
+
     test('preserves speech error subtypes from the worker', () async {
       await expectLater(
         backend.synthesizeTextToSpeech(
@@ -596,6 +757,41 @@ void main() {
     });
   });
 
+  test(
+    'cancel reaches a worker blocked in a native step through the flag',
+    () async {
+      final events = ReceivePort();
+      final eventQueue = StreamIterator<Object?>(events);
+      final isolate = await Isolate.spawn(_flagPollingWorker, events.sendPort);
+      addTearDown(() {
+        isolate.kill(priority: Isolate.immediate);
+        events.close();
+      });
+      expect(await eventQueue.moveNext(), isTrue);
+      final backend = NativeLlamaBackend(
+        initialSendPort: eventQueue.current as SendPort,
+      );
+      final pending = backend.synthesizeTextToSpeech(
+        22,
+        33,
+        const BackendTextToSpeechRequest(text: 'Blocked.'),
+      );
+      expect(await eventQueue.moveNext(), isTrue);
+      expect(eventQueue.current, 'in-step');
+      backend.cancelTextToSpeech();
+      await expectLater(
+        pending,
+        throwsA(
+          isA<LlamaTextToSpeechException>().having(
+            (error) => error.message,
+            'message',
+            'flag observed',
+          ),
+        ),
+      );
+    },
+  );
+
   test('modelFree and contextFree are no-op without worker port', () async {
     final backend = NativeLlamaBackend();
 
@@ -1031,6 +1227,26 @@ void main() {
   });
 }
 
+void _delayedTextToSpeechRejectingWorkerEntry(SendPort initialSendPort) {
+  final receivePort = ReceivePort();
+  initialSendPort.send(receivePort.sendPort);
+  receivePort.listen((message) async {
+    switch (message) {
+      case WorkerHandshake():
+        await Future<void>.delayed(const Duration(milliseconds: 50));
+        message.sendPort.send(DoneResponse());
+      case TextToSpeechSynthesizeRequest():
+        message.sendPort.send(
+          ErrorResponse('rejected', kind: WorkerErrorKind.textToSpeech),
+        );
+      case DisposeRequest():
+        message.sendPort.send(null);
+        receivePort.close();
+        Isolate.exit();
+    }
+  });
+}
+
 void _failingInitializationWorkerEntry(SendPort initialSendPort) {
   runLlamaWorkerForTesting(
     initialSendPort,
@@ -1221,6 +1437,9 @@ class _FakeWorkerHarness {
   DisposeRequest? _heldDispose;
   Completer<void> textToSpeechStarted = Completer<void>();
   TextToSpeechSynthesizeRequest? _heldTextToSpeech;
+  int? cancelFlagAtRequest;
+  bool holdNextDispose = false;
+  final Completer<void> disposeReceived = Completer<void>();
 
   _FakeWorkerHarness() {
     _port.listen((message) {
@@ -1386,6 +1605,9 @@ class _FakeWorkerHarness {
             ),
           );
         case TextToSpeechSynthesizeRequest():
+          cancelFlagAtRequest = Pointer<Int8>.fromAddress(
+            message.cancelFlagAddress,
+          ).value;
           if (!textToSpeechStarted.isCompleted) {
             textToSpeechStarted.complete();
           }
@@ -1537,7 +1759,11 @@ class _FakeWorkerHarness {
             message.sendPort.send(DoneResponse());
           }
         case DisposeRequest():
-          if (holdDispose) {
+          if (holdNextDispose) {
+            holdNextDispose = false;
+            _heldDispose = message;
+            disposeReceived.complete();
+          } else if (holdDispose) {
             _heldDispose = message;
           } else {
             message.sendPort.send(DoneResponse());
@@ -1550,6 +1776,11 @@ class _FakeWorkerHarness {
 
   SendPort get sendPort => _port.sendPort;
 
