fix: Clear native output callbacks in output hooks on unmount

[Pap36]

What

useFrameOutput, useDepthOutput and useObjectOutput hand JS callbacks to a native output (setOnFrameCallback, setOnFrameDroppedCallback, and their depth/object equivalents) and never clear them. Nitro keeps those callbacks alive as GC roots for as long as the native output exists, and the callbacks close over the hook's frameOutput and its per-mount worklet runtime. That is a cycle across the JSI boundary, so no garbage collection can ever break it: every mount of a component using useFrameOutput leaks a full worklet runtime plus whatever the onFrame worklet captured.

In a dev build with Worklets Bundle Mode each leaked runtime holds its own compiled copy of the app bundle. On an iPhone 11 our scanner screen cost ~350 MB per open and nothing came back on unmount, so the second open hit the Jetsam limit. Release builds leak the same objects, just far smaller (bytecode is mmapped and shared), so it shows up as slow growth instead of a crash.

How

Return effect cleanups from the three hooks that set the native callbacks back to undefined. That is the same call the Harness tests already make in their finally blocks. Clearing on unmount (and on callback change, before the new one is set) lets the frame output, the runtime and the captured scope be collected normally.

Verified on device (iPhone 11, iOS 26.6.1, dev client, physFootprint sampled every 3 s):

Step	Before	After cleanup
Home screen baseline	1509 MB	1509 MB
First open of the frame-processor screen	1622 MB	1622 MB
Back	1636 MB, nothing freed	1631 MB
Forced GC (DevTools)	no change	−324 MB
Second open	1931 MB → Jetsam kill	survives

A WeakRef probe on the hook's runtime and frameOutput confirmed the mechanism: both stayed alive across a forced full GC before the change, and the runtime was collected after it. The heap snapshot retainer path goes (GC roots) → (Custom) → onFrameDropped closure → hook environment → frameOutput / runtime.

useDepthOutput and useObjectOutput have the identical pattern and get the same cleanup.

Test

Adds a Harness test to visioncamera.hooks.harness.tsx that mounts a component calling useFrameOutput, attaches its output to a CameraSession created outside React (so the pipeline outlives the component), waits for Frames, unmounts, then captures two photos as the pipeline clock and asserts that no further Frames reached the unmounted onFrame.

Ran locally on a Pixel 7 (Android 17) with bun run test:harness:android -- --testPathPatterns=hooks:

• With the fix: all 7 tests in the file pass (1 pre-existing skip), the new one in ~1.2 s.

• With main's three hook files restored and the test kept: the new test fails at expect(onFrameReceived).toHaveBeenCalledTimes(framesAfterUnmount), because Frames keep arriving after unmount(). In one of the two control runs the app instead died on the frame thread with

  FATAL EXCEPTION: com.margelo.camera.frame
  java.lang.RuntimeException: Cannot call SyncJSCallback<bool (std::shared_ptr<HybridFrameSpec> const&)> - the underlying `jsi::Function` has already been deleted!
      at com.margelo.nitro.camera.hybrids.outputs.HybridFrameOutput.updateAnalyzer$lambda$4(HybridFrameOutput.kt:126)
  

  That is the Android face of the same bug: setOnFrameDroppedCallback is a no-op there, so nothing pins the runtime, the GC eventually frees the worklet runtime, and CameraX's analyzer still calls into the deleted jsi::Function on the next Frame. On iOS the drop callback pins everything and it leaks instead. Clearing the analyzer on unmount fixes both.

Related

• fix: Memoize targetResolution in output hooks margelo/react-native-vision-camera#4166 made this worse when targetResolution was an inline literal: a fresh output, and therefore a fresh leaked runtime, on every render rather than every mount.
• 🐛 Android: every <Camera> mount leaks its SurfaceView and camera output stream margelo/react-native-vision-camera#4180 / fix(android): Destroy the session lifecycle owner on reconfigure margelo/react-native-vision-camera#4186 are the Android-native counterparts of "per mount, nothing released"; this one is in the JS hooks and affects both platforms.

[Pap36]

Preview only; submitted upstream as margelo#4192.