NZ: pin the Axiom engine at 753ce7d9 and close #1563's follow-ups (G8-x2)

packages/microcosm-build/src/microcosm/build/nz/axiom_input_closure.json

@@ -14,7 +14,7 @@
   "rulespec_commit": "6fe181fc4c65763150fe97501ef3443774e64869",
   "engine": {
    "repository": "TheAxiomFoundation/axiom-rules-engine",
-   "commit": "04315d94d04efd61e27665ab3aa300d64fdf3a66"
+   "commit": "753ce7d9ac4cfb161f6aa4f281f5a3980e67a57d"
   }
  },
  "rulespec_pin": {
@@ -47,28 +47,28 @@
    "A main benefit only where the rates module's income-tested net weekly payment for it is positive: entitlement treats the s 24 income test and the s 21 work gap as met (jobseeker_income_less_than_zero_rate_cutout true, jobseeker_in_full_time_employment false), and the Sole Parent Support and Supported Living Payment entitlement rules have no income term; the rates module tests the person's own income only."
   ]
  },
- "engine_optional_evidence": "The axiom-rules-engine commit named in input_surface.engine (the engine that generated the committed input surface) has no optional inputs. A RuleSpec module has no inputs section (RulesDocument holds format, imports, module, units, relations and rules; src/rulespec.rs:361-374), a rule-level default is rejected (RuleSpecError::UnsupportedDefault, src/rulespec.rs:91-94), and the dense batch call collects only the inputs the caller supplies (python-ext/src/lib.rs:458-466). A plain read of an absent input fails every row whose evaluation reaches it with EvalError::MissingInput (src/dense.rs:2790-2803), and that fails the request. The internal input_or_else node that would read a default (src/dense.rs:2805-2811) is never produced from RuleSpec text: formula lowering turns a name no rule defines into a plain input read (src/formula.rs:1400-1413). So no root input of these modules can be verified engine-optional, and this closure classes none as such: an input a run does not encode is defaulted here, with its reason.",
+ "engine_optional_evidence": "The axiom-rules-engine commit named in input_surface.engine (the engine that generated the committed input surface) gives these modules no optional inputs. A RuleSpec inputs block can't declare a default: the engine keeps it only to check only_related and selected-year reads and to read elections, and every other key stays inert (RulesDocument, src/rulespec.rs:663-687; declared_elections, src/rulespec.rs:1569-1611). No module the committed input surface covers has an inputs block, at the surface's RuleSpec commit or at the pin. The engine rejects a rule-level default (RuleSpecError::UnsupportedDefault, src/rulespec.rs:274-277), and the dense batch call collects only the inputs the caller supplies (python-ext/src/lib.rs:552-563). A plain read of an absent input fails every row whose evaluation reaches it with EvalError::MissingInput (src/dense.rs:5154-5168), and the first row error fails the whole request (src/dense.rs:1539, python-ext/src/lib.rs:404-409). The internal input_or_else node that would read a default (src/dense.rs:5169-5175) never comes from RuleSpec text: formula lowering turns a name that no parameter or rule defines into a plain input read (src/formula.rs:1833-1851). So no root input of these modules can be verified engine-optional, and this closure classes none as such: the closure defaults, with its reason, any input a run doesn't encode.",
  "engine_source_pins": {
   "note": "The axiom-rules-engine source files that engine_optional_evidence and undetermined_judgments cite by line, pinned at input_surface.engine.commit so each line citation names one exact file version.",
   "pins": [
    {
     "repository": "TheAxiomFoundation/axiom-rules-engine",
-    "commit": "04315d94d04efd61e27665ab3aa300d64fdf3a66",
+    "commit": "753ce7d9ac4cfb161f6aa4f281f5a3980e67a57d",
     "path": "python-ext/src/lib.rs"
    },
    {
     "repository": "TheAxiomFoundation/axiom-rules-engine",
-    "commit": "04315d94d04efd61e27665ab3aa300d64fdf3a66",
+    "commit": "753ce7d9ac4cfb161f6aa4f281f5a3980e67a57d",
     "path": "src/dense.rs"
    },
    {
     "repository": "TheAxiomFoundation/axiom-rules-engine",
-    "commit": "04315d94d04efd61e27665ab3aa300d64fdf3a66",
+    "commit": "753ce7d9ac4cfb161f6aa4f281f5a3980e67a57d",
     "path": "src/formula.rs"
    },
    {
     "repository": "TheAxiomFoundation/axiom-rules-engine",
-    "commit": "04315d94d04efd61e27665ab3aa300d64fdf3a66",
+    "commit": "753ce7d9ac4cfb161f6aa4f281f5a3980e67a57d",
     "path": "src/rulespec.rs"
    }
   ]
@@ -80,7 +80,7 @@
    "eligible"
   ],
   "new_default": true,
-  "reason": "Entitlement judgments come back as codes 1 (holds), -1 (does not hold) and 0 (undetermined). At the axiom-rules-engine commit named in input_surface.engine, no dense leaf yields an undetermined value: dense input columns are booleans, integers, floats or text, an absent input is an error rather than an unknown, and and/or/not only pass an undetermined item along. With every root input encoded or defaulted by this closure, a 0 cannot arise, so the receipt kernel takeup.assign@1, once implemented, is to refuse any 0 it sees instead of guessing: a 0 would mean the engine or the closure changed. The alternatives a review could choose are treating 0 as not entitled (understates receipt) or as entitled (overstates it). microcosm.frame.input_closure.resolve_judgments implements all three."
+  "reason": "Entitlement judgments come back as codes 1 (holds), -1 (does not hold) and 0 (undetermined). At the axiom-rules-engine commit named in input_surface.engine, no dense leaf yields an undetermined value: dense input columns hold booleans, integers, decimals, floats, text or dates, an absent input is an error rather than an unknown, and and/or/not only pass an undetermined item along. With every root input encoded or defaulted by this closure, a 0 cannot arise, so the receipt kernel takeup.assign@1, once implemented, is to refuse any 0 it sees instead of guessing: a 0 would mean the engine or the closure changed. The alternatives a review could choose are treating 0 as not entitled (understates receipt) or as entitled (overstates it). microcosm.frame.input_closure.resolve_judgments implements all three."
  },
  "knobs": [
   {

