chore: finish the CI cleanup round

.github/workflows/perf-regression.yml

@@ -54,7 +54,7 @@ jobs:
       - name: Install
         run: |
           python -m pip install -U pip
-          pip install -e ".[bench]"
+          pip install -c constraints/ci.txt -e ".[bench]"
       - name: Resolve base commit
         id: base
         env:

docs/community/contributor-roadmap.md

@@ -35,11 +35,9 @@ review does not depend on understanding the whole cleaning pipeline.
   [Great Expectations recipe (#8)](https://github.com/FreshCode-Org/freshdata/issues/8),
   [ydata-profiling comparison (#7)](https://github.com/FreshCode-Org/freshdata/issues/7).
   Files live in [`examples/`](https://github.com/FreshCode-Org/freshdata/tree/main/examples).
-- **Small isolated cleanups** —
-  [de-duplicate `_has_outliers` (#33)](https://github.com/FreshCode-Org/freshdata/issues/33)
-  and
-  [fix boolean coercion in `_coerce_series` (#31)](https://github.com/FreshCode-Org/freshdata/issues/31)
-  are both labeled `good first issue` and name the exact function and file.
+- **Small isolated cleanups** — look for open issues labeled
+  [`good first issue`](https://github.com/FreshCode-Org/freshdata/labels/good%20first%20issue);
+  each names the exact function and file.
 
 **Skills:** Python, pandas basics. **You'll touch:** `examples/`, `docs/`, or one
 named module + its test.

src/freshdata/imputation/missforest.py

@@ -285,7 +285,10 @@ def _convergence_delta(
             if plan.model_type == "regressor":
                 prev_num = pd.to_numeric(prev, errors="coerce")
                 cur_num = pd.to_numeric(cur, errors="coerce")
-                denom = float(np.nanstd(cur_num.to_numpy(dtype="float64"))) or 1.0
+                # na_value: nullable (masked) columns with missing cells refuse a
+                # plain float64 conversion on pandas < 2.
+                values = cur_num.to_numpy(dtype="float64", na_value=np.nan)
+                denom = float(np.nanstd(values)) or 1.0
                 deltas.append(float(np.nanmean(np.abs(cur_num - prev_num))) / denom)
             else:
                 deltas.append(float((cur.astype(object) != prev.astype(object)).mean()))

tests/test_missforest_imputation.py

@@ -227,3 +227,16 @@ def test_missforest_auto_indicator_uses_prefill_informative_missingness():
 
     assert "age_was_missing" in out.columns
     assert int(out["age_was_missing"].sum()) == 40
+
+
+def test_missforest_handles_nullable_integer_column_with_missing_values():
+    # Regression: pandas < 2 refuses a plain float64 conversion of a masked
+    # (nullable) integer column that still holds missing values.
+    df = _mixed_frame()
+    df["age"] = pd.array(df["age"].round().astype(int), dtype="Int16")
+    df.loc[5:14, "age"] = pd.NA
+
+    out, report = fd.clean(df, impute="missforest", return_report=True, **ISOLATE)
+
+    assert out["age"].isna().sum() == 0
+    assert _missforest_actions(report, "age")