cfg-0096

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. Stock 3653e6d6d Vulkan binary (~/src/llama.cpp-vk3653e6d/build/bin; house convention: reports build_commit 3653e6d6d, prefix-matches 3653e6d in protocol.json known_builds, same as cfg-0083/deepseek-v4-flash and laguna-s-21's cfg-0091). Depths 0/32768 ran N=3 fresh-process reps; 131072 ran N=1 (house convention). PARTIAL-SERIES caveat does NOT apply here, unlike laguna-s-21's cfg-0091: this candidate's Vulkan arm completed ALL THREE depths cleanly, including d131072 (238.5 pp / 32.4 tg t/s) -- no device loss on either allowed attempt, and only needed the first. Scatter across the 3 reps was clean at every depth (max CV 0.2%, d32768 prefill) -- tighter than the ROCm arm's own scatter at the same depths. No MTP/speculation: same from-scratch GGUF parse as cfg-0094, zero nextn/mtp/eagle/medusa/draft tensors among all 733. Output sanity gate passed via the chat endpoint before recording any throughput.