122b fast run part 3/3 — final batch: sanitize-git-repo PASS, sparql + filter-js FAIL; combined fast set 4/9

Final resume batch of the memory-aborted fast__…141049 (MinimalPi, scaled caps, llama-swap restarted before the batch).

• sanitize-git-repo PASS 6m56s agent, 584k input tokens — heavy multi-turn work but it landed.
• sparql-university FAIL 5m56s — the known one-assertion-away shape.
• filter-js-from-html FAIL 5m37s (clean rerun of part 1's thrash-induced ERR).

This batch rode the ratchet to 29 MiB MemAvailable / 13 GiB swap near the end but finished without OOM (bounded by the last trial's 900s cap; swap never dropped below ~6.7 GiB free).

Combined 122b fast-set verdict (parts 1–3): 4/9 — PASS fix-git, nginx-request-logging, openssl-selfsigned-cert, sanitize-git-repo. That is the 35b MinimalPi baseline pass set (5/9) minus regex-log, which flaked (smoke PASS, fast FAIL). At ~4–7× the wall clock per task, the 122b shows no fast-set advantage over the 35b so far; the smaller model's failures here are hidden-criterion/task-logic, and more parameters didn't buy those back. No trial hit its cap, so the scores are cap-clean.