Set normalized device coordinates Y direction to be up, which is standard in Unity, and properly detect non-zero motion vector inputs. This *seems* to be the correct recipe.

Packages/com.unity.postprocessing@3.2.2/PostProcessing/Runtime/Effects/Upscaling/SGSR2/Shaders/2_pass_cs/sgsr2_convert.compute

@@ -68,7 +68,7 @@ void CS(uint3 gl_GlobalInvocationID : SV_DispatchThreadID)
     float2 EncodedVelocity = LOAD_TEXTURE2D_X(InputVelocity, InputPos);
 
     float2 motion;
-    if (EncodedVelocity.x > 0.0)
+    if (any(abs(EncodedVelocity) > 0.0))
     {
         motion = decodeVelocityFromTexture(EncodedVelocity.xy);
     }

Packages/com.unity.postprocessing@3.2.2/PostProcessing/Runtime/Effects/Upscaling/SGSR2/Shaders/3_pass_cs/sgsr2_convert.compute

@@ -58,7 +58,7 @@ void CS(uint3 gl_GlobalInvocationID : SV_DispatchThreadID)
     float2 EncodedVelocity = LOAD_TEXTURE2D_X(InputVelocity, InputPos);
 
     float2 motion;
-    if (EncodedVelocity.x > 0.0)
+    if (any(abs(EncodedVelocity) > 0.0))
     {
         motion = decodeVelocityFromTexture(EncodedVelocity.xy);
     }

Packages/com.unity.postprocessing@3.2.2/PostProcessing/Runtime/Effects/Upscaling/SGSR2/Shaders/sgsr2_common.hlsl

@@ -1,5 +1,7 @@
 #define EPSILON 1.192e-07f
 
+#define REQUEST_NDC_Y_UP
+
 #ifdef UNITY_REVERSED_Z
 #define DEPTH_NEAREST(a, b)     max((a), (b))
 #define DEPTH_CLIP(depth)       ((depth) > 1.0e-05f)