Propagate friction geo-attrs through least_cost_corridor (#3446)

[brendancol]

Closes #3446

least_cost_corridor builds the corridor as cd_a + cd_b, where each cost-distance surface carries the attrs of its source raster. Source rasters are usually plain marker masks with no geo-attrs, so xarray's binary sum dropped res/crs/transform/nodatavals from the output even when the friction surface that defines the grid had them.

Changes

• Add _apply_geo_metadata to re-emit the friction surface's attrs and name on the corridor output.
• Wire it into the non-precomputed path (single, threshold, unreachable, and pairwise outputs). The precomputed path keeps its source-derived behavior, since there is no friction surface to draw from.
• Only .attrs and .name are set; data, coords, dims, and dtype are untouched, and dask stays lazy.

Backends

Verified live on numpy, cupy, dask+numpy, and dask+cupy (CUDA host).

Test plan

• test_corridor_inherits_friction_geo_attrs across all four backends
• test_corridor_threshold_keeps_geo_attrs across all four backends
• test_corridor_unreachable_keeps_geo_attrs
• test_pairwise_inherits_friction_geo_attrs
• test_precomputed_keeps_source_attrs_not_friction
• Full corridor suite: 43 passed

Found by the metadata propagation sweep (/sweep-metadata).

[brendancol]

PR Review: Propagate friction geo-attrs through least_cost_corridor (#3446)

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None.

Nits (optional improvements)

• In the pairwise branch, _apply_geo_metadata calls .rename(geo_source.name) on each corridor, but the DataArray then goes into xr.Dataset(corridors) keyed by corridor_i_j, and the dict key wins as the variable name. So the rename is a no-op for the Dataset variables; only the attrs propagation does anything there. Harmless, just slightly redundant, not worth a change on its own.

What looks good

• The fix goes at the root cause: the corridor is cd_a + cd_b, each surface carries its source raster's attrs, and the binary sum drops anything that differs between the two operands. Re-emitting friction's attrs and name is the right call, since friction is the raster that defines the grid.
• The precomputed path is left alone, which is correct. There is no friction surface to draw from there, and test_precomputed_keeps_source_attrs_not_friction pins that behavior.
• _apply_geo_metadata calls result.copy() before touching attrs, so it does not mutate the input and keeps dask arrays lazy (checked: no compute triggered). .rename(None) yields a None name when friction is unnamed.
• Test coverage is thorough: all four backends for the main and threshold paths, plus the unreachable all-NaN and pairwise paths. Full corridor suite passes (43 tests).

Checklist

• Algorithm matches reference/paper (N/A; metadata-only change, values untouched)
• All implemented backends produce consistent results (numpy/cupy/dask+numpy/dask+cupy verified live)
• NaN handling is correct (unreachable all-NaN path keeps attrs)
• Edge cases are covered by tests (unreachable, pairwise, precomputed)
• Dask chunk boundaries handled correctly (no neighborhood op; copy stays lazy)
• No premature materialization or unnecessary copies (.copy() is a shallow metadata copy)
• Benchmark exists or is not needed (not needed; no perf-sensitive change)
• README feature matrix updated (N/A; no new function or backend change)
• Docstrings present and accurate (helper docstring explains the root cause)