Fix the NZ bridge doc and tests the #1440 review flagged

changelog.d/nz-as-bridge-review-follow-up.fixed.md

@@ -0,0 +1 @@
+Correct the NZ Accommodation Supplement bridge's documentation and tests after the review of #1440. `docs/transport-entitlement.md` now names `nz.as.bridge` as the shared population that runs the masks, gross rate and cutout for S0, S1, S2, S3 and S5, and the entitlement generator's docstring drops the sentence that contradicted it. `as_rate_bridge.json` now states why the youngest-child override takes a mask: a null age would reach the engine as NaN. `benefit_unit_rule.json` now says the youngest-child age column is null for a childless family, as `unit_construction.py` writes it. The childless-family tests read the Axiom adapter's own batch builder, and a new test runs a mixed couple through S5's base rate.

docs/transport-entitlement.md

@@ -178,15 +178,18 @@ does this with its invented family-credit engine. It supplies no real WFF
 inputs or statutory claim.
 
 `build/nz/as_rate_bridge.json` declares both NZ bridge outputs as kernel
-calls, with no new kernel. On each population that runs the Accommodation
-Supplement engine, a `calibration.prepare@1` node writes the family masks the
-equality-only bridge masks compare, a `simulate.counterfactual@1` node writes
-the reg 18(3)(b) sole-parent Jobseeker Support rate, and a
+calls, with no new kernel. A `calibration.prepare@1` node writes the family
+masks the equality-only bridge masks compare, a `simulate.counterfactual@1`
+node writes the reg 18(3)(b) sole-parent Jobseeker Support rate, and a
 `simulate.solve_zero@1` node writes the reg 18 cutout: the zero of each
 unit's Jobseeker Support residual over weekly income on [0, 5000]. A sole
 parent's residual is the sole-parent rate minus the Income Test 3 reduction;
-every other unit's residual is its own net payment. The resource keeps the
-harness expressions as the statement of equivalence. Each branch runs its own
+every other unit's residual is its own net payment. The shared `nz.as.bridge`
+population, a FILTER branch of `nz.as`, runs these three nodes once for S0,
+S1, S2, S3 and S5, and each variant repeats them on its own housing
+population. A node on `nz.as` itself would re-key the skeleton's
+`nz.terminal`, which composition refuses. The resource keeps the harness
+expressions as the statement of equivalence. Each branch runs its own
 reg 17 base-rate counterfactual: zero-income Jobseeker Support once for a
 single unit and twice for a partnered one, plus the eldest-child family tax
 credit over 52 for a unit with a dependent child. The tests in

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

@@ -111,7 +111,7 @@
           "variable": "jobseeker_support_youngest_dependent_child_age",
           "unit_attribute": "family.youngest_dependent_child_age",
           "value_source": "unit",
-          "reason": "harness: family_shape youngest_child_age (run.py:1938,1951). The unit column is null for a childless family, and the engine adapter reads integer inputs as int64, so the override applies only where the unit has a dependent child; other rows keep the closure default 0.",
+          "reason": "harness: family_shape youngest_child_age (run.py:1938,1951). The unit column is null for a childless family, and that null would reach the engine input as NaN, so the override applies only where the unit has a dependent child; other rows keep the closure default 0.",
           "mask": {
             "unit_attribute": "family.accommodation_supplement_bridge_has_dependent_child",
             "equals": true
@@ -285,7 +285,7 @@
             "variable": "jobseeker_support_youngest_dependent_child_age",
             "unit_attribute": "family.youngest_dependent_child_age",
             "value_source": "unit",
-            "reason": "harness: family_shape youngest_child_age (run.py:1938,1951). The unit column is null for a childless family, and the engine adapter reads integer inputs as int64, so the override applies only where the unit has a dependent child; other rows keep the closure default 0.",
+            "reason": "harness: family_shape youngest_child_age (run.py:1938,1951). The unit column is null for a childless family, and that null would reach the engine input as NaN, so the override applies only where the unit has a dependent child; other rows keep the closure default 0.",
             "mask": {
               "unit_attribute": "family.accommodation_supplement_bridge_has_dependent_child",
               "equals": true

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

@@ -91,7 +91,7 @@
       "family.is_sole_parent": "single adult with at least one dependent child",
       "family.is_partnered": "the unit contains a partnered couple",
       "family.is_single": "the unit has one adult (not family.is_partnered)",
-      "family.youngest_dependent_child_age": "age of the youngest dependent child; 0 when there is none"
+      "family.youngest_dependent_child_age": "age of the youngest dependent child; null when there is none"
     }
   },
   "settled_date": "2026-10-12",

