diff --git a/src/domains/continuum/kernel.jl b/src/domains/continuum/kernel.jl
index 1b1ef61..ab2cb85 100644
--- a/src/domains/continuum/kernel.jl
+++ b/src/domains/continuum/kernel.jl
@@ -8,26 +8,25 @@ This module defines:
 1. The ContinuumKernel type
 2. The weak form: compute_stiffness_value (most atomic operation)
 3. Block builder: compute_stiffness_block (builds D×D blocks)
-4. NEW: Microkernel interface via evaluate() wrapper
 
-Everything else (geometry preprocessing, integration, assembly, DOF mapping) belongs elsewhere.
+The DOF-based / matrix-free assembler microkernel surface
+(`qpoint_buffer_eltype`, `update_qpoint_buffer!`, `evaluate_entry`,
+`evaluate_mass_entry`, `reference_fields`) is implemented further
+down in this file. Everything else (geometry preprocessing,
+integration, assembly, DOF mapping) belongs elsewhere.
 """
 
 using Tensors
 using Tensors: basevec  # For unit vector construction
 
-# Import for microkernel interface and material cache accessors
-if isdefined(Main, :JuliaFEM) && isdefined(Main.JuliaFEM, :evaluate)
-    import ..JuliaFEM: evaluate, get_tangent
-end
-
 """
     ContinuumKernel{Theory<:AbstractContinuumTheory, Mat<:AbstractMaterial} <: AbstractKernel
 
 Domain kernel for continuum mechanics (3D solid mechanics).
 
-Couples formulation theory, material model, and displacement field.
-Works with any assembler (COO, CSC, nodal).
+Couples formulation theory, material model, displacement field, and
+optional density (carried on the kernel rather than the material so
+existing material structs stay untouched).
 
 # Type Parameters
 - `Theory`: Continuum theory (FullThreeD, PlaneStress, PlaneStrain, Axisymmetric)
@@ -37,6 +36,10 @@ Works with any assembler (COO, CSC, nodal).
 - `formulation`: ContinuumFormulation{Theory}
 - `material`: Material model instance
 - `field`: Displacement{3}() field type
+- `density::Float64`: mass density `ρ` [kg/m³] used by the mass matrix
+  (`evaluate_mass_entry` / `apply_M!` / `assemble_M!`). Defaults to `0`,
+  in which case the kernel produces a structural-zero `M` and the
+  static-elasticity tests are unchanged.
 
