fix(session): prevent direct commands from consuming model context

[xiami762]

Summary

• keep direct slash-command messages visible while excluding them from LLM context
• prevent the builtin command pseudo-model from replacing the active model context window
• add regression coverage for persistence, prompt conversion, token estimates, and context-model resolution

Root cause

Direct command output was persisted as regular conversation text and the context meter resolved builtin/command as a 4,096-token model. Large outputs such as /tools therefore filled the estimated context and were replayed to the next model call.

Validation

• 252 related tests passed
• Ruff checks passed
• git diff --check passed