Edit (round 4, snapshot10e): 11/11 pass.
Changes data verified against regenerated seeds for every scenario (baselines = seed bytes, new files absent, versions match sha, newest = disk). b2 5 approvals sane; b3 denial wrote nothing, 'You declined this Bash call. Nothing was run.' accurate.
j3 $0.4298 hand-checked correct: 33 of 58 requests above 100K on the long-prompt card ($0.3796), cache-read ratio 0.976, cache writes ~= final prompt (no resend/miss). Cost from the model's long single turn, not G-Code.
Notes: denied/refused calls send tool-result with no tool-use frame (desktop adds a failed card without input, useChatSession.ts:311-323); b6/b7/j4 scenario medium effort forced to high by runner.
