feat(constitutive): power-mean yield soft-min smoother + rampable delta constant

[lmoresi]

What

Adds the power-mean soft-min yield smoother and makes the soft-min softness δ a constants[] atom (runtime-rampable with no JIT recompile) on the ViscousFlowModel base class, so the visco-plastic subclasses inherit one implementation. This is the generalisable, scalable substrate for a yield homotopy toward the sharp Min surface (companion to the solver difficulty/bounded-solve tooling in #377).

Base class (ViscousFlowModel):

• _combine_yield(eta_ve, eta_pl) -- the shared viscous/plastic combination, keyed on yield_mode: "min" (exact hard Min), "harmonic", or "softmin" (the δ soft-min in the family chosen by yield_smoother). Development's tri-modal yield_mode semantics are preserved exactly (the source branch had collapsed "softmin" into "min"; that is deliberately NOT adopted).
• _get_yield_softness / _get_yield_offset -- δ and the onset offset held as constants[] UWexpression atoms (the offset defined symbolically in terms of δ = one symbol in the stress tensor, no tensor blow-up; one δ update repacks both). Ramping δ via the atom + solver._update_constants() forces no recompile.
• yield_smoother property/setter -- "sqrt" (default; overshoots τ_y in the transition) or "powermean" (undershoots τ_y -- eta_eff <= Min always -- overflow-safe on geodynamic ranges).

Models (Stage 1 -- isotropic, test-covered):

• ViscoPlasticFlowModel (Drucker-Prager): defaults to yield_mode="min" == today's exact hard Min (zero behaviour change), and gains yield_mode / yield_softness knobs so it can opt into the softmin / power-mean homotopy -- the model the hard-case DP homotopy needs.
• ViscoElasticPlasticFlowModel (VEP): its inline sqrt soft-min routed through _combine_yield; δ moves float -> atom (value-identical); yield_softness setter syncs the atom.

Why

The power-mean smoother is the preferred, physically-admissible (undershoot) homotopy family; holding δ as a rampable constant is what makes a scalable δ-continuation cheap (no recompile per step).

Safety -- zero behaviour change at stock settings

_combine_yield at default settings is bit-identical to the previous inline law (verified numerically, diff 0.00e+00) -- δ merely moves from a baked float to a constants[] symbol evaluated to the same number. Power-mean and runtime-δ are entirely opt-in. The distrusted in-solve enable_yield_homotopy ramp driver is intentionally not part of this PR.

Scope

Stage 1 (this PR): base helper + δ-atom + power-mean, wired to the DP and VEP models. Stage 2 (follow-up): the TI-VEP / TI-VEP-Split anisotropic sites (4 sites feeding rank-4 split/lagged tensors), gated by the TI solver tests.

Tests

tests/test_1055_yield_smoother.py: δ=0 exact Min; power-mean undershoot + overflow-safety + ->Min as δ->0; smoother validation/default; runtime δ ramp no-recompile; DP opt-in.

level_1 and tier_a gate green: 402 passed, 0 failed, 0 errors (1 pre-existing unrelated xfail).

Note: 3 test_1052_VEP_stability_regression (level_2) tests are red, but they fail identically on the development baseline without this change (bisected) -- pre-existing, unrelated.

Underworld development team with AI support from Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[lmoresi]

Closed as fully superseded — audit against current development (698d89a):

• The power-mean soft-min smoother, δ and onset-offset as rampable constants[] atoms, _combine_yield tri-modal semantics, the yield_smoother property, and tests/test_1055_yield_smoother.py all landed on development (via the Nonlinear solver: automatic warm start, model-advertised yield homotopy, and an inf-safe power-mean yield law #441 line of work), so everything this PR introduced is already merged.
• Development's power-mean is the inf-safe harmonic-normalised form (constitutive_models.py ~1010-1024, with the inf/inf-at-rigid-points rationale documented inline); this branch still carries the older product-over-sum form that NaNs on a cold start — merging would regress that fix.
• The remaining lines unique to this branch are pre-refactor remnants development deliberately replaced (old -oo defaults, old Max(yield_stress_min, …) guards superseded by the smooth_max floor, old flux_jacobian docstring superseded by the settable version from fix(constitutive_model): make flux_jacobian settable for custom Jacobian tangents #362).

Nothing of value is lost by closing.

Underworld development team with AI support from Claude Code