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

@@ -31,7 +31,7 @@
 from microcosm.build.transport.entitlement import _parameter, _rows
 from microcosm.build.transport.target_kernels import CalibrationPrepareKernel
 from microcosm.frame import ExportContract, Frame, VariableMetadata, WeightKind, Weights
-from microcosm.frame.adapters.axiom import NZ_SCHEMA
+from microcosm.frame.adapters.axiom import NZ_SCHEMA, _batch_from_table
 from microcosm.frame.rules_kernels import (
     SimulateCounterfactualKernel,
     SimulateSolveZeroKernel,
@@ -373,6 +373,7 @@ class Unit:
     adult_age: int = 30
     cash_assets: float = 0.0
     rent: float = 400.0
+    partner_age: int | None = None
 
 
 _RECEIPT_FLAGS = {
@@ -414,7 +415,12 @@ def _frame(units):
         adults = 2 if unit.partnered else 1
         for member in range(adults + len(unit.children)):
             adult = member < adults
-            age = unit.adult_age if adult else unit.children[member - adults]
+            if not adult:
+                age = unit.children[member - adults]
+            elif member == 1 and unit.partner_age is not None:
+                age = unit.partner_age
+            else:
+                age = unit.adult_age
             row = {
                 "person_id": next_person,
                 "person_household_id": family_id,
@@ -724,28 +730,28 @@ def _youngest_received(units, *, masked=True):
         _run(frame, row, scenario, engines)
     finally:
         received = engines[_REFS["main_benefits_rates"]].received
-    return received[-1][_YOUNGEST]
+    table = received[-1][[_YOUNGEST]]
+    return _batch_from_table(table, [_YOUNGEST])[_YOUNGEST]
 
 
 def test_a_childless_family_keeps_the_closure_default_youngest_child_age():
     units = [Unit(), Unit(children=(4, 7))]
     received = _youngest_received(units)
     # 0 for the childless family's adult, 4 for the parent and both children.
-    assert pd.Series(received).to_numpy(dtype=np.int64).tolist() == [0, 4, 4, 4]
+    assert received.dtype == np.float64
+    assert received.tolist() == [0.0, 4.0, 4.0, 4.0]
 
 
 def test_unmasked_youngest_child_override_carries_a_missing_age_to_the_engine():
     """Without the mask the childless unit's null age reaches the engine input.
 
-    pandas 3 reads the nullable Int64 source as float64 with NaN, so the
-    adapter's int64 conversion raises under numpy's invalid-value check and
-    otherwise yields int64's minimum.
+    pandas 3 reads the nullable Int64 source as float64 with NaN, so the adapter
+    takes its float branch and hands the engine NaN for that adult.
     """
     units = [Unit(), Unit(children=(4, 7))]
     received = _youngest_received(units, masked=False)
-    assert received.isna().tolist() == [True, False, False, False]
-    with np.errstate(invalid="raise"), pytest.raises(FloatingPointError):
-        pd.Series(received).to_numpy(dtype=np.int64)
+    assert received.dtype == np.float64
+    assert np.isnan(received).tolist() == [True, False, False, False]
 
 
 def test_bridge_scratch_columns_stay_out_of_the_export_contract():
@@ -761,7 +767,7 @@ def test_bridge_scratch_columns_stay_out_of_the_export_contract():
     assert not persisted & (scratch | {_GROSS})
 
 
-def test_every_as_population_runs_its_own_bridge_and_every_branch_its_base_rate():
+def test_each_bridge_population_runs_the_bridge_and_every_branch_its_base_rate():
     rows = {row["id"]: row for row, _ in _rows(_SPEC)[0]}
     scenarios = _SPEC["resources"]["scenarios"]["scenarios"]
     for node_id in ("masks", "sole_parent_gross_rate", "cutout"):
@@ -926,6 +932,25 @@ def test_a_partnered_superannuitant_unit_takes_coefficient_two_in_s5():
     assert abs(base[101] - (SuperannuationEngine.weekly + credit)) <= 1e-9
 