 # Example
 
@@ -44,7 +47,8 @@ Works with any assembler (COO, CSC, nodal).
 kernel = ContinuumKernel(
     ContinuumFormulation{FullThreeD}(),
     LinearElastic(E=210e9, ν=0.3),
-    Displacement{3}()
+    Displacement{3}();
+    density = 7850.0,        # for mass matrix; omit for static-only
 )
 ```
 """
@@ -52,14 +56,26 @@ struct ContinuumKernel{Theory<:AbstractContinuumTheory,Mat<:AbstractMaterial} <:
     formulation::ContinuumFormulation{Theory}
     material::Mat
     field::Displacement{3}
+    density::Float64
 end
 
-# Convenience constructor without field (defaults to Displacement{3})
+# Outer constructor: positional 3-arg form, density defaults to 0.
 function ContinuumKernel(
     formulation::ContinuumFormulation{Theory},
-    material::Mat
+    material::Mat,
+    field::Displacement{3};
+    density::Float64 = 0.0,
 ) where {Theory<:AbstractContinuumTheory,Mat<:AbstractMaterial}
-    return ContinuumKernel(formulation, material, Displacement{3}())
+    return ContinuumKernel{Theory, Mat}(formulation, material, field, density)
+end
+
+# Convenience constructor without field (defaults to Displacement{3}).
+function ContinuumKernel(
+    formulation::ContinuumFormulation{Theory},
+    material::Mat;
+    density::Float64 = 0.0,
+) where {Theory<:AbstractContinuumTheory,Mat<:AbstractMaterial}
+    return ContinuumKernel{Theory, Mat}(formulation, material, Displacement{3}(), density)
 end
 
 # ============================================================================
@@ -89,9 +105,9 @@ get_field(kernel::ContinuumKernel) = kernel.field
         β::Int
     ) -> Float64
 
-Compute **single scalar** stiffness value K[k,l][α,β] at integration point.
+Compute single scalar stiffness value K[k,l][α,β] at integration point.
 
-This is the **most atomic kernel operation** - computes one DOF-pair contribution.
+This is the most atomic kernel operation - computes one DOF-pair contribution.
 
 # Theory
 
@@ -143,30 +159,70 @@ This is more atomic than 3×3 blocks. Caller can:
 2. Direct assembly: `K_global[dof_k_α, dof_l_β] += value * detJ * w`
 """
 @inline function compute_stiffness_value(
-    grad_k::Vec{D,Float64},
-    grad_l::Vec{D,Float64},
-    C::Tensor{4,D,Float64},
+    grad_k::Vec{D,F},
+    grad_l::Vec{D,F},
+    C::Tensor{4,D,F},
     α::Int,
     β::Int
-) where D
+) where {D,F<:AbstractFloat}
     # Build strain-displacement operators Bₖ,α and Bₗ,β
     # These are 2nd-order tensors (D×D matrices)
-    
-    # eα and eβ are unit vectors
-    e_α = basevec(Vec{D}, α)
-    e_β = basevec(Vec{D}, β)
-    
+
+    # eα and eβ are unit vectors of the same float type as the inputs,
+    # so the symmetrization below stays at precision F end-to-end.
+    e_α = basevec(Vec{D,F}, α)
+    e_β = basevec(Vec{D,F}, β)
+
+    half = F(0.5)
+
     # Bₖ,α = ½(∇Nₖ ⊗ eα + eα ⊗ ∇Nₖ)
-    B_k_α = 0.5 * (grad_k ⊗ e_α + e_α ⊗ grad_k)
-    
+    B_k_α = half * (grad_k ⊗ e_α + e_α ⊗ grad_k)
+
     # Bₗ,β = ½(∇Nₗ ⊗ eβ + eβ ⊗ ∇Nₗ)
-    B_l_β = 0.5 * (grad_l ⊗ e_β + e_β ⊗ grad_l)
-    
+    B_l_β = half * (grad_l ⊗ e_β + e_β ⊗ grad_l)
+
     # Compute weak form: K[α,β] = Bₖ,α : C : Bₗ,β
     # Double contraction: sum over all indices
     return dcontract(B_k_α, dcontract(C, B_l_β))
 end
 
+"""
+    compute_stiffness_value(
+        grad_k::Vec{D,F},
+        grad_l::Vec{D,F},
+        C::SymmetricTensor{4,D,F},
+        α::Int, β::Int,
+    ) -> F
+
+`SymmetricTensor` overload of the atomic stiffness microkernel — computes
+the same `B_k,α : C : B_l,β` value but without converting `C` from
+`SymmetricTensor` to the full `Tensor`. The full conversion goes
+through `Tensors.jl`'s general `Tensor{4,3}(::SymmetricTensor)`
+constructor, which carries an error-string branch that pulls in
+`Base.string` / `print_to_string` — call sites the Metal codegen can't
+prove dead and which trigger a `julia.new_gc_frame` IR error during
+device compilation.
+
+Building `B_k,α` and `B_l,β` as `SymmetricTensor`s via `symmetric(...)`
+keeps the entire chain inside the symmetric-tensor methods of
+`dcontract`, all of which are GPU-clean. Bit-identical to the
+`Tensor`-based variant on the CPU.
+"""
+@inline function compute_stiffness_value(
+    grad_k::Vec{D,F},
+    grad_l::Vec{D,F},
+    C::SymmetricTensor{4,D,F},
+    α::Int,
+    β::Int
+) where {D,F<:AbstractFloat}
+    e_α = basevec(Vec{D,F}, α)
+    e_β = basevec(Vec{D,F}, β)
+    half = F(0.5)
+    B_k_α = symmetric(half * (grad_k ⊗ e_α + e_α ⊗ grad_k))
+    B_l_β = symmetric(half * (grad_l ⊗ e_β + e_β ⊗ grad_l))
