Dissipation positivity check

[IvarStefansson]

Assert dissipation positivity

Stacked on #1768 — the base branch is dissipation-driven-history, not develop.

Adds two parameter guards and a runtime check that the frictional dissipation
D = (μ* − tanψ)σ_n is non-negative at every converged solve.

What it is for. μ* − tanψ = μ_b(1 + tan²ψ)/(1 − μ_b tanψ) + μ_p is an algebraic identity, so with admissible parameters the dissipation is non-negative at every reachable state: the clips hold tanψ ≤ tanψ_0, and the guards hold μ_b tanψ_0 < 1, μ_b ≥ 0 and μ_p⁰ ≥ 0. The new runtime check therefore cannot fire on parameters. It catches an implementation error, and specifically one that loses or negates the μ_b term, that being the only term large enough to carry the sign. It also guards a future law in which μ_b or ψ_0 becomes state dependent, where a pole crossed mid-run would pass a check made once at construction.

The check is opt-in, via its own mixin. constitutive_laws.py today contains zero lifecycle-hook overrides — constitutive laws are pure operator providers, and the file's only validation idiom is "evaluate a sub-operator inside the operator-returning method and raise" (_validate_sliding_composition, BartonBandis, SecondOrderTensorUtils). Putting after_nonlinear_convergence on AsperityStressPartition would make a law class join the solve lifecycle and would charge every user of the partition for the check. A separate mixin keeps the laws clean and puts "this model asserts its own admissibility" in the base list where a reader sees it.

Changes

1. AsperityStressPartition.frictional_dissipation — (μ* − tanψ)σ_n. Both factors are read through self, so it is the composed coefficient and the damaged, partitioned dilation angle the model actually uses. Distinct from the total frictional work μ*σ_n that drives the wear, which includes the dilational part deliberately.
2. μ_p⁰ ≥ 0, validated in ploughing_friction_coefficient where the value is produced, so that what is checked is unambiguously the intact coefficient rather than the damage-scaled one.
3. μ_b ≥ 0, alongside the existing pole check in _validate_sliding_composition. Both are parameter-level preconditions of the same identity.
4. DissipationPositivityCheck — overrides after_nonlinear_convergence and raises, naming the offending cell count, the minimum and its index. Mixed into FractureDamageMomentumBalance. Zero is admissible and not reported: it is the frictionless limit.

Tests

test_dissipation_is_positive tightened from the cell mean to the smallest cell value — identical on the uniform states it prescribes, but a positive mean can hide negative cells. New tests cover the dissipation against a closed form in Pa, the check firing when the basic friction term is removed, the check silent on a correct composition, the frictionless limit accepted, and both parameter guards.

Mutation-tested, since a check nobody has made fire is untested scaffolding: dropping the characteristic_contact_traction factor fails the closed-form cases, and the firing test asserts D < 0 before expecting the raise so it cannot pass vacuously.

The check adds no operator to any equation, so results are unchanged.