packages/microcosm-build/src/microcosm/build/nz/axiom_rules_bindings.json

@@ -1,7 +1,7 @@
 {
   "version": 1,
   "country": "nz",
-  "description": "Axiom rules bindings for the New Zealand v0 graph build. Each binding names one rulespec-nz module by repository-relative rulespec_path and SHA-256 at the pinned commit, the Frame entity it runs on, the period label it is evaluated for, and the variables the graph requests. No code interprets these bindings yet. The planned shared composer (package G6) will build one AxiomEngine per binding and register every binding in one simulate.rules_by_ref@1 kernel (to be added by package G1), because a run's kernel registry holds one kernel per ref. The Axiom adapter raises when a module declares relations; every bound module declares none at the pin. No kernel implementation hash is to bind the whole NZ spec.",
+  "description": "Axiom rules bindings for the New Zealand v0 graph build. Each binding names one rulespec-nz module by repository-relative rulespec_path and SHA-256 at the pinned commit, the Frame entity it runs on, the period label it runs for, and the variables the graph requests. build_transport_registry (transport/registry.py) builds one AxiomEngine per binding, and one simulate.rules_by_ref@1 router serves them all, because a run's kernel registry holds one kernel per ref. A binding that declares no relations goes through assert_no_relations, which refuses a module whose dense program on the binding's entity declares relations; every bound module declares none at the pin. No kernel implementation hash is to bind the whole NZ spec.",
   "rulespec": {
     "repository": "TheAxiomFoundation/rulespec-nz",
     "commit": "8dc2507d5f10a0467c285fed54d018ee37246798",
@@ -12,9 +12,9 @@
   },
   "engine": {
     "repository": "TheAxiomFoundation/axiom-rules-engine",
-    "commit": null,
-    "wheel_sha256": null,
-    "status": "awaiting the hub engine pin at M0 (D1 method card MC1); the composer (package G6, not yet written) is to refuse to run until both pins are set"
+    "commit": "753ce7d9ac4cfb161f6aa4f281f5a3980e67a57d",
+    "wheel_sha256": "054c9196d5371f13bb7c2b28f8161ac92294e5c8543aeb6e221fde243897ed42",
+    "status": "Pinned for G8-x (D1 method card MC1). commit is the axiom-rules-engine commit that NZ's input surface and input closure record. wheel_sha256 is the SHA-256 of the dense-extension wheel the NZ hub built at that commit (maturin build --release --locked; CPython 3.14, macOS arm64). axiom_engine_ref records both as declared and hashes no installed wheel."
   },
   "entity_names": {
     "person": "Person",

packages/microcosm-build/src/microcosm/build/nz/export_replay_bindings.json

@@ -1,4 +1,4 @@
 {
   "export_contract_sha256": "4d388236cc3d3382501f5cfae8eae44b661fba6919681965eaa3fe4e032efe86",
-  "axiom_input_closure_sha256": "e19c4f4e3903b1e23fdea46fe9227157d908c6ce943fbf53ee49531de9903714"
+  "axiom_input_closure_sha256": "acd994081b81c21f7b8da4ba80648d92cfd7ddd849db8bfa1db98c2fa2fe3b2e"
 }

