fix(tools): implement dynamic tool output truncation to prevent LLM context overflow

[Adar5]

The Problem

By default, LangChain attempts to inject the raw string output of any executed tool directly into the context window. If a tool returns a massive un-chunked payload (e.g., a large plugin document or a future Jenkins log fetcher), it instantly crashes the LLM provider with a TokenLimitExceeded or OOM error, freezing the chat session.

The Solution

This PR implements a lightweight, dynamically configured defensive wrapper (truncate_tool_output decorator) around the TOOL_REGISTRY.

Key Changes:

Configuration: Added max_tool_output_length to config.yaml (default: 4000), allowing server admins to control memory limits without touching Python code.

Interception: The decorator intercepts all tool executions before they reach the LangChain agent, hard-capping string outputs at the configured limit.

Observability: Added standard server-side logging (logger.warning) so DevOps is notified when a truncation event occurs.

Context Awareness: Appends a clean [SYSTEM WARNING] so the LLM context is explicitly aware that the data was truncated and can inform the user.

Testing

[x] Added test_tool_overflow.py to assert massive 50,000+ character strings are safely truncated based on the config.yaml limit without blocking the event loop.