test: `EncounterRecencyTransformer` coverage gap — timezone/overflow/ndarray-fallback branches untested (82% coverage, worst in risk tier)

[shivamlalakiya]

Summary

philanthropy/preprocessing/_encounter_recency.py is at 82% branch
coverage, the lowest of any file in the risk tier make riskcov enforces at
a 93% floor — this file alone is dragging that subtree's average down. Unlike
the other coverage-gap issues in this repo, which name one or two specific
guards, this one is a cluster spanning both fit and transform on
EncounterRecencyTransformer, plus its private helpers. Filing it as one
issue rather than several because the branches share a theme (timezone- and
NaN-handling edge cases) and a contributor working through them will end up
reading the whole fit/transform pair anyway.

Untested regions, grouped by what triggers them (confirm current line numbers
with the coverage command below — they will drift as other work touches this
file):

1. _parse_dates's tz-localise fallback. parsed.dt.tz_convert(self.timezone)
   is wrapped in try/except Exception, falling back to
   parsed.dt.tz_localize(self.timezone). No test parses a tz-naive series
   through a timezone-configured transformer, so the fallback never runs.
2. _compute_recency_features's tz-mismatch branches. The three-way branch
   comparing dates.dt.tz against ref.tzinfo (tz-aware dates vs tz-naive
   reference, and the reverse) — only the "both match" case is exercised today.
3. fit's reference-date inference paths. Both UserWarning branches —
   "no parseable dates found in training data" and "X is not a DataFrame" —
   defaulting reference_date_ to today.
4. transform's ndarray-fallback and missing-date-column paths. The
   branch that rebuilds a DataFrame from feature_names_in_ when X isn't
   already one, the "cannot resolve column names" branch returning an
   all-NaN array, and the per-column "date column not found in X" warning
   inside the for col in cols loop.
5. _days_since_day_resolution, the overflow-safe fallback for date spans
   beyond ~292 years — entirely unreached, since no test constructs dates that
   overflow int64 nanosecond resolution.

Where

File	Locate with
philanthropy/preprocessing/_encounter_recency.py	grep -n "def _parse_dates|def _compute_recency_features|def _days_since_day_resolution|def fit|def transform" philanthropy/preprocessing/_encounter_recency.py

What to change

No source change — this is test-only. Start by running coverage to see the
current exact Missing line/branch list, since the numbering above is
described by behaviour, not frozen lines:

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

Then, working through the five groups above:

1. Fit/transform with a tz-naive date series and a timezone param set, to
   force the tz_localize fallback in _parse_dates.
2. Construct one fixture with tz-aware dates and a tz-naive
   reference_date, and one with the reverse, to hit both branches in
   _compute_recency_features.
3. Fit with a DataFrame containing only unparseable dates in every date
   column (to hit "no parseable dates"), and fit with a bare ndarray (to
   hit "X is not a DataFrame") — both should warn and default to today.
4. Transform with a bare ndarray after fitting on a DataFrame (exercises the
   feature_names_in_ rebuild path), transform without ever having resolved
   column names (the all-NaN fallback), and transform a DataFrame missing
   one of the configured date_col columns (the per-column warning).
5. Construct two pd.Timestamps roughly 300+ years apart and call
   _days_since_day_resolution directly (it's a @staticmethod, callable
   without fitting anything) to confirm it returns a finite day count instead
   of raising OverflowError/OutOfBoundsTimedelta.

Tests to add or extend

• File: tests/test_encounter_timezone.py (for the tz-specific branches in
  groups 1–2) or tests/test_preprocessing.py (for groups 3–5) — check
  tests/test_encounter_timezone.py's existing scope first
  (grep -n "class \|def test_" tests/test_encounter_timezone.py) before
  deciding which file a given test belongs in.
• Add tests named to match the existing TestEncounterRecencyTransformer*
  conventions in whichever file you extend.

Done when

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

philanthropy/preprocessing/_encounter_recency.py reads at or above the 93%
risk-tier floor.

make riskcov