Render Two Thumbnails, Not Four

Write pipeline/thumbnails/thumb.py using Pillow.

It renders EXACTLY two thumbnails: 1280x720 and 1080x1080. One design, two
crops. Never four files.

Design:
- Grab a frame from the video at --t seconds, cropped to the --cam region, and
  use it full bleed, cover cropped, anchored so the subject stays in frame.
- Sharpen and slightly boost contrast on camera frames.
- Overlay a left to right dark gradient for text legibility.
- Text block on the left: --line1 in white, --line2 in the accent color below
  it, then --accent as a smaller gray line, then a brand tag at the bottom.
- Drop shadows behind every text line so it survives a busy frame.
- Support --bg to use a supplied background image instead of a camera frame,
  with a lighter gradient in that case since generated images usually already
  have negative space.

Usage:
python pipeline/thumbnails/thumb.py work/ep-test.mp4 --t 1612 \
  --line1 "AI BUILDS A" --line2 "WEBSITE LIVE" \
  --accent "FOR $3 IN AI FEES" --out episodes/<date>/thumbnails

Write both files and print the paths.

Create skills/openmic-copywriter/SKILL.md and run it.

It writes episodes/<date>/copy.md for human review and
episodes/<date>/schedule.json for the publisher.

Every caption serves exactly ONE goal, matching the split I set in step 1:
[YOUR SPLIT].

Copy rules:
- Headline is the clip's hook sharpened, 8 words or fewer, number or tension
  first.
- Caption structure: one line hook restate, then one or two lines of real
  substance with an actual number or takeaway from the clip and never anything
  vague, then the CTA, then 3 to 5 niche hashtags. No hashtag walls.
- Per platform variants: square and feed formats get the full caption, the
  short form vertical platform gets a shorter rawer version with no links,
  the video platform gets a title of 60 characters or fewer plus a one line
  description, the professional network gets one extra line of context and no
  hashtag spam.
- Longform: curiosity title 60 characters or fewer, description with the
  premise line, chapter timestamps, and my links.
- Voice: mine. Plain spoken, numbers forward, no corporate filler. Pull actual
  phrasing from the transcript.

schedule.json shape, one entry per asset:
{ "file", "type": "short|longform", "goal", "keyword": "WORD or null",
  "platforms": [...], "scheduleDate": ISO8601, "copy": { per platform } }

Spread roughly two shorts per day across the week with the longform on its
regular day. Show me copy.md before writing schedule.json.