Save prediction artifacts in TabArena's result format

[adrian-prior]

Generated by Codex

Save per-config prediction artifacts using TabArena's raw result schema so they can support offline analysis and later storage integration. Each completed results.pkl contains validation/test predictions, labels, config metadata, metric errors and timings; temporal row identities and fit provenance live under relarena. Public loaders validate row context and array integrity, and the writer rejects unsupported task types.

Format and scope

• Uses data/<model_config>/<dataset__task>/<repeat>_<fold>/results.pkl, with the seed identifying a repeat of temporal fold 0.
• Labels are included in each config artifact. Loading requires trusted pickle files.
• Validation-only outputs use validation.partial until test predictions are available, so pickle discovery does not treat them as completed results.
• Preserves temporal holdout semantics and existing run fingerprints without hashing input tables again.
• Checks expected row order and array shape; models must return predictions in input-table order. Plain prediction arrays cannot expose a model-side permutation.
• Local artifact I/O only. Upload/download integration and resumable prediction caching are outside this change.
• Runner/CLI integration and end-to-end tests are in Export prediction artifacts in TabArena's result format #42.

Testing

• Latest lower-layer check: 38 runner, CLI, tuner and dataset tests passed. All configured pre-commit hooks passed, including pinned Ruff 0.15.13.
• Latest complete-stack suite: 443 passed, 11 skipped. The stack covers binary/regression round trips, alignment and corruption checks, fit counts, separate extra-refit timings, and rejection of multiclass, multilabel and link-prediction task types.
• Ran three cheap LightGBM configs (1, 2 and 3 estimators) on rel-f1/driver-dnf, seed 7, with all-config refits. Verified 566 validation rows, 702 test rows and one additional nonwinning refit.
• Verified those three artifacts and 20 synthetic binary/regression artifacts with TabArena's actual BaselineResult.from_pickle / ConfigResult reader at revision de1be372095e23e72e2f42992eceea036b97dea6, without conversion. Recomputed metric errors and checked config metadata and simulation inputs.
• Added workflows/verify_tabarena_predictions.py to repeat the reader check in an environment with TabArena installed.
• No successful cloud upload/download round trip is claimed.