[Design]: Finalize supported LLM model list for CugaAgent

[sami-marreed]

What you want and why

We need to formally define and document the set of LLM models that CugaAgent officially supports. Today, model selection is ad-hoc — contributors and users independently try models without a clear canonical list. This creates ambiguity around what is tested, what is recommended, and what compatibility guarantees exist.

A curated, versioned model support matrix will:

• Give users clear guidance on which models are production-ready with CugaAgent
• Help maintainers scope testing and CI coverage
• Allow us to set tier expectations (e.g. fully supported, community-supported, experimental)

How it could work

Propose a tiered model support matrix covering at minimum:

OpenAI-compatible / OSS

• GPT-4o, GPT-4o-mini
• nvidia/llama-3.1-nemotron-70b-instruct and similar Nemotron variants
• Meta Llama 3.x family (e.g. 8B, 70B, 120B via OpenAI-compatible endpoints)

Google

• Gemini 1.5 Pro / Flash

Anthropic

• Claude 3.5 Sonnet / Haiku

IBM

• watsonx.ai foundation models (relevant for wxo integration)

The design artifact should include:

1. A model ID → provider → tier mapping (Markdown table or YAML in docs/)
2. Criteria for each tier (e.g. fully tested in CI, community-tested only, experimental)
3. Notes on any known limitations per model (tool-calling support, context window, structured output)
4. A process for requesting/adding new models

Links or extra context

• This is a prerequisite for scoping evaluation benchmarks (component: evaluation)
• Relevant to wxo integration (component: wxo) which has its own model constraints
• Should align with the LLM node in DynamicAgentGraph and any model config in the SDK