feat: discover entity-linked RunbookSet skills and knowledge

[qiansheng91]

Summary

Add entity-driven RunbookSet Skill and Knowledge discovery to the UnifiedModel Query Service, plus a bundled runner that loads one exact inline Skill and applies policy-controlled inline Knowledge without moving Tool authority into UnifiedModel.

Motivation

Agents should discover capabilities from the EntitySet they are operating on through list_method(), then progressively load only the related RunbookSet content. Skill instructions and Knowledge stay model-driven, while Tool schemas, authorization, confirmation, execution, and read-back remain owned by the caller runtime.

Changes

• Query: add list_skills/list_skill and canonical list_knowledge planning, direct visible RunbookLink resolution, exact-ID filtering, detail loading, stable deduplication, and filter_by_entity support.
• Discovery: advertise list_skills and list_knowledge independently from list_method() only when the EntitySet has matching visible content.
• Knowledge: return Assistant-compatible summary/detail rows, complete RunbookLink and RunbookSet context, policy metadata, inline content, and deterministic Knowledge IDs.
• Agent: extend umodel-skill-runner to load only the selected Skill RunbookSet Knowledge, apply always/auto/manual/custom policies, order numeric priority deterministically, reject cross-Runbook and URL/script content, and treat Knowledge as untrusted reference.
• Safety: keep effective Tools as allowed by Skill, available in the runtime, and explicitly user-authorized; runtime definitions remain authoritative for schemas, risk, confirmation, mutation, and read-back.
• Tests/docs: add Query, AgentGateway, MCP E2E, documentation, and committed reproducible runner evaluation evidence in English and Chinese surfaces.

Test plan

• Focused planner, executor, discovery, Skill regression, AgentGateway, and MCP E2E tests
• make guard
• make example-validate: 153/153 examples passed; seven existing schema warnings
• make ci on c130169, including race tests and Go, Python, and Java SDK checks
• Five fresh-context runner repetitions plus edge cases for policy fallbacks, priority, RunbookSet isolation, prompt injection, missing Tool inputs, authorization, confirmation, and mutation read-back
• Independent final review: no remaining Critical, Important, or Minor issues

Compatibility

Additive SPL behavior. list_skill remains an alias normalized to list_skills. list_knowledge is canonical only. Response headers, IDs, localization, optional serialization, and method return types align with the existing UModel Assistant contract. No new public REST endpoint or MCP tool is introduced.

Out of scope

• list_tools, list_toolkits, or server-side Tool execution
• Downloading skill_url or content_url
• Executing embedded scripts or code
• Multi-hop Runbook discovery or global Skill/Knowledge search