test: handle model instances in fixture discriminators

tests/conformance/harness/expectations.py

@@ -224,7 +224,21 @@ def _discriminate_expected(
     blocks themselves don't know that, so we discriminate on the keys
     that ARE distinctive and fall back to graph-engine for plain
     ``final_state``-only fixtures.
+
+    Pydantic invokes this callable on both the validation and
+    serialization paths. On serialization the value is already one of the
+    concrete variants, so route by ``isinstance`` first; otherwise the
+    dump path falls through to ``graph_engine`` and warns that the
+    serialized variant doesn't match the union's first arm.
     """
+    if isinstance(value, GraphEngineExpected):
+        return "graph_engine"
+    if isinstance(value, LlmProviderExpected):
+        return "llm_provider"
+    if isinstance(value, PipelineUtilitiesExpected):
+        return "pipeline_utilities"
+    if isinstance(value, ObservabilityExpected):
+        return "observability"
     if not isinstance(value, dict):
         return "graph_engine"
     keys: set[str] = {str(k) for k in cast("dict[str, Any]", value)}

tests/conformance/harness/fixtures.py

@@ -241,7 +241,17 @@ def _discriminate_fixture(value: Any) -> Literal["llm_provider", "cases", "graph
     llm-provider fixtures (e.g. 003-message-validation) have BOTH —
     ``mock_provider`` is the load-bearing discriminator, ``cases`` is just
     the table style for sub-cases.
+
+    Also handle the serialization path (where the value is a concrete
+    variant) so a future ``model_dump`` through the top-level union
+    doesn't fall through to ``graph`` and warn.
     """
+    if isinstance(value, LlmProviderFixture):
+        return "llm_provider"
+    if isinstance(value, CasesFixture):
+        return "cases"
+    if isinstance(value, GraphFixture):
+        return "graph"
     if isinstance(value, dict):
         if "mock_provider" in value:
             return "llm_provider"