Summary

Adds opt-in MLflow tracking to the jabs-cli cross-validation command. Each run can record aggregate cross-validation metrics, run parameters, descriptive tags, and the training report as an artifact, so cross-validation runs of a behavior can be compared over time. MLflow is a fully optional dependency — the base install and all existing behavior are unchanged when it isn't used.

Tracks KLAUS-444.

What's included

New module jabs.classifier.mlflow_logging

• log_cross_validation_to_mlflow(...) — creates one MLflow run, logs metrics/params/tags + the report artifact, returns (run_id, tracking_uri).
• Helpers: aggregate_cv_metrics, build_params, build_tags, resolve_experiment_name, parse_kv_tags, load_env_file, mlflow_available, and MlflowLoggingError.
• import mlflow is lazy (inside the logging function only), so the base package never depends on it.

Optional dependency — new mlflow extra: pip install 'jabs-behavior-classifier[mlflow]'.

CLI options on cross-validation

• --mlflow [ENV_FILE] — enable logging; optional .env file with MLFLOW_* connection settings (ambient env if omitted).
• --mlflow-experiment NAME — override the experiment (see below).
• --mlflow-tag KEY=VALUE — repeatable free-form run tags.
• --mlflow-no-report — skip the report artifact (metrics + params only).

Per-behavior experiments — runs default to experiment jabs-<behavior> so a behavior's runs form their own leaderboard (mixing behaviors isn't comparable). Precedence: --mlflow-experiment → MLFLOW_EXPERIMENT_NAME → jabs-<behavior>. The experiment is auto-created.

Leaderboard metrics — cv_f1_behavior_mean, cv_accuracy_mean, precision/recall (mean + std), iteration count, and dataset composition are logged as MLflow metrics, so the experiment's runs table is sortable by mean F1. Full per-fold detail rides along as the report artifact.

Exit codes & failure handling

• Logging runs after results are printed and the report is saved.
• --mlflow requested but the mlflow extra not installed → fail fast with an error before running cross-validation, exit 1. Logging was explicitly requested but can't be honored, so the command stops rather than silently producing a run with no logging; install the extra (or drop --mlflow) and re-run.
• --mlflow ENV_FILE path doesn't exist → fail fast before running cross-validation, exit 1. The env-file path is validated up front (with a leading ~ expanded), so a typo is caught immediately rather than after the run.
• Push fails (server/auth/TLS) → results/report preserved, warn, exit 3 (distinct from the generic 1).

Docs — both copies (online + in-app cli-tools.md) gain a jabs-cli cross-validation command section and a detailed MLflow integration section (install, enabling, connection config, experiment selection, leaderboard, tags, exit codes).

Example

jabs-cli cross-validation /path/to/project --behavior grooming \
    --mlflow settings.env --mlflow-tag purpose=baseline
# -> logs to experiment "jabs-grooming"

Testing

• New tests/classifier/test_mlflow_logging.py (logging module, with a fake mlflow injected — no server/network) and MLflow CLI option-parsing tests in tests/scripts/test_cross_validation_cli.py.
• Full tests/classifier/ + tests/scripts/: 305 passed. ruff check/format clean.