tests: cover the five materials that had none

[petlenz]

Closes #34.

Asserted against closed forms rather than recorded output:

material	assertion
exponential_isotropic_hardening	H = K_inf(1-e^{-δκ}), dH = K_inf δ e^{-δκ}, plus the saturation distinguishing it from linear hardening
linear_damage_law	all three branches, including the 0.999 cap that keeps (1-d) away from zero, and a zero range not dividing by zero
scalar_complement_weight	value = 1 - source
strain_energy_state_function	d(ε_eq)/dε against a central difference
vector_strain_state_function	|n·ε·m|, and that the derivative carries the sign the measure discards

The hardening and damage materials read a specific producer property, so the tests use a small stub publishing one scalar under a chosen name. That isolates the material rather than testing it behind a full plasticity model, and one context per value also shows it is a pure function of its input.

Two of my own assertions were wrong first

I read hardening_modulus as H when it is dH — the material publishes H as hardening_stress — and asserted H < K_inf at κ=20, which is mathematically true and numerically false once exp(-160) underflows. Both are fixed; recording them because a test written from a guessed property name is worth less than no test.

weighted_sum, also listed in #34, is covered on #29 and not duplicated here.

[petlenz]

Critical review. No findings. What I probed:

• Each assertion is against a closed form, not recorded output, so a changed formula fails rather than a changed number being blessed.
• The scalar_source_stub builds one context per value, so each material is shown to be a pure function of its input rather than of call order.
• The finite-difference check on strain_energy_state_function uses a central difference at h=1e-6 against an analytic derivative — tightening the tolerance to 1e-7 still passes, so it is not passing on a loose bound.
• The three-branch damage test crosses all three branches, asserted by construction rather than assumed from the increment.

The two errors worth keeping visible are already in the PR body: I read hardening_modulus as H when it is dH, and asserted H < K_inf at kappa=20, which is mathematically true and numerically false once exp(-160) underflows. Both failed loudly rather than passing wrongly, which is the only reason they were cheap.