Dramatic Intent Search

Add features that save days, not minutes. Reference: nobody does these. Legal boundary: we describe TECHNIQUES and reference films as inspiration in text, we never copy frames or label prompts with a director's name as a product.

1. SEARCH BY DRAMATIC INTENT (the biggest gap — people don't search for "pretty shots", they search for "how do I show X"):
   Add an optional field `dramatic_intent` + a filter axis with values like:
   - Show hidden fear / Reveal a lie / First meeting / Point of no return / Realization / Aftermath of violence / Building dread / Establish power / Show isolation / Moment of decision / False calm before the hit
   AI auto-fill should infer this from the prompt text. Filter + browse by it. This is director's knowledge no generator has.

2. SHOT PAIRS (cinema lives in the CUT, not the frame):
   Allow linking two prompts as a pair: "Shot A + Shot B = the meaning". Add `paired_prompt_id` + `pair_note` ("cut from the wide to this reaction — the delay is what sells the fear"). In the detail view show the paired shot with a "Copy both" button. Users generate two clips and get a working edit, not an orphan image.

3. SCENE BREAKDOWNS (highest value): a content type where a known cinematic MOMENT (described in our own words — never copying footage) is broken into 4-8 shots with prompts that reproduce the TECHNIQUE, plus a note on why it works ("the tension comes from holding the wide 3 seconds longer than comfortable").
   - Title format: technique-based, e.g. "The slow-burn standoff — how tension is built with silence", NOT a director's name.
   - Each breakdown = an ordered list of prompts + a short director's note per shot + "Copy all".
   - This is a teaching artifact AND a production shortcut.

Legal rules for CC: never upload/copy film footage, never use a director's or film's name as the product label, references to films appear only as inspiration text inside descriptions.

Start with 1 (fastest, uses existing auto-fill) then 3. Give me the SQL. npm run build, commit, deploy.