The clips were silent screen captures - the user's actual ask was for someone talking
through what's happening. Each clip step now carries a narration line synthesized
offline via espeak-ng; the recording paces itself to each line's real spoken duration,
and the final audio track is built from the real measured per-line start times (adelay +
amix) so narration and on-screen action stay in sync over a 30+ second clip instead of
drifting against an idealized schedule.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
788319d098