Evaluate and integrate Qwen 3.5 model defaults

[kovtcharov]

Summary

Lemonade v9.4.1 adds Qwen 3.5 family model support on ROCm and Vulkan backends. GAIA should evaluate whether Qwen 3.5 variants outperform current Qwen 3 defaults and update accordingly.

Reference

• Lemonade v9.4.1 release
• Scope doc

Changes Required

File	Change	Effort
src/gaia/llm/lemonade_client.py	Update default model constants and AGENT_PROFILES if Qwen 3.5 is preferred	Low
src/gaia/llm/lemonade_manager.py	Update agent-to-model mapping if defaults change	Low
docs/	Update model references throughout documentation	Low

Process

1. Benchmark Qwen 3.5 vs Qwen 3 on GAIA agent tasks (chat, code, blender, jira)
2. Compare quality, speed, and memory usage
3. If Qwen 3.5 is better → update defaults (config-only change)
4. Update documentation to reflect new model names

Acceptance Criteria

• Benchmark results documented
• Default models updated if Qwen 3.5 is superior
• Documentation reflects current model defaults