At some points compiled metal code doesn't have matrix typecast function. For example, float3x3 -> half3x3.

Project Steps:

1. Open project "MetalCompile" (Attached by LukasC);
2. Open scene "Test".
3. Build & Run.
4. When xCode project starts, you should get some metal compilation errors.

Steps without building project:

1. Select shader "NewShader".
2. Click compile (Only for metal).
3. Goto line 2116, you should see smth like:

...
float3x3 tmpvar_29;
half3x3 tmpvar_30;
tmpvar_30[0].x = tmpvar_27.x;
tmpvar_30[0].y = tmpvar_28.x;
tmpvar_30[0].z = tmpvar_25.x;
tmpvar_30[1].x = tmpvar_27.y;
tmpvar_30[1].y = tmpvar_28.y;
tmpvar_30[1].z = tmpvar_25.y;
tmpvar_30[2].x = tmpvar_27.z;
tmpvar_30[2].y = tmpvar_28.z;
tmpvar_30[2].z = tmpvar_25.z;
tmpvar_29 = tmpvar_30;
...

4. You should notice that "tmpvar_29 = tmpvar_30" that this matrices are in different types.
In metal this type conversion requires cast, however opengles20 doesn't require it.
As I know unity generates metal shaders from openglES20 shaders, so this could explain why this bug happens.
But the strange thing is that, if you would use miss-matched matrix types normally in shader:

float3x3 mat1;
half3x3 mat2;
mat2 = mat1;

In metal it will be compiled correctly.

float3x3 mat1;
half3x3 mat2;
mat2 = _xlcast_half4x4(mat1);

So it means this bug is more complicated and caused by multiple factors.