+    return dcontract(B_k_α, dcontract(C, B_l_β))
+end
+
 """
     compute_stiffness_block(
         grad_k::Vec{D},
@@ -217,8 +273,8 @@ end
 
 Computes 3×3 stiffness block from gradients and elasticity tensor.
 
-Maintains API compatibility with `src/assemblers/kernel_interface.jl` which
-uses SymmetricTensor. Internally uses dimension-generic `compute_stiffness_block()`.
+Accepts `C` as a `SymmetricTensor{4,3}`; internally uses the
+dimension-generic `compute_stiffness_block()`.
 """
 @inline function compute_block_at_point(
     grad_k::Vec{3,Float64},
@@ -231,51 +287,189 @@ uses SymmetricTensor. Internally uses dimension-generic `compute_stiffness_block
 end
 
 """
-    evaluate(
-        kernel::ContinuumKernel{Theory},
-        ::Displacement{3}, ::Displacement{3},
-        k::Int, l::Int, α::Int, β::Int,
-        material_cache, geometry_cache, q::Int
-    ) -> Float64
+    compute_block!(
+        K_blocks::Matrix{Tensor{2,3,Float64,9}},
+        ∇N_data::Matrix{Vec{3,Float64}},
+        detJ_w::Vector{Float64},
+        𝔻::Vector{SymmetricTensor{4,3,Float64,36}},
+        k_local::Int, l_local::Int,
+    ) -> Nothing
 
-Microkernel interface for continuum mechanics stiffness assembly.
+Integrate one continuum stiffness block `K[k_local, l_local]` over the
+quadrature points of an element using the precomputed material tangents
+`𝔻`. Writes directly into `K_blocks[k_local, l_local]`.
 
-Wraps `compute_stiffness_value()` to provide dispatch-based field coupling.
+Used by the element-based COO assembler in
+`src/assemblers/element_based/element_based_coo.jl`. The phase-separated
+design keeps material evaluations in `update_material_cache!` and lets
+this hot loop touch only `Vec{3}` / `SymmetricTensor{4,3}` arithmetic.
 
-# Implementation
-
-Extracts gradients and elasticity tensor from caches, then calls
-`compute_stiffness_value()` for the actual computation.
-
-# Example
-
-```julia
-# Direct computation
-grad_k = geometry_cache.∇N_data[q, k]
-grad_l = geometry_cache.∇N_data[q, l]
-    @inbounds C = get_tangent(material_workspace, q)
-value = compute_stiffness_value(grad_k, grad_l, C, α, β)
-
-# Microkernel interface
-value = evaluate(kernel, Displacement{3}(), Displacement{3}(),
-                k, l, α, β, material_workspace, geometry_cache, q)
-```
-
-# Performance
-
-No overhead - compiler inlines to identical code.
+Zero-allocation; the inner loop is `@inbounds`.
 """
-@inline function evaluate(
-    kernel::ContinuumKernel{Theory},
-    ::Displacement{3}, ::Displacement{3},
-    k::Int, l::Int, α::Int, β::Int,
-    material_cache, geometry_cache, q::Int
-) where {Theory<:AbstractContinuumTheory}
-    # Extract from caches using indices
-    grad_k = geometry_cache.∇N_data[q, k]
-    grad_l = geometry_cache.∇N_data[q, l]
-    @inbounds C = get_tangent(material_cache, q)
-    
-    # Call existing implementation (no duplication!)
-    return compute_stiffness_value(grad_k, grad_l, C, α, β)
+function compute_block!(
+    K_blocks::Matrix{Tensor{2,3,Float64,9}},
+    ∇N_data::Matrix{Vec{3,Float64}},
+    detJ_w::Vector{Float64},
+    𝔻::Vector{SymmetricTensor{4,3,Float64,36}},
+    k_local::Int,
+    l_local::Int,
+)
+    K_kl = zero(Tensor{2,3,Float64,9})
+
+    NIP = length(detJ_w)
+    @inbounds for q in 1:NIP
+        grad_k = ∇N_data[q, k_local]
+        grad_l = ∇N_data[q, l_local]
+        w = detJ_w[q]
+        D = 𝔻[q]
+
+        K_kl_ip = compute_block_at_point(grad_k, grad_l, D)
+        K_kl += K_kl_ip * w
+    end
+
+    K_blocks[k_local, l_local] = K_kl
+    return nothing
+end
+
+# ============================================================================
+# Microkernel contract for the DOF-based assembler
+# ============================================================================
+# `qpoint_buffer_eltype`, `update_qpoint_buffer!`, `evaluate_entry` are the
+# kernel-agnostic surface defined in `src/assemblers/microkernel.jl`. These
+# three methods opt `ContinuumKernel` in.
+#
+# The per-IP buffer is the elasticity tensor `𝔻` stored as a `SymmetricTensor`,
+# which is the only material data the displacement-only weak form needs.
+
+import ..JuliaFEM: qpoint_buffer_eltype, update_qpoint_buffer!, evaluate_entry,
+                   evaluate_mass_entry,
+                   reference_fields,
+                   DOFLayoutEntry, entity_local, component, extract_tangent!,
+                   AssemblyMaterialWorkspace, compute_stress
+using Tensors: SymmetricTensor
+
+@inline qpoint_buffer_eltype(::ContinuumKernel) = SymmetricTensor{4,3,Float64,36}
