WIP: add preprocessing presets kwarg to TabPFNTSPipeline

[LeoGrin]

Summary

Expose tabpfn's inference-time X/y preprocessing pipeline via a simple string enum on TabPFNTSPipeline. Today changing it requires threading inference_config through tabpfn_model_config with PreprocessorConfig objects imported from a nested tabpfn module — not discoverable.

New kwarg: preprocessing: Literal["default", "none", "squashing_scaler"] = "default".

• "default" keeps current behaviour (tabpfn library defaults).
• "none" disables X/y preprocessing entirely (PREPROCESS_TRANSFORMS=[PreprocessorConfig("none")], REGRESSION_Y_PREPROCESS_TRANSFORMS=[None]).
• "squashing_scaler" uses squashing_scaler_max10 + svd_quarter_components followed by a numeric "none" config.

An explicit inference_config in tabpfn_model_config still wins over the preset; a warning is emitted if both are supplied.

Motivation

Empirical sweep on fev-bench (3 checkpoints × 2 splits × 3 preprocs, 846 per-task results): on the non-lite/small-datasets split, "none" and "squashing_scaler" both give a +0.05 SQL skill boost over defaults for both the library default checkpoint and OOD-finetuned variants. On the lite split the library defaults usually win (small regression from removing preprocessing). Worth exposing as a user-facing knob.

Changes

• tabpfn_time_series/preprocessing_presets.py (new) — PreprocessingPreset type + build_preprocessing_inference_config() helper.
• tabpfn_time_series/pipeline.py — add preprocessing kwarg + _apply_preprocessing_preset() merging logic.
• tests/test_preprocessing_presets.py — 6 tests covering the helper and the pipeline integration (both with and without user-supplied inference_config).

Test plan

• New unit tests pass (pytest tests/test_preprocessing_presets.py).
• Existing test_preprocessing.py, test_predictor.py still pass.
• test_pipeline.py::test_predict_client_mode requires TabPFN-cloud credentials (prompts for login interactively), unrelated to this change.
• User-facing example in README / quickstart notebook — left for follow-up.

🤖 Generated with Claude Code

[CLAassistant]

All committers have signed the CLA.

[gemini-code-assist]

Code Review

This pull request introduces a preprocessing parameter to the TabPFNTSPipeline constructor, allowing users to select from predefined inference-time preprocessing presets. These presets are defined in a new preprocessing_presets module. A review comment pointed out that the _apply_preprocessing_preset method should consistently return a copy of the configuration dictionary to prevent accidental mutation of the global default settings and to match the method's docstring.

[gemini-code-assist]

The docstring for _apply_preprocessing_preset states that it injects the configuration into a copy of the input, but the implementation returns the original tabpfn_model_config object in several branches (lines 293 and 303). Since the default value for tabpfn_model_config in the constructor is a shared global dictionary (TABPFN_DEFAULT_CONFIG), returning it directly can lead to accidental mutation of the default configuration if the predictor or other components modify it in-place.

Always returning a copy ensures the original configuration remains immutable and consistent with the docstring.

        config = tabpfn_model_config.copy()
        preset_cfg = build_preprocessing_inference_config(preprocessing)
        if preset_cfg is None:
            return config
        if "inference_config" in config:
            # User-supplied inference_config takes precedence; warn so the
            # mismatch between kwargs is discoverable.
            warnings.warn(
                "Both `preprocessing` and `tabpfn_model_config['inference_config']` "
                "were provided. Using the explicit `inference_config` from "
                "`tabpfn_model_config` and ignoring the preset.",
                stacklevel=3,
            )
            return config
        config["inference_config"] = preset_cfg
        return config