[MaxGhenis]

Targets main and lands after #1382. The UK Budget hold on shared compile-module hashes, which applied because eleven kernel implementation hashes move (see Cross-country), ended at 2026-10-29 12:00Z. f700082 fixed the round-1 review, and its delta review approved it with wording nits (see Review). Head fca69c1: d27a8f5 merges #1382's fix4 head (Belgium OK'd it 10/26), and fca69c1 fixes the changelog nit. A batch delta review of the fix4 merges approved them with no findings (see Verification); the BE re-clear passed (see Landing).

What this does (G8-r2b)

Adds exact income-band intervals, the prepared indicators built from them, and fixed-weight boundary diagnostics. No target is activated: all 474 IRD rows and 235 P1 rows stay parked until G8-r2c.

• Shared module, reference_intervals.py. Country data declares everything per table, and the module has no defaults for semantics:
  • entity, currency, columns and periods;
  • the first cut;
  • positive-band closure (left or right), nil band (present or absent) and negative-income policy.
  • Cuts are exact rationals over signed integer-cent inputs.
  • When a nil band is present, zero is nil under either closure. With a nil band and a zero first cut, the first positive band starts above zero, so no record is in two bands. Without a nil band, zero falls in the first band.
  • Validation refuses unresolved universes, malformed cuts and repeated scaling.
• Scaling. r2a's scale_money_and_intervals transformation now scales each named interval table once, with its companions' selected factor. Fractional-cent cuts stay exact.
• Output binding. In calibration_columns.py, interval outputs are their own kind beside predicate outputs, and name their entity explicitly. _prepare_roster_counts is untouched.
• Transport.
  • compose.py: a calibration.prepare@1 node receives the profile view's exact effective interval companion, and refuses if it supplies a different one. Only a spec with a profile view that has intervals is affected, which today means NZ.
  • target_kernels.py, preparation:
    • recomputes indicators from each variant's cuts;
    • records interval input fingerprints in its evidence, only when bindings name interval outputs;
    • refuses an intervals parameter without a resolution receipt.
  • target_kernels.py, the problem kernel: re-derives the fingerprints, checks the interval hash, and adds fixed-weight boundary diagnostics to its metadata and receipt, never to targets.
• NZ.
  • 237 national and 183 age bands are generated and checked against every feed label.
  • The national table scales in all three ageing profiles.
  • The canonical golden changes only identity fields.
• Integration with Add feed-derived NZ reference ageing factors (G8-r2a) #1382's fix3: test_nz_default_factor_resolution_validates_after_declaration_expiry now passes the interval document that NZ's profiles require.

Invariants (tested)

1. Classification is exact: scaled classification equals classification of the deflated source amount (property, with an independent Fraction oracle).
2. Partition and conservation: included and excluded records partition the input. Counts and amounts are conserved across bands, including the below-first-cut bucket, with nil unchanged.
3. Closure: the declared closure governs exact cuts and one cent either side; the alternative closure appears only in diagnostics.
4. No rounding of aged cuts: fractional-cent cuts stay exact (mutants dollar_rounded_edge and cent_rounded_edge killed).
5. Entity-generic: weights resolve through resolve_weights(entity) (mutant person_only_entity killed).
6. Diagnostics never move targets: RHS values are unchanged (mutant change_rhs_for_diagnostics killed).
7. Determinism and order: results are invariant under joint record-and-weight permutations, and a mismatched order refuses.
8. No-interval paths are byte-stable: with no interval outputs, the prepare evidence and the problem's metadata and receipt bytes are unchanged (the review compared 36 payloads, base against head, byte for byte). Only the implementation-hash cache bust remains.
9. With a nil band, zero belongs to nil only: under both closures, unscaled and scaled, a zero income is nil and never in a positive band (hub mutants restore_guard and nil_after_cuts killed).

Cross-country

• Belgium OK'd every transport hunk (10/24). BE has no transport graph or profile view, so no BE output bytes move. The re-clear on this head follows once it's green.
• Cache bust: 11 of the 32 kernel refs the hub probe enumerated move. The probe imported every microcosm package and the transport run registry at the base (2a64132) and at this head, and compared the implementation_hash() of every kernel instance it reached, 32 distinct refs:
  • the four kernels defined in target_kernels.py: calibration.prepare@1, targets.compile@1, targets.problem@1, takeup.compare@1;
  • seven kernels that hash target_kernels.py as a dependency: transport.create@1, transport.quantile_map@1, geography.support_from_facts@1, gates.battery@1, takeup.assign@1, takeup.gap@1, diagnostics.calibration@1.
  • The other 21 are unchanged.
  • Outside the enumeration: kernels the country runtimes and delegates instantiate dynamically, for example us.post_transfer.create@1, uk.full.target_compilation@1 and simulate.rules@1. The delta review hashed the first two, and neither shows another changed implementation.
  • The three engine-backed simulate kernels were read rather than run. With the built-in Axiom bindings they hash microcosm-frame modules, which this PR doesn't touch, plus their bound adapter classes.
• What moves downstream. A node key binds its kernel's implementation hash and its inputs' artifact keys (graph/keys.py, node_key). So any graph, in any country, that runs one of these kernels recomputes that node and everything downstream of it once. Every node of a graph rooted at transport.create@1 recomputes. Outputs are shown unchanged for no-interval inputs only for prepare and problem (invariant 8). The seven dependency kernels' own modules are unchanged.

