Add NeurIPS 2026 archival paper and ICLR 2027 port

[johmathe]

What this PR adds

Two new top-level directories, both self-contained and building with tectonic bispectrum.tex:

• neurips_2026/ — archival copy of the NeurIPS 2026 submission (10902), byte-for-byte from the submitted zip except one fix: the invalid [anonymized] option on \usepackage{neurips_2026} was removed so the sources compile (anonymous is the sty default). Includes the full OpenReview reviews + rebuttal as reviews.md.
• iclr_2027/ — the ICLR 2027 version, targeting the Sep 25, 2026 full-paper deadline (abstract already registered on OpenReview).

Also: .gitignore gains LaTeX artifact entries, and .pre-commit-config.yaml excludes both paper directories (vendored style files and verbatim review text must not be auto-fixed).

NeurIPS → ICLR format changes

• neurips_2026.sty → iclr2027_conference.sty + times (official ICLR 2027 style zip; fancyhdr/natbib/.bst vendored); \iclrfinalcopy left commented for double-blind submission.
• Custom \authorblock tabular → standard ICLR \author{... \And ...} block (hidden until final copy; title page renders "Anonymous authors").
• Anonymity fix: the GitHub URL in the abstract → "Code is provided as anonymized supplementary material." (Verified: no author/affiliation strings or repo URLs anywhere in the built PDF; remaining name hits are third-person citations of published work.)
• \begin{ack} removed (commented Acknowledgments stub for camera-ready); checklist.tex dropped (ICLR has no checklist).
• Added page-limit-exempt sections before references, using the template's exact headings: AI use statement (mandatory — see author checklist), Ethics statement, Reproducibility statement.
• \bibliographystyle{plainnat} → iclr2027_conference.

Rebuttal commitments incorporated

Text corrections (all reviewers):

• Abstract/intro claims: "reduces the computational cost O(|G|²)→O(|G|)" → "coefficient count"; "consistently outperform" → data-efficiency-in-tested-settings wording; "sub-millisecond" scoped to tested GPU backends (H100, A100, Apple MPS) at batch 16.
• "Conjectured generic completeness, supported by Jacobian-rank and reconstruction evidence, not proved" stated in abstract, contribution bullet, Sec. 5 intro, Table 1 row, and Limitations; proven (finite groups/disk) vs. conjectured (sphere) split made explicit in Limitations.
• Terminology: "augmented selective invariant" for Φ_sel vs. "selective bispectrum" for its degree-3 subset, audited through Sec. 5 and the appendices.
• Sec. 2: operational definitions of Wigner-D, Clebsch–Gordan, and the parity rule; formal completeness definition added.
• Sec. 5: new "Why the finite-group argument does not transfer" paragraph (rank deficiency, Ω(L²) bound, parity rank collapse to ≈ℓ+2, rank repair); four-step proof roadmap with step (3) flagged as the open obstacle added to Appendix A.
• Related work: one-sentence differentiation from Edidin–Satriano (complete, cubic-size) and Bendory et al. (three-shell frequency marching) — honest that ours is not the first O(L²) construction.
• Organ3D "likely overfits" replaced by the optimization/capacity diagnosis with pointer to the new sweep; S²CNN rows labeled "(published, unmatched setup)" with no implied speed advantage; coverage 96.8% everywhere; TorusDnTorus→TorusOnTorus; Eq. (3) coefficient cross-reference fixed.

Rebuttal experiments (appendix, with main-text pointers):

• App. E: Spherical MNIST feature-set ablation (masked features, identical 768-channel input, 231,818 params; 94.09 → 94.48 → 95.00; full model 94.78; NR/R = NR/NR at every stage).
• App. F.5: reconstruction at the deployed L=15 (128×256 grid; MNIST 16/16 feature-converged, median 1.4e-3; 62.5% aligned; random signals 3/16 / 0/16, reported as-is). Sec. 5 footnote now points here.
• App. B.4: cross-backend timing table (A100 CUDA 13.0 / x86-64 CPU / M5 Pro MPS + CPU; batch 16, float32, median over 600 forwards; sub-millisecond on both GPUs).
• App. D.4: OrganMNIST3D 21-run regularization sweep at (16,32) channels — no overfitting signature, train–test gap 0.067–0.087, optimization/capacity conclusion. (Per-epoch curves from the rebuttal not reproduced; summary statistics only.)

Camera-ready-only commitments (kept open, stated in Limitations): tuned e3nn/MACE + scattering baselines; matched S²CNN re-run.

Page budget

Main text ends on page 9 (ICLR submission limit: 9). Statements + references are page-limit-exempt; 27 pages total. Zero overfull hboxes, no undefined references, line numbers render. Trims applied (in the plan's order): reduced figure widths, compressed the Related Work scattering paragraph, moved the supported-groups table to Appendix B with a pointer.

Author checklist before OpenReview upload

• Review and sign off on the AI use statement wording (marked with a TODO comment in the tex).
• Update the registered OpenReview abstract to the revised text (editable until Sep 25 AoE).
• Prepare the anonymized supplementary code zip referenced in the abstract and Reproducibility statement.
• Camera-ready only: uncomment \iclrfinalcopy, restore Acknowledgments.

Made with Cursor

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 95.46%. Comparing base (88ae4df) to head (df03243).

Additional details and impacted files

@@           Coverage Diff           @@
##             main      #51   +/-   ##
=======================================
  Coverage   95.46%   95.46%           
=======================================
  Files          11       11           
  Lines        2227     2227           
=======================================
  Hits         2126     2126           
  Misses        101      101           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.