The documentation says that, for directional lights, the calculated resolution is clamped to 2K and then scaled in function of the Quality settings. The problem is that those operations are not done in this order, the scale is done before the clamping.

Code from Runtime/Camera/Shadows.cpp, in function CalculateShadowMapSize:

{
int viewSize = int(std::max(cullData.viewWidth, cullData.viewHeight));
mapSize = NextPowerOfTwo(int(viewSize * kMultDir));
mapSize >>= qualityShift;
}
const int maxShadowSize = GetMaxShadowMapSize(kLightDirectional, customResolution, qualityShift);
mapSize = clamp<int>(mapSize, 32, maxShadowSize);
On iPhone Xs, for example, the resolution is high and the mapSize is 16384. The qualityShift is 3 when shadow resolution is set to low, so the mapSize end up being 2048 and the maxShadowSize is also 2048. The clamp does nothing and the size is always 2K!

Also note that even if the clamping was done before the shift, the end result would be wrong because the shift is 3 and that means a division by 8. The documentation says that resolution on low is a quarter of maxShadowSize so the shift should be 2.