Add bm25 prefilter stage; unify candidates cache; validate reasoning quality

Add timing and error logging throughout

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

Enhance JD extraction (Qwen3.5-2B + richer prompt) and Step-11 reasoning

Add ONNX granite/reranker and compute config

Add retrieval & soft/composite ranking stages