[professor3333]

Summary

• The second Kaggle run stopped when MLflow tracking started: UnsupportedModelRegistryStoreURIException ... unsupported URI 'sqlite:///…/mlflow.db'. MLflow 3 defaults to a SQLite store, and mlflow-skinny (the neural extra) cannot open one. MLflow 3.17 also refuses the file store unless MLFLOW_ALLOW_FILE_STORE=true.
• Unless MLFLOW_TRACKING_URI is set, the trainer now uses the local file store ./mlruns and allows it for this process only. That is the directory the runbook packs and the docs describe.
• Tests:
  • a new tracker test reproduces the exact failure before the fix;
  • the end-to-end CLI test (train → predict → experiment with a tiny model) now runs with MLflow on and checks mlruns/.
• docs/training_on_gpu.md: the mlflow ui command sets the same opt-in.

Checks

• ruff check, ruff format --check, mypy src/, pytest -m "not slow" pass.