Writing to an open Dataset corrupts the Dataset's memory

What happened?

When holding a reference to an open, lazily-loaded Dataset ds, overwriting the file that backs ds causes ds to reference invalid memory.

What's weird is that this happens when using "coords", and doesn't happen without them.

I understand this access pattern is a bit weird, but I'm raising this bug in case this isn't an expected behaviour.

What did you expect to happen?

I expected ds to continue existing in memory, with its original values. Failing that, I expected a warning or error, either when the file backing ds was overwritten or, more realistically, when I went to access ds after it had been overwritten and was no longer referencing valid memory.

Minimal Complete Verifiable Example

import xarray as xr

# Test without coordinates (works as expected)

print("Testing WITHOUT coordinates:")
xr.Dataset({"x": [1, 2], "y": [3, 4]}).to_netcdf("test1.nc")
lazy1 = xr.open_dataset("test1.nc")
# overwrite the file
xr.Dataset({"z": [5, 6]}).to_netcdf("test1.nc")

# lazy1 still holds valid data
print(f"Expected y: [3, 4], Actual: {lazy1['y'].values}")

# Test with coordinates (invalid memory access)
print("\nTesting WITH coordinates:")

xr.Dataset({"x": ("dim", [1, 2]), "y": ("dim", [3, 4])}, coords={"dim": [0, 1]}).to_netcdf("test2.nc")
lazy2 = xr.open_dataset("test2.nc")
# overwrite the file
xr.Dataset({"z": ("dim", [5, 6])}, coords={"dim": [0, 1]}).to_netcdf("test2.nc")

# Accesses invalid memory
print(f"Expected y: [3, 4], Actual: {lazy2['y'].values}")

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.
• Recent environment — the issue occurs with the latest version of xarray and its dependencies.

Relevant log output

Anything else we need to know?

No response

Environment

Start by running the reported minimal example with xarray.open_dataset and the two coordinate cases. Trace the dataset-opening and coordinate-handling paths to identify where the lazy reference becomes invalid after the backing file is overwritten. Done means the original dataset remains valid or access produces a defined warning or error.