QwenPi (qwen-native tools) — 35b fast, 5/9; tools work flawlessly, no score change

First benchmark run of the QwenPi harness (read_file / write_file / edit / run_shell_command, --no-builtin-tools, delegating to pi's builtins). Knobs otherwise current (loop_guard on, strip off, subagent_timeout 600).

Result 5/9 — identical pass-SET to the closest plain-MinimalPi baseline (fast__qwen3.6-35b-a3b__20260706-113104, also 5/9: fix-git, nginx, openssl, sanitize-git-repo, regex-log PASS; cancel-async, filter-js, query-optimize, sparql FAIL). The tool rename moved nothing on score.

Tool mechanism verdict — clean: 0 schema/unknown-tool errors across all 9 trials; the model used the qwen-named tools fluently and every wrapper delegated to the real builtin (fix-git 18s, nginx/openssl 16s agent). The suspiciously fast fails are NOT tool regressions: cancel-async (17s) made write_file+run_shell_command+read_file and a real solution+summary — the usual hidden-edge-case near-miss; filter-js (26s) likewise (the known parse-reserialize class). regex-log PASSED with heavy tool use and NO format leak.

Takeaway: on this task set the failures are task-logic / hidden-criterion, not tool-surface friction — so matching the surface (which the literature says helps 3-8B models) had nothing to remove here. See the ledger verdict.