Review

Round 1 (GPT-6.1 Sol, at 75c2768): REQUEST_CHANGES. All three findings are fixed in f700082:

• Major: test collection failed. The composed-graph interval config built Owned("person", "age") without its required dtype. It now passes "int64".
• Major: a combination the brief allows was refused. The brief says that with a nil band and a zero first cut, the first positive band also requires z > 0. The validator instead refused left closure with nil present and a zero first cut. The guard is removed (_position already puts zero in nil first), the refusal test is replaced by acceptance tests, the two first=1 workarounds are gone, and the Hypothesis generator now draws the combination.
• Minor: the cache-bust disclosure was understated. It is rewritten above from a probe of the 32 kernel refs the probe reaches, not only the four the review checked.

The review's other checks passed:

• exactness against an independent Fraction oracle (54,792 comparisons);
• partition and conservation, no semantic defaults, entity generality and output kind;
• scaling once, and eight transport tamper refusals;
• the no-interval byte differential, the NZ data and parking, and the fix3 integration.

Round 2 (GPT-6.1 Sol, delta 75c2768..f700082): APPROVE_WITH_NITS.

• Both majors are fixed. The composed-graph interval tests collect and pass, and 107 interval tests pass. 79,110 independent classifications agree, including left closure with a nil band and a zero first cut. The scaled diagnostics, transport, generator coverage and in-memory counterexamples also pass.
• The cache-bust impact is verified: all eleven changed kernels, and the downstream recompute.
• Its minor finding was that "every kernel" overstated the probe. That is fixed above: the claim now covers the 32 enumerated refs and names the dynamically instantiated kernels outside them.
• Its nit was that "zero is nil" holds only when a nil band is present. That is fixed above, and in the changelog at fca69c1.
• The test_transport_target_kernels.py selection was UNCHECKED, because the fleet disk hold blocked pytest; in-memory prepare and problem probes pass. Ruff, format, the diff checks and both NZ generators pass.
• Once disk allowed (2026-10-25), the hub ran test_transport_target_kernels.py at f700082 under pytest, in full, at -n 0: 67 passed. It passes again after the fix4 merge (see Verification).

Verification

• Lane (GPT-6.1 Sol, at d28817f, before the rebase): 8 files, 1,624 passed.
• Hub at 75c2768 plus the dtype fix:
  • all 18 baseline files pass, one process each. They include test_reference_profiles.py 120, test_transport_composed_graph.py 166, test_transport_entitlement_graph.py 67, test_nz_spec_package.py 1,201 and test_ledger_targets.py 451.
  • 12 of 12 mutants killed, each with a passing same-file control and byte-verified restoration: left_closed_cut, dollar_rounded_edge, cent_rounded_edge, drop_nil_band, all_person_universe, unknown_universe_as_false, double_applied_factor, change_rhs_for_diagnostics, omit_helper_source_hash, reuse_s0_indicators, person_only_entity, default_closure.
  • ci_test_plan.py verify, graph_acceptance_burndown.py --verify and both NZ generators' --check exit 0.
• Hub at f700082:
  • test_reference_intervals.py 107, test_nz_reference_intervals.py 14, test_transport_target_kernels.py 67 and test_transport_composed_graph.py 166 passed;
  • mutants restore_guard and nil_after_cuts killed (three new tests red, control green); left_closed_cut and default_closure rerun and still kill;
  • ruff check and format, and git diff --check, are clean.
• Hub at fca69c1, after the fix4 merge, one file per process at -n 0: test_country_spec.py 136, test_nz_spec_package.py 1,201, test_transport_target_kernels.py 67, test_transport_composed_graph.py 172, test_nz_reference_intervals.py 14, test_reference_intervals.py 107 and test_spec_engine_country_bundles.py 21 passed.
• Delta review of f700082 (GPT-6.1 Sol): APPROVE_WITH_NITS (see Review).
• Qualify the changelog's "zero is nil" with "when a nil band is present" (fca69c1).
• Merge Add feed-derived NZ reference ageing factors (G8-r2a) #1382's fix4 head (d27a8f5). Belgium OK'd it; the merged transport/cli.py is c4ddcb4's blob (6f42b3b0) and the only transport file the merge changes. The NZ golden conflicted and was regenerated from the merged tree; per-key attribution shows no mismatches. Reruns above.
• Batch delta review of the fix4 merges (GPT-6.1 Sol, at db02ae8 on Normalize donor money with exact declared factors #1395, which contains this head): APPROVE, no findings. It rebuilt each merge with git merge-tree (only the conflicted NZ golden differs), attributed the golden key by key, checked Belgium's two conditions and the wording commits, and ran test_country_spec.py 136, test_nz_money_normalization.py 17 and the 12 currency-check tests of test_transport_composed_graph.py.
• BE re-clear on the landing head (see Landing).

