MicroDataFrame.cov() and .corr() ignore weights #327

MicroDataFrame.cov() and MicroDataFrame.corr() return plain pandas results. The overrides at microdf/microdataframe.py:78 and :84 call pd.DataFrame(self, copy=False).cov(*args, **kwargs) / .corr(...) and use __finalize__ only to keep weights off the column-by-column result. MicroSeries.cov and MicroSeries.corr are frequency-weighted, so the same pair of columns gives a different answer depending on the entry point.

import numpy as np, pandas as pd, microdf as mdf

df = pd.DataFrame({"x": [1.0, 2.0, 3.0, 4.0], "y": [1.0, 4.0, 2.0, 8.0]})
w = np.array([1, 1, 1, 5])
m = mdf.MicroDataFrame(df, weights=w)
rep = df.loc[df.index.repeat(w)].reset_index(drop=True)  # frequency-replicated sample

m.cov().loc["x", "y"]     # 3.1667  == df.cov()  (unweighted)
m.x.cov(m.y)              # 3.1786  == rep.cov() (weighted)
m.corr().loc["x", "y"]    # 0.792   == df.corr() (unweighted)
rep.corr().loc["x", "y"]  # 0.896

Observed on 1.5.2 (main at de08443).

Expected. MicroDataFrame.cov() and .corr() return the pairwise frequency-weighted matrix, consistent with MicroSeries.cov / .corr and with the package convention that weights are frequency counts (sum(w) - ddof in the denominator), so integer weights agree with pandas on the replicated sample.

Context. Surfaced while reviewing #324, where the API reference described these two methods as frequency-weighted. The docs are being corrected there; this issue is the code side. Whichever way it goes, the reference rows for these two methods should state which behaviour they have.