release: Unsga3 0.1.5

[jkbennitt]

Summary

Release prep for Unsga3 0.1.5 (2026-09-22). This PR bumps the package version and moves the merged forensic-audit wave (G1–G10) out of [Unreleased] into ## [0.1.5]. It does not tag v0.1.5 and does not publish to GitHub Packages or NuGet. After merge, the owner tags v0.1.5 and publishes.

Package id stays Unsga3. Version sources: Directory.Build.props, CITATION.cff, README status line, bug-report placeholder. No Version / PackageVersion overrides in the csproj files.

Left under [Unreleased] (not part of 0.1.5): the optional TypeSafe pass, and the earlier docs wording for initialPopulation.

What landed in 0.1.5

Honesty notes and small edge fixes from the forensic wave. The published 15-seed table is not rewritten. The default tournament stays RankNicheDistance — this release does not switch mating to Seada & Deb Algorithm 2 or to PymooCompatible.

• G1. DTLZ2 pymoo oracle passes n_var=12 (k=10) to match Dtlz2Problem. The published table used pymoo’s default n_var=10 and stays as published. Seed 1 was remeasured at n_var=12.
• G2. ZDT IGD compares the C# full non-dominated front with pymoo res.F. ReferenceDirectionThinning.OnePerDirection is a cardinality aid, not a parity claim.
• G3. Collapsed-nadir fallback (nadir = ideal + 1 when the span stays ≤ 1e-6) is documented. pymoo 0.6.2 stops at the worst point in the population. Behavior is unchanged.
• G4. Default RankNicheDistance is not Seada and Deb Algorithm 2. PymooCompatible matches the paper’s same-niche split; p_c stays 1.0 (paper experiments use 0.9). Default tournament unchanged.
• G5. Infeasible points still enter the hyperplane. Fixture: feasible (1, 1) beside infeasible (0, 0) keeps ideal at (0, 0).
• G6. Mating calls PrepareForSelection, which re-associates survivors. pymoo keeps the niche ids from survival.
• G7. WithDasDennis(1, 1) throws because N must be at least 2. Single-objective runs pass an explicit population size. N = 1 is not accepted.
• G8. Euclidean distance rejects a shorter vector. Single-point Pareto samplers throw. IGD+ has a hand-case test. ZDT6’s 0.280775 floor and ZDT3’s 0.1822287280 endpoint stay locked.
• G9. Duplicate elimination no longer spins when mutation cannot change the decision vector. The key is G12 significant digits. After the attempt cap, remaining offspring slots may be duplicates.
• G10. Equivalence docs no longer say the published Wilcoxon table is within 1–2% of pymoo. The DTLZ2 seed-1 guard is 2× the published mismatched scalar 0.00350. Das–Dennis Count uses a checked 64-bit combination. Two-layer reference directions remain absent.

Next step (owner, after merge)

1. Tag v0.1.5 on the merge commit.
2. Let the existing v* publish workflow ship the Unsga3 package (GitHub Packages). nuget.org is still the planned path, not this PR.

This PR does not create the tag and does not publish a package.

Checklist

• CHANGELOG.md: forensic/docs/fix items under 0.1.5; unrelated Unreleased items kept
• Version 0.1.5 in Directory.Build.props, CITATION.cff, README status line
• Compare links: [Unreleased] → v0.1.5...HEAD, [0.1.5] → v0.1.4...v0.1.5
• dotnet test Unsga3.slnx -c Release — 66 passed, 0 failed (SDK 10.0.401)
• No tag and no package publish in this PR
• Public API: ReferenceDirectionThinning already documented in XML comments from the forensic wave

Test plan

• dotnet test Unsga3.slnx -c Release — Passed: 66, Failed: 0, Skipped: 0.
• Local pack check (not published): Unsga3.0.1.5.nupkg, id Unsga3, version 0.1.5.