Update agent CLI pins and add isolated OpenRouter evaluation routing

[hanhainebula]

Summary

• Update the Agent CLI pins across task images and the launcher: Codex 0.157.1, Claude Code 2.1.283, and Pi 0.87.1.
• Add opt-in --openrouter routing for Codex and Claude Code. It uses a dedicated AGENT_OPENROUTER_API_KEY, full OpenRouter model IDs, and the appropriate API endpoint for each CLI.
• Keep the Agent key separate from task-submission and verifier credentials. Verifier versions and configuration are unchanged.
• Add setup and evaluation examples to both READMEs and the evaluation guide.

Example

bash scripts/run_task.sh --task task-1-1 --agent codex --openrouter \
  --model openai/gpt-6-astra --dry-run

bash scripts/run_task.sh --task task-1-1 --agent claude-code --openrouter \
  --model anthropic/claude-opus-5.5 --dry-run

Set AGENT_OPENROUTER_API_KEY in .env before running an evaluation.

Validation

• 50 relevant unit tests passed.
• git diff --check origin/main...HEAD passed.
• A live OpenRouter evaluation and Docker image builds were not run; a real Agent key has not been configured yet.