fast2 35b STRIP-OFF arm — KILLED BY USER after 1/6 trials (long wall times); the one scored trial PASSED

Second half of the strip-off isolation batch (SubagentsPi, loop_guard=on subagent_nudges=on STRIP_THINKING=off KEEP_TOOL_RESULTS=0). Stopped by the user at ~13:53 during trial 2 (headless-terminal, ~5 min in, no result — its partial dir has no result.json; leaked env container was removed and the mid-flight transcript strip was run manually). The fast-set half of the batch had already completed and settled the A/B — see runs/fast__qwen3.6-35b-a3b__20260708-115953/NOTES.md.

reshard-c4-data: PASS, reward 1.0, but 32m23s session vs 5m43s agent on the 07-07 baseline pass — this trial is the wall-clock exhibit for the SUBAGENT_NUDGES ceremony cost (new ledger item): 07-07 did scout+planner then self-served (335s session, 42 turns); this one ran the full nudged workflow — scout, planner ×2 + resume, worker, reviewer ×3 (the model re-called reviewer twice on its own after a clean first review) — 70 turns, 8 child calls ≈ 500s of child wall time, each child a fresh prefill on the shared llama-swap slot, plus ~450s of bash dead-ends (a fully-burned self-set 300s timeout + a FILE-NOT-FOUND detour) on top of the legitimate ~10k-file compress/verify work. Nudge dedup itself worked (exactly 3 nudges, one per stage). Same verdict either way — the ceremony bought nothing here but 5.8× the wall clock.