jgrusewski
62adf13187
feat(rl): P1+P2 kernels — PopArt, spectral norm, outcome head, Q-bias, per-branch LR
9 new CUDA kernels + OutcomeHead Rust struct + ISV slots 553-572:
PopArt: Welford-EMA reward normalization + V-head correction
Spectral: power iteration σ_max + L2 logit decoupling penalty
Outcome: K=3 trade-outcome classifier (Profit/Timeout/Loss)
Q-bias: EMA correction clipped ±10 for Bellman targets
Per-branch LR: adaptive per-head [0.5, 2.0] scaling from loss improvement
All unconditional — no feature flags. ISV-driven magnitudes.
No atomicAdd. Pre-compiled cubins.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>