fix: E1 Q-value reality check uses actual avg_q_value not avg_pnl

E1 enrichment was computing q_corr = mean(predicted_q - pnl) but
predicted_q was set to avg_pnl → bias always ~0.  Now uses actual
avg_q_value from the training step, producing meaningful corrections
when Q-values drift away from realized returns.

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

crates/ml/src/trainers/dqn/trainer/enrichment.rs

@@ -101,6 +101,7 @@ pub(crate) fn extract_eval_trades_from_metrics(
     total_trades: f32,
     total_pnl: f32,
     win_rate: f32,
+    avg_q_value: f32,
     n_bars: usize,
 ) -> Vec<EvalTrade> {
     let n_trades = total_trades as usize;
@@ -126,7 +127,7 @@ pub(crate) fn extract_eval_trades_from_metrics(
             magnitude: 1,
             holding_bars: avg_holding,
             pnl,
-            predicted_q: avg_pnl,
+            predicted_q: avg_q_value,
             ensemble_var: 0.5,
         });
     }

crates/ml/src/trainers/dqn/trainer/training_loop.rs

@@ -633,6 +633,7 @@ impl DQNTrainer {
                     60000.0,  // approximate trade count
                     0.0,      // approximate P&L
                     0.5,      // approximate win rate
+                    log_output.avg_q_value as f32,  // actual Q-mean from training step
                     val_bars,
                 );