inclusion and mean_field_composite: dilute, Mori-Tanaka and self-consistent

[petlenz]

Closes #107, closes #143. Part of the uvwmaterial port (#94); stacked on #130.

mean_field_tangent: dilute, Mori–Tanaka and self-consistent effective stiffness as a tangent producer over phase nodes (#130). matrix and phases are phases (produced tangent, fraction and Eshelby tensor); kernel dilute | mori_tanaka | self_consistent. Outputs tangent, concentration_<phase> for every phase (the matrix included) and residual; the stress is a linear_stress on the tangent. Dilute: Tᵢ = (I + Sᵢ : Cₘ⁻¹ : (Cᵢ − Cₘ))⁻¹; Mori–Tanaka: Aᵢ = Tᵢ : (cₘ I + Σ cⱼ Tⱼ)⁻¹. The Eshelby tensors come in as properties, so their Poisson's ratio is whatever node produces it (isotropic_moduli of the matrix tangent, or a constant) — the composite reaches into no other node and isotropises nothing.

Self-consistent as a graph cycle. Every phase, the matrix included (its own Eshelby node — closes #143), is embedded in the effective medium, whose ν is isotropic_moduli on this node's tangent (tolerance: 0, the projection without the isotropy check). The Eshelby nodes consume that ν over a local edge ("nu_edge": "local", starting from nu_initial): EdgeKind::Local imposes no ordering, so one update() is one relaxed Picard step (relaxation, default 0.5) on the tangent of the previous pass, and residual = ‖C_new − C‖/‖C_new‖. The fixed point is run by whoever calls update():

• material_point_evaluator gets residual_property, residual_tolerance, max_passes and repeats the graph update per host call until the residual is below tolerance (report documents: "point": {..., "residual": "scs::residual", "tolerance": 1e-10, "max_passes": 200}).
• A solver's Newton loop re-evaluates the material anyway; the driver and the UMAT close the point through the evaluator.

The old version iterated inside the node with material_ref calls eshelby(nu) into the inclusions and a hard-coded sphere for the matrix; both are gone. isotropic_moduli with a non-positive tolerance projects without the check (needed for the effective medium).

Tests (test_mean_field_tangent, 13): a phase republishes sphere, cylinder and spheroid tensors from its node; a phase without a shape is rejected at finalize(); Mori–Tanaka spheres match the analytic K and G (= Hashin–Shtrikman lower bound) at c = 0.05, 0.3, 0.6 to 1e-10 with zero residual; dilute matches its closed-form K and the Mori–Tanaka limit as c → 0; two phases of one material equal one; voids soften with the exact K; Σ c A = I and the linear_stress gives C : ε for every kernel; the self-consistent fixed point is symmetric under swapping matrix and inclusion (1e-8) and stiffer than Mori–Tanaka, converges within the pass loop and stays put on another pass; the evaluator closes the fixed point in one call; aligned fibres are transversely isotropic with E_L near Voigt; overfull composites and unknown kernels throw; JSON.

Report docs/materials/mean_field_tangent/report.json: Mori–Tanaka, dilute, self-consistent (its own Eshelby node at ν_eff over the local edge, converged per step through the point's residual) and unidirectional on one point, the closed forms as checks.

862/862 tests pass (GCC 14, reports on).

JSON input

New material type: mean_field_tangent. A json_model document that loads (the FromJson test of the PR):

{"materials": [
  {"type": "external_strain_source", "name": "strain"},
  {"type": "constant_scalar", "name": "Km", "value": 3.0},  {"type": "constant_scalar", "name": "Gm", "value": 1.0},
  {"type": "constant_scalar", "name": "Ki", "value": 60.0}, {"type": "constant_scalar", "name": "Gi", "value": 30.0},
  {"type": "isotropic_tangent", "name": "Cm", "K_source": "Km", "G_source": "Gm"},
  {"type": "isotropic_tangent", "name": "Ci", "K_source": "Ki", "G_source": "Gi"},
  {"type": "isotropic_moduli", "name": "moduli", "tangent_source": "Cm"},
  {"type": "eshelby_sphere", "name": "sphere", "nu_source": "moduli", "nu_property": "poisson_ratio"},
  {"type": "constant_scalar", "name": "c", "value": 0.3},
  {"type": "constant_scalar", "name": "cm", "value": 0.7},
  {"type": "phase", "name": "spheres", "tangent_source": "Ci", "fraction_source": "c", "eshelby_source": "sphere"},
  {"type": "phase", "name": "matrix", "tangent_source": "Cm", "fraction_source": "cm"},
  {"type": "mean_field_tangent", "name": "mf", "matrix": "matrix", "phases": ["spheres"], "kernel": "mori_tanaka"},
  {"type": "linear_stress", "name": "composite", "tangent_source": "mf", "strain_source": "strain"}
]}

