Fix AGI-targeted geography assignment

[MaxGhenis]

Summary

• fix AGI-weighted block probabilities so district totals match AGI target shares instead of population-share times AGI target
• load district AGI targets through UnifiedMatrixBuilder using the caller-provided calibration database
• add focused tests for the AGI probability math and calibration wiring

Testing

• uv run pytest policyengine_us_data/tests/test_calibration/test_clone_and_assign.py policyengine_us_data/tests/test_calibration/test_unified_matrix_builder.py policyengine_us_data/tests/test_calibration/test_unified_calibration.py -q
• uv run ruff format --check policyengine_us_data/calibration/clone_and_assign.py policyengine_us_data/calibration/unified_calibration.py policyengine_us_data/calibration/unified_matrix_builder.py policyengine_us_data/tests/test_calibration/test_clone_and_assign.py policyengine_us_data/tests/test_calibration/test_unified_calibration.py policyengine_us_data/tests/test_calibration/test_unified_matrix_builder.py

Follow-up to #671.

[baogorek]

Review (on behalf of Ben)

Both bug fixes look correct:

Math fix in _build_agi_block_probs: The old code (pop_probs * agi_weights / sum) conflated population distribution with AGI targets — a district with 90% of blocks but a low AGI target would still dominate. Dividing by within-district population mass first makes district totals proportional to just the AGI targets while preserving within-district block shares. The new test case makes the expected behavior clear.

db_path shadowing fix: Line 944 on main hardcodes db_path = str(STORAGE_FOLDER / "calibration" / "policy_data.db"), which shadows the db_path parameter. Moving UnifiedMatrixBuilder construction earlier and routing through get_district_agi_targets() fixes this cleanly.

Questions / minor notes

1. Silent filter changes in get_district_agi_targets vs the old raw SQL. The old query returned all reform IDs, all domain scopes, and didn't do best-period selection (the dict just kept the last row arbitrarily). The new method via _query_targets adds three filters that weren't there before: baseline only (reform_id == 0), unconditional only (empty domain_variable), and most-recent-period selection. All three seem like improvements, but the PR description doesn't mention them — were these intentional?

2. builder.dataset_path mutation (line ~1027 in the new code): the builder is constructed without dataset_path, then it's set as an attribute later. This works fine today since __init__ just stores it, but it's a minor fragility — if construction-time validation is ever added, this pattern would break. Not a blocker, just flagging it.

[baogorek]

Heads up — CI is failing with:

ValueError: Cannot downsample cps_2021: found 9 dataset variables missing from the current country package (count_under_18, count_under_6, hourly_wage, is_union_member_or_covered, other_type_retirement_account_distributions, ...).

I looked into it but couldn't figure out the root cause. Wanted to flag it so you're aware.

[baogorek]

LGTM