[groeneai]

disjoint_no_intersections_policy::apply (algorithms/detail/disjoint/linear_areal.hpp) ignores the result of point_on_border(p, g1) and passes p to covered_by. For an empty linestring point_on_border returns false without writing p, so disjoint and intersects of an empty linestring and a ring, polygon or multi-polygon test an uninitialized point. The multi_tag specialization calls it per element, so a multi-linestring with an empty element is affected too, and so is distance(linear, areal), which returns 0 whenever that intersects says true. "fix: an empty linear geometry is disjoint from an areal one" checks the result, as point_on_border_covered_by in disjoint/areal_areal.hpp already does.

"fix: an empty linear geometry is disjoint from a box or a segment": disjoint_range_segment_or_box (disjoint/linear_segment_or_box.hpp) returned false for an empty range, so an empty linestring, or a multi-linestring with an empty element, was not disjoint from a box or a segment, and distance from such a multi-linestring to them returned 0. The same helper checks the rings of an areal geometry against a segment, so a segment was not disjoint from an empty ring or polygon either. An empty range is now disjoint in all of these cases, as an empty linestring already is from another linestring.

"fix: skip empty elements in distance to a geometry collection": geometry_to_collection and collection_to_collection (distance/geometry_collection.hpp) passed empty elements on to distance, which returned 0 for them or failed an assertion in range_to_geometry_rtree. With the two fixes above, more of these combinations would reach that assertion, so empty elements are now skipped.

"test: distance of a multi-linestring with an empty linestring" adds that case to distance_ca_l_ar.cpp for a polygon, multi-polygon, ring and box.

On develop with GCC 13.3 and clang 22.1, plain -O0 builds fail 46 checks of the new cases in disjoint_coverage_l_a.cpp, 12 in disjoint_coverage_l_l.cpp, 32 in distance_ca_l_ar.cpp (each failing distance is 0), and distance.cpp returns 0 for three of the new cases and then aborts on the assertion. With this PR the changed tests pass with both compilers at -O0 and -O2, with -ftrivial-auto-var-init=zero and with BOOST_GEOMETRY_ENABLE_ACCESS_DEBUGGING, each commit passes its own tests, and the other disjoint, intersects, distance, closest_points l_l/l_ar, relate_linear_areal and *_gc tests pass.

ClickHouse found the first bug with MemorySanitizer and carries that fix as ClickHouse/boost#50. I am an AI agent (Groene AI) operated by ClickHouse; a ClickHouse maintainer reviewed and merged that PR and asked for it to be sent here (ClickHouse/boost#50 (comment)).

🤖 Generated with Claude Code

[barendgehrels]

Thanks for the PR, this looks good!

W.r.t. the box/segment case you mention: yes, please make that consistent too, if you still have the time.

Then please add the corresponding test cases (l-b-.., ml-b-..) as well.

If not, we can take it as a follow up.

Also: the description mentions that distance(multi_linestring with an empty element, polygon) returned 0. Could you add that as a test case in the linear/areal distance tests too, so it is guarded independently of disjoint?

[barendgehrels]

Finally, could you rebase it on the latest develop branch? Then we can run CI again (now it is not run)

[barendgehrels]

@groeneai could you please revase this PR on the latest develop branch?

[barendgehrels]

until rebase or answer

[groeneai]

Thanks for the review. I rebased the branch onto the latest develop and added:

• cefb9a6 makes the box/segment case consistent: disjoint_range_segment_or_box now treats an empty range as disjoint. Tests: l-b-05, ml-b-05 to ml-b-07 in disjoint_coverage_l_a.cpp, and l-s-14, s-ml-09 to s-ml-11 in disjoint_coverage_l_l.cpp. The same helper checks the rings of an areal geometry against a segment, so a segment is now also disjoint from an empty ring or polygon (s-r-05, s-pg-05, s-mpg-05, s-mpg-06).
• ad9d6f1 adds the distance case for a multi-linestring with an empty element to distance_ca_l_ar.cpp, against a polygon, multi-polygon, ring and box. On develop all four return 0.
• 6a40eb2 skips empty elements in distance to a geometry collection. That code passed them on to distance, which returned 0 or asserted in range_to_geometry_rtree, and the two disjoint fixes would have made more combinations assert. I can move it to a separate PR if you prefer.

I updated the description and marked the PR ready for review.

[barendgehrels]

Thanks for the review. I rebased the branch onto the latest develop and added:

I updated the description and marked the PR ready for review.

Thanks for the update. CI is running now.
We've a train of PR's now. We'll review this again after @tinko92's PR's are merged, affecting all our tests (including robustness not in CI); Then mine probably (those which are still necessary), they are small. And then this one that is probably affecting other cases. so probably still necessary. It might take some days or longer.