+  void acknowledgeHeldDispose() {
+    _heldDispose?.sendPort.send(DoneResponse());
+    _heldDispose = null;
+  }
+
   void releaseDispose() {
     holdDispose = false;
     _heldDispose?.sendPort.send(DoneResponse());
@@ -1591,3 +1822,47 @@ class _FakeWorkerHarness {
     _port.close();
   }
 }
+
+void _flagPollingWorker(SendPort events) {
+  final requests = ReceivePort();
+  events.send(requests.sendPort);
+  requests.listen((message) {
+    if (message is! TextToSpeechSynthesizeRequest) {
+      return;
+    }
+    final flag = Pointer<Int8>.fromAddress(message.cancelFlagAddress);
+    events.send('in-step');
+    final watch = Stopwatch()..start();
+    while (flag.value == 0 && watch.elapsed < const Duration(seconds: 10)) {
+      sleep(const Duration(milliseconds: 1));
+    }
+    message.sendPort.send(
+      ErrorResponse(
+        flag.value == 0 ? 'flag never observed' : 'flag observed',
+        kind: WorkerErrorKind.textToSpeech,
+      ),
+    );
+  });
+}
+
+final class _RecordingAllocator implements Allocator {
+  final List<int> allocated = <int>[];
+  final List<int> freed = <int>[];
+
+  @override
+  Pointer<T> allocate<T extends NativeType>(int byteCount, {int? alignment}) {
+    final pointer = malloc.allocate<Uint8>(byteCount, alignment: alignment);
+    pointer.asTypedList(byteCount).fillRange(0, byteCount, 0x5a);
+    allocated.add(pointer.address);
+    return pointer.cast<T>();
+  }
+
+  @override
+  void free(Pointer<NativeType> pointer) => freed.add(pointer.address);
+
+  void release() {
+    for (final address in allocated) {
+      malloc.free(Pointer<Uint8>.fromAddress(address));
+    }
+  }
+}
diff --git a/test/unit/backends/llama_cpp/worker_test.dart b/test/unit/backends/llama_cpp/worker_test.dart
index 256e89a9d..4d6213bbc 100644
--- a/test/unit/backends/llama_cpp/worker_test.dart
+++ b/test/unit/backends/llama_cpp/worker_test.dart
@@ -434,10 +434,12 @@ void main() {
             2,
             3,
             const BackendTextToSpeechRequest(text: 'Hello.'),
+            0x5a5a0,
             responsePort.sendPort,
           ),
         );
         await service.synthesisStarted.future;
+        expect(service.receivedCancelFlagAddress, 0x5a5a0);
         worker.sendPort.send(TextToSpeechCancelRequest());
         await service.cancelObserved.future;
         service.releaseSynthesis();
@@ -486,6 +488,7 @@ void main() {
               2,
               3,
               const BackendTextToSpeechRequest(text: 'Hello.'),
+              0x5a5a0,
               sendPort,
             ),
           );
@@ -863,6 +866,7 @@ class _BlockingTextToSpeechService extends LlamaCppService {
   final Completer<void> cancelObserved = Completer<void>();
   final Completer<void> _releaseSynthesis = Completer<void>();
   int cancelCalls = 0;
+  int? receivedCancelFlagAddress;
 
   void releaseSynthesis() {
     if (!_releaseSynthesis.isCompleted) {
@@ -880,9 +884,11 @@ class _BlockingTextToSpeechService extends LlamaCppService {
   Future<BackendTextToSpeechResult> synthesizeTextToSpeech(
     int contextHandle,
     int mmContextHandle,
-    BackendTextToSpeechRequest request, {
+    BackendTextToSpeechRequest request,
+    int cancelFlagAddress, {
     void Function(BackendTextToSpeechProgress progress)? onProgress,
   }) async {
+    receivedCancelFlagAddress = cancelFlagAddress;
     if (!synthesisStarted.isCompleted) {
       synthesisStarted.complete();
     }
@@ -931,7 +937,8 @@ class _ThrowingTextToSpeechService extends LlamaCppService {
   Future<BackendTextToSpeechResult> synthesizeTextToSpeech(
     int contextHandle,
     int mmContextHandle,
-    BackendTextToSpeechRequest request, {
+    BackendTextToSpeechRequest request,
+    int cancelFlagAddress, {
     void Function(BackendTextToSpeechProgress progress)? onProgress,
   }) async {
     throw exception;
diff --git a/test/unit/tooling/check_webgpu_bridge_tag_test.dart b/test/unit/tooling/check_webgpu_bridge_tag_test.dart
index 0ed779fda..751ac5568 100644
--- a/test/unit/tooling/check_webgpu_bridge_tag_test.dart
+++ b/test/unit/tooling/check_webgpu_bridge_tag_test.dart
@@ -696,18 +696,33 @@ void main() {
       );
     });
 