+
+"""
+    reference_fields(kernel::ContinuumKernel)
+
+The continuum-mechanics weak form needs `(σ, 𝔻)` per IP. For the linear
+case both are the constitutive evaluation at zero strain — a one-shot
+constant pre-computed here so Pass 1 of the DOF-based assembler can fill
+the per-element material workspace by simple copy.
+"""
+@inline function reference_fields(kernel::ContinuumKernel)
+    E_ref = zero(SymmetricTensor{2,3,Float64,6})
+    σ_ref, 𝔻_ref, _ = compute_stress(kernel.material, E_ref, NamedTuple(), 0.0)
+    return ((σ = σ_ref, 𝔻 = 𝔻_ref), NamedTuple())
+end
+
+# `buffer` is `AbstractVector` so the assembler can pass either a plain
+# `Vector{Buf}` (legacy) or a column view into a `Matrix{Buf}` (the new
+# flattened layout) without going through a copy.
+@inline function update_qpoint_buffer!(
+    buffer::AbstractVector{SymmetricTensor{4,3,Float64,36}},
+    workspace::AssemblyMaterialWorkspace{FieldType, StateType},
+    ::ContinuumKernel,
+) where {FieldType, StateType}
+    fields = getfield(workspace, 1)
+    extract_tangent!(buffer, fields, FieldType)
+    return nothing
+end
+
+"""
+    evaluate_entry(kernel::ContinuumKernel, geometry_cache,
+                   𝔻_vec::Vector{SymmetricTensor{4,3,Float64,36}},
+                   layout_i::DOFLayoutEntry, layout_j::DOFLayoutEntry,
+                   elem_id::Int) -> Float64
+
+Continuum-mechanics microkernel for the DOF-based assembler.
+
+Single-field (displacement) so the field index in each `DOFLayoutEntry`
+is ignored; only `(entity_local, component)` matter. Sums
+`∇Nᵢ : 𝔻 : ∇Nⱼ * detJ·w` over quadrature points and returns the scalar.
+The volume kernel ignores `elem_id`.
+
+Allocation-free; the inner integration is `@inbounds`. The actual
+`compute_stiffness_value` call is the same atomic kernel that the
+element-based assembler also uses, so by construction the two assemblers
+must agree to round-off.
+"""
+@inline function evaluate_entry(
+    kernel::ContinuumKernel,
+    geometry_cache,
+    𝔻_vec::AbstractVector{<:SymmetricTensor{4,3}},
+    layout_i::DOFLayoutEntry,
+    layout_j::DOFLayoutEntry,
+    ::Int,
+)
+    node_i = entity_local(layout_i)
+    comp_i = component(layout_i)
+    node_j = entity_local(layout_j)
+    comp_j = component(layout_j)
+
+    F = eltype(geometry_cache.detJ_w)
+    K_ij = zero(F)
+    n_ips = length(geometry_cache.detJ_w)
+    @inbounds for q in 1:n_ips
+        ∇N_i  = geometry_cache.∇N_data[q, node_i]
+        ∇N_j  = geometry_cache.∇N_data[q, node_j]
+        detJw = geometry_cache.detJ_w[q]
+        # Pass the SymmetricTensor straight through — the SymmetricTensor
+        # overload of compute_stiffness_value avoids the
+        # `Tensor{4,3}(::SymmetricTensor)` conversion that pulls in
+        # `Base.string` and breaks Metal codegen with `julia.new_gc_frame`.
+        K_ij += compute_stiffness_value(∇N_i, ∇N_j, 𝔻_vec[q], comp_i, comp_j) * detJw
+    end
+    return K_ij
+end
+
+"""
+    evaluate_mass_entry(kernel::ContinuumKernel, geometry_cache, qp_buffer,
+                        layout_i, layout_j) -> Float64
+
+Continuum mass matrix microkernel. Returns
+
+    M[i, j] = δ_{α,β} · ρ · Σ_q N_i(q) · N_j(q) · detJ·w(q)
+
+i.e. the *consistent* (not lumped) mass matrix block-diagonal in the
+displacement components (`δ_{α,β}` zeros the off-component entries).
+Returns `0.0` when `kernel.density == 0`, so kernels constructed without
+a density behave exactly as before — callers who only ever assemble `K`
+pay no extra cost.
+
+`qp_buffer` is unused in the linear case but kept in the signature so
+variable-density materials drop in via a one-line change.
+"""
+@inline function evaluate_mass_entry(
+    kernel::ContinuumKernel,
+    geometry_cache,
+    qp_buffer,
+    layout_i::DOFLayoutEntry,
+    layout_j::DOFLayoutEntry,
+)
+    F = eltype(geometry_cache.detJ_w)
+    ρ = F(kernel.density)
+    if ρ == zero(F)
+        return zero(F)
+    end
+
+    comp_i = component(layout_i)
+    comp_j = component(layout_j)
+    if comp_i != comp_j           # mass matrix is block-diagonal in (α, β)
+        return zero(F)
+    end
+
+    node_i = entity_local(layout_i)
+    node_j = entity_local(layout_j)
+
+    M_ij  = zero(F)
+    n_ips = length(geometry_cache.detJ_w)
+    @inbounds for q in 1:n_ips
+        N_i   = geometry_cache.N_data[q, node_i]
+        N_j   = geometry_cache.N_data[q, node_j]
+        detJw = geometry_cache.detJ_w[q]
+        M_ij += N_i * N_j * detJw
+    end
+    return ρ * M_ij
 end