Bug: max_completion_tokens in llm_pool.defaults ignored when judge_panel is not configured #348

Root cause

• No judge_panel → manager.py:256 uses system_config.llm (legacy default LLMConfig), bypassing llm_pool entirely
• resolve_llm_config() — which merges defaults → model params — is never called
• Ragas falls back to InstructorModelArgs.max_tokens = 1024
  Works when: model-level parameters set + judge_panel configured (triggers resolve_llm_config())

Reproduce
Set llm_pool.defaults.parameters.max_completion_tokens: 500, no judge_panel
Run eval → effective limit is 1024, not 500

Fix: When llm_pool is present but no judge_panel, resolve primary model via resolve_llm_config() instead of falling back to system_config.llm.