classify: return all-NaN instead of crashing on all-non-finite input (#3510)

[brendancol]

Closes #3510

What

head_tail_breaks, percentiles, and box_plot raised an opaque reduction error when every pixel was non-finite (all NaN, or all +/-inf, which is mapped to NaN before binning). Each one took a max or percentile over the finite subset of the data, and on this input that subset is empty.

• head_tail_breaks -> ValueError: zero-size array to reduction operation fmax
• percentiles -> IndexError: index -1 is out of bounds for axis 0 with size 0
• box_plot -> ValueError: zero-size array to reduction operation fmax

They now guard the empty-finite case and return an all-NaN array, matching equal_interval, std_mean, and maximum_breaks.

Backends

numpy, cupy, dask+numpy, dask+cupy. The guard sits in the shared bin/percentile helpers and the eager box_plot branch, so it covers all four.

Test plan

• Flipped the xfail regression tests (added by the test-coverage sweep) to plain assertions
• Added all-inf coverage for the three functions
• Full test_classify.py passes (117)

[brendancol]

PR Review: classify all-non-finite handling (#3510)

Read the full classify.py context, not just the diff. The fix is correct and the backend coverage holds up.

Blockers

None.

Suggestions

• percentiles now reaches np.nanmax(clean) in the wrapper on all-NaN input (the crash that used to preempt it is gone), so it emits a benign RuntimeWarning: All-NaN slice encountered. std_mean has the same behaviour and its test suppresses it with warnings.catch_warnings(). For consistency, test_percentiles_all_nan / test_percentiles_all_inf could do the same (xrspatial/tests/test_classify.py). It does not fail CI -- RuntimeWarning is not promoted to error in setup.cfg -- so this is housekeeping, not a correctness issue. head_tail_breaks and box_plot short-circuit before any nan-reduction and stay warning-free.

Nits

None.

What looks good

• The guard sits in the shared helpers (_compute_head_tail_bins, _run_percentiles, _run_dask_percentiles, _box_plot_bins_from_sample) plus the eager _run_box_plot branch, so a single guard per code path covers all four backends without duplicating logic. Verified the box_plot mapper routes numpy/cupy through _run_box_plot and both dask paths through _box_plot_bins_from_sample.
• Returning a [nan] bin (or _bin(agg, [nan], [0])) is the same degradation equal_interval/std_mean/maximum_breaks already use, so behaviour is uniform across classifiers.
• all-inf coverage uses both +inf and -inf, exercising the isinf -> nan mapping.
• Full test_classify.py passes (117), including cupy and dask+cupy on a CUDA box.

Checklist

• Degradation matches sibling classifiers
• All four backends produce all-NaN consistently
• NaN/inf handling correct (isinf catches both signs)
• Edge case (all-non-finite) covered by tests
• Dask sample paths guarded
• No premature materialization introduced
• No new public API (no README/docstring change needed)

[brendancol]

Follow-up review (#3510)

Re-reviewed after the follow-up commit and the clean merge of origin/main.

Disposition of prior findings

• Suggestion (percentiles all-NaN RuntimeWarning): fixed in c94d36e. test_percentiles_all_nan and test_percentiles_all_inf now wrap the call in warnings.catch_warnings() the same way test_std_mean_all_nan does. Confirmed by running the all-NaN/all-inf subset with -W error::RuntimeWarning: 16 passed, no unsuppressed warnings.

New state

• Merged origin/main (geotiff symbology sidecar commit, unrelated to classify). No conflicts. Full test_classify.py still passes (117).

No remaining findings.