diff --git a/test/assemblers/runtests.jl b/test/assemblers/runtests.jl
new file mode 100644
index 0000000..650ce14
--- /dev/null
+++ b/test/assemblers/runtests.jl
@@ -0,0 +1,120 @@
+using Test
+using JuliaFEM
+
+@testset "Assemblers" begin
+    # Locks in zero-allocation + correctness + 0 LLVM allocation sites
+    # for the element-template-driven DOF-based assembler.
+    include("test_dof_based_zero_alloc.jl")
+
+    # Matrix-free `apply_K!` correctness vs assembled K, zero
+    # allocations, 0 LLVM gc-alloc sites, and a Krylov.cg validation
+    # via LinearOperators + IterativeSolvers.
+    include("test_dof_based_apply_K.jl")
+
+    # Disjoint element subsets + `apply_K_contributions!` sum to `apply_K!`
+    # (fake two-rank reference); partition metadata + multiply-buffer hook.
+    include("test_dof_based_partitioned_matvec.jl")
+
+    # Full matvec + row mask; disjoint vertex-owned rows sum to `apply_K!`.
+    include("test_apply_K_masked_rows.jl")
+
+    # Structured Hex8 slab partitions, DOF closures, contribution sums.
+    include("test_partitioning_helpers.jl")
+
+    include("test_partitioning_zero_alloc.jl")
+
+    # Partition adjacency, halo DOF lists, ReferenceMaskMultiplyLayout.
+    include("test_halo_exchange.jl")
+
+    # Per-partition packed DOF layout + gather/expand + matvec glue.
+    include("test_packed_layout.jl")
+
+    # partitioned_owned_matvec! orchestration (serial halo replica).
+    include("test_partitioned_matvec.jl")
+
+    # Backend-agnostic apply_K! via KernelAbstractions: same kernel
+    # on CPU(), CUDABackend(), MetalBackend(), AMDGPUBackend(),
+    # oneAPIBackend(). Locally validates the CPU() backend produces
+    # bit-equivalent output to the direct CPU apply_K!.
+    include("test_dof_based_apply_K_ka.jl")
+
+    # Mass-matrix microkernel through the DOF-based assembler.
+    # Validates `evaluate_mass_entry`, `apply_M!`, and `assemble_M!`
+    # for both `ContinuumKernel` and `HeatKernel`, including
+    # row-sum (= rho * V), block-diagonal-in-components structure for
+    # elasticity, density linearity, and the same zero-alloc
+    # contract as `apply_K!` / `assemble!`.
+    include("test_dof_based_mass.jl")
+
+    # Neumann loads (NodalForce + UniformBodyForce) through
+    # `apply_load!`. Locks in row-sum identity (int b dV = b dot V),
+    # additive composition, end-to-end Poisson with body source vs
+    # the analytical T(x) = Q x (L - x) / (2 k) solution, and
+    # zero allocations for both load types.
+    include("test_dof_based_loads.jl")
+
+    # `BlockJacobiPreconditioner{N}` for vector problems where the
+    # 3x3 nodal block has full off-diagonal coupling. Validates
+    # `compute_block_diagonal!` against assembled K, that
+    # `ldiv!(P, x)` is the exact block-diag inverse, that
+    # PenaltyDirichlet + BlockJacobi CG matches the direct solve,
+    # and that BlockJacobi reaches the same residual in <= as many
+    # iterations as scalar Jacobi.
+    include("test_block_jacobi.jl")
+
+    # Float32 (single-precision) `apply_K!` through the
+    # precision-parametric KernelAbstractions cache. Locks in the
+    # storage typing of `to_float32(cache)`, F32-vs-F64 single-
+    # precision agreement on both `ContinuumKernel` and
+    # `HeatKernel`, the precision-mismatch guard, and the round-
+    # trip back-compat of the default Float64 KA path.
+    include("test_dof_based_apply_K_f32.jl")
+
+    # `SurfaceLoad` distributed-traction integration via
+    # `apply_load!`. Validates the row-sum identity
+    # `Sigma f = t * area` for both quad (Hex8 face) and tri (Tet4
+    # face) faces, end-to-end pull (3D elasticity) and 1D heat-
+    # conduction problems, additive composition with
+    # `UniformBodyForce`, and zero-allocation hot path.
+    include("test_surface_load.jl")
+
+    # `ICholPreconditioner` (IC(0) -- incomplete Cholesky with
+    # zero fill-in). Locks in algebraic correctness on
+    # tridiagonal/dense SPD matrices, the diagonal-shift retry
+    # for near-indefinite inputs (without aliasing the input
+    # `K`), `ldiv!` agreement with `(L * L')^{-1} * x`, fewer
+    # PCG iterations than scalar Jacobi on a stiff elasticity
+    # cantilever, end-to-end PenaltyDirichlet matrix-free PCG
+    # via the `(cache, asm, kernel, mesh; dirichlet)` factory,
+    # and zero-allocation `ldiv!`.
+    include("test_ichol_preconditioner.jl")
+
+    # `LinearMPC` -- penalty-enforced linear multipoint
+    # constraints sharing the `apply_constraint_*` hook
+    # protocol with the Dirichlet types. Locks in tuple ->
+    # flat-CSR packing, assembled-vs-matrix-free agreement on
+    # heat (periodic) and elasticity (multi-master),
+    # end-to-end periodic-heat PCG matches the direct solve,
+    # composition with `PenaltyDirichlet` for an
+    # inhomogeneous-offset rigid-link elasticity solve, and
+    # zero-alloc `apply_constraint_post!` / `_diag!`.
+    include("test_linear_mpc.jl")
+
+    # `lowest_eigenpairs` / `solve_eigenproblem` -- matrix-free
+    # generalized eigensolve `K phi = lambda M phi` via subspace
+    # iteration with Rayleigh-Ritz. Locks in algebraic
+    # correctness on dense SPD test problems, matrix-free
+    # apply_K!/apply_M! agreement with assembled K, M, recovery
+    # of the analytical 1D heat spectrum, and the high-level
+    # wrapper's shift-invert path for free-free systems with
+    # rigid-body / null-space modes.
+    include("test_eigensolve.jl")
+
+    # Typed `MatrixFreeOperator` / `MatrixFreeMassOperator` --
+    # `LinearAlgebra.mul!`, `eltype`, `size`, `op(y, x)` callable,
+    # `op * x` allocating mat-vec, optional Dirichlet folding into
+    # every mat-vec, plug-in via `LinearOperators.LinearOperator`
+    # into `IterativeSolvers.cg!`, and zero allocations on `mul!`
+    # after warmup.
+    include("test_matrix_free_operator.jl")
+end