This is the main one:
df4e67fe75/ReactAndroid/src/main/java/com/facebook/react/ReactInstanceManager.java (L512)

Why this happens is a mystery wrapped in an enigma.

This commit refactors device selection (more heavily on iOS) to make it
consistent across platforms.

Due to its complexity I couldn't break out each step into separate commits,
apologies to the reviewer.

Changes made to device handling:

- speaker is always the default, regardless of the mode
- "Phone" shows as a selectable option, even in video call mode
- "Phone" is not displayed when wired headphones are present
- Shared device picker between iOS and Android
- Runtime device updates while the picker is open

* Adds live streaming sound notification.

* Adds ios resources for the new files.

Fixes: https://github.com/jitsi/jitsi-meet-sdk-samples/issues/11

Fixes running profile builds.

Set our own audio device manager so we can tweak it if need be (enabling /
disabling the HW AEC on specific devices).

Switch to using the software video encoder / decoder. This may feel like a
downgrade, but it has advantages:

- simulcast is now working (on par with iOS)
- certain devices have broken VP8 HW encoders (I'm looking at you Samsung Galaxy
S7) so this fixes that

It tends to close, so always open the reverse tunnel.

After calling startService we are supposed to have a bit of time before turning
the service into a foreground service, but certain devices seem to be more
spartan and we've seen the following failure:

Caused by java.lang.IllegalStateException: Not allowed to start service Intent { act=JitsiMeetOngoingConferenceService:START cmp=org.jitsi.meet/.sdk.JitsiMeetOngoingConferenceService }: app is in background uid UidRecord{f6778d5 u0a220 CAC  bg:+1m1s417ms idle change:idle procs:1 proclist:15604, seq(0,0,0)}
       at android.app.ContextImpl.startServiceCommon + 1600(ContextImpl.java:1600)
       at android.app.ContextImpl.startService + 1546(ContextImpl.java:1546)
       at android.content.ContextWrapper.startService + 669(ContextWrapper.java:669)
       at org.jitsi.meet.sdk.JitsiMeetOngoingConferenceService.launch + 50(JitsiMeetOngoingConferenceService.java:50)

Be expliocit and call startForegroundService, on supported platforms.

I cannot see a path leading to it being null, but Crashlytics demonstrated it's
possible (so far in a small subset of old devices). Be defensive then.

The app is about to crash at that stage so it was a moot point to try to leave
the conference anyway.

Stopping ConnectionServers is still a good idea though, since a crash may leave
the device in a bad state otherwise.

Fixes this issue:

~~~
    java.util.ConcurrentModificationException
        at java.util.HashMap$HashIterator.nextNode(HashMap.java:1441)
        at java.util.HashMap$KeyIterator.next(HashMap.java:1465)
        at org.jitsi.meet.sdk.OngoingConferenceTracker.updateListeners(OngoingConferenceTracker.java:89)
        at org.jitsi.meet.sdk.OngoingConferenceTracker.onExternalAPIEvent(OngoingConferenceTracker.java:74)
        at org.jitsi.meet.sdk.ExternalAPIModule.sendEvent(ExternalAPIModule.java:71)
        at java.lang.reflect.Method.invoke(Native Method)
        at com.facebook.react.bridge.JavaMethodWrapper.invoke(JavaMethodWrapper.java:372)
        at com.facebook.react.bridge.JavaModuleWrapper.invoke(JavaModuleWrapper.java:158)
        at com.facebook.react.bridge.queue.NativeRunnable.run(Native Method)
        at android.os.Handler.handleCallback(Handler.java:873)
        at android.os.Handler.dispatchMessage(Handler.java:99)
        at com.facebook.react.bridge.queue.MessageQueueThreadHandler.dispatchMessage(MessageQueueThreadHandler.java:29)
        at android.os.Looper.loop(Looper.java:214)
        at com.facebook.react.bridge.queue.MessageQueueThreadImpl$4.run(MessageQueueThreadImpl.java:232)
        at java.lang.Thread.run(Thread.java:764)
~~~

The notification is posted by a foreground service, which also has the nice
side-effect of keeping the app alive for a long time.

This helper method gets the current Activity attached to React Native (via the
ReactContext). This is useful for modules which need access to it, without being
actual React Native modules.

They are only meant to be used internally. This reduces the risk of SDK users
sending garbage when we expect a specific type.

If the user swipes the activity from the recents list there is no other chance,
onDestroy gets called directly.

ENTER_PIP_MODE, for example, does not have it.

Its main task is to cleanup conferences (specially the connection services
stuff) to make sure the system is left in a working state even when the
unexpected happens.

It may be called in the Activity instead of in the Fragment. Handle both.

At the moment it includes:

- display name
- email
- avatar URL

This information is used *only* if no token was specified.

The welcomePageEnabled and pictureInPictureEnabled props on mobile have been
converted to feature flags.

People run these in their own checkout and will run into problems because
tagging will fail.

The naming didn't match, so adjust it. @ cannot be used for maven artifact
names.

It was extracted from the RN core to a community maintained package.

Entering PiP mode while the permissions dialog is display will not only
fail, but also mess up the Activity lifecycle on some OS versions.
We may end up with two activity/fragment instances and a situation where
the onStop callback was not called yet on the instance #1 while
the onResume has been already called on instance #2.

A libre build will exclude the following:

- Analytics modules
- Google Play services GMS
- Crashlytics
- Firebase