Persistent AI chat history + token efficiency

The AI chat lived only in React state — closing the panel or refreshing lost it, every open re-ran the briefing, and every turn resent the full persona + board context uncached. Now the conversation persists per space and the repeated context is cached.

Added

• supabase/migrations/20260620000100_ai_messages.sql — ai_messages table: one rolling conversation per space (the space is the thread, no separate sessions table). RLS mirrors cards (members read, editors write). Indexed on (space_id, created_at) for chronological load.
• GET /api/ai?spaceId=… — loads a space's saved thread (RLS-scoped).

Changed

• app/api/ai/route.ts
  • Persists the opening briefing and each Q&A turn (assistant brief; then the new user question + assistant reply per follow-up — no duplicates). Best-effort and signed-in/cloud only; local spaces stay ephemeral, guarded by a uuid check + RLS.
  • Prompt caching — the static persona + board prefix is sent as an Anthropic cache_control: ephemeral block, so resending it each turn is billed at a fraction (the biggest cost lever, vs. trimming context and hurting answers).
  • History cap — replays only the last MAX_HISTORY (24) turns to the model.
• components/AIPanel.tsx — on open, resumes the saved thread if one exists (no re-briefing, no scan); only a genuinely fresh space runs the briefing + scan. Sends spaceId; caps replayed turns to MAX_TURNS (24).
• app/page.tsx — passes spaceId (cloud spaces only) into <AIPanel>.
• lib/supabase/types.ts — added the ai_messages table type.