Here’s what I’ve debugged on Windows desktop – in the ‘happy path’ (i.e. one in which we get a Event.Latency.ScrollBegin.Wheel.TimeToScrollUpdateSwapBegin4) we see this sequence:
1.	The first GSU is handled synchronously instead of getting queued in the compositor_event_queue_. 
2.	Handling the first GSU creates the scroll-offset animation (LayerTreeHostImpl::ScrollAnimationCreate)
3.	ScrollAnimationCreate calls LayerTreeHostImpl::SetNeedsOneBeginImplFrame
4.	Because this happens within the scope of InputHandlerProxy::DispatchSingleInputEvent, there’s a SwapPromiseMonitor on the LayerTreeHostImpl, which is notified of the need for redraw
5.	This will call LayerTreeImpl::QueuePinnedSwapPromise, which adds the swap promise to a list that is ‘finished’ when a compositor frame is generated
6.	LatencyInfoSwapPromise::WillSwap will add a LatencyInfo to the compositor frame metadata, which is shipped off to the display compositor, where eventually the Swap will happen and the LatencyInfo will be completed 

For subsequent GSUs generated by mouse wheel we instead see the following:
1.	Subsequent GSU is queued, and this code detects that we need to animate for the input (SetNeedsAnimateInput calls SetNeedsOneBeginImplFrame)
bool needs_animate_input = compositor_event_queue_->empty();
    compositor_event_queue_->Queue(std::move(event_with_callback),
                                   tick_clock_->NowTicks());
    if (needs_animate_input)
      input_handler_->SetNeedsAnimateInput();
    return;
2.	There is no SwapPromiseMonitor registered on LayerTreeHostImpl (not called within scope of InputHandlerProxy::DispatchSingleInputEvent)
3.	Later when the GSU is handled on BeginFrame for impl thread, the animation target is updated (LayerTreeHostImpl::ScrollAnimationUpdateTarget), but nothing calls to LayerTreeHostImpl::SetNeedsOneBeginImplFrame or SetNeedsRedraw

For this second sequence, the latency info’s associated with the GSU’s are not shipped off with the compositor frame, so we don’t get the data point. 


Note I see similar issues with main thread driven scroll animations, but I suspect the root cause is different. I'm not as familiar with those, so needs more investigation on my end