122b fast run part 2/3 — resume batch after memory abort: nginx PASS, openssl PASS, cancel-async FAIL
Continuation of the aborted fast__qwen3.5-122b-a10b__20260712-141049 (see
its NOTES for the llama-server RAM-ratchet story). llama-swap restarted
immediately before this batch (fresh 59 GiB available); same MinimalPi +
scaled cap map.
- nginx-request-logging PASS 1m27s agent.
- openssl-selfsigned-cert PASS 2m53s — NB 2.4× its stock 120s fast cap;
the scaled map is what kept this pass alive.
- cancel-async-tasks FAIL 2m18s — the known hidden-assertion near-miss
shape, consistent with the 35b/27b history on this task.
Ratchet datapoint: this 8m37s batch alone burned 59 GiB → 299 MiB
MemAvailable, confirming the burn is proportional to inference activity and
that per-batch llama-swap restarts are load-bearing for 122b benchmark runs.