feat: modernize default model policy

[ASRagab]

Summary

• centralize proposer and generated-evaluator defaults with CLI > environment > project-default precedence
• use provider sampling defaults unless temperature is explicitly supplied, and ensure generated judge/composite evaluators compile and run offline without credentials
• update CLI help, operational scripts, packaged skills, documentation, and stale-model drift guards for current OpenAI, Anthropic, and Gemini models

Test evidence

• uv run pytest -q — 375 passed, 8 skipped
• uv run python scripts/check.py --skip-smoke — pytest and score checks passed
• LiteLLM routing verified offline for openai/gpt-5.6-sol, openai/gpt-5.6-luna, anthropic/claude-sonnet-5, and gemini/gemini-3.6-flash

No paid provider calls or dependency upgrades were introduced.