Fix #86: normalize promoted dependent attributes for age realism

[DeveshParagiri]

Summary

• Seed dependent promotion sampling with fixed age/gender so downstream sampling evaluates with child context
• Add age-aware normalization for promoted minors (education/employment/occupation/income) using spec-driven attribute/options heuristics
• Keep household-scoped economics inherited from household context
• Add focused regression tests for promoted minor realism

Validation

pytest -q tests/test_agent_focus.py::TestPromotedDependentQuality

Closes #86

---

Recreated from closed PR #93 (base branch was deleted)

---

⚠️ CHANGES REQUESTED - DO NOT MERGE

This PR uses fragile keyword matching for attribute type detection:

_INCOME_KEYWORDS = ("income", "salary", "wage", "earnings", "compensation")
if any(token in key for token in _INCOME_KEYWORDS):
    # normalize income for minors

Problem: Silently fails if LLM names attributes differently:

• earning (singular) → ❌ not matched
• pay, annual_pay → ❌ not matched
• diploma, schooling → ❌ not matched

Suggested fix: Add semantic_type field to AttributeSpec for explicit declaration.

See full review: #93 (review)

[DeveshParagiri]

Superseded by #106 which implements a proper LLM-driven approach for semantic_type classification.