feat(intelligence): ADR-043 External Intelligence Providers for SONA Learning

Implement trait-based IntelligenceProvider extension point for external
quality signals. Addresses PR #190 proposal (renumbered from ADR-029 to
avoid collision with existing ADR-029-rvf-canonical-format).

- IntelligenceProvider trait with load_signals() and quality_weights()
- FileSignalProvider built-in for JSON file-based signal exchange
- IntelligenceLoader for multi-provider registration and aggregation
- QualitySignal, QualityFactors, ProviderQualityWeights types
- calibration_bias() on TaskComplexityAnalyzer for router feedback
- 12 unit tests (all passing)

Co-Authored-By: claude-flow <ruv@ruv.net>

Author: ruvnet

crates/ruvllm/src/claude_flow/model_router.rs

@@ -657,6 +657,29 @@ impl TaskComplexityAnalyzer {
         }
     }
 
+    /// Returns signed calibration error.
+    ///
+    /// Positive value = systematically over-predicting complexity,
+    /// negative value = systematically under-predicting.
+    /// Returns 0.0 if no feedback has been recorded.
+    pub fn calibration_bias(&self) -> f32 {
+        let with_feedback: Vec<_> = self
+            .accuracy_history
+            .iter()
+            .filter(|r| r.actual.is_some())
+            .collect();
+
+        if with_feedback.is_empty() {
+            return 0.0;
+        }
+
+        let sum: f32 = with_feedback
+            .iter()
+            .map(|r| r.predicted - r.actual.unwrap())
+            .sum();
+        sum / with_feedback.len() as f32
+    }
+
     /// Get accuracy statistics
     pub fn accuracy_stats(&self) -> AnalyzerStats {
         let with_feedback: Vec<_> = self