Add knowline_engine.py (one-command Decision Index runs); --backend hf without accelerate

knowline_server.py: accept chat roles the template rejects; return 500 instead of dropping the connection

INFERENCE.md and knowline_server.py header: describe the front end directly

Self-contained knowline_server.py front end; drop the llm2jev dependency

KnowLine-4B-Gen1: merged bf16 weights (E step 5650), tokenizer, serving files