Add ordinal + calibration metrics, CV reporting, and CI-built model artifact

[AmirhosseinHonardoust]

Closes the remaining polish gaps from the previous overhaul.

Modeling / metrics

• Report cross-validated accuracy (0.429 ± 0.036) alongside the single-split
  0.483 so the headline isn't a lucky fold.
• Add ordinal-aware metrics (MAE, quadratic-weighted kappa) + log loss to
  metrics.json.
• Add src/ordinal.py (Frank & Hall) and a reproducible flat-vs-ordinal
  comparison (reports/metrics/ordinal_comparison.json). Flat RF wins on every
  metric, so it stays default; ordinal kept as a tested alternative.
• Add reliability diagram + Expected Calibration Error (src/model_metrics.py).
  RF ECE=0.09; sigmoid recalibration documented as a reported trade-off.

Ops

• models/*.joblib is now a git-ignored build artifact (removed from the repo).
• CI adds a pipeline job that runs generate->prep->train->evaluate->explain and
  uploads the model + reports as artifacts, exercising the training path in CI.

Tests: 12 -> 20 (ordinal classifier, metric helpers). Gate green locally:
ruff/black/mypy clean, 20 passed. Full pipeline reproducible & deterministic.