fix(dedup): cap merged memory priority

[YOMXXX]

Summary

• Adds a deterministic guardrail that caps merged_priority for merge/update decisions at the highest priority among the new memory and targeted candidate memories.
• Updates the conflict-detection prompt so the LLM is told not to raise priority merely because memories were merged.
• Adds regression coverage for the [Bug] L1 指令提取缺少场景限定机制，用户的场景受限偏好被泛化为全局规则 #48 failure mode where moderate instructions can be merged into an over-prioritized global rule.

Why

Issue #48 reports priority amplification as one contributor to scene-limited instructions becoming dominant long-term persona rules. Prompt guidance helps, but this PR enforces the cap after parsing the LLM decision so an over-eager model cannot write a higher priority than the evidence supports.

Tests

• npm test -- src/core/record/l1-dedup.test.ts
• npm test
• npm run build

Refs #48

[Maxwell-Code07]

We have received multiple PRs submitted by you. Thank you for your consistent contributions. We will review them one by one and provide feedback accordingly.