122b fast run part 1/3 — ABORTED 4/9 at a trial boundary: llama-server RAM ratchet nearly OOMed the box

First fast-set run of qwen3.5-122b-a10b (MinimalPi, ~5×-scaled cap map fix-git=600…default=1800 because the stock caps are calibrated on 27b/35b speed and would cut 122b passes). llama-swap was restarted BEFORE this run (59 GiB available at start).

Deliberately aborted (SIGINT) at 4/9 after the documented 122b llama-server memory ratchet (see the model's entry in the llama-swap config.yaml, kernel OOM kills earlier the same day) burned 59 GiB → 92 MiB available + 11 GiB swap in ~25 minutes of benchmark load. Extrapolation said swap would exhaust before trial 9, i.e. a near-certain kernel OOM kill of llama-server mid-run. Aborted at the filter-js trial boundary; the remaining 5 tasks were rerun as two fresh-restart batches (fast__…143738, fast__…144656) whose results merge with this run on the dashboard.

Scored trials (all finished well under their caps, no cut@ markers):

• fix-git PASS 1m53s agent — matches smoke behavior.
• regex-log FAIL 4m26s — NB PASSED in the smoke run 20 min earlier (5m44s). K=1 variance on this model, not a regression.
• query-optimize FAIL 4m20s — the usual perf-bar miss.
• filter-js-from-html ERR CancelledError at 6m21s / 231k input tokens — died while the box was thrashing (MemAvailable double-digit MiB); treat as infrastructure, not model. Clean rerun in fast__…144656 scored FAIL.

Combined fast-set verdict across the 3 parts: 4/9 (fix-git, nginx, openssl, sanitize-git-repo) — the 35b MinimalPi baseline pass set minus regex-log (which flaked here but passed in smoke). See COMMENTS.md and the AGENT_TODOS "qwen3.5-122b-a10b operational" item.