<rdar://problem/65914326>

I can reproduce. I'm a bit surprised we get this wrong.

This happens because when we see a negative z-index child, we make its stacking context layer compositing eagerly:

    for (auto* childLayer : layer.negativeZOrderLayers()) {
        computeCompositingRequirements(&layer, *childLayer, overlapMap, currentState, backingSharingState, anyDescendantHas3DTransform);

        // If we have to make a layer for this child, make one now so we can have a contents layer
        // (since we need to ensure that the -ve z-order child renders underneath our contents).
        if (!willBeComposited && currentState.subtreeIsCompositing) {
            layer.setIndirectCompositingReason(IndirectCompositingReason::BackgroundLayer);
            layerWillComposite();
        }
    }

but this messes up the overlap map, so the second negative z-index child doesn't see the rects from the first one.

Fixing https://bugs.webkit.org/show_bug.cgi?id=217154 also seems to fix this.

This seems fixed these days.