V3 MCP Optimization

Implement connection pooling to reuse MCP connections, replace linear tool search with a hash-based registry for O(1) lookup, and add transport batching with compression. These changes target sub-100ms p95 response times and under 400ms startup.

Create a ConnectionPool class that stores pooled connections with lastUsed timestamps and usage counts, pre-warms a minimum number of connections at startup, and evicts the least-recently-used connection when capacity is reached. Health checks run on an interval to remove unhealthy connections.

Use a hash-based Map index for O(1) exact-match lookups instead of linear search, backed by an LRU cache for frequently used tools. A fuzzy matcher handles typos and name variations when exact matches fail.

The least-connections strategy routes requests to the server with the fewest active connections and works well as a default. Response-time selection favors the fastest server, while weighted scoring combines load, latency, and tool-category affinity.

Slow startup typically comes from building tool indexes and opening connections lazily on first request. Pre-warm the connection pool and pre-build the tool index during server initialization to cut startup from ~1.8s toward under 400ms.

Track request latency percentiles (p50, p95, p99), error rate, connection pool hit rate, tool lookup time, and memory usage. Alert when error rate exceeds 5%, pool hit rate drops below 70%, or p95 latency exceeds 200ms.