verify-bookendslisted

# verify-bookends Last-line defense before a short is committed to fit-vertical. Checks three things in one Claude call: 1. **Cleanliness** — partial words, breath cutoffs, co-speaker interjections, off-shot frames at either end. 2. **Opening-hook strength (VVSA gate)** — the first ~3s is where a scrolling stranger swipes away. If the opening is pure setup / throat-clearing and a stronger hook line exists within the first 3s, snap t0 inward to that line. Hook-weakness alone is never a drop — only an inward-trim suggestion. 3. **Payoff landing** — if the tail contains a clear payoff word followed by trailing filler ("yeah", "so anyway") or silence, snap t1 to ~80ms past the payoff word rather than the sentence boundary. Dead air after the payoff costs retention. Output remains a single `keep`/`trim`/`drop` decision — same I/O contract, same caller logic in `shorts.sh`. ## Invoke ``` .claude/skills/verify-bookends/verify-bookends.sh <input_clip> <input_trimmed_transcript> <out_decision_json> [VERIFY_BOOKENDS=0 to skip] ``` - `input_clip` — post tighten-pace clip - `input_trimmed_transcript` — post tighten-pace re-timed transcript - `out_decision_json` — written JSON: `{"action":"keep"}`, `{"action":"trim","t0":<sec>,"t1":<sec>,"reason":...}`, or `{"action":"drop","reason":...}` Env: `VERIFY_BOOKENDS=0` makes the skill emit `{"action":"keep","reason":"disabled"}` without calling Claude. ## How 1. ffprobe the clip duration. 2. Slice the trimmed tr