Test that attention implementations handle output_attentions correctly.

The AutoGluon path uses validate_inputs=False and must not emit the deprecation warning.

Lock down the legacy output contract: structure, item ordering, value preservation, and prediction timestamps (multi-target, with past-only + know

Test that eager and SDPA implementations produce identical outputs on full pipeline.

Series with fewer than 3 points are no longer rejected (freq inference happens elsewhere).

Bool covariates must be target-encoded as categories, not passed through as raw 0/1 floats.

Target encoding with single target should produce finite, per-item-aware encodings.

End-to-end: unseen future category should encode to item mean, not category 0's encoding.

NaN is a category with its own mean: distinct from "x", from the item mean, and matching itself across past and future.

context[-n_future_covariates:] must contain known-future rows (Chronos2Dataset invariant).

With multi-target, falls back to ordinal encoding (not target encoding).

{'a': None} in future_covariates should behave like known_covariates_names=['a'].

An unseen future string encodes to the item mean, not the NaN category.

Test that fine-tuning works with preprocessed inputs from a HuggingFace Dataset.

Test that the pipeline works with different attention implementations.

Series with fewer than 3 points are allowed as long as the frequency can be determined.

predict_df must produce the same output as manually preprocessing the df with from_data_frame + make_future_df and assembling the result (multi-ta

Unseen categories at inference must map to per-item mean, not category 0's encoding.