fix(prompts): preserve scoped instruction context

[YOMXXX]

Description | 描述

Add conservative L1 prompt guardrails so scene-limited user instructions are preserved as scoped rules instead of being generalized into unconditional global behavior.

This addresses the over-generalization path described in #48 without changing schema or storage:

• L1 extraction now tells the model to keep explicit applicability conditions in content for instruction memories.
• L1 conflict detection now compares instruction scope before merge/update and preserves valid scope qualifiers in merged_content.
• Dedup now warns against automatically raising merged_priority for scoped/diagnostic instructions during merge.
• Focused tests lock these prompt constraints in place.

Related Issue | 关联 Issue

Fixes #48

Verification | 验证

• RED: npm test -- src/core/prompts/l1-scoped-instruction.test.ts failed because guardrails were absent
• GREEN: npm test -- src/core/prompts/l1-scoped-instruction.test.ts
• npm test
• npm run build

[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.