The project has a firm decision to keep v1 retrieval deterministic and keyword-based
(mneme_retrieval_deterministic). Adding ML-based dense retrieval introduces a heavy
dependency that breaks the pip-install-in-30-seconds contract and makes results
non-reproducible.

The correct path is to extend the existing keyword scorer: add field-weighted scoring
(scope x 2, constraints x 1.5), improve stopword filtering, or add phrase-level
matching. Keep retrieval auditable and reproducible — ML-based retrieval is the v2 opt-in.
