Redesign calibration loss as capped weighted MAPE

[MaxGhenis]

Summary

Redesigns Populace calibration loss from squared bounded relative error to capped weighted MAPE:

loss = weighted_mean(min(abs((estimate - target) / scale), cap))
scale = max(abs(target), abs(initial_estimate), 1)
cap = 10  # 1000% per target

This fixes the a912 failure mode where zero-valued/tiny fiscal rows dominated the objective and prevented meaningful movement on major targets like CTC.

Changes

• Adds fixed target-loss scales in populace.calibrate, defaulting to max(abs(target), abs(initial_estimate), 1) after matrix compilation.
• Caps each target's scaled miss at 1000%.
• Lowers the default Adam learning rate to 0.02, which is better behaved for absolute-loss gradients.
• Removes the obsolete compiled-target -1 guard from matrix compilation; the new loss has no (target + 1) denominator.
• Changes the US fiscal refresh builder from raw dollar target weighting / CTC-only boost to semantic weights:
  • national critical fiscal roles get higher priority,
  • state rows are downweighted,
  • state rows do not inherit national role boosts.
• Updates PUF aggregate diagnostics and docs to report the new loss formula.

Sanity check

On the published a912 diagnostics, the old objective's top 10 rows accounted for ~99.9999989% of total loss. Under the new capped MAPE scoring on the same final estimates, the top 10 rows account for ~0.6%, so loss is no longer owned by a handful of zero/tiny rows.

Verification

• uv run python -m pytest -q passed
• uv run ruff check ... passed on changed Python files
• git diff --check passed
• /cycle review: first pass found state rows were still getting national role boosts; fixed with regression coverage. Second pass found no actionable issues.