fix(ml): correct DQNConfig defaults — widen v_min/v_max, reduce cql_alpha, disable IQN

- v_min/v_max: -2/+2 → -10/+10 (C51 needs room to separate 45 actions)
- entropy_coefficient: 0.01 → 0.05 (5x stronger anti-collapse)
- cql_alpha: 1.0 → 0.1 (was adding ~3.8 loss penalty per step)
- use_iqn: true → false (IQN trains q_network but inference uses dist_dueling)
- Added get_count_bonuses() method for batch action selection

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

crates/ml/src/dqn/dqn.rs

@@ -235,8 +235,8 @@ impl Default for DQNConfig {
             dueling_hidden_dim: 128,
             use_distributional: true,  // BUG #36 FIXED: C51 re-enabled
             num_atoms: 51,
-            v_min: -2.0,   // Bug #5: Corrected from -10.0
-            v_max: 2.0,    // Bug #5: Corrected from 10.0
+            v_min: -10.0,  // Widened: [-2,2] compressed Q-values, [-10,10] gives C51 room to separate 45 actions
+            v_max: 10.0,   // Widened: [-2,2] compressed Q-values, [-10,10] gives C51 room to separate 45 actions
             use_noisy_nets: false,
             noisy_sigma_init: 0.5,
             enable_q_value_clipping: true,
@@ -246,17 +246,17 @@ impl Default for DQNConfig {
             gradient_collapse_patience: 100,
 
             // Entropy regularization: prevent 45-action collapse
-            entropy_coefficient: 0.01,
+            entropy_coefficient: 0.05,  // 5x stronger anti-collapse for 45-action space
             noisy_epsilon_floor: 0.05,
             use_count_bonus: true,
             count_bonus_coefficient: 0.1,
 
             // CQL: Enabled by default for offline training
             use_cql: true,
-            cql_alpha: 1.0,
+            cql_alpha: 0.1,  // Reduced from 1.0: full strength added ~ln(45)=3.8 loss penalty, crushing Q-value differentiation
 
-            // IQN: Enabled by default (complementary to C51)
-            use_iqn: true,
+            // IQN: Disabled: IQN trains base q_network but inference uses dist_dueling network (zero gradients)
+            use_iqn: false,
             iqn_num_quantiles: 64,
             iqn_kappa: 1.0,
             iqn_embedding_dim: 64,
@@ -3150,6 +3150,13 @@ impl DQN {
         self.count_bonus.reset();
     }
 
+    /// Get count bonuses for all actions (UCB exploration).
+    /// Returns a Vec<f32> of length num_actions with exploration bonuses.
+    pub fn get_count_bonuses(&self) -> Vec<f32> {
+        let bonuses = self.count_bonus.bonuses();
+        bonuses.iter().map(|b| *b as f32).collect()
+    }
+
     /// Update target network by copying weights from main network
     fn update_target_network(&mut self) -> Result<(), MLError> {
         // Wave 11.6: Update target network (hybrid > dueling > standard)