MCP claude.ai-proxy connections cache legacy-peer discovery results

What

• A new async check probes whether a claudeai-proxy type MCP (Model Context Protocol, the way Claude Code talks to external tool servers) connection needs a legacy-discovery fallback. It skips stateless connections and ones already marked discoverSupport: 'supported', and caches ids already found to have 'burned' listing/discover priors.
• The connection logic gains a remembered_legacy negotiation-mode outcome and a matching log line: once an earlier probe on a connector has found a legacy (pre-modern) peer, later connects skip the server/discover round trip entirely instead of re-checking the cached /v1/mcp_servers projection every time.

Why

This avoids repeating the same discovery round trip on every connection to an older MCP server reached through claude.ai-proxy, reducing unnecessary network calls.