Fix llama aliases and roleplayed tool calls

[0xCheetah1]

Summary

• Add llama shortcuts for the current free Maverick model and reject unknown bare /model aliases with a local error
• Broaden roleplayed tool-call detection so free models can recover bare JSON, OpenAI envelopes, Hermes envelopes, and tool_call XML wrappers
• Add resolver and detector test coverage, including the resolver test in npm test

Tests

• npx -y -p node@22 npm test