ClawsBench open-source readiness: BenchFlow fixes for review

[kywch]

ClawsBench open-sourcing validation identified six scoped BenchFlow fixes:

• fix(litellm): provision Vertex sandbox runtime #985 — Vertex sandbox runtime
• Fix Gemini ACP model IDs with Google provider prefixes #1030 — Gemini/Gemma model ID normalization
• fix(judge): parse native edit targets #1069 — native edit-target judge false positive
• Add Z.ai Coding Plan routing #1074 — Z.ai Coding Plan routing
• Cap OpenClaw Claude output tokens #1075 — OpenClaw Claude output-token cap
• Honor Codex proxy-owned model selection #1076 — Codex proxy model selection

Combined validation completed all 10 selected arXiv conditions with zero sandbox/trial errors.

All six PRs have green CI and are mergeable. #1074 remains stacked on #1030 and should be retargeted to main after #1030 merges.

Please review these PRs to help unblock ClawsBench open-sourcing.

[kywch]

Merged #1075 to get the agent parity check CI work for new PRs.

[kywch]

This is resolved when these PRs get merged:

• refactor(agents): extract OpenClaw from BenchFlow core #1093
• feat(parity): support generic ACP launch capture agents#71
• feat(openclaw): move ACP shim ownership to agents agents#72
• fix(openclaw): stream ACP events during prompts agents#73