feat(bench): add terminal-bench supervisor arm

[drewstone]

Summary

• add a Terminal-Bench supervisor arm that mounts the coordination MCP into opencode
• run spawned workers inside the same graded container through the tb-container executor
• add repo-relative smoke/run scripts and ignore local A/B run output
• fail loudly if supervisor orchestration cannot start, instead of silently falling back to the raw arm

Verification

• python3 -m py_compile bench/tb_agents/opencode_supervisor_agent.py
• bash -n bench/run-ab-arm.sh bench/smoke-sidecar.sh
• bash bench/smoke-sidecar.sh (Docker smoke: tools/list, spawn_agent, await_event, same-container worker receipt)
• env -u TB_TARGET_CONTAINER pnpm exec tsx bench/src/tb-supervisor-sidecar.mts exits 2
• pnpm run typecheck
• pnpm run build
• pnpm test: 127 files passed, 1,254 tests passed, 2 skipped
• pnpm run verify:package
• git merge-tree --write-tree origin/main HEAD

[tangletools]

✅ Auto-approved drewstone PR — 3cac8968

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: drewstone_author · 2026-07-07T23:53:23Z}