Add robust PE-native loss and APG refit

[MaxGhenis]

Summary

• Add a shared robust PE-native loss helper using bucketed baseline-weighted Huber residuals.
• Use the same robust objective in PE-native scoring, target diagnostics, eCPS replacement comparison, and calibration benchmark sidecars.
• Replace the fixed-step quadratic projected-gradient refit path with monotone accelerated projected gradient plus adaptive line search so the refit loop can optimize the non-quadratic Huber objective while preserving the weight-budget simplex.
• Add --target-scope-filter national|state to the PE-native calibration benchmark/scoring path so the immediate eCPS-replacement run can optimize and rank on the national surface instead of being driven by state targets.

Evidence

On the alias-coalesced #170 H5, the robust national-only APG comparison artifact at /Users/maxghenis/CosilicoAI/microplex-us/artifacts/robust_loss_v1_alias_coalesce_national_apg_20260602/sound_ecps_replacement_comparison.json reported MP 0.013041312976726992 vs latest-local eCPS 0.016522169142140997, with objective/scoring identity and eCPS refit recovery both passing.

Tests

• uv run --python 3.13 --extra dev python -m pytest tests/pipelines/test_pe_native_calibration_benchmark.py tests/pipelines/test_pe_native_scores.py tests/pipelines/test_pe_native_loss.py tests/pipelines/test_ecps_replacement_comparison.py
• uv run --python 3.13 ruff check src/microplex_us/pipelines/pe_native_loss.py src/microplex_us/pipelines/pe_native_optimization.py src/microplex_us/pipelines/pe_native_scores.py src/microplex_us/pipelines/ecps_replacement_comparison.py src/microplex_us/pipelines/pe_native_calibration_benchmark.py tests/pipelines/test_pe_native_loss.py tests/pipelines/test_ecps_replacement_comparison.py tests/pipelines/test_pe_native_scores.py tests/pipelines/test_pe_native_calibration_benchmark.py