[ruccho]

Summary

On Android, the following warning was logged every frame while recording.

W Unity : [unienc] jni::wrapper::objects::global_ref: Dropping a GlobalRef in a detached thread. Fix your code if this message appears frequently (see the GlobalRef docs).

attach_current_thread() returned a scoped AttachGuard, which detaches a thread that was not attached before (e.g. a tokio worker) when the guard is dropped. Every GlobalRef dropped after the guard was therefore deleted on a detached thread, and the jni crate attached the thread again only to delete it.

 impl Drop for ImageWriterImage {
     fn drop(&mut self) {
         if let Ok(mut env) = attach_current_thread() {   // attach
             Image.close(self.image)
-        }                                                 // detach
-    }                                                     // drop self.image (GlobalRef) -> warning
+        }                                                 // pop local frame (thread stays attached)
+    }                                                     // drop self.image (GlobalRef) -> OK
 }

The same pattern exists in queue_input_image (the image argument is dropped after the locals), the Bgra32 input path, the audio input path and pull_encoded_data_with_codec.

attach_current_thread() now attaches the thread permanently; the jni crate detaches it when the thread terminates. Since detaching no longer releases the local references created on the thread, the function returns a LocalFrameEnv guard that pushes a local reference frame and pops it on drop. The guard dereferences to JNIEnv, so the call sites are unchanged.

-pub fn attach_current_thread() -> Result<AttachGuard<'static>> {
-    vm.attach_current_thread()
+pub fn attach_current_thread() -> Result<LocalFrameEnv> {
+    let env = vm.attach_current_thread_permanently()?;
+    env.push_local_frame(LOCAL_FRAME_CAPACITY)?;
+    Ok(LocalFrameEnv { env })   // pops the frame on drop
 }

Local references must not outlive the guard. This was already required, as they became invalid when the thread was detached.

Evidence

Measured on a real device (samsung SCG13, Android 16), counting the warning in logcat.

Scenario	Before	After
scripts/android-device-test.sh --release (Bgra32 input), warnings	1661	0
Same, Attached thread debug logs from the jni crate	7484	20
Unity 6000.3 sample scene (Vulkan, HardwareBuffer input), about 45 s of buffering, warnings	4946	0

The device test passes before and after the change. In the Unity build, MediaCodec ran normally after the change and no JNI error or crash was logged.