Add boolean2 and CrossSection regression and corpus tests by zmerlynn · Pull Request #1755 · elalish/manifold

zmerlynn · 2026-06-11T12:47:07Z

Test suite for the boolean2 cutover - #1707 staging item 5, following #1722/#1749/#1751/#1754. Adds 79 tests:

test/boolean2_test.cpp (new): 37 white-box tests of the engine in suite Boolean2 - graph order, vertex merge, edge-vertex lists, intersection insertion, the retained-graph validator, and winding filtering. The validator's geometric predicates are deliberately independent of the engine's (using IntersectSegments to check its own output would be circular); a comment in the file records how their epsilon band differs from CCW's.
test/cross_section_test.cpp: extended from its existing 15 tests with 41 new public-API tests - Boolean/BatchBoolean invariants (commutativity, distributivity, area conservation), Offset join types and edge cases, Decompose containment, Simplify, and degenerate/non-finite inputs.
test/cross_section_offset_corpus_test.cpp (new): offsets the polygon triangulation corpus across every join type and gates on clean triangulation; runs on WASM via the preloaded corpus.

Most regression cases are seeds from a fuzz campaign against the backend, kept as literal geometry with a short note on what each shape triggers; the fuzz harness itself follows in a separate PR. Two of these - tiny-feature point contacts near a host vertex - are marked DISABLED_; they expose an arrangement/winding-robustness issue tracked separately. White-box ../src includes follow the existing pattern (boolean_test, manifold_test, and three others). Suite runs green on Linux/Mac/Windows/WASM and under ASan+UBSan.

codecov · 2026-06-11T13:19:45Z

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 93.83%. Comparing base (0a68eae) to head (a2ee932).

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1755      +/-   ##
==========================================
+ Coverage   92.74%   93.83%   +1.08%     
==========================================
  Files          44       44              
  Lines        9990     9985       -5     
==========================================
+ Hits         9265     9369     +104     
+ Misses        725      616     -109

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

zmerlynn · 2026-06-11T18:51:41Z

Cherry-picked the fix from #1757 here so these regression tests pass on the branch. Merge #1757 first; rebasing afterward drops the cherry-pick (identical to base) and leaves just the tests.

zmerlynn · 2026-06-11T18:53:18Z

Also, sorry for the noise here: fuzz testing was stable for days, then of course the minute I put up the PR, it finds something. (ETA: but what it found was a good research project and I extended #1757 as well after that first comment.)

elalish

Thanks, really helpful to see all this. I've scanned over these pretty quickly, so please highlight any particularly interesting ones that I haven't commented on.

elalish · 2026-06-13T09:43:13Z

+
+// Read the polygon corpus (name, expectedNumTri, epsilon, numPolys, then the
+// polygons), keeping just the name and the polygon set.
+std::vector<std::pair<std::string, Polygons>> ReadCorpus(


We must already have this function and the next somewhere, right? Can we reuse?

Extracted a shared ReadPolygonCorpus into test/polygon_corpus.h; offset_corpus and polygon_test now use it (with a guard so a missing corpus still fails loudly rather than silently registering no tests). polygon_fuzz keeps its own copy (separate fuzz harness, doesn't pull the gtest test infra) - can adopt it in a follow-up.

elalish · 2026-06-13T09:45:27Z

+  for (const auto& entry : corpus) {
+    const CrossSection cs(entry.second);
+    const vec2 size = cs.Bounds().Size();
+    const double diag = std::sqrt(size.x * size.x + size.y * size.y);


la::length(size)?

Done - and while here I did a broader la:: pass across the boolean2 tests and source (length/cross/isfinite/maxelem, vector add/sub); an earlier review pass had missed a bunch of these.

elalish · 2026-06-13T09:52:07Z

+        SCOPED_TRACE(entry.first + " frac=" + std::to_string(frac));
+        const CrossSection off = cs.Offset(frac * diag, jt);
+        const Polygons result = off.ToPolygons();
+        if (!result.empty()) ++nonEmpty;


Could we tighten this further to say an outset has more area than the input and an inset less? In fact, can we also say the join types are always ordered by area? Seems like Miter should be the largest area, followed by Round, then Square and Bevel, though not quite sure what order those are in - can never remember the distinction. Makes me wonder if we should have both in our API...

