feat: redesign prompt compaction with 4-tier memory management

[chinkan]

This branch redesigns prompt compaction with a 4-tier approach, RAG-aware context retrieval, and multi-provider support.

Key changes

• 4-tier compaction: sync (truncation), reconstruct (few-shot struct), auto-compact (summarization), reactive (expensive) — each with model-specific context window awareness
• RAG-aware compaction: retrieve_context_for_compaction fetches semantically relevant messages before summarization
• vec0 metadata pre-filtering: is_summarized and role columns on the vec0 table for filtered similarity search with configurable RRF weights
• Context window cache: ContextWindowCache with backoff/warmup, model-specific window sizing per provider
• System prompt preservation: compaction-aware system message handling, EoT marker protocol
• Background cache warmup: async warmup at startup
• Configurable compaction: rrf_k, rrf_weight_fts, rrf_weight_vec, preserve_message_count, reactive_compact_pct

Files touched

config.example.toml |   1 +
src/agent.rs        | 214 +++-
src/agent_prompt.rs |  62 +-
src/config.rs       |  91 ++
src/main.rs         |  39 +-
src/memory/conversations.rs |  53 +-
src/memory/mod.rs   |  56 +-
src/memory/rag.rs   | 114 ++
src/provider.rs     | 294 +++-
9 files changed, 892 insertions(+), 118 deletions(-)
1 new design doc added