De-hardcode partner correlation: replace attribute-name branching with metadata/policy resolver

[RandomOscillations]

Summary

Partner correlation currently uses hardcoded attribute-name checks (age, race_ethnicity variants, country) to select correlation behavior. This is brittle and not generalizable across arbitrary specs.

Why This Matters

Behavior should depend on declared semantics/policy, not string names. Renaming equivalent attributes can silently change sampling dynamics.

Current Behavior (Code)

In /Users/adithyasrinivasan/Projects/extropy/extropy/population/sampler/households.py:

• attr_name == "age" -> gaussian offset (line ~102)
• attr_name in ("race_ethnicity", "ethnicity", "race") -> same-group rates (line ~114)
• attr_name == "country" -> same-country logic (line ~127)

Proposed Fix

1. Introduce policy-driven partner correlation resolution:

• policy keyed by metadata (semantic_type, identity_type, explicit partner policy)
• deterministic resolver chooses algorithm (gaussian_offset, same_group_rate, same_value_probability)

2. Keep backwards compatibility by mapping legacy names to policy in one migration layer, then deprecate name branching.

3. Update validation to ensure required policy metadata exists for scope=partner_correlated attributes.

Tests

• equivalent semantics with different attribute names produce equivalent partner-correlation behavior
• missing policy metadata is flagged in strict mode
• legacy specs continue to work with migration mapping

Acceptance Criteria

• No runtime behavior depends on hardcoded attribute name strings.
• Partner correlation is fully metadata/policy-driven.

Pipeline Impact

Critical for generalization across diverse populations and scenario-extended schemas.

[RandomOscillations]

Implemented on branch codex/steps1-3-hardening (see PR #124). Keeping this issue open until merge + integration verification completes.