-    test('a different wrapper in the same family is not the approved anchor', () {
+    test('a different wrapper in the same family is not the approved pin', () {
       final root = _fakeRuntimeRepo(bridgeTag: 'v0.2.0', nativeTag: 'v0.2.0-2');
 
       expect(
         findBridgeRuntimeDrift(root, 'v0.2.0'),
         contains(
-          '$nativeLlamaCppTagPath pins v0.2.0-2, expected the bridge-qualified '
-          'native anchor $bridgeNativeReleaseTag',
+          '$nativeLlamaCppTagPath pins v0.2.0-2, expected the bridge-approved '
+          'native pin $bridgeApprovedNativePin',
         ),
       );
     });
 
+    test(
+      'the qualified anchor is not accepted in place of the approved pin',
+      () {
+        final root = _fakeRuntimeRepo(
+          bridgeTag: bridgeLlamaCppTag,
+          nativeTag: bridgeNativeReleaseTag,
+        );
+
+        expect(findBridgeRuntimeDrift(root, bridgeLlamaCppTag), [
+          '$nativeLlamaCppTagPath pins $bridgeNativeReleaseTag, expected the '
+              'bridge-approved native pin $bridgeApprovedNativePin',
+        ]);
+      },
+    );
+
     test('stale divergence prose is reported', () {
       final root = _fakeRuntimeRepo(
         bridgeTag: 'v0.2.0',
@@ -724,7 +739,7 @@ void main() {
     test('parity prose is accepted for matching upstream tags', () {
       final root = _fakeRuntimeRepo(
         bridgeTag: bridgeLlamaCppTag,
-        nativeTag: bridgeNativeReleaseTag,
+        nativeTag: bridgeApprovedNativePin,
         parityWording: true,
       );
 
diff --git a/tool/testing/check_webgpu_bridge_tag.dart b/tool/testing/check_webgpu_bridge_tag.dart
index dbf127a71..c06108ac7 100644
--- a/tool/testing/check_webgpu_bridge_tag.dart
+++ b/tool/testing/check_webgpu_bridge_tag.dart
@@ -383,6 +383,11 @@ const String bridgeNativeRepository = 'leehack/llamadart-native';
 /// The native release tag the pinned bridge assets were qualified against.
 const String bridgeNativeReleaseTag = 'v0.4.1';
 
+/// The native pin approved beside the pinned bridge assets: either
+/// [bridgeNativeReleaseTag] or a native wrapper rebuild of it from the same
+/// upstream llama.cpp commit.
+const String bridgeApprovedNativePin = 'v0.4.1-1';
+
 /// The asset repository release that published the pinned bridge assets.
 const String bridgeAssetsReleaseId = '389783936';
 
