Collect the OCCT kernel issues n3xd works around in ISSUES.md

2026-10-14 16:16:25 +02:00
parent 9923577764
commit 5a1b0fdeb1
1 changed files with 58 additions and 0 deletions
@@ -0,0 +1,58 @@
+# OCCT issues
+
+Kernel behaviour n3xd works around. Each entry says how to reproduce it, what
+the app does about it and which OCCT version last showed it. Recheck them on
+every kernel bump. Missing bindings are not listed here; they live in
+`NOTEPAD.md`.
+
+## `BRepFilletAPI_MakeFillet` segfaults on a fillet as wide as its wall
+
+- **Symptom:** SIGSEGV inside the fillet build. It is a native fault, so no
+  exception reaches Python and the whole process dies.
+- **Repro:** a 23-face part. It has a 110 × 60 × 20 mm base, a 2 mm diagonal
+  slab and 2 mm side walls, and already carries one r = 2 fillet. A second
+  r = 2 fillet on any edge bounding a 2 mm side wall crashes, for example the
+  wall's outer diagonal edge. r = 1 on the same edges builds. Production
+  project `d0260694`, 2026-10-14.
+- **Workaround:** every fillet and chamfer builds in a forked child
+  (`app/backend/n3xd/cad/cad_pool.py`), where the fault becomes a "try a
+  smaller radius" error. An earlier gate built parts under 200 faces in the
+  server process, on the assumption that 8.0.1's `ChFi3d_Builder::StartSol()`
+  hardening had removed the fault class. It hadn't.
+- **Last seen:** OCCT 8.0.1 (`n3xd-ocp` 8.0.1.2).
+
+## Fillets at an exact tangent radius fail or come out invalid
+
+- **Symptom:** at a radius that makes the blend exactly tangent to a
+  neighbouring face, `BRepFilletAPI_MakeFillet` either reports
+  `IsDone() == false` or reports success and returns a solid that
+  `BRepCheck_Analyzer` rejects.
+- **Workaround:** retry with an imperceptibly smaller radius
+  (`app/backend/n3xd/cad/operations/fillet.py`).
+- **Last seen:** OCCT 8.0.1.
+
+## Tangential union along a closed curve returns an empty shape
+
+- **Symptom:** fusing two solids that touch along a full circle (a revolved
+  ring exactly filling a matching hole) gives `IsDone()` and an empty compound,
+  which `BRepCheck_Analyzer` calls valid. Exact, fuzzy, `BOPAlgo_GlueShift`,
+  OBB, `CellsBuilder` and `MakerVolume` all give an empty or invalid result.
+- **Workaround:** `boolean_combine` rejects the empty result, so the feature
+  reports `boolean_failed`. Moving the tool by 1e-4 mm before the fuse gives
+  the right union in both operand orders, but silently moving user geometry
+  was rejected as a default. The app-side opt-in idea is in
+  `app/NOTEPAD.md`.
+- **Test:** `app/backend/tests/test_boolean.py::test_tangential_fuse_is_rejected_not_silently_emptied`
+- **Last seen:** OCCT 8.0.1 (rechecked 2026-08-11).
+
+## `BRepGProp` mass properties drift on spline-bounded shapes
+
+- **Symptom:** with default settings, `SurfaceProperties_s` and
+  `VolumeProperties_s` misreport spline-bounded faces and solids, while the
+  geometry itself is exact. A closed Catmull-Rom region in `spline.n3xd` reads
+  110.64 against a true 110.01 (+0.57 %). Its extrusion reads about 1109
+  against 1100.1.
+- **Partial workaround:** the eps overload (`SurfaceProperties_s(shape, props,
+  1e-6)`) fixes the area (110.01) but makes the volume worse (about 1166), so
+  it is not a drop-in. The app-side follow-up is in `app/NOTEPAD.md`.
+- **Last seen:** OCCT 7.9 / 8.0 (not rechecked since).