diff --git a/tests/integration/_search_helpers.py b/tests/integration/_search_helpers.py
index 198ceea9..b7414095 100644
--- a/tests/integration/_search_helpers.py
+++ b/tests/integration/_search_helpers.py
@@ -49,9 +49,10 @@ async def document_is_searchable(
             if r.get("id") == note_id:
                 if r.get("doc_type") == "note":
                     return True
-                # id matched but not a note — surface possible schema drift
-                # rather than silently timing out.
-                logger.debug(
+                # id matched but not a note — surface possible schema drift at
+                # WARNING (CI runs --log-cli-level=WARN) instead of letting the
+                # caller time out with a generic message.
+                logger.warning(
                     "search hit id=%s has doc_type=%s (expected note)",
                     note_id,
                     r.get("doc_type"),
diff --git a/tests/integration/test_astrolabe_session_jwt_search.py b/tests/integration/test_astrolabe_session_jwt_search.py
index 782a0b9b..f3b288b2 100644
--- a/tests/integration/test_astrolabe_session_jwt_search.py
+++ b/tests/integration/test_astrolabe_session_jwt_search.py
@@ -56,7 +56,8 @@ async def _get_with_retry(
             logger.warning(
                 "GET %s failed (attempt %s/%s): %s", url, attempt, max_attempts, e
             )
-            await anyio.sleep(2)
+            if attempt < max_attempts:
+                await anyio.sleep(2)  # no point sleeping before we give up
     assert last_exc is not None  # loop ran at least once, so this is set
     raise last_exc
 
diff --git a/tests/integration/test_rag.py b/tests/integration/test_rag.py
index aaf99fd8..d36a2d59 100644
--- a/tests/integration/test_rag.py
+++ b/tests/integration/test_rag.py
@@ -432,13 +432,16 @@ async def test_no_results_for_unrelated_query(nc_mcp_client, indexed_manual_pdf)
     query happened to retrieve any chunk at all). Comparing against a relevant
     query on the same corpus is self-calibrating and stable.
     """
-    unrelated = await _top_score(
-        nc_mcp_client, "quantum entanglement hadron collider particle physics"
+    # No results for the nonsense query is the ideal outcome — treat as score
+    # 0.0 and fall through, so the comparison (and the manual-is-indexed check
+    # below) still runs instead of the test silently skipping every time the
+    # physics query finds nothing.
+    unrelated = (
+        await _top_score(
+            nc_mcp_client, "quantum entanglement hadron collider particle physics"
+        )
+        or 0.0
     )
-    if unrelated is None:
-        # No results for the nonsense query is the ideal outcome; skip (rather
-        # than a silent pass) so the test report shows the path was taken.
-        pytest.skip("No results for nonsense physics query — the ideal outcome")
 
     relevant = await _top_score(
         nc_mcp_client, "how do I enable two-factor authentication"
diff --git a/tests/integration/test_sampling.py b/tests/integration/test_sampling.py
index 293e5dc7..78340996 100644
--- a/tests/integration/test_sampling.py
+++ b/tests/integration/test_sampling.py
@@ -274,8 +274,12 @@ async def test_semantic_search_answer_with_limit(nc_mcp_client, temporary_note_f
         category="Development",
     )
 
-    # Wait for vector indexing to complete
-    await wait_for_vector_sync(nc_mcp_client)
+    # Wait until the batch is indexed — gate on the last note being searchable
+    # rather than a bare idle signal, which can fire before the new notes are
+    # even enqueued.
+    await wait_for_vector_sync(
+        nc_mcp_client, search_term="async context managers", note_id=_note3["id"]
+    )
 
     call_result = await nc_mcp_client.call_tool(
         "nc_semantic_search_answer",
@@ -315,8 +319,10 @@ async def test_semantic_search_answer_score_threshold(
         category="Test",
     )
 
-    # Wait for vector indexing to complete
-    await wait_for_vector_sync(nc_mcp_client)
+    # Gate on the new note being searchable (not a bare idle signal).
+    await wait_for_vector_sync(
+        nc_mcp_client, search_term="widget manufacturing", note_id=_note["id"]
+    )
 
     # Query with exact match
     call_result = await nc_mcp_client.call_tool(
@@ -362,8 +368,10 @@ async def test_semantic_search_answer_max_tokens(nc_mcp_client, temporary_note_f
         category="Test",
     )
 
-    # Wait for vector indexing to complete
-    await wait_for_vector_sync(nc_mcp_client)
+    # Gate on the new note being searchable (not a bare idle signal).
+    await wait_for_vector_sync(
+        nc_mcp_client, search_term="Long Document content", note_id=_note["id"]
+    )
 
     call_result = await nc_mcp_client.call_tool(
         "nc_semantic_search_answer",