Skip longitude normalisation for projected datasets in harmonise_dims (fixes #1476)

[gaoflow]

Problem

harmonise_dims unconditionally wraps 0-360 longitudes to -180-180 for any dataset whose x-coordinates exceed 180:

if np.any(lons > 180):
    ds[x_dim] = xr.Variable(x_dim, np.where(lons > 180, lons - 360, lons))
    ds = ds.sortby(x_dim)

For a projected dataset whose x is in metres (e.g. SVY21 / EPSG:3414, where every x value is > 180), this subtracts 360 from every coordinate, silently shifting the grid 360 m to the west. As the source CRS is still unset, the displaced grid is then reprojected incorrectly, producing spatially shifted output. No warning is raised. (Reported in #1476.)

Fix

Normalise longitudes only when the dataset has a geographic CRS:

• geographic CRS → wrap to -180-180 (unchanged behaviour);
• projected CRS → leave untouched (x is metres, not longitude);
• no CRS → leave unchanged and emit a warning, rather than assuming geographic — the caller should set crs in the data catalog entry to opt in.

This follows the direction in the issue discussion (thanks @alimeshgi, @michaelohanrahan): an unknown CRS should not be assumed geographic, since that is the exact scenario that corrupts projected data.

Reproduction (from #1476)

import numpy as np, pandas as pd, xarray as xr
from hydromt.data_catalog.drivers.preprocessing import harmonise_dims

x = np.arange(10490, 10490 + 30*30, 30, dtype=float)   # 30 m resolution, all > 180
y = np.arange(39010, 39010 + 30*30, 30, dtype=float)[::-1]
t = pd.date_range("2020-01-01", periods=3, freq="D")
ds = xr.Dataset({"precip": (["time","y","x"], np.random.rand(3, len(y), len(x)))},
                coords={"x": x, "y": y, "time": t})

print(ds.x.values[0])                    # 10490.0
print(harmonise_dims(ds.copy()).x.values[0])
# before fix: 10130.0  (shifted -360)
# after fix : 10490.0  (unchanged, with a warning that no CRS is set)

Verification

• New regression tests in tests/data_catalog/drivers/test_preprocessing.py cover all three cases (geographic normalises, no-CRS left unchanged + warns, projected-CRS left unchanged). The two bug-targeting tests fail on main and pass with this change; the geographic case is unchanged.
• Full tests/data_catalog/drivers/ suite passes (112 passed, 15 skipped).
• ruff check and ruff format --check clean.

[michaelohanrahan]

Thanks for your attention on this issue @gaoflow