Fix silent data corruption for invalid enum values

[MaxGhenis]

Summary

Invalid enum values passed to Enum.encode() were silently converted to index 0 (the first enum value) instead of raising an error. This caused silent data corruption in simulations.

The Problem

In Enum.encode(), numpy.select() is used to map string values to enum indices:

array = numpy.select(
    [array == item.name for item in cls],
    [item.index for item in cls],
)

When no condition matches (invalid value), numpy.select returns its default value of 0, which silently converts invalid values to the first enum item.

For example:

• 'MARRIED' passed to FilingStatus (which has SINGLE, JOINT, SEPARATE) would silently become SINGLE (index 0)
• Empty string '' would also silently become the first enum value

Real-World Impact

This bug was discovered in PolicyEngine/policyengine-us#6901 where Missouri test files had:

• filing_status: MARRIED (invalid - should be JOINT)
• filing_status: (empty/blank)

These tests passed silently because the invalid values were converted to SINGLE.

The Fix

Add validation after numpy.select to detect unmatched values and raise a descriptive error:

unmatched_mask = ~numpy.isin(original_array, valid_names)
if unmatched_mask.any():
    invalid_values = numpy.unique(original_array[unmatched_mask])
    raise ValueError(
        f"Invalid value(s) {invalid_values.tolist()} for enum "
        f"{cls.__name__}. Valid values are: {valid_names}"
    )

Test Plan

• Added tests/core/test_enum_encoding.py with tests for:
  • Valid single values encode correctly
  • Valid multiple values encode correctly
  • Invalid values raise ValueError with helpful message
  • Empty strings raise ValueError
  • Mixed valid/invalid values raise ValueError
  • Case sensitivity is enforced
  • Order is preserved in encoding

Fixes #410

🤖 Generated with Claude Code