[MaxGhenis]

Status

Draft. This PR fixes three findings from the cloud review of #1440. It changes text and tests only: one doc paragraph, one test file, the generator's docstring, the reason on two overrides in as_rate_bridge.json, one column description in benefit_unit_rule.json, the golden that pins both resources' hashes, and a changelog fragment. No kernel, node declaration or override changes, and all three builders' --check pass. No output value changes, but the PR re-keys the whole NZ graph. as_rate_bridge.json's new hash alone re-keys the 22 nodes that bind it, and with their descendants 92 of 253. benefit_unit_rule.json's new text and hash change the params of the 11 nodes that select it. nz.create is one of them and every node descends from it, so all 253 node keys move. The first real NZ run comes at G8-x, so these keys invalidate no stored NZ run.

What changes

• docs/transport-entitlement.md and the generator's docstring: both placed the masks, gross rate and cutout "on each population that runs" the Accommodation Supplement engine. The merged graph never did that: the shared nz.as.bridge population runs them once for S0, S1, S2, S3 and S5, and each variant repeats them on its own housing population. Both now say so, and one test's name drops the same wording.
• Childless-family tests: test_unmasked_youngest_child_override_carries_a_missing_age_to_the_engine claimed the adapter's int64 conversion raises. It doesn't on these frames. A childless family's null reads as NaN and turns the override column into float64, so _batch_from_table takes its float branch. With the mask, the engine receives 0.0 for that adult; without it, NaN. Both tests cast the column themselves, so neither checked what the adapter does. They now read _batch_from_table and pin the float64 dtype.
• as_rate_bridge.json: two reason strings gave the int64 read as why the override takes a mask. They now give the real reason: a null age would reach the engine input as NaN.
• benefit_unit_rule.json: the column description said the youngest-child age holds 0 for a childless family. unit_construction.py writes a null, so it now says null.
• Mixed couple through S5: as_rate_bridge.json says a partner on a main benefit takes the NZ Superannuation amount, through max over members. No test in test_nz_as_bridge.py ran that. A new test runs it in both person orders, using a new partner_age field on the fixture's Unit. It covers the max half of that statement. The fixture flags both adults receives_nz_super, which changes nothing here: these nodes read receipt_group and no receives_* column; test_receipt_any_member_priority_and_mixed_count pins how the classifier groups a mixed couple.

Invariant

S5 mixed couple. With no dependent child, a couple's S5 base rate equals twice the superannuitant's amount, whichever partner comes first in person order.

Results

The hub ran these with no pytest, because the host sits under its disk floor. CI runs the file as usual.

Run	Result
Control: six test cases	all pass
Mutant: S5 base rate aggregates with first	the (50, 70) case fails; (70, 50) passes, so the test needs both orders
Mutant: S5 base rate aggregates with sum	both mixed cases pass; test_a_partnered_superannuitant_unit_takes_coefficient_two_in_s5 fails, so the existing test covers sum

Left out

• The review's finding on cents at an exact tie goes to G8-x, because it depends on how the pinned engine computes nets. Declare the NZ transport and entitlement graphs #1440's body records it.
• The review's finding on an empty region string needs no change. _bind lets a layer read only code columns, and decode_atomic_support refuses empty codes, so geography.derive@1 can't write "".

🤖 Generated with Claude Code

[MaxGhenis]

Merged by the NZ hub (session local_e407a527) at 082f60f with --squash. Gates: gh pr checks exit 0; MERGEABLE; not draft; no CHANGES_REQUESTED review. Reviewed head: 082f60f. Independent review: an Opus 5.5 reviewer read d901f0d, 1a58826 and 082f60f, because the Codex lanes are closed while the host sits under its disk floor. Final verdict APPROVE_WITH_NITS, Needs Max no. The hub applied every finding; the last one corrected the PR body's re-key sentence and changed no file. Direct run at this head: six cases pass, and the first and sum mutants fail as the PR body states. All three NZ builders pass --check. The golden moves only the two resource hashes and the fingerprint.