@@ -439,11 +444,11 @@ final List<BridgeTagPin> bridgeProvenancePins = <BridgeTagPin>[
   BridgeTagPin(
     'doc/webgpu_bridge.md',
     RegExp(
-      r'^\(`(?<nativeReleaseTag>[^`]+)`, both built from upstream llama\.cpp '
+      r'^\(`(?<nativePin>[^`]+)`, both built from upstream llama\.cpp '
       r'`(?<upstreamTag>[^`@]+)@(?<upstreamCommit>[0-9a-f]{40})`\)\r?\n'
       r'even though the bridge asset tag `v\d+\.\d+\.\d+` differs from the '
       r'native runtime tag\r?\n'
-      r'`v\d+\.\d+\.\d+`\. Provenance for this immutable consumer artifact: '
+      r'`v\d+\.\d+\.\d+(?:-\d+)?`\. Provenance for this immutable consumer artifact: '
       r'release `(?<releaseId>\d+)`,\r?\n'
       r'tag commit `(?<tagCommit>[0-9a-f]{40})`, bridge source\r?\n'
       r'`(?<bridgeCommit>[0-9a-f]{40})`, and manifest SHA-256\r?\n'
@@ -454,11 +459,11 @@ final List<BridgeTagPin> bridgeProvenancePins = <BridgeTagPin>[
   BridgeTagPin(
     'website/docs/platforms/webgpu-bridge.md',
     RegExp(
-      r'^  \(`(?<nativeReleaseTag>[^`]+)`, both built from upstream '
+      r'^  \(`(?<nativePin>[^`]+)`, both built from upstream '
       r'`(?<upstreamTag>[^`@]+)@(?<upstreamCommit>[0-9a-f]{40})`\)\r?\n'
       r'  even though the bridge asset tag `v\d+\.\d+\.\d+` differs from the '
       r'native runtime tag\r?\n'
-      r'  `v\d+\.\d+\.\d+`\. Pinned artifact provenance: release '
+      r'  `v\d+\.\d+\.\d+(?:-\d+)?`\. Pinned artifact provenance: release '
       r'`(?<releaseId>\d+)`, tag commit\r?\n'
       r'  `(?<tagCommit>[0-9a-f]{40})`, bridge source\r?\n'
       r'  `(?<bridgeCommit>[0-9a-f]{40})`, manifest SHA-256\r?\n'
@@ -470,7 +475,7 @@ final List<BridgeTagPin> bridgeProvenancePins = <BridgeTagPin>[
 
 /// The expected value behind each [bridgeProvenancePins] group name.
 Map<String, String> get bridgeProvenanceValues => <String, String>{
-  'nativeReleaseTag': bridgeNativeReleaseTag,
+  'nativePin': bridgeApprovedNativePin,
   'upstreamTag': bridgeLlamaCppTag,
   'upstreamCommit': bridgeLlamaCppCommit,
   'releaseId': bridgeAssetsReleaseId,
@@ -532,10 +537,10 @@ List<String> findBridgeRuntimeDrift(Directory repoRoot, String bridgeTag) {
       'expected stable vMAJOR.MINOR.PATCH(-N) or nightly bNNNN(-N|-llamadart.N)',
     );
   } else {
-    if (nativeTag != bridgeNativeReleaseTag) {
+    if (nativeTag != bridgeApprovedNativePin) {
       problems.add(
-        '$nativeLlamaCppTagPath pins $nativeTag, expected the bridge-qualified '
-        'native anchor $bridgeNativeReleaseTag',
+        '$nativeLlamaCppTagPath pins $nativeTag, expected the bridge-approved '
+        'native pin $bridgeApprovedNativePin',
       );
     }
     if (normalizedNativeFamily != bridgeTag) {
diff --git a/website/docs/changelog/recent-releases.md b/website/docs/changelog/recent-releases.md
index 73e986915..127352928 100644
--- a/website/docs/changelog/recent-releases.md
+++ b/website/docs/changelog/recent-releases.md
@@ -9,6 +9,17 @@ For canonical full release notes, use:
 
