122b fast2 3/7 — mailman PASS (second first-ever), 22m49s: the cap clamp earned its keep

Third job of the 122b fast2 one-task-per-job sequence.

mailman PASS at 22m49s agent (1369s), 2.83M input tokens — the second task in this sequence the 35b never passed (its fail was 3m25s). NB the stock fast2 cap for mailman is 600s: the 122b-scaled 1800s cap is the only reason this pass exists — a stock-cap run would have cut it at 10 min and recorded a FAIL. Confirms the per-model cap-scaling ledger item.

Ratchet datapoint: ~23 min of solid inference in one job rode MemAvailable down and ~9 GiB into swap (watchdog untripped) — right at the edge of the ~33 min single-job budget estimated from the 2026-07-12 burn rate, which is what the 1800s clamp was sized against.