[MaxGhenis]

G8-r2c shared hunk D (tracking #1183): per-band support diagnostics and a release-blocking measurement_prefit gate (ruling R2C-3), with ruling R2C-12's line in country_spec.py. Stacked on hunk A (#1456, nzhub/g8r2c-a-summed-bands); the two share one re-key window.

B/ is packages/microcosm-build/src/microcosm/build/, T/ is B/transport/, F/ is packages/microcosm-frame/src/microcosm/frame/, and tests are in packages/microcosm-build/tests/engine_free/shared/. af038d04 merges A's EXPAND writer fix d916ea8d into D, which moves T/compose.py lines after :591 down by up to 17. 91bf4c0f then merges A's bf418af1, test lines only, and changes one number in a test docstring. Other lines are at 4c644db7, after the fix round. The round changed no src line in D, so every src line number below still holds.

The change

Per-band support (B/reference_intervals.py). A table that declares the optional diagnostics.support_household_entity gets, per band, the distinct source households and the source-household Kish ESS, for count and amount contributions separately (band_support, :1027; kish_ess, :1010).

• A record contributes w to its band's count and w × z to its band's amount. Here w is the record's weight as the diagnostics read it, the frame's weights on the table's entity, which are the weights entering calibration, and z is its floored amount in cents. So an amount contribution's weight is w × z, the record's share of the band's weighted amount.
• The code sums contributions per source household before the Kish ratio, so donor clones of one household add no support. A household counts only where its summed contribution is nonzero, and an empty band reports 0 households and an ESS of 0.
• Reaching the household (record_source_households, :1077). The frame declares membership only through the person table: each group entity has a person column EntitySchema.membership_column(group) (F/schema.py:138), which Frame._group_positions reads (F/bundle.py:649). A person reads its household column directly. A unit on another entity, such as a tax_unit, reads its members' households and must nest in exactly one, as Frame._aggregate_group_to_group requires when it places one group onto another (F/bundle.py:1070). So two tax units in one household count as one source household. The code counts source households, not units.
• It refuses by name an entity with no household membership: a household entity the frame doesn't declare, an entity that isn't declared, or a unit with no members or with members in several households. It also refuses a frame without household.source_household_id.
• Backward compatibility. A table that declares none of D's fields parses, classifies and reports exactly as before: boundary_diagnostics refuses source households for such a table, and its report carries no support key (:801, :933-945, :995).

The gate (T/gate_bindings.py:1664, _measurement_prefit; binding at :2044). It reads the problem artifact and the battery's reserved graph_evidence map. Every threshold comes from the country's gates.json entry, and none has a default. A country without the entry gets no gate, because the battery evaluates only the entries its manifest lists (B/gate_battery.py:705). The gate refuses on R2C-3's four checks:

• Roster totals (:1803). The gate sums the model total over every band's count row, and separately over every band's amount row, across all the entry's declared tables together (NZ declares one table, with 12 bands), and compares each sum with the summed reference. The model total is the problem's A @ w0 at the weights entering calibration (problem.estimates(initial_weights), :1726), and the reference is the matrix's right-hand side, the aged reference the matrix uses (target_vector, :1727). It refuses outside roster_total_bounds, inclusive: with R2C-3's [0.75, 1.25], exactly 75% and 125% pass.

• Concentrated discrepancy (:1809). A band is discrepant if its count or its amount falls outside band_bounds, and it counts once if both fail. The gate refuses when more than maximum_discrepant_bands bands are discrepant. R2C-3's "3 or more refuse" is maximum_discrepant_bands: 2.

• Support (:1777, :1782). It refuses a band with fewer than minimum_source_households source households, or a Kish ESS below minimum_kish_ess, checked for count and amount separately. Exactly 30 passes.

• Missing evidence (:1714 onward). Today the gate checks:

  • that each alias in required_evidence is wired into the battery as an artifact edge (graph_evidence);
  • that the problem carries interval diagnostics for each declared table, with per-band support;
  • that each band's count and amount references entered the matrix.

  So of R2C-3's evidence items, the native-band report, the omission register, the VSI, VEX and VPT runs and the identity receipts are checked for presence only, by the alias the country declares. The gate doesn't read their contents.

R2C-12. B/country_spec.py:134 adds "measurement_prefit" to ALLOWED_GATE_FUNCTIONS. That's the only line of that file this PR touches. The test that pinned the loader's refusal is gone, and test_a_gates_json_entry_naming_the_gate_loads loads a gates.json entry naming the gate.

R2C-3's numbers are pinned in test_support/microcosm_build/reference_intervals.py as R2C3_PREFIT_THRESHOLDS, and test_r2c3_thresholds_are_pinned pins them for the NZ session to reuse. NZ's own entry belongs to the NZ session.

