Fix Variable.clone() to preserve update_variable-inherited attributes (#502)

[anth-volk]

Fixes #502

Summary

Variable.clone() re-ran __init__ with baseline_variable=None, dropping every attribute a reform's update_variable merged in from the baseline (value_type, entity, formulas, …) and raising ValueError: Missing attribute 'value_type'. Because TaxBenefitSystem.clone() clones every variable — and the YAML test runner + branch calculations clone the system — any reform using update_variable crashed on clone.

Fix

clone() now mirrors TaxBenefitSystem.clone(): empty_clone + a __dict__ copy, preserving the variable's merged (baseline + override) state instead of re-deriving it. formulas/metadata are copied so the clone stays independent of the original (the prior __init__-based clone produced fresh containers). Not re-running __init__ also avoids a spurious check_computation_modes() re-validation.

Verified safe for all three Variable.clone() callers (get_annualised_variable, get_neutralized_variable, TaxBenefitSystem.clone) — they reassign containers or mutate scalars, never mutate formulas/metadata in place. No callers in policyengine-us/uk.

Commits (TDD)

1. Characterization tests for clone() (green on unchanged code).
2. Variable.clone() fails for variables registered via update_variable #502 regression tests as strict xfail (label-only + adds-only overrides).
3. The fix + un-xfail.
4. End-to-end: clone a reformed TaxBenefitSystem (the test-runner/branch path) + a runtime-attribute preservation/independence test.
5. Changelog fragment.

Verification

• New tests/core/variables/test_variable_clone.py: 7 tests.
• Full tests/core: 658 passed, 1 skipped, 1 xfailed (no regressions).
• Issue repro now succeeds:

  system = reform(CountryTaxBenefitSystem())
  system.get_variable("disposable_income").clone()   # was ValueError

🤖 Generated with Claude Code

[MaxGhenis]

Approving. Verified today (2026-09-02) against current master, not the 7/8 CI run:

• Merged this branch onto master @ 9125de0 locally (clean auto-merge; master gained #518's check_formula_determinism call in Variable.__init__, which a __dict__-copy clone correctly does not re-run). tests/core: 707 passed, 1 skipped, 1 xfailed.
• Ran the real callers of the changed path against the merged core: policyengine-us reform YAML suites that neutralize_variable (FISC Act, Family Security Act HoH removal, End Child Poverty Act, EDAA, state dependent-exemption repeal) — 28 passed. policyengine-uk test_behavioral_responses.py + test_labour_supply_response_formulas.py (branch → Simulation.clone → TaxBenefitSystem.clone) — 6 passed, 11 skipped. The 6 failures in test_labour_supply_responses.yaml reproduce identically on stock core (dict + list TypeError), so they are pre-existing and unrelated.

Semantics check: empty_clone + __dict__ copy is exactly the TaxBenefitSystem.clone pattern; formulas/metadata copied, adds/subtracts shared as before. One behaviour change worth naming because it is an improvement, not a regression: the old __init__-based clone reset instance state to class defaults, so system.clone() silently dropped is_neutralized=True (and any replaced formulas) from every variable. Clones now preserve runtime state, which is what every caller (get_neutralized_variable, get_annualized_variable, branches, the YAML runner) actually wants.

The CI on this PR is stale (July 8, still on the py3.9/3.10 matrix); a merge of master into the branch refreshes it.