Rule search that finds the right rules

search_rules used to return nothing for the kind of multi-word request an agent naturally forms from a task, and single words returned hundreds of unranked rules. Both are fixed, in one piece of work.

What changed#

• Synonyms. A curated list expands your terms, so a search for login also finds rules written about sign in and authentication. The list is versioned with the code, and an expanded term always ranks below the word you actually typed.
• Relevance ranking. Results are scored on how much of your query a rule really covers, rather than on matching any single common word.
• A floor, and an honest empty result. Below a confidence threshold the search returns no rules and names the terms that matched nothing, instead of filling the response with whatever scored above zero. An agent can now tell a miss from a hit, which it could not before.
• Scope and limit. The tool takes optional section and limit arguments, and results are capped, with required rules protected from the cap so a broad query cannot drop the mandatory ones.

A task description like "login form with email and password fields" now returns a ranked, capped set you can build from.