[shchur]

Summary

Shorten the public signatures by moving rarely used options out of them.

• Options moved to **kwargs, validated against per-method TypedDicts with Unpack, so IDEs still autocomplete them. Unknown keys raise a TypeError that lists the supported ones.
  • Training jobs (CP fit / fit_predict*, FM predict / predict_proba): job_name, volume_size, custom_image_uri, timeout
  • Batch transform (CP predict / predict_proba): job_name, instance_count, custom_image_uri
  • Deploy (CP / FM deploy): initial_instance_count, volume_size, custom_image_uri
• Removed instance_count and leaderboard from training methods. Training is always single-instance and the leaderboard is always saved. Passing either logs a warning and the value is ignored. FM predict / predict_proba handle instance_count the same way.
• Moved image_column / test_data_image_column to MultiModalCloudPredictor. On tabular, fit / predict / predict_proba raise an error pointing to MultiModalCloudPredictor.
• Keyword-only after the data arguments: CP predict / predict_proba, all of CP / FM deploy, TimeSeriesCloudPredictor.predict, TimeSeriesFoundationModel.predict, TimeSeriesEndpoint.predict.
• framework_version and backend_overrides stay named everywhere. backend_overrides is now also named on the FM methods, where it used to go through **backend_kwargs.
• timeout now also works for time series training.
• The fit_predict plumbing (test_data, extra_ag_args) moved from fit(**kwargs) into a private _fit().

Breaking changes

• Positional calls now raise TypeError, e.g. predict(df, "s3://..."), deploy("s3://..."), TimeSeriesFoundationModel.predict(data, "target"), TimeSeriesCloudPredictor.predict(data, static_features), TimeSeriesEndpoint.predict(data, known_covariates). Pass these by keyword.
• Options moved into **kwargs: same names, still keyword arguments, so keyword callers are unaffected. Keys that a method doesn't support now raise TypeError instead of being forwarded unchecked; this affects FM methods, which used to forward their extra keyword arguments to the backend without checking them.
• leaderboard=False on training methods is ignored with a warning; the leaderboard is always saved.
• instance_count on training methods (CP fit / fit_predict*, FM predict / predict_proba) is ignored with a warning. It already had no effect.
• image_column / test_data_image_column are only accepted by MultiModalCloudPredictor.
• FM backend_overrides is now a named parameter; passing it by keyword works as before.

Docs

The hidden options and backend_overrides are documented under a numpydoc Other Parameters section, after Returns. docs/conf.py registers it as a separate napoleon section; by default napoleon merges it into Parameters.

Testing

• New tests/unittests/general/test_backend_kwargs.py: forwarding, unknown-key errors, ignored-kwarg warnings (CP and FM), image-column errors and split, the legacy backend_kwargs={...} migration hint, keyword-only calls.
• fit_predict unit tests now mock _fit instead of fit.

[github-actions]

Job PR-294-7d664fa is done.
Docs are uploaded to https://d12sc05jpx1wj5.cloudfront.net/PR-294/7d664fa/index.html