Expand scientific validation beyond the planted cpnet comparison corpus

[simoneSantoni]

Motivation

The current audit uses three small planted networks and cpnet commit 6aad458a6d434a3617d33e74f7163d514a27fecb. It is useful causal evidence, but it is not broad scientific validation and cpnet is not a correctness oracle.

Proposed validation matrix

• Add direct mathematical or exhaustive small-graph oracles where feasible.
• Add canonical networks from the cited papers with provenance and expected summary statistics.
• Sweep synthetic size, density, noise, core fraction, pair count, weight heterogeneity, and directionality.
• Repeat stochastic algorithms over documented seed ensembles and report uncertainty.
• Compare against additional maintained implementations when mathematical estimands match.
• Separate correctness checks from concordance and recovery benchmarks.

Acceptance criteria

• Every algorithm has at least one independent oracle or literature-backed fixture.
• Dataset provenance, licenses, preprocessing, seeds, and package revisions are recorded.
• Validation reports score semantics before comparing outputs.
• A small deterministic subset runs in CI; the full study is reproducible offline.
• Results are summarized in versioned machine-readable data and human-readable documentation.

[simoneSantoni]

Resolved in 48d3047. Added the scientific validation matrix and provenance report, 1,056 deterministic exhaustive/direct oracle checks, algorithm-level fixtures, and versioned machine-readable plus human-readable comparison artifacts. Concordance is explicitly separated from correctness.