Minimum-lengthscale constraints in Python

[smartalecH]

Stacked on #11 (base branch is docs/supported-features) — review that one first.

Ports the geometric minimum-lengthscale constraints from the Julia package to ssp_topopt, closing one of the gaps recorded in docs/features.md.

What's here

ssp_topopt.constraints adds constraint_solid and constraint_void, following the Julia implementation in src/julia/SSP/src/constrain.jl:

• the geometric constraints of Zhou et al. (2015), with the hyperparameter-free thresholds derived in Arrieta, Romano & Johnson, arXiv:2507.16108 — decay rate c = 64 R^2, threshold eps = 1e-8, and the conic-filter threshold points eta_e / eta_d of Qian & Sigmund (2013) (eqs. 7-12 of that paper);
• the gradient of the filtered density is obtained from the same differentiable cubic interpolation (utils.gradient) that ssp2 uses;
• the return value is normalized by the constraint threshold, matching the Julia API, so it can go straight into an nlopt inequality constraint: nonpositive means feasible.

Any SSP order works. The constraints only take rho_filtered and rho_projected, so they compose with ssp1_bilinear, ssp2, or a plain tanh_projection; a test exercises all three end to end.

Example

examples/python/ssp_constrained_example.py has two parts:

1. A sweep over a stripe geometry, which shows the solid constraint crossing the threshold right at the target lengthscale (qualitatively fig. 1c of the paper).
2. A two-stage optimization against a target pattern with bars thinner than the target lengthscale. Stage 1 is unconstrained and reproduces the sub-lengthscale bars (constraints at ~1e6 times the threshold); stage 2 turns the constraints on and repairs the design (both constraints negative, i.e. feasible) at a cost of FOM 1.2e-4 -> 8.7e-2.

Testing

tests/python/test_constraints.py (14 tests, 20 subtests; the full suite is 21 passing):

• threshold functions against the closed-form values, plus the eta_e + eta_d = 1 invariant;
• thin stripes/gaps violate the constraint and thick ones satisfy it, with the physical width measured from the projected design rather than assumed;
• reverse-mode gradients w.r.t. both rho_filtered and rho_projected match central finite differences to ~1e-7 (in float64 — the normalized constraint is O(1/eps), so single-precision differencing is pure roundoff);
• composition with each projection, jit parity, and argument validation.

As an extra check outside the test suite, I compared against the Julia implementation on identical arrays: solid 1.4147e6 (Python) vs 1.4174e6 (Julia) and void 1.0382e6 vs 1.0388e6 — 0.2% and 0.06%, consistent with the two packages using different cubic interpolation schemes (interpax cubic2 vs FastInterpolations CubicFit) for the density gradient.

Not included

Dilation/erosion and cubic-interpolation SSP1 remain Julia-only; docs/features.md is updated to reflect only what this PR adds.

[romanodev]

LGTM. Just a few cosmetic changes.