Add epsilon-insensitive calibration target policy - #1054

Fixes #1053

Merge status

Do not merge this PR yet. It depends on the L0 epsilon-insensitive loss feature branch being merged and released, and we still need to decide whether this target-policy/tolerance pathway is the right long-term calibration design.

Summary

• Adds a resolved calibration target policy (target_policy.yaml + JSONL/summary artifacts) with explicit fail, warn, and diagnostic_only enforcement levels.
• Passes per-target weights, tolerances, and scale floors into the L0 relative_epsilon loss, while keeping grouped target loss disabled.
• Adds hard population anchors by deriving total person_count rows from age-bin person-count targets when no explicit total row exists.
• Publishes target-policy artifacts through the Modal calibration outputs and Stage 2 package contract metadata/parameters so builds can be interrogated later.
• Pins the US-data l0 extra to the temporary L0 feature branch until the L0 support is released.

Validation

• make lint
• /opt/miniconda3/bin/python -m pytest tests/unit/calibration/test_target_policy.py tests/unit/calibration/test_unified_calibration.py::TestDerivedPopulationAnchors tests/unit/calibration/test_unified_calibration.py::TestFitTargetGroups tests/unit/test_remote_calibration_runner.py::test_collect_outputs_returns_pipeline_artifact_bytes tests/unit/test_pipeline.py::test_calibration_package_parameters_track_matrix_mode tests/unit/test_pipeline.py::test_calibration_package_parameters_ignore_unused_matrix_options tests/unit/test_calibration_package_stage_contract.py -q (39 passed)
• PYTHONPATH=/private/tmp/policyengine-l0-epsilon /opt/miniconda3/bin/python scripts/run_epsilon_calibration_demo.py --epochs 300 --output /private/tmp/epsilon_calibration_demo_summary.json
• In the L0 feature branch: /opt/miniconda3/bin/python -m pytest tests/test_calibration.py -k 'relative_epsilon or zero_target_weight' -q (4 passed) and ruff check l0/calibration.py tests/test_calibration.py

Not run: docs guard tests, per explicit instruction in the implementation thread.

Follow-up

Replace the temporary git dependency with a released L0 package version after the L0 epsilon-insensitive loss branch is merged and published.