Python 3.14 compatibility: person() accessor returns unweighted values in Microsimulation

[MaxGhenis]

Summary

When running with Python 3.14, the person() accessor method in Microsimulation contexts returns unweighted values instead of weighted values. This causes all tax calculations based on survey microdata to return incorrect (near-zero) values.

Reproduction

from policyengine_uk import Microsimulation

sim = Microsimulation()
# In Python 3.13: returns ~£86.7bn (weighted, correct)
# In Python 3.14: returns ~£0.11bn (unweighted, wrong)
print(sim.calculate("taxed_dividend_income", 2027).sum() / 1e9)

Expected behavior

person() should return weighted values in Microsimulation contexts regardless of Python version.

Actual behavior

• Python 3.13: person('taxed_dividend_income', '2027').sum() → £86.7bn (correct)
• Python 3.14: person('taxed_dividend_income', '2027').sum() → £0.11bn (wrong)

Impact

This breaks all microsimulation-based tax calculations. For example, dividend_income_tax returns £0 instead of ~£22bn when running with Python 3.14.

Individual Simulation objects work correctly in both Python versions - only Microsimulation with survey weights is affected.

Workaround

Pin Python 3.13 until this is fixed.

Environment

• Python 3.14 (tested with 3.14.0a6)
• policyengine-core (latest)
• policyengine-uk (latest)

[MaxGhenis]

Root Cause Found & Fix Confirmed

This is NOT a policyengine-core bug - it's a known NumPy bug with Python 3.14:

• numpy/numpy#28681: BUG: temporary elision heuristics are broken on Python 3.14

Technical Details

NumPy uses "temporary elision" optimization where it checks refcount == 1 to determine if an array can be modified in-place instead of allocating new memory. Python 3.14's new LOAD_FAST_BORROW bytecode instruction changes reference counting semantics, causing NumPy to incorrectly identify arrays as temporary when they're not.

In our case, when the dividend_income_tax formula does:

other_income = earned + savings  # Creates new array ✓
combined = other_income + dividends  # Should create new array, but modifies other_income in-place ✗

The second addition incorrectly modifies other_income in place, causing both calc(combined) and calc(other_income) to return the same value, making the tax difference = 0.

The Fix

Upgrade NumPy to 2.3.5 (or later). The fix was merged in PR #28748.

NumPy Version	dividend_income_tax
2.1.3	£0.00bn ❌
2.3.5	£22.55bn ✅

Action Required

For Python 3.14 support, policyengine-core should require numpy>=2.3.0.

This also explains why the issue only affects Microsimulation (where arrays are cached and reused in holders) but not single-household Simulation (where arrays are freshly computed each time).

[MaxGhenis]

Closing - this was a NumPy bug, not policyengine-core.

Root cause: NumPy's temporary elision optimization is broken on Python 3.14 due to the new LOAD_FAST_BORROW bytecode instruction.

Fix: Upgrade to NumPy 2.3.0+

See my detailed analysis in the comments above. The setup.py has been updated to allow numpy>=2.1.0,<3 which will pick up the fixed version.