[MaxGhenis]

Tests only. This PR carries cloud review 220's nit 1 on #1507 and review 216's nits 2 and 3 on #1488; each needed its PR in main first.

What changes

• The P1 rounding test (test_nz_reference_aging.py, review 220's nit 1).

  • The gap. The old test_p1_rounds_once_and_the_truncated_numerator_refuses used p1_ty2026_to_ty2027_central. Truncation changes that factor's numerator (1029852 against 1029853), but rounding N to cents, rounding the projected quarters to cents and rounding q to six digits all give 1029853, so the test couldn't tell one rounding from any of them.
  • A new helper, _p1_numerators, computes a P1 profile's numerator in millionths at 60 digits under one half-even rounding, and under truncation and each of those three intermediate roundings.
  • test_p1_rounds_once_and_refuses_truncated_or_intermediate_numerators runs for the low, central and high profiles. It checks that one rounding gives _EXPECTED's numerator, then declares each alternative that differs and expects the resolver to refuse it ("expected factor mismatch"). It needs the feed, so CI skips it.
  • test_p1_profiles_separate_one_rounding_from_each_alternative needs no feed, so it runs in CI. It checks that across the three profiles every alternative differs from one rounding at least once.

  Profile	One rounding	Truncated	N to cents	Quarters to cents	q to six digits
  low	1028904	1028903	1028901	1028903	1028904
  central	1029853	1029852	1029853	1029853	1029853
  high	1030799	1030799	1030800	1030800	1030800

• The docstring of test_every_calibration_node_has_its_own_standin_as_the_only_writer (test_nz_transport_declarations.py, review 216's nit 2). It said composition checks the measurement proof on the skeleton alone. Since Prove extension calibration nodes' measurement producers in compose #1488, compose_transport_graph also runs _check_measurement_proofs on the extended graph with extension_nodes, where the compiled graph has to prove each nz.v.<V>.calibrate producer.

• add_measurement goes (nit 3), with its two calls in _nz_composition_spec_and_config. It inserted a compile-only producer "until G declares one". NZ's packaged transport_graph.json and all eight calibration scenarios in scenarios.json now declare ird_taxable_income_ty2027_v1 through their stand-ins (nz.income.standin, nz.v.<V>.income.standin), so its early return fired on every call.

• Formatting. NZ: re-declare the ageing factors from the September 2026 QES #1507 left one assert unformatted in each of test_nz_reference_aging.py and test_nz_money_normalization.py. ruff format (0.15.16, the version CI pins) now formats both files, and neither syntax tree changes. CI's lint job runs ruff check, which doesn't check formatting.

Invariants

• The resolver refuses differing numerators. For each P1 profile, the resolver accepts the numerator from one half-even rounding of the exact ratio and refuses every alternative scheme whose numerator differs.
• The three profiles together separate every alternative. The feed-free test checks this at the encoded 2.7% central growth, the PREFU 2026 rate that the generator requires (tools/generate_nz_reference_aging.py:131). A growth re-declaration has to update that literal in _p1_numerators too; the generator's check and test_root_enclosure_certifies_no_rounded_quarters_and_irrational_root carry the same 2.7. With the new numbers in place, the test fails if they stop separating any of the four schemes.
• No runtime file changes. _MEASUREMENT stays, since six other tests in the file use it. With add_measurement's early return firing on every call, removing it leaves each graph's node list as it was.

Checks

• Local, with NZ's pinned feed 89178aaf. The hub ran these one process at a time under its memory guard (-n 0). The tree was 78907367, a local commit of main 3e7a7045 plus NZ: re-declare the ageing factors from the September 2026 QES #1507 plus this change. Main has since gained Test low counts and digests, and anchors listed second, in later reports #1513 and Sign off the local cuts' target-fit rows (dbacb023, 83802786), and the company car base #1491, and neither changes an NZ file:
  • test_nz_reference_aging.py: 25 passed in 19.5 s, none skipped, so the feed-gated refusal tests ran;
  • test_nz_transport_declarations.py: 62 passed in 180 s.
• CI: run 41931933866 on e48af5a0 passed all 11 checks. The 3.13 engine-free leg's pytest command names all three changed files and reports 25,886 passed, 291 skipped, 3 xfailed and none failed. The feed-gated tests skip there; the separation test has no skip path, so it ran. Main has since gained Add binding-local views to the Axiom adapter, registry, compose and closure (S2b PR 1) #1482, which shares no file with this PR; the merge audit comment gives the check.
• Review: GPT-6.1 Sol (Subfleet job 20261104-170839-nz-nits-review) reviewed the same patch (c78112a7e..83b4fe2d9; this PR's patch-id matches). Verdict: APPROVE_WITH_NITS, needs Max: no.
  • It recomputed all 15 numerators in the table above and traced each refusal case to the resolver's factor comparison (reference_factors.py:611-614).
  • It confirmed that each incorrect resolver fails a profile: truncation fails low and central; N or the quarters rounded to cents fails low and high; q rounded to six digits fails high.
  • It confirmed that add_measurement returned early on all nine calls, and that the formatting hunks leave the syntax trees unchanged.
  • Its two nits concerned this body's wording, and this body carries both fixes.

🤖 Generated with Claude Code

[MaxGhenis]

Merged by the NZ hub (session local_e407a527) at e48af5a with --merge. Gates: gh pr checks exit 0; MERGEABLE; not draft; no CHANGES_REQUESTED review. Reviewed head: e48af5a. Review: the GPT-6.1 Sol review (Subfleet 20261104-170839-nz-nits-review) returned APPROVE_WITH_NITS with needs Max: no and asked for no code change; body r3 carries both prose fixes. Head e48af5a has patch-id 2d3aa4a50f3731b6, equal to the reviewed c78112a7e..83b4fe2d9. CI run 41931933866 on e48af5a, whose parent is main fc5bb90, so the run tested this head's own tree: 11 of 11 passed. Main has since gained #1482 (5c57c51), which shares no file with this PR, and the merge is clean (tree c9691037). The removed add_measurement helper returned at once on both fc5bb90 and 5c57c51: the transport graph and all eight calibration scenarios already hold the ird_taxable_income_ty2027_v1 row, in blobs #1482 leaves unchanged, and load_transport_spec returns that JSON as it stands. The rest is a docstring, ruff format's re-wrap of one assert in each of two files, and the P1 rounding tests, which import nothing #1482 changed. The Belgium hub ran main's copies of all three files on 5c57c51's tree abff51a7, and they pass; it also confirmed the docstring describes #1488 correctly. A merge commit keeps the reviewed commit message.