feat: add XAI inference for explainable predictions in analysis results

app/routes/analyze.py

@@ -434,6 +434,56 @@ async def _analyze_file(
         f"confidence={fusion.confidence}, source={fusion.decision_source}"
     )
 
+    # ── Step 5b: XAI inference (explainable) ────
+    xai_payload: Optional[XAIExplanation] = None
+    try:
+        if xai_service.available:
+            xai_result = xai_service.predict(features, vocals)
+            if xai_result is not None:
+                xai_dict = xai_service.to_dict(xai_result)
+                # Convert dicts to pydantic models
+                xai_payload = XAIExplanation(
+                    probability=xai_dict["probability"],
+                    threshold=xai_dict["threshold"],
+                    baseProbability=xai_dict["baseProbability"],
+                    confidenceBand=ConfidenceBandModel(**xai_dict["confidenceBand"]),
+                    modelVotes=[ModelVoteModel(**v) for v in xai_dict["modelVotes"]],
+                    bestModel=xai_dict["bestModel"],
+                    topContributions=[
+                        FeatureContributionModel(**c)
+                        for c in xai_dict["topContributions"]
+                    ],
+                    allFeatures={
+                        k: FeatureContributionModel(**v)
+                        for k, v in xai_dict["allFeatures"].items()
+                    },
+                    featureCount=xai_dict["featureCount"],
+                )
+                logger.info(
+                    f"[{request_id}] XAI: p={xai_result.probability:.3f}, "
+                    f"band={xai_result.confidence_band.tier}, "
+                    f"top={xai_result.top_contributions[0].name if xai_result.top_contributions else 'n/a'}"
+                )
+    except Exception as e:
+        logger.warning(f"[{request_id}] XAI failed: {e}")
+        warnings.append("xai_unavailable")
+
+    # Build tower scores dict for UI (multi-signal breakdown)
+    tower_scores: Dict[str, float] = {
+        "local_features": round(
+            (features.spectral_regularity + features.temporal_patterns
+             + features.harmonic_structure) / 3.0, 3,
+        ),
+    }
+    if vocals and vocals.has_vocals:
+        tower_scores["vocals"] = round(vocals.vocal_ai_score, 3)
+    if clap_result and clap_result.available:
+        tower_scores["clap"] = round(clap_result.confidence, 3)
+    if fst_result and fst_result.available:
+        tower_scores["fst"] = round(fst_result.confidence, 3)
+    if xai_payload is not None:
+        tower_scores["xai_ensemble"] = round(xai_payload.probability, 3)
+
     # ── Step 4: Build response ─────────────
     vocal_response = None
     if vocals and vocals.has_vocals: