feat(ml): wire PPO accumulation_steps from hyperparams to PPOConfig

Exposes gradient_accumulation_steps in PpoHyperparameters so hyperopt
and training binaries can configure effective batch scaling. The actual
accumulation logic already existed in PPO::update_mlp().

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

crates/ml/src/trainers/ppo.rs

@@ -73,6 +73,9 @@ pub struct PpoHyperparameters {
     pub initial_capital: f64,
     /// Average bid-ask spread for GPU portfolio simulation (default: 0.0001 = 1bp)
     pub avg_spread: f64,
+    /// Number of gradient accumulation steps (default: 1 = no accumulation).
+    /// effective_batch = mini_batch_size * accumulation_steps
+    pub accumulation_steps: usize,
 }
 
 // REMOVED: Default implementation for PpoHyperparameters
@@ -123,6 +126,7 @@ impl PpoHyperparameters {
             gpu_timesteps_per_episode: 500,       // Default: 500 timesteps per episode
             initial_capital: 1_000_000.0,         // Default: $1M
             avg_spread: 0.0001,                   // Default: 1bp (ES/NQ futures)
+            accumulation_steps: 1,                // Default: no accumulation
         }
     }
 }
@@ -164,7 +168,7 @@ impl From<PpoHyperparameters> for PPOConfig {
             lstm_hidden_dim: 128,
             lstm_num_layers: 1,
             lstm_sequence_length: 32,
-            accumulation_steps: 1,
+            accumulation_steps: params.accumulation_steps.max(1),
             clip_epsilon_high: None,
             mixed_precision: None, // Auto-detected at trainer initialization
         }

crates/ml/tests/ppo_recurrent_performance_tests.rs

@@ -100,6 +100,7 @@ fn create_test_hyperparams(_use_lstm: bool, sequence_length: usize) -> PpoHyperp
         gpu_timesteps_per_episode: 500,
         initial_capital: 1_000_000.0,
         avg_spread: 0.0001,
+        accumulation_steps: 1,
     }
 }