MCP Tool Callback Provider break when mcp server restart with change or stateless

Bug description

With the changes on:
https://github.com/spring-projects/spring-ai/commit/57831042ffa45f43e4dbb2dda3d2494796e6d0e1

Caches were introduced to avoid calling the MCP server. But this will break on below usecases:

• The MCP Server is Stateless -> means no event from the server
• The MCP Server restarted and has tool changes (I believe even if not stateless this also doesnt work)

Environment

Spring AI version 1.1.0

Steps to reproduce

Create a new Spring AI with MCP Client and connect it to STATELESS Spring AI mcp server

• Remove/Add tool at runtime, or
• Just remove the tool and redeploy mcp server

Expected behavior

There are multiple possible solutions:

• cache with ttl
• allow disabling the built-in cache, so the application can implement their own cache
• invalidate cache on error and retry at DefaultToolCallingManager? (probably too overkill)

Workaround

• Create a scheduled invalidate cache to act as ttl

Start by reviewing the cache changes in commit 57831042ffa45f43e4dbb2dda3d2494796e6d0e1 and the DefaultToolCallingManager entry point. Reproduce the issue with a stateless MCP server, then verify how tool changes after restart or redeployment are observed. Done means the chosen cache behavior keeps the available tools current in those scenarios.