Gate SkillsBench launch on runner readiness

[huangruiteng]

Summary

• let the existing runner-profile CLI probe task-free SSH from the launcher's current environment
• require a ready connectivity receipt before any non-dry-run remote preflight or supervisor start
• preserve dry-run as a no-network plan while reporting the pending readiness gate

Boundary

• emits only the existing compact public connectivity receipt on failure
• reads no task material and records no SSH output, destination, or profile values
• validation launches no benchmark job
• does not change scoring, task semantics, submissions, uploads, or permission boundaries

Validation

• pytest -q tests/test_skillsbench_turn_launcher.py tests/test_skillsbench_runner_profile.py tests/test_skillsbench_reverse_tunnel_supervisor.py (36 passed)
• python examples/skillsbench-runner-profile-smoke.py
• python -m mypy
• ruff check
• bash -n scripts/skillsbench-launch-goal-xhigh.sh
• loopx check on all changed public surfaces
• loopx canary premerge --from-git-diff: all direct checks, 6 catalog canaries, 8 risk-profile smokes, and the public boundary check passed; the generic benchmark_sensitive hold is covered by the standing owner authorization for narrow benchmark helper/runtime PR self-merge

[huangruiteng]

No actionable findings.

Open questions/assumptions:

• Exit code 3 remains the launcher's compact preflight-blocked class; the JSON schema identifies runner connectivity versus later sandbox preflight failure.
• --current-environment accepts only the SSH destination and option keys needed by the existing task-free probe.

Product/architecture judgment:

• Motivation: a typed readiness receipt is insufficient if the real launcher can bypass it and enter remote preflight or supervisor startup while task-free SSH is rejected.
• Solved: the existing runner-profile CLI owns current-environment probing, while the launcher only enforces the receipt and preserves dry-run planning.
• Operator impact: failed remote acceptance now returns one public-safe blocker before supervisor state or benchmark work exists; unrelated LoopX todos can continue.
• Main risk: one additional task-free SSH probe on non-dry-run launch. It replaces ambiguous downstream failure with an explicit gate and reuses the launcher's existing destination/options.
• Design: the shell remains thin, and readiness semantics stay in the existing adapter rather than being duplicated in the launcher or heartbeat.

Validation:

• 36 focused tests and the runner-profile smoke passed
• mypy, Ruff, shell syntax, diff hygiene, and public/private boundary checks passed
• premerge canary passed all direct checks, 6 catalog canaries, 8 risk-profile smokes, and the public boundary check
• the remaining generic benchmark_sensitive hold is acknowledged under the standing owner authorization for narrow benchmark helper/runtime PR self-merge

Merge decision: self-merge after GitHub Python Tests pass.