Sync from GitHub: f94f60c3463cc1866c793ac6b6c79e0d27e516e8

app.py

@@ -1286,6 +1286,20 @@ def on_interleaved_benchmark(
         raw_output = extraction_detail.get("raw_output", "")[:500]
         extractor_elapsed = extraction_detail.get("elapsed_s", 0.0)
 
+        # Qualitative content — pair each surfaced text with the
+        # category our centroid-similarity lookup tagged it as. Lets us
+        # eyeball "is this actually biology content for a biology query"
+        # without running another similarity pass at heatmap-build time.
+        # Truncate text to 200 chars so the JSON stays readable.
+        _surfaced_context = []
+        for _idx, _pid in enumerate(pith_ids):
+            _text = pith_context[_idx] if _idx < len(pith_context) else ""
+            _surfaced_context.append({
+                "id": _pid,
+                "category_tagged": _categorize_node(_pid),
+                "text": (_text[:200] + ("..." if len(_text) > 200 else "")),
+            })
+
         results.append({
             "turn": i + 1,
             "category": category,
@@ -1298,6 +1312,7 @@ def on_interleaved_benchmark(
             "deposit_node_id": deposit_nid,
             "ignition_size":   len(ignition_sets[i]),
             "pith_ids":        list(pith_ids),
+            "surfaced_context": _surfaced_context,
             "trees":           trees_for_turn,
             "raw_extractor_output": raw_output,
             "extractor_elapsed_s": extractor_elapsed,