Add source-only New Zealand Axiom pilot

[MaxGhenis]

Fixes #507

Adds the source-only integration slice of #506. The broader NZ population/release/scoring work remains open.

Scope

• Add opt-in NZ dataset and Axiom TaxBenefitModelVersion classes through the shared Simulation surface; no pe.nz bundle, new country-model package, or US/UK pin changes.
• Consume Microcosm's actual entity-table HDF5 codec, including exact Decimal and nullable-boolean inputs, with explicit NZ tax-year dates.
• Resolve effective person/family weights through Frame from household weights; validate membership, substantive inputs, and formula-owned exclusions.
• Execute the real Axiom official-budget module, preserve source files, and retain input/runtime hashes and the unresolved fiscal-comparability bridge in output receipts.
• Preserve direct NZ model/dataset JSON round-trips; reject saved model identity drift. Generic Simulation base-type JSON behavior is documented and unchanged.
• Include an executable source-artifact example and focused regression tests.

Dependencies and limits

Depends on PolicyEngine/microcosm#821 and TheAxiomFoundation/rulespec-nz#117. The latter still needs signed Treasury corpus publication/activation.

This pilot does not construct or certify a calibrated NZ population, impute missing family inputs, implement fiscal/payment/debt bridges, or publish national scores. Annual entitlement changes must not be attached as completed Treasury budget-score replications.

Verification

• make format and make lint pass.
• 36 focused tests pass under both pandas 2.3.2 and pandas 3.0.5 with actual Microcosm HDF5 writer/reader and native Axiom execution. Regression coverage includes exact large integers and string missing-value semantics through columns, indices, and categories. A separate pandas 3.0.5 source-only review run passes 35 with one opt-in runtime test skipped.
• Verified native Axiom core bb4b568 and rulespec-nz 3b663b3; the executable records source/binary hashes, not a wrapper version as proof of the native version.
• Real-runtime test uses two synthetic families with nonuniform household weights. It verifies integration, not a national estimate.
• Independent re-review approved all changes after fixes to saved identity, Decimal padding, JSON precision, and pandas 3 string missing-value preservation. The final 36-case real-runtime runs pass against Microcosm 9f7deadd after its upstream-root-authority merge. The pandas 3 correction addresses the two JSON round-trip failures found by full CI; the full remote suite is rerunning on the new commit.

Leave this PR draft. It is the bounded source-only adapter tracked in #507, not completion of #506.