Testing clicks on 3D objects with a browser-automation tool: miss = coordinate-frame mismatch, not a broken picker
While checking that clicking a three.js piece (inside a Matterport/R3F canvas) selects it, screenshot-coordinate clicks kept missing, and it looked like selection had regressed.
Cause: the screenshot's coordinate frame (e.g. 800 px wide) didn't match the page's CSS viewport (innerWidth 834), and small 3D objects leave little margin, so clicks landed beside the piece.
A reliable check: 1. Expose the picker in development only: if (process.env.NODE_ENV === 'development') Object.assign(window, { __fnRaycast: (x, y) => raycast({x, y}) }). 2. Find an NDC point that hits (e.g. __fnRaycast(0, -0.3) returns kind 'item'). 3. Convert it to client pixels from the canvas rect: x = left + width*(ndcX+1)/2, y = top + height*(1-ndcY)/2. 4. Dispatch pointerdown and pointerup on document.elementFromPoint(x, y) (the canvas host), then read the store's selected id; subscribe to the store to see transitions.
That proves the picker and handlers work; then only scale the coordinates for screenshot clicks (multiply by screenshotWidth / innerWidth).