This is a part of performance regression investigation task.
Every movement of vertex will request modifier stack to be re-run. This involves conversion of EditMesh to Mesh, no matter whether there are modifiers or not.
Since rB84cf670d1d86 it seems that we can simply delay creation of mesh_cage in editbmesh_calc_modifiers and only do it if there is actual modifier which is to be applied. We can then have runtime.is_original set to truth for all Mesh IDs which are covered by Copy-on-Write, and set it to false when modifier runs.
This will solve about 50% of slowdown in the attached file (where EditMesh to Mesh takes almost same amount of time as batch cache creation).