Seed titles (autocomplete + recommendations) + weighted-random deck order

[valeboth]

Seed titles + weighted-random deck order

Seeds ("titles you loved" — strongest signal)

• Autocomplete: GET /api/search?q=… → TMDb search (KV-cached), 300ms debounce; a dropdown with poster + title + year + Movie/TV. Picked titles become removable chips.
• Stored in profile.prefs.seeds ({tmdb_id, media_type}, max 6; ≤3/user used in the pool).
• Pool with seeds: recommendations (/recommendations, /similar fallback) of both users' seeds drive the pool, plus a common-ground genre slice so overlap always exists. avoid_genres enforced on the recommendations (the API doesn't filter them). Union across both users. No seeds → the genre-union pool (unchanged).
• Seeds only apply to the room's media type (recs of a movie are movies).

Randomizer

• Deck order is now a per-user weighted shuffle at serve time: taste-biased (your favourites lift toward the top) but randomized — different each time and between the two users, so you don't always see the same sequence. Same shared pool → you still converge and match.

Verified local

• Autocomplete "inception" → Inception (2010) ✅
• Seed = Inception → deck is its recommendations (Oblivion, Insurgent, …) ✅
• Seed + avoid Sci-Fi → zero Sci-Fi titles in the deck ✅
• Two users' seeds (Inception + The Notebook) → pool has both sci-fi and romance ✅
• Two /deck calls → different order ✅
• lint/typecheck/build ✅

Free tier

• Autocomplete: 1 TMDb call per debounced keystroke, KV-cached (repeats free).
• Seed pool: ≤ ~1–2 calls per seed (≤3/user) + 1 common-ground call, all cached per room → well under the 50-subrequest limit.