[FIX] Pre-compute schema before Ray Dataset transform to prevent empty schema error

[K2alyan]

Problem

forecast() with X_df on a Ray distributed Dataset crashed with:

triad.collections.schema.SchemaError: Schema can't be empty

The crash happened in _distributed_forecast when X_df is not None:

df = fa.transform(df, format_df, schema="*,_in_sample:bool")
X_df = fa.transform(
    X_df,
    format_X_df,
    schema=fa.get_schema(df),           # ← Bug: empty schema on post-transform Ray Dataset
    params={..., "df_cols": fa.get_column_names(df)},
)

fa.get_schema() called on a Ray Dataset after fa.transform can return an empty schema in some Fugue/Ray version combinations, because Ray Datasets do not always propagate Arrow schema metadata through Fugue's transform pipeline the same way Spark/Dask do. Passing an empty schema to the next fa.transform call triggers SchemaError: Schema can't be empty.

Fix

Pre-compute the schema and column list before transforming df (when the original Ray Dataset has a reliable schema), then append _in_sample:bool explicitly. This is the same .copy() + .append() pattern already used in _get_schema() in this file.

_df_schema = fa.get_schema(df).copy()
_df_schema.append("_in_sample:bool")
_df_cols = fa.get_column_names(df) + ["_in_sample"]
df = fa.transform(df, format_df, schema="*,_in_sample:bool")
X_df = fa.transform(
    X_df,
    format_X_df,
    schema=_df_schema,
    params={"target_col": target_col, "df_cols": _df_cols},
)

Behavior is identical for Spark and Dask (where fa.get_schema works post-transform). For Ray it removes the crash.

Test plan

• Removes @pytest.mark.xfail from test_forecast_x_dataframe in nixtla_tests/nixtla_client/test_ray.py
• Run pytest nixtla_tests/nixtla_client/test_ray.py -m ray_run — all tests including the previously-failing test_forecast_x_dataframe should pass