Skip to content

Add property testing suite - #368

Merged
gpeairs merged 17 commits into
mainfrom
gp/geomlaws-phase0
Nov 9, 2026
Merged

gpeairs merged 17 commits into
mainfrom
gp/geomlaws-phase0

Conversation

@gpeairs

@gpeairs gpeairs commented Nov 5, 2026

Copy link
Copy Markdown
Member

This adds a property test suite for DeviceLayout. See test/laws/README.md. Currently it covers the boolean operations (union2d, intersect2d, difference2d, xor2d, clip, clip_tiled, but not _curved or _layerwise variants), offset, interiorcuts, path halos, transformations, units, and conversion to and from the Clipper integer grid.

It's meant in the near term to validate migration to Clipper2 with higher confidence than our current test suite gives, but it has already caught some issues on our side, as well as some edge cases in Clipper1 and/or Clipper2 that we can patch or warn/error for (see e.g. #349, #350; test findings also drove building Clipper2 with HI_PRECISION for the Julia wrapper).

Instead of checking fixed inputs against expected outputs, each test states a property ("law") that should hold for every input, such as "union is commutative" or "a point is in A ∪ B exactly when it is in A or in B", and checks it on inputs generated by Supposition.jl. Outputs are compared as regions of the plane, using exact rational arithmetic that never calls the clipping library under test.

I am still thinking about review for a 5k line PR like this, mainly AI-authored with human feedback. It only adds tests, which seem generally good but do take time to run (2 min on normal CI pipelines, with a longer weekly run on main). They also add some maintenance burden if they turn out to be too strict -- we may not want to block on issues that only come up with extremely sharp corners, with points meant to be collinear with diagonal lines, or at Clipper grid scale (femtometers).

@codecov

codecov Bot commented Nov 5, 2026

Copy link
Copy Markdown

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

Adds test/laws/ with a LawsSetup testsnippet (Supposition.jl generators,
exact lattice predicates, region equivalence), smoke laws for union2d
(B1 commutativity, B3 idempotence, B7 membership oracle, B8 validity),
and a ledger of three shrunk Clipper1 integer-lattice counterexamples
asserted with @test_broken. Supposition is a test-only dependency.
- Made region comparison conservative
    - ≐ now passes only canonical boundary equality after collinear stripping.
    - Removed the unsound sampled-equality, area, boundary-distance, and containment logic.
    - Sampling remains only in B7, where inconclusive examples are rejected rather than passed.

- Removed clipping from generators
    - Deleted the clipping-based notched-rectangle generator.
    - Replaced it with directly constructed rectangle and star generators.

- Guaranteed degeneracy cases
    - Rectangle pairs now deterministically cover translated/overlapping, shared-edge, vertex-touching, collinear-vertex, and unit-sliver cases.
    - Added independent assertions for each relationship.

- Replaced circular tests
    - Added fixed fixtures for orientation, segment intersection, simplicity, point membership, area, and the thin-triangle disappearance.
    - Marked integer B3 as known-broken and updated L3 to assert its normal-form failure with @test_broken.
Extends test/laws/ from four smoke laws to the phase-1 catalog: B1-B11
(boolean algebra, fill rules, tiled clipping, keyhole cuts), O1-O8
(offset and halo), T1 (rigid-transform equivariance), T4 (coordinate-unit
invariance), and D1 (prescale round trip).

Setup changes:
- Generator families: thick and thin integer stars, rectilinear
  staircases, float stars at nm-100 um scales, fm-perturbed float pairs,
  triples, polygon sets, and hand-built ClippedPolygon trees with holes.
- ClippedPolygon results are compared through their tree contours so the
  keyhole cut is tested separately (B11).
- Exact comparison (≐) remains the verdict wherever calibration shows the
  law holds exactly (rectilinear inputs; commutativity, identities,
  single fill rules). `region_within(X, Y, tol)` states rounding laws:
  exact area budget of tol times the summed perimeters plus membership
  agreement at probes farther than tol from both boundaries.
- Float Quantity coordinates now map to the fm lattice correctly.

Ledger adds L4 (hole vertex rounded outside its parent), L5
(`interiorcuts` silently drops such a hole; kernel-independent DL bug),
L6 (zero-length keyhole cut), L7 (small offsets leave a sharp corner's
vertex on the output), and L8 (fm-scale self-crossing difference output).

151 pass / 24 broken in ~2 min at PR budget.
…ger L9

Offset laws that compose operations (O3, O5, O6) are stated only where
|δ|·sin(turn) ≥ 1 at every vertex (outside the L7 regime) and where the
first offset produced no edge shorter than 4 lattice units; L9 records
the micro-edge mechanism with a reproducer. Normal form (B3) compares
regions, since the kernel may split or merge contours touching at a
vertex depending on operand order.
An edge of length L with endpoints rounded by half a unit has its
direction wrong by ~1/(2L), so a second offset by δ moves the adjacent
intersections by ~|δ|/(2L). O5 and O6 use that bound instead of a fixed 2
and no longer reject inputs with short edges; L9 asserts the bound. T1 on
rectilinear inputs compares regions.
… joins only

Re-union can drop a sub-unit spike left by a rounded intersection, so the
normal-form law is exact only on rectilinear input. Erosion with miter or
square joins cuts deeper than a disk at concave vertices of the grown
polygon, so grow-then-shrink is not guaranteed to contain A with those
joins; O6 is stated for round joins.
Rational region certificates replacing sampled acceptance, including cancellation of internal seams
Correct float-offset scaling and enforced valid-example budgets
Whole-edge hole containment, winding checks, and narrowly scoped L6 handling
Full-chord O8 checks, with regression tests for the review counterexamples
@gpeairs
gpeairs force-pushed the gp/geomlaws-phase0 branch from 8af073d to 554b990 Compare November 9, 2026 13:48
@gpeairs
gpeairs merged commit 8cbce04 into main Nov 9, 2026
11 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant