test: fix CI regressions (Vertex/litellmrouter construction, dedup config, trace recorder leak)

hindsight-api-slim/tests/conftest.py

@@ -54,6 +54,29 @@ async def _teardown_memory_engine(mem: MemoryEngine) -> None:
         unregister_span_recorder(mem._llm_recorder)
 
 
+@pytest.fixture(autouse=True)
+def _cleanup_leaked_span_recorders():
+    """Fail-safe for the process-global LLM-trace recorder registry (#2229).
+
+    ``MemoryEngine.__init__`` registers its recorder in the shared registry, and
+    only ``close()`` removes it. Tests that construct an engine directly (without
+    ``_teardown_memory_engine``/``close()``) leak an *enabled* recorder; a later
+    test's LLM calls then get recorded into the shared DB, flaking
+    ``test_llm_trace::test_disabled_writes_no_rows`` (it observes rows for its
+    bank even though its own recorder is disabled). ``_teardown_memory_engine``
+    guards the fixtures; this guards everything else by dropping any recorder a
+    test added to the registry.
+    """
+    from hindsight_api.tracing import get_span_recorder
+
+    recorders = get_span_recorder()._recorders
+    before = {id(r) for r in recorders}
+    yield
+    for recorder in list(recorders):
+        if id(recorder) not in before:
+            recorders.remove(recorder)
+
+
 # Default pg0 instance configuration for tests
 DEFAULT_PG0_INSTANCE_NAME = "hindsight-test"
 DEFAULT_PG0_PORT = int(os.environ.get("HINDSIGHT_TEST_PG_PORT", "5556"))

hindsight-api-slim/tests/test_consolidation_dedup.py

@@ -84,7 +84,13 @@ def _ctx(threshold: float = 0.97):
         conn=conn,
         memory_engine=types.SimpleNamespace(embeddings=object()),
         bank_id="bank1",
-        config=types.SimpleNamespace(consolidation_dedup_threshold=threshold),
+        # The merge path builds a search_vector UPDATE clause from the text-search
+        # config, so these must be present (production defaults: native/english).
+        config=types.SimpleNamespace(
+            consolidation_dedup_threshold=threshold,
+            text_search_extension="native",
+            text_search_extension_native_language="english",
+        ),
         dedup_llm_config=llm,
         create_text="YouTube content in Uzbek is very rich.",
         create_source_ids=[uuid.uuid4()],
@@ -179,7 +185,13 @@ def _update_ctx(threshold: float = 0.97):
         conn=conn,
         memory_engine=types.SimpleNamespace(embeddings=object()),
         bank_id="bank1",
-        config=types.SimpleNamespace(consolidation_dedup_threshold=threshold),
+        # The merge path builds a search_vector UPDATE clause from the text-search
+        # config, so these must be present (production defaults: native/english).
+        config=types.SimpleNamespace(
+            consolidation_dedup_threshold=threshold,
+            text_search_extension="native",
+            text_search_extension_native_language="english",
+        ),
         dedup_llm_config=llm,
         updated_id=_UPDATED_ID,
         updated_text="Uzbek content on YouTube is very rich and growing.",

hindsight-api-slim/tests/test_fact_extraction_analysis.py

@@ -27,6 +27,13 @@ def llm_config():
         api_key=config.retain_llm_api_key or config.llm_api_key,
         model=config.retain_llm_model or config.llm_model,
         base_url=config.retain_llm_base_url or config.llm_base_url,
+        # LLMConfig uses these as-passed and no longer reads them from global config,
+        # so the caller must forward the Vertex AI settings (mirrors MemoryEngine's
+        # own LLMConfig construction). Without this, provider=vertexai raises
+        # "HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID is required" even when it is set.
+        vertexai_project_id=config.llm_vertexai_project_id,
+        vertexai_region=config.llm_vertexai_region,
+        vertexai_service_account_key=config.llm_vertexai_service_account_key,
     )
 
 

hindsight-api-slim/tests/test_llm_provider.py

@@ -45,11 +45,22 @@ def _get_api_key() -> str:
 
 
 def _make_llm() -> LLMProvider:
+    # LLMProvider uses provider-specific settings as-passed (it does not resolve
+    # them from global config), so forward the ones whose providers require them:
+    # Vertex AI needs project/region, and litellmrouter needs its router config.
+    # Without these, provider=vertexai/litellmrouter raise at construction.
+    from hindsight_api.config import get_config
+
+    config = get_config()
     return LLMProvider(
         provider=_PROVIDER,
         api_key=_get_api_key(),
         base_url=os.environ.get("HINDSIGHT_API_LLM_BASE_URL", ""),
         model=_MODEL,
+        vertexai_project_id=config.llm_vertexai_project_id,
+        vertexai_region=config.llm_vertexai_region,
+        vertexai_service_account_key=config.llm_vertexai_service_account_key,
+        litellmrouter_config=config.llm_litellmrouter_config,
     )