CodeGraph CLI MCP Server

## Development Context - Feature: Adaptive compression optimization (zstd, lz4), real-time response compression, and dictionary-based compression for repeated data - Technical Stack: Rust (workspace crates: axum, tower-http, sysinfo, zstd, lz4_flex), async (tokio), tracing - Constraints: - Keep interfaces stable in core crates - Avoid blocking in async paths - HTTP negotiation must respect `Accept-Encoding` - Balance compression ratio and CPU usage under load - Success Criteria: - LZ4 and Zstd supported end-to-end for vector/cache compression - Adaptive selection favors CPU efficiency vs. ratio based on load and payload size - API responses compressed (gzip/br/zstd) when clients advertise support - Optional Zstd dictionary can be trained and used for repeated data patterns ## Design Overview - In-memory/vector compression: - Extend `codegraph-cache` MemoryManager with lossless LZ4 and Zstd compression for `Vec<f32>` vectors - Add adaptive selection heuristic using `sysinfo` CPU load and payload size - Provide optional Zstd dictionary training from samples - HTTP response compression: - Add `tower_http::compression::CompressionLayer` to API router to negotiate gzip/deflate/brotli/zstd - Keep middleware order: CORS → Compression → headers → backpressure → tracing ## Trade-off Strategy - Small payloads (<4 KiB): use LZ4 to minimize CPU overhead - High CPU (>75%): prefer LZ4 (faster, lower ratio) - Large payloads (>64 KiB) + moderate CPU: use Zstd level 6 for better ratios - Default moderate payloads: use Zstd level 3 ## Acceptance Tests (conceptual) - Compression API: round-trip (compress → decompress) equality for LZ4/Zstd - Adaptive path: selection changes with synthetic CPU load and payload size thresholds - HTTP responses: Content-Encoding negotiated (gzip/br/zstd) and payload size reduction observed - Zstd dictionary: training succeeds and improves compression ratio on repeated samples ## Notes - LZ4 dictionaries are not wired; Zstd dictionary support added where it matters most - HTTP dictionary-based compression is not standardized; dictionaries are applied for internal data paths