fix: DCEL corruption crashes in trim/split and shape registration

[1egoman]

Summary

This branch fixes a family of DCEL (Doubly-Connected Edge List) corruption bugs discovered by the fuzzer. The crashes all surface the same way — linkNext: half-edge "he_..." does not exist, splitEdge/mergeEdges "edge not found", or TypeError — during undo/redo of trim / split operations on overlapping shapes.

The root cause is a single theme: re-registering polygon boundaries into the DCEL can mint degenerate edges (zero-length self-loops, duplicate half-edge pairs under one edge key), which orphans half-edge references and corrupts the ref-counted edge cache. Later undo/redo replay then trips over the dangling entries.

A fuzzer (added in the last commit) replays randomized sequences and checks invariants after each op. End-to-end result across the 253-artifact corpus:

State	Artifacts passing
Before this branch	0
After first 3 fixes	127 / 253
After all 5 fixes	182 / 253

Full Jest suite: 1052 passing tests across 37 suites (10 skipped).

---

Fix 1 — 3be97aa — Address DCELShapeIndex duplicate edge causing thrown error

Problem. When a new polygon edge passes within floating-point noise of an existing vertex, addVertex(split.point) minted a near-duplicate vertex, and two adjacent existing edges were both "split" at that phantom point — producing two half-edge pairs for the same edge key. The edge cache can only hold one pair per key, so the second overwrite orphaned the first while shapes still referenced it. On undo/redo, releaseEdge/mergeEdges culled a live pair → linkNext: half-edge "he_..." does not exist.

Fix. In _registerShape Phase 2, snap each computed intersection point onto the closest of the four relevant vertex positions (existing origin/dest, candidate origin/dest) when within SPLIT_ENDPOINT_EPSILON (1e-9). addVertex then returns the genuinely-shared vertex id, so the existing endpoint-touch guard fires and no phantom vertex / duplicate edge key is created.

Before

[undo-redo-symmetry] invariant 'undo-redo-symmetry' threw:
  linkNext: next half-edge "he_121" does not exist.

After — the minimized repro (ellipse + open polygon + trim-split) reports:

NOT REPRODUCED - the minimized sequence no longer triggers the bug.

---

Fix 2 — dce010a — Avoid empty undo transactions for no-op selection actions

Problem. ToggleLinkDimensionsAction and OpenClosePolygonAction iterate the selection and silently skip geometries outside their target type. With a selection containing only such shapes, applyTransaction captured zero entries but still pushed an empty undo frame — so undo() became a silent no-op, tripping undo() did not change the document at all.

Fix. Pass { omitIfEmpty: true } to both transactions (a pre-existing HistoryManager option already used by the selection inspector).

Before

[undo-redo-symmetry] undo() did not change the document at all (no-op undo on non-empty stack)

After — all 35 no-op-undo artifacts report NOT REPRODUCED.

---

Fix 3 — d11d8c7 — Guard main-polygon dedup against vertex count mismatch in trim

Problem. In TrimSplitTool.processCurrentTrim, the "does the main polygon already exist?" dedup compared points.every((p, i) => equals(p, mainPoints[i])). When an existing polygon had more vertices than the reconstructed boundary, mainPoints[i] was undefined and PolygonSegment.equals crashed reading .type.

Before

TypeError: Cannot read properties of undefined (reading 'type')
    at Object.equals (polygon.ts:563)
    at TrimSplitTool.processCurrentTrim (TrimSplitTool.tsx:430)

After — vertex counts are compared first; mismatched polygons are skipped. All 6 TypeError artifacts now report NOT REPRODUCED.

---

Fix 4 — 4185f6a — Skip trims whose boundary retraces an untouched shape

Problem. walkCombinedBoundary is a heuristic (see #63): it can produce a boundary that hugs an overlapping, untouched shape instead of the union of the shapes being trimmed. Re-registering that boundary as a new polygon duplicated the untouched shape's edges → shared-edge ref-count corruption → undo/redo crash.

Example (rectangle (5.497,25.771)-(7.409,36.34) + closed polygon [(3.126,54.906),(5.093,40.035),(6.522,9.195),(13.767,10.635)], trim at (8.44, 24.126)):

Before — the trim produced a corrupt merged region whose shape contained interleaved zero-length self-loop edges:

pairs: v_2->v_3, v_3->v_3, v_3->v_4, v_4->v_4, v_4->v_23, v_23->v_23, ...

and crashed on undo with linkNext: next half-edge "he_97" does not exist.

After — the trim is detected as ambiguous (none of the boundary vertices belong to an affected shape) and skipped with an explanatory console.warn; geometry is left untouched:

[trim-split] skipped: combined boundary ( 6 edges) does not touch any of the affected shapes ( ply_fuzz2 ).
             Likely trimming an overlapping shape by mistake — leaving geometry unchanged.
NOT REPRODUCED - the minimized sequence no longer triggers the bug.

---

Fix 5 — 377235f — Skip degenerate self-loop edges when registering shapes

Problem. In _registerShape Phase 5, when an intersection split point resolves to the candidate edge's own destination, the split loop already adds the real prevVId → destId edge, but the trailing final segment closed into a zero-length self-loop edge (destId → destId). Self-loop pairs orphan half-edges from the edge cache and corrupt ref counts → linkNext crashes on undo/redo.

Fix. When prevVId === candidate.destId and the last split was a colinear-duplicate contact (uOnExisting < 0.5), skip the redundant final segment. Shared-edge endpoint touches (u ≈ 1 — used by colinear-merges in the shared-edge removal tests) are preserved.

After — the guard plus the trim-skip above leaves 182 / 253 fuzzer artifacts passing and the shared-edge tests still green.

---

Last commit — 800c0d4 — Add the fuzzer harness, adapted to the current tool registry

The regression tests live in fuzz/__tests__/fuzz.test.ts. They need the harness (fuzz/app.ts, fuzz/ops.ts, fuzz/repro.ts, fuzz/invariants.ts, ...), which is added here. The harness is adapted to the current architecture where trim-split lives inside the edit multi-tool (instead of a flat top-level tool): setActiveTool('edit') + changeToolSubTool('edit', 'trim-split'). Also adds the tsx devDependency used to run fuzz/cli.ts.

Testing

• npm test — 37 suites, 1052 passing, 10 skipped.
• Fuzzer corpus rerun after all fixes via fuzz/cli.ts repro <artifact> → NOT REPRODUCED for the 182 fixed cases.
• Regression tests added for each fix in fuzz/__tests__/fuzz.test.ts and existing unit suites.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
cad2d	Ready	Preview	Sep 21, 2026 1:12am UTC