[sysradium]

On macOS in Firefox, the Innomaker U20CAM preview stayed grey even though macOS showed the camera as in use. It would open at 640×480 but send no frames until another app (e.g. FaceTime) started it at a larger size. OpenCV gets frames at 640×480, so recording was never affected; only the browser previews were.

Tested in Firefox with nothing else using the camera:

Requested size	Frames in 3 s
1920×1080	80
1280×720	80
640×480	0

So with this pr:

• useCameraStream takes the camera's saved capture size and requests it first. If no frame arrives within 2.5 s, it reopens at 1280×720, then 1920×1080. All sizes are ideal constraints, so a camera that can't meet them still opens. Cameras that stream at their saved size never reach the fallback.

• BrowserCameraPreview sizes its box from the saved capture size instead of the stream's actual size. cameraPreviewGeometry now uses object-fit: cover, so a stream at a different aspect ratio (e.g. a 16:9 fallback for a 4:3 recording) is center-cropped and the preview shows the framing that will be recorded.

• Removed the metadata state (setMetadata on loadedmetadata) from BrowserCameraPreview. It stored the stream's actual videoWidth/videoHeight and sized the box from them, which is what made a 16:9 fallback stream show a wider view than the 4:3 recording. Now that the box always takes the saved size's shape, nothing reads the stream's size. Its deviceId check, which stopped a reassigned preview from reusing the previous camera's size, is no longer needed either. One consequence: a preview no longer shows when the browser picked a different size, because the image is cropped to the recorded framing.

• Recording still captures at the saved size; LeRobot raises an error if it can't set it.

• The recording page's live view (/camera-feed) already showed the recorded frames and is untouched.

• Teleoperation doesn't open cameras on the backend.

Notes

• For the Innomaker, the 640×480 mode is a center crop of the sensor. Template matching OpenCV captures put it at the middle 1440×1080 of the 1920×1080 frame (x=236; exact center is 240). So the cropped fallback preview matches the recording for this camera. A camera that scales the full sensor down instead of cropping could differ slightly at the edges.
• A camera that stalls takes about 2.5 s longer before its preview appears.