A yielding model started from rest divides by zero: strainrate_inv_II_min defaults to 0 while the plastic viscosity is tau_y/(2 edot_II) #776

What

ViscoPlasticFlowModel / ViscoElasticPlasticFlowModel form the plastic viscosity as

eta_yield = yield_stress / (2 * edot_II)

and strainrate_inv_II_min defaults to 0. So a model with a yield stress, started from
rest, divides by zero on its first residual evaluation — before any iteration:

SNES solve diverged after 0 iterations: DIVERGED_FUNCTION_NANORINF

Starting from rest is not an exotic initial condition. It is what a standard benchmark
prescribes: Duretz, de Borst & Le Pourhiet (2019) state "All initial stress and strain
components are set equal to zero", and every incremental-loading setup begins the same way.

Why the default is the problem

The singularity is unconditional, not problem-dependent: tau_y/(2 edot_II) has no finite
limit as edot_II -> 0 for any non-zero yield stress. So the zero default guarantees a NaN
for a whole class of legitimate models, and the failure gives no hint of its cause — the
reason string names NANORINF, not the rate floor.

The fix is one line in a driver (Parameters.strainrate_inv_II_min = <small>), but the
user has to know it exists, and nothing in the error points there.

Note on what does NOT fix it

A Picard/fixed-point first iteration does not help: the NaN is in the RESIDUAL, not the
Jacobian, and Picard changes only the Jacobian. (Raised separately: solve(picard=0) is
the default, so there is no automatic fixed-point first step either — but that is a
distinct question.)

Suggested fix

Either:

• default strainrate_inv_II_min to a small positive value derived from the model's own
  scales rather than 0 — it exists precisely to regularise this and defaulting it off
  makes the singular case the default; or
• raise a clear error when a yield stress is set and the rate floor is zero, naming the
  parameter, instead of letting PETSc report NANORINF.

The second is the more conservative change and still removes the silent trap.

Acceptance

A test that builds a yielding model with zero initial conditions and solves: it must either
converge, or fail with a message that names strainrate_inv_II_min.