fix: widen E1 Q-correction clamp ±2.0 → ±10.0

Q-mean drifted to +4.6 but E1 correction was capped at -2.0.
The model needs full correction range to keep Q-values calibrated.

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

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

@@ -48,7 +48,7 @@ pub(crate) struct HindsightExperience {
 /// Output of one enrichment cycle.
 #[derive(Debug, Clone)]
 pub(crate) struct EnrichmentResult {
-    /// E1: additive correction to Q-targets (clipped ±2.0).
+    /// E1: additive correction to Q-targets (clipped ±10.0).
     pub q_correction: f32,
     /// E2: next-epoch epsilon.
     pub epsilon: f32,
@@ -148,7 +148,7 @@ fn compute_q_correction(trades: &[EvalTrade]) -> f32 {
         .map(|t| t.predicted_q - t.pnl)
         .sum::<f32>()
         / trades.len() as f32;
-    (-bias).clamp(-2.0, 2.0)
+    (-bias).clamp(-10.0, 10.0)
 }
 
 /// E2: Adaptive epsilon — performance-driven exploration rate.