fix(hermes): split recall context for prefetch

[RerankerGuo]

Description | 描述

Split Gateway recall responses into stable system context and dynamic L1 prefetch context for Hermes.

This fixes the Hermes path where /recall only returned appendSystemContext, so L1 memories recalled by auto-recall were not injected through prefetch(). The Gateway now returns:

• system_context for persona / scene navigation / tool guidance
• prepend_context for dynamic L1 memories
• context as a backward-compatible combined field for older clients

The Hermes provider prefers prepend_context in prefetch() and caches system_context for system_prompt_block(), while still falling back to the legacy context response.

Related Issue | 关联 Issue

Fix #101

Change Type | 修改类型

• Bug fix | Bug 修复
• New feature | 新功能
• Documentation update | 文档更新
• Code optimization | 代码优化

Self-test Checklist | 自测清单

• Verified locally | 本地验证通过
• No existing features affected | 无影响现有功能

Additional Notes | 其他说明

Verified with npx vitest run src/gateway/server.test.ts, npm test, and npm run build using Node v24.15.0.
Also ran python3 -m py_compile for the Hermes provider files. pytest is not installed in this environment, so the Python pytest suite could not be executed here.

[Maxwell-Code07]

Thanks for this PR! Splitting the recall response into stable and dynamic context makes a lot of sense for Hermes integration. We'll review it internally and get back to you shortly.