GGUFEngine runs the one-pass readout on llama.cpp via llama-cpp-python (extra gguf, >=0.3.35,<0.4); Decider(path.gguf) or
Decider(repo, gguf_file=...). One row per decode by default; context n_batch = n_ctx (40,960). Verified over the regression set
on decider-2b v11 Q4_K_M (identical to the research readout), unit tests against a fake llama_cpp. Issue #16.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>