fix(ci): stop flaky sleep test and compat evaluation errors

.github/workflows/integration-testing-regression.yml

@@ -151,6 +151,9 @@ jobs:
             extra-deps: ""
           - group: evaluation
             path: tests_integ/evaluation
+            # These tests need OpenAI and Braintrust API keys, which this job
+            # doesn't have. The secure integ workflow runs them.
+            extra-args: --ignore=tests_integ/evaluation/test_third_party_adapters.py
             timeout: 15
             extra-deps: "strands-agents-evals"
           - group: services
@@ -247,16 +250,17 @@ jobs:
           BASELINE_TAG: ${{ needs.resolve-tag.outputs.tag }}
           TEST_GROUP: ${{ matrix.group }}
           PYTEST_PATH: ${{ matrix.path }}
+          PYTEST_EXTRA_ARGS: ${{ matrix.extra-args || '' }}
         timeout-minutes: ${{ matrix.timeout }}
         working-directory: /tmp/baseline-tests
-        # PYTEST_PATH comes from the job matrix and holds space-separated pytest
-        # path args; left unquoted intentionally so it word-splits.
+        # PYTEST_PATH and PYTEST_EXTRA_ARGS are unquoted on purpose so they
+        # split into separate pytest args.
         run: |
           echo "================================================"
           echo "Library:    bedrock-agentcore (PR branch source)"
           echo "Test suite: $BASELINE_TAG tests_integ/$TEST_GROUP"
           echo "================================================"
-          pytest $PYTEST_PATH \
+          pytest $PYTEST_PATH $PYTEST_EXTRA_ARGS \
             -n auto --dist=loadscope \
             -s --log-cli-level=INFO
 

tests/conftest.py

@@ -0,0 +1,15 @@
+"""Shared pytest configuration."""
+
+import sys
+
+
+def pytest_runtest_teardown(item):
+    """Stop ragas' background telemetry thread once ragas is imported.
+
+    The thread calls ``time.sleep(1)`` in a loop. Patching ``time.sleep`` in a
+    test patches it for every thread, so the ragas thread's calls land on the
+    mock and break call-count assertions.
+    """
+    analytics = sys.modules.get("ragas._analytics")
+    if analytics is not None:
+        analytics._analytics_batcher._running = False