Add calibration package checkpoint to unified_calibration pipeline

[baogorek]

Context

unified_calibration.py's run_calibration() currently builds the matrix and fits weights in one monolithic call. This means any change to fitting hyperparams (lambda, lr, epochs) requires re-running the expensive clone-by-clone simulation that builds the sparse matrix.

Proposal

Save a calibration package artifact to disk as a natural intermediate step within run_calibration(), before calling fit_l0_weights. The package should contain everything needed to run fitting independently:

• X_sparse (the calibration matrix)
• targets_df
• initial_weights
• household_id_mapping
• cd_household_indices
• target_names

This is the same interface already used in the Kaggle-based calibration workflow (pickle dict), just produced automatically by the pipeline.

CLI changes

• Add --build-only flag: stop after saving the package, skip fitting
• Add --package-path flag: load a pre-built package and skip matrix construction

Why this matters

1. Iteration speed: Tweak hyperparams without rebuilding the matrix
2. Resilience: If the GPU fitting crashes, you don't redo the matrix
3. Flexibility: The package can be downloaded and used in Kaggle, Modal, or locally
4. Modal runner: remote_calibration_runner.py can split into CPU (build) and GPU (fit) functions, with the package as the handoff artifact

Related

• Add calibration log to Github build artifacts #310 (calibration log as build artifact — the log is a natural output of the fit stage)
• PR Add census-block-first calibration pipeline (from PR #516) #531 (census-block calibration pipeline, where this design was discussed)

Implementation notes

• ~20-30 lines added to run_calibration() in unified_calibration.py
• Package format: pickle dict (matching existing Kaggle workflow) or npz + parquet
• The automated end-to-end pipeline remains the default; --build-only is opt-in

[baogorek]

Related to #310 — the calibration log is a natural output of the fit stage once the pipeline is split at the checkpoint.