refactor(knowledge): build the recipe journal and KB writes from named parts

scripts/code_metrics_baseline.txt

@@ -30,12 +30,6 @@ cognitive-complexity src/hyperloom/orchestrator/kernel/conc_sweep.py::_sweep_one
 cognitive-complexity src/hyperloom/orchestrator/kernel/conc_sweep.py::run_conc_sweep 60
 cognitive-complexity src/hyperloom/orchestrator/kernel/forge_fusion.py::_normalize_manifest 44
 cognitive-complexity src/hyperloom/orchestrator/kernel/request_handlers.py::_resolve_forge_server_log 49
-cognitive-complexity src/hyperloom/orchestrator/knowledge/recipe_journal.py::RecipeJournalCollaborator._build_recipe_attrs_from_state 41
-cognitive-complexity src/hyperloom/orchestrator/knowledge/recipe_journal.py::RecipeJournalCollaborator._collect_attempt_provenance 34
-cognitive-complexity src/hyperloom/orchestrator/knowledge/recipe_journal.py::RecipeJournalCollaborator._collect_workload_tags 37
-cognitive-complexity src/hyperloom/orchestrator/knowledge/recipe_journal.py::RecipeJournalCollaborator.finalize_recipe_and_journal 48
-cognitive-complexity src/hyperloom/orchestrator/knowledge/recipe_journal.py::RecipeJournalCollaborator.kb_amend_recipe 49
-cognitive-complexity src/hyperloom/orchestrator/knowledge/recipe_journal.py::_record_config_attempts 36
 cognitive-complexity src/hyperloom/orchestrator/loop/writeback.py::WritebackCollaborator._drain_agent_keep_outbox 34
 cognitive-complexity src/hyperloom/orchestrator/loop/writeback.py::WritebackCollaborator._keep_patch_sources 42
 cognitive-complexity src/hyperloom/orchestrator/loop/writeback.py::WritebackCollaborator._promote_baseline 96
@@ -348,11 +342,6 @@ function-lines src/hyperloom/orchestrator/kernel/request_handlers.py::_run_forge
 function-lines src/hyperloom/orchestrator/kernel/request_handlers.py::_run_forge_gemm_tuning 117
 function-lines src/hyperloom/orchestrator/kernel/request_handlers.py::_run_integrate_rebaseline_with_lock_retry 106
 function-lines src/hyperloom/orchestrator/kernel/request_handlers.py::integrate_handler 157
-function-lines src/hyperloom/orchestrator/knowledge/recipe_journal.py::RecipeJournalCollaborator._build_recipe_attrs_from_state 99
-function-lines src/hyperloom/orchestrator/knowledge/recipe_journal.py::RecipeJournalCollaborator._record_fact_per_variant 114
-function-lines src/hyperloom/orchestrator/knowledge/recipe_journal.py::RecipeJournalCollaborator.finalize_recipe_and_journal 211
-function-lines src/hyperloom/orchestrator/knowledge/recipe_journal.py::RecipeJournalCollaborator.kb_amend_recipe 127
-function-lines src/hyperloom/orchestrator/knowledge/recipe_journal.py::_record_config_attempts 136
 function-lines src/hyperloom/orchestrator/knowledge/recipe_kb/local_store.py::LocalRecipeStore.put_recipe 102
 function-lines src/hyperloom/orchestrator/knowledge/remote_recipe/client.py::RemoteRecipeClient.read 95
 function-lines src/hyperloom/orchestrator/knowledge/remote_recipe/client.py::RemoteRecipeClient.write_if_better 82
@@ -491,7 +480,7 @@ module-lines src/hyperloom/orchestrator/actions/executors/roofline.py 1303
 module-lines src/hyperloom/orchestrator/kernel/apply_kernel_patch.py 2656
 module-lines src/hyperloom/orchestrator/kernel/conc_sweep.py 1429
 module-lines src/hyperloom/orchestrator/kernel/request_handlers.py 4690
-module-lines src/hyperloom/orchestrator/knowledge/recipe_journal.py 1671
+module-lines src/hyperloom/orchestrator/knowledge/recipe_journal.py 1269
 module-lines src/hyperloom/orchestrator/loop/coordinator.py 1337
 module-lines src/hyperloom/orchestrator/loop/dispatcher.py 1453
 module-lines src/hyperloom/orchestrator/loop/intent_router.py 1409

src/hyperloom/common/coerce.py

@@ -24,6 +24,18 @@ def to_float(value: Any, default: _T | None = None) -> float | _T | None:
     return result if math.isfinite(result) else default
 
 
+def recorded_float(value: Any, default: float = 0.0) -> float:
+    """Coerce a recorded value to ``float``, keeping ``inf``, ``nan`` and ``bool`` as they were recorded.
+
+    :func:`to_float` returns *default* for those three instead, which would
+    render a degenerate figure as one nobody can tell from a real measurement.
+    """
+    try:
+        return float(value or default)
+    except (TypeError, ValueError):
+        return default
+
+
 def to_int(value: Any, default: _T | None = None) -> int | _T | None:
     """Coerce *value* to ``int``, rejecting ``bool``, ``None``, and non-finite floats."""
     if value is None or isinstance(value, bool):
@@ -116,6 +128,7 @@ def to_list(value: Any) -> list[Any]:
 
