Physics: a renderable that places itself collides at its pos - #1694

Follow-up to #1693, which shipped a hole in its own fix, and a second defect it left behind.

A renderable that places itself

#1693 moved every backend onto the frame a renderable DRAWS in, subtracting size * anchorPoint from where a body is built. That assumed every renderable takes that offset, and the engine already carries a flag saying otherwise: applyAnchorTransform, which is what preDraw reads before shifting anything. GLTFModel clears it in its constructor, Mesh clears it on the Camera3d world-space path, both because they emit world coordinates and pivot about their own model origin rather than a bounds box. They draw at pos while their anchorPoint still holds the default (0.5, 0.5), where it means nothing.

Reading it anyway moved each of those bodies by half its OWN bounds box. A scene sizes that box per node, so two objects that overlap on screen were displaced by different amounts, and the contact between them was not merely shifted, it was never reported at all. The Jungle Rabbit example is built entirely from such renderables and lost its collisions completely.

The same condition now guards all five places that compute the offset:

The two external adapters are in scope despite never seeing a Box3d: the flag is a plain Renderable flag, so an ordinary Rect hitbox on such a renderable was mis-built there too. The parity case added to both is exactly that, and it fails on both without the fix.

raycast3d measured in a different frame

#1693 moved the narrowphase onto the drawn frame and left raycast3d building its world AABB from raw pos. For any anchored body the two then disagreed by half a renderable, so a floor probe placed a character on a surface the solver does not have there. Nothing failed, because every existing case in that spec was one body queried against itself and no test compared the query to the narrowphase. raycast3d-box3d.spec.js now authors in the collision frame and has a case pinning the two together.

Tests

Six new cases, each verified to fail with the defect put back:

• box3d-world.spec.js — a renderable that places itself collides where it draws, and one that does not still reads its anchor (mutants in both directions killed)
• builtin-resting.spec.js — the same contract through a full 2D drop
• raycast3d-box3d.spec.js — a ray hits the box where the narrowphase collides with it
• parity.spec.ts on both adapters — a Rect body on a renderable that places itself

The first draft of the Box3d pair was vacuous: sizing each box from its own renderable's bounds made the anchor shift scale with the box, so the contact survived it. The boxes are now sized independently of the bounds they hang off.

melonjs: 7148 passed. matter: 225. planck: 220. debug-plugin: 23.

Releases

@melonjs/matter-adapter 1.5.1 and @melonjs/planck-adapter 1.6.1, dated and ready to dispatch on merge, since 1.5.0 and 1.6.0 went to npm carrying this. melonjs 20.7.0 stays _unreleased_; its entry is amended rather than added to, since the hole never shipped there.

🤖 Generated with Claude Code

https://claude.ai/code/session_01NGvtaUNATVCVxD2qcbiY4t