fix: resolve memory leaks causing ~4GB memory growth

[Minitour]

Summary

Analysis of a 3.7 GB memory dump revealed several critical memory leaks in the MCP server infrastructure that caused unbounded heap growth:

• Double process spawn: StdioClientTransport already spawns its own child process internally, but MCPProxy.createStdioClient() was also calling SubprocessManager.getOrCreateSubprocess() — resulting in two processes per stdio MCP server. The first was orphaned and never cleaned up on shutdown.

• Leaked temporary CapaMCPServer instances: handleProjectConfigure (plugin enrichment + tool validation) and handleGetServerTools created temporary CapaMCPServer instances with their own MCPProxy and MCP client connections that were never closed. Each project reconfigure leaked more connections.

• close() not closing the proxy: CapaMCPServer.close() only closed the MCP Server but never called mcpProxy.closeAll(), leaking all client connections and their transport buffers.

Changes

• Remove SubprocessManager dependency from MCPProxy — let StdioClientTransport manage its own subprocess lifecycle
• Add getOrCreateMCPServer() helper to CapaServer to centralize CapaMCPServer instance management and ensure reuse via the mcpServers Map
• Fix CapaMCPServer.close() to call mcpProxy.closeAll() before closing the server
• Add onclose handlers on MCP clients to auto-remove dead entries from the cache

Test plan

• Start capa server, configure a project with stdio MCP servers, verify tools still work
• Monitor process memory over time — should remain stable instead of growing
• Reconfigure a project multiple times and verify no new leaked processes appear
• Stop the server and verify all child processes are cleaned up

Made with Cursor