AgentFEM 0.2.6 promotion candidate

AgentFEM 0.2.6 was the immutable platform-foundation candidate used to prove the seven 0.3 promotion gates. It fixes the ownership boundaries between the readable engineering model and the FEniCSx/PETSc runtime, while closing several workflows that now depend on those boundaries.

A stable middle layer

Seven responsibilities are now explicit and machine readable:

Model -> Constitutive -> State -> Operator
      -> Procedure -> Backend -> Result/Verification

Model remains the engineering facade. State owns accepted/trial history and restart boundaries; Operator owns mathematical contributions; Procedure owns increments and algorithms; the backend owns compilation, assembly, DOFs and algebra; Result/Verification owns evidence and acceptance. Architecture tests prevent selected upward dependencies from silently returning.

Inelastic RVE foundation

The experimental finite-strain logarithmic J2 route now enters the ordinary model.step(...) lifecycle under both strong kinematics and reviewed affine-periodic constraints. It includes committed/trial quadrature state, consistent tangents, cutback, MPI-portable restart, regional materials, non-proportional loading, deterministic void and multi-void realizations, and Golden RVE evidence.

This is a usable research foundation, not a claim of universally validated finite-strain plasticity. Independent external structural validation remains an explicit promotion gate.

Migration and extension boundaries

Abaqus project migration preserves the recursive source graph, Part/Instance scope, element formulation identity, sets, sections, materials, Steps and unsupported assets. Native lowering remains a separately reviewed, fail-closed operation.

The core also publishes framework-neutral neural-field contracts. PyTorch and XDEM implementations remain in the separately installed AgentFEM-Learning companion, demonstrating that an advanced provider can extend AgentFEM without modifying its core.

Evidence before promotion

The 0.3 promotion audit now binds Linux/macOS installed-wheel acceptance, companion-provider evidence and a genuinely fresh AI-agent trial to an exact AgentFEM version, source commit and wheel digest. Deterministic CI cannot impersonate independent behavioral evidence. WSL2 remains the recommended Windows route and is tracked separately rather than blocking this release.

Compatibility

The recommended language remains:

Study -> Model -> scientific assets -> model.step(...) -> SimulationResult

Historical 0.2.x helpers remain compatibility delegates. This release changes ownership and evidence boundaries without requiring users to rewrite valid existing projects mechanically.