Bug: failed recording leaves an empty dataset directory behind

`handle_start_recording` writes the dataset directory (meta/info.json etc.) as soon as recording starts, before the first episode is actually captured. If something fails right away — camera drop, arm disconnect — before any episode gets saved, that directory is left on disk. It shows up in the local dataset list and picker even though nothing was actually recorded.

Should clean up the dataset directory when a recording fails with 0 saved episodes (and it wasn't resuming an existing dataset — don't want to touch a dataset that already had data in it).

Start at the handle_start_recording entry point and trace where the dataset directory is created, episodes are counted, and recording failures are handled. Verify the resume-existing-dataset case before changing cleanup behavior; done means a failed recording with zero saved episodes removes its new directory, while existing datasets remain untouched and empty datasets no longer appear in the local list or picker.

Written by the indexing model from the issue text.