Identical TD3 training loop, schedule, n-step TD return (n = 4), and γ = 0.985 to v2.4.4. Capacity-bound knobs aligned with v2.4.18-mini.
Replay buffer. 1.5M transitions — exactly matched to v2.4.18-mini.
Sampler / pattern-memory. Sign-balanced sampling preserved (UP / DOWN / FLAT quotas).
Polyak target sync. Encoders, mixers, critics on the standard v2.4.4 schedule (no CSC running means to sync, since there are no CSC modules).