Further tweaked motion vector processing, now the camera motion vectors from Unity produce an exact match with the auto-generated motion from the clip-delta matrix, in both D3D, Vulkan and OpenGL. It also makes sense with regards to the flipped Y-coordinate in screen space now. The final result looks good now too.

Packages/com.unity.postprocessing@3.2.2/PostProcessing/Runtime/Effects/Upscaling/SGSR2/Shaders/sgsr2_birp.hlsl

@@ -21,10 +21,9 @@ SamplerState s_LinearClamp  : register(s1);
 
 inline float2 decodeVelocityFromTexture(float2 ev)
 {
-    // Nothing much to do, motion vectors are not encoded
 #if UNITY_UV_STARTS_AT_TOP
-    return -ev;
+    return float2(ev.x, -ev.y) * 2.0f;
 #else
-    return ev;
+    return ev * 2.0f;
 #endif
 }

Packages/com.unity.postprocessing@3.2.2/PostProcessing/Runtime/Effects/Upscaling/SGSR2/Shaders/sgsr2_hdrp.hlsl

@@ -9,10 +9,9 @@
 
 inline float2 decodeVelocityFromTexture(float2 ev)
 {
-    // Nothing much to do, motion vectors are not encoded
 #if UNITY_UV_STARTS_AT_TOP
-    return -ev;
+    return float2(ev.x, -ev.y) * 2.0f;
 #else
-    return ev;
+    return ev * 2.0f;
 #endif
 }