Numpy datetime and timedelta dtypes have a nan-like value called NaT. The NaT value internally represented by the 64 bit signed integer -9223372036854775808:

np.int64(-9223372036854775808).view('datetime64')
# np.datetime64('NaT')

The zarr v2 spec does not define a JSON-serializable fill value encoding for datetimes or timedeltas. In lieu of a spec, we can go by the implementation of zarr-python 2.x, which used the integer form of the datetime / timedelta for all fill value, including NaT:

# /// script
# requires-python = ">=3.11"
# dependencies = [
#     "zarr==2.18",
#     "numcodecs<0.16"
# ]
# ///


import zarr
import json
z = zarr.create((10,), dtype='>M8[s]', fill_value='NaT')
print(json.loads(z.store['.zarray'])['fill_value'])
# -9223372036854775808

But we have a property test that checks if the numpy NaT value is serialized to metadata as the string "NaT". This goes against the behavior of v2, according to which It should be serialized as an int.