 ## Unreleased
 
+- Updated the default llama.cpp native runtime pin to
+  `leehack/llamadart-native@v0.4.1-1`, keeping the `v0.4.1` llama.cpp
+  ABI/bindings while picking up wrapper-only native fixes. Refreshed the
+  `llamadart_llama_cpp_flutter` Apple SwiftPM checksum and aligned current
+  README/website native override docs.
+- Stop a Qwen3-TTS audio decode at its next chunk boundary when native
+  text-to-speech is cancelled, instead of finishing the native step in
+  progress first. This needs llamadart-native v0.4.1-1 or later; older
+  runtimes keep the previous behaviour
+  ([llamadart-native#86](https://github.com/leehack/llamadart-native/issues/86),
+  [#322](https://github.com/leehack/llamadart/issues/322)).
 - Add an experimental `DecisionEngine` for Laya-style decision models (a
   ModernBERT encoder GGUF plus a safetensors head) on native llama.cpp, with
   typed `ChoiceKey`, `ScoreKey` and `NoulKey` questions
diff --git a/website/docs/getting-started/installation.md b/website/docs/getting-started/installation.md
index 90681fedd..d300edd9a 100644
--- a/website/docs/getting-started/installation.md
+++ b/website/docs/getting-started/installation.md
@@ -88,7 +88,7 @@ hooks:
     llamadart:
       # Optional. Defaults to llamadart's tested native runtime pin.
       # Use a leehack/llamadart-native release tag when testing another build.
-      llamadart_native_tag: v0.4.1
+      llamadart_native_tag: v0.4.1-1
 
       # Optional. GitHub repository slug or github.com URL.
       llamadart_native_repository: leehack/llamadart-native
@@ -115,7 +115,7 @@ per-target module list.
 Native source overrides are for compatibility testing. They do not regenerate
 Dart FFI bindings or symbol lookups, so the selected binary still must be ABI-
 and symbol-compatible with the default
-`leehack/llamadart-native@v0.4.1` runtime.
+`leehack/llamadart-native@v0.4.1-1` runtime.
 
 Available native tags are published on the
 [`leehack/llamadart-native` releases page](https://github.com/leehack/llamadart-native/releases).
@@ -143,7 +143,7 @@ gh release list --repo leehack/llamadart-native --limit 20
 
 Before overriding, confirm the release includes the asset for your target. The
 hook downloads files named `llamadart-native-<bundle>-<tag>.tar.gz`, for example
-`llamadart-native-windows-x64-v0.4.1.tar.gz`.
+`llamadart-native-windows-x64-v0.4.1-1.tar.gz`.
 For local testing, `llamadart_native_path` may point directly at a bundle
 archive, at an extracted bundle directory, or at a directory containing
 `<tag>/<bundle>/`, `<bundle>/`, or the expected archive file.
diff --git a/website/docs/guides/performance-tuning.md b/website/docs/guides/performance-tuning.md
index aa592c672..4cdc66f52 100644
--- a/website/docs/guides/performance-tuning.md
+++ b/website/docs/guides/performance-tuning.md
@@ -148,7 +148,7 @@ Guidelines:
 - DSpark is an experimental, opt-in native llama.cpp external draft-model
   strategy. Use `SpeculativeDecodingConfig.draftDspark(draftModelPath: ...)`;
   it is never selected automatically, maps to upstream `draft-dspark`, requires
-  at least the `b10356-llamadart.1` wrapper fix; the package-pinned `v0.4.1`
+  at least the `b10356-llamadart.1` wrapper fix; the package-pinned `v0.4.1-1`
   runtime satisfies that ABI, supports speculators-format checkpoints, and
   adds LFM2 target/draft support. DSpark remains subject to
   target/draft/backend compatibility.
diff --git a/website/docs/guides/text-to-speech.md b/website/docs/guides/text-to-speech.md
index cec3034ad..a4c2c883f 100644
--- a/website/docs/guides/text-to-speech.md
+++ b/website/docs/guides/text-to-speech.md
@@ -129,7 +129,10 @@ responsible for its own files, byte buffers, permissions, and disclosures.
 ## Cancellation, concurrency, and buffering
 
 Call `task.cancel()` to request cooperative cancellation. Cancelling only the
-event-stream subscription does not cancel synthesis.
+event-stream subscription does not cancel synthesis. On native llama.cpp with
+llamadart-native v0.4.1-1 or later, a cancel stops a Qwen3-TTS audio decode in
+progress at its next chunk boundary. Older runtimes finish the native step in
+progress first, which can include the whole audio decode.
 
