The bench/quality README documents an A/B flow using \
un_quality.py --backend llama-server --llama-url\ and an orchestrator script, but the code path is missing.
This PR restores the documented functionality so the quality parity suite is reproducible from the repo docs (CPU-verifiable via the runner's self-test, no GPU required).
The bench/quality README documents an A/B flow using \
un_quality.py --backend llama-server --llama-url\ and an orchestrator script, but the code path is missing.
This PR restores the documented functionality so the quality parity suite is reproducible from the repo docs (CPU-verifiable via the runner's self-test, no GPU required).