Prefer capital gains components in PUF top-tail guard

[MaxGhenis]

Summary

• make the PUF support-clone top-tail guard prefer PE-used capital-gains components (long_term_capital_gains_before_response / long_term_capital_gains plus short_term_capital_gains)
• fall back to stored capital_gains only when component variables are absent
• add a regression test for the current failing shape: low stored capital_gains with high LTCG still entering the open-ended AGI bin

Why

The post-#167 rebuild still regressed on the same target:

• MP latest-local loss: 1.2450774476685147
• eCPS latest-local loss: 0.7685371816961414
• top target: state/MD/adjusted_gross_income/amount/1000000_inf

The guard had affected_rows=0 because it used stored capital_gains ~= $14.3M for the bad MD row. PolicyEngine-US does not treat stored capital_gains as authoritative here: during simulation setup it copies long_term_capital_gains into long_term_capital_gains_before_response and removes capital_gains from input variables, so formulas use LTCG/STCG components. The same row had long_term_capital_gains ~= $95.0M, which should have triggered the guard.

Validation

• uv run ruff check src/microplex_us/pipelines/us.py tests/pipelines/test_us.py
• uv run ruff format --check src/microplex_us/pipelines/us.py tests/pipelines/test_us.py
• uv run --extra dev python -m pytest tests/pipelines/test_us.py -k 'puf_support_clone_top_tail_guard' (5 passed)
• uv run --python 3.13 --extra dev --extra policyengine python -m pytest tests/pipelines/test_us.py (179 passed, 6 expected warnings)

[MaxGhenis]

Local proof against the completed #167 artifact seed data:

• input: /Users/maxghenis/CosilicoAI/microplex-us/artifacts/ecps_shaped_cps_puf_support_clone_export_top_tail_guard_no_acs_efc1fd8_20260602/.../seed_data.parquet
• patched rough AGI variables use LTCG/STCG components, not stored capital_gains
• guard summary on that seed: affected_rows=1
• max rough AGI before: 95,107,558.38
• max rough AGI after: 78,999,999.0
• rows above cap before: 1
• rows above cap after: 0

So this patch catches the same artifact shape that #167 missed before we spend another rebuild.