 
+@pytest.mark.parametrize(("first_age", "second_age"), [(70, 50), (50, 70)])
+def test_a_mixed_couple_in_s5_takes_the_superannuitants_amount_for_both_partners(
+    first_age, second_age
+):
+    # The fixture pays the partner under 65 nothing, so ``max`` over members
+    # gives the superannuitant's amount in either person order.
+    base = _s5_base_rate(
+        [
+            Unit(
+                receipt_group="superannuitant",
+                adult_age=first_age,
+                partner_age=second_age,
+                partnered=True,
+            )
+        ]
+    )
+    assert base[100] == 2 * SuperannuationEngine.weekly
+
+
 def test_the_central_base_rate_keeps_jobseeker_for_superannuitant_units():
     _, _, base, _ = _bridge(
         [Unit(receipt_group="superannuitant", adult_age=70, partnered=True)]

packages/microcosm-build/tests/golden/nz_country_spec.json

@@ -1,6 +1,6 @@
 {
   "country": "nz",
-  "fingerprint": "fc9e3f65841862016207a1baa368e8ecf0476d194675ad8ec771811a5ea54f4f",
+  "fingerprint": "4d6be0d9d5adc1ecc1c65a04c3fa97baa646665eeb9439b6a8100642a7723f79",
   "gate_ids": [
     "calibration_per_family_fit",
     "demography_per_family_fit",
@@ -40,12 +40,12 @@
   "resource_hashes": {
     "as_area_crosswalk.json": "56eeb638d77c73fd92ca541d9d2f249fd3062bece0503470e0e85dde70c4c05b",
     "as_encoding.json": "91e0e9972abfaaa202a13ede6e2dbc815a4c6c9514726c970928a47adfef0e32",
-    "as_rate_bridge.json": "25fa43210fd91655988933fe23ac6411bc30ddeb73c480c24da19999ffa2da22",
+    "as_rate_bridge.json": "1399321cc7ef44486f7ab407968b036930ff6ae8fbab4c30ce710ea0fff7301c",
     "atomic_geography_support.json": "6efd09d5cde2f16d4d01e522a721191a829a754d2ce019d75bb0407085ecb4b7",
     "axiom_input_closure.json": "1c237a316eecffd8d8eb9258a451b992373b804d72b411895337647c272731aa",
     "axiom_rules_bindings.json": "fb163e4842c8a910d63bd0ecf0fc81e5a55a2dddaa9cde66a52d9f4ff2c791f3",
     "benefit_gross_up.json": "bb0100927c3d666f497b24694109463de247d53d6e7ee0cfa092255e33c5514e",
-    "benefit_unit_rule.json": "a9e8cbfd38a64d2f8db0002fe8c31bb5a18fb34dbe59fa943336487ce95e2ca8",
+    "benefit_unit_rule.json": "c84efba8353c770131256a8c17b82c7d98f738e6dd8ef217e19d30014cb14604",
     "calibration_bindings.json": "abf5278b51a13530e2c589a932942a185f5dda095c1eb3e4bc703999c24852b3",
     "calibration_settings.json": "56e4d1da8468245da6bd41daf263425cc65e4e9c9f11ff45eee459d3aed1989e",
     "chronicle_feed.json": "60a80dd0042838b0412baf9dd0985c961b4253a6876d16bcdbb8a2c6c03de0c7",

tools/build_nz_entitlement_graph.py

@@ -1,16 +1,16 @@
 #!/usr/bin/env python3
 """Generate NZ's post-calibration entitlement branch declarations.
 
-The Accommodation Supplement bridge follows ``as_rate_bridge.json``. On each
-population that runs the AS engine, a ``calibration.prepare@1`` node writes
-the family masks, a ``simulate.counterfactual@1`` node writes the reg 18(3)(b)
-sole-parent gross rate and a ``simulate.solve_zero@1`` node writes the reg 18
-cutout. The shared ``nz.as.bridge`` population, a FILTER branch of ``nz.as``,
-runs these once for every entitlement scenario, and each variant repeats them
-on its own housing population. An extension node on ``nz.as`` itself would
-re-key the skeleton's ``nz.terminal``, which composition refuses. Each
-branch runs its own reg 17 base-rate counterfactual. Every coefficient, mask
-and override comes from the resource; this generator adds no schedule.
+The Accommodation Supplement bridge follows ``as_rate_bridge.json``. A
+``calibration.prepare@1`` node writes the family masks, a
+``simulate.counterfactual@1`` node writes the reg 18(3)(b) sole-parent gross
+rate and a ``simulate.solve_zero@1`` node writes the reg 18 cutout. The shared
+``nz.as.bridge`` population, a FILTER branch of ``nz.as``, runs these once for
+every entitlement scenario, and each variant repeats them on its own housing
+population. An extension node on ``nz.as`` itself would re-key the skeleton's
+``nz.terminal``, which composition refuses. Each branch runs its own reg 17
+base-rate counterfactual. Every bridge coefficient, mask and override comes
+from the resource; this generator adds no schedule.
 """
 
 from __future__ import annotations