Add concept of FACTS.md or canned, optimized tool calls

Problem: no matter how much you prompt in the context, agents will re-verify facts anyway. Worse yet, the few times that an agent does take a statement at face value, it may actually be wrong.

Idea for a solution:

• Create the concept of FACTS.md or similar
• This file contains factual statements coupled with verification script that can be automatically evaluated
• When loading FACTS.md into contex, Mecatl automatically evaluates them, and only the facts that actually stand get loaded into context
• The script that was ran, and it's stdout are also loaded into context, as if the agent had itself made the same those tool calls itself
• There is a strict character budget for verifications, and the number of facts is also limited -> incentive to use it only to improve the speed and token economics of flows that the agent would otherwise always evaluate using tool calls

Potential further improvements:

• Facts themselves can have trigger conditions, for instance certain regular expressions matching a tool call or a message -> the agent could receive a canned answer from a more efficient tool call rather than a multi-step expression
• A cheap model could actually classify tool calls automatically or even write rewrite rules automatically
• There could be a way to reject CI when some facts become outdated

The issue names no files, tests, or entry points, so begin by locating how Mecatl loads context and executes tool calls. Review the proposed FACTS.md format, verification output, character and fact limits, and trigger conditions; the work is done when the project has an agreed implementation scope and tests for accepted and rejected facts.