fix(llm): forward consolidation temperature

hindsight-api-slim/hindsight_api/engine/consolidation/consolidator.py

@@ -303,6 +303,7 @@ async def _dedup_adjudicate(
         await dedup_llm_config.call(
             messages=[{"role": "user", "content": _DEDUP_PROMPT.format(new=anchor_text, existing=best_text)}],
             response_format=_DedupDecision,
+            temperature=config.llm_temperature_consolidation,
             scope="consolidation_dedup",
             strict_schema=get_config().llm_strict_schema_consolidation,
         )
@@ -2845,6 +2846,7 @@ def _fact_line(m: dict[str, Any]) -> str:
                     {"role": "user", "content": user_content},
                 ],
                 "response_format": response_model,
+                "temperature": config.llm_temperature_consolidation,
                 "scope": "consolidation",
                 # Resolved per operation (HINDSIGHT_API_LLM_STRICT_SCHEMA_CONSOLIDATION, falling
                 # back to the global flag) so an operator can grammar-enforce consolidation's

hindsight-api-slim/tests/test_consolidation.py

@@ -2530,6 +2530,7 @@ async def test_unsupported_max_items_still_truncates_creates(
             consolidation_max_attempts=1,
             consolidation_llm_max_retries=None,
             consolidation_max_completion_tokens=None,
+            llm_temperature_consolidation=0.0,
         )
 
         result = await _consolidate_batch_with_llm(

hindsight-api-slim/tests/test_consolidation_dedup.py

@@ -188,6 +188,7 @@ def _ctx(threshold: float = 0.97):
         # config, so these must be present (production defaults: native/english).
         config=types.SimpleNamespace(
             consolidation_dedup_threshold=threshold,
+            llm_temperature_consolidation=0.0,
             text_search_extension="native",
             text_search_extension_native_language="english",
         ),
@@ -229,6 +230,7 @@ async def test_dedup_llm_keep_does_not_merge() -> None:
         result = await _dedup_reconcile_create(**kwargs)
     assert result is None
     llm.call.assert_awaited_once()
+    assert llm.call.await_args.kwargs["temperature"] == 0.0
     conn.fetchval.assert_not_called()  # kept distinct → no merge
 
 
@@ -389,6 +391,7 @@ def _update_ctx(threshold: float = 0.97):
         # config, so these must be present (production defaults: native/english).
         config=types.SimpleNamespace(
             consolidation_dedup_threshold=threshold,
+            llm_temperature_consolidation=0.0,
             text_search_extension="native",
             text_search_extension_native_language="english",
         ),

hindsight-api-slim/tests/test_consolidation_output_language.py

@@ -60,6 +60,7 @@ def _batch_config(llm_output_language: str | None) -> SimpleNamespace:
         consolidation_max_attempts=2,
         consolidation_llm_max_retries=None,
         consolidation_max_completion_tokens=None,
+        llm_temperature_consolidation=0.0,
     )
 
 

hindsight-api-slim/tests/test_consolidation_retry_budget.py

@@ -26,6 +26,7 @@ def mock_config():
     config.consolidation_llm_max_retries = None
     config.consolidation_max_completion_tokens = None
     config.llm_strict_schema_consolidation = False
+    config.llm_temperature_consolidation = 0.0
     return config
 
 
@@ -83,6 +84,19 @@ async def test_strict_schema_threaded_to_call(self, mock_llm_config, mock_config
         )
         assert mock_llm_config.call.call_args.kwargs.get("strict_schema") is True
 
+    @pytest.mark.asyncio
+    async def test_temperature_threaded_to_call(self, mock_llm_config, mock_config):
+        """llm_temperature_consolidation is passed to llm_config.call()."""
+        mock_config.llm_temperature_consolidation = 0.65
+        await _consolidate_batch_with_llm(
+            llm_config=mock_llm_config,
+            memories=[{"id": "m1", "text": "test"}],
+            union_observations=[],
+            union_source_facts={},
+            config=mock_config,
+        )
+        assert mock_llm_config.call.call_args.kwargs.get("temperature") == 0.65
+
     @pytest.mark.asyncio
     async def test_strict_schema_passed_as_explicit_false(self, mock_llm_config, mock_config):
         """A disabled per-operation flag is passed explicitly, not omitted.