diff --git a/src/assemblers/halo_exchange.jl b/src/assemblers/halo_exchange.jl
new file mode 100644
index 0000000..d487e2c
--- /dev/null
+++ b/src/assemblers/halo_exchange.jl
@@ -0,0 +1,204 @@
+# This file is a part of JuliaFEM.
+# License is MIT: see https://github.com/JuliaFEM/JuliaFEM.jl/blob/master/LICENSE.md
+
+# Mesh-partition adjacency and halo-exchange plumbing.
+#
+# Despite the filename, this file holds the partition / halo-exchange
+# plumbing in full: [`PartitionAdjacency`](@ref) +
+# [`build_partition_adjacency`](@ref) describe which partition pairs
+# share at least one mesh node, and [`RankHaloExchange`](@ref) carries
+# the precomputed sparse send/recv tables that the matrix-free `apply_K!`
+# / `apply_M!` paths use to exchange halo DOF data each application.
+
+"""
+    PartitionAdjacency(max_part_id, neighbors)
+
+`neighbors[p]::Vector{Int}` lists partition ids sharing at least one mesh node
+with partition `p` (undirected graph). Indexed by `p = 1:max_part_id`.
+
+Built by [`build_partition_adjacency`](@ref). Setup allocates; hot paths use
+precomputed [`RankHaloExchange`](@ref) data instead.
+"""
+struct PartitionAdjacency
+    max_part_id::Int
+    neighbors::Vector{Vector{Int}}
+end
+
+"""
+    build_partition_adjacency(layout::MeshPartitionLayout, mesh::Mesh) -> PartitionAdjacency
+
+Scan element-node incidence via `mesh.inverse_connectivity`: two partitions
+are adjacent if some node carries incident elements from both.
+"""
+function build_partition_adjacency(
+    layout::MeshPartitionLayout,
+    mesh::Mesh{N,T},
+) where {N,T}
+    validate_partition(layout, length(mesh.connectivity))
+    np = maximum(layout.element_part_id)
+    sets = [Set{Int}() for _ in 1:np]
+    @inbounds for e in eachindex(mesh.connectivity)
+        pe = layout.element_part_id[e]
+        conn = mesh.connectivity[e]
+        for k in 1:N
+            n = Int(conn[k])
+            for (e2, _) in mesh.inverse_connectivity[n]
+                p2 = layout.element_part_id[Int(e2)]
+                if p2 != pe
+                    push!(sets[pe], p2)
+                    push!(sets[p2], pe)
+                end
+            end
+        end
+    end
+    neighbors = [sort!(collect(sets[i])) for i in 1:np]
+    return PartitionAdjacency(np, neighbors)
+end
+
+"""
+    RankHaloExchange
+
+Per-part halo metadata for a replicated global vector (MPI precursor).
+
+# Fields
+- `part::Int`
+- `recv_neighbor`, `send_neighbor` — aligned neighbor part ids
+- `recv_dof`, `send_dof` — `recv_dof[k]` global DOFs this part receives from
+  `recv_neighbor[k]`; `send_dof[k]` global DOFs this part sends to
+  `send_neighbor[k]`
+
+Invariant (serial replica): `sort(send_dof[j]) == sort(recv_dof[q][…])` where
+`exchanges[q]` receives from `part` at the slot matching `send_neighbor[j]`.
+Checked in tests.
+"""
+struct RankHaloExchange
+    part::Int
+    recv_neighbor::Vector{Int}
+    recv_dof::Vector{Vector{Int}}
+    send_neighbor::Vector{Int}
+    send_dof::Vector{Vector{Int}}
+end
+
+"""
+    matvec_halo_mpi_request_count(exchange::RankHaloExchange) -> Int
+
+Number of concurrent MPI operations in one [`exchange_matvec_halos_mpi!`](@ref) round:
+[`length`](@ref)`(exchange.recv_neighbor)` receives plus [`length`](@ref)`(exchange.send_neighbor)`
+sends. Size persistent [`MPI.Request`](@ref) buffers with
+[`allocate_exchange_matvec_halo_mpi_requests`](@ref) (after `using MPI`).
+"""
+function matvec_halo_mpi_request_count(exchange::RankHaloExchange)::Int
+    return length(exchange.recv_neighbor) + length(exchange.send_neighbor)
+end
+
+@inline function _mask_intersect!(out::BitVector, a::BitVector, b::BitVector)
+    @inbounds for i in eachindex(out)
+        out[i] = a[i] & b[i]
+    end
+    return out
+end
+
+"""
+    build_rank_halo_exchanges(handler, layout, mesh, node_owner, elements)
+        -> Vector{RankHaloExchange}
+
+Setup-only routine (allocates masks and DOF lists). Uses
+[`node_partition_owner_min!`](@ref)-style ownership via `node_owner`, Vertex
+field layout from `handler`, and [`mark_referenced_dofs!`](@ref) /
+[`mark_owned_vertex_field_dofs!`](@ref) / [`ghost_dof_mask!`](@ref).
