Guard-validation attempt #1 — regex-log PASS in 2m14s, but the trial peaked at 28k so the 40k trigger never fired; guard still unvalidated

Purpose: force the COMPACT_AT_TOKENS overflow guard to actually fire by setting the threshold (40000) below regex-log's usual peak context (~39–51k in prior runs).

Result: PASS, 2m14s, 17 turns — but this draw was unusually efficient and peaked at only 28,046 context, under the 40k trigger. pi.txt shows context guard active (compact_at=40000) and nothing else: no forcing compaction line, no compaction event. Third run in a row where the guard is armed but never exercised.

Per-turn context trajectory: 1.8k → 28k, crossing 15k at turn 9 of 17. So the next validation attempt uses COMPACT_AT_TOKENS=15000 — every observed regex-log draw (28k / 39k / 51k peaks) crosses that mid-run, leaving several turns after the forced compaction to prove the task still completes:

make smoke MODELS=qwen3.6-35b-a3b SMOKE_TASKS=regex-log COMPACT_AT_TOKENS=15000

Side note: regex-log's peak context across draws is 28k/39k/51k — ~2× spread on the same task/model, another datapoint for the K=1 variance theme (H5).