[Enhancement] mobieusAI: vector embeddings + semantic search for the Q&A flagship

What

Replace the FULLTEXT + LIKE retrieval that powers the mobieusAI Q&A flagship (#40) with cosine-similarity over per-tenant embeddings. Today retrieval is keyword-driven, so "how do I add a mod" surfaces threads about "moderator" but not threads phrased "promote a member to staff." Semantic search closes that gap.

Where it lives

• Surface: Q&A drawer (every page when logged in)
• Backlog id: B-AI-RAG-1

Implementation notes

• BYOK embeddings provider (Voyage AI or equivalent) — same per-tenant key model as Anthropic
• New ai_assistant_embeddings table per tenant: (source_table, source_id, embedding BLOB, dim, updated_at)
• Backfill worker: nightly recompute deltas, plus on-write hook for new content
• Top-K retrieval at query time: cosine vs. the question embedding
• Falls back to the existing FULLTEXT path when no embeddings are present

Tracked as B-AI-RAG-1 in docs/BACKLOG.md.