Expose retry method and improve token budgeting

[aandersen2323]

Summary

• expose the FlowDex retry endpoint through the MCP HTTP bridge with stricter method routing
• restructure inference retries so error context is appended to the user turn instead of the system prompt
• add tiktoken-backed token counting with fallbacks, propagate usage data, adjust tool context handling, and cover the utilities with tests while adding the dependency

Testing

• pytest server/tests/test_utils.py

---

https://chatgpt.com/codex/tasks/task_e_68fa86b9540083319f4c4a32550b4d90