fix(record): retry unparsable L1 extraction output

[YOMXXX]

Description | 描述

Retry L1 extraction once when the model returns output that cannot be parsed as the required JSON scene array.

This addresses the L1 parse-failure part of seed data loss in #115. Non-OpenAI-compatible models can produce structurally invalid JSON; previously parseExtractionResult() returned [], so the batch ended as extracted=0, stored=0 with no second attempt.

This PR:

• distinguishes parse failures from a valid empty extraction array
• retries exactly once on parse failure with a stricter JSON-only system instruction
• keeps valid [] responses as-is and does not retry them
• reuses the same runner instance on the CleanContextRunner path
• adds focused regression coverage for retry success and valid-empty no-retry behavior

Related Issue | 关联 Issue

Fixes #115

Verification | 验证

• RED: npm test -- src/core/record/l1-extractor-retry.test.ts failed because only l1-extraction ran once
• GREEN: npm test -- src/core/record/l1-extractor-retry.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 🚀