Studio: merge grounded page excerpts with search snippets instead of replacing

When auto-scrape grounding retrieved page-body chunks, it replaced the raw
search-result text for that step. If the retrieved chunk was a distractor or
dropped the key fact, the answer-bearing search snippet was lost and grounded
runs regressed below snippet-only accuracy on factual questions (e.g. returning
Apache 2.0 instead of the Qwen License, 403 instead of 404, or a single mirror
diameter instead of the sum).

Keep the search snippets and append the grounded excerpts as supplementary
evidence via a small _merge_scraped_evidence helper. Grounding stays opt-in and
off by default, so legacy runs are unchanged. Adds regression tests.

studio/backend/core/research_runs.py

@@ -530,6 +530,25 @@ def _bounded_synthesis_evidence(
     return separator.join(bounded)[:max_chars]
 
 
+def _merge_scraped_evidence(raw_result: str, scraped_section: str) -> str:
+    """Combine the raw search snippets with grounded page-body chunks (additive).
+
+    Grounded auto-scrape used to REPLACE ``raw_result`` with ``scraped_section``.
+    When the retrieved chunk was a distractor or dropped the key fact, the
+    answer-bearing search snippet was lost and the grounded run regressed below
+    snippet-only accuracy. Keep the snippets first (they already carry the answer
+    for most factual queries) and append the grounded excerpts as supplementary
+    evidence. If either side is empty the other is returned unchanged.
+    """
+    raw = (raw_result or "").strip()
+    scraped = (scraped_section or "").strip()
+    if not scraped:
+        return raw_result
+    if not raw:
+        return scraped_section
+    return f"{raw}\n\nAdditional detail retrieved from the pages above:\n{scraped}"
+
+
 def _parse_json_object(text: str) -> dict:
     text = text.strip()
     if text.startswith("```"):
@@ -1842,9 +1861,10 @@ class ResearchSupervisor:
                 fetched_urls.update(scraped_urls)
                 await self._check_active(run["id"])
                 if scraped_section:
-                    # Replace raw search text with the retrieved chunks; sources are already
-                    # cataloged above, so nothing citable is lost.
-                    result = scraped_section
+                    # Additive merge (not replace): keep the answer-bearing search
+                    # snippets and append the grounded page-body chunks. See
+                    # _merge_scraped_evidence for why replacing regressed accuracy.
+                    result = _merge_scraped_evidence(result, scraped_section)
             note = (
                 f"### {action['title']} ({action['action']})\n"
                 f"Input: {argument}\nResult:\n{result[:12000]}\n\n"

studio/backend/tests/test_research_runs_storage.py

@@ -2790,3 +2790,28 @@ def test_route_accepts_max_tokens_without_treating_it_as_a_credential(research_h
     run = asyncio.run(create_research_run(payload, request, current_subject = "alice"))
 
     assert run["config"]["inferenceRequest"]["maxTokens"] == 1024
+
+
+def test_merge_scraped_evidence_keeps_snippet_and_chunk():
+    # Grounded auto-scrape must AUGMENT the raw search snippets, not replace them.
+    # Replacing dropped the answer-bearing snippet whenever the scraped chunk was a
+    # distractor, regressing grounded runs below snippet-only accuracy.
+    from core.research_runs import _merge_scraped_evidence
+
+    raw = "Qwen2.5-72B-Instruct is released under the Qwen License (see model card)."
+    scraped = "Most Qwen2.5 sizes such as 7B and 14B are licensed under Apache 2.0."
+    merged = _merge_scraped_evidence(raw, scraped)
+    # both the correct snippet and the grounded chunk survive
+    assert "Qwen License" in merged
+    assert "Apache 2.0" in merged
+    # snippet comes first so it is never truncated away by the evidence cap
+    assert merged.index("Qwen License") < merged.index("Apache 2.0")
+
+
+def test_merge_scraped_evidence_handles_empty_sides():
+    from core.research_runs import _merge_scraped_evidence
+
+    # no scraped chunk -> raw snippets returned unchanged (grounding produced nothing)
+    assert _merge_scraped_evidence("only snippets", "") == "only snippets"
+    # no raw snippets -> the scraped section is returned
+    assert _merge_scraped_evidence("", "only chunk") == "only chunk"