[codex] Harden safe registration diagnostics

[sandking]

What

• Fix safe-core audit semantics for synchronous raw invoker failures so they record RAW_TOOL_FAILED instead of being classified as SAFE_TOOL_FAILED.
• Move Spring AI raw callback identity reflection into a focused helper with tests for original MCP names, fallback behavior, blanks, and throwing accessors.
• Split Spring AI boot starter responsibilities by extracting external raw provider diagnostics and safe callback provider assembly helpers.
• Document the production integration checklist and Spring AI model-facing safe provider wiring guidance.

Why

FastMCP Java is a safe MCP registration/wrapping layer over Spring AI, AgentScope, and the underlying MCP SDKs. The production hardening path should make the safe provider the only model-facing tool path, keep reflection compatibility isolated, and make diagnostics and assembly easier to reason about.

Validation

• git diff --check
• mvn -B test with JDK 17.0.18
• mvn -B -Pexamples test with JDK 17.0.18
• JAVA_HOME=$(/usr/libexec/java_home -v 11) PATH="$JAVA_HOME/bin:$PATH" mvn -B -pl fastmcp-safe-core,fastmcp-safe-config -am test with JDK 11.0.30

Notes

This PR does not make FastMCP Java implement the MCP protocol itself. The documented production check is about verifying that Spring AI / AgentScope / MCP SDK integrations expose only safe virtual tools to the model.