Add benchmark route profile contract

[huangruiteng]

Summary

• Add benchmark_route_profile_v0 under loopx.benchmark_core for public-safe route/model/reasoning and execution-surface metadata.
• Encode no-local-fallback permissions for official remote benchmark routes, private auth/tunnel handle references, compact artifact policy, pacing/timeouts, failure-attribution vocabulary, and score-claim boundaries.
• Add a focused smoke for an xhigh Codex CLI goal route without changing runner, scoring, leaderboard, or submission behavior.

Validation

• python3 examples/benchmark-route-profile-contract-smoke.py passed.
• python3 -m py_compile loopx/benchmark_core/route_profile.py examples/benchmark-route-profile-contract-smoke.py passed.
• git diff --cached --check passed before commit.
• loopx check --scan-path loopx/benchmark_core/__init__.py --scan-path loopx/benchmark_core/route_profile.py --scan-path examples/benchmark-route-profile-contract-smoke.py passed with two private-state heuristic warnings caused by intentional private/auth boundary vocabulary in the contract and smoke.
• loopx canary premerge --from-git-diff passed direct checks, Python compile, and 6 selected catalog canaries, but did not pass the merge gate because examples/control_plane/check-public-boundary-smoke.py timed out at the canary 120s limit and benchmark-sensitive surfaces require a manual hold.
• python3 examples/control_plane/check-public-boundary-smoke.py passed when run directly after the canary timeout.

Merge Gate

• No benchmark jobs launched.
• No runner/scoring/leaderboard/submission behavior changed.
• self_merge_allowed=false from premerge canary because this is benchmark-sensitive and needs maintainer review.

[huangruiteng]

Findings:

• [P1] reverse_tunnel_reference_label can still publish a local tunnel endpoint. If a caller passes a raw reverse-tunnel URL such as http://127.0.0.1:18180, build_benchmark_route_profile() sanitizes it to 127.0.0.1:18180 and validate_benchmark_route_profile() accepts it as public-safe. That preserves local endpoint/port topology in the public route profile even though the contract says transport handles are private references, not tunnel values. Please reject loopback/private host:port patterns for both transport handle labels, or coerce unsafe values to a generic handle label and add a smoke assertion for the raw URL case.

Product/architecture judgment:

The PR is useful and scoped correctly as a contract-only route profile: it does not launch runs, alter scoring, or change benchmark runner behavior. The permission policy and countability/score-claim boundaries are directionally right. The main gap is the public/private boundary for transport handle labels above.

Validation run locally on the PR worktree:

• python3 examples/benchmark-route-profile-contract-smoke.py -> pass
• python3 -m py_compile loopx/benchmark_core/route_profile.py loopx/benchmark_core/__init__.py examples/benchmark-route-profile-contract-smoke.py -> pass
• git diff --check origin/main...HEAD -> pass

Merge decision: hold until the transport-handle boundary is fixed and reviewer approval is available.

[huangruiteng]

Validation/update after owner-authorized closeout:

• Fixed the transport-handle blocker I raised earlier: raw endpoint/URL values such as http://127.0.0.1:18180 now collapse to a generic private handle label in build_benchmark_route_profile(), and validator rejects manually supplied endpoint labels for both auth and reverse-tunnel handles.
• Added smoke coverage for builder fallback and validator rejection of loopback/URL transport values.

Validation run on the PR worktree:

• python3 examples/benchmark-route-profile-contract-smoke.py -> pass
• python3 -m py_compile loopx/benchmark_core/route_profile.py loopx/benchmark_core/__init__.py examples/benchmark-route-profile-contract-smoke.py -> pass
• git diff --check origin/main...HEAD -> pass
• loopx check --scan-path loopx/benchmark_core/route_profile.py --scan-path loopx/benchmark_core/__init__.py --scan-path examples/benchmark-route-profile-contract-smoke.py -> pass, public boundary scan clean
• python3 examples/control_plane/check-public-boundary-smoke.py -> pass
• loopx canary premerge --from-git-diff --timeout-seconds 300 -> all executed checks passed: direct diff checks, py_compile, 6 selected catalog canaries, and public-boundary smoke. It still reports manual_review_required only because benchmark-sensitive surfaces default to an explicit maintainer hold.

Merge decision: owner explicitly authorized self-merge after runnable validation in the control thread; I will treat the benchmark-sensitive manual hold as satisfied by that authorization and proceed only via admin-bypass merge, with no runner/scoring/leaderboard behavior changed.