feat: Evaluate Prompt Orchestration Conflict Resolution

[thinhlpg]

Description

Evaluate and validate that USER instructions correctly override SERVER instructions when conflicts occur

Context

The prompt orchestration system composes prompts from multiple sources:

• SERVER-side (developer-controlled): timing, system prompt, deep research, tools, code assistant, chain-of-thought
• USER-side (user-controlled): project instruction, custom instruction, memory, tone/style

When these conflict (e.g., server says "JSON" but user says "Markdown"), user settings should take priority.

Models to test:

• Jan 30B Instruct
• Jan 30B Thinking

Scope

• Test conflict scenarios: SERVER vs USER instruction priority
• Test toggle behavior: Enable/disable optional server prompts one-by-one
• Validate with LLM-as-judge evaluation
• Document findings and edge cases

Out of Scope

• Changing the actual prompt orchestration implementation
• Production deployment