G1–G10: document forensic-audit gaps without rewriting the 15-seed table

[jkbennitt]

Summary

Ten commits, one per verified finding from the 2026-09-22 forensic audit. Docs, protocol, fixtures, and guards come first. The default tournament stays RankNicheDistance. The 15-seed Wilcoxon number table is not regenerated. No version bump.

Commits

1. G1 — DTLZ2 oracle passes n_var=12 (k=10), matching Dtlz2Problem. pymoo’s default n_var=10 (k=8) is documented as the published mismatch. Seed 1 remeasured; docs/WILCOXON-RESULTS.md numbers left as published.
2. G2 — ZDT IGD set definition: C# scores the full non-dominated front; pymoo res.F is the niche optimum. Seed 1 remeasured on a shared reference front. ReferenceDirectionThinning.OnePerDirection is a cardinality aid, not a parity claim.
3. G3 — Collapsed nadir stays ideal+1 when the span remains ≤ 1e-6. Unit test locks {2, 2+1e-8} → nadir 3. pymoo 0.6.2 stops at worst-of-population.
4. G4 — RankNicheDistance is not Seada & Deb Algorithm 2. PymooCompatible matches the same-niche split and pymoo’s distance-tie coin flip. pc stays 1.0 (paper experiments use 0.9). Default tournament unchanged.
5. G5 — Fixture: feasible (1, 1) beside infeasible (0, 0) sets ideal to (0, 0). Ideal and worst still come from the whole pool.
6. G6 — PrepareForSelection re-normalizes survivors and overwrites niche ids. pymoo keeps the survival ids. Five-point fixture locks the current ids.
7. G7 — WithDasDennis(1, 1) throws because N must be ≥ 2. Single-objective runs pass populationSize. N=1 is not accepted.
8. G8 — Euclidean distance rejects a shorter vector. IGD+ hand case. Zdt1(1) and the other one-point samplers throw. ZDT6 floor 0.280775 and ZDT3 endpoint 0.1822287280 are documented and locked (pymoo writes 0.182228780).
9. G9 — Duplicate-elimination attempt cap when mutation cannot diversify; remaining slots may be duplicates. The key is G12 significant digits, not 12 decimal places.
10. G10 — Published Mann–Whitney figures replace the 1–2% sentence. DTLZ2 seed-1 bar is 2× the mismatched scalar 0.00350, so a regression to about 2.9× fails. ReferenceDirections.Count uses a checked long and throws past Int32. Two-layer reference directions stay absent.

Measured seed 1 (pasted from this run, pymoo 0.6.2)

DTLZ2, p=12, pop=92, 150 gen, PymooCompatible, n_var=12:

Solver	Console IGD	Meta IGD	Front
C#	0.00403168	0.004031675764658275	92
pymoo	0.00308392	0.003083921253245871	91

Ratio of the meta values: 1.30732. One seed, fronts still differ by one point. The 15-seed pymoo column remains the old n_var=10 run (seed 1 there is C# 0.00403168 / pymoo 0.00349879).

ZDT1, p=12, pop=52, 100 gen, default tournament. Shared 100-point pymoo PF:

Set	n	IGD
C# full non-dominated front	52	0.05119280568479224
pymoo population, non-dominated	52	0.05378307132263516
pymoo res.F	13	0.0628633417931784

C# on the library 500-point sampler reprinted 0.0514307. Not an algorithm ranking.

Docs vs code

• Docs / protocol / fixtures: G1 notes, G2 set definition, G3 nadir note, G4 mating note, G5–G7 fixtures, G6 re-association note, G8 endpoint notes, G10 equivalence wording and two-layer absence.
• Code: oracle n_var=12; optional thinning helper; Euclidean length check; single-point sampler guards; duplicate-elimination attempt cap; Das–Dennis Count checked long; DTLZ2 CI factor 3× → 2×. No survival/mating/nadir formula change. Default tournament unchanged.

Checklist

• dotnet test Unsga3.slnx -c Release passes locally (66 tests)
• CHANGELOG.md updated under Unreleased
• Public API changes documented (XML docs / README)
• Survival / normalization / metrics: behavior of the hyperplane and mating is unchanged; oracle impact is documented. Multi-seed table was not re-run.

Test plan

• dotnet test Unsga3.slnx -c Release — 66 passed.
• DTLZ2 and ZDT1 seed-1 oracles remeasured as tabulated above.
• docs/WILCOXON-RESULTS.md numeric rows were not regenerated.