Loading weights wasn't enough — the first forward still cost ~30s on torch's
one-time graph/thread init. Warm now runs a throwaway _embed_texts() so the first
USER query is ~1.5s, not 30s.