clm-0101

HU-004 reviewed the raw llama-server long-prompt guard for the genuine Qwen3.6-35B-A3B-MTP UD-Q4_K_M artifact on aihydra (runtime commit 7077abb, ROCm0/gfx1151, f16/f16 KV, context 204800). The guard passed ALL 20 cells across the plain (cfg-0141) and native-MTP n2 (cfg-0142) arms, at the sg (target 8K: 7881/7879 prompt tokens) and lp16k (15869/15867), lp175k (17353/17351), lp20k (19849/19847) and lp40k (39786/39784) brackets, each with a natural-language and a native tools fixture. No silent-empty-success, no empty-tool-arguments, no explicit context refusal, no HTTP error, no timeout, and no server crash was observed. ggml-org/llama.cpp#27442 was NOT reproducible on this ROCm0/gfx1151 f16-KV fingerprint.

Scope boundary: guard evidence only, from /home/aihydra/bench-results/hu004-qwen36-mtp-n2-longprompt-guard-r1 (kanban task t_6b1eba47). Passing this guard clears the silent-empty- success and empty-tool-argument terminal classes on THIS exact ROCm0/gfx1151 f16-KV n2 fingerprint up to the 40K prompt bracket. It does not admit a capability, tau2, throughput, or production result; it does not qualify any other backend, build, KV policy, or context; it does not establish n2 d262144; and it claims nothing about n_max>=4. One behavior recorded honestly in run-0552: the n2 lp20k-tool cell returned 79 repeated identical get_booking tool calls (finish_reason length, 57.1 wall s vs ~8 s plain) carrying the correct non-empty required booking_id, one call with a truncated argument string. No required tool argument was empty and no empty assistant turn occurred, so no guard terminal class fired and the cell remains guard-pass. Energy windows are preserved in run-meta timestamps but not joined — this hbreviewer profile lacks Home Assistant credentials.

Note — the record's own working

Cited by — computed at build time, never stored