Weight fiscal calibration loss by target value basis

[MaxGhenis]

Summary

• remove initial-estimate dependence from default target loss scales; default scale is now max(abs(target), 1)
• weight US fiscal refresh target rows by target value within structured measure bases (amount, return_count, person_count)
• keep semantic critical-national-total boosts and state-row downweighting, normalized to mean row weight 1
• add regressions for return/claim count classification and for ignoring initial estimates in default scales

Objective

For compiled target row i and weights w:

estimate_i(w) = A_i @ w
scale_i = max(abs(target_i), 1)
row_weight_i = normalize_to_mean_1(value_weight_i * semantic_i * geography_i)
value_weight_i = max(abs(target_i), 1) / mean_j_in_same_basis(max(abs(target_j), 1))
loss(w) = sum_i row_weight_i * min(abs(estimate_i(w) - target_i) / scale_i, 10) / sum_i row_weight_i

Initial estimates are diagnostics only; they no longer set permanent target scale.

Validation

• uv run --frozen --all-packages --python 3.13 python -m pytest -q
• uv run --frozen --all-packages --python 3.13 ruff check packages/populace-calibrate/src/populace/calibrate/solve.py packages/populace-calibrate/tests/test_solve.py tools/build_us_fiscal_refresh_release.py packages/populace-build/tests/test_us_fiscal_refresh_builder.py
• git diff --check
• /cycle read-only subagent review: first pass found SOI returns/claims basis bug; fixed. Second pass found no actionable issues.

External Review

Requested ChatGPT Pro review via in-app browser, but chatgpt.com is currently blocked behind Cloudflare human verification. Pending user confirmation to solve that verification.