Make zero-valued calibration targets intentionally nonzero-weighted

Zero-valued targets should still receive nonzero calibration loss weight.

Why:

• A zero target can be meaningful, especially at small geography levels or sparse breakdowns.
• If a state or local target is zero but the dataset estimates a positive value, the optimizer should see that error instead of silently ignoring the row.
• This also makes target diagnostics easier to reason about when expanding the target surface.

Current context:

• The current fiscal refresh loss weighting appears to use max(abs(target), 1.0) before applying sqrt target-value weights, so zero targets are not literally zero-weighted today.
• We should make that behavior explicit and protected by tests/diagnostics, not incidental.

Acceptance criteria:

• Zero-valued amount and count targets get a finite positive loss weight.
• A regression test covers zero-valued targets under the active amount/count 50/50 target-weighting scheme.
• The target-weight explorer or calibration diagnostics exposes the number of zero-valued targets and their assigned weight range.
• Document the intended zero-target treatment near the loss-weighting code.

Start at the current fiscal refresh loss-weighting code and trace the active amount/count 50/50 target-weighting scheme. Then inspect the target-weight explorer or calibration diagnostics; done means zero-valued amount and count targets have finite positive weights, regression coverage exists, zero-target counts and weight ranges are exposed, and the intended treatment is documented nearby.