fix: set c51_alpha_max=0.5 — prevent C51 gradient convergence on H100

H100 fold 2 showed grad_norm→0 when C51 fully replaced MSE (alpha=1.0).
MSE provides a non-vanishing gradient floor. Default 0.5 = 50/50 blend.
Hyperopt can search [0.3, 0.9].

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

config/training/dqn-production.toml

@@ -82,7 +82,7 @@ q_clip_max = 50.0
 n_steps = 5
 tau = 0.005
 c51_warmup_epochs = 5
-c51_alpha_max = 1.0
+c51_alpha_max = 0.5
 her_ratio = 0.2
 cql_alpha = 1.0
 curiosity_weight = 0.1

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

@@ -1865,7 +1865,7 @@ impl DQNHyperparameters {
 
             // C51 warmup: use MSE loss for first 5 epochs (stronger gradient than C51 cross-entropy)
             c51_warmup_epochs: 5,
-            c51_alpha_max: 1.0, // Default: pure C51 after warmup. Hyperopt can search [0.3, 0.9].
+            c51_alpha_max: 0.5, // MSE always contributes 50%. Hyperopt searches [0.3, 0.9].
 
             // Decision Transformer pre-training: disabled by default
             dt_pretrain_epochs: 0,    // 0 = disabled