Issue filed by an AI agent on behalf of @ohdearquant.

What happens

ruvector-maxsim scores a query against a document by calling cosine for every (query token, document token) pair, and cosine recomputes the query token's norm inside each call. For a document of m tokens the same sum of squares is computed m times, so the inner loop does three multiply-adds per dimension where two would do.

Fix

PR #764 computes each query token's norm once per document and reuses it. The result is bit-identical (the test compares f32::to_bits(), not a tolerance) and the PR carries before and after measurements. It is up to date with main.