History compression pipeline: async summarization of older turns

[theinterneti]

Part of #266

When session history > tier 5 budget: detect overflow, trigger async LLM summarization of oldest turns, store summary in Redis with TTL. Session start pre-computes if near threshold.

Acceptance Criteria

• Overflow detection: tier 5 history > 80% of budget triggers compression
• Async LLM summarization of oldest 20 turns → compact paragraph
• Summary stored in Redis: key session:{id}:history_summary, TTL=7200s
• Session start checks if near threshold and pre-computes summary
• Unit tests with mocked LLM; integration test with simulated 50-turn session