 __all__ = [
     "to_float",
+    "recorded_float",
     "to_int",
     "first_float",
     "first_int",

src/hyperloom/inference_optimizer/tests/test_coerce.py

@@ -13,6 +13,7 @@
     first_int,
     first_str,
     optional_positive_int,
+    recorded_float,
     to_float,
     to_int,
     to_list,
@@ -69,6 +70,36 @@ def test_custom_default(self):
         assert to_float(None, default=-1.0) == -1.0
 
 
+class TestRecordedFloat:
+    @pytest.mark.parametrize(
+        "value,expected",
+        [
+            (1, 1.0),
+            (1.5, 1.5),
+            ("2.5", 2.5),
+            (-4, -4.0),
+        ],
+    )
+    def test_parses(self, value, expected):
+        assert recorded_float(value) == expected
+
+    @pytest.mark.parametrize("value", [None, "", "abc", object(), [1]])
+    def test_absent_or_unparseable_reads_as_the_default(self, value):
+        assert recorded_float(value) == 0.0
+        assert recorded_float(value, -1.0) == -1.0
+
+    def test_infinity_survives_so_a_degenerate_measurement_stays_visible(self):
+        assert recorded_float(float("inf")) == math.inf
+        assert recorded_float("Infinity") == math.inf
+        assert math.isnan(recorded_float(float("nan")))
+
+    @pytest.mark.parametrize("value", [float("inf"), float("nan"), True])
+    def test_keeps_what_to_float_rejects(self, value):
+        """The three inputs on which the two coercions disagree, so a swap is never silent."""
+        assert to_float(value, 0.0) == 0.0
+        assert recorded_float(value) != 0.0 or math.isnan(recorded_float(value))
+
+
 class TestToInt:
     @pytest.mark.parametrize(
         "value,expected",

src/hyperloom/inference_optimizer/tests/test_coordinator_kb_writes.py

@@ -25,6 +25,9 @@
     recipe_canonical_id,
 )
 from hyperloom.orchestrator.knowledge.config import KnowledgeConfig
+from hyperloom.orchestrator.knowledge.recipe_journal import (
+    recipe_attrs_from_state,
+)
 from hyperloom.inference_optimizer.breakdown.recorder import close_out as _close_out
 
 
@@ -433,7 +436,7 @@ def test_best_config_reads_stack_args_from_canonical_server_key(
         }
     ]
     ss.cumulative_gain_validated = 10.0
-    attrs = coord.recipe_journal._build_recipe_attrs_from_state()
+    attrs = recipe_attrs_from_state(coord.shared_state, coord.session_dir.name)
     assert attrs["best_config"]["extra_server_args"] == ("--page-size 32 --schedule-policy lpm"), (
         "stack-layer launch args must be read from the canonical "
         "extra_server_args key; got "
@@ -557,7 +560,7 @@ def test_session_entry_carries_throughput_date_and_actions(
     ]
     ss.cumulative_gain_validated = 7.5
     ss.cumulative_gain_validated_stack_len = 2
-    attrs = coord.recipe_journal._build_recipe_attrs_from_state()
+    attrs = recipe_attrs_from_state(coord.shared_state, coord.session_dir.name)
     sessions = attrs.get("sessions") or []
     assert sessions, "no session entry emitted"
     s = sessions[0]

src/hyperloom/inference_optimizer/tests/test_coordinator_sync_helpers_coverage_unit.py

@@ -15,6 +15,9 @@
 from hyperloom.orchestrator.loop.coordinator import Coordinator
 from hyperloom.orchestrator.state.objective import TargetGainObjective
 from hyperloom.inference_optimizer.protocol.intent import Intent, IntentType
+from hyperloom.orchestrator.knowledge.recipe_journal import (
+    workload_kb_tags,
+)
 from hyperloom.orchestrator.loop.dispatcher import HeldLeases
 
 
@@ -560,7 +563,7 @@ def test_collect_workload_tags(coord: Coordinator, monkeypatch) -> None:
         "workload_mode": "serving",
     }
 
-    assert coord.recipe_journal._collect_workload_tags() == {
+    assert workload_kb_tags(coord.shared_state) == {
         "framework": "sglang",
         "model_class": "moe",
         "model_name": "Qwen3-32B",
@@ -591,19 +594,19 @@ def test_workload_tags_read_ep_from_the_env_only_without_a_shared_state_value(co
     ss = coord.shared_state
     ss.ep = 8
 
