_apply_vertical_shift can hang on non-finite coords; vertical CRS path is untested

[brendancol]

Two related problems in xrspatial/reproject around the vertical-datum-shift code path.

Hang on non-finite coords

_apply_vertical_shift in xrspatial/reproject/__init__.py (around lines 702-791) hands output-grid coordinates to _interp_geoid_2d directly. When the target CRS is projected, output coords are first inverse-projected to lon/lat at lines 773-776. Some projections (polar stereographic near the pole, the antimeridian on certain CRSes) emit inf or NaN from transformer.transform. Those values flow into _interp_geoid_point in xrspatial/reproject/_vertical.py (lines 127-130):

while lon_w < -180.0:
    lon_w += 360.0
while lon_w >= 180.0:
    lon_w -= 360.0

On inf, neither loop terminates and the worker hangs.

The public geoid_height API guards against non-finite input at _vertical.py:205-216, but _apply_vertical_shift calls the JIT batch directly and skips that guard.

Fix

After the optional inverse-projection step, replace any non-finite values in xx_strip and yy_strip with NaN before calling _interp_geoid_2d. The interpolator already returns math.nan for out-of-range input, so NaN-in NaN-out is fine. Update the per-strip is_valid mask to drop pixels with non-finite coordinates so NaN from the geoid does not contaminate the output.

The change is under ten lines, scoped to _apply_vertical_shift. _vertical.py is left alone.

No coverage for the vertical CRS path

The src_vertical_crs and tgt_vertical_crs parameters of reproject() (defined at __init__.py:507-516) have no end-to-end tests in xrspatial/tests/test_reproject.py. _apply_vertical_shift is roughly 90 lines and is currently exercised only through unit tests of geoid_height.

Tests to add

A TestVerticalShift class covering:

• EGM96 to ellipsoidal on a small lon/lat raster, verifying the shift matches the known undulation at a sample location.
• The reverse direction, confirming opposite sign.
• EGM96 to EGM2008, skipped if the EGM2008 grid is not available.
• Same-vertical-CRS no-op.
• Single-side vertical CRS no-op.
• A projected source/target (UTM 33N) to exercise the inverse-projection branch.
• A polar-stereographic case (EPSG:3413) on a 89-90 N raster to regression-test the hang. Assertion is that the call returns and produces finite output where the source is finite.