Fix IndexError in mut_param_dataset_correlation for 1-column groups

[jaredgalloway]

Summary

Fixes a latent IndexError in ModelCollection.mut_param_dataset_correlation that surfaces when a (mut_param, x) group has only one replicate with a non-zero entry — a regime that's reached under sparse shift solutions (e.g. continuation-strategy fits with strong fusion regularization).

The bug was at multidms/model_collection.py:1424:

"correlation": replicate_params_df.T.corr().iloc[0, 1] ** r,

When replicate_params_df.T has only one column, .corr() returns a 1×1 DataFrame and iloc[0, 1] raises:

IndexError: index 1 is out of bounds for axis 0 with size 1

What changed

• Extract the per-group correlation reduction into a small helper _pairwise_correlation(replicate_params_df, r) that returns NaN for 1-column groups instead of indexing past the matrix bounds.
• Add unit tests for both branches: the 2-column happy path and the 1-column regression case.

How this was discovered

Running the spike pipeline with strategy: continuation, tol=1e-5, maxiter=100, beta0_ridge=1e-3 against the prod fusionreg grid produced sparse shift solutions where some (mut_param, fusionreg) cells had a single surviving replicate, tripping the latent bug. The independent-strategy run on the same hyperparameters did not hit it because its converged solutions are denser. Fit pickles from the failed run were preserved, so re-running only the evaluate rule reproduces the issue cheaply.

Test plan

• pixi run fmt-check — clean
• pixi run lint — clean
• pixi run pytest tests/test_model_collection.py -k "pairwise_correlation or mut_param_dataset_correlation" — 5 passed
  • test_pairwise_correlation_two_replicates (new)
  • test_pairwise_correlation_one_replicate_returns_nan (new regression test)
  • test_mut_param_dataset_correlation (existing, still green)
  • test_mut_param_dataset_correlation_return_data_r1 (existing, still green)
  • test_mut_param_dataset_correlation_return_data_r2 (existing, still green)
• CI: ruff + black + full pytest suite + docs build
• End-to-end verification: re-run spike pipeline evaluate rule against the cached continuation fit_collection.pkl and confirm it now completes

🤖 Generated with Claude Code