Track: Track2; Team name: LoopCounter; Model: EHNN

[FoxHound0x00]

Add EHNN (Kim et al., ECCV 2022 (arXiv:2208.10428)) as a Track-2 TNN backbone for the 2026 TDL Challenge, with config, unit tests (100% backbone coverage), a pipeline-test entry, and the GraphUniverse evaluation results.json.

Checklist

• My pull request has a clear and explanatory title.
• My pull request passes the Linting test.
• I added appropriate unit tests and I made sure the code passes all unit tests.
• My PR follows PEP8 guidelines.
• My code is properly documented, using numpy docs conventions, and I made sure the documentation renders properly.
• I linked to issues and PRs that are relevant to this PR.

Description

Track-2 (TNN) submission for the 2026 TDL Challenge — team LoopCounter.

Adds EHNN — Equivariant Hypergraph Neural Networks (Kim, Oh, Cho, Hong — ECCV 2022, arXiv:2208.10428; reference code jw9730/ehnn) as a hypergraph backbone.

Included:

• topobench/nn/backbones/hypergraph/ehnn.py — the EHNN-Linear (MLP) variant: the V2E and E2V maximally-expressive equivariant aggregation blocks with order-conditioned positional hypernetworks, all derived from the incidence matrix and its hyperedge sizes / node degrees (Hᵀx / |e| and (x_v + H x_e)/(1+deg), plus global and order-conditioned bias terms).
• Reuses the existing HypergraphWrapper (no new wrapper needed).
• configs/model/hypergraph/ehnn.yaml
• test/nn/backbones/hypergraph/test_ehnn.py — 100% backbone coverage; covers the operator cache, hyperedge-order clamping, sparse/dense incidence, and the building-block modules.
• hypergraph/ehnn added to test/pipeline/test_pipeline.py.
• 2026_tdl_challenge/outputs/.../results.json — full 12×3×2 GraphUniverse grid (community detection + triangle counting, in-distribution + OOD).

Issue

N/A — this is a model contribution for the 2026 TDL Challenge (Track 2 / track-2-tnn), not a bug fix.

Additional context

The original input projection and MLP classifier head are provided by TopoBench's feature encoder and readout, so this backbone consumes pre-encoded node features and returns node embeddings (consistent with the existing EDGNN backbone).

results.json was generated via the evaluation notebook's own functions (run_challenge_grid + save_challenge_artifacts), because the published 2026_tdl_challenge/run_evaluation.ipynb currently fails its own SHA-256 integrity check on main (embedded expected_hash f87b2c… ≠ the actual hash 3c1d78… of the locked cells) — happy to open a separate issue for that.