Downsampling with resample generates longer time array than expected

MCVE Code Sample

import xarray as xr
import pandas as pd
from functools import reduce

# Create time dimension array for all climatological winters
indexes = [pd.DatetimeIndex(f'{year}-12-01', f'{year+1}-03-01', freq='12H') 
                 for year in range(1980, 2020)]
index_union = reduce(pd.Index.union, indexes)

# Create DataArray
ds = xr.Dataset({'var': ('time', np.arange(len(ix_union))), 'time': index_union})

Problem Description

From here we can check ds time dimension months:

>>> pd.DatetimeIndex(ds.time.values).month.unique()
 Int64Index([12, 1, 2, 3], dtype='int64')

Now, if we downsample our Dataset to a one week period:

>>> pd.DatetimeIndex(ds.resample(time='1W').mean().time.values).month.unique()
Int64Index([12, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11], dtype='int64')

Instead of getting the weeks within the months in the original Dataset, we obtain additional months with missing values in the var array. One way of solving this issue is to use the dropna() method, but is a slow approach.

Expected Output

We would expect something like this:

>>> pd.DatetimeIndex(ds.resample(time='1W').mean().time.values).month.unique()
 Int64Index([12, 1, 2, 3], dtype='int64')

Output of xr.show_versions()

Start by reproducing the MCVE around Dataset.resample(time='1W').mean(), correcting the apparent np and ix_union typos if needed. Trace the resampling entry point and compare the generated time bins with the non-missing var values. Done means resampling the seasonal dataset does not create empty bins for unrelated months, without requiring dropna().