feat: bottleneck 2→16 — 8× more market feature information for all branches

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

config/training/dqn-localdev.toml

@@ -90,7 +90,7 @@ causal_weight = 0.1
 causal_intervention_interval = 10
 adversarial_warmup_epochs = 50
 adversarial_checkpoint_interval = 50
-bottleneck_dim = 0
+bottleneck_dim = 16
 stochastic_depth_prob = 0.2
 asymmetric_dd_weight = 0.5
 time_reversal_mod = 5

config/training/dqn-production.toml

@@ -106,7 +106,7 @@ causal_weight = 0.1
 causal_intervention_interval = 10
 adversarial_warmup_epochs = 50
 adversarial_checkpoint_interval = 50
-bottleneck_dim = 0
+bottleneck_dim = 16
 stochastic_depth_prob = 0.2
 asymmetric_dd_weight = 0.5
 time_reversal_mod = 5

config/training/dqn-smoketest.toml

@@ -101,7 +101,7 @@ time_reversal_mod = 10         # 10% reversed vs production 20%
 position_entropy_weight = 0.02
 causal_weight = 0.05           # Half of production
 causal_intervention_interval = 50  # Less frequent than production (10)
-bottleneck_dim = 0             # Disabled for smoketest (tiny network can't afford bottleneck)
+bottleneck_dim = 16            # 16D bottleneck — 8× more market feature info for all branches
 adversarial_warmup_epochs = 100    # Beyond smoketest duration (don't activate saboteur)
 adversarial_checkpoint_interval = 100
 pruning_epoch = 100                # Beyond 50-epoch test (don't prune tiny network)

crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs

@@ -269,7 +269,7 @@ impl Default for GpuDqnTrainConfig {
             causal_weight: 0.1,
             causal_intervention_interval: 50,
             enable_gradient_vaccine: true,     // always on
-            bottleneck_dim: 2,
+            bottleneck_dim: 16,
             market_dim: 42, // Default: 42 base features. Overridden to 50 when OFI (MBP-10) enabled.
         }
     }

crates/ml/src/trainers/dqn/config.rs

@@ -1381,7 +1381,7 @@ impl DQNHyperparameters {
             hindsight_lookahead: 10,
             pruning_epoch: 50,                     // #20: compute mask at epoch 50
             pruning_fraction: 0.7,                 // #20: prune 70% of smallest weights
-            bottleneck_dim: 2,                     // #31: 2D information compression (gem of gems)
+            bottleneck_dim: 16,                    // #31: 16D bottleneck — 8× more market feature info (was 2, threw away 97.6%)
             stochastic_depth_prob: 0.2,            // #21: 20% drop probability per hidden layer
             mixup_alpha: 0.2,                      // Manifold Mixup: Beta(0.2,0.2) target distribution mixing
             // Core family intensity scalars: 1.0 = no change to defaults (neutral)