RR4/RR5: ML replay, model cards, and audio explanation

[ACFHarbinger]

Implement RR4 and RR5: deterministic ML policy/training replay with model card and human-in-loop events, plus optional local AnalyserNode audio enhancement with mute, reduced-motion, and no-autoplay behavior.

Roadmap IDs: RR4, RR5