Yield edge cases: units in sums (#758), a stated residual strength (#759), no silent smoother bump (#782)

[lmoresi]

Three yield-law defects, each with the regression test that fails on development.

#758: a dimensional constant in a sum entered unscaled. UWexpression.__add__ (and UWexpression + UWQuantity, __rsub__, UWQuantity.__sub__(UWexpression)) summed operand values eagerly; a quantity's value is its dimensional magnitude, so it entered a non-dimensional sum unscaled. smooth_max(a, b, 0) = (a + b + |a - b|)/2 hit it in a + b: a 0.35 MPa yield-stress floor compiled as 0.175 or 175000 instead of 0.0221 (in the issue's model a floor 83x below the yield stress stopped yielding). The sums are now symbolic unless both operands are quantities, so each term is non-dimensionalised by its own units. The same eager sum also froze an operand's value into the compiled kernel: changing a in a + b after the first solve changed nothing (test_0106).

#759: a tension cut-off at zero gave a singular operator. A yield stress that reads a field (Drucker-Prager through the pressure, a cohesion weakening with damage) floored at the zero default gives eta_pl = tau_y/(2 edot_II) = 0 wherever it is cut off. Per the maintainer ruling (2026-10-18), the VP, VEP and TI-VEP models now refuse it until the residual strength yield_stress_min > 0 is stated; -sympy.oo (bare or wrapped) is the opt-out for a yield stress positive by construction. The message names the model, its material and the fields read. Tests and examples whose yield stress reads a field now state a floor or the opt-out (15 tests, 9 docs/examples; Ex_Shear_Band_Plasticity_SS also had the floor misspelled yield_min, silently ignored).

#782: selecting yield_smoother="powermean" bumped a zero softness to 1.0, the harmonic mean, without a word. A family switch now leaves the stated softness alone; at delta = 0 the power mean runs at s = 1000, close to the exact Min. This is not the issue's preferred option (raise): a zero softness with the power mean is a legitimate, deliberate setting (test_1059_yield_anchor uses it), so the defect was only the silent change.

#776 is already fixed on development (VP and VEP started from rest converge: VP in 31 iterations, VEP in 3); not changed here.

Tests: test_1059_yield_stress_min_units (five spellings of one floor against the closed form eta = max(tau_y, tau_min)/(2R), inert and active; 7 of 10 fail on development), test_0106_a_sum_of_expressions_stays_live (fails on development: error 0.12 against 7e-10), test_1057::test_a_field_dependent_yield_needs_a_residual_strength (refusal, negative floor, closed form in tension, both opt-out spellings), test_1055::test_dp_model_smoother_optin. Tier A/B levels 1-2: 2621 passed; the one failure, test_1079_diffusion_disk_restart, fails on development too (#848).

Internal adversarial review, two rounds (findings in the comment below).

Closes #758
Closes #759
Closes #782

Checklist (UW3 Style Charter — docs/developer/UW3_STYLE_CHARTER.md)

• I have read the Style Charter and followed it over the surrounding code (S2)
• No drive-by refactors, renames, or "while I was here" cleanups (S9)
• Bug fixes ship the regression test, written first, with level_* and tier_* markers (S8)
• No hedging names (maybe_ / try_ / do_) and no commented-out code (S3, S4)
• Every exception swallow states its sanctioned failure mode in a comment (S4)
• New data access uses .array / mesh.X.coords — no with ....access(...), no mesh.data (S7)
• Parallel safety considered — np2/np4 checked where swarm/mesh/solver dispatch is touched (S11)
• No pixi.toml / pixi.lock or other dependency changes riding in an unrelated PR
• AI-assisted work carries the attribution line below

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]

Internal adversarial review — two rounds, probes against a build without the branch.

Round 1:

• Blocking, fixed: the branch's own test_1059 no-floor control read a field with the zero default and tripped Tension cutoff at yield_stress_min = 0 gives exactly zero viscosity and a singular operator (DIVERGED_LINEAR_SOLVE) #759; it now states -oo.
• Fixed: the Tension cutoff at yield_stress_min = 0 gives exactly zero viscosity and a singular operator (DIVERGED_LINEAR_SOLVE) #759 test's eta > 0 could not fail (unsolved state, viscous everywhere); now the closed form eta = tau_res/(2 edot_II) under prescribed tension and shear (1e-10).
• Fixed: a negative floor passed the check; any non-positive numeric floor is refused.
• Fixed: fault_network.py docstring examples (damage_yield, band_yield) now state the opt-out; nonlinear-solver trap list carries the rule.
• Fixed: the message names the model, _material_name and the variables read.
• Marked: TransverseIsotropicVEPSplitFlowModel's lagged estimate reads the unfloored yield stress (TODO(BUG), predates the branch).
• Found, now tested: the eager sum also froze an operand's value into the kernel (test_0106, fails on development).

Round 2 (no blockers): a wrapped -oo was refused with a message asking for -oo (and as a floor it evaluated to NaN before the branch) — _floor_is_off now reads it at the check and at every floor site; message prints the floor as given; a field-valued floor is documented as trusted.

Attacks that failed: 25 arithmetic combinations × units / no-units models all unwrap correctly (a units-model numexpr + qexpr was 302 on development, 2.0003 here); generated C hash identical under three PYTHONHASHSEEDs; no guard false positives (constant, constant expression, coordinate-only yield) in VP/VEP/TI-VEP; caught through nested expressions and swarm variables (proxy degree 0 and 1); zero and negative floors refused in every spelling incl. units; -oo opts out in all three models; the guard fires through MultiMaterialConstitutiveModel; the power mean at delta = 0 stays finite from 1e-30 to 1e25; CI reach (356 files), style gates clean.

Left: in a model without reference scales, mixed-unit sums still add magnitudes (29.1 MPa + 350 kPa unwraps to 379.1 on both builds) — predates the branch, separate from #758.

---

Generated by Claude Code