fix: IQN num_quantiles 64→32 in all binaries + production config

Default was hardcoded as 64 in train_baseline_rl.rs and evaluate_baseline.rs,
overriding the config default of 32. Added num_quantiles=32 to production
config so it's explicit.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

config/training/dqn-production.toml

@@ -91,6 +91,7 @@ her_ratio = 0.2
 cql_alpha = 1.0
 curiosity_weight = 0.1
 iqn_lambda = 0.25
+num_quantiles = 32
 spectral_norm_sigma_max = 1.5
 spectral_decoupling_lambda = 0.01
 # With per-component clipping removed, raw gradient norms are ~4000.

crates/ml/examples/evaluate_baseline.rs

@@ -918,7 +918,7 @@ fn evaluate_dqn_fold(
             let gamma = hp_f64(hp, "gamma").unwrap_or(0.95);
             (10.0 / (1.0 - gamma) * 1.2).clamp(20.0, 300.0)
         }) as f32,
-        iqn_num_quantiles: hp_usize(hp, "num_quantiles").unwrap_or(64),
+        iqn_num_quantiles: hp_usize(hp, "num_quantiles").unwrap_or(32),
         ..DQNConfig::default()
     };
 
@@ -1086,7 +1086,7 @@ fn evaluate_dqn_fold_gpu(
             let gamma = hp_f64(hp, "gamma").unwrap_or(0.95);
             (10.0 / (1.0 - gamma) * 1.2).clamp(20.0, 300.0)
         }) as f32,
-        iqn_num_quantiles: hp_usize(hp, "num_quantiles").unwrap_or(64),
+        iqn_num_quantiles: hp_usize(hp, "num_quantiles").unwrap_or(32),
         ..DQNConfig::default()
     };
 

crates/ml/examples/train_baseline_rl.rs

@@ -402,7 +402,7 @@ fn build_dqn_hyperparams(
             (10.0_f64 / (1.0 - gamma) * 1.2).clamp(20.0, 300.0)
         }),
         noisy_sigma_init: hp_f64(hp, "noisy_sigma_init").unwrap_or(0.5),
-        num_quantiles: hp_usize(hp, "num_quantiles").unwrap_or(64),
+        num_quantiles: hp_usize(hp, "num_quantiles").unwrap_or(32),
         noisy_epsilon_floor: hp_f64(hp, "noisy_epsilon_floor").unwrap_or(0.05).into(),
         hold_penalty_weight: hp_f64(hp, "hold_penalty_weight").unwrap_or(0.01),
         max_position_absolute: hp_f64(hp, "max_position_absolute").unwrap_or(2.0),