[akhilles]

This change fixes four tessellation defects that the board corpus finds (ENG-3712). NURBS projection stops after 65536 patches near a collapsed edge, the parser keeps the first items of an over-long bounded list, shapes and colour groups come out in a fixed order while shell faces mesh in parallel, and whole spheres bounded by a vertex loop mesh as rings of latitude. Of 885 corpus STEP models, 877 give the same mesh as master, 7 change only because they now tessellate with no failed faces, and one times out on both.

[akhilles]

Verification (orb, 16 threads, tessellate_step_bytes with parallel, median of 3 in-process runs):

Model	master	this PR
ASP-134486-01.step	aborts after 1 min 56 s: 11.6 GB resident, a 3.4 GB queue allocation fails	757 ms, 5167 faces, 0 failed
36907406S.stp	panics (CapacityError)	16 ms, 315 faces, 0 failed
TCM1-63AX.step	75 ms, exact output differs on every run (6 of 6)	15 ms, identical on every run
901-144-8RFX.step	634 ms	144 ms, same mesh
STM32H563IGK6.step	201 of 209 faces fail	0 failed, 13 ms
MT48LC16M16A2B4-6A_G.step	54 of 62 faces fail	0 failed, 6 ms
TPS22916BYFPR.step	4 of 12 faces fail	0 failed, 2 ms
STM32H747XIH6.step	265 of 273 faces fail	0 failed, 23 ms
125_3203_001_A.stp (not in the issue)	1183 of 26239 faces fail (all vertex-loop spheres)	0 failed, 747 ms

Board corpus: all 893 distinct embedded models from the 619 boards, of which 885 are STEP (8 are WRL). Each model ran in its own process with a 300 s timeout and an 8 GB address-space limit. The canonical mesh hash is the sorted set of triangles with colours.

• 877 models give the same mesh as master. 7 differ: the six from the issue plus 125_3203_001_A.stp. All 7 now tessellate with no failed faces.
• Two full runs of this branch give identical exact (ordered) output for all 884 models that finish. Before the colour-group tie-break, 7 models still changed order between runs.
• RedPitaya_3Dmodel.stp exceeds 300 s on both master and this branch. That is a separate, existing problem.
• Models still partial on both master and this branch, unchanged: ASP-184330-01 (6 faces), Multispectral Thermal Camera PAG7936 (2), BSS138BKT-TP (1).

Tests: cargo test --workspace --all-features passes at every commit. The triangulate tests also pass with rayon removed from the build, and the wasm32 check passes. Each new regression test fails on master: the sphere test with CouldNotLower, the parser test with CapacityError, and the order tests by changing order. The projection test does not finish within 60 s on master; with the cap it takes about 6 s in a debug build.

[cursor]

Not approved: this is a functional tessellation and mesh-output change, not a small fixup, so human review is required. Cursor Security Agent passed with no findings; Cursor Bugbot was not present. Human review is needed and no reviewers were assigned.

  

_{Sent by Cursor Approval Agent: Pull Request Router and Approver}