Alright, tentatively resolving this one again.

For some reason (not yet discovered) calling the Unity internal function 'WorldNormalVector' wrappped in a defined(UNITY_INSTANCING_ENABLED) doesn't yield a shader parsing that adds surfNeedsTangentSpace (which would populate the INTERNAL_DATA internalSurfaceTtoW0). The results is that 'WorldNormalVector' returns zero (which later causes a nan in:
i.viewDir = normalize(mul( t2w, (_WorldSpaceCameraPos - i.worldPos)));) since all three tangent vectors are're defined as '0' by Unity.

There is a workaround which is aligned with what MicroSplats does already: that is to explicitly use i.worldNormal _somwhere_ (which then causes the INTERNAL_DATA to be available)

For example:
l.Albedo *= saturate(normalize(i.viewDir + i.worldPos + i.worldNormal) + 9999);

One thing I found out during this investigation is that i.viewDir is calculated by Unity as:

half3 _unity_tbn_0 = IN.tSpace0.xyz;
half3 _unity_tbn_1 = IN.tSpace1.xyz;
half3 _unity_tbn_2 = IN.tSpace2.xyz;

float3 worldPos = float3 ( IN.tSpace0.w , IN.tSpace1.w , IN.tSpace2.w ) ;
float3 worldViewDir = normalize ( UnityWorldSpaceViewDir ( worldPos ) ) ;
float3 viewDir = _unity_tbn_0 * worldViewDir.x + _unity_tbn_1 * worldViewDir.y + _unity_tbn_2 * worldViewDir.z ;
surfIN.viewDir = viewDir;

Where tSpace* are populated from the VS as:
o . tSpace0 = float4 ( worldTangent . x , worldBinormal . x , worldNormal . x , worldPos . x ) ;
o . tSpace1 = float4 ( worldTangent . y , worldBinormal . y , worldNormal . y , worldPos . y ) ;
o . tSpace2 = float4 ( worldTangent . z , worldBinormal . z , worldNormal . z , worldPos . z ) ;

I noticed that the MicroSplat/Example shader's path for UNITY_INSTANCING_ENABLED sets the normal as float3(0,1,0) which is then used in pixel shader _before_ the custom surface function is called to update the normal data. One thing I tried was to populate to read and feed the normal data from the vertex shader and that seemed to give a valid i.viewDir value in MicroSplatLayer SurfImpl (so theoretically no need to recalculate it). It could be something to investigate further (I noticed some weird shadows when I tried it so would need to investigate why that is).

Hope this unblocks Jason, I will follow up with the core team to see if there would be a more elegant way to force the WorldNormalVector function to be valid without having to tickle i.worldNormal

Cheers