[obiot]

Engine fixes found while building a new isometric example (coming in its own PR), plus the skill updates and a Line of Sight example that shows the raycast fix.

Tiled

• Collision objects whose shape reaches left of or above their position, like every rectangle on an isometric map, now have bounds covering their shape. The shape object kept its bounds at pos .. pos + size, so getBounds() (what the quadtree files it by, and what queryAABB and raycast prefilter on) sat beside the drawn wall, and a character could walk through it wherever the quadtree split differently. The object now starts at its shapes' top left corner and the shapes are translated back, so the vertices are unchanged. Such an object's pos changes; point objects keep theirs.
• Tile objects on isometric maps collide over their image, as on other maps. parseTMXShapes ran the isometric projection on a tile object's image box, which Tiled already draws upright, giving a slanted body hanging off the image.

Physics

• The builtin raycast honours the anchor, as collisions have since 20.7 (Physics: collision shapes follow the renderable, and segments collide as segments #1693). It placed each shape at pos + shape.pos, so on a body anchored anywhere but its top left (a Sprite defaults to its centre) a ray hit half the body's size away from where it collides, or missed it. It now uses the same anchor offsets as SAT (anchorOffsetX / anchorOffsetY, exported from sat.js), honouring applyAnchorTransform = false. The 3D raycast already did this.

Camera

• A zoomed camera follows, clamps and focuses in world units. follow (deadzone and clamps), moveTo's bounds clamp and focusOn used the on-screen width / height, while a zoomed camera shows width / zoom of the world: zoomed out, a followed target sat off centre and the view scrolled past its bounds; zoomed in, it stopped short of them.
• localToWorld / worldToLocal take zoom and screenX / screenY into account, so a pointer on a zoomed camera, or on a second camera drawn in a corner, hits what it points at. world = pos + (screen - screenOffset) / zoom, then the existing world offset and camera transform steps; worldToLocal is the exact inverse.

Also

• Text: an empty or blank label gets a canvas at least one 32 px bucket wide. A zero-wide canvas made WebGPU refuse the upload and lose the frame.
• Typings: TextureAtlas#getAnimationSettings() is typed as the settings it returns, and Sprite3d accepts named animations (as an aseprite atlas carries them).
• Skills: tilemaps (what sortOn = "y" can and cannot sort, and the default bodies of class-less tile objects), drawing (create a CanvasRenderTarget once, since its GPU texture lives as long as the renderer, and invalidate() it after repainting), lighting (a sprite without a normal map is not shaded at all, not "lit flatly").
• Line of Sight example: round pillars with Ellipse bodies anchored at their centre, and patrolling intruders the cone spots with a raycast check.

Verification

• New specs in tmx-shape-factory, tmxtilemap, raycast, camera, pointer-world-offset and text_bucket; every one fails with its fix reverted, and each half of the conversion fix (zoom, screen offset) is caught separately.
• Full suite: 8,304 passed, 10 skipped. Lint clean.
• The isometric example that surfaced these runs on WebGL, Canvas and WebGPU: a wall probe walking into every wall from 8 directions finds no clipping, and its minimap camera now uses follow() with a zoom of 0.1.

🤖 Generated with Claude Code

https://claude.ai/code/session_01SeWCkETjW35ZbXDRswV1hx

[Copilot]

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