Release v0.4.0 (integration of #53–#57 + coverage)

[johmathe]

Integration PR for v0.4.0. Merges the five fix branches, bumps the version, adds versionchanged notes to the three affected class docstrings, and carries the coverage work.

Status: #53–#57 are merged. This PR is now just the version bump, the versionchanged docstring notes, and the coverage commit, rebased onto main.

Stack

PR	Branch	Fixes
#53	fix-bessel-j0-j1-accuracy	torch.special J₀/J₁ were ~5e-7 accurate; Miller recurrence for all orders
#54	fix-disk-basis-conditioning (on #53)	SO2onDisk default basis cond ~1e16 → ≈2 (Nyquist band-limit)
#55	fix-dn-odd-n-fold	DnonDn dropped the fold block for odd n ≥ 5; invariant was incomplete
#56	fix-so3-selective-rank	SO3onS2 selective map rank-deficient for L ≥ 8; CUDA-graph cache key
#57	fix-minor-numerics	rotation float64 path, CnonCn.invert pivots, OctaonOcta._lm_step acceptance

Breaking changes (0.4.0)

output_size / feature layout changes for SO3onS2 (all lmax ≥ 3; L=16: 430 → 426), DnonDn (odd n ≥ 5; +16 entries) and the default SO2onDisk truncation (L=16: 105 → 76). Cached features and heads trained on the old layouts must be recomputed. The on-disk sparse CG cache self-invalidates.

Coverage (8a469f4)

Local pytest --cov=src/bispectrum: 95.6% → 98.6% (31/2181 statements missing), no pragma: no cover.

• Removed dead code: scalar _bisect_newton/_jn_scalar/_djn_scalar, unreachable num_brackets <= 0, n2d == 0/n3 == 0, and an empty-slice guard in the SO3 power loop.
• Bug found by the new tests: _load_or_compute_reduced_cg didn't catch pickle.UnpicklingError; a corrupt reduced-CG cache file crashed SO3onS2 construction.
• New tests: version fallback; dense-vs-sparse selective parity (incl. CG power); reduced-CG disk cache compute/load/corrupt/write-failure; CUDA-graph capture wiring with torch.cuda stubbed on CPU; triangle-violating CG triples; order-0 batch roots.
• Remaining uncovered: defensive numerical fallbacks in the D_n / octahedral CG constructions and the f.is_cuda dispatch.

Verification

747 passed / 23 skipped, pre-commit --all-files clean, benchmarks/verify_linear_bootstrap.py --lmax 100 ALL PASSED.

Follow-ups (not in this stack)

• benchmarks/figures/table_benchmarks.tex and paper_benchmarks.tex hard-code the old counts (430 / 105).
• neurips-2026-rebuttal text relying on "full rank after augmentation" for L ≥ 8 needs revising.
• Experiments with cached SO3onS2 / odd-n DnonDn / default SO2onDisk features need recomputing.
• Clean up config, CI, correctness, and API surface #48 touches the same files as SO3onS2: fix selective-map rank deficiency for L >= 8; key CUDA graph on input signature #56/Fix minor numerics: rotation precision, Cn invert pivots, Octa LM acceptance #57 and will conflict; rebase or close.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.57%. Comparing base (0744394) to head (f03227c).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #58      +/-   ##
==========================================
+ Coverage   95.61%   98.57%   +2.96%     
==========================================
  Files          11       11              
  Lines        2235     2181      -54     
==========================================
+ Hits         2137     2150      +13     
+ Misses         98       31      -67     

☔ 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.