fix(🤖): release the Java Surface wrapper on the non-opaque path

[kushal-matiks]

Fixes #4045

Problem

On Android, every non-opaque <Canvas> leaks one android.view.Surface.

surfaceAvailable builds a Surface to wrap the SurfaceTexture, hands it to ANativeWindow_fromSurface, then drops it with DeleteLocalRef. That only releases the JNI local reference — Surface.release() is never called, so the underlying buffer producer is left to the finalizer and CloseGuard logs:

W System: A resource failed to call Surface.release.

once per Canvas.

surfaceDestroyed() releases _jSurfaceTexture and OpenGLWindowContext releases the ANativeWindow, but nothing owns the intermediate Surface.

Fix

Release the Java wrapper once ANativeWindow_fromSurface has taken its own reference on the producer. The lookup is null-guarded and the call is exception-guarded, so it degrades to the current behaviour rather than throwing if anything is unexpected.

This mirrors the ownership rule the WebGPU path already documents in JniWebGPUView.cpp:

// ANativeWindow_fromSurface acquires a reference; SurfaceInfo releases it
// (via the releaser below) once it is done with the window.

Only the !opaque branch is touched; the opaque branch never creates the wrapper.

Verification

Measured on an app with 6 <Canvas> components (a tab bar of Skia icons), counting Surface.release warnings per launch:

environment	before	after
Emulator API 37, debug	6	0
Samsung SM-M366B (Mali), API 36, debug	6	0
Samsung SM-M366B, minified release (R8)	6	0

The unpatched release build was measured first as an instrument check — CloseGuard is normally gated on the debuggable flag, so a patched 0 would be meaningless if the unpatched release were also silent. It reported 6, so the 0 is a real measurement.

No regressions across 20 navigations plus 3 background/foreground cycles on the release build:

• SIGSEGV / SIGABRT / Fatal signal / tombstone / ANR — 0
• BufferQueue has been abandoned — 0
• EGL_BAD_SURFACE, EGL_BAD_NATIVE_WINDOW — 0
• rendering unchanged; canvases still transparent, no visual diff

R8 does not affect the change: GetMethodID(surfaceClass, "release", "()V") resolves android.view.Surface.release from the platform framework at runtime, and minification never rewrites android.*.

Not addressed here

surfaceSizeChanged re-enters surfaceAvailable, which assigns _jSurfaceTexture = env->NewGlobalRef(...) without DeleteGlobalRef on the previous value — a separate JNI global-ref leak on the resize/rotation path. Left alone deliberately: it is a distinct defect, and I could not exercise that path to verify a fix (the app I reproduced on is portrait-locked). Noted in #4045.

Testing done

Verified by rebuilding the Android native library from source and measuring on device via logcat, in both debug and minified release builds, on a physical Mali device and an emulator. I did not run the repo's own example app or test suite locally — happy to do so, or to adjust the approach, if you'd prefer.

[kushal-matiks]

I have signed the CLA!