[udaykusupati]

Two data races in the parallel passes, the intermittent python_tracking_correspondences CI failure, and what preserve_topology does and does not keep.

Commits

1. Engine 3D: a vertex's tagged edges are counted from its own fan. The link condition asks the order of the link vertices of a collapsing edge; with feature edges tracked, the count read the fans of their neighbours, two steps from the edge and outside the collapse's lock. ThreadSanitizer flagged it on double_sphere at 8 threads. On macOS it threw "Tuple::eid ... empty tet vector": 6 runs in 25 before the fix, 0 in 25 after. It also accounts for about 1% of tetwild runs on the open bunny. Outputs are byte-identical to main at 1 thread. Introduced with the feature-edge tags (Preserve input feature edges and points #1049).

2. simwild 2D: the last split's vertex is per thread. A write-write race in the parallel split pass (ThreadSanitizer, 6 of 6 dense runs). Its readers are serial; the output is byte-identical at 1 thread.

3. tracking: a boundary vertex collapsed into one inside keeps the outline. SurfaceMap::flatten split an edge that was already there, and counted the edit as not_a_disk. That is python_tracking_correspondences' intermittent CI failure: about 1 in 160 Linux runs; on the mac-mini, 9 hits in 1400 tetwild runs, every one this collapse. Regression test: a 4-vertex history, with a 4-face control.

4. tetwild: preserve_topology reports whether the surface handed to the insertion crosses itself. The surface simplification before insertion can fold the surface through itself, and the insertion then changes the topology. The hidden test tetwild-surface-history-octocat failed about 1 run in 3–5 for this reason. This is now warned about and reported, not prevented: crossing pairs, degenerate triangles, and the Euler characteristics right after insertion. The octocat test checks by phase. The check costs 3–32 ms. The specs state the three phases.

5. simwild 2D: under preserve_topology, the remeshing simplifies after the insertion, as 3D does. Before, 2D did not simplify at all under preserve_topology. Topology is identical on every 2D case and across eps_simplify; the simwild suite is 49/49. One case gains vertices, through a separate, older weakness: stuck-refinement splits a triangle at its energy floor, and it does the same on main at 8 threads.

6. simwild 2D: the simplification relabels the envelope it hands on. Envelope IO: components write and read what they hold as tags #1156 labels the envelope simwild hands to the next component, and the 3D simplification refreshes those labels when it rebuilds the envelope. The 2D twin did not, so under envelope_reference "simplified" the envelope came out unlabelled. Measured: a 64-edge straight interface simplified to 23 edges with no labels; now 23 edges with 23 label rows.

Tests

ctest 234/234 on the tree merged with main, the new map test, the 4-vertex history through Python (0 warnings; main gives not_a_disk 1), and octocat 3/3 (macOS arm64, Release). ThreadSanitizer: 2D runs of triwild, simwild 2D and the 2D offset show no other race in the collapse path, with a 3D positive control.

🤖 Generated with Claude Code