Same TD3 training loop, target-policy noise smoothing, n-step TD return (n = 4) and γ = 0.985 as full v2.4.18. Only capacity-bound knobs are scaled.
Replay buffer. 1.5M transitions (vs 3.6M full).
Buffer-aligned with v2.4.4-mini at exactly the same 1.5M capacity to make the 2×2 matrix apples-to-apples.
Sampler / pattern-memory. Sign-balanced sampling preserved (UP / DOWN / FLAT quotas) with identical logic to the full model.
Polyak target sync. Encoders, mixers, critics, and CSC running means all polyak-targeted on the same schedule as v2.4.18.