Fix the red trunk: inverted inside-test in the fault-network line clip, and a gmsh finalize cascade

[lmoresi]

Fix the four CI-only test_0851 failures: one inverted sign, one missing finally

All four CI failures (Linux, conda PETSc 3.25.3) trace to a single source
bug plus one missing cleanup — not to geometry constants in the tests.

What was environment-dependent

_line_clip_interval (fault_network_3d.py) had an inverted inside test in
its parallel-edge branch: for an interior point, num = nrm.(A - p0) is
NEGATIVE, but the code returned "outside" on num < 0. The branch only
fires when an edge is EXACTLY parallel to the clipped line (den == 0.0,
guard 1e-30). CI's OpenBLAS returns exact zeros for the axis-aligned
patch's SVD plane basis, so den == 0.0 and the inverted branch made
crossing_segment report "no crossing" for genuinely crossing patches.
macOS Accelerate leaves ~1e-17 jitter in den, dodging the branch into
the sign-correct division path — the dev boxes passed by luck. Verified
by probe: with the module's own code, an exactly axis-aligned crossing
pair returns None pre-fix on macOS too.

Per-test diagnosis

1. test_preparer_trims_junior_and_records_segment and
2. test_network_3d_end_to_end: crossing_segment(Main, Cross) returned
   None on CI, so the preparer fell through to the near-miss ligament
   refusal (dmin 0.02233 is the true rim-rim distance of the test pair —
   correct data, wrong branch). Fixed by the sign fix.
3. test_embed_refuses_crossing_patches: BoxInternalPatch's crossing
   pre-check uses the same crossing_segment; the miss let the geometry
   reach gmsh/tetgen, whose raw PLC error escaped. Fixed by the sign fix
   (the designed ValueError now fires on every platform).
4. test_mesh_simple_array_view_std (PETSc error 79): a CASCADE from 2.
   The escaped PLC exception aborted BoxInternalPatch after
   gmsh.initialize() and before gmsh.finalize(); the poisoned session
   made the next StructuredQuadBox (first test of the next file in the
   same test_08* pytest batch) write an invalid .msh, which PETSc
   rejects: error 79, "expecting $Nodes, not $". Reproduced exactly on
   macOS by blinding the crossing check and replaying the sequence.
   Fixed by try/finally gmsh.finalize() around generate/write.
   The std test itself is sound (no solver is involved; the error was at
   mesh file read) and is unchanged.

How the tests now pin the property

New test_crossing_segment_exact_parallel_edges: axis-aligned patches
make den exactly 0.0 on EVERY platform — the CI condition becomes
deterministic everywhere. Asserts the crossing segment analytically
(x=0.42, y=0.5, z in [0.32, 0.68]).

Negative controls

• Pre-fix, the new regression test's crossing call returns None on macOS
  (probe-verified) — the test genuinely detects the bug on all platforms.
• Same test includes a non-crossing exact-parallel pair (plane x=0.82
  outside the square) that must STAY None — guards against a blanket
  sign inversion.
• test_preparer_refusals (near-miss NotImplementedError, non-convex
  refusal) still passes — the refusal arms still fire.
• Cascade probe: with the crossing check monkeypatched blind, the PLC
  error still escapes but gmsh.isInitialized() == 0 afterwards and the
  follow-on StructuredQuadBox succeeds (pre-fix: exact CI error 79).

Verification (sequential, worktree amr-dev)

• test_0851_fault_network_3d.py test_0851_std_reduction_method.py test_0851_surface_influence_edge.py: 17 passed.
• test_0845_fault_split.py test_0846_fault_contact.py test_0847_fault_api.py test_0848_fault_split_3d.py: 29 passed
  (unchanged set, stays green).

Remaining for the fault session

Patch sets the crossing pre-check cannot classify — coplanar overlaps,
patches touching along an edge/point — can still reach tetgen and fail
with a raw PLC error instead of the embed's own message. TODO(BUG) is
placed at the generate(3) call in BoxInternalPatch. The finally
means such a failure can no longer poison later meshes, but the refusal
message should become the preparer's own.

Underworld development team with AI support from Claude Code

[Copilot]

Pull request overview

Fixes CI-only failures in the 3-D fault network meshing/tests by correcting an inverted inside-test in the convex line-clipping routine (triggered when an edge is exactly parallel to the plane–plane intersection line), and by ensuring gmsh is finalized even when meshing/write fails to prevent cross-test session poisoning.

Changes:

• Fix _line_clip_interval() parallel-edge inside/outside classification in fault_network_3d.py.
• Add a deterministic regression test covering the exact-parallel-edge case.
• Wrap gmsh mesh generation / write in try/finally to guarantee gmsh.finalize() on failure.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
tests/test_0851_fault_network_3d.py	Adds a regression test that deterministically exercises the exact-parallel-edge branch and asserts the analytic crossing segment.
src/underworld3/meshing/fault_network_3d.py	Fixes the parallel-edge “inside” test in the line clip interval computation to avoid missing genuine crossings.
src/underworld3/meshing/cartesian.py	Adds try/finally around gmsh generate/write so mesher failures don’t leave gmsh initialized and poison later meshes.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.