[petlenz]

Part of #212 (does not close it: driver/material_point.h still pulls umat/ in, see below). Part of #221: the engine will throw cutback_error itself for a fixed point that does not converge (#214), so the error types cannot live under umat/.

What

• core/errors.h (new): fatal_error and cutback_error in numsim::materials. The header comment now describes any host; Abaqus is one example. It sits in core/, not driver/, because the property engine is its next user.
• umat/errors.h: includes core/errors.h and keeps umat::fatal_error / umat::cutback_error as using-declarations, so every UMAT catch site and downstream user keeps compiling. The Abaqus fatal path (xit_, set_fatal_handler, invoke_fatal) stays here because it is UMAT-specific.
• driver/external_state_source.h (moved from umat/): external_strain_source, external_deformation_gradient_source, external_scalar_source. umat/external_state_source.h remains as a forwarding header, because numsim-fft-homogenization includes it.
• The strain source no longer reads raw 6-slot buffers. The buffer overload of external_strain_source::bind decoded the UMAT's canonical Voigt order and pulled umat/tensor_conversion.h into the source. The evaluator now converts with umat::strain_from_buffer and binds tensors. The 9-slot row-major overload of the deformation-gradient source involves no Voigt convention, so it stays.
• driver/ throws fatal_error / cutback_error from numsim::materials and includes neither umat/errors.h nor umat/external_state_source.h.

Decided here

• umat/json_model.h stays in umat/. It includes umat_interface.h and binds the deck's *USER MATERIAL constants to parameters. report/case_runner.h uses only its builder; separating that builder from the constants binding is a separate change, if it is ever needed.
• Limit of this PR: driver/material_point.h still includes umat/statev_map.h, which includes umat/errors.h (and its weak xit_ declaration) and umat/tensor_conversion.h. So material_point.h is not yet free of umat/; core/errors.h and driver/external_state_source.h are. The remaining step is to move statev_map out of umat/ (noted on Move the host-neutral headers errors.h and external_state_source.h out of umat/ #212, which stays open).

Tests

test_host_neutral_headers, written first (red: core/errors.h missing):

• umat::fatal_error / umat::cutback_error are the same types as the numsim::materials ones;
• external_strain_source has no raw-buffer bind (a requires check);
• both errors are std::runtime_errors;
• the tensor bind(tensor2, tensor2) of the strain source stays callable (positive requires check);
• the forwarder umat/external_state_source.h is compiled in this test.

test_host_neutral_includes includes only core/errors.h and driver/external_state_source.h and #errors if UMAT_ERRORS_H is defined or the guards are not NUMSIM_MATERIALS_CORE_ERRORS_H / NUMSIM_MATERIALS_DRIVER_EXTERNAL_STATE_SOURCE_H. test_umat_forwarder_includes includes only the forwarder and #errors unless umat/tensor_conversion.h came with it.

test_external_state_source's buffer case now converts through umat::strain_from_buffer and checks the same values.

Gate

GCC 14, reports on, -DNUMSIM_PREFER_SIBLINGS=OFF: -O1 build 0 warnings, 1446/1446 tests pass.

[petlenz]

Review findings, addressed in b20e77d:

1. The test TU included umat/errors.h itself, so it proved nothing. New test_host_neutral_includes includes only core/errors.h and driver/external_state_source.h and #errors if UMAT_ERRORS_H is defined. Red before the fix: the two guard checks below (the UMAT_ERRORS_H check holds already and is a regression guard).
2. driver/material_point.h still includes umat/statev_map.h (-> umat/errors.h, the weak xit_). Not moved here. The PR body now states the limit and no longer says Closes #212; the remaining step is noted on Move the host-neutral headers errors.h and external_state_source.h out of umat/ #212, which stays open.
3. The forwarder umat/external_state_source.h had no coverage and had lost its transitive include of umat/tensor_conversion.h. It includes it again, and test_umat_forwarder_includes (red: #error "umat/external_state_source.h must still include umat/tensor_conversion.h") plus the forwarder include in test_host_neutral_headers cover it.
4. Guards are now NUMSIM_MATERIALS_CORE_ERRORS_H and NUMSIM_MATERIALS_DRIVER_EXTERNAL_STATE_SOURCE_H (red: #error "core/errors.h guard must be NUMSIM_MATERIALS_CORE_ERRORS_H", same for the driver header).
5. A positive requires { s.bind(t, t); } for two tensor2 sits next to the negative raw-buffer check. It passes before and after (a regression guard, not red).

Gate: GCC 14 -O1, build 0 warnings, 1446/1446 tests pass.

[petlenz]

Merged main (#231) into this branch. The only conflict was the include block of tests/test_j2_kinematic_plasticity.cpp (driver/external_state_source.h from here, umat/errors.h from #231).

Follow-up commit bc4cac5: the six materials #231 changed (J2, J2-RK, J2 kinematic, Drucker-Prager, neo_hooke, mean-field kernels) now include core/errors.h and throw the core cutback_error (same type as umat::cutback_error), so no material header pulls in umat/. test_host_neutral_includes.cpp includes the six headers; with the previous headers it fails with #error "a header above includes umat/errors.h".

Gate: GCC 14 -O1, BUILD=0 TEST=0, 0 warnings, 1460 tests.