Tightened - asserted outset >= input >= inset per join type, plus the outset join-type ordering, which the corpus confirms is Miter >= Square >= Round >= Bevel (not what I'd have guessed either). The Square-vs-Bevel API question is fair but out of scope here.

elalish · 2026-06-13T09:57:11Z

+}
+
+double AreaTol(const CrossSection& a, const CrossSection& b) {
+  return 1e-6 * (1.0 + std::fabs(a.Area()) + std::fabs(b.Area()));


If these functions are only called once, let's just inline them.

Right - ExpectBooleanInvariants and ExpectDistributesOverUnion were each single-use; inlined both. AreaTol stays (used across ~10 tests).

elalish · 2026-06-13T10:00:48Z

+    const double scale = 1.0 + std::fabs(a.Area()) + std::fabs(b.Area());
+    EXPECT_NEAR(aUb.Area(), a.Area() + b.Area() - aIb.Area(), 1e-3 * scale)
+        << "inclusion-exclusion violated at offset " << offset;
+    EXPECT_NEAR(soup.Area(), aUb.Area(), 1e-5 * scale)


Where do these various near tolerance values come from? If you're really just testing for the presence/absence of a feature, would a bounding box check be more precise?

The tolerances are area-relative (scale = 1+|a|+|b|): 1e-3 for the inclusion-exclusion identity (generous - it's a sub-eps near-coincidence) and 1e-5 for parity between the binary a + b and building one CrossSection from both polygon sets at once (same arrangement, so tighter). A bbox check wouldn't discriminate here - the tiny polygon sits right at the big one's corner, so dropping it barely moves the union's bbox; the lost area is what catches it.

Revising my earlier reply here - I've made the area check independent of the engine since.

The bbox answer still holds: the small polygon sits at the big one's corner, so dropping it barely moves the union's bbox; the lost area is what catches it. But I was measuring that area with inclusion-exclusion, every term from the same engine - self-referential, so a uniform collapse passes 0 == 0 + 0 - 0 vacuously. I added an independent floor: the union must retain at least the raw shoelace area of the larger input, computed from the input coordinates with no engine call (ExpectUnionRetainsArea). That catches both a dropped polygon and a full collapse.

elalish · 2026-06-13T10:18:03Z

+
+// Host-drop only at large offset (passes at the origin): the StarRing host plus
+// an 8-vertex feature anchored 1e-9 from host[1].
+TEST(CrossSection, DISABLED_TinyFeatureNearCornerHostDropAtOffset1024) {


Ditto here.

Same answers as above. This one's the offset-sweep variant: it passes at the origin and only drops the piece once everything's shifted out to a large coordinate, which is why it sweeps {0, 1024, 4096} - same winding instability, surfaced by absolute-coordinate magnitude rather than the merge details. Same loose bound and explicit intersect-empty EXPECT; stays DISABLED_ while I'm still digging into the winding issue.

elalish · 2026-06-13T10:24:37Z

+  const vec2 d = b - a;
+  const double len2 = dot(d, d);
+  if (len2 == 0.0) return false;
+  return std::fabs(la::cross(d, p - a)) <= eps * std::sqrt(len2);


What about when a and b are less than eps apart and p is distant from them? That's why CCW is built how it is - I believe this function will sometimes say they are not collinear. Unless you restrict the kind of input you give this function?

Good catch - you're right it'd call them non-collinear in that case. It's saved by the input: these oracle predicates only run on retained output, where the validator has already dropped any eps-zero edge, so |b - a| > eps. I added a note making that precondition explicit - it's not meant as a general collinearity test.

elalish · 2026-06-13T10:29:17Z

+  return std::fabs(la::cross(d, p - a)) <= eps * std::sqrt(len2);
+}
+
+::testing::AssertionResult CheckRetainedGraphValidity(


This looks pretty complicated - can we have a description of what it's doing? Also, I subscribe to the philosophy that simple testing is preferable, since otherwise there is uncertainty when a test fails as to whether the bug is in the algorithm or the test itself. However, you've been using these for awhile, so you probably know better than I what this has been useful in catching.

