Feasible niching, equal-CV sort, and two-layer reference directions

[jkbennitt]

Summary

Constrained survival now follows pymoo Survival.filter_infeasible (Jain & Deb, NSGA-III Part II): niche only feasible members, fill any shortfall by ascending constraint violation, and build the hyperplane from feasible objectives when any feasible member exists. An all-infeasible generation does not move the hyperplane. The locked whole-pool ideal fixture is replaced in the same change.

Equal constraint violation in CompareConstraintDominated is mutual non-domination (Deb). Tournament comparators, including default RankNicheDistance, are unchanged. TwoLayerDasDennis adds Part I’s inside layer; single-layer DasDennis / WithDasDennis stays the default, and the M≤3 oracles do not use the second layer. OSY, TNK, and C1-DTLZ1 have formulation and survival self-tests. No new IGD table, and the package version is not bumped.

Checklist

• dotnet test Unsga3.slnx -c Release passes locally
• CHANGELOG.md updated under Unreleased (if user-visible)
• Public API changes documented (XML docs / README)
• If survival / normalization / metrics changed: noted oracle impact (or re-ran multi-seed)

Test plan

• dotnet test Unsga3.slnx -c Release — 87 passed.
• Feasible (1, 1) beside infeasible (0, 0) now sets ideal to (1, 1), not (0, 0). An all-infeasible population leaves a previously learned hyperplane in place.
• Survival tests: enough feasible members exclude an infeasible point; a shortfall is filled by ascending CV with no niche id; equal CV keeps the earlier index; an all-infeasible pool truncates by CV.
• Equal-CV individuals that Pareto-dominate each other share a front.
• OSY / TNK / C1-DTLZ1 match the pymoo-shaped formulas at fixed points, exercise CV fill on evaluated individuals, and keep a feasible seed through a 3-generation run. No IGD assertion.
• Two-layer Das–Dennis: outer prefix, scale 0.5, centroid duplicate removed, scale 1 collapses to the outer layer. A 2-generation DTLZ2 run accepts the array. Existing single-layer counts are unchanged.
• Unconstrained niche fixture PrepareForSelection_overwrites_survival_niche_ids still locks the previous Das–Dennis p=1 survival ids. Published Wilcoxon tables are not rewritten.