fix(l1): drop ungrounded extracted memories

[YOMXXX]

Summary

• Adds a conservative post-LLM validation pass for L1 extraction output.
• Drops extracted memories whose significant tokens have too little overlap with their source messages, preventing obvious hallucinated facts from reaching L1 storage.
• Keeps normal paraphrases/rewrites by using a low overlap threshold and falling back to the whole qualified batch when source IDs are missing.

Why

Issue #82 notes that LLM extraction output currently goes directly to dedup/storage without quality validation. A model can emit plausible but unsupported persona/instruction facts, and those facts then become durable memory. This PR adds a narrow source-grounding gate before dedup/write.

Tests

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

Refs #82

[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.