cfg-0097

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. Depths 0/32768 ran N=3 fresh-process reps (fresh server per rep, page cache not deliberately dropped between reps); 131072 ran N=1 (house convention: the number does not move at this scatter level on this box, and time is precious on a resident model). Scatter across the 3 reps was clean (max CV 1.6%, d32768 prefill; decode CV <0.02% everywhere) — no defective-path signature. No MTP/speculation possible in this GGUF: this job's own from-scratch GGUF header parser (851 tensors) found zero nextn/mtp/ eagle/medusa/draft tensors, and a live --spec-type draft-mtp load attempt refuses to start on this exact binary ("model doesn't contain MTP layers") — plain decode throughout. Output sanity gate passed before recording any throughput (chat-endpoint coherence check, "the quick brown fox" echoed correctly).