packages/microcosm-build/src/microcosm/build/nz/transport_graph.json

@@ -4564,7 +4564,13 @@
           ],
           "encoding": "value"
         },
-        "measurement": "ird_taxable_income_ty2027_v1",
+        "measurement": {
+          "resource": "income_measurements",
+          "path": [
+            "measurement_id"
+          ],
+          "encoding": "value"
+        },
         "identities": {
           "resource": "income_measurements",
           "path": [
@@ -4636,14 +4642,25 @@
           ],
           "encoding": "value"
         },
-        "measurement": "ird_taxable_income_ty2027_v1",
+        "measurement": {
+          "resource": "income_measurements",
+          "path": [
+            "measurement_id"
+          ],
+          "encoding": "value"
+        },
         "weight_entity": "household"
       },
       "artifact_inputs": [
         {
           "name": "problem",
           "producer": "nz.targets.problem",
           "artifact": "problem"
+        },
+        {
+          "name": "solution",
+          "producer": "nz.calibrate",
+          "artifact": "solution"
         }
       ],
       "artifact_outputs": [
@@ -4702,7 +4719,13 @@
           "encoding": "json"
         },
         "weight_entity": "household",
-        "measurement": "ird_taxable_income_ty2027_v1"
+        "measurement": {
+          "resource": "income_measurements",
+          "path": [
+            "measurement_id"
+          ],
+          "encoding": "value"
+        }
       },
       "artifact_inputs": [
         {

packages/microcosm-build/src/microcosm/build/transport/income_evidence_kernels.py

@@ -15,7 +15,9 @@
     size exactly as declared), beside the totals of what the measure does
     include at the calibration problem's pre-fit weights. It adds no
     model-side figure for an omitted item, because the register's sizes mean
-    IRD's.
+    IRD's. It decodes the fit's solution against that problem and refuses a
+    frame whose weights aren't the solution's, so the problem it reads is the
+    one this run's fit solved.
 ``transport.native_band_report@1``
     Sets the model, at the run's pre-fit and post-fit weights, beside every
     native row of the publisher's table: the row's count and amount from the
@@ -31,9 +33,14 @@
 first-party module whose source its ``run()`` executes, and every first-party
 module outside ``microcosm.graph`` whose constants it reads.
 ``microcosm.graph.kernel`` and ``microcosm.graph.decl`` stay out, as in the
-evidence kernels' lists. The trace test checks the executed modules and can't
-see a constant read, so a comment marks each entry whose module a run only
-reads constants from.
+evidence kernels' lists. Two tests check each list. A trace of one run checks
+the executed modules, and can't see a constant read. A static walk of the
+kernel's class and ``run()`` from source
+(``test_support/microcosm_build/static_reach.py``) over-approximates what a run
+can reach, in the ways that module's docstring lists. Outside
+``microcosm.graph`` it must name exactly the listed modules, less one guarded
+branch that no decoded target opens. A comment marks each entry whose module a
+run only reads constants from.
 ``evidence_contract`` hashes only the kernels it lists, and this module stays
 outside the gate battery's binding closure, so adding it re-keys no battery.
 
@@ -744,7 +751,7 @@ class OmissionRegisterKernel(KernelBase):
         "measurement",
         "weight_entity",
     )
-    _inputs = MappingProxyType({"problem": PROBLEM_TYPE})
+    _inputs = MappingProxyType({"problem": PROBLEM_TYPE, "solution": SOLUTION_TYPE})
 
     def implementation_hash(self) -> str:
         return _implementation_hash(self.ref)
@@ -799,6 +806,29 @@ def run(self, context: KernelContext) -> KernelResult:
             entity=entity,
             weight_entity=weight_entity,
         )
+        # Tie the problem to this run's fit: the solution must be the one
+        # solved from it, every weight unit on its axis must hold a record, and
+        # each record's frame weight (its weight unit's, which the frame
+        # broadcasts) must equal the solution's as a float64 value.
+        solution = decode_solution(
+            payloads["solution"],
+            problem_sha256=problem.sha256,
+            entity_ids=problem.entity_ids,
+        )
+        if len(set(rows)) != len(problem.entity_ids):
+            raise ValueError(
+                f"{ref} frame has a {weight_entity} on the problem's axis with no "
+                f"{entity}."
+            )
+        frame_weights = context.weights.get(entity)
+        if frame_weights is None or not np.array_equal(
+            np.asarray(frame_weights.values, dtype=np.float64),
+            np.asarray(solution.weights, dtype=np.float64)[rows],
+        ):
+            raise ValueError(
+                f"{ref} frame {entity} weights differ from the calibration "
+                f"solution's {weight_entity} weights."
+            )
         # The problem's initial weights are the fit's pre-fit weights, on its
         # own entity axis; each record takes its weight unit's.
         initial = problem.problem.initial_weights

packages/microcosm-build/tests/engine_free/shared/test_nz_axiom_input_closure.py

@@ -218,15 +218,21 @@ def test_every_default_is_cited_and_listed_for_review(self) -> None:
         )
 
