Refactor the Human-in-the-Loop (HITL) tool-approval logic currently embedded in BasicReActAgent into a dedicated shared module, without changing runtime behavior.

This is a cleanup/architecture issue to improve readability, reuse, and future evolution of tool-level approval policies across agents.

Goals

• Extract HITL tool-approval concerns from BasicReActAgent
• Keep behavior strictly unchanged (no functional regression)
• Make approval rules reusable by other ReAct/LangGraph agents
• Improve code separation (prompting, tool loading, HITL gating, graph wiring)

Scope

Move the following out of agentic_backend/core/agents/basic_react_agent.py into a dedicated module (e.g. agentic_backend/core/interrupts/tool_approval.py):

• tool approval tuning fields (safety.enable_tool_approval, safety.approval_required_tools)
• read-only vs mutating tool prefix classification
• exact tool-name approval matching logic
• HITL approval policy text (FR/EN)
• localized HITL confirmation payload builder (FR/EN)
• parsing helpers for approval configuration

Keep in BasicReActAgent:

• interrupt() call
• build_tool_loop(...) wiring
• graph construction / runtime control flow
• tool assembly (MCP + local tools)

Non-goals

• No UI changes
• No capability model refactor
• No change to HITL behavior or wording (except internal relocation if needed)
• No changes to Mermaid prompt policies / Mermaid UI rendering

Acceptance criteria

• BasicReActAgent still behaves exactly the same for:
  • read-only tools (no HITL)
  • mutating tools (HITL confirmation required when enabled)
  • exact tool-name forced approvals
• Existing localized HITL cards (FR/EN) still render the same
• py_compile passes on refactored files
• BasicReActAgent code is visibly slimmer and more focused
• HITL tool-approval logic can be imported/reused from a shared module

Suggested files

• New: agentic_backend/core/interrupts/tool_approval.py
• Update: agentic_backend/core/agents/basic_react_agent.py

Why this matters

We are progressively structuring the platform around reusable agent capabilities/policies (similar to the Mermaid rendering policies). Extracting HITL approval logic now makes future reuse and refactoring safer, especially before introducing a capability-first model for local + MCP tools.