Fix the ViroARCamera crash on Android and multi-skin GLB animation

CHANGELOG.md

@@ -98,6 +98,8 @@
 - **Android: a failure while a recording is being stopped no longer leaves `getRecordingStatus()` at IOError (`ARSessionRecorder`).** `encoder.stop()` and `muxer.stop()` failures reported it to the scene after the stop had already set the status back to None, so the status read IOError with nothing recording. Such a failure is now reported in the stop's result only. `ARScene.stopRecording()` may be called off the UI thread.
 - **Android hit tests always deliver their results (`VRORendererARCore_JNI`, `VRORenderer_JNI`).** The AR point, ray and position hit tests, the Quest OpenXR hit tests and the scene hit test kept the caller's one-shot listener only as a weak global reference, so the garbage collector could take it before the renderer thread answered, and the results never arrived. It is now a global reference, deleted once the callback has run.
 - **iOS downloads from a signed URL are saved (`VROPlatformDownloadURLToFile`, `VROPlatformDownloadURLToFileAsync`).** The temporary file was named after the URL's last path component with its query, and a signed URL's token pushed the name past the 255-byte file name limit, so the write failed and the download still reported success with a file that did not exist. `VROPlatformLastPathComponent` drops the query and fragment, and a failed write now fails the download.
+- **Android: a camera added before the renderer no longer crashes (`VROPlatformUtil`, #597).** `libviro_renderer.so` now sets the `JavaVM` in `JNI_OnLoad`. It was set only when a renderer or loader started, and under Fabric `ViroARCamera`'s view is built first, so adding its node reached `getJNIEnv` with no VM.
+- **A GLB whose skins share joints animates every skin (`VROGLTFLoader`, #599).** Resampling runs once per skin and rewrote a shared joint's channels in place; the next skin's pass kept only the translation of the merged channel, so those joints lost their rotation and scale. Every property of a channel is now copied.
 
 ### Removed
 

RELEASE_NOTES.md

@@ -36,6 +36,8 @@
 - **Co-location:** `VROColocationSession::deviceId()`, sent as `x-rv-device`, so the relay bills a device once.
 - **A click goes to the node drawn on top,** and a click in open space reaches the controller's own delegates again.
 - **Text sits where its vertical alignment puts it,** and a back-culled draw after a front-culled one culls back faces again.
+- **GLB characters with one skin per mesh animate every part,** not only the first skin's.
+- **Android: a `ViroARCamera` no longer crashes the scene as it mounts.**
 - **Video:** a paused iOS video stays paused and its late callbacks no longer reach a freed texture; Android play and pause keep their order.
 - **iOS:** ViroKit targets iOS 15.1, and builds for the iOS Simulator (arm64 and x86_64) as well as the device. The copy into viro now makes `ViroKit.xcframework`, replacing only the slice just built. `ios/build_ios_simulator_libs.sh` adds the simulator builds of LinearMath, protobuf-lite and GVR Audio that ViroKit links.
 - **Removed:** the macOS target, which no longer linked and nothing built, and dead code (an unused metrics client, the API key validator, the OpenCV headers, an old OpenXR loader).

ViroRenderer/VROGLTFLoader.cpp

@@ -1516,19 +1516,25 @@ void VROGLTFLoader::resampleSkeletalChannelsToCommonGrid(std::map<int, std::pair
                     const VROKeyframeAnimationFrame *fB = frames[kB].get();
 
                     // STEP samplers hold the left key; LINEAR (and pre-baked CUBICSPLINE) interpolate.
+                    // Every property the channel has is copied, not only the first: a raw glTF
+                    // channel carries one, but a joint shared by several skins comes back here
+                    // already merged by an earlier skin's pass, with all of them.
                     if (chan->_hasTranslation) {
                         hasTrans = true;
                         mergedFrames[gi]->translation = step ? fA->translation
                                                              : fA->translation.interpolate(fB->translation, alpha);
-                    } else if (chan->_hasScale) {
+                    }
+                    if (chan->_hasScale) {
                         hasScale = true;
                         mergedFrames[gi]->scale = step ? fA->scale
                                                        : fA->scale.interpolate(fB->scale, alpha);
-                    } else if (chan->_hasRotation) {
+                    }
+                    if (chan->_hasRotation) {
                         hasRotation = true;
                         mergedFrames[gi]->rotation = step ? fA->rotation
                                                           : VROQuaternion::slerp(fA->rotation, fB->rotation, alpha);
-                    } else if (chan->_hasMorphWeights) {
+                    }
+                    if (chan->_hasMorphWeights) {
                         hasMorph = true;
                         std::map<std::string, float> weights;
                         for (const auto &kv : fA->morphWeights) {

ViroRenderer/VROPlatformUtil.cpp

@@ -553,6 +553,15 @@ static jobject sJavaAssetMgr = nullptr;
 static jobject sPlatformUtil = nullptr;
 static AAssetManager *sAssetMgr = nullptr;
 
+// The JavaVM is known from the moment System.loadLibrary("viro_renderer") loads this library, so
+// a JNI call that reaches getJNIEnv() before any renderer or loader has called VROPlatformSetEnv
+// still finds it. Under Fabric a view such as VRTCamera is built before the scene's renderer, and
+// adding its node already calls back into Java.
+extern "C" JNIEXPORT jint JNI_OnLoad(JavaVM *vm, void *reserved) {
+    sVM = vm;
+    return JNI_VERSION_1_6;
+}
+
 // Map (and mutex) holding native tasks waiting to be dispatched. On Android the threading
 // functionality is handled on the Java layer, so we need some mechanism for mapping IDs
 // to corresponding tasks. Both the ID generator and the map itself are protected by the