     def test_no_input_is_claimed_engine_optional(self) -> None:
-        # The axiom-rules-engine commit in input_surface.engine (04315d94) has
-        # no optional inputs; the closure says why and classes none as
-        # engine-optional.
+        # The axiom-rules-engine commit in input_surface.engine (753ce7d9, the
+        # bindings' engine pin) gives the modules no optional inputs; the
+        # closure says why and classes none as engine-optional.
         assert not any(
             entry.closure_class is ClosureClass.ENGINE_OPTIONAL
             for entry in CLOSURE.entries
         )
         assert "input_surface.engine" in CLOSURE.engine_optional_evidence
-        assert CLOSURE.surface_engine_commit.startswith("04315d94")
+        assert CLOSURE.surface_engine_commit.startswith("753ce7d9")
+        bindings = json.loads(
+            (CLOSURE_PATH.parent / "axiom_rules_bindings.json").read_text(
+                encoding="utf-8"
+            )
+        )
+        assert CLOSURE.surface_engine_commit == bindings["engine"]["commit"]
         assert CLOSURE.surface_engine_repository == (
             "TheAxiomFoundation/axiom-rules-engine"
         )

packages/microcosm-build/tests/engine_free/shared/test_nz_transport_declarations.py

@@ -677,21 +677,20 @@ def test_nz_executable_support_keeps_region99_separate_from_ta999():
 
 
 def _nz_composition_spec_and_config():
-    """Supply compiler-only toy engine pins and CREATE's inventory.
+    """Supply synthetic engine refs and CREATE's inventory.
 
-    The packaged graph declares every money producer, so the compiler checks
-    the complete packaged node and population structure with no synthetic
-    node. The toy pins certify no runtime result. CREATE's inventory holds
+    The packaged graph declares every money producer, and the packaged
+    bindings carry the engine pins, so the compiler checks the complete
+    packaged node and population structure with no synthetic node or pin. The
+    synthetic engine refs certify no runtime result. CREATE's inventory holds
     canonical concepts and unit outputs only: no bridge or region column
     enters as a root.
     """
     from microcosm.graph import Owned
 
     spec = _spec()
     resources = spec["resources"]
-    resources["axiom_rules_bindings"]["engine"].update(
-        commit="1" * 40, wheel_sha256="2" * 64
-    )
+    compose.validate_transport_activation(spec)
 
     rows = _instantiated_nz_rows(spec)
     produced = {
@@ -1630,10 +1629,13 @@ def test_the_income_evidence_readers_name_the_measurement_but_dont_declare_it():
 
     measurement_proofs counts every node upstream of a calibration that
     declares the measurement's id and needs exactly one. The three readers
-    name the stand-in's id under ``measurement``, and the register reads
-    the problem's household weights. The readers sit two boundary filters
-    after the reweight, and only stand-ins write the three proxy columns, so
-    nz.terminal holds what nz.income.standin wrote at nz.central.
+    name the stand-in's id under ``measurement``, selected from
+    income_measurements' own ``measurement_id``. The stand-in keeps the literal
+    that R2C-1's proof reads, and the composed readers resolve to it. The
+    register reads the problem's household weights and checks the frame's
+    against the fit's solution. The readers sit two boundary filters after the
+    reweight, and only stand-ins write the three proxy columns, so nz.terminal
+    holds what nz.income.standin wrote at nz.central.
     """
     spec = _spec()
     rows = spec["resources"]["transport_graph"]["nodes"]
@@ -1643,13 +1645,21 @@ def test_the_income_evidence_readers_name_the_measurement_but_dont_declare_it():
     ]
     assert declarers == ["nz.income.standin"]
     by_id = {row["id"]: row for row in rows}
+    assert by_id["nz.income.standin"]["params"]["measurement_id"] == _MEASUREMENT
+    selection = {
+        "resource": "income_measurements",
+        "path": ["measurement_id"],
+        "encoding": "value",
+    }
+    _, _, graph, _ = _nz_composed()
     for producer, _ in _INCOME_EVIDENCE.values():
-        assert by_id[producer]["params"]["measurement"] == _MEASUREMENT, producer
+        assert by_id[producer]["params"]["measurement"] == selection, producer
+        assert graph.node(producer).params["measurement"] == _MEASUREMENT, producer
     register = by_id["nz.evidence.omission_register"]
     assert register["params"]["weight_entity"] == "household"
     assert [
         (edge["name"], edge["producer"]) for edge in register["artifact_inputs"]
-    ] == [("problem", "nz.targets.problem")]
+    ] == [("problem", "nz.targets.problem"), ("solution", "nz.calibrate")]
     for name, base in (("nz.as", "nz.calibrate"), ("nz.terminal", "nz.as")):
         assert (by_id[name]["kernel"], by_id[name]["base"]) == (
             "transport.boundary@1",