Actually, batching grass together to clusters of maybe 5k to 10k tris is a pretty good idea. The only "issue" is that it needs to be aligned to the ground. So you will need many different meshes to cover everything. But for terrain if you´ve got a heightmap you could get away with just one mesh by doing the alignment in the vertex shader of the grass sampling the heightmap.
Also in case you did not do it yet deactivate collision mesh generation while creating passable entities by setting the global variable collision_mode to 0 and back to 1 when you are done.

Edit: You may want to consider even bigger grass meshes because the few additional polygons you save from more efficient culling with smaller ones are most probably cheaper than the overhead of more entities.
Or using instanced rendering might be quite simple to implement in Lite-C, the only harder part will be to provide the transformations to the shader, as you would then probably just create a custom grass renderer.