v0.9.1

• Fourier-Bessel filter basis; Hann window basis with per-type init factors via get_init_factors
• Standardized L2 normalization on the unit disk (harmonic, Zernike, Fourier-Bessel); on a disk of radius R the norm equals R via the Jacobian
• New DISCO basis normalization modes modal (mean-subtracted, reduces spectral leakage) and geometric (spherical cap area measure)
• Deprecated basis_norm_mode="individual" → "nodal" and "area ratio" → "geometric" (old names emit DeprecationWarning)
• Faster DISCO sparsity-pattern setup; OpenMP forward/backward kernels with up to ~55x speedup in some configurations
• Cross-attention (key != value != query) in AttentionS2, NeighborhoodAttentionS2, and DistributedNeighborhoodAttentionS2
• Serial attention upsampling when nlon_out % nlon_in == 0: CPU/CUDA/torch upsample kernels and matching reference
• DistributedNeighborhoodAttentionS2 for self-attention and downsampling (distributed upsample not yet implemented)
• Optional per-head QK RMS norm (use_qknorm) for AttentionS2 and NeighborhoodAttentionS2; shape checks across attention layers
• Fixed Q/K/V projection gain when input dim != embedding dim
• Breaking: default NeighborhoodAttentionS2 scale changed from 1/sqrt(k_channels) to 1/sqrt(k_channels // num_heads) to match standard MHA head-dim scaling (num_heads > 1)
• Faster Legendre coefficient precomputation for SHT layers
• Differentiable polar_halo_exchange and get_group_neighbors for distributed attention
• More robust distributed transpose; _reduce clones before all_reduce for torch.compile compatibility
• Fixed Galewsky initial condition NaN from overflow; convolution adapter for mismatched residual channel counts
• Midpoint rule for filter-basis L2 norm integration (O(h^2)); improved _precompute_convolution_tensor_s2 docstring
• Expanded attention tests (including upsample); new tests/test_filter_basis.py; broader layer integrity coverage