…nfig

PriorLabs/TabPFN#890 (which adds get_inference_config) is on TabPFN main
but not yet released — every released version (≤ v7.1.1) lacks the method.
For those, fall back to the historical hardcoded MAX_NUMBER_OF_CLASSES=10
when the base estimator's class lives in a tabpfn-prefixed module. Other
estimators (sklearn etc.) still get the explicit-alphabet ValueError.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…282)

* Auto-infer alphabet_size in ManyClassClassifier from the base estimator's checkpoint

The previous fallback (`estimator.max_num_classes_`) has been dead since the
initial commit — TabPFN core has never set that attribute on any version
(v2 / v2.5 / v2.6 / v3), so users always had to pass `alphabet_size` explicitly
or hit `ValueError`. Now resolution cascades:

  1. Explicit `alphabet_size=...` (unchanged)
  2. `estimator.inference_config_.MAX_NUMBER_OF_CLASSES` (post-fit)
  3. Probe: fit a clone on 4×2 synthetic rows to populate `inference_config_`,
     then read `MAX_NUMBER_OF_CLASSES`

The probe is cheap because TabPFN's `_load_checkpoint_cached` (`@lru_cache`)
makes the subsequent codebook fits reuse the already-loaded checkpoint — net
I/O is a single ckpt read. Verified with the v2.5, v2.6, and v3 default
checkpoints (alphabet auto-resolves to 10, 10, 160 respectively).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* Log resolved alphabet_size and base-fit count in ManyClassClassifier.fit

Surfaces "Base estimator supports up to N classes; data has M — …" at
verbose=1, in both the no-mapping and codebook branches, so users can see
how the alphabet was resolved and how many base fits will follow.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* Strip categorical_features_indices on the probe clone

The probe fits on a synthetic 2-feature matrix; any user-provided
categorical indices ≥ 2 (e.g. categorical_features_indices=[3]) on the
base estimator would trip TabPFN's index-bounds validation before
inference_config_ gets populated. Reset them on the clone.

Surfaced by codex review on PR #282.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* Restore clean ValueError when probe rejects the synthetic input

Pre-PR, a non-TabPFN base estimator with no alphabet_size always raised
the documented "alphabet_size must be specified ..." ValueError. After the
probe was added, that path could surface an arbitrary error from inside
the user's estimator instead.

Catch ValueError/TypeError around probe.fit (the typical sklearn
input-validation errors) and fall through to None so the documented
ValueError still fires. Heavier exceptions (RuntimeError, OSError, etc.)
still propagate so genuine bugs aren't masked.

Surfaced by cursor-bot review on PR #282.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* Fix ruff D209 on _probe_alphabet_size docstring

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* Use TabPFN's get_inference_config() instead of a probe fit

TabPFN exposes a public get_inference_config() method that loads the
checkpoint without fit data and returns the active InferenceConfig
(honoring any constructor override). Drop the probe fit, the synthetic
X_probe / y_probe construction, the categorical_features_indices reset,
and the (ValueError, TypeError) catch — all of it collapses into one
method call.

Suggested by adrian-prior on PR #282.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* Default alphabet_size to 10 for older TabPFN without get_inference_config

PriorLabs/TabPFN#890 (which adds get_inference_config) is on TabPFN main
but not yet released — every released version (≤ v7.1.1) lacks the method.
For those, fall back to the historical hardcoded MAX_NUMBER_OF_CLASSES=10
when the base estimator's class lives in a tabpfn-prefixed module. Other
estimators (sklearn etc.) still get the explicit-alphabet ValueError.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>