Add hyenaND-retrofit Claude skill

[farhadrgh]

Adds a Claude Code skill that automates the attention -> HyenaND swap. Two paths:

• Native: in-repo configs that already use the build_attention_net / build_hyena_net / build_hybrid_net builders. Swap is mechanical — change the builder call, drop compile_compatible_fftconv = False, pick a layer pattern for hybrids. Sibling files are bare config shims (no __main__ block, matching the canonical hyena_patch16.py / hybrid_hhha.py).
• Foreign: external PyTorch models (timm ViT, HF transformers, hand-written) using nn.MultiheadAttention or F.scaled_dot_product_attention. Skill emits a sibling file with a full Hyena module plus a drop-in HyenaAttnAdapter that bridges the three nn.MHA ↔ Hyena interface gaps: tuple return, kwarg surface (need_weights/attn_mask/key_padding_mask), and channel layout / CLS handling.

The foreign path is organized around four orthogonal axes the user fixes up front:

1. data_dim ∈ {1, 2, 3} — picks ConvNd, sets CKConvND / SIRENKernelND / GaussianModulationND.
2. causal ∈ {True, False} — sets fft_padding, use_rope, mask parametrization, omega_0.
3. Host layout — tokens [B, N, C] vs feature_map [B, C, *S].
4. Return contract — (out, None) tuple vs bare tensor.

One parameterized adapter covers all sixteen combinations. FFT backend is selected automatically from the axes: fft_backend="subq_ops" (optimized CUDA kernel from subquadratic_ops_torch) when data_dim == 2 and not causal, else fft_backend="torch_fft" — the latter is required for 1D, 3D, circular padding, fp16, and causal paths because subq_ops is 2D-only in the current nvSubquadratic wiring (the upstream package ships fft_causal_conv1d and causal_conv1d but they are not exposed).

A foot-guns section calls out the failures that bite in practice: INT32 im2col overflow in large 3D short conv (≥160³ trips it; fix is DepthwiseFFTConv3d), RoPE divisibility (% 2 / % 4 / % 6 for 1D / 2D / 3D, must hold at every hierarchical stage), L_cache memory blowup ((2L-1)^D × D × 4 bytes — D=3 / L=512 ≈ 12.8 GB), state_dict-prefix mismatch on pretrained-attention loading, and the subq_ops constraint set (asserts loudly on data_dim ≠ 2, fft_padding ≠ "zero", causal, or fp16).

Hierarchical encoders with a use_hyena=True flag are an explicit exception to the sibling-file rule — the natural API is an in-place edit.

Five eval cases under evals/ cover the matrix:

• native-pure-swap, native-hybrid — v5_patch and vit5_hybrid configs.
• foreign-timm-vit — 2D, tokens, with CLS.
• foreign-3d-feature-map — 3D, feature_map, no CLS, F.scaled_dot_product_attention on [B, C, D, H, W].
• foreign-1d-causal-lm — 1D, causal tokens, attn_mask call site.

Together they exercise every (data_dim, causal, host_layout, prefix-tokens) combination the skill is meant to handle.

[tshimko-nv]

Looks good - one small comment about Claude vs. universal skill implementation, but otherwise, should be good.