Use modelId in assemble() for model-aware context packing

[lightcone0]

Context

OpenClaw v2026.3.22 now passes modelId to context-engine assemble(). We accept it for compatibility (#6) but don't use it yet.

Opportunity

Use modelId to adapt context assembly per model:

• Token budget scaling — pack more memories for models with larger context windows (e.g., 200K vs 8K)
• Format adaptation — some models handle structured markdown better than others
• Embedding-aware retrieval — adjust minScore thresholds based on known model characteristics

Scope

• packages/apps/openclaw/src/context-engine.ts — assemble() method
• Potentially packages/core/src/components/context.ts — pack() could accept model hints

Priority

Low — current behavior works fine. This is an optimization for when we have data on how different models respond to different context formats.