v2.7 — Single PortfolioGeneralizer replaces the per-symbol actor. The policy now ingests feat_critic (B, N, 512) and the full distributional q_dist (B, N, H, 32) and emits portfolio weights jointly via a 2-layer Transformer (d = 512, 8 heads). Symbol Mixer is kept on the critic side only (applying it to the policy in v2.6 collapsed attention into uniform per-symbol outputs). CVaR-aware actor target mean(Q) + 0.4 × CVaR₄₀%(Q). Reward gains an env-side drawdown penalty + efficiency bonus.
v2.6 — Multi-horizon QR-TD3 critic: nstep_horizons [4, 8, 12, 16] × num_quantiles 32, critic output (B, N, 4, 32).
v2.4.4-mini / v2.4.18-mini — Pure-size compression cells of the 2×2 CSC matrix.
v2.4.18 — Symmetric CrossSampleCenter, world-model + CEM planner.
v2.4 — N-agnostic refactor: branch / V_global separation; global Q (q1g / q2g) removed.
v2.3 — Surgical fixes on v2.1: separate encoders, qg_new.detach().
v2.2 — Minimalist refactor under Simple_encoder.
v2.1 — Adapter rank-collapse fix; ResidualSkipHead actor.
v2 — Initial release.