Terminal-Bench 2.1 (Harness Leaderboard, avg@3)

The Harness Leaderboard's run of Terminal-Bench 2.1 (commit d49e28f1) on 87 of its 89 tasks, comparing coding-agent harnesses on fixed open models in Harbor; qemu-startup and qemu-alpine-ssh are left out because they need KVM, which Hugging Face's sandboxes lack. Each task scores the mean of three attempts and the rate is the mean over tasks. Infrastructure failures are retried and excluded, and trials that looked up a task's answer key count as failures. Kept as a component beside the 89-task average task accuracy.