cfg-0094

aged evidence — reconstructed from the archive. Ingested from llama-bench, which measures throughput and not footprint: the memory block, chat-template hash and model revision are absent by construction. Peak memory comes from a Tier 0 capability probe, not from this tool. Everything present here — build commit, backend, KV quant, batch sizes, flash attention — is machine-reported. llama-bench's own -ctk/-ctv fields report f16 by default (no override passed on the command line; matches this candidate's 2026-08-16 screen, cfg-0093). Depths 0/32768 ran N=3 fresh-process reps (fresh server per rep, page cache not deliberately dropped between reps -- see queue.log); 131072 ran N=1 (deepseek-v4-flash/laguna-s21 convention: the number does not move at this scatter level on this box, and time is precious on a ~35-37GB resident model). Scatter across the 3 reps was clean at every depth (max CV 0.9%, d0 prefill) -- no defective-path signature. No MTP/speculation: this job's own from-scratch GGUF header parser (no numpy/gguf-py dependency available on-box) read all 733 tensor names directly and found zero nextn/mtp/eagle/medusa/draft tensors -- plain decode throughout. Output sanity gate passed before recording any throughput (chat-endpoint coherence check, 'the quick brown fox' echoed correctly).