Module ops 

Pure scoring/transform kernels over SparseVector. No graph/runtime dependencies — this is the analogue of uni-btic’s interval math, kept in the type crate so every layer (index, rerank, brute-force oracle) calls one canonical implementation.

Functions

l2_norm

Euclidean (L2) norm of the weights.

prune_top_k

Keep only the k terms with the largest absolute weight, preserving the ascending-index invariant. This is the universal query-side latency lever for learned-sparse retrieval — high-DF / low-weight query terms dominate the posting-scan cost, so dropping them trades a little recall for large speedups.

sparse_dot

Dot product of two sparse vectors via a linear merge-join over their (ascending) term ids. This is the SPLADE/learned-sparse scoring primitive and the exact ground truth a brute-force oracle uses.