fix(waterdata): Force float64 dtype on physical-measurement columns

``pd.to_numeric`` infers ``int64`` when every value in a column is
integer-shaped — common in the USGS API response for small streams
whose daily discharge happens to be whole-CFS for the queried period.
Callers expect ``value``/``altitude``/``drainage_area``/depths to be
float because they're continuous quantities; integer dtype causes
surprise downstream (integer-division arithmetic, dtype upcast when
concatenated with a sibling query whose values are fractional).
Append ``.astype("float64")`` after the numeric coercion.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Author: thodson-usgs

dataretrieval/waterdata/utils.py

@@ -1169,7 +1169,15 @@ def _type_cols(df: pd.DataFrame) -> pd.DataFrame:
         df[col] = pd.to_datetime(df[col], errors="coerce")
 
     for col in cols.intersection(numerical_cols):
-        df[col] = pd.to_numeric(df[col], errors="coerce")
+        # ``pd.to_numeric`` infers ``int64`` when every value happens to
+        # be integer-shaped (e.g. all-whole-CFS discharge values at a
+        # small stream). These are physical measurements — discharge,
+        # altitude, drainage area — and are continuous quantities by
+        # nature, so force ``float64`` to avoid surprise arithmetic
+        # surfaces (e.g. integer division on a chained subsequent
+        # query) and to keep dtype stable across sites with mixed
+        # whole vs fractional observations.
+        df[col] = pd.to_numeric(df[col], errors="coerce").astype("float64")
 
     return df
 

tests/waterdata_utils_test.py

@@ -16,6 +16,7 @@
     _handle_stats_nesting,
     _parse_retry_after,
     _raise_for_non_200,
+    _type_cols,
     _walk_pages,
 )
 
@@ -490,6 +491,27 @@ def test_get_resp_data_always_materializes_id_column():
 # --- _arrange_cols ----------------------------------------------------------
 
 
+def test_type_cols_value_is_always_float_even_for_whole_numbers():
+    """Regression: ``pd.to_numeric`` infers ``int64`` when every value
+    is integer-shaped. The USGS API returns whole-CFS discharge as
+    string-typed integers at small streams (e.g. ``"12"`` for 12 cfs),
+    so ``_type_cols`` would coerce ``value`` to ``int64`` and a
+    downstream chained query mixing this site with another that has
+    fractional readings would surface a confusing dtype upcast. Force
+    ``float64`` for every physical-measurement column."""
+    df = pd.DataFrame(
+        {
+            "value": ["12", "15", "18", "14", "11"],
+            "altitude": ["100", "200"] + [None] * 3,
+            "drainage_area": ["5", "10", "15", "20", "25"],
+        }
+    )
+    out = _type_cols(df)
+    assert out["value"].dtype == "float64"
+    assert out["altitude"].dtype == "float64"
+    assert out["drainage_area"].dtype == "float64"
+
+
 def test_arrange_cols_does_not_mutate_caller_properties():
     """`_arrange_cols` must not mutate the caller's `properties` list.