Add a small Random Forest baseline demonstration

[kpernyer]

Add a runnable CPU-only Random Forest example for commitment-outcome classification, using scikit-learn in an optional pinned ML environment. The default run generates 300 explicitly fictional tabular snapshots, fits a fixed forest, and compares its probability estimates with the historical training fulfilment rate on a later test period.

An explicit feature allowlist excludes identities and outcome fields. Training labels must be available by the fit cutoff; unknown, disputed, censored, and unavailable labels are accounted for separately. The optional JSON report records generation provenance, split dates, model settings, exclusions, and scored test predictions. The generator's assumptions and limits are documented; it does not label the challenge cards or extract features from the event ledger.

The guide also positions TabFM and TabPFN as future tabular comparisons, with primary-source links. No PyTorch, GPU, provider API, model-weight download, real dataset, or empirical organisational claim is introduced.

Validation: 22 local tests passed, existing 11-event fixture validated, full demo completed, and documentation links/whitespace checked. A dedicated Python 3.14 ML CI job runs the optional dependencies and full suite; base event validation remains on Python 3.11 and 3.14.