fix: data quality query and suppression fallback in recommendation worker

- Fix _DATA_QUALITY_QUERY: remove nonexistent d.source_id/s.source_class,
  use d.source_type directly
- Fix LIMIT 1 applied after jsonb expansion by restructuring as CTE
- Fix fallback build_quality_context_from_summary returning empty
  source_types which always triggered LOW_SOURCE_DIVERSITY suppression
- Update test to reflect corrected fallback behavior

services/recommendation/suppression.py

@@ -107,7 +107,8 @@ def build_quality_context_from_summary(
 
     This is a fallback when full document-level quality metrics aren't
     available. It uses the trend summary's evidence counts and confidence
-    as proxies.
+    as proxies.  We assume at least one source type contributed so that
+    the fallback does not automatically trigger LOW_SOURCE_DIVERSITY.
     """
     total = len(summary.top_supporting_evidence) + len(summary.top_opposing_evidence)
     return DataQualityContext(
@@ -116,7 +117,7 @@ def build_quality_context_from_summary(
         failed_documents=0,
         avg_extraction_confidence=summary.confidence,
         newest_evidence_at=summary.generated_at,
-        source_types=set(),
+        source_types={"unknown"},
     )
 
 

services/recommendation/worker.py

@@ -58,26 +58,30 @@ logger = logging.getLogger(__name__)
 # ---------------------------------------------------------------------------
 
 _DATA_QUALITY_QUERY = """
+WITH latest_trend AS (
+    SELECT top_supporting_evidence, top_opposing_evidence
+    FROM trend_windows
+    WHERE entity_id = $1 AND "window" = $2
+    ORDER BY generated_at DESC
+    LIMIT 1
+),
+evidence_ids AS (
+    SELECT jsonb_array_elements_text(
+        COALESCE(lt.top_supporting_evidence, '[]'::jsonb)
+        || COALESCE(lt.top_opposing_evidence, '[]'::jsonb)
+    ) AS eid
+    FROM latest_trend lt
+)
 SELECT
     COUNT(*) AS total_documents,
     COUNT(*) FILTER (WHERE di.validation_status = 'valid') AS valid_documents,
     COUNT(*) FILTER (WHERE di.validation_status = 'failed') AS failed_documents,
     AVG(di.confidence) FILTER (WHERE di.validation_status = 'valid') AS avg_extraction_confidence,
     MAX(d.published_at) AS newest_evidence_at,
-    ARRAY_AGG(DISTINCT s.source_class) FILTER (WHERE s.source_class IS NOT NULL) AS source_types
+    ARRAY_AGG(DISTINCT d.source_type) FILTER (WHERE d.source_type IS NOT NULL) AS source_types
 FROM documents d
 JOIN document_intelligence di ON di.document_id = d.id
-LEFT JOIN sources s ON d.source_id = s.id
-WHERE d.id = ANY(
-    SELECT UNNEST(
-        COALESCE(tw.top_supporting_evidence, '[]'::jsonb)
-        || COALESCE(tw.top_opposing_evidence, '[]'::jsonb)
-    )::uuid
-    FROM trend_windows tw
-    WHERE tw.entity_id = $1 AND tw."window" = $2
-    ORDER BY tw.generated_at DESC
-    LIMIT 1
-)
+WHERE d.id::text IN (SELECT eid FROM evidence_ids)
 """
 
 

tests/test_suppression.py

@@ -128,12 +128,9 @@ def test_fallback_context_from_summary():
 def test_no_suppression_with_summary_fallback():
     """When no quality context is provided, summary-based fallback is used."""
     summary = _make_summary(confidence=0.7)
-    # Default config has min_source_types=1, but fallback has empty source_types.
-    # With min_source_types=1 and empty source_types, LOW_SOURCE_DIVERSITY fires
-    # only when total_documents > 0. But default min_source_types is 1 and
-    # len(set()) = 0 < 1, so it would fire. Let's use a config that relaxes this.
-    config = SuppressionConfig(min_source_types=0)
-    result = evaluate_suppression(summary, config=config, reference_time=NOW)
+    # Default config has min_source_types=1. The fallback now returns
+    # source_types={"unknown"} so LOW_SOURCE_DIVERSITY does not fire.
+    result = evaluate_suppression(summary, config=SuppressionConfig(), reference_time=NOW)
     assert result.suppressed is False