fix(dqn,ppo): comprehensive verification audit fixes — 14 files, 7 agents
DQN correctness:
- N-step Bellman target uses gamma^n (was gamma^1), fixing ~2% Q-value bias
- Cash reserve enforcement actually reduces position (was warn-only)
- CVaR sort_last_dim(true) for IQN random tau ordering
- Marsaglia-Tsang gamma sampling uses normal(0,1) not uniform(0,1)
- DQNConfig::default state_dim 51→54 (51 market + 3 portfolio)
PPO correctness:
- set_learning_rate preserves trained weights (was recreating entire model)
- update_value_only() for critic pretraining (was training both networks)
- PolicyNetwork::entropy() single forward pass (was 2x GPU compute)
- LSTM entropy uses proper H=-sum(p*log(p)) (was -mean(log_probs))
- grad_norm metric set to None (was reporting policy loss as gradient norm)
Config parity (train/eval/hyperopt/enhanced_ml):
- PPO hyperopt state_dim 51→54, value_hidden_dims 3→5 layer
- enhanced_ml feature_count 16→54, PPO policy_hidden_dims [128,64,32]→[128,64]
- Eval: tensor core alignment, Rainbow from hyperopt params, warmup alignment
- GPU batch model_state_dim 51→54 (matches kernel output)
Infrastructure:
- QNetwork dropout training mode (AtomicBool toggle, was always disabled)
- reward_history Vec→VecDeque (O(1) front removal, was O(n))
- Plateau detection distinguishes worsening from plateau in log messages
2732 tests pass, 0 clippy warnings.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
cdb32cd8bf