fix(mlx): drain the pool in a finally so a failing in-flight op releases it too; fix the mid-generation test's call count

Review follow-ups: the post-op drain now runs in a finally block for the
TTS, STT and LLM folded callables, so an exception escaping the generation
(e.g. the voice-clone fallback failing) still releases the buffers the
worker held. The mid-generation test asserted one drain call but the
path legitimately produces two (unload_model's own and the post-op one);
a second test covers the error path.

backend/tests/test_mlx_unload_clears_cache.py

@@ -107,4 +107,30 @@ async def test_mlx_tts_generate_drains_cache_when_unloaded_mid_generation(monkey
 
     assert len(audio) == 2
     assert backend.model is None
-    mock_clear.assert_called_once()
+    # Once from unload_model() itself (reference drop) and once more after the
+    # in-flight generation released the model it had bound locally.
+    assert mock_clear.call_count == 2
+
+
+@pytest.mark.asyncio
+async def test_mlx_tts_generate_drains_cache_when_unloaded_and_generation_fails(monkeypatch):
+    """The post-generation drain also runs when the in-flight op raises."""
+    mock_clear = MagicMock()
+    monkeypatch.setattr(mlx_backend, "empty_mlx_cache", mock_clear)
+
+    backend = mlx_backend.MLXTTSBackend(model_size="0.6B")
+
+    class _FakeModel:
+        def generate(self, text, **kwargs):
+            backend.unload_model()
+            raise RuntimeError("boom")
+            yield  # pragma: no cover - makes this a generator
+
+    backend.model = _FakeModel()
+    backend._current_model_size = "0.6B"
+
+    with pytest.raises(RuntimeError, match="boom"):
+        await backend.generate("hello", {}, "en")
+
+    assert backend.model is None
+    assert mock_clear.call_count == 2