+
+`length(exchanges) == maximum(layout.element_part_id)`; `exchanges[p]` is the
+spec for partition id `p`.
+
+For matrix-free [`apply_K_owned_rows!`](@ref) with stencil-packed layouts, see
+[`build_matvec_halo_exchanges`](@ref) in `packed_layout.jl` (included after this file).
+"""
+function build_rank_halo_exchanges(
+    handler::DOFHandler,
+    layout::MeshPartitionLayout,
+    mesh::Mesh{N,T},
+    node_owner::Vector{Int},
+    elements::AbstractVector{El},
+) where {N,T,El<:Element}
+    validate_partition(layout, length(mesh.connectivity))
+    ndofs = handler.total_dofs
+    np = maximum(layout.element_part_id)
+    length(node_owner) == length(mesh.nodes) ||
+        throw(DimensionMismatch("node_owner length $(length(node_owner)), nnodes $(length(mesh.nodes))"))
+
+    adj = build_partition_adjacency(layout, mesh)
+
+    elem_by_part = [element_indices_for_part(layout, p) for p in 1:np]
+
+    owned = Vector{BitVector}(undef, np)
+    ghost = Vector{BitVector}(undef, np)
+    ref_tmp = falses(ndofs)
+    @inbounds for p in 1:np
+        owned[p] = falses(ndofs)
+        mark_owned_vertex_field_dofs!(owned[p], handler, node_owner, p)
+        fill!(ref_tmp, false)
+        mark_referenced_dofs!(ref_tmp, elements, elem_by_part[p], ndofs)
+        ghost[p] = falses(ndofs)
+        ghost_dof_mask!(ghost[p], ref_tmp, owned[p])
+    end
+
+    tmp = falses(ndofs)
+    dof_buf = Vector{Int}(undef, ndofs)
+
+    exchanges = Vector{RankHaloExchange}(undef, np)
+    @inbounds for p in 1:np
+        recv_n = Int[]
+        recv_d = Vector{Int}[]
+        send_n = Int[]
+        send_d = Vector{Int}[]
+        for q in adj.neighbors[p]
+            _mask_intersect!(tmp, ghost[p], owned[q])
+            nloc = collect_true_indices!(dof_buf, tmp)
+            push!(recv_n, q)
+            push!(recv_d, collect(Int, view(dof_buf, 1:nloc)))
+        end
+        for q in adj.neighbors[p]
+            _mask_intersect!(tmp, ghost[q], owned[p])
+            nloc = collect_true_indices!(dof_buf, tmp)
+            push!(send_n, q)
+            push!(send_d, collect(Int, view(dof_buf, 1:nloc)))
+        end
+        exchanges[p] = RankHaloExchange(p, recv_n, recv_d, send_n, send_d)
+    end
+
+    return exchanges
+end
+
+"""
+    referenced_dof_mask_for_part!(
+        mask::BitVector, elements, element_ids::AbstractVector{Int}, n_total_dofs::Int) -> mask
+
+[`mark_referenced_dofs!`](@ref) alias for a fixed element-id list (reuse the
+same buffer across Krylov iterations — allocation-free after setup).
+
+    referenced_dof_mask_for_part!(
+        mask::BitVector, elements, layout::MeshPartitionLayout, part::Int,
+        n_total_dofs::Int) -> mask
+
+Convenience overload that calls [`element_indices_for_part`](@ref) (allocates a
+new index vector each call — fine for setup, not for inner Krylov loops).
+"""
+function referenced_dof_mask_for_part!(
+    mask::BitVector,
+    elements::AbstractVector{El},
+    element_ids::AbstractVector{Int},
+    n_total_dofs::Int,
+) where {El <: Element}
+    return mark_referenced_dofs!(mask, elements, element_ids, n_total_dofs)
+end
+
+function referenced_dof_mask_for_part!(
+    mask::BitVector,
+    elements::AbstractVector{El},
+    layout::MeshPartitionLayout,
+    part::Int,
+    n_total_dofs::Int,
+) where {El <: Element}
+    ep = element_indices_for_part(layout, part)
+    return mark_referenced_dofs!(mask, elements, ep, n_total_dofs)
+end