Self-consistent: the matrix phase carries an Eshelby node too, both nodes read ν of the effective medium over a local edge, and the point closes the fixed point:

{"materials": [
  {"type": "isotropic_moduli", "name": "moduli_eff", "tangent_source": "mf", "tolerance": 0},
  {"type": "eshelby_sphere", "name": "sphere", "nu_source": "moduli_eff", "nu_property": "poisson_ratio", "nu_edge": "local", "nu_initial": 0.35},
  {"type": "phase", "name": "spheres", "tangent_source": "Ci", "fraction_source": "c", "eshelby_source": "sphere"},
  {"type": "phase", "name": "matrix", "tangent_source": "Cm", "fraction_source": "cm", "eshelby_source": "sphere"},
  {"type": "mean_field_tangent", "name": "mf", "matrix": "matrix", "phases": ["spheres"], "kernel": "self_consistent", "relaxation": 0.5},
  {"type": "linear_stress", "name": "composite", "tangent_source": "mf", "strain_source": "strain"}
 ],
 "point": {"strain_source": "strain", "stress_source": "composite", "tangent_source": "mf", "residual": "mf::residual", "tolerance": 1e-10, "max_passes": 200}}

Outputs tangent, concentration_<phase> (matrix included), residual. Several inclusions: more phase entries (each with its own Eshelby node) in phases.

[petlenz]

Review findings (design, tmech expressions instead of component loops, modern C++, comments, tests, correctness). Severity: must = wrong result, hang, build break or unreachable feature; should = design, duplication, accuracy; nit = style.

inclusion + mean_field_composite — needs changes

• [must] SC fixed point with isotropised Eshelby tensors and mixed shapes → tangent without major symmetry: ‖C−Cᵀ‖/‖C‖ = 4.6e-2 (spheres 0.2 + spheroids ar 10, probe); MT/dilute ≤ 3e-16. Symmetrise each iterate and the result, or restrict SC to isotropic media and throw; add a major-symmetry test for mixed shapes.
• [should] C = (C + C_new)/2 unconditional: with normalised concentrations plain iteration converges in 17–27 iterations, relaxed in 46–67 (probe). Drop or expose relaxation (default 1). The divergence cited in the PR body was uvwmaterial's unnormalised kernel.
• [should] dilute sets A_matrix = Is (Σ c A ≠ I, test special-cases it) → (Is − Σ cᵢTᵢ)/cm as short_fibre does.
• [should] :111-116 duplicates MT start verbatim; poisson_of duplicated in short_fibre → isotropize.h: require_isotropic(C, tol); std::clamp(poisson, −0.99, 0.49) hides an iterate leaving the admissible set; Eshelby recomputed every update (memoise in inclusion::eshelby(ν)).
• [should] inclusion.h:53-54 direction copy loop → tmech::adaptor<value_type const,3,1,tmech::full<3>>(d.data()); volume_fraction required even with fraction_source (tests write a dummy 0.0).
• [should] static_assert(Dim == 3) inside types registered unconditionally → register_default_materials<material_policy_2d>() breaks; register under if constexpr (Dim == 3).
• [nit] convergence tested after relaxation; no non-convergence test; ellipsoid never built through inclusion; SC only checked by ordering/swap (Hill SC closed form for spheres would pin 1e-10).

[petlenz]

Rebased on the Eshelby-node refactor of #128: inclusion now names its eshelby node instead of carrying shape parameters (the uvwmaterial model, one tensor object per inclusion). Of the review findings this commit also covers:

• should, poisson_of duplicated: gone, require_isotropic (core/isotropize.h) for the matrix, isotropize + poisson_ratio for the self-consistent iterate.
• should, Eshelby recomputed every update: the node caches its tensor by ν and eshelby(nu) returns the cached value when ν is the matrix's.
• should, direction copy loop: moved into the node (tmech::adaptor<…, full<3>>).
• should, static_assert(Dim == 3) in unconditionally registered types: inclusion and mean_field_composite register under if constexpr (Dim == 3).

Still open from the review (not part of the refactor): symmetrising the self-consistent tangent for mixed shapes, the ½ relaxation, dilute's A_matrix, the Mori–Tanaka start duplicated as the SC start, volume_fraction required beside fraction_source (#137).

[petlenz]

Producer/consumer pass: inclusion is a phase of any composite — material, fraction_source (the only way to give a fraction; the volume_fraction parameter is gone, so the parameter-or-source duality the review flagged cannot occur) and an optional eshelby node. The Eshelby nodes themselves consume nu_source/nu_property (#128): fixtures and the report wire isotropic_moduli of the matrix tangent to them, so a matrix whose tangent evolves re-evaluates its Eshelby tensors through the engine. mean_field_composite rejects a phase without a shape at update.