[denelson1-dot]

Fixes #521. Standalone on main (no other PR stacked).

I checked every site #521 lists against the port at c9c5979. The issue was written from the fork. At c9c5979 none of these sites reads uninitialised memory:

• SmVector3d, SmPoint3d and SmPoint2d default to SM_UNDEF_DOUBLE (SmVector3d.h:56-58).
• SmExtent1d defaults to an empty interval.
• AngleBetween writes 0 before it can fail.

The port matches upstream at every site but one: SmVertex::AreEdgesTangent / IsTopologicalVertex. This PR fixes that site and records the others.

Fixed: AreEdgesTangent / IsTopologicalVertex after a failed GetEndTangent (WR-256)

Upstream. The ignored failure leaves the caller's tangent as it was. That is the SmVector3d default, or in IsTopologicalVertex the line vector IsLine wrote. AngleBetween sees a zero-length vector and reports 0, so both functions answer TRUE. UB-106 had called that tangent uninitialised; a correction line now points it to WR-256. Only the curveless edge is really UB.

What changes:

• Port. topology::vertex returned InvalidInput here. It now reads Vec3::UNDEF, as upstream does.
• Shim. The topology_vertex shim guard (ok_tangent, ok_topological) also refused these inputs, and called GetEndTangent to decide. It now refuses only the real crash: a null curve whose tangent upstream would read.

Reachability. No input the topology.vertex harness can build makes GetEndTangent fail, so no existing case changes value:

• The harness builds only line and B-spline edges.
• SmBSplineCurve::Evaluate clamps the parameter.
• N_BasisFindSpan skips end spans shorter than 1e-12, so sm_BasisDerivs never sees the tiny knot difference it rejects.
• A NaN parameter gives NaN, not an error.

An SmCompositeCurve edge always gets there: that class keeps SmCurve::Evaluate, which returns SM_ERR. The new unit test failed_end_tangent_reads_the_default_vector builds that case.

New fixtures. fixtures/topology.vertex/end_tangents.json adds 70 compared cases around this path:

• zero end tangents (a B-spline collapsed to a point), which do reach the AngleBetween failure
• 1e-30 end spans with matching parameters, which exercise the end-span skip

To be clear: these cases agree with or without the port change. They pin the neighbouring behaviour; the unit test is what covers the changed lines.

Recorded, no code change (plan/findings.md, "#521 audit")

• SmLoop::GetUVCornerBoundaries: "keeps the previous edgeuse's curve" doesn't hold at c9c5979. GetOrCreateUVTrimCurve sets its out-pointer before it builds (SmEdgeuse.cpp:4275), so a failed build leaves NULL and the drop-point branch runs. Stale UV ends after a failed EvaluatePoint / DropPoint are ported as they are.
• SmEdge::ReverseOrientation: the default interval after a failed ReverseParameterization is already ported and documented.
• SmLoop::ContainsUVPoint: the four ignored evaluations read UNDEF points or the previous step's derivatives. The port returns their error (its documented deviation, reached only by a failing surface evaluation, a NaN parameter or a degenerate span). The doc wording there said "uninitialised"; this PR corrects it.
• Edge-use loop-gap / dog-leg checks: GetStartPointUVLoopGap is faithful. GetEndPointUVLoopGap and HasDogLegNearPole are unreachable and not ported.
• sm_CalcDistAndDev: faithful, with one open difference noted for a follow-up. The port propagates an evaluate_normal error that upstream ignores. That matters only for a zero-radius sphere or a SrfInVolume whose evaluation fails.
• SmFaceProps::SetProps_Stage3: no unset reads. A failed Stage 3 leaves the stage at 3a, which the port copies.

The fork's status propagation isn't ported. As with #465 and #468, the port keeps c9c5979's results because difftest compares against them.

Preserved invariants

No tolerance loosened, no case dropped, generators unchanged. The guard change only lets through inputs the C++ handles without UB. topology.vertex stays green.

Difftest

PASS topology.vertex cases=5578 passed=5578 failed=0 deferred=0 max_rel_err=0e0 with_values=5576 no_values=2 vacuous=0
SENSITIVE topology.vertex perturbed=5578 detected=5578 unperturbed=0 deferred=0
PASS topology.vertex cases=7562 passed=7562 failed=0 deferred=0 max_rel_err=0e0 with_values=7560 no_values=2 vacuous=0   (--seeds 2000)
FUZZ-PASS topology.vertex cases=2000 rejected=244 malformed=391 ran=1365 differs=0 crashes=0 panics=0 timeouts=0
--green: 260 unit(s), 206 passed, 54 partial, 0 failed

Verified on head fb0aec152 (x86-64): cargo fmt --all --check (+ rustfmt on the touched FFI unit file), clippy -D warnings (plain and --features difftest/cpp), cargo test -p solidcore (2080 passed), the difftest runner tests, the wasm32 check, harness/plan/units.py check, and a clean git merge-tree against main. --green --jobs 4 above.

Since then: merged main (no conflicts locally; #703 took WR-255, so this PR's entry is now WR-256, renumbered in its own commit). Re-checked on f3df18881: fmt, units.py check, clippy, the vertex unit tests and topology.vertex 5578/5578.

🤖 Generated with Claude Code

[fkirill]

Update after #704 landed: #700 is being narrowed to the remaining topology/healing failure paths, while retaining #704’s tangent behavior, tests and fixture coverage.

Correction to my earlier comment: #700’s 20 constant-segment fixtures exercise failed angle calculations / zero tangents. They do not establish failed GetEndTangent evaluation in the existing harness and do not contradict the reachability analysis here. The standalone regression’s failed tangent evaluations use an empty B-spline in Rust. The proposed status-propagation policy in #700 was broader than this PR’s parity contract; that vertex change will be removed rather than overwrite the merged behavior.

The remaining follow-up covers failed edge reversal, UV corner/loop-gap/deviation evaluations and healing classification/retry state. Its updated native comparisons and explicit known records will be published after revalidation.