fix(core): scope L1 memory search in stores

[Dimon94]

Description | 描述

Make L1 memory search scope filters apply before result truncation:

• accept sessionKey / sessionId in searchMemories() and Gateway /search/memories
• pass scope filters through the L1 store interfaces
• apply SQLite FTS filters in SQL and vector filters before final topK results are returned
• apply TCVDB filters through the query filter expression
• keep a caller-side scope check as a final defense
• add regression coverage for vector and FTS paths where unscoped topK truncation could otherwise hide in-scope memories

This is a general shared-backend correctness fix. It is not tied to any specific downstream host.

Related Issue | 关联 Issue

None.

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

Local verification:

• npm test -- src/core/tools/memory-search.test.ts
• npm test
• npm run build:plugin
• git diff --cached --check before commit

[Maxwell-Code07]

Hi @Dimon94, thanks for this fix!

This is an important correctness improvement for scoped memory search. We'll review the implementation details and get back to you.
Thanks! 🙏

[Dimon94]

Thanks for the quick response.

Agreed. The key correctness point here is that scope filtering should happen before topK truncation, otherwise valid in-scope memories can be hidden by out-of-scope results that happen to rank higher globally.

I tried to keep the fix store-local and defensive: each backend applies the scope as early as its query path allows, and the caller-side check remains as a final guard. I also added regression tests for both vector and FTS paths to cover the truncation-before-filtering failure mode.

I will wait for the implementation review and adjust anything that does not fit the backend conventions.