This icon will need to be created.

It's not obvious to me why this tool can't work similar to the grab brush, which uses sculpt_orig_vert_data_init rather than allocating its own array for original coordinates and normals.

Is there a specific reason for that?

You can use something like BLI_hash_int_01(vd.index ^ random_seed) rather than a table with a fixed amount of random numbers.

Just using modulo can lead to visible patterns depending on the mesh topology.

Better to remove this here and use bool use_pbvh_draw = BKE_sculptsession_use_pbvh_draw(ob, v3d); below.

In general initializing immediately avoids use of uninitialized variables.

These 4 lines could be wrapped into a sculpt_vertex_random_access_init function as part of the abstraction API.

Is BM_mesh_elem_table_ensure needed?

This shouldn't happen as far as I know, was this to fix an actual issue?

Why does this depend on the usage of PBVH drawing? Does the PBVH get destroyed and rebuilt in that case?

These are not obviously related, so that makes this test prone to break in the future or may not cover all cases already.

Same comment as above.

true -> false

There is no need to create a mask for this tool, it only needs to use it if it's already there.

Instead of 3 separate booleans, define this as an array like e.g. sculpt symmetry mirror axes.

This layout looks bad in the tool settings tab in the properties editor. I think you can just do:

props = tool.operator_properties("sculpt.mesh_filter")
layout.prop(props, "type", expand=False)
layout.prop(props, "strength")
layout.prop(props, "deform_axis")