Fix fit() crashes in tuning: rare class missing from holdout, Generator random_state

[LeoGrin]

Summary

Two crashes in the tuning_config path of TabPFNClassifier.fit(), both found by auditing main:

1. Temperature calibration / threshold tuning crashes when a rare class is absent from the tuning holdout.
find_optimal_temperature called log_loss(y_true, probas) without labels=. probas always has n_classes columns, but the holdout labels come from only the first tuning_n_folds folds of the StratifiedKFold, so a rare class can be entirely absent and sklearn raises ValueError: y_true and y_prob contain different number of classes. Repro: 25,002 samples with a 2-sample class and tuning_config={'calibrate_temperature': True} (auto-resolves to 1 of 3 folds).

Notably the failure is deterministic, not seed-dependent: StratifiedKFold's per-fold class allocation depends only on the class counts, shuffle/random_state only permute which samples fill each fold's quota. A scheduled retrain can start crashing because two rows of a new rare class were added.

The log_loss entry of METRIC_NAME_TO_OBJECTIVE (used by threshold tuning on binarized one-vs-rest labels) had the same missing-labels pattern and failed with y_true contains only one label; fixed with labels=[0, 1].

2. fit() crashes when random_state is a np.random.Generator and tuning is enabled.
get_tuning_splits forwarded random_state raw to StratifiedKFold, whose check_random_state rejects np.random.Generator — even though Generator is a documented random_state type and works everywhere else in fit(). Now converted to a static seed via infer_random_state, matching what fit() itself does.

Behavior impact

None where the code previously worked:

• With every class present in the holdout, sklearn infers labels = np.unique(y_true) = np.arange(n_classes) (y is label-encoded before the tuning call), so the explicit labels= is bit-for-bit identical — verified, including the chosen temperature.
• The Generator conversion only touches the isinstance(random_state, np.random.Generator) case, which previously always crashed; int/RandomState/None are passed through unchanged.

Tests

Three regression tests in tests/test_inference_tuning.py, matching the file's unit-test style (no model needed). All three fail on main and pass with this change; the 24 existing tests in the file still pass.

🤖 Generated with Claude Code

[bejaeger]

what if we deal with multiclass? should we take the labels from y_true unless it's a single label?