Defer heavy model imports to reduce cold start and RAM at boot

[rosspeili]

Problem

OPSIIE_0_3_80_XP.py imports transformers (BLIP), diffusers, torch, torchaudio, face_recognition, DeepFace, etc. at module top (lines ~10–41). docs/ARCHITECTURE.md already calls out “heavy imports first.” This hurts time-to-interactive and headless/CI scenarios.

Goal

Lazy-import or factory-init for expensive stacks (e.g. BLIP URL vision, Stable Diffusion /video, music) on first use, with clear logging when a stack is loading.

Acceptance criteria

• Measure before/after: time from process start to first prompt (document in PR).
• No user-visible command regression without an opt-in “preload” if you add one.
• Document any new first-hit latency in docs/TROUBLESHOOTING.md or architecture doc.