Invariants

Invariant	How it is tested
Source-household Kish ESS ≤ distinct source households, and a band with none has ESS 0	Hypothesis, test_ess_is_at_most_distinct_households_and_clones_leave_it_unchanged: 7 source households over a nil band and 3 positive bands, so it reaches shared households and empty bands
Cloning records within a source household leaves per-band support unchanged	The same property. Each original record carries clones × w and each clone carries w, so the comparison is exact on every example
The diagnostics' weighted band counts equal Σ w × indicator over the prepared columns	Differential, test_weighted_band_counts_equal_weighted_indicator_sums, against interval_indicators
Two units in one household count once, with Kish ESS over source households	test_two_tax_units_in_one_household_count_once_with_source_household_ess, test_clone_households_share_one_source_household
A household whose summed contribution is zero adds no support	test_a_zero_weight_household_is_not_counted
The model side is the matrix at the weights entering calibration, A·w₀, not at unit weights	test_the_model_side_is_the_matrix_at_the_initial_weights; the test problem's estimates multiplies a matrix by the weights it receives, with w₀ = (2, 4)
A table without D's fields reports as before	test_a_table_without_support_fields_parses_and_reports_as_before
Each threshold refuses just outside and passes at its edge, on each side	test_roster_totals_refuse_outside_75_to_125_percent, test_three_discrepant_bands_refuse_and_two_pass (count-only and amount-only failures, 2, 3 and 4 bands), test_band_bounds_are_inclusive_at_50_and_200_percent, test_a_band_failing_count_and_amount_counts_once, test_support_refuses_below_30_households_or_ess_separately, test_missing_evidence_refuses_each_way

Mutants

Mutant	Red
Each record is its own household (no grouping)	4 tests, among them the clone property and the two-tax-unit test
Amount contribution weight w instead of w × z	test_two_tax_units_in_one_household_count_once_with_source_household_ess
Accept a unit with members in two households	test_an_entity_without_household_membership_refuses_by_name[frame0-None-None]
Drop the backward-compatibility check	test_a_table_without_support_fields_parses_and_reports_as_before
Household floor 29 instead of 30	4 cases of test_support_refuses_below_30_households_or_ess_separately. This mutant survived the pushed test, whose 29-household case also had ESS 29; the test now checks which check refused.
ESS floor <= instead of <	the exact-30 cases
Drop the totals check	the 74.99% and 125.01% cases
Concentrated rule >= instead of >	the 2-band pass cases and test_a_band_failing_count_and_amount_counts_once
Exclusive band bounds	8 cases, among them test_band_bounds_are_inclusive_at_50_and_200_percent[0.5-True]
Drop the evidence-alias check	test_missing_evidence_refuses_each_way
Remove R2C-12's line	test_a_gates_json_entry_naming_the_gate_loads
Fix round OD1: count a household whose contribution is zero	test_a_zero_weight_household_is_not_counted
Fix round OD2: only the count role marks a band discrepant	the 6 amount cases of test_three_discrepant_bands_refuse_and_two_pass
Fix round OD3: evaluate the model side at unit weights	test_the_model_side_is_the_matrix_at_the_initial_weights and the edge cases that pass at w₀

At 4c644db7 I reran two of the mutants above: no grouping (5 tests red, among them test_two_tax_units_in_one_household_count_once_with_source_household_ess and the clone property) and concentrated >= (test_a_band_failing_count_and_amount_counts_once and the four 2-band pass cases).

Merge-up

• 7bc34fdb merges A's merge-up 923b4de1, which carries w4d fcdb47395293955347ec2697feef1a736a76cc23, into D. Git merged without conflict. git diff 923b4de1 7bc34fdb shows D's five files and nothing else, and the lines D adds to T/gate_bindings.py match D's own diff against A's old head.
• 38cf1bf0 merges A's tested head 4d2f3b2b into D. One conflict, in test_support/microcosm_build/reference_intervals.py: A reformatted the end of its fixtures and D appends there. I kept both sides.
• 5e4c846a merges the fix round's A head 30ae4599 into D with --no-ff. Git merged without conflict, and the merge brings in only A's fix-round files.
• af038d04 merges A's EXPAND writer fix d916ea8d into D with --no-ff. Git merged without conflict, and git diff 4c644db7 af038d04 shows only that fix's two files.
• 40079dfa merges A's bf418af1 into D with --no-ff, and 91bf4c0f corrects the fake's docstring: columns model/4 and model/8 give 3/8 of the model at unit weights, not 5/12. The delta review (Activate the NZ graph against pinned Chronicle facts (NZ graph G8) #1183 comment 6592283997) asked for both. git diff 4c644db7 91bf4c0f shows the EXPAND fix's two files plus that docstring's test file, and the proof and prefit test files give 80 passed at 91bf4c0f.
• nzhub/g8w4d-nz-wiring is still at fcdb4739.

