test: `preprocessing._transformers` — CRMCleaner/FiscalYearTransformer defensive branches are untested (87% coverage)

[shivamlalakiya]

Summary

philanthropy/preprocessing/_transformers.py sits at 87% branch coverage —
the lowest of any file in the risk tier that make riskcov enforces at 93%.
The gap is a cluster of defensive branches in CRMCleaner and
FiscalYearTransformer that no test reaches:

1. _get_pandas_output's two fallback paths. The function first tries the
   modern sklearn.utils._set_output._get_output_config path, falling back to
   a manual _sklearn_output_config attribute check if that import or call
   fails, and within that fallback there's a further branch between a plain
   "pandas" string and a {"dense": "pandas"} dict shape. No test forces the
   ImportError/AttributeError path or exercises the dict-shaped config.
2. The except Exception re-raise / retry branches in fit and transform
   on both CRMCleaner and FiscalYearTransformer. Each of these four
   methods wraps validate_data(...) in a try/except Exception, re-raising
   only when the message contains "Complex data not supported" and
   otherwise retrying with X.astype(object). Existing tests only exercise
   the "no exception" and "complex data raises" cases — the retry-succeeds
   path (some other validate_data exception, e.g. a mixed-type object array
   that a plain dtype check trips on but .astype(object) fixes) is never
   taken.
3. The plain ValueError("Complex data not supported") raise inside
   CRMCleaner.fit/transform and FiscalYearTransformer.fit/transform
   when validation succeeds but the resulting array is complex-typed — as
   opposed to the case where validate_data itself raises on complex input.
   Confirm with coverage which of the two complex-data raise sites in each
   method is the untested one before writing the test.

Where

File	Locate with
philanthropy/preprocessing/_transformers.py	grep -n "_get_pandas_output|except Exception|Complex data not supported" philanthropy/preprocessing/_transformers.py

What to change

No source change — this is test-only. Run coverage first to see exactly which
lines are still red before writing tests, since some of the branches above may
shift as other issues in this file land:

python -m pytest tests/test_preprocessing.py --cov=philanthropy.preprocessing._transformers --cov-report=term-missing -q

1. For _get_pandas_output: add a test that monkeypatches
   sklearn.utils._set_output._get_output_config to raise AttributeError
   (or removes it) and sets estimator._sklearn_output_config = {"transform": {"dense": "pandas"}} directly, asserting the dict-shaped fallback still
   returns True.
2. For the retry-succeeds branch: find or construct an input that
   validate_data rejects on its first pass but accepts once cast to
   dtype=object (a CRMCleaner/FiscalYearTransformer sibling test already
   exists for the complex-data case — locate with
   grep -n "Complex data" tests/test_preprocessing.py — use it as the
   pattern for constructing tricky input, but pick a case that recovers
   instead of raising).
3. For the complex-result-after-successful-validation raise: construct X
   that validate_data accepts (e.g. object dtype) but whose array is
   np.iscomplexobj(...) true after validation.

Tests to add or extend

• File: tests/test_preprocessing.py
• Add: test_get_pandas_output_dict_shaped_config_fallback
• Add one retry-succeeds test per class (CRMCleaner, FiscalYearTransformer)
  for fit and/or transform, named to match the existing
  test_crm_cleaner_* / test_fiscal_year_transformer_* conventions in that
  file (locate with grep -n "class Test" tests/test_preprocessing.py).

Done when

python -m pytest tests/test_preprocessing.py --cov=philanthropy.preprocessing._transformers --cov-report=term-missing -q

philanthropy/preprocessing/_transformers.py reads at or above the 93%
risk-tier floor with no branches from the three groups above left in
Missing.

make riskcov