Landing (2026-10-29)

• The NZ hub retargeted this PR to main on 2026-10-29, so it ran CI for the first time; stacked PRs never ran it.
• 3069ec0 merges main (fddbc70) through update-branch. d012ddb and 59a21bd merge Add feed-derived NZ reference ageing factors (G8-r2a) #1382's CI import fix: two test files imported microcosm_build.reference_factors as a top-level module, a path CI doesn't provide. Both merges conflicted in test_reference_profiles.py, which also imports factor_fixture here; the resolution keeps both fixtures on the test_support.microcosm_build path, and each file imports exactly the names its tests use.
• Delta review of that fix (GPT-6.1 Sol, job 20261029-083505): APPROVE, no findings. Merging the fixed head into main differs from merging the previously reviewed head only in the fixture import lines of test_reference_factors.py and test_reference_profiles.py.
• CI then failed test_uk_target_compile_modules.py::test_every_direct_import_of_a_listed_module_is_listed_or_excluded: Add feed-derived NZ reference ageing factors (G8-r2a) #1382's 1d5fff4 made ledger_targets, a UK target compile module, import LexicalFloat from ledger_artifact. Add feed-derived NZ reference ageing factors (G8-r2a) #1382's b36c702 first added ledger_artifact to EXCLUDED_DIRECT_IMPORTS; 6d1252a merges it.
• 33c737d merges main after Allocate Census residuals through shared reference profiles #1375's landing. It changes no file, but Allocate Census residuals through shared reference profiles #1375 landed as 6af97bf, so this branch had two merge bases with main; GitHub marked the PR conflicting and ran no CI until this merge gave it one.
• The Sol delta review of that exclusion (job 20261029-095750) asked to list ledger_artifact instead: both UK target kernels decode their facts through it, so an edit to its decoder could move a UK target value without moving either kernel's implementation hash. Add feed-derived NZ reference ageing factors (G8-r2a) #1382's 7cd1d26 lists it in UK_TARGET_COMPILE_MODULE_NAMES (see Add feed-derived NZ reference ageing factors (G8-r2a) #1382); ae1f24b merges it.
• Delta review of the listing across all five heads (Opus 5.5 in session, since every Codex lane was held): APPROVE_WITH_NITS, no blocking findings (see Add feed-derived NZ reference ageing factors (G8-r2a) #1382).
• CI stops at the first failure, so about 358 of the 22,000-plus engine-free tests never ran in the failed rounds. They sit in the 29 files after the failing one in CI's list; the hub ran all 29 (711 tests) with CI's import path at the two stack tips, Add receipt contract v2 and selection-aware reference preflight #1397 and Reconcile regional Census family-type counts under d1932 #1398: 705 passed and 6 skipped at each.
• CI at ae1f24b: every job passed (gh pr checks exit 0; run 40896989678): engine-free, engine-uk and engine-us on Python 3.13 and 3.14, integration-uk, lint, select-countries and wheels.
• BE re-clear: Belgium's hub is away until 2026-10-31, so the NZ hub ran Belgium's own dump_be.py and its three BE suites (test_spec_engine_country_bundles.py, test_country_spec.py, test_spec_only_country_packages.py) at ae1f24b, alone and merged with Belgium's open Rewrite the Belgian country package for donor transport #1369 (dc157a7). The BE dump matches main's byte for byte; with Rewrite the Belgian country package for donor transport #1369 merged in, it matches main plus Rewrite the Belgian country package for donor transport #1369. The BE golden comparison holds in both, and all three suites pass with Rewrite the Belgian country package for donor transport #1369 merged in.

Merge order

#1375 → #1382 → this → #1395 (w2a, money normalization, stacked here) → w1 → S5 (#1392) → w2b.

🤖 Generated with Claude Code

[MaxGhenis]

Merged by the NZ hub (session local_e407a527) at ae1f24b with --merge. Gates: gh pr checks exit 0; MERGEABLE; not draft; no CHANGES_REQUESTED review. Reviewed head: ae1f24b. Independent delta review of the ledger_artifact listing across all five stack heads: APPROVE_WITH_NITS (Opus 5.5 in session), jobs/review-microcosm-fix5list/out-opus.md. BE re-clear clean at this head (jobs/be-reclear-1029e).