Streams with embedded closed captions (e.g. OBS's caption API feeding a live transcription) carry a trailing CEA-708 caption SEI after some frames' slices. When Packetize's h264parse re-parses the byte stream it splits that SEI into its own slice-less, timestamp-less AU, which then got counted and sent to WebRTC viewers as a standalone video "frame": strict decoders (iOS VideoToolbox) error on a picture-less access unit, and since playback ignores PLI the picture stayed broken until the next keyframe — with long-GoP sources, seconds away. The bogus frames also inflated the frame count that the sender's synthesized per-frame timing divides by. Fold slice-less buffers into the next real frame instead — a leading SEI is where captions normally live, so caption-aware receivers still get them. A remainder at segment end has no frame to ride with and is dropped. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>