feat: add automatic model routing

[SDU-W-ZhangLab]

Summary

• add a virtual Auto entry to the chat model picker
• route each turn to a user-tagged fast, balanced, or reasoning model while preserving the conversation-level Auto selection and existing context
• add deterministic tier fallback, provider/model ordering, settings UI, persistence, and Chinese/English copy
• keep manual model selection and gateway overrides unchanged

Routing policy

• attachments, prompts >= 1,200 characters, or 2+ complexity signals -> reasoning
• prompts >= 400 characters or one complexity signal -> balanced
• otherwise -> fast
• fallback order:
  • fast -> balanced -> reasoning
  • balanced -> reasoning -> fast
  • reasoning -> balanced -> fast

This is intentionally a low-cost Auto router, not a parallel Fusion panel/judge. A Fusion-style ensemble should be a separate opt-in feature because it multiplies latency/cost and can duplicate tool side effects.

Verification

• node --test test/chat/model-selection.test.mjs (16/16 passed)
• npm run build (TypeScript + Vite production build passed)
• Biome check on changed source files (no errors; existing warnings remain)

The full frontend suite could not finish reliably on this machine because Node v22.18.0 is below the dependency-declared v22.19.0 minimum; unrelated SSH/MCP/subagent workers retained open handles.