fix(ml): DQN hyperopt overhaul — DSR reward, dead features, C51/noisy/network fixes
22-task overhaul (6 phases) for DQN training quality:
- Differential Sharpe Ratio (DSR) reward replacing raw PnL
- Remove 11 dead features (3 regime + 8 OFI): state_dim 54→43, feature_dim 51→40
- C51 v_min/v_max aligned to DSR Q-value range (±25.0)
- IQN batch path fix — consistent network for train+inference
- RMSNorm in distributional-dueling network
- Noisy layer sigma_init wired through config
- Rainbow config unified with DSR/C51/noisy defaults
- All dimension constants, comments, CUDA buffers updated
- 2747 lib tests passing, 0 failures, 0 warnings
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
a35a564f45