Optimize prompt caching by splitting system messages

[simonferquel]

Summary

Refactors session message handling to improve LLM prompt caching by categorizing system messages based on caching characteristics.

Changes

• Split system messages into three categories: invariant (cacheable globally), context-specific (cacheable per user/project), and session summaries
• Added strategic cache control markers at category boundaries
• New tests for cache control behavior

Benefits

• Better cache hit rates → lower API costs and faster responses
• Fully backward compatible

Technical

Replaced monolithic GetMessages() with three focused functions for different message categories, enabling optimal cache utilization at multiple granularities.

[simonferquel]

/review