Added a contract. The aim is an independent oracle - predicates not reused from the engine - that's caught missed splits, dropped/stacked verts and sign errors area checks miss. Happy to trim anything murky.

elalish · 2026-06-13T10:29:58Z

+                          /*debug=*/false, WindRule::Add);
+}
+
+void ExpectSameFingerprint(const OverlapResult& a, const OverlapResult& b,


Some description here would be nice too.

Added one - a determinism/idempotence check (arrangements equal up to eps-quantization).

elalish · 2026-06-13T10:31:52Z

+}
+
+// RemoveOverlaps2D produces an edge-balance imbalance for some
+//   vertices on these mixed-scale inputs - 1e-6 alongside 1024.


What is edge balance?

Signed sum of incident edge multiplicities at a vertex (out minus in); a valid arrangement conserves it. Defined in the comment, and fixed the stale wording - the test asserts the balance holds.

zmerlynn · 2026-06-13T15:06:42Z

Thanks for the review. A few tests that didn't get their own comments but are worth a look:

SubtractInvariantsEmptyIntersectionDrop - standalone regression for the conservation law area(A - B) + area(A intersect B) == area(A), on a near-degenerate StarRing pair where the empty-intersection path used to drop area.
BooleanAssociativitySeeds and BooleanDistributivitySeeds - table-driven associativity of union and distributivity of intersection over union, over fuzzer-derived seeds. These are the strongest checks in the file: the algebraic identity is its own oracle, so there's no precomputed expected area to get wrong.
SubtractInvariantsSeeds - the subtraction conservation law across a seed table (one seed, DominantSpikeStars, checks the inclusion-exclusion identity instead).

They're all area-based algebraic invariants rather than fixed-value assertions, so they keep holding as the arrangement internals shift.

elalish

This is looking pretty good, perhaps we should get this merged so we can take a more serious look at #1707 (comment)?

elalish · 2026-06-15T10:31:26Z

+// overlap geometrically. (A - B) + (A ∩ B) ends up at 39605 versus
+// a.Area=90895, missing ~51290 area units. The tiny x translation puts B's
+// vertices just inside the eps band of A's; the intersect path appears to
+// mis-classify the overlap region as empty.


I'm confused - the comment says there's a problem, but the test appears to be passing. Did this get fixed? If so, perhaps update the comment?

Fixed - passes on the boolean2 backend this is built on. I didn't bisect which commit closed it, so I left the attribution out of the comment and just rewrote it to past tense.

elalish · 2026-06-15T10:36:26Z

+};
+}  // namespace
+
+TEST(CrossSection, BooleanAssociativitySeeds) {


What is the distinction between cross_section_test and boolean2_test, particularly if this test is here?

Good point - moved the one direct-engine test (the Boolean2D non-finite check) to boolean2_test in a separate commit, so this file is now just the CrossSection public API while boolean2_test keeps the engine internals. This seed's a public-API algebraic identity, so it stays here.

elalish · 2026-06-15T10:38:57Z

+    //   the TinyVsLargeStars needle case - this is two large spiky
+    //   shapes, not a tiny shape vs a needle. May share the off-axis
+    //   T-junction root cause diagnosed earlier in this table, but the spike-
+    //   collision geometry could be its own failure mode.


Can we update all these comments to their resolution?

Good call - trimmed each to the distinguishing geometry plus the past-tense outcome, dropped the speculative root-cause notes. All green now.

elalish · 2026-06-15T10:41:02Z

+    std::cerr << "[seed] " << c.name << std::endl;
+    const CrossSection a = MakeShape(c.a), b = MakeShape(c.b);
+    if (c.kind == SubtractKind::InclusionExclusion) {
+      const auto aIb = a.Boolean(b, OpType::Intersect);


Can we use a ^ b instead?

Done in the subtract-invariant tests (here and SubtractInvariantsEmptyIntersectionDrop). There are other Boolean(.., Intersect) sites in the file if you'd like the whole thing converted - happy to, just kept this focused.

