[ygyang11]

Summary

• Add llm.sub_model configuration and a dual-LLM lifecycle with identity aliasing when no auxiliary model is configured.
• Route compression and goal evaluation through the auxiliary LLM while preserving model-specific usage attribution and main-model context accounting.
• Let sub-agents explicitly select main or sub, propagate both models through internal agents and teams, and document the selection policy in model-facing prompts.
• Keep /model, /provider, and /effort sub consistent with runtime replacement, rollback, snapshots, compressors, and background tasks.
• Remove memory.compression.summary_model and the obsolete model_override factory path.

Validation

• pytest -q: 2755 passed, 3 skipped
• ruff check src tests
• mypy src
• git diff --check
• Real REPL smoke coverage for main/sub routing, compressor and goal-evaluator usage, background switching, provider/model commands, and model-selection scenarios.