[petlenz]

What / why

Hosts classify a failure by its type (umat/errors.h): fatal_error stops the analysis, cutback_error (a std::runtime_error) means "retry with a smaller increment". Return-map and Newton non-convergence in the library threw plain std::runtime_error, so a host that classifies by type could not tell a step-size problem from a setup fault.

Contract

• Step-size-recoverable failures (non-convergence, no admissible multiplier, non-positive det F) throw cutback_error with the unchanged message.
• Setup and misuse errors keep their type (runtime_error, invalid_argument, logic_error); a smaller step cannot fix them.
• cutback_error derives from runtime_error, so the UMAT catch order in umat_interface.h is untouched (fatal_error first, then std::exception -> pnewdt = 0.25). Callers that catch runtime_error still work.

Changed sites

Site	Failure
materials/j2_plasticity.h (2)	return-map Newton did not converge; negative plastic multiplier
materials/j2_rk_plasticity.h	implicit stage did not converge
materials/drucker_prager_plasticity.h (2)	negative multiplier; smooth and apex Newton both failed
materials/j2_kinematic_plasticity.h (2)	coupled return map did not converge; negative multiplier
mean_field/kernels.h	self_consistent did not converge
materials/neo_hooke.h	det F <= 0 (was std::domain_error)

Deviations from the brief

• The vector_newton.h lines named in the brief (275, 298, 530) are not step-size failures: solve_with_factorization() called without a converged solve (API misuse), check_names (input validation) and the zero_blocks check (model definition). vector_newton::solve() itself never throws; it reports converged() and the materials throw. They are left as they are.
• Deliberately unchanged: j2_kinematic_plasticity.h graph-driven solver check (configuration, with a test that it is not a cutback_error), orientation.h ACG fit (preprocessing, not part of a step), all weighted_sum/property_engine/material_context/IO errors.
• Added beyond the brief: self_consistent (non-convergence) and neo_hooke (det F <= 0, previously domain_error; an inverted element is the usual reason for a cutback, and std::domain_error would not be caught by a host catching runtime_error).
• Materials include umat/errors.h for now; the type is still defined there.

Independence from #222

This PR does not depend on #222. #222 moves cutback_error to core/errors.h and keeps umat::cutback_error through a using-declaration, so both spellings stay valid. If #222 merges first, the only conflict to expect is the added #include "numsim-materials/umat/errors.h" lines, which still resolve through the forwarder, so no change is needed beyond taking either side.

Tests

Each family forces the failure and expects cutback_error; all were red first (the plain runtime_error/domain_error was thrown) and are green now.

• J2Softening.UnstableSofteningIsRejectedRatherThanClamped, J2Softening.ANonConvergedReturnMapAsksForASmallerStep (zero tolerance)
• J2Kinematic.ANonConvergedReturnMapAsksForASmallerStep, ANegativeMultiplierAsksForASmallerStep, AGraphDrivenSolverIsAConfigurationErrorNotACutback
• RkConvergence.AJ2RkStageThatRunsOutOfIterationsThrows, AFailedStageDoesNotInfluenceTheNextCall
• DruckerPragerErrors.ANegativeMultiplierAsksForASmallerStep, WhenBothNewtonsFailTheStepIsRetryable
• MeanFieldKernels.SelfConsistentThrowsWhenItDoesNotConverge
• NeoHooke.ANonPositiveDeterminantAsksForASmallerStep

Gate

Full gate (GCC 14, -O1): BUILD=0 TEST=0 warnings=0, 1454 of 1454 tests passed.