Shared hunks

Against A's head 30ae4599, at 4c644db7:

• B/country_spec.py: 134
• B/reference_intervals.py: 49, 59, 67-68, 134-138, 171, 331-334, 336-338, 446, 782, 784-790, 801-805, 933-943, 945, 995-1134, 1156-1160
• T/gate_bindings.py: 46-49, 1636-1835, 1900-1905, 2044-2052
• tests/test_reference_interval_support.py: 1-386
• tests/test_transport_gate_bindings.py: 605
• tests/test_transport_measurement_prefit.py: 1-300
• test_support/microcosm_build/reference_intervals.py: 380-423

The fix round changed only D's two test files. The src ranges are as they were at 38ae4a74.

The pin in test_transport_gate_bindings.py:605 gains measurement_prefit, because the roster test lists every registered gate. No module-hash pin moved.

Implementation hashes

I computed implementation_hash() for 43 transport kernels (every kernel the composed toy registry installs plus every module-level transport kernel), 89 UK runtime kernels (the UK spine registry and every UK kernel object or no-argument kernel class) and 6 US runtime kernels (the post-transfer registry on a stub engine and the module-level kernels), each in its own head's tree. While each hash ran, a trace recorded every source file it read. No UK or US kernel reads reference_profiles.py, reference_intervals.py or country_spec.py, and none moves. That includes the UK target kernels (uk.full.target_compilation@1, uk.full.national_targets@1 and the rest), which leave country_spec out on purpose (uk_runtime/target_compile_modules.py:59, EXCLUDED_DIRECT_IMPORTS). Editing T/compose.py moves no source_hash: no kernel reads it. Editing T/gate_bindings.py moves gates.battery@1, whose implementation_hash folds in its bindings' sources (T/gate_kernels.py:747-775), so D's additions there move the battery too.

Kernel	fcdb4739	A	D	Why it moves: the changed modules it hashes
calibration.prepare@1	c13e76497b8d2dfb	6e214cef0e1455ff	180823901b9a0d99	reference_intervals, reference_profiles
diagnostics.calibration@1	d38f8fad075a9ace	30654dfc7038f560	30654dfc7038f560 (same)	reference_profiles
gates.battery@1	29f0505b6a0093f8	00216abb7158bf98	ac050bab09f48dc2	country_spec, gate_bindings, reference_profiles
geography.support_from_facts@1	60273ea8388098b8	ad0a8e73d20405be	ad0a8e73d20405be (same)	reference_profiles
takeup.assign@1	53caf4ae91fc4acb	b2a3a50cc376108b	b2a3a50cc376108b (same)	reference_profiles
takeup.compare@1	a318a19b0d147ec9	0231145ac4631652	1605d7e8519f4b0f	country_spec, reference_profiles
takeup.gap@1	3e15ad5e2903c31c	6b200d952fe499c8	54f79560d896fab4	country_spec, reference_profiles
targets.compile@1	f3d372c40418665b	3da9798c5334f495	dc91d1479f8f84b1	country_spec, reference_profiles
targets.problem@1	aed111cadf163873	b949922486c0813b	fe401c736551e351	reference_intervals, reference_profiles
transport.create@1	82dc0880aefba156	ef23135c5ec9d98a	865c21926cee0ac0	country_spec, reference_profiles
transport.create@2	82dc0880aefba156	ef23135c5ec9d98a	865c21926cee0ac0	country_spec, reference_profiles
transport.quantile_map@1	0966a4eb8a492216	4519e4bbfd966f97	b25bb6538b19dd64	country_spec, reference_profiles

The fix round moves no hash. I recomputed implementation_hash() at 38ae4a74 and at 4c644db7, each in its own worktree: 43 transport kernels (the composed toy registry plus module-level transport kernels), 37 UK and 3 US runtime kernels that construct without an engine. All 83 hashes match, as do A's at 4d2f3b2b and 30ae4599. uk.full.rotated-holdout@1 and us.post_transfer.derive@1 need an engine to hash, so this run skipped them.

D adds no kernel to A's set. It moves calibration.prepare@1 and targets.problem@1 again through reference_intervals, and R2C-12's line moves the seven kernels that hash country_spec again: targets.compile@1, takeup.compare@1, takeup.gap@1 (not in R2C-12's list), transport.create@1/@2, transport.quantile_map@1 and gates.battery@1. D's additions to T/gate_bindings.py also move the battery. The battery's binding identity (the SHA-256 of repr(GateBatteryKernel(TRANSPORT_GATE_REGISTRY)._identity[0]), first 16 hex digits) moves at D only, from 715e6202808aade7 at fcdb4739 and A to 55947149ba78ec86, because the registry gains the measurement_prefit binding.

