[qianlim]

Summary

Make the recommended Transfer-Auto configuration for 10-FPS RGB-only inputs explicit:

• Keep synchronized 10-FPS input frames unchanged.
• Keep model FPS conditioning at 30 (fps: 30 / --fps 30), matching training.
• Encode the generated frames at 10 FPS for the intended playback speed.

Replace the previous recommendation to match model FPS to source FPS and the "diagnostic only" framing. Explain that changing the model's FPS conditioning to 10 has been reported to reduce quality; this PR does not claim a new model-quality evaluation.

Include a frame-preserving FFmpeg playback conversion, distinguish caption frame bounds from textual timestamps, and retain the separate joint RGB/LiDAR clock/alignment requirement. Update the inference skill's short pointer so it no longer contradicts the reference document. Preserve the existing FPS section anchor. No runtime changes.

Validation

• Repository pre-commit checks passed for both edited Markdown files.
• git diff --check passed.
• Synthetic FFmpeg smoke test: 297 decoded frames before and after conversion; 30 FPS / 9.9 seconds becomes 10 FPS / 29.7 seconds.
• Checked the unchanged source-FPS/model-FPS/saver behavior against the current multiview inference implementation.
• No model inference or quality benchmark was run for this documentation-only change.