-    tags = coord.recipe_journal._collect_workload_tags()
+    tags = workload_kb_tags(coord.shared_state)
     assert tags["ep"] == 8
     assert tags["pp"] == 3
 
     ss.ep = 0
-    assert coord.recipe_journal._collect_workload_tags()["ep"] == 2
+    assert workload_kb_tags(coord.shared_state)["ep"] == 2
 
 
 def test_an_unparseable_or_non_positive_parallelism_env_is_not_tagged(coord: Coordinator, monkeypatch) -> None:
     monkeypatch.setenv("EP", "two")
     monkeypatch.setenv("PP", "0")
 
-    tags = coord.recipe_journal._collect_workload_tags()
+    tags = workload_kb_tags(coord.shared_state)
     assert "ep" not in tags
     assert "pp" not in tags
 
@@ -618,7 +621,7 @@ def test_workload_tags_survive_a_resumed_null_fingerprint_meta(coord: Coordinato
     ss.stack_fingerprint_meta = None
     ss.baseline_workload_extra = None
 
-    assert coord.recipe_journal._collect_workload_tags() == {"framework": "vllm"}
+    assert workload_kb_tags(coord.shared_state) == {"framework": "vllm"}
 
 
 def test_unknown_and_empty_fingerprint_entries_are_not_tagged(coord: Coordinator, monkeypatch) -> None:
@@ -634,7 +637,7 @@ def test_unknown_and_empty_fingerprint_entries_are_not_tagged(coord: Coordinator
         "image_digest": "sha256:abc",
     }
 
-    tags = coord.recipe_journal._collect_workload_tags()
+    tags = workload_kb_tags(coord.shared_state)
     assert tags == {"framework": "vllm", "image_digest": "sha256:abc"}
 
 

src/hyperloom/inference_optimizer/tests/test_recipe_sedimentation.py

@@ -17,6 +17,10 @@
 )
 from hyperloom.orchestrator.loop.coordinator import AdvisoryServices, Coordinator
 from hyperloom.orchestrator.knowledge.recipe_kb import LocalRecipeStore, RecipeKB
+from hyperloom.orchestrator.knowledge.recipe_journal import (
+    attempt_provenance,
+    recipe_attrs_from_state,
+)
 
 
 def _make_coordinator(tmp_path: Path) -> Coordinator:
@@ -67,7 +71,7 @@ def test_collect_attempt_provenance_maps_keep_and_revert(tmp_path):
         },
     )
 
-    kept, kept_by_gap, reverted = coord.recipe_journal._collect_attempt_provenance()
+    kept, kept_by_gap, reverted = attempt_provenance(coord.shared_state)
     assert kept == {"mtp_on": "https://pr/123"}
     assert kept_by_gap == {"gap.research_hint.0": "https://pr/123"}
     assert len(reverted) == 1
@@ -106,7 +110,7 @@ def test_provenance_resolves_by_gap_id_when_name_mismatches(tmp_path):
         },
     )
 
-    attrs = coord.recipe_journal._build_recipe_attrs_from_state()
+    attrs = recipe_attrs_from_state(coord.shared_state, coord.session_dir.name)
     row = next(x for x in attrs["what_worked"] if x["name"] == "k007")
     assert row["source"] == "https://pr/777"
 
@@ -142,7 +146,7 @@ def test_build_recipe_attrs_sediments_source_and_revert(tmp_path):
         },
     )
 
-    attrs = coord.recipe_journal._build_recipe_attrs_from_state()
+    attrs = recipe_attrs_from_state(coord.shared_state, coord.session_dir.name)
     ww = attrs["what_worked"]
     mtp = next(x for x in ww if x["name"] == "mtp_on")
     assert mtp["source"] == "https://pr/123"
@@ -174,7 +178,7 @@ def test_sediment_toggle_off_keeps_recipe_ephemeral(tmp_path):
         },
     )
 
-    attrs = coord.recipe_journal._build_recipe_attrs_from_state()
+    attrs = recipe_attrs_from_state(coord.shared_state, coord.session_dir.name)
     mtp = next(x for x in attrs["what_worked"] if x["name"] == "mtp_on")
     assert "source" not in mtp
 
@@ -219,7 +223,7 @@ def test_experience_rows_survive_the_kb_round_trip(tmp_path):
         model="m",
         hardware="h",
         framework_name="f",
-        # Exactly what recipe_journal._build_recipe_attrs_from_state emits.
+        # Exactly what recipe_journal.recipe_attrs_from_state emits.
         what_worked=[
             {
                 "name": "mtp_on",

src/hyperloom/inference_optimizer/tests/test_remote_recipe_v2.py

@@ -3126,7 +3126,7 @@ def test_remote_read_is_wired_into_cli_t0_and_close_write_remains() -> None:
             fromlist=["_bootstrap_recipe_kb"],
         )._bootstrap_recipe_kb
     )
-    close_source = inspect.getsource(RecipeJournalCollaborator.finalize_recipe_and_journal)
+    close_source = inspect.getsource(RecipeJournalCollaborator._publish_remote_recipe)
     assert "RemoteWarmRecipeAdapter" in bootstrap_source
     assert "HyperloomRemoteKB.from_env().write" in close_source
     assert "write_final_remote_recipe" not in close_source