Fix model selection and model overrides

[vboctor]

Problem

The Telegram /model command and the deployed CLI model selector send a model override in the request, but the Agent Worker replaces it with the orchestrator agent model. Users are told that they can switch models, but the selected model is not actually used in Agent Worker mode.

Options

• Honor a validated model override from the caller, with an allowlist or safe validation policy.
• Or remove/disable the model-selection commands in modes where overrides are not supported, and document the configured model clearly.

Acceptance criteria

• A user can verify which model is active.
• A valid model override is honored consistently from Telegram and the deployed CLI, or the selector is no longer exposed.
• Invalid or unauthorized model IDs are rejected safely.
• Direct OpenRouter mode and Agent Worker mode have clear, consistent model-selection behavior.
• Regression tests cover the request-to-Agent-to-LLM model propagation path.