feat: add long-context-prefer routing strategy

[cursor]

Summary

• Add long-context-prefer routing strategy for GPT-5.5 / Claude Sonnet 4.6 / Gemini 3.x / Kimi K2
• Prefer models whose context window meets metadata.min_context_tokens (or default 100000 when metadata.long_context is truthy)
• Rank by context window, then quality, then cost

Motivation

Long-prompt / RAG workloads need capability-aware preference for large-context models without hard-failing when no candidate meets the threshold.

Changes

• Feature / fix
• Tests added or updated
• Documentation updated
• CHANGELOG updated

Test Plan

ruff check tests/test_long_context_prefer_strategy.py src/router/strategies.py src/router/schemas.py
pytest tests/test_long_context_prefer_strategy.py -q

11 passed. README badge → 773.

Checklist

• Self-reviewed the diff
• No secrets committed
• CI passes