fit_post_processing() crashes on a minority class with only 1 member instead of degrading gracefully

[yakew7]

Where: faircode/strategies.py:155 (fit_post_processing's stratify guard).

The gap:

stratify = y_train if len(np.unique(y_train)) > 1 else None

This only guards the fully-degenerate single-class case. sklearn.train_test_split(..., stratify=y) requires every class to have at least 2 members, not just 2+ distinct classes to exist - so a y_train with exactly 2 classes but a minority class of size 1 passes this check and then crashes inside train_test_split with a raw sklearn error.

Repro:

>>> import numpy as np, pandas as pd
>>> from sklearn.ensemble import RandomForestClassifier
>>> from faircode.strategies import fit_post_processing
>>> X_train = pd.DataFrame({'a': np.random.default_rng(0).random(20), 'b': np.random.default_rng(1).random(20)})
>>> y_train = np.zeros(20, dtype=int); y_train[0] = 1   # 2 classes, minority has 1 member
>>> sensitive_train = np.array(['g1','g2'] * 10)
>>> fit_post_processing(RandomForestClassifier(random_state=42, n_estimators=10), X_train, y_train, sensitive_train, random_state=42)
Traceback (most recent call last):
  ...
ValueError: The least populated classes in y have only 1 member, which is too few. The minimum number of groups for any class cannot be less than 2. Classes with too few members are: [1]

Why it matters: S4 (post_processing) is one of the five strategies run for every (audit, model, protected-attribute) cell in the benchmark harness. A small audit, a heavily row-filtered subset, or an intersectional subgroup can plausibly leave a rare positive/negative class with only 1 training example, crashing the whole strategy with an unhandled sklearn internal error - unlike the graceful insufficient_data-style degradation faircode.metrics/faircode.significance already use elsewhere for other degenerate inputs.

Suggested fix: check the minority class count, not just the number of distinct classes - e.g. min(np.bincount(y_train)) >= 2 - or catch the ValueError from train_test_split and fall back to stratify=None, instead of len(np.unique(y_train)) > 1.

[be-student]

I’d like to work on this. I’ll enforce the calibration-split precondition from the observed class counts, preserve normal stratification when every class supports it, and add a regression covering the singleton minority case and state returned to the benchmark caller.

[yakew7]

alright when you're done open a pr and we will review it