Add obol-agent provider inference smoke coverage and model preference command

[bussyjd]

Problem

We need explicit smoke coverage for stack-managed obol-agent inference across supported AI provider paths: local Ollama, Anthropic, and OpenAI. Current validation is focused on the default/local path and provider setup, but does not clearly exercise the default agent end-to-end for each provider.

We also need an obol model prefer command so operators can change the preferred model order used by stack-managed agent config generation, especially when multiple local Ollama models are available.

Scope

• Add/confirm smoke coverage for obol-agent chat completions through the stack-managed agent for local/Ollama, Anthropic, and OpenAI.
• Add obol model prefer to reorder configured models and resync the stack-managed agent.
• Include unit tests for preference ordering and CLI behavior.
• Keep this separate from the Hermes runtime refactor PR.

Acceptance Criteria

• Provider smoke tests are discoverable and documented in flows/tests.
• obol model prefer <model> persists preference order and affects the default model selected for the agent.
• Local model preference supports multiple Ollama models.
• Tests do not require cloud credentials unless the relevant provider key is explicitly present.
• Existing OpenClaw/manual runtime behavior is not regressed.

[bussyjd]

Closing — superseded by #399.

The original ask (model-preference command + obol-agent provider smoke coverage) has been re-scoped: per-agent ranking duplicates work that LiteLLM model alias groups can do once at the gateway layer. #399 captures the redirected design (smart routing via LiteLLM alias groups, with per-agent ranking removed). Closing #379 with this issue rather than landing the preference command as originally specified.