Sample directed edges in Erdos-Renyi null models

[SkBlaz]

Erdős–Rényi null models for directed networks currently use an undirected sampler and the undirected density denominator. A complete four-node digraph with 12 arcs becomes a graph with only 6 arcs, all oriented according to node order. This biases the null distribution used for hypothesis testing.

Use the input graph's directionality for the sampling probability, NetworkX generator, and output network. Directed graphs sample ordered pairs at m/(n*(n-1)); undirected sampling remains unchanged. Document the expected-density semantics.

Validation:

• Original code: 5 new regression failures and 7 passing controls.
• Fixed code: all 12 new cases pass, including complete digraphs, seeded sparse graphs, isolated nodes, and the executor.
• Combined regression and related suites: 76 passing tests on Python 3.12 and Python 3.9. Three additional Erdős–Rényi property tests pass on Python 3.12.
• One unrelated seed-contract test fails identically on unchanged master and the fixed branch: tests/test_nullmodels_contracts.py::test_generate_null_model_uses_per_sample_seeds_consistently. It expects seed+i instead of the executor's spawned child seeds; deselected only for the Python 3.9 validation.
• Focused Ruff F/B checks and git diff --check pass.