Switch JD-parse/T1 polish to Qwen3-1.7B

Cache-first model loading for zero-network ranking

Add CPU threading, T1-only LLM polish, experiment script

Reuse Qwen3.5 for T2; drop LFM2.5 & factor ONNX

Explain: research-based outcome notes (prompt-engineered + verified)

Docker: pre-download all models at build for offline ranking + docs