Newton tangent is NaN at rest for a fractional (non-half) power of the strain-rate invariant: the Jacobian guard covers half-integer powers only #841

What happens. A viscosity written on the bare invariant with a fractional power, for example the bounded shear-thinning law

edot = stokes.Unknowns.Einv2          # a bare sqrt(g), not a named expression
eta = 1 / (1 + edot ** sympy.Rational(1, 3))

has a finite residual at rest, but its consistent (Newton) Jacobian is NaN there. SymPy merges sqrt(g)**(1/3) into g**(1/6); the guard in _jacobian_unwrap adds 1e-36 to the base of half-integer powers only (exp.q == 2, on both JIT routes: _jit_graph.guard_half_integer_powers and the expanded route's _guard_sqrts), so the derivative g**(-5/6) * dg/dL is Inf * 0 = NaN at v = 0. The solve raises PETSc error 77 ("Computed maximum singular value as zero"). The alpha-blended kernel inherits it at alpha = 0 (0 * NaN = NaN).

Measured (16-cell box, driven lid): consistent_jacobian=False converges in 5 iterations; True raises error 77 on both routes, J(0) norm nan. With the invariant wrapped in a named expression the graph route converges (its node body sqrt(g) is guarded) and the expanded route still raises.

Why it matters now. On development the Newton tangent is opt-in. PR #794 makes it the default (one automatic cold Picard step through the blended kernel), so every cold solve of such a law raises.

Fix. Guard every non-integer power whose base has free symbols, on both routes. In floating point x + 1e-36 == x for any base above about 1e-20, so values and tangents away from rest are unchanged; only kernels that contain such powers change text. Found by the adversarial review of the #794 merge.