bug: VLM input plugins crash on dropped camera frames (ret not checked)

[Ridwannurudeen]

Description

Three VLM input plugins call cap.read() but never check the ret return value before passing the frame to downstream processing. When the camera drops a frame (returns ret=False, frame=None), the plugins crash instead of gracefully skipping.

Affected files

1. src/inputs/plugins/vlm_local_yolo.py:221 — ret is ignored, frame (potentially None) is passed directly to self.model.predict(), which raises a runtime error.

2. src/inputs/plugins/vlm_coco_local.py:131 — ret is ignored, frame (potentially None) is returned and passed to PyTorch model inference.

3. src/inputs/plugins/webcam_to_face_emotion.py:89 — Same pattern, ret ignored.

When this happens

• Virtual camera devices in simulation (Gazebo/Isaac Sim) may not have frames ready immediately at startup
• Physical cameras that temporarily lose connection (USB disconnect, power fluctuation)
• High system load causing the camera driver to drop frames

Expected behavior

The plugins should check ret and return None (skip the frame) when cap.read() fails, matching the pattern already used in unitree_realsense_dev_vlm_provider.py:98-109.

Steps to reproduce

1. Configure VLM_Local_YOLO with a camera index that points to a virtual device not yet streaming
2. Start the OM1 runtime
3. The plugin crashes on the first _poll() call when frame is None

[farukk77]

#2235
Fix crashes reported in issue #2207 where cap.read() can return ret=False or frame=None and invalid frames were passed into model/cv pipelines. Prevent resource leaks from probe VideoCapture handles that were not released in check_webcam(...) helpers.