feat(coding_agent_rl): select claude_code/codex harness+adapter pair via SWE_AGENT

[jingshenghang]

What

Make the SWE rollout in examples/coding_agent_rl agent-agnostic. The generate() orchestrator no longer hard-codes the Claude Code harness + Anthropic adapter; instead the (harness, adapter) pair is selected at import time by the SWE_AGENT env var.

Changes

• generate.py:
  • Add an _AGENTS registry mapping the agent name to its (harness, adapter) pair:
    • claude_code -> (ClaudeCodeHarness, AnthropicAdapter)
    • codex -> (CodexHarness, OpenAIAdapter)
  • Resolve SWE_AGENT (default claude_code) into HARNESS_CLS / ADAPTER_CLS, with an explicit ValueError on an unknown value.
  • Use HARNESS_CLS / ADAPTER_CLS everywhere the concrete classes were previously referenced (CLI install, adapter service, rollout run).
  • Update the module docstring to describe the selector.
• run_qwen36_35b_a3b_swe_8nodes.sh:
  • Export SWE_AGENT (default claude_code) and propagate it into the Ray RUNTIME_ENV_JSON so workers see the same selection.

Behavior

• Default is unchanged (claude_code), so existing runs are not affected.
• Set SWE_AGENT=codex to switch the rollout to the Codex harness + OpenAI adapter.