Yes, I generally prefer to update all related code at once, simply so it's not forgotten. Not a big deal though.

elalish · 2026-06-15T10:42:54Z

+      const auto aUb = a + b;
+      const double tol = AreaTol(a, b);
+      EXPECT_NEAR(aUb.Area(), a.Area() + b.Area() - aIb.Area(), tol)
+          << "inclusion-exclusion violated";


Why does this branch test fewer things? Do they fail?

They hold - tested it (subtract residuals ~1e-14 against a ~3e-6 tol). Inclusion-exclusion was just the identity that originally broke for this seed, so its standalone test only checked that one. Unified the loop to run all three for every seed and dropped the SubtractKind split.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

zmerlynn · 2026-06-15T15:11:53Z

This is looking pretty good, perhaps we should get this merged so we can take a more serious look at #1707 (comment)?

Yup, I will work on cleaning this up. Ignore the force push just now, just reconciling branches- no new content yet.

…ents Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

zmerlynn · 2026-06-16T00:58:49Z

Sorry for the churn - I hadn't run the full style pass on these test files before putting the PR up, which is why a couple of your comments turned into broader cleanup. Done now: one anonymous namespace instead of six, the fuzz-diagnosis prose trimmed out of the seed comments, ^ in the subtract tests, and the SubtractInvariants kind split collapsed (every seed runs all three identities). Replies inline.

elalish

LGTM - I'll probably run across other things in here eventually, but better to have it in and iterate. Thanks!

zmerlynn mentioned this pull request Jun 11, 2026

Flatten the boolean2 engine into src/ and consolidate to fewer files #1754

Merged

zmerlynn mentioned this pull request Jun 11, 2026

Make the intersection-vertex merge eps-only #1757

Merged

elalish requested changes Jun 13, 2026

View reviewed changes

elalish approved these changes Jun 15, 2026

View reviewed changes

zmerlynn mentioned this pull request Jun 15, 2026

Add experimental boolean2 CrossSection backend #1707

Open

11 tasks

zmerlynn and others added 13 commits June 15, 2026 15:03

Add boolean2 and CrossSection regression and corpus tests

7e18547

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Add TinyFeatureNearCorner annihilation regression tests

fe62b9d

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Add regression tests for the eps-only new-to-old merge

510042e

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Drop tolerance from the boolean2 tests

d1ecf71

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Add DISABLED_ tiny-feature point-contact host-drop seeds

4a36387

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Add contracts to boolean2_test validators

96bbbde

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Share the polygon-corpus reader and inline single-use test helpers

254ec39

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Tighten offset, Simplify, and rhombi-union assertions

c131901

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Use linalg helpers instead of hand-rolled vector math

ef7f5fc

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Note the validator oracle's >eps precondition

56d3fcb

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Assert the empty intersection in the disabled host-drop seeds

8ebe4d5

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Strengthen cross-section area oracle tests

b5c680a

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Guard the offset corpus test under MANIFOLD_NO_IOSTREAM

116c530

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

zmerlynn force-pushed the pr/boolean2-tests branch from 4f8039b to 116c530 Compare June 15, 2026 15:06

zmerlynn and others added 2 commits June 16, 2026 00:44

Consolidate CrossSection test anonymous namespaces and tidy seed comm…

25e3c78

…ents Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Move the direct-engine NonFinite test out of cross_section_test

a2ee932

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

elalish approved these changes Jun 16, 2026

View reviewed changes

elalish merged commit ecc145f into elalish:master Jun 16, 2026
43 checks passed

Conversation

zmerlynn commented Jun 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

codecov Bot commented Jun 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

zmerlynn commented Jun 11, 2026

Uh oh!

zmerlynn commented Jun 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

elalish left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

zmerlynn commented Jun 13, 2026

Uh oh!

elalish left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

zmerlynn commented Jun 11, 2026 •

edited

Loading

codecov Bot commented Jun 11, 2026 •

edited

Loading

zmerlynn commented Jun 11, 2026 •

edited

Loading