Multi-shape bodies report the geometry that actually collides

[obiot]

A body built from several shapes came apart under rotation on matter, and swung away from the body on planck. One root cause, two faces, introduced by the getBodyShapes rotation fix in matter-adapter 1.2.1 / planck-adapter 1.3.1.

What was wrong

It rotated the authored shapes about the renderable's origin. Both halves of that were wrong:

• Polygon#rotate moves a shape's points without its pos, so a shape carrying its offset in pos spun about its own origin while the origin stayed put. Several such shapes visibly separated.
• The renderable's origin is not what either engine rotates about. Matter turns a body about its centre of mass; planck about its origin, which on a compound body is a third point again. (getLocalCenter() would have been a second wrong answer.)

The fix

Stop deriving the pose and read the engine's own geometry back. There is then no pivot to get wrong, and the report describes what actually collides rather than what was authored: an Ellipse is simulated as a circle of average radius, a concave polygon as convex pieces, a degenerate one as a box, and on planck a shape with isActive: false has no fixture at all. The overlay drew none of that truthfully before. Authored shapes are untouched on renderable.bodyDef.shapes.

Also fixed, all from review of the same code

• updateShape at an unchanged angle reported the previous geometry, and a removed renderable stayed pinned: the cache was keyed by angle and cleared on no path but one.
• The cache missed on every frame of a body that was actually turning (float equality on the angle), and each miss allocated a fresh shape set through the object pools without ever releasing it — on the debug overlay's per-body-per-frame path.
• matter simulated shapes flagged isActive: false, which planck and the builtin both honour.
• planck reported its fixtures newest-first, i.e. the reverse of def.shapes.

Engine, and the builtin

Polygon#rotate(angle, pivot) now turns the whole shape. Ellipse#rotate always rotated its pos; same signature, and Body#rotate calls both over one mixed list, so a body's ellipse parts orbited correctly and its polygon parts did not. Every Rect added to a body becomes a polygon offset by pos, so the documented feet-and-torso shape was among them. A polygon at the origin — every caller in the engine until now — rotates to exactly the numbers it always did.

The builtin's getBodyShapes is deliberately unchanged. Its SAT never reads body.angle, so an unrotated hitbox around a spinning sprite is the solver reporting what it genuinely collides as; making the overlay follow the visual angle would turn an honest inconsistency into a confident lie. Three tests pin that contract so it is not "fixed" later.

One claimed defect that did not survive measurement

planck's getBodyAABB reading an "enlarged and stale" broadphase proxy. Measured against the exact AABB across steps at 2000 px/s: a constant 0.001 m (0.03 px) that does not grow with speed. The larger gap is the polygon skin radius, which is real collision geometry. Left alone.

Verification

• matter 196, planck 194, melonJS 7034, biome clean.
• Every fix has a test, each verified to fail with its fix reverted, on each backend where it applies.
• Repro reproduced and re-verified on all three backends; every physics example swept clean (platformer, platformer-matter, pool-matter, plinko-planck, isometric-rpg, whac-a-mole, space-invaders, jungle-rabbit) — worth doing because Polygon#rotate is core, so the blast radius is wider than the adapters.

Two of my own first test attempts were wrong and were replaced: a "the centroid does not move" invariant that only holds when the pivot is the centroid (false on planck), and a rotation check that encoded matter's pivot. Both are pivot-agnostic now.

Versions

matter-adapter 1.3.0, planck-adapter 1.4.0 — minor rather than patch, because getBodyShapes now returns different object types. Peer stays >=20.0.0: nothing here needs a newer engine, and the readback never calls rotate, so it does not depend on the core fix either.

🤖 Generated with Claude Code

https://claude.ai/code/session_01NGvtaUNATVCVxD2qcbiY4t

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.