feat: adaptive deck — top-up recommends based on what you liked

[valeboth]

As you swipe right, the deck now adapts toward what's actually being liked — not just fixed onboarding tastes.

How it works

On each top-up (extendDeck), before the genre-discovery slice, it now:

1. Reads the room's most-recent likes (direction='like').
2. Expands them into TMDb recommendations (the same getRecommendations path onboarding seeds already use).
3. Mixes them in personalized-first, deduped into the shared pool.

Shared pool preserved (matches still work) + solo

Likes are read from both users, so the recommendations they generate land in the shared rooms.deck — both users can see and therefore match on them. In a solo room the same query naturally returns just that one user's likes, so browsing alone gets personalized recommendations too.

Free-tier safety

• Only the 3 most-recent likes are expanded per top-up, ≤8 recs each.
• All via the 24h-cached getRecommendations → a handful of mostly-cached TMDb reads per top-up.
• No extra KV/D1 writes beyond what top-up already did; still capped at MAX_DECK = 300.

Cleanups + tests

• Recs aren't genre-filtered by TMDb, so avoid_genres is re-applied via a new pure helper rejectAvoidGenres, which also de-duplicates the inline filter in generatePool.
• Unit tests for rejectAvoidGenres; typecheck, lint, vitest (8/8) and wrangler --dry-run all pass.