fast2 35b, tangled 4-knob night batch — 2/6-ish, run cancelled inside fix-ocaml-gc's 62-min planner child

Night batch 2026-07-08, SubagentsPi, loop_guard=on strip_thinking=on keep_tool_results=10 subagent_nudges=on. Baseline: 07-07 run (fast2__qwen3.6-35b-a3b__20260707-224105) went 4/6. Run is marked incomplete — the user cancelled it at 08:06 inside the fix-ocaml-gc trial.

Per task: - fix-ocaml-gc — ERR CancelledError at 1h28m, no verifier. The loop guard's original target shape (verbatim two-command A/B loop in the root) did NOT recur, but the failure adapted: a planner subagent child blocked the root for 62 minutes (uncapped, shares the llama-swap slot), internally hit ≥48 LOOP DETECTED blocks and evaded the guard with small argument variations before dying exit 1 without implementing; the root then near-looped on slight-variation re-reads (16 turns saturating the 8k thinking budget) until cancelled. Both dead planner children were reported isError=False to the root. Full post-mortem: analysis/qwen3.6-35b-a3b/pi-subagents/fix-ocaml-gc.md (2026-07-08 section). Input tokens 3.35M vs 59.8M on 07-07 — the strip's token win is real; the wall-clock got worse anyway. - headless-terminal — ERR AgentTimeoutError but verifier reward 1.0: the session finished normally at ~9 min (final answer, task solved) and the pi process then hung ~21 min until the 30m cap killed it — likely a leftover pi-subagents child keeping node alive. One-off (every other trial tore down in 3–18s); ledger item filed. - reshard-c4-data — ERR at 17s: infra flake, the nvm curl install script failed during pi install. Just rerun. - financial-document-processor PASS→FAIL, git-multibranch FAIL→FAIL, mailman PASS.

Together with the sibling fast run (3/9), 35b collapsed under the tangled knob set while 27b held — context-strip amnesia is the prime suspect (re-read/premature-done shapes throughout). See the untangling + child-cap + isError items in AGENT_TODOS.md.