Evaluate Qwen model fit for transformers.js runtime

[ScottMorris]

Goal

Determine which Qwen models are realistic for transformers.js in smdu.

Questions to answer

• Which Qwen checkpoints are currently compatible with our intended runtime path?
• Which quantisation levels are required for practical local use?
• How does quality degrade across smaller Qwen options for smdu tasks?

Focus areas

• Qwen2.5 Instruct family for summarisation and NL query handling
• Qwen2.5 Coder family for command/action suggestion semantics
• tokenizer/runtime compatibility constraints in Node.js context

Acceptance Criteria

• publish a compatibility table (model, quantisation, expected RAM, task fit)
• identify one primary and one fallback Qwen path for MVP
• list explicit reasons for any rejected Qwen options