Spike: instruction resolver with most-specificity rule

[SamMorrowDrums]

Summary

Exploratory spike to evaluate finer-grained instruction control by associating instructions with specific sets of tools rather than whole toolsets.

Design

Core Types

• InstructionRule - Associates an instruction with a set of tool names
• InstructionResolver - Resolves which instructions apply using the most-specificity rule

Specificity Algorithm

1. A rule matches when ALL its tools are present in active tools
2. When Rule A's tools are a proper superset of Rule B's tools (A ⊃ B), A shadows B
3. Partial overlaps (neither is superset) → both rules apply
4. Output is sorted for determinism

Example

rules := []InstructionRule{
    NewInstructionRule("issues-all", "All issues instruction", "get_issue", "list_issues", "create_issue"),
    NewInstructionRule("issues-read", "Read issues instruction", "get_issue", "list_issues"),
    NewInstructionRule("create-only", "Create only instruction", "create_issue"),
}

// Active tools: {get_issue, list_issues, create_issue}
// Result: Only "issues-all" applies (shadows the other two)

// Active tools: {get_issue, list_issues}  
// Result: Only "issues-read" applies

Testing

All tests pass:

• TestInstructionResolver_BasicMatching
• TestInstructionResolver_SupersetShadowing
• TestInstructionResolver_PartialOverlapNoShadowing
• TestInstructionResolver_ComplexHierarchy
• TestInstructionResolver_EmptyRules
• TestInstructionResolver_EmptyActiveTools

Next Steps (if we proceed)

• Integrate with existing GenerateInstructions() or create a V2
• Define actual instruction rules based on tool combinations
• Consider adding priority/weight for equal-size rule conflicts