[codex] Fix init provider detection and runtime support

[rendrag-git]

Summary

Fixes #1, #2, #4, #7, #9, and #10. Partially addresses the scoring cleanup from #3.

• make oc-bench init fail closed when detection finds only detect-only providers instead of silently writing default vLLM config
• route explicit --no-detect --vllm-* config through vLLM provider shaping so GPT-OSS gets reasoning_effort=low
• add --probe-hosts plumbing for init detection and scan OpenClaw profiles inside Incus/Docker runtimes
• support incus:<instance> command execution paths for OpenClaw version, gateway start/stop, preflight, provider health, and backend runs while preserving Docker container behavior
• clarify top-level help for benchmark preflight vs provider-route provider-preflight
• clean up scoring intent comments and the duplicate preserved-file failure branch

Review follow-up

• moved runtime parsing helpers into openclaw_bench.container to avoid duplicate backend/preflight logic
• made provider-health runtime parsing unconditional so legacy plain container names still resolve as Docker
• added runtime-aware gateway stop and profile-scoped Incus gateway logs
• confirmed _parse_csv already exists in openclaw_bench/cli.py and is not a new runtime crash
• expanded --probe-hosts help text to mention bridge/LAN addresses

Validation

• python3 -m pytest tests/test_providers_detect.py tests/test_providers_init_integration.py tests/test_quickstart.py tests/test_preflight.py tests/test_backend.py tests/test_scoring.py tests/test_container.py tests/test_providers_preflight.py -q -> 153 passed
• Full python3 -m pytest -q was also attempted before PR creation and hit an unrelated local HTTP fixture failure: tests/test_serve.py::ServeTests::test_api_base_context_window_probe_failure_is_classified raised ConnectionResetError: [Errno 54] Connection reset by peer; rerunning that single test reproduced the same failure.

Notes

• .github/ is intentionally not included because it was already untracked before this work.
• The task-suite expansion ideas from Behavior coverage thin spots in tier suites #5 are intentionally deferred until live calibration data exists; changing the benchmark surface now would be premature.