[hallelx2]

Found by reading advanced_traceback.txt before the rebuild reaches this step, rather than waiting for it to fail again.

Exception: Run with UUID 8ec8e4919f6d403bb6bc8ae0bb9a1acb is already active.
To start a new run, first end the current run with mlflow.end_run().
To start a nested run, call start_run with nested=True

  cancerag/ml/advanced.py:73        summary["phases"]["mlflow"] = log_all_runs()
  cancerag/ml/mlflow_logging.py:58  with mlflow.start_run(run_name=...)

cancerag.pipeline opens an MLflow run for the pipeline as a whole (MlflowTelemetry, on unless --no-mlflow). log_all_runs then opened its own top-level run per CV fold, and MLflow refuses a second active top-level run — so the first row raised and took the step with it.

This is a second terminal failure, independent of the figures one. It explains why --list reported advanced no while optuna_trials.csv, lightgbm_tuned_final.joblib and stacking_meta.json were all on disk: Optuna finished, then run_all fell over on its final MLflow phase, so advanced_summary.json was never written. The step looked un-run rather than failed.

Change

All five call sites go through a _record helper that nests when a run is active and stays top-level when none is. Nesting is the right relationship, not a workaround — these per-fold and per-experiment records genuinely are children of the pipeline run. Run standalone (python -m cancerag.ml.mlflow_logging), behaviour is unchanged.

The tracking URI and experiment are also no longer set under an active run: they are already fixed by whoever opened it, and a nested run belongs to its parent's experiment regardless, so setting them only misreported where the records land.

Verification

• The tests were checked against the unfixed code. Reverting the helper to a plain start_run reproduces the exact production exception:

  FAILED test_logs_runs_inside_an_active_run - Exception: Run with UUID … is already active
  FAILED test_nested_flag_follows_whether_a_run_is_active - KeyError: 'nested'
  

  A test that passes against the bug would not be worth having.
• pytest tests/unit — 474 passed, 31 skipped (was 470; +4 here)
• ruff check clean on both touched files (also dropped two pre-existing unused imports)
• test_every_call_site_goes_through_the_helper greps the module, so a sixth mlflow.start_run added later reintroduces nothing silently
• Asserts the outer run survives with the same run_id — log_all_runs must not close the pipeline's run

The rebuild is in model_training and reaches advanced in roughly an hour; this goes to the host before then.

Closes HAL-4159