[petrasovaa]

Context

test_v_class_where_no_matches fails intermittently on Windows CI but passes on Linux. It is not a flaky test: it was asserting on the contents of uninitialized memory.

When the where clause selects no records, nrec is 0, so v.class hands an empty array to the classification library, which reads data[0] and data[count - 1]:

Invalid read of size 8
   at AS_class_interval (class.c:66)
 Address 0xdc86008 is 8 bytes before a block of size 1 alloc'd

The printed breaks were whatever the heap held. Linux zeroes fresh heap pages, so they came out as 0.0 and the test passed; Windows reuses dirty memory, so sometimes not.

Change

v.class keeps its behavior (warning, zeroed breaks, exit 0) but now produces it deliberately: with no records it skips the classification and stats calls, leaving the zero-initialized breaks and frequencies in place. AS_class_apply_algorithm also rejects an empty array, removing the same latent bug in d.vect.thematic.

Commit message

An empty WHERE result left an empty data array, which the classification
algorithms then indexed at data[0] and data[count - 1]. The values printed
as class breaks were whatever the heap happened to hold, which is zeroed
on Linux but not on Windows, making the test flaky there.

Warn the user and keep the zero-initialized breaks when nothing is selected, and
reject an empty array in AS_class_apply_algorithm.

[wenzeslaus]

Change to make valid but "zero" output on empty selection makes sense for the tool.

I have some doubts about using warning for no selection, but @petrasovaa says it is appropriate given the context (unlike e.g. for v.db.select).

I assume more work in the library would be needed to do something else than fatal at this point.