Cold-Start Design for New Users and New Items

You're designing recommendations for a short-video app. Two cold-start cases: (a) a user who just signed up 30 seconds ago with zero interaction history, and (b) a video uploaded 2 minutes ago with zero views.

1. Design the candidate generation and ranking treatment for case (a).
2. Design the candidate generation and ranking treatment for case (b), including how it eventually stops being "cold."
3. Your two-tower embeddings are re-trained and the full catalogue re-embedded weekly. What specific problem does this create for case (b), and how would you fix it without retraining more often?