fix(generalization): production quality — replace unwraps + update plan
Replace 8 .unwrap() calls in bottleneck/causal code with proper
ok_or_else() error handling. These unwrap on Option<CudaSlice> buffers
that are only accessed when their feature is enabled, but production
code should never panic — return MLError instead.
Update plan doc: 23 of 34 tasks marked [DONE]:
- Tasks 1-4: domain rand, CQL, S&P, adversarial injection
- Tasks 5-6: anti-correlation (beta_penalty), curiosity Q-penalty
- Tasks 10-11: mirror universe, time-reversal
- Tasks 13, 17-19: vol norm, regret, DD loss, position entropy
- Tasks 21-25: stochastic depth, noise, masking, anti-LR, clustering
- Tasks 27: ensemble disagreement
- Tasks 31-34: bottleneck, vaccine, self-play, causal intervention
Remaining 11 tasks require deeper architecture changes:
- #7 counterfactual augmentation (2x replay buffer)
- #8 feature importance filtering (gradient×activation)
- #9 DANN regime adversarial (new network head)
- #12 price-level invariance (dual forward pass)
- #14 market maker adversary (GAN architecture)
- #16 phantom liquidity (synthetic data generation)
- #20 lottery ticket pruning (weight mask + retrain)
- #26 HER regime oversampling (replay buffer modification)
- #28 flattest selection (sharpness computation)
- #29 cross-fold consistency (walk-forward infrastructure)
- #30 FP32 experience weights (weight buffer type change)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
f20f36527b