Hook and cover variants

A variant changes only the hook scene and the cover; every other scene is shared with the base project, so the experiment isolates what it tests. Paired cuts (durations) add one more axis: every hook × cover pair is also retimed to each length, with the same proportional scaling as adapt.

1. Resolve the project folder (absolute). It needs a valid project/video-spec.json (run the validate skill first if unsure). A rendered base is best: its scene clips are reused.
2. Ask what to test only if the user did not say; default to 3 hooks × 2 covers. Write project/variants.json (schema_get experiment-plan):
   • id, a one-sentence hypothesis ("a question hook keeps more viewers past 3 s than a statement hook"), and metric (e.g. 3s_retention, completion_rate, saves).
   • hooks[]: {id, label, scene}. Each scene is a full hook scene (purpose hook, 1.5–4 s). Use different hook mechanisms from the brief's candidates (question, statistic, contrarian, promise…). Keep every claim grounded: reuse the base hook's claim_refs or cite other evidence; a hook with a number needs a ref. With voice.mode: none, voiceover stays "" and the words go in the props.
   • covers[] (optional): {id, label, cover: {headline (≤ 6 words), focal_time_sec (inside the hook)}}.
3. Call variants {project_dir} to prepare. Fix every variant listed as failed by editing project/variants.json, then call it again.
   • Paired cuts: when the user wants the same piece at several lengths (e.g. "a 15 and a 30"), pass durations: [15, 30]. Each pair then becomes variants/<hook>-<cover>-15s/ and -30s/, and the output lists duration_sec per variant. The rules:
     • at most 4 lengths, no duplicates;
     • each length inside the platform's norms (e.g. 7-90 s for Reels and Shorts) and every target contract's duration range;
     • each length long enough for every scene to keep 0.5 s. The engine refuses bad lengths with the reason.
   • Later calls without durations keep the experiment's cuts; durations: [] goes back to the base length only.
   • Words are not rewritten. For a narrated spec, the notes list every scene whose voiceover no longer fits a shorter cut. Don't render those; make the short cut with adapt instead (its own project, where you trim the narration), then run variants there. Paired cuts work best for voice.mode: none pieces (text over music, motion loops).
4. Call variants {project_dir, render: true} (add quality: "final" when the user wants final renders). Renders queue one at a time; poll job_status for each job id every 10–20 s, or variants {project_dir, status_only: true}. Calling it again is safe: variants already queued or running are not resubmitted, and a variant being rendered is left untouched (listed under skipped). A variant whose render failed or was cancelled shows failed with the job's error; render: true resubmits it.
5. Optionally run lint on one variant folder (variants/<id>) per hook.
6. Report: the hypothesis and metric, then one line per variant (id, hook label, cover headline, cut length if any, status, variants/<id>/dist/). Remind the user to post variants under comparable conditions (same time slot, audience) and to compare only the chosen metric; the plugin does not publish or fetch analytics.