fast 27b, tangled 4-knob night batch — 7/9, dense model tolerates the context strip

Night batch 2026-07-08, SubagentsPi with FOUR knobs flipped together vs the 07-07 runs: loop_guard=on strip_thinking=on keep_tool_results=10 subagent_nudges=on (plus the established bash_timeout=90). NOT a clean A/B — see the untangling item in AGENT_TODOS.md.

Result 7/9 (mean 0.78) vs 6/9 on the 07-06 SubagentsPi baseline (fast__qwen3.6-27b__20260706-231402): cancel-async-tasks and openssl-selfsigned-cert flipped FAIL→PASS, query-optimize PASS→FAIL, filter-js-from-html still FAIL. Within noise, but clearly NOT the 35b collapse (see the sibling 35b run's NOTES) — the dense model tolerates thinking-strip + tool-result stubbing.

Mechanism notes: zero loop-guard blocks in any root session; 4 trials made zero subagent calls despite the turn-6 fallback nudge (27b ignores the nudge and self-serves — delegation census barely moved). Nudge messages render correctly as user messages in pi.txt.