Training pipeline shared with v2: TD3 with distributional quantile critic, twin target networks with soft Polyak update, prioritized replay augmented by a sign-balanced sampler over PatternMemory KMeans clusters (UP/DOWN/FLAT quotas), policy-smoothing noise on target actions.
Encoder gradient source. encoder_grad_source switched from "both" to "td_only" — the encoder is updated by the TD signal only, not by the actor loss. Removes the late-cycle head degradation (state-agnostic actor) observed on prior runs.
New diagnostics. PR_eff / std for z and feat in actor and critic paths (early rank-collapse detection); per-cycle Cohen's d separating positive-reward vs negative-reward state distributions at z / feat / mean_raw / action levels; actor_head_gate (current tanh(gate), 0 = branch off, ±1 = fully active).