Rank associative memories by normalized cosine similarity

[sylvesterkaczmarek]

Summary

Fixes #392.

Normalize query and stored embeddings when computing retrieval scores. This implements the documented cosine ordering rather than allowing embedding magnitude to dominate relevance. Zero-norm vectors receive similarity zero. Stored embeddings, checkpoint format, recency retrieval and existing tie handling are unchanged.

Validation

python -m pytest -q -o addopts='' -o required_plugins='' concordia/associative_memory

15 tests passed. All seven existing memory tests and eight new cases pass. Four new cases fail on unchanged upstream; four controls pass. Tests use real memory banks and deterministic local embeddings, with exact-match ranking, independent rescaling, an explicit normalized-dot oracle, restore/nonmutation, zero vectors and empty/unit-vector controls. Pytest ran serially with the repository's required xdist plugin setting disabled; no parallel-only fixtures are involved. No embedding service or language-model requests were made. Handling nonfinite embeddings and extreme norm overflow/underflow is outside this change.

Tested locally on macOS CPU using real module imports. New-test formatting, scoped static checks, syntax checks and git diff --check pass. No dependency or workflow changes. The full repository suite was not run.