 All typed STT and TTS wrappers over one `LlamaEngine` share a one-task speech
 lease. Do not run chat generation, transcription, or another synthesis on the
diff --git a/website/docs/platforms/support-matrix.md b/website/docs/platforms/support-matrix.md
index f47997706..0c307c9ab 100644
--- a/website/docs/platforms/support-matrix.md
+++ b/website/docs/platforms/support-matrix.md
@@ -8,7 +8,7 @@ backend-module configuration for
 `llamadart`.
 
 The native-assets hook currently pins `llamadart-native` tag
-`v0.4.1` and
+`v0.4.1-1` and
 `litert-lm-native` release `v0.17.0-6` (`lib/src/hook/native_release_pins.dart`). Apps can
 override the llama.cpp native GitHub source with
 `hooks.user_defines.llamadart.llamadart_native_tag` and
@@ -312,7 +312,7 @@ device/model bundle, use `cpu` or `gpu` for that artifact.
   a web load failure as a package bug. The [WebGPU Bridge](./webgpu-bridge)
   page has the browser-console probe and Flutter Web smoke-test path.
 
-## Current llama.cpp module availability by bundle (`v0.4.1`)
+## Current llama.cpp module availability by bundle (`v0.4.1-1`)
 
 | Bundle key | Available backend modules in bundle |
 | --- | --- |
@@ -360,7 +360,7 @@ hooks:
   user_defines:
     llamadart:
       # Optional. Defaults to llamadart's tested native runtime pin.
-      llamadart_native_tag: v0.4.1
+      llamadart_native_tag: v0.4.1-1
 
       # Optional. GitHub repository slug or github.com URL.
       llamadart_native_repository: leehack/llamadart-native
diff --git a/website/docs/platforms/webgpu-bridge.md b/website/docs/platforms/webgpu-bridge.md
index 4b478d7a6..b526144a1 100644
--- a/website/docs/platforms/webgpu-bridge.md
+++ b/website/docs/platforms/webgpu-bridge.md
@@ -202,15 +202,17 @@ cannot report success before the bridge exposes `prefetchModelToCache(...)`.
   remains unsupported; use memory64.
 - `v0.1.39+` remains the compatibility floor for bridge asset capabilities.
 - The pinned `v0.1.44` bridge assets embed llama.cpp `v0.4.1`, matching the native runtime
-  (`v0.4.1`, both built from upstream `v0.4.1@b29c606e28a01b1bc8c1351026a0fa6e616bf6c4`)
+  (`v0.4.1-1`, both built from upstream `v0.4.1@b29c606e28a01b1bc8c1351026a0fa6e616bf6c4`)
   even though the bridge asset tag `v0.1.44` differs from the native runtime tag
-  `v0.4.1`. Pinned artifact provenance: release `389783936`, tag commit
+  `v0.4.1-1`. Pinned artifact provenance: release `389783936`, tag commit
   `fdafd9f8cbdb9bf99c359536595eff9a23095379`, bridge source
   `89178be67c3c84300bc1b129182bd5bc5a8e21fc`, manifest SHA-256
   `8d61f453753ac7a7d839ac12318b70986a814748d86029993118c19454293aa9`. The
-  bridge assets provision an explicit 1 MiB stack for both wasm32 and memory64,
-  preventing graph-parameter growth from overflowing Emscripten's 64 KiB
-  default during memory64 Qwen3-ASR context construction.
+  bridge assets were qualified against native `v0.4.1`; `v0.4.1-1` rebuilds it
+  from the same upstream commit. The bridge assets provision an explicit 1 MiB
+  stack for both wasm32 and memory64, preventing graph-parameter growth from
+  overflowing Emscripten's 64 KiB default during memory64 Qwen3-ASR context
+  construction.
 - `v0.1.12+` bridge assets forward native-compatible `ModelParams` load
   tuning fields, including multi-sequence slots, KV cache type, flash attention,
   RoPE overrides, split mode, and main GPU.