Tests

The closed list at 4c644db7 (one pytest run, -n 4, no --maxfail): 2714 passed, 84 skipped, 0 failed. Per file: country_spec 143, nz_as_area_crosswalk 67 (3 skipped), nz_as_bridge 34, nz_axiom_input_closure 19 (2 skipped), nz_benefit_gross_up 16, nz_family_reconciliation 7 (9 skipped), nz_joint_geography 21 (12 skipped), nz_money_normalization 56, nz_msd_benefit_rates 6, nz_msd_pre_activation_reconciliation 4 (1 skipped), nz_reference_aging 6 (16 skipped), nz_reference_intervals 13 (1 skipped), nz_spec_package 1203, nz_target_references 16 (38 skipped), nz_transport_declarations 29 (2 skipped), nz_transport_toy_graph 23, reference_interval_support 12, reference_intervals 107, reference_profiles 129, reference_summed_bands 23, spec_engine_country_bundles 22, spec_only_country_packages 37, transport_composed_graph 469, transport_gate_bindings 131, transport_graph 15, transport_measurement_prefit 47, transport_measurement_proof 27, transport_population_kernels 32. The repo has no test_be_*.py file.

Every skip needs data or an engine this environment doesn't have: nz_axiom_input_closure 2 (needs axiom_rules_engine with its dense extension and a rulespec-nz checkout); nz_transport_declarations 2 (pinned feed not staged); nz_as_area_crosswalk 3 (needs NZ_AS_AREA_SOURCES); nz_joint_geography 11 (needs NZ_AS_AREA_SOURCES and NZ_JOINT_GEOGRAPHY_SOURCES); nz_target_references 38, nz_reference_aging 16, nz_family_reconciliation 9, nz_joint_geography 1, nz_reference_intervals 1 and nz_msd_pre_activation_reconciliation 1 (need NZ_CHRONICLE_ARTIFACT).

Checks: python3 tools/ci_test_plan.py verify prints verification=ok at 4c644db7, and uv run ruff check packages/microcosm-build passes. I didn't rerun the hash pins outside the closed list or the NZ builders' --check. The round changed no file they read, and the earlier run at 38ae4a74 stands for them.

Fix round

The independent review (#1183, comment 6589309875) drove this round. Lines are at 4c644db7. No fix touched shared src.

Fix	Commit	Where	Red under its mutant	Shared hunks touched
Merge-up of A's fix round	5e4c846a	A's T/compose.py and tests	n/a	A's (see #1456)
D1/OD2: three or more amount-only discrepant bands refuse	4c644db7	tests/test_transport_measurement_prefit.py:157-174 (test_three_discrepant_bands_refuse_and_two_pass), against T/gate_bindings.py:1768-1769	the 6 amount cases	tests only
D1/OD3: _Problem.estimates uses its weights; w₀ = (2, 4)	4c644db7	tests/test_transport_measurement_prefit.py:40-62, :124-134 (test_the_model_side_is_the_matrix_at_the_initial_weights), against T/gate_bindings.py:1726	test_the_model_side_is_the_matrix_at_the_initial_weights and the edge cases	tests only
D1/OD1: a zero-weight household isn't counted	4c644db7	tests/test_reference_interval_support.py:190-206 (test_a_zero_weight_household_is_not_counted), against B/reference_intervals.py:1019	test_a_zero_weight_household_is_not_counted	tests only

Not done

• No NZ wiring. NZ's gates.json entry, its evidence aliases and the household entity its coarse table declares belong to hunk G. The problem frame must carry the household entity and source_household_id for the table's support to run; hunk G needs to wire that.
• The gate checks R2C-3's evidence items for presence by alias, not by content.
• The engine-present hash pins weren't run (no engine in this session).

https://claude.ai/code/session_01LwQq6JfFhwRFMYkokanqow

---

Generated by Claude Code

[MaxGhenis]

Merged by the NZ hub (session local_e407a527) at 91bf4c0 with --merge. Gates: gh pr checks exit 0; MERGEABLE; not draft; no CHANGES_REQUESTED review. Reviewed head: 91bf4c0. Gates: delta review 200 (#1183 comment 6592283997, APPROVE_WITH_NITS) after review 6589309875; its test-only nit fixed at 91bf4c0. Belgium's hub OK on D 91bf4c0 (NZ hub notes/1103/from-belgium-ad-nits-and-1477.md). R2C-11 recheck 2026-11-03 03:5xZ: no reservation in force. Merge commit, after A.