fix(record): persist L1 extraction failure diagnostics

[YOMXXX]

Description | 描述

Persist L1 extraction failure diagnostics when the extraction model returns empty output, content-filter prose, or malformed JSON instead of the expected scene array.

Before this change, parseExtractionResult() returned an empty array for these failures and the pipeline reported extracted=0, leaving operators with no durable evidence of what the model actually returned. This matches the failure mode in #104, where content filtering made L1 appear to silently stop.

This PR:

• records failed L1 extraction parses to .metadata/l1-extraction-failures.jsonl
• includes failure reason/message, session identifiers, model, source message ids, previous scene, raw response length, truncation flag, and a capped raw response
• preserves existing extraction semantics: parse failures still return zero scenes/memories and do not block the pipeline
• adds focused regression coverage using a content-filter style model response

Related Issue | 关联 Issue

Fixes #104

Verification | 验证

• RED: npm test -- src/core/record/l1-extractor-diagnostics.test.ts failed with missing diagnostics file
• GREEN: npm test -- src/core/record/l1-extractor-diagnostics.test.ts
• npm test
• npm run build

[Maxwell-Code07]

@YOMXXX 4 PRs in one day tackling L1 extraction robustness — JSON parsing tolerance, dedup hardening, failure diagnostics persistence, and auto-retry. Great coverage! We'll review them one by one 🚀