Phase C (#90): FD-verify the generated consistent tangent end-to-end

[petlenz]

What

Ticks the "point numsim-materials' numerical_diff_checker at a codegen-generated material" box of the verification spine (#90). Independently verifies the strain-coupled consistent tangent shipped in #101.

The prior tangent e2e (ResidualReturnMapConsistentTangentHasCouplingTerm) pins specific rank-4 components against a hand-derived closed form. This adds a check that finite-differences the whole dσ/dε through the real property graph: the checker perturbs the strain, re-solves the Newton (reverting the z history so each sample restarts from the same z_old), and compares every component against the emitted dstress_dstrain. Because the coupling is captured by the re-solve rather than by construction, this is a genuinely independent check — it would catch any component the closed-form assertions didn't enumerate.

Verification

• max_rel_error = 5.4e-11 across 8 loading steps (smooth response ⇒ tight central-difference agreement); gate at < 1e-6 (5 orders of margin, not flaky).
• Sabotage-verified: dropping the coupling correction in the emitter makes the FD check fail at rel = 0.71, and the checker's worst-component diagnostics point exactly at the coupling entries (C(0,0,1,1): analytical 0 vs numerical c·ε₀₀). A verification test that can't fail is worthless; this one fails on the exact defect it exists to catch.

Notes

• Lives in the existing g++-gated tensor e2e (NCG_TENSOR_E2E); reuses tangent_checker<policy> (the numerical_diff_checker<Traits, tensor2, tensor2> alias) exactly as the numsim-materials J2 tests do.
• Next on Verification spine: pull numerical-correctness verification (FD tangent + golden model + gates) to the front #90: a synthetic unit test for the verifier's rel-vs-abs worst-failing branch, and the validated J2 golden as a standing gate.