Encoding error when saving netcdf

What happened?

When I select a single point (or regional subset) from a netcdf file and save as a new netcdf file, the newly saved file has an encoding issue that causes the data to be incorrect.

How the data should look:

How the newly saved file looks when opened:

Additional context copied from original discussion:
I am trying to save a regional subset of a netcdf file as netcdf file. I am first opening some data with dimensions of time, latitude, and longitude and then slicing that data by latitude and longitude to produce a smaller subset of the data. I save that smaller subset with the to_netcdf command. But when I go to open the new netcdf, the timeseries definitely wrong (see figures). The figure named 'correct' is what the temperature timeseries looks like when plotting directly from the original dataset. The figure named 'wrong' is what the temperature timeseries looks like when plotting from the newly saved netcdf (hopefully both figures attached properly). This happens when I select just a single point and save the data as a netcdf and it also happens when I save as a zarr file. However, when I load a single netcdf with open_dataset (instead of open_mfdataset) and save it as a new netcdf, everything is correct. So the issue seems to be coming from open_mfdataset. I've also noticed that not all grid points are incorrect, only some grid points have this issue. This doesn't happen when I convert to a series then save as a CSV, just happens when saving as a netcdf or zarr.
Link to the original discussion: https://github.com/pydata/xarray/discussions/7025#discussion-4385791

What did you expect to happen?

The data should have looked exactly the same.

Minimal Complete Verifiable Example

`import numpy as np
import xarray as xr
import pandas as pd
import matplotlib.pyplot as plt

# I took the encoding data from the original data and applied it to a dummy dataset to reproduce the issue
original_encoding = {
'original_shape': (744, 109, 245),
'missing_value': -32767,
'_FillValue': -32767,
'scale_factor': 0.0011997040993123216,
'add_offset': 269.40377331689564}


# create dummy dataframe
times = pd.date_range(start='2000-01-01',freq='1H',periods=8760)

# create dataset
ds = xr.Dataset({
    't2m': xr.DataArray(
                data   = np.random.random(8760),   
                dims   = ['time'],
                coords = {'time': times},

                ),
    'tmax': xr.DataArray(
                data   = np.random.random(8760),   
                dims   = ['time'],
                coords = {'time': times},
                )
            },
    )

# apply original encoding
ds.t2m.encoding = original_encoding

# save dataset as netcdf
ds.to_netcdf(r"...\test_ds2.nc")

# load saved dataset
ds_test = xr.open_dataset(r'...\test_ds2.nc')

# Plot the difference between the two variables
plt.plot(ds.t2m - ds_test.t2m)`

MVCE confirmation

• Minimal example — the example is as focused as reasonably possible to demonstrate the underlying issue in xarray.
• Complete example — the example is self-contained, including all data and the text of any traceback.
• Verifiable example — the example copy & pastes into an IPython prompt or Binder notebook, returning the result.
• New issue — a search of GitHub Issues suggests this is not a duplicate.

Relevant log output

No response

Anything else we need to know?

No response

Environment

Start with the minimal example using xarray's Dataset.to_netcdf and compare it with the open_mfdataset workflow described in the report; the report also identifies zarr output as affected. Reproduce the incorrect values and trace how the original encoding is applied during serialization, then verify that saved and reopened subsets preserve the original data.