One engineer wants embedding fusion to control latency...

This is a conflict between cost optimization and regulatory compliance. Since a well-pruned subgraph typically consumes only a few hundred tokens, the verbalization approach is safe and compliant within the one-quarter deadline. You should prioritize the ability to provide human-checkable citations for clinical claims, using fusion only for non-critical context or deferring it entirely until a robust attribution mechanism for opaque vectors is available.