perlin: preserve input x/y coordinates on the output

[brendancol]

Summary

• perlin() carried the input's attrs (including crs/res) but built the result with dims=agg.dims and no coords. A georeferenced input produced an output that kept its crs attribute while dropping the coordinate values, an internally inconsistent DataArray that breaks downstream coord-based ops (clip, interp, reproject).
• Fix: pass coords=agg.coords when constructing the result. The output has the same shape as the input, so coords align directly.

Backend coverage

Verified on numpy, cupy, dask+numpy, and dask+cupy. The result is wrapped in one shared code path, so the fix applies to all four.

Test plan

• New test_perlin_preserves_coords asserts x/y coords and attrs pass through
• pytest xrspatial/tests/test_perlin.py -- 13 passed (GPU and dask+GPU tests ran live)

Note: GitHub issues are disabled on the source fork, so this PR has no linked issue. Found by the metadata propagation sweep.

[brendancol]

PR Review: perlin: preserve input x/y coordinates on the output

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

• The new test_perlin_preserves_coords only runs on numpy. The GPU and dask tests (test_perlin_gpu, test_perlin_dask_cpu, test_perlin_dask_gpu) build coord-less rasters, so nothing asserts coord pass-through on cupy/dask. The fix is in one shared wrap path so all backends behave the same, and it was verified manually, but a coord-bearing assertion on at least the dask backend would lock it in. Low priority.

Nits (optional improvements)

None.

What looks good

• One-line fix at the source of the bug (xrspatial/perlin.py:354): coords were dropped while attrs were kept, so the output claimed a crs with no coordinates. coords=agg.coords matches how attrs are already handled.
• Output shape always equals input shape (perlin writes into a same-shape buffer or rebuilds it), so coords align directly with no off-by-one or half-pixel risk.
• The numpy in-place write touches only .data, not .coords, so reusing agg.coords is safe.
• Regression test uses general_output_checks plus explicit coord-value and attrs assertions.

Checklist

• Algorithm unchanged; only result wrapping touched
• Four backends share the wrap path; coords verified on numpy/cupy/dask/dask+cupy
• NaN handling unchanged
• Coord pass-through covered on numpy; GPU/dask coverage relies on the shared path
• No premature materialization introduced
• No benchmark needed (no perf change)
• No README change needed (no API change)
• No new public API; docstring unchanged

[brendancol]

Follow-up review

The one suggestion from the prior pass is addressed: test_perlin_preserves_coords_dask now asserts coord pass-through on the dask backend with ragged chunks, alongside the numpy test. Full perlin suite passes (14 tests, GPU and dask+GPU ran live).

No remaining blockers, suggestions, or nits.