Modularize chat prompts and add PR test workflow

[anth-volk]

Summary

This tracks the current branch work to make the UK chat runtime easier to maintain and safer to test.

Scope

• Move runtime prompt text into a dedicated prompts module.
• Add factual-neutrality and row-level microdata guidance to prompt and tool contracts.
• Clarify exposed tool boundaries around run_python and generate_chart.
• Add backend/frontend PR checks for pull requests against main.
• Clean up backend tests so live model calls and compiled-engine requirements are gated.
• Add model-agnostic AI guidance docs under docs/engineering/skills/ with thin tool-specific adapters.